Index: head/sys/amd64/include/pcpu.h =================================================================== --- head/sys/amd64/include/pcpu.h (revision 347133) +++ head/sys/amd64/include/pcpu.h (revision 347134) @@ -1,266 +1,270 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) Peter Wemm * 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 _MACHINE_PCPU_H_ #define _MACHINE_PCPU_H_ #ifndef _SYS_CDEFS_H_ #error "sys/cdefs.h is a prerequisite for this file" #endif #define PC_PTI_STACK_SZ 16 struct monitorbuf { int idle_state; /* Used by cpu_idle_mwait. */ - char padding[128 - (1 * sizeof(int))]; + int stop_state; /* Used by cpustop_handler. */ + char padding[128 - (2 * sizeof(int))]; }; _Static_assert(sizeof(struct monitorbuf) == 128, "2x cache line"); /* * The SMP parts are setup in pmap.c and locore.s for the BSP, and * mp_machdep.c sets up the data for the AP's to "see" when they awake. * The reason for doing it via a struct is so that an array of pointers * to each CPU's data can be set up for things like "check curproc on all * other processors" */ #define PCPU_MD_FIELDS \ struct monitorbuf pc_monitorbuf __aligned(128); /* cache line */\ struct pcpu *pc_prvspace; /* Self-reference */ \ struct pmap *pc_curpmap; \ struct amd64tss *pc_tssp; /* TSS segment active on CPU */ \ struct amd64tss *pc_commontssp;/* Common TSS for the CPU */ \ uint64_t pc_kcr3; \ uint64_t pc_ucr3; \ uint64_t pc_saved_ucr3; \ register_t pc_rsp0; \ register_t pc_scratch_rsp; /* User %rsp in syscall */ \ register_t pc_scratch_rax; \ u_int pc_apic_id; \ u_int pc_acpi_id; /* ACPI CPU id */ \ /* Pointer to the CPU %fs descriptor */ \ struct user_segment_descriptor *pc_fs32p; \ /* Pointer to the CPU %gs descriptor */ \ struct user_segment_descriptor *pc_gs32p; \ /* Pointer to the CPU LDT descriptor */ \ struct system_segment_descriptor *pc_ldt; \ /* Pointer to the CPU TSS descriptor */ \ struct system_segment_descriptor *pc_tss; \ uint64_t pc_pm_save_cnt; \ u_int pc_cmci_mask; /* MCx banks for CMCI */ \ uint64_t pc_dbreg[16]; /* ddb debugging regs */ \ uint64_t pc_pti_stack[PC_PTI_STACK_SZ]; \ register_t pc_pti_rsp0; \ int pc_dbreg_cmd; /* ddb debugging reg cmd */ \ u_int pc_vcpu_id; /* Xen vCPU ID */ \ uint32_t pc_pcid_next; \ uint32_t pc_pcid_gen; \ uint32_t pc_smp_tlb_done; /* TLB op acknowledgement */ \ uint32_t pc_ibpb_set; \ char __pad[3288] /* pad to UMA_PCPU_ALLOC_SIZE */ #define PC_DBREG_CMD_NONE 0 #define PC_DBREG_CMD_LOAD 1 #ifdef _KERNEL + +#define MONITOR_STOPSTATE_RUNNING 0 +#define MONITOR_STOPSTATE_STOPPED 1 #if defined(__GNUCLIKE_ASM) && defined(__GNUCLIKE___TYPEOF) /* * Evaluates to the byte offset of the per-cpu variable name. */ #define __pcpu_offset(name) \ __offsetof(struct pcpu, name) /* * Evaluates to the type of the per-cpu variable name. */ #define __pcpu_type(name) \ __typeof(((struct pcpu *)0)->name) /* * Evaluates to the address of the per-cpu variable name. */ #define __PCPU_PTR(name) __extension__ ({ \ __pcpu_type(name) *__p; \ \ __asm __volatile("movq %%gs:%1,%0; addq %2,%0" \ : "=r" (__p) \ : "m" (*(struct pcpu *)(__pcpu_offset(pc_prvspace))), \ "i" (__pcpu_offset(name))); \ \ __p; \ }) /* * Evaluates to the value of the per-cpu variable name. */ #define __PCPU_GET(name) __extension__ ({ \ __pcpu_type(name) __res; \ struct __s { \ u_char __b[MIN(sizeof(__pcpu_type(name)), 8)]; \ } __s; \ \ if (sizeof(__res) == 1 || sizeof(__res) == 2 || \ sizeof(__res) == 4 || sizeof(__res) == 8) { \ __asm __volatile("mov %%gs:%1,%0" \ : "=r" (__s) \ : "m" (*(struct __s *)(__pcpu_offset(name)))); \ *(struct __s *)(void *)&__res = __s; \ } else { \ __res = *__PCPU_PTR(name); \ } \ __res; \ }) /* * Adds the value to the per-cpu counter name. The implementation * must be atomic with respect to interrupts. */ #define __PCPU_ADD(name, val) do { \ __pcpu_type(name) __val; \ struct __s { \ u_char __b[MIN(sizeof(__pcpu_type(name)), 8)]; \ } __s; \ \ __val = (val); \ if (sizeof(__val) == 1 || sizeof(__val) == 2 || \ sizeof(__val) == 4 || sizeof(__val) == 8) { \ __s = *(struct __s *)(void *)&__val; \ __asm __volatile("add %1,%%gs:%0" \ : "=m" (*(struct __s *)(__pcpu_offset(name))) \ : "r" (__s)); \ } else \ *__PCPU_PTR(name) += __val; \ } while (0) /* * Increments the value of the per-cpu counter name. The implementation * must be atomic with respect to interrupts. */ #define __PCPU_INC(name) do { \ CTASSERT(sizeof(__pcpu_type(name)) == 1 || \ sizeof(__pcpu_type(name)) == 2 || \ sizeof(__pcpu_type(name)) == 4 || \ sizeof(__pcpu_type(name)) == 8); \ if (sizeof(__pcpu_type(name)) == 1) { \ __asm __volatile("incb %%gs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } else if (sizeof(__pcpu_type(name)) == 2) { \ __asm __volatile("incw %%gs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } else if (sizeof(__pcpu_type(name)) == 4) { \ __asm __volatile("incl %%gs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } else if (sizeof(__pcpu_type(name)) == 8) { \ __asm __volatile("incq %%gs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } \ } while (0) /* * Sets the value of the per-cpu variable name to value val. */ #define __PCPU_SET(name, val) { \ __pcpu_type(name) __val; \ struct __s { \ u_char __b[MIN(sizeof(__pcpu_type(name)), 8)]; \ } __s; \ \ __val = (val); \ if (sizeof(__val) == 1 || sizeof(__val) == 2 || \ sizeof(__val) == 4 || sizeof(__val) == 8) { \ __s = *(struct __s *)(void *)&__val; \ __asm __volatile("mov %1,%%gs:%0" \ : "=m" (*(struct __s *)(__pcpu_offset(name))) \ : "r" (__s)); \ } else { \ *__PCPU_PTR(name) = __val; \ } \ } #define get_pcpu() __extension__ ({ \ struct pcpu *__pc; \ \ __asm __volatile("movq %%gs:%1,%0" \ : "=r" (__pc) \ : "m" (*(struct pcpu *)(__pcpu_offset(pc_prvspace)))); \ __pc; \ }) #define PCPU_GET(member) __PCPU_GET(pc_ ## member) #define PCPU_ADD(member, val) __PCPU_ADD(pc_ ## member, val) #define PCPU_INC(member) __PCPU_INC(pc_ ## member) #define PCPU_PTR(member) __PCPU_PTR(pc_ ## member) #define PCPU_SET(member, val) __PCPU_SET(pc_ ## member, val) #define OFFSETOF_CURTHREAD 0 #ifdef __clang__ #pragma clang diagnostic push #pragma clang diagnostic ignored "-Wnull-dereference" #endif static __inline __pure2 struct thread * __curthread(void) { struct thread *td; __asm("movq %%gs:%P1,%0" : "=r" (td) : "n" (OFFSETOF_CURTHREAD)); return (td); } #ifdef __clang__ #pragma clang diagnostic pop #endif #define curthread (__curthread()) #define OFFSETOF_CURPCB 32 static __inline __pure2 struct pcb * __curpcb(void) { struct pcb *pcb; __asm("movq %%gs:%P1,%0" : "=r" (pcb) : "n" (OFFSETOF_CURPCB)); return (pcb); } #define curpcb (__curpcb()) #define IS_BSP() (PCPU_GET(cpuid) == 0) #else /* !__GNUCLIKE_ASM || !__GNUCLIKE___TYPEOF */ #error "this file needs to be ported to your compiler" #endif /* __GNUCLIKE_ASM && __GNUCLIKE___TYPEOF */ #endif /* _KERNEL */ #endif /* !_MACHINE_PCPU_H_ */ Index: head/sys/i386/include/pcpu.h =================================================================== --- head/sys/i386/include/pcpu.h (revision 347133) +++ head/sys/i386/include/pcpu.h (revision 347134) @@ -1,262 +1,266 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) Peter Wemm * 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 _MACHINE_PCPU_H_ #define _MACHINE_PCPU_H_ #ifndef _SYS_CDEFS_H_ #error "sys/cdefs.h is a prerequisite for this file" #endif #include #include #include #include struct monitorbuf { int idle_state; /* Used by cpu_idle_mwait. */ - char padding[128 - (1 * sizeof(int))]; + int stop_state; /* Used by cpustop_handler. */ + char padding[128 - (2 * sizeof(int))]; }; _Static_assert(sizeof(struct monitorbuf) == 128, "2x cache line"); /* * The SMP parts are setup in pmap.c and machdep.c for the BSP, and * pmap.c and mp_machdep.c sets up the data for the AP's to "see" when * they awake. The reason for doing it via a struct is so that an * array of pointers to each CPU's data can be set up for things like * "check curproc on all other processors" */ #define PCPU_MD_FIELDS \ struct monitorbuf pc_monitorbuf __aligned(128); /* cache line */\ struct pcpu *pc_prvspace; /* Self-reference */ \ struct pmap *pc_curpmap; \ struct segment_descriptor pc_common_tssd; \ struct segment_descriptor *pc_tss_gdt; \ struct segment_descriptor *pc_fsgs_gdt; \ struct i386tss *pc_common_tssp; \ u_int pc_kesp0; \ u_int pc_trampstk; \ int pc_currentldt; \ u_int pc_acpi_id; /* ACPI CPU id */ \ u_int pc_apic_id; \ int pc_private_tss; /* Flag indicating private tss*/\ u_int pc_cmci_mask; /* MCx banks for CMCI */ \ u_int pc_vcpu_id; /* Xen vCPU ID */ \ struct mtx pc_cmap_lock; \ void *pc_cmap_pte1; \ void *pc_cmap_pte2; \ caddr_t pc_cmap_addr1; \ caddr_t pc_cmap_addr2; \ vm_offset_t pc_qmap_addr; /* KVA for temporary mappings */\ vm_offset_t pc_copyout_maddr; \ vm_offset_t pc_copyout_saddr; \ struct mtx pc_copyout_mlock; \ struct sx pc_copyout_slock; \ char *pc_copyout_buf; \ vm_offset_t pc_pmap_eh_va; \ caddr_t pc_pmap_eh_ptep; \ uint32_t pc_smp_tlb_done; /* TLB op acknowledgement */ \ uint32_t pc_ibpb_set; \ char __pad[3610] #ifdef _KERNEL + +#define MONITOR_STOPSTATE_RUNNING 0 +#define MONITOR_STOPSTATE_STOPPED 1 #if defined(__GNUCLIKE_ASM) && defined(__GNUCLIKE___TYPEOF) /* * Evaluates to the byte offset of the per-cpu variable name. */ #define __pcpu_offset(name) \ __offsetof(struct pcpu, name) /* * Evaluates to the type of the per-cpu variable name. */ #define __pcpu_type(name) \ __typeof(((struct pcpu *)0)->name) /* * Evaluates to the address of the per-cpu variable name. */ #define __PCPU_PTR(name) __extension__ ({ \ __pcpu_type(name) *__p; \ \ __asm __volatile("movl %%fs:%1,%0; addl %2,%0" \ : "=r" (__p) \ : "m" (*(struct pcpu *)(__pcpu_offset(pc_prvspace))), \ "i" (__pcpu_offset(name))); \ \ __p; \ }) /* * Evaluates to the value of the per-cpu variable name. */ #define __PCPU_GET(name) __extension__ ({ \ __pcpu_type(name) __res; \ struct __s { \ u_char __b[MIN(sizeof(__res), 4)]; \ } __s; \ \ if (sizeof(__res) == 1 || sizeof(__res) == 2 || \ sizeof(__res) == 4) { \ __asm __volatile("mov %%fs:%1,%0" \ : "=r" (__s) \ : "m" (*(struct __s *)(__pcpu_offset(name)))); \ *(struct __s *)(void *)&__res = __s; \ } else { \ __res = *__PCPU_PTR(name); \ } \ __res; \ }) /* * Adds a value of the per-cpu counter name. The implementation * must be atomic with respect to interrupts. */ #define __PCPU_ADD(name, val) do { \ __pcpu_type(name) __val; \ struct __s { \ u_char __b[MIN(sizeof(__val), 4)]; \ } __s; \ \ __val = (val); \ if (sizeof(__val) == 1 || sizeof(__val) == 2 || \ sizeof(__val) == 4) { \ __s = *(struct __s *)(void *)&__val; \ __asm __volatile("add %1,%%fs:%0" \ : "=m" (*(struct __s *)(__pcpu_offset(name))) \ : "r" (__s)); \ } else \ *__PCPU_PTR(name) += __val; \ } while (0) /* * Increments the value of the per-cpu counter name. The implementation * must be atomic with respect to interrupts. */ #define __PCPU_INC(name) do { \ CTASSERT(sizeof(__pcpu_type(name)) == 1 || \ sizeof(__pcpu_type(name)) == 2 || \ sizeof(__pcpu_type(name)) == 4); \ if (sizeof(__pcpu_type(name)) == 1) { \ __asm __volatile("incb %%fs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } else if (sizeof(__pcpu_type(name)) == 2) { \ __asm __volatile("incw %%fs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } else if (sizeof(__pcpu_type(name)) == 4) { \ __asm __volatile("incl %%fs:%0" \ : "=m" (*(__pcpu_type(name) *)(__pcpu_offset(name)))\ : "m" (*(__pcpu_type(name) *)(__pcpu_offset(name))));\ } \ } while (0) /* * Sets the value of the per-cpu variable name to value val. */ #define __PCPU_SET(name, val) do { \ __pcpu_type(name) __val; \ struct __s { \ u_char __b[MIN(sizeof(__val), 4)]; \ } __s; \ \ __val = (val); \ if (sizeof(__val) == 1 || sizeof(__val) == 2 || \ sizeof(__val) == 4) { \ __s = *(struct __s *)(void *)&__val; \ __asm __volatile("mov %1,%%fs:%0" \ : "=m" (*(struct __s *)(__pcpu_offset(name))) \ : "r" (__s)); \ } else { \ *__PCPU_PTR(name) = __val; \ } \ } while (0) #define get_pcpu() __extension__ ({ \ struct pcpu *__pc; \ \ __asm __volatile("movl %%fs:%1,%0" \ : "=r" (__pc) \ : "m" (*(struct pcpu *)(__pcpu_offset(pc_prvspace)))); \ __pc; \ }) #define PCPU_GET(member) __PCPU_GET(pc_ ## member) #define PCPU_ADD(member, val) __PCPU_ADD(pc_ ## member, val) #define PCPU_INC(member) __PCPU_INC(pc_ ## member) #define PCPU_PTR(member) __PCPU_PTR(pc_ ## member) #define PCPU_SET(member, val) __PCPU_SET(pc_ ## member, val) #define OFFSETOF_CURTHREAD 0 #ifdef __clang__ #pragma clang diagnostic push #pragma clang diagnostic ignored "-Wnull-dereference" #endif static __inline __pure2 struct thread * __curthread(void) { struct thread *td; __asm("movl %%fs:%1,%0" : "=r" (td) : "m" (*(char *)OFFSETOF_CURTHREAD)); return (td); } #ifdef __clang__ #pragma clang diagnostic pop #endif #define curthread (__curthread()) #define OFFSETOF_CURPCB 16 static __inline __pure2 struct pcb * __curpcb(void) { struct pcb *pcb; __asm("movl %%fs:%1,%0" : "=r" (pcb) : "m" (*(char *)OFFSETOF_CURPCB)); return (pcb); } #define curpcb (__curpcb()) #define IS_BSP() (PCPU_GET(cpuid) == 0) #else /* defined(__GNUCLIKE_ASM) && defined(__GNUCLIKE___TYPEOF) */ #error "this file needs to be ported to your compiler" #endif /* __GNUCLIKE_ASM etc. */ #endif /* _KERNEL */ #endif /* !