Index: head/sys/amd64/include/counter.h =================================================================== --- head/sys/amd64/include/counter.h (revision 302371) +++ head/sys/amd64/include/counter.h (revision 302372) @@ -1,89 +1,89 @@ /*- * Copyright (c) 2012 Konstantin Belousov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR 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_COUNTER_H__ #define __MACHINE_COUNTER_H__ #include extern struct pcpu __pcpu[1]; #define counter_enter() do {} while (0) #define counter_exit() do {} while (0) #ifdef IN_SUBR_COUNTER_C static inline uint64_t counter_u64_read_one(uint64_t *p, int cpu) { return (*(uint64_t *)((char *)p + sizeof(struct pcpu) * cpu)); } static inline uint64_t counter_u64_fetch_inline(uint64_t *p) { uint64_t r; int i; r = 0; - for (i = 0; i < mp_ncpus; i++) + CPU_FOREACH(i) r += counter_u64_read_one((uint64_t *)p, i); return (r); } static void counter_u64_zero_one_cpu(void *arg) { *((uint64_t *)((char *)arg + sizeof(struct pcpu) * PCPU_GET(cpuid))) = 0; } static inline void counter_u64_zero_inline(counter_u64_t c) { smp_rendezvous(smp_no_rendevous_barrier, counter_u64_zero_one_cpu, smp_no_rendevous_barrier, c); } #endif #define counter_u64_add_protected(c, i) counter_u64_add(c, i) static inline void counter_u64_add(counter_u64_t c, int64_t inc) { __asm __volatile("addq\t%1,%%gs:(%0)" : : "r" ((char *)c - (char *)&__pcpu[0]), "ri" (inc) : "memory", "cc"); } #endif /* ! __MACHINE_COUNTER_H__ */ Index: head/sys/cddl/compat/opensolaris/sys/proc.h =================================================================== --- head/sys/cddl/compat/opensolaris/sys/proc.h (revision 302371) +++ head/sys/cddl/compat/opensolaris/sys/proc.h (revision 302372) @@ -1,97 +1,97 @@ /*- * Copyright (c) 2007 Pawel Jakub Dawidek * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _OPENSOLARIS_SYS_PROC_H_ #define _OPENSOLARIS_SYS_PROC_H_ #include #include #include_next #include #include #include #include #include #include #include #ifdef _KERNEL #define CPU curcpu #define minclsyspri PRIBIO #define maxclsyspri PVM -#define max_ncpus mp_ncpus -#define boot_max_ncpus mp_ncpus +#define max_ncpus (mp_maxid + 1) +#define boot_max_ncpus (mp_maxid + 1) #define TS_RUN 0 #define p0 proc0 #define t_tid td_tid typedef short pri_t; typedef struct thread _kthread; typedef struct thread kthread_t; typedef struct thread *kthread_id_t; typedef struct proc proc_t; extern struct proc *zfsproc; static __inline kthread_t * do_thread_create(caddr_t stk, size_t stksize, void (*proc)(void *), void *arg, size_t len, proc_t *pp, int state, pri_t pri) { kthread_t *td = NULL; int error; /* * Be sure there are no surprises. */ ASSERT(stk == NULL); ASSERT(len == 0); ASSERT(state == TS_RUN); ASSERT(pp == &p0); error = kproc_kthread_add(proc, arg, &zfsproc, &td, RFSTOPPED, stksize / PAGE_SIZE, "zfskern", "solthread %p", proc); if (error == 0) { thread_lock(td); sched_prio(td, pri); sched_add(td, SRQ_BORING); thread_unlock(td); } return (td); } #define thread_create(stk, stksize, proc, arg, len, pp, state, pri) \ do_thread_create(stk, stksize, proc, arg, len, pp, state, pri) #define thread_exit() kthread_exit() #endif /* _KERNEL */ #endif /* _OPENSOLARIS_SYS_PROC_H_ */ Index: head/sys/dev/cpuctl/cpuctl.c =================================================================== --- head/sys/dev/cpuctl/cpuctl.c (revision 302371) +++ head/sys/dev/cpuctl/cpuctl.c (revision 302372) @@ -1,557 +1,557 @@ /*- * Copyright (c) 2006-2008 Stanislav Sedov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static d_open_t cpuctl_open; static d_ioctl_t cpuctl_ioctl; #define CPUCTL_VERSION 1 #ifdef DEBUG # define DPRINTF(format,...) printf(format, __VA_ARGS__); #else # define DPRINTF(...) #endif #define UCODE_SIZE_MAX (32 * 1024) static int cpuctl_do_msr(int cpu, cpuctl_msr_args_t *data, u_long cmd, struct thread *td); static int cpuctl_do_cpuid(int cpu, cpuctl_cpuid_args_t *data, struct thread *td); static int cpuctl_do_cpuid_count(int cpu, cpuctl_cpuid_count_args_t *data, struct thread *td); static int cpuctl_do_update(int cpu, cpuctl_update_args_t *data, struct thread *td); static int update_intel(int cpu, cpuctl_update_args_t *args, struct thread *td); static int update_amd(int cpu, cpuctl_update_args_t *args, struct thread *td); static int update_via(int cpu, cpuctl_update_args_t *args, struct thread *td); static struct cdev **cpuctl_devs; static MALLOC_DEFINE(M_CPUCTL, "cpuctl", "CPUCTL buffer"); static struct cdevsw cpuctl_cdevsw = { .d_version = D_VERSION, .d_open = cpuctl_open, .d_ioctl = cpuctl_ioctl, .d_name = "cpuctl", }; /* * This function checks if specified cpu enabled or not. */ static int cpu_enabled(int cpu) { return (pmc_cpu_is_disabled(cpu) == 0); } /* * Check if the current thread is bound to a specific cpu. */ static int cpu_sched_is_bound(struct thread *td) { int ret; thread_lock(td); ret = sched_is_bound(td); thread_unlock(td); return (ret); } /* * Switch to target cpu to run. */ static void set_cpu(int cpu, struct thread *td) { - KASSERT(cpu >= 0 && cpu < mp_ncpus && cpu_enabled(cpu), + KASSERT(cpu >= 0 && cpu <= mp_maxid && cpu_enabled(cpu), ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); thread_lock(td); sched_bind(td, cpu); thread_unlock(td); KASSERT(td->td_oncpu == cpu, ("[cpuctl,%d]: cannot bind to target cpu %d on cpu %d", __LINE__, cpu, td->td_oncpu)); } static void restore_cpu(int oldcpu, int is_bound, struct thread *td) { - KASSERT(oldcpu >= 0 && oldcpu < mp_ncpus && cpu_enabled(oldcpu), + KASSERT(oldcpu >= 0 && oldcpu <= mp_maxid && cpu_enabled(oldcpu), ("[cpuctl,%d]: bad cpu number %d", __LINE__, oldcpu)); thread_lock(td); if (is_bound == 0) sched_unbind(td); else sched_bind(td, oldcpu); thread_unlock(td); } int cpuctl_ioctl(struct cdev *dev, u_long cmd, caddr_t data, int flags, struct thread *td) { int ret; int cpu = dev2unit(dev); - if (cpu >= mp_ncpus || !cpu_enabled(cpu)) { + if (cpu > mp_maxid || !cpu_enabled(cpu)) { DPRINTF("[cpuctl,%d]: bad cpu number %d\n", __LINE__, cpu); return (ENXIO); } /* Require write flag for "write" requests. */ if ((cmd == CPUCTL_WRMSR || cmd == CPUCTL_UPDATE) && ((flags & FWRITE) == 0)) return (EPERM); switch (cmd) { case CPUCTL_RDMSR: ret = cpuctl_do_msr(cpu, (cpuctl_msr_args_t *)data, cmd, td); break; case CPUCTL_MSRSBIT: case CPUCTL_MSRCBIT: case CPUCTL_WRMSR: ret = priv_check(td, PRIV_CPUCTL_WRMSR); if (ret != 0) goto fail; ret = cpuctl_do_msr(cpu, (cpuctl_msr_args_t *)data, cmd, td); break; case CPUCTL_CPUID: ret = cpuctl_do_cpuid(cpu, (cpuctl_cpuid_args_t *)data, td); break; case CPUCTL_UPDATE: ret = priv_check(td, PRIV_CPUCTL_UPDATE); if (ret != 0) goto fail; ret = cpuctl_do_update(cpu, (cpuctl_update_args_t *)data, td); break; case CPUCTL_CPUID_COUNT: ret = cpuctl_do_cpuid_count(cpu, (cpuctl_cpuid_count_args_t *)data, td); break; default: ret = EINVAL; break; } fail: return (ret); } /* * Actually perform cpuid operation. */ static int cpuctl_do_cpuid_count(int cpu, cpuctl_cpuid_count_args_t *data, struct thread *td) { int is_bound = 0; int oldcpu; - KASSERT(cpu >= 0 && cpu < mp_ncpus, + KASSERT(cpu >= 0 && cpu <= mp_maxid, ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); /* Explicitly clear cpuid data to avoid returning stale info. */ bzero(data->data, sizeof(data->data)); DPRINTF("[cpuctl,%d]: retrieving cpuid lev %#0x type %#0x for %d cpu\n", __LINE__, data->level, data->level_type, cpu); #ifdef __i386__ if (cpu_id == 0) return (ENODEV); #endif oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); cpuid_count(data->level, data->level_type, data->data); restore_cpu(oldcpu, is_bound, td); return (0); } static int cpuctl_do_cpuid(int cpu, cpuctl_cpuid_args_t *data, struct thread *td) { cpuctl_cpuid_count_args_t cdata; int error; cdata.level = data->level; /* Override the level type. */ cdata.level_type = 0; error = cpuctl_do_cpuid_count(cpu, &cdata, td); bcopy(cdata.data, data->data, sizeof(data->data)); /* Ignore error */ return (error); } /* * Actually perform MSR operations. */ static int cpuctl_do_msr(int cpu, cpuctl_msr_args_t *data, u_long cmd, struct thread *td) { uint64_t reg; int is_bound = 0; int oldcpu; int ret; - KASSERT(cpu >= 0 && cpu < mp_ncpus, + KASSERT(cpu >= 0 && cpu <= mp_maxid, ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); /* * Explicitly clear cpuid data to avoid returning stale * info */ DPRINTF("[cpuctl,%d]: operating on MSR %#0x for %d cpu\n", __LINE__, data->msr, cpu); #ifdef __i386__ if ((cpu_feature & CPUID_MSR) == 0) return (ENODEV); #endif oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); if (cmd == CPUCTL_RDMSR) { data->data = 0; ret = rdmsr_safe(data->msr, &data->data); } else if (cmd == CPUCTL_WRMSR) { ret = wrmsr_safe(data->msr, data->data); } else if (cmd == CPUCTL_MSRSBIT) { critical_enter(); ret = rdmsr_safe(data->msr, ®); if (ret == 0) ret = wrmsr_safe(data->msr, reg | data->data); critical_exit(); } else if (cmd == CPUCTL_MSRCBIT) { critical_enter(); ret = rdmsr_safe(data->msr, ®); if (ret == 0) ret = wrmsr_safe(data->msr, reg & ~data->data); critical_exit(); } else panic("[cpuctl,%d]: unknown operation requested: %lu", __LINE__, cmd); restore_cpu(oldcpu, is_bound, td); return (ret); } /* * Actually perform microcode update. */ static int cpuctl_do_update(int cpu, cpuctl_update_args_t *data, struct thread *td) { cpuctl_cpuid_args_t args = { .level = 0, }; char vendor[13]; int ret; - KASSERT(cpu >= 0 && cpu < mp_ncpus, + KASSERT(cpu >= 0 && cpu <= mp_maxid, ("[cpuctl,%d]: bad cpu number %d", __LINE__, cpu)); DPRINTF("[cpuctl,%d]: XXX %d", __LINE__, cpu); ret = cpuctl_do_cpuid(cpu, &args, td); if (ret != 0) return (ret); ((uint32_t *)vendor)[0] = args.data[1]; ((uint32_t *)vendor)[1] = args.data[3]; ((uint32_t *)vendor)[2] = args.data[2]; vendor[12] = '\0'; if (strncmp(vendor, INTEL_VENDOR_ID, sizeof(INTEL_VENDOR_ID)) == 0) ret = update_intel(cpu, data, td); else if(strncmp(vendor, AMD_VENDOR_ID, sizeof(AMD_VENDOR_ID)) == 0) ret = update_amd(cpu, data, td); else if(strncmp(vendor, CENTAUR_VENDOR_ID, sizeof(CENTAUR_VENDOR_ID)) == 0) ret = update_via(cpu, data, td); else ret = ENXIO; return (ret); } static int update_intel(int cpu, cpuctl_update_args_t *args, struct thread *td) { void *ptr; uint64_t rev0, rev1; uint32_t tmp[4]; int is_bound; int oldcpu; int ret; if (args->size == 0 || args->data == NULL) { DPRINTF("[cpuctl,%d]: zero-sized firmware image", __LINE__); return (EINVAL); } if (args->size > UCODE_SIZE_MAX) { DPRINTF("[cpuctl,%d]: firmware image too large", __LINE__); return (EINVAL); } /* * 16 byte alignment required. Rely on the fact that * malloc(9) always returns the pointer aligned at least on * the size of the allocation. */ ptr = malloc(args->size + 16, M_CPUCTL, M_WAITOK); if (copyin(args->data, ptr, args->size) != 0) { DPRINTF("[cpuctl,%d]: copyin %p->%p of %zd bytes failed", __LINE__, args->data, ptr, args->size); ret = EFAULT; goto fail; } oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); critical_enter(); rdmsr_safe(MSR_BIOS_SIGN, &rev0); /* Get current microcode revision. */ /* * Perform update. */ wrmsr_safe(MSR_BIOS_UPDT_TRIG, (uintptr_t)(ptr)); wrmsr_safe(MSR_BIOS_SIGN, 0); /* * Serialize instruction flow. */ do_cpuid(0, tmp); critical_exit(); rdmsr_safe(MSR_BIOS_SIGN, &rev1); /* Get new microcode revision. */ restore_cpu(oldcpu, is_bound, td); if (rev1 > rev0) ret = 0; else ret = EEXIST; fail: free(ptr, M_CPUCTL); return (ret); } static int update_amd(int cpu, cpuctl_update_args_t *args, struct thread *td) { void *ptr = NULL; uint32_t tmp[4]; int is_bound = 0; int oldcpu; int ret; if (args->size == 0 || args->data == NULL) { DPRINTF("[cpuctl,%d]: zero-sized firmware image", __LINE__); return (EINVAL); } if (args->size > UCODE_SIZE_MAX) { DPRINTF("[cpuctl,%d]: firmware image too large", __LINE__); return (EINVAL); } /* * XXX Might not require contignous address space - needs check */ ptr = contigmalloc(args->size, M_CPUCTL, 0, 0, 0xffffffff, 16, 0); if (ptr == NULL) { DPRINTF("[cpuctl,%d]: cannot allocate %zd bytes of memory", __LINE__, args->size); return (ENOMEM); } if (copyin(args->data, ptr, args->size) != 0) { DPRINTF("[cpuctl,%d]: copyin %p->%p of %zd bytes failed", __LINE__, args->data, ptr, args->size); ret = EFAULT; goto fail; } oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); critical_enter(); /* * Perform update. */ wrmsr_safe(MSR_K8_UCODE_UPDATE, (uintptr_t)ptr); /* * Serialize instruction flow. */ do_cpuid(0, tmp); critical_exit(); restore_cpu(oldcpu, is_bound, td); ret = 0; fail: if (ptr != NULL) contigfree(ptr, args->size, M_CPUCTL); return (ret); } static int update_via(int cpu, cpuctl_update_args_t *args, struct thread *td) { void *ptr; uint64_t rev0, rev1, res; uint32_t tmp[4]; int is_bound; int oldcpu; int ret; if (args->size == 0 || args->data == NULL) { DPRINTF("[cpuctl,%d]: zero-sized firmware image", __LINE__); return (EINVAL); } if (args->size > UCODE_SIZE_MAX) { DPRINTF("[cpuctl,%d]: firmware image too large", __LINE__); return (EINVAL); } /* * 4 byte alignment required. */ ptr = malloc(args->size, M_CPUCTL, M_WAITOK); if (copyin(args->data, ptr, args->size) != 0) { DPRINTF("[cpuctl,%d]: copyin %p->%p of %zd bytes failed", __LINE__, args->data, ptr, args->size); ret = EFAULT; goto fail; } oldcpu = td->td_oncpu; is_bound = cpu_sched_is_bound(td); set_cpu(cpu, td); critical_enter(); rdmsr_safe(MSR_BIOS_SIGN, &rev0); /* Get current microcode revision. */ /* * Perform update. */ wrmsr_safe(MSR_BIOS_UPDT_TRIG, (uintptr_t)(ptr)); do_cpuid(1, tmp); /* * Result are in low byte of MSR FCR5: * 0x00: No update has been attempted since RESET. * 0x01: The last attempted update was successful. * 0x02: The last attempted update was unsuccessful due to a bad * environment. No update was loaded and any preexisting * patches are still active. * 0x03: The last attempted update was not applicable to this processor. * No update was loaded and any preexisting patches are still * active. * 0x04: The last attempted update was not successful due to an invalid * update data block. No update was loaded and any preexisting * patches are still active */ rdmsr_safe(0x1205, &res); res &= 0xff; critical_exit(); rdmsr_safe(MSR_BIOS_SIGN, &rev1); /* Get new microcode revision. */ restore_cpu(oldcpu, is_bound, td); DPRINTF("[cpu,%d]: rev0=%x rev1=%x res=%x\n", __LINE__, (unsigned)(rev0 >> 32), (unsigned)(rev1 >> 32), (unsigned)res); if (res != 0x01) ret = EINVAL; else ret = 0; fail: free(ptr, M_CPUCTL); return (ret); } int cpuctl_open(struct cdev *dev, int flags, int fmt __unused, struct thread *td) { int ret = 0; int cpu; cpu = dev2unit(dev); - if (cpu >= mp_ncpus || !cpu_enabled(cpu)) { + if (cpu > mp_maxid || !cpu_enabled(cpu)) { DPRINTF("[cpuctl,%d]: incorrect cpu number %d\n", __LINE__, cpu); return (ENXIO); } if (flags & FWRITE) ret = securelevel_gt(td->td_ucred, 0); return (ret); } static int cpuctl_modevent(module_t mod __unused, int type, void *data __unused) { int cpu; switch(type) { case MOD_LOAD: if (bootverbose) printf("cpuctl: access to MSR registers/cpuid info.\n"); - cpuctl_devs = malloc(sizeof(*cpuctl_devs) * mp_ncpus, M_CPUCTL, + cpuctl_devs = malloc(sizeof(*cpuctl_devs) * (mp_maxid + 1), M_CPUCTL, M_WAITOK | M_ZERO); - for (cpu = 0; cpu < mp_ncpus; cpu++) + CPU_FOREACH(cpu) if (cpu_enabled(cpu)) cpuctl_devs[cpu] = make_dev(&cpuctl_cdevsw, cpu, UID_ROOT, GID_KMEM, 0640, "cpuctl%d", cpu); break; case MOD_UNLOAD: - for (cpu = 0; cpu < mp_ncpus; cpu++) { + CPU_FOREACH(cpu) { if (cpuctl_devs[cpu] != NULL) destroy_dev(cpuctl_devs[cpu]); } free(cpuctl_devs, M_CPUCTL); break; case MOD_SHUTDOWN: break; default: return (EOPNOTSUPP); } return (0); } DEV_MODULE(cpuctl, cpuctl_modevent, NULL); MODULE_VERSION(cpuctl, CPUCTL_VERSION); Index: head/sys/i386/include/counter.h =================================================================== --- head/sys/i386/include/counter.h (revision 302371) +++ head/sys/i386/include/counter.h (revision 302372) @@ -1,178 +1,178 @@ /*- * Copyright (c) 2012 Konstantin Belousov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR 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_COUNTER_H__ #define __MACHINE_COUNTER_H__ #include #ifdef INVARIANTS #include #endif #include #include #define counter_enter() do { \ if ((cpu_feature & CPUID_CX8) == 0) \ critical_enter(); \ } while (0) #define counter_exit() do { \ if ((cpu_feature & CPUID_CX8) == 0) \ critical_exit(); \ } while (0) extern struct pcpu __pcpu[MAXCPU]; static inline void counter_64_inc_8b(uint64_t *p, int64_t inc) { __asm __volatile( "movl %%fs:(%%esi),%%eax\n\t" "movl %%fs:4(%%esi),%%edx\n" "1:\n\t" "movl %%eax,%%ebx\n\t" "movl %%edx,%%ecx\n\t" "addl (%%edi),%%ebx\n\t" "adcl 4(%%edi),%%ecx\n\t" "cmpxchg8b %%fs:(%%esi)\n\t" "jnz 1b" : : "S" ((char *)p - (char *)&__pcpu[0]), "D" (&inc) : "memory", "cc", "eax", "edx", "ebx", "ecx"); } #ifdef IN_SUBR_COUNTER_C static inline uint64_t counter_u64_read_one_8b(uint64_t *p) { uint32_t res_lo, res_high; __asm __volatile( "movl %%eax,%%ebx\n\t" "movl %%edx,%%ecx\n\t" "cmpxchg8b (%2)" : "=a" (res_lo), "=d"(res_high) : "SD" (p) : "cc", "ebx", "ecx"); return (res_lo + ((uint64_t)res_high << 32)); } static inline uint64_t counter_u64_fetch_inline(uint64_t *p) { uint64_t res; int i; res = 0; if ((cpu_feature & CPUID_CX8) == 0) { /* * The machines without cmpxchg8b are not SMP. * Disabling the preemption provides atomicity of the * counter reading, since update is done in the * critical section as well. */ critical_enter(); - for (i = 0; i < mp_ncpus; i++) { + CPU_FOREACH(i) { res += *(uint64_t *)((char *)p + sizeof(struct pcpu) * i); } critical_exit(); } else { - for (i = 0; i < mp_ncpus; i++) + CPU_FOREACH(i) res += counter_u64_read_one_8b((uint64_t *)((char *)p + sizeof(struct pcpu) * i)); } return (res); } static inline void counter_u64_zero_one_8b(uint64_t *p) { __asm __volatile( "movl (%0),%%eax\n\t" "movl 4(%0),%%edx\n" "xorl %%ebx,%%ebx\n\t" "xorl %%ecx,%%ecx\n\t" "1:\n\t" "cmpxchg8b (%0)\n\t" "jnz 1b" : : "SD" (p) : "memory", "cc", "eax", "edx", "ebx", "ecx"); } static void counter_u64_zero_one_cpu(void *arg) { uint64_t *p; p = (uint64_t *)((char *)arg + sizeof(struct pcpu) * PCPU_GET(cpuid)); counter_u64_zero_one_8b(p); } static inline void counter_u64_zero_inline(counter_u64_t c) { int i; if ((cpu_feature & CPUID_CX8) == 0) { critical_enter(); - for (i = 0; i < mp_ncpus; i++) + CPU_FOREACH(i) *(uint64_t *)((char *)c + sizeof(struct pcpu) * i) = 0; critical_exit(); } else { smp_rendezvous(smp_no_rendevous_barrier, counter_u64_zero_one_cpu, smp_no_rendevous_barrier, c); } } #endif #define counter_u64_add_protected(c, inc) do { \ if ((cpu_feature & CPUID_CX8) == 0) { \ CRITICAL_ASSERT(curthread); \ *(uint64_t *)zpcpu_get(c) += (inc); \ } else \ counter_64_inc_8b((c), (inc)); \ } while (0) static inline void counter_u64_add(counter_u64_t c, int64_t inc) { if ((cpu_feature & CPUID_CX8) == 0) { critical_enter(); *(uint64_t *)zpcpu_get(c) += inc; critical_exit(); } else { counter_64_inc_8b(c, inc); } } #endif /* ! __MACHINE_COUNTER_H__ */ Index: head/sys/kern/subr_pcpu.c =================================================================== --- head/sys/kern/subr_pcpu.c (revision 302371) +++ head/sys/kern/subr_pcpu.c (revision 302372) @@ -1,419 +1,419 @@ /*- * Copyright (c) 2001 Wind River Systems, Inc. * All rights reserved. * Written by: John Baldwin * * Copyright (c) 2009 Jeffrey Roberson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 4. Neither the name of the author nor the names of any co-contributors * may 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. */ /* * This module provides MI support for per-cpu data. * * Each architecture determines the mapping of logical CPU IDs to physical * CPUs. The requirements of this mapping are as follows: * - Logical CPU IDs must reside in the range 0 ... MAXCPU - 1. * - The mapping is not required to be dense. That is, there may be * gaps in the mappings. * - The platform sets the value of MAXCPU in . * - It is suggested, but not required, that in the non-SMP case, the * platform define MAXCPU to be 1 and define the logical ID of the * sole CPU as 0. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_PCPU, "Per-cpu", "Per-cpu resource accouting."); struct dpcpu_free { uintptr_t df_start; int df_len; TAILQ_ENTRY(dpcpu_free) df_link; }; static DPCPU_DEFINE(char, modspace[DPCPU_MODMIN]); static TAILQ_HEAD(, dpcpu_free) dpcpu_head = TAILQ_HEAD_INITIALIZER(dpcpu_head); static struct sx dpcpu_lock; uintptr_t dpcpu_off[MAXCPU]; struct pcpu *cpuid_to_pcpu[MAXCPU]; struct cpuhead cpuhead = STAILQ_HEAD_INITIALIZER(cpuhead); /* * Initialize the MI portions of a struct pcpu. */ void pcpu_init(struct pcpu *pcpu, int cpuid, size_t size) { bzero(pcpu, size); KASSERT(cpuid >= 0 && cpuid < MAXCPU, ("pcpu_init: invalid cpuid %d", cpuid)); pcpu->pc_cpuid = cpuid; cpuid_to_pcpu[cpuid] = pcpu; STAILQ_INSERT_TAIL(&cpuhead, pcpu, pc_allcpu); cpu_pcpu_init(pcpu, cpuid, size); pcpu->pc_rm_queue.rmq_next = &pcpu->pc_rm_queue; pcpu->pc_rm_queue.rmq_prev = &pcpu->pc_rm_queue; } void dpcpu_init(void *dpcpu, int cpuid) { struct pcpu *pcpu; pcpu = pcpu_find(cpuid); pcpu->pc_dynamic = (uintptr_t)dpcpu - DPCPU_START; /* * Initialize defaults from our linker section. */ memcpy(dpcpu, (void *)DPCPU_START, DPCPU_BYTES); /* * Place it in the global pcpu offset array. */ dpcpu_off[cpuid] = pcpu->pc_dynamic; } static void dpcpu_startup(void *dummy __unused) { struct dpcpu_free *df; df = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO); df->df_start = (uintptr_t)&DPCPU_NAME(modspace); df->df_len = DPCPU_MODMIN; TAILQ_INSERT_HEAD(&dpcpu_head, df, df_link); sx_init(&dpcpu_lock, "dpcpu alloc lock"); } SYSINIT(dpcpu, SI_SUB_KLD, SI_ORDER_FIRST, dpcpu_startup, 0); /* * UMA_PCPU_ZONE zones, that are available for all kernel * consumers. Right now 64 bit zone is used for counter(9) * and pointer zone is used by flowtable. */ uma_zone_t pcpu_zone_64; uma_zone_t pcpu_zone_ptr; static void pcpu_zones_startup(void) { pcpu_zone_64 = uma_zcreate("64 pcpu", sizeof(uint64_t), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU); if (sizeof(uint64_t) == sizeof(void *)) pcpu_zone_ptr = pcpu_zone_64; else pcpu_zone_ptr = uma_zcreate("ptr pcpu", sizeof(void *), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_PCPU); } SYSINIT(pcpu_zones, SI_SUB_KMEM, SI_ORDER_ANY, pcpu_zones_startup, NULL); /* * First-fit extent based allocator for allocating space in the per-cpu * region reserved for modules. This is only intended for use by the * kernel linkers to place module linker sets. */ void * dpcpu_alloc(int size) { struct dpcpu_free *df; void *s; s = NULL; size = roundup2(size, sizeof(void *)); sx_xlock(&dpcpu_lock); TAILQ_FOREACH(df, &dpcpu_head, df_link) { if (df->df_len < size) continue; if (df->df_len == size) { s = (void *)df->df_start; TAILQ_REMOVE(&dpcpu_head, df, df_link); free(df, M_PCPU); break; } s = (void *)df->df_start; df->df_len -= size; df->df_start = df->df_start + size; break; } sx_xunlock(&dpcpu_lock); return (s); } /* * Free dynamic per-cpu space at module unload time. */ void dpcpu_free(void *s, int size) { struct dpcpu_free *df; struct dpcpu_free *dn; uintptr_t start; uintptr_t end; size = roundup2(size, sizeof(void *)); start = (uintptr_t)s; end = start + size; /* * Free a region of space and merge it with as many neighbors as * possible. Keeping the list sorted simplifies this operation. */ sx_xlock(&dpcpu_lock); TAILQ_FOREACH(df, &dpcpu_head, df_link) { if (df->df_start > end) break; /* * If we expand at the end of an entry we may have to * merge it with the one following it as well. */ if (df->df_start + df->df_len == start) { df->df_len += size; dn = TAILQ_NEXT(df, df_link); if (df->df_start + df->df_len == dn->df_start) { df->df_len += dn->df_len; TAILQ_REMOVE(&dpcpu_head, dn, df_link); free(dn, M_PCPU); } sx_xunlock(&dpcpu_lock); return; } if (df->df_start == end) { df->df_start = start; df->df_len += size; sx_xunlock(&dpcpu_lock); return; } } dn = malloc(sizeof(*df), M_PCPU, M_WAITOK | M_ZERO); dn->df_start = start; dn->df_len = size; if (df) TAILQ_INSERT_BEFORE(df, dn, df_link); else TAILQ_INSERT_TAIL(&dpcpu_head, dn, df_link); sx_xunlock(&dpcpu_lock); } /* * Initialize the per-cpu storage from an updated linker-set region. */ void dpcpu_copy(void *s, int size) { #ifdef SMP uintptr_t dpcpu; int i; - for (i = 0; i < mp_ncpus; ++i) { + CPU_FOREACH(i) { dpcpu = dpcpu_off[i]; if (dpcpu == 0) continue; memcpy((void *)(dpcpu + (uintptr_t)s), s, size); } #else memcpy((void *)(dpcpu_off[0] + (uintptr_t)s), s, size); #endif } /* * Destroy a struct pcpu. */ void pcpu_destroy(struct pcpu *pcpu) { STAILQ_REMOVE(&cpuhead, pcpu, pcpu, pc_allcpu); cpuid_to_pcpu[pcpu->pc_cpuid] = NULL; dpcpu_off[pcpu->pc_cpuid] = 0; } /* * Locate a struct pcpu by cpu id. */ struct pcpu * pcpu_find(u_int cpuid) { return (cpuid_to_pcpu[cpuid]); } int sysctl_dpcpu_quad(SYSCTL_HANDLER_ARGS) { uintptr_t dpcpu; int64_t count; int i; count = 0; - for (i = 0; i < mp_ncpus; ++i) { + CPU_FOREACH(i) { dpcpu = dpcpu_off[i]; if (dpcpu == 0) continue; count += *(int64_t *)(dpcpu + (uintptr_t)arg1); } return (SYSCTL_OUT(req, &count, sizeof(count))); } int sysctl_dpcpu_long(SYSCTL_HANDLER_ARGS) { uintptr_t dpcpu; long count; int i; count = 0; - for (i = 0; i < mp_ncpus; ++i) { + CPU_FOREACH(i) { dpcpu = dpcpu_off[i]; if (dpcpu == 0) continue; count += *(long *)(dpcpu + (uintptr_t)arg1); } return (SYSCTL_OUT(req, &count, sizeof(count))); } int sysctl_dpcpu_int(SYSCTL_HANDLER_ARGS) { uintptr_t dpcpu; int count; int i; count = 0; - for (i = 0; i < mp_ncpus; ++i) { + CPU_FOREACH(i) { dpcpu = dpcpu_off[i]; if (dpcpu == 0) continue; count += *(int *)(dpcpu + (uintptr_t)arg1); } return (SYSCTL_OUT(req, &count, sizeof(count))); } #ifdef DDB DB_SHOW_COMMAND(dpcpu_off, db_show_dpcpu_off) { int id; CPU_FOREACH(id) { db_printf("dpcpu_off[%2d] = 0x%jx (+ DPCPU_START = %p)\n", id, (uintmax_t)dpcpu_off[id], (void *)(uintptr_t)(dpcpu_off[id] + DPCPU_START)); } } static void show_pcpu(struct pcpu *pc) { struct thread *td; db_printf("cpuid = %d\n", pc->pc_cpuid); db_printf("dynamic pcpu = %p\n", (void *)pc->pc_dynamic); db_printf("curthread = "); td = pc->pc_curthread; if (td != NULL) db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid, td->td_name); else db_printf("none\n"); db_printf("curpcb = %p\n", pc->pc_curpcb); db_printf("fpcurthread = "); td = pc->pc_fpcurthread; if (td != NULL) db_printf("%p: pid %d \"%s\"\n", td, td->td_proc->p_pid, td->td_name); else db_printf("none\n"); db_printf("idlethread = "); td = pc->pc_idlethread; if (td != NULL) db_printf("%p: tid %d \"%s\"\n", td, td->td_tid, td->td_name); else db_printf("none\n"); db_show_mdpcpu(pc); #ifdef VIMAGE db_printf("curvnet = %p\n", pc->pc_curthread->td_vnet); #endif #ifdef WITNESS db_printf("spin locks held:\n"); witness_list_locks(&pc->pc_spinlocks, db_printf); #endif } DB_SHOW_COMMAND(pcpu, db_show_pcpu) { struct pcpu *pc; int id; if (have_addr) id = ((addr >> 4) % 16) * 10 + (addr % 16); else id = PCPU_GET(cpuid); pc = pcpu_find(id); if (pc == NULL) { db_printf("CPU %d not found\n", id); return; } show_pcpu(pc); } DB_SHOW_ALL_COMMAND(pcpu, db_show_cpu_all) { struct pcpu *pc; int id; db_printf("Current CPU: %d\n\n", PCPU_GET(cpuid)); for (id = 0; id <= mp_maxid; id++) { pc = pcpu_find(id); if (pc != NULL) { show_pcpu(pc); db_printf("\n"); } } } DB_SHOW_ALIAS(allpcpu, db_show_cpu_all); #endif Index: head/sys/kern/subr_taskqueue.c =================================================================== --- head/sys/kern/subr_taskqueue.c (revision 302371) +++ head/sys/kern/subr_taskqueue.c (revision 302372) @@ -1,1139 +1,1152 @@ /*- * Copyright (c) 2000 Doug Rabson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_TASKQUEUE, "taskqueue", "Task Queues"); static void *taskqueue_giant_ih; static void *taskqueue_ih; static void taskqueue_fast_enqueue(void *); static void taskqueue_swi_enqueue(void *); static void taskqueue_swi_giant_enqueue(void *); struct taskqueue_busy { struct task *tb_running; TAILQ_ENTRY(taskqueue_busy) tb_link; }; struct task * const TB_DRAIN_WAITER = (struct task *)0x1; struct taskqueue { STAILQ_HEAD(, task) tq_queue; taskqueue_enqueue_fn tq_enqueue; void *tq_context; char *tq_name; TAILQ_HEAD(, taskqueue_busy) tq_active; struct mtx tq_mutex; struct thread **tq_threads; int tq_tcount; int tq_spin; int tq_flags; int tq_callouts; taskqueue_callback_fn tq_callbacks[TASKQUEUE_NUM_CALLBACKS]; void *tq_cb_contexts[TASKQUEUE_NUM_CALLBACKS]; }; #define TQ_FLAGS_ACTIVE (1 << 0) #define TQ_FLAGS_BLOCKED (1 << 1) #define TQ_FLAGS_UNLOCKED_ENQUEUE (1 << 2) #define DT_CALLOUT_ARMED (1 << 0) #define TQ_LOCK(tq) \ do { \ if ((tq)->tq_spin) \ mtx_lock_spin(&(tq)->tq_mutex); \ else \ mtx_lock(&(tq)->tq_mutex); \ } while (0) #define TQ_ASSERT_LOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_OWNED) #define TQ_UNLOCK(tq) \ do { \ if ((tq)->tq_spin) \ mtx_unlock_spin(&(tq)->tq_mutex); \ else \ mtx_unlock(&(tq)->tq_mutex); \ } while (0) #define TQ_ASSERT_UNLOCKED(tq) mtx_assert(&(tq)->tq_mutex, MA_NOTOWNED) void _timeout_task_init(struct taskqueue *queue, struct timeout_task *timeout_task, int priority, task_fn_t func, void *context) { TASK_INIT(&timeout_task->t, priority, func, context); callout_init_mtx(&timeout_task->c, &queue->tq_mutex, CALLOUT_RETURNUNLOCKED); timeout_task->q = queue; timeout_task->f = 0; } static __inline int TQ_SLEEP(struct taskqueue *tq, void *p, struct mtx *m, int pri, const char *wm, int t) { if (tq->tq_spin) return (msleep_spin(p, m, wm, t)); return (msleep(p, m, pri, wm, t)); } static struct taskqueue * _taskqueue_create(const char *name, int mflags, taskqueue_enqueue_fn enqueue, void *context, int mtxflags, const char *mtxname __unused) { struct taskqueue *queue; char *tq_name; tq_name = malloc(TASKQUEUE_NAMELEN, M_TASKQUEUE, mflags | M_ZERO); if (tq_name == NULL) return (NULL); queue = malloc(sizeof(struct taskqueue), M_TASKQUEUE, mflags | M_ZERO); if (queue == NULL) { free(tq_name, M_TASKQUEUE); return (NULL); } snprintf(tq_name, TASKQUEUE_NAMELEN, "%s", (name) ? name : "taskqueue"); STAILQ_INIT(&queue->tq_queue); TAILQ_INIT(&queue->tq_active); queue->tq_enqueue = enqueue; queue->tq_context = context; queue->tq_name = tq_name; queue->tq_spin = (mtxflags & MTX_SPIN) != 0; queue->tq_flags |= TQ_FLAGS_ACTIVE; if (enqueue == taskqueue_fast_enqueue || enqueue == taskqueue_swi_enqueue || enqueue == taskqueue_swi_giant_enqueue || enqueue == taskqueue_thread_enqueue) queue->tq_flags |= TQ_FLAGS_UNLOCKED_ENQUEUE; mtx_init(&queue->tq_mutex, tq_name, NULL, mtxflags); return (queue); } struct taskqueue * taskqueue_create(const char *name, int mflags, taskqueue_enqueue_fn enqueue, void *context) { return _taskqueue_create(name, mflags, enqueue, context, MTX_DEF, name); } void taskqueue_set_callback(struct taskqueue *queue, enum taskqueue_callback_type cb_type, taskqueue_callback_fn callback, void *context) { KASSERT(((cb_type >= TASKQUEUE_CALLBACK_TYPE_MIN) && (cb_type <= TASKQUEUE_CALLBACK_TYPE_MAX)), ("Callback type %d not valid, must be %d-%d", cb_type, TASKQUEUE_CALLBACK_TYPE_MIN, TASKQUEUE_CALLBACK_TYPE_MAX)); KASSERT((queue->tq_callbacks[cb_type] == NULL), ("Re-initialization of taskqueue callback?")); queue->tq_callbacks[cb_type] = callback; queue->tq_cb_contexts[cb_type] = context; } /* * Signal a taskqueue thread to terminate. */ static void taskqueue_terminate(struct thread **pp, struct taskqueue *tq) { while (tq->tq_tcount > 0 || tq->tq_callouts > 0) { wakeup(tq); TQ_SLEEP(tq, pp, &tq->tq_mutex, PWAIT, "taskqueue_destroy", 0); } } void taskqueue_free(struct taskqueue *queue) { TQ_LOCK(queue); queue->tq_flags &= ~TQ_FLAGS_ACTIVE; taskqueue_terminate(queue->tq_threads, queue); KASSERT(TAILQ_EMPTY(&queue->tq_active), ("Tasks still running?")); KASSERT(queue->tq_callouts == 0, ("Armed timeout tasks")); mtx_destroy(&queue->tq_mutex); free(queue->tq_threads, M_TASKQUEUE); free(queue->tq_name, M_TASKQUEUE); free(queue, M_TASKQUEUE); } static int taskqueue_enqueue_locked(struct taskqueue *queue, struct task *task) { struct task *ins; struct task *prev; KASSERT(task->ta_func != NULL, ("enqueueing task with NULL func")); /* * Count multiple enqueues. */ if (task->ta_pending) { if (task->ta_pending < USHRT_MAX) task->ta_pending++; TQ_UNLOCK(queue); return (0); } /* * Optimise the case when all tasks have the same priority. */ prev = STAILQ_LAST(&queue->tq_queue, task, ta_link); if (!prev || prev->ta_priority >= task->ta_priority) { STAILQ_INSERT_TAIL(&queue->tq_queue, task, ta_link); } else { prev = NULL; for (ins = STAILQ_FIRST(&queue->tq_queue); ins; prev = ins, ins = STAILQ_NEXT(ins, ta_link)) if (ins->ta_priority < task->ta_priority) break; if (prev) STAILQ_INSERT_AFTER(&queue->tq_queue, prev, task, ta_link); else STAILQ_INSERT_HEAD(&queue->tq_queue, task, ta_link); } task->ta_pending = 1; if ((queue->tq_flags & TQ_FLAGS_UNLOCKED_ENQUEUE) != 0) TQ_UNLOCK(queue); if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0) queue->tq_enqueue(queue->tq_context); if ((queue->tq_flags & TQ_FLAGS_UNLOCKED_ENQUEUE) == 0) TQ_UNLOCK(queue); /* Return with lock released. */ return (0); } int grouptaskqueue_enqueue(struct taskqueue *queue, struct task *task) { TQ_LOCK(queue); if (task->ta_pending) { TQ_UNLOCK(queue); return (0); } STAILQ_INSERT_TAIL(&queue->tq_queue, task, ta_link); task->ta_pending = 1; TQ_UNLOCK(queue); if ((queue->tq_flags & TQ_FLAGS_BLOCKED) == 0) queue->tq_enqueue(queue->tq_context); return (0); } int taskqueue_enqueue(struct taskqueue *queue, struct task *task) { int res; TQ_LOCK(queue); res = taskqueue_enqueue_locked(queue, task); /* The lock is released inside. */ return (res); } static void taskqueue_timeout_func(void *arg) { struct taskqueue *queue; struct timeout_task *timeout_task; timeout_task = arg; queue = timeout_task->q; KASSERT((timeout_task->f & DT_CALLOUT_ARMED) != 0, ("Stray timeout")); timeout_task->f &= ~DT_CALLOUT_ARMED; queue->tq_callouts--; taskqueue_enqueue_locked(timeout_task->q, &timeout_task->t); /* The lock is released inside. */ } int taskqueue_enqueue_timeout(struct taskqueue *queue, struct timeout_task *timeout_task, int ticks) { int res; TQ_LOCK(queue); KASSERT(timeout_task->q == NULL || timeout_task->q == queue, ("Migrated queue")); KASSERT(!queue->tq_spin, ("Timeout for spin-queue")); timeout_task->q = queue; res = timeout_task->t.ta_pending; if (ticks == 0) { taskqueue_enqueue_locked(queue, &timeout_task->t); /* The lock is released inside. */ } else { if ((timeout_task->f & DT_CALLOUT_ARMED) != 0) { res++; } else { queue->tq_callouts++; timeout_task->f |= DT_CALLOUT_ARMED; if (ticks < 0) ticks = -ticks; /* Ignore overflow. */ } if (ticks > 0) { callout_reset(&timeout_task->c, ticks, taskqueue_timeout_func, timeout_task); } TQ_UNLOCK(queue); } return (res); } static void taskqueue_task_nop_fn(void *context, int pending) { } /* * Block until all currently queued tasks in this taskqueue * have begun execution. Tasks queued during execution of * this function are ignored. */ static void taskqueue_drain_tq_queue(struct taskqueue *queue) { struct task t_barrier; if (STAILQ_EMPTY(&queue->tq_queue)) return; /* * Enqueue our barrier after all current tasks, but with * the highest priority so that newly queued tasks cannot * pass it. Because of the high priority, we can not use * taskqueue_enqueue_locked directly (which drops the lock * anyway) so just insert it at tail while we have the * queue lock. */ TASK_INIT(&t_barrier, USHRT_MAX, taskqueue_task_nop_fn, &t_barrier); STAILQ_INSERT_TAIL(&queue->tq_queue, &t_barrier, ta_link); t_barrier.ta_pending = 1; /* * Once the barrier has executed, all previously queued tasks * have completed or are currently executing. */ while (t_barrier.ta_pending != 0) TQ_SLEEP(queue, &t_barrier, &queue->tq_mutex, PWAIT, "-", 0); } /* * Block until all currently executing tasks for this taskqueue * complete. Tasks that begin execution during the execution * of this function are ignored. */ static void taskqueue_drain_tq_active(struct taskqueue *queue) { struct taskqueue_busy tb_marker, *tb_first; if (TAILQ_EMPTY(&queue->tq_active)) return; /* Block taskq_terminate().*/ queue->tq_callouts++; /* * Wait for all currently executing taskqueue threads * to go idle. */ tb_marker.tb_running = TB_DRAIN_WAITER; TAILQ_INSERT_TAIL(&queue->tq_active, &tb_marker, tb_link); while (TAILQ_FIRST(&queue->tq_active) != &tb_marker) TQ_SLEEP(queue, &tb_marker, &queue->tq_mutex, PWAIT, "-", 0); TAILQ_REMOVE(&queue->tq_active, &tb_marker, tb_link); /* * Wakeup any other drain waiter that happened to queue up * without any intervening active thread. */ tb_first = TAILQ_FIRST(&queue->tq_active); if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER) wakeup(tb_first); /* Release taskqueue_terminate(). */ queue->tq_callouts--; if ((queue->tq_flags & TQ_FLAGS_ACTIVE) == 0) wakeup_one(queue->tq_threads); } void taskqueue_block(struct taskqueue *queue) { TQ_LOCK(queue); queue->tq_flags |= TQ_FLAGS_BLOCKED; TQ_UNLOCK(queue); } void taskqueue_unblock(struct taskqueue *queue) { TQ_LOCK(queue); queue->tq_flags &= ~TQ_FLAGS_BLOCKED; if (!STAILQ_EMPTY(&queue->tq_queue)) queue->tq_enqueue(queue->tq_context); TQ_UNLOCK(queue); } static void taskqueue_run_locked(struct taskqueue *queue) { struct taskqueue_busy tb; struct taskqueue_busy *tb_first; struct task *task; int pending; KASSERT(queue != NULL, ("tq is NULL")); TQ_ASSERT_LOCKED(queue); tb.tb_running = NULL; while (STAILQ_FIRST(&queue->tq_queue)) { TAILQ_INSERT_TAIL(&queue->tq_active, &tb, tb_link); /* * Carefully remove the first task from the queue and * zero its pending count. */ task = STAILQ_FIRST(&queue->tq_queue); KASSERT(task != NULL, ("task is NULL")); STAILQ_REMOVE_HEAD(&queue->tq_queue, ta_link); pending = task->ta_pending; task->ta_pending = 0; tb.tb_running = task; TQ_UNLOCK(queue); KASSERT(task->ta_func != NULL, ("task->ta_func is NULL")); task->ta_func(task->ta_context, pending); TQ_LOCK(queue); tb.tb_running = NULL; wakeup(task); TAILQ_REMOVE(&queue->tq_active, &tb, tb_link); tb_first = TAILQ_FIRST(&queue->tq_active); if (tb_first != NULL && tb_first->tb_running == TB_DRAIN_WAITER) wakeup(tb_first); } } void taskqueue_run(struct taskqueue *queue) { TQ_LOCK(queue); taskqueue_run_locked(queue); TQ_UNLOCK(queue); } static int task_is_running(struct taskqueue *queue, struct task *task) { struct taskqueue_busy *tb; TQ_ASSERT_LOCKED(queue); TAILQ_FOREACH(tb, &queue->tq_active, tb_link) { if (tb->tb_running == task) return (1); } return (0); } static int taskqueue_cancel_locked(struct taskqueue *queue, struct task *task, u_int *pendp) { if (task->ta_pending > 0) STAILQ_REMOVE(&queue->tq_queue, task, task, ta_link); if (pendp != NULL) *pendp = task->ta_pending; task->ta_pending = 0; return (task_is_running(queue, task) ? EBUSY : 0); } int taskqueue_cancel(struct taskqueue *queue, struct task *task, u_int *pendp) { int error; TQ_LOCK(queue); error = taskqueue_cancel_locked(queue, task, pendp); TQ_UNLOCK(queue); return (error); } int taskqueue_cancel_timeout(struct taskqueue *queue, struct timeout_task *timeout_task, u_int *pendp) { u_int pending, pending1; int error; TQ_LOCK(queue); pending = !!