Index: projects/mips64-clang/lib/libc/sys/cap_enter.2 =================================================================== --- projects/mips64-clang/lib/libc/sys/cap_enter.2 (revision 305711) +++ projects/mips64-clang/lib/libc/sys/cap_enter.2 (revision 305712) @@ -1,125 +1,130 @@ .\" .\" Copyright (c) 2008-2009 Robert N. M. Watson .\" All rights reserved. .\" .\" This software was developed at the University of Cambridge Computer .\" Laboratory with support from a grant from Google, Inc. .\" .\" Redistribution and use in source and binary forms, with or without .\" modification, are permitted provided that the following conditions .\" are met: .\" 1. Redistributions of source code must retain the above copyright .\" notice, this list of conditions and the following disclaimer. .\" 2. Redistributions in binary form must reproduce the above copyright .\" notice, this list of conditions and the following disclaimer in the .\" documentation and/or other materials provided with the distribution. .\" .\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND .\" ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE .\" IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE .\" ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE .\" FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL .\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS .\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) .\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT .\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY .\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF .\" SUCH DAMAGE. .\" .\" $FreeBSD$ .\" -.Dd March 27, 2014 +.Dd September 10, 2016 .Dt CAP_ENTER 2 .Os .Sh NAME .Nm cap_enter , .Nm cap_getmode .Nd Capability mode system calls .Sh LIBRARY .Lb libc .Sh SYNOPSIS .In sys/capsicum.h .Ft int .Fn cap_enter "void" .Ft int .Fn cap_getmode "u_int *modep" .Sh DESCRIPTION .Fn cap_enter places the current process into capability mode, a mode of execution in which processes may only issue system calls operating on file descriptors or reading limited global system state. Access to global name spaces, such as file system or IPC name spaces, is prevented. If the process is already in a capability mode sandbox, the system call is a no-op. Future process descendants created with .Xr fork 2 or .Xr pdfork 2 will be placed in capability mode from inception. .Pp When combined with .Xr cap_rights_limit 2 , .Xr cap_ioctls_limit 2 , .Xr cap_fcntls_limit 2 , .Fn cap_enter may be used to create kernel-enforced sandboxes in which appropriately-crafted applications or application components may be run. .Pp .Fn cap_getmode returns a flag indicating whether or not the process is in a capability mode sandbox. .Sh CAVEAT Creating effective process sandboxes is a tricky process that involves identifying the least possible rights required by the process and then passing those rights into the process in a safe manner. Consumers of .Fn cap_enter should also be aware of other inherited rights, such as access to VM resources, memory contents, and other process properties that should be considered. It is advisable to use .Xr fexecve 2 to create a runtime environment inside the sandbox that has as few implicitly acquired rights as possible. .Sh RETURN VALUES .Rv -std cap_enter cap_getmode +.Pp +When the process is in capability mode, +.Fn cap_getmode +sets the flag to a non-zero value. +A zero value means the process is not in capability mode. .Sh ERRORS The .Fn cap_enter and .Fn cap_getmode system calls will fail if: .Bl -tag -width Er .It Bq Er ENOSYS The kernel is compiled without: .Pp .Cd "options CAPABILITY_MODE" .El .Pp The .Fn cap_getmode system call may also return the following error: .Bl -tag -width Er .It Bq Er EFAULT Pointer .Fa modep points outside the process's allocated address space. .El .Sh SEE ALSO .Xr cap_fcntls_limit 2 , .Xr cap_ioctls_limit 2 , .Xr cap_rights_limit 2 , .Xr fexecve 2 , .Xr cap_sandboxed 3 , .Xr capsicum 4 .Sh HISTORY Support for capabilities and capabilities mode was developed as part of the .Tn TrustedBSD Project. .Sh AUTHORS These functions and the capability facility were created by .An "Robert N. M. Watson" at the University of Cambridge Computer Laboratory with support from a grant from Google, Inc. Index: projects/mips64-clang/lib/libifconfig/Makefile =================================================================== --- projects/mips64-clang/lib/libifconfig/Makefile (revision 305711) +++ projects/mips64-clang/lib/libifconfig/Makefile (revision 305712) @@ -1,20 +1,21 @@ # $FreeBSD$ PACKAGE= lib${LIB} LIB= ifconfig +PRIVATELIB= true # Don't build shared library, for now. NO_PIC= SHLIBDIR?= /lib SHLIB_MAJOR= 1 SRCS= libifconfig.c libifconfig_internal.c INCSDIR= ${INCLUDEDIR} INCS= libifconfig.h #MAN= libifconfig.3 CFLAGS+= -I${.CURDIR} WARNS?=6 .include Index: projects/mips64-clang/share/man/man9/fpu_kern.9 =================================================================== --- projects/mips64-clang/share/man/man9/fpu_kern.9 (revision 305711) +++ projects/mips64-clang/share/man/man9/fpu_kern.9 (revision 305712) @@ -1,200 +1,210 @@ .\" Copyright (c) 2014 .\" Konstantin Belousov . All rights reserved. .\" .\" Redistribution and use in source and binary forms, with or without .\" modification, are permitted provided that the following conditions .\" are met: .\" 1. Redistributions of source code must retain the above copyright .\" notice, this list of conditions and the following disclaimer. .\" 2. Redistributions in binary form must reproduce the above copyright .\" notice, this list of conditions and the following disclaimer in the .\" documentation and/or other materials provided with the distribution. .\" .\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR .\" IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES .\" OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. .\" IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, .\" INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT .\" NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, .\" DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY .\" THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT .\" (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF .\" THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. .\" .\" $FreeBSD$ .\" .Dd October 23, 2014 .Dt FPU_KERN 9 .Os .Sh NAME .Nm fpu_kern .Nd "facility to use the FPU in the kernel" .Sh SYNOPSIS .Ft struct fpu_kern_ctx * .Fn fpu_kern_alloc_ctx "u_int flags" .Ft void .Fn fpu_kern_free_ctx "struct fpu_kern_ctx *ctx" .Ft int .Fn fpu_kern_enter "struct thread *td" "struct fpu_kern_ctx *ctx" "u_int flags" .Ft int .Fn fpu_kern_leave "struct thread *td" "struct fpu_kern_ctx *ctx" .Ft int .Fn fpu_kern_thread "u_int flags" .Ft int .Fn is_fpu_kern_thread "u_int flags" .Sh DESCRIPTION The .Nm family of functions allows the use of FPU hardware in kernel code. Modern FPUs are not limited to providing hardware implementation for floating point arithmetic; they offer advanced accelerators for cryptography and other computational-intensive algorithms. These facilities share registers with the FPU hardware. .Pp Typical kernel code does not need access to the FPU. Saving a large register file on each entry to the kernel would waste time. When kernel code uses the FPU, the current FPU state must be saved to avoid corrupting the user-mode state, and vice versa. .Pp The management of the save and restore is automatic. The processor catches accesses to the FPU registers when the non-current context tries to access them. Explicit calls are required for the allocation of the save area and the notification of the start and end of the code using the FPU. .Pp The .Fn fpu_kern_alloc_ctx function allocates the memory used by .Nm to track the use of the FPU hardware state and the related software state. The .Fn fpu_kern_alloc_ctx function requires the .Fa flags argument, which currently accepts the following flags: .Bl -tag -width ".Dv FPU_KERN_NOWAIT" -offset indent .It Dv FPU_KERN_NOWAIT Do not wait for the available memory if the request could not be satisfied without sleep. .It 0 No special handling is required. .El .Pp The function returns the allocated context area, or .Va NULL if the allocation failed. .Pp The .Fn fpu_kern_free_ctx function frees the context previously allocated by .Fn fpu_kern_alloc_ctx . .Pp The .Fn fpu_kern_enter function designates the start of the region of kernel code where the use of the FPU is allowed. Its arguments are: .Bl -tag -width ".Fa ctx" -offset indent .It Fa td Currently must be .Va curthread . .It Fa ctx The context save area previously allocated by .Fn fpu_kern_alloc_ctx and not currently in use by another call to .Fn fpu_kern_enter . .It Fa flags This argument currently accepts the following flags: .Bl -tag -width ".Dv FPU_KERN_NORMAL" -offset indent .It Dv FPU_KERN_NORMAL Indicates that the caller intends to access the full FPU state. Must be specified currently. .It Dv FPU_KERN_KTHR Indicates that no saving of the current FPU state should be performed, if the thread called .Xr fpu_kern_thread 9 function. This is intended to minimize code duplication in callers which could be used from both kernel thread and syscall contexts. The .Fn fpu_kern_leave function correctly handles such contexts. +.It Dv FPU_KERN_NOCTX +Avoid nesting save area. +If the flag is specified, the +.Fa ctx +must be passed as +.Va NULL . +The flag should only be used for really short code blocks +which can be executed in a critical section. +It avoids the need to allocate the FPU context by the cost +of increased system latency. .El .El .Pp The function does not sleep or block. It could cause the .Nm Device Not Available exception during execution, and on the first FPU access after the function returns, as well as after each context switch (see Intel Software Developer Manual for the reference). Currently, no errors are defined which can be returned by .Fn fpu_kern_enter to the caller. .Pp The .Fn fpu_kern_leave function ends the region started by .Fn fpu_kern_enter . The uses of FPU in the kernel after the call to .Fn fpu_kern_leave are erroneous until the next call to .Fn fpu_kern_enter is performed. The function takes the .Fa td thread argument, which currently must be .Va curthread , and the .Fa ctx context pointer, previously passed to .Fn fpu_kern_enter . After the function returns, the context may be freed or reused by other invocation of .Fn fpu_kern_enter . There are no errors defined for the function, it always returns 0. .Pp The .Fn fpu_kern_thread function enables an optimization for threads which never leave to the usermode. The current thread will reuse the usermode save area for the kernel FPU state instead of requiring an explicitly allocated context. There are no flags defined for the function, and no error states that the function returns. Once this function has been called, neither .Fn fpu_kern_enter nor .Fn fpu_kern_leave is required to be called and the fpu is available for use in the calling thread. .Pp The .Fn is_fpu_kern_thread function returns the boolean indicating whether the current thread entered the mode enabled by .Fn fpu_kern_thread . There is currently no flags defined for the function, the return value is true if the current thread have the permanent FPU save area, and false otherwise. .Sh NOTES The .Nm is currently implemented only for the i386 and amd64 architectures. .Pp There is no way to handle floating point exceptions raised from kernel mode. .Pp The unused .Fa flags arguments to the .Nm functions are to be extended to allow specification of the set of the FPU hardware state used by the code region. This would allow optimizations of saving and restoring the state. .Sh AUTHORS The .Nm facitily and this manual page were written by .An Konstantin Belousov Aq Mt kib@FreeBSD.org . Index: projects/mips64-clang/share/mk/bsd.subdir.mk =================================================================== --- projects/mips64-clang/share/mk/bsd.subdir.mk (revision 305711) +++ projects/mips64-clang/share/mk/bsd.subdir.mk (revision 305712) @@ -1,193 +1,194 @@ # from: @(#)bsd.subdir.mk 5.9 (Berkeley) 2/1/91 # $FreeBSD$ # # The include file contains the default targets # for building subdirectories. # # For all of the directories listed in the variable SUBDIRS, the # specified directory will be visited and the target made. There is # also a default target which allows the command "make subdir" where # subdir is any directory listed in the variable SUBDIRS. # # # +++ variables +++ # # DISTRIBUTION Name of distribution. [base] # # SUBDIR A list of subdirectories that should be built as well. # Each of the targets will execute the same target in the # subdirectories. SUBDIR.yes is automatically appended # to this list. # # +++ targets +++ # # distribute: # This is a variant of install, which will # put the stuff into the right "distribution". # # See SUBDIR_TARGETS for list of targets that will recurse. # # Targets defined in STANDALONE_SUBDIR_TARGETS will always be ran # with SUBDIR_PARALLEL and will not respect .WAIT or SUBDIR_DEPEND_ # values. # # SUBDIR_TARGETS and STANDALONE_SUBDIR_TARGETS can be appended to # via make.conf or src.conf. # .if !target(____) ____: SUBDIR_TARGETS+= \ all all-man analyze buildconfig buildfiles buildincludes \ checkdpadd clean cleandepend cleandir cleanilinks \ cleanobj depend distribute files includes installconfig \ installfiles installincludes print-dir realinstall lint \ maninstall manlint obj objlink tags \ # Described above. STANDALONE_SUBDIR_TARGETS+= \ all-man buildconfig buildfiles buildincludes check checkdpadd \ clean cleandepend cleandir cleanilinks cleanobj files includes \ installconfig installincludes installfiles print-dir \ maninstall manlint obj objlink # It is safe to install in parallel when staging. .if defined(NO_ROOT) STANDALONE_SUBDIR_TARGETS+= realinstall .endif .include .if make(print-dir) NEED_SUBDIR= 1 ECHODIR= : .SILENT: .if ${RELDIR:U.} != "." print-dir: .PHONY @echo ${RELDIR} .endif .endif .if !defined(NEED_SUBDIR) .if ${.MAKE.LEVEL} == 0 && ${MK_DIRDEPS_BUILD} == "yes" && !empty(SUBDIR) && !(make(clean*) || make(destroy*)) .include # ignore this _SUBDIR: .endif .endif DISTRIBUTION?= base .if !target(distribute) distribute: .MAKE .for dist in ${DISTRIBUTION} ${_+_}cd ${.CURDIR}; \ ${MAKE} install installconfig -DNO_SUBDIR DESTDIR=${DISTDIR}/${dist} SHARED=copies .endfor .endif # Convenience targets to run 'build${target}' and 'install${target}' when # calling 'make ${target}'. .for __target in files includes .if !target(${__target}) ${__target}: build${__target} install${__target} .ORDER: build${__target} install${__target} .endif .endfor # Make 'install' supports a before and after target. Actual install # hooks are placed in 'realinstall'. .if !target(install) .for __stage in before real after .if !target(${__stage}install) ${__stage}install: .endif .endfor install: beforeinstall realinstall afterinstall .ORDER: beforeinstall realinstall afterinstall .endif .ORDER: all install # SUBDIR recursing may be disabled for MK_DIRDEPS_BUILD .if !target(_SUBDIR) .if defined(SUBDIR) SUBDIR:=${SUBDIR} ${SUBDIR.yes} SUBDIR:=${SUBDIR:u} .endif # Subdir code shared among 'make ', 'make ' and SUBDIR_PARALLEL. _SUBDIR_SH= \ if test -d ${.CURDIR}/$${dir}.${MACHINE_ARCH}; then \ dir=$${dir}.${MACHINE_ARCH}; \ fi; \ ${ECHODIR} "===> ${DIRPRFX}$${dir} ($${target})"; \ cd ${.CURDIR}/$${dir}; \ ${MAKE} $${target} DIRPRFX=${DIRPRFX}$${dir}/ # This is kept for compatibility only. The normal handling of attaching to # SUBDIR_TARGETS will create a target for each directory. _SUBDIR: .USEBEFORE .if defined(SUBDIR) && !empty(SUBDIR) && !defined(NO_SUBDIR) @${_+_}target=${.TARGET:realinstall=install}; \ for dir in ${SUBDIR:N.WAIT}; do ( ${_SUBDIR_SH} ); done .endif # Create 'make subdir' targets to run the real 'all' target. .for __dir in ${SUBDIR:N.WAIT} ${__dir}: all_subdir_${DIRPRFX}${__dir} .PHONY .endfor .for __target in ${SUBDIR_TARGETS} # Can ordering be skipped for this and SUBDIR_PARALLEL forced? .if ${STANDALONE_SUBDIR_TARGETS:M${__target}} _is_standalone_target= 1 -SUBDIR:= ${SUBDIR:N.WAIT} +_subdir_filter= N.WAIT .else _is_standalone_target= 0 +_subdir_filter= .endif __subdir_targets= -.for __dir in ${SUBDIR} +.for __dir in ${SUBDIR:${_subdir_filter}} .if ${__dir} == .WAIT __subdir_targets+= .WAIT .else __deps= .if ${_is_standalone_target} == 0 .if defined(SUBDIR_PARALLEL) # Apply SUBDIR_DEPEND dependencies for SUBDIR_PARALLEL. .for __dep in ${SUBDIR_DEPEND_${__dir}} __deps+= ${__target}_subdir_${DIRPRFX}${__dep} .endfor .else # For non-parallel builds, directories depend on all targets before them. __deps:= ${__subdir_targets} .endif # defined(SUBDIR_PARALLEL) .endif # ${_is_standalone_target} == 0 ${__target}_subdir_${DIRPRFX}${__dir}: .PHONY .MAKE .SILENT ${__deps} @${_+_}target=${__target:realinstall=install}; \ dir=${__dir}; \ ${_SUBDIR_SH}; __subdir_targets+= ${__target}_subdir_${DIRPRFX}${__dir} .endif # ${__dir} == .WAIT .endfor # __dir in ${SUBDIR} # Attach the subdir targets to the real target. # Only recurse on directly-called targets. I.e., don't recurse on dependencies # such as 'install' becoming {before,real,after}install, just recurse # 'install'. Despite that, 'realinstall' is special due to ordering issues # with 'afterinstall'. .if !defined(NO_SUBDIR) && (make(${__target}) || \ (${__target} == realinstall && make(install))) ${__target}: ${__subdir_targets} .PHONY .endif # make(${__target}) .endfor # __target in ${SUBDIR_TARGETS} .endif # !target(_SUBDIR) # Ensure all targets exist .for __target in ${SUBDIR_TARGETS} .if !target(${__target}) ${__target}: .endif .endfor .endif Index: projects/mips64-clang/sys/amd64/amd64/fpu.c =================================================================== --- projects/mips64-clang/sys/amd64/amd64/fpu.c (revision 305711) +++ projects/mips64-clang/sys/amd64/amd64/fpu.c (revision 305712) @@ -1,1066 +1,1109 @@ /*- * Copyright (c) 1990 William Jolitz. * Copyright (c) 1991 The Regents of the University of California. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)npx.c 7.2 (Berkeley) 5/12/91 */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Floating point support. */ #if defined(__GNUCLIKE_ASM) && !defined(lint) #define fldcw(cw) __asm __volatile("fldcw %0" : : "m" (cw)) #define fnclex() __asm __volatile("fnclex") #define fninit() __asm __volatile("fninit") #define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr))) #define fnstsw(addr) __asm __volatile("fnstsw %0" : "=am" (*(addr))) #define fxrstor(addr) __asm __volatile("fxrstor %0" : : "m" (*(addr))) #define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr))) #define ldmxcsr(csr) __asm __volatile("ldmxcsr %0" : : "m" (csr)) #define stmxcsr(addr) __asm __volatile("stmxcsr %0" : : "m" (*(addr))) static __inline void xrstor(char *addr, uint64_t mask) { uint32_t low, hi; low = mask; hi = mask >> 32; __asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi)); } static __inline void xsave(char *addr, uint64_t mask) { uint32_t low, hi; low = mask; hi = mask >> 32; __asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) : "memory"); } #else /* !(__GNUCLIKE_ASM && !lint) */ void fldcw(u_short cw); void fnclex(void); void fninit(void); void fnstcw(caddr_t addr); void fnstsw(caddr_t addr); void fxsave(caddr_t addr); void fxrstor(caddr_t addr); void ldmxcsr(u_int csr); void stmxcsr(u_int *csr); void xrstor(char *addr, uint64_t mask); void xsave(char *addr, uint64_t mask); #endif /* __GNUCLIKE_ASM && !lint */ #define start_emulating() load_cr0(rcr0() | CR0_TS) #define stop_emulating() clts() CTASSERT(sizeof(struct savefpu) == 512); CTASSERT(sizeof(struct xstate_hdr) == 64); CTASSERT(sizeof(struct savefpu_ymm) == 832); /* * This requirement is to make it easier for asm code to calculate * offset of the fpu save area from the pcb address. FPU save area * must be 64-byte aligned. */ CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0); /* * Ensure the copy of XCR0 saved in a core is contained in the padding * area. */ CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savefpu, sv_pad) && X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savefpu)); static void fpu_clean_state(void); SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, 1, "Floating point instructions executed in hardware"); int use_xsave; /* non-static for cpu_switch.S */ uint64_t xsave_mask; /* the same */ static uma_zone_t fpu_save_area_zone; static struct savefpu *fpu_initialstate; struct xsave_area_elm_descr { u_int offset; u_int size; } *xsave_area_desc; void fpusave(void *addr) { if (use_xsave) xsave((char *)addr, xsave_mask); else fxsave((char *)addr); } void fpurestore(void *addr) { if (use_xsave) xrstor((char *)addr, xsave_mask); else fxrstor((char *)addr); } void fpususpend(void *addr) { u_long cr0; cr0 = rcr0(); stop_emulating(); fpusave(addr); load_cr0(cr0); } void fpuresume(void *addr) { u_long cr0; cr0 = rcr0(); stop_emulating(); fninit(); if (use_xsave) load_xcr(XCR0, xsave_mask); fpurestore(addr); load_cr0(cr0); } /* * Enable XSAVE if supported and allowed by user. * Calculate the xsave_mask. */ static void fpuinit_bsp1(void) { u_int cp[4]; uint64_t xsave_mask_user; if ((cpu_feature2 & CPUID2_XSAVE) != 0) { use_xsave = 1; TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave); } if (!use_xsave) return; cpuid_count(0xd, 0x0, cp); xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE; if ((cp[0] & xsave_mask) != xsave_mask) panic("CPU0 does not support X87 or SSE: %x", cp[0]); xsave_mask = ((uint64_t)cp[3] << 32) | cp[0]; xsave_mask_user = xsave_mask; TUNABLE_ULONG_FETCH("hw.xsave_mask", &xsave_mask_user); xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE; xsave_mask &= xsave_mask_user; if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512) xsave_mask &= ~XFEATURE_AVX512; if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX) xsave_mask &= ~XFEATURE_MPX; cpuid_count(0xd, 0x1, cp); if ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0) { /* * Patch the XSAVE instruction in the cpu_switch code * to XSAVEOPT. We assume that XSAVE encoding used * REX byte, and set the bit 4 of the r/m byte. */ ctx_switch_xsave[3] |= 0x10; } } /* * Calculate the fpu save area size. */ static void fpuinit_bsp2(void) { u_int cp[4]; if (use_xsave) { cpuid_count(0xd, 0x0, cp); cpu_max_ext_state_size = cp[1]; /* * Reload the cpu_feature2, since we enabled OSXSAVE. */ do_cpuid(1, cp); cpu_feature2 = cp[2]; } else cpu_max_ext_state_size = sizeof(struct savefpu); } /* * Initialize the floating point unit. */ void fpuinit(void) { register_t saveintr; u_int mxcsr; u_short control; if (IS_BSP()) fpuinit_bsp1(); if (use_xsave) { load_cr4(rcr4() | CR4_XSAVE); load_xcr(XCR0, xsave_mask); } /* * XCR0 shall be set up before CPU can report the save area size. */ if (IS_BSP()) fpuinit_bsp2(); /* * It is too early for critical_enter() to work on AP. */ saveintr = intr_disable(); stop_emulating(); fninit(); control = __INITIAL_FPUCW__; fldcw(control); mxcsr = __INITIAL_MXCSR__; ldmxcsr(mxcsr); start_emulating(); intr_restore(saveintr); } /* * On the boot CPU we generate a clean state that is used to * initialize the floating point unit when it is first used by a * process. */ static void fpuinitstate(void *arg __unused) { register_t saveintr; int cp[4], i, max_ext_n; fpu_initialstate = malloc(cpu_max_ext_state_size, M_DEVBUF, M_WAITOK | M_ZERO); saveintr = intr_disable(); stop_emulating(); fpusave(fpu_initialstate); if (fpu_initialstate->sv_env.en_mxcsr_mask) cpu_mxcsr_mask = fpu_initialstate->sv_env.en_mxcsr_mask; else cpu_mxcsr_mask = 0xFFBF; /* * The fninit instruction does not modify XMM registers or x87 * registers (MM/ST). The fpusave call dumped the garbage * contained in the registers after reset to the initial state * saved. Clear XMM and x87 registers file image to make the * startup program state and signal handler XMM/x87 register * content predictable. */ bzero(fpu_initialstate->sv_fp, sizeof(fpu_initialstate->sv_fp)); bzero(fpu_initialstate->sv_xmm, sizeof(fpu_initialstate->sv_xmm)); /* * Create a table describing the layout of the CPU Extended * Save Area. */ if (use_xsave) { max_ext_n = flsl(xsave_mask); xsave_area_desc = malloc(max_ext_n * sizeof(struct xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO); /* x87 state */ xsave_area_desc[0].offset = 0; xsave_area_desc[0].size = 160; /* XMM */ xsave_area_desc[1].offset = 160; xsave_area_desc[1].size = 288 - 160; for (i = 2; i < max_ext_n; i++) { cpuid_count(0xd, i, cp); xsave_area_desc[i].offset = cp[1]; xsave_area_desc[i].size = cp[0]; } } fpu_save_area_zone = uma_zcreate("FPU_save_area", cpu_max_ext_state_size, NULL, NULL, NULL, NULL, XSAVE_AREA_ALIGN - 1, 0); start_emulating(); intr_restore(saveintr); } SYSINIT(fpuinitstate, SI_SUB_DRIVERS, SI_ORDER_ANY, fpuinitstate, NULL); /* * Free coprocessor (if we have it). */ void fpuexit(struct thread *td) { critical_enter(); if (curthread == PCPU_GET(fpcurthread)) { stop_emulating(); fpusave(curpcb->pcb_save); start_emulating(); PCPU_SET(fpcurthread, NULL); } critical_exit(); } int fpuformat(void) { return (_MC_FPFMT_XMM); } /* * The following mechanism is used to ensure that the FPE_... value * that is passed as a trapcode to the signal handler of the user * process does not have more than one bit set. * * Multiple bits may be set if the user process modifies the control * word while a status word bit is already set. While this is a sign * of bad coding, we have no choise than to narrow them down to one * bit, since we must not send a trapcode that is not exactly one of * the FPE_ macros. * * The mechanism has a static table with 127 entries. Each combination * of the 7 FPU status word exception bits directly translates to a * position in this table, where a single FPE_... value is stored. * This FPE_... value stored there is considered the "most important" * of the exception bits and will be sent as the signal code. The * precedence of the bits is based upon Intel Document "Numerical * Applications", Chapter "Special Computational Situations". * * The macro to choose one of these values does these steps: 1) Throw * away status word bits that cannot be masked. 2) Throw away the bits * currently masked in the control word, assuming the user isn't * interested in them anymore. 3) Reinsert status word bit 7 (stack * fault) if it is set, which cannot be masked but must be presered. * 4) Use the remaining bits to point into the trapcode table. * * The 6 maskable bits in order of their preference, as stated in the * above referenced Intel manual: * 1 Invalid operation (FP_X_INV) * 1a Stack underflow * 1b Stack overflow * 1c Operand of unsupported format * 1d SNaN operand. * 2 QNaN operand (not an exception, irrelavant here) * 3 Any other invalid-operation not mentioned above or zero divide * (FP_X_INV, FP_X_DZ) * 4 Denormal operand (FP_X_DNML) * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL) * 6 Inexact result (FP_X_IMP) */ static char fpetable[128] = { 0, FPE_FLTINV, /* 1 - INV */ FPE_FLTUND, /* 2 - DNML */ FPE_FLTINV, /* 3 - INV | DNML */ FPE_FLTDIV, /* 4 - DZ */ FPE_FLTINV, /* 5 - INV | DZ */ FPE_FLTDIV, /* 6 - DNML | DZ */ FPE_FLTINV, /* 7 - INV | DNML | DZ */ FPE_FLTOVF, /* 8 - OFL */ FPE_FLTINV, /* 9 - INV | OFL */ FPE_FLTUND, /* A - DNML | OFL */ FPE_FLTINV, /* B - INV | DNML | OFL */ FPE_FLTDIV, /* C - DZ | OFL */ FPE_FLTINV, /* D - INV | DZ | OFL */ FPE_FLTDIV, /* E - DNML | DZ | OFL */ FPE_FLTINV, /* F - INV | DNML | DZ | OFL */ FPE_FLTUND, /* 10 - UFL */ FPE_FLTINV, /* 11 - INV | UFL */ FPE_FLTUND, /* 12 - DNML | UFL */ FPE_FLTINV, /* 13 - INV | DNML | UFL */ FPE_FLTDIV, /* 14 - DZ | UFL */ FPE_FLTINV, /* 15 - INV | DZ | UFL */ FPE_FLTDIV, /* 16 - DNML | DZ | UFL */ FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */ FPE_FLTOVF, /* 18 - OFL | UFL */ FPE_FLTINV, /* 19 - INV | OFL | UFL */ FPE_FLTUND, /* 1A - DNML | OFL | UFL */ FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */ FPE_FLTDIV, /* 1C - DZ | OFL | UFL */ FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */ FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */ FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */ FPE_FLTRES, /* 20 - IMP */ FPE_FLTINV, /* 21 - INV | IMP */ FPE_FLTUND, /* 22 - DNML | IMP */ FPE_FLTINV, /* 23 - INV | DNML | IMP */ FPE_FLTDIV, /* 24 - DZ | IMP */ FPE_FLTINV, /* 25 - INV | DZ | IMP */ FPE_FLTDIV, /* 26 - DNML | DZ | IMP */ FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */ FPE_FLTOVF, /* 28 - OFL | IMP */ FPE_FLTINV, /* 29 - INV | OFL | IMP */ FPE_FLTUND, /* 2A - DNML | OFL | IMP */ FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */ FPE_FLTDIV, /* 2C - DZ | OFL | IMP */ FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */ FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */ FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */ FPE_FLTUND, /* 30 - UFL | IMP */ FPE_FLTINV, /* 31 - INV | UFL | IMP */ FPE_FLTUND, /* 32 - DNML | UFL | IMP */ FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */ FPE_FLTDIV, /* 34 - DZ | UFL | IMP */ FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */ FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */ FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */ FPE_FLTOVF, /* 38 - OFL | UFL | IMP */ FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */ FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */ FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */ FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */ FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */ FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */ FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */ FPE_FLTSUB, /* 40 - STK */ FPE_FLTSUB, /* 41 - INV | STK */ FPE_FLTUND, /* 42 - DNML | STK */ FPE_FLTSUB, /* 43 - INV | DNML | STK */ FPE_FLTDIV, /* 44 - DZ | STK */ FPE_FLTSUB, /* 45 - INV | DZ | STK */ FPE_FLTDIV, /* 46 - DNML | DZ | STK */ FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */ FPE_FLTOVF, /* 48 - OFL | STK */ FPE_FLTSUB, /* 49 - INV | OFL | STK */ FPE_FLTUND, /* 4A - DNML | OFL | STK */ FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */ FPE_FLTDIV, /* 4C - DZ | OFL | STK */ FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */ FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */ FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */ FPE_FLTUND, /* 50 - UFL | STK */ FPE_FLTSUB, /* 51 - INV | UFL | STK */ FPE_FLTUND, /* 52 - DNML | UFL | STK */ FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */ FPE_FLTDIV, /* 54 - DZ | UFL | STK */ FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */ FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */ FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */ FPE_FLTOVF, /* 58 - OFL | UFL | STK */ FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */ FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */ FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */ FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */ FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */ FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */ FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */ FPE_FLTRES, /* 60 - IMP | STK */ FPE_FLTSUB, /* 61 - INV | IMP | STK */ FPE_FLTUND, /* 62 - DNML | IMP | STK */ FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */ FPE_FLTDIV, /* 64 - DZ | IMP | STK */ FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */ FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */ FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */ FPE_FLTOVF, /* 68 - OFL | IMP | STK */ FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */ FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */ FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */ FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */ FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */ FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */ FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */ FPE_FLTUND, /* 70 - UFL | IMP | STK */ FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */ FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */ FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */ FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */ FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */ FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */ FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */ FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */ FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */ FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */ FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */ FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */ FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */ FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */ FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */ }; /* * Read the FP status and control words, then generate si_code value * for SIGFPE. The error code chosen will be one of the * FPE_... macros. It will be sent as the second argument to old * BSD-style signal handlers and as "siginfo_t->si_code" (second * argument) to SA_SIGINFO signal handlers. * * Some time ago, we cleared the x87 exceptions with FNCLEX there. * Clearing exceptions was necessary mainly to avoid IRQ13 bugs. The * usermode code which understands the FPU hardware enough to enable * the exceptions, can also handle clearing the exception state in the * handler. The only consequence of not clearing the exception is the * rethrow of the SIGFPE on return from the signal handler and * reexecution of the corresponding instruction. * * For XMM traps, the exceptions were never cleared. */ int fputrap_x87(void) { struct savefpu *pcb_save; u_short control, status; critical_enter(); /* * Interrupt handling (for another interrupt) may have pushed the * state to memory. Fetch the relevant parts of the state from * wherever they are. */ if (PCPU_GET(fpcurthread) != curthread) { pcb_save = curpcb->pcb_save; control = pcb_save->sv_env.en_cw; status = pcb_save->sv_env.en_sw; } else { fnstcw(&control); fnstsw(&status); } critical_exit(); return (fpetable[status & ((~control & 0x3f) | 0x40)]); } int fputrap_sse(void) { u_int mxcsr; critical_enter(); if (PCPU_GET(fpcurthread) != curthread) mxcsr = curpcb->pcb_save->sv_env.en_mxcsr; else stmxcsr(&mxcsr); critical_exit(); return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]); } /* * Device Not Available (DNA, #NM) exception handler. * * It would be better to switch FP context here (if curthread != * fpcurthread) and not necessarily for every context switch, but it * is too hard to access foreign pcb's. */ void fpudna(void) { /* * This handler is entered with interrupts enabled, so context * switches may occur before critical_enter() is executed. If * a context switch occurs, then when we regain control, our * state will have been completely restored. The CPU may * change underneath us, but the only part of our context that * lives in the CPU is CR0.TS and that will be "restored" by * setting it on the new CPU. */ critical_enter(); + KASSERT((curpcb->pcb_flags & PCB_FPUNOSAVE) == 0, + ("fpudna while in fpu_kern_enter(FPU_KERN_NOCTX)")); if (PCPU_GET(fpcurthread) == curthread) { printf("fpudna: fpcurthread == curthread\n"); stop_emulating(); critical_exit(); return; } if (PCPU_GET(fpcurthread) != NULL) { panic("fpudna: fpcurthread = %p (%d), curthread = %p (%d)\n", PCPU_GET(fpcurthread), PCPU_GET(fpcurthread)->td_tid, curthread, curthread->td_tid); } stop_emulating(); /* * Record new context early in case frstor causes a trap. */ PCPU_SET(fpcurthread, curthread); fpu_clean_state(); if ((curpcb->pcb_flags & PCB_FPUINITDONE) == 0) { /* * This is the first time this thread has used the FPU or * the PCB doesn't contain a clean FPU state. Explicitly * load an initial state. * * We prefer to restore the state from the actual save * area in PCB instead of directly loading from * fpu_initialstate, to ignite the XSAVEOPT * tracking engine. */ bcopy(fpu_initialstate, curpcb->pcb_save, cpu_max_ext_state_size); fpurestore(curpcb->pcb_save); if (curpcb->pcb_initial_fpucw != __INITIAL_FPUCW__) fldcw(curpcb->pcb_initial_fpucw); if (PCB_USER_FPU(curpcb)) set_pcb_flags(curpcb, PCB_FPUINITDONE | PCB_USERFPUINITDONE); else set_pcb_flags(curpcb, PCB_FPUINITDONE); } else fpurestore(curpcb->pcb_save); critical_exit(); } void fpudrop(void) { struct thread *td; td = PCPU_GET(fpcurthread); KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread")); CRITICAL_ASSERT(td); PCPU_SET(fpcurthread, NULL); clear_pcb_flags(td->td_pcb, PCB_FPUINITDONE); start_emulating(); } /* * Get the user state of the FPU into pcb->pcb_user_save without * dropping ownership (if possible). It returns the FPU ownership * status. */ int fpugetregs(struct thread *td) { struct pcb *pcb; uint64_t *xstate_bv, bit; char *sa; int max_ext_n, i, owned; pcb = td->td_pcb; if ((pcb->pcb_flags & PCB_USERFPUINITDONE) == 0) { bcopy(fpu_initialstate, get_pcb_user_save_pcb(pcb), cpu_max_ext_state_size); get_pcb_user_save_pcb(pcb)->sv_env.en_cw = pcb->pcb_initial_fpucw; fpuuserinited(td); return (_MC_FPOWNED_PCB); } critical_enter(); if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) { fpusave(get_pcb_user_save_pcb(pcb)); owned = _MC_FPOWNED_FPU; } else { owned = _MC_FPOWNED_PCB; } critical_exit(); if (use_xsave) { /* * Handle partially saved state. */ sa = (char *)get_pcb_user_save_pcb(pcb); xstate_bv = (uint64_t *)(sa + sizeof(struct savefpu) + offsetof(struct xstate_hdr, xstate_bv)); max_ext_n = flsl(xsave_mask); for (i = 0; i < max_ext_n; i++) { bit = 1ULL << i; if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0) continue; bcopy((char *)fpu_initialstate + xsave_area_desc[i].offset, sa + xsave_area_desc[i].offset, xsave_area_desc[i].size); *xstate_bv |= bit; } } return (owned); } void fpuuserinited(struct thread *td) { struct pcb *pcb; pcb = td->td_pcb; if (PCB_USER_FPU(pcb)) set_pcb_flags(pcb, PCB_FPUINITDONE | PCB_USERFPUINITDONE); else set_pcb_flags(pcb, PCB_FPUINITDONE); } int fpusetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size) { struct xstate_hdr *hdr, *ehdr; size_t len, max_len; uint64_t bv; /* XXXKIB should we clear all extended state in xstate_bv instead ? */ if (xfpustate == NULL) return (0); if (!use_xsave) return (EOPNOTSUPP); len = xfpustate_size; if (len < sizeof(struct xstate_hdr)) return (EINVAL); max_len = cpu_max_ext_state_size - sizeof(struct savefpu); if (len > max_len) return (EINVAL); ehdr = (struct xstate_hdr *)xfpustate; bv = ehdr->xstate_bv; /* * Avoid #gp. */ if (bv & ~xsave_mask) return (EINVAL); hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1); hdr->xstate_bv = bv; bcopy(xfpustate + sizeof(struct xstate_hdr), (char *)(hdr + 1), len - sizeof(struct xstate_hdr)); return (0); } /* * Set the state of the FPU. */ int fpusetregs(struct thread *td, struct savefpu *addr, char *xfpustate, size_t xfpustate_size) { struct pcb *pcb; int error; pcb = td->td_pcb; critical_enter(); if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) { error = fpusetxstate(td, xfpustate, xfpustate_size); if (error != 0) { critical_exit(); return (error); } bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr)); fpurestore(get_pcb_user_save_td(td)); critical_exit(); set_pcb_flags(pcb, PCB_FPUINITDONE | PCB_USERFPUINITDONE); } else { critical_exit(); error = fpusetxstate(td, xfpustate, xfpustate_size); if (error != 0) return (error); bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr)); fpuuserinited(td); } return (0); } /* * On AuthenticAMD processors, the fxrstor instruction does not restore * the x87's stored last instruction pointer, last data pointer, and last * opcode values, except in the rare case in which the exception summary * (ES) bit in the x87 status word is set to 1. * * In order to avoid leaking this information across processes, we clean * these values by performing a dummy load before executing fxrstor(). */ static void fpu_clean_state(void) { static float dummy_variable = 0.0; u_short status; /* * Clear the ES bit in the x87 status word if it is currently * set, in order to avoid causing a fault in the upcoming load. */ fnstsw(&status); if (status & 0x80) fnclex(); /* * Load the dummy variable into the x87 stack. This mangles * the x87 stack, but we don't care since we're about to call * fxrstor() anyway. */ __asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable)); } /* * This really sucks. We want the acpi version only, but it requires * the isa_if.h file in order to get the definitions. */ #include "opt_isa.h" #ifdef DEV_ISA #include /* * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI. */ static struct isa_pnp_id fpupnp_ids[] = { { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */ { 0 } }; static int fpupnp_probe(device_t dev) { int result; result = ISA_PNP_PROBE(device_get_parent(dev), dev, fpupnp_ids); if (result <= 0) device_quiet(dev); return (result); } static int fpupnp_attach(device_t dev) { return (0); } static device_method_t fpupnp_methods[] = { /* Device interface */ DEVMETHOD(device_probe, fpupnp_probe), DEVMETHOD(device_attach, fpupnp_attach), DEVMETHOD(device_detach, bus_generic_detach), DEVMETHOD(device_shutdown, bus_generic_shutdown), DEVMETHOD(device_suspend, bus_generic_suspend), DEVMETHOD(device_resume, bus_generic_resume), { 0, 0 } }; static driver_t fpupnp_driver = { "fpupnp", fpupnp_methods, 1, /* no softc */ }; static devclass_t fpupnp_devclass; DRIVER_MODULE(fpupnp, acpi, fpupnp_driver, fpupnp_devclass, 0, 0); #endif /* DEV_ISA */ static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx", "Kernel contexts for FPU state"); #define FPU_KERN_CTX_FPUINITDONE 0x01 #define FPU_KERN_CTX_DUMMY 0x02 /* avoided save for the kern thread */ #define FPU_KERN_CTX_INUSE 0x04 struct fpu_kern_ctx { struct savefpu *prev; uint32_t flags; char hwstate1[]; }; struct fpu_kern_ctx * fpu_kern_alloc_ctx(u_int flags) { struct fpu_kern_ctx *res; size_t sz; sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN + cpu_max_ext_state_size; res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ? M_NOWAIT : M_WAITOK) | M_ZERO); return (res); } void fpu_kern_free_ctx(struct fpu_kern_ctx *ctx) { KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx")); /* XXXKIB clear the memory ? */ free(ctx, M_FPUKERN_CTX); } static struct savefpu * fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx) { vm_offset_t p; p = (vm_offset_t)&ctx->hwstate1; p = roundup2(p, XSAVE_AREA_ALIGN); return ((struct savefpu *)p); } int fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags) { struct pcb *pcb; - KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("using inuse ctx")); + pcb = td->td_pcb; + KASSERT((flags & FPU_KERN_NOCTX) != 0 || ctx != NULL, + ("ctx is required when !FPU_KERN_NOCTX")); + KASSERT(ctx == NULL || (ctx->flags & FPU_KERN_CTX_INUSE) == 0, + ("using inuse ctx")); + KASSERT((pcb->pcb_flags & PCB_FPUNOSAVE) == 0, + ("recursive fpu_kern_enter while in PCB_FPUNOSAVE state")); + if ((flags & FPU_KERN_NOCTX) != 0) { + critical_enter(); + stop_emulating(); + if (curthread == PCPU_GET(fpcurthread)) { + fpusave(curpcb->pcb_save); + PCPU_SET(fpcurthread, NULL); + } else { + KASSERT(PCPU_GET(fpcurthread) == NULL, + ("invalid fpcurthread")); + } + + /* + * This breaks XSAVEOPT tracker, but + * PCB_FPUNOSAVE state is supposed to never need to + * save FPU context at all. + */ + fpurestore(fpu_initialstate); + set_pcb_flags(pcb, PCB_KERNFPU | PCB_FPUNOSAVE | + PCB_FPUINITDONE); + return (0); + } if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) { ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE; return (0); } - pcb = td->td_pcb; KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save == get_pcb_user_save_pcb(pcb), ("mangled pcb_save")); ctx->flags = FPU_KERN_CTX_INUSE; if ((pcb->pcb_flags & PCB_FPUINITDONE) != 0) ctx->flags |= FPU_KERN_CTX_FPUINITDONE; fpuexit(td); ctx->prev = pcb->pcb_save; pcb->pcb_save = fpu_kern_ctx_savefpu(ctx); set_pcb_flags(pcb, PCB_KERNFPU); clear_pcb_flags(pcb, PCB_FPUINITDONE); return (0); } int fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx) { struct pcb *pcb; - KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0, - ("leaving not inuse ctx")); - ctx->flags &= ~FPU_KERN_CTX_INUSE; - - if (is_fpu_kern_thread(0) && (ctx->flags & FPU_KERN_CTX_DUMMY) != 0) - return (0); - KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0, ("dummy ctx")); pcb = td->td_pcb; - critical_enter(); - if (curthread == PCPU_GET(fpcurthread)) - fpudrop(); - critical_exit(); - pcb->pcb_save = ctx->prev; + + if ((pcb->pcb_flags & PCB_FPUNOSAVE) != 0) { + KASSERT(ctx == NULL, ("non-null ctx after FPU_KERN_NOCTX")); + KASSERT(PCPU_GET(fpcurthread) == NULL, + ("non-NULL fpcurthread for PCB_FPUNOSAVE")); + CRITICAL_ASSERT(td); + + clear_pcb_flags(pcb, PCB_FPUNOSAVE | PCB_FPUINITDONE); + start_emulating(); + critical_exit(); + } else { + KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0, + ("leaving not inuse ctx")); + ctx->flags &= ~FPU_KERN_CTX_INUSE; + + if (is_fpu_kern_thread(0) && + (ctx->flags & FPU_KERN_CTX_DUMMY) != 0) + return (0); + KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0, + ("dummy ctx")); + critical_enter(); + if (curthread == PCPU_GET(fpcurthread)) + fpudrop(); + critical_exit(); + pcb->pcb_save = ctx->prev; + } + if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) { if ((pcb->pcb_flags & PCB_USERFPUINITDONE) != 0) { set_pcb_flags(pcb, PCB_FPUINITDONE); clear_pcb_flags(pcb, PCB_KERNFPU); } else clear_pcb_flags(pcb, PCB_FPUINITDONE | PCB_KERNFPU); } else { if ((ctx->flags & FPU_KERN_CTX_FPUINITDONE) != 0) set_pcb_flags(pcb, PCB_FPUINITDONE); else clear_pcb_flags(pcb, PCB_FPUINITDONE); KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave")); } return (0); } int fpu_kern_thread(u_int flags) { KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0, ("Only kthread may use fpu_kern_thread")); KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb), ("mangled pcb_save")); KASSERT(PCB_USER_FPU(curpcb), ("recursive call")); set_pcb_flags(curpcb, PCB_KERNFPU); return (0); } int is_fpu_kern_thread(u_int flags) { if ((curthread->td_pflags & TDP_KTHREAD) == 0) return (0); return ((curpcb->pcb_flags & PCB_KERNFPU) != 0); } /* * FPU save area alloc/free/init utility routines */ struct savefpu * fpu_save_area_alloc(void) { return (uma_zalloc(fpu_save_area_zone, 0)); } void fpu_save_area_free(struct savefpu *fsa) { uma_zfree(fpu_save_area_zone, fsa); } void fpu_save_area_reset(struct savefpu *fsa) { bcopy(fpu_initialstate, fsa, cpu_max_ext_state_size); } Index: projects/mips64-clang/sys/amd64/amd64/pmap.c =================================================================== --- projects/mips64-clang/sys/amd64/amd64/pmap.c (revision 305711) +++ projects/mips64-clang/sys/amd64/amd64/pmap.c (revision 305712) @@ -1,7274 +1,7268 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2003 Peter Wemm * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #define AMD64_NPT_AWARE #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_pmap.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #endif static __inline boolean_t pmap_type_guest(pmap_t pmap) { return ((pmap->pm_type == PT_EPT) || (pmap->pm_type == PT_RVI)); } static __inline boolean_t pmap_emulate_ad_bits(pmap_t pmap) { return ((pmap->pm_flags & PMAP_EMULATE_AD_BITS) != 0); } static __inline pt_entry_t pmap_valid_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_V; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_EMUL_V; else mask = EPT_PG_READ; break; default: panic("pmap_valid_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_rw_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_RW; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_EMUL_RW; else mask = EPT_PG_WRITE; break; default: panic("pmap_rw_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_global_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: mask = X86_PG_G; break; case PT_RVI: case PT_EPT: mask = 0; break; default: panic("pmap_global_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_accessed_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_A; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_READ; else mask = EPT_PG_A; break; default: panic("pmap_accessed_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_modified_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_M; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_WRITE; else mask = EPT_PG_M; break; default: panic("pmap_modified_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } extern struct pcpu __pcpu[]; #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define pa_to_pvh(pa) (&pv_table[pa_index(pa)]) #define NPV_LIST_LOCKS MAXCPU #define PHYS_TO_PV_LIST_LOCK(pa) \ (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS]) #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \ struct rwlock **_lockp = (lockp); \ struct rwlock *_new_lock; \ \ _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \ if (_new_lock != *_lockp) { \ if (*_lockp != NULL) \ rw_wunlock(*_lockp); \ *_lockp = _new_lock; \ rw_wlock(*_lockp); \ } \ } while (0) #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \ CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m)) #define RELEASE_PV_LIST_LOCK(lockp) do { \ struct rwlock **_lockp = (lockp); \ \ if (*_lockp != NULL) { \ rw_wunlock(*_lockp); \ *_lockp = NULL; \ } \ } while (0) #define VM_PAGE_TO_PV_LIST_LOCK(m) \ PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m)) struct pmap kernel_pmap_store; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ int nkpt; SYSCTL_INT(_machdep, OID_AUTO, nkpt, CTLFLAG_RD, &nkpt, 0, "Number of kernel page table pages allocated on bootup"); static int ndmpdp; vm_paddr_t dmaplimit; vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS; pt_entry_t pg_nx; static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); static int pat_works = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pat_works, CTLFLAG_RD, &pat_works, 1, "Is page attribute table fully functional?"); static int pg_ps_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pg_ps_enabled, 0, "Are large page mappings enabled?"); #define PAT_INDEX_SIZE 8 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */ static u_int64_t KPTphys; /* phys addr of kernel level 1 */ static u_int64_t KPDphys; /* phys addr of kernel level 2 */ u_int64_t KPDPphys; /* phys addr of kernel level 3 */ u_int64_t KPML4phys; /* phys addr of kernel level 4 */ static u_int64_t DMPDphys; /* phys addr of direct mapped level 2 */ static u_int64_t DMPDPphys; /* phys addr of direct mapped level 3 */ static int ndmpdpphys; /* number of DMPDPphys pages */ /* * pmap_mapdev support pre initialization (i.e. console) */ #define PMAP_PREINIT_MAPPING_COUNT 8 static struct pmap_preinit_mapping { vm_paddr_t pa; vm_offset_t va; vm_size_t sz; int mode; } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT]; static int pmap_initialized; /* * Data for the pv entry allocation mechanism. * Updates to pv_invl_gen are protected by the pv_list_locks[] * elements, but reads are not. */ static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks); static struct mtx pv_chunks_mutex; static struct rwlock pv_list_locks[NPV_LIST_LOCKS]; static u_long pv_invl_gen[NPV_LIST_LOCKS]; static struct md_page *pv_table; static struct md_page pv_dummy; /* * All those kernel PT submaps that BSD is so fond of */ pt_entry_t *CMAP1 = 0; caddr_t CADDR1 = 0; static vm_offset_t qframe = 0; static struct mtx qframe_mtx; static int pmap_flags = PMAP_PDE_SUPERPAGE; /* flags for x86 pmaps */ int pmap_pcid_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pcid_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pmap_pcid_enabled, 0, "Is TLB Context ID enabled ?"); int invpcid_works = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, invpcid_works, CTLFLAG_RD, &invpcid_works, 0, "Is the invpcid instruction available ?"); static int pmap_pcid_save_cnt_proc(SYSCTL_HANDLER_ARGS) { int i; uint64_t res; res = 0; CPU_FOREACH(i) { res += cpuid_to_pcpu[i]->pc_pm_save_cnt; } return (sysctl_handle_64(oidp, &res, 0, req)); } SYSCTL_PROC(_vm_pmap, OID_AUTO, pcid_save_cnt, CTLTYPE_U64 | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, pmap_pcid_save_cnt_proc, "QU", "Count of saved TLB context on switch"); static LIST_HEAD(, pmap_invl_gen) pmap_invl_gen_tracker = LIST_HEAD_INITIALIZER(&pmap_invl_gen_tracker); static struct mtx invl_gen_mtx; static u_long pmap_invl_gen = 0; /* Fake lock object to satisfy turnstiles interface. */ static struct lock_object invl_gen_ts = { .lo_name = "invlts", }; #define PMAP_ASSERT_NOT_IN_DI() \ KASSERT(curthread->td_md.md_invl_gen.gen == 0, ("DI already started")) /* * Start a new Delayed Invalidation (DI) block of code, executed by * the current thread. Within a DI block, the current thread may * destroy both the page table and PV list entries for a mapping and * then release the corresponding PV list lock before ensuring that * the mapping is flushed from the TLBs of any processors with the * pmap active. */ static void pmap_delayed_invl_started(void) { struct pmap_invl_gen *invl_gen; u_long currgen; invl_gen = &curthread->td_md.md_invl_gen; PMAP_ASSERT_NOT_IN_DI(); mtx_lock(&invl_gen_mtx); if (LIST_EMPTY(&pmap_invl_gen_tracker)) currgen = pmap_invl_gen; else currgen = LIST_FIRST(&pmap_invl_gen_tracker)->gen; invl_gen->gen = currgen + 1; LIST_INSERT_HEAD(&pmap_invl_gen_tracker, invl_gen, link); mtx_unlock(&invl_gen_mtx); } /* * Finish the DI block, previously started by the current thread. All * required TLB flushes for the pages marked by * pmap_delayed_invl_page() must be finished before this function is * called. * * This function works by bumping the global DI generation number to * the generation number of the current thread's DI, unless there is a * pending DI that started earlier. In the latter case, bumping the * global DI generation number would incorrectly signal that the * earlier DI had finished. Instead, this function bumps the earlier * DI's generation number to match the generation number of the * current thread's DI. */ static void pmap_delayed_invl_finished(void) { struct pmap_invl_gen *invl_gen, *next; struct turnstile *ts; invl_gen = &curthread->td_md.md_invl_gen; KASSERT(invl_gen->gen != 0, ("missed invl_started")); mtx_lock(&invl_gen_mtx); next = LIST_NEXT(invl_gen, link); if (next == NULL) { turnstile_chain_lock(&invl_gen_ts); ts = turnstile_lookup(&invl_gen_ts); pmap_invl_gen = invl_gen->gen; if (ts != NULL) { turnstile_broadcast(ts, TS_SHARED_QUEUE); turnstile_unpend(ts, TS_SHARED_LOCK); } turnstile_chain_unlock(&invl_gen_ts); } else { next->gen = invl_gen->gen; } LIST_REMOVE(invl_gen, link); mtx_unlock(&invl_gen_mtx); invl_gen->gen = 0; } #ifdef PV_STATS static long invl_wait; SYSCTL_LONG(_vm_pmap, OID_AUTO, invl_wait, CTLFLAG_RD, &invl_wait, 0, "Number of times DI invalidation blocked pmap_remove_all/write"); #endif static u_long * pmap_delayed_invl_genp(vm_page_t m) { return (&pv_invl_gen[pa_index(VM_PAGE_TO_PHYS(m)) % NPV_LIST_LOCKS]); } /* * Ensure that all currently executing DI blocks, that need to flush * TLB for the given page m, actually flushed the TLB at the time the * function returned. If the page m has an empty PV list and we call * pmap_delayed_invl_wait(), upon its return we know that no CPU has a * valid mapping for the page m in either its page table or TLB. * * This function works by blocking until the global DI generation * number catches up with the generation number associated with the * given page m and its PV list. Since this function's callers * typically own an object lock and sometimes own a page lock, it * cannot sleep. Instead, it blocks on a turnstile to relinquish the * processor. */ static void pmap_delayed_invl_wait(vm_page_t m) { struct thread *td; struct turnstile *ts; u_long *m_gen; #ifdef PV_STATS bool accounted = false; #endif td = curthread; m_gen = pmap_delayed_invl_genp(m); while (*m_gen > pmap_invl_gen) { #ifdef PV_STATS if (!accounted) { atomic_add_long(&invl_wait, 1); accounted = true; } #endif ts = turnstile_trywait(&invl_gen_ts); if (*m_gen > pmap_invl_gen) turnstile_wait(ts, NULL, TS_SHARED_QUEUE); else turnstile_cancel(ts); } } /* * Mark the page m's PV list as participating in the current thread's * DI block. Any threads concurrently using m's PV list to remove or * restrict all mappings to m will wait for the current thread's DI * block to complete before proceeding. * * The function works by setting the DI generation number for m's PV * list to at least the DI generation number of the current thread. * This forces a caller of pmap_delayed_invl_wait() to block until * current thread calls pmap_delayed_invl_finished(). */ static void pmap_delayed_invl_page(vm_page_t m) { u_long gen, *m_gen; rw_assert(VM_PAGE_TO_PV_LIST_LOCK(m), RA_WLOCKED); gen = curthread->td_md.md_invl_gen.gen; if (gen == 0) return; m_gen = pmap_delayed_invl_genp(m); if (*m_gen < gen) *m_gen = gen; } /* * Crashdump maps. */ static caddr_t crashdumpmap; static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp); static int popcnt_pc_map_pq(uint64_t *map); static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp); static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp); static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); static boolean_t pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static int pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode); static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp); static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va); static boolean_t pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); static vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va); static void pmap_pde_attr(pd_entry_t *pde, int cache_bits, int mask); static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp); static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot); static void pmap_pte_attr(pt_entry_t *pte, int cache_bits, int mask); static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free, struct rwlock **lockp); static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp); static void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte); static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, struct spglist *free); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp); static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde); static void pmap_update_pde_invalidate(pmap_t, vm_offset_t va, pd_entry_t pde); static vm_page_t _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp); static vm_page_t pmap_allocpde(pmap_t pmap, vm_offset_t va, struct rwlock **lockp); static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp); static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free); static int pmap_unuse_pt(pmap_t, vm_offset_t, pd_entry_t, struct spglist *); static vm_offset_t pmap_kmem_choose(vm_offset_t addr); /* * Move the kernel virtual free pointer to the next * 2MB. This is used to help improve performance * by using a large (2MB) page for much of the kernel * (.text, .data, .bss) */ static vm_offset_t pmap_kmem_choose(vm_offset_t addr) { vm_offset_t newaddr = addr; newaddr = roundup2(addr, NBPDR); return (newaddr); } /********************/ /* Inline functions */ /********************/ /* Return a non-clipped PD index for a given VA */ static __inline vm_pindex_t pmap_pde_pindex(vm_offset_t va) { return (va >> PDRSHIFT); } /* Return various clipped indexes for a given VA */ static __inline vm_pindex_t pmap_pte_index(vm_offset_t va) { return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1)); } static __inline vm_pindex_t pmap_pde_index(vm_offset_t va) { return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1)); } static __inline vm_pindex_t pmap_pdpe_index(vm_offset_t va) { return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1)); } static __inline vm_pindex_t pmap_pml4e_index(vm_offset_t va) { return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1)); } /* Return a pointer to the PML4 slot that corresponds to a VA */ static __inline pml4_entry_t * pmap_pml4e(pmap_t pmap, vm_offset_t va) { return (&pmap->pm_pml4[pmap_pml4e_index(va)]); } /* Return a pointer to the PDP slot that corresponds to a VA */ static __inline pdp_entry_t * pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va) { pdp_entry_t *pdpe; pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME); return (&pdpe[pmap_pdpe_index(va)]); } /* Return a pointer to the PDP slot that corresponds to a VA */ static __inline pdp_entry_t * pmap_pdpe(pmap_t pmap, vm_offset_t va) { pml4_entry_t *pml4e; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pml4e = pmap_pml4e(pmap, va); if ((*pml4e & PG_V) == 0) return (NULL); return (pmap_pml4e_to_pdpe(pml4e, va)); } /* Return a pointer to the PD slot that corresponds to a VA */ static __inline pd_entry_t * pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va) { pd_entry_t *pde; pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME); return (&pde[pmap_pde_index(va)]); } /* Return a pointer to the PD slot that corresponds to a VA */ static __inline pd_entry_t * pmap_pde(pmap_t pmap, vm_offset_t va) { pdp_entry_t *pdpe; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pdpe = pmap_pdpe(pmap, va); if (pdpe == NULL || (*pdpe & PG_V) == 0) return (NULL); return (pmap_pdpe_to_pde(pdpe, va)); } /* Return a pointer to the PT slot that corresponds to a VA */ static __inline pt_entry_t * pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va) { pt_entry_t *pte; pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); return (&pte[pmap_pte_index(va)]); } /* Return a pointer to the PT slot that corresponds to a VA */ static __inline pt_entry_t * pmap_pte(pmap_t pmap, vm_offset_t va) { pd_entry_t *pde; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pde = pmap_pde(pmap, va); if (pde == NULL || (*pde & PG_V) == 0) return (NULL); if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */ return ((pt_entry_t *)pde); return (pmap_pde_to_pte(pde, va)); } static __inline void pmap_resident_count_inc(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); pmap->pm_stats.resident_count += count; } static __inline void pmap_resident_count_dec(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(pmap->pm_stats.resident_count >= count, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count -= count; } PMAP_INLINE pt_entry_t * vtopte(vm_offset_t va) { u_int64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopte on a uva/gpa 0x%0lx", va)); return (PTmap + ((va >> PAGE_SHIFT) & mask)); } static __inline pd_entry_t * vtopde(vm_offset_t va) { u_int64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopde on a uva/gpa 0x%0lx", va)); return (PDmap + ((va >> PDRSHIFT) & mask)); } static u_int64_t allocpages(vm_paddr_t *firstaddr, int n) { u_int64_t ret; ret = *firstaddr; bzero((void *)ret, n * PAGE_SIZE); *firstaddr += n * PAGE_SIZE; return (ret); } CTASSERT(powerof2(NDMPML4E)); /* number of kernel PDP slots */ #define NKPDPE(ptpgs) howmany(ptpgs, NPDEPG) static void nkpt_init(vm_paddr_t addr) { int pt_pages; #ifdef NKPT pt_pages = NKPT; #else pt_pages = howmany(addr, 1 << PDRSHIFT); pt_pages += NKPDPE(pt_pages); /* * Add some slop beyond the bare minimum required for bootstrapping * the kernel. * * This is quite important when allocating KVA for kernel modules. * The modules are required to be linked in the negative 2GB of * the address space. If we run out of KVA in this region then * pmap_growkernel() will need to allocate page table pages to map * the entire 512GB of KVA space which is an unnecessary tax on * physical memory. * * Secondly, device memory mapped as part of setting up the low- * level console(s) is taken from KVA, starting at virtual_avail. * This is because cninit() is called after pmap_bootstrap() but * before vm_init() and pmap_init(). 20MB for a frame buffer is * not uncommon. */ pt_pages += 32; /* 64MB additional slop. */ #endif nkpt = pt_pages; } static void create_pagetables(vm_paddr_t *firstaddr) { int i, j, ndm1g, nkpdpe; pt_entry_t *pt_p; pd_entry_t *pd_p; pdp_entry_t *pdp_p; pml4_entry_t *p4_p; /* Allocate page table pages for the direct map */ ndmpdp = howmany(ptoa(Maxmem), NBPDP); if (ndmpdp < 4) /* Minimum 4GB of dirmap */ ndmpdp = 4; ndmpdpphys = howmany(ndmpdp, NPDPEPG); if (ndmpdpphys > NDMPML4E) { /* * Each NDMPML4E allows 512 GB, so limit to that, * and then readjust ndmpdp and ndmpdpphys. */ printf("NDMPML4E limits system to %d GB\n", NDMPML4E * 512); Maxmem = atop(NDMPML4E * NBPML4); ndmpdpphys = NDMPML4E; ndmpdp = NDMPML4E * NPDEPG; } DMPDPphys = allocpages(firstaddr, ndmpdpphys); ndm1g = 0; if ((amd_feature & AMDID_PAGE1GB) != 0) ndm1g = ptoa(Maxmem) >> PDPSHIFT; if (ndm1g < ndmpdp) DMPDphys = allocpages(firstaddr, ndmpdp - ndm1g); dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT; /* Allocate pages */ KPML4phys = allocpages(firstaddr, 1); KPDPphys = allocpages(firstaddr, NKPML4E); /* * Allocate the initial number of kernel page table pages required to * bootstrap. We defer this until after all memory-size dependent * allocations are done (e.g. direct map), so that we don't have to * build in too much slop in our estimate. * * Note that when NKPML4E > 1, we have an empty page underneath * all but the KPML4I'th one, so we need NKPML4E-1 extra (zeroed) * pages. (pmap_enter requires a PD page to exist for each KPML4E.) */ nkpt_init(*firstaddr); nkpdpe = NKPDPE(nkpt); KPTphys = allocpages(firstaddr, nkpt); KPDphys = allocpages(firstaddr, nkpdpe); /* Fill in the underlying page table pages */ /* Nominally read-only (but really R/W) from zero to physfree */ /* XXX not fully used, underneath 2M pages */ pt_p = (pt_entry_t *)KPTphys; for (i = 0; ptoa(i) < *firstaddr; i++) pt_p[i] = ptoa(i) | X86_PG_RW | X86_PG_V | X86_PG_G; /* Now map the page tables at their location within PTmap */ pd_p = (pd_entry_t *)KPDphys; for (i = 0; i < nkpt; i++) pd_p[i] = (KPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V; /* Map from zero to end of allocations under 2M pages */ /* This replaces some of the KPTphys entries above */ for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) pd_p[i] = (i << PDRSHIFT) | X86_PG_RW | X86_PG_V | PG_PS | X86_PG_G; /* And connect up the PD to the PDP (leaving room for L4 pages) */ pdp_p = (pdp_entry_t *)(KPDPphys + ptoa(KPML4I - KPML4BASE)); for (i = 0; i < nkpdpe; i++) pdp_p[i + KPDPI] = (KPDphys + ptoa(i)) | X86_PG_RW | X86_PG_V | PG_U; /* * Now, set up the direct map region using 2MB and/or 1GB pages. If * the end of physical memory is not aligned to a 1GB page boundary, * then the residual physical memory is mapped with 2MB pages. Later, * if pmap_mapdev{_attr}() uses the direct map for non-write-back * memory, pmap_change_attr() will demote any 2MB or 1GB page mappings * that are partially used. */ pd_p = (pd_entry_t *)DMPDphys; for (i = NPDEPG * ndm1g, j = 0; i < NPDEPG * ndmpdp; i++, j++) { pd_p[j] = (vm_paddr_t)i << PDRSHIFT; /* Preset PG_M and PG_A because demotion expects it. */ pd_p[j] |= X86_PG_RW | X86_PG_V | PG_PS | X86_PG_G | X86_PG_M | X86_PG_A; } pdp_p = (pdp_entry_t *)DMPDPphys; for (i = 0; i < ndm1g; i++) { pdp_p[i] = (vm_paddr_t)i << PDPSHIFT; /* Preset PG_M and PG_A because demotion expects it. */ pdp_p[i] |= X86_PG_RW | X86_PG_V | PG_PS | X86_PG_G | X86_PG_M | X86_PG_A; } for (j = 0; i < ndmpdp; i++, j++) { pdp_p[i] = DMPDphys + ptoa(j); pdp_p[i] |= X86_PG_RW | X86_PG_V | PG_U; } /* And recursively map PML4 to itself in order to get PTmap */ p4_p = (pml4_entry_t *)KPML4phys; p4_p[PML4PML4I] = KPML4phys; p4_p[PML4PML4I] |= X86_PG_RW | X86_PG_V | PG_U; /* Connect the Direct Map slot(s) up to the PML4. */ for (i = 0; i < ndmpdpphys; i++) { p4_p[DMPML4I + i] = DMPDPphys + ptoa(i); p4_p[DMPML4I + i] |= X86_PG_RW | X86_PG_V | PG_U; } /* Connect the KVA slots up to the PML4 */ for (i = 0; i < NKPML4E; i++) { p4_p[KPML4BASE + i] = KPDPphys + ptoa(i); p4_p[KPML4BASE + i] |= X86_PG_RW | X86_PG_V | PG_U; } } /* * Bootstrap the system enough to run with virtual memory. * * On amd64 this is called after mapping has already been enabled * and just syncs the pmap module with what has already been done. * [We can't call it easily with mapping off since the kernel is not * mapped with PA == VA, hence we would have to relocate every address * from the linked base (virtual) address "KERNBASE" to the actual * (physical) address starting relative to 0] */ void pmap_bootstrap(vm_paddr_t *firstaddr) { vm_offset_t va; pt_entry_t *pte; int i; /* * Create an initial set of page tables to run the kernel in. */ create_pagetables(firstaddr); /* * Add a physical memory segment (vm_phys_seg) corresponding to the * preallocated kernel page table pages so that vm_page structures * representing these pages will be created. The vm_page structures * are required for promotion of the corresponding kernel virtual * addresses to superpage mappings. */ vm_phys_add_seg(KPTphys, KPTphys + ptoa(nkpt)); virtual_avail = (vm_offset_t) KERNBASE + *firstaddr; virtual_avail = pmap_kmem_choose(virtual_avail); virtual_end = VM_MAX_KERNEL_ADDRESS; /* XXX do %cr0 as well */ load_cr4(rcr4() | CR4_PGE); load_cr3(KPML4phys); if (cpu_stdext_feature & CPUID_STDEXT_SMEP) load_cr4(rcr4() | CR4_SMEP); /* * Initialize the kernel pmap (which is statically allocated). */ PMAP_LOCK_INIT(kernel_pmap); kernel_pmap->pm_pml4 = (pdp_entry_t *)PHYS_TO_DMAP(KPML4phys); kernel_pmap->pm_cr3 = KPML4phys; CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */ TAILQ_INIT(&kernel_pmap->pm_pvchunk); kernel_pmap->pm_flags = pmap_flags; /* * Initialize the TLB invalidations generation number lock. */ mtx_init(&invl_gen_mtx, "invlgn", NULL, MTX_DEF); /* * Reserve some special page table entries/VA space for temporary * mapping of pages. */ #define SYSMAP(c, p, v, n) \ v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); va = virtual_avail; pte = vtopte(va); /* * Crashdump maps. The first page is reused as CMAP1 for the * memory test. */ SYSMAP(caddr_t, CMAP1, crashdumpmap, MAXDUMPPGS) CADDR1 = crashdumpmap; virtual_avail = va; /* Initialize the PAT MSR. */ pmap_init_pat(); /* Initialize TLB Context Id. */ TUNABLE_INT_FETCH("vm.pmap.pcid_enabled", &pmap_pcid_enabled); if ((cpu_feature2 & CPUID2_PCID) != 0 && pmap_pcid_enabled) { /* Check for INVPCID support */ invpcid_works = (cpu_stdext_feature & CPUID_STDEXT_INVPCID) != 0; for (i = 0; i < MAXCPU; i++) { kernel_pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN; kernel_pmap->pm_pcids[i].pm_gen = 1; } __pcpu[0].pc_pcid_next = PMAP_PCID_KERN + 1; __pcpu[0].pc_pcid_gen = 1; /* * pcpu area for APs is zeroed during AP startup. * pc_pcid_next and pc_pcid_gen are initialized by AP * during pcpu setup. */ load_cr4(rcr4() | CR4_PCIDE); } else { pmap_pcid_enabled = 0; } } /* * Setup the PAT MSR. */ void pmap_init_pat(void) { int pat_table[PAT_INDEX_SIZE]; uint64_t pat_msr; u_long cr0, cr4; int i; /* Bail if this CPU doesn't implement PAT. */ if ((cpu_feature & CPUID_PAT) == 0) panic("no PAT??"); /* Set default PAT index table. */ for (i = 0; i < PAT_INDEX_SIZE; i++) pat_table[i] = -1; pat_table[PAT_WRITE_BACK] = 0; pat_table[PAT_WRITE_THROUGH] = 1; pat_table[PAT_UNCACHEABLE] = 3; pat_table[PAT_WRITE_COMBINING] = 3; pat_table[PAT_WRITE_PROTECTED] = 3; pat_table[PAT_UNCACHED] = 3; /* Initialize default PAT entries. */ pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | PAT_VALUE(1, PAT_WRITE_THROUGH) | PAT_VALUE(2, PAT_UNCACHED) | PAT_VALUE(3, PAT_UNCACHEABLE) | PAT_VALUE(4, PAT_WRITE_BACK) | PAT_VALUE(5, PAT_WRITE_THROUGH) | PAT_VALUE(6, PAT_UNCACHED) | PAT_VALUE(7, PAT_UNCACHEABLE); if (pat_works) { /* * Leave the indices 0-3 at the default of WB, WT, UC-, and UC. * Program 5 and 6 as WP and WC. * Leave 4 and 7 as WB and UC. */ pat_msr &= ~(PAT_MASK(5) | PAT_MASK(6)); pat_msr |= PAT_VALUE(5, PAT_WRITE_PROTECTED) | PAT_VALUE(6, PAT_WRITE_COMBINING); pat_table[PAT_UNCACHED] = 2; pat_table[PAT_WRITE_PROTECTED] = 5; pat_table[PAT_WRITE_COMBINING] = 6; } else { /* * Just replace PAT Index 2 with WC instead of UC-. */ pat_msr &= ~PAT_MASK(2); pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING); pat_table[PAT_WRITE_COMBINING] = 2; } /* Disable PGE. */ cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); /* Disable caches (CD = 1, NW = 0). */ cr0 = rcr0(); load_cr0((cr0 & ~CR0_NW) | CR0_CD); /* Flushes caches and TLBs. */ wbinvd(); invltlb(); /* Update PAT and index table. */ wrmsr(MSR_PAT, pat_msr); for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = pat_table[i]; /* Flush caches and TLBs again. */ wbinvd(); invltlb(); /* Restore caches and PGE. */ load_cr0(cr0); load_cr4(cr4); } /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pat_mode = PAT_WRITE_BACK; } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { struct pmap_preinit_mapping *ppim; vm_page_t mpte; vm_size_t s; int error, i, pv_npg; /* * Initialize the vm page array entries for the kernel pmap's * page table pages. */ for (i = 0; i < nkpt; i++) { mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT)); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_init: page table page is out of range")); mpte->pindex = pmap_pde_pindex(KERNBASE) + i; mpte->phys_addr = KPTphys + (i << PAGE_SHIFT); } /* * If the kernel is running on a virtual machine, then it must assume * that MCA is enabled by the hypervisor. Moreover, the kernel must * be prepared for the hypervisor changing the vendor and family that * are reported by CPUID. Consequently, the workaround for AMD Family * 10h Erratum 383 is enabled if the processor's feature set does not * include at least one feature that is only supported by older Intel * or newer AMD processors. */ if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 && (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI | CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP | AMDID2_FMA4)) == 0) workaround_erratum383 = 1; /* * Are large page mappings enabled? */ TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled); if (pg_ps_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = NBPDR; } /* * Initialize the pv chunk list mutex. */ mtx_init(&pv_chunks_mutex, "pmap pv chunk list", NULL, MTX_DEF); /* * Initialize the pool of pv list locks. */ for (i = 0; i < NPV_LIST_LOCKS; i++) rw_init(&pv_list_locks[i], "pmap pv list"); /* * Calculate the size of the pv head table for superpages. */ pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, NBPDR); /* * Allocate memory for the pv head table for superpages. */ s = (vm_size_t)(pv_npg * sizeof(struct md_page)); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(kernel_arena, s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); TAILQ_INIT(&pv_dummy.pv_list); pmap_initialized = 1; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) continue; /* Make the direct map consistent */ if (ppim->pa < dmaplimit && ppim->pa + ppim->sz < dmaplimit) { (void)pmap_change_attr(PHYS_TO_DMAP(ppim->pa), ppim->sz, ppim->mode); } if (!bootverbose) continue; printf("PPIM %u: PA=%#lx, VA=%#lx, size=%#lx, mode=%#x\n", i, ppim->pa, ppim->va, ppim->sz, ppim->mode); } mtx_init(&qframe_mtx, "qfrmlk", NULL, MTX_SPIN); error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, (vmem_addr_t *)&qframe); if (error != 0) panic("qframe allocation failed"); } static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD, 0, "2MB page mapping counters"); static u_long pmap_pde_demotions; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pde_demotions, 0, "2MB page demotions"); static u_long pmap_pde_mappings; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD, &pmap_pde_mappings, 0, "2MB page mappings"); static u_long pmap_pde_p_failures; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_pde_p_failures, 0, "2MB page promotion failures"); static u_long pmap_pde_promotions; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD, &pmap_pde_promotions, 0, "2MB page promotions"); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pdpe, CTLFLAG_RD, 0, "1GB page mapping counters"); static u_long pmap_pdpe_demotions; SYSCTL_ULONG(_vm_pmap_pdpe, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pdpe_demotions, 0, "1GB page demotions"); /*************************************************** * Low level helper routines..... ***************************************************/ static pt_entry_t pmap_swap_pat(pmap_t pmap, pt_entry_t entry) { int x86_pat_bits = X86_PG_PTE_PAT | X86_PG_PDE_PAT; switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* Verify that both PAT bits are not set at the same time */ KASSERT((entry & x86_pat_bits) != x86_pat_bits, ("Invalid PAT bits in entry %#lx", entry)); /* Swap the PAT bits if one of them is set */ if ((entry & x86_pat_bits) != 0) entry ^= x86_pat_bits; break; case PT_EPT: /* * Nothing to do - the memory attributes are represented * the same way for regular pages and superpages. */ break; default: panic("pmap_switch_pat_bits: bad pm_type %d", pmap->pm_type); } return (entry); } /* * Determine the appropriate bits to set in a PTE or PDE for a specified * caching mode. */ static int pmap_cache_bits(pmap_t pmap, int mode, boolean_t is_pde) { int cache_bits, pat_flag, pat_idx; if (mode < 0 || mode >= PAT_INDEX_SIZE || pat_index[mode] < 0) panic("Unknown caching mode %d\n", mode); switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT; /* Map the caching mode to a PAT index. */ pat_idx = pat_index[mode]; /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */ cache_bits = 0; if (pat_idx & 0x4) cache_bits |= pat_flag; if (pat_idx & 0x2) cache_bits |= PG_NC_PCD; if (pat_idx & 0x1) cache_bits |= PG_NC_PWT; break; case PT_EPT: cache_bits = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(mode); break; default: panic("unsupported pmap type %d", pmap->pm_type); } return (cache_bits); } static int pmap_cache_mask(pmap_t pmap, boolean_t is_pde) { int mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = is_pde ? X86_PG_PDE_CACHE : X86_PG_PTE_CACHE; break; case PT_EPT: mask = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(0x7); break; default: panic("pmap_cache_mask: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline boolean_t pmap_ps_enabled(pmap_t pmap) { return (pg_ps_enabled && (pmap->pm_flags & PMAP_PDE_SUPERPAGE) != 0); } static void pmap_update_pde_store(pmap_t pmap, pd_entry_t *pde, pd_entry_t newpde) { switch (pmap->pm_type) { case PT_X86: break; case PT_RVI: case PT_EPT: /* * XXX * This is a little bogus since the generation number is * supposed to be bumped up when a region of the address * space is invalidated in the page tables. * * In this case the old PDE entry is valid but yet we want * to make sure that any mappings using the old entry are * invalidated in the TLB. * * The reason this works as expected is because we rendezvous * "all" host cpus and force any vcpu context to exit as a * side-effect. */ atomic_add_acq_long(&pmap->pm_eptgen, 1); break; default: panic("pmap_update_pde_store: bad pm_type %d", pmap->pm_type); } pde_store(pde, newpde); } /* * After changing the page size for the specified virtual address in the page * table, flush the corresponding entries from the processor's TLB. Only the * calling processor's TLB is affected. * * The calling thread must be pinned to a processor. */ static void pmap_update_pde_invalidate(pmap_t pmap, vm_offset_t va, pd_entry_t newpde) { pt_entry_t PG_G; if (pmap_type_guest(pmap)) return; KASSERT(pmap->pm_type == PT_X86, ("pmap_update_pde_invalidate: invalid type %d", pmap->pm_type)); PG_G = pmap_global_bit(pmap); if ((newpde & PG_PS) == 0) /* Demotion: flush a specific 2MB page mapping. */ invlpg(va); else if ((newpde & PG_G) == 0) /* * Promotion: flush every 4KB page mapping from the TLB * because there are too many to flush individually. */ invltlb(); else { /* * Promotion: flush every 4KB page mapping from the TLB, * including any global (PG_G) mappings. */ invltlb_glob(); } } #ifdef SMP /* * For SMP, these functions have to use the IPI mechanism for coherence. * * N.B.: Before calling any of the following TLB invalidation functions, * the calling processor must ensure that all stores updating a non- * kernel page table are globally performed. Otherwise, another * processor could cache an old, pre-update entry without being * invalidated. This can happen one of two ways: (1) The pmap becomes * active on another processor after its pm_active field is checked by * one of the following functions but before a store updating the page * table is globally performed. (2) The pmap becomes active on another * processor before its pm_active field is checked but due to * speculative loads one of the following functions stills reads the * pmap as inactive on the other processor. * * The kernel page table is exempt because its pm_active field is * immutable. The kernel page table is always active on every * processor. */ /* * Interrupt the cpus that are executing in the guest context. * This will force the vcpu to exit and the cached EPT mappings * will be invalidated by the host before the next vmresume. */ static __inline void pmap_invalidate_ept(pmap_t pmap) { int ipinum; sched_pin(); KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active), ("pmap_invalidate_ept: absurd pm_active")); /* * The TLB mappings associated with a vcpu context are not * flushed each time a different vcpu is chosen to execute. * * This is in contrast with a process's vtop mappings that * are flushed from the TLB on each context switch. * * Therefore we need to do more than just a TLB shootdown on * the active cpus in 'pmap->pm_active'. To do this we keep * track of the number of invalidations performed on this pmap. * * Each vcpu keeps a cache of this counter and compares it * just before a vmresume. If the counter is out-of-date an * invept will be done to flush stale mappings from the TLB. */ atomic_add_acq_long(&pmap->pm_eptgen, 1); /* * Force the vcpu to exit and trap back into the hypervisor. */ ipinum = pmap->pm_flags & PMAP_NESTED_IPIMASK; ipi_selected(pmap->pm_active, ipinum); sched_unpin(); } void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { cpuset_t *mask; u_int cpuid, i; if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_page: invalid type %d", pmap->pm_type)); sched_pin(); if (pmap == kernel_pmap) { invlpg(va); mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); if (pmap == PCPU_GET(curpmap)) invlpg(va); else if (pmap_pcid_enabled) pmap->pm_pcids[cpuid].pm_gen = 0; if (pmap_pcid_enabled) { CPU_FOREACH(i) { if (cpuid != i) pmap->pm_pcids[i].pm_gen = 0; } } mask = &pmap->pm_active; } smp_masked_invlpg(*mask, va); sched_unpin(); } /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */ #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE) void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { cpuset_t *mask; vm_offset_t addr; u_int cpuid, i; if (eva - sva >= PMAP_INVLPG_THRESHOLD) { pmap_invalidate_all(pmap); return; } if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: invalid type %d", pmap->pm_type)); sched_pin(); cpuid = PCPU_GET(cpuid); if (pmap == kernel_pmap) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); mask = &all_cpus; } else { if (pmap == PCPU_GET(curpmap)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); } else if (pmap_pcid_enabled) { pmap->pm_pcids[cpuid].pm_gen = 0; } if (pmap_pcid_enabled) { CPU_FOREACH(i) { if (cpuid != i) pmap->pm_pcids[i].pm_gen = 0; } } mask = &pmap->pm_active; } smp_masked_invlpg_range(*mask, sva, eva); sched_unpin(); } void pmap_invalidate_all(pmap_t pmap) { cpuset_t *mask; struct invpcid_descr d; u_int cpuid, i; if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_all: invalid type %d", pmap->pm_type)); sched_pin(); if (pmap == kernel_pmap) { if (pmap_pcid_enabled && invpcid_works) { bzero(&d, sizeof(d)); invpcid(&d, INVPCID_CTXGLOB); } else { invltlb_glob(); } mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); if (pmap == PCPU_GET(curpmap)) { if (pmap_pcid_enabled) { if (invpcid_works) { d.pcid = pmap->pm_pcids[cpuid].pm_pcid; d.pad = 0; d.addr = 0; invpcid(&d, INVPCID_CTX); } else { load_cr3(pmap->pm_cr3 | pmap->pm_pcids [PCPU_GET(cpuid)].pm_pcid); } } else { invltlb(); } } else if (pmap_pcid_enabled) { pmap->pm_pcids[cpuid].pm_gen = 0; } if (pmap_pcid_enabled) { CPU_FOREACH(i) { if (cpuid != i) pmap->pm_pcids[i].pm_gen = 0; } } mask = &pmap->pm_active; } smp_masked_invltlb(*mask, pmap); sched_unpin(); } void pmap_invalidate_cache(void) { sched_pin(); wbinvd(); smp_cache_flush(); sched_unpin(); } struct pde_action { cpuset_t invalidate; /* processors that invalidate their TLB */ pmap_t pmap; vm_offset_t va; pd_entry_t *pde; pd_entry_t newpde; u_int store; /* processor that updates the PDE */ }; static void pmap_update_pde_action(void *arg) { struct pde_action *act = arg; if (act->store == PCPU_GET(cpuid)) pmap_update_pde_store(act->pmap, act->pde, act->newpde); } static void pmap_update_pde_teardown(void *arg) { struct pde_action *act = arg; if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate)) pmap_update_pde_invalidate(act->pmap, act->va, act->newpde); } /* * Change the page size for the specified virtual address in a way that * prevents any possibility of the TLB ever having two entries that map the * same virtual address using different page sizes. This is the recommended * workaround for Erratum 383 on AMD Family 10h processors. It prevents a * machine check exception for a TLB state that is improperly diagnosed as a * hardware error. */ static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { struct pde_action act; cpuset_t active, other_cpus; u_int cpuid; sched_pin(); cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (pmap == kernel_pmap || pmap_type_guest(pmap)) active = all_cpus; else { active = pmap->pm_active; } if (CPU_OVERLAP(&active, &other_cpus)) { act.store = cpuid; act.invalidate = active; act.va = va; act.pmap = pmap; act.pde = pde; act.newpde = newpde; CPU_SET(cpuid, &active); smp_rendezvous_cpus(active, smp_no_rendevous_barrier, pmap_update_pde_action, pmap_update_pde_teardown, &act); } else { pmap_update_pde_store(pmap, pde, newpde); if (CPU_ISSET(cpuid, &active)) pmap_update_pde_invalidate(pmap, va, newpde); } sched_unpin(); } #else /* !SMP */ /* * Normal, non-SMP, invalidation functions. */ void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) invlpg(va); else if (pmap_pcid_enabled) pmap->pm_pcids[0].pm_gen = 0; } void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t addr; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); } else if (pmap_pcid_enabled) { pmap->pm_pcids[0].pm_gen = 0; } } void pmap_invalidate_all(pmap_t pmap) { struct invpcid_descr d; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_all: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap) { if (pmap_pcid_enabled && invpcid_works) { bzero(&d, sizeof(d)); invpcid(&d, INVPCID_CTXGLOB); } else { invltlb_glob(); } } else if (pmap == PCPU_GET(curpmap)) { if (pmap_pcid_enabled) { if (invpcid_works) { d.pcid = pmap->pm_pcids[0].pm_pcid; d.pad = 0; d.addr = 0; invpcid(&d, INVPCID_CTX); } else { load_cr3(pmap->pm_cr3 | pmap->pm_pcids[0]. pm_pcid); } } else { invltlb(); } } else if (pmap_pcid_enabled) { pmap->pm_pcids[0].pm_gen = 0; } } PMAP_INLINE void pmap_invalidate_cache(void) { wbinvd(); } static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { pmap_update_pde_store(pmap, pde, newpde); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) pmap_update_pde_invalidate(pmap, va, newpde); else pmap->pm_pcids[0].pm_gen = 0; } #endif /* !SMP */ #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024) void pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva, boolean_t force) { if (force) { sva &= ~(vm_offset_t)cpu_clflush_line_size; } else { KASSERT((sva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: sva not page-aligned")); KASSERT((eva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: eva not page-aligned")); } if ((cpu_feature & CPUID_SS) != 0 && !force) ; /* If "Self Snoop" is supported and allowed, do nothing. */ else if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0 && eva - sva < PMAP_CLFLUSH_THRESHOLD) { /* * XXX: Some CPUs fault, hang, or trash the local APIC * registers if we use CLFLUSH on the local APIC * range. The local APIC is always uncached, so we * don't need to flush for that range anyway. */ if (pmap_kextract(sva) == lapic_paddr) return; /* * Otherwise, do per-cache line flush. Use the mfence * instruction to insure that previous stores are * included in the write-back. The processor * propagates flush to other processors in the cache * coherence domain. */ mfence(); for (; sva < eva; sva += cpu_clflush_line_size) clflushopt(sva); mfence(); } else if ((cpu_feature & CPUID_CLFSH) != 0 && eva - sva < PMAP_CLFLUSH_THRESHOLD) { if (pmap_kextract(sva) == lapic_paddr) return; /* * Writes are ordered by CLFLUSH on Intel CPUs. */ if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (; sva < eva; sva += cpu_clflush_line_size) clflush(sva); if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } else { /* * No targeted cache flush methods are supported by CPU, * or the supplied range is bigger than 2MB. * Globally invalidate cache. */ pmap_invalidate_cache(); } } /* * Remove the specified set of pages from the data and instruction caches. * * In contrast to pmap_invalidate_cache_range(), this function does not * rely on the CPU's self-snoop feature, because it is intended for use * when moving pages into a different cache domain. */ void pmap_invalidate_cache_pages(vm_page_t *pages, int count) { vm_offset_t daddr, eva; int i; bool useclflushopt; useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0; if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE || ((cpu_feature & CPUID_CLFSH) == 0 && !useclflushopt)) pmap_invalidate_cache(); else { if (useclflushopt || cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (i = 0; i < count; i++) { daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i])); eva = daddr + PAGE_SIZE; for (; daddr < eva; daddr += cpu_clflush_line_size) { if (useclflushopt) clflushopt(daddr); else clflush(daddr); } } if (useclflushopt || cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte, PG_V; vm_paddr_t pa; pa = 0; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pdpe = pmap_pdpe(pmap, va); if (pdpe != NULL && (*pdpe & PG_V) != 0) { if ((*pdpe & PG_PS) != 0) pa = (*pdpe & PG_PS_FRAME) | (va & PDPMASK); else { pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & PG_V) != 0) { if ((*pde & PG_PS) != 0) { pa = (*pde & PG_PS_FRAME) | (va & PDRMASK); } else { pte = pmap_pde_to_pte(pde, va); pa = (*pte & PG_FRAME) | (va & PAGE_MASK); } } } } PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pd_entry_t pde, *pdep; pt_entry_t pte, PG_RW, PG_V; vm_paddr_t pa; vm_page_t m; pa = 0; m = NULL; PG_RW = pmap_rw_bit(pmap); PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); retry: pdep = pmap_pde(pmap, va); if (pdep != NULL && (pde = *pdep)) { if (pde & PG_PS) { if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { if (vm_page_pa_tryrelock(pmap, (pde & PG_PS_FRAME) | (va & PDRMASK), &pa)) goto retry; m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) | (va & PDRMASK)); vm_page_hold(m); } } else { pte = *pmap_pde_to_pte(pdep, va); if ((pte & PG_V) && ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { if (vm_page_pa_tryrelock(pmap, pte & PG_FRAME, &pa)) goto retry; m = PHYS_TO_VM_PAGE(pte & PG_FRAME); vm_page_hold(m); } } } PA_UNLOCK_COND(pa); PMAP_UNLOCK(pmap); return (m); } vm_paddr_t pmap_kextract(vm_offset_t va) { pd_entry_t pde; vm_paddr_t pa; if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { pa = DMAP_TO_PHYS(va); } else { pde = *vtopde(va); if (pde & PG_PS) { pa = (pde & PG_PS_FRAME) | (va & PDRMASK); } else { /* * Beware of a concurrent promotion that changes the * PDE at this point! For example, vtopte() must not * be used to access the PTE because it would use the * new PDE. It is, however, safe to use the old PDE * because the page table page is preserved by the * promotion. */ pa = *pmap_pde_to_pte(&pde, va); pa = (pa & PG_FRAME) | (va & PAGE_MASK); } } return (pa); } /*************************************************** * Low level mapping routines..... ***************************************************/ /* * Add a wired page to the kva. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kenter(vm_offset_t va, vm_paddr_t pa) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | X86_PG_RW | X86_PG_V | X86_PG_G); } static __inline void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode) { pt_entry_t *pte; int cache_bits; pte = vtopte(va); cache_bits = pmap_cache_bits(kernel_pmap, mode, 0); pte_store(pte, pa | X86_PG_RW | X86_PG_V | X86_PG_G | cache_bits); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *pte; pte = vtopte(va); pte_clear(pte); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return PHYS_TO_DMAP(start); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pt_entry_t *endpte, oldpte, pa, *pte; vm_page_t m; int cache_bits; oldpte = 0; pte = vtopte(sva); endpte = pte + count; while (pte < endpte) { m = *ma++; cache_bits = pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); pa = VM_PAGE_TO_PHYS(m) | cache_bits; if ((*pte & (PG_FRAME | X86_PG_PTE_CACHE)) != pa) { oldpte |= *pte; pte_store(pte, pa | X86_PG_G | X86_PG_RW | X86_PG_V); } pte++; } if (__predict_false((oldpte & X86_PG_V) != 0)) pmap_invalidate_range(kernel_pmap, sva, sva + count * PAGE_SIZE); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { KASSERT(va >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", va)); pmap_kremove(va); va += PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ static __inline void pmap_free_zero_pages(struct spglist *free) { vm_page_t m; while ((m = SLIST_FIRST(free)) != NULL) { SLIST_REMOVE_HEAD(free, plinks.s.ss); /* Preserve the page's PG_ZERO setting. */ vm_page_free_toq(m); } } /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Looks for a page table page mapping the specified virtual address in the * specified pmap's collection of idle page table pages. Returns NULL if there * is no page table page corresponding to the specified virtual address. */ static __inline vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_lookup(&pmap->pm_root, pmap_pde_pindex(va))); } /* * Removes the specified page table page from the specified pmap's collection * of idle page table pages. The specified page table page must be a member of * the pmap's collection. */ static __inline void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); vm_radix_remove(&pmap->pm_root, mpte->pindex); } /* * Decrements a page table page's wire count, which is used to record the * number of valid page table entries within the page. If the wire count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { --m->wire_count; if (m->wire_count == 0) { _pmap_unwire_ptp(pmap, va, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * unmap the page table page */ if (m->pindex >= (NUPDE + NUPDPE)) { /* PDP page */ pml4_entry_t *pml4; pml4 = pmap_pml4e(pmap, va); *pml4 = 0; } else if (m->pindex >= NUPDE) { /* PD page */ pdp_entry_t *pdp; pdp = pmap_pdpe(pmap, va); *pdp = 0; } else { /* PTE page */ pd_entry_t *pd; pd = pmap_pde(pmap, va); *pd = 0; } pmap_resident_count_dec(pmap, 1); if (m->pindex < NUPDE) { /* We just released a PT, unhold the matching PD */ vm_page_t pdpg; pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pdpg, free); } if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) { /* We just released a PD, unhold the matching PDP */ vm_page_t pdppg; pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pdppg, free); } /* * This is a release store so that the ordinary store unmapping * the page table page is globally performed before TLB shoot- * down is begun. */ atomic_subtract_rel_int(&vm_cnt.v_wire_count, 1); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing a page table entry, this routine is used to * conditionally free the page, and manage the hold/wire counts. */ static int pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde, struct spglist *free) { vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (0); KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0")); mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); return (pmap_unwire_ptp(pmap, va, mpte, free)); } void pmap_pinit0(pmap_t pmap) { int i; PMAP_LOCK_INIT(pmap); pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); pmap->pm_cr3 = KPML4phys; pmap->pm_root.rt_root = 0; CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); pmap->pm_flags = pmap_flags; CPU_FOREACH(i) { pmap->pm_pcids[i].pm_pcid = PMAP_PCID_NONE; pmap->pm_pcids[i].pm_gen = 0; } PCPU_SET(curpmap, kernel_pmap); pmap_activate(curthread); CPU_FILL(&kernel_pmap->pm_active); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ int pmap_pinit_type(pmap_t pmap, enum pmap_type pm_type, int flags) { vm_page_t pml4pg; vm_paddr_t pml4phys; int i; /* * allocate the page directory page */ while ((pml4pg = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) VM_WAIT; pml4phys = VM_PAGE_TO_PHYS(pml4pg); pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(pml4phys); CPU_FOREACH(i) { pmap->pm_pcids[i].pm_pcid = PMAP_PCID_NONE; pmap->pm_pcids[i].pm_gen = 0; } pmap->pm_cr3 = ~0; /* initialize to an invalid value */ if ((pml4pg->flags & PG_ZERO) == 0) pagezero(pmap->pm_pml4); /* * Do not install the host kernel mappings in the nested page * tables. These mappings are meaningless in the guest physical * address space. */ if ((pmap->pm_type = pm_type) == PT_X86) { pmap->pm_cr3 = pml4phys; /* Wire in kernel global address entries. */ for (i = 0; i < NKPML4E; i++) { pmap->pm_pml4[KPML4BASE + i] = (KPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V | PG_U; } for (i = 0; i < ndmpdpphys; i++) { pmap->pm_pml4[DMPML4I + i] = (DMPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V | PG_U; } /* install self-referential address mapping entry(s) */ pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; } pmap->pm_root.rt_root = 0; CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); pmap->pm_flags = flags; pmap->pm_eptgen = 0; return (1); } int pmap_pinit(pmap_t pmap) { return (pmap_pinit_type(pmap, PT_X86, pmap_flags)); } /* * This routine is called if the desired page table page does not exist. * * If page table page allocation fails, this routine may sleep before * returning NULL. It sleeps only if a lock pointer was given. * * Note: If a page allocation fails at page table level two or three, * one or two pages may be held during the wait, only to be released * afterwards. This conservative approach is easily argued to avoid * race conditions. */ static vm_page_t _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp) { vm_page_t m, pdppg, pdpg; pt_entry_t PG_A, PG_M, PG_RW, PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); /* * Allocate a page table page. */ if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { if (lockp != NULL) { RELEASE_PV_LIST_LOCK(lockp); PMAP_UNLOCK(pmap); PMAP_ASSERT_NOT_IN_DI(); VM_WAIT; PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, the page table * page may have been allocated. */ return (NULL); } if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); /* * Map the pagetable page into the process address space, if * it isn't already there. */ if (ptepindex >= (NUPDE + NUPDPE)) { pml4_entry_t *pml4; vm_pindex_t pml4index; /* Wire up a new PDPE page */ pml4index = ptepindex - (NUPDE + NUPDPE); pml4 = &pmap->pm_pml4[pml4index]; *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } else if (ptepindex >= NUPDE) { vm_pindex_t pml4index; vm_pindex_t pdpindex; pml4_entry_t *pml4; pdp_entry_t *pdp; /* Wire up a new PDE page */ pdpindex = ptepindex - NUPDE; pml4index = pdpindex >> NPML4EPGSHIFT; pml4 = &pmap->pm_pml4[pml4index]; if ((*pml4 & PG_V) == 0) { /* Have to allocate a new pdp, recurse */ if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index, lockp) == NULL) { --m->wire_count; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } } else { /* Add reference to pdp page */ pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME); pdppg->wire_count++; } pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); /* Now find the pdp page */ pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } else { vm_pindex_t pml4index; vm_pindex_t pdpindex; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pd; /* Wire up a new PTE page */ pdpindex = ptepindex >> NPDPEPGSHIFT; pml4index = pdpindex >> NPML4EPGSHIFT; /* First, find the pdp and check that its valid. */ pml4 = &pmap->pm_pml4[pml4index]; if ((*pml4 & PG_V) == 0) { /* Have to allocate a new pd, recurse */ if (_pmap_allocpte(pmap, NUPDE + pdpindex, lockp) == NULL) { --m->wire_count; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; } else { pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)]; if ((*pdp & PG_V) == 0) { /* Have to allocate a new pd, recurse */ if (_pmap_allocpte(pmap, NUPDE + pdpindex, lockp) == NULL) { --m->wire_count; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } } else { /* Add reference to the pd page */ pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME); pdpg->wire_count++; } } pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME); /* Now we know where the page directory page is */ pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)]; *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } pmap_resident_count_inc(pmap, 1); return (m); } static vm_page_t pmap_allocpde(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t pdpindex, ptepindex; pdp_entry_t *pdpe, PG_V; vm_page_t pdpg; PG_V = pmap_valid_bit(pmap); retry: pdpe = pmap_pdpe(pmap, va); if (pdpe != NULL && (*pdpe & PG_V) != 0) { /* Add a reference to the pd page. */ pdpg = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME); pdpg->wire_count++; } else { /* Allocate a pd page. */ ptepindex = pmap_pde_pindex(va); pdpindex = ptepindex >> NPDPEPGSHIFT; pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex, lockp); if (pdpg == NULL && lockp != NULL) goto retry; } return (pdpg); } static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t ptepindex; pd_entry_t *pd, PG_V; vm_page_t m; PG_V = pmap_valid_bit(pmap); /* * Calculate pagetable page index */ ptepindex = pmap_pde_pindex(va); retry: /* * Get the page directory entry */ pd = pmap_pde(pmap, va); /* * This supports switching from a 2MB page to a * normal 4K page. */ if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) { if (!pmap_demote_pde_locked(pmap, pd, va, lockp)) { /* * Invalidation of the 2MB page mapping may have caused * the deallocation of the underlying PD page. */ pd = NULL; } } /* * If the page table page is mapped, we just increment the * hold count, and activate it. */ if (pd != NULL && (*pd & PG_V) != 0) { m = PHYS_TO_VM_PAGE(*pd & PG_FRAME); m->wire_count++; } else { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ m = _pmap_allocpte(pmap, ptepindex, lockp); if (m == NULL && lockp != NULL) goto retry; } return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { vm_page_t m; int i; KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap has reserved page table page(s)")); KASSERT(CPU_EMPTY(&pmap->pm_active), ("releasing active pmap %p", pmap)); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pml4)); for (i = 0; i < NKPML4E; i++) /* KVA */ pmap->pm_pml4[KPML4BASE + i] = 0; for (i = 0; i < ndmpdpphys; i++)/* Direct Map */ pmap->pm_pml4[DMPML4I + i] = 0; pmap->pm_pml4[PML4PML4I] = 0; /* Recursive Mapping */ m->wire_count--; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; return sysctl_handle_long(oidp, &ksize, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_size, "LU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return sysctl_handle_long(oidp, &kfree, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_free, "LU", "Amount of KVM free"); /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t paddr; vm_page_t nkpg; pd_entry_t *pde, newpdir; pdp_entry_t *pdpe; mtx_assert(&kernel_map->system_mtx, MA_OWNED); /* * Return if "addr" is within the range of kernel page table pages * that were preallocated during pmap bootstrap. Moreover, leave * "kernel_vm_end" and the kernel page table as they were. * * The correctness of this action is based on the following * argument: vm_map_insert() allocates contiguous ranges of the * kernel virtual address space. It calls this function if a range * ends after "kernel_vm_end". If the kernel is mapped between * "kernel_vm_end" and "addr", then the range cannot begin at * "kernel_vm_end". In fact, its beginning address cannot be less * than the kernel. Thus, there is no immediate need to allocate * any new kernel page table pages between "kernel_vm_end" and * "KERNBASE". */ if (KERNBASE < addr && addr <= KERNBASE + nkpt * NBPDR) return; addr = roundup2(addr, NBPDR); if (addr - 1 >= kernel_map->max_offset) addr = kernel_map->max_offset; while (kernel_vm_end < addr) { pdpe = pmap_pdpe(kernel_pmap, kernel_vm_end); if ((*pdpe & X86_PG_V) == 0) { /* We need a new PDP entry */ nkpg = vm_page_alloc(NULL, kernel_vm_end >> PDPSHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); if ((nkpg->flags & PG_ZERO) == 0) pmap_zero_page(nkpg); paddr = VM_PAGE_TO_PHYS(nkpg); *pdpe = (pdp_entry_t)(paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M); continue; /* try again */ } pde = pmap_pdpe_to_pde(pdpe, kernel_vm_end); if ((*pde & X86_PG_V) != 0) { kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } continue; } nkpg = vm_page_alloc(NULL, pmap_pde_pindex(kernel_vm_end), VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); if ((nkpg->flags & PG_ZERO) == 0) pmap_zero_page(nkpg); paddr = VM_PAGE_TO_PHYS(nkpg); newpdir = paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; pde_store(pde, newpdir); kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 3); CTASSERT(_NPCPV == 168); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0 0xfffffffffffffffful #define PC_FREE1 0xfffffffffffffffful #define PC_FREE2 0x000000fffffffffful static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 }; #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, "Current number of pv entry chunks"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, "Current number of pv entry chunks allocated"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, "Current number of pv entry chunks frees"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, "Number of times tried to get a chunk page but failed."); static long pv_entry_frees, pv_entry_allocs, pv_entry_count; static int pv_entry_spare; SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, "Current number of pv entry frees"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, "Current number of pv entry allocs"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, "Current number of pv entries"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, "Current number of spare pv entries"); #endif /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. * * Returns NULL if PV entries were reclaimed from the specified pmap. * * We do not, however, unmap 2mpages because subsequent accesses will * allocate per-page pv entries until repromotion occurs, thereby * exacerbating the shortage of free pv entries. */ static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) { struct pch new_tail; struct pv_chunk *pc; struct md_page *pvh; pd_entry_t *pde; pmap_t pmap; pt_entry_t *pte, tpte; pt_entry_t PG_G, PG_A, PG_M, PG_RW; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint64_t inuse; int bit, field, freed; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL")); pmap = NULL; m_pc = NULL; PG_G = PG_A = PG_M = PG_RW = 0; SLIST_INIT(&free); TAILQ_INIT(&new_tail); pmap_delayed_invl_started(); mtx_lock(&pv_chunks_mutex); while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && SLIST_EMPTY(&free)) { TAILQ_REMOVE(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); if (pmap != pc->pc_pmap) { if (pmap != NULL) { pmap_invalidate_all(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } pmap_delayed_invl_finished(); pmap_delayed_invl_started(); pmap = pc->pc_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) { RELEASE_PV_LIST_LOCK(lockp); PMAP_LOCK(pmap); } else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) { pmap = NULL; TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru); mtx_lock(&pv_chunks_mutex); continue; } PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = bsfq(inuse); pv = &pc->pc_pventry[field * 64 + bit]; va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_PS) != 0) continue; pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_W) != 0) continue; tpte = pte_load_clear(pte); if ((tpte & PG_G) != 0) pmap_invalidate_page(pmap, va); m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((tpte & PG_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pmap_delayed_invl_page(m); pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt(pmap, va, *pde, &free); freed++; } } if (freed == 0) { TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru); mtx_lock(&pv_chunks_mutex); continue; } /* Every freed mapping is for a 4 KB page. */ pmap_resident_count_dec(pmap, freed); PV_STAT(atomic_add_long(&pv_entry_frees, freed)); PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1 && pc->pc_map[2] == PC_FREE2) { PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m_pc->phys_addr); mtx_lock(&pv_chunks_mutex); break; } TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru); mtx_lock(&pv_chunks_mutex); /* One freed pv entry in locked_pmap is sufficient. */ if (pmap == locked_pmap) break; } TAILQ_CONCAT(&pv_chunks, &new_tail, pc_lru); mtx_unlock(&pv_chunks_mutex); if (pmap != NULL) { pmap_invalidate_all(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } pmap_delayed_invl_finished(); if (m_pc == NULL && !SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->wire_count = 1; atomic_add_int(&vm_cnt.v_wire_count, 1); } pmap_free_zero_pages(&free); return (m_pc); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_frees, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, 1)); PV_STAT(atomic_subtract_long(&pv_entry_count, 1)); pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 || pc->pc_map[2] != PC_FREE2) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk(struct pv_chunk *pc) { vm_page_t m; mtx_lock(&pv_chunks_mutex); TAILQ_REMOVE(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m->phys_addr); vm_page_unwire(m, PQ_NONE); vm_page_free(m); } /* * Returns a new PV entry, allocating a new PV chunk from the system when * needed. If this PV chunk allocation fails and a PV list lock pointer was * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is * returned. * * The given PV list lock may be released. */ static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp) { int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_allocs, 1)); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = bsfq(pc->pc_map[field]); break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 64 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, 1)); return (pv); } } /* No free items, allocate another chunk */ m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); if (m == NULL) { if (lockp == NULL) { PV_STAT(pc_chunk_tryfail++); return (NULL); } m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */ pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; mtx_lock(&pv_chunks_mutex); TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1)); return (pv); } /* * Returns the number of one bits within the given PV chunk map. * * The erratas for Intel processors state that "POPCNT Instruction May * Take Longer to Execute Than Expected". It is believed that the * issue is the spurious dependency on the destination register. * Provide a hint to the register rename logic that the destination * value is overwritten, by clearing it, as suggested in the * optimization manual. It should be cheap for unaffected processors * as well. * * Reference numbers for erratas are * 4th Gen Core: HSD146 * 5th Gen Core: BDM85 * 6th Gen Core: SKL029 */ static int popcnt_pc_map_pq(uint64_t *map) { u_long result, tmp; __asm __volatile("xorl %k0,%k0;popcntq %2,%0;" "xorl %k1,%k1;popcntq %3,%1;addl %k1,%k0;" "xorl %k1,%k1;popcntq %4,%1;addl %k1,%k0" : "=&r" (result), "=&r" (tmp) : "m" (map[0]), "m" (map[1]), "m" (map[2])); return (result); } /* * Ensure that the number of spare PV entries in the specified pmap meets or * exceeds the given count, "needed". * * The given PV list lock may be released. */ static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp) { struct pch new_tail; struct pv_chunk *pc; int avail, free; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL")); /* * Newly allocated PV chunks must be stored in a private list until * the required number of PV chunks have been allocated. Otherwise, * reclaim_pv_chunk() could recycle one of these chunks. In * contrast, these chunks must be added to the pmap upon allocation. */ TAILQ_INIT(&new_tail); retry: avail = 0; TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) { #ifndef __POPCNT__ if ((cpu_feature2 & CPUID2_POPCNT) == 0) bit_count((bitstr_t *)pc->pc_map, 0, sizeof(pc->pc_map) * NBBY, &free); else #endif free = popcnt_pc_map_pq(pc->pc_map); if (free == 0) break; avail += free; if (avail >= needed) break; } for (; avail < needed; avail += _NPCPV) { m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); if (m == NULL) { m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0; pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV)); } if (!TAILQ_EMPTY(&new_tail)) { mtx_lock(&pv_chunks_mutex); TAILQ_CONCAT(&pv_chunks, &new_tail, pc_lru); mtx_unlock(&pv_chunks_mutex); } } /* * First find and then remove the pv entry for the specified pmap and virtual * address from the specified pv list. Returns the pv entry if found and NULL * otherwise. This operation can be performed on pv lists for either 4KB or * 2MB page mappings. */ static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; break; } } return (pv); } /* * After demotion from a 2MB page mapping to 512 4KB page mappings, * destroy the pv entry for the 2MB page mapping and reinstantiate the pv * entries for each of the 4KB page mappings. */ static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; struct pv_chunk *pc; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; int bit, field; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_demote_pde: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the 2mpage's pv entry for this mapping to the first * page's pv list. Once this transfer begins, the pv list lock * must not be released until the last pv entry is reinstantiated. */ pvh = pa_to_pvh(pa); va = trunc_2mpage(va); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; /* Instantiate the remaining NPTEPG - 1 pv entries. */ PV_STAT(atomic_add_long(&pv_entry_allocs, NPTEPG - 1)); va_last = va + NBPDR - PAGE_SIZE; for (;;) { pc = TAILQ_FIRST(&pmap->pm_pvchunk); KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0 || pc->pc_map[2] != 0, ("pmap_pv_demote_pde: missing spare")); for (field = 0; field < _NPCM; field++) { while (pc->pc_map[field]) { bit = bsfq(pc->pc_map[field]); pc->pc_map[field] &= ~(1ul << bit); pv = &pc->pc_pventry[field * 64 + bit]; va += PAGE_SIZE; pv->pv_va = va; m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_pde: page %p is not managed", m)); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (va == va_last) goto out; } } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } out: if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, NPTEPG - 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, NPTEPG - 1)); } /* * After promotion from 512 4KB page mappings to a single 2MB page mapping, * replace the many pv entries for the 4KB page mappings by a single pv entry * for the 2MB page mapping. */ static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; KASSERT((pa & PDRMASK) == 0, ("pmap_pv_promote_pde: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the first page's pv entry for this mapping to the 2mpage's * pv list. Aside from avoiding the cost of a call to get_pv_entry(), * a transfer avoids the possibility that get_pv_entry() calls * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the * mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = trunc_2mpage(va); pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } /* * First find and then destroy the pv entry for the specified pmap and virtual * address. This operation can be performed on pv lists for either 4KB or 2MB * page mappings. */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } /* * Conditionally create the PV entry for a 4KB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp) { pv_entry_t pv; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; return (TRUE); } else return (FALSE); } /* * Conditionally create the PV entry for a 2MB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; return (TRUE); } else return (FALSE); } /* * Fills a page table page with mappings to consecutive physical pages. */ static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte) { pt_entry_t *pte; for (pte = firstpte; pte < firstpte + NPTEPG; pte++) { *pte = newpte; newpte += PAGE_SIZE; } } /* * Tries to demote a 2MB page mapping. If demotion fails, the 2MB page * mapping is invalidated. */ static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { struct rwlock *lock; boolean_t rv; lock = NULL; rv = pmap_demote_pde_locked(pmap, pde, va, &lock); if (lock != NULL) rw_wunlock(lock); return (rv); } static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp) { pd_entry_t newpde, oldpde; pt_entry_t *firstpte, newpte; pt_entry_t PG_A, PG_G, PG_M, PG_RW, PG_V; vm_paddr_t mptepa; vm_page_t mpte; struct spglist free; int PG_PTE_CACHE; PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_PTE_CACHE = pmap_cache_mask(pmap, 0); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpde = *pde; KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V")); if ((oldpde & PG_A) != 0 && (mpte = pmap_lookup_pt_page(pmap, va)) != NULL) pmap_remove_pt_page(pmap, mpte); else { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: page table page for a wired mapping" " is missing")); /* * Invalidate the 2MB page mapping and return "failure" if the * mapping was never accessed or the allocation of the new * page table page fails. If the 2MB page mapping belongs to * the direct map region of the kernel's address space, then * the page allocation request specifies the highest possible * priority (VM_ALLOC_INTERRUPT). Otherwise, the priority is * normal. Page table pages are preallocated for every other * part of the kernel address space, so the direct map region * is the only part of the kernel address space that must be * handled here. */ if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc(NULL, pmap_pde_pindex(va), (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS ? VM_ALLOC_INTERRUPT : VM_ALLOC_NORMAL) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { SLIST_INIT(&free); pmap_remove_pde(pmap, pde, trunc_2mpage(va), &free, lockp); pmap_invalidate_page(pmap, trunc_2mpage(va)); pmap_free_zero_pages(&free); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } if (va < VM_MAXUSER_ADDRESS) pmap_resident_count_inc(pmap, 1); } mptepa = VM_PAGE_TO_PHYS(mpte); firstpte = (pt_entry_t *)PHYS_TO_DMAP(mptepa); newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V; KASSERT((oldpde & PG_A) != 0, ("pmap_demote_pde: oldpde is missing PG_A")); KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pde: oldpde is missing PG_M")); newpte = oldpde & ~PG_PS; newpte = pmap_swap_pat(pmap, newpte); /* * If the page table page is new, initialize it. */ if (mpte->wire_count == 1) { mpte->wire_count = NPTEPG; pmap_fill_ptp(firstpte, newpte); } KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME), ("pmap_demote_pde: firstpte and newpte map different physical" " addresses")); /* * If the mapping has changed attributes, update the page table * entries. */ if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE)) pmap_fill_ptp(firstpte, newpte); /* * The spare PV entries must be reserved prior to demoting the * mapping, that is, prior to changing the PDE. Otherwise, the state * of the PDE and the PV lists will be inconsistent, which can result * in reclaim_pv_chunk() attempting to remove a PV entry from the * wrong PV list and pmap_pv_demote_pde() failing to find the expected * PV entry for the 2MB page mapping that is being demoted. */ if ((oldpde & PG_MANAGED) != 0) reserve_pv_entries(pmap, NPTEPG - 1, lockp); /* * Demote the mapping. This pmap is locked. The old PDE has * PG_A set. If the old PDE has PG_RW set, it also has PG_M * set. Thus, there is no danger of a race with another * processor changing the setting of PG_A and/or PG_M between * the read above and the store below. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pde_store(pde, newpde); /* * Invalidate a stale recursive mapping of the page table page. */ if (va >= VM_MAXUSER_ADDRESS) pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); /* * Demote the PV entry. */ if ((oldpde & PG_MANAGED) != 0) pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME, lockp); atomic_add_long(&pmap_pde_demotions, 1); CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#lx" " in pmap %p", va, pmap); return (TRUE); } /* * pmap_remove_kernel_pde: Remove a kernel superpage mapping. */ static void pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; vm_paddr_t mptepa; vm_page_t mpte; KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte = pmap_lookup_pt_page(pmap, va); if (mpte == NULL) panic("pmap_remove_kernel_pde: Missing pt page."); pmap_remove_pt_page(pmap, mpte); mptepa = VM_PAGE_TO_PHYS(mpte); newpde = mptepa | X86_PG_M | X86_PG_A | X86_PG_RW | X86_PG_V; /* * Initialize the page table page. */ pagezero((void *)PHYS_TO_DMAP(mptepa)); /* * Demote the mapping. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pde_store(pde, newpde); /* * Invalidate a stale recursive mapping of the page table page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); } /* * pmap_remove_pde: do the things to unmap a superpage in a process */ static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pd_entry_t oldpde; vm_offset_t eva, va; vm_page_t m, mpte; pt_entry_t PG_G, PG_A, PG_M, PG_RW; PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_remove_pde: sva is not 2mpage aligned")); oldpde = pte_load_clear(pdq); if (oldpde & PG_W) pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; /* * Machines that don't support invlpg, also don't support * PG_G. */ if (oldpde & PG_G) pmap_invalidate_page(kernel_pmap, sva); pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE); if (oldpde & PG_MANAGED) { CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME); pvh = pa_to_pvh(oldpde & PG_PS_FRAME); pmap_pvh_free(pvh, pmap, sva); eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) { if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpde & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); pmap_delayed_invl_page(m); } } if (pmap == kernel_pmap) { pmap_remove_kernel_pde(pmap, pdq, sva); } else { mpte = pmap_lookup_pt_page(pmap, sva); if (mpte != NULL) { pmap_remove_pt_page(pmap, mpte); pmap_resident_count_dec(pmap, 1); KASSERT(mpte->wire_count == NPTEPG, ("pmap_remove_pde: pte page wire count error")); mpte->wire_count = 0; pmap_add_delayed_free_list(mpte, free, FALSE); atomic_subtract_int(&vm_cnt.v_wire_count, 1); } } return (pmap_unuse_pt(pmap, sva, *pmap_pdpe(pmap, sva), free)); } /* * pmap_remove_pte: do the things to unmap a page in a process */ static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pt_entry_t oldpte, PG_A, PG_M, PG_RW; vm_page_t m; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpte = pte_load_clear(ptq); if (oldpte & PG_W) pmap->pm_stats.wired_count -= 1; pmap_resident_count_dec(pmap, 1); if (oldpte & PG_MANAGED) { m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME); if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } pmap_delayed_invl_page(m); } return (pmap_unuse_pt(pmap, va, ptepde, free)); } /* * Remove a single page from a process address space */ static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, struct spglist *free) { struct rwlock *lock; pt_entry_t *pte, PG_V; PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((*pde & PG_V) == 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_V) == 0) return; lock = NULL; pmap_remove_pte(pmap, pte, va, *pde, free, &lock); if (lock != NULL) rw_wunlock(lock); pmap_invalidate_page(pmap, va); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct rwlock *lock; vm_offset_t va, va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t ptpaddr, *pde; pt_entry_t *pte, PG_G, PG_V; struct spglist free; int anyvalid; PG_G = pmap_global_bit(pmap); PG_V = pmap_valid_bit(pmap); /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; anyvalid = 0; SLIST_INIT(&free); pmap_delayed_invl_started(); PMAP_LOCK(pmap); /* * special handling of removing one page. a very * common operation and easy to short circuit some * code. */ if (sva + PAGE_SIZE == eva) { pde = pmap_pde(pmap, sva); if (pde && (*pde & PG_PS) == 0) { pmap_remove_page(pmap, sva, pde, &free); goto out; } } lock = NULL; for (; sva < eva; sva = va_next) { if (pmap->pm_stats.resident_count == 0) break; pml4e = pmap_pml4e(pmap, sva); if ((*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) { va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; continue; } /* * Calculate index for next page table. */ va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); ptpaddr = *pde; /* * Weed out invalid mappings. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we removing the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_remove_pde(). */ if ((ptpaddr & PG_G) == 0) anyvalid = 1; pmap_remove_pde(pmap, pde, sva, &free, &lock); continue; } else if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) { /* The large page mapping was destroyed. */ continue; } else ptpaddr = *pde; } /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (va_next > eva) va_next = eva; va = va_next; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if (*pte == 0) { if (va != va_next) { pmap_invalidate_range(pmap, va, sva); va = va_next; } continue; } if ((*pte & PG_G) == 0) anyvalid = 1; else if (va == va_next) va = sva; if (pmap_remove_pte(pmap, pte, sva, ptpaddr, &free, &lock)) { sva += PAGE_SIZE; break; } } if (va != va_next) pmap_invalidate_range(pmap, va, sva); } if (lock != NULL) rw_wunlock(lock); out: if (anyvalid) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_delayed_invl_finished(); pmap_free_zero_pages(&free); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; struct rwlock *lock; pt_entry_t *pte, tpte, PG_A, PG_M, PG_RW; pd_entry_t *pde; vm_offset_t va; struct spglist free; int pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); retry: rw_wlock(lock); while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { rw_wunlock(lock); PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pde = pmap_pde(pmap, va); (void)pmap_demote_pde_locked(pmap, pde, va, &lock); PMAP_UNLOCK(pmap); } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { rw_wunlock(lock); PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pmap_resident_count_dec(pmap, 1); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found" " a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); tpte = pte_load_clear(pte); if (tpte & PG_W) pmap->pm_stats.wired_count--; if (tpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); pmap_unuse_pt(pmap, pv->pv_va, *pde, &free); pmap_invalidate_page(pmap, pv->pv_va); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(lock); pmap_delayed_invl_wait(m); pmap_free_zero_pages(&free); } /* * pmap_protect_pde: do the things to protect a 2mpage in a process */ static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot) { pd_entry_t newpde, oldpde; vm_offset_t eva, va; vm_page_t m; boolean_t anychanged; pt_entry_t PG_G, PG_M, PG_RW; PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_protect_pde: sva is not 2mpage aligned")); anychanged = FALSE; retry: oldpde = newpde = *pde; if (oldpde & PG_MANAGED) { eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); } if ((prot & VM_PROT_WRITE) == 0) newpde &= ~(PG_RW | PG_M); if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (newpde != oldpde) { if (!atomic_cmpset_long(pde, oldpde, newpde)) goto retry; if (oldpde & PG_G) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } return (anychanged); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t ptpaddr, *pde; pt_entry_t *pte, PG_G, PG_M, PG_RW, PG_V; boolean_t anychanged; KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) == (VM_PROT_WRITE|VM_PROT_EXECUTE)) return; PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); anychanged = FALSE; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if ((*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) { va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); ptpaddr = *pde; /* * Weed out invalid mappings. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we protecting the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_protect_pde(). */ if (pmap_protect_pde(pmap, pde, sva, prot)) anychanged = TRUE; continue; } else if (!pmap_demote_pde(pmap, pde, sva)) { /* * The large page mapping was destroyed. */ continue; } } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { pt_entry_t obits, pbits; vm_page_t m; retry: obits = pbits = *pte; if ((pbits & PG_V) == 0) continue; if ((prot & VM_PROT_WRITE) == 0) { if ((pbits & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); vm_page_dirty(m); } pbits &= ~(PG_RW | PG_M); } if ((prot & VM_PROT_EXECUTE) == 0) pbits |= pg_nx; if (pbits != obits) { if (!atomic_cmpset_long(pte, obits, pbits)) goto retry; if (obits & PG_G) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } } } if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); } /* * Tries to promote the 512, contiguous 4KB page mappings that are within a * single page table page (PTP) to a single 2MB page mapping. For promotion * to occur, two conditions must be met: (1) the 4KB page mappings must map * aligned, contiguous physical memory and (2) the 4KB page mappings must have * identical characteristics. */ static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp) { pd_entry_t newpde; pt_entry_t *firstpte, oldpte, pa, *pte; pt_entry_t PG_G, PG_A, PG_M, PG_RW, PG_V; vm_page_t mpte; int PG_PTE_CACHE; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PG_PTE_CACHE = pmap_cache_mask(pmap, 0); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Examine the first PTE in the specified PTP. Abort if this PTE is * either invalid, unused, or does not map the first 4KB physical page * within a 2MB page. */ firstpte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); setpde: newpde = *firstpte; if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V)) { atomic_add_long(&pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } if ((newpde & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared without * a TLB invalidation. */ if (!atomic_cmpset_long(firstpte, newpde, newpde & ~PG_RW)) goto setpde; newpde &= ~PG_RW; } /* * Examine each of the other PTEs in the specified PTP. Abort if this * PTE maps an unexpected 4KB physical page or does not have identical * characteristics to the first PTE. */ pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE; for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) { setpte: oldpte = *pte; if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) { atomic_add_long(&pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } if ((oldpte & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared * without a TLB invalidation. */ if (!atomic_cmpset_long(pte, oldpte, oldpte & ~PG_RW)) goto setpte; oldpte &= ~PG_RW; CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#lx" " in pmap %p", (oldpte & PG_FRAME & PDRMASK) | (va & ~PDRMASK), pmap); } if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) { atomic_add_long(&pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } pa -= PAGE_SIZE; } /* * Save the page table page in its current state until the PDE * mapping the superpage is demoted by pmap_demote_pde() or * destroyed by pmap_remove_pde(). */ mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == pmap_pde_pindex(va), ("pmap_promote_pde: page table page's pindex is wrong")); if (pmap_insert_pt_page(pmap, mpte)) { atomic_add_long(&pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx in pmap %p", va, pmap); return; } /* * Promote the pv entries. */ if ((newpde & PG_MANAGED) != 0) pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME, lockp); /* * Propagate the PAT index to its proper position. */ newpde = pmap_swap_pat(pmap, newpde); /* * Map the superpage. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else pde_store(pde, PG_PS | newpde); atomic_add_long(&pmap_pde_promotions, 1); CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#lx" " in pmap %p", va, pmap); } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. * * When destroying both a page table and PV entry, this function * performs the TLB invalidation before releasing the PV list * lock, so we do not need pmap_delayed_invl_page() calls here. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind __unused) { struct rwlock *lock; pd_entry_t *pde; pt_entry_t *pte, PG_G, PG_A, PG_M, PG_RW, PG_V; pt_entry_t newpte, origpte; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte, om; boolean_t nosleep; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = trunc_page(va); KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig")); KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS, ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va)); KASSERT((m->oflags & VPO_UNMANAGED) != 0 || va < kmi.clean_sva || va >= kmi.clean_eva, ("pmap_enter: managed mapping within the clean submap")); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); pa = VM_PAGE_TO_PHYS(m); newpte = (pt_entry_t)(pa | PG_A | PG_V); if ((flags & VM_PROT_WRITE) != 0) newpte |= PG_M; if ((prot & VM_PROT_WRITE) != 0) newpte |= PG_RW; KASSERT((newpte & (PG_M | PG_RW)) != PG_M, ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't")); if ((prot & VM_PROT_EXECUTE) == 0) newpte |= pg_nx; if ((flags & PMAP_ENTER_WIRED) != 0) newpte |= PG_W; if (va < VM_MAXUSER_ADDRESS) newpte |= PG_U; if (pmap == kernel_pmap) newpte |= PG_G; newpte |= pmap_cache_bits(pmap, m->md.pat_mode, 0); /* * Set modified bit gratuitously for writeable mappings if * the page is unmanaged. We do not want to take a fault * to do the dirty bit accounting for these mappings. */ if ((m->oflags & VPO_UNMANAGED) != 0) { if ((newpte & PG_RW) != 0) newpte |= PG_M; } mpte = NULL; lock = NULL; PMAP_LOCK(pmap); /* * In the case that a page table page is not * resident, we are creating it here. */ retry: pde = pmap_pde(pmap, va); if (pde != NULL && (*pde & PG_V) != 0 && ((*pde & PG_PS) == 0 || pmap_demote_pde_locked(pmap, pde, va, &lock))) { pte = pmap_pde_to_pte(pde, va); if (va < VM_MAXUSER_ADDRESS && mpte == NULL) { mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); mpte->wire_count++; } } else if (va < VM_MAXUSER_ADDRESS) { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0; mpte = _pmap_allocpte(pmap, pmap_pde_pindex(va), nosleep ? NULL : &lock); if (mpte == NULL && nosleep) { if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (KERN_RESOURCE_SHORTAGE); } goto retry; } else panic("pmap_enter: invalid page directory va=%#lx", va); origpte = *pte; /* * Is the specified virtual address already mapped? */ if ((origpte & PG_V) != 0) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0) pmap->pm_stats.wired_count++; else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->wire_count--; KASSERT(mpte->wire_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%lx", va)); } /* * Has the physical page changed? */ opa = origpte & PG_FRAME; if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((origpte & PG_MANAGED) != 0) { newpte |= PG_MANAGED; if ((newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0) goto unchanged; goto validate; } } else { /* * Increment the counters. */ if ((newpte & PG_W) != 0) pmap->pm_stats.wired_count++; pmap_resident_count_inc(pmap, 1); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { newpte |= PG_MANAGED; pv = get_pv_entry(pmap, &lock); pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } /* * Update the PTE. */ if ((origpte & PG_V) != 0) { validate: origpte = pte_load_store(pte, newpte); opa = origpte & PG_FRAME; if (opa != pa) { if ((origpte & PG_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(om); if ((origpte & PG_A) != 0) vm_page_aflag_set(om, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa); pmap_pvh_free(&om->md, pmap, va); if ((om->aflags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } } else if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((origpte & PG_MANAGED) != 0) vm_page_dirty(m); /* * Although the PTE may still have PG_RW set, TLB * invalidation may nonetheless be required because * the PTE no longer has PG_M set. */ } else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) { /* * This PTE change does not require TLB invalidation. */ goto unchanged; } if ((origpte & PG_A) != 0) pmap_invalidate_page(pmap, va); } else pte_store(pte, newpte); unchanged: /* * If both the page table page and the reservation are fully * populated, then attempt promotion. */ if ((mpte == NULL || mpte->wire_count == NPTEPG) && pmap_ps_enabled(pmap) && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) pmap_promote_pde(pmap, pde, va, &lock); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } /* * Tries to create a 2MB page mapping. Returns TRUE if successful and FALSE * otherwise. Fails if (1) a page table page cannot be allocated without * blocking, (2) a mapping already exists at the specified virtual address, or * (3) a pv entry cannot be allocated without reclaiming another pv entry. */ static boolean_t pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp) { pd_entry_t *pde, newpde; pt_entry_t PG_V; vm_page_t mpde; struct spglist free; PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((mpde = pmap_allocpde(pmap, va, NULL)) == NULL) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpde)); pde = &pde[pmap_pde_index(va)]; if ((*pde & PG_V) != 0) { KASSERT(mpde->wire_count > 1, ("pmap_enter_pde: mpde's wire count is too low")); mpde->wire_count--; CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 1) | PG_PS | PG_V; if ((m->oflags & VPO_UNMANAGED) == 0) { newpde |= PG_MANAGED; /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_pde(pmap, va, VM_PAGE_TO_PHYS(m), lockp)) { SLIST_INIT(&free); if (pmap_unwire_ptp(pmap, va, mpde, &free)) { /* * Although "va" is not mapped, paging- * structure caches could nonetheless have * entries that refer to the freed page table * pages. Invalidate those entries. */ pmap_invalidate_page(pmap, va); pmap_free_zero_pages(&free); } CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } } if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (va < VM_MAXUSER_ADDRESS) newpde |= PG_U; /* * Increment counters. */ pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE); /* * Map the superpage. */ pde_store(pde, newpde); atomic_add_long(&pmap_pde_mappings, 1); CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx" " in pmap %p", va, pmap); return (TRUE); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { struct rwlock *lock; vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; lock = NULL; PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & PDRMASK) == 0 && va + NBPDR <= end && m->psind == 1 && pmap_ps_enabled(pmap) && pmap_enter_pde(pmap, va, m, prot, &lock)) m = &m[NBPDR / PAGE_SIZE - 1]; else mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte, &lock); m = TAILQ_NEXT(m, listq); } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { struct rwlock *lock; lock = NULL; PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp) { struct spglist free; pt_entry_t *pte, PG_V; vm_paddr_t pa; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { vm_pindex_t ptepindex; pd_entry_t *ptepa; /* * Calculate pagetable page index */ ptepindex = pmap_pde_pindex(va); if (mpte && (mpte->pindex == ptepindex)) { mpte->wire_count++; } else { /* * Get the page directory entry */ ptepa = pmap_pde(pmap, va); /* * If the page table page is mapped, we just increment * the hold count, and activate it. Otherwise, we * attempt to allocate a page table page. If this * attempt fails, we don't retry. Instead, we give up. */ if (ptepa && (*ptepa & PG_V) != 0) { if (*ptepa & PG_PS) return (NULL); mpte = PHYS_TO_VM_PAGE(*ptepa & PG_FRAME); mpte->wire_count++; } else { /* * Pass NULL instead of the PV list lock * pointer, because we don't intend to sleep. */ mpte = _pmap_allocpte(pmap, ptepindex, NULL); if (mpte == NULL) return (mpte); } } pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte)); pte = &pte[pmap_pte_index(va)]; } else { mpte = NULL; pte = vtopte(va); } if (*pte) { if (mpte != NULL) { mpte->wire_count--; mpte = NULL; } return (mpte); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m, lockp)) { if (mpte != NULL) { SLIST_INIT(&free); if (pmap_unwire_ptp(pmap, va, mpte, &free)) { /* * Although "va" is not mapped, paging- * structure caches could nonetheless have * entries that refer to the freed page table * pages. Invalidate those entries. */ pmap_invalidate_page(pmap, va); pmap_free_zero_pages(&free); } mpte = NULL; } return (mpte); } /* * Increment counters */ pmap_resident_count_inc(pmap, 1); pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 0); if ((prot & VM_PROT_EXECUTE) == 0) pa |= pg_nx; /* * Now validate mapping with RO protection */ if ((m->oflags & VPO_UNMANAGED) != 0) pte_store(pte, pa | PG_V | PG_U); else pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); return (mpte); } /* * Make a temporary mapping for a physical address. This is only intended * to be used for panic dumps. */ void * pmap_kenter_temporary(vm_paddr_t pa, int i) { vm_offset_t va; va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); pmap_kenter(va, pa); invlpg(va); return ((void *)crashdumpmap); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { pd_entry_t *pde; pt_entry_t PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pa, ptepa; vm_page_t p, pdpg; int pat_mode; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); if ((addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) { if (!pmap_ps_enabled(pmap)) return; if (!vm_object_populate(object, pindex, pindex + atop(size))) return; p = vm_page_lookup(object, pindex); KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); pat_mode = p->md.pat_mode; /* * Abort the mapping if the first page is not physically * aligned to a 2MB page boundary. */ ptepa = VM_PAGE_TO_PHYS(p); if (ptepa & (NBPDR - 1)) return; /* * Skip the first page. Abort the mapping if the rest of * the pages are not physically contiguous or have differing * memory attributes. */ p = TAILQ_NEXT(p, listq); for (pa = ptepa + PAGE_SIZE; pa < ptepa + size; pa += PAGE_SIZE) { KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); if (pa != VM_PAGE_TO_PHYS(p) || pat_mode != p->md.pat_mode) return; p = TAILQ_NEXT(p, listq); } /* * Map using 2MB pages. Since "ptepa" is 2M aligned and * "size" is a multiple of 2M, adding the PAT setting to "pa" * will not affect the termination of this loop. */ PMAP_LOCK(pmap); for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, 1); pa < ptepa + size; pa += NBPDR) { pdpg = pmap_allocpde(pmap, addr, NULL); if (pdpg == NULL) { /* * The creation of mappings below is only an * optimization. If a page directory page * cannot be allocated without blocking, * continue on to the next mapping rather than * blocking. */ addr += NBPDR; continue; } pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg)); pde = &pde[pmap_pde_index(addr)]; if ((*pde & PG_V) == 0) { pde_store(pde, pa | PG_PS | PG_M | PG_A | PG_U | PG_RW | PG_V); pmap_resident_count_inc(pmap, NBPDR / PAGE_SIZE); atomic_add_long(&pmap_pde_mappings, 1); } else { /* Continue on if the PDE is already valid. */ pdpg->wire_count--; KASSERT(pdpg->wire_count > 0, ("pmap_object_init_pt: missing reference " "to page directory page, va: 0x%lx", addr)); } addr += NBPDR; } PMAP_UNLOCK(pmap); } } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware * feature, so there is no need to invalidate any TLB entries. * Since pmap_demote_pde() for the wired entry must never fail, * pmap_delayed_invl_started()/finished() calls around the * function are not needed. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte, PG_V; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if ((*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) { va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); if ((*pde & PG_V) == 0) continue; if ((*pde & PG_PS) != 0) { if ((*pde & PG_W) == 0) panic("pmap_unwire: pde %#jx is missing PG_W", (uintmax_t)*pde); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { atomic_clear_long(pde, PG_W); pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; continue; } else if (!pmap_demote_pde(pmap, pde, sva)) panic("pmap_unwire: demotion failed"); } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if ((*pte & PG_V) == 0) continue; if ((*pte & PG_W) == 0) panic("pmap_unwire: pte %#jx is missing PG_W", (uintmax_t)*pte); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. */ atomic_clear_long(pte, PG_W); pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct rwlock *lock; struct spglist free; vm_offset_t addr; vm_offset_t end_addr = src_addr + len; vm_offset_t va_next; pt_entry_t PG_A, PG_M, PG_V; if (dst_addr != src_addr) return; if (dst_pmap->pm_type != src_pmap->pm_type) return; /* * EPT page table entries that require emulation of A/D bits are * sensitive to clearing the PG_A bit (aka EPT_PG_READ). Although * we clear PG_M (aka EPT_PG_WRITE) concomitantly, the PG_U bit * (aka EPT_PG_EXECUTE) could still be set. Since some EPT * implementations flag an EPT misconfiguration for exec-only * mappings we skip this function entirely for emulated pmaps. */ if (pmap_emulate_ad_bits(dst_pmap)) return; lock = NULL; if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } PG_A = pmap_accessed_bit(dst_pmap); PG_M = pmap_modified_bit(dst_pmap); PG_V = pmap_valid_bit(dst_pmap); for (addr = src_addr; addr < end_addr; addr = va_next) { pt_entry_t *src_pte, *dst_pte; vm_page_t dstmpde, dstmpte, srcmpte; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t srcptepaddr, *pde; KASSERT(addr < UPT_MIN_ADDRESS, ("pmap_copy: invalid to pmap_copy page tables")); pml4e = pmap_pml4e(src_pmap, addr); if ((*pml4e & PG_V) == 0) { va_next = (addr + NBPML4) & ~PML4MASK; if (va_next < addr) va_next = end_addr; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, addr); if ((*pdpe & PG_V) == 0) { va_next = (addr + NBPDP) & ~PDPMASK; if (va_next < addr) va_next = end_addr; continue; } va_next = (addr + NBPDR) & ~PDRMASK; if (va_next < addr) va_next = end_addr; pde = pmap_pdpe_to_pde(pdpe, addr); srcptepaddr = *pde; if (srcptepaddr == 0) continue; if (srcptepaddr & PG_PS) { if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr) continue; dstmpde = pmap_allocpde(dst_pmap, addr, NULL); if (dstmpde == NULL) break; pde = (pd_entry_t *) PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpde)); pde = &pde[pmap_pde_index(addr)]; if (*pde == 0 && ((srcptepaddr & PG_MANAGED) == 0 || pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr & PG_PS_FRAME, &lock))) { *pde = srcptepaddr & ~PG_W; pmap_resident_count_inc(dst_pmap, NBPDR / PAGE_SIZE); atomic_add_long(&pmap_pde_mappings, 1); } else dstmpde->wire_count--; continue; } srcptepaddr &= PG_FRAME; srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); KASSERT(srcmpte->wire_count > 0, ("pmap_copy: source page table page is unused")); if (va_next > end_addr) va_next = end_addr; src_pte = (pt_entry_t *)PHYS_TO_DMAP(srcptepaddr); src_pte = &src_pte[pmap_pte_index(addr)]; dstmpte = NULL; while (addr < va_next) { pt_entry_t ptetemp; ptetemp = *src_pte; /* * we only virtual copy managed pages */ if ((ptetemp & PG_MANAGED) != 0) { if (dstmpte != NULL && dstmpte->pindex == pmap_pde_pindex(addr)) dstmpte->wire_count++; else if ((dstmpte = pmap_allocpte(dst_pmap, addr, NULL)) == NULL) goto out; dst_pte = (pt_entry_t *) PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte)); dst_pte = &dst_pte[pmap_pte_index(addr)]; if (*dst_pte == 0 && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(ptetemp & PG_FRAME), &lock)) { /* * Clear the wired, modified, and * accessed (referenced) bits * during the copy. */ *dst_pte = ptetemp & ~(PG_W | PG_M | PG_A); pmap_resident_count_inc(dst_pmap, 1); } else { SLIST_INIT(&free); if (pmap_unwire_ptp(dst_pmap, addr, dstmpte, &free)) { /* * Although "addr" is not * mapped, paging-structure * caches could nonetheless * have entries that refer to * the freed page table pages. * Invalidate those entries. */ pmap_invalidate_page(dst_pmap, addr); pmap_free_zero_pages(&free); } goto out; } if (dstmpte->wire_count >= srcmpte->wire_count) break; } addr += PAGE_SIZE; src_pte++; } } out: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } /* * Zero the specified hardware page. */ void pmap_zero_page(vm_page_t m) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); pagezero((void *)va); } /* * Zero an an area within a single hardware page. off and size must not * cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); if (off == 0 && size == PAGE_SIZE) pagezero((void *)va); else bzero((char *)va + off, size); } /* * Copy 1 specified hardware page to another. */ void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc)); vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst)); pagecopy((void *)src, (void *)dst); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_page_t pages[2]; vm_offset_t vaddr[2], a_pg_offset, b_pg_offset; int cnt; boolean_t mapped; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; pages[0] = ma[a_offset >> PAGE_SHIFT]; b_pg_offset = b_offset & PAGE_MASK; pages[1] = mb[b_offset >> PAGE_SHIFT]; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); cnt = min(cnt, PAGE_SIZE - b_pg_offset); mapped = pmap_map_io_transient(pages, vaddr, 2, FALSE); a_cp = (char *)vaddr[0] + a_pg_offset; b_cp = (char *)vaddr[1] + b_pg_offset; bcopy(a_cp, b_cp, cnt); if (__predict_false(mapped)) pmap_unmap_io_transient(pages, vaddr, 2, FALSE); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; struct rwlock *lock; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_runlock(lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct rwlock *lock; struct md_page *pvh; pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; int count, md_gen, pvh_gen; if ((m->oflags & VPO_UNMANAGED) != 0) return (0); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: count = 0; TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pde(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } } rw_runlock(lock); return (count); } /* * Returns TRUE if the given page is mapped individually or as part of * a 2mpage. Otherwise, returns FALSE. */ boolean_t pmap_page_is_mapped(vm_page_t m) { struct rwlock *lock; boolean_t rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (FALSE); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list)); rw_runlock(lock); return (rv); } /* * Destroy all managed, non-wired mappings in the given user-space * pmap. This pmap cannot be active on any processor besides the * caller. * * This function cannot be applied to the kernel pmap. Moreover, it * is not intended for general use. It is only to be used during * process termination. Consequently, it can be implemented in ways * that make it faster than pmap_remove(). First, it can more quickly * destroy mappings by iterating over the pmap's collection of PV * entries, rather than searching the page table. Second, it doesn't * have to test and clear the page table entries atomically, because * no processor is currently accessing the user address space. In * particular, a page table entry's dirty bit won't change state once * this function starts. */ void pmap_remove_pages(pmap_t pmap) { pd_entry_t ptepde; pt_entry_t *pte, tpte; pt_entry_t PG_M, PG_RW, PG_V; struct spglist free; vm_page_t m, mpte, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct rwlock *lock; int64_t bit; uint64_t inuse, bitmask; int allfree, field, freed, idx; boolean_t superpage; vm_paddr_t pa; /* * Assert that the given pmap is only active on the current * CPU. Unfortunately, we cannot block another CPU from * activating the pmap while this function is executing. */ KASSERT(pmap == PCPU_GET(curpmap), ("non-current pmap %p", pmap)); #ifdef INVARIANTS { cpuset_t other_cpus; other_cpus = all_cpus; critical_enter(); CPU_CLR(PCPU_GET(cpuid), &other_cpus); CPU_AND(&other_cpus, &pmap->pm_active); critical_exit(); KASSERT(CPU_EMPTY(&other_cpus), ("pmap active %p", pmap)); } #endif lock = NULL; PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); SLIST_INIT(&free); PMAP_LOCK(pmap); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { allfree = 1; freed = 0; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = bsfq(inuse); bitmask = 1UL << bit; idx = field * 64 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pte = pmap_pdpe(pmap, pv->pv_va); ptepde = *pte; pte = pmap_pdpe_to_pde(pte, pv->pv_va); tpte = *pte; if ((tpte & (PG_PS | PG_V)) == PG_V) { superpage = FALSE; ptepde = tpte; pte = (pt_entry_t *)PHYS_TO_DMAP(tpte & PG_FRAME); pte = &pte[pmap_pte_index(pv->pv_va)]; tpte = *pte; } else { /* * Keep track whether 'tpte' is a * superpage explicitly instead of * relying on PG_PS being set. * * This is because PG_PS is numerically * identical to PG_PTE_PAT and thus a * regular page could be mistaken for * a superpage. */ superpage = TRUE; } if ((tpte & PG_V) == 0) { panic("bad pte va %lx pte %lx", pv->pv_va, tpte); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & PG_W) { allfree = 0; continue; } if (superpage) pa = tpte & PG_PS_FRAME; else pa = tpte & PG_FRAME; m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); pte_clear(pte); /* * Update the vm_page_t clean/reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (superpage) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } else vm_page_dirty(m); } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); /* Mark free */ pc->pc_map[field] |= bitmask; if (superpage) { pmap_resident_count_dec(pmap, NBPDR / PAGE_SIZE); pvh = pa_to_pvh(tpte & PG_PS_FRAME); TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) if ((mt->aflags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } mpte = pmap_lookup_pt_page(pmap, pv->pv_va); if (mpte != NULL) { pmap_remove_pt_page(pmap, mpte); pmap_resident_count_dec(pmap, 1); KASSERT(mpte->wire_count == NPTEPG, ("pmap_remove_pages: pte page wire count error")); mpte->wire_count = 0; pmap_add_delayed_free_list(mpte, &free, FALSE); atomic_subtract_int(&vm_cnt.v_wire_count, 1); } } else { pmap_resident_count_dec(pmap, 1); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((m->aflags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } pmap_unuse_pt(pmap, pv->pv_va, ptepde, &free); freed++; } } PV_STAT(atomic_add_long(&pv_entry_frees, freed)); PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } } if (lock != NULL) rw_wunlock(lock); pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } static boolean_t pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified) { struct rwlock *lock; pv_entry_t pv; struct md_page *pvh; pt_entry_t *pte, mask; pt_entry_t PG_A, PG_M, PG_RW, PG_V; pmap_t pmap; int md_gen, pvh_gen; boolean_t rv; rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va); mask = 0; if (modified) { PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); mask |= PG_RW | PG_M; } if (accessed) { PG_A = pmap_accessed_bit(pmap); PG_V = pmap_valid_bit(pmap); mask |= PG_V | PG_A; } rv = (*pte & mask) == mask; PMAP_UNLOCK(pmap); if (rv) goto out; } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pde(pmap, pv->pv_va); mask = 0; if (modified) { PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); mask |= PG_RW | PG_M; } if (accessed) { PG_A = pmap_accessed_bit(pmap); PG_V = pmap_valid_bit(pmap); mask |= PG_V | PG_A; } rv = (*pte & mask) == mask; PMAP_UNLOCK(pmap); if (rv) goto out; } } out: rw_runlock(lock); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no PTEs can have PG_M set. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (FALSE); return (pmap_page_test_mappings(m, FALSE, TRUE)); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pd_entry_t *pde; pt_entry_t *pte, PG_V; boolean_t rv; PG_V = pmap_valid_bit(pmap); rv = FALSE; PMAP_LOCK(pmap); pde = pmap_pde(pmap, addr); if (pde != NULL && (*pde & (PG_PS | PG_V)) == PG_V) { pte = pmap_pde_to_pte(pde, addr); rv = (*pte & PG_V) == 0; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); return (pmap_page_test_mappings(m, TRUE, FALSE)); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pmap_t pmap; struct rwlock *lock; pv_entry_t next_pv, pv; pd_entry_t *pde; pt_entry_t oldpte, *pte, PG_M, PG_RW; vm_offset_t va; int pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); retry_pv_loop: rw_wlock(lock); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); rw_wunlock(lock); goto retry_pv_loop; } } PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_RW) != 0) (void)pmap_demote_pde_locked(pmap, pde, va, &lock); KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); rw_wunlock(lock); goto retry_pv_loop; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_write: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); retry: oldpte = *pte; if (oldpte & PG_RW) { if (!atomic_cmpset_long(pte, oldpte, oldpte & ~(PG_RW | PG_M))) goto retry; if ((oldpte & PG_M) != 0) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); vm_page_aflag_clear(m, PGA_WRITEABLE); pmap_delayed_invl_wait(m); } static __inline boolean_t safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte) { if (!pmap_emulate_ad_bits(pmap)) return (TRUE); KASSERT(pmap->pm_type == PT_EPT, ("invalid pm_type %d", pmap->pm_type)); /* * XWR = 010 or 110 will cause an unconditional EPT misconfiguration * so we don't let the referenced (aka EPT_PG_READ) bit to be cleared * if the EPT_PG_WRITE bit is set. */ if ((pte & EPT_PG_WRITE) != 0) return (FALSE); /* * XWR = 100 is allowed only if the PMAP_SUPPORTS_EXEC_ONLY is set. */ if ((pte & EPT_PG_EXECUTE) == 0 || ((pmap->pm_flags & PMAP_SUPPORTS_EXEC_ONLY) != 0)) return (TRUE); else return (FALSE); } -#define PMAP_TS_REFERENCED_MAX 5 - /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. - * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). * * A DI block is not needed within this function, because * invalidations are performed before the PV list lock is * released. */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; struct rwlock *lock; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_A, PG_M, PG_RW; vm_offset_t va; vm_paddr_t pa; int cleared, md_gen, not_cleared, pvh_gen; struct spglist free; boolean_t demoted; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); SLIST_INIT(&free); cleared = 0; pa = VM_PAGE_TO_PHYS(m); lock = PHYS_TO_PV_LIST_LOCK(pa); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa); rw_wlock(lock); retry: not_cleared = 0; if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, pv->pv_va); oldpde = *pde; if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Although "oldpde" is mapping a 2MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } - if ((*pde & PG_A) != 0) { + if ((oldpde & PG_A) != 0) { /* * Since this reference bit is shared by 512 4KB * pages, it should not be cleared every time it is * tested. Apply a simple "hash" function on the * physical page number, the virtual superpage number, * and the pmap address to select one 4KB page out of * the 512 on which testing the reference bit will * result in clearing that reference bit. This * function is designed to avoid the selection of the * same 4KB page for every 2MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^ (uintptr_t)pmap) & (NPTEPG - 1)) == 0 && - (*pde & PG_W) == 0) { + (oldpde & PG_W) == 0) { if (safe_to_clear_referenced(pmap, oldpde)) { atomic_clear_long(pde, PG_A); pmap_invalidate_page(pmap, pv->pv_va); demoted = FALSE; } else if (pmap_demote_pde_locked(pmap, pde, pv->pv_va, &lock)) { /* * Remove the mapping to a single page * so that a subsequent access may * repromote. Since the underlying * page table page is fully populated, * this removal never frees a page * table page. */ demoted = TRUE; va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pde_to_pte(pde, va); pmap_remove_pte(pmap, pte, va, *pde, NULL, &lock); pmap_invalidate_page(pmap, va); } else demoted = TRUE; if (demoted) { /* * The superpage mapping was removed * entirely and therefore 'pv' is no * longer valid. */ if (pvf == pv) pvf = NULL; pv = NULL; } cleared++; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; } if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((*pte & PG_A) != 0) { if (safe_to_clear_referenced(pmap, *pte)) { atomic_clear_long(pte, PG_A); pmap_invalidate_page(pmap, pv->pv_va); cleared++; } else if ((*pte & PG_W) == 0) { /* * Wired pages cannot be paged out so * doing accessed bit emulation for * them is wasted effort. We do the * hard work for unwired pages only. */ pmap_remove_pte(pmap, pte, pv->pv_va, *pde, &free, &lock); pmap_invalidate_page(pmap, pv->pv_va); cleared++; if (pvf == pv) pvf = NULL; pv = NULL; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared + not_cleared < PMAP_TS_REFERENCED_MAX); out: rw_wunlock(lock); pmap_free_zero_pages(&free); return (cleared + not_cleared); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { struct rwlock *lock; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_A, PG_G, PG_M, PG_RW, PG_V; vm_offset_t va_next; vm_page_t m; boolean_t anychanged; if (advice != MADV_DONTNEED && advice != MADV_FREE) return; /* * A/D bit emulation requires an alternate code path when clearing * the modified and accessed bits below. Since this function is * advisory in nature we skip it entirely for pmaps that require * A/D bit emulation. */ if (pmap_emulate_ad_bits(pmap)) return; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); anychanged = FALSE; pmap_delayed_invl_started(); PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if ((*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) { va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); oldpde = *pde; if ((oldpde & PG_V) == 0) continue; else if ((oldpde & PG_PS) != 0) { if ((oldpde & PG_MANAGED) == 0) continue; lock = NULL; if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) { if (lock != NULL) rw_wunlock(lock); /* * The large page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Since the underlying page * table page is fully populated, this removal never * frees a page table page. */ if ((oldpde & PG_W) == 0) { pte = pmap_pde_to_pte(pde, sva); KASSERT((*pte & PG_V) != 0, ("pmap_advise: invalid PTE")); pmap_remove_pte(pmap, pte, sva, *pde, NULL, &lock); anychanged = TRUE; } if (lock != NULL) rw_wunlock(lock); } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V)) continue; else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); vm_page_dirty(m); } atomic_clear_long(pte, PG_M | PG_A); } else if ((*pte & PG_A) != 0) atomic_clear_long(pte, PG_A); else continue; if ((*pte & PG_G) != 0) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } } if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_delayed_invl_finished(); } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { struct md_page *pvh; pmap_t pmap; pv_entry_t next_pv, pv; pd_entry_t oldpde, *pde; pt_entry_t oldpte, *pte, PG_M, PG_RW, PG_V; struct rwlock *lock; vm_offset_t va; int md_gen, pvh_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("pmap_clear_modify: page %p is exclusive busied", m)); /* * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set. * If the object containing the page is locked and the page is not * exclusive busied, then PGA_WRITEABLE cannot be concurrently set. */ if ((m->aflags & PGA_WRITEABLE) == 0) return; pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_wlock(lock); restart: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, va); oldpde = *pde; if ((oldpde & PG_RW) != 0) { if (pmap_demote_pde_locked(pmap, pde, va, &lock)) { if ((oldpde & PG_W) == 0) { /* * Write protect the mapping to a * single page so that a subsequent * write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pde_to_pte(pde, va); oldpte = *pte; if ((oldpte & PG_V) != 0) { while (!atomic_cmpset_long(pte, oldpte, oldpte & ~(PG_M | PG_RW))) oldpte = *pte; vm_page_dirty(m); pmap_invalidate_page(pmap, va); } } } } PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found" " a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { atomic_clear_long(pte, PG_M); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); } /* * Miscellaneous support routines follow */ /* Adjust the cache mode for a 4KB page mapped via a PTE. */ static __inline void pmap_pte_attr(pt_entry_t *pte, int cache_bits, int mask) { u_int opte, npte; /* * The cache mode bits are all in the low 32-bits of the * PTE, so we can just spin on updating the low 32-bits. */ do { opte = *(u_int *)pte; npte = opte & ~mask; npte |= cache_bits; } while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte)); } /* Adjust the cache mode for a 2MB page mapped via a PDE. */ static __inline void pmap_pde_attr(pd_entry_t *pde, int cache_bits, int mask) { u_int opde, npde; /* * The cache mode bits are all in the low 32-bits of the * PDE, so we can just spin on updating the low 32-bits. */ do { opde = *(u_int *)pde; npde = opde & ~mask; npde |= cache_bits; } while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde)); } /* * Map a set of physical memory pages into the kernel virtual * address space. Return a pointer to where it is mapped. This * routine is intended to be used for mapping device memory, * NOT real memory. */ void * pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode) { struct pmap_preinit_mapping *ppim; vm_offset_t va, offset; vm_size_t tmpsize; int i; offset = pa & PAGE_MASK; size = round_page(offset + size); pa = trunc_page(pa); if (!pmap_initialized) { va = 0; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) { ppim->pa = pa; ppim->sz = size; ppim->mode = mode; ppim->va = virtual_avail; virtual_avail += size; va = ppim->va; break; } } if (va == 0) panic("%s: too many preinit mappings", __func__); } else { /* * If we have a preinit mapping, re-use it. */ for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->pa == pa && ppim->sz == size && ppim->mode == mode) return ((void *)(ppim->va + offset)); } /* * If the specified range of physical addresses fits within * the direct map window, use the direct map. */ if (pa < dmaplimit && pa + size < dmaplimit) { va = PHYS_TO_DMAP(pa); if (!pmap_change_attr(va, size, mode)) return ((void *)(va + offset)); } va = kva_alloc(size); if (va == 0) panic("%s: Couldn't allocate KVA", __func__); } for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode); pmap_invalidate_range(kernel_pmap, va, va + tmpsize); pmap_invalidate_cache_range(va, va + tmpsize, FALSE); return ((void *)(va + offset)); } void * pmap_mapdev(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE)); } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK)); } void pmap_unmapdev(vm_offset_t va, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t offset; int i; /* If we gave a direct map region in pmap_mapdev, do nothing */ if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) return; offset = va & PAGE_MASK; size = round_page(offset + size); va = trunc_page(va); for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == va && ppim->sz == size) { if (pmap_initialized) return; ppim->pa = 0; ppim->va = 0; ppim->sz = 0; ppim->mode = 0; if (va + size == virtual_avail) virtual_avail = va; return; } } if (pmap_initialized) kva_free(va, size); } /* * Tries to demote a 1GB page mapping. */ static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va) { pdp_entry_t newpdpe, oldpdpe; pd_entry_t *firstpde, newpde, *pde; pt_entry_t PG_A, PG_M, PG_RW, PG_V; vm_paddr_t mpdepa; vm_page_t mpde; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpdpe = *pdpe; KASSERT((oldpdpe & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pdpe: oldpdpe is missing PG_PS and/or PG_V")); if ((mpde = vm_page_alloc(NULL, va >> PDPSHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { CTR2(KTR_PMAP, "pmap_demote_pdpe: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } mpdepa = VM_PAGE_TO_PHYS(mpde); firstpde = (pd_entry_t *)PHYS_TO_DMAP(mpdepa); newpdpe = mpdepa | PG_M | PG_A | (oldpdpe & PG_U) | PG_RW | PG_V; KASSERT((oldpdpe & PG_A) != 0, ("pmap_demote_pdpe: oldpdpe is missing PG_A")); KASSERT((oldpdpe & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pdpe: oldpdpe is missing PG_M")); newpde = oldpdpe; /* * Initialize the page directory page. */ for (pde = firstpde; pde < firstpde + NPDEPG; pde++) { *pde = newpde; newpde += NBPDR; } /* * Demote the mapping. */ *pdpe = newpdpe; /* * Invalidate a stale recursive mapping of the page directory page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopde(va)); pmap_pdpe_demotions++; CTR2(KTR_PMAP, "pmap_demote_pdpe: success for va %#lx" " in pmap %p", va, pmap); return (TRUE); } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pat_mode = ma; /* * If "m" is a normal page, update its direct mapping. This update * can be relied upon to perform any cache operations that are * required for data coherence. */ if ((m->flags & PG_FICTITIOUS) == 0 && pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, m->md.pat_mode)) panic("memory attribute change on the direct map failed"); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the direct map or the kernel map. If * the virtual address range is contained within the kernel map, then the * memory type for each of the corresponding ranges of the direct map is also * changed. (The corresponding ranges of the direct map are those ranges that * map the same physical pages as the specified virtual address range.) These * changes to the direct map are necessary because Intel describes the * behavior of their processors as "undefined" if two or more mappings to the * same physical page have different memory types. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. In the * latter case, the memory type may have been changed on some part of the * virtual address range or the direct map. */ int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) { int error; PMAP_LOCK(kernel_pmap); error = pmap_change_attr_locked(va, size, mode); PMAP_UNLOCK(kernel_pmap); return (error); } static int pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode) { vm_offset_t base, offset, tmpva; vm_paddr_t pa_start, pa_end, pa_end1; pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; int cache_bits_pte, cache_bits_pde, error; boolean_t changed; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); /* * Only supported on kernel virtual addresses, including the direct * map but excluding the recursive map. */ if (base < DMAP_MIN_ADDRESS) return (EINVAL); cache_bits_pde = pmap_cache_bits(kernel_pmap, mode, 1); cache_bits_pte = pmap_cache_bits(kernel_pmap, mode, 0); changed = FALSE; /* * Pages that aren't mapped aren't supported. Also break down 2MB pages * into 4KB pages if required. */ for (tmpva = base; tmpva < base + size; ) { pdpe = pmap_pdpe(kernel_pmap, tmpva); if (pdpe == NULL || *pdpe == 0) return (EINVAL); if (*pdpe & PG_PS) { /* * If the current 1GB page already has the required * memory type, then we need not demote this page. Just * increment tmpva to the next 1GB page frame. */ if ((*pdpe & X86_PG_PDE_CACHE) == cache_bits_pde) { tmpva = trunc_1gpage(tmpva) + NBPDP; continue; } /* * If the current offset aligns with a 1GB page frame * and there is at least 1GB left within the range, then * we need not break down this page into 2MB pages. */ if ((tmpva & PDPMASK) == 0 && tmpva + PDPMASK < base + size) { tmpva += NBPDP; continue; } if (!pmap_demote_pdpe(kernel_pmap, pdpe, tmpva)) return (ENOMEM); } pde = pmap_pdpe_to_pde(pdpe, tmpva); if (*pde == 0) return (EINVAL); if (*pde & PG_PS) { /* * If the current 2MB page already has the required * memory type, then we need not demote this page. Just * increment tmpva to the next 2MB page frame. */ if ((*pde & X86_PG_PDE_CACHE) == cache_bits_pde) { tmpva = trunc_2mpage(tmpva) + NBPDR; continue; } /* * If the current offset aligns with a 2MB page frame * and there is at least 2MB left within the range, then * we need not break down this page into 4KB pages. */ if ((tmpva & PDRMASK) == 0 && tmpva + PDRMASK < base + size) { tmpva += NBPDR; continue; } if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) return (ENOMEM); } pte = pmap_pde_to_pte(pde, tmpva); if (*pte == 0) return (EINVAL); tmpva += PAGE_SIZE; } error = 0; /* * Ok, all the pages exist, so run through them updating their * cache mode if required. */ pa_start = pa_end = 0; for (tmpva = base; tmpva < base + size; ) { pdpe = pmap_pdpe(kernel_pmap, tmpva); if (*pdpe & PG_PS) { if ((*pdpe & X86_PG_PDE_CACHE) != cache_bits_pde) { pmap_pde_attr(pdpe, cache_bits_pde, X86_PG_PDE_CACHE); changed = TRUE; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pdpe & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pdpe & PG_PS_FRAME; pa_end = pa_start + NBPDP; } else if (pa_end == (*pdpe & PG_PS_FRAME)) pa_end += NBPDP; else { /* Run ended, update direct map. */ error = pmap_change_attr_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, mode); if (error != 0) break; /* Start physical address run. */ pa_start = *pdpe & PG_PS_FRAME; pa_end = pa_start + NBPDP; } } tmpva = trunc_1gpage(tmpva) + NBPDP; continue; } pde = pmap_pdpe_to_pde(pdpe, tmpva); if (*pde & PG_PS) { if ((*pde & X86_PG_PDE_CACHE) != cache_bits_pde) { pmap_pde_attr(pde, cache_bits_pde, X86_PG_PDE_CACHE); changed = TRUE; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pde & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pde & PG_PS_FRAME; pa_end = pa_start + NBPDR; } else if (pa_end == (*pde & PG_PS_FRAME)) pa_end += NBPDR; else { /* Run ended, update direct map. */ error = pmap_change_attr_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, mode); if (error != 0) break; /* Start physical address run. */ pa_start = *pde & PG_PS_FRAME; pa_end = pa_start + NBPDR; } } tmpva = trunc_2mpage(tmpva) + NBPDR; } else { pte = pmap_pde_to_pte(pde, tmpva); if ((*pte & X86_PG_PTE_CACHE) != cache_bits_pte) { pmap_pte_attr(pte, cache_bits_pte, X86_PG_PTE_CACHE); changed = TRUE; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pte & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pte & PG_FRAME; pa_end = pa_start + PAGE_SIZE; } else if (pa_end == (*pte & PG_FRAME)) pa_end += PAGE_SIZE; else { /* Run ended, update direct map. */ error = pmap_change_attr_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, mode); if (error != 0) break; /* Start physical address run. */ pa_start = *pte & PG_FRAME; pa_end = pa_start + PAGE_SIZE; } } tmpva += PAGE_SIZE; } } if (error == 0 && pa_start != pa_end && pa_start < dmaplimit) { pa_end1 = MIN(pa_end, dmaplimit); if (pa_start != pa_end1) error = pmap_change_attr_locked(PHYS_TO_DMAP(pa_start), pa_end1 - pa_start, mode); } /* * Flush CPU caches if required to make sure any data isn't cached that * shouldn't be, etc. */ if (changed) { pmap_invalidate_range(kernel_pmap, base, tmpva); pmap_invalidate_cache_range(base, tmpva, FALSE); } return (error); } /* * Demotes any mapping within the direct map region that covers more than the * specified range of physical addresses. This range's size must be a power * of two and its starting address must be a multiple of its size. Since the * demotion does not change any attributes of the mapping, a TLB invalidation * is not mandatory. The caller may, however, request a TLB invalidation. */ void pmap_demote_DMAP(vm_paddr_t base, vm_size_t len, boolean_t invalidate) { pdp_entry_t *pdpe; pd_entry_t *pde; vm_offset_t va; boolean_t changed; if (len == 0) return; KASSERT(powerof2(len), ("pmap_demote_DMAP: len is not a power of 2")); KASSERT((base & (len - 1)) == 0, ("pmap_demote_DMAP: base is not a multiple of len")); if (len < NBPDP && base < dmaplimit) { va = PHYS_TO_DMAP(base); changed = FALSE; PMAP_LOCK(kernel_pmap); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) panic("pmap_demote_DMAP: invalid PDPE"); if ((*pdpe & PG_PS) != 0) { if (!pmap_demote_pdpe(kernel_pmap, pdpe, va)) panic("pmap_demote_DMAP: PDPE failed"); changed = TRUE; } if (len < NBPDR) { pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) panic("pmap_demote_DMAP: invalid PDE"); if ((*pde & PG_PS) != 0) { if (!pmap_demote_pde(kernel_pmap, pde, va)) panic("pmap_demote_DMAP: PDE failed"); changed = TRUE; } } if (changed && invalidate) pmap_invalidate_page(kernel_pmap, va); PMAP_UNLOCK(kernel_pmap); } } /* * perform the pmap work for mincore */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa) { pd_entry_t *pdep; pt_entry_t pte, PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pa; int val; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK(pmap); retry: pdep = pmap_pde(pmap, addr); if (pdep != NULL && (*pdep & PG_V)) { if (*pdep & PG_PS) { pte = *pdep; /* Compute the physical address of the 4KB page. */ pa = ((*pdep & PG_PS_FRAME) | (addr & PDRMASK)) & PG_FRAME; val = MINCORE_SUPER; } else { pte = *pmap_pde_to_pte(pdep, addr); pa = pte & PG_FRAME; val = 0; } } else { pte = 0; pa = 0; val = 0; } if ((pte & PG_V) != 0) { val |= MINCORE_INCORE; if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((pte & PG_A) != 0) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) { /* Ensure that "PHYS_TO_VM_PAGE(pa)->object" doesn't change. */ if (vm_page_pa_tryrelock(pmap, pa, locked_pa)) goto retry; } else PA_UNLOCK_COND(*locked_pa); PMAP_UNLOCK(pmap); return (val); } static uint64_t pmap_pcid_alloc(pmap_t pmap, u_int cpuid) { uint32_t gen, new_gen, pcid_next; CRITICAL_ASSERT(curthread); gen = PCPU_GET(pcid_gen); if (pmap->pm_pcids[cpuid].pm_pcid == PMAP_PCID_KERN || pmap->pm_pcids[cpuid].pm_gen == gen) return (CR3_PCID_SAVE); pcid_next = PCPU_GET(pcid_next); KASSERT(pcid_next <= PMAP_PCID_OVERMAX, ("cpu %d pcid_next %#x", cpuid, pcid_next)); if (pcid_next == PMAP_PCID_OVERMAX) { new_gen = gen + 1; if (new_gen == 0) new_gen = 1; PCPU_SET(pcid_gen, new_gen); pcid_next = PMAP_PCID_KERN + 1; } else { new_gen = gen; } pmap->pm_pcids[cpuid].pm_pcid = pcid_next; pmap->pm_pcids[cpuid].pm_gen = new_gen; PCPU_SET(pcid_next, pcid_next + 1); return (0); } void pmap_activate_sw(struct thread *td) { pmap_t oldpmap, pmap; uint64_t cached, cr3; u_int cpuid; oldpmap = PCPU_GET(curpmap); pmap = vmspace_pmap(td->td_proc->p_vmspace); if (oldpmap == pmap) return; cpuid = PCPU_GET(cpuid); #ifdef SMP CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_SET(cpuid, &pmap->pm_active); #endif cr3 = rcr3(); if (pmap_pcid_enabled) { cached = pmap_pcid_alloc(pmap, cpuid); KASSERT(pmap->pm_pcids[cpuid].pm_pcid >= 0 && pmap->pm_pcids[cpuid].pm_pcid < PMAP_PCID_OVERMAX, ("pmap %p cpu %d pcid %#x", pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid)); KASSERT(pmap->pm_pcids[cpuid].pm_pcid != PMAP_PCID_KERN || pmap == kernel_pmap, ("non-kernel pmap thread %p pmap %p cpu %d pcid %#x", td, pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid)); if (!cached || (cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3) { load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid | cached); if (cached) PCPU_INC(pm_save_cnt); } } else if (cr3 != pmap->pm_cr3) { load_cr3(pmap->pm_cr3); } PCPU_SET(curpmap, pmap); #ifdef SMP CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active); #else CPU_CLR(cpuid, &oldpmap->pm_active); #endif } void pmap_activate(struct thread *td) { critical_enter(); pmap_activate_sw(td); critical_exit(); } void pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < NBPDR) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & PDRMASK; if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR || (*addr & PDRMASK) == superpage_offset) return; if ((*addr & PDRMASK) < superpage_offset) *addr = (*addr & ~PDRMASK) + superpage_offset; else *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset; } #ifdef INVARIANTS static unsigned long num_dirty_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_dirty_emulations, CTLFLAG_RW, &num_dirty_emulations, 0, NULL); static unsigned long num_accessed_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_accessed_emulations, CTLFLAG_RW, &num_accessed_emulations, 0, NULL); static unsigned long num_superpage_accessed_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_superpage_accessed_emulations, CTLFLAG_RW, &num_superpage_accessed_emulations, 0, NULL); static unsigned long ad_emulation_superpage_promotions; SYSCTL_ULONG(_vm_pmap, OID_AUTO, ad_emulation_superpage_promotions, CTLFLAG_RW, &ad_emulation_superpage_promotions, 0, NULL); #endif /* INVARIANTS */ int pmap_emulate_accessed_dirty(pmap_t pmap, vm_offset_t va, int ftype) { int rv; struct rwlock *lock; vm_page_t m, mpte; pd_entry_t *pde; pt_entry_t *pte, PG_A, PG_M, PG_RW, PG_V; KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE, ("pmap_emulate_accessed_dirty: invalid fault type %d", ftype)); if (!pmap_emulate_ad_bits(pmap)) return (-1); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); rv = -1; lock = NULL; PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); if (pde == NULL || (*pde & PG_V) == 0) goto done; if ((*pde & PG_PS) != 0) { if (ftype == VM_PROT_READ) { #ifdef INVARIANTS atomic_add_long(&num_superpage_accessed_emulations, 1); #endif *pde |= PG_A; rv = 0; } goto done; } pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_V) == 0) goto done; if (ftype == VM_PROT_WRITE) { if ((*pte & PG_RW) == 0) goto done; /* * Set the modified and accessed bits simultaneously. * * Intel EPT PTEs that do software emulation of A/D bits map * PG_A and PG_M to EPT_PG_READ and EPT_PG_WRITE respectively. * An EPT misconfiguration is triggered if the PTE is writable * but not readable (WR=10). This is avoided by setting PG_A * and PG_M simultaneously. */ *pte |= PG_M | PG_A; } else { *pte |= PG_A; } /* try to promote the mapping */ if (va < VM_MAXUSER_ADDRESS) mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); else mpte = NULL; m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); if ((mpte == NULL || mpte->wire_count == NPTEPG) && pmap_ps_enabled(pmap) && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) { pmap_promote_pde(pmap, pde, va, &lock); #ifdef INVARIANTS atomic_add_long(&ad_emulation_superpage_promotions, 1); #endif } #ifdef INVARIANTS if (ftype == VM_PROT_WRITE) atomic_add_long(&num_dirty_emulations, 1); else atomic_add_long(&num_accessed_emulations, 1); #endif rv = 0; /* success */ done: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (rv); } void pmap_get_mapping(pmap_t pmap, vm_offset_t va, uint64_t *ptr, int *num) { pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pde; pt_entry_t *pte, PG_V; int idx; idx = 0; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pml4 = pmap_pml4e(pmap, va); ptr[idx++] = *pml4; if ((*pml4 & PG_V) == 0) goto done; pdp = pmap_pml4e_to_pdpe(pml4, va); ptr[idx++] = *pdp; if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) goto done; pde = pmap_pdpe_to_pde(pdp, va); ptr[idx++] = *pde; if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) goto done; pte = pmap_pde_to_pte(pde, va); ptr[idx++] = *pte; done: PMAP_UNLOCK(pmap); *num = idx; } /** * Get the kernel virtual address of a set of physical pages. If there are * physical addresses not covered by the DMAP perform a transient mapping * that will be removed when calling pmap_unmap_io_transient. * * \param page The pages the caller wishes to obtain the virtual * address on the kernel memory map. * \param vaddr On return contains the kernel virtual memory address * of the pages passed in the page parameter. * \param count Number of pages passed in. * \param can_fault TRUE if the thread using the mapped pages can take * page faults, FALSE otherwise. * * \returns TRUE if the caller must call pmap_unmap_io_transient when * finished or FALSE otherwise. * */ boolean_t pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; boolean_t needs_mapping; pt_entry_t *pte; int cache_bits, error, i; /* * Allocate any KVA space that we need, this is done in a separate * loop to prevent calling vmem_alloc while pinned. */ needs_mapping = FALSE; for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (__predict_false(paddr >= dmaplimit)) { error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr[i]); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); needs_mapping = TRUE; } else { vaddr[i] = PHYS_TO_DMAP(paddr); } } /* Exit early if everything is covered by the DMAP */ if (!needs_mapping) return (FALSE); /* * NB: The sequence of updating a page table followed by accesses * to the corresponding pages used in the !DMAP case is subject to * the situation described in the "AMD64 Architecture Programmer's * Manual Volume 2: System Programming" rev. 3.23, "7.3.1 Special * Coherency Considerations". Therefore, issuing the INVLPG right * after modifying the PTE bits is crucial. */ if (!can_fault) sched_pin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= dmaplimit) { if (can_fault) { /* * Slow path, since we can get page faults * while mappings are active don't pin the * thread to the CPU and instead add a global * mapping visible to all CPUs. */ pmap_qenter(vaddr[i], &page[i], 1); } else { pte = vtopte(vaddr[i]); cache_bits = pmap_cache_bits(kernel_pmap, page[i]->md.pat_mode, 0); pte_store(pte, paddr | X86_PG_RW | X86_PG_V | cache_bits); invlpg(vaddr[i]); } } } return (needs_mapping); } void pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; int i; if (!can_fault) sched_unpin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= dmaplimit) { if (can_fault) pmap_qremove(vaddr[i], 1); vmem_free(kernel_arena, vaddr[i], PAGE_SIZE); } } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { vm_paddr_t paddr; paddr = VM_PAGE_TO_PHYS(m); if (paddr < dmaplimit) return (PHYS_TO_DMAP(paddr)); mtx_lock_spin(&qframe_mtx); KASSERT(*vtopte(qframe) == 0, ("qframe busy")); pte_store(vtopte(qframe), paddr | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0)); return (qframe); } void pmap_quick_remove_page(vm_offset_t addr) { if (addr != qframe) return; pte_store(vtopte(qframe), 0); invlpg(qframe); mtx_unlock_spin(&qframe_mtx); } #include "opt_ddb.h" #ifdef DDB #include DB_SHOW_COMMAND(pte, pmap_print_pte) { pmap_t pmap; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pde; pt_entry_t *pte, PG_V; vm_offset_t va; if (have_addr) { va = (vm_offset_t)addr; pmap = PCPU_GET(curpmap); /* XXX */ } else { db_printf("show pte addr\n"); return; } PG_V = pmap_valid_bit(pmap); pml4 = pmap_pml4e(pmap, va); db_printf("VA %#016lx pml4e %#016lx", va, *pml4); if ((*pml4 & PG_V) == 0) { db_printf("\n"); return; } pdp = pmap_pml4e_to_pdpe(pml4, va); db_printf(" pdpe %#016lx", *pdp); if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) { db_printf("\n"); return; } pde = pmap_pdpe_to_pde(pdp, va); db_printf(" pde %#016lx", *pde); if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) { db_printf("\n"); return; } pte = pmap_pde_to_pte(pde, va); db_printf(" pte %#016lx\n", *pte); } DB_SHOW_COMMAND(phys2dmap, pmap_phys2dmap) { vm_paddr_t a; if (have_addr) { a = (vm_paddr_t)addr; db_printf("0x%jx\n", (uintmax_t)PHYS_TO_DMAP(a)); } else { db_printf("show phys2dmap addr\n"); } } #endif Index: projects/mips64-clang/sys/amd64/include/fpu.h =================================================================== --- projects/mips64-clang/sys/amd64/include/fpu.h (revision 305711) +++ projects/mips64-clang/sys/amd64/include/fpu.h (revision 305712) @@ -1,92 +1,93 @@ /*- * Copyright (c) 1990 The Regents of the University of California. * All rights reserved. * * This code is derived from software contributed to Berkeley by * William Jolitz. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)npx.h 5.3 (Berkeley) 1/18/91 * $FreeBSD$ */ /* * Floating Point Data Structures and Constants * W. Jolitz 1/90 */ #ifndef _MACHINE_FPU_H_ #define _MACHINE_FPU_H_ #include #ifdef _KERNEL struct fpu_kern_ctx; #define PCB_USER_FPU(pcb) (((pcb)->pcb_flags & PCB_KERNFPU) == 0) #define XSAVE_AREA_ALIGN 64 void fpudna(void); void fpudrop(void); void fpuexit(struct thread *td); int fpuformat(void); int fpugetregs(struct thread *td); void fpuinit(void); void fpurestore(void *addr); void fpuresume(void *addr); void fpusave(void *addr); int fpusetregs(struct thread *td, struct savefpu *addr, char *xfpustate, size_t xfpustate_size); int fpusetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size); void fpususpend(void *addr); int fputrap_sse(void); int fputrap_x87(void); void fpuuserinited(struct thread *td); struct fpu_kern_ctx *fpu_kern_alloc_ctx(u_int flags); void fpu_kern_free_ctx(struct fpu_kern_ctx *ctx); int fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags); int fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx); int fpu_kern_thread(u_int flags); int is_fpu_kern_thread(u_int flags); struct savefpu *fpu_save_area_alloc(void); void fpu_save_area_free(struct savefpu *fsa); void fpu_save_area_reset(struct savefpu *fsa); /* * Flags for fpu_kern_alloc_ctx(), fpu_kern_enter() and fpu_kern_thread(). */ #define FPU_KERN_NORMAL 0x0000 #define FPU_KERN_NOWAIT 0x0001 #define FPU_KERN_KTHR 0x0002 +#define FPU_KERN_NOCTX 0x0004 #endif #endif /* !_MACHINE_FPU_H_ */ Index: projects/mips64-clang/sys/amd64/include/pcb.h =================================================================== --- projects/mips64-clang/sys/amd64/include/pcb.h (revision 305711) +++ projects/mips64-clang/sys/amd64/include/pcb.h (revision 305712) @@ -1,157 +1,158 @@ /*- * Copyright (c) 2003 Peter Wemm. * Copyright (c) 1990 The Regents of the University of California. * All rights reserved. * * This code is derived from software contributed to Berkeley by * William Jolitz. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pcb.h 5.10 (Berkeley) 5/12/91 * $FreeBSD$ */ #ifndef _AMD64_PCB_H_ #define _AMD64_PCB_H_ /* * AMD64 process control block */ #include #include #ifdef __amd64__ /* * NB: The fields marked with (*) are used by kernel debuggers. Their * ABI should be preserved. */ struct pcb { register_t pcb_r15; /* (*) */ register_t pcb_r14; /* (*) */ register_t pcb_r13; /* (*) */ register_t pcb_r12; /* (*) */ register_t pcb_rbp; /* (*) */ register_t pcb_rsp; /* (*) */ register_t pcb_rbx; /* (*) */ register_t pcb_rip; /* (*) */ register_t pcb_fsbase; register_t pcb_gsbase; register_t pcb_kgsbase; register_t pcb_cr0; register_t pcb_cr2; register_t pcb_cr3; register_t pcb_cr4; register_t pcb_dr0; register_t pcb_dr1; register_t pcb_dr2; register_t pcb_dr3; register_t pcb_dr6; register_t pcb_dr7; struct region_descriptor pcb_gdt; struct region_descriptor pcb_idt; struct region_descriptor pcb_ldt; uint16_t pcb_tr; u_int pcb_flags; #define PCB_FULL_IRET 0x01 /* full iret is required */ #define PCB_DBREGS 0x02 /* process using debug registers */ #define PCB_KERNFPU 0x04 /* kernel uses fpu */ #define PCB_FPUINITDONE 0x08 /* fpu state is initialized */ #define PCB_USERFPUINITDONE 0x10 /* fpu user state is initialized */ #define PCB_32BIT 0x40 /* process has 32 bit context (segs etc) */ +#define PCB_FPUNOSAVE 0x80 /* no save area for current FPU ctx */ uint16_t pcb_initial_fpucw; /* copyin/out fault recovery */ caddr_t pcb_onfault; uint64_t pcb_pad0; /* local tss, with i/o bitmap; NULL for common */ struct amd64tss *pcb_tssp; /* model specific registers */ register_t pcb_efer; register_t pcb_star; register_t pcb_lstar; register_t pcb_cstar; register_t pcb_sfmask; struct savefpu *pcb_save; uint64_t pcb_pad[5]; }; /* Per-CPU state saved during suspend and resume. */ struct susppcb { struct pcb sp_pcb; /* fpu context for suspend/resume */ void *sp_fpususpend; }; #endif #ifdef _KERNEL struct trapframe; /* * The pcb_flags is only modified by current thread, or by other threads * when current thread is stopped. However, current thread may change it * from the interrupt context in cpu_switch(), or in the trap handler. * When we read-modify-write pcb_flags from C sources, compiler may generate * code that is not atomic regarding the interrupt handler. If a trap or * interrupt happens and any flag is modified from the handler, it can be * clobbered with the cached value later. Therefore, we implement setting * and clearing flags with single-instruction functions, which do not race * with possible modification of the flags from the trap or interrupt context, * because traps and interrupts are executed only on instruction boundary. */ static __inline void set_pcb_flags(struct pcb *pcb, const u_int flags) { __asm __volatile("orl %1,%0" : "=m" (pcb->pcb_flags) : "ir" (flags), "m" (pcb->pcb_flags) : "cc"); } static __inline void clear_pcb_flags(struct pcb *pcb, const u_int flags) { __asm __volatile("andl %1,%0" : "=m" (pcb->pcb_flags) : "ir" (~flags), "m" (pcb->pcb_flags) : "cc"); } void makectx(struct trapframe *, struct pcb *); int savectx(struct pcb *) __returns_twice; void resumectx(struct pcb *); #endif #endif /* _AMD64_PCB_H_ */ Index: projects/mips64-clang/sys/arm/allwinner/a10_mmc.c =================================================================== --- projects/mips64-clang/sys/arm/allwinner/a10_mmc.c (revision 305711) +++ projects/mips64-clang/sys/arm/allwinner/a10_mmc.c (revision 305712) @@ -1,971 +1,905 @@ /*- * Copyright (c) 2013 Alexander Fedorov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include -#include #include #include #include #define A10_MMC_MEMRES 0 #define A10_MMC_IRQRES 1 #define A10_MMC_RESSZ 2 -#define A10_MMC_DMA_SEGS 16 +#define A10_MMC_DMA_SEGS ((MAXPHYS / PAGE_SIZE) + 1) #define A10_MMC_DMA_MAX_SIZE 0x2000 #define A10_MMC_DMA_FTRGLEVEL 0x20070008 #define CARD_ID_FREQUENCY 400000 -static int a10_mmc_pio_mode = 0; - -TUNABLE_INT("hw.a10.mmc.pio_mode", &a10_mmc_pio_mode); - static struct ofw_compat_data compat_data[] = { {"allwinner,sun4i-a10-mmc", 1}, {"allwinner,sun5i-a13-mmc", 1}, {NULL, 0} }; struct a10_mmc_softc { - bus_space_handle_t a10_bsh; - bus_space_tag_t a10_bst; device_t a10_dev; clk_t a10_clk_ahb; clk_t a10_clk_mmc; hwreset_t a10_rst_ahb; int a10_bus_busy; - int a10_id; int a10_resid; int a10_timeout; struct callout a10_timeoutc; struct mmc_host a10_host; struct mmc_request * a10_req; struct mtx a10_mtx; struct resource * a10_res[A10_MMC_RESSZ]; uint32_t a10_intr; uint32_t a10_intr_wait; void * a10_intrhand; - bus_size_t a10_fifo_reg; /* Fields required for DMA access. */ bus_addr_t a10_dma_desc_phys; bus_dmamap_t a10_dma_map; bus_dma_tag_t a10_dma_tag; void * a10_dma_desc; bus_dmamap_t a10_dma_buf_map; bus_dma_tag_t a10_dma_buf_tag; - int a10_dma_inuse; int a10_dma_map_err; }; static struct resource_spec a10_mmc_res_spec[] = { { SYS_RES_MEMORY, 0, RF_ACTIVE }, { SYS_RES_IRQ, 0, RF_ACTIVE | RF_SHAREABLE }, { -1, 0, 0 } }; static int a10_mmc_probe(device_t); static int a10_mmc_attach(device_t); static int a10_mmc_detach(device_t); static int a10_mmc_setup_dma(struct a10_mmc_softc *); static int a10_mmc_reset(struct a10_mmc_softc *); static void a10_mmc_intr(void *); -static int a10_mmc_update_clock(struct a10_mmc_softc *); +static int a10_mmc_update_clock(struct a10_mmc_softc *, uint32_t); static int a10_mmc_update_ios(device_t, device_t); static int a10_mmc_request(device_t, device_t, struct mmc_request *); static int a10_mmc_get_ro(device_t, device_t); static int a10_mmc_acquire_host(device_t, device_t); static int a10_mmc_release_host(device_t, device_t); #define A10_MMC_LOCK(_sc) mtx_lock(&(_sc)->a10_mtx) #define A10_MMC_UNLOCK(_sc) mtx_unlock(&(_sc)->a10_mtx) #define A10_MMC_READ_4(_sc, _reg) \ - bus_space_read_4((_sc)->a10_bst, (_sc)->a10_bsh, _reg) + bus_read_4((_sc)->a10_res[A10_MMC_MEMRES], _reg) #define A10_MMC_WRITE_4(_sc, _reg, _value) \ - bus_space_write_4((_sc)->a10_bst, (_sc)->a10_bsh, _reg, _value) + bus_write_4((_sc)->a10_res[A10_MMC_MEMRES], _reg, _value) static int a10_mmc_probe(device_t dev) { if (!ofw_bus_status_okay(dev)) return (ENXIO); if (ofw_bus_search_compatible(dev, compat_data)->ocd_data == 0) return (ENXIO); device_set_desc(dev, "Allwinner Integrated MMC/SD controller"); return (BUS_PROBE_DEFAULT); } static int a10_mmc_attach(device_t dev) { device_t child; struct a10_mmc_softc *sc; struct sysctl_ctx_list *ctx; struct sysctl_oid_list *tree; uint32_t bus_width; phandle_t node; int error; node = ofw_bus_get_node(dev); sc = device_get_softc(dev); sc->a10_dev = dev; sc->a10_req = NULL; - sc->a10_id = device_get_unit(dev); - if (sc->a10_id > 3) { - device_printf(dev, "only 4 hosts are supported (0-3)\n"); - return (ENXIO); - } if (bus_alloc_resources(dev, a10_mmc_res_spec, sc->a10_res) != 0) { device_printf(dev, "cannot allocate device resources\n"); return (ENXIO); } - sc->a10_bst = rman_get_bustag(sc->a10_res[A10_MMC_MEMRES]); - sc->a10_bsh = rman_get_bushandle(sc->a10_res[A10_MMC_MEMRES]); if (bus_setup_intr(dev, sc->a10_res[A10_MMC_IRQRES], INTR_TYPE_MISC | INTR_MPSAFE, NULL, a10_mmc_intr, sc, &sc->a10_intrhand)) { bus_release_resources(dev, a10_mmc_res_spec, sc->a10_res); device_printf(dev, "cannot setup interrupt handler\n"); return (ENXIO); } mtx_init(&sc->a10_mtx, device_get_nameunit(sc->a10_dev), "a10_mmc", MTX_DEF); callout_init_mtx(&sc->a10_timeoutc, &sc->a10_mtx, 0); -#if defined(__arm__) - /* - * Later chips use a different FIFO offset. Unfortunately the FDT - * uses the same compatible string for old and new implementations. - */ - switch (allwinner_soc_family()) { - case ALLWINNERSOC_SUN4I: - case ALLWINNERSOC_SUN5I: - case ALLWINNERSOC_SUN7I: - sc->a10_fifo_reg = A10_MMC_FIFO; - break; - default: - sc->a10_fifo_reg = A31_MMC_FIFO; - break; - } -#else /* __aarch64__ */ - sc->a10_fifo_reg = A31_MMC_FIFO; -#endif - /* De-assert reset */ if (hwreset_get_by_ofw_name(dev, 0, "ahb", &sc->a10_rst_ahb) == 0) { error = hwreset_deassert(sc->a10_rst_ahb); if (error != 0) { device_printf(dev, "cannot de-assert reset\n"); - return (error); + goto fail; } } /* Activate the module clock. */ error = clk_get_by_ofw_name(dev, 0, "ahb", &sc->a10_clk_ahb); if (error != 0) { device_printf(dev, "cannot get ahb clock\n"); goto fail; } error = clk_enable(sc->a10_clk_ahb); if (error != 0) { device_printf(dev, "cannot enable ahb clock\n"); goto fail; } error = clk_get_by_ofw_name(dev, 0, "mmc", &sc->a10_clk_mmc); if (error != 0) { device_printf(dev, "cannot get mmc clock\n"); goto fail; } error = clk_set_freq(sc->a10_clk_mmc, CARD_ID_FREQUENCY, CLK_SET_ROUND_DOWN); if (error != 0) { device_printf(dev, "cannot init mmc clock\n"); goto fail; } error = clk_enable(sc->a10_clk_mmc); if (error != 0) { device_printf(dev, "cannot enable mmc clock\n"); goto fail; } sc->a10_timeout = 10; ctx = device_get_sysctl_ctx(dev); tree = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); SYSCTL_ADD_INT(ctx, tree, OID_AUTO, "req_timeout", CTLFLAG_RW, &sc->a10_timeout, 0, "Request timeout in seconds"); - /* Reset controller. */ + /* Hardware reset */ + A10_MMC_WRITE_4(sc, A10_MMC_HWRST, 1); + DELAY(100); + A10_MMC_WRITE_4(sc, A10_MMC_HWRST, 0); + DELAY(500); + + /* Soft Reset controller. */ if (a10_mmc_reset(sc) != 0) { device_printf(dev, "cannot reset the controller\n"); goto fail; } - if (a10_mmc_pio_mode == 0 && a10_mmc_setup_dma(sc) != 0) { + if (a10_mmc_setup_dma(sc) != 0) { device_printf(sc->a10_dev, "Couldn't setup DMA!\n"); - a10_mmc_pio_mode = 1; + goto fail; } - if (bootverbose) - device_printf(sc->a10_dev, "DMA status: %s\n", - a10_mmc_pio_mode ? "disabled" : "enabled"); if (OF_getencprop(node, "bus-width", &bus_width, sizeof(uint32_t)) <= 0) bus_width = 4; sc->a10_host.f_min = 400000; sc->a10_host.f_max = 50000000; sc->a10_host.host_ocr = MMC_OCR_320_330 | MMC_OCR_330_340; sc->a10_host.mode = mode_sd; sc->a10_host.caps = MMC_CAP_HSPEED; if (bus_width >= 4) sc->a10_host.caps |= MMC_CAP_4_BIT_DATA; if (bus_width >= 8) sc->a10_host.caps |= MMC_CAP_8_BIT_DATA; child = device_add_child(dev, "mmc", -1); if (child == NULL) { device_printf(dev, "attaching MMC bus failed!\n"); goto fail; } if (device_probe_and_attach(child) != 0) { device_printf(dev, "attaching MMC child failed!\n"); device_delete_child(dev, child); goto fail; } return (0); fail: callout_drain(&sc->a10_timeoutc); mtx_destroy(&sc->a10_mtx); bus_teardown_intr(dev, sc->a10_res[A10_MMC_IRQRES], sc->a10_intrhand); bus_release_resources(dev, a10_mmc_res_spec, sc->a10_res); return (ENXIO); } static int a10_mmc_detach(device_t dev) { return (EBUSY); } static void a10_dma_desc_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int err) { struct a10_mmc_softc *sc; sc = (struct a10_mmc_softc *)arg; if (err) { sc->a10_dma_map_err = err; return; } sc->a10_dma_desc_phys = segs[0].ds_addr; } static int a10_mmc_setup_dma(struct a10_mmc_softc *sc) { int dma_desc_size, error; /* Allocate the DMA descriptor memory. */ dma_desc_size = sizeof(struct a10_mmc_dma_desc) * A10_MMC_DMA_SEGS; error = bus_dma_tag_create(bus_get_dma_tag(sc->a10_dev), A10_MMC_DMA_ALIGN, 0, BUS_SPACE_MAXADDR_32BIT, BUS_SPACE_MAXADDR, NULL, NULL, dma_desc_size, 1, dma_desc_size, 0, NULL, NULL, &sc->a10_dma_tag); if (error) return (error); error = bus_dmamem_alloc(sc->a10_dma_tag, &sc->a10_dma_desc, BUS_DMA_WAITOK | BUS_DMA_ZERO, &sc->a10_dma_map); if (error) return (error); error = bus_dmamap_load(sc->a10_dma_tag, sc->a10_dma_map, sc->a10_dma_desc, dma_desc_size, a10_dma_desc_cb, sc, 0); if (error) return (error); if (sc->a10_dma_map_err) return (sc->a10_dma_map_err); /* Create the DMA map for data transfers. */ error = bus_dma_tag_create(bus_get_dma_tag(sc->a10_dev), A10_MMC_DMA_ALIGN, 0, BUS_SPACE_MAXADDR_32BIT, BUS_SPACE_MAXADDR, NULL, NULL, A10_MMC_DMA_MAX_SIZE * A10_MMC_DMA_SEGS, A10_MMC_DMA_SEGS, A10_MMC_DMA_MAX_SIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &sc->a10_dma_buf_tag); if (error) return (error); error = bus_dmamap_create(sc->a10_dma_buf_tag, 0, &sc->a10_dma_buf_map); if (error) return (error); return (0); } static void a10_dma_cb(void *arg, bus_dma_segment_t *segs, int nsegs, int err) { int i; struct a10_mmc_dma_desc *dma_desc; struct a10_mmc_softc *sc; sc = (struct a10_mmc_softc *)arg; sc->a10_dma_map_err = err; if (err) return; dma_desc = sc->a10_dma_desc; - /* Note nsegs is guaranteed to be zero if err is non-zero. */ for (i = 0; i < nsegs; i++) { dma_desc[i].buf_size = segs[i].ds_len; dma_desc[i].buf_addr = segs[i].ds_addr; dma_desc[i].config = A10_MMC_DMA_CONFIG_CH | A10_MMC_DMA_CONFIG_OWN; if (i == 0) dma_desc[i].config |= A10_MMC_DMA_CONFIG_FD; if (i < (nsegs - 1)) { dma_desc[i].config |= A10_MMC_DMA_CONFIG_DIC; dma_desc[i].next = sc->a10_dma_desc_phys + ((i + 1) * sizeof(struct a10_mmc_dma_desc)); } else { dma_desc[i].config |= A10_MMC_DMA_CONFIG_LD | A10_MMC_DMA_CONFIG_ER; dma_desc[i].next = 0; } - } + } } static int a10_mmc_prepare_dma(struct a10_mmc_softc *sc) { bus_dmasync_op_t sync_op; int error; struct mmc_command *cmd; uint32_t val; cmd = sc->a10_req->cmd; if (cmd->data->len > A10_MMC_DMA_MAX_SIZE * A10_MMC_DMA_SEGS) return (EFBIG); error = bus_dmamap_load(sc->a10_dma_buf_tag, sc->a10_dma_buf_map, - cmd->data->data, cmd->data->len, a10_dma_cb, sc, BUS_DMA_NOWAIT); + cmd->data->data, cmd->data->len, a10_dma_cb, sc, 0); if (error) return (error); if (sc->a10_dma_map_err) return (sc->a10_dma_map_err); - sc->a10_dma_inuse = 1; if (cmd->data->flags & MMC_DATA_WRITE) sync_op = BUS_DMASYNC_PREWRITE; else sync_op = BUS_DMASYNC_PREREAD; bus_dmamap_sync(sc->a10_dma_buf_tag, sc->a10_dma_buf_map, sync_op); bus_dmamap_sync(sc->a10_dma_tag, sc->a10_dma_map, BUS_DMASYNC_PREWRITE); - val = A10_MMC_READ_4(sc, A10_MMC_IMASK); - val &= ~(A10_MMC_RX_DATA_REQ | A10_MMC_TX_DATA_REQ); - A10_MMC_WRITE_4(sc, A10_MMC_IMASK, val); - val = A10_MMC_READ_4(sc, A10_MMC_GCTRL); - val &= ~A10_MMC_ACCESS_BY_AHB; - val |= A10_MMC_DMA_ENABLE; - A10_MMC_WRITE_4(sc, A10_MMC_GCTRL, val); - val |= A10_MMC_DMA_RESET; - A10_MMC_WRITE_4(sc, A10_MMC_GCTRL, val); - A10_MMC_WRITE_4(sc, A10_MMC_DMAC, A10_MMC_IDMAC_SOFT_RST); + /* Enable DMA */ + val = A10_MMC_READ_4(sc, A10_MMC_GCTL); + val &= ~A10_MMC_CTRL_FIFO_AC_MOD; + val |= A10_MMC_CTRL_DMA_ENB; + A10_MMC_WRITE_4(sc, A10_MMC_GCTL, val); + + /* Reset DMA */ + val |= A10_MMC_CTRL_DMA_RST; + A10_MMC_WRITE_4(sc, A10_MMC_GCTL, val); + + A10_MMC_WRITE_4(sc, A10_MMC_DMAC, A10_MMC_DMAC_IDMAC_SOFT_RST); A10_MMC_WRITE_4(sc, A10_MMC_DMAC, - A10_MMC_IDMAC_IDMA_ON | A10_MMC_IDMAC_FIX_BURST); - val = A10_MMC_READ_4(sc, A10_MMC_IDIE); - val &= ~(A10_MMC_IDMAC_RECEIVE_INT | A10_MMC_IDMAC_TRANSMIT_INT); + A10_MMC_DMAC_IDMAC_IDMA_ON | A10_MMC_DMAC_IDMAC_FIX_BURST); + + /* Enable RX or TX DMA interrupt */ if (cmd->data->flags & MMC_DATA_WRITE) - val |= A10_MMC_IDMAC_TRANSMIT_INT; + val |= A10_MMC_IDST_TX_INT; else - val |= A10_MMC_IDMAC_RECEIVE_INT; + val |= A10_MMC_IDST_RX_INT; A10_MMC_WRITE_4(sc, A10_MMC_IDIE, val); + + /* Set DMA descritptor list address */ A10_MMC_WRITE_4(sc, A10_MMC_DLBA, sc->a10_dma_desc_phys); - A10_MMC_WRITE_4(sc, A10_MMC_FTRGL, A10_MMC_DMA_FTRGLEVEL); + /* FIFO trigger level */ + A10_MMC_WRITE_4(sc, A10_MMC_FWLR, A10_MMC_DMA_FTRGLEVEL); + return (0); } static int a10_mmc_reset(struct a10_mmc_softc *sc) { int timeout; - A10_MMC_WRITE_4(sc, A10_MMC_GCTRL, - A10_MMC_READ_4(sc, A10_MMC_GCTRL) | A10_MMC_RESET); + A10_MMC_WRITE_4(sc, A10_MMC_GCTL, A10_MMC_RESET); timeout = 1000; while (--timeout > 0) { - if ((A10_MMC_READ_4(sc, A10_MMC_GCTRL) & A10_MMC_RESET) == 0) + if ((A10_MMC_READ_4(sc, A10_MMC_GCTL) & A10_MMC_RESET) == 0) break; DELAY(100); } if (timeout == 0) return (ETIMEDOUT); /* Set the timeout. */ - A10_MMC_WRITE_4(sc, A10_MMC_TIMEOUT, 0xffffffff); + A10_MMC_WRITE_4(sc, A10_MMC_TMOR, + A10_MMC_TMOR_DTO_LMT_SHIFT(A10_MMC_TMOR_DTO_LMT_MASK) | + A10_MMC_TMOR_RTO_LMT_SHIFT(A10_MMC_TMOR_RTO_LMT_MASK)); /* Clear pending interrupts. */ - A10_MMC_WRITE_4(sc, A10_MMC_RINTR, 0xffffffff); + A10_MMC_WRITE_4(sc, A10_MMC_RISR, 0xffffffff); A10_MMC_WRITE_4(sc, A10_MMC_IDST, 0xffffffff); /* Unmask interrupts. */ - A10_MMC_WRITE_4(sc, A10_MMC_IMASK, - A10_MMC_CMD_DONE | A10_MMC_INT_ERR_BIT | - A10_MMC_DATA_OVER | A10_MMC_AUTOCMD_DONE); + A10_MMC_WRITE_4(sc, A10_MMC_IMKR, + A10_MMC_INT_CMD_DONE | A10_MMC_INT_ERR_BIT | + A10_MMC_INT_DATA_OVER | A10_MMC_INT_AUTO_STOP_DONE); /* Enable interrupts and AHB access. */ - A10_MMC_WRITE_4(sc, A10_MMC_GCTRL, - A10_MMC_READ_4(sc, A10_MMC_GCTRL) | A10_MMC_INT_ENABLE); + A10_MMC_WRITE_4(sc, A10_MMC_GCTL, + A10_MMC_READ_4(sc, A10_MMC_GCTL) | A10_MMC_CTRL_INT_ENB); return (0); } static void a10_mmc_req_done(struct a10_mmc_softc *sc) { struct mmc_command *cmd; struct mmc_request *req; cmd = sc->a10_req->cmd; if (cmd->error != MMC_ERR_NONE) { /* Reset the controller. */ a10_mmc_reset(sc); - a10_mmc_update_clock(sc); } - if (sc->a10_dma_inuse == 0) { - /* Reset the FIFO. */ - A10_MMC_WRITE_4(sc, A10_MMC_GCTRL, - A10_MMC_READ_4(sc, A10_MMC_GCTRL) | A10_MMC_FIFO_RESET); - } req = sc->a10_req; callout_stop(&sc->a10_timeoutc); sc->a10_req = NULL; sc->a10_intr = 0; sc->a10_resid = 0; - sc->a10_dma_inuse = 0; sc->a10_dma_map_err = 0; sc->a10_intr_wait = 0; req->done(req); } static void a10_mmc_req_ok(struct a10_mmc_softc *sc) { int timeout; struct mmc_command *cmd; uint32_t status; timeout = 1000; while (--timeout > 0) { - status = A10_MMC_READ_4(sc, A10_MMC_STAS); - if ((status & A10_MMC_CARD_DATA_BUSY) == 0) + status = A10_MMC_READ_4(sc, A10_MMC_STAR); + if ((status & A10_MMC_STAR_CARD_BUSY) == 0) break; DELAY(1000); } cmd = sc->a10_req->cmd; if (timeout == 0) { cmd->error = MMC_ERR_FAILED; a10_mmc_req_done(sc); return; } if (cmd->flags & MMC_RSP_PRESENT) { if (cmd->flags & MMC_RSP_136) { cmd->resp[0] = A10_MMC_READ_4(sc, A10_MMC_RESP3); cmd->resp[1] = A10_MMC_READ_4(sc, A10_MMC_RESP2); cmd->resp[2] = A10_MMC_READ_4(sc, A10_MMC_RESP1); cmd->resp[3] = A10_MMC_READ_4(sc, A10_MMC_RESP0); } else cmd->resp[0] = A10_MMC_READ_4(sc, A10_MMC_RESP0); } /* All data has been transferred ? */ if (cmd->data != NULL && (sc->a10_resid << 2) < cmd->data->len) cmd->error = MMC_ERR_FAILED; a10_mmc_req_done(sc); } static void a10_mmc_timeout(void *arg) { struct a10_mmc_softc *sc; sc = (struct a10_mmc_softc *)arg; if (sc->a10_req != NULL) { device_printf(sc->a10_dev, "controller timeout\n"); sc->a10_req->cmd->error = MMC_ERR_TIMEOUT; a10_mmc_req_done(sc); } else device_printf(sc->a10_dev, "Spurious timeout - no active request\n"); } -static int -a10_mmc_pio_transfer(struct a10_mmc_softc *sc, struct mmc_data *data) -{ - int i, write; - uint32_t bit, *buf; - - buf = (uint32_t *)data->data; - write = (data->flags & MMC_DATA_WRITE) ? 