_MACHINE_PCPU_H_ */ Index: head/sys/kern/subr_smp.c =================================================================== --- head/sys/kern/subr_smp.c (revision 347133) +++ head/sys/kern/subr_smp.c (revision 347134) @@ -1,1173 +1,1199 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2001, John Baldwin . * * 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. */ /* * This module holds the global variables and machine independent functions * used for the kernel SMP support. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "opt_sched.h" #ifdef SMP MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data"); volatile cpuset_t stopped_cpus; volatile cpuset_t started_cpus; volatile cpuset_t suspended_cpus; cpuset_t hlt_cpus_mask; cpuset_t logical_cpus_mask; void (*cpustop_restartfunc)(void); #endif static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS); /* This is used in modules that need to work in both SMP and UP. */ cpuset_t all_cpus; int mp_ncpus; /* export this for libkvm consumers. */ int mp_maxcpus = MAXCPU; volatile int smp_started; u_int mp_maxid; static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD|CTLFLAG_CAPRD, NULL, "Kernel SMP"); SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0, "Max CPU ID."); SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus, 0, "Max number of CPUs that the system was compiled for."); SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD|CTLTYPE_INT|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_smp_active, "I", "Indicates system is running in SMP mode"); int smp_disabled = 0; /* has smp been disabled? */ SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD, &smp_disabled, 0, "SMP has been disabled from the loader"); int smp_cpus = 1; /* how many cpu's running */ SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0, "Number of CPUs online"); int smp_threads_per_core = 1; /* how many SMT threads are running per core */ SYSCTL_INT(_kern_smp, OID_AUTO, threads_per_core, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_threads_per_core, 0, "Number of SMT threads online per core"); int mp_ncores = -1; /* how many physical cores running */ SYSCTL_INT(_kern_smp, OID_AUTO, cores, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_ncores, 0, "Number of CPUs online"); int smp_topology = 0; /* Which topology we're using. */ SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0, "Topology override setting; 0 is default provided by hardware."); #ifdef SMP /* Enable forwarding of a signal to a process running on a different CPU */ static int forward_signal_enabled = 1; SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW, &forward_signal_enabled, 0, "Forwarding of a signal to a process on a different CPU"); /* Variables needed for SMP rendezvous. */ static volatile int smp_rv_ncpus; static void (*volatile smp_rv_setup_func)(void *arg); static void (*volatile smp_rv_action_func)(void *arg); static void (*volatile smp_rv_teardown_func)(void *arg); static void *volatile smp_rv_func_arg; static volatile int smp_rv_waiters[4]; /* * Shared mutex to restrict busywaits between smp_rendezvous() and * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these * functions trigger at once and cause multiple CPUs to busywait with * interrupts disabled. */ struct mtx smp_ipi_mtx; /* * Let the MD SMP code initialize mp_maxid very early if it can. */ static void mp_setmaxid(void *dummy) { cpu_mp_setmaxid(); KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__)); KASSERT(mp_ncpus > 1 || mp_maxid == 0, ("%s: one CPU but mp_maxid is not zero", __func__)); KASSERT(mp_maxid >= mp_ncpus - 1, ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid, mp_ncpus)); } SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL); /* * Call the MD SMP initialization code. */ static void mp_start(void *dummy) { mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN); /* Probe for MP hardware. */ if (smp_disabled != 0 || cpu_mp_probe() == 0) { mp_ncores = 1; mp_ncpus = 1; CPU_SETOF(PCPU_GET(cpuid), &all_cpus); return; } cpu_mp_start(); printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n", mp_ncpus); /* Provide a default for most architectures that don't have SMT/HTT. */ if (mp_ncores < 0) mp_ncores = mp_ncpus; cpu_mp_announce(); } SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL); void forward_signal(struct thread *td) { int id; /* * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on * this thread, so all we need to do is poke it if it is currently * executing so that it executes ast(). */ THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT(TD_IS_RUNNING(td), ("forward_signal: thread is not TDS_RUNNING")); CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc); if (!smp_started || cold || panicstr) return; if (!forward_signal_enabled) return; /* No need to IPI ourself. */ if (td == curthread) return; id = td->td_oncpu; if (id == NOCPU) return; ipi_cpu(id, IPI_AST); } /* * When called the executing CPU will send an IPI to all other CPUs * requesting that they halt execution. * * Usually (but not necessarily) called with 'other_cpus' as its arg. * * - Signals all CPUs in map to stop. * - Waits for each to stop. * * Returns: * -1: error * 0: NA * 1: ok * */ #if defined(__amd64__) || defined(__i386__) #define X86 1 #else #define X86 0 #endif static int generic_stop_cpus(cpuset_t map, u_int type) { #ifdef KTR char cpusetbuf[CPUSETBUFSIZ]; #endif static volatile u_int stopping_cpu = NOCPU; int i; volatile cpuset_t *cpus; KASSERT( type == IPI_STOP || type == IPI_STOP_HARD #if X86 || type == IPI_SUSPEND #endif , ("%s: invalid stop type", __func__)); if (!smp_started) return (0); CTR2(KTR_SMP, "stop_cpus(%s) with %u type", cpusetobj_strprint(cpusetbuf, &map), type); #if X86 /* * When suspending, ensure there are are no IPIs in progress. * IPIs that have been issued, but not yet delivered (e.g. * not pending on a vCPU when running under virtualization) * will be lost, violating FreeBSD's assumption of reliable * IPI delivery. */ if (type == IPI_SUSPEND) mtx_lock_spin(&smp_ipi_mtx); #endif #if X86 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { #endif if (stopping_cpu != PCPU_GET(cpuid)) while (atomic_cmpset_int(&stopping_cpu, NOCPU, PCPU_GET(cpuid)) == 0) while (stopping_cpu != NOCPU) cpu_spinwait(); /* spin */ /* send the stop IPI to all CPUs in map */ ipi_selected(map, type); #if X86 } #endif #if X86 if (type == IPI_SUSPEND) cpus = &suspended_cpus; else #endif cpus = &stopped_cpus; i = 0; while (!CPU_SUBSET(cpus, &map)) { /* spin */ cpu_spinwait(); i++; if (i == 100000000) { printf("timeout stopping cpus\n"); break; } } #if X86 if (type == IPI_SUSPEND) mtx_unlock_spin(&smp_ipi_mtx); #endif stopping_cpu = NOCPU; return (1); } int stop_cpus(cpuset_t map) { return (generic_stop_cpus(map, IPI_STOP)); } int stop_cpus_hard(cpuset_t map) { return (generic_stop_cpus(map, IPI_STOP_HARD)); } #if X86 int suspend_cpus(cpuset_t map) { return (generic_stop_cpus(map, IPI_SUSPEND)); } #endif /* * Called by a CPU to restart stopped CPUs. * * Usually (but not necessarily) called with 'stopped_cpus' as its arg. * * - Signals all CPUs in map to restart. * - Waits for each to restart. * * Returns: * -1: error * 0: NA * 1: ok */ static int generic_restart_cpus(cpuset_t map, u_int type) { #ifdef KTR char cpusetbuf[CPUSETBUFSIZ]; #endif volatile cpuset_t *cpus; - KASSERT(type == IPI_STOP || type == IPI_STOP_HARD #if X86 - || type == IPI_SUSPEND -#endif - , ("%s: invalid stop type", __func__)); + KASSERT(type == IPI_STOP || type == IPI_STOP_HARD + || type == IPI_SUSPEND, ("%s: invalid stop type", __func__)); if (!smp_started) return (0); CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); -#if X86 if (type == IPI_SUSPEND) cpus = &resuming_cpus; else -#endif cpus = &stopped_cpus; /* signal other cpus to restart */ -#if X86 if (type == IPI_SUSPEND) CPU_COPY_STORE_REL(&map, &toresume_cpus); else -#endif CPU_COPY_STORE_REL(&map, &started_cpus); -#if X86 + /* + * Wake up any CPUs stopped with MWAIT. From MI code we can't tell if + * MONITOR/MWAIT is enabled, but the potentially redundant writes are + * relatively inexpensive. + */ + if (type == IPI_STOP) { + struct monitorbuf *mb; + u_int id; + + CPU_FOREACH(id) { + if (!CPU_ISSET(id, &map)) + continue; + + mb = &pcpu_find(id)->pc_monitorbuf; + atomic_store_int(&mb->stop_state, + MONITOR_STOPSTATE_RUNNING); + } + } + if (!nmi_is_broadcast || nmi_kdb_lock == 0) { -#endif + /* wait for each to clear its bit */ + while (CPU_OVERLAP(cpus, &map)) + cpu_spinwait(); + } +#else /* !X86 */ + KASSERT(type == IPI_STOP || type == IPI_STOP_HARD, + ("%s: invalid stop type", __func__)); + + if (!smp_started) + return (0); + + CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); + + cpus = &stopped_cpus; + + /* signal other cpus to restart */ + CPU_COPY_STORE_REL(&map, &started_cpus); + /* wait for each to clear its bit */ while (CPU_OVERLAP(cpus, &map)) cpu_spinwait(); -#if X86 - } #endif - return (1); } int restart_cpus(cpuset_t map) { return (generic_restart_cpus(map, IPI_STOP)); } #if X86 int resume_cpus(cpuset_t map) { return (generic_restart_cpus(map, IPI_SUSPEND)); } #endif #undef X86 /* * All-CPU rendezvous. CPUs are signalled, all execute the setup function * (if specified), rendezvous, execute the action function (if specified), * rendezvous again, execute the teardown function (if specified), and then * resume. * * Note that the supplied external functions _must_ be reentrant and aware * that they are running in parallel and in an unknown lock context. */ void smp_rendezvous_action(void) { struct thread *td; void *local_func_arg; void (*local_setup_func)(void*); void (*local_action_func)(void*); void (*local_teardown_func)(void*); #ifdef INVARIANTS int owepreempt; #endif /* Ensure we have up-to-date values. */ atomic_add_acq_int(&smp_rv_waiters[0], 1); while (smp_rv_waiters[0] < smp_rv_ncpus) cpu_spinwait(); /* Fetch rendezvous parameters after acquire barrier. */ local_func_arg = smp_rv_func_arg; local_setup_func = smp_rv_setup_func; local_action_func = smp_rv_action_func; local_teardown_func = smp_rv_teardown_func; /* * Use a nested critical section to prevent any preemptions * from occurring during a rendezvous action routine. * Specifically, if a rendezvous handler is invoked via an IPI * and the interrupted thread was in the critical_exit() * function after setting td_critnest to 0 but before * performing a deferred preemption, this routine can be * invoked with td_critnest set to 0 and td_owepreempt true. * In that case, a critical_exit() during the rendezvous * action would trigger a preemption which is not permitted in * a rendezvous action. To fix this, wrap all of the * rendezvous action handlers in a critical section. We * cannot use a regular critical section however as having * critical_exit() preempt from this routine would also be * problematic (the preemption must not occur before the IPI * has been acknowledged via an EOI). Instead, we * intentionally ignore td_owepreempt when leaving the * critical section. This should be harmless because we do * not permit rendezvous action routines to schedule threads, * and thus td_owepreempt should never transition from 0 to 1 * during this routine. */ td = curthread; td->td_critnest++; #ifdef INVARIANTS owepreempt = td->td_owepreempt; #endif /* * If requested, run a setup function before the main action * function. Ensure all CPUs have completed the setup * function before moving on to the action function. */ if (local_setup_func != smp_no_rendezvous_barrier) { if (smp_rv_setup_func != NULL) smp_rv_setup_func(smp_rv_func_arg); atomic_add_int(&smp_rv_waiters[1], 1); while (smp_rv_waiters[1] < smp_rv_ncpus) cpu_spinwait(); } if (local_action_func != NULL) local_action_func(local_func_arg); if (local_teardown_func != smp_no_rendezvous_barrier) { /* * Signal that the main action has been completed. If a * full exit rendezvous is requested, then all CPUs will * wait here until all CPUs have finished the main action. */ atomic_add_int(&smp_rv_waiters[2], 1); while (smp_rv_waiters[2] < smp_rv_ncpus) cpu_spinwait(); if (local_teardown_func != NULL) local_teardown_func(local_func_arg); } /* * Signal that the rendezvous is fully completed by this CPU. * This means that no member of smp_rv_* pseudo-structure will be * accessed by this target CPU after this point; in particular, * memory pointed by smp_rv_func_arg. * * The release semantic ensures that all accesses performed by * the current CPU are visible when smp_rendezvous_cpus() * returns, by synchronizing with the * atomic_load_acq_int(&smp_rv_waiters[3]). */ atomic_add_rel_int(&smp_rv_waiters[3], 1); td->td_critnest--; KASSERT(owepreempt == td->td_owepreempt, ("rendezvous action changed td_owepreempt")); } void smp_rendezvous_cpus(cpuset_t map, void (* setup_func)(void *), void (* action_func)(void *), void (* teardown_func)(void *), void *arg) { int curcpumap, i, ncpus = 0; /* Look comments in the !SMP case. */ if (!smp_started) { spinlock_enter(); if (setup_func != NULL) setup_func(arg); if (action_func != NULL) action_func(arg); if (teardown_func != NULL) teardown_func(arg); spinlock_exit(); return; } CPU_FOREACH(i) { if (CPU_ISSET(i, &map)) ncpus++; } if (ncpus == 0) panic("ncpus is 0 with non-zero map"); mtx_lock_spin(&smp_ipi_mtx); /* Pass rendezvous parameters via global variables. */ smp_rv_ncpus = ncpus; smp_rv_setup_func = setup_func; smp_rv_action_func = action_func; smp_rv_teardown_func = teardown_func; smp_rv_func_arg = arg; smp_rv_waiters[1] = 0; smp_rv_waiters[2] = 0; smp_rv_waiters[3] = 0; atomic_store_rel_int(&smp_rv_waiters[0], 0); /* * Signal other processors, which will enter the IPI with * interrupts off. */ curcpumap = CPU_ISSET(curcpu, &map); CPU_CLR(curcpu, &map); ipi_selected(map, IPI_RENDEZVOUS); /* Check if the current CPU is in the map */ if (curcpumap != 0) smp_rendezvous_action(); /* * Ensure that the master CPU waits for all the other * CPUs to finish the rendezvous, so that smp_rv_* * pseudo-structure and the arg are guaranteed to not * be in use. * * Load acquire synchronizes with the release add in * smp_rendezvous_action(), which ensures that our caller sees * all memory actions done by the called functions on other * CPUs. */ while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus) cpu_spinwait(); mtx_unlock_spin(&smp_ipi_mtx); } void smp_rendezvous(void (* setup_func)(void *), void (* action_func)(void *), void (* teardown_func)(void *), void *arg) { smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); } static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1]; struct cpu_group * smp_topo(void) { char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; struct cpu_group *top; /* * Check for a fake topology request for debugging purposes. */ switch (smp_topology) { case 1: /* Dual core with no sharing. */ top = smp_topo_1level(CG_SHARE_NONE, 2, 0); break; case 2: /* No topology, all cpus are equal. */ top = smp_topo_none(); break; case 3: /* Dual core with shared L2. */ top = smp_topo_1level(CG_SHARE_L2, 2, 0); break; case 4: /* quad core, shared l3 among each package, private l2. */ top = smp_topo_1level(CG_SHARE_L3, 4, 0); break; case 5: /* quad core, 2 dualcore parts on each package share l2. */ top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0); break; case 6: /* Single-core 2xHTT */ top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT); break; case 7: /* quad core with a shared l3, 8 threads sharing L2. */ top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8, CG_FLAG_SMT); break; default: /* Default, ask the system what it wants. */ top = cpu_topo(); break; } /* * Verify the returned topology. */ if (top->cg_count != mp_ncpus) panic("Built bad topology at %p. CPU count %d != %d", top, top->cg_count, mp_ncpus); if (CPU_CMP(&top->cg_mask, &all_cpus)) panic("Built bad topology at %p. CPU mask (%s) != (%s)", top, cpusetobj_strprint(cpusetbuf, &top->cg_mask), cpusetobj_strprint(cpusetbuf2, &all_cpus)); /* * Collapse nonsense levels that may be created out of convenience by * the MD layers. They cause extra work in the search functions. */ while (top->cg_children == 1) { top = &top->cg_child[0]; top->cg_parent = NULL; } return (top); } struct cpu_group * smp_topo_alloc(u_int count) { static u_int index; u_int curr; curr = index; index += count; return (&group[curr]); } struct cpu_group * smp_topo_none(void) { struct cpu_group *top; top = &group[0]; top->cg_parent = NULL; top->cg_child = NULL; top->cg_mask = all_cpus; top->cg_count = mp_ncpus; top->cg_children = 0; top->cg_level = CG_SHARE_NONE; top->cg_flags = 0; return (top); } static int smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share, int count, int flags, int start) { char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; cpuset_t mask; int i; CPU_ZERO(&mask); for (i = 0; i < count; i++, start++) CPU_SET(start, &mask); child->cg_parent = parent; child->cg_child = NULL; child->cg_children = 0; child->cg_level = share; child->cg_count = count; child->cg_flags = flags; child->cg_mask = mask; parent->cg_children++; for (; parent != NULL; parent = parent->cg_parent) { if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask)) panic("Duplicate children in %p. mask (%s) child (%s)", parent, cpusetobj_strprint(cpusetbuf, &parent->cg_mask), cpusetobj_strprint(cpusetbuf2, &child->cg_mask)); CPU_OR(&parent->cg_mask, &child->cg_mask); parent->cg_count += child->cg_count; } return (start); } struct cpu_group * smp_topo_1level(int share, int count, int flags) { struct cpu_group *child; struct cpu_group *top; int packages; int cpu; int i; cpu = 0; top = &group[0]; packages = mp_ncpus / count; top->cg_child = child = &group[1]; top->cg_level = CG_SHARE_NONE; for (i = 0; i < packages; i++, child++) cpu = smp_topo_addleaf(top, child, share, count, flags, cpu); return (top); } struct cpu_group * smp_topo_2level(int l2share, int l2count, int l1share, int l1count, int l1flags) { struct cpu_group *top; struct cpu_group *l1g; struct cpu_group *l2g; int cpu; int i; int j; cpu = 0; top = &group[0]; l2g = &group[1]; top->cg_child = l2g; top->cg_level = CG_SHARE_NONE; top->cg_children = mp_ncpus / (l2count * l1count); l1g = l2g + top->cg_children; for (i = 0; i < top->cg_children; i++, l2g++) { l2g->cg_parent = top; l2g->cg_child = l1g; l2g->cg_level = l2share; for (j = 0; j < l2count; j++, l1g++) cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count, l1flags, cpu); } return (top); } struct cpu_group * smp_topo_find(struct cpu_group *top, int cpu) { struct cpu_group *cg; cpuset_t mask; int children; int i; CPU_SETOF(cpu, &mask); cg = top; for (;;) { if (!CPU_OVERLAP(&cg->cg_mask, &mask)) return (NULL); if (cg->cg_children == 0) return (cg); children = cg->cg_children; for (i = 0, cg = cg->cg_child; i < children; cg++, i++) if (CPU_OVERLAP(&cg->cg_mask, &mask)) break; } return (NULL); } #else /* !SMP */ void smp_rendezvous_cpus(cpuset_t map, void (*setup_func)(void *), void (*action_func)(void *), void (*teardown_func)(void *), void *arg) { /* * In the !SMP case we just need to ensure the same initial conditions * as the SMP case. */ spinlock_enter(); if (setup_func != NULL) setup_func(arg); if (action_func != NULL) action_func(arg); if (teardown_func != NULL) teardown_func(arg); spinlock_exit(); } void smp_rendezvous(void (*setup_func)(void *), void (*action_func)(void *), void (*teardown_func)(void *), void *arg) { smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); } /* * Provide dummy SMP support for UP kernels. Modules that need to use SMP * APIs will still work using this dummy support. */ static void mp_setvariables_for_up(void *dummy) { mp_ncpus = 1; mp_ncores = 1; mp_maxid = PCPU_GET(cpuid); CPU_SETOF(mp_maxid, &all_cpus); KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero")); } SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setvariables_for_up, NULL); #endif /* SMP */ void smp_no_rendezvous_barrier(void *dummy) { #ifdef SMP KASSERT((!smp_started),("smp_no_rendezvous called and smp is started")); #endif } /* * Wait for specified idle threads to switch once. This ensures that even * preempted threads have cycled through the switch function once, * exiting their codepaths. This allows us to change global pointers * with no other synchronization. */ int quiesce_cpus(cpuset_t map, const char *wmesg, int prio) { struct pcpu *pcpu; u_int gen[MAXCPU]; int error; int cpu; error = 0; for (cpu = 0; cpu <= mp_maxid; cpu++) { if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) continue; pcpu = pcpu_find(cpu); gen[cpu] = pcpu->pc_idlethread->td_generation; } for (cpu = 0; cpu <= mp_maxid; cpu++) { if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) continue; pcpu = pcpu_find(cpu); thread_lock(curthread); sched_bind(curthread, cpu); thread_unlock(curthread); while (gen[cpu] == pcpu->pc_idlethread->td_generation) { error = tsleep(quiesce_cpus, prio, wmesg, 1); if (error != EWOULDBLOCK) goto out; error = 0; } } out: thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); return (error); } int quiesce_all_cpus(const char *wmesg, int prio) { return quiesce_cpus(all_cpus, wmesg, prio); } /* Extra care is taken with this sysctl because the data type is volatile */ static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS) { int error, active; active = smp_started; error = SYSCTL_OUT(req, &active, sizeof(active)); return (error); } #ifdef SMP void topo_init_node(struct topo_node *node) { bzero(node, sizeof(*node)); TAILQ_INIT(&node->children); } void topo_init_root(struct topo_node *root) { topo_init_node(root); root->type = TOPO_TYPE_SYSTEM; } /* * Add a child node with the given ID under the given parent. * Do nothing if there is already a child with that ID. */ struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid, topo_node_type type, uintptr_t subtype) { struct topo_node *node; TAILQ_FOREACH_REVERSE(node, &parent->children, topo_children, siblings) { if (node->hwid == hwid && node->type == type && node->subtype == subtype) { return (node); } } node = malloc(sizeof(*node), M_TOPO, M_WAITOK); topo_init_node(node); node->parent = parent; node->hwid = hwid; node->type = type; node->subtype = subtype; TAILQ_INSERT_TAIL(&parent->children, node, siblings); parent->nchildren++; return (node); } /* * Find a child node with the given ID under the given parent. */ struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid, topo_node_type type, uintptr_t subtype) { struct topo_node *node; TAILQ_FOREACH(node, &parent->children, siblings) { if (node->hwid == hwid && node->type == type && node->subtype == subtype) { return (node); } } return (NULL); } /* * Given a node change the order of its parent's child nodes such * that the node becomes the firt child while preserving the cyclic * order of the children. In other words, the given node is promoted * by rotation. */ void topo_promote_child(struct topo_node *child) { struct topo_node *next; struct topo_node *node; struct topo_node *parent; parent = child->parent; next = TAILQ_NEXT(child, siblings); TAILQ_REMOVE(&parent->children, child, siblings); TAILQ_INSERT_HEAD(&parent->children, child, siblings); while (next != NULL) { node = next; next = TAILQ_NEXT(node, siblings); TAILQ_REMOVE(&parent->children, node, siblings); TAILQ_INSERT_AFTER(&parent->children, child, node, siblings); child = node; } } /* * Iterate to the next node in the depth-first search (traversal) of * the topology tree. */ struct topo_node * topo_next_node(struct topo_node *top, struct topo_node *node) { struct topo_node *next; if ((next = TAILQ_FIRST(&node->children)) != NULL) return (next); if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); while (node != top && (node = node->parent) != top) if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); return (NULL); } /* * Iterate to the next node in the depth-first search of the topology tree, * but without descending below the current node. */ struct topo_node * topo_next_nonchild_node(struct topo_node *top, struct topo_node *node) { struct topo_node *next; if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); while (node != top && (node = node->parent) != top) if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); return (NULL); } /* * Assign the given ID to the given topology node that represents a logical * processor. */ void topo_set_pu_id(struct topo_node *node, cpuid_t id) { KASSERT(node->type == TOPO_TYPE_PU, ("topo_set_pu_id: wrong node type: %u", node->type)); KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0, ("topo_set_pu_id: cpuset already not empty")); node->id = id; CPU_SET(id, &node->cpuset); node->cpu_count = 1; node->subtype = 1; while ((node = node->parent) != NULL) { KASSERT(!CPU_ISSET(id, &node->cpuset), ("logical ID %u is already set in node %p", id, node)); CPU_SET(id, &node->cpuset); node->cpu_count++; } } static struct topology_spec { topo_node_type type; bool match_subtype; uintptr_t subtype; } topology_level_table[TOPO_LEVEL_COUNT] = { [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, }, [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, }, [TOPO_LEVEL_CACHEGROUP] = { .type = TOPO_TYPE_CACHE, .match_subtype = true, .subtype = CG_SHARE_L3, }, [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, }, [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, }, }; static bool topo_analyze_table(struct topo_node *root, int all, enum topo_level level, struct topo_analysis *results) { struct topology_spec *spec; struct topo_node *node; int count; if (level >= TOPO_LEVEL_COUNT) return (true); spec = &topology_level_table[level]; count = 0; node = topo_next_node(root, root); while (node != NULL) { if (node->type != spec->type || (spec->match_subtype && node->subtype != spec->subtype)) { node = topo_next_node(root, node); continue; } if (!all && CPU_EMPTY(&node->cpuset)) { node = topo_next_nonchild_node(root, node); continue; } count++; if (!topo_analyze_table(node, all, level + 1, results)) return (false); node = topo_next_nonchild_node(root, node); } /* No explicit subgroups is essentially one subgroup. */ if (count == 0) { count = 1; if (!topo_analyze_table(root, all, level + 1, results)) return (false); } if (results->entities[level] == -1) results->entities[level] = count; else if (results->entities[level] != count) return (false); return (true); } /* * Check if the topology is uniform, that is, each package has the same number * of cores in it and each core has the same number of threads (logical * processors) in it. If so, calculate the number of packages, the number of * groups per package, the number of cachegroups per group, and the number of * logical processors per cachegroup. 