(callout_stop(&timeout_task->c) > 0); error = taskqueue_cancel_locked(queue, &timeout_task->t, &pending1); if ((timeout_task->f & DT_CALLOUT_ARMED) != 0) { timeout_task->f &= ~DT_CALLOUT_ARMED; queue->tq_callouts--; } TQ_UNLOCK(queue); if (pendp != NULL) *pendp = pending + pending1; return (error); } void taskqueue_drain(struct taskqueue *queue, struct task *task) { if (!queue->tq_spin) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); TQ_LOCK(queue); while (task->ta_pending != 0 || task_is_running(queue, task)) TQ_SLEEP(queue, task, &queue->tq_mutex, PWAIT, "-", 0); TQ_UNLOCK(queue); } void taskqueue_drain_all(struct taskqueue *queue) { if (!queue->tq_spin) WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); TQ_LOCK(queue); taskqueue_drain_tq_queue(queue); taskqueue_drain_tq_active(queue); TQ_UNLOCK(queue); } void taskqueue_drain_timeout(struct taskqueue *queue, struct timeout_task *timeout_task) { callout_drain(&timeout_task->c); taskqueue_drain(queue, &timeout_task->t); } static void taskqueue_swi_enqueue(void *context) { swi_sched(taskqueue_ih, 0); } static void taskqueue_swi_run(void *dummy) { taskqueue_run(taskqueue_swi); } static void taskqueue_swi_giant_enqueue(void *context) { swi_sched(taskqueue_giant_ih, 0); } static void taskqueue_swi_giant_run(void *dummy) { taskqueue_run(taskqueue_swi_giant); } static int _taskqueue_start_threads(struct taskqueue **tqp, int count, int pri, cpuset_t *mask, const char *name, va_list ap) { char ktname[MAXCOMLEN + 1]; struct thread *td; struct taskqueue *tq; int i, error; if (count <= 0) return (EINVAL); vsnprintf(ktname, sizeof(ktname), name, ap); tq = *tqp; tq->tq_threads = malloc(sizeof(struct thread *) * count, M_TASKQUEUE, M_NOWAIT | M_ZERO); if (tq->tq_threads == NULL) { printf("%s: no memory for %s threads\n", __func__, ktname); return (ENOMEM); } for (i = 0; i < count; i++) { if (count == 1) error = kthread_add(taskqueue_thread_loop, tqp, NULL, &tq->tq_threads[i], RFSTOPPED, 0, "%s", ktname); else error = kthread_add(taskqueue_thread_loop, tqp, NULL, &tq->tq_threads[i], RFSTOPPED, 0, "%s_%d", ktname, i); if (error) { /* should be ok to continue, taskqueue_free will dtrt */ printf("%s: kthread_add(%s): error %d", __func__, ktname, error); tq->tq_threads[i] = NULL; /* paranoid */ } else tq->tq_tcount++; } for (i = 0; i < count; i++) { if (tq->tq_threads[i] == NULL) continue; td = tq->tq_threads[i]; if (mask) { error = cpuset_setthread(td->td_tid, mask); /* * Failing to pin is rarely an actual fatal error; * it'll just affect performance. */ if (error) printf("%s: curthread=%llu: can't pin; " "error=%d\n", __func__, (unsigned long long) td->td_tid, error); } thread_lock(td); sched_prio(td, pri); sched_add(td, SRQ_BORING); thread_unlock(td); } return (0); } int taskqueue_start_threads(struct taskqueue **tqp, int count, int pri, const char *name, ...) { va_list ap; int error; va_start(ap, name); error = _taskqueue_start_threads(tqp, count, pri, NULL, name, ap); va_end(ap); return (error); } int taskqueue_start_threads_cpuset(struct taskqueue **tqp, int count, int pri, cpuset_t *mask, const char *name, ...) { va_list ap; int error; va_start(ap, name); error = _taskqueue_start_threads(tqp, count, pri, mask, name, ap); va_end(ap); return (error); } static inline void taskqueue_run_callback(struct taskqueue *tq, enum taskqueue_callback_type cb_type) { taskqueue_callback_fn tq_callback; TQ_ASSERT_UNLOCKED(tq); tq_callback = tq->tq_callbacks[cb_type]; if (tq_callback != NULL) tq_callback(tq->tq_cb_contexts[cb_type]); } void taskqueue_thread_loop(void *arg) { struct taskqueue **tqp, *tq; tqp = arg; tq = *tqp; taskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_INIT); TQ_LOCK(tq); while ((tq->tq_flags & TQ_FLAGS_ACTIVE) != 0) { /* XXX ? */ taskqueue_run_locked(tq); /* * Because taskqueue_run() can drop tq_mutex, we need to * check if the TQ_FLAGS_ACTIVE flag wasn't removed in the * meantime, which means we missed a wakeup. */ if ((tq->tq_flags & TQ_FLAGS_ACTIVE) == 0) break; TQ_SLEEP(tq, tq, &tq->tq_mutex, 0, "-", 0); } taskqueue_run_locked(tq); /* * This thread is on its way out, so just drop the lock temporarily * in order to call the shutdown callback. This allows the callback * to look at the taskqueue, even just before it dies. */ TQ_UNLOCK(tq); taskqueue_run_callback(tq, TASKQUEUE_CALLBACK_TYPE_SHUTDOWN); TQ_LOCK(tq); /* rendezvous with thread that asked us to terminate */ tq->tq_tcount--; wakeup_one(tq->tq_threads); TQ_UNLOCK(tq); kthread_exit(); } void taskqueue_thread_enqueue(void *context) { struct taskqueue **tqp, *tq; tqp = context; tq = *tqp; wakeup_one(tq); } TASKQUEUE_DEFINE(swi, taskqueue_swi_enqueue, NULL, swi_add(NULL, "task queue", taskqueue_swi_run, NULL, SWI_TQ, INTR_MPSAFE, &taskqueue_ih)); TASKQUEUE_DEFINE(swi_giant, taskqueue_swi_giant_enqueue, NULL, swi_add(NULL, "Giant taskq", taskqueue_swi_giant_run, NULL, SWI_TQ_GIANT, 0, &taskqueue_giant_ih)); TASKQUEUE_DEFINE_THREAD(thread); struct taskqueue * taskqueue_create_fast(const char *name, int mflags, taskqueue_enqueue_fn enqueue, void *context) { return _taskqueue_create(name, mflags, enqueue, context, MTX_SPIN, "fast_taskqueue"); } static void *taskqueue_fast_ih; static void taskqueue_fast_enqueue(void *context) { swi_sched(taskqueue_fast_ih, 0); } static void taskqueue_fast_run(void *dummy) { taskqueue_run(taskqueue_fast); } TASKQUEUE_FAST_DEFINE(fast, taskqueue_fast_enqueue, NULL, swi_add(NULL, "fast taskq", taskqueue_fast_run, NULL, SWI_TQ_FAST, INTR_MPSAFE, &taskqueue_fast_ih)); int taskqueue_member(struct taskqueue *queue, struct thread *td) { int i, j, ret = 0; for (i = 0, j = 0; ; i++) { if (queue->tq_threads[i] == NULL) continue; if (queue->tq_threads[i] == td) { ret = 1; break; } if (++j >= queue->tq_tcount) break; } return (ret); } struct taskqgroup_cpu { LIST_HEAD(, grouptask) tgc_tasks; struct taskqueue *tgc_taskq; int tgc_cnt; int tgc_cpu; }; struct taskqgroup { struct taskqgroup_cpu tqg_queue[MAXCPU]; struct mtx tqg_lock; char * tqg_name; int tqg_adjusting; int tqg_stride; int tqg_cnt; }; struct taskq_bind_task { struct task bt_task; int bt_cpuid; }; static void taskqgroup_cpu_create(struct taskqgroup *qgroup, int idx) { struct taskqgroup_cpu *qcpu; + int i, j; qcpu = &qgroup->tqg_queue[idx]; LIST_INIT(&qcpu->tgc_tasks); qcpu->tgc_taskq = taskqueue_create_fast(NULL, M_WAITOK, taskqueue_thread_enqueue, &qcpu->tgc_taskq); taskqueue_start_threads(&qcpu->tgc_taskq, 1, PI_SOFT, "%s_%d", qgroup->tqg_name, idx); - qcpu->tgc_cpu = idx * qgroup->tqg_stride; + + for (i = CPU_FIRST(), j = 0; j < idx * qgroup->tqg_stride; + j++, i = CPU_NEXT(i)) { + /* + * Wait: evaluate the idx * qgroup->tqg_stride'th CPU, + * potentially wrapping the actual count + */ + } + qcpu->tgc_cpu = i; } static void taskqgroup_cpu_remove(struct taskqgroup *qgroup, int idx) { taskqueue_free(qgroup->tqg_queue[idx].tgc_taskq); } /* * Find the taskq with least # of tasks that doesn't currently have any * other queues from the uniq identifier. */ static int taskqgroup_find(struct taskqgroup *qgroup, void *uniq) { struct grouptask *n; int i, idx, mincnt; int strict; mtx_assert(&qgroup->tqg_lock, MA_OWNED); if (qgroup->tqg_cnt == 0) return (0); idx = -1; mincnt = INT_MAX; /* * Two passes; First scan for a queue with the least tasks that * does not already service this uniq id. If that fails simply find * the queue with the least total tasks; */ for (strict = 1; mincnt == INT_MAX; strict = 0) { for (i = 0; i < qgroup->tqg_cnt; i++) { if (qgroup->tqg_queue[i].tgc_cnt > mincnt) continue; if (strict) { LIST_FOREACH(n, &qgroup->tqg_queue[i].tgc_tasks, gt_list) if (n->gt_uniq == uniq) break; if (n != NULL) continue; } mincnt = qgroup->tqg_queue[i].tgc_cnt; idx = i; } } if (idx == -1) panic("taskqgroup_find: Failed to pick a qid."); return (idx); } void taskqgroup_attach(struct taskqgroup *qgroup, struct grouptask *gtask, void *uniq, int irq, char *name) { cpuset_t mask; int qid; gtask->gt_uniq = uniq; gtask->gt_name = name; gtask->gt_irq = irq; gtask->gt_cpu = -1; mtx_lock(&qgroup->tqg_lock); qid = taskqgroup_find(qgroup, uniq); qgroup->tqg_queue[qid].tgc_cnt++; LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; if (irq != -1 && smp_started) { CPU_ZERO(&mask); CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask); mtx_unlock(&qgroup->tqg_lock); intr_setaffinity(irq, &mask); } else mtx_unlock(&qgroup->tqg_lock); } int taskqgroup_attach_cpu(struct taskqgroup *qgroup, struct grouptask *gtask, void *uniq, int cpu, int irq, char *name) { cpuset_t mask; int i, qid; qid = -1; gtask->gt_uniq = uniq; gtask->gt_name = name; gtask->gt_irq = irq; gtask->gt_cpu = cpu; mtx_lock(&qgroup->tqg_lock); if (smp_started) { for (i = 0; i < qgroup->tqg_cnt; i++) if (qgroup->tqg_queue[i].tgc_cpu == cpu) { qid = i; break; } if (qid == -1) { mtx_unlock(&qgroup->tqg_lock); return (EINVAL); } } else qid = 0; qgroup->tqg_queue[qid].tgc_cnt++; LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; if (irq != -1 && smp_started) { CPU_ZERO(&mask); CPU_SET(qgroup->tqg_queue[qid].tgc_cpu, &mask); mtx_unlock(&qgroup->tqg_lock); intr_setaffinity(irq, &mask); } else mtx_unlock(&qgroup->tqg_lock); return (0); } void taskqgroup_detach(struct taskqgroup *qgroup, struct grouptask *gtask) { int i; mtx_lock(&qgroup->tqg_lock); for (i = 0; i < qgroup->tqg_cnt; i++) if (qgroup->tqg_queue[i].tgc_taskq == gtask->gt_taskqueue) break; if (i == qgroup->tqg_cnt) panic("taskqgroup_detach: task not in group\n"); qgroup->tqg_queue[i].tgc_cnt--; LIST_REMOVE(gtask, gt_list); mtx_unlock(&qgroup->tqg_lock); gtask->gt_taskqueue = NULL; } static void taskqgroup_binder(void *ctx, int pending) { struct taskq_bind_task *task = (struct taskq_bind_task *)ctx; cpuset_t mask; int error; CPU_ZERO(&mask); CPU_SET(task->bt_cpuid, &mask); error = cpuset_setthread(curthread->td_tid, &mask); thread_lock(curthread); sched_bind(curthread, task->bt_cpuid); thread_unlock(curthread); if (error) printf("taskqgroup_binder: setaffinity failed: %d\n", error); free(task, M_DEVBUF); } static void taskqgroup_bind(struct taskqgroup *qgroup) { struct taskq_bind_task *task; int i; /* * Bind taskqueue threads to specific CPUs, if they have been assigned * one. */ for (i = 0; i < qgroup->tqg_cnt; i++) { task = malloc(sizeof (*task), M_DEVBUF, M_NOWAIT); TASK_INIT(&task->bt_task, 0, taskqgroup_binder, task); task->bt_cpuid = qgroup->tqg_queue[i].tgc_cpu; taskqueue_enqueue(qgroup->tqg_queue[i].tgc_taskq, &task->bt_task); } } static int _taskqgroup_adjust(struct taskqgroup *qgroup, int cnt, int stride) { LIST_HEAD(, grouptask) gtask_head = LIST_HEAD_INITIALIZER(NULL); cpuset_t mask; struct grouptask *gtask; - int i, old_cnt, qid; + int i, k, old_cnt, qid, cpu; mtx_assert(&qgroup->tqg_lock, MA_OWNED); if (cnt < 1 || cnt * stride > mp_ncpus || !smp_started) { - printf("taskqgroup_adjust failed cnt: %d stride: %d mp_ncpus: %d smp_started: %d\n", - cnt, stride, mp_ncpus, smp_started); + printf("taskqgroup_adjust failed cnt: %d stride: %d " + "mp_ncpus: %d smp_started: %d\n", cnt, stride, mp_ncpus, + smp_started); return (EINVAL); } if (qgroup->tqg_adjusting) { printf("taskqgroup_adjust failed: adjusting\n"); return (EBUSY); } qgroup->tqg_adjusting = 1; old_cnt = qgroup->tqg_cnt; mtx_unlock(&qgroup->tqg_lock); /* * Set up queue for tasks added before boot. */ if (old_cnt == 0) { LIST_SWAP(>ask_head, &qgroup->tqg_queue[0].tgc_tasks, grouptask, gt_list); qgroup->tqg_queue[0].tgc_cnt = 0; } /* * If new taskq threads have been added. */ for (i = old_cnt; i < cnt; i++) taskqgroup_cpu_create(qgroup, i); mtx_lock(&qgroup->tqg_lock); qgroup->tqg_cnt = cnt; qgroup->tqg_stride = stride; /* * Adjust drivers to use new taskqs. */ for (i = 0; i < old_cnt; i++) { while ((gtask = LIST_FIRST(&qgroup->tqg_queue[i].tgc_tasks))) { LIST_REMOVE(gtask, gt_list); qgroup->tqg_queue[i].tgc_cnt--; LIST_INSERT_HEAD(>ask_head, gtask, gt_list); } } while ((gtask = LIST_FIRST(>ask_head))) { LIST_REMOVE(gtask, gt_list); if (gtask->gt_cpu == -1) qid = taskqgroup_find(qgroup, gtask->gt_uniq); else { for (i = 0; i < qgroup->tqg_cnt; i++) if (qgroup->tqg_queue[i].tgc_cpu == gtask->gt_cpu) { qid = i; break; } } qgroup->tqg_queue[qid].tgc_cnt++; LIST_INSERT_HEAD(&qgroup->tqg_queue[qid].tgc_tasks, gtask, gt_list); gtask->gt_taskqueue = qgroup->tqg_queue[qid].tgc_taskq; } /* * Set new CPU and IRQ affinity */ + cpu = CPU_FIRST(); for (i = 0; i < cnt; i++) { - qgroup->tqg_queue[i].tgc_cpu = i * qgroup->tqg_stride; + qgroup->tqg_queue[i].tgc_cpu = cpu; + for (k = 0; k < qgroup->tqg_stride; k++) + cpu = CPU_NEXT(cpu); CPU_ZERO(&mask); CPU_SET(qgroup->tqg_queue[i].tgc_cpu, &mask); LIST_FOREACH(gtask, &qgroup->tqg_queue[i].tgc_tasks, gt_list) { if (gtask->gt_irq == -1) continue; intr_setaffinity(gtask->gt_irq, &mask); } } mtx_unlock(&qgroup->tqg_lock); /* * If taskq thread count has been reduced. */ for (i = cnt; i < old_cnt; i++) taskqgroup_cpu_remove(qgroup, i); mtx_lock(&qgroup->tqg_lock); qgroup->tqg_adjusting = 0; taskqgroup_bind(qgroup); return (0); } int taskqgroup_adjust(struct taskqgroup *qgroup, int cpu, int stride) { int error; mtx_lock(&qgroup->tqg_lock); error = _taskqgroup_adjust(qgroup, cpu, stride); mtx_unlock(&qgroup->tqg_lock); return (error); } struct taskqgroup * taskqgroup_create(char *name) { struct taskqgroup *qgroup; qgroup = malloc(sizeof(*qgroup), M_TASKQUEUE, M_WAITOK | M_ZERO); mtx_init(&qgroup->tqg_lock, "taskqgroup", NULL, MTX_DEF); qgroup->tqg_name = name; LIST_INIT(&qgroup->tqg_queue[0].tgc_tasks); return (qgroup); } void taskqgroup_destroy(struct taskqgroup *qgroup) { } Index: head/sys/net/flowtable.c =================================================================== --- head/sys/net/flowtable.c (revision 302371) +++ head/sys/net/flowtable.c (revision 302372) @@ -1,1184 +1,1184 @@ /*- * Copyright (c) 2014 Gleb Smirnoff * Copyright (c) 2008-2010, BitGravity Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * * 2. Neither the name of the BitGravity Corporation 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 COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ #include "opt_route.h" #include "opt_mpath.h" #include "opt_ddb.h" #include "opt_inet.h" #include "opt_inet6.h" #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #ifdef FLOWTABLE_HASH_ALL #include #include #include #endif #include #ifdef FLOWTABLE_HASH_ALL #define KEY_PORTS (sizeof(uint16_t) * 2) #define KEY_ADDRS 2 #else #define KEY_PORTS 0 #define KEY_ADDRS 1 #endif #ifdef INET6 #define KEY_ADDR_LEN sizeof(struct in6_addr) #else #define KEY_ADDR_LEN sizeof(struct in_addr) #endif #define KEYLEN ((KEY_ADDR_LEN * KEY_ADDRS + KEY_PORTS) / sizeof(uint32_t)) struct flentry { uint32_t f_hash; /* hash flowing forward */ uint32_t f_key[KEYLEN]; /* address(es and ports) */ uint32_t f_uptime; /* uptime at last access */ uint16_t f_fibnum; /* fib index */ #ifdef FLOWTABLE_HASH_ALL uint8_t f_proto; /* protocol */ uint8_t f_flags; /* stale? */ #define FL_STALE 1 #endif SLIST_ENTRY(flentry) f_next; /* pointer to collision entry */ struct rtentry *f_rt; /* rtentry for flow */ struct llentry *f_lle; /* llentry for flow */ }; #undef KEYLEN SLIST_HEAD(flist, flentry); /* Make sure we can use pcpu_zone_ptr for struct flist. */ CTASSERT(sizeof(struct flist) == sizeof(void *)); struct flowtable { counter_u64_t *ft_stat; int ft_size; /* * ft_table is a malloc(9)ed array of pointers. Pointers point to * memory from UMA_ZONE_PCPU zone. * ft_masks is per-cpu pointer itself. Each instance points * to a malloc(9)ed bitset, that is private to corresponding CPU. */ struct flist **ft_table; bitstr_t **ft_masks; bitstr_t *ft_tmpmask; }; #define FLOWSTAT_ADD(ft, name, v) \ counter_u64_add((ft)->ft_stat[offsetof(struct flowtable_stat, name) / sizeof(uint64_t)], (v)) #define FLOWSTAT_INC(ft, name) FLOWSTAT_ADD(ft, name, 1) static struct proc *flowcleanerproc; static uint32_t flow_hashjitter; static struct cv flowclean_f_cv; static struct cv flowclean_c_cv; static struct mtx flowclean_lock; static uint32_t flowclean_cycles; /* * TODO: * - add sysctls to resize && flush flow tables * - Add per flowtable sysctls for statistics and configuring timeouts * - add saturation counter to rtentry to support per-packet load-balancing * add flag to indicate round-robin flow, add list lookup from head for flows * - add sysctl / device node / syscall to support exporting and importing * of flows with flag to indicate that a flow was imported so should * not be considered for auto-cleaning * - support explicit connection state (currently only ad-hoc for DSR) * - idetach() cleanup for options VIMAGE builds. */ #ifdef INET static VNET_DEFINE(struct flowtable, ip4_ft); #define V_ip4_ft VNET(ip4_ft) #endif #ifdef INET6 static VNET_DEFINE(struct flowtable, ip6_ft); #define V_ip6_ft VNET(ip6_ft) #endif static uma_zone_t flow_zone; static VNET_DEFINE(int, flowtable_enable) = 1; #define V_flowtable_enable VNET(flowtable_enable) static SYSCTL_NODE(_net, OID_AUTO, flowtable, CTLFLAG_RD, NULL, "flowtable"); SYSCTL_INT(_net_flowtable, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(flowtable_enable), 0, "enable flowtable caching."); SYSCTL_UMA_MAX(_net_flowtable, OID_AUTO, maxflows, CTLFLAG_RW, &flow_zone, "Maximum number of flows allowed"); static MALLOC_DEFINE(M_FTABLE, "flowtable", "flowtable hashes and bitstrings"); static struct flentry * flowtable_lookup_common(struct flowtable *, uint32_t *, int, uint32_t); #ifdef INET static struct flentry * flowtable_lookup_ipv4(struct mbuf *m, struct route *ro) { struct flentry *fle; struct sockaddr_in *sin; struct ip *ip; uint32_t fibnum; #ifdef FLOWTABLE_HASH_ALL uint32_t key[3]; int iphlen; uint16_t sport, dport; uint8_t proto; #endif ip = mtod(m, struct ip *); if (ip->ip_src.s_addr == ip->ip_dst.s_addr || (ntohl(ip->ip_dst.s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET || (ntohl(ip->ip_src.s_addr) >> IN_CLASSA_NSHIFT) == IN_LOOPBACKNET) return (NULL); fibnum = M_GETFIB(m); #ifdef FLOWTABLE_HASH_ALL iphlen = ip->ip_hl << 2; proto = ip->ip_p; switch (proto) { case IPPROTO_TCP: { struct tcphdr *th; th = (struct tcphdr *)((char *)ip + iphlen); sport = th->th_sport; dport = th->th_dport; if (th->th_flags & (TH_RST|TH_FIN)) fibnum |= (FL_STALE << 24); break; } case IPPROTO_UDP: { struct udphdr *uh; uh = (struct udphdr *)((char *)ip + iphlen); sport = uh->uh_sport; dport = uh->uh_dport; break; } case IPPROTO_SCTP: { struct sctphdr *sh; sh = (struct sctphdr *)((char *)ip + iphlen); sport = sh->src_port; dport = sh->dest_port; /* XXXGL: handle stale? */ break; } default: sport = dport = 0; break; } key[0] = ip->ip_dst.s_addr; key[1] = ip->ip_src.s_addr; key[2] = (dport << 16) | sport; fibnum |= proto << 16; fle = flowtable_lookup_common(&V_ip4_ft, key, 3 * sizeof(uint32_t), fibnum); #else /* !FLOWTABLE_HASH_ALL */ fle = flowtable_lookup_common(&V_ip4_ft, (uint32_t *)&ip->ip_dst, sizeof(struct in_addr), fibnum); #endif /* FLOWTABLE_HASH_ALL */ if (fle == NULL) return (NULL); sin = (struct sockaddr_in *)&ro->ro_dst; sin->sin_family = AF_INET; sin->sin_len = sizeof(*sin); sin->sin_addr = ip->ip_dst; return (fle); } #endif /* INET */ #ifdef INET6 /* * PULLUP_TO(len, p, T) makes sure that len + sizeof(T) is contiguous, * then it sets p to point at the offset "len" in the mbuf. WARNING: the * pointer might become stale after other pullups (but we never use it * this way). */ #define PULLUP_TO(_len, p, T) \ do { \ int x = (_len) + sizeof(T); \ if ((m)->m_len < x) \ return (NULL); \ p = (mtod(m, char *) + (_len)); \ } while (0) #define TCP(p) ((struct tcphdr *)(p)) #define SCTP(p) ((struct sctphdr *)(p)) #define UDP(p) ((struct udphdr *)(p)) static struct flentry * flowtable_lookup_ipv6(struct mbuf *m, struct route *ro) { struct flentry *fle; struct sockaddr_in6 *sin6; struct ip6_hdr *ip6; uint32_t fibnum; #ifdef FLOWTABLE_HASH_ALL uint32_t key[9]; void *ulp; int hlen; uint16_t sport, dport; u_short offset; uint8_t proto; #else uint32_t key[4]; #endif ip6 = mtod(m, struct ip6_hdr *); if (in6_localaddr(&ip6->ip6_dst)) return (NULL); fibnum = M_GETFIB(m); #ifdef FLOWTABLE_HASH_ALL hlen = sizeof(struct ip6_hdr); proto = ip6->ip6_nxt; offset = sport = dport = 0; ulp = NULL; while (ulp == NULL) { switch (proto) { case IPPROTO_ICMPV6: case IPPROTO_OSPFIGP: case IPPROTO_PIM: case IPPROTO_CARP: case IPPROTO_ESP: case IPPROTO_NONE: ulp = ip6; break; case IPPROTO_TCP: PULLUP_TO(hlen, ulp, struct tcphdr); dport = TCP(ulp)->th_dport; sport = TCP(ulp)->th_sport; if (TCP(ulp)->th_flags & (TH_RST|TH_FIN)) fibnum |= (FL_STALE << 24); break; case IPPROTO_SCTP: PULLUP_TO(hlen, ulp, struct sctphdr); dport = SCTP(ulp)->src_port; sport = SCTP(ulp)->dest_port; /* XXXGL: handle stale? */ break; case IPPROTO_UDP: PULLUP_TO(hlen, ulp, struct udphdr); dport = UDP(ulp)->uh_dport; sport = UDP(ulp)->uh_sport; break; case IPPROTO_HOPOPTS: /* RFC 2460 */ PULLUP_TO(hlen, ulp, struct ip6_hbh); hlen += (((struct ip6_hbh *)ulp)->ip6h_len + 1) << 3; proto = ((struct ip6_hbh *)ulp)->ip6h_nxt; ulp = NULL; break; case IPPROTO_ROUTING: /* RFC 2460 */ PULLUP_TO(hlen, ulp, struct ip6_rthdr); hlen += (((struct ip6_rthdr *)ulp)->ip6r_len + 1) << 3; proto = ((struct ip6_rthdr *)ulp)->ip6r_nxt; ulp = NULL; break; case IPPROTO_FRAGMENT: /* RFC 2460 */ PULLUP_TO(hlen, ulp, struct ip6_frag); hlen += sizeof (struct ip6_frag); proto = ((struct ip6_frag *)ulp)->ip6f_nxt; offset = ((struct ip6_frag *)ulp)->ip6f_offlg & IP6F_OFF_MASK; ulp = NULL; break; case IPPROTO_DSTOPTS: /* RFC 2460 */ PULLUP_TO(hlen, ulp, struct ip6_hbh); hlen += (((struct ip6_hbh *)ulp)->ip6h_len + 1) << 3; proto = ((struct ip6_hbh *)ulp)->ip6h_nxt; ulp = NULL; break; case IPPROTO_AH: /* RFC 2402 */ PULLUP_TO(hlen, ulp, struct ip6_ext); hlen += (((struct ip6_ext *)ulp)->ip6e_len + 2) << 2; proto = ((struct ip6_ext *)ulp)->ip6e_nxt; ulp = NULL; break; default: PULLUP_TO(hlen, ulp, struct ip6_ext); break; } } bcopy(&ip6->ip6_dst, &key[0], sizeof(struct in6_addr)); bcopy(&ip6->ip6_src, &key[4], sizeof(struct in6_addr)); key[8] = (dport << 16) | sport; fibnum |= proto << 16; fle = flowtable_lookup_common(&V_ip6_ft, key, 9 * sizeof(uint32_t), fibnum); #else /* !FLOWTABLE_HASH_ALL */ bcopy(&ip6->ip6_dst, &key[0], sizeof(struct in6_addr)); fle = flowtable_lookup_common(&V_ip6_ft, key, sizeof(struct in6_addr), fibnum); #endif /* FLOWTABLE_HASH_ALL */ if (fle == NULL) return (NULL); sin6 = (struct sockaddr_in6 *)&ro->ro_dst; sin6->sin6_family = AF_INET6; sin6->sin6_len = sizeof(*sin6); bcopy(&ip6->ip6_dst, &sin6->sin6_addr, sizeof(struct in6_addr)); return (fle); } #endif /* INET6 */ static bitstr_t * flowtable_mask(struct flowtable *ft) { /* * flowtable_free_stale() calls w/o critical section, but * with sched_bind(). Since pointer is stable throughout * ft lifetime, it is safe, otherwise... * * CRITICAL_ASSERT(curthread); */ return (*(bitstr_t **)zpcpu_get(ft->ft_masks)); } static struct flist * flowtable_list(struct flowtable *ft, uint32_t hash) { CRITICAL_ASSERT(curthread); return (zpcpu_get(ft->ft_table[hash % ft->ft_size])); } static int flow_stale(struct flowtable *ft, struct flentry *fle, int maxidle) { if (((fle->f_rt->rt_flags & RTF_UP) == 0) || (fle->f_rt->rt_ifp == NULL) || !RT_LINK_IS_UP(fle->f_rt->rt_ifp) || (fle->f_lle->la_flags & LLE_VALID) == 0) return (1); if (time_uptime - fle->f_uptime > maxidle) return (1); #ifdef FLOWTABLE_HASH_ALL if (fle->f_flags & FL_STALE) return (1); #endif return (0); } static int flow_full(void) { int count, max; count = uma_zone_get_cur(flow_zone); max = uma_zone_get_max(flow_zone); return (count > (max - (max >> 3))); } static int flow_matches(struct flentry *fle, uint32_t *key, int keylen, uint32_t fibnum) { #ifdef FLOWTABLE_HASH_ALL uint8_t proto; proto = (fibnum >> 16) & 0xff; fibnum &= 0xffff; #endif CRITICAL_ASSERT(curthread); /* Microoptimization for IPv4: don't use bcmp(). */ if (((keylen == sizeof(uint32_t) && (fle->f_key[0] == key[0])) || (bcmp(fle->f_key, key, keylen) == 0)) && fibnum == fle->f_fibnum && #ifdef FLOWTABLE_HASH_ALL proto == fle->f_proto && #endif (fle->f_rt->rt_flags & RTF_UP) && fle->f_rt->rt_ifp != NULL && (fle->f_lle->la_flags & LLE_VALID)) return (1); return (0); } static struct flentry * flowtable_insert(struct flowtable *ft, uint32_t hash, uint32_t *key, int keylen, uint32_t fibnum0) { #ifdef INET6 struct route_in6 sro6; #endif #ifdef INET struct route sro; #endif struct route *ro = NULL; struct rtentry *rt; struct lltable *lt = NULL; struct llentry *lle; struct sockaddr_storage *l3addr; struct ifnet *ifp; struct flist *flist; struct flentry *fle, *iter; bitstr_t *mask; uint16_t fibnum = fibnum0; #ifdef FLOWTABLE_HASH_ALL uint8_t proto; proto = (fibnum0 >> 16) & 0xff; fibnum = fibnum0 & 0xffff; #endif /* * This bit of code ends up locking the * same route 3 times (just like ip_output + ether_output) * - at lookup * - in rt_check when called by arpresolve * - dropping the refcount for the rtentry * * This could be consolidated to one if we wrote a variant * of arpresolve with an rt_check variant that expected to * receive the route locked */ #ifdef INET if (ft == &V_ip4_ft) { struct sockaddr_in *sin; ro = &sro; bzero(&sro.ro_dst, sizeof(sro.ro_dst)); sin = (struct sockaddr_in *)&sro.ro_dst; sin->sin_family = AF_INET; sin->sin_len = sizeof(*sin); sin->sin_addr.s_addr = key[0]; } #endif #ifdef INET6 if (ft == &V_ip6_ft) { struct sockaddr_in6 *sin6; ro = (struct route *)&sro6; sin6 = &sro6.ro_dst; bzero(sin6, sizeof(*sin6)); sin6->sin6_family = AF_INET6; sin6->sin6_len = sizeof(*sin6); bcopy(key, &sin6->sin6_addr, sizeof(struct in6_addr)); } #endif ro->ro_rt = NULL; #ifdef RADIX_MPATH rtalloc_mpath_fib(ro, hash, fibnum); #else rtalloc_ign_fib(ro, 0, fibnum); #endif if (ro->ro_rt == NULL) return (NULL); rt = ro->ro_rt; ifp = rt->rt_ifp; if (ifp->if_flags & (IFF_POINTOPOINT | IFF_LOOPBACK)) { RTFREE(rt); return (NULL); } #ifdef INET if (ft == &V_ip4_ft) lt = LLTABLE(ifp); #endif #ifdef INET6 if (ft == &V_ip6_ft) lt = LLTABLE6(ifp); #endif if (rt->rt_flags & RTF_GATEWAY) l3addr = (struct sockaddr_storage *)rt->rt_gateway; else l3addr = (struct sockaddr_storage *)&ro->ro_dst; lle = llentry_alloc(ifp, lt, l3addr); if (lle == NULL) { RTFREE(rt); return (NULL); } /* Don't insert the entry if the ARP hasn't yet finished resolving. */ if ((lle->la_flags & LLE_VALID) == 0) { RTFREE(rt); LLE_FREE(lle); FLOWSTAT_INC(ft, ft_fail_lle_invalid); return (NULL); } fle = uma_zalloc(flow_zone, M_NOWAIT | M_ZERO); if (fle == NULL) { RTFREE(rt); LLE_FREE(lle); return (NULL); } fle->f_hash = hash; bcopy(key, &fle->f_key, keylen); fle->f_rt = rt; fle->f_lle = lle; fle->f_fibnum = fibnum; fle->f_uptime = time_uptime; #ifdef FLOWTABLE_HASH_ALL fle->f_proto = proto; fle->f_flags = fibnum0 >> 24; #endif critical_enter(); mask = flowtable_mask(ft); flist = flowtable_list(ft, hash); if (SLIST_EMPTY(flist)) { bit_set(mask, (hash % ft->ft_size)); SLIST_INSERT_HEAD(flist, fle, f_next); goto skip; } /* * find end of list and make sure that we were not * preempted by another thread handling this flow */ SLIST_FOREACH(iter, flist, f_next) { KASSERT(iter->f_hash % ft->ft_size == hash % ft->ft_size, ("%s: wrong hash", __func__)); if (flow_matches(iter, key, keylen, fibnum)) { /* * We probably migrated to an other CPU after * lookup in flowtable_lookup_common() failed. * It appeared that this CPU already has flow * entry. */ iter->f_uptime = time_uptime; #ifdef FLOWTABLE_HASH_ALL iter->f_flags |= fibnum >> 24; #endif critical_exit(); FLOWSTAT_INC(ft, ft_collisions); uma_zfree(flow_zone, fle); return (iter); } } SLIST_INSERT_HEAD(flist, fle, f_next); skip: critical_exit(); FLOWSTAT_INC(ft, ft_inserts); return (fle); } int flowtable_lookup(sa_family_t sa, struct mbuf *m, struct route *ro) { struct flentry *fle; struct llentry *lle; if (V_flowtable_enable == 0) return (ENXIO); switch (sa) { #ifdef INET case AF_INET: fle = flowtable_lookup_ipv4(m, ro); break; #endif #ifdef INET6 case AF_INET6: fle = flowtable_lookup_ipv6(m, ro); break; #endif default: panic("%s: sa %d", __func__, sa); } if (fle == NULL) return (EHOSTUNREACH); if (M_HASHTYPE_GET(m) == M_HASHTYPE_NONE) { M_HASHTYPE_SET(m, M_HASHTYPE_OPAQUE_HASH); m->m_pkthdr.flowid = fle->f_hash; } ro->ro_rt = fle->f_rt; ro->ro_flags |= RT_NORTREF; lle = fle->f_lle; if (lle != NULL && (lle->la_flags & LLE_VALID)) ro->ro_lle = lle; /* share ref with fle->f_lle */ return (0); } static struct flentry * flowtable_lookup_common(struct flowtable *ft, uint32_t *key, int keylen, uint32_t fibnum) { struct flist *flist; struct flentry *fle; uint32_t hash; FLOWSTAT_INC(ft, ft_lookups); hash = jenkins_hash32(key, keylen / sizeof(uint32_t), flow_hashjitter); critical_enter(); flist = flowtable_list(ft, hash); SLIST_FOREACH(fle, flist, f_next) { KASSERT(fle->f_hash % ft->ft_size == hash % ft->ft_size, ("%s: wrong hash", __func__)); if (flow_matches(fle, key, keylen, fibnum)) { fle->f_uptime = time_uptime; #ifdef FLOWTABLE_HASH_ALL fle->f_flags |= fibnum >> 24; #endif critical_exit(); FLOWSTAT_INC(ft, ft_hits); return (fle); } } critical_exit(); FLOWSTAT_INC(ft, ft_misses); return (flowtable_insert(ft, hash, key, keylen, fibnum)); } static void flowtable_alloc(struct flowtable *ft) { ft->ft_table = malloc(ft->ft_size * sizeof(struct flist), M_FTABLE, M_WAITOK); for (int i = 0; i < ft->ft_size; i++) ft->ft_table[i] = uma_zalloc(pcpu_zone_ptr, M_WAITOK | M_ZERO); ft->ft_masks = uma_zalloc(pcpu_zone_ptr, M_WAITOK); - for (int i = 0; i < mp_ncpus; i++) { + CPU_FOREACH(i) { bitstr_t **b; b = zpcpu_get_cpu(ft->ft_masks, i); *b = bit_alloc(ft->ft_size, M_FTABLE, M_WAITOK); } ft->ft_tmpmask = bit_alloc(ft->ft_size, M_FTABLE, M_WAITOK); } static void flowtable_free_stale(struct flowtable *ft, struct rtentry *rt, int maxidle) { struct flist *flist, freelist; struct flentry *fle, *fle1, *fleprev; bitstr_t *mask, *tmpmask; int curbit, tmpsize; SLIST_INIT(&freelist); mask = flowtable_mask(ft); tmpmask = ft->ft_tmpmask; tmpsize = ft->ft_size; memcpy(tmpmask, mask, ft->ft_size/8); curbit = 0; fleprev = NULL; /* pacify gcc */ /* * XXX Note to self, bit_ffs operates at the byte level * and thus adds gratuitous overhead */ bit_ffs(tmpmask, ft->ft_size, &curbit); while (curbit != -1) { if (curbit >= ft->ft_size || curbit < -1) { log(LOG_ALERT, "warning: bad curbit value %d \n", curbit); break; } FLOWSTAT_INC(ft, ft_free_checks); critical_enter(); flist = flowtable_list(ft, curbit); #ifdef DIAGNOSTIC if (SLIST_EMPTY(flist) && curbit > 0) { log(LOG_ALERT, "warning bit=%d set, but no fle found\n", curbit); } #endif SLIST_FOREACH_SAFE(fle, flist, f_next, fle1) { if (rt != NULL && fle->f_rt != rt) { fleprev = fle; continue; } if (!flow_stale(ft, fle, maxidle)) { fleprev = fle; continue; } if (fle == SLIST_FIRST(flist)) SLIST_REMOVE_HEAD(flist, f_next); else SLIST_REMOVE_AFTER(fleprev, f_next); SLIST_INSERT_HEAD(&freelist, fle, f_next); } if (SLIST_EMPTY(flist)) bit_clear(mask, curbit); critical_exit(); bit_clear(tmpmask, curbit); bit_ffs(tmpmask, tmpsize, &curbit); } SLIST_FOREACH_SAFE(fle, &freelist, f_next, fle1) { FLOWSTAT_INC(ft, ft_frees); if (fle->f_rt != NULL) RTFREE(fle->f_rt); if (fle->f_lle != NULL) LLE_FREE(fle->f_lle); uma_zfree(flow_zone, fle); } } static void flowtable_clean_vnet(struct flowtable *ft, struct rtentry *rt, int maxidle) { int i; CPU_FOREACH(i) { if (smp_started == 1) { thread_lock(curthread); sched_bind(curthread, i); thread_unlock(curthread); } flowtable_free_stale(ft, rt, maxidle); if (smp_started == 1) { thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); } } } void flowtable_route_flush(sa_family_t sa, struct rtentry *rt) { struct flowtable *ft; switch (sa) { #ifdef INET case AF_INET: ft = &V_ip4_ft; break; #endif #ifdef INET6 case AF_INET6: ft = &V_ip6_ft; break; #endif default: panic("%s: sa %d", __func__, sa); } flowtable_clean_vnet(ft, rt, 0); } static void flowtable_cleaner(void) { VNET_ITERATOR_DECL(vnet_iter); struct thread *td; if (bootverbose) log(LOG_INFO, "flowtable cleaner started\n"); td = curthread; while (1) { uint32_t flowclean_freq, maxidle; /* * The maximum idle time, as well as frequency are arbitrary. */ if (flow_full()) maxidle = 5; else maxidle = 30; VNET_LIST_RLOCK(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); #ifdef INET flowtable_clean_vnet(&V_ip4_ft, NULL, maxidle); #endif #ifdef INET6 flowtable_clean_vnet(&V_ip6_ft, NULL, maxidle); #endif CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); if (flow_full()) flowclean_freq = 4*hz; else flowclean_freq = 20*hz; mtx_lock(&flowclean_lock); thread_lock(td); sched_prio(td, PPAUSE); thread_unlock(td); flowclean_cycles++; cv_broadcast(&flowclean_f_cv); cv_timedwait(&flowclean_c_cv, &flowclean_lock, flowclean_freq); mtx_unlock(&flowclean_lock); } } static void flowtable_flush(void *unused __unused) { uint64_t start; mtx_lock(&flowclean_lock); start = flowclean_cycles; while (start == flowclean_cycles) { cv_broadcast(&flowclean_c_cv); cv_wait(&flowclean_f_cv, &flowclean_lock); } mtx_unlock(&flowclean_lock); } static struct kproc_desc flow_kp = { "flowcleaner", flowtable_cleaner, &flowcleanerproc }; SYSINIT(flowcleaner, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, kproc_start, &flow_kp); static int flowtable_get_size(char *name) { int size; if (TUNABLE_INT_FETCH(name, &size)) { if (size < 256) size = 256; if (!powerof2(size)) { printf("%s must be power of 2\n", name); size = 2048; } } else { /* * round up to the next power of 2 */ size = 1 << fls((1024 + maxusers * 64) - 1); } return (size); } static void flowtable_init(const void *unused __unused) { flow_hashjitter = arc4random(); flow_zone = uma_zcreate("flows", sizeof(struct flentry), NULL, NULL, NULL, NULL, (64-1), UMA_ZONE_MAXBUCKET); uma_zone_set_max(flow_zone, 1024 + maxusers * 64 * mp_ncpus); cv_init(&flowclean_c_cv, "c_flowcleanwait"); cv_init(&flowclean_f_cv, "f_flowcleanwait"); mtx_init(&flowclean_lock, "flowclean lock", NULL, MTX_DEF); EVENTHANDLER_REGISTER(ifnet_departure_event, flowtable_flush, NULL, EVENTHANDLER_PRI_ANY); } SYSINIT(flowtable_init, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, flowtable_init, NULL); #ifdef INET static SYSCTL_NODE(_net_flowtable, OID_AUTO, ip4, CTLFLAG_RD, NULL, "Flowtable for IPv4"); static VNET_PCPUSTAT_DEFINE(struct flowtable_stat, ip4_ftstat); VNET_PCPUSTAT_SYSINIT(ip4_ftstat); VNET_PCPUSTAT_SYSUNINIT(ip4_ftstat); SYSCTL_VNET_PCPUSTAT(_net_flowtable_ip4, OID_AUTO, stat, struct flowtable_stat, ip4_ftstat, "Flowtable statistics for IPv4 " "(struct flowtable_stat, net/flowtable.h)"); static void flowtable_init_vnet_v4(const void *unused __unused) { V_ip4_ft.ft_size = flowtable_get_size("net.flowtable.ip4.size"); V_ip4_ft.ft_stat = VNET(ip4_ftstat); flowtable_alloc(&V_ip4_ft); } VNET_SYSINIT(ft_vnet_v4, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, flowtable_init_vnet_v4, NULL); #endif /* INET */ #ifdef INET6 static SYSCTL_NODE(_net_flowtable, OID_AUTO, ip6, CTLFLAG_RD, NULL, "Flowtable for IPv6"); static VNET_PCPUSTAT_DEFINE(struct flowtable_stat, ip6_ftstat); VNET_PCPUSTAT_SYSINIT(ip6_ftstat); VNET_PCPUSTAT_SYSUNINIT(ip6_ftstat); SYSCTL_VNET_PCPUSTAT(_net_flowtable_ip6, OID_AUTO, stat, struct flowtable_stat, ip6_ftstat, "Flowtable statistics for IPv6 " "(struct flowtable_stat, net/flowtable.h)"); static void flowtable_init_vnet_v6(const void *unused __unused) { V_ip6_ft.ft_size = flowtable_get_size("net.flowtable.ip6.size"); V_ip6_ft.ft_stat = VNET(ip6_ftstat); flowtable_alloc(&V_ip6_ft); } VNET_SYSINIT(flowtable_init_vnet_v6, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, flowtable_init_vnet_v6, NULL); #endif /* INET6 */ #ifdef DDB static bitstr_t * flowtable_mask_pcpu(struct flowtable *ft, int cpuid) { return (zpcpu_get_cpu(*ft->ft_masks, cpuid)); } static struct flist * flowtable_list_pcpu(struct flowtable *ft, uint32_t hash, int cpuid) { return (zpcpu_get_cpu(&ft->ft_table[hash % ft->ft_size], cpuid)); } static void flow_show(struct flowtable *ft, struct flentry *fle) { int idle_time; int rt_valid, ifp_valid; volatile struct rtentry *rt; struct ifnet *ifp = NULL; uint32_t *hashkey = fle->f_key; idle_time = (int)(time_uptime - fle->f_uptime); rt = fle->f_rt; rt_valid = rt != NULL; if (rt_valid) ifp = rt->rt_ifp; ifp_valid = ifp != NULL; #ifdef INET if (ft == &V_ip4_ft) { char daddr[4*sizeof "123"]; #ifdef FLOWTABLE_HASH_ALL char saddr[4*sizeof "123"]; uint16_t sport, dport; #endif inet_ntoa_r(*(struct in_addr *) &hashkey[0], daddr); #ifdef FLOWTABLE_HASH_ALL inet_ntoa_r(*(struct in_addr *) &hashkey[1], saddr); dport = ntohs((uint16_t)(hashkey[2] >> 16)); sport = ntohs((uint16_t)(hashkey[2] & 0xffff)); db_printf("%s:%d->%s:%d", saddr, sport, daddr, dport); #else db_printf("%s ", daddr); #endif } #endif /* INET */ #ifdef INET6 if (ft == &V_ip6_ft) { #ifdef FLOWTABLE_HASH_ALL db_printf("\n\tkey=%08x:%08x:%08x%08x:%08x:%08x%08x:%08x:%08x", hashkey[0], hashkey[1], hashkey[2], hashkey[3], hashkey[4], hashkey[5], hashkey[6], hashkey[7], hashkey[8]); #else db_printf("\n\tkey=%08x:%08x:%08x ", hashkey[0], hashkey[1], hashkey[2]); #endif } #endif /* INET6 */ db_printf("hash=%08x idle_time=%03d" "\n\tfibnum=%02d rt=%p", fle->f_hash, idle_time, fle->f_fibnum, fle->f_rt); #ifdef FLOWTABLE_HASH_ALL if (fle->f_flags & FL_STALE) db_printf(" FL_STALE "); #endif if (rt_valid) { if (rt->rt_flags & RTF_UP) db_printf(" RTF_UP "); } if (ifp_valid) { if (ifp->if_flags & IFF_LOOPBACK) db_printf(" IFF_LOOPBACK "); if (ifp->if_flags & IFF_UP) db_printf(" IFF_UP "); if (ifp->if_flags & IFF_POINTOPOINT) db_printf(" IFF_POINTOPOINT "); } db_printf("\n"); } static void flowtable_show(struct flowtable *ft, int cpuid) { int curbit = 0; bitstr_t *mask, *tmpmask; if (cpuid != -1) db_printf("cpu: %d\n", cpuid); mask = flowtable_mask_pcpu(ft, cpuid); tmpmask = ft->ft_tmpmask; memcpy(tmpmask, mask, ft->ft_size/8); /* * XXX Note to self, bit_ffs operates at the byte level * and thus adds gratuitous overhead */ bit_ffs(tmpmask, ft->ft_size, &curbit); while (curbit != -1) { struct flist *flist; struct flentry *fle; if (curbit >= ft->ft_size || curbit < -1) { db_printf("warning: bad curbit value %d \n", curbit); break; } flist = flowtable_list_pcpu(ft, curbit, cpuid); SLIST_FOREACH(fle, flist, f_next) flow_show(ft, fle); bit_clear(tmpmask, curbit); bit_ffs(tmpmask, ft->ft_size, &curbit); } } static void flowtable_show_vnet(struct flowtable *ft) { int i; CPU_FOREACH(i) flowtable_show(ft, i); } DB_SHOW_COMMAND(flowtables, db_show_flowtables) { VNET_ITERATOR_DECL(vnet_iter); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); #ifdef VIMAGE db_printf("vnet %p\n", vnet_iter); #endif #ifdef INET printf("IPv4:\n"); flowtable_show_vnet(&V_ip4_ft); #endif #ifdef INET6 printf("IPv6:\n"); flowtable_show_vnet(&V_ip6_ft); #endif CURVNET_RESTORE(); } } #endif Index: head/sys/net/iflib.c =================================================================== --- head/sys/net/iflib.c (revision 302371) +++ head/sys/net/iflib.c (revision 302372) @@ -1,4806 +1,4806 @@ /*- * Copyright (c) 2014-2016, Matthew Macy * 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. Neither the name of Matthew Macy 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 COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_acpi.