1 : 0; - bit = write ? A10_MMC_FIFO_FULL : A10_MMC_FIFO_EMPTY; - for (i = sc->a10_resid; i < (data->len >> 2); i++) { - if ((A10_MMC_READ_4(sc, A10_MMC_STAS) & bit)) - return (1); - if (write) - A10_MMC_WRITE_4(sc, sc->a10_fifo_reg, buf[i]); - else - buf[i] = A10_MMC_READ_4(sc, sc->a10_fifo_reg); - sc->a10_resid = i + 1; - } - - return (0); -} - static void a10_mmc_intr(void *arg) { bus_dmasync_op_t sync_op; struct a10_mmc_softc *sc; struct mmc_data *data; uint32_t idst, imask, rint; sc = (struct a10_mmc_softc *)arg; A10_MMC_LOCK(sc); - rint = A10_MMC_READ_4(sc, A10_MMC_RINTR); + rint = A10_MMC_READ_4(sc, A10_MMC_RISR); idst = A10_MMC_READ_4(sc, A10_MMC_IDST); - imask = A10_MMC_READ_4(sc, A10_MMC_IMASK); + imask = A10_MMC_READ_4(sc, A10_MMC_IMKR); if (idst == 0 && imask == 0 && rint == 0) { A10_MMC_UNLOCK(sc); return; } #ifdef DEBUG device_printf(sc->a10_dev, "idst: %#x, imask: %#x, rint: %#x\n", idst, imask, rint); #endif if (sc->a10_req == NULL) { device_printf(sc->a10_dev, "Spurious interrupt - no active request, rint: 0x%08X\n", rint); goto end; } if (rint & A10_MMC_INT_ERR_BIT) { device_printf(sc->a10_dev, "error rint: 0x%08X\n", rint); - if (rint & A10_MMC_RESP_TIMEOUT) + if (rint & A10_MMC_INT_RESP_TIMEOUT) sc->a10_req->cmd->error = MMC_ERR_TIMEOUT; else sc->a10_req->cmd->error = MMC_ERR_FAILED; a10_mmc_req_done(sc); goto end; } - if (idst & A10_MMC_IDMAC_ERROR) { + if (idst & A10_MMC_IDST_ERROR) { device_printf(sc->a10_dev, "error idst: 0x%08x\n", idst); sc->a10_req->cmd->error = MMC_ERR_FAILED; a10_mmc_req_done(sc); goto end; } sc->a10_intr |= rint; data = sc->a10_req->cmd->data; - if (data != NULL && sc->a10_dma_inuse == 1 && - (idst & A10_MMC_IDMAC_COMPLETE)) { + if (data != NULL && (idst & A10_MMC_IDST_COMPLETE) != 0) { if (data->flags & MMC_DATA_WRITE) sync_op = BUS_DMASYNC_POSTWRITE; else sync_op = BUS_DMASYNC_POSTREAD; bus_dmamap_sync(sc->a10_dma_buf_tag, sc->a10_dma_buf_map, sync_op); bus_dmamap_sync(sc->a10_dma_tag, sc->a10_dma_map, BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(sc->a10_dma_buf_tag, sc->a10_dma_buf_map); sc->a10_resid = data->len >> 2; - } else if (data != NULL && sc->a10_dma_inuse == 0 && - (rint & (A10_MMC_DATA_OVER | A10_MMC_RX_DATA_REQ | - A10_MMC_TX_DATA_REQ)) != 0) - a10_mmc_pio_transfer(sc, data); + } if ((sc->a10_intr & sc->a10_intr_wait) == sc->a10_intr_wait) a10_mmc_req_ok(sc); end: A10_MMC_WRITE_4(sc, A10_MMC_IDST, idst); - A10_MMC_WRITE_4(sc, A10_MMC_RINTR, rint); + A10_MMC_WRITE_4(sc, A10_MMC_RISR, rint); A10_MMC_UNLOCK(sc); } static int a10_mmc_request(device_t bus, device_t child, struct mmc_request *req) { int blksz; struct a10_mmc_softc *sc; struct mmc_command *cmd; - uint32_t cmdreg, val; + uint32_t cmdreg; + int err; sc = device_get_softc(bus); A10_MMC_LOCK(sc); if (sc->a10_req) { A10_MMC_UNLOCK(sc); return (EBUSY); } sc->a10_req = req; cmd = req->cmd; - cmdreg = A10_MMC_START; + cmdreg = A10_MMC_CMDR_LOAD; if (cmd->opcode == MMC_GO_IDLE_STATE) - cmdreg |= A10_MMC_SEND_INIT_SEQ; + cmdreg |= A10_MMC_CMDR_SEND_INIT_SEQ; if (cmd->flags & MMC_RSP_PRESENT) - cmdreg |= A10_MMC_RESP_EXP; + cmdreg |= A10_MMC_CMDR_RESP_RCV; if (cmd->flags & MMC_RSP_136) - cmdreg |= A10_MMC_LONG_RESP; + cmdreg |= A10_MMC_CMDR_LONG_RESP; if (cmd->flags & MMC_RSP_CRC) - cmdreg |= A10_MMC_CHECK_RESP_CRC; + cmdreg |= A10_MMC_CMDR_CHK_RESP_CRC; sc->a10_intr = 0; sc->a10_resid = 0; - sc->a10_intr_wait = A10_MMC_CMD_DONE; + sc->a10_intr_wait = A10_MMC_INT_CMD_DONE; cmd->error = MMC_ERR_NONE; if (cmd->data != NULL) { - sc->a10_intr_wait |= A10_MMC_DATA_OVER; - cmdreg |= A10_MMC_DATA_EXP | A10_MMC_WAIT_PREOVER; + sc->a10_intr_wait |= A10_MMC_INT_DATA_OVER; + cmdreg |= A10_MMC_CMDR_DATA_TRANS | A10_MMC_CMDR_WAIT_PRE_OVER; if (cmd->data->flags & MMC_DATA_MULTI) { - cmdreg |= A10_MMC_SEND_AUTOSTOP; - sc->a10_intr_wait |= A10_MMC_AUTOCMD_DONE; + cmdreg |= A10_MMC_CMDR_STOP_CMD_FLAG; + sc->a10_intr_wait |= A10_MMC_INT_AUTO_STOP_DONE; } if (cmd->data->flags & MMC_DATA_WRITE) - cmdreg |= A10_MMC_WRITE; + cmdreg |= A10_MMC_CMDR_DIR_WRITE; blksz = min(cmd->data->len, MMC_SECTOR_SIZE); - A10_MMC_WRITE_4(sc, A10_MMC_BLKSZ, blksz); - A10_MMC_WRITE_4(sc, A10_MMC_BCNTR, cmd->data->len); + A10_MMC_WRITE_4(sc, A10_MMC_BKSR, blksz); + A10_MMC_WRITE_4(sc, A10_MMC_BYCR, cmd->data->len); - if (a10_mmc_pio_mode == 0) - a10_mmc_prepare_dma(sc); - /* Enable PIO access if sc->a10_dma_inuse is not set. */ - if (sc->a10_dma_inuse == 0) { - val = A10_MMC_READ_4(sc, A10_MMC_GCTRL); - val &= ~A10_MMC_DMA_ENABLE; - val |= A10_MMC_ACCESS_BY_AHB; - A10_MMC_WRITE_4(sc, A10_MMC_GCTRL, val); - val = A10_MMC_READ_4(sc, A10_MMC_IMASK); - val |= A10_MMC_RX_DATA_REQ | A10_MMC_TX_DATA_REQ; - A10_MMC_WRITE_4(sc, A10_MMC_IMASK, val); - } + err = a10_mmc_prepare_dma(sc); + if (err != 0) + device_printf(sc->a10_dev, "prepare_dma failed: %d\n", err); } - A10_MMC_WRITE_4(sc, A10_MMC_CARG, cmd->arg); + A10_MMC_WRITE_4(sc, A10_MMC_CAGR, cmd->arg); A10_MMC_WRITE_4(sc, A10_MMC_CMDR, cmdreg | cmd->opcode); callout_reset(&sc->a10_timeoutc, sc->a10_timeout * hz, a10_mmc_timeout, sc); A10_MMC_UNLOCK(sc); return (0); } static int a10_mmc_read_ivar(device_t bus, device_t child, int which, uintptr_t *result) { struct a10_mmc_softc *sc; sc = device_get_softc(bus); switch (which) { default: return (EINVAL); case MMCBR_IVAR_BUS_MODE: *(int *)result = sc->a10_host.ios.bus_mode; break; case MMCBR_IVAR_BUS_WIDTH: *(int *)result = sc->a10_host.ios.bus_width; break; case MMCBR_IVAR_CHIP_SELECT: *(int *)result = sc->a10_host.ios.chip_select; break; case MMCBR_IVAR_CLOCK: *(int *)result = sc->a10_host.ios.clock; break; case MMCBR_IVAR_F_MIN: *(int *)result = sc->a10_host.f_min; break; case MMCBR_IVAR_F_MAX: *(int *)result = sc->a10_host.f_max; break; case MMCBR_IVAR_HOST_OCR: *(int *)result = sc->a10_host.host_ocr; break; case MMCBR_IVAR_MODE: *(int *)result = sc->a10_host.mode; break; case MMCBR_IVAR_OCR: *(int *)result = sc->a10_host.ocr; break; case MMCBR_IVAR_POWER_MODE: *(int *)result = sc->a10_host.ios.power_mode; break; case MMCBR_IVAR_VDD: *(int *)result = sc->a10_host.ios.vdd; break; case MMCBR_IVAR_CAPS: *(int *)result = sc->a10_host.caps; break; case MMCBR_IVAR_MAX_DATA: *(int *)result = 65535; break; } return (0); } static int a10_mmc_write_ivar(device_t bus, device_t child, int which, uintptr_t value) { struct a10_mmc_softc *sc; sc = device_get_softc(bus); switch (which) { default: return (EINVAL); case MMCBR_IVAR_BUS_MODE: sc->a10_host.ios.bus_mode = value; break; case MMCBR_IVAR_BUS_WIDTH: sc->a10_host.ios.bus_width = value; break; case MMCBR_IVAR_CHIP_SELECT: sc->a10_host.ios.chip_select = value; break; case MMCBR_IVAR_CLOCK: sc->a10_host.ios.clock = value; break; case MMCBR_IVAR_MODE: sc->a10_host.mode = value; break; case MMCBR_IVAR_OCR: sc->a10_host.ocr = value; break; case MMCBR_IVAR_POWER_MODE: sc->a10_host.ios.power_mode = value; break; case MMCBR_IVAR_VDD: sc->a10_host.ios.vdd = value; break; /* These are read-only */ case MMCBR_IVAR_CAPS: case MMCBR_IVAR_HOST_OCR: case MMCBR_IVAR_F_MIN: case MMCBR_IVAR_F_MAX: case MMCBR_IVAR_MAX_DATA: return (EINVAL); } return (0); } static int -a10_mmc_update_clock(struct a10_mmc_softc *sc) +a10_mmc_update_clock(struct a10_mmc_softc *sc, uint32_t clkon) { uint32_t cmdreg; int retry; + uint32_t ckcr; - cmdreg = A10_MMC_START | A10_MMC_UPCLK_ONLY | - A10_MMC_WAIT_PREOVER; + ckcr = A10_MMC_READ_4(sc, A10_MMC_CKCR); + ckcr &= ~(A10_MMC_CKCR_CCLK_ENB | A10_MMC_CKCR_CCLK_CTRL); + + if (clkon) + ckcr |= A10_MMC_CKCR_CCLK_ENB; + + A10_MMC_WRITE_4(sc, A10_MMC_CKCR, ckcr); + + cmdreg = A10_MMC_CMDR_LOAD | A10_MMC_CMDR_PRG_CLK | + A10_MMC_CMDR_WAIT_PRE_OVER; A10_MMC_WRITE_4(sc, A10_MMC_CMDR, cmdreg); retry = 0xfffff; while (--retry > 0) { - if ((A10_MMC_READ_4(sc, A10_MMC_CMDR) & A10_MMC_START) == 0) { - A10_MMC_WRITE_4(sc, A10_MMC_RINTR, 0xffffffff); + if ((A10_MMC_READ_4(sc, A10_MMC_CMDR) & A10_MMC_CMDR_LOAD) == 0) { + A10_MMC_WRITE_4(sc, A10_MMC_RISR, 0xffffffff); return (0); } DELAY(10); } - A10_MMC_WRITE_4(sc, A10_MMC_RINTR, 0xffffffff); + A10_MMC_WRITE_4(sc, A10_MMC_RISR, 0xffffffff); device_printf(sc->a10_dev, "timeout updating clock\n"); return (ETIMEDOUT); } static int a10_mmc_update_ios(device_t bus, device_t child) { int error; struct a10_mmc_softc *sc; struct mmc_ios *ios; - uint32_t clkcr; + uint32_t ckcr; sc = device_get_softc(bus); - clkcr = A10_MMC_READ_4(sc, A10_MMC_CLKCR); - if (clkcr & A10_MMC_CARD_CLK_ON) { - /* Disable clock. */ - clkcr &= ~A10_MMC_CARD_CLK_ON; - A10_MMC_WRITE_4(sc, A10_MMC_CLKCR, clkcr); - error = a10_mmc_update_clock(sc); - if (error != 0) - return (error); - } ios = &sc->a10_host.ios; + + /* Set the bus width. */ + switch (ios->bus_width) { + case bus_width_1: + A10_MMC_WRITE_4(sc, A10_MMC_BWDR, A10_MMC_BWDR1); + break; + case bus_width_4: + A10_MMC_WRITE_4(sc, A10_MMC_BWDR, A10_MMC_BWDR4); + break; + case bus_width_8: + A10_MMC_WRITE_4(sc, A10_MMC_BWDR, A10_MMC_BWDR8); + break; + } + if (ios->clock) { - /* Reset the divider. */ - clkcr &= ~A10_MMC_CLKCR_DIV; - A10_MMC_WRITE_4(sc, A10_MMC_CLKCR, clkcr); - error = a10_mmc_update_clock(sc); + + /* Disable clock */ + error = a10_mmc_update_clock(sc, 0); if (error != 0) return (error); + /* Reset the divider. */ + ckcr = A10_MMC_READ_4(sc, A10_MMC_CKCR); + ckcr &= ~A10_MMC_CKCR_CCLK_DIV; + A10_MMC_WRITE_4(sc, A10_MMC_CKCR, ckcr); + /* Set the MMC clock. */ error = clk_set_freq(sc->a10_clk_mmc, ios->clock, CLK_SET_ROUND_DOWN); if (error != 0) { device_printf(sc->a10_dev, "failed to set frequency to %u Hz: %d\n", ios->clock, error); return (error); } /* Enable clock. */ - clkcr |= A10_MMC_CARD_CLK_ON; - A10_MMC_WRITE_4(sc, A10_MMC_CLKCR, clkcr); - error = a10_mmc_update_clock(sc); + error = a10_mmc_update_clock(sc, 1); if (error != 0) return (error); } - /* Set the bus width. */ - switch (ios->bus_width) { - case bus_width_1: - A10_MMC_WRITE_4(sc, A10_MMC_WIDTH, A10_MMC_WIDTH1); - break; - case bus_width_4: - A10_MMC_WRITE_4(sc, A10_MMC_WIDTH, A10_MMC_WIDTH4); - break; - case bus_width_8: - A10_MMC_WRITE_4(sc, A10_MMC_WIDTH, A10_MMC_WIDTH8); - break; - } return (0); } static int a10_mmc_get_ro(device_t bus, device_t child) { return (0); } static int a10_mmc_acquire_host(device_t bus, device_t child) { struct a10_mmc_softc *sc; int error; sc = device_get_softc(bus); A10_MMC_LOCK(sc); while (sc->a10_bus_busy) { error = msleep(sc, &sc->a10_mtx, PCATCH, "mmchw", 0); if (error != 0) { A10_MMC_UNLOCK(sc); return (error); } } sc->a10_bus_busy++; A10_MMC_UNLOCK(sc); return (0); } static int a10_mmc_release_host(device_t bus, device_t child) { struct a10_mmc_softc *sc; sc = device_get_softc(bus); A10_MMC_LOCK(sc); sc->a10_bus_busy--; wakeup(sc); A10_MMC_UNLOCK(sc); return (0); } static device_method_t a10_mmc_methods[] = { /* Device interface */ DEVMETHOD(device_probe, a10_mmc_probe), DEVMETHOD(device_attach, a10_mmc_attach), DEVMETHOD(device_detach, a10_mmc_detach), /* Bus interface */ DEVMETHOD(bus_read_ivar, a10_mmc_read_ivar), DEVMETHOD(bus_write_ivar, a10_mmc_write_ivar), DEVMETHOD(bus_print_child, bus_generic_print_child), /* MMC bridge interface */ DEVMETHOD(mmcbr_update_ios, a10_mmc_update_ios), DEVMETHOD(mmcbr_request, a10_mmc_request), DEVMETHOD(mmcbr_get_ro, a10_mmc_get_ro), DEVMETHOD(mmcbr_acquire_host, a10_mmc_acquire_host), DEVMETHOD(mmcbr_release_host, a10_mmc_release_host), DEVMETHOD_END }; static devclass_t a10_mmc_devclass; static driver_t a10_mmc_driver = { "a10_mmc", a10_mmc_methods, sizeof(struct a10_mmc_softc), }; DRIVER_MODULE(a10_mmc, simplebus, a10_mmc_driver, a10_mmc_devclass, 0, 0); DRIVER_MODULE(mmc, a10_mmc, mmc_driver, mmc_devclass, NULL, NULL); MODULE_DEPEND(a10_mmc, mmc, 1, 1, 1); Index: projects/mips64-clang/sys/arm/allwinner/a10_mmc.h =================================================================== --- projects/mips64-clang/sys/arm/allwinner/a10_mmc.h (revision 305711) +++ projects/mips64-clang/sys/arm/allwinner/a10_mmc.h (revision 305712) @@ -1,202 +1,204 @@ /*- * Copyright (c) 2013 Alexander Fedorov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _A10_MMC_H_ #define _A10_MMC_H_ -#define A10_MMC_GCTRL 0x00 /* Global Control Register */ -#define A10_MMC_CLKCR 0x04 /* Clock Control Register */ -#define A10_MMC_TIMEOUT 0x08 /* Timeout Register */ -#define A10_MMC_WIDTH 0x0C /* Bus Width Register */ -#define A10_MMC_BLKSZ 0x10 /* Block Size Register */ -#define A10_MMC_BCNTR 0x14 /* Byte Count Register */ +#define A10_MMC_GCTL 0x00 /* Control Register */ +#define A10_MMC_CKCR 0x04 /* Clock Control Register */ +#define A10_MMC_TMOR 0x08 /* Timeout Register */ +#define A10_MMC_BWDR 0x0C /* Bus Width Register */ +#define A10_MMC_BKSR 0x10 /* Block Size Register */ +#define A10_MMC_BYCR 0x14 /* Byte Count Register */ #define A10_MMC_CMDR 0x18 /* Command Register */ -#define A10_MMC_CARG 0x1C /* Argument Register */ +#define A10_MMC_CAGR 0x1C /* Argument Register */ #define A10_MMC_RESP0 0x20 /* Response Register 0 */ #define A10_MMC_RESP1 0x24 /* Response Register 1 */ #define A10_MMC_RESP2 0x28 /* Response Register 2 */ #define A10_MMC_RESP3 0x2C /* Response Register 3 */ -#define A10_MMC_IMASK 0x30 /* Interrupt Mask Register */ -#define A10_MMC_MISTA 0x34 /* Masked Interrupt Status Register */ -#define A10_MMC_RINTR 0x38 /* Raw Interrupt Status Register */ -#define A10_MMC_STAS 0x3C /* Status Register */ -#define A10_MMC_FTRGL 0x40 /* FIFO Threshold Watermark Register */ +#define A10_MMC_IMKR 0x30 /* Interrupt Mask Register */ +#define A10_MMC_MISR 0x34 /* Masked Interrupt Status Register */ +#define A10_MMC_RISR 0x38 /* Raw Interrupt Status Register */ +#define A10_MMC_STAR 0x3C /* Status Register */ +#define A10_MMC_FWLR 0x40 /* FIFO Threshold Watermark Register */ #define A10_MMC_FUNS 0x44 /* Function Select Register */ -#define A10_MMC_CBCR 0x48 /* CIU Byte Count Register */ -#define A10_MMC_BBCR 0x4C /* BIU Byte Count Register */ -#define A10_MMC_DBGC 0x50 /* Debug Enable Register */ +#define A10_MMC_HWRST 0x78 /* Hardware reset (not documented) */ #define A10_MMC_DMAC 0x80 /* IDMAC Control Register */ #define A10_MMC_DLBA 0x84 /* IDMAC Desc List Base Address Reg */ #define A10_MMC_IDST 0x88 /* IDMAC Status Register */ #define A10_MMC_IDIE 0x8C /* IDMAC Interrupt Enable Register */ -#define A10_MMC_CHDA 0x90 -#define A10_MMC_CBDA 0x94 -#define A10_MMC_FIFO 0x100 /* FIFO Access Address (A10/A20) */ -#define A31_MMC_FIFO 0x200 /* FIFO Access Address (A31) */ +#define A10_MMC_FIFO 0x100 /* FIFO Access Address (A10/A20) */ +#define A31_MMC_FIFO 0x200 /* FIFO Access Address (A31) */ -/* A10_MMC_GCTRL */ -#define A10_MMC_SOFT_RESET (1U << 0) -#define A10_MMC_FIFO_RESET (1U << 1) -#define A10_MMC_DMA_RESET (1U << 2) -#define A10_MMC_INT_ENABLE (1U << 4) -#define A10_MMC_DMA_ENABLE (1U << 5) -#define A10_MMC_DEBOUNCE_ENABLE (1U << 8) -#define A10_MMC_DDR_MODE (1U << 10) -#define A10_MMC_ACCESS_BY_AHB (1U << 31) +/* A10_MMC_GCTL */ +#define A10_MMC_CTRL_SOFT_RST (1U << 0) +#define A10_MMC_CTRL_FIFO_RST (1U << 1) +#define A10_MMC_CTRL_DMA_RST (1U << 2) +#define A10_MMC_CTRL_INT_ENB (1U << 4) +#define A10_MMC_CTRL_DMA_ENB (1U << 5) +#define A10_MMC_CTRL_CD_DBC_ENB (1U << 8) +#define A10_MMC_CTRL_DDR_MOD_SEL (1U << 10) +#define A10_MMC_CTRL_FIFO_AC_MOD (1U << 31) #define A10_MMC_RESET \ - (A10_MMC_SOFT_RESET | A10_MMC_FIFO_RESET | A10_MMC_DMA_RESET) + (A10_MMC_CTRL_SOFT_RST | A10_MMC_CTRL_FIFO_RST | A10_MMC_CTRL_DMA_RST) -/* A10_MMC_CLKCR */ -#define A10_MMC_CARD_CLK_ON (1U << 16) -#define A10_MMC_LOW_POWER_ON (1U << 17) -#define A10_MMC_CLKCR_DIV 0xff +/* A10_MMC_CKCR */ +#define A10_MMC_CKCR_CCLK_ENB (1U << 16) +#define A10_MMC_CKCR_CCLK_CTRL (1U << 17) +#define A10_MMC_CKCR_CCLK_DIV 0xff -/* A10_MMC_WIDTH */ -#define A10_MMC_WIDTH1 0 -#define A10_MMC_WIDTH4 1 -#define A10_MMC_WIDTH8 2 +/* A10_MMC_TMOR */ +#define A10_MMC_TMOR_RTO_LMT_SHIFT(x) x /* Response timeout limit */ +#define A10_MMC_TMOR_RTO_LMT_MASK 0xff +#define A10_MMC_TMOR_DTO_LMT_SHIFT(x) (x << 8) /* Data timeout limit */ +#define A10_MMC_TMOR_DTO_LMT_MASK 0xffffff +/* A10_MMC_BWDR */ +#define A10_MMC_BWDR1 0 +#define A10_MMC_BWDR4 1 +#define A10_MMC_BWDR8 2 + /* A10_MMC_CMDR */ -#define A10_MMC_RESP_EXP (1U << 6) -#define A10_MMC_LONG_RESP (1U << 7) -#define A10_MMC_CHECK_RESP_CRC (1U << 8) -#define A10_MMC_DATA_EXP (1U << 9) -#define A10_MMC_WRITE (1U << 10) -#define A10_MMC_SEQ_MODE (1U << 11) -#define A10_MMC_SEND_AUTOSTOP (1U << 12) -#define A10_MMC_WAIT_PREOVER (1U << 13) -#define A10_MMC_STOP_ABORT_CMD (1U << 14) -#define A10_MMC_SEND_INIT_SEQ (1U << 15) -#define A10_MMC_UPCLK_ONLY (1U << 21) -#define A10_MMC_RDCEATADEV (1U << 22) -#define A10_MMC_CCS_EXP (1U << 23) -#define A10_MMC_ENB_BOOT (1U << 24) -#define A10_MMC_ALT_BOOT_OPT (1U << 25) -#define A10_MMC_BOOT_ACK_EXP (1U << 26) -#define A10_MMC_DISABLE_BOOT (1U << 27) -#define A10_MMC_VOL_SWITCH (1U << 28) -#define A10_MMC_START (1U << 31) +#define A10_MMC_CMDR_RESP_RCV (1U << 6) +#define A10_MMC_CMDR_LONG_RESP (1U << 7) +#define A10_MMC_CMDR_CHK_RESP_CRC (1U << 8) +#define A10_MMC_CMDR_DATA_TRANS (1U << 9) +#define A10_MMC_CMDR_DIR_WRITE (1U << 10) +#define A10_MMC_CMDR_TRANS_MODE_STREAM (1U << 11) +#define A10_MMC_CMDR_STOP_CMD_FLAG (1U << 12) +#define A10_MMC_CMDR_WAIT_PRE_OVER (1U << 13) +#define A10_MMC_CMDR_STOP_ABT_CMD (1U << 14) +#define A10_MMC_CMDR_SEND_INIT_SEQ (1U << 15) +#define A10_MMC_CMDR_PRG_CLK (1U << 21) +#define A10_MMC_CMDR_RD_CEDATA_DEV (1U << 22) +#define A10_MMC_CMDR_CCS_EXP (1U << 23) +#define A10_MMC_CMDR_BOOT_MOD_SHIFT 24 +#define A10_MMC_CMDR_BOOT_MOD_NORMAL 0 +#define A10_MMC_CMDR_BOOT_MOD_MANDATORY 1 +#define A10_MMC_CMDR_BOOT_MOD_ALT 2 +#define A10_MMC_CMDR_EXP_BOOT_ACK (1U << 26) +#define A10_MMC_CMDR_BOOT_ABT (1U << 27) +#define A10_MMC_CMDR_VOL_SW (1U << 28) +#define A10_MMC_CMDR_LOAD (1U << 31) -/* A10_MMC_IMASK and A10_MMC_RINTR */ -#define A10_MMC_RESP_ERR (1U << 1) -#define A10_MMC_CMD_DONE (1U << 2) -#define A10_MMC_DATA_OVER (1U << 3) -#define A10_MMC_TX_DATA_REQ (1U << 4) -#define A10_MMC_RX_DATA_REQ (1U << 5) -#define A10_MMC_RESP_CRC_ERR (1U << 6) -#define A10_MMC_DATA_CRC_ERR (1U << 7) -#define A10_MMC_RESP_TIMEOUT (1U << 8) -#define A10_MMC_ACK_RECV (1U << 8) -#define A10_MMC_DATA_TIMEOUT (1U << 9) -#define A10_MMC_BOOT_START (1U << 9) -#define A10_MMC_DATA_STARVE (1U << 10) -#define A10_MMC_VOL_CHG_DONE (1U << 10) -#define A10_MMC_FIFO_RUN_ERR (1U << 11) -#define A10_MMC_HARDW_LOCKED (1U << 12) -#define A10_MMC_START_BIT_ERR (1U << 13) -#define A10_MMC_AUTOCMD_DONE (1U << 14) -#define A10_MMC_END_BIT_ERR (1U << 15) -#define A10_MMC_SDIO_INT (1U << 16) -#define A10_MMC_CARD_INSERT (1U << 30) -#define A10_MMC_CARD_REMOVE (1U << 31) +/* A10_MMC_IMKR and A10_MMC_RISR */ +#define A10_MMC_INT_RESP_ERR (1U << 1) +#define A10_MMC_INT_CMD_DONE (1U << 2) +#define A10_MMC_INT_DATA_OVER (1U << 3) +#define A10_MMC_INT_TX_DATA_REQ (1U << 4) +#define A10_MMC_INT_RX_DATA_REQ (1U << 5) +#define A10_MMC_INT_RESP_CRC_ERR (1U << 6) +#define A10_MMC_INT_DATA_CRC_ERR (1U << 7) +#define A10_MMC_INT_RESP_TIMEOUT (1U << 8) +#define A10_MMC_INT_BOOT_ACK_RECV (1U << 8) +#define A10_MMC_INT_DATA_TIMEOUT (1U << 9) +#define A10_MMC_INT_BOOT_START (1U << 9) +#define A10_MMC_INT_DATA_STARVE (1U << 10) +#define A10_MMC_INT_VOL_CHG_DONE (1U << 10) +#define A10_MMC_INT_FIFO_RUN_ERR (1U << 11) +#define A10_MMC_INT_CMD_BUSY (1U << 12) +#define A10_MMC_INT_DATA_START_ERR (1U << 13) +#define A10_MMC_INT_AUTO_STOP_DONE (1U << 14) +#define A10_MMC_INT_DATA_END_BIT_ERR (1U << 15) +#define A10_MMC_INT_SDIO (1U << 16) +#define A10_MMC_INT_CARD_INSERT (1U << 30) +#define A10_MMC_INT_CARD_REMOVE (1U << 31) #define A10_MMC_INT_ERR_BIT \ - (A10_MMC_RESP_ERR | A10_MMC_RESP_CRC_ERR | \ - A10_MMC_DATA_CRC_ERR | A10_MMC_RESP_TIMEOUT | \ - A10_MMC_FIFO_RUN_ERR | A10_MMC_HARDW_LOCKED | \ - A10_MMC_START_BIT_ERR | A10_MMC_END_BIT_ERR) + (A10_MMC_INT_RESP_ERR | A10_MMC_INT_RESP_CRC_ERR | \ + A10_MMC_INT_DATA_CRC_ERR | A10_MMC_INT_RESP_TIMEOUT | \ + A10_MMC_INT_FIFO_RUN_ERR | A10_MMC_INT_CMD_BUSY | \ + A10_MMC_INT_DATA_START_ERR | A10_MMC_INT_DATA_END_BIT_ERR) -/* A10_MMC_STAS */ -#define A10_MMC_RX_WLFLAG (1U << 0) -#define A10_MMC_TX_WLFLAG (1U << 1) -#define A10_MMC_FIFO_EMPTY (1U << 2) -#define A10_MMC_FIFO_FULL (1U << 3) -#define A10_MMC_CARD_PRESENT (1U << 8) -#define A10_MMC_CARD_DATA_BUSY (1U << 9) -#define A10_MMC_DATA_FSM_BUSY (1U << 10) -#define A10_MMC_DMA_REQ (1U << 31) -#define A10_MMC_FIFO_SIZE 16 +/* A10_MMC_STAR */ +#define A10_MMC_STAR_FIFO_RX_LEVEL (1U << 0) +#define A10_MMC_STAR_FIFO_TX_LEVEL (1U << 1) +#define A10_MMC_STAR_FIFO_EMPTY (1U << 2) +#define A10_MMC_STAR_FIFO_FULL (1U << 3) +#define A10_MMC_STAR_CARD_PRESENT (1U << 8) +#define A10_MMC_STAR_CARD_BUSY (1U << 9) +#define A10_MMC_STAR_FSM_BUSY (1U << 10) +#define A10_MMC_STAR_DMA_REQ (1U << 31) /* A10_MMC_FUNS */ #define A10_MMC_CE_ATA_ON (0xceaaU << 16) #define A10_MMC_SEND_IRQ_RESP (1U << 0) #define A10_MMC_SDIO_RD_WAIT (1U << 1) #define A10_MMC_ABT_RD_DATA (1U << 2) #define A10_MMC_SEND_CC_SD (1U << 8) #define A10_MMC_SEND_AUTOSTOP_CC_SD (1U << 9) #define A10_MMC_CE_ATA_DEV_INT_ENB (1U << 10) /* IDMA CONTROLLER BUS MOD BIT FIELD */ -#define A10_MMC_IDMAC_SOFT_RST (1U << 0) -#define A10_MMC_IDMAC_FIX_BURST (1U << 1) -#define A10_MMC_IDMAC_IDMA_ON (1U << 7) -#define A10_MMC_IDMAC_REFETCH_DES (1U << 31) +#define A10_MMC_DMAC_IDMAC_SOFT_RST (1U << 0) +#define A10_MMC_DMAC_IDMAC_FIX_BURST (1U << 1) +#define A10_MMC_DMAC_IDMAC_IDMA_ON (1U << 7) +#define A10_MMC_DMAC_IDMAC_REFETCH_DES (1U << 31) /* A10_MMC_IDST */ -#define A10_MMC_IDMAC_TRANSMIT_INT (1U << 0) -#define A10_MMC_IDMAC_RECEIVE_INT (1U << 1) -#define A10_MMC_IDMAC_FATAL_BUS_ERR (1U << 2) -#define A10_MMC_IDMAC_DES_INVALID (1U << 4) -#define A10_MMC_IDMAC_CARD_ERR_SUM (1U << 5) -#define A10_MMC_IDMAC_NORMAL_INT_SUM (1U << 8) -#define A10_MMC_IDMAC_ABNORMAL_INT_SUM (1U << 9) -#define A10_MMC_IDMAC_HOST_ABT_INTX (1U << 10) -#define A10_MMC_IDMAC_HOST_ABT_INRX (1U << 10) -#define A10_MMC_IDMAC_IDLE (0U << 13) -#define A10_MMC_IDMAC_SUSPEND (1U << 13) -#define A10_MMC_IDMAC_DESC_RD (2U << 13) -#define A10_MMC_IDMAC_DESC_CHECK (3U << 13) -#define A10_MMC_IDMAC_RD_REQ_WAIT (4U << 13) -#define A10_MMC_IDMAC_WR_REQ_WAIT (5U << 13) -#define A10_MMC_IDMAC_RD (6U << 13) -#define A10_MMC_IDMAC_WR (7U << 13) -#define A10_MMC_IDMAC_DESC_CLOSE (8U << 13) -#define A10_MMC_IDMAC_ERROR \ - (A10_MMC_IDMAC_FATAL_BUS_ERR | A10_MMC_IDMAC_CARD_ERR_SUM | \ - A10_MMC_IDMAC_DES_INVALID | A10_MMC_IDMAC_ABNORMAL_INT_SUM) -#define A10_MMC_IDMAC_COMPLETE \ - (A10_MMC_IDMAC_TRANSMIT_INT | A10_MMC_IDMAC_RECEIVE_INT) +#define A10_MMC_IDST_TX_INT (1U << 0) +#define A10_MMC_IDST_RX_INT (1U << 1) +#define A10_MMC_IDST_FATAL_BERR_INT (1U << 2) +#define A10_MMC_IDST_DES_UNAVL_INT (1U << 4) +#define A10_MMC_IDST_ERR_FLAG_SUM (1U << 5) +#define A10_MMC_IDST_NOR_INT_SUM (1U << 8) +#define A10_MMC_IDST_ABN_INT_SUM (1U << 9) +#define A10_MMC_IDST_HOST_ABT_INTX (1U << 10) +#define A10_MMC_IDST_HOST_ABT_INRX (1U << 10) +#define A10_MMC_IDST_IDLE (0U << 13) +#define A10_MMC_IDST_SUSPEND (1U << 13) +#define A10_MMC_IDST_DESC_RD (2U << 13) +#define A10_MMC_IDST_DESC_CHECK (3U << 13) +#define A10_MMC_IDST_RD_REQ_WAIT (4U << 13) +#define A10_MMC_IDST_WR_REQ_WAIT (5U << 13) +#define A10_MMC_IDST_RD (6U << 13) +#define A10_MMC_IDST_WR (7U << 13) +#define A10_MMC_IDST_DESC_CLOSE (8U << 13) +#define A10_MMC_IDST_ERROR \ + (A10_MMC_IDST_FATAL_BERR_INT | A10_MMC_IDST_ERR_FLAG_SUM | \ + A10_MMC_IDST_DES_UNAVL_INT | A10_MMC_IDST_ABN_INT_SUM) +#define A10_MMC_IDST_COMPLETE \ + (A10_MMC_IDST_TX_INT | A10_MMC_IDST_RX_INT) /* The DMA descriptor table. */ struct a10_mmc_dma_desc { uint32_t config; -#define A10_MMC_DMA_CONFIG_DIC (1U << 1) -#define A10_MMC_DMA_CONFIG_LD (1U << 2) -#define A10_MMC_DMA_CONFIG_FD (1U << 3) -#define A10_MMC_DMA_CONFIG_CH (1U << 4) -#define A10_MMC_DMA_CONFIG_ER (1U << 5) -#define A10_MMC_DMA_CONFIG_CES (1U << 30) -#define A10_MMC_DMA_CONFIG_OWN (1U << 31) +#define A10_MMC_DMA_CONFIG_DIC (1U << 1) /* Disable Interrupt Completion */ +#define A10_MMC_DMA_CONFIG_LD (1U << 2) /* Last DES */ +#define A10_MMC_DMA_CONFIG_FD (1U << 3) /* First DES */ +#define A10_MMC_DMA_CONFIG_CH (1U << 4) /* CHAIN MOD */ +#define A10_MMC_DMA_CONFIG_ER (1U << 5) /* End of Ring (undocumented register) */ +#define A10_MMC_DMA_CONFIG_CES (1U << 30) /* Card Error Summary */ +#define A10_MMC_DMA_CONFIG_OWN (1U << 31) /* DES Own Flag */ uint32_t buf_size; uint32_t buf_addr; uint32_t next; }; -/* DMA descriptors and data buffers must be aligned to 32-bits */ -#define A10_MMC_DMA_ALIGN 4 +#define A10_MMC_DMA_ALIGN 4 #endif /* _A10_MMC_H_ */ Index: projects/mips64-clang/sys/arm/arm/pmap-v6.c =================================================================== --- projects/mips64-clang/sys/arm/arm/pmap-v6.c (revision 305711) +++ projects/mips64-clang/sys/arm/arm/pmap-v6.c (revision 305712) @@ -1,6806 +1,6800 @@ /*- * Copyright (c) 1991 Regents of the University of California. * Copyright (c) 1994 John S. Dyson * Copyright (c) 1994 David Greenman * Copyright (c) 2005-2010 Alan L. Cox * Copyright (c) 2014-2016 Svatopluk Kraus * Copyright (c) 2014-2016 Michal Meloun * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_vm.h" #include "opt_pmap.h" #include "opt_ddb.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #else #include #endif #ifdef DDB #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #endif #ifndef PMAP_SHPGPERPROC #define PMAP_SHPGPERPROC 200 #endif #ifndef DIAGNOSTIC #define PMAP_INLINE __inline #else #define PMAP_INLINE #endif #ifdef PMAP_DEBUG static void pmap_zero_page_check(vm_page_t m); void pmap_debug(int level); int pmap_pid_dump(int pid); #define PDEBUG(_lev_,_stat_) \ if (pmap_debug_level >= (_lev_)) \ ((_stat_)) #define dprintf printf int pmap_debug_level = 1; #else /* PMAP_DEBUG */ #define PDEBUG(_lev_,_stat_) /* Nothing */ #define dprintf(x, arg...) #endif /* PMAP_DEBUG */ /* * Level 2 page tables map definion ('max' is excluded). */ #define PT2V_MIN_ADDRESS ((vm_offset_t)PT2MAP) #define PT2V_MAX_ADDRESS ((vm_offset_t)PT2MAP + PT2MAP_SIZE) #define UPT2V_MIN_ADDRESS ((vm_offset_t)PT2MAP) #define UPT2V_MAX_ADDRESS \ ((vm_offset_t)(PT2MAP + (KERNBASE >> PT2MAP_SHIFT))) /* * Promotion to a 1MB (PTE1) page mapping requires that the corresponding * 4KB (PTE2) page mappings have identical settings for the following fields: */ #define PTE2_PROMOTE (PTE2_V | PTE2_A | PTE2_NM | PTE2_S | PTE2_NG | \ PTE2_NX | PTE2_RO | PTE2_U | PTE2_W | \ PTE2_ATTR_MASK) #define PTE1_PROMOTE (PTE1_V | PTE1_A | PTE1_NM | PTE1_S | PTE1_NG | \ PTE1_NX | PTE1_RO | PTE1_U | PTE1_W | \ PTE1_ATTR_MASK) #define ATTR_TO_L1(l2_attr) ((((l2_attr) & L2_TEX0) ? L1_S_TEX0 : 0) | \ (((l2_attr) & L2_C) ? L1_S_C : 0) | \ (((l2_attr) & L2_B) ? L1_S_B : 0) | \ (((l2_attr) & PTE2_A) ? PTE1_A : 0) | \ (((l2_attr) & PTE2_NM) ? PTE1_NM : 0) | \ (((l2_attr) & PTE2_S) ? PTE1_S : 0) | \ (((l2_attr) & PTE2_NG) ? PTE1_NG : 0) | \ (((l2_attr) & PTE2_NX) ? PTE1_NX : 0) | \ (((l2_attr) & PTE2_RO) ? PTE1_RO : 0) | \ (((l2_attr) & PTE2_U) ? PTE1_U : 0) | \ (((l2_attr) & PTE2_W) ? PTE1_W : 0)) #define ATTR_TO_L2(l1_attr) ((((l1_attr) & L1_S_TEX0) ? L2_TEX0 : 0) | \ (((l1_attr) & L1_S_C) ? L2_C : 0) | \ (((l1_attr) & L1_S_B) ? L2_B : 0) | \ (((l1_attr) & PTE1_A) ? PTE2_A : 0) | \ (((l1_attr) & PTE1_NM) ? PTE2_NM : 0) | \ (((l1_attr) & PTE1_S) ? PTE2_S : 0) | \ (((l1_attr) & PTE1_NG) ? PTE2_NG : 0) | \ (((l1_attr) & PTE1_NX) ? PTE2_NX : 0) | \ (((l1_attr) & PTE1_RO) ? PTE2_RO : 0) | \ (((l1_attr) & PTE1_U) ? PTE2_U : 0) | \ (((l1_attr) & PTE1_W) ? PTE2_W : 0)) /* * PTE2 descriptors creation macros. */ #define PTE2_ATTR_DEFAULT vm_memattr_to_pte2(VM_MEMATTR_DEFAULT) #define PTE2_ATTR_PT vm_memattr_to_pte2(pt_memattr) #define PTE2_KPT(pa) PTE2_KERN(pa, PTE2_AP_KRW, PTE2_ATTR_PT) #define PTE2_KPT_NG(pa) PTE2_KERN_NG(pa, PTE2_AP_KRW, PTE2_ATTR_PT) #define PTE2_KRW(pa) PTE2_KERN(pa, PTE2_AP_KRW, PTE2_ATTR_DEFAULT) #define PTE2_KRO(pa) PTE2_KERN(pa, PTE2_AP_KR, PTE2_ATTR_DEFAULT) #define PV_STATS #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif /* * The boot_pt1 is used temporary in very early boot stage as L1 page table. * We can init many things with no memory allocation thanks to its static * allocation and this brings two main advantages: * (1) other cores can be started very simply, * (2) various boot loaders can be supported as its arguments can be processed * in virtual address space and can be moved to safe location before * first allocation happened. * Only disadvantage is that boot_pt1 is used only in very early boot stage. * However, the table is uninitialized and so lays in bss. Therefore kernel * image size is not influenced. * * QQQ: In the future, maybe, boot_pt1 can be used for soft reset and * CPU suspend/resume game. */ extern pt1_entry_t boot_pt1[]; vm_paddr_t base_pt1; pt1_entry_t *kern_pt1; pt2_entry_t *kern_pt2tab; pt2_entry_t *PT2MAP; static uint32_t ttb_flags; static vm_memattr_t pt_memattr; ttb_entry_t pmap_kern_ttb; struct pmap kernel_pmap_store; LIST_HEAD(pmaplist, pmap); static struct pmaplist allpmaps; static struct mtx allpmaps_lock; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ static vm_offset_t kernel_vm_end_new; vm_offset_t kernel_vm_end = KERNBASE + NKPT2PG * NPT2_IN_PG * PTE1_SIZE; vm_offset_t vm_max_kernel_address; vm_paddr_t kernel_l1pa; static struct rwlock __aligned(CACHE_LINE_SIZE) pvh_global_lock; /* * Data for the pv entry allocation mechanism */ static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks); static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; static struct md_page *pv_table; /* XXX: Is it used only the list in md_page? */ static int shpgperproc = PMAP_SHPGPERPROC; struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */ int pv_maxchunks; /* How many chunks we have KVA for */ vm_offset_t pv_vafree; /* freelist stored in the PTE */ vm_paddr_t first_managed_pa; #define pa_to_pvh(pa) (&pv_table[pte1_index(pa - first_managed_pa)]) /* * All those kernel PT submaps that BSD is so fond of */ struct sysmaps { struct mtx lock; pt2_entry_t *CMAP1; pt2_entry_t *CMAP2; pt2_entry_t *CMAP3; caddr_t CADDR1; caddr_t CADDR2; caddr_t CADDR3; }; static struct sysmaps sysmaps_pcpu[MAXCPU]; caddr_t _tmppt = 0; struct msgbuf *msgbufp = NULL; /* XXX move it to machdep.c */ /* * Crashdump maps. */ static caddr_t crashdumpmap; static pt2_entry_t *PMAP1 = NULL, *PMAP2; static pt2_entry_t *PADDR1 = NULL, *PADDR2; #ifdef DDB static pt2_entry_t *PMAP3; static pt2_entry_t *PADDR3; static int PMAP3cpu __unused; /* for SMP only */ #endif #ifdef SMP static int PMAP1cpu; static int PMAP1changedcpu; SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, &PMAP1changedcpu, 0, "Number of times pmap_pte2_quick changed CPU with same PMAP1"); #endif static int PMAP1changed; SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, &PMAP1changed, 0, "Number of times pmap_pte2_quick changed PMAP1"); static int PMAP1unchanged; SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, &PMAP1unchanged, 0, "Number of times pmap_pte2_quick didn't change PMAP1"); static struct mtx PMAP2mutex; static __inline void pt2_wirecount_init(vm_page_t m); static boolean_t pmap_demote_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va); void cache_icache_sync_fresh(vm_offset_t va, vm_paddr_t pa, vm_size_t size); /* * Function to set the debug level of the pmap code. */ #ifdef PMAP_DEBUG void pmap_debug(int level) { pmap_debug_level = level; dprintf("pmap_debug: level=%d\n", pmap_debug_level); } #endif /* PMAP_DEBUG */ /* * This table must corespond with memory attribute configuration in vm.h. * First entry is used for normal system mapping. * * Device memory is always marked as shared. * Normal memory is shared only in SMP . * Not outer shareable bits are not used yet. * Class 6 cannot be used on ARM11. */ #define TEXDEF_TYPE_SHIFT 0 #define TEXDEF_TYPE_MASK 0x3 #define TEXDEF_INNER_SHIFT 2 #define TEXDEF_INNER_MASK 0x3 #define TEXDEF_OUTER_SHIFT 4 #define TEXDEF_OUTER_MASK 0x3 #define TEXDEF_NOS_SHIFT 6 #define TEXDEF_NOS_MASK 0x1 #define TEX(t, i, o, s) \ ((t) << TEXDEF_TYPE_SHIFT) | \ ((i) << TEXDEF_INNER_SHIFT) | \ ((o) << TEXDEF_OUTER_SHIFT | \ ((s) << TEXDEF_NOS_SHIFT)) static uint32_t tex_class[8] = { /* type inner cache outer cache */ TEX(PRRR_MEM, NMRR_WB_WA, NMRR_WB_WA, 0), /* 0 - ATTR_WB_WA */ TEX(PRRR_MEM, NMRR_NC, NMRR_NC, 0), /* 1 - ATTR_NOCACHE */ TEX(PRRR_DEV, NMRR_NC, NMRR_NC, 0), /* 2 - ATTR_DEVICE */ TEX(PRRR_SO, NMRR_NC, NMRR_NC, 0), /* 3 - ATTR_SO */ TEX(PRRR_MEM, NMRR_WT, NMRR_WT, 0), /* 4 - ATTR_WT */ TEX(PRRR_MEM, NMRR_NC, NMRR_NC, 0), /* 5 - NOT USED YET */ TEX(PRRR_MEM, NMRR_NC, NMRR_NC, 0), /* 6 - NOT USED YET */ TEX(PRRR_MEM, NMRR_NC, NMRR_NC, 0), /* 7 - NOT USED YET */ }; #undef TEX static uint32_t pte2_attr_tab[8] = { PTE2_ATTR_WB_WA, /* 0 - VM_MEMATTR_WB_WA */ PTE2_ATTR_NOCACHE, /* 1 - VM_MEMATTR_NOCACHE */ PTE2_ATTR_DEVICE, /* 2 - VM_MEMATTR_DEVICE */ PTE2_ATTR_SO, /* 3 - VM_MEMATTR_SO */ PTE2_ATTR_WT, /* 4 - VM_MEMATTR_WRITE_THROUGH */ 0, /* 5 - NOT USED YET */ 0, /* 6 - NOT USED YET */ 0 /* 7 - NOT USED YET */ }; CTASSERT(VM_MEMATTR_WB_WA == 0); CTASSERT(VM_MEMATTR_NOCACHE == 1); CTASSERT(VM_MEMATTR_DEVICE == 2); CTASSERT(VM_MEMATTR_SO == 3); CTASSERT(VM_MEMATTR_WRITE_THROUGH == 4); static inline uint32_t vm_memattr_to_pte2(vm_memattr_t ma) { KASSERT((u_int)ma < 5, ("%s: bad vm_memattr_t %d", __func__, ma)); return (pte2_attr_tab[(u_int)ma]); } static inline uint32_t vm_page_pte2_attr(vm_page_t m) { return (vm_memattr_to_pte2(m->md.pat_mode)); } /* * Convert TEX definition entry to TTB flags. */ static uint32_t encode_ttb_flags(int idx) { uint32_t inner, outer, nos, reg; inner = (tex_class[idx] >> TEXDEF_INNER_SHIFT) & TEXDEF_INNER_MASK; outer = (tex_class[idx] >> TEXDEF_OUTER_SHIFT) & TEXDEF_OUTER_MASK; nos = (tex_class[idx] >> TEXDEF_NOS_SHIFT) & TEXDEF_NOS_MASK; reg = nos << 5; reg |= outer << 3; if (cpuinfo.coherent_walk) reg |= (inner & 0x1) << 6; reg |= (inner & 0x2) >> 1; #ifdef SMP reg |= 1 << 1; #endif return reg; } /* * Set TEX remapping registers in current CPU. */ void pmap_set_tex(void) { uint32_t prrr, nmrr; uint32_t type, inner, outer, nos; int i; #ifdef PMAP_PTE_NOCACHE /* XXX fixme */ if (cpuinfo.coherent_walk) { pt_memattr = VM_MEMATTR_WB_WA; ttb_flags = encode_ttb_flags(0); } else { pt_memattr = VM_MEMATTR_NOCACHE; ttb_flags = encode_ttb_flags(1); } #else pt_memattr = VM_MEMATTR_WB_WA; ttb_flags = encode_ttb_flags(0); #endif prrr = 0; nmrr = 0; /* Build remapping register from TEX classes. */ for (i = 0; i < 8; i++) { type = (tex_class[i] >> TEXDEF_TYPE_SHIFT) & TEXDEF_TYPE_MASK; inner = (tex_class[i] >> TEXDEF_INNER_SHIFT) & TEXDEF_INNER_MASK; outer = (tex_class[i] >> TEXDEF_OUTER_SHIFT) & TEXDEF_OUTER_MASK; nos = (tex_class[i] >> TEXDEF_NOS_SHIFT) & TEXDEF_NOS_MASK; prrr |= type << (i * 2); prrr |= nos << (i + 24); nmrr |= inner << (i * 2); nmrr |= outer << (i * 2 + 16); } /* Add shareable bits for device memory. */ prrr |= PRRR_DS0 | PRRR_DS1; /* Add shareable bits for normal memory in SMP case. */ #ifdef SMP prrr |= PRRR_NS1; #endif cp15_prrr_set(prrr); cp15_nmrr_set(nmrr); /* Caches are disabled, so full TLB flush should be enough. */ tlb_flush_all_local(); } /* * KERNBASE must be multiple of NPT2_IN_PG * PTE1_SIZE. In other words, * KERNBASE is mapped by first L2 page table in L2 page table page. It * meets same constrain due to PT2MAP being placed just under KERNBASE. */ CTASSERT((KERNBASE & (NPT2_IN_PG * PTE1_SIZE - 1)) == 0); CTASSERT((KERNBASE - VM_MAXUSER_ADDRESS) >= PT2MAP_SIZE); /* * In crazy dreams, PAGE_SIZE could be a multiple of PTE2_SIZE in general. * For now, anyhow, the following check must be fulfilled. */ CTASSERT(PAGE_SIZE == PTE2_SIZE); /* * We don't want to mess up MI code with all MMU and PMAP definitions, * so some things, which depend on other ones, are defined independently. * Now, it is time to check that we don't screw up something. */ CTASSERT(PDRSHIFT == PTE1_SHIFT); /* * Check L1 and L2 page table entries definitions consistency. */ CTASSERT(NB_IN_PT1 == (sizeof(pt1_entry_t) * NPTE1_IN_PT1)); CTASSERT(NB_IN_PT2 == (sizeof(pt2_entry_t) * NPTE2_IN_PT2)); /* * Check L2 page tables page consistency. */ CTASSERT(PAGE_SIZE == (NPT2_IN_PG * NB_IN_PT2)); CTASSERT((1 << PT2PG_SHIFT) == NPT2_IN_PG); /* * Check PT2TAB consistency. * PT2TAB_ENTRIES is defined as a division of NPTE1_IN_PT1 by NPT2_IN_PG. * This should be done without remainder. */ CTASSERT(NPTE1_IN_PT1 == (PT2TAB_ENTRIES * NPT2_IN_PG)); /* * A PT2MAP magic. * * All level 2 page tables (PT2s) are mapped continuously and accordingly * into PT2MAP address space. As PT2 size is less than PAGE_SIZE, this can * be done only if PAGE_SIZE is a multiple of PT2 size. All PT2s in one page * must be used together, but not necessary at once. The first PT2 in a page * must map things on correctly aligned address and the others must follow * in right order. */ #define NB_IN_PT2TAB (PT2TAB_ENTRIES * sizeof(pt2_entry_t)) #define NPT2_IN_PT2TAB (NB_IN_PT2TAB / NB_IN_PT2) #define NPG_IN_PT2TAB (NB_IN_PT2TAB / PAGE_SIZE) /* * Check PT2TAB consistency. * NPT2_IN_PT2TAB is defined as a division of NB_IN_PT2TAB by NB_IN_PT2. * NPG_IN_PT2TAB is defined as a division of NB_IN_PT2TAB by PAGE_SIZE. * The both should be done without remainder. */ CTASSERT(NB_IN_PT2TAB == (NPT2_IN_PT2TAB * NB_IN_PT2)); CTASSERT(NB_IN_PT2TAB == (NPG_IN_PT2TAB * PAGE_SIZE)); /* * The implementation was made general, however, with the assumption * bellow in mind. In case of another value of NPG_IN_PT2TAB, * the code should be once more rechecked. */ CTASSERT(NPG_IN_PT2TAB == 1); /* * Get offset of PT2 in a page * associated with given PT1 index. */ static __inline u_int page_pt2off(u_int pt1_idx) { return ((pt1_idx & PT2PG_MASK) * NB_IN_PT2); } /* * Get physical address of PT2 * associated with given PT2s page and PT1 index. */ static __inline vm_paddr_t page_pt2pa(vm_paddr_t pgpa, u_int pt1_idx) { return (pgpa + page_pt2off(pt1_idx)); } /* * Get first entry of PT2 * associated with given PT2s page and PT1 index. */ static __inline pt2_entry_t * page_pt2(vm_offset_t pgva, u_int pt1_idx) { return ((pt2_entry_t *)(pgva + page_pt2off(pt1_idx))); } /* * Get virtual address of PT2s page (mapped in PT2MAP) * which holds PT2 which holds entry which maps given virtual address. */ static __inline vm_offset_t pt2map_pt2pg(vm_offset_t va) { va &= ~(NPT2_IN_PG * PTE1_SIZE - 1); return ((vm_offset_t)pt2map_entry(va)); } /***************************************************************************** * * THREE pmap initialization milestones exist: * * locore.S * -> fundamental init (including MMU) in ASM * * initarm() * -> fundamental init continues in C * -> first available physical address is known * * pmap_bootstrap_prepare() -> FIRST PMAP MILESTONE (first epoch begins) * -> basic (safe) interface for physical address allocation is made * -> basic (safe) interface for virtual mapping is made * -> limited not SMP coherent work is possible * * -> more fundamental init continues in C * -> locks and some more things are available * -> all fundamental allocations and mappings are done * * pmap_bootstrap() -> SECOND PMAP MILESTONE (second epoch begins) * -> phys_avail[] and virtual_avail is set * -> control is passed to vm subsystem * -> physical and virtual address allocation are off limit * -> low level mapping functions, some SMP coherent, * are available, which cannot be used before vm subsystem * is being inited * * mi_startup() * -> vm subsystem is being inited * * pmap_init() -> THIRD PMAP MILESTONE (third epoch begins) * -> pmap is fully inited * *****************************************************************************/ /***************************************************************************** * * PMAP first stage initialization and utility functions * for pre-bootstrap epoch. * * After pmap_bootstrap_prepare() is called, the following functions * can be used: * * (1) strictly only for this stage functions for physical page allocations, * virtual space allocations, and mappings: * * vm_paddr_t pmap_preboot_get_pages(u_int num); * void pmap_preboot_map_pages(vm_paddr_t pa, vm_offset_t va, u_int num); * vm_offset_t pmap_preboot_reserve_pages(u_int num); * vm_offset_t pmap_preboot_get_vpages(u_int num); * void pmap_preboot_map_attr(vm_paddr_t pa, vm_offset_t va, vm_size_t size, * vm_prot_t prot, vm_memattr_t attr); * * (2) for all stages: * * vm_paddr_t pmap_kextract(vm_offset_t va); * * NOTE: This is not SMP coherent stage. * *****************************************************************************/ #define KERNEL_P2V(pa) \ ((vm_offset_t)((pa) - arm_physmem_kernaddr + KERNVIRTADDR)) #define KERNEL_V2P(va) \ ((vm_paddr_t)((va) - KERNVIRTADDR + arm_physmem_kernaddr)) static vm_paddr_t last_paddr; /* * Pre-bootstrap epoch page allocator. */ vm_paddr_t pmap_preboot_get_pages(u_int num) { vm_paddr_t ret; ret = last_paddr; last_paddr += num * PAGE_SIZE; return (ret); } /* * The fundamental initialization of PMAP stuff. * * Some things already happened in locore.S and some things could happen * before pmap_bootstrap_prepare() is called, so let's recall what is done: * 1. Caches are disabled. * 2. We are running on virtual addresses already with 'boot_pt1' * as L1 page table. * 3. So far, all virtual addresses can be converted to physical ones and * vice versa by the following macros: * KERNEL_P2V(pa) .... physical to virtual ones, * KERNEL_V2P(va) .... virtual to physical ones. * * What is done herein: * 1. The 'boot_pt1' is replaced by real kernel L1 page table 'kern_pt1'. * 2. PT2MAP magic is brought to live. * 3. Basic preboot functions for page allocations and mappings can be used. * 4. Everything is prepared for L1 cache enabling. * * Variations: * 1. To use second TTB register, so kernel and users page tables will be * separated. This way process forking - pmap_pinit() - could be faster, * it saves physical pages and KVA per a process, and it's simple change. * However, it will lead, due to hardware matter, to the following: * (a) 2G space for kernel and 2G space for users. * (b) 1G space for kernel in low addresses and 3G for users above it. * A question is: Is the case (b) really an option? Note that case (b) * does save neither physical memory and KVA. */ void pmap_bootstrap_prepare(vm_paddr_t last) { vm_paddr_t pt2pg_pa, pt2tab_pa, pa, size; vm_offset_t pt2pg_va; pt1_entry_t *pte1p; pt2_entry_t *pte2p; u_int i; uint32_t actlr_mask, actlr_set, l1_attr; /* * Now, we are going to make real kernel mapping. Note that we are * already running on some mapping made in locore.S and we expect * that it's large enough to ensure nofault access to physical memory * allocated herein before switch. * * As kernel image and everything needed before are and will be mapped * by section mappings, we align last physical address to PTE1_SIZE. */ last_paddr = pte1_roundup(last); /* * Allocate and zero page(s) for kernel L1 page table. * * Note that it's first allocation on space which was PTE1_SIZE * aligned and as such base_pt1 is aligned to NB_IN_PT1 too. */ base_pt1 = pmap_preboot_get_pages(NPG_IN_PT1); kern_pt1 = (pt1_entry_t *)KERNEL_P2V(base_pt1); bzero((void*)kern_pt1, NB_IN_PT1); pte1_sync_range(kern_pt1, NB_IN_PT1); /* Allocate and zero page(s) for kernel PT2TAB. */ pt2tab_pa = pmap_preboot_get_pages(NPG_IN_PT2TAB); kern_pt2tab = (pt2_entry_t *)KERNEL_P2V(pt2tab_pa); bzero(kern_pt2tab, NB_IN_PT2TAB); pte2_sync_range(kern_pt2tab, NB_IN_PT2TAB); /* Allocate and zero page(s) for kernel L2 page tables. */ pt2pg_pa = pmap_preboot_get_pages(NKPT2PG); pt2pg_va = KERNEL_P2V(pt2pg_pa); size = NKPT2PG * PAGE_SIZE; bzero((void*)pt2pg_va, size); pte2_sync_range((pt2_entry_t *)pt2pg_va, size); /* * Add a physical memory segment (vm_phys_seg) corresponding to the * preallocated pages for kernel L2 page tables so that vm_page * structures representing these pages will be created. The vm_page * structures are required for promotion of the corresponding kernel * virtual addresses to section mappings. */ vm_phys_add_seg(pt2tab_pa, pmap_preboot_get_pages(0)); /* * Insert allocated L2 page table pages to PT2TAB and make * link to all PT2s in L1 page table. See how kernel_vm_end * is initialized. * * We play simple and safe. So every KVA will have underlaying * L2 page table, even kernel image mapped by sections. */ pte2p = kern_pt2tab_entry(KERNBASE); for (pa = pt2pg_pa; pa < pt2pg_pa + size; pa += PTE2_SIZE) pt2tab_store(pte2p++, PTE2_KPT(pa)); pte1p = kern_pte1(KERNBASE); for (pa = pt2pg_pa; pa < pt2pg_pa + size; pa += NB_IN_PT2) pte1_store(pte1p++, PTE1_LINK(pa)); /* Make section mappings for kernel. */ l1_attr = ATTR_TO_L1(PTE2_ATTR_DEFAULT); pte1p = kern_pte1(KERNBASE); for (pa = KERNEL_V2P(KERNBASE); pa < last; pa += PTE1_SIZE) pte1_store(pte1p++, PTE1_KERN(pa, PTE1_AP_KRW, l1_attr)); /* * Get free and aligned space for PT2MAP and make L1 page table links * to L2 page tables held in PT2TAB. * * Note that pages holding PT2s are stored in PT2TAB as pt2_entry_t * descriptors and PT2TAB page(s) itself is(are) used as PT2s. Thus * each entry in PT2TAB maps all PT2s in a page. This implies that * virtual address of PT2MAP must be aligned to NPT2_IN_PG * PTE1_SIZE. */ PT2MAP = (pt2_entry_t *)(KERNBASE - PT2MAP_SIZE); pte1p = kern_pte1((vm_offset_t)PT2MAP); for (pa = pt2tab_pa, i = 0; i < NPT2_IN_PT2TAB; i++, pa += NB_IN_PT2) { pte1_store(pte1p++, PTE1_LINK(pa)); } /* * Store PT2TAB in PT2TAB itself, i.e. self reference mapping. * Each pmap will hold own PT2TAB, so the mapping should be not global. */ pte2p = kern_pt2tab_entry((vm_offset_t)PT2MAP); for (pa = pt2tab_pa, i = 0; i < NPG_IN_PT2TAB; i++, pa += PTE2_SIZE) { pt2tab_store(pte2p++, PTE2_KPT_NG(pa)); } /* * Choose correct L2 page table and make mappings for allocations * made herein which replaces temporary locore.S mappings after a while. * Note that PT2MAP cannot be used until we switch to kern_pt1. * * Note, that these allocations started aligned on 1M section and * kernel PT1 was allocated first. Making of mappings must follow * order of physical allocations as we've used KERNEL_P2V() macro * for virtual addresses resolution. */ pte2p = kern_pt2tab_entry((vm_offset_t)kern_pt1); pt2pg_va = KERNEL_P2V(pte2_pa(pte2_load(pte2p))); pte2p = page_pt2(pt2pg_va, pte1_index((vm_offset_t)kern_pt1)); /* Make mapping for kernel L1 page table. */ for (pa = base_pt1, i = 0; i < NPG_IN_PT1; i++, pa += PTE2_SIZE) pte2_store(pte2p++, PTE2_KPT(pa)); /* Make mapping for kernel PT2TAB. */ for (pa = pt2tab_pa, i = 0; i < NPG_IN_PT2TAB; i++, pa += PTE2_SIZE) pte2_store(pte2p++, PTE2_KPT(pa)); /* Finally, switch from 'boot_pt1' to 'kern_pt1'. */ pmap_kern_ttb = base_pt1 | ttb_flags; cpuinfo_get_actlr_modifier(&actlr_mask, &actlr_set); reinit_mmu(pmap_kern_ttb, actlr_mask, actlr_set); /* * Initialize the first available KVA. As kernel image is mapped by * sections, we are leaving some gap behind. */ virtual_avail = (vm_offset_t)kern_pt2tab + NPG_IN_PT2TAB * PAGE_SIZE; } /* * Setup L2 page table page for given KVA. * Used in pre-bootstrap epoch. * * Note that we have allocated NKPT2PG pages for L2 page tables in advance * and used them for mapping KVA starting from KERNBASE. However, this is not * enough. Vectors and devices need L2 page tables too. Note that they are * even above VM_MAX_KERNEL_ADDRESS. */ static __inline vm_paddr_t pmap_preboot_pt2pg_setup(vm_offset_t va) { pt2_entry_t *pte2p, pte2; vm_paddr_t pt2pg_pa; /* Get associated entry in PT2TAB. */ pte2p = kern_pt2tab_entry(va); /* Just return, if PT2s page exists already. */ pte2 = pt2tab_load(pte2p); if (pte2_is_valid(pte2)) return (pte2_pa(pte2)); KASSERT(va >= VM_MAX_KERNEL_ADDRESS, ("%s: NKPT2PG too small", __func__)); /* * Allocate page for PT2s and insert it to PT2TAB. * In other words, map it into PT2MAP space. */ pt2pg_pa = pmap_preboot_get_pages(1); pt2tab_store(pte2p, PTE2_KPT(pt2pg_pa)); /* Zero all PT2s in allocated page. */ bzero((void*)pt2map_pt2pg(va), PAGE_SIZE); pte2_sync_range((pt2_entry_t *)pt2map_pt2pg(va), PAGE_SIZE); return (pt2pg_pa); } /* * Setup L2 page table for given KVA. * Used in pre-bootstrap epoch. */ static void pmap_preboot_pt2_setup(vm_offset_t va) { pt1_entry_t *pte1p; vm_paddr_t pt2pg_pa, pt2_pa; /* Setup PT2's page. */ pt2pg_pa = pmap_preboot_pt2pg_setup(va); pt2_pa = page_pt2pa(pt2pg_pa, pte1_index(va)); /* Insert PT2 to PT1. */ pte1p = kern_pte1(va); pte1_store(pte1p, PTE1_LINK(pt2_pa)); } /* * Get L2 page entry associated with given KVA. * Used in pre-bootstrap epoch. */ static __inline pt2_entry_t* pmap_preboot_vtopte2(vm_offset_t va) { pt1_entry_t *pte1p; /* Setup PT2 if needed. */ pte1p = kern_pte1(va); if (!pte1_is_valid(pte1_load(pte1p))) /* XXX - sections ?! */ pmap_preboot_pt2_setup(va); return (pt2map_entry(va)); } /* * Pre-bootstrap epoch page(s) mapping(s). */ void pmap_preboot_map_pages(vm_paddr_t pa, vm_offset_t va, u_int num) { u_int i; pt2_entry_t *pte2p; /* Map all the pages. */ for (i = 0; i < num; i++) { pte2p = pmap_preboot_vtopte2(va); pte2_store(pte2p, PTE2_KRW(pa)); va += PAGE_SIZE; pa += PAGE_SIZE; } } /* * Pre-bootstrap epoch virtual space alocator. */ vm_offset_t pmap_preboot_reserve_pages(u_int num) { u_int i; vm_offset_t start, va; pt2_entry_t *pte2p; /* Allocate virtual space. */ start = va = virtual_avail; virtual_avail += num * PAGE_SIZE; /* Zero the mapping. */ for (i = 0; i < num; i++) { pte2p = pmap_preboot_vtopte2(va); pte2_store(pte2p, 0); va += PAGE_SIZE; } return (start); } /* * Pre-bootstrap epoch page(s) allocation and mapping(s). */ vm_offset_t pmap_preboot_get_vpages(u_int num) { vm_paddr_t pa; vm_offset_t va; /* Allocate physical page(s). */ pa = pmap_preboot_get_pages(num); /* Allocate virtual space. */ va = virtual_avail; virtual_avail += num * PAGE_SIZE; /* Map and zero all. */ pmap_preboot_map_pages(pa, va, num); bzero((void *)va, num * PAGE_SIZE); return (va); } /* * Pre-bootstrap epoch page mapping(s) with attributes. */ void pmap_preboot_map_attr(vm_paddr_t pa, vm_offset_t va, vm_size_t size, vm_prot_t prot, vm_memattr_t attr) { u_int num; u_int l1_attr, l1_prot, l2_prot, l2_attr; pt1_entry_t *pte1p; pt2_entry_t *pte2p; l2_prot = prot & VM_PROT_WRITE ? PTE2_AP_KRW : PTE2_AP_KR; l2_prot |= (prot & VM_PROT_EXECUTE) ? PTE2_X : PTE2_NX; l2_attr = vm_memattr_to_pte2(attr); l1_prot = ATTR_TO_L1(l2_prot); l1_attr = ATTR_TO_L1(l2_attr); /* Map all the pages. */ num = round_page(size); while (num > 0) { if ((((va | pa) & PTE1_OFFSET) == 0) && (num >= PTE1_SIZE)) { pte1p = kern_pte1(va); pte1_store(pte1p, PTE1_KERN(pa, l1_prot, l1_attr)); va += PTE1_SIZE; pa += PTE1_SIZE; num -= PTE1_SIZE; } else { pte2p = pmap_preboot_vtopte2(va); pte2_store(pte2p, PTE2_KERN(pa, l2_prot, l2_attr)); va += PAGE_SIZE; pa += PAGE_SIZE; num -= PAGE_SIZE; } } } /* * Extract from the kernel page table the physical address * that is mapped by the given virtual address "va". */ vm_paddr_t pmap_kextract(vm_offset_t va) { vm_paddr_t pa; pt1_entry_t pte1; pt2_entry_t pte2; pte1 = pte1_load(kern_pte1(va)); if (pte1_is_section(pte1)) { pa = pte1_pa(pte1) | (va & PTE1_OFFSET); } else if (pte1_is_link(pte1)) { /* * We should beware of concurrent promotion that changes * pte1 at this point. However, it's not a problem as PT2 * page is preserved by promotion in PT2TAB. So even if * it happens, using of PT2MAP is still safe. * * QQQ: However, concurrent removing is a problem which * ends in abort on PT2MAP space. Locking must be used * to deal with this. */ pte2 = pte2_load(pt2map_entry(va)); pa = pte2_pa(pte2) | (va & PTE2_OFFSET); } else { panic("%s: va %#x pte1 %#x", __func__, va, pte1); } return (pa); } /* * Extract from the kernel page table the physical address * that is mapped by the given virtual address "va". Also * return L2 page table entry which maps the address. * * This is only intended to be used for panic dumps. */ vm_paddr_t pmap_dump_kextract(vm_offset_t va, pt2_entry_t *pte2p) { vm_paddr_t pa; pt1_entry_t pte1; pt2_entry_t pte2; pte1 = pte1_load(kern_pte1(va)); if (pte1_is_section(pte1)) { pa = pte1_pa(pte1) | (va & PTE1_OFFSET); pte2 = pa | ATTR_TO_L2(pte1) | PTE2_V; } else if (pte1_is_link(pte1)) { pte2 = pte2_load(pt2map_entry(va)); pa = pte2_pa(pte2); } else { pte2 = 0; pa = 0; } if (pte2p != NULL) *pte2p = pte2; return (pa); } /***************************************************************************** * * PMAP second stage initialization and utility functions * for bootstrap epoch. * * After pmap_bootstrap() is called, the following functions for * mappings can be used: * * void pmap_kenter(vm_offset_t va, vm_paddr_t pa); * void pmap_kremove(vm_offset_t va); * vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, * int prot); * * NOTE: This is not SMP coherent stage. And physical page allocation is not * allowed during this stage. * *****************************************************************************/ /* * Initialize kernel PMAP locks and lists, kernel_pmap itself, and * reserve various virtual spaces for temporary mappings. */ void pmap_bootstrap(vm_offset_t firstaddr) { pt2_entry_t *unused __unused; struct sysmaps *sysmaps; u_int i; /* * Initialize the kernel pmap (which is statically allocated). */ PMAP_LOCK_INIT(kernel_pmap); kernel_l1pa = (vm_paddr_t)kern_pt1; /* for libkvm */ kernel_pmap->pm_pt1 = kern_pt1; kernel_pmap->pm_pt2tab = kern_pt2tab; CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */ TAILQ_INIT(&kernel_pmap->pm_pvchunk); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); LIST_INIT(&allpmaps); /* * Request a spin mutex so that changes to allpmaps cannot be * preempted by smp_rendezvous_cpus(). */ mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); mtx_lock_spin(&allpmaps_lock); LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); mtx_unlock_spin(&allpmaps_lock); /* * Reserve some special page table entries/VA space for temporary * mapping of pages. */ #define SYSMAP(c, p, v, n) do { \ v = (c)pmap_preboot_reserve_pages(n); \ p = pt2map_entry((vm_offset_t)v); \ } while (0) /* * Local CMAP1/CMAP2 are used for zeroing and copying pages. * Local CMAP3 is used for data cache cleaning. */ for (i = 0; i < MAXCPU; i++) { sysmaps = &sysmaps_pcpu[i]; mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1); SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1); SYSMAP(caddr_t, sysmaps->CMAP3, sysmaps->CADDR3, 1); } /* * Crashdump maps. */ SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS); /* * _tmppt is used for reading arbitrary physical pages via /dev/mem. */ SYSMAP(caddr_t, unused, _tmppt, 1); /* * PADDR1 and PADDR2 are used by pmap_pte2_quick() and pmap_pte2(), * respectively. PADDR3 is used by pmap_pte2_ddb(). */ SYSMAP(pt2_entry_t *, PMAP1, PADDR1, 1); SYSMAP(pt2_entry_t *, PMAP2, PADDR2, 1); #ifdef DDB SYSMAP(pt2_entry_t *, PMAP3, PADDR3, 1); #endif mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); /* * Note that in very short time in initarm(), we are going to * initialize phys_avail[] array and no further page allocation * can happen after that until vm subsystem will be initialized. */ kernel_vm_end_new = kernel_vm_end; virtual_end = vm_max_kernel_address; } static void pmap_init_qpages(void) { struct pcpu *pc; int i; CPU_FOREACH(i) { pc = pcpu_find(i); pc->pc_qmap_addr = kva_alloc(PAGE_SIZE); if (pc->pc_qmap_addr == 0) panic("%s: unable to allocate KVA", __func__); } } SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, pmap_init_qpages, NULL); /* * The function can already be use in second initialization stage. * As such, the function DOES NOT call pmap_growkernel() where PT2 * allocation can happen. So if used, be sure that PT2 for given * virtual address is allocated already! * * Add a wired page to the kva. * Note: not SMP coherent. */ static __inline void pmap_kenter_prot_attr(vm_offset_t va, vm_paddr_t pa, uint32_t prot, uint32_t attr) { pt1_entry_t *pte1p; pt2_entry_t *pte2p; pte1p = kern_pte1(va); if (!pte1_is_valid(pte1_load(pte1p))) { /* XXX - sections ?! */ /* * This is a very low level function, so PT2 and particularly * PT2PG associated with given virtual address must be already * allocated. It's a pain mainly during pmap initialization * stage. However, called after pmap initialization with * virtual address not under kernel_vm_end will lead to * the same misery. */ if (!pte2_is_valid(pte2_load(kern_pt2tab_entry(va)))) panic("%s: kernel PT2 not allocated!", __func__); } pte2p = pt2map_entry(va); pte2_store(pte2p, PTE2_KERN(pa, prot, attr)); } PMAP_INLINE void pmap_kenter(vm_offset_t va, vm_paddr_t pa) { pmap_kenter_prot_attr(va, pa, PTE2_AP_KRW, PTE2_ATTR_DEFAULT); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt2_entry_t *pte2p; pte2p = pt2map_entry(va); pte2_clear(pte2p); } /* * Share new kernel PT2PG with all pmaps. * The caller is responsible for maintaining TLB consistency. */ static void pmap_kenter_pt2tab(vm_offset_t va, pt2_entry_t npte2) { pmap_t pmap; pt2_entry_t *pte2p; mtx_lock_spin(&allpmaps_lock); LIST_FOREACH(pmap, &allpmaps, pm_list) { pte2p = pmap_pt2tab_entry(pmap, va); pt2tab_store(pte2p, npte2); } mtx_unlock_spin(&allpmaps_lock); } /* * Share new kernel PTE1 with all pmaps. * The caller is responsible for maintaining TLB consistency. */ static void pmap_kenter_pte1(vm_offset_t va, pt1_entry_t npte1) { pmap_t pmap; pt1_entry_t *pte1p; mtx_lock_spin(&allpmaps_lock); LIST_FOREACH(pmap, &allpmaps, pm_list) { pte1p = pmap_pte1(pmap, va); pte1_store(pte1p, npte1); } mtx_unlock_spin(&allpmaps_lock); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. * * NOTE: Read the comments above pmap_kenter_prot_attr() as * the function is used herein! */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { vm_offset_t va, sva; vm_paddr_t pte1_offset; pt1_entry_t npte1; uint32_t l1prot, l2prot; uint32_t l1attr, l2attr; PDEBUG(1, printf("%s: virt = %#x, start = %#x, end = %#x (size = %#x)," " prot = %d\n", __func__, *virt, start, end, end - start, prot)); l2prot = (prot & VM_PROT_WRITE) ? PTE2_AP_KRW : PTE2_AP_KR; l2prot |= (prot & VM_PROT_EXECUTE) ? PTE2_X : PTE2_NX; l1prot = ATTR_TO_L1(l2prot); l2attr = PTE2_ATTR_DEFAULT; l1attr = ATTR_TO_L1(l2attr); va = *virt; /* * Does the physical address range's size and alignment permit at * least one section mapping to be created? */ pte1_offset = start & PTE1_OFFSET; if ((end - start) - ((PTE1_SIZE - pte1_offset) & PTE1_OFFSET) >= PTE1_SIZE) { /* * Increase the starting virtual address so that its alignment * does not preclude the use of section mappings. */ if ((va & PTE1_OFFSET) < pte1_offset) va = pte1_trunc(va) + pte1_offset; else if ((va & PTE1_OFFSET) > pte1_offset) va = pte1_roundup(va) + pte1_offset; } sva = va; while (start < end) { if ((start & PTE1_OFFSET) == 0 && end - start >= PTE1_SIZE) { KASSERT((va & PTE1_OFFSET) == 0, ("%s: misaligned va %#x", __func__, va)); npte1 = PTE1_KERN(start, l1prot, l1attr); pmap_kenter_pte1(va, npte1); va += PTE1_SIZE; start += PTE1_SIZE; } else { pmap_kenter_prot_attr(va, start, l2prot, l2attr); va += PAGE_SIZE; start += PAGE_SIZE; } } tlb_flush_range(sva, va - sva); *virt = va; return (sva); } /* * Make a temporary mapping for a physical address. * This is only intended to be used for panic dumps. */ void * pmap_kenter_temporary(vm_paddr_t pa, int i) { vm_offset_t va; /* QQQ: 'i' should be less or equal to MAXDUMPPGS. */ va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); pmap_kenter(va, pa); tlb_flush_local(va); return ((void *)crashdumpmap); } /************************************* * * TLB & cache maintenance routines. * *************************************/ /* * We inline these within pmap.c for speed. */ PMAP_INLINE void pmap_tlb_flush(pmap_t pmap, vm_offset_t va) { if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active)) tlb_flush(va); } PMAP_INLINE void pmap_tlb_flush_range(pmap_t pmap, vm_offset_t sva, vm_size_t size) { if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active)) tlb_flush_range(sva, size); } /* * Abuse the pte2 nodes for unmapped kva to thread a kva freelist through. * Requirements: * - Must deal with pages in order to ensure that none of the PTE2_* bits * are ever set, PTE2_V in particular. * - Assumes we can write to pte2s without pte2_store() atomic ops. * - Assumes nothing will ever test these addresses for 0 to indicate * no mapping instead of correctly checking PTE2_V. * - Assumes a vm_offset_t will fit in a pte2 (true for arm). * Because PTE2_V is never set, there can be no mappings to invalidate. */ static vm_offset_t pmap_pte2list_alloc(vm_offset_t *head) { pt2_entry_t *pte2p; vm_offset_t va; va = *head; if (va == 0) panic("pmap_ptelist_alloc: exhausted ptelist KVA"); pte2p = pt2map_entry(va); *head = *pte2p; if (*head & PTE2_V) panic("%s: va with PTE2_V set!", __func__); *pte2p = 0; return (va); } static void pmap_pte2list_free(vm_offset_t *head, vm_offset_t va) { pt2_entry_t *pte2p; if (va & PTE2_V) panic("%s: freeing va with PTE2_V set!", __func__); pte2p = pt2map_entry(va); *pte2p = *head; /* virtual! PTE2_V is 0 though */ *head = va; } static void pmap_pte2list_init(vm_offset_t *head, void *base, int npages) { int i; vm_offset_t va; *head = 0; for (i = npages - 1; i >= 0; i--) { va = (vm_offset_t)base + i * PAGE_SIZE; pmap_pte2list_free(head, va); } } /***************************************************************************** * * PMAP third and final stage initialization. * * After pmap_init() is called, PMAP subsystem is fully initialized. * *****************************************************************************/ SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0, "Max number of PV entries"); SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0, "Page share factor per proc"); static u_long nkpt2pg = NKPT2PG; SYSCTL_ULONG(_vm_pmap, OID_AUTO, nkpt2pg, CTLFLAG_RD, &nkpt2pg, 0, "Pre-allocated pages for kernel PT2s"); static int sp_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, sp_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &sp_enabled, 0, "Are large page mappings enabled?"); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pte1, CTLFLAG_RD, 0, "1MB page mapping counters"); static u_long pmap_pte1_demotions; SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pte1_demotions, 0, "1MB page demotions"); static u_long pmap_pte1_mappings; SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, mappings, CTLFLAG_RD, &pmap_pte1_mappings, 0, "1MB page mappings"); static u_long pmap_pte1_p_failures; SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_pte1_p_failures, 0, "1MB page promotion failures"); static u_long pmap_pte1_promotions; SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, promotions, CTLFLAG_RD, &pmap_pte1_promotions, 0, "1MB page promotions"); static u_long pmap_pte1_kern_demotions; SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, kern_demotions, CTLFLAG_RD, &pmap_pte1_kern_demotions, 0, "1MB page kernel demotions"); static u_long pmap_pte1_kern_promotions; SYSCTL_ULONG(_vm_pmap_pte1, OID_AUTO, kern_promotions, CTLFLAG_RD, &pmap_pte1_kern_promotions, 0, "1MB page kernel promotions"); static __inline ttb_entry_t pmap_ttb_get(pmap_t pmap) { return (vtophys(pmap->pm_pt1) | ttb_flags); } /* * Initialize a vm_page's machine-dependent fields. * * Variations: * 1. Pages for L2 page tables are always not managed. So, pv_list and * pt2_wirecount can share same physical space. However, proper * initialization on a page alloc for page tables and reinitialization * on the page free must be ensured. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); pt2_wirecount_init(m); m->md.pat_mode = VM_MEMATTR_DEFAULT; } /* * Virtualization for faster way how to zero whole page. */ static __inline void pagezero(void *page) { bzero(page, PAGE_SIZE); } /* * Zero L2 page table page. * Use same KVA as in pmap_zero_page(). */ static __inline vm_paddr_t pmap_pt2pg_zero(vm_page_t m) { vm_paddr_t pa; struct sysmaps *sysmaps; pa = VM_PAGE_TO_PHYS(m); /* * XXX: For now, we map whole page even if it's already zero, * to sync it even if the sync is only DSB. */ sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (pte2_load(sysmaps->CMAP2) != 0) panic("%s: CMAP2 busy", __func__); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(pa, PTE2_AP_KRW, vm_page_pte2_attr(m))); /* Even VM_ALLOC_ZERO request is only advisory. */ if ((m->flags & PG_ZERO) == 0) pagezero(sysmaps->CADDR2); pte2_sync_range((pt2_entry_t *)sysmaps->CADDR2, PAGE_SIZE); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); return (pa); } /* * Init just allocated page as L2 page table(s) holder * and return its physical address. */ static __inline vm_paddr_t pmap_pt2pg_init(pmap_t pmap, vm_offset_t va, vm_page_t m) { vm_paddr_t pa; pt2_entry_t *pte2p; /* Check page attributes. */ if (m->md.pat_mode != pt_memattr) pmap_page_set_memattr(m, pt_memattr); /* Zero page and init wire counts. */ pa = pmap_pt2pg_zero(m); pt2_wirecount_init(m); /* * Map page to PT2MAP address space for given pmap. * Note that PT2MAP space is shared with all pmaps. */ if (pmap == kernel_pmap) pmap_kenter_pt2tab(va, PTE2_KPT(pa)); else { pte2p = pmap_pt2tab_entry(pmap, va); pt2tab_store(pte2p, PTE2_KPT_NG(pa)); } return (pa); } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { vm_size_t s; pt2_entry_t *pte2p, pte2; u_int i, pte1_idx, pv_npg; PDEBUG(1, printf("%s: phys_start = %#x\n", __func__, PHYSADDR)); /* * Initialize the vm page array entries for kernel pmap's * L2 page table pages allocated in advance. */ pte1_idx = pte1_index(KERNBASE - PT2MAP_SIZE); pte2p = kern_pt2tab_entry(KERNBASE - PT2MAP_SIZE); for (i = 0; i < nkpt2pg + NPG_IN_PT2TAB; i++, pte2p++) { vm_paddr_t pa; vm_page_t m; pte2 = pte2_load(pte2p); KASSERT(pte2_is_valid(pte2), ("%s: no valid entry", __func__)); pa = pte2_pa(pte2); m = PHYS_TO_VM_PAGE(pa); KASSERT(m >= vm_page_array && m < &vm_page_array[vm_page_array_size], ("%s: L2 page table page is out of range", __func__)); m->pindex = pte1_idx; m->phys_addr = pa; pte1_idx += NPT2_IN_PG; } /* * Initialize the address space (zone) for the pv entries. Set a * high water mark so that the system can recover from excessive * numbers of pv entries. */ TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count; TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); pv_entry_max = roundup(pv_entry_max, _NPCPV); pv_entry_high_water = 9 * (pv_entry_max / 10); /* * Are large page mappings enabled? */ TUNABLE_INT_FETCH("vm.pmap.sp_enabled", &sp_enabled); if (sp_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("%s: can't assign to pagesizes[1]", __func__)); pagesizes[1] = PTE1_SIZE; } /* * Calculate the size of the pv head table for sections. * Handle the possibility that "vm_phys_segs[...].end" is zero. * Note that the table is only for sections which could be promoted. */ first_managed_pa = pte1_trunc(vm_phys_segs[0].start); pv_npg = (pte1_trunc(vm_phys_segs[vm_phys_nsegs - 1].end - PAGE_SIZE) - first_managed_pa) / PTE1_SIZE + 1; /* * Allocate memory for the pv head table for sections. */ s = (vm_size_t)(pv_npg * sizeof(struct md_page)); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(kernel_arena, s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc); pv_chunkbase = (struct pv_chunk *)kva_alloc(PAGE_SIZE * pv_maxchunks); if (pv_chunkbase == NULL) panic("%s: not enough kvm for pv chunks", __func__); pmap_pte2list_init(&pv_vafree, pv_chunkbase, pv_maxchunks); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { u_int anychanged; pt2_entry_t *epte2p, *pte2p, pte2; vm_page_t m; vm_paddr_t pa; anychanged = 0; pte2p = pt2map_entry(sva); epte2p = pte2p + count; while (pte2p < epte2p) { m = *ma++; pa = VM_PAGE_TO_PHYS(m); pte2 = pte2_load(pte2p); if ((pte2_pa(pte2) != pa) || (pte2_attr(pte2) != vm_page_pte2_attr(m))) { anychanged++; pte2_store(pte2p, PTE2_KERN(pa, PTE2_AP_KRW, vm_page_pte2_attr(m))); } pte2p++; } if (__predict_false(anychanged)) tlb_flush_range(sva, count * PAGE_SIZE); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { pmap_kremove(va); va += PAGE_SIZE; } tlb_flush_range(sva, va - sva); } /* * Are we current address space or kernel? */ static __inline int pmap_is_current(pmap_t pmap) { return (pmap == kernel_pmap || (pmap == vmspace_pmap(curthread->td_proc->p_vmspace))); } /* * If the given pmap is not the current or kernel pmap, the returned * pte2 must be released by passing it to pmap_pte2_release(). */ static pt2_entry_t * pmap_pte2(pmap_t pmap, vm_offset_t va) { pt1_entry_t pte1; vm_paddr_t pt2pg_pa; pte1 = pte1_load(pmap_pte1(pmap, va)); if (pte1_is_section(pte1)) panic("%s: attempt to map PTE1", __func__); if (pte1_is_link(pte1)) { /* Are we current address space or kernel? */ if (pmap_is_current(pmap)) return (pt2map_entry(va)); /* Note that L2 page table size is not equal to PAGE_SIZE. */ pt2pg_pa = trunc_page(pte1_link_pa(pte1)); mtx_lock(&PMAP2mutex); if (pte2_pa(pte2_load(PMAP2)) != pt2pg_pa) { pte2_store(PMAP2, PTE2_KPT(pt2pg_pa)); tlb_flush((vm_offset_t)PADDR2); } return (PADDR2 + (arm32_btop(va) & (NPTE2_IN_PG - 1))); } return (NULL); } /* * Releases a pte2 that was obtained from pmap_pte2(). * Be prepared for the pte2p being NULL. */ static __inline void pmap_pte2_release(pt2_entry_t *pte2p) { if ((pt2_entry_t *)(trunc_page((vm_offset_t)pte2p)) == PADDR2) { mtx_unlock(&PMAP2mutex); } } /* * Super fast pmap_pte2 routine best used when scanning * the pv lists. This eliminates many coarse-grained * invltlb calls. Note that many of the pv list * scans are across different pmaps. It is very wasteful * to do an entire tlb flush for checking a single mapping. * * If the given pmap is not the current pmap, pvh_global_lock * must be held and curthread pinned to a CPU. */ static pt2_entry_t * pmap_pte2_quick(pmap_t pmap, vm_offset_t va) { pt1_entry_t pte1; vm_paddr_t pt2pg_pa; pte1 = pte1_load(pmap_pte1(pmap, va)); if (pte1_is_section(pte1)) panic("%s: attempt to map PTE1", __func__); if (pte1_is_link(pte1)) { /* Are we current address space or kernel? */ if (pmap_is_current(pmap)) return (pt2map_entry(va)); rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("%s: curthread not pinned", __func__)); /* Note that L2 page table size is not equal to PAGE_SIZE. */ pt2pg_pa = trunc_page(pte1_link_pa(pte1)); if (pte2_pa(pte2_load(PMAP1)) != pt2pg_pa) { pte2_store(PMAP1, PTE2_KPT(pt2pg_pa)); #ifdef SMP PMAP1cpu = PCPU_GET(cpuid); #endif tlb_flush_local((vm_offset_t)PADDR1); PMAP1changed++; } else #ifdef SMP if (PMAP1cpu != PCPU_GET(cpuid)) { PMAP1cpu = PCPU_GET(cpuid); tlb_flush_local((vm_offset_t)PADDR1); PMAP1changedcpu++; } else #endif PMAP1unchanged++; return (PADDR1 + (arm32_btop(va) & (NPTE2_IN_PG - 1))); } return (NULL); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { vm_paddr_t pa; pt1_entry_t pte1; pt2_entry_t *pte2p; PMAP_LOCK(pmap); pte1 = pte1_load(pmap_pte1(pmap, va)); if (pte1_is_section(pte1)) pa = pte1_pa(pte1) | (va & PTE1_OFFSET); else if (pte1_is_link(pte1)) { pte2p = pmap_pte2(pmap, va); pa = pte2_pa(pte2_load(pte2p)) | (va & PTE2_OFFSET); pmap_pte2_release(pte2p); } else pa = 0; PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { vm_paddr_t pa, lockpa; pt1_entry_t pte1; pt2_entry_t pte2, *pte2p; vm_page_t m; lockpa = 0; m = NULL; PMAP_LOCK(pmap); retry: pte1 = pte1_load(pmap_pte1(pmap, va)); if (pte1_is_section(pte1)) { if (!(pte1 & PTE1_RO) || !(prot & VM_PROT_WRITE)) { pa = pte1_pa(pte1) | (va & PTE1_OFFSET); if (vm_page_pa_tryrelock(pmap, pa, &lockpa)) goto retry; m = PHYS_TO_VM_PAGE(pa); vm_page_hold(m); } } else if (pte1_is_link(pte1)) { pte2p = pmap_pte2(pmap, va); pte2 = pte2_load(pte2p); pmap_pte2_release(pte2p); if (pte2_is_valid(pte2) && (!(pte2 & PTE2_RO) || !(prot & VM_PROT_WRITE))) { pa = pte2_pa(pte2); if (vm_page_pa_tryrelock(pmap, pa, &lockpa)) goto retry; m = PHYS_TO_VM_PAGE(pa); vm_page_hold(m); } } PA_UNLOCK_COND(lockpa); PMAP_UNLOCK(pmap); return (m); } /* * Grow the number of kernel L2 page table entries, if needed. */ void pmap_growkernel(vm_offset_t addr) { vm_page_t m; vm_paddr_t pt2pg_pa, pt2_pa; pt1_entry_t pte1; pt2_entry_t pte2; PDEBUG(1, printf("%s: addr = %#x\n", __func__, addr)); /* * All the time kernel_vm_end is first KVA for which underlying * L2 page table is either not allocated or linked from L1 page table * (not considering sections). Except for two possible cases: * * (1) in the very beginning as long as pmap_growkernel() was * not called, it could be first unused KVA (which is not * rounded up to PTE1_SIZE), * * (2) when all KVA space is mapped and kernel_map->max_offset * address is not rounded up to PTE1_SIZE. (For example, * it could be 0xFFFFFFFF.) */ kernel_vm_end = pte1_roundup(kernel_vm_end); mtx_assert(&kernel_map->system_mtx, MA_OWNED); addr = roundup2(addr, PTE1_SIZE); if (addr - 1 >= kernel_map->max_offset) addr = kernel_map->max_offset; while (kernel_vm_end < addr) { pte1 = pte1_load(kern_pte1(kernel_vm_end)); if (pte1_is_valid(pte1)) { kernel_vm_end += PTE1_SIZE; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } continue; } /* * kernel_vm_end_new is used in pmap_pinit() when kernel * mappings are entered to new pmap all at once to avoid race * between pmap_kenter_pte1() and kernel_vm_end increase. * The same aplies to pmap_kenter_pt2tab(). */ kernel_vm_end_new = kernel_vm_end + PTE1_SIZE; pte2 = pt2tab_load(kern_pt2tab_entry(kernel_vm_end)); if (!pte2_is_valid(pte2)) { /* * Install new PT2s page into kernel PT2TAB. */ m = vm_page_alloc(NULL, pte1_index(kernel_vm_end) & ~PT2PG_MASK, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) panic("%s: no memory to grow kernel", __func__); /* * QQQ: To link all new L2 page tables from L1 page * table now and so pmap_kenter_pte1() them * at once together with pmap_kenter_pt2tab() * could be nice speed up. However, * pmap_growkernel() does not happen so often... * QQQ: The other TTBR is another option. */ pt2pg_pa = pmap_pt2pg_init(kernel_pmap, kernel_vm_end, m); } else pt2pg_pa = pte2_pa(pte2); pt2_pa = page_pt2pa(pt2pg_pa, pte1_index(kernel_vm_end)); pmap_kenter_pte1(kernel_vm_end, PTE1_LINK(pt2_pa)); kernel_vm_end = kernel_vm_end_new; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } } } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = vm_max_kernel_address - KERNBASE; return (sysctl_handle_long(oidp, &ksize, 0, req)); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_size, "IU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = vm_max_kernel_address - kernel_vm_end; return (sysctl_handle_long(oidp, &kfree, 0, req)); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_free, "IU", "Amount of KVM free"); /*********************************************** * * Pmap allocation/deallocation routines. * ***********************************************/ /* * Initialize the pmap for the swapper process. */ void pmap_pinit0(pmap_t pmap) { PDEBUG(1, printf("%s: pmap = %p\n", __func__, pmap)); PMAP_LOCK_INIT(pmap); /* * Kernel page table directory and pmap stuff around is already * initialized, we are using it right now and here. So, finish * only PMAP structures initialization for process0 ... * * Since the L1 page table and PT2TAB is shared with the kernel pmap, * which is already included in the list "allpmaps", this pmap does * not need to be inserted into that list. */ pmap->pm_pt1 = kern_pt1; pmap->pm_pt2tab = kern_pt2tab; CPU_ZERO(&pmap->pm_active); PCPU_SET(curpmap, pmap); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); CPU_SET(0, &pmap->pm_active); } static __inline void pte1_copy_nosync(pt1_entry_t *spte1p, pt1_entry_t *dpte1p, vm_offset_t sva, vm_offset_t eva) { u_int idx, count; idx = pte1_index(sva); count = (pte1_index(eva) - idx + 1) * sizeof(pt1_entry_t); bcopy(spte1p + idx, dpte1p + idx, count); } static __inline void pt2tab_copy_nosync(pt2_entry_t *spte2p, pt2_entry_t *dpte2p, vm_offset_t sva, vm_offset_t eva) { u_int idx, count; idx = pt2tab_index(sva); count = (pt2tab_index(eva) - idx + 1) * sizeof(pt2_entry_t); bcopy(spte2p + idx, dpte2p + idx, count); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ int pmap_pinit(pmap_t pmap) { pt1_entry_t *pte1p; pt2_entry_t *pte2p; vm_paddr_t pa, pt2tab_pa; u_int i; PDEBUG(6, printf("%s: pmap = %p, pm_pt1 = %p\n", __func__, pmap, pmap->pm_pt1)); /* * No need to allocate L2 page table space yet but we do need * a valid L1 page table and PT2TAB table. * * Install shared kernel mappings to these tables. It's a little * tricky as some parts of KVA are reserved for vectors, devices, * and whatever else. These parts are supposed to be above * vm_max_kernel_address. Thus two regions should be installed: * * (1) . * * QQQ: The second region should be stable enough to be installed * only once in time when the tables are allocated. * QQQ: Maybe copy of both regions at once could be faster ... * QQQ: Maybe the other TTBR is an option. * * Finally, install own PT2TAB table to these tables. */ if (pmap->pm_pt1 == NULL) { pmap->pm_pt1 = (pt1_entry_t *)kmem_alloc_contig(kernel_arena, NB_IN_PT1, M_NOWAIT | M_ZERO, 0, -1UL, NB_IN_PT1, 0, pt_memattr); if (pmap->pm_pt1 == NULL) return (0); } if (pmap->pm_pt2tab == NULL) { /* * QQQ: (1) PT2TAB must be contiguous. If PT2TAB is one page * only, what should be the only size for 32 bit systems, * then we could allocate it with vm_page_alloc() and all * the stuff needed as other L2 page table pages. * (2) Note that a process PT2TAB is special L2 page table * page. Its mapping in kernel_arena is permanent and can * be used no matter which process is current. Its mapping * in PT2MAP can be used only for current process. */ pmap->pm_pt2tab = (pt2_entry_t *)kmem_alloc_attr(kernel_arena, NB_IN_PT2TAB, M_NOWAIT | M_ZERO, 0, -1UL, pt_memattr); if (pmap->pm_pt2tab == NULL) { /* * QQQ: As struct pmap is allocated from UMA with * UMA_ZONE_NOFREE flag, it's important to leave * no allocation in pmap if initialization failed. */ kmem_free(kernel_arena, (vm_offset_t)pmap->pm_pt1, NB_IN_PT1); pmap->pm_pt1 = NULL; return (0); } /* * QQQ: Each L2 page table page vm_page_t has pindex set to * pte1 index of virtual address mapped by this page. * It's not valid for non kernel PT2TABs themselves. * The pindex of these pages can not be altered because * of the way how they are allocated now. However, it * should not be a problem. */ } mtx_lock_spin(&allpmaps_lock); /* * To avoid race with pmap_kenter_pte1() and pmap_kenter_pt2tab(), * kernel_vm_end_new is used here instead of kernel_vm_end. */ pte1_copy_nosync(kern_pt1, pmap->pm_pt1, KERNBASE, kernel_vm_end_new - 1); pte1_copy_nosync(kern_pt1, pmap->pm_pt1, vm_max_kernel_address, 0xFFFFFFFF); pt2tab_copy_nosync(kern_pt2tab, pmap->pm_pt2tab, KERNBASE, kernel_vm_end_new - 1); pt2tab_copy_nosync(kern_pt2tab, pmap->pm_pt2tab, vm_max_kernel_address, 0xFFFFFFFF); LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); mtx_unlock_spin(&allpmaps_lock); /* * Store PT2MAP PT2 pages (a.k.a. PT2TAB) in PT2TAB itself. * I.e. self reference mapping. The PT2TAB is private, however mapped * into shared PT2MAP space, so the mapping should be not global. */ pt2tab_pa = vtophys(pmap->pm_pt2tab); pte2p = pmap_pt2tab_entry(pmap, (vm_offset_t)PT2MAP); for (pa = pt2tab_pa, i = 0; i < NPG_IN_PT2TAB; i++, pa += PTE2_SIZE) { pt2tab_store(pte2p++, PTE2_KPT_NG(pa)); } /* Insert PT2MAP PT2s into pmap PT1. */ pte1p = pmap_pte1(pmap, (vm_offset_t)PT2MAP); for (pa = pt2tab_pa, i = 0; i < NPT2_IN_PT2TAB; i++, pa += NB_IN_PT2) { pte1_store(pte1p++, PTE1_LINK(pa)); } /* * Now synchronize new mapping which was made above. */ pte1_sync_range(pmap->pm_pt1, NB_IN_PT1); pte2_sync_range(pmap->pm_pt2tab, NB_IN_PT2TAB); CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); return (1); } #ifdef INVARIANTS static boolean_t pt2tab_user_is_empty(pt2_entry_t *tab) { u_int i, end; end = pt2tab_index(VM_MAXUSER_ADDRESS); for (i = 0; i < end; i++) if (tab[i] != 0) return (FALSE); return (TRUE); } #endif /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { #ifdef INVARIANTS vm_offset_t start, end; #endif KASSERT(pmap->pm_stats.resident_count == 0, ("%s: pmap resident count %ld != 0", __func__, pmap->pm_stats.resident_count)); KASSERT(pt2tab_user_is_empty(pmap->pm_pt2tab), ("%s: has allocated user PT2(s)", __func__)); KASSERT(CPU_EMPTY(&pmap->pm_active), ("%s: pmap %p is active on some CPU(s)", __func__, pmap)); mtx_lock_spin(&allpmaps_lock); LIST_REMOVE(pmap, pm_list); mtx_unlock_spin(&allpmaps_lock); #ifdef INVARIANTS start = pte1_index(KERNBASE) * sizeof(pt1_entry_t); end = (pte1_index(0xFFFFFFFF) + 1) * sizeof(pt1_entry_t); bzero((char *)pmap->pm_pt1 + start, end - start); start = pt2tab_index(KERNBASE) * sizeof(pt2_entry_t); end = (pt2tab_index(0xFFFFFFFF) + 1) * sizeof(pt2_entry_t); bzero((char *)pmap->pm_pt2tab + start, end - start); #endif /* * We are leaving PT1 and PT2TAB allocated on released pmap, * so hopefully UMA vmspace_zone will always be inited with * UMA_ZONE_NOFREE flag. */ } /********************************************************* * * L2 table pages and their pages management routines. * *********************************************************/ /* * Virtual interface for L2 page table wire counting. * * Each L2 page table in a page has own counter which counts a number of * valid mappings in a table. Global page counter counts mappings in all * tables in a page plus a single itself mapping in PT2TAB. * * During a promotion we leave the associated L2 page table counter * untouched, so the table (strictly speaking a page which holds it) * is never freed if promoted. * * If a page m->wire_count == 1 then no valid mappings exist in any L2 page * table in the page and the page itself is only mapped in PT2TAB. */ static __inline void pt2_wirecount_init(vm_page_t m) { u_int i; /* * Note: A page m is allocated with VM_ALLOC_WIRED flag and * m->wire_count should be already set correctly. * So, there is no need to set it again herein. */ for (i = 0; i < NPT2_IN_PG; i++) m->md.pt2_wirecount[i] = 0; } static __inline void pt2_wirecount_inc(vm_page_t m, uint32_t pte1_idx) { /* * Note: A just modificated pte2 (i.e. already allocated) * is acquiring one extra reference which must be * explicitly cleared. It influences the KASSERTs herein. * All L2 page tables in a page always belong to the same * pmap, so we allow only one extra reference for the page. */ KASSERT(m->md.pt2_wirecount[pte1_idx & PT2PG_MASK] < (NPTE2_IN_PT2 + 1), ("%s: PT2 is overflowing ...", __func__)); KASSERT(m->wire_count <= (NPTE2_IN_PG + 1), ("%s: PT2PG is overflowing ...", __func__)); m->wire_count++; m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]++; } static __inline void pt2_wirecount_dec(vm_page_t m, uint32_t pte1_idx) { KASSERT(m->md.pt2_wirecount[pte1_idx & PT2PG_MASK] != 0, ("%s: PT2 is underflowing ...", __func__)); KASSERT(m->wire_count > 1, ("%s: PT2PG is underflowing ...", __func__)); m->wire_count--; m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]--; } static __inline void pt2_wirecount_set(vm_page_t m, uint32_t pte1_idx, uint16_t count) { KASSERT(count <= NPTE2_IN_PT2, ("%s: invalid count %u", __func__, count)); KASSERT(m->wire_count > m->md.pt2_wirecount[pte1_idx & PT2PG_MASK], ("%s: PT2PG corrupting (%u, %u) ...", __func__, m->wire_count, m->md.pt2_wirecount[pte1_idx & PT2PG_MASK])); m->wire_count -= m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]; m->wire_count += count; m->md.pt2_wirecount[pte1_idx & PT2PG_MASK] = count; KASSERT(m->wire_count <= (NPTE2_IN_PG + 1), ("%s: PT2PG is overflowed (%u) ...", __func__, m->wire_count)); } static __inline uint32_t pt2_wirecount_get(vm_page_t m, uint32_t pte1_idx) { return (m->md.pt2_wirecount[pte1_idx & PT2PG_MASK]); } static __inline boolean_t pt2_is_empty(vm_page_t m, vm_offset_t va) { return (m->md.pt2_wirecount[pte1_index(va) & PT2PG_MASK] == 0); } static __inline boolean_t pt2_is_full(vm_page_t m, vm_offset_t va) { return (m->md.pt2_wirecount[pte1_index(va) & PT2PG_MASK] == NPTE2_IN_PT2); } static __inline boolean_t pt2pg_is_empty(vm_page_t m) { return (m->wire_count == 1); } /* * This routine is called if the L2 page table * is not mapped correctly. */ static vm_page_t _pmap_allocpte2(pmap_t pmap, vm_offset_t va, u_int flags) { uint32_t pte1_idx; pt1_entry_t *pte1p; pt2_entry_t pte2; vm_page_t m; vm_paddr_t pt2pg_pa, pt2_pa; pte1_idx = pte1_index(va); pte1p = pmap->pm_pt1 + pte1_idx; KASSERT(pte1_load(pte1p) == 0, ("%s: pm_pt1[%#x] is not zero: %#x", __func__, pte1_idx, pte1_load(pte1p))); pte2 = pt2tab_load(pmap_pt2tab_entry(pmap, va)); if (!pte2_is_valid(pte2)) { /* * Install new PT2s page into pmap PT2TAB. */ m = vm_page_alloc(NULL, pte1_idx & ~PT2PG_MASK, VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) { if ((flags & PMAP_ENTER_NOSLEEP) == 0) { PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); VM_WAIT; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, * the L2 page table page may have been allocated. */ return (NULL); } pmap->pm_stats.resident_count++; pt2pg_pa = pmap_pt2pg_init(pmap, va, m); } else { pt2pg_pa = pte2_pa(pte2); m = PHYS_TO_VM_PAGE(pt2pg_pa); } pt2_wirecount_inc(m, pte1_idx); pt2_pa = page_pt2pa(pt2pg_pa, pte1_idx); pte1_store(pte1p, PTE1_LINK(pt2_pa)); return (m); } static vm_page_t pmap_allocpte2(pmap_t pmap, vm_offset_t va, u_int flags) { u_int pte1_idx; pt1_entry_t *pte1p, pte1; vm_page_t m; pte1_idx = pte1_index(va); retry: pte1p = pmap->pm_pt1 + pte1_idx; pte1 = pte1_load(pte1p); /* * This supports switching from a 1MB page to a * normal 4K page. */ if (pte1_is_section(pte1)) { (void)pmap_demote_pte1(pmap, pte1p, va); /* * Reload pte1 after demotion. * * Note: Demotion can even fail as either PT2 is not find for * the virtual address or PT2PG can not be allocated. */ pte1 = pte1_load(pte1p); } /* * If the L2 page table page is mapped, we just increment the * hold count, and activate it. */ if (pte1_is_link(pte1)) { m = PHYS_TO_VM_PAGE(pte1_link_pa(pte1)); pt2_wirecount_inc(m, pte1_idx); } else { /* * Here if the PT2 isn't mapped, or if it has * been deallocated. */ m = _pmap_allocpte2(pmap, va, flags); if (m == NULL && (flags & PMAP_ENTER_NOSLEEP) == 0) goto retry; } return (m); } static __inline void pmap_free_zero_pages(struct spglist *free) { vm_page_t m; while ((m = SLIST_FIRST(free)) != NULL) { SLIST_REMOVE_HEAD(free, plinks.s.ss); /* Preserve the page's PG_ZERO setting. */ vm_page_free_toq(m); } } /* * Schedule the specified unused L2 page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free) { /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ #ifdef PMAP_DEBUG pmap_zero_page_check(m); #endif m->flags |= PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Unwire L2 page tables page. */ static void pmap_unwire_pt2pg(pmap_t pmap, vm_offset_t va, vm_page_t m) { pt1_entry_t *pte1p, opte1 __unused; pt2_entry_t *pte2p; uint32_t i; KASSERT(pt2pg_is_empty(m), ("%s: pmap %p PT2PG %p wired", __func__, pmap, m)); /* * Unmap all L2 page tables in the page from L1 page table. * * QQQ: Individual L2 page tables (except the last one) can be unmapped * earlier. However, we are doing that this way. */ KASSERT(m->pindex == (pte1_index(va) & ~PT2PG_MASK), ("%s: pmap %p va %#x PT2PG %p bad index", __func__, pmap, va, m)); pte1p = pmap->pm_pt1 + m->pindex; for (i = 0; i < NPT2_IN_PG; i++, pte1p++) { KASSERT(m->md.pt2_wirecount[i] == 0, ("%s: pmap %p PT2 %u (PG %p) wired", __func__, pmap, i, m)); opte1 = pte1_load(pte1p); if (pte1_is_link(opte1)) { pte1_clear(pte1p); /* * Flush intermediate TLB cache. */ pmap_tlb_flush(pmap, (m->pindex + i) << PTE1_SHIFT); } #ifdef INVARIANTS else KASSERT((opte1 == 0) || pte1_is_section(opte1), ("%s: pmap %p va %#x bad pte1 %x at %u", __func__, pmap, va, opte1, i)); #endif } /* * Unmap the page from PT2TAB. */ pte2p = pmap_pt2tab_entry(pmap, va); (void)pt2tab_load_clear(pte2p); pmap_tlb_flush(pmap, pt2map_pt2pg(va)); m->wire_count = 0; pmap->pm_stats.resident_count--; /* * This is a release store so that the ordinary store unmapping * the L2 page table page is globally performed before TLB shoot- * down is begun. */ atomic_subtract_rel_int(&vm_cnt.v_wire_count, 1); } /* * Decrements a L2 page table page's wire count, which is used to record the * number of valid page table entries within the page. If the wire count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static __inline boolean_t pmap_unwire_pt2(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { pt2_wirecount_dec(m, pte1_index(va)); if (pt2pg_is_empty(m)) { /* * QQQ: Wire count is zero, so whole page should be zero and * we can set PG_ZERO flag to it. * Note that when promotion is enabled, it takes some * more efforts. See pmap_unwire_pt2_all() below. */ pmap_unwire_pt2pg(pmap, va, m); pmap_add_delayed_free_list(m, free); return (TRUE); } else return (FALSE); } /* * Drop a L2 page table page's wire count at once, which is used to record * the number of valid L2 page table entries within the page. If the wire * count drops to zero, then the L2 page table page is unmapped. */ static __inline void pmap_unwire_pt2_all(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { u_int pte1_idx = pte1_index(va); KASSERT(m->pindex == (pte1_idx & ~PT2PG_MASK), ("%s: PT2 page's pindex is wrong", __func__)); KASSERT(m->wire_count > pt2_wirecount_get(m, pte1_idx), ("%s: bad pt2 wire count %u > %u", __func__, m->wire_count, pt2_wirecount_get(m, pte1_idx))); /* * It's possible that the L2 page table was never used. * It happened in case that a section was created without promotion. */ if (pt2_is_full(m, va)) { pt2_wirecount_set(m, pte1_idx, 0); /* * QQQ: We clear L2 page table now, so when L2 page table page * is going to be freed, we can set it PG_ZERO flag ... * This function is called only on section mappings, so * hopefully it's not to big overload. * * XXX: If pmap is current, existing PT2MAP mapping could be * used for zeroing. */ pmap_zero_page_area(m, page_pt2off(pte1_idx), NB_IN_PT2); } #ifdef INVARIANTS else KASSERT(pt2_is_empty(m, va), ("%s: PT2 is not empty (%u)", __func__, pt2_wirecount_get(m, pte1_idx))); #endif if (pt2pg_is_empty(m)) { pmap_unwire_pt2pg(pmap, va, m); pmap_add_delayed_free_list(m, free); } } /* * After removing a L2 page table entry, this routine is used to * conditionally free the page, and manage the hold/wire counts. */ static boolean_t pmap_unuse_pt2(pmap_t pmap, vm_offset_t va, struct spglist *free) { pt1_entry_t pte1; vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (FALSE); pte1 = pte1_load(pmap_pte1(pmap, va)); mpte = PHYS_TO_VM_PAGE(pte1_link_pa(pte1)); return (pmap_unwire_pt2(pmap, va, mpte, free)); } /************************************* * * Page management routines. * *************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 11); CTASSERT(_NPCPV == 336); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */ #define PC_FREE10 0x0000fffful /* Free values for index 10 */ static const uint32_t pc_freemask[_NPCM] = { PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE10 }; SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, "Current number of pv entries"); #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, "Current number of pv entry chunks"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, "Current number of pv entry chunks allocated"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, "Current number of pv entry chunks frees"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, "Number of times tried to get a chunk page but failed."); static long pv_entry_frees, pv_entry_allocs; static int pv_entry_spare; SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, "Current number of pv entry frees"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, "Current number of pv entry allocs"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, "Current number of spare pv entries"); #endif /* * Is given page managed? */ static __inline boolean_t is_managed(vm_paddr_t pa) { vm_offset_t pgnum; vm_page_t m; pgnum = atop(pa); if (pgnum >= first_page) { m = PHYS_TO_VM_PAGE(pa); if (m == NULL) return (FALSE); if ((m->oflags & VPO_UNMANAGED) == 0) return (TRUE); } return (FALSE); } static __inline boolean_t pte1_is_managed(pt1_entry_t pte1) { return (is_managed(pte1_pa(pte1))); } static __inline boolean_t pte2_is_managed(pt2_entry_t pte2) { return (is_managed(pte2_pa(pte2))); } /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. */ static vm_page_t pmap_pv_reclaim(pmap_t locked_pmap) { struct pch newtail; struct pv_chunk *pc; struct md_page *pvh; pt1_entry_t *pte1p; pmap_t pmap; pt2_entry_t *pte2p, tpte2; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint32_t inuse; int bit, field, freed; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); pmap = NULL; m_pc = NULL; SLIST_INIT(&free); TAILQ_INIT(&newtail); while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && (pv_vafree == 0 || SLIST_EMPTY(&free))) { TAILQ_REMOVE(&pv_chunks, pc, pc_lru); if (pmap != pc->pc_pmap) { if (pmap != NULL) { if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } pmap = pc->pc_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) PMAP_LOCK(pmap); else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) { pmap = NULL; TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); continue; } } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = ffs(inuse) - 1; pv = &pc->pc_pventry[field * 32 + bit]; va = pv->pv_va; pte1p = pmap_pte1(pmap, va); if (pte1_is_section(pte1_load(pte1p))) continue; pte2p = pmap_pte2(pmap, va); tpte2 = pte2_load(pte2p); if ((tpte2 & PTE2_W) == 0) tpte2 = pte2_load_clear(pte2p); pmap_pte2_release(pte2p); if ((tpte2 & PTE2_W) != 0) continue; KASSERT(tpte2 != 0, ("pmap_pv_reclaim: pmap %p va %#x zero pte", pmap, va)); pmap_tlb_flush(pmap, va); m = PHYS_TO_VM_PAGE(pte2_pa(tpte2)); if (pte2_is_dirty(tpte2)) vm_page_dirty(m); if ((tpte2 & PTE2_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt2(pmap, va, &free); freed++; } } if (freed == 0) { TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); continue; } /* Every freed mapping is for a 4 KB page. */ pmap->pm_stats.resident_count -= freed; PV_STAT(pv_entry_frees += freed); PV_STAT(pv_entry_spare += freed); pv_entry_count -= freed; TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); for (field = 0; field < _NPCM; field++) if (pc->pc_map[field] != pc_freemask[field]) { TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); /* * One freed pv entry in locked_pmap is * sufficient. */ if (pmap == locked_pmap) goto out; break; } if (field == _NPCM) { PV_STAT(pv_entry_spare -= _NPCPV); PV_STAT(pc_chunk_count--); PV_STAT(pc_chunk_frees++); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); pmap_qremove((vm_offset_t)pc, 1); pmap_pte2list_free(&pv_vafree, (vm_offset_t)pc); break; } } out: TAILQ_CONCAT(&pv_chunks, &newtail, pc_lru); if (pmap != NULL) { if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } if (m_pc == NULL && pv_vafree != 0 && SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->wire_count = 1; atomic_add_int(&vm_cnt.v_wire_count, 1); } pmap_free_zero_pages(&free); return (m_pc); } static void free_pv_chunk(struct pv_chunk *pc) { vm_page_t m; TAILQ_REMOVE(&pv_chunks, pc, pc_lru); PV_STAT(pv_entry_spare -= _NPCPV); PV_STAT(pc_chunk_count--); PV_STAT(pc_chunk_frees++); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); pmap_qremove((vm_offset_t)pc, 1); vm_page_unwire(m, PQ_NONE); vm_page_free(m); pmap_pte2list_free(&pv_vafree, (vm_offset_t)pc); } /* * Free the pv_entry back to the free list. */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(pv_entry_frees++); PV_STAT(pv_entry_spare++); pv_entry_count--; pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 32; bit = idx % 32; pc->pc_map[field] |= 1ul << bit; for (idx = 0; idx < _NPCM; idx++) if (pc->pc_map[idx] != pc_freemask[idx]) { /* * 98% of the time, pc is already at the head of the * list. If it isn't already, move it to the head. */ if (__predict_false(TAILQ_FIRST(&pmap->pm_pvchunk) != pc)) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } /* * Get a new pv_entry, allocating a block from the system * when needed. */ static pv_entry_t get_pv_entry(pmap_t pmap, boolean_t try) { static const struct timeval printinterval = { 60, 0 }; static struct timeval lastprint; int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(pv_entry_allocs++); pv_entry_count++; if (pv_entry_count > pv_entry_high_water) if (ratecheck(&lastprint, &printinterval)) printf("Approaching the limit on PV entries, consider " "increasing either the vm.pmap.shpgperproc or the " "vm.pmap.pv_entry_max tunable.\n"); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = ffs(pc->pc_map[field]) - 1; break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 32 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ for (field = 0; field < _NPCM; field++) if (pc->pc_map[field] != 0) { PV_STAT(pv_entry_spare--); return (pv); /* not full, return */ } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(pv_entry_spare--); return (pv); } } /* * Access to the pte2list "pv_vafree" is synchronized by the pvh * global lock. If "pv_vafree" is currently non-empty, it will * remain non-empty until pmap_pte2list_alloc() completes. */ if (pv_vafree == 0 || (m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { if (try) { pv_entry_count--; PV_STAT(pc_chunk_tryfail++); return (NULL); } m = pmap_pv_reclaim(pmap); if (m == NULL) goto retry; } PV_STAT(pc_chunk_count++); PV_STAT(pc_chunk_allocs++); pc = (struct pv_chunk *)pmap_pte2list_alloc(&pv_vafree); pmap_qenter((vm_offset_t)pc, &m, 1); pc->pc_pmap = pmap; pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */ for (field = 1; field < _NPCM; field++) pc->pc_map[field] = pc_freemask[field]; TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(pv_entry_spare += _NPCPV - 1); return (pv); } /* * Create a pv entry for page at pa for * (pmap, va). */ static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); pv = get_pv_entry(pmap, FALSE); pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); break; } } return (pv); } static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } static void pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) { struct md_page *pvh; rw_assert(&pvh_global_lock, RA_WLOCKED); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } static void pmap_pv_demote_pte1(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((pa & PTE1_OFFSET) == 0, ("pmap_pv_demote_pte1: pa is not 1mpage aligned")); /* * Transfer the 1mpage's pv entry for this mapping to the first * page's pv list. */ pvh = pa_to_pvh(pa); va = pte1_trunc(va); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_pte1: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); /* Instantiate the remaining NPTE2_IN_PT2 - 1 pv entries. */ va_last = va + PTE1_SIZE - PAGE_SIZE; do { m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_pte1: page %p is not managed", m)); va += PAGE_SIZE; pmap_insert_entry(pmap, va, m); } while (va < va_last); } static void pmap_pv_promote_pte1(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((pa & PTE1_OFFSET) == 0, ("pmap_pv_promote_pte1: pa is not 1mpage aligned")); /* * Transfer the first page's pv entry for this mapping to the * 1mpage's pv list. Aside from avoiding the cost of a call * to get_pv_entry(), a transfer avoids the possibility that * get_pv_entry() calls pmap_pv_reclaim() and that pmap_pv_reclaim() * removes one of the mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = pte1_trunc(va); pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_pte1: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); /* Free the remaining NPTE2_IN_PT2 - 1 pv entries. */ va_last = va + PTE1_SIZE - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } /* * Conditionally create a pv entry. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pv_entry_count < pv_entry_high_water && (pv = get_pv_entry(pmap, TRUE)) != NULL) { pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); return (TRUE); } else return (FALSE); } /* * Create the pv entries for each of the pages within a section. */ static boolean_t pmap_pv_insert_pte1(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); if (pv_entry_count < pv_entry_high_water && (pv = get_pv_entry(pmap, TRUE)) != NULL) { pv->pv_va = va; pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); return (TRUE); } else return (FALSE); } static inline void pmap_tlb_flush_pte1(pmap_t pmap, vm_offset_t va, pt1_entry_t npte1) { /* Kill all the small mappings or the big one only. */ if (pte1_is_section(npte1)) pmap_tlb_flush_range(pmap, pte1_trunc(va), PTE1_SIZE); else pmap_tlb_flush(pmap, pte1_trunc(va)); } /* * Update kernel pte1 on all pmaps. * * The following function is called only on one cpu with disabled interrupts. * In SMP case, smp_rendezvous_cpus() is used to stop other cpus. This way * nobody can invoke explicit hardware table walk during the update of pte1. * Unsolicited hardware table walk can still happen, invoked by speculative * data or instruction prefetch or even by speculative hardware table walk. * * The break-before-make approach should be implemented here. However, it's * not so easy to do that for kernel mappings as it would be unhappy to unmap * itself unexpectedly but voluntarily. */ static void pmap_update_pte1_kernel(vm_offset_t va, pt1_entry_t npte1) { pmap_t pmap; pt1_entry_t *pte1p; /* * Get current pmap. Interrupts should be disabled here * so PCPU_GET() is done atomically. */ pmap = PCPU_GET(curpmap); if (pmap == NULL) pmap = kernel_pmap; /* * (1) Change pte1 on current pmap. * (2) Flush all obsolete TLB entries on current CPU. * (3) Change pte1 on all pmaps. * (4) Flush all obsolete TLB entries on all CPUs in SMP case. */ pte1p = pmap_pte1(pmap, va); pte1_store(pte1p, npte1); /* Kill all the small mappings or the big one only. */ if (pte1_is_section(npte1)) { pmap_pte1_kern_promotions++; tlb_flush_range_local(pte1_trunc(va), PTE1_SIZE); } else { pmap_pte1_kern_demotions++; tlb_flush_local(pte1_trunc(va)); } /* * In SMP case, this function is called when all cpus are at smp * rendezvous, so there is no need to use 'allpmaps_lock' lock here. * In UP case, the function is called with this lock locked. */ LIST_FOREACH(pmap, &allpmaps, pm_list) { pte1p = pmap_pte1(pmap, va); pte1_store(pte1p, npte1); } #ifdef SMP /* Kill all the small mappings or the big one only. */ if (pte1_is_section(npte1)) tlb_flush_range(pte1_trunc(va), PTE1_SIZE); else tlb_flush(pte1_trunc(va)); #endif } #ifdef SMP struct pte1_action { vm_offset_t va; pt1_entry_t npte1; u_int update; /* CPU that updates the PTE1 */ }; static void pmap_update_pte1_action(void *arg) { struct pte1_action *act = arg; if (act->update == PCPU_GET(cpuid)) pmap_update_pte1_kernel(act->va, act->npte1); } /* * Change pte1 on current pmap. * Note that kernel pte1 must be changed on all pmaps. * * According to the architecture reference manual published by ARM, * the behaviour is UNPREDICTABLE when two or more TLB entries map the same VA. * According to this manual, UNPREDICTABLE behaviours must never happen in * a viable system. In contrast, on x86 processors, it is not specified which * TLB entry mapping the virtual address will be used, but the MMU doesn't * generate a bogus translation the way it does on Cortex-A8 rev 2 (Beaglebone * Black). * * It's a problem when either promotion or demotion is being done. The pte1 * update and appropriate TLB flush must be done atomically in general. */ static void pmap_change_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va, pt1_entry_t npte1) { if (pmap == kernel_pmap) { struct pte1_action act; sched_pin(); act.va = va; act.npte1 = npte1; act.update = PCPU_GET(cpuid); smp_rendezvous_cpus(all_cpus, smp_no_rendevous_barrier, pmap_update_pte1_action, NULL, &act); sched_unpin(); } else { register_t cspr; /* * Use break-before-make approach for changing userland * mappings. It can cause L1 translation aborts on other * cores in SMP case. So, special treatment is implemented * in pmap_fault(). To reduce the likelihood that another core * will be affected by the broken mapping, disable interrupts * until the mapping change is completed. */ cspr = disable_interrupts(PSR_I | PSR_F); pte1_clear(pte1p); pmap_tlb_flush_pte1(pmap, va, npte1); pte1_store(pte1p, npte1); restore_interrupts(cspr); } } #else static void pmap_change_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va, pt1_entry_t npte1) { if (pmap == kernel_pmap) { mtx_lock_spin(&allpmaps_lock); pmap_update_pte1_kernel(va, npte1); mtx_unlock_spin(&allpmaps_lock); } else { register_t cspr; /* * Use break-before-make approach for changing userland * mappings. It's absolutely safe in UP case when interrupts * are disabled. */ cspr = disable_interrupts(PSR_I | PSR_F); pte1_clear(pte1p); pmap_tlb_flush_pte1(pmap, va, npte1); pte1_store(pte1p, npte1); restore_interrupts(cspr); } } #endif /* * Tries to promote the NPTE2_IN_PT2, contiguous 4KB page mappings that are * within a single page table page (PT2) to a single 1MB page mapping. * For promotion to occur, two conditions must be met: (1) the 4KB page * mappings must map aligned, contiguous physical memory and (2) the 4KB page * mappings must have identical characteristics. * * Managed (PG_MANAGED) mappings within the kernel address space are not * promoted. The reason is that kernel PTE1s are replicated in each pmap but * pmap_remove_write(), pmap_clear_modify(), and pmap_clear_reference() only * read the PTE1 from the kernel pmap. */ static void pmap_promote_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va) { pt1_entry_t npte1; pt2_entry_t *fpte2p, fpte2, fpte2_fav; pt2_entry_t *pte2p, pte2; vm_offset_t pteva __unused; vm_page_t m __unused; PDEBUG(6, printf("%s(%p): try for va %#x pte1 %#x at %p\n", __func__, pmap, va, pte1_load(pte1p), pte1p)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Examine the first PTE2 in the specified PT2. Abort if this PTE2 is * either invalid, unused, or does not map the first 4KB physical page * within a 1MB page. */ fpte2p = pmap_pte2_quick(pmap, pte1_trunc(va)); fpte2 = pte2_load(fpte2p); if ((fpte2 & ((PTE2_FRAME & PTE1_OFFSET) | PTE2_A | PTE2_V)) != (PTE2_A | PTE2_V)) { pmap_pte1_p_failures++; CTR3(KTR_PMAP, "%s: failure(1) for va %#x in pmap %p", __func__, va, pmap); return; } if (pte2_is_managed(fpte2) && pmap == kernel_pmap) { pmap_pte1_p_failures++; CTR3(KTR_PMAP, "%s: failure(2) for va %#x in pmap %p", __func__, va, pmap); return; } if ((fpte2 & (PTE2_NM | PTE2_RO)) == PTE2_NM) { /* * When page is not modified, PTE2_RO can be set without * a TLB invalidation. */ fpte2 |= PTE2_RO; pte2_store(fpte2p, fpte2); } /* * Examine each of the other PTE2s in the specified PT2. Abort if this * PTE2 maps an unexpected 4KB physical page or does not have identical * characteristics to the first PTE2. */ fpte2_fav = (fpte2 & (PTE2_FRAME | PTE2_A | PTE2_V)); fpte2_fav += PTE1_SIZE - PTE2_SIZE; /* examine from the end */ for (pte2p = fpte2p + NPTE2_IN_PT2 - 1; pte2p > fpte2p; pte2p--) { pte2 = pte2_load(pte2p); if ((pte2 & (PTE2_FRAME | PTE2_A | PTE2_V)) != fpte2_fav) { pmap_pte1_p_failures++; CTR3(KTR_PMAP, "%s: failure(3) for va %#x in pmap %p", __func__, va, pmap); return; } if ((pte2 & (PTE2_NM | PTE2_RO)) == PTE2_NM) { /* * When page is not modified, PTE2_RO can be set * without a TLB invalidation. See note above. */ pte2 |= PTE2_RO; pte2_store(pte2p, pte2); pteva = pte1_trunc(va) | (pte2 & PTE1_OFFSET & PTE2_FRAME); CTR3(KTR_PMAP, "%s: protect for va %#x in pmap %p", __func__, pteva, pmap); } if ((pte2 & PTE2_PROMOTE) != (fpte2 & PTE2_PROMOTE)) { pmap_pte1_p_failures++; CTR3(KTR_PMAP, "%s: failure(4) for va %#x in pmap %p", __func__, va, pmap); return; } fpte2_fav -= PTE2_SIZE; } /* * The page table page in its current state will stay in PT2TAB * until the PTE1 mapping the section is demoted by pmap_demote_pte1() * or destroyed by pmap_remove_pte1(). * * Note that L2 page table size is not equal to PAGE_SIZE. */ m = PHYS_TO_VM_PAGE(trunc_page(pte1_link_pa(pte1_load(pte1p)))); KASSERT(m >= vm_page_array && m < &vm_page_array[vm_page_array_size], ("%s: PT2 page is out of range", __func__)); KASSERT(m->pindex == (pte1_index(va) & ~PT2PG_MASK), ("%s: PT2 page's pindex is wrong", __func__)); /* * Get pte1 from pte2 format. */ npte1 = (fpte2 & PTE1_FRAME) | ATTR_TO_L1(fpte2) | PTE1_V; /* * Promote the pv entries. */ if (pte2_is_managed(fpte2)) pmap_pv_promote_pte1(pmap, va, pte1_pa(npte1)); /* * Promote the mappings. */ pmap_change_pte1(pmap, pte1p, va, npte1); pmap_pte1_promotions++; CTR3(KTR_PMAP, "%s: success for va %#x in pmap %p", __func__, va, pmap); PDEBUG(6, printf("%s(%p): success for va %#x pte1 %#x(%#x) at %p\n", __func__, pmap, va, npte1, pte1_load(pte1p), pte1p)); } /* * Zero L2 page table page. */ static __inline void pmap_clear_pt2(pt2_entry_t *fpte2p) { pt2_entry_t *pte2p; for (pte2p = fpte2p; pte2p < fpte2p + NPTE2_IN_PT2; pte2p++) pte2_clear(pte2p); } /* * Removes a 1MB page mapping from the kernel pmap. */ static void pmap_remove_kernel_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va) { vm_page_t m; uint32_t pte1_idx; pt2_entry_t *fpte2p; vm_paddr_t pt2_pa; PMAP_LOCK_ASSERT(pmap, MA_OWNED); m = pmap_pt2_page(pmap, va); if (m == NULL) /* * QQQ: Is this function called only on promoted pte1? * We certainly do section mappings directly * (without promotion) in kernel !!! */ panic("%s: missing pt2 page", __func__); pte1_idx = pte1_index(va); /* * Initialize the L2 page table. */ fpte2p = page_pt2(pt2map_pt2pg(va), pte1_idx); pmap_clear_pt2(fpte2p); /* * Remove the mapping. */ pt2_pa = page_pt2pa(VM_PAGE_TO_PHYS(m), pte1_idx); pmap_kenter_pte1(va, PTE1_LINK(pt2_pa)); /* * QQQ: We do not need to invalidate PT2MAP mapping * as we did not change it. I.e. the L2 page table page * was and still is mapped the same way. */ } /* * Do the things to unmap a section in a process */ static void pmap_remove_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t sva, struct spglist *free) { pt1_entry_t opte1; struct md_page *pvh; vm_offset_t eva, va; vm_page_t m; PDEBUG(6, printf("%s(%p): va %#x pte1 %#x at %p\n", __func__, pmap, sva, pte1_load(pte1p), pte1p)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PTE1_OFFSET) == 0, ("%s: sva is not 1mpage aligned", __func__)); /* * Clear and invalidate the mapping. It should occupy one and only TLB * entry. So, pmap_tlb_flush() called with aligned address should be * sufficient. */ opte1 = pte1_load_clear(pte1p); pmap_tlb_flush(pmap, sva); if (pte1_is_wired(opte1)) pmap->pm_stats.wired_count -= PTE1_SIZE / PAGE_SIZE; pmap->pm_stats.resident_count -= PTE1_SIZE / PAGE_SIZE; if (pte1_is_managed(opte1)) { pvh = pa_to_pvh(pte1_pa(opte1)); pmap_pvh_free(pvh, pmap, sva); eva = sva + PTE1_SIZE; for (va = sva, m = PHYS_TO_VM_PAGE(pte1_pa(opte1)); va < eva; va += PAGE_SIZE, m++) { if (pte1_is_dirty(opte1)) vm_page_dirty(m); if (opte1 & PTE1_A) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } if (pmap == kernel_pmap) { /* * L2 page table(s) can't be removed from kernel map as * kernel counts on it (stuff around pmap_growkernel()). */ pmap_remove_kernel_pte1(pmap, pte1p, sva); } else { /* * Get associated L2 page table page. * It's possible that the page was never allocated. */ m = pmap_pt2_page(pmap, sva); if (m != NULL) pmap_unwire_pt2_all(pmap, sva, m, free); } } /* * Fills L2 page table page with mappings to consecutive physical pages. */ static __inline void pmap_fill_pt2(pt2_entry_t *fpte2p, pt2_entry_t npte2) { pt2_entry_t *pte2p; for (pte2p = fpte2p; pte2p < fpte2p + NPTE2_IN_PT2; pte2p++) { pte2_store(pte2p, npte2); npte2 += PTE2_SIZE; } } /* * Tries to demote a 1MB page mapping. If demotion fails, the * 1MB page mapping is invalidated. */ static boolean_t pmap_demote_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t va) { pt1_entry_t opte1, npte1; pt2_entry_t *fpte2p, npte2; vm_paddr_t pt2pg_pa, pt2_pa; vm_page_t m; struct spglist free; uint32_t pte1_idx, isnew = 0; PDEBUG(6, printf("%s(%p): try for va %#x pte1 %#x at %p\n", __func__, pmap, va, pte1_load(pte1p), pte1p)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); opte1 = pte1_load(pte1p); KASSERT(pte1_is_section(opte1), ("%s: opte1 not a section", __func__)); if ((opte1 & PTE1_A) == 0 || (m = pmap_pt2_page(pmap, va)) == NULL) { KASSERT(!pte1_is_wired(opte1), ("%s: PT2 page for a wired mapping is missing", __func__)); /* * Invalidate the 1MB page mapping and return * "failure" if the mapping was never accessed or the * allocation of the new page table page fails. */ if ((opte1 & PTE1_A) == 0 || (m = vm_page_alloc(NULL, pte1_index(va) & ~PT2PG_MASK, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED)) == NULL) { SLIST_INIT(&free); pmap_remove_pte1(pmap, pte1p, pte1_trunc(va), &free); pmap_free_zero_pages(&free); CTR3(KTR_PMAP, "%s: failure for va %#x in pmap %p", __func__, va, pmap); return (FALSE); } if (va < VM_MAXUSER_ADDRESS) pmap->pm_stats.resident_count++; isnew = 1; /* * We init all L2 page tables in the page even if * we are going to change everything for one L2 page * table in a while. */ pt2pg_pa = pmap_pt2pg_init(pmap, va, m); } else { if (va < VM_MAXUSER_ADDRESS) { if (pt2_is_empty(m, va)) isnew = 1; /* Demoting section w/o promotion. */ #ifdef INVARIANTS else KASSERT(pt2_is_full(m, va), ("%s: bad PT2 wire" " count %u", __func__, pt2_wirecount_get(m, pte1_index(va)))); #endif } } pt2pg_pa = VM_PAGE_TO_PHYS(m); pte1_idx = pte1_index(va); /* * If the pmap is current, then the PT2MAP can provide access to * the page table page (promoted L2 page tables are not unmapped). * Otherwise, temporarily map the L2 page table page (m) into * the kernel's address space at either PADDR1 or PADDR2. * * Note that L2 page table size is not equal to PAGE_SIZE. */ if (pmap_is_current(pmap)) fpte2p = page_pt2(pt2map_pt2pg(va), pte1_idx); else if (curthread->td_pinned > 0 && rw_wowned(&pvh_global_lock)) { if (pte2_pa(pte2_load(PMAP1)) != pt2pg_pa) { pte2_store(PMAP1, PTE2_KPT(pt2pg_pa)); #ifdef SMP PMAP1cpu = PCPU_GET(cpuid); #endif tlb_flush_local((vm_offset_t)PADDR1); PMAP1changed++; } else #ifdef SMP if (PMAP1cpu != PCPU_GET(cpuid)) { PMAP1cpu = PCPU_GET(cpuid); tlb_flush_local((vm_offset_t)PADDR1); PMAP1changedcpu++; } else #endif PMAP1unchanged++; fpte2p = page_pt2((vm_offset_t)PADDR1, pte1_idx); } else { mtx_lock(&PMAP2mutex); if (pte2_pa(pte2_load(PMAP2)) != pt2pg_pa) { pte2_store(PMAP2, PTE2_KPT(pt2pg_pa)); tlb_flush((vm_offset_t)PADDR2); } fpte2p = page_pt2((vm_offset_t)PADDR2, pte1_idx); } pt2_pa = page_pt2pa(pt2pg_pa, pte1_idx); npte1 = PTE1_LINK(pt2_pa); KASSERT((opte1 & PTE1_A) != 0, ("%s: opte1 is missing PTE1_A", __func__)); KASSERT((opte1 & (PTE1_NM | PTE1_RO)) != PTE1_NM, ("%s: opte1 has PTE1_NM", __func__)); /* * Get pte2 from pte1 format. */ npte2 = pte1_pa(opte1) | ATTR_TO_L2(opte1) | PTE2_V; /* * If the L2 page table page is new, initialize it. If the mapping * has changed attributes, update the page table entries. */ if (isnew != 0) { pt2_wirecount_set(m, pte1_idx, NPTE2_IN_PT2); pmap_fill_pt2(fpte2p, npte2); } else if ((pte2_load(fpte2p) & PTE2_PROMOTE) != (npte2 & PTE2_PROMOTE)) pmap_fill_pt2(fpte2p, npte2); KASSERT(pte2_pa(pte2_load(fpte2p)) == pte2_pa(npte2), ("%s: fpte2p and npte2 map different physical addresses", __func__)); if (fpte2p == PADDR2) mtx_unlock(&PMAP2mutex); /* * Demote the mapping. This pmap is locked. The old PTE1 has * PTE1_A set. If the old PTE1 has not PTE1_RO set, it also * has not PTE1_NM set. Thus, there is no danger of a race with * another processor changing the setting of PTE1_A and/or PTE1_NM * between the read above and the store below. */ pmap_change_pte1(pmap, pte1p, va, npte1); /* * Demote the pv entry. This depends on the earlier demotion * of the mapping. Specifically, the (re)creation of a per- * page pv entry might trigger the execution of pmap_pv_reclaim(), * which might reclaim a newly (re)created per-page pv entry * and destroy the associated mapping. In order to destroy * the mapping, the PTE1 must have already changed from mapping * the 1mpage to referencing the page table page. */ if (pte1_is_managed(opte1)) pmap_pv_demote_pte1(pmap, va, pte1_pa(opte1)); pmap_pte1_demotions++; CTR3(KTR_PMAP, "%s: success for va %#x in pmap %p", __func__, va, pmap); PDEBUG(6, printf("%s(%p): success for va %#x pte1 %#x(%#x) at %p\n", __func__, pmap, va, npte1, pte1_load(pte1p), pte1p)); return (TRUE); } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { pt1_entry_t *pte1p; pt2_entry_t *pte2p; pt2_entry_t npte2, opte2; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte2, om; boolean_t wired; va = trunc_page(va); mpte2 = NULL; wired = (flags & PMAP_ENTER_WIRED) != 0; KASSERT(va <= vm_max_kernel_address, ("%s: toobig", __func__)); KASSERT(va < UPT2V_MIN_ADDRESS || va >= UPT2V_MAX_ADDRESS, ("%s: invalid to pmap_enter page table pages (va: 0x%x)", __func__, va)); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); sched_pin(); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { mpte2 = pmap_allocpte2(pmap, va, flags); if (mpte2 == NULL) { KASSERT((flags & PMAP_ENTER_NOSLEEP) != 0, ("pmap_allocpte2 failed with sleep allowed")); sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (KERN_RESOURCE_SHORTAGE); } } pte1p = pmap_pte1(pmap, va); if (pte1_is_section(pte1_load(pte1p))) panic("%s: attempted on 1MB page", __func__); pte2p = pmap_pte2_quick(pmap, va); if (pte2p == NULL) panic("%s: invalid L1 page table entry va=%#x", __func__, va); om = NULL; pa = VM_PAGE_TO_PHYS(m); opte2 = pte2_load(pte2p); opa = pte2_pa(opte2); /* * Mapping has not changed, must be protection or wiring change. */ if (pte2_is_valid(opte2) && (opa == pa)) { /* * Wiring change, just update stats. We don't worry about * wiring PT2 pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT2 page will be also. */ if (wired && !pte2_is_wired(opte2)) pmap->pm_stats.wired_count++; else if (!wired && pte2_is_wired(opte2)) pmap->pm_stats.wired_count--; /* * Remove extra pte2 reference */ if (mpte2) pt2_wirecount_dec(mpte2, pte1_index(va)); if (pte2_is_managed(opte2)) om = m; goto validate; } /* * QQQ: We think that changing physical address on writeable mapping * is not safe. Well, maybe on kernel address space with correct * locking, it can make a sense. However, we have no idea why * anyone should do that on user address space. Are we wrong? */ KASSERT((opa == 0) || (opa == pa) || !pte2_is_valid(opte2) || ((opte2 & PTE2_RO) != 0), ("%s: pmap %p va %#x(%#x) opa %#x pa %#x - gotcha %#x %#x!", __func__, pmap, va, opte2, opa, pa, flags, prot)); pv = NULL; /* * Mapping has changed, invalidate old range and fall through to * handle validating new mapping. */ if (opa) { if (pte2_is_wired(opte2)) pmap->pm_stats.wired_count--; if (pte2_is_managed(opte2)) { om = PHYS_TO_VM_PAGE(opa); pv = pmap_pvh_remove(&om->md, pmap, va); } /* * Remove extra pte2 reference */ if (mpte2 != NULL) pt2_wirecount_dec(mpte2, va >> PTE1_SHIFT); } else pmap->pm_stats.resident_count++; /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva, ("%s: managed mapping within the clean submap", __func__)); if (pv == NULL) pv = get_pv_entry(pmap, FALSE); pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } else if (pv != NULL) free_pv_entry(pmap, pv); /* * Increment counters */ if (wired) pmap->pm_stats.wired_count++; validate: /* * Now validate mapping with desired protection/wiring. */ npte2 = PTE2(pa, PTE2_NM, vm_page_pte2_attr(m)); if (prot & VM_PROT_WRITE) { if (pte2_is_managed(npte2)) vm_page_aflag_set(m, PGA_WRITEABLE); } else npte2 |= PTE2_RO; if ((prot & VM_PROT_EXECUTE) == 0) npte2 |= PTE2_NX; if (wired) npte2 |= PTE2_W; if (va < VM_MAXUSER_ADDRESS) npte2 |= PTE2_U; if (pmap != kernel_pmap) npte2 |= PTE2_NG; /* * If the mapping or permission bits are different, we need * to update the pte2. * * QQQ: Think again and again what to do * if the mapping is going to be changed! */ if ((opte2 & ~(PTE2_NM | PTE2_A)) != (npte2 & ~(PTE2_NM | PTE2_A))) { /* * Sync icache if exec permission and attribute VM_MEMATTR_WB_WA * is set. Do it now, before the mapping is stored and made * valid for hardware table walk. If done later, there is a race * for other threads of current process in lazy loading case. * Don't do it for kernel memory which is mapped with exec * permission even if the memory isn't going to hold executable * code. The only time when icache sync is needed is after * kernel module is loaded and the relocation info is processed. * And it's done in elf_cpu_load_file(). * * QQQ: (1) Does it exist any better way where * or how to sync icache? * (2) Now, we do it on a page basis. */ if ((prot & VM_PROT_EXECUTE) && pmap != kernel_pmap && m->md.pat_mode == VM_MEMATTR_WB_WA && (opa != pa || (opte2 & PTE2_NX))) cache_icache_sync_fresh(va, pa, PAGE_SIZE); npte2 |= PTE2_A; if (flags & VM_PROT_WRITE) npte2 &= ~PTE2_NM; if (opte2 & PTE2_V) { /* Change mapping with break-before-make approach. */ opte2 = pte2_load_clear(pte2p); pmap_tlb_flush(pmap, va); pte2_store(pte2p, npte2); if (opte2 & PTE2_A) { if (pte2_is_managed(opte2)) vm_page_aflag_set(om, PGA_REFERENCED); } if (pte2_is_dirty(opte2)) { if (pte2_is_managed(opte2)) vm_page_dirty(om); } if (pte2_is_managed(opte2) && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } else pte2_store(pte2p, npte2); } #if 0 else { /* * QQQ: In time when both access and not mofified bits are * emulated by software, this should not happen. Some * analysis is need, if this really happen. Missing * tlb flush somewhere could be the reason. */ panic("%s: pmap %p va %#x opte2 %x npte2 %x !!", __func__, pmap, va, opte2, npte2); } #endif /* * If both the L2 page table page and the reservation are fully * populated, then attempt promotion. */ if ((mpte2 == NULL || pt2_is_full(mpte2, va)) && sp_enabled && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) pmap_promote_pte1(pmap, pte1p, va); sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } /* * Do the things to unmap a page in a process. */ static int pmap_remove_pte2(pmap_t pmap, pt2_entry_t *pte2p, vm_offset_t va, struct spglist *free) { pt2_entry_t opte2; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Clear and invalidate the mapping. */ opte2 = pte2_load_clear(pte2p); pmap_tlb_flush(pmap, va); KASSERT(pte2_is_valid(opte2), ("%s: pmap %p va %#x not link pte2 %#x", __func__, pmap, va, opte2)); if (opte2 & PTE2_W) pmap->pm_stats.wired_count -= 1; pmap->pm_stats.resident_count -= 1; if (pte2_is_managed(opte2)) { m = PHYS_TO_VM_PAGE(pte2_pa(opte2)); if (pte2_is_dirty(opte2)) vm_page_dirty(m); if (opte2 & PTE2_A) vm_page_aflag_set(m, PGA_REFERENCED); pmap_remove_entry(pmap, m, va); } return (pmap_unuse_pt2(pmap, va, free)); } /* * Remove a single page from a process address space. */ static void pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free) { pt2_entry_t *pte2p; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("%s: curthread not pinned", __func__)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((pte2p = pmap_pte2_quick(pmap, va)) == NULL || !pte2_is_valid(pte2_load(pte2p))) return; pmap_remove_pte2(pmap, pte2p, va, free); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t nextva; pt1_entry_t *pte1p, pte1; pt2_entry_t *pte2p, pte2; struct spglist free; /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; SLIST_INIT(&free); rw_wlock(&pvh_global_lock); sched_pin(); PMAP_LOCK(pmap); /* * Special handling of removing one page. A very common * operation and easy to short circuit some code. */ if (sva + PAGE_SIZE == eva) { pte1 = pte1_load(pmap_pte1(pmap, sva)); if (pte1_is_link(pte1)) { pmap_remove_page(pmap, sva, &free); goto out; } } for (; sva < eva; sva = nextva) { /* * Calculate address for next L2 page table. */ nextva = pte1_trunc(sva + PTE1_SIZE); if (nextva < sva) nextva = eva; if (pmap->pm_stats.resident_count == 0) break; pte1p = pmap_pte1(pmap, sva); pte1 = pte1_load(pte1p); /* * Weed out invalid mappings. Note: we assume that the L1 page * table is always allocated, and in kernel virtual. */ if (pte1 == 0) continue; if (pte1_is_section(pte1)) { /* * Are we removing the entire large page? If not, * demote the mapping and fall through. */ if (sva + PTE1_SIZE == nextva && eva >= nextva) { pmap_remove_pte1(pmap, pte1p, sva, &free); continue; } else if (!pmap_demote_pte1(pmap, pte1p, sva)) { /* The large page mapping was destroyed. */ continue; } #ifdef INVARIANTS else { /* Update pte1 after demotion. */ pte1 = pte1_load(pte1p); } #endif } KASSERT(pte1_is_link(pte1), ("%s: pmap %p va %#x pte1 %#x at %p" " is not link", __func__, pmap, sva, pte1, pte1p)); /* * Limit our scan to either the end of the va represented * by the current L2 page table page, or to the end of the * range being removed. */ if (nextva > eva) nextva = eva; for (pte2p = pmap_pte2_quick(pmap, sva); sva != nextva; pte2p++, sva += PAGE_SIZE) { pte2 = pte2_load(pte2p); if (!pte2_is_valid(pte2)) continue; if (pmap_remove_pte2(pmap, pte2p, sva, &free)) break; } } out: sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; pt2_entry_t *pte2p, opte2; pt1_entry_t *pte1p; vm_offset_t va; struct spglist free; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); SLIST_INIT(&free); rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, va); (void)pmap_demote_pte1(pmap, pte1p, va); PMAP_UNLOCK(pmap); } small_mappings: while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pmap->pm_stats.resident_count--; pte1p = pmap_pte1(pmap, pv->pv_va); KASSERT(!pte1_is_section(pte1_load(pte1p)), ("%s: found " "a 1mpage in page %p's pv list", __func__, m)); pte2p = pmap_pte2_quick(pmap, pv->pv_va); opte2 = pte2_load_clear(pte2p); pmap_tlb_flush(pmap, pv->pv_va); KASSERT(pte2_is_valid(opte2), ("%s: pmap %p va %x zero pte2", __func__, pmap, pv->pv_va)); if (pte2_is_wired(opte2)) pmap->pm_stats.wired_count--; if (opte2 & PTE2_A) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if (pte2_is_dirty(opte2)) vm_page_dirty(m); pmap_unuse_pt2(pmap, pv->pv_va, &free); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); sched_unpin(); rw_wunlock(&pvh_global_lock); pmap_free_zero_pages(&free); } /* * Just subroutine for pmap_remove_pages() to reasonably satisfy * good coding style, a.k.a. 80 character line width limit hell. */ static __inline void pmap_remove_pte1_quick(pmap_t pmap, pt1_entry_t pte1, pv_entry_t pv, struct spglist *free) { vm_paddr_t pa; vm_page_t m, mt, mpt2pg; struct md_page *pvh; pa = pte1_pa(pte1); m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("%s: vm_page_t %p addr mismatch %#x %#x", __func__, m, m->phys_addr, pa)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("%s: bad pte1 %#x", __func__, pte1)); if (pte1_is_dirty(pte1)) { for (mt = m; mt < &m[PTE1_SIZE / PAGE_SIZE]; mt++) vm_page_dirty(mt); } pmap->pm_stats.resident_count -= PTE1_SIZE / PAGE_SIZE; pvh = pa_to_pvh(pa); TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[PTE1_SIZE / PAGE_SIZE]; mt++) if (TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } mpt2pg = pmap_pt2_page(pmap, pv->pv_va); if (mpt2pg != NULL) pmap_unwire_pt2_all(pmap, pv->pv_va, mpt2pg, free); } /* * Just subroutine for pmap_remove_pages() to reasonably satisfy * good coding style, a.k.a. 80 character line width limit hell. */ static __inline void pmap_remove_pte2_quick(pmap_t pmap, pt2_entry_t pte2, pv_entry_t pv, struct spglist *free) { vm_paddr_t pa; vm_page_t m; struct md_page *pvh; pa = pte2_pa(pte2); m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("%s: vm_page_t %p addr mismatch %#x %#x", __func__, m, m->phys_addr, pa)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("%s: bad pte2 %#x", __func__, pte2)); if (pte2_is_dirty(pte2)) vm_page_dirty(m); pmap->pm_stats.resident_count--; TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(pa); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } pmap_unuse_pt2(pmap, pv->pv_va, free); } /* * Remove all pages from specified address space this aids process * exit speeds. Also, this code is special cased for current process * only, but can have the more generic (and slightly slower) mode enabled. * This is much faster than pmap_remove in the case of running down * an entire address space. */ void pmap_remove_pages(pmap_t pmap) { pt1_entry_t *pte1p, pte1; pt2_entry_t *pte2p, pte2; pv_entry_t pv; struct pv_chunk *pc, *npc; struct spglist free; int field, idx; int32_t bit; uint32_t inuse, bitmask; boolean_t allfree; /* * Assert that the given pmap is only active on the current * CPU. Unfortunately, we cannot block another CPU from * activating the pmap while this function is executing. */ KASSERT(pmap == vmspace_pmap(curthread->td_proc->p_vmspace), ("%s: non-current pmap %p", __func__, pmap)); #if defined(SMP) && defined(INVARIANTS) { cpuset_t other_cpus; sched_pin(); other_cpus = pmap->pm_active; CPU_CLR(PCPU_GET(cpuid), &other_cpus); sched_unpin(); KASSERT(CPU_EMPTY(&other_cpus), ("%s: pmap %p active on other cpus", __func__, pmap)); } #endif SLIST_INIT(&free); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); sched_pin(); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { KASSERT(pc->pc_pmap == pmap, ("%s: wrong pmap %p %p", __func__, pmap, pc->pc_pmap)); allfree = TRUE; for (field = 0; field < _NPCM; field++) { inuse = (~(pc->pc_map[field])) & pc_freemask[field]; while (inuse != 0) { bit = ffs(inuse) - 1; bitmask = 1UL << bit; idx = field * 32 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; /* * Note that we cannot remove wired pages * from a process' mapping at this time */ pte1p = pmap_pte1(pmap, pv->pv_va); pte1 = pte1_load(pte1p); if (pte1_is_section(pte1)) { if (pte1_is_wired(pte1)) { allfree = FALSE; continue; } pte1_clear(pte1p); pmap_remove_pte1_quick(pmap, pte1, pv, &free); } else if (pte1_is_link(pte1)) { pte2p = pt2map_entry(pv->pv_va); pte2 = pte2_load(pte2p); if (!pte2_is_valid(pte2)) { printf("%s: pmap %p va %#x " "pte2 %#x\n", __func__, pmap, pv->pv_va, pte2); panic("bad pte2"); } if (pte2_is_wired(pte2)) { allfree = FALSE; continue; } pte2_clear(pte2p); pmap_remove_pte2_quick(pmap, pte2, pv, &free); } else { printf("%s: pmap %p va %#x pte1 %#x\n", __func__, pmap, pv->pv_va, pte1); panic("bad pte1"); } /* Mark free */ PV_STAT(pv_entry_frees++); PV_STAT(pv_entry_spare++); pv_entry_count--; pc->pc_map[field] |= bitmask; } } if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } } tlb_flush_all_ng_local(); sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * This code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No L2 page table pages. * but is *MUCH* faster than pmap_enter... */ static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpt2pg) { pt2_entry_t *pte2p, pte2; vm_paddr_t pa; struct spglist free; uint32_t l2prot; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("%s: managed mapping within the clean submap", __func__)); rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * In the case that a L2 page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { u_int pte1_idx; pt1_entry_t pte1, *pte1p; vm_paddr_t pt2_pa; /* * Get L1 page table things. */ pte1_idx = pte1_index(va); pte1p = pmap_pte1(pmap, va); pte1 = pte1_load(pte1p); if (mpt2pg && (mpt2pg->pindex == (pte1_idx & ~PT2PG_MASK))) { /* * Each of NPT2_IN_PG L2 page tables on the page can * come here. Make sure that associated L1 page table * link is established. * * QQQ: It comes that we don't establish all links to * L2 page tables for newly allocated L2 page * tables page. */ KASSERT(!pte1_is_section(pte1), ("%s: pte1 %#x is section", __func__, pte1)); if (!pte1_is_link(pte1)) { pt2_pa = page_pt2pa(VM_PAGE_TO_PHYS(mpt2pg), pte1_idx); pte1_store(pte1p, PTE1_LINK(pt2_pa)); } pt2_wirecount_inc(mpt2pg, pte1_idx); } else { /* * If the L2 page table page is mapped, we just * increment the hold count, and activate it. */ if (pte1_is_section(pte1)) { return (NULL); } else if (pte1_is_link(pte1)) { mpt2pg = PHYS_TO_VM_PAGE(pte1_link_pa(pte1)); pt2_wirecount_inc(mpt2pg, pte1_idx); } else { mpt2pg = _pmap_allocpte2(pmap, va, PMAP_ENTER_NOSLEEP); if (mpt2pg == NULL) return (NULL); } } } else { mpt2pg = NULL; } /* * This call to pt2map_entry() makes the assumption that we are * entering the page into the current pmap. In order to support * quick entry into any pmap, one would likely use pmap_pte2_quick(). * But that isn't as quick as pt2map_entry(). */ pte2p = pt2map_entry(va); pte2 = pte2_load(pte2p); if (pte2_is_valid(pte2)) { if (mpt2pg != NULL) { /* * Remove extra pte2 reference */ pt2_wirecount_dec(mpt2pg, pte1_index(va)); mpt2pg = NULL; } return (NULL); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m)) { if (mpt2pg != NULL) { SLIST_INIT(&free); if (pmap_unwire_pt2(pmap, va, mpt2pg, &free)) { pmap_tlb_flush(pmap, va); pmap_free_zero_pages(&free); } mpt2pg = NULL; } return (NULL); } /* * Increment counters */ pmap->pm_stats.resident_count++; /* * Now validate mapping with RO protection */ pa = VM_PAGE_TO_PHYS(m); l2prot = PTE2_RO | PTE2_NM; if (va < VM_MAXUSER_ADDRESS) l2prot |= PTE2_U | PTE2_NG; if ((prot & VM_PROT_EXECUTE) == 0) l2prot |= PTE2_NX; else if (m->md.pat_mode == VM_MEMATTR_WB_WA && pmap != kernel_pmap) { /* * Sync icache if exec permission and attribute VM_MEMATTR_WB_WA * is set. QQQ: For more info, see comments in pmap_enter(). */ cache_icache_sync_fresh(va, pa, PAGE_SIZE); } pte2_store(pte2p, PTE2(pa, l2prot, vm_page_pte2_attr(m))); return (mpt2pg); } void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * Tries to create 1MB page mapping. Returns TRUE if successful and * FALSE otherwise. Fails if (1) a page table page cannot be allocated without * blocking, (2) a mapping already exists at the specified virtual address, or * (3) a pv entry cannot be allocated without reclaiming another pv entry. */ static boolean_t pmap_enter_pte1(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { pt1_entry_t *pte1p; vm_paddr_t pa; uint32_t l1prot; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); pte1p = pmap_pte1(pmap, va); if (pte1_is_valid(pte1_load(pte1p))) { CTR3(KTR_PMAP, "%s: failure for va %#lx in pmap %p", __func__, va, pmap); return (FALSE); } if ((m->oflags & VPO_UNMANAGED) == 0) { /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_pte1(pmap, va, VM_PAGE_TO_PHYS(m))) { CTR3(KTR_PMAP, "%s: failure for va %#lx in pmap %p", __func__, va, pmap); return (FALSE); } } /* * Increment counters. */ pmap->pm_stats.resident_count += PTE1_SIZE / PAGE_SIZE; /* * Map the section. * * QQQ: Why VM_PROT_WRITE is not evaluated and the mapping is * made readonly? */ pa = VM_PAGE_TO_PHYS(m); l1prot = PTE1_RO | PTE1_NM; if (va < VM_MAXUSER_ADDRESS) l1prot |= PTE1_U | PTE1_NG; if ((prot & VM_PROT_EXECUTE) == 0) l1prot |= PTE1_NX; else if (m->md.pat_mode == VM_MEMATTR_WB_WA && pmap != kernel_pmap) { /* * Sync icache if exec permission and attribute VM_MEMATTR_WB_WA * is set. QQQ: For more info, see comments in pmap_enter(). */ cache_icache_sync_fresh(va, pa, PTE1_SIZE); } pte1_store(pte1p, PTE1(pa, l1prot, ATTR_TO_L1(vm_page_pte2_attr(m)))); pmap_pte1_mappings++; CTR3(KTR_PMAP, "%s: success for va %#lx in pmap %p", __func__, va, pmap); return (TRUE); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_offset_t va; vm_page_t m, mpt2pg; vm_pindex_t diff, psize; PDEBUG(6, printf("%s: pmap %p start %#x end %#x m %p prot %#x\n", __func__, pmap, start, end, m_start, prot)); VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpt2pg = NULL; m = m_start; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & PTE1_OFFSET) == 0 && va + PTE1_SIZE <= end && m->psind == 1 && sp_enabled && pmap_enter_pte1(pmap, va, m, prot)) m = &m[PTE1_SIZE / PAGE_SIZE - 1]; else mpt2pg = pmap_enter_quick_locked(pmap, va, m, prot, mpt2pg); m = TAILQ_NEXT(m, listq); } rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { pt1_entry_t *pte1p; vm_paddr_t pa, pte2_pa; vm_page_t p; vm_memattr_t pat_mode; u_int l1attr, l1prot; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("%s: non-device object", __func__)); if ((addr & PTE1_OFFSET) == 0 && (size & PTE1_OFFSET) == 0) { if (!vm_object_populate(object, pindex, pindex + atop(size))) return; p = vm_page_lookup(object, pindex); KASSERT(p->valid == VM_PAGE_BITS_ALL, ("%s: invalid page %p", __func__, p)); pat_mode = p->md.pat_mode; /* * Abort the mapping if the first page is not physically * aligned to a 1MB page boundary. */ pte2_pa = VM_PAGE_TO_PHYS(p); if (pte2_pa & PTE1_OFFSET) return; /* * Skip the first page. Abort the mapping if the rest of * the pages are not physically contiguous or have differing * memory attributes. */ p = TAILQ_NEXT(p, listq); for (pa = pte2_pa + PAGE_SIZE; pa < pte2_pa + size; pa += PAGE_SIZE) { KASSERT(p->valid == VM_PAGE_BITS_ALL, ("%s: invalid page %p", __func__, p)); if (pa != VM_PAGE_TO_PHYS(p) || pat_mode != p->md.pat_mode) return; p = TAILQ_NEXT(p, listq); } /* * Map using 1MB pages. * * QQQ: Well, we are mapping a section, so same condition must * be hold like during promotion. It looks that only RW mapping * is done here, so readonly mapping must be done elsewhere. */ l1prot = PTE1_U | PTE1_NG | PTE1_RW | PTE1_M | PTE1_A; l1attr = ATTR_TO_L1(vm_memattr_to_pte2(pat_mode)); PMAP_LOCK(pmap); for (pa = pte2_pa; pa < pte2_pa + size; pa += PTE1_SIZE) { pte1p = pmap_pte1(pmap, addr); if (!pte1_is_valid(pte1_load(pte1p))) { pte1_store(pte1p, PTE1(pa, l1prot, l1attr)); pmap->pm_stats.resident_count += PTE1_SIZE / PAGE_SIZE; pmap_pte1_mappings++; } /* Else continue on if the PTE1 is already valid. */ addr += PTE1_SIZE; } PMAP_UNLOCK(pmap); } } /* * Do the things to protect a 1mpage in a process. */ static void pmap_protect_pte1(pmap_t pmap, pt1_entry_t *pte1p, vm_offset_t sva, vm_prot_t prot) { pt1_entry_t npte1, opte1; vm_offset_t eva, va; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PTE1_OFFSET) == 0, ("%s: sva is not 1mpage aligned", __func__)); opte1 = npte1 = pte1_load(pte1p); if (pte1_is_managed(opte1)) { eva = sva + PTE1_SIZE; for (va = sva, m = PHYS_TO_VM_PAGE(pte1_pa(opte1)); va < eva; va += PAGE_SIZE, m++) if (pte1_is_dirty(opte1)) vm_page_dirty(m); } if ((prot & VM_PROT_WRITE) == 0) npte1 |= PTE1_RO | PTE1_NM; if ((prot & VM_PROT_EXECUTE) == 0) npte1 |= PTE1_NX; /* * QQQ: Herein, execute permission is never set. * It only can be cleared. So, no icache * syncing is needed. */ if (npte1 != opte1) { pte1_store(pte1p, npte1); pmap_tlb_flush(pmap, sva); } } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { boolean_t pv_lists_locked; vm_offset_t nextva; pt1_entry_t *pte1p, pte1; pt2_entry_t *pte2p, opte2, npte2; KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } if ((prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE)) return; if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } PMAP_LOCK(pmap); for (; sva < eva; sva = nextva) { /* * Calculate address for next L2 page table. */ nextva = pte1_trunc(sva + PTE1_SIZE); if (nextva < sva) nextva = eva; pte1p = pmap_pte1(pmap, sva); pte1 = pte1_load(pte1p); /* * Weed out invalid mappings. Note: we assume that L1 page * page table is always allocated, and in kernel virtual. */ if (pte1 == 0) continue; if (pte1_is_section(pte1)) { /* * Are we protecting the entire large page? If not, * demote the mapping and fall through. */ if (sva + PTE1_SIZE == nextva && eva >= nextva) { pmap_protect_pte1(pmap, pte1p, sva, prot); continue; } else { if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { PMAP_UNLOCK(pmap); goto resume; } sched_pin(); } if (!pmap_demote_pte1(pmap, pte1p, sva)) { /* * The large page mapping * was destroyed. */ continue; } #ifdef INVARIANTS else { /* Update pte1 after demotion */ pte1 = pte1_load(pte1p); } #endif } } KASSERT(pte1_is_link(pte1), ("%s: pmap %p va %#x pte1 %#x at %p" " is not link", __func__, pmap, sva, pte1, pte1p)); /* * Limit our scan to either the end of the va represented * by the current L2 page table page, or to the end of the * range being protected. */ if (nextva > eva) nextva = eva; for (pte2p = pmap_pte2_quick(pmap, sva); sva != nextva; pte2p++, sva += PAGE_SIZE) { vm_page_t m; opte2 = npte2 = pte2_load(pte2p); if (!pte2_is_valid(opte2)) continue; if ((prot & VM_PROT_WRITE) == 0) { if (pte2_is_managed(opte2) && pte2_is_dirty(opte2)) { m = PHYS_TO_VM_PAGE(pte2_pa(opte2)); vm_page_dirty(m); } npte2 |= PTE2_RO | PTE2_NM; } if ((prot & VM_PROT_EXECUTE) == 0) npte2 |= PTE2_NX; /* * QQQ: Herein, execute permission is never set. * It only can be cleared. So, no icache * syncing is needed. */ if (npte2 != opte2) { pte2_store(pte2p, npte2); pmap_tlb_flush(pmap, sva); } } } if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * pmap_pvh_wired_mappings: * * Return the updated number "count" of managed mappings that are wired. */ static int pmap_pvh_wired_mappings(struct md_page *pvh, int count) { pmap_t pmap; pt1_entry_t pte1; pt2_entry_t pte2; pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1 = pte1_load(pmap_pte1(pmap, pv->pv_va)); if (pte1_is_section(pte1)) { if (pte1_is_wired(pte1)) count++; } else { KASSERT(pte1_is_link(pte1), ("%s: pte1 %#x is not link", __func__, pte1)); pte2 = pte2_load(pmap_pte2_quick(pmap, pv->pv_va)); if (pte2_is_wired(pte2)) count++; } PMAP_UNLOCK(pmap); } sched_unpin(); return (count); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { int count; count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&pvh_global_lock); count = pmap_pvh_wired_mappings(&m->md, count); if ((m->flags & PG_FICTITIOUS) == 0) { count = pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)), count); } rw_wunlock(&pvh_global_lock); return (count); } /* * Returns TRUE if any of the given mappings were used to modify * physical memory. Otherwise, returns FALSE. Both page and 1mpage * mappings are supported. */ static boolean_t pmap_is_modified_pvh(struct md_page *pvh) { pv_entry_t pv; pt1_entry_t pte1; pt2_entry_t pte2; pmap_t pmap; boolean_t rv; rw_assert(&pvh_global_lock, RA_WLOCKED); rv = FALSE; sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1 = pte1_load(pmap_pte1(pmap, pv->pv_va)); if (pte1_is_section(pte1)) { rv = pte1_is_dirty(pte1); } else { KASSERT(pte1_is_link(pte1), ("%s: pte1 %#x is not link", __func__, pte1)); pte2 = pte2_load(pmap_pte2_quick(pmap, pv->pv_va)); rv = pte2_is_dirty(pte2); } PMAP_UNLOCK(pmap); if (rv) break; } sched_unpin(); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no PTE2s can have PG_M set. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (FALSE); rw_wlock(&pvh_global_lock); rv = pmap_is_modified_pvh(&m->md) || ((m->flags & PG_FICTITIOUS) == 0 && pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); rw_wunlock(&pvh_global_lock); return (rv); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pt1_entry_t pte1; pt2_entry_t pte2; boolean_t rv; rv = FALSE; PMAP_LOCK(pmap); pte1 = pte1_load(pmap_pte1(pmap, addr)); if (pte1_is_link(pte1)) { pte2 = pte2_load(pt2map_entry(addr)); rv = !pte2_is_valid(pte2) ; } PMAP_UNLOCK(pmap); return (rv); } /* * Returns TRUE if any of the given mappings were referenced and FALSE * otherwise. Both page and 1mpage mappings are supported. */ static boolean_t pmap_is_referenced_pvh(struct md_page *pvh) { pv_entry_t pv; pt1_entry_t pte1; pt2_entry_t pte2; pmap_t pmap; boolean_t rv; rw_assert(&pvh_global_lock, RA_WLOCKED); rv = FALSE; sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1 = pte1_load(pmap_pte1(pmap, pv->pv_va)); if (pte1_is_section(pte1)) { rv = (pte1 & (PTE1_A | PTE1_V)) == (PTE1_A | PTE1_V); } else { pte2 = pte2_load(pmap_pte2_quick(pmap, pv->pv_va)); rv = (pte2 & (PTE2_A | PTE2_V)) == (PTE2_A | PTE2_V); } PMAP_UNLOCK(pmap); if (rv) break; } sched_unpin(); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); rw_wlock(&pvh_global_lock); rv = pmap_is_referenced_pvh(&m->md) || ((m->flags & PG_FICTITIOUS) == 0 && pmap_is_referenced_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); rw_wunlock(&pvh_global_lock); return (rv); } -#define PMAP_TS_REFERENCED_MAX 5 - /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. - * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; pt1_entry_t *pte1p, opte1; pt2_entry_t *pte2p, opte2; vm_paddr_t pa; int rtval = 0; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); pa = VM_PAGE_TO_PHYS(m); pvh = pa_to_pvh(pa); rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0 || (pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, pv->pv_va); opte1 = pte1_load(pte1p); if (pte1_is_dirty(opte1)) { /* * Although "opte1" is mapping a 1MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } if ((opte1 & PTE1_A) != 0) { /* * Since this reference bit is shared by 256 4KB pages, * it should not be cleared every time it is tested. * Apply a simple "hash" function on the physical page * number, the virtual section number, and the pmap * address to select one 4KB page out of the 256 * on which testing the reference bit will result * in clearing that bit. This function is designed * to avoid the selection of the same 4KB page * for every 1MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the section is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PTE1_SHIFT) ^ (uintptr_t)pmap) & (NPTE2_IN_PG - 1)) == 0 && !pte1_is_wired(opte1)) { pte1_clear_bit(pte1p, PTE1_A); pmap_tlb_flush(pmap, pv->pv_va); } rtval++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); } if (rtval >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, pv->pv_va); KASSERT(pte1_is_link(pte1_load(pte1p)), ("%s: not found a link in page %p's pv list", __func__, m)); pte2p = pmap_pte2_quick(pmap, pv->pv_va); opte2 = pte2_load(pte2p); if (pte2_is_dirty(opte2)) vm_page_dirty(m); if ((opte2 & PTE2_A) != 0) { pte2_clear_bit(pte2p, PTE2_A); pmap_tlb_flush(pmap, pv->pv_va); rtval++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && rtval < PMAP_TS_REFERENCED_MAX); out: sched_unpin(); rw_wunlock(&pvh_global_lock); return (rtval); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, * so there is no need to invalidate any TLB entries. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t nextva; pt1_entry_t *pte1p, pte1; pt2_entry_t *pte2p, pte2; boolean_t pv_lists_locked; if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } PMAP_LOCK(pmap); for (; sva < eva; sva = nextva) { nextva = pte1_trunc(sva + PTE1_SIZE); if (nextva < sva) nextva = eva; pte1p = pmap_pte1(pmap, sva); pte1 = pte1_load(pte1p); /* * Weed out invalid mappings. Note: we assume that L1 page * page table is always allocated, and in kernel virtual. */ if (pte1 == 0) continue; if (pte1_is_section(pte1)) { if (!pte1_is_wired(pte1)) panic("%s: pte1 %#x not wired", __func__, pte1); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + PTE1_SIZE == nextva && eva >= nextva) { pte1_clear_bit(pte1p, PTE1_W); pmap->pm_stats.wired_count -= PTE1_SIZE / PAGE_SIZE; continue; } else { if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { PMAP_UNLOCK(pmap); /* Repeat sva. */ goto resume; } sched_pin(); } if (!pmap_demote_pte1(pmap, pte1p, sva)) panic("%s: demotion failed", __func__); #ifdef INVARIANTS else { /* Update pte1 after demotion */ pte1 = pte1_load(pte1p); } #endif } } KASSERT(pte1_is_link(pte1), ("%s: pmap %p va %#x pte1 %#x at %p" " is not link", __func__, pmap, sva, pte1, pte1p)); /* * Limit our scan to either the end of the va represented * by the current L2 page table page, or to the end of the * range being protected. */ if (nextva > eva) nextva = eva; for (pte2p = pmap_pte2_quick(pmap, sva); sva != nextva; pte2p++, sva += PAGE_SIZE) { pte2 = pte2_load(pte2p); if (!pte2_is_valid(pte2)) continue; if (!pte2_is_wired(pte2)) panic("%s: pte2 %#x is missing PTE2_W", __func__, pte2); /* * PTE2_W must be cleared atomically. Although the pmap * lock synchronizes access to PTE2_W, another processor * could be changing PTE2_NM and/or PTE2_A concurrently. */ pte2_clear_bit(pte2p, PTE2_W); pmap->pm_stats.wired_count--; } } if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pv_entry_t next_pv, pv; pmap_t pmap; pt1_entry_t *pte1p; pt2_entry_t *pte2p, opte2; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, va); if (!(pte1_load(pte1p) & PTE1_RO)) (void)pmap_demote_pte1(pmap, pte1p, va); PMAP_UNLOCK(pmap); } small_mappings: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, pv->pv_va); KASSERT(!pte1_is_section(pte1_load(pte1p)), ("%s: found" " a section in page %p's pv list", __func__, m)); pte2p = pmap_pte2_quick(pmap, pv->pv_va); opte2 = pte2_load(pte2p); if (!(opte2 & PTE2_RO)) { pte2_store(pte2p, opte2 | PTE2_RO | PTE2_NM); if (pte2_is_dirty(opte2)) vm_page_dirty(m); pmap_tlb_flush(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); sched_unpin(); rw_wunlock(&pvh_global_lock); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { pt1_entry_t *pte1p, opte1; pt2_entry_t *pte2p, pte2; vm_offset_t pdnxt; vm_page_t m; boolean_t pv_lists_locked; if (advice != MADV_DONTNEED && advice != MADV_FREE) return; if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pdnxt = pte1_trunc(sva + PTE1_SIZE); if (pdnxt < sva) pdnxt = eva; pte1p = pmap_pte1(pmap, sva); opte1 = pte1_load(pte1p); if (!pte1_is_valid(opte1)) /* XXX */ continue; else if (pte1_is_section(opte1)) { if (!pte1_is_managed(opte1)) continue; if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { PMAP_UNLOCK(pmap); goto resume; } sched_pin(); } if (!pmap_demote_pte1(pmap, pte1p, sva)) { /* * The large page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Since the underlying L2 page * table is fully populated, this removal never * frees a L2 page table page. */ if (!pte1_is_wired(opte1)) { pte2p = pmap_pte2_quick(pmap, sva); KASSERT(pte2_is_valid(pte2_load(pte2p)), ("%s: invalid PTE2", __func__)); pmap_remove_pte2(pmap, pte2p, sva, NULL); } } if (pdnxt > eva) pdnxt = eva; for (pte2p = pmap_pte2_quick(pmap, sva); sva != pdnxt; pte2p++, sva += PAGE_SIZE) { pte2 = pte2_load(pte2p); if (!pte2_is_valid(pte2) || !pte2_is_managed(pte2)) continue; else if (pte2_is_dirty(pte2)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(pte2_pa(pte2)); vm_page_dirty(m); } pte2_set_bit(pte2p, PTE2_NM); pte2_clear_bit(pte2p, PTE2_A); } else if ((pte2 & PTE2_A) != 0) pte2_clear_bit(pte2p, PTE2_A); else continue; pmap_tlb_flush(pmap, sva); } } if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { struct md_page *pvh; pv_entry_t next_pv, pv; pmap_t pmap; pt1_entry_t *pte1p, opte1; pt2_entry_t *pte2p, opte2; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("%s: page %p is exclusive busy", __func__, m)); /* * If the page is not PGA_WRITEABLE, then no PTE2s can have PTE2_NM * cleared. If the object containing the page is locked and the page * is not exclusive busied, then PGA_WRITEABLE cannot be concurrently * set. */ if ((m->flags & PGA_WRITEABLE) == 0) return; rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, va); opte1 = pte1_load(pte1p); if (!(opte1 & PTE1_RO)) { if (pmap_demote_pte1(pmap, pte1p, va) && !pte1_is_wired(opte1)) { /* * Write protect the mapping to a * single page so that a subsequent * write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - pte1_pa(opte1); pte2p = pmap_pte2_quick(pmap, va); opte2 = pte2_load(pte2p); if ((opte2 & PTE2_V)) { pte2_set_bit(pte2p, PTE2_NM | PTE2_RO); vm_page_dirty(m); pmap_tlb_flush(pmap, va); } } } PMAP_UNLOCK(pmap); } small_mappings: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte1p = pmap_pte1(pmap, pv->pv_va); KASSERT(!pte1_is_section(pte1_load(pte1p)), ("%s: found" " a section in page %p's pv list", __func__, m)); pte2p = pmap_pte2_quick(pmap, pv->pv_va); if (pte2_is_dirty(pte2_load(pte2p))) { pte2_set_bit(pte2p, PTE2_NM); pmap_tlb_flush(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } sched_unpin(); rw_wunlock(&pvh_global_lock); } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { struct sysmaps *sysmaps; vm_memattr_t oma; vm_paddr_t pa; oma = m->md.pat_mode; m->md.pat_mode = ma; CTR5(KTR_PMAP, "%s: page %p - 0x%08X oma: %d, ma: %d", __func__, m, VM_PAGE_TO_PHYS(m), oma, ma); if ((m->flags & PG_FICTITIOUS) != 0) return; #if 0 /* * If "m" is a normal page, flush it from the cache. * * First, try to find an existing mapping of the page by sf * buffer. sf_buf_invalidate_cache() modifies mapping and * flushes the cache. */ if (sf_buf_invalidate_cache(m, oma)) return; #endif /* * If page is not mapped by sf buffer, map the page * transient and do invalidation. */ if (ma != oma) { pa = VM_PAGE_TO_PHYS(m); sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP2) panic("%s: CMAP2 busy", __func__); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(pa, PTE2_AP_KRW, vm_memattr_to_pte2(ma))); dcache_wbinv_poc((vm_offset_t)sysmaps->CADDR2, pa, PAGE_SIZE); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); } } /* * Miscellaneous support routines follow */ /* * Returns TRUE if the given page is mapped individually or as part of * a 1mpage. Otherwise, returns FALSE. */ boolean_t pmap_page_is_mapped(vm_page_t m) { boolean_t rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (FALSE); rw_wlock(&pvh_global_lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list)); rw_wunlock(&pvh_global_lock); return (rv); } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("%s: page %p is not managed", __func__, m)); rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_wunlock(&pvh_global_lock); return (rv); } /* * pmap_zero_page zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. */ void pmap_zero_page(vm_page_t m) { struct sysmaps *sysmaps; sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (pte2_load(sysmaps->CMAP2) != 0) panic("%s: CMAP2 busy", __func__); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW, vm_page_pte2_attr(m))); pagezero(sysmaps->CADDR2); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); } /* * pmap_zero_page_area zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. * * off and size may not cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { struct sysmaps *sysmaps; sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (pte2_load(sysmaps->CMAP2) != 0) panic("%s: CMAP2 busy", __func__); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW, vm_page_pte2_attr(m))); if (off == 0 && size == PAGE_SIZE) pagezero(sysmaps->CADDR2); else bzero(sysmaps->CADDR2 + off, size); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); } /* * pmap_copy_page copies the specified (machine independent) * page by mapping the page into virtual memory and using * bcopy to copy the page, one machine dependent page at a * time. */ void pmap_copy_page(vm_page_t src, vm_page_t dst) { struct sysmaps *sysmaps; sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (pte2_load(sysmaps->CMAP1) != 0) panic("%s: CMAP1 busy", __func__); if (pte2_load(sysmaps->CMAP2) != 0) panic("%s: CMAP2 busy", __func__); pte2_store(sysmaps->CMAP1, PTE2_KERN_NG(VM_PAGE_TO_PHYS(src), PTE2_AP_KR | PTE2_NM, vm_page_pte2_attr(src))); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(VM_PAGE_TO_PHYS(dst), PTE2_AP_KRW, vm_page_pte2_attr(dst))); bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); pte2_clear(sysmaps->CMAP1); tlb_flush((vm_offset_t)sysmaps->CADDR1); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { struct sysmaps *sysmaps; vm_page_t a_pg, b_pg; char *a_cp, *b_cp; vm_offset_t a_pg_offset, b_pg_offset; int cnt; sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP1 != 0) panic("pmap_copy_pages: CMAP1 busy"); if (*sysmaps->CMAP2 != 0) panic("pmap_copy_pages: CMAP2 busy"); while (xfersize > 0) { a_pg = ma[a_offset >> PAGE_SHIFT]; a_pg_offset = a_offset & PAGE_MASK; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); b_pg = mb[b_offset >> PAGE_SHIFT]; b_pg_offset = b_offset & PAGE_MASK; cnt = min(cnt, PAGE_SIZE - b_pg_offset); pte2_store(sysmaps->CMAP1, PTE2_KERN_NG(VM_PAGE_TO_PHYS(a_pg), PTE2_AP_KR | PTE2_NM, vm_page_pte2_attr(a_pg))); tlb_flush_local((vm_offset_t)sysmaps->CADDR1); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(VM_PAGE_TO_PHYS(b_pg), PTE2_AP_KRW, vm_page_pte2_attr(b_pg))); tlb_flush_local((vm_offset_t)sysmaps->CADDR2); a_cp = sysmaps->CADDR1 + a_pg_offset; b_cp = sysmaps->CADDR2 + b_pg_offset; bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } pte2_clear(sysmaps->CMAP1); tlb_flush((vm_offset_t)sysmaps->CADDR1); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); } vm_offset_t pmap_quick_enter_page(vm_page_t m) { pt2_entry_t *pte2p; vm_offset_t qmap_addr; critical_enter(); qmap_addr = PCPU_GET(qmap_addr); pte2p = pt2map_entry(qmap_addr); KASSERT(pte2_load(pte2p) == 0, ("%s: PTE2 busy", __func__)); pte2_store(pte2p, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW, vm_page_pte2_attr(m))); return (qmap_addr); } void pmap_quick_remove_page(vm_offset_t addr) { pt2_entry_t *pte2p; vm_offset_t qmap_addr; qmap_addr = PCPU_GET(qmap_addr); pte2p = pt2map_entry(qmap_addr); KASSERT(addr == qmap_addr, ("%s: invalid address", __func__)); KASSERT(pte2_load(pte2p) != 0, ("%s: PTE2 not in use", __func__)); pte2_clear(pte2p); tlb_flush(qmap_addr); critical_exit(); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct spglist free; vm_offset_t addr; vm_offset_t end_addr = src_addr + len; vm_offset_t nextva; if (dst_addr != src_addr) return; if (!pmap_is_current(src_pmap)) return; rw_wlock(&pvh_global_lock); if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } sched_pin(); for (addr = src_addr; addr < end_addr; addr = nextva) { pt2_entry_t *src_pte2p, *dst_pte2p; vm_page_t dst_mpt2pg, src_mpt2pg; pt1_entry_t src_pte1; u_int pte1_idx; KASSERT(addr < VM_MAXUSER_ADDRESS, ("%s: invalid to pmap_copy page tables", __func__)); nextva = pte1_trunc(addr + PTE1_SIZE); if (nextva < addr) nextva = end_addr; pte1_idx = pte1_index(addr); src_pte1 = src_pmap->pm_pt1[pte1_idx]; if (pte1_is_section(src_pte1)) { if ((addr & PTE1_OFFSET) != 0 || (addr + PTE1_SIZE) > end_addr) continue; if (dst_pmap->pm_pt1[pte1_idx] == 0 && (!pte1_is_managed(src_pte1) || pmap_pv_insert_pte1(dst_pmap, addr, pte1_pa(src_pte1)))) { dst_pmap->pm_pt1[pte1_idx] = src_pte1 & ~PTE1_W; dst_pmap->pm_stats.resident_count += PTE1_SIZE / PAGE_SIZE; pmap_pte1_mappings++; } continue; } else if (!pte1_is_link(src_pte1)) continue; src_mpt2pg = PHYS_TO_VM_PAGE(pte1_link_pa(src_pte1)); /* * We leave PT2s to be linked from PT1 even if they are not * referenced until all PT2s in a page are without reference. * * QQQ: It could be changed ... */ #if 0 /* single_pt2_link_is_cleared */ KASSERT(pt2_wirecount_get(src_mpt2pg, pte1_idx) > 0, ("%s: source page table page is unused", __func__)); #else if (pt2_wirecount_get(src_mpt2pg, pte1_idx) == 0) continue; #endif if (nextva > end_addr) nextva = end_addr; src_pte2p = pt2map_entry(addr); while (addr < nextva) { pt2_entry_t temp_pte2; temp_pte2 = pte2_load(src_pte2p); /* * we only virtual copy managed pages */ if (pte2_is_managed(temp_pte2)) { dst_mpt2pg = pmap_allocpte2(dst_pmap, addr, PMAP_ENTER_NOSLEEP); if (dst_mpt2pg == NULL) goto out; dst_pte2p = pmap_pte2_quick(dst_pmap, addr); if (!pte2_is_valid(pte2_load(dst_pte2p)) && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(pte2_pa(temp_pte2)))) { /* * Clear the wired, modified, and * accessed (referenced) bits * during the copy. */ temp_pte2 &= ~(PTE2_W | PTE2_A); temp_pte2 |= PTE2_NM; pte2_store(dst_pte2p, temp_pte2); dst_pmap->pm_stats.resident_count++; } else { SLIST_INIT(&free); if (pmap_unwire_pt2(dst_pmap, addr, dst_mpt2pg, &free)) { pmap_tlb_flush(dst_pmap, addr); pmap_free_zero_pages(&free); } goto out; } if (pt2_wirecount_get(dst_mpt2pg, pte1_idx) >= pt2_wirecount_get(src_mpt2pg, pte1_idx)) break; } addr += PAGE_SIZE; src_pte2p++; } } out: sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more section mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t pte1_offset; if (size < PTE1_SIZE) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); pte1_offset = offset & PTE1_OFFSET; if (size - ((PTE1_SIZE - pte1_offset) & PTE1_OFFSET) < PTE1_SIZE || (*addr & PTE1_OFFSET) == pte1_offset) return; if ((*addr & PTE1_OFFSET) < pte1_offset) *addr = pte1_trunc(*addr) + pte1_offset; else *addr = pte1_roundup(*addr) + pte1_offset; } void pmap_activate(struct thread *td) { pmap_t pmap, oldpmap; u_int cpuid, ttb; PDEBUG(9, printf("%s: td = %08x\n", __func__, (uint32_t)td)); critical_enter(); pmap = vmspace_pmap(td->td_proc->p_vmspace); oldpmap = PCPU_GET(curpmap); cpuid = PCPU_GET(cpuid); #if defined(SMP) CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active); CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_CLR(cpuid, &oldpmap->pm_active); CPU_SET(cpuid, &pmap->pm_active); #endif ttb = pmap_ttb_get(pmap); /* * pmap_activate is for the current thread on the current cpu */ td->td_pcb->pcb_pagedir = ttb; cp15_ttbr_set(ttb); PCPU_SET(curpmap, pmap); critical_exit(); } /* * Perform the pmap work for mincore. */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa) { pt1_entry_t *pte1p, pte1; pt2_entry_t *pte2p, pte2; vm_paddr_t pa; boolean_t managed; int val; PMAP_LOCK(pmap); retry: pte1p = pmap_pte1(pmap, addr); pte1 = pte1_load(pte1p); if (pte1_is_section(pte1)) { pa = trunc_page(pte1_pa(pte1) | (addr & PTE1_OFFSET)); managed = pte1_is_managed(pte1); val = MINCORE_SUPER | MINCORE_INCORE; if (pte1_is_dirty(pte1)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if (pte1 & PTE1_A) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } else if (pte1_is_link(pte1)) { pte2p = pmap_pte2(pmap, addr); pte2 = pte2_load(pte2p); pmap_pte2_release(pte2p); pa = pte2_pa(pte2); managed = pte2_is_managed(pte2); val = MINCORE_INCORE; if (pte2_is_dirty(pte2)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if (pte2 & PTE2_A) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } else { managed = FALSE; val = 0; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && managed) { /* Ensure that "PHYS_TO_VM_PAGE(pa)->object" doesn't change. */ if (vm_page_pa_tryrelock(pmap, pa, locked_pa)) goto retry; } else PA_UNLOCK_COND(*locked_pa); PMAP_UNLOCK(pmap); return (val); } void pmap_kenter_device(vm_offset_t va, vm_size_t size, vm_paddr_t pa) { vm_offset_t sva; uint32_t l2attr; KASSERT((size & PAGE_MASK) == 0, ("%s: device mapping not page-sized", __func__)); sva = va; l2attr = vm_memattr_to_pte2(VM_MEMATTR_DEVICE); while (size != 0) { pmap_kenter_prot_attr(va, pa, PTE2_AP_KRW, l2attr); va += PAGE_SIZE; pa += PAGE_SIZE; size -= PAGE_SIZE; } tlb_flush_range(sva, va - sva); } void pmap_kremove_device(vm_offset_t va, vm_size_t size) { vm_offset_t sva; KASSERT((size & PAGE_MASK) == 0, ("%s: device mapping not page-sized", __func__)); sva = va; while (size != 0) { pmap_kremove(va); va += PAGE_SIZE; size -= PAGE_SIZE; } tlb_flush_range(sva, va - sva); } void pmap_set_pcb_pagedir(pmap_t pmap, struct pcb *pcb) { pcb->pcb_pagedir = pmap_ttb_get(pmap); } /* * Clean L1 data cache range by physical address. * The range must be within a single page. */ static void pmap_dcache_wb_pou(vm_paddr_t pa, vm_size_t size, uint32_t attr) { struct sysmaps *sysmaps; KASSERT(((pa & PAGE_MASK) + size) <= PAGE_SIZE, ("%s: not on single page", __func__)); sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP3) panic("%s: CMAP3 busy", __func__); pte2_store(sysmaps->CMAP3, PTE2_KERN_NG(pa, PTE2_AP_KRW, attr)); dcache_wb_pou((vm_offset_t)sysmaps->CADDR3 + (pa & PAGE_MASK), size); pte2_clear(sysmaps->CMAP3); tlb_flush((vm_offset_t)sysmaps->CADDR3); sched_unpin(); mtx_unlock(&sysmaps->lock); } /* * Sync instruction cache range which is not mapped yet. */ void cache_icache_sync_fresh(vm_offset_t va, vm_paddr_t pa, vm_size_t size) { uint32_t len, offset; vm_page_t m; /* Write back d-cache on given address range. */ offset = pa & PAGE_MASK; for ( ; size != 0; size -= len, pa += len, offset = 0) { len = min(PAGE_SIZE - offset, size); m = PHYS_TO_VM_PAGE(pa); KASSERT(m != NULL, ("%s: vm_page_t is null for %#x", __func__, pa)); pmap_dcache_wb_pou(pa, len, vm_page_pte2_attr(m)); } /* * I-cache is VIPT. Only way how to flush all virtual mappings * on given physical address is to invalidate all i-cache. */ icache_inv_all(); } void pmap_sync_icache(pmap_t pmap, vm_offset_t va, vm_size_t size) { /* Write back d-cache on given address range. */ if (va >= VM_MIN_KERNEL_ADDRESS) { dcache_wb_pou(va, size); } else { uint32_t len, offset; vm_paddr_t pa; vm_page_t m; offset = va & PAGE_MASK; for ( ; size != 0; size -= len, va += len, offset = 0) { pa = pmap_extract(pmap, va); /* offset is preserved */ len = min(PAGE_SIZE - offset, size); m = PHYS_TO_VM_PAGE(pa); KASSERT(m != NULL, ("%s: vm_page_t is null for %#x", __func__, pa)); pmap_dcache_wb_pou(pa, len, vm_page_pte2_attr(m)); } } /* * I-cache is VIPT. Only way how to flush all virtual mappings * on given physical address is to invalidate all i-cache. */ icache_inv_all(); } /* * The implementation of pmap_fault() uses IN_RANGE2() macro which * depends on the fact that given range size is a power of 2. */ CTASSERT(powerof2(NB_IN_PT1)); CTASSERT(powerof2(PT2MAP_SIZE)); #define IN_RANGE2(addr, start, size) \ ((vm_offset_t)(start) == ((vm_offset_t)(addr) & ~((size) - 1))) /* * Handle access and R/W emulation faults. */ int pmap_fault(pmap_t pmap, vm_offset_t far, uint32_t fsr, int idx, bool usermode) { pt1_entry_t *pte1p, pte1; pt2_entry_t *pte2p, pte2; if (pmap == NULL) pmap = kernel_pmap; /* * In kernel, we should never get abort with FAR which is in range of * pmap->pm_pt1 or PT2MAP address spaces. If it happens, stop here * and print out a useful abort message and even get to the debugger * otherwise it likely ends with never ending loop of aborts. */ if (__predict_false(IN_RANGE2(far, pmap->pm_pt1, NB_IN_PT1))) { /* * All L1 tables should always be mapped and present. * However, we check only current one herein. For user mode, * only permission abort from malicious user is not fatal. * And alignment abort as it may have higher priority. */ if (!usermode || (idx != FAULT_ALIGN && idx != FAULT_PERM_L2)) { CTR4(KTR_PMAP, "%s: pmap %#x pm_pt1 %#x far %#x", __func__, pmap, pmap->pm_pt1, far); panic("%s: pm_pt1 abort", __func__); } return (KERN_INVALID_ADDRESS); } if (__predict_false(IN_RANGE2(far, PT2MAP, PT2MAP_SIZE))) { /* * PT2MAP should be always mapped and present in current * L1 table. However, only existing L2 tables are mapped * in PT2MAP. For user mode, only L2 translation abort and * permission abort from malicious user is not fatal. * And alignment abort as it may have higher priority. */ if (!usermode || (idx != FAULT_ALIGN && idx != FAULT_TRAN_L2 && idx != FAULT_PERM_L2)) { CTR4(KTR_PMAP, "%s: pmap %#x PT2MAP %#x far %#x", __func__, pmap, PT2MAP, far); panic("%s: PT2MAP abort", __func__); } return (KERN_INVALID_ADDRESS); } /* * A pmap lock is used below for handling of access and R/W emulation * aborts. They were handled by atomic operations before so some * analysis of new situation is needed to answer the following question: * Is it safe to use the lock even for these aborts? * * There may happen two cases in general: * * (1) Aborts while the pmap lock is locked already - this should not * happen as pmap lock is not recursive. However, under pmap lock only * internal kernel data should be accessed and such data should be * mapped with A bit set and NM bit cleared. If double abort happens, * then a mapping of data which has caused it must be fixed. Further, * all new mappings are always made with A bit set and the bit can be * cleared only on managed mappings. * * (2) Aborts while another lock(s) is/are locked - this already can * happen. However, there is no difference here if it's either access or * R/W emulation abort, or if it's some other abort. */ PMAP_LOCK(pmap); #ifdef SMP /* * Special treatment is due to break-before-make approach done when * pte1 is updated for userland mapping during section promotion or * demotion. If not caught here, pmap_enter() can find a section * mapping on faulting address. That is not allowed. */ if (idx == FAULT_TRAN_L1 && usermode && cp15_ats1cur_check(far) == 0) { PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } #endif /* * Accesss bits for page and section. Note that the entry * is not in TLB yet, so TLB flush is not necessary. * * QQQ: This is hardware emulation, we do not call userret() * for aborts from user mode. */ if (idx == FAULT_ACCESS_L2) { pte2p = pt2map_entry(far); pte2 = pte2_load(pte2p); if (pte2_is_valid(pte2)) { pte2_store(pte2p, pte2 | PTE2_A); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } } if (idx == FAULT_ACCESS_L1) { pte1p = pmap_pte1(pmap, far); pte1 = pte1_load(pte1p); if (pte1_is_section(pte1)) { pte1_store(pte1p, pte1 | PTE1_A); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } } /* * Handle modify bits for page and section. Note that the modify * bit is emulated by software. So PTEx_RO is software read only * bit and PTEx_NM flag is real hardware read only bit. * * QQQ: This is hardware emulation, we do not call userret() * for aborts from user mode. */ if ((fsr & FSR_WNR) && (idx == FAULT_PERM_L2)) { pte2p = pt2map_entry(far); pte2 = pte2_load(pte2p); if (pte2_is_valid(pte2) && !(pte2 & PTE2_RO) && (pte2 & PTE2_NM)) { pte2_store(pte2p, pte2 & ~PTE2_NM); tlb_flush(trunc_page(far)); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } } if ((fsr & FSR_WNR) && (idx == FAULT_PERM_L1)) { pte1p = pmap_pte1(pmap, far); pte1 = pte1_load(pte1p); if (pte1_is_section(pte1) && !(pte1 & PTE1_RO) && (pte1 & PTE1_NM)) { pte1_store(pte1p, pte1 & ~PTE1_NM); tlb_flush(pte1_trunc(far)); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } } /* * QQQ: The previous code, mainly fast handling of access and * modify bits aborts, could be moved to ASM. Now we are * starting to deal with not fast aborts. */ #ifdef INVARIANTS /* * Read an entry in PT2TAB associated with both pmap and far. * It's safe because PT2TAB is always mapped. */ pte2 = pt2tab_load(pmap_pt2tab_entry(pmap, far)); if (pte2_is_valid(pte2)) { /* * Now, when we know that L2 page table is allocated, * we can use PT2MAP to get L2 page table entry. */ pte2 = pte2_load(pt2map_entry(far)); if (pte2_is_valid(pte2)) { /* * If L2 page table entry is valid, make sure that * L1 page table entry is valid too. Note that we * leave L2 page entries untouched when promoted. */ pte1 = pte1_load(pmap_pte1(pmap, far)); if (!pte1_is_valid(pte1)) { panic("%s: missing L1 page entry (%p, %#x)", __func__, pmap, far); } } } #endif PMAP_UNLOCK(pmap); return (KERN_FAILURE); } #if defined(PMAP_DEBUG) /* * Reusing of KVA used in pmap_zero_page function !!! */ static void pmap_zero_page_check(vm_page_t m) { uint32_t *p, *end; struct sysmaps *sysmaps; sched_pin(); sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (pte2_load(sysmaps->CMAP2) != 0) panic("%s: CMAP2 busy", __func__); pte2_store(sysmaps->CMAP2, PTE2_KERN_NG(VM_PAGE_TO_PHYS(m), PTE2_AP_KRW, vm_page_pte2_attr(m))); end = (uint32_t*)(sysmaps->CADDR2 + PAGE_SIZE); for (p = (uint32_t*)sysmaps->CADDR2; p < end; p++) if (*p != 0) panic("%s: page %p not zero, va: %p", __func__, m, sysmaps->CADDR2); pte2_clear(sysmaps->CMAP2); tlb_flush((vm_offset_t)sysmaps->CADDR2); sched_unpin(); mtx_unlock(&sysmaps->lock); } int pmap_pid_dump(int pid) { pmap_t pmap; struct proc *p; int npte2 = 0; int i, j, index; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { if (p->p_pid != pid || p->p_vmspace == NULL) continue; index = 0; pmap = vmspace_pmap(p->p_vmspace); for (i = 0; i < NPTE1_IN_PT1; i++) { pt1_entry_t pte1; pt2_entry_t *pte2p, pte2; vm_offset_t base, va; vm_paddr_t pa; vm_page_t m; base = i << PTE1_SHIFT; pte1 = pte1_load(&pmap->pm_pt1[i]); if (pte1_is_section(pte1)) { /* * QQQ: Do something here! */ } else if (pte1_is_link(pte1)) { for (j = 0; j < NPTE2_IN_PT2; j++) { va = base + (j << PAGE_SHIFT); if (va >= VM_MIN_KERNEL_ADDRESS) { if (index) { index = 0; printf("\n"); } sx_sunlock(&allproc_lock); return (npte2); } pte2p = pmap_pte2(pmap, va); pte2 = pte2_load(pte2p); pmap_pte2_release(pte2p); if (!pte2_is_valid(pte2)) continue; pa = pte2_pa(pte2); m = PHYS_TO_VM_PAGE(pa); printf("va: 0x%x, pa: 0x%x, h: %d, w:" " %d, f: 0x%x", va, pa, m->hold_count, m->wire_count, m->flags); npte2++; index++; if (index >= 2) { index = 0; printf("\n"); } else { printf(" "); } } } } } sx_sunlock(&allproc_lock); return (npte2); } #endif #ifdef DDB static pt2_entry_t * pmap_pte2_ddb(pmap_t pmap, vm_offset_t va) { pt1_entry_t pte1; vm_paddr_t pt2pg_pa; pte1 = pte1_load(pmap_pte1(pmap, va)); if (!pte1_is_link(pte1)) return (NULL); if (pmap_is_current(pmap)) return (pt2map_entry(va)); /* Note that L2 page table size is not equal to PAGE_SIZE. */ pt2pg_pa = trunc_page(pte1_link_pa(pte1)); if (pte2_pa(pte2_load(PMAP3)) != pt2pg_pa) { pte2_store(PMAP3, PTE2_KPT(pt2pg_pa)); #ifdef SMP PMAP3cpu = PCPU_GET(cpuid); #endif tlb_flush_local((vm_offset_t)PADDR3); } #ifdef SMP else if (PMAP3cpu != PCPU_GET(cpuid)) { PMAP3cpu = PCPU_GET(cpuid); tlb_flush_local((vm_offset_t)PADDR3); } #endif return (PADDR3 + (arm32_btop(va) & (NPTE2_IN_PG - 1))); } static void dump_pmap(pmap_t pmap) { printf("pmap %p\n", pmap); printf(" pm_pt1: %p\n", pmap->pm_pt1); printf(" pm_pt2tab: %p\n", pmap->pm_pt2tab); printf(" pm_active: 0x%08lX\n", pmap->pm_active.__bits[0]); } DB_SHOW_COMMAND(pmaps, pmap_list_pmaps) { pmap_t pmap; LIST_FOREACH(pmap, &allpmaps, pm_list) { dump_pmap(pmap); } } static int pte2_class(pt2_entry_t pte2) { int cls; cls = (pte2 >> 2) & 0x03; cls |= (pte2 >> 4) & 0x04; return (cls); } static void dump_section(pmap_t pmap, uint32_t pte1_idx) { } static void dump_link(pmap_t pmap, uint32_t pte1_idx, boolean_t invalid_ok) { uint32_t i; vm_offset_t va; pt2_entry_t *pte2p, pte2; vm_page_t m; va = pte1_idx << PTE1_SHIFT; pte2p = pmap_pte2_ddb(pmap, va); for (i = 0; i < NPTE2_IN_PT2; i++, pte2p++, va += PAGE_SIZE) { pte2 = pte2_load(pte2p); if (pte2 == 0) continue; if (!pte2_is_valid(pte2)) { printf(" 0x%08X: 0x%08X", va, pte2); if (!invalid_ok) printf(" - not valid !!!"); printf("\n"); continue; } m = PHYS_TO_VM_PAGE(pte2_pa(pte2)); printf(" 0x%08X: 0x%08X, TEX%d, s:%d, g:%d, m:%p", va , pte2, pte2_class(pte2), !!(pte2 & PTE2_S), !(pte2 & PTE2_NG), m); if (m != NULL) { printf(" v:%d h:%d w:%d f:0x%04X\n", m->valid, m->hold_count, m->wire_count, m->flags); } else { printf("\n"); } } } static __inline boolean_t is_pv_chunk_space(vm_offset_t va) { if ((((vm_offset_t)pv_chunkbase) <= va) && (va < ((vm_offset_t)pv_chunkbase + PAGE_SIZE * pv_maxchunks))) return (TRUE); return (FALSE); } DB_SHOW_COMMAND(pmap, pmap_pmap_print) { /* XXX convert args. */ pmap_t pmap = (pmap_t)addr; pt1_entry_t pte1; pt2_entry_t pte2; vm_offset_t va, eva; vm_page_t m; uint32_t i; boolean_t invalid_ok, dump_link_ok, dump_pv_chunk; if (have_addr) { pmap_t pm; LIST_FOREACH(pm, &allpmaps, pm_list) if (pm == pmap) break; if (pm == NULL) { printf("given pmap %p is not in allpmaps list\n", pmap); return; } } else pmap = PCPU_GET(curpmap); eva = (modif[0] == 'u') ? VM_MAXUSER_ADDRESS : 0xFFFFFFFF; dump_pv_chunk = FALSE; /* XXX evaluate from modif[] */ printf("pmap: 0x%08X\n", (uint32_t)pmap); printf("PT2MAP: 0x%08X\n", (uint32_t)PT2MAP); printf("pt2tab: 0x%08X\n", (uint32_t)pmap->pm_pt2tab); for(i = 0; i < NPTE1_IN_PT1; i++) { pte1 = pte1_load(&pmap->pm_pt1[i]); if (pte1 == 0) continue; va = i << PTE1_SHIFT; if (va >= eva) break; if (pte1_is_section(pte1)) { printf("0x%08X: Section 0x%08X, s:%d g:%d\n", va, pte1, !!(pte1 & PTE1_S), !(pte1 & PTE1_NG)); dump_section(pmap, i); } else if (pte1_is_link(pte1)) { dump_link_ok = TRUE; invalid_ok = FALSE; pte2 = pte2_load(pmap_pt2tab_entry(pmap, va)); m = PHYS_TO_VM_PAGE(pte1_link_pa(pte1)); printf("0x%08X: Link 0x%08X, pt2tab: 0x%08X m: %p", va, pte1, pte2, m); if (is_pv_chunk_space(va)) { printf(" - pv_chunk space"); if (dump_pv_chunk) invalid_ok = TRUE; else dump_link_ok = FALSE; } else if (m != NULL) printf(" w:%d w2:%u", m->wire_count, pt2_wirecount_get(m, pte1_index(va))); if (pte2 == 0) printf(" !!! pt2tab entry is ZERO"); else if (pte2_pa(pte1) != pte2_pa(pte2)) printf(" !!! pt2tab entry is DIFFERENT - m: %p", PHYS_TO_VM_PAGE(pte2_pa(pte2))); printf("\n"); if (dump_link_ok) dump_link(pmap, i, invalid_ok); } else printf("0x%08X: Invalid entry 0x%08X\n", va, pte1); } } static void dump_pt2tab(pmap_t pmap) { uint32_t i; pt2_entry_t pte2; vm_offset_t va; vm_paddr_t pa; vm_page_t m; printf("PT2TAB:\n"); for (i = 0; i < PT2TAB_ENTRIES; i++) { pte2 = pte2_load(&pmap->pm_pt2tab[i]); if (!pte2_is_valid(pte2)) continue; va = i << PT2TAB_SHIFT; pa = pte2_pa(pte2); m = PHYS_TO_VM_PAGE(pa); printf(" 0x%08X: 0x%08X, TEX%d, s:%d, m:%p", va, pte2, pte2_class(pte2), !!(pte2 & PTE2_S), m); if (m != NULL) printf(" , h: %d, w: %d, f: 0x%04X pidx: %lld", m->hold_count, m->wire_count, m->flags, m->pindex); printf("\n"); } } DB_SHOW_COMMAND(pmap_pt2tab, pmap_pt2tab_print) { /* XXX convert args. */ pmap_t pmap = (pmap_t)addr; pt1_entry_t pte1; pt2_entry_t pte2; vm_offset_t va; uint32_t i, start; if (have_addr) { printf("supported only on current pmap\n"); return; } pmap = PCPU_GET(curpmap); printf("curpmap: 0x%08X\n", (uint32_t)pmap); printf("PT2MAP: 0x%08X\n", (uint32_t)PT2MAP); printf("pt2tab: 0x%08X\n", (uint32_t)pmap->pm_pt2tab); start = pte1_index((vm_offset_t)PT2MAP); for (i = start; i < (start + NPT2_IN_PT2TAB); i++) { pte1 = pte1_load(&pmap->pm_pt1[i]); if (pte1 == 0) continue; va = i << PTE1_SHIFT; if (pte1_is_section(pte1)) { printf("0x%08X: Section 0x%08X, s:%d\n", va, pte1, !!(pte1 & PTE1_S)); dump_section(pmap, i); } else if (pte1_is_link(pte1)) { pte2 = pte2_load(pmap_pt2tab_entry(pmap, va)); printf("0x%08X: Link 0x%08X, pt2tab: 0x%08X\n", va, pte1, pte2); if (pte2 == 0) printf(" !!! pt2tab entry is ZERO\n"); } else printf("0x%08X: Invalid entry 0x%08X\n", va, pte1); } dump_pt2tab(pmap); } #endif Index: projects/mips64-clang/sys/arm64/arm64/pmap.c =================================================================== --- projects/mips64-clang/sys/arm64/arm64/pmap.c (revision 305711) +++ projects/mips64-clang/sys/arm64/arm64/pmap.c (revision 305712) @@ -1,4732 +1,4744 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2003 Peter Wemm * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * Copyright (c) 2014 Andrew Turner * All rights reserved. * Copyright (c) 2014-2016 The FreeBSD Foundation * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * This software was developed by Andrew Turner under sponsorship from * the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define NL0PG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NL1PG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NL2PG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NL3PG (PAGE_SIZE/(sizeof (pt_entry_t))) #define NUL0E L0_ENTRIES #define NUL1E (NUL0E * NL1PG) #define NUL2E (NUL1E * NL2PG) #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif /* * These are configured by the mair_el1 register. This is set up in locore.S */ #define DEVICE_MEMORY 0 #define UNCACHED_MEMORY 1 #define CACHED_MEMORY 2 #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #define pmap_l2_pindex(v) ((v) >> L2_SHIFT) #define pa_to_pvh(pa) (&pv_table[pmap_l2_pindex(pa)]) #define NPV_LIST_LOCKS MAXCPU #define PHYS_TO_PV_LIST_LOCK(pa) \ (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS]) #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \ struct rwlock **_lockp = (lockp); \ struct rwlock *_new_lock; \ \ _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \ if (_new_lock != *_lockp) { \ if (*_lockp != NULL) \ rw_wunlock(*_lockp); \ *_lockp = _new_lock; \ rw_wlock(*_lockp); \ } \ } while (0) #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \ CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m)) #define RELEASE_PV_LIST_LOCK(lockp) do { \ struct rwlock **_lockp = (lockp); \ \ if (*_lockp != NULL) { \ rw_wunlock(*_lockp); \ *_lockp = NULL; \ } \ } while (0) #define VM_PAGE_TO_PV_LIST_LOCK(m) \ PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m)) struct pmap kernel_pmap_store; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ vm_offset_t kernel_vm_end = 0; struct msgbuf *msgbufp = NULL; /* * Data for the pv entry allocation mechanism. * Updates to pv_invl_gen are protected by the pv_list_locks[] * elements, but reads are not. */ static struct md_page *pv_table; static struct md_page pv_dummy; vm_paddr_t dmap_phys_base; /* The start of the dmap region */ vm_paddr_t dmap_phys_max; /* The limit of the dmap region */ vm_offset_t dmap_max_addr; /* The virtual address limit of the dmap */ /* This code assumes all L1 DMAP entries will be used */ CTASSERT((DMAP_MIN_ADDRESS & ~L0_OFFSET) == DMAP_MIN_ADDRESS); CTASSERT((DMAP_MAX_ADDRESS & ~L0_OFFSET) == DMAP_MAX_ADDRESS); #define DMAP_TABLES ((DMAP_MAX_ADDRESS - DMAP_MIN_ADDRESS) >> L0_SHIFT) extern pt_entry_t pagetable_dmap[]; static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); static int superpages_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, superpages_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &superpages_enabled, 0, "Are large page mappings enabled?"); /* * Data for the pv entry allocation mechanism */ static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks); static struct mtx pv_chunks_mutex; static struct rwlock pv_list_locks[NPV_LIST_LOCKS]; static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp); static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp); static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode); static int pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode); static pt_entry_t *pmap_demote_l1(pmap_t pmap, pt_entry_t *l1, vm_offset_t va); static pt_entry_t *pmap_demote_l2_locked(pmap_t pmap, pt_entry_t *l2, vm_offset_t va, struct rwlock **lockp); static pt_entry_t *pmap_demote_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t va); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static int pmap_remove_l3(pmap_t pmap, pt_entry_t *l3, vm_offset_t sva, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp); static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp); static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free); static int pmap_unuse_l3(pmap_t, vm_offset_t, pd_entry_t, struct spglist *); /* * These load the old table data and store the new value. * They need to be atomic as the System MMU may write to the table at * the same time as the CPU. */ #define pmap_load_store(table, entry) atomic_swap_64(table, entry) #define pmap_set(table, mask) atomic_set_64(table, mask) #define pmap_load_clear(table) atomic_swap_64(table, 0) #define pmap_load(table) (*table) /********************/ /* Inline functions */ /********************/ static __inline void pagecopy(void *s, void *d) { memcpy(d, s, PAGE_SIZE); } #define pmap_l0_index(va) (((va) >> L0_SHIFT) & L0_ADDR_MASK) #define pmap_l1_index(va) (((va) >> L1_SHIFT) & Ln_ADDR_MASK) #define pmap_l2_index(va) (((va) >> L2_SHIFT) & Ln_ADDR_MASK) #define pmap_l3_index(va) (((va) >> L3_SHIFT) & Ln_ADDR_MASK) static __inline pd_entry_t * pmap_l0(pmap_t pmap, vm_offset_t va) { return (&pmap->pm_l0[pmap_l0_index(va)]); } static __inline pd_entry_t * pmap_l0_to_l1(pd_entry_t *l0, vm_offset_t va) { pd_entry_t *l1; l1 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l0) & ~ATTR_MASK); return (&l1[pmap_l1_index(va)]); } static __inline pd_entry_t * pmap_l1(pmap_t pmap, vm_offset_t va) { pd_entry_t *l0; l0 = pmap_l0(pmap, va); if ((pmap_load(l0) & ATTR_DESCR_MASK) != L0_TABLE) return (NULL); return (pmap_l0_to_l1(l0, va)); } static __inline pd_entry_t * pmap_l1_to_l2(pd_entry_t *l1, vm_offset_t va) { pd_entry_t *l2; l2 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l1) & ~ATTR_MASK); return (&l2[pmap_l2_index(va)]); } static __inline pd_entry_t * pmap_l2(pmap_t pmap, vm_offset_t va) { pd_entry_t *l1; l1 = pmap_l1(pmap, va); if ((pmap_load(l1) & ATTR_DESCR_MASK) != L1_TABLE) return (NULL); return (pmap_l1_to_l2(l1, va)); } static __inline pt_entry_t * pmap_l2_to_l3(pd_entry_t *l2, vm_offset_t va) { pt_entry_t *l3; l3 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l2) & ~ATTR_MASK); return (&l3[pmap_l3_index(va)]); } /* * Returns the lowest valid pde for a given virtual address. * The next level may or may not point to a valid page or block. */ static __inline pd_entry_t * pmap_pde(pmap_t pmap, vm_offset_t va, int *level) { pd_entry_t *l0, *l1, *l2, desc; l0 = pmap_l0(pmap, va); desc = pmap_load(l0) & ATTR_DESCR_MASK; if (desc != L0_TABLE) { *level = -1; return (NULL); } l1 = pmap_l0_to_l1(l0, va); desc = pmap_load(l1) & ATTR_DESCR_MASK; if (desc != L1_TABLE) { *level = 0; return (l0); } l2 = pmap_l1_to_l2(l1, va); desc = pmap_load(l2) & ATTR_DESCR_MASK; if (desc != L2_TABLE) { *level = 1; return (l1); } *level = 2; return (l2); } /* * Returns the lowest valid pte block or table entry for a given virtual * address. If there are no valid entries return NULL and set the level to * the first invalid level. */ static __inline pt_entry_t * pmap_pte(pmap_t pmap, vm_offset_t va, int *level) { pd_entry_t *l1, *l2, desc; pt_entry_t *l3; l1 = pmap_l1(pmap, va); if (l1 == NULL) { *level = 0; return (NULL); } desc = pmap_load(l1) & ATTR_DESCR_MASK; if (desc == L1_BLOCK) { *level = 1; return (l1); } if (desc != L1_TABLE) { *level = 1; return (NULL); } l2 = pmap_l1_to_l2(l1, va); desc = pmap_load(l2) & ATTR_DESCR_MASK; if (desc == L2_BLOCK) { *level = 2; return (l2); } if (desc != L2_TABLE) { *level = 2; return (NULL); } *level = 3; l3 = pmap_l2_to_l3(l2, va); if ((pmap_load(l3) & ATTR_DESCR_MASK) != L3_PAGE) return (NULL); return (l3); } static inline bool pmap_superpages_enabled(void) { return (superpages_enabled != 0); } bool pmap_get_tables(pmap_t pmap, vm_offset_t va, pd_entry_t **l0, pd_entry_t **l1, pd_entry_t **l2, pt_entry_t **l3) { pd_entry_t *l0p, *l1p, *l2p; if (pmap->pm_l0 == NULL) return (false); l0p = pmap_l0(pmap, va); *l0 = l0p; if ((pmap_load(l0p) & ATTR_DESCR_MASK) != L0_TABLE) return (false); l1p = pmap_l0_to_l1(l0p, va); *l1 = l1p; if ((pmap_load(l1p) & ATTR_DESCR_MASK) == L1_BLOCK) { *l2 = NULL; *l3 = NULL; return (true); } if ((pmap_load(l1p) & ATTR_DESCR_MASK) != L1_TABLE) return (false); l2p = pmap_l1_to_l2(l1p, va); *l2 = l2p; if ((pmap_load(l2p) & ATTR_DESCR_MASK) == L2_BLOCK) { *l3 = NULL; return (true); } *l3 = pmap_l2_to_l3(l2p, va); return (true); } static __inline int pmap_is_current(pmap_t pmap) { return ((pmap == pmap_kernel()) || (pmap == curthread->td_proc->p_vmspace->vm_map.pmap)); } static __inline int pmap_l3_valid(pt_entry_t l3) { return ((l3 & ATTR_DESCR_MASK) == L3_PAGE); } /* Is a level 1 or 2entry a valid block and cacheable */ CTASSERT(L1_BLOCK == L2_BLOCK); static __inline int pmap_pte_valid_cacheable(pt_entry_t pte) { return (((pte & ATTR_DESCR_MASK) == L1_BLOCK) && ((pte & ATTR_IDX_MASK) == ATTR_IDX(CACHED_MEMORY))); } static __inline int pmap_l3_valid_cacheable(pt_entry_t l3) { return (((l3 & ATTR_DESCR_MASK) == L3_PAGE) && ((l3 & ATTR_IDX_MASK) == ATTR_IDX(CACHED_MEMORY))); } #define PTE_SYNC(pte) cpu_dcache_wb_range((vm_offset_t)pte, sizeof(*pte)) /* * Checks if the page is dirty. We currently lack proper tracking of this on * arm64 so for now assume is a page mapped as rw was accessed it is. */ static inline int pmap_page_dirty(pt_entry_t pte) { return ((pte & (ATTR_AF | ATTR_AP_RW_BIT)) == (ATTR_AF | ATTR_AP(ATTR_AP_RW))); } static __inline void pmap_resident_count_inc(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); pmap->pm_stats.resident_count += count; } static __inline void pmap_resident_count_dec(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(pmap->pm_stats.resident_count >= count, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count -= count; } static pt_entry_t * pmap_early_page_idx(vm_offset_t l1pt, vm_offset_t va, u_int *l1_slot, u_int *l2_slot) { pt_entry_t *l2; pd_entry_t *l1; l1 = (pd_entry_t *)l1pt; *l1_slot = (va >> L1_SHIFT) & Ln_ADDR_MASK; /* Check locore has used a table L1 map */ KASSERT((l1[*l1_slot] & ATTR_DESCR_MASK) == L1_TABLE, ("Invalid bootstrap L1 table")); /* Find the address of the L2 table */ l2 = (pt_entry_t *)init_pt_va; *l2_slot = pmap_l2_index(va); return (l2); } static vm_paddr_t pmap_early_vtophys(vm_offset_t l1pt, vm_offset_t va) { u_int l1_slot, l2_slot; pt_entry_t *l2; l2 = pmap_early_page_idx(l1pt, va, &l1_slot, &l2_slot); return ((l2[l2_slot] & ~ATTR_MASK) + (va & L2_OFFSET)); } static void pmap_bootstrap_dmap(vm_offset_t kern_l1, vm_paddr_t min_pa, vm_paddr_t max_pa) { vm_offset_t va; vm_paddr_t pa; u_int l1_slot; pa = dmap_phys_base = min_pa & ~L1_OFFSET; va = DMAP_MIN_ADDRESS; for (; va < DMAP_MAX_ADDRESS && pa < max_pa; pa += L1_SIZE, va += L1_SIZE, l1_slot++) { l1_slot = ((va - DMAP_MIN_ADDRESS) >> L1_SHIFT); pmap_load_store(&pagetable_dmap[l1_slot], (pa & ~L1_OFFSET) | ATTR_DEFAULT | ATTR_IDX(CACHED_MEMORY) | L1_BLOCK); } /* Set the upper limit of the DMAP region */ dmap_phys_max = pa; dmap_max_addr = va; cpu_dcache_wb_range((vm_offset_t)pagetable_dmap, PAGE_SIZE * DMAP_TABLES); cpu_tlb_flushID(); } static vm_offset_t pmap_bootstrap_l2(vm_offset_t l1pt, vm_offset_t va, vm_offset_t l2_start) { vm_offset_t l2pt; vm_paddr_t pa; pd_entry_t *l1; u_int l1_slot; KASSERT((va & L1_OFFSET) == 0, ("Invalid virtual address")); l1 = (pd_entry_t *)l1pt; l1_slot = pmap_l1_index(va); l2pt = l2_start; for (; va < VM_MAX_KERNEL_ADDRESS; l1_slot++, va += L1_SIZE) { KASSERT(l1_slot < Ln_ENTRIES, ("Invalid L1 index")); pa = pmap_early_vtophys(l1pt, l2pt); pmap_load_store(&l1[l1_slot], (pa & ~Ln_TABLE_MASK) | L1_TABLE); l2pt += PAGE_SIZE; } /* Clean the L2 page table */ memset((void *)l2_start, 0, l2pt - l2_start); cpu_dcache_wb_range(l2_start, l2pt - l2_start); /* Flush the l1 table to ram */ cpu_dcache_wb_range((vm_offset_t)l1, PAGE_SIZE); return l2pt; } static vm_offset_t pmap_bootstrap_l3(vm_offset_t l1pt, vm_offset_t va, vm_offset_t l3_start) { vm_offset_t l2pt, l3pt; vm_paddr_t pa; pd_entry_t *l2; u_int l2_slot; KASSERT((va & L2_OFFSET) == 0, ("Invalid virtual address")); l2 = pmap_l2(kernel_pmap, va); l2 = (pd_entry_t *)rounddown2((uintptr_t)l2, PAGE_SIZE); l2pt = (vm_offset_t)l2; l2_slot = pmap_l2_index(va); l3pt = l3_start; for (; va < VM_MAX_KERNEL_ADDRESS; l2_slot++, va += L2_SIZE) { KASSERT(l2_slot < Ln_ENTRIES, ("Invalid L2 index")); pa = pmap_early_vtophys(l1pt, l3pt); pmap_load_store(&l2[l2_slot], (pa & ~Ln_TABLE_MASK) | L2_TABLE); l3pt += PAGE_SIZE; } /* Clean the L2 page table */ memset((void *)l3_start, 0, l3pt - l3_start); cpu_dcache_wb_range(l3_start, l3pt - l3_start); cpu_dcache_wb_range((vm_offset_t)l2, PAGE_SIZE); return l3pt; } /* * Bootstrap the system enough to run with virtual memory. */ void pmap_bootstrap(vm_offset_t l0pt, vm_offset_t l1pt, vm_paddr_t kernstart, vm_size_t kernlen) { u_int l1_slot, l2_slot, avail_slot, map_slot, used_map_slot; uint64_t kern_delta; pt_entry_t *l2; vm_offset_t va, freemempos; vm_offset_t dpcpu, msgbufpv; vm_paddr_t pa, max_pa, min_pa; int i; kern_delta = KERNBASE - kernstart; physmem = 0; printf("pmap_bootstrap %lx %lx %lx\n", l1pt, kernstart, kernlen); printf("%lx\n", l1pt); printf("%lx\n", (KERNBASE >> L1_SHIFT) & Ln_ADDR_MASK); /* Set this early so we can use the pagetable walking functions */ kernel_pmap_store.pm_l0 = (pd_entry_t *)l0pt; PMAP_LOCK_INIT(kernel_pmap); /* Assume the address we were loaded to is a valid physical address */ min_pa = max_pa = KERNBASE - kern_delta; /* * Find the minimum physical address. physmap is sorted, * but may contain empty ranges. */ for (i = 0; i < (physmap_idx * 2); i += 2) { if (physmap[i] == physmap[i + 1]) continue; if (physmap[i] <= min_pa) min_pa = physmap[i]; if (physmap[i + 1] > max_pa) max_pa = physmap[i + 1]; } /* Create a direct map region early so we can use it for pa -> va */ pmap_bootstrap_dmap(l1pt, min_pa, max_pa); va = KERNBASE; pa = KERNBASE - kern_delta; /* * Start to initialise phys_avail by copying from physmap * up to the physical address KERNBASE points at. */ map_slot = avail_slot = 0; for (; map_slot < (physmap_idx * 2) && avail_slot < (PHYS_AVAIL_SIZE - 2); map_slot += 2) { if (physmap[map_slot] == physmap[map_slot + 1]) continue; if (physmap[map_slot] <= pa && physmap[map_slot + 1] > pa) break; phys_avail[avail_slot] = physmap[map_slot]; phys_avail[avail_slot + 1] = physmap[map_slot + 1]; physmem += (phys_avail[avail_slot + 1] - phys_avail[avail_slot]) >> PAGE_SHIFT; avail_slot += 2; } /* Add the memory before the kernel */ if (physmap[avail_slot] < pa && avail_slot < (PHYS_AVAIL_SIZE - 2)) { phys_avail[avail_slot] = physmap[map_slot]; phys_avail[avail_slot + 1] = pa; physmem += (phys_avail[avail_slot + 1] - phys_avail[avail_slot]) >> PAGE_SHIFT; avail_slot += 2; } used_map_slot = map_slot; /* * Read the page table to find out what is already mapped. * This assumes we have mapped a block of memory from KERNBASE * using a single L1 entry. */ l2 = pmap_early_page_idx(l1pt, KERNBASE, &l1_slot, &l2_slot); /* Sanity check the index, KERNBASE should be the first VA */ KASSERT(l2_slot == 0, ("The L2 index is non-zero")); /* Find how many pages we have mapped */ for (; l2_slot < Ln_ENTRIES; l2_slot++) { if ((l2[l2_slot] & ATTR_DESCR_MASK) == 0) break; /* Check locore used L2 blocks */ KASSERT((l2[l2_slot] & ATTR_DESCR_MASK) == L2_BLOCK, ("Invalid bootstrap L2 table")); KASSERT((l2[l2_slot] & ~ATTR_MASK) == pa, ("Incorrect PA in L2 table")); va += L2_SIZE; pa += L2_SIZE; } va = roundup2(va, L1_SIZE); freemempos = KERNBASE + kernlen; freemempos = roundup2(freemempos, PAGE_SIZE); /* Create the l2 tables up to VM_MAX_KERNEL_ADDRESS */ freemempos = pmap_bootstrap_l2(l1pt, va, freemempos); /* And the l3 tables for the early devmap */ freemempos = pmap_bootstrap_l3(l1pt, VM_MAX_KERNEL_ADDRESS - L2_SIZE, freemempos); cpu_tlb_flushID(); #define alloc_pages(var, np) \ (var) = freemempos; \ freemempos += (np * PAGE_SIZE); \ memset((char *)(var), 0, ((np) * PAGE_SIZE)); /* Allocate dynamic per-cpu area. */ alloc_pages(dpcpu, DPCPU_SIZE / PAGE_SIZE); dpcpu_init((void *)dpcpu, 0); /* Allocate memory for the msgbuf, e.g. for /sbin/dmesg */ alloc_pages(msgbufpv, round_page(msgbufsize) / PAGE_SIZE); msgbufp = (void *)msgbufpv; virtual_avail = roundup2(freemempos, L1_SIZE); virtual_end = VM_MAX_KERNEL_ADDRESS - L2_SIZE; kernel_vm_end = virtual_avail; pa = pmap_early_vtophys(l1pt, freemempos); /* Finish initialising physmap */ map_slot = used_map_slot; for (; avail_slot < (PHYS_AVAIL_SIZE - 2) && map_slot < (physmap_idx * 2); map_slot += 2) { if (physmap[map_slot] == physmap[map_slot + 1]) continue; /* Have we used the current range? */ if (physmap[map_slot + 1] <= pa) continue; /* Do we need to split the entry? */ if (physmap[map_slot] < pa) { phys_avail[avail_slot] = pa; phys_avail[avail_slot + 1] = physmap[map_slot + 1]; } else { phys_avail[avail_slot] = physmap[map_slot]; phys_avail[avail_slot + 1] = physmap[map_slot + 1]; } physmem += (phys_avail[avail_slot + 1] - phys_avail[avail_slot]) >> PAGE_SHIFT; avail_slot += 2; } phys_avail[avail_slot] = 0; phys_avail[avail_slot + 1] = 0; /* * Maxmem isn't the "maximum memory", it's one larger than the * highest page of the physical address space. It should be * called something like "Maxphyspage". */ Maxmem = atop(phys_avail[avail_slot - 1]); cpu_tlb_flushID(); } /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pv_memattr = VM_MEMATTR_WRITE_BACK; } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { vm_size_t s; int i, pv_npg; /* * Are large page mappings enabled? */ TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled); /* * Initialize the pv chunk list mutex. */ mtx_init(&pv_chunks_mutex, "pmap pv chunk list", NULL, MTX_DEF); /* * Initialize the pool of pv list locks. */ for (i = 0; i < NPV_LIST_LOCKS; i++) rw_init(&pv_list_locks[i], "pmap pv list"); /* * Calculate the size of the pv head table for superpages. */ pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, L2_SIZE); /* * Allocate memory for the pv head table for superpages. */ s = (vm_size_t)(pv_npg * sizeof(struct md_page)); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(kernel_arena, s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); TAILQ_INIT(&pv_dummy.pv_list); } static SYSCTL_NODE(_vm_pmap, OID_AUTO, l2, CTLFLAG_RD, 0, "2MB page mapping counters"); static u_long pmap_l2_demotions; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, demotions, CTLFLAG_RD, &pmap_l2_demotions, 0, "2MB page demotions"); static u_long pmap_l2_p_failures; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_l2_p_failures, 0, "2MB page promotion failures"); static u_long pmap_l2_promotions; SYSCTL_ULONG(_vm_pmap_l2, OID_AUTO, promotions, CTLFLAG_RD, &pmap_l2_promotions, 0, "2MB page promotions"); /* * Invalidate a single TLB entry. */ PMAP_INLINE void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { sched_pin(); __asm __volatile( "dsb ishst \n" "tlbi vaae1is, %0 \n" "dsb ish \n" "isb \n" : : "r"(va >> PAGE_SHIFT)); sched_unpin(); } PMAP_INLINE void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t addr; sched_pin(); dsb(ishst); for (addr = sva; addr < eva; addr += PAGE_SIZE) { __asm __volatile( "tlbi vaae1is, %0" : : "r"(addr >> PAGE_SHIFT)); } __asm __volatile( "dsb ish \n" "isb \n"); sched_unpin(); } PMAP_INLINE void pmap_invalidate_all(pmap_t pmap) { sched_pin(); __asm __volatile( "dsb ishst \n" "tlbi vmalle1is \n" "dsb ish \n" "isb \n"); sched_unpin(); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { pt_entry_t *pte, tpte; vm_paddr_t pa; int lvl; pa = 0; PMAP_LOCK(pmap); /* * Find the block or page map for this virtual address. pmap_pte * will return either a valid block/page entry, or NULL. */ pte = pmap_pte(pmap, va, &lvl); if (pte != NULL) { tpte = pmap_load(pte); pa = tpte & ~ATTR_MASK; switch(lvl) { case 1: KASSERT((tpte & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_extract: Invalid L1 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L1_OFFSET); break; case 2: KASSERT((tpte & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_extract: Invalid L2 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L2_OFFSET); break; case 3: KASSERT((tpte & ATTR_DESCR_MASK) == L3_PAGE, ("pmap_extract: Invalid L3 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L3_OFFSET); break; } } PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pt_entry_t *pte, tpte; vm_offset_t off; vm_paddr_t pa; vm_page_t m; int lvl; pa = 0; m = NULL; PMAP_LOCK(pmap); retry: pte = pmap_pte(pmap, va, &lvl); if (pte != NULL) { tpte = pmap_load(pte); KASSERT(lvl > 0 && lvl <= 3, ("pmap_extract_and_hold: Invalid level %d", lvl)); CTASSERT(L1_BLOCK == L2_BLOCK); KASSERT((lvl == 3 && (tpte & ATTR_DESCR_MASK) == L3_PAGE) || (lvl < 3 && (tpte & ATTR_DESCR_MASK) == L1_BLOCK), ("pmap_extract_and_hold: Invalid pte at L%d: %lx", lvl, tpte & ATTR_DESCR_MASK)); if (((tpte & ATTR_AP_RW_BIT) == ATTR_AP(ATTR_AP_RW)) || ((prot & VM_PROT_WRITE) == 0)) { switch(lvl) { case 1: off = va & L1_OFFSET; break; case 2: off = va & L2_OFFSET; break; case 3: default: off = 0; } if (vm_page_pa_tryrelock(pmap, (tpte & ~ATTR_MASK) | off, &pa)) goto retry; m = PHYS_TO_VM_PAGE((tpte & ~ATTR_MASK) | off); vm_page_hold(m); } } PA_UNLOCK_COND(pa); PMAP_UNLOCK(pmap); return (m); } vm_paddr_t pmap_kextract(vm_offset_t va) { pt_entry_t *pte, tpte; vm_paddr_t pa; int lvl; if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { pa = DMAP_TO_PHYS(va); } else { pa = 0; pte = pmap_pte(kernel_pmap, va, &lvl); if (pte != NULL) { tpte = pmap_load(pte); pa = tpte & ~ATTR_MASK; switch(lvl) { case 1: KASSERT((tpte & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_kextract: Invalid L1 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L1_OFFSET); break; case 2: KASSERT((tpte & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_kextract: Invalid L2 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L2_OFFSET); break; case 3: KASSERT((tpte & ATTR_DESCR_MASK) == L3_PAGE, ("pmap_kextract: Invalid L3 pte found: %lx", tpte & ATTR_DESCR_MASK)); pa |= (va & L3_OFFSET); break; } } } return (pa); } /*************************************************** * Low level mapping routines..... ***************************************************/ static void pmap_kenter(vm_offset_t sva, vm_size_t size, vm_paddr_t pa, int mode) { pd_entry_t *pde; pt_entry_t *pte; vm_offset_t va; int lvl; KASSERT((pa & L3_OFFSET) == 0, ("pmap_kenter: Invalid physical address")); KASSERT((sva & L3_OFFSET) == 0, ("pmap_kenter: Invalid virtual address")); KASSERT((size & PAGE_MASK) == 0, ("pmap_kenter: Mapping is not page-sized")); va = sva; while (size != 0) { pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_kenter: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_kenter: Invalid level %d", lvl)); pte = pmap_l2_to_l3(pde, va); pmap_load_store(pte, (pa & ~L3_OFFSET) | ATTR_DEFAULT | ATTR_IDX(mode) | L3_PAGE); PTE_SYNC(pte); va += PAGE_SIZE; pa += PAGE_SIZE; size -= PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } void pmap_kenter_device(vm_offset_t sva, vm_size_t size, vm_paddr_t pa) { pmap_kenter(sva, size, pa, DEVICE_MEMORY); } /* * Remove a page from the kernel pagetables. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *pte; int lvl; pte = pmap_pte(kernel_pmap, va, &lvl); KASSERT(pte != NULL, ("pmap_kremove: Invalid address")); KASSERT(lvl == 3, ("pmap_kremove: Invalid pte level %d", lvl)); if (pmap_l3_valid_cacheable(pmap_load(pte))) cpu_dcache_wb_range(va, L3_SIZE); pmap_load_clear(pte); PTE_SYNC(pte); pmap_invalidate_page(kernel_pmap, va); } void pmap_kremove_device(vm_offset_t sva, vm_size_t size) { pt_entry_t *pte; vm_offset_t va; int lvl; KASSERT((sva & L3_OFFSET) == 0, ("pmap_kremove_device: Invalid virtual address")); KASSERT((size & PAGE_MASK) == 0, ("pmap_kremove_device: Mapping is not page-sized")); va = sva; while (size != 0) { pte = pmap_pte(kernel_pmap, va, &lvl); KASSERT(pte != NULL, ("Invalid page table, va: 0x%lx", va)); KASSERT(lvl == 3, ("Invalid device pagetable level: %d != 3", lvl)); pmap_load_clear(pte); PTE_SYNC(pte); va += PAGE_SIZE; size -= PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return PHYS_TO_DMAP(start); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pd_entry_t *pde; pt_entry_t *pte, pa; vm_offset_t va; vm_page_t m; int i, lvl; va = sva; for (i = 0; i < count; i++) { pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_qenter: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_qenter: Invalid level %d", lvl)); m = ma[i]; pa = VM_PAGE_TO_PHYS(m) | ATTR_DEFAULT | ATTR_AP(ATTR_AP_RW) | ATTR_IDX(m->md.pv_memattr) | L3_PAGE; pte = pmap_l2_to_l3(pde, va); pmap_load_store(pte, pa); PTE_SYNC(pte); va += L3_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. */ void pmap_qremove(vm_offset_t sva, int count) { pt_entry_t *pte; vm_offset_t va; int lvl; KASSERT(sva >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", sva)); va = sva; while (count-- > 0) { pte = pmap_pte(kernel_pmap, va, &lvl); KASSERT(lvl == 3, ("Invalid device pagetable level: %d != 3", lvl)); if (pte != NULL) { if (pmap_l3_valid_cacheable(pmap_load(pte))) cpu_dcache_wb_range(va, L3_SIZE); pmap_load_clear(pte); PTE_SYNC(pte); } va += PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ static __inline void pmap_free_zero_pages(struct spglist *free) { vm_page_t m; while ((m = SLIST_FIRST(free)) != NULL) { SLIST_REMOVE_HEAD(free, plinks.s.ss); /* Preserve the page's PG_ZERO setting. */ vm_page_free_toq(m); } } /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Decrements a page table page's wire count, which is used to record the * number of valid page table entries within the page. If the wire count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { --m->wire_count; if (m->wire_count == 0) { _pmap_unwire_l3(pmap, va, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * unmap the page table page */ if (m->pindex >= (NUL2E + NUL1E)) { /* l1 page */ pd_entry_t *l0; l0 = pmap_l0(pmap, va); pmap_load_clear(l0); PTE_SYNC(l0); } else if (m->pindex >= NUL2E) { /* l2 page */ pd_entry_t *l1; l1 = pmap_l1(pmap, va); pmap_load_clear(l1); PTE_SYNC(l1); } else { /* l3 page */ pd_entry_t *l2; l2 = pmap_l2(pmap, va); pmap_load_clear(l2); PTE_SYNC(l2); } pmap_resident_count_dec(pmap, 1); if (m->pindex < NUL2E) { /* We just released an l3, unhold the matching l2 */ pd_entry_t *l1, tl1; vm_page_t l2pg; l1 = pmap_l1(pmap, va); tl1 = pmap_load(l1); l2pg = PHYS_TO_VM_PAGE(tl1 & ~ATTR_MASK); pmap_unwire_l3(pmap, va, l2pg, free); } else if (m->pindex < (NUL2E + NUL1E)) { /* We just released an l2, unhold the matching l1 */ pd_entry_t *l0, tl0; vm_page_t l1pg; l0 = pmap_l0(pmap, va); tl0 = pmap_load(l0); l1pg = PHYS_TO_VM_PAGE(tl0 & ~ATTR_MASK); pmap_unwire_l3(pmap, va, l1pg, free); } pmap_invalidate_page(pmap, va); /* * This is a release store so that the ordinary store unmapping * the page table page is globally performed before TLB shoot- * down is begun. */ atomic_subtract_rel_int(&vm_cnt.v_wire_count, 1); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing an l3 entry, this routine is used to * conditionally free the page, and manage the hold/wire counts. */ static int pmap_unuse_l3(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde, struct spglist *free) { vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (0); KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0")); mpte = PHYS_TO_VM_PAGE(ptepde & ~ATTR_MASK); return (pmap_unwire_l3(pmap, va, mpte, free)); } void pmap_pinit0(pmap_t pmap) { PMAP_LOCK_INIT(pmap); bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); pmap->pm_l0 = kernel_pmap->pm_l0; pmap->pm_root.rt_root = 0; } int pmap_pinit(pmap_t pmap) { vm_paddr_t l0phys; vm_page_t l0pt; /* * allocate the l0 page */ while ((l0pt = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) VM_WAIT; l0phys = VM_PAGE_TO_PHYS(l0pt); pmap->pm_l0 = (pd_entry_t *)PHYS_TO_DMAP(l0phys); if ((l0pt->flags & PG_ZERO) == 0) pagezero(pmap->pm_l0); pmap->pm_root.rt_root = 0; bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); return (1); } /* * This routine is called if the desired page table page does not exist. * * If page table page allocation fails, this routine may sleep before * returning NULL. It sleeps only if a lock pointer was given. * * Note: If a page allocation fails at page table level two or three, * one or two pages may be held during the wait, only to be released * afterwards. This conservative approach is easily argued to avoid * race conditions. */ static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp) { vm_page_t m, l1pg, l2pg; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Allocate a page table page. */ if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { if (lockp != NULL) { RELEASE_PV_LIST_LOCK(lockp); PMAP_UNLOCK(pmap); VM_WAIT; PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, the page table * page may have been allocated. */ return (NULL); } if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); /* * Map the pagetable page into the process address space, if * it isn't already there. */ if (ptepindex >= (NUL2E + NUL1E)) { pd_entry_t *l0; vm_pindex_t l0index; l0index = ptepindex - (NUL2E + NUL1E); l0 = &pmap->pm_l0[l0index]; pmap_load_store(l0, VM_PAGE_TO_PHYS(m) | L0_TABLE); PTE_SYNC(l0); } else if (ptepindex >= NUL2E) { vm_pindex_t l0index, l1index; pd_entry_t *l0, *l1; pd_entry_t tl0; l1index = ptepindex - NUL2E; l0index = l1index >> L0_ENTRIES_SHIFT; l0 = &pmap->pm_l0[l0index]; tl0 = pmap_load(l0); if (tl0 == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUL2E + NUL1E + l0index, lockp) == NULL) { --m->wire_count; /* XXX: release mem barrier? */ atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } } else { l1pg = PHYS_TO_VM_PAGE(tl0 & ~ATTR_MASK); l1pg->wire_count++; } l1 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l0) & ~ATTR_MASK); l1 = &l1[ptepindex & Ln_ADDR_MASK]; pmap_load_store(l1, VM_PAGE_TO_PHYS(m) | L1_TABLE); PTE_SYNC(l1); } else { vm_pindex_t l0index, l1index; pd_entry_t *l0, *l1, *l2; pd_entry_t tl0, tl1; l1index = ptepindex >> Ln_ENTRIES_SHIFT; l0index = l1index >> L0_ENTRIES_SHIFT; l0 = &pmap->pm_l0[l0index]; tl0 = pmap_load(l0); if (tl0 == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUL2E + l1index, lockp) == NULL) { --m->wire_count; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } tl0 = pmap_load(l0); l1 = (pd_entry_t *)PHYS_TO_DMAP(tl0 & ~ATTR_MASK); l1 = &l1[l1index & Ln_ADDR_MASK]; } else { l1 = (pd_entry_t *)PHYS_TO_DMAP(tl0 & ~ATTR_MASK); l1 = &l1[l1index & Ln_ADDR_MASK]; tl1 = pmap_load(l1); if (tl1 == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUL2E + l1index, lockp) == NULL) { --m->wire_count; /* XXX: release mem barrier? */ atomic_subtract_int( &vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } } else { l2pg = PHYS_TO_VM_PAGE(tl1 & ~ATTR_MASK); l2pg->wire_count++; } } l2 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l1) & ~ATTR_MASK); l2 = &l2[ptepindex & Ln_ADDR_MASK]; pmap_load_store(l2, VM_PAGE_TO_PHYS(m) | L2_TABLE); PTE_SYNC(l2); } pmap_resident_count_inc(pmap, 1); return (m); } static vm_page_t pmap_alloc_l3(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t ptepindex; pd_entry_t *pde, tpde; #ifdef INVARIANTS pt_entry_t *pte; #endif vm_page_t m; int lvl; /* * Calculate pagetable page index */ ptepindex = pmap_l2_pindex(va); retry: /* * Get the page directory entry */ pde = pmap_pde(pmap, va, &lvl); /* * If the page table page is mapped, we just increment the hold count, * and activate it. If we get a level 2 pde it will point to a level 3 * table. */ switch (lvl) { case -1: break; case 0: #ifdef INVARIANTS pte = pmap_l0_to_l1(pde, va); KASSERT(pmap_load(pte) == 0, ("pmap_alloc_l3: TODO: l0 superpages")); #endif break; case 1: #ifdef INVARIANTS pte = pmap_l1_to_l2(pde, va); KASSERT(pmap_load(pte) == 0, ("pmap_alloc_l3: TODO: l1 superpages")); #endif break; case 2: tpde = pmap_load(pde); if (tpde != 0) { m = PHYS_TO_VM_PAGE(tpde & ~ATTR_MASK); m->wire_count++; return (m); } break; default: panic("pmap_alloc_l3: Invalid level %d", lvl); } /* * Here if the pte page isn't mapped, or if it has been deallocated. */ m = _pmap_alloc_l3(pmap, ptepindex, lockp); if (m == NULL && lockp != NULL) goto retry; return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { vm_page_t m; KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap has reserved page table page(s)")); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_l0)); m->wire_count--; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; return sysctl_handle_long(oidp, &ksize, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_size, "LU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return sysctl_handle_long(oidp, &kfree, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_free, "LU", "Amount of KVM free"); /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t paddr; vm_page_t nkpg; pd_entry_t *l0, *l1, *l2; mtx_assert(&kernel_map->system_mtx, MA_OWNED); addr = roundup2(addr, L2_SIZE); if (addr - 1 >= kernel_map->max_offset) addr = kernel_map->max_offset; while (kernel_vm_end < addr) { l0 = pmap_l0(kernel_pmap, kernel_vm_end); KASSERT(pmap_load(l0) != 0, ("pmap_growkernel: No level 0 kernel entry")); l1 = pmap_l0_to_l1(l0, kernel_vm_end); if (pmap_load(l1) == 0) { /* We need a new PDP entry */ nkpg = vm_page_alloc(NULL, kernel_vm_end >> L1_SHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); if ((nkpg->flags & PG_ZERO) == 0) pmap_zero_page(nkpg); paddr = VM_PAGE_TO_PHYS(nkpg); pmap_load_store(l1, paddr | L1_TABLE); PTE_SYNC(l1); continue; /* try again */ } l2 = pmap_l1_to_l2(l1, kernel_vm_end); if ((pmap_load(l2) & ATTR_AF) != 0) { kernel_vm_end = (kernel_vm_end + L2_SIZE) & ~L2_OFFSET; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } continue; } nkpg = vm_page_alloc(NULL, kernel_vm_end >> L2_SHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); if ((nkpg->flags & PG_ZERO) == 0) pmap_zero_page(nkpg); paddr = VM_PAGE_TO_PHYS(nkpg); pmap_load_store(l2, paddr | L2_TABLE); PTE_SYNC(l2); pmap_invalidate_page(kernel_pmap, kernel_vm_end); kernel_vm_end = (kernel_vm_end + L2_SIZE) & ~L2_OFFSET; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 3); CTASSERT(_NPCPV == 168); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0 0xfffffffffffffffful #define PC_FREE1 0xfffffffffffffffful #define PC_FREE2 0x000000fffffffffful static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 }; #if 0 #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, "Current number of pv entry chunks"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, "Current number of pv entry chunks allocated"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, "Current number of pv entry chunks frees"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, "Number of times tried to get a chunk page but failed."); static long pv_entry_frees, pv_entry_allocs, pv_entry_count; static int pv_entry_spare; SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, "Current number of pv entry frees"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, "Current number of pv entry allocs"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, "Current number of pv entries"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, "Current number of spare pv entries"); #endif #endif /* 0 */ /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. * * Returns NULL if PV entries were reclaimed from the specified pmap. * * We do not, however, unmap 2mpages because subsequent accesses will * allocate per-page pv entries until repromotion occurs, thereby * exacerbating the shortage of free pv entries. */ static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) { panic("ARM64TODO: reclaim_pv_chunk"); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_frees, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, 1)); PV_STAT(atomic_subtract_long(&pv_entry_count, 1)); pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 || pc->pc_map[2] != PC_FREE2) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk(struct pv_chunk *pc) { vm_page_t m; mtx_lock(&pv_chunks_mutex); TAILQ_REMOVE(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m->phys_addr); vm_page_unwire(m, PQ_NONE); vm_page_free(m); } /* * Returns a new PV entry, allocating a new PV chunk from the system when * needed. If this PV chunk allocation fails and a PV list lock pointer was * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is * returned. * * The given PV list lock may be released. */ static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp) { int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_allocs, 1)); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = ffsl(pc->pc_map[field]) - 1; break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 64 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, 1)); return (pv); } } /* No free items, allocate another chunk */ m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); if (m == NULL) { if (lockp == NULL) { PV_STAT(pc_chunk_tryfail++); return (NULL); } m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */ pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; mtx_lock(&pv_chunks_mutex); TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1)); return (pv); } /* * Ensure that the number of spare PV entries in the specified pmap meets or * exceeds the given count, "needed". * * The given PV list lock may be released. */ static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp) { struct pch new_tail; struct pv_chunk *pc; int avail, free; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL")); /* * Newly allocated PV chunks must be stored in a private list until * the required number of PV chunks have been allocated. Otherwise, * reclaim_pv_chunk() could recycle one of these chunks. In * contrast, these chunks must be added to the pmap upon allocation. */ TAILQ_INIT(&new_tail); retry: avail = 0; TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) { bit_count((bitstr_t *)pc->pc_map, 0, sizeof(pc->pc_map) * NBBY, &free); if (free == 0) break; avail += free; if (avail >= needed) break; } for (; avail < needed; avail += _NPCPV) { m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); if (m == NULL) { m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0; pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail, pc, pc_lru); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV)); } if (!TAILQ_EMPTY(&new_tail)) { mtx_lock(&pv_chunks_mutex); TAILQ_CONCAT(&pv_chunks, &new_tail, pc_lru); mtx_unlock(&pv_chunks_mutex); } } /* * First find and then remove the pv entry for the specified pmap and virtual * address from the specified pv list. Returns the pv entry if found and NULL * otherwise. This operation can be performed on pv lists for either 4KB or * 2MB page mappings. */ static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; break; } } return (pv); } /* * After demotion from a 2MB page mapping to 512 4KB page mappings, * destroy the pv entry for the 2MB page mapping and reinstantiate the pv * entries for each of the 4KB page mappings. */ static void pmap_pv_demote_l2(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; struct pv_chunk *pc; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; int bit, field; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((pa & L2_OFFSET) == 0, ("pmap_pv_demote_l2: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the 2mpage's pv entry for this mapping to the first * page's pv list. Once this transfer begins, the pv list lock * must not be released until the last pv entry is reinstantiated. */ pvh = pa_to_pvh(pa); va = va & ~L2_OFFSET; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_l2: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; /* Instantiate the remaining Ln_ENTRIES - 1 pv entries. */ PV_STAT(atomic_add_long(&pv_entry_allocs, Ln_ENTRIES - 1)); va_last = va + L2_SIZE - PAGE_SIZE; for (;;) { pc = TAILQ_FIRST(&pmap->pm_pvchunk); KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0 || pc->pc_map[2] != 0, ("pmap_pv_demote_l2: missing spare")); for (field = 0; field < _NPCM; field++) { while (pc->pc_map[field]) { bit = ffsl(pc->pc_map[field]) - 1; pc->pc_map[field] &= ~(1ul << bit); pv = &pc->pc_pventry[field * 64 + bit]; va += PAGE_SIZE; pv->pv_va = va; m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_l2: page %p is not managed", m)); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (va == va_last) goto out; } } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } out: if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, Ln_ENTRIES - 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, Ln_ENTRIES - 1)); } /* * First find and then destroy the pv entry for the specified pmap and virtual * address. This operation can be performed on pv lists for either 4KB or 2MB * page mappings. */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } /* * Conditionally create the PV entry for a 4KB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp) { pv_entry_t pv; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; return (TRUE); } else return (FALSE); } /* * pmap_remove_l3: do the things to unmap a page in a process */ static int pmap_remove_l3(pmap_t pmap, pt_entry_t *l3, vm_offset_t va, pd_entry_t l2e, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pt_entry_t old_l3; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(pmap_load(l3))) cpu_dcache_wb_range(va, L3_SIZE); old_l3 = pmap_load_clear(l3); PTE_SYNC(l3); pmap_invalidate_page(pmap, va); if (old_l3 & ATTR_SW_WIRED) pmap->pm_stats.wired_count -= 1; pmap_resident_count_dec(pmap, 1); if (old_l3 & ATTR_SW_MANAGED) { m = PHYS_TO_VM_PAGE(old_l3 & ~ATTR_MASK); if (pmap_page_dirty(old_l3)) vm_page_dirty(m); if (old_l3 & ATTR_AF) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } return (pmap_unuse_l3(pmap, va, l2e, free)); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct rwlock *lock; vm_offset_t va, va_next; pd_entry_t *l0, *l1, *l2; pt_entry_t l3_paddr, *l3; struct spglist free; int anyvalid; /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; anyvalid = 0; SLIST_INIT(&free); PMAP_LOCK(pmap); lock = NULL; for (; sva < eva; sva = va_next) { if (pmap->pm_stats.resident_count == 0) break; l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } l1 = pmap_l0_to_l1(l0, sva); if (pmap_load(l1) == 0) { va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; continue; } /* * Calculate index for next page table. */ va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (l2 == NULL) continue; l3_paddr = pmap_load(l2); if ((l3_paddr & ATTR_DESCR_MASK) == L2_BLOCK) { /* TODO: Add pmap_remove_l2 */ if (pmap_demote_l2_locked(pmap, l2, sva & ~L2_OFFSET, &lock) == NULL) continue; l3_paddr = pmap_load(l2); } /* * Weed out invalid mappings. */ if ((l3_paddr & ATTR_DESCR_MASK) != L2_TABLE) continue; /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (va_next > eva) va_next = eva; va = va_next; for (l3 = pmap_l2_to_l3(l2, sva); sva != va_next; l3++, sva += L3_SIZE) { if (l3 == NULL) panic("l3 == NULL"); if (pmap_load(l3) == 0) { if (va != va_next) { pmap_invalidate_range(pmap, va, sva); va = va_next; } continue; } if (va == va_next) va = sva; if (pmap_remove_l3(pmap, l3, sva, l3_paddr, &free, &lock)) { sva += L3_SIZE; break; } } if (va != va_next) pmap_invalidate_range(pmap, va, sva); } if (lock != NULL) rw_wunlock(lock); if (anyvalid) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; struct rwlock *lock; pd_entry_t *pde, tpde; pt_entry_t *pte, tpte; vm_offset_t va; struct spglist free; int lvl, pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); retry: rw_wlock(lock); while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { rw_wunlock(lock); PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pte = pmap_pte(pmap, va, &lvl); KASSERT(pte != NULL, ("pmap_remove_all: no page table entry found")); KASSERT(lvl == 2, ("pmap_remove_all: invalid pte level %d", lvl)); pmap_demote_l2_locked(pmap, pte, va, &lock); PMAP_UNLOCK(pmap); } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { rw_wunlock(lock); PMAP_UNLOCK(pmap); goto retry; } } pmap_resident_count_dec(pmap, 1); pde = pmap_pde(pmap, pv->pv_va, &lvl); KASSERT(pde != NULL, ("pmap_remove_all: no page directory entry found")); KASSERT(lvl == 2, ("pmap_remove_all: invalid pde level %d", lvl)); tpde = pmap_load(pde); pte = pmap_l2_to_l3(pde, pv->pv_va); tpte = pmap_load(pte); if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(tpte)) cpu_dcache_wb_range(pv->pv_va, L3_SIZE); pmap_load_clear(pte); PTE_SYNC(pte); pmap_invalidate_page(pmap, pv->pv_va); if (tpte & ATTR_SW_WIRED) pmap->pm_stats.wired_count--; if ((tpte & ATTR_AF) != 0) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if (pmap_page_dirty(tpte)) vm_page_dirty(m); pmap_unuse_l3(pmap, pv->pv_va, tpde, &free); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(lock); pmap_free_zero_pages(&free); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t va, va_next; pd_entry_t *l0, *l1, *l2; pt_entry_t *l3p, l3; if ((prot & VM_PROT_READ) == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } if ((prot & VM_PROT_WRITE) == VM_PROT_WRITE) return; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } l1 = pmap_l0_to_l1(l0, sva); if (pmap_load(l1) == 0) { va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (pmap_load(l2) == 0) continue; if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) { l3p = pmap_demote_l2(pmap, l2, sva); if (l3p == NULL) continue; } KASSERT((pmap_load(l2) & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_protect: Invalid L2 entry after demotion")); if (va_next > eva) va_next = eva; va = va_next; for (l3p = pmap_l2_to_l3(l2, sva); sva != va_next; l3p++, sva += L3_SIZE) { l3 = pmap_load(l3p); if (pmap_l3_valid(l3)) { pmap_set(l3p, ATTR_AP(ATTR_AP_RO)); PTE_SYNC(l3p); /* XXX: Use pmap_invalidate_range */ pmap_invalidate_page(pmap, va); } } } PMAP_UNLOCK(pmap); /* TODO: Only invalidate entries we are touching */ pmap_invalidate_all(pmap); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Looks for a page table page mapping the specified virtual address in the * specified pmap's collection of idle page table pages. Returns NULL if there * is no page table page corresponding to the specified virtual address. */ static __inline vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_lookup(&pmap->pm_root, pmap_l2_pindex(va))); } /* * Removes the specified page table page from the specified pmap's collection * of idle page table pages. The specified page table page must be a member of * the pmap's collection. */ static __inline void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); vm_radix_remove(&pmap->pm_root, mpte->pindex); } /* * Performs a break-before-make update of a pmap entry. This is needed when * either promoting or demoting pages to ensure the TLB doesn't get into an * inconsistent state. */ static void pmap_update_entry(pmap_t pmap, pd_entry_t *pte, pd_entry_t newpte, vm_offset_t va, vm_size_t size) { register_t intr; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Ensure we don't get switched out with the page table in an * inconsistent state. We also need to ensure no interrupts fire * as they may make use of an address we are about to invalidate. */ intr = intr_disable(); critical_enter(); /* Clear the old mapping */ pmap_load_clear(pte); PTE_SYNC(pte); pmap_invalidate_range(pmap, va, va + size); /* Create the new mapping */ pmap_load_store(pte, newpte); PTE_SYNC(pte); critical_exit(); intr_restore(intr); } /* * After promotion from 512 4KB page mappings to a single 2MB page mapping, * replace the many pv entries for the 4KB page mappings by a single pv entry * for the 2MB page mapping. */ static void pmap_pv_promote_l2(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; KASSERT((pa & L2_OFFSET) == 0, ("pmap_pv_promote_l2: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the first page's pv entry for this mapping to the 2mpage's * pv list. Aside from avoiding the cost of a call to get_pv_entry(), * a transfer avoids the possibility that get_pv_entry() calls * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the * mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = va & ~L2_OFFSET; pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_l2: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + L2_SIZE - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } /* * Tries to promote the 512, contiguous 4KB page mappings that are within a * single level 2 table entry to a single 2MB page mapping. For promotion * to occur, two conditions must be met: (1) the 4KB page mappings must map * aligned, contiguous physical memory and (2) the 4KB page mappings must have * identical characteristics. */ static void pmap_promote_l2(pmap_t pmap, pd_entry_t *l2, vm_offset_t va, struct rwlock **lockp) { pt_entry_t *firstl3, *l3, newl2, oldl3, pa; vm_page_t mpte; vm_offset_t sva; PMAP_LOCK_ASSERT(pmap, MA_OWNED); sva = va & ~L2_OFFSET; firstl3 = pmap_l2_to_l3(l2, sva); newl2 = pmap_load(firstl3); /* Check the alingment is valid */ if (((newl2 & ~ATTR_MASK) & L2_OFFSET) != 0) { atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx" " in pmap %p", va, pmap); return; } pa = newl2 + L2_SIZE - PAGE_SIZE; for (l3 = firstl3 + NL3PG - 1; l3 > firstl3; l3--) { oldl3 = pmap_load(l3); if (oldl3 != pa) { atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx" " in pmap %p", va, pmap); return; } pa -= PAGE_SIZE; } /* * Save the page table page in its current state until the L2 * mapping the superpage is demoted by pmap_demote_l2() or * destroyed by pmap_remove_l3(). */ mpte = PHYS_TO_VM_PAGE(pmap_load(l2) & ~ATTR_MASK); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_l2: page table page is out of range")); KASSERT(mpte->pindex == pmap_l2_pindex(va), ("pmap_promote_l2: page table page's pindex is wrong")); if (pmap_insert_pt_page(pmap, mpte)) { atomic_add_long(&pmap_l2_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_l2: failure for va %#lx in pmap %p", va, pmap); return; } if ((newl2 & ATTR_SW_MANAGED) != 0) pmap_pv_promote_l2(pmap, va, newl2 & ~ATTR_MASK, lockp); newl2 &= ~ATTR_DESCR_MASK; newl2 |= L2_BLOCK; pmap_update_entry(pmap, l2, newl2, sva, L2_SIZE); atomic_add_long(&pmap_l2_promotions, 1); CTR2(KTR_PMAP, "pmap_promote_l2: success for va %#lx in pmap %p", va, pmap); } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind __unused) { struct rwlock *lock; pd_entry_t *pde; pt_entry_t new_l3, orig_l3; pt_entry_t *l2, *l3; pv_entry_t pv; vm_paddr_t opa, pa, l1_pa, l2_pa, l3_pa; vm_page_t mpte, om, l1_m, l2_m, l3_m; boolean_t nosleep; int lvl; va = trunc_page(va); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); pa = VM_PAGE_TO_PHYS(m); new_l3 = (pt_entry_t)(pa | ATTR_DEFAULT | ATTR_IDX(m->md.pv_memattr) | L3_PAGE); if ((prot & VM_PROT_WRITE) == 0) new_l3 |= ATTR_AP(ATTR_AP_RO); if ((flags & PMAP_ENTER_WIRED) != 0) new_l3 |= ATTR_SW_WIRED; if ((va >> 63) == 0) new_l3 |= ATTR_AP(ATTR_AP_USER); CTR2(KTR_PMAP, "pmap_enter: %.16lx -> %.16lx", va, pa); mpte = NULL; lock = NULL; PMAP_LOCK(pmap); pde = pmap_pde(pmap, va, &lvl); if (pde != NULL && lvl == 1) { l2 = pmap_l1_to_l2(pde, va); if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK && (l3 = pmap_demote_l2_locked(pmap, l2, va & ~L2_OFFSET, &lock)) != NULL) { l3 = &l3[pmap_l3_index(va)]; if (va < VM_MAXUSER_ADDRESS) { mpte = PHYS_TO_VM_PAGE( pmap_load(l2) & ~ATTR_MASK); mpte->wire_count++; } goto havel3; } } if (va < VM_MAXUSER_ADDRESS) { nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0; mpte = pmap_alloc_l3(pmap, va, nosleep ? NULL : &lock); if (mpte == NULL && nosleep) { CTR0(KTR_PMAP, "pmap_enter: mpte == NULL"); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (KERN_RESOURCE_SHORTAGE); } pde = pmap_pde(pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_enter: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_enter: Invalid level %d", lvl)); l3 = pmap_l2_to_l3(pde, va); } else { /* * If we get a level 2 pde it must point to a level 3 entry * otherwise we will need to create the intermediate tables */ if (lvl < 2) { switch(lvl) { default: case -1: /* Get the l0 pde to update */ pde = pmap_l0(pmap, va); KASSERT(pde != NULL, ("...")); l1_m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (l1_m == NULL) panic("pmap_enter: l1 pte_m == NULL"); if ((l1_m->flags & PG_ZERO) == 0) pmap_zero_page(l1_m); l1_pa = VM_PAGE_TO_PHYS(l1_m); pmap_load_store(pde, l1_pa | L0_TABLE); PTE_SYNC(pde); /* FALLTHROUGH */ case 0: /* Get the l1 pde to update */ pde = pmap_l1_to_l2(pde, va); KASSERT(pde != NULL, ("...")); l2_m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (l2_m == NULL) panic("pmap_enter: l2 pte_m == NULL"); if ((l2_m->flags & PG_ZERO) == 0) pmap_zero_page(l2_m); l2_pa = VM_PAGE_TO_PHYS(l2_m); pmap_load_store(pde, l2_pa | L1_TABLE); PTE_SYNC(pde); /* FALLTHROUGH */ case 1: /* Get the l2 pde to update */ pde = pmap_l1_to_l2(pde, va); l3_m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (l3_m == NULL) panic("pmap_enter: l3 pte_m == NULL"); if ((l3_m->flags & PG_ZERO) == 0) pmap_zero_page(l3_m); l3_pa = VM_PAGE_TO_PHYS(l3_m); pmap_load_store(pde, l3_pa | L2_TABLE); PTE_SYNC(pde); break; } } l3 = pmap_l2_to_l3(pde, va); pmap_invalidate_page(pmap, va); } havel3: om = NULL; orig_l3 = pmap_load(l3); opa = orig_l3 & ~ATTR_MASK; /* * Is the specified virtual address already mapped? */ if (pmap_l3_valid(orig_l3)) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((flags & PMAP_ENTER_WIRED) != 0 && (orig_l3 & ATTR_SW_WIRED) == 0) pmap->pm_stats.wired_count++; else if ((flags & PMAP_ENTER_WIRED) == 0 && (orig_l3 & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->wire_count--; KASSERT(mpte->wire_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%lx", va)); } /* * Has the physical page changed? */ if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((orig_l3 & ATTR_SW_MANAGED) != 0) { new_l3 |= ATTR_SW_MANAGED; if ((new_l3 & ATTR_AP(ATTR_AP_RW)) == ATTR_AP(ATTR_AP_RW)) { vm_page_aflag_set(m, PGA_WRITEABLE); } } goto validate; } /* Flush the cache, there might be uncommitted data in it */ if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(orig_l3)) cpu_dcache_wb_range(va, L3_SIZE); } else { /* * Increment the counters. */ if ((new_l3 & ATTR_SW_WIRED) != 0) pmap->pm_stats.wired_count++; pmap_resident_count_inc(pmap, 1); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { new_l3 |= ATTR_SW_MANAGED; pv = get_pv_entry(pmap, &lock); pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((new_l3 & ATTR_AP_RW_BIT) == ATTR_AP(ATTR_AP_RW)) vm_page_aflag_set(m, PGA_WRITEABLE); } /* * Update the L3 entry. */ if (orig_l3 != 0) { validate: orig_l3 = pmap_load(l3); opa = orig_l3 & ~ATTR_MASK; if (opa != pa) { pmap_update_entry(pmap, l3, new_l3, va, PAGE_SIZE); if ((orig_l3 & ATTR_SW_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); if (pmap_page_dirty(orig_l3)) vm_page_dirty(om); if ((orig_l3 & ATTR_AF) != 0) vm_page_aflag_set(om, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa); pmap_pvh_free(&om->md, pmap, va); if ((om->aflags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } } else { pmap_load_store(l3, new_l3); PTE_SYNC(l3); pmap_invalidate_page(pmap, va); if (pmap_page_dirty(orig_l3) && (orig_l3 & ATTR_SW_MANAGED) != 0) vm_page_dirty(m); } } else { pmap_load_store(l3, new_l3); } PTE_SYNC(l3); pmap_invalidate_page(pmap, va); if (pmap != pmap_kernel()) { if (pmap == &curproc->p_vmspace->vm_pmap && (prot & VM_PROT_EXECUTE) != 0) cpu_icache_sync_range(va, PAGE_SIZE); if ((mpte == NULL || mpte->wire_count == NL3PG) && pmap_superpages_enabled() && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) { pmap_promote_l2(pmap, pde, va, &lock); } } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { struct rwlock *lock; vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; lock = NULL; PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte, &lock); m = TAILQ_NEXT(m, listq); } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { struct rwlock *lock; lock = NULL; PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp) { struct spglist free; pd_entry_t *pde; pt_entry_t *l2, *l3; vm_paddr_t pa; int lvl; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); PMAP_LOCK_ASSERT(pmap, MA_OWNED); CTR2(KTR_PMAP, "pmap_enter_quick_locked: %p %lx", pmap, va); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { vm_pindex_t l2pindex; /* * Calculate pagetable page index */ l2pindex = pmap_l2_pindex(va); if (mpte && (mpte->pindex == l2pindex)) { mpte->wire_count++; } else { /* * Get the l2 entry */ pde = pmap_pde(pmap, va, &lvl); /* * If the page table page is mapped, we just increment * the hold count, and activate it. Otherwise, we * attempt to allocate a page table page. If this * attempt fails, we don't retry. Instead, we give up. */ if (lvl == 1) { l2 = pmap_l1_to_l2(pde, va); if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) return (NULL); } if (lvl == 2 && pmap_load(pde) != 0) { mpte = PHYS_TO_VM_PAGE(pmap_load(pde) & ~ATTR_MASK); mpte->wire_count++; } else { /* * Pass NULL instead of the PV list lock * pointer, because we don't intend to sleep. */ mpte = _pmap_alloc_l3(pmap, l2pindex, NULL); if (mpte == NULL) return (mpte); } } l3 = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte)); l3 = &l3[pmap_l3_index(va)]; } else { mpte = NULL; pde = pmap_pde(kernel_pmap, va, &lvl); KASSERT(pde != NULL, ("pmap_enter_quick_locked: Invalid page entry, va: 0x%lx", va)); KASSERT(lvl == 2, ("pmap_enter_quick_locked: Invalid level %d", lvl)); l3 = pmap_l2_to_l3(pde, va); } if (pmap_load(l3) != 0) { if (mpte != NULL) { mpte->wire_count--; mpte = NULL; } return (mpte); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m, lockp)) { if (mpte != NULL) { SLIST_INIT(&free); if (pmap_unwire_l3(pmap, va, mpte, &free)) { pmap_invalidate_page(pmap, va); pmap_free_zero_pages(&free); } mpte = NULL; } return (mpte); } /* * Increment counters */ pmap_resident_count_inc(pmap, 1); pa = VM_PAGE_TO_PHYS(m) | ATTR_DEFAULT | ATTR_IDX(m->md.pv_memattr) | ATTR_AP(ATTR_AP_RO) | L3_PAGE; /* * Now validate mapping with RO protection */ if ((m->oflags & VPO_UNMANAGED) == 0) pa |= ATTR_SW_MANAGED; pmap_load_store(l3, pa); PTE_SYNC(l3); pmap_invalidate_page(pmap, va); return (mpte); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, * so there is no need to invalidate any TLB entries. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va_next; pd_entry_t *l0, *l1, *l2; pt_entry_t *l3; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { l0 = pmap_l0(pmap, sva); if (pmap_load(l0) == 0) { va_next = (sva + L0_SIZE) & ~L0_OFFSET; if (va_next < sva) va_next = eva; continue; } l1 = pmap_l0_to_l1(l0, sva); if (pmap_load(l1) == 0) { va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (pmap_load(l2) == 0) continue; if ((pmap_load(l2) & ATTR_DESCR_MASK) == L2_BLOCK) { l3 = pmap_demote_l2(pmap, l2, sva); if (l3 == NULL) continue; } KASSERT((pmap_load(l2) & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_unwire: Invalid l2 entry after demotion")); if (va_next > eva) va_next = eva; for (l3 = pmap_l2_to_l3(l2, sva); sva != va_next; l3++, sva += L3_SIZE) { if (pmap_load(l3) == 0) continue; if ((pmap_load(l3) & ATTR_SW_WIRED) == 0) panic("pmap_unwire: l3 %#jx is missing " "ATTR_SW_WIRED", (uintmax_t)pmap_load(l3)); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. */ atomic_clear_long(l3, ATTR_SW_WIRED); pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { } /* * pmap_zero_page zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. */ void pmap_zero_page(vm_page_t m) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); pagezero((void *)va); } /* * pmap_zero_page_area zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. * * off and size may not cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); if (off == 0 && size == PAGE_SIZE) pagezero((void *)va); else bzero((char *)va + off, size); } /* * pmap_copy_page copies the specified (machine independent) * page by mapping the page into virtual memory and using * bcopy to copy the page, one machine dependent page at a * time. */ void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc)); vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst)); pagecopy((void *)src, (void *)dst); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_page_t m_a, m_b; vm_paddr_t p_a, p_b; vm_offset_t a_pg_offset, b_pg_offset; int cnt; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; m_a = ma[a_offset >> PAGE_SHIFT]; p_a = m_a->phys_addr; b_pg_offset = b_offset & PAGE_MASK; m_b = mb[b_offset >> PAGE_SHIFT]; p_b = m_b->phys_addr; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); cnt = min(cnt, PAGE_SIZE - b_pg_offset); if (__predict_false(!PHYS_IN_DMAP(p_a))) { panic("!DMAP a %lx", p_a); } else { a_cp = (char *)PHYS_TO_DMAP(p_a) + a_pg_offset; } if (__predict_false(!PHYS_IN_DMAP(p_b))) { panic("!DMAP b %lx", p_b); } else { b_cp = (char *)PHYS_TO_DMAP(p_b) + b_pg_offset; } bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { return (PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m))); } void pmap_quick_remove_page(vm_offset_t addr) { } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; struct rwlock *lock; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_runlock(lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct rwlock *lock; struct md_page *pvh; pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; int count, lvl, md_gen, pvh_gen; if ((m->oflags & VPO_UNMANAGED) != 0) return (0); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: count = 0; TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va, &lvl); if (pte != NULL && (pmap_load(pte) & ATTR_SW_WIRED) != 0) count++; PMAP_UNLOCK(pmap); } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va, &lvl); if (pte != NULL && (pmap_load(pte) & ATTR_SW_WIRED) != 0) count++; PMAP_UNLOCK(pmap); } } rw_runlock(lock); return (count); } /* * Destroy all managed, non-wired mappings in the given user-space * pmap. This pmap cannot be active on any processor besides the * caller. * * This function cannot be applied to the kernel pmap. Moreover, it * is not intended for general use. It is only to be used during * process termination. Consequently, it can be implemented in ways * that make it faster than pmap_remove(). First, it can more quickly * destroy mappings by iterating over the pmap's collection of PV * entries, rather than searching the page table. Second, it doesn't * have to test and clear the page table entries atomically, because * no processor is currently accessing the user address space. In * particular, a page table entry's dirty bit won't change state once * this function starts. */ void pmap_remove_pages(pmap_t pmap) { pd_entry_t *pde; pt_entry_t *pte, tpte; struct spglist free; vm_page_t m, ml3, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct rwlock *lock; int64_t bit; uint64_t inuse, bitmask; int allfree, field, freed, idx, lvl; vm_paddr_t pa; lock = NULL; SLIST_INIT(&free); PMAP_LOCK(pmap); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { allfree = 1; freed = 0; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = ffsl(inuse) - 1; bitmask = 1UL << bit; idx = field * 64 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pde = pmap_pde(pmap, pv->pv_va, &lvl); KASSERT(pde != NULL, ("Attempting to remove an unmapped page")); switch(lvl) { case 1: pte = pmap_l1_to_l2(pde, pv->pv_va); tpte = pmap_load(pte); KASSERT((tpte & ATTR_DESCR_MASK) == L2_BLOCK, ("Attempting to remove an invalid " "block: %lx", tpte)); tpte = pmap_load(pte); break; case 2: pte = pmap_l2_to_l3(pde, pv->pv_va); tpte = pmap_load(pte); KASSERT((tpte & ATTR_DESCR_MASK) == L3_PAGE, ("Attempting to remove an invalid " "page: %lx", tpte)); break; default: panic( "Invalid page directory level: %d", lvl); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & ATTR_SW_WIRED) { allfree = 0; continue; } pa = tpte & ~ATTR_MASK; m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad pte %#jx", (uintmax_t)tpte)); if (pmap_is_current(pmap)) { if (lvl == 2 && pmap_l3_valid_cacheable(tpte)) { cpu_dcache_wb_range(pv->pv_va, L3_SIZE); } else if (lvl == 1 && pmap_pte_valid_cacheable(tpte)) { cpu_dcache_wb_range(pv->pv_va, L2_SIZE); } } pmap_load_clear(pte); PTE_SYNC(pte); pmap_invalidate_page(pmap, pv->pv_va); /* * Update the vm_page_t clean/reference bits. */ if ((tpte & ATTR_AP_RW_BIT) == ATTR_AP(ATTR_AP_RW)) { switch (lvl) { case 1: for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) vm_page_dirty(m); break; case 2: vm_page_dirty(m); break; } } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); /* Mark free */ pc->pc_map[field] |= bitmask; switch (lvl) { case 1: pmap_resident_count_dec(pmap, L2_SIZE / PAGE_SIZE); pvh = pa_to_pvh(tpte & ~ATTR_MASK); TAILQ_REMOVE(&pvh->pv_list, pv,pv_next); pvh->pv_gen++; if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[L2_SIZE / PAGE_SIZE]; mt++) if ((mt->aflags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } ml3 = pmap_lookup_pt_page(pmap, pv->pv_va); if (ml3 != NULL) { pmap_remove_pt_page(pmap, ml3); pmap_resident_count_dec(pmap,1); KASSERT(ml3->wire_count == NL3PG, ("pmap_remove_pages: l3 page wire count error")); ml3->wire_count = 0; pmap_add_delayed_free_list(ml3, &free, FALSE); atomic_subtract_int( &vm_cnt.v_wire_count, 1); } break; case 2: pmap_resident_count_dec(pmap, 1); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((m->aflags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh( VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } break; } pmap_unuse_l3(pmap, pv->pv_va, pmap_load(pde), &free); freed++; } } PV_STAT(atomic_add_long(&pv_entry_frees, freed)); PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } } pmap_invalidate_all(pmap); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * This is used to check if a page has been accessed or modified. As we * don't have a bit to see if it has been modified we have to assume it * has been if the page is read/write. */ static boolean_t pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified) { struct rwlock *lock; pv_entry_t pv; struct md_page *pvh; pt_entry_t *pte, mask, value; pmap_t pmap; int lvl, md_gen, pvh_gen; boolean_t rv; rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va, &lvl); KASSERT(lvl == 3, ("pmap_page_test_mappings: Invalid level %d", lvl)); mask = 0; value = 0; if (modified) { mask |= ATTR_AP_RW_BIT; value |= ATTR_AP(ATTR_AP_RW); } if (accessed) { mask |= ATTR_AF | ATTR_DESCR_MASK; value |= ATTR_AF | L3_PAGE; } rv = (pmap_load(pte) & mask) == value; PMAP_UNLOCK(pmap); if (rv) goto out; } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va, &lvl); KASSERT(lvl == 2, ("pmap_page_test_mappings: Invalid level %d", lvl)); mask = 0; value = 0; if (modified) { mask |= ATTR_AP_RW_BIT; value |= ATTR_AP(ATTR_AP_RW); } if (accessed) { mask |= ATTR_AF | ATTR_DESCR_MASK; value |= ATTR_AF | L2_BLOCK; } rv = (pmap_load(pte) & mask) == value; PMAP_UNLOCK(pmap); if (rv) goto out; } } out: rw_runlock(lock); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no PTEs can have PG_M set. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (FALSE); return (pmap_page_test_mappings(m, FALSE, TRUE)); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pt_entry_t *pte; boolean_t rv; int lvl; rv = FALSE; PMAP_LOCK(pmap); pte = pmap_pte(pmap, addr, &lvl); if (pte != NULL && pmap_load(pte) != 0) { rv = TRUE; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); return (pmap_page_test_mappings(m, TRUE, FALSE)); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pmap_t pmap; struct rwlock *lock; pv_entry_t next_pv, pv; pt_entry_t oldpte, *pte; vm_offset_t va; int lvl, md_gen, pvh_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); retry_pv_loop: rw_wlock(lock); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); rw_wunlock(lock); goto retry_pv_loop; } } va = pv->pv_va; pte = pmap_pte(pmap, pv->pv_va, &lvl); if ((pmap_load(pte) & ATTR_AP_RW_BIT) == ATTR_AP(ATTR_AP_RW)) pmap_demote_l2_locked(pmap, pte, va & ~L2_OFFSET, &lock); KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); rw_wunlock(lock); goto retry_pv_loop; } } pte = pmap_pte(pmap, pv->pv_va, &lvl); retry: oldpte = pmap_load(pte); if ((oldpte & ATTR_AP_RW_BIT) == ATTR_AP(ATTR_AP_RW)) { if (!atomic_cmpset_long(pte, oldpte, oldpte | ATTR_AP(ATTR_AP_RO))) goto retry; if ((oldpte & ATTR_AF) != 0) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); vm_page_aflag_clear(m, PGA_WRITEABLE); } static __inline boolean_t safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte) { return (FALSE); } -#define PMAP_TS_REFERENCED_MAX 5 - /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; struct rwlock *lock; pd_entry_t *pde, tpde; pt_entry_t *pte, tpte; pt_entry_t *l3; vm_offset_t va; vm_paddr_t pa; int cleared, md_gen, not_cleared, lvl, pvh_gen; struct spglist free; bool demoted; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); SLIST_INIT(&free); cleared = 0; pa = VM_PAGE_TO_PHYS(m); lock = PHYS_TO_PV_LIST_LOCK(pa); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa); rw_wlock(lock); retry: not_cleared = 0; if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pde = pmap_pde(pmap, pv->pv_va, &lvl); KASSERT(pde != NULL, ("pmap_ts_referenced: no l1 table found")); KASSERT(lvl == 1, ("pmap_ts_referenced: invalid pde level %d", lvl)); tpde = pmap_load(pde); KASSERT((tpde & ATTR_DESCR_MASK) == L1_TABLE, ("pmap_ts_referenced: found an invalid l1 table")); pte = pmap_l1_to_l2(pde, pv->pv_va); tpte = pmap_load(pte); + if (pmap_page_dirty(tpte)) { + /* + * Although "tpte" is mapping a 2MB page, because + * this function is called at a 4KB page granularity, + * we only update the 4KB page under test. + */ + vm_page_dirty(m); + } if ((tpte & ATTR_AF) != 0) { /* * Since this reference bit is shared by 512 4KB * pages, it should not be cleared every time it is * tested. Apply a simple "hash" function on the * physical page number, the virtual superpage number, * and the pmap address to select one 4KB page out of * the 512 on which testing the reference bit will * result in clearing that reference bit. This * function is designed to avoid the selection of the * same 4KB page for every 2MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> L2_SHIFT) ^ (uintptr_t)pmap) & (Ln_ENTRIES - 1)) == 0 && (tpte & ATTR_SW_WIRED) == 0) { if (safe_to_clear_referenced(pmap, tpte)) { /* * TODO: We don't handle the access * flag at all. We need to be able * to set it in the exception handler. */ panic("ARM64TODO: " "safe_to_clear_referenced\n"); } else if (pmap_demote_l2_locked(pmap, pte, pv->pv_va, &lock) != NULL) { demoted = true; va += VM_PAGE_TO_PHYS(m) - (tpte & ~ATTR_MASK); l3 = pmap_l2_to_l3(pte, va); pmap_remove_l3(pmap, l3, va, pmap_load(pte), NULL, &lock); } else demoted = true; if (demoted) { /* * The superpage mapping was removed * entirely and therefore 'pv' is no * longer valid. */ if (pvf == pv) pvf = NULL; pv = NULL; } cleared++; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; } if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } pde = pmap_pde(pmap, pv->pv_va, &lvl); KASSERT(pde != NULL, ("pmap_ts_referenced: no l2 table found")); KASSERT(lvl == 2, ("pmap_ts_referenced: invalid pde level %d", lvl)); tpde = pmap_load(pde); KASSERT((tpde & ATTR_DESCR_MASK) == L2_TABLE, ("pmap_ts_referenced: found an invalid l2 table")); pte = pmap_l2_to_l3(pde, pv->pv_va); tpte = pmap_load(pte); + if (pmap_page_dirty(tpte)) + vm_page_dirty(m); if ((tpte & ATTR_AF) != 0) { if (safe_to_clear_referenced(pmap, tpte)) { /* * TODO: We don't handle the access flag * at all. We need to be able to set it in * the exception handler. */ panic("ARM64TODO: safe_to_clear_referenced\n"); } else if ((tpte & ATTR_SW_WIRED) == 0) { /* * Wired pages cannot be paged out so * doing accessed bit emulation for * them is wasted effort. We do the * hard work for unwired pages only. */ pmap_remove_l3(pmap, pte, pv->pv_va, tpde, &free, &lock); pmap_invalidate_page(pmap, pv->pv_va); cleared++; if (pvf == pv) pvf = NULL; pv = NULL; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared + not_cleared < PMAP_TS_REFERENCED_MAX); out: rw_wunlock(lock); pmap_free_zero_pages(&free); return (cleared + not_cleared); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("pmap_clear_modify: page %p is exclusive busied", m)); /* * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set. * If the object containing the page is locked and the page is not * exclusive busied, then PGA_WRITEABLE cannot be concurrently set. */ if ((m->aflags & PGA_WRITEABLE) == 0) return; /* ARM64TODO: We lack support for tracking if a page is modified */ } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { return ((void *)PHYS_TO_DMAP(pa)); } void pmap_unmapbios(vm_paddr_t pa, vm_size_t size) { } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pv_memattr = ma; /* * If "m" is a normal page, update its direct mapping. This update * can be relied upon to perform any cache operations that are * required for data coherence. */ if ((m->flags & PG_FICTITIOUS) == 0 && pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, m->md.pv_memattr) != 0) panic("memory attribute change on the direct map failed"); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the direct map or the kernel map. If * the virtual address range is contained within the kernel map, then the * memory type for each of the corresponding ranges of the direct map is also * changed. (The corresponding ranges of the direct map are those ranges that * map the same physical pages as the specified virtual address range.) These * changes to the direct map are necessary because Intel describes the * behavior of their processors as "undefined" if two or more mappings to the * same physical page have different memory types. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. In the * latter case, the memory type may have been changed on some part of the * virtual address range or the direct map. */ static int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) { int error; PMAP_LOCK(kernel_pmap); error = pmap_change_attr_locked(va, size, mode); PMAP_UNLOCK(kernel_pmap); return (error); } static int pmap_change_attr_locked(vm_offset_t va, vm_size_t size, int mode) { vm_offset_t base, offset, tmpva; pt_entry_t l3, *pte, *newpte; int lvl; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); if (!VIRT_IN_DMAP(base)) return (EINVAL); for (tmpva = base; tmpva < base + size; ) { pte = pmap_pte(kernel_pmap, va, &lvl); if (pte == NULL) return (EINVAL); if ((pmap_load(pte) & ATTR_IDX_MASK) == ATTR_IDX(mode)) { /* * We already have the correct attribute, * ignore this entry. */ switch (lvl) { default: panic("Invalid DMAP table level: %d\n", lvl); case 1: tmpva = (tmpva & ~L1_OFFSET) + L1_SIZE; break; case 2: tmpva = (tmpva & ~L2_OFFSET) + L2_SIZE; break; case 3: tmpva += PAGE_SIZE; break; } } else { /* * Split the entry to an level 3 table, then * set the new attribute. */ switch (lvl) { default: panic("Invalid DMAP table level: %d\n", lvl); case 1: newpte = pmap_demote_l1(kernel_pmap, pte, tmpva & ~L1_OFFSET); if (newpte == NULL) return (EINVAL); pte = pmap_l1_to_l2(pte, tmpva); case 2: newpte = pmap_demote_l2(kernel_pmap, pte, tmpva & ~L2_OFFSET); if (newpte == NULL) return (EINVAL); pte = pmap_l2_to_l3(pte, tmpva); case 3: /* Update the entry */ l3 = pmap_load(pte); l3 &= ~ATTR_IDX_MASK; l3 |= ATTR_IDX(mode); pmap_update_entry(kernel_pmap, pte, l3, tmpva, PAGE_SIZE); /* * If moving to a non-cacheable entry flush * the cache. */ if (mode == VM_MEMATTR_UNCACHEABLE) cpu_dcache_wbinv_range(tmpva, L3_SIZE); break; } tmpva += PAGE_SIZE; } } return (0); } /* * Create an L2 table to map all addresses within an L1 mapping. */ static pt_entry_t * pmap_demote_l1(pmap_t pmap, pt_entry_t *l1, vm_offset_t va) { pt_entry_t *l2, newl2, oldl1; vm_offset_t tmpl1; vm_paddr_t l2phys, phys; vm_page_t ml2; int i; PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldl1 = pmap_load(l1); KASSERT((oldl1 & ATTR_DESCR_MASK) == L1_BLOCK, ("pmap_demote_l1: Demoting a non-block entry")); KASSERT((va & L1_OFFSET) == 0, ("pmap_demote_l1: Invalid virtual address %#lx", va)); KASSERT((oldl1 & ATTR_SW_MANAGED) == 0, ("pmap_demote_l1: Level 1 table shouldn't be managed")); tmpl1 = 0; if (va <= (vm_offset_t)l1 && va + L1_SIZE > (vm_offset_t)l1) { tmpl1 = kva_alloc(PAGE_SIZE); if (tmpl1 == 0) return (NULL); } if ((ml2 = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { CTR2(KTR_PMAP, "pmap_demote_l1: failure for va %#lx" " in pmap %p", va, pmap); return (NULL); } l2phys = VM_PAGE_TO_PHYS(ml2); l2 = (pt_entry_t *)PHYS_TO_DMAP(l2phys); /* Address the range points at */ phys = oldl1 & ~ATTR_MASK; /* The attributed from the old l1 table to be copied */ newl2 = oldl1 & ATTR_MASK; /* Create the new entries */ for (i = 0; i < Ln_ENTRIES; i++) { l2[i] = newl2 | phys; phys += L2_SIZE; } cpu_dcache_wb_range((vm_offset_t)l2, PAGE_SIZE); KASSERT(l2[0] == ((oldl1 & ~ATTR_DESCR_MASK) | L2_BLOCK), ("Invalid l2 page (%lx != %lx)", l2[0], (oldl1 & ~ATTR_DESCR_MASK) | L2_BLOCK)); if (tmpl1 != 0) { pmap_kenter(tmpl1, PAGE_SIZE, DMAP_TO_PHYS((vm_offset_t)l1) & ~L3_OFFSET, CACHED_MEMORY); l1 = (pt_entry_t *)(tmpl1 + ((vm_offset_t)l1 & PAGE_MASK)); } pmap_update_entry(pmap, l1, l2phys | L1_TABLE, va, PAGE_SIZE); if (tmpl1 != 0) { pmap_kremove(tmpl1); kva_free(tmpl1, PAGE_SIZE); } return (l2); } /* * Create an L3 table to map all addresses within an L2 mapping. */ static pt_entry_t * pmap_demote_l2_locked(pmap_t pmap, pt_entry_t *l2, vm_offset_t va, struct rwlock **lockp) { pt_entry_t *l3, newl3, oldl2; vm_offset_t tmpl2; vm_paddr_t l3phys, phys; vm_page_t ml3; int i; PMAP_LOCK_ASSERT(pmap, MA_OWNED); l3 = NULL; oldl2 = pmap_load(l2); KASSERT((oldl2 & ATTR_DESCR_MASK) == L2_BLOCK, ("pmap_demote_l2: Demoting a non-block entry")); KASSERT((va & L2_OFFSET) == 0, ("pmap_demote_l2: Invalid virtual address %#lx", va)); tmpl2 = 0; if (va <= (vm_offset_t)l2 && va + L2_SIZE > (vm_offset_t)l2) { tmpl2 = kva_alloc(PAGE_SIZE); if (tmpl2 == 0) return (NULL); } if ((ml3 = pmap_lookup_pt_page(pmap, va)) != NULL) { pmap_remove_pt_page(pmap, ml3); } else { ml3 = vm_page_alloc(NULL, pmap_l2_pindex(va), (VIRT_IN_DMAP(va) ? VM_ALLOC_INTERRUPT : VM_ALLOC_NORMAL) | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); if (ml3 == NULL) { CTR2(KTR_PMAP, "pmap_demote_l2: failure for va %#lx" " in pmap %p", va, pmap); goto fail; } if (va < VM_MAXUSER_ADDRESS) pmap_resident_count_inc(pmap, 1); } l3phys = VM_PAGE_TO_PHYS(ml3); l3 = (pt_entry_t *)PHYS_TO_DMAP(l3phys); /* Address the range points at */ phys = oldl2 & ~ATTR_MASK; /* The attributed from the old l2 table to be copied */ newl3 = (oldl2 & (ATTR_MASK & ~ATTR_DESCR_MASK)) | L3_PAGE; /* * If the page table page is new, initialize it. */ if (ml3->wire_count == 1) { for (i = 0; i < Ln_ENTRIES; i++) { l3[i] = newl3 | phys; phys += L3_SIZE; } cpu_dcache_wb_range((vm_offset_t)l3, PAGE_SIZE); } KASSERT(l3[0] == ((oldl2 & ~ATTR_DESCR_MASK) | L3_PAGE), ("Invalid l3 page (%lx != %lx)", l3[0], (oldl2 & ~ATTR_DESCR_MASK) | L3_PAGE)); /* * Map the temporary page so we don't lose access to the l2 table. */ if (tmpl2 != 0) { pmap_kenter(tmpl2, PAGE_SIZE, DMAP_TO_PHYS((vm_offset_t)l2) & ~L3_OFFSET, CACHED_MEMORY); l2 = (pt_entry_t *)(tmpl2 + ((vm_offset_t)l2 & PAGE_MASK)); } /* * The spare PV entries must be reserved prior to demoting the * mapping, that is, prior to changing the PDE. Otherwise, the state * of the L2 and the PV lists will be inconsistent, which can result * in reclaim_pv_chunk() attempting to remove a PV entry from the * wrong PV list and pmap_pv_demote_l2() failing to find the expected * PV entry for the 2MB page mapping that is being demoted. */ if ((oldl2 & ATTR_SW_MANAGED) != 0) reserve_pv_entries(pmap, Ln_ENTRIES - 1, lockp); pmap_update_entry(pmap, l2, l3phys | L2_TABLE, va, PAGE_SIZE); /* * Demote the PV entry. */ if ((oldl2 & ATTR_SW_MANAGED) != 0) pmap_pv_demote_l2(pmap, va, oldl2 & ~ATTR_MASK, lockp); atomic_add_long(&pmap_l2_demotions, 1); CTR3(KTR_PMAP, "pmap_demote_l2: success for va %#lx" " in pmap %p %lx", va, pmap, l3[0]); fail: if (tmpl2 != 0) { pmap_kremove(tmpl2); kva_free(tmpl2, PAGE_SIZE); } return (l3); } static pt_entry_t * pmap_demote_l2(pmap_t pmap, pt_entry_t *l2, vm_offset_t va) { struct rwlock *lock; pt_entry_t *l3; lock = NULL; l3 = pmap_demote_l2_locked(pmap, l2, va, &lock); if (lock != NULL) rw_wunlock(lock); return (l3); } /* * perform the pmap work for mincore */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa) { pd_entry_t *l1p, l1; pd_entry_t *l2p, l2; pt_entry_t *l3p, l3; vm_paddr_t pa; bool managed; int val; PMAP_LOCK(pmap); retry: pa = 0; val = 0; managed = false; l1p = pmap_l1(pmap, addr); if (l1p == NULL) /* No l1 */ goto done; l1 = pmap_load(l1p); if ((l1 & ATTR_DESCR_MASK) == L1_INVAL) goto done; if ((l1 & ATTR_DESCR_MASK) == L1_BLOCK) { pa = (l1 & ~ATTR_MASK) | (addr & L1_OFFSET); managed = (l1 & ATTR_SW_MANAGED) == ATTR_SW_MANAGED; val = MINCORE_SUPER | MINCORE_INCORE; if (pmap_page_dirty(l1)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((l1 & ATTR_AF) == ATTR_AF) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; goto done; } l2p = pmap_l1_to_l2(l1p, addr); if (l2p == NULL) /* No l2 */ goto done; l2 = pmap_load(l2p); if ((l2 & ATTR_DESCR_MASK) == L2_INVAL) goto done; if ((l2 & ATTR_DESCR_MASK) == L2_BLOCK) { pa = (l2 & ~ATTR_MASK) | (addr & L2_OFFSET); managed = (l2 & ATTR_SW_MANAGED) == ATTR_SW_MANAGED; val = MINCORE_SUPER | MINCORE_INCORE; if (pmap_page_dirty(l2)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((l2 & ATTR_AF) == ATTR_AF) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; goto done; } l3p = pmap_l2_to_l3(l2p, addr); if (l3p == NULL) /* No l3 */ goto done; l3 = pmap_load(l2p); if ((l3 & ATTR_DESCR_MASK) == L3_INVAL) goto done; if ((l3 & ATTR_DESCR_MASK) == L3_PAGE) { pa = (l3 & ~ATTR_MASK) | (addr & L3_OFFSET); managed = (l3 & ATTR_SW_MANAGED) == ATTR_SW_MANAGED; val = MINCORE_INCORE; if (pmap_page_dirty(l3)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((l3 & ATTR_AF) == ATTR_AF) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } done: if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && managed) { /* Ensure that "PHYS_TO_VM_PAGE(pa)->object" doesn't change. */ if (vm_page_pa_tryrelock(pmap, pa, locked_pa)) goto retry; } else PA_UNLOCK_COND(*locked_pa); PMAP_UNLOCK(pmap); return (val); } void pmap_activate(struct thread *td) { pmap_t pmap; critical_enter(); pmap = vmspace_pmap(td->td_proc->p_vmspace); td->td_pcb->pcb_l0addr = vtophys(pmap->pm_l0); __asm __volatile("msr ttbr0_el1, %0" : : "r"(td->td_pcb->pcb_l0addr)); pmap_invalidate_all(pmap); critical_exit(); } void pmap_sync_icache(pmap_t pmap, vm_offset_t va, vm_size_t sz) { if (va >= VM_MIN_KERNEL_ADDRESS) { cpu_icache_sync_range(va, sz); } else { u_int len, offset; vm_paddr_t pa; /* Find the length of data in this page to flush */ offset = va & PAGE_MASK; len = imin(PAGE_SIZE - offset, sz); while (sz != 0) { /* Extract the physical address & find it in the DMAP */ pa = pmap_extract(pmap, va); if (pa != 0) cpu_icache_sync_range(PHYS_TO_DMAP(pa), len); /* Move to the next page */ sz -= len; va += len; /* Set the length for the next iteration */ len = imin(PAGE_SIZE, sz); } } } int pmap_fault(pmap_t pmap, uint64_t esr, uint64_t far) { #ifdef SMP uint64_t par; #endif switch (ESR_ELx_EXCEPTION(esr)) { case EXCP_DATA_ABORT_L: case EXCP_DATA_ABORT: break; default: return (KERN_FAILURE); } #ifdef SMP PMAP_LOCK(pmap); switch (esr & ISS_DATA_DFSC_MASK) { case ISS_DATA_DFSC_TF_L0: case ISS_DATA_DFSC_TF_L1: case ISS_DATA_DFSC_TF_L2: case ISS_DATA_DFSC_TF_L3: /* Ask the MMU to check the address */ if (pmap == kernel_pmap) par = arm64_address_translate_s1e1r(far); else par = arm64_address_translate_s1e0r(far); /* * If the translation was successful the address was invalid * due to a break-before-make sequence. We can unlock and * return success to the trap handler. */ if (PAR_SUCCESS(par)) { PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } break; default: break; } PMAP_UNLOCK(pmap); #endif return (KERN_FAILURE); } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < L2_SIZE) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & L2_OFFSET; if (size - ((L2_SIZE - superpage_offset) & L2_OFFSET) < L2_SIZE || (*addr & L2_OFFSET) == superpage_offset) return; if ((*addr & L2_OFFSET) < superpage_offset) *addr = (*addr & ~L2_OFFSET) + superpage_offset; else *addr = ((*addr + L2_OFFSET) & ~L2_OFFSET) + superpage_offset; } /** * Get the kernel virtual address of a set of physical pages. If there are * physical addresses not covered by the DMAP perform a transient mapping * that will be removed when calling pmap_unmap_io_transient. * * \param page The pages the caller wishes to obtain the virtual * address on the kernel memory map. * \param vaddr On return contains the kernel virtual memory address * of the pages passed in the page parameter. * \param count Number of pages passed in. * \param can_fault TRUE if the thread using the mapped pages can take * page faults, FALSE otherwise. * * \returns TRUE if the caller must call pmap_unmap_io_transient when * finished or FALSE otherwise. * */ boolean_t pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; boolean_t needs_mapping; int error, i; /* * Allocate any KVA space that we need, this is done in a separate * loop to prevent calling vmem_alloc while pinned. */ needs_mapping = FALSE; for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (__predict_false(!PHYS_IN_DMAP(paddr))) { error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr[i]); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); needs_mapping = TRUE; } else { vaddr[i] = PHYS_TO_DMAP(paddr); } } /* Exit early if everything is covered by the DMAP */ if (!needs_mapping) return (FALSE); if (!can_fault) sched_pin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (!PHYS_IN_DMAP(paddr)) { panic( "pmap_map_io_transient: TODO: Map out of DMAP data"); } } return (needs_mapping); } void pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; int i; if (!can_fault) sched_unpin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (!PHYS_IN_DMAP(paddr)) { panic("ARM64TODO: pmap_unmap_io_transient: Unmap data"); } } } Index: projects/mips64-clang/sys/cddl/compat/opensolaris/sys/random.h =================================================================== --- projects/mips64-clang/sys/cddl/compat/opensolaris/sys/random.h (revision 305711) +++ projects/mips64-clang/sys/cddl/compat/opensolaris/sys/random.h (revision 305712) @@ -1,37 +1,37 @@ /*- * Copyright (c) 2007 Pawel Jakub Dawidek * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _OPENSOLARIS_SYS_RANDOM_H_ #define _OPENSOLARIS_SYS_RANDOM_H_ #include_next #define random_get_bytes(p, s) read_random((p), (int)(s)) -#define random_get_pseudo_bytes(p, s) read_random((p), (int)(s)) +#define random_get_pseudo_bytes(p, s) arc4rand((p), (int)(s), 0) #endif /* !_OPENSOLARIS_SYS_RANDOM_H_ */ Index: projects/mips64-clang/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/lz4.c =================================================================== --- projects/mips64-clang/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/lz4.c (revision 305711) +++ projects/mips64-clang/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/lz4.c (revision 305712) @@ -1,1025 +1,1025 @@ /* * LZ4 - Fast LZ compression algorithm * Header File * Copyright (C) 2011-2013, Yann Collet. * BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php) * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are * met: * * * Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * * Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following disclaimer * in the documentation and/or other materials provided with the * distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * You can contact the author at : * - LZ4 homepage : http://fastcompression.blogspot.com/p/lz4.html * - LZ4 source repository : http://code.google.com/p/lz4/ */ #include static int real_LZ4_compress(const char *source, char *dest, int isize, int osize); static int LZ4_compressBound(int isize); static int LZ4_uncompress_unknownOutputSize(const char *source, char *dest, int isize, int maxOutputSize); static int LZ4_compressCtx(void *ctx, const char *source, char *dest, int isize, int osize); static int LZ4_compress64kCtx(void *ctx, const char *source, char *dest, int isize, int osize); static kmem_cache_t *lz4_ctx_cache; /*ARGSUSED*/ size_t lz4_compress(void *s_start, void *d_start, size_t s_len, size_t d_len, int n) { uint32_t bufsiz; char *dest = d_start; ASSERT(d_len >= sizeof (bufsiz)); bufsiz = real_LZ4_compress(s_start, &dest[sizeof (bufsiz)], s_len, d_len - sizeof (bufsiz)); /* Signal an error if the compression routine returned zero. */ if (bufsiz == 0) return (s_len); /* * Encode the compresed buffer size at the start. We'll need this in * decompression to counter the effects of padding which might be * added to the compressed buffer and which, if unhandled, would * confuse the hell out of our decompression function. */ *(uint32_t *)dest = BE_32(bufsiz); return (bufsiz + sizeof (bufsiz)); } /*ARGSUSED*/ int lz4_decompress(void *s_start, void *d_start, size_t s_len, size_t d_len, int n) { const char *src = s_start; uint32_t bufsiz = BE_IN32(src); /* invalid compressed buffer size encoded at start */ if (bufsiz + sizeof (bufsiz) > s_len) return (1); /* * Returns 0 on success (decompression function returned non-negative) * and non-zero on failure (decompression function returned negative. */ return (LZ4_uncompress_unknownOutputSize(&src[sizeof (bufsiz)], d_start, bufsiz, d_len) < 0); } /* * LZ4 API Description: * * Simple Functions: * real_LZ4_compress() : * isize : is the input size. Max supported value is ~1.9GB * return : the number of bytes written in buffer dest * or 0 if the compression fails (if LZ4_COMPRESSMIN is set). * note : destination buffer must be already allocated. * destination buffer must be sized to handle worst cases * situations (input data not compressible) worst case size * evaluation is provided by function LZ4_compressBound(). * * Advanced Functions * * LZ4_compressBound() : * Provides the maximum size that LZ4 may output in a "worst case" * scenario (input data not compressible) primarily useful for memory * allocation of output buffer. * * isize : is the input size. Max supported value is ~1.9GB * return : maximum output size in a "worst case" scenario * note : this function is limited by "int" range (2^31-1) * * LZ4_uncompress_unknownOutputSize() : * isize : is the input size, therefore the compressed size * maxOutputSize : is the size of the destination buffer (which must be * already allocated) * return : the number of bytes decoded in the destination buffer * (necessarily <= maxOutputSize). If the source stream is * malformed, the function will stop decoding and return a * negative result, indicating the byte position of the faulty * instruction. This function never writes beyond dest + * maxOutputSize, and is therefore protected against malicious * data packets. * note : Destination buffer must be already allocated. * * LZ4_compressCtx() : * This function explicitly handles the CTX memory structure. * * ILLUMOS CHANGES: the CTX memory structure must be explicitly allocated * by the caller (either on the stack or using kmem_zalloc). Passing NULL * isn't valid. * * LZ4_compress64kCtx() : * Same as LZ4_compressCtx(), but specific to small inputs (<64KB). * isize *Must* be <64KB, otherwise the output will be corrupted. * * ILLUMOS CHANGES: the CTX memory structure must be explicitly allocated * by the caller (either on the stack or using kmem_zalloc). Passing NULL * isn't valid. */ /* * Tuning parameters */ /* * COMPRESSIONLEVEL: Increasing this value improves compression ratio * Lowering this value reduces memory usage. Reduced memory usage * typically improves speed, due to cache effect (ex: L1 32KB for Intel, * L1 64KB for AMD). Memory usage formula : N->2^(N+2) Bytes * (examples : 12 -> 16KB ; 17 -> 512KB) */ #define COMPRESSIONLEVEL 12 /* * NOTCOMPRESSIBLE_CONFIRMATION: Decreasing this value will make the * algorithm skip faster data segments considered "incompressible". * This may decrease compression ratio dramatically, but will be * faster on incompressible data. Increasing this value will make * the algorithm search more before declaring a segment "incompressible". * This could improve compression a bit, but will be slower on * incompressible data. The default value (6) is recommended. */ #define NOTCOMPRESSIBLE_CONFIRMATION 6 /* * BIG_ENDIAN_NATIVE_BUT_INCOMPATIBLE: This will provide a boost to * performance for big endian cpu, but the resulting compressed stream * will be incompatible with little-endian CPU. You can set this option * to 1 in situations where data will stay within closed environment. * This option is useless on Little_Endian CPU (such as x86). */ /* #define BIG_ENDIAN_NATIVE_BUT_INCOMPATIBLE 1 */ /* * CPU Feature Detection */ /* 32 or 64 bits ? */ #if (defined(__x86_64__) || defined(__x86_64) || defined(__amd64__) || \ defined(__amd64) || defined(__ppc64__) || defined(_WIN64) || \ defined(__LP64__) || defined(_LP64)) #define LZ4_ARCH64 1 -/* - * Illumos: On amd64 we have 20k of stack and 24k on sun4u and sun4v, so we - * can spend 16k on the algorithm - */ -/* FreeBSD: Use heap for all platforms for now */ -#define STACKLIMIT 0 #else #define LZ4_ARCH64 0 +#endif + /* - * Illumos: On i386 we only have 12k of stack, so in order to maintain the - * same COMPRESSIONLEVEL we have to use heap allocation. Performance will - * suck, but alas, it's ZFS on 32-bit we're talking about, so... + * Limits the amount of stack space that the algorithm may consume to hold + * the compression lookup table. The value `9' here means we'll never use + * more than 2k of stack (see above for a description of COMPRESSIONLEVEL). + * If more memory is needed, it is allocated from the heap. */ +/* FreeBSD: Use heap for all platforms for now */ #define STACKLIMIT 0 -#endif /* * Little Endian or Big Endian? * Note: overwrite the below #define if you know your architecture endianess. */ #if BYTE_ORDER == BIG_ENDIAN #define LZ4_BIG_ENDIAN 1 #else /* * Little Endian assumed. PDP Endian and other very rare endian format * are unsupported. */ #endif /* * Unaligned memory access is automatically enabled for "common" CPU, * such as x86. For others CPU, the compiler will be more cautious, and * insert extra code to ensure aligned access is respected. If you know * your target CPU supports unaligned memory access, you may want to * force this option manually to improve performance */ #if defined(__ARM_FEATURE_UNALIGNED) #define LZ4_FORCE_UNALIGNED_ACCESS 1 #endif /* * FreeBSD: can't use GCC's __builtin_ctz when using sparc64 because * gcc currently rely on libcompiler_rt. * * TODO: revisit this when situation changes. */ #if defined(__sparc64__) #define LZ4_FORCE_SW_BITCOUNT #endif /* * Compiler Options */ #if __STDC_VERSION__ >= 199901L /* C99 */ /* "restrict" is a known keyword */ #else /* Disable restrict */ #define restrict #endif #define lz4_bswap16(x) ((unsigned short int) ((((x) >> 8) & 0xffu) | \ (((x) & 0xffu) << 8))) #define expect(expr, value) (__builtin_expect((expr), (value))) #if defined(likely) #undef likely #endif #if defined(unlikely) #undef unlikely #endif #define likely(expr) expect((expr) != 0, 1) #define unlikely(expr) expect((expr) != 0, 0) /* Basic types */ #define BYTE uint8_t #define U16 uint16_t #define U32 uint32_t #define S32 int32_t #define U64 uint64_t #ifndef LZ4_FORCE_UNALIGNED_ACCESS #pragma pack(1) #endif typedef struct _U16_S { U16 v; } U16_S; typedef struct _U32_S { U32 v; } U32_S; typedef struct _U64_S { U64 v; } U64_S; #ifndef LZ4_FORCE_UNALIGNED_ACCESS #pragma pack() #endif #define A64(x) (((U64_S *)(x))->v) #define A32(x) (((U32_S *)(x))->v) #define A16(x) (((U16_S *)(x))->v) /* * Constants */ #define MINMATCH 4 #define HASH_LOG COMPRESSIONLEVEL #define HASHTABLESIZE (1 << HASH_LOG) #define HASH_MASK (HASHTABLESIZE - 1) #define SKIPSTRENGTH (NOTCOMPRESSIBLE_CONFIRMATION > 2 ? \ NOTCOMPRESSIBLE_CONFIRMATION : 2) /* * Defines if memory is allocated into the stack (local variable), * or into the heap (kmem_alloc()). */ #define HEAPMODE (HASH_LOG > STACKLIMIT) #define COPYLENGTH 8 #define LASTLITERALS 5 #define MFLIMIT (COPYLENGTH + MINMATCH) #define MINLENGTH (MFLIMIT + 1) #define MAXD_LOG 16 #define MAX_DISTANCE ((1 << MAXD_LOG) - 1) #define ML_BITS 4 #define ML_MASK ((1U<> ((MINMATCH * 8) - \ HASH_LOG)) #define LZ4_HASH_VALUE(p) LZ4_HASH_FUNCTION(A32(p)) #define LZ4_WILDCOPY(s, d, e) do { LZ4_COPYPACKET(s, d) } while (d < e); #define LZ4_BLINDCOPY(s, d, l) { BYTE* e = (d) + l; LZ4_WILDCOPY(s, d, e); \ d = e; } /* Private functions */ #if LZ4_ARCH64 static inline int LZ4_NbCommonBytes(register U64 val) { #if defined(LZ4_BIG_ENDIAN) #if !defined(LZ4_FORCE_SW_BITCOUNT) return (__builtin_clzll(val) >> 3); #else int r; if (!(val >> 32)) { r = 4; } else { r = 0; val >>= 32; } if (!(val >> 16)) { r += 2; val >>= 8; } else { val >>= 24; } r += (!val); return (r); #endif #else #if !defined(LZ4_FORCE_SW_BITCOUNT) return (__builtin_ctzll(val) >> 3); #else static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2, 0, 3, 1, 3, 1, 4, 2, 7, 0, 2, 3, 6, 1, 5, 3, 5, 1, 3, 4, 4, 2, 5, 6, 7, 7, 0, 1, 2, 3, 3, 4, 6, 2, 6, 5, 5, 3, 4, 5, 6, 7, 1, 2, 4, 6, 4, 4, 5, 7, 2, 6, 5, 7, 6, 7, 7 }; return DeBruijnBytePos[((U64) ((val & -val) * 0x0218A392CDABBD3F)) >> 58]; #endif #endif } #else static inline int LZ4_NbCommonBytes(register U32 val) { #if defined(LZ4_BIG_ENDIAN) #if !defined(LZ4_FORCE_SW_BITCOUNT) return (__builtin_clz(val) >> 3); #else int r; if (!(val >> 16)) { r = 2; val >>= 8; } else { r = 0; val >>= 24; } r += (!val); return (r); #endif #else #if !defined(LZ4_FORCE_SW_BITCOUNT) return (__builtin_ctz(val) >> 3); #else static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0, 3, 2, 2, 1, 3, 2, 0, 1, 3, 3, 1, 2, 2, 2, 2, 0, 3, 1, 2, 0, 1, 0, 1, 1 }; return DeBruijnBytePos[((U32) ((val & -(S32) val) * 0x077CB531U)) >> 27]; #endif #endif } #endif /* Public functions */ static int LZ4_compressBound(int isize) { return (isize + (isize / 255) + 16); } /* Compression functions */ /*ARGSUSED*/ static int LZ4_compressCtx(void *ctx, const char *source, char *dest, int isize, int osize) { #if HEAPMODE struct refTables *srt = (struct refTables *)ctx; HTYPE *HashTable = (HTYPE *) (srt->hashTable); #else HTYPE HashTable[HASHTABLESIZE] = { 0 }; #endif const BYTE *ip = (BYTE *) source; INITBASE(base); const BYTE *anchor = ip; const BYTE *const iend = ip + isize; const BYTE *const oend = (BYTE *) dest + osize; const BYTE *const mflimit = iend - MFLIMIT; #define matchlimit (iend - LASTLITERALS) BYTE *op = (BYTE *) dest; int len, length; const int skipStrength = SKIPSTRENGTH; U32 forwardH; /* Init */ if (isize < MINLENGTH) goto _last_literals; /* First Byte */ HashTable[LZ4_HASH_VALUE(ip)] = ip - base; ip++; forwardH = LZ4_HASH_VALUE(ip); /* Main Loop */ for (;;) { int findMatchAttempts = (1U << skipStrength) + 3; const BYTE *forwardIp = ip; const BYTE *ref; BYTE *token; /* Find a match */ do { U32 h = forwardH; int step = findMatchAttempts++ >> skipStrength; ip = forwardIp; forwardIp = ip + step; if unlikely(forwardIp > mflimit) { goto _last_literals; } forwardH = LZ4_HASH_VALUE(forwardIp); ref = base + HashTable[h]; HashTable[h] = ip - base; } while ((ref < ip - MAX_DISTANCE) || (A32(ref) != A32(ip))); /* Catch up */ while ((ip > anchor) && (ref > (BYTE *) source) && unlikely(ip[-1] == ref[-1])) { ip--; ref--; } /* Encode Literal length */ length = ip - anchor; token = op++; /* Check output limit */ if unlikely(op + length + (2 + 1 + LASTLITERALS) + (length >> 8) > oend) return (0); if (length >= (int)RUN_MASK) { *token = (RUN_MASK << ML_BITS); len = length - RUN_MASK; for (; len > 254; len -= 255) *op++ = 255; *op++ = (BYTE)len; } else *token = (length << ML_BITS); /* Copy Literals */ LZ4_BLINDCOPY(anchor, op, length); _next_match: /* Encode Offset */ LZ4_WRITE_LITTLEENDIAN_16(op, ip - ref); /* Start Counting */ ip += MINMATCH; ref += MINMATCH; /* MinMatch verified */ anchor = ip; while likely(ip < matchlimit - (STEPSIZE - 1)) { UARCH diff = AARCH(ref) ^ AARCH(ip); if (!diff) { ip += STEPSIZE; ref += STEPSIZE; continue; } ip += LZ4_NbCommonBytes(diff); goto _endCount; } #if LZ4_ARCH64 if ((ip < (matchlimit - 3)) && (A32(ref) == A32(ip))) { ip += 4; ref += 4; } #endif if ((ip < (matchlimit - 1)) && (A16(ref) == A16(ip))) { ip += 2; ref += 2; } if ((ip < matchlimit) && (*ref == *ip)) ip++; _endCount: /* Encode MatchLength */ len = (ip - anchor); /* Check output limit */ if unlikely(op + (1 + LASTLITERALS) + (len >> 8) > oend) return (0); if (len >= (int)ML_MASK) { *token += ML_MASK; len -= ML_MASK; for (; len > 509; len -= 510) { *op++ = 255; *op++ = 255; } if (len > 254) { len -= 255; *op++ = 255; } *op++ = (BYTE)len; } else *token += len; /* Test end of chunk */ if (ip > mflimit) { anchor = ip; break; } /* Fill table */ HashTable[LZ4_HASH_VALUE(ip - 2)] = ip - 2 - base; /* Test next position */ ref = base + HashTable[LZ4_HASH_VALUE(ip)]; HashTable[LZ4_HASH_VALUE(ip)] = ip - base; if ((ref > ip - (MAX_DISTANCE + 1)) && (A32(ref) == A32(ip))) { token = op++; *token = 0; goto _next_match; } /* Prepare next loop */ anchor = ip++; forwardH = LZ4_HASH_VALUE(ip); } _last_literals: /* Encode Last Literals */ { int lastRun = iend - anchor; if (op + lastRun + 1 + ((lastRun + 255 - RUN_MASK) / 255) > oend) return (0); if (lastRun >= (int)RUN_MASK) { *op++ = (RUN_MASK << ML_BITS); lastRun -= RUN_MASK; for (; lastRun > 254; lastRun -= 255) { *op++ = 255; } *op++ = (BYTE)lastRun; } else *op++ = (lastRun << ML_BITS); (void) memcpy(op, anchor, iend - anchor); op += iend - anchor; } /* End */ return (int)(((char *)op) - dest); } /* Note : this function is valid only if isize < LZ4_64KLIMIT */ #define LZ4_64KLIMIT ((1 << 16) + (MFLIMIT - 1)) #define HASHLOG64K (HASH_LOG + 1) #define HASH64KTABLESIZE (1U << HASHLOG64K) #define LZ4_HASH64K_FUNCTION(i) (((i) * 2654435761U) >> ((MINMATCH*8) - \ HASHLOG64K)) #define LZ4_HASH64K_VALUE(p) LZ4_HASH64K_FUNCTION(A32(p)) /*ARGSUSED*/ static int LZ4_compress64kCtx(void *ctx, const char *source, char *dest, int isize, int osize) { #if HEAPMODE struct refTables *srt = (struct refTables *)ctx; U16 *HashTable = (U16 *) (srt->hashTable); #else U16 HashTable[HASH64KTABLESIZE] = { 0 }; #endif const BYTE *ip = (BYTE *) source; const BYTE *anchor = ip; const BYTE *const base = ip; const BYTE *const iend = ip + isize; const BYTE *const oend = (BYTE *) dest + osize; const BYTE *const mflimit = iend - MFLIMIT; #define matchlimit (iend - LASTLITERALS) BYTE *op = (BYTE *) dest; int len, length; const int skipStrength = SKIPSTRENGTH; U32 forwardH; /* Init */ if (isize < MINLENGTH) goto _last_literals; /* First Byte */ ip++; forwardH = LZ4_HASH64K_VALUE(ip); /* Main Loop */ for (;;) { int findMatchAttempts = (1U << skipStrength) + 3; const BYTE *forwardIp = ip; const BYTE *ref; BYTE *token; /* Find a match */ do { U32 h = forwardH; int step = findMatchAttempts++ >> skipStrength; ip = forwardIp; forwardIp = ip + step; if (forwardIp > mflimit) { goto _last_literals; } forwardH = LZ4_HASH64K_VALUE(forwardIp); ref = base + HashTable[h]; HashTable[h] = ip - base; } while (A32(ref) != A32(ip)); /* Catch up */ while ((ip > anchor) && (ref > (BYTE *) source) && (ip[-1] == ref[-1])) { ip--; ref--; } /* Encode Literal length */ length = ip - anchor; token = op++; /* Check output limit */ if unlikely(op + length + (2 + 1 + LASTLITERALS) + (length >> 8) > oend) return (0); if (length >= (int)RUN_MASK) { *token = (RUN_MASK << ML_BITS); len = length - RUN_MASK; for (; len > 254; len -= 255) *op++ = 255; *op++ = (BYTE)len; } else *token = (length << ML_BITS); /* Copy Literals */ LZ4_BLINDCOPY(anchor, op, length); _next_match: /* Encode Offset */ LZ4_WRITE_LITTLEENDIAN_16(op, ip - ref); /* Start Counting */ ip += MINMATCH; ref += MINMATCH; /* MinMatch verified */ anchor = ip; while (ip < matchlimit - (STEPSIZE - 1)) { UARCH diff = AARCH(ref) ^ AARCH(ip); if (!diff) { ip += STEPSIZE; ref += STEPSIZE; continue; } ip += LZ4_NbCommonBytes(diff); goto _endCount; } #if LZ4_ARCH64 if ((ip < (matchlimit - 3)) && (A32(ref) == A32(ip))) { ip += 4; ref += 4; } #endif if ((ip < (matchlimit - 1)) && (A16(ref) == A16(ip))) { ip += 2; ref += 2; } if ((ip < matchlimit) && (*ref == *ip)) ip++; _endCount: /* Encode MatchLength */ len = (ip - anchor); /* Check output limit */ if unlikely(op + (1 + LASTLITERALS) + (len >> 8) > oend) return (0); if (len >= (int)ML_MASK) { *token += ML_MASK; len -= ML_MASK; for (; len > 509; len -= 510) { *op++ = 255; *op++ = 255; } if (len > 254) { len -= 255; *op++ = 255; } *op++ = (BYTE)len; } else *token += len; /* Test end of chunk */ if (ip > mflimit) { anchor = ip; break; } /* Fill table */ HashTable[LZ4_HASH64K_VALUE(ip - 2)] = ip - 2 - base; /* Test next position */ ref = base + HashTable[LZ4_HASH64K_VALUE(ip)]; HashTable[LZ4_HASH64K_VALUE(ip)] = ip - base; if (A32(ref) == A32(ip)) { token = op++; *token = 0; goto _next_match; } /* Prepare next loop */ anchor = ip++; forwardH = LZ4_HASH64K_VALUE(ip); } _last_literals: /* Encode Last Literals */ { int lastRun = iend - anchor; if (op + lastRun + 1 + ((lastRun + 255 - RUN_MASK) / 255) > oend) return (0); if (lastRun >= (int)RUN_MASK) { *op++ = (RUN_MASK << ML_BITS); lastRun -= RUN_MASK; for (; lastRun > 254; lastRun -= 255) *op++ = 255; *op++ = (BYTE)lastRun; } else *op++ = (lastRun << ML_BITS); (void) memcpy(op, anchor, iend - anchor); op += iend - anchor; } /* End */ return (int)(((char *)op) - dest); } static int real_LZ4_compress(const char *source, char *dest, int isize, int osize) { #if HEAPMODE void *ctx = kmem_cache_alloc(lz4_ctx_cache, KM_NOSLEEP); int result; /* * out of kernel memory, gently fall through - this will disable * compression in zio_compress_data */ if (ctx == NULL) return (0); bzero(ctx, sizeof(struct refTables)); if (isize < LZ4_64KLIMIT) result = LZ4_compress64kCtx(ctx, source, dest, isize, osize); else result = LZ4_compressCtx(ctx, source, dest, isize, osize); kmem_cache_free(lz4_ctx_cache, ctx); return (result); #else if (isize < (int)LZ4_64KLIMIT) return (LZ4_compress64kCtx(NULL, source, dest, isize, osize)); return (LZ4_compressCtx(NULL, source, dest, isize, osize)); #endif } /* Decompression functions */ /* - * Note: The decoding functionLZ4_uncompress_unknownOutputSize() is safe + * Note: The decoding function LZ4_uncompress_unknownOutputSize() is safe * against "buffer overflow" attack type. They will never write nor * read outside of the provided output buffers. * LZ4_uncompress_unknownOutputSize() also insures that it will never * read outside of the input buffer. A corrupted input will produce * an error result, a negative int, indicating the position of the * error within input stream. */ static int LZ4_uncompress_unknownOutputSize(const char *source, char *dest, int isize, int maxOutputSize) { /* Local Variables */ const BYTE *restrict ip = (const BYTE *) source; const BYTE *const iend = ip + isize; const BYTE *ref; BYTE *op = (BYTE *) dest; BYTE *const oend = op + maxOutputSize; BYTE *cpy; size_t dec32table[] = {0, 3, 2, 3, 0, 0, 0, 0}; #if LZ4_ARCH64 size_t dec64table[] = {0, 0, 0, (size_t)-1, 0, 1, 2, 3}; #endif /* Main Loop */ while (ip < iend) { unsigned token; size_t length; /* get runlength */ token = *ip++; if ((length = (token >> ML_BITS)) == RUN_MASK) { int s = 255; while ((ip < iend) && (s == 255)) { s = *ip++; length += s; } } /* copy literals */ cpy = op + length; + /* CORNER-CASE: cpy might overflow. */ + if (cpy < op) + goto _output_error; /* cpy was overflowed, bail! */ if ((cpy > oend - COPYLENGTH) || (ip + length > iend - COPYLENGTH)) { if (cpy > oend) /* Error: writes beyond output buffer */ goto _output_error; if (ip + length != iend) /* * Error: LZ4 format requires to consume all * input at this stage */ goto _output_error; (void) memcpy(op, ip, length); op += length; /* Necessarily EOF, due to parsing restrictions */ break; } LZ4_WILDCOPY(ip, op, cpy); ip -= (op - cpy); op = cpy; /* get offset */ LZ4_READ_LITTLEENDIAN_16(ref, cpy, ip); ip += 2; if (ref < (BYTE * const) dest) /* * Error: offset creates reference outside of * destination buffer */ goto _output_error; /* get matchlength */ if ((length = (token & ML_MASK)) == ML_MASK) { while (ip < iend) { int s = *ip++; length += s; if (s == 255) continue; break; } } /* copy repeated sequence */ if unlikely(op - ref < STEPSIZE) { #if LZ4_ARCH64 size_t dec64 = dec64table[op-ref]; #else const int dec64 = 0; #endif op[0] = ref[0]; op[1] = ref[1]; op[2] = ref[2]; op[3] = ref[3]; op += 4; ref += 4; ref -= dec32table[op-ref]; A32(op) = A32(ref); op += STEPSIZE - 4; ref -= dec64; } else { LZ4_COPYSTEP(ref, op); } cpy = op + length - (STEPSIZE - 4); if (cpy > oend - COPYLENGTH) { if (cpy > oend) /* * Error: request to write outside of * destination buffer */ goto _output_error; LZ4_SECURECOPY(ref, op, (oend - COPYLENGTH)); while (op < cpy) *op++ = *ref++; op = cpy; if (op == oend) /* * Check EOF (should never happen, since * last 5 bytes are supposed to be literals) */ goto _output_error; continue; } LZ4_SECURECOPY(ref, op, cpy); op = cpy; /* correction */ } /* end of decoding */ return (int)(((char *)op) - dest); /* write overflow error detected */ _output_error: return (int)(-(((char *)ip) - source)); } extern void lz4_init(void) { #if HEAPMODE lz4_ctx_cache = kmem_cache_create("lz4_ctx", sizeof(struct refTables), 0, NULL, NULL, NULL, NULL, NULL, 0); #endif } extern void lz4_fini(void) { #if HEAPMODE kmem_cache_destroy(lz4_ctx_cache); #endif } Index: projects/mips64-clang/sys/cddl/contrib/opensolaris =================================================================== --- projects/mips64-clang/sys/cddl/contrib/opensolaris (revision 305711) +++ projects/mips64-clang/sys/cddl/contrib/opensolaris (revision 305712) Property changes on: projects/mips64-clang/sys/cddl/contrib/opensolaris ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head/sys/cddl/contrib/opensolaris:r305667-305705 Index: projects/mips64-clang/sys/dev/cxgbe/common/t4_hw.c =================================================================== --- projects/mips64-clang/sys/dev/cxgbe/common/t4_hw.c (revision 305711) +++ projects/mips64-clang/sys/dev/cxgbe/common/t4_hw.c (revision 305712) @@ -1,9382 +1,9390 @@ /*- * Copyright (c) 2012, 2016 Chelsio Communications, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include #include #include "common.h" #include "t4_regs.h" #include "t4_regs_values.h" #include "firmware/t4fw_interface.h" #undef msleep #define msleep(x) do { \ if (cold) \ DELAY((x) * 1000); \ else \ pause("t4hw", (x) * hz / 1000); \ } while (0) /** * t4_wait_op_done_val - wait until an operation is completed * @adapter: the adapter performing the operation * @reg: the register to check for completion * @mask: a single-bit field within @reg that indicates completion * @polarity: the value of the field when the operation is completed * @attempts: number of check iterations * @delay: delay in usecs between iterations * @valp: where to store the value of the register at completion time * * Wait until an operation is completed by checking a bit in a register * up to @attempts times. If @valp is not NULL the value of the register * at the time it indicated completion is stored there. Returns 0 if the * operation completes and -EAGAIN otherwise. */ static int t4_wait_op_done_val(struct adapter *adapter, int reg, u32 mask, int polarity, int attempts, int delay, u32 *valp) { while (1) { u32 val = t4_read_reg(adapter, reg); if (!!(val & mask) == polarity) { if (valp) *valp = val; return 0; } if (--attempts == 0) return -EAGAIN; if (delay) udelay(delay); } } static inline int t4_wait_op_done(struct adapter *adapter, int reg, u32 mask, int polarity, int attempts, int delay) { return t4_wait_op_done_val(adapter, reg, mask, polarity, attempts, delay, NULL); } /** * t4_set_reg_field - set a register field to a value * @adapter: the adapter to program * @addr: the register address * @mask: specifies the portion of the register to modify * @val: the new value for the register field * * Sets a register field specified by the supplied mask to the * given value. */ void t4_set_reg_field(struct adapter *adapter, unsigned int addr, u32 mask, u32 val) { u32 v = t4_read_reg(adapter, addr) & ~mask; t4_write_reg(adapter, addr, v | val); (void) t4_read_reg(adapter, addr); /* flush */ } /** * t4_read_indirect - read indirectly addressed registers * @adap: the adapter * @addr_reg: register holding the indirect address * @data_reg: register holding the value of the indirect register * @vals: where the read register values are stored * @nregs: how many indirect registers to read * @start_idx: index of first indirect register to read * * Reads registers that are accessed indirectly through an address/data * register pair. */ void t4_read_indirect(struct adapter *adap, unsigned int addr_reg, unsigned int data_reg, u32 *vals, unsigned int nregs, unsigned int start_idx) { while (nregs--) { t4_write_reg(adap, addr_reg, start_idx); *vals++ = t4_read_reg(adap, data_reg); start_idx++; } } /** * t4_write_indirect - write indirectly addressed registers * @adap: the adapter * @addr_reg: register holding the indirect addresses * @data_reg: register holding the value for the indirect registers * @vals: values to write * @nregs: how many indirect registers to write * @start_idx: address of first indirect register to write * * Writes a sequential block of registers that are accessed indirectly * through an address/data register pair. */ void t4_write_indirect(struct adapter *adap, unsigned int addr_reg, unsigned int data_reg, const u32 *vals, unsigned int nregs, unsigned int start_idx) { while (nregs--) { t4_write_reg(adap, addr_reg, start_idx++); t4_write_reg(adap, data_reg, *vals++); } } /* * Read a 32-bit PCI Configuration Space register via the PCI-E backdoor * mechanism. This guarantees that we get the real value even if we're * operating within a Virtual Machine and the Hypervisor is trapping our * Configuration Space accesses. * * N.B. This routine should only be used as a last resort: the firmware uses * the backdoor registers on a regular basis and we can end up * conflicting with it's uses! */ u32 t4_hw_pci_read_cfg4(adapter_t *adap, int reg) { u32 req = V_FUNCTION(adap->pf) | V_REGISTER(reg); u32 val; if (chip_id(adap) <= CHELSIO_T5) req |= F_ENABLE; else req |= F_T6_ENABLE; if (is_t4(adap)) req |= F_LOCALCFG; t4_write_reg(adap, A_PCIE_CFG_SPACE_REQ, req); val = t4_read_reg(adap, A_PCIE_CFG_SPACE_DATA); /* * Reset F_ENABLE to 0 so reads of PCIE_CFG_SPACE_DATA won't cause a * Configuration Space read. (None of the other fields matter when * F_ENABLE is 0 so a simple register write is easier than a * read-modify-write via t4_set_reg_field().) */ t4_write_reg(adap, A_PCIE_CFG_SPACE_REQ, 0); return val; } /* * t4_report_fw_error - report firmware error * @adap: the adapter * * The adapter firmware can indicate error conditions to the host. * If the firmware has indicated an error, print out the reason for * the firmware error. */ static void t4_report_fw_error(struct adapter *adap) { static const char *const reason[] = { "Crash", /* PCIE_FW_EVAL_CRASH */ "During Device Preparation", /* PCIE_FW_EVAL_PREP */ "During Device Configuration", /* PCIE_FW_EVAL_CONF */ "During Device Initialization", /* PCIE_FW_EVAL_INIT */ "Unexpected Event", /* PCIE_FW_EVAL_UNEXPECTEDEVENT */ "Insufficient Airflow", /* PCIE_FW_EVAL_OVERHEAT */ "Device Shutdown", /* PCIE_FW_EVAL_DEVICESHUTDOWN */ "Reserved", /* reserved */ }; u32 pcie_fw; pcie_fw = t4_read_reg(adap, A_PCIE_FW); if (pcie_fw & F_PCIE_FW_ERR) CH_ERR(adap, "Firmware reports adapter error: %s\n", reason[G_PCIE_FW_EVAL(pcie_fw)]); } /* * Get the reply to a mailbox command and store it in @rpl in big-endian order. */ static void get_mbox_rpl(struct adapter *adap, __be64 *rpl, int nflit, u32 mbox_addr) { for ( ; nflit; nflit--, mbox_addr += 8) *rpl++ = cpu_to_be64(t4_read_reg64(adap, mbox_addr)); } /* * Handle a FW assertion reported in a mailbox. */ static void fw_asrt(struct adapter *adap, struct fw_debug_cmd *asrt) { CH_ALERT(adap, "FW assertion at %.16s:%u, val0 %#x, val1 %#x\n", asrt->u.assert.filename_0_7, be32_to_cpu(asrt->u.assert.line), be32_to_cpu(asrt->u.assert.x), be32_to_cpu(asrt->u.assert.y)); } #define X_CIM_PF_NOACCESS 0xeeeeeeee /** * t4_wr_mbox_meat_timeout - send a command to FW through the given mailbox * @adap: the adapter * @mbox: index of the mailbox to use * @cmd: the command to write * @size: command length in bytes * @rpl: where to optionally store the reply * @sleep_ok: if true we may sleep while awaiting command completion * @timeout: time to wait for command to finish before timing out * (negative implies @sleep_ok=false) * * Sends the given command to FW through the selected mailbox and waits * for the FW to execute the command. If @rpl is not %NULL it is used to * store the FW's reply to the command. The command and its optional * reply are of the same length. Some FW commands like RESET and * INITIALIZE can take a considerable amount of time to execute. * @sleep_ok determines whether we may sleep while awaiting the response. * If sleeping is allowed we use progressive backoff otherwise we spin. * Note that passing in a negative @timeout is an alternate mechanism * for specifying @sleep_ok=false. This is useful when a higher level * interface allows for specification of @timeout but not @sleep_ok ... * * The return value is 0 on success or a negative errno on failure. A * failure can happen either because we are not able to execute the * command or FW executes it but signals an error. In the latter case * the return value is the error code indicated by FW (negated). */ int t4_wr_mbox_meat_timeout(struct adapter *adap, int mbox, const void *cmd, int size, void *rpl, bool sleep_ok, int timeout) { /* * We delay in small increments at first in an effort to maintain * responsiveness for simple, fast executing commands but then back * off to larger delays to a maximum retry delay. */ static const int delay[] = { 1, 1, 3, 5, 10, 10, 20, 50, 100 }; u32 v; u64 res; int i, ms, delay_idx, ret; const __be64 *p = cmd; u32 data_reg = PF_REG(mbox, A_CIM_PF_MAILBOX_DATA); u32 ctl_reg = PF_REG(mbox, A_CIM_PF_MAILBOX_CTRL); u32 ctl; __be64 cmd_rpl[MBOX_LEN/8]; u32 pcie_fw; if ((size & 15) || size > MBOX_LEN) return -EINVAL; if (adap->flags & IS_VF) { if (is_t6(adap)) data_reg = FW_T6VF_MBDATA_BASE_ADDR; else data_reg = FW_T4VF_MBDATA_BASE_ADDR; ctl_reg = VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_CTRL); } /* * If we have a negative timeout, that implies that we can't sleep. */ if (timeout < 0) { sleep_ok = false; timeout = -timeout; } /* * Attempt to gain access to the mailbox. */ for (i = 0; i < 4; i++) { ctl = t4_read_reg(adap, ctl_reg); v = G_MBOWNER(ctl); if (v != X_MBOWNER_NONE) break; } /* * If we were unable to gain access, dequeue ourselves from the * mailbox atomic access list and report the error to our caller. */ if (v != X_MBOWNER_PL) { t4_report_fw_error(adap); ret = (v == X_MBOWNER_FW) ? -EBUSY : -ETIMEDOUT; return ret; } /* * If we gain ownership of the mailbox and there's a "valid" message * in it, this is likely an asynchronous error message from the * firmware. So we'll report that and then proceed on with attempting * to issue our own command ... which may well fail if the error * presaged the firmware crashing ... */ if (ctl & F_MBMSGVALID) { CH_ERR(adap, "found VALID command in mbox %u: " "%llx %llx %llx %llx %llx %llx %llx %llx\n", mbox, (unsigned long long)t4_read_reg64(adap, data_reg), (unsigned long long)t4_read_reg64(adap, data_reg + 8), (unsigned long long)t4_read_reg64(adap, data_reg + 16), (unsigned long long)t4_read_reg64(adap, data_reg + 24), (unsigned long long)t4_read_reg64(adap, data_reg + 32), (unsigned long long)t4_read_reg64(adap, data_reg + 40), (unsigned long long)t4_read_reg64(adap, data_reg + 48), (unsigned long long)t4_read_reg64(adap, data_reg + 56)); } /* * Copy in the new mailbox command and send it on its way ... */ for (i = 0; i < size; i += 8, p++) t4_write_reg64(adap, data_reg + i, be64_to_cpu(*p)); if (adap->flags & IS_VF) { /* * For the VFs, the Mailbox Data "registers" are * actually backed by T4's "MA" interface rather than * PL Registers (as is the case for the PFs). Because * these are in different coherency domains, the write * to the VF's PL-register-backed Mailbox Control can * race in front of the writes to the MA-backed VF * Mailbox Data "registers". So we need to do a * read-back on at least one byte of the VF Mailbox * Data registers before doing the write to the VF * Mailbox Control register. */ t4_read_reg(adap, data_reg); } CH_DUMP_MBOX(adap, mbox, data_reg); t4_write_reg(adap, ctl_reg, F_MBMSGVALID | V_MBOWNER(X_MBOWNER_FW)); t4_read_reg(adap, ctl_reg); /* flush write */ delay_idx = 0; ms = delay[0]; /* * Loop waiting for the reply; bail out if we time out or the firmware * reports an error. */ pcie_fw = 0; for (i = 0; i < timeout; i += ms) { if (!(adap->flags & IS_VF)) { pcie_fw = t4_read_reg(adap, A_PCIE_FW); if (pcie_fw & F_PCIE_FW_ERR) break; } if (sleep_ok) { ms = delay[delay_idx]; /* last element may repeat */ if (delay_idx < ARRAY_SIZE(delay) - 1) delay_idx++; msleep(ms); } else { mdelay(ms); } v = t4_read_reg(adap, ctl_reg); if (v == X_CIM_PF_NOACCESS) continue; if (G_MBOWNER(v) == X_MBOWNER_PL) { if (!(v & F_MBMSGVALID)) { t4_write_reg(adap, ctl_reg, V_MBOWNER(X_MBOWNER_NONE)); continue; } /* * Retrieve the command reply and release the mailbox. */ get_mbox_rpl(adap, cmd_rpl, MBOX_LEN/8, data_reg); t4_write_reg(adap, ctl_reg, V_MBOWNER(X_MBOWNER_NONE)); CH_DUMP_MBOX(adap, mbox, data_reg); res = be64_to_cpu(cmd_rpl[0]); if (G_FW_CMD_OP(res >> 32) == FW_DEBUG_CMD) { fw_asrt(adap, (struct fw_debug_cmd *)cmd_rpl); res = V_FW_CMD_RETVAL(EIO); } else if (rpl) memcpy(rpl, cmd_rpl, size); return -G_FW_CMD_RETVAL((int)res); } } /* * We timed out waiting for a reply to our mailbox command. Report * the error and also check to see if the firmware reported any * errors ... */ ret = (pcie_fw & F_PCIE_FW_ERR) ? -ENXIO : -ETIMEDOUT; CH_ERR(adap, "command %#x in mailbox %d timed out\n", *(const u8 *)cmd, mbox); t4_report_fw_error(adap); t4_fatal_err(adap); return ret; } int t4_wr_mbox_meat(struct adapter *adap, int mbox, const void *cmd, int size, void *rpl, bool sleep_ok) { return t4_wr_mbox_meat_timeout(adap, mbox, cmd, size, rpl, sleep_ok, FW_CMD_MAX_TIMEOUT); } static int t4_edc_err_read(struct adapter *adap, int idx) { u32 edc_ecc_err_addr_reg; u32 edc_bist_status_rdata_reg; if (is_t4(adap)) { CH_WARN(adap, "%s: T4 NOT supported.\n", __func__); return 0; } if (idx != 0 && idx != 1) { CH_WARN(adap, "%s: idx %d NOT supported.\n", __func__, idx); return 0; } edc_ecc_err_addr_reg = EDC_T5_REG(A_EDC_H_ECC_ERR_ADDR, idx); edc_bist_status_rdata_reg = EDC_T5_REG(A_EDC_H_BIST_STATUS_RDATA, idx); CH_WARN(adap, "edc%d err addr 0x%x: 0x%x.\n", idx, edc_ecc_err_addr_reg, t4_read_reg(adap, edc_ecc_err_addr_reg)); CH_WARN(adap, "bist: 0x%x, status %llx %llx %llx %llx %llx %llx %llx %llx %llx.\n", edc_bist_status_rdata_reg, (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 8), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 16), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 24), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 32), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 40), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 48), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 56), (unsigned long long)t4_read_reg64(adap, edc_bist_status_rdata_reg + 64)); return 0; } /** * t4_mc_read - read from MC through backdoor accesses * @adap: the adapter * @idx: which MC to access * @addr: address of first byte requested * @data: 64 bytes of data containing the requested address * @ecc: where to store the corresponding 64-bit ECC word * * Read 64 bytes of data from MC starting at a 64-byte-aligned address * that covers the requested address @addr. If @parity is not %NULL it * is assigned the 64-bit ECC word for the read data. */ int t4_mc_read(struct adapter *adap, int idx, u32 addr, __be32 *data, u64 *ecc) { int i; u32 mc_bist_cmd_reg, mc_bist_cmd_addr_reg, mc_bist_cmd_len_reg; u32 mc_bist_status_rdata_reg, mc_bist_data_pattern_reg; if (is_t4(adap)) { mc_bist_cmd_reg = A_MC_BIST_CMD; mc_bist_cmd_addr_reg = A_MC_BIST_CMD_ADDR; mc_bist_cmd_len_reg = A_MC_BIST_CMD_LEN; mc_bist_status_rdata_reg = A_MC_BIST_STATUS_RDATA; mc_bist_data_pattern_reg = A_MC_BIST_DATA_PATTERN; } else { mc_bist_cmd_reg = MC_REG(A_MC_P_BIST_CMD, idx); mc_bist_cmd_addr_reg = MC_REG(A_MC_P_BIST_CMD_ADDR, idx); mc_bist_cmd_len_reg = MC_REG(A_MC_P_BIST_CMD_LEN, idx); mc_bist_status_rdata_reg = MC_REG(A_MC_P_BIST_STATUS_RDATA, idx); mc_bist_data_pattern_reg = MC_REG(A_MC_P_BIST_DATA_PATTERN, idx); } if (t4_read_reg(adap, mc_bist_cmd_reg) & F_START_BIST) return -EBUSY; t4_write_reg(adap, mc_bist_cmd_addr_reg, addr & ~0x3fU); t4_write_reg(adap, mc_bist_cmd_len_reg, 64); t4_write_reg(adap, mc_bist_data_pattern_reg, 0xc); t4_write_reg(adap, mc_bist_cmd_reg, V_BIST_OPCODE(1) | F_START_BIST | V_BIST_CMD_GAP(1)); i = t4_wait_op_done(adap, mc_bist_cmd_reg, F_START_BIST, 0, 10, 1); if (i) return i; #define MC_DATA(i) MC_BIST_STATUS_REG(mc_bist_status_rdata_reg, i) for (i = 15; i >= 0; i--) *data++ = ntohl(t4_read_reg(adap, MC_DATA(i))); if (ecc) *ecc = t4_read_reg64(adap, MC_DATA(16)); #undef MC_DATA return 0; } /** * t4_edc_read - read from EDC through backdoor accesses * @adap: the adapter * @idx: which EDC to access * @addr: address of first byte requested * @data: 64 bytes of data containing the requested address * @ecc: where to store the corresponding 64-bit ECC word * * Read 64 bytes of data from EDC starting at a 64-byte-aligned address * that covers the requested address @addr. If @parity is not %NULL it * is assigned the 64-bit ECC word for the read data. */ int t4_edc_read(struct adapter *adap, int idx, u32 addr, __be32 *data, u64 *ecc) { int i; u32 edc_bist_cmd_reg, edc_bist_cmd_addr_reg, edc_bist_cmd_len_reg; u32 edc_bist_cmd_data_pattern, edc_bist_status_rdata_reg; if (is_t4(adap)) { edc_bist_cmd_reg = EDC_REG(A_EDC_BIST_CMD, idx); edc_bist_cmd_addr_reg = EDC_REG(A_EDC_BIST_CMD_ADDR, idx); edc_bist_cmd_len_reg = EDC_REG(A_EDC_BIST_CMD_LEN, idx); edc_bist_cmd_data_pattern = EDC_REG(A_EDC_BIST_DATA_PATTERN, idx); edc_bist_status_rdata_reg = EDC_REG(A_EDC_BIST_STATUS_RDATA, idx); } else { /* * These macro are missing in t4_regs.h file. * Added temporarily for testing. */ #define EDC_STRIDE_T5 (EDC_T51_BASE_ADDR - EDC_T50_BASE_ADDR) #define EDC_REG_T5(reg, idx) (reg + EDC_STRIDE_T5 * idx) edc_bist_cmd_reg = EDC_REG_T5(A_EDC_H_BIST_CMD, idx); edc_bist_cmd_addr_reg = EDC_REG_T5(A_EDC_H_BIST_CMD_ADDR, idx); edc_bist_cmd_len_reg = EDC_REG_T5(A_EDC_H_BIST_CMD_LEN, idx); edc_bist_cmd_data_pattern = EDC_REG_T5(A_EDC_H_BIST_DATA_PATTERN, idx); edc_bist_status_rdata_reg = EDC_REG_T5(A_EDC_H_BIST_STATUS_RDATA, idx); #undef EDC_REG_T5 #undef EDC_STRIDE_T5 } if (t4_read_reg(adap, edc_bist_cmd_reg) & F_START_BIST) return -EBUSY; t4_write_reg(adap, edc_bist_cmd_addr_reg, addr & ~0x3fU); t4_write_reg(adap, edc_bist_cmd_len_reg, 64); t4_write_reg(adap, edc_bist_cmd_data_pattern, 0xc); t4_write_reg(adap, edc_bist_cmd_reg, V_BIST_OPCODE(1) | V_BIST_CMD_GAP(1) | F_START_BIST); i = t4_wait_op_done(adap, edc_bist_cmd_reg, F_START_BIST, 0, 10, 1); if (i) return i; #define EDC_DATA(i) EDC_BIST_STATUS_REG(edc_bist_status_rdata_reg, i) for (i = 15; i >= 0; i--) *data++ = ntohl(t4_read_reg(adap, EDC_DATA(i))); if (ecc) *ecc = t4_read_reg64(adap, EDC_DATA(16)); #undef EDC_DATA return 0; } /** * t4_mem_read - read EDC 0, EDC 1 or MC into buffer * @adap: the adapter * @mtype: memory type: MEM_EDC0, MEM_EDC1 or MEM_MC * @addr: address within indicated memory type * @len: amount of memory to read * @buf: host memory buffer * * Reads an [almost] arbitrary memory region in the firmware: the * firmware memory address, length and host buffer must be aligned on * 32-bit boudaries. The memory is returned as a raw byte sequence from * the firmware's memory. If this memory contains data structures which * contain multi-byte integers, it's the callers responsibility to * perform appropriate byte order conversions. */ int t4_mem_read(struct adapter *adap, int mtype, u32 addr, u32 len, __be32 *buf) { u32 pos, start, end, offset; int ret; /* * Argument sanity checks ... */ if ((addr & 0x3) || (len & 0x3)) return -EINVAL; /* * The underlaying EDC/MC read routines read 64 bytes at a time so we * need to round down the start and round up the end. We'll start * copying out of the first line at (addr - start) a word at a time. */ start = rounddown2(addr, 64); end = roundup2(addr + len, 64); offset = (addr - start)/sizeof(__be32); for (pos = start; pos < end; pos += 64, offset = 0) { __be32 data[16]; /* * Read the chip's memory block and bail if there's an error. */ if ((mtype == MEM_MC) || (mtype == MEM_MC1)) ret = t4_mc_read(adap, mtype - MEM_MC, pos, data, NULL); else ret = t4_edc_read(adap, mtype, pos, data, NULL); if (ret) return ret; /* * Copy the data into the caller's memory buffer. */ while (offset < 16 && len > 0) { *buf++ = data[offset++]; len -= sizeof(__be32); } } return 0; } /* * Return the specified PCI-E Configuration Space register from our Physical * Function. We try first via a Firmware LDST Command (if fw_attach != 0) * since we prefer to let the firmware own all of these registers, but if that * fails we go for it directly ourselves. */ u32 t4_read_pcie_cfg4(struct adapter *adap, int reg, int drv_fw_attach) { /* * If fw_attach != 0, construct and send the Firmware LDST Command to * retrieve the specified PCI-E Configuration Space register. */ if (drv_fw_attach != 0) { struct fw_ldst_cmd ldst_cmd; int ret; memset(&ldst_cmd, 0, sizeof(ldst_cmd)); ldst_cmd.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_FUNC_PCIE)); ldst_cmd.cycles_to_len16 = cpu_to_be32(FW_LEN16(ldst_cmd)); ldst_cmd.u.pcie.select_naccess = V_FW_LDST_CMD_NACCESS(1); ldst_cmd.u.pcie.ctrl_to_fn = (F_FW_LDST_CMD_LC | V_FW_LDST_CMD_FN(adap->pf)); ldst_cmd.u.pcie.r = reg; /* * If the LDST Command succeeds, return the result, otherwise * fall through to reading it directly ourselves ... */ ret = t4_wr_mbox(adap, adap->mbox, &ldst_cmd, sizeof(ldst_cmd), &ldst_cmd); if (ret == 0) return be32_to_cpu(ldst_cmd.u.pcie.data[0]); CH_WARN(adap, "Firmware failed to return " "Configuration Space register %d, err = %d\n", reg, -ret); } /* * Read the desired Configuration Space register via the PCI-E * Backdoor mechanism. */ return t4_hw_pci_read_cfg4(adap, reg); } /** * t4_get_regs_len - return the size of the chips register set * @adapter: the adapter * * Returns the size of the chip's BAR0 register space. */ unsigned int t4_get_regs_len(struct adapter *adapter) { unsigned int chip_version = chip_id(adapter); switch (chip_version) { case CHELSIO_T4: if (adapter->flags & IS_VF) return FW_T4VF_REGMAP_SIZE; return T4_REGMAP_SIZE; case CHELSIO_T5: case CHELSIO_T6: if (adapter->flags & IS_VF) return FW_T4VF_REGMAP_SIZE; return T5_REGMAP_SIZE; } CH_ERR(adapter, "Unsupported chip version %d\n", chip_version); return 0; } /** * t4_get_regs - read chip registers into provided buffer * @adap: the adapter * @buf: register buffer * @buf_size: size (in bytes) of register buffer * * If the provided register buffer isn't large enough for the chip's * full register range, the register dump will be truncated to the * register buffer's size. */ void t4_get_regs(struct adapter *adap, u8 *buf, size_t buf_size) { static const unsigned int t4_reg_ranges[] = { 0x1008, 0x1108, 0x1180, 0x1184, 0x1190, 0x1194, 0x11a0, 0x11a4, 0x11b0, 0x11b4, 0x11fc, 0x123c, 0x1300, 0x173c, 0x1800, 0x18fc, 0x3000, 0x30d8, 0x30e0, 0x30e4, 0x30ec, 0x5910, 0x5920, 0x5924, 0x5960, 0x5960, 0x5968, 0x5968, 0x5970, 0x5970, 0x5978, 0x5978, 0x5980, 0x5980, 0x5988, 0x5988, 0x5990, 0x5990, 0x5998, 0x5998, 0x59a0, 0x59d4, 0x5a00, 0x5ae0, 0x5ae8, 0x5ae8, 0x5af0, 0x5af0, 0x5af8, 0x5af8, 0x6000, 0x6098, 0x6100, 0x6150, 0x6200, 0x6208, 0x6240, 0x6248, 0x6280, 0x62b0, 0x62c0, 0x6338, 0x6370, 0x638c, 0x6400, 0x643c, 0x6500, 0x6524, 0x6a00, 0x6a04, 0x6a14, 0x6a38, 0x6a60, 0x6a70, 0x6a78, 0x6a78, 0x6b00, 0x6b0c, 0x6b1c, 0x6b84, 0x6bf0, 0x6bf8, 0x6c00, 0x6c0c, 0x6c1c, 0x6c84, 0x6cf0, 0x6cf8, 0x6d00, 0x6d0c, 0x6d1c, 0x6d84, 0x6df0, 0x6df8, 0x6e00, 0x6e0c, 0x6e1c, 0x6e84, 0x6ef0, 0x6ef8, 0x6f00, 0x6f0c, 0x6f1c, 0x6f84, 0x6ff0, 0x6ff8, 0x7000, 0x700c, 0x701c, 0x7084, 0x70f0, 0x70f8, 0x7100, 0x710c, 0x711c, 0x7184, 0x71f0, 0x71f8, 0x7200, 0x720c, 0x721c, 0x7284, 0x72f0, 0x72f8, 0x7300, 0x730c, 0x731c, 0x7384, 0x73f0, 0x73f8, 0x7400, 0x7450, 0x7500, 0x7530, 0x7600, 0x760c, 0x7614, 0x761c, 0x7680, 0x76cc, 0x7700, 0x7798, 0x77c0, 0x77fc, 0x7900, 0x79fc, 0x7b00, 0x7b58, 0x7b60, 0x7b84, 0x7b8c, 0x7c38, 0x7d00, 0x7d38, 0x7d40, 0x7d80, 0x7d8c, 0x7ddc, 0x7de4, 0x7e04, 0x7e10, 0x7e1c, 0x7e24, 0x7e38, 0x7e40, 0x7e44, 0x7e4c, 0x7e78, 0x7e80, 0x7ea4, 0x7eac, 0x7edc, 0x7ee8, 0x7efc, 0x8dc0, 0x8e04, 0x8e10, 0x8e1c, 0x8e30, 0x8e78, 0x8ea0, 0x8eb8, 0x8ec0, 0x8f6c, 0x8fc0, 0x9008, 0x9010, 0x9058, 0x9060, 0x9060, 0x9068, 0x9074, 0x90fc, 0x90fc, 0x9400, 0x9408, 0x9410, 0x9458, 0x9600, 0x9600, 0x9608, 0x9638, 0x9640, 0x96bc, 0x9800, 0x9808, 0x9820, 0x983c, 0x9850, 0x9864, 0x9c00, 0x9c6c, 0x9c80, 0x9cec, 0x9d00, 0x9d6c, 0x9d80, 0x9dec, 0x9e00, 0x9e6c, 0x9e80, 0x9eec, 0x9f00, 0x9f6c, 0x9f80, 0x9fec, 0xd004, 0xd004, 0xd010, 0xd03c, 0xdfc0, 0xdfe0, 0xe000, 0xea7c, 0xf000, 0x11190, 0x19040, 0x1906c, 0x19078, 0x19080, 0x1908c, 0x190e4, 0x190f0, 0x190f8, 0x19100, 0x19110, 0x19120, 0x19124, 0x19150, 0x19194, 0x1919c, 0x191b0, 0x191d0, 0x191e8, 0x19238, 0x1924c, 0x193f8, 0x1943c, 0x1944c, 0x19474, 0x19490, 0x194e0, 0x194f0, 0x194f8, 0x19800, 0x19c08, 0x19c10, 0x19c90, 0x19ca0, 0x19ce4, 0x19cf0, 0x19d40, 0x19d50, 0x19d94, 0x19da0, 0x19de8, 0x19df0, 0x19e40, 0x19e50, 0x19e90, 0x19ea0, 0x19f4c, 0x1a000, 0x1a004, 0x1a010, 0x1a06c, 0x1a0b0, 0x1a0e4, 0x1a0ec, 0x1a0f4, 0x1a100, 0x1a108, 0x1a114, 0x1a120, 0x1a128, 0x1a130, 0x1a138, 0x1a138, 0x1a190, 0x1a1c4, 0x1a1fc, 0x1a1fc, 0x1e040, 0x1e04c, 0x1e284, 0x1e28c, 0x1e2c0, 0x1e2c0, 0x1e2e0, 0x1e2e0, 0x1e300, 0x1e384, 0x1e3c0, 0x1e3c8, 0x1e440, 0x1e44c, 0x1e684, 0x1e68c, 0x1e6c0, 0x1e6c0, 0x1e6e0, 0x1e6e0, 0x1e700, 0x1e784, 0x1e7c0, 0x1e7c8, 0x1e840, 0x1e84c, 0x1ea84, 0x1ea8c, 0x1eac0, 0x1eac0, 0x1eae0, 0x1eae0, 0x1eb00, 0x1eb84, 0x1ebc0, 0x1ebc8, 0x1ec40, 0x1ec4c, 0x1ee84, 0x1ee8c, 0x1eec0, 0x1eec0, 0x1eee0, 0x1eee0, 0x1ef00, 0x1ef84, 0x1efc0, 0x1efc8, 0x1f040, 0x1f04c, 0x1f284, 0x1f28c, 0x1f2c0, 0x1f2c0, 0x1f2e0, 0x1f2e0, 0x1f300, 0x1f384, 0x1f3c0, 0x1f3c8, 0x1f440, 0x1f44c, 0x1f684, 0x1f68c, 0x1f6c0, 0x1f6c0, 0x1f6e0, 0x1f6e0, 0x1f700, 0x1f784, 0x1f7c0, 0x1f7c8, 0x1f840, 0x1f84c, 0x1fa84, 0x1fa8c, 0x1fac0, 0x1fac0, 0x1fae0, 0x1fae0, 0x1fb00, 0x1fb84, 0x1fbc0, 0x1fbc8, 0x1fc40, 0x1fc4c, 0x1fe84, 0x1fe8c, 0x1fec0, 0x1fec0, 0x1fee0, 0x1fee0, 0x1ff00, 0x1ff84, 0x1ffc0, 0x1ffc8, 0x20000, 0x2002c, 0x20100, 0x2013c, 0x20190, 0x201a0, 0x201a8, 0x201b8, 0x201c4, 0x201c8, 0x20200, 0x20318, 0x20400, 0x204b4, 0x204c0, 0x20528, 0x20540, 0x20614, 0x21000, 0x21040, 0x2104c, 0x21060, 0x210c0, 0x210ec, 0x21200, 0x21268, 0x21270, 0x21284, 0x212fc, 0x21388, 0x21400, 0x21404, 0x21500, 0x21500, 0x21510, 0x21518, 0x2152c, 0x21530, 0x2153c, 0x2153c, 0x21550, 0x21554, 0x21600, 0x21600, 0x21608, 0x2161c, 0x21624, 0x21628, 0x21630, 0x21634, 0x2163c, 0x2163c, 0x21700, 0x2171c, 0x21780, 0x2178c, 0x21800, 0x21818, 0x21820, 0x21828, 0x21830, 0x21848, 0x21850, 0x21854, 0x21860, 0x21868, 0x21870, 0x21870, 0x21878, 0x21898, 0x218a0, 0x218a8, 0x218b0, 0x218c8, 0x218d0, 0x218d4, 0x218e0, 0x218e8, 0x218f0, 0x218f0, 0x218f8, 0x21a18, 0x21a20, 0x21a28, 0x21a30, 0x21a48, 0x21a50, 0x21a54, 0x21a60, 0x21a68, 0x21a70, 0x21a70, 0x21a78, 0x21a98, 0x21aa0, 0x21aa8, 0x21ab0, 0x21ac8, 0x21ad0, 0x21ad4, 0x21ae0, 0x21ae8, 0x21af0, 0x21af0, 0x21af8, 0x21c18, 0x21c20, 0x21c20, 0x21c28, 0x21c30, 0x21c38, 0x21c38, 0x21c80, 0x21c98, 0x21ca0, 0x21ca8, 0x21cb0, 0x21cc8, 0x21cd0, 0x21cd4, 0x21ce0, 0x21ce8, 0x21cf0, 0x21cf0, 0x21cf8, 0x21d7c, 0x21e00, 0x21e04, 0x22000, 0x2202c, 0x22100, 0x2213c, 0x22190, 0x221a0, 0x221a8, 0x221b8, 0x221c4, 0x221c8, 0x22200, 0x22318, 0x22400, 0x224b4, 0x224c0, 0x22528, 0x22540, 0x22614, 0x23000, 0x23040, 0x2304c, 0x23060, 0x230c0, 0x230ec, 0x23200, 0x23268, 0x23270, 0x23284, 0x232fc, 0x23388, 0x23400, 0x23404, 0x23500, 0x23500, 0x23510, 0x23518, 0x2352c, 0x23530, 0x2353c, 0x2353c, 0x23550, 0x23554, 0x23600, 0x23600, 0x23608, 0x2361c, 0x23624, 0x23628, 0x23630, 0x23634, 0x2363c, 0x2363c, 0x23700, 0x2371c, 0x23780, 0x2378c, 0x23800, 0x23818, 0x23820, 0x23828, 0x23830, 0x23848, 0x23850, 0x23854, 0x23860, 0x23868, 0x23870, 0x23870, 0x23878, 0x23898, 0x238a0, 0x238a8, 0x238b0, 0x238c8, 0x238d0, 0x238d4, 0x238e0, 0x238e8, 0x238f0, 0x238f0, 0x238f8, 0x23a18, 0x23a20, 0x23a28, 0x23a30, 0x23a48, 0x23a50, 0x23a54, 0x23a60, 0x23a68, 0x23a70, 0x23a70, 0x23a78, 0x23a98, 0x23aa0, 0x23aa8, 0x23ab0, 0x23ac8, 0x23ad0, 0x23ad4, 0x23ae0, 0x23ae8, 0x23af0, 0x23af0, 0x23af8, 0x23c18, 0x23c20, 0x23c20, 0x23c28, 0x23c30, 0x23c38, 0x23c38, 0x23c80, 0x23c98, 0x23ca0, 0x23ca8, 0x23cb0, 0x23cc8, 0x23cd0, 0x23cd4, 0x23ce0, 0x23ce8, 0x23cf0, 0x23cf0, 0x23cf8, 0x23d7c, 0x23e00, 0x23e04, 0x24000, 0x2402c, 0x24100, 0x2413c, 0x24190, 0x241a0, 0x241a8, 0x241b8, 0x241c4, 0x241c8, 0x24200, 0x24318, 0x24400, 0x244b4, 0x244c0, 0x24528, 0x24540, 0x24614, 0x25000, 0x25040, 0x2504c, 0x25060, 0x250c0, 0x250ec, 0x25200, 0x25268, 0x25270, 0x25284, 0x252fc, 0x25388, 0x25400, 0x25404, 0x25500, 0x25500, 0x25510, 0x25518, 0x2552c, 0x25530, 0x2553c, 0x2553c, 0x25550, 0x25554, 0x25600, 0x25600, 0x25608, 0x2561c, 0x25624, 0x25628, 0x25630, 0x25634, 0x2563c, 0x2563c, 0x25700, 0x2571c, 0x25780, 0x2578c, 0x25800, 0x25818, 0x25820, 0x25828, 0x25830, 0x25848, 0x25850, 0x25854, 0x25860, 0x25868, 0x25870, 0x25870, 0x25878, 0x25898, 0x258a0, 0x258a8, 0x258b0, 0x258c8, 0x258d0, 0x258d4, 0x258e0, 0x258e8, 0x258f0, 0x258f0, 0x258f8, 0x25a18, 0x25a20, 0x25a28, 0x25a30, 0x25a48, 0x25a50, 0x25a54, 0x25a60, 0x25a68, 0x25a70, 0x25a70, 0x25a78, 0x25a98, 0x25aa0, 0x25aa8, 0x25ab0, 0x25ac8, 0x25ad0, 0x25ad4, 0x25ae0, 0x25ae8, 0x25af0, 0x25af0, 0x25af8, 0x25c18, 0x25c20, 0x25c20, 0x25c28, 0x25c30, 0x25c38, 0x25c38, 0x25c80, 0x25c98, 0x25ca0, 0x25ca8, 0x25cb0, 0x25cc8, 0x25cd0, 0x25cd4, 0x25ce0, 0x25ce8, 0x25cf0, 0x25cf0, 0x25cf8, 0x25d7c, 0x25e00, 0x25e04, 0x26000, 0x2602c, 0x26100, 0x2613c, 0x26190, 0x261a0, 0x261a8, 0x261b8, 0x261c4, 0x261c8, 0x26200, 0x26318, 0x26400, 0x264b4, 0x264c0, 0x26528, 0x26540, 0x26614, 0x27000, 0x27040, 0x2704c, 0x27060, 0x270c0, 0x270ec, 0x27200, 0x27268, 0x27270, 0x27284, 0x272fc, 0x27388, 0x27400, 0x27404, 0x27500, 0x27500, 0x27510, 0x27518, 0x2752c, 0x27530, 0x2753c, 0x2753c, 0x27550, 0x27554, 0x27600, 0x27600, 0x27608, 0x2761c, 0x27624, 0x27628, 0x27630, 0x27634, 0x2763c, 0x2763c, 0x27700, 0x2771c, 0x27780, 0x2778c, 0x27800, 0x27818, 0x27820, 0x27828, 0x27830, 0x27848, 0x27850, 0x27854, 0x27860, 0x27868, 0x27870, 0x27870, 0x27878, 0x27898, 0x278a0, 0x278a8, 0x278b0, 0x278c8, 0x278d0, 0x278d4, 0x278e0, 0x278e8, 0x278f0, 0x278f0, 0x278f8, 0x27a18, 0x27a20, 0x27a28, 0x27a30, 0x27a48, 0x27a50, 0x27a54, 0x27a60, 0x27a68, 0x27a70, 0x27a70, 0x27a78, 0x27a98, 0x27aa0, 0x27aa8, 0x27ab0, 0x27ac8, 0x27ad0, 0x27ad4, 0x27ae0, 0x27ae8, 0x27af0, 0x27af0, 0x27af8, 0x27c18, 0x27c20, 0x27c20, 0x27c28, 0x27c30, 0x27c38, 0x27c38, 0x27c80, 0x27c98, 0x27ca0, 0x27ca8, 0x27cb0, 0x27cc8, 0x27cd0, 0x27cd4, 0x27ce0, 0x27ce8, 0x27cf0, 0x27cf0, 0x27cf8, 0x27d7c, 0x27e00, 0x27e04, }; static const unsigned int t4vf_reg_ranges[] = { VF_SGE_REG(A_SGE_VF_KDOORBELL), VF_SGE_REG(A_SGE_VF_GTS), VF_MPS_REG(A_MPS_VF_CTL), VF_MPS_REG(A_MPS_VF_STAT_RX_VF_ERR_FRAMES_H), VF_PL_REG(A_PL_VF_WHOAMI), VF_PL_REG(A_PL_VF_WHOAMI), VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_CTRL), VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_STATUS), FW_T4VF_MBDATA_BASE_ADDR, FW_T4VF_MBDATA_BASE_ADDR + ((NUM_CIM_PF_MAILBOX_DATA_INSTANCES - 1) * 4), }; static const unsigned int t5_reg_ranges[] = { 0x1008, 0x10c0, 0x10cc, 0x10f8, 0x1100, 0x1100, 0x110c, 0x1148, 0x1180, 0x1184, 0x1190, 0x1194, 0x11a0, 0x11a4, 0x11b0, 0x11b4, 0x11fc, 0x123c, 0x1280, 0x173c, 0x1800, 0x18fc, 0x3000, 0x3028, 0x3060, 0x30b0, 0x30b8, 0x30d8, 0x30e0, 0x30fc, 0x3140, 0x357c, 0x35a8, 0x35cc, 0x35ec, 0x35ec, 0x3600, 0x5624, 0x56cc, 0x56ec, 0x56f4, 0x5720, 0x5728, 0x575c, 0x580c, 0x5814, 0x5890, 0x589c, 0x58a4, 0x58ac, 0x58b8, 0x58bc, 0x5940, 0x59c8, 0x59d0, 0x59dc, 0x59fc, 0x5a18, 0x5a60, 0x5a70, 0x5a80, 0x5a9c, 0x5b94, 0x5bfc, 0x6000, 0x6020, 0x6028, 0x6040, 0x6058, 0x609c, 0x60a8, 0x614c, 0x7700, 0x7798, 0x77c0, 0x78fc, 0x7b00, 0x7b58, 0x7b60, 0x7b84, 0x7b8c, 0x7c54, 0x7d00, 0x7d38, 0x7d40, 0x7d80, 0x7d8c, 0x7ddc, 0x7de4, 0x7e04, 0x7e10, 0x7e1c, 0x7e24, 0x7e38, 0x7e40, 0x7e44, 0x7e4c, 0x7e78, 0x7e80, 0x7edc, 0x7ee8, 0x7efc, 0x8dc0, 0x8de0, 0x8df8, 0x8e04, 0x8e10, 0x8e84, 0x8ea0, 0x8f84, 0x8fc0, 0x9058, 0x9060, 0x9060, 0x9068, 0x90f8, 0x9400, 0x9408, 0x9410, 0x9470, 0x9600, 0x9600, 0x9608, 0x9638, 0x9640, 0x96f4, 0x9800, 0x9808, 0x9820, 0x983c, 0x9850, 0x9864, 0x9c00, 0x9c6c, 0x9c80, 0x9cec, 0x9d00, 0x9d6c, 0x9d80, 0x9dec, 0x9e00, 0x9e6c, 0x9e80, 0x9eec, 0x9f00, 0x9f6c, 0x9f80, 0xa020, 0xd004, 0xd004, 0xd010, 0xd03c, 0xdfc0, 0xdfe0, 0xe000, 0x1106c, 0x11074, 0x11088, 0x1109c, 0x1117c, 0x11190, 0x11204, 0x19040, 0x1906c, 0x19078, 0x19080, 0x1908c, 0x190e8, 0x190f0, 0x190f8, 0x19100, 0x19110, 0x19120, 0x19124, 0x19150, 0x19194, 0x1919c, 0x191b0, 0x191d0, 0x191e8, 0x19238, 0x19290, 0x193f8, 0x19428, 0x19430, 0x19444, 0x1944c, 0x1946c, 0x19474, 0x19474, 0x19490, 0x194cc, 0x194f0, 0x194f8, 0x19c00, 0x19c08, 0x19c10, 0x19c60, 0x19c94, 0x19ce4, 0x19cf0, 0x19d40, 0x19d50, 0x19d94, 0x19da0, 0x19de8, 0x19df0, 0x19e10, 0x19e50, 0x19e90, 0x19ea0, 0x19f24, 0x19f34, 0x19f34, 0x19f40, 0x19f50, 0x19f90, 0x19fb4, 0x19fc4, 0x19fe4, 0x1a000, 0x1a004, 0x1a010, 0x1a06c, 0x1a0b0, 0x1a0e4, 0x1a0ec, 0x1a0f8, 0x1a100, 0x1a108, 0x1a114, 0x1a120, 0x1a128, 0x1a130, 0x1a138, 0x1a138, 0x1a190, 0x1a1c4, 0x1a1fc, 0x1a1fc, 0x1e008, 0x1e00c, 0x1e040, 0x1e044, 0x1e04c, 0x1e04c, 0x1e284, 0x1e290, 0x1e2c0, 0x1e2c0, 0x1e2e0, 0x1e2e0, 0x1e300, 0x1e384, 0x1e3c0, 0x1e3c8, 0x1e408, 0x1e40c, 0x1e440, 0x1e444, 0x1e44c, 0x1e44c, 0x1e684, 0x1e690, 0x1e6c0, 0x1e6c0, 0x1e6e0, 0x1e6e0, 0x1e700, 0x1e784, 0x1e7c0, 0x1e7c8, 0x1e808, 0x1e80c, 0x1e840, 0x1e844, 0x1e84c, 0x1e84c, 0x1ea84, 0x1ea90, 0x1eac0, 0x1eac0, 0x1eae0, 0x1eae0, 0x1eb00, 0x1eb84, 0x1ebc0, 0x1ebc8, 0x1ec08, 0x1ec0c, 0x1ec40, 0x1ec44, 0x1ec4c, 0x1ec4c, 0x1ee84, 0x1ee90, 0x1eec0, 0x1eec0, 0x1eee0, 0x1eee0, 0x1ef00, 0x1ef84, 0x1efc0, 0x1efc8, 0x1f008, 0x1f00c, 0x1f040, 0x1f044, 0x1f04c, 0x1f04c, 0x1f284, 0x1f290, 0x1f2c0, 0x1f2c0, 0x1f2e0, 0x1f2e0, 0x1f300, 0x1f384, 0x1f3c0, 0x1f3c8, 0x1f408, 0x1f40c, 0x1f440, 0x1f444, 0x1f44c, 0x1f44c, 0x1f684, 0x1f690, 0x1f6c0, 0x1f6c0, 0x1f6e0, 0x1f6e0, 0x1f700, 0x1f784, 0x1f7c0, 0x1f7c8, 0x1f808, 0x1f80c, 0x1f840, 0x1f844, 0x1f84c, 0x1f84c, 0x1fa84, 0x1fa90, 0x1fac0, 0x1fac0, 0x1fae0, 0x1fae0, 0x1fb00, 0x1fb84, 0x1fbc0, 0x1fbc8, 0x1fc08, 0x1fc0c, 0x1fc40, 0x1fc44, 0x1fc4c, 0x1fc4c, 0x1fe84, 0x1fe90, 0x1fec0, 0x1fec0, 0x1fee0, 0x1fee0, 0x1ff00, 0x1ff84, 0x1ffc0, 0x1ffc8, 0x30000, 0x30030, 0x30038, 0x30038, 0x30040, 0x30040, 0x30100, 0x30144, 0x30190, 0x301a0, 0x301a8, 0x301b8, 0x301c4, 0x301c8, 0x301d0, 0x301d0, 0x30200, 0x30318, 0x30400, 0x304b4, 0x304c0, 0x3052c, 0x30540, 0x3061c, 0x30800, 0x30828, 0x30834, 0x30834, 0x308c0, 0x30908, 0x30910, 0x309ac, 0x30a00, 0x30a14, 0x30a1c, 0x30a2c, 0x30a44, 0x30a50, 0x30a74, 0x30a74, 0x30a7c, 0x30afc, 0x30b08, 0x30c24, 0x30d00, 0x30d00, 0x30d08, 0x30d14, 0x30d1c, 0x30d20, 0x30d3c, 0x30d3c, 0x30d48, 0x30d50, 0x31200, 0x3120c, 0x31220, 0x31220, 0x31240, 0x31240, 0x31600, 0x3160c, 0x31a00, 0x31a1c, 0x31e00, 0x31e20, 0x31e38, 0x31e3c, 0x31e80, 0x31e80, 0x31e88, 0x31ea8, 0x31eb0, 0x31eb4, 0x31ec8, 0x31ed4, 0x31fb8, 0x32004, 0x32200, 0x32200, 0x32208, 0x32240, 0x32248, 0x32280, 0x32288, 0x322c0, 0x322c8, 0x322fc, 0x32600, 0x32630, 0x32a00, 0x32abc, 0x32b00, 0x32b10, 0x32b20, 0x32b30, 0x32b40, 0x32b50, 0x32b60, 0x32b70, 0x33000, 0x33028, 0x33030, 0x33048, 0x33060, 0x33068, 0x33070, 0x3309c, 0x330f0, 0x33128, 0x33130, 0x33148, 0x33160, 0x33168, 0x33170, 0x3319c, 0x331f0, 0x33238, 0x33240, 0x33240, 0x33248, 0x33250, 0x3325c, 0x33264, 0x33270, 0x332b8, 0x332c0, 0x332e4, 0x332f8, 0x33338, 0x33340, 0x33340, 0x33348, 0x33350, 0x3335c, 0x33364, 0x33370, 0x333b8, 0x333c0, 0x333e4, 0x333f8, 0x33428, 0x33430, 0x33448, 0x33460, 0x33468, 0x33470, 0x3349c, 0x334f0, 0x33528, 0x33530, 0x33548, 0x33560, 0x33568, 0x33570, 0x3359c, 0x335f0, 0x33638, 0x33640, 0x33640, 0x33648, 0x33650, 0x3365c, 0x33664, 0x33670, 0x336b8, 0x336c0, 0x336e4, 0x336f8, 0x33738, 0x33740, 0x33740, 0x33748, 0x33750, 0x3375c, 0x33764, 0x33770, 0x337b8, 0x337c0, 0x337e4, 0x337f8, 0x337fc, 0x33814, 0x33814, 0x3382c, 0x3382c, 0x33880, 0x3388c, 0x338e8, 0x338ec, 0x33900, 0x33928, 0x33930, 0x33948, 0x33960, 0x33968, 0x33970, 0x3399c, 0x339f0, 0x33a38, 0x33a40, 0x33a40, 0x33a48, 0x33a50, 0x33a5c, 0x33a64, 0x33a70, 0x33ab8, 0x33ac0, 0x33ae4, 0x33af8, 0x33b10, 0x33b28, 0x33b28, 0x33b3c, 0x33b50, 0x33bf0, 0x33c10, 0x33c28, 0x33c28, 0x33c3c, 0x33c50, 0x33cf0, 0x33cfc, 0x34000, 0x34030, 0x34038, 0x34038, 0x34040, 0x34040, 0x34100, 0x34144, 0x34190, 0x341a0, 0x341a8, 0x341b8, 0x341c4, 0x341c8, 0x341d0, 0x341d0, 0x34200, 0x34318, 0x34400, 0x344b4, 0x344c0, 0x3452c, 0x34540, 0x3461c, 0x34800, 0x34828, 0x34834, 0x34834, 0x348c0, 0x34908, 0x34910, 0x349ac, 0x34a00, 0x34a14, 0x34a1c, 0x34a2c, 0x34a44, 0x34a50, 0x34a74, 0x34a74, 0x34a7c, 0x34afc, 0x34b08, 0x34c24, 0x34d00, 0x34d00, 0x34d08, 0x34d14, 0x34d1c, 0x34d20, 0x34d3c, 0x34d3c, 0x34d48, 0x34d50, 0x35200, 0x3520c, 0x35220, 0x35220, 0x35240, 0x35240, 0x35600, 0x3560c, 0x35a00, 0x35a1c, 0x35e00, 0x35e20, 0x35e38, 0x35e3c, 0x35e80, 0x35e80, 0x35e88, 0x35ea8, 0x35eb0, 0x35eb4, 0x35ec8, 0x35ed4, 0x35fb8, 0x36004, 0x36200, 0x36200, 0x36208, 0x36240, 0x36248, 0x36280, 0x36288, 0x362c0, 0x362c8, 0x362fc, 0x36600, 0x36630, 0x36a00, 0x36abc, 0x36b00, 0x36b10, 0x36b20, 0x36b30, 0x36b40, 0x36b50, 0x36b60, 0x36b70, 0x37000, 0x37028, 0x37030, 0x37048, 0x37060, 0x37068, 0x37070, 0x3709c, 0x370f0, 0x37128, 0x37130, 0x37148, 0x37160, 0x37168, 0x37170, 0x3719c, 0x371f0, 0x37238, 0x37240, 0x37240, 0x37248, 0x37250, 0x3725c, 0x37264, 0x37270, 0x372b8, 0x372c0, 0x372e4, 0x372f8, 0x37338, 0x37340, 0x37340, 0x37348, 0x37350, 0x3735c, 0x37364, 0x37370, 0x373b8, 0x373c0, 0x373e4, 0x373f8, 0x37428, 0x37430, 0x37448, 0x37460, 0x37468, 0x37470, 0x3749c, 0x374f0, 0x37528, 0x37530, 0x37548, 0x37560, 0x37568, 0x37570, 0x3759c, 0x375f0, 0x37638, 0x37640, 0x37640, 0x37648, 0x37650, 0x3765c, 0x37664, 0x37670, 0x376b8, 0x376c0, 0x376e4, 0x376f8, 0x37738, 0x37740, 0x37740, 0x37748, 0x37750, 0x3775c, 0x37764, 0x37770, 0x377b8, 0x377c0, 0x377e4, 0x377f8, 0x377fc, 0x37814, 0x37814, 0x3782c, 0x3782c, 0x37880, 0x3788c, 0x378e8, 0x378ec, 0x37900, 0x37928, 0x37930, 0x37948, 0x37960, 0x37968, 0x37970, 0x3799c, 0x379f0, 0x37a38, 0x37a40, 0x37a40, 0x37a48, 0x37a50, 0x37a5c, 0x37a64, 0x37a70, 0x37ab8, 0x37ac0, 0x37ae4, 0x37af8, 0x37b10, 0x37b28, 0x37b28, 0x37b3c, 0x37b50, 0x37bf0, 0x37c10, 0x37c28, 0x37c28, 0x37c3c, 0x37c50, 0x37cf0, 0x37cfc, 0x38000, 0x38030, 0x38038, 0x38038, 0x38040, 0x38040, 0x38100, 0x38144, 0x38190, 0x381a0, 0x381a8, 0x381b8, 0x381c4, 0x381c8, 0x381d0, 0x381d0, 0x38200, 0x38318, 0x38400, 0x384b4, 0x384c0, 0x3852c, 0x38540, 0x3861c, 0x38800, 0x38828, 0x38834, 0x38834, 0x388c0, 0x38908, 0x38910, 0x389ac, 0x38a00, 0x38a14, 0x38a1c, 0x38a2c, 0x38a44, 0x38a50, 0x38a74, 0x38a74, 0x38a7c, 0x38afc, 0x38b08, 0x38c24, 0x38d00, 0x38d00, 0x38d08, 0x38d14, 0x38d1c, 0x38d20, 0x38d3c, 0x38d3c, 0x38d48, 0x38d50, 0x39200, 0x3920c, 0x39220, 0x39220, 0x39240, 0x39240, 0x39600, 0x3960c, 0x39a00, 0x39a1c, 0x39e00, 0x39e20, 0x39e38, 0x39e3c, 0x39e80, 0x39e80, 0x39e88, 0x39ea8, 0x39eb0, 0x39eb4, 0x39ec8, 0x39ed4, 0x39fb8, 0x3a004, 0x3a200, 0x3a200, 0x3a208, 0x3a240, 0x3a248, 0x3a280, 0x3a288, 0x3a2c0, 0x3a2c8, 0x3a2fc, 0x3a600, 0x3a630, 0x3aa00, 0x3aabc, 0x3ab00, 0x3ab10, 0x3ab20, 0x3ab30, 0x3ab40, 0x3ab50, 0x3ab60, 0x3ab70, 0x3b000, 0x3b028, 0x3b030, 0x3b048, 0x3b060, 0x3b068, 0x3b070, 0x3b09c, 0x3b0f0, 0x3b128, 0x3b130, 0x3b148, 0x3b160, 0x3b168, 0x3b170, 0x3b19c, 0x3b1f0, 0x3b238, 0x3b240, 0x3b240, 0x3b248, 0x3b250, 0x3b25c, 0x3b264, 0x3b270, 0x3b2b8, 0x3b2c0, 0x3b2e4, 0x3b2f8, 0x3b338, 0x3b340, 0x3b340, 0x3b348, 0x3b350, 0x3b35c, 0x3b364, 0x3b370, 0x3b3b8, 0x3b3c0, 0x3b3e4, 0x3b3f8, 0x3b428, 0x3b430, 0x3b448, 0x3b460, 0x3b468, 0x3b470, 0x3b49c, 0x3b4f0, 0x3b528, 0x3b530, 0x3b548, 0x3b560, 0x3b568, 0x3b570, 0x3b59c, 0x3b5f0, 0x3b638, 0x3b640, 0x3b640, 0x3b648, 0x3b650, 0x3b65c, 0x3b664, 0x3b670, 0x3b6b8, 0x3b6c0, 0x3b6e4, 0x3b6f8, 0x3b738, 0x3b740, 0x3b740, 0x3b748, 0x3b750, 0x3b75c, 0x3b764, 0x3b770, 0x3b7b8, 0x3b7c0, 0x3b7e4, 0x3b7f8, 0x3b7fc, 0x3b814, 0x3b814, 0x3b82c, 0x3b82c, 0x3b880, 0x3b88c, 0x3b8e8, 0x3b8ec, 0x3b900, 0x3b928, 0x3b930, 0x3b948, 0x3b960, 0x3b968, 0x3b970, 0x3b99c, 0x3b9f0, 0x3ba38, 0x3ba40, 0x3ba40, 0x3ba48, 0x3ba50, 0x3ba5c, 0x3ba64, 0x3ba70, 0x3bab8, 0x3bac0, 0x3bae4, 0x3baf8, 0x3bb10, 0x3bb28, 0x3bb28, 0x3bb3c, 0x3bb50, 0x3bbf0, 0x3bc10, 0x3bc28, 0x3bc28, 0x3bc3c, 0x3bc50, 0x3bcf0, 0x3bcfc, 0x3c000, 0x3c030, 0x3c038, 0x3c038, 0x3c040, 0x3c040, 0x3c100, 0x3c144, 0x3c190, 0x3c1a0, 0x3c1a8, 0x3c1b8, 0x3c1c4, 0x3c1c8, 0x3c1d0, 0x3c1d0, 0x3c200, 0x3c318, 0x3c400, 0x3c4b4, 0x3c4c0, 0x3c52c, 0x3c540, 0x3c61c, 0x3c800, 0x3c828, 0x3c834, 0x3c834, 0x3c8c0, 0x3c908, 0x3c910, 0x3c9ac, 0x3ca00, 0x3ca14, 0x3ca1c, 0x3ca2c, 0x3ca44, 0x3ca50, 0x3ca74, 0x3ca74, 0x3ca7c, 0x3cafc, 0x3cb08, 0x3cc24, 0x3cd00, 0x3cd00, 0x3cd08, 0x3cd14, 0x3cd1c, 0x3cd20, 0x3cd3c, 0x3cd3c, 0x3cd48, 0x3cd50, 0x3d200, 0x3d20c, 0x3d220, 0x3d220, 0x3d240, 0x3d240, 0x3d600, 0x3d60c, 0x3da00, 0x3da1c, 0x3de00, 0x3de20, 0x3de38, 0x3de3c, 0x3de80, 0x3de80, 0x3de88, 0x3dea8, 0x3deb0, 0x3deb4, 0x3dec8, 0x3ded4, 0x3dfb8, 0x3e004, 0x3e200, 0x3e200, 0x3e208, 0x3e240, 0x3e248, 0x3e280, 0x3e288, 0x3e2c0, 0x3e2c8, 0x3e2fc, 0x3e600, 0x3e630, 0x3ea00, 0x3eabc, 0x3eb00, 0x3eb10, 0x3eb20, 0x3eb30, 0x3eb40, 0x3eb50, 0x3eb60, 0x3eb70, 0x3f000, 0x3f028, 0x3f030, 0x3f048, 0x3f060, 0x3f068, 0x3f070, 0x3f09c, 0x3f0f0, 0x3f128, 0x3f130, 0x3f148, 0x3f160, 0x3f168, 0x3f170, 0x3f19c, 0x3f1f0, 0x3f238, 0x3f240, 0x3f240, 0x3f248, 0x3f250, 0x3f25c, 0x3f264, 0x3f270, 0x3f2b8, 0x3f2c0, 0x3f2e4, 0x3f2f8, 0x3f338, 0x3f340, 0x3f340, 0x3f348, 0x3f350, 0x3f35c, 0x3f364, 0x3f370, 0x3f3b8, 0x3f3c0, 0x3f3e4, 0x3f3f8, 0x3f428, 0x3f430, 0x3f448, 0x3f460, 0x3f468, 0x3f470, 0x3f49c, 0x3f4f0, 0x3f528, 0x3f530, 0x3f548, 0x3f560, 0x3f568, 0x3f570, 0x3f59c, 0x3f5f0, 0x3f638, 0x3f640, 0x3f640, 0x3f648, 0x3f650, 0x3f65c, 0x3f664, 0x3f670, 0x3f6b8, 0x3f6c0, 0x3f6e4, 0x3f6f8, 0x3f738, 0x3f740, 0x3f740, 0x3f748, 0x3f750, 0x3f75c, 0x3f764, 0x3f770, 0x3f7b8, 0x3f7c0, 0x3f7e4, 0x3f7f8, 0x3f7fc, 0x3f814, 0x3f814, 0x3f82c, 0x3f82c, 0x3f880, 0x3f88c, 0x3f8e8, 0x3f8ec, 0x3f900, 0x3f928, 0x3f930, 0x3f948, 0x3f960, 0x3f968, 0x3f970, 0x3f99c, 0x3f9f0, 0x3fa38, 0x3fa40, 0x3fa40, 0x3fa48, 0x3fa50, 0x3fa5c, 0x3fa64, 0x3fa70, 0x3fab8, 0x3fac0, 0x3fae4, 0x3faf8, 0x3fb10, 0x3fb28, 0x3fb28, 0x3fb3c, 0x3fb50, 0x3fbf0, 0x3fc10, 0x3fc28, 0x3fc28, 0x3fc3c, 0x3fc50, 0x3fcf0, 0x3fcfc, 0x40000, 0x4000c, 0x40040, 0x40050, 0x40060, 0x40068, 0x4007c, 0x4008c, 0x40094, 0x400b0, 0x400c0, 0x40144, 0x40180, 0x4018c, 0x40200, 0x40254, 0x40260, 0x40264, 0x40270, 0x40288, 0x40290, 0x40298, 0x402ac, 0x402c8, 0x402d0, 0x402e0, 0x402f0, 0x402f0, 0x40300, 0x4033c, 0x403f8, 0x403fc, 0x41304, 0x413c4, 0x41400, 0x4140c, 0x41414, 0x4141c, 0x41480, 0x414d0, 0x44000, 0x44054, 0x4405c, 0x44078, 0x440c0, 0x44174, 0x44180, 0x441ac, 0x441b4, 0x441b8, 0x441c0, 0x44254, 0x4425c, 0x44278, 0x442c0, 0x44374, 0x44380, 0x443ac, 0x443b4, 0x443b8, 0x443c0, 0x44454, 0x4445c, 0x44478, 0x444c0, 0x44574, 0x44580, 0x445ac, 0x445b4, 0x445b8, 0x445c0, 0x44654, 0x4465c, 0x44678, 0x446c0, 0x44774, 0x44780, 0x447ac, 0x447b4, 0x447b8, 0x447c0, 0x44854, 0x4485c, 0x44878, 0x448c0, 0x44974, 0x44980, 0x449ac, 0x449b4, 0x449b8, 0x449c0, 0x449fc, 0x45000, 0x45004, 0x45010, 0x45030, 0x45040, 0x45060, 0x45068, 0x45068, 0x45080, 0x45084, 0x450a0, 0x450b0, 0x45200, 0x45204, 0x45210, 0x45230, 0x45240, 0x45260, 0x45268, 0x45268, 0x45280, 0x45284, 0x452a0, 0x452b0, 0x460c0, 0x460e4, 0x47000, 0x4703c, 0x47044, 0x4708c, 0x47200, 0x47250, 0x47400, 0x47408, 0x47414, 0x47420, 0x47600, 0x47618, 0x47800, 0x47814, 0x48000, 0x4800c, 0x48040, 0x48050, 0x48060, 0x48068, 0x4807c, 0x4808c, 0x48094, 0x480b0, 0x480c0, 0x48144, 0x48180, 0x4818c, 0x48200, 0x48254, 0x48260, 0x48264, 0x48270, 0x48288, 0x48290, 0x48298, 0x482ac, 0x482c8, 0x482d0, 0x482e0, 0x482f0, 0x482f0, 0x48300, 0x4833c, 0x483f8, 0x483fc, 0x49304, 0x493c4, 0x49400, 0x4940c, 0x49414, 0x4941c, 0x49480, 0x494d0, 0x4c000, 0x4c054, 0x4c05c, 0x4c078, 0x4c0c0, 0x4c174, 0x4c180, 0x4c1ac, 0x4c1b4, 0x4c1b8, 0x4c1c0, 0x4c254, 0x4c25c, 0x4c278, 0x4c2c0, 0x4c374, 0x4c380, 0x4c3ac, 0x4c3b4, 0x4c3b8, 0x4c3c0, 0x4c454, 0x4c45c, 0x4c478, 0x4c4c0, 0x4c574, 0x4c580, 0x4c5ac, 0x4c5b4, 0x4c5b8, 0x4c5c0, 0x4c654, 0x4c65c, 0x4c678, 0x4c6c0, 0x4c774, 0x4c780, 0x4c7ac, 0x4c7b4, 0x4c7b8, 0x4c7c0, 0x4c854, 0x4c85c, 0x4c878, 0x4c8c0, 0x4c974, 0x4c980, 0x4c9ac, 0x4c9b4, 0x4c9b8, 0x4c9c0, 0x4c9fc, 0x4d000, 0x4d004, 0x4d010, 0x4d030, 0x4d040, 0x4d060, 0x4d068, 0x4d068, 0x4d080, 0x4d084, 0x4d0a0, 0x4d0b0, 0x4d200, 0x4d204, 0x4d210, 0x4d230, 0x4d240, 0x4d260, 0x4d268, 0x4d268, 0x4d280, 0x4d284, 0x4d2a0, 0x4d2b0, 0x4e0c0, 0x4e0e4, 0x4f000, 0x4f03c, 0x4f044, 0x4f08c, 0x4f200, 0x4f250, 0x4f400, 0x4f408, 0x4f414, 0x4f420, 0x4f600, 0x4f618, 0x4f800, 0x4f814, 0x50000, 0x50084, 0x50090, 0x500cc, 0x50400, 0x50400, 0x50800, 0x50884, 0x50890, 0x508cc, 0x50c00, 0x50c00, 0x51000, 0x5101c, 0x51300, 0x51308, }; static const unsigned int t5vf_reg_ranges[] = { VF_SGE_REG(A_SGE_VF_KDOORBELL), VF_SGE_REG(A_SGE_VF_GTS), VF_MPS_REG(A_MPS_VF_CTL), VF_MPS_REG(A_MPS_VF_STAT_RX_VF_ERR_FRAMES_H), VF_PL_REG(A_PL_VF_WHOAMI), VF_PL_REG(A_PL_VF_REVISION), VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_CTRL), VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_STATUS), FW_T4VF_MBDATA_BASE_ADDR, FW_T4VF_MBDATA_BASE_ADDR + ((NUM_CIM_PF_MAILBOX_DATA_INSTANCES - 1) * 4), }; static const unsigned int t6_reg_ranges[] = { 0x1008, 0x101c, 0x1024, 0x10a8, 0x10b4, 0x10f8, 0x1100, 0x1114, 0x111c, 0x112c, 0x1138, 0x113c, 0x1144, 0x114c, 0x1180, 0x1184, 0x1190, 0x1194, 0x11a0, 0x11a4, 0x11b0, 0x11b4, 0x11fc, 0x1274, 0x1280, 0x133c, 0x1800, 0x18fc, 0x3000, 0x302c, 0x3060, 0x30b0, 0x30b8, 0x30d8, 0x30e0, 0x30fc, 0x3140, 0x357c, 0x35a8, 0x35cc, 0x35ec, 0x35ec, 0x3600, 0x5624, 0x56cc, 0x56ec, 0x56f4, 0x5720, 0x5728, 0x575c, 0x580c, 0x5814, 0x5890, 0x589c, 0x58a4, 0x58ac, 0x58b8, 0x58bc, 0x5940, 0x595c, 0x5980, 0x598c, 0x59b0, 0x59c8, 0x59d0, 0x59dc, 0x59fc, 0x5a18, 0x5a60, 0x5a6c, 0x5a80, 0x5a8c, 0x5a94, 0x5a9c, 0x5b94, 0x5bfc, 0x5c10, 0x5e48, 0x5e50, 0x5e94, 0x5ea0, 0x5eb0, 0x5ec0, 0x5ec0, 0x5ec8, 0x5ed0, 0x5ee0, 0x5ee0, 0x5ef0, 0x5ef0, 0x5f00, 0x5f00, 0x6000, 0x6020, 0x6028, 0x6040, 0x6058, 0x609c, 0x60a8, 0x619c, 0x7700, 0x7798, 0x77c0, 0x7880, 0x78cc, 0x78fc, 0x7b00, 0x7b58, 0x7b60, 0x7b84, 0x7b8c, 0x7c54, 0x7d00, 0x7d38, 0x7d40, 0x7d84, 0x7d8c, 0x7ddc, 0x7de4, 0x7e04, 0x7e10, 0x7e1c, 0x7e24, 0x7e38, 0x7e40, 0x7e44, 0x7e4c, 0x7e78, 0x7e80, 0x7edc, 0x7ee8, 0x7efc, 0x8dc0, 0x8de4, 0x8df8, 0x8e04, 0x8e10, 0x8e84, 0x8ea0, 0x8f88, 0x8fb8, 0x9058, 0x9060, 0x9060, 0x9068, 0x90f8, 0x9100, 0x9124, 0x9400, 0x9470, 0x9600, 0x9600, 0x9608, 0x9638, 0x9640, 0x9704, 0x9710, 0x971c, 0x9800, 0x9808, 0x9820, 0x983c, 0x9850, 0x9864, 0x9c00, 0x9c6c, 0x9c80, 0x9cec, 0x9d00, 0x9d6c, 0x9d80, 0x9dec, 0x9e00, 0x9e6c, 0x9e80, 0x9eec, 0x9f00, 0x9f6c, 0x9f80, 0xa020, 0xd004, 0xd03c, 0xd100, 0xd118, 0xd200, 0xd214, 0xd220, 0xd234, 0xd240, 0xd254, 0xd260, 0xd274, 0xd280, 0xd294, 0xd2a0, 0xd2b4, 0xd2c0, 0xd2d4, 0xd2e0, 0xd2f4, 0xd300, 0xd31c, 0xdfc0, 0xdfe0, 0xe000, 0xf008, 0xf010, 0xf018, 0xf020, 0xf028, 0x11000, 0x11014, 0x11048, 0x1106c, 0x11074, 0x11088, 0x11098, 0x11120, 0x1112c, 0x1117c, 0x11190, 0x112e0, 0x11300, 0x1130c, 0x12000, 0x1206c, 0x19040, 0x1906c, 0x19078, 0x19080, 0x1908c, 0x190e8, 0x190f0, 0x190f8, 0x19100, 0x19110, 0x19120, 0x19124, 0x19150, 0x19194, 0x1919c, 0x191b0, 0x191d0, 0x191e8, 0x19238, 0x19290, 0x192a4, 0x192b0, 0x192bc, 0x192bc, 0x19348, 0x1934c, 0x193f8, 0x19418, 0x19420, 0x19428, 0x19430, 0x19444, 0x1944c, 0x1946c, 0x19474, 0x19474, 0x19490, 0x194cc, 0x194f0, 0x194f8, 0x19c00, 0x19c48, 0x19c50, 0x19c80, 0x19c94, 0x19c98, 0x19ca0, 0x19cbc, 0x19ce4, 0x19ce4, 0x19cf0, 0x19cf8, 0x19d00, 0x19d28, 0x19d50, 0x19d78, 0x19d94, 0x19d98, 0x19da0, 0x19dc8, 0x19df0, 0x19e10, 0x19e50, 0x19e6c, 0x19ea0, 0x19ebc, 0x19ec4, 0x19ef4, 0x19f04, 0x19f2c, 0x19f34, 0x19f34, 0x19f40, 0x19f50, 0x19f90, 0x19fac, 0x19fc4, 0x19fc8, 0x19fd0, 0x19fe4, 0x1a000, 0x1a004, 0x1a010, 0x1a06c, 0x1a0b0, 0x1a0e4, 0x1a0ec, 0x1a0f8, 0x1a100, 0x1a108, 0x1a114, 0x1a120, 0x1a128, 0x1a130, 0x1a138, 0x1a138, 0x1a190, 0x1a1c4, 0x1a1fc, 0x1a1fc, 0x1e008, 0x1e00c, 0x1e040, 0x1e044, 0x1e04c, 0x1e04c, 0x1e284, 0x1e290, 0x1e2c0, 0x1e2c0, 0x1e2e0, 0x1e2e0, 0x1e300, 0x1e384, 0x1e3c0, 0x1e3c8, 0x1e408, 0x1e40c, 0x1e440, 0x1e444, 0x1e44c, 0x1e44c, 0x1e684, 0x1e690, 0x1e6c0, 0x1e6c0, 0x1e6e0, 0x1e6e0, 0x1e700, 0x1e784, 0x1e7c0, 0x1e7c8, 0x1e808, 0x1e80c, 0x1e840, 0x1e844, 0x1e84c, 0x1e84c, 0x1ea84, 0x1ea90, 0x1eac0, 0x1eac0, 0x1eae0, 0x1eae0, 0x1eb00, 0x1eb84, 0x1ebc0, 0x1ebc8, 0x1ec08, 0x1ec0c, 0x1ec40, 0x1ec44, 0x1ec4c, 0x1ec4c, 0x1ee84, 0x1ee90, 0x1eec0, 0x1eec0, 0x1eee0, 0x1eee0, 0x1ef00, 0x1ef84, 0x1efc0, 0x1efc8, 0x1f008, 0x1f00c, 0x1f040, 0x1f044, 0x1f04c, 0x1f04c, 0x1f284, 0x1f290, 0x1f2c0, 0x1f2c0, 0x1f2e0, 0x1f2e0, 0x1f300, 0x1f384, 0x1f3c0, 0x1f3c8, 0x1f408, 0x1f40c, 0x1f440, 0x1f444, 0x1f44c, 0x1f44c, 0x1f684, 0x1f690, 0x1f6c0, 0x1f6c0, 0x1f6e0, 0x1f6e0, 0x1f700, 0x1f784, 0x1f7c0, 0x1f7c8, 0x1f808, 0x1f80c, 0x1f840, 0x1f844, 0x1f84c, 0x1f84c, 0x1fa84, 0x1fa90, 0x1fac0, 0x1fac0, 0x1fae0, 0x1fae0, 0x1fb00, 0x1fb84, 0x1fbc0, 0x1fbc8, 0x1fc08, 0x1fc0c, 0x1fc40, 0x1fc44, 0x1fc4c, 0x1fc4c, 0x1fe84, 0x1fe90, 0x1fec0, 0x1fec0, 0x1fee0, 0x1fee0, 0x1ff00, 0x1ff84, 0x1ffc0, 0x1ffc8, 0x30000, 0x30030, 0x30038, 0x30038, 0x30040, 0x30040, 0x30048, 0x30048, 0x30050, 0x30050, 0x3005c, 0x30060, 0x30068, 0x30068, 0x30070, 0x30070, 0x30100, 0x30168, 0x30190, 0x301a0, 0x301a8, 0x301b8, 0x301c4, 0x301c8, 0x301d0, 0x301d0, 0x30200, 0x30320, 0x30400, 0x304b4, 0x304c0, 0x3052c, 0x30540, 0x3061c, 0x30800, 0x308a0, 0x308c0, 0x30908, 0x30910, 0x309b8, 0x30a00, 0x30a04, 0x30a0c, 0x30a14, 0x30a1c, 0x30a2c, 0x30a44, 0x30a50, 0x30a74, 0x30a74, 0x30a7c, 0x30afc, 0x30b08, 0x30c24, 0x30d00, 0x30d14, 0x30d1c, 0x30d3c, 0x30d44, 0x30d4c, 0x30d54, 0x30d74, 0x30d7c, 0x30d7c, 0x30de0, 0x30de0, 0x30e00, 0x30ed4, 0x30f00, 0x30fa4, 0x30fc0, 0x30fc4, 0x31000, 0x31004, 0x31080, 0x310fc, 0x31208, 0x31220, 0x3123c, 0x31254, 0x31300, 0x31300, 0x31308, 0x3131c, 0x31338, 0x3133c, 0x31380, 0x31380, 0x31388, 0x313a8, 0x313b4, 0x313b4, 0x31400, 0x31420, 0x31438, 0x3143c, 0x31480, 0x31480, 0x314a8, 0x314a8, 0x314b0, 0x314b4, 0x314c8, 0x314d4, 0x31a40, 0x31a4c, 0x31af0, 0x31b20, 0x31b38, 0x31b3c, 0x31b80, 0x31b80, 0x31ba8, 0x31ba8, 0x31bb0, 0x31bb4, 0x31bc8, 0x31bd4, 0x32140, 0x3218c, 0x321f0, 0x321f4, 0x32200, 0x32200, 0x32218, 0x32218, 0x32400, 0x32400, 0x32408, 0x3241c, 0x32618, 0x32620, 0x32664, 0x32664, 0x326a8, 0x326a8, 0x326ec, 0x326ec, 0x32a00, 0x32abc, 0x32b00, 0x32b38, 0x32b40, 0x32b58, 0x32b60, 0x32b78, 0x32c00, 0x32c00, 0x32c08, 0x32c3c, 0x32e00, 0x32e2c, 0x32f00, 0x32f2c, 0x33000, 0x3302c, 0x33034, 0x33050, 0x33058, 0x33058, 0x33060, 0x3308c, 0x3309c, 0x330ac, 0x330c0, 0x330c0, 0x330c8, 0x330d0, 0x330d8, 0x330e0, 0x330ec, 0x3312c, 0x33134, 0x33150, 0x33158, 0x33158, 0x33160, 0x3318c, 0x3319c, 0x331ac, 0x331c0, 0x331c0, 0x331c8, 0x331d0, 0x331d8, 0x331e0, 0x331ec, 0x33290, 0x33298, 0x332c4, 0x332e4, 0x33390, 0x33398, 0x333c4, 0x333e4, 0x3342c, 0x33434, 0x33450, 0x33458, 0x33458, 0x33460, 0x3348c, 0x3349c, 0x334ac, 0x334c0, 0x334c0, 0x334c8, 0x334d0, 0x334d8, 0x334e0, 0x334ec, 0x3352c, 0x33534, 0x33550, 0x33558, 0x33558, 0x33560, 0x3358c, 0x3359c, 0x335ac, 0x335c0, 0x335c0, 0x335c8, 0x335d0, 0x335d8, 0x335e0, 0x335ec, 0x33690, 0x33698, 0x336c4, 0x336e4, 0x33790, 0x33798, 0x337c4, 0x337e4, 0x337fc, 0x33814, 0x33814, 0x33854, 0x33868, 0x33880, 0x3388c, 0x338c0, 0x338d0, 0x338e8, 0x338ec, 0x33900, 0x3392c, 0x33934, 0x33950, 0x33958, 0x33958, 0x33960, 0x3398c, 0x3399c, 0x339ac, 0x339c0, 0x339c0, 0x339c8, 0x339d0, 0x339d8, 0x339e0, 0x339ec, 0x33a90, 0x33a98, 0x33ac4, 0x33ae4, 0x33b10, 0x33b24, 0x33b28, 0x33b38, 0x33b50, 0x33bf0, 0x33c10, 0x33c24, 0x33c28, 0x33c38, 0x33c50, 0x33cf0, 0x33cfc, 0x34000, 0x34030, 0x34038, 0x34038, 0x34040, 0x34040, 0x34048, 0x34048, 0x34050, 0x34050, 0x3405c, 0x34060, 0x34068, 0x34068, 0x34070, 0x34070, 0x34100, 0x34168, 0x34190, 0x341a0, 0x341a8, 0x341b8, 0x341c4, 0x341c8, 0x341d0, 0x341d0, 0x34200, 0x34320, 0x34400, 0x344b4, 0x344c0, 0x3452c, 0x34540, 0x3461c, 0x34800, 0x348a0, 0x348c0, 0x34908, 0x34910, 0x349b8, 0x34a00, 0x34a04, 0x34a0c, 0x34a14, 0x34a1c, 0x34a2c, 0x34a44, 0x34a50, 0x34a74, 0x34a74, 0x34a7c, 0x34afc, 0x34b08, 0x34c24, 0x34d00, 0x34d14, 0x34d1c, 0x34d3c, 0x34d44, 0x34d4c, 0x34d54, 0x34d74, 0x34d7c, 0x34d7c, 0x34de0, 0x34de0, 0x34e00, 0x34ed4, 0x34f00, 0x34fa4, 0x34fc0, 0x34fc4, 0x35000, 0x35004, 0x35080, 0x350fc, 0x35208, 0x35220, 0x3523c, 0x35254, 0x35300, 0x35300, 0x35308, 0x3531c, 0x35338, 0x3533c, 0x35380, 0x35380, 0x35388, 0x353a8, 0x353b4, 0x353b4, 0x35400, 0x35420, 0x35438, 0x3543c, 0x35480, 0x35480, 0x354a8, 0x354a8, 0x354b0, 0x354b4, 0x354c8, 0x354d4, 0x35a40, 0x35a4c, 0x35af0, 0x35b20, 0x35b38, 0x35b3c, 0x35b80, 0x35b80, 0x35ba8, 0x35ba8, 0x35bb0, 0x35bb4, 0x35bc8, 0x35bd4, 0x36140, 0x3618c, 0x361f0, 0x361f4, 0x36200, 0x36200, 0x36218, 0x36218, 0x36400, 0x36400, 0x36408, 0x3641c, 0x36618, 0x36620, 0x36664, 0x36664, 0x366a8, 0x366a8, 0x366ec, 0x366ec, 0x36a00, 0x36abc, 0x36b00, 0x36b38, 0x36b40, 0x36b58, 0x36b60, 0x36b78, 0x36c00, 0x36c00, 0x36c08, 0x36c3c, 0x36e00, 0x36e2c, 0x36f00, 0x36f2c, 0x37000, 0x3702c, 0x37034, 0x37050, 0x37058, 0x37058, 0x37060, 0x3708c, 0x3709c, 0x370ac, 0x370c0, 0x370c0, 0x370c8, 0x370d0, 0x370d8, 0x370e0, 0x370ec, 0x3712c, 0x37134, 0x37150, 0x37158, 0x37158, 0x37160, 0x3718c, 0x3719c, 0x371ac, 0x371c0, 0x371c0, 0x371c8, 0x371d0, 0x371d8, 0x371e0, 0x371ec, 0x37290, 0x37298, 0x372c4, 0x372e4, 0x37390, 0x37398, 0x373c4, 0x373e4, 0x3742c, 0x37434, 0x37450, 0x37458, 0x37458, 0x37460, 0x3748c, 0x3749c, 0x374ac, 0x374c0, 0x374c0, 0x374c8, 0x374d0, 0x374d8, 0x374e0, 0x374ec, 0x3752c, 0x37534, 0x37550, 0x37558, 0x37558, 0x37560, 0x3758c, 0x3759c, 0x375ac, 0x375c0, 0x375c0, 0x375c8, 0x375d0, 0x375d8, 0x375e0, 0x375ec, 0x37690, 0x37698, 0x376c4, 0x376e4, 0x37790, 0x37798, 0x377c4, 0x377e4, 0x377fc, 0x37814, 0x37814, 0x37854, 0x37868, 0x37880, 0x3788c, 0x378c0, 0x378d0, 0x378e8, 0x378ec, 0x37900, 0x3792c, 0x37934, 0x37950, 0x37958, 0x37958, 0x37960, 0x3798c, 0x3799c, 0x379ac, 0x379c0, 0x379c0, 0x379c8, 0x379d0, 0x379d8, 0x379e0, 0x379ec, 0x37a90, 0x37a98, 0x37ac4, 0x37ae4, 0x37b10, 0x37b24, 0x37b28, 0x37b38, 0x37b50, 0x37bf0, 0x37c10, 0x37c24, 0x37c28, 0x37c38, 0x37c50, 0x37cf0, 0x37cfc, 0x40040, 0x40040, 0x40080, 0x40084, 0x40100, 0x40100, 0x40140, 0x401bc, 0x40200, 0x40214, 0x40228, 0x40228, 0x40240, 0x40258, 0x40280, 0x40280, 0x40304, 0x40304, 0x40330, 0x4033c, 0x41304, 0x413c8, 0x413d0, 0x413dc, 0x413f0, 0x413f0, 0x41400, 0x4140c, 0x41414, 0x4141c, 0x41480, 0x414d0, 0x44000, 0x4407c, 0x440c0, 0x441ac, 0x441b4, 0x4427c, 0x442c0, 0x443ac, 0x443b4, 0x4447c, 0x444c0, 0x445ac, 0x445b4, 0x4467c, 0x446c0, 0x447ac, 0x447b4, 0x4487c, 0x448c0, 0x449ac, 0x449b4, 0x44a7c, 0x44ac0, 0x44bac, 0x44bb4, 0x44c7c, 0x44cc0, 0x44dac, 0x44db4, 0x44e7c, 0x44ec0, 0x44fac, 0x44fb4, 0x4507c, 0x450c0, 0x451ac, 0x451b4, 0x451fc, 0x45800, 0x45804, 0x45810, 0x45830, 0x45840, 0x45860, 0x45868, 0x45868, 0x45880, 0x45884, 0x458a0, 0x458b0, 0x45a00, 0x45a04, 0x45a10, 0x45a30, 0x45a40, 0x45a60, 0x45a68, 0x45a68, 0x45a80, 0x45a84, 0x45aa0, 0x45ab0, 0x460c0, 0x460e4, 0x47000, 0x4703c, 0x47044, 0x4708c, 0x47200, 0x47250, 0x47400, 0x47408, 0x47414, 0x47420, 0x47600, 0x47618, 0x47800, 0x47814, 0x47820, 0x4782c, 0x50000, 0x50084, 0x50090, 0x500cc, 0x50300, 0x50384, 0x50400, 0x50400, 0x50800, 0x50884, 0x50890, 0x508cc, 0x50b00, 0x50b84, 0x50c00, 0x50c00, 0x51000, 0x51020, 0x51028, 0x510b0, 0x51300, 0x51324, }; static const unsigned int t6vf_reg_ranges[] = { VF_SGE_REG(A_SGE_VF_KDOORBELL), VF_SGE_REG(A_SGE_VF_GTS), VF_MPS_REG(A_MPS_VF_CTL), VF_MPS_REG(A_MPS_VF_STAT_RX_VF_ERR_FRAMES_H), VF_PL_REG(A_PL_VF_WHOAMI), VF_PL_REG(A_PL_VF_REVISION), VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_CTRL), VF_CIM_REG(A_CIM_VF_EXT_MAILBOX_STATUS), FW_T6VF_MBDATA_BASE_ADDR, FW_T6VF_MBDATA_BASE_ADDR + ((NUM_CIM_PF_MAILBOX_DATA_INSTANCES - 1) * 4), }; u32 *buf_end = (u32 *)(buf + buf_size); const unsigned int *reg_ranges; int reg_ranges_size, range; unsigned int chip_version = chip_id(adap); /* * Select the right set of register ranges to dump depending on the * adapter chip type. */ switch (chip_version) { case CHELSIO_T4: if (adap->flags & IS_VF) { reg_ranges = t4vf_reg_ranges; reg_ranges_size = ARRAY_SIZE(t4vf_reg_ranges); } else { reg_ranges = t4_reg_ranges; reg_ranges_size = ARRAY_SIZE(t4_reg_ranges); } break; case CHELSIO_T5: if (adap->flags & IS_VF) { reg_ranges = t5vf_reg_ranges; reg_ranges_size = ARRAY_SIZE(t5vf_reg_ranges); } else { reg_ranges = t5_reg_ranges; reg_ranges_size = ARRAY_SIZE(t5_reg_ranges); } break; case CHELSIO_T6: if (adap->flags & IS_VF) { reg_ranges = t6vf_reg_ranges; reg_ranges_size = ARRAY_SIZE(t6vf_reg_ranges); } else { reg_ranges = t6_reg_ranges; reg_ranges_size = ARRAY_SIZE(t6_reg_ranges); } break; default: CH_ERR(adap, "Unsupported chip version %d\n", chip_version); return; } /* * Clear the register buffer and insert the appropriate register * values selected by the above register ranges. */ memset(buf, 0, buf_size); for (range = 0; range < reg_ranges_size; range += 2) { unsigned int reg = reg_ranges[range]; unsigned int last_reg = reg_ranges[range + 1]; u32 *bufp = (u32 *)(buf + reg); /* * Iterate across the register range filling in the register * buffer but don't write past the end of the register buffer. */ while (reg <= last_reg && bufp < buf_end) { *bufp++ = t4_read_reg(adap, reg); reg += sizeof(u32); } } } /* * Partial EEPROM Vital Product Data structure. Includes only the ID and * VPD-R sections. */ struct t4_vpd_hdr { u8 id_tag; u8 id_len[2]; u8 id_data[ID_LEN]; u8 vpdr_tag; u8 vpdr_len[2]; }; /* * EEPROM reads take a few tens of us while writes can take a bit over 5 ms. */ #define EEPROM_DELAY 10 /* 10us per poll spin */ #define EEPROM_MAX_POLL 5000 /* x 5000 == 50ms */ #define EEPROM_STAT_ADDR 0x7bfc #define VPD_BASE 0x400 #define VPD_BASE_OLD 0 #define VPD_LEN 1024 #define VPD_INFO_FLD_HDR_SIZE 3 #define CHELSIO_VPD_UNIQUE_ID 0x82 /* * Small utility function to wait till any outstanding VPD Access is complete. * We have a per-adapter state variable "VPD Busy" to indicate when we have a * VPD Access in flight. This allows us to handle the problem of having a * previous VPD Access time out and prevent an attempt to inject a new VPD * Request before any in-flight VPD reguest has completed. */ static int t4_seeprom_wait(struct adapter *adapter) { unsigned int base = adapter->params.pci.vpd_cap_addr; int max_poll; /* * If no VPD Access is in flight, we can just return success right * away. */ if (!adapter->vpd_busy) return 0; /* * Poll the VPD Capability Address/Flag register waiting for it * to indicate that the operation is complete. */ max_poll = EEPROM_MAX_POLL; do { u16 val; udelay(EEPROM_DELAY); t4_os_pci_read_cfg2(adapter, base + PCI_VPD_ADDR, &val); /* * If the operation is complete, mark the VPD as no longer * busy and return success. */ if ((val & PCI_VPD_ADDR_F) == adapter->vpd_flag) { adapter->vpd_busy = 0; return 0; } } while (--max_poll); /* * Failure! Note that we leave the VPD Busy status set in order to * avoid pushing a new VPD Access request into the VPD Capability till * the current operation eventually succeeds. It's a bug to issue a * new request when an existing request is in flight and will result * in corrupt hardware state. */ return -ETIMEDOUT; } /** * t4_seeprom_read - read a serial EEPROM location * @adapter: adapter to read * @addr: EEPROM virtual address * @data: where to store the read data * * Read a 32-bit word from a location in serial EEPROM using the card's PCI * VPD capability. Note that this function must be called with a virtual * address. */ int t4_seeprom_read(struct adapter *adapter, u32 addr, u32 *data) { unsigned int base = adapter->params.pci.vpd_cap_addr; int ret; /* * VPD Accesses must alway be 4-byte aligned! */ if (addr >= EEPROMVSIZE || (addr & 3)) return -EINVAL; /* * Wait for any previous operation which may still be in flight to * complete. */ ret = t4_seeprom_wait(adapter); if (ret) { CH_ERR(adapter, "VPD still busy from previous operation\n"); return ret; } /* * Issue our new VPD Read request, mark the VPD as being busy and wait * for our request to complete. If it doesn't complete, note the * error and return it to our caller. Note that we do not reset the * VPD Busy status! */ t4_os_pci_write_cfg2(adapter, base + PCI_VPD_ADDR, (u16)addr); adapter->vpd_busy = 1; adapter->vpd_flag = PCI_VPD_ADDR_F; ret = t4_seeprom_wait(adapter); if (ret) { CH_ERR(adapter, "VPD read of address %#x failed\n", addr); return ret; } /* * Grab the returned data, swizzle it into our endianness and * return success. */ t4_os_pci_read_cfg4(adapter, base + PCI_VPD_DATA, data); *data = le32_to_cpu(*data); return 0; } /** * t4_seeprom_write - write a serial EEPROM location * @adapter: adapter to write * @addr: virtual EEPROM address * @data: value to write * * Write a 32-bit word to a location in serial EEPROM using the card's PCI * VPD capability. Note that this function must be called with a virtual * address. */ int t4_seeprom_write(struct adapter *adapter, u32 addr, u32 data) { unsigned int base = adapter->params.pci.vpd_cap_addr; int ret; u32 stats_reg; int max_poll; /* * VPD Accesses must alway be 4-byte aligned! */ if (addr >= EEPROMVSIZE || (addr & 3)) return -EINVAL; /* * Wait for any previous operation which may still be in flight to * complete. */ ret = t4_seeprom_wait(adapter); if (ret) { CH_ERR(adapter, "VPD still busy from previous operation\n"); return ret; } /* * Issue our new VPD Read request, mark the VPD as being busy and wait * for our request to complete. If it doesn't complete, note the * error and return it to our caller. Note that we do not reset the * VPD Busy status! */ t4_os_pci_write_cfg4(adapter, base + PCI_VPD_DATA, cpu_to_le32(data)); t4_os_pci_write_cfg2(adapter, base + PCI_VPD_ADDR, (u16)addr | PCI_VPD_ADDR_F); adapter->vpd_busy = 1; adapter->vpd_flag = 0; ret = t4_seeprom_wait(adapter); if (ret) { CH_ERR(adapter, "VPD write of address %#x failed\n", addr); return ret; } /* * Reset PCI_VPD_DATA register after a transaction and wait for our * request to complete. If it doesn't complete, return error. */ t4_os_pci_write_cfg4(adapter, base + PCI_VPD_DATA, 0); max_poll = EEPROM_MAX_POLL; do { udelay(EEPROM_DELAY); t4_seeprom_read(adapter, EEPROM_STAT_ADDR, &stats_reg); } while ((stats_reg & 0x1) && --max_poll); if (!max_poll) return -ETIMEDOUT; /* Return success! */ return 0; } /** * t4_eeprom_ptov - translate a physical EEPROM address to virtual * @phys_addr: the physical EEPROM address * @fn: the PCI function number * @sz: size of function-specific area * * Translate a physical EEPROM address to virtual. The first 1K is * accessed through virtual addresses starting at 31K, the rest is * accessed through virtual addresses starting at 0. * * The mapping is as follows: * [0..1K) -> [31K..32K) * [1K..1K+A) -> [ES-A..ES) * [1K+A..ES) -> [0..ES-A-1K) * * where A = @fn * @sz, and ES = EEPROM size. */ int t4_eeprom_ptov(unsigned int phys_addr, unsigned int fn, unsigned int sz) { fn *= sz; if (phys_addr < 1024) return phys_addr + (31 << 10); if (phys_addr < 1024 + fn) return EEPROMSIZE - fn + phys_addr - 1024; if (phys_addr < EEPROMSIZE) return phys_addr - 1024 - fn; return -EINVAL; } /** * t4_seeprom_wp - enable/disable EEPROM write protection * @adapter: the adapter * @enable: whether to enable or disable write protection * * Enables or disables write protection on the serial EEPROM. */ int t4_seeprom_wp(struct adapter *adapter, int enable) { return t4_seeprom_write(adapter, EEPROM_STAT_ADDR, enable ? 0xc : 0); } /** * get_vpd_keyword_val - Locates an information field keyword in the VPD * @v: Pointer to buffered vpd data structure * @kw: The keyword to search for * * Returns the value of the information field keyword or * -ENOENT otherwise. */ static int get_vpd_keyword_val(const struct t4_vpd_hdr *v, const char *kw) { int i; unsigned int offset , len; const u8 *buf = (const u8 *)v; const u8 *vpdr_len = &v->vpdr_len[0]; offset = sizeof(struct t4_vpd_hdr); len = (u16)vpdr_len[0] + ((u16)vpdr_len[1] << 8); if (len + sizeof(struct t4_vpd_hdr) > VPD_LEN) { return -ENOENT; } for (i = offset; i + VPD_INFO_FLD_HDR_SIZE <= offset + len;) { if(memcmp(buf + i , kw , 2) == 0){ i += VPD_INFO_FLD_HDR_SIZE; return i; } i += VPD_INFO_FLD_HDR_SIZE + buf[i+2]; } return -ENOENT; } /** * get_vpd_params - read VPD parameters from VPD EEPROM * @adapter: adapter to read * @p: where to store the parameters * @vpd: caller provided temporary space to read the VPD into * * Reads card parameters stored in VPD EEPROM. */ static int get_vpd_params(struct adapter *adapter, struct vpd_params *p, u8 *vpd) { int i, ret, addr; int ec, sn, pn, na; u8 csum; const struct t4_vpd_hdr *v; /* * Card information normally starts at VPD_BASE but early cards had * it at 0. */ ret = t4_seeprom_read(adapter, VPD_BASE, (u32 *)(vpd)); if (ret) return (ret); /* * The VPD shall have a unique identifier specified by the PCI SIG. * For chelsio adapters, the identifier is 0x82. The first byte of a VPD * shall be CHELSIO_VPD_UNIQUE_ID (0x82). The VPD programming software * is expected to automatically put this entry at the * beginning of the VPD. */ addr = *vpd == CHELSIO_VPD_UNIQUE_ID ? VPD_BASE : VPD_BASE_OLD; for (i = 0; i < VPD_LEN; i += 4) { ret = t4_seeprom_read(adapter, addr + i, (u32 *)(vpd + i)); if (ret) return ret; } v = (const struct t4_vpd_hdr *)vpd; #define FIND_VPD_KW(var,name) do { \ var = get_vpd_keyword_val(v , name); \ if (var < 0) { \ CH_ERR(adapter, "missing VPD keyword " name "\n"); \ return -EINVAL; \ } \ } while (0) FIND_VPD_KW(i, "RV"); for (csum = 0; i >= 0; i--) csum += vpd[i]; if (csum) { CH_ERR(adapter, "corrupted VPD EEPROM, actual csum %u\n", csum); return -EINVAL; } FIND_VPD_KW(ec, "EC"); FIND_VPD_KW(sn, "SN"); FIND_VPD_KW(pn, "PN"); FIND_VPD_KW(na, "NA"); #undef FIND_VPD_KW memcpy(p->id, v->id_data, ID_LEN); strstrip(p->id); memcpy(p->ec, vpd + ec, EC_LEN); strstrip(p->ec); i = vpd[sn - VPD_INFO_FLD_HDR_SIZE + 2]; memcpy(p->sn, vpd + sn, min(i, SERNUM_LEN)); strstrip(p->sn); i = vpd[pn - VPD_INFO_FLD_HDR_SIZE + 2]; memcpy(p->pn, vpd + pn, min(i, PN_LEN)); strstrip((char *)p->pn); i = vpd[na - VPD_INFO_FLD_HDR_SIZE + 2]; memcpy(p->na, vpd + na, min(i, MACADDR_LEN)); strstrip((char *)p->na); return 0; } /* serial flash and firmware constants and flash config file constants */ enum { SF_ATTEMPTS = 10, /* max retries for SF operations */ /* flash command opcodes */ SF_PROG_PAGE = 2, /* program page */ SF_WR_DISABLE = 4, /* disable writes */ SF_RD_STATUS = 5, /* read status register */ SF_WR_ENABLE = 6, /* enable writes */ SF_RD_DATA_FAST = 0xb, /* read flash */ SF_RD_ID = 0x9f, /* read ID */ SF_ERASE_SECTOR = 0xd8, /* erase sector */ }; /** * sf1_read - read data from the serial flash * @adapter: the adapter * @byte_cnt: number of bytes to read * @cont: whether another operation will be chained * @lock: whether to lock SF for PL access only * @valp: where to store the read data * * Reads up to 4 bytes of data from the serial flash. The location of * the read needs to be specified prior to calling this by issuing the * appropriate commands to the serial flash. */ static int sf1_read(struct adapter *adapter, unsigned int byte_cnt, int cont, int lock, u32 *valp) { int ret; if (!byte_cnt || byte_cnt > 4) return -EINVAL; if (t4_read_reg(adapter, A_SF_OP) & F_BUSY) return -EBUSY; t4_write_reg(adapter, A_SF_OP, V_SF_LOCK(lock) | V_CONT(cont) | V_BYTECNT(byte_cnt - 1)); ret = t4_wait_op_done(adapter, A_SF_OP, F_BUSY, 0, SF_ATTEMPTS, 5); if (!ret) *valp = t4_read_reg(adapter, A_SF_DATA); return ret; } /** * sf1_write - write data to the serial flash * @adapter: the adapter * @byte_cnt: number of bytes to write * @cont: whether another operation will be chained * @lock: whether to lock SF for PL access only * @val: value to write * * Writes up to 4 bytes of data to the serial flash. The location of * the write needs to be specified prior to calling this by issuing the * appropriate commands to the serial flash. */ static int sf1_write(struct adapter *adapter, unsigned int byte_cnt, int cont, int lock, u32 val) { if (!byte_cnt || byte_cnt > 4) return -EINVAL; if (t4_read_reg(adapter, A_SF_OP) & F_BUSY) return -EBUSY; t4_write_reg(adapter, A_SF_DATA, val); t4_write_reg(adapter, A_SF_OP, V_SF_LOCK(lock) | V_CONT(cont) | V_BYTECNT(byte_cnt - 1) | V_OP(1)); return t4_wait_op_done(adapter, A_SF_OP, F_BUSY, 0, SF_ATTEMPTS, 5); } /** * flash_wait_op - wait for a flash operation to complete * @adapter: the adapter * @attempts: max number of polls of the status register * @delay: delay between polls in ms * * Wait for a flash operation to complete by polling the status register. */ static int flash_wait_op(struct adapter *adapter, int attempts, int delay) { int ret; u32 status; while (1) { if ((ret = sf1_write(adapter, 1, 1, 1, SF_RD_STATUS)) != 0 || (ret = sf1_read(adapter, 1, 0, 1, &status)) != 0) return ret; if (!(status & 1)) return 0; if (--attempts == 0) return -EAGAIN; if (delay) msleep(delay); } } /** * t4_read_flash - read words from serial flash * @adapter: the adapter * @addr: the start address for the read * @nwords: how many 32-bit words to read * @data: where to store the read data * @byte_oriented: whether to store data as bytes or as words * * Read the specified number of 32-bit words from the serial flash. * If @byte_oriented is set the read data is stored as a byte array * (i.e., big-endian), otherwise as 32-bit words in the platform's * natural endianness. */ int t4_read_flash(struct adapter *adapter, unsigned int addr, unsigned int nwords, u32 *data, int byte_oriented) { int ret; if (addr + nwords * sizeof(u32) > adapter->params.sf_size || (addr & 3)) return -EINVAL; addr = swab32(addr) | SF_RD_DATA_FAST; if ((ret = sf1_write(adapter, 4, 1, 0, addr)) != 0 || (ret = sf1_read(adapter, 1, 1, 0, data)) != 0) return ret; for ( ; nwords; nwords--, data++) { ret = sf1_read(adapter, 4, nwords > 1, nwords == 1, data); if (nwords == 1) t4_write_reg(adapter, A_SF_OP, 0); /* unlock SF */ if (ret) return ret; if (byte_oriented) *data = (__force __u32)(cpu_to_be32(*data)); } return 0; } /** * t4_write_flash - write up to a page of data to the serial flash * @adapter: the adapter * @addr: the start address to write * @n: length of data to write in bytes * @data: the data to write * @byte_oriented: whether to store data as bytes or as words * * Writes up to a page of data (256 bytes) to the serial flash starting * at the given address. All the data must be written to the same page. * If @byte_oriented is set the write data is stored as byte stream * (i.e. matches what on disk), otherwise in big-endian. */ int t4_write_flash(struct adapter *adapter, unsigned int addr, unsigned int n, const u8 *data, int byte_oriented) { int ret; u32 buf[SF_PAGE_SIZE / 4]; unsigned int i, c, left, val, offset = addr & 0xff; if (addr >= adapter->params.sf_size || offset + n > SF_PAGE_SIZE) return -EINVAL; val = swab32(addr) | SF_PROG_PAGE; if ((ret = sf1_write(adapter, 1, 0, 1, SF_WR_ENABLE)) != 0 || (ret = sf1_write(adapter, 4, 1, 1, val)) != 0) goto unlock; for (left = n; left; left -= c) { c = min(left, 4U); for (val = 0, i = 0; i < c; ++i) val = (val << 8) + *data++; if (!byte_oriented) val = cpu_to_be32(val); ret = sf1_write(adapter, c, c != left, 1, val); if (ret) goto unlock; } ret = flash_wait_op(adapter, 8, 1); if (ret) goto unlock; t4_write_reg(adapter, A_SF_OP, 0); /* unlock SF */ /* Read the page to verify the write succeeded */ ret = t4_read_flash(adapter, addr & ~0xff, ARRAY_SIZE(buf), buf, byte_oriented); if (ret) return ret; if (memcmp(data - n, (u8 *)buf + offset, n)) { CH_ERR(adapter, "failed to correctly write the flash page at %#x\n", addr); return -EIO; } return 0; unlock: t4_write_reg(adapter, A_SF_OP, 0); /* unlock SF */ return ret; } /** * t4_get_fw_version - read the firmware version * @adapter: the adapter * @vers: where to place the version * * Reads the FW version from flash. */ int t4_get_fw_version(struct adapter *adapter, u32 *vers) { return t4_read_flash(adapter, FLASH_FW_START + offsetof(struct fw_hdr, fw_ver), 1, vers, 0); } /** * t4_get_bs_version - read the firmware bootstrap version * @adapter: the adapter * @vers: where to place the version * * Reads the FW Bootstrap version from flash. */ int t4_get_bs_version(struct adapter *adapter, u32 *vers) { return t4_read_flash(adapter, FLASH_FWBOOTSTRAP_START + offsetof(struct fw_hdr, fw_ver), 1, vers, 0); } /** * t4_get_tp_version - read the TP microcode version * @adapter: the adapter * @vers: where to place the version * * Reads the TP microcode version from flash. */ int t4_get_tp_version(struct adapter *adapter, u32 *vers) { return t4_read_flash(adapter, FLASH_FW_START + offsetof(struct fw_hdr, tp_microcode_ver), 1, vers, 0); } /** * t4_get_exprom_version - return the Expansion ROM version (if any) * @adapter: the adapter * @vers: where to place the version * * Reads the Expansion ROM header from FLASH and returns the version * number (if present) through the @vers return value pointer. We return * this in the Firmware Version Format since it's convenient. Return * 0 on success, -ENOENT if no Expansion ROM is present. */ int t4_get_exprom_version(struct adapter *adap, u32 *vers) { struct exprom_header { unsigned char hdr_arr[16]; /* must start with 0x55aa */ unsigned char hdr_ver[4]; /* Expansion ROM version */ } *hdr; u32 exprom_header_buf[DIV_ROUND_UP(sizeof(struct exprom_header), sizeof(u32))]; int ret; ret = t4_read_flash(adap, FLASH_EXP_ROM_START, ARRAY_SIZE(exprom_header_buf), exprom_header_buf, 0); if (ret) return ret; hdr = (struct exprom_header *)exprom_header_buf; if (hdr->hdr_arr[0] != 0x55 || hdr->hdr_arr[1] != 0xaa) return -ENOENT; *vers = (V_FW_HDR_FW_VER_MAJOR(hdr->hdr_ver[0]) | V_FW_HDR_FW_VER_MINOR(hdr->hdr_ver[1]) | V_FW_HDR_FW_VER_MICRO(hdr->hdr_ver[2]) | V_FW_HDR_FW_VER_BUILD(hdr->hdr_ver[3])); return 0; } /** * t4_get_scfg_version - return the Serial Configuration version * @adapter: the adapter * @vers: where to place the version * * Reads the Serial Configuration Version via the Firmware interface * (thus this can only be called once we're ready to issue Firmware * commands). The format of the Serial Configuration version is * adapter specific. Returns 0 on success, an error on failure. * * Note that early versions of the Firmware didn't include the ability * to retrieve the Serial Configuration version, so we zero-out the * return-value parameter in that case to avoid leaving it with * garbage in it. * * Also note that the Firmware will return its cached copy of the Serial * Initialization Revision ID, not the actual Revision ID as written in * the Serial EEPROM. This is only an issue if a new VPD has been written * and the Firmware/Chip haven't yet gone through a RESET sequence. So * it's best to defer calling this routine till after a FW_RESET_CMD has * been issued if the Host Driver will be performing a full adapter * initialization. */ int t4_get_scfg_version(struct adapter *adapter, u32 *vers) { u32 scfgrev_param; int ret; scfgrev_param = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_SCFGREV)); ret = t4_query_params(adapter, adapter->mbox, adapter->pf, 0, 1, &scfgrev_param, vers); if (ret) *vers = 0; return ret; } /** * t4_get_vpd_version - return the VPD version * @adapter: the adapter * @vers: where to place the version * * Reads the VPD via the Firmware interface (thus this can only be called * once we're ready to issue Firmware commands). The format of the * VPD version is adapter specific. Returns 0 on success, an error on * failure. * * Note that early versions of the Firmware didn't include the ability * to retrieve the VPD version, so we zero-out the return-value parameter * in that case to avoid leaving it with garbage in it. * * Also note that the Firmware will return its cached copy of the VPD * Revision ID, not the actual Revision ID as written in the Serial * EEPROM. This is only an issue if a new VPD has been written and the * Firmware/Chip haven't yet gone through a RESET sequence. So it's best * to defer calling this routine till after a FW_RESET_CMD has been issued * if the Host Driver will be performing a full adapter initialization. */ int t4_get_vpd_version(struct adapter *adapter, u32 *vers) { u32 vpdrev_param; int ret; vpdrev_param = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_VPDREV)); ret = t4_query_params(adapter, adapter->mbox, adapter->pf, 0, 1, &vpdrev_param, vers); if (ret) *vers = 0; return ret; } /** * t4_get_version_info - extract various chip/firmware version information * @adapter: the adapter * * Reads various chip/firmware version numbers and stores them into the * adapter Adapter Parameters structure. If any of the efforts fails * the first failure will be returned, but all of the version numbers * will be read. */ int t4_get_version_info(struct adapter *adapter) { int ret = 0; #define FIRST_RET(__getvinfo) \ do { \ int __ret = __getvinfo; \ if (__ret && !ret) \ ret = __ret; \ } while (0) FIRST_RET(t4_get_fw_version(adapter, &adapter->params.fw_vers)); FIRST_RET(t4_get_bs_version(adapter, &adapter->params.bs_vers)); FIRST_RET(t4_get_tp_version(adapter, &adapter->params.tp_vers)); FIRST_RET(t4_get_exprom_version(adapter, &adapter->params.er_vers)); FIRST_RET(t4_get_scfg_version(adapter, &adapter->params.scfg_vers)); FIRST_RET(t4_get_vpd_version(adapter, &adapter->params.vpd_vers)); #undef FIRST_RET return ret; } /** * t4_flash_erase_sectors - erase a range of flash sectors * @adapter: the adapter * @start: the first sector to erase * @end: the last sector to erase * * Erases the sectors in the given inclusive range. */ int t4_flash_erase_sectors(struct adapter *adapter, int start, int end) { int ret = 0; if (end >= adapter->params.sf_nsec) return -EINVAL; while (start <= end) { if ((ret = sf1_write(adapter, 1, 0, 1, SF_WR_ENABLE)) != 0 || (ret = sf1_write(adapter, 4, 0, 1, SF_ERASE_SECTOR | (start << 8))) != 0 || (ret = flash_wait_op(adapter, 14, 500)) != 0) { CH_ERR(adapter, "erase of flash sector %d failed, error %d\n", start, ret); break; } start++; } t4_write_reg(adapter, A_SF_OP, 0); /* unlock SF */ return ret; } /** * t4_flash_cfg_addr - return the address of the flash configuration file * @adapter: the adapter * * Return the address within the flash where the Firmware Configuration * File is stored, or an error if the device FLASH is too small to contain * a Firmware Configuration File. */ int t4_flash_cfg_addr(struct adapter *adapter) { /* * If the device FLASH isn't large enough to hold a Firmware * Configuration File, return an error. */ if (adapter->params.sf_size < FLASH_CFG_START + FLASH_CFG_MAX_SIZE) return -ENOSPC; return FLASH_CFG_START; } /* * Return TRUE if the specified firmware matches the adapter. I.e. T4 * firmware for T4 adapters, T5 firmware for T5 adapters, etc. We go ahead * and emit an error message for mismatched firmware to save our caller the * effort ... */ static int t4_fw_matches_chip(struct adapter *adap, const struct fw_hdr *hdr) { /* * The expression below will return FALSE for any unsupported adapter * which will keep us "honest" in the future ... */ if ((is_t4(adap) && hdr->chip == FW_HDR_CHIP_T4) || (is_t5(adap) && hdr->chip == FW_HDR_CHIP_T5) || (is_t6(adap) && hdr->chip == FW_HDR_CHIP_T6)) return 1; CH_ERR(adap, "FW image (%d) is not suitable for this adapter (%d)\n", hdr->chip, chip_id(adap)); return 0; } /** * t4_load_fw - download firmware * @adap: the adapter * @fw_data: the firmware image to write * @size: image size * * Write the supplied firmware image to the card's serial flash. */ int t4_load_fw(struct adapter *adap, const u8 *fw_data, unsigned int size) { u32 csum; int ret, addr; unsigned int i; u8 first_page[SF_PAGE_SIZE]; const u32 *p = (const u32 *)fw_data; const struct fw_hdr *hdr = (const struct fw_hdr *)fw_data; unsigned int sf_sec_size = adap->params.sf_size / adap->params.sf_nsec; unsigned int fw_start_sec; unsigned int fw_start; unsigned int fw_size; if (ntohl(hdr->magic) == FW_HDR_MAGIC_BOOTSTRAP) { fw_start_sec = FLASH_FWBOOTSTRAP_START_SEC; fw_start = FLASH_FWBOOTSTRAP_START; fw_size = FLASH_FWBOOTSTRAP_MAX_SIZE; } else { fw_start_sec = FLASH_FW_START_SEC; fw_start = FLASH_FW_START; fw_size = FLASH_FW_MAX_SIZE; } if (!size) { CH_ERR(adap, "FW image has no data\n"); return -EINVAL; } if (size & 511) { CH_ERR(adap, "FW image size not multiple of 512 bytes\n"); return -EINVAL; } if ((unsigned int) be16_to_cpu(hdr->len512) * 512 != size) { CH_ERR(adap, "FW image size differs from size in FW header\n"); return -EINVAL; } if (size > fw_size) { CH_ERR(adap, "FW image too large, max is %u bytes\n", fw_size); return -EFBIG; } if (!t4_fw_matches_chip(adap, hdr)) return -EINVAL; for (csum = 0, i = 0; i < size / sizeof(csum); i++) csum += be32_to_cpu(p[i]); if (csum != 0xffffffff) { CH_ERR(adap, "corrupted firmware image, checksum %#x\n", csum); return -EINVAL; } i = DIV_ROUND_UP(size, sf_sec_size); /* # of sectors spanned */ ret = t4_flash_erase_sectors(adap, fw_start_sec, fw_start_sec + i - 1); if (ret) goto out; /* * We write the correct version at the end so the driver can see a bad * version if the FW write fails. Start by writing a copy of the * first page with a bad version. */ memcpy(first_page, fw_data, SF_PAGE_SIZE); ((struct fw_hdr *)first_page)->fw_ver = cpu_to_be32(0xffffffff); ret = t4_write_flash(adap, fw_start, SF_PAGE_SIZE, first_page, 1); if (ret) goto out; addr = fw_start; for (size -= SF_PAGE_SIZE; size; size -= SF_PAGE_SIZE) { addr += SF_PAGE_SIZE; fw_data += SF_PAGE_SIZE; ret = t4_write_flash(adap, addr, SF_PAGE_SIZE, fw_data, 1); if (ret) goto out; } ret = t4_write_flash(adap, fw_start + offsetof(struct fw_hdr, fw_ver), sizeof(hdr->fw_ver), (const u8 *)&hdr->fw_ver, 1); out: if (ret) CH_ERR(adap, "firmware download failed, error %d\n", ret); return ret; } /** * t4_fwcache - firmware cache operation * @adap: the adapter * @op : the operation (flush or flush and invalidate) */ int t4_fwcache(struct adapter *adap, enum fw_params_param_dev_fwcache op) { struct fw_params_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_PARAMS_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_PARAMS_CMD_PFN(adap->pf) | V_FW_PARAMS_CMD_VFN(0)); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); c.param[0].mnem = cpu_to_be32(V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_FWCACHE)); c.param[0].val = (__force __be32)op; return t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), NULL); } void t4_cim_read_pif_la(struct adapter *adap, u32 *pif_req, u32 *pif_rsp, unsigned int *pif_req_wrptr, unsigned int *pif_rsp_wrptr) { int i, j; u32 cfg, val, req, rsp; cfg = t4_read_reg(adap, A_CIM_DEBUGCFG); if (cfg & F_LADBGEN) t4_write_reg(adap, A_CIM_DEBUGCFG, cfg ^ F_LADBGEN); val = t4_read_reg(adap, A_CIM_DEBUGSTS); req = G_POLADBGWRPTR(val); rsp = G_PILADBGWRPTR(val); if (pif_req_wrptr) *pif_req_wrptr = req; if (pif_rsp_wrptr) *pif_rsp_wrptr = rsp; for (i = 0; i < CIM_PIFLA_SIZE; i++) { for (j = 0; j < 6; j++) { t4_write_reg(adap, A_CIM_DEBUGCFG, V_POLADBGRDPTR(req) | V_PILADBGRDPTR(rsp)); *pif_req++ = t4_read_reg(adap, A_CIM_PO_LA_DEBUGDATA); *pif_rsp++ = t4_read_reg(adap, A_CIM_PI_LA_DEBUGDATA); req++; rsp++; } req = (req + 2) & M_POLADBGRDPTR; rsp = (rsp + 2) & M_PILADBGRDPTR; } t4_write_reg(adap, A_CIM_DEBUGCFG, cfg); } void t4_cim_read_ma_la(struct adapter *adap, u32 *ma_req, u32 *ma_rsp) { u32 cfg; int i, j, idx; cfg = t4_read_reg(adap, A_CIM_DEBUGCFG); if (cfg & F_LADBGEN) t4_write_reg(adap, A_CIM_DEBUGCFG, cfg ^ F_LADBGEN); for (i = 0; i < CIM_MALA_SIZE; i++) { for (j = 0; j < 5; j++) { idx = 8 * i + j; t4_write_reg(adap, A_CIM_DEBUGCFG, V_POLADBGRDPTR(idx) | V_PILADBGRDPTR(idx)); *ma_req++ = t4_read_reg(adap, A_CIM_PO_LA_MADEBUGDATA); *ma_rsp++ = t4_read_reg(adap, A_CIM_PI_LA_MADEBUGDATA); } } t4_write_reg(adap, A_CIM_DEBUGCFG, cfg); } void t4_ulprx_read_la(struct adapter *adap, u32 *la_buf) { unsigned int i, j; for (i = 0; i < 8; i++) { u32 *p = la_buf + i; t4_write_reg(adap, A_ULP_RX_LA_CTL, i); j = t4_read_reg(adap, A_ULP_RX_LA_WRPTR); t4_write_reg(adap, A_ULP_RX_LA_RDPTR, j); for (j = 0; j < ULPRX_LA_SIZE; j++, p += 8) *p = t4_read_reg(adap, A_ULP_RX_LA_RDDATA); } } #define ADVERT_MASK (FW_PORT_CAP_SPEED_100M | FW_PORT_CAP_SPEED_1G |\ FW_PORT_CAP_SPEED_10G | FW_PORT_CAP_SPEED_40G | \ FW_PORT_CAP_SPEED_100G | FW_PORT_CAP_ANEG) /** * t4_link_l1cfg - apply link configuration to MAC/PHY * @phy: the PHY to setup * @mac: the MAC to setup * @lc: the requested link configuration * * Set up a port's MAC and PHY according to a desired link configuration. * - If the PHY can auto-negotiate first decide what to advertise, then * enable/disable auto-negotiation as desired, and reset. * - If the PHY does not auto-negotiate just reset it. * - If auto-negotiation is off set the MAC to the proper speed/duplex/FC, * otherwise do it later based on the outcome of auto-negotiation. */ int t4_link_l1cfg(struct adapter *adap, unsigned int mbox, unsigned int port, struct link_config *lc) { struct fw_port_cmd c; unsigned int fc = 0, mdi = V_FW_PORT_CAP_MDI(FW_PORT_CAP_MDI_AUTO); lc->link_ok = 0; if (lc->requested_fc & PAUSE_RX) fc |= FW_PORT_CAP_FC_RX; if (lc->requested_fc & PAUSE_TX) fc |= FW_PORT_CAP_FC_TX; memset(&c, 0, sizeof(c)); c.op_to_portid = cpu_to_be32(V_FW_CMD_OP(FW_PORT_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_PORT_CMD_PORTID(port)); c.action_to_len16 = cpu_to_be32(V_FW_PORT_CMD_ACTION(FW_PORT_ACTION_L1_CFG) | FW_LEN16(c)); if (!(lc->supported & FW_PORT_CAP_ANEG)) { c.u.l1cfg.rcap = cpu_to_be32((lc->supported & ADVERT_MASK) | fc); lc->fc = lc->requested_fc & (PAUSE_RX | PAUSE_TX); } else if (lc->autoneg == AUTONEG_DISABLE) { c.u.l1cfg.rcap = cpu_to_be32(lc->requested_speed | fc | mdi); lc->fc = lc->requested_fc & (PAUSE_RX | PAUSE_TX); } else c.u.l1cfg.rcap = cpu_to_be32(lc->advertising | fc | mdi); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_restart_aneg - restart autonegotiation * @adap: the adapter * @mbox: mbox to use for the FW command * @port: the port id * * Restarts autonegotiation for the selected port. */ int t4_restart_aneg(struct adapter *adap, unsigned int mbox, unsigned int port) { struct fw_port_cmd c; memset(&c, 0, sizeof(c)); c.op_to_portid = cpu_to_be32(V_FW_CMD_OP(FW_PORT_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_PORT_CMD_PORTID(port)); c.action_to_len16 = cpu_to_be32(V_FW_PORT_CMD_ACTION(FW_PORT_ACTION_L1_CFG) | FW_LEN16(c)); c.u.l1cfg.rcap = cpu_to_be32(FW_PORT_CAP_ANEG); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } typedef void (*int_handler_t)(struct adapter *adap); struct intr_info { unsigned int mask; /* bits to check in interrupt status */ const char *msg; /* message to print or NULL */ short stat_idx; /* stat counter to increment or -1 */ unsigned short fatal; /* whether the condition reported is fatal */ int_handler_t int_handler; /* platform-specific int handler */ }; /** * t4_handle_intr_status - table driven interrupt handler * @adapter: the adapter that generated the interrupt * @reg: the interrupt status register to process * @acts: table of interrupt actions * * A table driven interrupt handler that applies a set of masks to an * interrupt status word and performs the corresponding actions if the * interrupts described by the mask have occurred. The actions include * optionally emitting a warning or alert message. The table is terminated * by an entry specifying mask 0. Returns the number of fatal interrupt * conditions. */ static int t4_handle_intr_status(struct adapter *adapter, unsigned int reg, const struct intr_info *acts) { int fatal = 0; unsigned int mask = 0; unsigned int status = t4_read_reg(adapter, reg); for ( ; acts->mask; ++acts) { if (!(status & acts->mask)) continue; if (acts->fatal) { fatal++; CH_ALERT(adapter, "%s (0x%x)\n", acts->msg, status & acts->mask); } else if (acts->msg) CH_WARN_RATELIMIT(adapter, "%s (0x%x)\n", acts->msg, status & acts->mask); if (acts->int_handler) acts->int_handler(adapter); mask |= acts->mask; } status &= mask; if (status) /* clear processed interrupts */ t4_write_reg(adapter, reg, status); return fatal; } /* * Interrupt handler for the PCIE module. */ static void pcie_intr_handler(struct adapter *adapter) { static const struct intr_info sysbus_intr_info[] = { { F_RNPP, "RXNP array parity error", -1, 1 }, { F_RPCP, "RXPC array parity error", -1, 1 }, { F_RCIP, "RXCIF array parity error", -1, 1 }, { F_RCCP, "Rx completions control array parity error", -1, 1 }, { F_RFTP, "RXFT array parity error", -1, 1 }, { 0 } }; static const struct intr_info pcie_port_intr_info[] = { { F_TPCP, "TXPC array parity error", -1, 1 }, { F_TNPP, "TXNP array parity error", -1, 1 }, { F_TFTP, "TXFT array parity error", -1, 1 }, { F_TCAP, "TXCA array parity error", -1, 1 }, { F_TCIP, "TXCIF array parity error", -1, 1 }, { F_RCAP, "RXCA array parity error", -1, 1 }, { F_OTDD, "outbound request TLP discarded", -1, 1 }, { F_RDPE, "Rx data parity error", -1, 1 }, { F_TDUE, "Tx uncorrectable data error", -1, 1 }, { 0 } }; static const struct intr_info pcie_intr_info[] = { { F_MSIADDRLPERR, "MSI AddrL parity error", -1, 1 }, { F_MSIADDRHPERR, "MSI AddrH parity error", -1, 1 }, { F_MSIDATAPERR, "MSI data parity error", -1, 1 }, { F_MSIXADDRLPERR, "MSI-X AddrL parity error", -1, 1 }, { F_MSIXADDRHPERR, "MSI-X AddrH parity error", -1, 1 }, { F_MSIXDATAPERR, "MSI-X data parity error", -1, 1 }, { F_MSIXDIPERR, "MSI-X DI parity error", -1, 1 }, { F_PIOCPLPERR, "PCI PIO completion FIFO parity error", -1, 1 }, { F_PIOREQPERR, "PCI PIO request FIFO parity error", -1, 1 }, { F_TARTAGPERR, "PCI PCI target tag FIFO parity error", -1, 1 }, { F_CCNTPERR, "PCI CMD channel count parity error", -1, 1 }, { F_CREQPERR, "PCI CMD channel request parity error", -1, 1 }, { F_CRSPPERR, "PCI CMD channel response parity error", -1, 1 }, { F_DCNTPERR, "PCI DMA channel count parity error", -1, 1 }, { F_DREQPERR, "PCI DMA channel request parity error", -1, 1 }, { F_DRSPPERR, "PCI DMA channel response parity error", -1, 1 }, { F_HCNTPERR, "PCI HMA channel count parity error", -1, 1 }, { F_HREQPERR, "PCI HMA channel request parity error", -1, 1 }, { F_HRSPPERR, "PCI HMA channel response parity error", -1, 1 }, { F_CFGSNPPERR, "PCI config snoop FIFO parity error", -1, 1 }, { F_FIDPERR, "PCI FID parity error", -1, 1 }, { F_INTXCLRPERR, "PCI INTx clear parity error", -1, 1 }, { F_MATAGPERR, "PCI MA tag parity error", -1, 1 }, { F_PIOTAGPERR, "PCI PIO tag parity error", -1, 1 }, { F_RXCPLPERR, "PCI Rx completion parity error", -1, 1 }, { F_RXWRPERR, "PCI Rx write parity error", -1, 1 }, { F_RPLPERR, "PCI replay buffer parity error", -1, 1 }, { F_PCIESINT, "PCI core secondary fault", -1, 1 }, { F_PCIEPINT, "PCI core primary fault", -1, 1 }, { F_UNXSPLCPLERR, "PCI unexpected split completion error", -1, 0 }, { 0 } }; static const struct intr_info t5_pcie_intr_info[] = { { F_MSTGRPPERR, "Master Response Read Queue parity error", -1, 1 }, { F_MSTTIMEOUTPERR, "Master Timeout FIFO parity error", -1, 1 }, { F_MSIXSTIPERR, "MSI-X STI SRAM parity error", -1, 1 }, { F_MSIXADDRLPERR, "MSI-X AddrL parity error", -1, 1 }, { F_MSIXADDRHPERR, "MSI-X AddrH parity error", -1, 1 }, { F_MSIXDATAPERR, "MSI-X data parity error", -1, 1 }, { F_MSIXDIPERR, "MSI-X DI parity error", -1, 1 }, { F_PIOCPLGRPPERR, "PCI PIO completion Group FIFO parity error", -1, 1 }, { F_PIOREQGRPPERR, "PCI PIO request Group FIFO parity error", -1, 1 }, { F_TARTAGPERR, "PCI PCI target tag FIFO parity error", -1, 1 }, { F_MSTTAGQPERR, "PCI master tag queue parity error", -1, 1 }, { F_CREQPERR, "PCI CMD channel request parity error", -1, 1 }, { F_CRSPPERR, "PCI CMD channel response parity error", -1, 1 }, { F_DREQWRPERR, "PCI DMA channel write request parity error", -1, 1 }, { F_DREQPERR, "PCI DMA channel request parity error", -1, 1 }, { F_DRSPPERR, "PCI DMA channel response parity error", -1, 1 }, { F_HREQWRPERR, "PCI HMA channel count parity error", -1, 1 }, { F_HREQPERR, "PCI HMA channel request parity error", -1, 1 }, { F_HRSPPERR, "PCI HMA channel response parity error", -1, 1 }, { F_CFGSNPPERR, "PCI config snoop FIFO parity error", -1, 1 }, { F_FIDPERR, "PCI FID parity error", -1, 1 }, { F_VFIDPERR, "PCI INTx clear parity error", -1, 1 }, { F_MAGRPPERR, "PCI MA group FIFO parity error", -1, 1 }, { F_PIOTAGPERR, "PCI PIO tag parity error", -1, 1 }, { F_IPRXHDRGRPPERR, "PCI IP Rx header group parity error", -1, 1 }, { F_IPRXDATAGRPPERR, "PCI IP Rx data group parity error", -1, 1 }, { F_RPLPERR, "PCI IP replay buffer parity error", -1, 1 }, { F_IPSOTPERR, "PCI IP SOT buffer parity error", -1, 1 }, { F_TRGT1GRPPERR, "PCI TRGT1 group FIFOs parity error", -1, 1 }, { F_READRSPERR, "Outbound read error", -1, 0 }, { 0 } }; int fat; if (is_t4(adapter)) fat = t4_handle_intr_status(adapter, A_PCIE_CORE_UTL_SYSTEM_BUS_AGENT_STATUS, sysbus_intr_info) + t4_handle_intr_status(adapter, A_PCIE_CORE_UTL_PCI_EXPRESS_PORT_STATUS, pcie_port_intr_info) + t4_handle_intr_status(adapter, A_PCIE_INT_CAUSE, pcie_intr_info); else fat = t4_handle_intr_status(adapter, A_PCIE_INT_CAUSE, t5_pcie_intr_info); if (fat) t4_fatal_err(adapter); } /* * TP interrupt handler. */ static void tp_intr_handler(struct adapter *adapter) { static const struct intr_info tp_intr_info[] = { { 0x3fffffff, "TP parity error", -1, 1 }, { F_FLMTXFLSTEMPTY, "TP out of Tx pages", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adapter, A_TP_INT_CAUSE, tp_intr_info)) t4_fatal_err(adapter); } /* * SGE interrupt handler. */ static void sge_intr_handler(struct adapter *adapter) { u64 v; u32 err; static const struct intr_info sge_intr_info[] = { { F_ERR_CPL_EXCEED_IQE_SIZE, "SGE received CPL exceeding IQE size", -1, 1 }, { F_ERR_INVALID_CIDX_INC, "SGE GTS CIDX increment too large", -1, 0 }, { F_ERR_CPL_OPCODE_0, "SGE received 0-length CPL", -1, 0 }, { F_DBFIFO_LP_INT, NULL, -1, 0, t4_db_full }, { F_ERR_DATA_CPL_ON_HIGH_QID1 | F_ERR_DATA_CPL_ON_HIGH_QID0, "SGE IQID > 1023 received CPL for FL", -1, 0 }, { F_ERR_BAD_DB_PIDX3, "SGE DBP 3 pidx increment too large", -1, 0 }, { F_ERR_BAD_DB_PIDX2, "SGE DBP 2 pidx increment too large", -1, 0 }, { F_ERR_BAD_DB_PIDX1, "SGE DBP 1 pidx increment too large", -1, 0 }, { F_ERR_BAD_DB_PIDX0, "SGE DBP 0 pidx increment too large", -1, 0 }, { F_ERR_ING_CTXT_PRIO, "SGE too many priority ingress contexts", -1, 0 }, { F_INGRESS_SIZE_ERR, "SGE illegal ingress QID", -1, 0 }, { F_EGRESS_SIZE_ERR, "SGE illegal egress QID", -1, 0 }, { 0 } }; static const struct intr_info t4t5_sge_intr_info[] = { { F_ERR_DROPPED_DB, NULL, -1, 0, t4_db_dropped }, { F_DBFIFO_HP_INT, NULL, -1, 0, t4_db_full }, { F_ERR_EGR_CTXT_PRIO, "SGE too many priority egress contexts", -1, 0 }, { 0 } }; /* * For now, treat below interrupts as fatal so that we disable SGE and * get better debug */ static const struct intr_info t6_sge_intr_info[] = { { F_ERR_PCIE_ERROR0 | F_ERR_PCIE_ERROR1, "SGE PCIe error for a DBP thread", -1, 1 }, { F_FATAL_WRE_LEN, "SGE Actual WRE packet is less than advertized length", -1, 1 }, { 0 } }; v = (u64)t4_read_reg(adapter, A_SGE_INT_CAUSE1) | ((u64)t4_read_reg(adapter, A_SGE_INT_CAUSE2) << 32); if (v) { CH_ALERT(adapter, "SGE parity error (%#llx)\n", (unsigned long long)v); t4_write_reg(adapter, A_SGE_INT_CAUSE1, v); t4_write_reg(adapter, A_SGE_INT_CAUSE2, v >> 32); } v |= t4_handle_intr_status(adapter, A_SGE_INT_CAUSE3, sge_intr_info); if (chip_id(adapter) <= CHELSIO_T5) v |= t4_handle_intr_status(adapter, A_SGE_INT_CAUSE3, t4t5_sge_intr_info); else v |= t4_handle_intr_status(adapter, A_SGE_INT_CAUSE3, t6_sge_intr_info); err = t4_read_reg(adapter, A_SGE_ERROR_STATS); if (err & F_ERROR_QID_VALID) { CH_ERR(adapter, "SGE error for queue %u\n", G_ERROR_QID(err)); if (err & F_UNCAPTURED_ERROR) CH_ERR(adapter, "SGE UNCAPTURED_ERROR set (clearing)\n"); t4_write_reg(adapter, A_SGE_ERROR_STATS, F_ERROR_QID_VALID | F_UNCAPTURED_ERROR); } if (v != 0) t4_fatal_err(adapter); } #define CIM_OBQ_INTR (F_OBQULP0PARERR | F_OBQULP1PARERR | F_OBQULP2PARERR |\ F_OBQULP3PARERR | F_OBQSGEPARERR | F_OBQNCSIPARERR) #define CIM_IBQ_INTR (F_IBQTP0PARERR | F_IBQTP1PARERR | F_IBQULPPARERR |\ F_IBQSGEHIPARERR | F_IBQSGELOPARERR | F_IBQNCSIPARERR) /* * CIM interrupt handler. */ static void cim_intr_handler(struct adapter *adapter) { static const struct intr_info cim_intr_info[] = { { F_PREFDROPINT, "CIM control register prefetch drop", -1, 1 }, { CIM_OBQ_INTR, "CIM OBQ parity error", -1, 1 }, { CIM_IBQ_INTR, "CIM IBQ parity error", -1, 1 }, { F_MBUPPARERR, "CIM mailbox uP parity error", -1, 1 }, { F_MBHOSTPARERR, "CIM mailbox host parity error", -1, 1 }, { F_TIEQINPARERRINT, "CIM TIEQ outgoing parity error", -1, 1 }, { F_TIEQOUTPARERRINT, "CIM TIEQ incoming parity error", -1, 1 }, { 0 } }; static const struct intr_info cim_upintr_info[] = { { F_RSVDSPACEINT, "CIM reserved space access", -1, 1 }, { F_ILLTRANSINT, "CIM illegal transaction", -1, 1 }, { F_ILLWRINT, "CIM illegal write", -1, 1 }, { F_ILLRDINT, "CIM illegal read", -1, 1 }, { F_ILLRDBEINT, "CIM illegal read BE", -1, 1 }, { F_ILLWRBEINT, "CIM illegal write BE", -1, 1 }, { F_SGLRDBOOTINT, "CIM single read from boot space", -1, 1 }, { F_SGLWRBOOTINT, "CIM single write to boot space", -1, 1 }, { F_BLKWRBOOTINT, "CIM block write to boot space", -1, 1 }, { F_SGLRDFLASHINT, "CIM single read from flash space", -1, 1 }, { F_SGLWRFLASHINT, "CIM single write to flash space", -1, 1 }, { F_BLKWRFLASHINT, "CIM block write to flash space", -1, 1 }, { F_SGLRDEEPROMINT, "CIM single EEPROM read", -1, 1 }, { F_SGLWREEPROMINT, "CIM single EEPROM write", -1, 1 }, { F_BLKRDEEPROMINT, "CIM block EEPROM read", -1, 1 }, { F_BLKWREEPROMINT, "CIM block EEPROM write", -1, 1 }, { F_SGLRDCTLINT , "CIM single read from CTL space", -1, 1 }, { F_SGLWRCTLINT , "CIM single write to CTL space", -1, 1 }, { F_BLKRDCTLINT , "CIM block read from CTL space", -1, 1 }, { F_BLKWRCTLINT , "CIM block write to CTL space", -1, 1 }, { F_SGLRDPLINT , "CIM single read from PL space", -1, 1 }, { F_SGLWRPLINT , "CIM single write to PL space", -1, 1 }, { F_BLKRDPLINT , "CIM block read from PL space", -1, 1 }, { F_BLKWRPLINT , "CIM block write to PL space", -1, 1 }, { F_REQOVRLOOKUPINT , "CIM request FIFO overwrite", -1, 1 }, { F_RSPOVRLOOKUPINT , "CIM response FIFO overwrite", -1, 1 }, { F_TIMEOUTINT , "CIM PIF timeout", -1, 1 }, { F_TIMEOUTMAINT , "CIM PIF MA timeout", -1, 1 }, { 0 } }; int fat; if (t4_read_reg(adapter, A_PCIE_FW) & F_PCIE_FW_ERR) t4_report_fw_error(adapter); fat = t4_handle_intr_status(adapter, A_CIM_HOST_INT_CAUSE, cim_intr_info) + t4_handle_intr_status(adapter, A_CIM_HOST_UPACC_INT_CAUSE, cim_upintr_info); if (fat) t4_fatal_err(adapter); } /* * ULP RX interrupt handler. */ static void ulprx_intr_handler(struct adapter *adapter) { static const struct intr_info ulprx_intr_info[] = { { F_CAUSE_CTX_1, "ULPRX channel 1 context error", -1, 1 }, { F_CAUSE_CTX_0, "ULPRX channel 0 context error", -1, 1 }, { 0x7fffff, "ULPRX parity error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adapter, A_ULP_RX_INT_CAUSE, ulprx_intr_info)) t4_fatal_err(adapter); } /* * ULP TX interrupt handler. */ static void ulptx_intr_handler(struct adapter *adapter) { static const struct intr_info ulptx_intr_info[] = { { F_PBL_BOUND_ERR_CH3, "ULPTX channel 3 PBL out of bounds", -1, 0 }, { F_PBL_BOUND_ERR_CH2, "ULPTX channel 2 PBL out of bounds", -1, 0 }, { F_PBL_BOUND_ERR_CH1, "ULPTX channel 1 PBL out of bounds", -1, 0 }, { F_PBL_BOUND_ERR_CH0, "ULPTX channel 0 PBL out of bounds", -1, 0 }, { 0xfffffff, "ULPTX parity error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adapter, A_ULP_TX_INT_CAUSE, ulptx_intr_info)) t4_fatal_err(adapter); } /* * PM TX interrupt handler. */ static void pmtx_intr_handler(struct adapter *adapter) { static const struct intr_info pmtx_intr_info[] = { { F_PCMD_LEN_OVFL0, "PMTX channel 0 pcmd too large", -1, 1 }, { F_PCMD_LEN_OVFL1, "PMTX channel 1 pcmd too large", -1, 1 }, { F_PCMD_LEN_OVFL2, "PMTX channel 2 pcmd too large", -1, 1 }, { F_ZERO_C_CMD_ERROR, "PMTX 0-length pcmd", -1, 1 }, { 0xffffff0, "PMTX framing error", -1, 1 }, { F_OESPI_PAR_ERROR, "PMTX oespi parity error", -1, 1 }, { F_DB_OPTIONS_PAR_ERROR, "PMTX db_options parity error", -1, 1 }, { F_ICSPI_PAR_ERROR, "PMTX icspi parity error", -1, 1 }, { F_C_PCMD_PAR_ERROR, "PMTX c_pcmd parity error", -1, 1}, { 0 } }; if (t4_handle_intr_status(adapter, A_PM_TX_INT_CAUSE, pmtx_intr_info)) t4_fatal_err(adapter); } /* * PM RX interrupt handler. */ static void pmrx_intr_handler(struct adapter *adapter) { static const struct intr_info pmrx_intr_info[] = { { F_ZERO_E_CMD_ERROR, "PMRX 0-length pcmd", -1, 1 }, { 0x3ffff0, "PMRX framing error", -1, 1 }, { F_OCSPI_PAR_ERROR, "PMRX ocspi parity error", -1, 1 }, { F_DB_OPTIONS_PAR_ERROR, "PMRX db_options parity error", -1, 1 }, { F_IESPI_PAR_ERROR, "PMRX iespi parity error", -1, 1 }, { F_E_PCMD_PAR_ERROR, "PMRX e_pcmd parity error", -1, 1}, { 0 } }; if (t4_handle_intr_status(adapter, A_PM_RX_INT_CAUSE, pmrx_intr_info)) t4_fatal_err(adapter); } /* * CPL switch interrupt handler. */ static void cplsw_intr_handler(struct adapter *adapter) { static const struct intr_info cplsw_intr_info[] = { { F_CIM_OP_MAP_PERR, "CPLSW CIM op_map parity error", -1, 1 }, { F_CIM_OVFL_ERROR, "CPLSW CIM overflow", -1, 1 }, { F_TP_FRAMING_ERROR, "CPLSW TP framing error", -1, 1 }, { F_SGE_FRAMING_ERROR, "CPLSW SGE framing error", -1, 1 }, { F_CIM_FRAMING_ERROR, "CPLSW CIM framing error", -1, 1 }, { F_ZERO_SWITCH_ERROR, "CPLSW no-switch error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adapter, A_CPL_INTR_CAUSE, cplsw_intr_info)) t4_fatal_err(adapter); } /* * LE interrupt handler. */ static void le_intr_handler(struct adapter *adap) { unsigned int chip_ver = chip_id(adap); static const struct intr_info le_intr_info[] = { { F_LIPMISS, "LE LIP miss", -1, 0 }, { F_LIP0, "LE 0 LIP error", -1, 0 }, { F_PARITYERR, "LE parity error", -1, 1 }, { F_UNKNOWNCMD, "LE unknown command", -1, 1 }, { F_REQQPARERR, "LE request queue parity error", -1, 1 }, { 0 } }; static const struct intr_info t6_le_intr_info[] = { { F_T6_LIPMISS, "LE LIP miss", -1, 0 }, { F_T6_LIP0, "LE 0 LIP error", -1, 0 }, { F_TCAMINTPERR, "LE parity error", -1, 1 }, { F_T6_UNKNOWNCMD, "LE unknown command", -1, 1 }, { F_SSRAMINTPERR, "LE request queue parity error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adap, A_LE_DB_INT_CAUSE, (chip_ver <= CHELSIO_T5) ? le_intr_info : t6_le_intr_info)) t4_fatal_err(adap); } /* * MPS interrupt handler. */ static void mps_intr_handler(struct adapter *adapter) { static const struct intr_info mps_rx_intr_info[] = { { 0xffffff, "MPS Rx parity error", -1, 1 }, { 0 } }; static const struct intr_info mps_tx_intr_info[] = { { V_TPFIFO(M_TPFIFO), "MPS Tx TP FIFO parity error", -1, 1 }, { F_NCSIFIFO, "MPS Tx NC-SI FIFO parity error", -1, 1 }, { V_TXDATAFIFO(M_TXDATAFIFO), "MPS Tx data FIFO parity error", -1, 1 }, { V_TXDESCFIFO(M_TXDESCFIFO), "MPS Tx desc FIFO parity error", -1, 1 }, { F_BUBBLE, "MPS Tx underflow", -1, 1 }, { F_SECNTERR, "MPS Tx SOP/EOP error", -1, 1 }, { F_FRMERR, "MPS Tx framing error", -1, 1 }, { 0 } }; static const struct intr_info mps_trc_intr_info[] = { { V_FILTMEM(M_FILTMEM), "MPS TRC filter parity error", -1, 1 }, { V_PKTFIFO(M_PKTFIFO), "MPS TRC packet FIFO parity error", -1, 1 }, { F_MISCPERR, "MPS TRC misc parity error", -1, 1 }, { 0 } }; static const struct intr_info mps_stat_sram_intr_info[] = { { 0x1fffff, "MPS statistics SRAM parity error", -1, 1 }, { 0 } }; static const struct intr_info mps_stat_tx_intr_info[] = { { 0xfffff, "MPS statistics Tx FIFO parity error", -1, 1 }, { 0 } }; static const struct intr_info mps_stat_rx_intr_info[] = { { 0xffffff, "MPS statistics Rx FIFO parity error", -1, 1 }, { 0 } }; static const struct intr_info mps_cls_intr_info[] = { { F_MATCHSRAM, "MPS match SRAM parity error", -1, 1 }, { F_MATCHTCAM, "MPS match TCAM parity error", -1, 1 }, { F_HASHSRAM, "MPS hash SRAM parity error", -1, 1 }, { 0 } }; int fat; fat = t4_handle_intr_status(adapter, A_MPS_RX_PERR_INT_CAUSE, mps_rx_intr_info) + t4_handle_intr_status(adapter, A_MPS_TX_INT_CAUSE, mps_tx_intr_info) + t4_handle_intr_status(adapter, A_MPS_TRC_INT_CAUSE, mps_trc_intr_info) + t4_handle_intr_status(adapter, A_MPS_STAT_PERR_INT_CAUSE_SRAM, mps_stat_sram_intr_info) + t4_handle_intr_status(adapter, A_MPS_STAT_PERR_INT_CAUSE_TX_FIFO, mps_stat_tx_intr_info) + t4_handle_intr_status(adapter, A_MPS_STAT_PERR_INT_CAUSE_RX_FIFO, mps_stat_rx_intr_info) + t4_handle_intr_status(adapter, A_MPS_CLS_INT_CAUSE, mps_cls_intr_info); t4_write_reg(adapter, A_MPS_INT_CAUSE, 0); t4_read_reg(adapter, A_MPS_INT_CAUSE); /* flush */ if (fat) t4_fatal_err(adapter); } #define MEM_INT_MASK (F_PERR_INT_CAUSE | F_ECC_CE_INT_CAUSE | \ F_ECC_UE_INT_CAUSE) /* * EDC/MC interrupt handler. */ static void mem_intr_handler(struct adapter *adapter, int idx) { static const char name[4][7] = { "EDC0", "EDC1", "MC/MC0", "MC1" }; unsigned int addr, cnt_addr, v; if (idx <= MEM_EDC1) { addr = EDC_REG(A_EDC_INT_CAUSE, idx); cnt_addr = EDC_REG(A_EDC_ECC_STATUS, idx); } else if (idx == MEM_MC) { if (is_t4(adapter)) { addr = A_MC_INT_CAUSE; cnt_addr = A_MC_ECC_STATUS; } else { addr = A_MC_P_INT_CAUSE; cnt_addr = A_MC_P_ECC_STATUS; } } else { addr = MC_REG(A_MC_P_INT_CAUSE, 1); cnt_addr = MC_REG(A_MC_P_ECC_STATUS, 1); } v = t4_read_reg(adapter, addr) & MEM_INT_MASK; if (v & F_PERR_INT_CAUSE) CH_ALERT(adapter, "%s FIFO parity error\n", name[idx]); if (v & F_ECC_CE_INT_CAUSE) { u32 cnt = G_ECC_CECNT(t4_read_reg(adapter, cnt_addr)); t4_edc_err_read(adapter, idx); t4_write_reg(adapter, cnt_addr, V_ECC_CECNT(M_ECC_CECNT)); CH_WARN_RATELIMIT(adapter, "%u %s correctable ECC data error%s\n", cnt, name[idx], cnt > 1 ? "s" : ""); } if (v & F_ECC_UE_INT_CAUSE) CH_ALERT(adapter, "%s uncorrectable ECC data error\n", name[idx]); t4_write_reg(adapter, addr, v); if (v & (F_PERR_INT_CAUSE | F_ECC_UE_INT_CAUSE)) t4_fatal_err(adapter); } /* * MA interrupt handler. */ static void ma_intr_handler(struct adapter *adapter) { u32 v, status = t4_read_reg(adapter, A_MA_INT_CAUSE); if (status & F_MEM_PERR_INT_CAUSE) { CH_ALERT(adapter, "MA parity error, parity status %#x\n", t4_read_reg(adapter, A_MA_PARITY_ERROR_STATUS1)); if (is_t5(adapter)) CH_ALERT(adapter, "MA parity error, parity status %#x\n", t4_read_reg(adapter, A_MA_PARITY_ERROR_STATUS2)); } if (status & F_MEM_WRAP_INT_CAUSE) { v = t4_read_reg(adapter, A_MA_INT_WRAP_STATUS); CH_ALERT(adapter, "MA address wrap-around error by " "client %u to address %#x\n", G_MEM_WRAP_CLIENT_NUM(v), G_MEM_WRAP_ADDRESS(v) << 4); } t4_write_reg(adapter, A_MA_INT_CAUSE, status); t4_fatal_err(adapter); } /* * SMB interrupt handler. */ static void smb_intr_handler(struct adapter *adap) { static const struct intr_info smb_intr_info[] = { { F_MSTTXFIFOPARINT, "SMB master Tx FIFO parity error", -1, 1 }, { F_MSTRXFIFOPARINT, "SMB master Rx FIFO parity error", -1, 1 }, { F_SLVFIFOPARINT, "SMB slave FIFO parity error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adap, A_SMB_INT_CAUSE, smb_intr_info)) t4_fatal_err(adap); } /* * NC-SI interrupt handler. */ static void ncsi_intr_handler(struct adapter *adap) { static const struct intr_info ncsi_intr_info[] = { { F_CIM_DM_PRTY_ERR, "NC-SI CIM parity error", -1, 1 }, { F_MPS_DM_PRTY_ERR, "NC-SI MPS parity error", -1, 1 }, { F_TXFIFO_PRTY_ERR, "NC-SI Tx FIFO parity error", -1, 1 }, { F_RXFIFO_PRTY_ERR, "NC-SI Rx FIFO parity error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adap, A_NCSI_INT_CAUSE, ncsi_intr_info)) t4_fatal_err(adap); } /* * XGMAC interrupt handler. */ static void xgmac_intr_handler(struct adapter *adap, int port) { u32 v, int_cause_reg; if (is_t4(adap)) int_cause_reg = PORT_REG(port, A_XGMAC_PORT_INT_CAUSE); else int_cause_reg = T5_PORT_REG(port, A_MAC_PORT_INT_CAUSE); v = t4_read_reg(adap, int_cause_reg); v &= (F_TXFIFO_PRTY_ERR | F_RXFIFO_PRTY_ERR); if (!v) return; if (v & F_TXFIFO_PRTY_ERR) CH_ALERT(adap, "XGMAC %d Tx FIFO parity error\n", port); if (v & F_RXFIFO_PRTY_ERR) CH_ALERT(adap, "XGMAC %d Rx FIFO parity error\n", port); t4_write_reg(adap, int_cause_reg, v); t4_fatal_err(adap); } /* * PL interrupt handler. */ static void pl_intr_handler(struct adapter *adap) { static const struct intr_info pl_intr_info[] = { { F_FATALPERR, "Fatal parity error", -1, 1 }, { F_PERRVFID, "PL VFID_MAP parity error", -1, 1 }, { 0 } }; static const struct intr_info t5_pl_intr_info[] = { { F_FATALPERR, "Fatal parity error", -1, 1 }, { 0 } }; if (t4_handle_intr_status(adap, A_PL_PL_INT_CAUSE, is_t4(adap) ? pl_intr_info : t5_pl_intr_info)) t4_fatal_err(adap); } #define PF_INTR_MASK (F_PFSW | F_PFCIM) /** * t4_slow_intr_handler - control path interrupt handler * @adapter: the adapter * * T4 interrupt handler for non-data global interrupt events, e.g., errors. * The designation 'slow' is because it involves register reads, while * data interrupts typically don't involve any MMIOs. */ int t4_slow_intr_handler(struct adapter *adapter) { u32 cause = t4_read_reg(adapter, A_PL_INT_CAUSE); if (!(cause & GLBL_INTR_MASK)) return 0; if (cause & F_CIM) cim_intr_handler(adapter); if (cause & F_MPS) mps_intr_handler(adapter); if (cause & F_NCSI) ncsi_intr_handler(adapter); if (cause & F_PL) pl_intr_handler(adapter); if (cause & F_SMB) smb_intr_handler(adapter); if (cause & F_MAC0) xgmac_intr_handler(adapter, 0); if (cause & F_MAC1) xgmac_intr_handler(adapter, 1); if (cause & F_MAC2) xgmac_intr_handler(adapter, 2); if (cause & F_MAC3) xgmac_intr_handler(adapter, 3); if (cause & F_PCIE) pcie_intr_handler(adapter); if (cause & F_MC0) mem_intr_handler(adapter, MEM_MC); if (is_t5(adapter) && (cause & F_MC1)) mem_intr_handler(adapter, MEM_MC1); if (cause & F_EDC0) mem_intr_handler(adapter, MEM_EDC0); if (cause & F_EDC1) mem_intr_handler(adapter, MEM_EDC1); if (cause & F_LE) le_intr_handler(adapter); if (cause & F_TP) tp_intr_handler(adapter); if (cause & F_MA) ma_intr_handler(adapter); if (cause & F_PM_TX) pmtx_intr_handler(adapter); if (cause & F_PM_RX) pmrx_intr_handler(adapter); if (cause & F_ULP_RX) ulprx_intr_handler(adapter); if (cause & F_CPL_SWITCH) cplsw_intr_handler(adapter); if (cause & F_SGE) sge_intr_handler(adapter); if (cause & F_ULP_TX) ulptx_intr_handler(adapter); /* Clear the interrupts just processed for which we are the master. */ t4_write_reg(adapter, A_PL_INT_CAUSE, cause & GLBL_INTR_MASK); (void)t4_read_reg(adapter, A_PL_INT_CAUSE); /* flush */ return 1; } /** * t4_intr_enable - enable interrupts * @adapter: the adapter whose interrupts should be enabled * * Enable PF-specific interrupts for the calling function and the top-level * interrupt concentrator for global interrupts. Interrupts are already * enabled at each module, here we just enable the roots of the interrupt * hierarchies. * * Note: this function should be called only when the driver manages * non PF-specific interrupts from the various HW modules. Only one PCI * function at a time should be doing this. */ void t4_intr_enable(struct adapter *adapter) { u32 val = 0; u32 whoami = t4_read_reg(adapter, A_PL_WHOAMI); u32 pf = (chip_id(adapter) <= CHELSIO_T5 ? G_SOURCEPF(whoami) : G_T6_SOURCEPF(whoami)); if (chip_id(adapter) <= CHELSIO_T5) val = F_ERR_DROPPED_DB | F_ERR_EGR_CTXT_PRIO | F_DBFIFO_HP_INT; else val = F_ERR_PCIE_ERROR0 | F_ERR_PCIE_ERROR1 | F_FATAL_WRE_LEN; t4_write_reg(adapter, A_SGE_INT_ENABLE3, F_ERR_CPL_EXCEED_IQE_SIZE | F_ERR_INVALID_CIDX_INC | F_ERR_CPL_OPCODE_0 | F_ERR_DATA_CPL_ON_HIGH_QID1 | F_INGRESS_SIZE_ERR | F_ERR_DATA_CPL_ON_HIGH_QID0 | F_ERR_BAD_DB_PIDX3 | F_ERR_BAD_DB_PIDX2 | F_ERR_BAD_DB_PIDX1 | F_ERR_BAD_DB_PIDX0 | F_ERR_ING_CTXT_PRIO | F_DBFIFO_LP_INT | F_EGRESS_SIZE_ERR | val); t4_write_reg(adapter, MYPF_REG(A_PL_PF_INT_ENABLE), PF_INTR_MASK); t4_set_reg_field(adapter, A_PL_INT_MAP0, 0, 1 << pf); } /** * t4_intr_disable - disable interrupts * @adapter: the adapter whose interrupts should be disabled * * Disable interrupts. We only disable the top-level interrupt * concentrators. The caller must be a PCI function managing global * interrupts. */ void t4_intr_disable(struct adapter *adapter) { u32 whoami = t4_read_reg(adapter, A_PL_WHOAMI); u32 pf = (chip_id(adapter) <= CHELSIO_T5 ? G_SOURCEPF(whoami) : G_T6_SOURCEPF(whoami)); t4_write_reg(adapter, MYPF_REG(A_PL_PF_INT_ENABLE), 0); t4_set_reg_field(adapter, A_PL_INT_MAP0, 1 << pf, 0); } /** * t4_intr_clear - clear all interrupts * @adapter: the adapter whose interrupts should be cleared * * Clears all interrupts. The caller must be a PCI function managing * global interrupts. */ void t4_intr_clear(struct adapter *adapter) { static const unsigned int cause_reg[] = { A_SGE_INT_CAUSE1, A_SGE_INT_CAUSE2, A_SGE_INT_CAUSE3, A_PCIE_NONFAT_ERR, A_PCIE_INT_CAUSE, A_MA_INT_WRAP_STATUS, A_MA_PARITY_ERROR_STATUS1, A_MA_INT_CAUSE, A_EDC_INT_CAUSE, EDC_REG(A_EDC_INT_CAUSE, 1), A_CIM_HOST_INT_CAUSE, A_CIM_HOST_UPACC_INT_CAUSE, MYPF_REG(A_CIM_PF_HOST_INT_CAUSE), A_TP_INT_CAUSE, A_ULP_RX_INT_CAUSE, A_ULP_TX_INT_CAUSE, A_PM_RX_INT_CAUSE, A_PM_TX_INT_CAUSE, A_MPS_RX_PERR_INT_CAUSE, A_CPL_INTR_CAUSE, MYPF_REG(A_PL_PF_INT_CAUSE), A_PL_PL_INT_CAUSE, A_LE_DB_INT_CAUSE, }; unsigned int i; for (i = 0; i < ARRAY_SIZE(cause_reg); ++i) t4_write_reg(adapter, cause_reg[i], 0xffffffff); t4_write_reg(adapter, is_t4(adapter) ? A_MC_INT_CAUSE : A_MC_P_INT_CAUSE, 0xffffffff); if (is_t4(adapter)) { t4_write_reg(adapter, A_PCIE_CORE_UTL_SYSTEM_BUS_AGENT_STATUS, 0xffffffff); t4_write_reg(adapter, A_PCIE_CORE_UTL_PCI_EXPRESS_PORT_STATUS, 0xffffffff); } else t4_write_reg(adapter, A_MA_PARITY_ERROR_STATUS2, 0xffffffff); t4_write_reg(adapter, A_PL_INT_CAUSE, GLBL_INTR_MASK); (void) t4_read_reg(adapter, A_PL_INT_CAUSE); /* flush */ } /** * hash_mac_addr - return the hash value of a MAC address * @addr: the 48-bit Ethernet MAC address * * Hashes a MAC address according to the hash function used by HW inexact * (hash) address matching. */ static int hash_mac_addr(const u8 *addr) { u32 a = ((u32)addr[0] << 16) | ((u32)addr[1] << 8) | addr[2]; u32 b = ((u32)addr[3] << 16) | ((u32)addr[4] << 8) | addr[5]; a ^= b; a ^= (a >> 12); a ^= (a >> 6); return a & 0x3f; } /** * t4_config_rss_range - configure a portion of the RSS mapping table * @adapter: the adapter * @mbox: mbox to use for the FW command * @viid: virtual interface whose RSS subtable is to be written * @start: start entry in the table to write * @n: how many table entries to write * @rspq: values for the "response queue" (Ingress Queue) lookup table * @nrspq: number of values in @rspq * * Programs the selected part of the VI's RSS mapping table with the * provided values. If @nrspq < @n the supplied values are used repeatedly * until the full table range is populated. * * The caller must ensure the values in @rspq are in the range allowed for * @viid. */ int t4_config_rss_range(struct adapter *adapter, int mbox, unsigned int viid, int start, int n, const u16 *rspq, unsigned int nrspq) { int ret; const u16 *rsp = rspq; const u16 *rsp_end = rspq + nrspq; struct fw_rss_ind_tbl_cmd cmd; memset(&cmd, 0, sizeof(cmd)); cmd.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_RSS_IND_TBL_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_RSS_IND_TBL_CMD_VIID(viid)); cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd)); /* * Each firmware RSS command can accommodate up to 32 RSS Ingress * Queue Identifiers. These Ingress Queue IDs are packed three to * a 32-bit word as 10-bit values with the upper remaining 2 bits * reserved. */ while (n > 0) { int nq = min(n, 32); int nq_packed = 0; __be32 *qp = &cmd.iq0_to_iq2; /* * Set up the firmware RSS command header to send the next * "nq" Ingress Queue IDs to the firmware. */ cmd.niqid = cpu_to_be16(nq); cmd.startidx = cpu_to_be16(start); /* * "nq" more done for the start of the next loop. */ start += nq; n -= nq; /* * While there are still Ingress Queue IDs to stuff into the * current firmware RSS command, retrieve them from the * Ingress Queue ID array and insert them into the command. */ while (nq > 0) { /* * Grab up to the next 3 Ingress Queue IDs (wrapping * around the Ingress Queue ID array if necessary) and * insert them into the firmware RSS command at the * current 3-tuple position within the commad. */ u16 qbuf[3]; u16 *qbp = qbuf; int nqbuf = min(3, nq); nq -= nqbuf; qbuf[0] = qbuf[1] = qbuf[2] = 0; while (nqbuf && nq_packed < 32) { nqbuf--; nq_packed++; *qbp++ = *rsp++; if (rsp >= rsp_end) rsp = rspq; } *qp++ = cpu_to_be32(V_FW_RSS_IND_TBL_CMD_IQ0(qbuf[0]) | V_FW_RSS_IND_TBL_CMD_IQ1(qbuf[1]) | V_FW_RSS_IND_TBL_CMD_IQ2(qbuf[2])); } /* * Send this portion of the RRS table update to the firmware; * bail out on any errors. */ ret = t4_wr_mbox(adapter, mbox, &cmd, sizeof(cmd), NULL); if (ret) return ret; } return 0; } /** * t4_config_glbl_rss - configure the global RSS mode * @adapter: the adapter * @mbox: mbox to use for the FW command * @mode: global RSS mode * @flags: mode-specific flags * * Sets the global RSS mode. */ int t4_config_glbl_rss(struct adapter *adapter, int mbox, unsigned int mode, unsigned int flags) { struct fw_rss_glb_config_cmd c; memset(&c, 0, sizeof(c)); c.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_RSS_GLB_CONFIG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); if (mode == FW_RSS_GLB_CONFIG_CMD_MODE_MANUAL) { c.u.manual.mode_pkd = cpu_to_be32(V_FW_RSS_GLB_CONFIG_CMD_MODE(mode)); } else if (mode == FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL) { c.u.basicvirtual.mode_pkd = cpu_to_be32(V_FW_RSS_GLB_CONFIG_CMD_MODE(mode)); c.u.basicvirtual.synmapen_to_hashtoeplitz = cpu_to_be32(flags); } else return -EINVAL; return t4_wr_mbox(adapter, mbox, &c, sizeof(c), NULL); } /** * t4_config_vi_rss - configure per VI RSS settings * @adapter: the adapter * @mbox: mbox to use for the FW command * @viid: the VI id * @flags: RSS flags * @defq: id of the default RSS queue for the VI. * * Configures VI-specific RSS properties. */ int t4_config_vi_rss(struct adapter *adapter, int mbox, unsigned int viid, unsigned int flags, unsigned int defq) { struct fw_rss_vi_config_cmd c; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_RSS_VI_CONFIG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_RSS_VI_CONFIG_CMD_VIID(viid)); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); c.u.basicvirtual.defaultq_to_udpen = cpu_to_be32(flags | V_FW_RSS_VI_CONFIG_CMD_DEFAULTQ(defq)); return t4_wr_mbox(adapter, mbox, &c, sizeof(c), NULL); } /* Read an RSS table row */ static int rd_rss_row(struct adapter *adap, int row, u32 *val) { t4_write_reg(adap, A_TP_RSS_LKP_TABLE, 0xfff00000 | row); return t4_wait_op_done_val(adap, A_TP_RSS_LKP_TABLE, F_LKPTBLROWVLD, 1, 5, 0, val); } /** * t4_read_rss - read the contents of the RSS mapping table * @adapter: the adapter * @map: holds the contents of the RSS mapping table * * Reads the contents of the RSS hash->queue mapping table. */ int t4_read_rss(struct adapter *adapter, u16 *map) { u32 val; int i, ret; for (i = 0; i < RSS_NENTRIES / 2; ++i) { ret = rd_rss_row(adapter, i, &val); if (ret) return ret; *map++ = G_LKPTBLQUEUE0(val); *map++ = G_LKPTBLQUEUE1(val); } return 0; } /** * t4_fw_tp_pio_rw - Access TP PIO through LDST * @adap: the adapter * @vals: where the indirect register values are stored/written * @nregs: how many indirect registers to read/write * @start_idx: index of first indirect register to read/write * @rw: Read (1) or Write (0) * * Access TP PIO registers through LDST */ void t4_fw_tp_pio_rw(struct adapter *adap, u32 *vals, unsigned int nregs, unsigned int start_index, unsigned int rw) { int ret, i; int cmd = FW_LDST_ADDRSPC_TP_PIO; struct fw_ldst_cmd c; for (i = 0 ; i < nregs; i++) { memset(&c, 0, sizeof(c)); c.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | (rw ? F_FW_CMD_READ : F_FW_CMD_WRITE) | V_FW_LDST_CMD_ADDRSPACE(cmd)); c.cycles_to_len16 = cpu_to_be32(FW_LEN16(c)); c.u.addrval.addr = cpu_to_be32(start_index + i); c.u.addrval.val = rw ? 