'all' parameter tells whether to include * administratively disabled logical processors into the analysis. */ int topo_analyze(struct topo_node *topo_root, int all, struct topo_analysis *results) { results->entities[TOPO_LEVEL_PKG] = -1; results->entities[TOPO_LEVEL_CORE] = -1; results->entities[TOPO_LEVEL_THREAD] = -1; results->entities[TOPO_LEVEL_GROUP] = -1; results->entities[TOPO_LEVEL_CACHEGROUP] = -1; if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results)) return (0); KASSERT(results->entities[TOPO_LEVEL_PKG] > 0, ("bug in topology or analysis")); return (1); } #endif /* SMP */ Index: head/sys/x86/include/x86_smp.h =================================================================== --- head/sys/x86/include/x86_smp.h (revision 347133) +++ head/sys/x86/include/x86_smp.h (revision 347134) @@ -1,111 +1,116 @@ /*- * SPDX-License-Identifier: Beerware * * ---------------------------------------------------------------------------- * "THE BEER-WARE LICENSE" (Revision 42): * wrote this file. As long as you retain this notice you * can do whatever you want with this stuff. If we meet some day, and you think * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp * ---------------------------------------------------------------------------- * * $FreeBSD$ * */ #ifndef _X86_X86_SMP_H_ #define _X86_X86_SMP_H_ #include #include #include #include #include struct pmap; /* global data in mp_x86.c */ extern int mp_naps; extern int boot_cpu_id; extern struct pcb stoppcbs[]; extern int cpu_apic_ids[]; extern int bootAP; extern void *dpcpu; extern char *bootSTK; extern void *bootstacks[]; extern unsigned int boot_address; extern unsigned int bootMP_size; extern volatile u_int cpu_ipi_pending[]; extern volatile int aps_ready; extern struct mtx ap_boot_mtx; extern int cpu_logical; extern int cpu_cores; extern volatile uint32_t smp_tlb_generation; extern struct pmap *smp_tlb_pmap; extern vm_offset_t smp_tlb_addr1, smp_tlb_addr2; extern u_int xhits_gbl[]; extern u_int xhits_pg[]; extern u_int xhits_rng[]; extern u_int ipi_global; extern u_int ipi_page; extern u_int ipi_range; extern u_int ipi_range_size; extern int nmi_kdb_lock; extern int nmi_is_broadcast; struct cpu_info { int cpu_present:1; int cpu_bsp:1; int cpu_disabled:1; int cpu_hyperthread:1; }; extern struct cpu_info *cpu_info; +/* + * Set if MWAIT does not reliably wake when the MONITORed address is written. + */ +extern bool mwait_cpustop_broken; + #ifdef COUNT_IPIS extern u_long *ipi_invltlb_counts[MAXCPU]; extern u_long *ipi_invlrng_counts[MAXCPU]; extern u_long *ipi_invlpg_counts[MAXCPU]; extern u_long *ipi_invlcache_counts[MAXCPU]; extern u_long *ipi_rendezvous_counts[MAXCPU]; #endif /* IPI handlers */ inthand_t IDTVEC(invltlb), /* TLB shootdowns - global */ IDTVEC(invlpg), /* TLB shootdowns - 1 page */ IDTVEC(invlrng), /* TLB shootdowns - page range */ IDTVEC(invlcache), /* Write back and invalidate cache */ IDTVEC(ipi_intr_bitmap_handler), /* Bitmap based IPIs */ IDTVEC(cpustop), /* CPU stops & waits to be restarted */ IDTVEC(cpususpend), /* CPU suspends & waits to be resumed */ IDTVEC(rendezvous); /* handle CPU rendezvous */ /* functions in x86_mp.c */ void assign_cpu_ids(void); void cpu_add(u_int apic_id, char boot_cpu); void cpustop_handler(void); void cpususpend_handler(void); void alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx); void init_secondary_tail(void); void invltlb_handler(void); void invlpg_handler(void); void invlrng_handler(void); void invlcache_handler(void); void init_secondary(void); void ipi_startup(int apic_id, int vector); void ipi_all_but_self(u_int ipi); void ipi_bitmap_handler(struct trapframe frame); void ipi_cpu(int cpu, u_int ipi); int ipi_nmi_handler(void); void ipi_selected(cpuset_t cpus, u_int ipi); void set_interrupt_apic_ids(void); void smp_cache_flush(void); void smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, struct pmap *pmap); void smp_masked_invlpg_range(cpuset_t mask, vm_offset_t startva, vm_offset_t endva, struct pmap *pmap); void smp_masked_invltlb(cpuset_t mask, struct pmap *pmap); void mem_range_AP_init(void); void topo_probe(void); void ipi_send_cpu(int cpu, u_int ipi); #endif Index: head/sys/x86/x86/cpu_machdep.c =================================================================== --- head/sys/x86/x86/cpu_machdep.c (revision 347133) +++ head/sys/x86/x86/cpu_machdep.c (revision 347134) @@ -1,964 +1,977 @@ /*- * Copyright (c) 2003 Peter Wemm. * Copyright (c) 1992 Terrence R. Lambert. * Copyright (c) 1982, 1987, 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. * 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: @(#)machdep.c 7.4 (Berkeley) 6/3/91 */ #include __FBSDID("$FreeBSD$"); #include "opt_acpi.h" #include "opt_atpic.h" #include "opt_cpu.h" #include "opt_ddb.h" #include "opt_inet.h" #include "opt_isa.h" #include "opt_kdb.h" #include "opt_kstack_pages.h" #include "opt_maxmem.h" #include "opt_mp_watchdog.h" #include "opt_platform.h" #ifdef __i386__ #include "opt_apic.h" #endif #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 #ifdef CPU_ELAN #include #endif #include #include #include #include #include #include #include #include #include #include #include #define STATE_RUNNING 0x0 #define STATE_MWAIT 0x1 #define STATE_SLEEPING 0x2 #ifdef SMP static u_int cpu_reset_proxyid; static volatile u_int cpu_reset_proxy_active; #endif +/* + * Automatically initialized per CPU errata in cpu_idle_tun below. + */ +bool mwait_cpustop_broken = false; +SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN, + &mwait_cpustop_broken, 0, + "Can not reliably wake MONITOR/MWAIT cpus without interrupts"); /* * Machine dependent boot() routine * * I haven't seen anything to put here yet * Possibly some stuff might be grafted back here from boot() */ void cpu_boot(int howto) { } /* * Flush the D-cache for non-DMA I/O so that the I-cache can * be made coherent later. */ void cpu_flush_dcache(void *ptr, size_t len) { /* Not applicable */ } void acpi_cpu_c1(void) { __asm __volatile("sti; hlt"); } /* * Use mwait to pause execution while waiting for an interrupt or * another thread to signal that there is more work. * * NOTE: Interrupts will cause a wakeup; however, this function does * not enable interrupt handling. The caller is responsible to enable * interrupts. */ void acpi_cpu_idle_mwait(uint32_t mwait_hint) { int *state; uint64_t v; /* * A comment in Linux patch claims that 'CPUs run faster with * speculation protection disabled. All CPU threads in a core * must disable speculation protection for it to be * disabled. Disable it while we are idle so the other * hyperthread can run fast.' * * XXXKIB. Software coordination mode should be supported, * but all Intel CPUs provide hardware coordination. */ state = &PCPU_PTR(monitorbuf)->idle_state; KASSERT(atomic_load_int(state) == STATE_SLEEPING, ("cpu_mwait_cx: wrong monitorbuf state")); atomic_store_int(state, STATE_MWAIT); if (PCPU_GET(ibpb_set) || hw_ssb_active) { v = rdmsr(MSR_IA32_SPEC_CTRL); wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS | IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD)); } else { v = 0; } cpu_monitor(state, 0, 0); if (atomic_load_int(state) == STATE_MWAIT) cpu_mwait(MWAIT_INTRBREAK, mwait_hint); /* * SSB cannot be disabled while we sleep, or rather, if it was * disabled, the sysctl thread will bind to our cpu to tweak * MSR. */ if (v != 0) wrmsr(MSR_IA32_SPEC_CTRL, v); /* * We should exit on any event that interrupts mwait, because * that event might be a wanted interrupt. */ atomic_store_int(state, STATE_RUNNING); } /* Get current clock frequency for the given cpu id. */ int cpu_est_clockrate(int cpu_id, uint64_t *rate) { uint64_t tsc1, tsc2; uint64_t acnt, mcnt, perf; register_t reg; if (pcpu_find(cpu_id) == NULL || rate == NULL) return (EINVAL); #ifdef __i386__ if ((cpu_feature & CPUID_TSC) == 0) return (EOPNOTSUPP); #endif /* * If TSC is P-state invariant and APERF/MPERF MSRs do not exist, * DELAY(9) based logic fails. */ if (tsc_is_invariant && !tsc_perf_stat) return (EOPNOTSUPP); #ifdef SMP if (smp_cpus > 1) { /* Schedule ourselves on the indicated cpu. */ thread_lock(curthread); sched_bind(curthread, cpu_id); thread_unlock(curthread); } #endif /* Calibrate by measuring a short delay. */ reg = intr_disable(); if (tsc_is_invariant) { wrmsr(MSR_MPERF, 0); wrmsr(MSR_APERF, 0); tsc1 = rdtsc(); DELAY(1000); mcnt = rdmsr(MSR_MPERF); acnt = rdmsr(MSR_APERF); tsc2 = rdtsc(); intr_restore(reg); perf = 1000 * acnt / mcnt; *rate = (tsc2 - tsc1) * perf; } else { tsc1 = rdtsc(); DELAY(1000); tsc2 = rdtsc(); intr_restore(reg); *rate = (tsc2 - tsc1) * 1000; } #ifdef SMP if (smp_cpus > 1) { thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); } #endif return (0); } /* * Shutdown the CPU as much as possible */ void cpu_halt(void) { for (;;) halt(); } static void cpu_reset_real(void) { struct region_descriptor null_idt; int b; disable_intr(); #ifdef CPU_ELAN if (elan_mmcr != NULL) elan_mmcr->RESCFG = 1; #endif #ifdef __i386__ if (cpu == CPU_GEODE1100) { /* Attempt Geode's own reset */ outl(0xcf8, 0x80009044ul); outl(0xcfc, 0xf); } #endif #if !defined(BROKEN_KEYBOARD_RESET) /* * Attempt to do a CPU reset via the keyboard controller, * do not turn off GateA20, as any machine that fails * to do the reset here would then end up in no man's land. */ outb(IO_KBD + 4, 0xFE); DELAY(500000); /* wait 0.5 sec to see if that did it */ #endif /* * Attempt to force a reset via the Reset Control register at * I/O port 0xcf9. Bit 2 forces a system reset when it * transitions from 0 to 1. Bit 1 selects the type of reset * to attempt: 0 selects a "soft" reset, and 1 selects a * "hard" reset. We try a "hard" reset. The first write sets * bit 1 to select a "hard" reset and clears bit 2. The * second write forces a 0 -> 1 transition in bit 2 to trigger * a reset. */ outb(0xcf9, 0x2); outb(0xcf9, 0x6); DELAY(500000); /* wait 0.5 sec to see if that did it */ /* * Attempt to force a reset via the Fast A20 and Init register * at I/O port 0x92. Bit 1 serves as an alternate A20 gate. * Bit 0 asserts INIT# when set to 1. We are careful to only * preserve bit 1 while setting bit 0. We also must clear bit * 0 before setting it if it isn't already clear. */ b = inb(0x92); if (b != 0xff) { if ((b & 0x1) != 0) outb(0x92, b & 0xfe); outb(0x92, b | 0x1); DELAY(500000); /* wait 0.5 sec to see if that did it */ } printf("No known reset method worked, attempting CPU shutdown\n"); DELAY(1000000); /* wait 1 sec for printf to complete */ /* Wipe the IDT. */ null_idt.rd_limit = 0; null_idt.rd_base = 0; lidt(&null_idt); /* "good night, sweet prince .... " */ breakpoint(); /* NOTREACHED */ while(1); } #ifdef SMP static void cpu_reset_proxy(void) { cpu_reset_proxy_active = 1; while (cpu_reset_proxy_active == 1) ia32_pause(); /* Wait for other cpu to see that we've started */ printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid); DELAY(1000000); cpu_reset_real(); } #endif void cpu_reset(void) { #ifdef SMP + struct monitorbuf *mb; cpuset_t map; u_int cnt; if (smp_started) { map = all_cpus; CPU_CLR(PCPU_GET(cpuid), &map); CPU_NAND(&map, &stopped_cpus); if (!CPU_EMPTY(&map)) { printf("cpu_reset: Stopping other CPUs\n"); stop_cpus(map); } if (PCPU_GET(cpuid) != 0) { cpu_reset_proxyid = PCPU_GET(cpuid); cpustop_restartfunc = cpu_reset_proxy; cpu_reset_proxy_active = 0; printf("cpu_reset: Restarting BSP\n"); /* Restart CPU #0. */ CPU_SETOF(0, &started_cpus); + mb = &pcpu_find(0)->pc_monitorbuf; + atomic_store_int(&mb->stop_state, + MONITOR_STOPSTATE_RUNNING); wmb(); cnt = 0; while (cpu_reset_proxy_active == 0 && cnt < 10000000) { ia32_pause(); cnt++; /* Wait for BSP to announce restart */ } if (cpu_reset_proxy_active == 0) { printf("cpu_reset: Failed to restart BSP\n"); } else { cpu_reset_proxy_active = 2; while (1) ia32_pause(); /* NOTREACHED */ } } DELAY(1000000); } #endif cpu_reset_real(); /* NOTREACHED */ } bool cpu_mwait_usable(void) { return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags & (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) == (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK))); } void (*cpu_idle_hook)(sbintime_t) = NULL; /* ACPI idle hook. */ static int cpu_ident_amdc1e = 0; /* AMD C1E supported. */ static int idle_mwait = 1; /* Use MONITOR/MWAIT for short idle. */ SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait, 0, "Use MONITOR/MWAIT for short idle"); static void cpu_idle_acpi(sbintime_t sbt) { int *state; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_SLEEPING); /* See comments in cpu_idle_hlt(). */ disable_intr(); if (sched_runnable()) enable_intr(); else if (cpu_idle_hook) cpu_idle_hook(sbt); else acpi_cpu_c1(); atomic_store_int(state, STATE_RUNNING); } static void cpu_idle_hlt(sbintime_t sbt) { int *state; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_SLEEPING); /* * Since we may be in a critical section from cpu_idle(), if * an interrupt fires during that critical section we may have * a pending preemption. If the CPU halts, then that thread * may not execute until a later interrupt awakens the CPU. * To handle this race, check for a runnable thread after * disabling interrupts and immediately return if one is * found. Also, we must absolutely guarentee that hlt is * the next instruction after sti. This ensures that any * interrupt that fires after the call to disable_intr() will * immediately awaken the CPU from hlt. Finally, please note * that on x86 this works fine because of interrupts enabled only * after the instruction following sti takes place, while IF is set * to 1 immediately, allowing hlt instruction to acknowledge the * interrupt. */ disable_intr(); if (sched_runnable()) enable_intr(); else acpi_cpu_c1(); atomic_store_int(state, STATE_RUNNING); } static void cpu_idle_mwait(sbintime_t sbt) { int *state; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_MWAIT); /* See comments in cpu_idle_hlt(). */ disable_intr(); if (sched_runnable()) { atomic_store_int(state, STATE_RUNNING); enable_intr(); return; } cpu_monitor(state, 0, 0); if (atomic_load_int(state) == STATE_MWAIT) __asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0)); else enable_intr(); atomic_store_int(state, STATE_RUNNING); } static void cpu_idle_spin(sbintime_t sbt) { int *state; int i; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_RUNNING); /* * The sched_runnable() call is racy but as long as there is * a loop missing it one time will have just a little impact if any * (and it is much better than missing the check at all). */ for (i = 0; i < 1000; i++) { if (sched_runnable()) return; cpu_spinwait(); } } /* * C1E renders the local APIC timer dead, so we disable it by * reading the Interrupt Pending Message register and clearing * both C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27). * * Reference: * "BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors" * #32559 revision 3.00+ */ #define MSR_AMDK8_IPM 0xc0010055 #define AMDK8_SMIONCMPHALT (1ULL << 27) #define AMDK8_C1EONCMPHALT (1ULL << 28) #define AMDK8_CMPHALT (AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT) void cpu_probe_amdc1e(void) { /* * Detect the presence of C1E capability mostly on latest * dual-cores (or future) k8 family. */ if (cpu_vendor_id == CPU_VENDOR_AMD && (cpu_id & 0x00000f00) == 0x00000f00 && (cpu_id & 0x0fff0000) >= 0x00040000) { cpu_ident_amdc1e = 1; } } void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi; void cpu_idle(int busy) { uint64_t msr; sbintime_t sbt = -1; CTR2(KTR_SPARE2, "cpu_idle(%d) at %d", busy, curcpu); #ifdef MP_WATCHDOG ap_watchdog(PCPU_GET(cpuid)); #endif /* If we are busy - try to use fast methods. */ if (busy) { if ((cpu_feature2 & CPUID2_MON) && idle_mwait) { cpu_idle_mwait(busy); goto out; } } /* If we have time - switch timers into idle mode. */ if (!busy) { critical_enter(); sbt = cpu_idleclock(); } /* Apply AMD APIC timer C1E workaround. */ if (cpu_ident_amdc1e && cpu_disable_c3_sleep) { msr = rdmsr(MSR_AMDK8_IPM); if (msr & AMDK8_CMPHALT) wrmsr(MSR_AMDK8_IPM, msr & ~AMDK8_CMPHALT); } /* Call main idle method. */ cpu_idle_fn(sbt); /* Switch timers back into active mode. */ if (!busy) { cpu_activeclock(); critical_exit(); } out: CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done", busy, curcpu); } static int cpu_idle_apl31_workaround; SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW, &cpu_idle_apl31_workaround, 0, "Apollo Lake APL31 MWAIT bug workaround"); int cpu_idle_wakeup(int cpu) { struct monitorbuf *mb; int *state; mb = &pcpu_find(cpu)->pc_monitorbuf; state = &mb->idle_state; switch (atomic_load_int(state)) { case STATE_SLEEPING: return (0); case STATE_MWAIT: atomic_store_int(state, STATE_RUNNING); return (cpu_idle_apl31_workaround ? 0 : 1); case STATE_RUNNING: return (1); default: panic("bad monitor state"); return (1); } } /* * Ordered by speed/power consumption. */ static struct { void *id_fn; char *id_name; int id_cpuid2_flag; } idle_tbl[] = { { .id_fn = cpu_idle_spin, .id_name = "spin" }, { .id_fn = cpu_idle_mwait, .id_name = "mwait", .id_cpuid2_flag = CPUID2_MON }, { .id_fn = cpu_idle_hlt, .id_name = "hlt" }, { .id_fn = cpu_idle_acpi, .id_name = "acpi" }, }; static int idle_sysctl_available(SYSCTL_HANDLER_ARGS) { char *avail, *p; int error; int i; avail = malloc(256, M_TEMP, M_WAITOK); p = avail; for (i = 0; i < nitems(idle_tbl); i++) { if (idle_tbl[i].id_cpuid2_flag != 0 && (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0) continue; if (strcmp(idle_tbl[i].id_name, "acpi") == 0 && cpu_idle_hook == NULL) continue; p += sprintf(p, "%s%s", p != avail ? ", " : "", idle_tbl[i].id_name); } error = sysctl_handle_string(oidp, avail, 0, req); free(avail, M_TEMP); return (error); } SYSCTL_PROC(_machdep, OID_AUTO, idle_available, CTLTYPE_STRING | CTLFLAG_RD, 0, 0, idle_sysctl_available, "A", "list of available idle functions"); static bool cpu_idle_selector(const char *new_idle_name) { int i; for (i = 0; i < nitems(idle_tbl); i++) { if (idle_tbl[i].id_cpuid2_flag != 0 && (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0) continue; if (strcmp(idle_tbl[i].id_name, "acpi") == 0 && cpu_idle_hook == NULL) continue; if (strcmp(idle_tbl[i].id_name, new_idle_name)) continue; cpu_idle_fn = idle_tbl[i].id_fn; if (bootverbose) printf("CPU idle set to %s\n", idle_tbl[i].id_name); return (true); } return (false); } static int cpu_idle_sysctl(SYSCTL_HANDLER_ARGS) { char buf[16], *p; int error, i; p = "unknown"; for (i = 0; i < nitems(idle_tbl); i++) { if (idle_tbl[i].id_fn == cpu_idle_fn) { p = idle_tbl[i].id_name; break; } } strncpy(buf, p, sizeof(buf)); error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); return (cpu_idle_selector(buf) ? 0 : EINVAL); } SYSCTL_PROC(_machdep, OID_AUTO, idle, CTLTYPE_STRING | CTLFLAG_RW, 0, 0, cpu_idle_sysctl, "A", "currently selected idle function"); static void cpu_idle_tun(void *unused __unused) { char tunvar[16]; if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar))) cpu_idle_selector(tunvar); else if (cpu_vendor_id == CPU_VENDOR_AMD && CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) { /* Ryzen erratas 1057, 1109. */ cpu_idle_selector("hlt"); idle_mwait = 0; + mwait_cpustop_broken = true; } if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_id == 0x506c9) { /* * Apollo Lake errata APL31 (public errata APL30). * Stores to the armed address range may not trigger * MWAIT to resume execution. OS needs to use * interrupts to wake processors from MWAIT-induced * sleep states. */ cpu_idle_apl31_workaround = 1; + mwait_cpustop_broken = true; } TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround); } SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL); static int panic_on_nmi = 1; SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN, &panic_on_nmi, 0, "Panic on NMI raised by hardware failure"); int nmi_is_broadcast = 1; SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN, &nmi_is_broadcast, 0, "Chipset NMI is broadcast"); #ifdef KDB int kdb_on_nmi = 1; SYSCTL_INT(_machdep, OID_AUTO, kdb_on_nmi, CTLFLAG_RWTUN, &kdb_on_nmi, 0, "Go to KDB on NMI with unknown source"); #endif void nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame) { bool claimed = false; #ifdef DEV_ISA /* machine/parity/power fail/"kitchen sink" faults */ if (isa_nmi(frame->tf_err)) { claimed = true; if (panic_on_nmi) panic("NMI indicates hardware failure"); } #endif /* DEV_ISA */ #ifdef KDB if (!claimed && kdb_on_nmi) { /* * NMI can be hooked up to a pushbutton for debugging. */ printf("NMI/cpu%d ... going to debugger\n", cpu); kdb_trap(type, 0, frame); } #endif /* KDB */ } void nmi_handle_intr(u_int type, struct trapframe *frame) { #ifdef SMP if (nmi_is_broadcast) { nmi_call_kdb_smp(type, frame); return; } #endif nmi_call_kdb(PCPU_GET(cpuid), type, frame); } int hw_ibrs_active; int hw_ibrs_disable = 1; SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active"); void hw_ibrs_recalculate(void) { uint64_t v; if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) { if (hw_ibrs_disable) { v = rdmsr(MSR_IA32_SPEC_CTRL); v &= ~(uint64_t)IA32_SPEC_CTRL_IBRS; wrmsr(MSR_IA32_SPEC_CTRL, v); } else { v = rdmsr(MSR_IA32_SPEC_CTRL); v |= IA32_SPEC_CTRL_IBRS; wrmsr(MSR_IA32_SPEC_CTRL, v); } return; } hw_ibrs_active = (cpu_stdext_feature3 & CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable; } static int hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = hw_ibrs_disable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); hw_ibrs_disable = val != 0; hw_ibrs_recalculate(); return (0); } SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I", "Disable Indirect Branch Restricted Speculation"); int hw_ssb_active; int hw_ssb_disable; SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD, &hw_ssb_active, 0, "Speculative Store Bypass Disable active"); static void hw_ssb_set_one(bool enable) { uint64_t v; v = rdmsr(MSR_IA32_SPEC_CTRL); if (enable) v |= (uint64_t)IA32_SPEC_CTRL_SSBD; else v &= ~(uint64_t)IA32_SPEC_CTRL_SSBD; wrmsr(MSR_IA32_SPEC_CTRL, v); } static void hw_ssb_set(bool enable, bool for_all_cpus) { struct thread *td; int bound_cpu, i, is_bound; if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) { hw_ssb_active = 0; return; } hw_ssb_active = enable; if (for_all_cpus) { td = curthread; thread_lock(td); is_bound = sched_is_bound(td); bound_cpu = td->td_oncpu; CPU_FOREACH(i) { sched_bind(td, i); hw_ssb_set_one(enable); } if (is_bound) sched_bind(td, bound_cpu); else sched_unbind(td); thread_unlock(td); } else { hw_ssb_set_one(enable); } } void hw_ssb_recalculate(bool all_cpus) { switch (hw_ssb_disable) { default: hw_ssb_disable = 0; /* FALLTHROUGH */ case 0: /* off */ hw_ssb_set(false, all_cpus); break; case 1: /* on */ hw_ssb_set(true, all_cpus); break; case 2: /* auto */ hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ? false : true, all_cpus); break; } } static int hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = hw_ssb_disable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); hw_ssb_disable = val; hw_ssb_recalculate(true); return (0); } SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ssb_disable_handler, "I", "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto"); /* * Enable and restore kernel text write permissions. * Callers must ensure that disable_wp()/restore_wp() are executed * without rescheduling on the same core. */ bool disable_wp(void) { u_int cr0; cr0 = rcr0(); if ((cr0 & CR0_WP) == 0) return (false); load_cr0(cr0 & ~CR0_WP); return (true); } void restore_wp(bool old_wp) { if (old_wp) load_cr0(rcr0() | CR0_WP); } bool acpi_get_fadt_bootflags(uint16_t *flagsp) { #ifdef DEV_ACPI ACPI_TABLE_FADT *fadt; vm_paddr_t physaddr; physaddr = acpi_find_table(ACPI_SIG_FADT); if (physaddr == 0) return (false); fadt = acpi_map_table(physaddr, ACPI_SIG_FADT); if (fadt == NULL) return (false); *flagsp = fadt->BootFlags; acpi_unmap_table(fadt); return (true); #else return (false); #endif } Index: head/sys/x86/x86/mp_x86.c =================================================================== --- head/sys/x86/x86/mp_x86.c (revision 347133) +++ head/sys/x86/x86/mp_x86.c (revision 347134) @@ -1,1799 +1,1821 @@ /*- * Copyright (c) 1996, by Steve Passe * Copyright (c) 2003, by Peter Wemm * 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. The name of the developer may NOT be used to endorse or promote products * derived from this software without specific prior written permission. * * 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("$FreeBSD$"); #ifdef __i386__ #include "opt_apic.h" #endif #include "opt_cpu.h" #include "opt_kstack_pages.h" #include "opt_pmap.h" #include "opt_sched.h" #include "opt_smp.h" #include #include #include #include /* cngetc() */ #include #ifdef GPROF #include #endif #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 static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items"); /* lock region used by kernel profiling */ int mcount_lock; int mp_naps; /* # of Applications processors */ int boot_cpu_id = -1; /* designated BSP */ /* AP uses this during bootstrap. Do not staticize. */ char *bootSTK; int bootAP; /* Free these after use */ void *bootstacks[MAXCPU]; void *dpcpu; struct pcb stoppcbs[MAXCPU]; struct susppcb **susppcbs; #ifdef COUNT_IPIS /* Interrupt counts. */ static u_long *ipi_preempt_counts[MAXCPU]; static u_long *ipi_ast_counts[MAXCPU]; u_long *ipi_invltlb_counts[MAXCPU]; u_long *ipi_invlrng_counts[MAXCPU]; u_long *ipi_invlpg_counts[MAXCPU]; u_long *ipi_invlcache_counts[MAXCPU]; u_long *ipi_rendezvous_counts[MAXCPU]; static u_long *ipi_hardclock_counts[MAXCPU]; #endif /* Default cpu_ops implementation. */ struct cpu_ops cpu_ops; /* * Local data and functions. */ static volatile cpuset_t ipi_stop_nmi_pending; volatile cpuset_t resuming_cpus; volatile cpuset_t