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "ifdi_if.h" #if defined(__i386__) || defined(__amd64__) #include #include #include #include #include #include #endif /* * enable accounting of every mbuf as it comes in to and goes out of iflib's software descriptor references */ #define MEMORY_LOGGING 0 /* * Enable mbuf vectors for compressing long mbuf chains */ /* * NB: * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead * we prefetch needs to be determined by the time spent in m_free vis a vis * the cost of a prefetch. This will of course vary based on the workload: * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which * is quite expensive, thus suggesting very little prefetch. * - small packet forwarding which is just returning a single mbuf to * UMA will typically be very fast vis a vis the cost of a memory * access. */ /* * File organization: * - private structures * - iflib private utility functions * - ifnet functions * - vlan registry and other exported functions * - iflib public core functions * * */ static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library"); struct iflib_txq; typedef struct iflib_txq *iflib_txq_t; struct iflib_rxq; typedef struct iflib_rxq *iflib_rxq_t; struct iflib_fl; typedef struct iflib_fl *iflib_fl_t; typedef struct iflib_filter_info { driver_filter_t *ifi_filter; void *ifi_filter_arg; struct grouptask *ifi_task; } *iflib_filter_info_t; struct iflib_ctx { KOBJ_FIELDS; /* * Pointer to hardware driver's softc */ void *ifc_softc; device_t ifc_dev; if_t ifc_ifp; cpuset_t ifc_cpus; if_shared_ctx_t ifc_sctx; struct if_softc_ctx ifc_softc_ctx; struct mtx ifc_mtx; uint16_t ifc_nhwtxqs; uint16_t ifc_nhwrxqs; iflib_txq_t ifc_txqs; iflib_rxq_t ifc_rxqs; uint32_t ifc_if_flags; uint32_t ifc_flags; uint32_t ifc_max_fl_buf_size; int ifc_in_detach; int ifc_link_state; int ifc_link_irq; int ifc_pause_frames; int ifc_watchdog_events; struct cdev *ifc_led_dev; struct resource *ifc_msix_mem; struct if_irq ifc_legacy_irq; struct grouptask ifc_admin_task; struct grouptask ifc_vflr_task; struct iflib_filter_info ifc_filter_info; struct ifmedia ifc_media; struct sysctl_oid *ifc_sysctl_node; uint16_t ifc_sysctl_ntxqs; uint16_t ifc_sysctl_nrxqs; uint16_t ifc_sysctl_ntxds; uint16_t ifc_sysctl_nrxds; struct if_txrx ifc_txrx; #define isc_txd_encap ifc_txrx.ift_txd_encap #define isc_txd_flush ifc_txrx.ift_txd_flush #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update #define isc_rxd_available ifc_txrx.ift_rxd_available #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_rxd_flush ifc_txrx.ift_rxd_flush #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_legacy_intr ifc_txrx.ift_legacy_intr eventhandler_tag ifc_vlan_attach_event; eventhandler_tag ifc_vlan_detach_event; uint8_t ifc_mac[ETHER_ADDR_LEN]; char ifc_mtx_name[16]; }; void * iflib_get_softc(if_ctx_t ctx) { return (ctx->ifc_softc); } device_t iflib_get_dev(if_ctx_t ctx) { return (ctx->ifc_dev); } if_t iflib_get_ifp(if_ctx_t ctx) { return (ctx->ifc_ifp); } struct ifmedia * iflib_get_media(if_ctx_t ctx) { return (&ctx->ifc_media); } void iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN]) { bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN); } if_softc_ctx_t iflib_get_softc_ctx(if_ctx_t ctx) { return (&ctx->ifc_softc_ctx); } if_shared_ctx_t iflib_get_sctx(if_ctx_t ctx) { return (ctx->ifc_sctx); } #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*)) #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP) #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF) #define RX_SW_DESC_MAP_CREATED (1 << 0) #define TX_SW_DESC_MAP_CREATED (1 << 1) #define RX_SW_DESC_INUSE (1 << 3) #define TX_SW_DESC_MAPPED (1 << 4) typedef struct iflib_sw_rx_desc { bus_dmamap_t ifsd_map; /* bus_dma map for packet */ struct mbuf *ifsd_m; /* rx: uninitialized mbuf */ caddr_t ifsd_cl; /* direct cluster pointer for rx */ uint16_t ifsd_flags; } *iflib_rxsd_t; typedef struct iflib_sw_tx_desc_val { bus_dmamap_t ifsd_map; /* bus_dma map for packet */ struct mbuf *ifsd_m; /* pkthdr mbuf */ uint8_t ifsd_flags; } *iflib_txsd_val_t; typedef struct iflib_sw_tx_desc_array { bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ struct mbuf **ifsd_m; /* pkthdr mbufs */ uint8_t *ifsd_flags; } iflib_txsd_array_t; /* magic number that should be high enough for any hardware */ #define IFLIB_MAX_TX_SEGS 128 #define IFLIB_MAX_RX_SEGS 32 #define IFLIB_RX_COPY_THRESH 128 #define IFLIB_MAX_RX_REFRESH 32 #define IFLIB_QUEUE_IDLE 0 #define IFLIB_QUEUE_HUNG 1 #define IFLIB_QUEUE_WORKING 2 /* this should really scale with ring size - 32 is a fairly arbitrary value for this */ #define TX_BATCH_SIZE 16 #define IFLIB_RESTART_BUDGET 8 #define IFC_LEGACY 0x1 #define IFC_QFLUSH 0x2 #define IFC_MULTISEG 0x4 #define IFC_DMAR 0x8 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \ CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \ CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP) struct iflib_txq { uint16_t ift_in_use; uint16_t ift_cidx; uint16_t ift_cidx_processed; uint16_t ift_pidx; uint8_t ift_gen; uint8_t ift_db_pending; uint8_t ift_db_pending_queued; uint8_t ift_npending; /* implicit pad */ uint64_t ift_processed; uint64_t ift_cleaned; #if MEMORY_LOGGING uint64_t ift_enqueued; uint64_t ift_dequeued; #endif uint64_t ift_no_tx_dma_setup; uint64_t ift_no_desc_avail; uint64_t ift_mbuf_defrag_failed; uint64_t ift_mbuf_defrag; uint64_t ift_map_failed; uint64_t ift_txd_encap_efbig; uint64_t ift_pullups; struct mtx ift_mtx; struct mtx ift_db_mtx; /* constant values */ if_ctx_t ift_ctx; struct ifmp_ring **ift_br; struct grouptask ift_task; uint16_t ift_size; uint16_t ift_id; struct callout ift_timer; struct callout ift_db_check; iflib_txsd_array_t ift_sds; uint8_t ift_nbr; uint8_t ift_qstatus; uint8_t ift_active; uint8_t ift_closed; int ift_watchdog_time; struct iflib_filter_info ift_filter_info; bus_dma_tag_t ift_desc_tag; bus_dma_tag_t ift_tso_desc_tag; iflib_dma_info_t ift_ifdi; #define MTX_NAME_LEN 16 char ift_mtx_name[MTX_NAME_LEN]; char ift_db_mtx_name[MTX_NAME_LEN]; bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE); } __aligned(CACHE_LINE_SIZE); struct iflib_fl { uint16_t ifl_cidx; uint16_t ifl_pidx; uint16_t ifl_credits; uint8_t ifl_gen; #if MEMORY_LOGGING uint64_t ifl_m_enqueued; uint64_t ifl_m_dequeued; uint64_t ifl_cl_enqueued; uint64_t ifl_cl_dequeued; #endif /* implicit pad */ /* constant */ uint16_t ifl_size; uint16_t ifl_buf_size; uint16_t ifl_cltype; uma_zone_t ifl_zone; iflib_rxsd_t ifl_sds; iflib_rxq_t ifl_rxq; uint8_t ifl_id; bus_dma_tag_t ifl_desc_tag; iflib_dma_info_t ifl_ifdi; uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE); caddr_t ifl_vm_addrs[IFLIB_MAX_RX_REFRESH]; } __aligned(CACHE_LINE_SIZE); static inline int get_inuse(int size, int cidx, int pidx, int gen) { int used; if (pidx > cidx) used = pidx - cidx; else if (pidx < cidx) used = size - cidx + pidx; else if (gen == 0 && pidx == cidx) used = 0; else if (gen == 1 && pidx == cidx) used = size; else panic("bad state"); return (used); } #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen)) #define IDXDIFF(head, tail, wrap) \ ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head)) struct iflib_rxq { /* If there is a separate completion queue - * these are the cq cidx and pidx. Otherwise * these are unused. */ uint16_t ifr_size; uint16_t ifr_cq_cidx; uint16_t ifr_cq_pidx; uint8_t ifr_cq_gen; if_ctx_t ifr_ctx; iflib_fl_t ifr_fl; uint64_t ifr_rx_irq; uint16_t ifr_id; uint8_t ifr_lro_enabled; uint8_t ifr_nfl; struct lro_ctrl ifr_lc; struct grouptask ifr_task; struct iflib_filter_info ifr_filter_info; iflib_dma_info_t ifr_ifdi; /* dynamically allocate if any drivers need a value substantially larger than this */ struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE); } __aligned(CACHE_LINE_SIZE); /* * Only allow a single packet to take up most 1/nth of the tx ring */ #define MAX_SINGLE_PACKET_FRACTION 12 #define IF_BAD_DMA (bus_addr_t)-1 static int enable_msix = 1; #define mtx_held(m) (((m)->mtx_lock & ~MTX_FLAGMASK) != (uintptr_t)0) #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING)) #define CTX_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF) #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx) #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx) #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx) #define TXDB_LOCK_INIT(txq) mtx_init(&(txq)->ift_db_mtx, (txq)->ift_db_mtx_name, NULL, MTX_DEF) #define TXDB_TRYLOCK(txq) mtx_trylock(&(txq)->ift_db_mtx) #define TXDB_LOCK(txq) mtx_lock(&(txq)->ift_db_mtx) #define TXDB_UNLOCK(txq) mtx_unlock(&(txq)->ift_db_mtx) #define TXDB_LOCK_DESTROY(txq) mtx_destroy(&(txq)->ift_db_mtx) #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx) #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx) /* Our boot-time initialization hook */ static int iflib_module_event_handler(module_t, int, void *); static moduledata_t iflib_moduledata = { "iflib", iflib_module_event_handler, NULL }; DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY); MODULE_VERSION(iflib, 1); MODULE_DEPEND(iflib, pci, 1, 1, 1); MODULE_DEPEND(iflib, ether, 1, 1, 1); TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1); TASKQGROUP_DEFINE(if_config_tqg, 1, 1); #ifndef IFLIB_DEBUG_COUNTERS #ifdef INVARIANTS #define IFLIB_DEBUG_COUNTERS 1 #else #define IFLIB_DEBUG_COUNTERS 0 #endif /* !INVARIANTS */ #endif static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0, "iflib driver parameters"); /* * XXX need to ensure that this can't accidentally cause the head to be moved backwards */ static int iflib_min_tx_latency = 0; SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW, &iflib_min_tx_latency, 0, "minimize transmit latency at the possibel expense of throughput"); #if IFLIB_DEBUG_COUNTERS static int iflib_tx_seen; static int iflib_tx_sent; static int iflib_tx_encap; static int iflib_rx_allocs; static int iflib_fl_refills; static int iflib_fl_refills_large; static int iflib_tx_frees; SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0, "# tx mbufs seen"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0, "# tx mbufs sent"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0, "# tx mbufs encapped"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0, "# tx frees"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0, "# rx allocations"); SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0, "# refills"); SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD, &iflib_fl_refills_large, 0, "# large refills"); static int iflib_txq_drain_flushing; static int iflib_txq_drain_oactive; static int iflib_txq_drain_notready; static int iflib_txq_drain_encapfail; SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD, &iflib_txq_drain_flushing, 0, "# drain flushes"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD, &iflib_txq_drain_oactive, 0, "# drain oactives"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD, &iflib_txq_drain_notready, 0, "# drain notready"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD, &iflib_txq_drain_encapfail, 0, "# drain encap fails"); static int iflib_encap_load_mbuf_fail; static int iflib_encap_txq_avail_fail; static int iflib_encap_txd_encap_fail; SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD, &iflib_encap_load_mbuf_fail, 0, "# busdma load failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD, &iflib_encap_txq_avail_fail, 0, "# txq avail failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD, &iflib_encap_txd_encap_fail, 0, "# driver encap failures"); static int iflib_task_fn_rxs; static int iflib_rx_intr_enables; static int iflib_fast_intrs; static int iflib_intr_link; static int iflib_intr_msix; static int iflib_rx_unavail; static int iflib_rx_ctx_inactive; static int iflib_rx_zero_len; static int iflib_rx_if_input; static int iflib_rx_mbuf_null; static int iflib_rxd_flush; static int iflib_verbose_debug; SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD, &iflib_intr_link, 0, "# intr link calls"); SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD, &iflib_intr_msix, 0, "# intr msix calls"); SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0, "# task_fn_rx calls"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD, &iflib_rx_intr_enables, 0, "# rx intr enables"); SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0, "# fast_intr calls"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0, "# times rxeof called with no available data"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD, &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD, &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input, 0, "# times rxeof called if_input"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD, &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf"); SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0, "# times rxd_flush called"); SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW, &iflib_verbose_debug, 0, "enable verbose debugging"); #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1) #else #define DBG_COUNTER_INC(name) #endif #define IFLIB_DEBUG 0 static void iflib_tx_structures_free(if_ctx_t ctx); static void iflib_rx_structures_free(if_ctx_t ctx); static int iflib_queues_alloc(if_ctx_t ctx); static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq); static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx); static int iflib_qset_structures_setup(if_ctx_t ctx); static int iflib_msix_init(if_ctx_t ctx); static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str); static void iflib_txq_check_drain(iflib_txq_t txq, int budget); static uint32_t iflib_txq_can_drain(struct ifmp_ring *); static int iflib_register(if_ctx_t); static void iflib_init_locked(if_ctx_t ctx); static void iflib_add_device_sysctl_pre(if_ctx_t ctx); static void iflib_add_device_sysctl_post(if_ctx_t ctx); #ifdef DEV_NETMAP #include #include #include MODULE_DEPEND(iflib, netmap, 1, 1, 1); /* * device-specific sysctl variables: * * ixl_crcstrip: 0: keep CRC in rx frames (default), 1: strip it. * During regular operations the CRC is stripped, but on some * hardware reception of frames not multiple of 64 is slower, * so using crcstrip=0 helps in benchmarks. * * ixl_rx_miss, ixl_rx_miss_bufs: * count packets that might be missed due to lost interrupts. */ SYSCTL_DECL(_dev_netmap); /* * The xl driver by default strips CRCs and we do not override it. */ int iflib_crcstrip = 1; SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip, CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames"); int iflib_rx_miss, iflib_rx_miss_bufs; SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss, CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr"); SYSCTL_INT(_dev_netmap, OID_AUTO, ixl_rx_miss_bufs, CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs"); /* * Register/unregister. We are already under netmap lock. * Only called on the first register or the last unregister. */ static int iflib_netmap_register(struct netmap_adapter *na, int onoff) { struct ifnet *ifp = na->ifp; if_ctx_t ctx = ifp->if_softc; CTX_LOCK(ctx); IFDI_INTR_DISABLE(ctx); /* Tell the stack that the interface is no longer active */ ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE); if (!CTX_IS_VF(ctx)) IFDI_CRCSTRIP_SET(ctx, onoff); /* enable or disable flags and callbacks in na and ifp */ if (onoff) { nm_set_native_flags(na); } else { nm_clear_native_flags(na); } IFDI_INIT(ctx); IFDI_CRCSTRIP_SET(ctx, onoff); // XXX why twice ? CTX_UNLOCK(ctx); return (ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1); } /* * Reconcile kernel and user view of the transmit ring. * * All information is in the kring. * Userspace wants to send packets up to the one before kring->rhead, * kernel knows kring->nr_hwcur is the first unsent packet. * * Here we push packets out (as many as possible), and possibly * reclaim buffers from previously completed transmission. * * The caller (netmap) guarantees that there is only one instance * running at any time. Any interference with other driver * methods should be handled by the individual drivers. */ static int iflib_netmap_txsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct ifnet *ifp = na->ifp; struct netmap_ring *ring = kring->ring; u_int nm_i; /* index into the netmap ring */ u_int nic_i; /* index into the NIC ring */ u_int n; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; struct if_pkt_info pi; /* * interrupts on every tx packet are expensive so request * them every half ring, or where NS_REPORT is set */ u_int report_frequency = kring->nkr_num_slots >> 1; /* device-specific */ if_ctx_t ctx = ifp->if_softc; iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id]; pi.ipi_segs = txq->ift_segs; pi.ipi_qsidx = kring->ring_id; pi.ipi_ndescs = 0; bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); /* * First part: process new packets to send. * nm_i is the current index in the netmap ring, * nic_i is the corresponding index in the NIC ring. * * If we have packets to send (nm_i != head) * iterate over the netmap ring, fetch length and update * the corresponding slot in the NIC ring. Some drivers also * need to update the buffer's physical address in the NIC slot * even NS_BUF_CHANGED is not set (PNMB computes the addresses). * * The netmap_reload_map() calls is especially expensive, * even when (as in this case) the tag is 0, so do only * when the buffer has actually changed. * * If possible do not set the report/intr bit on all slots, * but only a few times per ring or when NS_REPORT is set. * * Finally, on 10G and faster drivers, it might be useful * to prefetch the next slot and txr entry. */ nm_i = kring->nr_hwcur; if (nm_i != head) { /* we have new packets to send */ nic_i = netmap_idx_k2n(kring, nm_i); __builtin_prefetch(&ring->slot[nm_i]); __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]); __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]); for (n = 0; nm_i != head; n++) { struct netmap_slot *slot = &ring->slot[nm_i]; u_int len = slot->len; uint64_t paddr; void *addr = PNMB(na, slot, &paddr); int flags = (slot->flags & NS_REPORT || nic_i == 0 || nic_i == report_frequency) ? IPI_TX_INTR : 0; /* device-specific */ pi.ipi_pidx = nic_i; pi.ipi_flags = flags; /* Fill the slot in the NIC ring. */ ctx->isc_txd_encap(ctx->ifc_softc, &pi); /* prefetch for next round */ __builtin_prefetch(&ring->slot[nm_i + 1]); __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]); __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]); NM_CHECK_ADDR_LEN(na, addr, len); if (slot->flags & NS_BUF_CHANGED) { /* buffer has changed, reload map */ netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr); } slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED); /* make sure changes to the buffer are synced */ bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i], BUS_DMASYNC_PREWRITE); nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } kring->nr_hwcur = head; /* synchronize the NIC ring */ bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); /* (re)start the tx unit up to slot nic_i (excluded) */ ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i); } /* * Second part: reclaim buffers for completed transmissions. */ if (iflib_tx_credits_update(ctx, txq)) { /* some tx completed, increment avail */ nic_i = txq->ift_cidx_processed; kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim); } return (0); } /* * Reconcile kernel and user view of the receive ring. * Same as for the txsync, this routine must be efficient. * The caller guarantees a single invocations, but races against * the rest of the driver should be handled here. * * On call, kring->rhead is the first packet that userspace wants * to keep, and kring->rcur is the wakeup point. * The kernel has previously reported packets up to kring->rtail. * * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective * of whether or not we received an interrupt. */ static int iflib_netmap_rxsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct ifnet *ifp = na->ifp; struct netmap_ring *ring = kring->ring; u_int nm_i; /* index into the netmap ring */ u_int nic_i; /* index into the NIC ring */ u_int i, n; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR; struct if_rxd_info ri; /* device-specific */ if_ctx_t ctx = ifp->if_softc; iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id]; iflib_fl_t fl = rxq->ifr_fl; if (head > lim) return netmap_ring_reinit(kring); bzero(&ri, sizeof(ri)); ri.iri_qsidx = kring->ring_id; ri.iri_ifp = ctx->ifc_ifp; /* XXX check sync modes */ for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); /* * First part: import newly received packets. * * nm_i is the index of the next free slot in the netmap ring, * nic_i is the index of the next received packet in the NIC ring, * and they may differ in case if_init() has been called while * in netmap mode. For the receive ring we have * * nic_i = rxr->next_check; * nm_i = kring->nr_hwtail (previous) * and * nm_i == (nic_i + kring->nkr_hwofs) % ring_size * * rxr->next_check is set to 0 on a ring reinit */ if (netmap_no_pendintr || force_update) { int crclen = iflib_crcstrip ? 0 : 4; int error, avail; uint16_t slot_flags = kring->nkr_slot_flags; for (fl = rxq->ifr_fl, i = 0; i < rxq->ifr_nfl; i++, fl++) { nic_i = fl->ifl_cidx; nm_i = netmap_idx_n2k(kring, nic_i); avail = ctx->isc_rxd_available(ctx->ifc_softc, kring->ring_id, nic_i); for (n = 0; avail > 0; n++, avail--) { error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); if (error) ring->slot[nm_i].len = 0; else ring->slot[nm_i].len = ri.iri_len - crclen; ring->slot[nm_i].flags = slot_flags; bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, BUS_DMASYNC_POSTREAD); nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } if (n) { /* update the state variables */ if (netmap_no_pendintr && !force_update) { /* diagnostics */ iflib_rx_miss ++; iflib_rx_miss_bufs += n; } fl->ifl_cidx = nic_i; kring->nr_hwtail = nm_i; } kring->nr_kflags &= ~NKR_PENDINTR; } } /* * Second part: skip past packets that userspace has released. * (kring->nr_hwcur to head excluded), * and make the buffers available for reception. * As usual nm_i is the index in the netmap ring, * nic_i is the index in the NIC ring, and * nm_i == (nic_i + kring->nkr_hwofs) % ring_size */ /* XXX not sure how this will work with multiple free lists */ nm_i = kring->nr_hwcur; if (nm_i != head) { nic_i = netmap_idx_k2n(kring, nm_i); for (n = 0; nm_i != head; n++) { struct netmap_slot *slot = &ring->slot[nm_i]; uint64_t paddr; caddr_t vaddr; void *addr = PNMB(na, slot, &paddr); if (addr == NETMAP_BUF_BASE(na)) /* bad buf */ goto ring_reset; vaddr = addr; if (slot->flags & NS_BUF_CHANGED) { /* buffer has changed, reload map */ netmap_reload_map(na, fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, addr); slot->flags &= ~NS_BUF_CHANGED; } /* * XXX we should be batching this operation - TODO */ ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i, &paddr, &vaddr, 1); bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_sds[nic_i].ifsd_map, BUS_DMASYNC_PREREAD); nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } kring->nr_hwcur = head; bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); /* * IMPORTANT: we must leave one free slot in the ring, * so move nic_i back by one unit */ nic_i = nm_prev(nic_i, lim); ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i); } return 0; ring_reset: return netmap_ring_reinit(kring); } static int iflib_netmap_attach(if_ctx_t ctx) { struct netmap_adapter na; bzero(&na, sizeof(na)); na.ifp = ctx->ifc_ifp; na.na_flags = NAF_BDG_MAYSLEEP; MPASS(ctx->ifc_softc_ctx.isc_ntxqsets); MPASS(ctx->ifc_softc_ctx.isc_nrxqsets); na.num_tx_desc = ctx->ifc_sctx->isc_ntxd; na.num_rx_desc = ctx->ifc_sctx->isc_ntxd; na.nm_txsync = iflib_netmap_txsync; na.nm_rxsync = iflib_netmap_rxsync; na.nm_register = iflib_netmap_register; na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; return (netmap_attach(&na)); } static void iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq) { struct netmap_adapter *na = NA(ctx->ifc_ifp); struct netmap_slot *slot; slot = netmap_reset(na, NR_TX, txq->ift_id, 0); if (slot == 0) return; for (int i = 0; i < ctx->ifc_sctx->isc_ntxd; i++) { /* * In netmap mode, set the map for the packet buffer. * NOTE: Some drivers (not this one) also need to set * the physical buffer address in the NIC ring. * netmap_idx_n2k() maps a nic index, i, into the corresponding * netmap slot index, si */ int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i); netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si)); } } static void iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq) { struct netmap_adapter *na = NA(ctx->ifc_ifp); struct netmap_slot *slot; iflib_rxsd_t sd; int nrxd; slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0); if (slot == 0) return; sd = rxq->ifr_fl[0].ifl_sds; nrxd = ctx->ifc_sctx->isc_nrxd; for (int i = 0; i < nrxd; i++, sd++) { int sj = netmap_idx_n2k(&na->rx_rings[rxq->ifr_id], i); uint64_t paddr; void *addr; caddr_t vaddr; vaddr = addr = PNMB(na, slot + sj, &paddr); netmap_load_map(na, rxq->ifr_fl[0].ifl_ifdi->idi_tag, sd->ifsd_map, addr); /* Update descriptor and the cached value */ ctx->isc_rxd_refill(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, i, &paddr, &vaddr, 1); } /* preserve queue */ if (ctx->ifc_ifp->if_capenable & IFCAP_NETMAP) { struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id]; int t = na->num_rx_desc - 1 - nm_kr_rxspace(kring); ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, t); } else ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, 0 /* fl_id */, nrxd-1); } #define iflib_netmap_detach(ifp) netmap_detach(ifp) #else #define iflib_netmap_txq_init(ctx, txq) #define iflib_netmap_rxq_init(ctx, rxq) #define iflib_netmap_detach(ifp) #define iflib_netmap_attach(ctx) (0) #define netmap_rx_irq(ifp, qid, budget) (0) #endif #if defined(__i386__) || defined(__amd64__) static __inline void prefetch(void *x) { __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); } #else #define prefetch(x) #endif static void _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err) { if (err) return; *(bus_addr_t *) arg = segs[0].ds_addr; } int iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags) { int err; if_shared_ctx_t sctx = ctx->ifc_sctx; device_t dev = ctx->ifc_dev; KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized")); err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ sctx->isc_q_align, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ size, /* maxsize */ 1, /* nsegments */ size, /* maxsegsize */ BUS_DMA_ALLOCNOW, /* flags */ NULL, /* lockfunc */ NULL, /* lockarg */ &dma->idi_tag); if (err) { device_printf(dev, "%s: bus_dma_tag_create failed: %d\n", __func__, err); goto fail_0; } err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr, BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map); if (err) { device_printf(dev, "%s: bus_dmamem_alloc(%ju) failed: %d\n", __func__, (uintmax_t)size, err); goto fail_1; } dma->idi_paddr = IF_BAD_DMA; err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr, size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT); if (err || dma->idi_paddr == IF_BAD_DMA) { device_printf(dev, "%s: bus_dmamap_load failed: %d\n", __func__, err); goto fail_2; } dma->idi_size = size; return (0); fail_2: bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); fail_1: bus_dma_tag_destroy(dma->idi_tag); fail_0: dma->idi_tag = NULL; return (err); } int iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count) { int i, err; iflib_dma_info_t *dmaiter; dmaiter = dmalist; for (i = 0; i < count; i++, dmaiter++) { if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0) break; } if (err) iflib_dma_free_multi(dmalist, i); return (err); } void iflib_dma_free(iflib_dma_info_t dma) { if (dma->idi_tag == NULL) return; if (dma->idi_paddr != IF_BAD_DMA) { bus_dmamap_sync(dma->idi_tag, dma->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(dma->idi_tag, dma->idi_map); dma->idi_paddr = IF_BAD_DMA; } if (dma->idi_vaddr != NULL) { bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); dma->idi_vaddr = NULL; } bus_dma_tag_destroy(dma->idi_tag); dma->idi_tag = NULL; } void iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count) { int i; iflib_dma_info_t *dmaiter = dmalist; for (i = 0; i < count; i++, dmaiter++) iflib_dma_free(*dmaiter); } static int iflib_fast_intr(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED) return (FILTER_HANDLED); GROUPTASK_ENQUEUE(gtask); return (FILTER_HANDLED); } static int _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, driver_filter_t filter, driver_intr_t handler, void *arg, char *name) { int rc; struct resource *res; void *tag; device_t dev = ctx->ifc_dev; MPASS(rid < 512); irq->ii_rid = rid; res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, RF_SHAREABLE | RF_ACTIVE); if (res == NULL) { device_printf(dev, "failed to allocate IRQ for rid %d, name %s.\n", rid, name); return (ENOMEM); } irq->ii_res = res; KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL")); rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET, filter, handler, arg, &tag); if (rc != 0) { device_printf(dev, "failed to setup interrupt for rid %d, name %s: %d\n", rid, name ? name : "unknown", rc); return (rc); } else if (name) bus_describe_intr(dev, res, tag, name); irq->ii_tag = tag; return (0); } /********************************************************************* * * Allocate memory for tx_buffer structures. The tx_buffer stores all * the information needed to transmit a packet on the wire. This is * called only once at attach, setup is done every reset. * **********************************************************************/ static int iflib_txsd_alloc(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; int err, nsegments, ntsosegments; nsegments = scctx->isc_tx_nsegments; ntsosegments = scctx->isc_tx_tso_segments_max; MPASS(sctx->isc_ntxd > 0); MPASS(nsegments > 0); MPASS(ntsosegments > 0); /* * Setup DMA descriptor areas. */ if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ sctx->isc_tx_maxsize, /* maxsize */ nsegments, /* nsegments */ sctx->isc_tx_maxsegsize, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockfuncarg */ &txq->ift_desc_tag))) { device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err); device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n", sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize); goto fail; } #ifdef INVARIANTS device_printf(dev,"maxsize: %zd nsegments: %d maxsegsize: %zd\n", sctx->isc_tx_maxsize, nsegments, sctx->isc_tx_maxsegsize); #endif device_printf(dev,"TSO maxsize: %d ntsosegments: %d maxsegsize: %d\n", scctx->isc_tx_tso_size_max, ntsosegments, scctx->isc_tx_tso_segsize_max); if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ scctx->isc_tx_tso_size_max, /* maxsize */ ntsosegments, /* nsegments */ scctx->isc_tx_tso_segsize_max, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockfuncarg */ &txq->ift_tso_desc_tag))) { device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err); goto fail; } #ifdef INVARIANTS device_printf(dev,"TSO maxsize: %d ntsosegments: %d maxsegsize: %d\n", scctx->isc_tx_tso_size_max, ntsosegments, scctx->isc_tx_tso_segsize_max); #endif if (!(txq->ift_sds.ifsd_flags = (uint8_t *) malloc(sizeof(uint8_t) * sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate tx_buffer memory\n"); err = ENOMEM; goto fail; } if (!(txq->ift_sds.ifsd_m = (struct mbuf **) malloc(sizeof(struct mbuf *) * sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate tx_buffer memory\n"); err = ENOMEM; goto fail; } /* Create the descriptor buffer dma maps */ #if defined(ACPI_DMAR) || (!(defined(__i386__) && !defined(__amd64__))) if ((ctx->ifc_flags & IFC_DMAR) == 0) return (0); if (!(txq->ift_sds.ifsd_map = (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * sctx->isc_ntxd, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate tx_buffer map memory\n"); err = ENOMEM; goto fail; } for (int i = 0; i < sctx->isc_ntxd; i++) { err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]); if (err != 0) { device_printf(dev, "Unable to create TX DMA map\n"); goto fail; } } #endif return (0); fail: /* We free all, it handles case where we are in the middle */ iflib_tx_structures_free(ctx); return (err); } static void iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i) { bus_dmamap_t map; map = NULL; if (txq->ift_sds.ifsd_map != NULL) map = txq->ift_sds.ifsd_map[i]; if (map != NULL) { bus_dmamap_unload(txq->ift_desc_tag, map); bus_dmamap_destroy(txq->ift_desc_tag, map); txq->ift_sds.ifsd_map[i] = NULL; } } static void iflib_txq_destroy(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; for (int i = 0; i < sctx->isc_ntxd; i++) iflib_txsd_destroy(ctx, txq, i); if (txq->ift_sds.ifsd_map != NULL) { free(txq->ift_sds.ifsd_map, M_IFLIB); txq->ift_sds.ifsd_map = NULL; } if (txq->ift_sds.ifsd_m != NULL) { free(txq->ift_sds.ifsd_m, M_IFLIB); txq->ift_sds.ifsd_m = NULL; } if (txq->ift_sds.ifsd_flags != NULL) { free(txq->ift_sds.ifsd_flags, M_IFLIB); txq->ift_sds.ifsd_flags = NULL; } if (txq->ift_desc_tag != NULL) { bus_dma_tag_destroy(txq->ift_desc_tag); txq->ift_desc_tag = NULL; } if (txq->ift_tso_desc_tag != NULL) { bus_dma_tag_destroy(txq->ift_tso_desc_tag); txq->ift_tso_desc_tag = NULL; } } static void iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i) { struct mbuf **mp; mp = &txq->ift_sds.ifsd_m[i]; if (*mp == NULL) return; if (txq->ift_sds.ifsd_map != NULL) { bus_dmamap_sync(txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_desc_tag, txq->ift_sds.ifsd_map[i]); } m_freem(*mp); DBG_COUNTER_INC(tx_frees); *mp = NULL; } static int iflib_txq_setup(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; int i; /* Set number of descriptors available */ txq->ift_qstatus = IFLIB_QUEUE_IDLE; /* Reset indices */ txq->ift_cidx_processed = txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0; txq->ift_size = sctx->isc_ntxd; for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++) bzero((void *)di->idi_vaddr, di->idi_size); IFDI_TXQ_SETUP(ctx, txq->ift_id); for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++) bus_dmamap_sync(di->idi_tag, di->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); return (0); } /********************************************************************* * * Allocate memory for rx_buffer structures. Since we use one * rx_buffer per received packet, the maximum number of rx_buffer's * that we'll need is equal to the number of receive descriptors * that we've allocated. * **********************************************************************/ static int iflib_rxsd_alloc(iflib_rxq_t rxq) { if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; device_t dev = ctx->ifc_dev; iflib_fl_t fl; iflib_rxsd_t rxsd; int err; MPASS(sctx->isc_nrxd > 0); fl = rxq->ifr_fl; for (int i = 0; i < rxq->ifr_nfl; i++, fl++) { fl->ifl_sds = malloc(sizeof(struct iflib_sw_rx_desc) * sctx->isc_nrxd, M_IFLIB, M_WAITOK | M_ZERO); if (fl->ifl_sds == NULL) { device_printf(dev, "Unable to allocate rx sw desc memory\n"); return (ENOMEM); } fl->ifl_size = sctx->isc_nrxd; /* this isn't necessarily the same */ err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ sctx->isc_rx_maxsize, /* maxsize */ sctx->isc_rx_nsegments, /* nsegments */ sctx->isc_rx_maxsegsize, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockarg */ &fl->ifl_desc_tag); if (err) { device_printf(dev, "%s: bus_dma_tag_create failed %d\n", __func__, err); goto fail; } rxsd = fl->ifl_sds; for (int i = 0; i < sctx->isc_nrxd; i++, rxsd++) { err = bus_dmamap_create(fl->ifl_desc_tag, 0, &rxsd->ifsd_map); if (err) { device_printf(dev, "%s: bus_dmamap_create failed: %d\n", __func__, err); goto fail; } } } return (0); fail: iflib_rx_structures_free(ctx); return (err); } /* * Internal service routines */ struct rxq_refill_cb_arg { int error; bus_dma_segment_t seg; int nseg; }; static void _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) { struct rxq_refill_cb_arg *cb_arg = arg; cb_arg->error = error; cb_arg->seg = segs[0]; cb_arg->nseg = nseg; } #ifdef ACPI_DMAR #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR) #else #define IS_DMAR(ctx) (0) #endif /** * rxq_refill - refill an rxq free-buffer list * @ctx: the iflib context * @rxq: the free-list to refill * @n: the number of new buffers to allocate * * (Re)populate an rxq free-buffer list with up to @n new packet buffers. * The caller must assure that @n does not exceed the queue's capacity. */ static void _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count) { struct mbuf *m; int pidx = fl->ifl_pidx; iflib_rxsd_t rxsd = &fl->ifl_sds[pidx]; caddr_t cl; int n, i = 0; uint64_t bus_addr; int err; n = count; MPASS(n > 0); MPASS(fl->ifl_credits + n <= fl->ifl_size); if (pidx < fl->ifl_cidx) MPASS(pidx + n <= fl->ifl_cidx); if (pidx == fl->ifl_cidx && (fl->ifl_credits < fl->ifl_size)) MPASS(fl->ifl_gen == 0); if (pidx > fl->ifl_cidx) MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx); DBG_COUNTER_INC(fl_refills); if (n > 8) DBG_COUNTER_INC(fl_refills_large); while (n--) { /* * We allocate an uninitialized mbuf + cluster, mbuf is * initialized after rx. * * If the cluster is still set then we know a minimum sized packet was received */ if ((cl = rxsd->ifsd_cl) == NULL) { if ((cl = rxsd->ifsd_cl = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL) break; #if MEMORY_LOGGING fl->ifl_cl_enqueued++; #endif } if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) { break; } #if MEMORY_LOGGING fl->ifl_m_enqueued++; #endif DBG_COUNTER_INC(rx_allocs); #ifdef notyet if ((rxsd->ifsd_flags & RX_SW_DESC_MAP_CREATED) == 0) { int err; if ((err = bus_dmamap_create(fl->ifl_ifdi->idi_tag, 0, &rxsd->ifsd_map))) { log(LOG_WARNING, "bus_dmamap_create failed %d\n", err); uma_zfree(fl->ifl_zone, cl); n = 0; goto done; } rxsd->ifsd_flags |= RX_SW_DESC_MAP_CREATED; } #endif #if defined(__i386__) || defined(__amd64__) if (!IS_DMAR(ctx)) { bus_addr = pmap_kextract((vm_offset_t)cl); } else #endif { struct rxq_refill_cb_arg cb_arg; iflib_rxq_t q; cb_arg.error = 0; q = fl->ifl_rxq; err = bus_dmamap_load(fl->ifl_desc_tag, rxsd->ifsd_map, cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0); if (err != 0 || cb_arg.error) { /* * !zone_pack ? */ if (fl->ifl_zone == zone_pack) uma_zfree(fl->ifl_zone, cl); m_free(m); n = 0; goto done; } bus_addr = cb_arg.seg.ds_addr; } rxsd->ifsd_flags |= RX_SW_DESC_INUSE; MPASS(rxsd->ifsd_m == NULL); rxsd->ifsd_cl = cl; rxsd->ifsd_m = m; fl->ifl_bus_addrs[i] = bus_addr; fl->ifl_vm_addrs[i] = cl; rxsd++; fl->ifl_credits++; i++; MPASS(fl->ifl_credits <= fl->ifl_size); if (++fl->ifl_pidx == fl->ifl_size) { fl->ifl_pidx = 0; fl->ifl_gen = 1; rxsd = fl->ifl_sds; } if (n == 0 || i == IFLIB_MAX_RX_REFRESH) { ctx->isc_rxd_refill(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx, fl->ifl_bus_addrs, fl->ifl_vm_addrs, i); i = 0; pidx = fl->ifl_pidx; } } done: DBG_COUNTER_INC(rxd_flush); if (fl->ifl_pidx == 0) pidx = fl->ifl_size - 1; else pidx = fl->ifl_pidx - 1; ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx); } static __inline void __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max) { /* we avoid allowing pidx to catch up with cidx as it confuses ixl */ int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1; #ifdef INVARIANTS int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1; #endif MPASS(fl->ifl_credits <= fl->ifl_size); MPASS(reclaimable == delta); if (reclaimable > 0) _iflib_fl_refill(ctx, fl, min(max, reclaimable)); } static void iflib_fl_bufs_free(iflib_fl_t fl) { iflib_dma_info_t idi = fl->ifl_ifdi; uint32_t i; for (i = 0; i < fl->ifl_size; i++) { iflib_rxsd_t d = &fl->ifl_sds[i]; if (d->ifsd_flags & RX_SW_DESC_INUSE) { bus_dmamap_unload(fl->ifl_desc_tag, d->ifsd_map); bus_dmamap_destroy(fl->ifl_desc_tag, d->ifsd_map); if (d->ifsd_m != NULL) { m_init(d->ifsd_m, M_NOWAIT, MT_DATA, 0); uma_zfree(zone_mbuf, d->ifsd_m); } if (d->ifsd_cl != NULL) uma_zfree(fl->ifl_zone, d->ifsd_cl); d->ifsd_flags = 0; } else { MPASS(d->ifsd_cl == NULL); MPASS(d->ifsd_m == NULL); } #if MEMORY_LOGGING fl->ifl_m_dequeued++; fl->ifl_cl_dequeued++; #endif d->ifsd_cl = NULL; d->ifsd_m = NULL; } /* * Reset free list values */ fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = 0;; bzero(idi->idi_vaddr, idi->idi_size); } /********************************************************************* * * Initialize a receive ring and its buffers. * **********************************************************************/ static int iflib_fl_setup(iflib_fl_t fl) { iflib_rxq_t rxq = fl->ifl_rxq; if_ctx_t ctx = rxq->ifr_ctx; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; /* ** Free current RX buffer structs and their mbufs */ iflib_fl_bufs_free(fl); /* Now replenish the mbufs */ MPASS(fl->ifl_credits == 0); /* * XXX don't set the max_frame_size to larger * than the hardware can handle */ if (sctx->isc_max_frame_size <= 2048) fl->ifl_buf_size = MCLBYTES; else if (sctx->isc_max_frame_size <= 4096) fl->ifl_buf_size = MJUMPAGESIZE; else if (sctx->isc_max_frame_size <= 9216) fl->ifl_buf_size = MJUM9BYTES; else fl->ifl_buf_size = MJUM16BYTES; if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size) ctx->ifc_max_fl_buf_size = fl->ifl_buf_size; fl->ifl_cltype = m_gettype(fl->ifl_buf_size); fl->ifl_zone = m_getzone(fl->ifl_buf_size); /* avoid pre-allocating zillions of clusters to an idle card * potentially speeding up attach */ _iflib_fl_refill(ctx, fl, min(128, fl->ifl_size)); MPASS(min(128, fl->ifl_size) == fl->ifl_credits); if (min(128, fl->ifl_size) != fl->ifl_credits) return (ENOBUFS); /* * handle failure */ MPASS(rxq != NULL); MPASS(fl->ifl_ifdi != NULL); bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); return (0); } /********************************************************************* * * Free receive ring data structures * **********************************************************************/ static void iflib_rx_sds_free(iflib_rxq_t rxq) { iflib_fl_t fl; int i; if (rxq->ifr_fl != NULL) { for (i = 0; i < rxq->ifr_nfl; i++) { fl = &rxq->ifr_fl[i]; if (fl->ifl_desc_tag != NULL) { bus_dma_tag_destroy(fl->ifl_desc_tag); fl->ifl_desc_tag = NULL; } } if (rxq->ifr_fl->ifl_sds != NULL) free(rxq->ifr_fl->ifl_sds, M_IFLIB); free(rxq->ifr_fl, M_IFLIB); rxq->ifr_fl = NULL; rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0; } } /* * MI independent logic * */ static void iflib_timer(void *arg) { iflib_txq_t txq = arg; if_ctx_t ctx = txq->ift_ctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; /* ** Check on the state of the TX queue(s), this ** can be done without the lock because its RO ** and the HUNG state will be static if set. */ IFDI_TIMER(ctx, txq->ift_id); if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) && (ctx->ifc_pause_frames == 0)) goto hung; if (TXQ_AVAIL(txq) <= 2*scctx->isc_tx_nsegments || ifmp_ring_is_stalled(txq->ift_br[0])) GROUPTASK_ENQUEUE(&txq->ift_task); ctx->ifc_pause_frames = 0; if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu); return; hung: CTX_LOCK(ctx); if_setdrvflagbits(ctx->ifc_ifp, 0, IFF_DRV_RUNNING); device_printf(ctx->ifc_dev, "TX(%d) desc avail = %d, pidx = %d\n", txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx); IFDI_WATCHDOG_RESET(ctx); ctx->ifc_watchdog_events++; ctx->ifc_pause_frames = 0; iflib_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_init_locked(if_ctx_t ctx) { if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; if_t ifp = ctx->ifc_ifp; iflib_fl_t fl; iflib_txq_t txq; iflib_rxq_t rxq; int i, j; if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); IFDI_INTR_DISABLE(ctx); /* Set hardware offload abilities */ if_clearhwassist(ifp); if (if_getcapenable(ifp) & IFCAP_TXCSUM) if_sethwassistbits(ifp, CSUM_IP | CSUM_TCP | CSUM_UDP, 0); if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6) if_sethwassistbits(ifp, (CSUM_TCP_IPV6 | CSUM_UDP_IPV6), 0); if (if_getcapenable(ifp) & IFCAP_TSO4) if_sethwassistbits(ifp, CSUM_IP_TSO, 0); if (if_getcapenable(ifp) & IFCAP_TSO6) if_sethwassistbits(ifp, CSUM_IP6_TSO, 0); for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) { CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); callout_stop(&txq->ift_db_check); CALLOUT_UNLOCK(txq); iflib_netmap_txq_init(ctx, txq); } for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { iflib_netmap_rxq_init(ctx, rxq); } IFDI_INIT(ctx); for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { if (iflib_fl_setup(fl)) { device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n"); goto done; } } } done: if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); IFDI_INTR_ENABLE(ctx); txq = ctx->ifc_txqs; for (i = 0; i < sctx->isc_ntxqsets; i++, txq++) callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu); } static int iflib_media_change(if_t ifp) { if_ctx_t ctx = if_getsoftc(ifp); int err; CTX_LOCK(ctx); if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0) iflib_init_locked(ctx); CTX_UNLOCK(ctx); return (err); } static void iflib_media_status(if_t ifp, struct ifmediareq *ifmr) { if_ctx_t ctx = if_getsoftc(ifp); CTX_LOCK(ctx); IFDI_UPDATE_ADMIN_STATUS(ctx); IFDI_MEDIA_STATUS(ctx, ifmr); CTX_UNLOCK(ctx); } static void iflib_stop(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; iflib_fl_t fl; int i, j; /* Tell the stack that the interface is no longer active */ if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); IFDI_INTR_DISABLE(ctx); msleep(ctx, &ctx->ifc_mtx, PUSER, "iflib_init", hz); /* Wait for current tx queue users to exit to disarm watchdog timer. */ for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) { /* make sure all transmitters have completed before proceeding XXX */ /* clean any enqueued buffers */ iflib_txq_check_drain(txq, 0); /* Free any existing tx buffers. */ for (j = 0; j < sctx->isc_ntxd; j++) { iflib_txsd_free(ctx, txq, j); } txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0; txq->ift_in_use = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0; txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0; txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0; txq->ift_pullups = 0; ifmp_ring_reset_stats(txq->ift_br[0]); for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++) bzero((void *)di->idi_vaddr, di->idi_size); } for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) { /* make sure all transmitters have completed before proceeding XXX */ for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++) bzero((void *)di->idi_vaddr, di->idi_size); /* also resets the free lists pidx/cidx */ for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) iflib_fl_bufs_free(fl); } IFDI_STOP(ctx); } static iflib_rxsd_t rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int *cltype, int unload) { int flid, cidx; iflib_rxsd_t sd; iflib_fl_t fl; iflib_dma_info_t di; flid = irf->irf_flid; cidx = irf->irf_idx; fl = &rxq->ifr_fl[flid]; fl->ifl_credits--; #if MEMORY_LOGGING fl->ifl_m_dequeued++; if (cltype) fl->ifl_cl_dequeued++; #endif sd = &fl->ifl_sds[cidx]; di = fl->ifl_ifdi; bus_dmamap_sync(di->idi_tag, di->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); /* not valid assert if bxe really does SGE from non-contiguous elements */ MPASS(fl->ifl_cidx == cidx); if (unload) bus_dmamap_unload(fl->ifl_desc_tag, sd->ifsd_map); if (__predict_false(++fl->ifl_cidx == fl->ifl_size)) { fl->ifl_cidx = 0; fl->ifl_gen = 0; } /* YES ick */ if (cltype) *cltype = fl->ifl_cltype; return (sd); } static struct mbuf * assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri) { int i, padlen , flags, cltype; struct mbuf *m, *mh, *mt; iflib_rxsd_t sd; caddr_t cl; i = 0; do { sd = rxd_frag_to_sd(rxq, &ri->iri_frags[i], &cltype, TRUE); MPASS(sd->ifsd_cl != NULL); MPASS(sd->ifsd_m != NULL); m = sd->ifsd_m; if (i == 0) { flags = M_PKTHDR|M_EXT; mh = mt = m; padlen = ri->iri_pad; } else { flags = M_EXT; mt->m_next = m; mt = m; /* assuming padding is only on the first fragment */ padlen = 0; } sd->ifsd_m = NULL; cl = sd->ifsd_cl; sd->ifsd_cl = NULL; /* Can these two be made one ? */ m_init(m, M_NOWAIT, MT_DATA, flags); m_cljset(m, cl, cltype); /* * These must follow m_init and m_cljset */ m->m_data += padlen; ri->iri_len -= padlen; m->m_len = ri->iri_len; } while (++i < ri->iri_nfrags); return (mh); } /* * Process one software descriptor */ static struct mbuf * iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri) { struct mbuf *m; iflib_rxsd_t sd; /* should I merge this back in now that the two paths are basically duplicated? */ if (ri->iri_len <= IFLIB_RX_COPY_THRESH) { sd = rxd_frag_to_sd(rxq, &ri->iri_frags[0], NULL, FALSE); m = sd->ifsd_m; sd->ifsd_m = NULL; m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR); memcpy(m->m_data, sd->ifsd_cl, ri->iri_len); m->m_len = ri->iri_len; } else { m = assemble_segments(rxq, ri); } m->m_pkthdr.len = ri->iri_len; m->m_pkthdr.rcvif = ri->iri_ifp; m->m_flags |= ri->iri_flags; m->m_pkthdr.ether_vtag = ri->iri_vtag; m->m_pkthdr.flowid = ri->iri_flowid; M_HASHTYPE_SET(m, ri->iri_rsstype); m->m_pkthdr.csum_flags = ri->iri_csum_flags; m->m_pkthdr.csum_data = ri->iri_csum_data; return (m); } static bool iflib_rxeof(iflib_rxq_t rxq, int budget) { if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; int avail, i; uint16_t *cidxp; struct if_rxd_info ri; int err, budget_left, rx_bytes, rx_pkts; iflib_fl_t fl; struct ifnet *ifp; struct lro_entry *queued; int lro_enabled; /* * XXX early demux data packets so that if_input processing only handles * acks in interrupt context */ struct mbuf *m, *mh, *mt; if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &budget)) { return (FALSE); } mh = mt = NULL; MPASS(budget > 0); rx_pkts = rx_bytes = 0; if (sctx->isc_flags & IFLIB_HAS_CQ) cidxp = &rxq->ifr_cq_cidx; else cidxp = &rxq->ifr_fl[0].ifl_cidx; if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp)) == 0) { for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) __iflib_fl_refill_lt(ctx, fl, budget + 8); DBG_COUNTER_INC(rx_unavail); return (false); } for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) { if (__predict_false(!CTX_ACTIVE(ctx))) { DBG_COUNTER_INC(rx_ctx_inactive); break; } /* * Reset client set fields to their default values */ bzero(&ri, sizeof(ri)); ri.iri_qsidx = rxq->ifr_id; ri.iri_cidx = *cidxp; ri.iri_ifp = ctx->ifc_ifp; ri.iri_frags = rxq->ifr_frags; err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); /* in lieu of handling correctly - make sure it isn't being unhandled */ MPASS(err == 0); if (sctx->isc_flags & IFLIB_HAS_CQ) { /* we know we consumed _one_ CQ entry */ if (++rxq->ifr_cq_cidx == sctx->isc_nrxd) { rxq->ifr_cq_cidx = 0; rxq->ifr_cq_gen = 0; } /* was this only a completion queue message? */ if (__predict_false(ri.iri_nfrags == 0)) continue; } MPASS(ri.iri_nfrags != 0); MPASS(ri.iri_len != 0); /* will advance the cidx on the corresponding free lists */ m = iflib_rxd_pkt_get(rxq, &ri); if (avail == 0 && budget_left) avail = iflib_rxd_avail(ctx, rxq, *cidxp); if (__predict_false(m == NULL)) { DBG_COUNTER_INC(rx_mbuf_null); continue; } /* imm_pkt: -- cxgb */ if (mh == NULL) mh = mt = m; else { mt->m_nextpkt = m; mt = m; } } /* make sure that we can refill faster than drain */ for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) __iflib_fl_refill_lt(ctx, fl, budget + 8); ifp = ctx->ifc_ifp; lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO); while (mh != NULL) { m = mh; mh = mh->m_nextpkt; m->m_nextpkt = NULL; rx_bytes += m->m_pkthdr.len; rx_pkts++; #if defined(INET6) || defined(INET) if (lro_enabled && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0) continue; #endif DBG_COUNTER_INC(rx_if_input); ifp->if_input(ifp, m); } if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); /* * Flush any outstanding LRO work */ while ((queued = LIST_FIRST(&rxq->ifr_lc.lro_active)) != NULL) { LIST_REMOVE(queued, next); #if defined(INET6) || defined(INET) tcp_lro_flush(&rxq->ifr_lc, queued); #endif } return (iflib_rxd_avail(ctx, rxq, *cidxp)); } #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags) #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG) #define TXQ_MAX_DB_DEFERRED(ctx) (ctx->ifc_sctx->isc_ntxd >> 5) #define TXQ_MAX_DB_CONSUMED(ctx) (ctx->ifc_sctx->isc_ntxd >> 4) static __inline void iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring) { uint32_t dbval; if (ring || txq->ift_db_pending >= TXQ_MAX_DB_DEFERRED(ctx)) { /* the lock will only ever be contended in the !min_latency case */ if (!TXDB_TRYLOCK(txq)) return; dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx; ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval); txq->ift_db_pending = txq->ift_npending = 0; TXDB_UNLOCK(txq); } } static void iflib_txd_deferred_db_check(void * arg) { iflib_txq_t txq = arg; /* simple non-zero boolean so use bitwise OR */ if ((txq->ift_db_pending | txq->ift_npending) && txq->ift_db_pending >= txq->ift_db_pending_queued) iflib_txd_db_check(txq->ift_ctx, txq, TRUE); txq->ift_db_pending_queued = 0; if (ifmp_ring_is_stalled(txq->ift_br[0])) iflib_txq_check_drain(txq, 4); } #ifdef PKT_DEBUG static void print_pkt(if_pkt_info_t pi) { printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n", pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx); printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n", pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag); printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n", pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto); } #endif #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO) #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO) static int iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp) { struct ether_vlan_header *eh; struct mbuf *m; m = *mp; /* * Determine where frame payload starts. * Jump over vlan headers if already present, * helpful for QinQ too. */ if (__predict_false(m->m_len < sizeof(*eh))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL)) return (ENOMEM); } eh = mtod(m, struct ether_vlan_header *); if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { pi->ipi_etype = ntohs(eh->evl_proto); pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; } else { pi->ipi_etype = ntohs(eh->evl_encap_proto); pi->ipi_ehdrlen = ETHER_HDR_LEN; } switch (pi->ipi_etype) { #ifdef INET case ETHERTYPE_IP: { struct ip *ip = NULL; struct tcphdr *th = NULL; struct mbuf *n; int minthlen; minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th)); if (__predict_false(m->m_len < minthlen)) { /* * if this code bloat is causing too much of a hit * move it to a separate function and mark it noinline */ if (m->m_len == pi->ipi_ehdrlen) { n = m->m_next; MPASS(n); if (n->m_len >= sizeof(*ip)) { ip = (struct ip *)n->m_data; if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } else { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) return (ENOMEM); ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); } } else { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) return (ENOMEM); ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } } else { ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } pi->ipi_ip_hlen = ip->ip_hl << 2; pi->ipi_ipproto = ip->ip_p; pi->ipi_flags |= IPI_TX_IPV4; if (pi->ipi_csum_flags & CSUM_IP) ip->ip_sum = 0; if (pi->ipi_ipproto == IPPROTO_TCP) { if (__predict_false(th == NULL)) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL)) return (ENOMEM); th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen); } pi->ipi_tcp_hflags = th->th_flags; pi->ipi_tcp_hlen = th->th_off << 2; pi->ipi_tcp_seq = th->th_seq; } if (IS_TSO4(pi)) { if (__predict_false(ip->ip_p != IPPROTO_TCP)) return (ENXIO); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(IPPROTO_TCP)); pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; } break; } #endif #ifdef INET6 case ETHERTYPE_IPV6: { struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen); struct tcphdr *th; pi->ipi_ip_hlen = sizeof(struct ip6_hdr); if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) { if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL)) return (ENOMEM); } th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen); /* XXX-BZ this will go badly in case of ext hdrs. */ pi->ipi_ipproto = ip6->ip6_nxt; pi->ipi_flags |= IPI_TX_IPV6; if (pi->ipi_ipproto == IPPROTO_TCP) { if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) { if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL)) return (ENOMEM); } pi->ipi_tcp_hflags = th->th_flags; pi->ipi_tcp_hlen = th->th_off << 2; } if (IS_TSO6(pi)) { if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP)) return (ENXIO); /* * The corresponding flag is set by the stack in the IPv4 * TSO case, but not in IPv6 (at least in FreeBSD 10.2). * So, set it here because the rest of the flow requires it. */ pi->ipi_csum_flags |= CSUM_TCP_IPV6; th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; } break; } #endif default: pi->ipi_csum_flags &= ~CSUM_OFFLOAD; pi->ipi_ip_hlen = 0; break; } *mp = m; return (0); } static __noinline struct mbuf * collapse_pkthdr(struct mbuf *m0) { struct mbuf *m, *m_next, *tmp; m = m0; m_next = m->m_next; while (m_next != NULL && m_next->m_len == 0) { m = m_next; m->m_next = NULL; m_free(m); m_next = m_next->m_next; } m = m0; m->m_next = m_next; if ((m_next->m_flags & M_EXT) == 0) { m = m_defrag(m, M_NOWAIT); } else { tmp = m_next->m_next; memcpy(m_next, m, MPKTHSIZE); m = m_next; m->m_next = tmp; } return (m); } /* * If dodgy hardware rejects the scatter gather chain we've handed it * we'll need to rebuild the mbuf chain before we can call m_defrag */ static __noinline struct mbuf * iflib_rebuild_mbuf(iflib_txq_t txq) { int ntxd, mhlen, len, i, pidx; struct mbuf *m, *mh, **ifsd_m; if_shared_ctx_t sctx; pidx = txq->ift_pidx; ifsd_m = txq->ift_sds.ifsd_m; sctx = txq->ift_ctx->ifc_sctx; ntxd = sctx->isc_ntxd; mh = m = ifsd_m[pidx]; ifsd_m[pidx] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif len = m->m_len; mhlen = m->m_pkthdr.len; i = 1; while (len < mhlen && (m->m_next == NULL)) { m->m_next = ifsd_m[(pidx + i) & (ntxd-1)]; ifsd_m[(pidx + i) & (ntxd -1)] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif m = m->m_next; len += m->m_len; i++; } return (mh); } static int iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map, struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs, int max_segs, int flags) { if_ctx_t ctx; if_shared_ctx_t sctx; int i, next, pidx, mask, err, maxsegsz, ntxd, count; struct mbuf *m, *tmp, **ifsd_m, **mp; m = *m0; /* * Please don't ever do this */ if (__predict_false(m->m_len == 0)) *m0 = m = collapse_pkthdr(m); ctx = txq->ift_ctx; sctx = ctx->ifc_sctx; ifsd_m = txq->ift_sds.ifsd_m; ntxd = sctx->isc_ntxd; pidx = txq->ift_pidx; if (map != NULL) { uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags; err = bus_dmamap_load_mbuf_sg(tag, map, *m0, segs, nsegs, BUS_DMA_NOWAIT); if (err) return (err); ifsd_flags[pidx] |= TX_SW_DESC_MAPPED; i = 0; next = pidx; mask = (sctx->isc_ntxd-1); m = *m0; do { mp = &ifsd_m[next]; *mp = m; m = m->m_next; (*mp)->m_next = NULL; if (__predict_false((*mp)->m_len == 0)) { m_free(*mp); *mp = NULL; } else next = (pidx + i) & (ntxd-1); } while (m != NULL); } else { int buflen, sgsize, max_sgsize; vm_offset_t vaddr; vm_paddr_t curaddr; count = i = 0; maxsegsz = sctx->isc_tx_maxsize; m = *m0; do { if (__predict_false(m->m_len <= 0)) { tmp = m; m = m->m_next; tmp->m_next = NULL; m_free(tmp); continue; } buflen = m->m_len; vaddr = (vm_offset_t)m->m_data; /* * see if we can't be smarter about physically * contiguous mappings */ next = (pidx + count) & (ntxd-1); MPASS(ifsd_m[next] == NULL); #if MEMORY_LOGGING txq->ift_enqueued++; #endif ifsd_m[next] = m; while (buflen > 0) { max_sgsize = MIN(buflen, maxsegsz); curaddr = pmap_kextract(vaddr); sgsize = PAGE_SIZE - (curaddr & PAGE_MASK); sgsize = MIN(sgsize, max_sgsize); segs[i].ds_addr = curaddr; segs[i].ds_len = sgsize; vaddr += sgsize; buflen -= sgsize; i++; if (i >= max_segs) goto err; } count++; tmp = m; m = m->m_next; tmp->m_next = NULL; } while (m != NULL); *nsegs = i; } return (0); err: *m0 = iflib_rebuild_mbuf(txq); return (EFBIG); } static int iflib_encap(iflib_txq_t txq, struct mbuf **m_headp) { if_ctx_t ctx; if_shared_ctx_t sctx; if_softc_ctx_t scctx; bus_dma_segment_t *segs; struct mbuf *m_head; bus_dmamap_t map; struct if_pkt_info pi; int remap = 0; int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd; bus_dma_tag_t desc_tag; segs = txq->ift_segs; ctx = txq->ift_ctx; sctx = ctx->ifc_sctx; scctx = &ctx->ifc_softc_ctx; segs = txq->ift_segs; ntxd = sctx->isc_ntxd; m_head = *m_headp; map = NULL; /* * If we're doing TSO the next descriptor to clean may be quite far ahead */ cidx = txq->ift_cidx; pidx = txq->ift_pidx; next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1); /* prefetch the next cache line of mbuf pointers and flags */ prefetch(&txq->ift_sds.ifsd_m[next]); if (txq->ift_sds.ifsd_map != NULL) { prefetch(&txq->ift_sds.ifsd_map[next]); map = txq->ift_sds.ifsd_map[pidx]; next = (cidx + CACHE_LINE_SIZE) & (ntxd-1); prefetch(&txq->ift_sds.ifsd_flags[next]); } if (m_head->m_pkthdr.csum_flags & CSUM_TSO) { desc_tag = txq->ift_tso_desc_tag; max_segs = scctx->isc_tx_tso_segments_max; } else { desc_tag = txq->ift_desc_tag; max_segs = scctx->isc_tx_nsegments; } m_head = *m_headp; bzero(&pi, sizeof(pi)); pi.ipi_len = m_head->m_pkthdr.len; pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST)); pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags; pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0; pi.ipi_pidx = pidx; pi.ipi_qsidx = txq->ift_id; /* deliberate bitwise OR to make one condition */ if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) { if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) return (err); m_head = *m_headp; } retry: err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT); defrag: if (__predict_false(err)) { switch (err) { case EFBIG: /* try collapse once and defrag once */ if (remap == 0) m_head = m_collapse(*m_headp, M_NOWAIT, max_segs); if (remap == 1) m_head = m_defrag(*m_headp, M_NOWAIT); remap++; if (__predict_false(m_head == NULL)) goto defrag_failed; txq->ift_mbuf_defrag++; *m_headp = m_head; goto retry; break; case ENOMEM: txq->ift_no_tx_dma_setup++; break; default: txq->ift_no_tx_dma_setup++; m_freem(*m_headp); DBG_COUNTER_INC(tx_frees); *m_headp = NULL; break; } txq->ift_map_failed++; DBG_COUNTER_INC(encap_load_mbuf_fail); return (err); } /* * XXX assumes a 1 to 1 relationship between segments and * descriptors - this does not hold true on all drivers, e.g. * cxgb */ if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) { txq->ift_no_desc_avail++; if (map != NULL) bus_dmamap_unload(desc_tag, map); DBG_COUNTER_INC(encap_txq_avail_fail); if (txq->ift_task.gt_task.ta_pending == 0) GROUPTASK_ENQUEUE(&txq->ift_task); return (ENOBUFS); } pi.ipi_segs = segs; pi.ipi_nsegs = nsegs; MPASS(pidx >= 0 && pidx < sctx->isc_ntxd); #ifdef PKT_DEBUG print_pkt(&pi); #endif if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) { bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); DBG_COUNTER_INC(tx_encap); MPASS(pi.ipi_new_pidx >= 0 && pi.ipi_new_pidx < sctx->isc_ntxd); ndesc = pi.ipi_new_pidx - pi.ipi_pidx; if (pi.ipi_new_pidx < pi.ipi_pidx) { ndesc += sctx->isc_ntxd; txq->ift_gen = 1; } MPASS(pi.ipi_new_pidx != pidx); MPASS(ndesc > 0); txq->ift_in_use += ndesc; /* * We update the last software descriptor again here because there may * be a sentinel and/or there may be more mbufs than segments */ txq->ift_pidx = pi.ipi_new_pidx; txq->ift_npending += pi.ipi_ndescs; } else if (__predict_false(err == EFBIG && remap < 2)) { *m_headp = m_head = iflib_rebuild_mbuf(txq); remap = 1; txq->ift_txd_encap_efbig++; goto defrag; } else DBG_COUNTER_INC(encap_txd_encap_fail); return (err); defrag_failed: txq->ift_mbuf_defrag_failed++; txq->ift_map_failed++; m_freem(*m_headp); DBG_COUNTER_INC(tx_frees); *m_headp = NULL; return (ENOMEM); } /* forward compatibility for cxgb */ #define FIRST_QSET(ctx) 0 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets) #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets) #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NRXQSETS(ctx)) + FIRST_QSET(ctx)) #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments)) #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh) #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max) /* if there are more than TXQ_MIN_OCCUPANCY packets pending we consider deferring * doorbell writes * * ORing with 2 assures that min occupancy is never less than 2 without any conditional logic */ #define TXQ_MIN_OCCUPANCY(ctx) ((ctx->ifc_sctx->isc_ntxd >> 6)| 0x2) static inline int iflib_txq_min_occupancy(iflib_txq_t txq) { if_ctx_t ctx; ctx = txq->ift_ctx; return (get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen) < TXQ_MIN_OCCUPANCY(ctx) + MAX_TX_DESC(ctx)); } static void iflib_tx_desc_free(iflib_txq_t txq, int n) { int hasmap; uint32_t qsize, cidx, mask, gen; struct mbuf *m, **ifsd_m; uint8_t *ifsd_flags; bus_dmamap_t *ifsd_map; cidx = txq->ift_cidx; gen = txq->ift_gen; qsize = txq->ift_ctx->ifc_sctx->isc_ntxd; mask = qsize-1; hasmap = txq->ift_sds.ifsd_map != NULL; ifsd_flags = txq->ift_sds.ifsd_flags; ifsd_m = txq->ift_sds.ifsd_m; ifsd_map = txq->ift_sds.ifsd_map; while (n--) { prefetch(ifsd_m[(cidx + 3) & mask]); prefetch(ifsd_m[(cidx + 4) & mask]); if (ifsd_m[cidx] != NULL) { prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]); prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]); if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) { /* * does it matter if it's not the TSO tag? If so we'll * have to add the type to flags */ bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]); ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED; } if ((m = ifsd_m[cidx]) != NULL) { /* XXX we don't support any drivers that batch packets yet */ MPASS(m->m_nextpkt == NULL); m_freem(m); ifsd_m[cidx] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif DBG_COUNTER_INC(tx_frees); } } if (__predict_false(++cidx == qsize)) { cidx = 0; gen = 0; } } txq->ift_cidx = cidx; txq->ift_gen = gen; } static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh) { int reclaim; if_ctx_t ctx = txq->ift_ctx; KASSERT(thresh >= 0, ("invalid threshold to reclaim")); MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size); /* * Need a rate-limiting check so that this isn't called every time */ iflib_tx_credits_update(ctx, txq); reclaim = DESC_RECLAIMABLE(txq); if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) { #ifdef INVARIANTS if (iflib_verbose_debug) { printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__, txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments, reclaim, thresh); } #endif return (0); } iflib_tx_desc_free(txq, reclaim); txq->ift_cleaned += reclaim; txq->ift_in_use -= reclaim; if (txq->ift_active == FALSE) txq->ift_active = TRUE; return (reclaim); } static struct mbuf ** _ring_peek_one(struct ifmp_ring *r, int cidx, int offset) { return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (r->size-1)])); } static void iflib_txq_check_drain(iflib_txq_t txq, int budget) { ifmp_ring_check_drainage(txq->ift_br[0], budget); } static uint32_t iflib_txq_can_drain(struct ifmp_ring *r) { iflib_txq_t txq = r->cookie; if_ctx_t ctx = txq->ift_ctx; return ((TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx)) || ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, false)); } static uint32_t iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) { iflib_txq_t txq = r->cookie; if_ctx_t ctx = txq->ift_ctx; if_t ifp = ctx->ifc_ifp; struct mbuf **mp, *m; int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail, err, in_use_prev, desc_used; if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(txq_drain_notready); return (0); } avail = IDXDIFF(pidx, cidx, r->size); if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) { DBG_COUNTER_INC(txq_drain_flushing); for (i = 0; i < avail; i++) { m_freem(r->items[(cidx + i) & (r->size-1)]); r->items[(cidx + i) & (r->size-1)] = NULL; } return (avail); } iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) { txq->ift_qstatus = IFLIB_QUEUE_IDLE; CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); callout_stop(&txq->ift_db_check); CALLOUT_UNLOCK(txq); DBG_COUNTER_INC(txq_drain_oactive); return (0); } consumed = mcast_sent = bytes_sent = pkt_sent = 0; count = MIN(avail, TX_BATCH_SIZE); for (desc_used = i = 0; i < count && TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2; i++) { mp = _ring_peek_one(r, cidx, i); in_use_prev = txq->ift_in_use; err = iflib_encap(txq, mp); /* * What other errors should we bail out for? */ if (err == ENOBUFS) { DBG_COUNTER_INC(txq_drain_encapfail); break; } consumed++; if (err) continue; pkt_sent++; m = *mp; DBG_COUNTER_INC(tx_sent); bytes_sent += m->m_pkthdr.len; if (m->m_flags & M_MCAST) mcast_sent++; txq->ift_db_pending += (txq->ift_in_use - in_use_prev); desc_used += (txq->ift_in_use - in_use_prev); iflib_txd_db_check(ctx, txq, FALSE); ETHER_BPF_MTAP(ifp, m); if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) break; if (desc_used > TXQ_MAX_DB_CONSUMED(ctx)) break; } if ((iflib_min_tx_latency || iflib_txq_min_occupancy(txq)) && txq->ift_db_pending) iflib_txd_db_check(ctx, txq, TRUE); else if ((txq->ift_db_pending || TXQ_AVAIL(txq) < MAX_TX_DESC(ctx)) && (callout_pending(&txq->ift_db_check) == 0)) { txq->ift_db_pending_queued = txq->ift_db_pending; callout_reset_on(&txq->ift_db_check, 1, iflib_txd_deferred_db_check, txq, txq->ift_db_check.c_cpu); } if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent); if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent); if (mcast_sent) if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent); return (consumed); } static void _task_fn_tx(void *context, int pending) { iflib_txq_t txq = context; if_ctx_t ctx = txq->ift_ctx; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE); } static void _task_fn_rx(void *context, int pending) { iflib_rxq_t rxq = context; if_ctx_t ctx = rxq->ifr_ctx; bool more; DBG_COUNTER_INC(task_fn_rxs); if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) return; if ((more = iflib_rxeof(rxq, 16 /* XXX */)) == false) { if (ctx->ifc_flags & IFC_LEGACY) IFDI_INTR_ENABLE(ctx); else { DBG_COUNTER_INC(rx_intr_enables); IFDI_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); } } if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) return; if (more) GROUPTASK_ENQUEUE(&rxq->ifr_task); } static void _task_fn_admin(void *context, int pending) { if_ctx_t ctx = context; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; iflib_txq_t txq; int i; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; CTX_LOCK(ctx); for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); } IFDI_UPDATE_ADMIN_STATUS(ctx); for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu); IFDI_LINK_INTR_ENABLE(ctx); CTX_UNLOCK(ctx); if (LINK_ACTIVE(ctx) == 0) return; for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); } static void _task_fn_iov(void *context, int pending) { if_ctx_t ctx = context; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; CTX_LOCK(ctx); IFDI_VFLR_HANDLE(ctx); CTX_UNLOCK(ctx); } static int iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS) { int err; if_int_delay_info_t info; if_ctx_t ctx; info = (if_int_delay_info_t)arg1; ctx = info->iidi_ctx; info->iidi_req = req; info->iidi_oidp = oidp; CTX_LOCK(ctx); err = IFDI_SYSCTL_INT_DELAY(ctx, info); CTX_UNLOCK(ctx); return (err); } /********************************************************************* * * IFNET FUNCTIONS * **********************************************************************/ static void iflib_if_init_locked(if_ctx_t ctx) { iflib_stop(ctx); iflib_init_locked(ctx); } static void iflib_if_init(void *arg) { if_ctx_t ctx = arg; CTX_LOCK(ctx); iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); } static int iflib_if_transmit(if_t ifp, struct mbuf *m) { if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq; struct mbuf *marr[8], **mp, *next; int err, i, count, qidx; if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(tx_frees); m_freem(m); return (0); } qidx = 0; if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m)) qidx = QIDX(ctx, m); /* * XXX calculate buf_ring based on flowid (divvy up bits?) */ txq = &ctx->ifc_txqs[qidx]; #ifdef DRIVER_BACKPRESSURE if (txq->ift_closed) { while (m != NULL) { next = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); m = next; } return (ENOBUFS); } #endif qidx = count = 0; mp = marr; next = m; do { count++; next = next->m_nextpkt; } while (next != NULL); if (count > nitems(marr)) if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) { /* XXX check nextpkt */ m_freem(m); /* XXX simplify for now */ DBG_COUNTER_INC(tx_frees); return (ENOBUFS); } for (next = m, i = 0; next != NULL; i++) { mp[i] = next; next = next->m_nextpkt; mp[i]->m_nextpkt = NULL; } DBG_COUNTER_INC(tx_seen); err = ifmp_ring_enqueue(txq->ift_br[0], (void **)mp, count, TX_BATCH_SIZE); if (iflib_txq_can_drain(txq->ift_br[0])) GROUPTASK_ENQUEUE(&txq->ift_task); if (err) { /* support forthcoming later */ #ifdef DRIVER_BACKPRESSURE txq->ift_closed = TRUE; #endif for (i = 0; i < count; i++) m_freem(mp[i]); ifmp_ring_check_drainage(txq->ift_br[0], TX_BATCH_SIZE); } if (count > nitems(marr)) free(mp, M_IFLIB); return (err); } static void iflib_if_qflush(if_t ifp) { if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq = ctx->ifc_txqs; int i; CTX_LOCK(ctx); ctx->ifc_flags |= IFC_QFLUSH; CTX_UNLOCK(ctx); for (i = 0; i < NTXQSETS(ctx); i++, txq++) while (!(ifmp_ring_is_idle(txq->ift_br[0]) || ifmp_ring_is_stalled(txq->ift_br[0]))) iflib_txq_check_drain(txq, 0); CTX_LOCK(ctx); ctx->ifc_flags &= ~IFC_QFLUSH; CTX_UNLOCK(ctx); if_qflush(ifp); } #define IFCAP_REINIT (IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_TSO6|IFCAP_VLAN_HWTAGGING|IFCAP_VLAN_MTU | \ IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO) #define IFCAP_FLAGS (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | \ IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO) static int iflib_if_ioctl(if_t ifp, u_long command, caddr_t data) { if_ctx_t ctx = if_getsoftc(ifp); struct ifreq *ifr = (struct ifreq *)data; #if defined(INET) || defined(INET6) struct ifaddr *ifa = (struct ifaddr *)data; #endif bool avoid_reset = FALSE; int err = 0, reinit = 0, bits; switch (command) { case SIOCSIFADDR: #ifdef INET if (ifa->ifa_addr->sa_family == AF_INET) avoid_reset = TRUE; #endif #ifdef INET6 if (ifa->ifa_addr->sa_family == AF_INET6) avoid_reset = TRUE; #endif /* ** Calling init results in link renegotiation, ** so we avoid doing it when possible. */ if (avoid_reset) { if_setflagbits(ifp, IFF_UP,0); if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING)) reinit = 1; #ifdef INET if (!