0 : cpu_to_be32(vals[i]); ret = t4_wr_mbox(adap, adap->mbox, &c, sizeof(c), &c); if (ret == 0) { if (rw) vals[i] = be32_to_cpu(c.u.addrval.val); } } } /** * t4_read_rss_key - read the global RSS key * @adap: the adapter * @key: 10-entry array holding the 320-bit RSS key * * Reads the global 320-bit RSS key. */ void t4_read_rss_key(struct adapter *adap, u32 *key) { if (t4_use_ldst(adap)) t4_fw_tp_pio_rw(adap, key, 10, A_TP_RSS_SECRET_KEY0, 1); else t4_read_indirect(adap, A_TP_PIO_ADDR, A_TP_PIO_DATA, key, 10, A_TP_RSS_SECRET_KEY0); } /** * t4_write_rss_key - program one of the RSS keys * @adap: the adapter * @key: 10-entry array holding the 320-bit RSS key * @idx: which RSS key to write * * Writes one of the RSS keys with the given 320-bit value. If @idx is * 0..15 the corresponding entry in the RSS key table is written, * otherwise the global RSS key is written. */ void t4_write_rss_key(struct adapter *adap, u32 *key, int idx) { u8 rss_key_addr_cnt = 16; u32 vrt = t4_read_reg(adap, A_TP_RSS_CONFIG_VRT); /* * T6 and later: for KeyMode 3 (per-vf and per-vf scramble), * allows access to key addresses 16-63 by using KeyWrAddrX * as index[5:4](upper 2) into key table */ if ((chip_id(adap) > CHELSIO_T5) && (vrt & F_KEYEXTEND) && (G_KEYMODE(vrt) == 3)) rss_key_addr_cnt = 32; if (t4_use_ldst(adap)) t4_fw_tp_pio_rw(adap, key, 10, A_TP_RSS_SECRET_KEY0, 0); else t4_write_indirect(adap, A_TP_PIO_ADDR, A_TP_PIO_DATA, key, 10, A_TP_RSS_SECRET_KEY0); if (idx >= 0 && idx < rss_key_addr_cnt) { if (rss_key_addr_cnt > 16) t4_write_reg(adap, A_TP_RSS_CONFIG_VRT, V_KEYWRADDRX(idx >> 4) | V_T6_VFWRADDR(idx) | F_KEYWREN); else t4_write_reg(adap, A_TP_RSS_CONFIG_VRT, V_KEYWRADDR(idx) | F_KEYWREN); } } /** * t4_read_rss_pf_config - read PF RSS Configuration Table * @adapter: the adapter * @index: the entry in the PF RSS table to read * @valp: where to store the returned value * * Reads the PF RSS Configuration Table at the specified index and returns * the value found there. */ void t4_read_rss_pf_config(struct adapter *adapter, unsigned int index, u32 *valp) { if (t4_use_ldst(adapter)) t4_fw_tp_pio_rw(adapter, valp, 1, A_TP_RSS_PF0_CONFIG + index, 1); else t4_read_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, valp, 1, A_TP_RSS_PF0_CONFIG + index); } /** * t4_write_rss_pf_config - write PF RSS Configuration Table * @adapter: the adapter * @index: the entry in the VF RSS table to read * @val: the value to store * * Writes the PF RSS Configuration Table at the specified index with the * specified value. */ void t4_write_rss_pf_config(struct adapter *adapter, unsigned int index, u32 val) { if (t4_use_ldst(adapter)) t4_fw_tp_pio_rw(adapter, &val, 1, A_TP_RSS_PF0_CONFIG + index, 0); else t4_write_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &val, 1, A_TP_RSS_PF0_CONFIG + index); } /** * t4_read_rss_vf_config - read VF RSS Configuration Table * @adapter: the adapter * @index: the entry in the VF RSS table to read * @vfl: where to store the returned VFL * @vfh: where to store the returned VFH * * Reads the VF RSS Configuration Table at the specified index and returns * the (VFL, VFH) values found there. */ void t4_read_rss_vf_config(struct adapter *adapter, unsigned int index, u32 *vfl, u32 *vfh) { u32 vrt, mask, data; if (chip_id(adapter) <= CHELSIO_T5) { mask = V_VFWRADDR(M_VFWRADDR); data = V_VFWRADDR(index); } else { mask = V_T6_VFWRADDR(M_T6_VFWRADDR); data = V_T6_VFWRADDR(index); } /* * Request that the index'th VF Table values be read into VFL/VFH. */ vrt = t4_read_reg(adapter, A_TP_RSS_CONFIG_VRT); vrt &= ~(F_VFRDRG | F_VFWREN | F_KEYWREN | mask); vrt |= data | F_VFRDEN; t4_write_reg(adapter, A_TP_RSS_CONFIG_VRT, vrt); /* * Grab the VFL/VFH values ... */ if (t4_use_ldst(adapter)) { t4_fw_tp_pio_rw(adapter, vfl, 1, A_TP_RSS_VFL_CONFIG, 1); t4_fw_tp_pio_rw(adapter, vfh, 1, A_TP_RSS_VFH_CONFIG, 1); } else { t4_read_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, vfl, 1, A_TP_RSS_VFL_CONFIG); t4_read_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, vfh, 1, A_TP_RSS_VFH_CONFIG); } } /** * t4_write_rss_vf_config - write VF RSS Configuration Table * * @adapter: the adapter * @index: the entry in the VF RSS table to write * @vfl: the VFL to store * @vfh: the VFH to store * * Writes the VF RSS Configuration Table at the specified index with the * specified (VFL, VFH) values. */ void t4_write_rss_vf_config(struct adapter *adapter, unsigned int index, u32 vfl, u32 vfh) { u32 vrt, mask, data; if (chip_id(adapter) <= CHELSIO_T5) { mask = V_VFWRADDR(M_VFWRADDR); data = V_VFWRADDR(index); } else { mask = V_T6_VFWRADDR(M_T6_VFWRADDR); data = V_T6_VFWRADDR(index); } /* * Load up VFL/VFH with the values to be written ... */ if (t4_use_ldst(adapter)) { t4_fw_tp_pio_rw(adapter, &vfl, 1, A_TP_RSS_VFL_CONFIG, 0); t4_fw_tp_pio_rw(adapter, &vfh, 1, A_TP_RSS_VFH_CONFIG, 0); } else { t4_write_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &vfl, 1, A_TP_RSS_VFL_CONFIG); t4_write_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &vfh, 1, A_TP_RSS_VFH_CONFIG); } /* * Write the VFL/VFH into the VF Table at index'th location. */ vrt = t4_read_reg(adapter, A_TP_RSS_CONFIG_VRT); vrt &= ~(F_VFRDRG | F_VFWREN | F_KEYWREN | mask); vrt |= data | F_VFRDEN; t4_write_reg(adapter, A_TP_RSS_CONFIG_VRT, vrt); } /** * t4_read_rss_pf_map - read PF RSS Map * @adapter: the adapter * * Reads the PF RSS Map register and returns its value. */ u32 t4_read_rss_pf_map(struct adapter *adapter) { u32 pfmap; if (t4_use_ldst(adapter)) t4_fw_tp_pio_rw(adapter, &pfmap, 1, A_TP_RSS_PF_MAP, 1); else t4_read_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &pfmap, 1, A_TP_RSS_PF_MAP); return pfmap; } /** * t4_write_rss_pf_map - write PF RSS Map * @adapter: the adapter * @pfmap: PF RSS Map value * * Writes the specified value to the PF RSS Map register. */ void t4_write_rss_pf_map(struct adapter *adapter, u32 pfmap) { if (t4_use_ldst(adapter)) t4_fw_tp_pio_rw(adapter, &pfmap, 1, A_TP_RSS_PF_MAP, 0); else t4_write_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &pfmap, 1, A_TP_RSS_PF_MAP); } /** * t4_read_rss_pf_mask - read PF RSS Mask * @adapter: the adapter * * Reads the PF RSS Mask register and returns its value. */ u32 t4_read_rss_pf_mask(struct adapter *adapter) { u32 pfmask; if (t4_use_ldst(adapter)) t4_fw_tp_pio_rw(adapter, &pfmask, 1, A_TP_RSS_PF_MSK, 1); else t4_read_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &pfmask, 1, A_TP_RSS_PF_MSK); return pfmask; } /** * t4_write_rss_pf_mask - write PF RSS Mask * @adapter: the adapter * @pfmask: PF RSS Mask value * * Writes the specified value to the PF RSS Mask register. */ void t4_write_rss_pf_mask(struct adapter *adapter, u32 pfmask) { if (t4_use_ldst(adapter)) t4_fw_tp_pio_rw(adapter, &pfmask, 1, A_TP_RSS_PF_MSK, 0); else t4_write_indirect(adapter, A_TP_PIO_ADDR, A_TP_PIO_DATA, &pfmask, 1, A_TP_RSS_PF_MSK); } /** * t4_tp_get_tcp_stats - read TP's TCP MIB counters * @adap: the adapter * @v4: holds the TCP/IP counter values * @v6: holds the TCP/IPv6 counter values * * Returns the values of TP's TCP/IP and TCP/IPv6 MIB counters. * Either @v4 or @v6 may be %NULL to skip the corresponding stats. */ void t4_tp_get_tcp_stats(struct adapter *adap, struct tp_tcp_stats *v4, struct tp_tcp_stats *v6) { u32 val[A_TP_MIB_TCP_RXT_SEG_LO - A_TP_MIB_TCP_OUT_RST + 1]; #define STAT_IDX(x) ((A_TP_MIB_TCP_##x) - A_TP_MIB_TCP_OUT_RST) #define STAT(x) val[STAT_IDX(x)] #define STAT64(x) (((u64)STAT(x##_HI) << 32) | STAT(x##_LO)) if (v4) { t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, val, ARRAY_SIZE(val), A_TP_MIB_TCP_OUT_RST); v4->tcp_out_rsts = STAT(OUT_RST); v4->tcp_in_segs = STAT64(IN_SEG); v4->tcp_out_segs = STAT64(OUT_SEG); v4->tcp_retrans_segs = STAT64(RXT_SEG); } if (v6) { t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, val, ARRAY_SIZE(val), A_TP_MIB_TCP_V6OUT_RST); v6->tcp_out_rsts = STAT(OUT_RST); v6->tcp_in_segs = STAT64(IN_SEG); v6->tcp_out_segs = STAT64(OUT_SEG); v6->tcp_retrans_segs = STAT64(RXT_SEG); } #undef STAT64 #undef STAT #undef STAT_IDX } /** * t4_tp_get_err_stats - read TP's error MIB counters * @adap: the adapter * @st: holds the counter values * * Returns the values of TP's error counters. */ void t4_tp_get_err_stats(struct adapter *adap, struct tp_err_stats *st) { int nchan = adap->chip_params->nchan; t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->mac_in_errs, nchan, A_TP_MIB_MAC_IN_ERR_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->hdr_in_errs, nchan, A_TP_MIB_HDR_IN_ERR_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->tcp_in_errs, nchan, A_TP_MIB_TCP_IN_ERR_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->tnl_cong_drops, nchan, A_TP_MIB_TNL_CNG_DROP_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->ofld_chan_drops, nchan, A_TP_MIB_OFD_CHN_DROP_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->tnl_tx_drops, nchan, A_TP_MIB_TNL_DROP_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->ofld_vlan_drops, nchan, A_TP_MIB_OFD_VLN_DROP_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->tcp6_in_errs, nchan, A_TP_MIB_TCP_V6IN_ERR_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, &st->ofld_no_neigh, 2, A_TP_MIB_OFD_ARP_DROP); } /** * t4_tp_get_proxy_stats - read TP's proxy MIB counters * @adap: the adapter * @st: holds the counter values * * Returns the values of TP's proxy counters. */ void t4_tp_get_proxy_stats(struct adapter *adap, struct tp_proxy_stats *st) { int nchan = adap->chip_params->nchan; t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->proxy, nchan, A_TP_MIB_TNL_LPBK_0); } /** * t4_tp_get_cpl_stats - read TP's CPL MIB counters * @adap: the adapter * @st: holds the counter values * * Returns the values of TP's CPL counters. */ void t4_tp_get_cpl_stats(struct adapter *adap, struct tp_cpl_stats *st) { int nchan = adap->chip_params->nchan; t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->req, nchan, A_TP_MIB_CPL_IN_REQ_0); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, st->rsp, nchan, A_TP_MIB_CPL_OUT_RSP_0); } /** * t4_tp_get_rdma_stats - read TP's RDMA MIB counters * @adap: the adapter * @st: holds the counter values * * Returns the values of TP's RDMA counters. */ void t4_tp_get_rdma_stats(struct adapter *adap, struct tp_rdma_stats *st) { t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, &st->rqe_dfr_pkt, 2, A_TP_MIB_RQE_DFR_PKT); } /** * t4_get_fcoe_stats - read TP's FCoE MIB counters for a port * @adap: the adapter * @idx: the port index * @st: holds the counter values * * Returns the values of TP's FCoE counters for the selected port. */ void t4_get_fcoe_stats(struct adapter *adap, unsigned int idx, struct tp_fcoe_stats *st) { u32 val[2]; t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, &st->frames_ddp, 1, A_TP_MIB_FCOE_DDP_0 + idx); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, &st->frames_drop, 1, A_TP_MIB_FCOE_DROP_0 + idx); t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, val, 2, A_TP_MIB_FCOE_BYTE_0_HI + 2 * idx); st->octets_ddp = ((u64)val[0] << 32) | val[1]; } /** * t4_get_usm_stats - read TP's non-TCP DDP MIB counters * @adap: the adapter * @st: holds the counter values * * Returns the values of TP's counters for non-TCP directly-placed packets. */ void t4_get_usm_stats(struct adapter *adap, struct tp_usm_stats *st) { u32 val[4]; t4_read_indirect(adap, A_TP_MIB_INDEX, A_TP_MIB_DATA, val, 4, A_TP_MIB_USM_PKTS); st->frames = val[0]; st->drops = val[1]; st->octets = ((u64)val[2] << 32) | val[3]; } /** * t4_read_mtu_tbl - returns the values in the HW path MTU table * @adap: the adapter * @mtus: where to store the MTU values * @mtu_log: where to store the MTU base-2 log (may be %NULL) * * Reads the HW path MTU table. */ void t4_read_mtu_tbl(struct adapter *adap, u16 *mtus, u8 *mtu_log) { u32 v; int i; for (i = 0; i < NMTUS; ++i) { t4_write_reg(adap, A_TP_MTU_TABLE, V_MTUINDEX(0xff) | V_MTUVALUE(i)); v = t4_read_reg(adap, A_TP_MTU_TABLE); mtus[i] = G_MTUVALUE(v); if (mtu_log) mtu_log[i] = G_MTUWIDTH(v); } } /** * t4_read_cong_tbl - reads the congestion control table * @adap: the adapter * @incr: where to store the alpha values * * Reads the additive increments programmed into the HW congestion * control table. */ void t4_read_cong_tbl(struct adapter *adap, u16 incr[NMTUS][NCCTRL_WIN]) { unsigned int mtu, w; for (mtu = 0; mtu < NMTUS; ++mtu) for (w = 0; w < NCCTRL_WIN; ++w) { t4_write_reg(adap, A_TP_CCTRL_TABLE, V_ROWINDEX(0xffff) | (mtu << 5) | w); incr[mtu][w] = (u16)t4_read_reg(adap, A_TP_CCTRL_TABLE) & 0x1fff; } } /** * t4_tp_wr_bits_indirect - set/clear bits in an indirect TP register * @adap: the adapter * @addr: the indirect TP register address * @mask: specifies the field within the register to modify * @val: new value for the field * * Sets a field of an indirect TP register to the given value. */ void t4_tp_wr_bits_indirect(struct adapter *adap, unsigned int addr, unsigned int mask, unsigned int val) { t4_write_reg(adap, A_TP_PIO_ADDR, addr); val |= t4_read_reg(adap, A_TP_PIO_DATA) & ~mask; t4_write_reg(adap, A_TP_PIO_DATA, val); } /** * init_cong_ctrl - initialize congestion control parameters * @a: the alpha values for congestion control * @b: the beta values for congestion control * * Initialize the congestion control parameters. */ static void init_cong_ctrl(unsigned short *a, unsigned short *b) { a[0] = a[1] = a[2] = a[3] = a[4] = a[5] = a[6] = a[7] = a[8] = 1; a[9] = 2; a[10] = 3; a[11] = 4; a[12] = 5; a[13] = 6; a[14] = 7; a[15] = 8; a[16] = 9; a[17] = 10; a[18] = 14; a[19] = 17; a[20] = 21; a[21] = 25; a[22] = 30; a[23] = 35; a[24] = 45; a[25] = 60; a[26] = 80; a[27] = 100; a[28] = 200; a[29] = 300; a[30] = 400; a[31] = 500; b[0] = b[1] = b[2] = b[3] = b[4] = b[5] = b[6] = b[7] = b[8] = 0; b[9] = b[10] = 1; b[11] = b[12] = 2; b[13] = b[14] = b[15] = b[16] = 3; b[17] = b[18] = b[19] = b[20] = b[21] = 4; b[22] = b[23] = b[24] = b[25] = b[26] = b[27] = 5; b[28] = b[29] = 6; b[30] = b[31] = 7; } /* The minimum additive increment value for the congestion control table */ #define CC_MIN_INCR 2U /** * t4_load_mtus - write the MTU and congestion control HW tables * @adap: the adapter * @mtus: the values for the MTU table * @alpha: the values for the congestion control alpha parameter * @beta: the values for the congestion control beta parameter * * Write the HW MTU table with the supplied MTUs and the high-speed * congestion control table with the supplied alpha, beta, and MTUs. * We write the two tables together because the additive increments * depend on the MTUs. */ void t4_load_mtus(struct adapter *adap, const unsigned short *mtus, const unsigned short *alpha, const unsigned short *beta) { static const unsigned int avg_pkts[NCCTRL_WIN] = { 2, 6, 10, 14, 20, 28, 40, 56, 80, 112, 160, 224, 320, 448, 640, 896, 1281, 1792, 2560, 3584, 5120, 7168, 10240, 14336, 20480, 28672, 40960, 57344, 81920, 114688, 163840, 229376 }; unsigned int i, w; for (i = 0; i < NMTUS; ++i) { unsigned int mtu = mtus[i]; unsigned int log2 = fls(mtu); if (!(mtu & ((1 << log2) >> 2))) /* round */ log2--; t4_write_reg(adap, A_TP_MTU_TABLE, V_MTUINDEX(i) | V_MTUWIDTH(log2) | V_MTUVALUE(mtu)); for (w = 0; w < NCCTRL_WIN; ++w) { unsigned int inc; inc = max(((mtu - 40) * alpha[w]) / avg_pkts[w], CC_MIN_INCR); t4_write_reg(adap, A_TP_CCTRL_TABLE, (i << 21) | (w << 16) | (beta[w] << 13) | inc); } } } /** * t4_set_pace_tbl - set the pace table * @adap: the adapter * @pace_vals: the pace values in microseconds * @start: index of the first entry in the HW pace table to set * @n: how many entries to set * * Sets (a subset of the) HW pace table. */ int t4_set_pace_tbl(struct adapter *adap, const unsigned int *pace_vals, unsigned int start, unsigned int n) { unsigned int vals[NTX_SCHED], i; unsigned int tick_ns = dack_ticks_to_usec(adap, 1000); if (n > NTX_SCHED) return -ERANGE; /* convert values from us to dack ticks, rounding to closest value */ for (i = 0; i < n; i++, pace_vals++) { vals[i] = (1000 * *pace_vals + tick_ns / 2) / tick_ns; if (vals[i] > 0x7ff) return -ERANGE; if (*pace_vals && vals[i] == 0) return -ERANGE; } for (i = 0; i < n; i++, start++) t4_write_reg(adap, A_TP_PACE_TABLE, (start << 16) | vals[i]); return 0; } /** * t4_set_sched_bps - set the bit rate for a HW traffic scheduler * @adap: the adapter * @kbps: target rate in Kbps * @sched: the scheduler index * * Configure a Tx HW scheduler for the target rate. */ int t4_set_sched_bps(struct adapter *adap, int sched, unsigned int kbps) { unsigned int v, tps, cpt, bpt, delta, mindelta = ~0; unsigned int clk = adap->params.vpd.cclk * 1000; unsigned int selected_cpt = 0, selected_bpt = 0; if (kbps > 0) { kbps *= 125; /* -> bytes */ for (cpt = 1; cpt <= 255; cpt++) { tps = clk / cpt; bpt = (kbps + tps / 2) / tps; if (bpt > 0 && bpt <= 255) { v = bpt * tps; delta = v >= kbps ? v - kbps : kbps - v; if (delta < mindelta) { mindelta = delta; selected_cpt = cpt; selected_bpt = bpt; } } else if (selected_cpt) break; } if (!selected_cpt) return -EINVAL; } t4_write_reg(adap, A_TP_TM_PIO_ADDR, A_TP_TX_MOD_Q1_Q0_RATE_LIMIT - sched / 2); v = t4_read_reg(adap, A_TP_TM_PIO_DATA); if (sched & 1) v = (v & 0xffff) | (selected_cpt << 16) | (selected_bpt << 24); else v = (v & 0xffff0000) | selected_cpt | (selected_bpt << 8); t4_write_reg(adap, A_TP_TM_PIO_DATA, v); return 0; } /** * t4_set_sched_ipg - set the IPG for a Tx HW packet rate scheduler * @adap: the adapter * @sched: the scheduler index * @ipg: the interpacket delay in tenths of nanoseconds * * Set the interpacket delay for a HW packet rate scheduler. */ int t4_set_sched_ipg(struct adapter *adap, int sched, unsigned int ipg) { unsigned int v, addr = A_TP_TX_MOD_Q1_Q0_TIMER_SEPARATOR - sched / 2; /* convert ipg to nearest number of core clocks */ ipg *= core_ticks_per_usec(adap); ipg = (ipg + 5000) / 10000; if (ipg > M_TXTIMERSEPQ0) return -EINVAL; t4_write_reg(adap, A_TP_TM_PIO_ADDR, addr); v = t4_read_reg(adap, A_TP_TM_PIO_DATA); if (sched & 1) v = (v & V_TXTIMERSEPQ0(M_TXTIMERSEPQ0)) | V_TXTIMERSEPQ1(ipg); else v = (v & V_TXTIMERSEPQ1(M_TXTIMERSEPQ1)) | V_TXTIMERSEPQ0(ipg); t4_write_reg(adap, A_TP_TM_PIO_DATA, v); t4_read_reg(adap, A_TP_TM_PIO_DATA); return 0; } /* * Calculates a rate in bytes/s given the number of 256-byte units per 4K core * clocks. The formula is * * bytes/s = bytes256 * 256 * ClkFreq / 4096 * * which is equivalent to * * bytes/s = 62.5 * bytes256 * ClkFreq_ms */ static u64 chan_rate(struct adapter *adap, unsigned int bytes256) { u64 v = bytes256 * adap->params.vpd.cclk; return v * 62 + v / 2; } /** * t4_get_chan_txrate - get the current per channel Tx rates * @adap: the adapter * @nic_rate: rates for NIC traffic * @ofld_rate: rates for offloaded traffic * * Return the current Tx rates in bytes/s for NIC and offloaded traffic * for each channel. */ void t4_get_chan_txrate(struct adapter *adap, u64 *nic_rate, u64 *ofld_rate) { u32 v; v = t4_read_reg(adap, A_TP_TX_TRATE); nic_rate[0] = chan_rate(adap, G_TNLRATE0(v)); nic_rate[1] = chan_rate(adap, G_TNLRATE1(v)); if (adap->chip_params->nchan > 2) { nic_rate[2] = chan_rate(adap, G_TNLRATE2(v)); nic_rate[3] = chan_rate(adap, G_TNLRATE3(v)); } v = t4_read_reg(adap, A_TP_TX_ORATE); ofld_rate[0] = chan_rate(adap, G_OFDRATE0(v)); ofld_rate[1] = chan_rate(adap, G_OFDRATE1(v)); if (adap->chip_params->nchan > 2) { ofld_rate[2] = chan_rate(adap, G_OFDRATE2(v)); ofld_rate[3] = chan_rate(adap, G_OFDRATE3(v)); } } /** * t4_set_trace_filter - configure one of the tracing filters * @adap: the adapter * @tp: the desired trace filter parameters * @idx: which filter to configure * @enable: whether to enable or disable the filter * * Configures one of the tracing filters available in HW. If @tp is %NULL * it indicates that the filter is already written in the register and it * just needs to be enabled or disabled. */ int t4_set_trace_filter(struct adapter *adap, const struct trace_params *tp, int idx, int enable) { int i, ofst = idx * 4; u32 data_reg, mask_reg, cfg; u32 multitrc = F_TRCMULTIFILTER; u32 en = is_t4(adap) ? F_TFEN : F_T5_TFEN; if (idx < 0 || idx >= NTRACE) return -EINVAL; if (tp == NULL || !enable) { t4_set_reg_field(adap, A_MPS_TRC_FILTER_MATCH_CTL_A + ofst, en, enable ? en : 0); return 0; } /* * TODO - After T4 data book is updated, specify the exact * section below. * * See T4 data book - MPS section for a complete description * of the below if..else handling of A_MPS_TRC_CFG register * value. */ cfg = t4_read_reg(adap, A_MPS_TRC_CFG); if (cfg & F_TRCMULTIFILTER) { /* * If multiple tracers are enabled, then maximum * capture size is 2.5KB (FIFO size of a single channel) * minus 2 flits for CPL_TRACE_PKT header. */ if (tp->snap_len > ((10 * 1024 / 4) - (2 * 8))) return -EINVAL; } else { /* * If multiple tracers are disabled, to avoid deadlocks * maximum packet capture size of 9600 bytes is recommended. * Also in this mode, only trace0 can be enabled and running. */ multitrc = 0; if (tp->snap_len > 9600 || idx) return -EINVAL; } if (tp->port > (is_t4(adap) ? 11 : 19) || tp->invert > 1 || tp->skip_len > M_TFLENGTH || tp->skip_ofst > M_TFOFFSET || tp->min_len > M_TFMINPKTSIZE) return -EINVAL; /* stop the tracer we'll be changing */ t4_set_reg_field(adap, A_MPS_TRC_FILTER_MATCH_CTL_A + ofst, en, 0); idx *= (A_MPS_TRC_FILTER1_MATCH - A_MPS_TRC_FILTER0_MATCH); data_reg = A_MPS_TRC_FILTER0_MATCH + idx; mask_reg = A_MPS_TRC_FILTER0_DONT_CARE + idx; for (i = 0; i < TRACE_LEN / 4; i++, data_reg += 4, mask_reg += 4) { t4_write_reg(adap, data_reg, tp->data[i]); t4_write_reg(adap, mask_reg, ~tp->mask[i]); } t4_write_reg(adap, A_MPS_TRC_FILTER_MATCH_CTL_B + ofst, V_TFCAPTUREMAX(tp->snap_len) | V_TFMINPKTSIZE(tp->min_len)); t4_write_reg(adap, A_MPS_TRC_FILTER_MATCH_CTL_A + ofst, V_TFOFFSET(tp->skip_ofst) | V_TFLENGTH(tp->skip_len) | en | (is_t4(adap) ? V_TFPORT(tp->port) | V_TFINVERTMATCH(tp->invert) : V_T5_TFPORT(tp->port) | V_T5_TFINVERTMATCH(tp->invert))); return 0; } /** * t4_get_trace_filter - query one of the tracing filters * @adap: the adapter * @tp: the current trace filter parameters * @idx: which trace filter to query * @enabled: non-zero if the filter is enabled * * Returns the current settings of one of the HW tracing filters. */ void t4_get_trace_filter(struct adapter *adap, struct trace_params *tp, int idx, int *enabled) { u32 ctla, ctlb; int i, ofst = idx * 4; u32 data_reg, mask_reg; ctla = t4_read_reg(adap, A_MPS_TRC_FILTER_MATCH_CTL_A + ofst); ctlb = t4_read_reg(adap, A_MPS_TRC_FILTER_MATCH_CTL_B + ofst); if (is_t4(adap)) { *enabled = !!(ctla & F_TFEN); tp->port = G_TFPORT(ctla); tp->invert = !!(ctla & F_TFINVERTMATCH); } else { *enabled = !!(ctla & F_T5_TFEN); tp->port = G_T5_TFPORT(ctla); tp->invert = !!(ctla & F_T5_TFINVERTMATCH); } tp->snap_len = G_TFCAPTUREMAX(ctlb); tp->min_len = G_TFMINPKTSIZE(ctlb); tp->skip_ofst = G_TFOFFSET(ctla); tp->skip_len = G_TFLENGTH(ctla); ofst = (A_MPS_TRC_FILTER1_MATCH - A_MPS_TRC_FILTER0_MATCH) * idx; data_reg = A_MPS_TRC_FILTER0_MATCH + ofst; mask_reg = A_MPS_TRC_FILTER0_DONT_CARE + ofst; for (i = 0; i < TRACE_LEN / 4; i++, data_reg += 4, mask_reg += 4) { tp->mask[i] = ~t4_read_reg(adap, mask_reg); tp->data[i] = t4_read_reg(adap, data_reg) & tp->mask[i]; } } /** * t4_pmtx_get_stats - returns the HW stats from PMTX * @adap: the adapter * @cnt: where to store the count statistics * @cycles: where to store the cycle statistics * * Returns performance statistics from PMTX. */ void t4_pmtx_get_stats(struct adapter *adap, u32 cnt[], u64 cycles[]) { int i; u32 data[2]; for (i = 0; i < adap->chip_params->pm_stats_cnt; i++) { t4_write_reg(adap, A_PM_TX_STAT_CONFIG, i + 1); cnt[i] = t4_read_reg(adap, A_PM_TX_STAT_COUNT); if (is_t4(adap)) cycles[i] = t4_read_reg64(adap, A_PM_TX_STAT_LSB); else { t4_read_indirect(adap, A_PM_TX_DBG_CTRL, A_PM_TX_DBG_DATA, data, 2, A_PM_TX_DBG_STAT_MSB); cycles[i] = (((u64)data[0] << 32) | data[1]); } } } /** * t4_pmrx_get_stats - returns the HW stats from PMRX * @adap: the adapter * @cnt: where to store the count statistics * @cycles: where to store the cycle statistics * * Returns performance statistics from PMRX. */ void t4_pmrx_get_stats(struct adapter *adap, u32 cnt[], u64 cycles[]) { int i; u32 data[2]; for (i = 0; i < adap->chip_params->pm_stats_cnt; i++) { t4_write_reg(adap, A_PM_RX_STAT_CONFIG, i + 1); cnt[i] = t4_read_reg(adap, A_PM_RX_STAT_COUNT); if (is_t4(adap)) { cycles[i] = t4_read_reg64(adap, A_PM_RX_STAT_LSB); } else { t4_read_indirect(adap, A_PM_RX_DBG_CTRL, A_PM_RX_DBG_DATA, data, 2, A_PM_RX_DBG_STAT_MSB); cycles[i] = (((u64)data[0] << 32) | data[1]); } } } /** * t4_get_mps_bg_map - return the buffer groups associated with a port * @adap: the adapter * @idx: the port index * * Returns a bitmap indicating which MPS buffer groups are associated * with the given port. Bit i is set if buffer group i is used by the * port. */ static unsigned int t4_get_mps_bg_map(struct adapter *adap, int idx) { u32 n = G_NUMPORTS(t4_read_reg(adap, A_MPS_CMN_CTL)); if (n == 0) return idx == 0 ? 0xf : 0; if (n == 1 && chip_id(adap) <= CHELSIO_T5) return idx < 2 ? (3 << (2 * idx)) : 0; return 1 << idx; } /** * t4_get_port_type_description - return Port Type string description * @port_type: firmware Port Type enumeration */ const char *t4_get_port_type_description(enum fw_port_type port_type) { static const char *const port_type_description[] = { "Fiber_XFI", "Fiber_XAUI", "BT_SGMII", "BT_XFI", "BT_XAUI", "KX4", "CX4", "KX", "KR", "SFP", "BP_AP", "BP4_AP", "QSFP_10G", "QSA", "QSFP", "BP40_BA", }; if (port_type < ARRAY_SIZE(port_type_description)) return port_type_description[port_type]; return "UNKNOWN"; } /** * t4_get_port_stats_offset - collect port stats relative to a previous * snapshot * @adap: The adapter * @idx: The port * @stats: Current stats to fill * @offset: Previous stats snapshot */ void t4_get_port_stats_offset(struct adapter *adap, int idx, struct port_stats *stats, struct port_stats *offset) { u64 *s, *o; int i; t4_get_port_stats(adap, idx, stats); for (i = 0, s = (u64 *)stats, o = (u64 *)offset ; i < (sizeof(struct port_stats)/sizeof(u64)) ; i++, s++, o++) *s -= *o; } /** * t4_get_port_stats - collect port statistics * @adap: the adapter * @idx: the port index * @p: the stats structure to fill * * Collect statistics related to the given port from HW. */ void t4_get_port_stats(struct adapter *adap, int idx, struct port_stats *p) { u32 bgmap = t4_get_mps_bg_map(adap, idx); u32 stat_ctl; #define GET_STAT(name) \ t4_read_reg64(adap, \ (is_t4(adap) ? PORT_REG(idx, A_MPS_PORT_STAT_##name##_L) : \ T5_PORT_REG(idx, A_MPS_PORT_STAT_##name##_L))) #define GET_STAT_COM(name) t4_read_reg64(adap, A_MPS_STAT_##name##_L) stat_ctl = t4_read_reg(adap, A_MPS_STAT_CTL); p->tx_pause = GET_STAT(TX_PORT_PAUSE); p->tx_octets = GET_STAT(TX_PORT_BYTES); p->tx_frames = GET_STAT(TX_PORT_FRAMES); p->tx_bcast_frames = GET_STAT(TX_PORT_BCAST); p->tx_mcast_frames = GET_STAT(TX_PORT_MCAST); p->tx_ucast_frames = GET_STAT(TX_PORT_UCAST); p->tx_error_frames = GET_STAT(TX_PORT_ERROR); p->tx_frames_64 = GET_STAT(TX_PORT_64B); p->tx_frames_65_127 = GET_STAT(TX_PORT_65B_127B); p->tx_frames_128_255 = GET_STAT(TX_PORT_128B_255B); p->tx_frames_256_511 = GET_STAT(TX_PORT_256B_511B); p->tx_frames_512_1023 = GET_STAT(TX_PORT_512B_1023B); p->tx_frames_1024_1518 = GET_STAT(TX_PORT_1024B_1518B); p->tx_frames_1519_max = GET_STAT(TX_PORT_1519B_MAX); p->tx_drop = GET_STAT(TX_PORT_DROP); p->tx_ppp0 = GET_STAT(TX_PORT_PPP0); p->tx_ppp1 = GET_STAT(TX_PORT_PPP1); p->tx_ppp2 = GET_STAT(TX_PORT_PPP2); p->tx_ppp3 = GET_STAT(TX_PORT_PPP3); p->tx_ppp4 = GET_STAT(TX_PORT_PPP4); p->tx_ppp5 = GET_STAT(TX_PORT_PPP5); p->tx_ppp6 = GET_STAT(TX_PORT_PPP6); p->tx_ppp7 = GET_STAT(TX_PORT_PPP7); if (stat_ctl & F_COUNTPAUSESTATTX) { p->tx_frames -= p->tx_pause; p->tx_octets -= p->tx_pause * 64; p->tx_mcast_frames -= p->tx_pause; } p->rx_pause = GET_STAT(RX_PORT_PAUSE); p->rx_octets = GET_STAT(RX_PORT_BYTES); p->rx_frames = GET_STAT(RX_PORT_FRAMES); p->rx_bcast_frames = GET_STAT(RX_PORT_BCAST); p->rx_mcast_frames = GET_STAT(RX_PORT_MCAST); p->rx_ucast_frames = GET_STAT(RX_PORT_UCAST); p->rx_too_long = GET_STAT(RX_PORT_MTU_ERROR); p->rx_jabber = GET_STAT(RX_PORT_MTU_CRC_ERROR); p->rx_fcs_err = GET_STAT(RX_PORT_CRC_ERROR); p->rx_len_err = GET_STAT(RX_PORT_LEN_ERROR); p->rx_symbol_err = GET_STAT(RX_PORT_SYM_ERROR); p->rx_runt = GET_STAT(RX_PORT_LESS_64B); p->rx_frames_64 = GET_STAT(RX_PORT_64B); p->rx_frames_65_127 = GET_STAT(RX_PORT_65B_127B); p->rx_frames_128_255 = GET_STAT(RX_PORT_128B_255B); p->rx_frames_256_511 = GET_STAT(RX_PORT_256B_511B); p->rx_frames_512_1023 = GET_STAT(RX_PORT_512B_1023B); p->rx_frames_1024_1518 = GET_STAT(RX_PORT_1024B_1518B); p->rx_frames_1519_max = GET_STAT(RX_PORT_1519B_MAX); p->rx_ppp0 = GET_STAT(RX_PORT_PPP0); p->rx_ppp1 = GET_STAT(RX_PORT_PPP1); p->rx_ppp2 = GET_STAT(RX_PORT_PPP2); p->rx_ppp3 = GET_STAT(RX_PORT_PPP3); p->rx_ppp4 = GET_STAT(RX_PORT_PPP4); p->rx_ppp5 = GET_STAT(RX_PORT_PPP5); p->rx_ppp6 = GET_STAT(RX_PORT_PPP6); p->rx_ppp7 = GET_STAT(RX_PORT_PPP7); if (stat_ctl & F_COUNTPAUSESTATRX) { p->rx_frames -= p->rx_pause; p->rx_octets -= p->rx_pause * 64; p->rx_mcast_frames -= p->rx_pause; } p->rx_ovflow0 = (bgmap & 1) ? GET_STAT_COM(RX_BG_0_MAC_DROP_FRAME) : 0; p->rx_ovflow1 = (bgmap & 2) ? GET_STAT_COM(RX_BG_1_MAC_DROP_FRAME) : 0; p->rx_ovflow2 = (bgmap & 4) ? GET_STAT_COM(RX_BG_2_MAC_DROP_FRAME) : 0; p->rx_ovflow3 = (bgmap & 8) ? GET_STAT_COM(RX_BG_3_MAC_DROP_FRAME) : 0; p->rx_trunc0 = (bgmap & 1) ? GET_STAT_COM(RX_BG_0_MAC_TRUNC_FRAME) : 0; p->rx_trunc1 = (bgmap & 2) ? GET_STAT_COM(RX_BG_1_MAC_TRUNC_FRAME) : 0; p->rx_trunc2 = (bgmap & 4) ? GET_STAT_COM(RX_BG_2_MAC_TRUNC_FRAME) : 0; p->rx_trunc3 = (bgmap & 8) ? GET_STAT_COM(RX_BG_3_MAC_TRUNC_FRAME) : 0; #undef GET_STAT #undef GET_STAT_COM } /** * t4_get_lb_stats - collect loopback port statistics * @adap: the adapter * @idx: the loopback port index * @p: the stats structure to fill * * Return HW statistics for the given loopback port. */ void t4_get_lb_stats(struct adapter *adap, int idx, struct lb_port_stats *p) { u32 bgmap = t4_get_mps_bg_map(adap, idx); #define GET_STAT(name) \ t4_read_reg64(adap, \ (is_t4(adap) ? \ PORT_REG(idx, A_MPS_PORT_STAT_LB_PORT_##name##_L) : \ T5_PORT_REG(idx, A_MPS_PORT_STAT_LB_PORT_##name##_L))) #define GET_STAT_COM(name) t4_read_reg64(adap, A_MPS_STAT_##name##_L) p->octets = GET_STAT(BYTES); p->frames = GET_STAT(FRAMES); p->bcast_frames = GET_STAT(BCAST); p->mcast_frames = GET_STAT(MCAST); p->ucast_frames = GET_STAT(UCAST); p->error_frames = GET_STAT(ERROR); p->frames_64 = GET_STAT(64B); p->frames_65_127 = GET_STAT(65B_127B); p->frames_128_255 = GET_STAT(128B_255B); p->frames_256_511 = GET_STAT(256B_511B); p->frames_512_1023 = GET_STAT(512B_1023B); p->frames_1024_1518 = GET_STAT(1024B_1518B); p->frames_1519_max = GET_STAT(1519B_MAX); p->drop = GET_STAT(DROP_FRAMES); p->ovflow0 = (bgmap & 1) ? GET_STAT_COM(RX_BG_0_LB_DROP_FRAME) : 0; p->ovflow1 = (bgmap & 2) ? GET_STAT_COM(RX_BG_1_LB_DROP_FRAME) : 0; p->ovflow2 = (bgmap & 4) ? GET_STAT_COM(RX_BG_2_LB_DROP_FRAME) : 0; p->ovflow3 = (bgmap & 8) ? GET_STAT_COM(RX_BG_3_LB_DROP_FRAME) : 0; p->trunc0 = (bgmap & 1) ? GET_STAT_COM(RX_BG_0_LB_TRUNC_FRAME) : 0; p->trunc1 = (bgmap & 2) ? GET_STAT_COM(RX_BG_1_LB_TRUNC_FRAME) : 0; p->trunc2 = (bgmap & 4) ? GET_STAT_COM(RX_BG_2_LB_TRUNC_FRAME) : 0; p->trunc3 = (bgmap & 8) ? GET_STAT_COM(RX_BG_3_LB_TRUNC_FRAME) : 0; #undef GET_STAT #undef GET_STAT_COM } /** * t4_wol_magic_enable - enable/disable magic packet WoL * @adap: the adapter * @port: the physical port index * @addr: MAC address expected in magic packets, %NULL to disable * * Enables/disables magic packet wake-on-LAN for the selected port. */ void t4_wol_magic_enable(struct adapter *adap, unsigned int port, const u8 *addr) { u32 mag_id_reg_l, mag_id_reg_h, port_cfg_reg; if (is_t4(adap)) { mag_id_reg_l = PORT_REG(port, A_XGMAC_PORT_MAGIC_MACID_LO); mag_id_reg_h = PORT_REG(port, A_XGMAC_PORT_MAGIC_MACID_HI); port_cfg_reg = PORT_REG(port, A_XGMAC_PORT_CFG2); } else { mag_id_reg_l = T5_PORT_REG(port, A_MAC_PORT_MAGIC_MACID_LO); mag_id_reg_h = T5_PORT_REG(port, A_MAC_PORT_MAGIC_MACID_HI); port_cfg_reg = T5_PORT_REG(port, A_MAC_PORT_CFG2); } if (addr) { t4_write_reg(adap, mag_id_reg_l, (addr[2] << 24) | (addr[3] << 16) | (addr[4] << 8) | addr[5]); t4_write_reg(adap, mag_id_reg_h, (addr[0] << 8) | addr[1]); } t4_set_reg_field(adap, port_cfg_reg, F_MAGICEN, V_MAGICEN(addr != NULL)); } /** * t4_wol_pat_enable - enable/disable pattern-based WoL * @adap: the adapter * @port: the physical port index * @map: bitmap of which HW pattern filters to set * @mask0: byte mask for bytes 0-63 of a packet * @mask1: byte mask for bytes 64-127 of a packet * @crc: Ethernet CRC for selected bytes * @enable: enable/disable switch * * Sets the pattern filters indicated in @map to mask out the bytes * specified in @mask0/@mask1 in received packets and compare the CRC of * the resulting packet against @crc. If @enable is %true pattern-based * WoL is enabled, otherwise disabled. */ int t4_wol_pat_enable(struct adapter *adap, unsigned int port, unsigned int map, u64 mask0, u64 mask1, unsigned int crc, bool enable) { int i; u32 port_cfg_reg; if (is_t4(adap)) port_cfg_reg = PORT_REG(port, A_XGMAC_PORT_CFG2); else port_cfg_reg = T5_PORT_REG(port, A_MAC_PORT_CFG2); if (!enable) { t4_set_reg_field(adap, port_cfg_reg, F_PATEN, 0); return 0; } if (map > 0xff) return -EINVAL; #define EPIO_REG(name) \ (is_t4(adap) ? PORT_REG(port, A_XGMAC_PORT_EPIO_##name) : \ T5_PORT_REG(port, A_MAC_PORT_EPIO_##name)) t4_write_reg(adap, EPIO_REG(DATA1), mask0 >> 32); t4_write_reg(adap, EPIO_REG(DATA2), mask1); t4_write_reg(adap, EPIO_REG(DATA3), mask1 >> 32); for (i = 0; i < NWOL_PAT; i++, map >>= 1) { if (!(map & 1)) continue; /* write byte masks */ t4_write_reg(adap, EPIO_REG(DATA0), mask0); t4_write_reg(adap, EPIO_REG(OP), V_ADDRESS(i) | F_EPIOWR); t4_read_reg(adap, EPIO_REG(OP)); /* flush */ if (t4_read_reg(adap, EPIO_REG(OP)) & F_BUSY) return -ETIMEDOUT; /* write CRC */ t4_write_reg(adap, EPIO_REG(DATA0), crc); t4_write_reg(adap, EPIO_REG(OP), V_ADDRESS(i + 32) | F_EPIOWR); t4_read_reg(adap, EPIO_REG(OP)); /* flush */ if (t4_read_reg(adap, EPIO_REG(OP)) & F_BUSY) return -ETIMEDOUT; } #undef EPIO_REG t4_set_reg_field(adap, port_cfg_reg, 0, F_PATEN); return 0; } /* t4_mk_filtdelwr - create a delete filter WR * @ftid: the filter ID * @wr: the filter work request to populate * @qid: ingress queue to receive the delete notification * * Creates a filter work request to delete the supplied filter. If @qid is * negative the delete notification is suppressed. */ void t4_mk_filtdelwr(unsigned int ftid, struct fw_filter_wr *wr, int qid) { memset(wr, 0, sizeof(*wr)); wr->op_pkd = cpu_to_be32(V_FW_WR_OP(FW_FILTER_WR)); wr->len16_pkd = cpu_to_be32(V_FW_WR_LEN16(sizeof(*wr) / 16)); wr->tid_to_iq = cpu_to_be32(V_FW_FILTER_WR_TID(ftid) | V_FW_FILTER_WR_NOREPLY(qid < 0)); wr->del_filter_to_l2tix = cpu_to_be32(F_FW_FILTER_WR_DEL_FILTER); if (qid >= 0) wr->rx_chan_rx_rpl_iq = cpu_to_be16(V_FW_FILTER_WR_RX_RPL_IQ(qid)); } #define INIT_CMD(var, cmd, rd_wr) do { \ (var).op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_##cmd##_CMD) | \ F_FW_CMD_REQUEST | \ F_FW_CMD_##rd_wr); \ (var).retval_len16 = cpu_to_be32(FW_LEN16(var)); \ } while (0) int t4_fwaddrspace_write(struct adapter *adap, unsigned int mbox, u32 addr, u32 val) { u32 ldst_addrspace; struct fw_ldst_cmd c; memset(&c, 0, sizeof(c)); ldst_addrspace = V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_FIRMWARE); c.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | ldst_addrspace); c.cycles_to_len16 = cpu_to_be32(FW_LEN16(c)); c.u.addrval.addr = cpu_to_be32(addr); c.u.addrval.val = cpu_to_be32(val); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_mdio_rd - read a PHY register through MDIO * @adap: the adapter * @mbox: mailbox to use for the FW command * @phy_addr: the PHY address * @mmd: the PHY MMD to access (0 for clause 22 PHYs) * @reg: the register to read * @valp: where to store the value * * Issues a FW command through the given mailbox to read a PHY register. */ int t4_mdio_rd(struct adapter *adap, unsigned int mbox, unsigned int phy_addr, unsigned int mmd, unsigned int reg, unsigned int *valp) { int ret; u32 ldst_addrspace; struct fw_ldst_cmd c; memset(&c, 0, sizeof(c)); ldst_addrspace = V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_MDIO); c.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | ldst_addrspace); c.cycles_to_len16 = cpu_to_be32(FW_LEN16(c)); c.u.mdio.paddr_mmd = cpu_to_be16(V_FW_LDST_CMD_PADDR(phy_addr) | V_FW_LDST_CMD_MMD(mmd)); c.u.mdio.raddr = cpu_to_be16(reg); ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret == 0) *valp = be16_to_cpu(c.u.mdio.rval); return ret; } /** * t4_mdio_wr - write a PHY register through MDIO * @adap: the adapter * @mbox: mailbox to use for the FW command * @phy_addr: the PHY address * @mmd: the PHY MMD to access (0 for clause 22 PHYs) * @reg: the register to write * @valp: value to write * * Issues a FW command through the given mailbox to write a PHY register. */ int t4_mdio_wr(struct adapter *adap, unsigned int mbox, unsigned int phy_addr, unsigned int mmd, unsigned int reg, unsigned int val) { u32 ldst_addrspace; struct fw_ldst_cmd c; memset(&c, 0, sizeof(c)); ldst_addrspace = V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_MDIO); c.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | ldst_addrspace); c.cycles_to_len16 = cpu_to_be32(FW_LEN16(c)); c.u.mdio.paddr_mmd = cpu_to_be16(V_FW_LDST_CMD_PADDR(phy_addr) | V_FW_LDST_CMD_MMD(mmd)); c.u.mdio.raddr = cpu_to_be16(reg); c.u.mdio.rval = cpu_to_be16(val); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * * t4_sge_decode_idma_state - decode the idma state * @adap: the adapter * @state: the state idma is stuck in */ void t4_sge_decode_idma_state(struct adapter *adapter, int state) { static const char * const t4_decode[] = { "IDMA_IDLE", "IDMA_PUSH_MORE_CPL_FIFO", "IDMA_PUSH_CPL_MSG_HEADER_TO_FIFO", "Not used", "IDMA_PHYSADDR_SEND_PCIEHDR", "IDMA_PHYSADDR_SEND_PAYLOAD_FIRST", "IDMA_PHYSADDR_SEND_PAYLOAD", "IDMA_SEND_FIFO_TO_IMSG", "IDMA_FL_REQ_DATA_FL_PREP", "IDMA_FL_REQ_DATA_FL", "IDMA_FL_DROP", "IDMA_FL_H_REQ_HEADER_FL", "IDMA_FL_H_SEND_PCIEHDR", "IDMA_FL_H_PUSH_CPL_FIFO", "IDMA_FL_H_SEND_CPL", "IDMA_FL_H_SEND_IP_HDR_FIRST", "IDMA_FL_H_SEND_IP_HDR", "IDMA_FL_H_REQ_NEXT_HEADER_FL", "IDMA_FL_H_SEND_NEXT_PCIEHDR", "IDMA_FL_H_SEND_IP_HDR_PADDING", "IDMA_FL_D_SEND_PCIEHDR", "IDMA_FL_D_SEND_CPL_AND_IP_HDR", "IDMA_FL_D_REQ_NEXT_DATA_FL", "IDMA_FL_SEND_PCIEHDR", "IDMA_FL_PUSH_CPL_FIFO", "IDMA_FL_SEND_CPL", "IDMA_FL_SEND_PAYLOAD_FIRST", "IDMA_FL_SEND_PAYLOAD", "IDMA_FL_REQ_NEXT_DATA_FL", "IDMA_FL_SEND_NEXT_PCIEHDR", "IDMA_FL_SEND_PADDING", "IDMA_FL_SEND_COMPLETION_TO_IMSG", "IDMA_FL_SEND_FIFO_TO_IMSG", "IDMA_FL_REQ_DATAFL_DONE", "IDMA_FL_REQ_HEADERFL_DONE", }; static const char * const t5_decode[] = { "IDMA_IDLE", "IDMA_ALMOST_IDLE", "IDMA_PUSH_MORE_CPL_FIFO", "IDMA_PUSH_CPL_MSG_HEADER_TO_FIFO", "IDMA_SGEFLRFLUSH_SEND_PCIEHDR", "IDMA_PHYSADDR_SEND_PCIEHDR", "IDMA_PHYSADDR_SEND_PAYLOAD_FIRST", "IDMA_PHYSADDR_SEND_PAYLOAD", "IDMA_SEND_FIFO_TO_IMSG", "IDMA_FL_REQ_DATA_FL", "IDMA_FL_DROP", "IDMA_FL_DROP_SEND_INC", "IDMA_FL_H_REQ_HEADER_FL", "IDMA_FL_H_SEND_PCIEHDR", "IDMA_FL_H_PUSH_CPL_FIFO", "IDMA_FL_H_SEND_CPL", "IDMA_FL_H_SEND_IP_HDR_FIRST", "IDMA_FL_H_SEND_IP_HDR", "IDMA_FL_H_REQ_NEXT_HEADER_FL", "IDMA_FL_H_SEND_NEXT_PCIEHDR", "IDMA_FL_H_SEND_IP_HDR_PADDING", "IDMA_FL_D_SEND_PCIEHDR", "IDMA_FL_D_SEND_CPL_AND_IP_HDR", "IDMA_FL_D_REQ_NEXT_DATA_FL", "IDMA_FL_SEND_PCIEHDR", "IDMA_FL_PUSH_CPL_FIFO", "IDMA_FL_SEND_CPL", "IDMA_FL_SEND_PAYLOAD_FIRST", "IDMA_FL_SEND_PAYLOAD", "IDMA_FL_REQ_NEXT_DATA_FL", "IDMA_FL_SEND_NEXT_PCIEHDR", "IDMA_FL_SEND_PADDING", "IDMA_FL_SEND_COMPLETION_TO_IMSG", }; static const char * const t6_decode[] = { "IDMA_IDLE", "IDMA_PUSH_MORE_CPL_FIFO", "IDMA_PUSH_CPL_MSG_HEADER_TO_FIFO", "IDMA_SGEFLRFLUSH_SEND_PCIEHDR", "IDMA_PHYSADDR_SEND_PCIEHDR", "IDMA_PHYSADDR_SEND_PAYLOAD_FIRST", "IDMA_PHYSADDR_SEND_PAYLOAD", "IDMA_FL_REQ_DATA_FL", "IDMA_FL_DROP", "IDMA_FL_DROP_SEND_INC", "IDMA_FL_H_REQ_HEADER_FL", "IDMA_FL_H_SEND_PCIEHDR", "IDMA_FL_H_PUSH_CPL_FIFO", "IDMA_FL_H_SEND_CPL", "IDMA_FL_H_SEND_IP_HDR_FIRST", "IDMA_FL_H_SEND_IP_HDR", "IDMA_FL_H_REQ_NEXT_HEADER_FL", "IDMA_FL_H_SEND_NEXT_PCIEHDR", "IDMA_FL_H_SEND_IP_HDR_PADDING", "IDMA_FL_D_SEND_PCIEHDR", "IDMA_FL_D_SEND_CPL_AND_IP_HDR", "IDMA_FL_D_REQ_NEXT_DATA_FL", "IDMA_FL_SEND_PCIEHDR", "IDMA_FL_PUSH_CPL_FIFO", "IDMA_FL_SEND_CPL", "IDMA_FL_SEND_PAYLOAD_FIRST", "IDMA_FL_SEND_PAYLOAD", "IDMA_FL_REQ_NEXT_DATA_FL", "IDMA_FL_SEND_NEXT_PCIEHDR", "IDMA_FL_SEND_PADDING", "IDMA_FL_SEND_COMPLETION_TO_IMSG", }; static const u32 sge_regs[] = { A_SGE_DEBUG_DATA_LOW_INDEX_2, A_SGE_DEBUG_DATA_LOW_INDEX_3, A_SGE_DEBUG_DATA_HIGH_INDEX_10, }; const char * const *sge_idma_decode; int sge_idma_decode_nstates; int i; unsigned int chip_version = chip_id(adapter); /* Select the right set of decode strings to dump depending on the * adapter chip type. */ switch (chip_version) { case CHELSIO_T4: sge_idma_decode = (const char * const *)t4_decode; sge_idma_decode_nstates = ARRAY_SIZE(t4_decode); break; case CHELSIO_T5: sge_idma_decode = (const char * const *)t5_decode; sge_idma_decode_nstates = ARRAY_SIZE(t5_decode); break; case CHELSIO_T6: sge_idma_decode = (const char * const *)t6_decode; sge_idma_decode_nstates = ARRAY_SIZE(t6_decode); break; default: CH_ERR(adapter, "Unsupported chip version %d\n", chip_version); return; } if (state < sge_idma_decode_nstates) CH_WARN(adapter, "idma state %s\n", sge_idma_decode[state]); else CH_WARN(adapter, "idma state %d unknown\n", state); for (i = 0; i < ARRAY_SIZE(sge_regs); i++) CH_WARN(adapter, "SGE register %#x value %#x\n", sge_regs[i], t4_read_reg(adapter, sge_regs[i])); } /** * t4_sge_ctxt_flush - flush the SGE context cache * @adap: the adapter * @mbox: mailbox to use for the FW command * * Issues a FW command through the given mailbox to flush the * SGE context cache. */ int t4_sge_ctxt_flush(struct adapter *adap, unsigned int mbox) { int ret; u32 ldst_addrspace; struct fw_ldst_cmd c; memset(&c, 0, sizeof(c)); ldst_addrspace = V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_SGE_EGRC); c.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | ldst_addrspace); c.cycles_to_len16 = cpu_to_be32(FW_LEN16(c)); c.u.idctxt.msg_ctxtflush = cpu_to_be32(F_FW_LDST_CMD_CTXTFLUSH); ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); return ret; } /** * t4_fw_hello - establish communication with FW * @adap: the adapter * @mbox: mailbox to use for the FW command * @evt_mbox: mailbox to receive async FW events * @master: specifies the caller's willingness to be the device master * @state: returns the current device state (if non-NULL) * * Issues a command to establish communication with FW. Returns either * an error (negative integer) or the mailbox of the Master PF. */ int t4_fw_hello(struct adapter *adap, unsigned int mbox, unsigned int evt_mbox, enum dev_master master, enum dev_state *state) { int ret; struct fw_hello_cmd c; u32 v; unsigned int master_mbox; int retries = FW_CMD_HELLO_RETRIES; retry: memset(&c, 0, sizeof(c)); INIT_CMD(c, HELLO, WRITE); c.err_to_clearinit = cpu_to_be32( V_FW_HELLO_CMD_MASTERDIS(master == MASTER_CANT) | V_FW_HELLO_CMD_MASTERFORCE(master == MASTER_MUST) | V_FW_HELLO_CMD_MBMASTER(master == MASTER_MUST ? mbox : M_FW_HELLO_CMD_MBMASTER) | V_FW_HELLO_CMD_MBASYNCNOT(evt_mbox) | V_FW_HELLO_CMD_STAGE(FW_HELLO_CMD_STAGE_OS) | F_FW_HELLO_CMD_CLEARINIT); /* * Issue the HELLO command to the firmware. If it's not successful * but indicates that we got a "busy" or "timeout" condition, retry * the HELLO until we exhaust our retry limit. If we do exceed our * retry limit, check to see if the firmware left us any error * information and report that if so ... */ ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret != FW_SUCCESS) { if ((ret == -EBUSY || ret == -ETIMEDOUT) && retries-- > 0) goto retry; if (t4_read_reg(adap, A_PCIE_FW) & F_PCIE_FW_ERR) t4_report_fw_error(adap); return ret; } v = be32_to_cpu(c.err_to_clearinit); master_mbox = G_FW_HELLO_CMD_MBMASTER(v); if (state) { if (v & F_FW_HELLO_CMD_ERR) *state = DEV_STATE_ERR; else if (v & F_FW_HELLO_CMD_INIT) *state = DEV_STATE_INIT; else *state = DEV_STATE_UNINIT; } /* * If we're not the Master PF then we need to wait around for the * Master PF Driver to finish setting up the adapter. * * Note that we also do this wait if we're a non-Master-capable PF and * there is no current Master PF; a Master PF may show up momentarily * and we wouldn't want to fail pointlessly. (This can happen when an * OS loads lots of different drivers rapidly at the same time). In * this case, the Master PF returned by the firmware will be * M_PCIE_FW_MASTER so the test below will work ... */ if ((v & (F_FW_HELLO_CMD_ERR|F_FW_HELLO_CMD_INIT)) == 0 && master_mbox != mbox) { int waiting = FW_CMD_HELLO_TIMEOUT; /* * Wait for the firmware to either indicate an error or * initialized state. If we see either of these we bail out * and report the issue to the caller. If we exhaust the * "hello timeout" and we haven't exhausted our retries, try * again. Otherwise bail with a timeout error. */ for (;;) { u32 pcie_fw; msleep(50); waiting -= 50; /* * If neither Error nor Initialialized are indicated * by the firmware keep waiting till we exhaust our * timeout ... and then retry if we haven't exhausted * our retries ... */ pcie_fw = t4_read_reg(adap, A_PCIE_FW); if (!(pcie_fw & (F_PCIE_FW_ERR|F_PCIE_FW_INIT))) { if (waiting <= 0) { if (retries-- > 0) goto retry; return -ETIMEDOUT; } continue; } /* * We either have an Error or Initialized condition * report errors preferentially. */ if (state) { if (pcie_fw & F_PCIE_FW_ERR) *state = DEV_STATE_ERR; else if (pcie_fw & F_PCIE_FW_INIT) *state = DEV_STATE_INIT; } /* * If we arrived before a Master PF was selected and * there's not a valid Master PF, grab its identity * for our caller. */ if (master_mbox == M_PCIE_FW_MASTER && (pcie_fw & F_PCIE_FW_MASTER_VLD)) master_mbox = G_PCIE_FW_MASTER(pcie_fw); break; } } return master_mbox; } /** * t4_fw_bye - end communication with FW * @adap: the adapter * @mbox: mailbox to use for the FW command * * Issues a command to terminate communication with FW. */ int t4_fw_bye(struct adapter *adap, unsigned int mbox) { struct fw_bye_cmd c; memset(&c, 0, sizeof(c)); INIT_CMD(c, BYE, WRITE); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_fw_reset - issue a reset to FW * @adap: the adapter * @mbox: mailbox to use for the FW command * @reset: specifies the type of reset to perform * * Issues a reset command of the specified type to FW. */ int t4_fw_reset(struct adapter *adap, unsigned int mbox, int reset) { struct fw_reset_cmd c; memset(&c, 0, sizeof(c)); INIT_CMD(c, RESET, WRITE); c.val = cpu_to_be32(reset); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_fw_halt - issue a reset/halt to FW and put uP into RESET * @adap: the adapter * @mbox: mailbox to use for the FW RESET command (if desired) * @force: force uP into RESET even if FW RESET command fails * * Issues a RESET command to firmware (if desired) with a HALT indication * and then puts the microprocessor into RESET state. The RESET command * will only be issued if a legitimate mailbox is provided (mbox <= * M_PCIE_FW_MASTER). * * This is generally used in order for the host to safely manipulate the * adapter without fear of conflicting with whatever the firmware might * be doing. The only way out of this state is to RESTART the firmware * ... */ int t4_fw_halt(struct adapter *adap, unsigned int mbox, int force) { int ret = 0; /* * If a legitimate mailbox is provided, issue a RESET command * with a HALT indication. */ if (mbox <= M_PCIE_FW_MASTER) { struct fw_reset_cmd c; memset(&c, 0, sizeof(c)); INIT_CMD(c, RESET, WRITE); c.val = cpu_to_be32(F_PIORST | F_PIORSTMODE); c.halt_pkd = cpu_to_be32(F_FW_RESET_CMD_HALT); ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /* * Normally we won't complete the operation if the firmware RESET * command fails but if our caller insists we'll go ahead and put the * uP into RESET. This can be useful if the firmware is hung or even * missing ... We'll have to take the risk of putting the uP into * RESET without the cooperation of firmware in that case. * * We also force the firmware's HALT flag to be on in case we bypassed * the firmware RESET command above or we're dealing with old firmware * which doesn't have the HALT capability. This will serve as a flag * for the incoming firmware to know that it's coming out of a HALT * rather than a RESET ... if it's new enough to understand that ... */ if (ret == 0 || force) { t4_set_reg_field(adap, A_CIM_BOOT_CFG, F_UPCRST, F_UPCRST); t4_set_reg_field(adap, A_PCIE_FW, F_PCIE_FW_HALT, F_PCIE_FW_HALT); } /* * And we always return the result of the firmware RESET command * even when we force the uP into RESET ... */ return ret; } /** * t4_fw_restart - restart the firmware by taking the uP out of RESET * @adap: the adapter * @reset: if we want to do a RESET to restart things * * Restart firmware previously halted by t4_fw_halt(). On successful * return the previous PF Master remains as the new PF Master and there * is no need to issue a new HELLO command, etc. * * We do this in two ways: * * 1. If we're dealing with newer firmware we'll simply want to take * the chip's microprocessor out of RESET. This will cause the * firmware to start up from its start vector. And then we'll loop * until the firmware indicates it's started again (PCIE_FW.HALT * reset to 0) or we timeout. * * 2. If we're dealing with older firmware then we'll need to RESET * the chip since older firmware won't recognize the PCIE_FW.HALT * flag and automatically RESET itself on startup. */ int t4_fw_restart(struct adapter *adap, unsigned int mbox, int reset) { if (reset) { /* * Since we're directing the RESET instead of the firmware * doing it automatically, we need to clear the PCIE_FW.HALT * bit. */ t4_set_reg_field(adap, A_PCIE_FW, F_PCIE_FW_HALT, 0); /* * If we've been given a valid mailbox, first try to get the * firmware to do the RESET. If that works, great and we can * return success. Otherwise, if we haven't been given a * valid mailbox or the RESET command failed, fall back to * hitting the chip with a hammer. */ if (mbox <= M_PCIE_FW_MASTER) { t4_set_reg_field(adap, A_CIM_BOOT_CFG, F_UPCRST, 0); msleep(100); if (t4_fw_reset(adap, mbox, F_PIORST | F_PIORSTMODE) == 0) return 0; } t4_write_reg(adap, A_PL_RST, F_PIORST | F_PIORSTMODE); msleep(2000); } else { int ms; t4_set_reg_field(adap, A_CIM_BOOT_CFG, F_UPCRST, 0); for (ms = 0; ms < FW_CMD_MAX_TIMEOUT; ) { if (!(t4_read_reg(adap, A_PCIE_FW) & F_PCIE_FW_HALT)) return FW_SUCCESS; msleep(100); ms += 100; } return -ETIMEDOUT; } return 0; } /** * t4_fw_upgrade - perform all of the steps necessary to upgrade FW * @adap: the adapter * @mbox: mailbox to use for the FW RESET command (if desired) * @fw_data: the firmware image to write * @size: image size * @force: force upgrade even if firmware doesn't cooperate * * Perform all of the steps necessary for upgrading an adapter's * firmware image. Normally this requires the cooperation of the * existing firmware in order to halt all existing activities * but if an invalid mailbox token is passed in we skip that step * (though we'll still put the adapter microprocessor into RESET in * that case). * * On successful return the new firmware will have been loaded and * the adapter will have been fully RESET losing all previous setup * state. On unsuccessful return the adapter may be completely hosed ... * positive errno indicates that the adapter is ~probably~ intact, a * negative errno indicates that things are looking bad ... */ int t4_fw_upgrade(struct adapter *adap, unsigned int mbox, const u8 *fw_data, unsigned int size, int force) { const struct fw_hdr *fw_hdr = (const struct fw_hdr *)fw_data; unsigned int bootstrap = be32_to_cpu(fw_hdr->magic) == FW_HDR_MAGIC_BOOTSTRAP; int reset, ret; if (!t4_fw_matches_chip(adap, fw_hdr)) return -EINVAL; if (!bootstrap) { ret = t4_fw_halt(adap, mbox, force); if (ret < 0 && !force) return ret; } ret = t4_load_fw(adap, fw_data, size); if (ret < 0 || bootstrap) return ret; /* * Older versions of the firmware don't understand the new * PCIE_FW.HALT flag and so won't know to perform a RESET when they * restart. So for newly loaded older firmware we'll have to do the * RESET for it so it starts up on a clean slate. We can tell if * the newly loaded firmware will handle this right by checking * its header flags to see if it advertises the capability. */ reset = ((be32_to_cpu(fw_hdr->flags) & FW_HDR_FLAGS_RESET_HALT) == 0); return t4_fw_restart(adap, mbox, reset); } /** * t4_fw_initialize - ask FW to initialize the device * @adap: the adapter * @mbox: mailbox to use for the FW command * * Issues a command to FW to partially initialize the device. This * performs initialization that generally doesn't depend on user input. */ int t4_fw_initialize(struct adapter *adap, unsigned int mbox) { struct fw_initialize_cmd c; memset(&c, 0, sizeof(c)); INIT_CMD(c, INITIALIZE, WRITE); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_query_params_rw - query FW or device parameters * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF * @vf: the VF * @nparams: the number of parameters * @params: the parameter names * @val: the parameter values * @rw: Write and read flag * * Reads the value of FW or device parameters. Up to 7 parameters can be * queried at once. */ int t4_query_params_rw(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int nparams, const u32 *params, u32 *val, int rw) { int i, ret; struct fw_params_cmd c; __be32 *p = &c.param[0].mnem; if (nparams > 7) return -EINVAL; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_PARAMS_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | V_FW_PARAMS_CMD_PFN(pf) | V_FW_PARAMS_CMD_VFN(vf)); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); for (i = 0; i < nparams; i++) { *p++ = cpu_to_be32(*params++); if (rw) *p = cpu_to_be32(*(val + i)); p++; } ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret == 0) for (i = 0, p = &c.param[0].val; i < nparams; i++, p += 2) *val++ = be32_to_cpu(*p); return ret; } int t4_query_params(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int nparams, const u32 *params, u32 *val) { return t4_query_params_rw(adap, mbox, pf, vf, nparams, params, val, 0); } /** * t4_set_params_timeout - sets FW or device parameters * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF * @vf: the VF * @nparams: the number of parameters * @params: the parameter names * @val: the parameter values * @timeout: the timeout time * * Sets the value of FW or device parameters. Up to 7 parameters can be * specified at once. */ int t4_set_params_timeout(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int nparams, const u32 *params, const u32 *val, int timeout) { struct fw_params_cmd c; __be32 *p = &c.param[0].mnem; if (nparams > 7) return -EINVAL; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_PARAMS_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_PARAMS_CMD_PFN(pf) | V_FW_PARAMS_CMD_VFN(vf)); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); while (nparams--) { *p++ = cpu_to_be32(*params++); *p++ = cpu_to_be32(*val++); } return t4_wr_mbox_timeout(adap, mbox, &c, sizeof(c), NULL, timeout); } /** * t4_set_params - sets FW or device parameters * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF * @vf: the VF * @nparams: the number of parameters * @params: the parameter names * @val: the parameter values * * Sets the value of FW or device parameters. Up to 7 parameters can be * specified at once. */ int t4_set_params(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int nparams, const u32 *params, const u32 *val) { return t4_set_params_timeout(adap, mbox, pf, vf, nparams, params, val, FW_CMD_MAX_TIMEOUT); } /** * t4_cfg_pfvf - configure PF/VF resource limits * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF being configured * @vf: the VF being configured * @txq: the max number of egress queues * @txq_eth_ctrl: the max number of egress Ethernet or control queues * @rxqi: the max number of interrupt-capable ingress queues * @rxq: the max number of interruptless ingress queues * @tc: the PCI traffic class * @vi: the max number of virtual interfaces * @cmask: the channel access rights mask for the PF/VF * @pmask: the port access rights mask for the PF/VF * @nexact: the maximum number of exact MPS filters * @rcaps: read capabilities * @wxcaps: write/execute capabilities * * Configures resource limits and capabilities for a physical or virtual * function. */ int t4_cfg_pfvf(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int txq, unsigned int txq_eth_ctrl, unsigned int rxqi, unsigned int rxq, unsigned int tc, unsigned int vi, unsigned int cmask, unsigned int pmask, unsigned int nexact, unsigned int rcaps, unsigned int wxcaps) { struct fw_pfvf_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_PFVF_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_PFVF_CMD_PFN(pf) | V_FW_PFVF_CMD_VFN(vf)); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); c.niqflint_niq = cpu_to_be32(V_FW_PFVF_CMD_NIQFLINT(rxqi) | V_FW_PFVF_CMD_NIQ(rxq)); c.type_to_neq = cpu_to_be32(V_FW_PFVF_CMD_CMASK(cmask) | V_FW_PFVF_CMD_PMASK(pmask) | V_FW_PFVF_CMD_NEQ(txq)); c.tc_to_nexactf = cpu_to_be32(V_FW_PFVF_CMD_TC(tc) | V_FW_PFVF_CMD_NVI(vi) | V_FW_PFVF_CMD_NEXACTF(nexact)); c.r_caps_to_nethctrl = cpu_to_be32(V_FW_PFVF_CMD_R_CAPS(rcaps) | V_FW_PFVF_CMD_WX_CAPS(wxcaps) | V_FW_PFVF_CMD_NETHCTRL(txq_eth_ctrl)); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_alloc_vi_func - allocate a virtual interface * @adap: the adapter * @mbox: mailbox to use for the FW command * @port: physical port associated with the VI * @pf: the PF owning the VI * @vf: the VF owning the VI * @nmac: number of MAC addresses needed (1 to 5) * @mac: the MAC addresses of the VI * @rss_size: size of RSS table slice associated with this VI * @portfunc: which Port Application Function MAC Address is desired * @idstype: Intrusion Detection Type * * Allocates a virtual interface for the given physical port. If @mac is * not %NULL it contains the MAC addresses of the VI as assigned by FW. * If @rss_size is %NULL the VI is not assigned any RSS slice by FW. * @mac should be large enough to hold @nmac Ethernet addresses, they are * stored consecutively so the space needed is @nmac * 6 bytes. * Returns a negative error number or the non-negative VI id. */ int t4_alloc_vi_func(struct adapter *adap, unsigned int mbox, unsigned int port, unsigned int pf, unsigned int vf, unsigned int nmac, u8 *mac, u16 *rss_size, unsigned int portfunc, unsigned int idstype) { int ret; struct fw_vi_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_VI_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_VI_CMD_PFN(pf) | V_FW_VI_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_VI_CMD_ALLOC | FW_LEN16(c)); c.type_to_viid = cpu_to_be16(V_FW_VI_CMD_TYPE(idstype) | V_FW_VI_CMD_FUNC(portfunc)); c.portid_pkd = V_FW_VI_CMD_PORTID(port); c.nmac = nmac - 1; if(!rss_size) c.norss_rsssize = F_FW_VI_CMD_NORSS; ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret) return ret; if (mac) { memcpy(mac, c.mac, sizeof(c.mac)); switch (nmac) { case 5: memcpy(mac + 24, c.nmac3, sizeof(c.nmac3)); case 4: memcpy(mac + 18, c.nmac2, sizeof(c.nmac2)); case 3: memcpy(mac + 12, c.nmac1, sizeof(c.nmac1)); case 2: memcpy(mac + 6, c.nmac0, sizeof(c.nmac0)); } } if (rss_size) *rss_size = G_FW_VI_CMD_RSSSIZE(be16_to_cpu(c.norss_rsssize)); return G_FW_VI_CMD_VIID(be16_to_cpu(c.type_to_viid)); } /** * t4_alloc_vi - allocate an [Ethernet Function] virtual interface * @adap: the adapter * @mbox: mailbox to use for the FW command * @port: physical port associated with the VI * @pf: the PF owning the VI * @vf: the VF owning the VI * @nmac: number of MAC addresses needed (1 to 5) * @mac: the MAC addresses of the VI * @rss_size: size of RSS table slice associated with this VI * * backwards compatible and convieniance routine to allocate a Virtual * Interface with a Ethernet Port Application Function and Intrustion * Detection System disabled. */ int t4_alloc_vi(struct adapter *adap, unsigned int mbox, unsigned int port, unsigned int pf, unsigned int vf, unsigned int nmac, u8 *mac, u16 *rss_size) { return t4_alloc_vi_func(adap, mbox, port, pf, vf, nmac, mac, rss_size, FW_VI_FUNC_ETH, 0); } /** * t4_free_vi - free a virtual interface * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the VI * @vf: the VF owning the VI * @viid: virtual interface identifiler * * Free a previously allocated virtual interface. */ int t4_free_vi(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int viid) { struct fw_vi_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_VI_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_VI_CMD_PFN(pf) | V_FW_VI_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_VI_CMD_FREE | FW_LEN16(c)); c.type_to_viid = cpu_to_be16(V_FW_VI_CMD_VIID(viid)); return t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); } /** * t4_set_rxmode - set Rx properties of a virtual interface * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @mtu: the new MTU or -1 * @promisc: 1 to enable promiscuous mode, 0 to disable it, -1 no change * @all_multi: 1 to enable all-multi mode, 0 to disable it, -1 no change * @bcast: 1 to enable broadcast Rx, 0 to disable it, -1 no change * @vlanex: 1 to enable HW VLAN extraction, 0 to disable it, -1 no change * @sleep_ok: if true we may sleep while awaiting command completion * * Sets Rx properties of a virtual interface. */ int t4_set_rxmode(struct adapter *adap, unsigned int mbox, unsigned int viid, int mtu, int promisc, int all_multi, int bcast, int vlanex, bool sleep_ok) { struct fw_vi_rxmode_cmd c; /* convert to FW values */ if (mtu < 0) mtu = M_FW_VI_RXMODE_CMD_MTU; if (promisc < 0) promisc = M_FW_VI_RXMODE_CMD_PROMISCEN; if (all_multi < 0) all_multi = M_FW_VI_RXMODE_CMD_ALLMULTIEN; if (bcast < 0) bcast = M_FW_VI_RXMODE_CMD_BROADCASTEN; if (vlanex < 0) vlanex = M_FW_VI_RXMODE_CMD_VLANEXEN; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_VI_RXMODE_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_VI_RXMODE_CMD_VIID(viid)); c.retval_len16 = cpu_to_be32(FW_LEN16(c)); c.mtu_to_vlanexen = cpu_to_be32(V_FW_VI_RXMODE_CMD_MTU(mtu) | V_FW_VI_RXMODE_CMD_PROMISCEN(promisc) | V_FW_VI_RXMODE_CMD_ALLMULTIEN(all_multi) | V_FW_VI_RXMODE_CMD_BROADCASTEN(bcast) | V_FW_VI_RXMODE_CMD_VLANEXEN(vlanex)); return t4_wr_mbox_meat(adap, mbox, &c, sizeof(c), NULL, sleep_ok); } /** * t4_alloc_mac_filt - allocates exact-match filters for MAC addresses * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @free: if true any existing filters for this VI id are first removed * @naddr: the number of MAC addresses to allocate filters for (up to 7) * @addr: the MAC address(es) * @idx: where to store the index of each allocated filter * @hash: pointer to hash address filter bitmap * @sleep_ok: call is allowed to sleep * * Allocates an exact-match filter for each of the supplied addresses and * sets it to the corresponding address. If @idx is not %NULL it should * have at least @naddr entries, each of which will be set to the index of * the filter allocated for the corresponding MAC address. If a filter * could not be allocated for an address its index is set to 0xffff. * If @hash is not %NULL addresses that fail to allocate an exact filter * are hashed and update the hash filter bitmap pointed at by @hash. * * Returns a negative error number or the number of filters allocated. */ int t4_alloc_mac_filt(struct adapter *adap, unsigned int mbox, unsigned int viid, bool free, unsigned int naddr, const u8 **addr, u16 *idx, u64 *hash, bool sleep_ok) { int offset, ret = 0; struct fw_vi_mac_cmd c; unsigned int nfilters = 0; unsigned int max_naddr = adap->chip_params->mps_tcam_size; unsigned int rem = naddr; if (naddr > max_naddr) return -EINVAL; for (offset = 0; offset < naddr ; /**/) { unsigned int fw_naddr = (rem < ARRAY_SIZE(c.u.exact) ? rem : ARRAY_SIZE(c.u.exact)); size_t len16 = DIV_ROUND_UP(offsetof(struct fw_vi_mac_cmd, u.exact[fw_naddr]), 16); struct fw_vi_mac_exact *p; int i; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_VI_MAC_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_CMD_EXEC(free) | V_FW_VI_MAC_CMD_VIID(viid)); c.freemacs_to_len16 = cpu_to_be32(V_FW_VI_MAC_CMD_FREEMACS(free) | V_FW_CMD_LEN16(len16)); for (i = 0, p = c.u.exact; i < fw_naddr; i++, p++) { p->valid_to_idx = cpu_to_be16(F_FW_VI_MAC_CMD_VALID | V_FW_VI_MAC_CMD_IDX(FW_VI_MAC_ADD_MAC)); memcpy(p->macaddr, addr[offset+i], sizeof(p->macaddr)); } /* * It's okay if we run out of space in our MAC address arena. * Some of the addresses we submit may get stored so we need * to run through the reply to see what the results were ... */ ret = t4_wr_mbox_meat(adap, mbox, &c, sizeof(c), &c, sleep_ok); if (ret && ret != -FW_ENOMEM) break; for (i = 0, p = c.u.exact; i < fw_naddr; i++, p++) { u16 index = G_FW_VI_MAC_CMD_IDX( be16_to_cpu(p->valid_to_idx)); if (idx) idx[offset+i] = (index >= max_naddr ? 0xffff : index); if (index < max_naddr) nfilters++; else if (hash) *hash |= (1ULL << hash_mac_addr(addr[offset+i])); } free = false; offset += fw_naddr; rem -= fw_naddr; } if (ret == 0 || ret == -FW_ENOMEM) ret = nfilters; return ret; } /** * t4_change_mac - modifies the exact-match filter for a MAC address * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @idx: index of existing filter for old value of MAC address, or -1 * @addr: the new MAC address value * @persist: whether a new MAC allocation should be persistent * @add_smt: if true also add the address to the HW SMT * * Modifies an exact-match filter and sets it to the new MAC address if * @idx >= 0, or adds the MAC address to a new filter if @idx < 0. In the * latter case the address is added persistently if @persist is %true. * * Note that in general it is not possible to modify the value of a given * filter so the generic way to modify an address filter is to free the one * being used by the old address value and allocate a new filter for the * new address value. * * Returns a negative error number or the index of the filter with the new * MAC value. Note that this index may differ from @idx. */ int t4_change_mac(struct adapter *adap, unsigned int mbox, unsigned int viid, int idx, const u8 *addr, bool persist, bool add_smt) { int ret, mode; struct fw_vi_mac_cmd c; struct fw_vi_mac_exact *p = c.u.exact; unsigned int max_mac_addr = adap->chip_params->mps_tcam_size; if (idx < 0) /* new allocation */ idx = persist ? FW_VI_MAC_ADD_PERSIST_MAC : FW_VI_MAC_ADD_MAC; mode = add_smt ? FW_VI_MAC_SMT_AND_MPSTCAM : FW_VI_MAC_MPS_TCAM_ENTRY; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_VI_MAC_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_VI_MAC_CMD_VIID(viid)); c.freemacs_to_len16 = cpu_to_be32(V_FW_CMD_LEN16(1)); p->valid_to_idx = cpu_to_be16(F_FW_VI_MAC_CMD_VALID | V_FW_VI_MAC_CMD_SMAC_RESULT(mode) | V_FW_VI_MAC_CMD_IDX(idx)); memcpy(p->macaddr, addr, sizeof(p->macaddr)); ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret == 0) { ret = G_FW_VI_MAC_CMD_IDX(be16_to_cpu(p->valid_to_idx)); if (ret >= max_mac_addr) ret = -ENOMEM; } return ret; } /** * t4_set_addr_hash - program the MAC inexact-match hash filter * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @ucast: whether the hash filter should also match unicast addresses * @vec: the value to be written to the hash filter * @sleep_ok: call is allowed to sleep * * Sets the 64-bit inexact-match hash filter for a virtual interface. */ int t4_set_addr_hash(struct adapter *adap, unsigned int mbox, unsigned int viid, bool ucast, u64 vec, bool sleep_ok) { struct fw_vi_mac_cmd c; u32 val; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_VI_MAC_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_VI_ENABLE_CMD_VIID(viid)); val = V_FW_VI_MAC_CMD_ENTRY_TYPE(FW_VI_MAC_TYPE_HASHVEC) | V_FW_VI_MAC_CMD_HASHUNIEN(ucast) | V_FW_CMD_LEN16(1); c.freemacs_to_len16 = cpu_to_be32(val); c.u.hash.hashvec = cpu_to_be64(vec); return t4_wr_mbox_meat(adap, mbox, &c, sizeof(c), NULL, sleep_ok); } /** * t4_enable_vi_params - enable/disable a virtual interface * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @rx_en: 1=enable Rx, 0=disable Rx * @tx_en: 1=enable Tx, 0=disable Tx * @dcb_en: 1=enable delivery of Data Center Bridging messages. * * Enables/disables a virtual interface. Note that setting DCB Enable * only makes sense when enabling a Virtual Interface ... */ int t4_enable_vi_params(struct adapter *adap, unsigned int mbox, unsigned int viid, bool rx_en, bool tx_en, bool dcb_en) { struct fw_vi_enable_cmd c; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_VI_ENABLE_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_VI_ENABLE_CMD_VIID(viid)); c.ien_to_len16 = cpu_to_be32(V_FW_VI_ENABLE_CMD_IEN(rx_en) | V_FW_VI_ENABLE_CMD_EEN(tx_en) | V_FW_VI_ENABLE_CMD_DCB_INFO(dcb_en) | FW_LEN16(c)); return t4_wr_mbox_ns(adap, mbox, &c, sizeof(c), NULL); } /** * t4_enable_vi - enable/disable a virtual interface * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @rx_en: 1=enable Rx, 0=disable Rx * @tx_en: 1=enable Tx, 0=disable Tx * * Enables/disables a virtual interface. Note that setting DCB Enable * only makes sense when enabling a Virtual Interface ... */ int t4_enable_vi(struct adapter *adap, unsigned int mbox, unsigned int viid, bool rx_en, bool tx_en) { return t4_enable_vi_params(adap, mbox, viid, rx_en, tx_en, 0); } /** * t4_identify_port - identify a VI's port by blinking its LED * @adap: the adapter * @mbox: mailbox to use for the FW command * @viid: the VI id * @nblinks: how many times to blink LED at 2.5 Hz * * Identifies a VI's port by blinking its LED. */ int t4_identify_port(struct adapter *adap, unsigned int mbox, unsigned int viid, unsigned int nblinks) { struct fw_vi_enable_cmd c; memset(&c, 0, sizeof(c)); c.op_to_viid = cpu_to_be32(V_FW_CMD_OP(FW_VI_ENABLE_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_VI_ENABLE_CMD_VIID(viid)); c.ien_to_len16 = cpu_to_be32(F_FW_VI_ENABLE_CMD_LED | FW_LEN16(c)); c.blinkdur = cpu_to_be16(nblinks); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_iq_stop - stop an ingress queue and its FLs * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the queues * @vf: the VF owning the queues * @iqtype: the ingress queue type (FW_IQ_TYPE_FL_INT_CAP, etc.) * @iqid: ingress queue id * @fl0id: FL0 queue id or 0xffff if no attached FL0 * @fl1id: FL1 queue id or 0xffff if no attached FL1 * * Stops an ingress queue and its associated FLs, if any. This causes * any current or future data/messages destined for these queues to be * tossed. */ int t4_iq_stop(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int iqtype, unsigned int iqid, unsigned int fl0id, unsigned int fl1id) { struct fw_iq_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(pf) | V_FW_IQ_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_IQSTOP | FW_LEN16(c)); c.type_to_iqandstindex = cpu_to_be32(V_FW_IQ_CMD_TYPE(iqtype)); c.iqid = cpu_to_be16(iqid); c.fl0id = cpu_to_be16(fl0id); c.fl1id = cpu_to_be16(fl1id); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_iq_free - free an ingress queue and its FLs * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the queues * @vf: the VF owning the queues * @iqtype: the ingress queue type (FW_IQ_TYPE_FL_INT_CAP, etc.) * @iqid: ingress queue id * @fl0id: FL0 queue id or 0xffff if no attached FL0 * @fl1id: FL1 queue id or 0xffff if no attached FL1 * * Frees an ingress queue and its associated FLs, if any. */ int t4_iq_free(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int iqtype, unsigned int iqid, unsigned int fl0id, unsigned int fl1id) { struct fw_iq_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(pf) | V_FW_IQ_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_IQ_CMD_FREE | FW_LEN16(c)); c.type_to_iqandstindex = cpu_to_be32(V_FW_IQ_CMD_TYPE(iqtype)); c.iqid = cpu_to_be16(iqid); c.fl0id = cpu_to_be16(fl0id); c.fl1id = cpu_to_be16(fl1id); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_eth_eq_free - free an Ethernet egress queue * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the queue * @vf: the VF owning the queue * @eqid: egress queue id * * Frees an Ethernet egress queue. */ int t4_eth_eq_free(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int eqid) { struct fw_eq_eth_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(pf) | V_FW_EQ_ETH_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_EQ_ETH_CMD_FREE | FW_LEN16(c)); c.eqid_pkd = cpu_to_be32(V_FW_EQ_ETH_CMD_EQID(eqid)); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_ctrl_eq_free - free a control egress queue * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the queue * @vf: the VF owning the queue * @eqid: egress queue id * * Frees a control egress queue. */ int t4_ctrl_eq_free(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int eqid) { struct fw_eq_ctrl_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(pf) | V_FW_EQ_CTRL_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_EQ_CTRL_CMD_FREE | FW_LEN16(c)); c.cmpliqid_eqid = cpu_to_be32(V_FW_EQ_CTRL_CMD_EQID(eqid)); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_ofld_eq_free - free an offload egress queue * @adap: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the queue * @vf: the VF owning the queue * @eqid: egress queue id * * Frees a control egress queue. */ int t4_ofld_eq_free(struct adapter *adap, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int eqid) { struct fw_eq_ofld_cmd c; memset(&c, 0, sizeof(c)); c.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(pf) | V_FW_EQ_OFLD_CMD_VFN(vf)); c.alloc_to_len16 = cpu_to_be32(F_FW_EQ_OFLD_CMD_FREE | FW_LEN16(c)); c.eqid_pkd = cpu_to_be32(V_FW_EQ_OFLD_CMD_EQID(eqid)); return t4_wr_mbox(adap, mbox, &c, sizeof(c), NULL); } /** * t4_link_down_rc_str - return a string for a Link Down Reason Code * @link_down_rc: Link Down Reason Code * * Returns a string representation of the Link Down Reason Code. */ const char *t4_link_down_rc_str(unsigned char link_down_rc) { static const char *reason[] = { "Link Down", "Remote Fault", "Auto-negotiation Failure", "Reserved3", "Insufficient Airflow", "Unable To Determine Reason", "No RX Signal Detected", "Reserved7", }; if (link_down_rc >= ARRAY_SIZE(reason)) return "Bad Reason Code"; return reason[link_down_rc]; } /** * t4_handle_fw_rpl - process a FW reply message * @adap: the adapter * @rpl: start of the FW message * * Processes a FW message, such as link state change messages. */ int t4_handle_fw_rpl(struct adapter *adap, const __be64 *rpl) { u8 opcode = *(const u8 *)rpl; const struct fw_port_cmd *p = (const void *)rpl; unsigned int action = G_FW_PORT_CMD_ACTION(be32_to_cpu(p->action_to_len16)); if (opcode == FW_PORT_CMD && action == FW_PORT_ACTION_GET_PORT_INFO) { /* link/module state change message */ int speed = 0, fc = 0, i; int chan = G_FW_PORT_CMD_PORTID(be32_to_cpu(p->op_to_portid)); struct port_info *pi = NULL; struct link_config *lc; u32 stat = be32_to_cpu(p->u.info.lstatus_to_modtype); int link_ok = (stat & F_FW_PORT_CMD_LSTATUS) != 0; u32 mod = G_FW_PORT_CMD_MODTYPE(stat); if (stat & F_FW_PORT_CMD_RXPAUSE) fc |= PAUSE_RX; if (stat & F_FW_PORT_CMD_TXPAUSE) fc |= PAUSE_TX; if (stat & V_FW_PORT_CMD_LSPEED(FW_PORT_CAP_SPEED_100M)) speed = 100; else if (stat & V_FW_PORT_CMD_LSPEED(FW_PORT_CAP_SPEED_1G)) speed = 1000; else if (stat & V_FW_PORT_CMD_LSPEED(FW_PORT_CAP_SPEED_10G)) speed = 10000; else if (stat & V_FW_PORT_CMD_LSPEED(FW_PORT_CAP_SPEED_40G)) speed = 40000; for_each_port(adap, i) { pi = adap2pinfo(adap, i); if (pi->tx_chan == chan) break; } lc = &pi->link_cfg; if (mod != pi->mod_type) { pi->mod_type = mod; t4_os_portmod_changed(adap, i); } if (link_ok != lc->link_ok || speed != lc->speed || fc != lc->fc) { /* something changed */ int reason; if (!link_ok && lc->link_ok) reason = G_FW_PORT_CMD_LINKDNRC(stat); else reason = -1; lc->link_ok = link_ok; lc->speed = speed; lc->fc = fc; lc->supported = be16_to_cpu(p->u.info.pcap); t4_os_link_changed(adap, i, link_ok, reason); } } else { CH_WARN_RATELIMIT(adap, "Unknown firmware reply %d\n", opcode); return -EINVAL; } return 0; } /** * get_pci_mode - determine a card's PCI mode * @adapter: the adapter * @p: where to store the PCI settings * * Determines a card's PCI mode and associated parameters, such as speed * and width. */ static void get_pci_mode(struct adapter *adapter, struct pci_params *p) { u16 val; u32 pcie_cap; pcie_cap = t4_os_find_pci_capability(adapter, PCI_CAP_ID_EXP); if (pcie_cap) { t4_os_pci_read_cfg2(adapter, pcie_cap + PCI_EXP_LNKSTA, &val); p->speed = val & PCI_EXP_LNKSTA_CLS; p->width = (val & PCI_EXP_LNKSTA_NLW) >> 4; } } /** * init_link_config - initialize a link's SW state * @lc: structure holding the link state * @caps: link capabilities * * Initializes the SW state maintained for each link, including the link's * capabilities and default speed/flow-control/autonegotiation settings. */ static void init_link_config(struct link_config *lc, unsigned int caps) { lc->supported = caps; lc->requested_speed = 0; lc->speed = 0; lc->requested_fc = lc->fc = PAUSE_RX | PAUSE_TX; if (lc->supported & FW_PORT_CAP_ANEG) { lc->advertising = lc->supported & ADVERT_MASK; lc->autoneg = AUTONEG_ENABLE; lc->requested_fc |= PAUSE_AUTONEG; } else { lc->advertising = 0; lc->autoneg = AUTONEG_DISABLE; } } struct flash_desc { u32 vendor_and_model_id; u32 size_mb; }; int t4_get_flash_params(struct adapter *adapter) { /* * Table for non-Numonix supported flash parts. Numonix parts are left * to the preexisting well-tested code. All flash parts have 64KB * sectors. */ static struct flash_desc supported_flash[] = { { 0x150201, 4 << 20 }, /* Spansion 4MB S25FL032P */ }; int ret; u32 info = 0; ret = sf1_write(adapter, 1, 1, 0, SF_RD_ID); if (!ret) ret = sf1_read(adapter, 3, 0, 1, &info); t4_write_reg(adapter, A_SF_OP, 0); /* unlock SF */ if (ret < 0) return ret; for (ret = 0; ret < ARRAY_SIZE(supported_flash); ++ret) if (supported_flash[ret].vendor_and_model_id == info) { adapter->params.sf_size = supported_flash[ret].size_mb; adapter->params.sf_nsec = adapter->params.sf_size / SF_SEC_SIZE; return 0; } if ((info & 0xff) != 0x20) /* not a Numonix flash */ return -EINVAL; info >>= 16; /* log2 of size */ if (info >= 0x14 && info < 0x18) adapter->params.sf_nsec = 1 << (info - 16); else if (info == 0x18) adapter->params.sf_nsec = 64; else return -EINVAL; adapter->params.sf_size = 1 << info; /* * We should ~probably~ reject adapters with FLASHes which are too * small but we have some legacy FPGAs with small FLASHes that we'd * still like to use. So instead we emit a scary message ... */ if (adapter->params.sf_size < FLASH_MIN_SIZE) CH_WARN(adapter, "WARNING!!! FLASH size %#x < %#x!!!\n", adapter->params.sf_size, FLASH_MIN_SIZE); return 0; } static void set_pcie_completion_timeout(struct adapter *adapter, u8 range) { u16 val; u32 pcie_cap; pcie_cap = t4_os_find_pci_capability(adapter, PCI_CAP_ID_EXP); if (pcie_cap) { t4_os_pci_read_cfg2(adapter, pcie_cap + PCI_EXP_DEVCTL2, &val); val &= 0xfff0; val |= range ; t4_os_pci_write_cfg2(adapter, pcie_cap + PCI_EXP_DEVCTL2, val); } } const struct chip_params *t4_get_chip_params(int chipid) { static const struct chip_params chip_params[] = { { /* T4 */ .nchan = NCHAN, .pm_stats_cnt = PM_NSTATS, .cng_ch_bits_log = 2, .nsched_cls = 15, .cim_num_obq = CIM_NUM_OBQ, .mps_rplc_size = 128, .vfcount = 128, .sge_fl_db = F_DBPRIO, .mps_tcam_size = NUM_MPS_CLS_SRAM_L_INSTANCES, }, { /* T5 */ .nchan = NCHAN, .pm_stats_cnt = PM_NSTATS, .cng_ch_bits_log = 2, .nsched_cls = 16, .cim_num_obq = CIM_NUM_OBQ_T5, .mps_rplc_size = 128, .vfcount = 128, .sge_fl_db = F_DBPRIO | F_DBTYPE, .mps_tcam_size = NUM_MPS_T5_CLS_SRAM_L_INSTANCES, }, { /* T6 */ .nchan = T6_NCHAN, .pm_stats_cnt = T6_PM_NSTATS, .cng_ch_bits_log = 3, .nsched_cls = 16, .cim_num_obq = CIM_NUM_OBQ_T5, .mps_rplc_size = 256, .vfcount = 256, .sge_fl_db = 0, .mps_tcam_size = NUM_MPS_T5_CLS_SRAM_L_INSTANCES, }, }; chipid -= CHELSIO_T4; if (chipid < 0 || chipid >= ARRAY_SIZE(chip_params)) return NULL; return &chip_params[chipid]; } /** * t4_prep_adapter - prepare SW and HW for operation * @adapter: the adapter * @buf: temporary space of at least VPD_LEN size provided by the caller. * * Initialize adapter SW state for the various HW modules, set initial * values for some adapter tunables, take PHYs out of reset, and * initialize the MDIO interface. */ int t4_prep_adapter(struct adapter *adapter, u8 *buf) { int ret; uint16_t device_id; uint32_t pl_rev; get_pci_mode(adapter, &adapter->params.pci); pl_rev = t4_read_reg(adapter, A_PL_REV); adapter->params.chipid = G_CHIPID(pl_rev); adapter->params.rev = G_REV(pl_rev); if (adapter->params.chipid == 0) { /* T4 did not have chipid in PL_REV (T5 onwards do) */ adapter->params.chipid = CHELSIO_T4; /* T4A1 chip is not supported */ if (adapter->params.rev == 1) { CH_ALERT(adapter, "T4 rev 1 chip is not supported.\n"); return -EINVAL; } } adapter->chip_params = t4_get_chip_params(chip_id(adapter)); if (adapter->chip_params == NULL) return -EINVAL; adapter->params.pci.vpd_cap_addr = t4_os_find_pci_capability(adapter, PCI_CAP_ID_VPD); ret = t4_get_flash_params(adapter); if (ret < 0) return ret; ret = get_vpd_params(adapter, &adapter->params.vpd, buf); if (ret < 0) return ret; /* Cards with real ASICs have the chipid in the PCIe device id */ t4_os_pci_read_cfg2(adapter, PCI_DEVICE_ID, &device_id); if (device_id >> 12 == chip_id(adapter)) adapter->params.cim_la_size = CIMLA_SIZE; else { /* FPGA */ adapter->params.fpga = 1; adapter->params.cim_la_size = 2 * CIMLA_SIZE; } init_cong_ctrl(adapter->params.a_wnd, adapter->params.b_wnd); /* * Default port and clock for debugging in case we can't reach FW. */ adapter->params.nports = 1; adapter->params.portvec = 1; adapter->params.vpd.cclk = 50000; /* Set pci completion timeout value to 4 seconds. */ set_pcie_completion_timeout(adapter, 0xd); return 0; } /** * t4_shutdown_adapter - shut down adapter, host & wire * @adapter: the adapter * * Perform an emergency shutdown of the adapter and stop it from * continuing any further communication on the ports or DMA to the * host. This is typically used when the adapter and/or firmware * have crashed and we want to prevent any further accidental * communication with the rest of the world. This will also force * the port Link Status to go down -- if register writes work -- * which should help our peers figure out that we're down. */ int t4_shutdown_adapter(struct adapter *adapter) { int port; t4_intr_disable(adapter); t4_write_reg(adapter, A_DBG_GPIO_EN, 0); for_each_port(adapter, port) { u32 a_port_cfg = PORT_REG(port, is_t4(adapter) ? A_XGMAC_PORT_CFG : A_MAC_PORT_CFG); t4_write_reg(adapter, a_port_cfg, t4_read_reg(adapter, a_port_cfg) & ~V_SIGNAL_DET(1)); } t4_set_reg_field(adapter, A_SGE_CONTROL, F_GLOBALENABLE, 0); return 0; } /** * t4_init_devlog_params - initialize adapter->params.devlog * @adap: the adapter * @fw_attach: whether we can talk to the firmware * * Initialize various fields of the adapter's Firmware Device Log * Parameters structure. */ int t4_init_devlog_params(struct adapter *adap, int fw_attach) { struct devlog_params *dparams = &adap->params.devlog; u32 pf_dparams; unsigned int devlog_meminfo; struct fw_devlog_cmd devlog_cmd; int ret; /* If we're dealing with newer firmware, the Device Log Paramerters * are stored in a designated register which allows us to access the * Device Log even if we can't talk to the firmware. */ pf_dparams = t4_read_reg(adap, PCIE_FW_REG(A_PCIE_FW_PF, PCIE_FW_PF_DEVLOG)); if (pf_dparams) { unsigned int nentries, nentries128; dparams->memtype = G_PCIE_FW_PF_DEVLOG_MEMTYPE(pf_dparams); dparams->start = G_PCIE_FW_PF_DEVLOG_ADDR16(pf_dparams) << 4; nentries128 = G_PCIE_FW_PF_DEVLOG_NENTRIES128(pf_dparams); nentries = (nentries128 + 1) * 128; dparams->size = nentries * sizeof(struct fw_devlog_e); return 0; } /* * For any failing returns ... */ memset(dparams, 0, sizeof *dparams); /* * If we can't talk to the firmware, there's really nothing we can do * at this point. */ if (!fw_attach) return -ENXIO; /* Otherwise, ask the firmware for it's Device Log Parameters. */ memset(&devlog_cmd, 0, sizeof devlog_cmd); devlog_cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_DEVLOG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); devlog_cmd.retval_len16 = cpu_to_be32(FW_LEN16(devlog_cmd)); ret = t4_wr_mbox(adap, adap->mbox, &devlog_cmd, sizeof(devlog_cmd), &devlog_cmd); if (ret) return ret; devlog_meminfo = be32_to_cpu(devlog_cmd.memtype_devlog_memaddr16_devlog); dparams->memtype = G_FW_DEVLOG_CMD_MEMTYPE_DEVLOG(devlog_meminfo); dparams->start = G_FW_DEVLOG_CMD_MEMADDR16_DEVLOG(devlog_meminfo) << 4; dparams->size = be32_to_cpu(devlog_cmd.memsize_devlog); return 0; } /** * t4_init_sge_params - initialize adap->params.sge * @adapter: the adapter * * Initialize various fields of the adapter's SGE Parameters structure. */ int t4_init_sge_params(struct adapter *adapter) { u32 r; struct sge_params *sp = &adapter->params.sge; unsigned i; r = t4_read_reg(adapter, A_SGE_INGRESS_RX_THRESHOLD); sp->counter_val[0] = G_THRESHOLD_0(r); sp->counter_val[1] = G_THRESHOLD_1(r); sp->counter_val[2] = G_THRESHOLD_2(r); sp->counter_val[3] = G_THRESHOLD_3(r); r = t4_read_reg(adapter, A_SGE_TIMER_VALUE_0_AND_1); sp->timer_val[0] = core_ticks_to_us(adapter, G_TIMERVALUE0(r)); sp->timer_val[1] = core_ticks_to_us(adapter, G_TIMERVALUE1(r)); r = t4_read_reg(adapter, A_SGE_TIMER_VALUE_2_AND_3); sp->timer_val[2] = core_ticks_to_us(adapter, G_TIMERVALUE2(r)); sp->timer_val[3] = core_ticks_to_us(adapter, G_TIMERVALUE3(r)); r = t4_read_reg(adapter, A_SGE_TIMER_VALUE_4_AND_5); sp->timer_val[4] = core_ticks_to_us(adapter, G_TIMERVALUE4(r)); sp->timer_val[5] = core_ticks_to_us(adapter, G_TIMERVALUE5(r)); r = t4_read_reg(adapter, A_SGE_CONM_CTRL); sp->fl_starve_threshold = G_EGRTHRESHOLD(r) * 2 + 1; if (is_t4(adapter)) sp->fl_starve_threshold2 = sp->fl_starve_threshold; - else + else if (is_t5(adapter)) sp->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(r) * 2 + 1; + else + sp->fl_starve_threshold2 = G_T6_EGRTHRESHOLDPACKING(r) * 2 + 1; /* egress queues: log2 of # of doorbells per BAR2 page */ r = t4_read_reg(adapter, A_SGE_EGRESS_QUEUES_PER_PAGE_PF); r >>= S_QUEUESPERPAGEPF0 + (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * adapter->pf; sp->eq_s_qpp = r & M_QUEUESPERPAGEPF0; /* ingress queues: log2 of # of doorbells per BAR2 page */ r = t4_read_reg(adapter, A_SGE_INGRESS_QUEUES_PER_PAGE_PF); r >>= S_QUEUESPERPAGEPF0 + (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * adapter->pf; sp->iq_s_qpp = r & M_QUEUESPERPAGEPF0; r = t4_read_reg(adapter, A_SGE_HOST_PAGE_SIZE); r >>= S_HOSTPAGESIZEPF0 + (S_HOSTPAGESIZEPF1 - S_HOSTPAGESIZEPF0) * adapter->pf; sp->page_shift = (r & M_HOSTPAGESIZEPF0) + 10; r = t4_read_reg(adapter, A_SGE_CONTROL); sp->sge_control = r; sp->spg_len = r & F_EGRSTATUSPAGESIZE ? 128 : 64; sp->fl_pktshift = G_PKTSHIFT(r); - sp->pad_boundary = 1 << (G_INGPADBOUNDARY(r) + 5); + if (chip_id(adapter) <= CHELSIO_T5) { + sp->pad_boundary = 1 << (G_INGPADBOUNDARY(r) + + X_INGPADBOUNDARY_SHIFT); + } else { + sp->pad_boundary = 1 << (G_INGPADBOUNDARY(r) + + X_T6_INGPADBOUNDARY_SHIFT); + } if (is_t4(adapter)) sp->pack_boundary = sp->pad_boundary; else { r = t4_read_reg(adapter, A_SGE_CONTROL2); if (G_INGPACKBOUNDARY(r) == 0) sp->pack_boundary = 16; else sp->pack_boundary = 1 << (G_INGPACKBOUNDARY(r) + 5); } for (i = 0; i < SGE_FLBUF_SIZES; i++) sp->sge_fl_buffer_size[i] = t4_read_reg(adapter, A_SGE_FL_BUFFER_SIZE0 + (4 * i)); return 0; } /* * Read and cache the adapter's compressed filter mode and ingress config. */ static void read_filter_mode_and_ingress_config(struct adapter *adap) { struct tp_params *tpp = &adap->params.tp; if (t4_use_ldst(adap)) { t4_fw_tp_pio_rw(adap, &tpp->vlan_pri_map, 1, A_TP_VLAN_PRI_MAP, 1); t4_fw_tp_pio_rw(adap, &tpp->ingress_config, 1, A_TP_INGRESS_CONFIG, 1); } else { t4_read_indirect(adap, A_TP_PIO_ADDR, A_TP_PIO_DATA, &tpp->vlan_pri_map, 1, A_TP_VLAN_PRI_MAP); t4_read_indirect(adap, A_TP_PIO_ADDR, A_TP_PIO_DATA, &tpp->ingress_config, 1, A_TP_INGRESS_CONFIG); } /* * Now that we have TP_VLAN_PRI_MAP cached, we can calculate the field * shift positions of several elements of the Compressed Filter Tuple * for this adapter which we need frequently ... */ tpp->fcoe_shift = t4_filter_field_shift(adap, F_FCOE); tpp->port_shift = t4_filter_field_shift(adap, F_PORT); tpp->vnic_shift = t4_filter_field_shift(adap, F_VNIC_ID); tpp->vlan_shift = t4_filter_field_shift(adap, F_VLAN); tpp->tos_shift = t4_filter_field_shift(adap, F_TOS); tpp->protocol_shift = t4_filter_field_shift(adap, F_PROTOCOL); tpp->ethertype_shift = t4_filter_field_shift(adap, F_ETHERTYPE); tpp->macmatch_shift = t4_filter_field_shift(adap, F_MACMATCH); tpp->matchtype_shift = t4_filter_field_shift(adap, F_MPSHITTYPE); tpp->frag_shift = t4_filter_field_shift(adap, F_FRAGMENTATION); /* * If TP_INGRESS_CONFIG.VNID == 0, then TP_VLAN_PRI_MAP.VNIC_ID * represents the presence of an Outer VLAN instead of a VNIC ID. */ if ((tpp->ingress_config & F_VNIC) == 0) tpp->vnic_shift = -1; } /** * t4_init_tp_params - initialize adap->params.tp * @adap: the adapter * * Initialize various fields of the adapter's TP Parameters structure. */ int t4_init_tp_params(struct adapter *adap) { int chan; u32 v; struct tp_params *tpp = &adap->params.tp; v = t4_read_reg(adap, A_TP_TIMER_RESOLUTION); tpp->tre = G_TIMERRESOLUTION(v); tpp->dack_re = G_DELAYEDACKRESOLUTION(v); /* MODQ_REQ_MAP defaults to setting queues 0-3 to chan 0-3 */ for (chan = 0; chan < MAX_NCHAN; chan++) tpp->tx_modq[chan] = chan; read_filter_mode_and_ingress_config(adap); /* * For T6, cache the adapter's compressed error vector * and passing outer header info for encapsulated packets. */ if (chip_id(adap) > CHELSIO_T5) { v = t4_read_reg(adap, A_TP_OUT_CONFIG); tpp->rx_pkt_encap = (v & F_CRXPKTENC) ? 1 : 0; } return 0; } /** * t4_filter_field_shift - calculate filter field shift * @adap: the adapter * @filter_sel: the desired field (from TP_VLAN_PRI_MAP bits) * * Return the shift position of a filter field within the Compressed * Filter Tuple. The filter field is specified via its selection bit * within TP_VLAN_PRI_MAL (filter mode). E.g. F_VLAN. */ int t4_filter_field_shift(const struct adapter *adap, int filter_sel) { unsigned int filter_mode = adap->params.tp.vlan_pri_map; unsigned int sel; int field_shift; if ((filter_mode & filter_sel) == 0) return -1; for (sel = 1, field_shift = 0; sel < filter_sel; sel <<= 1) { switch (filter_mode & sel) { case F_FCOE: field_shift += W_FT_FCOE; break; case F_PORT: field_shift += W_FT_PORT; break; case F_VNIC_ID: field_shift += W_FT_VNIC_ID; break; case F_VLAN: field_shift += W_FT_VLAN; break; case F_TOS: field_shift += W_FT_TOS; break; case F_PROTOCOL: field_shift += W_FT_PROTOCOL; break; case F_ETHERTYPE: field_shift += W_FT_ETHERTYPE; break; case F_MACMATCH: field_shift += W_FT_MACMATCH; break; case F_MPSHITTYPE: field_shift += W_FT_MPSHITTYPE; break; case F_FRAGMENTATION: field_shift += W_FT_FRAGMENTATION; break; } } return field_shift; } int t4_port_init(struct adapter *adap, int mbox, int pf, int vf, int port_id) { u8 addr[6]; int ret, i, j; struct fw_port_cmd c; u16 rss_size; struct port_info *p = adap2pinfo(adap, port_id); u32 param, val; memset(&c, 0, sizeof(c)); for (i = 0, j = -1; i <= p->port_id; i++) { do { j++; } while ((adap->params.portvec & (1 << j)) == 0); } if (!(adap->flags & IS_VF) || adap->params.vfres.r_caps & FW_CMD_CAP_PORT) { c.op_to_portid = htonl(V_FW_CMD_OP(FW_PORT_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | V_FW_PORT_CMD_PORTID(j)); c.action_to_len16 = htonl( V_FW_PORT_CMD_ACTION(FW_PORT_ACTION_GET_PORT_INFO) | FW_LEN16(c)); ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret) return ret; ret = be32_to_cpu(c.u.info.lstatus_to_modtype); p->mdio_addr = (ret & F_FW_PORT_CMD_MDIOCAP) ? G_FW_PORT_CMD_MDIOADDR(ret) : -1; p->port_type = G_FW_PORT_CMD_PTYPE(ret); p->mod_type = G_FW_PORT_CMD_MODTYPE(ret); init_link_config(&p->link_cfg, be16_to_cpu(c.u.info.pcap)); } ret = t4_alloc_vi(adap, mbox, j, pf, vf, 1, addr, &rss_size); if (ret < 0) return ret; p->vi[0].viid = ret; p->tx_chan = j; p->rx_chan_map = t4_get_mps_bg_map(adap, j); p->lport = j; p->vi[0].rss_size = rss_size; t4_os_set_hw_addr(adap, p->port_id, addr); param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_RSSINFO) | V_FW_PARAMS_PARAM_YZ(p->vi[0].viid); ret = t4_query_params(adap, mbox, pf, vf, 1, ¶m, &val); if (ret) p->vi[0].rss_base = 0xffff; else { /* MPASS((val >> 16) == rss_size); */ p->vi[0].rss_base = val & 0xffff; } return 0; } /** * t4_read_cimq_cfg - read CIM queue configuration * @adap: the adapter * @base: holds the queue base addresses in bytes * @size: holds the queue sizes in bytes * @thres: holds the queue full thresholds in bytes * * Returns the current configuration of the CIM queues, starting with * the IBQs, then the OBQs. */ void t4_read_cimq_cfg(struct adapter *adap, u16 *base, u16 *size, u16 *thres) { unsigned int i, v; int cim_num_obq = adap->chip_params->cim_num_obq; for (i = 0; i < CIM_NUM_IBQ; i++) { t4_write_reg(adap, A_CIM_QUEUE_CONFIG_REF, F_IBQSELECT | V_QUENUMSELECT(i)); v = t4_read_reg(adap, A_CIM_QUEUE_CONFIG_CTRL); /* value is in 256-byte units */ *base++ = G_CIMQBASE(v) * 256; *size++ = G_CIMQSIZE(v) * 256; *thres++ = G_QUEFULLTHRSH(v) * 8; /* 8-byte unit */ } for (i = 0; i < cim_num_obq; i++) { t4_write_reg(adap, A_CIM_QUEUE_CONFIG_REF, F_OBQSELECT | V_QUENUMSELECT(i)); v = t4_read_reg(adap, A_CIM_QUEUE_CONFIG_CTRL); /* value is in 256-byte units */ *base++ = G_CIMQBASE(v) * 256; *size++ = G_CIMQSIZE(v) * 256; } } /** * t4_read_cim_ibq - read the contents of a CIM inbound queue * @adap: the adapter * @qid: the queue index * @data: where to store the queue contents * @n: capacity of @data in 32-bit words * * Reads the contents of the selected CIM queue starting at address 0 up * to the capacity of @data. @n must be a multiple of 4. Returns < 0 on * error and the number of 32-bit words actually read on success. */ int t4_read_cim_ibq(struct adapter *adap, unsigned int qid, u32 *data, size_t n) { int i, err, attempts; unsigned int addr; const unsigned int nwords = CIM_IBQ_SIZE * 4; if (qid > 5 || (n & 3)) return -EINVAL; addr = qid * nwords; if (n > nwords) n = nwords; /* It might take 3-10ms before the IBQ debug read access is allowed. * Wait for 1 Sec with a delay of 1 usec. */ attempts = 1000000; for (i = 0; i < n; i++, addr++) { t4_write_reg(adap, A_CIM_IBQ_DBG_CFG, V_IBQDBGADDR(addr) | F_IBQDBGEN); err = t4_wait_op_done(adap, A_CIM_IBQ_DBG_CFG, F_IBQDBGBUSY, 0, attempts, 1); if (err) return err; *data++ = t4_read_reg(adap, A_CIM_IBQ_DBG_DATA); } t4_write_reg(adap, A_CIM_IBQ_DBG_CFG, 0); return i; } /** * t4_read_cim_obq - read the contents of a CIM outbound queue * @adap: the adapter * @qid: the queue index * @data: where to store the queue contents * @n: capacity of @data in 32-bit words * * Reads the contents of the selected CIM queue starting at address 0 up * to the capacity of @data. @n must be a multiple of 4. Returns < 0 on * error and the number of 32-bit words actually read on success. */ int t4_read_cim_obq(struct adapter *adap, unsigned int qid, u32 *data, size_t n) { int i, err; unsigned int addr, v, nwords; int cim_num_obq = adap->chip_params->cim_num_obq; if ((qid > (cim_num_obq - 1)) || (n & 3)) return -EINVAL; t4_write_reg(adap, A_CIM_QUEUE_CONFIG_REF, F_OBQSELECT | V_QUENUMSELECT(qid)); v = t4_read_reg(adap, A_CIM_QUEUE_CONFIG_CTRL); addr = G_CIMQBASE(v) * 64; /* muliple of 256 -> muliple of 4 */ nwords = G_CIMQSIZE(v) * 64; /* same */ if (n > nwords) n = nwords; for (i = 0; i < n; i++, addr++) { t4_write_reg(adap, A_CIM_OBQ_DBG_CFG, V_OBQDBGADDR(addr) | F_OBQDBGEN); err = t4_wait_op_done(adap, A_CIM_OBQ_DBG_CFG, F_OBQDBGBUSY, 0, 2, 1); if (err) return err; *data++ = t4_read_reg(adap, A_CIM_OBQ_DBG_DATA); } t4_write_reg(adap, A_CIM_OBQ_DBG_CFG, 0); return i; } enum { CIM_QCTL_BASE = 0, CIM_CTL_BASE = 0x2000, CIM_PBT_ADDR_BASE = 0x2800, CIM_PBT_LRF_BASE = 0x3000, CIM_PBT_DATA_BASE = 0x3800 }; /** * t4_cim_read - read a block from CIM internal address space * @adap: the adapter * @addr: the start address within the CIM address space * @n: number of words to read * @valp: where to store the result * * Reads a block of 4-byte words from the CIM intenal address space. */ int t4_cim_read(struct adapter *adap, unsigned int addr, unsigned int n, unsigned int *valp) { int ret = 0; if (t4_read_reg(adap, A_CIM_HOST_ACC_CTRL) & F_HOSTBUSY) return -EBUSY; for ( ; !ret && n--; addr += 4) { t4_write_reg(adap, A_CIM_HOST_ACC_CTRL, addr); ret = t4_wait_op_done(adap, A_CIM_HOST_ACC_CTRL, F_HOSTBUSY, 0, 5, 2); if (!ret) *valp++ = t4_read_reg(adap, A_CIM_HOST_ACC_DATA); } return ret; } /** * t4_cim_write - write a block into CIM internal address space * @adap: the adapter * @addr: the start address within the CIM address space * @n: number of words to write * @valp: set of values to write * * Writes a block of 4-byte words into the CIM intenal address space. */ int t4_cim_write(struct adapter *adap, unsigned int addr, unsigned int n, const unsigned int *valp) { int ret = 0; if (t4_read_reg(adap, A_CIM_HOST_ACC_CTRL) & F_HOSTBUSY) return -EBUSY; for ( ; !ret && n--; addr += 4) { t4_write_reg(adap, A_CIM_HOST_ACC_DATA, *valp++); t4_write_reg(adap, A_CIM_HOST_ACC_CTRL, addr | F_HOSTWRITE); ret = t4_wait_op_done(adap, A_CIM_HOST_ACC_CTRL, F_HOSTBUSY, 0, 5, 2); } return ret; } static int t4_cim_write1(struct adapter *adap, unsigned int addr, unsigned int val) { return t4_cim_write(adap, addr, 1, &val); } /** * t4_cim_ctl_read - read a block from CIM control region * @adap: the adapter * @addr: the start address within the CIM control region * @n: number of words to read * @valp: where to store the result * * Reads a block of 4-byte words from the CIM control region. */ int t4_cim_ctl_read(struct adapter *adap, unsigned int addr, unsigned int n, unsigned int *valp) { return t4_cim_read(adap, addr + CIM_CTL_BASE, n, valp); } /** * t4_cim_read_la - read CIM LA capture buffer * @adap: the adapter * @la_buf: where to store the LA data * @wrptr: the HW write pointer within the capture buffer * * Reads the contents of the CIM LA buffer with the most recent entry at * the end of the returned data and with the entry at @wrptr first. * We try to leave the LA in the running state we find it in. */ int t4_cim_read_la(struct adapter *adap, u32 *la_buf, unsigned int *wrptr) { int i, ret; unsigned int cfg, val, idx; ret = t4_cim_read(adap, A_UP_UP_DBG_LA_CFG, 1, &cfg); if (ret) return ret; if (cfg & F_UPDBGLAEN) { /* LA is running, freeze it */ ret = t4_cim_write1(adap, A_UP_UP_DBG_LA_CFG, 0); if (ret) return ret; } ret = t4_cim_read(adap, A_UP_UP_DBG_LA_CFG, 1, &val); if (ret) goto restart; idx = G_UPDBGLAWRPTR(val); if (wrptr) *wrptr = idx; for (i = 0; i < adap->params.cim_la_size; i++) { ret = t4_cim_write1(adap, A_UP_UP_DBG_LA_CFG, V_UPDBGLARDPTR(idx) | F_UPDBGLARDEN); if (ret) break; ret = t4_cim_read(adap, A_UP_UP_DBG_LA_CFG, 1, &val); if (ret) break; if (val & F_UPDBGLARDEN) { ret = -ETIMEDOUT; break; } ret = t4_cim_read(adap, A_UP_UP_DBG_LA_DATA, 1, &la_buf[i]); if (ret) break; /* address can't exceed 0xfff (UpDbgLaRdPtr is of 12-bits) */ idx = (idx + 1) & M_UPDBGLARDPTR; /* * Bits 0-3 of UpDbgLaRdPtr can be between 0000 to 1001 to * identify the 32-bit portion of the full 312-bit data */ if (is_t6(adap)) while ((idx & 0xf) > 9) idx = (idx + 1) % M_UPDBGLARDPTR; } restart: if (cfg & F_UPDBGLAEN) { int r = t4_cim_write1(adap, A_UP_UP_DBG_LA_CFG, cfg & ~F_UPDBGLARDEN); if (!ret) ret = r; } return ret; } /** * t4_tp_read_la - read TP LA capture buffer * @adap: the adapter * @la_buf: where to store the LA data * @wrptr: the HW write pointer within the capture buffer * * Reads the contents of the TP LA buffer with the most recent entry at * the end of the returned data and with the entry at @wrptr first. * We leave the LA in the running state we find it in. */ void t4_tp_read_la(struct adapter *adap, u64 *la_buf, unsigned int *wrptr) { bool last_incomplete; unsigned int i, cfg, val, idx; cfg = t4_read_reg(adap, A_TP_DBG_LA_CONFIG) & 0xffff; if (cfg & F_DBGLAENABLE) /* freeze LA */ t4_write_reg(adap, A_TP_DBG_LA_CONFIG, adap->params.tp.la_mask | (cfg ^ F_DBGLAENABLE)); val = t4_read_reg(adap, A_TP_DBG_LA_CONFIG); idx = G_DBGLAWPTR(val); last_incomplete = G_DBGLAMODE(val) >= 2 && (val & F_DBGLAWHLF) == 0; if (last_incomplete) idx = (idx + 1) & M_DBGLARPTR; if (wrptr) *wrptr = idx; val &= 0xffff; val &= ~V_DBGLARPTR(M_DBGLARPTR); val |= adap->params.tp.la_mask; for (i = 0; i < TPLA_SIZE; i++) { t4_write_reg(adap, A_TP_DBG_LA_CONFIG, V_DBGLARPTR(idx) | val); la_buf[i] = t4_read_reg64(adap, A_TP_DBG_LA_DATAL); idx = (idx + 1) & M_DBGLARPTR; } /* Wipe out last entry if it isn't valid */ if (last_incomplete) la_buf[TPLA_SIZE - 1] = ~0ULL; if (cfg & F_DBGLAENABLE) /* restore running state */ t4_write_reg(adap, A_TP_DBG_LA_CONFIG, cfg | adap->params.tp.la_mask); } /* * SGE Hung Ingress DMA Warning Threshold time and Warning Repeat Rate (in * seconds). If we find one of the SGE Ingress DMA State Machines in the same * state for more than the Warning Threshold then we'll issue a warning about * a potential hang. We'll repeat the warning as the SGE Ingress DMA Channel * appears to be hung every Warning Repeat second till the situation clears. * If the situation clears, we'll note that as well. */ #define SGE_IDMA_WARN_THRESH 1 #define SGE_IDMA_WARN_REPEAT 300 /** * t4_idma_monitor_init - initialize SGE Ingress DMA Monitor * @adapter: the adapter * @idma: the adapter IDMA Monitor state * * Initialize the state of an SGE Ingress DMA Monitor. */ void t4_idma_monitor_init(struct adapter *adapter, struct sge_idma_monitor_state *idma) { /* Initialize the state variables for detecting an SGE Ingress DMA * hang. The SGE has internal counters which count up on each clock * tick whenever the SGE finds its Ingress DMA State Engines in the * same state they were on the previous clock tick. The clock used is * the Core Clock so we have a limit on the maximum "time" they can * record; typically a very small number of seconds. For instance, * with a 600MHz Core Clock, we can only count up to a bit more than * 7s. So we'll synthesize a larger counter in order to not run the * risk of having the "timers" overflow and give us the flexibility to * maintain a Hung SGE State Machine of our own which operates across * a longer time frame. */ idma->idma_1s_thresh = core_ticks_per_usec(adapter) * 1000000; /* 1s */ idma->idma_stalled[0] = idma->idma_stalled[1] = 0; } /** * t4_idma_monitor - monitor SGE Ingress DMA state * @adapter: the adapter * @idma: the adapter IDMA Monitor state * @hz: number of ticks/second * @ticks: number of ticks since the last IDMA Monitor call */ void t4_idma_monitor(struct adapter *adapter, struct sge_idma_monitor_state *idma, int hz, int ticks) { int i, idma_same_state_cnt[2]; /* Read the SGE Debug Ingress DMA Same State Count registers. These * are counters inside the SGE which count up on each clock when the * SGE finds its Ingress DMA State Engines in the same states they * were in the previous clock. The counters will peg out at * 0xffffffff without wrapping around so once they pass the 1s * threshold they'll stay above that till the IDMA state changes. */ t4_write_reg(adapter, A_SGE_DEBUG_INDEX, 13); idma_same_state_cnt[0] = t4_read_reg(adapter, A_SGE_DEBUG_DATA_HIGH); idma_same_state_cnt[1] = t4_read_reg(adapter, A_SGE_DEBUG_DATA_LOW); for (i = 0; i < 2; i++) { u32 debug0, debug11; /* If the Ingress DMA Same State Counter ("timer") is less * than 1s, then we can reset our synthesized Stall Timer and * continue. If we have previously emitted warnings about a * potential stalled Ingress Queue, issue a note indicating * that the Ingress Queue has resumed forward progress. */ if (idma_same_state_cnt[i] < idma->idma_1s_thresh) { if (idma->idma_stalled[i] >= SGE_IDMA_WARN_THRESH*hz) CH_WARN(adapter, "SGE idma%d, queue %u, " "resumed after %d seconds\n", i, idma->idma_qid[i], idma->idma_stalled[i]/hz); idma->idma_stalled[i] = 0; continue; } /* Synthesize an SGE Ingress DMA Same State Timer in the Hz * domain. The first time we get here it'll be because we * passed the 1s Threshold; each additional time it'll be * because the RX Timer Callback is being fired on its regular * schedule. * * If the stall is below our Potential Hung Ingress Queue * Warning Threshold, continue. */ if (idma->idma_stalled[i] == 0) { idma->idma_stalled[i] = hz; idma->idma_warn[i] = 0; } else { idma->idma_stalled[i] += ticks; idma->idma_warn[i] -= ticks; } if (idma->idma_stalled[i] < SGE_IDMA_WARN_THRESH*hz) continue; /* We'll issue a warning every SGE_IDMA_WARN_REPEAT seconds. */ if (idma->idma_warn[i] > 0) continue; idma->idma_warn[i] = SGE_IDMA_WARN_REPEAT*hz; /* Read and save the SGE IDMA State and Queue ID information. * We do this every time in case it changes across time ... * can't be too careful ... */ t4_write_reg(adapter, A_SGE_DEBUG_INDEX, 0); debug0 = t4_read_reg(adapter, A_SGE_DEBUG_DATA_LOW); idma->idma_state[i] = (debug0 >> (i * 9)) & 0x3f; t4_write_reg(adapter, A_SGE_DEBUG_INDEX, 11); debug11 = t4_read_reg(adapter, A_SGE_DEBUG_DATA_LOW); idma->idma_qid[i] = (debug11 >> (i * 16)) & 0xffff; CH_WARN(adapter, "SGE idma%u, queue %u, potentially stuck in " " state %u for %d seconds (debug0=%#x, debug11=%#x)\n", i, idma->idma_qid[i], idma->idma_state[i], idma->idma_stalled[i]/hz, debug0, debug11); t4_sge_decode_idma_state(adapter, idma->idma_state[i]); } } /** * t4_read_pace_tbl - read the pace table * @adap: the adapter * @pace_vals: holds the returned values * * Returns the values of TP's pace table in microseconds. */ void t4_read_pace_tbl(struct adapter *adap, unsigned int pace_vals[NTX_SCHED]) { unsigned int i, v; for (i = 0; i < NTX_SCHED; i++) { t4_write_reg(adap, A_TP_PACE_TABLE, 0xffff0000 + i); v = t4_read_reg(adap, A_TP_PACE_TABLE); pace_vals[i] = dack_ticks_to_usec(adap, v); } } /** * t4_get_tx_sched - get the configuration of a Tx HW traffic scheduler * @adap: the adapter * @sched: the scheduler index * @kbps: the byte rate in Kbps * @ipg: the interpacket delay in tenths of nanoseconds * * Return the current configuration of a HW Tx scheduler. */ void t4_get_tx_sched(struct adapter *adap, unsigned int sched, unsigned int *kbps, unsigned int *ipg) { unsigned int v, addr, bpt, cpt; if (kbps) { addr = A_TP_TX_MOD_Q1_Q0_RATE_LIMIT - sched / 2; t4_write_reg(adap, A_TP_TM_PIO_ADDR, addr); v = t4_read_reg(adap, A_TP_TM_PIO_DATA); if (sched & 1) v >>= 16; bpt = (v >> 8) & 0xff; cpt = v & 0xff; if (!cpt) *kbps = 0; /* scheduler disabled */ else { v = (adap->params.vpd.cclk * 1000) / cpt; /* ticks/s */ *kbps = (v * bpt) / 125; } } if (ipg) { addr = A_TP_TX_MOD_Q1_Q0_TIMER_SEPARATOR - sched / 2; t4_write_reg(adap, A_TP_TM_PIO_ADDR, addr); v = t4_read_reg(adap, A_TP_TM_PIO_DATA); if (sched & 1) v >>= 16; v &= 0xffff; *ipg = (10000 * v) / core_ticks_per_usec(adap); } } /** * t4_load_cfg - download config file * @adap: the adapter * @cfg_data: the cfg text file to write * @size: text file size * * Write the supplied config text file to the card's serial flash. */ int t4_load_cfg(struct adapter *adap, const u8 *cfg_data, unsigned int size) { int ret, i, n, cfg_addr; unsigned int addr; unsigned int flash_cfg_start_sec; unsigned int sf_sec_size = adap->params.sf_size / adap->params.sf_nsec; cfg_addr = t4_flash_cfg_addr(adap); if (cfg_addr < 0) return cfg_addr; addr = cfg_addr; flash_cfg_start_sec = addr / SF_SEC_SIZE; if (size > FLASH_CFG_MAX_SIZE) { CH_ERR(adap, "cfg file too large, max is %u bytes\n", FLASH_CFG_MAX_SIZE); return -EFBIG; } i = DIV_ROUND_UP(FLASH_CFG_MAX_SIZE, /* # of sectors spanned */ sf_sec_size); ret = t4_flash_erase_sectors(adap, flash_cfg_start_sec, flash_cfg_start_sec + i - 1); /* * If size == 0 then we're simply erasing the FLASH sectors associated * with the on-adapter Firmware Configuration File. */ if (ret || size == 0) goto out; /* this will write to the flash up to SF_PAGE_SIZE at a time */ for (i = 0; i< size; i+= SF_PAGE_SIZE) { if ( (size - i) < SF_PAGE_SIZE) n = size - i; else n = SF_PAGE_SIZE; ret = t4_write_flash(adap, addr, n, cfg_data, 1); if (ret) goto out; addr += SF_PAGE_SIZE; cfg_data += SF_PAGE_SIZE; } out: if (ret) CH_ERR(adap, "config file %s failed %d\n", (size == 0 ? "clear" : "download"), ret); return ret; } /** * t5_fw_init_extern_mem - initialize the external memory * @adap: the adapter * * Initializes the external memory on T5. */ int t5_fw_init_extern_mem(struct adapter *adap) { u32 params[1], val[1]; int ret; if (!is_t5(adap)) return 0; val[0] = 0xff; /* Initialize all MCs */ params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_MCINIT)); ret = t4_set_params_timeout(adap, adap->mbox, adap->pf, 0, 1, params, val, FW_CMD_MAX_TIMEOUT); return ret; } /* BIOS boot headers */ typedef struct pci_expansion_rom_header { u8 signature[2]; /* ROM Signature. Should be 0xaa55 */ u8 reserved[22]; /* Reserved per processor Architecture data */ u8 pcir_offset[2]; /* Offset to PCI Data Structure */ } pci_exp_rom_header_t; /* PCI_EXPANSION_ROM_HEADER */ /* Legacy PCI Expansion ROM Header */ typedef struct legacy_pci_expansion_rom_header { u8 signature[2]; /* ROM Signature. Should be 0xaa55 */ u8 size512; /* Current Image Size in units of 512 bytes */ u8 initentry_point[4]; u8 cksum; /* Checksum computed on the entire Image */ u8 reserved[16]; /* Reserved */ u8 pcir_offset[2]; /* Offset to PCI Data Struture */ } legacy_pci_exp_rom_header_t; /* LEGACY_PCI_EXPANSION_ROM_HEADER */ /* EFI PCI Expansion ROM Header */ typedef struct efi_pci_expansion_rom_header { u8 signature[2]; // ROM signature. The value 0xaa55 u8 initialization_size[2]; /* Units 512. Includes this header */ u8 efi_signature[4]; /* Signature from EFI image header. 0x0EF1 */ u8 efi_subsystem[2]; /* Subsystem value for EFI image header */ u8 efi_machine_type[2]; /* Machine type from EFI image header */ u8 compression_type[2]; /* Compression type. */ /* * Compression type definition * 0x0: uncompressed * 0x1: Compressed * 0x2-0xFFFF: Reserved */ u8 reserved[8]; /* Reserved */ u8 efi_image_header_offset[2]; /* Offset to EFI Image */ u8 pcir_offset[2]; /* Offset to PCI Data Structure */ } efi_pci_exp_rom_header_t; /* EFI PCI Expansion ROM Header */ /* PCI Data Structure Format */ typedef struct pcir_data_structure { /* PCI Data Structure */ u8 signature[4]; /* Signature. The string "PCIR" */ u8 vendor_id[2]; /* Vendor Identification */ u8 device_id[2]; /* Device Identification */ u8 vital_product[2]; /* Pointer to Vital Product Data */ u8 length[2]; /* PCIR Data Structure Length */ u8 revision; /* PCIR Data Structure Revision */ u8 class_code[3]; /* Class Code */ u8 image_length[2]; /* Image Length. Multiple of 512B */ u8 code_revision[2]; /* Revision Level of Code/Data */ u8 code_type; /* Code Type. */ /* * PCI Expansion ROM Code Types * 0x00: Intel IA-32, PC-AT compatible. Legacy * 0x01: Open Firmware standard for PCI. FCODE * 0x02: Hewlett-Packard PA RISC. HP reserved * 0x03: EFI Image. EFI * 0x04-0xFF: Reserved. */ u8 indicator; /* Indicator. Identifies the last image in the ROM */ u8 reserved[2]; /* Reserved */ } pcir_data_t; /* PCI__DATA_STRUCTURE */ /* BOOT constants */ enum { BOOT_FLASH_BOOT_ADDR = 0x0,/* start address of boot image in flash */ BOOT_SIGNATURE = 0xaa55, /* signature of BIOS boot ROM */ BOOT_SIZE_INC = 512, /* image size measured in 512B chunks */ BOOT_MIN_SIZE = sizeof(pci_exp_rom_header_t), /* basic header */ BOOT_MAX_SIZE = 1024*BOOT_SIZE_INC, /* 1 byte * length increment */ VENDOR_ID = 0x1425, /* Vendor ID */ PCIR_SIGNATURE = 0x52494350 /* PCIR signature */ }; /* * modify_device_id - Modifies the device ID of the Boot BIOS image * @adatper: the device ID to write. * @boot_data: the boot image to modify. * * Write the supplied device ID to the boot BIOS image. */ static void modify_device_id(int device_id, u8 *boot_data) { legacy_pci_exp_rom_header_t *header; pcir_data_t *pcir_header; u32 cur_header = 0; /* * Loop through all chained images and change the device ID's */ while (1) { header = (legacy_pci_exp_rom_header_t *) &boot_data[cur_header]; pcir_header = (pcir_data_t *) &boot_data[cur_header + le16_to_cpu(*(u16*)header->pcir_offset)]; /* * Only modify the Device ID if code type is Legacy or HP. * 0x00: Okay to modify * 0x01: FCODE. Do not be modify * 0x03: Okay to modify * 0x04-0xFF: Do not modify */ if (pcir_header->code_type == 0x00) { u8 csum = 0; int i; /* * Modify Device ID to match current adatper */ *(u16*) pcir_header->device_id = device_id; /* * Set checksum temporarily to 0. * We will recalculate it later. */ header->cksum = 0x0; /* * Calculate and update checksum */ for (i = 0; i < (header->size512 * 512); i++) csum += (u8)boot_data[cur_header + i]; /* * Invert summed value to create the checksum * Writing new checksum value directly to the boot data */ boot_data[cur_header + 7] = -csum; } else if (pcir_header->code_type == 0x03) { /* * Modify Device ID to match current adatper */ *(u16*) pcir_header->device_id = device_id; } /* * Check indicator element to identify if this is the last * image in the ROM. */ if (pcir_header->indicator & 0x80) break; /* * Move header pointer up to the next image in the ROM. */ cur_header += header->size512 * 512; } } /* * t4_load_boot - download boot flash * @adapter: the adapter * @boot_data: the boot image to write * @boot_addr: offset in flash to write boot_data * @size: image size * * Write the supplied boot image to the card's serial flash. * The boot image has the following sections: a 28-byte header and the * boot image. */ int t4_load_boot(struct adapter *adap, u8 *boot_data, unsigned int boot_addr, unsigned int size) { pci_exp_rom_header_t *header; int pcir_offset ; pcir_data_t *pcir_header; int ret, addr; uint16_t device_id; unsigned int i; unsigned int boot_sector = (boot_addr * 1024 ); unsigned int sf_sec_size = adap->params.sf_size / adap->params.sf_nsec; /* * Make sure the boot image does not encroach on the firmware region */ if ((boot_sector + size) >> 16 > FLASH_FW_START_SEC) { CH_ERR(adap, "boot image encroaching on firmware region\n"); return -EFBIG; } /* * The boot sector is comprised of the Expansion-ROM boot, iSCSI boot, * and Boot configuration data sections. These 3 boot sections span * sectors 0 to 7 in flash and live right before the FW image location. */ i = DIV_ROUND_UP(size ? size : FLASH_FW_START, sf_sec_size); ret = t4_flash_erase_sectors(adap, boot_sector >> 16, (boot_sector >> 16) + i - 1); /* * If size == 0 then we're simply erasing the FLASH sectors associated * with the on-adapter option ROM file */ if (ret || (size == 0)) goto out; /* Get boot header */ header = (pci_exp_rom_header_t *)boot_data; pcir_offset = le16_to_cpu(*(u16 *)header->pcir_offset); /* PCIR Data Structure */ pcir_header = (pcir_data_t *) &boot_data[pcir_offset]; /* * Perform some primitive sanity testing to avoid accidentally * writing garbage over the boot sectors. We ought to check for * more but it's not worth it for now ... */ if (size < BOOT_MIN_SIZE || size > BOOT_MAX_SIZE) { CH_ERR(adap, "boot image too small/large\n"); return -EFBIG; } #ifndef CHELSIO_T4_DIAGS /* * Check BOOT ROM header signature */ if (le16_to_cpu(*(u16*)header->signature) != BOOT_SIGNATURE ) { CH_ERR(adap, "Boot image missing signature\n"); return -EINVAL; } /* * Check PCI header signature */ if (le32_to_cpu(*(u32*)pcir_header->signature) != PCIR_SIGNATURE) { CH_ERR(adap, "PCI header missing signature\n"); return -EINVAL; } /* * Check Vendor ID matches Chelsio ID */ if (le16_to_cpu(*(u16*)pcir_header->vendor_id) != VENDOR_ID) { CH_ERR(adap, "Vendor ID missing signature\n"); return -EINVAL; } #endif /* * Retrieve adapter's device ID */ t4_os_pci_read_cfg2(adap, PCI_DEVICE_ID, &device_id); /* Want to deal with PF 0 so I strip off PF 4 indicator */ device_id = device_id & 0xf0ff; /* * Check PCIE Device ID */ if (le16_to_cpu(*(u16*)pcir_header->device_id) != device_id) { /* * Change the device ID in the Boot BIOS image to match * the Device ID of the current adapter. */ modify_device_id(device_id, boot_data); } /* * Skip over the first SF_PAGE_SIZE worth of data and write it after * we finish copying the rest of the boot image. This will ensure * that the BIOS boot header will only be written if the boot image * was written in full. */ addr = boot_sector; for (size -= SF_PAGE_SIZE; size; size -= SF_PAGE_SIZE) { addr += SF_PAGE_SIZE; boot_data += SF_PAGE_SIZE; ret = t4_write_flash(adap, addr, SF_PAGE_SIZE, boot_data, 0); if (ret) goto out; } ret = t4_write_flash(adap, boot_sector, SF_PAGE_SIZE, (const u8 *)header, 0); out: if (ret) CH_ERR(adap, "boot image download failed, error %d\n", ret); return ret; } /* * t4_flash_bootcfg_addr - return the address of the flash optionrom configuration * @adapter: the adapter * * Return the address within the flash where the OptionROM Configuration * is stored, or an error if the device FLASH is too small to contain * a OptionROM Configuration. */ static int t4_flash_bootcfg_addr(struct adapter *adapter) { /* * If the device FLASH isn't large enough to hold a Firmware * Configuration File, return an error. */ if (adapter->params.sf_size < FLASH_BOOTCFG_START + FLASH_BOOTCFG_MAX_SIZE) return -ENOSPC; return FLASH_BOOTCFG_START; } int t4_load_bootcfg(struct adapter *adap,const u8 *cfg_data, unsigned int size) { int ret, i, n, cfg_addr; unsigned int addr; unsigned int flash_cfg_start_sec; unsigned int sf_sec_size = adap->params.sf_size / adap->params.sf_nsec; cfg_addr = t4_flash_bootcfg_addr(adap); if (cfg_addr < 0) return cfg_addr; addr = cfg_addr; flash_cfg_start_sec = addr / SF_SEC_SIZE; if (size > FLASH_BOOTCFG_MAX_SIZE) { CH_ERR(adap, "bootcfg file too large, max is %u bytes\n", FLASH_BOOTCFG_MAX_SIZE); return -EFBIG; } i = DIV_ROUND_UP(FLASH_BOOTCFG_MAX_SIZE,/* # of sectors spanned */ sf_sec_size); ret = t4_flash_erase_sectors(adap, flash_cfg_start_sec, flash_cfg_start_sec + i - 1); /* * If size == 0 then we're simply erasing the FLASH sectors associated * with the on-adapter OptionROM Configuration File. */ if (ret || size == 0) goto out; /* this will write to the flash up to SF_PAGE_SIZE at a time */ for (i = 0; i< size; i+= SF_PAGE_SIZE) { if ( (size - i) < SF_PAGE_SIZE) n = size - i; else n = SF_PAGE_SIZE; ret = t4_write_flash(adap, addr, n, cfg_data, 0); if (ret) goto out; addr += SF_PAGE_SIZE; cfg_data += SF_PAGE_SIZE; } out: if (ret) CH_ERR(adap, "boot config data %s failed %d\n", (size == 0 ? "clear" : "download"), ret); return ret; } /** * t4_set_filter_mode - configure the optional components of filter tuples * @adap: the adapter * @mode_map: a bitmap selcting which optional filter components to enable * * Sets the filter mode by selecting the optional components to enable * in filter tuples. Returns 0 on success and a negative error if the * requested mode needs more bits than are available for optional * components. */ int t4_set_filter_mode(struct adapter *adap, unsigned int mode_map) { static u8 width[] = { 1, 3, 17, 17, 8, 8, 16, 9, 3, 1 }; int i, nbits = 0; for (i = S_FCOE; i <= S_FRAGMENTATION; i++) if (mode_map & (1 << i)) nbits += width[i]; if (nbits > FILTER_OPT_LEN) return -EINVAL; if (t4_use_ldst(adap)) t4_fw_tp_pio_rw(adap, &mode_map, 1, A_TP_VLAN_PRI_MAP, 0); else t4_write_indirect(adap, A_TP_PIO_ADDR, A_TP_PIO_DATA, &mode_map, 1, A_TP_VLAN_PRI_MAP); read_filter_mode_and_ingress_config(adap); return 0; } /** * t4_clr_port_stats - clear port statistics * @adap: the adapter * @idx: the port index * * Clear HW statistics for the given port. */ void t4_clr_port_stats(struct adapter *adap, int idx) { unsigned int i; u32 bgmap = t4_get_mps_bg_map(adap, idx); u32 port_base_addr; if (is_t4(adap)) port_base_addr = PORT_BASE(idx); else port_base_addr = T5_PORT_BASE(idx); for (i = A_MPS_PORT_STAT_TX_PORT_BYTES_L; i <= A_MPS_PORT_STAT_TX_PORT_PPP7_H; i += 8) t4_write_reg(adap, port_base_addr + i, 0); for (i = A_MPS_PORT_STAT_RX_PORT_BYTES_L; i <= A_MPS_PORT_STAT_RX_PORT_LESS_64B_H; i += 8) t4_write_reg(adap, port_base_addr + i, 0); for (i = 0; i < 4; i++) if (bgmap & (1 << i)) { t4_write_reg(adap, A_MPS_STAT_RX_BG_0_MAC_DROP_FRAME_L + i * 8, 0); t4_write_reg(adap, A_MPS_STAT_RX_BG_0_MAC_TRUNC_FRAME_L + i * 8, 0); } } /** * t4_i2c_rd - read I2C data from adapter * @adap: the adapter * @port: Port number if per-port device; <0 if not * @devid: per-port device ID or absolute device ID * @offset: byte offset into device I2C space * @len: byte length of I2C space data * @buf: buffer in which to return I2C data * * Reads the I2C data from the indicated device and location. */ int t4_i2c_rd(struct adapter *adap, unsigned int mbox, int port, unsigned int devid, unsigned int offset, unsigned int len, u8 *buf) { u32 ldst_addrspace; struct fw_ldst_cmd ldst; int ret; if (port >= 4 || devid >= 256 || offset >= 256 || len > sizeof ldst.u.i2c.data) return -EINVAL; memset(&ldst, 0, sizeof ldst); ldst_addrspace = V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_I2C); ldst.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | ldst_addrspace); ldst.cycles_to_len16 = cpu_to_be32(FW_LEN16(ldst)); ldst.u.i2c.pid = (port < 0 ? 0xff : port); ldst.u.i2c.did = devid; ldst.u.i2c.boffset = offset; ldst.u.i2c.blen = len; ret = t4_wr_mbox(adap, mbox, &ldst, sizeof ldst, &ldst); if (!ret) memcpy(buf, ldst.u.i2c.data, len); return ret; } /** * t4_i2c_wr - write I2C data to adapter * @adap: the adapter * @port: Port number if per-port device; <0 if not * @devid: per-port device ID or absolute device ID * @offset: byte offset into device I2C space * @len: byte length of I2C space data * @buf: buffer containing new I2C data * * Write the I2C data to the indicated device and location. */ int t4_i2c_wr(struct adapter *adap, unsigned int mbox, int port, unsigned int devid, unsigned int offset, unsigned int len, u8 *buf) { u32 ldst_addrspace; struct fw_ldst_cmd ldst; if (port >= 4 || devid >= 256 || offset >= 256 || len > sizeof ldst.u.i2c.data) return -EINVAL; memset(&ldst, 0, sizeof ldst); ldst_addrspace = V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_I2C); ldst.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | ldst_addrspace); ldst.cycles_to_len16 = cpu_to_be32(FW_LEN16(ldst)); ldst.u.i2c.pid = (port < 0 ? 0xff : port); ldst.u.i2c.did = devid; ldst.u.i2c.boffset = offset; ldst.u.i2c.blen = len; memcpy(ldst.u.i2c.data, buf, len); return t4_wr_mbox(adap, mbox, &ldst, sizeof ldst, &ldst); } /** * t4_sge_ctxt_rd - read an SGE context through FW * @adap: the adapter * @mbox: mailbox to use for the FW command * @cid: the context id * @ctype: the context type * @data: where to store the context data * * Issues a FW command through the given mailbox to read an SGE context. */ int t4_sge_ctxt_rd(struct adapter *adap, unsigned int mbox, unsigned int cid, enum ctxt_type ctype, u32 *data) { int ret; struct fw_ldst_cmd c; if (ctype == CTXT_EGRESS) ret = FW_LDST_ADDRSPC_SGE_EGRC; else if (ctype == CTXT_INGRESS) ret = FW_LDST_ADDRSPC_SGE_INGC; else if (ctype == CTXT_FLM) ret = FW_LDST_ADDRSPC_SGE_FLMC; else ret = FW_LDST_ADDRSPC_SGE_CONMC; memset(&c, 0, sizeof(c)); c.op_to_addrspace = cpu_to_be32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | V_FW_LDST_CMD_ADDRSPACE(ret)); c.cycles_to_len16 = cpu_to_be32(FW_LEN16(c)); c.u.idctxt.physid = cpu_to_be32(cid); ret = t4_wr_mbox(adap, mbox, &c, sizeof(c), &c); if (ret == 0) { data[0] = be32_to_cpu(c.u.idctxt.ctxt_data0); data[1] = be32_to_cpu(c.u.idctxt.ctxt_data1); data[2] = be32_to_cpu(c.u.idctxt.ctxt_data2); data[3] = be32_to_cpu(c.u.idctxt.ctxt_data3); data[4] = be32_to_cpu(c.u.idctxt.ctxt_data4); data[5] = be32_to_cpu(c.u.idctxt.ctxt_data5); } return ret; } /** * t4_sge_ctxt_rd_bd - read an SGE context bypassing FW * @adap: the adapter * @cid: the context id * @ctype: the context type * @data: where to store the context data * * Reads an SGE context directly, bypassing FW. This is only for * debugging when FW is unavailable. */ int t4_sge_ctxt_rd_bd(struct adapter *adap, unsigned int cid, enum ctxt_type ctype, u32 *data) { int i, ret; t4_write_reg(adap, A_SGE_CTXT_CMD, V_CTXTQID(cid) | V_CTXTTYPE(ctype)); ret = t4_wait_op_done(adap, A_SGE_CTXT_CMD, F_BUSY, 0, 3, 1); if (!ret) for (i = A_SGE_CTXT_DATA0; i <= A_SGE_CTXT_DATA5; i += 4) *data++ = t4_read_reg(adap, i); return ret; } int t4_sched_config(struct adapter *adapter, int type, int minmaxen, int sleep_ok) { struct fw_sched_cmd cmd; memset(&cmd, 0, sizeof(cmd)); cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_SCHED_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE); cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd)); cmd.u.config.sc = FW_SCHED_SC_CONFIG; cmd.u.config.type = type; cmd.u.config.minmaxen = minmaxen; return t4_wr_mbox_meat(adapter,adapter->mbox, &cmd, sizeof(cmd), NULL, sleep_ok); } int t4_sched_params(struct adapter *adapter, int type, int level, int mode, int rateunit, int ratemode, int channel, int cl, int minrate, int maxrate, int weight, int pktsize, int sleep_ok) { struct fw_sched_cmd cmd; memset(&cmd, 0, sizeof(cmd)); cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_SCHED_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE); cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd)); cmd.u.params.sc = FW_SCHED_SC_PARAMS; cmd.u.params.type = type; cmd.u.params.level = level; cmd.u.params.mode = mode; cmd.u.params.ch = channel; cmd.u.params.cl = cl; cmd.u.params.unit = rateunit; cmd.u.params.rate = ratemode; cmd.u.params.min = cpu_to_be32(minrate); cmd.u.params.max = cpu_to_be32(maxrate); cmd.u.params.weight = cpu_to_be16(weight); cmd.u.params.pktsize = cpu_to_be16(pktsize); return t4_wr_mbox_meat(adapter,adapter->mbox, &cmd, sizeof(cmd), NULL, sleep_ok); } /* * t4_config_watchdog - configure (enable/disable) a watchdog timer * @adapter: the adapter * @mbox: mailbox to use for the FW command * @pf: the PF owning the queue * @vf: the VF owning the queue * @timeout: watchdog timeout in ms * @action: watchdog timer / action * * There are separate watchdog timers for each possible watchdog * action. Configure one of the watchdog timers by setting a non-zero * timeout. Disable a watchdog timer by using a timeout of zero. */ int t4_config_watchdog(struct adapter *adapter, unsigned int mbox, unsigned int pf, unsigned int vf, unsigned int timeout, unsigned int action) { struct fw_watchdog_cmd wdog; unsigned int ticks; /* * The watchdog command expects a timeout in units of 10ms so we need * to convert it here (via rounding) and force a minimum of one 10ms * "tick" if the timeout is non-zero but the conversion results in 0 * ticks. */ ticks = (timeout + 5)/10; if (timeout && !ticks) ticks = 1; memset(&wdog, 0, sizeof wdog); wdog.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_WATCHDOG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | V_FW_PARAMS_CMD_PFN(pf) | V_FW_PARAMS_CMD_VFN(vf)); wdog.retval_len16 = cpu_to_be32(FW_LEN16(wdog)); wdog.timeout = cpu_to_be32(ticks); wdog.action = cpu_to_be32(action); return t4_wr_mbox(adapter, mbox, &wdog, sizeof wdog, NULL); } int t4_get_devlog_level(struct adapter *adapter, unsigned int *level) { struct fw_devlog_cmd devlog_cmd; int ret; memset(&devlog_cmd, 0, sizeof(devlog_cmd)); devlog_cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_DEVLOG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); devlog_cmd.retval_len16 = cpu_to_be32(FW_LEN16(devlog_cmd)); ret = t4_wr_mbox(adapter, adapter->mbox, &devlog_cmd, sizeof(devlog_cmd), &devlog_cmd); if (ret) return ret; *level = devlog_cmd.level; return 0; } int t4_set_devlog_level(struct adapter *adapter, unsigned int level) { struct fw_devlog_cmd devlog_cmd; memset(&devlog_cmd, 0, sizeof(devlog_cmd)); devlog_cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_DEVLOG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE); devlog_cmd.level = level; devlog_cmd.retval_len16 = cpu_to_be32(FW_LEN16(devlog_cmd)); return t4_wr_mbox(adapter, adapter->mbox, &devlog_cmd, sizeof(devlog_cmd), &devlog_cmd); } Index: projects/mips64-clang/sys/dev/cxgbe/common/t4vf_hw.c =================================================================== --- projects/mips64-clang/sys/dev/cxgbe/common/t4vf_hw.c (revision 305711) +++ projects/mips64-clang/sys/dev/cxgbe/common/t4vf_hw.c (revision 305712) @@ -1,376 +1,384 @@ /*- * Copyright (c) 2016 Chelsio Communications, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "common.h" #include "t4_regs.h" +#include "t4_regs_values.h" #undef msleep #define msleep(x) do { \ if (cold) \ DELAY((x) * 1000); \ else \ pause("t4hw", (x) * hz / 1000); \ } while (0) /* * Wait for the device to become ready (signified by our "who am I" register * returning a value other than all 1's). Return an error if it doesn't * become ready ... */ int t4vf_wait_dev_ready(struct adapter *adapter) { const u32 whoami = VF_PL_REG(A_PL_VF_WHOAMI); const u32 notready1 = 0xffffffff; const u32 notready2 = 0xeeeeeeee; u32 val; val = t4_read_reg(adapter, whoami); if (val != notready1 && val != notready2) return 0; msleep(500); val = t4_read_reg(adapter, whoami); if (val != notready1 && val != notready2) return 0; else return -EIO; } /** * t4vf_fw_reset - issue a reset to FW * @adapter: the adapter * * Issues a reset command to FW. For a Physical Function this would * result in the Firmware reseting all of its state. For a Virtual * Function this just resets the state associated with the VF. */ int t4vf_fw_reset(struct adapter *adapter) { struct fw_reset_cmd cmd; memset(&cmd, 0, sizeof(cmd)); cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_RESET_CMD) | F_FW_CMD_WRITE); cmd.retval_len16 = cpu_to_be32(V_FW_CMD_LEN16(FW_LEN16(cmd))); return t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), NULL); } /** * t4vf_get_sge_params - retrieve adapter Scatter gather Engine parameters * @adapter: the adapter * * Retrieves various core SGE parameters in the form of hardware SGE * register values. The caller is responsible for decoding these as * needed. The SGE parameters are stored in @adapter->params.sge. */ int t4vf_get_sge_params(struct adapter *adapter) { struct sge_params *sp = &adapter->params.sge; u32 params[7], vals[7]; u32 whoami; unsigned int pf, s_hps; int i, v; params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_CONTROL)); params[1] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_HOST_PAGE_SIZE)); params[2] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_TIMER_VALUE_0_AND_1)); params[3] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_TIMER_VALUE_2_AND_3)); params[4] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_TIMER_VALUE_4_AND_5)); params[5] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_CONM_CTRL)); params[6] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_INGRESS_RX_THRESHOLD)); v = t4vf_query_params(adapter, 7, params, vals); if (v != FW_SUCCESS) return v; sp->sge_control = vals[0]; sp->counter_val[0] = G_THRESHOLD_0(vals[6]); sp->counter_val[1] = G_THRESHOLD_1(vals[6]); sp->counter_val[2] = G_THRESHOLD_2(vals[6]); sp->counter_val[3] = G_THRESHOLD_3(vals[6]); sp->timer_val[0] = core_ticks_to_us(adapter, G_TIMERVALUE0(vals[2])); sp->timer_val[1] = core_ticks_to_us(adapter, G_TIMERVALUE1(vals[2])); sp->timer_val[2] = core_ticks_to_us(adapter, G_TIMERVALUE2(vals[3])); sp->timer_val[3] = core_ticks_to_us(adapter, G_TIMERVALUE3(vals[3])); sp->timer_val[4] = core_ticks_to_us(adapter, G_TIMERVALUE4(vals[4])); sp->timer_val[5] = core_ticks_to_us(adapter, G_TIMERVALUE5(vals[4])); sp->fl_starve_threshold = G_EGRTHRESHOLD(vals[5]) * 2 + 1; if (is_t4(adapter)) sp->fl_starve_threshold2 = sp->fl_starve_threshold; + else if (is_t5(adapter)) + sp->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(vals[5]) * 2 + 1; else - sp->fl_starve_threshold2 = G_EGRTHRESHOLDPACKING(vals[5]) * 2 + - 1; + sp->fl_starve_threshold2 = G_T6_EGRTHRESHOLDPACKING(vals[5]) * 2 + 1; /* * We need the Queues/Page and Host Page Size for our VF. * This is based on the PF from which we're instantiated. */ whoami = t4_read_reg(adapter, VF_PL_REG(A_PL_VF_WHOAMI)); pf = G_SOURCEPF(whoami); s_hps = (S_HOSTPAGESIZEPF0 + (S_HOSTPAGESIZEPF1 - S_HOSTPAGESIZEPF0) * pf); sp->page_shift = ((vals[1] >> s_hps) & M_HOSTPAGESIZEPF0) + 10; for (i = 0; i < SGE_FLBUF_SIZES; i++) { params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_FL_BUFFER_SIZE0 + (4 * i))); v = t4vf_query_params(adapter, 1, params, vals); if (v != FW_SUCCESS) return v; sp->sge_fl_buffer_size[i] = vals[0]; } /* * T4 uses a single control field to specify both the PCIe Padding and * Packing Boundary. T5 introduced the ability to specify these * separately with the Padding Boundary in SGE_CONTROL and and Packing * Boundary in SGE_CONTROL2. So for T5 and later we need to grab * SGE_CONTROL in order to determine how ingress packet data will be * laid out in Packed Buffer Mode. Unfortunately, older versions of * the firmware won't let us retrieve SGE_CONTROL2 so if we get a * failure grabbing it we throw an error since we can't figure out the * right value. */ sp->spg_len = sp->sge_control & F_EGRSTATUSPAGESIZE ? 128 : 64; sp->fl_pktshift = G_PKTSHIFT(sp->sge_control); - sp->pad_boundary = 1 << (G_INGPADBOUNDARY(sp->sge_control) + 5); + if (chip_id(adapter) <= CHELSIO_T5) { + sp->pad_boundary = 1 << (G_INGPADBOUNDARY(sp->sge_control) + + X_INGPADBOUNDARY_SHIFT); + } else { + sp->pad_boundary = 1 << (G_INGPADBOUNDARY(sp->sge_control) + + X_T6_INGPADBOUNDARY_SHIFT); + } if (is_t4(adapter)) sp->pack_boundary = sp->pad_boundary; else { params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_CONTROL2)); v = t4vf_query_params(adapter, 1, params, vals); if (v != FW_SUCCESS) { CH_ERR(adapter, "Unable to get SGE Control2; " "probably old firmware.\n"); return v; } if (G_INGPACKBOUNDARY(vals[0]) == 0) sp->pack_boundary = 16; else sp->pack_boundary = 1 << (G_INGPACKBOUNDARY(vals[0]) + 5); } /* * For T5 and later we want to use the new BAR2 Doorbells. * Unfortunately, older firmware didn't allow the this register to be * read. */ if (!is_t4(adapter)) { unsigned int s_qpp; params[0] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_EGRESS_QUEUES_PER_PAGE_VF)); params[1] = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_REG) | V_FW_PARAMS_PARAM_XYZ(A_SGE_INGRESS_QUEUES_PER_PAGE_VF)); v = t4vf_query_params(adapter, 2, params, vals); if (v != FW_SUCCESS) { CH_WARN(adapter, "Unable to get VF SGE Queues/Page; " "probably old firmware.\n"); return v; } s_qpp = (S_QUEUESPERPAGEPF0 + (S_QUEUESPERPAGEPF1 - S_QUEUESPERPAGEPF0) * pf); sp->eq_s_qpp = ((vals[0] >> s_qpp) & M_QUEUESPERPAGEPF0); sp->iq_s_qpp = ((vals[1] >> s_qpp) & M_QUEUESPERPAGEPF0); } return 0; } /** * t4vf_get_rss_glb_config - retrieve adapter RSS Global Configuration * @adapter: the adapter * * Retrieves global RSS mode and parameters with which we have to live * and stores them in the @adapter's RSS parameters. */ int t4vf_get_rss_glb_config(struct adapter *adapter) { struct rss_params *rss = &adapter->params.rss; struct fw_rss_glb_config_cmd cmd, rpl; int v; /* * Execute an RSS Global Configuration read command to retrieve * our RSS configuration. */ memset(&cmd, 0, sizeof(cmd)); cmd.op_to_write = cpu_to_be32(V_FW_CMD_OP(FW_RSS_GLB_CONFIG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd)); v = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl); if (v != FW_SUCCESS) return v; /* * Transate the big-endian RSS Global Configuration into our * cpu-endian format based on the RSS mode. We also do first level * filtering at this point to weed out modes which don't support * VF Drivers ... */ rss->mode = G_FW_RSS_GLB_CONFIG_CMD_MODE( be32_to_cpu(rpl.u.manual.mode_pkd)); switch (rss->mode) { case FW_RSS_GLB_CONFIG_CMD_MODE_BASICVIRTUAL: { u32 word = be32_to_cpu( rpl.u.basicvirtual.synmapen_to_hashtoeplitz); rss->u.basicvirtual.synmapen = ((word & F_FW_RSS_GLB_CONFIG_CMD_SYNMAPEN) != 0); rss->u.basicvirtual.syn4tupenipv6 = ((word & F_FW_RSS_GLB_CONFIG_CMD_SYN4TUPENIPV6) != 0); rss->u.basicvirtual.syn2tupenipv6 = ((word & F_FW_RSS_GLB_CONFIG_CMD_SYN2TUPENIPV6) != 0); rss->u.basicvirtual.syn4tupenipv4 = ((word & F_FW_RSS_GLB_CONFIG_CMD_SYN4TUPENIPV4) != 0); rss->u.basicvirtual.syn2tupenipv4 = ((word & F_FW_RSS_GLB_CONFIG_CMD_SYN2TUPENIPV4) != 0); rss->u.basicvirtual.ofdmapen = ((word & F_FW_RSS_GLB_CONFIG_CMD_OFDMAPEN) != 0); rss->u.basicvirtual.tnlmapen = ((word & F_FW_RSS_GLB_CONFIG_CMD_TNLMAPEN) != 0); rss->u.basicvirtual.tnlalllookup = ((word & F_FW_RSS_GLB_CONFIG_CMD_TNLALLLKP) != 0); rss->u.basicvirtual.hashtoeplitz = ((word & F_FW_RSS_GLB_CONFIG_CMD_HASHTOEPLITZ) != 0); /* we need at least Tunnel Map Enable to be set */ if (!rss->u.basicvirtual.tnlmapen) return -EINVAL; break; } default: /* all unknown/unsupported RSS modes result in an error */ return -EINVAL; } return 0; } /** * t4vf_get_vfres - retrieve VF resource limits * @adapter: the adapter * * Retrieves configured resource limits and capabilities for a virtual * function. The results are stored in @adapter->vfres. */ int t4vf_get_vfres(struct adapter *adapter) { struct vf_resources *vfres = &adapter->params.vfres; struct fw_pfvf_cmd cmd, rpl; int v; u32 word; /* * Execute PFVF Read command to get VF resource limits; bail out early * with error on command failure. */ memset(&cmd, 0, sizeof(cmd)); cmd.op_to_vfn = cpu_to_be32(V_FW_CMD_OP(FW_PFVF_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); cmd.retval_len16 = cpu_to_be32(FW_LEN16(cmd)); v = t4vf_wr_mbox(adapter, &cmd, sizeof(cmd), &rpl); if (v != FW_SUCCESS) return v; /* * Extract VF resource limits and return success. */ word = be32_to_cpu(rpl.niqflint_niq); vfres->niqflint = G_FW_PFVF_CMD_NIQFLINT(word); vfres->niq = G_FW_PFVF_CMD_NIQ(word); word = be32_to_cpu(rpl.type_to_neq); vfres->neq = G_FW_PFVF_CMD_NEQ(word); vfres->pmask = G_FW_PFVF_CMD_PMASK(word); word = be32_to_cpu(rpl.tc_to_nexactf); vfres->tc = G_FW_PFVF_CMD_TC(word); vfres->nvi = G_FW_PFVF_CMD_NVI(word); vfres->nexactf = G_FW_PFVF_CMD_NEXACTF(word); word = be32_to_cpu(rpl.r_caps_to_nethctrl); vfres->r_caps = G_FW_PFVF_CMD_R_CAPS(word); vfres->wx_caps = G_FW_PFVF_CMD_WX_CAPS(word); vfres->nethctrl = G_FW_PFVF_CMD_NETHCTRL(word); return 0; } /** */ int t4vf_prep_adapter(struct adapter *adapter) { int err; /* * Wait for the device to become ready before proceeding ... */ err = t4vf_wait_dev_ready(adapter); if (err) return err; adapter->params.chipid = pci_get_device(adapter->dev) >> 12; if (adapter->params.chipid >= 0xa) { adapter->params.chipid -= (0xa - 0x4); adapter->params.fpga = 1; } /* * Default port and clock for debugging in case we can't reach * firmware. */ adapter->params.nports = 1; adapter->params.vfres.pmask = 1; adapter->params.vpd.cclk = 50000; adapter->chip_params = t4_get_chip_params(chip_id(adapter)); if (adapter->chip_params == NULL) return -EINVAL; return 0; } Index: projects/mips64-clang/sys/dev/cxgbe/t4_main.c =================================================================== --- projects/mips64-clang/sys/dev/cxgbe/t4_main.c (revision 305711) +++ projects/mips64-clang/sys/dev/cxgbe/t4_main.c (revision 305712) @@ -1,9575 +1,9585 @@ /*- * Copyright (c) 2011 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_inet.h" #include "opt_inet6.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef RSS #include #endif #if defined(__i386__) || defined(__amd64__) #include #include #endif #ifdef DDB #include #include #endif #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "common/t4_regs_values.h" #include "t4_ioctl.h" #include "t4_l2t.h" #include "t4_mp_ring.h" #include "t4_if.h" /* T4 bus driver interface */ static int t4_probe(device_t); static int t4_attach(device_t); static int t4_detach(device_t); static int t4_ready(device_t); static int t4_read_port_device(device_t, int, device_t *); static device_method_t t4_methods[] = { DEVMETHOD(device_probe, t4_probe), DEVMETHOD(device_attach, t4_attach), DEVMETHOD(device_detach, t4_detach), DEVMETHOD(t4_is_main_ready, t4_ready), DEVMETHOD(t4_read_port_device, t4_read_port_device), DEVMETHOD_END }; static driver_t t4_driver = { "t4nex", t4_methods, sizeof(struct adapter) }; /* T4 port (cxgbe) interface */ static int cxgbe_probe(device_t); static int cxgbe_attach(device_t); static int cxgbe_detach(device_t); device_method_t cxgbe_methods[] = { DEVMETHOD(device_probe, cxgbe_probe), DEVMETHOD(device_attach, cxgbe_attach), DEVMETHOD(device_detach, cxgbe_detach), { 0, 0 } }; static driver_t cxgbe_driver = { "cxgbe", cxgbe_methods, sizeof(struct port_info) }; /* T4 VI (vcxgbe) interface */ static int vcxgbe_probe(device_t); static int vcxgbe_attach(device_t); static int vcxgbe_detach(device_t); static device_method_t vcxgbe_methods[] = { DEVMETHOD(device_probe, vcxgbe_probe), DEVMETHOD(device_attach, vcxgbe_attach), DEVMETHOD(device_detach, vcxgbe_detach), { 0, 0 } }; static driver_t vcxgbe_driver = { "vcxgbe", vcxgbe_methods, sizeof(struct vi_info) }; static d_ioctl_t t4_ioctl; static struct cdevsw t4_cdevsw = { .d_version = D_VERSION, .d_ioctl = t4_ioctl, .d_name = "t4nex", }; /* T5 bus driver interface */ static int t5_probe(device_t); static device_method_t t5_methods[] = { DEVMETHOD(device_probe, t5_probe), DEVMETHOD(device_attach, t4_attach), DEVMETHOD(device_detach, t4_detach), DEVMETHOD(t4_is_main_ready, t4_ready), DEVMETHOD(t4_read_port_device, t4_read_port_device), DEVMETHOD_END }; static driver_t t5_driver = { "t5nex", t5_methods, sizeof(struct adapter) }; /* T5 port (cxl) interface */ static driver_t cxl_driver = { "cxl", cxgbe_methods, sizeof(struct port_info) }; /* T5 VI (vcxl) interface */ static driver_t vcxl_driver = { "vcxl", vcxgbe_methods, sizeof(struct vi_info) }; /* ifnet + media interface */ static void cxgbe_init(void *); static int cxgbe_ioctl(struct ifnet *, unsigned long, caddr_t); static int cxgbe_transmit(struct ifnet *, struct mbuf *); static void cxgbe_qflush(struct ifnet *); static int cxgbe_media_change(struct ifnet *); static void cxgbe_media_status(struct ifnet *, struct ifmediareq *); MALLOC_DEFINE(M_CXGBE, "cxgbe", "Chelsio T4/T5 Ethernet driver and services"); /* * Correct lock order when you need to acquire multiple locks is t4_list_lock, * then ADAPTER_LOCK, then t4_uld_list_lock. */ static struct sx t4_list_lock; SLIST_HEAD(, adapter) t4_list; #ifdef TCP_OFFLOAD static struct sx t4_uld_list_lock; SLIST_HEAD(, uld_info) t4_uld_list; #endif /* * Tunables. See tweak_tunables() too. * * Each tunable is set to a default value here if it's known at compile-time. * Otherwise it is set to -1 as an indication to tweak_tunables() that it should * provide a reasonable default when the driver is loaded. * * Tunables applicable to both T4 and T5 are under hw.cxgbe. Those specific to * T5 are under hw.cxl. */ /* * Number of queues for tx and rx, 10G and 1G, NIC and offload. */ #define NTXQ_10G 16 int t4_ntxq10g = -1; TUNABLE_INT("hw.cxgbe.ntxq10g", &t4_ntxq10g); #define NRXQ_10G 8 int t4_nrxq10g = -1; TUNABLE_INT("hw.cxgbe.nrxq10g", &t4_nrxq10g); #define NTXQ_1G 4 int t4_ntxq1g = -1; TUNABLE_INT("hw.cxgbe.ntxq1g", &t4_ntxq1g); #define NRXQ_1G 2 int t4_nrxq1g = -1; TUNABLE_INT("hw.cxgbe.nrxq1g", &t4_nrxq1g); #define NTXQ_VI 1 static int t4_ntxq_vi = -1; TUNABLE_INT("hw.cxgbe.ntxq_vi", &t4_ntxq_vi); #define NRXQ_VI 1 static int t4_nrxq_vi = -1; TUNABLE_INT("hw.cxgbe.nrxq_vi", &t4_nrxq_vi); static int t4_rsrv_noflowq = 0; TUNABLE_INT("hw.cxgbe.rsrv_noflowq", &t4_rsrv_noflowq); #ifdef TCP_OFFLOAD #define NOFLDTXQ_10G 8 static int t4_nofldtxq10g = -1; TUNABLE_INT("hw.cxgbe.nofldtxq10g", &t4_nofldtxq10g); #define NOFLDRXQ_10G 2 static int t4_nofldrxq10g = -1; TUNABLE_INT("hw.cxgbe.nofldrxq10g", &t4_nofldrxq10g); #define NOFLDTXQ_1G 2 static int t4_nofldtxq1g = -1; TUNABLE_INT("hw.cxgbe.nofldtxq1g", &t4_nofldtxq1g); #define NOFLDRXQ_1G 1 static int t4_nofldrxq1g = -1; TUNABLE_INT("hw.cxgbe.nofldrxq1g", &t4_nofldrxq1g); #define NOFLDTXQ_VI 1 static int t4_nofldtxq_vi = -1; TUNABLE_INT("hw.cxgbe.nofldtxq_vi", &t4_nofldtxq_vi); #define NOFLDRXQ_VI 1 static int t4_nofldrxq_vi = -1; TUNABLE_INT("hw.cxgbe.nofldrxq_vi", &t4_nofldrxq_vi); #endif #ifdef DEV_NETMAP #define NNMTXQ_VI 2 static int t4_nnmtxq_vi = -1; TUNABLE_INT("hw.cxgbe.nnmtxq_vi", &t4_nnmtxq_vi); #define NNMRXQ_VI 2 static int t4_nnmrxq_vi = -1; TUNABLE_INT("hw.cxgbe.nnmrxq_vi", &t4_nnmrxq_vi); #endif /* * Holdoff parameters for 10G and 1G ports. */ #define TMR_IDX_10G 1 int t4_tmr_idx_10g = TMR_IDX_10G; TUNABLE_INT("hw.cxgbe.holdoff_timer_idx_10G", &t4_tmr_idx_10g); #define PKTC_IDX_10G (-1) int t4_pktc_idx_10g = PKTC_IDX_10G; TUNABLE_INT("hw.cxgbe.holdoff_pktc_idx_10G", &t4_pktc_idx_10g); #define TMR_IDX_1G 1 int t4_tmr_idx_1g = TMR_IDX_1G; TUNABLE_INT("hw.cxgbe.holdoff_timer_idx_1G", &t4_tmr_idx_1g); #define PKTC_IDX_1G (-1) int t4_pktc_idx_1g = PKTC_IDX_1G; TUNABLE_INT("hw.cxgbe.holdoff_pktc_idx_1G", &t4_pktc_idx_1g); /* * Size (# of entries) of each tx and rx queue. */ unsigned int t4_qsize_txq = TX_EQ_QSIZE; TUNABLE_INT("hw.cxgbe.qsize_txq", &t4_qsize_txq); unsigned int t4_qsize_rxq = RX_IQ_QSIZE; TUNABLE_INT("hw.cxgbe.qsize_rxq", &t4_qsize_rxq); /* * Interrupt types allowed (bits 0, 1, 2 = INTx, MSI, MSI-X respectively). */ int t4_intr_types = INTR_MSIX | INTR_MSI | INTR_INTX; TUNABLE_INT("hw.cxgbe.interrupt_types", &t4_intr_types); /* * Configuration file. */ #define DEFAULT_CF "default" #define FLASH_CF "flash" #define UWIRE_CF "uwire" #define FPGA_CF "fpga" static char t4_cfg_file[32] = DEFAULT_CF; TUNABLE_STR("hw.cxgbe.config_file", t4_cfg_file, sizeof(t4_cfg_file)); /* * PAUSE settings (bit 0, 1 = rx_pause, tx_pause respectively). * rx_pause = 1 to heed incoming PAUSE frames, 0 to ignore them. * tx_pause = 1 to emit PAUSE frames when the rx FIFO reaches its high water * mark or when signalled to do so, 0 to never emit PAUSE. */ static int t4_pause_settings = PAUSE_TX | PAUSE_RX; TUNABLE_INT("hw.cxgbe.pause_settings", &t4_pause_settings); /* * Firmware auto-install by driver during attach (0, 1, 2 = prohibited, allowed, * encouraged respectively). */ static unsigned int t4_fw_install = 1; TUNABLE_INT("hw.cxgbe.fw_install", &t4_fw_install); /* * ASIC features that will be used. Disable the ones you don't want so that the * chip resources aren't wasted on features that will not be used. */ static int t4_nbmcaps_allowed = 0; TUNABLE_INT("hw.cxgbe.nbmcaps_allowed", &t4_nbmcaps_allowed); static int t4_linkcaps_allowed = 0; /* No DCBX, PPP, etc. by default */ TUNABLE_INT("hw.cxgbe.linkcaps_allowed", &t4_linkcaps_allowed); static int t4_switchcaps_allowed = FW_CAPS_CONFIG_SWITCH_INGRESS | FW_CAPS_CONFIG_SWITCH_EGRESS; TUNABLE_INT("hw.cxgbe.switchcaps_allowed", &t4_switchcaps_allowed); static int t4_niccaps_allowed = FW_CAPS_CONFIG_NIC; TUNABLE_INT("hw.cxgbe.niccaps_allowed", &t4_niccaps_allowed); static int t4_toecaps_allowed = -1; TUNABLE_INT("hw.cxgbe.toecaps_allowed", &t4_toecaps_allowed); static int t4_rdmacaps_allowed = -1; TUNABLE_INT("hw.cxgbe.rdmacaps_allowed", &t4_rdmacaps_allowed); static int t4_tlscaps_allowed = 0; TUNABLE_INT("hw.cxgbe.tlscaps_allowed", &t4_tlscaps_allowed); static int t4_iscsicaps_allowed = -1; TUNABLE_INT("hw.cxgbe.iscsicaps_allowed", &t4_iscsicaps_allowed); static int t4_fcoecaps_allowed = 0; TUNABLE_INT("hw.cxgbe.fcoecaps_allowed", &t4_fcoecaps_allowed); static int t5_write_combine = 0; TUNABLE_INT("hw.cxl.write_combine", &t5_write_combine); static int t4_num_vis = 1; TUNABLE_INT("hw.cxgbe.num_vis", &t4_num_vis); /* Functions used by extra VIs to obtain unique MAC addresses for each VI. */ static int vi_mac_funcs[] = { FW_VI_FUNC_OFLD, FW_VI_FUNC_IWARP, FW_VI_FUNC_OPENISCSI, FW_VI_FUNC_OPENFCOE, FW_VI_FUNC_FOISCSI, FW_VI_FUNC_FOFCOE, }; struct intrs_and_queues { uint16_t intr_type; /* INTx, MSI, or MSI-X */ uint16_t nirq; /* Total # of vectors */ uint16_t intr_flags_10g;/* Interrupt flags for each 10G port */ uint16_t intr_flags_1g; /* Interrupt flags for each 1G port */ uint16_t ntxq10g; /* # of NIC txq's for each 10G port */ uint16_t nrxq10g; /* # of NIC rxq's for each 10G port */ uint16_t ntxq1g; /* # of NIC txq's for each 1G port */ uint16_t nrxq1g; /* # of NIC rxq's for each 1G port */ uint16_t rsrv_noflowq; /* Flag whether to reserve queue 0 */ uint16_t nofldtxq10g; /* # of TOE txq's for each 10G port */ uint16_t nofldrxq10g; /* # of TOE rxq's for each 10G port */ uint16_t nofldtxq1g; /* # of TOE txq's for each 1G port */ uint16_t nofldrxq1g; /* # of TOE rxq's for each 1G port */ /* The vcxgbe/vcxl interfaces use these and not the ones above. */ uint16_t ntxq_vi; /* # of NIC txq's */ uint16_t nrxq_vi; /* # of NIC rxq's */ uint16_t nofldtxq_vi; /* # of TOE txq's */ uint16_t nofldrxq_vi; /* # of TOE rxq's */ uint16_t nnmtxq_vi; /* # of netmap txq's */ uint16_t nnmrxq_vi; /* # of netmap rxq's */ }; struct filter_entry { uint32_t valid:1; /* filter allocated and valid */ uint32_t locked:1; /* filter is administratively locked */ uint32_t pending:1; /* filter action is pending firmware reply */ uint32_t smtidx:8; /* Source MAC Table index for smac */ struct l2t_entry *l2t; /* Layer Two Table entry for dmac */ struct t4_filter_specification fs; }; static void setup_memwin(struct adapter *); static void position_memwin(struct adapter *, int, uint32_t); static int rw_via_memwin(struct adapter *, int, uint32_t, uint32_t *, int, int); static inline int read_via_memwin(struct adapter *, int, uint32_t, uint32_t *, int); static inline int write_via_memwin(struct adapter *, int, uint32_t, const uint32_t *, int); static int validate_mem_range(struct adapter *, uint32_t, int); static int fwmtype_to_hwmtype(int); static int validate_mt_off_len(struct adapter *, int, uint32_t, int, uint32_t *); static int fixup_devlog_params(struct adapter *); static int cfg_itype_and_nqueues(struct adapter *, int, int, int, struct intrs_and_queues *); static int prep_firmware(struct adapter *); static int partition_resources(struct adapter *, const struct firmware *, const char *); static int get_params__pre_init(struct adapter *); static int get_params__post_init(struct adapter *); static int set_params__post_init(struct adapter *); static void t4_set_desc(struct adapter *); static void build_medialist(struct port_info *, struct ifmedia *); static int cxgbe_init_synchronized(struct vi_info *); static int cxgbe_uninit_synchronized(struct vi_info *); static void quiesce_txq(struct adapter *, struct sge_txq *); static void quiesce_wrq(struct adapter *, struct sge_wrq *); static void quiesce_iq(struct adapter *, struct sge_iq *); static void quiesce_fl(struct adapter *, struct sge_fl *); static int t4_alloc_irq(struct adapter *, struct irq *, int rid, driver_intr_t *, void *, char *); static int t4_free_irq(struct adapter *, struct irq *); static void get_regs(struct adapter *, struct t4_regdump *, uint8_t *); static void vi_refresh_stats(struct adapter *, struct vi_info *); static void cxgbe_refresh_stats(struct adapter *, struct port_info *); static void cxgbe_tick(void *); static void cxgbe_vlan_config(void *, struct ifnet *, uint16_t); static void cxgbe_sysctls(struct port_info *); static int sysctl_int_array(SYSCTL_HANDLER_ARGS); static int sysctl_bitfield(SYSCTL_HANDLER_ARGS); static int sysctl_btphy(SYSCTL_HANDLER_ARGS); static int sysctl_noflowq(SYSCTL_HANDLER_ARGS); static int sysctl_holdoff_tmr_idx(SYSCTL_HANDLER_ARGS); static int sysctl_holdoff_pktc_idx(SYSCTL_HANDLER_ARGS); static int sysctl_qsize_rxq(SYSCTL_HANDLER_ARGS); static int sysctl_qsize_txq(SYSCTL_HANDLER_ARGS); static int sysctl_pause_settings(SYSCTL_HANDLER_ARGS); static int sysctl_handle_t4_reg64(SYSCTL_HANDLER_ARGS); static int sysctl_temperature(SYSCTL_HANDLER_ARGS); #ifdef SBUF_DRAIN static int sysctl_cctrl(SYSCTL_HANDLER_ARGS); static int sysctl_cim_ibq_obq(SYSCTL_HANDLER_ARGS); static int sysctl_cim_la(SYSCTL_HANDLER_ARGS); static int sysctl_cim_la_t6(SYSCTL_HANDLER_ARGS); static int sysctl_cim_ma_la(SYSCTL_HANDLER_ARGS); static int sysctl_cim_pif_la(SYSCTL_HANDLER_ARGS); static int sysctl_cim_qcfg(SYSCTL_HANDLER_ARGS); static int sysctl_cpl_stats(SYSCTL_HANDLER_ARGS); static int sysctl_ddp_stats(SYSCTL_HANDLER_ARGS); static int sysctl_devlog(SYSCTL_HANDLER_ARGS); static int sysctl_fcoe_stats(SYSCTL_HANDLER_ARGS); static int sysctl_hw_sched(SYSCTL_HANDLER_ARGS); static int sysctl_lb_stats(SYSCTL_HANDLER_ARGS); static int sysctl_linkdnrc(SYSCTL_HANDLER_ARGS); static int sysctl_meminfo(SYSCTL_HANDLER_ARGS); static int sysctl_mps_tcam(SYSCTL_HANDLER_ARGS); static int sysctl_mps_tcam_t6(SYSCTL_HANDLER_ARGS); static int sysctl_path_mtus(SYSCTL_HANDLER_ARGS); static int sysctl_pm_stats(SYSCTL_HANDLER_ARGS); static int sysctl_rdma_stats(SYSCTL_HANDLER_ARGS); static int sysctl_tcp_stats(SYSCTL_HANDLER_ARGS); static int sysctl_tids(SYSCTL_HANDLER_ARGS); static int sysctl_tp_err_stats(SYSCTL_HANDLER_ARGS); static int sysctl_tp_la_mask(SYSCTL_HANDLER_ARGS); static int sysctl_tp_la(SYSCTL_HANDLER_ARGS); static int sysctl_tx_rate(SYSCTL_HANDLER_ARGS); static int sysctl_ulprx_la(SYSCTL_HANDLER_ARGS); static int sysctl_wcwr_stats(SYSCTL_HANDLER_ARGS); static int sysctl_tc_params(SYSCTL_HANDLER_ARGS); #endif #ifdef TCP_OFFLOAD static int sysctl_tp_tick(SYSCTL_HANDLER_ARGS); static int sysctl_tp_dack_timer(SYSCTL_HANDLER_ARGS); static int sysctl_tp_timer(SYSCTL_HANDLER_ARGS); #endif static uint32_t fconf_iconf_to_mode(uint32_t, uint32_t); static uint32_t mode_to_fconf(uint32_t); static uint32_t mode_to_iconf(uint32_t); static int check_fspec_against_fconf_iconf(struct adapter *, struct t4_filter_specification *); static int get_filter_mode(struct adapter *, uint32_t *); static int set_filter_mode(struct adapter *, uint32_t); static inline uint64_t get_filter_hits(struct adapter *, uint32_t); static int get_filter(struct adapter *, struct t4_filter *); static int set_filter(struct adapter *, struct t4_filter *); static int del_filter(struct adapter *, struct t4_filter *); static void clear_filter(struct filter_entry *); static int set_filter_wr(struct adapter *, int); static int del_filter_wr(struct adapter *, int); static int set_tcb_rpl(struct sge_iq *, const struct rss_header *, struct mbuf *); static int get_sge_context(struct adapter *, struct t4_sge_context *); static int load_fw(struct adapter *, struct t4_data *); static int read_card_mem(struct adapter *, int, struct t4_mem_range *); static int read_i2c(struct adapter *, struct t4_i2c_data *); #ifdef TCP_OFFLOAD static int toe_capability(struct vi_info *, int); #endif static int mod_event(module_t, int, void *); static int notify_siblings(device_t, int); struct { uint16_t device; char *desc; } t4_pciids[] = { {0xa000, "Chelsio Terminator 4 FPGA"}, {0x4400, "Chelsio T440-dbg"}, {0x4401, "Chelsio T420-CR"}, {0x4402, "Chelsio T422-CR"}, {0x4403, "Chelsio T440-CR"}, {0x4404, "Chelsio T420-BCH"}, {0x4405, "Chelsio T440-BCH"}, {0x4406, "Chelsio T440-CH"}, {0x4407, "Chelsio T420-SO"}, {0x4408, "Chelsio T420-CX"}, {0x4409, "Chelsio T420-BT"}, {0x440a, "Chelsio T404-BT"}, {0x440e, "Chelsio T440-LP-CR"}, }, t5_pciids[] = { {0xb000, "Chelsio Terminator 5 FPGA"}, {0x5400, "Chelsio T580-dbg"}, {0x5401, "Chelsio T520-CR"}, /* 2 x 10G */ {0x5402, "Chelsio T522-CR"}, /* 2 x 10G, 2 X 1G */ {0x5403, "Chelsio T540-CR"}, /* 4 x 10G */ {0x5407, "Chelsio T520-SO"}, /* 2 x 10G, nomem */ {0x5409, "Chelsio T520-BT"}, /* 2 x 10GBaseT */ {0x540a, "Chelsio T504-BT"}, /* 4 x 1G */ {0x540d, "Chelsio T580-CR"}, /* 2 x 40G */ {0x540e, "Chelsio T540-LP-CR"}, /* 4 x 10G */ {0x5410, "Chelsio T580-LP-CR"}, /* 2 x 40G */ {0x5411, "Chelsio T520-LL-CR"}, /* 2 x 10G */ {0x5412, "Chelsio T560-CR"}, /* 1 x 40G, 2 x 10G */ {0x5414, "Chelsio T580-LP-SO-CR"}, /* 2 x 40G, nomem */ {0x5415, "Chelsio T502-BT"}, /* 2 x 1G */ #ifdef notyet {0x5404, "Chelsio T520-BCH"}, {0x5405, "Chelsio T540-BCH"}, {0x5406, "Chelsio T540-CH"}, {0x5408, "Chelsio T520-CX"}, {0x540b, "Chelsio B520-SR"}, {0x540c, "Chelsio B504-BT"}, {0x540f, "Chelsio Amsterdam"}, {0x5413, "Chelsio T580-CHR"}, #endif }; #ifdef TCP_OFFLOAD /* * service_iq() has an iq and needs the fl. Offset of fl from the iq should be * exactly the same for both rxq and ofld_rxq. */ CTASSERT(offsetof(struct sge_ofld_rxq, iq) == offsetof(struct sge_rxq, iq)); CTASSERT(offsetof(struct sge_ofld_rxq, fl) == offsetof(struct sge_rxq, fl)); #endif CTASSERT(sizeof(struct cluster_metadata) <= CL_METADATA_SIZE); static int t4_probe(device_t dev) { int i; uint16_t v = pci_get_vendor(dev); uint16_t d = pci_get_device(dev); uint8_t f = pci_get_function(dev); if (v != PCI_VENDOR_ID_CHELSIO) return (ENXIO); /* Attach only to PF0 of the FPGA */ if (d == 0xa000 && f != 0) return (ENXIO); for (i = 0; i < nitems(t4_pciids); i++) { if (d == t4_pciids[i].device) { device_set_desc(dev, t4_pciids[i].desc); return (BUS_PROBE_DEFAULT); } } return (ENXIO); } static int t5_probe(device_t dev) { int i; uint16_t v = pci_get_vendor(dev); uint16_t d = pci_get_device(dev); uint8_t f = pci_get_function(dev); if (v != PCI_VENDOR_ID_CHELSIO) return (ENXIO); /* Attach only to PF0 of the FPGA */ if (d == 0xb000 && f != 0) return (ENXIO); for (i = 0; i < nitems(t5_pciids); i++) { if (d == t5_pciids[i].device) { device_set_desc(dev, t5_pciids[i].desc); return (BUS_PROBE_DEFAULT); } } return (ENXIO); } static void t5_attribute_workaround(device_t dev) { device_t root_port; uint32_t v; /* * The T5 chips do not properly echo the No Snoop and Relaxed * Ordering attributes when replying to a TLP from a Root * Port. As a workaround, find the parent Root Port and * disable No Snoop and Relaxed Ordering. Note that this * affects all devices under this root port. */ root_port = pci_find_pcie_root_port(dev); if (root_port == NULL) { device_printf(dev, "Unable to find parent root port\n"); return; } v = pcie_adjust_config(root_port, PCIER_DEVICE_CTL, PCIEM_CTL_RELAXED_ORD_ENABLE | PCIEM_CTL_NOSNOOP_ENABLE, 0, 2); if ((v & (PCIEM_CTL_RELAXED_ORD_ENABLE | PCIEM_CTL_NOSNOOP_ENABLE)) != 0) device_printf(dev, "Disabled No Snoop/Relaxed Ordering on %s\n", device_get_nameunit(root_port)); } static int t4_attach(device_t dev) { struct adapter *sc; int rc = 0, i, j, n10g, n1g, rqidx, tqidx; struct make_dev_args mda; struct intrs_and_queues iaq; struct sge *s; uint8_t *buf; #ifdef TCP_OFFLOAD int ofld_rqidx, ofld_tqidx; #endif #ifdef DEV_NETMAP int nm_rqidx, nm_tqidx; #endif int num_vis; sc = device_get_softc(dev); sc->dev = dev; - TUNABLE_INT_FETCH("hw.cxgbe.debug_flags", &sc->debug_flags); + TUNABLE_INT_FETCH("hw.cxgbe.dflags", &sc->debug_flags); if ((pci_get_device(dev) & 0xff00) == 0x5400) t5_attribute_workaround(dev); pci_enable_busmaster(dev); if (pci_find_cap(dev, PCIY_EXPRESS, &i) == 0) { uint32_t v; pci_set_max_read_req(dev, 4096); v = pci_read_config(dev, i + PCIER_DEVICE_CTL, 2); v |= PCIEM_CTL_RELAXED_ORD_ENABLE; pci_write_config(dev, i + PCIER_DEVICE_CTL, v, 2); sc->params.pci.mps = 128 << ((v & PCIEM_CTL_MAX_PAYLOAD) >> 5); } sc->sge_gts_reg = MYPF_REG(A_SGE_PF_GTS); sc->sge_kdoorbell_reg = MYPF_REG(A_SGE_PF_KDOORBELL); sc->traceq = -1; mtx_init(&sc->ifp_lock, sc->ifp_lockname, 0, MTX_DEF); snprintf(sc->ifp_lockname, sizeof(sc->ifp_lockname), "%s tracer", device_get_nameunit(dev)); snprintf(sc->lockname, sizeof(sc->lockname), "%s", device_get_nameunit(dev)); mtx_init(&sc->sc_lock, sc->lockname, 0, MTX_DEF); t4_add_adapter(sc); mtx_init(&sc->sfl_lock, "starving freelists", 0, MTX_DEF); TAILQ_INIT(&sc->sfl); callout_init_mtx(&sc->sfl_callout, &sc->sfl_lock, 0); mtx_init(&sc->reg_lock, "indirect register access", 0, MTX_DEF); rc = t4_map_bars_0_and_4(sc); if (rc != 0) goto done; /* error message displayed already */ /* * This is the real PF# to which we're attaching. Works from within PCI * passthrough environments too, where pci_get_function() could return a * different PF# depending on the passthrough configuration. We need to * use the real PF# in all our communication with the firmware. */ sc->pf = G_SOURCEPF(t4_read_reg(sc, A_PL_WHOAMI)); sc->mbox = sc->pf; memset(sc->chan_map, 0xff, sizeof(sc->chan_map)); /* Prepare the adapter for operation. */ buf = malloc(PAGE_SIZE, M_CXGBE, M_ZERO | M_WAITOK); rc = -t4_prep_adapter(sc, buf); free(buf, M_CXGBE); if (rc != 0) { device_printf(dev, "failed to prepare adapter: %d.\n", rc); goto done; } /* * Do this really early, with the memory windows set up even before the * character device. The userland tool's register i/o and mem read * will work even in "recovery mode". */ setup_memwin(sc); if (t4_init_devlog_params(sc, 0) == 0) fixup_devlog_params(sc); make_dev_args_init(&mda); mda.mda_devsw = &t4_cdevsw; mda.mda_uid = UID_ROOT; mda.mda_gid = GID_WHEEL; mda.mda_mode = 0600; mda.mda_si_drv1 = sc; rc = make_dev_s(&mda, &sc->cdev, "%s", device_get_nameunit(dev)); if (rc != 0) device_printf(dev, "failed to create nexus char device: %d.\n", rc); /* Go no further if recovery mode has been requested. */ if (TUNABLE_INT_FETCH("hw.cxgbe.sos", &i) && i != 0) { device_printf(dev, "recovery mode.\n"); goto done; } #if defined(__i386__) if ((cpu_feature & CPUID_CX8) == 0) { device_printf(dev, "64 bit atomics not available.\n"); rc = ENOTSUP; goto done; } #endif /* Prepare the firmware for operation */ rc = prep_firmware(sc); if (rc != 0) goto done; /* error message displayed already */ rc = get_params__post_init(sc); if (rc != 0) goto done; /* error message displayed already */ rc = set_params__post_init(sc); if (rc != 0) goto done; /* error message displayed already */ rc = t4_map_bar_2(sc); if (rc != 0) goto done; /* error message displayed already */ rc = t4_create_dma_tag(sc); if (rc != 0) goto done; /* error message displayed already */ /* * Number of VIs to create per-port. The first VI is the "main" regular * VI for the port. The rest are additional virtual interfaces on the * same physical port. Note that the main VI does not have native * netmap support but the extra VIs do. * * Limit the number of VIs per port to the number of available * MAC addresses per port. */ if (t4_num_vis >= 1) num_vis = t4_num_vis; else num_vis = 1; if (num_vis > nitems(vi_mac_funcs)) { num_vis = nitems(vi_mac_funcs); device_printf(dev, "Number of VIs limited to %d\n", num_vis); } /* * First pass over all the ports - allocate VIs and initialize some * basic parameters like mac address, port type, etc. We also figure * out whether a port is 10G or 1G and use that information when * calculating how many interrupts to attempt to allocate. */ n10g = n1g = 0; for_each_port(sc, i) { struct port_info *pi; pi = malloc(sizeof(*pi), M_CXGBE, M_ZERO | M_WAITOK); sc->port[i] = pi; /* These must be set before t4_port_init */ pi->adapter = sc; pi->port_id = i; /* * XXX: vi[0] is special so we can't delay this allocation until * pi->nvi's final value is known. */ pi->vi = malloc(sizeof(struct vi_info) * num_vis, M_CXGBE, M_ZERO | M_WAITOK); /* * Allocate the "main" VI and initialize parameters * like mac addr. */ rc = -t4_port_init(sc, sc->mbox, sc->pf, 0, i); if (rc != 0) { device_printf(dev, "unable to initialize port %d: %d\n", i, rc); free(pi->vi, M_CXGBE); free(pi, M_CXGBE); sc->port[i] = NULL; goto done; } pi->link_cfg.requested_fc &= ~(PAUSE_TX | PAUSE_RX); pi->link_cfg.requested_fc |= t4_pause_settings; pi->link_cfg.fc &= ~(PAUSE_TX | PAUSE_RX); pi->link_cfg.fc |= t4_pause_settings; rc = -t4_link_l1cfg(sc, sc->mbox, pi->tx_chan, &pi->link_cfg); if (rc != 0) { device_printf(dev, "port %d l1cfg failed: %d\n", i, rc); free(pi->vi, M_CXGBE); free(pi, M_CXGBE); sc->port[i] = NULL; goto done; } snprintf(pi->lockname, sizeof(pi->lockname), "%sp%d", device_get_nameunit(dev), i); mtx_init(&pi->pi_lock, pi->lockname, 0, MTX_DEF); sc->chan_map[pi->tx_chan] = i; pi->tc = malloc(sizeof(struct tx_sched_class) * sc->chip_params->nsched_cls, M_CXGBE, M_ZERO | M_WAITOK); if (is_10G_port(pi) || is_40G_port(pi)) { n10g++; } else { n1g++; } pi->linkdnrc = -1; pi->dev = device_add_child(dev, is_t4(sc) ? "cxgbe" : "cxl", -1); if (pi->dev == NULL) { device_printf(dev, "failed to add device for port %d.\n", i); rc = ENXIO; goto done; } pi->vi[0].dev = pi->dev; device_set_softc(pi->dev, pi); } /* * Interrupt type, # of interrupts, # of rx/tx queues, etc. */ rc = cfg_itype_and_nqueues(sc, n10g, n1g, num_vis, &iaq); if (rc != 0) goto done; /* error message displayed already */ if (iaq.nrxq_vi + iaq.nofldrxq_vi + iaq.nnmrxq_vi == 0) num_vis = 1; sc->intr_type = iaq.intr_type; sc->intr_count = iaq.nirq; s = &sc->sge; s->nrxq = n10g * iaq.nrxq10g + n1g * iaq.nrxq1g; s->ntxq = n10g * iaq.ntxq10g + n1g * iaq.ntxq1g; if (num_vis > 1) { s->nrxq += (n10g + n1g) * (num_vis - 1) * iaq.nrxq_vi; s->ntxq += (n10g + n1g) * (num_vis - 1) * iaq.ntxq_vi; } s->neq = s->ntxq + s->nrxq; /* the free list in an rxq is an eq */ s->neq += sc->params.nports + 1;/* ctrl queues: 1 per port + 1 mgmt */ s->niq = s->nrxq + 1; /* 1 extra for firmware event queue */ #ifdef TCP_OFFLOAD if (is_offload(sc)) { s->nofldrxq = n10g * iaq.nofldrxq10g + n1g * iaq.nofldrxq1g; s->nofldtxq = n10g * iaq.nofldtxq10g + n1g * iaq.nofldtxq1g; if (num_vis > 1) { s->nofldrxq += (n10g + n1g) * (num_vis - 1) * iaq.nofldrxq_vi; s->nofldtxq += (n10g + n1g) * (num_vis - 1) * iaq.nofldtxq_vi; } s->neq += s->nofldtxq + s->nofldrxq; s->niq += s->nofldrxq; s->ofld_rxq = malloc(s->nofldrxq * sizeof(struct sge_ofld_rxq), M_CXGBE, M_ZERO | M_WAITOK); s->ofld_txq = malloc(s->nofldtxq * sizeof(struct sge_wrq), M_CXGBE, M_ZERO | M_WAITOK); } #endif #ifdef DEV_NETMAP if (num_vis > 1) { s->nnmrxq = (n10g + n1g) * (num_vis - 1) * iaq.nnmrxq_vi; s->nnmtxq = (n10g + n1g) * (num_vis - 1) * iaq.nnmtxq_vi; } s->neq += s->nnmtxq + s->nnmrxq; s->niq += s->nnmrxq; s->nm_rxq = malloc(s->nnmrxq * sizeof(struct sge_nm_rxq), M_CXGBE, M_ZERO | M_WAITOK); s->nm_txq = malloc(s->nnmtxq * sizeof(struct sge_nm_txq), M_CXGBE, M_ZERO | M_WAITOK); #endif s->ctrlq = malloc(sc->params.nports * sizeof(struct sge_wrq), M_CXGBE, M_ZERO | M_WAITOK); s->rxq = malloc(s->nrxq * sizeof(struct sge_rxq), M_CXGBE, M_ZERO | M_WAITOK); s->txq = malloc(s->ntxq * sizeof(struct sge_txq), M_CXGBE, M_ZERO | M_WAITOK); s->iqmap = malloc(s->niq * sizeof(struct sge_iq *), M_CXGBE, M_ZERO | M_WAITOK); s->eqmap = malloc(s->neq * sizeof(struct sge_eq *), M_CXGBE, M_ZERO | M_WAITOK); sc->irq = malloc(sc->intr_count * sizeof(struct irq), M_CXGBE, M_ZERO | M_WAITOK); t4_init_l2t(sc, M_WAITOK); /* * Second pass over the ports. This time we know the number of rx and * tx queues that each port should get. */ rqidx = tqidx = 0; #ifdef TCP_OFFLOAD ofld_rqidx = ofld_tqidx = 0; #endif #ifdef DEV_NETMAP nm_rqidx = nm_tqidx = 0; #endif for_each_port(sc, i) { struct port_info *pi = sc->port[i]; struct vi_info *vi; if (pi == NULL) continue; pi->nvi = num_vis; for_each_vi(pi, j, vi) { vi->pi = pi; vi->qsize_rxq = t4_qsize_rxq; vi->qsize_txq = t4_qsize_txq; vi->first_rxq = rqidx; vi->first_txq = tqidx; if (is_10G_port(pi) || is_40G_port(pi)) { vi->tmr_idx = t4_tmr_idx_10g; vi->pktc_idx = t4_pktc_idx_10g; vi->flags |= iaq.intr_flags_10g & INTR_RXQ; vi->nrxq = j == 0 ? iaq.nrxq10g : iaq.nrxq_vi; vi->ntxq = j == 0 ? iaq.ntxq10g : iaq.ntxq_vi; } else { vi->tmr_idx = t4_tmr_idx_1g; vi->pktc_idx = t4_pktc_idx_1g; vi->flags |= iaq.intr_flags_1g & INTR_RXQ; vi->nrxq = j == 0 ? iaq.nrxq1g : iaq.nrxq_vi; vi->ntxq = j == 0 ? iaq.ntxq1g : iaq.ntxq_vi; } rqidx += vi->nrxq; tqidx += vi->ntxq; if (j == 0 && vi->ntxq > 1) vi->rsrv_noflowq = iaq.rsrv_noflowq ? 1 : 0; else vi->rsrv_noflowq = 0; #ifdef TCP_OFFLOAD vi->first_ofld_rxq = ofld_rqidx; vi->first_ofld_txq = ofld_tqidx; if (is_10G_port(pi) || is_40G_port(pi)) { vi->flags |= iaq.intr_flags_10g & INTR_OFLD_RXQ; vi->nofldrxq = j == 0 ? iaq.nofldrxq10g : iaq.nofldrxq_vi; vi->nofldtxq = j == 0 ? iaq.nofldtxq10g : iaq.nofldtxq_vi; } else { vi->flags |= iaq.intr_flags_1g & INTR_OFLD_RXQ; vi->nofldrxq = j == 0 ? iaq.nofldrxq1g : iaq.nofldrxq_vi; vi->nofldtxq = j == 0 ? iaq.nofldtxq1g : iaq.nofldtxq_vi; } ofld_rqidx += vi->nofldrxq; ofld_tqidx += vi->nofldtxq; #endif #ifdef DEV_NETMAP if (j > 0) { vi->first_nm_rxq = nm_rqidx; vi->first_nm_txq = nm_tqidx; vi->nnmrxq = iaq.nnmrxq_vi; vi->nnmtxq = iaq.nnmtxq_vi; nm_rqidx += vi->nnmrxq; nm_tqidx += vi->nnmtxq; } #endif } } rc = t4_setup_intr_handlers(sc); if (rc != 0) { device_printf(dev, "failed to setup interrupt handlers: %d\n", rc); goto done; } rc = bus_generic_attach(dev); if (rc != 0) { device_printf(dev, "failed to attach all child ports: %d\n", rc); goto done; } device_printf(dev, "PCIe gen%d x%d, %d ports, %d %s interrupt%s, %d eq, %d iq\n", sc->params.pci.speed, sc->params.pci.width, sc->params.nports, sc->intr_count, sc->intr_type == INTR_MSIX ? "MSI-X" : (sc->intr_type == INTR_MSI ? "MSI" : "INTx"), sc->intr_count > 1 ? "s" : "", sc->sge.neq, sc->sge.niq); t4_set_desc(sc); notify_siblings(dev, 0); done: if (rc != 0 && sc->cdev) { /* cdev was created and so cxgbetool works; recover that way. */ device_printf(dev, "error during attach, adapter is now in recovery mode.\n"); rc = 0; } if (rc != 0) t4_detach_common(dev); else t4_sysctls(sc); return (rc); } static int t4_ready(device_t dev) { struct adapter *sc; sc = device_get_softc(dev); if (sc->flags & FW_OK) return (0); return (ENXIO); } static int t4_read_port_device(device_t dev, int port, device_t *child) { struct adapter *sc; struct port_info *pi; sc = device_get_softc(dev); if (port < 0 || port >= MAX_NPORTS) return (EINVAL); pi = sc->port[port]; if (pi == NULL || pi->dev == NULL) return (ENXIO); *child = pi->dev; return (0); } static int notify_siblings(device_t dev, int detaching) { device_t sibling; int error, i; error = 0; for (i = 0; i < PCI_FUNCMAX; i++) { if (i == pci_get_function(dev)) continue; sibling = pci_find_dbsf(pci_get_domain(dev), pci_get_bus(dev), pci_get_slot(dev), i); if (sibling == NULL || !device_is_attached(sibling)) continue; if (detaching) error = T4_DETACH_CHILD(sibling); else (void)T4_ATTACH_CHILD(sibling); if (error) break; } return (error); } /* * Idempotent */ static int t4_detach(device_t dev) { struct adapter *sc; int rc; sc = device_get_softc(dev); rc = notify_siblings(dev, 1); if (rc) { device_printf(dev, "failed to detach sibling devices: %d\n", rc); return (rc); } return (t4_detach_common(dev)); } int t4_detach_common(device_t dev) { struct adapter *sc; struct port_info *pi; int i, rc; sc = device_get_softc(dev); if (sc->flags & FULL_INIT_DONE) { if (!(sc->flags & IS_VF)) t4_intr_disable(sc); } if (sc->cdev) { destroy_dev(sc->cdev); sc->cdev = NULL; } if (device_is_attached(dev)) { rc = bus_generic_detach(dev); if (rc) { device_printf(dev, "failed to detach child devices: %d\n", rc); return (rc); } } for (i = 0; i < sc->intr_count; i++) t4_free_irq(sc, &sc->irq[i]); for (i = 0; i < MAX_NPORTS; i++) { pi = sc->port[i]; if (pi) { t4_free_vi(sc, sc->mbox, sc->pf, 0, pi->vi[0].viid); if (pi->dev) device_delete_child(dev, pi->dev); mtx_destroy(&pi->pi_lock); free(pi->vi, M_CXGBE); free(pi->tc, M_CXGBE); free(pi, M_CXGBE); } } if (sc->flags & FULL_INIT_DONE) adapter_full_uninit(sc); if ((sc->flags & (IS_VF | FW_OK)) == FW_OK) t4_fw_bye(sc, sc->mbox); if (sc->intr_type == INTR_MSI || sc->intr_type == INTR_MSIX) pci_release_msi(dev); if (sc->regs_res) bus_release_resource(dev, SYS_RES_MEMORY, sc->regs_rid, sc->regs_res); if (sc->udbs_res) bus_release_resource(dev, SYS_RES_MEMORY, sc->udbs_rid, sc->udbs_res); if (sc->msix_res) bus_release_resource(dev, SYS_RES_MEMORY, sc->msix_rid, sc->msix_res); if (sc->l2t) t4_free_l2t(sc->l2t); #ifdef TCP_OFFLOAD free(sc->sge.ofld_rxq, M_CXGBE); free(sc->sge.ofld_txq, M_CXGBE); #endif #ifdef DEV_NETMAP free(sc->sge.nm_rxq, M_CXGBE); free(sc->sge.nm_txq, M_CXGBE); #endif free(sc->irq, M_CXGBE); free(sc->sge.rxq, M_CXGBE); free(sc->sge.txq, M_CXGBE); free(sc->sge.ctrlq, M_CXGBE); free(sc->sge.iqmap, M_CXGBE); free(sc->sge.eqmap, M_CXGBE); free(sc->tids.ftid_tab, M_CXGBE); t4_destroy_dma_tag(sc); if (mtx_initialized(&sc->sc_lock)) { sx_xlock(&t4_list_lock); SLIST_REMOVE(&t4_list, sc, adapter, link); sx_xunlock(&t4_list_lock); mtx_destroy(&sc->sc_lock); } callout_drain(&sc->sfl_callout); if (mtx_initialized(&sc->tids.ftid_lock)) mtx_destroy(&sc->tids.ftid_lock); if (mtx_initialized(&sc->sfl_lock)) mtx_destroy(&sc->sfl_lock); if (mtx_initialized(&sc->ifp_lock)) mtx_destroy(&sc->ifp_lock); if (mtx_initialized(&sc->reg_lock)) mtx_destroy(&sc->reg_lock); for (i = 0; i < NUM_MEMWIN; i++) { struct memwin *mw = &sc->memwin[i]; if (rw_initialized(&mw->mw_lock)) rw_destroy(&mw->mw_lock); } bzero(sc, sizeof(*sc)); return (0); } static int cxgbe_probe(device_t dev) { char buf[128]; struct port_info *pi = device_get_softc(dev); snprintf(buf, sizeof(buf), "port %d", pi->port_id); device_set_desc_copy(dev, buf); return (BUS_PROBE_DEFAULT); } #define T4_CAP (IFCAP_VLAN_HWTAGGING | IFCAP_VLAN_MTU | IFCAP_HWCSUM | \ IFCAP_VLAN_HWCSUM | IFCAP_TSO | IFCAP_JUMBO_MTU | IFCAP_LRO | \ IFCAP_VLAN_HWTSO | IFCAP_LINKSTATE | IFCAP_HWCSUM_IPV6 | IFCAP_HWSTATS) #define T4_CAP_ENABLE (T4_CAP) static int cxgbe_vi_attach(device_t dev, struct vi_info *vi) { struct ifnet *ifp; struct sbuf *sb; vi->xact_addr_filt = -1; callout_init(&vi->tick, 1); /* Allocate an ifnet and set it up */ ifp = if_alloc(IFT_ETHER); if (ifp == NULL) { device_printf(dev, "Cannot allocate ifnet\n"); return (ENOMEM); } vi->ifp = ifp; ifp->if_softc = vi; if_initname(ifp, device_get_name(dev), device_get_unit(dev)); ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST; ifp->if_init = cxgbe_init; ifp->if_ioctl = cxgbe_ioctl; ifp->if_transmit = cxgbe_transmit; ifp->if_qflush = cxgbe_qflush; ifp->if_get_counter = cxgbe_get_counter; ifp->if_capabilities = T4_CAP; #ifdef TCP_OFFLOAD if (vi->nofldrxq != 0) ifp->if_capabilities |= IFCAP_TOE; #endif ifp->if_capenable = T4_CAP_ENABLE; ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO | CSUM_UDP_IPV6 | CSUM_TCP_IPV6; ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN); ifp->if_hw_tsomaxsegcount = TX_SGL_SEGS; ifp->if_hw_tsomaxsegsize = 65536; /* Initialize ifmedia for this VI */ ifmedia_init(&vi->media, IFM_IMASK, cxgbe_media_change, cxgbe_media_status); build_medialist(vi->pi, &vi->media); vi->vlan_c = EVENTHANDLER_REGISTER(vlan_config, cxgbe_vlan_config, ifp, EVENTHANDLER_PRI_ANY); ether_ifattach(ifp, vi->hw_addr); #ifdef DEV_NETMAP if (vi->nnmrxq != 0) cxgbe_nm_attach(vi); #endif sb = sbuf_new_auto(); sbuf_printf(sb, "%d txq, %d rxq (NIC)", vi->ntxq, vi->nrxq); #ifdef TCP_OFFLOAD if (ifp->if_capabilities & IFCAP_TOE) sbuf_printf(sb, "; %d txq, %d rxq (TOE)", vi->nofldtxq, vi->nofldrxq); #endif #ifdef DEV_NETMAP if (ifp->if_capabilities & IFCAP_NETMAP) sbuf_printf(sb, "; %d txq, %d rxq (netmap)", vi->nnmtxq, vi->nnmrxq); #endif sbuf_finish(sb); device_printf(dev, "%s\n", sbuf_data(sb)); sbuf_delete(sb); vi_sysctls(vi); return (0); } static int cxgbe_attach(device_t dev) { struct port_info *pi = device_get_softc(dev); struct vi_info *vi; int i, rc; callout_init_mtx(&pi->tick, &pi->pi_lock, 0); rc = cxgbe_vi_attach(dev, &pi->vi[0]); if (rc) return (rc); for_each_vi(pi, i, vi) { if (i == 0) continue; vi->dev = device_add_child(dev, is_t4(pi->adapter) ? "vcxgbe" : "vcxl", -1); if (vi->dev == NULL) { device_printf(dev, "failed to add VI %d\n", i); continue; } device_set_softc(vi->dev, vi); } cxgbe_sysctls(pi); bus_generic_attach(dev); return (0); } static void cxgbe_vi_detach(struct vi_info *vi) { struct ifnet *ifp = vi->ifp; ether_ifdetach(ifp); if (vi->vlan_c) EVENTHANDLER_DEREGISTER(vlan_config, vi->vlan_c); /* Let detach proceed even if these fail. */ #ifdef DEV_NETMAP if (ifp->if_capabilities & IFCAP_NETMAP) cxgbe_nm_detach(vi); #endif cxgbe_uninit_synchronized(vi); callout_drain(&vi->tick); vi_full_uninit(vi); ifmedia_removeall(&vi->media); if_free(vi->ifp); vi->ifp = NULL; } static int cxgbe_detach(device_t dev) { struct port_info *pi = device_get_softc(dev); struct adapter *sc = pi->adapter; int rc; /* Detach the extra VIs first. */ rc = bus_generic_detach(dev); if (rc) return (rc); device_delete_children(dev); doom_vi(sc, &pi->vi[0]); if (pi->flags & HAS_TRACEQ) { sc->traceq = -1; /* cloner should not create ifnet */ t4_tracer_port_detach(sc); } cxgbe_vi_detach(&pi->vi[0]); callout_drain(&pi->tick); end_synchronized_op(sc, 0); return (0); } static void cxgbe_init(void *arg) { struct vi_info *vi = arg; struct adapter *sc = vi->pi->adapter; if (begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4init") != 0) return; cxgbe_init_synchronized(vi); end_synchronized_op(sc, 0); } static int cxgbe_ioctl(struct ifnet *ifp, unsigned long cmd, caddr_t data) { int rc = 0, mtu, flags, can_sleep; struct vi_info *vi = ifp->if_softc; struct adapter *sc = vi->pi->adapter; struct ifreq *ifr = (struct ifreq *)data; uint32_t mask; switch (cmd) { case SIOCSIFMTU: mtu = ifr->ifr_mtu; if ((mtu < ETHERMIN) || (mtu > ETHERMTU_JUMBO)) return (EINVAL); rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4mtu"); if (rc) return (rc); ifp->if_mtu = mtu; if (vi->flags & VI_INIT_DONE) { t4_update_fl_bufsize(ifp); if (ifp->if_drv_flags & IFF_DRV_RUNNING) rc = update_mac_settings(ifp, XGMAC_MTU); } end_synchronized_op(sc, 0); break; case SIOCSIFFLAGS: can_sleep = 0; redo_sifflags: rc = begin_synchronized_op(sc, vi, can_sleep ? (SLEEP_OK | INTR_OK) : HOLD_LOCK, "t4flg"); if (rc) return (rc); if (ifp->if_flags & IFF_UP) { if (ifp->if_drv_flags & IFF_DRV_RUNNING) { flags = vi->if_flags; if ((ifp->if_flags ^ flags) & (IFF_PROMISC | IFF_ALLMULTI)) { if (can_sleep == 1) { end_synchronized_op(sc, 0); can_sleep = 0; goto redo_sifflags; } rc = update_mac_settings(ifp, XGMAC_PROMISC | XGMAC_ALLMULTI); } } else { if (can_sleep == 0) { end_synchronized_op(sc, LOCK_HELD); can_sleep = 1; goto redo_sifflags; } rc = cxgbe_init_synchronized(vi); } vi->if_flags = ifp->if_flags; } else if (ifp->if_drv_flags & IFF_DRV_RUNNING) { if (can_sleep == 0) { end_synchronized_op(sc, LOCK_HELD); can_sleep = 1; goto redo_sifflags; } rc = cxgbe_uninit_synchronized(vi); } end_synchronized_op(sc, can_sleep ? 0 : LOCK_HELD); break; case SIOCADDMULTI: case SIOCDELMULTI: /* these two are called with a mutex held :-( */ rc = begin_synchronized_op(sc, vi, HOLD_LOCK, "t4multi"); if (rc) return (rc); if (ifp->if_drv_flags & IFF_DRV_RUNNING) rc = update_mac_settings(ifp, XGMAC_MCADDRS); end_synchronized_op(sc, LOCK_HELD); break; case SIOCSIFCAP: rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4cap"); if (rc) return (rc); mask = ifr->ifr_reqcap ^ ifp->if_capenable; if (mask & IFCAP_TXCSUM) { ifp->if_capenable ^= IFCAP_TXCSUM; ifp->if_hwassist ^= (CSUM_TCP | CSUM_UDP | CSUM_IP); if (IFCAP_TSO4 & ifp->if_capenable && !(IFCAP_TXCSUM & ifp->if_capenable)) { ifp->if_capenable &= ~IFCAP_TSO4; if_printf(ifp, "tso4 disabled due to -txcsum.\n"); } } if (mask & IFCAP_TXCSUM_IPV6) { ifp->if_capenable ^= IFCAP_TXCSUM_IPV6; ifp->if_hwassist ^= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6); if (IFCAP_TSO6 & ifp->if_capenable && !(IFCAP_TXCSUM_IPV6 & ifp->if_capenable)) { ifp->if_capenable &= ~IFCAP_TSO6; if_printf(ifp, "tso6 disabled due to -txcsum6.\n"); } } if (mask & IFCAP_RXCSUM) ifp->if_capenable ^= IFCAP_RXCSUM; if (mask & IFCAP_RXCSUM_IPV6) ifp->if_capenable ^= IFCAP_RXCSUM_IPV6; /* * Note that we leave CSUM_TSO alone (it is always set). The * kernel takes both IFCAP_TSOx and CSUM_TSO into account before * sending a TSO request our way, so it's sufficient to toggle * IFCAP_TSOx only. */ if (mask & IFCAP_TSO4) { if (!(IFCAP_TSO4 & ifp->if_capenable) && !(IFCAP_TXCSUM & ifp->if_capenable)) { if_printf(ifp, "enable txcsum first.\n"); rc = EAGAIN; goto fail; } ifp->if_capenable ^= IFCAP_TSO4; } if (mask & IFCAP_TSO6) { if (!(IFCAP_TSO6 & ifp->if_capenable) && !(IFCAP_TXCSUM_IPV6 & ifp->if_capenable)) { if_printf(ifp, "enable txcsum6 first.\n"); rc = EAGAIN; goto fail; } ifp->if_capenable ^= IFCAP_TSO6; } if (mask & IFCAP_LRO) { #if defined(INET) || defined(INET6) int i; struct sge_rxq *rxq; ifp->if_capenable ^= IFCAP_LRO; for_each_rxq(vi, i, rxq) { if (ifp->if_capenable & IFCAP_LRO) rxq->iq.flags |= IQ_LRO_ENABLED; else rxq->iq.flags &= ~IQ_LRO_ENABLED; } #endif } #ifdef TCP_OFFLOAD if (mask & IFCAP_TOE) { int enable = (ifp->if_capenable ^ mask) & IFCAP_TOE; rc = toe_capability(vi, enable); if (rc != 0) goto fail; ifp->if_capenable ^= mask; } #endif if (mask & IFCAP_VLAN_HWTAGGING) { ifp->if_capenable ^= IFCAP_VLAN_HWTAGGING; if (ifp->if_drv_flags & IFF_DRV_RUNNING) rc = update_mac_settings(ifp, XGMAC_VLANEX); } if (mask & IFCAP_VLAN_MTU) { ifp->if_capenable ^= IFCAP_VLAN_MTU; /* Need to find out how to disable auto-mtu-inflation */ } if (mask & IFCAP_VLAN_HWTSO) ifp->if_capenable ^= IFCAP_VLAN_HWTSO; if (mask & IFCAP_VLAN_HWCSUM) ifp->if_capenable ^= IFCAP_VLAN_HWCSUM; #ifdef VLAN_CAPABILITIES VLAN_CAPABILITIES(ifp); #endif fail: end_synchronized_op(sc, 0); break; case SIOCSIFMEDIA: case SIOCGIFMEDIA: ifmedia_ioctl(ifp, ifr, &vi->media, cmd); break; case SIOCGI2C: { struct ifi2creq i2c; rc = copyin(ifr->ifr_data, &i2c, sizeof(i2c)); if (rc != 0) break; if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) { rc = EPERM; break; } if (i2c.len > sizeof(i2c.data)) { rc = EINVAL; break; } rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4i2c"); if (rc) return (rc); rc = -t4_i2c_rd(sc, sc->mbox, vi->pi->port_id, i2c.dev_addr, i2c.offset, i2c.len, &i2c.data[0]); end_synchronized_op(sc, 0); if (rc == 0) rc = copyout(&i2c, ifr->ifr_data, sizeof(i2c)); break; } default: rc = ether_ioctl(ifp, cmd, data); } return (rc); } static int cxgbe_transmit(struct ifnet *ifp, struct mbuf *m) { struct vi_info *vi = ifp->if_softc; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct sge_txq *txq; void *items[1]; int rc; M_ASSERTPKTHDR(m); MPASS(m->m_nextpkt == NULL); /* not quite ready for this yet */ if (__predict_false(pi->link_cfg.link_ok == 0)) { m_freem(m); return (ENETDOWN); } rc = parse_pkt(sc, &m); if (__predict_false(rc != 0)) { MPASS(m == NULL); /* was freed already */ atomic_add_int(&pi->tx_parse_error, 1); /* rare, atomic is ok */ return (rc); } /* Select a txq. */ txq = &sc->sge.txq[vi->first_txq]; if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) txq += ((m->m_pkthdr.flowid % (vi->ntxq - vi->rsrv_noflowq)) + vi->rsrv_noflowq); items[0] = m; rc = mp_ring_enqueue(txq->r, items, 1, 4096); if (__predict_false(rc != 0)) m_freem(m); return (rc); } static void cxgbe_qflush(struct ifnet *ifp) { struct vi_info *vi = ifp->if_softc; struct sge_txq *txq; int i; /* queues do not exist if !VI_INIT_DONE. */ if (vi->flags & VI_INIT_DONE) { for_each_txq(vi, i, txq) { TXQ_LOCK(txq); txq->eq.flags &= ~EQ_ENABLED; TXQ_UNLOCK(txq); while (!mp_ring_is_idle(txq->r)) { mp_ring_check_drainage(txq->r, 0); pause("qflush", 1); } } } if_qflush(ifp); } static uint64_t vi_get_counter(struct ifnet *ifp, ift_counter c) { struct vi_info *vi = ifp->if_softc; struct fw_vi_stats_vf *s = &vi->stats; vi_refresh_stats(vi->pi->adapter, vi); switch (c) { case IFCOUNTER_IPACKETS: return (s->rx_bcast_frames + s->rx_mcast_frames + s->rx_ucast_frames); case IFCOUNTER_IERRORS: return (s->rx_err_frames); case IFCOUNTER_OPACKETS: return (s->tx_bcast_frames + s->tx_mcast_frames + s->tx_ucast_frames + s->tx_offload_frames); case IFCOUNTER_OERRORS: return (s->tx_drop_frames); case IFCOUNTER_IBYTES: return (s->rx_bcast_bytes + s->rx_mcast_bytes + s->rx_ucast_bytes); case IFCOUNTER_OBYTES: return (s->tx_bcast_bytes + s->tx_mcast_bytes + s->tx_ucast_bytes + s->tx_offload_bytes); case IFCOUNTER_IMCASTS: return (s->rx_mcast_frames); case IFCOUNTER_OMCASTS: return (s->tx_mcast_frames); case IFCOUNTER_OQDROPS: { uint64_t drops; drops = 0; if (vi->flags & VI_INIT_DONE) { int i; struct sge_txq *txq; for_each_txq(vi, i, txq) drops += counter_u64_fetch(txq->r->drops); } return (drops); } default: return (if_get_counter_default(ifp, c)); } } uint64_t cxgbe_get_counter(struct ifnet *ifp, ift_counter c) { struct vi_info *vi = ifp->if_softc; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct port_stats *s = &pi->stats; if (pi->nvi > 1 || sc->flags & IS_VF) return (vi_get_counter(ifp, c)); cxgbe_refresh_stats(sc, pi); switch (c) { case IFCOUNTER_IPACKETS: return (s->rx_frames); case IFCOUNTER_IERRORS: return (s->rx_jabber + s->rx_runt + s->rx_too_long + s->rx_fcs_err + s->rx_len_err); case IFCOUNTER_OPACKETS: return (s->tx_frames); case IFCOUNTER_OERRORS: return (s->tx_error_frames); case IFCOUNTER_IBYTES: return (s->rx_octets); case IFCOUNTER_OBYTES: return (s->tx_octets); case IFCOUNTER_IMCASTS: return (s->rx_mcast_frames); case IFCOUNTER_OMCASTS: return (s->tx_mcast_frames); case IFCOUNTER_IQDROPS: return (s->rx_ovflow0 + s->rx_ovflow1 + s->rx_ovflow2 + s->rx_ovflow3 + s->rx_trunc0 + s->rx_trunc1 + s->rx_trunc2 + s->rx_trunc3 + pi->tnl_cong_drops); case IFCOUNTER_OQDROPS: { uint64_t drops; drops = s->tx_drop; if (vi->flags & VI_INIT_DONE) { int i; struct sge_txq *txq; for_each_txq(vi, i, txq) drops += counter_u64_fetch(txq->r->drops); } return (drops); } default: return (if_get_counter_default(ifp, c)); } } static int cxgbe_media_change(struct ifnet *ifp) { struct vi_info *vi = ifp->if_softc; device_printf(vi->dev, "%s unimplemented.\n", __func__); return (EOPNOTSUPP); } static void cxgbe_media_status(struct ifnet *ifp, struct ifmediareq *ifmr) { struct vi_info *vi = ifp->if_softc; struct port_info *pi = vi->pi; struct ifmedia_entry *cur; int speed = pi->link_cfg.speed; cur = vi->media.ifm_cur; ifmr->ifm_status = IFM_AVALID; if (!pi->link_cfg.link_ok) return; ifmr->ifm_status |= IFM_ACTIVE; /* active and current will differ iff current media is autoselect. */ if (IFM_SUBTYPE(cur->ifm_media) != IFM_AUTO) return; ifmr->ifm_active = IFM_ETHER | IFM_FDX; if (speed == 10000) ifmr->ifm_active |= IFM_10G_T; else if (speed == 1000) ifmr->ifm_active |= IFM_1000_T; else if (speed == 100) ifmr->ifm_active |= IFM_100_TX; else if (speed == 10) ifmr->ifm_active |= IFM_10_T; else KASSERT(0, ("%s: link up but speed unknown (%u)", __func__, speed)); } static int vcxgbe_probe(device_t dev) { char buf[128]; struct vi_info *vi = device_get_softc(dev); snprintf(buf, sizeof(buf), "port %d vi %td", vi->pi->port_id, vi - vi->pi->vi); device_set_desc_copy(dev, buf); return (BUS_PROBE_DEFAULT); } static int vcxgbe_attach(device_t dev) { struct vi_info *vi; struct port_info *pi; struct adapter *sc; int func, index, rc; u32 param, val; vi = device_get_softc(dev); pi = vi->pi; sc = pi->adapter; index = vi - pi->vi; KASSERT(index < nitems(vi_mac_funcs), ("%s: VI %s doesn't have a MAC func", __func__, device_get_nameunit(dev))); func = vi_mac_funcs[index]; rc = t4_alloc_vi_func(sc, sc->mbox, pi->tx_chan, sc->pf, 0, 1, vi->hw_addr, &vi->rss_size, func, 0); if (rc < 0) { device_printf(dev, "Failed to allocate virtual interface " "for port %d: %d\n", pi->port_id, -rc); return (-rc); } vi->viid = rc; param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_RSSINFO) | V_FW_PARAMS_PARAM_YZ(vi->viid); rc = t4_query_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); if (rc) vi->rss_base = 0xffff; else { /* MPASS((val >> 16) == rss_size); */ vi->rss_base = val & 0xffff; } rc = cxgbe_vi_attach(dev, vi); if (rc) { t4_free_vi(sc, sc->mbox, sc->pf, 0, vi->viid); return (rc); } return (0); } static int vcxgbe_detach(device_t dev) { struct vi_info *vi; struct adapter *sc; vi = device_get_softc(dev); sc = vi->pi->adapter; doom_vi(sc, vi); cxgbe_vi_detach(vi); t4_free_vi(sc, sc->mbox, sc->pf, 0, vi->viid); end_synchronized_op(sc, 0); return (0); } void t4_fatal_err(struct adapter *sc) { t4_set_reg_field(sc, A_SGE_CONTROL, F_GLOBALENABLE, 0); t4_intr_disable(sc); log(LOG_EMERG, "%s: encountered fatal error, adapter stopped.\n", device_get_nameunit(sc->dev)); } void t4_add_adapter(struct adapter *sc) { sx_xlock(&t4_list_lock); SLIST_INSERT_HEAD(&t4_list, sc, link); sx_xunlock(&t4_list_lock); } int t4_map_bars_0_and_4(struct adapter *sc) { sc->regs_rid = PCIR_BAR(0); sc->regs_res = bus_alloc_resource_any(sc->dev, SYS_RES_MEMORY, &sc->regs_rid, RF_ACTIVE); if (sc->regs_res == NULL) { device_printf(sc->dev, "cannot map registers.\n"); return (ENXIO); } sc->bt = rman_get_bustag(sc->regs_res); sc->bh = rman_get_bushandle(sc->regs_res); sc->mmio_len = rman_get_size(sc->regs_res); setbit(&sc->doorbells, DOORBELL_KDB); sc->msix_rid = PCIR_BAR(4); sc->msix_res = bus_alloc_resource_any(sc->dev, SYS_RES_MEMORY, &sc->msix_rid, RF_ACTIVE); if (sc->msix_res == NULL) { device_printf(sc->dev, "cannot map MSI-X BAR.\n"); return (ENXIO); } return (0); } int t4_map_bar_2(struct adapter *sc) { /* * T4: only iWARP driver uses the userspace doorbells. There is no need * to map it if RDMA is disabled. */ if (is_t4(sc) && sc->rdmacaps == 0) return (0); sc->udbs_rid = PCIR_BAR(2); sc->udbs_res = bus_alloc_resource_any(sc->dev, SYS_RES_MEMORY, &sc->udbs_rid, RF_ACTIVE); if (sc->udbs_res == NULL) { device_printf(sc->dev, "cannot map doorbell BAR.\n"); return (ENXIO); } sc->udbs_base = rman_get_virtual(sc->udbs_res); - if (is_t5(sc)) { + if (chip_id(sc) >= CHELSIO_T5) { setbit(&sc->doorbells, DOORBELL_UDB); #if defined(__i386__) || defined(__amd64__) if (t5_write_combine) { - int rc; + int rc, mode; /* * Enable write combining on BAR2. This is the * userspace doorbell BAR and is split into 128B * (UDBS_SEG_SIZE) doorbell regions, each associated * with an egress queue. The first 64B has the doorbell * and the second 64B can be used to submit a tx work * request with an implicit doorbell. */ rc = pmap_change_attr((vm_offset_t)sc->udbs_base, rman_get_size(sc->udbs_res), PAT_WRITE_COMBINING); if (rc == 0) { clrbit(&sc->doorbells, DOORBELL_UDB); setbit(&sc->doorbells, DOORBELL_WCWR); setbit(&sc->doorbells, DOORBELL_UDBWC); } else { device_printf(sc->dev, "couldn't enable write combining: %d\n", rc); } + mode = is_t5(sc) ? V_STATMODE(0) : V_T6_STATMODE(0); t4_write_reg(sc, A_SGE_STAT_CFG, - V_STATSOURCE_T5(7) | V_STATMODE(0)); + V_STATSOURCE_T5(7) | mode); } #endif } return (0); } struct memwin_init { uint32_t base; uint32_t aperture; }; static const struct memwin_init t4_memwin[NUM_MEMWIN] = { { MEMWIN0_BASE, MEMWIN0_APERTURE }, { MEMWIN1_BASE, MEMWIN1_APERTURE }, { MEMWIN2_BASE_T4, MEMWIN2_APERTURE_T4 } }; static const struct memwin_init t5_memwin[NUM_MEMWIN] = { { MEMWIN0_BASE, MEMWIN0_APERTURE }, { MEMWIN1_BASE, MEMWIN1_APERTURE }, { MEMWIN2_BASE_T5, MEMWIN2_APERTURE_T5 }, }; static void setup_memwin(struct adapter *sc) { const struct memwin_init *mw_init; struct memwin *mw; int i; uint32_t bar0; if (is_t4(sc)) { /* * Read low 32b of bar0 indirectly via the hardware backdoor * mechanism. Works from within PCI passthrough environments * too, where rman_get_start() can return a different value. We * need to program the T4 memory window decoders with the actual * addresses that will be coming across the PCIe link. */ bar0 = t4_hw_pci_read_cfg4(sc, PCIR_BAR(0)); bar0 &= (uint32_t) PCIM_BAR_MEM_BASE; mw_init = &t4_memwin[0]; } else { /* T5+ use the relative offset inside the PCIe BAR */ bar0 = 0; mw_init = &t5_memwin[0]; } for (i = 0, mw = &sc->memwin[0]; i < NUM_MEMWIN; i++, mw_init++, mw++) { rw_init(&mw->mw_lock, "memory window access"); mw->mw_base = mw_init->base; mw->mw_aperture = mw_init->aperture; mw->mw_curpos = 0; t4_write_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_BASE_WIN, i), (mw->mw_base + bar0) | V_BIR(0) | V_WINDOW(ilog2(mw->mw_aperture) - 10)); rw_wlock(&mw->mw_lock); position_memwin(sc, i, 0); rw_wunlock(&mw->mw_lock); } /* flush */ t4_read_reg(sc, PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_BASE_WIN, 2)); } /* * Positions the memory window at the given address in the card's address space. * There are some alignment requirements and the actual position may be at an * address prior to the requested address. mw->mw_curpos always has the actual * position of the window. */ static void position_memwin(struct adapter *sc, int idx, uint32_t addr) { struct memwin *mw; uint32_t pf; uint32_t reg; MPASS(idx >= 0 && idx < NUM_MEMWIN); mw = &sc->memwin[idx]; rw_assert(&mw->mw_lock, RA_WLOCKED); if (is_t4(sc)) { pf = 0; mw->mw_curpos = addr & ~0xf; /* start must be 16B aligned */ } else { pf = V_PFNUM(sc->pf); mw->mw_curpos = addr & ~0x7f; /* start must be 128B aligned */ } reg = PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, idx); t4_write_reg(sc, reg, mw->mw_curpos | pf); t4_read_reg(sc, reg); /* flush */ } static int rw_via_memwin(struct adapter *sc, int idx, uint32_t addr, uint32_t *val, int len, int rw) { struct memwin *mw; uint32_t mw_end, v; MPASS(idx >= 0 && idx < NUM_MEMWIN); /* Memory can only be accessed in naturally aligned 4 byte units */ if (addr & 3 || len & 3 || len <= 0) return (EINVAL); mw = &sc->memwin[idx]; while (len > 0) { rw_rlock(&mw->mw_lock); mw_end = mw->mw_curpos + mw->mw_aperture; if (addr >= mw_end || addr < mw->mw_curpos) { /* Will need to reposition the window */ if (!rw_try_upgrade(&mw->mw_lock)) { rw_runlock(&mw->mw_lock); rw_wlock(&mw->mw_lock); } rw_assert(&mw->mw_lock, RA_WLOCKED); position_memwin(sc, idx, addr); rw_downgrade(&mw->mw_lock); mw_end = mw->mw_curpos + mw->mw_aperture; } rw_assert(&mw->mw_lock, RA_RLOCKED); while (addr < mw_end && len > 0) { if (rw == 0) { v = t4_read_reg(sc, mw->mw_base + addr - mw->mw_curpos); *val++ = le32toh(v); } else { v = *val++; t4_write_reg(sc, mw->mw_base + addr - mw->mw_curpos, htole32(v)); } addr += 4; len -= 4; } rw_runlock(&mw->mw_lock); } return (0); } static inline int read_via_memwin(struct adapter *sc, int idx, uint32_t addr, uint32_t *val, int len) { return (rw_via_memwin(sc, idx, addr, val, len, 0)); } static inline int write_via_memwin(struct adapter *sc, int idx, uint32_t addr, const uint32_t *val, int len) { return (rw_via_memwin(sc, idx, addr, (void *)(uintptr_t)val, len, 1)); } static int t4_range_cmp(const void *a, const void *b) { return ((const struct t4_range *)a)->start - ((const struct t4_range *)b)->start; } /* * Verify that the memory range specified by the addr/len pair is valid within * the card's address space. */ static int validate_mem_range(struct adapter *sc, uint32_t addr, int len) { struct t4_range mem_ranges[4], *r, *next; uint32_t em, addr_len; int i, n, remaining; /* Memory can only be accessed in naturally aligned 4 byte units */ if (addr & 3 || len & 3 || len <= 0) return (EINVAL); /* Enabled memories */ em = t4_read_reg(sc, A_MA_TARGET_MEM_ENABLE); r = &mem_ranges[0]; n = 0; bzero(r, sizeof(mem_ranges)); if (em & F_EDRAM0_ENABLE) { addr_len = t4_read_reg(sc, A_MA_EDRAM0_BAR); r->size = G_EDRAM0_SIZE(addr_len) << 20; if (r->size > 0) { r->start = G_EDRAM0_BASE(addr_len) << 20; if (addr >= r->start && addr + len <= r->start + r->size) return (0); r++; n++; } } if (em & F_EDRAM1_ENABLE) { addr_len = t4_read_reg(sc, A_MA_EDRAM1_BAR); r->size = G_EDRAM1_SIZE(addr_len) << 20; if (r->size > 0) { r->start = G_EDRAM1_BASE(addr_len) << 20; if (addr >= r->start && addr + len <= r->start + r->size) return (0); r++; n++; } } if (em & F_EXT_MEM_ENABLE) { addr_len = t4_read_reg(sc, A_MA_EXT_MEMORY_BAR); r->size = G_EXT_MEM_SIZE(addr_len) << 20; if (r->size > 0) { r->start = G_EXT_MEM_BASE(addr_len) << 20; if (addr >= r->start && addr + len <= r->start + r->size) return (0); r++; n++; } } if (is_t5(sc) && em & F_EXT_MEM1_ENABLE) { addr_len = t4_read_reg(sc, A_MA_EXT_MEMORY1_BAR); r->size = G_EXT_MEM1_SIZE(addr_len) << 20; if (r->size > 0) { r->start = G_EXT_MEM1_BASE(addr_len) << 20; if (addr >= r->start && addr + len <= r->start + r->size) return (0); r++; n++; } } MPASS(n <= nitems(mem_ranges)); if (n > 1) { /* Sort and merge the ranges. */ qsort(mem_ranges, n, sizeof(struct t4_range), t4_range_cmp); /* Start from index 0 and examine the next n - 1 entries. */ r = &mem_ranges[0]; for (remaining = n - 1; remaining > 0; remaining--, r++) { MPASS(r->size > 0); /* r is a valid entry. */ next = r + 1; MPASS(next->size > 0); /* and so is the next one. */ while (r->start + r->size >= next->start) { /* Merge the next one into the current entry. */ r->size = max(r->start + r->size, next->start + next->size) - r->start; n--; /* One fewer entry in total. */ if (--remaining == 0) goto done; /* short circuit */ next++; } if (next != r + 1) { /* * Some entries were merged into r and next * points to the first valid entry that couldn't * be merged. */ MPASS(next->size > 0); /* must be valid */ memcpy(r + 1, next, remaining * sizeof(*r)); #ifdef INVARIANTS /* * This so that the foo->size assertion in the * next iteration of the loop do the right * thing for entries that were pulled up and are * no longer valid. */ MPASS(n < nitems(mem_ranges)); bzero(&mem_ranges[n], (nitems(mem_ranges) - n) * sizeof(struct t4_range)); #endif } } done: /* Done merging the ranges. */ MPASS(n > 0); r = &mem_ranges[0]; for (i = 0; i < n; i++, r++) { if (addr >= r->start && addr + len <= r->start + r->size) return (0); } } return (EFAULT); } static int fwmtype_to_hwmtype(int mtype) { switch (mtype) { case FW_MEMTYPE_EDC0: return (MEM_EDC0); case FW_MEMTYPE_EDC1: return (MEM_EDC1); case FW_MEMTYPE_EXTMEM: return (MEM_MC0); case FW_MEMTYPE_EXTMEM1: return (MEM_MC1); default: panic("%s: cannot translate fw mtype %d.", __func__, mtype); } } /* * Verify that the memory range specified by the memtype/offset/len pair is * valid and lies entirely within the memtype specified. The global address of * the start of the range is returned in addr. */ static int validate_mt_off_len(struct adapter *sc, int mtype, uint32_t off, int len, uint32_t *addr) { uint32_t em, addr_len, maddr; /* Memory can only be accessed in naturally aligned 4 byte units */ if (off & 3 || len & 3 || len == 0) return (EINVAL); em = t4_read_reg(sc, A_MA_TARGET_MEM_ENABLE); switch (fwmtype_to_hwmtype(mtype)) { case MEM_EDC0: if (!(em & F_EDRAM0_ENABLE)) return (EINVAL); addr_len = t4_read_reg(sc, A_MA_EDRAM0_BAR); maddr = G_EDRAM0_BASE(addr_len) << 20; break; case MEM_EDC1: if (!(em & F_EDRAM1_ENABLE)) return (EINVAL); addr_len = t4_read_reg(sc, A_MA_EDRAM1_BAR); maddr = G_EDRAM1_BASE(addr_len) << 20; break; case MEM_MC: if (!(em & F_EXT_MEM_ENABLE)) return (EINVAL); addr_len = t4_read_reg(sc, A_MA_EXT_MEMORY_BAR); maddr = G_EXT_MEM_BASE(addr_len) << 20; break; case MEM_MC1: if (!is_t5(sc) || !(em & F_EXT_MEM1_ENABLE)) return (EINVAL); addr_len = t4_read_reg(sc, A_MA_EXT_MEMORY1_BAR); maddr = G_EXT_MEM1_BASE(addr_len) << 20; break; default: return (EINVAL); } *addr = maddr + off; /* global address */ return (validate_mem_range(sc, *addr, len)); } static int fixup_devlog_params(struct adapter *sc) { struct devlog_params *dparams = &sc->params.devlog; int rc; rc = validate_mt_off_len(sc, dparams->memtype, dparams->start, dparams->size, &dparams->addr); return (rc); } static int cfg_itype_and_nqueues(struct adapter *sc, int n10g, int n1g, int num_vis, struct intrs_and_queues *iaq) { int rc, itype, navail, nrxq10g, nrxq1g, n; int nofldrxq10g = 0, nofldrxq1g = 0; bzero(iaq, sizeof(*iaq)); iaq->ntxq10g = t4_ntxq10g; iaq->ntxq1g = t4_ntxq1g; iaq->ntxq_vi = t4_ntxq_vi; iaq->nrxq10g = nrxq10g = t4_nrxq10g; iaq->nrxq1g = nrxq1g = t4_nrxq1g; iaq->nrxq_vi = t4_nrxq_vi; iaq->rsrv_noflowq = t4_rsrv_noflowq; #ifdef TCP_OFFLOAD if (is_offload(sc)) { iaq->nofldtxq10g = t4_nofldtxq10g; iaq->nofldtxq1g = t4_nofldtxq1g; iaq->nofldtxq_vi = t4_nofldtxq_vi; iaq->nofldrxq10g = nofldrxq10g = t4_nofldrxq10g; iaq->nofldrxq1g = nofldrxq1g = t4_nofldrxq1g; iaq->nofldrxq_vi = t4_nofldrxq_vi; } #endif #ifdef DEV_NETMAP iaq->nnmtxq_vi = t4_nnmtxq_vi; iaq->nnmrxq_vi = t4_nnmrxq_vi; #endif for (itype = INTR_MSIX; itype; itype >>= 1) { if ((itype & t4_intr_types) == 0) continue; /* not allowed */ if (itype == INTR_MSIX) navail = pci_msix_count(sc->dev); else if (itype == INTR_MSI) navail = pci_msi_count(sc->dev); else navail = 1; restart: if (navail == 0) continue; iaq->intr_type = itype; iaq->intr_flags_10g = 0; iaq->intr_flags_1g = 0; /* * Best option: an interrupt vector for errors, one for the * firmware event queue, and one for every rxq (NIC and TOE) of * every VI. The VIs that support netmap use the same * interrupts for the NIC rx queues and the netmap rx queues * because only one set of queues is active at a time. */ iaq->nirq = T4_EXTRA_INTR; iaq->nirq += n10g * (nrxq10g + nofldrxq10g); iaq->nirq += n1g * (nrxq1g + nofldrxq1g); iaq->nirq += (n10g + n1g) * (num_vis - 1) * max(iaq->nrxq_vi, iaq->nnmrxq_vi); /* See comment above. */ iaq->nirq += (n10g + n1g) * (num_vis - 1) * iaq->nofldrxq_vi; if (iaq->nirq <= navail && (itype != INTR_MSI || powerof2(iaq->nirq))) { iaq->intr_flags_10g = INTR_ALL; iaq->intr_flags_1g = INTR_ALL; goto allocate; } /* Disable the VIs (and netmap) if there aren't enough intrs */ if (num_vis > 1) { device_printf(sc->dev, "virtual interfaces disabled " "because num_vis=%u with current settings " "(nrxq10g=%u, nrxq1g=%u, nofldrxq10g=%u, " "nofldrxq1g=%u, nrxq_vi=%u nofldrxq_vi=%u, " "nnmrxq_vi=%u) would need %u interrupts but " "only %u are available.\n", num_vis, nrxq10g, nrxq1g, nofldrxq10g, nofldrxq1g, iaq->nrxq_vi, iaq->nofldrxq_vi, iaq->nnmrxq_vi, iaq->nirq, navail); num_vis = 1; iaq->ntxq_vi = iaq->nrxq_vi = 0; iaq->nofldtxq_vi = iaq->nofldrxq_vi = 0; iaq->nnmtxq_vi = iaq->nnmrxq_vi = 0; goto restart; } /* * Second best option: a vector for errors, one for the firmware * event queue, and vectors for either all the NIC rx queues or * all the TOE rx queues. The queues that don't get vectors * will forward their interrupts to those that do. */ iaq->nirq = T4_EXTRA_INTR; if (nrxq10g >= nofldrxq10g) { iaq->intr_flags_10g = INTR_RXQ; iaq->nirq += n10g * nrxq10g; } else { iaq->intr_flags_10g = INTR_OFLD_RXQ; iaq->nirq += n10g * nofldrxq10g; } if (nrxq1g >= nofldrxq1g) { iaq->intr_flags_1g = INTR_RXQ; iaq->nirq += n1g * nrxq1g; } else { iaq->intr_flags_1g = INTR_OFLD_RXQ; iaq->nirq += n1g * nofldrxq1g; } if (iaq->nirq <= navail && (itype != INTR_MSI || powerof2(iaq->nirq))) goto allocate; /* * Next best option: an interrupt vector for errors, one for the * firmware event queue, and at least one per main-VI. At this * point we know we'll have to downsize nrxq and/or nofldrxq to * fit what's available to us. */ iaq->nirq = T4_EXTRA_INTR; iaq->nirq += n10g + n1g; if (iaq->nirq <= navail) { int leftover = navail - iaq->nirq; if (n10g > 0) { int target = max(nrxq10g, nofldrxq10g); iaq->intr_flags_10g = nrxq10g >= nofldrxq10g ? INTR_RXQ : INTR_OFLD_RXQ; n = 1; while (n < target && leftover >= n10g) { leftover -= n10g; iaq->nirq += n10g; n++; } iaq->nrxq10g = min(n, nrxq10g); #ifdef TCP_OFFLOAD iaq->nofldrxq10g = min(n, nofldrxq10g); #endif } if (n1g > 0) { int target = max(nrxq1g, nofldrxq1g); iaq->intr_flags_1g = nrxq1g >= nofldrxq1g ? INTR_RXQ : INTR_OFLD_RXQ; n = 1; while (n < target && leftover >= n1g) { leftover -= n1g; iaq->nirq += n1g; n++; } iaq->nrxq1g = min(n, nrxq1g); #ifdef TCP_OFFLOAD iaq->nofldrxq1g = min(n, nofldrxq1g); #endif } if (itype != INTR_MSI || powerof2(iaq->nirq)) goto allocate; } /* * Least desirable option: one interrupt vector for everything. */ iaq->nirq = iaq->nrxq10g = iaq->nrxq1g = 1; iaq->intr_flags_10g = iaq->intr_flags_1g = 0; #ifdef TCP_OFFLOAD if (is_offload(sc)) iaq->nofldrxq10g = iaq->nofldrxq1g = 1; #endif allocate: navail = iaq->nirq; rc = 0; if (itype == INTR_MSIX) rc = pci_alloc_msix(sc->dev, &navail); else if (itype == INTR_MSI) rc = pci_alloc_msi(sc->dev, &navail); if (rc == 0) { if (navail == iaq->nirq) return (0); /* * Didn't get the number requested. Use whatever number * the kernel is willing to allocate (it's in navail). */ device_printf(sc->dev, "fewer vectors than requested, " "type=%d, req=%d, rcvd=%d; will downshift req.\n", itype, iaq->nirq, navail); pci_release_msi(sc->dev); goto restart; } device_printf(sc->dev, "failed to allocate vectors:%d, type=%d, req=%d, rcvd=%d\n", itype, rc, iaq->nirq, navail); } device_printf(sc->dev, "failed to find a usable interrupt type. " "allowed=%d, msi-x=%d, msi=%d, intx=1", t4_intr_types, pci_msix_count(sc->dev), pci_msi_count(sc->dev)); return (ENXIO); } #define FW_VERSION(chip) ( \ V_FW_HDR_FW_VER_MAJOR(chip##FW_VERSION_MAJOR) | \ V_FW_HDR_FW_VER_MINOR(chip##FW_VERSION_MINOR) | \ V_FW_HDR_FW_VER_MICRO(chip##FW_VERSION_MICRO) | \ V_FW_HDR_FW_VER_BUILD(chip##FW_VERSION_BUILD)) #define FW_INTFVER(chip, intf) (chip##FW_HDR_INTFVER_##intf) struct fw_info { uint8_t chip; char *kld_name; char *fw_mod_name; struct fw_hdr fw_hdr; /* XXX: waste of space, need a sparse struct */ } fw_info[] = { { .chip = CHELSIO_T4, .kld_name = "t4fw_cfg", .fw_mod_name = "t4fw", .fw_hdr = { .chip = FW_HDR_CHIP_T4, .fw_ver = htobe32_const(FW_VERSION(T4)), .intfver_nic = FW_INTFVER(T4, NIC), .intfver_vnic = FW_INTFVER(T4, VNIC), .intfver_ofld = FW_INTFVER(T4, OFLD), .intfver_ri = FW_INTFVER(T4, RI), .intfver_iscsipdu = FW_INTFVER(T4, ISCSIPDU), .intfver_iscsi = FW_INTFVER(T4, ISCSI), .intfver_fcoepdu = FW_INTFVER(T4, FCOEPDU), .intfver_fcoe = FW_INTFVER(T4, FCOE), }, }, { .chip = CHELSIO_T5, .kld_name = "t5fw_cfg", .fw_mod_name = "t5fw", .fw_hdr = { .chip = FW_HDR_CHIP_T5, .fw_ver = htobe32_const(FW_VERSION(T5)), .intfver_nic = FW_INTFVER(T5, NIC), .intfver_vnic = FW_INTFVER(T5, VNIC), .intfver_ofld = FW_INTFVER(T5, OFLD), .intfver_ri = FW_INTFVER(T5, RI), .intfver_iscsipdu = FW_INTFVER(T5, ISCSIPDU), .intfver_iscsi = FW_INTFVER(T5, ISCSI), .intfver_fcoepdu = FW_INTFVER(T5, FCOEPDU), .intfver_fcoe = FW_INTFVER(T5, FCOE), }, } }; static struct fw_info * find_fw_info(int chip) { int i; for (i = 0; i < nitems(fw_info); i++) { if (fw_info[i].chip == chip) return (&fw_info[i]); } return (NULL); } /* * Is the given firmware API compatible with the one the driver was compiled * with? */ static int fw_compatible(const struct fw_hdr *hdr1, const struct fw_hdr *hdr2) { /* short circuit if it's the exact same firmware version */ if (hdr1->chip == hdr2->chip && hdr1->fw_ver == hdr2->fw_ver) return (1); /* * XXX: Is this too conservative? Perhaps I should limit this to the * features that are supported in the driver. */ #define SAME_INTF(x) (hdr1->intfver_##x == hdr2->intfver_##x) if (hdr1->chip == hdr2->chip && SAME_INTF(nic) && SAME_INTF(vnic) && SAME_INTF(ofld) && SAME_INTF(ri) && SAME_INTF(iscsipdu) && SAME_INTF(iscsi) && SAME_INTF(fcoepdu) && SAME_INTF(fcoe)) return (1); #undef SAME_INTF return (0); } /* * The firmware in the KLD is usable, but should it be installed? This routine * explains itself in detail if it indicates the KLD firmware should be * installed. */ static int should_install_kld_fw(struct adapter *sc, int card_fw_usable, int k, int c) { const char *reason; if (!card_fw_usable) { reason = "incompatible or unusable"; goto install; } if (k > c) { reason = "older than the version bundled with this driver"; goto install; } if (t4_fw_install == 2 && k != c) { reason = "different than the version bundled with this driver"; goto install; } return (0); install: if (t4_fw_install == 0) { device_printf(sc->dev, "firmware on card (%u.%u.%u.%u) is %s, " "but the driver is prohibited from installing a different " "firmware on the card.\n", G_FW_HDR_FW_VER_MAJOR(c), G_FW_HDR_FW_VER_MINOR(c), G_FW_HDR_FW_VER_MICRO(c), G_FW_HDR_FW_VER_BUILD(c), reason); return (0); } device_printf(sc->dev, "firmware on card (%u.%u.%u.%u) is %s, " "installing firmware %u.%u.%u.%u on card.\n", G_FW_HDR_FW_VER_MAJOR(c), G_FW_HDR_FW_VER_MINOR(c), G_FW_HDR_FW_VER_MICRO(c), G_FW_HDR_FW_VER_BUILD(c), reason, G_FW_HDR_FW_VER_MAJOR(k), G_FW_HDR_FW_VER_MINOR(k), G_FW_HDR_FW_VER_MICRO(k), G_FW_HDR_FW_VER_BUILD(k)); return (1); } /* * Establish contact with the firmware and determine if we are the master driver * or not, and whether we are responsible for chip initialization. */ static int prep_firmware(struct adapter *sc) { const struct firmware *fw = NULL, *default_cfg; int rc, pf, card_fw_usable, kld_fw_usable, need_fw_reset = 1; enum dev_state state; struct fw_info *fw_info; struct fw_hdr *card_fw; /* fw on the card */ const struct fw_hdr *kld_fw; /* fw in the KLD */ const struct fw_hdr *drv_fw; /* fw header the driver was compiled against */ /* Contact firmware. */ rc = t4_fw_hello(sc, sc->mbox, sc->mbox, MASTER_MAY, &state); if (rc < 0 || state == DEV_STATE_ERR) { rc = -rc; device_printf(sc->dev, "failed to connect to the firmware: %d, %d.\n", rc, state); return (rc); } pf = rc; if (pf == sc->mbox) sc->flags |= MASTER_PF; else if (state == DEV_STATE_UNINIT) { /* * We didn't get to be the master so we definitely won't be * configuring the chip. It's a bug if someone else hasn't * configured it already. */ device_printf(sc->dev, "couldn't be master(%d), " "device not already initialized either(%d).\n", rc, state); return (EDOOFUS); } /* This is the firmware whose headers the driver was compiled against */ fw_info = find_fw_info(chip_id(sc)); if (fw_info == NULL) { device_printf(sc->dev, "unable to look up firmware information for chip %d.\n", chip_id(sc)); return (EINVAL); } drv_fw = &fw_info->fw_hdr; /* * The firmware KLD contains many modules. The KLD name is also the * name of the module that contains the default config file. */ default_cfg = firmware_get(fw_info->kld_name); /* Read the header of the firmware on the card */ card_fw = malloc(sizeof(*card_fw), M_CXGBE, M_ZERO | M_WAITOK); rc = -t4_read_flash(sc, FLASH_FW_START, sizeof (*card_fw) / sizeof (uint32_t), (uint32_t *)card_fw, 1); if (rc == 0) card_fw_usable = fw_compatible(drv_fw, (const void*)card_fw); else { device_printf(sc->dev, "Unable to read card's firmware header: %d\n", rc); card_fw_usable = 0; } /* This is the firmware in the KLD */ fw = firmware_get(fw_info->fw_mod_name); if (fw != NULL) { kld_fw = (const void *)fw->data; kld_fw_usable = fw_compatible(drv_fw, kld_fw); } else { kld_fw = NULL; kld_fw_usable = 0; } if (card_fw_usable && card_fw->fw_ver == drv_fw->fw_ver && (!kld_fw_usable || kld_fw->fw_ver == drv_fw->fw_ver)) { /* * Common case: the firmware on the card is an exact match and * the KLD is an exact match too, or the KLD is * absent/incompatible. Note that t4_fw_install = 2 is ignored * here -- use cxgbetool loadfw if you want to reinstall the * same firmware as the one on the card. */ } else if (kld_fw_usable && state == DEV_STATE_UNINIT && should_install_kld_fw(sc, card_fw_usable, be32toh(kld_fw->fw_ver), be32toh(card_fw->fw_ver))) { rc = -t4_fw_upgrade(sc, sc->mbox, fw->data, fw->datasize, 0); if (rc != 0) { device_printf(sc->dev, "failed to install firmware: %d\n", rc); goto done; } /* Installed successfully, update the cached header too. */ memcpy(card_fw, kld_fw, sizeof(*card_fw)); card_fw_usable = 1; need_fw_reset = 0; /* already reset as part of load_fw */ } if (!card_fw_usable) { uint32_t d, c, k; d = ntohl(drv_fw->fw_ver); c = ntohl(card_fw->fw_ver); k = kld_fw ? ntohl(kld_fw->fw_ver) : 0; device_printf(sc->dev, "Cannot find a usable firmware: " "fw_install %d, chip state %d, " "driver compiled with %d.