toresume_cpus; /* used to hold the AP's until we are ready to release them */ struct mtx ap_boot_mtx; /* Set to 1 once we're ready to let the APs out of the pen. */ volatile int aps_ready = 0; /* * Store data from cpu_add() until later in the boot when we actually setup * the APs. */ struct cpu_info *cpu_info; int *apic_cpuids; int cpu_apic_ids[MAXCPU]; _Static_assert(MAXCPU <= MAX_APIC_ID, "MAXCPU cannot be larger that MAX_APIC_ID"); _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID, "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID"); /* Holds pending bitmap based IPIs per CPU */ volatile u_int cpu_ipi_pending[MAXCPU]; static void release_aps(void *dummy); static void cpustop_handler_post(u_int cpu); static int hyperthreading_allowed = 1; SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN, &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs"); static struct topo_node topo_root; static int pkg_id_shift; static int node_id_shift; static int core_id_shift; static int disabled_cpus; struct cache_info { int id_shift; int present; } static caches[MAX_CACHE_LEVELS]; unsigned int boot_address; +static bool stop_mwait = false; +SYSCTL_BOOL(_machdep, OID_AUTO, stop_mwait, CTLFLAG_RWTUN, &stop_mwait, 0, + "Use MONITOR/MWAIT when stopping CPU, if available"); + #define MiB(v) (v ## ULL << 20) void mem_range_AP_init(void) { if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP) mem_range_softc.mr_op->initAP(&mem_range_softc); } /* * Round up to the next power of two, if necessary, and then * take log2. * Returns -1 if argument is zero. */ static __inline int mask_width(u_int x) { return (fls(x << (1 - powerof2(x))) - 1); } /* * Add a cache level to the cache topology description. */ static int add_deterministic_cache(int type, int level, int share_count) { if (type == 0) return (0); if (type > 3) { printf("unexpected cache type %d\n", type); return (1); } if (type == 2) /* ignore instruction cache */ return (1); if (level == 0 || level > MAX_CACHE_LEVELS) { printf("unexpected cache level %d\n", type); return (1); } if (caches[level - 1].present) { printf("WARNING: multiple entries for L%u data cache\n", level); printf("%u => %u\n", caches[level - 1].id_shift, mask_width(share_count)); } caches[level - 1].id_shift = mask_width(share_count); caches[level - 1].present = 1; if (caches[level - 1].id_shift > pkg_id_shift) { printf("WARNING: L%u data cache covers more " "APIC IDs than a package (%u > %u)\n", level, caches[level - 1].id_shift, pkg_id_shift); caches[level - 1].id_shift = pkg_id_shift; } if (caches[level - 1].id_shift < core_id_shift) { printf("WARNING: L%u data cache covers fewer " "APIC IDs than a core (%u < %u)\n", level, caches[level - 1].id_shift, core_id_shift); caches[level - 1].id_shift = core_id_shift; } return (1); } /* * Determine topology of processing units and caches for AMD CPUs. * See: * - AMD CPUID Specification (Publication # 25481) * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559) * - BKDG For AMD Family 10h Processors (Publication # 31116) * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301) * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751) * - PPR For AMD Family 17h Models 00h-0Fh Processors (Publication # 54945) */ static void topo_probe_amd(void) { u_int p[4]; uint64_t v; int level; int nodes_per_socket; int share_count; int type; int i; /* No multi-core capability. */ if ((amd_feature2 & AMDID2_CMP) == 0) return; /* For families 10h and newer. */ pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >> AMDID_COREID_SIZE_SHIFT; /* For 0Fh family. */ if (pkg_id_shift == 0) pkg_id_shift = mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1); /* * Families prior to 16h define the following value as * cores per compute unit and we don't really care about the AMD * compute units at the moment. Perhaps we should treat them as * cores and cores within the compute units as hardware threads, * but that's up for debate. * Later families define the value as threads per compute unit, * so we are following AMD's nomenclature here. */ if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 && CPUID_TO_FAMILY(cpu_id) >= 0x16) { cpuid_count(0x8000001e, 0, p); share_count = ((p[1] >> 8) & 0xff) + 1; core_id_shift = mask_width(share_count); /* * For Zen (17h), gather Nodes per Processor. Each node is a * Zeppelin die; TR and EPYC CPUs will have multiple dies per * package. Communication latency between dies is higher than * within them. */ nodes_per_socket = ((p[2] >> 8) & 0x7) + 1; node_id_shift = pkg_id_shift - mask_width(nodes_per_socket); } if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) { for (i = 0; ; i++) { cpuid_count(0x8000001d, i, p); type = p[0] & 0x1f; level = (p[0] >> 5) & 0x7; share_count = 1 + ((p[0] >> 14) & 0xfff); if (!add_deterministic_cache(type, level, share_count)) break; } } else { if (cpu_exthigh >= 0x80000005) { cpuid_count(0x80000005, 0, p); if (((p[2] >> 24) & 0xff) != 0) { caches[0].id_shift = 0; caches[0].present = 1; } } if (cpu_exthigh >= 0x80000006) { cpuid_count(0x80000006, 0, p); if (((p[2] >> 16) & 0xffff) != 0) { caches[1].id_shift = 0; caches[1].present = 1; } if (((p[3] >> 18) & 0x3fff) != 0) { nodes_per_socket = 1; if ((amd_feature2 & AMDID2_NODE_ID) != 0) { /* * Handle multi-node processors that * have multiple chips, each with its * own L3 cache, on the same die. */ v = rdmsr(0xc001100c); nodes_per_socket = 1 + ((v >> 3) & 0x7); } caches[2].id_shift = pkg_id_shift - mask_width(nodes_per_socket); caches[2].present = 1; } } } } /* * Determine topology of processing units for Intel CPUs * using CPUID Leaf 1 and Leaf 4, if supported. * See: * - Intel 64 Architecture Processor Topology Enumeration * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual, * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS */ static void topo_probe_intel_0x4(void) { u_int p[4]; int max_cores; int max_logical; /* Both zero and one here mean one logical processor per package. */ max_logical = (cpu_feature & CPUID_HTT) != 0 ? (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1; if (max_logical <= 1) return; if (cpu_high >= 0x4) { cpuid_count(0x04, 0, p); max_cores = ((p[0] >> 26) & 0x3f) + 1; } else max_cores = 1; core_id_shift = mask_width(max_logical/max_cores); KASSERT(core_id_shift >= 0, ("intel topo: max_cores > max_logical\n")); pkg_id_shift = core_id_shift + mask_width(max_cores); } /* * Determine topology of processing units for Intel CPUs * using CPUID Leaf 11, if supported. * See: * - Intel 64 Architecture Processor Topology Enumeration * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual, * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS */ static void topo_probe_intel_0xb(void) { u_int p[4]; int bits; int type; int i; /* Fall back if CPU leaf 11 doesn't really exist. */ cpuid_count(0x0b, 0, p); if (p[1] == 0) { topo_probe_intel_0x4(); return; } /* We only support three levels for now. */ for (i = 0; ; i++) { cpuid_count(0x0b, i, p); bits = p[0] & 0x1f; type = (p[2] >> 8) & 0xff; if (type == 0) break; /* TODO: check for duplicate (re-)assignment */ if (type == CPUID_TYPE_SMT) core_id_shift = bits; else if (type == CPUID_TYPE_CORE) pkg_id_shift = bits; else printf("unknown CPU level type %d\n", type); } if (pkg_id_shift < core_id_shift) { printf("WARNING: core covers more APIC IDs than a package\n"); core_id_shift = pkg_id_shift; } } /* * Determine topology of caches for Intel CPUs. * See: * - Intel 64 Architecture Processor Topology Enumeration * - Intel 64 and IA-32 Architectures Software Developer’s Manual * Volume 2A: Instruction Set Reference, A-M, * CPUID instruction */ static void topo_probe_intel_caches(void) { u_int p[4]; int level; int share_count; int type; int i; if (cpu_high < 0x4) { /* * Available cache level and sizes can be determined * via CPUID leaf 2, but that requires a huge table of hardcoded * values, so for now just assume L1 and L2 caches potentially * shared only by HTT processing units, if HTT is present. */ caches[0].id_shift = pkg_id_shift; caches[0].present = 1; caches[1].id_shift = pkg_id_shift; caches[1].present = 1; return; } for (i = 0; ; i++) { cpuid_count(0x4, i, p); type = p[0] & 0x1f; level = (p[0] >> 5) & 0x7; share_count = 1 + ((p[0] >> 14) & 0xfff); if (!add_deterministic_cache(type, level, share_count)) break; } } /* * Determine topology of processing units and caches for Intel CPUs. * See: * - Intel 64 Architecture Processor Topology Enumeration */ static void topo_probe_intel(void) { /* * Note that 0x1 <= cpu_high < 4 case should be * compatible with topo_probe_intel_0x4() logic when * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1) * or it should trigger the fallback otherwise. */ if (cpu_high >= 0xb) topo_probe_intel_0xb(); else if (cpu_high >= 0x1) topo_probe_intel_0x4(); topo_probe_intel_caches(); } /* * Topology information is queried only on BSP, on which this * code runs and for which it can query CPUID information. * Then topology is extrapolated on all packages using an * assumption that APIC ID to hardware component ID mapping is * homogenious. * That doesn't necesserily imply that the topology is uniform. */ void topo_probe(void) { static int cpu_topo_probed = 0; struct x86_topo_layer { int type; int subtype; int id_shift; } topo_layers[MAX_CACHE_LEVELS + 4]; struct topo_node *parent; struct topo_node *node; int layer; int nlayers; int node_id; int i; if (cpu_topo_probed) return; CPU_ZERO(&logical_cpus_mask); if (mp_ncpus <= 1) ; /* nothing */ else if (cpu_vendor_id == CPU_VENDOR_AMD) topo_probe_amd(); else if (cpu_vendor_id == CPU_VENDOR_INTEL) topo_probe_intel(); KASSERT(pkg_id_shift >= core_id_shift, ("bug in APIC topology discovery")); nlayers = 0; bzero(topo_layers, sizeof(topo_layers)); topo_layers[nlayers].type = TOPO_TYPE_PKG; topo_layers[nlayers].id_shift = pkg_id_shift; if (bootverbose) printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift); nlayers++; if (pkg_id_shift > node_id_shift && node_id_shift != 0) { topo_layers[nlayers].type = TOPO_TYPE_GROUP; topo_layers[nlayers].id_shift = node_id_shift; if (bootverbose) printf("Node ID shift: %u\n", topo_layers[nlayers].id_shift); nlayers++; } /* * Consider all caches to be within a package/chip * and "in front" of all sub-components like * cores and hardware threads. */ for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) { if (caches[i].present) { if (node_id_shift != 0) KASSERT(caches[i].id_shift <= node_id_shift, ("bug in APIC topology discovery")); KASSERT(caches[i].id_shift <= pkg_id_shift, ("bug in APIC topology discovery")); KASSERT(caches[i].id_shift >= core_id_shift, ("bug in APIC topology discovery")); topo_layers[nlayers].type = TOPO_TYPE_CACHE; topo_layers[nlayers].subtype = i + 1; topo_layers[nlayers].id_shift = caches[i].id_shift; if (bootverbose) printf("L%u cache ID shift: %u\n", topo_layers[nlayers].subtype, topo_layers[nlayers].id_shift); nlayers++; } } if (pkg_id_shift > core_id_shift) { topo_layers[nlayers].type = TOPO_TYPE_CORE; topo_layers[nlayers].id_shift = core_id_shift; if (bootverbose) printf("Core ID shift: %u\n", topo_layers[nlayers].id_shift); nlayers++; } topo_layers[nlayers].type = TOPO_TYPE_PU; topo_layers[nlayers].id_shift = 0; nlayers++; topo_init_root(&topo_root); for (i = 0; i <= max_apic_id; ++i) { if (!cpu_info[i].cpu_present) continue; parent = &topo_root; for (layer = 0; layer < nlayers; ++layer) { node_id = i >> topo_layers[layer].id_shift; parent = topo_add_node_by_hwid(parent, node_id, topo_layers[layer].type, topo_layers[layer].subtype); } } parent = &topo_root; for (layer = 0; layer < nlayers; ++layer) { node_id = boot_cpu_id >> topo_layers[layer].id_shift; node = topo_find_node_by_hwid(parent, node_id, topo_layers[layer].type, topo_layers[layer].subtype); topo_promote_child(node); parent = node; } cpu_topo_probed = 1; } /* * Assign logical CPU IDs to local APICs. */ void assign_cpu_ids(void) { struct topo_node *node; u_int smt_mask; int nhyper; smt_mask = (1u << core_id_shift) - 1; /* * Assign CPU IDs to local APIC IDs and disable any CPUs * beyond MAXCPU. CPU 0 is always assigned to the BSP. */ mp_ncpus = 0; nhyper = 0; TOPO_FOREACH(node, &topo_root) { if (node->type != TOPO_TYPE_PU) continue; if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask)) cpu_info[node->hwid].cpu_hyperthread = 1; if (resource_disabled("lapic", node->hwid)) { if (node->hwid != boot_cpu_id) cpu_info[node->hwid].cpu_disabled = 1; else printf("Cannot disable BSP, APIC ID = %d\n", node->hwid); } if (!hyperthreading_allowed && cpu_info[node->hwid].cpu_hyperthread) cpu_info[node->hwid].cpu_disabled = 1; if (mp_ncpus >= MAXCPU) cpu_info[node->hwid].cpu_disabled = 1; if (cpu_info[node->hwid].cpu_disabled) { disabled_cpus++; continue; } if (cpu_info[node->hwid].cpu_hyperthread) nhyper++; cpu_apic_ids[mp_ncpus] = node->hwid; apic_cpuids[node->hwid] = mp_ncpus; topo_set_pu_id(node, mp_ncpus); mp_ncpus++; } KASSERT(mp_maxid >= mp_ncpus - 1, ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid, mp_ncpus)); mp_ncores = mp_ncpus - nhyper; smp_threads_per_core = mp_ncpus / mp_ncores; } /* * Print various