(if_getflags(ifp) & IFF_NOARP)) arp_ifinit(ifp, ifa); #endif } else err = ether_ioctl(ifp, command, data); break; case SIOCSIFMTU: CTX_LOCK(ctx); if (ifr->ifr_mtu == if_getmtu(ifp)) { CTX_UNLOCK(ctx); break; } bits = if_getdrvflags(ifp); /* stop the driver and free any clusters before proceeding */ iflib_stop(ctx); if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) { if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size) ctx->ifc_flags |= IFC_MULTISEG; else ctx->ifc_flags &= ~IFC_MULTISEG; err = if_setmtu(ifp, ifr->ifr_mtu); } iflib_init_locked(ctx); if_setdrvflags(ifp, bits); CTX_UNLOCK(ctx); break; case SIOCSIFFLAGS: CTX_LOCK(ctx); if (if_getflags(ifp) & IFF_UP) { if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { if ((if_getflags(ifp) ^ ctx->ifc_if_flags) & (IFF_PROMISC | IFF_ALLMULTI)) { err = IFDI_PROMISC_SET(ctx, if_getflags(ifp)); } } else reinit = 1; } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { iflib_stop(ctx); } ctx->ifc_if_flags = if_getflags(ifp); CTX_UNLOCK(ctx); break; break; case SIOCADDMULTI: case SIOCDELMULTI: if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { CTX_LOCK(ctx); IFDI_INTR_DISABLE(ctx); IFDI_MULTI_SET(ctx); IFDI_INTR_ENABLE(ctx); CTX_UNLOCK(ctx); } break; case SIOCSIFMEDIA: CTX_LOCK(ctx); IFDI_MEDIA_SET(ctx); CTX_UNLOCK(ctx); /* falls thru */ case SIOCGIFMEDIA: err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command); break; case SIOCGI2C: { struct ifi2creq i2c; err = copyin(ifr->ifr_data, &i2c, sizeof(i2c)); if (err != 0) break; if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) { err = EINVAL; break; } if (i2c.len > sizeof(i2c.data)) { err = EINVAL; break; } if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0) err = copyout(&i2c, ifr->ifr_data, sizeof(i2c)); break; } case SIOCSIFCAP: { int mask, setmask; mask = ifr->ifr_reqcap ^ if_getcapenable(ifp); setmask = 0; #ifdef TCP_OFFLOAD setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6); #endif setmask |= (mask & IFCAP_FLAGS); if ((mask & IFCAP_WOL) && (if_getcapabilities(ifp) & IFCAP_WOL) != 0) setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC)); if_vlancap(ifp); /* * want to ensure that traffic has stopped before we change any of the flags */ if (setmask) { CTX_LOCK(ctx); bits = if_getdrvflags(ifp); if (setmask & IFCAP_REINIT) iflib_stop(ctx); if_togglecapenable(ifp, setmask); if (setmask & IFCAP_REINIT) iflib_init_locked(ctx); if_setdrvflags(ifp, bits); CTX_UNLOCK(ctx); } break; } case SIOCGPRIVATE_0: case SIOCSDRVSPEC: case SIOCGDRVSPEC: CTX_LOCK(ctx); err = IFDI_PRIV_IOCTL(ctx, command, data); CTX_UNLOCK(ctx); break; default: err = ether_ioctl(ifp, command, data); break; } if (reinit) iflib_if_init(ctx); return (err); } static uint64_t iflib_if_get_counter(if_t ifp, ift_counter cnt) { if_ctx_t ctx = if_getsoftc(ifp); return (IFDI_GET_COUNTER(ctx, cnt)); } /********************************************************************* * * OTHER FUNCTIONS EXPORTED TO THE STACK * **********************************************************************/ static void iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag) { if_ctx_t ctx = if_getsoftc(ifp); if ((void *)ctx != arg) return; if ((vtag == 0) || (vtag > 4095)) return; CTX_LOCK(ctx); IFDI_VLAN_REGISTER(ctx, vtag); /* Re-init to load the changes */ if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER) iflib_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag) { if_ctx_t ctx = if_getsoftc(ifp); if ((void *)ctx != arg) return; if ((vtag == 0) || (vtag > 4095)) return; CTX_LOCK(ctx); IFDI_VLAN_UNREGISTER(ctx, vtag); /* Re-init to load the changes */ if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER) iflib_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_led_func(void *arg, int onoff) { if_ctx_t ctx = arg; CTX_LOCK(ctx); IFDI_LED_FUNC(ctx, onoff); CTX_UNLOCK(ctx); } /********************************************************************* * * BUS FUNCTION DEFINITIONS * **********************************************************************/ int iflib_device_probe(device_t dev) { pci_vendor_info_t *ent; uint16_t pci_vendor_id, pci_device_id; uint16_t pci_subvendor_id, pci_subdevice_id; uint16_t pci_rev_id; if_shared_ctx_t sctx; if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) return (ENOTSUP); pci_vendor_id = pci_get_vendor(dev); pci_device_id = pci_get_device(dev); pci_subvendor_id = pci_get_subvendor(dev); pci_subdevice_id = pci_get_subdevice(dev); pci_rev_id = pci_get_revid(dev); if (sctx->isc_parse_devinfo != NULL) sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id); ent = sctx->isc_vendor_info; while (ent->pvi_vendor_id != 0) { if (pci_vendor_id != ent->pvi_vendor_id) { ent++; continue; } if ((pci_device_id == ent->pvi_device_id) && ((pci_subvendor_id == ent->pvi_subvendor_id) || (ent->pvi_subvendor_id == 0)) && ((pci_subdevice_id == ent->pvi_subdevice_id) || (ent->pvi_subdevice_id == 0)) && ((pci_rev_id == ent->pvi_rev_id) || (ent->pvi_rev_id == 0))) { device_set_desc_copy(dev, ent->pvi_name); /* this needs to be changed to zero if the bus probing code * ever stops re-probing on best match because the sctx * may have its values over written by register calls * in subsequent probes */ return (BUS_PROBE_DEFAULT); } ent++; } return (ENXIO); } int iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp) { int err, rid, msix, msix_bar; if_ctx_t ctx; if_t ifp; if_softc_ctx_t scctx; ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO); if (sc == NULL) { sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); device_set_softc(dev, ctx); } ctx->ifc_sctx = sctx; ctx->ifc_dev = dev; ctx->ifc_txrx = *sctx->isc_txrx; ctx->ifc_softc = sc; if ((err = iflib_register(ctx)) != 0) { device_printf(dev, "iflib_register failed %d\n", err); return (err); } iflib_add_device_sysctl_pre(ctx); if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); return (err); } #ifdef ACPI_DMAR if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL) ctx->ifc_flags |= IFC_DMAR; #endif scctx = &ctx->ifc_softc_ctx; msix_bar = scctx->isc_msix_bar; if (scctx->isc_tx_nsegments > sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_nsegments = max(1, sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION); if (scctx->isc_tx_tso_segments_max > sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_tso_segments_max = max(1, sctx->isc_ntxd / MAX_SINGLE_PACKET_FRACTION); ifp = ctx->ifc_ifp; /* * XXX sanity check that ntxd & nrxd are a power of 2 */ /* * Protect the stack against modern hardware */ if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX) scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX; /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max; ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max; ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max; if (scctx->isc_rss_table_size == 0) scctx->isc_rss_table_size = 64; scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;; /* ** Now setup MSI or MSI/X, should ** return us the number of supported ** vectors. (Will be 1 for MSI) */ if (sctx->isc_flags & IFLIB_SKIP_MSIX) { msix = scctx->isc_vectors; } else if (scctx->isc_msix_bar != 0) msix = iflib_msix_init(ctx); else { scctx->isc_vectors = 1; scctx->isc_ntxqsets = 1; scctx->isc_nrxqsets = 1; scctx->isc_intr = IFLIB_INTR_LEGACY; msix = 0; } /* Get memory for the station queues */ if ((err = iflib_queues_alloc(ctx))) { device_printf(dev, "Unable to allocate queue memory\n"); goto fail; } if ((err = iflib_qset_structures_setup(ctx))) { device_printf(dev, "qset structure setup failed %d\n", err); goto fail_queues; } if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) { device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err); goto fail_intr_free; } if (msix <= 1) { rid = 0; if (scctx->isc_intr == IFLIB_INTR_MSI) { MPASS(msix == 1); rid = 1; } if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx, &rid, "irq0")) != 0) { device_printf(dev, "iflib_legacy_setup failed %d\n", err); goto fail_intr_free; } } ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac); if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } if ((err = iflib_netmap_attach(ctx))) { device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err); goto fail_detach; } *ctxp = ctx; iflib_add_device_sysctl_post(ctx); return (0); fail_detach: ether_ifdetach(ctx->ifc_ifp); fail_intr_free: if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI) pci_release_msi(ctx->ifc_dev); fail_queues: /* XXX free queues */ fail: IFDI_DETACH(ctx); return (err); } int iflib_device_attach(device_t dev) { if_ctx_t ctx; if_shared_ctx_t sctx; if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) return (ENOTSUP); pci_enable_busmaster(dev); return (iflib_device_register(dev, NULL, sctx, &ctx)); } int iflib_device_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq; iflib_rxq_t rxq; device_t dev = ctx->ifc_dev; int i; struct taskqgroup *tqg; /* Make sure VLANS are not using driver */ if (if_vlantrunkinuse(ifp)) { device_printf(dev,"Vlan in use, detach first\n"); return (EBUSY); } CTX_LOCK(ctx); ctx->ifc_in_detach = 1; iflib_stop(ctx); CTX_UNLOCK(ctx); /* Unregister VLAN events */ if (ctx->ifc_vlan_attach_event != NULL) EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event); if (ctx->ifc_vlan_detach_event != NULL) EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event); iflib_netmap_detach(ifp); ether_ifdetach(ifp); /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ CTX_LOCK_DESTROY(ctx); if (ctx->ifc_led_dev != NULL) led_destroy(ctx->ifc_led_dev); /* XXX drain any dependent tasks */ tqg = qgroup_if_io_tqg; for (txq = ctx->ifc_txqs, i = 0, rxq = ctx->ifc_rxqs; i < NTXQSETS(ctx); i++, txq++) { callout_drain(&txq->ift_timer); callout_drain(&txq->ift_db_check); if (txq->ift_task.gt_uniq != NULL) taskqgroup_detach(tqg, &txq->ift_task); } for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { if (rxq->ifr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &rxq->ifr_task); } tqg = qgroup_if_config_tqg; if (ctx->ifc_admin_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_admin_task); if (ctx->ifc_vflr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_vflr_task); IFDI_DETACH(ctx); if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) { pci_release_msi(dev); } if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) { iflib_irq_free(ctx, &ctx->ifc_legacy_irq); } if (ctx->ifc_msix_mem != NULL) { bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY, ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } bus_generic_detach(dev); if_free(ifp); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); return (0); } int iflib_device_detach(device_t dev) { if_ctx_t ctx = device_get_softc(dev); return (iflib_device_deregister(ctx)); } int iflib_device_suspend(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_SUSPEND(ctx); CTX_UNLOCK(ctx); return bus_generic_suspend(dev); } int iflib_device_shutdown(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_SHUTDOWN(ctx); CTX_UNLOCK(ctx); return bus_generic_suspend(dev); } int iflib_device_resume(device_t dev) { if_ctx_t ctx = device_get_softc(dev); iflib_txq_t txq = ctx->ifc_txqs; CTX_LOCK(ctx); IFDI_RESUME(ctx); iflib_init_locked(ctx); CTX_UNLOCK(ctx); for (int i = 0; i < NTXQSETS(ctx); i++, txq++) iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); return (bus_generic_resume(dev)); } int iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params) { int error; if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); error = IFDI_IOV_INIT(ctx, num_vfs, params); CTX_UNLOCK(ctx); return (error); } void iflib_device_iov_uninit(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_IOV_UNINIT(ctx); CTX_UNLOCK(ctx); } int iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params) { int error; if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); error = IFDI_IOV_VF_ADD(ctx, vfnum, params); CTX_UNLOCK(ctx); return (error); } /********************************************************************* * * MODULE FUNCTION DEFINITIONS * **********************************************************************/ /* * - Start a fast taskqueue thread for each core * - Start a taskqueue for control operations */ static int iflib_module_init(void) { return (0); } static int iflib_module_event_handler(module_t mod, int what, void *arg) { int err; switch (what) { case MOD_LOAD: if ((err = iflib_module_init()) != 0) return (err); break; case MOD_UNLOAD: return (EBUSY); default: return (EOPNOTSUPP); } return (0); } /********************************************************************* * * PUBLIC FUNCTION DEFINITIONS * ordered as in iflib.h * **********************************************************************/ static void _iflib_assert(if_shared_ctx_t sctx) { MPASS(sctx->isc_tx_maxsize); MPASS(sctx->isc_tx_maxsegsize); MPASS(sctx->isc_rx_maxsize); MPASS(sctx->isc_rx_nsegments); MPASS(sctx->isc_rx_maxsegsize); MPASS(sctx->isc_txrx->ift_txd_encap); MPASS(sctx->isc_txrx->ift_txd_flush); MPASS(sctx->isc_txrx->ift_txd_credits_update); MPASS(sctx->isc_txrx->ift_rxd_available); MPASS(sctx->isc_txrx->ift_rxd_pkt_get); MPASS(sctx->isc_txrx->ift_rxd_refill); MPASS(sctx->isc_txrx->ift_rxd_flush); MPASS(sctx->isc_nrxd); } static int iflib_register(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; driver_t *driver = sctx->isc_driver; device_t dev = ctx->ifc_dev; if_t ifp; _iflib_assert(sctx); CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev)); MPASS(ctx->ifc_flags == 0); ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER); if (ifp == NULL) { device_printf(dev, "can not allocate ifnet structure\n"); return (ENOMEM); } /* * Initialize our context's device specific methods */ kobj_init((kobj_t) ctx, (kobj_class_t) driver); kobj_class_compile((kobj_class_t) driver); driver->refs++; if_initname(ifp, device_get_name(dev), device_get_unit(dev)); if_setsoftc(ifp, ctx); if_setdev(ifp, dev); if_setinitfn(ifp, iflib_if_init); if_setioctlfn(ifp, iflib_if_ioctl); if_settransmitfn(ifp, iflib_if_transmit); if_setqflushfn(ifp, iflib_if_qflush); if_setgetcounterfn(ifp, iflib_if_get_counter); if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); if_setcapabilities(ifp, 0); if_setcapenable(ifp, 0); ctx->ifc_vlan_attach_event = EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx, EVENTHANDLER_PRI_FIRST); ctx->ifc_vlan_detach_event = EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx, EVENTHANDLER_PRI_FIRST); ifmedia_init(&ctx->ifc_media, IFM_IMASK, iflib_media_change, iflib_media_status); return (0); } static int iflib_queues_alloc(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; device_t dev = ctx->ifc_dev; int nrxqsets = ctx->ifc_softc_ctx.isc_nrxqsets; int ntxqsets = ctx->ifc_softc_ctx.isc_ntxqsets; iflib_txq_t txq; iflib_rxq_t rxq; iflib_fl_t fl = NULL; - int i, j, err, txconf, rxconf, fl_ifdi_offset; + int i, j, cpu, err, txconf, rxconf, fl_ifdi_offset; iflib_dma_info_t ifdip; uint32_t *rxqsizes = sctx->isc_rxqsizes; uint32_t *txqsizes = sctx->isc_txqsizes; uint8_t nrxqs = sctx->isc_nrxqs; uint8_t ntxqs = sctx->isc_ntxqs; int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1; caddr_t *vaddrs; uint64_t *paddrs; struct ifmp_ring **brscp; int nbuf_rings = 1; /* XXX determine dynamically */ KASSERT(ntxqs > 0, ("number of queues must be at least 1")); KASSERT(nrxqs > 0, ("number of queues must be at least 1")); brscp = NULL; rxq = NULL; /* Allocate the TX ring struct memory */ if (!(txq = (iflib_txq_t) malloc(sizeof(struct iflib_txq) * ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate TX ring memory\n"); err = ENOMEM; goto fail; } /* Now allocate the RX */ if (!(rxq = (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX ring memory\n"); err = ENOMEM; goto rx_fail; } if (!(brscp = malloc(sizeof(void *) * nbuf_rings * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to buf_ring_sc * memory\n"); err = ENOMEM; goto rx_fail; } ctx->ifc_txqs = txq; ctx->ifc_rxqs = rxq; txq = NULL; rxq = NULL; /* * XXX handle allocation failure */ - for (txconf = i = 0; i < ntxqsets; i++, txconf++, txq++) { + for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) { /* Set up some basics */ if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) { device_printf(dev, "failed to allocate iflib_dma_info\n"); err = ENOMEM; goto err_tx_desc; } txq->ift_ifdi = ifdip; for (j = 0; j < ntxqs; j++, ifdip++) { if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) { device_printf(dev, "Unable to allocate Descriptor memory\n"); err = ENOMEM; goto err_tx_desc; } bzero((void *)ifdip->idi_vaddr, txqsizes[j]); } txq->ift_ctx = ctx; txq->ift_id = i; /* XXX fix this */ - txq->ift_timer.c_cpu = i % mp_ncpus; - txq->ift_db_check.c_cpu = i % mp_ncpus; + txq->ift_timer.c_cpu = cpu; + txq->ift_db_check.c_cpu = cpu; txq->ift_nbr = nbuf_rings; if (iflib_txsd_alloc(txq)) { device_printf(dev, "Critical Failure setting up TX buffers\n"); err = ENOMEM; goto err_tx_desc; } /* Initialize the TX lock */ snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout", device_get_nameunit(dev), txq->ift_id); mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF); callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0); callout_init_mtx(&txq->ift_db_check, &txq->ift_mtx, 0); snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db", device_get_nameunit(dev), txq->ift_id); TXDB_LOCK_INIT(txq); txq->ift_br = brscp + i*nbuf_rings; for (j = 0; j < nbuf_rings; j++) { err = ifmp_ring_alloc(&txq->ift_br[j], 2048, txq, iflib_txq_drain, iflib_txq_can_drain, M_IFLIB, M_WAITOK); if (err) { /* XXX free any allocated rings */ device_printf(dev, "Unable to allocate buf_ring\n"); goto err_tx_desc; } } } for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) { /* Set up some basics */ if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) { device_printf(dev, "failed to allocate iflib_dma_info\n"); err = ENOMEM; goto err_tx_desc; } rxq->ifr_ifdi = ifdip; for (j = 0; j < nrxqs; j++, ifdip++) { if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) { device_printf(dev, "Unable to allocate Descriptor memory\n"); err = ENOMEM; goto err_tx_desc; } bzero((void *)ifdip->idi_vaddr, rxqsizes[j]); } rxq->ifr_ctx = ctx; rxq->ifr_id = i; if (sctx->isc_flags & IFLIB_HAS_CQ) { fl_ifdi_offset = 1; } else { fl_ifdi_offset = 0; } rxq->ifr_nfl = nfree_lists; if (!(fl = (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate free list memory\n"); err = ENOMEM; goto err_tx_desc; } rxq->ifr_fl = fl; for (j = 0; j < nfree_lists; j++) { rxq->ifr_fl[j].ifl_rxq = rxq; rxq->ifr_fl[j].ifl_id = j; rxq->ifr_fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + fl_ifdi_offset]; } /* Allocate receive buffers for the ring*/ if (iflib_rxsd_alloc(rxq)) { device_printf(dev, "Critical Failure setting up receive buffers\n"); err = ENOMEM; goto err_rx_desc; } } /* TXQs */ vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); for (i = 0; i < ntxqsets; i++) { iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi; for (j = 0; j < ntxqs; j++, di++) { vaddrs[i*ntxqs + j] = di->idi_vaddr; paddrs[i*ntxqs + j] = di->idi_paddr; } } if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) { device_printf(ctx->ifc_dev, "device queue allocation failed\n"); iflib_tx_structures_free(ctx); free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); goto err_rx_desc; } free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); /* RXQs */ vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); for (i = 0; i < nrxqsets; i++) { iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi; for (j = 0; j < nrxqs; j++, di++) { vaddrs[i*nrxqs + j] = di->idi_vaddr; paddrs[i*nrxqs + j] = di->idi_paddr; } } if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) { device_printf(ctx->ifc_dev, "device queue allocation failed\n"); iflib_tx_structures_free(ctx); free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); goto err_rx_desc; } free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); return (0); /* XXX handle allocation failure changes */ err_rx_desc: err_tx_desc: if (ctx->ifc_rxqs != NULL) free(ctx->ifc_rxqs, M_IFLIB); ctx->ifc_rxqs = NULL; if (ctx->ifc_txqs != NULL) free(ctx->ifc_txqs, M_IFLIB); ctx->ifc_txqs = NULL; rx_fail: if (brscp != NULL) free(brscp, M_IFLIB); if (rxq != NULL) free(rxq, M_IFLIB); if (txq != NULL) free(txq, M_IFLIB); fail: return (err); } static int iflib_tx_structures_setup(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; int i; for (i = 0; i < NTXQSETS(ctx); i++, txq++) iflib_txq_setup(txq); return (0); } static void iflib_tx_structures_free(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; int i, j; for (i = 0; i < NTXQSETS(ctx); i++, txq++) { iflib_txq_destroy(txq); for (j = 0; j < ctx->ifc_nhwtxqs; j++) iflib_dma_free(&txq->ift_ifdi[j]); } free(ctx->ifc_txqs, M_IFLIB); ctx->ifc_txqs = NULL; IFDI_QUEUES_FREE(ctx); } /********************************************************************* * * Initialize all receive rings. * **********************************************************************/ static int iflib_rx_structures_setup(if_ctx_t ctx) { iflib_rxq_t rxq = ctx->ifc_rxqs; int q; #if defined(INET6) || defined(INET) int i, err; #endif for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) { #if defined(INET6) || defined(INET) tcp_lro_free(&rxq->ifr_lc); if ((err = tcp_lro_init(&rxq->ifr_lc)) != 0) { device_printf(ctx->ifc_dev, "LRO Initialization failed!\n"); goto fail; } rxq->ifr_lro_enabled = TRUE; rxq->ifr_lc.ifp = ctx->ifc_ifp; #endif IFDI_RXQ_SETUP(ctx, rxq->ifr_id); } return (0); #if defined(INET6) || defined(INET) fail: /* * Free RX software descriptors allocated so far, we will only handle * the rings that completed, the failing case will have * cleaned up for itself. 'q' failed, so its the terminus. */ rxq = ctx->ifc_rxqs; for (i = 0; i < q; ++i, rxq++) { iflib_rx_sds_free(rxq); rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0; } return (err); #endif } /********************************************************************* * * Free all receive rings. * **********************************************************************/ static void iflib_rx_structures_free(if_ctx_t ctx) { iflib_rxq_t rxq = ctx->ifc_rxqs; for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, rxq++) { iflib_rx_sds_free(rxq); } } static int iflib_qset_structures_setup(if_ctx_t ctx) { int err; if ((err = iflib_tx_structures_setup(ctx)) != 0) return (err); if ((err = iflib_rx_structures_setup(ctx)) != 0) { device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); } return (err); } int iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name) { return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name)); } static void find_nth(if_ctx_t ctx, cpuset_t *cpus, int qid) { int i, cpuid; CPU_COPY(&ctx->ifc_cpus, cpus); /* clear up to the qid'th bit */ for (i = 0; i < qid; i++) { cpuid = CPU_FFS(cpus); CPU_CLR(cpuid, cpus); } } int iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid, iflib_intr_type_t type, driver_filter_t *filter, void *filter_arg, int qid, char *name) { struct grouptask *gtask; struct taskqgroup *tqg; iflib_filter_info_t info; cpuset_t cpus; task_fn_t *fn; int tqrid, err; void *q; info = &ctx->ifc_filter_info; switch (type) { /* XXX merge tx/rx for netmap? */ case IFLIB_INTR_TX: q = &ctx->ifc_txqs[qid]; info = &ctx->ifc_txqs[qid].ift_filter_info; gtask = &ctx->ifc_txqs[qid].ift_task; tqg = qgroup_if_io_tqg; tqrid = irq->ii_rid; fn = _task_fn_tx; break; case IFLIB_INTR_RX: q = &ctx->ifc_rxqs[qid]; info = &ctx->ifc_rxqs[qid].ifr_filter_info; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; tqrid = irq->ii_rid; fn = _task_fn_rx; break; case IFLIB_INTR_ADMIN: q = ctx; info = &ctx->ifc_filter_info; gtask = &ctx->ifc_admin_task; tqg = qgroup_if_config_tqg; tqrid = -1; fn = _task_fn_admin; break; default: panic("unknown net intr type"); } GROUPTASK_INIT(gtask, 0, fn, q); info->ifi_filter = filter; info->ifi_filter_arg = filter_arg; info->ifi_task = gtask; /* XXX query cpu that rid belongs to */ err = _iflib_irq_alloc(ctx, irq, rid, iflib_fast_intr, NULL, info, name); if (err != 0) return (err); if (tqrid != -1) { find_nth(ctx, &cpus, qid); taskqgroup_attach_cpu(tqg, gtask, q, CPU_FFS(&cpus), irq->ii_rid, name); } else taskqgroup_attach(tqg, gtask, q, tqrid, name); return (0); } void iflib_softirq_alloc_generic(if_ctx_t ctx, int rid, iflib_intr_type_t type, void *arg, int qid, char *name) { struct grouptask *gtask; struct taskqgroup *tqg; task_fn_t *fn; void *q; switch (type) { case IFLIB_INTR_TX: q = &ctx->ifc_txqs[qid]; gtask = &ctx->ifc_txqs[qid].ift_task; tqg = qgroup_if_io_tqg; fn = _task_fn_tx; break; case IFLIB_INTR_RX: q = &ctx->ifc_rxqs[qid]; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; break; case IFLIB_INTR_ADMIN: q = ctx; gtask = &ctx->ifc_admin_task; tqg = qgroup_if_config_tqg; rid = -1; fn = _task_fn_admin; break; case IFLIB_INTR_IOV: q = ctx; gtask = &ctx->ifc_vflr_task; tqg = qgroup_if_config_tqg; rid = -1; fn = _task_fn_iov; break; default: panic("unknown net intr type"); } GROUPTASK_INIT(gtask, 0, fn, q); taskqgroup_attach(tqg, gtask, q, rid, name); } void iflib_irq_free(if_ctx_t ctx, if_irq_t irq) { if (irq->ii_tag) bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag); if (irq->ii_res) bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res); } static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name) { iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_irq_t irq = &ctx->ifc_legacy_irq; iflib_filter_info_t info; struct grouptask *gtask; struct taskqgroup *tqg; task_fn_t *fn; int tqrid; void *q; int err; q = &ctx->ifc_rxqs[0]; info = &rxq[0].ifr_filter_info; gtask = &rxq[0].ifr_task; tqg = qgroup_if_io_tqg; tqrid = irq->ii_rid = *rid; fn = _task_fn_rx; ctx->ifc_flags |= IFC_LEGACY; info->ifi_filter = filter; info->ifi_filter_arg = filter_arg; info->ifi_task = gtask; /* We allocate a single interrupt resource */ if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr, NULL, info, name)) != 0) return (err); GROUPTASK_INIT(gtask, 0, fn, q); taskqgroup_attach(tqg, gtask, q, tqrid, name); GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq); taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, tqrid, "tx"); GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin/link"); return (0); } void iflib_led_create(if_ctx_t ctx) { ctx->ifc_led_dev = led_create(iflib_led_func, ctx, device_get_nameunit(ctx->ifc_dev)); } void iflib_tx_intr_deferred(if_ctx_t ctx, int txqid) { GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task); } void iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid) { GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task); } void iflib_admin_intr_deferred(if_ctx_t ctx) { GROUPTASK_ENQUEUE(&ctx->ifc_admin_task); } void iflib_iov_intr_deferred(if_ctx_t ctx) { GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task); } void iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name) { taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name); } void iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, task_fn_t *fn, char *name) { GROUPTASK_INIT(gtask, 0, fn, ctx); taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name); } void iflib_link_state_change(if_ctx_t ctx, int link_state) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq = ctx->ifc_txqs; #if 0 if_setbaudrate(ifp, baudrate); #endif /* If link down, disable watchdog */ if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) { for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++) txq->ift_qstatus = IFLIB_QUEUE_IDLE; } ctx->ifc_link_state = link_state; if_link_state_change(ifp, link_state); } static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq) { int credits; if (ctx->isc_txd_credits_update == NULL) return (0); if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, txq->ift_cidx_processed, true)) == 0) return (0); txq->ift_processed += credits; txq->ift_cidx_processed += credits; if (txq->ift_cidx_processed >= txq->ift_size) txq->ift_cidx_processed -= txq->ift_size; return (credits); } static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, int cidx) { return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx)); } void iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name, const char *description, if_int_delay_info_t info, int offset, int value) { info->iidi_ctx = ctx; info->iidi_offset = offset; info->iidi_value = value; SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev), SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)), OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW, info, 0, iflib_sysctl_int_delay, "I", description); } struct mtx * iflib_ctx_lock_get(if_ctx_t ctx) { return (&ctx->ifc_mtx); } static int iflib_msix_init(if_ctx_t ctx) { device_t dev = ctx->ifc_dev; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs; int iflib_num_tx_queues, iflib_num_rx_queues; int err, admincnt, bar; iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs; iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs; bar = ctx->ifc_softc_ctx.isc_msix_bar; admincnt = sctx->isc_admin_intrcnt; /* Override by tuneable */ if (enable_msix == 0) goto msi; /* ** When used in a virtualized environment ** PCI BUSMASTER capability may not be set ** so explicity set it here and rewrite ** the ENABLE in the MSIX control register ** at this point to cause the host to ** successfully initialize us. */ { uint16_t pci_cmd_word; int msix_ctrl, rid; rid = 0; pci_cmd_word = pci_read_config(dev, PCIR_COMMAND, 2); pci_cmd_word |= PCIM_CMD_BUSMASTEREN; pci_write_config(dev, PCIR_COMMAND, pci_cmd_word, 2); pci_find_cap(dev, PCIY_MSIX, &rid); rid += PCIR_MSIX_CTRL; msix_ctrl = pci_read_config(dev, rid, 2); msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE; pci_write_config(dev, rid, msix_ctrl, 2); } /* * bar == -1 => "trust me I know what I'm doing" * https://www.youtube.com/watch?v=nnwWKkNau4I * Some drivers are for hardware that is so shoddily * documented that no one knows which bars are which * so the developer has to map all bars. This hack * allows shoddy garbage to use msix in this framework. */ if (bar != -1) { ctx->ifc_msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &bar, RF_ACTIVE); if (ctx->ifc_msix_mem == NULL) { /* May not be enabled */ device_printf(dev, "Unable to map MSIX table \n"); goto msi; } } /* First try MSI/X */ if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */ device_printf(dev, "System has MSIX disabled \n"); bus_release_resource(dev, SYS_RES_MEMORY, bar, ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; goto msi; } #if IFLIB_DEBUG /* use only 1 qset in debug mode */ queuemsgs = min(msgs - admincnt, 1); #else queuemsgs = msgs - admincnt; #endif if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) == 0) { #ifdef RSS queues = imin(queuemsgs, rss_getnumbuckets()); #else queues = queuemsgs; #endif queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues); device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n", CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt); } else { device_printf(dev, "Unable to fetch CPU list\n"); /* Figure out a reasonable auto config value */ queues = min(queuemsgs, mp_ncpus); } #ifdef RSS /* If we're doing RSS, clamp at the number of RSS buckets */ if (queues > rss_getnumbuckets()) queues = rss_getnumbuckets(); #endif if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queues) queues = rx_queues = iflib_num_rx_queues; else rx_queues = queues; if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues) tx_queues = iflib_num_tx_queues; else tx_queues = queues; device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues); vectors = queues + admincnt; if ((err = pci_alloc_msix(dev, &vectors)) == 0) { device_printf(dev, "Using MSIX interrupts with %d vectors\n", vectors); scctx->isc_vectors = vectors; scctx->isc_nrxqsets = rx_queues; scctx->isc_ntxqsets = tx_queues; scctx->isc_intr = IFLIB_INTR_MSIX; return (vectors); } else { device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err); } msi: vectors = pci_msi_count(dev); scctx->isc_nrxqsets = 1; scctx->isc_ntxqsets = 1; scctx->isc_vectors = vectors; if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) { device_printf(dev,"Using an MSI interrupt\n"); scctx->isc_intr = IFLIB_INTR_MSI; } else { device_printf(dev,"Using a Legacy interrupt\n"); scctx->isc_intr = IFLIB_INTR_LEGACY; } return (vectors); } char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" }; static int mp_ring_state_handler(SYSCTL_HANDLER_ARGS) { int rc; uint16_t *state = ((uint16_t *)oidp->oid_arg1); struct sbuf *sb; char *ring_state = "UNKNOWN"; /* XXX needed ? */ rc = sysctl_wire_old_buffer(req, 0); MPASS(rc == 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 80, req); MPASS(sb != NULL); if (sb == NULL) return (ENOMEM); if (state[3] <= 3) ring_state = ring_states[state[3]]; sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s", state[0], state[1], state[2], ring_state); rc = sbuf_finish(sb); sbuf_delete(sb); return(rc); } #define NAME_BUFLEN 32 static void iflib_add_device_sysctl_pre(if_ctx_t ctx) { device_t dev = iflib_get_dev(ctx); struct sysctl_oid_list *child, *oid_list; struct sysctl_ctx_list *ctx_list; struct sysctl_oid *node; ctx_list = device_get_sysctl_ctx(dev); child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib", CTLFLAG_RD, NULL, "IFLIB fields"); oid_list = SYSCTL_CHILDREN(node); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, "# of txqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, "# of txqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxds", CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxds, 0, "# of tx descriptors to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxds", CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxds, 0, "# of rx descriptors to use, 0 => use default #"); } static void iflib_add_device_sysctl_post(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = iflib_get_dev(ctx); struct sysctl_oid_list *child; struct sysctl_ctx_list *ctx_list; iflib_fl_t fl; iflib_txq_t txq; iflib_rxq_t rxq; int i, j; char namebuf[NAME_BUFLEN]; char *qfmt; struct sysctl_oid *queue_node, *fl_node, *node; struct sysctl_oid_list *queue_list, *fl_list; ctx_list = device_get_sysctl_ctx(dev); node = ctx->ifc_sysctl_node; child = SYSCTL_CHILDREN(node); if (scctx->isc_ntxqsets > 100) qfmt = "txq%03d"; else if (scctx->isc_ntxqsets > 10) qfmt = "txq%02d"; else qfmt = "txq%d"; for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) { snprintf(namebuf, NAME_BUFLEN, qfmt, i); queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); #if MEMORY_LOGGING SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued", CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued", CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued"); #endif SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag", CTLFLAG_RD, &txq->ift_mbuf_defrag, "# of times m_defrag was called"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups", CTLFLAG_RD, &txq->ift_pullups, "# of times m_pullup was called"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed", CTLFLAG_RD, &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail", CTLFLAG_RD, &txq->ift_mbuf_defrag_failed, "# of times no descriptors were available"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed, "# of times dma map failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx", CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx", CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use", CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed", CTLFLAG_RD, &txq->ift_processed, "descriptors procesed for clean"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned", CTLFLAG_RD, &txq->ift_cleaned, "total cleaned"); SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state", CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br[0]->state), 0, mp_ring_state_handler, "A", "soft ring state"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues", CTLFLAG_RD, &txq->ift_br[0]->enqueues, "# of enqueues to the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops", CTLFLAG_RD, &txq->ift_br[0]->drops, "# of drops in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts", CTLFLAG_RD, &txq->ift_br[0]->starts, "# of normal consumer starts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls", CTLFLAG_RD, &txq->ift_br[0]->stalls, "# of consumer stalls in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts", CTLFLAG_RD, &txq->ift_br[0]->restarts, "# of consumer restarts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications", CTLFLAG_RD, &txq->ift_br[0]->abdications, "# of consumer abdications in the mp_ring for this queue"); } if (scctx->isc_nrxqsets > 100) qfmt = "rxq%03d"; else if (scctx->isc_nrxqsets > 10) qfmt = "rxq%02d"; else qfmt = "rxq%d"; for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) { snprintf(namebuf, NAME_BUFLEN, qfmt, i); queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); if (sctx->isc_flags & IFLIB_HAS_CQ) { SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx", CTLFLAG_RD, &rxq->ifr_cq_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1, "Consumer Index"); } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j); fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf, CTLFLAG_RD, NULL, "freelist Name"); fl_list = SYSCTL_CHILDREN(fl_node); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx", CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx", CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits", CTLFLAG_RD, &fl->ifl_credits, 1, "credits available"); #if MEMORY_LOGGING SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued, "mbufs allocated"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued, "mbufs freed"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued, "clusters allocated"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued, "clusters freed"); #endif } } } Index: head/sys/netinet/ip_id.c =================================================================== --- head/sys/netinet/ip_id.c (revision 302371) +++ head/sys/netinet/ip_id.c (revision 302372) @@ -1,297 +1,299 @@ /*- * Copyright (c) 2008 Michael J. Silbersack. * 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 unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * IP ID generation is a fascinating topic. * * In order to avoid ID collisions during packet reassembly, common sense * dictates that the period between reuse of IDs be as large as possible. * This leads to the classic implementation of a system-wide counter, thereby * ensuring that IDs repeat only once every 2^16 packets. * * Subsequent security researchers have pointed out that using a global * counter makes ID values predictable. This predictability allows traffic * analysis, idle scanning, and even packet injection in specific cases. * These results suggest that IP IDs should be as random as possible. * * The "searchable queues" algorithm used in this IP ID implementation was * proposed by Amit Klein. It is a compromise between the above two * viewpoints that has provable behavior that can be tuned to the user's * requirements. * * The basic concept is that we supplement a standard random number generator * with a queue of the last L IDs that we have handed out to ensure that all * IDs have a period of at least L. * * To efficiently implement this idea, we keep two data structures: a * circular array of IDs of size L and a bitstring of 65536 bits. * * To start, we ask the RNG for a new ID. A quick index into the bitstring * is used to determine if this is a recently used value. The process is * repeated until a value is returned that is not in the bitstring. * * Having found a usable ID, we remove the ID stored at the current position * in the queue from the bitstring and replace it with our new ID. Our new * ID is then added to the bitstring and the queue pointer is incremented. * * The lower limit of 512 was chosen because there doesn't seem to be much * point to having a smaller value. The upper limit of 32768 was chosen for * two reasons. First, every step above 32768 decreases the entropy. Taken * to an extreme, 65533 would offer 1 bit of entropy. Second, the number of * attempts it takes the algorithm to find an unused ID drastically * increases, killing performance. The default value of 8192 was chosen * because it provides a good tradeoff between randomness and non-repetition. * * With L=8192, the queue will use 16K of memory. The bitstring always * uses 8K of memory. No memory is allocated until the use of random ids is * enabled. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* * By default we generate IP ID only for non-atomic datagrams, as * suggested by RFC6864. We use per-CPU counter for that, or if * user wants to, we can turn on random ID generation. */ static VNET_DEFINE(int, ip_rfc6864) = 1; static VNET_DEFINE(int, ip_do_randomid) = 0; #define V_ip_rfc6864 VNET(ip_rfc6864) #define V_ip_do_randomid VNET(ip_do_randomid) /* * Random ID state engine. */ static MALLOC_DEFINE(M_IPID, "ipid", "randomized ip id state"); static VNET_DEFINE(uint16_t *, id_array); static VNET_DEFINE(bitstr_t *, id_bits); static VNET_DEFINE(int, array_ptr); static VNET_DEFINE(int, array_size); static VNET_DEFINE(int, random_id_collisions); static VNET_DEFINE(int, random_id_total); static VNET_DEFINE(struct mtx, ip_id_mtx); #define V_id_array VNET(id_array) #define V_id_bits VNET(id_bits) #define V_array_ptr VNET(array_ptr) #define V_array_size VNET(array_size) #define V_random_id_collisions VNET(random_id_collisions) #define V_random_id_total VNET(random_id_total) #define V_ip_id_mtx VNET(ip_id_mtx) /* * Non-random ID state engine is simply a per-cpu counter. */ static VNET_DEFINE(counter_u64_t, ip_id); #define V_ip_id VNET(ip_id) static int sysctl_ip_randomid(SYSCTL_HANDLER_ARGS); static int sysctl_ip_id_change(SYSCTL_HANDLER_ARGS); static void ip_initid(int); static uint16_t ip_randomid(void); static void ipid_sysinit(void); static void ipid_sysuninit(void); SYSCTL_DECL(_net_inet_ip); SYSCTL_PROC(_net_inet_ip, OID_AUTO, random_id, CTLTYPE_INT | CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip_do_randomid), 0, sysctl_ip_randomid, "IU", "Assign random ip_id values"); SYSCTL_INT(_net_inet_ip, OID_AUTO, rfc6864, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ip_rfc6864), 0, "Use constant IP ID for atomic datagrams"); SYSCTL_PROC(_net_inet_ip, OID_AUTO, random_id_period, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET, &VNET_NAME(array_size), 0, sysctl_ip_id_change, "IU", "IP ID Array size"); SYSCTL_INT(_net_inet_ip, OID_AUTO, random_id_collisions, CTLFLAG_RD | CTLFLAG_VNET, &VNET_NAME(random_id_collisions), 0, "Count of IP ID collisions"); SYSCTL_INT(_net_inet_ip, OID_AUTO, random_id_total, CTLFLAG_RD | CTLFLAG_VNET, &VNET_NAME(random_id_total), 0, "Count of IP IDs created"); static int sysctl_ip_randomid(SYSCTL_HANDLER_ARGS) { int error, new; new = V_ip_do_randomid; error = sysctl_handle_int(oidp, &new, 0, req); if (error || req->newptr == NULL) return (error); if (new != 0 && new != 1) return (EINVAL); if (new == V_ip_do_randomid) return (0); if (new == 1 && V_ip_do_randomid == 0) ip_initid(8192); /* We don't free memory when turning random ID off, due to race. */ V_ip_do_randomid = new; return (0); } static int sysctl_ip_id_change(SYSCTL_HANDLER_ARGS) { int error, new; new = V_array_size; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new >= 512 && new <= 32768) ip_initid(new); else error = EINVAL; } return (error); } static void ip_initid(int new_size) { uint16_t *new_array; bitstr_t *new_bits; new_array = malloc(new_size * sizeof(uint16_t), M_IPID, M_WAITOK | M_ZERO); new_bits = malloc(bitstr_size(65536), M_IPID, M_WAITOK | M_ZERO); mtx_lock(&V_ip_id_mtx); if (V_id_array != NULL) { free(V_id_array, M_IPID); free(V_id_bits, M_IPID); } V_id_array = new_array; V_id_bits = new_bits; V_array_size = new_size; V_array_ptr = 0; V_random_id_collisions = 0; V_random_id_total = 0; mtx_unlock(&V_ip_id_mtx); } static uint16_t ip_randomid(void) { uint16_t new_id; mtx_lock(&V_ip_id_mtx); /* * To avoid a conflict with the zeros that the array is initially * filled with, we never hand out an id of zero. */ new_id = 0; do { if (new_id != 0) V_random_id_collisions++; arc4rand(&new_id, sizeof(new_id), 0); } while (bit_test(V_id_bits, new_id) || new_id == 0); bit_clear(V_id_bits, V_id_array[V_array_ptr]); bit_set(V_id_bits, new_id); V_id_array[V_array_ptr] = new_id; V_array_ptr++; if (V_array_ptr == V_array_size) V_array_ptr = 0; V_random_id_total++; mtx_unlock(&V_ip_id_mtx); return (new_id); } void ip_fillid(struct ip *ip) { /* * Per RFC6864 Section 4 * * o Atomic datagrams: (DF==1) && (MF==0) && (frag_offset==0) * o Non-atomic datagrams: (DF==0) || (MF==1) || (frag_offset>0) */ if (V_ip_rfc6864 && (ip->ip_off & htons(IP_DF)) == htons(IP_DF)) ip->ip_id = 0; else if (V_ip_do_randomid) ip->ip_id = ip_randomid(); else { counter_u64_add(V_ip_id, 1); /* * There are two issues about this trick, to be kept in mind. * 1) We can migrate between counter_u64_add() and next * line, and grab counter from other CPU, resulting in too * quick ID reuse. This is tolerable in our particular case, * since probability of such event is much lower then reuse * of ID due to legitimate overflow, that at modern Internet * speeds happens all the time. * 2) We are relying on the fact that counter(9) is based on * UMA_ZONE_PCPU uma(9) zone. We also take only last * sixteen bits of a counter, so we don't care about the * fact that machines with 32-bit word update their counters * not atomically. */ ip->ip_id = htons((*(uint64_t *)zpcpu_get(V_ip_id)) & 0xffff); } } static void ipid_sysinit(void) { + int i; mtx_init(&V_ip_id_mtx, "ip_id_mtx", NULL, MTX_DEF); V_ip_id = counter_u64_alloc(M_WAITOK); - for (int i = 0; i < mp_ncpus; i++) + + CPU_FOREACH(i) arc4rand(zpcpu_get_cpu(V_ip_id, i), sizeof(uint64_t), 0); } VNET_SYSINIT(ip_id, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY, ipid_sysinit, NULL); static void ipid_sysuninit(void) { if (V_id_array != NULL) { free(V_id_array, M_IPID); free(V_id_bits, M_IPID); } counter_u64_free(V_ip_id); mtx_destroy(&V_ip_id_mtx); } VNET_SYSUNINIT(ip_id, SI_SUB_PROTO_DOMAIN, SI_ORDER_THIRD, ipid_sysuninit, NULL); Index: head/sys/powerpc/include/counter.h =================================================================== --- head/sys/powerpc/include/counter.h (revision 302371) +++ head/sys/powerpc/include/counter.h (revision 302372) @@ -1,162 +1,162 @@ /*- * Copyright (c) 2012, 2013 Konstantin Belousov * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR 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_COUNTER_H__ #define __MACHINE_COUNTER_H__ #include #ifdef INVARIANTS #include #endif #ifdef __powerpc64__ #define counter_enter() do {} while (0) #define counter_exit() do {} while (0) #ifdef IN_SUBR_COUNTER_C static inline uint64_t counter_u64_read_one(uint64_t *p, int cpu) { return (*(uint64_t *)((char *)p + sizeof(struct pcpu) * cpu)); } static inline uint64_t counter_u64_fetch_inline(uint64_t *p) { uint64_t r; int i; r = 0; - for (i = 0; i < mp_ncpus; i++) + CPU_FOREACH(i) r += counter_u64_read_one((uint64_t *)p, i); return (r); } static void counter_u64_zero_one_cpu(void *arg) { *((uint64_t *)((char *)arg + sizeof(struct pcpu) * PCPU_GET(cpuid))) = 0; } static inline void counter_u64_zero_inline(counter_u64_t c) { smp_rendezvous(smp_no_rendevous_barrier, counter_u64_zero_one_cpu, smp_no_rendevous_barrier, c); } #endif #define counter_u64_add_protected(c, i) counter_u64_add(c, i) extern struct pcpu __pcpu[MAXCPU]; static inline void counter_u64_add(counter_u64_t c, int64_t inc) { uint64_t ccpu, old; __asm __volatile("\n" "1:\n\t" "mfsprg %0, 0\n\t" "ldarx %1, %0, %2\n\t" "add %1, %1, %3\n\t" "stdcx. %1, %0, %2\n\t" "bne- 1b" : "=&b" (ccpu), "=&r" (old) : "r" ((char *)c - (char *)&__pcpu[0]), "r" (inc) : "cr0", "memory"); } #else /* !64bit */ #define counter_enter() critical_enter() #define counter_exit() critical_exit() #ifdef IN_SUBR_COUNTER_C /* XXXKIB non-atomic 64bit read */ static inline uint64_t counter_u64_read_one(uint64_t *p, int cpu) { return (*(uint64_t *)((char *)p + sizeof(struct pcpu) * cpu)); } static inline uint64_t counter_u64_fetch_inline(uint64_t *p) { uint64_t r; int i; r = 0; for (i = 0; i < mp_ncpus; i++) r += counter_u64_read_one((uint64_t *)p, i); return (r); } /* XXXKIB non-atomic 64bit store, might interrupt increment */ static void counter_u64_zero_one_cpu(void *arg) { *((uint64_t *)((char *)arg + sizeof(struct pcpu) * PCPU_GET(cpuid))) = 0; } static inline void counter_u64_zero_inline(counter_u64_t c) { smp_rendezvous(smp_no_rendevous_barrier, counter_u64_zero_one_cpu, smp_no_rendevous_barrier, c); } #endif #define counter_u64_add_protected(c, inc) do { \ CRITICAL_ASSERT(curthread); \ *(uint64_t *)zpcpu_get(c) += (inc); \ } while (0) static inline void counter_u64_add(counter_u64_t c, int64_t inc) { counter_enter(); counter_u64_add_protected(c, inc); counter_exit(); } #endif /* 64bit */ #endif /* ! __MACHINE_COUNTER_H__ */ Index: head/sys/powerpc/powerpc/mp_machdep.c =================================================================== --- head/sys/powerpc/powerpc/mp_machdep.c (revision 302371) +++ head/sys/powerpc/powerpc/mp_machdep.c (revision 302372) @@ -1,377 +1,373 @@ /*- * Copyright (c) 2008 Marcel Moolenaar * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "pic_if.h" extern struct pcpu __pcpu[MAXCPU]; volatile static int ap_awake; volatile static u_int ap_letgo; volatile static u_quad_t ap_timebase; static u_int ipi_msg_cnt[32]; static struct mtx ap_boot_mtx; struct pcb stoppcbs[MAXCPU]; void machdep_ap_bootstrap(void) { /* Set PIR */ PCPU_SET(pir, mfspr(SPR_PIR)); PCPU_SET(awake, 1); __asm __volatile("msync; isync"); while (ap_letgo == 0) ; /* Initialize DEC and TB, sync with the BSP values */ #ifdef __powerpc64__ /* Writing to the time base register is hypervisor-privileged */ if (mfmsr() & PSL_HV) mttb(ap_timebase); #else mttb(ap_timebase); #endif decr_ap_init(); /* Give platform code a chance to do anything necessary */ platform_smp_ap_init(); /* Serialize console output and AP count increment */ mtx_lock_spin(&ap_boot_mtx); ap_awake++; printf("SMP: AP CPU #%d launched\n", PCPU_GET(cpuid)); mtx_unlock_spin(&ap_boot_mtx); /* Start per-CPU event timers. */ cpu_initclocks_ap(); /* Announce ourselves awake, and enter the scheduler */ sched_throw(NULL); } void cpu_mp_setmaxid(void) { struct cpuref cpuref; int error; mp_ncpus = 0; + mp_maxid = 0; error = platform_smp_first_cpu(&cpuref); while (!error) { mp_ncpus++; + mp_maxid = max(cpuref.cr_cpuid, mp_maxid); error = platform_smp_next_cpu(&cpuref); } /* Sanity. */ if (mp_ncpus == 0) mp_ncpus = 1; - - /* - * Set the largest cpuid we're going to use. This is necessary - * for VM initialization. - */ - mp_maxid = min(mp_ncpus, MAXCPU) - 1; } int cpu_mp_probe(void) { /* * We're not going to enable SMP if there's only 1 processor. */ return (mp_ncpus > 1); } void cpu_mp_start(void) { struct cpuref bsp, cpu; struct pcpu *pc; int error; error = platform_smp_get_bsp(&bsp); KASSERT(error == 0, ("Don't know BSP")); KASSERT(bsp.cr_cpuid == 0, ("%s: cpuid != 0", __func__)); error = platform_smp_first_cpu(&cpu); while (!error) { if (cpu.cr_cpuid >= MAXCPU) { printf("SMP: cpu%d: skipped -- ID out of range\n", cpu.cr_cpuid); goto next; } if (CPU_ISSET(cpu.cr_cpuid, &all_cpus)) { printf("SMP: cpu%d: skipped - duplicate ID\n", cpu.cr_cpuid); goto next; } if (cpu.cr_cpuid != bsp.cr_cpuid) { void *dpcpu; pc = &__pcpu[cpu.cr_cpuid]; dpcpu = (void *)kmem_malloc(kernel_arena, DPCPU_SIZE, M_WAITOK | M_ZERO); pcpu_init(pc, cpu.cr_cpuid, sizeof(*pc)); dpcpu_init(dpcpu, cpu.cr_cpuid); } else { pc = pcpup; pc->pc_cpuid = bsp.cr_cpuid; pc->pc_bsp = 1; } pc->pc_hwref = cpu.cr_hwref; CPU_SET(pc->pc_cpuid, &all_cpus); next: error = platform_smp_next_cpu(&cpu); } } void cpu_mp_announce(void) { struct pcpu *pc; int i; for (i = 0; i <= mp_maxid; i++) { pc = pcpu_find(i); if (pc == NULL) continue; printf("cpu%d: dev=%x", i, (int)pc->pc_hwref); if (pc->pc_bsp) printf(" (BSP)"); printf("\n"); } } static void cpu_mp_unleash(void *dummy) { struct pcpu *pc; int cpus, timeout; if (mp_ncpus <= 1) return; mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN); cpus = 0; smp_cpus = 0; #ifdef BOOKE tlb1_ap_prep(); #endif STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { cpus++; if (!pc->pc_bsp) { if (bootverbose) printf("Waking up CPU %d (dev=%x)\n", pc->pc_cpuid, (int)pc->pc_hwref); platform_smp_start_cpu(pc); timeout = 2000; /* wait 2sec for the AP */ while (!pc->pc_awake && --timeout > 0) DELAY(1000); } else { PCPU_SET(pir, mfspr(SPR_PIR)); pc->pc_awake = 1; } if (pc->pc_awake) { if (bootverbose) printf("Adding CPU %d, pir=%x, awake=%x\n", pc->pc_cpuid, pc->pc_pir, pc->pc_awake); smp_cpus++; } else CPU_SET(pc->pc_cpuid, &stopped_cpus); } ap_awake = 1; /* Provide our current DEC and TB values for APs */ ap_timebase = mftb() + 10; __asm __volatile("msync; isync"); /* Let APs continue */ atomic_store_rel_int(&ap_letgo, 1); #ifdef __powerpc64__ /* Writing to the time base register is hypervisor-privileged */ if (mfmsr() & PSL_HV) mttb(ap_timebase); #else mttb(ap_timebase); #endif while (ap_awake < smp_cpus) ; if (smp_cpus != cpus || cpus != mp_ncpus) { printf("SMP: %d CPUs found; %d CPUs usable; %d CPUs woken\n", mp_ncpus, cpus, smp_cpus); } /* Let the APs get into the scheduler */ DELAY(10000); /* XXX Atomic set operation? */ smp_started = 1; } SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, cpu_mp_unleash, NULL); int powerpc_ipi_handler(void *arg) { u_int cpuid; uint32_t ipimask; int msg; CTR2(KTR_SMP, "%s: MSR 0x%08x", __func__, mfmsr()); ipimask = atomic_readandclear_32(&(pcpup->pc_ipimask)); if (ipimask == 0) return (FILTER_STRAY); while ((msg = ffs(ipimask) - 1) != -1) { ipimask &= ~(1u << msg); ipi_msg_cnt[msg]++; switch (msg) { case IPI_AST: CTR1(KTR_SMP, "%s: IPI_AST", __func__); break; case IPI_PREEMPT: CTR1(KTR_SMP, "%s: IPI_PREEMPT", __func__); sched_preempt(curthread); break; case IPI_RENDEZVOUS: CTR1(KTR_SMP, "%s: IPI_RENDEZVOUS", __func__); smp_rendezvous_action(); break; case IPI_STOP: /* * IPI_STOP_HARD is mapped to IPI_STOP so it is not * necessary to add such case in the switch. */ CTR1(KTR_SMP, "%s: IPI_STOP or IPI_STOP_HARD (stop)", __func__); cpuid = PCPU_GET(cpuid); savectx(&stoppcbs[cpuid]); savectx(PCPU_GET(curpcb)); CPU_SET_ATOMIC(cpuid, &stopped_cpus); while (!CPU_ISSET(cpuid, &started_cpus)) cpu_spinwait(); CPU_CLR_ATOMIC(cpuid, &stopped_cpus); CPU_CLR_ATOMIC(cpuid, &started_cpus); CTR1(KTR_SMP, "%s: IPI_STOP (restart)", __func__); break; case IPI_HARDCLOCK: CTR1(KTR_SMP, "%s: IPI_HARDCLOCK", __func__); hardclockintr(); break; } } return (FILTER_HANDLED); } static void ipi_send(struct pcpu *pc, int ipi) { CTR4(KTR_SMP, "%s: pc=%p, targetcpu=%d, IPI=%d", __func__, pc, pc->pc_cpuid, ipi); atomic_set_32(&pc->pc_ipimask, (1 << ipi)); powerpc_sync(); PIC_IPI(root_pic, pc->pc_cpuid); CTR1(KTR_SMP, "%s: sent", __func__); } /* Send an IPI to a set of cpus. */ void ipi_selected(cpuset_t cpus, int ipi) { struct pcpu *pc; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (CPU_ISSET(pc->pc_cpuid, &cpus)) ipi_send(pc, ipi); } } /* Send an IPI to a specific CPU. */ void ipi_cpu(int cpu, u_int ipi) { ipi_send(cpuid_to_pcpu[cpu], ipi); } /* Send an IPI to all CPUs EXCEPT myself. */ void ipi_all_but_self(int ipi) { struct pcpu *pc; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc != pcpup) ipi_send(pc, ipi); } } Index: head/sys/vm/uma.h =================================================================== --- head/sys/vm/uma.h (revision 302371) +++ head/sys/vm/uma.h (revision 302372) @@ -1,694 +1,694 @@ /*- * Copyright (c) 2002, 2003, 2004, 2005 Jeffrey Roberson * Copyright (c) 2004, 2005 Bosko Milekic * 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 unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * $FreeBSD$ * */ /* * uma.h - External definitions for the Universal Memory Allocator * */ #ifndef _VM_UMA_H_ #define _VM_UMA_H_ #include /* For NULL */ #include /* For M_* */ /* User visible parameters */ #define UMA_SMALLEST_UNIT (PAGE_SIZE / 256) /* Smallest item allocated */ /* Types and type defs */ struct uma_zone; /* Opaque type used as a handle to the zone */ typedef struct uma_zone * uma_zone_t; void zone_drain(uma_zone_t); /* * Item constructor * * Arguments: * item A pointer to the memory which has been allocated. * arg The arg field passed to uma_zalloc_arg * size The size of the allocated item * flags See zalloc flags * * Returns: * 0 on success * errno on failure * * Discussion: * The constructor is called just before the memory is returned * to the user. It may block if necessary. */ typedef int (*uma_ctor)(void *mem, int size, void *arg, int flags); /* * Item destructor * * Arguments: * item A pointer to the memory which has been allocated. * size The size of the item being destructed. * arg Argument passed through uma_zfree_arg * * Returns: * Nothing * * Discussion: * The destructor may perform operations that differ from those performed * by the initializer, but it must leave the object in the same state. * This IS type stable storage. This is called after EVERY zfree call. */ typedef void (*uma_dtor)(void *mem, int size, void *arg); /* * Item initializer * * Arguments: * item A pointer to the memory which has been allocated. * size The size of the item being initialized. * flags See zalloc flags * * Returns: * 0 on success * errno on failure * * Discussion: * The initializer is called when the memory is cached in the uma zone. * The initializer and the destructor should leave the object in the same * state. */ typedef int (*uma_init)(void *mem, int size, int flags); /* * Item discard function * * Arguments: * item A pointer to memory which has been 'freed' but has not left the * zone's cache. * size The size of the item being discarded. * * Returns: * Nothing * * Discussion: * This routine is called when memory leaves a zone and is returned to the * system for other uses. It is the counter-part to the init function. */ typedef void (*uma_fini)(void *mem, int size); /* * Import new memory into a cache zone. */ typedef int (*uma_import)(void *arg, void **store, int count, int flags); /* * Free memory from a cache zone. */ typedef void (*uma_release)(void *arg, void **store, int count); /* * What's the difference between initializing and constructing? * * The item is initialized when it is cached, and this is the state that the * object should be in when returned to the allocator. The purpose of this is * to remove some code which would otherwise be called on each allocation by * utilizing a known, stable state. This differs from the constructor which * will be called on EVERY allocation. * * For example, in the initializer you may want to initialize embedded locks, * NULL list pointers, set up initial states, magic numbers, etc. This way if * the object is held in the allocator and re-used it won't be necessary to * re-initialize it. * * The constructor may be used to lock a data structure, link it on to lists, * bump reference counts or total counts of outstanding structures, etc. * */ /* Function proto types */ /* * Create a new uma zone * * Arguments: * name The text name of the zone for debugging and stats. This memory * should not be freed until the zone has been deallocated. * size The size of the object that is being created. * ctor The constructor that is called when the object is allocated. * dtor The destructor that is called when the object is freed. * init An initializer that sets up the initial state of the memory. * fini A discard function that undoes initialization done by init. * ctor/dtor/init/fini may all be null, see notes above. * align A bitmask that corresponds to the requested alignment * eg 4 would be 0x3 * flags A set of parameters that control the behavior of the zone. * * Returns: * A pointer to a structure which is intended to be opaque to users of * the interface. The value may be null if the wait flag is not set. */ uma_zone_t uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, uma_init uminit, uma_fini fini, int align, uint32_t flags); /* * Create a secondary uma zone * * Arguments: * name The text name of the zone for debugging and stats. This memory * should not be freed until the zone has been deallocated. * ctor The constructor that is called when the object is allocated. * dtor The destructor that is called when the object is freed. * zinit An initializer that sets up the initial state of the memory * as the object passes from the Keg's slab to the Zone's cache. * zfini A discard function that undoes initialization done by init * as the object passes from the Zone's cache to the Keg's slab. * * ctor/dtor/zinit/zfini may all be null, see notes above. * Note that the zinit and zfini specified here are NOT * exactly the same as the init/fini specified to uma_zcreate() * when creating a master zone. These zinit/zfini are called * on the TRANSITION from keg to zone (and vice-versa). Once * these are set, the primary zone may alter its init/fini * (which are called when the object passes from VM to keg) * using uma_zone_set_init/fini()) as well as its own * zinit/zfini (unset by default for master zone) with * uma_zone_set_zinit/zfini() (note subtle 'z' prefix). * * master A reference to this zone's Master Zone (Primary Zone), * which contains the backing Keg for the Secondary Zone * being added. * * Returns: * A pointer to a structure which is intended to be opaque to users of * the interface. The value may be null if the wait flag is not set. */ uma_zone_t uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, uma_init zinit, uma_fini zfini, uma_zone_t master); /* * Add a second master to a secondary zone. This provides multiple data * backends for objects with the same size. Both masters must have * compatible allocation flags. Presently, UMA_ZONE_MALLOC type zones are * the only supported. * * Returns: * Error on failure, 0 on success. */ int uma_zsecond_add(uma_zone_t zone, uma_zone_t master); /* * Create cache-only zones. * * This allows uma's per-cpu cache facilities to handle arbitrary * pointers. Consumers must specify the import and release functions to * fill and destroy caches. UMA does not allocate any memory for these * zones. The 'arg' parameter is passed to import/release and is caller * specific. */ uma_zone_t uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease, void *arg, int flags); /* * Definitions for uma_zcreate flags * * These flags share space with UMA_ZFLAGs in uma_int.h. Be careful not to * overlap when adding new features. 0xf0000000 is in use by uma_int.h. */ #define UMA_ZONE_PAGEABLE 0x0001 /* Return items not fully backed by physical memory XXX Not yet */ #define UMA_ZONE_ZINIT 0x0002 /* Initialize with zeros */ #define UMA_ZONE_STATIC 0x0004 /* Statically sized zone */ #define UMA_ZONE_OFFPAGE 0x0008 /* Force the slab structure allocation off of the real memory */ #define UMA_ZONE_MALLOC 0x0010 /* For use by malloc(9) only! */ #define UMA_ZONE_NOFREE 0x0020 /* Do not free slabs of this type! */ #define UMA_ZONE_MTXCLASS 0x0040 /* Create a new lock class */ #define UMA_ZONE_VM 0x0080 /* * Used for internal vm datastructures * only. */ #define UMA_ZONE_HASH 0x0100 /* * Use a hash table instead of caching * information in the vm_page. */ #define UMA_ZONE_SECONDARY 0x0200 /* Zone is a Secondary Zone */ /* 0x0400 Unused */ #define UMA_ZONE_MAXBUCKET 0x0800 /* Use largest buckets */ #define UMA_ZONE_CACHESPREAD 0x1000 /* * Spread memory start locations across * all possible cache lines. May * require many virtually contiguous * backend pages and can fail early. */ #define UMA_ZONE_VTOSLAB 0x2000 /* Zone uses vtoslab for lookup. */ #define UMA_ZONE_NODUMP 0x4000 /* * Zone's pages will not be included in * mini-dumps. */ #define UMA_ZONE_PCPU 0x8000 /* - * Allocates mp_ncpus slabs sized to + * Allocates mp_maxid + 1 slabs sized to * sizeof(struct pcpu). */ /* * These flags are shared between the keg and zone. In zones wishing to add * new kegs these flags must be compatible. Some are determined based on * physical parameters of the request and may not be provided by the consumer. */ #define UMA_ZONE_INHERIT \ (UMA_ZONE_OFFPAGE | UMA_ZONE_MALLOC | UMA_ZONE_NOFREE | \ UMA_ZONE_HASH | UMA_ZONE_VTOSLAB | UMA_ZONE_PCPU) /* Definitions for align */ #define UMA_ALIGN_PTR (sizeof(void *) - 1) /* Alignment fit for ptr */ #define UMA_ALIGN_LONG (sizeof(long) - 1) /* "" long */ #define UMA_ALIGN_INT (sizeof(int) - 1) /* "" int */ #define UMA_ALIGN_SHORT (sizeof(short) - 1) /* "" short */ #define UMA_ALIGN_CHAR (sizeof(char) - 1) /* "" char */ #define UMA_ALIGN_CACHE (0 - 1) /* Cache line size align */ /* * Destroys an empty uma zone. If the zone is not empty uma complains loudly. * * Arguments: * zone The zone we want to destroy. * */ void uma_zdestroy(uma_zone_t zone); /* * Allocates an item out of a zone * * Arguments: * zone The zone we are allocating from * arg This data is passed to the ctor function * flags See sys/malloc.h for available flags. * * Returns: * A non-null pointer to an initialized element from the zone is * guaranteed if the wait flag is M_WAITOK. Otherwise a null pointer * may be returned if the zone is empty or the ctor failed. */ void *uma_zalloc_arg(uma_zone_t zone, void *arg, int flags); /* * Allocates an item out of a zone without supplying an argument * * This is just a wrapper for uma_zalloc_arg for convenience. * */ static __inline void *uma_zalloc(uma_zone_t zone, int flags); static __inline void * uma_zalloc(uma_zone_t zone, int flags) { return uma_zalloc_arg(zone, NULL, flags); } /* * Frees an item back into the specified zone. * * Arguments: * zone The zone the item was originally allocated out of. * item The memory to be freed. * arg Argument passed to the destructor * * Returns: * Nothing. */ void uma_zfree_arg(uma_zone_t zone, void *item, void *arg); /* * Frees an item back to a zone without supplying an argument * * This is just a wrapper for uma_zfree_arg for convenience. * */ static __inline void uma_zfree(uma_zone_t zone, void *item); static __inline void uma_zfree(uma_zone_t zone, void *item) { uma_zfree_arg(zone, item, NULL); } /* * XXX The rest of the prototypes in this header are h0h0 magic for the VM. * If you think you need to use it for a normal zone you're probably incorrect. */ /* * Backend page supplier routines * * Arguments: * zone The zone that is requesting pages. * size The number of bytes being requested. * pflag Flags for these memory pages, see below. * wait Indicates our willingness to block. * * Returns: * A pointer to the allocated memory or NULL on failure. */ typedef void *(*uma_alloc)(uma_zone_t zone, vm_size_t size, uint8_t *pflag, int wait); /* * Backend page free routines * * Arguments: * item A pointer to the previously allocated pages. * size The original size of the allocation. * pflag The flags for the slab. See UMA_SLAB_* below. * * Returns: * None */ typedef void (*uma_free)(void *item, vm_size_t size, uint8_t pflag); /* * Sets up the uma allocator. (Called by vm_mem_init) * * Arguments: * bootmem A pointer to memory used to bootstrap the system. * * Returns: * Nothing * * Discussion: * This memory is used for zones which allocate things before the * backend page supplier can give us pages. It should be * UMA_SLAB_SIZE * boot_pages bytes. (see uma_int.h) * */ void uma_startup(void *bootmem, int boot_pages); /* * Finishes starting up the allocator. This should * be called when kva is ready for normal allocs. * * Arguments: * None * * Returns: * Nothing * * Discussion: * uma_startup2 is called by kmeminit() to enable us of uma for malloc. */ void uma_startup2(void); /* * Reclaims unused memory for all zones * * Arguments: * None * Returns: * None * * This should only be called by the page out daemon. */ void uma_reclaim(void); /* * Sets the alignment mask to be used for all zones requesting cache * alignment. Should be called by MD boot code prior to starting VM/UMA. * * Arguments: * align The alignment mask * * Returns: * Nothing */ void uma_set_align(int align); /* * Set a reserved number of items to hold for M_USE_RESERVE allocations. All * other requests must allocate new backing pages. */ void uma_zone_reserve(uma_zone_t zone, int nitems); /* * Reserves the maximum KVA space required by the zone and configures the zone * to use a VM_ALLOC_NOOBJ-based backend allocator. * * Arguments: * zone The zone to update. * nitems The upper limit on the number of items that can be allocated. * * Returns: * 0 if KVA space can not be allocated * 1 if successful * * Discussion: * When the machine supports a direct map and the zone's items are smaller * than a page, the zone will use the direct map instead of allocating KVA * space. */ int uma_zone_reserve_kva(uma_zone_t zone, int nitems); /* * Sets a high limit on the number of items allowed in a zone * * Arguments: * zone The zone to limit * nitems The requested upper limit on the number of items allowed * * Returns: * int The effective value of nitems after rounding up based on page size */ int uma_zone_set_max(uma_zone_t zone, int nitems); /* * Obtains the effective limit on the number of items in a zone * * Arguments: * zone The zone to obtain the effective limit from * * Return: * 0 No limit * int The effective limit of the zone */ int uma_zone_get_max(uma_zone_t zone); /* * Sets a warning to be printed when limit is reached * * Arguments: * zone The zone we will warn about * warning Warning content * * Returns: * Nothing */ void uma_zone_set_warning(uma_zone_t zone, const char *warning); /* * Sets a function to run when limit is reached * * Arguments: * zone The zone to which this applies * fx The function ro run * * Returns: * Nothing */ typedef void (*uma_maxaction_t)(uma_zone_t, int); void uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t); /* * Obtains the approximate current number of items allocated from a zone * * Arguments: * zone The zone to obtain the current allocation count from * * Return: * int The approximate current number of items allocated from the zone */ int uma_zone_get_cur(uma_zone_t zone); /* * The following two routines (uma_zone_set_init/fini) * are used to set the backend init/fini pair which acts on an * object as it becomes allocated and is placed in a slab within * the specified zone's backing keg. These should probably not * be changed once allocations have already begun, but only be set * immediately upon zone creation. */ void uma_zone_set_init(uma_zone_t zone, uma_init uminit); void uma_zone_set_fini(uma_zone_t zone, uma_fini fini); /* * The following two routines (uma_zone_set_zinit/zfini) are * used to set the zinit/zfini pair which acts on an object as * it passes from the backing Keg's slab cache to the * specified Zone's bucket cache. These should probably not * be changed once allocations have already begun, but only be set * immediately upon zone creation. */ void uma_zone_set_zinit(uma_zone_t zone, uma_init zinit); void uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini); /* * Replaces the standard backend allocator for this zone. * * Arguments: * zone The zone whose backend allocator is being changed. * allocf A pointer to the allocation function * * Returns: * Nothing * * Discussion: * This could be used to implement pageable allocation, or perhaps * even DMA allocators if used in conjunction with the OFFPAGE * zone flag. */ void uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf); /* * Used for freeing memory provided by the allocf above * * Arguments: * zone The zone that intends to use this free routine. * freef The page freeing routine. * * Returns: * Nothing */ void uma_zone_set_freef(uma_zone_t zone, uma_free freef); /* * These flags are setable in the allocf and visible in the freef. */ #define UMA_SLAB_BOOT 0x01 /* Slab alloced from boot pages */ #define UMA_SLAB_KMEM 0x02 /* Slab alloced from kmem_map */ #define UMA_SLAB_KERNEL 0x04 /* Slab alloced from kernel_map */ #define UMA_SLAB_PRIV 0x08 /* Slab alloced from priv allocator */ #define UMA_SLAB_OFFP 0x10 /* Slab is managed separately */ #define UMA_SLAB_MALLOC 0x20 /* Slab is a large malloc slab */ /* 0x40 and 0x80 are available */ /* * Used to pre-fill a zone with some number of items * * Arguments: * zone The zone to fill * itemcnt The number of items to reserve * * Returns: * Nothing * * NOTE: This is blocking and should only be done at startup */ void uma_prealloc(uma_zone_t zone, int itemcnt); /* * Used to determine if a fixed-size zone is exhausted. * * Arguments: * zone The zone to check * * Returns: * Non-zero if zone is exhausted. */ int uma_zone_exhausted(uma_zone_t zone); int uma_zone_exhausted_nolock(uma_zone_t zone); /* * Common UMA_ZONE_PCPU zones. */ extern uma_zone_t pcpu_zone_64; extern uma_zone_t pcpu_zone_ptr; /* * Exported statistics structures to be used by user space monitoring tools. * Statistics stream consists of a uma_stream_header, followed by a series of * alternative uma_type_header and uma_type_stat structures. */ #define UMA_STREAM_VERSION 0x00000001 struct uma_stream_header { uint32_t ush_version; /* Stream format version. */ uint32_t ush_maxcpus; /* Value of MAXCPU for stream. */ uint32_t ush_count; /* Number of records. */ uint32_t _ush_pad; /* Pad/reserved field. */ }; #define UTH_MAX_NAME 32 #define UTH_ZONE_SECONDARY 0x00000001 struct uma_type_header { /* * Static per-zone data, some extracted from the supporting keg. */ char uth_name[UTH_MAX_NAME]; uint32_t uth_align; /* Keg: alignment. */ uint32_t uth_size; /* Keg: requested size of item. */ uint32_t uth_rsize; /* Keg: real size of item. */ uint32_t uth_maxpages; /* Keg: maximum number of pages. */ uint32_t uth_limit; /* Keg: max items to allocate. */ /* * Current dynamic zone/keg-derived statistics. */ uint32_t uth_pages; /* Keg: pages allocated. */ uint32_t uth_keg_free; /* Keg: items free. */ uint32_t uth_zone_free; /* Zone: items free. */ uint32_t uth_bucketsize; /* Zone: desired bucket size. */ uint32_t uth_zone_flags; /* Zone: flags. */ uint64_t uth_allocs; /* Zone: number of allocations. */ uint64_t uth_frees; /* Zone: number of frees. */ uint64_t uth_fails; /* Zone: number of alloc failures. */ uint64_t uth_sleeps; /* Zone: number of alloc sleeps. */ uint64_t _uth_reserved1[2]; /* Reserved. */ }; struct uma_percpu_stat { uint64_t ups_allocs; /* Cache: number of allocations. */ uint64_t ups_frees; /* Cache: number of frees. */ uint64_t ups_cache_free; /* Cache: free items in cache. */ uint64_t _ups_reserved[5]; /* Reserved. */ }; void uma_reclaim_wakeup(void); void uma_reclaim_worker(void *); #endif /* _VM_UMA_H_ */ Index: head/sys/vm/uma_core.c =================================================================== --- head/sys/vm/uma_core.c (revision 302371) +++ head/sys/vm/uma_core.c (revision 302372) @@ -1,3684 +1,3685 @@ /*- * Copyright (c) 2002-2005, 2009, 2013 Jeffrey Roberson * Copyright (c) 2004, 2005 Bosko Milekic * Copyright (c) 2004-2006 Robert N. M. Watson * 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 unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ /* * uma_core.c Implementation of the Universal Memory allocator * * This allocator is intended to replace the multitude of similar object caches * in the standard FreeBSD kernel. The intent is to be flexible as well as * efficient. A primary design goal is to return unused memory to the rest of * the system. This will make the system as a whole more flexible due to the * ability to move memory to subsystems which most need it instead of leaving * pools of reserved memory unused. * * The basic ideas stem from similar slab/zone based allocators whose algorithms * are well known. * */ /* * TODO: * - Improve memory usage for large allocations * - Investigate cache size adjustments */ #include __FBSDID("$FreeBSD$"); /* I should really use ktr.. */ /* #define UMA_DEBUG 1 #define UMA_DEBUG_ALLOC 1 #define UMA_DEBUG_ALLOC_1 1 */ #include "opt_ddb.h" #include "opt_param.