%d.%d.%d, " "card has %d.%d.%d.%d, KLD has %d.%d.%d.%d\n", t4_fw_install, state, G_FW_HDR_FW_VER_MAJOR(d), G_FW_HDR_FW_VER_MINOR(d), G_FW_HDR_FW_VER_MICRO(d), G_FW_HDR_FW_VER_BUILD(d), G_FW_HDR_FW_VER_MAJOR(c), G_FW_HDR_FW_VER_MINOR(c), G_FW_HDR_FW_VER_MICRO(c), G_FW_HDR_FW_VER_BUILD(c), G_FW_HDR_FW_VER_MAJOR(k), G_FW_HDR_FW_VER_MINOR(k), G_FW_HDR_FW_VER_MICRO(k), G_FW_HDR_FW_VER_BUILD(k)); rc = EINVAL; goto done; } /* Reset device */ if (need_fw_reset && (rc = -t4_fw_reset(sc, sc->mbox, F_PIORSTMODE | F_PIORST)) != 0) { device_printf(sc->dev, "firmware reset failed: %d.\n", rc); if (rc != ETIMEDOUT && rc != EIO) t4_fw_bye(sc, sc->mbox); goto done; } sc->flags |= FW_OK; rc = get_params__pre_init(sc); if (rc != 0) goto done; /* error message displayed already */ /* Partition adapter resources as specified in the config file. */ if (state == DEV_STATE_UNINIT) { KASSERT(sc->flags & MASTER_PF, ("%s: trying to change chip settings when not master.", __func__)); rc = partition_resources(sc, default_cfg, fw_info->kld_name); if (rc != 0) goto done; /* error message displayed already */ t4_tweak_chip_settings(sc); /* get basic stuff going */ rc = -t4_fw_initialize(sc, sc->mbox); if (rc != 0) { device_printf(sc->dev, "fw init failed: %d.\n", rc); goto done; } } else { snprintf(sc->cfg_file, sizeof(sc->cfg_file), "pf%d", pf); sc->cfcsum = 0; } done: free(card_fw, M_CXGBE); if (fw != NULL) firmware_put(fw, FIRMWARE_UNLOAD); if (default_cfg != NULL) firmware_put(default_cfg, FIRMWARE_UNLOAD); return (rc); } #define FW_PARAM_DEV(param) \ (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | \ V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_##param)) #define FW_PARAM_PFVF(param) \ (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_PFVF) | \ V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_PFVF_##param)) /* * Partition chip resources for use between various PFs, VFs, etc. */ static int partition_resources(struct adapter *sc, const struct firmware *default_cfg, const char *name_prefix) { const struct firmware *cfg = NULL; int rc = 0; struct fw_caps_config_cmd caps; uint32_t mtype, moff, finicsum, cfcsum; /* * Figure out what configuration file to use. Pick the default config * file for the card if the user hasn't specified one explicitly. */ snprintf(sc->cfg_file, sizeof(sc->cfg_file), "%s", t4_cfg_file); if (strncmp(t4_cfg_file, DEFAULT_CF, sizeof(t4_cfg_file)) == 0) { /* Card specific overrides go here. */ if (pci_get_device(sc->dev) == 0x440a) snprintf(sc->cfg_file, sizeof(sc->cfg_file), UWIRE_CF); if (is_fpga(sc)) snprintf(sc->cfg_file, sizeof(sc->cfg_file), FPGA_CF); } /* * We need to load another module if the profile is anything except * "default" or "flash". */ if (strncmp(sc->cfg_file, DEFAULT_CF, sizeof(sc->cfg_file)) != 0 && strncmp(sc->cfg_file, FLASH_CF, sizeof(sc->cfg_file)) != 0) { char s[32]; snprintf(s, sizeof(s), "%s_%s", name_prefix, sc->cfg_file); cfg = firmware_get(s); if (cfg == NULL) { if (default_cfg != NULL) { device_printf(sc->dev, "unable to load module \"%s\" for " "configuration profile \"%s\", will use " "the default config file instead.\n", s, sc->cfg_file); snprintf(sc->cfg_file, sizeof(sc->cfg_file), "%s", DEFAULT_CF); } else { device_printf(sc->dev, "unable to load module \"%s\" for " "configuration profile \"%s\", will use " "the config file on the card's flash " "instead.\n", s, sc->cfg_file); snprintf(sc->cfg_file, sizeof(sc->cfg_file), "%s", FLASH_CF); } } } if (strncmp(sc->cfg_file, DEFAULT_CF, sizeof(sc->cfg_file)) == 0 && default_cfg == NULL) { device_printf(sc->dev, "default config file not available, will use the config " "file on the card's flash instead.\n"); snprintf(sc->cfg_file, sizeof(sc->cfg_file), "%s", FLASH_CF); } if (strncmp(sc->cfg_file, FLASH_CF, sizeof(sc->cfg_file)) != 0) { u_int cflen; const uint32_t *cfdata; uint32_t param, val, addr; KASSERT(cfg != NULL || default_cfg != NULL, ("%s: no config to upload", __func__)); /* * Ask the firmware where it wants us to upload the config file. */ param = FW_PARAM_DEV(CF); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); if (rc != 0) { /* No support for config file? Shouldn't happen. */ device_printf(sc->dev, "failed to query config file location: %d.\n", rc); goto done; } mtype = G_FW_PARAMS_PARAM_Y(val); moff = G_FW_PARAMS_PARAM_Z(val) << 16; /* * XXX: sheer laziness. We deliberately added 4 bytes of * useless stuffing/comments at the end of the config file so * it's ok to simply throw away the last remaining bytes when * the config file is not an exact multiple of 4. This also * helps with the validate_mt_off_len check. */ if (cfg != NULL) { cflen = cfg->datasize & ~3; cfdata = cfg->data; } else { cflen = default_cfg->datasize & ~3; cfdata = default_cfg->data; } if (cflen > FLASH_CFG_MAX_SIZE) { device_printf(sc->dev, "config file too long (%d, max allowed is %d). " "Will try to use the config on the card, if any.\n", cflen, FLASH_CFG_MAX_SIZE); goto use_config_on_flash; } rc = validate_mt_off_len(sc, mtype, moff, cflen, &addr); if (rc != 0) { device_printf(sc->dev, "%s: addr (%d/0x%x) or len %d is not valid: %d. " "Will try to use the config on the card, if any.\n", __func__, mtype, moff, cflen, rc); goto use_config_on_flash; } write_via_memwin(sc, 2, addr, cfdata, cflen); } else { use_config_on_flash: mtype = FW_MEMTYPE_FLASH; moff = t4_flash_cfg_addr(sc); } bzero(&caps, sizeof(caps)); caps.op_to_write = htobe32(V_FW_CMD_OP(FW_CAPS_CONFIG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); caps.cfvalid_to_len16 = htobe32(F_FW_CAPS_CONFIG_CMD_CFVALID | V_FW_CAPS_CONFIG_CMD_MEMTYPE_CF(mtype) | V_FW_CAPS_CONFIG_CMD_MEMADDR64K_CF(moff >> 16) | FW_LEN16(caps)); rc = -t4_wr_mbox(sc, sc->mbox, &caps, sizeof(caps), &caps); if (rc != 0) { device_printf(sc->dev, "failed to pre-process config file: %d " "(mtype %d, moff 0x%x).\n", rc, mtype, moff); goto done; } finicsum = be32toh(caps.finicsum); cfcsum = be32toh(caps.cfcsum); if (finicsum != cfcsum) { device_printf(sc->dev, "WARNING: config file checksum mismatch: %08x %08x\n", finicsum, cfcsum); } sc->cfcsum = cfcsum; #define LIMIT_CAPS(x) do { \ caps.x &= htobe16(t4_##x##_allowed); \ } while (0) /* * Let the firmware know what features will (not) be used so it can tune * things accordingly. */ LIMIT_CAPS(nbmcaps); LIMIT_CAPS(linkcaps); LIMIT_CAPS(switchcaps); LIMIT_CAPS(niccaps); LIMIT_CAPS(toecaps); LIMIT_CAPS(rdmacaps); LIMIT_CAPS(tlscaps); LIMIT_CAPS(iscsicaps); LIMIT_CAPS(fcoecaps); #undef LIMIT_CAPS caps.op_to_write = htobe32(V_FW_CMD_OP(FW_CAPS_CONFIG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE); caps.cfvalid_to_len16 = htobe32(FW_LEN16(caps)); rc = -t4_wr_mbox(sc, sc->mbox, &caps, sizeof(caps), NULL); if (rc != 0) { device_printf(sc->dev, "failed to process config file: %d.\n", rc); } done: if (cfg != NULL) firmware_put(cfg, FIRMWARE_UNLOAD); return (rc); } /* * Retrieve parameters that are needed (or nice to have) very early. */ static int get_params__pre_init(struct adapter *sc) { int rc; uint32_t param[2], val[2]; t4_get_version_info(sc); snprintf(sc->fw_version, sizeof(sc->fw_version), "%u.%u.%u.%u", G_FW_HDR_FW_VER_MAJOR(sc->params.fw_vers), G_FW_HDR_FW_VER_MINOR(sc->params.fw_vers), G_FW_HDR_FW_VER_MICRO(sc->params.fw_vers), G_FW_HDR_FW_VER_BUILD(sc->params.fw_vers)); snprintf(sc->bs_version, sizeof(sc->bs_version), "%u.%u.%u.%u", G_FW_HDR_FW_VER_MAJOR(sc->params.bs_vers), G_FW_HDR_FW_VER_MINOR(sc->params.bs_vers), G_FW_HDR_FW_VER_MICRO(sc->params.bs_vers), G_FW_HDR_FW_VER_BUILD(sc->params.bs_vers)); snprintf(sc->tp_version, sizeof(sc->tp_version), "%u.%u.%u.%u", G_FW_HDR_FW_VER_MAJOR(sc->params.tp_vers), G_FW_HDR_FW_VER_MINOR(sc->params.tp_vers), G_FW_HDR_FW_VER_MICRO(sc->params.tp_vers), G_FW_HDR_FW_VER_BUILD(sc->params.tp_vers)); snprintf(sc->er_version, sizeof(sc->er_version), "%u.%u.%u.%u", G_FW_HDR_FW_VER_MAJOR(sc->params.er_vers), G_FW_HDR_FW_VER_MINOR(sc->params.er_vers), G_FW_HDR_FW_VER_MICRO(sc->params.er_vers), G_FW_HDR_FW_VER_BUILD(sc->params.er_vers)); param[0] = FW_PARAM_DEV(PORTVEC); param[1] = FW_PARAM_DEV(CCLK); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 2, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query parameters (pre_init): %d.\n", rc); return (rc); } sc->params.portvec = val[0]; sc->params.nports = bitcount32(val[0]); sc->params.vpd.cclk = val[1]; /* Read device log parameters. */ rc = -t4_init_devlog_params(sc, 1); if (rc == 0) fixup_devlog_params(sc); else { device_printf(sc->dev, "failed to get devlog parameters: %d.\n", rc); rc = 0; /* devlog isn't critical for device operation */ } return (rc); } /* * Retrieve various parameters that are of interest to the driver. The device * has been initialized by the firmware at this point. */ static int get_params__post_init(struct adapter *sc) { int rc; uint32_t param[7], val[7]; struct fw_caps_config_cmd caps; param[0] = FW_PARAM_PFVF(IQFLINT_START); param[1] = FW_PARAM_PFVF(EQ_START); param[2] = FW_PARAM_PFVF(FILTER_START); param[3] = FW_PARAM_PFVF(FILTER_END); param[4] = FW_PARAM_PFVF(L2T_START); param[5] = FW_PARAM_PFVF(L2T_END); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 6, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query parameters (post_init): %d.\n", rc); return (rc); } sc->sge.iq_start = val[0]; sc->sge.eq_start = val[1]; sc->tids.ftid_base = val[2]; sc->tids.nftids = val[3] - val[2] + 1; sc->params.ftid_min = val[2]; sc->params.ftid_max = val[3]; sc->vres.l2t.start = val[4]; sc->vres.l2t.size = val[5] - val[4] + 1; KASSERT(sc->vres.l2t.size <= L2T_SIZE, ("%s: L2 table size (%u) larger than expected (%u)", __func__, sc->vres.l2t.size, L2T_SIZE)); /* get capabilites */ bzero(&caps, sizeof(caps)); caps.op_to_write = htobe32(V_FW_CMD_OP(FW_CAPS_CONFIG_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ); caps.cfvalid_to_len16 = htobe32(FW_LEN16(caps)); rc = -t4_wr_mbox(sc, sc->mbox, &caps, sizeof(caps), &caps); if (rc != 0) { device_printf(sc->dev, "failed to get card capabilities: %d.\n", rc); return (rc); } #define READ_CAPS(x) do { \ sc->x = htobe16(caps.x); \ } while (0) READ_CAPS(nbmcaps); READ_CAPS(linkcaps); READ_CAPS(switchcaps); READ_CAPS(niccaps); READ_CAPS(toecaps); READ_CAPS(rdmacaps); READ_CAPS(tlscaps); READ_CAPS(iscsicaps); READ_CAPS(fcoecaps); if (sc->niccaps & FW_CAPS_CONFIG_NIC_ETHOFLD) { param[0] = FW_PARAM_PFVF(ETHOFLD_START); param[1] = FW_PARAM_PFVF(ETHOFLD_END); param[2] = FW_PARAM_DEV(FLOWC_BUFFIFO_SZ); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 3, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query NIC parameters: %d.\n", rc); return (rc); } sc->tids.etid_base = val[0]; sc->params.etid_min = val[0]; sc->tids.netids = val[1] - val[0] + 1; sc->params.netids = sc->tids.netids; sc->params.eo_wr_cred = val[2]; sc->params.ethoffload = 1; } if (sc->toecaps) { /* query offload-related parameters */ param[0] = FW_PARAM_DEV(NTID); param[1] = FW_PARAM_PFVF(SERVER_START); param[2] = FW_PARAM_PFVF(SERVER_END); param[3] = FW_PARAM_PFVF(TDDP_START); param[4] = FW_PARAM_PFVF(TDDP_END); param[5] = FW_PARAM_DEV(FLOWC_BUFFIFO_SZ); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 6, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query TOE parameters: %d.\n", rc); return (rc); } sc->tids.ntids = val[0]; sc->tids.natids = min(sc->tids.ntids / 2, MAX_ATIDS); sc->tids.stid_base = val[1]; sc->tids.nstids = val[2] - val[1] + 1; sc->vres.ddp.start = val[3]; sc->vres.ddp.size = val[4] - val[3] + 1; sc->params.ofldq_wr_cred = val[5]; sc->params.offload = 1; } if (sc->rdmacaps) { param[0] = FW_PARAM_PFVF(STAG_START); param[1] = FW_PARAM_PFVF(STAG_END); param[2] = FW_PARAM_PFVF(RQ_START); param[3] = FW_PARAM_PFVF(RQ_END); param[4] = FW_PARAM_PFVF(PBL_START); param[5] = FW_PARAM_PFVF(PBL_END); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 6, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query RDMA parameters(1): %d.\n", rc); return (rc); } sc->vres.stag.start = val[0]; sc->vres.stag.size = val[1] - val[0] + 1; sc->vres.rq.start = val[2]; sc->vres.rq.size = val[3] - val[2] + 1; sc->vres.pbl.start = val[4]; sc->vres.pbl.size = val[5] - val[4] + 1; param[0] = FW_PARAM_PFVF(SQRQ_START); param[1] = FW_PARAM_PFVF(SQRQ_END); param[2] = FW_PARAM_PFVF(CQ_START); param[3] = FW_PARAM_PFVF(CQ_END); param[4] = FW_PARAM_PFVF(OCQ_START); param[5] = FW_PARAM_PFVF(OCQ_END); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 6, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query RDMA parameters(2): %d.\n", rc); return (rc); } sc->vres.qp.start = val[0]; sc->vres.qp.size = val[1] - val[0] + 1; sc->vres.cq.start = val[2]; sc->vres.cq.size = val[3] - val[2] + 1; sc->vres.ocq.start = val[4]; sc->vres.ocq.size = val[5] - val[4] + 1; } if (sc->iscsicaps) { param[0] = FW_PARAM_PFVF(ISCSI_START); param[1] = FW_PARAM_PFVF(ISCSI_END); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 2, param, val); if (rc != 0) { device_printf(sc->dev, "failed to query iSCSI parameters: %d.\n", rc); return (rc); } sc->vres.iscsi.start = val[0]; sc->vres.iscsi.size = val[1] - val[0] + 1; } t4_init_sge_params(sc); /* * We've got the params we wanted to query via the firmware. Now grab * some others directly from the chip. */ rc = t4_read_chip_settings(sc); return (rc); } static int set_params__post_init(struct adapter *sc) { uint32_t param, val; /* ask for encapsulated CPLs */ param = FW_PARAM_PFVF(CPLFW4MSG_ENCAP); val = 1; (void)t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); return (0); } #undef FW_PARAM_PFVF #undef FW_PARAM_DEV static void t4_set_desc(struct adapter *sc) { char buf[128]; struct adapter_params *p = &sc->params; snprintf(buf, sizeof(buf), "Chelsio %s", p->vpd.id); device_set_desc_copy(sc->dev, buf); } static void build_medialist(struct port_info *pi, struct ifmedia *media) { int m; PORT_LOCK(pi); ifmedia_removeall(media); m = IFM_ETHER | IFM_FDX; switch(pi->port_type) { case FW_PORT_TYPE_BT_XFI: case FW_PORT_TYPE_BT_XAUI: ifmedia_add(media, m | IFM_10G_T, 0, NULL); /* fall through */ case FW_PORT_TYPE_BT_SGMII: ifmedia_add(media, m | IFM_1000_T, 0, NULL); ifmedia_add(media, m | IFM_100_TX, 0, NULL); ifmedia_add(media, IFM_ETHER | IFM_AUTO, 0, NULL); ifmedia_set(media, IFM_ETHER | IFM_AUTO); break; case FW_PORT_TYPE_CX4: ifmedia_add(media, m | IFM_10G_CX4, 0, NULL); ifmedia_set(media, m | IFM_10G_CX4); break; case FW_PORT_TYPE_QSFP_10G: case FW_PORT_TYPE_SFP: case FW_PORT_TYPE_FIBER_XFI: case FW_PORT_TYPE_FIBER_XAUI: switch (pi->mod_type) { case FW_PORT_MOD_TYPE_LR: ifmedia_add(media, m | IFM_10G_LR, 0, NULL); ifmedia_set(media, m | IFM_10G_LR); break; case FW_PORT_MOD_TYPE_SR: ifmedia_add(media, m | IFM_10G_SR, 0, NULL); ifmedia_set(media, m | IFM_10G_SR); break; case FW_PORT_MOD_TYPE_LRM: ifmedia_add(media, m | IFM_10G_LRM, 0, NULL); ifmedia_set(media, m | IFM_10G_LRM); break; case FW_PORT_MOD_TYPE_TWINAX_PASSIVE: case FW_PORT_MOD_TYPE_TWINAX_ACTIVE: ifmedia_add(media, m | IFM_10G_TWINAX, 0, NULL); ifmedia_set(media, m | IFM_10G_TWINAX); break; case FW_PORT_MOD_TYPE_NONE: m &= ~IFM_FDX; ifmedia_add(media, m | IFM_NONE, 0, NULL); ifmedia_set(media, m | IFM_NONE); break; case FW_PORT_MOD_TYPE_NA: case FW_PORT_MOD_TYPE_ER: default: device_printf(pi->dev, "unknown port_type (%d), mod_type (%d)\n", pi->port_type, pi->mod_type); ifmedia_add(media, m | IFM_UNKNOWN, 0, NULL); ifmedia_set(media, m | IFM_UNKNOWN); break; } break; case FW_PORT_TYPE_QSFP: switch (pi->mod_type) { case FW_PORT_MOD_TYPE_LR: ifmedia_add(media, m | IFM_40G_LR4, 0, NULL); ifmedia_set(media, m | IFM_40G_LR4); break; case FW_PORT_MOD_TYPE_SR: ifmedia_add(media, m | IFM_40G_SR4, 0, NULL); ifmedia_set(media, m | IFM_40G_SR4); break; case FW_PORT_MOD_TYPE_TWINAX_PASSIVE: case FW_PORT_MOD_TYPE_TWINAX_ACTIVE: ifmedia_add(media, m | IFM_40G_CR4, 0, NULL); ifmedia_set(media, m | IFM_40G_CR4); break; case FW_PORT_MOD_TYPE_NONE: m &= ~IFM_FDX; ifmedia_add(media, m | IFM_NONE, 0, NULL); ifmedia_set(media, m | IFM_NONE); break; default: device_printf(pi->dev, "unknown port_type (%d), mod_type (%d)\n", pi->port_type, pi->mod_type); ifmedia_add(media, m | IFM_UNKNOWN, 0, NULL); ifmedia_set(media, m | IFM_UNKNOWN); break; } break; default: device_printf(pi->dev, "unknown port_type (%d), mod_type (%d)\n", pi->port_type, pi->mod_type); ifmedia_add(media, m | IFM_UNKNOWN, 0, NULL); ifmedia_set(media, m | IFM_UNKNOWN); break; } PORT_UNLOCK(pi); } #define FW_MAC_EXACT_CHUNK 7 /* * Program the port's XGMAC based on parameters in ifnet. The caller also * indicates which parameters should be programmed (the rest are left alone). */ int update_mac_settings(struct ifnet *ifp, int flags) { int rc = 0; struct vi_info *vi = ifp->if_softc; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; int mtu = -1, promisc = -1, allmulti = -1, vlanex = -1; ASSERT_SYNCHRONIZED_OP(sc); KASSERT(flags, ("%s: not told what to update.", __func__)); if (flags & XGMAC_MTU) mtu = ifp->if_mtu; if (flags & XGMAC_PROMISC) promisc = ifp->if_flags & IFF_PROMISC ? 1 : 0; if (flags & XGMAC_ALLMULTI) allmulti = ifp->if_flags & IFF_ALLMULTI ? 1 : 0; if (flags & XGMAC_VLANEX) vlanex = ifp->if_capenable & IFCAP_VLAN_HWTAGGING ? 1 : 0; if (flags & (XGMAC_MTU|XGMAC_PROMISC|XGMAC_ALLMULTI|XGMAC_VLANEX)) { rc = -t4_set_rxmode(sc, sc->mbox, vi->viid, mtu, promisc, allmulti, 1, vlanex, false); if (rc) { if_printf(ifp, "set_rxmode (%x) failed: %d\n", flags, rc); return (rc); } } if (flags & XGMAC_UCADDR) { uint8_t ucaddr[ETHER_ADDR_LEN]; bcopy(IF_LLADDR(ifp), ucaddr, sizeof(ucaddr)); rc = t4_change_mac(sc, sc->mbox, vi->viid, vi->xact_addr_filt, ucaddr, true, true); if (rc < 0) { rc = -rc; if_printf(ifp, "change_mac failed: %d\n", rc); return (rc); } else { vi->xact_addr_filt = rc; rc = 0; } } if (flags & XGMAC_MCADDRS) { const uint8_t *mcaddr[FW_MAC_EXACT_CHUNK]; int del = 1; uint64_t hash = 0; struct ifmultiaddr *ifma; int i = 0, j; if_maddr_rlock(ifp); TAILQ_FOREACH(ifma, &ifp->if_multiaddrs, ifma_link) { if (ifma->ifma_addr->sa_family != AF_LINK) continue; mcaddr[i] = LLADDR((struct sockaddr_dl *)ifma->ifma_addr); MPASS(ETHER_IS_MULTICAST(mcaddr[i])); i++; if (i == FW_MAC_EXACT_CHUNK) { rc = t4_alloc_mac_filt(sc, sc->mbox, vi->viid, del, i, mcaddr, NULL, &hash, 0); if (rc < 0) { rc = -rc; for (j = 0; j < i; j++) { if_printf(ifp, "failed to add mc address" " %02x:%02x:%02x:" "%02x:%02x:%02x rc=%d\n", mcaddr[j][0], mcaddr[j][1], mcaddr[j][2], mcaddr[j][3], mcaddr[j][4], mcaddr[j][5], rc); } goto mcfail; } del = 0; i = 0; } } if (i > 0) { rc = t4_alloc_mac_filt(sc, sc->mbox, vi->viid, del, i, mcaddr, NULL, &hash, 0); if (rc < 0) { rc = -rc; for (j = 0; j < i; j++) { if_printf(ifp, "failed to add mc address" " %02x:%02x:%02x:" "%02x:%02x:%02x rc=%d\n", mcaddr[j][0], mcaddr[j][1], mcaddr[j][2], mcaddr[j][3], mcaddr[j][4], mcaddr[j][5], rc); } goto mcfail; } } rc = -t4_set_addr_hash(sc, sc->mbox, vi->viid, 0, hash, 0); if (rc != 0) if_printf(ifp, "failed to set mc address hash: %d", rc); mcfail: if_maddr_runlock(ifp); } return (rc); } /* * {begin|end}_synchronized_op must be called from the same thread. */ int begin_synchronized_op(struct adapter *sc, struct vi_info *vi, int flags, char *wmesg) { int rc, pri; #ifdef WITNESS /* the caller thinks it's ok to sleep, but is it really? */ if (flags & SLEEP_OK) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "begin_synchronized_op"); #endif if (INTR_OK) pri = PCATCH; else pri = 0; ADAPTER_LOCK(sc); for (;;) { if (vi && IS_DOOMED(vi)) { rc = ENXIO; goto done; } if (!IS_BUSY(sc)) { rc = 0; break; } if (!(flags & SLEEP_OK)) { rc = EBUSY; goto done; } if (mtx_sleep(&sc->flags, &sc->sc_lock, pri, wmesg, 0)) { rc = EINTR; goto done; } } KASSERT(!IS_BUSY(sc), ("%s: controller busy.", __func__)); SET_BUSY(sc); #ifdef INVARIANTS sc->last_op = wmesg; sc->last_op_thr = curthread; sc->last_op_flags = flags; #endif done: if (!(flags & HOLD_LOCK) || rc) ADAPTER_UNLOCK(sc); return (rc); } /* * Tell if_ioctl and if_init that the VI is going away. This is * special variant of begin_synchronized_op and must be paired with a * call to end_synchronized_op. */ void doom_vi(struct adapter *sc, struct vi_info *vi) { ADAPTER_LOCK(sc); SET_DOOMED(vi); wakeup(&sc->flags); while (IS_BUSY(sc)) mtx_sleep(&sc->flags, &sc->sc_lock, 0, "t4detach", 0); SET_BUSY(sc); #ifdef INVARIANTS sc->last_op = "t4detach"; sc->last_op_thr = curthread; sc->last_op_flags = 0; #endif ADAPTER_UNLOCK(sc); } /* * {begin|end}_synchronized_op must be called from the same thread. */ void end_synchronized_op(struct adapter *sc, int flags) { if (flags & LOCK_HELD) ADAPTER_LOCK_ASSERT_OWNED(sc); else ADAPTER_LOCK(sc); KASSERT(IS_BUSY(sc), ("%s: controller not busy.", __func__)); CLR_BUSY(sc); wakeup(&sc->flags); ADAPTER_UNLOCK(sc); } static int cxgbe_init_synchronized(struct vi_info *vi) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct ifnet *ifp = vi->ifp; int rc = 0, i; struct sge_txq *txq; ASSERT_SYNCHRONIZED_OP(sc); if (ifp->if_drv_flags & IFF_DRV_RUNNING) return (0); /* already running */ if (!(sc->flags & FULL_INIT_DONE) && ((rc = adapter_full_init(sc)) != 0)) return (rc); /* error message displayed already */ if (!(vi->flags & VI_INIT_DONE) && ((rc = vi_full_init(vi)) != 0)) return (rc); /* error message displayed already */ rc = update_mac_settings(ifp, XGMAC_ALL); if (rc) goto done; /* error message displayed already */ rc = -t4_enable_vi(sc, sc->mbox, vi->viid, true, true); if (rc != 0) { if_printf(ifp, "enable_vi failed: %d\n", rc); goto done; } /* * Can't fail from this point onwards. Review cxgbe_uninit_synchronized * if this changes. */ for_each_txq(vi, i, txq) { TXQ_LOCK(txq); txq->eq.flags |= EQ_ENABLED; TXQ_UNLOCK(txq); } /* * The first iq of the first port to come up is used for tracing. */ if (sc->traceq < 0 && IS_MAIN_VI(vi)) { sc->traceq = sc->sge.rxq[vi->first_rxq].iq.abs_id; t4_write_reg(sc, is_t4(sc) ? A_MPS_TRC_RSS_CONTROL : A_MPS_T5_TRC_RSS_CONTROL, V_RSSCONTROL(pi->tx_chan) | V_QUEUENUMBER(sc->traceq)); pi->flags |= HAS_TRACEQ; } /* all ok */ PORT_LOCK(pi); ifp->if_drv_flags |= IFF_DRV_RUNNING; pi->up_vis++; if (pi->nvi > 1 || sc->flags & IS_VF) callout_reset(&vi->tick, hz, vi_tick, vi); else callout_reset(&pi->tick, hz, cxgbe_tick, pi); PORT_UNLOCK(pi); done: if (rc != 0) cxgbe_uninit_synchronized(vi); return (rc); } /* * Idempotent. */ static int cxgbe_uninit_synchronized(struct vi_info *vi) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct ifnet *ifp = vi->ifp; int rc, i; struct sge_txq *txq; ASSERT_SYNCHRONIZED_OP(sc); if (!(vi->flags & VI_INIT_DONE)) { KASSERT(!(ifp->if_drv_flags & IFF_DRV_RUNNING), ("uninited VI is running")); return (0); } /* * Disable the VI so that all its data in either direction is discarded * by the MPS. Leave everything else (the queues, interrupts, and 1Hz * tick) intact as the TP can deliver negative advice or data that it's * holding in its RAM (for an offloaded connection) even after the VI is * disabled. */ rc = -t4_enable_vi(sc, sc->mbox, vi->viid, false, false); if (rc) { if_printf(ifp, "disable_vi failed: %d\n", rc); return (rc); } for_each_txq(vi, i, txq) { TXQ_LOCK(txq); txq->eq.flags &= ~EQ_ENABLED; TXQ_UNLOCK(txq); } PORT_LOCK(pi); if (pi->nvi > 1 || sc->flags & IS_VF) callout_stop(&vi->tick); else callout_stop(&pi->tick); if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) { PORT_UNLOCK(pi); return (0); } ifp->if_drv_flags &= ~IFF_DRV_RUNNING; pi->up_vis--; if (pi->up_vis > 0) { PORT_UNLOCK(pi); return (0); } PORT_UNLOCK(pi); pi->link_cfg.link_ok = 0; pi->link_cfg.speed = 0; pi->linkdnrc = -1; t4_os_link_changed(sc, pi->port_id, 0, -1); return (0); } /* * It is ok for this function to fail midway and return right away. t4_detach * will walk the entire sc->irq list and clean up whatever is valid. */ int t4_setup_intr_handlers(struct adapter *sc) { int rc, rid, p, q, v; char s[8]; struct irq *irq; struct port_info *pi; struct vi_info *vi; struct sge *sge = &sc->sge; struct sge_rxq *rxq; #ifdef TCP_OFFLOAD struct sge_ofld_rxq *ofld_rxq; #endif #ifdef DEV_NETMAP struct sge_nm_rxq *nm_rxq; #endif #ifdef RSS int nbuckets = rss_getnumbuckets(); #endif /* * Setup interrupts. */ irq = &sc->irq[0]; rid = sc->intr_type == INTR_INTX ? 0 : 1; if (sc->intr_count == 1) return (t4_alloc_irq(sc, irq, rid, t4_intr_all, sc, "all")); /* Multiple interrupts. */ if (sc->flags & IS_VF) KASSERT(sc->intr_count >= T4VF_EXTRA_INTR + sc->params.nports, ("%s: too few intr.", __func__)); else KASSERT(sc->intr_count >= T4_EXTRA_INTR + sc->params.nports, ("%s: too few intr.", __func__)); /* The first one is always error intr on PFs */ if (!(sc->flags & IS_VF)) { rc = t4_alloc_irq(sc, irq, rid, t4_intr_err, sc, "err"); if (rc != 0) return (rc); irq++; rid++; } /* The second one is always the firmware event queue (first on VFs) */ rc = t4_alloc_irq(sc, irq, rid, t4_intr_evt, &sge->fwq, "evt"); if (rc != 0) return (rc); irq++; rid++; for_each_port(sc, p) { pi = sc->port[p]; for_each_vi(pi, v, vi) { vi->first_intr = rid - 1; if (vi->nnmrxq > 0) { int n = max(vi->nrxq, vi->nnmrxq); MPASS(vi->flags & INTR_RXQ); rxq = &sge->rxq[vi->first_rxq]; #ifdef DEV_NETMAP nm_rxq = &sge->nm_rxq[vi->first_nm_rxq]; #endif for (q = 0; q < n; q++) { snprintf(s, sizeof(s), "%x%c%x", p, 'a' + v, q); if (q < vi->nrxq) irq->rxq = rxq++; #ifdef DEV_NETMAP if (q < vi->nnmrxq) irq->nm_rxq = nm_rxq++; #endif rc = t4_alloc_irq(sc, irq, rid, t4_vi_intr, irq, s); if (rc != 0) return (rc); irq++; rid++; vi->nintr++; } } else if (vi->flags & INTR_RXQ) { for_each_rxq(vi, q, rxq) { snprintf(s, sizeof(s), "%x%c%x", p, 'a' + v, q); rc = t4_alloc_irq(sc, irq, rid, t4_intr, rxq, s); if (rc != 0) return (rc); #ifdef RSS bus_bind_intr(sc->dev, irq->res, rss_getcpu(q % nbuckets)); #endif irq++; rid++; vi->nintr++; } } #ifdef TCP_OFFLOAD if (vi->flags & INTR_OFLD_RXQ) { for_each_ofld_rxq(vi, q, ofld_rxq) { snprintf(s, sizeof(s), "%x%c%x", p, 'A' + v, q); rc = t4_alloc_irq(sc, irq, rid, t4_intr, ofld_rxq, s); if (rc != 0) return (rc); irq++; rid++; vi->nintr++; } } #endif } } MPASS(irq == &sc->irq[sc->intr_count]); return (0); } int adapter_full_init(struct adapter *sc) { int rc, i; ASSERT_SYNCHRONIZED_OP(sc); ADAPTER_LOCK_ASSERT_NOTOWNED(sc); KASSERT((sc->flags & FULL_INIT_DONE) == 0, ("%s: FULL_INIT_DONE already", __func__)); /* * queues that belong to the adapter (not any particular port). */ rc = t4_setup_adapter_queues(sc); if (rc != 0) goto done; for (i = 0; i < nitems(sc->tq); i++) { sc->tq[i] = taskqueue_create("t4 taskq", M_NOWAIT, taskqueue_thread_enqueue, &sc->tq[i]); if (sc->tq[i] == NULL) { device_printf(sc->dev, "failed to allocate task queue %d\n", i); rc = ENOMEM; goto done; } taskqueue_start_threads(&sc->tq[i], 1, PI_NET, "%s tq%d", device_get_nameunit(sc->dev), i); } if (!(sc->flags & IS_VF)) t4_intr_enable(sc); sc->flags |= FULL_INIT_DONE; done: if (rc != 0) adapter_full_uninit(sc); return (rc); } int adapter_full_uninit(struct adapter *sc) { int i; ADAPTER_LOCK_ASSERT_NOTOWNED(sc); t4_teardown_adapter_queues(sc); for (i = 0; i < nitems(sc->tq) && sc->tq[i]; i++) { taskqueue_free(sc->tq[i]); sc->tq[i] = NULL; } sc->flags &= ~FULL_INIT_DONE; return (0); } #ifdef RSS #define SUPPORTED_RSS_HASHTYPES (RSS_HASHTYPE_RSS_IPV4 | \ RSS_HASHTYPE_RSS_TCP_IPV4 | RSS_HASHTYPE_RSS_IPV6 | \ RSS_HASHTYPE_RSS_TCP_IPV6 | RSS_HASHTYPE_RSS_UDP_IPV4 | \ RSS_HASHTYPE_RSS_UDP_IPV6) /* Translates kernel hash types to hardware. */ static int hashconfig_to_hashen(int hashconfig) { int hashen = 0; if (hashconfig & RSS_HASHTYPE_RSS_IPV4) hashen |= F_FW_RSS_VI_CONFIG_CMD_IP4TWOTUPEN; if (hashconfig & RSS_HASHTYPE_RSS_IPV6) hashen |= F_FW_RSS_VI_CONFIG_CMD_IP6TWOTUPEN; if (hashconfig & RSS_HASHTYPE_RSS_UDP_IPV4) { hashen |= F_FW_RSS_VI_CONFIG_CMD_UDPEN | F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN; } if (hashconfig & RSS_HASHTYPE_RSS_UDP_IPV6) { hashen |= F_FW_RSS_VI_CONFIG_CMD_UDPEN | F_FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN; } if (hashconfig & RSS_HASHTYPE_RSS_TCP_IPV4) hashen |= F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN; if (hashconfig & RSS_HASHTYPE_RSS_TCP_IPV6) hashen |= F_FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN; return (hashen); } /* Translates hardware hash types to kernel. */ static int hashen_to_hashconfig(int hashen) { int hashconfig = 0; if (hashen & F_FW_RSS_VI_CONFIG_CMD_UDPEN) { /* * If UDP hashing was enabled it must have been enabled for * either IPv4 or IPv6 (inclusive or). Enabling UDP without * enabling any 4-tuple hash is nonsense configuration. */ MPASS(hashen & (F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN | F_FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN)); if (hashen & F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN) hashconfig |= RSS_HASHTYPE_RSS_UDP_IPV4; if (hashen & F_FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN) hashconfig |= RSS_HASHTYPE_RSS_UDP_IPV6; } if (hashen & F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN) hashconfig |= RSS_HASHTYPE_RSS_TCP_IPV4; if (hashen & F_FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN) hashconfig |= RSS_HASHTYPE_RSS_TCP_IPV6; if (hashen & F_FW_RSS_VI_CONFIG_CMD_IP4TWOTUPEN) hashconfig |= RSS_HASHTYPE_RSS_IPV4; if (hashen & F_FW_RSS_VI_CONFIG_CMD_IP6TWOTUPEN) hashconfig |= RSS_HASHTYPE_RSS_IPV6; return (hashconfig); } #endif int vi_full_init(struct vi_info *vi) { struct adapter *sc = vi->pi->adapter; struct ifnet *ifp = vi->ifp; uint16_t *rss; struct sge_rxq *rxq; int rc, i, j, hashen; #ifdef RSS int nbuckets = rss_getnumbuckets(); int hashconfig = rss_gethashconfig(); int extra; uint32_t raw_rss_key[RSS_KEYSIZE / sizeof(uint32_t)]; uint32_t rss_key[RSS_KEYSIZE / sizeof(uint32_t)]; #endif ASSERT_SYNCHRONIZED_OP(sc); KASSERT((vi->flags & VI_INIT_DONE) == 0, ("%s: VI_INIT_DONE already", __func__)); sysctl_ctx_init(&vi->ctx); vi->flags |= VI_SYSCTL_CTX; /* * Allocate tx/rx/fl queues for this VI. */ rc = t4_setup_vi_queues(vi); if (rc != 0) goto done; /* error message displayed already */ /* * Setup RSS for this VI. Save a copy of the RSS table for later use. */ if (vi->nrxq > vi->rss_size) { if_printf(ifp, "nrxq (%d) > hw RSS table size (%d); " "some queues will never receive traffic.\n", vi->nrxq, vi->rss_size); } else if (vi->rss_size % vi->nrxq) { if_printf(ifp, "nrxq (%d), hw RSS table size (%d); " "expect uneven traffic distribution.\n", vi->nrxq, vi->rss_size); } #ifdef RSS MPASS(RSS_KEYSIZE == 40); if (vi->nrxq != nbuckets) { if_printf(ifp, "nrxq (%d) != kernel RSS buckets (%d);" "performance will be impacted.\n", vi->nrxq, nbuckets); } rss_getkey((void *)&raw_rss_key[0]); for (i = 0; i < nitems(rss_key); i++) { rss_key[i] = htobe32(raw_rss_key[nitems(rss_key) - 1 - i]); } t4_write_rss_key(sc, &rss_key[0], -1); #endif rss = malloc(vi->rss_size * sizeof (*rss), M_CXGBE, M_ZERO | M_WAITOK); for (i = 0; i < vi->rss_size;) { #ifdef RSS j = rss_get_indirection_to_bucket(i); j %= vi->nrxq; rxq = &sc->sge.rxq[vi->first_rxq + j]; rss[i++] = rxq->iq.abs_id; #else for_each_rxq(vi, j, rxq) { rss[i++] = rxq->iq.abs_id; if (i == vi->rss_size) break; } #endif } rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, rss, vi->rss_size); if (rc != 0) { if_printf(ifp, "rss_config failed: %d\n", rc); goto done; } #ifdef RSS hashen = hashconfig_to_hashen(hashconfig); /* * We may have had to enable some hashes even though the global config * wants them disabled. This is a potential problem that must be * reported to the user. */ extra = hashen_to_hashconfig(hashen) ^ hashconfig; /* * If we consider only the supported hash types, then the enabled hashes * are a superset of the requested hashes. In other words, there cannot * be any supported hash that was requested but not enabled, but there * can be hashes that were not requested but had to be enabled. */ extra &= SUPPORTED_RSS_HASHTYPES; MPASS((extra & hashconfig) == 0); if (extra) { if_printf(ifp, "global RSS config (0x%x) cannot be accommodated.\n", hashconfig); } if (extra & RSS_HASHTYPE_RSS_IPV4) if_printf(ifp, "IPv4 2-tuple hashing forced on.\n"); if (extra & RSS_HASHTYPE_RSS_TCP_IPV4) if_printf(ifp, "TCP/IPv4 4-tuple hashing forced on.\n"); if (extra & RSS_HASHTYPE_RSS_IPV6) if_printf(ifp, "IPv6 2-tuple hashing forced on.\n"); if (extra & RSS_HASHTYPE_RSS_TCP_IPV6) if_printf(ifp, "TCP/IPv6 4-tuple hashing forced on.\n"); if (extra & RSS_HASHTYPE_RSS_UDP_IPV4) if_printf(ifp, "UDP/IPv4 4-tuple hashing forced on.\n"); if (extra & RSS_HASHTYPE_RSS_UDP_IPV6) if_printf(ifp, "UDP/IPv6 4-tuple hashing forced on.\n"); #else hashen = F_FW_RSS_VI_CONFIG_CMD_IP6FOURTUPEN | F_FW_RSS_VI_CONFIG_CMD_IP6TWOTUPEN | F_FW_RSS_VI_CONFIG_CMD_IP4FOURTUPEN | F_FW_RSS_VI_CONFIG_CMD_IP4TWOTUPEN | F_FW_RSS_VI_CONFIG_CMD_UDPEN; #endif rc = -t4_config_vi_rss(sc, sc->mbox, vi->viid, hashen, rss[0]); if (rc != 0) { if_printf(ifp, "rss hash/defaultq config failed: %d\n", rc); goto done; } vi->rss = rss; vi->flags |= VI_INIT_DONE; done: if (rc != 0) vi_full_uninit(vi); return (rc); } /* * Idempotent. */ int vi_full_uninit(struct vi_info *vi) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; int i; struct sge_rxq *rxq; struct sge_txq *txq; #ifdef TCP_OFFLOAD struct sge_ofld_rxq *ofld_rxq; struct sge_wrq *ofld_txq; #endif if (vi->flags & VI_INIT_DONE) { /* Need to quiesce queues. */ /* XXX: Only for the first VI? */ if (IS_MAIN_VI(vi) && !(sc->flags & IS_VF)) quiesce_wrq(sc, &sc->sge.ctrlq[pi->port_id]); for_each_txq(vi, i, txq) { quiesce_txq(sc, txq); } #ifdef TCP_OFFLOAD for_each_ofld_txq(vi, i, ofld_txq) { quiesce_wrq(sc, ofld_txq); } #endif for_each_rxq(vi, i, rxq) { quiesce_iq(sc, &rxq->iq); quiesce_fl(sc, &rxq->fl); } #ifdef TCP_OFFLOAD for_each_ofld_rxq(vi, i, ofld_rxq) { quiesce_iq(sc, &ofld_rxq->iq); quiesce_fl(sc, &ofld_rxq->fl); } #endif free(vi->rss, M_CXGBE); free(vi->nm_rss, M_CXGBE); } t4_teardown_vi_queues(vi); vi->flags &= ~VI_INIT_DONE; return (0); } static void quiesce_txq(struct adapter *sc, struct sge_txq *txq) { struct sge_eq *eq = &txq->eq; struct sge_qstat *spg = (void *)&eq->desc[eq->sidx]; (void) sc; /* unused */ #ifdef INVARIANTS TXQ_LOCK(txq); MPASS((eq->flags & EQ_ENABLED) == 0); TXQ_UNLOCK(txq); #endif /* Wait for the mp_ring to empty. */ while (!mp_ring_is_idle(txq->r)) { mp_ring_check_drainage(txq->r, 0); pause("rquiesce", 1); } /* Then wait for the hardware to finish. */ while (spg->cidx != htobe16(eq->pidx)) pause("equiesce", 1); /* Finally, wait for the driver to reclaim all descriptors. */ while (eq->cidx != eq->pidx) pause("dquiesce", 1); } static void quiesce_wrq(struct adapter *sc, struct sge_wrq *wrq) { /* XXXTX */ } static void quiesce_iq(struct adapter *sc, struct sge_iq *iq) { (void) sc; /* unused */ /* Synchronize with the interrupt handler */ while (!atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_DISABLED)) pause("iqfree", 1); } static void quiesce_fl(struct adapter *sc, struct sge_fl *fl) { mtx_lock(&sc->sfl_lock); FL_LOCK(fl); fl->flags |= FL_DOOMED; FL_UNLOCK(fl); callout_stop(&sc->sfl_callout); mtx_unlock(&sc->sfl_lock); KASSERT((fl->flags & FL_STARVING) == 0, ("%s: still starving", __func__)); } static int t4_alloc_irq(struct adapter *sc, struct irq *irq, int rid, driver_intr_t *handler, void *arg, char *name) { int rc; irq->rid = rid; irq->res = bus_alloc_resource_any(sc->dev, SYS_RES_IRQ, &irq->rid, RF_SHAREABLE | RF_ACTIVE); if (irq->res == NULL) { device_printf(sc->dev, "failed to allocate IRQ for rid %d, name %s.\n", rid, name); return (ENOMEM); } rc = bus_setup_intr(sc->dev, irq->res, INTR_MPSAFE | INTR_TYPE_NET, NULL, handler, arg, &irq->tag); if (rc != 0) { device_printf(sc->dev, "failed to setup interrupt for rid %d, name %s: %d\n", rid, name, rc); } else if (name) bus_describe_intr(sc->dev, irq->res, irq->tag, "%s", name); return (rc); } static int t4_free_irq(struct adapter *sc, struct irq *irq) { if (irq->tag) bus_teardown_intr(sc->dev, irq->res, irq->tag); if (irq->res) bus_release_resource(sc->dev, SYS_RES_IRQ, irq->rid, irq->res); bzero(irq, sizeof(*irq)); return (0); } static void get_regs(struct adapter *sc, struct t4_regdump *regs, uint8_t *buf) { regs->version = chip_id(sc) | chip_rev(sc) << 10; t4_get_regs(sc, buf, regs->len); } #define A_PL_INDIR_CMD 0x1f8 #define S_PL_AUTOINC 31 #define M_PL_AUTOINC 0x1U #define V_PL_AUTOINC(x) ((x) << S_PL_AUTOINC) #define G_PL_AUTOINC(x) (((x) >> S_PL_AUTOINC) & M_PL_AUTOINC) #define S_PL_VFID 20 #define M_PL_VFID 0xffU #define V_PL_VFID(x) ((x) << S_PL_VFID) #define G_PL_VFID(x) (((x) >> S_PL_VFID) & M_PL_VFID) #define S_PL_ADDR 0 #define M_PL_ADDR 0xfffffU #define V_PL_ADDR(x) ((x) << S_PL_ADDR) #define G_PL_ADDR(x) (((x) >> S_PL_ADDR) & M_PL_ADDR) #define A_PL_INDIR_DATA 0x1fc static uint64_t read_vf_stat(struct adapter *sc, unsigned int viid, int reg) { u32 stats[2]; mtx_assert(&sc->reg_lock, MA_OWNED); if (sc->flags & IS_VF) { stats[0] = t4_read_reg(sc, VF_MPS_REG(reg)); stats[1] = t4_read_reg(sc, VF_MPS_REG(reg + 4)); } else { t4_write_reg(sc, A_PL_INDIR_CMD, V_PL_AUTOINC(1) | V_PL_VFID(G_FW_VIID_VIN(viid)) | V_PL_ADDR(VF_MPS_REG(reg))); stats[0] = t4_read_reg(sc, A_PL_INDIR_DATA); stats[1] = t4_read_reg(sc, A_PL_INDIR_DATA); } return (((uint64_t)stats[1]) << 32 | stats[0]); } static void t4_get_vi_stats(struct adapter *sc, unsigned int viid, struct fw_vi_stats_vf *stats) { #define GET_STAT(name) \ read_vf_stat(sc, viid, A_MPS_VF_STAT_##name##_L) stats->tx_bcast_bytes = GET_STAT(TX_VF_BCAST_BYTES); stats->tx_bcast_frames = GET_STAT(TX_VF_BCAST_FRAMES); stats->tx_mcast_bytes = GET_STAT(TX_VF_MCAST_BYTES); stats->tx_mcast_frames = GET_STAT(TX_VF_MCAST_FRAMES); stats->tx_ucast_bytes = GET_STAT(TX_VF_UCAST_BYTES); stats->tx_ucast_frames = GET_STAT(TX_VF_UCAST_FRAMES); stats->tx_drop_frames = GET_STAT(TX_VF_DROP_FRAMES); stats->tx_offload_bytes = GET_STAT(TX_VF_OFFLOAD_BYTES); stats->tx_offload_frames = GET_STAT(TX_VF_OFFLOAD_FRAMES); stats->rx_bcast_bytes = GET_STAT(RX_VF_BCAST_BYTES); stats->rx_bcast_frames = GET_STAT(RX_VF_BCAST_FRAMES); stats->rx_mcast_bytes = GET_STAT(RX_VF_MCAST_BYTES); stats->rx_mcast_frames = GET_STAT(RX_VF_MCAST_FRAMES); stats->rx_ucast_bytes = GET_STAT(RX_VF_UCAST_BYTES); stats->rx_ucast_frames = GET_STAT(RX_VF_UCAST_FRAMES); stats->rx_err_frames = GET_STAT(RX_VF_ERR_FRAMES); #undef GET_STAT } static void t4_clr_vi_stats(struct adapter *sc, unsigned int viid) { int reg; t4_write_reg(sc, A_PL_INDIR_CMD, V_PL_AUTOINC(1) | V_PL_VFID(G_FW_VIID_VIN(viid)) | V_PL_ADDR(VF_MPS_REG(A_MPS_VF_STAT_TX_VF_BCAST_BYTES_L))); for (reg = A_MPS_VF_STAT_TX_VF_BCAST_BYTES_L; reg <= A_MPS_VF_STAT_RX_VF_ERR_FRAMES_H; reg += 4) t4_write_reg(sc, A_PL_INDIR_DATA, 0); } static void vi_refresh_stats(struct adapter *sc, struct vi_info *vi) { struct timeval tv; const struct timeval interval = {0, 250000}; /* 250ms */ if (!(vi->flags & VI_INIT_DONE)) return; getmicrotime(&tv); timevalsub(&tv, &interval); if (timevalcmp(&tv, &vi->last_refreshed, <)) return; mtx_lock(&sc->reg_lock); t4_get_vi_stats(sc, vi->viid, &vi->stats); getmicrotime(&vi->last_refreshed); mtx_unlock(&sc->reg_lock); } static void cxgbe_refresh_stats(struct adapter *sc, struct port_info *pi) { int i; u_int v, tnl_cong_drops; struct timeval tv; const struct timeval interval = {0, 250000}; /* 250ms */ getmicrotime(&tv); timevalsub(&tv, &interval); if (timevalcmp(&tv, &pi->last_refreshed, <)) return; tnl_cong_drops = 0; t4_get_port_stats(sc, pi->tx_chan, &pi->stats); for (i = 0; i < sc->chip_params->nchan; i++) { if (pi->rx_chan_map & (1 << i)) { mtx_lock(&sc->reg_lock); t4_read_indirect(sc, A_TP_MIB_INDEX, A_TP_MIB_DATA, &v, 1, A_TP_MIB_TNL_CNG_DROP_0 + i); mtx_unlock(&sc->reg_lock); tnl_cong_drops += v; } } pi->tnl_cong_drops = tnl_cong_drops; getmicrotime(&pi->last_refreshed); } static void cxgbe_tick(void *arg) { struct port_info *pi = arg; struct adapter *sc = pi->adapter; PORT_LOCK_ASSERT_OWNED(pi); cxgbe_refresh_stats(sc, pi); callout_schedule(&pi->tick, hz); } void vi_tick(void *arg) { struct vi_info *vi = arg; struct adapter *sc = vi->pi->adapter; vi_refresh_stats(sc, vi); callout_schedule(&vi->tick, hz); } static void cxgbe_vlan_config(void *arg, struct ifnet *ifp, uint16_t vid) { struct ifnet *vlan; if (arg != ifp || ifp->if_type != IFT_ETHER) return; vlan = VLAN_DEVAT(ifp, vid); VLAN_SETCOOKIE(vlan, ifp); } /* * Should match fw_caps_config_ enums in t4fw_interface.h */ static char *caps_decoder[] = { "\20\001IPMI\002NCSI", /* 0: NBM */ "\20\001PPP\002QFC\003DCBX", /* 1: link */ "\20\001INGRESS\002EGRESS", /* 2: switch */ "\20\001NIC\002VM\003IDS\004UM\005UM_ISGL" /* 3: NIC */ "\006HASHFILTER\007ETHOFLD", "\20\001TOE", /* 4: TOE */ "\20\001RDDP\002RDMAC", /* 5: RDMA */ "\20\001INITIATOR_PDU\002TARGET_PDU" /* 6: iSCSI */ "\003INITIATOR_CNXOFLD\004TARGET_CNXOFLD" "\005INITIATOR_SSNOFLD\006TARGET_SSNOFLD" "\007T10DIF" "\010INITIATOR_CMDOFLD\011TARGET_CMDOFLD", "\20\00KEYS", /* 7: TLS */ "\20\001INITIATOR\002TARGET\003CTRL_OFLD" /* 8: FCoE */ "\004PO_INITIATOR\005PO_TARGET", }; void t4_sysctls(struct adapter *sc) { struct sysctl_ctx_list *ctx; struct sysctl_oid *oid; struct sysctl_oid_list *children, *c0; static char *doorbells = {"\20\1UDB\2WCWR\3UDBWC\4KDB"}; ctx = device_get_sysctl_ctx(sc->dev); /* * dev.t4nex.X. */ oid = device_get_sysctl_tree(sc->dev); c0 = children = SYSCTL_CHILDREN(oid); sc->sc_do_rxcopy = 1; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "do_rx_copy", CTLFLAG_RW, &sc->sc_do_rxcopy, 1, "Do RX copy of small frames"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nports", CTLFLAG_RD, NULL, sc->params.nports, "# of ports"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "doorbells", CTLTYPE_STRING | CTLFLAG_RD, doorbells, sc->doorbells, sysctl_bitfield, "A", "available doorbells"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "core_clock", CTLFLAG_RD, NULL, sc->params.vpd.cclk, "core clock frequency (in KHz)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "holdoff_timers", CTLTYPE_STRING | CTLFLAG_RD, sc->params.sge.timer_val, sizeof(sc->params.sge.timer_val), sysctl_int_array, "A", "interrupt holdoff timer values (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "holdoff_pkt_counts", CTLTYPE_STRING | CTLFLAG_RD, sc->params.sge.counter_val, sizeof(sc->params.sge.counter_val), sysctl_int_array, "A", "interrupt holdoff packet counter values"); t4_sge_sysctls(sc, ctx, children); sc->lro_timeout = 100; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "lro_timeout", CTLFLAG_RW, &sc->lro_timeout, 0, "lro inactive-flush timeout (in us)"); - SYSCTL_ADD_INT(ctx, children, OID_AUTO, "debug_flags", CTLFLAG_RW, + SYSCTL_ADD_INT(ctx, children, OID_AUTO, "dflags", CTLFLAG_RW, &sc->debug_flags, 0, "flags to enable runtime debugging"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "tp_version", CTLFLAG_RD, sc->tp_version, 0, "TP microcode version"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "firmware_version", CTLFLAG_RD, sc->fw_version, 0, "firmware version"); if (sc->flags & IS_VF) return; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "hw_revision", CTLFLAG_RD, NULL, chip_rev(sc), "chip hardware revision"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "sn", CTLFLAG_RD, sc->params.vpd.sn, 0, "serial number"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "pn", CTLFLAG_RD, sc->params.vpd.pn, 0, "part number"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "ec", CTLFLAG_RD, sc->params.vpd.ec, 0, "engineering change"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "na", CTLFLAG_RD, sc->params.vpd.na, 0, "network address"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "er_version", CTLFLAG_RD, sc->er_version, 0, "expansion ROM version"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "bs_version", CTLFLAG_RD, sc->bs_version, 0, "bootstrap firmware version"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "scfg_version", CTLFLAG_RD, NULL, sc->params.scfg_vers, "serial config version"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "vpd_version", CTLFLAG_RD, NULL, sc->params.vpd_vers, "VPD version"); SYSCTL_ADD_STRING(ctx, children, OID_AUTO, "cf", CTLFLAG_RD, sc->cfg_file, 0, "configuration file"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cfcsum", CTLFLAG_RD, NULL, sc->cfcsum, "config file checksum"); #define SYSCTL_CAP(name, n, text) \ SYSCTL_ADD_PROC(ctx, children, OID_AUTO, #name, \ CTLTYPE_STRING | CTLFLAG_RD, caps_decoder[n], sc->name, \ sysctl_bitfield, "A", "available " text " capabilities") SYSCTL_CAP(nbmcaps, 0, "NBM"); SYSCTL_CAP(linkcaps, 1, "link"); SYSCTL_CAP(switchcaps, 2, "switch"); SYSCTL_CAP(niccaps, 3, "NIC"); SYSCTL_CAP(toecaps, 4, "TCP offload"); SYSCTL_CAP(rdmacaps, 5, "RDMA"); SYSCTL_CAP(iscsicaps, 6, "iSCSI"); SYSCTL_CAP(tlscaps, 7, "TLS"); SYSCTL_CAP(fcoecaps, 8, "FCoE"); #undef SYSCTL_CAP SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nfilters", CTLFLAG_RD, NULL, sc->tids.nftids, "number of filters"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "temperature", CTLTYPE_INT | CTLFLAG_RD, sc, 0, sysctl_temperature, "I", "chip temperature (in Celsius)"); #ifdef SBUF_DRAIN /* * dev.t4nex.X.misc. Marked CTLFLAG_SKIP to avoid information overload. */ oid = SYSCTL_ADD_NODE(ctx, c0, OID_AUTO, "misc", CTLFLAG_RD | CTLFLAG_SKIP, NULL, "logs and miscellaneous information"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cctrl", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_cctrl, "A", "congestion control"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ibq_tp0", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_cim_ibq_obq, "A", "CIM IBQ 0 (TP0)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ibq_tp1", CTLTYPE_STRING | CTLFLAG_RD, sc, 1, sysctl_cim_ibq_obq, "A", "CIM IBQ 1 (TP1)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ibq_ulp", CTLTYPE_STRING | CTLFLAG_RD, sc, 2, sysctl_cim_ibq_obq, "A", "CIM IBQ 2 (ULP)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ibq_sge0", CTLTYPE_STRING | CTLFLAG_RD, sc, 3, sysctl_cim_ibq_obq, "A", "CIM IBQ 3 (SGE0)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ibq_sge1", CTLTYPE_STRING | CTLFLAG_RD, sc, 4, sysctl_cim_ibq_obq, "A", "CIM IBQ 4 (SGE1)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ibq_ncsi", CTLTYPE_STRING | CTLFLAG_RD, sc, 5, sysctl_cim_ibq_obq, "A", "CIM IBQ 5 (NCSI)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_la", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, chip_id(sc) <= CHELSIO_T5 ? sysctl_cim_la : sysctl_cim_la_t6, "A", "CIM logic analyzer"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_ma_la", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_cim_ma_la, "A", "CIM MA logic analyzer"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_ulp0", CTLTYPE_STRING | CTLFLAG_RD, sc, 0 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 0 (ULP0)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_ulp1", CTLTYPE_STRING | CTLFLAG_RD, sc, 1 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 1 (ULP1)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_ulp2", CTLTYPE_STRING | CTLFLAG_RD, sc, 2 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 2 (ULP2)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_ulp3", CTLTYPE_STRING | CTLFLAG_RD, sc, 3 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 3 (ULP3)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_sge", CTLTYPE_STRING | CTLFLAG_RD, sc, 4 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 4 (SGE)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_ncsi", CTLTYPE_STRING | CTLFLAG_RD, sc, 5 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 5 (NCSI)"); if (chip_id(sc) > CHELSIO_T4) { SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_sge0_rx", CTLTYPE_STRING | CTLFLAG_RD, sc, 6 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 6 (SGE0-RX)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_obq_sge1_rx", CTLTYPE_STRING | CTLFLAG_RD, sc, 7 + CIM_NUM_IBQ, sysctl_cim_ibq_obq, "A", "CIM OBQ 7 (SGE1-RX)"); } SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_pif_la", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_cim_pif_la, "A", "CIM PIF logic analyzer"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cim_qcfg", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_cim_qcfg, "A", "CIM queue configuration"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cpl_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_cpl_stats, "A", "CPL statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "ddp_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_ddp_stats, "A", "non-TCP DDP statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "devlog", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_devlog, "A", "firmware's device log"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "fcoe_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_fcoe_stats, "A", "FCoE statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "hw_sched", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_hw_sched, "A", "hardware scheduler "); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "l2t", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_l2t, "A", "hardware L2 table"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "lb_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_lb_stats, "A", "loopback statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "meminfo", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_meminfo, "A", "memory regions"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "mps_tcam", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, chip_id(sc) <= CHELSIO_T5 ? sysctl_mps_tcam : sysctl_mps_tcam_t6, "A", "MPS TCAM entries"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "path_mtus", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_path_mtus, "A", "path MTUs"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pm_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_pm_stats, "A", "PM statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "rdma_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_rdma_stats, "A", "RDMA statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tcp_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_tcp_stats, "A", "TCP statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tids", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_tids, "A", "TID information"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tp_err_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_tp_err_stats, "A", "TP error statistics"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tp_la_mask", CTLTYPE_INT | CTLFLAG_RW, sc, 0, sysctl_tp_la_mask, "I", "TP logic analyzer event capture mask"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tp_la", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_tp_la, "A", "TP logic analyzer"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "tx_rate", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_tx_rate, "A", "Tx rate"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "ulprx_la", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_ulprx_la, "A", "ULPRX logic analyzer"); if (is_t5(sc)) { SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "wcwr_stats", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_wcwr_stats, "A", "write combined work requests"); } #endif #ifdef TCP_OFFLOAD if (is_offload(sc)) { /* * dev.t4nex.X.toe. */ oid = SYSCTL_ADD_NODE(ctx, c0, OID_AUTO, "toe", CTLFLAG_RD, NULL, "TOE parameters"); children = SYSCTL_CHILDREN(oid); sc->tt.sndbuf = 256 * 1024; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "sndbuf", CTLFLAG_RW, &sc->tt.sndbuf, 0, "max hardware send buffer size"); sc->tt.ddp = 0; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "ddp", CTLFLAG_RW, &sc->tt.ddp, 0, "DDP allowed"); sc->tt.rx_coalesce = 1; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "rx_coalesce", CTLFLAG_RW, &sc->tt.rx_coalesce, 0, "receive coalescing"); sc->tt.tx_align = 1; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "tx_align", CTLFLAG_RW, &sc->tt.tx_align, 0, "chop and align payload"); sc->tt.tx_zcopy = 0; SYSCTL_ADD_INT(ctx, children, OID_AUTO, "tx_zcopy", CTLFLAG_RW, &sc->tt.tx_zcopy, 0, "Enable zero-copy aio_write(2)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "timer_tick", CTLTYPE_STRING | CTLFLAG_RD, sc, 0, sysctl_tp_tick, "A", "TP timer tick (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "timestamp_tick", CTLTYPE_STRING | CTLFLAG_RD, sc, 1, sysctl_tp_tick, "A", "TCP timestamp tick (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "dack_tick", CTLTYPE_STRING | CTLFLAG_RD, sc, 2, sysctl_tp_tick, "A", "DACK tick (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "dack_timer", CTLTYPE_UINT | CTLFLAG_RD, sc, 0, sysctl_tp_dack_timer, "IU", "DACK timer (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "rexmt_min", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_RXT_MIN, sysctl_tp_timer, "LU", "Retransmit min (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "rexmt_max", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_RXT_MAX, sysctl_tp_timer, "LU", "Retransmit max (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "persist_min", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_PERS_MIN, sysctl_tp_timer, "LU", "Persist timer min (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "persist_max", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_PERS_MAX, sysctl_tp_timer, "LU", "Persist timer max (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "keepalive_idle", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_KEEP_IDLE, sysctl_tp_timer, "LU", "Keepidle idle timer (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "keepalive_intvl", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_KEEP_INTVL, sysctl_tp_timer, "LU", "Keepidle interval (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "initial_srtt", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_INIT_SRTT, sysctl_tp_timer, "LU", "Initial SRTT (us)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "finwait2_timer", CTLTYPE_ULONG | CTLFLAG_RD, sc, A_TP_FINWAIT2_TIMER, sysctl_tp_timer, "LU", "FINWAIT2 timer (us)"); } #endif } void vi_sysctls(struct vi_info *vi) { struct sysctl_ctx_list *ctx; struct sysctl_oid *oid; struct sysctl_oid_list *children; ctx = device_get_sysctl_ctx(vi->dev); /* * dev.v?(cxgbe|cxl).X. */ oid = device_get_sysctl_tree(vi->dev); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "viid", CTLFLAG_RD, NULL, vi->viid, "VI identifer"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nrxq", CTLFLAG_RD, &vi->nrxq, 0, "# of rx queues"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "ntxq", CTLFLAG_RD, &vi->ntxq, 0, "# of tx queues"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "first_rxq", CTLFLAG_RD, &vi->first_rxq, 0, "index of first rx queue"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "first_txq", CTLFLAG_RD, &vi->first_txq, 0, "index of first tx queue"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rss_size", CTLFLAG_RD, NULL, vi->rss_size, "size of RSS indirection table"); if (IS_MAIN_VI(vi)) { SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "rsrv_noflowq", CTLTYPE_INT | CTLFLAG_RW, vi, 0, sysctl_noflowq, "IU", "Reserve queue 0 for non-flowid packets"); } #ifdef TCP_OFFLOAD if (vi->nofldrxq != 0) { SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nofldrxq", CTLFLAG_RD, &vi->nofldrxq, 0, "# of rx queues for offloaded TCP connections"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nofldtxq", CTLFLAG_RD, &vi->nofldtxq, 0, "# of tx queues for offloaded TCP connections"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "first_ofld_rxq", CTLFLAG_RD, &vi->first_ofld_rxq, 0, "index of first TOE rx queue"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "first_ofld_txq", CTLFLAG_RD, &vi->first_ofld_txq, 0, "index of first TOE tx queue"); } #endif #ifdef DEV_NETMAP if (vi->nnmrxq != 0) { SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nnmrxq", CTLFLAG_RD, &vi->nnmrxq, 0, "# of netmap rx queues"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "nnmtxq", CTLFLAG_RD, &vi->nnmtxq, 0, "# of netmap tx queues"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "first_nm_rxq", CTLFLAG_RD, &vi->first_nm_rxq, 0, "index of first netmap rx queue"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "first_nm_txq", CTLFLAG_RD, &vi->first_nm_txq, 0, "index of first netmap tx queue"); } #endif SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "holdoff_tmr_idx", CTLTYPE_INT | CTLFLAG_RW, vi, 0, sysctl_holdoff_tmr_idx, "I", "holdoff timer index"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "holdoff_pktc_idx", CTLTYPE_INT | CTLFLAG_RW, vi, 0, sysctl_holdoff_pktc_idx, "I", "holdoff packet counter index"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "qsize_rxq", CTLTYPE_INT | CTLFLAG_RW, vi, 0, sysctl_qsize_rxq, "I", "rx queue size"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "qsize_txq", CTLTYPE_INT | CTLFLAG_RW, vi, 0, sysctl_qsize_txq, "I", "tx queue size"); } static void cxgbe_sysctls(struct port_info *pi) { struct sysctl_ctx_list *ctx; struct sysctl_oid *oid; struct sysctl_oid_list *children, *children2; struct adapter *sc = pi->adapter; int i; char name[16]; ctx = device_get_sysctl_ctx(pi->dev); /* * dev.cxgbe.X. */ oid = device_get_sysctl_tree(pi->dev); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "linkdnrc", CTLTYPE_STRING | CTLFLAG_RD, pi, 0, sysctl_linkdnrc, "A", "reason why link is down"); if (pi->port_type == FW_PORT_TYPE_BT_XAUI) { SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "temperature", CTLTYPE_INT | CTLFLAG_RD, pi, 0, sysctl_btphy, "I", "PHY temperature (in Celsius)"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "fw_version", CTLTYPE_INT | CTLFLAG_RD, pi, 1, sysctl_btphy, "I", "PHY firmware version"); } SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pause_settings", CTLTYPE_STRING | CTLFLAG_RW, pi, PAUSE_TX, sysctl_pause_settings, "A", "PAUSE settings (bit 0 = rx_pause, bit 1 = tx_pause)"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "max_speed", CTLFLAG_RD, NULL, port_top_speed(pi), "max speed (in Gbps)"); if (sc->flags & IS_VF) return; /* * dev.(cxgbe|cxl).X.tc. */ oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "tc", CTLFLAG_RD, NULL, "Tx scheduler traffic classes"); for (i = 0; i < sc->chip_params->nsched_cls; i++) { struct tx_sched_class *tc = &pi->tc[i]; snprintf(name, sizeof(name), "%d", i); children2 = SYSCTL_CHILDREN(SYSCTL_ADD_NODE(ctx, SYSCTL_CHILDREN(oid), OID_AUTO, name, CTLFLAG_RD, NULL, "traffic class")); SYSCTL_ADD_UINT(ctx, children2, OID_AUTO, "flags", CTLFLAG_RD, &tc->flags, 0, "flags"); SYSCTL_ADD_UINT(ctx, children2, OID_AUTO, "refcount", CTLFLAG_RD, &tc->refcount, 0, "references to this class"); #ifdef SBUF_DRAIN SYSCTL_ADD_PROC(ctx, children2, OID_AUTO, "params", CTLTYPE_STRING | CTLFLAG_RD, sc, (pi->port_id << 16) | i, sysctl_tc_params, "A", "traffic class parameters"); #endif } /* * dev.cxgbe.X.stats. */ oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "stats", CTLFLAG_RD, NULL, "port statistics"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "tx_parse_error", CTLFLAG_RD, &pi->tx_parse_error, 0, "# of tx packets with invalid length or # of segments"); #define SYSCTL_ADD_T4_REG64(pi, name, desc, reg) \ SYSCTL_ADD_OID(ctx, children, OID_AUTO, name, \ CTLTYPE_U64 | CTLFLAG_RD, sc, reg, \ sysctl_handle_t4_reg64, "QU", desc) SYSCTL_ADD_T4_REG64(pi, "tx_octets", "# of octets in good frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_BYTES_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames", "total # of good frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_FRAMES_L)); SYSCTL_ADD_T4_REG64(pi, "tx_bcast_frames", "# of broadcast frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_BCAST_L)); SYSCTL_ADD_T4_REG64(pi, "tx_mcast_frames", "# of multicast frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_MCAST_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ucast_frames", "# of unicast frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_UCAST_L)); SYSCTL_ADD_T4_REG64(pi, "tx_error_frames", "# of error frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_ERROR_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_64", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_64B_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_65_127", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_65B_127B_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_128_255", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_128B_255B_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_256_511", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_256B_511B_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_512_1023", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_512B_1023B_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_1024_1518", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_1024B_1518B_L)); SYSCTL_ADD_T4_REG64(pi, "tx_frames_1519_max", "# of tx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_1519B_MAX_L)); SYSCTL_ADD_T4_REG64(pi, "tx_drop", "# of dropped tx frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_DROP_L)); SYSCTL_ADD_T4_REG64(pi, "tx_pause", "# of pause frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PAUSE_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp0", "# of PPP prio 0 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP0_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp1", "# of PPP prio 1 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP1_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp2", "# of PPP prio 2 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP2_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp3", "# of PPP prio 3 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP3_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp4", "# of PPP prio 4 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP4_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp5", "# of PPP prio 5 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP5_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp6", "# of PPP prio 6 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP6_L)); SYSCTL_ADD_T4_REG64(pi, "tx_ppp7", "# of PPP prio 7 frames transmitted", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_TX_PORT_PPP7_L)); SYSCTL_ADD_T4_REG64(pi, "rx_octets", "# of octets in good frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_BYTES_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames", "total # of good frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_FRAMES_L)); SYSCTL_ADD_T4_REG64(pi, "rx_bcast_frames", "# of broadcast frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_BCAST_L)); SYSCTL_ADD_T4_REG64(pi, "rx_mcast_frames", "# of multicast frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_MCAST_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ucast_frames", "# of unicast frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_UCAST_L)); SYSCTL_ADD_T4_REG64(pi, "rx_too_long", "# of frames exceeding MTU", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_MTU_ERROR_L)); SYSCTL_ADD_T4_REG64(pi, "rx_jabber", "# of jabber frames", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_MTU_CRC_ERROR_L)); SYSCTL_ADD_T4_REG64(pi, "rx_fcs_err", "# of frames received with bad FCS", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_CRC_ERROR_L)); SYSCTL_ADD_T4_REG64(pi, "rx_len_err", "# of frames received with length error", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_LEN_ERROR_L)); SYSCTL_ADD_T4_REG64(pi, "rx_symbol_err", "symbol errors", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_SYM_ERROR_L)); SYSCTL_ADD_T4_REG64(pi, "rx_runt", "# of short frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_LESS_64B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_64", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_64B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_65_127", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_65B_127B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_128_255", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_128B_255B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_256_511", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_256B_511B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_512_1023", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_512B_1023B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_1024_1518", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_1024B_1518B_L)); SYSCTL_ADD_T4_REG64(pi, "rx_frames_1519_max", "# of rx frames in this range", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_1519B_MAX_L)); SYSCTL_ADD_T4_REG64(pi, "rx_pause", "# of pause frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PAUSE_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp0", "# of PPP prio 0 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP0_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp1", "# of PPP prio 1 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP1_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp2", "# of PPP prio 2 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP2_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp3", "# of PPP prio 3 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP3_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp4", "# of PPP prio 4 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP4_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp5", "# of PPP prio 5 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP5_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp6", "# of PPP prio 6 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP6_L)); SYSCTL_ADD_T4_REG64(pi, "rx_ppp7", "# of PPP prio 7 frames received", PORT_REG(pi->tx_chan, A_MPS_PORT_STAT_RX_PORT_PPP7_L)); #undef SYSCTL_ADD_T4_REG64 #define SYSCTL_ADD_T4_PORTSTAT(name, desc) \ SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, #name, CTLFLAG_RD, \ &pi->stats.name, desc) /* We get these from port_stats and they may be stale by up to 1s */ SYSCTL_ADD_T4_PORTSTAT(rx_ovflow0, "# drops due to buffer-group 0 overflows"); SYSCTL_ADD_T4_PORTSTAT(rx_ovflow1, "# drops due to buffer-group 1 overflows"); SYSCTL_ADD_T4_PORTSTAT(rx_ovflow2, "# drops due to buffer-group 2 overflows"); SYSCTL_ADD_T4_PORTSTAT(rx_ovflow3, "# drops due to buffer-group 3 overflows"); SYSCTL_ADD_T4_PORTSTAT(rx_trunc0, "# of buffer-group 0 truncated packets"); SYSCTL_ADD_T4_PORTSTAT(rx_trunc1, "# of buffer-group 1 truncated packets"); SYSCTL_ADD_T4_PORTSTAT(rx_trunc2, "# of buffer-group 2 truncated packets"); SYSCTL_ADD_T4_PORTSTAT(rx_trunc3, "# of buffer-group 3 truncated packets"); #undef SYSCTL_ADD_T4_PORTSTAT } static int sysctl_int_array(SYSCTL_HANDLER_ARGS) { int rc, *i, space = 0; struct sbuf sb; sbuf_new_for_sysctl(&sb, NULL, 64, req); for (i = arg1; arg2; arg2 -= sizeof(int), i++) { if (space) sbuf_printf(&sb, " "); sbuf_printf(&sb, "%d", *i); space = 1; } rc = sbuf_finish(&sb); sbuf_delete(&sb); return (rc); } static int sysctl_bitfield(SYSCTL_HANDLER_ARGS) { int rc; struct sbuf *sb; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return(rc); sb = sbuf_new_for_sysctl(NULL, NULL, 128, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "%b", (int)arg2, (char *)arg1); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_btphy(SYSCTL_HANDLER_ARGS) { struct port_info *pi = arg1; int op = arg2; struct adapter *sc = pi->adapter; u_int v; int rc; rc = begin_synchronized_op(sc, &pi->vi[0], SLEEP_OK | INTR_OK, "t4btt"); if (rc) return (rc); /* XXX: magic numbers */ rc = -t4_mdio_rd(sc, sc->mbox, pi->mdio_addr, 0x1e, op ? 0x20 : 0xc820, &v); end_synchronized_op(sc, 0); if (rc) return (rc); if (op == 0) v /= 256; rc = sysctl_handle_int(oidp, &v, 0, req); return (rc); } static int sysctl_noflowq(SYSCTL_HANDLER_ARGS) { struct vi_info *vi = arg1; int rc, val; val = vi->rsrv_noflowq; rc = sysctl_handle_int(oidp, &val, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); if ((val >= 1) && (vi->ntxq > 1)) vi->rsrv_noflowq = 1; else vi->rsrv_noflowq = 0; return (rc); } static int sysctl_holdoff_tmr_idx(SYSCTL_HANDLER_ARGS) { struct vi_info *vi = arg1; struct adapter *sc = vi->pi->adapter; int idx, rc, i; struct sge_rxq *rxq; #ifdef TCP_OFFLOAD struct sge_ofld_rxq *ofld_rxq; #endif uint8_t v; idx = vi->tmr_idx; rc = sysctl_handle_int(oidp, &idx, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); if (idx < 0 || idx >= SGE_NTIMERS) return (EINVAL); rc = begin_synchronized_op(sc, vi, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4tmr"); if (rc) return (rc); v = V_QINTR_TIMER_IDX(idx) | V_QINTR_CNT_EN(vi->pktc_idx != -1); for_each_rxq(vi, i, rxq) { #ifdef atomic_store_rel_8 atomic_store_rel_8(&rxq->iq.intr_params, v); #else rxq->iq.intr_params = v; #endif } #ifdef TCP_OFFLOAD for_each_ofld_rxq(vi, i, ofld_rxq) { #ifdef atomic_store_rel_8 atomic_store_rel_8(&ofld_rxq->iq.intr_params, v); #else ofld_rxq->iq.intr_params = v; #endif } #endif vi->tmr_idx = idx; end_synchronized_op(sc, LOCK_HELD); return (0); } static int sysctl_holdoff_pktc_idx(SYSCTL_HANDLER_ARGS) { struct vi_info *vi = arg1; struct adapter *sc = vi->pi->adapter; int idx, rc; idx = vi->pktc_idx; rc = sysctl_handle_int(oidp, &idx, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); if (idx < -1 || idx >= SGE_NCOUNTERS) return (EINVAL); rc = begin_synchronized_op(sc, vi, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4pktc"); if (rc) return (rc); if (vi->flags & VI_INIT_DONE) rc = EBUSY; /* cannot be changed once the queues are created */ else vi->pktc_idx = idx; end_synchronized_op(sc, LOCK_HELD); return (rc); } static int sysctl_qsize_rxq(SYSCTL_HANDLER_ARGS) { struct vi_info *vi = arg1; struct adapter *sc = vi->pi->adapter; int qsize, rc; qsize = vi->qsize_rxq; rc = sysctl_handle_int(oidp, &qsize, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); if (qsize < 128 || (qsize & 7)) return (EINVAL); rc = begin_synchronized_op(sc, vi, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4rxqs"); if (rc) return (rc); if (vi->flags & VI_INIT_DONE) rc = EBUSY; /* cannot be changed once the queues are created */ else vi->qsize_rxq = qsize; end_synchronized_op(sc, LOCK_HELD); return (rc); } static int sysctl_qsize_txq(SYSCTL_HANDLER_ARGS) { struct vi_info *vi = arg1; struct adapter *sc = vi->pi->adapter; int qsize, rc; qsize = vi->qsize_txq; rc = sysctl_handle_int(oidp, &qsize, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); if (qsize < 128 || qsize > 65536) return (EINVAL); rc = begin_synchronized_op(sc, vi, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4txqs"); if (rc) return (rc); if (vi->flags & VI_INIT_DONE) rc = EBUSY; /* cannot be changed once the queues are created */ else vi->qsize_txq = qsize; end_synchronized_op(sc, LOCK_HELD); return (rc); } static int sysctl_pause_settings(SYSCTL_HANDLER_ARGS) { struct port_info *pi = arg1; struct adapter *sc = pi->adapter; struct link_config *lc = &pi->link_cfg; int rc; if (req->newptr == NULL) { struct sbuf *sb; static char *bits = "\20\1PAUSE_RX\2PAUSE_TX"; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return(rc); sb = sbuf_new_for_sysctl(NULL, NULL, 128, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "%b", lc->fc & (PAUSE_TX | PAUSE_RX), bits); rc = sbuf_finish(sb); sbuf_delete(sb); } else { char s[2]; int n; s[0] = '0' + (lc->requested_fc & (PAUSE_TX | PAUSE_RX)); s[1] = 0; rc = sysctl_handle_string(oidp, s, sizeof(s), req); if (rc != 0) return(rc); if (s[1] != 0) return (EINVAL); if (s[0] < '0' || s[0] > '9') return (EINVAL); /* not a number */ n = s[0] - '0'; if (n & ~(PAUSE_TX | PAUSE_RX)) return (EINVAL); /* some other bit is set too */ rc = begin_synchronized_op(sc, &pi->vi[0], SLEEP_OK | INTR_OK, "t4PAUSE"); if (rc) return (rc); if ((lc->requested_fc & (PAUSE_TX | PAUSE_RX)) != n) { int link_ok = lc->link_ok; lc->requested_fc &= ~(PAUSE_TX | PAUSE_RX); lc->requested_fc |= n; rc = -t4_link_l1cfg(sc, sc->mbox, pi->tx_chan, lc); lc->link_ok = link_ok; /* restore */ } end_synchronized_op(sc, 0); } return (rc); } static int sysctl_handle_t4_reg64(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; int reg = arg2; uint64_t val; val = t4_read_reg64(sc, reg); return (sysctl_handle_64(oidp, &val, 0, req)); } static int sysctl_temperature(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; int rc, t; uint32_t param, val; rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4temp"); if (rc) return (rc); param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DEV) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DEV_DIAG) | V_FW_PARAMS_PARAM_Y(FW_PARAM_DEV_DIAG_TMP); rc = -t4_query_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); end_synchronized_op(sc, 0); if (rc) return (rc); /* unknown is returned as 0 but we display -1 in that case */ t = val == 0 ? -1 : val; rc = sysctl_handle_int(oidp, &t, 0, req); return (rc); } #ifdef SBUF_DRAIN static int sysctl_cctrl(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i; uint16_t incr[NMTUS][NCCTRL_WIN]; static const char *dec_fac[] = { "0.5", "0.5625", "0.625", "0.6875", "0.75", "0.8125", "0.875", "0.9375" }; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); t4_read_cong_tbl(sc, incr); for (i = 0; i < NCCTRL_WIN; ++i) { sbuf_printf(sb, "%2d: %4u %4u %4u %4u %4u %4u %4u %4u\n", i, incr[0][i], incr[1][i], incr[2][i], incr[3][i], incr[4][i], incr[5][i], incr[6][i], incr[7][i]); sbuf_printf(sb, "%8u %4u %4u %4u %4u %4u %4u %4u %5u %s\n", incr[8][i], incr[9][i], incr[10][i], incr[11][i], incr[12][i], incr[13][i], incr[14][i], incr[15][i], sc->params.a_wnd[i], dec_fac[sc->params.b_wnd[i]]); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static const char *qname[CIM_NUM_IBQ + CIM_NUM_OBQ_T5] = { "TP0", "TP1", "ULP", "SGE0", "SGE1", "NC-SI", /* ibq's */ "ULP0", "ULP1", "ULP2", "ULP3", "SGE", "NC-SI", /* obq's */ "SGE0-RX", "SGE1-RX" /* additional obq's (T5 onwards) */ }; static int sysctl_cim_ibq_obq(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i, n, qid = arg2; uint32_t *buf, *p; char *qtype; u_int cim_num_obq = sc->chip_params->cim_num_obq; KASSERT(qid >= 0 && qid < CIM_NUM_IBQ + cim_num_obq, ("%s: bad qid %d\n", __func__, qid)); if (qid < CIM_NUM_IBQ) { /* inbound queue */ qtype = "IBQ"; n = 4 * CIM_IBQ_SIZE; buf = malloc(n * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); rc = t4_read_cim_ibq(sc, qid, buf, n); } else { /* outbound queue */ qtype = "OBQ"; qid -= CIM_NUM_IBQ; n = 4 * cim_num_obq * CIM_OBQ_SIZE; buf = malloc(n * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); rc = t4_read_cim_obq(sc, qid, buf, n); } if (rc < 0) { rc = -rc; goto done; } n = rc * sizeof(uint32_t); /* rc has # of words actually read */ rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) goto done; sb = sbuf_new_for_sysctl(NULL, NULL, PAGE_SIZE, req); if (sb == NULL) { rc = ENOMEM; goto done; } sbuf_printf(sb, "%s%d %s", qtype , qid, qname[arg2]); for (i = 0, p = buf; i < n; i += 16, p += 4) sbuf_printf(sb, "\n%#06x: %08x %08x %08x %08x", i, p[0], p[1], p[2], p[3]); rc = sbuf_finish(sb); sbuf_delete(sb); done: free(buf, M_CXGBE); return (rc); } static int sysctl_cim_la(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; u_int cfg; struct sbuf *sb; uint32_t *buf, *p; int rc; MPASS(chip_id(sc) <= CHELSIO_T5); rc = -t4_cim_read(sc, A_UP_UP_DBG_LA_CFG, 1, &cfg); if (rc != 0) return (rc); rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); buf = malloc(sc->params.cim_la_size * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); rc = -t4_cim_read_la(sc, buf, NULL); if (rc != 0) goto done; sbuf_printf(sb, "Status Data PC%s", cfg & F_UPDBGLACAPTPCONLY ? "" : " LS0Stat LS0Addr LS0Data"); for (p = buf; p <= &buf[sc->params.cim_la_size - 8]; p += 8) { if (cfg & F_UPDBGLACAPTPCONLY) { sbuf_printf(sb, "\n %02x %08x %08x", p[5] & 0xff, p[6], p[7]); sbuf_printf(sb, "\n %02x %02x%06x %02x%06x", (p[3] >> 8) & 0xff, p[3] & 0xff, p[4] >> 8, p[4] & 0xff, p[5] >> 8); sbuf_printf(sb, "\n %02x %x%07x %x%07x", (p[0] >> 4) & 0xff, p[0] & 0xf, p[1] >> 4, p[1] & 0xf, p[2] >> 4); } else { sbuf_printf(sb, "\n %02x %x%07x %x%07x %08x %08x " "%08x%08x%08x%08x", (p[0] >> 4) & 0xff, p[0] & 0xf, p[1] >> 4, p[1] & 0xf, p[2] >> 4, p[2] & 0xf, p[3], p[4], p[5], p[6], p[7]); } } rc = sbuf_finish(sb); sbuf_delete(sb); done: free(buf, M_CXGBE); return (rc); } static int sysctl_cim_la_t6(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; u_int cfg; struct sbuf *sb; uint32_t *buf, *p; int rc; MPASS(chip_id(sc) > CHELSIO_T5); rc = -t4_cim_read(sc, A_UP_UP_DBG_LA_CFG, 1, &cfg); if (rc != 0) return (rc); rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); buf = malloc(sc->params.cim_la_size * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); rc = -t4_cim_read_la(sc, buf, NULL); if (rc != 0) goto done; sbuf_printf(sb, "Status Inst Data PC%s", cfg & F_UPDBGLACAPTPCONLY ? "" : " LS0Stat LS0Addr LS0Data LS1Stat LS1Addr LS1Data"); for (p = buf; p <= &buf[sc->params.cim_la_size - 10]; p += 10) { if (cfg & F_UPDBGLACAPTPCONLY) { sbuf_printf(sb, "\n %02x %08x %08x %08x", p[3] & 0xff, p[2], p[1], p[0]); sbuf_printf(sb, "\n %02x %02x%06x %02x%06x %02x%06x", (p[6] >> 8) & 0xff, p[6] & 0xff, p[5] >> 8, p[5] & 0xff, p[4] >> 8, p[4] & 0xff, p[3] >> 8); sbuf_printf(sb, "\n %02x %04x%04x %04x%04x %04x%04x", (p[9] >> 16) & 0xff, p[9] & 0xffff, p[8] >> 16, p[8] & 0xffff, p[7] >> 16, p[7] & 0xffff, p[6] >> 16); } else { sbuf_printf(sb, "\n %02x %04x%04x %04x%04x %04x%04x " "%08x %08x %08x %08x %08x %08x", (p[9] >> 16) & 0xff, p[9] & 0xffff, p[8] >> 16, p[8] & 0xffff, p[7] >> 16, p[7] & 0xffff, p[6] >> 16, p[2], p[1], p[0], p[5], p[4], p[3]); } } rc = sbuf_finish(sb); sbuf_delete(sb); done: free(buf, M_CXGBE); return (rc); } static int sysctl_cim_ma_la(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; u_int i; struct sbuf *sb; uint32_t *buf, *p; int rc; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); buf = malloc(2 * CIM_MALA_SIZE * 5 * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); t4_cim_read_ma_la(sc, buf, buf + 5 * CIM_MALA_SIZE); p = buf; for (i = 0; i < CIM_MALA_SIZE; i++, p += 5) { sbuf_printf(sb, "\n%02x%08x%08x%08x%08x", p[4], p[3], p[2], p[1], p[0]); } sbuf_printf(sb, "\n\nCnt ID Tag UE Data RDY VLD"); for (i = 0; i < CIM_MALA_SIZE; i++, p += 5) { sbuf_printf(sb, "\n%3u %2u %x %u %08x%08x %u %u", (p[2] >> 10) & 0xff, (p[2] >> 7) & 7, (p[2] >> 3) & 0xf, (p[2] >> 2) & 1, (p[1] >> 2) | ((p[2] & 3) << 30), (p[0] >> 2) | ((p[1] & 3) << 30), (p[0] >> 1) & 1, p[0] & 1); } rc = sbuf_finish(sb); sbuf_delete(sb); free(buf, M_CXGBE); return (rc); } static int sysctl_cim_pif_la(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; u_int i; struct sbuf *sb; uint32_t *buf, *p; int rc; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); buf = malloc(2 * CIM_PIFLA_SIZE * 6 * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); t4_cim_read_pif_la(sc, buf, buf + 6 * CIM_PIFLA_SIZE, NULL, NULL); p = buf; sbuf_printf(sb, "Cntl ID DataBE Addr Data"); for (i = 0; i < CIM_PIFLA_SIZE; i++, p += 6) { sbuf_printf(sb, "\n %02x %02x %04x %08x %08x%08x%08x%08x", (p[5] >> 22) & 0xff, (p[5] >> 16) & 0x3f, p[5] & 0xffff, p[4], p[3], p[2], p[1], p[0]); } sbuf_printf(sb, "\n\nCntl ID Data"); for (i = 0; i < CIM_PIFLA_SIZE; i++, p += 6) { sbuf_printf(sb, "\n %02x %02x %08x%08x%08x%08x", (p[4] >> 6) & 0xff, p[4] & 0x3f, p[3], p[2], p[1], p[0]); } rc = sbuf_finish(sb); sbuf_delete(sb); free(buf, M_CXGBE); return (rc); } static int sysctl_cim_qcfg(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i; uint16_t base[CIM_NUM_IBQ + CIM_NUM_OBQ_T5]; uint16_t size[CIM_NUM_IBQ + CIM_NUM_OBQ_T5]; uint16_t thres[CIM_NUM_IBQ]; uint32_t obq_wr[2 * CIM_NUM_OBQ_T5], *wr = obq_wr; uint32_t stat[4 * (CIM_NUM_IBQ + CIM_NUM_OBQ_T5)], *p = stat; u_int cim_num_obq, ibq_rdaddr, obq_rdaddr, nq; cim_num_obq = sc->chip_params->cim_num_obq; if (is_t4(sc)) { ibq_rdaddr = A_UP_IBQ_0_RDADDR; obq_rdaddr = A_UP_OBQ_0_REALADDR; } else { ibq_rdaddr = A_UP_IBQ_0_SHADOW_RDADDR; obq_rdaddr = A_UP_OBQ_0_SHADOW_REALADDR; } nq = CIM_NUM_IBQ + cim_num_obq; rc = -t4_cim_read(sc, ibq_rdaddr, 4 * nq, stat); if (rc == 0) rc = -t4_cim_read(sc, obq_rdaddr, 2 * cim_num_obq, obq_wr); if (rc != 0) return (rc); t4_read_cimq_cfg(sc, base, size, thres); rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, PAGE_SIZE, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "Queue Base Size Thres RdPtr WrPtr SOP EOP Avail"); for (i = 0; i < CIM_NUM_IBQ; i++, p += 4) sbuf_printf(sb, "\n%7s %5x %5u %5u %6x %4x %4u %4u %5u", qname[i], base[i], size[i], thres[i], G_IBQRDADDR(p[0]), G_IBQWRADDR(p[1]), G_QUESOPCNT(p[3]), G_QUEEOPCNT(p[3]), G_QUEREMFLITS(p[2]) * 16); for ( ; i < nq; i++, p += 4, wr += 2) sbuf_printf(sb, "\n%7s %5x %5u %12x %4x %4u %4u %5u", qname[i], base[i], size[i], G_QUERDADDR(p[0]) & 0x3fff, wr[0] - base[i], G_QUESOPCNT(p[3]), G_QUEEOPCNT(p[3]), G_QUEREMFLITS(p[2]) * 16); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_cpl_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tp_cpl_stats stats; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); mtx_lock(&sc->reg_lock); t4_tp_get_cpl_stats(sc, &stats); mtx_unlock(&sc->reg_lock); if (sc->chip_params->nchan > 2) { sbuf_printf(sb, " channel 0 channel 1" " channel 2 channel 3"); sbuf_printf(sb, "\nCPL requests: %10u %10u %10u %10u", stats.req[0], stats.req[1], stats.req[2], stats.req[3]); sbuf_printf(sb, "\nCPL responses: %10u %10u %10u %10u", stats.rsp[0], stats.rsp[1], stats.rsp[2], stats.rsp[3]); } else { sbuf_printf(sb, " channel 0 channel 1"); sbuf_printf(sb, "\nCPL requests: %10u %10u", stats.req[0], stats.req[1]); sbuf_printf(sb, "\nCPL responses: %10u %10u", stats.rsp[0], stats.rsp[1]); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_ddp_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tp_usm_stats stats; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return(rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); t4_get_usm_stats(sc, &stats); sbuf_printf(sb, "Frames: %u\n", stats.frames); sbuf_printf(sb, "Octets: %ju\n", stats.octets); sbuf_printf(sb, "Drops: %u", stats.drops); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static const char * const devlog_level_strings[] = { [FW_DEVLOG_LEVEL_EMERG] = "EMERG", [FW_DEVLOG_LEVEL_CRIT] = "CRIT", [FW_DEVLOG_LEVEL_ERR] = "ERR", [FW_DEVLOG_LEVEL_NOTICE] = "NOTICE", [FW_DEVLOG_LEVEL_INFO] = "INFO", [FW_DEVLOG_LEVEL_DEBUG] = "DEBUG" }; static const char * const devlog_facility_strings[] = { [FW_DEVLOG_FACILITY_CORE] = "CORE", [FW_DEVLOG_FACILITY_CF] = "CF", [FW_DEVLOG_FACILITY_SCHED] = "SCHED", [FW_DEVLOG_FACILITY_TIMER] = "TIMER", [FW_DEVLOG_FACILITY_RES] = "RES", [FW_DEVLOG_FACILITY_HW] = "HW", [FW_DEVLOG_FACILITY_FLR] = "FLR", [FW_DEVLOG_FACILITY_DMAQ] = "DMAQ", [FW_DEVLOG_FACILITY_PHY] = "PHY", [FW_DEVLOG_FACILITY_MAC] = "MAC", [FW_DEVLOG_FACILITY_PORT] = "PORT", [FW_DEVLOG_FACILITY_VI] = "VI", [FW_DEVLOG_FACILITY_FILTER] = "FILTER", [FW_DEVLOG_FACILITY_ACL] = "ACL", [FW_DEVLOG_FACILITY_TM] = "TM", [FW_DEVLOG_FACILITY_QFC] = "QFC", [FW_DEVLOG_FACILITY_DCB] = "DCB", [FW_DEVLOG_FACILITY_ETH] = "ETH", [FW_DEVLOG_FACILITY_OFLD] = "OFLD", [FW_DEVLOG_FACILITY_RI] = "RI", [FW_DEVLOG_FACILITY_ISCSI] = "ISCSI", [FW_DEVLOG_FACILITY_FCOE] = "FCOE", [FW_DEVLOG_FACILITY_FOISCSI] = "FOISCSI", [FW_DEVLOG_FACILITY_FOFCOE] = "FOFCOE", [FW_DEVLOG_FACILITY_CHNET] = "CHNET", }; static int sysctl_devlog(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct devlog_params *dparams = &sc->params.devlog; struct fw_devlog_e *buf, *e; int i, j, rc, nentries, first = 0; struct sbuf *sb; uint64_t ftstamp = UINT64_MAX; if (dparams->addr == 0) return (ENXIO); buf = malloc(dparams->size, M_CXGBE, M_NOWAIT); if (buf == NULL) return (ENOMEM); rc = read_via_memwin(sc, 1, dparams->addr, (void *)buf, dparams->size); if (rc != 0) goto done; nentries = dparams->size / sizeof(struct fw_devlog_e); for (i = 0; i < nentries; i++) { e = &buf[i]; if (e->timestamp == 0) break; /* end */ e->timestamp = be64toh(e->timestamp); e->seqno = be32toh(e->seqno); for (j = 0; j < 8; j++) e->params[j] = be32toh(e->params[j]); if (e->timestamp < ftstamp) { ftstamp = e->timestamp; first = i; } } if (buf[first].timestamp == 0) goto done; /* nothing in the log */ rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) goto done; sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) { rc = ENOMEM; goto done; } sbuf_printf(sb, "%10s %15s %8s %8s %s\n", "Seq#", "Tstamp", "Level", "Facility", "Message"); i = first; do { e = &buf[i]; if (e->timestamp == 0) break; /* end */ sbuf_printf(sb, "%10d %15ju %8s %8s ", e->seqno, e->timestamp, (e->level < nitems(devlog_level_strings) ? devlog_level_strings[e->level] : "UNKNOWN"), (e->facility < nitems(devlog_facility_strings) ? devlog_facility_strings[e->facility] : "UNKNOWN")); sbuf_printf(sb, e->fmt, e->params[0], e->params[1], e->params[2], e->params[3], e->params[4], e->params[5], e->params[6], e->params[7]); if (++i == nentries) i = 0; } while (i != first); rc = sbuf_finish(sb); sbuf_delete(sb); done: free(buf, M_CXGBE); return (rc); } static int sysctl_fcoe_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tp_fcoe_stats stats[MAX_NCHAN]; int i, nchan = sc->chip_params->nchan; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); for (i = 0; i < nchan; i++) t4_get_fcoe_stats(sc, i, &stats[i]); if (nchan > 2) { sbuf_printf(sb, " channel 0 channel 1" " channel 2 channel 3"); sbuf_printf(sb, "\noctetsDDP: %16ju %16ju %16ju %16ju", stats[0].octets_ddp, stats[1].octets_ddp, stats[2].octets_ddp, stats[3].octets_ddp); sbuf_printf(sb, "\nframesDDP: %16u %16u %16u %16u", stats[0].frames_ddp, stats[1].frames_ddp, stats[2].frames_ddp, stats[3].frames_ddp); sbuf_printf(sb, "\nframesDrop: %16u %16u %16u %16u", stats[0].frames_drop, stats[1].frames_drop, stats[2].frames_drop, stats[3].frames_drop); } else { sbuf_printf(sb, " channel 0 channel 1"); sbuf_printf(sb, "\noctetsDDP: %16ju %16ju", stats[0].octets_ddp, stats[1].octets_ddp); sbuf_printf(sb, "\nframesDDP: %16u %16u", stats[0].frames_ddp, stats[1].frames_ddp); sbuf_printf(sb, "\nframesDrop: %16u %16u", stats[0].frames_drop, stats[1].frames_drop); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_hw_sched(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i; unsigned int map, kbps, ipg, mode; unsigned int pace_tab[NTX_SCHED]; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); map = t4_read_reg(sc, A_TP_TX_MOD_QUEUE_REQ_MAP); mode = G_TIMERMODE(t4_read_reg(sc, A_TP_MOD_CONFIG)); t4_read_pace_tbl(sc, pace_tab); sbuf_printf(sb, "Scheduler Mode Channel Rate (Kbps) " "Class IPG (0.1 ns) Flow IPG (us)"); for (i = 0; i < NTX_SCHED; ++i, map >>= 2) { t4_get_tx_sched(sc, i, &kbps, &ipg); sbuf_printf(sb, "\n %u %-5s %u ", i, (mode & (1 << i)) ? "flow" : "class", map & 3); if (kbps) sbuf_printf(sb, "%9u ", kbps); else sbuf_printf(sb, " disabled "); if (ipg) sbuf_printf(sb, "%13u ", ipg); else sbuf_printf(sb, " disabled "); if (pace_tab[i]) sbuf_printf(sb, "%10u", pace_tab[i]); else sbuf_printf(sb, " disabled"); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_lb_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i, j; uint64_t *p0, *p1; struct lb_port_stats s[2]; static const char *stat_name[] = { "OctetsOK:", "FramesOK:", "BcastFrames:", "McastFrames:", "UcastFrames:", "ErrorFrames:", "Frames64:", "Frames65To127:", "Frames128To255:", "Frames256To511:", "Frames512To1023:", "Frames1024To1518:", "Frames1519ToMax:", "FramesDropped:", "BG0FramesDropped:", "BG1FramesDropped:", "BG2FramesDropped:", "BG3FramesDropped:", "BG0FramesTrunc:", "BG1FramesTrunc:", "BG2FramesTrunc:", "BG3FramesTrunc:" }; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); memset(s, 0, sizeof(s)); for (i = 0; i < sc->chip_params->nchan; i += 2) { t4_get_lb_stats(sc, i, &s[0]); t4_get_lb_stats(sc, i + 1, &s[1]); p0 = &s[0].octets; p1 = &s[1].octets; sbuf_printf(sb, "%s Loopback %u" " Loopback %u", i == 0 ? "" : "\n", i, i + 1); for (j = 0; j < nitems(stat_name); j++) sbuf_printf(sb, "\n%-17s %20ju %20ju", stat_name[j], *p0++, *p1++); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_linkdnrc(SYSCTL_HANDLER_ARGS) { int rc = 0; struct port_info *pi = arg1; struct sbuf *sb; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return(rc); sb = sbuf_new_for_sysctl(NULL, NULL, 64, req); if (sb == NULL) return (ENOMEM); if (pi->linkdnrc < 0) sbuf_printf(sb, "n/a"); else sbuf_printf(sb, "%s", t4_link_down_rc_str(pi->linkdnrc)); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } struct mem_desc { unsigned int base; unsigned int limit; unsigned int idx; }; static int mem_desc_cmp(const void *a, const void *b) { return ((const struct mem_desc *)a)->base - ((const struct mem_desc *)b)->base; } static void mem_region_show(struct sbuf *sb, const char *name, unsigned int from, unsigned int to) { unsigned int size; if (from == to) return; size = to - from + 1; if (size == 0) return; /* XXX: need humanize_number(3) in libkern for a more readable 'size' */ sbuf_printf(sb, "%-15s %#x-%#x [%u]\n", name, from, to, size); } static int sysctl_meminfo(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i, n; uint32_t lo, hi, used, alloc; static const char *memory[] = {"EDC0:", "EDC1:", "MC:", "MC0:", "MC1:"}; static const char *region[] = { "DBQ contexts:", "IMSG contexts:", "FLM cache:", "TCBs:", "Pstructs:", "Timers:", "Rx FL:", "Tx FL:", "Pstruct FL:", "Tx payload:", "Rx payload:", "LE hash:", "iSCSI region:", "TDDP region:", "TPT region:", "STAG region:", "RQ region:", "RQUDP region:", "PBL region:", "TXPBL region:", "DBVFIFO region:", "ULPRX state:", "ULPTX state:", "On-chip queues:" }; struct mem_desc avail[4]; struct mem_desc mem[nitems(region) + 3]; /* up to 3 holes */ struct mem_desc *md = mem; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); for (i = 0; i < nitems(mem); i++) { mem[i].limit = 0; mem[i].idx = i; } /* Find and sort the populated memory ranges */ i = 0; lo = t4_read_reg(sc, A_MA_TARGET_MEM_ENABLE); if (lo & F_EDRAM0_ENABLE) { hi = t4_read_reg(sc, A_MA_EDRAM0_BAR); avail[i].base = G_EDRAM0_BASE(hi) << 20; avail[i].limit = avail[i].base + (G_EDRAM0_SIZE(hi) << 20); avail[i].idx = 0; i++; } if (lo & F_EDRAM1_ENABLE) { hi = t4_read_reg(sc, A_MA_EDRAM1_BAR); avail[i].base = G_EDRAM1_BASE(hi) << 20; avail[i].limit = avail[i].base + (G_EDRAM1_SIZE(hi) << 20); avail[i].idx = 1; i++; } if (lo & F_EXT_MEM_ENABLE) { hi = t4_read_reg(sc, A_MA_EXT_MEMORY_BAR); avail[i].base = G_EXT_MEM_BASE(hi) << 20; avail[i].limit = avail[i].base + (G_EXT_MEM_SIZE(hi) << 20); avail[i].idx = is_t5(sc) ? 3 : 2; /* Call it MC0 for T5 */ i++; } if (is_t5(sc) && lo & F_EXT_MEM1_ENABLE) { hi = t4_read_reg(sc, A_MA_EXT_MEMORY1_BAR); avail[i].base = G_EXT_MEM1_BASE(hi) << 20; avail[i].limit = avail[i].base + (G_EXT_MEM1_SIZE(hi) << 20); avail[i].idx = 4; i++; } if (!i) /* no memory available */ return 0; qsort(avail, i, sizeof(struct mem_desc), mem_desc_cmp); (md++)->base = t4_read_reg(sc, A_SGE_DBQ_CTXT_BADDR); (md++)->base = t4_read_reg(sc, A_SGE_IMSG_CTXT_BADDR); (md++)->base = t4_read_reg(sc, A_SGE_FLM_CACHE_BADDR); (md++)->base = t4_read_reg(sc, A_TP_CMM_TCB_BASE); (md++)->base = t4_read_reg(sc, A_TP_CMM_MM_BASE); (md++)->base = t4_read_reg(sc, A_TP_CMM_TIMER_BASE); (md++)->base = t4_read_reg(sc, A_TP_CMM_MM_RX_FLST_BASE); (md++)->base = t4_read_reg(sc, A_TP_CMM_MM_TX_FLST_BASE); (md++)->base = t4_read_reg(sc, A_TP_CMM_MM_PS_FLST_BASE); /* the next few have explicit upper bounds */ md->base = t4_read_reg(sc, A_TP_PMM_TX_BASE); md->limit = md->base - 1 + t4_read_reg(sc, A_TP_PMM_TX_PAGE_SIZE) * G_PMTXMAXPAGE(t4_read_reg(sc, A_TP_PMM_TX_MAX_PAGE)); md++; md->base = t4_read_reg(sc, A_TP_PMM_RX_BASE); md->limit = md->base - 1 + t4_read_reg(sc, A_TP_PMM_RX_PAGE_SIZE) * G_PMRXMAXPAGE(t4_read_reg(sc, A_TP_PMM_RX_MAX_PAGE)); md++; if (t4_read_reg(sc, A_LE_DB_CONFIG) & F_HASHEN) { if (chip_id(sc) <= CHELSIO_T5) md->base = t4_read_reg(sc, A_LE_DB_HASH_TID_BASE); else md->base = t4_read_reg(sc, A_LE_DB_HASH_TBL_BASE_ADDR); md->limit = 0; } else { md->base = 0; md->idx = nitems(region); /* hide it */ } md++; #define ulp_region(reg) \ md->base = t4_read_reg(sc, A_ULP_ ## reg ## _LLIMIT);\ (md++)->limit = t4_read_reg(sc, A_ULP_ ## reg ## _ULIMIT) ulp_region(RX_ISCSI); ulp_region(RX_TDDP); ulp_region(TX_TPT); ulp_region(RX_STAG); ulp_region(RX_RQ); ulp_region(RX_RQUDP); ulp_region(RX_PBL); ulp_region(TX_PBL); #undef ulp_region md->base = 0; md->idx = nitems(region); if (!is_t4(sc)) { uint32_t size = 0; uint32_t sge_ctrl = t4_read_reg(sc, A_SGE_CONTROL2); uint32_t fifo_size = t4_read_reg(sc, A_SGE_DBVFIFO_SIZE); if (is_t5(sc)) { if (sge_ctrl & F_VFIFO_ENABLE) size = G_DBVFIFO_SIZE(fifo_size); } else size = G_T6_DBVFIFO_SIZE(fifo_size); if (size) { md->base = G_BASEADDR(t4_read_reg(sc, A_SGE_DBVFIFO_BADDR)); md->limit = md->base + (size << 2) - 1; } } md++; md->base = t4_read_reg(sc, A_ULP_RX_CTX_BASE); md->limit = 0; md++; md->base = t4_read_reg(sc, A_ULP_TX_ERR_TABLE_BASE); md->limit = 0; md++; md->base = sc->vres.ocq.start; if (sc->vres.ocq.size) md->limit = md->base + sc->vres.ocq.size - 1; else md->idx = nitems(region); /* hide it */ md++; /* add any address-space holes, there can be up to 3 */ for (n = 0; n < i - 1; n++) if (avail[n].limit < avail[n + 1].base) (md++)->base = avail[n].limit; if (avail[n].limit) (md++)->base = avail[n].limit; n = md - mem; qsort(mem, n, sizeof(struct mem_desc), mem_desc_cmp); for (lo = 0; lo < i; lo++) mem_region_show(sb, memory[avail[lo].idx], avail[lo].base, avail[lo].limit - 1); sbuf_printf(sb, "\n"); for (i = 0; i < n; i++) { if (mem[i].idx >= nitems(region)) continue; /* skip holes */ if (!mem[i].limit) mem[i].limit = i < n - 1 ? mem[i + 1].base - 1 : ~0; mem_region_show(sb, region[mem[i].idx], mem[i].base, mem[i].limit); } sbuf_printf(sb, "\n"); lo = t4_read_reg(sc, A_CIM_SDRAM_BASE_ADDR); hi = t4_read_reg(sc, A_CIM_SDRAM_ADDR_SIZE) + lo - 1; mem_region_show(sb, "uP RAM:", lo, hi); lo = t4_read_reg(sc, A_CIM_EXTMEM2_BASE_ADDR); hi = t4_read_reg(sc, A_CIM_EXTMEM2_ADDR_SIZE) + lo - 1; mem_region_show(sb, "uP Extmem2:", lo, hi); lo = t4_read_reg(sc, A_TP_PMM_RX_MAX_PAGE); sbuf_printf(sb, "\n%u Rx pages of size %uKiB for %u channels\n", G_PMRXMAXPAGE(lo), t4_read_reg(sc, A_TP_PMM_RX_PAGE_SIZE) >> 10, (lo & F_PMRXNUMCHN) ? 2 : 1); lo = t4_read_reg(sc, A_TP_PMM_TX_MAX_PAGE); hi = t4_read_reg(sc, A_TP_PMM_TX_PAGE_SIZE); sbuf_printf(sb, "%u Tx pages of size %u%ciB for %u channels\n", G_PMTXMAXPAGE(lo), hi >= (1 << 20) ? (hi >> 20) : (hi >> 10), hi >= (1 << 20) ? 'M' : 'K', 1 << G_PMTXNUMCHN(lo)); sbuf_printf(sb, "%u p-structs\n", t4_read_reg(sc, A_TP_CMM_MM_MAX_PSTRUCT)); for (i = 0; i < 4; i++) { if (chip_id(sc) > CHELSIO_T5) lo = t4_read_reg(sc, A_MPS_RX_MAC_BG_PG_CNT0 + i * 4); else lo = t4_read_reg(sc, A_MPS_RX_PG_RSV0 + i * 4); if (is_t5(sc)) { used = G_T5_USED(lo); alloc = G_T5_ALLOC(lo); } else { used = G_USED(lo); alloc = G_ALLOC(lo); } /* For T6 these are MAC buffer groups */ sbuf_printf(sb, "\nPort %d using %u pages out of %u allocated", i, used, alloc); } for (i = 0; i < sc->chip_params->nchan; i++) { if (chip_id(sc) > CHELSIO_T5) lo = t4_read_reg(sc, A_MPS_RX_LPBK_BG_PG_CNT0 + i * 4); else lo = t4_read_reg(sc, A_MPS_RX_PG_RSV4 + i * 4); if (is_t5(sc)) { used = G_T5_USED(lo); alloc = G_T5_ALLOC(lo); } else { used = G_USED(lo); alloc = G_ALLOC(lo); } /* For T6 these are MAC buffer groups */ sbuf_printf(sb, "\nLoopback %d using %u pages out of %u allocated", i, used, alloc); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static inline void tcamxy2valmask(uint64_t x, uint64_t y, uint8_t *addr, uint64_t *mask) { *mask = x | y; y = htobe64(y); memcpy(addr, (char *)&y + 2, ETHER_ADDR_LEN); } static int sysctl_mps_tcam(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i; MPASS(chip_id(sc) <= CHELSIO_T5); rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "Idx Ethernet address Mask Vld Ports PF" " VF Replication P0 P1 P2 P3 ML"); for (i = 0; i < sc->chip_params->mps_tcam_size; i++) { uint64_t tcamx, tcamy, mask; uint32_t cls_lo, cls_hi; uint8_t addr[ETHER_ADDR_LEN]; tcamy = t4_read_reg64(sc, MPS_CLS_TCAM_Y_L(i)); tcamx = t4_read_reg64(sc, MPS_CLS_TCAM_X_L(i)); if (tcamx & tcamy) continue; tcamxy2valmask(tcamx, tcamy, addr, &mask); cls_lo = t4_read_reg(sc, MPS_CLS_SRAM_L(i)); cls_hi = t4_read_reg(sc, MPS_CLS_SRAM_H(i)); sbuf_printf(sb, "\n%3u %02x:%02x:%02x:%02x:%02x:%02x %012jx" " %c %#x%4u%4d", i, addr[0], addr[1], addr[2], addr[3], addr[4], addr[5], (uintmax_t)mask, (cls_lo & F_SRAM_VLD) ? 'Y' : 'N', G_PORTMAP(cls_hi), G_PF(cls_lo), (cls_lo & F_VF_VALID) ? G_VF(cls_lo) : -1); if (cls_lo & F_REPLICATE) { struct fw_ldst_cmd ldst_cmd; memset(&ldst_cmd, 0, sizeof(ldst_cmd)); ldst_cmd.op_to_addrspace = htobe32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_MPS)); ldst_cmd.cycles_to_len16 = htobe32(FW_LEN16(ldst_cmd)); ldst_cmd.u.mps.rplc.fid_idx = htobe16(V_FW_LDST_CMD_FID(FW_LDST_MPS_RPLC) | V_FW_LDST_CMD_IDX(i)); rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4mps"); if (rc) break; rc = -t4_wr_mbox(sc, sc->mbox, &ldst_cmd, sizeof(ldst_cmd), &ldst_cmd); end_synchronized_op(sc, 0); if (rc != 0) { sbuf_printf(sb, "%36d", rc); rc = 0; } else { sbuf_printf(sb, " %08x %08x %08x %08x", be32toh(ldst_cmd.u.mps.rplc.rplc127_96), be32toh(ldst_cmd.u.mps.rplc.rplc95_64), be32toh(ldst_cmd.u.mps.rplc.rplc63_32), be32toh(ldst_cmd.u.mps.rplc.rplc31_0)); } } else sbuf_printf(sb, "%36s", ""); sbuf_printf(sb, "%4u%3u%3u%3u %#3x", G_SRAM_PRIO0(cls_lo), G_SRAM_PRIO1(cls_lo), G_SRAM_PRIO2(cls_lo), G_SRAM_PRIO3(cls_lo), (cls_lo >> S_MULTILISTEN0) & 0xf); } if (rc) (void) sbuf_finish(sb); else rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_mps_tcam_t6(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i; MPASS(chip_id(sc) > CHELSIO_T5); rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "Idx Ethernet address Mask VNI Mask" " IVLAN Vld DIP_Hit Lookup Port Vld Ports PF VF" " Replication" " P0 P1 P2 P3 ML\n"); for (i = 0; i < sc->chip_params->mps_tcam_size; i++) { uint8_t dip_hit, vlan_vld, lookup_type, port_num; uint16_t ivlan; uint64_t tcamx, tcamy, val, mask; uint32_t cls_lo, cls_hi, ctl, data2, vnix, vniy; uint8_t addr[ETHER_ADDR_LEN]; ctl = V_CTLREQID(1) | V_CTLCMDTYPE(0) | V_CTLXYBITSEL(0); if (i < 256) ctl |= V_CTLTCAMINDEX(i) | V_CTLTCAMSEL(0); else ctl |= V_CTLTCAMINDEX(i - 256) | V_CTLTCAMSEL(1); t4_write_reg(sc, A_MPS_CLS_TCAM_DATA2_CTL, ctl); val = t4_read_reg(sc, A_MPS_CLS_TCAM_RDATA1_REQ_ID1); tcamy = G_DMACH(val) << 32; tcamy |= t4_read_reg(sc, A_MPS_CLS_TCAM_RDATA0_REQ_ID1); data2 = t4_read_reg(sc, A_MPS_CLS_TCAM_RDATA2_REQ_ID1); lookup_type = G_DATALKPTYPE(data2); port_num = G_DATAPORTNUM(data2); if (lookup_type && lookup_type != M_DATALKPTYPE) { /* Inner header VNI */ vniy = ((data2 & F_DATAVIDH2) << 23) | (G_DATAVIDH1(data2) << 16) | G_VIDL(val); dip_hit = data2 & F_DATADIPHIT; vlan_vld = 0; } else { vniy = 0; dip_hit = 0; vlan_vld = data2 & F_DATAVIDH2; ivlan = G_VIDL(val); } ctl |= V_CTLXYBITSEL(1); t4_write_reg(sc, A_MPS_CLS_TCAM_DATA2_CTL, ctl); val = t4_read_reg(sc, A_MPS_CLS_TCAM_RDATA1_REQ_ID1); tcamx = G_DMACH(val) << 32; tcamx |= t4_read_reg(sc, A_MPS_CLS_TCAM_RDATA0_REQ_ID1); data2 = t4_read_reg(sc, A_MPS_CLS_TCAM_RDATA2_REQ_ID1); if (lookup_type && lookup_type != M_DATALKPTYPE) { /* Inner header VNI mask */ vnix = ((data2 & F_DATAVIDH2) << 23) | (G_DATAVIDH1(data2) << 16) | G_VIDL(val); } else vnix = 0; if (tcamx & tcamy) continue; tcamxy2valmask(tcamx, tcamy, addr, &mask); cls_lo = t4_read_reg(sc, MPS_CLS_SRAM_L(i)); cls_hi = t4_read_reg(sc, MPS_CLS_SRAM_H(i)); if (lookup_type && lookup_type != M_DATALKPTYPE) { sbuf_printf(sb, "\n%3u %02x:%02x:%02x:%02x:%02x:%02x " "%012jx %06x %06x - - %3c" " 'I' %4x %3c %#x%4u%4d", i, addr[0], addr[1], addr[2], addr[3], addr[4], addr[5], (uintmax_t)mask, vniy, vnix, dip_hit ? 'Y' : 'N', port_num, cls_lo & F_T6_SRAM_VLD ? 'Y' : 'N', G_PORTMAP(cls_hi), G_T6_PF(cls_lo), cls_lo & F_T6_VF_VALID ? G_T6_VF(cls_lo) : -1); } else { sbuf_printf(sb, "\n%3u %02x:%02x:%02x:%02x:%02x:%02x " "%012jx - - ", i, addr[0], addr[1], addr[2], addr[3], addr[4], addr[5], (uintmax_t)mask); if (vlan_vld) sbuf_printf(sb, "%4u Y ", ivlan); else sbuf_printf(sb, " - N "); sbuf_printf(sb, "- %3c %4x %3c %#x%4u%4d", lookup_type ? 'I' : 'O', port_num, cls_lo & F_T6_SRAM_VLD ? 'Y' : 'N', G_PORTMAP(cls_hi), G_T6_PF(cls_lo), cls_lo & F_T6_VF_VALID ? G_T6_VF(cls_lo) : -1); } if (cls_lo & F_T6_REPLICATE) { struct fw_ldst_cmd ldst_cmd; memset(&ldst_cmd, 0, sizeof(ldst_cmd)); ldst_cmd.op_to_addrspace = htobe32(V_FW_CMD_OP(FW_LDST_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_READ | V_FW_LDST_CMD_ADDRSPACE(FW_LDST_ADDRSPC_MPS)); ldst_cmd.cycles_to_len16 = htobe32(FW_LEN16(ldst_cmd)); ldst_cmd.u.mps.rplc.fid_idx = htobe16(V_FW_LDST_CMD_FID(FW_LDST_MPS_RPLC) | V_FW_LDST_CMD_IDX(i)); rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t6mps"); if (rc) break; rc = -t4_wr_mbox(sc, sc->mbox, &ldst_cmd, sizeof(ldst_cmd), &ldst_cmd); end_synchronized_op(sc, 0); if (rc != 0) { sbuf_printf(sb, "%72d", rc); rc = 0; } else { sbuf_printf(sb, " %08x %08x %08x %08x" " %08x %08x %08x %08x", be32toh(ldst_cmd.u.mps.rplc.rplc255_224), be32toh(ldst_cmd.u.mps.rplc.rplc223_192), be32toh(ldst_cmd.u.mps.rplc.rplc191_160), be32toh(ldst_cmd.u.mps.rplc.rplc159_128), be32toh(ldst_cmd.u.mps.rplc.rplc127_96), be32toh(ldst_cmd.u.mps.rplc.rplc95_64), be32toh(ldst_cmd.u.mps.rplc.rplc63_32), be32toh(ldst_cmd.u.mps.rplc.rplc31_0)); } } else sbuf_printf(sb, "%72s", ""); sbuf_printf(sb, "%4u%3u%3u%3u %#x", G_T6_SRAM_PRIO0(cls_lo), G_T6_SRAM_PRIO1(cls_lo), G_T6_SRAM_PRIO2(cls_lo), G_T6_SRAM_PRIO3(cls_lo), (cls_lo >> S_T6_MULTILISTEN0) & 0xf); } if (rc) (void) sbuf_finish(sb); else rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_path_mtus(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; uint16_t mtus[NMTUS]; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); t4_read_mtu_tbl(sc, mtus, NULL); sbuf_printf(sb, "%u %u %u %u %u %u %u %u %u %u %u %u %u %u %u %u", mtus[0], mtus[1], mtus[2], mtus[3], mtus[4], mtus[5], mtus[6], mtus[7], mtus[8], mtus[9], mtus[10], mtus[11], mtus[12], mtus[13], mtus[14], mtus[15]); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_pm_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, i; uint32_t tx_cnt[MAX_PM_NSTATS], rx_cnt[MAX_PM_NSTATS]; uint64_t tx_cyc[MAX_PM_NSTATS], rx_cyc[MAX_PM_NSTATS]; static const char *tx_stats[MAX_PM_NSTATS] = { "Read:", "Write bypass:", "Write mem:", "Bypass + mem:", "Tx FIFO wait", NULL, "Tx latency" }; static const char *rx_stats[MAX_PM_NSTATS] = { "Read:", "Write bypass:", "Write mem:", "Flush:", " Rx FIFO wait", NULL, "Rx latency" }; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); t4_pmtx_get_stats(sc, tx_cnt, tx_cyc); t4_pmrx_get_stats(sc, rx_cnt, rx_cyc); sbuf_printf(sb, " Tx pcmds Tx bytes"); for (i = 0; i < 4; i++) { sbuf_printf(sb, "\n%-13s %10u %20ju", tx_stats[i], tx_cnt[i], tx_cyc[i]); } sbuf_printf(sb, "\n Rx pcmds Rx bytes"); for (i = 0; i < 4; i++) { sbuf_printf(sb, "\n%-13s %10u %20ju", rx_stats[i], rx_cnt[i], rx_cyc[i]); } if (chip_id(sc) > CHELSIO_T5) { sbuf_printf(sb, "\n Total wait Total occupancy"); sbuf_printf(sb, "\n%-13s %10u %20ju", tx_stats[i], tx_cnt[i], tx_cyc[i]); sbuf_printf(sb, "\n%-13s %10u %20ju", rx_stats[i], rx_cnt[i], rx_cyc[i]); i += 2; MPASS(i < nitems(tx_stats)); sbuf_printf(sb, "\n Reads Total wait"); sbuf_printf(sb, "\n%-13s %10u %20ju", tx_stats[i], tx_cnt[i], tx_cyc[i]); sbuf_printf(sb, "\n%-13s %10u %20ju", rx_stats[i], rx_cnt[i], rx_cyc[i]); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_rdma_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tp_rdma_stats stats; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); mtx_lock(&sc->reg_lock); t4_tp_get_rdma_stats(sc, &stats); mtx_unlock(&sc->reg_lock); sbuf_printf(sb, "NoRQEModDefferals: %u\n", stats.rqe_dfr_mod); sbuf_printf(sb, "NoRQEPktDefferals: %u", stats.rqe_dfr_pkt); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_tcp_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tp_tcp_stats v4, v6; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); mtx_lock(&sc->reg_lock); t4_tp_get_tcp_stats(sc, &v4, &v6); mtx_unlock(&sc->reg_lock); sbuf_printf(sb, " IP IPv6\n"); sbuf_printf(sb, "OutRsts: %20u %20u\n", v4.tcp_out_rsts, v6.tcp_out_rsts); sbuf_printf(sb, "InSegs: %20ju %20ju\n", v4.tcp_in_segs, v6.tcp_in_segs); sbuf_printf(sb, "OutSegs: %20ju %20ju\n", v4.tcp_out_segs, v6.tcp_out_segs); sbuf_printf(sb, "RetransSegs: %20ju %20ju", v4.tcp_retrans_segs, v6.tcp_retrans_segs); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_tids(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tid_info *t = &sc->tids; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); if (t->natids) { sbuf_printf(sb, "ATID range: 0-%u, in use: %u\n", t->natids - 1, t->atids_in_use); } if (t->ntids) { if (t4_read_reg(sc, A_LE_DB_CONFIG) & F_HASHEN) { uint32_t b = t4_read_reg(sc, A_LE_DB_SERVER_INDEX) / 4; if (b) { sbuf_printf(sb, "TID range: 0-%u, %u-%u", b - 1, t4_read_reg(sc, A_LE_DB_TID_HASHBASE) / 4, t->ntids - 1); } else { sbuf_printf(sb, "TID range: %u-%u", t4_read_reg(sc, A_LE_DB_TID_HASHBASE) / 4, t->ntids - 1); } } else sbuf_printf(sb, "TID range: 0-%u", t->ntids - 1); sbuf_printf(sb, ", in use: %u\n", atomic_load_acq_int(&t->tids_in_use)); } if (t->nstids) { sbuf_printf(sb, "STID range: %u-%u, in use: %u\n", t->stid_base, t->stid_base + t->nstids - 1, t->stids_in_use); } if (t->nftids) { sbuf_printf(sb, "FTID range: %u-%u\n", t->ftid_base, t->ftid_base + t->nftids - 1); } if (t->netids) { sbuf_printf(sb, "ETID range: %u-%u\n", t->etid_base, t->etid_base + t->netids - 1); } sbuf_printf(sb, "HW TID usage: %u IP users, %u IPv6 users", t4_read_reg(sc, A_LE_DB_ACT_CNT_IPV4), t4_read_reg(sc, A_LE_DB_ACT_CNT_IPV6)); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_tp_err_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; struct tp_err_stats stats; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); mtx_lock(&sc->reg_lock); t4_tp_get_err_stats(sc, &stats); mtx_unlock(&sc->reg_lock); if (sc->chip_params->nchan > 2) { sbuf_printf(sb, " channel 0 channel 1" " channel 2 channel 3\n"); sbuf_printf(sb, "macInErrs: %10u %10u %10u %10u\n", stats.mac_in_errs[0], stats.mac_in_errs[1], stats.mac_in_errs[2], stats.mac_in_errs[3]); sbuf_printf(sb, "hdrInErrs: %10u %10u %10u %10u\n", stats.hdr_in_errs[0], stats.hdr_in_errs[1], stats.hdr_in_errs[2], stats.hdr_in_errs[3]); sbuf_printf(sb, "tcpInErrs: %10u %10u %10u %10u\n", stats.tcp_in_errs[0], stats.tcp_in_errs[1], stats.tcp_in_errs[2], stats.tcp_in_errs[3]); sbuf_printf(sb, "tcp6InErrs: %10u %10u %10u %10u\n", stats.tcp6_in_errs[0], stats.tcp6_in_errs[1], stats.tcp6_in_errs[2], stats.tcp6_in_errs[3]); sbuf_printf(sb, "tnlCongDrops: %10u %10u %10u %10u\n", stats.tnl_cong_drops[0], stats.tnl_cong_drops[1], stats.tnl_cong_drops[2], stats.tnl_cong_drops[3]); sbuf_printf(sb, "tnlTxDrops: %10u %10u %10u %10u\n", stats.tnl_tx_drops[0], stats.tnl_tx_drops[1], stats.tnl_tx_drops[2], stats.tnl_tx_drops[3]); sbuf_printf(sb, "ofldVlanDrops: %10u %10u %10u %10u\n", stats.ofld_vlan_drops[0], stats.ofld_vlan_drops[1], stats.ofld_vlan_drops[2], stats.ofld_vlan_drops[3]); sbuf_printf(sb, "ofldChanDrops: %10u %10u %10u %10u\n\n", stats.ofld_chan_drops[0], stats.ofld_chan_drops[1], stats.ofld_chan_drops[2], stats.ofld_chan_drops[3]); } else { sbuf_printf(sb, " channel 0 channel 1\n"); sbuf_printf(sb, "macInErrs: %10u %10u\n", stats.mac_in_errs[0], stats.mac_in_errs[1]); sbuf_printf(sb, "hdrInErrs: %10u %10u\n", stats.hdr_in_errs[0], stats.hdr_in_errs[1]); sbuf_printf(sb, "tcpInErrs: %10u %10u\n", stats.tcp_in_errs[0], stats.tcp_in_errs[1]); sbuf_printf(sb, "tcp6InErrs: %10u %10u\n", stats.tcp6_in_errs[0], stats.tcp6_in_errs[1]); sbuf_printf(sb, "tnlCongDrops: %10u %10u\n", stats.tnl_cong_drops[0], stats.tnl_cong_drops[1]); sbuf_printf(sb, "tnlTxDrops: %10u %10u\n", stats.tnl_tx_drops[0], stats.tnl_tx_drops[1]); sbuf_printf(sb, "ofldVlanDrops: %10u %10u\n", stats.ofld_vlan_drops[0], stats.ofld_vlan_drops[1]); sbuf_printf(sb, "ofldChanDrops: %10u %10u\n\n", stats.ofld_chan_drops[0], stats.ofld_chan_drops[1]); } sbuf_printf(sb, "ofldNoNeigh: %u\nofldCongDefer: %u", stats.ofld_no_neigh, stats.ofld_cong_defer); rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_tp_la_mask(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct tp_params *tpp = &sc->params.tp; u_int mask; int rc; mask = tpp->la_mask >> 16; rc = sysctl_handle_int(oidp, &mask, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); if (mask > 0xffff) return (EINVAL); tpp->la_mask = mask << 16; t4_set_reg_field(sc, A_TP_DBG_LA_CONFIG, 0xffff0000U, tpp->la_mask); return (0); } struct field_desc { const char *name; u_int start; u_int width; }; static void field_desc_show(struct sbuf *sb, uint64_t v, const struct field_desc *f) { char buf[32]; int line_size = 0; while (f->name) { uint64_t mask = (1ULL << f->width) - 1; int len = snprintf(buf, sizeof(buf), "%s: %ju", f->name, ((uintmax_t)v >> f->start) & mask); if (line_size + len >= 79) { line_size = 8; sbuf_printf(sb, "\n "); } sbuf_printf(sb, "%s ", buf); line_size += len + 1; f++; } sbuf_printf(sb, "\n"); } static const struct field_desc tp_la0[] = { { "RcfOpCodeOut", 60, 4 }, { "State", 56, 4 }, { "WcfState", 52, 4 }, { "RcfOpcSrcOut", 50, 2 }, { "CRxError", 49, 1 }, { "ERxError", 48, 1 }, { "SanityFailed", 47, 1 }, { "SpuriousMsg", 46, 1 }, { "FlushInputMsg", 45, 1 }, { "FlushInputCpl", 44, 1 }, { "RssUpBit", 43, 1 }, { "RssFilterHit", 42, 1 }, { "Tid", 32, 10 }, { "InitTcb", 31, 1 }, { "LineNumber", 24, 7 }, { "Emsg", 23, 1 }, { "EdataOut", 22, 1 }, { "Cmsg", 21, 1 }, { "CdataOut", 20, 1 }, { "EreadPdu", 19, 1 }, { "CreadPdu", 18, 1 }, { "TunnelPkt", 17, 1 }, { "RcfPeerFin", 16, 1 }, { "RcfReasonOut", 12, 4 }, { "TxCchannel", 10, 2 }, { "RcfTxChannel", 8, 2 }, { "RxEchannel", 6, 2 }, { "RcfRxChannel", 5, 1 }, { "RcfDataOutSrdy", 4, 1 }, { "RxDvld", 3, 1 }, { "RxOoDvld", 2, 1 }, { "RxCongestion", 1, 1 }, { "TxCongestion", 0, 1 }, { NULL } }; static const struct field_desc tp_la1[] = { { "CplCmdIn", 56, 8 }, { "CplCmdOut", 48, 8 }, { "ESynOut", 47, 1 }, { "EAckOut", 46, 1 }, { "EFinOut", 45, 1 }, { "ERstOut", 44, 1 }, { "SynIn", 43, 1 }, { "AckIn", 42, 1 }, { "FinIn", 41, 1 }, { "RstIn", 40, 1 }, { "DataIn", 39, 1 }, { "DataInVld", 38, 1 }, { "PadIn", 37, 1 }, { "RxBufEmpty", 36, 1 }, { "RxDdp", 35, 1 }, { "RxFbCongestion", 34, 1 }, { "TxFbCongestion", 33, 1 }, { "TxPktSumSrdy", 32, 1 }, { "RcfUlpType", 28, 4 }, { "Eread", 27, 1 }, { "Ebypass", 26, 1 }, { "Esave", 25, 1 }, { "Static0", 24, 1 }, { "Cread", 23, 1 }, { "Cbypass", 22, 1 }, { "Csave", 21, 1 }, { "CPktOut", 20, 1 }, { "RxPagePoolFull", 18, 2 }, { "RxLpbkPkt", 17, 1 }, { "TxLpbkPkt", 16, 1 }, { "RxVfValid", 15, 1 }, { "SynLearned", 14, 1 }, { "SetDelEntry", 13, 1 }, { "SetInvEntry", 12, 1 }, { "CpcmdDvld", 11, 1 }, { "CpcmdSave", 10, 1 }, { "RxPstructsFull", 8, 2 }, { "EpcmdDvld", 7, 1 }, { "EpcmdFlush", 6, 1 }, { "EpcmdTrimPrefix", 5, 1 }, { "EpcmdTrimPostfix", 4, 1 }, { "ERssIp4Pkt", 3, 1 }, { "ERssIp6Pkt", 2, 1 }, { "ERssTcpUdpPkt", 1, 1 }, { "ERssFceFipPkt", 0, 1 }, { NULL } }; static const struct field_desc tp_la2[] = { { "CplCmdIn", 56, 8 }, { "MpsVfVld", 55, 1 }, { "MpsPf", 52, 3 }, { "MpsVf", 44, 8 }, { "SynIn", 43, 1 }, { "AckIn", 42, 1 }, { "FinIn", 41, 1 }, { "RstIn", 40, 1 }, { "DataIn", 39, 1 }, { "DataInVld", 38, 1 }, { "PadIn", 37, 1 }, { "RxBufEmpty", 36, 1 }, { "RxDdp", 35, 1 }, { "RxFbCongestion", 34, 1 }, { "TxFbCongestion", 33, 1 }, { "TxPktSumSrdy", 32, 1 }, { "RcfUlpType", 28, 4 }, { "Eread", 27, 1 }, { "Ebypass", 26, 1 }, { "Esave", 25, 1 }, { "Static0", 24, 1 }, { "Cread", 23, 1 }, { "Cbypass", 22, 1 }, { "Csave", 21, 1 }, { "CPktOut", 20, 1 }, { "RxPagePoolFull", 18, 2 }, { "RxLpbkPkt", 17, 1 }, { "TxLpbkPkt", 16, 1 }, { "RxVfValid", 15, 1 }, { "SynLearned", 14, 1 }, { "SetDelEntry", 13, 1 }, { "SetInvEntry", 12, 1 }, { "CpcmdDvld", 11, 1 }, { "CpcmdSave", 10, 1 }, { "RxPstructsFull", 8, 2 }, { "EpcmdDvld", 7, 1 }, { "EpcmdFlush", 6, 1 }, { "EpcmdTrimPrefix", 5, 1 }, { "EpcmdTrimPostfix", 4, 1 }, { "ERssIp4Pkt", 3, 1 }, { "ERssIp6Pkt", 2, 1 }, { "ERssTcpUdpPkt", 1, 1 }, { "ERssFceFipPkt", 0, 1 }, { NULL } }; static void tp_la_show(struct sbuf *sb, uint64_t *p, int idx) { field_desc_show(sb, *p, tp_la0); } static void tp_la_show2(struct sbuf *sb, uint64_t *p, int idx) { if (idx) sbuf_printf(sb, "\n"); field_desc_show(sb, p[0], tp_la0); if (idx < (TPLA_SIZE / 2 - 1) || p[1] != ~0ULL) field_desc_show(sb, p[1], tp_la0); } static void tp_la_show3(struct sbuf *sb, uint64_t *p, int idx) { if (idx) sbuf_printf(sb, "\n"); field_desc_show(sb, p[0], tp_la0); if (idx < (TPLA_SIZE / 2 - 1) || p[1] != ~0ULL) field_desc_show(sb, p[1], (p[0] & (1 << 17)) ? tp_la2 : tp_la1); } static int sysctl_tp_la(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; uint64_t *buf, *p; int rc; u_int i, inc; void (*show_func)(struct sbuf *, uint64_t *, int); rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); buf = malloc(TPLA_SIZE * sizeof(uint64_t), M_CXGBE, M_ZERO | M_WAITOK); t4_tp_read_la(sc, buf, NULL); p = buf; switch (G_DBGLAMODE(t4_read_reg(sc, A_TP_DBG_LA_CONFIG))) { case 2: inc = 2; show_func = tp_la_show2; break; case 3: inc = 2; show_func = tp_la_show3; break; default: inc = 1; show_func = tp_la_show; } for (i = 0; i < TPLA_SIZE / inc; i++, p += inc) (*show_func)(sb, p, i); rc = sbuf_finish(sb); sbuf_delete(sb); free(buf, M_CXGBE); return (rc); } static int sysctl_tx_rate(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc; u64 nrate[MAX_NCHAN], orate[MAX_NCHAN]; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 256, req); if (sb == NULL) return (ENOMEM); t4_get_chan_txrate(sc, nrate, orate); if (sc->chip_params->nchan > 2) { sbuf_printf(sb, " channel 0 channel 1" " channel 2 channel 3\n"); sbuf_printf(sb, "NIC B/s: %10ju %10ju %10ju %10ju\n", nrate[0], nrate[1], nrate[2], nrate[3]); sbuf_printf(sb, "Offload B/s: %10ju %10ju %10ju %10ju", orate[0], orate[1], orate[2], orate[3]); } else { sbuf_printf(sb, " channel 0 channel 1\n"); sbuf_printf(sb, "NIC B/s: %10ju %10ju\n", nrate[0], nrate[1]); sbuf_printf(sb, "Offload B/s: %10ju %10ju", orate[0], orate[1]); } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_ulprx_la(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; uint32_t *buf, *p; int rc, i; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); buf = malloc(ULPRX_LA_SIZE * 8 * sizeof(uint32_t), M_CXGBE, M_ZERO | M_WAITOK); t4_ulprx_read_la(sc, buf); p = buf; sbuf_printf(sb, " Pcmd Type Message" " Data"); for (i = 0; i < ULPRX_LA_SIZE; i++, p += 8) { sbuf_printf(sb, "\n%08x%08x %4x %08x %08x%08x%08x%08x", p[1], p[0], p[2], p[3], p[7], p[6], p[5], p[4]); } rc = sbuf_finish(sb); sbuf_delete(sb); free(buf, M_CXGBE); return (rc); } static int sysctl_wcwr_stats(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct sbuf *sb; int rc, v; + MPASS(chip_id(sc) >= CHELSIO_T5); + rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); v = t4_read_reg(sc, A_SGE_STAT_CFG); if (G_STATSOURCE_T5(v) == 7) { - if (G_STATMODE(v) == 0) { + int mode; + + mode = is_t5(sc) ? G_STATMODE(v) : G_T6_STATMODE(v); + if (mode == 0) { sbuf_printf(sb, "total %d, incomplete %d", t4_read_reg(sc, A_SGE_STAT_TOTAL), t4_read_reg(sc, A_SGE_STAT_MATCH)); - } else if (G_STATMODE(v) == 1) { + } else if (mode == 1) { sbuf_printf(sb, "total %d, data overflow %d", t4_read_reg(sc, A_SGE_STAT_TOTAL), t4_read_reg(sc, A_SGE_STAT_MATCH)); + } else { + sbuf_printf(sb, "unknown mode %d", mode); } } rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } static int sysctl_tc_params(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; struct tx_sched_class *tc; struct t4_sched_class_params p; struct sbuf *sb; int i, rc, port_id, flags, mbps, gbps; rc = sysctl_wire_old_buffer(req, 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 4096, req); if (sb == NULL) return (ENOMEM); port_id = arg2 >> 16; MPASS(port_id < sc->params.nports); MPASS(sc->port[port_id] != NULL); i = arg2 & 0xffff; MPASS(i < sc->chip_params->nsched_cls); tc = &sc->port[port_id]->tc[i]; rc = begin_synchronized_op(sc, NULL, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4tc_p"); if (rc) goto done; flags = tc->flags; p = tc->params; end_synchronized_op(sc, LOCK_HELD); if ((flags & TX_SC_OK) == 0) { sbuf_printf(sb, "none"); goto done; } if (p.level == SCHED_CLASS_LEVEL_CL_WRR) { sbuf_printf(sb, "cl-wrr weight %u", p.weight); goto done; } else if (p.level == SCHED_CLASS_LEVEL_CL_RL) sbuf_printf(sb, "cl-rl"); else if (p.level == SCHED_CLASS_LEVEL_CH_RL) sbuf_printf(sb, "ch-rl"); else { rc = ENXIO; goto done; } if (p.ratemode == SCHED_CLASS_RATEMODE_REL) { /* XXX: top speed or actual link speed? */ gbps = port_top_speed(sc->port[port_id]); sbuf_printf(sb, " %u%% of %uGbps", p.maxrate, gbps); } else if (p.ratemode == SCHED_CLASS_RATEMODE_ABS) { switch (p.rateunit) { case SCHED_CLASS_RATEUNIT_BITS: mbps = p.maxrate / 1000; gbps = p.maxrate / 1000000; if (p.maxrate == gbps * 1000000) sbuf_printf(sb, " %uGbps", gbps); else if (p.maxrate == mbps * 1000) sbuf_printf(sb, " %uMbps", mbps); else sbuf_printf(sb, " %uKbps", p.maxrate); break; case SCHED_CLASS_RATEUNIT_PKTS: sbuf_printf(sb, " %upps", p.maxrate); break; default: rc = ENXIO; goto done; } } switch (p.mode) { case SCHED_CLASS_MODE_CLASS: sbuf_printf(sb, " aggregate"); break; case SCHED_CLASS_MODE_FLOW: sbuf_printf(sb, " per-flow"); break; default: rc = ENXIO; goto done; } done: if (rc == 0) rc = sbuf_finish(sb); sbuf_delete(sb); return (rc); } #endif #ifdef TCP_OFFLOAD static void unit_conv(char *buf, size_t len, u_int val, u_int factor) { u_int rem = val % factor; if (rem == 0) snprintf(buf, len, "%u", val / factor); else { while (rem % 10 == 0) rem /= 10; snprintf(buf, len, "%u.%u", val / factor, rem); } } static int sysctl_tp_tick(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; char buf[16]; u_int res, re; u_int cclk_ps = 1000000000 / sc->params.vpd.cclk; res = t4_read_reg(sc, A_TP_TIMER_RESOLUTION); switch (arg2) { case 0: /* timer_tick */ re = G_TIMERRESOLUTION(res); break; case 1: /* TCP timestamp tick */ re = G_TIMESTAMPRESOLUTION(res); break; case 2: /* DACK tick */ re = G_DELAYEDACKRESOLUTION(res); break; default: return (EDOOFUS); } unit_conv(buf, sizeof(buf), (cclk_ps << re), 1000000); return (sysctl_handle_string(oidp, buf, sizeof(buf), req)); } static int sysctl_tp_dack_timer(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; u_int res, dack_re, v; u_int cclk_ps = 1000000000 / sc->params.vpd.cclk; res = t4_read_reg(sc, A_TP_TIMER_RESOLUTION); dack_re = G_DELAYEDACKRESOLUTION(res); v = ((cclk_ps << dack_re) / 1000000) * t4_read_reg(sc, A_TP_DACK_TIMER); return (sysctl_handle_int(oidp, &v, 0, req)); } static int sysctl_tp_timer(SYSCTL_HANDLER_ARGS) { struct adapter *sc = arg1; int reg = arg2; u_int tre; u_long tp_tick_us, v; u_int cclk_ps = 1000000000 / sc->params.vpd.cclk; MPASS(reg == A_TP_RXT_MIN || reg == A_TP_RXT_MAX || reg == A_TP_PERS_MIN || reg == A_TP_PERS_MAX || reg == A_TP_KEEP_IDLE || A_TP_KEEP_INTVL || reg == A_TP_INIT_SRTT || reg == A_TP_FINWAIT2_TIMER); tre = G_TIMERRESOLUTION(t4_read_reg(sc, A_TP_TIMER_RESOLUTION)); tp_tick_us = (cclk_ps << tre) / 1000000; if (reg == A_TP_INIT_SRTT) v = tp_tick_us * G_INITSRTT(t4_read_reg(sc, reg)); else v = tp_tick_us * t4_read_reg(sc, reg); return (sysctl_handle_long(oidp, &v, 0, req)); } #endif static uint32_t fconf_iconf_to_mode(uint32_t fconf, uint32_t iconf) { uint32_t mode; mode = T4_FILTER_IPv4 | T4_FILTER_IPv6 | T4_FILTER_IP_SADDR | T4_FILTER_IP_DADDR | T4_FILTER_IP_SPORT | T4_FILTER_IP_DPORT; if (fconf & F_FRAGMENTATION) mode |= T4_FILTER_IP_FRAGMENT; if (fconf & F_MPSHITTYPE) mode |= T4_FILTER_MPS_HIT_TYPE; if (fconf & F_MACMATCH) mode |= T4_FILTER_MAC_IDX; if (fconf & F_ETHERTYPE) mode |= T4_FILTER_ETH_TYPE; if (fconf & F_PROTOCOL) mode |= T4_FILTER_IP_PROTO; if (fconf & F_TOS) mode |= T4_FILTER_IP_TOS; if (fconf & F_VLAN) mode |= T4_FILTER_VLAN; if (fconf & F_VNIC_ID) { mode |= T4_FILTER_VNIC; if (iconf & F_VNIC) mode |= T4_FILTER_IC_VNIC; } if (fconf & F_PORT) mode |= T4_FILTER_PORT; if (fconf & F_FCOE) mode |= T4_FILTER_FCoE; return (mode); } static uint32_t mode_to_fconf(uint32_t mode) { uint32_t fconf = 0; if (mode & T4_FILTER_IP_FRAGMENT) fconf |= F_FRAGMENTATION; if (mode & T4_FILTER_MPS_HIT_TYPE) fconf |= F_MPSHITTYPE; if (mode & T4_FILTER_MAC_IDX) fconf |= F_MACMATCH; if (mode & T4_FILTER_ETH_TYPE) fconf |= F_ETHERTYPE; if (mode & T4_FILTER_IP_PROTO) fconf |= F_PROTOCOL; if (mode & T4_FILTER_IP_TOS) fconf |= F_TOS; if (mode & T4_FILTER_VLAN) fconf |= F_VLAN; if (mode & T4_FILTER_VNIC) fconf |= F_VNIC_ID; if (mode & T4_FILTER_PORT) fconf |= F_PORT; if (mode & T4_FILTER_FCoE) fconf |= F_FCOE; return (fconf); } static uint32_t mode_to_iconf(uint32_t mode) { if (mode & T4_FILTER_IC_VNIC) return (F_VNIC); return (0); } static int check_fspec_against_fconf_iconf(struct adapter *sc, struct t4_filter_specification *fs) { struct tp_params *tpp = &sc->params.tp; uint32_t fconf = 0; if (fs->val.frag || fs->mask.frag) fconf |= F_FRAGMENTATION; if (fs->val.matchtype || fs->mask.matchtype) fconf |= F_MPSHITTYPE; if (fs->val.macidx || fs->mask.macidx) fconf |= F_MACMATCH; if (fs->val.ethtype || fs->mask.ethtype) fconf |= F_ETHERTYPE; if (fs->val.proto || fs->mask.proto) fconf |= F_PROTOCOL; if (fs->val.tos || fs->mask.tos) fconf |= F_TOS; if (fs->val.vlan_vld || fs->mask.vlan_vld) fconf |= F_VLAN; if (fs->val.ovlan_vld || fs->mask.ovlan_vld) { fconf |= F_VNIC_ID; if (tpp->ingress_config & F_VNIC) return (EINVAL); } if (fs->val.pfvf_vld || fs->mask.pfvf_vld) { fconf |= F_VNIC_ID; if ((tpp->ingress_config & F_VNIC) == 0) return (EINVAL); } if (fs->val.iport || fs->mask.iport) fconf |= F_PORT; if (fs->val.fcoe || fs->mask.fcoe) fconf |= F_FCOE; if ((tpp->vlan_pri_map | fconf) != tpp->vlan_pri_map) return (E2BIG); return (0); } static int get_filter_mode(struct adapter *sc, uint32_t *mode) { struct tp_params *tpp = &sc->params.tp; /* * We trust the cached values of the relevant TP registers. This means * things work reliably only if writes to those registers are always via * t4_set_filter_mode. */ *mode = fconf_iconf_to_mode(tpp->vlan_pri_map, tpp->ingress_config); return (0); } static int set_filter_mode(struct adapter *sc, uint32_t mode) { struct tp_params *tpp = &sc->params.tp; uint32_t fconf, iconf; int rc; iconf = mode_to_iconf(mode); if ((iconf ^ tpp->ingress_config) & F_VNIC) { /* * For now we just complain if A_TP_INGRESS_CONFIG is not * already set to the correct value for the requested filter * mode. It's not clear if it's safe to write to this register * on the fly. (And we trust the cached value of the register). */ return (EBUSY); } fconf = mode_to_fconf(mode); rc = begin_synchronized_op(sc, NULL, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4setfm"); if (rc) return (rc); if (sc->tids.ftids_in_use > 0) { rc = EBUSY; goto done; } #ifdef TCP_OFFLOAD if (uld_active(sc, ULD_TOM)) { rc = EBUSY; goto done; } #endif rc = -t4_set_filter_mode(sc, fconf); done: end_synchronized_op(sc, LOCK_HELD); return (rc); } static inline uint64_t get_filter_hits(struct adapter *sc, uint32_t fid) { uint32_t tcb_addr; tcb_addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + (fid + sc->tids.ftid_base) * TCB_SIZE; if (is_t4(sc)) { uint64_t hits; read_via_memwin(sc, 0, tcb_addr + 16, (uint32_t *)&hits, 8); return (be64toh(hits)); } else { uint32_t hits; read_via_memwin(sc, 0, tcb_addr + 24, &hits, 4); return (be32toh(hits)); } } static int get_filter(struct adapter *sc, struct t4_filter *t) { int i, rc, nfilters = sc->tids.nftids; struct filter_entry *f; rc = begin_synchronized_op(sc, NULL, HOLD_LOCK | SLEEP_OK | INTR_OK, "t4getf"); if (rc) return (rc); if (sc->tids.ftids_in_use == 0 || sc->tids.ftid_tab == NULL || t->idx >= nfilters) { t->idx = 0xffffffff; goto done; } f = &sc->tids.ftid_tab[t->idx]; for (i = t->idx; i < nfilters; i++, f++) { if (f->valid) { t->idx = i; t->l2tidx = f->l2t ? f->l2t->idx : 0; t->smtidx = f->smtidx; if (f->fs.hitcnts) t->hits = get_filter_hits(sc, t->idx); else t->hits = UINT64_MAX; t->fs = f->fs; goto done; } } t->idx = 0xffffffff; done: end_synchronized_op(sc, LOCK_HELD); return (0); } static int set_filter(struct adapter *sc, struct t4_filter *t) { unsigned int nfilters, nports; struct filter_entry *f; int i, rc; rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4setf"); if (rc) return (rc); nfilters = sc->tids.nftids; nports = sc->params.nports; if (nfilters == 0) { rc = ENOTSUP; goto done; } if (t->idx >= nfilters) { rc = EINVAL; goto done; } /* Validate against the global filter mode and ingress config */ rc = check_fspec_against_fconf_iconf(sc, &t->fs); if (rc != 0) goto done; if (t->fs.action == FILTER_SWITCH && t->fs.eport >= nports) { rc = EINVAL; goto done; } if (t->fs.val.iport >= nports) { rc = EINVAL; goto done; } /* Can't specify an iq if not steering to it */ if (!t->fs.dirsteer && t->fs.iq) { rc = EINVAL; goto done; } /* IPv6 filter idx must be 4 aligned */ if (t->fs.type == 1 && ((t->idx & 0x3) || t->idx + 4 >= nfilters)) { rc = EINVAL; goto done; } if (!(sc->flags & FULL_INIT_DONE) && ((rc = adapter_full_init(sc)) != 0)) goto done; if (sc->tids.ftid_tab == NULL) { KASSERT(sc->tids.ftids_in_use == 0, ("%s: no memory allocated but filters_in_use > 0", __func__)); sc->tids.ftid_tab = malloc(sizeof (struct filter_entry) * nfilters, M_CXGBE, M_NOWAIT | M_ZERO); if (sc->tids.ftid_tab == NULL) { rc = ENOMEM; goto done; } mtx_init(&sc->tids.ftid_lock, "T4 filters", 0, MTX_DEF); } for (i = 0; i < 4; i++) { f = &sc->tids.ftid_tab[t->idx + i]; if (f->pending || f->valid) { rc = EBUSY; goto done; } if (f->locked) { rc = EPERM; goto done; } if (t->fs.type == 0) break; } f = &sc->tids.ftid_tab[t->idx]; f->fs = t->fs; rc = set_filter_wr(sc, t->idx); done: end_synchronized_op(sc, 0); if (rc == 0) { mtx_lock(&sc->tids.ftid_lock); for (;;) { if (f->pending == 0) { rc = f->valid ? 0 : EIO; break; } if (mtx_sleep(&sc->tids.ftid_tab, &sc->tids.ftid_lock, PCATCH, "t4setfw", 0)) { rc = EINPROGRESS; break; } } mtx_unlock(&sc->tids.ftid_lock); } return (rc); } static int del_filter(struct adapter *sc, struct t4_filter *t) { unsigned int nfilters; struct filter_entry *f; int rc; rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4delf"); if (rc) return (rc); nfilters = sc->tids.nftids; if (nfilters == 0) { rc = ENOTSUP; goto done; } if (sc->tids.ftid_tab == NULL || sc->tids.ftids_in_use == 0 || t->idx >= nfilters) { rc = EINVAL; goto done; } if (!(sc->flags & FULL_INIT_DONE)) { rc = EAGAIN; goto done; } f = &sc->tids.ftid_tab[t->idx]; if (f->pending) { rc = EBUSY; goto done; } if (f->locked) { rc = EPERM; goto done; } if (f->valid) { t->fs = f->fs; /* extra info for the caller */ rc = del_filter_wr(sc, t->idx); } done: end_synchronized_op(sc, 0); if (rc == 0) { mtx_lock(&sc->tids.ftid_lock); for (;;) { if (f->pending == 0) { rc = f->valid ? EIO : 0; break; } if (mtx_sleep(&sc->tids.ftid_tab, &sc->tids.ftid_lock, PCATCH, "t4delfw", 0)) { rc = EINPROGRESS; break; } } mtx_unlock(&sc->tids.ftid_lock); } return (rc); } static void clear_filter(struct filter_entry *f) { if (f->l2t) t4_l2t_release(f->l2t); bzero(f, sizeof (*f)); } static int set_filter_wr(struct adapter *sc, int fidx) { struct filter_entry *f = &sc->tids.ftid_tab[fidx]; struct fw_filter_wr *fwr; unsigned int ftid, vnic_vld, vnic_vld_mask; struct wrq_cookie cookie; ASSERT_SYNCHRONIZED_OP(sc); if (f->fs.newdmac || f->fs.newvlan) { /* This filter needs an L2T entry; allocate one. */ f->l2t = t4_l2t_alloc_switching(sc->l2t); if (f->l2t == NULL) return (EAGAIN); if (t4_l2t_set_switching(sc, f->l2t, f->fs.vlan, f->fs.eport, f->fs.dmac)) { t4_l2t_release(f->l2t); f->l2t = NULL; return (ENOMEM); } } /* Already validated against fconf, iconf */ MPASS((f->fs.val.pfvf_vld & f->fs.val.ovlan_vld) == 0); MPASS((f->fs.mask.pfvf_vld & f->fs.mask.ovlan_vld) == 0); if (f->fs.val.pfvf_vld || f->fs.val.ovlan_vld) vnic_vld = 1; else vnic_vld = 0; if (f->fs.mask.pfvf_vld || f->fs.mask.ovlan_vld) vnic_vld_mask = 1; else vnic_vld_mask = 0; ftid = sc->tids.ftid_base + fidx; fwr = start_wrq_wr(&sc->sge.mgmtq, howmany(sizeof(*fwr), 16), &cookie); if (fwr == NULL) return (ENOMEM); bzero(fwr, sizeof(*fwr)); fwr->op_pkd = htobe32(V_FW_WR_OP(FW_FILTER_WR)); fwr->len16_pkd = htobe32(FW_LEN16(*fwr)); fwr->tid_to_iq = htobe32(V_FW_FILTER_WR_TID(ftid) | V_FW_FILTER_WR_RQTYPE(f->fs.type) | V_FW_FILTER_WR_NOREPLY(0) | V_FW_FILTER_WR_IQ(f->fs.iq)); fwr->del_filter_to_l2tix = htobe32(V_FW_FILTER_WR_RPTTID(f->fs.rpttid) | V_FW_FILTER_WR_DROP(f->fs.action == FILTER_DROP) | V_FW_FILTER_WR_DIRSTEER(f->fs.dirsteer) | V_FW_FILTER_WR_MASKHASH(f->fs.maskhash) | V_FW_FILTER_WR_DIRSTEERHASH(f->fs.dirsteerhash) | V_FW_FILTER_WR_LPBK(f->fs.action == FILTER_SWITCH) | V_FW_FILTER_WR_DMAC(f->fs.newdmac) | V_FW_FILTER_WR_SMAC(f->fs.newsmac) | V_FW_FILTER_WR_INSVLAN(f->fs.newvlan == VLAN_INSERT || f->fs.newvlan == VLAN_REWRITE) | V_FW_FILTER_WR_RMVLAN(f->fs.newvlan == VLAN_REMOVE || f->fs.newvlan == VLAN_REWRITE) | V_FW_FILTER_WR_HITCNTS(f->fs.hitcnts) | V_FW_FILTER_WR_TXCHAN(f->fs.eport) | V_FW_FILTER_WR_PRIO(f->fs.prio) | V_FW_FILTER_WR_L2TIX(f->l2t ? f->l2t->idx : 0)); fwr->ethtype = htobe16(f->fs.val.ethtype); fwr->ethtypem = htobe16(f->fs.mask.ethtype); fwr->frag_to_ovlan_vldm = (V_FW_FILTER_WR_FRAG(f->fs.val.frag) | V_FW_FILTER_WR_FRAGM(f->fs.mask.frag) | V_FW_FILTER_WR_IVLAN_VLD(f->fs.val.vlan_vld) | V_FW_FILTER_WR_OVLAN_VLD(vnic_vld) | V_FW_FILTER_WR_IVLAN_VLDM(f->fs.mask.vlan_vld) | V_FW_FILTER_WR_OVLAN_VLDM(vnic_vld_mask)); fwr->smac_sel = 0; fwr->rx_chan_rx_rpl_iq = htobe16(V_FW_FILTER_WR_RX_CHAN(0) | V_FW_FILTER_WR_RX_RPL_IQ(sc->sge.fwq.abs_id)); fwr->maci_to_matchtypem = htobe32(V_FW_FILTER_WR_MACI(f->fs.val.macidx) | V_FW_FILTER_WR_MACIM(f->fs.mask.macidx) | V_FW_FILTER_WR_FCOE(f->fs.val.fcoe) | V_FW_FILTER_WR_FCOEM(f->fs.mask.fcoe) | V_FW_FILTER_WR_PORT(f->fs.val.iport) | V_FW_FILTER_WR_PORTM(f->fs.mask.iport) | V_FW_FILTER_WR_MATCHTYPE(f->fs.val.matchtype) | V_FW_FILTER_WR_MATCHTYPEM(f->fs.mask.matchtype)); fwr->ptcl = f->fs.val.proto; fwr->ptclm = f->fs.mask.proto; fwr->ttyp = f->fs.val.tos; fwr->ttypm = f->fs.mask.tos; fwr->ivlan = htobe16(f->fs.val.vlan); fwr->ivlanm = htobe16(f->fs.mask.vlan); fwr->ovlan = htobe16(f->fs.val.vnic); fwr->ovlanm = htobe16(f->fs.mask.vnic); bcopy(f->fs.val.dip, fwr->lip, sizeof (fwr->lip)); bcopy(f->fs.mask.dip, fwr->lipm, sizeof (fwr->lipm)); bcopy(f->fs.val.sip, fwr->fip, sizeof (fwr->fip)); bcopy(f->fs.mask.sip, fwr->fipm, sizeof (fwr->fipm)); fwr->lp = htobe16(f->fs.val.dport); fwr->lpm = htobe16(f->fs.mask.dport); fwr->fp = htobe16(f->fs.val.sport); fwr->fpm = htobe16(f->fs.mask.sport); if (f->fs.newsmac) bcopy(f->fs.smac, fwr->sma, sizeof (fwr->sma)); f->pending = 1; sc->tids.ftids_in_use++; commit_wrq_wr(&sc->sge.mgmtq, fwr, &cookie); return (0); } static int del_filter_wr(struct adapter *sc, int fidx) { struct filter_entry *f = &sc->tids.ftid_tab[fidx]; struct fw_filter_wr *fwr; unsigned int ftid; struct wrq_cookie cookie; ftid = sc->tids.ftid_base + fidx; fwr = start_wrq_wr(&sc->sge.mgmtq, howmany(sizeof(*fwr), 16), &cookie); if (fwr == NULL) return (ENOMEM); bzero(fwr, sizeof (*fwr)); t4_mk_filtdelwr(ftid, fwr, sc->sge.fwq.abs_id); f->pending = 1; commit_wrq_wr(&sc->sge.mgmtq, fwr, &cookie); return (0); } int t4_filter_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_set_tcb_rpl *rpl = (const void *)(rss + 1); unsigned int idx = GET_TID(rpl); unsigned int rc; struct filter_entry *f; KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, rss->opcode)); MPASS(iq == &sc->sge.fwq); MPASS(is_ftid(sc, idx)); idx -= sc->tids.ftid_base; f = &sc->tids.ftid_tab[idx]; rc = G_COOKIE(rpl->cookie); mtx_lock(&sc->tids.ftid_lock); if (rc == FW_FILTER_WR_FLT_ADDED) { KASSERT(f->pending, ("%s: filter[%u] isn't pending.", __func__, idx)); f->smtidx = (be64toh(rpl->oldval) >> 24) & 0xff; f->pending = 0; /* asynchronous setup completed */ f->valid = 1; } else { if (rc != FW_FILTER_WR_FLT_DELETED) { /* Add or delete failed, display an error */ log(LOG_ERR, "filter %u setup failed with error %u\n", idx, rc); } clear_filter(f); sc->tids.ftids_in_use--; } wakeup(&sc->tids.ftid_tab); mtx_unlock(&sc->tids.ftid_lock); return (0); } static int set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { MPASS(iq->set_tcb_rpl != NULL); return (iq->set_tcb_rpl(iq, rss, m)); } static int l2t_write_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { MPASS(iq->l2t_write_rpl != NULL); return (iq->l2t_write_rpl(iq, rss, m)); } static int get_sge_context(struct adapter *sc, struct t4_sge_context *cntxt) { int rc; if (cntxt->cid > M_CTXTQID) return (EINVAL); if (cntxt->mem_id != CTXT_EGRESS && cntxt->mem_id != CTXT_INGRESS && cntxt->mem_id != CTXT_FLM && cntxt->mem_id != CTXT_CNM) return (EINVAL); rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4ctxt"); if (rc) return (rc); if (sc->flags & FW_OK) { rc = -t4_sge_ctxt_rd(sc, sc->mbox, cntxt->cid, cntxt->mem_id, &cntxt->data[0]); if (rc == 0) goto done; } /* * Read via firmware failed or wasn't even attempted. Read directly via * the backdoor. */ rc = -t4_sge_ctxt_rd_bd(sc, cntxt->cid, cntxt->mem_id, &cntxt->data[0]); done: end_synchronized_op(sc, 0); return (rc); } static int load_fw(struct adapter *sc, struct t4_data *fw) { int rc; uint8_t *fw_data; rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4ldfw"); if (rc) return (rc); if (sc->flags & FULL_INIT_DONE) { rc = EBUSY; goto done; } fw_data = malloc(fw->len, M_CXGBE, M_WAITOK); if (fw_data == NULL) { rc = ENOMEM; goto done; } rc = copyin(fw->data, fw_data, fw->len); if (rc == 0) rc = -t4_load_fw(sc, fw_data, fw->len); free(fw_data, M_CXGBE); done: end_synchronized_op(sc, 0); return (rc); } #define MAX_READ_BUF_SIZE (128 * 1024) static int read_card_mem(struct adapter *sc, int win, struct t4_mem_range *mr) { uint32_t addr, remaining, n; uint32_t *buf; int rc; uint8_t *dst; rc = validate_mem_range(sc, mr->addr, mr->len); if (rc != 0) return (rc); buf = malloc(min(mr->len, MAX_READ_BUF_SIZE), M_CXGBE, M_WAITOK); addr = mr->addr; remaining = mr->len; dst = (void *)mr->data; while (remaining) { n = min(remaining, MAX_READ_BUF_SIZE); read_via_memwin(sc, 2, addr, buf, n); rc = copyout(buf, dst, n); if (rc != 0) break; dst += n; remaining -= n; addr += n; } free(buf, M_CXGBE); return (rc); } #undef MAX_READ_BUF_SIZE static int read_i2c(struct adapter *sc, struct t4_i2c_data *i2cd) { int rc; if (i2cd->len == 0 || i2cd->port_id >= sc->params.nports) return (EINVAL); if (i2cd->len > sizeof(i2cd->data)) return (EFBIG); rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4i2crd"); if (rc) return (rc); rc = -t4_i2c_rd(sc, sc->mbox, i2cd->port_id, i2cd->dev_addr, i2cd->offset, i2cd->len, &i2cd->data[0]); end_synchronized_op(sc, 0); return (rc); } static int in_range(int val, int lo, int hi) { return (val < 0 || (val <= hi && val >= lo)); } static int set_sched_class_config(struct adapter *sc, int minmax) { int rc; if (minmax < 0) return (EINVAL); rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4sscc"); if (rc) return (rc); rc = -t4_sched_config(sc, FW_SCHED_TYPE_PKTSCHED, minmax, 1); end_synchronized_op(sc, 0); return (rc); } static int set_sched_class_params(struct adapter *sc, struct t4_sched_class_params *p, int sleep_ok) { int rc, top_speed, fw_level, fw_mode, fw_rateunit, fw_ratemode; struct port_info *pi; struct tx_sched_class *tc; if (p->level == SCHED_CLASS_LEVEL_CL_RL) fw_level = FW_SCHED_PARAMS_LEVEL_CL_RL; else if (p->level == SCHED_CLASS_LEVEL_CL_WRR) fw_level = FW_SCHED_PARAMS_LEVEL_CL_WRR; else if (p->level == SCHED_CLASS_LEVEL_CH_RL) fw_level = FW_SCHED_PARAMS_LEVEL_CH_RL; else return (EINVAL); if (p->mode == SCHED_CLASS_MODE_CLASS) fw_mode = FW_SCHED_PARAMS_MODE_CLASS; else if (p->mode == SCHED_CLASS_MODE_FLOW) fw_mode = FW_SCHED_PARAMS_MODE_FLOW; else return (EINVAL); if (p->rateunit == SCHED_CLASS_RATEUNIT_BITS) fw_rateunit = FW_SCHED_PARAMS_UNIT_BITRATE; else if (p->rateunit == SCHED_CLASS_RATEUNIT_PKTS) fw_rateunit = FW_SCHED_PARAMS_UNIT_PKTRATE; else return (EINVAL); if (p->ratemode == SCHED_CLASS_RATEMODE_REL) fw_ratemode = FW_SCHED_PARAMS_RATE_REL; else if (p->ratemode == SCHED_CLASS_RATEMODE_ABS) fw_ratemode = FW_SCHED_PARAMS_RATE_ABS; else return (EINVAL); /* Vet our parameters ... */ if (!in_range(p->channel, 0, sc->chip_params->nchan - 1)) return (ERANGE); pi = sc->port[sc->chan_map[p->channel]]; if (pi == NULL) return (ENXIO); MPASS(pi->tx_chan == p->channel); top_speed = port_top_speed(pi) * 1000000; /* Gbps -> Kbps */ if (!in_range(p->cl, 0, sc->chip_params->nsched_cls) || !in_range(p->minrate, 0, top_speed) || !in_range(p->maxrate, 0, top_speed) || !in_range(p->weight, 0, 100)) return (ERANGE); /* * Translate any unset parameters into the firmware's * nomenclature and/or fail the call if the parameters * are required ... */ if (p->rateunit < 0 || p->ratemode < 0 || p->channel < 0 || p->cl < 0) return (EINVAL); if (p->minrate < 0) p->minrate = 0; if (p->maxrate < 0) { if (p->level == SCHED_CLASS_LEVEL_CL_RL || p->level == SCHED_CLASS_LEVEL_CH_RL) return (EINVAL); else p->maxrate = 0; } if (p->weight < 0) { if (p->level == SCHED_CLASS_LEVEL_CL_WRR) return (EINVAL); else p->weight = 0; } if (p->pktsize < 0) { if (p->level == SCHED_CLASS_LEVEL_CL_RL || p->level == SCHED_CLASS_LEVEL_CH_RL) return (EINVAL); else p->pktsize = 0; } rc = begin_synchronized_op(sc, NULL, sleep_ok ? (SLEEP_OK | INTR_OK) : HOLD_LOCK, "t4sscp"); if (rc) return (rc); tc = &pi->tc[p->cl]; tc->params = *p; rc = -t4_sched_params(sc, FW_SCHED_TYPE_PKTSCHED, fw_level, fw_mode, fw_rateunit, fw_ratemode, p->channel, p->cl, p->minrate, p->maxrate, p->weight, p->pktsize, sleep_ok); if (rc == 0) tc->flags |= TX_SC_OK; else { /* * Unknown state at this point, see tc->params for what was * attempted. */ tc->flags &= ~TX_SC_OK; } end_synchronized_op(sc, sleep_ok ? 0 : LOCK_HELD); return (rc); } int t4_set_sched_class(struct adapter *sc, struct t4_sched_params *p) { if (p->type != SCHED_CLASS_TYPE_PACKET) return (EINVAL); if (p->subcmd == SCHED_CLASS_SUBCMD_CONFIG) return (set_sched_class_config(sc, p->u.config.minmax)); if (p->subcmd == SCHED_CLASS_SUBCMD_PARAMS) return (set_sched_class_params(sc, &p->u.params, 1)); return (EINVAL); } int t4_set_sched_queue(struct adapter *sc, struct t4_sched_queue *p) { struct port_info *pi = NULL; struct vi_info *vi; struct sge_txq *txq; uint32_t fw_mnem, fw_queue, fw_class; int i, rc; rc = begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4setsq"); if (rc) return (rc); if (p->port >= sc->params.nports) { rc = EINVAL; goto done; } /* XXX: Only supported for the main VI. */ pi = sc->port[p->port]; vi = &pi->vi[0]; if (!(vi->flags & VI_INIT_DONE)) { /* tx queues not set up yet */ rc = EAGAIN; goto done; } if (!in_range(p->queue, 0, vi->ntxq - 1) || !in_range(p->cl, 0, sc->chip_params->nsched_cls - 1)) { rc = EINVAL; goto done; } /* * Create a template for the FW_PARAMS_CMD mnemonic and value (TX * Scheduling Class in this case). */ fw_mnem = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH)); fw_class = p->cl < 0 ? 0xffffffff : p->cl; /* * If op.queue is non-negative, then we're only changing the scheduling * on a single specified TX queue. */ if (p->queue >= 0) { txq = &sc->sge.txq[vi->first_txq + p->queue]; fw_queue = (fw_mnem | V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id)); rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class); goto done; } /* * Change the scheduling on all the TX queues for the * interface. */ for_each_txq(vi, i, txq) { fw_queue = (fw_mnem | V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id)); rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class); if (rc) goto done; } rc = 0; done: end_synchronized_op(sc, 0); return (rc); } int t4_os_find_pci_capability(struct adapter *sc, int cap) { int i; return (pci_find_cap(sc->dev, cap, &i) == 0 ? i : 0); } int t4_os_pci_save_state(struct adapter *sc) { device_t dev; struct pci_devinfo *dinfo; dev = sc->dev; dinfo = device_get_ivars(dev); pci_cfg_save(dev, dinfo, 0); return (0); } int t4_os_pci_restore_state(struct adapter *sc) { device_t dev; struct pci_devinfo *dinfo; dev = sc->dev; dinfo = device_get_ivars(dev); pci_cfg_restore(dev, dinfo); return (0); } void t4_os_portmod_changed(const struct adapter *sc, int idx) { struct port_info *pi = sc->port[idx]; struct vi_info *vi; struct ifnet *ifp; int v; static const char *mod_str[] = { NULL, "LR", "SR", "ER", "TWINAX", "active TWINAX", "LRM" }; for_each_vi(pi, v, vi) { build_medialist(pi, &vi->media); } ifp = pi->vi[0].ifp; if (pi->mod_type == FW_PORT_MOD_TYPE_NONE) if_printf(ifp, "transceiver unplugged.\n"); else if (pi->mod_type == FW_PORT_MOD_TYPE_UNKNOWN) if_printf(ifp, "unknown transceiver inserted.\n"); else if (pi->mod_type == FW_PORT_MOD_TYPE_NOTSUPPORTED) if_printf(ifp, "unsupported transceiver inserted.\n"); else if (pi->mod_type > 0 && pi->mod_type < nitems(mod_str)) { if_printf(ifp, "%s transceiver inserted.\n", mod_str[pi->mod_type]); } else { if_printf(ifp, "transceiver (type %d) inserted.\n", pi->mod_type); } } void t4_os_link_changed(struct adapter *sc, int idx, int link_stat, int reason) { struct port_info *pi = sc->port[idx]; struct vi_info *vi; struct ifnet *ifp; int v; if (link_stat) pi->linkdnrc = -1; else { if (reason >= 0) pi->linkdnrc = reason; } for_each_vi(pi, v, vi) { ifp = vi->ifp; if (ifp == NULL) continue; if (link_stat) { ifp->if_baudrate = IF_Mbps(pi->link_cfg.speed); if_link_state_change(ifp, LINK_STATE_UP); } else { if_link_state_change(ifp, LINK_STATE_DOWN); } } } void t4_iterate(void (*func)(struct adapter *, void *), void *arg) { struct adapter *sc; sx_slock(&t4_list_lock); SLIST_FOREACH(sc, &t4_list, link) { /* * func should not make any assumptions about what state sc is * in - the only guarantee is that sc->sc_lock is a valid lock. */ func(sc, arg); } sx_sunlock(&t4_list_lock); } static int t4_ioctl(struct cdev *dev, unsigned long cmd, caddr_t data, int fflag, struct thread *td) { int rc; struct adapter *sc = dev->si_drv1; rc = priv_check(td, PRIV_DRIVER); if (rc != 0) return (rc); switch (cmd) { case CHELSIO_T4_GETREG: { struct t4_reg *edata = (struct t4_reg *)data; if ((edata->addr & 0x3) != 0 || edata->addr >= sc->mmio_len) return (EFAULT); if (edata->size == 4) edata->val = t4_read_reg(sc, edata->addr); else if (edata->size == 8) edata->val = t4_read_reg64(sc, edata->addr); else return (EINVAL); break; } case CHELSIO_T4_SETREG: { struct t4_reg *edata = (struct t4_reg *)data; if ((edata->addr & 0x3) != 0 || edata->addr >= sc->mmio_len) return (EFAULT); if (edata->size == 4) { if (edata->val & 0xffffffff00000000) return (EINVAL); t4_write_reg(sc, edata->addr, (uint32_t) edata->val); } else if (edata->size == 8) t4_write_reg64(sc, edata->addr, edata->val); else return (EINVAL); break; } case CHELSIO_T4_REGDUMP: { struct t4_regdump *regs = (struct t4_regdump *)data; int reglen = t4_get_regs_len(sc); uint8_t *buf; if (regs->len < reglen) { regs->len = reglen; /* hint to the caller */ return (ENOBUFS); } regs->len = reglen; buf = malloc(reglen, M_CXGBE, M_WAITOK | M_ZERO); get_regs(sc, regs, buf); rc = copyout(buf, regs->data, reglen); free(buf, M_CXGBE); break; } case CHELSIO_T4_GET_FILTER_MODE: rc = get_filter_mode(sc, (uint32_t *)data); break; case CHELSIO_T4_SET_FILTER_MODE: rc = set_filter_mode(sc, *(uint32_t *)data); break; case CHELSIO_T4_GET_FILTER: rc = get_filter(sc, (struct t4_filter *)data); break; case CHELSIO_T4_SET_FILTER: rc = set_filter(sc, (struct t4_filter *)data); break; case CHELSIO_T4_DEL_FILTER: rc = del_filter(sc, (struct t4_filter *)data); break; case CHELSIO_T4_GET_SGE_CONTEXT: rc = get_sge_context(sc, (struct t4_sge_context *)data); break; case CHELSIO_T4_LOAD_FW: rc = load_fw(sc, (struct t4_data *)data); break; case CHELSIO_T4_GET_MEM: rc = read_card_mem(sc, 2, (struct t4_mem_range *)data); break; case CHELSIO_T4_GET_I2C: rc = read_i2c(sc, (struct t4_i2c_data *)data); break; case CHELSIO_T4_CLEAR_STATS: { int i, v; u_int port_id = *(uint32_t *)data; struct port_info *pi; struct vi_info *vi; if (port_id >= sc->params.nports) return (EINVAL); pi = sc->port[port_id]; + if (pi == NULL) + return (EIO); /* MAC stats */ t4_clr_port_stats(sc, pi->tx_chan); pi->tx_parse_error = 0; mtx_lock(&sc->reg_lock); for_each_vi(pi, v, vi) { if (vi->flags & VI_INIT_DONE) t4_clr_vi_stats(sc, vi->viid); } mtx_unlock(&sc->reg_lock); /* * Since this command accepts a port, clear stats for * all VIs on this port. */ for_each_vi(pi, v, vi) { if (vi->flags & VI_INIT_DONE) { struct sge_rxq *rxq; struct sge_txq *txq; struct sge_wrq *wrq; for_each_rxq(vi, i, rxq) { #if defined(INET) || defined(INET6) rxq->lro.lro_queued = 0; rxq->lro.lro_flushed = 0; #endif rxq->rxcsum = 0; rxq->vlan_extraction = 0; } for_each_txq(vi, i, txq) { txq->txcsum = 0; txq->tso_wrs = 0; txq->vlan_insertion = 0; txq->imm_wrs = 0; txq->sgl_wrs = 0; txq->txpkt_wrs = 0; txq->txpkts0_wrs = 0; txq->txpkts1_wrs = 0; txq->txpkts0_pkts = 0; txq->txpkts1_pkts = 0; mp_ring_reset_stats(txq->r); } #ifdef TCP_OFFLOAD /* nothing to clear for each ofld_rxq */ for_each_ofld_txq(vi, i, wrq) { wrq->tx_wrs_direct = 0; wrq->tx_wrs_copied = 0; } #endif if (IS_MAIN_VI(vi)) { wrq = &sc->sge.ctrlq[pi->port_id]; wrq->tx_wrs_direct = 0; wrq->tx_wrs_copied = 0; } } } break; } case CHELSIO_T4_SCHED_CLASS: rc = t4_set_sched_class(sc, (struct t4_sched_params *)data); break; case CHELSIO_T4_SCHED_QUEUE: rc = t4_set_sched_queue(sc, (struct t4_sched_queue *)data); break; case CHELSIO_T4_GET_TRACER: rc = t4_get_tracer(sc, (struct t4_tracer *)data); break; case CHELSIO_T4_SET_TRACER: rc = t4_set_tracer(sc, (struct t4_tracer *)data); break; default: rc = ENOTTY; } return (rc); } void t4_db_full(struct adapter *sc) { CXGBE_UNIMPLEMENTED(__func__); } void t4_db_dropped(struct adapter *sc) { CXGBE_UNIMPLEMENTED(__func__); } #ifdef TCP_OFFLOAD static int toe_capability(struct vi_info *vi, int enable) { int rc; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; ASSERT_SYNCHRONIZED_OP(sc); if (!is_offload(sc)) return (ENODEV); if (enable) { if ((vi->ifp->if_capenable & IFCAP_TOE) != 0) { /* TOE is already enabled. */ return (0); } /* * We need the port's queues around so that we're able to send * and receive CPLs to/from the TOE even if the ifnet for this * port has never been UP'd administratively. */ if (!(vi->flags & VI_INIT_DONE)) { rc = vi_full_init(vi); if (rc) return (rc); } if (!(pi->vi[0].flags & VI_INIT_DONE)) { rc = vi_full_init(&pi->vi[0]); if (rc) return (rc); } if (isset(&sc->offload_map, pi->port_id)) { /* TOE is enabled on another VI of this port. */ pi->uld_vis++; return (0); } if (!uld_active(sc, ULD_TOM)) { rc = t4_activate_uld(sc, ULD_TOM); if (rc == EAGAIN) { log(LOG_WARNING, "You must kldload t4_tom.ko before trying " "to enable TOE on a cxgbe interface.\n"); } if (rc != 0) return (rc); KASSERT(sc->tom_softc != NULL, ("%s: TOM activated but softc NULL", __func__)); KASSERT(uld_active(sc, ULD_TOM), ("%s: TOM activated but flag not set", __func__)); } /* Activate iWARP and iSCSI too, if the modules are loaded. */ if (!uld_active(sc, ULD_IWARP)) (void) t4_activate_uld(sc, ULD_IWARP); if (!uld_active(sc, ULD_ISCSI)) (void) t4_activate_uld(sc, ULD_ISCSI); pi->uld_vis++; setbit(&sc->offload_map, pi->port_id); } else { pi->uld_vis--; if (!isset(&sc->offload_map, pi->port_id) || pi->uld_vis > 0) return (0); KASSERT(uld_active(sc, ULD_TOM), ("%s: TOM never initialized?", __func__)); clrbit(&sc->offload_map, pi->port_id); } return (0); } /* * Add an upper layer driver to the global list. */ int t4_register_uld(struct uld_info *ui) { int rc = 0; struct uld_info *u; sx_xlock(&t4_uld_list_lock); SLIST_FOREACH(u, &t4_uld_list, link) { if (u->uld_id == ui->uld_id) { rc = EEXIST; goto done; } } SLIST_INSERT_HEAD(&t4_uld_list, ui, link); ui->refcount = 0; done: sx_xunlock(&t4_uld_list_lock); return (rc); } int t4_unregister_uld(struct uld_info *ui) { int rc = EINVAL; struct uld_info *u; sx_xlock(&t4_uld_list_lock); SLIST_FOREACH(u, &t4_uld_list, link) { if (u == ui) { if (ui->refcount > 0) { rc = EBUSY; goto done; } SLIST_REMOVE(&t4_uld_list, ui, uld_info, link); rc = 0; goto done; } } done: sx_xunlock(&t4_uld_list_lock); return (rc); } int t4_activate_uld(struct adapter *sc, int id) { int rc; struct uld_info *ui; ASSERT_SYNCHRONIZED_OP(sc); if (id < 0 || id > ULD_MAX) return (EINVAL); rc = EAGAIN; /* kldoad the module with this ULD and try again. */ sx_slock(&t4_uld_list_lock); SLIST_FOREACH(ui, &t4_uld_list, link) { if (ui->uld_id == id) { if (!(sc->flags & FULL_INIT_DONE)) { rc = adapter_full_init(sc); if (rc != 0) break; } rc = ui->activate(sc); if (rc == 0) { setbit(&sc->active_ulds, id); ui->refcount++; } break; } } sx_sunlock(&t4_uld_list_lock); return (rc); } int t4_deactivate_uld(struct adapter *sc, int id) { int rc; struct uld_info *ui; ASSERT_SYNCHRONIZED_OP(sc); if (id < 0 || id > ULD_MAX) return (EINVAL); rc = ENXIO; sx_slock(&t4_uld_list_lock); SLIST_FOREACH(ui, &t4_uld_list, link) { if (ui->uld_id == id) { rc = ui->deactivate(sc); if (rc == 0) { clrbit(&sc->active_ulds, id); ui->refcount--; } break; } } sx_sunlock(&t4_uld_list_lock); return (rc); } int uld_active(struct adapter *sc, int uld_id) { MPASS(uld_id >= 0 && uld_id <= ULD_MAX); return (isset(&sc->active_ulds, uld_id)); } #endif /* * Come up with reasonable defaults for some of the tunables, provided they're * not set by the user (in which case we'll use the values as is). */ static void tweak_tunables(void) { int nc = mp_ncpus; /* our snapshot of the number of CPUs */ if (t4_ntxq10g < 1) { #ifdef RSS t4_ntxq10g = rss_getnumbuckets(); #else t4_ntxq10g = min(nc, NTXQ_10G); #endif } if (t4_ntxq1g < 1) { #ifdef RSS /* XXX: way too many for 1GbE? */ t4_ntxq1g = rss_getnumbuckets(); #else t4_ntxq1g = min(nc, NTXQ_1G); #endif } if (t4_ntxq_vi < 1) t4_ntxq_vi = min(nc, NTXQ_VI); if (t4_nrxq10g < 1) { #ifdef RSS t4_nrxq10g = rss_getnumbuckets(); #else t4_nrxq10g = min(nc, NRXQ_10G); #endif } if (t4_nrxq1g < 1) { #ifdef RSS /* XXX: way too many for 1GbE? */ t4_nrxq1g = rss_getnumbuckets(); #else t4_nrxq1g = min(nc, NRXQ_1G); #endif } if (t4_nrxq_vi < 1) t4_nrxq_vi = min(nc, NRXQ_VI); #ifdef TCP_OFFLOAD if (t4_nofldtxq10g < 1) t4_nofldtxq10g = min(nc, NOFLDTXQ_10G); if (t4_nofldtxq1g < 1) t4_nofldtxq1g = min(nc, NOFLDTXQ_1G); if (t4_nofldtxq_vi < 1) t4_nofldtxq_vi = min(nc, NOFLDTXQ_VI); if (t4_nofldrxq10g < 1) t4_nofldrxq10g = min(nc, NOFLDRXQ_10G); if (t4_nofldrxq1g < 1) t4_nofldrxq1g = min(nc, NOFLDRXQ_1G); if (t4_nofldrxq_vi < 1) t4_nofldrxq_vi = min(nc, NOFLDRXQ_VI); if (t4_toecaps_allowed == -1) t4_toecaps_allowed = FW_CAPS_CONFIG_TOE; if (t4_rdmacaps_allowed == -1) { t4_rdmacaps_allowed = FW_CAPS_CONFIG_RDMA_RDDP | FW_CAPS_CONFIG_RDMA_RDMAC; } if (t4_iscsicaps_allowed == -1) { t4_iscsicaps_allowed = FW_CAPS_CONFIG_ISCSI_INITIATOR_PDU | FW_CAPS_CONFIG_ISCSI_TARGET_PDU | FW_CAPS_CONFIG_ISCSI_T10DIF; } #else if (t4_toecaps_allowed == -1) t4_toecaps_allowed = 0; if (t4_rdmacaps_allowed == -1) t4_rdmacaps_allowed = 0; if (t4_iscsicaps_allowed == -1) t4_iscsicaps_allowed = 0; #endif #ifdef DEV_NETMAP if (t4_nnmtxq_vi < 1) t4_nnmtxq_vi = min(nc, NNMTXQ_VI); if (t4_nnmrxq_vi < 1) t4_nnmrxq_vi = min(nc, NNMRXQ_VI); #endif if (t4_tmr_idx_10g < 0 || t4_tmr_idx_10g >= SGE_NTIMERS) t4_tmr_idx_10g = TMR_IDX_10G; if (t4_pktc_idx_10g < -1 || t4_pktc_idx_10g >= SGE_NCOUNTERS) t4_pktc_idx_10g = PKTC_IDX_10G; if (t4_tmr_idx_1g < 0 || t4_tmr_idx_1g >= SGE_NTIMERS) t4_tmr_idx_1g = TMR_IDX_1G; if (t4_pktc_idx_1g < -1 || t4_pktc_idx_1g >= SGE_NCOUNTERS) t4_pktc_idx_1g = PKTC_IDX_1G; if (t4_qsize_txq < 128) t4_qsize_txq = 128; if (t4_qsize_rxq < 128) t4_qsize_rxq = 128; while (t4_qsize_rxq & 7) t4_qsize_rxq++; t4_intr_types &= INTR_MSIX | INTR_MSI | INTR_INTX; } #ifdef DDB static void t4_dump_tcb(struct adapter *sc, int tid) { uint32_t base, i, j, off, pf, reg, save, tcb_addr, win_pos; reg = PCIE_MEM_ACCESS_REG(A_PCIE_MEM_ACCESS_OFFSET, 2); save = t4_read_reg(sc, reg); base = sc->memwin[2].mw_base; /* Dump TCB for the tid */ tcb_addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE); tcb_addr += tid * TCB_SIZE; if (is_t4(sc)) { pf = 0; win_pos = tcb_addr & ~0xf; /* start must be 16B aligned */ } else { pf = V_PFNUM(sc->pf); win_pos = tcb_addr & ~0x7f; /* start must be 128B aligned */ } t4_write_reg(sc, reg, win_pos | pf); t4_read_reg(sc, reg); off = tcb_addr - win_pos; for (i = 0; i < 4; i++) { uint32_t buf[8]; for (j = 0; j < 8; j++, off += 4) buf[j] = htonl(t4_read_reg(sc, base + off)); db_printf("%08x %08x %08x %08x %08x %08x %08x %08x\n", buf[0], buf[1], buf[2], buf[3], buf[4], buf[5], buf[6], buf[7]); } t4_write_reg(sc, reg, save); t4_read_reg(sc, reg); } static void t4_dump_devlog(struct adapter *sc) { struct devlog_params *dparams = &sc->params.devlog; struct fw_devlog_e e; int i, first, j, m, nentries, rc; uint64_t ftstamp = UINT64_MAX; if (dparams->start == 0) { db_printf("devlog params not valid\n"); return; } nentries = dparams->size / sizeof(struct fw_devlog_e); m = fwmtype_to_hwmtype(dparams->memtype); /* Find the first entry. */ first = -1; for (i = 0; i < nentries && !db_pager_quit; i++) { rc = -t4_mem_read(sc, m, dparams->start + i * sizeof(e), sizeof(e), (void *)&e); if (rc != 0) break; if (e.timestamp == 0) break; e.timestamp = be64toh(e.timestamp); if (e.timestamp < ftstamp) { ftstamp = e.timestamp; first = i; } } if (first == -1) return; i = first; do { rc = -t4_mem_read(sc, m, dparams->start + i * sizeof(e), sizeof(e), (void *)&e); if (rc != 0) return; if (e.timestamp == 0) return; e.timestamp = be64toh(e.timestamp); e.seqno = be32toh(e.seqno); for (j = 0; j < 8; j++) e.params[j] = be32toh(e.params[j]); db_printf("%10d %15ju %8s %8s ", e.seqno, e.timestamp, (e.level < nitems(devlog_level_strings) ? devlog_level_strings[e.level] : "UNKNOWN"), (e.facility < nitems(devlog_facility_strings) ? devlog_facility_strings[e.facility] : "UNKNOWN")); db_printf(e.fmt, e.params[0], e.params[1], e.params[2], e.params[3], e.params[4], e.params[5], e.params[6], e.params[7]); if (++i == nentries) i = 0; } while (i != first && !db_pager_quit); } static struct command_table db_t4_table = LIST_HEAD_INITIALIZER(db_t4_table); _DB_SET(_show, t4, NULL, db_show_table, 0, &db_t4_table); DB_FUNC(devlog, db_show_devlog, db_t4_table, CS_OWN, NULL) { device_t dev; int t; bool valid; valid = false; t = db_read_token(); if (t == tIDENT) { dev = device_lookup_by_name(db_tok_string); valid = true; } db_skip_to_eol(); if (!valid) { db_printf("usage: show t4 devlog \n"); return; } if (dev == NULL) { db_printf("device not found\n"); return; } t4_dump_devlog(device_get_softc(dev)); } DB_FUNC(tcb, db_show_t4tcb, db_t4_table, CS_OWN, NULL) { device_t dev; int radix, tid, t; bool valid; valid = false; radix = db_radix; db_radix = 10; t = db_read_token(); if (t == tIDENT) { dev = device_lookup_by_name(db_tok_string); t = db_read_token(); if (t == tNUMBER) { tid = db_tok_number; valid = true; } } db_radix = radix; db_skip_to_eol(); if (!valid) { db_printf("usage: show t4 tcb \n"); return; } if (dev == NULL) { db_printf("device not found\n"); return; } if (tid < 0) { db_printf("invalid tid\n"); return; } t4_dump_tcb(device_get_softc(dev), tid); } #endif static struct sx mlu; /* mod load unload */ SX_SYSINIT(cxgbe_mlu, &mlu, "cxgbe mod load/unload"); static int mod_event(module_t mod, int cmd, void *arg) { int rc = 0; static int loaded = 0; switch (cmd) { case MOD_LOAD: sx_xlock(&mlu); if (loaded++ == 0) { t4_sge_modload(); t4_register_cpl_handler(CPL_SET_TCB_RPL, set_tcb_rpl); t4_register_cpl_handler(CPL_L2T_WRITE_RPL, l2t_write_rpl); t4_register_cpl_handler(CPL_TRACE_PKT, t4_trace_pkt); t4_register_cpl_handler(CPL_T5_TRACE_PKT, t5_trace_pkt); sx_init(&t4_list_lock, "T4/T5 adapters"); SLIST_INIT(&t4_list); #ifdef TCP_OFFLOAD sx_init(&t4_uld_list_lock, "T4/T5 ULDs"); SLIST_INIT(&t4_uld_list); #endif t4_tracer_modload(); tweak_tunables(); } sx_xunlock(&mlu); break; case MOD_UNLOAD: sx_xlock(&mlu); if (--loaded == 0) { int tries; sx_slock(&t4_list_lock); if (!SLIST_EMPTY(&t4_list)) { rc = EBUSY; sx_sunlock(&t4_list_lock); goto done_unload; } #ifdef TCP_OFFLOAD sx_slock(&t4_uld_list_lock); if (!SLIST_EMPTY(&t4_uld_list)) { rc = EBUSY; sx_sunlock(&t4_uld_list_lock); sx_sunlock(&t4_list_lock); goto done_unload; } #endif tries = 0; while (tries++ < 5 && t4_sge_extfree_refs() != 0) { uprintf("%ju clusters with custom free routine " "still is use.\n", t4_sge_extfree_refs()); pause("t4unload", 2 * hz); } #ifdef TCP_OFFLOAD sx_sunlock(&t4_uld_list_lock); #endif sx_sunlock(&t4_list_lock); if (t4_sge_extfree_refs() == 0) { t4_tracer_modunload(); #ifdef TCP_OFFLOAD sx_destroy(&t4_uld_list_lock); #endif sx_destroy(&t4_list_lock); t4_sge_modunload(); loaded = 0; } else { rc = EBUSY; loaded++; /* undo earlier decrement */ } } done_unload: sx_xunlock(&mlu); break; } return (rc); } static devclass_t t4_devclass, t5_devclass; static devclass_t cxgbe_devclass, cxl_devclass; static devclass_t vcxgbe_devclass, vcxl_devclass; DRIVER_MODULE(t4nex, pci, t4_driver, t4_devclass, mod_event, 0); MODULE_VERSION(t4nex, 1); MODULE_DEPEND(t4nex, firmware, 1, 1, 1); #ifdef DEV_NETMAP MODULE_DEPEND(t4nex, netmap, 1, 1, 1); #endif /* DEV_NETMAP */ DRIVER_MODULE(t5nex, pci, t5_driver, t5_devclass, mod_event, 0); MODULE_VERSION(t5nex, 1); MODULE_DEPEND(t5nex, firmware, 1, 1, 1); #ifdef DEV_NETMAP MODULE_DEPEND(t5nex, netmap, 1, 1, 1); #endif /* DEV_NETMAP */ DRIVER_MODULE(cxgbe, t4nex, cxgbe_driver, cxgbe_devclass, 0, 0); MODULE_VERSION(cxgbe, 1); DRIVER_MODULE(cxl, t5nex, cxl_driver, cxl_devclass, 0, 0); MODULE_VERSION(cxl, 1); DRIVER_MODULE(vcxgbe, cxgbe, vcxgbe_driver, vcxgbe_devclass, 0, 0); MODULE_VERSION(vcxgbe, 1); DRIVER_MODULE(vcxl, cxl, vcxl_driver, vcxl_devclass, 0, 0); MODULE_VERSION(vcxl, 1); Index: projects/mips64-clang/sys/dev/cxgbe/t4_netmap.c =================================================================== --- projects/mips64-clang/sys/dev/cxgbe/t4_netmap.c (revision 305711) +++ projects/mips64-clang/sys/dev/cxgbe/t4_netmap.c (revision 305712) @@ -1,1011 +1,1013 @@ /*- * Copyright (c) 2014 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #ifdef DEV_NETMAP #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "common/common.h" #include "common/t4_regs.h" #include "common/t4_regs_values.h" extern int fl_pad; /* XXXNM */ SYSCTL_NODE(_hw, OID_AUTO, cxgbe, CTLFLAG_RD, 0, "cxgbe netmap parameters"); /* * 0 = normal netmap rx * 1 = black hole * 2 = supermassive black hole (buffer packing enabled) */ int black_hole = 0; SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_black_hole, CTLFLAG_RDTUN, &black_hole, 0, "Sink incoming packets."); int rx_ndesc = 256; SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_rx_ndesc, CTLFLAG_RWTUN, &rx_ndesc, 0, "# of rx descriptors after which the hw cidx is updated."); int holdoff_tmr_idx = 2; SYSCTL_INT(_hw_cxgbe, OID_AUTO, nm_holdoff_tmr_idx, CTLFLAG_RWTUN, &holdoff_tmr_idx, 0, "Holdoff timer index for netmap rx queues."); /* * Congestion drops. * -1: no congestion feedback (not recommended). * 0: backpressure the channel instead of dropping packets right away. * 1: no backpressure, drop packets for the congested queue immediately. */ static int nm_cong_drop = 1; TUNABLE_INT("hw.cxgbe.nm_cong_drop", &nm_cong_drop); static int alloc_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int cong) { int rc, cntxt_id, i; __be32 v; struct adapter *sc = vi->pi->adapter; struct sge_params *sp = &sc->params.sge; struct netmap_adapter *na = NA(vi->ifp); struct fw_iq_cmd c; MPASS(na != NULL); MPASS(nm_rxq->iq_desc != NULL); MPASS(nm_rxq->fl_desc != NULL); bzero(nm_rxq->iq_desc, vi->qsize_rxq * IQ_ESIZE); bzero(nm_rxq->fl_desc, na->num_rx_desc * EQ_ESIZE + sp->spg_len); bzero(&c, sizeof(c)); c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) | V_FW_IQ_CMD_VFN(0)); c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART | FW_LEN16(c)); if (vi->flags & INTR_RXQ) { KASSERT(nm_rxq->intr_idx < sc->intr_count, ("%s: invalid direct intr_idx %d", __func__, nm_rxq->intr_idx)); v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx); } else { CXGBE_UNIMPLEMENTED(__func__); /* XXXNM: needs review */ v = V_FW_IQ_CMD_IQANDSTINDEX(nm_rxq->intr_idx) | F_FW_IQ_CMD_IQANDST; } c.type_to_iqandstindex = htobe32(v | V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) | V_FW_IQ_CMD_VIID(vi->viid) | V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT)); c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(vi->pi->tx_chan) | F_FW_IQ_CMD_IQGTSMODE | V_FW_IQ_CMD_IQINTCNTTHRESH(0) | V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4)); c.iqsize = htobe16(vi->qsize_rxq); c.iqaddr = htobe64(nm_rxq->iq_ba); if (cong >= 0) { c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN | V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF | F_FW_IQ_CMD_FL0CONGEN); } c.iqns_to_fl0congen |= htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) | F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO | (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) | (black_hole == 2 ? F_FW_IQ_CMD_FL0PACKEN : 0)); c.fl0dcaen_to_fl0cidxfthresh = - htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) | - V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B)); + htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ? + X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) | + V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ? + X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B)); c.fl0size = htobe16(na->num_rx_desc / 8 + sp->spg_len / EQ_ESIZE); c.fl0addr = htobe64(nm_rxq->fl_ba); rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); if (rc != 0) { device_printf(sc->dev, "failed to create netmap ingress queue: %d\n", rc); return (rc); } nm_rxq->iq_cidx = 0; MPASS(nm_rxq->iq_sidx == vi->qsize_rxq - sp->spg_len / IQ_ESIZE); nm_rxq->iq_gen = F_RSPD_GEN; nm_rxq->iq_cntxt_id = be16toh(c.iqid); nm_rxq->iq_abs_id = be16toh(c.physiqid); cntxt_id = nm_rxq->iq_cntxt_id - sc->sge.iq_start; if (cntxt_id >= sc->sge.niq) { panic ("%s: nm_rxq->iq_cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.niq - 1); } sc->sge.iqmap[cntxt_id] = (void *)nm_rxq; nm_rxq->fl_cntxt_id = be16toh(c.fl0id); nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0; MPASS(nm_rxq->fl_sidx == na->num_rx_desc); cntxt_id = nm_rxq->fl_cntxt_id - sc->sge.eq_start; if (cntxt_id >= sc->sge.neq) { panic("%s: nm_rxq->fl_cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.neq - 1); } sc->sge.eqmap[cntxt_id] = (void *)nm_rxq; nm_rxq->fl_db_val = V_QID(nm_rxq->fl_cntxt_id) | sc->chip_params->sge_fl_db; if (is_t5(sc) && cong >= 0) { uint32_t param, val; param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) | V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id); param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) | V_FW_PARAMS_PARAM_YZ(nm_rxq->iq_cntxt_id); if (cong == 0) val = 1 << 19; else { val = 2 << 19; for (i = 0; i < 4; i++) { if (cong & (1 << i)) val |= 1 << (i << 2); } } rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); if (rc != 0) { /* report error but carry on */ device_printf(sc->dev, "failed to set congestion manager context for " "ingress queue %d: %d\n", nm_rxq->iq_cntxt_id, rc); } } t4_write_reg(sc, sc->sge_gts_reg, V_INGRESSQID(nm_rxq->iq_cntxt_id) | V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx))); return (rc); } static int free_nm_rxq_hwq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq) { struct adapter *sc = vi->pi->adapter; int rc; rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, 0xffff); if (rc != 0) device_printf(sc->dev, "%s: failed for iq %d, fl %d: %d\n", __func__, nm_rxq->iq_cntxt_id, nm_rxq->fl_cntxt_id, rc); return (rc); } static int alloc_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq) { int rc, cntxt_id; size_t len; struct adapter *sc = vi->pi->adapter; struct netmap_adapter *na = NA(vi->ifp); struct fw_eq_eth_cmd c; MPASS(na != NULL); MPASS(nm_txq->desc != NULL); len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len; bzero(nm_txq->desc, len); bzero(&c, sizeof(c)); c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) | V_FW_EQ_ETH_CMD_VFN(0)); c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC | F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c)); c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE | F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid)); c.fetchszm_to_iqid = htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) | V_FW_EQ_ETH_CMD_PCIECHN(vi->pi->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO | V_FW_EQ_ETH_CMD_IQID(sc->sge.nm_rxq[nm_txq->iqidx].iq_cntxt_id)); c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) | V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) | V_FW_EQ_ETH_CMD_EQSIZE(len / EQ_ESIZE)); c.eqaddr = htobe64(nm_txq->ba); rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); if (rc != 0) { device_printf(vi->dev, "failed to create netmap egress queue: %d\n", rc); return (rc); } nm_txq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd)); cntxt_id = nm_txq->cntxt_id - sc->sge.eq_start; if (cntxt_id >= sc->sge.neq) panic("%s: nm_txq->cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.neq - 1); sc->sge.eqmap[cntxt_id] = (void *)nm_txq; nm_txq->pidx = nm_txq->cidx = 0; MPASS(nm_txq->sidx == na->num_tx_desc); nm_txq->equiqidx = nm_txq->equeqidx = nm_txq->dbidx = 0; nm_txq->doorbells = sc->doorbells; if (isset(&nm_txq->doorbells, DOORBELL_UDB) || isset(&nm_txq->doorbells, DOORBELL_UDBWC) || isset(&nm_txq->doorbells, DOORBELL_WCWR)) { uint32_t s_qpp = sc->params.sge.eq_s_qpp; uint32_t mask = (1 << s_qpp) - 1; volatile uint8_t *udb; udb = sc->udbs_base + UDBS_DB_OFFSET; udb += (nm_txq->cntxt_id >> s_qpp) << PAGE_SHIFT; nm_txq->udb_qid = nm_txq->cntxt_id & mask; if (nm_txq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE) clrbit(&nm_txq->doorbells, DOORBELL_WCWR); else { udb += nm_txq->udb_qid << UDBS_SEG_SHIFT; nm_txq->udb_qid = 0; } nm_txq->udb = (volatile void *)udb; } return (rc); } static int free_nm_txq_hwq(struct vi_info *vi, struct sge_nm_txq *nm_txq) { struct adapter *sc = vi->pi->adapter; int rc; rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, nm_txq->cntxt_id); if (rc != 0) device_printf(sc->dev, "%s: failed for eq %d: %d\n", __func__, nm_txq->cntxt_id, rc); return (rc); } static int cxgbe_netmap_on(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp, struct netmap_adapter *na) { struct netmap_slot *slot; struct sge_nm_rxq *nm_rxq; struct sge_nm_txq *nm_txq; int rc, i, j, hwidx; struct hw_buf_info *hwb; ASSERT_SYNCHRONIZED_OP(sc); if ((vi->flags & VI_INIT_DONE) == 0 || (ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) return (EAGAIN); hwb = &sc->sge.hw_buf_info[0]; for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) { if (hwb->size == NETMAP_BUF_SIZE(na)) break; } if (i >= SGE_FLBUF_SIZES) { if_printf(ifp, "no hwidx for netmap buffer size %d.\n", NETMAP_BUF_SIZE(na)); return (ENXIO); } hwidx = i; /* Must set caps before calling netmap_reset */ nm_set_native_flags(na); for_each_nm_rxq(vi, i, nm_rxq) { struct irq *irq = &sc->irq[vi->first_intr + i]; alloc_nm_rxq_hwq(vi, nm_rxq, tnl_cong(vi->pi, nm_cong_drop)); nm_rxq->fl_hwidx = hwidx; slot = netmap_reset(na, NR_RX, i, 0); MPASS(slot != NULL); /* XXXNM: error check, not assert */ /* We deal with 8 bufs at a time */ MPASS((na->num_rx_desc & 7) == 0); MPASS(na->num_rx_desc == nm_rxq->fl_sidx); for (j = 0; j < nm_rxq->fl_sidx; j++) { uint64_t ba; PNMB(na, &slot[j], &ba); MPASS(ba != 0); nm_rxq->fl_desc[j] = htobe64(ba | hwidx); } j = nm_rxq->fl_pidx = nm_rxq->fl_sidx - 8; MPASS((j & 7) == 0); j /= 8; /* driver pidx to hardware pidx */ wmb(); t4_write_reg(sc, sc->sge_kdoorbell_reg, nm_rxq->fl_db_val | V_PIDX(j)); atomic_cmpset_int(&irq->nm_state, NM_OFF, NM_ON); } for_each_nm_txq(vi, i, nm_txq) { alloc_nm_txq_hwq(vi, nm_txq); slot = netmap_reset(na, NR_TX, i, 0); MPASS(slot != NULL); /* XXXNM: error check, not assert */ } if (vi->nm_rss == NULL) { vi->nm_rss = malloc(vi->rss_size * sizeof(uint16_t), M_CXGBE, M_ZERO | M_WAITOK); } for (i = 0; i < vi->rss_size;) { for_each_nm_rxq(vi, j, nm_rxq) { vi->nm_rss[i++] = nm_rxq->iq_abs_id; if (i == vi->rss_size) break; } } rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, vi->nm_rss, vi->rss_size); if (rc != 0) if_printf(ifp, "netmap rss_config failed: %d\n", rc); return (rc); } static int cxgbe_netmap_off(struct adapter *sc, struct vi_info *vi, struct ifnet *ifp, struct netmap_adapter *na) { int rc, i; struct sge_nm_txq *nm_txq; struct sge_nm_rxq *nm_rxq; ASSERT_SYNCHRONIZED_OP(sc); if ((vi->flags & VI_INIT_DONE) == 0) return (0); rc = -t4_config_rss_range(sc, sc->mbox, vi->viid, 0, vi->rss_size, vi->rss, vi->rss_size); if (rc != 0) if_printf(ifp, "failed to restore RSS config: %d\n", rc); nm_clear_native_flags(na); for_each_nm_txq(vi, i, nm_txq) { struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx]; /* Wait for hw pidx to catch up ... */ while (be16toh(nm_txq->pidx) != spg->pidx) pause("nmpidx", 1); /* ... and then for the cidx. */ while (spg->pidx != spg->cidx) pause("nmcidx", 1); free_nm_txq_hwq(vi, nm_txq); } for_each_nm_rxq(vi, i, nm_rxq) { struct irq *irq = &sc->irq[vi->first_intr + i]; while (!atomic_cmpset_int(&irq->nm_state, NM_ON, NM_OFF)) pause("nmst", 1); free_nm_rxq_hwq(vi, nm_rxq); } return (rc); } static int cxgbe_netmap_reg(struct netmap_adapter *na, int on) { struct ifnet *ifp = na->ifp; struct vi_info *vi = ifp->if_softc; struct adapter *sc = vi->pi->adapter; int rc; rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4nmreg"); if (rc != 0) return (rc); if (on) rc = cxgbe_netmap_on(sc, vi, ifp, na); else rc = cxgbe_netmap_off(sc, vi, ifp, na); end_synchronized_op(sc, 0); return (rc); } /* How many packets can a single type1 WR carry in n descriptors */ static inline int ndesc_to_npkt(const int n) { MPASS(n > 0 && n <= SGE_MAX_WR_NDESC); return (n * 2 - 1); } #define MAX_NPKT_IN_TYPE1_WR (ndesc_to_npkt(SGE_MAX_WR_NDESC)) /* Space (in descriptors) needed for a type1 WR that carries n packets */ static inline int npkt_to_ndesc(const int n) { MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR); return ((n + 2) / 2); } /* Space (in 16B units) needed for a type1 WR that carries n packets */ static inline int npkt_to_len16(const int n) { MPASS(n > 0 && n <= MAX_NPKT_IN_TYPE1_WR); return (n * 2 + 1); } #define NMIDXDIFF(q, idx) IDXDIFF((q)->pidx, (q)->idx, (q)->sidx) static void ring_nm_txq_db(struct adapter *sc, struct sge_nm_txq *nm_txq) { int n; u_int db = nm_txq->doorbells; MPASS(nm_txq->pidx != nm_txq->dbidx); n = NMIDXDIFF(nm_txq, dbidx); if (n > 1) clrbit(&db, DOORBELL_WCWR); wmb(); switch (ffs(db) - 1) { case DOORBELL_UDB: *nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n)); break; case DOORBELL_WCWR: { volatile uint64_t *dst, *src; /* * Queues whose 128B doorbell segment fits in the page do not * use relative qid (udb_qid is always 0). Only queues with * doorbell segments can do WCWR. */ KASSERT(nm_txq->udb_qid == 0 && n == 1, ("%s: inappropriate doorbell (0x%x, %d, %d) for nm_txq %p", __func__, nm_txq->doorbells, n, nm_txq->pidx, nm_txq)); dst = (volatile void *)((uintptr_t)nm_txq->udb + UDBS_WR_OFFSET - UDBS_DB_OFFSET); src = (void *)&nm_txq->desc[nm_txq->dbidx]; while (src != (void *)&nm_txq->desc[nm_txq->dbidx + 1]) *dst++ = *src++; wmb(); break; } case DOORBELL_UDBWC: *nm_txq->udb = htole32(V_QID(nm_txq->udb_qid) | V_PIDX(n)); wmb(); break; case DOORBELL_KDB: t4_write_reg(sc, sc->sge_kdoorbell_reg, V_QID(nm_txq->cntxt_id) | V_PIDX(n)); break; } nm_txq->dbidx = nm_txq->pidx; } int lazy_tx_credit_flush = 1; /* * Write work requests to send 'npkt' frames and ring the doorbell to send them * on their way. No need to check for wraparound. */ static void cxgbe_nm_tx(struct adapter *sc, struct sge_nm_txq *nm_txq, struct netmap_kring *kring, int npkt, int npkt_remaining, int txcsum) { struct netmap_ring *ring = kring->ring; struct netmap_slot *slot; const u_int lim = kring->nkr_num_slots - 1; struct fw_eth_tx_pkts_wr *wr = (void *)&nm_txq->desc[nm_txq->pidx]; uint16_t len; uint64_t ba; struct cpl_tx_pkt_core *cpl; struct ulptx_sgl *usgl; int i, n; while (npkt) { n = min(npkt, MAX_NPKT_IN_TYPE1_WR); len = 0; wr = (void *)&nm_txq->desc[nm_txq->pidx]; wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)); wr->equiq_to_len16 = htobe32(V_FW_WR_LEN16(npkt_to_len16(n))); wr->npkt = n; wr->r3 = 0; wr->type = 1; cpl = (void *)(wr + 1); for (i = 0; i < n; i++) { slot = &ring->slot[kring->nr_hwcur]; PNMB(kring->na, slot, &ba); MPASS(ba != 0); cpl->ctrl0 = nm_txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(slot->len); /* * netmap(4) says "netmap does not use features such as * checksum offloading, TCP segmentation offloading, * encryption, VLAN encapsulation/decapsulation, etc." * * So the ncxl interfaces have tx hardware checksumming * disabled by default. But you can override netmap by * enabling IFCAP_TXCSUM on the interface manully. */ cpl->ctrl1 = txcsum ? 0 : htobe64(F_TXPKT_IPCSUM_DIS | F_TXPKT_L4CSUM_DIS); usgl = (void *)(cpl + 1); usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | V_ULPTX_NSGE(1)); usgl->len0 = htobe32(slot->len); usgl->addr0 = htobe64(ba); slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED); cpl = (void *)(usgl + 1); MPASS(slot->len + len <= UINT16_MAX); len += slot->len; kring->nr_hwcur = nm_next(kring->nr_hwcur, lim); } wr->plen = htobe16(len); npkt -= n; nm_txq->pidx += npkt_to_ndesc(n); MPASS(nm_txq->pidx <= nm_txq->sidx); if (__predict_false(nm_txq->pidx == nm_txq->sidx)) { /* * This routine doesn't know how to write WRs that wrap * around. Make sure it wasn't asked to. */ MPASS(npkt == 0); nm_txq->pidx = 0; } if (npkt == 0 && npkt_remaining == 0) { /* All done. */ if (lazy_tx_credit_flush == 0) { wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ); nm_txq->equeqidx = nm_txq->pidx; nm_txq->equiqidx = nm_txq->pidx; } ring_nm_txq_db(sc, nm_txq); return; } if (NMIDXDIFF(nm_txq, equiqidx) >= nm_txq->sidx / 2) { wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ | F_FW_WR_EQUIQ); nm_txq->equeqidx = nm_txq->pidx; nm_txq->equiqidx = nm_txq->pidx; } else if (NMIDXDIFF(nm_txq, equeqidx) >= 64) { wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ); nm_txq->equeqidx = nm_txq->pidx; } if (NMIDXDIFF(nm_txq, dbidx) >= 2 * SGE_MAX_WR_NDESC) ring_nm_txq_db(sc, nm_txq); } /* Will get called again. */ MPASS(npkt_remaining); } /* How many contiguous free descriptors starting at pidx */ static inline int contiguous_ndesc_available(struct sge_nm_txq *nm_txq) { if (nm_txq->cidx > nm_txq->pidx) return (nm_txq->cidx - nm_txq->pidx - 1); else if (nm_txq->cidx > 0) return (nm_txq->sidx - nm_txq->pidx); else return (nm_txq->sidx - nm_txq->pidx - 1); } static int reclaim_nm_tx_desc(struct sge_nm_txq *nm_txq) { struct sge_qstat *spg = (void *)&nm_txq->desc[nm_txq->sidx]; uint16_t hw_cidx = spg->cidx; /* snapshot */ struct fw_eth_tx_pkts_wr *wr; int n = 0; hw_cidx = be16toh(hw_cidx); while (nm_txq->cidx != hw_cidx) { wr = (void *)&nm_txq->desc[nm_txq->cidx]; MPASS(wr->op_pkd == htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR))); MPASS(wr->type == 1); MPASS(wr->npkt > 0 && wr->npkt <= MAX_NPKT_IN_TYPE1_WR); n += wr->npkt; nm_txq->cidx += npkt_to_ndesc(wr->npkt); /* * We never sent a WR that wrapped around so the credits coming * back, WR by WR, should never cause the cidx to wrap around * either. */ MPASS(nm_txq->cidx <= nm_txq->sidx); if (__predict_false(nm_txq->cidx == nm_txq->sidx)) nm_txq->cidx = 0; } return (n); } static int cxgbe_netmap_txsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct ifnet *ifp = na->ifp; struct vi_info *vi = ifp->if_softc; struct adapter *sc = vi->pi->adapter; struct sge_nm_txq *nm_txq = &sc->sge.nm_txq[vi->first_nm_txq + kring->ring_id]; const u_int head = kring->rhead; u_int reclaimed = 0; int n, d, npkt_remaining, ndesc_remaining, txcsum; /* * Tx was at kring->nr_hwcur last time around and now we need to advance * to kring->rhead. Note that the driver's pidx moves independent of * netmap's kring->nr_hwcur (pidx counts descriptors and the relation * between descriptors and frames isn't 1:1). */ npkt_remaining = head >= kring->nr_hwcur ? head - kring->nr_hwcur : kring->nkr_num_slots - kring->nr_hwcur + head; txcsum = ifp->if_capenable & (IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6); while (npkt_remaining) { reclaimed += reclaim_nm_tx_desc(nm_txq); ndesc_remaining = contiguous_ndesc_available(nm_txq); /* Can't run out of descriptors with packets still remaining */ MPASS(ndesc_remaining > 0); /* # of desc needed to tx all remaining packets */ d = (npkt_remaining / MAX_NPKT_IN_TYPE1_WR) * SGE_MAX_WR_NDESC; if (npkt_remaining % MAX_NPKT_IN_TYPE1_WR) d += npkt_to_ndesc(npkt_remaining % MAX_NPKT_IN_TYPE1_WR); if (d <= ndesc_remaining) n = npkt_remaining; else { /* Can't send all, calculate how many can be sent */ n = (ndesc_remaining / SGE_MAX_WR_NDESC) * MAX_NPKT_IN_TYPE1_WR; if (ndesc_remaining % SGE_MAX_WR_NDESC) n += ndesc_to_npkt(ndesc_remaining % SGE_MAX_WR_NDESC); } /* Send n packets and update nm_txq->pidx and kring->nr_hwcur */ npkt_remaining -= n; cxgbe_nm_tx(sc, nm_txq, kring, n, npkt_remaining, txcsum); } MPASS(npkt_remaining == 0); MPASS(kring->nr_hwcur == head); MPASS(nm_txq->dbidx == nm_txq->pidx); /* * Second part: reclaim buffers for completed transmissions. */ if (reclaimed || flags & NAF_FORCE_RECLAIM || nm_kr_txempty(kring)) { reclaimed += reclaim_nm_tx_desc(nm_txq); kring->nr_hwtail += reclaimed; if (kring->nr_hwtail >= kring->nkr_num_slots) kring->nr_hwtail -= kring->nkr_num_slots; } return (0); } static int cxgbe_netmap_rxsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct netmap_ring *ring = kring->ring; struct ifnet *ifp = na->ifp; struct vi_info *vi = ifp->if_softc; struct adapter *sc = vi->pi->adapter; struct sge_nm_rxq *nm_rxq = &sc->sge.nm_rxq[vi->first_nm_rxq + kring->ring_id]; u_int const head = kring->rhead; u_int n; int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR; if (black_hole) return (0); /* No updates ever. */ if (netmap_no_pendintr || force_update) { kring->nr_hwtail = atomic_load_acq_32(&nm_rxq->fl_cidx); kring->nr_kflags &= ~NKR_PENDINTR; } /* Userspace done with buffers from kring->nr_hwcur to head */ n = head >= kring->nr_hwcur ? head - kring->nr_hwcur : kring->nkr_num_slots - kring->nr_hwcur + head; n &= ~7U; if (n > 0) { u_int fl_pidx = nm_rxq->fl_pidx; struct netmap_slot *slot = &ring->slot[fl_pidx]; uint64_t ba; int i, dbinc = 0, hwidx = nm_rxq->fl_hwidx; /* * We always deal with 8 buffers at a time. We must have * stopped at an 8B boundary (fl_pidx) last time around and we * must have a multiple of 8B buffers to give to the freelist. */ MPASS((fl_pidx & 7) == 0); MPASS((n & 7) == 0); IDXINCR(kring->nr_hwcur, n, kring->nkr_num_slots); IDXINCR(nm_rxq->fl_pidx, n, nm_rxq->fl_sidx); while (n > 0) { for (i = 0; i < 8; i++, fl_pidx++, slot++) { PNMB(na, slot, &ba); MPASS(ba != 0); nm_rxq->fl_desc[fl_pidx] = htobe64(ba | hwidx); slot->flags &= ~NS_BUF_CHANGED; MPASS(fl_pidx <= nm_rxq->fl_sidx); } n -= 8; if (fl_pidx == nm_rxq->fl_sidx) { fl_pidx = 0; slot = &ring->slot[0]; } if (++dbinc == 8 && n >= 32) { wmb(); t4_write_reg(sc, sc->sge_kdoorbell_reg, nm_rxq->fl_db_val | V_PIDX(dbinc)); dbinc = 0; } } MPASS(nm_rxq->fl_pidx == fl_pidx); if (dbinc > 0) { wmb(); t4_write_reg(sc, sc->sge_kdoorbell_reg, nm_rxq->fl_db_val | V_PIDX(dbinc)); } } return (0); } void cxgbe_nm_attach(struct vi_info *vi) { struct port_info *pi; struct adapter *sc; struct netmap_adapter na; MPASS(vi->nnmrxq > 0); MPASS(vi->ifp != NULL); pi = vi->pi; sc = pi->adapter; bzero(&na, sizeof(na)); na.ifp = vi->ifp; na.na_flags = NAF_BDG_MAYSLEEP; /* Netmap doesn't know about the space reserved for the status page. */ na.num_tx_desc = vi->qsize_txq - sc->params.sge.spg_len / EQ_ESIZE; /* * The freelist's cidx/pidx drives netmap's rx cidx/pidx. So * num_rx_desc is based on the number of buffers that can be held in the * freelist, and not the number of entries in the iq. (These two are * not exactly the same due to the space taken up by the status page). */ na.num_rx_desc = rounddown(vi->qsize_rxq, 8); na.nm_txsync = cxgbe_netmap_txsync; na.nm_rxsync = cxgbe_netmap_rxsync; na.nm_register = cxgbe_netmap_reg; na.num_tx_rings = vi->nnmtxq; na.num_rx_rings = vi->nnmrxq; netmap_attach(&na); /* This adds IFCAP_NETMAP to if_capabilities */ } void cxgbe_nm_detach(struct vi_info *vi) { MPASS(vi->nnmrxq > 0); MPASS(vi->ifp != NULL); netmap_detach(vi->ifp); } static void handle_nm_fw6_msg(struct adapter *sc, struct ifnet *ifp, const struct cpl_fw6_msg *cpl) { const struct cpl_sge_egr_update *egr; uint32_t oq; struct sge_nm_txq *nm_txq; if (cpl->type != FW_TYPE_RSSCPL && cpl->type != FW6_TYPE_RSSCPL) panic("%s: FW_TYPE 0x%x on nm_rxq.", __func__, cpl->type); /* data[0] is RSS header */ egr = (const void *)&cpl->data[1]; oq = be32toh(egr->opcode_qid); MPASS(G_CPL_OPCODE(oq) == CPL_SGE_EGR_UPDATE); nm_txq = (void *)sc->sge.eqmap[G_EGR_QID(oq) - sc->sge.eq_start]; netmap_tx_irq(ifp, nm_txq->nid); } void t4_nm_intr(void *arg) { struct sge_nm_rxq *nm_rxq = arg; struct vi_info *vi = nm_rxq->vi; struct adapter *sc = vi->pi->adapter; struct ifnet *ifp = vi->ifp; struct netmap_adapter *na = NA(ifp); struct netmap_kring *kring = &na->rx_rings[nm_rxq->nid]; struct netmap_ring *ring = kring->ring; struct iq_desc *d = &nm_rxq->iq_desc[nm_rxq->iq_cidx]; uint32_t lq; u_int n = 0, work = 0; uint8_t opcode; uint32_t fl_cidx = atomic_load_acq_32(&nm_rxq->fl_cidx); u_int fl_credits = fl_cidx & 7; while ((d->rsp.u.type_gen & F_RSPD_GEN) == nm_rxq->iq_gen) { rmb(); lq = be32toh(d->rsp.pldbuflen_qid); opcode = d->rss.opcode; switch (G_RSPD_TYPE(d->rsp.u.type_gen)) { case X_RSPD_TYPE_FLBUF: if (black_hole != 2) { /* No buffer packing so new buf every time */ MPASS(lq & F_RSPD_NEWBUF); } /* fall through */ case X_RSPD_TYPE_CPL: MPASS(opcode < NUM_CPL_CMDS); switch (opcode) { case CPL_FW4_MSG: case CPL_FW6_MSG: handle_nm_fw6_msg(sc, ifp, (const void *)&d->cpl[0]); break; case CPL_RX_PKT: ring->slot[fl_cidx].len = G_RSPD_LEN(lq) - sc->params.sge.fl_pktshift; ring->slot[fl_cidx].flags = kring->nkr_slot_flags; fl_cidx += (lq & F_RSPD_NEWBUF) ? 1 : 0; fl_credits += (lq & F_RSPD_NEWBUF) ? 1 : 0; if (__predict_false(fl_cidx == nm_rxq->fl_sidx)) fl_cidx = 0; break; default: panic("%s: unexpected opcode 0x%x on nm_rxq %p", __func__, opcode, nm_rxq); } break; case X_RSPD_TYPE_INTR: /* Not equipped to handle forwarded interrupts. */ panic("%s: netmap queue received interrupt for iq %u\n", __func__, lq); default: panic("%s: illegal response type %d on nm_rxq %p", __func__, G_RSPD_TYPE(d->rsp.u.type_gen), nm_rxq); } d++; if (__predict_false(++nm_rxq->iq_cidx == nm_rxq->iq_sidx)) { nm_rxq->iq_cidx = 0; d = &nm_rxq->iq_desc[0]; nm_rxq->iq_gen ^= F_RSPD_GEN; } if (__predict_false(++n == rx_ndesc)) { atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx); if (black_hole && fl_credits >= 8) { fl_credits /= 8; IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx); t4_write_reg(sc, sc->sge_kdoorbell_reg, nm_rxq->fl_db_val | V_PIDX(fl_credits)); fl_credits = fl_cidx & 7; } else if (!black_hole) { netmap_rx_irq(ifp, nm_rxq->nid, &work); MPASS(work != 0); } t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(n) | V_INGRESSQID(nm_rxq->iq_cntxt_id) | V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX))); n = 0; } } atomic_store_rel_32(&nm_rxq->fl_cidx, fl_cidx); if (black_hole) { fl_credits /= 8; IDXINCR(nm_rxq->fl_pidx, fl_credits * 8, nm_rxq->fl_sidx); t4_write_reg(sc, sc->sge_kdoorbell_reg, nm_rxq->fl_db_val | V_PIDX(fl_credits)); } else netmap_rx_irq(ifp, nm_rxq->nid, &work); t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(n) | V_INGRESSQID((u32)nm_rxq->iq_cntxt_id) | V_SEINTARM(V_QINTR_TIMER_IDX(holdoff_tmr_idx))); } #endif Index: projects/mips64-clang/sys/dev/cxgbe/t4_sge.c =================================================================== --- projects/mips64-clang/sys/dev/cxgbe/t4_sge.c (revision 305711) +++ projects/mips64-clang/sys/dev/cxgbe/t4_sge.c (revision 305712) @@ -1,5210 +1,5221 @@ /*- * Copyright (c) 2011 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DEV_NETMAP #include #include #include #include #include #endif #include "common/common.h" #include "common/t4_regs.h" #include "common/t4_regs_values.h" #include "common/t4_msg.h" #include "t4_l2t.h" #include "t4_mp_ring.h" #ifdef T4_PKT_TIMESTAMP #define RX_COPY_THRESHOLD (MINCLSIZE - 8) #else #define RX_COPY_THRESHOLD MINCLSIZE #endif /* * Ethernet frames are DMA'd at this byte offset into the freelist buffer. * 0-7 are valid values. */ static int fl_pktshift = 2; TUNABLE_INT("hw.cxgbe.fl_pktshift", &fl_pktshift); /* * Pad ethernet payload up to this boundary. * -1: driver should figure out a good value. * 0: disable padding. * Any power of 2 from 32 to 4096 (both inclusive) is also a valid value. */ int fl_pad = -1; TUNABLE_INT("hw.cxgbe.fl_pad", &fl_pad); /* * Status page length. * -1: driver should figure out a good value. * 64 or 128 are the only other valid values. */ static int spg_len = -1; TUNABLE_INT("hw.cxgbe.spg_len", &spg_len); /* * Congestion drops. * -1: no congestion feedback (not recommended). * 0: backpressure the channel instead of dropping packets right away. * 1: no backpressure, drop packets for the congested queue immediately. */ static int cong_drop = 0; TUNABLE_INT("hw.cxgbe.cong_drop", &cong_drop); /* * Deliver multiple frames in the same free list buffer if they fit. * -1: let the driver decide whether to enable buffer packing or not. * 0: disable buffer packing. * 1: enable buffer packing. */ static int buffer_packing = -1; TUNABLE_INT("hw.cxgbe.buffer_packing", &buffer_packing); /* * Start next frame in a packed buffer at this boundary. * -1: driver should figure out a good value. * T4: driver will ignore this and use the same value as fl_pad above. * T5: 16, or a power of 2 from 64 to 4096 (both inclusive) is a valid value. */ static int fl_pack = -1; TUNABLE_INT("hw.cxgbe.fl_pack", &fl_pack); /* * Allow the driver to create mbuf(s) in a cluster allocated for rx. * 0: never; always allocate mbufs from the zone_mbuf UMA zone. * 1: ok to create mbuf(s) within a cluster if there is room. */ static int allow_mbufs_in_cluster = 1; TUNABLE_INT("hw.cxgbe.allow_mbufs_in_cluster", &allow_mbufs_in_cluster); /* * Largest rx cluster size that the driver is allowed to allocate. */ static int largest_rx_cluster = MJUM16BYTES; TUNABLE_INT("hw.cxgbe.largest_rx_cluster", &largest_rx_cluster); /* * Size of cluster allocation that's most likely to succeed. The driver will * fall back to this size if it fails to allocate clusters larger than this. */ static int safest_rx_cluster = PAGE_SIZE; TUNABLE_INT("hw.cxgbe.safest_rx_cluster", &safest_rx_cluster); struct txpkts { u_int wr_type; /* type 0 or type 1 */ u_int npkt; /* # of packets in this work request */ u_int plen; /* total payload (sum of all packets) */ u_int len16; /* # of 16B pieces used by this work request */ }; /* A packet's SGL. This + m_pkthdr has all info needed for tx */ struct sgl { struct sglist sg; struct sglist_seg seg[TX_SGL_SEGS]; }; static int service_iq(struct sge_iq *, int); static struct mbuf *get_fl_payload(struct adapter *, struct sge_fl *, uint32_t); static int t4_eth_rx(struct sge_iq *, const struct rss_header *, struct mbuf *); static inline void init_iq(struct sge_iq *, struct adapter *, int, int, int); static inline void init_fl(struct adapter *, struct sge_fl *, int, int, char *); static inline void init_eq(struct adapter *, struct sge_eq *, int, int, uint8_t, uint16_t, char *); static int alloc_ring(struct adapter *, size_t, bus_dma_tag_t *, bus_dmamap_t *, bus_addr_t *, void **); static int free_ring(struct adapter *, bus_dma_tag_t, bus_dmamap_t, bus_addr_t, void *); static int alloc_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *, int, int); static int free_iq_fl(struct vi_info *, struct sge_iq *, struct sge_fl *); static void add_fl_sysctls(struct sysctl_ctx_list *, struct sysctl_oid *, struct sge_fl *); static int alloc_fwq(struct adapter *); static int free_fwq(struct adapter *); static int alloc_mgmtq(struct adapter *); static int free_mgmtq(struct adapter *); static int alloc_rxq(struct vi_info *, struct sge_rxq *, int, int, struct sysctl_oid *); static int free_rxq(struct vi_info *, struct sge_rxq *); #ifdef TCP_OFFLOAD static int alloc_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *, int, int, struct sysctl_oid *); static int free_ofld_rxq(struct vi_info *, struct sge_ofld_rxq *); #endif #ifdef DEV_NETMAP static int alloc_nm_rxq(struct vi_info *, struct sge_nm_rxq *, int, int, struct sysctl_oid *); static int free_nm_rxq(struct vi_info *, struct sge_nm_rxq *); static int alloc_nm_txq(struct vi_info *, struct sge_nm_txq *, int, int, struct sysctl_oid *); static int free_nm_txq(struct vi_info *, struct sge_nm_txq *); #endif static int ctrl_eq_alloc(struct adapter *, struct sge_eq *); static int eth_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *); #ifdef TCP_OFFLOAD static int ofld_eq_alloc(struct adapter *, struct vi_info *, struct sge_eq *); #endif static int alloc_eq(struct adapter *, struct vi_info *, struct sge_eq *); static int free_eq(struct adapter *, struct sge_eq *); static int alloc_wrq(struct adapter *, struct vi_info *, struct sge_wrq *, struct sysctl_oid *); static int free_wrq(struct adapter *, struct sge_wrq *); static int alloc_txq(struct vi_info *, struct sge_txq *, int, struct sysctl_oid *); static int free_txq(struct vi_info *, struct sge_txq *); static void oneseg_dma_callback(void *, bus_dma_segment_t *, int, int); static inline void ring_fl_db(struct adapter *, struct sge_fl *); static int refill_fl(struct adapter *, struct sge_fl *, int); static void refill_sfl(void *); static int alloc_fl_sdesc(struct sge_fl *); static void free_fl_sdesc(struct adapter *, struct sge_fl *); static void find_best_refill_source(struct adapter *, struct sge_fl *, int); static void find_safe_refill_source(struct adapter *, struct sge_fl *); static void add_fl_to_sfl(struct adapter *, struct sge_fl *); static inline void get_pkt_gl(struct mbuf *, struct sglist *); static inline u_int txpkt_len16(u_int, u_int); static inline u_int txpkt_vm_len16(u_int, u_int); static inline u_int txpkts0_len16(u_int); static inline u_int txpkts1_len16(void); static u_int write_txpkt_wr(struct sge_txq *, struct fw_eth_tx_pkt_wr *, struct mbuf *, u_int); -static u_int write_txpkt_vm_wr(struct sge_txq *, struct fw_eth_tx_pkt_vm_wr *, - struct mbuf *, u_int); +static u_int write_txpkt_vm_wr(struct adapter *, struct sge_txq *, + struct fw_eth_tx_pkt_vm_wr *, struct mbuf *, u_int); static int try_txpkts(struct mbuf *, struct mbuf *, struct txpkts *, u_int); static int add_to_txpkts(struct mbuf *, struct txpkts *, u_int); static u_int write_txpkts_wr(struct sge_txq *, struct fw_eth_tx_pkts_wr *, struct mbuf *, const struct txpkts *, u_int); static void write_gl_to_txd(struct sge_txq *, struct mbuf *, caddr_t *, int); static inline void copy_to_txd(struct sge_eq *, caddr_t, caddr_t *, int); static inline void ring_eq_db(struct adapter *, struct sge_eq *, u_int); static inline uint16_t read_hw_cidx(struct sge_eq *); static inline u_int reclaimable_tx_desc(struct sge_eq *); static inline u_int total_available_tx_desc(struct sge_eq *); static u_int reclaim_tx_descs(struct sge_txq *, u_int); static void tx_reclaim(void *, int); static __be64 get_flit(struct sglist_seg *, int, int); static int handle_sge_egr_update(struct sge_iq *, const struct rss_header *, struct mbuf *); static int handle_fw_msg(struct sge_iq *, const struct rss_header *, struct mbuf *); static int t4_handle_wrerr_rpl(struct adapter *, const __be64 *); static void wrq_tx_drain(void *, int); static void drain_wrq_wr_list(struct adapter *, struct sge_wrq *); static int sysctl_uint16(SYSCTL_HANDLER_ARGS); static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS); static int sysctl_tc(SYSCTL_HANDLER_ARGS); static counter_u64_t extfree_refs; static counter_u64_t extfree_rels; an_handler_t t4_an_handler; fw_msg_handler_t t4_fw_msg_handler[NUM_FW6_TYPES]; cpl_handler_t t4_cpl_handler[NUM_CPL_CMDS]; static int an_not_handled(struct sge_iq *iq, const struct rsp_ctrl *ctrl) { #ifdef INVARIANTS panic("%s: async notification on iq %p (ctrl %p)", __func__, iq, ctrl); #else log(LOG_ERR, "%s: async notification on iq %p (ctrl %p)\n", __func__, iq, ctrl); #endif return (EDOOFUS); } int t4_register_an_handler(an_handler_t h) { uintptr_t *loc, new; new = h ? (uintptr_t)h : (uintptr_t)an_not_handled; loc = (uintptr_t *) &t4_an_handler; atomic_store_rel_ptr(loc, new); return (0); } static int fw_msg_not_handled(struct adapter *sc, const __be64 *rpl) { const struct cpl_fw6_msg *cpl = __containerof(rpl, struct cpl_fw6_msg, data[0]); #ifdef INVARIANTS panic("%s: fw_msg type %d", __func__, cpl->type); #else log(LOG_ERR, "%s: fw_msg type %d\n", __func__, cpl->type); #endif return (EDOOFUS); } int t4_register_fw_msg_handler(int type, fw_msg_handler_t h) { uintptr_t *loc, new; if (type >= nitems(t4_fw_msg_handler)) return (EINVAL); /* * These are dispatched by the handler for FW{4|6}_CPL_MSG using the CPL * handler dispatch table. Reject any attempt to install a handler for * this subtype. */ if (type == FW_TYPE_RSSCPL || type == FW6_TYPE_RSSCPL) return (EINVAL); new = h ? (uintptr_t)h : (uintptr_t)fw_msg_not_handled; loc = (uintptr_t *) &t4_fw_msg_handler[type]; atomic_store_rel_ptr(loc, new); return (0); } static int cpl_not_handled(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { #ifdef INVARIANTS panic("%s: opcode 0x%02x on iq %p with payload %p", __func__, rss->opcode, iq, m); #else log(LOG_ERR, "%s: opcode 0x%02x on iq %p with payload %p\n", __func__, rss->opcode, iq, m); m_freem(m); #endif return (EDOOFUS); } int t4_register_cpl_handler(int opcode, cpl_handler_t h) { uintptr_t *loc, new; if (opcode >= nitems(t4_cpl_handler)) return (EINVAL); new = h ? (uintptr_t)h : (uintptr_t)cpl_not_handled; loc = (uintptr_t *) &t4_cpl_handler[opcode]; atomic_store_rel_ptr(loc, new); return (0); } /* * Called on MOD_LOAD. Validates and calculates the SGE tunables. */ void t4_sge_modload(void) { int i; if (fl_pktshift < 0 || fl_pktshift > 7) { printf("Invalid hw.cxgbe.fl_pktshift value (%d)," " using 2 instead.\n", fl_pktshift); fl_pktshift = 2; } if (spg_len != 64 && spg_len != 128) { int len; #if defined(__i386__) || defined(__amd64__) len = cpu_clflush_line_size > 64 ? 128 : 64; #else len = 64; #endif if (spg_len != -1) { printf("Invalid hw.cxgbe.spg_len value (%d)," " using %d instead.\n", spg_len, len); } spg_len = len; } if (cong_drop < -1 || cong_drop > 1) { printf("Invalid hw.cxgbe.cong_drop value (%d)," " using 0 instead.\n", cong_drop); cong_drop = 0; } extfree_refs = counter_u64_alloc(M_WAITOK); extfree_rels = counter_u64_alloc(M_WAITOK); counter_u64_zero(extfree_refs); counter_u64_zero(extfree_rels); t4_an_handler = an_not_handled; for (i = 0; i < nitems(t4_fw_msg_handler); i++) t4_fw_msg_handler[i] = fw_msg_not_handled; for (i = 0; i < nitems(t4_cpl_handler); i++) t4_cpl_handler[i] = cpl_not_handled; t4_register_cpl_handler(CPL_FW4_MSG, handle_fw_msg); t4_register_cpl_handler(CPL_FW6_MSG, handle_fw_msg); t4_register_cpl_handler(CPL_SGE_EGR_UPDATE, handle_sge_egr_update); t4_register_cpl_handler(CPL_RX_PKT, t4_eth_rx); t4_register_fw_msg_handler(FW6_TYPE_CMD_RPL, t4_handle_fw_rpl); t4_register_fw_msg_handler(FW6_TYPE_WRERR_RPL, t4_handle_wrerr_rpl); } void t4_sge_modunload(void) { counter_u64_free(extfree_refs); counter_u64_free(extfree_rels); } uint64_t t4_sge_extfree_refs(void) { uint64_t refs, rels; rels = counter_u64_fetch(extfree_rels); refs = counter_u64_fetch(extfree_refs); return (refs - rels); } static inline void setup_pad_and_pack_boundaries(struct adapter *sc) { uint32_t v, m; - int pad, pack; + int pad, pack, pad_shift; + pad_shift = chip_id(sc) > CHELSIO_T5 ? X_T6_INGPADBOUNDARY_SHIFT : + X_INGPADBOUNDARY_SHIFT; pad = fl_pad; - if (fl_pad < 32 || fl_pad > 4096 || !powerof2(fl_pad)) { + if (fl_pad < (1 << pad_shift) || + fl_pad > (1 << (pad_shift + M_INGPADBOUNDARY)) || + !powerof2(fl_pad)) { /* * If there is any chance that we might use buffer packing and * the chip is a T4, then pick 64 as the pad/pack boundary. Set - * it to 32 in all other cases. + * it to the minimum allowed in all other cases. */ - pad = is_t4(sc) && buffer_packing ? 64 : 32; + pad = is_t4(sc) && buffer_packing ? 64 : 1 << pad_shift; /* * For fl_pad = 0 we'll still write a reasonable value to the * register but all the freelists will opt out of padding. * We'll complain here only if the user tried to set it to a * value greater than 0 that was invalid. */ if (fl_pad > 0) { device_printf(sc->dev, "Invalid hw.cxgbe.fl_pad value" " (%d), using %d instead.\n", fl_pad, pad); } } m = V_INGPADBOUNDARY(M_INGPADBOUNDARY); - v = V_INGPADBOUNDARY(ilog2(pad) - 5); + v = V_INGPADBOUNDARY(ilog2(pad) - pad_shift); t4_set_reg_field(sc, A_SGE_CONTROL, m, v); if (is_t4(sc)) { if (fl_pack != -1 && fl_pack != pad) { /* Complain but carry on. */ device_printf(sc->dev, "hw.cxgbe.fl_pack (%d) ignored," " using %d instead.\n", fl_pack, pad); } return; } pack = fl_pack; if (fl_pack < 16 || fl_pack == 32 || fl_pack > 4096 || !powerof2(fl_pack)) { pack = max(sc->params.pci.mps, CACHE_LINE_SIZE); MPASS(powerof2(pack)); if (pack < 16) pack = 16; if (pack == 32) pack = 64; if (pack > 4096) pack = 4096; if (fl_pack != -1) { device_printf(sc->dev, "Invalid hw.cxgbe.fl_pack value" " (%d), using %d instead.\n", fl_pack, pack); } } m = V_INGPACKBOUNDARY(M_INGPACKBOUNDARY); if (pack == 16) v = V_INGPACKBOUNDARY(0); else v = V_INGPACKBOUNDARY(ilog2(pack) - 5); MPASS(!is_t4(sc)); /* T4 doesn't have SGE_CONTROL2 */ t4_set_reg_field(sc, A_SGE_CONTROL2, m, v); } /* * adap->params.vpd.cclk must be set up before this is called. */ void t4_tweak_chip_settings(struct adapter *sc) { int i; uint32_t v, m; int intr_timer[SGE_NTIMERS] = {1, 5, 10, 50, 100, 200}; int timer_max = M_TIMERVALUE0 * 1000 / sc->params.vpd.cclk; int intr_pktcount[SGE_NCOUNTERS] = {1, 8, 16, 32}; /* 63 max */ uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE); static int sge_flbuf_sizes[] = { MCLBYTES, #if MJUMPAGESIZE != MCLBYTES MJUMPAGESIZE, MJUMPAGESIZE - CL_METADATA_SIZE, MJUMPAGESIZE - 2 * MSIZE - CL_METADATA_SIZE, #endif MJUM9BYTES, MJUM16BYTES, MCLBYTES - MSIZE - CL_METADATA_SIZE, MJUM9BYTES - CL_METADATA_SIZE, MJUM16BYTES - CL_METADATA_SIZE, }; KASSERT(sc->flags & MASTER_PF, ("%s: trying to change chip settings when not master.", __func__)); m = V_PKTSHIFT(M_PKTSHIFT) | F_RXPKTCPLMODE | F_EGRSTATUSPAGESIZE; v = V_PKTSHIFT(fl_pktshift) | F_RXPKTCPLMODE | V_EGRSTATUSPAGESIZE(spg_len == 128); t4_set_reg_field(sc, A_SGE_CONTROL, m, v); setup_pad_and_pack_boundaries(sc); v = V_HOSTPAGESIZEPF0(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF1(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF2(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF3(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF4(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF5(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF6(PAGE_SHIFT - 10) | V_HOSTPAGESIZEPF7(PAGE_SHIFT - 10); t4_write_reg(sc, A_SGE_HOST_PAGE_SIZE, v); KASSERT(nitems(sge_flbuf_sizes) <= SGE_FLBUF_SIZES, ("%s: hw buffer size table too big", __func__)); for (i = 0; i < min(nitems(sge_flbuf_sizes), SGE_FLBUF_SIZES); i++) { t4_write_reg(sc, A_SGE_FL_BUFFER_SIZE0 + (4 * i), sge_flbuf_sizes[i]); } v = V_THRESHOLD_0(intr_pktcount[0]) | V_THRESHOLD_1(intr_pktcount[1]) | V_THRESHOLD_2(intr_pktcount[2]) | V_THRESHOLD_3(intr_pktcount[3]); t4_write_reg(sc, A_SGE_INGRESS_RX_THRESHOLD, v); KASSERT(intr_timer[0] <= timer_max, ("%s: not a single usable timer (%d, %d)", __func__, intr_timer[0], timer_max)); for (i = 1; i < nitems(intr_timer); i++) { KASSERT(intr_timer[i] >= intr_timer[i - 1], ("%s: timers not listed in increasing order (%d)", __func__, i)); while (intr_timer[i] > timer_max) { if (i == nitems(intr_timer) - 1) { intr_timer[i] = timer_max; break; } intr_timer[i] += intr_timer[i - 1]; intr_timer[i] /= 2; } } v = V_TIMERVALUE0(us_to_core_ticks(sc, intr_timer[0])) | V_TIMERVALUE1(us_to_core_ticks(sc, intr_timer[1])); t4_write_reg(sc, A_SGE_TIMER_VALUE_0_AND_1, v); v = V_TIMERVALUE2(us_to_core_ticks(sc, intr_timer[2])) | V_TIMERVALUE3(us_to_core_ticks(sc, intr_timer[3])); t4_write_reg(sc, A_SGE_TIMER_VALUE_2_AND_3, v); v = V_TIMERVALUE4(us_to_core_ticks(sc, intr_timer[4])) | V_TIMERVALUE5(us_to_core_ticks(sc, intr_timer[5])); t4_write_reg(sc, A_SGE_TIMER_VALUE_4_AND_5, v); /* 4K, 16K, 64K, 256K DDP "page sizes" for TDDP */ v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6); t4_write_reg(sc, A_ULP_RX_TDDP_PSZ, v); /* * 4K, 8K, 16K, 64K DDP "page sizes" for iSCSI DDP. These have been * chosen with MAXPHYS = 128K in mind. The largest DDP buffer that we * may have to deal with is MAXPHYS + 1 page. */ v = V_HPZ0(0) | V_HPZ1(1) | V_HPZ2(2) | V_HPZ3(4); t4_write_reg(sc, A_ULP_RX_ISCSI_PSZ, v); /* We use multiple DDP page sizes both in plain-TOE and ISCSI modes. */ m = v = F_TDDPTAGTCB | F_ISCSITAGTCB; t4_set_reg_field(sc, A_ULP_RX_CTL, m, v); m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; t4_set_reg_field(sc, A_TP_PARA_REG5, m, v); } /* * SGE wants the buffer to be at least 64B and then a multiple of 16. If * padding is in use, the buffer's start and end need to be aligned to the pad * boundary as well. We'll just make sure that the size is a multiple of the * boundary here, it is up to the buffer allocation code to make sure the start * of the buffer is aligned as well. */ static inline int hwsz_ok(struct adapter *sc, int hwsz) { int mask = fl_pad ? sc->params.sge.pad_boundary - 1 : 16 - 1; return (hwsz >= 64 && (hwsz & mask) == 0); } /* * XXX: driver really should be able to deal with unexpected settings. */ int t4_read_chip_settings(struct adapter *sc) { struct sge *s = &sc->sge; struct sge_params *sp = &sc->params.sge; int i, j, n, rc = 0; uint32_t m, v, r; uint16_t indsz = min(RX_COPY_THRESHOLD - 1, M_INDICATESIZE); static int sw_buf_sizes[] = { /* Sorted by size */ MCLBYTES, #if MJUMPAGESIZE != MCLBYTES MJUMPAGESIZE, #endif MJUM9BYTES, MJUM16BYTES }; struct sw_zone_info *swz, *safe_swz; struct hw_buf_info *hwb; m = F_RXPKTCPLMODE; v = F_RXPKTCPLMODE; r = sc->params.sge.sge_control; if ((r & m) != v) { device_printf(sc->dev, "invalid SGE_CONTROL(0x%x)\n", r); rc = EINVAL; } /* * If this changes then every single use of PAGE_SHIFT in the driver * needs to be carefully reviewed for PAGE_SHIFT vs sp->page_shift. */ if (sp->page_shift != PAGE_SHIFT) { device_printf(sc->dev, "invalid SGE_HOST_PAGE_SIZE(0x%x)\n", r); rc = EINVAL; } /* Filter out unusable hw buffer sizes entirely (mark with -2). */ hwb = &s->hw_buf_info[0]; for (i = 0; i < nitems(s->hw_buf_info); i++, hwb++) { r = sc->params.sge.sge_fl_buffer_size[i]; hwb->size = r; hwb->zidx = hwsz_ok(sc, r) ? -1 : -2; hwb->next = -1; } /* * Create a sorted list in decreasing order of hw buffer sizes (and so * increasing order of spare area) for each software zone. * * If padding is enabled then the start and end of the buffer must align * to the pad boundary; if packing is enabled then they must align with * the pack boundary as well. Allocations from the cluster zones are * aligned to min(size, 4K), so the buffer starts at that alignment and * ends at hwb->size alignment. If mbuf inlining is allowed the * starting alignment will be reduced to MSIZE and the driver will * exercise appropriate caution when deciding on the best buffer layout * to use. */ n = 0; /* no usable buffer size to begin with */ swz = &s->sw_zone_info[0]; safe_swz = NULL; for (i = 0; i < SW_ZONE_SIZES; i++, swz++) { int8_t head = -1, tail = -1; swz->size = sw_buf_sizes[i]; swz->zone = m_getzone(swz->size); swz->type = m_gettype(swz->size); if (swz->size < PAGE_SIZE) { MPASS(powerof2(swz->size)); if (fl_pad && (swz->size % sp->pad_boundary != 0)) continue; } if (swz->size == safest_rx_cluster) safe_swz = swz; hwb = &s->hw_buf_info[0]; for (j = 0; j < SGE_FLBUF_SIZES; j++, hwb++) { if (hwb->zidx != -1 || hwb->size > swz->size) continue; #ifdef INVARIANTS if (fl_pad) MPASS(hwb->size % sp->pad_boundary == 0); #endif hwb->zidx = i; if (head == -1) head = tail = j; else if (hwb->size < s->hw_buf_info[tail].size) { s->hw_buf_info[tail].next = j; tail = j; } else { int8_t *cur; struct hw_buf_info *t; for (cur = &head; *cur != -1; cur = &t->next) { t = &s->hw_buf_info[*cur]; if (hwb->size == t->size) { hwb->zidx = -2; break; } if (hwb->size > t->size) { hwb->next = *cur; *cur = j; break; } } } } swz->head_hwidx = head; swz->tail_hwidx = tail; if (tail != -1) { n++; if (swz->size - s->hw_buf_info[tail].size >= CL_METADATA_SIZE) sc->flags |= BUF_PACKING_OK; } } if (n == 0) { device_printf(sc->dev, "no usable SGE FL buffer size.\n"); rc = EINVAL; } s->safe_hwidx1 = -1; s->safe_hwidx2 = -1; if (safe_swz != NULL) { s->safe_hwidx1 = safe_swz->head_hwidx; for (i = safe_swz->head_hwidx; i != -1; i = hwb->next) { int spare; hwb = &s->hw_buf_info[i]; #ifdef INVARIANTS if (fl_pad) MPASS(hwb->size % sp->pad_boundary == 0); #endif spare = safe_swz->size - hwb->size; if (spare >= CL_METADATA_SIZE) { s->safe_hwidx2 = i; break; } } } if (sc->flags & IS_VF) return (0); v = V_HPZ0(0) | V_HPZ1(2) | V_HPZ2(4) | V_HPZ3(6); r = t4_read_reg(sc, A_ULP_RX_TDDP_PSZ); if (r != v) { device_printf(sc->dev, "invalid ULP_RX_TDDP_PSZ(0x%x)\n", r); rc = EINVAL; } m = v = F_TDDPTAGTCB; r = t4_read_reg(sc, A_ULP_RX_CTL); if ((r & m) != v) { device_printf(sc->dev, "invalid ULP_RX_CTL(0x%x)\n", r); rc = EINVAL; } m = V_INDICATESIZE(M_INDICATESIZE) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; v = V_INDICATESIZE(indsz) | F_REARMDDPOFFSET | F_RESETDDPOFFSET; r = t4_read_reg(sc, A_TP_PARA_REG5); if ((r & m) != v) { device_printf(sc->dev, "invalid TP_PARA_REG5(0x%x)\n", r); rc = EINVAL; } t4_init_tp_params(sc); t4_read_mtu_tbl(sc, sc->params.mtus, NULL); t4_load_mtus(sc, sc->params.mtus, sc->params.a_wnd, sc->params.b_wnd); return (rc); } int t4_create_dma_tag(struct adapter *sc) { int rc; rc = bus_dma_tag_create(bus_get_dma_tag(sc->dev), 1, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, BUS_SPACE_MAXSIZE, BUS_SPACE_UNRESTRICTED, BUS_SPACE_MAXSIZE, BUS_DMA_ALLOCNOW, NULL, NULL, &sc->dmat); if (rc != 0) { device_printf(sc->dev, "failed to create main DMA tag: %d\n", rc); } return (rc); } void t4_sge_sysctls(struct adapter *sc, struct sysctl_ctx_list *ctx, struct sysctl_oid_list *children) { struct sge_params *sp = &sc->params.sge; SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "buffer_sizes", CTLTYPE_STRING | CTLFLAG_RD, &sc->sge, 0, sysctl_bufsizes, "A", "freelist buffer sizes"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pktshift", CTLFLAG_RD, NULL, sp->fl_pktshift, "payload DMA offset in rx buffer (bytes)"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pad", CTLFLAG_RD, NULL, sp->pad_boundary, "payload pad boundary (bytes)"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "spg_len", CTLFLAG_RD, NULL, sp->spg_len, "status page size (bytes)"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "cong_drop", CTLFLAG_RD, NULL, cong_drop, "congestion drop setting"); SYSCTL_ADD_INT(ctx, children, OID_AUTO, "fl_pack", CTLFLAG_RD, NULL, sp->pack_boundary, "payload pack boundary (bytes)"); } int t4_destroy_dma_tag(struct adapter *sc) { if (sc->dmat) bus_dma_tag_destroy(sc->dmat); return (0); } /* * Allocate and initialize the firmware event queue and the management queue. * * Returns errno on failure. Resources allocated up to that point may still be * allocated. Caller is responsible for cleanup in case this function fails. */ int t4_setup_adapter_queues(struct adapter *sc) { int rc; ADAPTER_LOCK_ASSERT_NOTOWNED(sc); sysctl_ctx_init(&sc->ctx); sc->flags |= ADAP_SYSCTL_CTX; /* * Firmware event queue */ rc = alloc_fwq(sc); if (rc != 0) return (rc); /* * Management queue. This is just a control queue that uses the fwq as * its associated iq. */ if (!