information about the SMP system hardware and setup. */ void cpu_mp_announce(void) { struct topo_node *node; const char *hyperthread; struct topo_analysis topology; printf("FreeBSD/SMP: "); if (topo_analyze(&topo_root, 1, &topology)) { printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]); if (topology.entities[TOPO_LEVEL_GROUP] > 1) printf(" x %d groups", topology.entities[TOPO_LEVEL_GROUP]); if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1) printf(" x %d cache groups", topology.entities[TOPO_LEVEL_CACHEGROUP]); if (topology.entities[TOPO_LEVEL_CORE] > 0) printf(" x %d core(s)", topology.entities[TOPO_LEVEL_CORE]); if (topology.entities[TOPO_LEVEL_THREAD] > 1) printf(" x %d hardware threads", topology.entities[TOPO_LEVEL_THREAD]); } else { printf("Non-uniform topology"); } printf("\n"); if (disabled_cpus) { printf("FreeBSD/SMP Online: "); if (topo_analyze(&topo_root, 0, &topology)) { printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]); if (topology.entities[TOPO_LEVEL_GROUP] > 1) printf(" x %d groups", topology.entities[TOPO_LEVEL_GROUP]); if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1) printf(" x %d cache groups", topology.entities[TOPO_LEVEL_CACHEGROUP]); if (topology.entities[TOPO_LEVEL_CORE] > 0) printf(" x %d core(s)", topology.entities[TOPO_LEVEL_CORE]); if (topology.entities[TOPO_LEVEL_THREAD] > 1) printf(" x %d hardware threads", topology.entities[TOPO_LEVEL_THREAD]); } else { printf("Non-uniform topology"); } printf("\n"); } if (!bootverbose) return; TOPO_FOREACH(node, &topo_root) { switch (node->type) { case TOPO_TYPE_PKG: printf("Package HW ID = %u\n", node->hwid); break; case TOPO_TYPE_CORE: printf("\tCore HW ID = %u\n", node->hwid); break; case TOPO_TYPE_PU: if (cpu_info[node->hwid].cpu_hyperthread) hyperthread = "/HT"; else hyperthread = ""; if (node->subtype == 0) printf("\t\tCPU (AP%s): APIC ID: %u" "(disabled)\n", hyperthread, node->hwid); else if (node->id == 0) printf("\t\tCPU0 (BSP): APIC ID: %u\n", node->hwid); else printf("\t\tCPU%u (AP%s): APIC ID: %u\n", node->id, hyperthread, node->hwid); break; default: /* ignored */ break; } } } /* * Add a scheduling group, a group of logical processors sharing * a particular cache (and, thus having an affinity), to the scheduling * topology. * This function recursively works on lower level caches. */ static void x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root) { struct topo_node *node; int nchildren; int ncores; int i; KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE || root->type == TOPO_TYPE_GROUP, ("x86topo_add_sched_group: bad type: %u", root->type)); CPU_COPY(&root->cpuset, &cg_root->cg_mask); cg_root->cg_count = root->cpu_count; if (root->type == TOPO_TYPE_SYSTEM) cg_root->cg_level = CG_SHARE_NONE; else cg_root->cg_level = root->subtype; /* * Check how many core nodes we have under the given root node. * If we have multiple logical processors, but not multiple * cores, then those processors must be hardware threads. */ ncores = 0; node = root; while (node != NULL) { if (node->type != TOPO_TYPE_CORE) { node = topo_next_node(root, node); continue; } ncores++; node = topo_next_nonchild_node(root, node); } if (cg_root->cg_level != CG_SHARE_NONE && root->cpu_count > 1 && ncores < 2) cg_root->cg_flags = CG_FLAG_SMT; /* * Find out how many cache nodes we have under the given root node. * We ignore cache nodes that cover all the same processors as the * root node. Also, we do not descend below found cache nodes. * That is, we count top-level "non-redundant" caches under the root * node. */ nchildren = 0; node = root; while (node != NULL) { if ((node->type != TOPO_TYPE_GROUP && node->type != TOPO_TYPE_CACHE) || (root->type != TOPO_TYPE_SYSTEM && CPU_CMP(&node->cpuset, &root->cpuset) == 0)) { node = topo_next_node(root, node); continue; } nchildren++; node = topo_next_nonchild_node(root, node); } cg_root->cg_child = smp_topo_alloc(nchildren); cg_root->cg_children = nchildren; /* * Now find again the same cache nodes as above and recursively * build scheduling topologies for them. */ node = root; i = 0; while (node != NULL) { if ((node->type != TOPO_TYPE_GROUP && node->type != TOPO_TYPE_CACHE) || (root->type != TOPO_TYPE_SYSTEM && CPU_CMP(&node->cpuset, &root->cpuset) == 0)) { node = topo_next_node(root, node); continue; } cg_root->cg_child[i].cg_parent = cg_root; x86topo_add_sched_group(node, &cg_root->cg_child[i]); i++; node = topo_next_nonchild_node(root, node); } } /* * Build the MI scheduling topology from the discovered hardware topology. */ struct cpu_group * cpu_topo(void) { struct cpu_group *cg_root; if (mp_ncpus <= 1) return (smp_topo_none()); cg_root = smp_topo_alloc(1); x86topo_add_sched_group(&topo_root, cg_root); return (cg_root); } static void cpu_alloc(void *dummy __unused) { /* * Dynamically allocate the arrays that depend on the * maximum APIC ID. */ cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS, M_WAITOK | M_ZERO); apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS, M_WAITOK | M_ZERO); } SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL); /* * Add a logical CPU to the topology. */ void cpu_add(u_int apic_id, char boot_cpu) { if (apic_id > max_apic_id) { panic("SMP: APIC ID %d too high", apic_id); return; } KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice", apic_id)); cpu_info[apic_id].cpu_present = 1; if (boot_cpu) { KASSERT(boot_cpu_id == -1, ("CPU %u claims to be BSP, but CPU %u already is", apic_id, boot_cpu_id)); boot_cpu_id = apic_id; cpu_info[apic_id].cpu_bsp = 1; } if (bootverbose) printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" : "AP"); } void cpu_mp_setmaxid(void) { /* * mp_ncpus and mp_maxid should be already set by calls to cpu_add(). * If there were no calls to cpu_add() assume this is a UP system. */ if (mp_ncpus == 0) mp_ncpus = 1; } int cpu_mp_probe(void) { /* * Always record BSP in CPU map so that the mbuf init code works * correctly. */ CPU_SETOF(0, &all_cpus); return (mp_ncpus > 1); } /* Allocate memory for the AP trampoline. */ void alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx) { unsigned int i; bool allocated; allocated = false; for (i = *physmap_idx; i <= *physmap_idx; i -= 2) { /* * Find a memory region big enough and below the 1MB boundary * for the trampoline code. * NB: needs to be page aligned. */ if (physmap[i] >= MiB(1) || (trunc_page(physmap[i + 1]) - round_page(physmap[i])) < round_page(bootMP_size)) continue; allocated = true; /* * Try to steal from the end of the region to mimic previous * behaviour, else fallback to steal from the start. */ if (physmap[i + 1] < MiB(1)) { boot_address = trunc_page(physmap[i + 1]); if ((physmap[i + 1] - boot_address) < bootMP_size) boot_address -= round_page(bootMP_size); physmap[i + 1] = boot_address; } else { boot_address = round_page(physmap[i]); physmap[i] = boot_address + round_page(bootMP_size); } if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) { memmove(&physmap[i], &physmap[i + 2], sizeof(*physmap) * (*physmap_idx - i + 2)); *physmap_idx -= 2; } break; } if (!allocated) { boot_address = basemem * 1024 - bootMP_size; if (bootverbose) printf( "Cannot find enough space for the boot trampoline, placing it at %#x", boot_address); } } /* * AP CPU's call this to initialize themselves. */ void init_secondary_tail(void) { u_int cpuid; pmap_activate_boot(vmspace_pmap(proc0.p_vmspace)); /* * On real hardware, switch to x2apic mode if possible. Do it * after aps_ready was signalled, to avoid manipulating the * mode while BSP might still want to send some IPI to us * (second startup IPI is ignored on modern hardware etc). */ lapic_xapic_mode(); /* Initialize the PAT MSR. */ pmap_init_pat(); /* set up CPU registers and state */ cpu_setregs(); /* set up SSE/NX */ initializecpu(); /* set up FPU state on the AP */ #ifdef __amd64__ fpuinit(); #else npxinit(false); #endif if (cpu_ops.cpu_init) cpu_ops.cpu_init(); /* A quick check from sanity claus */ cpuid = PCPU_GET(cpuid); if (PCPU_GET(apic_id) != lapic_id()) { printf("SMP: cpuid = %d\n", cpuid); printf("SMP: actual apic_id = %d\n", lapic_id()); printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id)); panic("cpuid mismatch! boom!!"); } /* Initialize curthread. */ KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread")); PCPU_SET(curthread, PCPU_GET(idlethread)); mtx_lock_spin(&ap_boot_mtx); mca_init(); /* Init local apic for irq's */ lapic_setup(1); /* Set memory range attributes for this CPU to match the BSP */ mem_range_AP_init(); smp_cpus++; CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid); if (bootverbose) printf("SMP: AP CPU #%d Launched!\n", cpuid); else printf("%s%d%s", smp_cpus == 2 ? "Launching APs: " : "", cpuid, smp_cpus == mp_ncpus ? "\n" : " "); /* Determine if we are a logical CPU. */ if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread) CPU_SET(cpuid, &logical_cpus_mask); if (bootverbose) lapic_dump("AP"); if (smp_cpus == mp_ncpus) { /* enable IPI's, tlb shootdown, freezes etc */ atomic_store_rel_int(&smp_started, 1); } #ifdef __amd64__ /* * Enable global pages TLB extension * This also implicitly flushes the TLB */ load_cr4(rcr4() | CR4_PGE); if (pmap_pcid_enabled) load_cr4(rcr4() | CR4_PCIDE); load_ds(_udatasel); load_es(_udatasel); load_fs(_ufssel); #endif mtx_unlock_spin(&ap_boot_mtx); /* Wait until all the AP's are up. */ while (atomic_load_acq_int(&smp_started) == 0) ia32_pause(); #ifndef EARLY_AP_STARTUP /* Start per-CPU event timers. */ cpu_initclocks_ap(); #endif sched_throw(NULL); panic("scheduler returned us to %s", __func__); /* NOTREACHED */ } static void smp_after_idle_runnable(void *arg __unused) { struct thread *idle_td; int cpu; for (cpu = 1; cpu < mp_ncpus; cpu++) { idle_td = pcpu_find(cpu)->pc_idlethread; while (atomic_load_int(&idle_td->td_lastcpu) == NOCPU && atomic_load_int(&idle_td->td_oncpu) == NOCPU) cpu_spinwait(); kmem_free((vm_offset_t)bootstacks[cpu], kstack_pages * PAGE_SIZE); } } SYSINIT(smp_after_idle_runnable, SI_SUB_SMP, SI_ORDER_ANY, smp_after_idle_runnable, NULL); /* * We tell the I/O APIC code about all the CPUs we want to receive * interrupts. If we don't want certain CPUs to receive IRQs we * can simply not tell the I/O APIC code about them in this function. * We also do not tell it about the BSP since it tells itself about * the BSP internally to work with UP kernels and on UP machines. */ void set_interrupt_apic_ids(void) { u_int i, apic_id; for (i = 0; i < MAXCPU; i++) { apic_id = cpu_apic_ids[i]; if (apic_id == -1) continue; if (cpu_info[apic_id].cpu_bsp) continue; if (cpu_info[apic_id].cpu_disabled) continue; /* Don't let hyperthreads service interrupts. */ if (cpu_info[apic_id].cpu_hyperthread) continue; intr_add_cpu(i); } } #ifdef COUNT_XINVLTLB_HITS u_int xhits_gbl[MAXCPU]; u_int xhits_pg[MAXCPU]; u_int xhits_rng[MAXCPU]; static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, ""); SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl, sizeof(xhits_gbl), "IU", ""); SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg, sizeof(xhits_pg), "IU", ""); SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng, sizeof(xhits_rng), "IU", ""); u_int ipi_global; u_int ipi_page; u_int ipi_range; u_int ipi_range_size; SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, ""); SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, ""); SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, ""); SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size, 0, ""); #endif /* COUNT_XINVLTLB_HITS */ /* * Init and startup IPI. */ void ipi_startup(int apic_id, int vector) { /* * This attempts to follow the algorithm described in the * Intel Multiprocessor Specification v1.4 in section B.4. * For each IPI, we allow the local APIC ~20us to deliver the * IPI. If that times out, we panic. */ /* * first we do an INIT IPI: this INIT IPI might be run, resetting * and running the target CPU. OR this INIT IPI might be latched (P5 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be * ignored. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL | APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id); lapic_ipi_wait(100); /* Explicitly deassert the INIT IPI. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL | APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id); DELAY(10000); /* wait ~10mS */ /* * next we do a STARTUP IPI: the previous INIT IPI might still be * latched, (P5 bug) this 1st STARTUP would then terminate * immediately, and the previously started INIT IPI would continue. OR * the previous INIT IPI has already run. and this STARTUP IPI will * run. OR the previous INIT IPI was ignored. and this STARTUP IPI * will run. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP | vector, apic_id); if (!lapic_ipi_wait(100)) panic("Failed to deliver first STARTUP IPI to APIC %d", apic_id); DELAY(200); /* wait ~200uS */ /* * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is * recognized after hardware RESET or INIT IPI. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP | vector, apic_id); if (!lapic_ipi_wait(100)) panic("Failed to deliver second STARTUP IPI to APIC %d", apic_id); DELAY(200); /* wait ~200uS */ } /* * Send an IPI to specified CPU handling the bitmap logic. */ void ipi_send_cpu(int cpu, u_int ipi) { u_int bitmap, old_pending, new_pending; KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu)); if (IPI_IS_BITMAPED(ipi)) { bitmap = 1 << ipi; ipi = IPI_BITMAP_VECTOR; do { old_pending = cpu_ipi_pending[cpu]; new_pending = old_pending | bitmap; } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu], old_pending, new_pending)); if (old_pending) return; } lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]); } void ipi_bitmap_handler(struct trapframe frame) { struct trapframe *oldframe; struct thread *td; int cpu = PCPU_GET(cpuid); u_int ipi_bitmap; critical_enter(); td = curthread; td->td_intr_nesting_level++; oldframe = td->td_intr_frame; td->td_intr_frame = &frame; ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]); if (ipi_bitmap & (1 << IPI_PREEMPT)) { #ifdef COUNT_IPIS (*ipi_preempt_counts[cpu])++; #endif sched_preempt(td); } if (ipi_bitmap & (1 << IPI_AST)) { #ifdef COUNT_IPIS (*ipi_ast_counts[cpu])++; #endif /* Nothing to do for AST */ } if (ipi_bitmap & (1 << IPI_HARDCLOCK)) { #ifdef COUNT_IPIS (*ipi_hardclock_counts[cpu])++; #endif hardclockintr(); } td->td_intr_frame = oldframe; td->td_intr_nesting_level--; critical_exit(); } /* * send an IPI to a set of cpus. */ void ipi_selected(cpuset_t cpus, u_int ipi) { int cpu; /* * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit * of help in order to understand what is the source. * Set the mask of receiving CPUs for this purpose. */ if (ipi == IPI_STOP_HARD) CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus); while ((cpu = CPU_FFS(&cpus)) != 0) { cpu--; CPU_CLR(cpu, &cpus); CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi); ipi_send_cpu(cpu, ipi); } } /* * send an IPI to a specific CPU. */ void ipi_cpu(int cpu, u_int ipi) { /* * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit * of help in order to understand what is the source. * Set the mask of receiving CPUs for this purpose. */ if (ipi == IPI_STOP_HARD) CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending); CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi); ipi_send_cpu(cpu, ipi); } /* * send an IPI to all CPUs EXCEPT myself */ void ipi_all_but_self(u_int ipi) { cpuset_t other_cpus; other_cpus = all_cpus; CPU_CLR(PCPU_GET(cpuid), &other_cpus); if (IPI_IS_BITMAPED(ipi)) { ipi_selected(other_cpus, ipi); return; } /* * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit * of help in order to understand what is the source. * Set the mask of receiving CPUs for this purpose. */ if (ipi == IPI_STOP_HARD) CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus); CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi); lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS); } int ipi_nmi_handler(void) { u_int cpuid; /* * As long as there is not a simple way to know about a NMI's * source, if the bitmask for the current CPU is present in * the global pending bitword an IPI_STOP_HARD has been issued * and should be handled. */ cpuid = PCPU_GET(cpuid); if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending)) return (1); CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending); cpustop_handler(); return (0); } int nmi_kdb_lock; void nmi_call_kdb_smp(u_int type, struct trapframe *frame) { int cpu; bool call_post; cpu = PCPU_GET(cpuid); if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) { nmi_call_kdb(cpu, type, frame); call_post = false; } else { savectx(&stoppcbs[cpu]); CPU_SET_ATOMIC(cpu, &stopped_cpus); while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) ia32_pause(); call_post = true; } atomic_store_rel_int(&nmi_kdb_lock, 0); if (call_post) cpustop_handler_post(cpu); } /* - * Handle an IPI_STOP by saving our current context and spinning until we - * are resumed. + * Handle an IPI_STOP by saving our current context and spinning (or mwaiting, + * if available) until we are resumed. */ void cpustop_handler(void) { + struct monitorbuf *mb; u_int cpu; + bool use_mwait; cpu = PCPU_GET(cpuid); savectx(&stoppcbs[cpu]); + use_mwait = (stop_mwait && (cpu_feature2 & CPUID2_MON) != 0 && + !mwait_cpustop_broken); + if (use_mwait) { + mb = PCPU_PTR(monitorbuf); + atomic_store_int(&mb->stop_state, + MONITOR_STOPSTATE_STOPPED); + } + /* Indicate that we are stopped */ CPU_SET_ATOMIC(cpu, &stopped_cpus); /* Wait for restart */ while (!CPU_ISSET(cpu, &started_cpus)) { + if (use_mwait) { + cpu_monitor(mb, 0, 0); + if (atomic_load_int(&mb->stop_state) == + MONITOR_STOPSTATE_STOPPED) + cpu_mwait(0, MWAIT_C1); + continue; + } + ia32_pause(); /* * Halt non-BSP CPUs on panic -- we're never going to need them * again, and might as well save power / release resources * (e.g., overprovisioned VM infrastructure). */ while (__predict_false(!IS_BSP() && panicstr != NULL)) halt(); } cpustop_handler_post(cpu); } static void cpustop_handler_post(u_int cpu) { CPU_CLR_ATOMIC(cpu, &started_cpus); CPU_CLR_ATOMIC(cpu, &stopped_cpus); /* * We don't broadcast TLB invalidations to other CPUs when they are * stopped. Hence, we clear the TLB before resuming. */ invltlb_glob(); #if defined(__amd64__) && defined(DDB) amd64_db_resume_dbreg(); #endif if (cpu == 0 && cpustop_restartfunc != NULL) { cpustop_restartfunc(); cpustop_restartfunc = NULL; } } /* * Handle an IPI_SUSPEND by saving our current context and spinning until we * are resumed. */ void cpususpend_handler(void) { u_int cpu; mtx_assert(&smp_ipi_mtx, MA_NOTOWNED); cpu = PCPU_GET(cpuid); if (savectx(&susppcbs[cpu]->sp_pcb)) { #ifdef __amd64__ fpususpend(susppcbs[cpu]->sp_fpususpend); #else npxsuspend(susppcbs[cpu]->sp_fpususpend); #endif /* * suspended_cpus is cleared shortly after each AP is restarted * by a Startup IPI, so that the BSP can proceed to restarting * the next AP. * * resuming_cpus gets cleared when the AP completes * initialization after having been released by the BSP. * resuming_cpus is probably not the best name for the * variable, because it is actually a set of processors that * haven't resumed yet and haven't necessarily started resuming. * * Note that suspended_cpus is meaningful only for ACPI suspend * as it's not really used for Xen suspend since the APs are * automatically restored to the running state and the correct * context. For the same reason resumectx is never called in * that case. */ CPU_SET_ATOMIC(cpu, &suspended_cpus); CPU_SET_ATOMIC(cpu, &resuming_cpus); /* * Invalidate the cache after setting the global status bits. * The last AP to set its bit may end up being an Owner of the * corresponding cache line in MOESI protocol. The AP may be * stopped before the cache line is written to the main memory. */ wbinvd(); } else { #ifdef __amd64__ fpuresume(susppcbs[cpu]->sp_fpususpend); #else npxresume(susppcbs[cpu]->sp_fpususpend); #endif pmap_init_pat(); initializecpu(); PCPU_SET(switchtime, 0); PCPU_SET(switchticks, ticks); /* Indicate that we have restarted and restored the context. */ CPU_CLR_ATOMIC(cpu, &suspended_cpus); } /* Wait for resume directive */ while (!CPU_ISSET(cpu, &toresume_cpus)) ia32_pause(); /* Re-apply microcode updates. */ ucode_reload(); #ifdef __i386__ /* Finish removing the identity mapping of low memory for this AP. */ invltlb_glob(); #endif if (cpu_ops.cpu_resume) cpu_ops.cpu_resume(); #ifdef __amd64__ if (vmm_resume_p) vmm_resume_p(); #endif /* Resume MCA and local APIC */ lapic_xapic_mode(); mca_resume(); lapic_setup(0); /* Indicate that we are resumed */ CPU_CLR_ATOMIC(cpu, &resuming_cpus); CPU_CLR_ATOMIC(cpu, &suspended_cpus); CPU_CLR_ATOMIC(cpu, &toresume_cpus); } void invlcache_handler(void) { uint32_t generation; #ifdef COUNT_IPIS (*ipi_invlcache_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ /* * Reading the generation here allows greater parallelism * since wbinvd is a serializing instruction. Without the * temporary, we'd wait for wbinvd to complete, then the read * would execute, then the dependent write, which must then * complete before return from interrupt. */ generation = smp_tlb_generation; wbinvd(); PCPU_SET(smp_tlb_done, generation); } /* * This is called once the rest of the system is up and running and we're * ready to let the AP's out of the pen. */ static void release_aps(void *dummy __unused) { if (mp_ncpus == 1) return; atomic_store_rel_int(&aps_ready, 1); while (smp_started == 0) ia32_pause(); } SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL); #ifdef COUNT_IPIS /* * Setup interrupt counters for IPI handlers. */ static void mp_ipi_intrcnt(void *dummy) { char buf[64]; int i; CPU_FOREACH(i) { snprintf(buf, sizeof(buf), "cpu%d:invltlb", i); intrcnt_add(buf, &ipi_invltlb_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:invlrng", i); intrcnt_add(buf, &ipi_invlrng_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:invlpg", i); intrcnt_add(buf, &ipi_invlpg_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:invlcache", i); intrcnt_add(buf, &ipi_invlcache_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:preempt", i); intrcnt_add(buf, &ipi_preempt_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:ast", i); intrcnt_add(buf, &ipi_ast_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i); intrcnt_add(buf, &ipi_rendezvous_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:hardclock", i); intrcnt_add(buf, &ipi_hardclock_counts[i]); } } SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL); #endif /* * Flush the TLB on other CPU's */ /* Variables needed for SMP tlb shootdown. */ vm_offset_t smp_tlb_addr1, smp_tlb_addr2; pmap_t smp_tlb_pmap; volatile uint32_t smp_tlb_generation; #ifdef __amd64__ #define read_eflags() read_rflags() #endif static void smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap, vm_offset_t addr1, vm_offset_t addr2) { cpuset_t other_cpus; volatile uint32_t *p_cpudone; uint32_t generation; int cpu; /* It is not necessary to signal other CPUs while in the debugger. */ if (kdb_active || panicstr != NULL) return; /* * Check for other cpus. Return if none. */ if (CPU_ISFULLSET(&mask)) { if (mp_ncpus <= 1) return; } else { CPU_CLR(PCPU_GET(cpuid), &mask); if (CPU_EMPTY(&mask)) return; } if (!(read_eflags() & PSL_I)) panic("%s: interrupts disabled", __func__); mtx_lock_spin(&smp_ipi_mtx); smp_tlb_addr1 = addr1; smp_tlb_addr2 = addr2; smp_tlb_pmap = pmap; generation = ++smp_tlb_generation; if (CPU_ISFULLSET(&mask)) { ipi_all_but_self(vector); other_cpus = all_cpus; CPU_CLR(PCPU_GET(cpuid), &other_cpus); } else { other_cpus = mask; while ((cpu = CPU_FFS(&mask)) != 0) { cpu--; CPU_CLR(cpu, &mask); CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, vector); ipi_send_cpu(cpu, vector); } } while ((cpu = CPU_FFS(&other_cpus)) != 0) { cpu--; CPU_CLR(cpu, &other_cpus); p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done; while (*p_cpudone != generation) ia32_pause(); } mtx_unlock_spin(&smp_ipi_mtx); } void smp_masked_invltlb(cpuset_t mask, pmap_t pmap) { if (smp_started) { smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0); #ifdef COUNT_XINVLTLB_HITS ipi_global++; #endif } } void smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap) { if (smp_started) { smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0); #ifdef COUNT_XINVLTLB_HITS ipi_page++; #endif } } void smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2, pmap_t pmap) { if (smp_started) { smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap, addr1, addr2); #ifdef COUNT_XINVLTLB_HITS ipi_range++; ipi_range_size += (addr2 - addr1) / PAGE_SIZE; #endif } } void smp_cache_flush(void) { if (smp_started) { smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL, 0, 0); } } /* * Handlers for TLB related IPIs */ void invltlb_handler(void) { uint32_t generation; #ifdef COUNT_XINVLTLB_HITS xhits_gbl[PCPU_GET(cpuid)]++; #endif /* COUNT_XINVLTLB_HITS */ #ifdef COUNT_IPIS (*ipi_invltlb_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ /* * Reading the generation here allows greater parallelism * since invalidating the TLB is a serializing operation. */ generation = smp_tlb_generation; if (smp_tlb_pmap == kernel_pmap) invltlb_glob(); #ifdef __amd64__ else invltlb(); #endif PCPU_SET(smp_tlb_done, generation); } void invlpg_handler(void) { uint32_t generation; #ifdef COUNT_XINVLTLB_HITS xhits_pg[PCPU_GET(cpuid)]++; #endif /* COUNT_XINVLTLB_HITS */ #ifdef COUNT_IPIS (*ipi_invlpg_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ generation = smp_tlb_generation; /* Overlap with serialization */ #ifdef __i386__ if (smp_tlb_pmap == kernel_pmap) #endif invlpg(smp_tlb_addr1); PCPU_SET(smp_tlb_done, generation); } void invlrng_handler(void) { vm_offset_t addr, addr2; uint32_t generation; #ifdef COUNT_XINVLTLB_HITS xhits_rng[PCPU_GET(cpuid)]++; #endif /* COUNT_XINVLTLB_HITS */ #ifdef COUNT_IPIS (*ipi_invlrng_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ addr = smp_tlb_addr1; addr2 = smp_tlb_addr2; generation = smp_tlb_generation; /* Overlap with serialization */ #ifdef __i386__ if (smp_tlb_pmap == kernel_pmap) #endif do { invlpg(addr); addr += PAGE_SIZE; } while (addr < addr2); PCPU_SET(smp_tlb_done, generation); }