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DEBUG_MEMGUARD #include #endif /* * This is the zone and keg from which all zones are spawned. The idea is that * even the zone & keg heads are allocated from the allocator, so we use the * bss section to bootstrap us. */ static struct uma_keg masterkeg; static struct uma_zone masterzone_k; static struct uma_zone masterzone_z; static uma_zone_t kegs = &masterzone_k; static uma_zone_t zones = &masterzone_z; /* This is the zone from which all of uma_slab_t's are allocated. */ static uma_zone_t slabzone; /* * The initial hash tables come out of this zone so they can be allocated * prior to malloc coming up. */ static uma_zone_t hashzone; /* The boot-time adjusted value for cache line alignment. */ int uma_align_cache = 64 - 1; static MALLOC_DEFINE(M_UMAHASH, "UMAHash", "UMA Hash Buckets"); /* * Are we allowed to allocate buckets? */ static int bucketdisable = 1; /* Linked list of all kegs in the system */ static LIST_HEAD(,uma_keg) uma_kegs = LIST_HEAD_INITIALIZER(uma_kegs); /* Linked list of all cache-only zones in the system */ static LIST_HEAD(,uma_zone) uma_cachezones = LIST_HEAD_INITIALIZER(uma_cachezones); /* This RW lock protects the keg list */ static struct rwlock_padalign uma_rwlock; /* Linked list of boot time pages */ static LIST_HEAD(,uma_slab) uma_boot_pages = LIST_HEAD_INITIALIZER(uma_boot_pages); /* This mutex protects the boot time pages list */ static struct mtx_padalign uma_boot_pages_mtx; static struct sx uma_drain_lock; /* Is the VM done starting up? */ static int booted = 0; #define UMA_STARTUP 1 #define UMA_STARTUP2 2 /* * This is the handle used to schedule events that need to happen * outside of the allocation fast path. */ static struct callout uma_callout; #define UMA_TIMEOUT 20 /* Seconds for callout interval. */ /* * This structure is passed as the zone ctor arg so that I don't have to create * a special allocation function just for zones. */ struct uma_zctor_args { const char *name; size_t size; uma_ctor ctor; uma_dtor dtor; uma_init uminit; uma_fini fini; uma_import import; uma_release release; void *arg; uma_keg_t keg; int align; uint32_t flags; }; struct uma_kctor_args { uma_zone_t zone; size_t size; uma_init uminit; uma_fini fini; int align; uint32_t flags; }; struct uma_bucket_zone { uma_zone_t ubz_zone; char *ubz_name; int ubz_entries; /* Number of items it can hold. */ int ubz_maxsize; /* Maximum allocation size per-item. */ }; /* * Compute the actual number of bucket entries to pack them in power * of two sizes for more efficient space utilization. */ #define BUCKET_SIZE(n) \ (((sizeof(void *) * (n)) - sizeof(struct uma_bucket)) / sizeof(void *)) #define BUCKET_MAX BUCKET_SIZE(256) struct uma_bucket_zone bucket_zones[] = { { NULL, "4 Bucket", BUCKET_SIZE(4), 4096 }, { NULL, "6 Bucket", BUCKET_SIZE(6), 3072 }, { NULL, "8 Bucket", BUCKET_SIZE(8), 2048 }, { NULL, "12 Bucket", BUCKET_SIZE(12), 1536 }, { NULL, "16 Bucket", BUCKET_SIZE(16), 1024 }, { NULL, "32 Bucket", BUCKET_SIZE(32), 512 }, { NULL, "64 Bucket", BUCKET_SIZE(64), 256 }, { NULL, "128 Bucket", BUCKET_SIZE(128), 128 }, { NULL, "256 Bucket", BUCKET_SIZE(256), 64 }, { NULL, NULL, 0} }; /* * Flags and enumerations to be passed to internal functions. */ enum zfreeskip { SKIP_NONE = 0, SKIP_DTOR, SKIP_FINI }; /* Prototypes.. */ static void *noobj_alloc(uma_zone_t, vm_size_t, uint8_t *, int); static void *page_alloc(uma_zone_t, vm_size_t, uint8_t *, int); static void *startup_alloc(uma_zone_t, vm_size_t, uint8_t *, int); static void page_free(void *, vm_size_t, uint8_t); static uma_slab_t keg_alloc_slab(uma_keg_t, uma_zone_t, int); static void cache_drain(uma_zone_t); static void bucket_drain(uma_zone_t, uma_bucket_t); static void bucket_cache_drain(uma_zone_t zone); static int keg_ctor(void *, int, void *, int); static void keg_dtor(void *, int, void *); static int zone_ctor(void *, int, void *, int); static void zone_dtor(void *, int, void *); static int zero_init(void *, int, int); static void keg_small_init(uma_keg_t keg); static void keg_large_init(uma_keg_t keg); static void zone_foreach(void (*zfunc)(uma_zone_t)); static void zone_timeout(uma_zone_t zone); static int hash_alloc(struct uma_hash *); static int hash_expand(struct uma_hash *, struct uma_hash *); static void hash_free(struct uma_hash *hash); static void uma_timeout(void *); static void uma_startup3(void); static void *zone_alloc_item(uma_zone_t, void *, int); static void zone_free_item(uma_zone_t, void *, void *, enum zfreeskip); static void bucket_enable(void); static void bucket_init(void); static uma_bucket_t bucket_alloc(uma_zone_t zone, void *, int); static void bucket_free(uma_zone_t zone, uma_bucket_t, void *); static void bucket_zone_drain(void); static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *, int flags); static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t last, int flags); static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int flags); static void *slab_alloc_item(uma_keg_t keg, uma_slab_t slab); static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item); static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, int align, uint32_t flags); static int zone_import(uma_zone_t zone, void **bucket, int max, int flags); static void zone_release(uma_zone_t zone, void **bucket, int cnt); static void uma_zero_item(void *item, uma_zone_t zone); void uma_print_zone(uma_zone_t); void uma_print_stats(void); static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS); static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS); #ifdef INVARIANTS static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item); static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item); #endif SYSINIT(uma_startup3, SI_SUB_VM_CONF, SI_ORDER_SECOND, uma_startup3, NULL); SYSCTL_PROC(_vm, OID_AUTO, zone_count, CTLFLAG_RD|CTLTYPE_INT, 0, 0, sysctl_vm_zone_count, "I", "Number of UMA zones"); SYSCTL_PROC(_vm, OID_AUTO, zone_stats, CTLFLAG_RD|CTLTYPE_STRUCT, 0, 0, sysctl_vm_zone_stats, "s,struct uma_type_header", "Zone Stats"); static int zone_warnings = 1; SYSCTL_INT(_vm, OID_AUTO, zone_warnings, CTLFLAG_RWTUN, &zone_warnings, 0, "Warn when UMA zones becomes full"); /* * This routine checks to see whether or not it's safe to enable buckets. */ static void bucket_enable(void) { bucketdisable = vm_page_count_min(); } /* * Initialize bucket_zones, the array of zones of buckets of various sizes. * * For each zone, calculate the memory required for each bucket, consisting * of the header and an array of pointers. */ static void bucket_init(void) { struct uma_bucket_zone *ubz; int size; for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) { size = roundup(sizeof(struct uma_bucket), sizeof(void *)); size += sizeof(void *) * ubz->ubz_entries; ubz->ubz_zone = uma_zcreate(ubz->ubz_name, size, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_MTXCLASS | UMA_ZFLAG_BUCKET); } } /* * Given a desired number of entries for a bucket, return the zone from which * to allocate the bucket. */ static struct uma_bucket_zone * bucket_zone_lookup(int entries) { struct uma_bucket_zone *ubz; for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) if (ubz->ubz_entries >= entries) return (ubz); ubz--; return (ubz); } static int bucket_select(int size) { struct uma_bucket_zone *ubz; ubz = &bucket_zones[0]; if (size > ubz->ubz_maxsize) return MAX((ubz->ubz_maxsize * ubz->ubz_entries) / size, 1); for (; ubz->ubz_entries != 0; ubz++) if (ubz->ubz_maxsize < size) break; ubz--; return (ubz->ubz_entries); } static uma_bucket_t bucket_alloc(uma_zone_t zone, void *udata, int flags) { struct uma_bucket_zone *ubz; uma_bucket_t bucket; /* * This is to stop us from allocating per cpu buckets while we're * running out of vm.boot_pages. Otherwise, we would exhaust the * boot pages. This also prevents us from allocating buckets in * low memory situations. */ if (bucketdisable) return (NULL); /* * To limit bucket recursion we store the original zone flags * in a cookie passed via zalloc_arg/zfree_arg. This allows the * NOVM flag to persist even through deep recursions. We also * store ZFLAG_BUCKET once we have recursed attempting to allocate * a bucket for a bucket zone so we do not allow infinite bucket * recursion. This cookie will even persist to frees of unused * buckets via the allocation path or bucket allocations in the * free path. */ if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) udata = (void *)(uintptr_t)zone->uz_flags; else { if ((uintptr_t)udata & UMA_ZFLAG_BUCKET) return (NULL); udata = (void *)((uintptr_t)udata | UMA_ZFLAG_BUCKET); } if ((uintptr_t)udata & UMA_ZFLAG_CACHEONLY) flags |= M_NOVM; ubz = bucket_zone_lookup(zone->uz_count); if (ubz->ubz_zone == zone && (ubz + 1)->ubz_entries != 0) ubz++; bucket = uma_zalloc_arg(ubz->ubz_zone, udata, flags); if (bucket) { #ifdef INVARIANTS bzero(bucket->ub_bucket, sizeof(void *) * ubz->ubz_entries); #endif bucket->ub_cnt = 0; bucket->ub_entries = ubz->ubz_entries; } return (bucket); } static void bucket_free(uma_zone_t zone, uma_bucket_t bucket, void *udata) { struct uma_bucket_zone *ubz; KASSERT(bucket->ub_cnt == 0, ("bucket_free: Freeing a non free bucket.")); if ((zone->uz_flags & UMA_ZFLAG_BUCKET) == 0) udata = (void *)(uintptr_t)zone->uz_flags; ubz = bucket_zone_lookup(bucket->ub_entries); uma_zfree_arg(ubz->ubz_zone, bucket, udata); } static void bucket_zone_drain(void) { struct uma_bucket_zone *ubz; for (ubz = &bucket_zones[0]; ubz->ubz_entries != 0; ubz++) zone_drain(ubz->ubz_zone); } static void zone_log_warning(uma_zone_t zone) { static const struct timeval warninterval = { 300, 0 }; if (!zone_warnings || zone->uz_warning == NULL) return; if (ratecheck(&zone->uz_ratecheck, &warninterval)) printf("[zone: %s] %s\n", zone->uz_name, zone->uz_warning); } static inline void zone_maxaction(uma_zone_t zone) { if (zone->uz_maxaction.ta_func != NULL) taskqueue_enqueue(taskqueue_thread, &zone->uz_maxaction); } static void zone_foreach_keg(uma_zone_t zone, void (*kegfn)(uma_keg_t)) { uma_klink_t klink; LIST_FOREACH(klink, &zone->uz_kegs, kl_link) kegfn(klink->kl_keg); } /* * Routine called by timeout which is used to fire off some time interval * based calculations. (stats, hash size, etc.) * * Arguments: * arg Unused * * Returns: * Nothing */ static void uma_timeout(void *unused) { bucket_enable(); zone_foreach(zone_timeout); /* Reschedule this event */ callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); } /* * Routine to perform timeout driven calculations. This expands the * hashes and does per cpu statistics aggregation. * * Returns nothing. */ static void keg_timeout(uma_keg_t keg) { KEG_LOCK(keg); /* * Expand the keg hash table. * * This is done if the number of slabs is larger than the hash size. * What I'm trying to do here is completely reduce collisions. This * may be a little aggressive. Should I allow for two collisions max? */ if (keg->uk_flags & UMA_ZONE_HASH && keg->uk_pages / keg->uk_ppera >= keg->uk_hash.uh_hashsize) { struct uma_hash newhash; struct uma_hash oldhash; int ret; /* * This is so involved because allocating and freeing * while the keg lock is held will lead to deadlock. * I have to do everything in stages and check for * races. */ newhash = keg->uk_hash; KEG_UNLOCK(keg); ret = hash_alloc(&newhash); KEG_LOCK(keg); if (ret) { if (hash_expand(&keg->uk_hash, &newhash)) { oldhash = keg->uk_hash; keg->uk_hash = newhash; } else oldhash = newhash; KEG_UNLOCK(keg); hash_free(&oldhash); return; } } KEG_UNLOCK(keg); } static void zone_timeout(uma_zone_t zone) { zone_foreach_keg(zone, &keg_timeout); } /* * Allocate and zero fill the next sized hash table from the appropriate * backing store. * * Arguments: * hash A new hash structure with the old hash size in uh_hashsize * * Returns: * 1 on success and 0 on failure. */ static int hash_alloc(struct uma_hash *hash) { int oldsize; int alloc; oldsize = hash->uh_hashsize; /* We're just going to go to a power of two greater */ if (oldsize) { hash->uh_hashsize = oldsize * 2; alloc = sizeof(hash->uh_slab_hash[0]) * hash->uh_hashsize; hash->uh_slab_hash = (struct slabhead *)malloc(alloc, M_UMAHASH, M_NOWAIT); } else { alloc = sizeof(hash->uh_slab_hash[0]) * UMA_HASH_SIZE_INIT; hash->uh_slab_hash = zone_alloc_item(hashzone, NULL, M_WAITOK); hash->uh_hashsize = UMA_HASH_SIZE_INIT; } if (hash->uh_slab_hash) { bzero(hash->uh_slab_hash, alloc); hash->uh_hashmask = hash->uh_hashsize - 1; return (1); } return (0); } /* * Expands the hash table for HASH zones. This is done from zone_timeout * to reduce collisions. This must not be done in the regular allocation * path, otherwise, we can recurse on the vm while allocating pages. * * Arguments: * oldhash The hash you want to expand * newhash The hash structure for the new table * * Returns: * Nothing * * Discussion: */ static int hash_expand(struct uma_hash *oldhash, struct uma_hash *newhash) { uma_slab_t slab; int hval; int i; if (!newhash->uh_slab_hash) return (0); if (oldhash->uh_hashsize >= newhash->uh_hashsize) return (0); /* * I need to investigate hash algorithms for resizing without a * full rehash. */ for (i = 0; i < oldhash->uh_hashsize; i++) while (!SLIST_EMPTY(&oldhash->uh_slab_hash[i])) { slab = SLIST_FIRST(&oldhash->uh_slab_hash[i]); SLIST_REMOVE_HEAD(&oldhash->uh_slab_hash[i], us_hlink); hval = UMA_HASH(newhash, slab->us_data); SLIST_INSERT_HEAD(&newhash->uh_slab_hash[hval], slab, us_hlink); } return (1); } /* * Free the hash bucket to the appropriate backing store. * * Arguments: * slab_hash The hash bucket we're freeing * hashsize The number of entries in that hash bucket * * Returns: * Nothing */ static void hash_free(struct uma_hash *hash) { if (hash->uh_slab_hash == NULL) return; if (hash->uh_hashsize == UMA_HASH_SIZE_INIT) zone_free_item(hashzone, hash->uh_slab_hash, NULL, SKIP_NONE); else free(hash->uh_slab_hash, M_UMAHASH); } /* * Frees all outstanding items in a bucket * * Arguments: * zone The zone to free to, must be unlocked. * bucket The free/alloc bucket with items, cpu queue must be locked. * * Returns: * Nothing */ static void bucket_drain(uma_zone_t zone, uma_bucket_t bucket) { int i; if (bucket == NULL) return; if (zone->uz_fini) for (i = 0; i < bucket->ub_cnt; i++) zone->uz_fini(bucket->ub_bucket[i], zone->uz_size); zone->uz_release(zone->uz_arg, bucket->ub_bucket, bucket->ub_cnt); bucket->ub_cnt = 0; } /* * Drains the per cpu caches for a zone. * * NOTE: This may only be called while the zone is being turn down, and not * during normal operation. This is necessary in order that we do not have * to migrate CPUs to drain the per-CPU caches. * * Arguments: * zone The zone to drain, must be unlocked. * * Returns: * Nothing */ static void cache_drain(uma_zone_t zone) { uma_cache_t cache; int cpu; /* * XXX: It is safe to not lock the per-CPU caches, because we're * tearing down the zone anyway. I.e., there will be no further use * of the caches at this point. * * XXX: It would good to be able to assert that the zone is being * torn down to prevent improper use of cache_drain(). * * XXX: We lock the zone before passing into bucket_cache_drain() as * it is used elsewhere. Should the tear-down path be made special * there in some form? */ CPU_FOREACH(cpu) { cache = &zone->uz_cpu[cpu]; bucket_drain(zone, cache->uc_allocbucket); bucket_drain(zone, cache->uc_freebucket); if (cache->uc_allocbucket != NULL) bucket_free(zone, cache->uc_allocbucket, NULL); if (cache->uc_freebucket != NULL) bucket_free(zone, cache->uc_freebucket, NULL); cache->uc_allocbucket = cache->uc_freebucket = NULL; } ZONE_LOCK(zone); bucket_cache_drain(zone); ZONE_UNLOCK(zone); } static void cache_shrink(uma_zone_t zone) { if (zone->uz_flags & UMA_ZFLAG_INTERNAL) return; ZONE_LOCK(zone); zone->uz_count = (zone->uz_count_min + zone->uz_count) / 2; ZONE_UNLOCK(zone); } static void cache_drain_safe_cpu(uma_zone_t zone) { uma_cache_t cache; uma_bucket_t b1, b2; if (zone->uz_flags & UMA_ZFLAG_INTERNAL) return; b1 = b2 = NULL; ZONE_LOCK(zone); critical_enter(); cache = &zone->uz_cpu[curcpu]; if (cache->uc_allocbucket) { if (cache->uc_allocbucket->ub_cnt != 0) LIST_INSERT_HEAD(&zone->uz_buckets, cache->uc_allocbucket, ub_link); else b1 = cache->uc_allocbucket; cache->uc_allocbucket = NULL; } if (cache->uc_freebucket) { if (cache->uc_freebucket->ub_cnt != 0) LIST_INSERT_HEAD(&zone->uz_buckets, cache->uc_freebucket, ub_link); else b2 = cache->uc_freebucket; cache->uc_freebucket = NULL; } critical_exit(); ZONE_UNLOCK(zone); if (b1) bucket_free(zone, b1, NULL); if (b2) bucket_free(zone, b2, NULL); } /* * Safely drain per-CPU caches of a zone(s) to alloc bucket. * This is an expensive call because it needs to bind to all CPUs * one by one and enter a critical section on each of them in order * to safely access their cache buckets. * Zone lock must not be held on call this function. */ static void cache_drain_safe(uma_zone_t zone) { int cpu; /* * Polite bucket sizes shrinking was not enouth, shrink aggressively. */ if (zone) cache_shrink(zone); else zone_foreach(cache_shrink); CPU_FOREACH(cpu) { thread_lock(curthread); sched_bind(curthread, cpu); thread_unlock(curthread); if (zone) cache_drain_safe_cpu(zone); else zone_foreach(cache_drain_safe_cpu); } thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); } /* * Drain the cached buckets from a zone. Expects a locked zone on entry. */ static void bucket_cache_drain(uma_zone_t zone) { uma_bucket_t bucket; /* * Drain the bucket queues and free the buckets, we just keep two per * cpu (alloc/free). */ while ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { LIST_REMOVE(bucket, ub_link); ZONE_UNLOCK(zone); bucket_drain(zone, bucket); bucket_free(zone, bucket, NULL); ZONE_LOCK(zone); } /* * Shrink further bucket sizes. Price of single zone lock collision * is probably lower then price of global cache drain. */ if (zone->uz_count > zone->uz_count_min) zone->uz_count--; } static void keg_free_slab(uma_keg_t keg, uma_slab_t slab, int start) { uint8_t *mem; int i; uint8_t flags; mem = slab->us_data; flags = slab->us_flags; i = start; if (keg->uk_fini != NULL) { for (i--; i > -1; i--) keg->uk_fini(slab->us_data + (keg->uk_rsize * i), keg->uk_size); } if (keg->uk_flags & UMA_ZONE_OFFPAGE) zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); #ifdef UMA_DEBUG printf("%s: Returning %d bytes.\n", keg->uk_name, PAGE_SIZE * keg->uk_ppera); #endif keg->uk_freef(mem, PAGE_SIZE * keg->uk_ppera, flags); } /* * Frees pages from a keg back to the system. This is done on demand from * the pageout daemon. * * Returns nothing. */ static void keg_drain(uma_keg_t keg) { struct slabhead freeslabs = { 0 }; uma_slab_t slab; uma_slab_t n; /* * We don't want to take pages from statically allocated kegs at this * time */ if (keg->uk_flags & UMA_ZONE_NOFREE || keg->uk_freef == NULL) return; #ifdef UMA_DEBUG printf("%s free items: %u\n", keg->uk_name, keg->uk_free); #endif KEG_LOCK(keg); if (keg->uk_free == 0) goto finished; slab = LIST_FIRST(&keg->uk_free_slab); while (slab) { n = LIST_NEXT(slab, us_link); /* We have no where to free these to */ if (slab->us_flags & UMA_SLAB_BOOT) { slab = n; continue; } LIST_REMOVE(slab, us_link); keg->uk_pages -= keg->uk_ppera; keg->uk_free -= keg->uk_ipers; if (keg->uk_flags & UMA_ZONE_HASH) UMA_HASH_REMOVE(&keg->uk_hash, slab, slab->us_data); SLIST_INSERT_HEAD(&freeslabs, slab, us_hlink); slab = n; } finished: KEG_UNLOCK(keg); while ((slab = SLIST_FIRST(&freeslabs)) != NULL) { SLIST_REMOVE(&freeslabs, slab, uma_slab, us_hlink); keg_free_slab(keg, slab, keg->uk_ipers); } } static void zone_drain_wait(uma_zone_t zone, int waitok) { /* * Set draining to interlock with zone_dtor() so we can release our * locks as we go. Only dtor() should do a WAITOK call since it * is the only call that knows the structure will still be available * when it wakes up. */ ZONE_LOCK(zone); while (zone->uz_flags & UMA_ZFLAG_DRAINING) { if (waitok == M_NOWAIT) goto out; msleep(zone, zone->uz_lockptr, PVM, "zonedrain", 1); } zone->uz_flags |= UMA_ZFLAG_DRAINING; bucket_cache_drain(zone); ZONE_UNLOCK(zone); /* * The DRAINING flag protects us from being freed while * we're running. Normally the uma_rwlock would protect us but we * must be able to release and acquire the right lock for each keg. */ zone_foreach_keg(zone, &keg_drain); ZONE_LOCK(zone); zone->uz_flags &= ~UMA_ZFLAG_DRAINING; wakeup(zone); out: ZONE_UNLOCK(zone); } void zone_drain(uma_zone_t zone) { zone_drain_wait(zone, M_NOWAIT); } /* * Allocate a new slab for a keg. This does not insert the slab onto a list. * * Arguments: * wait Shall we wait? * * Returns: * The slab that was allocated or NULL if there is no memory and the * caller specified M_NOWAIT. */ static uma_slab_t keg_alloc_slab(uma_keg_t keg, uma_zone_t zone, int wait) { uma_alloc allocf; uma_slab_t slab; uint8_t *mem; uint8_t flags; int i; mtx_assert(&keg->uk_lock, MA_OWNED); slab = NULL; mem = NULL; #ifdef UMA_DEBUG printf("alloc_slab: Allocating a new slab for %s\n", keg->uk_name); #endif allocf = keg->uk_allocf; KEG_UNLOCK(keg); if (keg->uk_flags & UMA_ZONE_OFFPAGE) { slab = zone_alloc_item(keg->uk_slabzone, NULL, wait); if (slab == NULL) goto out; } /* * This reproduces the old vm_zone behavior of zero filling pages the * first time they are added to a zone. * * Malloced items are zeroed in uma_zalloc. */ if ((keg->uk_flags & UMA_ZONE_MALLOC) == 0) wait |= M_ZERO; else wait &= ~M_ZERO; if (keg->uk_flags & UMA_ZONE_NODUMP) wait |= M_NODUMP; /* zone is passed for legacy reasons. */ mem = allocf(zone, keg->uk_ppera * PAGE_SIZE, &flags, wait); if (mem == NULL) { if (keg->uk_flags & UMA_ZONE_OFFPAGE) zone_free_item(keg->uk_slabzone, slab, NULL, SKIP_NONE); slab = NULL; goto out; } /* Point the slab into the allocated memory */ if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) slab = (uma_slab_t )(mem + keg->uk_pgoff); if (keg->uk_flags & UMA_ZONE_VTOSLAB) for (i = 0; i < keg->uk_ppera; i++) vsetslab((vm_offset_t)mem + (i * PAGE_SIZE), slab); slab->us_keg = keg; slab->us_data = mem; slab->us_freecount = keg->uk_ipers; slab->us_flags = flags; BIT_FILL(SLAB_SETSIZE, &slab->us_free); #ifdef INVARIANTS BIT_ZERO(SLAB_SETSIZE, &slab->us_debugfree); #endif if (keg->uk_init != NULL) { for (i = 0; i < keg->uk_ipers; i++) if (keg->uk_init(slab->us_data + (keg->uk_rsize * i), keg->uk_size, wait) != 0) break; if (i != keg->uk_ipers) { keg_free_slab(keg, slab, i); slab = NULL; goto out; } } out: KEG_LOCK(keg); if (slab != NULL) { if (keg->uk_flags & UMA_ZONE_HASH) UMA_HASH_INSERT(&keg->uk_hash, slab, mem); keg->uk_pages += keg->uk_ppera; keg->uk_free += keg->uk_ipers; } return (slab); } /* * This function is intended to be used early on in place of page_alloc() so * that we may use the boot time page cache to satisfy allocations before * the VM is ready. */ static void * startup_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) { uma_keg_t keg; uma_slab_t tmps; int pages, check_pages; keg = zone_first_keg(zone); pages = howmany(bytes, PAGE_SIZE); check_pages = pages - 1; KASSERT(pages > 0, ("startup_alloc can't reserve 0 pages\n")); /* * Check our small startup cache to see if it has pages remaining. */ mtx_lock(&uma_boot_pages_mtx); /* First check if we have enough room. */ tmps = LIST_FIRST(&uma_boot_pages); while (tmps != NULL && check_pages-- > 0) tmps = LIST_NEXT(tmps, us_link); if (tmps != NULL) { /* * It's ok to lose tmps references. The last one will * have tmps->us_data pointing to the start address of * "pages" contiguous pages of memory. */ while (pages-- > 0) { tmps = LIST_FIRST(&uma_boot_pages); LIST_REMOVE(tmps, us_link); } mtx_unlock(&uma_boot_pages_mtx); *pflag = tmps->us_flags; return (tmps->us_data); } mtx_unlock(&uma_boot_pages_mtx); if (booted < UMA_STARTUP2) panic("UMA: Increase vm.boot_pages"); /* * Now that we've booted reset these users to their real allocator. */ #ifdef UMA_MD_SMALL_ALLOC keg->uk_allocf = (keg->uk_ppera > 1) ? page_alloc : uma_small_alloc; #else keg->uk_allocf = page_alloc; #endif return keg->uk_allocf(zone, bytes, pflag, wait); } /* * Allocates a number of pages from the system * * Arguments: * bytes The number of bytes requested * wait Shall we wait? * * Returns: * A pointer to the alloced memory or possibly * NULL if M_NOWAIT is set. */ static void * page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait) { void *p; /* Returned page */ *pflag = UMA_SLAB_KMEM; p = (void *) kmem_malloc(kmem_arena, bytes, wait); return (p); } /* * Allocates a number of pages from within an object * * Arguments: * bytes The number of bytes requested * wait Shall we wait? * * Returns: * A pointer to the alloced memory or possibly * NULL if M_NOWAIT is set. */ static void * noobj_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags, int wait) { TAILQ_HEAD(, vm_page) alloctail; u_long npages; vm_offset_t retkva, zkva; vm_page_t p, p_next; uma_keg_t keg; TAILQ_INIT(&alloctail); keg = zone_first_keg(zone); npages = howmany(bytes, PAGE_SIZE); while (npages > 0) { p = vm_page_alloc(NULL, 0, VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ); if (p != NULL) { /* * Since the page does not belong to an object, its * listq is unused. */ TAILQ_INSERT_TAIL(&alloctail, p, listq); npages--; continue; } if (wait & M_WAITOK) { VM_WAIT; continue; } /* * Page allocation failed, free intermediate pages and * exit. */ TAILQ_FOREACH_SAFE(p, &alloctail, listq, p_next) { vm_page_unwire(p, PQ_NONE); vm_page_free(p); } return (NULL); } *flags = UMA_SLAB_PRIV; zkva = keg->uk_kva + atomic_fetchadd_long(&keg->uk_offset, round_page(bytes)); retkva = zkva; TAILQ_FOREACH(p, &alloctail, listq) { pmap_qenter(zkva, &p, 1); zkva += PAGE_SIZE; } return ((void *)retkva); } /* * Frees a number of pages to the system * * Arguments: * mem A pointer to the memory to be freed * size The size of the memory being freed * flags The original p->us_flags field * * Returns: * Nothing */ static void page_free(void *mem, vm_size_t size, uint8_t flags) { struct vmem *vmem; if (flags & UMA_SLAB_KMEM) vmem = kmem_arena; else if (flags & UMA_SLAB_KERNEL) vmem = kernel_arena; else panic("UMA: page_free used with invalid flags %d", flags); kmem_free(vmem, (vm_offset_t)mem, size); } /* * Zero fill initializer * * Arguments/Returns follow uma_init specifications */ static int zero_init(void *mem, int size, int flags) { bzero(mem, size); return (0); } /* * Finish creating a small uma keg. This calculates ipers, and the keg size. * * Arguments * keg The zone we should initialize * * Returns * Nothing */ static void keg_small_init(uma_keg_t keg) { u_int rsize; u_int memused; u_int wastedspace; u_int shsize; if (keg->uk_flags & UMA_ZONE_PCPU) { - u_int ncpus = mp_ncpus ? mp_ncpus : MAXCPU; + u_int ncpus = (mp_maxid + 1) ? (mp_maxid + 1) : MAXCPU; keg->uk_slabsize = sizeof(struct pcpu); keg->uk_ppera = howmany(ncpus * sizeof(struct pcpu), PAGE_SIZE); } else { keg->uk_slabsize = UMA_SLAB_SIZE; keg->uk_ppera = 1; } /* * Calculate the size of each allocation (rsize) according to * alignment. If the requested size is smaller than we have * allocation bits for we round it up. */ rsize = keg->uk_size; if (rsize < keg->uk_slabsize / SLAB_SETSIZE) rsize = keg->uk_slabsize / SLAB_SETSIZE; if (rsize & keg->uk_align) rsize = (rsize & ~keg->uk_align) + (keg->uk_align + 1); keg->uk_rsize = rsize; KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0 || keg->uk_rsize < sizeof(struct pcpu), ("%s: size %u too large", __func__, keg->uk_rsize)); if (keg->uk_flags & UMA_ZONE_OFFPAGE) shsize = 0; else shsize = sizeof(struct uma_slab); keg->uk_ipers = (keg->uk_slabsize - shsize) / rsize; KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); memused = keg->uk_ipers * rsize + shsize; wastedspace = keg->uk_slabsize - memused; /* * We can't do OFFPAGE if we're internal or if we've been * asked to not go to the VM for buckets. If we do this we * may end up going to the VM for slabs which we do not * want to do if we're UMA_ZFLAG_CACHEONLY as a result * of UMA_ZONE_VM, which clearly forbids it. */ if ((keg->uk_flags & UMA_ZFLAG_INTERNAL) || (keg->uk_flags & UMA_ZFLAG_CACHEONLY)) return; /* * See if using an OFFPAGE slab will limit our waste. Only do * this if it permits more items per-slab. * * XXX We could try growing slabsize to limit max waste as well. * Historically this was not done because the VM could not * efficiently handle contiguous allocations. */ if ((wastedspace >= keg->uk_slabsize / UMA_MAX_WASTE) && (keg->uk_ipers < (keg->uk_slabsize / keg->uk_rsize))) { keg->uk_ipers = keg->uk_slabsize / keg->uk_rsize; KASSERT(keg->uk_ipers > 0 && keg->uk_ipers <= SLAB_SETSIZE, ("%s: keg->uk_ipers %u", __func__, keg->uk_ipers)); #ifdef UMA_DEBUG printf("UMA decided we need offpage slab headers for " "keg: %s, calculated wastedspace = %d, " "maximum wasted space allowed = %d, " "calculated ipers = %d, " "new wasted space = %d\n", keg->uk_name, wastedspace, keg->uk_slabsize / UMA_MAX_WASTE, keg->uk_ipers, keg->uk_slabsize - keg->uk_ipers * keg->uk_rsize); #endif keg->uk_flags |= UMA_ZONE_OFFPAGE; } if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) keg->uk_flags |= UMA_ZONE_HASH; } /* * Finish creating a large (> UMA_SLAB_SIZE) uma kegs. Just give in and do * OFFPAGE for now. When I can allow for more dynamic slab sizes this will be * more complicated. * * Arguments * keg The keg we should initialize * * Returns * Nothing */ static void keg_large_init(uma_keg_t keg) { u_int shsize; KASSERT(keg != NULL, ("Keg is null in keg_large_init")); KASSERT((keg->uk_flags & UMA_ZFLAG_CACHEONLY) == 0, ("keg_large_init: Cannot large-init a UMA_ZFLAG_CACHEONLY keg")); KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, ("%s: Cannot large-init a UMA_ZONE_PCPU keg", __func__)); keg->uk_ppera = howmany(keg->uk_size, PAGE_SIZE); keg->uk_slabsize = keg->uk_ppera * PAGE_SIZE; keg->uk_ipers = 1; keg->uk_rsize = keg->uk_size; /* We can't do OFFPAGE if we're internal, bail out here. */ if (keg->uk_flags & UMA_ZFLAG_INTERNAL) return; /* Check whether we have enough space to not do OFFPAGE. */ if ((keg->uk_flags & UMA_ZONE_OFFPAGE) == 0) { shsize = sizeof(struct uma_slab); if (shsize & UMA_ALIGN_PTR) shsize = (shsize & ~UMA_ALIGN_PTR) + (UMA_ALIGN_PTR + 1); if ((PAGE_SIZE * keg->uk_ppera) - keg->uk_rsize < shsize) keg->uk_flags |= UMA_ZONE_OFFPAGE; } if ((keg->uk_flags & UMA_ZONE_OFFPAGE) && (keg->uk_flags & UMA_ZONE_VTOSLAB) == 0) keg->uk_flags |= UMA_ZONE_HASH; } static void keg_cachespread_init(uma_keg_t keg) { int alignsize; int trailer; int pages; int rsize; KASSERT((keg->uk_flags & UMA_ZONE_PCPU) == 0, ("%s: Cannot cachespread-init a UMA_ZONE_PCPU keg", __func__)); alignsize = keg->uk_align + 1; rsize = keg->uk_size; /* * We want one item to start on every align boundary in a page. To * do this we will span pages. We will also extend the item by the * size of align if it is an even multiple of align. Otherwise, it * would fall on the same boundary every time. */ if (rsize & keg->uk_align) rsize = (rsize & ~keg->uk_align) + alignsize; if ((rsize & alignsize) == 0) rsize += alignsize; trailer = rsize - keg->uk_size; pages = (rsize * (PAGE_SIZE / alignsize)) / PAGE_SIZE; pages = MIN(pages, (128 * 1024) / PAGE_SIZE); keg->uk_rsize = rsize; keg->uk_ppera = pages; keg->uk_slabsize = UMA_SLAB_SIZE; keg->uk_ipers = ((pages * PAGE_SIZE) + trailer) / rsize; keg->uk_flags |= UMA_ZONE_OFFPAGE | UMA_ZONE_VTOSLAB; KASSERT(keg->uk_ipers <= SLAB_SETSIZE, ("%s: keg->uk_ipers too high(%d) increase max_ipers", __func__, keg->uk_ipers)); } /* * Keg header ctor. This initializes all fields, locks, etc. And inserts * the keg onto the global keg list. * * Arguments/Returns follow uma_ctor specifications * udata Actually uma_kctor_args */ static int keg_ctor(void *mem, int size, void *udata, int flags) { struct uma_kctor_args *arg = udata; uma_keg_t keg = mem; uma_zone_t zone; bzero(keg, size); keg->uk_size = arg->size; keg->uk_init = arg->uminit; keg->uk_fini = arg->fini; keg->uk_align = arg->align; keg->uk_free = 0; keg->uk_reserve = 0; keg->uk_pages = 0; keg->uk_flags = arg->flags; keg->uk_allocf = page_alloc; keg->uk_freef = page_free; keg->uk_slabzone = NULL; /* * The master zone is passed to us at keg-creation time. */ zone = arg->zone; keg->uk_name = zone->uz_name; if (arg->flags & UMA_ZONE_VM) keg->uk_flags |= UMA_ZFLAG_CACHEONLY; if (arg->flags & UMA_ZONE_ZINIT) keg->uk_init = zero_init; if (arg->flags & UMA_ZONE_MALLOC) keg->uk_flags |= UMA_ZONE_VTOSLAB; if (arg->flags & UMA_ZONE_PCPU) #ifdef SMP keg->uk_flags |= UMA_ZONE_OFFPAGE; #else keg->uk_flags &= ~UMA_ZONE_PCPU; #endif if (keg->uk_flags & UMA_ZONE_CACHESPREAD) { keg_cachespread_init(keg); } else { if (keg->uk_size > (UMA_SLAB_SIZE - sizeof(struct uma_slab))) keg_large_init(keg); else keg_small_init(keg); } if (keg->uk_flags & UMA_ZONE_OFFPAGE) keg->uk_slabzone = slabzone; /* * If we haven't booted yet we need allocations to go through the * startup cache until the vm is ready. */ if (keg->uk_ppera == 1) { #ifdef UMA_MD_SMALL_ALLOC keg->uk_allocf = uma_small_alloc; keg->uk_freef = uma_small_free; if (booted < UMA_STARTUP) keg->uk_allocf = startup_alloc; #else if (booted < UMA_STARTUP2) keg->uk_allocf = startup_alloc; #endif } else if (booted < UMA_STARTUP2 && (keg->uk_flags & UMA_ZFLAG_INTERNAL)) keg->uk_allocf = startup_alloc; /* * Initialize keg's lock */ KEG_LOCK_INIT(keg, (arg->flags & UMA_ZONE_MTXCLASS)); /* * If we're putting the slab header in the actual page we need to * figure out where in each page it goes. This calculates a right * justified offset into the memory on an ALIGN_PTR boundary. */ if (!(keg->uk_flags & UMA_ZONE_OFFPAGE)) { u_int totsize; /* Size of the slab struct and free list */ totsize = sizeof(struct uma_slab); if (totsize & UMA_ALIGN_PTR) totsize = (totsize & ~UMA_ALIGN_PTR) + (UMA_ALIGN_PTR + 1); keg->uk_pgoff = (PAGE_SIZE * keg->uk_ppera) - totsize; /* * The only way the following is possible is if with our * UMA_ALIGN_PTR adjustments we are now bigger than * UMA_SLAB_SIZE. I haven't checked whether this is * mathematically possible for all cases, so we make * sure here anyway. */ totsize = keg->uk_pgoff + sizeof(struct uma_slab); if (totsize > PAGE_SIZE * keg->uk_ppera) { printf("zone %s ipers %d rsize %d size %d\n", zone->uz_name, keg->uk_ipers, keg->uk_rsize, keg->uk_size); panic("UMA slab won't fit."); } } if (keg->uk_flags & UMA_ZONE_HASH) hash_alloc(&keg->uk_hash); #ifdef UMA_DEBUG printf("UMA: %s(%p) size %d(%d) flags %#x ipers %d ppera %d out %d free %d\n", zone->uz_name, zone, keg->uk_size, keg->uk_rsize, keg->uk_flags, keg->uk_ipers, keg->uk_ppera, (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free); #endif LIST_INSERT_HEAD(&keg->uk_zones, zone, uz_link); rw_wlock(&uma_rwlock); LIST_INSERT_HEAD(&uma_kegs, keg, uk_link); rw_wunlock(&uma_rwlock); return (0); } /* * Zone header ctor. This initializes all fields, locks, etc. * * Arguments/Returns follow uma_ctor specifications * udata Actually uma_zctor_args */ static int zone_ctor(void *mem, int size, void *udata, int flags) { struct uma_zctor_args *arg = udata; uma_zone_t zone = mem; uma_zone_t z; uma_keg_t keg; bzero(zone, size); zone->uz_name = arg->name; zone->uz_ctor = arg->ctor; zone->uz_dtor = arg->dtor; zone->uz_slab = zone_fetch_slab; zone->uz_init = NULL; zone->uz_fini = NULL; zone->uz_allocs = 0; zone->uz_frees = 0; zone->uz_fails = 0; zone->uz_sleeps = 0; zone->uz_count = 0; zone->uz_count_min = 0; zone->uz_flags = 0; zone->uz_warning = NULL; timevalclear(&zone->uz_ratecheck); keg = arg->keg; ZONE_LOCK_INIT(zone, (arg->flags & UMA_ZONE_MTXCLASS)); /* * This is a pure cache zone, no kegs. */ if (arg->import) { if (arg->flags & UMA_ZONE_VM) arg->flags |= UMA_ZFLAG_CACHEONLY; zone->uz_flags = arg->flags; zone->uz_size = arg->size; zone->uz_import = arg->import; zone->uz_release = arg->release; zone->uz_arg = arg->arg; zone->uz_lockptr = &zone->uz_lock; rw_wlock(&uma_rwlock); LIST_INSERT_HEAD(&uma_cachezones, zone, uz_link); rw_wunlock(&uma_rwlock); goto out; } /* * Use the regular zone/keg/slab allocator. */ zone->uz_import = (uma_import)zone_import; zone->uz_release = (uma_release)zone_release; zone->uz_arg = zone; if (arg->flags & UMA_ZONE_SECONDARY) { KASSERT(arg->keg != NULL, ("Secondary zone on zero'd keg")); zone->uz_init = arg->uminit; zone->uz_fini = arg->fini; zone->uz_lockptr = &keg->uk_lock; zone->uz_flags |= UMA_ZONE_SECONDARY; rw_wlock(&uma_rwlock); ZONE_LOCK(zone); LIST_FOREACH(z, &keg->uk_zones, uz_link) { if (LIST_NEXT(z, uz_link) == NULL) { LIST_INSERT_AFTER(z, zone, uz_link); break; } } ZONE_UNLOCK(zone); rw_wunlock(&uma_rwlock); } else if (keg == NULL) { if ((keg = uma_kcreate(zone, arg->size, arg->uminit, arg->fini, arg->align, arg->flags)) == NULL) return (ENOMEM); } else { struct uma_kctor_args karg; int error; /* We should only be here from uma_startup() */ karg.size = arg->size; karg.uminit = arg->uminit; karg.fini = arg->fini; karg.align = arg->align; karg.flags = arg->flags; karg.zone = zone; error = keg_ctor(arg->keg, sizeof(struct uma_keg), &karg, flags); if (error) return (error); } /* * Link in the first keg. */ zone->uz_klink.kl_keg = keg; LIST_INSERT_HEAD(&zone->uz_kegs, &zone->uz_klink, kl_link); zone->uz_lockptr = &keg->uk_lock; zone->uz_size = keg->uk_size; zone->uz_flags |= (keg->uk_flags & (UMA_ZONE_INHERIT | UMA_ZFLAG_INHERIT)); /* * Some internal zones don't have room allocated for the per cpu * caches. If we're internal, bail out here. */ if (keg->uk_flags & UMA_ZFLAG_INTERNAL) { KASSERT((zone->uz_flags & UMA_ZONE_SECONDARY) == 0, ("Secondary zone requested UMA_ZFLAG_INTERNAL")); return (0); } out: if ((arg->flags & UMA_ZONE_MAXBUCKET) == 0) zone->uz_count = bucket_select(zone->uz_size); else zone->uz_count = BUCKET_MAX; zone->uz_count_min = zone->uz_count; return (0); } /* * Keg header dtor. This frees all data, destroys locks, frees the hash * table and removes the keg from the global list. * * Arguments/Returns follow uma_dtor specifications * udata unused */ static void keg_dtor(void *arg, int size, void *udata) { uma_keg_t keg; keg = (uma_keg_t)arg; KEG_LOCK(keg); if (keg->uk_free != 0) { printf("Freed UMA keg (%s) was not empty (%d items). " " Lost %d pages of memory.\n", keg->uk_name ? keg->uk_name : "", keg->uk_free, keg->uk_pages); } KEG_UNLOCK(keg); hash_free(&keg->uk_hash); KEG_LOCK_FINI(keg); } /* * Zone header dtor. * * Arguments/Returns follow uma_dtor specifications * udata unused */ static void zone_dtor(void *arg, int size, void *udata) { uma_klink_t klink; uma_zone_t zone; uma_keg_t keg; zone = (uma_zone_t)arg; keg = zone_first_keg(zone); if (!(zone->uz_flags & UMA_ZFLAG_INTERNAL)) cache_drain(zone); rw_wlock(&uma_rwlock); LIST_REMOVE(zone, uz_link); rw_wunlock(&uma_rwlock); /* * XXX there are some races here where * the zone can be drained but zone lock * released and then refilled before we * remove it... we dont care for now */ zone_drain_wait(zone, M_WAITOK); /* * Unlink all of our kegs. */ while ((klink = LIST_FIRST(&zone->uz_kegs)) != NULL) { klink->kl_keg = NULL; LIST_REMOVE(klink, kl_link); if (klink == &zone->uz_klink) continue; free(klink, M_TEMP); } /* * We only destroy kegs from non secondary zones. */ if (keg != NULL && (zone->uz_flags & UMA_ZONE_SECONDARY) == 0) { rw_wlock(&uma_rwlock); LIST_REMOVE(keg, uk_link); rw_wunlock(&uma_rwlock); zone_free_item(kegs, keg, NULL, SKIP_NONE); } ZONE_LOCK_FINI(zone); } /* * Traverses every zone in the system and calls a callback * * Arguments: * zfunc A pointer to a function which accepts a zone * as an argument. * * Returns: * Nothing */ static void zone_foreach(void (*zfunc)(uma_zone_t)) { uma_keg_t keg; uma_zone_t zone; rw_rlock(&uma_rwlock); LIST_FOREACH(keg, &uma_kegs, uk_link) { LIST_FOREACH(zone, &keg->uk_zones, uz_link) zfunc(zone); } rw_runlock(&uma_rwlock); } /* Public functions */ /* See uma.h */ void uma_startup(void *bootmem, int boot_pages) { struct uma_zctor_args args; uma_slab_t slab; int i; #ifdef UMA_DEBUG printf("Creating uma keg headers zone and keg.\n"); #endif rw_init(&uma_rwlock, "UMA lock"); /* "manually" create the initial zone */ memset(&args, 0, sizeof(args)); args.name = "UMA Kegs"; args.size = sizeof(struct uma_keg); args.ctor = keg_ctor; args.dtor = keg_dtor; args.uminit = zero_init; args.fini = NULL; args.keg = &masterkeg; args.align = 32 - 1; args.flags = UMA_ZFLAG_INTERNAL; /* The initial zone has no Per cpu queues so it's smaller */ zone_ctor(kegs, sizeof(struct uma_zone), &args, M_WAITOK); #ifdef UMA_DEBUG printf("Filling boot free list.\n"); #endif for (i = 0; i < boot_pages; i++) { slab = (uma_slab_t)((uint8_t *)bootmem + (i * UMA_SLAB_SIZE)); slab->us_data = (uint8_t *)slab; slab->us_flags = UMA_SLAB_BOOT; LIST_INSERT_HEAD(&uma_boot_pages, slab, us_link); } mtx_init(&uma_boot_pages_mtx, "UMA boot pages", NULL, MTX_DEF); #ifdef UMA_DEBUG printf("Creating uma zone headers zone and keg.\n"); #endif args.name = "UMA Zones"; args.size = sizeof(struct uma_zone) + (sizeof(struct uma_cache) * (mp_maxid + 1)); args.ctor = zone_ctor; args.dtor = zone_dtor; args.uminit = zero_init; args.fini = NULL; args.keg = NULL; args.align = 32 - 1; args.flags = UMA_ZFLAG_INTERNAL; /* The initial zone has no Per cpu queues so it's smaller */ zone_ctor(zones, sizeof(struct uma_zone), &args, M_WAITOK); #ifdef UMA_DEBUG printf("Creating slab and hash zones.\n"); #endif /* Now make a zone for slab headers */ slabzone = uma_zcreate("UMA Slabs", sizeof(struct uma_slab), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); hashzone = uma_zcreate("UMA Hash", sizeof(struct slabhead *) * UMA_HASH_SIZE_INIT, NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZFLAG_INTERNAL); bucket_init(); booted = UMA_STARTUP; #ifdef UMA_DEBUG printf("UMA startup complete.\n"); #endif } /* see uma.h */ void uma_startup2(void) { booted = UMA_STARTUP2; bucket_enable(); sx_init(&uma_drain_lock, "umadrain"); #ifdef UMA_DEBUG printf("UMA startup2 complete.\n"); #endif } /* * Initialize our callout handle * */ static void uma_startup3(void) { #ifdef UMA_DEBUG printf("Starting callout.\n"); #endif callout_init(&uma_callout, 1); callout_reset(&uma_callout, UMA_TIMEOUT * hz, uma_timeout, NULL); #ifdef UMA_DEBUG printf("UMA startup3 complete.