(sc->flags & IS_VF)) rc = alloc_mgmtq(sc); return (rc); } /* * Idempotent */ int t4_teardown_adapter_queues(struct adapter *sc) { ADAPTER_LOCK_ASSERT_NOTOWNED(sc); /* Do this before freeing the queue */ if (sc->flags & ADAP_SYSCTL_CTX) { sysctl_ctx_free(&sc->ctx); sc->flags &= ~ADAP_SYSCTL_CTX; } free_mgmtq(sc); free_fwq(sc); return (0); } static inline int first_vector(struct vi_info *vi) { struct adapter *sc = vi->pi->adapter; if (sc->intr_count == 1) return (0); return (vi->first_intr); } /* * Given an arbitrary "index," come up with an iq that can be used by other * queues (of this VI) for interrupt forwarding, SGE egress updates, etc. * The iq returned is guaranteed to be something that takes direct interrupts. */ static struct sge_iq * vi_intr_iq(struct vi_info *vi, int idx) { struct adapter *sc = vi->pi->adapter; struct sge *s = &sc->sge; struct sge_iq *iq = NULL; int nintr, i; if (sc->intr_count == 1) return (&sc->sge.fwq); nintr = vi->nintr; KASSERT(nintr != 0, ("%s: vi %p has no exclusive interrupts, total interrupts = %d", __func__, vi, sc->intr_count)); i = idx % nintr; if (vi->flags & INTR_RXQ) { if (i < vi->nrxq) { iq = &s->rxq[vi->first_rxq + i].iq; goto done; } i -= vi->nrxq; } #ifdef TCP_OFFLOAD if (vi->flags & INTR_OFLD_RXQ) { if (i < vi->nofldrxq) { iq = &s->ofld_rxq[vi->first_ofld_rxq + i].iq; goto done; } i -= vi->nofldrxq; } #endif panic("%s: vi %p, intr_flags 0x%lx, idx %d, total intr %d\n", __func__, vi, vi->flags & INTR_ALL, idx, nintr); done: MPASS(iq != NULL); KASSERT(iq->flags & IQ_INTR, ("%s: iq %p (vi %p, intr_flags 0x%lx, idx %d)", __func__, iq, vi, vi->flags & INTR_ALL, idx)); return (iq); } /* Maximum payload that can be delivered with a single iq descriptor */ static inline int mtu_to_max_payload(struct adapter *sc, int mtu, const int toe) { int payload; #ifdef TCP_OFFLOAD if (toe) { payload = sc->tt.rx_coalesce ? G_RXCOALESCESIZE(t4_read_reg(sc, A_TP_PARA_REG2)) : mtu; } else { #endif /* large enough even when hw VLAN extraction is disabled */ payload = sc->params.sge.fl_pktshift + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN + mtu; #ifdef TCP_OFFLOAD } #endif return (payload); } int t4_setup_vi_queues(struct vi_info *vi) { int rc = 0, i, j, intr_idx, iqid; struct sge_rxq *rxq; struct sge_txq *txq; struct sge_wrq *ctrlq; #ifdef TCP_OFFLOAD struct sge_ofld_rxq *ofld_rxq; struct sge_wrq *ofld_txq; #endif #ifdef DEV_NETMAP int saved_idx; struct sge_nm_rxq *nm_rxq; struct sge_nm_txq *nm_txq; #endif char name[16]; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct ifnet *ifp = vi->ifp; struct sysctl_oid *oid = device_get_sysctl_tree(vi->dev); struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); int maxp, mtu = ifp->if_mtu; /* Interrupt vector to start from (when using multiple vectors) */ intr_idx = first_vector(vi); #ifdef DEV_NETMAP saved_idx = intr_idx; if (ifp->if_capabilities & IFCAP_NETMAP) { /* netmap is supported with direct interrupts only. */ MPASS(vi->flags & INTR_RXQ); /* * We don't have buffers to back the netmap rx queues * right now so we create the queues in a way that * doesn't set off any congestion signal in the chip. */ oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_rxq", CTLFLAG_RD, NULL, "rx queues"); for_each_nm_rxq(vi, i, nm_rxq) { rc = alloc_nm_rxq(vi, nm_rxq, intr_idx, i, oid); if (rc != 0) goto done; intr_idx++; } oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "nm_txq", CTLFLAG_RD, NULL, "tx queues"); for_each_nm_txq(vi, i, nm_txq) { iqid = vi->first_nm_rxq + (i % vi->nnmrxq); rc = alloc_nm_txq(vi, nm_txq, iqid, i, oid); if (rc != 0) goto done; } } /* Normal rx queues and netmap rx queues share the same interrupts. */ intr_idx = saved_idx; #endif /* * First pass over all NIC and TOE rx queues: * a) initialize iq and fl * b) allocate queue iff it will take direct interrupts. */ maxp = mtu_to_max_payload(sc, mtu, 0); if (vi->flags & INTR_RXQ) { oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq", CTLFLAG_RD, NULL, "rx queues"); } for_each_rxq(vi, i, rxq) { init_iq(&rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq); snprintf(name, sizeof(name), "%s rxq%d-fl", device_get_nameunit(vi->dev), i); init_fl(sc, &rxq->fl, vi->qsize_rxq / 8, maxp, name); if (vi->flags & INTR_RXQ) { rxq->iq.flags |= IQ_INTR; rc = alloc_rxq(vi, rxq, intr_idx, i, oid); if (rc != 0) goto done; intr_idx++; } } #ifdef DEV_NETMAP if (ifp->if_capabilities & IFCAP_NETMAP) intr_idx = saved_idx + max(vi->nrxq, vi->nnmrxq); #endif #ifdef TCP_OFFLOAD maxp = mtu_to_max_payload(sc, mtu, 1); if (vi->flags & INTR_OFLD_RXQ) { oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq", CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections"); } for_each_ofld_rxq(vi, i, ofld_rxq) { init_iq(&ofld_rxq->iq, sc, vi->tmr_idx, vi->pktc_idx, vi->qsize_rxq); snprintf(name, sizeof(name), "%s ofld_rxq%d-fl", device_get_nameunit(vi->dev), i); init_fl(sc, &ofld_rxq->fl, vi->qsize_rxq / 8, maxp, name); if (vi->flags & INTR_OFLD_RXQ) { ofld_rxq->iq.flags |= IQ_INTR; rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid); if (rc != 0) goto done; intr_idx++; } } #endif /* * Second pass over all NIC and TOE rx queues. The queues forwarding * their interrupts are allocated now. */ j = 0; if (!(vi->flags & INTR_RXQ)) { oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "rxq", CTLFLAG_RD, NULL, "rx queues"); for_each_rxq(vi, i, rxq) { MPASS(!(rxq->iq.flags & IQ_INTR)); intr_idx = vi_intr_iq(vi, j)->abs_id; rc = alloc_rxq(vi, rxq, intr_idx, i, oid); if (rc != 0) goto done; j++; } } #ifdef TCP_OFFLOAD if (vi->nofldrxq != 0 && !(vi->flags & INTR_OFLD_RXQ)) { oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_rxq", CTLFLAG_RD, NULL, "rx queues for offloaded TCP connections"); for_each_ofld_rxq(vi, i, ofld_rxq) { MPASS(!(ofld_rxq->iq.flags & IQ_INTR)); intr_idx = vi_intr_iq(vi, j)->abs_id; rc = alloc_ofld_rxq(vi, ofld_rxq, intr_idx, i, oid); if (rc != 0) goto done; j++; } } #endif /* * Now the tx queues. Only one pass needed. */ oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "txq", CTLFLAG_RD, NULL, "tx queues"); j = 0; for_each_txq(vi, i, txq) { iqid = vi_intr_iq(vi, j)->cntxt_id; snprintf(name, sizeof(name), "%s txq%d", device_get_nameunit(vi->dev), i); init_eq(sc, &txq->eq, EQ_ETH, vi->qsize_txq, pi->tx_chan, iqid, name); rc = alloc_txq(vi, txq, i, oid); if (rc != 0) goto done; j++; } #ifdef TCP_OFFLOAD oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ofld_txq", CTLFLAG_RD, NULL, "tx queues for offloaded TCP connections"); for_each_ofld_txq(vi, i, ofld_txq) { struct sysctl_oid *oid2; iqid = vi_intr_iq(vi, j)->cntxt_id; snprintf(name, sizeof(name), "%s ofld_txq%d", device_get_nameunit(vi->dev), i); init_eq(sc, &ofld_txq->eq, EQ_OFLD, vi->qsize_txq, pi->tx_chan, iqid, name); snprintf(name, sizeof(name), "%d", i); oid2 = SYSCTL_ADD_NODE(&vi->ctx, SYSCTL_CHILDREN(oid), OID_AUTO, name, CTLFLAG_RD, NULL, "offload tx queue"); rc = alloc_wrq(sc, vi, ofld_txq, oid2); if (rc != 0) goto done; j++; } #endif /* * Finally, the control queue. */ if (!IS_MAIN_VI(vi) || sc->flags & IS_VF) goto done; oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, "ctrlq", CTLFLAG_RD, NULL, "ctrl queue"); ctrlq = &sc->sge.ctrlq[pi->port_id]; iqid = vi_intr_iq(vi, 0)->cntxt_id; snprintf(name, sizeof(name), "%s ctrlq", device_get_nameunit(vi->dev)); init_eq(sc, &ctrlq->eq, EQ_CTRL, CTRL_EQ_QSIZE, pi->tx_chan, iqid, name); rc = alloc_wrq(sc, vi, ctrlq, oid); done: if (rc) t4_teardown_vi_queues(vi); return (rc); } /* * Idempotent */ int t4_teardown_vi_queues(struct vi_info *vi) { int i; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct sge_rxq *rxq; struct sge_txq *txq; #ifdef TCP_OFFLOAD struct sge_ofld_rxq *ofld_rxq; struct sge_wrq *ofld_txq; #endif #ifdef DEV_NETMAP struct sge_nm_rxq *nm_rxq; struct sge_nm_txq *nm_txq; #endif /* Do this before freeing the queues */ if (vi->flags & VI_SYSCTL_CTX) { sysctl_ctx_free(&vi->ctx); vi->flags &= ~VI_SYSCTL_CTX; } #ifdef DEV_NETMAP if (vi->ifp->if_capabilities & IFCAP_NETMAP) { for_each_nm_txq(vi, i, nm_txq) { free_nm_txq(vi, nm_txq); } for_each_nm_rxq(vi, i, nm_rxq) { free_nm_rxq(vi, nm_rxq); } } #endif /* * Take down all the tx queues first, as they reference the rx queues * (for egress updates, etc.). */ if (IS_MAIN_VI(vi) && !(sc->flags & IS_VF)) free_wrq(sc, &sc->sge.ctrlq[pi->port_id]); for_each_txq(vi, i, txq) { free_txq(vi, txq); } #ifdef TCP_OFFLOAD for_each_ofld_txq(vi, i, ofld_txq) { free_wrq(sc, ofld_txq); } #endif /* * Then take down the rx queues that forward their interrupts, as they * reference other rx queues. */ for_each_rxq(vi, i, rxq) { if ((rxq->iq.flags & IQ_INTR) == 0) free_rxq(vi, rxq); } #ifdef TCP_OFFLOAD for_each_ofld_rxq(vi, i, ofld_rxq) { if ((ofld_rxq->iq.flags & IQ_INTR) == 0) free_ofld_rxq(vi, ofld_rxq); } #endif /* * Then take down the rx queues that take direct interrupts. */ for_each_rxq(vi, i, rxq) { if (rxq->iq.flags & IQ_INTR) free_rxq(vi, rxq); } #ifdef TCP_OFFLOAD for_each_ofld_rxq(vi, i, ofld_rxq) { if (ofld_rxq->iq.flags & IQ_INTR) free_ofld_rxq(vi, ofld_rxq); } #endif return (0); } /* * Deals with errors and the firmware event queue. All data rx queues forward * their interrupt to the firmware event queue. */ void t4_intr_all(void *arg) { struct adapter *sc = arg; struct sge_iq *fwq = &sc->sge.fwq; t4_intr_err(arg); if (atomic_cmpset_int(&fwq->state, IQS_IDLE, IQS_BUSY)) { service_iq(fwq, 0); atomic_cmpset_int(&fwq->state, IQS_BUSY, IQS_IDLE); } } /* Deals with error interrupts */ void t4_intr_err(void *arg) { struct adapter *sc = arg; t4_write_reg(sc, MYPF_REG(A_PCIE_PF_CLI), 0); t4_slow_intr_handler(sc); } void t4_intr_evt(void *arg) { struct sge_iq *iq = arg; if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) { service_iq(iq, 0); atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE); } } void t4_intr(void *arg) { struct sge_iq *iq = arg; if (atomic_cmpset_int(&iq->state, IQS_IDLE, IQS_BUSY)) { service_iq(iq, 0); atomic_cmpset_int(&iq->state, IQS_BUSY, IQS_IDLE); } } void t4_vi_intr(void *arg) { struct irq *irq = arg; #ifdef DEV_NETMAP if (atomic_cmpset_int(&irq->nm_state, NM_ON, NM_BUSY)) { t4_nm_intr(irq->nm_rxq); atomic_cmpset_int(&irq->nm_state, NM_BUSY, NM_ON); } #endif if (irq->rxq != NULL) t4_intr(irq->rxq); } /* * Deals with anything and everything on the given ingress queue. */ static int service_iq(struct sge_iq *iq, int budget) { struct sge_iq *q; struct sge_rxq *rxq = iq_to_rxq(iq); /* Use iff iq is part of rxq */ struct sge_fl *fl; /* Use iff IQ_HAS_FL */ struct adapter *sc = iq->adapter; struct iq_desc *d = &iq->desc[iq->cidx]; int ndescs = 0, limit; int rsp_type, refill; uint32_t lq; uint16_t fl_hw_cidx; struct mbuf *m0; STAILQ_HEAD(, sge_iq) iql = STAILQ_HEAD_INITIALIZER(iql); #if defined(INET) || defined(INET6) const struct timeval lro_timeout = {0, sc->lro_timeout}; #endif KASSERT(iq->state == IQS_BUSY, ("%s: iq %p not BUSY", __func__, iq)); limit = budget ? budget : iq->qsize / 16; if (iq->flags & IQ_HAS_FL) { fl = &rxq->fl; fl_hw_cidx = fl->hw_cidx; /* stable snapshot */ } else { fl = NULL; fl_hw_cidx = 0; /* to silence gcc warning */ } /* * We always come back and check the descriptor ring for new indirect * interrupts and other responses after running a single handler. */ for (;;) { while ((d->rsp.u.type_gen & F_RSPD_GEN) == iq->gen) { rmb(); refill = 0; m0 = NULL; rsp_type = G_RSPD_TYPE(d->rsp.u.type_gen); lq = be32toh(d->rsp.pldbuflen_qid); switch (rsp_type) { case X_RSPD_TYPE_FLBUF: KASSERT(iq->flags & IQ_HAS_FL, ("%s: data for an iq (%p) with no freelist", __func__, iq)); m0 = get_fl_payload(sc, fl, lq); if (__predict_false(m0 == NULL)) goto process_iql; refill = IDXDIFF(fl->hw_cidx, fl_hw_cidx, fl->sidx) > 2; #ifdef T4_PKT_TIMESTAMP /* * 60 bit timestamp for the payload is * *(uint64_t *)m0->m_pktdat. Note that it is * in the leading free-space in the mbuf. The * kernel can clobber it during a pullup, * m_copymdata, etc. You need to make sure that * the mbuf reaches you unmolested if you care * about the timestamp. */ *(uint64_t *)m0->m_pktdat = be64toh(ctrl->u.last_flit) & 0xfffffffffffffff; #endif /* fall through */ case X_RSPD_TYPE_CPL: KASSERT(d->rss.opcode < NUM_CPL_CMDS, ("%s: bad opcode %02x.", __func__, d->rss.opcode)); t4_cpl_handler[d->rss.opcode](iq, &d->rss, m0); break; case X_RSPD_TYPE_INTR: /* * Interrupts should be forwarded only to queues * that are not forwarding their interrupts. * This means service_iq can recurse but only 1 * level deep. */ KASSERT(budget == 0, ("%s: budget %u, rsp_type %u", __func__, budget, rsp_type)); /* * There are 1K interrupt-capable queues (qids 0 * through 1023). A response type indicating a * forwarded interrupt with a qid >= 1K is an * iWARP async notification. */ if (lq >= 1024) { t4_an_handler(iq, &d->rsp); break; } q = sc->sge.iqmap[lq - sc->sge.iq_start - sc->sge.iq_base]; if (atomic_cmpset_int(&q->state, IQS_IDLE, IQS_BUSY)) { if (service_iq(q, q->qsize / 16) == 0) { atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE); } else { STAILQ_INSERT_TAIL(&iql, q, link); } } break; default: KASSERT(0, ("%s: illegal response type %d on iq %p", __func__, rsp_type, iq)); log(LOG_ERR, "%s: illegal response type %d on iq %p", device_get_nameunit(sc->dev), rsp_type, iq); break; } d++; if (__predict_false(++iq->cidx == iq->sidx)) { iq->cidx = 0; iq->gen ^= F_RSPD_GEN; d = &iq->desc[0]; } if (__predict_false(++ndescs == limit)) { t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) | V_INGRESSQID(iq->cntxt_id) | V_SEINTARM(V_QINTR_TIMER_IDX(X_TIMERREG_UPDATE_CIDX))); ndescs = 0; #if defined(INET) || defined(INET6) if (iq->flags & IQ_LRO_ENABLED && sc->lro_timeout != 0) { tcp_lro_flush_inactive(&rxq->lro, &lro_timeout); } #endif if (budget) { if (iq->flags & IQ_HAS_FL) { FL_LOCK(fl); refill_fl(sc, fl, 32); FL_UNLOCK(fl); } return (EINPROGRESS); } } if (refill) { FL_LOCK(fl); refill_fl(sc, fl, 32); FL_UNLOCK(fl); fl_hw_cidx = fl->hw_cidx; } } process_iql: if (STAILQ_EMPTY(&iql)) break; /* * Process the head only, and send it to the back of the list if * it's still not done. */ q = STAILQ_FIRST(&iql); STAILQ_REMOVE_HEAD(&iql, link); if (service_iq(q, q->qsize / 8) == 0) atomic_cmpset_int(&q->state, IQS_BUSY, IQS_IDLE); else STAILQ_INSERT_TAIL(&iql, q, link); } #if defined(INET) || defined(INET6) if (iq->flags & IQ_LRO_ENABLED) { struct lro_ctrl *lro = &rxq->lro; tcp_lro_flush_all(lro); } #endif t4_write_reg(sc, sc->sge_gts_reg, V_CIDXINC(ndescs) | V_INGRESSQID((u32)iq->cntxt_id) | V_SEINTARM(iq->intr_params)); if (iq->flags & IQ_HAS_FL) { int starved; FL_LOCK(fl); starved = refill_fl(sc, fl, 64); FL_UNLOCK(fl); if (__predict_false(starved != 0)) add_fl_to_sfl(sc, fl); } return (0); } static inline int cl_has_metadata(struct sge_fl *fl, struct cluster_layout *cll) { int rc = fl->flags & FL_BUF_PACKING || cll->region1 > 0; if (rc) MPASS(cll->region3 >= CL_METADATA_SIZE); return (rc); } static inline struct cluster_metadata * cl_metadata(struct adapter *sc, struct sge_fl *fl, struct cluster_layout *cll, caddr_t cl) { if (cl_has_metadata(fl, cll)) { struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx]; return ((struct cluster_metadata *)(cl + swz->size) - 1); } return (NULL); } static void rxb_free(struct mbuf *m, void *arg1, void *arg2) { uma_zone_t zone = arg1; caddr_t cl = arg2; uma_zfree(zone, cl); counter_u64_add(extfree_rels, 1); } /* * The mbuf returned by this function could be allocated from zone_mbuf or * constructed in spare room in the cluster. * * The mbuf carries the payload in one of these ways * a) frame inside the mbuf (mbuf from zone_mbuf) * b) m_cljset (for clusters without metadata) zone_mbuf * c) m_extaddref (cluster with metadata) inline mbuf * d) m_extaddref (cluster with metadata) zone_mbuf */ static struct mbuf * get_scatter_segment(struct adapter *sc, struct sge_fl *fl, int fr_offset, int remaining) { struct mbuf *m; struct fl_sdesc *sd = &fl->sdesc[fl->cidx]; struct cluster_layout *cll = &sd->cll; struct sw_zone_info *swz = &sc->sge.sw_zone_info[cll->zidx]; struct hw_buf_info *hwb = &sc->sge.hw_buf_info[cll->hwidx]; struct cluster_metadata *clm = cl_metadata(sc, fl, cll, sd->cl); int len, blen; caddr_t payload; blen = hwb->size - fl->rx_offset; /* max possible in this buf */ len = min(remaining, blen); payload = sd->cl + cll->region1 + fl->rx_offset; if (fl->flags & FL_BUF_PACKING) { const u_int l = fr_offset + len; const u_int pad = roundup2(l, fl->buf_boundary) - l; if (fl->rx_offset + len + pad < hwb->size) blen = len + pad; MPASS(fl->rx_offset + blen <= hwb->size); } else { MPASS(fl->rx_offset == 0); /* not packing */ } if (sc->sc_do_rxcopy && len < RX_COPY_THRESHOLD) { /* * Copy payload into a freshly allocated mbuf. */ m = fr_offset == 0 ? m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA); if (m == NULL) return (NULL); fl->mbuf_allocated++; #ifdef T4_PKT_TIMESTAMP /* Leave room for a timestamp */ m->m_data += 8; #endif /* copy data to mbuf */ bcopy(payload, mtod(m, caddr_t), len); } else if (sd->nmbuf * MSIZE < cll->region1) { /* * There's spare room in the cluster for an mbuf. Create one * and associate it with the payload that's in the cluster. */ MPASS(clm != NULL); m = (struct mbuf *)(sd->cl + sd->nmbuf * MSIZE); /* No bzero required */ if (m_init(m, M_NOWAIT, MT_DATA, fr_offset == 0 ? M_PKTHDR | M_NOFREE : M_NOFREE)) return (NULL); fl->mbuf_inlined++; m_extaddref(m, payload, blen, &clm->refcount, rxb_free, swz->zone, sd->cl); if (sd->nmbuf++ == 0) counter_u64_add(extfree_refs, 1); } else { /* * Grab an mbuf from zone_mbuf and associate it with the * payload in the cluster. */ m = fr_offset == 0 ? m_gethdr(M_NOWAIT, MT_DATA) : m_get(M_NOWAIT, MT_DATA); if (m == NULL) return (NULL); fl->mbuf_allocated++; if (clm != NULL) { m_extaddref(m, payload, blen, &clm->refcount, rxb_free, swz->zone, sd->cl); if (sd->nmbuf++ == 0) counter_u64_add(extfree_refs, 1); } else { m_cljset(m, sd->cl, swz->type); sd->cl = NULL; /* consumed, not a recycle candidate */ } } if (fr_offset == 0) m->m_pkthdr.len = remaining; m->m_len = len; if (fl->flags & FL_BUF_PACKING) { fl->rx_offset += blen; MPASS(fl->rx_offset <= hwb->size); if (fl->rx_offset < hwb->size) return (m); /* without advancing the cidx */ } if (__predict_false(++fl->cidx % 8 == 0)) { uint16_t cidx = fl->cidx / 8; if (__predict_false(cidx == fl->sidx)) fl->cidx = cidx = 0; fl->hw_cidx = cidx; } fl->rx_offset = 0; return (m); } static struct mbuf * get_fl_payload(struct adapter *sc, struct sge_fl *fl, uint32_t len_newbuf) { struct mbuf *m0, *m, **pnext; u_int remaining; const u_int total = G_RSPD_LEN(len_newbuf); if (__predict_false(fl->flags & FL_BUF_RESUME)) { M_ASSERTPKTHDR(fl->m0); MPASS(fl->m0->m_pkthdr.len == total); MPASS(fl->remaining < total); m0 = fl->m0; pnext = fl->pnext; remaining = fl->remaining; fl->flags &= ~FL_BUF_RESUME; goto get_segment; } if (fl->rx_offset > 0 && len_newbuf & F_RSPD_NEWBUF) { fl->rx_offset = 0; if (__predict_false(++fl->cidx % 8 == 0)) { uint16_t cidx = fl->cidx / 8; if (__predict_false(cidx == fl->sidx)) fl->cidx = cidx = 0; fl->hw_cidx = cidx; } } /* * Payload starts at rx_offset in the current hw buffer. Its length is * 'len' and it may span multiple hw buffers. */ m0 = get_scatter_segment(sc, fl, 0, total); if (m0 == NULL) return (NULL); remaining = total - m0->m_len; pnext = &m0->m_next; while (remaining > 0) { get_segment: MPASS(fl->rx_offset == 0); m = get_scatter_segment(sc, fl, total - remaining, remaining); if (__predict_false(m == NULL)) { fl->m0 = m0; fl->pnext = pnext; fl->remaining = remaining; fl->flags |= FL_BUF_RESUME; return (NULL); } *pnext = m; pnext = &m->m_next; remaining -= m->m_len; } *pnext = NULL; M_ASSERTPKTHDR(m0); return (m0); } static int t4_eth_rx(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m0) { struct sge_rxq *rxq = iq_to_rxq(iq); struct ifnet *ifp = rxq->ifp; struct adapter *sc = iq->adapter; const struct cpl_rx_pkt *cpl = (const void *)(rss + 1); #if defined(INET) || defined(INET6) struct lro_ctrl *lro = &rxq->lro; #endif static const int sw_hashtype[4][2] = { {M_HASHTYPE_NONE, M_HASHTYPE_NONE}, {M_HASHTYPE_RSS_IPV4, M_HASHTYPE_RSS_IPV6}, {M_HASHTYPE_RSS_TCP_IPV4, M_HASHTYPE_RSS_TCP_IPV6}, {M_HASHTYPE_RSS_UDP_IPV4, M_HASHTYPE_RSS_UDP_IPV6}, }; KASSERT(m0 != NULL, ("%s: no payload with opcode %02x", __func__, rss->opcode)); m0->m_pkthdr.len -= sc->params.sge.fl_pktshift; m0->m_len -= sc->params.sge.fl_pktshift; m0->m_data += sc->params.sge.fl_pktshift; m0->m_pkthdr.rcvif = ifp; M_HASHTYPE_SET(m0, sw_hashtype[rss->hash_type][rss->ipv6]); m0->m_pkthdr.flowid = be32toh(rss->hash_val); if (cpl->csum_calc && !cpl->err_vec) { if (ifp->if_capenable & IFCAP_RXCSUM && cpl->l2info & htobe32(F_RXF_IP)) { m0->m_pkthdr.csum_flags = (CSUM_IP_CHECKED | CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR); rxq->rxcsum++; } else if (ifp->if_capenable & IFCAP_RXCSUM_IPV6 && cpl->l2info & htobe32(F_RXF_IP6)) { m0->m_pkthdr.csum_flags = (CSUM_DATA_VALID_IPV6 | CSUM_PSEUDO_HDR); rxq->rxcsum++; } if (__predict_false(cpl->ip_frag)) m0->m_pkthdr.csum_data = be16toh(cpl->csum); else m0->m_pkthdr.csum_data = 0xffff; } if (cpl->vlan_ex) { m0->m_pkthdr.ether_vtag = be16toh(cpl->vlan); m0->m_flags |= M_VLANTAG; rxq->vlan_extraction++; } #if defined(INET) || defined(INET6) if (iq->flags & IQ_LRO_ENABLED && tcp_lro_rx(lro, m0, 0) == 0) { /* queued for LRO */ } else #endif ifp->if_input(ifp, m0); return (0); } /* * Must drain the wrq or make sure that someone else will. */ static void wrq_tx_drain(void *arg, int n) { struct sge_wrq *wrq = arg; struct sge_eq *eq = &wrq->eq; EQ_LOCK(eq); if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list)) drain_wrq_wr_list(wrq->adapter, wrq); EQ_UNLOCK(eq); } static void drain_wrq_wr_list(struct adapter *sc, struct sge_wrq *wrq) { struct sge_eq *eq = &wrq->eq; u_int available, dbdiff; /* # of hardware descriptors */ u_int n; struct wrqe *wr; struct fw_eth_tx_pkt_wr *dst; /* any fw WR struct will do */ EQ_LOCK_ASSERT_OWNED(eq); MPASS(TAILQ_EMPTY(&wrq->incomplete_wrs)); wr = STAILQ_FIRST(&wrq->wr_list); MPASS(wr != NULL); /* Must be called with something useful to do */ MPASS(eq->pidx == eq->dbidx); dbdiff = 0; do { eq->cidx = read_hw_cidx(eq); if (eq->pidx == eq->cidx) available = eq->sidx - 1; else available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; MPASS(wr->wrq == wrq); n = howmany(wr->wr_len, EQ_ESIZE); if (available < n) break; dst = (void *)&eq->desc[eq->pidx]; if (__predict_true(eq->sidx - eq->pidx > n)) { /* Won't wrap, won't end exactly at the status page. */ bcopy(&wr->wr[0], dst, wr->wr_len); eq->pidx += n; } else { int first_portion = (eq->sidx - eq->pidx) * EQ_ESIZE; bcopy(&wr->wr[0], dst, first_portion); if (wr->wr_len > first_portion) { bcopy(&wr->wr[first_portion], &eq->desc[0], wr->wr_len - first_portion); } eq->pidx = n - (eq->sidx - eq->pidx); } if (available < eq->sidx / 4 && atomic_cmpset_int(&eq->equiq, 0, 1)) { dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ | F_FW_WR_EQUEQ); eq->equeqidx = eq->pidx; } else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) { dst->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ); eq->equeqidx = eq->pidx; } dbdiff += n; if (dbdiff >= 16) { ring_eq_db(sc, eq, dbdiff); dbdiff = 0; } STAILQ_REMOVE_HEAD(&wrq->wr_list, link); free_wrqe(wr); MPASS(wrq->nwr_pending > 0); wrq->nwr_pending--; MPASS(wrq->ndesc_needed >= n); wrq->ndesc_needed -= n; } while ((wr = STAILQ_FIRST(&wrq->wr_list)) != NULL); if (dbdiff) ring_eq_db(sc, eq, dbdiff); } /* * Doesn't fail. Holds on to work requests it can't send right away. */ void t4_wrq_tx_locked(struct adapter *sc, struct sge_wrq *wrq, struct wrqe *wr) { #ifdef INVARIANTS struct sge_eq *eq = &wrq->eq; #endif EQ_LOCK_ASSERT_OWNED(eq); MPASS(wr != NULL); MPASS(wr->wr_len > 0 && wr->wr_len <= SGE_MAX_WR_LEN); MPASS((wr->wr_len & 0x7) == 0); STAILQ_INSERT_TAIL(&wrq->wr_list, wr, link); wrq->nwr_pending++; wrq->ndesc_needed += howmany(wr->wr_len, EQ_ESIZE); if (!TAILQ_EMPTY(&wrq->incomplete_wrs)) return; /* commit_wrq_wr will drain wr_list as well. */ drain_wrq_wr_list(sc, wrq); /* Doorbell must have caught up to the pidx. */ MPASS(eq->pidx == eq->dbidx); } void t4_update_fl_bufsize(struct ifnet *ifp) { struct vi_info *vi = ifp->if_softc; struct adapter *sc = vi->pi->adapter; struct sge_rxq *rxq; #ifdef TCP_OFFLOAD struct sge_ofld_rxq *ofld_rxq; #endif struct sge_fl *fl; int i, maxp, mtu = ifp->if_mtu; maxp = mtu_to_max_payload(sc, mtu, 0); for_each_rxq(vi, i, rxq) { fl = &rxq->fl; FL_LOCK(fl); find_best_refill_source(sc, fl, maxp); FL_UNLOCK(fl); } #ifdef TCP_OFFLOAD maxp = mtu_to_max_payload(sc, mtu, 1); for_each_ofld_rxq(vi, i, ofld_rxq) { fl = &ofld_rxq->fl; FL_LOCK(fl); find_best_refill_source(sc, fl, maxp); FL_UNLOCK(fl); } #endif } static inline int mbuf_nsegs(struct mbuf *m) { M_ASSERTPKTHDR(m); KASSERT(m->m_pkthdr.l5hlen > 0, ("%s: mbuf %p missing information on # of segments.", __func__, m)); return (m->m_pkthdr.l5hlen); } static inline void set_mbuf_nsegs(struct mbuf *m, uint8_t nsegs) { M_ASSERTPKTHDR(m); m->m_pkthdr.l5hlen = nsegs; } static inline int mbuf_len16(struct mbuf *m) { int n; M_ASSERTPKTHDR(m); n = m->m_pkthdr.PH_loc.eight[0]; MPASS(n > 0 && n <= SGE_MAX_WR_LEN / 16); return (n); } static inline void set_mbuf_len16(struct mbuf *m, uint8_t len16) { M_ASSERTPKTHDR(m); m->m_pkthdr.PH_loc.eight[0] = len16; } static inline int needs_tso(struct mbuf *m) { M_ASSERTPKTHDR(m); if (m->m_pkthdr.csum_flags & CSUM_TSO) { KASSERT(m->m_pkthdr.tso_segsz > 0, ("%s: TSO requested in mbuf %p but MSS not provided", __func__, m)); return (1); } return (0); } static inline int needs_l3_csum(struct mbuf *m) { M_ASSERTPKTHDR(m); if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TSO)) return (1); return (0); } static inline int needs_l4_csum(struct mbuf *m) { M_ASSERTPKTHDR(m); if (m->m_pkthdr.csum_flags & (CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) return (1); return (0); } static inline int needs_vlan_insertion(struct mbuf *m) { M_ASSERTPKTHDR(m); if (m->m_flags & M_VLANTAG) { KASSERT(m->m_pkthdr.ether_vtag != 0, ("%s: HWVLAN requested in mbuf %p but tag not provided", __func__, m)); return (1); } return (0); } static void * m_advance(struct mbuf **pm, int *poffset, int len) { struct mbuf *m = *pm; int offset = *poffset; uintptr_t p = 0; MPASS(len > 0); for (;;) { if (offset + len < m->m_len) { offset += len; p = mtod(m, uintptr_t) + offset; break; } len -= m->m_len - offset; m = m->m_next; offset = 0; MPASS(m != NULL); } *poffset = offset; *pm = m; return ((void *)p); } static inline int same_paddr(char *a, char *b) { if (a == b) return (1); else if (a != NULL && b != NULL) { vm_offset_t x = (vm_offset_t)a; vm_offset_t y = (vm_offset_t)b; if ((x & PAGE_MASK) == (y & PAGE_MASK) && pmap_kextract(x) == pmap_kextract(y)) return (1); } return (0); } /* * Can deal with empty mbufs in the chain that have m_len = 0, but the chain * must have at least one mbuf that's not empty. */ static inline int count_mbuf_nsegs(struct mbuf *m) { char *prev_end, *start; int len, nsegs; MPASS(m != NULL); nsegs = 0; prev_end = NULL; for (; m; m = m->m_next) { len = m->m_len; if (__predict_false(len == 0)) continue; start = mtod(m, char *); nsegs += sglist_count(start, len); if (same_paddr(prev_end, start)) nsegs--; prev_end = start + len; } MPASS(nsegs > 0); return (nsegs); } /* * Analyze the mbuf to determine its tx needs. The mbuf passed in may change: * a) caller can assume it's been freed if this function returns with an error. * b) it may get defragged up if the gather list is too long for the hardware. */ int parse_pkt(struct adapter *sc, struct mbuf **mp) { struct mbuf *m0 = *mp, *m; int rc, nsegs, defragged = 0, offset; struct ether_header *eh; void *l3hdr; #if defined(INET) || defined(INET6) struct tcphdr *tcp; #endif uint16_t eh_type; M_ASSERTPKTHDR(m0); if (__predict_false(m0->m_pkthdr.len < ETHER_HDR_LEN)) { rc = EINVAL; fail: m_freem(m0); *mp = NULL; return (rc); } restart: /* * First count the number of gather list segments in the payload. * Defrag the mbuf if nsegs exceeds the hardware limit. */ M_ASSERTPKTHDR(m0); MPASS(m0->m_pkthdr.len > 0); nsegs = count_mbuf_nsegs(m0); if (nsegs > (needs_tso(m0) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)) { if (defragged++ > 0 || (m = m_defrag(m0, M_NOWAIT)) == NULL) { rc = EFBIG; goto fail; } *mp = m0 = m; /* update caller's copy after defrag */ goto restart; } if (__predict_false(nsegs > 2 && m0->m_pkthdr.len <= MHLEN)) { m0 = m_pullup(m0, m0->m_pkthdr.len); if (m0 == NULL) { /* Should have left well enough alone. */ rc = EFBIG; goto fail; } *mp = m0; /* update caller's copy after pullup */ goto restart; } set_mbuf_nsegs(m0, nsegs); if (sc->flags & IS_VF) set_mbuf_len16(m0, txpkt_vm_len16(nsegs, needs_tso(m0))); else set_mbuf_len16(m0, txpkt_len16(nsegs, needs_tso(m0))); if (!needs_tso(m0) && !(sc->flags & IS_VF && (needs_l3_csum(m0) || needs_l4_csum(m0)))) return (0); m = m0; eh = mtod(m, struct ether_header *); eh_type = ntohs(eh->ether_type); if (eh_type == ETHERTYPE_VLAN) { struct ether_vlan_header *evh = (void *)eh; eh_type = ntohs(evh->evl_proto); m0->m_pkthdr.l2hlen = sizeof(*evh); } else m0->m_pkthdr.l2hlen = sizeof(*eh); offset = 0; l3hdr = m_advance(&m, &offset, m0->m_pkthdr.l2hlen); switch (eh_type) { #ifdef INET6 case ETHERTYPE_IPV6: { struct ip6_hdr *ip6 = l3hdr; MPASS(!needs_tso(m0) || ip6->ip6_nxt == IPPROTO_TCP); m0->m_pkthdr.l3hlen = sizeof(*ip6); break; } #endif #ifdef INET case ETHERTYPE_IP: { struct ip *ip = l3hdr; m0->m_pkthdr.l3hlen = ip->ip_hl * 4; break; } #endif default: panic("%s: ethertype 0x%04x unknown. if_cxgbe must be compiled" " with the same INET/INET6 options as the kernel.", __func__, eh_type); } #if defined(INET) || defined(INET6) if (needs_tso(m0)) { tcp = m_advance(&m, &offset, m0->m_pkthdr.l3hlen); m0->m_pkthdr.l4hlen = tcp->th_off * 4; } #endif MPASS(m0 == *mp); return (0); } void * start_wrq_wr(struct sge_wrq *wrq, int len16, struct wrq_cookie *cookie) { struct sge_eq *eq = &wrq->eq; struct adapter *sc = wrq->adapter; int ndesc, available; struct wrqe *wr; void *w; MPASS(len16 > 0); ndesc = howmany(len16, EQ_ESIZE / 16); MPASS(ndesc > 0 && ndesc <= SGE_MAX_WR_NDESC); EQ_LOCK(eq); if (!STAILQ_EMPTY(&wrq->wr_list)) drain_wrq_wr_list(sc, wrq); if (!STAILQ_EMPTY(&wrq->wr_list)) { slowpath: EQ_UNLOCK(eq); wr = alloc_wrqe(len16 * 16, wrq); if (__predict_false(wr == NULL)) return (NULL); cookie->pidx = -1; cookie->ndesc = ndesc; return (&wr->wr); } eq->cidx = read_hw_cidx(eq); if (eq->pidx == eq->cidx) available = eq->sidx - 1; else available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; if (available < ndesc) goto slowpath; cookie->pidx = eq->pidx; cookie->ndesc = ndesc; TAILQ_INSERT_TAIL(&wrq->incomplete_wrs, cookie, link); w = &eq->desc[eq->pidx]; IDXINCR(eq->pidx, ndesc, eq->sidx); if (__predict_false(eq->pidx < ndesc - 1)) { w = &wrq->ss[0]; wrq->ss_pidx = cookie->pidx; wrq->ss_len = len16 * 16; } EQ_UNLOCK(eq); return (w); } void commit_wrq_wr(struct sge_wrq *wrq, void *w, struct wrq_cookie *cookie) { struct sge_eq *eq = &wrq->eq; struct adapter *sc = wrq->adapter; int ndesc, pidx; struct wrq_cookie *prev, *next; if (cookie->pidx == -1) { struct wrqe *wr = __containerof(w, struct wrqe, wr); t4_wrq_tx(sc, wr); return; } ndesc = cookie->ndesc; /* Can be more than SGE_MAX_WR_NDESC here. */ pidx = cookie->pidx; MPASS(pidx >= 0 && pidx < eq->sidx); if (__predict_false(w == &wrq->ss[0])) { int n = (eq->sidx - wrq->ss_pidx) * EQ_ESIZE; MPASS(wrq->ss_len > n); /* WR had better wrap around. */ bcopy(&wrq->ss[0], &eq->desc[wrq->ss_pidx], n); bcopy(&wrq->ss[n], &eq->desc[0], wrq->ss_len - n); wrq->tx_wrs_ss++; } else wrq->tx_wrs_direct++; EQ_LOCK(eq); prev = TAILQ_PREV(cookie, wrq_incomplete_wrs, link); next = TAILQ_NEXT(cookie, link); if (prev == NULL) { MPASS(pidx == eq->dbidx); if (next == NULL || ndesc >= 16) ring_eq_db(wrq->adapter, eq, ndesc); else { MPASS(IDXDIFF(next->pidx, pidx, eq->sidx) == ndesc); next->pidx = pidx; next->ndesc += ndesc; } } else { MPASS(IDXDIFF(pidx, prev->pidx, eq->sidx) == prev->ndesc); prev->ndesc += ndesc; } TAILQ_REMOVE(&wrq->incomplete_wrs, cookie, link); if (TAILQ_EMPTY(&wrq->incomplete_wrs) && !STAILQ_EMPTY(&wrq->wr_list)) drain_wrq_wr_list(sc, wrq); #ifdef INVARIANTS if (TAILQ_EMPTY(&wrq->incomplete_wrs)) { /* Doorbell must have caught up to the pidx. */ MPASS(wrq->eq.pidx == wrq->eq.dbidx); } #endif EQ_UNLOCK(eq); } static u_int can_resume_eth_tx(struct mp_ring *r) { struct sge_eq *eq = r->cookie; return (total_available_tx_desc(eq) > eq->sidx / 8); } static inline int cannot_use_txpkts(struct mbuf *m) { /* maybe put a GL limit too, to avoid silliness? */ return (needs_tso(m)); } /* * r->items[cidx] to r->items[pidx], with a wraparound at r->size, are ready to * be consumed. Return the actual number consumed. 0 indicates a stall. */ static u_int eth_tx(struct mp_ring *r, u_int cidx, u_int pidx) { struct sge_txq *txq = r->cookie; struct sge_eq *eq = &txq->eq; struct ifnet *ifp = txq->ifp; struct vi_info *vi = ifp->if_softc; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; u_int total, remaining; /* # of packets */ u_int available, dbdiff; /* # of hardware descriptors */ u_int n, next_cidx; struct mbuf *m0, *tail; struct txpkts txp; struct fw_eth_tx_pkts_wr *wr; /* any fw WR struct will do */ remaining = IDXDIFF(pidx, cidx, r->size); MPASS(remaining > 0); /* Must not be called without work to do. */ total = 0; TXQ_LOCK(txq); if (__predict_false((eq->flags & EQ_ENABLED) == 0)) { while (cidx != pidx) { m0 = r->items[cidx]; m_freem(m0); if (++cidx == r->size) cidx = 0; } reclaim_tx_descs(txq, 2048); total = remaining; goto done; } /* How many hardware descriptors do we have readily available. */ if (eq->pidx == eq->cidx) available = eq->sidx - 1; else available = IDXDIFF(eq->cidx, eq->pidx, eq->sidx) - 1; dbdiff = IDXDIFF(eq->pidx, eq->dbidx, eq->sidx); while (remaining > 0) { m0 = r->items[cidx]; M_ASSERTPKTHDR(m0); MPASS(m0->m_nextpkt == NULL); if (available < SGE_MAX_WR_NDESC) { available += reclaim_tx_descs(txq, 64); if (available < howmany(mbuf_len16(m0), EQ_ESIZE / 16)) break; /* out of descriptors */ } next_cidx = cidx + 1; if (__predict_false(next_cidx == r->size)) next_cidx = 0; wr = (void *)&eq->desc[eq->pidx]; if (sc->flags & IS_VF) { total++; remaining--; ETHER_BPF_MTAP(ifp, m0); - n = write_txpkt_vm_wr(txq, (void *)wr, m0, available); + n = write_txpkt_vm_wr(sc, txq, (void *)wr, m0, + available); } else if (remaining > 1 && try_txpkts(m0, r->items[next_cidx], &txp, available) == 0) { /* pkts at cidx, next_cidx should both be in txp. */ MPASS(txp.npkt == 2); tail = r->items[next_cidx]; MPASS(tail->m_nextpkt == NULL); ETHER_BPF_MTAP(ifp, m0); ETHER_BPF_MTAP(ifp, tail); m0->m_nextpkt = tail; if (__predict_false(++next_cidx == r->size)) next_cidx = 0; while (next_cidx != pidx) { if (add_to_txpkts(r->items[next_cidx], &txp, available) != 0) break; tail->m_nextpkt = r->items[next_cidx]; tail = tail->m_nextpkt; ETHER_BPF_MTAP(ifp, tail); if (__predict_false(++next_cidx == r->size)) next_cidx = 0; } n = write_txpkts_wr(txq, wr, m0, &txp, available); total += txp.npkt; remaining -= txp.npkt; } else { total++; remaining--; ETHER_BPF_MTAP(ifp, m0); n = write_txpkt_wr(txq, (void *)wr, m0, available); } MPASS(n >= 1 && n <= available && n <= SGE_MAX_WR_NDESC); available -= n; dbdiff += n; IDXINCR(eq->pidx, n, eq->sidx); if (total_available_tx_desc(eq) < eq->sidx / 4 && atomic_cmpset_int(&eq->equiq, 0, 1)) { wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUIQ | F_FW_WR_EQUEQ); eq->equeqidx = eq->pidx; } else if (IDXDIFF(eq->pidx, eq->equeqidx, eq->sidx) >= 32) { wr->equiq_to_len16 |= htobe32(F_FW_WR_EQUEQ); eq->equeqidx = eq->pidx; } if (dbdiff >= 16 && remaining >= 4) { ring_eq_db(sc, eq, dbdiff); available += reclaim_tx_descs(txq, 4 * dbdiff); dbdiff = 0; } cidx = next_cidx; } if (dbdiff != 0) { ring_eq_db(sc, eq, dbdiff); reclaim_tx_descs(txq, 32); } done: TXQ_UNLOCK(txq); return (total); } static inline void init_iq(struct sge_iq *iq, struct adapter *sc, int tmr_idx, int pktc_idx, int qsize) { KASSERT(tmr_idx >= 0 && tmr_idx < SGE_NTIMERS, ("%s: bad tmr_idx %d", __func__, tmr_idx)); KASSERT(pktc_idx < SGE_NCOUNTERS, /* -ve is ok, means don't use */ ("%s: bad pktc_idx %d", __func__, pktc_idx)); iq->flags = 0; iq->adapter = sc; iq->intr_params = V_QINTR_TIMER_IDX(tmr_idx); iq->intr_pktc_idx = SGE_NCOUNTERS - 1; if (pktc_idx >= 0) { iq->intr_params |= F_QINTR_CNT_EN; iq->intr_pktc_idx = pktc_idx; } iq->qsize = roundup2(qsize, 16); /* See FW_IQ_CMD/iqsize */ iq->sidx = iq->qsize - sc->params.sge.spg_len / IQ_ESIZE; } static inline void init_fl(struct adapter *sc, struct sge_fl *fl, int qsize, int maxp, char *name) { fl->qsize = qsize; fl->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE; strlcpy(fl->lockname, name, sizeof(fl->lockname)); if (sc->flags & BUF_PACKING_OK && ((!is_t4(sc) && buffer_packing) || /* T5+: enabled unless 0 */ (is_t4(sc) && buffer_packing == 1)))/* T4: disabled unless 1 */ fl->flags |= FL_BUF_PACKING; find_best_refill_source(sc, fl, maxp); find_safe_refill_source(sc, fl); } static inline void init_eq(struct adapter *sc, struct sge_eq *eq, int eqtype, int qsize, uint8_t tx_chan, uint16_t iqid, char *name) { KASSERT(eqtype <= EQ_TYPEMASK, ("%s: bad qtype %d", __func__, eqtype)); eq->flags = eqtype & EQ_TYPEMASK; eq->tx_chan = tx_chan; eq->iqid = iqid; eq->sidx = qsize - sc->params.sge.spg_len / EQ_ESIZE; strlcpy(eq->lockname, name, sizeof(eq->lockname)); } static int alloc_ring(struct adapter *sc, size_t len, bus_dma_tag_t *tag, bus_dmamap_t *map, bus_addr_t *pa, void **va) { int rc; rc = bus_dma_tag_create(sc->dmat, 512, 0, BUS_SPACE_MAXADDR, BUS_SPACE_MAXADDR, NULL, NULL, len, 1, len, 0, NULL, NULL, tag); if (rc != 0) { device_printf(sc->dev, "cannot allocate DMA tag: %d\n", rc); goto done; } rc = bus_dmamem_alloc(*tag, va, BUS_DMA_WAITOK | BUS_DMA_COHERENT | BUS_DMA_ZERO, map); if (rc != 0) { device_printf(sc->dev, "cannot allocate DMA memory: %d\n", rc); goto done; } rc = bus_dmamap_load(*tag, *map, *va, len, oneseg_dma_callback, pa, 0); if (rc != 0) { device_printf(sc->dev, "cannot load DMA map: %d\n", rc); goto done; } done: if (rc) free_ring(sc, *tag, *map, *pa, *va); return (rc); } static int free_ring(struct adapter *sc, bus_dma_tag_t tag, bus_dmamap_t map, bus_addr_t pa, void *va) { if (pa) bus_dmamap_unload(tag, map); if (va) bus_dmamem_free(tag, va, map); if (tag) bus_dma_tag_destroy(tag); return (0); } /* * Allocates the ring for an ingress queue and an optional freelist. If the * freelist is specified it will be allocated and then associated with the * ingress queue. * * Returns errno on failure. Resources allocated up to that point may still be * allocated. Caller is responsible for cleanup in case this function fails. * * If the ingress queue will take interrupts directly (iq->flags & IQ_INTR) then * the intr_idx specifies the vector, starting from 0. Otherwise it specifies * the abs_id of the ingress queue to which its interrupts should be forwarded. */ static int alloc_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl, int intr_idx, int cong) { int rc, i, cntxt_id; size_t len; struct fw_iq_cmd c; struct port_info *pi = vi->pi; struct adapter *sc = iq->adapter; struct sge_params *sp = &sc->params.sge; __be32 v = 0; len = iq->qsize * IQ_ESIZE; rc = alloc_ring(sc, len, &iq->desc_tag, &iq->desc_map, &iq->ba, (void **)&iq->desc); if (rc != 0) return (rc); bzero(&c, sizeof(c)); c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_IQ_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_IQ_CMD_PFN(sc->pf) | V_FW_IQ_CMD_VFN(0)); c.alloc_to_len16 = htobe32(F_FW_IQ_CMD_ALLOC | F_FW_IQ_CMD_IQSTART | FW_LEN16(c)); /* Special handling for firmware event queue */ if (iq == &sc->sge.fwq) v |= F_FW_IQ_CMD_IQASYNCH; if (iq->flags & IQ_INTR) { KASSERT(intr_idx < sc->intr_count, ("%s: invalid direct intr_idx %d", __func__, intr_idx)); } else v |= F_FW_IQ_CMD_IQANDST; v |= V_FW_IQ_CMD_IQANDSTINDEX(intr_idx); c.type_to_iqandstindex = htobe32(v | V_FW_IQ_CMD_TYPE(FW_IQ_TYPE_FL_INT_CAP) | V_FW_IQ_CMD_VIID(vi->viid) | V_FW_IQ_CMD_IQANUD(X_UPDATEDELIVERY_INTERRUPT)); c.iqdroprss_to_iqesize = htobe16(V_FW_IQ_CMD_IQPCIECH(pi->tx_chan) | F_FW_IQ_CMD_IQGTSMODE | V_FW_IQ_CMD_IQINTCNTTHRESH(iq->intr_pktc_idx) | V_FW_IQ_CMD_IQESIZE(ilog2(IQ_ESIZE) - 4)); c.iqsize = htobe16(iq->qsize); c.iqaddr = htobe64(iq->ba); if (cong >= 0) c.iqns_to_fl0congen = htobe32(F_FW_IQ_CMD_IQFLINTCONGEN); if (fl) { mtx_init(&fl->fl_lock, fl->lockname, NULL, MTX_DEF); len = fl->qsize * EQ_ESIZE; rc = alloc_ring(sc, len, &fl->desc_tag, &fl->desc_map, &fl->ba, (void **)&fl->desc); if (rc) return (rc); /* Allocate space for one software descriptor per buffer. */ rc = alloc_fl_sdesc(fl); if (rc != 0) { device_printf(sc->dev, "failed to setup fl software descriptors: %d\n", rc); return (rc); } if (fl->flags & FL_BUF_PACKING) { fl->lowat = roundup2(sp->fl_starve_threshold2, 8); fl->buf_boundary = sp->pack_boundary; } else { fl->lowat = roundup2(sp->fl_starve_threshold, 8); fl->buf_boundary = 16; } if (fl_pad && fl->buf_boundary < sp->pad_boundary) fl->buf_boundary = sp->pad_boundary; c.iqns_to_fl0congen |= htobe32(V_FW_IQ_CMD_FL0HOSTFCMODE(X_HOSTFCMODE_NONE) | F_FW_IQ_CMD_FL0FETCHRO | F_FW_IQ_CMD_FL0DATARO | (fl_pad ? F_FW_IQ_CMD_FL0PADEN : 0) | (fl->flags & FL_BUF_PACKING ? F_FW_IQ_CMD_FL0PACKEN : 0)); if (cong >= 0) { c.iqns_to_fl0congen |= htobe32(V_FW_IQ_CMD_FL0CNGCHMAP(cong) | F_FW_IQ_CMD_FL0CONGCIF | F_FW_IQ_CMD_FL0CONGEN); } c.fl0dcaen_to_fl0cidxfthresh = - htobe16(V_FW_IQ_CMD_FL0FBMIN(X_FETCHBURSTMIN_128B) | - V_FW_IQ_CMD_FL0FBMAX(X_FETCHBURSTMAX_512B)); + htobe16(V_FW_IQ_CMD_FL0FBMIN(chip_id(sc) <= CHELSIO_T5 ? + X_FETCHBURSTMIN_128B : X_FETCHBURSTMIN_64B) | + V_FW_IQ_CMD_FL0FBMAX(chip_id(sc) <= CHELSIO_T5 ? + X_FETCHBURSTMAX_512B : X_FETCHBURSTMAX_256B)); c.fl0size = htobe16(fl->qsize); c.fl0addr = htobe64(fl->ba); } rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); if (rc != 0) { device_printf(sc->dev, "failed to create ingress queue: %d\n", rc); return (rc); } iq->cidx = 0; iq->gen = F_RSPD_GEN; iq->intr_next = iq->intr_params; iq->cntxt_id = be16toh(c.iqid); iq->abs_id = be16toh(c.physiqid); iq->flags |= IQ_ALLOCATED; cntxt_id = iq->cntxt_id - sc->sge.iq_start; if (cntxt_id >= sc->sge.niq) { panic ("%s: iq->cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.niq - 1); } sc->sge.iqmap[cntxt_id] = iq; if (fl) { u_int qid; iq->flags |= IQ_HAS_FL; fl->cntxt_id = be16toh(c.fl0id); fl->pidx = fl->cidx = 0; cntxt_id = fl->cntxt_id - sc->sge.eq_start; if (cntxt_id >= sc->sge.neq) { panic("%s: fl->cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.neq - 1); } sc->sge.eqmap[cntxt_id] = (void *)fl; qid = fl->cntxt_id; if (isset(&sc->doorbells, DOORBELL_UDB)) { uint32_t s_qpp = sc->params.sge.eq_s_qpp; uint32_t mask = (1 << s_qpp) - 1; volatile uint8_t *udb; udb = sc->udbs_base + UDBS_DB_OFFSET; udb += (qid >> s_qpp) << PAGE_SHIFT; qid &= mask; if (qid < PAGE_SIZE / UDBS_SEG_SIZE) { udb += qid << UDBS_SEG_SHIFT; qid = 0; } fl->udb = (volatile void *)udb; } fl->dbval = V_QID(qid) | sc->chip_params->sge_fl_db; FL_LOCK(fl); /* Enough to make sure the SGE doesn't think it's starved */ refill_fl(sc, fl, fl->lowat); FL_UNLOCK(fl); } if (is_t5(sc) && !(sc->flags & IS_VF) && cong >= 0) { uint32_t param, val; param = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_CONM_CTXT) | V_FW_PARAMS_PARAM_YZ(iq->cntxt_id); if (cong == 0) val = 1 << 19; else { val = 2 << 19; for (i = 0; i < 4; i++) { if (cong & (1 << i)) val |= 1 << (i << 2); } } rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, ¶m, &val); if (rc != 0) { /* report error but carry on */ device_printf(sc->dev, "failed to set congestion manager context for " "ingress queue %d: %d\n", iq->cntxt_id, rc); } } /* Enable IQ interrupts */ atomic_store_rel_int(&iq->state, IQS_IDLE); t4_write_reg(sc, sc->sge_gts_reg, V_SEINTARM(iq->intr_params) | V_INGRESSQID(iq->cntxt_id)); return (0); } static int free_iq_fl(struct vi_info *vi, struct sge_iq *iq, struct sge_fl *fl) { int rc; struct adapter *sc = iq->adapter; device_t dev; if (sc == NULL) return (0); /* nothing to do */ dev = vi ? vi->dev : sc->dev; if (iq->flags & IQ_ALLOCATED) { rc = -t4_iq_free(sc, sc->mbox, sc->pf, 0, FW_IQ_TYPE_FL_INT_CAP, iq->cntxt_id, fl ? fl->cntxt_id : 0xffff, 0xffff); if (rc != 0) { device_printf(dev, "failed to free queue %p: %d\n", iq, rc); return (rc); } iq->flags &= ~IQ_ALLOCATED; } free_ring(sc, iq->desc_tag, iq->desc_map, iq->ba, iq->desc); bzero(iq, sizeof(*iq)); if (fl) { free_ring(sc, fl->desc_tag, fl->desc_map, fl->ba, fl->desc); if (fl->sdesc) free_fl_sdesc(sc, fl); if (mtx_initialized(&fl->fl_lock)) mtx_destroy(&fl->fl_lock); bzero(fl, sizeof(*fl)); } return (0); } static void add_fl_sysctls(struct sysctl_ctx_list *ctx, struct sysctl_oid *oid, struct sge_fl *fl) { struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL, "freelist"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id", CTLTYPE_INT | CTLFLAG_RD, &fl->cntxt_id, 0, sysctl_uint16, "I", "SGE context id of the freelist"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "padding", CTLFLAG_RD, NULL, fl_pad ? 1 : 0, "padding enabled"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "packing", CTLFLAG_RD, NULL, fl->flags & FL_BUF_PACKING ? 1 : 0, "packing enabled"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &fl->cidx, 0, "consumer index"); if (fl->flags & FL_BUF_PACKING) { SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "rx_offset", CTLFLAG_RD, &fl->rx_offset, 0, "packing rx offset"); } SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &fl->pidx, 0, "producer index"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_allocated", CTLFLAG_RD, &fl->mbuf_allocated, "# of mbuf allocated"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "mbuf_inlined", CTLFLAG_RD, &fl->mbuf_inlined, "# of mbuf inlined in clusters"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_allocated", CTLFLAG_RD, &fl->cl_allocated, "# of clusters allocated"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_recycled", CTLFLAG_RD, &fl->cl_recycled, "# of clusters recycled"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "cluster_fast_recycled", CTLFLAG_RD, &fl->cl_fast_recycled, "# of clusters recycled (fast)"); } static int alloc_fwq(struct adapter *sc) { int rc, intr_idx; struct sge_iq *fwq = &sc->sge.fwq; struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev); struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); init_iq(fwq, sc, 0, 0, FW_IQ_QSIZE); fwq->flags |= IQ_INTR; /* always */ if (sc->flags & IS_VF) intr_idx = 0; else { intr_idx = sc->intr_count > 1 ? 1 : 0; fwq->set_tcb_rpl = t4_filter_rpl; fwq->l2t_write_rpl = do_l2t_write_rpl; } rc = alloc_iq_fl(&sc->port[0]->vi[0], fwq, NULL, intr_idx, -1); if (rc != 0) { device_printf(sc->dev, "failed to create firmware event queue: %d\n", rc); return (rc); } oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "fwq", CTLFLAG_RD, NULL, "firmware event queue"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "abs_id", CTLTYPE_INT | CTLFLAG_RD, &fwq->abs_id, 0, sysctl_uint16, "I", "absolute id of the queue"); SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cntxt_id", CTLTYPE_INT | CTLFLAG_RD, &fwq->cntxt_id, 0, sysctl_uint16, "I", "SGE context id of the queue"); SYSCTL_ADD_PROC(&sc->ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &fwq->cidx, 0, sysctl_uint16, "I", "consumer index"); return (0); } static int free_fwq(struct adapter *sc) { return free_iq_fl(NULL, &sc->sge.fwq, NULL); } static int alloc_mgmtq(struct adapter *sc) { int rc; struct sge_wrq *mgmtq = &sc->sge.mgmtq; char name[16]; struct sysctl_oid *oid = device_get_sysctl_tree(sc->dev); struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); oid = SYSCTL_ADD_NODE(&sc->ctx, children, OID_AUTO, "mgmtq", CTLFLAG_RD, NULL, "management queue"); snprintf(name, sizeof(name), "%s mgmtq", device_get_nameunit(sc->dev)); init_eq(sc, &mgmtq->eq, EQ_CTRL, CTRL_EQ_QSIZE, sc->port[0]->tx_chan, sc->sge.fwq.cntxt_id, name); rc = alloc_wrq(sc, NULL, mgmtq, oid); if (rc != 0) { device_printf(sc->dev, "failed to create management queue: %d\n", rc); return (rc); } return (0); } static int free_mgmtq(struct adapter *sc) { return free_wrq(sc, &sc->sge.mgmtq); } int tnl_cong(struct port_info *pi, int drop) { if (drop == -1) return (-1); else if (drop == 1) return (0); else return (pi->rx_chan_map); } static int alloc_rxq(struct vi_info *vi, struct sge_rxq *rxq, int intr_idx, int idx, struct sysctl_oid *oid) { int rc; struct adapter *sc = vi->pi->adapter; struct sysctl_oid_list *children; char name[16]; rc = alloc_iq_fl(vi, &rxq->iq, &rxq->fl, intr_idx, tnl_cong(vi->pi, cong_drop)); if (rc != 0) return (rc); if (idx == 0) sc->sge.iq_base = rxq->iq.abs_id - rxq->iq.cntxt_id; else KASSERT(rxq->iq.cntxt_id + sc->sge.iq_base == rxq->iq.abs_id, ("iq_base mismatch")); KASSERT(sc->sge.iq_base == 0 || sc->flags & IS_VF, ("PF with non-zero iq_base")); /* * The freelist is just barely above the starvation threshold right now, * fill it up a bit more. */ FL_LOCK(&rxq->fl); refill_fl(sc, &rxq->fl, 128); FL_UNLOCK(&rxq->fl); #if defined(INET) || defined(INET6) rc = tcp_lro_init(&rxq->lro); if (rc != 0) return (rc); rxq->lro.ifp = vi->ifp; /* also indicates LRO init'ed */ if (vi->ifp->if_capenable & IFCAP_LRO) rxq->iq.flags |= IQ_LRO_ENABLED; #endif rxq->ifp = vi->ifp; children = SYSCTL_CHILDREN(oid); snprintf(name, sizeof(name), "%d", idx); oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL, "rx queue"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id", CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.abs_id, 0, sysctl_uint16, "I", "absolute id of the queue"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cntxt_id, 0, sysctl_uint16, "I", "SGE context id of the queue"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &rxq->iq.cidx, 0, sysctl_uint16, "I", "consumer index"); #if defined(INET) || defined(INET6) SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_queued", CTLFLAG_RD, &rxq->lro.lro_queued, 0, NULL); SYSCTL_ADD_U64(&vi->ctx, children, OID_AUTO, "lro_flushed", CTLFLAG_RD, &rxq->lro.lro_flushed, 0, NULL); #endif SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "rxcsum", CTLFLAG_RD, &rxq->rxcsum, "# of times hardware assisted with checksum"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_extraction", CTLFLAG_RD, &rxq->vlan_extraction, "# of times hardware extracted 802.1Q tag"); add_fl_sysctls(&vi->ctx, oid, &rxq->fl); return (rc); } static int free_rxq(struct vi_info *vi, struct sge_rxq *rxq) { int rc; #if defined(INET) || defined(INET6) if (rxq->lro.ifp) { tcp_lro_free(&rxq->lro); rxq->lro.ifp = NULL; } #endif rc = free_iq_fl(vi, &rxq->iq, &rxq->fl); if (rc == 0) bzero(rxq, sizeof(*rxq)); return (rc); } #ifdef TCP_OFFLOAD static int alloc_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq, int intr_idx, int idx, struct sysctl_oid *oid) { int rc; struct sysctl_oid_list *children; char name[16]; rc = alloc_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl, intr_idx, vi->pi->rx_chan_map); if (rc != 0) return (rc); children = SYSCTL_CHILDREN(oid); snprintf(name, sizeof(name), "%d", idx); oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL, "rx queue"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "abs_id", CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.abs_id, 0, sysctl_uint16, "I", "absolute id of the queue"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cntxt_id, 0, sysctl_uint16, "I", "SGE context id of the queue"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &ofld_rxq->iq.cidx, 0, sysctl_uint16, "I", "consumer index"); add_fl_sysctls(&vi->ctx, oid, &ofld_rxq->fl); return (rc); } static int free_ofld_rxq(struct vi_info *vi, struct sge_ofld_rxq *ofld_rxq) { int rc; rc = free_iq_fl(vi, &ofld_rxq->iq, &ofld_rxq->fl); if (rc == 0) bzero(ofld_rxq, sizeof(*ofld_rxq)); return (rc); } #endif #ifdef DEV_NETMAP static int alloc_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq, int intr_idx, int idx, struct sysctl_oid *oid) { int rc; struct sysctl_oid_list *children; struct sysctl_ctx_list *ctx; char name[16]; size_t len; struct adapter *sc = vi->pi->adapter; struct netmap_adapter *na = NA(vi->ifp); MPASS(na != NULL); len = vi->qsize_rxq * IQ_ESIZE; rc = alloc_ring(sc, len, &nm_rxq->iq_desc_tag, &nm_rxq->iq_desc_map, &nm_rxq->iq_ba, (void **)&nm_rxq->iq_desc); if (rc != 0) return (rc); len = na->num_rx_desc * EQ_ESIZE + sc->params.sge.spg_len; rc = alloc_ring(sc, len, &nm_rxq->fl_desc_tag, &nm_rxq->fl_desc_map, &nm_rxq->fl_ba, (void **)&nm_rxq->fl_desc); if (rc != 0) return (rc); nm_rxq->vi = vi; nm_rxq->nid = idx; nm_rxq->iq_cidx = 0; nm_rxq->iq_sidx = vi->qsize_rxq - sc->params.sge.spg_len / IQ_ESIZE; nm_rxq->iq_gen = F_RSPD_GEN; nm_rxq->fl_pidx = nm_rxq->fl_cidx = 0; nm_rxq->fl_sidx = na->num_rx_desc; nm_rxq->intr_idx = intr_idx; ctx = &vi->ctx; children = SYSCTL_CHILDREN(oid); snprintf(name, sizeof(name), "%d", idx); oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL, "rx queue"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "abs_id", CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_abs_id, 0, sysctl_uint16, "I", "absolute id of the queue"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id", CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cntxt_id, 0, sysctl_uint16, "I", "SGE context id of the queue"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->iq_cidx, 0, sysctl_uint16, "I", "consumer index"); children = SYSCTL_CHILDREN(oid); oid = SYSCTL_ADD_NODE(ctx, children, OID_AUTO, "fl", CTLFLAG_RD, NULL, "freelist"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cntxt_id", CTLTYPE_INT | CTLFLAG_RD, &nm_rxq->fl_cntxt_id, 0, sysctl_uint16, "I", "SGE context id of the freelist"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cidx", CTLFLAG_RD, &nm_rxq->fl_cidx, 0, "consumer index"); SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "pidx", CTLFLAG_RD, &nm_rxq->fl_pidx, 0, "producer index"); return (rc); } static int free_nm_rxq(struct vi_info *vi, struct sge_nm_rxq *nm_rxq) { struct adapter *sc = vi->pi->adapter; free_ring(sc, nm_rxq->iq_desc_tag, nm_rxq->iq_desc_map, nm_rxq->iq_ba, nm_rxq->iq_desc); free_ring(sc, nm_rxq->fl_desc_tag, nm_rxq->fl_desc_map, nm_rxq->fl_ba, nm_rxq->fl_desc); return (0); } static int alloc_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq, int iqidx, int idx, struct sysctl_oid *oid) { int rc; size_t len; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct netmap_adapter *na = NA(vi->ifp); char name[16]; struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); len = na->num_tx_desc * EQ_ESIZE + sc->params.sge.spg_len; rc = alloc_ring(sc, len, &nm_txq->desc_tag, &nm_txq->desc_map, &nm_txq->ba, (void **)&nm_txq->desc); if (rc) return (rc); nm_txq->pidx = nm_txq->cidx = 0; nm_txq->sidx = na->num_tx_desc; nm_txq->nid = idx; nm_txq->iqidx = iqidx; nm_txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) | V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) | V_TXPKT_VF(vi->viid)); snprintf(name, sizeof(name), "%d", idx); oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL, "netmap tx queue"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, &nm_txq->cntxt_id, 0, "SGE context id of the queue"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &nm_txq->cidx, 0, sysctl_uint16, "I", "consumer index"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx", CTLTYPE_INT | CTLFLAG_RD, &nm_txq->pidx, 0, sysctl_uint16, "I", "producer index"); return (rc); } static int free_nm_txq(struct vi_info *vi, struct sge_nm_txq *nm_txq) { struct adapter *sc = vi->pi->adapter; free_ring(sc, nm_txq->desc_tag, nm_txq->desc_map, nm_txq->ba, nm_txq->desc); return (0); } #endif static int ctrl_eq_alloc(struct adapter *sc, struct sge_eq *eq) { int rc, cntxt_id; struct fw_eq_ctrl_cmd c; int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; bzero(&c, sizeof(c)); c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_CTRL_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_CTRL_CMD_PFN(sc->pf) | V_FW_EQ_CTRL_CMD_VFN(0)); c.alloc_to_len16 = htobe32(F_FW_EQ_CTRL_CMD_ALLOC | F_FW_EQ_CTRL_CMD_EQSTART | FW_LEN16(c)); c.cmpliqid_eqid = htonl(V_FW_EQ_CTRL_CMD_CMPLIQID(eq->iqid)); c.physeqid_pkd = htobe32(0); c.fetchszm_to_iqid = htobe32(V_FW_EQ_CTRL_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) | V_FW_EQ_CTRL_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_CTRL_CMD_FETCHRO | V_FW_EQ_CTRL_CMD_IQID(eq->iqid)); c.dcaen_to_eqsize = htobe32(V_FW_EQ_CTRL_CMD_FBMIN(X_FETCHBURSTMIN_64B) | V_FW_EQ_CTRL_CMD_FBMAX(X_FETCHBURSTMAX_512B) | V_FW_EQ_CTRL_CMD_EQSIZE(qsize)); c.eqaddr = htobe64(eq->ba); rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); if (rc != 0) { device_printf(sc->dev, "failed to create control queue %d: %d\n", eq->tx_chan, rc); return (rc); } eq->flags |= EQ_ALLOCATED; eq->cntxt_id = G_FW_EQ_CTRL_CMD_EQID(be32toh(c.cmpliqid_eqid)); cntxt_id = eq->cntxt_id - sc->sge.eq_start; if (cntxt_id >= sc->sge.neq) panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.neq - 1); sc->sge.eqmap[cntxt_id] = eq; return (rc); } static int eth_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq) { int rc, cntxt_id; struct fw_eq_eth_cmd c; int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; bzero(&c, sizeof(c)); c.op_to_vfn = htobe32(V_FW_CMD_OP(FW_EQ_ETH_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_ETH_CMD_PFN(sc->pf) | V_FW_EQ_ETH_CMD_VFN(0)); c.alloc_to_len16 = htobe32(F_FW_EQ_ETH_CMD_ALLOC | F_FW_EQ_ETH_CMD_EQSTART | FW_LEN16(c)); c.autoequiqe_to_viid = htobe32(F_FW_EQ_ETH_CMD_AUTOEQUIQE | F_FW_EQ_ETH_CMD_AUTOEQUEQE | V_FW_EQ_ETH_CMD_VIID(vi->viid)); c.fetchszm_to_iqid = htobe32(V_FW_EQ_ETH_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) | V_FW_EQ_ETH_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_ETH_CMD_FETCHRO | V_FW_EQ_ETH_CMD_IQID(eq->iqid)); c.dcaen_to_eqsize = htobe32(V_FW_EQ_ETH_CMD_FBMIN(X_FETCHBURSTMIN_64B) | V_FW_EQ_ETH_CMD_FBMAX(X_FETCHBURSTMAX_512B) | V_FW_EQ_ETH_CMD_EQSIZE(qsize)); c.eqaddr = htobe64(eq->ba); rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); if (rc != 0) { device_printf(vi->dev, "failed to create Ethernet egress queue: %d\n", rc); return (rc); } eq->flags |= EQ_ALLOCATED; eq->cntxt_id = G_FW_EQ_ETH_CMD_EQID(be32toh(c.eqid_pkd)); eq->abs_id = G_FW_EQ_ETH_CMD_PHYSEQID(be32toh(c.physeqid_pkd)); cntxt_id = eq->cntxt_id - sc->sge.eq_start; if (cntxt_id >= sc->sge.neq) panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.neq - 1); sc->sge.eqmap[cntxt_id] = eq; return (rc); } #ifdef TCP_OFFLOAD static int ofld_eq_alloc(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq) { int rc, cntxt_id; struct fw_eq_ofld_cmd c; int qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; bzero(&c, sizeof(c)); c.op_to_vfn = htonl(V_FW_CMD_OP(FW_EQ_OFLD_CMD) | F_FW_CMD_REQUEST | F_FW_CMD_WRITE | F_FW_CMD_EXEC | V_FW_EQ_OFLD_CMD_PFN(sc->pf) | V_FW_EQ_OFLD_CMD_VFN(0)); c.alloc_to_len16 = htonl(F_FW_EQ_OFLD_CMD_ALLOC | F_FW_EQ_OFLD_CMD_EQSTART | FW_LEN16(c)); c.fetchszm_to_iqid = htonl(V_FW_EQ_OFLD_CMD_HOSTFCMODE(X_HOSTFCMODE_NONE) | V_FW_EQ_OFLD_CMD_PCIECHN(eq->tx_chan) | F_FW_EQ_OFLD_CMD_FETCHRO | V_FW_EQ_OFLD_CMD_IQID(eq->iqid)); c.dcaen_to_eqsize = htobe32(V_FW_EQ_OFLD_CMD_FBMIN(X_FETCHBURSTMIN_64B) | V_FW_EQ_OFLD_CMD_FBMAX(X_FETCHBURSTMAX_512B) | V_FW_EQ_OFLD_CMD_EQSIZE(qsize)); c.eqaddr = htobe64(eq->ba); rc = -t4_wr_mbox(sc, sc->mbox, &c, sizeof(c), &c); if (rc != 0) { device_printf(vi->dev, "failed to create egress queue for TCP offload: %d\n", rc); return (rc); } eq->flags |= EQ_ALLOCATED; eq->cntxt_id = G_FW_EQ_OFLD_CMD_EQID(be32toh(c.eqid_pkd)); cntxt_id = eq->cntxt_id - sc->sge.eq_start; if (cntxt_id >= sc->sge.neq) panic("%s: eq->cntxt_id (%d) more than the max (%d)", __func__, cntxt_id, sc->sge.neq - 1); sc->sge.eqmap[cntxt_id] = eq; return (rc); } #endif static int alloc_eq(struct adapter *sc, struct vi_info *vi, struct sge_eq *eq) { int rc, qsize; size_t len; mtx_init(&eq->eq_lock, eq->lockname, NULL, MTX_DEF); qsize = eq->sidx + sc->params.sge.spg_len / EQ_ESIZE; len = qsize * EQ_ESIZE; rc = alloc_ring(sc, len, &eq->desc_tag, &eq->desc_map, &eq->ba, (void **)&eq->desc); if (rc) return (rc); eq->pidx = eq->cidx = 0; eq->equeqidx = eq->dbidx = 0; eq->doorbells = sc->doorbells; switch (eq->flags & EQ_TYPEMASK) { case EQ_CTRL: rc = ctrl_eq_alloc(sc, eq); break; case EQ_ETH: rc = eth_eq_alloc(sc, vi, eq); break; #ifdef TCP_OFFLOAD case EQ_OFLD: rc = ofld_eq_alloc(sc, vi, eq); break; #endif default: panic("%s: invalid eq type %d.", __func__, eq->flags & EQ_TYPEMASK); } if (rc != 0) { device_printf(sc->dev, "failed to allocate egress queue(%d): %d\n", eq->flags & EQ_TYPEMASK, rc); } if (isset(&eq->doorbells, DOORBELL_UDB) || isset(&eq->doorbells, DOORBELL_UDBWC) || isset(&eq->doorbells, DOORBELL_WCWR)) { uint32_t s_qpp = sc->params.sge.eq_s_qpp; uint32_t mask = (1 << s_qpp) - 1; volatile uint8_t *udb; udb = sc->udbs_base + UDBS_DB_OFFSET; udb += (eq->cntxt_id >> s_qpp) << PAGE_SHIFT; /* pg offset */ eq->udb_qid = eq->cntxt_id & mask; /* id in page */ if (eq->udb_qid >= PAGE_SIZE / UDBS_SEG_SIZE) clrbit(&eq->doorbells, DOORBELL_WCWR); else { udb += eq->udb_qid << UDBS_SEG_SHIFT; /* seg offset */ eq->udb_qid = 0; } eq->udb = (volatile void *)udb; } return (rc); } static int free_eq(struct adapter *sc, struct sge_eq *eq) { int rc; if (eq->flags & EQ_ALLOCATED) { switch (eq->flags & EQ_TYPEMASK) { case EQ_CTRL: rc = -t4_ctrl_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id); break; case EQ_ETH: rc = -t4_eth_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id); break; #ifdef TCP_OFFLOAD case EQ_OFLD: rc = -t4_ofld_eq_free(sc, sc->mbox, sc->pf, 0, eq->cntxt_id); break; #endif default: panic("%s: invalid eq type %d.", __func__, eq->flags & EQ_TYPEMASK); } if (rc != 0) { device_printf(sc->dev, "failed to free egress queue (%d): %d\n", eq->flags & EQ_TYPEMASK, rc); return (rc); } eq->flags &= ~EQ_ALLOCATED; } free_ring(sc, eq->desc_tag, eq->desc_map, eq->ba, eq->desc); if (mtx_initialized(&eq->eq_lock)) mtx_destroy(&eq->eq_lock); bzero(eq, sizeof(*eq)); return (0); } static int alloc_wrq(struct adapter *sc, struct vi_info *vi, struct sge_wrq *wrq, struct sysctl_oid *oid) { int rc; struct sysctl_ctx_list *ctx = vi ? &vi->ctx : &sc->ctx; struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); rc = alloc_eq(sc, vi, &wrq->eq); if (rc) return (rc); wrq->adapter = sc; TASK_INIT(&wrq->wrq_tx_task, 0, wrq_tx_drain, wrq); TAILQ_INIT(&wrq->incomplete_wrs); STAILQ_INIT(&wrq->wr_list); wrq->nwr_pending = 0; wrq->ndesc_needed = 0; SYSCTL_ADD_UINT(ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, &wrq->eq.cntxt_id, 0, "SGE context id of the queue"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.cidx, 0, sysctl_uint16, "I", "consumer index"); SYSCTL_ADD_PROC(ctx, children, OID_AUTO, "pidx", CTLTYPE_INT | CTLFLAG_RD, &wrq->eq.pidx, 0, sysctl_uint16, "I", "producer index"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_direct", CTLFLAG_RD, &wrq->tx_wrs_direct, "# of work requests (direct)"); SYSCTL_ADD_UQUAD(ctx, children, OID_AUTO, "tx_wrs_copied", CTLFLAG_RD, &wrq->tx_wrs_copied, "# of work requests (copied)"); return (rc); } static int free_wrq(struct adapter *sc, struct sge_wrq *wrq) { int rc; rc = free_eq(sc, &wrq->eq); if (rc) return (rc); bzero(wrq, sizeof(*wrq)); return (0); } static int alloc_txq(struct vi_info *vi, struct sge_txq *txq, int idx, struct sysctl_oid *oid) { int rc; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct sge_eq *eq = &txq->eq; char name[16]; struct sysctl_oid_list *children = SYSCTL_CHILDREN(oid); rc = mp_ring_alloc(&txq->r, eq->sidx, txq, eth_tx, can_resume_eth_tx, M_CXGBE, M_WAITOK); if (rc != 0) { device_printf(sc->dev, "failed to allocate mp_ring: %d\n", rc); return (rc); } rc = alloc_eq(sc, vi, eq); if (rc != 0) { mp_ring_free(txq->r); txq->r = NULL; return (rc); } /* Can't fail after this point. */ if (idx == 0) sc->sge.eq_base = eq->abs_id - eq->cntxt_id; else KASSERT(eq->cntxt_id + sc->sge.eq_base == eq->abs_id, ("eq_base mismatch")); KASSERT(sc->sge.eq_base == 0 || sc->flags & IS_VF, ("PF with non-zero eq_base")); TASK_INIT(&txq->tx_reclaim_task, 0, tx_reclaim, eq); txq->ifp = vi->ifp; txq->gl = sglist_alloc(TX_SGL_SEGS, M_WAITOK); if (sc->flags & IS_VF) txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT_XT) | V_TXPKT_INTF(pi->tx_chan)); else txq->cpl_ctrl0 = htobe32(V_TXPKT_OPCODE(CPL_TX_PKT) | V_TXPKT_INTF(pi->tx_chan) | V_TXPKT_VF_VLD(1) | V_TXPKT_VF(vi->viid)); txq->tc_idx = -1; txq->sdesc = malloc(eq->sidx * sizeof(struct tx_sdesc), M_CXGBE, M_ZERO | M_WAITOK); snprintf(name, sizeof(name), "%d", idx); oid = SYSCTL_ADD_NODE(&vi->ctx, children, OID_AUTO, name, CTLFLAG_RD, NULL, "tx queue"); children = SYSCTL_CHILDREN(oid); SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "abs_id", CTLFLAG_RD, &eq->abs_id, 0, "absolute id of the queue"); SYSCTL_ADD_UINT(&vi->ctx, children, OID_AUTO, "cntxt_id", CTLFLAG_RD, &eq->cntxt_id, 0, "SGE context id of the queue"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "cidx", CTLTYPE_INT | CTLFLAG_RD, &eq->cidx, 0, sysctl_uint16, "I", "consumer index"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "pidx", CTLTYPE_INT | CTLFLAG_RD, &eq->pidx, 0, sysctl_uint16, "I", "producer index"); SYSCTL_ADD_PROC(&vi->ctx, children, OID_AUTO, "tc", CTLTYPE_INT | CTLFLAG_RW, vi, idx, sysctl_tc, "I", "traffic class (-1 means none)"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txcsum", CTLFLAG_RD, &txq->txcsum, "# of times hardware assisted with checksum"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "vlan_insertion", CTLFLAG_RD, &txq->vlan_insertion, "# of times hardware inserted 802.1Q tag"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "tso_wrs", CTLFLAG_RD, &txq->tso_wrs, "# of TSO work requests"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "imm_wrs", CTLFLAG_RD, &txq->imm_wrs, "# of work requests with immediate data"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "sgl_wrs", CTLFLAG_RD, &txq->sgl_wrs, "# of work requests with direct SGL"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkt_wrs", CTLFLAG_RD, &txq->txpkt_wrs, "# of txpkt work requests (one pkt/WR)"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_wrs", CTLFLAG_RD, &txq->txpkts0_wrs, "# of txpkts (type 0) work requests"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_wrs", CTLFLAG_RD, &txq->txpkts1_wrs, "# of txpkts (type 1) work requests"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts0_pkts", CTLFLAG_RD, &txq->txpkts0_pkts, "# of frames tx'd using type0 txpkts work requests"); SYSCTL_ADD_UQUAD(&vi->ctx, children, OID_AUTO, "txpkts1_pkts", CTLFLAG_RD, &txq->txpkts1_pkts, "# of frames tx'd using type1 txpkts work requests"); SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_enqueues", CTLFLAG_RD, &txq->r->enqueues, "# of enqueues to the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_drops", CTLFLAG_RD, &txq->r->drops, "# of drops in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_starts", CTLFLAG_RD, &txq->r->starts, "# of normal consumer starts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_stalls", CTLFLAG_RD, &txq->r->stalls, "# of consumer stalls in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_restarts", CTLFLAG_RD, &txq->r->restarts, "# of consumer restarts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(&vi->ctx, children, OID_AUTO, "r_abdications", CTLFLAG_RD, &txq->r->abdications, "# of consumer abdications in the mp_ring for this queue"); return (0); } static int free_txq(struct vi_info *vi, struct sge_txq *txq) { int rc; struct adapter *sc = vi->pi->adapter; struct sge_eq *eq = &txq->eq; rc = free_eq(sc, eq); if (rc) return (rc); sglist_free(txq->gl); free(txq->sdesc, M_CXGBE); mp_ring_free(txq->r); bzero(txq, sizeof(*txq)); return (0); } static void oneseg_dma_callback(void *arg, bus_dma_segment_t *segs, int nseg, int error) { bus_addr_t *ba = arg; KASSERT(nseg == 1, ("%s meant for single segment mappings only.", __func__)); *ba = error ? 0 : segs->ds_addr; } static inline void ring_fl_db(struct adapter *sc, struct sge_fl *fl) { uint32_t n, v; n = IDXDIFF(fl->pidx / 8, fl->dbidx, fl->sidx); MPASS(n > 0); wmb(); v = fl->dbval | V_PIDX(n); if (fl->udb) *fl->udb = htole32(v); else t4_write_reg(sc, sc->sge_kdoorbell_reg, v); IDXINCR(fl->dbidx, n, fl->sidx); } /* * Fills up the freelist by allocating up to 'n' buffers. Buffers that are * recycled do not count towards this allocation budget. * * Returns non-zero to indicate that this freelist should be added to the list * of starving freelists. */ static int refill_fl(struct adapter *sc, struct sge_fl *fl, int n) { __be64 *d; struct fl_sdesc *sd; uintptr_t pa; caddr_t cl; struct cluster_layout *cll; struct sw_zone_info *swz; struct cluster_metadata *clm; uint16_t max_pidx; uint16_t hw_cidx = fl->hw_cidx; /* stable snapshot */ FL_LOCK_ASSERT_OWNED(fl); /* * We always stop at the beginning of the hardware descriptor that's just * before the one with the hw cidx. This is to avoid hw pidx = hw cidx, * which would mean an empty freelist to the chip. */ max_pidx = __predict_false(hw_cidx == 0) ? fl->sidx - 1 : hw_cidx - 1; if (fl->pidx == max_pidx * 8) return (0); d = &fl->desc[fl->pidx]; sd = &fl->sdesc[fl->pidx]; cll = &fl->cll_def; /* default layout */ swz = &sc->sge.sw_zone_info[cll->zidx]; while (n > 0) { if (sd->cl != NULL) { if (sd->nmbuf == 0) { /* * Fast recycle without involving any atomics on * the cluster's metadata (if the cluster has * metadata). This happens when all frames * received in the cluster were small enough to * fit within a single mbuf each. */ fl->cl_fast_recycled++; #ifdef INVARIANTS clm = cl_metadata(sc, fl, &sd->cll, sd->cl); if (clm != NULL) MPASS(clm->refcount == 1); #endif goto recycled_fast; } /* * Cluster is guaranteed to have metadata. Clusters * without metadata always take the fast recycle path * when they're recycled. */ clm = cl_metadata(sc, fl, &sd->cll, sd->cl); MPASS(clm != NULL); if (atomic_fetchadd_int(&clm->refcount, -1) == 1) { fl->cl_recycled++; counter_u64_add(extfree_rels, 1); goto recycled; } sd->cl = NULL; /* gave up my reference */ } MPASS(sd->cl == NULL); alloc: cl = uma_zalloc(swz->zone, M_NOWAIT); if (__predict_false(cl == NULL)) { if (cll == &fl->cll_alt || fl->cll_alt.zidx == -1 || fl->cll_def.zidx == fl->cll_alt.zidx) break; /* fall back to the safe zone */ cll = &fl->cll_alt; swz = &sc->sge.sw_zone_info[cll->zidx]; goto alloc; } fl->cl_allocated++; n--; pa = pmap_kextract((vm_offset_t)cl); pa += cll->region1; sd->cl = cl; sd->cll = *cll; *d = htobe64(pa | cll->hwidx); clm = cl_metadata(sc, fl, cll, cl); if (clm != NULL) { recycled: #ifdef INVARIANTS clm->sd = sd; #endif clm->refcount = 1; } sd->nmbuf = 0; recycled_fast: d++; sd++; if (__predict_false(++fl->pidx % 8 == 0)) { uint16_t pidx = fl->pidx / 8; if (__predict_false(pidx == fl->sidx)) { fl->pidx = 0; pidx = 0; sd = fl->sdesc; d = fl->desc; } if (pidx == max_pidx) break; if (IDXDIFF(pidx, fl->dbidx, fl->sidx) >= 4) ring_fl_db(sc, fl); } } if (fl->pidx / 8 != fl->dbidx) ring_fl_db(sc, fl); return (FL_RUNNING_LOW(fl) && !(fl->flags & FL_STARVING)); } /* * Attempt to refill all starving freelists. */ static void refill_sfl(void *arg) { struct adapter *sc = arg; struct sge_fl *fl, *fl_temp; mtx_assert(&sc->sfl_lock, MA_OWNED); TAILQ_FOREACH_SAFE(fl, &sc->sfl, link, fl_temp) { FL_LOCK(fl); refill_fl(sc, fl, 64); if (FL_NOT_RUNNING_LOW(fl) || fl->flags & FL_DOOMED) { TAILQ_REMOVE(&sc->sfl, fl, link); fl->flags &= ~FL_STARVING; } FL_UNLOCK(fl); } if (!TAILQ_EMPTY(&sc->sfl)) callout_schedule(&sc->sfl_callout, hz / 5); } static int alloc_fl_sdesc(struct sge_fl *fl) { fl->sdesc = malloc(fl->sidx * 8 * sizeof(struct fl_sdesc), M_CXGBE, M_ZERO | M_WAITOK); return (0); } static void free_fl_sdesc(struct adapter *sc, struct sge_fl *fl) { struct fl_sdesc *sd; struct cluster_metadata *clm; struct cluster_layout *cll; int i; sd = fl->sdesc; for (i = 0; i < fl->sidx * 8; i++, sd++) { if (sd->cl == NULL) continue; cll = &sd->cll; clm = cl_metadata(sc, fl, cll, sd->cl); if (sd->nmbuf == 0) uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl); else if (clm && atomic_fetchadd_int(&clm->refcount, -1) == 1) { uma_zfree(sc->sge.sw_zone_info[cll->zidx].zone, sd->cl); counter_u64_add(extfree_rels, 1); } sd->cl = NULL; } free(fl->sdesc, M_CXGBE); fl->sdesc = NULL; } static inline void get_pkt_gl(struct mbuf *m, struct sglist *gl) { int rc; M_ASSERTPKTHDR(m); sglist_reset(gl); rc = sglist_append_mbuf(gl, m); if (__predict_false(rc != 0)) { panic("%s: mbuf %p (%d segs) was vetted earlier but now fails " "with %d.", __func__, m, mbuf_nsegs(m), rc); } KASSERT(gl->sg_nseg == mbuf_nsegs(m), ("%s: nsegs changed for mbuf %p from %d to %d", __func__, m, mbuf_nsegs(m), gl->sg_nseg)); KASSERT(gl->sg_nseg > 0 && gl->sg_nseg <= (needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS), ("%s: %d segments, should have been 1 <= nsegs <= %d", __func__, gl->sg_nseg, needs_tso(m) ? TX_SGL_SEGS_TSO : TX_SGL_SEGS)); } /* * len16 for a txpkt WR with a GL. Includes the firmware work request header. */ static inline u_int txpkt_len16(u_int nsegs, u_int tso) { u_int n; MPASS(nsegs > 0); nsegs--; /* first segment is part of ulptx_sgl */ n = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1)); if (tso) n += sizeof(struct cpl_tx_pkt_lso_core); return (howmany(n, 16)); } /* * len16 for a txpkt_vm WR with a GL. Includes the firmware work * request header. */ static inline u_int txpkt_vm_len16(u_int nsegs, u_int tso) { u_int n; MPASS(nsegs > 0); nsegs--; /* first segment is part of ulptx_sgl */ n = sizeof(struct fw_eth_tx_pkt_vm_wr) + sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1)); if (tso) n += sizeof(struct cpl_tx_pkt_lso_core); return (howmany(n, 16)); } /* * len16 for a txpkts type 0 WR with a GL. Does not include the firmware work * request header. */ static inline u_int txpkts0_len16(u_int nsegs) { u_int n; MPASS(nsegs > 0); nsegs--; /* first segment is part of ulptx_sgl */ n = sizeof(struct ulp_txpkt) + sizeof(struct ulptx_idata) + sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl) + 8 * ((3 * nsegs) / 2 + (nsegs & 1)); return (howmany(n, 16)); } /* * len16 for a txpkts type 1 WR with a GL. Does not include the firmware work * request header. */ static inline u_int txpkts1_len16(void) { u_int n; n = sizeof(struct cpl_tx_pkt_core) + sizeof(struct ulptx_sgl); return (howmany(n, 16)); } static inline u_int imm_payload(u_int ndesc) { u_int n; n = ndesc * EQ_ESIZE - sizeof(struct fw_eth_tx_pkt_wr) - sizeof(struct cpl_tx_pkt_core); return (n); } /* * Write a VM txpkt WR for this packet to the hardware descriptors, update the * software descriptor, and advance the pidx. It is guaranteed that enough * descriptors are available. * * The return value is the # of hardware descriptors used. */ static u_int -write_txpkt_vm_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_vm_wr *wr, - struct mbuf *m0, u_int available) +write_txpkt_vm_wr(struct adapter *sc, struct sge_txq *txq, + struct fw_eth_tx_pkt_vm_wr *wr, struct mbuf *m0, u_int available) { struct sge_eq *eq = &txq->eq; struct tx_sdesc *txsd; struct cpl_tx_pkt_core *cpl; uint32_t ctrl; /* used in many unrelated places */ uint64_t ctrl1; int csum_type, len16, ndesc, pktlen, nsegs; caddr_t dst; TXQ_LOCK_ASSERT_OWNED(txq); M_ASSERTPKTHDR(m0); MPASS(available > 0 && available < eq->sidx); len16 = mbuf_len16(m0); nsegs = mbuf_nsegs(m0); pktlen = m0->m_pkthdr.len; ctrl = sizeof(struct cpl_tx_pkt_core); if (needs_tso(m0)) ctrl += sizeof(struct cpl_tx_pkt_lso_core); ndesc = howmany(len16, EQ_ESIZE / 16); MPASS(ndesc <= available); /* Firmware work request header */ MPASS(wr == (void *)&eq->desc[eq->pidx]); wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_VM_WR) | V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); ctrl = V_FW_WR_LEN16(len16); wr->equiq_to_len16 = htobe32(ctrl); wr->r3[0] = 0; wr->r3[1] = 0; /* * Copy over ethmacdst, ethmacsrc, ethtype, and vlantci. * vlantci is ignored unless the ethtype is 0x8100, so it's * simpler to always copy it rather than making it * conditional. Also, it seems that we do not have to set * vlantci or fake the ethtype when doing VLAN tag insertion. */ m_copydata(m0, 0, sizeof(struct ether_header) + 2, wr->ethmacdst); csum_type = -1; if (needs_tso(m0)) { struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1); KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 && m0->m_pkthdr.l4hlen > 0, ("%s: mbuf %p needs TSO but missing header lengths", __func__, m0)); ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2); if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header)) ctrl |= V_LSO_ETHHDR_LEN(1); if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr)) ctrl |= F_LSO_IPV6; lso->lso_ctrl = htobe32(ctrl); lso->ipid_ofst = htobe16(0); lso->mss = htobe16(m0->m_pkthdr.tso_segsz); lso->seqno_offset = htobe32(0); lso->len = htobe32(pktlen); if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr)) csum_type = TX_CSUM_TCPIP6; else csum_type = TX_CSUM_TCPIP; cpl = (void *)(lso + 1); txq->tso_wrs++; } else { if (m0->m_pkthdr.csum_flags & CSUM_IP_TCP) csum_type = TX_CSUM_TCPIP; else if (m0->m_pkthdr.csum_flags & CSUM_IP_UDP) csum_type = TX_CSUM_UDPIP; else if (m0->m_pkthdr.csum_flags & CSUM_IP6_TCP) csum_type = TX_CSUM_TCPIP6; else if (m0->m_pkthdr.csum_flags & CSUM_IP6_UDP) csum_type = TX_CSUM_UDPIP6; #if defined(INET) else if (m0->m_pkthdr.csum_flags & CSUM_IP) { /* * XXX: The firmware appears to stomp on the * fragment/flags field of the IP header when * using TX_CSUM_IP. Fall back to doing * software checksums. */ u_short *sump; struct mbuf *m; int offset; m = m0; offset = 0; sump = m_advance(&m, &offset, m0->m_pkthdr.l2hlen + offsetof(struct ip, ip_sum)); *sump = in_cksum_skip(m0, m0->m_pkthdr.l2hlen + m0->m_pkthdr.l3hlen, m0->m_pkthdr.l2hlen); m0->m_pkthdr.csum_flags &= ~CSUM_IP; } #endif cpl = (void *)(wr + 1); } /* Checksum offload */ ctrl1 = 0; if (needs_l3_csum(m0) == 0) ctrl1 |= F_TXPKT_IPCSUM_DIS; if (csum_type >= 0) { KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0, ("%s: mbuf %p needs checksum offload but missing header lengths", __func__, m0)); - /* XXX: T6 */ - ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen - - ETHER_HDR_LEN); + if (chip_id(sc) <= CHELSIO_T5) { + ctrl1 |= V_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen - + ETHER_HDR_LEN); + } else { + ctrl1 |= V_T6_TXPKT_ETHHDR_LEN(m0->m_pkthdr.l2hlen - + ETHER_HDR_LEN); + } ctrl1 |= V_TXPKT_IPHDR_LEN(m0->m_pkthdr.l3hlen); ctrl1 |= V_TXPKT_CSUM_TYPE(csum_type); } else ctrl1 |= F_TXPKT_L4CSUM_DIS; if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) txq->txcsum++; /* some hardware assistance provided */ /* VLAN tag insertion */ if (needs_vlan_insertion(m0)) { ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag); txq->vlan_insertion++; } /* CPL header */ cpl->ctrl0 = txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(pktlen); cpl->ctrl1 = htobe64(ctrl1); /* SGL */ dst = (void *)(cpl + 1); /* * A packet using TSO will use up an entire descriptor for the * firmware work request header, LSO CPL, and TX_PKT_XT CPL. * If this descriptor is the last descriptor in the ring, wrap * around to the front of the ring explicitly for the start of * the sgl. */ if (dst == (void *)&eq->desc[eq->sidx]) { dst = (void *)&eq->desc[0]; write_gl_to_txd(txq, m0, &dst, 0); } else write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx); txq->sgl_wrs++; txq->txpkt_wrs++; txsd = &txq->sdesc[eq->pidx]; txsd->m = m0; txsd->desc_used = ndesc; return (ndesc); } /* * Write a txpkt WR for this packet to the hardware descriptors, update the * software descriptor, and advance the pidx. It is guaranteed that enough * descriptors are available. * * The return value is the # of hardware descriptors used. */ static u_int write_txpkt_wr(struct sge_txq *txq, struct fw_eth_tx_pkt_wr *wr, struct mbuf *m0, u_int available) { struct sge_eq *eq = &txq->eq; struct tx_sdesc *txsd; struct cpl_tx_pkt_core *cpl; uint32_t ctrl; /* used in many unrelated places */ uint64_t ctrl1; int len16, ndesc, pktlen, nsegs; caddr_t dst; TXQ_LOCK_ASSERT_OWNED(txq); M_ASSERTPKTHDR(m0); MPASS(available > 0 && available < eq->sidx); len16 = mbuf_len16(m0); nsegs = mbuf_nsegs(m0); pktlen = m0->m_pkthdr.len; ctrl = sizeof(struct cpl_tx_pkt_core); if (needs_tso(m0)) ctrl += sizeof(struct cpl_tx_pkt_lso_core); else if (pktlen <= imm_payload(2) && available >= 2) { /* Immediate data. Recalculate len16 and set nsegs to 0. */ ctrl += pktlen; len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) + pktlen, 16); nsegs = 0; } ndesc = howmany(len16, EQ_ESIZE / 16); MPASS(ndesc <= available); /* Firmware work request header */ MPASS(wr == (void *)&eq->desc[eq->pidx]); wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) | V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); ctrl = V_FW_WR_LEN16(len16); wr->equiq_to_len16 = htobe32(ctrl); wr->r3 = 0; if (needs_tso(m0)) { struct cpl_tx_pkt_lso_core *lso = (void *)(wr + 1); KASSERT(m0->m_pkthdr.l2hlen > 0 && m0->m_pkthdr.l3hlen > 0 && m0->m_pkthdr.l4hlen > 0, ("%s: mbuf %p needs TSO but missing header lengths", __func__, m0)); ctrl = V_LSO_OPCODE(CPL_TX_PKT_LSO) | F_LSO_FIRST_SLICE | F_LSO_LAST_SLICE | V_LSO_IPHDR_LEN(m0->m_pkthdr.l3hlen >> 2) | V_LSO_TCPHDR_LEN(m0->m_pkthdr.l4hlen >> 2); if (m0->m_pkthdr.l2hlen == sizeof(struct ether_vlan_header)) ctrl |= V_LSO_ETHHDR_LEN(1); if (m0->m_pkthdr.l3hlen == sizeof(struct ip6_hdr)) ctrl |= F_LSO_IPV6; lso->lso_ctrl = htobe32(ctrl); lso->ipid_ofst = htobe16(0); lso->mss = htobe16(m0->m_pkthdr.tso_segsz); lso->seqno_offset = htobe32(0); lso->len = htobe32(pktlen); cpl = (void *)(lso + 1); txq->tso_wrs++; } else cpl = (void *)(wr + 1); /* Checksum offload */ ctrl1 = 0; if (needs_l3_csum(m0) == 0) ctrl1 |= F_TXPKT_IPCSUM_DIS; if (needs_l4_csum(m0) == 0) ctrl1 |= F_TXPKT_L4CSUM_DIS; if (m0->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) txq->txcsum++; /* some hardware assistance provided */ /* VLAN tag insertion */ if (needs_vlan_insertion(m0)) { ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m0->m_pkthdr.ether_vtag); txq->vlan_insertion++; } /* CPL header */ cpl->ctrl0 = txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(pktlen); cpl->ctrl1 = htobe64(ctrl1); /* SGL */ dst = (void *)(cpl + 1); if (nsegs > 0) { write_gl_to_txd(txq, m0, &dst, eq->sidx - ndesc < eq->pidx); txq->sgl_wrs++; } else { struct mbuf *m; for (m = m0; m != NULL; m = m->m_next) { copy_to_txd(eq, mtod(m, caddr_t), &dst, m->m_len); #ifdef INVARIANTS pktlen -= m->m_len; #endif } #ifdef INVARIANTS KASSERT(pktlen == 0, ("%s: %d bytes left.", __func__, pktlen)); #endif txq->imm_wrs++; } txq->txpkt_wrs++; txsd = &txq->sdesc[eq->pidx]; txsd->m = m0; txsd->desc_used = ndesc; return (ndesc); } static int try_txpkts(struct mbuf *m, struct mbuf *n, struct txpkts *txp, u_int available) { u_int needed, nsegs1, nsegs2, l1, l2; if (cannot_use_txpkts(m) || cannot_use_txpkts(n)) return (1); nsegs1 = mbuf_nsegs(m); nsegs2 = mbuf_nsegs(n); if (nsegs1 + nsegs2 == 2) { txp->wr_type = 1; l1 = l2 = txpkts1_len16(); } else { txp->wr_type = 0; l1 = txpkts0_len16(nsegs1); l2 = txpkts0_len16(nsegs2); } txp->len16 = howmany(sizeof(struct fw_eth_tx_pkts_wr), 16) + l1 + l2; needed = howmany(txp->len16, EQ_ESIZE / 16); if (needed > SGE_MAX_WR_NDESC || needed > available) return (1); txp->plen = m->m_pkthdr.len + n->m_pkthdr.len; if (txp->plen > 65535) return (1); txp->npkt = 2; set_mbuf_len16(m, l1); set_mbuf_len16(n, l2); return (0); } static int add_to_txpkts(struct mbuf *m, struct txpkts *txp, u_int available) { u_int plen, len16, needed, nsegs; MPASS(txp->wr_type == 0 || txp->wr_type == 1); nsegs = mbuf_nsegs(m); if (needs_tso(m) || (txp->wr_type == 1 && nsegs != 1)) return (1); plen = txp->plen + m->m_pkthdr.len; if (plen > 65535) return (1); if (txp->wr_type == 0) len16 = txpkts0_len16(nsegs); else len16 = txpkts1_len16(); needed = howmany(txp->len16 + len16, EQ_ESIZE / 16); if (needed > SGE_MAX_WR_NDESC || needed > available) return (1); txp->npkt++; txp->plen = plen; txp->len16 += len16; set_mbuf_len16(m, len16); return (0); } /* * Write a txpkts WR for the packets in txp to the hardware descriptors, update * the software descriptor, and advance the pidx. It is guaranteed that enough * descriptors are available. * * The return value is the # of hardware descriptors used. */ static u_int write_txpkts_wr(struct sge_txq *txq, struct fw_eth_tx_pkts_wr *wr, struct mbuf *m0, const struct txpkts *txp, u_int available) { struct sge_eq *eq = &txq->eq; struct tx_sdesc *txsd; struct cpl_tx_pkt_core *cpl; uint32_t ctrl; uint64_t ctrl1; int ndesc, checkwrap; struct mbuf *m; void *flitp; TXQ_LOCK_ASSERT_OWNED(txq); MPASS(txp->npkt > 0); MPASS(txp->plen < 65536); MPASS(m0 != NULL); MPASS(m0->m_nextpkt != NULL); MPASS(txp->len16 <= howmany(SGE_MAX_WR_LEN, 16)); MPASS(available > 0 && available < eq->sidx); ndesc = howmany(txp->len16, EQ_ESIZE / 16); MPASS(ndesc <= available); MPASS(wr == (void *)&eq->desc[eq->pidx]); wr->op_pkd = htobe32(V_FW_WR_OP(FW_ETH_TX_PKTS_WR)); ctrl = V_FW_WR_LEN16(txp->len16); wr->equiq_to_len16 = htobe32(ctrl); wr->plen = htobe16(txp->plen); wr->npkt = txp->npkt; wr->r3 = 0; wr->type = txp->wr_type; flitp = wr + 1; /* * At this point we are 16B into a hardware descriptor. If checkwrap is * set then we know the WR is going to wrap around somewhere. We'll * check for that at appropriate points. */ checkwrap = eq->sidx - ndesc < eq->pidx; for (m = m0; m != NULL; m = m->m_nextpkt) { if (txp->wr_type == 0) { struct ulp_txpkt *ulpmc; struct ulptx_idata *ulpsc; /* ULP master command */ ulpmc = flitp; ulpmc->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(eq->iqid)); ulpmc->len = htobe32(mbuf_len16(m)); /* ULP subcommand */ ulpsc = (void *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) | F_ULP_TX_SC_MORE); ulpsc->len = htobe32(sizeof(struct cpl_tx_pkt_core)); cpl = (void *)(ulpsc + 1); if (checkwrap && (uintptr_t)cpl == (uintptr_t)&eq->desc[eq->sidx]) cpl = (void *)&eq->desc[0]; txq->txpkts0_pkts += txp->npkt; txq->txpkts0_wrs++; } else { cpl = flitp; txq->txpkts1_pkts += txp->npkt; txq->txpkts1_wrs++; } /* Checksum offload */ ctrl1 = 0; if (needs_l3_csum(m) == 0) ctrl1 |= F_TXPKT_IPCSUM_DIS; if (needs_l4_csum(m) == 0) ctrl1 |= F_TXPKT_L4CSUM_DIS; if (m->m_pkthdr.csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_UDP_IPV6 | CSUM_TCP_IPV6 | CSUM_TSO)) txq->txcsum++; /* some hardware assistance provided */ /* VLAN tag insertion */ if (needs_vlan_insertion(m)) { ctrl1 |= F_TXPKT_VLAN_VLD | V_TXPKT_VLAN(m->m_pkthdr.ether_vtag); txq->vlan_insertion++; } /* CPL header */ cpl->ctrl0 = txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(m->m_pkthdr.len); cpl->ctrl1 = htobe64(ctrl1); flitp = cpl + 1; if (checkwrap && (uintptr_t)flitp == (uintptr_t)&eq->desc[eq->sidx]) flitp = (void *)&eq->desc[0]; write_gl_to_txd(txq, m, (caddr_t *)(&flitp), checkwrap); } txsd = &txq->sdesc[eq->pidx]; txsd->m = m0; txsd->desc_used = ndesc; return (ndesc); } /* * If the SGL ends on an address that is not 16 byte aligned, this function will * add a 0 filled flit at the end. */ static void write_gl_to_txd(struct sge_txq *txq, struct mbuf *m, caddr_t *to, int checkwrap) { struct sge_eq *eq = &txq->eq; struct sglist *gl = txq->gl; struct sglist_seg *seg; __be64 *flitp, *wrap; struct ulptx_sgl *usgl; int i, nflits, nsegs; KASSERT(((uintptr_t)(*to) & 0xf) == 0, ("%s: SGL must start at a 16 byte boundary: %p", __func__, *to)); MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]); MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]); get_pkt_gl(m, gl); nsegs = gl->sg_nseg; MPASS(nsegs > 0); nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2; flitp = (__be64 *)(*to); wrap = (__be64 *)(&eq->desc[eq->sidx]); seg = &gl->sg_segs[0]; usgl = (void *)flitp; /* * We start at a 16 byte boundary somewhere inside the tx descriptor * ring, so we're at least 16 bytes away from the status page. There is * no chance of a wrap around in the middle of usgl (which is 16 bytes). */ usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | V_ULPTX_NSGE(nsegs)); usgl->len0 = htobe32(seg->ss_len); usgl->addr0 = htobe64(seg->ss_paddr); seg++; if (checkwrap == 0 || (uintptr_t)(flitp + nflits) <= (uintptr_t)wrap) { /* Won't wrap around at all */ for (i = 0; i < nsegs - 1; i++, seg++) { usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len); usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr); } if (i & 1) usgl->sge[i / 2].len[1] = htobe32(0); flitp += nflits; } else { /* Will wrap somewhere in the rest of the SGL */ /* 2 flits already written, write the rest flit by flit */ flitp = (void *)(usgl + 1); for (i = 0; i < nflits - 2; i++) { if (flitp == wrap) flitp = (void *)eq->desc; *flitp++ = get_flit(seg, nsegs - 1, i); } } if (nflits & 1) { MPASS(((uintptr_t)flitp) & 0xf); *flitp++ = 0; } MPASS((((uintptr_t)flitp) & 0xf) == 0); if (__predict_false(flitp == wrap)) *to = (void *)eq->desc; else *to = (void *)flitp; } static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len) { MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]); MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]); if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)&eq->desc[eq->sidx])) { bcopy(from, *to, len); (*to) += len; } else { int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to); bcopy(from, *to, portion); from += portion; portion = len - portion; /* remaining */ bcopy(from, (void *)eq->desc, portion); (*to) = (caddr_t)eq->desc + portion; } } static inline void ring_eq_db(struct adapter *sc, struct sge_eq *eq, u_int n) { u_int db; MPASS(n > 0); db = eq->doorbells; if (n > 1) clrbit(&db, DOORBELL_WCWR); wmb(); switch (ffs(db) - 1) { case DOORBELL_UDB: *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n)); break; case DOORBELL_WCWR: { volatile uint64_t *dst, *src; int i; /* * Queues whose 128B doorbell segment fits in the page do not * use relative qid (udb_qid is always 0). Only queues with * doorbell segments can do WCWR. */ KASSERT(eq->udb_qid == 0 && n == 1, ("%s: inappropriate doorbell (0x%x, %d, %d) for eq %p", __func__, eq->doorbells, n, eq->dbidx, eq)); dst = (volatile void *)((uintptr_t)eq->udb + UDBS_WR_OFFSET - UDBS_DB_OFFSET); i = eq->dbidx; src = (void *)&eq->desc[i]; while (src != (void *)&eq->desc[i + 1]) *dst++ = *src++; wmb(); break; } case DOORBELL_UDBWC: *eq->udb = htole32(V_QID(eq->udb_qid) | V_PIDX(n)); wmb(); break; case DOORBELL_KDB: t4_write_reg(sc, sc->sge_kdoorbell_reg, V_QID(eq->cntxt_id) | V_PIDX(n)); break; } IDXINCR(eq->dbidx, n, eq->sidx); } static inline u_int reclaimable_tx_desc(struct sge_eq *eq) { uint16_t hw_cidx; hw_cidx = read_hw_cidx(eq); return (IDXDIFF(hw_cidx, eq->cidx, eq->sidx)); } static inline u_int total_available_tx_desc(struct sge_eq *eq) { uint16_t hw_cidx, pidx; hw_cidx = read_hw_cidx(eq); pidx = eq->pidx; if (pidx == hw_cidx) return (eq->sidx - 1); else return (IDXDIFF(hw_cidx, pidx, eq->sidx) - 1); } static inline uint16_t read_hw_cidx(struct sge_eq *eq) { struct sge_qstat *spg = (void *)&eq->desc[eq->sidx]; uint16_t cidx = spg->cidx; /* stable snapshot */ return (be16toh(cidx)); } /* * Reclaim 'n' descriptors approximately. */ static u_int reclaim_tx_descs(struct sge_txq *txq, u_int n) { struct tx_sdesc *txsd; struct sge_eq *eq = &txq->eq; u_int can_reclaim, reclaimed; TXQ_LOCK_ASSERT_OWNED(txq); MPASS(n > 0); reclaimed = 0; can_reclaim = reclaimable_tx_desc(eq); while (can_reclaim && reclaimed < n) { int ndesc; struct mbuf *m, *nextpkt; txsd = &txq->sdesc[eq->cidx]; ndesc = txsd->desc_used; /* Firmware doesn't return "partial" credits. */ KASSERT(can_reclaim >= ndesc, ("%s: unexpected number of credits: %d, %d", __func__, can_reclaim, ndesc)); for (m = txsd->m; m != NULL; m = nextpkt) { nextpkt = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); } reclaimed += ndesc; can_reclaim -= ndesc; IDXINCR(eq->cidx, ndesc, eq->sidx); } return (reclaimed); } static void tx_reclaim(void *arg, int n) { struct sge_txq *txq = arg; struct sge_eq *eq = &txq->eq; do { if (TXQ_TRYLOCK(txq) == 0) break; n = reclaim_tx_descs(txq, 32); if (eq->cidx == eq->pidx) eq->equeqidx = eq->pidx; TXQ_UNLOCK(txq); } while (n > 0); } static __be64 get_flit(struct sglist_seg *segs, int nsegs, int idx) { int i = (idx / 3) * 2; switch (idx % 3) { case 0: { __be64 rc; rc = htobe32(segs[i].ss_len); if (i + 1 < nsegs) rc |= (uint64_t)htobe32(segs[i + 1].ss_len) << 32; return (rc); } case 1: return (htobe64(segs[i].ss_paddr)); case 2: return (htobe64(segs[i + 1].ss_paddr)); } return (0); } static void find_best_refill_source(struct adapter *sc, struct sge_fl *fl, int maxp) { int8_t zidx, hwidx, idx; uint16_t region1, region3; int spare, spare_needed, n; struct sw_zone_info *swz; struct hw_buf_info *hwb, *hwb_list = &sc->sge.hw_buf_info[0]; /* * Buffer Packing: Look for PAGE_SIZE or larger zone which has a bufsize * large enough for the max payload and cluster metadata. Otherwise * settle for the largest bufsize that leaves enough room in the cluster * for metadata. * * Without buffer packing: Look for the smallest zone which has a * bufsize large enough for the max payload. Settle for the largest * bufsize available if there's nothing big enough for max payload. */ spare_needed = fl->flags & FL_BUF_PACKING ? CL_METADATA_SIZE : 0; swz = &sc->sge.sw_zone_info[0]; hwidx = -1; for (zidx = 0; zidx < SW_ZONE_SIZES; zidx++, swz++) { if (swz->size > largest_rx_cluster) { if (__predict_true(hwidx != -1)) break; /* * This is a misconfiguration. largest_rx_cluster is * preventing us from finding a refill source. See * dev.t5nex..buffer_sizes to figure out why. */ device_printf(sc->dev, "largest_rx_cluster=%u leaves no" " refill source for fl %p (dma %u). Ignored.\n", largest_rx_cluster, fl, maxp); } for (idx = swz->head_hwidx; idx != -1; idx = hwb->next) { hwb = &hwb_list[idx]; spare = swz->size - hwb->size; if (spare < spare_needed) continue; hwidx = idx; /* best option so far */ if (hwb->size >= maxp) { if ((fl->flags & FL_BUF_PACKING) == 0) goto done; /* stop looking (not packing) */ if (swz->size >= safest_rx_cluster) goto done; /* stop looking (packing) */ } break; /* keep looking, next zone */ } } done: /* A usable hwidx has been located. */ MPASS(hwidx != -1); hwb = &hwb_list[hwidx]; zidx = hwb->zidx; swz = &sc->sge.sw_zone_info[zidx]; region1 = 0; region3 = swz->size - hwb->size; /* * Stay within this zone and see if there is a better match when mbuf * inlining is allowed. Remember that the hwidx's are sorted in * decreasing order of size (so in increasing order of spare area). */ for (idx = hwidx; idx != -1; idx = hwb->next) { hwb = &hwb_list[idx]; spare = swz->size - hwb->size; if (allow_mbufs_in_cluster == 0 || hwb->size < maxp) break; /* * Do not inline mbufs if doing so would violate the pad/pack * boundary alignment requirement. */ if (fl_pad && (MSIZE % sc->params.sge.pad_boundary) != 0) continue; if (fl->flags & FL_BUF_PACKING && (MSIZE % sc->params.sge.pack_boundary) != 0) continue; if (spare < CL_METADATA_SIZE + MSIZE) continue; n = (spare - CL_METADATA_SIZE) / MSIZE; if (n > howmany(hwb->size, maxp)) break; hwidx = idx; if (fl->flags & FL_BUF_PACKING) { region1 = n * MSIZE; region3 = spare - region1; } else { region1 = MSIZE; region3 = spare - region1; break; } } KASSERT(zidx >= 0 && zidx < SW_ZONE_SIZES, ("%s: bad zone %d for fl %p, maxp %d", __func__, zidx, fl, maxp)); KASSERT(hwidx >= 0 && hwidx <= SGE_FLBUF_SIZES, ("%s: bad hwidx %d for fl %p, maxp %d", __func__, hwidx, fl, maxp)); KASSERT(region1 + sc->sge.hw_buf_info[hwidx].size + region3 == sc->sge.sw_zone_info[zidx].size, ("%s: bad buffer layout for fl %p, maxp %d. " "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp, sc->sge.sw_zone_info[zidx].size, region1, sc->sge.hw_buf_info[hwidx].size, region3)); if (fl->flags & FL_BUF_PACKING || region1 > 0) { KASSERT(region3 >= CL_METADATA_SIZE, ("%s: no room for metadata. fl %p, maxp %d; " "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp, sc->sge.sw_zone_info[zidx].size, region1, sc->sge.hw_buf_info[hwidx].size, region3)); KASSERT(region1 % MSIZE == 0, ("%s: bad mbuf region for fl %p, maxp %d. " "cl %d; r1 %d, payload %d, r3 %d", __func__, fl, maxp, sc->sge.sw_zone_info[zidx].size, region1, sc->sge.hw_buf_info[hwidx].size, region3)); } fl->cll_def.zidx = zidx; fl->cll_def.hwidx = hwidx; fl->cll_def.region1 = region1; fl->cll_def.region3 = region3; } static void find_safe_refill_source(struct adapter *sc, struct sge_fl *fl) { struct sge *s = &sc->sge; struct hw_buf_info *hwb; struct sw_zone_info *swz; int spare; int8_t hwidx; if (fl->flags & FL_BUF_PACKING) hwidx = s->safe_hwidx2; /* with room for metadata */ else if (allow_mbufs_in_cluster && s->safe_hwidx2 != -1) { hwidx = s->safe_hwidx2; hwb = &s->hw_buf_info[hwidx]; swz = &s->sw_zone_info[hwb->zidx]; spare = swz->size - hwb->size; /* no good if there isn't room for an mbuf as well */ if (spare < CL_METADATA_SIZE + MSIZE) hwidx = s->safe_hwidx1; } else hwidx = s->safe_hwidx1; if (hwidx == -1) { /* No fallback source */ fl->cll_alt.hwidx = -1; fl->cll_alt.zidx = -1; return; } hwb = &s->hw_buf_info[hwidx]; swz = &s->sw_zone_info[hwb->zidx]; spare = swz->size - hwb->size; fl->cll_alt.hwidx = hwidx; fl->cll_alt.zidx = hwb->zidx; if (allow_mbufs_in_cluster && (fl_pad == 0 || (MSIZE % sc->params.sge.pad_boundary) == 0)) fl->cll_alt.region1 = ((spare - CL_METADATA_SIZE) / MSIZE) * MSIZE; else fl->cll_alt.region1 = 0; fl->cll_alt.region3 = spare - fl->cll_alt.region1; } static void add_fl_to_sfl(struct adapter *sc, struct sge_fl *fl) { mtx_lock(&sc->sfl_lock); FL_LOCK(fl); if ((fl->flags & FL_DOOMED) == 0) { fl->flags |= FL_STARVING; TAILQ_INSERT_TAIL(&sc->sfl, fl, link); callout_reset(&sc->sfl_callout, hz / 5, refill_sfl, sc); } FL_UNLOCK(fl); mtx_unlock(&sc->sfl_lock); } static void handle_wrq_egr_update(struct adapter *sc, struct sge_eq *eq) { struct sge_wrq *wrq = (void *)eq; atomic_readandclear_int(&eq->equiq); taskqueue_enqueue(sc->tq[eq->tx_chan], &wrq->wrq_tx_task); } static void handle_eth_egr_update(struct adapter *sc, struct sge_eq *eq) { struct sge_txq *txq = (void *)eq; MPASS((eq->flags & EQ_TYPEMASK) == EQ_ETH); atomic_readandclear_int(&eq->equiq); mp_ring_check_drainage(txq->r, 0); taskqueue_enqueue(sc->tq[eq->tx_chan], &txq->tx_reclaim_task); } static int handle_sge_egr_update(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { const struct cpl_sge_egr_update *cpl = (const void *)(rss + 1); unsigned int qid = G_EGR_QID(ntohl(cpl->opcode_qid)); struct adapter *sc = iq->adapter; struct sge *s = &sc->sge; struct sge_eq *eq; static void (*h[])(struct adapter *, struct sge_eq *) = {NULL, &handle_wrq_egr_update, &handle_eth_egr_update, &handle_wrq_egr_update}; KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, rss->opcode)); eq = s->eqmap[qid - s->eq_start - s->eq_base]; (*h[eq->flags & EQ_TYPEMASK])(sc, eq); return (0); } /* handle_fw_msg works for both fw4_msg and fw6_msg because this is valid */ CTASSERT(offsetof(struct cpl_fw4_msg, data) == \ offsetof(struct cpl_fw6_msg, data)); static int handle_fw_msg(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_fw6_msg *cpl = (const void *)(rss + 1); KASSERT(m == NULL, ("%s: payload with opcode %02x", __func__, rss->opcode)); if (cpl->type == FW_TYPE_RSSCPL || cpl->type == FW6_TYPE_RSSCPL) { const struct rss_header *rss2; rss2 = (const struct rss_header *)&cpl->data[0]; return (t4_cpl_handler[rss2->opcode](iq, rss2, m)); } return (t4_fw_msg_handler[cpl->type](sc, &cpl->data[0])); } /** * t4_handle_wrerr_rpl - process a FW work request error message * @adap: the adapter * @rpl: start of the FW message */ static int t4_handle_wrerr_rpl(struct adapter *adap, const __be64 *rpl) { u8 opcode = *(const u8 *)rpl; const struct fw_error_cmd *e = (const void *)rpl; unsigned int i; if (opcode != FW_ERROR_CMD) { log(LOG_ERR, "%s: Received WRERR_RPL message with opcode %#x\n", device_get_nameunit(adap->dev), opcode); return (EINVAL); } log(LOG_ERR, "%s: FW_ERROR (%s) ", device_get_nameunit(adap->dev), G_FW_ERROR_CMD_FATAL(be32toh(e->op_to_type)) ? "fatal" : "non-fatal"); switch (G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))) { case FW_ERROR_TYPE_EXCEPTION: log(LOG_ERR, "exception info:\n"); for (i = 0; i < nitems(e->u.exception.info); i++) log(LOG_ERR, "%s%08x", i == 0 ? "\t" : " ", be32toh(e->u.exception.info[i])); log(LOG_ERR, "\n"); break; case FW_ERROR_TYPE_HWMODULE: log(LOG_ERR, "HW module regaddr %08x regval %08x\n", be32toh(e->u.hwmodule.regaddr), be32toh(e->u.hwmodule.regval)); break; case FW_ERROR_TYPE_WR: log(LOG_ERR, "WR cidx %d PF %d VF %d eqid %d hdr:\n", be16toh(e->u.wr.cidx), G_FW_ERROR_CMD_PFN(be16toh(e->u.wr.pfn_vfn)), G_FW_ERROR_CMD_VFN(be16toh(e->u.wr.pfn_vfn)), be32toh(e->u.wr.eqid)); for (i = 0; i < nitems(e->u.wr.wrhdr); i++) log(LOG_ERR, "%s%02x", i == 0 ? "\t" : " ", e->u.wr.wrhdr[i]); log(LOG_ERR, "\n"); break; case FW_ERROR_TYPE_ACL: log(LOG_ERR, "ACL cidx %d PF %d VF %d eqid %d %s", be16toh(e->u.acl.cidx), G_FW_ERROR_CMD_PFN(be16toh(e->u.acl.pfn_vfn)), G_FW_ERROR_CMD_VFN(be16toh(e->u.acl.pfn_vfn)), be32toh(e->u.acl.eqid), G_FW_ERROR_CMD_MV(be16toh(e->u.acl.mv_pkd)) ? "vlanid" : "MAC"); for (i = 0; i < nitems(e->u.acl.val); i++) log(LOG_ERR, " %02x", e->u.acl.val[i]); log(LOG_ERR, "\n"); break; default: log(LOG_ERR, "type %#x\n", G_FW_ERROR_CMD_TYPE(be32toh(e->op_to_type))); return (EINVAL); } return (0); } static int sysctl_uint16(SYSCTL_HANDLER_ARGS) { uint16_t *id = arg1; int i = *id; return sysctl_handle_int(oidp, &i, 0, req); } static int sysctl_bufsizes(SYSCTL_HANDLER_ARGS) { struct sge *s = arg1; struct hw_buf_info *hwb = &s->hw_buf_info[0]; struct sw_zone_info *swz = &s->sw_zone_info[0]; int i, rc; struct sbuf sb; char c; sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND); for (i = 0; i < SGE_FLBUF_SIZES; i++, hwb++) { if (hwb->zidx >= 0 && swz[hwb->zidx].size <= largest_rx_cluster) c = '*'; else c = '\0'; sbuf_printf(&sb, "%u%c ", hwb->size, c); } sbuf_trim(&sb); sbuf_finish(&sb); rc = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); sbuf_delete(&sb); return (rc); } static int sysctl_tc(SYSCTL_HANDLER_ARGS) { struct vi_info *vi = arg1; struct port_info *pi; struct adapter *sc; struct sge_txq *txq; struct tx_sched_class *tc; int qidx = arg2, rc, tc_idx; uint32_t fw_queue, fw_class; MPASS(qidx >= 0 && qidx < vi->ntxq); pi = vi->pi; sc = pi->adapter; txq = &sc->sge.txq[vi->first_txq + qidx]; tc_idx = txq->tc_idx; rc = sysctl_handle_int(oidp, &tc_idx, 0, req); if (rc != 0 || req->newptr == NULL) return (rc); /* Note that -1 is legitimate input (it means unbind). */ if (tc_idx < -1 || tc_idx >= sc->chip_params->nsched_cls) return (EINVAL); rc = begin_synchronized_op(sc, vi, SLEEP_OK | INTR_OK, "t4stc"); if (rc) return (rc); if (tc_idx == txq->tc_idx) { rc = 0; /* No change, nothing to do. */ goto done; } fw_queue = V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) | V_FW_PARAMS_PARAM_X(FW_PARAMS_PARAM_DMAQ_EQ_SCHEDCLASS_ETH) | V_FW_PARAMS_PARAM_YZ(txq->eq.cntxt_id); if (tc_idx == -1) fw_class = 0xffffffff; /* Unbind. */ else { /* * Bind to a different class. Ethernet txq's are only allowed * to bind to cl-rl mode-class for now. XXX: too restrictive. */ tc = &pi->tc[tc_idx]; if (tc->flags & TX_SC_OK && tc->params.level == SCHED_CLASS_LEVEL_CL_RL && tc->params.mode == SCHED_CLASS_MODE_CLASS) { /* Ok to proceed. */ fw_class = tc_idx; } else { rc = tc->flags & TX_SC_OK ? EBUSY : ENXIO; goto done; } } rc = -t4_set_params(sc, sc->mbox, sc->pf, 0, 1, &fw_queue, &fw_class); if (rc == 0) { if (txq->tc_idx != -1) { tc = &pi->tc[txq->tc_idx]; MPASS(tc->refcount > 0); tc->refcount--; } if (tc_idx != -1) { tc = &pi->tc[tc_idx]; tc->refcount++; } txq->tc_idx = tc_idx; } done: end_synchronized_op(sc, 0); return (rc); } Index: projects/mips64-clang/sys/i386/i386/pmap.c =================================================================== --- projects/mips64-clang/sys/i386/i386/pmap.c (revision 305711) +++ projects/mips64-clang/sys/i386/i386/pmap.c (revision 305712) @@ -1,5622 +1,5616 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_apic.h" #include "opt_cpu.h" #include "opt_pmap.h" #include "opt_smp.h" #include "opt_xbox.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DEV_APIC #include #include #include #endif #include #include #include #include #include #ifdef SMP #include #endif #ifdef XBOX #include #endif #if !defined(CPU_DISABLE_SSE) && defined(I686_CPU) #define CPU_ENABLE_SSE #endif #ifndef PMAP_SHPGPERPROC #define PMAP_SHPGPERPROC 200 #endif #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define pa_to_pvh(pa) (&pv_table[pa_index(pa)]) /* * Get PDEs and PTEs for user/kernel address space */ #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) #define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ atomic_clear_int((u_int *)(pte), PG_W)) #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) struct pmap kernel_pmap_store; LIST_HEAD(pmaplist, pmap); static struct pmaplist allpmaps; static struct mtx allpmaps_lock; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ int pgeflag = 0; /* PG_G or-in */ int pseflag = 0; /* PG_PS or-in */ static int nkpt = NKPT; vm_offset_t kernel_vm_end = KERNBASE + NKPT * NBPDR; extern u_int32_t KERNend; extern u_int32_t KPTphys; #if defined(PAE) || defined(PAE_TABLES) pt_entry_t pg_nx; static uma_zone_t pdptzone; #endif static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD, 0, "VM/pmap parameters"); static int pat_works = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pat_works, CTLFLAG_RD, &pat_works, 1, "Is page attribute table fully functional?"); static int pg_ps_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pg_ps_enabled, 0, "Are large page mappings enabled?"); #define PAT_INDEX_SIZE 8 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */ /* * pmap_mapdev support pre initialization (i.e. console) */ #define PMAP_PREINIT_MAPPING_COUNT 8 static struct pmap_preinit_mapping { vm_paddr_t pa; vm_offset_t va; vm_size_t sz; int mode; } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT]; static int pmap_initialized; static struct rwlock_padalign pvh_global_lock; /* * Data for the pv entry allocation mechanism */ static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks); static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; static struct md_page *pv_table; static int shpgperproc = PMAP_SHPGPERPROC; struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */ int pv_maxchunks; /* How many chunks we have KVA for */ vm_offset_t pv_vafree; /* freelist stored in the PTE */ /* * All those kernel PT submaps that BSD is so fond of */ struct sysmaps { struct mtx lock; pt_entry_t *CMAP1; pt_entry_t *CMAP2; caddr_t CADDR1; caddr_t CADDR2; }; static struct sysmaps sysmaps_pcpu[MAXCPU]; pt_entry_t *CMAP3; static pd_entry_t *KPTD; caddr_t ptvmmap = 0; caddr_t CADDR3; struct msgbuf *msgbufp = NULL; /* * Crashdump maps. */ static caddr_t crashdumpmap; static pt_entry_t *PMAP1 = NULL, *PMAP2; static pt_entry_t *PADDR1 = NULL, *PADDR2; #ifdef SMP static int PMAP1cpu; static int PMAP1changedcpu; SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD, &PMAP1changedcpu, 0, "Number of times pmap_pte_quick changed CPU with same PMAP1"); #endif static int PMAP1changed; SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD, &PMAP1changed, 0, "Number of times pmap_pte_quick changed PMAP1"); static int PMAP1unchanged; SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD, &PMAP1unchanged, 0, "Number of times pmap_pte_quick didn't change PMAP1"); static struct mtx PMAP2mutex; static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, boolean_t try); static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); static boolean_t pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static int pmap_pvh_wired_mappings(struct md_page *pvh, int count); static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); static boolean_t pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte); static void pmap_flush_page(vm_page_t m); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static boolean_t pmap_is_modified_pvh(struct md_page *pvh); static boolean_t pmap_is_referenced_pvh(struct md_page *pvh); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); static void pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde); static vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va); static void pmap_pde_attr(pd_entry_t *pde, int cache_bits); static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot); static void pmap_pte_attr(pt_entry_t *pte, int cache_bits); static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free); static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva, struct spglist *free); static void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte); static void pmap_remove_page(struct pmap *pmap, vm_offset_t va, struct spglist *free); static void pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va); static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde); static void pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde); static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, u_int flags); static vm_page_t _pmap_allocpte(pmap_t pmap, u_int ptepindex, u_int flags); static void _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free); static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); static void pmap_pte_release(pt_entry_t *pte); static int pmap_unuse_pt(pmap_t, vm_offset_t, struct spglist *); #if defined(PAE) || defined(PAE_TABLES) static void *pmap_pdpt_allocf(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait); #endif static void pmap_set_pg(void); static __inline void pagezero(void *page); CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); /* * If you get an error here, then you set KVA_PAGES wrong! See the * description of KVA_PAGES in sys/i386/include/pmap.h. It must be * multiple of 4 for a normal kernel, or a multiple of 8 for a PAE. */ CTASSERT(KERNBASE % (1 << 24) == 0); /* * Bootstrap the system enough to run with virtual memory. * * On the i386 this is called after mapping has already been enabled * and just syncs the pmap module with what has already been done. * [We can't call it easily with mapping off since the kernel is not * mapped with PA == VA, hence we would have to relocate every address * from the linked base (virtual) address "KERNBASE" to the actual * (physical) address starting relative to 0] */ void pmap_bootstrap(vm_paddr_t firstaddr) { vm_offset_t va; pt_entry_t *pte, *unused; struct sysmaps *sysmaps; int i; /* * Add a physical memory segment (vm_phys_seg) corresponding to the * preallocated kernel page table pages so that vm_page structures * representing these pages will be created. The vm_page structures * are required for promotion of the corresponding kernel virtual * addresses to superpage mappings. */ vm_phys_add_seg(KPTphys, KPTphys + ptoa(nkpt)); /* * Initialize the first available kernel virtual address. However, * using "firstaddr" may waste a few pages of the kernel virtual * address space, because locore may not have mapped every physical * page that it allocated. Preferably, locore would provide a first * unused virtual address in addition to "firstaddr". */ virtual_avail = (vm_offset_t) KERNBASE + firstaddr; virtual_end = VM_MAX_KERNEL_ADDRESS; /* * Initialize the kernel pmap (which is statically allocated). */ PMAP_LOCK_INIT(kernel_pmap); kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD); #if defined(PAE) || defined(PAE_TABLES) kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT); #endif CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */ TAILQ_INIT(&kernel_pmap->pm_pvchunk); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); LIST_INIT(&allpmaps); /* * Request a spin mutex so that changes to allpmaps cannot be * preempted by smp_rendezvous_cpus(). Otherwise, * pmap_update_pde_kernel() could access allpmaps while it is * being changed. */ mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN); mtx_lock_spin(&allpmaps_lock); LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list); mtx_unlock_spin(&allpmaps_lock); /* * Reserve some special page table entries/VA space for temporary * mapping of pages. */ #define SYSMAP(c, p, v, n) \ v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); va = virtual_avail; pte = vtopte(va); /* * CMAP1/CMAP2 are used for zeroing and copying pages. * CMAP3 is used for the boot-time memory test. */ for (i = 0; i < MAXCPU; i++) { sysmaps = &sysmaps_pcpu[i]; mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF); SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1) SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1) } SYSMAP(caddr_t, CMAP3, CADDR3, 1); /* * Crashdump maps. */ SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) /* * ptvmmap is used for reading arbitrary physical pages via /dev/mem. */ SYSMAP(caddr_t, unused, ptvmmap, 1) /* * msgbufp is used to map the system message buffer. */ SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(msgbufsize))) /* * KPTmap is used by pmap_kextract(). * * KPTmap is first initialized by locore. However, that initial * KPTmap can only support NKPT page table pages. Here, a larger * KPTmap is created that can support KVA_PAGES page table pages. */ SYSMAP(pt_entry_t *, KPTD, KPTmap, KVA_PAGES) for (i = 0; i < NKPT; i++) KPTD[i] = (KPTphys + (i << PAGE_SHIFT)) | pgeflag | PG_RW | PG_V; /* * Adjust the start of the KPTD and KPTmap so that the implementation * of pmap_kextract() and pmap_growkernel() can be made simpler. */ KPTD -= KPTDI; KPTmap -= i386_btop(KPTDI << PDRSHIFT); /* * PADDR1 and PADDR2 are used by pmap_pte_quick() and pmap_pte(), * respectively. */ SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1) SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1) mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); virtual_avail = va; /* * Leave in place an identity mapping (virt == phys) for the low 1 MB * physical memory region that is used by the ACPI wakeup code. This * mapping must not have PG_G set. */ #ifdef XBOX /* FIXME: This is gross, but needed for the XBOX. Since we are in such * an early stadium, we cannot yet neatly map video memory ... :-( * Better fixes are very welcome! */ if (!arch_i386_is_xbox) #endif for (i = 1; i < NKPT; i++) PTD[i] = 0; /* Initialize the PAT MSR if present. */ pmap_init_pat(); /* Turn on PG_G on kernel page(s) */ pmap_set_pg(); } static void pmap_init_qpages(void) { struct pcpu *pc; int i; CPU_FOREACH(i) { pc = pcpu_find(i); pc->pc_qmap_addr = kva_alloc(PAGE_SIZE); if (pc->pc_qmap_addr == 0) panic("pmap_init_qpages: unable to allocate KVA"); } } SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, pmap_init_qpages, NULL); /* * Setup the PAT MSR. */ void pmap_init_pat(void) { int pat_table[PAT_INDEX_SIZE]; uint64_t pat_msr; u_long cr0, cr4; int i; /* Set default PAT index table. */ for (i = 0; i < PAT_INDEX_SIZE; i++) pat_table[i] = -1; pat_table[PAT_WRITE_BACK] = 0; pat_table[PAT_WRITE_THROUGH] = 1; pat_table[PAT_UNCACHEABLE] = 3; pat_table[PAT_WRITE_COMBINING] = 3; pat_table[PAT_WRITE_PROTECTED] = 3; pat_table[PAT_UNCACHED] = 3; /* Bail if this CPU doesn't implement PAT. */ if ((cpu_feature & CPUID_PAT) == 0) { for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = pat_table[i]; pat_works = 0; return; } /* * Due to some Intel errata, we can only safely use the lower 4 * PAT entries. * * Intel Pentium III Processor Specification Update * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B * or Mode C Paging) * * Intel Pentium IV Processor Specification Update * Errata N46 (PAT Index MSB May Be Calculated Incorrectly) */ if (cpu_vendor_id == CPU_VENDOR_INTEL && !(CPUID_TO_FAMILY(cpu_id) == 6 && CPUID_TO_MODEL(cpu_id) >= 0xe)) pat_works = 0; /* Initialize default PAT entries. */ pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | PAT_VALUE(1, PAT_WRITE_THROUGH) | PAT_VALUE(2, PAT_UNCACHED) | PAT_VALUE(3, PAT_UNCACHEABLE) | PAT_VALUE(4, PAT_WRITE_BACK) | PAT_VALUE(5, PAT_WRITE_THROUGH) | PAT_VALUE(6, PAT_UNCACHED) | PAT_VALUE(7, PAT_UNCACHEABLE); if (pat_works) { /* * Leave the indices 0-3 at the default of WB, WT, UC-, and UC. * Program 5 and 6 as WP and WC. * Leave 4 and 7 as WB and UC. */ pat_msr &= ~(PAT_MASK(5) | PAT_MASK(6)); pat_msr |= PAT_VALUE(5, PAT_WRITE_PROTECTED) | PAT_VALUE(6, PAT_WRITE_COMBINING); pat_table[PAT_UNCACHED] = 2; pat_table[PAT_WRITE_PROTECTED] = 5; pat_table[PAT_WRITE_COMBINING] = 6; } else { /* * Just replace PAT Index 2 with WC instead of UC-. */ pat_msr &= ~PAT_MASK(2); pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING); pat_table[PAT_WRITE_COMBINING] = 2; } /* Disable PGE. */ cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); /* Disable caches (CD = 1, NW = 0). */ cr0 = rcr0(); load_cr0((cr0 & ~CR0_NW) | CR0_CD); /* Flushes caches and TLBs. */ wbinvd(); invltlb(); /* Update PAT and index table. */ wrmsr(MSR_PAT, pat_msr); for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = pat_table[i]; /* Flush caches and TLBs again. */ wbinvd(); invltlb(); /* Restore caches and PGE. */ load_cr0(cr0); load_cr4(cr4); } /* * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on. */ static void pmap_set_pg(void) { pt_entry_t *pte; vm_offset_t va, endva; if (pgeflag == 0) return; endva = KERNBASE + KERNend; if (pseflag) { va = KERNBASE + KERNLOAD; while (va < endva) { pdir_pde(PTD, va) |= pgeflag; invltlb(); /* Flush non-PG_G entries. */ va += NBPDR; } } else { va = (vm_offset_t)btext; while (va < endva) { pte = vtopte(va); if (*pte) *pte |= pgeflag; invltlb(); /* Flush non-PG_G entries. */ va += PAGE_SIZE; } } } /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pat_mode = PAT_WRITE_BACK; } #if defined(PAE) || defined(PAE_TABLES) static void * pmap_pdpt_allocf(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) { /* Inform UMA that this allocator uses kernel_map/object. */ *flags = UMA_SLAB_KERNEL; return ((void *)kmem_alloc_contig(kernel_arena, bytes, wait, 0x0ULL, 0xffffffffULL, 1, 0, VM_MEMATTR_DEFAULT)); } #endif /* * Abuse the pte nodes for unmapped kva to thread a kva freelist through. * Requirements: * - Must deal with pages in order to ensure that none of the PG_* bits * are ever set, PG_V in particular. * - Assumes we can write to ptes without pte_store() atomic ops, even * on PAE systems. This should be ok. * - Assumes nothing will ever test these addresses for 0 to indicate * no mapping instead of correctly checking PG_V. * - Assumes a vm_offset_t will fit in a pte (true for i386). * Because PG_V is never set, there can be no mappings to invalidate. */ static vm_offset_t pmap_ptelist_alloc(vm_offset_t *head) { pt_entry_t *pte; vm_offset_t va; va = *head; if (va == 0) panic("pmap_ptelist_alloc: exhausted ptelist KVA"); pte = vtopte(va); *head = *pte; if (*head & PG_V) panic("pmap_ptelist_alloc: va with PG_V set!"); *pte = 0; return (va); } static void pmap_ptelist_free(vm_offset_t *head, vm_offset_t va) { pt_entry_t *pte; if (va & PG_V) panic("pmap_ptelist_free: freeing va with PG_V set!"); pte = vtopte(va); *pte = *head; /* virtual! PG_V is 0 though */ *head = va; } static void pmap_ptelist_init(vm_offset_t *head, void *base, int npages) { int i; vm_offset_t va; *head = 0; for (i = npages - 1; i >= 0; i--) { va = (vm_offset_t)base + i * PAGE_SIZE; pmap_ptelist_free(head, va); } } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { struct pmap_preinit_mapping *ppim; vm_page_t mpte; vm_size_t s; int i, pv_npg; /* * Initialize the vm page array entries for the kernel pmap's * page table pages. */ for (i = 0; i < NKPT; i++) { mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT)); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_init: page table page is out of range")); mpte->pindex = i + KPTDI; mpte->phys_addr = KPTphys + (i << PAGE_SHIFT); } /* * Initialize the address space (zone) for the pv entries. Set a * high water mark so that the system can recover from excessive * numbers of pv entries. */ TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count; TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); pv_entry_max = roundup(pv_entry_max, _NPCPV); pv_entry_high_water = 9 * (pv_entry_max / 10); /* * If the kernel is running on a virtual machine, then it must assume * that MCA is enabled by the hypervisor. Moreover, the kernel must * be prepared for the hypervisor changing the vendor and family that * are reported by CPUID. Consequently, the workaround for AMD Family * 10h Erratum 383 is enabled if the processor's feature set does not * include at least one feature that is only supported by older Intel * or newer AMD processors. */ if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 && (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI | CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP | AMDID2_FMA4)) == 0) workaround_erratum383 = 1; /* * Are large page mappings supported and enabled? */ TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled); if (pseflag == 0) pg_ps_enabled = 0; else if (pg_ps_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = NBPDR; } /* * Calculate the size of the pv head table for superpages. * Handle the possibility that "vm_phys_segs[...].end" is zero. */ pv_npg = trunc_4mpage(vm_phys_segs[vm_phys_nsegs - 1].end - PAGE_SIZE) / NBPDR + 1; /* * Allocate memory for the pv head table for superpages. */ s = (vm_size_t)(pv_npg * sizeof(struct md_page)); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(kernel_arena, s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc); pv_chunkbase = (struct pv_chunk *)kva_alloc(PAGE_SIZE * pv_maxchunks); if (pv_chunkbase == NULL) panic("pmap_init: not enough kvm for pv chunks"); pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks); #if defined(PAE) || defined(PAE_TABLES) pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, UMA_ZONE_VM | UMA_ZONE_NOFREE); uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); #endif pmap_initialized = 1; if (!bootverbose) return; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) continue; printf("PPIM %u: PA=%#jx, VA=%#x, size=%#x, mode=%#x\n", i, (uintmax_t)ppim->pa, ppim->va, ppim->sz, ppim->mode); } } SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_max, CTLFLAG_RD, &pv_entry_max, 0, "Max number of PV entries"); SYSCTL_INT(_vm_pmap, OID_AUTO, shpgperproc, CTLFLAG_RD, &shpgperproc, 0, "Page share factor per proc"); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD, 0, "2/4MB page mapping counters"); static u_long pmap_pde_demotions; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pde_demotions, 0, "2/4MB page demotions"); static u_long pmap_pde_mappings; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD, &pmap_pde_mappings, 0, "2/4MB page mappings"); static u_long pmap_pde_p_failures; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_pde_p_failures, 0, "2/4MB page promotion failures"); static u_long pmap_pde_promotions; SYSCTL_ULONG(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD, &pmap_pde_promotions, 0, "2/4MB page promotions"); /*************************************************** * Low level helper routines..... ***************************************************/ /* * Determine the appropriate bits to set in a PTE or PDE for a specified * caching mode. */ int pmap_cache_bits(int mode, boolean_t is_pde) { int cache_bits, pat_flag, pat_idx; if (mode < 0 || mode >= PAT_INDEX_SIZE || pat_index[mode] < 0) panic("Unknown caching mode %d\n", mode); /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT; /* Map the caching mode to a PAT index. */ pat_idx = pat_index[mode]; /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */ cache_bits = 0; if (pat_idx & 0x4) cache_bits |= pat_flag; if (pat_idx & 0x2) cache_bits |= PG_NC_PCD; if (pat_idx & 0x1) cache_bits |= PG_NC_PWT; return (cache_bits); } /* * The caller is responsible for maintaining TLB consistency. */ static void pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde) { pd_entry_t *pde; pmap_t pmap; boolean_t PTD_updated; PTD_updated = FALSE; mtx_lock_spin(&allpmaps_lock); LIST_FOREACH(pmap, &allpmaps, pm_list) { if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) PTD_updated = TRUE; pde = pmap_pde(pmap, va); pde_store(pde, newpde); } mtx_unlock_spin(&allpmaps_lock); KASSERT(PTD_updated, ("pmap_kenter_pde: current page table is not in allpmaps")); } /* * After changing the page size for the specified virtual address in the page * table, flush the corresponding entries from the processor's TLB. Only the * calling processor's TLB is affected. * * The calling thread must be pinned to a processor. */ static void pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde) { u_long cr4; if ((newpde & PG_PS) == 0) /* Demotion: flush a specific 2MB page mapping. */ invlpg(va); else if ((newpde & PG_G) == 0) /* * Promotion: flush every 4KB page mapping from the TLB * because there are too many to flush individually. */ invltlb(); else { /* * Promotion: flush every 4KB page mapping from the TLB, * including any global (PG_G) mappings. */ cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); /* * Although preemption at this point could be detrimental to * performance, it would not lead to an error. PG_G is simply * ignored if CR4.PGE is clear. Moreover, in case this block * is re-entered, the load_cr4() either above or below will * modify CR4.PGE flushing the TLB. */ load_cr4(cr4 | CR4_PGE); } } void invltlb_glob(void) { uint64_t cr4; if (pgeflag == 0) { invltlb(); } else { cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); load_cr4(cr4 | CR4_PGE); } } #ifdef SMP /* * For SMP, these functions have to use the IPI mechanism for coherence. * * N.B.: Before calling any of the following TLB invalidation functions, * the calling processor must ensure that all stores updating a non- * kernel page table are globally performed. Otherwise, another * processor could cache an old, pre-update entry without being * invalidated. This can happen one of two ways: (1) The pmap becomes * active on another processor after its pm_active field is checked by * one of the following functions but before a store updating the page * table is globally performed. (2) The pmap becomes active on another * processor before its pm_active field is checked but due to * speculative loads one of the following functions stills reads the * pmap as inactive on the other processor. * * The kernel page table is exempt because its pm_active field is * immutable. The kernel page table is always active on every * processor. */ void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { cpuset_t *mask, other_cpus; u_int cpuid; sched_pin(); if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) { invlpg(va); mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (CPU_ISSET(cpuid, &pmap->pm_active)) invlpg(va); CPU_AND(&other_cpus, &pmap->pm_active); mask = &other_cpus; } smp_masked_invlpg(*mask, va); sched_unpin(); } /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */ #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE) void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { cpuset_t *mask, other_cpus; vm_offset_t addr; u_int cpuid; if (eva - sva >= PMAP_INVLPG_THRESHOLD) { pmap_invalidate_all(pmap); return; } sched_pin(); if (pmap == kernel_pmap || !CPU_CMP(&pmap->pm_active, &all_cpus)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (CPU_ISSET(cpuid, &pmap->pm_active)) for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); CPU_AND(&other_cpus, &pmap->pm_active); mask = &other_cpus; } smp_masked_invlpg_range(*mask, sva, eva); sched_unpin(); } void pmap_invalidate_all(pmap_t pmap) { cpuset_t *mask, other_cpus; u_int cpuid; sched_pin(); if (pmap == kernel_pmap) { invltlb_glob(); mask = &all_cpus; } else if (!CPU_CMP(&pmap->pm_active, &all_cpus)) { invltlb(); mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (CPU_ISSET(cpuid, &pmap->pm_active)) invltlb(); CPU_AND(&other_cpus, &pmap->pm_active); mask = &other_cpus; } smp_masked_invltlb(*mask, pmap); sched_unpin(); } void pmap_invalidate_cache(void) { sched_pin(); wbinvd(); smp_cache_flush(); sched_unpin(); } struct pde_action { cpuset_t invalidate; /* processors that invalidate their TLB */ vm_offset_t va; pd_entry_t *pde; pd_entry_t newpde; u_int store; /* processor that updates the PDE */ }; static void pmap_update_pde_kernel(void *arg) { struct pde_action *act = arg; pd_entry_t *pde; pmap_t pmap; if (act->store == PCPU_GET(cpuid)) { /* * Elsewhere, this operation requires allpmaps_lock for * synchronization. Here, it does not because it is being * performed in the context of an all_cpus rendezvous. */ LIST_FOREACH(pmap, &allpmaps, pm_list) { pde = pmap_pde(pmap, act->va); pde_store(pde, act->newpde); } } } static void pmap_update_pde_user(void *arg) { struct pde_action *act = arg; if (act->store == PCPU_GET(cpuid)) pde_store(act->pde, act->newpde); } static void pmap_update_pde_teardown(void *arg) { struct pde_action *act = arg; if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate)) pmap_update_pde_invalidate(act->va, act->newpde); } /* * Change the page size for the specified virtual address in a way that * prevents any possibility of the TLB ever having two entries that map the * same virtual address using different page sizes. This is the recommended * workaround for Erratum 383 on AMD Family 10h processors. It prevents a * machine check exception for a TLB state that is improperly diagnosed as a * hardware error. */ static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { struct pde_action act; cpuset_t active, other_cpus; u_int cpuid; sched_pin(); cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (pmap == kernel_pmap) active = all_cpus; else active = pmap->pm_active; if (CPU_OVERLAP(&active, &other_cpus)) { act.store = cpuid; act.invalidate = active; act.va = va; act.pde = pde; act.newpde = newpde; CPU_SET(cpuid, &active); smp_rendezvous_cpus(active, smp_no_rendevous_barrier, pmap == kernel_pmap ? pmap_update_pde_kernel : pmap_update_pde_user, pmap_update_pde_teardown, &act); } else { if (pmap == kernel_pmap) pmap_kenter_pde(va, newpde); else pde_store(pde, newpde); if (CPU_ISSET(cpuid, &active)) pmap_update_pde_invalidate(va, newpde); } sched_unpin(); } #else /* !SMP */ /* * Normal, non-SMP, 486+ invalidation functions. * We inline these within pmap.c for speed. */ PMAP_INLINE void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active)) invlpg(va); } PMAP_INLINE void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t addr; if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active)) for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); } PMAP_INLINE void pmap_invalidate_all(pmap_t pmap) { if (pmap == kernel_pmap) invltlb_glob(); else if (!CPU_EMPTY(&pmap->pm_active)) invltlb(); } PMAP_INLINE void pmap_invalidate_cache(void) { wbinvd(); } static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { if (pmap == kernel_pmap) pmap_kenter_pde(va, newpde); else pde_store(pde, newpde); if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active)) pmap_update_pde_invalidate(va, newpde); } #endif /* !SMP */ #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024) void pmap_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva, boolean_t force) { if (force) { sva &= ~(vm_offset_t)cpu_clflush_line_size; } else { KASSERT((sva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: sva not page-aligned")); KASSERT((eva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: eva not page-aligned")); } if ((cpu_feature & CPUID_SS) != 0 && !force) ; /* If "Self Snoop" is supported and allowed, do nothing. */ else if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0 && eva - sva < PMAP_CLFLUSH_THRESHOLD) { #ifdef DEV_APIC /* * XXX: Some CPUs fault, hang, or trash the local APIC * registers if we use CLFLUSH on the local APIC * range. The local APIC is always uncached, so we * don't need to flush for that range anyway. */ if (pmap_kextract(sva) == lapic_paddr) return; #endif /* * Otherwise, do per-cache line flush. Use the mfence * instruction to insure that previous stores are * included in the write-back. The processor * propagates flush to other processors in the cache * coherence domain. */ mfence(); for (; sva < eva; sva += cpu_clflush_line_size) clflushopt(sva); mfence(); } else if ((cpu_feature & CPUID_CLFSH) != 0 && eva - sva < PMAP_CLFLUSH_THRESHOLD) { #ifdef DEV_APIC if (pmap_kextract(sva) == lapic_paddr) return; #endif /* * Writes are ordered by CLFLUSH on Intel CPUs. */ if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (; sva < eva; sva += cpu_clflush_line_size) clflush(sva); if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } else { /* * No targeted cache flush methods are supported by CPU, * or the supplied range is bigger than 2MB. * Globally invalidate cache. */ pmap_invalidate_cache(); } } void pmap_invalidate_cache_pages(vm_page_t *pages, int count) { int i; if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE || (cpu_feature & CPUID_CLFSH) == 0) { pmap_invalidate_cache(); } else { for (i = 0; i < count; i++) pmap_flush_page(pages[i]); } } /* * Are we current address space or kernel? */ static __inline int pmap_is_current(pmap_t pmap) { return (pmap == kernel_pmap || pmap == vmspace_pmap(curthread->td_proc->p_vmspace)); } /* * If the given pmap is not the current or kernel pmap, the returned pte must * be released by passing it to pmap_pte_release(). */ pt_entry_t * pmap_pte(pmap_t pmap, vm_offset_t va) { pd_entry_t newpf; pd_entry_t *pde; pde = pmap_pde(pmap, va); if (*pde & PG_PS) return (pde); if (*pde != 0) { /* are we current address space or kernel? */ if (pmap_is_current(pmap)) return (vtopte(va)); mtx_lock(&PMAP2mutex); newpf = *pde & PG_FRAME; if ((*PMAP2 & PG_FRAME) != newpf) { *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); } return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); } return (NULL); } /* * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte * being NULL. */ static __inline void pmap_pte_release(pt_entry_t *pte) { if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) mtx_unlock(&PMAP2mutex); } /* * NB: The sequence of updating a page table followed by accesses to the * corresponding pages is subject to the situation described in the "AMD64 * Architecture Programmer's Manual Volume 2: System Programming" rev. 3.23, * "7.3.1 Special Coherency Considerations". Therefore, issuing the INVLPG * right after modifying the PTE bits is crucial. */ static __inline void invlcaddr(void *caddr) { invlpg((u_int)caddr); } /* * Super fast pmap_pte routine best used when scanning * the pv lists. This eliminates many coarse-grained * invltlb calls. Note that many of the pv list * scans are across different pmaps. It is very wasteful * to do an entire invltlb for checking a single mapping. * * If the given pmap is not the current pmap, pvh_global_lock * must be held and curthread pinned to a CPU. */ static pt_entry_t * pmap_pte_quick(pmap_t pmap, vm_offset_t va) { pd_entry_t newpf; pd_entry_t *pde; pde = pmap_pde(pmap, va); if (*pde & PG_PS) return (pde); if (*pde != 0) { /* are we current address space or kernel? */ if (pmap_is_current(pmap)) return (vtopte(va)); rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); newpf = *pde & PG_FRAME; if ((*PMAP1 & PG_FRAME) != newpf) { *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; #ifdef SMP PMAP1cpu = PCPU_GET(cpuid); #endif invlcaddr(PADDR1); PMAP1changed++; } else #ifdef SMP if (PMAP1cpu != PCPU_GET(cpuid)) { PMAP1cpu = PCPU_GET(cpuid); invlcaddr(PADDR1); PMAP1changedcpu++; } else #endif PMAP1unchanged++; return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); } return (0); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { vm_paddr_t rtval; pt_entry_t *pte; pd_entry_t pde; rtval = 0; PMAP_LOCK(pmap); pde = pmap->pm_pdir[va >> PDRSHIFT]; if (pde != 0) { if ((pde & PG_PS) != 0) rtval = (pde & PG_PS_FRAME) | (va & PDRMASK); else { pte = pmap_pte(pmap, va); rtval = (*pte & PG_FRAME) | (va & PAGE_MASK); pmap_pte_release(pte); } } PMAP_UNLOCK(pmap); return (rtval); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pd_entry_t pde; pt_entry_t pte, *ptep; vm_page_t m; vm_paddr_t pa; pa = 0; m = NULL; PMAP_LOCK(pmap); retry: pde = *pmap_pde(pmap, va); if (pde != 0) { if (pde & PG_PS) { if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) { if (vm_page_pa_tryrelock(pmap, (pde & PG_PS_FRAME) | (va & PDRMASK), &pa)) goto retry; m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) | (va & PDRMASK)); vm_page_hold(m); } } else { ptep = pmap_pte(pmap, va); pte = *ptep; pmap_pte_release(ptep); if (pte != 0 && ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) { if (vm_page_pa_tryrelock(pmap, pte & PG_FRAME, &pa)) goto retry; m = PHYS_TO_VM_PAGE(pte & PG_FRAME); vm_page_hold(m); } } } PA_UNLOCK_COND(pa); PMAP_UNLOCK(pmap); return (m); } /*************************************************** * Low level mapping routines..... ***************************************************/ /* * Add a wired page to the kva. * Note: not SMP coherent. * * This function may be used before pmap_bootstrap() is called. */ PMAP_INLINE void pmap_kenter(vm_offset_t va, vm_paddr_t pa) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | PG_RW | PG_V | pgeflag); } static __inline void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | PG_RW | PG_V | pgeflag | pmap_cache_bits(mode, 0)); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. * * This function may be used before pmap_bootstrap() is called. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *pte; pte = vtopte(va); pte_clear(pte); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { vm_offset_t va, sva; vm_paddr_t superpage_offset; pd_entry_t newpde; va = *virt; /* * Does the physical address range's size and alignment permit at * least one superpage mapping to be created? */ superpage_offset = start & PDRMASK; if ((end - start) - ((NBPDR - superpage_offset) & PDRMASK) >= NBPDR) { /* * Increase the starting virtual address so that its alignment * does not preclude the use of superpage mappings. */ if ((va & PDRMASK) < superpage_offset) va = (va & ~PDRMASK) + superpage_offset; else if ((va & PDRMASK) > superpage_offset) va = ((va + PDRMASK) & ~PDRMASK) + superpage_offset; } sva = va; while (start < end) { if ((start & PDRMASK) == 0 && end - start >= NBPDR && pseflag) { KASSERT((va & PDRMASK) == 0, ("pmap_map: misaligned va %#x", va)); newpde = start | PG_PS | pgeflag | PG_RW | PG_V; pmap_kenter_pde(va, newpde); va += NBPDR; start += NBPDR; } else { pmap_kenter(va, start); va += PAGE_SIZE; start += PAGE_SIZE; } } pmap_invalidate_range(kernel_pmap, sva, va); *virt = va; return (sva); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pt_entry_t *endpte, oldpte, pa, *pte; vm_page_t m; oldpte = 0; pte = vtopte(sva); endpte = pte + count; while (pte < endpte) { m = *ma++; pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 0); if ((*pte & (PG_FRAME | PG_PTE_CACHE)) != pa) { oldpte |= *pte; pte_store(pte, pa | pgeflag | PG_RW | PG_V); } pte++; } if (__predict_false((oldpte & PG_V) != 0)) pmap_invalidate_range(kernel_pmap, sva, sva + count * PAGE_SIZE); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { pmap_kremove(va); va += PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ static __inline void pmap_free_zero_pages(struct spglist *free) { vm_page_t m; while ((m = SLIST_FIRST(free)) != NULL) { SLIST_REMOVE_HEAD(free, plinks.s.ss); /* Preserve the page's PG_ZERO setting. */ vm_page_free_toq(m); } } /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Looks for a page table page mapping the specified virtual address in the * specified pmap's collection of idle page table pages. Returns NULL if there * is no page table page corresponding to the specified virtual address. */ static __inline vm_page_t pmap_lookup_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_lookup(&pmap->pm_root, va >> PDRSHIFT)); } /* * Removes the specified page table page from the specified pmap's collection * of idle page table pages. The specified page table page must be a member of * the pmap's collection. */ static __inline void pmap_remove_pt_page(pmap_t pmap, vm_page_t mpte) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); vm_radix_remove(&pmap->pm_root, mpte->pindex); } /* * Decrements a page table page's wire count, which is used to record the * number of valid page table entries within the page. If the wire count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free) { --m->wire_count; if (m->wire_count == 0) { _pmap_unwire_ptp(pmap, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free) { vm_offset_t pteva; /* * unmap the page table page */ pmap->pm_pdir[m->pindex] = 0; --pmap->pm_stats.resident_count; /* * This is a release store so that the ordinary store unmapping * the page table page is globally performed before TLB shoot- * down is begun. */ atomic_subtract_rel_int(&vm_cnt.v_wire_count, 1); /* * Do an invltlb to make the invalidated mapping * take effect immediately. */ pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex); pmap_invalidate_page(pmap, pteva); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing a page table entry, this routine is used to * conditionally free the page, and manage the hold/wire counts. */ static int pmap_unuse_pt(pmap_t pmap, vm_offset_t va, struct spglist *free) { pd_entry_t ptepde; vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (0); ptepde = *pmap_pde(pmap, va); mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); return (pmap_unwire_ptp(pmap, mpte, free)); } /* * Initialize the pmap for the swapper process. */ void pmap_pinit0(pmap_t pmap) { PMAP_LOCK_INIT(pmap); /* * Since the page table directory is shared with the kernel pmap, * which is already included in the list "allpmaps", this pmap does * not need to be inserted into that list. */ pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD); #if defined(PAE) || defined(PAE_TABLES) pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT); #endif pmap->pm_root.rt_root = 0; CPU_ZERO(&pmap->pm_active); PCPU_SET(curpmap, pmap); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ int pmap_pinit(pmap_t pmap) { vm_page_t m, ptdpg[NPGPTD]; vm_paddr_t pa; int i; /* * No need to allocate page table space yet but we do need a valid * page directory table. */ if (pmap->pm_pdir == NULL) { pmap->pm_pdir = (pd_entry_t *)kva_alloc(NBPTD); if (pmap->pm_pdir == NULL) return (0); #if defined(PAE) || defined(PAE_TABLES) pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); KASSERT(((vm_offset_t)pmap->pm_pdpt & ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, ("pmap_pinit: pdpt misaligned")); KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), ("pmap_pinit: pdpt above 4g")); #endif pmap->pm_root.rt_root = 0; } KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_pinit: pmap has reserved page table page(s)")); /* * allocate the page directory page(s) */ for (i = 0; i < NPGPTD;) { m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) VM_WAIT; else { ptdpg[i++] = m; } } pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD); for (i = 0; i < NPGPTD; i++) if ((ptdpg[i]->flags & PG_ZERO) == 0) pagezero(pmap->pm_pdir + (i * NPDEPG)); mtx_lock_spin(&allpmaps_lock); LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); /* Copy the kernel page table directory entries. */ bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t)); mtx_unlock_spin(&allpmaps_lock); /* install self-referential address mapping entry(s) */ for (i = 0; i < NPGPTD; i++) { pa = VM_PAGE_TO_PHYS(ptdpg[i]); pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M; #if defined(PAE) || defined(PAE_TABLES) pmap->pm_pdpt[i] = pa | PG_V; #endif } CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); return (1); } /* * this routine is called if the page table page is not * mapped correctly. */ static vm_page_t _pmap_allocpte(pmap_t pmap, u_int ptepindex, u_int flags) { vm_paddr_t ptepa; vm_page_t m; /* * Allocate a page table page. */ if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { if ((flags & PMAP_ENTER_NOSLEEP) == 0) { PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); VM_WAIT; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, the page table * page may have been allocated. */ return (NULL); } if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); /* * Map the pagetable page into the process address space, if * it isn't already there. */ pmap->pm_stats.resident_count++; ptepa = VM_PAGE_TO_PHYS(m); pmap->pm_pdir[ptepindex] = (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); return (m); } static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, u_int flags) { u_int ptepindex; pd_entry_t ptepa; vm_page_t m; /* * Calculate pagetable page index */ ptepindex = va >> PDRSHIFT; retry: /* * Get the page directory entry */ ptepa = pmap->pm_pdir[ptepindex]; /* * This supports switching from a 4MB page to a * normal 4K page. */ if (ptepa & PG_PS) { (void)pmap_demote_pde(pmap, &pmap->pm_pdir[ptepindex], va); ptepa = pmap->pm_pdir[ptepindex]; } /* * If the page table page is mapped, we just increment the * hold count, and activate it. */ if (ptepa) { m = PHYS_TO_VM_PAGE(ptepa & PG_FRAME); m->wire_count++; } else { /* * Here if the pte page isn't mapped, or if it has * been deallocated. */ m = _pmap_allocpte(pmap, ptepindex, flags); if (m == NULL && (flags & PMAP_ENTER_NOSLEEP) == 0) goto retry; } return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { vm_page_t m, ptdpg[NPGPTD]; int i; KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap has reserved page table page(s)")); KASSERT(CPU_EMPTY(&pmap->pm_active), ("releasing active pmap %p", pmap)); mtx_lock_spin(&allpmaps_lock); LIST_REMOVE(pmap, pm_list); mtx_unlock_spin(&allpmaps_lock); for (i = 0; i < NPGPTD; i++) ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i] & PG_FRAME); bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * sizeof(*pmap->pm_pdir)); pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); for (i = 0; i < NPGPTD; i++) { m = ptdpg[i]; #if defined(PAE) || defined(PAE_TABLES) KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), ("pmap_release: got wrong ptd page")); #endif m->wire_count--; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); } } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE; return (sysctl_handle_long(oidp, &ksize, 0, req)); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_size, "IU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return (sysctl_handle_long(oidp, &kfree, 0, req)); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_free, "IU", "Amount of KVM free"); /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t ptppaddr; vm_page_t nkpg; pd_entry_t newpdir; mtx_assert(&kernel_map->system_mtx, MA_OWNED); addr = roundup2(addr, NBPDR); if (addr - 1 >= kernel_map->max_offset) addr = kernel_map->max_offset; while (kernel_vm_end < addr) { if (pdir_pde(PTD, kernel_vm_end)) { kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } continue; } nkpg = vm_page_alloc(NULL, kernel_vm_end >> PDRSHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); nkpt++; if ((nkpg->flags & PG_ZERO) == 0) pmap_zero_page(nkpg); ptppaddr = VM_PAGE_TO_PHYS(nkpg); newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); pdir_pde(KPTD, kernel_vm_end) = pgeflag | newpdir; pmap_kenter_pde(kernel_vm_end, newpdir); kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 11); CTASSERT(_NPCPV == 336); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */ #define PC_FREE10 0x0000fffful /* Free values for index 10 */ static const uint32_t pc_freemask[_NPCM] = { PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE10 }; SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, "Current number of pv entries"); #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, "Current number of pv entry chunks"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, "Current number of pv entry chunks allocated"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, "Current number of pv entry chunks frees"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, "Number of times tried to get a chunk page but failed."); static long pv_entry_frees, pv_entry_allocs; static int pv_entry_spare; SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, "Current number of pv entry frees"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, "Current number of pv entry allocs"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, "Current number of spare pv entries"); #endif /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. */ static vm_page_t pmap_pv_reclaim(pmap_t locked_pmap) { struct pch newtail; struct pv_chunk *pc; struct md_page *pvh; pd_entry_t *pde; pmap_t pmap; pt_entry_t *pte, tpte; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint32_t inuse; int bit, field, freed; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); pmap = NULL; m_pc = NULL; SLIST_INIT(&free); TAILQ_INIT(&newtail); while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && (pv_vafree == 0 || SLIST_EMPTY(&free))) { TAILQ_REMOVE(&pv_chunks, pc, pc_lru); if (pmap != pc->pc_pmap) { if (pmap != NULL) { pmap_invalidate_all(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } pmap = pc->pc_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) PMAP_LOCK(pmap); else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) { pmap = NULL; TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); continue; } } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = bsfl(inuse); pv = &pc->pc_pventry[field * 32 + bit]; va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_PS) != 0) continue; pte = pmap_pte(pmap, va); tpte = *pte; if ((tpte & PG_W) == 0) tpte = pte_load_clear(pte); pmap_pte_release(pte); if ((tpte & PG_W) != 0) continue; KASSERT(tpte != 0, ("pmap_pv_reclaim: pmap %p va %x zero pte", pmap, va)); if ((tpte & PG_G) != 0) pmap_invalidate_page(pmap, va); m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((tpte & PG_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt(pmap, va, &free); freed++; } } if (freed == 0) { TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); continue; } /* Every freed mapping is for a 4 KB page. */ pmap->pm_stats.resident_count -= freed; PV_STAT(pv_entry_frees += freed); PV_STAT(pv_entry_spare += freed); pv_entry_count -= freed; TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); for (field = 0; field < _NPCM; field++) if (pc->pc_map[field] != pc_freemask[field]) { TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); /* * One freed pv entry in locked_pmap is * sufficient. */ if (pmap == locked_pmap) goto out; break; } if (field == _NPCM) { PV_STAT(pv_entry_spare -= _NPCPV); PV_STAT(pc_chunk_count--); PV_STAT(pc_chunk_frees++); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); pmap_qremove((vm_offset_t)pc, 1); pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); break; } } out: TAILQ_CONCAT(&pv_chunks, &newtail, pc_lru); if (pmap != NULL) { pmap_invalidate_all(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } if (m_pc == NULL && pv_vafree != 0 && SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->wire_count = 1; atomic_add_int(&vm_cnt.v_wire_count, 1); } pmap_free_zero_pages(&free); return (m_pc); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(pv_entry_frees++); PV_STAT(pv_entry_spare++); pv_entry_count--; pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 32; bit = idx % 32; pc->pc_map[field] |= 1ul << bit; for (idx = 0; idx < _NPCM; idx++) if (pc->pc_map[idx] != pc_freemask[idx]) { /* * 98% of the time, pc is already at the head of the * list. If it isn't already, move it to the head. */ if (__predict_false(TAILQ_FIRST(&pmap->pm_pvchunk) != pc)) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk(struct pv_chunk *pc) { vm_page_t m; TAILQ_REMOVE(&pv_chunks, pc, pc_lru); PV_STAT(pv_entry_spare -= _NPCPV); PV_STAT(pc_chunk_count--); PV_STAT(pc_chunk_frees++); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); pmap_qremove((vm_offset_t)pc, 1); vm_page_unwire(m, PQ_NONE); vm_page_free(m); pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); } /* * get a new pv_entry, allocating a block from the system * when needed. */ static pv_entry_t get_pv_entry(pmap_t pmap, boolean_t try) { static const struct timeval printinterval = { 60, 0 }; static struct timeval lastprint; int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(pv_entry_allocs++); pv_entry_count++; if (pv_entry_count > pv_entry_high_water) if (ratecheck(&lastprint, &printinterval)) printf("Approaching the limit on PV entries, consider " "increasing either the vm.pmap.shpgperproc or the " "vm.pmap.pv_entry_max tunable.\n"); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = bsfl(pc->pc_map[field]); break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 32 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ for (field = 0; field < _NPCM; field++) if (pc->pc_map[field] != 0) { PV_STAT(pv_entry_spare--); return (pv); /* not full, return */ } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(pv_entry_spare--); return (pv); } } /* * Access to the ptelist "pv_vafree" is synchronized by the pvh * global lock. If "pv_vafree" is currently non-empty, it will * remain non-empty until pmap_ptelist_alloc() completes. */ if (pv_vafree == 0 || (m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { if (try) { pv_entry_count--; PV_STAT(pc_chunk_tryfail++); return (NULL); } m = pmap_pv_reclaim(pmap); if (m == NULL) goto retry; } PV_STAT(pc_chunk_count++); PV_STAT(pc_chunk_allocs++); pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree); pmap_qenter((vm_offset_t)pc, &m, 1); pc->pc_pmap = pmap; pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */ for (field = 1; field < _NPCM; field++) pc->pc_map[field] = pc_freemask[field]; TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(pv_entry_spare += _NPCPV - 1); return (pv); } static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); break; } } return (pv); } static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_demote_pde: pa is not 4mpage aligned")); /* * Transfer the 4mpage's pv entry for this mapping to the first * page's pv list. */ pvh = pa_to_pvh(pa); va = trunc_4mpage(va); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); /* Instantiate the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_pde: page %p is not managed", m)); va += PAGE_SIZE; pmap_insert_entry(pmap, va, m); } while (va < va_last); } static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_promote_pde: pa is not 4mpage aligned")); /* * Transfer the first page's pv entry for this mapping to the * 4mpage's pv list. Aside from avoiding the cost of a call * to get_pv_entry(), a transfer avoids the possibility that * get_pv_entry() calls pmap_collect() and that pmap_collect() * removes one of the mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = trunc_4mpage(va); pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } static void pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) { struct md_page *pvh; rw_assert(&pvh_global_lock, RA_WLOCKED); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } /* * Create a pv entry for page at pa for * (pmap, va). */ static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); pv = get_pv_entry(pmap, FALSE); pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } /* * Conditionally create a pv entry. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pv_entry_count < pv_entry_high_water && (pv = get_pv_entry(pmap, TRUE)) != NULL) { pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); return (TRUE); } else return (FALSE); } /* * Create the pv entries for each of the pages within a superpage. */ static boolean_t pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); if (pv_entry_count < pv_entry_high_water && (pv = get_pv_entry(pmap, TRUE)) != NULL) { pv->pv_va = va; pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); return (TRUE); } else return (FALSE); } /* * Fills a page table page with mappings to consecutive physical pages. */ static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte) { pt_entry_t *pte; for (pte = firstpte; pte < firstpte + NPTEPG; pte++) { *pte = newpte; newpte += PAGE_SIZE; } } /* * Tries to demote a 2- or 4MB page mapping. If demotion fails, the * 2- or 4MB page mapping is invalidated. */ static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde, oldpde; pt_entry_t *firstpte, newpte; vm_paddr_t mptepa; vm_page_t mpte; struct spglist free; PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpde = *pde; KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V")); if ((oldpde & PG_A) != 0 && (mpte = pmap_lookup_pt_page(pmap, va)) != NULL) pmap_remove_pt_page(pmap, mpte); else { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: page table page for a wired mapping" " is missing")); /* * Invalidate the 2- or 4MB page mapping and return * "failure" if the mapping was never accessed or the * allocation of the new page table page fails. */ if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc(NULL, va >> PDRSHIFT, VM_ALLOC_NOOBJ | VM_ALLOC_NORMAL | VM_ALLOC_WIRED)) == NULL) { SLIST_INIT(&free); pmap_remove_pde(pmap, pde, trunc_4mpage(va), &free); pmap_invalidate_page(pmap, trunc_4mpage(va)); pmap_free_zero_pages(&free); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#x" " in pmap %p", va, pmap); return (FALSE); } if (va < VM_MAXUSER_ADDRESS) pmap->pm_stats.resident_count++; } mptepa = VM_PAGE_TO_PHYS(mpte); /* * If the page mapping is in the kernel's address space, then the * KPTmap can provide access to the page table page. Otherwise, * temporarily map the page table page (mpte) into the kernel's * address space at either PADDR1 or PADDR2. */ if (va >= KERNBASE) firstpte = &KPTmap[i386_btop(trunc_4mpage(va))]; else if (curthread->td_pinned > 0 && rw_wowned(&pvh_global_lock)) { if ((*PMAP1 & PG_FRAME) != mptepa) { *PMAP1 = mptepa | PG_RW | PG_V | PG_A | PG_M; #ifdef SMP PMAP1cpu = PCPU_GET(cpuid); #endif invlcaddr(PADDR1); PMAP1changed++; } else #ifdef SMP if (PMAP1cpu != PCPU_GET(cpuid)) { PMAP1cpu = PCPU_GET(cpuid); invlcaddr(PADDR1); PMAP1changedcpu++; } else #endif PMAP1unchanged++; firstpte = PADDR1; } else { mtx_lock(&PMAP2mutex); if ((*PMAP2 & PG_FRAME) != mptepa) { *PMAP2 = mptepa | PG_RW | PG_V | PG_A | PG_M; pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2); } firstpte = PADDR2; } newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V; KASSERT((oldpde & PG_A) != 0, ("pmap_demote_pde: oldpde is missing PG_A")); KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pde: oldpde is missing PG_M")); newpte = oldpde & ~PG_PS; if ((newpte & PG_PDE_PAT) != 0) newpte ^= PG_PDE_PAT | PG_PTE_PAT; /* * If the page table page is new, initialize it. */ if (mpte->wire_count == 1) { mpte->wire_count = NPTEPG; pmap_fill_ptp(firstpte, newpte); } KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME), ("pmap_demote_pde: firstpte and newpte map different physical" " addresses")); /* * If the mapping has changed attributes, update the page table * entries. */ if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE)) pmap_fill_ptp(firstpte, newpte); /* * Demote the mapping. This pmap is locked. The old PDE has * PG_A set. If the old PDE has PG_RW set, it also has PG_M * set. Thus, there is no danger of a race with another * processor changing the setting of PG_A and/or PG_M between * the read above and the store below. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else if (pmap == kernel_pmap) pmap_kenter_pde(va, newpde); else pde_store(pde, newpde); if (firstpte == PADDR2) mtx_unlock(&PMAP2mutex); /* * Invalidate the recursive mapping of the page table page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); /* * Demote the pv entry. This depends on the earlier demotion * of the mapping. Specifically, the (re)creation of a per- * page pv entry might trigger the execution of pmap_collect(), * which might reclaim a newly (re)created per-page pv entry * and destroy the associated mapping. In order to destroy * the mapping, the PDE must have already changed from mapping * the 2mpage to referencing the page table page. */ if ((oldpde & PG_MANAGED) != 0) pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME); pmap_pde_demotions++; CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#x" " in pmap %p", va, pmap); return (TRUE); } /* * Removes a 2- or 4MB page mapping from the kernel pmap. */ static void pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; vm_paddr_t mptepa; vm_page_t mpte; PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte = pmap_lookup_pt_page(pmap, va); if (mpte == NULL) panic("pmap_remove_kernel_pde: Missing pt page."); pmap_remove_pt_page(pmap, mpte); mptepa = VM_PAGE_TO_PHYS(mpte); newpde = mptepa | PG_M | PG_A | PG_RW | PG_V; /* * Initialize the page table page. */ pagezero((void *)&KPTmap[i386_btop(trunc_4mpage(va))]); /* * Remove the mapping. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pmap_kenter_pde(va, newpde); /* * Invalidate the recursive mapping of the page table page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); } /* * pmap_remove_pde: do the things to unmap a superpage in a process */ static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free) { struct md_page *pvh; pd_entry_t oldpde; vm_offset_t eva, va; vm_page_t m, mpte; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_remove_pde: sva is not 4mpage aligned")); oldpde = pte_load_clear(pdq); if (oldpde & PG_W) pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; /* * Machines that don't support invlpg, also don't support * PG_G. */ if (oldpde & PG_G) pmap_invalidate_page(kernel_pmap, sva); pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; if (oldpde & PG_MANAGED) { pvh = pa_to_pvh(oldpde & PG_PS_FRAME); pmap_pvh_free(pvh, pmap, sva); eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) { if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpde & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } if (pmap == kernel_pmap) { pmap_remove_kernel_pde(pmap, pdq, sva); } else { mpte = pmap_lookup_pt_page(pmap, sva); if (mpte != NULL) { pmap_remove_pt_page(pmap, mpte); pmap->pm_stats.resident_count--; KASSERT(mpte->wire_count == NPTEPG, ("pmap_remove_pde: pte page wire count error")); mpte->wire_count = 0; pmap_add_delayed_free_list(mpte, free, FALSE); atomic_subtract_int(&vm_cnt.v_wire_count, 1); } } } /* * pmap_remove_pte: do the things to unmap a page in a process */ static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, struct spglist *free) { pt_entry_t oldpte; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpte = pte_load_clear(ptq); KASSERT(oldpte != 0, ("pmap_remove_pte: pmap %p va %x zero pte", pmap, va)); if (oldpte & PG_W) pmap->pm_stats.wired_count -= 1; /* * Machines that don't support invlpg, also don't support * PG_G. */ if (oldpte & PG_G) pmap_invalidate_page(kernel_pmap, va); pmap->pm_stats.resident_count -= 1; if (oldpte & PG_MANAGED) { m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME); if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); pmap_remove_entry(pmap, m, va); } return (pmap_unuse_pt(pmap, va, free)); } /* * Remove a single page from a process address space */ static void pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free) { pt_entry_t *pte; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) return; pmap_remove_pte(pmap, pte, va, free); pmap_invalidate_page(pmap, va); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t pdnxt; pd_entry_t ptpaddr; pt_entry_t *pte; struct spglist free; int anyvalid; /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; anyvalid = 0; SLIST_INIT(&free); rw_wlock(&pvh_global_lock); sched_pin(); PMAP_LOCK(pmap); /* * special handling of removing one page. a very * common operation and easy to short circuit some * code. */ if ((sva + PAGE_SIZE == eva) && ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { pmap_remove_page(pmap, sva, &free); goto out; } for (; sva < eva; sva = pdnxt) { u_int pdirindex; /* * Calculate index for next page table. */ pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; if (pmap->pm_stats.resident_count == 0) break; pdirindex = sva >> PDRSHIFT; ptpaddr = pmap->pm_pdir[pdirindex]; /* * Weed out invalid mappings. Note: we assume that the page * directory table is always allocated, and in kernel virtual. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we removing the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == pdnxt && eva >= pdnxt) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_remove_pde(). */ if ((ptpaddr & PG_G) == 0) anyvalid = 1; pmap_remove_pde(pmap, &pmap->pm_pdir[pdirindex], sva, &free); continue; } else if (!pmap_demote_pde(pmap, &pmap->pm_pdir[pdirindex], sva)) { /* The large page mapping was destroyed. */ continue; } } /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (pdnxt > eva) pdnxt = eva; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { if (*pte == 0) continue; /* * The TLB entry for a PG_G mapping is invalidated * by pmap_remove_pte(). */ if ((*pte & PG_G) == 0) anyvalid = 1; if (pmap_remove_pte(pmap, pte, sva, &free)) break; } } out: sched_unpin(); if (anyvalid) pmap_invalidate_all(pmap); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; pt_entry_t *pte, tpte; pd_entry_t *pde; vm_offset_t va; struct spglist free; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); (void)pmap_demote_pde(pmap, pde, va); PMAP_UNLOCK(pmap); } small_mappings: while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pmap->pm_stats.resident_count--; pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found" " a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); tpte = pte_load_clear(pte); KASSERT(tpte != 0, ("pmap_remove_all: pmap %p va %x zero pte", pmap, pv->pv_va)); if (tpte & PG_W) pmap->pm_stats.wired_count--; if (tpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); pmap_unuse_pt(pmap, pv->pv_va, &free); pmap_invalidate_page(pmap, pv->pv_va); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); sched_unpin(); rw_wunlock(&pvh_global_lock); pmap_free_zero_pages(&free); } /* * pmap_protect_pde: do the things to protect a 4mpage in a process */ static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot) { pd_entry_t newpde, oldpde; vm_offset_t eva, va; vm_page_t m; boolean_t anychanged; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_protect_pde: sva is not 4mpage aligned")); anychanged = FALSE; retry: oldpde = newpde = *pde; if (oldpde & PG_MANAGED) { eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); } if ((prot & VM_PROT_WRITE) == 0) newpde &= ~(PG_RW | PG_M); #if defined(PAE) || defined(PAE_TABLES) if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; #endif if (newpde != oldpde) { if (!pde_cmpset(pde, oldpde, newpde)) goto retry; if (oldpde & PG_G) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } return (anychanged); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t pdnxt; pd_entry_t ptpaddr; pt_entry_t *pte; boolean_t anychanged, pv_lists_locked; KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } #if defined(PAE) || defined(PAE_TABLES) if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) == (VM_PROT_WRITE|VM_PROT_EXECUTE)) return; #else if (prot & VM_PROT_WRITE) return; #endif if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } anychanged = FALSE; PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pt_entry_t obits, pbits; u_int pdirindex; pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; pdirindex = sva >> PDRSHIFT; ptpaddr = pmap->pm_pdir[pdirindex]; /* * Weed out invalid mappings. Note: we assume that the page * directory table is always allocated, and in kernel virtual. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we protecting the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == pdnxt && eva >= pdnxt) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_protect_pde(). */ if (pmap_protect_pde(pmap, &pmap->pm_pdir[pdirindex], sva, prot)) anychanged = TRUE; continue; } else { if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { if (anychanged) pmap_invalidate_all( pmap); PMAP_UNLOCK(pmap); goto resume; } sched_pin(); } if (!pmap_demote_pde(pmap, &pmap->pm_pdir[pdirindex], sva)) { /* * The large page mapping was * destroyed. */ continue; } } } if (pdnxt > eva) pdnxt = eva; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { vm_page_t m; retry: /* * Regardless of whether a pte is 32 or 64 bits in * size, PG_RW, PG_A, and PG_M are among the least * significant 32 bits. */ obits = pbits = *pte; if ((pbits & PG_V) == 0) continue; if ((prot & VM_PROT_WRITE) == 0) { if ((pbits & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); vm_page_dirty(m); } pbits &= ~(PG_RW | PG_M); } #if defined(PAE) || defined(PAE_TABLES) if ((prot & VM_PROT_EXECUTE) == 0) pbits |= pg_nx; #endif if (pbits != obits) { #if defined(PAE) || defined(PAE_TABLES) if (!atomic_cmpset_64(pte, obits, pbits)) goto retry; #else if (!atomic_cmpset_int((u_int *)pte, obits, pbits)) goto retry; #endif if (obits & PG_G) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } } } if (anychanged) pmap_invalidate_all(pmap); if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Tries to promote the 512 or 1024, contiguous 4KB page mappings that are * within a single page table page (PTP) to a single 2- or 4MB page mapping. * For promotion to occur, two conditions must be met: (1) the 4KB page * mappings must map aligned, contiguous physical memory and (2) the 4KB page * mappings must have identical characteristics. * * Managed (PG_MANAGED) mappings within the kernel address space are not * promoted. The reason is that kernel PDEs are replicated in each pmap but * pmap_clear_ptes() and pmap_ts_referenced() only read the PDE from the kernel * pmap. */ static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; pt_entry_t *firstpte, oldpte, pa, *pte; vm_offset_t oldpteva; vm_page_t mpte; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Examine the first PTE in the specified PTP. Abort if this PTE is * either invalid, unused, or does not map the first 4KB physical page * within a 2- or 4MB page. */ firstpte = pmap_pte_quick(pmap, trunc_4mpage(va)); setpde: newpde = *firstpte; if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V)) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } if ((*firstpte & PG_MANAGED) != 0 && pmap == kernel_pmap) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } if ((newpde & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared without * a TLB invalidation. */ if (!atomic_cmpset_int((u_int *)firstpte, newpde, newpde & ~PG_RW)) goto setpde; newpde &= ~PG_RW; } /* * Examine each of the other PTEs in the specified PTP. Abort if this * PTE maps an unexpected 4KB physical page or does not have identical * characteristics to the first PTE. */ pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE; for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) { setpte: oldpte = *pte; if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } if ((oldpte & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared * without a TLB invalidation. */ if (!atomic_cmpset_int((u_int *)pte, oldpte, oldpte & ~PG_RW)) goto setpte; oldpte &= ~PG_RW; oldpteva = (oldpte & PG_FRAME & PDRMASK) | (va & ~PDRMASK); CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#x" " in pmap %p", oldpteva, pmap); } if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } pa -= PAGE_SIZE; } /* * Save the page table page in its current state until the PDE * mapping the superpage is demoted by pmap_demote_pde() or * destroyed by pmap_remove_pde(). */ mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == va >> PDRSHIFT, ("pmap_promote_pde: page table page's pindex is wrong")); if (pmap_insert_pt_page(pmap, mpte)) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x in pmap %p", va, pmap); return; } /* * Promote the pv entries. */ if ((newpde & PG_MANAGED) != 0) pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME); /* * Propagate the PAT index to its proper position. */ if ((newpde & PG_PTE_PAT) != 0) newpde ^= PG_PDE_PAT | PG_PTE_PAT; /* * Map the superpage. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else if (pmap == kernel_pmap) pmap_kenter_pde(va, PG_PS | newpde); else pde_store(pde, PG_PS | newpde); pmap_pde_promotions++; CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#x" " in pmap %p", va, pmap); } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { pd_entry_t *pde; pt_entry_t *pte; pt_entry_t newpte, origpte; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte, om; boolean_t invlva, wired; va = trunc_page(va); mpte = NULL; wired = (flags & PMAP_ENTER_WIRED) != 0; KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig")); KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS, ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va)); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); sched_pin(); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { mpte = pmap_allocpte(pmap, va, flags); if (mpte == NULL) { KASSERT((flags & PMAP_ENTER_NOSLEEP) != 0, ("pmap_allocpte failed with sleep allowed")); sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (KERN_RESOURCE_SHORTAGE); } } pde = pmap_pde(pmap, va); if ((*pde & PG_PS) != 0) panic("pmap_enter: attempted pmap_enter on 4MB page"); pte = pmap_pte_quick(pmap, va); /* * Page Directory table entry not valid, we need a new PT page */ if (pte == NULL) { panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x", (uintmax_t)pmap->pm_pdir[PTDPTDI], va); } pa = VM_PAGE_TO_PHYS(m); om = NULL; origpte = *pte; opa = origpte & PG_FRAME; /* * Mapping has not changed, must be protection or wiring change. */ if (origpte && (opa == pa)) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if (wired && ((origpte & PG_W) == 0)) pmap->pm_stats.wired_count++; else if (!wired && (origpte & PG_W)) pmap->pm_stats.wired_count--; /* * Remove extra pte reference */ if (mpte) mpte->wire_count--; if (origpte & PG_MANAGED) { om = m; pa |= PG_MANAGED; } goto validate; } pv = NULL; /* * Mapping has changed, invalidate old range and fall through to * handle validating new mapping. */ if (opa) { if (origpte & PG_W) pmap->pm_stats.wired_count--; if (origpte & PG_MANAGED) { om = PHYS_TO_VM_PAGE(opa); pv = pmap_pvh_remove(&om->md, pmap, va); } if (mpte != NULL) { mpte->wire_count--; KASSERT(mpte->wire_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%x", va)); } } else pmap->pm_stats.resident_count++; /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva, ("pmap_enter: managed mapping within the clean submap")); if (pv == NULL) pv = get_pv_entry(pmap, FALSE); pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); pa |= PG_MANAGED; } else if (pv != NULL) free_pv_entry(pmap, pv); /* * Increment counters */ if (wired) pmap->pm_stats.wired_count++; validate: /* * Now validate mapping with desired protection/wiring. */ newpte = (pt_entry_t)(pa | pmap_cache_bits(m->md.pat_mode, 0) | PG_V); if ((prot & VM_PROT_WRITE) != 0) { newpte |= PG_RW; if ((newpte & PG_MANAGED) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } #if defined(PAE) || defined(PAE_TABLES) if ((prot & VM_PROT_EXECUTE) == 0) newpte |= pg_nx; #endif if (wired) newpte |= PG_W; if (va < VM_MAXUSER_ADDRESS) newpte |= PG_U; if (pmap == kernel_pmap) newpte |= pgeflag; /* * if the mapping or permission bits are different, we need * to update the pte. */ if ((origpte & ~(PG_M|PG_A)) != newpte) { newpte |= PG_A; if ((flags & VM_PROT_WRITE) != 0) newpte |= PG_M; if (origpte & PG_V) { invlva = FALSE; origpte = pte_load_store(pte, newpte); if (origpte & PG_A) { if (origpte & PG_MANAGED) vm_page_aflag_set(om, PGA_REFERENCED); if (opa != VM_PAGE_TO_PHYS(m)) invlva = TRUE; #if defined(PAE) || defined(PAE_TABLES) if ((origpte & PG_NX) == 0 && (newpte & PG_NX) != 0) invlva = TRUE; #endif } if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((origpte & PG_MANAGED) != 0) vm_page_dirty(om); if ((prot & VM_PROT_WRITE) == 0) invlva = TRUE; } if ((origpte & PG_MANAGED) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); if (invlva) pmap_invalidate_page(pmap, va); } else pte_store(pte, newpte); } /* * If both the page table page and the reservation are fully * populated, then attempt promotion. */ if ((mpte == NULL || mpte->wire_count == NPTEPG) && pg_ps_enabled && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) pmap_promote_pde(pmap, pde, va); sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } /* * Tries to create a 2- or 4MB page mapping. Returns TRUE if successful and * FALSE otherwise. Fails if (1) a page table page cannot be allocated without * blocking, (2) a mapping already exists at the specified virtual address, or * (3) a pv entry cannot be allocated without reclaiming another pv entry. */ static boolean_t pmap_enter_pde(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { pd_entry_t *pde, newpde; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); pde = pmap_pde(pmap, va); if (*pde != 0) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 1) | PG_PS | PG_V; if ((m->oflags & VPO_UNMANAGED) == 0) { newpde |= PG_MANAGED; /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_pde(pmap, va, VM_PAGE_TO_PHYS(m))) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } } #if defined(PAE) || defined(PAE_TABLES) if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; #endif if (va < VM_MAXUSER_ADDRESS) newpde |= PG_U; /* * Increment counters. */ pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; /* * Map the superpage. */ pde_store(pde, newpde); pmap_pde_mappings++; CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx" " in pmap %p", va, pmap); return (TRUE); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & PDRMASK) == 0 && va + NBPDR <= end && m->psind == 1 && pg_ps_enabled && pmap_enter_pde(pmap, va, m, prot)) m = &m[NBPDR / PAGE_SIZE - 1]; else mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte); m = TAILQ_NEXT(m, listq); } rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte) { pt_entry_t *pte; vm_paddr_t pa; struct spglist free; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { u_int ptepindex; pd_entry_t ptepa; /* * Calculate pagetable page index */ ptepindex = va >> PDRSHIFT; if (mpte && (mpte->pindex == ptepindex)) { mpte->wire_count++; } else { /* * Get the page directory entry */ ptepa = pmap->pm_pdir[ptepindex]; /* * If the page table page is mapped, we just increment * the hold count, and activate it. */ if (ptepa) { if (ptepa & PG_PS) return (NULL); mpte = PHYS_TO_VM_PAGE(ptepa & PG_FRAME); mpte->wire_count++; } else { mpte = _pmap_allocpte(pmap, ptepindex, PMAP_ENTER_NOSLEEP); if (mpte == NULL) return (mpte); } } } else { mpte = NULL; } /* * This call to vtopte makes the assumption that we are * entering the page into the current pmap. In order to support * quick entry into any pmap, one would likely use pmap_pte_quick. * But that isn't as quick as vtopte. */ pte = vtopte(va); if (*pte) { if (mpte != NULL) { mpte->wire_count--; mpte = NULL; } return (mpte); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m)) { if (mpte != NULL) { SLIST_INIT(&free); if (pmap_unwire_ptp(pmap, mpte, &free)) { pmap_invalidate_page(pmap, va); pmap_free_zero_pages(&free); } mpte = NULL; } return (mpte); } /* * Increment counters */ pmap->pm_stats.resident_count++; pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(m->md.pat_mode, 0); #if defined(PAE) || defined(PAE_TABLES) if ((prot & VM_PROT_EXECUTE) == 0) pa |= pg_nx; #endif /* * Now validate mapping with RO protection */ if ((m->oflags & VPO_UNMANAGED) != 0) pte_store(pte, pa | PG_V | PG_U); else pte_store(pte, pa | PG_V | PG_U | PG_MANAGED); return (mpte); } /* * Make a temporary mapping for a physical address. This is only intended * to be used for panic dumps. */ void * pmap_kenter_temporary(vm_paddr_t pa, int i) { vm_offset_t va; va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); pmap_kenter(va, pa); invlpg(va); return ((void *)crashdumpmap); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { pd_entry_t *pde; vm_paddr_t pa, ptepa; vm_page_t p; int pat_mode; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); if (pseflag && (addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) { if (!vm_object_populate(object, pindex, pindex + atop(size))) return; p = vm_page_lookup(object, pindex); KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); pat_mode = p->md.pat_mode; /* * Abort the mapping if the first page is not physically * aligned to a 2/4MB page boundary. */ ptepa = VM_PAGE_TO_PHYS(p); if (ptepa & (NBPDR - 1)) return; /* * Skip the first page. Abort the mapping if the rest of * the pages are not physically contiguous or have differing * memory attributes. */ p = TAILQ_NEXT(p, listq); for (pa = ptepa + PAGE_SIZE; pa < ptepa + size; pa += PAGE_SIZE) { KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); if (pa != VM_PAGE_TO_PHYS(p) || pat_mode != p->md.pat_mode) return; p = TAILQ_NEXT(p, listq); } /* * Map using 2/4MB pages. Since "ptepa" is 2/4M aligned and * "size" is a multiple of 2/4M, adding the PAT setting to * "pa" will not affect the termination of this loop. */ PMAP_LOCK(pmap); for (pa = ptepa | pmap_cache_bits(pat_mode, 1); pa < ptepa + size; pa += NBPDR) { pde = pmap_pde(pmap, addr); if (*pde == 0) { pde_store(pde, pa | PG_PS | PG_M | PG_A | PG_U | PG_RW | PG_V); pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; pmap_pde_mappings++; } /* Else continue on if the PDE is already valid. */ addr += NBPDR; } PMAP_UNLOCK(pmap); } } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, * so there is no need to invalidate any TLB entries. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t pdnxt; pd_entry_t *pde; pt_entry_t *pte; boolean_t pv_lists_locked; if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; pde = pmap_pde(pmap, sva); if ((*pde & PG_V) == 0) continue; if ((*pde & PG_PS) != 0) { if ((*pde & PG_W) == 0) panic("pmap_unwire: pde %#jx is missing PG_W", (uintmax_t)*pde); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == pdnxt && eva >= pdnxt) { /* * Regardless of whether a pde (or pte) is 32 * or 64 bits in size, PG_W is among the least * significant 32 bits. */ atomic_clear_int((u_int *)pde, PG_W); pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; continue; } else { if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { PMAP_UNLOCK(pmap); /* Repeat sva. */ goto resume; } sched_pin(); } if (!pmap_demote_pde(pmap, pde, sva)) panic("pmap_unwire: demotion failed"); } } if (pdnxt > eva) pdnxt = eva; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { if ((*pte & PG_V) == 0) continue; if ((*pte & PG_W) == 0) panic("pmap_unwire: pte %#jx is missing PG_W", (uintmax_t)*pte); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. * * PG_W is among the least significant 32 bits. */ atomic_clear_int((u_int *)pte, PG_W); pmap->pm_stats.wired_count--; } } if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct spglist free; vm_offset_t addr; vm_offset_t end_addr = src_addr + len; vm_offset_t pdnxt; if (dst_addr != src_addr) return; if (!pmap_is_current(src_pmap)) return; rw_wlock(&pvh_global_lock); if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } sched_pin(); for (addr = src_addr; addr < end_addr; addr = pdnxt) { pt_entry_t *src_pte, *dst_pte; vm_page_t dstmpte, srcmpte; pd_entry_t srcptepaddr; u_int ptepindex; KASSERT(addr < UPT_MIN_ADDRESS, ("pmap_copy: invalid to pmap_copy page tables")); pdnxt = (addr + NBPDR) & ~PDRMASK; if (pdnxt < addr) pdnxt = end_addr; ptepindex = addr >> PDRSHIFT; srcptepaddr = src_pmap->pm_pdir[ptepindex]; if (srcptepaddr == 0) continue; if (srcptepaddr & PG_PS) { if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr) continue; if (dst_pmap->pm_pdir[ptepindex] == 0 && ((srcptepaddr & PG_MANAGED) == 0 || pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr & PG_PS_FRAME))) { dst_pmap->pm_pdir[ptepindex] = srcptepaddr & ~PG_W; dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; pmap_pde_mappings++; } continue; } srcmpte = PHYS_TO_VM_PAGE(srcptepaddr & PG_FRAME); KASSERT(srcmpte->wire_count > 0, ("pmap_copy: source page table page is unused")); if (pdnxt > end_addr) pdnxt = end_addr; src_pte = vtopte(addr); while (addr < pdnxt) { pt_entry_t ptetemp; ptetemp = *src_pte; /* * we only virtual copy managed pages */ if ((ptetemp & PG_MANAGED) != 0) { dstmpte = pmap_allocpte(dst_pmap, addr, PMAP_ENTER_NOSLEEP); if (dstmpte == NULL) goto out; dst_pte = pmap_pte_quick(dst_pmap, addr); if (*dst_pte == 0 && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) { /* * Clear the wired, modified, and * accessed (referenced) bits * during the copy. */ *dst_pte = ptetemp & ~(PG_W | PG_M | PG_A); dst_pmap->pm_stats.resident_count++; } else { SLIST_INIT(&free); if (pmap_unwire_ptp(dst_pmap, dstmpte, &free)) { pmap_invalidate_page(dst_pmap, addr); pmap_free_zero_pages(&free); } goto out; } if (dstmpte->wire_count >= srcmpte->wire_count) break; } addr += PAGE_SIZE; src_pte++; } } out: sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } /* * Zero 1 page of virtual memory mapped from a hardware page by the caller. */ static __inline void pagezero(void *page) { #if defined(I686_CPU) if (cpu_class == CPUCLASS_686) { #if defined(CPU_ENABLE_SSE) if (cpu_feature & CPUID_SSE2) sse2_pagezero(page); else #endif i686_pagezero(page); } else #endif bzero(page, PAGE_SIZE); } /* * Zero the specified hardware page. */ void pmap_zero_page(vm_page_t m) { struct sysmaps *sysmaps; sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP2) panic("pmap_zero_page: CMAP2 busy"); sched_pin(); *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(m->md.pat_mode, 0); invlcaddr(sysmaps->CADDR2); pagezero(sysmaps->CADDR2); *sysmaps->CMAP2 = 0; sched_unpin(); mtx_unlock(&sysmaps->lock); } /* * Zero an an area within a single hardware page. off and size must not * cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { struct sysmaps *sysmaps; sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP2) panic("pmap_zero_page_area: CMAP2 busy"); sched_pin(); *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(m->md.pat_mode, 0); invlcaddr(sysmaps->CADDR2); if (off == 0 && size == PAGE_SIZE) pagezero(sysmaps->CADDR2); else bzero((char *)sysmaps->CADDR2 + off, size); *sysmaps->CMAP2 = 0; sched_unpin(); mtx_unlock(&sysmaps->lock); } /* * Copy 1 specified hardware page to another. */ void pmap_copy_page(vm_page_t src, vm_page_t dst) { struct sysmaps *sysmaps; sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP1) panic("pmap_copy_page: CMAP1 busy"); if (*sysmaps->CMAP2) panic("pmap_copy_page: CMAP2 busy"); sched_pin(); *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A | pmap_cache_bits(src->md.pat_mode, 0); invlcaddr(sysmaps->CADDR1); *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M | pmap_cache_bits(dst->md.pat_mode, 0); invlcaddr(sysmaps->CADDR2); bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE); *sysmaps->CMAP1 = 0; *sysmaps->CMAP2 = 0; sched_unpin(); mtx_unlock(&sysmaps->lock); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { struct sysmaps *sysmaps; vm_page_t a_pg, b_pg; char *a_cp, *b_cp; vm_offset_t a_pg_offset, b_pg_offset; int cnt; sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP1 != 0) panic("pmap_copy_pages: CMAP1 busy"); if (*sysmaps->CMAP2 != 0) panic("pmap_copy_pages: CMAP2 busy"); sched_pin(); while (xfersize > 0) { a_pg = ma[a_offset >> PAGE_SHIFT]; a_pg_offset = a_offset & PAGE_MASK; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); b_pg = mb[b_offset >> PAGE_SHIFT]; b_pg_offset = b_offset & PAGE_MASK; cnt = min(cnt, PAGE_SIZE - b_pg_offset); *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(a_pg) | PG_A | pmap_cache_bits(a_pg->md.pat_mode, 0); invlcaddr(sysmaps->CADDR1); *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(b_pg) | PG_A | PG_M | pmap_cache_bits(b_pg->md.pat_mode, 0); invlcaddr(sysmaps->CADDR2); a_cp = sysmaps->CADDR1 + a_pg_offset; b_cp = sysmaps->CADDR2 + b_pg_offset; bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } *sysmaps->CMAP1 = 0; *sysmaps->CMAP2 = 0; sched_unpin(); mtx_unlock(&sysmaps->lock); } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_wunlock(&pvh_global_lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { int count; count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&pvh_global_lock); count = pmap_pvh_wired_mappings(&m->md, count); if ((m->flags & PG_FICTITIOUS) == 0) { count = pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)), count); } rw_wunlock(&pvh_global_lock); return (count); } /* * pmap_pvh_wired_mappings: * * Return the updated number "count" of managed mappings that are wired. */ static int pmap_pvh_wired_mappings(struct md_page *pvh, int count) { pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte = pmap_pte_quick(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } sched_unpin(); return (count); } /* * Returns TRUE if the given page is mapped individually or as part of * a 4mpage. Otherwise, returns FALSE. */ boolean_t pmap_page_is_mapped(vm_page_t m) { boolean_t rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (FALSE); rw_wlock(&pvh_global_lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list)); rw_wunlock(&pvh_global_lock); return (rv); } /* * Remove all pages from specified address space * this aids process exit speeds. Also, this code * is special cased for current process only, but * can have the more generic (and slightly slower) * mode enabled. This is much faster than pmap_remove * in the case of running down an entire address space. */ void pmap_remove_pages(pmap_t pmap) { pt_entry_t *pte, tpte; vm_page_t m, mpte, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct spglist free; int field, idx; int32_t bit; uint32_t inuse, bitmask; int allfree; if (pmap != PCPU_GET(curpmap)) { printf("warning: pmap_remove_pages called with non-current pmap\n"); return; } SLIST_INIT(&free); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); sched_pin(); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { KASSERT(pc->pc_pmap == pmap, ("Wrong pmap %p %p", pmap, pc->pc_pmap)); allfree = 1; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = bsfl(inuse); bitmask = 1UL << bit; idx = field * 32 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pte = pmap_pde(pmap, pv->pv_va); tpte = *pte; if ((tpte & PG_PS) == 0) { pte = vtopte(pv->pv_va); tpte = *pte & ~PG_PTE_PAT; } if (tpte == 0) { printf( "TPTE at %p IS ZERO @ VA %08x\n", pte, pv->pv_va); panic("bad pte"); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & PG_W) { allfree = 0; continue; } m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); KASSERT(m->phys_addr == (tpte & PG_FRAME), ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); pte_clear(pte); /* * Update the vm_page_t clean/reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((tpte & PG_PS) != 0) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } else vm_page_dirty(m); } /* Mark free */ PV_STAT(pv_entry_frees++); PV_STAT(pv_entry_spare++); pv_entry_count--; pc->pc_map[field] |= bitmask; if ((tpte & PG_PS) != 0) { pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; pvh = pa_to_pvh(tpte & PG_PS_FRAME); TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) if (TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } mpte = pmap_lookup_pt_page(pmap, pv->pv_va); if (mpte != NULL) { pmap_remove_pt_page(pmap, mpte); pmap->pm_stats.resident_count--; KASSERT(mpte->wire_count == NPTEPG, ("pmap_remove_pages: pte page wire count error")); mpte->wire_count = 0; pmap_add_delayed_free_list(mpte, &free, FALSE); atomic_subtract_int(&vm_cnt.v_wire_count, 1); } } else { pmap->pm_stats.resident_count--; TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } pmap_unuse_pt(pmap, pv->pv_va, &free); } } } if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } } sched_unpin(); pmap_invalidate_all(pmap); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no PTEs can have PG_M set. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (FALSE); rw_wlock(&pvh_global_lock); rv = pmap_is_modified_pvh(&m->md) || ((m->flags & PG_FICTITIOUS) == 0 && pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); rw_wunlock(&pvh_global_lock); return (rv); } /* * Returns TRUE if any of the given mappings were used to modify * physical memory. Otherwise, returns FALSE. Both page and 2mpage * mappings are supported. */ static boolean_t pmap_is_modified_pvh(struct md_page *pvh) { pv_entry_t pv; pt_entry_t *pte; pmap_t pmap; boolean_t rv; rw_assert(&pvh_global_lock, RA_WLOCKED); rv = FALSE; sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte = pmap_pte_quick(pmap, pv->pv_va); rv = (*pte & (PG_M | PG_RW)) == (PG_M | PG_RW); PMAP_UNLOCK(pmap); if (rv) break; } sched_unpin(); return (rv); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is elgible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pd_entry_t *pde; pt_entry_t *pte; boolean_t rv; rv = FALSE; PMAP_LOCK(pmap); pde = pmap_pde(pmap, addr); if (*pde != 0 && (*pde & PG_PS) == 0) { pte = vtopte(addr); rv = *pte == 0; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); rw_wlock(&pvh_global_lock); rv = pmap_is_referenced_pvh(&m->md) || ((m->flags & PG_FICTITIOUS) == 0 && pmap_is_referenced_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); rw_wunlock(&pvh_global_lock); return (rv); } /* * Returns TRUE if any of the given mappings were referenced and FALSE * otherwise. Both page and 4mpage mappings are supported. */ static boolean_t pmap_is_referenced_pvh(struct md_page *pvh) { pv_entry_t pv; pt_entry_t *pte; pmap_t pmap; boolean_t rv; rw_assert(&pvh_global_lock, RA_WLOCKED); rv = FALSE; sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte = pmap_pte_quick(pmap, pv->pv_va); rv = (*pte & (PG_A | PG_V)) == (PG_A | PG_V); PMAP_UNLOCK(pmap); if (rv) break; } sched_unpin(); return (rv); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pv_entry_t next_pv, pv; pmap_t pmap; pd_entry_t *pde; pt_entry_t oldpte, *pte; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); if ((*pde & PG_RW) != 0) (void)pmap_demote_pde(pmap, pde, va); PMAP_UNLOCK(pmap); } small_mappings: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_write: found" " a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); retry: oldpte = *pte; if ((oldpte & PG_RW) != 0) { /* * Regardless of whether a pte is 32 or 64 bits * in size, PG_RW and PG_M are among the least * significant 32 bits. */ if (!atomic_cmpset_int((u_int *)pte, oldpte, oldpte & ~(PG_RW | PG_M))) goto retry; if ((oldpte & PG_M) != 0) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); sched_unpin(); rw_wunlock(&pvh_global_lock); } -#define PMAP_TS_REFERENCED_MAX 5 - /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. - * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; pd_entry_t *pde; pt_entry_t *pte; vm_paddr_t pa; int rtval = 0; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); pa = VM_PAGE_TO_PHYS(m); pvh = pa_to_pvh(pa); rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0 || (pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); if ((*pde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Although "*pde" is mapping a 2/4MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } if ((*pde & PG_A) != 0) { /* * Since this reference bit is shared by either 1024 * or 512 4KB pages, it should not be cleared every * time it is tested. Apply a simple "hash" function * on the physical page number, the virtual superpage * number, and the pmap address to select one 4KB page * out of the 1024 or 512 on which testing the * reference bit will result in clearing that bit. * This function is designed to avoid the selection of * the same 4KB page for every 2- or 4MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^ (uintptr_t)pmap) & (NPTEPG - 1)) == 0 && (*pde & PG_W) == 0) { atomic_clear_int((u_int *)pde, PG_A); pmap_invalidate_page(pmap, pv->pv_va); } rtval++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); } if (rtval >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced: found a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((*pte & PG_A) != 0) { atomic_clear_int((u_int *)pte, PG_A); pmap_invalidate_page(pmap, pv->pv_va); rtval++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && rtval < PMAP_TS_REFERENCED_MAX); out: sched_unpin(); rw_wunlock(&pvh_global_lock); return (rtval); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { pd_entry_t oldpde, *pde; pt_entry_t *pte; vm_offset_t pdnxt; vm_page_t m; boolean_t anychanged, pv_lists_locked; if (advice != MADV_DONTNEED && advice != MADV_FREE) return; if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } anychanged = FALSE; PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; pde = pmap_pde(pmap, sva); oldpde = *pde; if ((oldpde & PG_V) == 0) continue; else if ((oldpde & PG_PS) != 0) { if ((oldpde & PG_MANAGED) == 0) continue; if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); goto resume; } sched_pin(); } if (!pmap_demote_pde(pmap, pde, sva)) { /* * The large page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Since the underlying page * table page is fully populated, this removal never * frees a page table page. */ if ((oldpde & PG_W) == 0) { pte = pmap_pte_quick(pmap, sva); KASSERT((*pte & PG_V) != 0, ("pmap_advise: invalid PTE")); pmap_remove_pte(pmap, pte, sva, NULL); anychanged = TRUE; } } if (pdnxt > eva) pdnxt = eva; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V)) continue; else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); vm_page_dirty(m); } atomic_clear_int((u_int *)pte, PG_M | PG_A); } else if ((*pte & PG_A) != 0) atomic_clear_int((u_int *)pte, PG_A); else continue; if ((*pte & PG_G) != 0) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } } if (anychanged) pmap_invalidate_all(pmap); if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { struct md_page *pvh; pv_entry_t next_pv, pv; pmap_t pmap; pd_entry_t oldpde, *pde; pt_entry_t oldpte, *pte; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("pmap_clear_modify: page %p is exclusive busied", m)); /* * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set. * If the object containing the page is locked and the page is not * exclusive busied, then PGA_WRITEABLE cannot be concurrently set. */ if ((m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); oldpde = *pde; if ((oldpde & PG_RW) != 0) { if (pmap_demote_pde(pmap, pde, va)) { if ((oldpde & PG_W) == 0) { /* * Write protect the mapping to a * single page so that a subsequent * write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pte_quick(pmap, va); oldpte = *pte; if ((oldpte & PG_V) != 0) { /* * Regardless of whether a pte is 32 or 64 bits * in size, PG_RW and PG_M are among the least * significant 32 bits. */ while (!atomic_cmpset_int((u_int *)pte, oldpte, oldpte & ~(PG_M | PG_RW))) oldpte = *pte; vm_page_dirty(m); pmap_invalidate_page(pmap, va); } } } } PMAP_UNLOCK(pmap); } small_mappings: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found" " a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Regardless of whether a pte is 32 or 64 bits * in size, PG_M is among the least significant * 32 bits. */ atomic_clear_int((u_int *)pte, PG_M); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } sched_unpin(); rw_wunlock(&pvh_global_lock); } /* * Miscellaneous support routines follow */ /* Adjust the cache mode for a 4KB page mapped via a PTE. */ static __inline void pmap_pte_attr(pt_entry_t *pte, int cache_bits) { u_int opte, npte; /* * The cache mode bits are all in the low 32-bits of the * PTE, so we can just spin on updating the low 32-bits. */ do { opte = *(u_int *)pte; npte = opte & ~PG_PTE_CACHE; npte |= cache_bits; } while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte)); } /* Adjust the cache mode for a 2/4MB page mapped via a PDE. */ static __inline void pmap_pde_attr(pd_entry_t *pde, int cache_bits) { u_int opde, npde; /* * The cache mode bits are all in the low 32-bits of the * PDE, so we can just spin on updating the low 32-bits. */ do { opde = *(u_int *)pde; npde = opde & ~PG_PDE_CACHE; npde |= cache_bits; } while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde)); } /* * Map a set of physical memory pages into the kernel virtual * address space. Return a pointer to where it is mapped. This * routine is intended to be used for mapping device memory, * NOT real memory. */ void * pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode) { struct pmap_preinit_mapping *ppim; vm_offset_t va, offset; vm_size_t tmpsize; int i; offset = pa & PAGE_MASK; size = round_page(offset + size); pa = pa & PG_FRAME; if (pa < KERNLOAD && pa + size <= KERNLOAD) va = KERNBASE + pa; else if (!pmap_initialized) { va = 0; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) { ppim->pa = pa; ppim->sz = size; ppim->mode = mode; ppim->va = virtual_avail; virtual_avail += size; va = ppim->va; break; } } if (va == 0) panic("%s: too many preinit mappings", __func__); } else { /* * If we have a preinit mapping, re-use it. */ for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->pa == pa && ppim->sz == size && ppim->mode == mode) return ((void *)(ppim->va + offset)); } va = kva_alloc(size); if (va == 0) panic("%s: Couldn't allocate KVA", __func__); } for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode); pmap_invalidate_range(kernel_pmap, va, va + tmpsize); pmap_invalidate_cache_range(va, va + size, FALSE); return ((void *)(va + offset)); } void * pmap_mapdev(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE)); } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_attr(pa, size, PAT_WRITE_BACK)); } void pmap_unmapdev(vm_offset_t va, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t offset; int i; if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD) return; offset = va & PAGE_MASK; size = round_page(offset + size); va = trunc_page(va); for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == va && ppim->sz == size) { if (pmap_initialized) return; ppim->pa = 0; ppim->va = 0; ppim->sz = 0; ppim->mode = 0; if (va + size == virtual_avail) virtual_avail = va; return; } } if (pmap_initialized) kva_free(va, size); } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pat_mode = ma; if ((m->flags & PG_FICTITIOUS) != 0) return; /* * If "m" is a normal page, flush it from the cache. * See pmap_invalidate_cache_range(). * * First, try to find an existing mapping of the page by sf * buffer. sf_buf_invalidate_cache() modifies mapping and * flushes the cache. */ if (sf_buf_invalidate_cache(m)) return; /* * If page is not mapped by sf buffer, but CPU does not * support self snoop, map the page transient and do * invalidation. In the worst case, whole cache is flushed by * pmap_invalidate_cache_range(). */ if ((cpu_feature & CPUID_SS) == 0) pmap_flush_page(m); } static void pmap_flush_page(vm_page_t m) { struct sysmaps *sysmaps; vm_offset_t sva, eva; bool useclflushopt; useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0; if (useclflushopt || (cpu_feature & CPUID_CLFSH) != 0) { sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)]; mtx_lock(&sysmaps->lock); if (*sysmaps->CMAP2) panic("pmap_flush_page: CMAP2 busy"); sched_pin(); *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(m->md.pat_mode, 0); invlcaddr(sysmaps->CADDR2); sva = (vm_offset_t)sysmaps->CADDR2; eva = sva + PAGE_SIZE; /* * Use mfence despite the ordering implied by * mtx_{un,}lock() because clflush on non-Intel CPUs * and clflushopt are not guaranteed to be ordered by * any other instruction. */ if (useclflushopt || cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (; sva < eva; sva += cpu_clflush_line_size) { if (useclflushopt) clflushopt(sva); else clflush(sva); } if (useclflushopt || cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); *sysmaps->CMAP2 = 0; sched_unpin(); mtx_unlock(&sysmaps->lock); } else pmap_invalidate_cache(); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the kernel map. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. */ int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) { vm_offset_t base, offset, tmpva; pd_entry_t *pde; pt_entry_t *pte; int cache_bits_pte, cache_bits_pde; boolean_t changed; base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); /* * Only supported on kernel virtual addresses above the recursive map. */ if (base < VM_MIN_KERNEL_ADDRESS) return (EINVAL); cache_bits_pde = pmap_cache_bits(mode, 1); cache_bits_pte = pmap_cache_bits(mode, 0); changed = FALSE; /* * Pages that aren't mapped aren't supported. Also break down * 2/4MB pages into 4KB pages if required. */ PMAP_LOCK(kernel_pmap); for (tmpva = base; tmpva < base + size; ) { pde = pmap_pde(kernel_pmap, tmpva); if (*pde == 0) { PMAP_UNLOCK(kernel_pmap); return (EINVAL); } if (*pde & PG_PS) { /* * If the current 2/4MB page already has * the required memory type, then we need not * demote this page. Just increment tmpva to * the next 2/4MB page frame. */ if ((*pde & PG_PDE_CACHE) == cache_bits_pde) { tmpva = trunc_4mpage(tmpva) + NBPDR; continue; } /* * If the current offset aligns with a 2/4MB * page frame and there is at least 2/4MB left * within the range, then we need not break * down this page into 4KB pages. */ if ((tmpva & PDRMASK) == 0 && tmpva + PDRMASK < base + size) { tmpva += NBPDR; continue; } if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) { PMAP_UNLOCK(kernel_pmap); return (ENOMEM); } } pte = vtopte(tmpva); if (*pte == 0) { PMAP_UNLOCK(kernel_pmap); return (EINVAL); } tmpva += PAGE_SIZE; } PMAP_UNLOCK(kernel_pmap); /* * Ok, all the pages exist, so run through them updating their * cache mode if required. */ for (tmpva = base; tmpva < base + size; ) { pde = pmap_pde(kernel_pmap, tmpva); if (*pde & PG_PS) { if ((*pde & PG_PDE_CACHE) != cache_bits_pde) { pmap_pde_attr(pde, cache_bits_pde); changed = TRUE; } tmpva = trunc_4mpage(tmpva) + NBPDR; } else { pte = vtopte(tmpva); if ((*pte & PG_PTE_CACHE) != cache_bits_pte) { pmap_pte_attr(pte, cache_bits_pte); changed = TRUE; } tmpva += PAGE_SIZE; } } /* * Flush CPU caches to make sure any data isn't cached that * shouldn't be, etc. */ if (changed) { pmap_invalidate_range(kernel_pmap, base, tmpva); pmap_invalidate_cache_range(base, tmpva, FALSE); } return (0); } /* * perform the pmap work for mincore */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa) { pd_entry_t *pdep; pt_entry_t *ptep, pte; vm_paddr_t pa; int val; PMAP_LOCK(pmap); retry: pdep = pmap_pde(pmap, addr); if (*pdep != 0) { if (*pdep & PG_PS) { pte = *pdep; /* Compute the physical address of the 4KB page. */ pa = ((*pdep & PG_PS_FRAME) | (addr & PDRMASK)) & PG_FRAME; val = MINCORE_SUPER; } else { ptep = pmap_pte(pmap, addr); pte = *ptep; pmap_pte_release(ptep); pa = pte & PG_FRAME; val = 0; } } else { pte = 0; pa = 0; val = 0; } if ((pte & PG_V) != 0) { val |= MINCORE_INCORE; if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((pte & PG_A) != 0) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) { /* Ensure that "PHYS_TO_VM_PAGE(pa)->object" doesn't change. */ if (vm_page_pa_tryrelock(pmap, pa, locked_pa)) goto retry; } else PA_UNLOCK_COND(*locked_pa); PMAP_UNLOCK(pmap); return (val); } void pmap_activate(struct thread *td) { pmap_t pmap, oldpmap; u_int cpuid; u_int32_t cr3; critical_enter(); pmap = vmspace_pmap(td->td_proc->p_vmspace); oldpmap = PCPU_GET(curpmap); cpuid = PCPU_GET(cpuid); #if defined(SMP) CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active); CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_CLR(cpuid, &oldpmap->pm_active); CPU_SET(cpuid, &pmap->pm_active); #endif #if defined(PAE) || defined(PAE_TABLES) cr3 = vtophys(pmap->pm_pdpt); #else cr3 = vtophys(pmap->pm_pdir); #endif /* * pmap_activate is for the current thread on the current cpu */ td->td_pcb->pcb_cr3 = cr3; load_cr3(cr3); PCPU_SET(curpmap, pmap); critical_exit(); } void pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < NBPDR) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & PDRMASK; if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR || (*addr & PDRMASK) == superpage_offset) return; if ((*addr & PDRMASK) < superpage_offset) *addr = (*addr & ~PDRMASK) + superpage_offset; else *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset; } vm_offset_t pmap_quick_enter_page(vm_page_t m) { vm_offset_t qaddr; pt_entry_t *pte; critical_enter(); qaddr = PCPU_GET(qmap_addr); pte = vtopte(qaddr); KASSERT(*pte == 0, ("pmap_quick_enter_page: PTE busy")); *pte = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(pmap_page_get_memattr(m), 0); invlpg(qaddr); return (qaddr); } void pmap_quick_remove_page(vm_offset_t addr) { vm_offset_t qaddr; pt_entry_t *pte; qaddr = PCPU_GET(qmap_addr); pte = vtopte(qaddr); KASSERT(*pte != 0, ("pmap_quick_remove_page: PTE not in use")); KASSERT(addr == qaddr, ("pmap_quick_remove_page: invalid address")); *pte = 0; critical_exit(); } #if defined(PMAP_DEBUG) pmap_pid_dump(int pid) { pmap_t pmap; struct proc *p; int npte = 0; int index; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { if (p->p_pid != pid) continue; if (p->p_vmspace) { int i,j; index = 0; pmap = vmspace_pmap(p->p_vmspace); for (i = 0; i < NPDEPTD; i++) { pd_entry_t *pde; pt_entry_t *pte; vm_offset_t base = i << PDRSHIFT; pde = &pmap->pm_pdir[i]; if (pde && pmap_pde_v(pde)) { for (j = 0; j < NPTEPG; j++) { vm_offset_t va = base + (j << PAGE_SHIFT); if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) { if (index) { index = 0; printf("\n"); } sx_sunlock(&allproc_lock); return (npte); } pte = pmap_pte(pmap, va); if (pte && pmap_pte_v(pte)) { pt_entry_t pa; vm_page_t m; pa = *pte; m = PHYS_TO_VM_PAGE(pa & PG_FRAME); printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x", va, pa, m->hold_count, m->wire_count, m->flags); npte++; index++; if (index >= 2) { index = 0; printf("\n"); } else { printf(" "); } } } } } } } sx_sunlock(&allproc_lock); return (npte); } #endif Index: projects/mips64-clang/sys/kern/kern_exit.c =================================================================== --- projects/mips64-clang/sys/kern/kern_exit.c (revision 305711) +++ projects/mips64-clang/sys/kern/kern_exit.c (revision 305712) @@ -1,1326 +1,1326 @@ /*- * Copyright (c) 1982, 1986, 1989, 1991, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)kern_exit.c 8.7 (Berkeley) 2/12/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_compat.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* for acct_process() function prototype */ #include #include #include #include #include #ifdef KTRACE #include #endif #include #include #include #include #include #include #include #include #include #include #ifdef KDTRACE_HOOKS #include dtrace_execexit_func_t dtrace_fasttrap_exit; #endif SDT_PROVIDER_DECLARE(proc); SDT_PROBE_DEFINE1(proc, , , exit, "int"); /* Hook for NFS teardown procedure. */ void (*nlminfo_release_p)(struct proc *p); struct proc * proc_realparent(struct proc *child) { struct proc *p, *parent; sx_assert(&proctree_lock, SX_LOCKED); if ((child->p_treeflag & P_TREE_ORPHANED) == 0) { if (child->p_oppid == 0 || child->p_pptr->p_pid == child->p_oppid) parent = child->p_pptr; else parent = initproc; return (parent); } for (p = child; (p->p_treeflag & P_TREE_FIRST_ORPHAN) == 0;) { /* Cannot use LIST_PREV(), since the list head is not known. */ p = __containerof(p->p_orphan.le_prev, struct proc, p_orphan.le_next); KASSERT((p->p_treeflag & P_TREE_ORPHANED) != 0, ("missing P_ORPHAN %p", p)); } parent = __containerof(p->p_orphan.le_prev, struct proc, p_orphans.lh_first); return (parent); } void reaper_abandon_children(struct proc *p, bool exiting) { struct proc *p1, *p2, *ptmp; sx_assert(&proctree_lock, SX_LOCKED); KASSERT(p != initproc, ("reaper_abandon_children for initproc")); if ((p->p_treeflag & P_TREE_REAPER) == 0) return; p1 = p->p_reaper; LIST_FOREACH_SAFE(p2, &p->p_reaplist, p_reapsibling, ptmp) { LIST_REMOVE(p2, p_reapsibling); p2->p_reaper = p1; p2->p_reapsubtree = p->p_reapsubtree; LIST_INSERT_HEAD(&p1->p_reaplist, p2, p_reapsibling); if (exiting && p2->p_pptr == p) { PROC_LOCK(p2); proc_reparent(p2, p1); PROC_UNLOCK(p2); } } KASSERT(LIST_EMPTY(&p->p_reaplist), ("p_reaplist not empty")); p->p_treeflag &= ~P_TREE_REAPER; } static void clear_orphan(struct proc *p) { struct proc *p1; sx_assert(&proctree_lock, SA_XLOCKED); if ((p->p_treeflag & P_TREE_ORPHANED) == 0) return; if ((p->p_treeflag & P_TREE_FIRST_ORPHAN) != 0) { p1 = LIST_NEXT(p, p_orphan); if (p1 != NULL) p1->p_treeflag |= P_TREE_FIRST_ORPHAN; p->p_treeflag &= ~P_TREE_FIRST_ORPHAN; } LIST_REMOVE(p, p_orphan); p->p_treeflag &= ~P_TREE_ORPHANED; } /* * exit -- death of process. */ void sys_sys_exit(struct thread *td, struct sys_exit_args *uap) { exit1(td, uap->rval, 0); /* NOTREACHED */ } /* * Exit: deallocate address space and other resources, change proc state to * zombie, and unlink proc from allproc and parent's lists. Save exit status * and rusage for wait(). Check for child processes and orphan them. */ void exit1(struct thread *td, int rval, int signo) { struct proc *p, *nq, *q, *t; struct thread *tdt; mtx_assert(&Giant, MA_NOTOWNED); KASSERT(rval == 0 || signo == 0, ("exit1 rv %d sig %d", rval, signo)); p = td->td_proc; /* * XXX in case we're rebooting we just let init die in order to * work around an unsolved stack overflow seen very late during * shutdown on sparc64 when the gmirror worker process exists. */ if (p == initproc && rebooting == 0) { printf("init died (signal %d, exit %d)\n", signo, rval); panic("Going nowhere without my init!"); } /* * Deref SU mp, since the thread does not return to userspace. */ if (softdep_ast_cleanup != NULL) softdep_ast_cleanup(); /* * MUST abort all other threads before proceeding past here. */ PROC_LOCK(p); /* * First check if some other thread or external request got * here before us. If so, act appropriately: exit or suspend. * We must ensure that stop requests are handled before we set * P_WEXIT. */ thread_suspend_check(0); while (p->p_flag & P_HADTHREADS) { /* * Kill off the other threads. This requires * some co-operation from other parts of the kernel * so it may not be instantaneous. With this state set * any thread entering the kernel from userspace will * thread_exit() in trap(). Any thread attempting to * sleep will return immediately with EINTR or EWOULDBLOCK * which will hopefully force them to back out to userland * freeing resources as they go. Any thread attempting * to return to userland will thread_exit() from userret(). * thread_exit() will unsuspend us when the last of the * other threads exits. * If there is already a thread singler after resumption, * calling thread_single will fail; in that case, we just * re-check all suspension request, the thread should * either be suspended there or exit. */ if (!thread_single(p, SINGLE_EXIT)) /* * All other activity in this process is now * stopped. Threading support has been turned * off. */ break; /* * Recheck for new stop or suspend requests which * might appear while process lock was dropped in * thread_single(). */ thread_suspend_check(0); } KASSERT(p->p_numthreads == 1, ("exit1: proc %p exiting with %d threads", p, p->p_numthreads)); racct_sub(p, RACCT_NTHR, 1); /* Let event handler change exit status */ p->p_xexit = rval; p->p_xsig = signo; /* * Wakeup anyone in procfs' PIOCWAIT. They should have a hold * on our vmspace, so we should block below until they have * released their reference to us. Note that if they have * requested S_EXIT stops we will block here until they ack * via PIOCCONT. */ _STOPEVENT(p, S_EXIT, 0); /* * Ignore any pending request to stop due to a stop signal. * Once P_WEXIT is set, future requests will be ignored as * well. */ p->p_flag &= ~P_STOPPED_SIG; KASSERT(!P_SHOULDSTOP(p), ("exiting process is stopped")); /* * Note that we are exiting and do another wakeup of anyone in * PIOCWAIT in case they aren't listening for S_EXIT stops or * decided to wait again after we told them we are exiting. */ p->p_flag |= P_WEXIT; wakeup(&p->p_stype); /* * Wait for any processes that have a hold on our vmspace to * release their reference. */ while (p->p_lock > 0) msleep(&p->p_lock, &p->p_mtx, PWAIT, "exithold", 0); PROC_UNLOCK(p); /* Drain the limit callout while we don't have the proc locked */ callout_drain(&p->p_limco); #ifdef AUDIT /* * The Sun BSM exit token contains two components: an exit status as * passed to exit(), and a return value to indicate what sort of exit * it was. The exit status is WEXITSTATUS(rv), but it's not clear * what the return value is. */ AUDIT_ARG_EXIT(rval, 0); AUDIT_SYSCALL_EXIT(0, td); #endif /* Are we a task leader with peers? */ if (p->p_peers != NULL && p == p->p_leader) { mtx_lock(&ppeers_lock); q = p->p_peers; while (q != NULL) { PROC_LOCK(q); kern_psignal(q, SIGKILL); PROC_UNLOCK(q); q = q->p_peers; } while (p->p_peers != NULL) msleep(p, &ppeers_lock, PWAIT, "exit1", 0); mtx_unlock(&ppeers_lock); } /* * Check if any loadable modules need anything done at process exit. * E.g. SYSV IPC stuff. * Event handler could change exit status. * XXX what if one of these generates an error? */ EVENTHANDLER_INVOKE(process_exit, p); /* * If parent is waiting for us to exit or exec, * P_PPWAIT is set; we will wakeup the parent below. */ PROC_LOCK(p); stopprofclock(p); p->p_flag &= ~(P_TRACED | P_PPWAIT | P_PPTRACE); p->p_ptevents = 0; /* * Stop the real interval timer. If the handler is currently * executing, prevent it from rearming itself and let it finish. */ if (timevalisset(&p->p_realtimer.it_value) && _callout_stop_safe(&p->p_itcallout, CS_EXECUTING, NULL) == 0) { timevalclear(&p->p_realtimer.it_interval); msleep(&p->p_itcallout, &p->p_mtx, PWAIT, "ritwait", 0); KASSERT(!timevalisset(&p->p_realtimer.it_value), ("realtime timer is still armed")); } PROC_UNLOCK(p); umtx_thread_exit(td); /* * Reset any sigio structures pointing to us as a result of * F_SETOWN with our pid. */ funsetownlst(&p->p_sigiolst); /* * If this process has an nlminfo data area (for lockd), release it */ if (nlminfo_release_p != NULL && p->p_nlminfo != NULL) (*nlminfo_release_p)(p); /* * Close open files and release open-file table. * This may block! */ fdescfree(td); /* * If this thread tickled GEOM, we need to wait for the giggling to * stop before we return to userland */ if (td->td_pflags & TDP_GEOM) g_waitidle(); /* * Remove ourself from our leader's peer list and wake our leader. */ if (p->p_leader->p_peers != NULL) { mtx_lock(&ppeers_lock); if (p->p_leader->p_peers != NULL) { q = p->p_leader; while (q->p_peers != p) q = q->p_peers; q->p_peers = p->p_peers; wakeup(p->p_leader); } mtx_unlock(&ppeers_lock); } vmspace_exit(td); killjobc(); (void)acct_process(td); #ifdef KTRACE ktrprocexit(td); #endif /* * Release reference to text vnode */ if (p->p_textvp != NULL) { vrele(p->p_textvp); p->p_textvp = NULL; } /* * Release our limits structure. */ lim_free(p->p_limit); p->p_limit = NULL; tidhash_remove(td); /* * Remove proc from allproc queue and pidhash chain. * Place onto zombproc. Unlink from parent's child list. */ sx_xlock(&allproc_lock); LIST_REMOVE(p, p_list); LIST_INSERT_HEAD(&zombproc, p, p_list); LIST_REMOVE(p, p_hash); sx_xunlock(&allproc_lock); /* * Call machine-dependent code to release any * machine-dependent resources other than the address space. * The address space is released by "vmspace_exitfree(p)" in * vm_waitproc(). */ cpu_exit(td); WITNESS_WARN(WARN_PANIC, NULL, "process (pid %d) exiting", p->p_pid); /* * Reparent all children processes: * - traced ones to the original parent (or init if we are that parent) * - the rest to init */ sx_xlock(&proctree_lock); q = LIST_FIRST(&p->p_children); if (q != NULL) /* only need this if any child is S_ZOMB */ wakeup(q->p_reaper); for (; q != NULL; q = nq) { nq = LIST_NEXT(q, p_sibling); PROC_LOCK(q); q->p_sigparent = SIGCHLD; if (!(q->p_flag & P_TRACED)) { proc_reparent(q, q->p_reaper); } else { /* * Traced processes are killed since their existence * means someone is screwing up. */ t = proc_realparent(q); if (t == p) { proc_reparent(q, q->p_reaper); } else { PROC_LOCK(t); proc_reparent(q, t); PROC_UNLOCK(t); } /* * Since q was found on our children list, the * proc_reparent() call moved q to the orphan * list due to present P_TRACED flag. Clear * orphan link for q now while q is locked. */ clear_orphan(q); q->p_flag &= ~(P_TRACED | P_STOPPED_TRACE); q->p_flag2 &= ~P2_PTRACE_FSTP; q->p_ptevents = 0; FOREACH_THREAD_IN_PROC(q, tdt) { tdt->td_dbgflags &= ~(TDB_SUSPEND | TDB_XSIG | TDB_FSTP); } kern_psignal(q, SIGKILL); } PROC_UNLOCK(q); } /* * Also get rid of our orphans. */ while ((q = LIST_FIRST(&p->p_orphans)) != NULL) { PROC_LOCK(q); CTR2(KTR_PTRACE, "exit: pid %d, clearing orphan %d", p->p_pid, q->p_pid); clear_orphan(q); PROC_UNLOCK(q); } /* Save exit status. */ PROC_LOCK(p); p->p_xthread = td; /* Tell the prison that we are gone. */ prison_proc_free(p->p_ucred->cr_prison); #ifdef KDTRACE_HOOKS /* * Tell the DTrace fasttrap provider about the exit if it * has declared an interest. */ if (dtrace_fasttrap_exit) dtrace_fasttrap_exit(p); #endif /* * Notify interested parties of our demise. */ KNOTE_LOCKED(p->p_klist, NOTE_EXIT); #ifdef KDTRACE_HOOKS int reason = CLD_EXITED; if (WCOREDUMP(signo)) reason = CLD_DUMPED; else if (WIFSIGNALED(signo)) reason = CLD_KILLED; SDT_PROBE1(proc, , , exit, reason); #endif /* * If this is a process with a descriptor, we may not need to deliver * a signal to the parent. proctree_lock is held over * procdesc_exit() to serialize concurrent calls to close() and * exit(). */ if (p->p_procdesc == NULL || procdesc_exit(p)) { /* * Notify parent that we're gone. If parent has the * PS_NOCLDWAIT flag set, or if the handler is set to SIG_IGN, * notify process 1 instead (and hope it will handle this * situation). */ PROC_LOCK(p->p_pptr); mtx_lock(&p->p_pptr->p_sigacts->ps_mtx); if (p->p_pptr->p_sigacts->ps_flag & (PS_NOCLDWAIT | PS_CLDSIGIGN)) { struct proc *pp; mtx_unlock(&p->p_pptr->p_sigacts->ps_mtx); pp = p->p_pptr; PROC_UNLOCK(pp); proc_reparent(p, p->p_reaper); p->p_sigparent = SIGCHLD; PROC_LOCK(p->p_pptr); /* * Notify parent, so in case he was wait(2)ing or * executing waitpid(2) with our pid, he will * continue. */ wakeup(pp); } else mtx_unlock(&p->p_pptr->p_sigacts->ps_mtx); if (p->p_pptr == p->p_reaper || p->p_pptr == initproc) childproc_exited(p); else if (p->p_sigparent != 0) { if (p->p_sigparent == SIGCHLD) childproc_exited(p); else /* LINUX thread */ kern_psignal(p->p_pptr, p->p_sigparent); } } else PROC_LOCK(p->p_pptr); sx_xunlock(&proctree_lock); /* * The state PRS_ZOMBIE prevents other proesses from sending * signal to the process, to avoid memory leak, we free memory * for signal queue at the time when the state is set. */ sigqueue_flush(&p->p_sigqueue); sigqueue_flush(&td->td_sigqueue); /* * We have to wait until after acquiring all locks before * changing p_state. We need to avoid all possible context * switches (including ones from blocking on a mutex) while * marked as a zombie. We also have to set the zombie state * before we release the parent process' proc lock to avoid * a lost wakeup. So, we first call wakeup, then we grab the * sched lock, update the state, and release the parent process' * proc lock. */ wakeup(p->p_pptr); cv_broadcast(&p->p_pwait); sched_exit(p->p_pptr, td); PROC_SLOCK(p); p->p_state = PRS_ZOMBIE; PROC_UNLOCK(p->p_pptr); /* * Save our children's rusage information in our exit rusage. */ PROC_STATLOCK(p); ruadd(&p->p_ru, &p->p_rux, &p->p_stats->p_cru, &p->p_crux); PROC_STATUNLOCK(p); /* * Make sure the scheduler takes this thread out of its tables etc. * This will also release this thread's reference to the ucred. * Other thread parts to release include pcb bits and such. */ thread_exit(); } #ifndef _SYS_SYSPROTO_H_ struct abort2_args { char *why; int nargs; void **args; }; #endif int sys_abort2(struct thread *td, struct abort2_args *uap) { struct proc *p = td->td_proc; struct sbuf *sb; void *uargs[16]; int error, i, sig; /* * Do it right now so we can log either proper call of abort2(), or * note, that invalid argument was passed. 512 is big enough to * handle 16 arguments' descriptions with additional comments. */ sb = sbuf_new(NULL, NULL, 512, SBUF_FIXEDLEN); sbuf_clear(sb); sbuf_printf(sb, "%s(pid %d uid %d) aborted: ", p->p_comm, p->p_pid, td->td_ucred->cr_uid); /* * Since we can't return from abort2(), send SIGKILL in cases, where * abort2() was called improperly */ sig = SIGKILL; /* Prevent from DoSes from user-space. */ if (uap->nargs < 0 || uap->nargs > 16) goto out; if (uap->nargs > 0) { if (uap->args == NULL) goto out; error = copyin(uap->args, uargs, uap->nargs * sizeof(void *)); if (error != 0) goto out; } /* * Limit size of 'reason' string to 128. Will fit even when * maximal number of arguments was chosen to be logged. */ if (uap->why != NULL) { error = sbuf_copyin(sb, uap->why, 128); if (error < 0) goto out; } else { sbuf_printf(sb, "(null)"); } if (uap->nargs > 0) { sbuf_printf(sb, "("); for (i = 0;i < uap->nargs; i++) sbuf_printf(sb, "%s%p", i == 0 ? "" : ", ", uargs[i]); sbuf_printf(sb, ")"); } /* * Final stage: arguments were proper, string has been * successfully copied from userspace, and copying pointers * from user-space succeed. */ sig = SIGABRT; out: if (sig == SIGKILL) { sbuf_trim(sb); sbuf_printf(sb, " (Reason text inaccessible)"); } sbuf_cat(sb, "\n"); sbuf_finish(sb); log(LOG_INFO, "%s", sbuf_data(sb)); sbuf_delete(sb); exit1(td, 0, sig); return (0); } #ifdef COMPAT_43 /* * The dirty work is handled by kern_wait(). */ int owait(struct thread *td, struct owait_args *uap __unused) { int error, status; error = kern_wait(td, WAIT_ANY, &status, 0, NULL); if (error == 0) td->td_retval[1] = status; return (error); } #endif /* COMPAT_43 */ /* * The dirty work is handled by kern_wait(). */ int sys_wait4(struct thread *td, struct wait4_args *uap) { struct rusage ru, *rup; int error, status; if (uap->rusage != NULL) rup = &ru; else rup = NULL; error = kern_wait(td, uap->pid, &status, uap->options, rup); - if (uap->status != NULL && error == 0) + if (uap->status != NULL && error == 0 && td->td_retval[0] != 0) error = copyout(&status, uap->status, sizeof(status)); - if (uap->rusage != NULL && error == 0) + if (uap->rusage != NULL && error == 0 && td->td_retval[0] != 0) error = copyout(&ru, uap->rusage, sizeof(struct rusage)); return (error); } int sys_wait6(struct thread *td, struct wait6_args *uap) { struct __wrusage wru, *wrup; siginfo_t si, *sip; idtype_t idtype; id_t id; int error, status; idtype = uap->idtype; id = uap->id; if (uap->wrusage != NULL) wrup = &wru; else wrup = NULL; if (uap->info != NULL) { sip = &si; bzero(sip, sizeof(*sip)); } else sip = NULL; /* * We expect all callers of wait6() to know about WEXITED and * WTRAPPED. */ error = kern_wait6(td, idtype, id, &status, uap->options, wrup, sip); - if (uap->status != NULL && error == 0) + if (uap->status != NULL && error == 0 && td->td_retval[0] != 0) error = copyout(&status, uap->status, sizeof(status)); - if (uap->wrusage != NULL && error == 0) + if (uap->wrusage != NULL && error == 0 && td->td_retval[0] != 0) error = copyout(&wru, uap->wrusage, sizeof(wru)); if (uap->info != NULL && error == 0) error = copyout(&si, uap->info, sizeof(si)); return (error); } /* * Reap the remains of a zombie process and optionally return status and * rusage. Asserts and will release both the proctree_lock and the process * lock as part of its work. */ void proc_reap(struct thread *td, struct proc *p, int *status, int options) { struct proc *q, *t; sx_assert(&proctree_lock, SA_XLOCKED); PROC_LOCK_ASSERT(p, MA_OWNED); PROC_SLOCK_ASSERT(p, MA_OWNED); KASSERT(p->p_state == PRS_ZOMBIE, ("proc_reap: !PRS_ZOMBIE")); q = td->td_proc; PROC_SUNLOCK(p); if (status) *status = KW_EXITCODE(p->p_xexit, p->p_xsig); if (options & WNOWAIT) { /* * Only poll, returning the status. Caller does not wish to * release the proc struct just yet. */ PROC_UNLOCK(p); sx_xunlock(&proctree_lock); return; } PROC_LOCK(q); sigqueue_take(p->p_ksi); PROC_UNLOCK(q); /* * If we got the child via a ptrace 'attach', we need to give it back * to the old parent. */ if (p->p_oppid != 0 && p->p_oppid != p->p_pptr->p_pid) { PROC_UNLOCK(p); t = proc_realparent(p); PROC_LOCK(t); PROC_LOCK(p); CTR2(KTR_PTRACE, "wait: traced child %d moved back to parent %d", p->p_pid, t->p_pid); proc_reparent(p, t); p->p_oppid = 0; PROC_UNLOCK(p); pksignal(t, SIGCHLD, p->p_ksi); wakeup(t); cv_broadcast(&p->p_pwait); PROC_UNLOCK(t); sx_xunlock(&proctree_lock); return; } p->p_oppid = 0; PROC_UNLOCK(p); /* * Remove other references to this process to ensure we have an * exclusive reference. */ sx_xlock(&allproc_lock); LIST_REMOVE(p, p_list); /* off zombproc */ sx_xunlock(&allproc_lock); LIST_REMOVE(p, p_sibling); reaper_abandon_children(p, true); LIST_REMOVE(p, p_reapsibling); PROC_LOCK(p); clear_orphan(p); PROC_UNLOCK(p); leavepgrp(p); if (p->p_procdesc != NULL) procdesc_reap(p); sx_xunlock(&proctree_lock); PROC_LOCK(p); knlist_detach(p->p_klist); p->p_klist = NULL; PROC_UNLOCK(p); /* * Removal from allproc list and process group list paired with * PROC_LOCK which was executed during that time should guarantee * nothing can reach this process anymore. As such further locking * is unnecessary. */ p->p_xexit = p->p_xsig = 0; /* XXX: why? */ PROC_LOCK(q); ruadd(&q->p_stats->p_cru, &q->p_crux, &p->p_ru, &p->p_rux); PROC_UNLOCK(q); /* * Decrement the count of procs running with this uid. */ (void)chgproccnt(p->p_ucred->cr_ruidinfo, -1, 0); /* * Destroy resource accounting information associated with the process. */ #ifdef RACCT if (racct_enable) { PROC_LOCK(p); racct_sub(p, RACCT_NPROC, 1); PROC_UNLOCK(p); } #endif racct_proc_exit(p); /* * Free credentials, arguments, and sigacts. */ crfree(p->p_ucred); proc_set_cred(p, NULL); pargs_drop(p->p_args); p->p_args = NULL; sigacts_free(p->p_sigacts); p->p_sigacts = NULL; /* * Do any thread-system specific cleanups. */ thread_wait(p); /* * Give vm and machine-dependent layer a chance to free anything that * cpu_exit couldn't release while still running in process context. */ vm_waitproc(p); #ifdef MAC mac_proc_destroy(p); #endif /* * Free any domain policy that's still hiding around. */ vm_domain_policy_cleanup(&p->p_vm_dom_policy); KASSERT(FIRST_THREAD_IN_PROC(p), ("proc_reap: no residual thread!")); uma_zfree(proc_zone, p); atomic_add_int(&nprocs, -1); } static int proc_to_reap(struct thread *td, struct proc *p, idtype_t idtype, id_t id, int *status, int options, struct __wrusage *wrusage, siginfo_t *siginfo, int check_only) { struct rusage *rup; sx_assert(&proctree_lock, SA_XLOCKED); PROC_LOCK(p); switch (idtype) { case P_ALL: if (p->p_procdesc != NULL) { PROC_UNLOCK(p); return (0); } break; case P_PID: if (p->p_pid != (pid_t)id) { PROC_UNLOCK(p); return (0); } break; case P_PGID: if (p->p_pgid != (pid_t)id) { PROC_UNLOCK(p); return (0); } break; case P_SID: if (p->p_session->s_sid != (pid_t)id) { PROC_UNLOCK(p); return (0); } break; case P_UID: if (p->p_ucred->cr_uid != (uid_t)id) { PROC_UNLOCK(p); return (0); } break; case P_GID: if (p->p_ucred->cr_gid != (gid_t)id) { PROC_UNLOCK(p); return (0); } break; case P_JAILID: if (p->p_ucred->cr_prison->pr_id != (int)id) { PROC_UNLOCK(p); return (0); } break; /* * It seems that the thread structures get zeroed out * at process exit. This makes it impossible to * support P_SETID, P_CID or P_CPUID. */ default: PROC_UNLOCK(p); return (0); } if (p_canwait(td, p)) { PROC_UNLOCK(p); return (0); } if (((options & WEXITED) == 0) && (p->p_state == PRS_ZOMBIE)) { PROC_UNLOCK(p); return (0); } /* * This special case handles a kthread spawned by linux_clone * (see linux_misc.c). The linux_wait4 and linux_waitpid * functions need to be able to distinguish between waiting * on a process and waiting on a thread. It is a thread if * p_sigparent is not SIGCHLD, and the WLINUXCLONE option * signifies we want to wait for threads and not processes. */ if ((p->p_sigparent != SIGCHLD) ^ ((options & WLINUXCLONE) != 0)) { PROC_UNLOCK(p); return (0); } if (siginfo != NULL) { bzero(siginfo, sizeof(*siginfo)); siginfo->si_errno = 0; /* * SUSv4 requires that the si_signo value is always * SIGCHLD. Obey it despite the rfork(2) interface * allows to request other signal for child exit * notification. */ siginfo->si_signo = SIGCHLD; /* * This is still a rough estimate. We will fix the * cases TRAPPED, STOPPED, and CONTINUED later. */ if (WCOREDUMP(p->p_xsig)) { siginfo->si_code = CLD_DUMPED; siginfo->si_status = WTERMSIG(p->p_xsig); } else if (WIFSIGNALED(p->p_xsig)) { siginfo->si_code = CLD_KILLED; siginfo->si_status = WTERMSIG(p->p_xsig); } else { siginfo->si_code = CLD_EXITED; siginfo->si_status = p->p_xexit; } siginfo->si_pid = p->p_pid; siginfo->si_uid = p->p_ucred->cr_uid; /* * The si_addr field would be useful additional * detail, but apparently the PC value may be lost * when we reach this point. bzero() above sets * siginfo->si_addr to NULL. */ } /* * There should be no reason to limit resources usage info to * exited processes only. A snapshot about any resources used * by a stopped process may be exactly what is needed. */ if (wrusage != NULL) { rup = &wrusage->wru_self; *rup = p->p_ru; PROC_STATLOCK(p); calcru(p, &rup->ru_utime, &rup->ru_stime); PROC_STATUNLOCK(p); rup = &wrusage->wru_children; *rup = p->p_stats->p_cru; calccru(p, &rup->ru_utime, &rup->ru_stime); } if (p->p_state == PRS_ZOMBIE && !check_only) { PROC_SLOCK(p); proc_reap(td, p, status, options); return (-1); } PROC_UNLOCK(p); return (1); } int kern_wait(struct thread *td, pid_t pid, int *status, int options, struct rusage *rusage) { struct __wrusage wru, *wrup; idtype_t idtype; id_t id; int ret; /* * Translate the special pid values into the (idtype, pid) * pair for kern_wait6. The WAIT_MYPGRP case is handled by * kern_wait6() on its own. */ if (pid == WAIT_ANY) { idtype = P_ALL; id = 0; } else if (pid < 0) { idtype = P_PGID; id = (id_t)-pid; } else { idtype = P_PID; id = (id_t)pid; } if (rusage != NULL) wrup = &wru; else wrup = NULL; /* * For backward compatibility we implicitly add flags WEXITED * and WTRAPPED here. */ options |= WEXITED | WTRAPPED; ret = kern_wait6(td, idtype, id, status, options, wrup, NULL); if (rusage != NULL) *rusage = wru.wru_self; return (ret); } int kern_wait6(struct thread *td, idtype_t idtype, id_t id, int *status, int options, struct __wrusage *wrusage, siginfo_t *siginfo) { struct proc *p, *q; pid_t pid; int error, nfound, ret; AUDIT_ARG_VALUE((int)idtype); /* XXX - This is likely wrong! */ AUDIT_ARG_PID((pid_t)id); /* XXX - This may be wrong! */ AUDIT_ARG_VALUE(options); q = td->td_proc; if ((pid_t)id == WAIT_MYPGRP && (idtype == P_PID || idtype == P_PGID)) { PROC_LOCK(q); id = (id_t)q->p_pgid; PROC_UNLOCK(q); idtype = P_PGID; } /* If we don't know the option, just return. */ if ((options & ~(WUNTRACED | WNOHANG | WCONTINUED | WNOWAIT | WEXITED | WTRAPPED | WLINUXCLONE)) != 0) return (EINVAL); if ((options & (WEXITED | WUNTRACED | WCONTINUED | WTRAPPED)) == 0) { /* * We will be unable to find any matching processes, * because there are no known events to look for. * Prefer to return error instead of blocking * indefinitely. */ return (EINVAL); } loop: if (q->p_flag & P_STATCHILD) { PROC_LOCK(q); q->p_flag &= ~P_STATCHILD; PROC_UNLOCK(q); } nfound = 0; sx_xlock(&proctree_lock); LIST_FOREACH(p, &q->p_children, p_sibling) { pid = p->p_pid; ret = proc_to_reap(td, p, idtype, id, status, options, wrusage, siginfo, 0); if (ret == 0) continue; else if (ret == 1) nfound++; else { td->td_retval[0] = pid; return (0); } PROC_LOCK(p); PROC_SLOCK(p); if ((options & WTRAPPED) != 0 && (p->p_flag & P_TRACED) != 0 && (p->p_flag & (P_STOPPED_TRACE | P_STOPPED_SIG)) != 0 && (p->p_suspcount == p->p_numthreads) && ((p->p_flag & P_WAITED) == 0)) { PROC_SUNLOCK(p); if ((options & WNOWAIT) == 0) p->p_flag |= P_WAITED; sx_xunlock(&proctree_lock); if (status != NULL) *status = W_STOPCODE(p->p_xsig); if (siginfo != NULL) { siginfo->si_status = p->p_xsig; siginfo->si_code = CLD_TRAPPED; } if ((options & WNOWAIT) == 0) { PROC_LOCK(q); sigqueue_take(p->p_ksi); PROC_UNLOCK(q); } CTR4(KTR_PTRACE, "wait: returning trapped pid %d status %#x (xstat %d) xthread %d", p->p_pid, W_STOPCODE(p->p_xsig), p->p_xsig, p->p_xthread != NULL ? p->p_xthread->td_tid : -1); PROC_UNLOCK(p); td->td_retval[0] = pid; return (0); } if ((options & WUNTRACED) != 0 && (p->p_flag & P_STOPPED_SIG) != 0 && (p->p_suspcount == p->p_numthreads) && ((p->p_flag & P_WAITED) == 0)) { PROC_SUNLOCK(p); if ((options & WNOWAIT) == 0) p->p_flag |= P_WAITED; sx_xunlock(&proctree_lock); if (status != NULL) *status = W_STOPCODE(p->p_xsig); if (siginfo != NULL) { siginfo->si_status = p->p_xsig; siginfo->si_code = CLD_STOPPED; } if ((options & WNOWAIT) == 0) { PROC_LOCK(q); sigqueue_take(p->p_ksi); PROC_UNLOCK(q); } PROC_UNLOCK(p); td->td_retval[0] = pid; return (0); } PROC_SUNLOCK(p); if ((options & WCONTINUED) != 0 && (p->p_flag & P_CONTINUED) != 0) { sx_xunlock(&proctree_lock); if ((options & WNOWAIT) == 0) { p->p_flag &= ~P_CONTINUED; PROC_LOCK(q); sigqueue_take(p->p_ksi); PROC_UNLOCK(q); } PROC_UNLOCK(p); if (status != NULL) *status = SIGCONT; if (siginfo != NULL) { siginfo->si_status = SIGCONT; siginfo->si_code = CLD_CONTINUED; } td->td_retval[0] = pid; return (0); } PROC_UNLOCK(p); } /* * Look in the orphans list too, to allow the parent to * collect it's child exit status even if child is being * debugged. * * Debugger detaches from the parent upon successful * switch-over from parent to child. At this point due to * re-parenting the parent loses the child to debugger and a * wait4(2) call would report that it has no children to wait * for. By maintaining a list of orphans we allow the parent * to successfully wait until the child becomes a zombie. */ if (nfound == 0) { LIST_FOREACH(p, &q->p_orphans, p_orphan) { ret = proc_to_reap(td, p, idtype, id, NULL, options, NULL, NULL, 1); if (ret != 0) { KASSERT(ret != -1, ("reaped an orphan (pid %d)", (int)td->td_retval[0])); nfound++; break; } } } if (nfound == 0) { sx_xunlock(&proctree_lock); return (ECHILD); } if (options & WNOHANG) { sx_xunlock(&proctree_lock); td->td_retval[0] = 0; return (0); } PROC_LOCK(q); sx_xunlock(&proctree_lock); if (q->p_flag & P_STATCHILD) { q->p_flag &= ~P_STATCHILD; error = 0; } else error = msleep(q, &q->p_mtx, PWAIT | PCATCH, "wait", 0); PROC_UNLOCK(q); if (error) return (error); goto loop; } /* * Make process 'parent' the new parent of process 'child'. * Must be called with an exclusive hold of proctree lock. */ void proc_reparent(struct proc *child, struct proc *parent) { sx_assert(&proctree_lock, SX_XLOCKED); PROC_LOCK_ASSERT(child, MA_OWNED); if (child->p_pptr == parent) return; PROC_LOCK(child->p_pptr); sigqueue_take(child->p_ksi); PROC_UNLOCK(child->p_pptr); LIST_REMOVE(child, p_sibling); LIST_INSERT_HEAD(&parent->p_children, child, p_sibling); clear_orphan(child); if (child->p_flag & P_TRACED) { if (LIST_EMPTY(&child->p_pptr->p_orphans)) { child->p_treeflag |= P_TREE_FIRST_ORPHAN; LIST_INSERT_HEAD(&child->p_pptr->p_orphans, child, p_orphan); } else { LIST_INSERT_AFTER(LIST_FIRST(&child->p_pptr->p_orphans), child, p_orphan); } child->p_treeflag |= P_TREE_ORPHANED; } child->p_pptr = parent; } Index: projects/mips64-clang/sys/kern/kern_mutex.c =================================================================== --- projects/mips64-clang/sys/kern/kern_mutex.c (revision 305711) +++ projects/mips64-clang/sys/kern/kern_mutex.c (revision 305712) @@ -1,1082 +1,1119 @@ /*- * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Berkeley Software Design Inc's name may not be used to endorse or * promote products derived from this software without specific prior * written permission. * * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ */ /* * Machine independent bits of mutex implementation. */ #include __FBSDID("$FreeBSD$"); #include "opt_adaptive_mutexes.h" #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) #define ADAPTIVE_MUTEXES #endif #ifdef HWPMC_HOOKS #include PMC_SOFT_DEFINE( , , lock, failed); #endif /* * Return the mutex address when the lock cookie address is provided. * This functionality assumes that struct mtx* have a member named mtx_lock. */ #define mtxlock2mtx(c) (__containerof(c, struct mtx, mtx_lock)) /* * Internal utility macros. */ #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) #define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED) #define mtx_owner(m) ((struct thread *)((m)->mtx_lock & ~MTX_FLAGMASK)) static void assert_mtx(const struct lock_object *lock, int what); #ifdef DDB static void db_show_mtx(const struct lock_object *lock); #endif static void lock_mtx(struct lock_object *lock, uintptr_t how); static void lock_spin(struct lock_object *lock, uintptr_t how); #ifdef KDTRACE_HOOKS static int owner_mtx(const struct lock_object *lock, struct thread **owner); #endif static uintptr_t unlock_mtx(struct lock_object *lock); static uintptr_t unlock_spin(struct lock_object *lock); /* * Lock classes for sleep and spin mutexes. */ struct lock_class lock_class_mtx_sleep = { .lc_name = "sleep mutex", .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE, .lc_assert = assert_mtx, #ifdef DDB .lc_ddb_show = db_show_mtx, #endif .lc_lock = lock_mtx, .lc_unlock = unlock_mtx, #ifdef KDTRACE_HOOKS .lc_owner = owner_mtx, #endif }; struct lock_class lock_class_mtx_spin = { .lc_name = "spin mutex", .lc_flags = LC_SPINLOCK | LC_RECURSABLE, .lc_assert = assert_mtx, #ifdef DDB .lc_ddb_show = db_show_mtx, #endif .lc_lock = lock_spin, .lc_unlock = unlock_spin, #ifdef KDTRACE_HOOKS .lc_owner = owner_mtx, #endif }; #ifdef ADAPTIVE_MUTEXES static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging"); static struct lock_delay_config mtx_delay = { .initial = 1000, .step = 500, .min = 100, .max = 5000, }; SYSCTL_INT(_debug_mtx, OID_AUTO, delay_initial, CTLFLAG_RW, &mtx_delay.initial, 0, ""); SYSCTL_INT(_debug_mtx, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_delay.step, 0, ""); SYSCTL_INT(_debug_mtx, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_delay.min, 0, ""); SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max, 0, ""); static void mtx_delay_sysinit(void *dummy) { mtx_delay.initial = mp_ncpus * 25; mtx_delay.step = (mp_ncpus * 25) / 2; mtx_delay.min = mp_ncpus * 5; mtx_delay.max = mp_ncpus * 25 * 10; } LOCK_DELAY_SYSINIT(mtx_delay_sysinit); #endif +static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin, CTLFLAG_RD, NULL, + "mtx spin debugging"); + +static struct lock_delay_config mtx_spin_delay = { + .initial = 1000, + .step = 500, + .min = 100, + .max = 5000, +}; + +SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_initial, CTLFLAG_RW, + &mtx_spin_delay.initial, 0, ""); +SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_step, CTLFLAG_RW, &mtx_spin_delay.step, + 0, ""); +SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_min, CTLFLAG_RW, &mtx_spin_delay.min, + 0, ""); +SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_spin_delay.max, + 0, ""); + +static void +mtx_spin_delay_sysinit(void *dummy) +{ + + mtx_spin_delay.initial = mp_ncpus * 25; + mtx_spin_delay.step = (mp_ncpus * 25) / 2; + mtx_spin_delay.min = mp_ncpus * 5; + mtx_spin_delay.max = mp_ncpus * 25 * 10; +} +LOCK_DELAY_SYSINIT(mtx_spin_delay_sysinit); + /* * System-wide mutexes */ struct mtx blocked_lock; struct mtx Giant; void assert_mtx(const struct lock_object *lock, int what) { mtx_assert((const struct mtx *)lock, what); } void lock_mtx(struct lock_object *lock, uintptr_t how) { mtx_lock((struct mtx *)lock); } void lock_spin(struct lock_object *lock, uintptr_t how) { panic("spin locks can only use msleep_spin"); } uintptr_t unlock_mtx(struct lock_object *lock) { struct mtx *m; m = (struct mtx *)lock; mtx_assert(m, MA_OWNED | MA_NOTRECURSED); mtx_unlock(m); return (0); } uintptr_t unlock_spin(struct lock_object *lock) { panic("spin locks can only use msleep_spin"); } #ifdef KDTRACE_HOOKS int owner_mtx(const struct lock_object *lock, struct thread **owner) { const struct mtx *m = (const struct mtx *)lock; *owner = mtx_owner(m); return (mtx_unowned(m) == 0); } #endif /* * Function versions of the inlined __mtx_* macros. These are used by * modules and can also be called from assembly language if needed. */ void __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d", curthread, m->lock_object.lo_name, file, line)); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); __mtx_lock(m, curthread, opts, file, line); LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, line); WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE, file, line); TD_LOCKS_INC(curthread); } void __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, line); mtx_assert(m, MA_OWNED); __mtx_unlock(m, curthread, opts, file, line); TD_LOCKS_DEC(curthread); } void __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("mtx_lock_spin() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); if (mtx_owned(m)) KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || (opts & MTX_RECURSE) != 0, ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); opts &= ~MTX_RECURSE; WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); __mtx_lock_spin(m, curthread, opts, file, line); LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, line); WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); } int __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return (1); m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("mtx_trylock_spin() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); KASSERT((opts & MTX_RECURSE) == 0, ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); if (__mtx_trylock_spin(m, curthread, opts, file, line)) { LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line); WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); return (1); } LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line); return (0); } void __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, line); mtx_assert(m, MA_OWNED); __mtx_unlock_spin(m); } /* * The important part of mtx_trylock{,_flags}() * Tries to acquire lock `m.' If this function is called on a mutex that * is already owned, it will recursively acquire the lock. */ int _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif int rval; if (SCHEDULER_STOPPED()) return (1); m = mtxlock2mtx(c); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d", curthread, m->lock_object.lo_name, file, line)); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); if (mtx_owned(m) && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || (opts & MTX_RECURSE) != 0)) { m->mtx_recurse++; atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); rval = 1; } else rval = _mtx_obtain_lock(m, (uintptr_t)curthread); opts &= ~MTX_RECURSE; LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line); if (rval) { WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, file, line); TD_LOCKS_INC(curthread); if (m->mtx_recurse == 0) LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, waittime, file, line); } return (rval); } /* * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. * * We call this if the lock is either contested (i.e. we need to go to * sleep waiting for it), or if we need to recurse on it. */ void __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts, const char *file, int line) { struct mtx *m; struct turnstile *ts; uintptr_t v; #ifdef ADAPTIVE_MUTEXES volatile struct thread *owner; #endif #ifdef KTR int cont_logged = 0; #endif #ifdef LOCK_PROFILING int contested = 0; uint64_t waittime = 0; #endif #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS) struct lock_delay_arg lda; #endif #ifdef KDTRACE_HOOKS u_int sleep_cnt = 0; int64_t sleep_time = 0; int64_t all_time = 0; #endif if (SCHEDULER_STOPPED()) return; #if defined(ADAPTIVE_MUTEXES) lock_delay_arg_init(&lda, &mtx_delay); #elif defined(KDTRACE_HOOKS) lock_delay_arg_init(&lda, NULL); #endif m = mtxlock2mtx(c); if (mtx_owned(m)) { KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || (opts & MTX_RECURSE) != 0, ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); opts &= ~MTX_RECURSE; m->mtx_recurse++; atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); return; } opts &= ~MTX_RECURSE; #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR4(KTR_LOCK, "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", m->lock_object.lo_name, (void *)m->mtx_lock, file, line); #ifdef KDTRACE_HOOKS all_time -= lockstat_nsecs(&m->lock_object); #endif for (;;) { if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) break; #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif #ifdef ADAPTIVE_MUTEXES /* * If the owner is running on another CPU, spin until the * owner stops running or the state of the lock changes. */ v = m->mtx_lock; if (v != MTX_UNOWNED) { owner = (struct thread *)(v & ~MTX_FLAGMASK); if (TD_IS_RUNNING(owner)) { if (LOCK_LOG_TEST(&m->lock_object, 0)) CTR3(KTR_LOCK, "%s: spinning on %p held by %p", __func__, m, owner); KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "spinning", "lockname:\"%s\"", m->lock_object.lo_name); while (mtx_owner(m) == owner && TD_IS_RUNNING(owner)) lock_delay(&lda); KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "running"); continue; } } #endif ts = turnstile_trywait(&m->lock_object); v = m->mtx_lock; /* * Check if the lock has been released while spinning for * the turnstile chain lock. */ if (v == MTX_UNOWNED) { turnstile_cancel(ts); continue; } #ifdef ADAPTIVE_MUTEXES /* * The current lock owner might have started executing * on another CPU (or the lock could have changed * owners) while we were waiting on the turnstile * chain lock. If so, drop the turnstile lock and try * again. */ owner = (struct thread *)(v & ~MTX_FLAGMASK); if (TD_IS_RUNNING(owner)) { turnstile_cancel(ts); continue; } #endif /* * If the mutex isn't already contested and a failure occurs * setting the contested bit, the mutex was either released * or the state of the MTX_RECURSED bit changed. */ if ((v & MTX_CONTESTED) == 0 && !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { turnstile_cancel(ts); continue; } /* * We definitely must sleep for this lock. */ mtx_assert(m, MA_NOTOWNED); #ifdef KTR if (!cont_logged) { CTR6(KTR_CONTENTION, "contention: %p at %s:%d wants %s, taken by %s:%d", (void *)tid, file, line, m->lock_object.lo_name, WITNESS_FILE(&m->lock_object), WITNESS_LINE(&m->lock_object)); cont_logged = 1; } #endif /* * Block on the turnstile. */ #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&m->lock_object); #endif turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&m->lock_object); sleep_cnt++; #endif } #ifdef KDTRACE_HOOKS all_time += lockstat_nsecs(&m->lock_object); #endif #ifdef KTR if (cont_logged) { CTR4(KTR_CONTENTION, "contention end: %s acquired by %p at %s:%d", m->lock_object.lo_name, (void *)tid, file, line); } #endif LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, waittime, file, line); #ifdef KDTRACE_HOOKS if (sleep_time) LOCKSTAT_RECORD1(adaptive__block, m, sleep_time); /* * Only record the loops spinning and not sleeping. */ if (lda.spin_cnt > sleep_cnt) LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time); #endif } static void _mtx_lock_spin_failed(struct mtx *m) { struct thread *td; td = mtx_owner(m); /* If the mutex is unlocked, try again. */ if (td == NULL) return; printf( "spin lock %p (%s) held by %p (tid %d) too long\n", m, m->lock_object.lo_name, td, td->td_tid); #ifdef WITNESS witness_display_spinlock(&m->lock_object, td, printf); #endif panic("spin lock held too long"); } #ifdef SMP /* * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock. * * This is only called if we need to actually spin for the lock. Recursion * is handled inline. */ void _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts, const char *file, int line) { struct mtx *m; - int i = 0; + struct lock_delay_arg lda; #ifdef LOCK_PROFILING int contested = 0; uint64_t waittime = 0; #endif #ifdef KDTRACE_HOOKS int64_t spin_time = 0; #endif if (SCHEDULER_STOPPED()) return; + lock_delay_arg_init(&lda, &mtx_spin_delay); m = mtxlock2mtx(c); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "spinning", "lockname:\"%s\"", m->lock_object.lo_name); #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); #ifdef KDTRACE_HOOKS spin_time -= lockstat_nsecs(&m->lock_object); #endif for (;;) { if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) break; /* Give interrupts a chance while we spin. */ spinlock_exit(); while (m->mtx_lock != MTX_UNOWNED) { - if (i++ < 10000000) { - cpu_spinwait(); + if (lda.spin_cnt < 10000000) { + lock_delay(&lda); continue; } - if (i < 60000000 || kdb_active || panicstr != NULL) + lda.spin_cnt++; + if (lda.spin_cnt < 60000000 || kdb_active || + panicstr != NULL) DELAY(1); else _mtx_lock_spin_failed(m); cpu_spinwait(); } spinlock_enter(); } #ifdef KDTRACE_HOOKS spin_time += lockstat_nsecs(&m->lock_object); #endif if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "running"); #ifdef KDTRACE_HOOKS LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, contested, waittime, file, line); if (spin_time != 0) LOCKSTAT_RECORD1(spin__spin, m, spin_time); #endif } #endif /* SMP */ void thread_lock_flags_(struct thread *td, int opts, const char *file, int line) { struct mtx *m; uintptr_t tid; - int i; + struct lock_delay_arg lda; #ifdef LOCK_PROFILING int contested = 0; uint64_t waittime = 0; #endif #ifdef KDTRACE_HOOKS int64_t spin_time = 0; #endif - i = 0; tid = (uintptr_t)curthread; if (SCHEDULER_STOPPED()) { /* * Ensure that spinlock sections are balanced even when the * scheduler is stopped, since we may otherwise inadvertently * re-enable interrupts while dumping core. */ spinlock_enter(); return; } + lock_delay_arg_init(&lda, &mtx_spin_delay); + #ifdef KDTRACE_HOOKS spin_time -= lockstat_nsecs(&td->td_lock->lock_object); #endif for (;;) { retry: spinlock_enter(); m = td->td_lock; KASSERT(m->mtx_lock != MTX_DESTROYED, ("thread_lock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("thread_lock() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); if (mtx_owned(m)) KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0, ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); for (;;) { if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid)) break; if (m->mtx_lock == tid) { m->mtx_recurse++; break; } #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); /* Give interrupts a chance while we spin. */ spinlock_exit(); while (m->mtx_lock != MTX_UNOWNED) { - if (i++ < 10000000) + if (lda.spin_cnt < 10000000) { + lock_delay(&lda); + } else { + lda.spin_cnt++; + if (lda.spin_cnt < 60000000 || + kdb_active || panicstr != NULL) + DELAY(1); + else + _mtx_lock_spin_failed(m); cpu_spinwait(); - else if (i < 60000000 || - kdb_active || panicstr != NULL) - DELAY(1); - else - _mtx_lock_spin_failed(m); - cpu_spinwait(); + } if (m != td->td_lock) goto retry; } spinlock_enter(); } if (m == td->td_lock) break; __mtx_unlock_spin(m); /* does spinlock_exit() */ } #ifdef KDTRACE_HOOKS spin_time += lockstat_nsecs(&m->lock_object); #endif if (m->mtx_recurse == 0) LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, contested, waittime, file, line); LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, line); WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); #ifdef KDTRACE_HOOKS if (spin_time != 0) LOCKSTAT_RECORD1(thread__spin, m, spin_time); #endif } struct mtx * thread_lock_block(struct thread *td) { struct mtx *lock; THREAD_LOCK_ASSERT(td, MA_OWNED); lock = td->td_lock; td->td_lock = &blocked_lock; mtx_unlock_spin(lock); return (lock); } void thread_lock_unblock(struct thread *td, struct mtx *new) { mtx_assert(new, MA_OWNED); MPASS(td->td_lock == &blocked_lock); atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new); } void thread_lock_set(struct thread *td, struct mtx *new) { struct mtx *lock; mtx_assert(new, MA_OWNED); THREAD_LOCK_ASSERT(td, MA_OWNED); lock = td->td_lock; td->td_lock = new; mtx_unlock_spin(lock); } /* * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. * * We are only called here if the lock is recursed or contested (i.e. we * need to wake up a blocked thread). */ void __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; struct turnstile *ts; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); if (mtx_recursed(m)) { if (--(m->mtx_recurse) == 0) atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); return; } /* * We have to lock the chain before the turnstile so this turnstile * can be removed from the hash list if it is empty. */ turnstile_chain_lock(&m->lock_object); ts = turnstile_lookup(&m->lock_object); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); MPASS(ts != NULL); turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); _mtx_release_lock_quick(m); /* * This turnstile is now no longer associated with the mutex. We can * unlock the chain lock so a new turnstile may take it's place. */ turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); turnstile_chain_unlock(&m->lock_object); } /* * All the unlocking of MTX_SPIN locks is done inline. * See the __mtx_unlock_spin() macro for the details. */ /* * The backing function for the INVARIANTS-enabled mtx_assert() */ #ifdef INVARIANT_SUPPORT void __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line) { const struct mtx *m; if (panicstr != NULL || dumping) return; m = mtxlock2mtx(c); switch (what) { case MA_OWNED: case MA_OWNED | MA_RECURSED: case MA_OWNED | MA_NOTRECURSED: if (!mtx_owned(m)) panic("mutex %s not owned at %s:%d", m->lock_object.lo_name, file, line); if (mtx_recursed(m)) { if ((what & MA_NOTRECURSED) != 0) panic("mutex %s recursed at %s:%d", m->lock_object.lo_name, file, line); } else if ((what & MA_RECURSED) != 0) { panic("mutex %s unrecursed at %s:%d", m->lock_object.lo_name, file, line); } break; case MA_NOTOWNED: if (mtx_owned(m)) panic("mutex %s owned at %s:%d", m->lock_object.lo_name, file, line); break; default: panic("unknown mtx_assert at %s:%d", file, line); } } #endif /* * General init routine used by the MTX_SYSINIT() macro. */ void mtx_sysinit(void *arg) { struct mtx_args *margs = arg; mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); } /* * Mutex initialization routine; initialize lock `m' of type contained in * `opts' with options contained in `opts' and name `name.' The optional * lock type `type' is used as a general lock category name for use with * witness. */ void _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts) { struct mtx *m; struct lock_class *class; int flags; m = mtxlock2mtx(c); MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0); ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock, ("%s: mtx_lock not aligned for %s: %p", __func__, name, &m->mtx_lock)); /* Determine lock class and lock flags. */ if (opts & MTX_SPIN) class = &lock_class_mtx_spin; else class = &lock_class_mtx_sleep; flags = 0; if (opts & MTX_QUIET) flags |= LO_QUIET; if (opts & MTX_RECURSE) flags |= LO_RECURSABLE; if ((opts & MTX_NOWITNESS) == 0) flags |= LO_WITNESS; if (opts & MTX_DUPOK) flags |= LO_DUPOK; if (opts & MTX_NOPROFILE) flags |= LO_NOPROFILE; if (opts & MTX_NEW) flags |= LO_NEW; /* Initialize mutex. */ lock_init(&m->lock_object, class, name, type, flags); m->mtx_lock = MTX_UNOWNED; m->mtx_recurse = 0; } /* * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be * passed in as a flag here because if the corresponding mtx_init() was * called with MTX_QUIET set, then it will already be set in the mutex's * flags. */ void _mtx_destroy(volatile uintptr_t *c) { struct mtx *m; m = mtxlock2mtx(c); if (!mtx_owned(m)) MPASS(mtx_unowned(m)); else { MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); /* Perform the non-mtx related part of mtx_unlock_spin(). */ if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) spinlock_exit(); else TD_LOCKS_DEC(curthread); lock_profile_release_lock(&m->lock_object); /* Tell witness this isn't locked to make it happy. */ WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__, __LINE__); } m->mtx_lock = MTX_DESTROYED; lock_destroy(&m->lock_object); } /* * Intialize the mutex code and system mutexes. This is called from the MD * startup code prior to mi_startup(). The per-CPU data space needs to be * setup before this is called. */ void mutex_init(void) { /* Setup turnstiles so that sleep mutexes work. */ init_turnstiles(); /* * Initialize mutexes. */ mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN); blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */ mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN); mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN); mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN); mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN); mtx_init(&devmtx, "cdev", NULL, MTX_DEF); mtx_lock(&Giant); } #ifdef DDB void db_show_mtx(const struct lock_object *lock) { struct thread *td; const struct mtx *m; m = (const struct mtx *)lock; db_printf(" flags: {"); if (LOCK_CLASS(lock) == &lock_class_mtx_spin) db_printf("SPIN"); else db_printf("DEF"); if (m->lock_object.lo_flags & LO_RECURSABLE) db_printf(", RECURSE"); if (m->lock_object.lo_flags & LO_DUPOK) db_printf(", DUPOK"); db_printf("}\n"); db_printf(" state: {"); if (mtx_unowned(m)) db_printf("UNOWNED"); else if (mtx_destroyed(m)) db_printf("DESTROYED"); else { db_printf("OWNED"); if (m->mtx_lock & MTX_CONTESTED) db_printf(", CONTESTED"); if (m->mtx_lock & MTX_RECURSED) db_printf(", RECURSED"); } db_printf("}\n"); if (!mtx_unowned(m) && !mtx_destroyed(m)) { td = mtx_owner(m); db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, td->td_tid, td->td_proc->p_pid, td->td_name); if (mtx_recursed(m)) db_printf(" recursed: %d\n", m->mtx_recurse); } } #endif Index: projects/mips64-clang/sys/kern/subr_witness.c =================================================================== --- projects/mips64-clang/sys/kern/subr_witness.c (revision 305711) +++ projects/mips64-clang/sys/kern/subr_witness.c (revision 305712) @@ -1,3017 +1,3025 @@ /*- * Copyright (c) 2008 Isilon Systems, Inc. * Copyright (c) 2008 Ilya Maykov * Copyright (c) 1998 Berkeley Software Design, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Berkeley Software Design Inc's name may not be used to endorse or * promote products derived from this software without specific prior * written permission. * * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ */ /* * Implementation of the `witness' lock verifier. Originally implemented for * mutexes in BSD/OS. Extended to handle generic lock objects and lock * classes in FreeBSD. */ /* * Main Entry: witness * Pronunciation: 'wit-n&s * Function: noun * Etymology: Middle English witnesse, from Old English witnes knowledge, * testimony, witness, from 2wit * Date: before 12th century * 1 : attestation of a fact or event : TESTIMONY * 2 : one that gives evidence; specifically : one who testifies in * a cause or before a judicial tribunal * 3 : one asked to be present at a transaction so as to be able to * testify to its having taken place * 4 : one who has personal knowledge of something * 5 a : something serving as evidence or proof : SIGN * b : public affirmation by word or example of usually * religious faith or conviction * 6 capitalized : a member of the Jehovah's Witnesses */ /* * Special rules concerning Giant and lock orders: * * 1) Giant must be acquired before any other mutexes. Stated another way, * no other mutex may be held when Giant is acquired. * * 2) Giant must be released when blocking on a sleepable lock. * * This rule is less obvious, but is a result of Giant providing the same * semantics as spl(). Basically, when a thread sleeps, it must release * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule * 2). * * 3) Giant may be acquired before or after sleepable locks. * * This rule is also not quite as obvious. Giant may be acquired after * a sleepable lock because it is a non-sleepable lock and non-sleepable * locks may always be acquired while holding a sleepable lock. The second * case, Giant before a sleepable lock, follows from rule 2) above. Suppose * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to * execute. Thus, acquiring Giant both before and after a sleepable lock * will not result in a lock order reversal. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include "opt_stack.h" #include "opt_witness.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif #include #if !defined(DDB) && !defined(STACK) #error "DDB or STACK options are required for WITNESS" #endif /* Note that these traces do not work with KTR_ALQ. */ #if 0 #define KTR_WITNESS KTR_SUBSYS #else #define KTR_WITNESS 0 #endif #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */ #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */ #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */ /* Define this to check for blessed mutexes */ #undef BLESSING #ifndef WITNESS_COUNT #define WITNESS_COUNT 1536 #endif #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */ #define WITNESS_PENDLIST (1024 + MAXCPU) /* Allocate 256 KB of stack data space */ #define WITNESS_LO_DATA_COUNT 2048 /* Prime, gives load factor of ~2 at full load */ #define WITNESS_LO_HASH_SIZE 1021 /* * XXX: This is somewhat bogus, as we assume here that at most 2048 threads * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should * probably be safe for the most part, but it's still a SWAG. */ #define LOCK_NCHILDREN 5 #define LOCK_CHILDCOUNT 2048 #define MAX_W_NAME 64 #define FULLGRAPH_SBUF_SIZE 512 /* * These flags go in the witness relationship matrix and describe the * relationship between any two struct witness objects. */ #define WITNESS_UNRELATED 0x00 /* No lock order relation. */ #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) #define WITNESS_RELATED_MASK \ (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been * observed. */ #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ /* Descendant to ancestor flags */ #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) /* Ancestor to descendant flags */ #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) #define WITNESS_INDEX_ASSERT(i) \ MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count) static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness"); /* * Lock instances. A lock instance is the data associated with a lock while * it is held by witness. For example, a lock instance will hold the * recursion count of a lock. Lock instances are held in lists. Spin locks * are held in a per-cpu list while sleep locks are held in per-thread list. */ struct lock_instance { struct lock_object *li_lock; const char *li_file; int li_line; u_int li_flags; }; /* * A simple list type used to build the list of locks held by a thread * or CPU. We can't simply embed the list in struct lock_object since a * lock may be held by more than one thread if it is a shared lock. Locks * are added to the head of the list, so we fill up each list entry from * "the back" logically. To ease some of the arithmetic, we actually fill * in each list entry the normal way (children[0] then children[1], etc.) but * when we traverse the list we read children[count-1] as the first entry * down to children[0] as the final entry. */ struct lock_list_entry { struct lock_list_entry *ll_next; struct lock_instance ll_children[LOCK_NCHILDREN]; u_int ll_count; }; /* * The main witness structure. One of these per named lock type in the system * (for example, "vnode interlock"). */ struct witness { char w_name[MAX_W_NAME]; uint32_t w_index; /* Index in the relationship matrix */ struct lock_class *w_class; STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ struct witness *w_hash_next; /* Linked list in hash buckets. */ const char *w_file; /* File where last acquired */ uint32_t w_line; /* Line where last acquired */ uint32_t w_refcount; uint16_t w_num_ancestors; /* direct/indirect * ancestor count */ uint16_t w_num_descendants; /* direct/indirect * descendant count */ int16_t w_ddb_level; unsigned w_displayed:1; unsigned w_reversed:1; }; STAILQ_HEAD(witness_list, witness); /* * The witness hash table. Keys are witness names (const char *), elements are * witness objects (struct witness *). */ struct witness_hash { struct witness *wh_array[WITNESS_HASH_SIZE]; uint32_t wh_size; uint32_t wh_count; }; /* * Key type for the lock order data hash table. */ struct witness_lock_order_key { uint16_t from; uint16_t to; }; struct witness_lock_order_data { struct stack wlod_stack; struct witness_lock_order_key wlod_key; struct witness_lock_order_data *wlod_next; }; /* * The witness lock order data hash table. Keys are witness index tuples * (struct witness_lock_order_key), elements are lock order data objects * (struct witness_lock_order_data). */ struct witness_lock_order_hash { struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE]; u_int wloh_size; u_int wloh_count; }; #ifdef BLESSING struct witness_blessed { const char *b_lock1; const char *b_lock2; }; #endif struct witness_pendhelp { const char *wh_type; struct lock_object *wh_lock; }; struct witness_order_list_entry { const char *w_name; struct lock_class *w_class; }; /* * Returns 0 if one of the locks is a spin lock and the other is not. * Returns 1 otherwise. */ static __inline int witness_lock_type_equal(struct witness *w1, struct witness *w2) { return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) == (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))); } static __inline int witness_lock_order_key_equal(const struct witness_lock_order_key *a, const struct witness_lock_order_key *b) { return (a->from == b->from && a->to == b->to); } static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname); static void adopt(struct witness *parent, struct witness *child); #ifdef BLESSING static int blessed(struct witness *, struct witness *); #endif static void depart(struct witness *w); static struct witness *enroll(const char *description, struct lock_class *lock_class); static struct lock_instance *find_instance(struct lock_list_entry *list, const struct lock_object *lock); static int isitmychild(struct witness *parent, struct witness *child); static int isitmydescendant(struct witness *parent, struct witness *child); static void itismychild(struct witness *parent, struct witness *child); static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS); static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS); static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS); static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent); #ifdef DDB static void witness_ddb_compute_levels(void); static void witness_ddb_display(int(*)(const char *fmt, ...)); static void witness_ddb_display_descendants(int(*)(const char *fmt, ...), struct witness *, int indent); static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), struct witness_list *list); static void witness_ddb_level_descendants(struct witness *parent, int l); static void witness_ddb_list(struct thread *td); #endif static void witness_debugger(int cond, const char *msg); static void witness_free(struct witness *m); static struct witness *witness_get(void); static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size); static struct witness *witness_hash_get(const char *key); static void witness_hash_put(struct witness *w); static void witness_init_hash_tables(void); static void witness_increment_graph_generation(void); static void witness_lock_list_free(struct lock_list_entry *lle); static struct lock_list_entry *witness_lock_list_get(void); static int witness_lock_order_add(struct witness *parent, struct witness *child); static int witness_lock_order_check(struct witness *parent, struct witness *child); static struct witness_lock_order_data *witness_lock_order_get( struct witness *parent, struct witness *child); static void witness_list_lock(struct lock_instance *instance, int (*prnt)(const char *fmt, ...)); static int witness_output(const char *fmt, ...) __printflike(1, 2); static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0); static void witness_setflag(struct lock_object *lock, int flag, int set); static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking"); /* * If set to 0, lock order checking is disabled. If set to -1, * witness is completely disabled. Otherwise witness performs full * lock order checking for all locks. At runtime, lock order checking * may be toggled. However, witness cannot be reenabled once it is * completely disabled. */ static int witness_watch = 1; SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0, sysctl_debug_witness_watch, "I", "witness is watching lock operations"); #ifdef KDB /* * When KDB is enabled and witness_kdb is 1, it will cause the system * to drop into kdebug() when: * - a lock hierarchy violation occurs * - locks are held when going to sleep. */ #ifdef WITNESS_KDB int witness_kdb = 1; #else int witness_kdb = 0; #endif SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, ""); #endif /* KDB */ #if defined(DDB) || defined(KDB) /* * When DDB or KDB is enabled and witness_trace is 1, it will cause the system * to print a stack trace: * - a lock hierarchy violation occurs * - locks are held when going to sleep. */ int witness_trace = 1; SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, ""); #endif /* DDB || KDB */ #ifdef WITNESS_SKIPSPIN int witness_skipspin = 1; #else int witness_skipspin = 0; #endif SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, ""); int badstack_sbuf_size; int witness_count = WITNESS_COUNT; SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN, &witness_count, 0, ""); /* * Output channel for witness messages. By default we print to the console. */ enum witness_channel { WITNESS_CONSOLE, WITNESS_LOG, WITNESS_NONE, }; static enum witness_channel witness_channel = WITNESS_CONSOLE; SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING | CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A", "Output channel for warnings"); /* * Call this to print out the relations between locks. */ SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs"); /* * Call this to print out the witness faulty stacks. */ SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks"); static struct mtx w_mtx; /* w_list */ static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); /* w_typelist */ static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); /* lock list */ static struct lock_list_entry *w_lock_list_free = NULL; static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; static u_int pending_cnt; static int w_free_cnt, w_spin_cnt, w_sleep_cnt; SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, ""); static struct witness *w_data; static uint8_t **w_rmatrix; static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; static struct witness_hash w_hash; /* The witness hash table. */ /* The lock order data hash */ static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; static struct witness_lock_order_data *w_lofree = NULL; static struct witness_lock_order_hash w_lohash; static int w_max_used_index = 0; static unsigned int w_generation = 0; static const char w_notrunning[] = "Witness not running\n"; static const char w_stillcold[] = "Witness is still cold\n"; static struct witness_order_list_entry order_lists[] = { /* * sx locks */ { "proctree", &lock_class_sx }, { "allproc", &lock_class_sx }, { "allprison", &lock_class_sx }, { NULL, NULL }, /* * Various mutexes */ { "Giant", &lock_class_mtx_sleep }, { "pipe mutex", &lock_class_mtx_sleep }, { "sigio lock", &lock_class_mtx_sleep }, { "process group", &lock_class_mtx_sleep }, { "process lock", &lock_class_mtx_sleep }, { "session", &lock_class_mtx_sleep }, { "uidinfo hash", &lock_class_rw }, #ifdef HWPMC_HOOKS { "pmc-sleep", &lock_class_mtx_sleep }, #endif { "time lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * umtx */ { "umtx lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * Sockets */ { "accept", &lock_class_mtx_sleep }, { "so_snd", &lock_class_mtx_sleep }, { "so_rcv", &lock_class_mtx_sleep }, { "sellck", &lock_class_mtx_sleep }, { NULL, NULL }, /* * Routing */ { "so_rcv", &lock_class_mtx_sleep }, { "radix node head", &lock_class_rw }, { "rtentry", &lock_class_mtx_sleep }, { "ifaddr", &lock_class_mtx_sleep }, { NULL, NULL }, /* * IPv4 multicast: * protocol locks before interface locks, after UDP locks. */ { "udpinp", &lock_class_rw }, { "in_multi_mtx", &lock_class_mtx_sleep }, { "igmp_mtx", &lock_class_mtx_sleep }, { "if_addr_lock", &lock_class_rw }, { NULL, NULL }, /* * IPv6 multicast: * protocol locks before interface locks, after UDP locks. */ { "udpinp", &lock_class_rw }, { "in6_multi_mtx", &lock_class_mtx_sleep }, { "mld_mtx", &lock_class_mtx_sleep }, { "if_addr_lock", &lock_class_rw }, { NULL, NULL }, /* * UNIX Domain Sockets */ { "unp_link_rwlock", &lock_class_rw }, { "unp_list_lock", &lock_class_mtx_sleep }, { "unp", &lock_class_mtx_sleep }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * UDP/IP */ { "udp", &lock_class_rw }, { "udpinp", &lock_class_rw }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * TCP/IP */ { "tcp", &lock_class_rw }, { "tcpinp", &lock_class_rw }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * BPF */ { "bpf global lock", &lock_class_mtx_sleep }, { "bpf interface lock", &lock_class_rw }, { "bpf cdev lock", &lock_class_mtx_sleep }, { NULL, NULL }, /* * NFS server */ { "nfsd_mtx", &lock_class_mtx_sleep }, { "so_snd", &lock_class_mtx_sleep }, { NULL, NULL }, /* * IEEE 802.11 */ { "802.11 com lock", &lock_class_mtx_sleep}, { NULL, NULL }, /* * Network drivers */ { "network driver", &lock_class_mtx_sleep}, { NULL, NULL }, /* * Netgraph */ { "ng_node", &lock_class_mtx_sleep }, { "ng_worklist", &lock_class_mtx_sleep }, { NULL, NULL }, /* * CDEV */ { "vm map (system)", &lock_class_mtx_sleep }, { "vm page queue", &lock_class_mtx_sleep }, { "vnode interlock", &lock_class_mtx_sleep }, { "cdev", &lock_class_mtx_sleep }, { NULL, NULL }, /* * VM */ { "vm map (user)", &lock_class_sx }, { "vm object", &lock_class_rw }, { "vm page", &lock_class_mtx_sleep }, { "vm page queue", &lock_class_mtx_sleep }, { "pmap pv global", &lock_class_rw }, { "pmap", &lock_class_mtx_sleep }, { "pmap pv list", &lock_class_rw }, { "vm page free queue", &lock_class_mtx_sleep }, { NULL, NULL }, /* * kqueue/VFS interaction */ { "kqueue", &lock_class_mtx_sleep }, { "struct mount mtx", &lock_class_mtx_sleep }, { "vnode interlock", &lock_class_mtx_sleep }, { NULL, NULL }, /* + * VFS namecache + */ + { "ncglobal", &lock_class_rw }, + { "ncbuc", &lock_class_rw }, + { "vnode interlock", &lock_class_mtx_sleep }, + { "ncneg", &lock_class_mtx_sleep }, + { NULL, NULL }, + /* * ZFS locking */ { "dn->dn_mtx", &lock_class_sx }, { "dr->dt.di.dr_mtx", &lock_class_sx }, { "db->db_mtx", &lock_class_sx }, { NULL, NULL }, /* * spin locks */ #ifdef SMP { "ap boot", &lock_class_mtx_spin }, #endif { "rm.mutex_mtx", &lock_class_mtx_spin }, { "sio", &lock_class_mtx_spin }, #ifdef __i386__ { "cy", &lock_class_mtx_spin }, #endif #ifdef __sparc64__ { "pcib_mtx", &lock_class_mtx_spin }, { "rtc_mtx", &lock_class_mtx_spin }, #endif { "scc_hwmtx", &lock_class_mtx_spin }, { "uart_hwmtx", &lock_class_mtx_spin }, { "fast_taskqueue", &lock_class_mtx_spin }, { "intr table", &lock_class_mtx_spin }, #ifdef HWPMC_HOOKS { "pmc-per-proc", &lock_class_mtx_spin }, #endif { "process slock", &lock_class_mtx_spin }, { "syscons video lock", &lock_class_mtx_spin }, { "sleepq chain", &lock_class_mtx_spin }, { "rm_spinlock", &lock_class_mtx_spin }, { "turnstile chain", &lock_class_mtx_spin }, { "turnstile lock", &lock_class_mtx_spin }, { "sched lock", &lock_class_mtx_spin }, { "td_contested", &lock_class_mtx_spin }, { "callout", &lock_class_mtx_spin }, { "entropy harvest mutex", &lock_class_mtx_spin }, #ifdef SMP { "smp rendezvous", &lock_class_mtx_spin }, #endif #ifdef __powerpc__ { "tlb0", &lock_class_mtx_spin }, #endif /* * leaf locks */ { "intrcnt", &lock_class_mtx_spin }, { "icu", &lock_class_mtx_spin }, #if defined(SMP) && defined(__sparc64__) { "ipi", &lock_class_mtx_spin }, #endif #ifdef __i386__ { "allpmaps", &lock_class_mtx_spin }, { "descriptor tables", &lock_class_mtx_spin }, #endif { "clk", &lock_class_mtx_spin }, { "cpuset", &lock_class_mtx_spin }, { "mprof lock", &lock_class_mtx_spin }, { "zombie lock", &lock_class_mtx_spin }, { "ALD Queue", &lock_class_mtx_spin }, #if defined(__i386__) || defined(__amd64__) { "pcicfg", &lock_class_mtx_spin }, { "NDIS thread lock", &lock_class_mtx_spin }, #endif { "tw_osl_io_lock", &lock_class_mtx_spin }, { "tw_osl_q_lock", &lock_class_mtx_spin }, { "tw_cl_io_lock", &lock_class_mtx_spin }, { "tw_cl_intr_lock", &lock_class_mtx_spin }, { "tw_cl_gen_lock", &lock_class_mtx_spin }, #ifdef HWPMC_HOOKS { "pmc-leaf", &lock_class_mtx_spin }, #endif { "blocked lock", &lock_class_mtx_spin }, { NULL, NULL }, { NULL, NULL } }; #ifdef BLESSING /* * Pairs of locks which have been blessed * Don't complain about order problems with blessed locks */ static struct witness_blessed blessed_list[] = { }; #endif /* * This global is set to 0 once it becomes safe to use the witness code. */ static int witness_cold = 1; /* * This global is set to 1 once the static lock orders have been enrolled * so that a warning can be issued for any spin locks enrolled later. */ static int witness_spin_warn = 0; /* Trim useless garbage from filenames. */ static const char * fixup_filename(const char *file) { if (file == NULL) return (NULL); while (strncmp(file, "../", 3) == 0) file += 3; return (file); } /* * The WITNESS-enabled diagnostic code. Note that the witness code does * assume that the early boot is single-threaded at least until after this * routine is completed. */ static void witness_initialize(void *dummy __unused) { struct lock_object *lock; struct witness_order_list_entry *order; struct witness *w, *w1; int i; w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS, M_WAITOK | M_ZERO); w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO); for (i = 0; i < witness_count + 1; i++) { w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) * (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO); } badstack_sbuf_size = witness_count * 256; /* * We have to release Giant before initializing its witness * structure so that WITNESS doesn't get confused. */ mtx_unlock(&Giant); mtx_assert(&Giant, MA_NOTOWNED); CTR1(KTR_WITNESS, "%s: initializing witness", __func__); mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | MTX_NOWITNESS | MTX_NOPROFILE); for (i = witness_count - 1; i >= 0; i--) { w = &w_data[i]; memset(w, 0, sizeof(*w)); w_data[i].w_index = i; /* Witness index never changes. */ witness_free(w); } KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, ("%s: Invalid list of free witness objects", __func__)); /* Witness with index 0 is not used to aid in debugging. */ STAILQ_REMOVE_HEAD(&w_free, w_list); w_free_cnt--; for (i = 0; i < witness_count; i++) { memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) * (witness_count + 1)); } for (i = 0; i < LOCK_CHILDCOUNT; i++) witness_lock_list_free(&w_locklistdata[i]); witness_init_hash_tables(); /* First add in all the specified order lists. */ for (order = order_lists; order->w_name != NULL; order++) { w = enroll(order->w_name, order->w_class); if (w == NULL) continue; w->w_file = "order list"; for (order++; order->w_name != NULL; order++) { w1 = enroll(order->w_name, order->w_class); if (w1 == NULL) continue; w1->w_file = "order list"; itismychild(w, w1); w = w1; } } witness_spin_warn = 1; /* Iterate through all locks and add them to witness. */ for (i = 0; pending_locks[i].wh_lock != NULL; i++) { lock = pending_locks[i].wh_lock; KASSERT(lock->lo_flags & LO_WITNESS, ("%s: lock %s is on pending list but not LO_WITNESS", __func__, lock->lo_name)); lock->lo_witness = enroll(pending_locks[i].wh_type, LOCK_CLASS(lock)); } /* Mark the witness code as being ready for use. */ witness_cold = 0; mtx_lock(&Giant); } SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL); void witness_init(struct lock_object *lock, const char *type) { struct lock_class *class; /* Various sanity checks. */ class = LOCK_CLASS(lock); if ((lock->lo_flags & LO_RECURSABLE) != 0 && (class->lc_flags & LC_RECURSABLE) == 0) kassert_panic("%s: lock (%s) %s can not be recursable", __func__, class->lc_name, lock->lo_name); if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (class->lc_flags & LC_SLEEPABLE) == 0) kassert_panic("%s: lock (%s) %s can not be sleepable", __func__, class->lc_name, lock->lo_name); if ((lock->lo_flags & LO_UPGRADABLE) != 0 && (class->lc_flags & LC_UPGRADABLE) == 0) kassert_panic("%s: lock (%s) %s can not be upgradable", __func__, class->lc_name, lock->lo_name); /* * If we shouldn't watch this lock, then just clear lo_witness. * Otherwise, if witness_cold is set, then it is too early to * enroll this lock, so defer it to witness_initialize() by adding * it to the pending_locks list. If it is not too early, then enroll * the lock now. */ if (witness_watch < 1 || panicstr != NULL || (lock->lo_flags & LO_WITNESS) == 0) lock->lo_witness = NULL; else if (witness_cold) { pending_locks[pending_cnt].wh_lock = lock; pending_locks[pending_cnt++].wh_type = type; if (pending_cnt > WITNESS_PENDLIST) panic("%s: pending locks list is too small, " "increase WITNESS_PENDLIST\n", __func__); } else lock->lo_witness = enroll(type, class); } void witness_destroy(struct lock_object *lock) { struct lock_class *class; struct witness *w; class = LOCK_CLASS(lock); if (witness_cold) panic("lock (%s) %s destroyed while witness_cold", class->lc_name, lock->lo_name); /* XXX: need to verify that no one holds the lock */ if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) return; w = lock->lo_witness; mtx_lock_spin(&w_mtx); MPASS(w->w_refcount > 0); w->w_refcount--; if (w->w_refcount == 0) depart(w); mtx_unlock_spin(&w_mtx); } #ifdef DDB static void witness_ddb_compute_levels(void) { struct witness *w; /* * First clear all levels. */ STAILQ_FOREACH(w, &w_all, w_list) w->w_ddb_level = -1; /* * Look for locks with no parents and level all their descendants. */ STAILQ_FOREACH(w, &w_all, w_list) { /* If the witness has ancestors (is not a root), skip it. */ if (w->w_num_ancestors > 0) continue; witness_ddb_level_descendants(w, 0); } } static void witness_ddb_level_descendants(struct witness *w, int l) { int i; if (w->w_ddb_level >= l) return; w->w_ddb_level = l; l++; for (i = 1; i <= w_max_used_index; i++) { if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) witness_ddb_level_descendants(&w_data[i], l); } } static void witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...), struct witness *w, int indent) { int i; for (i = 0; i < indent; i++) prnt(" "); prnt("%s (type: %s, depth: %d, active refs: %d)", w->w_name, w->w_class->lc_name, w->w_ddb_level, w->w_refcount); if (w->w_displayed) { prnt(" -- (already displayed)\n"); return; } w->w_displayed = 1; if (w->w_file != NULL && w->w_line != 0) prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file), w->w_line); else prnt(" -- never acquired\n"); indent++; WITNESS_INDEX_ASSERT(w->w_index); for (i = 1; i <= w_max_used_index; i++) { if (db_pager_quit) return; if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) witness_ddb_display_descendants(prnt, &w_data[i], indent); } } static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), struct witness_list *list) { struct witness *w; STAILQ_FOREACH(w, list, w_typelist) { if (w->w_file == NULL || w->w_ddb_level > 0) continue; /* This lock has no anscestors - display its descendants. */ witness_ddb_display_descendants(prnt, w, 0); if (db_pager_quit) return; } } static void witness_ddb_display(int(*prnt)(const char *fmt, ...)) { struct witness *w; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); witness_ddb_compute_levels(); /* Clear all the displayed flags. */ STAILQ_FOREACH(w, &w_all, w_list) w->w_displayed = 0; /* * First, handle sleep locks which have been acquired at least * once. */ prnt("Sleep locks:\n"); witness_ddb_display_list(prnt, &w_sleep); if (db_pager_quit) return; /* * Now do spin locks which have been acquired at least once. */ prnt("\nSpin locks:\n"); witness_ddb_display_list(prnt, &w_spin); if (db_pager_quit) return; /* * Finally, any locks which have not been acquired yet. */ prnt("\nLocks which were never acquired:\n"); STAILQ_FOREACH(w, &w_all, w_list) { if (w->w_file != NULL || w->w_refcount == 0) continue; prnt("%s (type: %s, depth: %d)\n", w->w_name, w->w_class->lc_name, w->w_ddb_level); if (db_pager_quit) return; } } #endif /* DDB */ int witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) { if (witness_watch == -1 || panicstr != NULL) return (0); /* Require locks that witness knows about. */ if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || lock2->lo_witness == NULL) return (EINVAL); mtx_assert(&w_mtx, MA_NOTOWNED); mtx_lock_spin(&w_mtx); /* * If we already have either an explicit or implied lock order that * is the other way around, then return an error. */ if (witness_watch && isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { mtx_unlock_spin(&w_mtx); return (EDOOFUS); } /* Try to add the new order. */ CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, lock2->lo_witness->w_name, lock1->lo_witness->w_name); itismychild(lock1->lo_witness, lock2->lo_witness); mtx_unlock_spin(&w_mtx); return (0); } void witness_checkorder(struct lock_object *lock, int flags, const char *file, int line, struct lock_object *interlock) { struct lock_list_entry *lock_list, *lle; struct lock_instance *lock1, *lock2, *plock; struct lock_class *class, *iclass; struct witness *w, *w1; struct thread *td; int i, j; if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || panicstr != NULL) return; w = lock->lo_witness; class = LOCK_CLASS(lock); td = curthread; if (class->lc_flags & LC_SLEEPLOCK) { /* * Since spin locks include a critical section, this check * implicitly enforces a lock order of all sleep locks before * all spin locks. */ if (td->td_critnest != 0 && !kdb_active) kassert_panic("acquiring blockable sleep lock with " "spinlock or critical section held (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); /* * If this is the first lock acquired then just return as * no order checking is needed. */ lock_list = td->td_sleeplocks; if (lock_list == NULL || lock_list->ll_count == 0) return; } else { /* * If this is the first lock, just return as no order * checking is needed. Avoid problems with thread * migration pinning the thread while checking if * spinlocks are held. If at least one spinlock is held * the thread is in a safe path and it is allowed to * unpin it. */ sched_pin(); lock_list = PCPU_GET(spinlocks); if (lock_list == NULL || lock_list->ll_count == 0) { sched_unpin(); return; } sched_unpin(); } /* * Check to see if we are recursing on a lock we already own. If * so, make sure that we don't mismatch exclusive and shared lock * acquires. */ lock1 = find_instance(lock_list, lock); if (lock1 != NULL) { if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && (flags & LOP_EXCLUSIVE) == 0) { witness_output("shared lock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while exclusively locked from %s:%d\n", fixup_filename(lock1->li_file), lock1->li_line); kassert_panic("excl->share"); } if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && (flags & LOP_EXCLUSIVE) != 0) { witness_output("exclusive lock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while share locked from %s:%d\n", fixup_filename(lock1->li_file), lock1->li_line); kassert_panic("share->excl"); } return; } /* Warn if the interlock is not locked exactly once. */ if (interlock != NULL) { iclass = LOCK_CLASS(interlock); lock1 = find_instance(lock_list, interlock); if (lock1 == NULL) kassert_panic("interlock (%s) %s not locked @ %s:%d", iclass->lc_name, interlock->lo_name, fixup_filename(file), line); else if ((lock1->li_flags & LI_RECURSEMASK) != 0) kassert_panic("interlock (%s) %s recursed @ %s:%d", iclass->lc_name, interlock->lo_name, fixup_filename(file), line); } /* * Find the previously acquired lock, but ignore interlocks. */ plock = &lock_list->ll_children[lock_list->ll_count - 1]; if (interlock != NULL && plock->li_lock == interlock) { if (lock_list->ll_count > 1) plock = &lock_list->ll_children[lock_list->ll_count - 2]; else { lle = lock_list->ll_next; /* * The interlock is the only lock we hold, so * simply return. */ if (lle == NULL) return; plock = &lle->ll_children[lle->ll_count - 1]; } } /* * Try to perform most checks without a lock. If this succeeds we * can skip acquiring the lock and return success. Otherwise we redo * the check with the lock held to handle races with concurrent updates. */ w1 = plock->li_lock->lo_witness; if (witness_lock_order_check(w1, w)) return; mtx_lock_spin(&w_mtx); if (witness_lock_order_check(w1, w)) { mtx_unlock_spin(&w_mtx); return; } witness_lock_order_add(w1, w); /* * Check for duplicate locks of the same type. Note that we only * have to check for this on the last lock we just acquired. Any * other cases will be caught as lock order violations. */ if (w1 == w) { i = w->w_index; if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { w_rmatrix[i][i] |= WITNESS_REVERSAL; w->w_reversed = 1; mtx_unlock_spin(&w_mtx); witness_output( "acquiring duplicate lock of same type: \"%s\"\n", w->w_name); witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, fixup_filename(plock->li_file), plock->li_line); witness_output(" 2nd %s @ %s:%d\n", lock->lo_name, fixup_filename(file), line); witness_debugger(1, __func__); } else mtx_unlock_spin(&w_mtx); return; } mtx_assert(&w_mtx, MA_OWNED); /* * If we know that the lock we are acquiring comes after * the lock we most recently acquired in the lock order tree, * then there is no need for any further checks. */ if (isitmychild(w1, w)) goto out; for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { for (i = lle->ll_count - 1; i >= 0; i--, j++) { MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN); lock1 = &lle->ll_children[i]; /* * Ignore the interlock. */ if (interlock == lock1->li_lock) continue; /* * If this lock doesn't undergo witness checking, * then skip it. */ w1 = lock1->li_lock->lo_witness; if (w1 == NULL) { KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, ("lock missing witness structure")); continue; } /* * If we are locking Giant and this is a sleepable * lock, then skip it. */ if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && lock == &Giant.lock_object) continue; /* * If we are locking a sleepable lock and this lock * is Giant, then skip it. */ if ((lock->lo_flags & LO_SLEEPABLE) != 0 && lock1->li_lock == &Giant.lock_object) continue; /* * If we are locking a sleepable lock and this lock * isn't sleepable, we want to treat it as a lock * order violation to enfore a general lock order of * sleepable locks before non-sleepable locks. */ if (((lock->lo_flags & LO_SLEEPABLE) != 0 && (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) goto reversal; /* * If we are locking Giant and this is a non-sleepable * lock, then treat it as a reversal. */ if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && lock == &Giant.lock_object) goto reversal; /* * Check the lock order hierarchy for a reveresal. */ if (!isitmydescendant(w, w1)) continue; reversal: /* * We have a lock order violation, check to see if it * is allowed or has already been yelled about. */ #ifdef BLESSING /* * If the lock order is blessed, just bail. We don't * look for other lock order violations though, which * may be a bug. */ if (blessed(w, w1)) goto out; #endif /* Bail if this violation is known */ if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) goto out; /* Record this as a violation */ w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; w->w_reversed = w1->w_reversed = 1; witness_increment_graph_generation(); mtx_unlock_spin(&w_mtx); #ifdef WITNESS_NO_VNODE /* * There are known LORs between VNODE locks. They are * not an indication of a bug. VNODE locks are flagged * as such (LO_IS_VNODE) and we don't yell if the LOR * is between 2 VNODE locks. */ if ((lock->lo_flags & LO_IS_VNODE) != 0 && (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0) return; #endif /* * Ok, yell about it. */ if (((lock->lo_flags & LO_SLEEPABLE) != 0 && (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) witness_output( "lock order reversal: (sleepable after non-sleepable)\n"); else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && lock == &Giant.lock_object) witness_output( "lock order reversal: (Giant after non-sleepable)\n"); else witness_output("lock order reversal:\n"); /* * Try to locate an earlier lock with * witness w in our list. */ do { lock2 = &lle->ll_children[i]; MPASS(lock2->li_lock != NULL); if (lock2->li_lock->lo_witness == w) break; if (i == 0 && lle->ll_next != NULL) { lle = lle->ll_next; i = lle->ll_count - 1; MPASS(i >= 0 && i < LOCK_NCHILDREN); } else i--; } while (i >= 0); if (i < 0) { witness_output(" 1st %p %s (%s) @ %s:%d\n", lock1->li_lock, lock1->li_lock->lo_name, w1->w_name, fixup_filename(lock1->li_file), lock1->li_line); witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock, lock->lo_name, w->w_name, fixup_filename(file), line); } else { witness_output(" 1st %p %s (%s) @ %s:%d\n", lock2->li_lock, lock2->li_lock->lo_name, lock2->li_lock->lo_witness->w_name, fixup_filename(lock2->li_file), lock2->li_line); witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock1->li_lock, lock1->li_lock->lo_name, w1->w_name, fixup_filename(lock1->li_file), lock1->li_line); witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock, lock->lo_name, w->w_name, fixup_filename(file), line); } witness_debugger(1, __func__); return; } } /* * If requested, build a new lock order. However, don't build a new * relationship between a sleepable lock and Giant if it is in the * wrong direction. The correct lock order is that sleepable locks * always come before Giant. */ if (flags & LOP_NEWORDER && !(plock->li_lock == &Giant.lock_object && (lock->lo_flags & LO_SLEEPABLE) != 0)) { CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, w->w_name, plock->li_lock->lo_witness->w_name); itismychild(plock->li_lock->lo_witness, w); } out: mtx_unlock_spin(&w_mtx); } void witness_lock(struct lock_object *lock, int flags, const char *file, int line) { struct lock_list_entry **lock_list, *lle; struct lock_instance *instance; struct witness *w; struct thread *td; if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || panicstr != NULL) return; w = lock->lo_witness; td = curthread; /* Determine lock list for this lock. */ if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) lock_list = &td->td_sleeplocks; else lock_list = PCPU_PTR(spinlocks); /* Check to see if we are recursing on a lock we already own. */ instance = find_instance(*lock_list, lock); if (instance != NULL) { instance->li_flags++; CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, td->td_proc->p_pid, lock->lo_name, instance->li_flags & LI_RECURSEMASK); instance->li_file = file; instance->li_line = line; return; } /* Update per-witness last file and line acquire. */ w->w_file = file; w->w_line = line; /* Find the next open lock instance in the list and fill it. */ lle = *lock_list; if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { lle = witness_lock_list_get(); if (lle == NULL) return; lle->ll_next = *lock_list; CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, td->td_proc->p_pid, lle); *lock_list = lle; } instance = &lle->ll_children[lle->ll_count++]; instance->li_lock = lock; instance->li_line = line; instance->li_file = file; if ((flags & LOP_EXCLUSIVE) != 0) instance->li_flags = LI_EXCLUSIVE; else instance->li_flags = 0; CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); } void witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) { struct lock_instance *instance; struct lock_class *class; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (witness_watch) { if ((lock->lo_flags & LO_UPGRADABLE) == 0) kassert_panic( "upgrade of non-upgradable lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((class->lc_flags & LC_SLEEPLOCK) == 0) kassert_panic( "upgrade of non-sleep lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); } instance = find_instance(curthread->td_sleeplocks, lock); if (instance == NULL) { kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); return; } if (witness_watch) { if ((instance->li_flags & LI_EXCLUSIVE) != 0) kassert_panic( "upgrade of exclusive lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((instance->li_flags & LI_RECURSEMASK) != 0) kassert_panic( "upgrade of recursed lock (%s) %s r=%d @ %s:%d", class->lc_name, lock->lo_name, instance->li_flags & LI_RECURSEMASK, fixup_filename(file), line); } instance->li_flags |= LI_EXCLUSIVE; } void witness_downgrade(struct lock_object *lock, int flags, const char *file, int line) { struct lock_instance *instance; struct lock_class *class; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (witness_watch) { if ((lock->lo_flags & LO_UPGRADABLE) == 0) kassert_panic( "downgrade of non-upgradable lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((class->lc_flags & LC_SLEEPLOCK) == 0) kassert_panic( "downgrade of non-sleep lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); } instance = find_instance(curthread->td_sleeplocks, lock); if (instance == NULL) { kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); return; } if (witness_watch) { if ((instance->li_flags & LI_EXCLUSIVE) == 0) kassert_panic( "downgrade of shared lock (%s) %s @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((instance->li_flags & LI_RECURSEMASK) != 0) kassert_panic( "downgrade of recursed lock (%s) %s r=%d @ %s:%d", class->lc_name, lock->lo_name, instance->li_flags & LI_RECURSEMASK, fixup_filename(file), line); } instance->li_flags &= ~LI_EXCLUSIVE; } void witness_unlock(struct lock_object *lock, int flags, const char *file, int line) { struct lock_list_entry **lock_list, *lle; struct lock_instance *instance; struct lock_class *class; struct thread *td; register_t s; int i, j; if (witness_cold || lock->lo_witness == NULL || panicstr != NULL) return; td = curthread; class = LOCK_CLASS(lock); /* Find lock instance associated with this lock. */ if (class->lc_flags & LC_SLEEPLOCK) lock_list = &td->td_sleeplocks; else lock_list = PCPU_PTR(spinlocks); lle = *lock_list; for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) for (i = 0; i < (*lock_list)->ll_count; i++) { instance = &(*lock_list)->ll_children[i]; if (instance->li_lock == lock) goto found; } /* * When disabling WITNESS through witness_watch we could end up in * having registered locks in the td_sleeplocks queue. * We have to make sure we flush these queues, so just search for * eventual register locks and remove them. */ if (witness_watch > 0) { kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, fixup_filename(file), line); return; } else { return; } found: /* First, check for shared/exclusive mismatches. */ if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && (flags & LOP_EXCLUSIVE) == 0) { witness_output("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while exclusively locked from %s:%d\n", fixup_filename(instance->li_file), instance->li_line); kassert_panic("excl->ushare"); } if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && (flags & LOP_EXCLUSIVE) != 0) { witness_output("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); witness_output("while share locked from %s:%d\n", fixup_filename(instance->li_file), instance->li_line); kassert_panic("share->uexcl"); } /* If we are recursed, unrecurse. */ if ((instance->li_flags & LI_RECURSEMASK) > 0) { CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, td->td_proc->p_pid, instance->li_lock->lo_name, instance->li_flags); instance->li_flags--; return; } /* The lock is now being dropped, check for NORELEASE flag */ if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { witness_output("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name, lock->lo_name, fixup_filename(file), line); kassert_panic("lock marked norelease"); } /* Otherwise, remove this item from the list. */ s = intr_disable(); CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, td->td_proc->p_pid, instance->li_lock->lo_name, (*lock_list)->ll_count - 1); for (j = i; j < (*lock_list)->ll_count - 1; j++) (*lock_list)->ll_children[j] = (*lock_list)->ll_children[j + 1]; (*lock_list)->ll_count--; intr_restore(s); /* * In order to reduce contention on w_mtx, we want to keep always an * head object into lists so that frequent allocation from the * free witness pool (and subsequent locking) is avoided. * In order to maintain the current code simple, when the head * object is totally unloaded it means also that we do not have * further objects in the list, so the list ownership needs to be * hand over to another object if the current head needs to be freed. */ if ((*lock_list)->ll_count == 0) { if (*lock_list == lle) { if (lle->ll_next == NULL) return; } else lle = *lock_list; *lock_list = lle->ll_next; CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, td->td_proc->p_pid, lle); witness_lock_list_free(lle); } } void witness_thread_exit(struct thread *td) { struct lock_list_entry *lle; int i, n; lle = td->td_sleeplocks; if (lle == NULL || panicstr != NULL) return; if (lle->ll_count != 0) { for (n = 0; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { if (n == 0) witness_output( "Thread %p exiting with the following locks held:\n", td); n++; witness_list_lock(&lle->ll_children[i], witness_output); } kassert_panic( "Thread %p cannot exit while holding sleeplocks\n", td); } witness_lock_list_free(lle); } /* * Warn if any locks other than 'lock' are held. Flags can be passed in to * exempt Giant and sleepable locks from the checks as well. If any * non-exempt locks are held, then a supplied message is printed to the * output channel along with a list of the offending locks. If indicated in the * flags then a failure results in a panic as well. */ int witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) { struct lock_list_entry *lock_list, *lle; struct lock_instance *lock1; struct thread *td; va_list ap; int i, n; if (witness_cold || witness_watch < 1 || panicstr != NULL) return (0); n = 0; td = curthread; for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { lock1 = &lle->ll_children[i]; if (lock1->li_lock == lock) continue; if (flags & WARN_GIANTOK && lock1->li_lock == &Giant.lock_object) continue; if (flags & WARN_SLEEPOK && (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) continue; if (n == 0) { va_start(ap, fmt); witness_voutput(fmt, ap); va_end(ap); witness_output( " with the following %slocks held:\n", (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : ""); } n++; witness_list_lock(lock1, witness_output); } /* * Pin the thread in order to avoid problems with thread migration. * Once that all verifies are passed about spinlocks ownership, * the thread is in a safe path and it can be unpinned. */ sched_pin(); lock_list = PCPU_GET(spinlocks); if (lock_list != NULL && lock_list->ll_count != 0) { sched_unpin(); /* * We should only have one spinlock and as long as * the flags cannot match for this locks class, * check if the first spinlock is the one curthread * should hold. */ lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && lock1->li_lock == lock && n == 0) return (0); va_start(ap, fmt); witness_voutput(fmt, ap); va_end(ap); witness_output(" with the following %slocks held:\n", (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : ""); n += witness_list_locks(&lock_list, witness_output); } else sched_unpin(); if (flags & WARN_PANIC && n) kassert_panic("%s", __func__); else witness_debugger(n, __func__); return (n); } const char * witness_file(struct lock_object *lock) { struct witness *w; if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) return ("?"); w = lock->lo_witness; return (w->w_file); } int witness_line(struct lock_object *lock) { struct witness *w; if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) return (0); w = lock->lo_witness; return (w->w_line); } static struct witness * enroll(const char *description, struct lock_class *lock_class) { struct witness *w; struct witness_list *typelist; MPASS(description != NULL); if (witness_watch == -1 || panicstr != NULL) return (NULL); if ((lock_class->lc_flags & LC_SPINLOCK)) { if (witness_skipspin) return (NULL); else typelist = &w_spin; } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) { typelist = &w_sleep; } else { kassert_panic("lock class %s is not sleep or spin", lock_class->lc_name); return (NULL); } mtx_lock_spin(&w_mtx); w = witness_hash_get(description); if (w) goto found; if ((w = witness_get()) == NULL) return (NULL); MPASS(strlen(description) < MAX_W_NAME); strcpy(w->w_name, description); w->w_class = lock_class; w->w_refcount = 1; STAILQ_INSERT_HEAD(&w_all, w, w_list); if (lock_class->lc_flags & LC_SPINLOCK) { STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); w_spin_cnt++; } else if (lock_class->lc_flags & LC_SLEEPLOCK) { STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); w_sleep_cnt++; } /* Insert new witness into the hash */ witness_hash_put(w); witness_increment_graph_generation(); mtx_unlock_spin(&w_mtx); return (w); found: w->w_refcount++; mtx_unlock_spin(&w_mtx); if (lock_class != w->w_class) kassert_panic( "lock (%s) %s does not match earlier (%s) lock", description, lock_class->lc_name, w->w_class->lc_name); return (w); } static void depart(struct witness *w) { struct witness_list *list; MPASS(w->w_refcount == 0); if (w->w_class->lc_flags & LC_SLEEPLOCK) { list = &w_sleep; w_sleep_cnt--; } else { list = &w_spin; w_spin_cnt--; } /* * Set file to NULL as it may point into a loadable module. */ w->w_file = NULL; w->w_line = 0; witness_increment_graph_generation(); } static void adopt(struct witness *parent, struct witness *child) { int pi, ci, i, j; if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); /* If the relationship is already known, there's no work to be done. */ if (isitmychild(parent, child)) return; /* When the structure of the graph changes, bump up the generation. */ witness_increment_graph_generation(); /* * The hard part ... create the direct relationship, then propagate all * indirect relationships. */ pi = parent->w_index; ci = child->w_index; WITNESS_INDEX_ASSERT(pi); WITNESS_INDEX_ASSERT(ci); MPASS(pi != ci); w_rmatrix[pi][ci] |= WITNESS_PARENT; w_rmatrix[ci][pi] |= WITNESS_CHILD; /* * If parent was not already an ancestor of child, * then we increment the descendant and ancestor counters. */ if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { parent->w_num_descendants++; child->w_num_ancestors++; } /* * Find each ancestor of 'pi'. Note that 'pi' itself is counted as * an ancestor of 'pi' during this loop. */ for (i = 1; i <= w_max_used_index; i++) { if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && (i != pi)) continue; /* Find each descendant of 'i' and mark it as a descendant. */ for (j = 1; j <= w_max_used_index; j++) { /* * Skip children that are already marked as * descendants of 'i'. */ if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) continue; /* * We are only interested in descendants of 'ci'. Note * that 'ci' itself is counted as a descendant of 'ci'. */ if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && (j != ci)) continue; w_rmatrix[i][j] |= WITNESS_ANCESTOR; w_rmatrix[j][i] |= WITNESS_DESCENDANT; w_data[i].w_num_descendants++; w_data[j].w_num_ancestors++; /* * Make sure we aren't marking a node as both an * ancestor and descendant. We should have caught * this as a lock order reversal earlier. */ if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { printf("witness rmatrix paradox! [%d][%d]=%d " "both ancestor and descendant\n", i, j, w_rmatrix[i][j]); kdb_backtrace(); printf("Witness disabled.\n"); witness_watch = -1; } if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { printf("witness rmatrix paradox! [%d][%d]=%d " "both ancestor and descendant\n", j, i, w_rmatrix[j][i]); kdb_backtrace(); printf("Witness disabled.\n"); witness_watch = -1; } } } } static void itismychild(struct witness *parent, struct witness *child) { int unlocked; MPASS(child != NULL && parent != NULL); if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); if (!witness_lock_type_equal(parent, child)) { if (witness_cold == 0) { unlocked = 1; mtx_unlock_spin(&w_mtx); } else { unlocked = 0; } kassert_panic( "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " "the same lock type", __func__, parent->w_name, parent->w_class->lc_name, child->w_name, child->w_class->lc_name); if (unlocked) mtx_lock_spin(&w_mtx); } adopt(parent, child); } /* * Generic code for the isitmy*() functions. The rmask parameter is the * expected relationship of w1 to w2. */ static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) { unsigned char r1, r2; int i1, i2; i1 = w1->w_index; i2 = w2->w_index; WITNESS_INDEX_ASSERT(i1); WITNESS_INDEX_ASSERT(i2); r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; /* The flags on one better be the inverse of the flags on the other */ if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { /* Don't squawk if we're potentially racing with an update. */ if (!mtx_owned(&w_mtx)) return (0); printf("%s: rmatrix mismatch between %s (index %d) and %s " "(index %d): w_rmatrix[%d][%d] == %hhx but " "w_rmatrix[%d][%d] == %hhx\n", fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, i2, i1, r2); kdb_backtrace(); printf("Witness disabled.\n"); witness_watch = -1; } return (r1 & rmask); } /* * Checks if @child is a direct child of @parent. */ static int isitmychild(struct witness *parent, struct witness *child) { return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); } /* * Checks if @descendant is a direct or inderect descendant of @ancestor. */ static int isitmydescendant(struct witness *ancestor, struct witness *descendant) { return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, __func__)); } #ifdef BLESSING static int blessed(struct witness *w1, struct witness *w2) { int i; struct witness_blessed *b; for (i = 0; i < nitems(blessed_list); i++) { b = &blessed_list[i]; if (strcmp(w1->w_name, b->b_lock1) == 0) { if (strcmp(w2->w_name, b->b_lock2) == 0) return (1); continue; } if (strcmp(w1->w_name, b->b_lock2) == 0) if (strcmp(w2->w_name, b->b_lock1) == 0) return (1); } return (0); } #endif static struct witness * witness_get(void) { struct witness *w; int index; if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); if (witness_watch == -1) { mtx_unlock_spin(&w_mtx); return (NULL); } if (STAILQ_EMPTY(&w_free)) { witness_watch = -1; mtx_unlock_spin(&w_mtx); printf("WITNESS: unable to allocate a new witness object\n"); return (NULL); } w = STAILQ_FIRST(&w_free); STAILQ_REMOVE_HEAD(&w_free, w_list); w_free_cnt--; index = w->w_index; MPASS(index > 0 && index == w_max_used_index+1 && index < witness_count); bzero(w, sizeof(*w)); w->w_index = index; if (index > w_max_used_index) w_max_used_index = index; return (w); } static void witness_free(struct witness *w) { STAILQ_INSERT_HEAD(&w_free, w, w_list); w_free_cnt++; } static struct lock_list_entry * witness_lock_list_get(void) { struct lock_list_entry *lle; if (witness_watch == -1) return (NULL); mtx_lock_spin(&w_mtx); lle = w_lock_list_free; if (lle == NULL) { witness_watch = -1; mtx_unlock_spin(&w_mtx); printf("%s: witness exhausted\n", __func__); return (NULL); } w_lock_list_free = lle->ll_next; mtx_unlock_spin(&w_mtx); bzero(lle, sizeof(*lle)); return (lle); } static void witness_lock_list_free(struct lock_list_entry *lle) { mtx_lock_spin(&w_mtx); lle->ll_next = w_lock_list_free; w_lock_list_free = lle; mtx_unlock_spin(&w_mtx); } static struct lock_instance * find_instance(struct lock_list_entry *list, const struct lock_object *lock) { struct lock_list_entry *lle; struct lock_instance *instance; int i; for (lle = list; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { instance = &lle->ll_children[i]; if (instance->li_lock == lock) return (instance); } return (NULL); } static void witness_list_lock(struct lock_instance *instance, int (*prnt)(const char *fmt, ...)) { struct lock_object *lock; lock = instance->li_lock; prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); if (lock->lo_witness->w_name != lock->lo_name) prnt(" (%s)", lock->lo_witness->w_name); prnt(" r = %d (%p) locked @ %s:%d\n", instance->li_flags & LI_RECURSEMASK, lock, fixup_filename(instance->li_file), instance->li_line); } static int witness_output(const char *fmt, ...) { va_list ap; int ret; va_start(ap, fmt); ret = witness_voutput(fmt, ap); va_end(ap); return (ret); } static int witness_voutput(const char *fmt, va_list ap) { int ret; ret = 0; switch (witness_channel) { case WITNESS_CONSOLE: ret = vprintf(fmt, ap); break; case WITNESS_LOG: vlog(LOG_NOTICE, fmt, ap); break; case WITNESS_NONE: break; } return (ret); } #ifdef DDB static int witness_thread_has_locks(struct thread *td) { if (td->td_sleeplocks == NULL) return (0); return (td->td_sleeplocks->ll_count != 0); } static int witness_proc_has_locks(struct proc *p) { struct thread *td; FOREACH_THREAD_IN_PROC(p, td) { if (witness_thread_has_locks(td)) return (1); } return (0); } #endif int witness_list_locks(struct lock_list_entry **lock_list, int (*prnt)(const char *fmt, ...)) { struct lock_list_entry *lle; int i, nheld; nheld = 0; for (lle = *lock_list; lle != NULL; lle = lle->ll_next) for (i = lle->ll_count - 1; i >= 0; i--) { witness_list_lock(&lle->ll_children[i], prnt); nheld++; } return (nheld); } /* * This is a bit risky at best. We call this function when we have timed * out acquiring a spin lock, and we assume that the other CPU is stuck * with this lock held. So, we go groveling around in the other CPU's * per-cpu data to try to find the lock instance for this spin lock to * see when it was last acquired. */ void witness_display_spinlock(struct lock_object *lock, struct thread *owner, int (*prnt)(const char *fmt, ...)) { struct lock_instance *instance; struct pcpu *pc; if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) return; pc = pcpu_find(owner->td_oncpu); instance = find_instance(pc->pc_spinlocks, lock); if (instance != NULL) witness_list_lock(instance, prnt); } void witness_save(struct lock_object *lock, const char **filep, int *linep) { struct lock_list_entry *lock_list; struct lock_instance *instance; struct lock_class *class; /* * This function is used independently in locking code to deal with * Giant, SCHEDULER_STOPPED() check can be removed here after Giant * is gone. */ if (SCHEDULER_STOPPED()) return; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (class->lc_flags & LC_SLEEPLOCK) lock_list = curthread->td_sleeplocks; else { if (witness_skipspin) return; lock_list = PCPU_GET(spinlocks); } instance = find_instance(lock_list, lock); if (instance == NULL) { kassert_panic("%s: lock (%s) %s not locked", __func__, class->lc_name, lock->lo_name); return; } *filep = instance->li_file; *linep = instance->li_line; } void witness_restore(struct lock_object *lock, const char *file, int line) { struct lock_list_entry *lock_list; struct lock_instance *instance; struct lock_class *class; /* * This function is used independently in locking code to deal with * Giant, SCHEDULER_STOPPED() check can be removed here after Giant * is gone. */ if (SCHEDULER_STOPPED()) return; KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (class->lc_flags & LC_SLEEPLOCK) lock_list = curthread->td_sleeplocks; else { if (witness_skipspin) return; lock_list = PCPU_GET(spinlocks); } instance = find_instance(lock_list, lock); if (instance == NULL) kassert_panic("%s: lock (%s) %s not locked", __func__, class->lc_name, lock->lo_name); lock->lo_witness->w_file = file; lock->lo_witness->w_line = line; if (instance == NULL) return; instance->li_file = file; instance->li_line = line; } void witness_assert(const struct lock_object *lock, int flags, const char *file, int line) { #ifdef INVARIANT_SUPPORT struct lock_instance *instance; struct lock_class *class; if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if ((class->lc_flags & LC_SLEEPLOCK) != 0) instance = find_instance(curthread->td_sleeplocks, lock); else if ((class->lc_flags & LC_SPINLOCK) != 0) instance = find_instance(PCPU_GET(spinlocks), lock); else { kassert_panic("Lock (%s) %s is not sleep or spin!", class->lc_name, lock->lo_name); return; } switch (flags) { case LA_UNLOCKED: if (instance != NULL) kassert_panic("Lock (%s) %s locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); break; case LA_LOCKED: case LA_LOCKED | LA_RECURSED: case LA_LOCKED | LA_NOTRECURSED: case LA_SLOCKED: case LA_SLOCKED | LA_RECURSED: case LA_SLOCKED | LA_NOTRECURSED: case LA_XLOCKED: case LA_XLOCKED | LA_RECURSED: case LA_XLOCKED | LA_NOTRECURSED: if (instance == NULL) { kassert_panic("Lock (%s) %s not locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); break; } if ((flags & LA_XLOCKED) != 0 && (instance->li_flags & LI_EXCLUSIVE) == 0) kassert_panic( "Lock (%s) %s not exclusively locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((flags & LA_SLOCKED) != 0 && (instance->li_flags & LI_EXCLUSIVE) != 0) kassert_panic( "Lock (%s) %s exclusively locked @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((flags & LA_RECURSED) != 0 && (instance->li_flags & LI_RECURSEMASK) == 0) kassert_panic("Lock (%s) %s not recursed @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); if ((flags & LA_NOTRECURSED) != 0 && (instance->li_flags & LI_RECURSEMASK) != 0) kassert_panic("Lock (%s) %s recursed @ %s:%d.", class->lc_name, lock->lo_name, fixup_filename(file), line); break; default: kassert_panic("Invalid lock assertion at %s:%d.", fixup_filename(file), line); } #endif /* INVARIANT_SUPPORT */ } static void witness_setflag(struct lock_object *lock, int flag, int set) { struct lock_list_entry *lock_list; struct lock_instance *instance; struct lock_class *class; if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) return; class = LOCK_CLASS(lock); if (class->lc_flags & LC_SLEEPLOCK) lock_list = curthread->td_sleeplocks; else { if (witness_skipspin) return; lock_list = PCPU_GET(spinlocks); } instance = find_instance(lock_list, lock); if (instance == NULL) { kassert_panic("%s: lock (%s) %s not locked", __func__, class->lc_name, lock->lo_name); return; } if (set) instance->li_flags |= flag; else instance->li_flags &= ~flag; } void witness_norelease(struct lock_object *lock) { witness_setflag(lock, LI_NORELEASE, 1); } void witness_releaseok(struct lock_object *lock) { witness_setflag(lock, LI_NORELEASE, 0); } #ifdef DDB static void witness_ddb_list(struct thread *td) { KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); KASSERT(kdb_active, ("%s: not in the debugger", __func__)); if (witness_watch < 1) return; witness_list_locks(&td->td_sleeplocks, db_printf); /* * We only handle spinlocks if td == curthread. This is somewhat broken * if td is currently executing on some other CPU and holds spin locks * as we won't display those locks. If we had a MI way of getting * the per-cpu data for a given cpu then we could use * td->td_oncpu to get the list of spinlocks for this thread * and "fix" this. * * That still wouldn't really fix this unless we locked the scheduler * lock or stopped the other CPU to make sure it wasn't changing the * list out from under us. It is probably best to just not try to * handle threads on other CPU's for now. */ if (td == curthread && PCPU_GET(spinlocks) != NULL) witness_list_locks(PCPU_PTR(spinlocks), db_printf); } DB_SHOW_COMMAND(locks, db_witness_list) { struct thread *td; if (have_addr) td = db_lookup_thread(addr, true); else td = kdb_thread; witness_ddb_list(td); } DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) { struct thread *td; struct proc *p; /* * It would be nice to list only threads and processes that actually * held sleep locks, but that information is currently not exported * by WITNESS. */ FOREACH_PROC_IN_SYSTEM(p) { if (!witness_proc_has_locks(p)) continue; FOREACH_THREAD_IN_PROC(p, td) { if (!witness_thread_has_locks(td)) continue; db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, p->p_comm, td, td->td_tid); witness_ddb_list(td); if (db_pager_quit) return; } } } DB_SHOW_ALIAS(alllocks, db_witness_list_all) DB_SHOW_COMMAND(witness, db_witness_display) { witness_ddb_display(db_printf); } #endif static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) { struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; struct witness *tmp_w1, *tmp_w2, *w1, *w2; struct sbuf *sb; u_int w_rmatrix1, w_rmatrix2; int error, generation, i, j; tmp_data1 = NULL; tmp_data2 = NULL; tmp_w1 = NULL; tmp_w2 = NULL; if (witness_watch < 1) { error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); return (error); } if (witness_cold) { error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); return (error); } error = 0; sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND); if (sb == NULL) return (ENOMEM); /* Allocate and init temporary storage space. */ tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, M_WAITOK | M_ZERO); tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, M_WAITOK | M_ZERO); stack_zero(&tmp_data1->wlod_stack); stack_zero(&tmp_data2->wlod_stack); restart: mtx_lock_spin(&w_mtx); generation = w_generation; mtx_unlock_spin(&w_mtx); sbuf_printf(sb, "Number of known direct relationships is %d\n", w_lohash.wloh_count); for (i = 1; i < w_max_used_index; i++) { mtx_lock_spin(&w_mtx); if (generation != w_generation) { mtx_unlock_spin(&w_mtx); /* The graph has changed, try again. */ req->oldidx = 0; sbuf_clear(sb); goto restart; } w1 = &w_data[i]; if (w1->w_reversed == 0) { mtx_unlock_spin(&w_mtx); continue; } /* Copy w1 locally so we can release the spin lock. */ *tmp_w1 = *w1; mtx_unlock_spin(&w_mtx); if (tmp_w1->w_reversed == 0) continue; for (j = 1; j < w_max_used_index; j++) { if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) continue; mtx_lock_spin(&w_mtx); if (generation != w_generation) { mtx_unlock_spin(&w_mtx); /* The graph has changed, try again. */ req->oldidx = 0; sbuf_clear(sb); goto restart; } w2 = &w_data[j]; data1 = witness_lock_order_get(w1, w2); data2 = witness_lock_order_get(w2, w1); /* * Copy information locally so we can release the * spin lock. */ *tmp_w2 = *w2; w_rmatrix1 = (unsigned int)w_rmatrix[i][j]; w_rmatrix2 = (unsigned int)w_rmatrix[j][i]; if (data1) { stack_zero(&tmp_data1->wlod_stack); stack_copy(&data1->wlod_stack, &tmp_data1->wlod_stack); } if (data2 && data2 != data1) { stack_zero(&tmp_data2->wlod_stack); stack_copy(&data2->wlod_stack, &tmp_data2->wlod_stack); } mtx_unlock_spin(&w_mtx); sbuf_printf(sb, "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", tmp_w1->w_name, tmp_w1->w_class->lc_name, tmp_w2->w_name, tmp_w2->w_class->lc_name); if (data1) { sbuf_printf(sb, "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", tmp_w1->w_name, tmp_w1->w_class->lc_name, tmp_w2->w_name, tmp_w2->w_class->lc_name); stack_sbuf_print(sb, &tmp_data1->wlod_stack); sbuf_printf(sb, "\n"); } if (data2 && data2 != data1) { sbuf_printf(sb, "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", tmp_w2->w_name, tmp_w2->w_class->lc_name, tmp_w1->w_name, tmp_w1->w_class->lc_name); stack_sbuf_print(sb, &tmp_data2->wlod_stack); sbuf_printf(sb, "\n"); } } } mtx_lock_spin(&w_mtx); if (generation != w_generation) { mtx_unlock_spin(&w_mtx); /* * The graph changed while we were printing stack data, * try again. */ req->oldidx = 0; sbuf_clear(sb); goto restart; } mtx_unlock_spin(&w_mtx); /* Free temporary storage space. */ free(tmp_data1, M_TEMP); free(tmp_data2, M_TEMP); free(tmp_w1, M_TEMP); free(tmp_w2, M_TEMP); sbuf_finish(sb); error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); sbuf_delete(sb); return (error); } static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS) { static const struct { enum witness_channel channel; const char *name; } channels[] = { { WITNESS_CONSOLE, "console" }, { WITNESS_LOG, "log" }, { WITNESS_NONE, "none" }, }; char buf[16]; u_int i; int error; buf[0] = '\0'; for (i = 0; i < nitems(channels); i++) if (witness_channel == channels[i].channel) { snprintf(buf, sizeof(buf), "%s", channels[i].name); break; } error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); error = EINVAL; for (i = 0; i < nitems(channels); i++) if (strcmp(channels[i].name, buf) == 0) { witness_channel = channels[i].channel; error = 0; break; } return (error); } static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) { struct witness *w; struct sbuf *sb; int error; if (witness_watch < 1) { error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); return (error); } if (witness_cold) { error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); return (error); } error = 0; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); if (sb == NULL) return (ENOMEM); sbuf_printf(sb, "\n"); mtx_lock_spin(&w_mtx); STAILQ_FOREACH(w, &w_all, w_list) w->w_displayed = 0; STAILQ_FOREACH(w, &w_all, w_list) witness_add_fullgraph(sb, w); mtx_unlock_spin(&w_mtx); /* * Close the sbuf and return to userland. */ error = sbuf_finish(sb); sbuf_delete(sb); return (error); } static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) { int error, value; value = witness_watch; error = sysctl_handle_int(oidp, &value, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (value > 1 || value < -1 || (witness_watch == -1 && value != witness_watch)) return (EINVAL); witness_watch = value; return (0); } static void witness_add_fullgraph(struct sbuf *sb, struct witness *w) { int i; if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) return; w->w_displayed = 1; WITNESS_INDEX_ASSERT(w->w_index); for (i = 1; i <= w_max_used_index; i++) { if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, w_data[i].w_name); witness_add_fullgraph(sb, &w_data[i]); } } } /* * A simple hash function. Takes a key pointer and a key size. If size == 0, * interprets the key as a string and reads until the null * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit * hash value computed from the key. */ static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size) { unsigned int hash = 5381; int i; /* hash = hash * 33 + key[i] */ if (size) for (i = 0; i < size; i++) hash = ((hash << 5) + hash) + (unsigned int)key[i]; else for (i = 0; key[i] != 0; i++) hash = ((hash << 5) + hash) + (unsigned int)key[i]; return (hash); } /* * Initializes the two witness hash tables. Called exactly once from * witness_initialize(). */ static void witness_init_hash_tables(void) { int i; MPASS(witness_cold); /* Initialize the hash tables. */ for (i = 0; i < WITNESS_HASH_SIZE; i++) w_hash.wh_array[i] = NULL; w_hash.wh_size = WITNESS_HASH_SIZE; w_hash.wh_count = 0; /* Initialize the lock order data hash. */ w_lofree = NULL; for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { memset(&w_lodata[i], 0, sizeof(w_lodata[i])); w_lodata[i].wlod_next = w_lofree; w_lofree = &w_lodata[i]; } w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; w_lohash.wloh_count = 0; for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) w_lohash.wloh_array[i] = NULL; } static struct witness * witness_hash_get(const char *key) { struct witness *w; uint32_t hash; MPASS(key != NULL); if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); hash = witness_hash_djb2(key, 0) % w_hash.wh_size; w = w_hash.wh_array[hash]; while (w != NULL) { if (strcmp(w->w_name, key) == 0) goto out; w = w->w_hash_next; } out: return (w); } static void witness_hash_put(struct witness *w) { uint32_t hash; MPASS(w != NULL); MPASS(w->w_name != NULL); if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); KASSERT(witness_hash_get(w->w_name) == NULL, ("%s: trying to add a hash entry that already exists!", __func__)); KASSERT(w->w_hash_next == NULL, ("%s: w->w_hash_next != NULL", __func__)); hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; w->w_hash_next = w_hash.wh_array[hash]; w_hash.wh_array[hash] = w; w_hash.wh_count++; } static struct witness_lock_order_data * witness_lock_order_get(struct witness *parent, struct witness *child) { struct witness_lock_order_data *data = NULL; struct witness_lock_order_key key; unsigned int hash; MPASS(parent != NULL && child != NULL); key.from = parent->w_index; key.to = child->w_index; WITNESS_INDEX_ASSERT(key.from); WITNESS_INDEX_ASSERT(key.to); if ((w_rmatrix[parent->w_index][child->w_index] & WITNESS_LOCK_ORDER_KNOWN) == 0) goto out; hash = witness_hash_djb2((const char*)&key, sizeof(key)) % w_lohash.wloh_size; data = w_lohash.wloh_array[hash]; while (data != NULL) { if (witness_lock_order_key_equal(&data->wlod_key, &key)) break; data = data->wlod_next; } out: return (data); } /* * Verify that parent and child have a known relationship, are not the same, * and child is actually a child of parent. This is done without w_mtx * to avoid contention in the common case. */ static int witness_lock_order_check(struct witness *parent, struct witness *child) { if (parent != child && w_rmatrix[parent->w_index][child->w_index] & WITNESS_LOCK_ORDER_KNOWN && isitmychild(parent, child)) return (1); return (0); } static int witness_lock_order_add(struct witness *parent, struct witness *child) { struct witness_lock_order_data *data = NULL; struct witness_lock_order_key key; unsigned int hash; MPASS(parent != NULL && child != NULL); key.from = parent->w_index; key.to = child->w_index; WITNESS_INDEX_ASSERT(key.from); WITNESS_INDEX_ASSERT(key.to); if (w_rmatrix[parent->w_index][child->w_index] & WITNESS_LOCK_ORDER_KNOWN) return (1); hash = witness_hash_djb2((const char*)&key, sizeof(key)) % w_lohash.wloh_size; w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; data = w_lofree; if (data == NULL) return (0); w_lofree = data->wlod_next; data->wlod_next = w_lohash.wloh_array[hash]; data->wlod_key = key; w_lohash.wloh_array[hash] = data; w_lohash.wloh_count++; stack_zero(&data->wlod_stack); stack_save(&data->wlod_stack); return (1); } /* Call this whenever the structure of the witness graph changes. */ static void witness_increment_graph_generation(void) { if (witness_cold == 0) mtx_assert(&w_mtx, MA_OWNED); w_generation++; } static int witness_output_drain(void *arg __unused, const char *data, int len) { witness_output("%.*s", len, data); return (len); } static void witness_debugger(int cond, const char *msg) { char buf[32]; struct sbuf sb; struct stack st; if (!cond) return; if (witness_trace) { sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN); sbuf_set_drain(&sb, witness_output_drain, NULL); stack_zero(&st); stack_save(&st); witness_output("stack backtrace:\n"); stack_sbuf_print_ddb(&sb, &st); sbuf_finish(&sb); } #ifdef KDB if (witness_kdb) kdb_enter(KDB_WHY_WITNESS, msg); #endif } Index: projects/mips64-clang/sys/kern/uipc_syscalls.c =================================================================== --- projects/mips64-clang/sys/kern/uipc_syscalls.c (revision 305711) +++ projects/mips64-clang/sys/kern/uipc_syscalls.c (revision 305712) @@ -1,1755 +1,1575 @@ /*- * Copyright (c) 1982, 1986, 1989, 1990, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)uipc_syscalls.c 8.4 (Berkeley) 2/21/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_capsicum.h" #include "opt_inet.h" #include "opt_inet6.h" #include "opt_compat.h" #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef KTRACE #include #endif #ifdef COMPAT_FREEBSD32 #include #endif #include #include #include /* * Flags for accept1() and kern_accept4(), in addition to SOCK_CLOEXEC * and SOCK_NONBLOCK. */ #define ACCEPT4_INHERIT 0x1 #define ACCEPT4_COMPAT 0x2 static int sendit(struct thread *td, int s, struct msghdr *mp, int flags); static int recvit(struct thread *td, int s, struct msghdr *mp, void *namelenp); static int accept1(struct thread *td, int s, struct sockaddr *uname, socklen_t *anamelen, int flags); static int getsockname1(struct thread *td, struct getsockname_args *uap, int compat); static int getpeername1(struct thread *td, struct getpeername_args *uap, int compat); static int sockargs(struct mbuf **, char *, socklen_t, int); /* * Convert a user file descriptor to a kernel file entry and check if required * capability rights are present. * A reference on the file entry is held upon returning. */ int getsock_cap(struct thread *td, int fd, cap_rights_t *rightsp, struct file **fpp, u_int *fflagp) { struct file *fp; int error; error = fget_unlocked(td->td_proc->p_fd, fd, rightsp, &fp, NULL); if (error != 0) return (error); if (fp->f_type != DTYPE_SOCKET) { fdrop(fp, td); return (ENOTSOCK); } if (fflagp != NULL) *fflagp = fp->f_flag; *fpp = fp; return (0); } /* * System call interface to the socket abstraction. */ #if defined(COMPAT_43) #define COMPAT_OLDSOCK #endif int -sys_socket(td, uap) - struct thread *td; - struct socket_args /* { - int domain; - int type; - int protocol; - } */ *uap; +sys_socket(struct thread *td, struct socket_args *uap) { struct socket *so; struct file *fp; int fd, error, type, oflag, fflag; AUDIT_ARG_SOCKET(uap->domain, uap->type, uap->protocol); type = uap->type; oflag = 0; fflag = 0; if ((type & SOCK_CLOEXEC) != 0) { type &= ~SOCK_CLOEXEC; oflag |= O_CLOEXEC; } if ((type & SOCK_NONBLOCK) != 0) { type &= ~SOCK_NONBLOCK; fflag |= FNONBLOCK; } #ifdef MAC error = mac_socket_check_create(td->td_ucred, uap->domain, type, uap->protocol); if (error != 0) return (error); #endif error = falloc(td, &fp, &fd, oflag); if (error != 0) return (error); /* An extra reference on `fp' has been held for us by falloc(). */ error = socreate(uap->domain, &so, type, uap->protocol, td->td_ucred, td); if (error != 0) { fdclose(td, fp, fd); } else { finit(fp, FREAD | FWRITE | fflag, DTYPE_SOCKET, so, &socketops); if ((fflag & FNONBLOCK) != 0) (void) fo_ioctl(fp, FIONBIO, &fflag, td->td_ucred, td); td->td_retval[0] = fd; } fdrop(fp, td); return (error); } -/* ARGSUSED */ int -sys_bind(td, uap) - struct thread *td; - struct bind_args /* { - int s; - caddr_t name; - int namelen; - } */ *uap; +sys_bind(struct thread *td, struct bind_args *uap) { struct sockaddr *sa; int error; error = getsockaddr(&sa, uap->name, uap->namelen); if (error == 0) { error = kern_bindat(td, AT_FDCWD, uap->s, sa); free(sa, M_SONAME); } return (error); } int kern_bindat(struct thread *td, int dirfd, int fd, struct sockaddr *sa) { struct socket *so; struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(fd); AUDIT_ARG_SOCKADDR(td, dirfd, sa); error = getsock_cap(td, fd, cap_rights_init(&rights, CAP_BIND), &fp, NULL); if (error != 0) return (error); so = fp->f_data; #ifdef KTRACE if (KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(sa); #endif #ifdef MAC error = mac_socket_check_bind(td->td_ucred, so, sa); if (error == 0) { #endif if (dirfd == AT_FDCWD) error = sobind(so, sa, td); else error = sobindat(dirfd, so, sa, td); #ifdef MAC } #endif fdrop(fp, td); return (error); } -/* ARGSUSED */ int -sys_bindat(td, uap) - struct thread *td; - struct bindat_args /* { - int fd; - int s; - caddr_t name; - int namelen; - } */ *uap; +sys_bindat(struct thread *td, struct bindat_args *uap) { struct sockaddr *sa; int error; error = getsockaddr(&sa, uap->name, uap->namelen); if (error == 0) { error = kern_bindat(td, uap->fd, uap->s, sa); free(sa, M_SONAME); } return (error); } -/* ARGSUSED */ int -sys_listen(td, uap) - struct thread *td; - struct listen_args /* { - int s; - int backlog; - } */ *uap; +sys_listen(struct thread *td, struct listen_args *uap) { struct socket *so; struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(uap->s); error = getsock_cap(td, uap->s, cap_rights_init(&rights, CAP_LISTEN), &fp, NULL); if (error == 0) { so = fp->f_data; #ifdef MAC error = mac_socket_check_listen(td->td_ucred, so); if (error == 0) #endif error = solisten(so, uap->backlog, td); fdrop(fp, td); } return(error); } /* * accept1() */ static int accept1(td, s, uname, anamelen, flags) struct thread *td; int s; struct sockaddr *uname; socklen_t *anamelen; int flags; { struct sockaddr *name; socklen_t namelen; struct file *fp; int error; if (uname == NULL) return (kern_accept4(td, s, NULL, NULL, flags, NULL)); error = copyin(anamelen, &namelen, sizeof (namelen)); if (error != 0) return (error); error = kern_accept4(td, s, &name, &namelen, flags, &fp); if (error != 0) return (error); if (error == 0 && uname != NULL) { #ifdef COMPAT_OLDSOCK if (flags & ACCEPT4_COMPAT) ((struct osockaddr *)name)->sa_family = name->sa_family; #endif error = copyout(name, uname, namelen); } if (error == 0) error = copyout(&namelen, anamelen, sizeof(namelen)); if (error != 0) fdclose(td, fp, td->td_retval[0]); fdrop(fp, td); free(name, M_SONAME); return (error); } int kern_accept(struct thread *td, int s, struct sockaddr **name, socklen_t *namelen, struct file **fp) { return (kern_accept4(td, s, name, namelen, ACCEPT4_INHERIT, fp)); } int kern_accept4(struct thread *td, int s, struct sockaddr **name, socklen_t *namelen, int flags, struct file **fp) { struct file *headfp, *nfp = NULL; struct sockaddr *sa = NULL; struct socket *head, *so; cap_rights_t rights; u_int fflag; pid_t pgid; int error, fd, tmp; if (name != NULL) *name = NULL; AUDIT_ARG_FD(s); error = getsock_cap(td, s, cap_rights_init(&rights, CAP_ACCEPT), &headfp, &fflag); if (error != 0) return (error); head = headfp->f_data; if ((head->so_options & SO_ACCEPTCONN) == 0) { error = EINVAL; goto done; } #ifdef MAC error = mac_socket_check_accept(td->td_ucred, head); if (error != 0) goto done; #endif error = falloc(td, &nfp, &fd, (flags & SOCK_CLOEXEC) ? O_CLOEXEC : 0); if (error != 0) goto done; ACCEPT_LOCK(); if ((head->so_state & SS_NBIO) && TAILQ_EMPTY(&head->so_comp)) { ACCEPT_UNLOCK(); error = EWOULDBLOCK; goto noconnection; } while (TAILQ_EMPTY(&head->so_comp) && head->so_error == 0) { if (head->so_rcv.sb_state & SBS_CANTRCVMORE) { head->so_error = ECONNABORTED; break; } error = msleep(&head->so_timeo, &accept_mtx, PSOCK | PCATCH, "accept", 0); if (error != 0) { ACCEPT_UNLOCK(); goto noconnection; } } if (head->so_error) { error = head->so_error; head->so_error = 0; ACCEPT_UNLOCK(); goto noconnection; } so = TAILQ_FIRST(&head->so_comp); KASSERT(!(so->so_qstate & SQ_INCOMP), ("accept1: so SQ_INCOMP")); KASSERT(so->so_qstate & SQ_COMP, ("accept1: so not SQ_COMP")); /* * Before changing the flags on the socket, we have to bump the * reference count. Otherwise, if the protocol calls sofree(), * the socket will be released due to a zero refcount. */ SOCK_LOCK(so); /* soref() and so_state update */ soref(so); /* file descriptor reference */ TAILQ_REMOVE(&head->so_comp, so, so_list); head->so_qlen--; if (flags & ACCEPT4_INHERIT) so->so_state |= (head->so_state & SS_NBIO); else so->so_state |= (flags & SOCK_NONBLOCK) ? SS_NBIO : 0; so->so_qstate &= ~SQ_COMP; so->so_head = NULL; SOCK_UNLOCK(so); ACCEPT_UNLOCK(); /* An extra reference on `nfp' has been held for us by falloc(). */ td->td_retval[0] = fd; /* connection has been removed from the listen queue */ KNOTE_UNLOCKED(&head->so_rcv.sb_sel.si_note, 0); if (flags & ACCEPT4_INHERIT) { pgid = fgetown(&head->so_sigio); if (pgid != 0) fsetown(pgid, &so->so_sigio); } else { fflag &= ~(FNONBLOCK | FASYNC); if (flags & SOCK_NONBLOCK) fflag |= FNONBLOCK; } finit(nfp, fflag, DTYPE_SOCKET, so, &socketops); /* Sync socket nonblocking/async state with file flags */ tmp = fflag & FNONBLOCK; (void) fo_ioctl(nfp, FIONBIO, &tmp, td->td_ucred, td); tmp = fflag & FASYNC; (void) fo_ioctl(nfp, FIOASYNC, &tmp, td->td_ucred, td); sa = NULL; error = soaccept(so, &sa); if (error != 0) goto noconnection; if (sa == NULL) { if (name) *namelen = 0; goto done; } AUDIT_ARG_SOCKADDR(td, AT_FDCWD, sa); if (name) { /* check sa_len before it is destroyed */ if (*namelen > sa->sa_len) *namelen = sa->sa_len; #ifdef KTRACE if (KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(sa); #endif *name = sa; sa = NULL; } noconnection: free(sa, M_SONAME); /* * close the new descriptor, assuming someone hasn't ripped it * out from under us. */ if (error != 0) fdclose(td, nfp, fd); /* * Release explicitly held references before returning. We return * a reference on nfp to the caller on success if they request it. */ done: if (fp != NULL) { if (error == 0) { *fp = nfp; nfp = NULL; } else *fp = NULL; } if (nfp != NULL) fdrop(nfp, td); fdrop(headfp, td); return (error); } int sys_accept(td, uap) struct thread *td; struct accept_args *uap; { return (accept1(td, uap->s, uap->name, uap->anamelen, ACCEPT4_INHERIT)); } int sys_accept4(td, uap) struct thread *td; struct accept4_args *uap; { if (uap->flags & ~(SOCK_CLOEXEC | SOCK_NONBLOCK)) return (EINVAL); return (accept1(td, uap->s, uap->name, uap->anamelen, uap->flags)); } #ifdef COMPAT_OLDSOCK int oaccept(td, uap) struct thread *td; struct accept_args *uap; { return (accept1(td, uap->s, uap->name, uap->anamelen, ACCEPT4_INHERIT | ACCEPT4_COMPAT)); } #endif /* COMPAT_OLDSOCK */ -/* ARGSUSED */ int -sys_connect(td, uap) - struct thread *td; - struct connect_args /* { - int s; - caddr_t name; - int namelen; - } */ *uap; +sys_connect(struct thread *td, struct connect_args *uap) { struct sockaddr *sa; int error; error = getsockaddr(&sa, uap->name, uap->namelen); if (error == 0) { error = kern_connectat(td, AT_FDCWD, uap->s, sa); free(sa, M_SONAME); } return (error); } int kern_connectat(struct thread *td, int dirfd, int fd, struct sockaddr *sa) { struct socket *so; struct file *fp; cap_rights_t rights; int error, interrupted = 0; AUDIT_ARG_FD(fd); AUDIT_ARG_SOCKADDR(td, dirfd, sa); error = getsock_cap(td, fd, cap_rights_init(&rights, CAP_CONNECT), &fp, NULL); if (error != 0) return (error); so = fp->f_data; if (so->so_state & SS_ISCONNECTING) { error = EALREADY; goto done1; } #ifdef KTRACE if (KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(sa); #endif #ifdef MAC error = mac_socket_check_connect(td->td_ucred, so, sa); if (error != 0) goto bad; #endif if (dirfd == AT_FDCWD) error = soconnect(so, sa, td); else error = soconnectat(dirfd, so, sa, td); if (error != 0) goto bad; if ((so->so_state & SS_NBIO) && (so->so_state & SS_ISCONNECTING)) { error = EINPROGRESS; goto done1; } SOCK_LOCK(so); while ((so->so_state & SS_ISCONNECTING) && so->so_error == 0) { error = msleep(&so->so_timeo, SOCK_MTX(so), PSOCK | PCATCH, "connec", 0); if (error != 0) { if (error == EINTR || error == ERESTART) interrupted = 1; break; } } if (error == 0) { error = so->so_error; so->so_error = 0; } SOCK_UNLOCK(so); bad: if (!interrupted) so->so_state &= ~SS_ISCONNECTING; if (error == ERESTART) error = EINTR; done1: fdrop(fp, td); return (error); } -/* ARGSUSED */ int -sys_connectat(td, uap) - struct thread *td; - struct connectat_args /* { - int fd; - int s; - caddr_t name; - int namelen; - } */ *uap; +sys_connectat(struct thread *td, struct connectat_args *uap) { struct sockaddr *sa; int error; error = getsockaddr(&sa, uap->name, uap->namelen); if (error == 0) { error = kern_connectat(td, uap->fd, uap->s, sa); free(sa, M_SONAME); } return (error); } int kern_socketpair(struct thread *td, int domain, int type, int protocol, int *rsv) { struct file *fp1, *fp2; struct socket *so1, *so2; int fd, error, oflag, fflag; AUDIT_ARG_SOCKET(domain, type, protocol); oflag = 0; fflag = 0; if ((type & SOCK_CLOEXEC) != 0) { type &= ~SOCK_CLOEXEC; oflag |= O_CLOEXEC; } if ((type & SOCK_NONBLOCK) != 0) { type &= ~SOCK_NONBLOCK; fflag |= FNONBLOCK; } #ifdef MAC /* We might want to have a separate check for socket pairs. */ error = mac_socket_check_create(td->td_ucred, domain, type, protocol); if (error != 0) return (error); #endif error = socreate(domain, &so1, type, protocol, td->td_ucred, td); if (error != 0) return (error); error = socreate(domain, &so2, type, protocol, td->td_ucred, td); if (error != 0) goto free1; /* On success extra reference to `fp1' and 'fp2' is set by falloc. */ error = falloc(td, &fp1, &fd, oflag); if (error != 0) goto free2; rsv[0] = fd; fp1->f_data = so1; /* so1 already has ref count */ error = falloc(td, &fp2, &fd, oflag); if (error != 0) goto free3; fp2->f_data = so2; /* so2 already has ref count */ rsv[1] = fd; error = soconnect2(so1, so2); if (error != 0) goto free4; if (type == SOCK_DGRAM) { /* * Datagram socket connection is asymmetric. */ error = soconnect2(so2, so1); if (error != 0) goto free4; } finit(fp1, FREAD | FWRITE | fflag, DTYPE_SOCKET, fp1->f_data, &socketops); finit(fp2, FREAD | FWRITE | fflag, DTYPE_SOCKET, fp2->f_data, &socketops); if ((fflag & FNONBLOCK) != 0) { (void) fo_ioctl(fp1, FIONBIO, &fflag, td->td_ucred, td); (void) fo_ioctl(fp2, FIONBIO, &fflag, td->td_ucred, td); } fdrop(fp1, td); fdrop(fp2, td); return (0); free4: fdclose(td, fp2, rsv[1]); fdrop(fp2, td); free3: fdclose(td, fp1, rsv[0]); fdrop(fp1, td); free2: if (so2 != NULL) (void)soclose(so2); free1: if (so1 != NULL) (void)soclose(so1); return (error); } int sys_socketpair(struct thread *td, struct socketpair_args *uap) { int error, sv[2]; error = kern_socketpair(td, uap->domain, uap->type, uap->protocol, sv); if (error != 0) return (error); error = copyout(sv, uap->rsv, 2 * sizeof(int)); if (error != 0) { (void)kern_close(td, sv[0]); (void)kern_close(td, sv[1]); } return (error); } static int -sendit(td, s, mp, flags) - struct thread *td; - int s; - struct msghdr *mp; - int flags; +sendit(struct thread *td, int s, struct msghdr *mp, int flags) { struct mbuf *control; struct sockaddr *to; int error; #ifdef CAPABILITY_MODE if (IN_CAPABILITY_MODE(td) && (mp->msg_name != NULL)) return (ECAPMODE); #endif if (mp->msg_name != NULL) { error = getsockaddr(&to, mp->msg_name, mp->msg_namelen); if (error != 0) { to = NULL; goto bad; } mp->msg_name = to; } else { to = NULL; } if (mp->msg_control) { if (mp->msg_controllen < sizeof(struct cmsghdr) #ifdef COMPAT_OLDSOCK && mp->msg_flags != MSG_COMPAT #endif ) { error = EINVAL; goto bad; } error = sockargs(&control, mp->msg_control, mp->msg_controllen, MT_CONTROL); if (error != 0) goto bad; #ifdef COMPAT_OLDSOCK if (mp->msg_flags == MSG_COMPAT) { struct cmsghdr *cm; M_PREPEND(control, sizeof(*cm), M_WAITOK); cm = mtod(control, struct cmsghdr *); cm->cmsg_len = control->m_len; cm->cmsg_level = SOL_SOCKET; cm->cmsg_type = SCM_RIGHTS; } #endif } else { control = NULL; } error = kern_sendit(td, s, mp, flags, control, UIO_USERSPACE); bad: free(to, M_SONAME); return (error); } int -kern_sendit(td, s, mp, flags, control, segflg) - struct thread *td; - int s; - struct msghdr *mp; - int flags; - struct mbuf *control; - enum uio_seg segflg; +kern_sendit(struct thread *td, int s, struct msghdr *mp, int flags, + struct mbuf *control, enum uio_seg segflg) { struct file *fp; struct uio auio; struct iovec *iov; struct socket *so; cap_rights_t rights; #ifdef KTRACE struct uio *ktruio = NULL; #endif ssize_t len; int i, error; AUDIT_ARG_FD(s); cap_rights_init(&rights, CAP_SEND); if (mp->msg_name != NULL) { AUDIT_ARG_SOCKADDR(td, AT_FDCWD, mp->msg_name); cap_rights_set(&rights, CAP_CONNECT); } error = getsock_cap(td, s, &rights, &fp, NULL); if (error != 0) return (error); so = (struct socket *)fp->f_data; #ifdef KTRACE if (mp->msg_name != NULL && KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(mp->msg_name); #endif #ifdef MAC if (mp->msg_name != NULL) { error = mac_socket_check_connect(td->td_ucred, so, mp->msg_name); if (error != 0) goto bad; } error = mac_socket_check_send(td->td_ucred, so); if (error != 0) goto bad; #endif auio.uio_iov = mp->msg_iov; auio.uio_iovcnt = mp->msg_iovlen; auio.uio_segflg = segflg; auio.uio_rw = UIO_WRITE; auio.uio_td = td; auio.uio_offset = 0; /* XXX */ auio.uio_resid = 0; iov = mp->msg_iov; for (i = 0; i < mp->msg_iovlen; i++, iov++) { if ((auio.uio_resid += iov->iov_len) < 0) { error = EINVAL; goto bad; } } #ifdef KTRACE if (KTRPOINT(td, KTR_GENIO)) ktruio = cloneuio(&auio); #endif len = auio.uio_resid; error = sosend(so, mp->msg_name, &auio, 0, control, flags, td); if (error != 0) { if (auio.uio_resid != len && (error == ERESTART || error == EINTR || error == EWOULDBLOCK)) error = 0; /* Generation of SIGPIPE can be controlled per socket */ if (error == EPIPE && !(so->so_options & SO_NOSIGPIPE) && !(flags & MSG_NOSIGNAL)) { PROC_LOCK(td->td_proc); tdsignal(td, SIGPIPE); PROC_UNLOCK(td->td_proc); } } if (error == 0) td->td_retval[0] = len - auio.uio_resid; #ifdef KTRACE if (ktruio != NULL) { ktruio->uio_resid = td->td_retval[0]; ktrgenio(s, UIO_WRITE, ktruio, error); } #endif bad: fdrop(fp, td); return (error); } int -sys_sendto(td, uap) - struct thread *td; - struct sendto_args /* { - int s; - caddr_t buf; - size_t len; - int flags; - caddr_t to; - int tolen; - } */ *uap; +sys_sendto(struct thread *td, struct sendto_args *uap) { struct msghdr msg; struct iovec aiov; msg.msg_name = uap->to; msg.msg_namelen = uap->tolen; msg.msg_iov = &aiov; msg.msg_iovlen = 1; msg.msg_control = 0; #ifdef COMPAT_OLDSOCK msg.msg_flags = 0; #endif aiov.iov_base = uap->buf; aiov.iov_len = uap->len; return (sendit(td, uap->s, &msg, uap->flags)); } #ifdef COMPAT_OLDSOCK int -osend(td, uap) - struct thread *td; - struct osend_args /* { - int s; - caddr_t buf; - int len; - int flags; - } */ *uap; +osend(struct thread *td, struct osend_args *uap) { struct msghdr msg; struct iovec aiov; msg.msg_name = 0; msg.msg_namelen = 0; msg.msg_iov = &aiov; msg.msg_iovlen = 1; aiov.iov_base = uap->buf; aiov.iov_len = uap->len; msg.msg_control = 0; msg.msg_flags = 0; return (sendit(td, uap->s, &msg, uap->flags)); } int -osendmsg(td, uap) - struct thread *td; - struct osendmsg_args /* { - int s; - caddr_t msg; - int flags; - } */ *uap; +osendmsg(struct thread *td, struct osendmsg_args *uap) { struct msghdr msg; struct iovec *iov; int error; error = copyin(uap->msg, &msg, sizeof (struct omsghdr)); if (error != 0) return (error); error = copyiniov(msg.msg_iov, msg.msg_iovlen, &iov, EMSGSIZE); if (error != 0) return (error); msg.msg_iov = iov; msg.msg_flags = MSG_COMPAT; error = sendit(td, uap->s, &msg, uap->flags); free(iov, M_IOV); return (error); } #endif int -sys_sendmsg(td, uap) - struct thread *td; - struct sendmsg_args /* { - int s; - caddr_t msg; - int flags; - } */ *uap; +sys_sendmsg(struct thread *td, struct sendmsg_args *uap) { struct msghdr msg; struct iovec *iov; int error; error = copyin(uap->msg, &msg, sizeof (msg)); if (error != 0) return (error); error = copyiniov(msg.msg_iov, msg.msg_iovlen, &iov, EMSGSIZE); if (error != 0) return (error); msg.msg_iov = iov; #ifdef COMPAT_OLDSOCK msg.msg_flags = 0; #endif error = sendit(td, uap->s, &msg, uap->flags); free(iov, M_IOV); return (error); } int -kern_recvit(td, s, mp, fromseg, controlp) - struct thread *td; - int s; - struct msghdr *mp; - enum uio_seg fromseg; - struct mbuf **controlp; +kern_recvit(struct thread *td, int s, struct msghdr *mp, enum uio_seg fromseg, + struct mbuf **controlp) { struct uio auio; struct iovec *iov; struct mbuf *m, *control = NULL; caddr_t ctlbuf; struct file *fp; struct socket *so; struct sockaddr *fromsa = NULL; cap_rights_t rights; #ifdef KTRACE struct uio *ktruio = NULL; #endif ssize_t len; int error, i; if (controlp != NULL) *controlp = NULL; AUDIT_ARG_FD(s); error = getsock_cap(td, s, cap_rights_init(&rights, CAP_RECV), &fp, NULL); if (error != 0) return (error); so = fp->f_data; #ifdef MAC error = mac_socket_check_receive(td->td_ucred, so); if (error != 0) { fdrop(fp, td); return (error); } #endif auio.uio_iov = mp->msg_iov; auio.uio_iovcnt = mp->msg_iovlen; auio.uio_segflg = UIO_USERSPACE; auio.uio_rw = UIO_READ; auio.uio_td = td; auio.uio_offset = 0; /* XXX */ auio.uio_resid = 0; iov = mp->msg_iov; for (i = 0; i < mp->msg_iovlen; i++, iov++) { if ((auio.uio_resid += iov->iov_len) < 0) { fdrop(fp, td); return (EINVAL); } } #ifdef KTRACE if (KTRPOINT(td, KTR_GENIO)) ktruio = cloneuio(&auio); #endif len = auio.uio_resid; error = soreceive(so, &fromsa, &auio, NULL, (mp->msg_control || controlp) ? &control : NULL, &mp->msg_flags); if (error != 0) { if (auio.uio_resid != len && (error == ERESTART || error == EINTR || error == EWOULDBLOCK)) error = 0; } if (fromsa != NULL) AUDIT_ARG_SOCKADDR(td, AT_FDCWD, fromsa); #ifdef KTRACE if (ktruio != NULL) { ktruio->uio_resid = len - auio.uio_resid; ktrgenio(s, UIO_READ, ktruio, error); } #endif if (error != 0) goto out; td->td_retval[0] = len - auio.uio_resid; if (mp->msg_name) { len = mp->msg_namelen; if (len <= 0 || fromsa == NULL) len = 0; else { /* save sa_len before it is destroyed by MSG_COMPAT */ len = MIN(len, fromsa->sa_len); #ifdef COMPAT_OLDSOCK if (mp->msg_flags & MSG_COMPAT) ((struct osockaddr *)fromsa)->sa_family = fromsa->sa_family; #endif if (fromseg == UIO_USERSPACE) { error = copyout(fromsa, mp->msg_name, (unsigned)len); if (error != 0) goto out; } else bcopy(fromsa, mp->msg_name, len); } mp->msg_namelen = len; } if (mp->msg_control && controlp == NULL) { #ifdef COMPAT_OLDSOCK /* * We assume that old recvmsg calls won't receive access * rights and other control info, esp. as control info * is always optional and those options didn't exist in 4.3. * If we receive rights, trim the cmsghdr; anything else * is tossed. */ if (control && mp->msg_flags & MSG_COMPAT) { if (mtod(control, struct cmsghdr *)->cmsg_level != SOL_SOCKET || mtod(control, struct cmsghdr *)->cmsg_type != SCM_RIGHTS) { mp->msg_controllen = 0; goto out; } control->m_len -= sizeof (struct cmsghdr); control->m_data += sizeof (struct cmsghdr); } #endif len = mp->msg_controllen; m = control; mp->msg_controllen = 0; ctlbuf = mp->msg_control; while (m && len > 0) { unsigned int tocopy; if (len >= m->m_len) tocopy = m->m_len; else { mp->msg_flags |= MSG_CTRUNC; tocopy = len; } if ((error = copyout(mtod(m, caddr_t), ctlbuf, tocopy)) != 0) goto out; ctlbuf += tocopy; len -= tocopy; m = m->m_next; } mp->msg_controllen = ctlbuf - (caddr_t)mp->msg_control; } out: fdrop(fp, td); #ifdef KTRACE if (fromsa && KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(fromsa); #endif free(fromsa, M_SONAME); if (error == 0 && controlp != NULL) *controlp = control; else if (control) m_freem(control); return (error); } static int -recvit(td, s, mp, namelenp) - struct thread *td; - int s; - struct msghdr *mp; - void *namelenp; +recvit(struct thread *td, int s, struct msghdr *mp, void *namelenp) { int error; error = kern_recvit(td, s, mp, UIO_USERSPACE, NULL); if (error != 0) return (error); if (namelenp != NULL) { error = copyout(&mp->msg_namelen, namelenp, sizeof (socklen_t)); #ifdef COMPAT_OLDSOCK if (mp->msg_flags & MSG_COMPAT) error = 0; /* old recvfrom didn't check */ #endif } return (error); } int -sys_recvfrom(td, uap) - struct thread *td; - struct recvfrom_args /* { - int s; - caddr_t buf; - size_t len; - int flags; - struct sockaddr * __restrict from; - socklen_t * __restrict fromlenaddr; - } */ *uap; +sys_recvfrom(struct thread *td, struct recvfrom_args *uap) { struct msghdr msg; struct iovec aiov; int error; if (uap->fromlenaddr) { error = copyin(uap->fromlenaddr, &msg.msg_namelen, sizeof (msg.msg_namelen)); if (error != 0) goto done2; } else { msg.msg_namelen = 0; } msg.msg_name = uap->from; msg.msg_iov = &aiov; msg.msg_iovlen = 1; aiov.iov_base = uap->buf; aiov.iov_len = uap->len; msg.msg_control = 0; msg.msg_flags = uap->flags; error = recvit(td, uap->s, &msg, uap->fromlenaddr); done2: return (error); } #ifdef COMPAT_OLDSOCK int -orecvfrom(td, uap) - struct thread *td; - struct recvfrom_args *uap; +orecvfrom(struct thread *td, struct recvfrom_args *uap) { uap->flags |= MSG_COMPAT; return (sys_recvfrom(td, uap)); } #endif #ifdef COMPAT_OLDSOCK int -orecv(td, uap) - struct thread *td; - struct orecv_args /* { - int s; - caddr_t buf; - int len; - int flags; - } */ *uap; +orecv(struct thread *td, struct orecv_args *uap) { struct msghdr msg; struct iovec aiov; msg.msg_name = 0; msg.msg_namelen = 0; msg.msg_iov = &aiov; msg.msg_iovlen = 1; aiov.iov_base = uap->buf; aiov.iov_len = uap->len; msg.msg_control = 0; msg.msg_flags = uap->flags; return (recvit(td, uap->s, &msg, NULL)); } /* * Old recvmsg. This code takes advantage of the fact that the old msghdr * overlays the new one, missing only the flags, and with the (old) access * rights where the control fields are now. */ int -orecvmsg(td, uap) - struct thread *td; - struct orecvmsg_args /* { - int s; - struct omsghdr *msg; - int flags; - } */ *uap; +orecvmsg(struct thread *td, struct orecvmsg_args *uap) { struct msghdr msg; struct iovec *iov; int error; error = copyin(uap->msg, &msg, sizeof (struct omsghdr)); if (error != 0) return (error); error = copyiniov(msg.msg_iov, msg.msg_iovlen, &iov, EMSGSIZE); if (error != 0) return (error); msg.msg_flags = uap->flags | MSG_COMPAT; msg.msg_iov = iov; error = recvit(td, uap->s, &msg, &uap->msg->msg_namelen); if (msg.msg_controllen && error == 0) error = copyout(&msg.msg_controllen, &uap->msg->msg_accrightslen, sizeof (int)); free(iov, M_IOV); return (error); } #endif int -sys_recvmsg(td, uap) - struct thread *td; - struct recvmsg_args /* { - int s; - struct msghdr *msg; - int flags; - } */ *uap; +sys_recvmsg(struct thread *td, struct recvmsg_args *uap) { struct msghdr msg; struct iovec *uiov, *iov; int error; error = copyin(uap->msg, &msg, sizeof (msg)); if (error != 0) return (error); error = copyiniov(msg.msg_iov, msg.msg_iovlen, &iov, EMSGSIZE); if (error != 0) return (error); msg.msg_flags = uap->flags; #ifdef COMPAT_OLDSOCK msg.msg_flags &= ~MSG_COMPAT; #endif uiov = msg.msg_iov; msg.msg_iov = iov; error = recvit(td, uap->s, &msg, NULL); if (error == 0) { msg.msg_iov = uiov; error = copyout(&msg, uap->msg, sizeof(msg)); } free(iov, M_IOV); return (error); } -/* ARGSUSED */ int -sys_shutdown(td, uap) - struct thread *td; - struct shutdown_args /* { - int s; - int how; - } */ *uap; +sys_shutdown(struct thread *td, struct shutdown_args *uap) { struct socket *so; struct file *fp; cap_rights_t rights; int error; AUDIT_ARG_FD(uap->s); error = getsock_cap(td, uap->s, cap_rights_init(&rights, CAP_SHUTDOWN), &fp, NULL); if (error == 0) { so = fp->f_data; error = soshutdown(so, uap->how); /* * Previous versions did not return ENOTCONN, but 0 in * case the socket was not connected. Some important * programs like syslogd up to r279016, 2015-02-19, * still depend on this behavior. */ if (error == ENOTCONN && td->td_proc->p_osrel < P_OSREL_SHUTDOWN_ENOTCONN) error = 0; fdrop(fp, td); } return (error); } -/* ARGSUSED */ int -sys_setsockopt(td, uap) - struct thread *td; - struct setsockopt_args /* { - int s; - int level; - int name; - caddr_t val; - int valsize; - } */ *uap; +sys_setsockopt(struct thread *td, struct setsockopt_args *uap) { return (kern_setsockopt(td, uap->s, uap->level, uap->name, uap->val, UIO_USERSPACE, uap->valsize)); } int -kern_setsockopt(td, s, level, name, val, valseg, valsize) - struct thread *td; - int s; - int level; - int name; - void *val; - enum uio_seg valseg; - socklen_t valsize; +kern_setsockopt(struct thread *td, int s, int level, int name, void *val, + enum uio_seg valseg, socklen_t valsize) { struct socket *so; struct file *fp; struct sockopt sopt; cap_rights_t rights; int error; if (val == NULL && valsize != 0) return (EFAULT); if ((int)valsize < 0) return (EINVAL); sopt.sopt_dir = SOPT_SET; sopt.sopt_level = level; sopt.sopt_name = name; sopt.sopt_val = val; sopt.sopt_valsize = valsize; switch (valseg) { case UIO_USERSPACE: sopt.sopt_td = td; break; case UIO_SYSSPACE: sopt.sopt_td = NULL; break; default: panic("kern_setsockopt called with bad valseg"); } AUDIT_ARG_FD(s); error = getsock_cap(td, s, cap_rights_init(&rights, CAP_SETSOCKOPT), &fp, NULL); if (error == 0) { so = fp->f_data; error = sosetopt(so, &sopt); fdrop(fp, td); } return(error); } -/* ARGSUSED */ int -sys_getsockopt(td, uap) - struct thread *td; - struct getsockopt_args /* { - int s; - int level; - int name; - void * __restrict val; - socklen_t * __restrict avalsize; - } */ *uap; +sys_getsockopt(struct thread *td, struct getsockopt_args *uap) { socklen_t valsize; int error; if (uap->val) { error = copyin(uap->avalsize, &valsize, sizeof (valsize)); if (error != 0) return (error); } error = kern_getsockopt(td, uap->s, uap->level, uap->name, uap->val, UIO_USERSPACE, &valsize); if (error == 0) error = copyout(&valsize, uap->avalsize, sizeof (valsize)); return (error); } /* * Kernel version of getsockopt. * optval can be a userland or userspace. optlen is always a kernel pointer. */ int -kern_getsockopt(td, s, level, name, val, valseg, valsize) - struct thread *td; - int s; - int level; - int name; - void *val; - enum uio_seg valseg; - socklen_t *valsize; +kern_getsockopt(struct thread *td, int s, int level, int name, void *val, + enum uio_seg valseg, socklen_t *valsize) { struct socket *so; struct file *fp; struct sockopt sopt; cap_rights_t rights; int error; if (val == NULL) *valsize = 0; if ((int)*valsize < 0) return (EINVAL); sopt.sopt_dir = SOPT_GET; sopt.sopt_level = level; sopt.sopt_name = name; sopt.sopt_val = val; sopt.sopt_valsize = (size_t)*valsize; /* checked non-negative above */ switch (valseg) { case UIO_USERSPACE: sopt.sopt_td = td; break; case UIO_SYSSPACE: sopt.sopt_td = NULL; break; default: panic("kern_getsockopt called with bad valseg"); } AUDIT_ARG_FD(s); error = getsock_cap(td, s, cap_rights_init(&rights, CAP_GETSOCKOPT), &fp, NULL); if (error == 0) { so = fp->f_data; error = sogetopt(so, &sopt); *valsize = sopt.sopt_valsize; fdrop(fp, td); } return (error); } /* * getsockname1() - Get socket name. */ -/* ARGSUSED */ static int -getsockname1(td, uap, compat) - struct thread *td; - struct getsockname_args /* { - int fdes; - struct sockaddr * __restrict asa; - socklen_t * __restrict alen; - } */ *uap; - int compat; +getsockname1(struct thread *td, struct getsockname_args *uap, int compat) { struct sockaddr *sa; socklen_t len; int error; error = copyin(uap->alen, &len, sizeof(len)); if (error != 0) return (error); error = kern_getsockname(td, uap->fdes, &sa, &len); if (error != 0) return (error); if (len != 0) { #ifdef COMPAT_OLDSOCK if (compat) ((struct osockaddr *)sa)->sa_family = sa->sa_family; #endif error = copyout(sa, uap->asa, (u_int)len); } free(sa, M_SONAME); if (error == 0) error = copyout(&len, uap->alen, sizeof(len)); return (error); } int kern_getsockname(struct thread *td, int fd, struct sockaddr **sa, socklen_t *alen) { struct socket *so; struct file *fp; cap_rights_t rights; socklen_t len; int error; AUDIT_ARG_FD(fd); error = getsock_cap(td, fd, cap_rights_init(&rights, CAP_GETSOCKNAME), &fp, NULL); if (error != 0) return (error); so = fp->f_data; *sa = NULL; CURVNET_SET(so->so_vnet); error = (*so->so_proto->pr_usrreqs->pru_sockaddr)(so, sa); CURVNET_RESTORE(); if (error != 0) goto bad; if (*sa == NULL) len = 0; else len = MIN(*alen, (*sa)->sa_len); *alen = len; #ifdef KTRACE if (KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(*sa); #endif bad: fdrop(fp, td); if (error != 0 && *sa != NULL) { free(*sa, M_SONAME); *sa = NULL; } return (error); } int -sys_getsockname(td, uap) - struct thread *td; - struct getsockname_args *uap; +sys_getsockname(struct thread *td, struct getsockname_args *uap) { return (getsockname1(td, uap, 0)); } #ifdef COMPAT_OLDSOCK int -ogetsockname(td, uap) - struct thread *td; - struct getsockname_args *uap; +ogetsockname(struct thread *td, struct getsockname_args *uap) { return (getsockname1(td, uap, 1)); } #endif /* COMPAT_OLDSOCK */ /* * getpeername1() - Get name of peer for connected socket. */ -/* ARGSUSED */ static int -getpeername1(td, uap, compat) - struct thread *td; - struct getpeername_args /* { - int fdes; - struct sockaddr * __restrict asa; - socklen_t * __restrict alen; - } */ *uap; - int compat; +getpeername1(struct thread *td, struct getpeername_args *uap, int compat) { struct sockaddr *sa; socklen_t len; int error; error = copyin(uap->alen, &len, sizeof (len)); if (error != 0) return (error); error = kern_getpeername(td, uap->fdes, &sa, &len); if (error != 0) return (error); if (len != 0) { #ifdef COMPAT_OLDSOCK if (compat) ((struct osockaddr *)sa)->sa_family = sa->sa_family; #endif error = copyout(sa, uap->asa, (u_int)len); } free(sa, M_SONAME); if (error == 0) error = copyout(&len, uap->alen, sizeof(len)); return (error); } int kern_getpeername(struct thread *td, int fd, struct sockaddr **sa, socklen_t *alen) { struct socket *so; struct file *fp; cap_rights_t rights; socklen_t len; int error; AUDIT_ARG_FD(fd); error = getsock_cap(td, fd, cap_rights_init(&rights, CAP_GETPEERNAME), &fp, NULL); if (error != 0) return (error); so = fp->f_data; if ((so->so_state & (SS_ISCONNECTED|SS_ISCONFIRMING)) == 0) { error = ENOTCONN; goto done; } *sa = NULL; CURVNET_SET(so->so_vnet); error = (*so->so_proto->pr_usrreqs->pru_peeraddr)(so, sa); CURVNET_RESTORE(); if (error != 0) goto bad; if (*sa == NULL) len = 0; else len = MIN(*alen, (*sa)->sa_len); *alen = len; #ifdef KTRACE if (KTRPOINT(td, KTR_STRUCT)) ktrsockaddr(*sa); #endif bad: if (error != 0 && *sa != NULL) { free(*sa, M_SONAME); *sa = NULL; } done: fdrop(fp, td); return (error); } int -sys_getpeername(td, uap) - struct thread *td; - struct getpeername_args *uap; +sys_getpeername(struct thread *td, struct getpeername_args *uap) { return (getpeername1(td, uap, 0)); } #ifdef COMPAT_OLDSOCK int -ogetpeername(td, uap) - struct thread *td; - struct ogetpeername_args *uap; +ogetpeername(struct thread *td, struct ogetpeername_args *uap) { /* XXX uap should have type `getpeername_args *' to begin with. */ return (getpeername1(td, (struct getpeername_args *)uap, 1)); } #endif /* COMPAT_OLDSOCK */ static int sockargs(struct mbuf **mp, char *buf, socklen_t buflen, int type) { struct sockaddr *sa; struct mbuf *m; int error; if (buflen > MLEN) { #ifdef COMPAT_OLDSOCK if (type == MT_SONAME && buflen <= 112) buflen = MLEN; /* unix domain compat. hack */ else #endif if (buflen > MCLBYTES) return (EINVAL); } m = m_get2(buflen, M_WAITOK, type, 0); m->m_len = buflen; error = copyin(buf, mtod(m, void *), buflen); if (error != 0) (void) m_free(m); else { *mp = m; if (type == MT_SONAME) { sa = mtod(m, struct sockaddr *); #if defined(COMPAT_OLDSOCK) && BYTE_ORDER != BIG_ENDIAN if (sa->sa_family == 0 && sa->sa_len < AF_MAX) sa->sa_family = sa->sa_len; #endif sa->sa_len = buflen; } } return (error); } int -getsockaddr(namp, uaddr, len) - struct sockaddr **namp; - caddr_t uaddr; - size_t len; +getsockaddr(struct sockaddr **namp, caddr_t uaddr, size_t len) { struct sockaddr *sa; int error; if (len > SOCK_MAXADDRLEN) return (ENAMETOOLONG); if (len < offsetof(struct sockaddr, sa_data[0])) return (EINVAL); sa = malloc(len, M_SONAME, M_WAITOK); error = copyin(uaddr, sa, len); if (error != 0) { free(sa, M_SONAME); } else { #if defined(COMPAT_OLDSOCK) && BYTE_ORDER != BIG_ENDIAN if (sa->sa_family == 0 && sa->sa_len < AF_MAX) sa->sa_family = sa->sa_len; #endif sa->sa_len = len; *namp = sa; } return (error); } Index: projects/mips64-clang/sys/kern/vfs_cache.c =================================================================== --- projects/mips64-clang/sys/kern/vfs_cache.c (revision 305711) +++ projects/mips64-clang/sys/kern/vfs_cache.c (revision 305712) @@ -1,1591 +1,1739 @@ /*- * Copyright (c) 1989, 1993, 1995 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * Poul-Henning Kamp of the FreeBSD Project. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_cache.c 8.5 (Berkeley) 3/22/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_ktrace.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include +#include #include #include #include #include #ifdef KTRACE #include #endif #include SDT_PROVIDER_DECLARE(vfs); SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *", "char *"); SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *"); SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *"); SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int", "struct vnode *", "char *"); SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative, "struct vnode *", "char *"); SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *", "char *"); SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *"); SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *"); SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *"); SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *", "struct vnode *"); SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *", "char *"); /* * This structure describes the elements in the cache of recent * names looked up by namei. */ struct namecache { LIST_ENTRY(namecache) nc_hash; /* hash chain */ LIST_ENTRY(namecache) nc_src; /* source vnode list */ TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ struct vnode *nc_dvp; /* vnode of parent of name */ struct vnode *nc_vp; /* vnode the name refers to */ u_char nc_flag; /* flag bits */ u_char nc_nlen; /* length of name */ char nc_name[0]; /* segment name + nul */ }; /* * struct namecache_ts repeats struct namecache layout up to the * nc_nlen member. * struct namecache_ts is used in place of struct namecache when time(s) need * to be stored. The nc_dotdottime field is used when a cache entry is mapping * both a non-dotdot directory name plus dotdot for the directory's * parent. */ struct namecache_ts { LIST_ENTRY(namecache) nc_hash; /* hash chain */ LIST_ENTRY(namecache) nc_src; /* source vnode list */ TAILQ_ENTRY(namecache) nc_dst; /* destination vnode list */ struct vnode *nc_dvp; /* vnode of parent of name */ struct vnode *nc_vp; /* vnode the name refers to */ u_char nc_flag; /* flag bits */ u_char nc_nlen; /* length of name */ struct timespec nc_time; /* timespec provided by fs */ struct timespec nc_dotdottime; /* dotdot timespec provided by fs */ int nc_ticks; /* ticks value when entry was added */ char nc_name[0]; /* segment name + nul */ }; /* * Flags in namecache.nc_flag */ #define NCF_WHITE 0x01 #define NCF_ISDOTDOT 0x02 #define NCF_TS 0x04 #define NCF_DTS 0x08 #define NCF_DVDROP 0x10 /* * Name caching works as follows: * * Names found by directory scans are retained in a cache * for future reference. It is managed LRU, so frequently * used names will hang around. Cache is indexed by hash value * obtained from (vp, name) where vp refers to the directory * containing name. * * If it is a "negative" entry, (i.e. for a name that is known NOT to * exist) the vnode pointer will be NULL. * * Upon reaching the last segment of a path, if the reference * is for DELETE, or NOCACHE is set (rewrite), and the * name is located in the cache, it will be dropped. + * + * These locks are used (in the order in which they can be taken): + * NAME TYPE ROLE + * cache_lock rwlock global, needed for all modifications + * bucketlock rwlock for access to given hash bucket + * ncneg_mtx mtx negative entry LRU management + * + * A name -> vnode lookup can be safely performed by either locking cache_lock + * or the relevant hash bucket. + * + * ".." and vnode -> name lookups require cache_lock. + * + * Modifications require both cache_lock and relevant bucketlock taken for + * writing. + * + * Negative entry LRU management requires ncneg_mtx taken on top of either + * cache_lock or bucketlock. */ /* * Structures associated with name caching. */ #define NCHHASH(hash) \ (&nchashtbl[(hash) & nchash]) static LIST_HEAD(nchashhead, namecache) *nchashtbl; /* Hash Table */ static TAILQ_HEAD(, namecache) ncneg; /* Hash Table */ static u_long nchash; /* size of hash table */ SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0, "Size of namecache hash table"); static u_long ncnegfactor = 16; /* ratio of negative entries */ SYSCTL_ULONG(_vfs, OID_AUTO, ncnegfactor, CTLFLAG_RW, &ncnegfactor, 0, "Ratio of negative namecache entries"); static u_long numneg; /* number of negative entries allocated */ SYSCTL_ULONG(_debug, OID_AUTO, numneg, CTLFLAG_RD, &numneg, 0, "Number of negative entries in namecache"); static u_long numcache; /* number of cache entries allocated */ SYSCTL_ULONG(_debug, OID_AUTO, numcache, CTLFLAG_RD, &numcache, 0, "Number of namecache entries"); static u_long numcachehv; /* number of cache entries with vnodes held */ SYSCTL_ULONG(_debug, OID_AUTO, numcachehv, CTLFLAG_RD, &numcachehv, 0, "Number of namecache entries with vnodes held"); u_int ncsizefactor = 2; SYSCTL_UINT(_vfs, OID_AUTO, ncsizefactor, CTLFLAG_RW, &ncsizefactor, 0, "Size factor for namecache"); struct nchstats nchstats; /* cache effectiveness statistics */ static struct rwlock cache_lock; -RW_SYSINIT(vfscache, &cache_lock, "Name Cache"); +RW_SYSINIT(vfscache, &cache_lock, "ncglobal"); +#define CACHE_TRY_WLOCK() rw_try_wlock(&cache_lock) #define CACHE_UPGRADE_LOCK() rw_try_upgrade(&cache_lock) #define CACHE_RLOCK() rw_rlock(&cache_lock) #define CACHE_RUNLOCK() rw_runlock(&cache_lock) #define CACHE_WLOCK() rw_wlock(&cache_lock) #define CACHE_WUNLOCK() rw_wunlock(&cache_lock) static struct mtx_padalign ncneg_mtx; -MTX_SYSINIT(vfscache_neg, &ncneg_mtx, "Name Cache neg", MTX_DEF); +MTX_SYSINIT(vfscache_neg, &ncneg_mtx, "ncneg", MTX_DEF); +static u_int numbucketlocks; +static struct rwlock_padalign *bucketlocks; +#define HASH2BUCKETLOCK(hash) \ + ((struct rwlock *)(&bucketlocks[((hash) % numbucketlocks)])) + /* * UMA zones for the VFS cache. * * The small cache is used for entries with short names, which are the * most common. The large cache is used for entries which are too big to * fit in the small cache. */ static uma_zone_t cache_zone_small; static uma_zone_t cache_zone_small_ts; static uma_zone_t cache_zone_large; static uma_zone_t cache_zone_large_ts; #define CACHE_PATH_CUTOFF 35 static struct namecache * cache_alloc(int len, int ts) { if (len > CACHE_PATH_CUTOFF) { if (ts) return (uma_zalloc(cache_zone_large_ts, M_WAITOK)); else return (uma_zalloc(cache_zone_large, M_WAITOK)); } if (ts) return (uma_zalloc(cache_zone_small_ts, M_WAITOK)); else return (uma_zalloc(cache_zone_small, M_WAITOK)); } static void cache_free(struct namecache *ncp) { int ts; if (ncp == NULL) return; ts = ncp->nc_flag & NCF_TS; if ((ncp->nc_flag & NCF_DVDROP) != 0) vdrop(ncp->nc_dvp); if (ncp->nc_nlen <= CACHE_PATH_CUTOFF) { if (ts) uma_zfree(cache_zone_small_ts, ncp); else uma_zfree(cache_zone_small, ncp); } else if (ts) uma_zfree(cache_zone_large_ts, ncp); else uma_zfree(cache_zone_large, ncp); } static char * nc_get_name(struct namecache *ncp) { struct namecache_ts *ncp_ts; if ((ncp->nc_flag & NCF_TS) == 0) return (ncp->nc_name); ncp_ts = (struct namecache_ts *)ncp; return (ncp_ts->nc_name); } static void cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp) { KASSERT((ncp->nc_flag & NCF_TS) != 0 || (tsp == NULL && ticksp == NULL), ("No NCF_TS")); if (tsp != NULL) *tsp = ((struct namecache_ts *)ncp)->nc_time; if (ticksp != NULL) *ticksp = ((struct namecache_ts *)ncp)->nc_ticks; } static int doingcache = 1; /* 1 => enable the cache */ SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0, "VFS namecache enabled"); /* Export size information to userland */ SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR, sizeof(struct namecache), "sizeof(struct namecache)"); /* * The new name cache statistics */ static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW, 0, "Name cache statistics"); #define STATNODE_ULONG(name, descr) \ SYSCTL_ULONG(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, 0, descr); #define STATNODE_COUNTER(name, descr) \ static counter_u64_t name; \ SYSCTL_COUNTER_U64(_vfs_cache, OID_AUTO, name, CTLFLAG_RD, &name, descr); STATNODE_ULONG(numneg, "Number of negative cache entries"); STATNODE_ULONG(numcache, "Number of cache entries"); STATNODE_COUNTER(numcalls, "Number of cache lookups"); STATNODE_COUNTER(dothits, "Number of '.' hits"); STATNODE_COUNTER(dotdothits, "Number of '..' hits"); STATNODE_COUNTER(numchecks, "Number of checks in lookup"); STATNODE_COUNTER(nummiss, "Number of cache misses"); STATNODE_COUNTER(nummisszap, "Number of cache misses we do not want to cache"); STATNODE_COUNTER(numposzaps, "Number of cache hits (positive) we do not want to cache"); STATNODE_COUNTER(numposhits, "Number of cache hits (positive)"); STATNODE_COUNTER(numnegzaps, "Number of cache hits (negative) we do not want to cache"); STATNODE_COUNTER(numneghits, "Number of cache hits (negative)"); /* These count for kern___getcwd(), too. */ STATNODE_COUNTER(numfullpathcalls, "Number of fullpath search calls"); STATNODE_COUNTER(numfullpathfail1, "Number of fullpath search errors (ENOTDIR)"); STATNODE_COUNTER(numfullpathfail2, "Number of fullpath search errors (VOP_VPTOCNP failures)"); STATNODE_COUNTER(numfullpathfail4, "Number of fullpath search errors (ENOMEM)"); STATNODE_COUNTER(numfullpathfound, "Number of successful fullpath calls"); static long numupgrades; STATNODE_ULONG(numupgrades, "Number of updates of the cache after lookup (write lock + retry)"); +static long zap_and_exit_bucket_fail; STATNODE_ULONG(zap_and_exit_bucket_fail, + "Number of times bucketlocked zap_and_exit case failed to writelock"); static void cache_zap(struct namecache *ncp); static int vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen); static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, u_int buflen); static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries"); static uint32_t cache_get_hash(char *name, u_char len, struct vnode *dvp) { uint32_t hash; hash = fnv_32_buf(name, len, FNV1_32_INIT); hash = fnv_32_buf(&dvp, sizeof(dvp), hash); return (hash); } +#ifdef INVARIANTS +static void +cache_assert_bucket_locked(struct namecache *ncp, int mode) +{ + struct rwlock *bucketlock; + uint32_t hash; + + hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp); + bucketlock = HASH2BUCKETLOCK(hash); + rw_assert(bucketlock, mode); +} +#else +#define cache_assert_bucket_locked(x, y) do { } while (0) +#endif + +static void +cache_lock_all_buckets(void) +{ + u_int i; + + for (i = 0; i < numbucketlocks; i++) + rw_wlock(&bucketlocks[i]); +} + +static void +cache_unlock_all_buckets(void) +{ + u_int i; + + for (i = 0; i < numbucketlocks; i++) + rw_wunlock(&bucketlocks[i]); +} + static int sysctl_nchstats(SYSCTL_HANDLER_ARGS) { struct nchstats snap; if (req->oldptr == NULL) return (SYSCTL_OUT(req, 0, sizeof(snap))); snap = nchstats; snap.ncs_goodhits = counter_u64_fetch(numposhits); snap.ncs_neghits = counter_u64_fetch(numneghits); snap.ncs_badhits = counter_u64_fetch(numposzaps) + counter_u64_fetch(numnegzaps); snap.ncs_miss = counter_u64_fetch(nummisszap) + counter_u64_fetch(nummiss); return (SYSCTL_OUT(req, &snap, sizeof(snap))); } SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU", "VFS cache effectiveness statistics"); #ifdef DIAGNOSTIC /* * Grab an atomic snapshot of the name cache hash chain lengths */ static SYSCTL_NODE(_debug, OID_AUTO, hashstat, CTLFLAG_RW, NULL, "hash table stats"); static int sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS) { struct nchashhead *ncpp; struct namecache *ncp; int i, error, n_nchash, *cntbuf; retry: n_nchash = nchash + 1; /* nchash is max index, not count */ if (req->oldptr == NULL) return SYSCTL_OUT(req, 0, n_nchash * sizeof(int)); cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK); CACHE_RLOCK(); if (n_nchash != nchash + 1) { CACHE_RUNLOCK(); free(cntbuf, M_TEMP); goto retry; } /* Scan hash tables counting entries */ for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++) LIST_FOREACH(ncp, ncpp, nc_hash) cntbuf[i]++; CACHE_RUNLOCK(); for (error = 0, i = 0; i < n_nchash; i++) if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0) break; free(cntbuf, M_TEMP); return (error); } SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD| CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int", "nchash chain lengths"); static int sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS) { int error; struct nchashhead *ncpp; struct namecache *ncp; int n_nchash; int count, maxlength, used, pct; if (!req->oldptr) return SYSCTL_OUT(req, 0, 4 * sizeof(int)); CACHE_RLOCK(); n_nchash = nchash + 1; /* nchash is max index, not count */ used = 0; maxlength = 0; /* Scan hash tables for applicable entries */ for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) { count = 0; LIST_FOREACH(ncp, ncpp, nc_hash) { count++; } if (count) used++; if (maxlength < count) maxlength = count; } n_nchash = nchash + 1; CACHE_RUNLOCK(); pct = (used * 100) / (n_nchash / 100); error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash)); if (error) return (error); error = SYSCTL_OUT(req, &used, sizeof(used)); if (error) return (error); error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength)); if (error) return (error); error = SYSCTL_OUT(req, &pct, sizeof(pct)); if (error) return (error); return (0); } SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD| CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I", "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)"); #endif /* * Negative entries management */ static void -cache_negative_hit(struct namecache *ncp, int wlocked) +cache_negative_hit(struct namecache *ncp) { - if (!wlocked) { - rw_assert(&cache_lock, RA_RLOCKED); - mtx_lock(&ncneg_mtx); - } else { - rw_assert(&cache_lock, RA_WLOCKED); - } - + mtx_lock(&ncneg_mtx); TAILQ_REMOVE(&ncneg, ncp, nc_dst); TAILQ_INSERT_TAIL(&ncneg, ncp, nc_dst); - - if (!wlocked) - mtx_unlock(&ncneg_mtx); + mtx_unlock(&ncneg_mtx); } static void cache_negative_insert(struct namecache *ncp) { rw_assert(&cache_lock, RA_WLOCKED); + cache_assert_bucket_locked(ncp, RA_WLOCKED); MPASS(ncp->nc_vp == NULL); + mtx_lock(&ncneg_mtx); TAILQ_INSERT_TAIL(&ncneg, ncp, nc_dst); numneg++; + mtx_unlock(&ncneg_mtx); } static void cache_negative_remove(struct namecache *ncp) { rw_assert(&cache_lock, RA_WLOCKED); + cache_assert_bucket_locked(ncp, RA_WLOCKED); MPASS(ncp->nc_vp == NULL); + mtx_lock(&ncneg_mtx); TAILQ_REMOVE(&ncneg, ncp, nc_dst); numneg--; + mtx_unlock(&ncneg_mtx); } static struct namecache * cache_negative_zap_one(void) { struct namecache *ncp; rw_assert(&cache_lock, RA_WLOCKED); ncp = TAILQ_FIRST(&ncneg); KASSERT(ncp->nc_vp == NULL, ("ncp %p vp %p on ncneg", ncp, ncp->nc_vp)); cache_zap(ncp); return (ncp); } /* * cache_zap(): * * Removes a namecache entry from cache, whether it contains an actual * pointer to a vnode or if it is just a negative cache entry. */ static void -cache_zap(struct namecache *ncp) +cache_zap_locked(struct namecache *ncp) { rw_assert(&cache_lock, RA_WLOCKED); + cache_assert_bucket_locked(ncp, RA_WLOCKED); CTR2(KTR_VFS, "cache_zap(%p) vp %p", ncp, ncp->nc_vp); if (ncp->nc_vp != NULL) { SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp, nc_get_name(ncp), ncp->nc_vp); } else { SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp, nc_get_name(ncp)); } LIST_REMOVE(ncp, nc_hash); if (ncp->nc_flag & NCF_ISDOTDOT) { if (ncp == ncp->nc_dvp->v_cache_dd) ncp->nc_dvp->v_cache_dd = NULL; } else { LIST_REMOVE(ncp, nc_src); if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) { ncp->nc_flag |= NCF_DVDROP; numcachehv--; } } if (ncp->nc_vp) { TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst); if (ncp == ncp->nc_vp->v_cache_dd) ncp->nc_vp->v_cache_dd = NULL; } else { cache_negative_remove(ncp); } numcache--; } +static void +cache_zap(struct namecache *ncp) +{ + struct rwlock *bucketlock; + uint32_t hash; + + rw_assert(&cache_lock, RA_WLOCKED); + + hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp); + bucketlock = HASH2BUCKETLOCK(hash); + rw_wlock(bucketlock); + cache_zap_locked(ncp); + rw_wunlock(bucketlock); +} + /* * Lookup an entry in the cache * * Lookup is called with dvp pointing to the directory to search, * cnp pointing to the name of the entry being sought. If the lookup * succeeds, the vnode is returned in *vpp, and a status of -1 is * returned. If the lookup determines that the name does not exist * (negative caching), a status of ENOENT is returned. If the lookup * fails, a status of zero is returned. If the directory vnode is * recycled out from under us due to a forced unmount, a status of * ENOENT is returned. * * vpp is locked and ref'd on return. If we're looking up DOTDOT, dvp is * unlocked. If we're looking up . an extra ref is taken, but the lock is * not recursively acquired. */ +enum { UNLOCKED, WLOCKED, RLOCKED }; + +static void +cache_unlock(int cache_locked) +{ + + switch (cache_locked) { + case UNLOCKED: + break; + case WLOCKED: + CACHE_WUNLOCK(); + break; + case RLOCKED: + CACHE_RUNLOCK(); + break; + } +} + int cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp, struct timespec *tsp, int *ticksp) { + struct rwlock *bucketlock; struct namecache *ncp; uint32_t hash; - int error, ltype, wlocked; + int error, ltype, cache_locked; if (!doingcache) { cnp->cn_flags &= ~MAKEENTRY; return (0); } retry: - wlocked = 0; - counter_u64_add(numcalls, 1); + bucketlock = NULL; + cache_locked = UNLOCKED; error = 0; + counter_u64_add(numcalls, 1); retry_wlocked: if (cnp->cn_nameptr[0] == '.') { if (cnp->cn_namelen == 1) { *vpp = dvp; CTR2(KTR_VFS, "cache_lookup(%p, %s) found via .", dvp, cnp->cn_nameptr); counter_u64_add(dothits, 1); SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp); if (tsp != NULL) timespecclear(tsp); if (ticksp != NULL) *ticksp = ticks; VREF(*vpp); /* * When we lookup "." we still can be asked to lock it * differently... */ ltype = cnp->cn_lkflags & LK_TYPE_MASK; if (ltype != VOP_ISLOCKED(*vpp)) { if (ltype == LK_EXCLUSIVE) { vn_lock(*vpp, LK_UPGRADE | LK_RETRY); if ((*vpp)->v_iflag & VI_DOOMED) { /* forced unmount */ vrele(*vpp); *vpp = NULL; return (ENOENT); } } else vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY); } return (-1); } - if (!wlocked) - CACHE_RLOCK(); if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { counter_u64_add(dotdothits, 1); + if (cache_locked == UNLOCKED) { + CACHE_RLOCK(); + cache_locked = RLOCKED; + } + if (dvp->v_cache_dd == NULL) { SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL); goto unlock; } if ((cnp->cn_flags & MAKEENTRY) == 0) { - if (!wlocked && !CACHE_UPGRADE_LOCK()) + if (cache_locked != WLOCKED && + !CACHE_UPGRADE_LOCK()) goto wlock; ncp = NULL; if (dvp->v_cache_dd->nc_flag & NCF_ISDOTDOT) { ncp = dvp->v_cache_dd; cache_zap(ncp); } dvp->v_cache_dd = NULL; CACHE_WUNLOCK(); cache_free(ncp); return (0); } ncp = dvp->v_cache_dd; if (ncp->nc_flag & NCF_ISDOTDOT) *vpp = ncp->nc_vp; else *vpp = ncp->nc_dvp; /* Return failure if negative entry was found. */ if (*vpp == NULL) goto negative_success; CTR3(KTR_VFS, "cache_lookup(%p, %s) found %p via ..", dvp, cnp->cn_nameptr, *vpp); SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp); cache_out_ts(ncp, tsp, ticksp); if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) == NCF_DTS && tsp != NULL) *tsp = ((struct namecache_ts *)ncp)-> nc_dotdottime; goto success; } - } else if (!wlocked) - CACHE_RLOCK(); + } hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp); + if (cache_locked == UNLOCKED) { + bucketlock = HASH2BUCKETLOCK(hash); + rw_rlock(bucketlock); + } + LIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) { counter_u64_add(numchecks, 1); if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen && !bcmp(nc_get_name(ncp), cnp->cn_nameptr, ncp->nc_nlen)) break; } /* We failed to find an entry */ if (ncp == NULL) { SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr, NULL); if ((cnp->cn_flags & MAKEENTRY) == 0) { counter_u64_add(nummisszap, 1); } else { counter_u64_add(nummiss, 1); } goto unlock; } /* We don't want to have an entry, so dump it */ if ((cnp->cn_flags & MAKEENTRY) == 0) { counter_u64_add(numposzaps, 1); - if (!wlocked && !CACHE_UPGRADE_LOCK()) - goto wlock; - cache_zap(ncp); - CACHE_WUNLOCK(); - cache_free(ncp); - return (0); + goto zap_and_exit; } /* We found a "positive" match, return the vnode */ if (ncp->nc_vp) { counter_u64_add(numposhits, 1); *vpp = ncp->nc_vp; CTR4(KTR_VFS, "cache_lookup(%p, %s) found %p via ncp %p", dvp, cnp->cn_nameptr, *vpp, ncp); SDT_PROBE3(vfs, namecache, lookup, hit, dvp, nc_get_name(ncp), *vpp); cache_out_ts(ncp, tsp, ticksp); goto success; } negative_success: /* We found a negative match, and want to create it, so purge */ if (cnp->cn_nameiop == CREATE) { counter_u64_add(numnegzaps, 1); - if (!wlocked && !CACHE_UPGRADE_LOCK()) - goto wlock; - cache_zap(ncp); - CACHE_WUNLOCK(); - cache_free(ncp); - return (0); + goto zap_and_exit; } counter_u64_add(numneghits, 1); - cache_negative_hit(ncp, wlocked); + cache_negative_hit(ncp); if (ncp->nc_flag & NCF_WHITE) cnp->cn_flags |= ISWHITEOUT; SDT_PROBE2(vfs, namecache, lookup, hit__negative, dvp, nc_get_name(ncp)); cache_out_ts(ncp, tsp, ticksp); - if (wlocked) - CACHE_WUNLOCK(); - else - CACHE_RUNLOCK(); + MPASS(bucketlock != NULL || cache_locked != UNLOCKED); + if (bucketlock != NULL) + rw_runlock(bucketlock); + cache_unlock(cache_locked); return (ENOENT); wlock: /* * We need to update the cache after our lookup, so upgrade to * a write lock and retry the operation. */ CACHE_RUNLOCK(); +wlock_unlocked: CACHE_WLOCK(); numupgrades++; - wlocked = 1; + cache_locked = WLOCKED; goto retry_wlocked; success: /* * On success we return a locked and ref'd vnode as per the lookup * protocol. */ MPASS(dvp != *vpp); ltype = 0; /* silence gcc warning */ if (cnp->cn_flags & ISDOTDOT) { ltype = VOP_ISLOCKED(dvp); VOP_UNLOCK(dvp, 0); } vhold(*vpp); - if (wlocked) - CACHE_WUNLOCK(); - else - CACHE_RUNLOCK(); + MPASS(bucketlock != NULL || cache_locked != UNLOCKED); + if (bucketlock != NULL) + rw_runlock(bucketlock); + cache_unlock(cache_locked); error = vget(*vpp, cnp->cn_lkflags | LK_VNHELD, cnp->cn_thread); if (cnp->cn_flags & ISDOTDOT) { vn_lock(dvp, ltype | LK_RETRY); if (dvp->v_iflag & VI_DOOMED) { if (error == 0) vput(*vpp); *vpp = NULL; return (ENOENT); } } if (error) { *vpp = NULL; goto retry; } if ((cnp->cn_flags & ISLASTCN) && (cnp->cn_lkflags & LK_TYPE_MASK) == LK_EXCLUSIVE) { ASSERT_VOP_ELOCKED(*vpp, "cache_lookup"); } return (-1); unlock: - if (wlocked) - CACHE_WUNLOCK(); - else - CACHE_RUNLOCK(); + MPASS(bucketlock != NULL || cache_locked != UNLOCKED); + if (bucketlock != NULL) + rw_runlock(bucketlock); + cache_unlock(cache_locked); return (0); + +zap_and_exit: + if (bucketlock != NULL) { + rw_assert(&cache_lock, RA_UNLOCKED); + if (!CACHE_TRY_WLOCK()) { + rw_runlock(bucketlock); + bucketlock = NULL; + zap_and_exit_bucket_fail++; + goto wlock_unlocked; + } + cache_locked = WLOCKED; + rw_runlock(bucketlock); + bucketlock = NULL; + } else if (cache_locked != WLOCKED && !CACHE_UPGRADE_LOCK()) + goto wlock; + cache_zap(ncp); + CACHE_WUNLOCK(); + cache_free(ncp); + return (0); } /* * Add an entry to the cache. */ void cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp, struct timespec *tsp, struct timespec *dtsp) { + struct rwlock *bucketlock; struct namecache *ncp, *n2, *ndd, *nneg; struct namecache_ts *n3; struct nchashhead *ncpp; uint32_t hash; int flag; int len; CTR3(KTR_VFS, "cache_enter(%p, %p, %s)", dvp, vp, cnp->cn_nameptr); VNASSERT(vp == NULL || (vp->v_iflag & VI_DOOMED) == 0, vp, ("cache_enter: Adding a doomed vnode")); VNASSERT(dvp == NULL || (dvp->v_iflag & VI_DOOMED) == 0, dvp, ("cache_enter: Doomed vnode used as src")); if (!doingcache) return; /* * Avoid blowout in namecache entries. */ if (numcache >= desiredvnodes * ncsizefactor) return; ndd = nneg = NULL; flag = 0; if (cnp->cn_nameptr[0] == '.') { if (cnp->cn_namelen == 1) return; if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') { CACHE_WLOCK(); /* * If dotdot entry already exists, just retarget it * to new parent vnode, otherwise continue with new * namecache entry allocation. */ if ((ncp = dvp->v_cache_dd) != NULL && ncp->nc_flag & NCF_ISDOTDOT) { KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent")); if (ncp->nc_vp != NULL) { TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst); } else { cache_negative_remove(ncp); } if (vp != NULL) { TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst); } else { cache_negative_insert(ncp); } ncp->nc_vp = vp; CACHE_WUNLOCK(); return; } dvp->v_cache_dd = NULL; SDT_PROBE3(vfs, namecache, enter, done, dvp, "..", vp); CACHE_WUNLOCK(); flag = NCF_ISDOTDOT; } } /* * Calculate the hash key and setup as much of the new * namecache entry as possible before acquiring the lock. */ ncp = cache_alloc(cnp->cn_namelen, tsp != NULL); ncp->nc_vp = vp; ncp->nc_dvp = dvp; ncp->nc_flag = flag; if (tsp != NULL) { n3 = (struct namecache_ts *)ncp; n3->nc_time = *tsp; n3->nc_ticks = ticks; n3->nc_flag |= NCF_TS; if (dtsp != NULL) { n3->nc_dotdottime = *dtsp; n3->nc_flag |= NCF_DTS; } } len = ncp->nc_nlen = cnp->cn_namelen; hash = cache_get_hash(cnp->cn_nameptr, len, dvp); strlcpy(nc_get_name(ncp), cnp->cn_nameptr, len + 1); CACHE_WLOCK(); /* * See if this vnode or negative entry is already in the cache * with this name. This can happen with concurrent lookups of * the same path name. */ ncpp = NCHHASH(hash); LIST_FOREACH(n2, ncpp, nc_hash) { if (n2->nc_dvp == dvp && n2->nc_nlen == cnp->cn_namelen && !bcmp(nc_get_name(n2), cnp->cn_nameptr, n2->nc_nlen)) { if (tsp != NULL) { KASSERT((n2->nc_flag & NCF_TS) != 0, ("no NCF_TS")); n3 = (struct namecache_ts *)n2; n3->nc_time = ((struct namecache_ts *)ncp)->nc_time; n3->nc_ticks = ((struct namecache_ts *)ncp)->nc_ticks; if (dtsp != NULL) { n3->nc_dotdottime = ((struct namecache_ts *)ncp)-> nc_dotdottime; n3->nc_flag |= NCF_DTS; } } CACHE_WUNLOCK(); cache_free(ncp); return; } } if (flag == NCF_ISDOTDOT) { /* * See if we are trying to add .. entry, but some other lookup * has populated v_cache_dd pointer already. */ if (dvp->v_cache_dd != NULL) { CACHE_WUNLOCK(); cache_free(ncp); return; } KASSERT(vp == NULL || vp->v_type == VDIR, ("wrong vnode type %p", vp)); dvp->v_cache_dd = ncp; } numcache++; if (vp != NULL) { if (vp->v_type == VDIR) { if (flag != NCF_ISDOTDOT) { /* * For this case, the cache entry maps both the * directory name in it and the name ".." for the * directory's parent. */ if ((ndd = vp->v_cache_dd) != NULL) { if ((ndd->nc_flag & NCF_ISDOTDOT) != 0) cache_zap(ndd); else ndd = NULL; } vp->v_cache_dd = ncp; } } else { vp->v_cache_dd = NULL; } } - /* - * Insert the new namecache entry into the appropriate chain - * within the cache entries table. - */ - LIST_INSERT_HEAD(ncpp, ncp, nc_hash); if (flag != NCF_ISDOTDOT) { if (LIST_EMPTY(&dvp->v_cache_src)) { vhold(dvp); numcachehv++; } LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src); } + bucketlock = HASH2BUCKETLOCK(hash); + rw_wlock(bucketlock); + /* + * Insert the new namecache entry into the appropriate chain + * within the cache entries table. + */ + LIST_INSERT_HEAD(ncpp, ncp, nc_hash); + + /* * If the entry is "negative", we place it into the * "negative" cache queue, otherwise, we place it into the * destination vnode's cache entries queue. */ if (vp != NULL) { TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst); SDT_PROBE3(vfs, namecache, enter, done, dvp, nc_get_name(ncp), vp); } else { if (cnp->cn_flags & ISWHITEOUT) ncp->nc_flag |= NCF_WHITE; cache_negative_insert(ncp); SDT_PROBE2(vfs, namecache, enter_negative, done, dvp, nc_get_name(ncp)); } + rw_wunlock(bucketlock); if (numneg * ncnegfactor > numcache) nneg = cache_negative_zap_one(); CACHE_WUNLOCK(); cache_free(ndd); cache_free(nneg); } +static u_int +cache_roundup_2(u_int val) +{ + u_int res; + + for (res = 1; res <= val; res <<= 1) + continue; + + return (res); +} + /* * Name cache initialization, from vfs_init() when we are booting */ static void nchinit(void *dummy __unused) { + u_int i; TAILQ_INIT(&ncneg); cache_zone_small = uma_zcreate("S VFS Cache", sizeof(struct namecache) + CACHE_PATH_CUTOFF + 1, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); cache_zone_small_ts = uma_zcreate("STS VFS Cache", sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); cache_zone_large = uma_zcreate("L VFS Cache", sizeof(struct namecache) + NAME_MAX + 1, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); cache_zone_large_ts = uma_zcreate("LTS VFS Cache", sizeof(struct namecache_ts) + NAME_MAX + 1, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_ZINIT); nchashtbl = hashinit(desiredvnodes * 2, M_VFSCACHE, &nchash); + numbucketlocks = cache_roundup_2(mp_ncpus * 16); + if (numbucketlocks > nchash) + numbucketlocks = nchash; + bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE, + M_WAITOK | M_ZERO); + for (i = 0; i < numbucketlocks; i++) + rw_init_flags(&bucketlocks[i], "ncbuc", RW_DUPOK); numcalls = counter_u64_alloc(M_WAITOK); dothits = counter_u64_alloc(M_WAITOK); dotdothits = counter_u64_alloc(M_WAITOK); numchecks = counter_u64_alloc(M_WAITOK); nummiss = counter_u64_alloc(M_WAITOK); nummisszap = counter_u64_alloc(M_WAITOK); numposzaps = counter_u64_alloc(M_WAITOK); numposhits = counter_u64_alloc(M_WAITOK); numnegzaps = counter_u64_alloc(M_WAITOK); numneghits = counter_u64_alloc(M_WAITOK); numfullpathcalls = counter_u64_alloc(M_WAITOK); numfullpathfail1 = counter_u64_alloc(M_WAITOK); numfullpathfail2 = counter_u64_alloc(M_WAITOK); numfullpathfail4 = counter_u64_alloc(M_WAITOK); numfullpathfound = counter_u64_alloc(M_WAITOK); } SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL); void cache_changesize(int newmaxvnodes) { struct nchashhead *new_nchashtbl, *old_nchashtbl; u_long new_nchash, old_nchash; struct namecache *ncp; uint32_t hash; int i; new_nchashtbl = hashinit(newmaxvnodes * 2, M_VFSCACHE, &new_nchash); /* If same hash table size, nothing to do */ if (nchash == new_nchash) { free(new_nchashtbl, M_VFSCACHE); return; } /* * Move everything from the old hash table to the new table. * None of the namecache entries in the table can be removed * because to do so, they have to be removed from the hash table. */ CACHE_WLOCK(); + cache_lock_all_buckets(); old_nchashtbl = nchashtbl; old_nchash = nchash; nchashtbl = new_nchashtbl; nchash = new_nchash; for (i = 0; i <= old_nchash; i++) { while ((ncp = LIST_FIRST(&old_nchashtbl[i])) != NULL) { hash = cache_get_hash(nc_get_name(ncp), ncp->nc_nlen, ncp->nc_dvp); LIST_REMOVE(ncp, nc_hash); LIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash); } } + cache_unlock_all_buckets(); CACHE_WUNLOCK(); free(old_nchashtbl, M_VFSCACHE); } /* * Invalidate all entries to a particular vnode. */ void cache_purge(struct vnode *vp) { TAILQ_HEAD(, namecache) ncps; struct namecache *ncp, *nnp; CTR1(KTR_VFS, "cache_purge(%p)", vp); SDT_PROBE1(vfs, namecache, purge, done, vp); TAILQ_INIT(&ncps); CACHE_WLOCK(); while (!LIST_EMPTY(&vp->v_cache_src)) { ncp = LIST_FIRST(&vp->v_cache_src); cache_zap(ncp); TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); } while (!TAILQ_EMPTY(&vp->v_cache_dst)) { ncp = TAILQ_FIRST(&vp->v_cache_dst); cache_zap(ncp); TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); } if (vp->v_cache_dd != NULL) { ncp = vp->v_cache_dd; KASSERT(ncp->nc_flag & NCF_ISDOTDOT, ("lost dotdot link")); cache_zap(ncp); TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); } KASSERT(vp->v_cache_dd == NULL, ("incomplete purge")); CACHE_WUNLOCK(); TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) { cache_free(ncp); } } /* * Invalidate all negative entries for a particular directory vnode. */ void cache_purge_negative(struct vnode *vp) { TAILQ_HEAD(, namecache) ncps; struct namecache *ncp, *nnp; CTR1(KTR_VFS, "cache_purge_negative(%p)", vp); SDT_PROBE1(vfs, namecache, purge_negative, done, vp); TAILQ_INIT(&ncps); CACHE_WLOCK(); LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) { if (ncp->nc_vp != NULL) continue; cache_zap(ncp); TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); } CACHE_WUNLOCK(); TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) { cache_free(ncp); } } /* * Flush all entries referencing a particular filesystem. */ void cache_purgevfs(struct mount *mp) { TAILQ_HEAD(, namecache) ncps; - struct nchashhead *ncpp; + struct rwlock *bucketlock; + struct nchashhead *bucket; struct namecache *ncp, *nnp; + u_long i, j, n_nchash; /* Scan hash tables for applicable entries */ SDT_PROBE1(vfs, namecache, purgevfs, done, mp); TAILQ_INIT(&ncps); CACHE_WLOCK(); - for (ncpp = &nchashtbl[nchash]; ncpp >= nchashtbl; ncpp--) { - LIST_FOREACH_SAFE(ncp, ncpp, nc_hash, nnp) { - if (ncp->nc_dvp->v_mount != mp) - continue; - cache_zap(ncp); - TAILQ_INSERT_TAIL(&ncps, ncp, nc_dst); + n_nchash = nchash + 1; + for (i = 0; i < numbucketlocks; i++) { + bucketlock = (struct rwlock *)&bucketlocks[i]; + rw_wlock(bucketlock); + for (j = i; j < n_nchash; j += numbucketlocks) { + bucket = &nchashtbl[j]; + LIST_FOREACH_SAFE(ncp, bucket, nc_hash, nnp) { + cache_assert_bucket_locked(ncp, RA_WLOCKED); + if (ncp->nc_dvp->v_mount != mp) + continue; + cache_zap_locked(ncp); + TAILQ_INSERT_HEAD(&ncps, ncp, nc_dst); + } } + rw_wunlock(bucketlock); } CACHE_WUNLOCK(); TAILQ_FOREACH_SAFE(ncp, &ncps, nc_dst, nnp) { cache_free(ncp); } } /* * Perform canonical checks and cache lookup and pass on to filesystem * through the vop_cachedlookup only if needed. */ int vfs_cache_lookup(struct vop_lookup_args *ap) { struct vnode *dvp; int error; struct vnode **vpp = ap->a_vpp; struct componentname *cnp = ap->a_cnp; struct ucred *cred = cnp->cn_cred; int flags = cnp->cn_flags; struct thread *td = cnp->cn_thread; *vpp = NULL; dvp = ap->a_dvp; if (dvp->v_type != VDIR) return (ENOTDIR); if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) && (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)) return (EROFS); error = VOP_ACCESS(dvp, VEXEC, cred, td); if (error) return (error); error = cache_lookup(dvp, vpp, cnp, NULL, NULL); if (error == 0) return (VOP_CACHEDLOOKUP(dvp, vpp, cnp)); if (error == -1) return (0); return (error); } /* * XXX All of these sysctls would probably be more productive dead. */ static int disablecwd; SYSCTL_INT(_debug, OID_AUTO, disablecwd, CTLFLAG_RW, &disablecwd, 0, "Disable the getcwd syscall"); /* Implementation of the getcwd syscall. */ int sys___getcwd(struct thread *td, struct __getcwd_args *uap) { return (kern___getcwd(td, uap->buf, UIO_USERSPACE, uap->buflen, MAXPATHLEN)); } int kern___getcwd(struct thread *td, char *buf, enum uio_seg bufseg, u_int buflen, u_int path_max) { char *bp, *tmpbuf; struct filedesc *fdp; struct vnode *cdir, *rdir; int error; if (disablecwd) return (ENODEV); if (buflen < 2) return (EINVAL); if (buflen > path_max) buflen = path_max; tmpbuf = malloc(buflen, M_TEMP, M_WAITOK); fdp = td->td_proc->p_fd; FILEDESC_SLOCK(fdp); cdir = fdp->fd_cdir; VREF(cdir); rdir = fdp->fd_rdir; VREF(rdir); FILEDESC_SUNLOCK(fdp); error = vn_fullpath1(td, cdir, rdir, tmpbuf, &bp, buflen); vrele(rdir); vrele(cdir); if (!error) { if (bufseg == UIO_SYSSPACE) bcopy(bp, buf, strlen(bp) + 1); else error = copyout(bp, buf, strlen(bp) + 1); #ifdef KTRACE if (KTRPOINT(curthread, KTR_NAMEI)) ktrnamei(bp); #endif } free(tmpbuf, M_TEMP); return (error); } /* * Thus begins the fullpath magic. */ static int disablefullpath; SYSCTL_INT(_debug, OID_AUTO, disablefullpath, CTLFLAG_RW, &disablefullpath, 0, "Disable the vn_fullpath function"); /* * Retrieve the full filesystem path that correspond to a vnode from the name * cache (if available) */ int vn_fullpath(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf) { char *buf; struct filedesc *fdp; struct vnode *rdir; int error; if (disablefullpath) return (ENODEV); if (vn == NULL) return (EINVAL); buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); fdp = td->td_proc->p_fd; FILEDESC_SLOCK(fdp); rdir = fdp->fd_rdir; VREF(rdir); FILEDESC_SUNLOCK(fdp); error = vn_fullpath1(td, vn, rdir, buf, retbuf, MAXPATHLEN); vrele(rdir); if (!error) *freebuf = buf; else free(buf, M_TEMP); return (error); } /* * This function is similar to vn_fullpath, but it attempts to lookup the * pathname relative to the global root mount point. This is required for the * auditing sub-system, as audited pathnames must be absolute, relative to the * global root mount point. */ int vn_fullpath_global(struct thread *td, struct vnode *vn, char **retbuf, char **freebuf) { char *buf; int error; if (disablefullpath) return (ENODEV); if (vn == NULL) return (EINVAL); buf = malloc(MAXPATHLEN, M_TEMP, M_WAITOK); error = vn_fullpath1(td, vn, rootvnode, buf, retbuf, MAXPATHLEN); if (!error) *freebuf = buf; else free(buf, M_TEMP); return (error); } int vn_vptocnp(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen) { int error; CACHE_RLOCK(); error = vn_vptocnp_locked(vp, cred, buf, buflen); if (error == 0) CACHE_RUNLOCK(); return (error); } static int vn_vptocnp_locked(struct vnode **vp, struct ucred *cred, char *buf, u_int *buflen) { struct vnode *dvp; struct namecache *ncp; int error; TAILQ_FOREACH(ncp, &((*vp)->v_cache_dst), nc_dst) { if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) break; } if (ncp != NULL) { if (*buflen < ncp->nc_nlen) { CACHE_RUNLOCK(); vrele(*vp); counter_u64_add(numfullpathfail4, 1); error = ENOMEM; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); return (error); } *buflen -= ncp->nc_nlen; memcpy(buf + *buflen, nc_get_name(ncp), ncp->nc_nlen); SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp, nc_get_name(ncp), vp); dvp = *vp; *vp = ncp->nc_dvp; vref(*vp); CACHE_RUNLOCK(); vrele(dvp); CACHE_RLOCK(); return (0); } SDT_PROBE1(vfs, namecache, fullpath, miss, vp); CACHE_RUNLOCK(); vn_lock(*vp, LK_SHARED | LK_RETRY); error = VOP_VPTOCNP(*vp, &dvp, cred, buf, buflen); vput(*vp); if (error) { counter_u64_add(numfullpathfail2, 1); SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); return (error); } *vp = dvp; CACHE_RLOCK(); if (dvp->v_iflag & VI_DOOMED) { /* forced unmount */ CACHE_RUNLOCK(); vrele(dvp); error = ENOENT; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); return (error); } /* * *vp has its use count incremented still. */ return (0); } /* * The magic behind kern___getcwd() and vn_fullpath(). */ static int vn_fullpath1(struct thread *td, struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf, u_int buflen) { int error, slash_prefixed; #ifdef KDTRACE_HOOKS struct vnode *startvp = vp; #endif struct vnode *vp1; buflen--; buf[buflen] = '\0'; error = 0; slash_prefixed = 0; SDT_PROBE1(vfs, namecache, fullpath, entry, vp); counter_u64_add(numfullpathcalls, 1); vref(vp); CACHE_RLOCK(); if (vp->v_type != VDIR) { error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen); if (error) return (error); if (buflen == 0) { CACHE_RUNLOCK(); vrele(vp); return (ENOMEM); } buf[--buflen] = '/'; slash_prefixed = 1; } while (vp != rdir && vp != rootvnode) { if (vp->v_vflag & VV_ROOT) { if (vp->v_iflag & VI_DOOMED) { /* forced unmount */ CACHE_RUNLOCK(); vrele(vp); error = ENOENT; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); break; } vp1 = vp->v_mount->mnt_vnodecovered; vref(vp1); CACHE_RUNLOCK(); vrele(vp); vp = vp1; CACHE_RLOCK(); continue; } if (vp->v_type != VDIR) { CACHE_RUNLOCK(); vrele(vp); counter_u64_add(numfullpathfail1, 1); error = ENOTDIR; SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL); break; } error = vn_vptocnp_locked(&vp, td->td_ucred, buf, &buflen); if (error) break; if (buflen == 0) { CACHE_RUNLOCK(); vrele(vp); error = ENOMEM; SDT_PROBE3(vfs, namecache, fullpath, return, error, startvp, NULL); break; } buf[--buflen] = '/'; slash_prefixed = 1; } if (error) return (error); if (!slash_prefixed) { if (buflen == 0) { CACHE_RUNLOCK(); vrele(vp); counter_u64_add(numfullpathfail4, 1); SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM, startvp, NULL); return (ENOMEM); } buf[--buflen] = '/'; } counter_u64_add(numfullpathfound, 1); CACHE_RUNLOCK(); vrele(vp); SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, buf + buflen); *retbuf = buf + buflen; return (0); } struct vnode * vn_dir_dd_ino(struct vnode *vp) { struct namecache *ncp; struct vnode *ddvp; ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino"); CACHE_RLOCK(); TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) { if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) continue; ddvp = ncp->nc_dvp; vhold(ddvp); CACHE_RUNLOCK(); if (vget(ddvp, LK_SHARED | LK_NOWAIT | LK_VNHELD, curthread)) return (NULL); return (ddvp); } CACHE_RUNLOCK(); return (NULL); } int vn_commname(struct vnode *vp, char *buf, u_int buflen) { struct namecache *ncp; int l; CACHE_RLOCK(); TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) if ((ncp->nc_flag & NCF_ISDOTDOT) == 0) break; if (ncp == NULL) { CACHE_RUNLOCK(); return (ENOENT); } l = min(ncp->nc_nlen, buflen - 1); memcpy(buf, nc_get_name(ncp), l); CACHE_RUNLOCK(); buf[l] = '\0'; return (0); } /* ABI compat shims for old kernel modules. */ #undef cache_enter void cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp); void cache_enter(struct vnode *dvp, struct vnode *vp, struct componentname *cnp) { cache_enter_time(dvp, vp, cnp, NULL, NULL); } /* * This function updates path string to vnode's full global path * and checks the size of the new path string against the pathlen argument. * * Requires a locked, referenced vnode. * Vnode is re-locked on success or ENODEV, otherwise unlocked. * * If sysctl debug.disablefullpath is set, ENODEV is returned, * vnode is left locked and path remain untouched. * * If vp is a directory, the call to vn_fullpath_global() always succeeds * because it falls back to the ".." lookup if the namecache lookup fails. */ int vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path, u_int pathlen) { struct nameidata nd; struct vnode *vp1; char *rpath, *fbuf; int error; ASSERT_VOP_ELOCKED(vp, __func__); /* Return ENODEV if sysctl debug.disablefullpath==1 */ if (disablefullpath) return (ENODEV); /* Construct global filesystem path from vp. */ VOP_UNLOCK(vp, 0); error = vn_fullpath_global(td, vp, &rpath, &fbuf); if (error != 0) { vrele(vp); return (error); } if (strlen(rpath) >= pathlen) { vrele(vp); error = ENAMETOOLONG; goto out; } /* * Re-lookup the vnode by path to detect a possible rename. * As a side effect, the vnode is relocked. * If vnode was renamed, return ENOENT. */ NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path, td); error = namei(&nd); if (error != 0) { vrele(vp); goto out; } NDFREE(&nd, NDF_ONLY_PNBUF); vp1 = nd.ni_vp; vrele(vp); if (vp1 == vp) strcpy(path, rpath); else { vput(vp1); error = ENOENT; } out: free(fbuf, M_TEMP); return (error); } Index: projects/mips64-clang/sys/powerpc/booke/pmap.c =================================================================== --- projects/mips64-clang/sys/powerpc/booke/pmap.c (revision 305711) +++ projects/mips64-clang/sys/powerpc/booke/pmap.c (revision 305712) @@ -1,3601 +1,3607 @@ /*- * Copyright (C) 2007-2009 Semihalf, Rafal Jaworowski * Copyright (C) 2006 Semihalf, Marian Balakowicz * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN * NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * Some hw specific parts of this pmap were derived or influenced * by NetBSD's ibm4xx pmap module. More generic code is shared with * a few other pmap modules from the FreeBSD tree. */ /* * VM layout notes: * * Kernel and user threads run within one common virtual address space * defined by AS=0. * * Virtual address space layout: * ----------------------------- * 0x0000_0000 - 0xafff_ffff : user process * 0xb000_0000 - 0xbfff_ffff : pmap_mapdev()-ed area (PCI/PCIE etc.) * 0xc000_0000 - 0xc0ff_ffff : kernel reserved * 0xc000_0000 - data_end : kernel code+data, env, metadata etc. * 0xc100_0000 - 0xfeef_ffff : KVA * 0xc100_0000 - 0xc100_3fff : reserved for page zero/copy * 0xc100_4000 - 0xc200_3fff : reserved for ptbl bufs * 0xc200_4000 - 0xc200_8fff : guard page + kstack0 * 0xc200_9000 - 0xfeef_ffff : actual free KVA space * 0xfef0_0000 - 0xffff_ffff : I/O devices region */ #include __FBSDID("$FreeBSD$"); #include "opt_kstack_pages.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "mmu_if.h" #define SPARSE_MAPDEV #ifdef DEBUG #define debugf(fmt, args...) printf(fmt, ##args) #else #define debugf(fmt, args...) #endif #define TODO panic("%s: not implemented", __func__); extern unsigned char _etext[]; extern unsigned char _end[]; extern uint32_t *bootinfo; vm_paddr_t kernload; vm_offset_t kernstart; vm_size_t kernsize; /* Message buffer and tables. */ static vm_offset_t data_start; static vm_size_t data_end; /* Phys/avail memory regions. */ static struct mem_region *availmem_regions; static int availmem_regions_sz; static struct mem_region *physmem_regions; static int physmem_regions_sz; /* Reserved KVA space and mutex for mmu_booke_zero_page. */ static vm_offset_t zero_page_va; static struct mtx zero_page_mutex; static struct mtx tlbivax_mutex; /* Reserved KVA space and mutex for mmu_booke_copy_page. */ static vm_offset_t copy_page_src_va; static vm_offset_t copy_page_dst_va; static struct mtx copy_page_mutex; /**************************************************************************/ /* PMAP */ /**************************************************************************/ static int mmu_booke_enter_locked(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int flags, int8_t psind); unsigned int kptbl_min; /* Index of the first kernel ptbl. */ unsigned int kernel_ptbls; /* Number of KVA ptbls. */ /* * If user pmap is processed with mmu_booke_remove and the resident count * drops to 0, there are no more pages to remove, so we need not continue. */ #define PMAP_REMOVE_DONE(pmap) \ ((pmap) != kernel_pmap && (pmap)->pm_stats.resident_count == 0) extern int elf32_nxstack; /**************************************************************************/ /* TLB and TID handling */ /**************************************************************************/ /* Translation ID busy table */ static volatile pmap_t tidbusy[MAXCPU][TID_MAX + 1]; /* * TLB0 capabilities (entry, way numbers etc.). These can vary between e500 * core revisions and should be read from h/w registers during early config. */ uint32_t tlb0_entries; uint32_t tlb0_ways; uint32_t tlb0_entries_per_way; uint32_t tlb1_entries; #define TLB0_ENTRIES (tlb0_entries) #define TLB0_WAYS (tlb0_ways) #define TLB0_ENTRIES_PER_WAY (tlb0_entries_per_way) #define TLB1_ENTRIES (tlb1_entries) #define TLB1_MAXENTRIES 64 static vm_offset_t tlb1_map_base = VM_MAXUSER_ADDRESS + PAGE_SIZE; static tlbtid_t tid_alloc(struct pmap *); static void tid_flush(tlbtid_t tid); static void tlb_print_entry(int, uint32_t, uint32_t, uint32_t, uint32_t); static void tlb1_read_entry(tlb_entry_t *, unsigned int); static void tlb1_write_entry(tlb_entry_t *, unsigned int); static int tlb1_iomapped(int, vm_paddr_t, vm_size_t, vm_offset_t *); static vm_size_t tlb1_mapin_region(vm_offset_t, vm_paddr_t, vm_size_t); static vm_size_t tsize2size(unsigned int); static unsigned int size2tsize(vm_size_t); static unsigned int ilog2(unsigned int); static void set_mas4_defaults(void); static inline void tlb0_flush_entry(vm_offset_t); static inline unsigned int tlb0_tableidx(vm_offset_t, unsigned int); /**************************************************************************/ /* Page table management */ /**************************************************************************/ static struct rwlock_padalign pvh_global_lock; /* Data for the pv entry allocation mechanism */ static uma_zone_t pvzone; static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0; #define PV_ENTRY_ZONE_MIN 2048 /* min pv entries in uma zone */ #ifndef PMAP_SHPGPERPROC #define PMAP_SHPGPERPROC 200 #endif static void ptbl_init(void); static struct ptbl_buf *ptbl_buf_alloc(void); static void ptbl_buf_free(struct ptbl_buf *); static void ptbl_free_pmap_ptbl(pmap_t, pte_t *); static pte_t *ptbl_alloc(mmu_t, pmap_t, unsigned int, boolean_t); static void ptbl_free(mmu_t, pmap_t, unsigned int); static void ptbl_hold(mmu_t, pmap_t, unsigned int); static int ptbl_unhold(mmu_t, pmap_t, unsigned int); static vm_paddr_t pte_vatopa(mmu_t, pmap_t, vm_offset_t); static pte_t *pte_find(mmu_t, pmap_t, vm_offset_t); static int pte_enter(mmu_t, pmap_t, vm_page_t, vm_offset_t, uint32_t, boolean_t); static int pte_remove(mmu_t, pmap_t, vm_offset_t, uint8_t); static void kernel_pte_alloc(vm_offset_t data_end, vm_offset_t addr, vm_offset_t pdir); static pv_entry_t pv_alloc(void); static void pv_free(pv_entry_t); static void pv_insert(pmap_t, vm_offset_t, vm_page_t); static void pv_remove(pmap_t, vm_offset_t, vm_page_t); static void booke_pmap_init_qpages(void); /* Number of kva ptbl buffers, each covering one ptbl (PTBL_PAGES). */ #define PTBL_BUFS (128 * 16) struct ptbl_buf { TAILQ_ENTRY(ptbl_buf) link; /* list link */ vm_offset_t kva; /* va of mapping */ }; /* ptbl free list and a lock used for access synchronization. */ static TAILQ_HEAD(, ptbl_buf) ptbl_buf_freelist; static struct mtx ptbl_buf_freelist_lock; /* Base address of kva space allocated fot ptbl bufs. */ static vm_offset_t ptbl_buf_pool_vabase; /* Pointer to ptbl_buf structures. */ static struct ptbl_buf *ptbl_bufs; #ifdef SMP extern tlb_entry_t __boot_tlb1[]; void pmap_bootstrap_ap(volatile uint32_t *); #endif /* * Kernel MMU interface */ static void mmu_booke_clear_modify(mmu_t, vm_page_t); static void mmu_booke_copy(mmu_t, pmap_t, pmap_t, vm_offset_t, vm_size_t, vm_offset_t); static void mmu_booke_copy_page(mmu_t, vm_page_t, vm_page_t); static void mmu_booke_copy_pages(mmu_t, vm_page_t *, vm_offset_t, vm_page_t *, vm_offset_t, int); static int mmu_booke_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int flags, int8_t psind); static void mmu_booke_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t, vm_prot_t); static void mmu_booke_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t); static vm_paddr_t mmu_booke_extract(mmu_t, pmap_t, vm_offset_t); static vm_page_t mmu_booke_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t); static void mmu_booke_init(mmu_t); static boolean_t mmu_booke_is_modified(mmu_t, vm_page_t); static boolean_t mmu_booke_is_prefaultable(mmu_t, pmap_t, vm_offset_t); static boolean_t mmu_booke_is_referenced(mmu_t, vm_page_t); static int mmu_booke_ts_referenced(mmu_t, vm_page_t); static vm_offset_t mmu_booke_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int); static int mmu_booke_mincore(mmu_t, pmap_t, vm_offset_t, vm_paddr_t *); static void mmu_booke_object_init_pt(mmu_t, pmap_t, vm_offset_t, vm_object_t, vm_pindex_t, vm_size_t); static boolean_t mmu_booke_page_exists_quick(mmu_t, pmap_t, vm_page_t); static void mmu_booke_page_init(mmu_t, vm_page_t); static int mmu_booke_page_wired_mappings(mmu_t, vm_page_t); static void mmu_booke_pinit(mmu_t, pmap_t); static void mmu_booke_pinit0(mmu_t, pmap_t); static void mmu_booke_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t); static void mmu_booke_qenter(mmu_t, vm_offset_t, vm_page_t *, int); static void mmu_booke_qremove(mmu_t, vm_offset_t, int); static void mmu_booke_release(mmu_t, pmap_t); static void mmu_booke_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t); static void mmu_booke_remove_all(mmu_t, vm_page_t); static void mmu_booke_remove_write(mmu_t, vm_page_t); static void mmu_booke_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t); static void mmu_booke_zero_page(mmu_t, vm_page_t); static void mmu_booke_zero_page_area(mmu_t, vm_page_t, int, int); static void mmu_booke_activate(mmu_t, struct thread *); static void mmu_booke_deactivate(mmu_t, struct thread *); static void mmu_booke_bootstrap(mmu_t, vm_offset_t, vm_offset_t); static void *mmu_booke_mapdev(mmu_t, vm_paddr_t, vm_size_t); static void *mmu_booke_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t); static void mmu_booke_unmapdev(mmu_t, vm_offset_t, vm_size_t); static vm_paddr_t mmu_booke_kextract(mmu_t, vm_offset_t); static void mmu_booke_kenter(mmu_t, vm_offset_t, vm_paddr_t); static void mmu_booke_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t); static void mmu_booke_kremove(mmu_t, vm_offset_t); static boolean_t mmu_booke_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t); static void mmu_booke_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t); static void mmu_booke_dumpsys_map(mmu_t, vm_paddr_t pa, size_t, void **); static void mmu_booke_dumpsys_unmap(mmu_t, vm_paddr_t pa, size_t, void *); static void mmu_booke_scan_init(mmu_t); static vm_offset_t mmu_booke_quick_enter_page(mmu_t mmu, vm_page_t m); static void mmu_booke_quick_remove_page(mmu_t mmu, vm_offset_t addr); static int mmu_booke_change_attr(mmu_t mmu, vm_offset_t addr, vm_size_t sz, vm_memattr_t mode); static mmu_method_t mmu_booke_methods[] = { /* pmap dispatcher interface */ MMUMETHOD(mmu_clear_modify, mmu_booke_clear_modify), MMUMETHOD(mmu_copy, mmu_booke_copy), MMUMETHOD(mmu_copy_page, mmu_booke_copy_page), MMUMETHOD(mmu_copy_pages, mmu_booke_copy_pages), MMUMETHOD(mmu_enter, mmu_booke_enter), MMUMETHOD(mmu_enter_object, mmu_booke_enter_object), MMUMETHOD(mmu_enter_quick, mmu_booke_enter_quick), MMUMETHOD(mmu_extract, mmu_booke_extract), MMUMETHOD(mmu_extract_and_hold, mmu_booke_extract_and_hold), MMUMETHOD(mmu_init, mmu_booke_init), MMUMETHOD(mmu_is_modified, mmu_booke_is_modified), MMUMETHOD(mmu_is_prefaultable, mmu_booke_is_prefaultable), MMUMETHOD(mmu_is_referenced, mmu_booke_is_referenced), MMUMETHOD(mmu_ts_referenced, mmu_booke_ts_referenced), MMUMETHOD(mmu_map, mmu_booke_map), MMUMETHOD(mmu_mincore, mmu_booke_mincore), MMUMETHOD(mmu_object_init_pt, mmu_booke_object_init_pt), MMUMETHOD(mmu_page_exists_quick,mmu_booke_page_exists_quick), MMUMETHOD(mmu_page_init, mmu_booke_page_init), MMUMETHOD(mmu_page_wired_mappings, mmu_booke_page_wired_mappings), MMUMETHOD(mmu_pinit, mmu_booke_pinit), MMUMETHOD(mmu_pinit0, mmu_booke_pinit0), MMUMETHOD(mmu_protect, mmu_booke_protect), MMUMETHOD(mmu_qenter, mmu_booke_qenter), MMUMETHOD(mmu_qremove, mmu_booke_qremove), MMUMETHOD(mmu_release, mmu_booke_release), MMUMETHOD(mmu_remove, mmu_booke_remove), MMUMETHOD(mmu_remove_all, mmu_booke_remove_all), MMUMETHOD(mmu_remove_write, mmu_booke_remove_write), MMUMETHOD(mmu_sync_icache, mmu_booke_sync_icache), MMUMETHOD(mmu_unwire, mmu_booke_unwire), MMUMETHOD(mmu_zero_page, mmu_booke_zero_page), MMUMETHOD(mmu_zero_page_area, mmu_booke_zero_page_area), MMUMETHOD(mmu_activate, mmu_booke_activate), MMUMETHOD(mmu_deactivate, mmu_booke_deactivate), MMUMETHOD(mmu_quick_enter_page, mmu_booke_quick_enter_page), MMUMETHOD(mmu_quick_remove_page, mmu_booke_quick_remove_page), /* Internal interfaces */ MMUMETHOD(mmu_bootstrap, mmu_booke_bootstrap), MMUMETHOD(mmu_dev_direct_mapped,mmu_booke_dev_direct_mapped), MMUMETHOD(mmu_mapdev, mmu_booke_mapdev), MMUMETHOD(mmu_mapdev_attr, mmu_booke_mapdev_attr), MMUMETHOD(mmu_kenter, mmu_booke_kenter), MMUMETHOD(mmu_kenter_attr, mmu_booke_kenter_attr), MMUMETHOD(mmu_kextract, mmu_booke_kextract), MMUMETHOD(mmu_kremove, mmu_booke_kremove), MMUMETHOD(mmu_unmapdev, mmu_booke_unmapdev), MMUMETHOD(mmu_change_attr, mmu_booke_change_attr), /* dumpsys() support */ MMUMETHOD(mmu_dumpsys_map, mmu_booke_dumpsys_map), MMUMETHOD(mmu_dumpsys_unmap, mmu_booke_dumpsys_unmap), MMUMETHOD(mmu_scan_init, mmu_booke_scan_init), { 0, 0 } }; MMU_DEF(booke_mmu, MMU_TYPE_BOOKE, mmu_booke_methods, 0); static __inline uint32_t tlb_calc_wimg(vm_paddr_t pa, vm_memattr_t ma) { uint32_t attrib; int i; if (ma != VM_MEMATTR_DEFAULT) { switch (ma) { case VM_MEMATTR_UNCACHEABLE: return (MAS2_I | MAS2_G); case VM_MEMATTR_WRITE_COMBINING: case VM_MEMATTR_WRITE_BACK: case VM_MEMATTR_PREFETCHABLE: return (MAS2_I); case VM_MEMATTR_WRITE_THROUGH: return (MAS2_W | MAS2_M); case VM_MEMATTR_CACHEABLE: return (MAS2_M); } } /* * Assume the page is cache inhibited and access is guarded unless * it's in our available memory array. */ attrib = _TLB_ENTRY_IO; for (i = 0; i < physmem_regions_sz; i++) { if ((pa >= physmem_regions[i].mr_start) && (pa < (physmem_regions[i].mr_start + physmem_regions[i].mr_size))) { attrib = _TLB_ENTRY_MEM; break; } } return (attrib); } static inline void tlb_miss_lock(void) { #ifdef SMP struct pcpu *pc; if (!smp_started) return; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc != pcpup) { CTR3(KTR_PMAP, "%s: tlb miss LOCK of CPU=%d, " "tlb_lock=%p", __func__, pc->pc_cpuid, pc->pc_booke_tlb_lock); KASSERT((pc->pc_cpuid != PCPU_GET(cpuid)), ("tlb_miss_lock: tried to lock self")); tlb_lock(pc->pc_booke_tlb_lock); CTR1(KTR_PMAP, "%s: locked", __func__); } } #endif } static inline void tlb_miss_unlock(void) { #ifdef SMP struct pcpu *pc; if (!smp_started) return; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc != pcpup) { CTR2(KTR_PMAP, "%s: tlb miss UNLOCK of CPU=%d", __func__, pc->pc_cpuid); tlb_unlock(pc->pc_booke_tlb_lock); CTR1(KTR_PMAP, "%s: unlocked", __func__); } } #endif } /* Return number of entries in TLB0. */ static __inline void tlb0_get_tlbconf(void) { uint32_t tlb0_cfg; tlb0_cfg = mfspr(SPR_TLB0CFG); tlb0_entries = tlb0_cfg & TLBCFG_NENTRY_MASK; tlb0_ways = (tlb0_cfg & TLBCFG_ASSOC_MASK) >> TLBCFG_ASSOC_SHIFT; tlb0_entries_per_way = tlb0_entries / tlb0_ways; } /* Return number of entries in TLB1. */ static __inline void tlb1_get_tlbconf(void) { uint32_t tlb1_cfg; tlb1_cfg = mfspr(SPR_TLB1CFG); tlb1_entries = tlb1_cfg & TLBCFG_NENTRY_MASK; } /**************************************************************************/ /* Page table related */ /**************************************************************************/ /* Initialize pool of kva ptbl buffers. */ static void ptbl_init(void) { int i; CTR3(KTR_PMAP, "%s: s (ptbl_bufs = 0x%08x size 0x%08x)", __func__, (uint32_t)ptbl_bufs, sizeof(struct ptbl_buf) * PTBL_BUFS); CTR3(KTR_PMAP, "%s: s (ptbl_buf_pool_vabase = 0x%08x size = 0x%08x)", __func__, ptbl_buf_pool_vabase, PTBL_BUFS * PTBL_PAGES * PAGE_SIZE); mtx_init(&ptbl_buf_freelist_lock, "ptbl bufs lock", NULL, MTX_DEF); TAILQ_INIT(&ptbl_buf_freelist); for (i = 0; i < PTBL_BUFS; i++) { ptbl_bufs[i].kva = ptbl_buf_pool_vabase + i * PTBL_PAGES * PAGE_SIZE; TAILQ_INSERT_TAIL(&ptbl_buf_freelist, &ptbl_bufs[i], link); } } /* Get a ptbl_buf from the freelist. */ static struct ptbl_buf * ptbl_buf_alloc(void) { struct ptbl_buf *buf; mtx_lock(&ptbl_buf_freelist_lock); buf = TAILQ_FIRST(&ptbl_buf_freelist); if (buf != NULL) TAILQ_REMOVE(&ptbl_buf_freelist, buf, link); mtx_unlock(&ptbl_buf_freelist_lock); CTR2(KTR_PMAP, "%s: buf = %p", __func__, buf); return (buf); } /* Return ptbl buff to free pool. */ static void ptbl_buf_free(struct ptbl_buf *buf) { CTR2(KTR_PMAP, "%s: buf = %p", __func__, buf); mtx_lock(&ptbl_buf_freelist_lock); TAILQ_INSERT_TAIL(&ptbl_buf_freelist, buf, link); mtx_unlock(&ptbl_buf_freelist_lock); } /* * Search the list of allocated ptbl bufs and find on list of allocated ptbls */ static void ptbl_free_pmap_ptbl(pmap_t pmap, pte_t *ptbl) { struct ptbl_buf *pbuf; CTR2(KTR_PMAP, "%s: ptbl = %p", __func__, ptbl); PMAP_LOCK_ASSERT(pmap, MA_OWNED); TAILQ_FOREACH(pbuf, &pmap->pm_ptbl_list, link) if (pbuf->kva == (vm_offset_t)ptbl) { /* Remove from pmap ptbl buf list. */ TAILQ_REMOVE(&pmap->pm_ptbl_list, pbuf, link); /* Free corresponding ptbl buf. */ ptbl_buf_free(pbuf); break; } } /* Allocate page table. */ static pte_t * ptbl_alloc(mmu_t mmu, pmap_t pmap, unsigned int pdir_idx, boolean_t nosleep) { vm_page_t mtbl[PTBL_PAGES]; vm_page_t m; struct ptbl_buf *pbuf; unsigned int pidx; pte_t *ptbl; int i, j; CTR4(KTR_PMAP, "%s: pmap = %p su = %d pdir_idx = %d", __func__, pmap, (pmap == kernel_pmap), pdir_idx); KASSERT((pdir_idx <= (VM_MAXUSER_ADDRESS / PDIR_SIZE)), ("ptbl_alloc: invalid pdir_idx")); KASSERT((pmap->pm_pdir[pdir_idx] == NULL), ("pte_alloc: valid ptbl entry exists!")); pbuf = ptbl_buf_alloc(); if (pbuf == NULL) panic("pte_alloc: couldn't alloc kernel virtual memory"); ptbl = (pte_t *)pbuf->kva; CTR2(KTR_PMAP, "%s: ptbl kva = %p", __func__, ptbl); /* Allocate ptbl pages, this will sleep! */ for (i = 0; i < PTBL_PAGES; i++) { pidx = (PTBL_PAGES * pdir_idx) + i; while ((m = vm_page_alloc(NULL, pidx, VM_ALLOC_NOOBJ | VM_ALLOC_WIRED)) == NULL) { PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); if (nosleep) { ptbl_free_pmap_ptbl(pmap, ptbl); for (j = 0; j < i; j++) vm_page_free(mtbl[j]); atomic_subtract_int(&vm_cnt.v_wire_count, i); return (NULL); } VM_WAIT; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); } mtbl[i] = m; } /* Map allocated pages into kernel_pmap. */ mmu_booke_qenter(mmu, (vm_offset_t)ptbl, mtbl, PTBL_PAGES); /* Zero whole ptbl. */ bzero((caddr_t)ptbl, PTBL_PAGES * PAGE_SIZE); /* Add pbuf to the pmap ptbl bufs list. */ TAILQ_INSERT_TAIL(&pmap->pm_ptbl_list, pbuf, link); return (ptbl); } /* Free ptbl pages and invalidate pdir entry. */ static void ptbl_free(mmu_t mmu, pmap_t pmap, unsigned int pdir_idx) { pte_t *ptbl; vm_paddr_t pa; vm_offset_t va; vm_page_t m; int i; CTR4(KTR_PMAP, "%s: pmap = %p su = %d pdir_idx = %d", __func__, pmap, (pmap == kernel_pmap), pdir_idx); KASSERT((pdir_idx <= (VM_MAXUSER_ADDRESS / PDIR_SIZE)), ("ptbl_free: invalid pdir_idx")); ptbl = pmap->pm_pdir[pdir_idx]; CTR2(KTR_PMAP, "%s: ptbl = %p", __func__, ptbl); KASSERT((ptbl != NULL), ("ptbl_free: null ptbl")); /* * Invalidate the pdir entry as soon as possible, so that other CPUs * don't attempt to look up the page tables we are releasing. */ mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); pmap->pm_pdir[pdir_idx] = NULL; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); for (i = 0; i < PTBL_PAGES; i++) { va = ((vm_offset_t)ptbl + (i * PAGE_SIZE)); pa = pte_vatopa(mmu, kernel_pmap, va); m = PHYS_TO_VM_PAGE(pa); vm_page_free_zero(m); atomic_subtract_int(&vm_cnt.v_wire_count, 1); mmu_booke_kremove(mmu, va); } ptbl_free_pmap_ptbl(pmap, ptbl); } /* * Decrement ptbl pages hold count and attempt to free ptbl pages. * Called when removing pte entry from ptbl. * * Return 1 if ptbl pages were freed. */ static int ptbl_unhold(mmu_t mmu, pmap_t pmap, unsigned int pdir_idx) { pte_t *ptbl; vm_paddr_t pa; vm_page_t m; int i; CTR4(KTR_PMAP, "%s: pmap = %p su = %d pdir_idx = %d", __func__, pmap, (pmap == kernel_pmap), pdir_idx); KASSERT((pdir_idx <= (VM_MAXUSER_ADDRESS / PDIR_SIZE)), ("ptbl_unhold: invalid pdir_idx")); KASSERT((pmap != kernel_pmap), ("ptbl_unhold: unholding kernel ptbl!")); ptbl = pmap->pm_pdir[pdir_idx]; //debugf("ptbl_unhold: ptbl = 0x%08x\n", (u_int32_t)ptbl); KASSERT(((vm_offset_t)ptbl >= VM_MIN_KERNEL_ADDRESS), ("ptbl_unhold: non kva ptbl")); /* decrement hold count */ for (i = 0; i < PTBL_PAGES; i++) { pa = pte_vatopa(mmu, kernel_pmap, (vm_offset_t)ptbl + (i * PAGE_SIZE)); m = PHYS_TO_VM_PAGE(pa); m->wire_count--; } /* * Free ptbl pages if there are no pte etries in this ptbl. * wire_count has the same value for all ptbl pages, so check the last * page. */ if (m->wire_count == 0) { ptbl_free(mmu, pmap, pdir_idx); //debugf("ptbl_unhold: e (freed ptbl)\n"); return (1); } return (0); } /* * Increment hold count for ptbl pages. This routine is used when a new pte * entry is being inserted into the ptbl. */ static void ptbl_hold(mmu_t mmu, pmap_t pmap, unsigned int pdir_idx) { vm_paddr_t pa; pte_t *ptbl; vm_page_t m; int i; CTR3(KTR_PMAP, "%s: pmap = %p pdir_idx = %d", __func__, pmap, pdir_idx); KASSERT((pdir_idx <= (VM_MAXUSER_ADDRESS / PDIR_SIZE)), ("ptbl_hold: invalid pdir_idx")); KASSERT((pmap != kernel_pmap), ("ptbl_hold: holding kernel ptbl!")); ptbl = pmap->pm_pdir[pdir_idx]; KASSERT((ptbl != NULL), ("ptbl_hold: null ptbl")); for (i = 0; i < PTBL_PAGES; i++) { pa = pte_vatopa(mmu, kernel_pmap, (vm_offset_t)ptbl + (i * PAGE_SIZE)); m = PHYS_TO_VM_PAGE(pa); m->wire_count++; } } /* Allocate pv_entry structure. */ pv_entry_t pv_alloc(void) { pv_entry_t pv; pv_entry_count++; if (pv_entry_count > pv_entry_high_water) pagedaemon_wakeup(); pv = uma_zalloc(pvzone, M_NOWAIT); return (pv); } /* Free pv_entry structure. */ static __inline void pv_free(pv_entry_t pve) { pv_entry_count--; uma_zfree(pvzone, pve); } /* Allocate and initialize pv_entry structure. */ static void pv_insert(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pve; //int su = (pmap == kernel_pmap); //debugf("pv_insert: s (su = %d pmap = 0x%08x va = 0x%08x m = 0x%08x)\n", su, // (u_int32_t)pmap, va, (u_int32_t)m); pve = pv_alloc(); if (pve == NULL) panic("pv_insert: no pv entries!"); pve->pv_pmap = pmap; pve->pv_va = va; /* add to pv_list */ PMAP_LOCK_ASSERT(pmap, MA_OWNED); rw_assert(&pvh_global_lock, RA_WLOCKED); TAILQ_INSERT_TAIL(&m->md.pv_list, pve, pv_link); //debugf("pv_insert: e\n"); } /* Destroy pv entry. */ static void pv_remove(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pve; //int su = (pmap == kernel_pmap); //debugf("pv_remove: s (su = %d pmap = 0x%08x va = 0x%08x)\n", su, (u_int32_t)pmap, va); PMAP_LOCK_ASSERT(pmap, MA_OWNED); rw_assert(&pvh_global_lock, RA_WLOCKED); /* find pv entry */ TAILQ_FOREACH(pve, &m->md.pv_list, pv_link) { if ((pmap == pve->pv_pmap) && (va == pve->pv_va)) { /* remove from pv_list */ TAILQ_REMOVE(&m->md.pv_list, pve, pv_link); if (TAILQ_EMPTY(&m->md.pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); /* free pv entry struct */ pv_free(pve); break; } } //debugf("pv_remove: e\n"); } /* * Clean pte entry, try to free page table page if requested. * * Return 1 if ptbl pages were freed, otherwise return 0. */ static int pte_remove(mmu_t mmu, pmap_t pmap, vm_offset_t va, uint8_t flags) { unsigned int pdir_idx = PDIR_IDX(va); unsigned int ptbl_idx = PTBL_IDX(va); vm_page_t m; pte_t *ptbl; pte_t *pte; //int su = (pmap == kernel_pmap); //debugf("pte_remove: s (su = %d pmap = 0x%08x va = 0x%08x flags = %d)\n", // su, (u_int32_t)pmap, va, flags); ptbl = pmap->pm_pdir[pdir_idx]; KASSERT(ptbl, ("pte_remove: null ptbl")); pte = &ptbl[ptbl_idx]; if (pte == NULL || !PTE_ISVALID(pte)) return (0); if (PTE_ISWIRED(pte)) pmap->pm_stats.wired_count--; /* Handle managed entry. */ if (PTE_ISMANAGED(pte)) { /* Get vm_page_t for mapped pte. */ m = PHYS_TO_VM_PAGE(PTE_PA(pte)); if (PTE_ISMODIFIED(pte)) vm_page_dirty(m); if (PTE_ISREFERENCED(pte)) vm_page_aflag_set(m, PGA_REFERENCED); pv_remove(pmap, va, m); } mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); tlb0_flush_entry(va); *pte = 0; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); pmap->pm_stats.resident_count--; if (flags & PTBL_UNHOLD) { //debugf("pte_remove: e (unhold)\n"); return (ptbl_unhold(mmu, pmap, pdir_idx)); } //debugf("pte_remove: e\n"); return (0); } /* * Insert PTE for a given page and virtual address. */ static int pte_enter(mmu_t mmu, pmap_t pmap, vm_page_t m, vm_offset_t va, uint32_t flags, boolean_t nosleep) { unsigned int pdir_idx = PDIR_IDX(va); unsigned int ptbl_idx = PTBL_IDX(va); pte_t *ptbl, *pte; CTR4(KTR_PMAP, "%s: su = %d pmap = %p va = %p", __func__, pmap == kernel_pmap, pmap, va); /* Get the page table pointer. */ ptbl = pmap->pm_pdir[pdir_idx]; if (ptbl == NULL) { /* Allocate page table pages. */ ptbl = ptbl_alloc(mmu, pmap, pdir_idx, nosleep); if (ptbl == NULL) { KASSERT(nosleep, ("nosleep and NULL ptbl")); return (ENOMEM); } } else { /* * Check if there is valid mapping for requested * va, if there is, remove it. */ pte = &pmap->pm_pdir[pdir_idx][ptbl_idx]; if (PTE_ISVALID(pte)) { pte_remove(mmu, pmap, va, PTBL_HOLD); } else { /* * pte is not used, increment hold count * for ptbl pages. */ if (pmap != kernel_pmap) ptbl_hold(mmu, pmap, pdir_idx); } } /* * Insert pv_entry into pv_list for mapped page if part of managed * memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { flags |= PTE_MANAGED; /* Create and insert pv entry. */ pv_insert(pmap, va, m); } pmap->pm_stats.resident_count++; mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); tlb0_flush_entry(va); if (pmap->pm_pdir[pdir_idx] == NULL) { /* * If we just allocated a new page table, hook it in * the pdir. */ pmap->pm_pdir[pdir_idx] = ptbl; } pte = &(pmap->pm_pdir[pdir_idx][ptbl_idx]); *pte = PTE_RPN_FROM_PA(VM_PAGE_TO_PHYS(m)); *pte |= (PTE_VALID | flags | PTE_PS_4KB); /* 4KB pages only */ tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); return (0); } /* Return the pa for the given pmap/va. */ static vm_paddr_t pte_vatopa(mmu_t mmu, pmap_t pmap, vm_offset_t va) { vm_paddr_t pa = 0; pte_t *pte; pte = pte_find(mmu, pmap, va); if ((pte != NULL) && PTE_ISVALID(pte)) pa = (PTE_PA(pte) | (va & PTE_PA_MASK)); return (pa); } /* Get a pointer to a PTE in a page table. */ static pte_t * pte_find(mmu_t mmu, pmap_t pmap, vm_offset_t va) { unsigned int pdir_idx = PDIR_IDX(va); unsigned int ptbl_idx = PTBL_IDX(va); KASSERT((pmap != NULL), ("pte_find: invalid pmap")); if (pmap->pm_pdir[pdir_idx]) return (&(pmap->pm_pdir[pdir_idx][ptbl_idx])); return (NULL); } /* Set up kernel page tables. */ static void kernel_pte_alloc(vm_offset_t data_end, vm_offset_t addr, vm_offset_t pdir) { int i; vm_offset_t va; pte_t *pte; /* Initialize kernel pdir */ for (i = 0; i < kernel_ptbls; i++) kernel_pmap->pm_pdir[kptbl_min + i] = (pte_t *)(pdir + (i * PAGE_SIZE * PTBL_PAGES)); /* * Fill in PTEs covering kernel code and data. They are not required * for address translation, as this area is covered by static TLB1 * entries, but for pte_vatopa() to work correctly with kernel area * addresses. */ for (va = addr; va < data_end; va += PAGE_SIZE) { pte = &(kernel_pmap->pm_pdir[PDIR_IDX(va)][PTBL_IDX(va)]); *pte = PTE_RPN_FROM_PA(kernload + (va - kernstart)); *pte |= PTE_M | PTE_SR | PTE_SW | PTE_SX | PTE_WIRED | PTE_VALID | PTE_PS_4KB; } } /**************************************************************************/ /* PMAP related */ /**************************************************************************/ /* * This is called during booke_init, before the system is really initialized. */ static void mmu_booke_bootstrap(mmu_t mmu, vm_offset_t start, vm_offset_t kernelend) { vm_paddr_t phys_kernelend; struct mem_region *mp, *mp1; int cnt, i, j; vm_paddr_t s, e, sz; vm_paddr_t physsz, hwphyssz; u_int phys_avail_count; vm_size_t kstack0_sz; vm_offset_t kernel_pdir, kstack0; vm_paddr_t kstack0_phys; void *dpcpu; debugf("mmu_booke_bootstrap: entered\n"); /* Set interesting system properties */ hw_direct_map = 0; elf32_nxstack = 1; /* Initialize invalidation mutex */ mtx_init(&tlbivax_mutex, "tlbivax", NULL, MTX_SPIN); /* Read TLB0 size and associativity. */ tlb0_get_tlbconf(); /* * Align kernel start and end address (kernel image). * Note that kernel end does not necessarily relate to kernsize. * kernsize is the size of the kernel that is actually mapped. */ kernstart = trunc_page(start); data_start = round_page(kernelend); data_end = data_start; /* * Addresses of preloaded modules (like file systems) use * physical addresses. Make sure we relocate those into * virtual addresses. */ preload_addr_relocate = kernstart - kernload; /* Allocate the dynamic per-cpu area. */ dpcpu = (void *)data_end; data_end += DPCPU_SIZE; /* Allocate space for the message buffer. */ msgbufp = (struct msgbuf *)data_end; data_end += msgbufsize; debugf(" msgbufp at 0x%08x end = 0x%08x\n", (uint32_t)msgbufp, data_end); data_end = round_page(data_end); /* Allocate space for ptbl_bufs. */ ptbl_bufs = (struct ptbl_buf *)data_end; data_end += sizeof(struct ptbl_buf) * PTBL_BUFS; debugf(" ptbl_bufs at 0x%08x end = 0x%08x\n", (uint32_t)ptbl_bufs, data_end); data_end = round_page(data_end); /* Allocate PTE tables for kernel KVA. */ kernel_pdir = data_end; kernel_ptbls = howmany(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS, PDIR_SIZE); data_end += kernel_ptbls * PTBL_PAGES * PAGE_SIZE; debugf(" kernel ptbls: %d\n", kernel_ptbls); debugf(" kernel pdir at 0x%08x end = 0x%08x\n", kernel_pdir, data_end); debugf(" data_end: 0x%08x\n", data_end); if (data_end - kernstart > kernsize) { kernsize += tlb1_mapin_region(kernstart + kernsize, kernload + kernsize, (data_end - kernstart) - kernsize); } data_end = kernstart + kernsize; debugf(" updated data_end: 0x%08x\n", data_end); /* * Clear the structures - note we can only do it safely after the * possible additional TLB1 translations are in place (above) so that * all range up to the currently calculated 'data_end' is covered. */ dpcpu_init(dpcpu, 0); memset((void *)ptbl_bufs, 0, sizeof(struct ptbl_buf) * PTBL_SIZE); memset((void *)kernel_pdir, 0, kernel_ptbls * PTBL_PAGES * PAGE_SIZE); /*******************************************************/ /* Set the start and end of kva. */ /*******************************************************/ virtual_avail = round_page(data_end); virtual_end = VM_MAX_KERNEL_ADDRESS; /* Allocate KVA space for page zero/copy operations. */ zero_page_va = virtual_avail; virtual_avail += PAGE_SIZE; copy_page_src_va = virtual_avail; virtual_avail += PAGE_SIZE; copy_page_dst_va = virtual_avail; virtual_avail += PAGE_SIZE; debugf("zero_page_va = 0x%08x\n", zero_page_va); debugf("copy_page_src_va = 0x%08x\n", copy_page_src_va); debugf("copy_page_dst_va = 0x%08x\n", copy_page_dst_va); /* Initialize page zero/copy mutexes. */ mtx_init(&zero_page_mutex, "mmu_booke_zero_page", NULL, MTX_DEF); mtx_init(©_page_mutex, "mmu_booke_copy_page", NULL, MTX_DEF); /* Allocate KVA space for ptbl bufs. */ ptbl_buf_pool_vabase = virtual_avail; virtual_avail += PTBL_BUFS * PTBL_PAGES * PAGE_SIZE; debugf("ptbl_buf_pool_vabase = 0x%08x end = 0x%08x\n", ptbl_buf_pool_vabase, virtual_avail); /* Calculate corresponding physical addresses for the kernel region. */ phys_kernelend = kernload + kernsize; debugf("kernel image and allocated data:\n"); debugf(" kernload = 0x%09llx\n", (uint64_t)kernload); debugf(" kernstart = 0x%08x\n", kernstart); debugf(" kernsize = 0x%08x\n", kernsize); if (sizeof(phys_avail) / sizeof(phys_avail[0]) < availmem_regions_sz) panic("mmu_booke_bootstrap: phys_avail too small"); /* * Remove kernel physical address range from avail regions list. Page * align all regions. Non-page aligned memory isn't very interesting * to us. Also, sort the entries for ascending addresses. */ /* Retrieve phys/avail mem regions */ mem_regions(&physmem_regions, &physmem_regions_sz, &availmem_regions, &availmem_regions_sz); sz = 0; cnt = availmem_regions_sz; debugf("processing avail regions:\n"); for (mp = availmem_regions; mp->mr_size; mp++) { s = mp->mr_start; e = mp->mr_start + mp->mr_size; debugf(" %09jx-%09jx -> ", (uintmax_t)s, (uintmax_t)e); /* Check whether this region holds all of the kernel. */ if (s < kernload && e > phys_kernelend) { availmem_regions[cnt].mr_start = phys_kernelend; availmem_regions[cnt++].mr_size = e - phys_kernelend; e = kernload; } /* Look whether this regions starts within the kernel. */ if (s >= kernload && s < phys_kernelend) { if (e <= phys_kernelend) goto empty; s = phys_kernelend; } /* Now look whether this region ends within the kernel. */ if (e > kernload && e <= phys_kernelend) { if (s >= kernload) goto empty; e = kernload; } /* Now page align the start and size of the region. */ s = round_page(s); e = trunc_page(e); if (e < s) e = s; sz = e - s; debugf("%09jx-%09jx = %jx\n", (uintmax_t)s, (uintmax_t)e, (uintmax_t)sz); /* Check whether some memory is left here. */ if (sz == 0) { empty: memmove(mp, mp + 1, (cnt - (mp - availmem_regions)) * sizeof(*mp)); cnt--; mp--; continue; } /* Do an insertion sort. */ for (mp1 = availmem_regions; mp1 < mp; mp1++) if (s < mp1->mr_start) break; if (mp1 < mp) { memmove(mp1 + 1, mp1, (char *)mp - (char *)mp1); mp1->mr_start = s; mp1->mr_size = sz; } else { mp->mr_start = s; mp->mr_size = sz; } } availmem_regions_sz = cnt; /*******************************************************/ /* Steal physical memory for kernel stack from the end */ /* of the first avail region */ /*******************************************************/ kstack0_sz = kstack_pages * PAGE_SIZE; kstack0_phys = availmem_regions[0].mr_start + availmem_regions[0].mr_size; kstack0_phys -= kstack0_sz; availmem_regions[0].mr_size -= kstack0_sz; /*******************************************************/ /* Fill in phys_avail table, based on availmem_regions */ /*******************************************************/ phys_avail_count = 0; physsz = 0; hwphyssz = 0; TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz); debugf("fill in phys_avail:\n"); for (i = 0, j = 0; i < availmem_regions_sz; i++, j += 2) { debugf(" region: 0x%jx - 0x%jx (0x%jx)\n", (uintmax_t)availmem_regions[i].mr_start, (uintmax_t)availmem_regions[i].mr_start + availmem_regions[i].mr_size, (uintmax_t)availmem_regions[i].mr_size); if (hwphyssz != 0 && (physsz + availmem_regions[i].mr_size) >= hwphyssz) { debugf(" hw.physmem adjust\n"); if (physsz < hwphyssz) { phys_avail[j] = availmem_regions[i].mr_start; phys_avail[j + 1] = availmem_regions[i].mr_start + hwphyssz - physsz; physsz = hwphyssz; phys_avail_count++; } break; } phys_avail[j] = availmem_regions[i].mr_start; phys_avail[j + 1] = availmem_regions[i].mr_start + availmem_regions[i].mr_size; phys_avail_count++; physsz += availmem_regions[i].mr_size; } physmem = btoc(physsz); /* Calculate the last available physical address. */ for (i = 0; phys_avail[i + 2] != 0; i += 2) ; Maxmem = powerpc_btop(phys_avail[i + 1]); debugf("Maxmem = 0x%08lx\n", Maxmem); debugf("phys_avail_count = %d\n", phys_avail_count); debugf("physsz = 0x%09jx physmem = %jd (0x%09jx)\n", (uintmax_t)physsz, (uintmax_t)physmem, (uintmax_t)physmem); /*******************************************************/ /* Initialize (statically allocated) kernel pmap. */ /*******************************************************/ PMAP_LOCK_INIT(kernel_pmap); kptbl_min = VM_MIN_KERNEL_ADDRESS / PDIR_SIZE; debugf("kernel_pmap = 0x%08x\n", (uint32_t)kernel_pmap); debugf("kptbl_min = %d, kernel_ptbls = %d\n", kptbl_min, kernel_ptbls); debugf("kernel pdir range: 0x%08x - 0x%08x\n", kptbl_min * PDIR_SIZE, (kptbl_min + kernel_ptbls) * PDIR_SIZE - 1); kernel_pte_alloc(data_end, kernstart, kernel_pdir); for (i = 0; i < MAXCPU; i++) { kernel_pmap->pm_tid[i] = TID_KERNEL; /* Initialize each CPU's tidbusy entry 0 with kernel_pmap */ tidbusy[i][TID_KERNEL] = kernel_pmap; } /* Mark kernel_pmap active on all CPUs */ CPU_FILL(&kernel_pmap->pm_active); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); /*******************************************************/ /* Final setup */ /*******************************************************/ /* Enter kstack0 into kernel map, provide guard page */ kstack0 = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE; thread0.td_kstack = kstack0; thread0.td_kstack_pages = kstack_pages; debugf("kstack_sz = 0x%08x\n", kstack0_sz); debugf("kstack0_phys at 0x%09llx - 0x%09llx\n", kstack0_phys, kstack0_phys + kstack0_sz); debugf("kstack0 at 0x%08x - 0x%08x\n", kstack0, kstack0 + kstack0_sz); virtual_avail += KSTACK_GUARD_PAGES * PAGE_SIZE + kstack0_sz; for (i = 0; i < kstack_pages; i++) { mmu_booke_kenter(mmu, kstack0, kstack0_phys); kstack0 += PAGE_SIZE; kstack0_phys += PAGE_SIZE; } pmap_bootstrapped = 1; debugf("virtual_avail = %08x\n", virtual_avail); debugf("virtual_end = %08x\n", virtual_end); debugf("mmu_booke_bootstrap: exit\n"); } #ifdef SMP void tlb1_ap_prep(void) { tlb_entry_t *e, tmp; unsigned int i; /* Prepare TLB1 image for AP processors */ e = __boot_tlb1; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&tmp, i); if ((tmp.mas1 & MAS1_VALID) && (tmp.mas2 & _TLB_ENTRY_SHARED)) memcpy(e++, &tmp, sizeof(tmp)); } } void pmap_bootstrap_ap(volatile uint32_t *trcp __unused) { int i; /* * Finish TLB1 configuration: the BSP already set up its TLB1 and we * have the snapshot of its contents in the s/w __boot_tlb1[] table * created by tlb1_ap_prep(), so use these values directly to * (re)program AP's TLB1 hardware. * * Start at index 1 because index 0 has the kernel map. */ for (i = 1; i < TLB1_ENTRIES; i++) { if (__boot_tlb1[i].mas1 & MAS1_VALID) tlb1_write_entry(&__boot_tlb1[i], i); } set_mas4_defaults(); } #endif static void booke_pmap_init_qpages(void) { struct pcpu *pc; int i; CPU_FOREACH(i) { pc = pcpu_find(i); pc->pc_qmap_addr = kva_alloc(PAGE_SIZE); if (pc->pc_qmap_addr == 0) panic("pmap_init_qpages: unable to allocate KVA"); } } SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, booke_pmap_init_qpages, NULL); /* * Get the physical page address for the given pmap/virtual address. */ static vm_paddr_t mmu_booke_extract(mmu_t mmu, pmap_t pmap, vm_offset_t va) { vm_paddr_t pa; PMAP_LOCK(pmap); pa = pte_vatopa(mmu, pmap, va); PMAP_UNLOCK(pmap); return (pa); } /* * Extract the physical page address associated with the given * kernel virtual address. */ static vm_paddr_t mmu_booke_kextract(mmu_t mmu, vm_offset_t va) { tlb_entry_t e; int i; /* Check TLB1 mappings */ for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (va >= e.virt && va < e.virt + e.size) return (e.phys + (va - e.virt)); } return (pte_vatopa(mmu, kernel_pmap, va)); } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ static void mmu_booke_init(mmu_t mmu) { int shpgperproc = PMAP_SHPGPERPROC; /* * Initialize the address space (zone) for the pv entries. Set a * high water mark so that the system can recover from excessive * numbers of pv entries. */ pvzone = uma_zcreate("PV ENTRY", sizeof(struct pv_entry), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE); TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count; TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); pv_entry_high_water = 9 * (pv_entry_max / 10); uma_zone_reserve_kva(pvzone, pv_entry_max); /* Pre-fill pvzone with initial number of pv entries. */ uma_prealloc(pvzone, PV_ENTRY_ZONE_MIN); /* Initialize ptbl allocation. */ ptbl_init(); } /* * Map a list of wired pages into kernel virtual address space. This is * intended for temporary mappings which do not need page modification or * references recorded. Existing mappings in the region are overwritten. */ static void mmu_booke_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count) { vm_offset_t va; va = sva; while (count-- > 0) { mmu_booke_kenter(mmu, va, VM_PAGE_TO_PHYS(*m)); va += PAGE_SIZE; m++; } } /* * Remove page mappings from kernel virtual address space. Intended for * temporary mappings entered by mmu_booke_qenter. */ static void mmu_booke_qremove(mmu_t mmu, vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { mmu_booke_kremove(mmu, va); va += PAGE_SIZE; } } /* * Map a wired page into kernel virtual address space. */ static void mmu_booke_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa) { mmu_booke_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT); } static void mmu_booke_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma) { uint32_t flags; pte_t *pte; KASSERT(((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_kenter: invalid va")); flags = PTE_SR | PTE_SW | PTE_SX | PTE_WIRED | PTE_VALID; flags |= tlb_calc_wimg(pa, ma) << PTE_MAS2_SHIFT; flags |= PTE_PS_4KB; pte = pte_find(mmu, kernel_pmap, va); mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); if (PTE_ISVALID(pte)) { CTR1(KTR_PMAP, "%s: replacing entry!", __func__); /* Flush entry from TLB0 */ tlb0_flush_entry(va); } *pte = PTE_RPN_FROM_PA(pa) | flags; //debugf("mmu_booke_kenter: pdir_idx = %d ptbl_idx = %d va=0x%08x " // "pa=0x%08x rpn=0x%08x flags=0x%08x\n", // pdir_idx, ptbl_idx, va, pa, pte->rpn, pte->flags); /* Flush the real memory from the instruction cache. */ if ((flags & (PTE_I | PTE_G)) == 0) __syncicache((void *)va, PAGE_SIZE); tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } /* * Remove a page from kernel page table. */ static void mmu_booke_kremove(mmu_t mmu, vm_offset_t va) { pte_t *pte; CTR2(KTR_PMAP,"%s: s (va = 0x%08x)\n", __func__, va); KASSERT(((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_kremove: invalid va")); pte = pte_find(mmu, kernel_pmap, va); if (!PTE_ISVALID(pte)) { CTR1(KTR_PMAP, "%s: invalid pte", __func__); return; } mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); /* Invalidate entry in TLB0, update PTE. */ tlb0_flush_entry(va); *pte = 0; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } /* * Initialize pmap associated with process 0. */ static void mmu_booke_pinit0(mmu_t mmu, pmap_t pmap) { PMAP_LOCK_INIT(pmap); mmu_booke_pinit(mmu, pmap); PCPU_SET(curpmap, pmap); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ static void mmu_booke_pinit(mmu_t mmu, pmap_t pmap) { int i; CTR4(KTR_PMAP, "%s: pmap = %p, proc %d '%s'", __func__, pmap, curthread->td_proc->p_pid, curthread->td_proc->p_comm); KASSERT((pmap != kernel_pmap), ("pmap_pinit: initializing kernel_pmap")); for (i = 0; i < MAXCPU; i++) pmap->pm_tid[i] = TID_NONE; CPU_ZERO(&kernel_pmap->pm_active); bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); bzero(&pmap->pm_pdir, sizeof(pte_t *) * PDIR_NENTRIES); TAILQ_INIT(&pmap->pm_ptbl_list); } /* * Release any resources held by the given physical map. * Called when a pmap initialized by mmu_booke_pinit is being released. * Should only be called if the map contains no valid mappings. */ static void mmu_booke_release(mmu_t mmu, pmap_t pmap) { KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); } /* * Insert the given physical page at the specified virtual address in the * target physical map with the protection requested. If specified the page * will be wired down. */ static int mmu_booke_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { int error; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); error = mmu_booke_enter_locked(mmu, pmap, va, m, prot, flags, psind); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (error); } static int mmu_booke_enter_locked(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int pmap_flags, int8_t psind __unused) { pte_t *pte; vm_paddr_t pa; uint32_t flags; int error, su, sync; pa = VM_PAGE_TO_PHYS(m); su = (pmap == kernel_pmap); sync = 0; //debugf("mmu_booke_enter_locked: s (pmap=0x%08x su=%d tid=%d m=0x%08x va=0x%08x " // "pa=0x%08x prot=0x%08x flags=%#x)\n", // (u_int32_t)pmap, su, pmap->pm_tid, // (u_int32_t)m, va, pa, prot, flags); if (su) { KASSERT(((va >= virtual_avail) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_enter_locked: kernel pmap, non kernel va")); } else { KASSERT((va <= VM_MAXUSER_ADDRESS), ("mmu_booke_enter_locked: user pmap, non user va")); } if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * If there is an existing mapping, and the physical address has not * changed, must be protection or wiring change. */ if (((pte = pte_find(mmu, pmap, va)) != NULL) && (PTE_ISVALID(pte)) && (PTE_PA(pte) == pa)) { /* * Before actually updating pte->flags we calculate and * prepare its new value in a helper var. */ flags = *pte; flags &= ~(PTE_UW | PTE_UX | PTE_SW | PTE_SX | PTE_MODIFIED); /* Wiring change, just update stats. */ if ((pmap_flags & PMAP_ENTER_WIRED) != 0) { if (!PTE_ISWIRED(pte)) { flags |= PTE_WIRED; pmap->pm_stats.wired_count++; } } else { if (PTE_ISWIRED(pte)) { flags &= ~PTE_WIRED; pmap->pm_stats.wired_count--; } } if (prot & VM_PROT_WRITE) { /* Add write permissions. */ flags |= PTE_SW; if (!su) flags |= PTE_UW; if ((flags & PTE_MANAGED) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } else { /* Handle modified pages, sense modify status. */ /* * The PTE_MODIFIED flag could be set by underlying * TLB misses since we last read it (above), possibly * other CPUs could update it so we check in the PTE * directly rather than rely on that saved local flags * copy. */ if (PTE_ISMODIFIED(pte)) vm_page_dirty(m); } if (prot & VM_PROT_EXECUTE) { flags |= PTE_SX; if (!su) flags |= PTE_UX; /* * Check existing flags for execute permissions: if we * are turning execute permissions on, icache should * be flushed. */ if ((*pte & (PTE_UX | PTE_SX)) == 0) sync++; } flags &= ~PTE_REFERENCED; /* * The new flags value is all calculated -- only now actually * update the PTE. */ mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); tlb0_flush_entry(va); *pte &= ~PTE_FLAGS_MASK; *pte |= flags; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } else { /* * If there is an existing mapping, but it's for a different * physical address, pte_enter() will delete the old mapping. */ //if ((pte != NULL) && PTE_ISVALID(pte)) // debugf("mmu_booke_enter_locked: replace\n"); //else // debugf("mmu_booke_enter_locked: new\n"); /* Now set up the flags and install the new mapping. */ flags = (PTE_SR | PTE_VALID); flags |= PTE_M; if (!su) flags |= PTE_UR; if (prot & VM_PROT_WRITE) { flags |= PTE_SW; if (!su) flags |= PTE_UW; if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_WRITEABLE); } if (prot & VM_PROT_EXECUTE) { flags |= PTE_SX; if (!su) flags |= PTE_UX; } /* If its wired update stats. */ if ((pmap_flags & PMAP_ENTER_WIRED) != 0) flags |= PTE_WIRED; error = pte_enter(mmu, pmap, m, va, flags, (pmap_flags & PMAP_ENTER_NOSLEEP) != 0); if (error != 0) return (KERN_RESOURCE_SHORTAGE); if ((flags & PMAP_ENTER_WIRED) != 0) pmap->pm_stats.wired_count++; /* Flush the real memory from the instruction cache. */ if (prot & VM_PROT_EXECUTE) sync++; } if (sync && (su || pmap == PCPU_GET(curpmap))) { __syncicache((void *)va, PAGE_SIZE); sync = 0; } return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ static void mmu_booke_enter_object(mmu_t mmu, pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_page_t m; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); m = m_start; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { mmu_booke_enter_locked(mmu, pmap, start + ptoa(diff), m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP, 0); m = TAILQ_NEXT(m, listq); } rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } static void mmu_booke_enter_quick(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); mmu_booke_enter_locked(mmu, pmap, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP, 0); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly rounded to the page size. */ static void mmu_booke_remove(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_offset_t endva) { pte_t *pte; uint8_t hold_flag; int su = (pmap == kernel_pmap); //debugf("mmu_booke_remove: s (su = %d pmap=0x%08x tid=%d va=0x%08x endva=0x%08x)\n", // su, (u_int32_t)pmap, pmap->pm_tid, va, endva); if (su) { KASSERT(((va >= virtual_avail) && (va <= VM_MAX_KERNEL_ADDRESS)), ("mmu_booke_remove: kernel pmap, non kernel va")); } else { KASSERT((va <= VM_MAXUSER_ADDRESS), ("mmu_booke_remove: user pmap, non user va")); } if (PMAP_REMOVE_DONE(pmap)) { //debugf("mmu_booke_remove: e (empty)\n"); return; } hold_flag = PTBL_HOLD_FLAG(pmap); //debugf("mmu_booke_remove: hold_flag = %d\n", hold_flag); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); for (; va < endva; va += PAGE_SIZE) { pte = pte_find(mmu, pmap, va); if ((pte != NULL) && PTE_ISVALID(pte)) pte_remove(mmu, pmap, va, hold_flag); } PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); //debugf("mmu_booke_remove: e\n"); } /* * Remove physical page from all pmaps in which it resides. */ static void mmu_booke_remove_all(mmu_t mmu, vm_page_t m) { pv_entry_t pv, pvn; uint8_t hold_flag; rw_wlock(&pvh_global_lock); for (pv = TAILQ_FIRST(&m->md.pv_list); pv != NULL; pv = pvn) { pvn = TAILQ_NEXT(pv, pv_link); PMAP_LOCK(pv->pv_pmap); hold_flag = PTBL_HOLD_FLAG(pv->pv_pmap); pte_remove(mmu, pv->pv_pmap, pv->pv_va, hold_flag); PMAP_UNLOCK(pv->pv_pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); } /* * Map a range of physical addresses into kernel virtual address space. */ static vm_offset_t mmu_booke_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start, vm_paddr_t pa_end, int prot) { vm_offset_t sva = *virt; vm_offset_t va = sva; //debugf("mmu_booke_map: s (sva = 0x%08x pa_start = 0x%08x pa_end = 0x%08x)\n", // sva, pa_start, pa_end); while (pa_start < pa_end) { mmu_booke_kenter(mmu, va, pa_start); va += PAGE_SIZE; pa_start += PAGE_SIZE; } *virt = va; //debugf("mmu_booke_map: e (va = 0x%08x)\n", va); return (sva); } /* * The pmap must be activated before it's address space can be accessed in any * way. */ static void mmu_booke_activate(mmu_t mmu, struct thread *td) { pmap_t pmap; u_int cpuid; pmap = &td->td_proc->p_vmspace->vm_pmap; CTR5(KTR_PMAP, "%s: s (td = %p, proc = '%s', id = %d, pmap = 0x%08x)", __func__, td, td->td_proc->p_comm, td->td_proc->p_pid, pmap); KASSERT((pmap != kernel_pmap), ("mmu_booke_activate: kernel_pmap!")); sched_pin(); cpuid = PCPU_GET(cpuid); CPU_SET_ATOMIC(cpuid, &pmap->pm_active); PCPU_SET(curpmap, pmap); if (pmap->pm_tid[cpuid] == TID_NONE) tid_alloc(pmap); /* Load PID0 register with pmap tid value. */ mtspr(SPR_PID0, pmap->pm_tid[cpuid]); __asm __volatile("isync"); mtspr(SPR_DBCR0, td->td_pcb->pcb_cpu.booke.dbcr0); sched_unpin(); CTR3(KTR_PMAP, "%s: e (tid = %d for '%s')", __func__, pmap->pm_tid[PCPU_GET(cpuid)], td->td_proc->p_comm); } /* * Deactivate the specified process's address space. */ static void mmu_booke_deactivate(mmu_t mmu, struct thread *td) { pmap_t pmap; pmap = &td->td_proc->p_vmspace->vm_pmap; CTR5(KTR_PMAP, "%s: td=%p, proc = '%s', id = %d, pmap = 0x%08x", __func__, td, td->td_proc->p_comm, td->td_proc->p_pid, pmap); td->td_pcb->pcb_cpu.booke.dbcr0 = mfspr(SPR_DBCR0); CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmap->pm_active); PCPU_SET(curpmap, NULL); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ static void mmu_booke_copy(mmu_t mmu, pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { } /* * Set the physical protection on the specified range of this map as requested. */ static void mmu_booke_protect(mmu_t mmu, pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t va; vm_page_t m; pte_t *pte; if ((prot & VM_PROT_READ) == VM_PROT_NONE) { mmu_booke_remove(mmu, pmap, sva, eva); return; } if (prot & VM_PROT_WRITE) return; PMAP_LOCK(pmap); for (va = sva; va < eva; va += PAGE_SIZE) { if ((pte = pte_find(mmu, pmap, va)) != NULL) { if (PTE_ISVALID(pte)) { m = PHYS_TO_VM_PAGE(PTE_PA(pte)); mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); /* Handle modified pages. */ if (PTE_ISMODIFIED(pte) && PTE_ISMANAGED(pte)) vm_page_dirty(m); tlb0_flush_entry(va); *pte &= ~(PTE_UW | PTE_SW | PTE_MODIFIED); tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } } } PMAP_UNLOCK(pmap); } /* * Clear the write and modified bits in each of the given page's mappings. */ static void mmu_booke_remove_write(mmu_t mmu, vm_page_t m) { pv_entry_t pv; pte_t *pte; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_remove_write: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(mmu, pv->pv_pmap, pv->pv_va)) != NULL) { if (PTE_ISVALID(pte)) { m = PHYS_TO_VM_PAGE(PTE_PA(pte)); mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); /* Handle modified pages. */ if (PTE_ISMODIFIED(pte)) vm_page_dirty(m); /* Flush mapping from TLB0. */ *pte &= ~(PTE_UW | PTE_SW | PTE_MODIFIED); tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } } PMAP_UNLOCK(pv->pv_pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); } static void mmu_booke_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz) { pte_t *pte; pmap_t pmap; vm_page_t m; vm_offset_t addr; vm_paddr_t pa = 0; int active, valid; va = trunc_page(va); sz = round_page(sz); rw_wlock(&pvh_global_lock); pmap = PCPU_GET(curpmap); active = (pm == kernel_pmap || pm == pmap) ? 1 : 0; while (sz > 0) { PMAP_LOCK(pm); pte = pte_find(mmu, pm, va); valid = (pte != NULL && PTE_ISVALID(pte)) ? 1 : 0; if (valid) pa = PTE_PA(pte); PMAP_UNLOCK(pm); if (valid) { if (!active) { /* Create a mapping in the active pmap. */ addr = 0; m = PHYS_TO_VM_PAGE(pa); PMAP_LOCK(pmap); pte_enter(mmu, pmap, m, addr, PTE_SR | PTE_VALID | PTE_UR, FALSE); __syncicache((void *)addr, PAGE_SIZE); pte_remove(mmu, pmap, addr, PTBL_UNHOLD); PMAP_UNLOCK(pmap); } else __syncicache((void *)va, PAGE_SIZE); } va += PAGE_SIZE; sz -= PAGE_SIZE; } rw_wunlock(&pvh_global_lock); } /* * Atomically extract and hold the physical page with the given * pmap and virtual address pair if that mapping permits the given * protection. */ static vm_page_t mmu_booke_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pte_t *pte; vm_page_t m; uint32_t pte_wbit; vm_paddr_t pa; m = NULL; pa = 0; PMAP_LOCK(pmap); retry: pte = pte_find(mmu, pmap, va); if ((pte != NULL) && PTE_ISVALID(pte)) { if (pmap == kernel_pmap) pte_wbit = PTE_SW; else pte_wbit = PTE_UW; if ((*pte & pte_wbit) || ((prot & VM_PROT_WRITE) == 0)) { if (vm_page_pa_tryrelock(pmap, PTE_PA(pte), &pa)) goto retry; m = PHYS_TO_VM_PAGE(PTE_PA(pte)); vm_page_hold(m); } } PA_UNLOCK_COND(pa); PMAP_UNLOCK(pmap); return (m); } /* * Initialize a vm_page's machine-dependent fields. */ static void mmu_booke_page_init(mmu_t mmu, vm_page_t m) { TAILQ_INIT(&m->md.pv_list); } /* * mmu_booke_zero_page_area zeros the specified hardware page by * mapping it into virtual memory and using bzero to clear * its contents. * * off and size must reside within a single page. */ static void mmu_booke_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size) { vm_offset_t va; /* XXX KASSERT off and size are within a single page? */ mtx_lock(&zero_page_mutex); va = zero_page_va; mmu_booke_kenter(mmu, va, VM_PAGE_TO_PHYS(m)); bzero((caddr_t)va + off, size); mmu_booke_kremove(mmu, va); mtx_unlock(&zero_page_mutex); } /* * mmu_booke_zero_page zeros the specified hardware page. */ static void mmu_booke_zero_page(mmu_t mmu, vm_page_t m) { vm_offset_t off, va; mtx_lock(&zero_page_mutex); va = zero_page_va; mmu_booke_kenter(mmu, va, VM_PAGE_TO_PHYS(m)); for (off = 0; off < PAGE_SIZE; off += cacheline_size) __asm __volatile("dcbz 0,%0" :: "r"(va + off)); mmu_booke_kremove(mmu, va); mtx_unlock(&zero_page_mutex); } /* * mmu_booke_copy_page copies the specified (machine independent) page by * mapping the page into virtual memory and using memcopy to copy the page, * one machine dependent page at a time. */ static void mmu_booke_copy_page(mmu_t mmu, vm_page_t sm, vm_page_t dm) { vm_offset_t sva, dva; sva = copy_page_src_va; dva = copy_page_dst_va; mtx_lock(©_page_mutex); mmu_booke_kenter(mmu, sva, VM_PAGE_TO_PHYS(sm)); mmu_booke_kenter(mmu, dva, VM_PAGE_TO_PHYS(dm)); memcpy((caddr_t)dva, (caddr_t)sva, PAGE_SIZE); mmu_booke_kremove(mmu, dva); mmu_booke_kremove(mmu, sva); mtx_unlock(©_page_mutex); } static inline void mmu_booke_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset, vm_page_t *mb, vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_offset_t a_pg_offset, b_pg_offset; int cnt; mtx_lock(©_page_mutex); while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); mmu_booke_kenter(mmu, copy_page_src_va, VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])); a_cp = (char *)copy_page_src_va + a_pg_offset; b_pg_offset = b_offset & PAGE_MASK; cnt = min(cnt, PAGE_SIZE - b_pg_offset); mmu_booke_kenter(mmu, copy_page_dst_va, VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])); b_cp = (char *)copy_page_dst_va + b_pg_offset; bcopy(a_cp, b_cp, cnt); mmu_booke_kremove(mmu, copy_page_dst_va); mmu_booke_kremove(mmu, copy_page_src_va); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } mtx_unlock(©_page_mutex); } static vm_offset_t mmu_booke_quick_enter_page(mmu_t mmu, vm_page_t m) { vm_paddr_t paddr; vm_offset_t qaddr; uint32_t flags; pte_t *pte; paddr = VM_PAGE_TO_PHYS(m); flags = PTE_SR | PTE_SW | PTE_SX | PTE_WIRED | PTE_VALID; flags |= tlb_calc_wimg(paddr, pmap_page_get_memattr(m)) << PTE_MAS2_SHIFT; flags |= PTE_PS_4KB; critical_enter(); qaddr = PCPU_GET(qmap_addr); pte = pte_find(mmu, kernel_pmap, qaddr); KASSERT(*pte == 0, ("mmu_booke_quick_enter_page: PTE busy")); /* * XXX: tlbivax is broadcast to other cores, but qaddr should * not be present in other TLBs. Is there a better instruction * sequence to use? Or just forget it & use mmu_booke_kenter()... */ __asm __volatile("tlbivax 0, %0" :: "r"(qaddr & MAS2_EPN_MASK)); __asm __volatile("isync; msync"); *pte = PTE_RPN_FROM_PA(paddr) | flags; /* Flush the real memory from the instruction cache. */ if ((flags & (PTE_I | PTE_G)) == 0) __syncicache((void *)qaddr, PAGE_SIZE); return (qaddr); } static void mmu_booke_quick_remove_page(mmu_t mmu, vm_offset_t addr) { pte_t *pte; pte = pte_find(mmu, kernel_pmap, addr); KASSERT(PCPU_GET(qmap_addr) == addr, ("mmu_booke_quick_remove_page: invalid address")); KASSERT(*pte != 0, ("mmu_booke_quick_remove_page: PTE not in use")); *pte = 0; critical_exit(); } /* * Return whether or not the specified physical page was modified * in any of physical maps. */ static boolean_t mmu_booke_is_modified(mmu_t mmu, vm_page_t m) { pte_t *pte; pv_entry_t pv; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_is_modified: page %p is not managed", m)); rv = FALSE; /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no PTEs can be modified. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (rv); rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(mmu, pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { if (PTE_ISMODIFIED(pte)) rv = TRUE; } PMAP_UNLOCK(pv->pv_pmap); if (rv) break; } rw_wunlock(&pvh_global_lock); return (rv); } /* * Return whether or not the specified virtual address is eligible * for prefault. */ static boolean_t mmu_booke_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t addr) { return (FALSE); } /* * Return whether or not the specified physical page was referenced * in any physical maps. */ static boolean_t mmu_booke_is_referenced(mmu_t mmu, vm_page_t m) { pte_t *pte; pv_entry_t pv; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_is_referenced: page %p is not managed", m)); rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(mmu, pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { if (PTE_ISREFERENCED(pte)) rv = TRUE; } PMAP_UNLOCK(pv->pv_pmap); if (rv) break; } rw_wunlock(&pvh_global_lock); return (rv); } /* * Clear the modify bits on the specified physical page. */ static void mmu_booke_clear_modify(mmu_t mmu, vm_page_t m) { pte_t *pte; pv_entry_t pv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_clear_modify: page %p is not managed", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("mmu_booke_clear_modify: page %p is exclusive busied", m)); /* * If the page is not PG_AWRITEABLE, then no PTEs can be modified. * If the object containing the page is locked and the page is not * exclusive busied, then PG_AWRITEABLE cannot be concurrently set. */ if ((m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(mmu, pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); if (*pte & (PTE_SW | PTE_UW | PTE_MODIFIED)) { tlb0_flush_entry(pv->pv_va); *pte &= ~(PTE_SW | PTE_UW | PTE_MODIFIED | PTE_REFERENCED); } tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); } PMAP_UNLOCK(pv->pv_pmap); } rw_wunlock(&pvh_global_lock); } /* * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). */ static int mmu_booke_ts_referenced(mmu_t mmu, vm_page_t m) { pte_t *pte; pv_entry_t pv; int count; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_ts_referenced: page %p is not managed", m)); count = 0; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(mmu, pv->pv_pmap, pv->pv_va)) != NULL && PTE_ISVALID(pte)) { + if (PTE_ISMODIFIED(pte)) + vm_page_dirty(m); if (PTE_ISREFERENCED(pte)) { mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); tlb0_flush_entry(pv->pv_va); *pte &= ~PTE_REFERENCED; tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); - if (++count > 4) { + if (++count >= PMAP_TS_REFERENCED_MAX) { PMAP_UNLOCK(pv->pv_pmap); break; } } } PMAP_UNLOCK(pv->pv_pmap); } rw_wunlock(&pvh_global_lock); return (count); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range must * have the wired attribute set. In contrast, invalid mappings cannot have * the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, so * there is no need to invalidate any TLB entries. */ static void mmu_booke_unwire(mmu_t mmu, pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va; pte_t *pte; PMAP_LOCK(pmap); for (va = sva; va < eva; va += PAGE_SIZE) { if ((pte = pte_find(mmu, pmap, va)) != NULL && PTE_ISVALID(pte)) { if (!PTE_ISWIRED(pte)) panic("mmu_booke_unwire: pte %p isn't wired", pte); *pte &= ~PTE_WIRED; pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Return true if the pmap's pv is one of the first 16 pvs linked to from this * page. This count may be changed upwards or downwards in the future; it is * only necessary that true be returned for a small subset of pmaps for proper * page aging. */ static boolean_t mmu_booke_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m) { pv_entry_t pv; int loops; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("mmu_booke_page_exists_quick: page %p is not managed", m)); loops = 0; rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { if (pv->pv_pmap == pmap) { rv = TRUE; break; } if (++loops >= 16) break; } rw_wunlock(&pvh_global_lock); return (rv); } /* * Return the number of managed mappings to the given physical page that are * wired. */ static int mmu_booke_page_wired_mappings(mmu_t mmu, vm_page_t m) { pv_entry_t pv; pte_t *pte; int count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_link) { PMAP_LOCK(pv->pv_pmap); if ((pte = pte_find(mmu, pv->pv_pmap, pv->pv_va)) != NULL) if (PTE_ISVALID(pte) && PTE_ISWIRED(pte)) count++; PMAP_UNLOCK(pv->pv_pmap); } rw_wunlock(&pvh_global_lock); return (count); } static int mmu_booke_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size) { int i; vm_offset_t va; /* * This currently does not work for entries that * overlap TLB1 entries. */ for (i = 0; i < TLB1_ENTRIES; i ++) { if (tlb1_iomapped(i, pa, size, &va) == 0) return (0); } return (EFAULT); } void mmu_booke_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va) { vm_paddr_t ppa; vm_offset_t ofs; vm_size_t gran; /* Minidumps are based on virtual memory addresses. */ if (do_minidump) { *va = (void *)(vm_offset_t)pa; return; } /* Raw physical memory dumps don't have a virtual address. */ /* We always map a 256MB page at 256M. */ gran = 256 * 1024 * 1024; ppa = rounddown2(pa, gran); ofs = pa - ppa; *va = (void *)gran; tlb1_set_entry((vm_offset_t)va, ppa, gran, _TLB_ENTRY_IO); if (sz > (gran - ofs)) tlb1_set_entry((vm_offset_t)(va + gran), ppa + gran, gran, _TLB_ENTRY_IO); } void mmu_booke_dumpsys_unmap(mmu_t mmu, vm_paddr_t pa, size_t sz, void *va) { vm_paddr_t ppa; vm_offset_t ofs; vm_size_t gran; tlb_entry_t e; int i; /* Minidumps are based on virtual memory addresses. */ /* Nothing to do... */ if (do_minidump) return; for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) break; } /* Raw physical memory dumps don't have a virtual address. */ i--; e.mas1 = 0; e.mas2 = 0; e.mas3 = 0; tlb1_write_entry(&e, i); gran = 256 * 1024 * 1024; ppa = rounddown2(pa, gran); ofs = pa - ppa; if (sz > (gran - ofs)) { i--; e.mas1 = 0; e.mas2 = 0; e.mas3 = 0; tlb1_write_entry(&e, i); } } extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1]; void mmu_booke_scan_init(mmu_t mmu) { vm_offset_t va; pte_t *pte; int i; if (!do_minidump) { /* Initialize phys. segments for dumpsys(). */ memset(&dump_map, 0, sizeof(dump_map)); mem_regions(&physmem_regions, &physmem_regions_sz, &availmem_regions, &availmem_regions_sz); for (i = 0; i < physmem_regions_sz; i++) { dump_map[i].pa_start = physmem_regions[i].mr_start; dump_map[i].pa_size = physmem_regions[i].mr_size; } return; } /* Virtual segments for minidumps: */ memset(&dump_map, 0, sizeof(dump_map)); /* 1st: kernel .data and .bss. */ dump_map[0].pa_start = trunc_page((uintptr_t)_etext); dump_map[0].pa_size = round_page((uintptr_t)_end) - dump_map[0].pa_start; /* 2nd: msgbuf and tables (see pmap_bootstrap()). */ dump_map[1].pa_start = data_start; dump_map[1].pa_size = data_end - data_start; /* 3rd: kernel VM. */ va = dump_map[1].pa_start + dump_map[1].pa_size; /* Find start of next chunk (from va). */ while (va < virtual_end) { /* Don't dump the buffer cache. */ if (va >= kmi.buffer_sva && va < kmi.buffer_eva) { va = kmi.buffer_eva; continue; } pte = pte_find(mmu, kernel_pmap, va); if (pte != NULL && PTE_ISVALID(pte)) break; va += PAGE_SIZE; } if (va < virtual_end) { dump_map[2].pa_start = va; va += PAGE_SIZE; /* Find last page in chunk. */ while (va < virtual_end) { /* Don't run into the buffer cache. */ if (va == kmi.buffer_sva) break; pte = pte_find(mmu, kernel_pmap, va); if (pte == NULL || !PTE_ISVALID(pte)) break; va += PAGE_SIZE; } dump_map[2].pa_size = va - dump_map[2].pa_start; } } /* * Map a set of physical memory pages into the kernel virtual address space. * Return a pointer to where it is mapped. This routine is intended to be used * for mapping device memory, NOT real memory. */ static void * mmu_booke_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size) { return (mmu_booke_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT)); } static void * mmu_booke_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma) { tlb_entry_t e; void *res; uintptr_t va, tmpva; vm_size_t sz; int i; /* * Check if this is premapped in TLB1. Note: this should probably also * check whether a sequence of TLB1 entries exist that match the * requirement, but now only checks the easy case. */ if (ma == VM_MEMATTR_DEFAULT) { for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (pa >= e.phys && (pa + size) <= (e.phys + e.size)) return (void *)(e.virt + (vm_offset_t)(pa - e.phys)); } } size = roundup(size, PAGE_SIZE); /* * The device mapping area is between VM_MAXUSER_ADDRESS and * VM_MIN_KERNEL_ADDRESS. This gives 1GB of device addressing. */ #ifdef SPARSE_MAPDEV /* * With a sparse mapdev, align to the largest starting region. This * could feasibly be optimized for a 'best-fit' alignment, but that * calculation could be very costly. */ do { tmpva = tlb1_map_base; va = roundup(tlb1_map_base, 1 << flsl(size)); } while (!atomic_cmpset_int(&tlb1_map_base, tmpva, va + size)); #else va = atomic_fetchadd_int(&tlb1_map_base, size); #endif res = (void *)va; do { sz = 1 << (ilog2(size) & ~1); if (va % sz != 0) { do { sz >>= 2; } while (va % sz != 0); } if (bootverbose) printf("Wiring VA=%x to PA=%jx (size=%x)\n", va, (uintmax_t)pa, sz); tlb1_set_entry(va, pa, sz, _TLB_ENTRY_SHARED | tlb_calc_wimg(pa, ma)); size -= sz; pa += sz; va += sz; } while (size > 0); return (res); } /* * 'Unmap' a range mapped by mmu_booke_mapdev(). */ static void mmu_booke_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size) { #ifdef SUPPORTS_SHRINKING_TLB1 vm_offset_t base, offset; /* * Unmap only if this is inside kernel virtual space. */ if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)) { base = trunc_page(va); offset = va & PAGE_MASK; size = roundup(offset + size, PAGE_SIZE); kva_free(base, size); } #endif } /* * mmu_booke_object_init_pt preloads the ptes for a given object into the * specified pmap. This eliminates the blast of soft faults on process startup * and immediately after an mmap. */ static void mmu_booke_object_init_pt(mmu_t mmu, pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("mmu_booke_object_init_pt: non-device object")); } /* * Perform the pmap work for mincore. */ static int mmu_booke_mincore(mmu_t mmu, pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa) { /* XXX: this should be implemented at some point */ return (0); } static int mmu_booke_change_attr(mmu_t mmu, vm_offset_t addr, vm_size_t sz, vm_memattr_t mode) { vm_offset_t va; pte_t *pte; int i, j; tlb_entry_t e; /* Check TLB1 mappings */ for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (addr >= e.virt && addr < e.virt + e.size) break; } if (i < TLB1_ENTRIES) { /* Only allow full mappings to be modified for now. */ /* Validate the range. */ for (j = i, va = addr; va < addr + sz; va += e.size, j++) { tlb1_read_entry(&e, j); if (va != e.virt || (sz - (va - addr) < e.size)) return (EINVAL); } for (va = addr; va < addr + sz; va += e.size, i++) { tlb1_read_entry(&e, i); e.mas2 &= ~MAS2_WIMGE_MASK; e.mas2 |= tlb_calc_wimg(e.phys, mode); /* * Write it out to the TLB. Should really re-sync with other * cores. */ tlb1_write_entry(&e, i); } return (0); } /* Not in TLB1, try through pmap */ /* First validate the range. */ for (va = addr; va < addr + sz; va += PAGE_SIZE) { pte = pte_find(mmu, kernel_pmap, va); if (pte == NULL || !PTE_ISVALID(pte)) return (EINVAL); } mtx_lock_spin(&tlbivax_mutex); tlb_miss_lock(); for (va = addr; va < addr + sz; va += PAGE_SIZE) { pte = pte_find(mmu, kernel_pmap, va); *pte &= ~(PTE_MAS2_MASK << PTE_MAS2_SHIFT); *pte |= tlb_calc_wimg(PTE_PA(pte), mode << PTE_MAS2_SHIFT); tlb0_flush_entry(va); } tlb_miss_unlock(); mtx_unlock_spin(&tlbivax_mutex); return (pte_vatopa(mmu, kernel_pmap, va)); } /**************************************************************************/ /* TID handling */ /**************************************************************************/ /* * Allocate a TID. If necessary, steal one from someone else. * The new TID is flushed from the TLB before returning. */ static tlbtid_t tid_alloc(pmap_t pmap) { tlbtid_t tid; int thiscpu; KASSERT((pmap != kernel_pmap), ("tid_alloc: kernel pmap")); CTR2(KTR_PMAP, "%s: s (pmap = %p)", __func__, pmap); thiscpu = PCPU_GET(cpuid); tid = PCPU_GET(tid_next); if (tid > TID_MAX) tid = TID_MIN; PCPU_SET(tid_next, tid + 1); /* If we are stealing TID then clear the relevant pmap's field */ if (tidbusy[thiscpu][tid] != NULL) { CTR2(KTR_PMAP, "%s: warning: stealing tid %d", __func__, tid); tidbusy[thiscpu][tid]->pm_tid[thiscpu] = TID_NONE; /* Flush all entries from TLB0 matching this TID. */ tid_flush(tid); } tidbusy[thiscpu][tid] = pmap; pmap->pm_tid[thiscpu] = tid; __asm __volatile("msync; isync"); CTR3(KTR_PMAP, "%s: e (%02d next = %02d)", __func__, tid, PCPU_GET(tid_next)); return (tid); } /**************************************************************************/ /* TLB0 handling */ /**************************************************************************/ static void tlb_print_entry(int i, uint32_t mas1, uint32_t mas2, uint32_t mas3, uint32_t mas7) { int as; char desc[3]; tlbtid_t tid; vm_size_t size; unsigned int tsize; desc[2] = '\0'; if (mas1 & MAS1_VALID) desc[0] = 'V'; else desc[0] = ' '; if (mas1 & MAS1_IPROT) desc[1] = 'P'; else desc[1] = ' '; as = (mas1 & MAS1_TS_MASK) ? 1 : 0; tid = MAS1_GETTID(mas1); tsize = (mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT; size = 0; if (tsize) size = tsize2size(tsize); debugf("%3d: (%s) [AS=%d] " "sz = 0x%08x tsz = %d tid = %d mas1 = 0x%08x " "mas2(va) = 0x%08x mas3(pa) = 0x%08x mas7 = 0x%08x\n", i, desc, as, size, tsize, tid, mas1, mas2, mas3, mas7); } /* Convert TLB0 va and way number to tlb0[] table index. */ static inline unsigned int tlb0_tableidx(vm_offset_t va, unsigned int way) { unsigned int idx; idx = (way * TLB0_ENTRIES_PER_WAY); idx += (va & MAS2_TLB0_ENTRY_IDX_MASK) >> MAS2_TLB0_ENTRY_IDX_SHIFT; return (idx); } /* * Invalidate TLB0 entry. */ static inline void tlb0_flush_entry(vm_offset_t va) { CTR2(KTR_PMAP, "%s: s va=0x%08x", __func__, va); mtx_assert(&tlbivax_mutex, MA_OWNED); __asm __volatile("tlbivax 0, %0" :: "r"(va & MAS2_EPN_MASK)); __asm __volatile("isync; msync"); __asm __volatile("tlbsync; msync"); CTR1(KTR_PMAP, "%s: e", __func__); } /* Print out contents of the MAS registers for each TLB0 entry */ void tlb0_print_tlbentries(void) { uint32_t mas0, mas1, mas2, mas3, mas7; int entryidx, way, idx; debugf("TLB0 entries:\n"); for (way = 0; way < TLB0_WAYS; way ++) for (entryidx = 0; entryidx < TLB0_ENTRIES_PER_WAY; entryidx++) { mas0 = MAS0_TLBSEL(0) | MAS0_ESEL(way); mtspr(SPR_MAS0, mas0); __asm __volatile("isync"); mas2 = entryidx << MAS2_TLB0_ENTRY_IDX_SHIFT; mtspr(SPR_MAS2, mas2); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); mas2 = mfspr(SPR_MAS2); mas3 = mfspr(SPR_MAS3); mas7 = mfspr(SPR_MAS7); idx = tlb0_tableidx(mas2, way); tlb_print_entry(idx, mas1, mas2, mas3, mas7); } } /**************************************************************************/ /* TLB1 handling */ /**************************************************************************/ /* * TLB1 mapping notes: * * TLB1[0] Kernel text and data. * TLB1[1-15] Additional kernel text and data mappings (if required), PCI * windows, other devices mappings. */ /* * Read an entry from given TLB1 slot. */ void tlb1_read_entry(tlb_entry_t *entry, unsigned int slot) { uint32_t mas0; KASSERT((entry != NULL), ("%s(): Entry is NULL!", __func__)); mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(slot); mtspr(SPR_MAS0, mas0); __asm __volatile("isync; tlbre"); entry->mas1 = mfspr(SPR_MAS1); entry->mas2 = mfspr(SPR_MAS2); entry->mas3 = mfspr(SPR_MAS3); switch ((mfpvr() >> 16) & 0xFFFF) { case FSL_E500v2: case FSL_E500mc: case FSL_E5500: case FSL_E6500: entry->mas7 = mfspr(SPR_MAS7); break; default: entry->mas7 = 0; break; } entry->virt = entry->mas2 & MAS2_EPN_MASK; entry->phys = ((vm_paddr_t)(entry->mas7 & MAS7_RPN) << 32) | (entry->mas3 & MAS3_RPN); entry->size = tsize2size((entry->mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT); } /* * Write given entry to TLB1 hardware. * Use 32 bit pa, clear 4 high-order bits of RPN (mas7). */ static void tlb1_write_entry(tlb_entry_t *e, unsigned int idx) { uint32_t mas0; //debugf("tlb1_write_entry: s\n"); /* Select entry */ mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(idx); //debugf("tlb1_write_entry: mas0 = 0x%08x\n", mas0); mtspr(SPR_MAS0, mas0); __asm __volatile("isync"); mtspr(SPR_MAS1, e->mas1); __asm __volatile("isync"); mtspr(SPR_MAS2, e->mas2); __asm __volatile("isync"); mtspr(SPR_MAS3, e->mas3); __asm __volatile("isync"); switch ((mfpvr() >> 16) & 0xFFFF) { case FSL_E500mc: case FSL_E5500: case FSL_E6500: mtspr(SPR_MAS8, 0); __asm __volatile("isync"); /* FALLTHROUGH */ case FSL_E500v2: mtspr(SPR_MAS7, e->mas7); __asm __volatile("isync"); break; default: break; } __asm __volatile("tlbwe; isync; msync"); //debugf("tlb1_write_entry: e\n"); } /* * Return the largest uint value log such that 2^log <= num. */ static unsigned int ilog2(unsigned int num) { int lz; __asm ("cntlzw %0, %1" : "=r" (lz) : "r" (num)); return (31 - lz); } /* * Convert TLB TSIZE value to mapped region size. */ static vm_size_t tsize2size(unsigned int tsize) { /* * size = 4^tsize KB * size = 4^tsize * 2^10 = 2^(2 * tsize - 10) */ return ((1 << (2 * tsize)) * 1024); } /* * Convert region size (must be power of 4) to TLB TSIZE value. */ static unsigned int size2tsize(vm_size_t size) { return (ilog2(size) / 2 - 5); } /* * Register permanent kernel mapping in TLB1. * * Entries are created starting from index 0 (current free entry is * kept in tlb1_idx) and are not supposed to be invalidated. */ int tlb1_set_entry(vm_offset_t va, vm_paddr_t pa, vm_size_t size, uint32_t flags) { tlb_entry_t e; uint32_t ts, tid; int tsize, index; for (index = 0; index < TLB1_ENTRIES; index++) { tlb1_read_entry(&e, index); if ((e.mas1 & MAS1_VALID) == 0) break; /* Check if we're just updating the flags, and update them. */ if (e.phys == pa && e.virt == va && e.size == size) { e.mas2 = (va & MAS2_EPN_MASK) | flags; tlb1_write_entry(&e, index); return (0); } } if (index >= TLB1_ENTRIES) { printf("tlb1_set_entry: TLB1 full!\n"); return (-1); } /* Convert size to TSIZE */ tsize = size2tsize(size); tid = (TID_KERNEL << MAS1_TID_SHIFT) & MAS1_TID_MASK; /* XXX TS is hard coded to 0 for now as we only use single address space */ ts = (0 << MAS1_TS_SHIFT) & MAS1_TS_MASK; e.phys = pa; e.virt = va; e.size = size; e.mas1 = MAS1_VALID | MAS1_IPROT | ts | tid; e.mas1 |= ((tsize << MAS1_TSIZE_SHIFT) & MAS1_TSIZE_MASK); e.mas2 = (va & MAS2_EPN_MASK) | flags; /* Set supervisor RWX permission bits */ e.mas3 = (pa & MAS3_RPN) | MAS3_SR | MAS3_SW | MAS3_SX; e.mas7 = (pa >> 32) & MAS7_RPN; tlb1_write_entry(&e, index); /* * XXX in general TLB1 updates should be propagated between CPUs, * since current design assumes to have the same TLB1 set-up on all * cores. */ return (0); } /* * Map in contiguous RAM region into the TLB1 using maximum of * KERNEL_REGION_MAX_TLB_ENTRIES entries. * * If necessary round up last entry size and return total size * used by all allocated entries. */ vm_size_t tlb1_mapin_region(vm_offset_t va, vm_paddr_t pa, vm_size_t size) { vm_size_t pgs[KERNEL_REGION_MAX_TLB_ENTRIES]; vm_size_t mapped, pgsz, base, mask; int idx, nents; /* Round up to the next 1M */ size = roundup2(size, 1 << 20); mapped = 0; idx = 0; base = va; pgsz = 64*1024*1024; while (mapped < size) { while (mapped < size && idx < KERNEL_REGION_MAX_TLB_ENTRIES) { while (pgsz > (size - mapped)) pgsz >>= 2; pgs[idx++] = pgsz; mapped += pgsz; } /* We under-map. Correct for this. */ if (mapped < size) { while (pgs[idx - 1] == pgsz) { idx--; mapped -= pgsz; } /* XXX We may increase beyond out starting point. */ pgsz <<= 2; pgs[idx++] = pgsz; mapped += pgsz; } } nents = idx; mask = pgs[0] - 1; /* Align address to the boundary */ if (va & mask) { va = (va + mask) & ~mask; pa = (pa + mask) & ~mask; } for (idx = 0; idx < nents; idx++) { pgsz = pgs[idx]; debugf("%u: %llx -> %x, size=%x\n", idx, pa, va, pgsz); tlb1_set_entry(va, pa, pgsz, _TLB_ENTRY_SHARED | _TLB_ENTRY_MEM); pa += pgsz; va += pgsz; } mapped = (va - base); #ifdef __powerpc64__ printf("mapped size 0x%016lx (wasted space 0x%16lx)\n", #else printf("mapped size 0x%08x (wasted space 0x%08x)\n", #endif mapped, mapped - size); return (mapped); } /* * TLB1 initialization routine, to be called after the very first * assembler level setup done in locore.S. */ void tlb1_init() { uint32_t mas0, mas1, mas2, mas3, mas7; uint32_t tsz; tlb1_get_tlbconf(); mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(0); mtspr(SPR_MAS0, mas0); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); mas2 = mfspr(SPR_MAS2); mas3 = mfspr(SPR_MAS3); mas7 = mfspr(SPR_MAS7); kernload = ((vm_paddr_t)(mas7 & MAS7_RPN) << 32) | (mas3 & MAS3_RPN); tsz = (mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT; kernsize += (tsz > 0) ? tsize2size(tsz) : 0; /* Setup TLB miss defaults */ set_mas4_defaults(); } /* * pmap_early_io_unmap() should be used in short conjunction with * pmap_early_io_map(), as in the following snippet: * * x = pmap_early_io_map(...); * * pmap_early_io_unmap(x, size); * * And avoiding more allocations between. */ void pmap_early_io_unmap(vm_offset_t va, vm_size_t size) { int i; tlb_entry_t e; vm_size_t isize; size = roundup(size, PAGE_SIZE); isize = size; for (i = 0; i < TLB1_ENTRIES && size > 0; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (va <= e.virt && (va + isize) >= (e.virt + e.size)) { size -= e.size; e.mas1 &= ~MAS1_VALID; tlb1_write_entry(&e, i); } } if (tlb1_map_base == va + isize) tlb1_map_base -= isize; } vm_offset_t pmap_early_io_map(vm_paddr_t pa, vm_size_t size) { vm_paddr_t pa_base; vm_offset_t va, sz; int i; tlb_entry_t e; KASSERT(!pmap_bootstrapped, ("Do not use after PMAP is up!")); for (i = 0; i < TLB1_ENTRIES; i++) { tlb1_read_entry(&e, i); if (!(e.mas1 & MAS1_VALID)) continue; if (pa >= e.phys && (pa + size) <= (e.phys + e.size)) return (e.virt + (pa - e.phys)); } pa_base = rounddown(pa, PAGE_SIZE); size = roundup(size + (pa - pa_base), PAGE_SIZE); tlb1_map_base = roundup2(tlb1_map_base, 1 << (ilog2(size) & ~1)); va = tlb1_map_base + (pa - pa_base); do { sz = 1 << (ilog2(size) & ~1); tlb1_set_entry(tlb1_map_base, pa_base, sz, _TLB_ENTRY_SHARED | _TLB_ENTRY_IO); size -= sz; pa_base += sz; tlb1_map_base += sz; } while (size > 0); return (va); } /* * Setup MAS4 defaults. * These values are loaded to MAS0-2 on a TLB miss. */ static void set_mas4_defaults(void) { uint32_t mas4; /* Defaults: TLB0, PID0, TSIZED=4K */ mas4 = MAS4_TLBSELD0; mas4 |= (TLB_SIZE_4K << MAS4_TSIZED_SHIFT) & MAS4_TSIZED_MASK; #ifdef SMP mas4 |= MAS4_MD; #endif mtspr(SPR_MAS4, mas4); __asm __volatile("isync"); } /* * Print out contents of the MAS registers for each TLB1 entry */ void tlb1_print_tlbentries(void) { uint32_t mas0, mas1, mas2, mas3, mas7; int i; debugf("TLB1 entries:\n"); for (i = 0; i < TLB1_ENTRIES; i++) { mas0 = MAS0_TLBSEL(1) | MAS0_ESEL(i); mtspr(SPR_MAS0, mas0); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); mas2 = mfspr(SPR_MAS2); mas3 = mfspr(SPR_MAS3); mas7 = mfspr(SPR_MAS7); tlb_print_entry(i, mas1, mas2, mas3, mas7); } } /* * Return 0 if the physical IO range is encompassed by one of the * the TLB1 entries, otherwise return related error code. */ static int tlb1_iomapped(int i, vm_paddr_t pa, vm_size_t size, vm_offset_t *va) { uint32_t prot; vm_paddr_t pa_start; vm_paddr_t pa_end; unsigned int entry_tsize; vm_size_t entry_size; tlb_entry_t e; *va = (vm_offset_t)NULL; tlb1_read_entry(&e, i); /* Skip invalid entries */ if (!(e.mas1 & MAS1_VALID)) return (EINVAL); /* * The entry must be cache-inhibited, guarded, and r/w * so it can function as an i/o page */ prot = e.mas2 & (MAS2_I | MAS2_G); if (prot != (MAS2_I | MAS2_G)) return (EPERM); prot = e.mas3 & (MAS3_SR | MAS3_SW); if (prot != (MAS3_SR | MAS3_SW)) return (EPERM); /* The address should be within the entry range. */ entry_tsize = (e.mas1 & MAS1_TSIZE_MASK) >> MAS1_TSIZE_SHIFT; KASSERT((entry_tsize), ("tlb1_iomapped: invalid entry tsize")); entry_size = tsize2size(entry_tsize); pa_start = (((vm_paddr_t)e.mas7 & MAS7_RPN) << 32) | (e.mas3 & MAS3_RPN); pa_end = pa_start + entry_size; if ((pa < pa_start) || ((pa + size) > pa_end)) return (ERANGE); /* Return virtual address of this mapping. */ *va = (e.mas2 & MAS2_EPN_MASK) + (pa - pa_start); return (0); } /* * Invalidate all TLB0 entries which match the given TID. Note this is * dedicated for cases when invalidations should NOT be propagated to other * CPUs. */ static void tid_flush(tlbtid_t tid) { register_t msr; uint32_t mas0, mas1, mas2; int entry, way; /* Don't evict kernel translations */ if (tid == TID_KERNEL) return; msr = mfmsr(); __asm __volatile("wrteei 0"); for (way = 0; way < TLB0_WAYS; way++) for (entry = 0; entry < TLB0_ENTRIES_PER_WAY; entry++) { mas0 = MAS0_TLBSEL(0) | MAS0_ESEL(way); mtspr(SPR_MAS0, mas0); __asm __volatile("isync"); mas2 = entry << MAS2_TLB0_ENTRY_IDX_SHIFT; mtspr(SPR_MAS2, mas2); __asm __volatile("isync; tlbre"); mas1 = mfspr(SPR_MAS1); if (!(mas1 & MAS1_VALID)) continue; if (((mas1 & MAS1_TID_MASK) >> MAS1_TID_SHIFT) != tid) continue; mas1 &= ~MAS1_VALID; mtspr(SPR_MAS1, mas1); __asm __volatile("isync; tlbwe; isync; msync"); } mtmsr(msr); } Index: projects/mips64-clang/sys/powerpc/conf/MPC85XX =================================================================== --- projects/mips64-clang/sys/powerpc/conf/MPC85XX (revision 305711) +++ projects/mips64-clang/sys/powerpc/conf/MPC85XX (revision 305712) @@ -1,94 +1,95 @@ # # Custom kernel for Freescale MPC85XX development boards like the CDS etc. # # $FreeBSD$ # cpu BOOKE cpu BOOKE_E500 ident MPC85XX machine powerpc powerpc include "dpaa/config.dpaa" makeoptions DEBUG="-Wa,-me500 -g" makeoptions WERROR="-Werror -Wno-format -Wno-redundant-decls" makeoptions NO_MODULES=yes options FPU_EMU options _KPOSIX_PRIORITY_SCHEDULING options ALT_BREAK_TO_DEBUGGER options BREAK_TO_DEBUGGER options BOOTP options BOOTP_NFSROOT #options BOOTP_NFSV3 options CD9660 options COMPAT_43 options DDB #options DEADLKRES options DEVICE_POLLING #options DIAGNOSTIC options FDT #makeoptions FDT_DTS_FILE=mpc8555cds.dts options FFS options GDB options GEOM_PART_GPT options INET options INET6 options INVARIANTS options INVARIANT_SUPPORT options KDB options KTRACE options MD_ROOT options MPC85XX options MSDOSFS options NFS_ROOT options NFSCL options NFSLOCKD options PROCFS options PSEUDOFS options SCHED_ULE options CAPABILITIES options CAPABILITY_MODE options SMP options SYSVMSG options SYSVSEM options SYSVSHM options WITNESS options WITNESS_SKIPSPIN device ata device bpf device cfi device crypto device cryptodev device da device ds1553 device em device alc device ether device fxp device gpio device iic device iicbus #device isa device loop device md device miibus device pass device pci device quicc device random #device rl device scbus device scc device sec device tsec device tun device uart options USB_DEBUG # enable debug msgs #device uhci +device ehci device umass device usb device vlan Index: projects/mips64-clang/sys/riscv/riscv/pmap.c =================================================================== --- projects/mips64-clang/sys/riscv/riscv/pmap.c (revision 305711) +++ projects/mips64-clang/sys/riscv/riscv/pmap.c (revision 305712) @@ -1,3279 +1,3284 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2003 Peter Wemm * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * Copyright (c) 2014 Andrew Turner * All rights reserved. * Copyright (c) 2014 The FreeBSD Foundation * All rights reserved. * Copyright (c) 2015-2016 Ruslan Bukin * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Portions of this software were developed by Andrew Turner under * sponsorship from The FreeBSD Foundation. * * Portions of this software were developed by SRI International and the * University of Cambridge Computer Laboratory under DARPA/AFRL contract * FA8750-10-C-0237 ("CTSRD"), as part of the DARPA CRASH research programme. * * Portions of this software were developed by the University of Cambridge * Computer Laboratory as part of the CTSRD Project, with support from the * UK Higher Education Innovation Fund (HEIF). * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define NPDEPG (PAGE_SIZE/(sizeof (pd_entry_t))) #define NUPDE (NPDEPG * NPDEPG) #define NUSERPGTBLS (NUPDE + NPDEPG) #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #define pmap_l2_pindex(v) ((v) >> L2_SHIFT) #define NPV_LIST_LOCKS MAXCPU #define PHYS_TO_PV_LIST_LOCK(pa) \ (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS]) #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \ struct rwlock **_lockp = (lockp); \ struct rwlock *_new_lock; \ \ _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \ if (_new_lock != *_lockp) { \ if (*_lockp != NULL) \ rw_wunlock(*_lockp); \ *_lockp = _new_lock; \ rw_wlock(*_lockp); \ } \ } while (0) #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \ CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m)) #define RELEASE_PV_LIST_LOCK(lockp) do { \ struct rwlock **_lockp = (lockp); \ \ if (*_lockp != NULL) { \ rw_wunlock(*_lockp); \ *_lockp = NULL; \ } \ } while (0) #define VM_PAGE_TO_PV_LIST_LOCK(m) \ PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m)) /* The list of all the user pmaps */ LIST_HEAD(pmaplist, pmap); static struct pmaplist allpmaps; static MALLOC_DEFINE(M_VMPMAP, "pmap", "PMAP L1"); struct pmap kernel_pmap_store; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ vm_offset_t kernel_vm_end = 0; struct msgbuf *msgbufp = NULL; vm_paddr_t dmap_phys_base; /* The start of the dmap region */ vm_paddr_t dmap_phys_max; /* The limit of the dmap region */ vm_offset_t dmap_max_addr; /* The virtual address limit of the dmap */ /* This code assumes all L1 DMAP entries will be used */ CTASSERT((DMAP_MIN_ADDRESS & ~L1_OFFSET) == DMAP_MIN_ADDRESS); CTASSERT((DMAP_MAX_ADDRESS & ~L1_OFFSET) == DMAP_MAX_ADDRESS); static struct rwlock_padalign pvh_global_lock; /* * Data for the pv entry allocation mechanism */ static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks); static struct mtx pv_chunks_mutex; static struct rwlock pv_list_locks[NPV_LIST_LOCKS]; static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp); static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp); static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static int pmap_remove_l3(pmap_t pmap, pt_entry_t *l3, vm_offset_t sva, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp); static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp); static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free); static int pmap_unuse_l3(pmap_t, vm_offset_t, pd_entry_t, struct spglist *); /* * These load the old table data and store the new value. * They need to be atomic as the System MMU may write to the table at * the same time as the CPU. */ #define pmap_load_store(table, entry) atomic_swap_64(table, entry) #define pmap_set(table, mask) atomic_set_64(table, mask) #define pmap_load_clear(table) atomic_swap_64(table, 0) #define pmap_load(table) (*table) /********************/ /* Inline functions */ /********************/ static __inline void pagecopy(void *s, void *d) { memcpy(d, s, PAGE_SIZE); } static __inline void pagezero(void *p) { bzero(p, PAGE_SIZE); } #define pmap_l1_index(va) (((va) >> L1_SHIFT) & Ln_ADDR_MASK) #define pmap_l2_index(va) (((va) >> L2_SHIFT) & Ln_ADDR_MASK) #define pmap_l3_index(va) (((va) >> L3_SHIFT) & Ln_ADDR_MASK) #define PTE_TO_PHYS(pte) ((pte >> PTE_PPN0_S) * PAGE_SIZE) static __inline pd_entry_t * pmap_l1(pmap_t pmap, vm_offset_t va) { return (&pmap->pm_l1[pmap_l1_index(va)]); } static __inline pd_entry_t * pmap_l1_to_l2(pd_entry_t *l1, vm_offset_t va) { vm_paddr_t phys; pd_entry_t *l2; phys = PTE_TO_PHYS(pmap_load(l1)); l2 = (pd_entry_t *)PHYS_TO_DMAP(phys); return (&l2[pmap_l2_index(va)]); } static __inline pd_entry_t * pmap_l2(pmap_t pmap, vm_offset_t va) { pd_entry_t *l1; l1 = pmap_l1(pmap, va); if (l1 == NULL) return (NULL); if ((pmap_load(l1) & PTE_V) == 0) return (NULL); if ((pmap_load(l1) & PTE_RX) != 0) return (NULL); return (pmap_l1_to_l2(l1, va)); } static __inline pt_entry_t * pmap_l2_to_l3(pd_entry_t *l2, vm_offset_t va) { vm_paddr_t phys; pt_entry_t *l3; phys = PTE_TO_PHYS(pmap_load(l2)); l3 = (pd_entry_t *)PHYS_TO_DMAP(phys); return (&l3[pmap_l3_index(va)]); } static __inline pt_entry_t * pmap_l3(pmap_t pmap, vm_offset_t va) { pd_entry_t *l2; l2 = pmap_l2(pmap, va); if (l2 == NULL) return (NULL); if ((pmap_load(l2) & PTE_V) == 0) return (NULL); if ((pmap_load(l2) & PTE_RX) != 0) return (NULL); return (pmap_l2_to_l3(l2, va)); } static __inline int pmap_is_write(pt_entry_t entry) { return (entry & PTE_W); } static __inline int pmap_is_current(pmap_t pmap) { return ((pmap == pmap_kernel()) || (pmap == curthread->td_proc->p_vmspace->vm_map.pmap)); } static __inline int pmap_l3_valid(pt_entry_t l3) { return (l3 & PTE_V); } static __inline int pmap_l3_valid_cacheable(pt_entry_t l3) { /* TODO */ return (0); } #define PTE_SYNC(pte) cpu_dcache_wb_range((vm_offset_t)pte, sizeof(*pte)) /* Checks if the page is dirty. */ static inline int pmap_page_dirty(pt_entry_t pte) { return (pte & PTE_D); } static __inline void pmap_resident_count_inc(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); pmap->pm_stats.resident_count += count; } static __inline void pmap_resident_count_dec(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(pmap->pm_stats.resident_count >= count, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count -= count; } static void pmap_distribute_l1(struct pmap *pmap, vm_pindex_t l1index, pt_entry_t entry) { struct pmap *user_pmap; pd_entry_t *l1; /* Distribute new kernel L1 entry to all the user pmaps */ if (pmap != kernel_pmap) return; LIST_FOREACH(user_pmap, &allpmaps, pm_list) { l1 = &user_pmap->pm_l1[l1index]; if (entry) pmap_load_store(l1, entry); else pmap_load_clear(l1); } } static pt_entry_t * pmap_early_page_idx(vm_offset_t l1pt, vm_offset_t va, u_int *l1_slot, u_int *l2_slot) { pt_entry_t *l2; pd_entry_t *l1; l1 = (pd_entry_t *)l1pt; *l1_slot = (va >> L1_SHIFT) & Ln_ADDR_MASK; /* Check locore has used a table L1 map */ KASSERT((l1[*l1_slot] & PTE_RX) == 0, ("Invalid bootstrap L1 table")); /* Find the address of the L2 table */ l2 = (pt_entry_t *)init_pt_va; *l2_slot = pmap_l2_index(va); return (l2); } static vm_paddr_t pmap_early_vtophys(vm_offset_t l1pt, vm_offset_t va) { u_int l1_slot, l2_slot; pt_entry_t *l2; u_int ret; l2 = pmap_early_page_idx(l1pt, va, &l1_slot, &l2_slot); /* Check locore has used L2 superpages */ KASSERT((l2[l2_slot] & PTE_RX) != 0, ("Invalid bootstrap L2 table")); /* L2 is superpages */ ret = (l2[l2_slot] >> PTE_PPN1_S) << L2_SHIFT; ret += (va & L2_OFFSET); return (ret); } static void pmap_bootstrap_dmap(vm_offset_t kern_l1, vm_paddr_t min_pa, vm_paddr_t max_pa) { vm_offset_t va; vm_paddr_t pa; pd_entry_t *l1; u_int l1_slot; pt_entry_t entry; pn_t pn; pa = dmap_phys_base = min_pa & ~L1_OFFSET; va = DMAP_MIN_ADDRESS; l1 = (pd_entry_t *)kern_l1; l1_slot = pmap_l1_index(DMAP_MIN_ADDRESS); for (; va < DMAP_MAX_ADDRESS && pa < max_pa; pa += L1_SIZE, va += L1_SIZE, l1_slot++) { KASSERT(l1_slot < Ln_ENTRIES, ("Invalid L1 index")); /* superpages */ pn = (pa / PAGE_SIZE); entry = (PTE_V | PTE_RWX); entry |= (pn << PTE_PPN0_S); pmap_load_store(&l1[l1_slot], entry); } /* Set the upper limit of the DMAP region */ dmap_phys_max = pa; dmap_max_addr = va; cpu_dcache_wb_range((vm_offset_t)l1, PAGE_SIZE); cpu_tlb_flushID(); } static vm_offset_t pmap_bootstrap_l3(vm_offset_t l1pt, vm_offset_t va, vm_offset_t l3_start) { vm_offset_t l2pt, l3pt; pt_entry_t entry; pd_entry_t *l2; vm_paddr_t pa; u_int l2_slot; pn_t pn; KASSERT((va & L2_OFFSET) == 0, ("Invalid virtual address")); l2 = pmap_l2(kernel_pmap, va); l2 = (pd_entry_t *)((uintptr_t)l2 & ~(PAGE_SIZE - 1)); l2pt = (vm_offset_t)l2; l2_slot = pmap_l2_index(va); l3pt = l3_start; for (; va < VM_MAX_KERNEL_ADDRESS; l2_slot++, va += L2_SIZE) { KASSERT(l2_slot < Ln_ENTRIES, ("Invalid L2 index")); pa = pmap_early_vtophys(l1pt, l3pt); pn = (pa / PAGE_SIZE); entry = (PTE_V); entry |= (pn << PTE_PPN0_S); pmap_load_store(&l2[l2_slot], entry); l3pt += PAGE_SIZE; } /* Clean the L2 page table */ memset((void *)l3_start, 0, l3pt - l3_start); cpu_dcache_wb_range(l3_start, l3pt - l3_start); cpu_dcache_wb_range((vm_offset_t)l2, PAGE_SIZE); return (l3pt); } /* * Bootstrap the system enough to run with virtual memory. */ void pmap_bootstrap(vm_offset_t l1pt, vm_paddr_t kernstart, vm_size_t kernlen) { u_int l1_slot, l2_slot, avail_slot, map_slot, used_map_slot; uint64_t kern_delta; pt_entry_t *l2; vm_offset_t va, freemempos; vm_offset_t dpcpu, msgbufpv; vm_paddr_t pa, min_pa, max_pa; int i; kern_delta = KERNBASE - kernstart; physmem = 0; printf("pmap_bootstrap %lx %lx %lx\n", l1pt, kernstart, kernlen); printf("%lx\n", l1pt); printf("%lx\n", (KERNBASE >> L1_SHIFT) & Ln_ADDR_MASK); /* Set this early so we can use the pagetable walking functions */ kernel_pmap_store.pm_l1 = (pd_entry_t *)l1pt; PMAP_LOCK_INIT(kernel_pmap); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); LIST_INIT(&allpmaps); /* Assume the address we were loaded to is a valid physical address */ min_pa = max_pa = KERNBASE - kern_delta; /* * Find the minimum physical address. physmap is sorted, * but may contain empty ranges. */ for (i = 0; i < (physmap_idx * 2); i += 2) { if (physmap[i] == physmap[i + 1]) continue; if (physmap[i] <= min_pa) min_pa = physmap[i]; if (physmap[i + 1] > max_pa) max_pa = physmap[i + 1]; break; } /* Create a direct map region early so we can use it for pa -> va */ pmap_bootstrap_dmap(l1pt, min_pa, max_pa); va = KERNBASE; pa = KERNBASE - kern_delta; /* * Start to initialize phys_avail by copying from physmap * up to the physical address KERNBASE points at. */ map_slot = avail_slot = 0; for (; map_slot < (physmap_idx * 2); map_slot += 2) { if (physmap[map_slot] == physmap[map_slot + 1]) continue; if (physmap[map_slot] <= pa && physmap[map_slot + 1] > pa) break; phys_avail[avail_slot] = physmap[map_slot]; phys_avail[avail_slot + 1] = physmap[map_slot + 1]; physmem += (phys_avail[avail_slot + 1] - phys_avail[avail_slot]) >> PAGE_SHIFT; avail_slot += 2; } /* Add the memory before the kernel */ if (physmap[avail_slot] < pa) { phys_avail[avail_slot] = physmap[map_slot]; phys_avail[avail_slot + 1] = pa; physmem += (phys_avail[avail_slot + 1] - phys_avail[avail_slot]) >> PAGE_SHIFT; avail_slot += 2; } used_map_slot = map_slot; /* * Read the page table to find out what is already mapped. * This assumes we have mapped a block of memory from KERNBASE * using a single L1 entry. */ l2 = pmap_early_page_idx(l1pt, KERNBASE, &l1_slot, &l2_slot); /* Sanity check the index, KERNBASE should be the first VA */ KASSERT(l2_slot == 0, ("The L2 index is non-zero")); /* Find how many pages we have mapped */ for (; l2_slot < Ln_ENTRIES; l2_slot++) { if ((l2[l2_slot] & PTE_V) == 0) break; /* Check locore used L2 superpages */ KASSERT((l2[l2_slot] & PTE_RX) != 0, ("Invalid bootstrap L2 table")); va += L2_SIZE; pa += L2_SIZE; } va = roundup2(va, L2_SIZE); freemempos = KERNBASE + kernlen; freemempos = roundup2(freemempos, PAGE_SIZE); /* Create the l3 tables for the early devmap */ freemempos = pmap_bootstrap_l3(l1pt, VM_MAX_KERNEL_ADDRESS - L2_SIZE, freemempos); cpu_tlb_flushID(); #define alloc_pages(var, np) \ (var) = freemempos; \ freemempos += (np * PAGE_SIZE); \ memset((char *)(var), 0, ((np) * PAGE_SIZE)); /* Allocate dynamic per-cpu area. */ alloc_pages(dpcpu, DPCPU_SIZE / PAGE_SIZE); dpcpu_init((void *)dpcpu, 0); /* Allocate memory for the msgbuf, e.g. for /sbin/dmesg */ alloc_pages(msgbufpv, round_page(msgbufsize) / PAGE_SIZE); msgbufp = (void *)msgbufpv; virtual_avail = roundup2(freemempos, L2_SIZE); virtual_end = VM_MAX_KERNEL_ADDRESS - L2_SIZE; kernel_vm_end = virtual_avail; pa = pmap_early_vtophys(l1pt, freemempos); /* Finish initialising physmap */ map_slot = used_map_slot; for (; avail_slot < (PHYS_AVAIL_SIZE - 2) && map_slot < (physmap_idx * 2); map_slot += 2) { if (physmap[map_slot] == physmap[map_slot + 1]) { continue; } /* Have we used the current range? */ if (physmap[map_slot + 1] <= pa) { continue; } /* Do we need to split the entry? */ if (physmap[map_slot] < pa) { phys_avail[avail_slot] = pa; phys_avail[avail_slot + 1] = physmap[map_slot + 1]; } else { phys_avail[avail_slot] = physmap[map_slot]; phys_avail[avail_slot + 1] = physmap[map_slot + 1]; } physmem += (phys_avail[avail_slot + 1] - phys_avail[avail_slot]) >> PAGE_SHIFT; avail_slot += 2; } phys_avail[avail_slot] = 0; phys_avail[avail_slot + 1] = 0; /* * Maxmem isn't the "maximum memory", it's one larger than the * highest page of the physical address space. It should be * called something like "Maxphyspage". */ Maxmem = atop(phys_avail[avail_slot - 1]); cpu_tlb_flushID(); } /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pv_memattr = VM_MEMATTR_WRITE_BACK; } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { int i; /* * Initialize the pv chunk list mutex. */ mtx_init(&pv_chunks_mutex, "pmap pv chunk list", NULL, MTX_DEF); /* * Initialize the pool of pv list locks. */ for (i = 0; i < NPV_LIST_LOCKS; i++) rw_init(&pv_list_locks[i], "pmap pv list"); } /* * Normal, non-SMP, invalidation functions. * We inline these within pmap.c for speed. */ PMAP_INLINE void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { /* TODO */ sched_pin(); __asm __volatile("sfence.vm"); sched_unpin(); } PMAP_INLINE void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { /* TODO */ sched_pin(); __asm __volatile("sfence.vm"); sched_unpin(); } PMAP_INLINE void pmap_invalidate_all(pmap_t pmap) { /* TODO */ sched_pin(); __asm __volatile("sfence.vm"); sched_unpin(); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { pd_entry_t *l2p, l2; pt_entry_t *l3p, l3; vm_paddr_t pa; pa = 0; PMAP_LOCK(pmap); /* * Start with the l2 tabel. We are unable to allocate * pages in the l1 table. */ l2p = pmap_l2(pmap, va); if (l2p != NULL) { l2 = pmap_load(l2p); if ((l2 & PTE_RX) == 0) { l3p = pmap_l2_to_l3(l2p, va); if (l3p != NULL) { l3 = pmap_load(l3p); pa = PTE_TO_PHYS(l3); pa |= (va & L3_OFFSET); } } else { /* L2 is superpages */ pa = (l2 >> PTE_PPN1_S) << L2_SHIFT; pa |= (va & L2_OFFSET); } } PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pt_entry_t *l3p, l3; vm_paddr_t phys; vm_paddr_t pa; vm_page_t m; pa = 0; m = NULL; PMAP_LOCK(pmap); retry: l3p = pmap_l3(pmap, va); if (l3p != NULL && (l3 = pmap_load(l3p)) != 0) { if ((pmap_is_write(l3)) || ((prot & VM_PROT_WRITE) == 0)) { phys = PTE_TO_PHYS(l3); if (vm_page_pa_tryrelock(pmap, phys, &pa)) goto retry; m = PHYS_TO_VM_PAGE(phys); vm_page_hold(m); } } PA_UNLOCK_COND(pa); PMAP_UNLOCK(pmap); return (m); } vm_paddr_t pmap_kextract(vm_offset_t va) { pd_entry_t *l2; pt_entry_t *l3; vm_paddr_t pa; if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { pa = DMAP_TO_PHYS(va); } else { l2 = pmap_l2(kernel_pmap, va); if (l2 == NULL) panic("pmap_kextract: No l2"); if ((pmap_load(l2) & PTE_RX) != 0) { /* superpages */ pa = (pmap_load(l2) >> PTE_PPN1_S) << L2_SHIFT; pa |= (va & L2_OFFSET); return (pa); } l3 = pmap_l2_to_l3(l2, va); if (l3 == NULL) panic("pmap_kextract: No l3..."); pa = PTE_TO_PHYS(pmap_load(l3)); pa |= (va & PAGE_MASK); } return (pa); } /*************************************************** * Low level mapping routines..... ***************************************************/ void pmap_kenter_device(vm_offset_t sva, vm_size_t size, vm_paddr_t pa) { pt_entry_t entry; pt_entry_t *l3; vm_offset_t va; pn_t pn; KASSERT((pa & L3_OFFSET) == 0, ("pmap_kenter_device: Invalid physical address")); KASSERT((sva & L3_OFFSET) == 0, ("pmap_kenter_device: Invalid virtual address")); KASSERT((size & PAGE_MASK) == 0, ("pmap_kenter_device: Mapping is not page-sized")); va = sva; while (size != 0) { l3 = pmap_l3(kernel_pmap, va); KASSERT(l3 != NULL, ("Invalid page table, va: 0x%lx", va)); pn = (pa / PAGE_SIZE); entry = (PTE_V | PTE_RWX); entry |= (pn << PTE_PPN0_S); pmap_load_store(l3, entry); PTE_SYNC(l3); va += PAGE_SIZE; pa += PAGE_SIZE; size -= PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *l3; l3 = pmap_l3(kernel_pmap, va); KASSERT(l3 != NULL, ("pmap_kremove: Invalid address")); if (pmap_l3_valid_cacheable(pmap_load(l3))) cpu_dcache_wb_range(va, L3_SIZE); pmap_load_clear(l3); PTE_SYNC(l3); pmap_invalidate_page(kernel_pmap, va); } void pmap_kremove_device(vm_offset_t sva, vm_size_t size) { pt_entry_t *l3; vm_offset_t va; KASSERT((sva & L3_OFFSET) == 0, ("pmap_kremove_device: Invalid virtual address")); KASSERT((size & PAGE_MASK) == 0, ("pmap_kremove_device: Mapping is not page-sized")); va = sva; while (size != 0) { l3 = pmap_l3(kernel_pmap, va); KASSERT(l3 != NULL, ("Invalid page table, va: 0x%lx", va)); pmap_load_clear(l3); PTE_SYNC(l3); va += PAGE_SIZE; size -= PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return PHYS_TO_DMAP(start); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pt_entry_t *l3, pa; vm_offset_t va; vm_page_t m; pt_entry_t entry; pn_t pn; int i; va = sva; for (i = 0; i < count; i++) { m = ma[i]; pa = VM_PAGE_TO_PHYS(m); pn = (pa / PAGE_SIZE); l3 = pmap_l3(kernel_pmap, va); entry = (PTE_V | PTE_RWX); entry |= (pn << PTE_PPN0_S); pmap_load_store(l3, entry); PTE_SYNC(l3); va += L3_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qremove(vm_offset_t sva, int count) { pt_entry_t *l3; vm_offset_t va; KASSERT(sva >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", sva)); va = sva; while (count-- > 0) { l3 = pmap_l3(kernel_pmap, va); KASSERT(l3 != NULL, ("pmap_kremove: Invalid address")); if (pmap_l3_valid_cacheable(pmap_load(l3))) cpu_dcache_wb_range(va, L3_SIZE); pmap_load_clear(l3); PTE_SYNC(l3); va += PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ static __inline void pmap_free_zero_pages(struct spglist *free) { vm_page_t m; while ((m = SLIST_FIRST(free)) != NULL) { SLIST_REMOVE_HEAD(free, plinks.s.ss); /* Preserve the page's PG_ZERO setting. */ vm_page_free_toq(m); } } /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Decrements a page table page's wire count, which is used to record the * number of valid page table entries within the page. If the wire count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { --m->wire_count; if (m->wire_count == 0) { _pmap_unwire_l3(pmap, va, m, free); return (TRUE); } else { return (FALSE); } } static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { vm_paddr_t phys; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * unmap the page table page */ if (m->pindex >= NUPDE) { /* PD page */ pd_entry_t *l1; l1 = pmap_l1(pmap, va); pmap_load_clear(l1); pmap_distribute_l1(pmap, pmap_l1_index(va), 0); PTE_SYNC(l1); } else { /* PTE page */ pd_entry_t *l2; l2 = pmap_l2(pmap, va); pmap_load_clear(l2); PTE_SYNC(l2); } pmap_resident_count_dec(pmap, 1); if (m->pindex < NUPDE) { pd_entry_t *l1; /* We just released a PT, unhold the matching PD */ vm_page_t pdpg; l1 = pmap_l1(pmap, va); phys = PTE_TO_PHYS(pmap_load(l1)); pdpg = PHYS_TO_VM_PAGE(phys); pmap_unwire_l3(pmap, va, pdpg, free); } pmap_invalidate_page(pmap, va); /* * This is a release store so that the ordinary store unmapping * the page table page is globally performed before TLB shoot- * down is begun. */ atomic_subtract_rel_int(&vm_cnt.v_wire_count, 1); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing an l3 entry, this routine is used to * conditionally free the page, and manage the hold/wire counts. */ static int pmap_unuse_l3(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde, struct spglist *free) { vm_paddr_t phys; vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (0); KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0")); phys = PTE_TO_PHYS(ptepde); mpte = PHYS_TO_VM_PAGE(phys); return (pmap_unwire_l3(pmap, va, mpte, free)); } void pmap_pinit0(pmap_t pmap) { PMAP_LOCK_INIT(pmap); bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); pmap->pm_l1 = kernel_pmap->pm_l1; } int pmap_pinit(pmap_t pmap) { vm_paddr_t l1phys; vm_page_t l1pt; /* * allocate the l1 page */ while ((l1pt = vm_page_alloc(NULL, 0xdeadbeef, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) VM_WAIT; l1phys = VM_PAGE_TO_PHYS(l1pt); pmap->pm_l1 = (pd_entry_t *)PHYS_TO_DMAP(l1phys); if ((l1pt->flags & PG_ZERO) == 0) pagezero(pmap->pm_l1); bzero(&pmap->pm_stats, sizeof(pmap->pm_stats)); /* Install kernel pagetables */ memcpy(pmap->pm_l1, kernel_pmap->pm_l1, PAGE_SIZE); /* Add to the list of all user pmaps */ LIST_INSERT_HEAD(&allpmaps, pmap, pm_list); return (1); } /* * This routine is called if the desired page table page does not exist. * * If page table page allocation fails, this routine may sleep before * returning NULL. It sleeps only if a lock pointer was given. * * Note: If a page allocation fails at page table level two or three, * one or two pages may be held during the wait, only to be released * afterwards. This conservative approach is easily argued to avoid * race conditions. */ static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp) { vm_page_t m, /*pdppg, */pdpg; pt_entry_t entry; vm_paddr_t phys; pn_t pn; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Allocate a page table page. */ if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { if (lockp != NULL) { RELEASE_PV_LIST_LOCK(lockp); PMAP_UNLOCK(pmap); rw_runlock(&pvh_global_lock); VM_WAIT; rw_rlock(&pvh_global_lock); PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, the page table * page may have been allocated. */ return (NULL); } if ((m->flags & PG_ZERO) == 0) pmap_zero_page(m); /* * Map the pagetable page into the process address space, if * it isn't already there. */ if (ptepindex >= NUPDE) { pd_entry_t *l1; vm_pindex_t l1index; l1index = ptepindex - NUPDE; l1 = &pmap->pm_l1[l1index]; pn = (VM_PAGE_TO_PHYS(m) / PAGE_SIZE); entry = (PTE_V); entry |= (pn << PTE_PPN0_S); pmap_load_store(l1, entry); pmap_distribute_l1(pmap, l1index, entry); PTE_SYNC(l1); } else { vm_pindex_t l1index; pd_entry_t *l1, *l2; l1index = ptepindex >> (L1_SHIFT - L2_SHIFT); l1 = &pmap->pm_l1[l1index]; if (pmap_load(l1) == 0) { /* recurse for allocating page dir */ if (_pmap_alloc_l3(pmap, NUPDE + l1index, lockp) == NULL) { --m->wire_count; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); return (NULL); } } else { phys = PTE_TO_PHYS(pmap_load(l1)); pdpg = PHYS_TO_VM_PAGE(phys); pdpg->wire_count++; } phys = PTE_TO_PHYS(pmap_load(l1)); l2 = (pd_entry_t *)PHYS_TO_DMAP(phys); l2 = &l2[ptepindex & Ln_ADDR_MASK]; pn = (VM_PAGE_TO_PHYS(m) / PAGE_SIZE); entry = (PTE_V); entry |= (pn << PTE_PPN0_S); pmap_load_store(l2, entry); PTE_SYNC(l2); } pmap_resident_count_inc(pmap, 1); return (m); } static vm_page_t pmap_alloc_l3(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t ptepindex; pd_entry_t *l2; vm_paddr_t phys; vm_page_t m; /* * Calculate pagetable page index */ ptepindex = pmap_l2_pindex(va); retry: /* * Get the page directory entry */ l2 = pmap_l2(pmap, va); /* * If the page table page is mapped, we just increment the * hold count, and activate it. */ if (l2 != NULL && pmap_load(l2) != 0) { phys = PTE_TO_PHYS(pmap_load(l2)); m = PHYS_TO_VM_PAGE(phys); m->wire_count++; } else { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ m = _pmap_alloc_l3(pmap, ptepindex, lockp); if (m == NULL && lockp != NULL) goto retry; } return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { vm_page_t m; KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_l1)); m->wire_count--; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); /* Remove pmap from the allpmaps list */ LIST_REMOVE(pmap, pm_list); /* Remove kernel pagetables */ bzero(pmap->pm_l1, PAGE_SIZE); } #if 0 static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; return sysctl_handle_long(oidp, &ksize, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_size, "LU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return sysctl_handle_long(oidp, &kfree, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD, 0, 0, kvm_free, "LU", "Amount of KVM free"); #endif /* 0 */ /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t paddr; vm_page_t nkpg; pd_entry_t *l1, *l2; pt_entry_t entry; pn_t pn; mtx_assert(&kernel_map->system_mtx, MA_OWNED); addr = roundup2(addr, L2_SIZE); if (addr - 1 >= kernel_map->max_offset) addr = kernel_map->max_offset; while (kernel_vm_end < addr) { l1 = pmap_l1(kernel_pmap, kernel_vm_end); if (pmap_load(l1) == 0) { /* We need a new PDP entry */ nkpg = vm_page_alloc(NULL, kernel_vm_end >> L1_SHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); if ((nkpg->flags & PG_ZERO) == 0) pmap_zero_page(nkpg); paddr = VM_PAGE_TO_PHYS(nkpg); pn = (paddr / PAGE_SIZE); entry = (PTE_V); entry |= (pn << PTE_PPN0_S); pmap_load_store(l1, entry); pmap_distribute_l1(kernel_pmap, pmap_l1_index(kernel_vm_end), entry); PTE_SYNC(l1); continue; /* try again */ } l2 = pmap_l1_to_l2(l1, kernel_vm_end); if ((pmap_load(l2) & PTE_A) != 0) { kernel_vm_end = (kernel_vm_end + L2_SIZE) & ~L2_OFFSET; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } continue; } nkpg = vm_page_alloc(NULL, kernel_vm_end >> L2_SHIFT, VM_ALLOC_INTERRUPT | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); if ((nkpg->flags & PG_ZERO) == 0) { pmap_zero_page(nkpg); } paddr = VM_PAGE_TO_PHYS(nkpg); pn = (paddr / PAGE_SIZE); entry = (PTE_V); entry |= (pn << PTE_PPN0_S); pmap_load_store(l2, entry); PTE_SYNC(l2); pmap_invalidate_page(kernel_pmap, kernel_vm_end); kernel_vm_end = (kernel_vm_end + L2_SIZE) & ~L2_OFFSET; if (kernel_vm_end - 1 >= kernel_map->max_offset) { kernel_vm_end = kernel_map->max_offset; break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 3); CTASSERT(_NPCPV == 168); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0 0xfffffffffffffffful #define PC_FREE1 0xfffffffffffffffful #define PC_FREE2 0x000000fffffffffful static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 }; #if 0 #ifdef PV_STATS static int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, 0, "Current number of pv entry chunks"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, 0, "Current number of pv entry chunks allocated"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, 0, "Current number of pv entry chunks frees"); SYSCTL_INT(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, 0, "Number of times tried to get a chunk page but failed."); static long pv_entry_frees, pv_entry_allocs, pv_entry_count; static int pv_entry_spare; SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, 0, "Current number of pv entry frees"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, 0, "Current number of pv entry allocs"); SYSCTL_LONG(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, 0, "Current number of pv entries"); SYSCTL_INT(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, 0, "Current number of spare pv entries"); #endif #endif /* 0 */ /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. * * Returns NULL if PV entries were reclaimed from the specified pmap. * * We do not, however, unmap 2mpages because subsequent accesses will * allocate per-page pv entries until repromotion occurs, thereby * exacerbating the shortage of free pv entries. */ static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) { panic("RISCVTODO: reclaim_pv_chunk"); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; rw_assert(&pvh_global_lock, RA_LOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_frees, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, 1)); PV_STAT(atomic_subtract_long(&pv_entry_count, 1)); pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 || pc->pc_map[2] != PC_FREE2) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk(struct pv_chunk *pc) { vm_page_t m; mtx_lock(&pv_chunks_mutex); TAILQ_REMOVE(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); PV_STAT(atomic_subtract_int(&pv_entry_spare, _NPCPV)); PV_STAT(atomic_subtract_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_frees, 1)); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); #if 0 /* TODO: For minidump */ dump_drop_page(m->phys_addr); #endif vm_page_unwire(m, PQ_INACTIVE); vm_page_free(m); } /* * Returns a new PV entry, allocating a new PV chunk from the system when * needed. If this PV chunk allocation fails and a PV list lock pointer was * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is * returned. * * The given PV list lock may be released. */ static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp) { int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; rw_assert(&pvh_global_lock, RA_LOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(atomic_add_long(&pv_entry_allocs, 1)); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = ffsl(pc->pc_map[field]) - 1; break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 64 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_subtract_int(&pv_entry_spare, 1)); return (pv); } } /* No free items, allocate another chunk */ m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED); if (m == NULL) { if (lockp == NULL) { PV_STAT(pc_chunk_tryfail++); return (NULL); } m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } PV_STAT(atomic_add_int(&pc_chunk_count, 1)); PV_STAT(atomic_add_int(&pc_chunk_allocs, 1)); #if 0 /* TODO: This is for minidump */ dump_add_page(m->phys_addr); #endif pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */ pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; mtx_lock(&pv_chunks_mutex); TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); mtx_unlock(&pv_chunks_mutex); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(atomic_add_long(&pv_entry_count, 1)); PV_STAT(atomic_add_int(&pv_entry_spare, _NPCPV - 1)); return (pv); } /* * First find and then remove the pv entry for the specified pmap and virtual * address from the specified pv list. Returns the pv entry if found and NULL * otherwise. This operation can be performed on pv lists for either 4KB or * 2MB page mappings. */ static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_LOCKED); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; break; } } return (pv); } /* * First find and then destroy the pv entry for the specified pmap and virtual * address. This operation can be performed on pv lists for either 4KB or 2MB * page mappings. */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } /* * Conditionally create the PV entry for a 4KB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_LOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; return (TRUE); } else return (FALSE); } /* * pmap_remove_l3: do the things to unmap a page in a process */ static int pmap_remove_l3(pmap_t pmap, pt_entry_t *l3, vm_offset_t va, pd_entry_t l2e, struct spglist *free, struct rwlock **lockp) { pt_entry_t old_l3; vm_paddr_t phys; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(pmap_load(l3))) cpu_dcache_wb_range(va, L3_SIZE); old_l3 = pmap_load_clear(l3); PTE_SYNC(l3); pmap_invalidate_page(pmap, va); if (old_l3 & PTE_SW_WIRED) pmap->pm_stats.wired_count -= 1; pmap_resident_count_dec(pmap, 1); if (old_l3 & PTE_SW_MANAGED) { phys = PTE_TO_PHYS(old_l3); m = PHYS_TO_VM_PAGE(phys); if (pmap_page_dirty(old_l3)) vm_page_dirty(m); if (old_l3 & PTE_A) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(&m->md, pmap, va); } return (pmap_unuse_l3(pmap, va, l2e, free)); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct rwlock *lock; vm_offset_t va, va_next; pd_entry_t *l1, *l2; pt_entry_t l3_pte, *l3; struct spglist free; int anyvalid; /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; anyvalid = 0; SLIST_INIT(&free); rw_rlock(&pvh_global_lock); PMAP_LOCK(pmap); lock = NULL; for (; sva < eva; sva = va_next) { if (pmap->pm_stats.resident_count == 0) break; l1 = pmap_l1(pmap, sva); if (pmap_load(l1) == 0) { va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; continue; } /* * Calculate index for next page table. */ va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (l2 == NULL) continue; l3_pte = pmap_load(l2); /* * Weed out invalid mappings. */ if (l3_pte == 0) continue; if ((pmap_load(l2) & PTE_RX) != 0) continue; /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (va_next > eva) va_next = eva; va = va_next; for (l3 = pmap_l2_to_l3(l2, sva); sva != va_next; l3++, sva += L3_SIZE) { if (l3 == NULL) panic("l3 == NULL"); if (pmap_load(l3) == 0) { if (va != va_next) { pmap_invalidate_range(pmap, va, sva); va = va_next; } continue; } if (va == va_next) va = sva; if (pmap_remove_l3(pmap, l3, sva, l3_pte, &free, &lock)) { sva += L3_SIZE; break; } } if (va != va_next) pmap_invalidate_range(pmap, va, sva); } if (lock != NULL) rw_wunlock(lock); if (anyvalid) pmap_invalidate_all(pmap); rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { pv_entry_t pv; pmap_t pmap; pt_entry_t *l3, tl3; pd_entry_t *l2, tl2; struct spglist free; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); rw_wlock(&pvh_global_lock); while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pmap_resident_count_dec(pmap, 1); l2 = pmap_l2(pmap, pv->pv_va); KASSERT(l2 != NULL, ("pmap_remove_all: no l2 table found")); tl2 = pmap_load(l2); KASSERT((tl2 & PTE_RX) == 0, ("pmap_remove_all: found a table when expecting " "a block in %p's pv list", m)); l3 = pmap_l2_to_l3(l2, pv->pv_va); if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(pmap_load(l3))) cpu_dcache_wb_range(pv->pv_va, L3_SIZE); tl3 = pmap_load_clear(l3); PTE_SYNC(l3); pmap_invalidate_page(pmap, pv->pv_va); if (tl3 & PTE_SW_WIRED) pmap->pm_stats.wired_count--; if ((tl3 & PTE_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if (pmap_page_dirty(tl3)) vm_page_dirty(m); pmap_unuse_l3(pmap, pv->pv_va, pmap_load(l2), &free); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&pvh_global_lock); pmap_free_zero_pages(&free); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t va, va_next; pd_entry_t *l1, *l2; pt_entry_t *l3p, l3; pt_entry_t entry; if ((prot & VM_PROT_READ) == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } if ((prot & VM_PROT_WRITE) == VM_PROT_WRITE) return; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { l1 = pmap_l1(pmap, sva); if (pmap_load(l1) == 0) { va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (l2 == NULL) continue; if ((pmap_load(l2) & PTE_RX) != 0) continue; if (va_next > eva) va_next = eva; va = va_next; for (l3p = pmap_l2_to_l3(l2, sva); sva != va_next; l3p++, sva += L3_SIZE) { l3 = pmap_load(l3p); if (pmap_l3_valid(l3)) { entry = pmap_load(l3p); entry &= ~(PTE_W); pmap_load_store(l3p, entry); PTE_SYNC(l3p); /* XXX: Use pmap_invalidate_range */ pmap_invalidate_page(pmap, va); } } } PMAP_UNLOCK(pmap); /* TODO: Only invalidate entries we are touching */ pmap_invalidate_all(pmap); } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind __unused) { struct rwlock *lock; pd_entry_t *l1, *l2; pt_entry_t new_l3, orig_l3; pt_entry_t *l3; pv_entry_t pv; vm_paddr_t opa, pa, l2_pa, l3_pa; vm_page_t mpte, om, l2_m, l3_m; boolean_t nosleep; pt_entry_t entry; pn_t l2_pn; pn_t l3_pn; pn_t pn; va = trunc_page(va); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); pa = VM_PAGE_TO_PHYS(m); pn = (pa / PAGE_SIZE); new_l3 = PTE_V | PTE_R | PTE_X; if (prot & VM_PROT_WRITE) new_l3 |= PTE_W; if ((va >> 63) == 0) new_l3 |= PTE_U; new_l3 |= (pn << PTE_PPN0_S); if ((flags & PMAP_ENTER_WIRED) != 0) new_l3 |= PTE_SW_WIRED; CTR2(KTR_PMAP, "pmap_enter: %.16lx -> %.16lx", va, pa); mpte = NULL; lock = NULL; rw_rlock(&pvh_global_lock); PMAP_LOCK(pmap); if (va < VM_MAXUSER_ADDRESS) { nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0; mpte = pmap_alloc_l3(pmap, va, nosleep ? NULL : &lock); if (mpte == NULL && nosleep) { CTR0(KTR_PMAP, "pmap_enter: mpte == NULL"); if (lock != NULL) rw_wunlock(lock); rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (KERN_RESOURCE_SHORTAGE); } l3 = pmap_l3(pmap, va); } else { l3 = pmap_l3(pmap, va); /* TODO: This is not optimal, but should mostly work */ if (l3 == NULL) { l2 = pmap_l2(pmap, va); if (l2 == NULL) { l2_m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (l2_m == NULL) panic("pmap_enter: l2 pte_m == NULL"); if ((l2_m->flags & PG_ZERO) == 0) pmap_zero_page(l2_m); l2_pa = VM_PAGE_TO_PHYS(l2_m); l2_pn = (l2_pa / PAGE_SIZE); l1 = pmap_l1(pmap, va); entry = (PTE_V); entry |= (l2_pn << PTE_PPN0_S); pmap_load_store(l1, entry); pmap_distribute_l1(pmap, pmap_l1_index(va), entry); PTE_SYNC(l1); l2 = pmap_l1_to_l2(l1, va); } KASSERT(l2 != NULL, ("No l2 table after allocating one")); l3_m = vm_page_alloc(NULL, 0, VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (l3_m == NULL) panic("pmap_enter: l3 pte_m == NULL"); if ((l3_m->flags & PG_ZERO) == 0) pmap_zero_page(l3_m); l3_pa = VM_PAGE_TO_PHYS(l3_m); l3_pn = (l3_pa / PAGE_SIZE); entry = (PTE_V); entry |= (l3_pn << PTE_PPN0_S); pmap_load_store(l2, entry); PTE_SYNC(l2); l3 = pmap_l2_to_l3(l2, va); } pmap_invalidate_page(pmap, va); } om = NULL; orig_l3 = pmap_load(l3); opa = PTE_TO_PHYS(orig_l3); /* * Is the specified virtual address already mapped? */ if (pmap_l3_valid(orig_l3)) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((flags & PMAP_ENTER_WIRED) != 0 && (orig_l3 & PTE_SW_WIRED) == 0) pmap->pm_stats.wired_count++; else if ((flags & PMAP_ENTER_WIRED) == 0 && (orig_l3 & PTE_SW_WIRED) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->wire_count--; KASSERT(mpte->wire_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%lx", va)); } /* * Has the physical page changed? */ if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((orig_l3 & PTE_SW_MANAGED) != 0) { new_l3 |= PTE_SW_MANAGED; if (pmap_is_write(new_l3)) vm_page_aflag_set(m, PGA_WRITEABLE); } goto validate; } /* Flush the cache, there might be uncommitted data in it */ if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(orig_l3)) cpu_dcache_wb_range(va, L3_SIZE); } else { /* * Increment the counters. */ if ((new_l3 & PTE_SW_WIRED) != 0) pmap->pm_stats.wired_count++; pmap_resident_count_inc(pmap, 1); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0) { new_l3 |= PTE_SW_MANAGED; pv = get_pv_entry(pmap, &lock); pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (pmap_is_write(new_l3)) vm_page_aflag_set(m, PGA_WRITEABLE); } /* * Update the L3 entry. */ if (orig_l3 != 0) { validate: orig_l3 = pmap_load_store(l3, new_l3); PTE_SYNC(l3); opa = PTE_TO_PHYS(orig_l3); if (opa != pa) { if ((orig_l3 & PTE_SW_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); if (pmap_page_dirty(orig_l3)) vm_page_dirty(om); if ((orig_l3 & PTE_A) != 0) vm_page_aflag_set(om, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa); pmap_pvh_free(&om->md, pmap, va); } } else if (pmap_page_dirty(orig_l3)) { if ((orig_l3 & PTE_SW_MANAGED) != 0) vm_page_dirty(m); } } else { pmap_load_store(l3, new_l3); PTE_SYNC(l3); } pmap_invalidate_page(pmap, va); if ((pmap != pmap_kernel()) && (pmap == &curproc->p_vmspace->vm_pmap)) cpu_icache_sync_range(va, PAGE_SIZE); if (lock != NULL) rw_wunlock(lock); rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { struct rwlock *lock; vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; lock = NULL; rw_rlock(&pvh_global_lock); PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte, &lock); m = TAILQ_NEXT(m, listq); } if (lock != NULL) rw_wunlock(lock); rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { struct rwlock *lock; lock = NULL; rw_rlock(&pvh_global_lock); PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock); if (lock != NULL) rw_wunlock(lock); rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp) { struct spglist free; vm_paddr_t phys; pd_entry_t *l2; pt_entry_t *l3; vm_paddr_t pa; pt_entry_t entry; pn_t pn; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); rw_assert(&pvh_global_lock, RA_LOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); CTR2(KTR_PMAP, "pmap_enter_quick_locked: %p %lx", pmap, va); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { vm_pindex_t l2pindex; /* * Calculate pagetable page index */ l2pindex = pmap_l2_pindex(va); if (mpte && (mpte->pindex == l2pindex)) { mpte->wire_count++; } else { /* * Get the l2 entry */ l2 = pmap_l2(pmap, va); /* * If the page table page is mapped, we just increment * the hold count, and activate it. Otherwise, we * attempt to allocate a page table page. If this * attempt fails, we don't retry. Instead, we give up. */ if (l2 != NULL && pmap_load(l2) != 0) { phys = PTE_TO_PHYS(pmap_load(l2)); mpte = PHYS_TO_VM_PAGE(phys); mpte->wire_count++; } else { /* * Pass NULL instead of the PV list lock * pointer, because we don't intend to sleep. */ mpte = _pmap_alloc_l3(pmap, l2pindex, NULL); if (mpte == NULL) return (mpte); } } l3 = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte)); l3 = &l3[pmap_l3_index(va)]; } else { mpte = NULL; l3 = pmap_l3(kernel_pmap, va); } if (l3 == NULL) panic("pmap_enter_quick_locked: No l3"); if (pmap_load(l3) != 0) { if (mpte != NULL) { mpte->wire_count--; mpte = NULL; } return (mpte); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m, lockp)) { if (mpte != NULL) { SLIST_INIT(&free); if (pmap_unwire_l3(pmap, va, mpte, &free)) { pmap_invalidate_page(pmap, va); pmap_free_zero_pages(&free); } mpte = NULL; } return (mpte); } /* * Increment counters */ pmap_resident_count_inc(pmap, 1); pa = VM_PAGE_TO_PHYS(m); pn = (pa / PAGE_SIZE); /* RISCVTODO: check permissions */ entry = (PTE_V | PTE_RWX); entry |= (pn << PTE_PPN0_S); /* * Now validate mapping with RO protection */ if ((m->oflags & VPO_UNMANAGED) == 0) entry |= PTE_SW_MANAGED; pmap_load_store(l3, entry); PTE_SYNC(l3); pmap_invalidate_page(pmap, va); return (mpte); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, * so there is no need to invalidate any TLB entries. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va_next; pd_entry_t *l1, *l2; pt_entry_t *l3; boolean_t pv_lists_locked; pv_lists_locked = FALSE; PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { l1 = pmap_l1(pmap, sva); if (pmap_load(l1) == 0) { va_next = (sva + L1_SIZE) & ~L1_OFFSET; if (va_next < sva) va_next = eva; continue; } va_next = (sva + L2_SIZE) & ~L2_OFFSET; if (va_next < sva) va_next = eva; l2 = pmap_l1_to_l2(l1, sva); if (pmap_load(l2) == 0) continue; if (va_next > eva) va_next = eva; for (l3 = pmap_l2_to_l3(l2, sva); sva != va_next; l3++, sva += L3_SIZE) { if (pmap_load(l3) == 0) continue; if ((pmap_load(l3) & PTE_SW_WIRED) == 0) panic("pmap_unwire: l3 %#jx is missing " "PTE_SW_WIRED", (uintmax_t)*l3); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. */ atomic_clear_long(l3, PTE_SW_WIRED); pmap->pm_stats.wired_count--; } } if (pv_lists_locked) rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { } /* * pmap_zero_page zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. */ void pmap_zero_page(vm_page_t m) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); pagezero((void *)va); } /* * pmap_zero_page_area zeros the specified hardware page by mapping * the page into KVM and using bzero to clear its contents. * * off and size may not cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); if (off == 0 && size == PAGE_SIZE) pagezero((void *)va); else bzero((char *)va + off, size); } /* * pmap_copy_page copies the specified (machine independent) * page by mapping the page into virtual memory and using * bcopy to copy the page, one machine dependent page at a * time. */ void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc)); vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst)); pagecopy((void *)src, (void *)dst); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_page_t m_a, m_b; vm_paddr_t p_a, p_b; vm_offset_t a_pg_offset, b_pg_offset; int cnt; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; m_a = ma[a_offset >> PAGE_SHIFT]; p_a = m_a->phys_addr; b_pg_offset = b_offset & PAGE_MASK; m_b = mb[b_offset >> PAGE_SHIFT]; p_b = m_b->phys_addr; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); cnt = min(cnt, PAGE_SIZE - b_pg_offset); if (__predict_false(!PHYS_IN_DMAP(p_a))) { panic("!DMAP a %lx", p_a); } else { a_cp = (char *)PHYS_TO_DMAP(p_a) + a_pg_offset; } if (__predict_false(!PHYS_IN_DMAP(p_b))) { panic("!DMAP b %lx", p_b); } else { b_cp = (char *)PHYS_TO_DMAP(p_b) + b_pg_offset; } bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { return (PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m))); } void pmap_quick_remove_page(vm_offset_t addr) { } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct rwlock *lock; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; rw_rlock(&pvh_global_lock); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } rw_runlock(lock); rw_runlock(&pvh_global_lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct rwlock *lock; pmap_t pmap; pt_entry_t *l3; pv_entry_t pv; int count, md_gen; if ((m->oflags & VPO_UNMANAGED) != 0) return (0); rw_rlock(&pvh_global_lock); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: count = 0; TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } l3 = pmap_l3(pmap, pv->pv_va); if (l3 != NULL && (pmap_load(l3) & PTE_SW_WIRED) != 0) count++; PMAP_UNLOCK(pmap); } rw_runlock(lock); rw_runlock(&pvh_global_lock); return (count); } /* * Destroy all managed, non-wired mappings in the given user-space * pmap. This pmap cannot be active on any processor besides the * caller. * * This function cannot be applied to the kernel pmap. Moreover, it * is not intended for general use. It is only to be used during * process termination. Consequently, it can be implemented in ways * that make it faster than pmap_remove(). First, it can more quickly * destroy mappings by iterating over the pmap's collection of PV * entries, rather than searching the page table. Second, it doesn't * have to test and clear the page table entries atomically, because * no processor is currently accessing the user address space. In * particular, a page table entry's dirty bit won't change state once * this function starts. */ void pmap_remove_pages(pmap_t pmap) { pd_entry_t ptepde, *l2; pt_entry_t *l3, tl3; struct spglist free; vm_page_t m; pv_entry_t pv; struct pv_chunk *pc, *npc; struct rwlock *lock; int64_t bit; uint64_t inuse, bitmask; int allfree, field, freed, idx; vm_paddr_t pa; lock = NULL; SLIST_INIT(&free); rw_rlock(&pvh_global_lock); PMAP_LOCK(pmap); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { allfree = 1; freed = 0; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = ffsl(inuse) - 1; bitmask = 1UL << bit; idx = field * 64 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; l2 = pmap_l2(pmap, pv->pv_va); ptepde = pmap_load(l2); l3 = pmap_l2_to_l3(l2, pv->pv_va); tl3 = pmap_load(l3); /* * We cannot remove wired pages from a process' mapping at this time */ if (tl3 & PTE_SW_WIRED) { allfree = 0; continue; } pa = PTE_TO_PHYS(tl3); m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tl3)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad l3 %#jx", (uintmax_t)tl3)); if (pmap_is_current(pmap) && pmap_l3_valid_cacheable(pmap_load(l3))) cpu_dcache_wb_range(pv->pv_va, L3_SIZE); pmap_load_clear(l3); PTE_SYNC(l3); pmap_invalidate_page(pmap, pv->pv_va); /* * Update the vm_page_t clean/reference bits. */ if (pmap_page_dirty(tl3)) vm_page_dirty(m); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); /* Mark free */ pc->pc_map[field] |= bitmask; pmap_resident_count_dec(pmap, 1); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; pmap_unuse_l3(pmap, pv->pv_va, ptepde, &free); freed++; } } PV_STAT(atomic_add_long(&pv_entry_frees, freed)); PV_STAT(atomic_add_int(&pv_entry_spare, freed)); PV_STAT(atomic_subtract_long(&pv_entry_count, freed)); if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } } pmap_invalidate_all(pmap); if (lock != NULL) rw_wunlock(lock); rw_runlock(&pvh_global_lock); PMAP_UNLOCK(pmap); pmap_free_zero_pages(&free); } /* * This is used to check if a page has been accessed or modified. As we * don't have a bit to see if it has been modified we have to assume it * has been if the page is read/write. */ static boolean_t pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified) { struct rwlock *lock; pv_entry_t pv; pt_entry_t *l3, mask, value; pmap_t pmap; int md_gen; boolean_t rv; rv = FALSE; rw_rlock(&pvh_global_lock); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } l3 = pmap_l3(pmap, pv->pv_va); mask = 0; value = 0; if (modified) { mask |= PTE_D; value |= PTE_D; } if (accessed) { mask |= PTE_A; value |= PTE_A; } #if 0 if (modified) { mask |= ATTR_AP_RW_BIT; value |= ATTR_AP(ATTR_AP_RW); } if (accessed) { mask |= ATTR_AF | ATTR_DESCR_MASK; value |= ATTR_AF | L3_PAGE; } #endif rv = (pmap_load(l3) & mask) == value; PMAP_UNLOCK(pmap); if (rv) goto out; } out: rw_runlock(lock); rw_runlock(&pvh_global_lock); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no PTEs can have PG_M set. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (FALSE); return (pmap_page_test_mappings(m, FALSE, TRUE)); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pt_entry_t *l3; boolean_t rv; rv = FALSE; PMAP_LOCK(pmap); l3 = pmap_l3(pmap, addr); if (l3 != NULL && pmap_load(l3) != 0) { rv = TRUE; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); return (pmap_page_test_mappings(m, TRUE, FALSE)); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { pmap_t pmap; struct rwlock *lock; pv_entry_t pv; pt_entry_t *l3, oldl3; pt_entry_t newl3; int md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; rw_rlock(&pvh_global_lock); lock = VM_PAGE_TO_PV_LIST_LOCK(m); retry_pv_loop: rw_wlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); rw_wunlock(lock); goto retry_pv_loop; } } l3 = pmap_l3(pmap, pv->pv_va); retry: oldl3 = pmap_load(l3); if (pmap_is_write(oldl3)) { newl3 = oldl3 & ~(PTE_W); if (!atomic_cmpset_long(l3, oldl3, newl3)) goto retry; /* TODO: use pmap_page_dirty(oldl3) ? */ if ((oldl3 & PTE_A) != 0) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); vm_page_aflag_clear(m, PGA_WRITEABLE); rw_runlock(&pvh_global_lock); } static __inline boolean_t safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte) { return (FALSE); } -#define PMAP_TS_REFERENCED_MAX 5 - /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. + * As an optimization, update the page's dirty field if a modified bit is + * found while counting reference bits. This opportunistic update can be + * performed at low cost and can eliminate the need for some future calls + * to pmap_is_modified(). However, since this function stops after + * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some + * dirty pages. Those dirty pages will only be detected by a future call + * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) { pv_entry_t pv, pvf; pmap_t pmap; struct rwlock *lock; pd_entry_t *l2; - pt_entry_t *l3; + pt_entry_t *l3, old_l3; vm_paddr_t pa; int cleared, md_gen, not_cleared; struct spglist free; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); SLIST_INIT(&free); cleared = 0; pa = VM_PAGE_TO_PHYS(m); lock = PHYS_TO_PV_LIST_LOCK(pa); rw_rlock(&pvh_global_lock); rw_wlock(lock); retry: not_cleared = 0; if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } l2 = pmap_l2(pmap, pv->pv_va); KASSERT((pmap_load(l2) & PTE_RX) == 0, ("pmap_ts_referenced: found an invalid l2 table")); l3 = pmap_l2_to_l3(l2, pv->pv_va); - if ((pmap_load(l3) & PTE_A) != 0) { - if (safe_to_clear_referenced(pmap, pmap_load(l3))) { + old_l3 = pmap_load(l3); + if (pmap_page_dirty(old_l3)) + vm_page_dirty(m); + if ((old_l3 & PTE_A) != 0) { + if (safe_to_clear_referenced(pmap, old_l3)) { /* * TODO: We don't handle the access flag * at all. We need to be able to set it in * the exception handler. */ panic("RISCVTODO: safe_to_clear_referenced\n"); - } else if ((pmap_load(l3) & PTE_SW_WIRED) == 0) { + } else if ((old_l3 & PTE_SW_WIRED) == 0) { /* * Wired pages cannot be paged out so * doing accessed bit emulation for * them is wasted effort. We do the * hard work for unwired pages only. */ pmap_remove_l3(pmap, l3, pv->pv_va, pmap_load(l2), &free, &lock); pmap_invalidate_page(pmap, pv->pv_va); cleared++; if (pvf == pv) pvf = NULL; pv = NULL; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared + not_cleared < PMAP_TS_REFERENCED_MAX); out: rw_wunlock(lock); rw_runlock(&pvh_global_lock); pmap_free_zero_pages(&free); return (cleared + not_cleared); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("pmap_clear_modify: page %p is exclusive busied", m)); /* * If the page is not PGA_WRITEABLE, then no PTEs can have PG_M set. * If the object containing the page is locked and the page is not * exclusive busied, then PGA_WRITEABLE cannot be concurrently set. */ if ((m->aflags & PGA_WRITEABLE) == 0) return; /* RISCVTODO: We lack support for tracking if a page is modified */ } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { return ((void *)PHYS_TO_DMAP(pa)); } void pmap_unmapbios(vm_paddr_t pa, vm_size_t size) { } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pv_memattr = ma; /* * RISCVTODO: Implement the below (from the amd64 pmap) * If "m" is a normal page, update its direct mapping. This update * can be relied upon to perform any cache operations that are * required for data coherence. */ if ((m->flags & PG_FICTITIOUS) == 0 && PHYS_IN_DMAP(VM_PAGE_TO_PHYS(m))) panic("RISCVTODO: pmap_page_set_memattr"); } /* * perform the pmap work for mincore */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa) { panic("RISCVTODO: pmap_mincore"); } void pmap_activate(struct thread *td) { pmap_t pmap; critical_enter(); pmap = vmspace_pmap(td->td_proc->p_vmspace); td->td_pcb->pcb_l1addr = vtophys(pmap->pm_l1); __asm __volatile("csrw sptbr, %0" :: "r"(td->td_pcb->pcb_l1addr >> PAGE_SHIFT)); pmap_invalidate_all(pmap); critical_exit(); } void pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { panic("RISCVTODO: pmap_sync_icache"); } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { } /** * Get the kernel virtual address of a set of physical pages. If there are * physical addresses not covered by the DMAP perform a transient mapping * that will be removed when calling pmap_unmap_io_transient. * * \param page The pages the caller wishes to obtain the virtual * address on the kernel memory map. * \param vaddr On return contains the kernel virtual memory address * of the pages passed in the page parameter. * \param count Number of pages passed in. * \param can_fault TRUE if the thread using the mapped pages can take * page faults, FALSE otherwise. * * \returns TRUE if the caller must call pmap_unmap_io_transient when * finished or FALSE otherwise. * */ boolean_t pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; boolean_t needs_mapping; int error, i; /* * Allocate any KVA space that we need, this is done in a separate * loop to prevent calling vmem_alloc while pinned. */ needs_mapping = FALSE; for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (__predict_false(paddr >= DMAP_MAX_PHYSADDR)) { error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr[i]); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); needs_mapping = TRUE; } else { vaddr[i] = PHYS_TO_DMAP(paddr); } } /* Exit early if everything is covered by the DMAP */ if (!needs_mapping) return (FALSE); if (!can_fault) sched_pin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= DMAP_MAX_PHYSADDR) { panic( "pmap_map_io_transient: TODO: Map out of DMAP data"); } } return (needs_mapping); } void pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; int i; if (!can_fault) sched_unpin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= DMAP_MAX_PHYSADDR) { panic("RISCVTODO: pmap_unmap_io_transient: Unmap data"); } } } Index: projects/mips64-clang/sys/sparc64/sparc64/pmap.c =================================================================== --- projects/mips64-clang/sys/sparc64/sparc64/pmap.c (revision 305711) +++ projects/mips64-clang/sys/sparc64/sparc64/pmap.c (revision 305712) @@ -1,2334 +1,2328 @@ /*- * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is also stored by the * logical address mapping module, this module may throw away valid virtual * to physical mappings at almost any time. However, invalidations of * mappings must be done as requested. * * In order to cope with hardware architectures which make virtual to * physical map invalidates expensive, this module may delay invalidate * reduced protection operations until such time as they are actually * necessary. This module is given full information as to which processors * are currently using which maps, and to when physical maps must be made * correct. */ #include "opt_kstack_pages.h" #include "opt_pmap.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * Virtual address of message buffer */ struct msgbuf *msgbufp; /* * Map of physical memory reagions */ vm_paddr_t phys_avail[128]; static struct ofw_mem_region mra[128]; struct ofw_mem_region sparc64_memreg[128]; int sparc64_nmemreg; static struct ofw_map translations[128]; static int translations_size; static vm_offset_t pmap_idle_map; static vm_offset_t pmap_temp_map_1; static vm_offset_t pmap_temp_map_2; /* * First and last available kernel virtual addresses */ vm_offset_t virtual_avail; vm_offset_t virtual_end; vm_offset_t kernel_vm_end; vm_offset_t vm_max_kernel_address; /* * Kernel pmap */ struct pmap kernel_pmap_store; struct rwlock_padalign tte_list_global_lock; /* * Allocate physical memory for use in pmap_bootstrap. */ static vm_paddr_t pmap_bootstrap_alloc(vm_size_t size, uint32_t colors); static void pmap_bootstrap_set_tte(struct tte *tp, u_long vpn, u_long data); static void pmap_cache_remove(vm_page_t m, vm_offset_t va); static int pmap_protect_tte(struct pmap *pm1, struct pmap *pm2, struct tte *tp, vm_offset_t va); static int pmap_unwire_tte(pmap_t pm, pmap_t pm2, struct tte *tp, vm_offset_t va); static void pmap_init_qpages(void); /* * Map the given physical page at the specified virtual address in the * target pmap with the protection requested. If specified the page * will be wired down. * * The page queues and pmap must be locked. */ static int pmap_enter_locked(pmap_t pm, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind); extern int tl1_dmmu_miss_direct_patch_tsb_phys_1[]; extern int tl1_dmmu_miss_direct_patch_tsb_phys_end_1[]; extern int tl1_dmmu_miss_patch_asi_1[]; extern int tl1_dmmu_miss_patch_quad_ldd_1[]; extern int tl1_dmmu_miss_patch_tsb_1[]; extern int tl1_dmmu_miss_patch_tsb_2[]; extern int tl1_dmmu_miss_patch_tsb_mask_1[]; extern int tl1_dmmu_miss_patch_tsb_mask_2[]; extern int tl1_dmmu_prot_patch_asi_1[]; extern int tl1_dmmu_prot_patch_quad_ldd_1[]; extern int tl1_dmmu_prot_patch_tsb_1[]; extern int tl1_dmmu_prot_patch_tsb_2[]; extern int tl1_dmmu_prot_patch_tsb_mask_1[]; extern int tl1_dmmu_prot_patch_tsb_mask_2[]; extern int tl1_immu_miss_patch_asi_1[]; extern int tl1_immu_miss_patch_quad_ldd_1[]; extern int tl1_immu_miss_patch_tsb_1[]; extern int tl1_immu_miss_patch_tsb_2[]; extern int tl1_immu_miss_patch_tsb_mask_1[]; extern int tl1_immu_miss_patch_tsb_mask_2[]; /* * If user pmap is processed with pmap_remove and with pmap_remove and the * resident count drops to 0, there are no more pages to remove, so we * need not continue. */ #define PMAP_REMOVE_DONE(pm) \ ((pm) != kernel_pmap && (pm)->pm_stats.resident_count == 0) /* * The threshold (in bytes) above which tsb_foreach() is used in pmap_remove() * and pmap_protect() instead of trying each virtual address. */ #define PMAP_TSB_THRESH ((TSB_SIZE / 2) * PAGE_SIZE) SYSCTL_NODE(_debug, OID_AUTO, pmap_stats, CTLFLAG_RD, 0, ""); PMAP_STATS_VAR(pmap_nenter); PMAP_STATS_VAR(pmap_nenter_update); PMAP_STATS_VAR(pmap_nenter_replace); PMAP_STATS_VAR(pmap_nenter_new); PMAP_STATS_VAR(pmap_nkenter); PMAP_STATS_VAR(pmap_nkenter_oc); PMAP_STATS_VAR(pmap_nkenter_stupid); PMAP_STATS_VAR(pmap_nkremove); PMAP_STATS_VAR(pmap_nqenter); PMAP_STATS_VAR(pmap_nqremove); PMAP_STATS_VAR(pmap_ncache_enter); PMAP_STATS_VAR(pmap_ncache_enter_c); PMAP_STATS_VAR(pmap_ncache_enter_oc); PMAP_STATS_VAR(pmap_ncache_enter_cc); PMAP_STATS_VAR(pmap_ncache_enter_coc); PMAP_STATS_VAR(pmap_ncache_enter_nc); PMAP_STATS_VAR(pmap_ncache_enter_cnc); PMAP_STATS_VAR(pmap_ncache_remove); PMAP_STATS_VAR(pmap_ncache_remove_c); PMAP_STATS_VAR(pmap_ncache_remove_oc); PMAP_STATS_VAR(pmap_ncache_remove_cc); PMAP_STATS_VAR(pmap_ncache_remove_coc); PMAP_STATS_VAR(pmap_ncache_remove_nc); PMAP_STATS_VAR(pmap_nzero_page); PMAP_STATS_VAR(pmap_nzero_page_c); PMAP_STATS_VAR(pmap_nzero_page_oc); PMAP_STATS_VAR(pmap_nzero_page_nc); PMAP_STATS_VAR(pmap_nzero_page_area); PMAP_STATS_VAR(pmap_nzero_page_area_c); PMAP_STATS_VAR(pmap_nzero_page_area_oc); PMAP_STATS_VAR(pmap_nzero_page_area_nc); PMAP_STATS_VAR(pmap_ncopy_page); PMAP_STATS_VAR(pmap_ncopy_page_c); PMAP_STATS_VAR(pmap_ncopy_page_oc); PMAP_STATS_VAR(pmap_ncopy_page_nc); PMAP_STATS_VAR(pmap_ncopy_page_dc); PMAP_STATS_VAR(pmap_ncopy_page_doc); PMAP_STATS_VAR(pmap_ncopy_page_sc); PMAP_STATS_VAR(pmap_ncopy_page_soc); PMAP_STATS_VAR(pmap_nnew_thread); PMAP_STATS_VAR(pmap_nnew_thread_oc); static inline u_long dtlb_get_data(u_int tlb, u_int slot); /* * Quick sort callout for comparing memory regions */ static int mr_cmp(const void *a, const void *b); static int om_cmp(const void *a, const void *b); static int mr_cmp(const void *a, const void *b) { const struct ofw_mem_region *mra; const struct ofw_mem_region *mrb; mra = a; mrb = b; if (mra->mr_start < mrb->mr_start) return (-1); else if (mra->mr_start > mrb->mr_start) return (1); else return (0); } static int om_cmp(const void *a, const void *b) { const struct ofw_map *oma; const struct ofw_map *omb; oma = a; omb = b; if (oma->om_start < omb->om_start) return (-1); else if (oma->om_start > omb->om_start) return (1); else return (0); } static inline u_long dtlb_get_data(u_int tlb, u_int slot) { u_long data; register_t s; slot = TLB_DAR_SLOT(tlb, slot); /* * We read ASI_DTLB_DATA_ACCESS_REG twice back-to-back in order to * work around errata of USIII and beyond. */ s = intr_disable(); (void)ldxa(slot, ASI_DTLB_DATA_ACCESS_REG); data = ldxa(slot, ASI_DTLB_DATA_ACCESS_REG); intr_restore(s); return (data); } /* * Bootstrap the system enough to run with virtual memory. */ void pmap_bootstrap(u_int cpu_impl) { struct pmap *pm; struct tte *tp; vm_offset_t off; vm_offset_t va; vm_paddr_t pa; vm_size_t physsz; vm_size_t virtsz; u_long data; u_long vpn; phandle_t pmem; phandle_t vmem; u_int dtlb_slots_avail; int i; int j; int sz; uint32_t asi; uint32_t colors; uint32_t ldd; /* * Set the kernel context. */ pmap_set_kctx(); colors = dcache_color_ignore != 0 ? 1 : DCACHE_COLORS; /* * Find out what physical memory is available from the PROM and * initialize the phys_avail array. This must be done before * pmap_bootstrap_alloc is called. */ if ((pmem = OF_finddevice("/memory")) == -1) OF_panic("%s: finddevice /memory", __func__); if ((sz = OF_getproplen(pmem, "available")) == -1) OF_panic("%s: getproplen /memory/available", __func__); if (sizeof(phys_avail) < sz) OF_panic("%s: phys_avail too small", __func__); if (sizeof(mra) < sz) OF_panic("%s: mra too small", __func__); bzero(mra, sz); if (OF_getprop(pmem, "available", mra, sz) == -1) OF_panic("%s: getprop /memory/available", __func__); sz /= sizeof(*mra); #ifdef DIAGNOSTIC OF_printf("pmap_bootstrap: physical memory\n"); #endif qsort(mra, sz, sizeof (*mra), mr_cmp); physsz = 0; getenv_quad("hw.physmem", &physmem); physmem = btoc(physmem); for (i = 0, j = 0; i < sz; i++, j += 2) { #ifdef DIAGNOSTIC OF_printf("start=%#lx size=%#lx\n", mra[i].mr_start, mra[i].mr_size); #endif if (physmem != 0 && btoc(physsz + mra[i].mr_size) >= physmem) { if (btoc(physsz) < physmem) { phys_avail[j] = mra[i].mr_start; phys_avail[j + 1] = mra[i].mr_start + (ctob(physmem) - physsz); physsz = ctob(physmem); } break; } phys_avail[j] = mra[i].mr_start; phys_avail[j + 1] = mra[i].mr_start + mra[i].mr_size; physsz += mra[i].mr_size; } physmem = btoc(physsz); /* * Calculate the size of kernel virtual memory, and the size and mask * for the kernel TSB based on the phsyical memory size but limited * by the amount of dTLB slots available for locked entries if we have * to lock the TSB in the TLB (given that for spitfire-class CPUs all * of the dt64 slots can hold locked entries but there is no large * dTLB for unlocked ones, we don't use more than half of it for the * TSB). * Note that for reasons unknown OpenSolaris doesn't take advantage of * ASI_ATOMIC_QUAD_LDD_PHYS on UltraSPARC-III. However, given that no * public documentation is available for these, the latter just might * not support it, yet. */ if (cpu_impl == CPU_IMPL_SPARC64V || cpu_impl >= CPU_IMPL_ULTRASPARCIIIp) { tsb_kernel_ldd_phys = 1; virtsz = roundup(5 / 3 * physsz, PAGE_SIZE_4M << (PAGE_SHIFT - TTE_SHIFT)); } else { dtlb_slots_avail = 0; for (i = 0; i < dtlb_slots; i++) { data = dtlb_get_data(cpu_impl == CPU_IMPL_ULTRASPARCIII ? TLB_DAR_T16 : TLB_DAR_T32, i); if ((data & (TD_V | TD_L)) != (TD_V | TD_L)) dtlb_slots_avail++; } #ifdef SMP dtlb_slots_avail -= PCPU_PAGES; #endif if (cpu_impl >= CPU_IMPL_ULTRASPARCI && cpu_impl < CPU_IMPL_ULTRASPARCIII) dtlb_slots_avail /= 2; virtsz = roundup(physsz, PAGE_SIZE_4M << (PAGE_SHIFT - TTE_SHIFT)); virtsz = MIN(virtsz, (dtlb_slots_avail * PAGE_SIZE_4M) << (PAGE_SHIFT - TTE_SHIFT)); } vm_max_kernel_address = VM_MIN_KERNEL_ADDRESS + virtsz; tsb_kernel_size = virtsz >> (PAGE_SHIFT - TTE_SHIFT); tsb_kernel_mask = (tsb_kernel_size >> TTE_SHIFT) - 1; /* * Allocate the kernel TSB and lock it in the TLB if necessary. */ pa = pmap_bootstrap_alloc(tsb_kernel_size, colors); if (pa & PAGE_MASK_4M) OF_panic("%s: TSB unaligned", __func__); tsb_kernel_phys = pa; if (tsb_kernel_ldd_phys == 0) { tsb_kernel = (struct tte *)(VM_MIN_KERNEL_ADDRESS - tsb_kernel_size); pmap_map_tsb(); bzero(tsb_kernel, tsb_kernel_size); } else { tsb_kernel = (struct tte *)TLB_PHYS_TO_DIRECT(tsb_kernel_phys); aszero(ASI_PHYS_USE_EC, tsb_kernel_phys, tsb_kernel_size); } /* * Allocate and map the dynamic per-CPU area for the BSP. */ pa = pmap_bootstrap_alloc(DPCPU_SIZE, colors); dpcpu0 = (void *)TLB_PHYS_TO_DIRECT(pa); /* * Allocate and map the message buffer. */ pa = pmap_bootstrap_alloc(msgbufsize, colors); msgbufp = (struct msgbuf *)TLB_PHYS_TO_DIRECT(pa); /* * Patch the TSB addresses and mask as well as the ASIs used to load * it into the trap table. */ #define LDDA_R_I_R(rd, imm_asi, rs1, rs2) \ (EIF_OP(IOP_LDST) | EIF_F3_RD(rd) | EIF_F3_OP3(INS3_LDDA) | \ EIF_F3_RS1(rs1) | EIF_F3_I(0) | EIF_F3_IMM_ASI(imm_asi) | \ EIF_F3_RS2(rs2)) #define OR_R_I_R(rd, imm13, rs1) \ (EIF_OP(IOP_MISC) | EIF_F3_RD(rd) | EIF_F3_OP3(INS2_OR) | \ EIF_F3_RS1(rs1) | EIF_F3_I(1) | EIF_IMM(imm13, 13)) #define SETHI(rd, imm22) \ (EIF_OP(IOP_FORM2) | EIF_F2_RD(rd) | EIF_F2_OP2(INS0_SETHI) | \ EIF_IMM((imm22) >> 10, 22)) #define WR_R_I(rd, imm13, rs1) \ (EIF_OP(IOP_MISC) | EIF_F3_RD(rd) | EIF_F3_OP3(INS2_WR) | \ EIF_F3_RS1(rs1) | EIF_F3_I(1) | EIF_IMM(imm13, 13)) #define PATCH_ASI(addr, asi) do { \ if (addr[0] != WR_R_I(IF_F3_RD(addr[0]), 0x0, \ IF_F3_RS1(addr[0]))) \ OF_panic("%s: patched instructions have changed", \ __func__); \ addr[0] |= EIF_IMM((asi), 13); \ flush(addr); \ } while (0) #define PATCH_LDD(addr, asi) do { \ if (addr[0] != LDDA_R_I_R(IF_F3_RD(addr[0]), 0x0, \ IF_F3_RS1(addr[0]), IF_F3_RS2(addr[0]))) \ OF_panic("%s: patched instructions have changed", \ __func__); \ addr[0] |= EIF_F3_IMM_ASI(asi); \ flush(addr); \ } while (0) #define PATCH_TSB(addr, val) do { \ if (addr[0] != SETHI(IF_F2_RD(addr[0]), 0x0) || \ addr[1] != OR_R_I_R(IF_F3_RD(addr[1]), 0x0, \ IF_F3_RS1(addr[1])) || \ addr[3] != SETHI(IF_F2_RD(addr[3]), 0x0)) \ OF_panic("%s: patched instructions have changed", \ __func__); \ addr[0] |= EIF_IMM((val) >> 42, 22); \ addr[1] |= EIF_IMM((val) >> 32, 10); \ addr[3] |= EIF_IMM((val) >> 10, 22); \ flush(addr); \ flush(addr + 1); \ flush(addr + 3); \ } while (0) #define PATCH_TSB_MASK(addr, val) do { \ if (addr[0] != SETHI(IF_F2_RD(addr[0]), 0x0) || \ addr[1] != OR_R_I_R(IF_F3_RD(addr[1]), 0x0, \ IF_F3_RS1(addr[1]))) \ OF_panic("%s: patched instructions have changed", \ __func__); \ addr[0] |= EIF_IMM((val) >> 10, 22); \ addr[1] |= EIF_IMM((val), 10); \ flush(addr); \ flush(addr + 1); \ } while (0) if (tsb_kernel_ldd_phys == 0) { asi = ASI_N; ldd = ASI_NUCLEUS_QUAD_LDD; off = (vm_offset_t)tsb_kernel; } else { asi = ASI_PHYS_USE_EC; ldd = ASI_ATOMIC_QUAD_LDD_PHYS; off = (vm_offset_t)tsb_kernel_phys; } PATCH_TSB(tl1_dmmu_miss_direct_patch_tsb_phys_1, tsb_kernel_phys); PATCH_TSB(tl1_dmmu_miss_direct_patch_tsb_phys_end_1, tsb_kernel_phys + tsb_kernel_size - 1); PATCH_ASI(tl1_dmmu_miss_patch_asi_1, asi); PATCH_LDD(tl1_dmmu_miss_patch_quad_ldd_1, ldd); PATCH_TSB(tl1_dmmu_miss_patch_tsb_1, off); PATCH_TSB(tl1_dmmu_miss_patch_tsb_2, off); PATCH_TSB_MASK(tl1_dmmu_miss_patch_tsb_mask_1, tsb_kernel_mask); PATCH_TSB_MASK(tl1_dmmu_miss_patch_tsb_mask_2, tsb_kernel_mask); PATCH_ASI(tl1_dmmu_prot_patch_asi_1, asi); PATCH_LDD(tl1_dmmu_prot_patch_quad_ldd_1, ldd); PATCH_TSB(tl1_dmmu_prot_patch_tsb_1, off); PATCH_TSB(tl1_dmmu_prot_patch_tsb_2, off); PATCH_TSB_MASK(tl1_dmmu_prot_patch_tsb_mask_1, tsb_kernel_mask); PATCH_TSB_MASK(tl1_dmmu_prot_patch_tsb_mask_2, tsb_kernel_mask); PATCH_ASI(tl1_immu_miss_patch_asi_1, asi); PATCH_LDD(tl1_immu_miss_patch_quad_ldd_1, ldd); PATCH_TSB(tl1_immu_miss_patch_tsb_1, off); PATCH_TSB(tl1_immu_miss_patch_tsb_2, off); PATCH_TSB_MASK(tl1_immu_miss_patch_tsb_mask_1, tsb_kernel_mask); PATCH_TSB_MASK(tl1_immu_miss_patch_tsb_mask_2, tsb_kernel_mask); /* * Enter fake 8k pages for the 4MB kernel pages, so that * pmap_kextract() will work for them. */ for (i = 0; i < kernel_tlb_slots; i++) { pa = kernel_tlbs[i].te_pa; va = kernel_tlbs[i].te_va; for (off = 0; off < PAGE_SIZE_4M; off += PAGE_SIZE) { tp = tsb_kvtotte(va + off); vpn = TV_VPN(va + off, TS_8K); data = TD_V | TD_8K | TD_PA(pa + off) | TD_REF | TD_SW | TD_CP | TD_CV | TD_P | TD_W; pmap_bootstrap_set_tte(tp, vpn, data); } } /* * Set the start and end of KVA. The kernel is loaded starting * at the first available 4MB super page, so we advance to the * end of the last one used for it. */ virtual_avail = KERNBASE + kernel_tlb_slots * PAGE_SIZE_4M; virtual_end = vm_max_kernel_address; kernel_vm_end = vm_max_kernel_address; /* * Allocate kva space for temporary mappings. */ pmap_idle_map = virtual_avail; virtual_avail += PAGE_SIZE * colors; pmap_temp_map_1 = virtual_avail; virtual_avail += PAGE_SIZE * colors; pmap_temp_map_2 = virtual_avail; virtual_avail += PAGE_SIZE * colors; /* * Allocate a kernel stack with guard page for thread0 and map it * into the kernel TSB. We must ensure that the virtual address is * colored properly for corresponding CPUs, since we're allocating * from phys_avail so the memory won't have an associated vm_page_t. */ pa = pmap_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, colors); kstack0_phys = pa; virtual_avail += roundup(KSTACK_GUARD_PAGES, colors) * PAGE_SIZE; kstack0 = virtual_avail; virtual_avail += roundup(KSTACK_PAGES, colors) * PAGE_SIZE; if (dcache_color_ignore == 0) KASSERT(DCACHE_COLOR(kstack0) == DCACHE_COLOR(kstack0_phys), ("pmap_bootstrap: kstack0 miscolored")); for (i = 0; i < KSTACK_PAGES; i++) { pa = kstack0_phys + i * PAGE_SIZE; va = kstack0 + i * PAGE_SIZE; tp = tsb_kvtotte(va); vpn = TV_VPN(va, TS_8K); data = TD_V | TD_8K | TD_PA(pa) | TD_REF | TD_SW | TD_CP | TD_CV | TD_P | TD_W; pmap_bootstrap_set_tte(tp, vpn, data); } /* * Calculate the last available physical address. */ for (i = 0; phys_avail[i + 2] != 0; i += 2) ; Maxmem = sparc64_btop(phys_avail[i + 1]); /* * Add the PROM mappings to the kernel TSB. */ if ((vmem = OF_finddevice("/virtual-memory")) == -1) OF_panic("%s: finddevice /virtual-memory", __func__); if ((sz = OF_getproplen(vmem, "translations")) == -1) OF_panic("%s: getproplen translations", __func__); if (sizeof(translations) < sz) OF_panic("%s: translations too small", __func__); bzero(translations, sz); if (OF_getprop(vmem, "translations", translations, sz) == -1) OF_panic("%s: getprop /virtual-memory/translations", __func__); sz /= sizeof(*translations); translations_size = sz; #ifdef DIAGNOSTIC OF_printf("pmap_bootstrap: translations\n"); #endif qsort(translations, sz, sizeof (*translations), om_cmp); for (i = 0; i < sz; i++) { #ifdef DIAGNOSTIC OF_printf("translation: start=%#lx size=%#lx tte=%#lx\n", translations[i].om_start, translations[i].om_size, translations[i].om_tte); #endif if ((translations[i].om_tte & TD_V) == 0) continue; if (translations[i].om_start < VM_MIN_PROM_ADDRESS || translations[i].om_start > VM_MAX_PROM_ADDRESS) continue; for (off = 0; off < translations[i].om_size; off += PAGE_SIZE) { va = translations[i].om_start + off; tp = tsb_kvtotte(va); vpn = TV_VPN(va, TS_8K); data = ((translations[i].om_tte & ~((TD_SOFT2_MASK << TD_SOFT2_SHIFT) | (cpu_impl >= CPU_IMPL_ULTRASPARCI && cpu_impl < CPU_IMPL_ULTRASPARCIII ? (TD_DIAG_SF_MASK << TD_DIAG_SF_SHIFT) : (TD_RSVD_CH_MASK << TD_RSVD_CH_SHIFT)) | (TD_SOFT_MASK << TD_SOFT_SHIFT))) | TD_EXEC) + off; pmap_bootstrap_set_tte(tp, vpn, data); } } /* * Get the available physical memory ranges from /memory/reg. These * are only used for kernel dumps, but it may not be wise to do PROM * calls in that situation. */ if ((sz = OF_getproplen(pmem, "reg")) == -1) OF_panic("%s: getproplen /memory/reg", __func__); if (sizeof(sparc64_memreg) < sz) OF_panic("%s: sparc64_memreg too small", __func__); if (OF_getprop(pmem, "reg", sparc64_memreg, sz) == -1) OF_panic("%s: getprop /memory/reg", __func__); sparc64_nmemreg = sz / sizeof(*sparc64_memreg); /* * Initialize the kernel pmap (which is statically allocated). */ pm = kernel_pmap; PMAP_LOCK_INIT(pm); for (i = 0; i < MAXCPU; i++) pm->pm_context[i] = TLB_CTX_KERNEL; CPU_FILL(&pm->pm_active); /* * Initialize the global tte list lock, which is more commonly * known as the pmap pv global lock. */ rw_init(&tte_list_global_lock, "pmap pv global"); /* * Flush all non-locked TLB entries possibly left over by the * firmware. */ tlb_flush_nonlocked(); } static void pmap_init_qpages(void) { struct pcpu *pc; int i; if (dcache_color_ignore != 0) return; CPU_FOREACH(i) { pc = pcpu_find(i); pc->pc_qmap_addr = kva_alloc(PAGE_SIZE * DCACHE_COLORS); if (pc->pc_qmap_addr == 0) panic("pmap_init_qpages: unable to allocate KVA"); } } SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, pmap_init_qpages, NULL); /* * Map the 4MB kernel TSB pages. */ void pmap_map_tsb(void) { vm_offset_t va; vm_paddr_t pa; u_long data; int i; for (i = 0; i < tsb_kernel_size; i += PAGE_SIZE_4M) { va = (vm_offset_t)tsb_kernel + i; pa = tsb_kernel_phys + i; data = TD_V | TD_4M | TD_PA(pa) | TD_L | TD_CP | TD_CV | TD_P | TD_W; stxa(AA_DMMU_TAR, ASI_DMMU, TLB_TAR_VA(va) | TLB_TAR_CTX(TLB_CTX_KERNEL)); stxa_sync(0, ASI_DTLB_DATA_IN_REG, data); } } /* * Set the secondary context to be the kernel context (needed for FP block * operations in the kernel). */ void pmap_set_kctx(void) { stxa(AA_DMMU_SCXR, ASI_DMMU, (ldxa(AA_DMMU_SCXR, ASI_DMMU) & TLB_CXR_PGSZ_MASK) | TLB_CTX_KERNEL); flush(KERNBASE); } /* * Allocate a physical page of memory directly from the phys_avail map. * Can only be called from pmap_bootstrap before avail start and end are * calculated. */ static vm_paddr_t pmap_bootstrap_alloc(vm_size_t size, uint32_t colors) { vm_paddr_t pa; int i; size = roundup(size, PAGE_SIZE * colors); for (i = 0; phys_avail[i + 1] != 0; i += 2) { if (phys_avail[i + 1] - phys_avail[i] < size) continue; pa = phys_avail[i]; phys_avail[i] += size; return (pa); } OF_panic("%s: no suitable region found", __func__); } /* * Set a TTE. This function is intended as a helper when tsb_kernel is * direct-mapped but we haven't taken over the trap table, yet, as it's the * case when we are taking advantage of ASI_ATOMIC_QUAD_LDD_PHYS to access * the kernel TSB. */ void pmap_bootstrap_set_tte(struct tte *tp, u_long vpn, u_long data) { if (tsb_kernel_ldd_phys == 0) { tp->tte_vpn = vpn; tp->tte_data = data; } else { stxa((vm_paddr_t)tp + offsetof(struct tte, tte_vpn), ASI_PHYS_USE_EC, vpn); stxa((vm_paddr_t)tp + offsetof(struct tte, tte_data), ASI_PHYS_USE_EC, data); } } /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.tte_list); m->md.color = DCACHE_COLOR(VM_PAGE_TO_PHYS(m)); m->md.pmap = NULL; } /* * Initialize the pmap module. */ void pmap_init(void) { vm_offset_t addr; vm_size_t size; int result; int i; for (i = 0; i < translations_size; i++) { addr = translations[i].om_start; size = translations[i].om_size; if ((translations[i].om_tte & TD_V) == 0) continue; if (addr < VM_MIN_PROM_ADDRESS || addr > VM_MAX_PROM_ADDRESS) continue; result = vm_map_find(kernel_map, NULL, 0, &addr, size, 0, VMFS_NO_SPACE, VM_PROT_ALL, VM_PROT_ALL, MAP_NOFAULT); if (result != KERN_SUCCESS || addr != translations[i].om_start) panic("pmap_init: vm_map_find"); } } /* * Extract the physical page address associated with the given * map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pm, vm_offset_t va) { struct tte *tp; vm_paddr_t pa; if (pm == kernel_pmap) return (pmap_kextract(va)); PMAP_LOCK(pm); tp = tsb_tte_lookup(pm, va); if (tp == NULL) pa = 0; else pa = TTE_GET_PA(tp) | (va & TTE_GET_PAGE_MASK(tp)); PMAP_UNLOCK(pm); return (pa); } /* * Atomically extract and hold the physical page with the given * pmap and virtual address pair if that mapping permits the given * protection. */ vm_page_t pmap_extract_and_hold(pmap_t pm, vm_offset_t va, vm_prot_t prot) { struct tte *tp; vm_page_t m; vm_paddr_t pa; m = NULL; pa = 0; PMAP_LOCK(pm); retry: if (pm == kernel_pmap) { if (va >= VM_MIN_DIRECT_ADDRESS) { tp = NULL; m = PHYS_TO_VM_PAGE(TLB_DIRECT_TO_PHYS(va)); (void)vm_page_pa_tryrelock(pm, TLB_DIRECT_TO_PHYS(va), &pa); vm_page_hold(m); } else { tp = tsb_kvtotte(va); if ((tp->tte_data & TD_V) == 0) tp = NULL; } } else tp = tsb_tte_lookup(pm, va); if (tp != NULL && ((tp->tte_data & TD_SW) || (prot & VM_PROT_WRITE) == 0)) { if (vm_page_pa_tryrelock(pm, TTE_GET_PA(tp), &pa)) goto retry; m = PHYS_TO_VM_PAGE(TTE_GET_PA(tp)); vm_page_hold(m); } PA_UNLOCK_COND(pa); PMAP_UNLOCK(pm); return (m); } /* * Extract the physical page address associated with the given kernel virtual * address. */ vm_paddr_t pmap_kextract(vm_offset_t va) { struct tte *tp; if (va >= VM_MIN_DIRECT_ADDRESS) return (TLB_DIRECT_TO_PHYS(va)); tp = tsb_kvtotte(va); if ((tp->tte_data & TD_V) == 0) return (0); return (TTE_GET_PA(tp) | (va & TTE_GET_PAGE_MASK(tp))); } int pmap_cache_enter(vm_page_t m, vm_offset_t va) { struct tte *tp; int color; rw_assert(&tte_list_global_lock, RA_WLOCKED); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("pmap_cache_enter: fake page")); PMAP_STATS_INC(pmap_ncache_enter); if (dcache_color_ignore != 0) return (1); /* * Find the color for this virtual address and note the added mapping. */ color = DCACHE_COLOR(va); m->md.colors[color]++; /* * If all existing mappings have the same color, the mapping is * cacheable. */ if (m->md.color == color) { KASSERT(m->md.colors[DCACHE_OTHER_COLOR(color)] == 0, ("pmap_cache_enter: cacheable, mappings of other color")); if (m->md.color == DCACHE_COLOR(VM_PAGE_TO_PHYS(m))) PMAP_STATS_INC(pmap_ncache_enter_c); else PMAP_STATS_INC(pmap_ncache_enter_oc); return (1); } /* * If there are no mappings of the other color, and the page still has * the wrong color, this must be a new mapping. Change the color to * match the new mapping, which is cacheable. We must flush the page * from the cache now. */ if (m->md.colors[DCACHE_OTHER_COLOR(color)] == 0) { KASSERT(m->md.colors[color] == 1, ("pmap_cache_enter: changing color, not new mapping")); dcache_page_inval(VM_PAGE_TO_PHYS(m)); m->md.color = color; if (m->md.color == DCACHE_COLOR(VM_PAGE_TO_PHYS(m))) PMAP_STATS_INC(pmap_ncache_enter_cc); else PMAP_STATS_INC(pmap_ncache_enter_coc); return (1); } /* * If the mapping is already non-cacheable, just return. */ if (m->md.color == -1) { PMAP_STATS_INC(pmap_ncache_enter_nc); return (0); } PMAP_STATS_INC(pmap_ncache_enter_cnc); /* * Mark all mappings as uncacheable, flush any lines with the other * color out of the dcache, and set the color to none (-1). */ TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { atomic_clear_long(&tp->tte_data, TD_CV); tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); } dcache_page_inval(VM_PAGE_TO_PHYS(m)); m->md.color = -1; return (0); } static void pmap_cache_remove(vm_page_t m, vm_offset_t va) { struct tte *tp; int color; rw_assert(&tte_list_global_lock, RA_WLOCKED); CTR3(KTR_PMAP, "pmap_cache_remove: m=%p va=%#lx c=%d", m, va, m->md.colors[DCACHE_COLOR(va)]); KASSERT((m->flags & PG_FICTITIOUS) == 0, ("pmap_cache_remove: fake page")); PMAP_STATS_INC(pmap_ncache_remove); if (dcache_color_ignore != 0) return; KASSERT(m->md.colors[DCACHE_COLOR(va)] > 0, ("pmap_cache_remove: no mappings %d <= 0", m->md.colors[DCACHE_COLOR(va)])); /* * Find the color for this virtual address and note the removal of * the mapping. */ color = DCACHE_COLOR(va); m->md.colors[color]--; /* * If the page is cacheable, just return and keep the same color, even * if there are no longer any mappings. */ if (m->md.color != -1) { if (m->md.color == DCACHE_COLOR(VM_PAGE_TO_PHYS(m))) PMAP_STATS_INC(pmap_ncache_remove_c); else PMAP_STATS_INC(pmap_ncache_remove_oc); return; } KASSERT(m->md.colors[DCACHE_OTHER_COLOR(color)] != 0, ("pmap_cache_remove: uncacheable, no mappings of other color")); /* * If the page is not cacheable (color is -1), and the number of * mappings for this color is not zero, just return. There are * mappings of the other color still, so remain non-cacheable. */ if (m->md.colors[color] != 0) { PMAP_STATS_INC(pmap_ncache_remove_nc); return; } /* * The number of mappings for this color is now zero. Recache the * other colored mappings, and change the page color to the other * color. There should be no lines in the data cache for this page, * so flushing should not be needed. */ TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { atomic_set_long(&tp->tte_data, TD_CV); tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); } m->md.color = DCACHE_OTHER_COLOR(color); if (m->md.color == DCACHE_COLOR(VM_PAGE_TO_PHYS(m))) PMAP_STATS_INC(pmap_ncache_remove_cc); else PMAP_STATS_INC(pmap_ncache_remove_coc); } /* * Map a wired page into kernel virtual address space. */ void pmap_kenter(vm_offset_t va, vm_page_t m) { vm_offset_t ova; struct tte *tp; vm_page_t om; u_long data; rw_assert(&tte_list_global_lock, RA_WLOCKED); PMAP_STATS_INC(pmap_nkenter); tp = tsb_kvtotte(va); CTR4(KTR_PMAP, "pmap_kenter: va=%#lx pa=%#lx tp=%p data=%#lx", va, VM_PAGE_TO_PHYS(m), tp, tp->tte_data); if (DCACHE_COLOR(VM_PAGE_TO_PHYS(m)) != DCACHE_COLOR(va)) { CTR5(KTR_SPARE2, "pmap_kenter: off color va=%#lx pa=%#lx o=%p ot=%d pi=%#lx", va, VM_PAGE_TO_PHYS(m), m->object, m->object ? m->object->type : -1, m->pindex); PMAP_STATS_INC(pmap_nkenter_oc); } if ((tp->tte_data & TD_V) != 0) { om = PHYS_TO_VM_PAGE(TTE_GET_PA(tp)); ova = TTE_GET_VA(tp); if (m == om && va == ova) { PMAP_STATS_INC(pmap_nkenter_stupid); return; } TAILQ_REMOVE(&om->md.tte_list, tp, tte_link); pmap_cache_remove(om, ova); if (va != ova) tlb_page_demap(kernel_pmap, ova); } data = TD_V | TD_8K | VM_PAGE_TO_PHYS(m) | TD_REF | TD_SW | TD_CP | TD_P | TD_W; if (pmap_cache_enter(m, va) != 0) data |= TD_CV; tp->tte_vpn = TV_VPN(va, TS_8K); tp->tte_data = data; TAILQ_INSERT_TAIL(&m->md.tte_list, tp, tte_link); } /* * Map a wired page into kernel virtual address space. This additionally * takes a flag argument which is or'ed to the TTE data. This is used by * sparc64_bus_mem_map(). * NOTE: if the mapping is non-cacheable, it's the caller's responsibility * to flush entries that might still be in the cache, if applicable. */ void pmap_kenter_flags(vm_offset_t va, vm_paddr_t pa, u_long flags) { struct tte *tp; tp = tsb_kvtotte(va); CTR4(KTR_PMAP, "pmap_kenter_flags: va=%#lx pa=%#lx tp=%p data=%#lx", va, pa, tp, tp->tte_data); tp->tte_vpn = TV_VPN(va, TS_8K); tp->tte_data = TD_V | TD_8K | TD_PA(pa) | TD_REF | TD_P | flags; } /* * Remove a wired page from kernel virtual address space. */ void pmap_kremove(vm_offset_t va) { struct tte *tp; vm_page_t m; rw_assert(&tte_list_global_lock, RA_WLOCKED); PMAP_STATS_INC(pmap_nkremove); tp = tsb_kvtotte(va); CTR3(KTR_PMAP, "pmap_kremove: va=%#lx tp=%p data=%#lx", va, tp, tp->tte_data); if ((tp->tte_data & TD_V) == 0) return; m = PHYS_TO_VM_PAGE(TTE_GET_PA(tp)); TAILQ_REMOVE(&m->md.tte_list, tp, tte_link); pmap_cache_remove(m, va); TTE_ZERO(tp); } /* * Inverse of pmap_kenter_flags, used by bus_space_unmap(). */ void pmap_kremove_flags(vm_offset_t va) { struct tte *tp; tp = tsb_kvtotte(va); CTR3(KTR_PMAP, "pmap_kremove_flags: va=%#lx tp=%p data=%#lx", va, tp, tp->tte_data); TTE_ZERO(tp); } /* * Map a range of physical addresses into kernel virtual address space. * * The value passed in *virt is a suggested virtual address for the mapping. * Architectures which can support a direct-mapped physical to virtual region * can return the appropriate address within that region, leaving '*virt' * unchanged. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return (TLB_PHYS_TO_DIRECT(start)); } /* * Map a list of wired pages into kernel virtual address space. This is * intended for temporary mappings which do not need page modification or * references recorded. Existing mappings in the region are overwritten. */ void pmap_qenter(vm_offset_t sva, vm_page_t *m, int count) { vm_offset_t va; PMAP_STATS_INC(pmap_nqenter); va = sva; rw_wlock(&tte_list_global_lock); while (count-- > 0) { pmap_kenter(va, *m); va += PAGE_SIZE; m++; } rw_wunlock(&tte_list_global_lock); tlb_range_demap(kernel_pmap, sva, va); } /* * Remove page mappings from kernel virtual address space. Intended for * temporary mappings entered by pmap_qenter. */ void pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; PMAP_STATS_INC(pmap_nqremove); va = sva; rw_wlock(&tte_list_global_lock); while (count-- > 0) { pmap_kremove(va); va += PAGE_SIZE; } rw_wunlock(&tte_list_global_lock); tlb_range_demap(kernel_pmap, sva, va); } /* * Initialize the pmap associated with process 0. */ void pmap_pinit0(pmap_t pm) { int i; PMAP_LOCK_INIT(pm); for (i = 0; i < MAXCPU; i++) pm->pm_context[i] = TLB_CTX_KERNEL; CPU_ZERO(&pm->pm_active); pm->pm_tsb = NULL; pm->pm_tsb_obj = NULL; bzero(&pm->pm_stats, sizeof(pm->pm_stats)); } /* * Initialize a preallocated and zeroed pmap structure, such as one in a * vmspace structure. */ int pmap_pinit(pmap_t pm) { vm_page_t ma[TSB_PAGES]; vm_page_t m; int i; /* * Allocate KVA space for the TSB. */ if (pm->pm_tsb == NULL) { pm->pm_tsb = (struct tte *)kva_alloc(TSB_BSIZE); if (pm->pm_tsb == NULL) return (0); } /* * Allocate an object for it. */ if (pm->pm_tsb_obj == NULL) pm->pm_tsb_obj = vm_object_allocate(OBJT_PHYS, TSB_PAGES); for (i = 0; i < MAXCPU; i++) pm->pm_context[i] = -1; CPU_ZERO(&pm->pm_active); VM_OBJECT_WLOCK(pm->pm_tsb_obj); for (i = 0; i < TSB_PAGES; i++) { m = vm_page_grab(pm->pm_tsb_obj, i, VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | VM_ALLOC_ZERO); m->valid = VM_PAGE_BITS_ALL; m->md.pmap = pm; ma[i] = m; } VM_OBJECT_WUNLOCK(pm->pm_tsb_obj); pmap_qenter((vm_offset_t)pm->pm_tsb, ma, TSB_PAGES); bzero(&pm->pm_stats, sizeof(pm->pm_stats)); return (1); } /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pm) { vm_object_t obj; vm_page_t m; #ifdef SMP struct pcpu *pc; #endif CTR2(KTR_PMAP, "pmap_release: ctx=%#x tsb=%p", pm->pm_context[curcpu], pm->pm_tsb); KASSERT(pmap_resident_count(pm) == 0, ("pmap_release: resident pages %ld != 0", pmap_resident_count(pm))); /* * After the pmap was freed, it might be reallocated to a new process. * When switching, this might lead us to wrongly assume that we need * not switch contexts because old and new pmap pointer are equal. * Therefore, make sure that this pmap is not referenced by any PCPU * pointer any more. This could happen in two cases: * - A process that referenced the pmap is currently exiting on a CPU. * However, it is guaranteed to not switch in any more after setting * its state to PRS_ZOMBIE. * - A process that referenced this pmap ran on a CPU, but we switched * to a kernel thread, leaving the pmap pointer unchanged. */ #ifdef SMP sched_pin(); STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) atomic_cmpset_rel_ptr((uintptr_t *)&pc->pc_pmap, (uintptr_t)pm, (uintptr_t)NULL); sched_unpin(); #else critical_enter(); if (PCPU_GET(pmap) == pm) PCPU_SET(pmap, NULL); critical_exit(); #endif pmap_qremove((vm_offset_t)pm->pm_tsb, TSB_PAGES); obj = pm->pm_tsb_obj; VM_OBJECT_WLOCK(obj); KASSERT(obj->ref_count == 1, ("pmap_release: tsbobj ref count != 1")); while (!TAILQ_EMPTY(&obj->memq)) { m = TAILQ_FIRST(&obj->memq); m->md.pmap = NULL; m->wire_count--; atomic_subtract_int(&vm_cnt.v_wire_count, 1); vm_page_free_zero(m); } VM_OBJECT_WUNLOCK(obj); } /* * Grow the number of kernel page table entries. Unneeded. */ void pmap_growkernel(vm_offset_t addr) { panic("pmap_growkernel: can't grow kernel"); } int pmap_remove_tte(struct pmap *pm, struct pmap *pm2, struct tte *tp, vm_offset_t va) { vm_page_t m; u_long data; rw_assert(&tte_list_global_lock, RA_WLOCKED); data = atomic_readandclear_long(&tp->tte_data); if ((data & TD_FAKE) == 0) { m = PHYS_TO_VM_PAGE(TD_PA(data)); TAILQ_REMOVE(&m->md.tte_list, tp, tte_link); if ((data & TD_WIRED) != 0) pm->pm_stats.wired_count--; if ((data & TD_PV) != 0) { if ((data & TD_W) != 0) vm_page_dirty(m); if ((data & TD_REF) != 0) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.tte_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); pm->pm_stats.resident_count--; } pmap_cache_remove(m, va); } TTE_ZERO(tp); if (PMAP_REMOVE_DONE(pm)) return (0); return (1); } /* * Remove the given range of addresses from the specified map. */ void pmap_remove(pmap_t pm, vm_offset_t start, vm_offset_t end) { struct tte *tp; vm_offset_t va; CTR3(KTR_PMAP, "pmap_remove: ctx=%#lx start=%#lx end=%#lx", pm->pm_context[curcpu], start, end); if (PMAP_REMOVE_DONE(pm)) return; rw_wlock(&tte_list_global_lock); PMAP_LOCK(pm); if (end - start > PMAP_TSB_THRESH) { tsb_foreach(pm, NULL, start, end, pmap_remove_tte); tlb_context_demap(pm); } else { for (va = start; va < end; va += PAGE_SIZE) if ((tp = tsb_tte_lookup(pm, va)) != NULL && !pmap_remove_tte(pm, NULL, tp, va)) break; tlb_range_demap(pm, start, end - 1); } PMAP_UNLOCK(pm); rw_wunlock(&tte_list_global_lock); } void pmap_remove_all(vm_page_t m) { struct pmap *pm; struct tte *tpn; struct tte *tp; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); rw_wlock(&tte_list_global_lock); for (tp = TAILQ_FIRST(&m->md.tte_list); tp != NULL; tp = tpn) { tpn = TAILQ_NEXT(tp, tte_link); if ((tp->tte_data & TD_PV) == 0) continue; pm = TTE_GET_PMAP(tp); va = TTE_GET_VA(tp); PMAP_LOCK(pm); if ((tp->tte_data & TD_WIRED) != 0) pm->pm_stats.wired_count--; if ((tp->tte_data & TD_REF) != 0) vm_page_aflag_set(m, PGA_REFERENCED); if ((tp->tte_data & TD_W) != 0) vm_page_dirty(m); tp->tte_data &= ~TD_V; tlb_page_demap(pm, va); TAILQ_REMOVE(&m->md.tte_list, tp, tte_link); pm->pm_stats.resident_count--; pmap_cache_remove(m, va); TTE_ZERO(tp); PMAP_UNLOCK(pm); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&tte_list_global_lock); } static int pmap_protect_tte(struct pmap *pm, struct pmap *pm2, struct tte *tp, vm_offset_t va) { u_long data; vm_page_t m; PMAP_LOCK_ASSERT(pm, MA_OWNED); data = atomic_clear_long(&tp->tte_data, TD_SW | TD_W); if ((data & (TD_PV | TD_W)) == (TD_PV | TD_W)) { m = PHYS_TO_VM_PAGE(TD_PA(data)); vm_page_dirty(m); } return (1); } /* * Set the physical protection on the specified range of this map as requested. */ void pmap_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t va; struct tte *tp; CTR4(KTR_PMAP, "pmap_protect: ctx=%#lx sva=%#lx eva=%#lx prot=%#lx", pm->pm_context[curcpu], sva, eva, prot); if ((prot & VM_PROT_READ) == VM_PROT_NONE) { pmap_remove(pm, sva, eva); return; } if (prot & VM_PROT_WRITE) return; PMAP_LOCK(pm); if (eva - sva > PMAP_TSB_THRESH) { tsb_foreach(pm, NULL, sva, eva, pmap_protect_tte); tlb_context_demap(pm); } else { for (va = sva; va < eva; va += PAGE_SIZE) if ((tp = tsb_tte_lookup(pm, va)) != NULL) pmap_protect_tte(pm, NULL, tp, va); tlb_range_demap(pm, sva, eva - 1); } PMAP_UNLOCK(pm); } /* * Map the given physical page at the specified virtual address in the * target pmap with the protection requested. If specified the page * will be wired down. */ int pmap_enter(pmap_t pm, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { int rv; rw_wlock(&tte_list_global_lock); PMAP_LOCK(pm); rv = pmap_enter_locked(pm, va, m, prot, flags, psind); rw_wunlock(&tte_list_global_lock); PMAP_UNLOCK(pm); return (rv); } /* * Map the given physical page at the specified virtual address in the * target pmap with the protection requested. If specified the page * will be wired down. * * The page queues and pmap must be locked. */ static int pmap_enter_locked(pmap_t pm, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind __unused) { struct tte *tp; vm_paddr_t pa; vm_page_t real; u_long data; boolean_t wired; rw_assert(&tte_list_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pm, MA_OWNED); if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m)) VM_OBJECT_ASSERT_LOCKED(m->object); PMAP_STATS_INC(pmap_nenter); pa = VM_PAGE_TO_PHYS(m); wired = (flags & PMAP_ENTER_WIRED) != 0; /* * If this is a fake page from the device_pager, but it covers actual * physical memory, convert to the real backing page. */ if ((m->flags & PG_FICTITIOUS) != 0) { real = vm_phys_paddr_to_vm_page(pa); if (real != NULL) m = real; } CTR6(KTR_PMAP, "pmap_enter_locked: ctx=%p m=%p va=%#lx pa=%#lx prot=%#x wired=%d", pm->pm_context[curcpu], m, va, pa, prot, wired); /* * If there is an existing mapping, and the physical address has not * changed, must be protection or wiring change. */ if ((tp = tsb_tte_lookup(pm, va)) != NULL && TTE_GET_PA(tp) == pa) { CTR0(KTR_PMAP, "pmap_enter_locked: update"); PMAP_STATS_INC(pmap_nenter_update); /* * Wiring change, just update stats. */ if (wired) { if ((tp->tte_data & TD_WIRED) == 0) { tp->tte_data |= TD_WIRED; pm->pm_stats.wired_count++; } } else { if ((tp->tte_data & TD_WIRED) != 0) { tp->tte_data &= ~TD_WIRED; pm->pm_stats.wired_count--; } } /* * Save the old bits and clear the ones we're interested in. */ data = tp->tte_data; tp->tte_data &= ~(TD_EXEC | TD_SW | TD_W); /* * If we're turning off write permissions, sense modify status. */ if ((prot & VM_PROT_WRITE) != 0) { tp->tte_data |= TD_SW; if (wired) tp->tte_data |= TD_W; if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_WRITEABLE); } else if ((data & TD_W) != 0) vm_page_dirty(m); /* * If we're turning on execute permissions, flush the icache. */ if ((prot & VM_PROT_EXECUTE) != 0) { if ((data & TD_EXEC) == 0) icache_page_inval(pa); tp->tte_data |= TD_EXEC; } /* * Delete the old mapping. */ tlb_page_demap(pm, TTE_GET_VA(tp)); } else { /* * If there is an existing mapping, but its for a different * physical address, delete the old mapping. */ if (tp != NULL) { CTR0(KTR_PMAP, "pmap_enter_locked: replace"); PMAP_STATS_INC(pmap_nenter_replace); pmap_remove_tte(pm, NULL, tp, va); tlb_page_demap(pm, va); } else { CTR0(KTR_PMAP, "pmap_enter_locked: new"); PMAP_STATS_INC(pmap_nenter_new); } /* * Now set up the data and install the new mapping. */ data = TD_V | TD_8K | TD_PA(pa); if (pm == kernel_pmap) data |= TD_P; if ((prot & VM_PROT_WRITE) != 0) { data |= TD_SW; if ((m->oflags & VPO_UNMANAGED) == 0) vm_page_aflag_set(m, PGA_WRITEABLE); } if (prot & VM_PROT_EXECUTE) { data |= TD_EXEC; icache_page_inval(pa); } /* * If its wired update stats. We also don't need reference or * modify tracking for wired mappings, so set the bits now. */ if (wired) { pm->pm_stats.wired_count++; data |= TD_REF | TD_WIRED; if ((prot & VM_PROT_WRITE) != 0) data |= TD_W; } tsb_tte_enter(pm, m, va, TS_8K, data); } return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_page_t m; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); m = m_start; rw_wlock(&tte_list_global_lock); PMAP_LOCK(pm); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { pmap_enter_locked(pm, start + ptoa(diff), m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), 0, 0); m = TAILQ_NEXT(m, listq); } rw_wunlock(&tte_list_global_lock); PMAP_UNLOCK(pm); } void pmap_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&tte_list_global_lock); PMAP_LOCK(pm); pmap_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), 0, 0); rw_wunlock(&tte_list_global_lock); PMAP_UNLOCK(pm); } void pmap_object_init_pt(pmap_t pm, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); } static int pmap_unwire_tte(pmap_t pm, pmap_t pm2, struct tte *tp, vm_offset_t va) { PMAP_LOCK_ASSERT(pm, MA_OWNED); if ((tp->tte_data & TD_WIRED) == 0) panic("pmap_unwire_tte: tp %p is missing TD_WIRED", tp); atomic_clear_long(&tp->tte_data, TD_WIRED); pm->pm_stats.wired_count--; return (1); } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range must * have the wired attribute set. In contrast, invalid mappings cannot have * the wired attribute set, so they are ignored. * * The wired attribute of the translation table entry is not a hardware * feature, so there is no need to invalidate any TLB entries. */ void pmap_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va; struct tte *tp; PMAP_LOCK(pm); if (eva - sva > PMAP_TSB_THRESH) tsb_foreach(pm, NULL, sva, eva, pmap_unwire_tte); else { for (va = sva; va < eva; va += PAGE_SIZE) if ((tp = tsb_tte_lookup(pm, va)) != NULL) pmap_unwire_tte(pm, NULL, tp, va); } PMAP_UNLOCK(pm); } static int pmap_copy_tte(pmap_t src_pmap, pmap_t dst_pmap, struct tte *tp, vm_offset_t va) { vm_page_t m; u_long data; if ((tp->tte_data & TD_FAKE) != 0) return (1); if (tsb_tte_lookup(dst_pmap, va) == NULL) { data = tp->tte_data & ~(TD_PV | TD_REF | TD_SW | TD_CV | TD_W); m = PHYS_TO_VM_PAGE(TTE_GET_PA(tp)); tsb_tte_enter(dst_pmap, m, va, TS_8K, data); } return (1); } void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct tte *tp; vm_offset_t va; if (dst_addr != src_addr) return; rw_wlock(&tte_list_global_lock); if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } if (len > PMAP_TSB_THRESH) { tsb_foreach(src_pmap, dst_pmap, src_addr, src_addr + len, pmap_copy_tte); tlb_context_demap(dst_pmap); } else { for (va = src_addr; va < src_addr + len; va += PAGE_SIZE) if ((tp = tsb_tte_lookup(src_pmap, va)) != NULL) pmap_copy_tte(src_pmap, dst_pmap, tp, va); tlb_range_demap(dst_pmap, src_addr, src_addr + len - 1); } rw_wunlock(&tte_list_global_lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } void pmap_zero_page(vm_page_t m) { struct tte *tp; vm_offset_t va; vm_paddr_t pa; KASSERT((m->flags & PG_FICTITIOUS) == 0, ("pmap_zero_page: fake page")); PMAP_STATS_INC(pmap_nzero_page); pa = VM_PAGE_TO_PHYS(m); if (dcache_color_ignore != 0 || m->md.color == DCACHE_COLOR(pa)) { PMAP_STATS_INC(pmap_nzero_page_c); va = TLB_PHYS_TO_DIRECT(pa); cpu_block_zero((void *)va, PAGE_SIZE); } else if (m->md.color == -1) { PMAP_STATS_INC(pmap_nzero_page_nc); aszero(ASI_PHYS_USE_EC, pa, PAGE_SIZE); } else { PMAP_STATS_INC(pmap_nzero_page_oc); PMAP_LOCK(kernel_pmap); va = pmap_temp_map_1 + (m->md.color * PAGE_SIZE); tp = tsb_kvtotte(va); tp->tte_data = TD_V | TD_8K | TD_PA(pa) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(va, TS_8K); cpu_block_zero((void *)va, PAGE_SIZE); tlb_page_demap(kernel_pmap, va); PMAP_UNLOCK(kernel_pmap); } } void pmap_zero_page_area(vm_page_t m, int off, int size) { struct tte *tp; vm_offset_t va; vm_paddr_t pa; KASSERT((m->flags & PG_FICTITIOUS) == 0, ("pmap_zero_page_area: fake page")); KASSERT(off + size <= PAGE_SIZE, ("pmap_zero_page_area: bad off/size")); PMAP_STATS_INC(pmap_nzero_page_area); pa = VM_PAGE_TO_PHYS(m); if (dcache_color_ignore != 0 || m->md.color == DCACHE_COLOR(pa)) { PMAP_STATS_INC(pmap_nzero_page_area_c); va = TLB_PHYS_TO_DIRECT(pa); bzero((void *)(va + off), size); } else if (m->md.color == -1) { PMAP_STATS_INC(pmap_nzero_page_area_nc); aszero(ASI_PHYS_USE_EC, pa + off, size); } else { PMAP_STATS_INC(pmap_nzero_page_area_oc); PMAP_LOCK(kernel_pmap); va = pmap_temp_map_1 + (m->md.color * PAGE_SIZE); tp = tsb_kvtotte(va); tp->tte_data = TD_V | TD_8K | TD_PA(pa) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(va, TS_8K); bzero((void *)(va + off), size); tlb_page_demap(kernel_pmap, va); PMAP_UNLOCK(kernel_pmap); } } void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t vdst; vm_offset_t vsrc; vm_paddr_t pdst; vm_paddr_t psrc; struct tte *tp; KASSERT((mdst->flags & PG_FICTITIOUS) == 0, ("pmap_copy_page: fake dst page")); KASSERT((msrc->flags & PG_FICTITIOUS) == 0, ("pmap_copy_page: fake src page")); PMAP_STATS_INC(pmap_ncopy_page); pdst = VM_PAGE_TO_PHYS(mdst); psrc = VM_PAGE_TO_PHYS(msrc); if (dcache_color_ignore != 0 || (msrc->md.color == DCACHE_COLOR(psrc) && mdst->md.color == DCACHE_COLOR(pdst))) { PMAP_STATS_INC(pmap_ncopy_page_c); vdst = TLB_PHYS_TO_DIRECT(pdst); vsrc = TLB_PHYS_TO_DIRECT(psrc); cpu_block_copy((void *)vsrc, (void *)vdst, PAGE_SIZE); } else if (msrc->md.color == -1 && mdst->md.color == -1) { PMAP_STATS_INC(pmap_ncopy_page_nc); ascopy(ASI_PHYS_USE_EC, psrc, pdst, PAGE_SIZE); } else if (msrc->md.color == -1) { if (mdst->md.color == DCACHE_COLOR(pdst)) { PMAP_STATS_INC(pmap_ncopy_page_dc); vdst = TLB_PHYS_TO_DIRECT(pdst); ascopyfrom(ASI_PHYS_USE_EC, psrc, (void *)vdst, PAGE_SIZE); } else { PMAP_STATS_INC(pmap_ncopy_page_doc); PMAP_LOCK(kernel_pmap); vdst = pmap_temp_map_1 + (mdst->md.color * PAGE_SIZE); tp = tsb_kvtotte(vdst); tp->tte_data = TD_V | TD_8K | TD_PA(pdst) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(vdst, TS_8K); ascopyfrom(ASI_PHYS_USE_EC, psrc, (void *)vdst, PAGE_SIZE); tlb_page_demap(kernel_pmap, vdst); PMAP_UNLOCK(kernel_pmap); } } else if (mdst->md.color == -1) { if (msrc->md.color == DCACHE_COLOR(psrc)) { PMAP_STATS_INC(pmap_ncopy_page_sc); vsrc = TLB_PHYS_TO_DIRECT(psrc); ascopyto((void *)vsrc, ASI_PHYS_USE_EC, pdst, PAGE_SIZE); } else { PMAP_STATS_INC(pmap_ncopy_page_soc); PMAP_LOCK(kernel_pmap); vsrc = pmap_temp_map_1 + (msrc->md.color * PAGE_SIZE); tp = tsb_kvtotte(vsrc); tp->tte_data = TD_V | TD_8K | TD_PA(psrc) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(vsrc, TS_8K); ascopyto((void *)vsrc, ASI_PHYS_USE_EC, pdst, PAGE_SIZE); tlb_page_demap(kernel_pmap, vsrc); PMAP_UNLOCK(kernel_pmap); } } else { PMAP_STATS_INC(pmap_ncopy_page_oc); PMAP_LOCK(kernel_pmap); vdst = pmap_temp_map_1 + (mdst->md.color * PAGE_SIZE); tp = tsb_kvtotte(vdst); tp->tte_data = TD_V | TD_8K | TD_PA(pdst) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(vdst, TS_8K); vsrc = pmap_temp_map_2 + (msrc->md.color * PAGE_SIZE); tp = tsb_kvtotte(vsrc); tp->tte_data = TD_V | TD_8K | TD_PA(psrc) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(vsrc, TS_8K); cpu_block_copy((void *)vsrc, (void *)vdst, PAGE_SIZE); tlb_page_demap(kernel_pmap, vdst); tlb_page_demap(kernel_pmap, vsrc); PMAP_UNLOCK(kernel_pmap); } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { vm_paddr_t pa; vm_offset_t qaddr; struct tte *tp; pa = VM_PAGE_TO_PHYS(m); if (dcache_color_ignore != 0 || m->md.color == DCACHE_COLOR(pa)) return (TLB_PHYS_TO_DIRECT(pa)); critical_enter(); qaddr = PCPU_GET(qmap_addr); qaddr += (PAGE_SIZE * ((DCACHE_COLORS + DCACHE_COLOR(pa) - DCACHE_COLOR(qaddr)) % DCACHE_COLORS)); tp = tsb_kvtotte(qaddr); KASSERT(tp->tte_data == 0, ("pmap_quick_enter_page: PTE busy")); tp->tte_data = TD_V | TD_8K | TD_PA(pa) | TD_CP | TD_CV | TD_W; tp->tte_vpn = TV_VPN(qaddr, TS_8K); return (qaddr); } void pmap_quick_remove_page(vm_offset_t addr) { vm_offset_t qaddr; struct tte *tp; if (addr >= VM_MIN_DIRECT_ADDRESS) return; tp = tsb_kvtotte(addr); qaddr = PCPU_GET(qmap_addr); KASSERT((addr >= qaddr) && (addr < (qaddr + (PAGE_SIZE * DCACHE_COLORS))), ("pmap_quick_remove_page: invalid address")); KASSERT(tp->tte_data != 0, ("pmap_quick_remove_page: PTE not in use")); stxa(TLB_DEMAP_VA(addr) | TLB_DEMAP_NUCLEUS | TLB_DEMAP_PAGE, ASI_DMMU_DEMAP, 0); stxa(TLB_DEMAP_VA(addr) | TLB_DEMAP_NUCLEUS | TLB_DEMAP_PAGE, ASI_IMMU_DEMAP, 0); flush(KERNBASE); TTE_ZERO(tp); critical_exit(); } int unmapped_buf_allowed; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { panic("pmap_copy_pages: not implemented"); } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pm, vm_page_t m) { struct tte *tp; int loops; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); loops = 0; rv = FALSE; rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; if (TTE_GET_PMAP(tp) == pm) { rv = TRUE; break; } if (++loops >= 16) break; } rw_wunlock(&tte_list_global_lock); return (rv); } /* * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct tte *tp; int count; count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) if ((tp->tte_data & (TD_PV | TD_WIRED)) == (TD_PV | TD_WIRED)) count++; rw_wunlock(&tte_list_global_lock); return (count); } /* * Remove all pages from specified address space, this aids process exit * speeds. This is much faster than pmap_remove in the case of running down * an entire address space. Only works for the current pmap. */ void pmap_remove_pages(pmap_t pm) { } /* * Returns TRUE if the given page has a managed mapping. */ boolean_t pmap_page_is_mapped(vm_page_t m) { struct tte *tp; boolean_t rv; rv = FALSE; if ((m->oflags & VPO_UNMANAGED) != 0) return (rv); rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) if ((tp->tte_data & TD_PV) != 0) { rv = TRUE; break; } rw_wunlock(&tte_list_global_lock); return (rv); } -#define PMAP_TS_REFERENCED_MAX 5 - /* * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. - * - * XXX: The exact number of bits to check and clear is a matter that - * should be tested and standardized at some point in the future for - * optimal aging of shared pages. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). */ int pmap_ts_referenced(vm_page_t m) { struct tte *tpf; struct tte *tpn; struct tte *tp; u_long data; int count; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); count = 0; rw_wlock(&tte_list_global_lock); if ((tp = TAILQ_FIRST(&m->md.tte_list)) != NULL) { tpf = tp; do { tpn = TAILQ_NEXT(tp, tte_link); TAILQ_REMOVE(&m->md.tte_list, tp, tte_link); TAILQ_INSERT_TAIL(&m->md.tte_list, tp, tte_link); if ((tp->tte_data & TD_PV) == 0) continue; data = atomic_clear_long(&tp->tte_data, TD_REF); if ((data & TD_W) != 0) vm_page_dirty(m); if ((data & TD_REF) != 0 && ++count >= PMAP_TS_REFERENCED_MAX) break; } while ((tp = tpn) != NULL && tp != tpf); } rw_wunlock(&tte_list_global_lock); return (count); } boolean_t pmap_is_modified(vm_page_t m) { struct tte *tp; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); rv = FALSE; /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * concurrently set while the object is locked. Thus, if PGA_WRITEABLE * is clear, no TTEs can have TD_W set. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return (rv); rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; if ((tp->tte_data & TD_W) != 0) { rv = TRUE; break; } } rw_wunlock(&tte_list_global_lock); return (rv); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is elgible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { boolean_t rv; PMAP_LOCK(pmap); rv = tsb_tte_lookup(pmap, addr) == NULL; PMAP_UNLOCK(pmap); return (rv); } /* * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { struct tte *tp; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); rv = FALSE; rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; if ((tp->tte_data & TD_REF) != 0) { rv = TRUE; break; } } rw_wunlock(&tte_list_global_lock); return (rv); } /* * This function is advisory. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { } void pmap_clear_modify(vm_page_t m) { struct tte *tp; u_long data; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); VM_OBJECT_ASSERT_WLOCKED(m->object); KASSERT(!vm_page_xbusied(m), ("pmap_clear_modify: page %p is exclusive busied", m)); /* * If the page is not PGA_WRITEABLE, then no TTEs can have TD_W set. * If the object containing the page is locked and the page is not * exclusive busied, then PGA_WRITEABLE cannot be concurrently set. */ if ((m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; data = atomic_clear_long(&tp->tte_data, TD_W); if ((data & TD_W) != 0) tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); } rw_wunlock(&tte_list_global_lock); } void pmap_remove_write(vm_page_t m) { struct tte *tp; u_long data; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); /* * If the page is not exclusive busied, then PGA_WRITEABLE cannot be * set by another thread while the object is locked. Thus, * if PGA_WRITEABLE is clear, no page table entries need updating. */ VM_OBJECT_ASSERT_WLOCKED(m->object); if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0) return; rw_wlock(&tte_list_global_lock); TAILQ_FOREACH(tp, &m->md.tte_list, tte_link) { if ((tp->tte_data & TD_PV) == 0) continue; data = atomic_clear_long(&tp->tte_data, TD_SW | TD_W); if ((data & TD_W) != 0) { vm_page_dirty(m); tlb_page_demap(TTE_GET_PMAP(tp), TTE_GET_VA(tp)); } } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(&tte_list_global_lock); } int pmap_mincore(pmap_t pm, vm_offset_t addr, vm_paddr_t *locked_pa) { /* TODO; */ return (0); } /* * Activate a user pmap. The pmap must be activated before its address space * can be accessed in any way. */ void pmap_activate(struct thread *td) { struct vmspace *vm; struct pmap *pm; int context; critical_enter(); vm = td->td_proc->p_vmspace; pm = vmspace_pmap(vm); context = PCPU_GET(tlb_ctx); if (context == PCPU_GET(tlb_ctx_max)) { tlb_flush_user(); context = PCPU_GET(tlb_ctx_min); } PCPU_SET(tlb_ctx, context + 1); pm->pm_context[curcpu] = context; #ifdef SMP CPU_SET_ATOMIC(PCPU_GET(cpuid), &pm->pm_active); atomic_store_acq_ptr((uintptr_t *)PCPU_PTR(pmap), (uintptr_t)pm); #else CPU_SET(PCPU_GET(cpuid), &pm->pm_active); PCPU_SET(pmap, pm); #endif stxa(AA_DMMU_TSB, ASI_DMMU, pm->pm_tsb); stxa(AA_IMMU_TSB, ASI_IMMU, pm->pm_tsb); stxa(AA_DMMU_PCXR, ASI_DMMU, (ldxa(AA_DMMU_PCXR, ASI_DMMU) & TLB_CXR_PGSZ_MASK) | context); flush(KERNBASE); critical_exit(); } void pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { } Index: projects/mips64-clang/sys/vm/pmap.h =================================================================== --- projects/mips64-clang/sys/vm/pmap.h (revision 305711) +++ projects/mips64-clang/sys/vm/pmap.h (revision 305712) @@ -1,161 +1,171 @@ /*- * Copyright (c) 1991, 1993 * The Regents of the University of California. All rights reserved. * * This code is derived from software contributed to Berkeley by * The Mach Operating System project at Carnegie-Mellon University. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.h 8.1 (Berkeley) 6/11/93 * * * Copyright (c) 1987, 1990 Carnegie-Mellon University. * All rights reserved. * * Author: Avadis Tevanian, Jr. * * Permission to use, copy, modify and distribute this software and * its documentation is hereby granted, provided that both the copyright * notice and this permission notice appear in all copies of the * software, derivative works or modified versions, and any portions * thereof, and that both notices appear in supporting documentation. * * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. * * Carnegie Mellon requests users of this software to return to * * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU * School of Computer Science * Carnegie Mellon University * Pittsburgh PA 15213-3890 * * any improvements or extensions that they make and grant Carnegie the * rights to redistribute these changes. * * $FreeBSD$ */ /* * Machine address mapping definitions -- machine-independent * section. [For machine-dependent section, see "machine/pmap.h".] */ #ifndef _PMAP_VM_ #define _PMAP_VM_ /* * Each machine dependent implementation is expected to * keep certain statistics. They may do this anyway they * so choose, but are expected to return the statistics * in the following structure. */ struct pmap_statistics { long resident_count; /* # of pages mapped (total) */ long wired_count; /* # of pages wired */ }; typedef struct pmap_statistics *pmap_statistics_t; /* * Each machine-dependent implementation is required to provide: * * vm_memattr_t pmap_page_get_memattr(vm_page_t); * boolean_t pmap_page_is_mapped(vm_page_t); * boolean_t pmap_page_is_write_mapped(vm_page_t); * void pmap_page_set_memattr(vm_page_t, vm_memattr_t); */ #include #ifdef _KERNEL struct thread; /* * Updates to kernel_vm_end are synchronized by the kernel_map's system mutex. */ extern vm_offset_t kernel_vm_end; /* * Flags for pmap_enter(). The bits in the low-order byte are reserved * for the protection code (vm_prot_t) that describes the fault type. */ #define PMAP_ENTER_NOSLEEP 0x0100 #define PMAP_ENTER_WIRED 0x0200 +/* + * Define the maximum number of machine-dependent reference bits that are + * cleared by a call to pmap_ts_referenced(). This limit serves two purposes. + * First, it bounds the cost of reference bit maintenance on widely shared + * pages. Second, it prevents numeric overflow during maintenance of a + * widely shared page's "act_count" field. An overflow could result in the + * premature deactivation of the page. + */ +#define PMAP_TS_REFERENCED_MAX 5 + void pmap_activate(struct thread *td); void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice); void pmap_align_superpage(vm_object_t, vm_ooffset_t, vm_offset_t *, vm_size_t); void pmap_clear_modify(vm_page_t m); void pmap_copy(pmap_t, pmap_t, vm_offset_t, vm_size_t, vm_offset_t); void pmap_copy_page(vm_page_t, vm_page_t); void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize); int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind); void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot); void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot); vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va); vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot); void pmap_growkernel(vm_offset_t); void pmap_init(void); boolean_t pmap_is_modified(vm_page_t m); boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t va); boolean_t pmap_is_referenced(vm_page_t m); vm_offset_t pmap_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int); int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *locked_pa); void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size); boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m); void pmap_page_init(vm_page_t m); int pmap_page_wired_mappings(vm_page_t m); int pmap_pinit(pmap_t); void pmap_pinit0(pmap_t); void pmap_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t); void pmap_qenter(vm_offset_t, vm_page_t *, int); void pmap_qremove(vm_offset_t, int); vm_offset_t pmap_quick_enter_page(vm_page_t); void pmap_quick_remove_page(vm_offset_t); void pmap_release(pmap_t); void pmap_remove(pmap_t, vm_offset_t, vm_offset_t); void pmap_remove_all(vm_page_t m); void pmap_remove_pages(pmap_t); void pmap_remove_write(vm_page_t m); void pmap_sync_icache(pmap_t, vm_offset_t, vm_size_t); boolean_t pmap_ts_referenced(vm_page_t m); void pmap_unwire(pmap_t pmap, vm_offset_t start, vm_offset_t end); void pmap_zero_page(vm_page_t); void pmap_zero_page_area(vm_page_t, int off, int size); #define pmap_resident_count(pm) ((pm)->pm_stats.resident_count) #define pmap_wired_count(pm) ((pm)->pm_stats.wired_count) #endif /* _KERNEL */ #endif /* _PMAP_VM_ */ Index: projects/mips64-clang/tests/sys/kern/Makefile =================================================================== --- projects/mips64-clang/tests/sys/kern/Makefile (revision 305711) +++ projects/mips64-clang/tests/sys/kern/Makefile (revision 305712) @@ -1,39 +1,40 @@ # $FreeBSD$ TESTSRC= ${SRCTOP}/contrib/netbsd-tests/kernel .PATH: ${SRCTOP}/sys/kern TESTSDIR= ${TESTSBASE}/sys/kern ATF_TESTS_C+= kern_copyin ATF_TESTS_C+= kern_descrip_test ATF_TESTS_C+= ptrace_test PLAIN_TESTS_C+= subr_unit_test ATF_TESTS_C+= unix_seqpacket_test ATF_TESTS_C+= unix_passfd_test TEST_METADATA.unix_seqpacket_test+= timeout="15" +ATF_TESTS_C+= waitpid_nohang LIBADD.ptrace_test+= pthread LIBADD.unix_seqpacket_test+= pthread NETBSD_ATF_TESTS_C+= lockf_test NETBSD_ATF_TESTS_C+= mqueue_test CFLAGS.mqueue_test+= -I${SRCTOP}/tests LIBADD.mqueue_test+= rt # subr_unit.c contains functions whose prototypes lie in headers that cannot be # included in userland. But as far as subr_unit_test goes, they're effectively # static. So it's ok to disable -Wmissing-prototypes for this program. CFLAGS.subr_unit.c+= -Wno-missing-prototypes SRCS.subr_unit_test+= subr_unit.c WARNS?= 5 TESTS_SUBDIRS+= acct TESTS_SUBDIRS+= execve TESTS_SUBDIRS+= pipe .include .include Index: projects/mips64-clang/tests/sys/kern/waitpid_nohang.c =================================================================== --- projects/mips64-clang/tests/sys/kern/waitpid_nohang.c (nonexistent) +++ projects/mips64-clang/tests/sys/kern/waitpid_nohang.c (revision 305712) @@ -0,0 +1,70 @@ +/*- + * Copyright (c) 2016 Jilles Tjoelker + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * 1. Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * 2. Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF + * SUCH DAMAGE. + */ + +#include +__FBSDID("$FreeBSD$"); + +#include + +#include +#include +#include + +ATF_TC_WITHOUT_HEAD(waitpid_nohang); +ATF_TC_BODY(waitpid_nohang, tc) +{ + pid_t child, pid; + int status, r; + + child = fork(); + ATF_REQUIRE(child != -1); + if (child == 0) { + sleep(10); + _exit(1); + } + + status = 42; + pid = waitpid(child, &status, WNOHANG); + ATF_REQUIRE(pid == 0); + ATF_CHECK(status == 42); + + r = kill(child, SIGTERM); + ATF_REQUIRE(r == 0); + r = waitid(P_PID, child, NULL, WEXITED | WNOWAIT); + ATF_REQUIRE(r == 0); + + status = -1; + pid = waitpid(child, &status, WNOHANG); + ATF_REQUIRE(pid == child); + ATF_CHECK(WIFSIGNALED(status) && WTERMSIG(status) == SIGTERM); +} + +ATF_TP_ADD_TCS(tp) +{ + + ATF_TP_ADD_TC(tp, waitpid_nohang); + return (atf_no_error()); +} Property changes on: projects/mips64-clang/tests/sys/kern/waitpid_nohang.c ___________________________________________________________________ Added: svn:eol-style ## -0,0 +1 ## +native \ No newline at end of property Added: svn:keywords ## -0,0 +1 ## +FreeBSD=%H \ No newline at end of property Added: svn:mime-type ## -0,0 +1 ## +text/plain \ No newline at end of property Index: projects/mips64-clang =================================================================== --- projects/mips64-clang (revision 305711) +++ projects/mips64-clang (revision 305712) Property changes on: projects/mips64-clang ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head:r305667-305705