\n"); #endif } static uma_keg_t uma_kcreate(uma_zone_t zone, size_t size, uma_init uminit, uma_fini fini, int align, uint32_t flags) { struct uma_kctor_args args; args.size = size; args.uminit = uminit; args.fini = fini; args.align = (align == UMA_ALIGN_CACHE) ? uma_align_cache : align; args.flags = flags; args.zone = zone; return (zone_alloc_item(kegs, &args, M_WAITOK)); } /* See uma.h */ void uma_set_align(int align) { if (align != UMA_ALIGN_CACHE) uma_align_cache = align; } /* See uma.h */ uma_zone_t uma_zcreate(const char *name, size_t size, uma_ctor ctor, uma_dtor dtor, uma_init uminit, uma_fini fini, int align, uint32_t flags) { struct uma_zctor_args args; uma_zone_t res; bool locked; /* This stuff is essential for the zone ctor */ memset(&args, 0, sizeof(args)); args.name = name; args.size = size; args.ctor = ctor; args.dtor = dtor; args.uminit = uminit; args.fini = fini; #ifdef INVARIANTS /* * If a zone is being created with an empty constructor and * destructor, pass UMA constructor/destructor which checks for * memory use after free. */ if ((!(flags & (UMA_ZONE_ZINIT | UMA_ZONE_NOFREE))) && ctor == NULL && dtor == NULL && uminit == NULL && fini == NULL) { args.ctor = trash_ctor; args.dtor = trash_dtor; args.uminit = trash_init; args.fini = trash_fini; } #endif args.align = align; args.flags = flags; args.keg = NULL; if (booted < UMA_STARTUP2) { locked = false; } else { sx_slock(&uma_drain_lock); locked = true; } res = zone_alloc_item(zones, &args, M_WAITOK); if (locked) sx_sunlock(&uma_drain_lock); return (res); } /* See uma.h */ uma_zone_t uma_zsecond_create(char *name, uma_ctor ctor, uma_dtor dtor, uma_init zinit, uma_fini zfini, uma_zone_t master) { struct uma_zctor_args args; uma_keg_t keg; uma_zone_t res; bool locked; keg = zone_first_keg(master); memset(&args, 0, sizeof(args)); args.name = name; args.size = keg->uk_size; args.ctor = ctor; args.dtor = dtor; args.uminit = zinit; args.fini = zfini; args.align = keg->uk_align; args.flags = keg->uk_flags | UMA_ZONE_SECONDARY; args.keg = keg; if (booted < UMA_STARTUP2) { locked = false; } else { sx_slock(&uma_drain_lock); locked = true; } /* XXX Attaches only one keg of potentially many. */ res = zone_alloc_item(zones, &args, M_WAITOK); if (locked) sx_sunlock(&uma_drain_lock); return (res); } /* See uma.h */ uma_zone_t uma_zcache_create(char *name, int size, uma_ctor ctor, uma_dtor dtor, uma_init zinit, uma_fini zfini, uma_import zimport, uma_release zrelease, void *arg, int flags) { struct uma_zctor_args args; memset(&args, 0, sizeof(args)); args.name = name; args.size = size; args.ctor = ctor; args.dtor = dtor; args.uminit = zinit; args.fini = zfini; args.import = zimport; args.release = zrelease; args.arg = arg; args.align = 0; args.flags = flags; return (zone_alloc_item(zones, &args, M_WAITOK)); } static void zone_lock_pair(uma_zone_t a, uma_zone_t b) { if (a < b) { ZONE_LOCK(a); mtx_lock_flags(b->uz_lockptr, MTX_DUPOK); } else { ZONE_LOCK(b); mtx_lock_flags(a->uz_lockptr, MTX_DUPOK); } } static void zone_unlock_pair(uma_zone_t a, uma_zone_t b) { ZONE_UNLOCK(a); ZONE_UNLOCK(b); } int uma_zsecond_add(uma_zone_t zone, uma_zone_t master) { uma_klink_t klink; uma_klink_t kl; int error; error = 0; klink = malloc(sizeof(*klink), M_TEMP, M_WAITOK | M_ZERO); zone_lock_pair(zone, master); /* * zone must use vtoslab() to resolve objects and must already be * a secondary. */ if ((zone->uz_flags & (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) != (UMA_ZONE_VTOSLAB | UMA_ZONE_SECONDARY)) { error = EINVAL; goto out; } /* * The new master must also use vtoslab(). */ if ((zone->uz_flags & UMA_ZONE_VTOSLAB) != UMA_ZONE_VTOSLAB) { error = EINVAL; goto out; } /* * The underlying object must be the same size. rsize * may be different. */ if (master->uz_size != zone->uz_size) { error = E2BIG; goto out; } /* * Put it at the end of the list. */ klink->kl_keg = zone_first_keg(master); LIST_FOREACH(kl, &zone->uz_kegs, kl_link) { if (LIST_NEXT(kl, kl_link) == NULL) { LIST_INSERT_AFTER(kl, klink, kl_link); break; } } klink = NULL; zone->uz_flags |= UMA_ZFLAG_MULTI; zone->uz_slab = zone_fetch_slab_multi; out: zone_unlock_pair(zone, master); if (klink != NULL) free(klink, M_TEMP); return (error); } /* See uma.h */ void uma_zdestroy(uma_zone_t zone) { sx_slock(&uma_drain_lock); zone_free_item(zones, zone, NULL, SKIP_NONE); sx_sunlock(&uma_drain_lock); } /* See uma.h */ void * uma_zalloc_arg(uma_zone_t zone, void *udata, int flags) { void *item; uma_cache_t cache; uma_bucket_t bucket; int lockfail; int cpu; /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); /* This is the fast path allocation */ #ifdef UMA_DEBUG_ALLOC_1 printf("Allocating one item from %s(%p)\n", zone->uz_name, zone); #endif CTR3(KTR_UMA, "uma_zalloc_arg thread %x zone %s flags %d", curthread, zone->uz_name, flags); if (flags & M_WAITOK) { WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "uma_zalloc_arg: zone \"%s\"", zone->uz_name); } KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), ("uma_zalloc_arg: called with spinlock or critical section held")); #ifdef DEBUG_MEMGUARD if (memguard_cmp_zone(zone)) { item = memguard_alloc(zone->uz_size, flags); if (item != NULL) { if (zone->uz_init != NULL && zone->uz_init(item, zone->uz_size, flags) != 0) return (NULL); if (zone->uz_ctor != NULL && zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { zone->uz_fini(item, zone->uz_size); return (NULL); } return (item); } /* This is unfortunate but should not be fatal. */ } #endif /* * If possible, allocate from the per-CPU cache. There are two * requirements for safe access to the per-CPU cache: (1) the thread * accessing the cache must not be preempted or yield during access, * and (2) the thread must not migrate CPUs without switching which * cache it accesses. We rely on a critical section to prevent * preemption and migration. We release the critical section in * order to acquire the zone mutex if we are unable to allocate from * the current cache; when we re-acquire the critical section, we * must detect and handle migration if it has occurred. */ critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; zalloc_start: bucket = cache->uc_allocbucket; if (bucket != NULL && bucket->ub_cnt > 0) { bucket->ub_cnt--; item = bucket->ub_bucket[bucket->ub_cnt]; #ifdef INVARIANTS bucket->ub_bucket[bucket->ub_cnt] = NULL; #endif KASSERT(item != NULL, ("uma_zalloc: Bucket pointer mangled.")); cache->uc_allocs++; critical_exit(); if (zone->uz_ctor != NULL && zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { atomic_add_long(&zone->uz_fails, 1); zone_free_item(zone, item, udata, SKIP_DTOR); return (NULL); } #ifdef INVARIANTS uma_dbg_alloc(zone, NULL, item); #endif if (flags & M_ZERO) uma_zero_item(item, zone); return (item); } /* * We have run out of items in our alloc bucket. * See if we can switch with our free bucket. */ bucket = cache->uc_freebucket; if (bucket != NULL && bucket->ub_cnt > 0) { #ifdef UMA_DEBUG_ALLOC printf("uma_zalloc: Swapping empty with alloc.\n"); #endif cache->uc_freebucket = cache->uc_allocbucket; cache->uc_allocbucket = bucket; goto zalloc_start; } /* * Discard any empty allocation bucket while we hold no locks. */ bucket = cache->uc_allocbucket; cache->uc_allocbucket = NULL; critical_exit(); if (bucket != NULL) bucket_free(zone, bucket, udata); /* Short-circuit for zones without buckets and low memory. */ if (zone->uz_count == 0 || bucketdisable) goto zalloc_item; /* * Attempt to retrieve the item from the per-CPU cache has failed, so * we must go back to the zone. This requires the zone lock, so we * must drop the critical section, then re-acquire it when we go back * to the cache. Since the critical section is released, we may be * preempted or migrate. As such, make sure not to maintain any * thread-local state specific to the cache from prior to releasing * the critical section. */ lockfail = 0; if (ZONE_TRYLOCK(zone) == 0) { /* Record contention to size the buckets. */ ZONE_LOCK(zone); lockfail = 1; } critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; /* * Since we have locked the zone we may as well send back our stats. */ atomic_add_long(&zone->uz_allocs, cache->uc_allocs); atomic_add_long(&zone->uz_frees, cache->uc_frees); cache->uc_allocs = 0; cache->uc_frees = 0; /* See if we lost the race to fill the cache. */ if (cache->uc_allocbucket != NULL) { ZONE_UNLOCK(zone); goto zalloc_start; } /* * Check the zone's cache of buckets. */ if ((bucket = LIST_FIRST(&zone->uz_buckets)) != NULL) { KASSERT(bucket->ub_cnt != 0, ("uma_zalloc_arg: Returning an empty bucket.")); LIST_REMOVE(bucket, ub_link); cache->uc_allocbucket = bucket; ZONE_UNLOCK(zone); goto zalloc_start; } /* We are no longer associated with this CPU. */ critical_exit(); /* * We bump the uz count when the cache size is insufficient to * handle the working set. */ if (lockfail && zone->uz_count < BUCKET_MAX) zone->uz_count++; ZONE_UNLOCK(zone); /* * Now lets just fill a bucket and put it on the free list. If that * works we'll restart the allocation from the beginning and it * will use the just filled bucket. */ bucket = zone_alloc_bucket(zone, udata, flags); if (bucket != NULL) { ZONE_LOCK(zone); critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; /* * See if we lost the race or were migrated. Cache the * initialized bucket to make this less likely or claim * the memory directly. */ if (cache->uc_allocbucket == NULL) cache->uc_allocbucket = bucket; else LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); ZONE_UNLOCK(zone); goto zalloc_start; } /* * We may not be able to get a bucket so return an actual item. */ #ifdef UMA_DEBUG printf("uma_zalloc_arg: Bucketzone returned NULL\n"); #endif zalloc_item: item = zone_alloc_item(zone, udata, flags); return (item); } static uma_slab_t keg_fetch_slab(uma_keg_t keg, uma_zone_t zone, int flags) { uma_slab_t slab; int reserve; mtx_assert(&keg->uk_lock, MA_OWNED); slab = NULL; reserve = 0; if ((flags & M_USE_RESERVE) == 0) reserve = keg->uk_reserve; for (;;) { /* * Find a slab with some space. Prefer slabs that are partially * used over those that are totally full. This helps to reduce * fragmentation. */ if (keg->uk_free > reserve) { if (!LIST_EMPTY(&keg->uk_part_slab)) { slab = LIST_FIRST(&keg->uk_part_slab); } else { slab = LIST_FIRST(&keg->uk_free_slab); LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); } MPASS(slab->us_keg == keg); return (slab); } /* * M_NOVM means don't ask at all! */ if (flags & M_NOVM) break; if (keg->uk_maxpages && keg->uk_pages >= keg->uk_maxpages) { keg->uk_flags |= UMA_ZFLAG_FULL; /* * If this is not a multi-zone, set the FULL bit. * Otherwise slab_multi() takes care of it. */ if ((zone->uz_flags & UMA_ZFLAG_MULTI) == 0) { zone->uz_flags |= UMA_ZFLAG_FULL; zone_log_warning(zone); zone_maxaction(zone); } if (flags & M_NOWAIT) break; zone->uz_sleeps++; msleep(keg, &keg->uk_lock, PVM, "keglimit", 0); continue; } slab = keg_alloc_slab(keg, zone, flags); /* * If we got a slab here it's safe to mark it partially used * and return. We assume that the caller is going to remove * at least one item. */ if (slab) { MPASS(slab->us_keg == keg); LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); return (slab); } /* * We might not have been able to get a slab but another cpu * could have while we were unlocked. Check again before we * fail. */ flags |= M_NOVM; } return (slab); } static uma_slab_t zone_fetch_slab(uma_zone_t zone, uma_keg_t keg, int flags) { uma_slab_t slab; if (keg == NULL) { keg = zone_first_keg(zone); KEG_LOCK(keg); } for (;;) { slab = keg_fetch_slab(keg, zone, flags); if (slab) return (slab); if (flags & (M_NOWAIT | M_NOVM)) break; } KEG_UNLOCK(keg); return (NULL); } /* * uma_zone_fetch_slab_multi: Fetches a slab from one available keg. Returns * with the keg locked. On NULL no lock is held. * * The last pointer is used to seed the search. It is not required. */ static uma_slab_t zone_fetch_slab_multi(uma_zone_t zone, uma_keg_t last, int rflags) { uma_klink_t klink; uma_slab_t slab; uma_keg_t keg; int flags; int empty; int full; /* * Don't wait on the first pass. This will skip limit tests * as well. We don't want to block if we can find a provider * without blocking. */ flags = (rflags & ~M_WAITOK) | M_NOWAIT; /* * Use the last slab allocated as a hint for where to start * the search. */ if (last != NULL) { slab = keg_fetch_slab(last, zone, flags); if (slab) return (slab); KEG_UNLOCK(last); } /* * Loop until we have a slab incase of transient failures * while M_WAITOK is specified. I'm not sure this is 100% * required but we've done it for so long now. */ for (;;) { empty = 0; full = 0; /* * Search the available kegs for slabs. Be careful to hold the * correct lock while calling into the keg layer. */ LIST_FOREACH(klink, &zone->uz_kegs, kl_link) { keg = klink->kl_keg; KEG_LOCK(keg); if ((keg->uk_flags & UMA_ZFLAG_FULL) == 0) { slab = keg_fetch_slab(keg, zone, flags); if (slab) return (slab); } if (keg->uk_flags & UMA_ZFLAG_FULL) full++; else empty++; KEG_UNLOCK(keg); } if (rflags & (M_NOWAIT | M_NOVM)) break; flags = rflags; /* * All kegs are full. XXX We can't atomically check all kegs * and sleep so just sleep for a short period and retry. */ if (full && !empty) { ZONE_LOCK(zone); zone->uz_flags |= UMA_ZFLAG_FULL; zone->uz_sleeps++; zone_log_warning(zone); zone_maxaction(zone); msleep(zone, zone->uz_lockptr, PVM, "zonelimit", hz/100); zone->uz_flags &= ~UMA_ZFLAG_FULL; ZONE_UNLOCK(zone); continue; } } return (NULL); } static void * slab_alloc_item(uma_keg_t keg, uma_slab_t slab) { void *item; uint8_t freei; MPASS(keg == slab->us_keg); mtx_assert(&keg->uk_lock, MA_OWNED); freei = BIT_FFS(SLAB_SETSIZE, &slab->us_free) - 1; BIT_CLR(SLAB_SETSIZE, freei, &slab->us_free); item = slab->us_data + (keg->uk_rsize * freei); slab->us_freecount--; keg->uk_free--; /* Move this slab to the full list */ if (slab->us_freecount == 0) { LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&keg->uk_full_slab, slab, us_link); } return (item); } static int zone_import(uma_zone_t zone, void **bucket, int max, int flags) { uma_slab_t slab; uma_keg_t keg; int i; slab = NULL; keg = NULL; /* Try to keep the buckets totally full */ for (i = 0; i < max; ) { if ((slab = zone->uz_slab(zone, keg, flags)) == NULL) break; keg = slab->us_keg; while (slab->us_freecount && i < max) { bucket[i++] = slab_alloc_item(keg, slab); if (keg->uk_free <= keg->uk_reserve) break; } /* Don't grab more than one slab at a time. */ flags &= ~M_WAITOK; flags |= M_NOWAIT; } if (slab != NULL) KEG_UNLOCK(keg); return i; } static uma_bucket_t zone_alloc_bucket(uma_zone_t zone, void *udata, int flags) { uma_bucket_t bucket; int max; /* Don't wait for buckets, preserve caller's NOVM setting. */ bucket = bucket_alloc(zone, udata, M_NOWAIT | (flags & M_NOVM)); if (bucket == NULL) return (NULL); max = MIN(bucket->ub_entries, zone->uz_count); bucket->ub_cnt = zone->uz_import(zone->uz_arg, bucket->ub_bucket, max, flags); /* * Initialize the memory if necessary. */ if (bucket->ub_cnt != 0 && zone->uz_init != NULL) { int i; for (i = 0; i < bucket->ub_cnt; i++) if (zone->uz_init(bucket->ub_bucket[i], zone->uz_size, flags) != 0) break; /* * If we couldn't initialize the whole bucket, put the * rest back onto the freelist. */ if (i != bucket->ub_cnt) { zone->uz_release(zone->uz_arg, &bucket->ub_bucket[i], bucket->ub_cnt - i); #ifdef INVARIANTS bzero(&bucket->ub_bucket[i], sizeof(void *) * (bucket->ub_cnt - i)); #endif bucket->ub_cnt = i; } } if (bucket->ub_cnt == 0) { bucket_free(zone, bucket, udata); atomic_add_long(&zone->uz_fails, 1); return (NULL); } return (bucket); } /* * Allocates a single item from a zone. * * Arguments * zone The zone to alloc for. * udata The data to be passed to the constructor. * flags M_WAITOK, M_NOWAIT, M_ZERO. * * Returns * NULL if there is no memory and M_NOWAIT is set * An item if successful */ static void * zone_alloc_item(uma_zone_t zone, void *udata, int flags) { void *item; item = NULL; #ifdef UMA_DEBUG_ALLOC printf("INTERNAL: Allocating one item from %s(%p)\n", zone->uz_name, zone); #endif if (zone->uz_import(zone->uz_arg, &item, 1, flags) != 1) goto fail; atomic_add_long(&zone->uz_allocs, 1); /* * We have to call both the zone's init (not the keg's init) * and the zone's ctor. This is because the item is going from * a keg slab directly to the user, and the user is expecting it * to be both zone-init'd as well as zone-ctor'd. */ if (zone->uz_init != NULL) { if (zone->uz_init(item, zone->uz_size, flags) != 0) { zone_free_item(zone, item, udata, SKIP_FINI); goto fail; } } if (zone->uz_ctor != NULL) { if (zone->uz_ctor(item, zone->uz_size, udata, flags) != 0) { zone_free_item(zone, item, udata, SKIP_DTOR); goto fail; } } #ifdef INVARIANTS uma_dbg_alloc(zone, NULL, item); #endif if (flags & M_ZERO) uma_zero_item(item, zone); return (item); fail: atomic_add_long(&zone->uz_fails, 1); return (NULL); } /* See uma.h */ void uma_zfree_arg(uma_zone_t zone, void *item, void *udata) { uma_cache_t cache; uma_bucket_t bucket; int lockfail; int cpu; /* Enable entropy collection for RANDOM_ENABLE_UMA kernel option */ random_harvest_fast_uma(&zone, sizeof(zone), 1, RANDOM_UMA); #ifdef UMA_DEBUG_ALLOC_1 printf("Freeing item %p to %s(%p)\n", item, zone->uz_name, zone); #endif CTR2(KTR_UMA, "uma_zfree_arg thread %x zone %s", curthread, zone->uz_name); KASSERT(curthread->td_critnest == 0 || SCHEDULER_STOPPED(), ("uma_zfree_arg: called with spinlock or critical section held")); /* uma_zfree(..., NULL) does nothing, to match free(9). */ if (item == NULL) return; #ifdef DEBUG_MEMGUARD if (is_memguard_addr(item)) { if (zone->uz_dtor != NULL) zone->uz_dtor(item, zone->uz_size, udata); if (zone->uz_fini != NULL) zone->uz_fini(item, zone->uz_size); memguard_free(item); return; } #endif #ifdef INVARIANTS if (zone->uz_flags & UMA_ZONE_MALLOC) uma_dbg_free(zone, udata, item); else uma_dbg_free(zone, NULL, item); #endif if (zone->uz_dtor != NULL) zone->uz_dtor(item, zone->uz_size, udata); /* * The race here is acceptable. If we miss it we'll just have to wait * a little longer for the limits to be reset. */ if (zone->uz_flags & UMA_ZFLAG_FULL) goto zfree_item; /* * If possible, free to the per-CPU cache. There are two * requirements for safe access to the per-CPU cache: (1) the thread * accessing the cache must not be preempted or yield during access, * and (2) the thread must not migrate CPUs without switching which * cache it accesses. We rely on a critical section to prevent * preemption and migration. We release the critical section in * order to acquire the zone mutex if we are unable to free to the * current cache; when we re-acquire the critical section, we must * detect and handle migration if it has occurred. */ zfree_restart: critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; zfree_start: /* * Try to free into the allocbucket first to give LIFO ordering * for cache-hot datastructures. Spill over into the freebucket * if necessary. Alloc will swap them if one runs dry. */ bucket = cache->uc_allocbucket; if (bucket == NULL || bucket->ub_cnt >= bucket->ub_entries) bucket = cache->uc_freebucket; if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { KASSERT(bucket->ub_bucket[bucket->ub_cnt] == NULL, ("uma_zfree: Freeing to non free bucket index.")); bucket->ub_bucket[bucket->ub_cnt] = item; bucket->ub_cnt++; cache->uc_frees++; critical_exit(); return; } /* * We must go back the zone, which requires acquiring the zone lock, * which in turn means we must release and re-acquire the critical * section. Since the critical section is released, we may be * preempted or migrate. As such, make sure not to maintain any * thread-local state specific to the cache from prior to releasing * the critical section. */ critical_exit(); if (zone->uz_count == 0 || bucketdisable) goto zfree_item; lockfail = 0; if (ZONE_TRYLOCK(zone) == 0) { /* Record contention to size the buckets. */ ZONE_LOCK(zone); lockfail = 1; } critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; /* * Since we have locked the zone we may as well send back our stats. */ atomic_add_long(&zone->uz_allocs, cache->uc_allocs); atomic_add_long(&zone->uz_frees, cache->uc_frees); cache->uc_allocs = 0; cache->uc_frees = 0; bucket = cache->uc_freebucket; if (bucket != NULL && bucket->ub_cnt < bucket->ub_entries) { ZONE_UNLOCK(zone); goto zfree_start; } cache->uc_freebucket = NULL; /* Can we throw this on the zone full list? */ if (bucket != NULL) { #ifdef UMA_DEBUG_ALLOC printf("uma_zfree: Putting old bucket on the free list.\n"); #endif /* ub_cnt is pointing to the last free item */ KASSERT(bucket->ub_cnt != 0, ("uma_zfree: Attempting to insert an empty bucket onto the full list.\n")); LIST_INSERT_HEAD(&zone->uz_buckets, bucket, ub_link); } /* We are no longer associated with this CPU. */ critical_exit(); /* * We bump the uz count when the cache size is insufficient to * handle the working set. */ if (lockfail && zone->uz_count < BUCKET_MAX) zone->uz_count++; ZONE_UNLOCK(zone); #ifdef UMA_DEBUG_ALLOC printf("uma_zfree: Allocating new free bucket.\n"); #endif bucket = bucket_alloc(zone, udata, M_NOWAIT); if (bucket) { critical_enter(); cpu = curcpu; cache = &zone->uz_cpu[cpu]; if (cache->uc_freebucket == NULL) { cache->uc_freebucket = bucket; goto zfree_start; } /* * We lost the race, start over. We have to drop our * critical section to free the bucket. */ critical_exit(); bucket_free(zone, bucket, udata); goto zfree_restart; } /* * If nothing else caught this, we'll just do an internal free. */ zfree_item: zone_free_item(zone, item, udata, SKIP_DTOR); return; } static void slab_free_item(uma_keg_t keg, uma_slab_t slab, void *item) { uint8_t freei; mtx_assert(&keg->uk_lock, MA_OWNED); MPASS(keg == slab->us_keg); /* Do we need to remove from any lists? */ if (slab->us_freecount+1 == keg->uk_ipers) { LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); } else if (slab->us_freecount == 0) { LIST_REMOVE(slab, us_link); LIST_INSERT_HEAD(&keg->uk_part_slab, slab, us_link); } /* Slab management. */ freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; BIT_SET(SLAB_SETSIZE, freei, &slab->us_free); slab->us_freecount++; /* Keg statistics. */ keg->uk_free++; } static void zone_release(uma_zone_t zone, void **bucket, int cnt) { void *item; uma_slab_t slab; uma_keg_t keg; uint8_t *mem; int clearfull; int i; clearfull = 0; keg = zone_first_keg(zone); KEG_LOCK(keg); for (i = 0; i < cnt; i++) { item = bucket[i]; if (!(zone->uz_flags & UMA_ZONE_VTOSLAB)) { mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); if (zone->uz_flags & UMA_ZONE_HASH) { slab = hash_sfind(&keg->uk_hash, mem); } else { mem += keg->uk_pgoff; slab = (uma_slab_t)mem; } } else { slab = vtoslab((vm_offset_t)item); if (slab->us_keg != keg) { KEG_UNLOCK(keg); keg = slab->us_keg; KEG_LOCK(keg); } } slab_free_item(keg, slab, item); if (keg->uk_flags & UMA_ZFLAG_FULL) { if (keg->uk_pages < keg->uk_maxpages) { keg->uk_flags &= ~UMA_ZFLAG_FULL; clearfull = 1; } /* * We can handle one more allocation. Since we're * clearing ZFLAG_FULL, wake up all procs blocked * on pages. This should be uncommon, so keeping this * simple for now (rather than adding count of blocked * threads etc). */ wakeup(keg); } } KEG_UNLOCK(keg); if (clearfull) { ZONE_LOCK(zone); zone->uz_flags &= ~UMA_ZFLAG_FULL; wakeup(zone); ZONE_UNLOCK(zone); } } /* * Frees a single item to any zone. * * Arguments: * zone The zone to free to * item The item we're freeing * udata User supplied data for the dtor * skip Skip dtors and finis */ static void zone_free_item(uma_zone_t zone, void *item, void *udata, enum zfreeskip skip) { #ifdef INVARIANTS if (skip == SKIP_NONE) { if (zone->uz_flags & UMA_ZONE_MALLOC) uma_dbg_free(zone, udata, item); else uma_dbg_free(zone, NULL, item); } #endif if (skip < SKIP_DTOR && zone->uz_dtor) zone->uz_dtor(item, zone->uz_size, udata); if (skip < SKIP_FINI && zone->uz_fini) zone->uz_fini(item, zone->uz_size); atomic_add_long(&zone->uz_frees, 1); zone->uz_release(zone->uz_arg, &item, 1); } /* See uma.h */ int uma_zone_set_max(uma_zone_t zone, int nitems) { uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return (0); KEG_LOCK(keg); keg->uk_maxpages = (nitems / keg->uk_ipers) * keg->uk_ppera; if (keg->uk_maxpages * keg->uk_ipers < nitems) keg->uk_maxpages += keg->uk_ppera; nitems = keg->uk_maxpages * keg->uk_ipers; KEG_UNLOCK(keg); return (nitems); } /* See uma.h */ int uma_zone_get_max(uma_zone_t zone) { int nitems; uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return (0); KEG_LOCK(keg); nitems = keg->uk_maxpages * keg->uk_ipers; KEG_UNLOCK(keg); return (nitems); } /* See uma.h */ void uma_zone_set_warning(uma_zone_t zone, const char *warning) { ZONE_LOCK(zone); zone->uz_warning = warning; ZONE_UNLOCK(zone); } /* See uma.h */ void uma_zone_set_maxaction(uma_zone_t zone, uma_maxaction_t maxaction) { ZONE_LOCK(zone); TASK_INIT(&zone->uz_maxaction, 0, (task_fn_t *)maxaction, zone); ZONE_UNLOCK(zone); } /* See uma.h */ int uma_zone_get_cur(uma_zone_t zone) { int64_t nitems; u_int i; ZONE_LOCK(zone); nitems = zone->uz_allocs - zone->uz_frees; CPU_FOREACH(i) { /* * See the comment in sysctl_vm_zone_stats() regarding the * safety of accessing the per-cpu caches. With the zone lock * held, it is safe, but can potentially result in stale data. */ nitems += zone->uz_cpu[i].uc_allocs - zone->uz_cpu[i].uc_frees; } ZONE_UNLOCK(zone); return (nitems < 0 ? 0 : nitems); } /* See uma.h */ void uma_zone_set_init(uma_zone_t zone, uma_init uminit) { uma_keg_t keg; keg = zone_first_keg(zone); KASSERT(keg != NULL, ("uma_zone_set_init: Invalid zone type")); KEG_LOCK(keg); KASSERT(keg->uk_pages == 0, ("uma_zone_set_init on non-empty keg")); keg->uk_init = uminit; KEG_UNLOCK(keg); } /* See uma.h */ void uma_zone_set_fini(uma_zone_t zone, uma_fini fini) { uma_keg_t keg; keg = zone_first_keg(zone); KASSERT(keg != NULL, ("uma_zone_set_fini: Invalid zone type")); KEG_LOCK(keg); KASSERT(keg->uk_pages == 0, ("uma_zone_set_fini on non-empty keg")); keg->uk_fini = fini; KEG_UNLOCK(keg); } /* See uma.h */ void uma_zone_set_zinit(uma_zone_t zone, uma_init zinit) { ZONE_LOCK(zone); KASSERT(zone_first_keg(zone)->uk_pages == 0, ("uma_zone_set_zinit on non-empty keg")); zone->uz_init = zinit; ZONE_UNLOCK(zone); } /* See uma.h */ void uma_zone_set_zfini(uma_zone_t zone, uma_fini zfini) { ZONE_LOCK(zone); KASSERT(zone_first_keg(zone)->uk_pages == 0, ("uma_zone_set_zfini on non-empty keg")); zone->uz_fini = zfini; ZONE_UNLOCK(zone); } /* See uma.h */ /* XXX uk_freef is not actually used with the zone locked */ void uma_zone_set_freef(uma_zone_t zone, uma_free freef) { uma_keg_t keg; keg = zone_first_keg(zone); KASSERT(keg != NULL, ("uma_zone_set_freef: Invalid zone type")); KEG_LOCK(keg); keg->uk_freef = freef; KEG_UNLOCK(keg); } /* See uma.h */ /* XXX uk_allocf is not actually used with the zone locked */ void uma_zone_set_allocf(uma_zone_t zone, uma_alloc allocf) { uma_keg_t keg; keg = zone_first_keg(zone); KEG_LOCK(keg); keg->uk_allocf = allocf; KEG_UNLOCK(keg); } /* See uma.h */ void uma_zone_reserve(uma_zone_t zone, int items) { uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return; KEG_LOCK(keg); keg->uk_reserve = items; KEG_UNLOCK(keg); return; } /* See uma.h */ int uma_zone_reserve_kva(uma_zone_t zone, int count) { uma_keg_t keg; vm_offset_t kva; u_int pages; keg = zone_first_keg(zone); if (keg == NULL) return (0); pages = count / keg->uk_ipers; if (pages * keg->uk_ipers < count) pages++; #ifdef UMA_MD_SMALL_ALLOC if (keg->uk_ppera > 1) { #else if (1) { #endif kva = kva_alloc((vm_size_t)pages * UMA_SLAB_SIZE); if (kva == 0) return (0); } else kva = 0; KEG_LOCK(keg); keg->uk_kva = kva; keg->uk_offset = 0; keg->uk_maxpages = pages; #ifdef UMA_MD_SMALL_ALLOC keg->uk_allocf = (keg->uk_ppera > 1) ? noobj_alloc : uma_small_alloc; #else keg->uk_allocf = noobj_alloc; #endif keg->uk_flags |= UMA_ZONE_NOFREE; KEG_UNLOCK(keg); return (1); } /* See uma.h */ void uma_prealloc(uma_zone_t zone, int items) { int slabs; uma_slab_t slab; uma_keg_t keg; keg = zone_first_keg(zone); if (keg == NULL) return; KEG_LOCK(keg); slabs = items / keg->uk_ipers; if (slabs * keg->uk_ipers < items) slabs++; while (slabs > 0) { slab = keg_alloc_slab(keg, zone, M_WAITOK); if (slab == NULL) break; MPASS(slab->us_keg == keg); LIST_INSERT_HEAD(&keg->uk_free_slab, slab, us_link); slabs--; } KEG_UNLOCK(keg); } /* See uma.h */ static void uma_reclaim_locked(bool kmem_danger) { #ifdef UMA_DEBUG printf("UMA: vm asked us to release pages!\n"); #endif sx_assert(&uma_drain_lock, SA_XLOCKED); bucket_enable(); zone_foreach(zone_drain); if (vm_page_count_min() || kmem_danger) { cache_drain_safe(NULL); zone_foreach(zone_drain); } /* * Some slabs may have been freed but this zone will be visited early * we visit again so that we can free pages that are empty once other * zones are drained. We have to do the same for buckets. */ zone_drain(slabzone); bucket_zone_drain(); } void uma_reclaim(void) { sx_xlock(&uma_drain_lock); uma_reclaim_locked(false); sx_xunlock(&uma_drain_lock); } static int uma_reclaim_needed; void uma_reclaim_wakeup(void) { uma_reclaim_needed = 1; wakeup(&uma_reclaim_needed); } void uma_reclaim_worker(void *arg __unused) { sx_xlock(&uma_drain_lock); for (;;) { sx_sleep(&uma_reclaim_needed, &uma_drain_lock, PVM, "umarcl", 0); if (uma_reclaim_needed) { uma_reclaim_needed = 0; uma_reclaim_locked(true); } } } /* See uma.h */ int uma_zone_exhausted(uma_zone_t zone) { int full; ZONE_LOCK(zone); full = (zone->uz_flags & UMA_ZFLAG_FULL); ZONE_UNLOCK(zone); return (full); } int uma_zone_exhausted_nolock(uma_zone_t zone) { return (zone->uz_flags & UMA_ZFLAG_FULL); } void * uma_large_malloc(vm_size_t size, int wait) { void *mem; uma_slab_t slab; uint8_t flags; slab = zone_alloc_item(slabzone, NULL, wait); if (slab == NULL) return (NULL); mem = page_alloc(NULL, size, &flags, wait); if (mem) { vsetslab((vm_offset_t)mem, slab); slab->us_data = mem; slab->us_flags = flags | UMA_SLAB_MALLOC; slab->us_size = size; } else { zone_free_item(slabzone, slab, NULL, SKIP_NONE); } return (mem); } void uma_large_free(uma_slab_t slab) { page_free(slab->us_data, slab->us_size, slab->us_flags); zone_free_item(slabzone, slab, NULL, SKIP_NONE); } static void uma_zero_item(void *item, uma_zone_t zone) { + int i; if (zone->uz_flags & UMA_ZONE_PCPU) { - for (int i = 0; i < mp_ncpus; i++) + CPU_FOREACH(i) bzero(zpcpu_get_cpu(item, i), zone->uz_size); } else bzero(item, zone->uz_size); } void uma_print_stats(void) { zone_foreach(uma_print_zone); } static void slab_print(uma_slab_t slab) { printf("slab: keg %p, data %p, freecount %d\n", slab->us_keg, slab->us_data, slab->us_freecount); } static void cache_print(uma_cache_t cache) { printf("alloc: %p(%d), free: %p(%d)\n", cache->uc_allocbucket, cache->uc_allocbucket?cache->uc_allocbucket->ub_cnt:0, cache->uc_freebucket, cache->uc_freebucket?cache->uc_freebucket->ub_cnt:0); } static void uma_print_keg(uma_keg_t keg) { uma_slab_t slab; printf("keg: %s(%p) size %d(%d) flags %#x ipers %d ppera %d " "out %d free %d limit %d\n", keg->uk_name, keg, keg->uk_size, keg->uk_rsize, keg->uk_flags, keg->uk_ipers, keg->uk_ppera, (keg->uk_ipers * keg->uk_pages) - keg->uk_free, keg->uk_free, (keg->uk_maxpages / keg->uk_ppera) * keg->uk_ipers); printf("Part slabs:\n"); LIST_FOREACH(slab, &keg->uk_part_slab, us_link) slab_print(slab); printf("Free slabs:\n"); LIST_FOREACH(slab, &keg->uk_free_slab, us_link) slab_print(slab); printf("Full slabs:\n"); LIST_FOREACH(slab, &keg->uk_full_slab, us_link) slab_print(slab); } void uma_print_zone(uma_zone_t zone) { uma_cache_t cache; uma_klink_t kl; int i; printf("zone: %s(%p) size %d flags %#x\n", zone->uz_name, zone, zone->uz_size, zone->uz_flags); LIST_FOREACH(kl, &zone->uz_kegs, kl_link) uma_print_keg(kl->kl_keg); CPU_FOREACH(i) { cache = &zone->uz_cpu[i]; printf("CPU %d Cache:\n", i); cache_print(cache); } } #ifdef DDB /* * Generate statistics across both the zone and its per-cpu cache's. Return * desired statistics if the pointer is non-NULL for that statistic. * * Note: does not update the zone statistics, as it can't safely clear the * per-CPU cache statistic. * * XXXRW: Following the uc_allocbucket and uc_freebucket pointers here isn't * safe from off-CPU; we should modify the caches to track this information * directly so that we don't have to. */ static void uma_zone_sumstat(uma_zone_t z, int *cachefreep, uint64_t *allocsp, uint64_t *freesp, uint64_t *sleepsp) { uma_cache_t cache; uint64_t allocs, frees, sleeps; int cachefree, cpu; allocs = frees = sleeps = 0; cachefree = 0; CPU_FOREACH(cpu) { cache = &z->uz_cpu[cpu]; if (cache->uc_allocbucket != NULL) cachefree += cache->uc_allocbucket->ub_cnt; if (cache->uc_freebucket != NULL) cachefree += cache->uc_freebucket->ub_cnt; allocs += cache->uc_allocs; frees += cache->uc_frees; } allocs += z->uz_allocs; frees += z->uz_frees; sleeps += z->uz_sleeps; if (cachefreep != NULL) *cachefreep = cachefree; if (allocsp != NULL) *allocsp = allocs; if (freesp != NULL) *freesp = frees; if (sleepsp != NULL) *sleepsp = sleeps; } #endif /* DDB */ static int sysctl_vm_zone_count(SYSCTL_HANDLER_ARGS) { uma_keg_t kz; uma_zone_t z; int count; count = 0; rw_rlock(&uma_rwlock); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) count++; } rw_runlock(&uma_rwlock); return (sysctl_handle_int(oidp, &count, 0, req)); } static int sysctl_vm_zone_stats(SYSCTL_HANDLER_ARGS) { struct uma_stream_header ush; struct uma_type_header uth; struct uma_percpu_stat ups; uma_bucket_t bucket; struct sbuf sbuf; uma_cache_t cache; uma_klink_t kl; uma_keg_t kz; uma_zone_t z; uma_keg_t k; int count, error, i; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sbuf_new_for_sysctl(&sbuf, NULL, 128, req); sbuf_clear_flags(&sbuf, SBUF_INCLUDENUL); count = 0; rw_rlock(&uma_rwlock); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) count++; } /* * Insert stream header. */ bzero(&ush, sizeof(ush)); ush.ush_version = UMA_STREAM_VERSION; ush.ush_maxcpus = (mp_maxid + 1); ush.ush_count = count; (void)sbuf_bcat(&sbuf, &ush, sizeof(ush)); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) { bzero(&uth, sizeof(uth)); ZONE_LOCK(z); strlcpy(uth.uth_name, z->uz_name, UTH_MAX_NAME); uth.uth_align = kz->uk_align; uth.uth_size = kz->uk_size; uth.uth_rsize = kz->uk_rsize; LIST_FOREACH(kl, &z->uz_kegs, kl_link) { k = kl->kl_keg; uth.uth_maxpages += k->uk_maxpages; uth.uth_pages += k->uk_pages; uth.uth_keg_free += k->uk_free; uth.uth_limit = (k->uk_maxpages / k->uk_ppera) * k->uk_ipers; } /* * A zone is secondary is it is not the first entry * on the keg's zone list. */ if ((z->uz_flags & UMA_ZONE_SECONDARY) && (LIST_FIRST(&kz->uk_zones) != z)) uth.uth_zone_flags = UTH_ZONE_SECONDARY; LIST_FOREACH(bucket, &z->uz_buckets, ub_link) uth.uth_zone_free += bucket->ub_cnt; uth.uth_allocs = z->uz_allocs; uth.uth_frees = z->uz_frees; uth.uth_fails = z->uz_fails; uth.uth_sleeps = z->uz_sleeps; (void)sbuf_bcat(&sbuf, &uth, sizeof(uth)); /* * While it is not normally safe to access the cache * bucket pointers while not on the CPU that owns the * cache, we only allow the pointers to be exchanged * without the zone lock held, not invalidated, so * accept the possible race associated with bucket * exchange during monitoring. */ for (i = 0; i < (mp_maxid + 1); i++) { bzero(&ups, sizeof(ups)); if (kz->uk_flags & UMA_ZFLAG_INTERNAL) goto skip; if (CPU_ABSENT(i)) goto skip; cache = &z->uz_cpu[i]; if (cache->uc_allocbucket != NULL) ups.ups_cache_free += cache->uc_allocbucket->ub_cnt; if (cache->uc_freebucket != NULL) ups.ups_cache_free += cache->uc_freebucket->ub_cnt; ups.ups_allocs = cache->uc_allocs; ups.ups_frees = cache->uc_frees; skip: (void)sbuf_bcat(&sbuf, &ups, sizeof(ups)); } ZONE_UNLOCK(z); } } rw_runlock(&uma_rwlock); error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } int sysctl_handle_uma_zone_max(SYSCTL_HANDLER_ARGS) { uma_zone_t zone = *(uma_zone_t *)arg1; int error, max; max = uma_zone_get_max(zone); error = sysctl_handle_int(oidp, &max, 0, req); if (error || !req->newptr) return (error); uma_zone_set_max(zone, max); return (0); } int sysctl_handle_uma_zone_cur(SYSCTL_HANDLER_ARGS) { uma_zone_t zone = *(uma_zone_t *)arg1; int cur; cur = uma_zone_get_cur(zone); return (sysctl_handle_int(oidp, &cur, 0, req)); } #ifdef INVARIANTS static uma_slab_t uma_dbg_getslab(uma_zone_t zone, void *item) { uma_slab_t slab; uma_keg_t keg; uint8_t *mem; mem = (uint8_t *)((uintptr_t)item & (~UMA_SLAB_MASK)); if (zone->uz_flags & UMA_ZONE_VTOSLAB) { slab = vtoslab((vm_offset_t)mem); } else { /* * It is safe to return the slab here even though the * zone is unlocked because the item's allocation state * essentially holds a reference. */ ZONE_LOCK(zone); keg = LIST_FIRST(&zone->uz_kegs)->kl_keg; if (keg->uk_flags & UMA_ZONE_HASH) slab = hash_sfind(&keg->uk_hash, mem); else slab = (uma_slab_t)(mem + keg->uk_pgoff); ZONE_UNLOCK(zone); } return (slab); } /* * Set up the slab's freei data such that uma_dbg_free can function. * */ static void uma_dbg_alloc(uma_zone_t zone, uma_slab_t slab, void *item) { uma_keg_t keg; int freei; if (zone_first_keg(zone) == NULL) return; if (slab == NULL) { slab = uma_dbg_getslab(zone, item); if (slab == NULL) panic("uma: item %p did not belong to zone %s\n", item, zone->uz_name); } keg = slab->us_keg; freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; if (BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) panic("Duplicate alloc of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); BIT_SET_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); return; } /* * Verifies freed addresses. Checks for alignment, valid slab membership * and duplicate frees. * */ static void uma_dbg_free(uma_zone_t zone, uma_slab_t slab, void *item) { uma_keg_t keg; int freei; if (zone_first_keg(zone) == NULL) return; if (slab == NULL) { slab = uma_dbg_getslab(zone, item); if (slab == NULL) panic("uma: Freed item %p did not belong to zone %s\n", item, zone->uz_name); } keg = slab->us_keg; freei = ((uintptr_t)item - (uintptr_t)slab->us_data) / keg->uk_rsize; if (freei >= keg->uk_ipers) panic("Invalid free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); if (((freei * keg->uk_rsize) + slab->us_data) != item) panic("Unaligned free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); if (!BIT_ISSET(SLAB_SETSIZE, freei, &slab->us_debugfree)) panic("Duplicate free of %p from zone %p(%s) slab %p(%d)\n", item, zone, zone->uz_name, slab, freei); BIT_CLR_ATOMIC(SLAB_SETSIZE, freei, &slab->us_debugfree); } #endif /* INVARIANTS */ #ifdef DDB DB_SHOW_COMMAND(uma, db_show_uma) { uint64_t allocs, frees, sleeps; uma_bucket_t bucket; uma_keg_t kz; uma_zone_t z; int cachefree; db_printf("%18s %8s %8s %8s %12s %8s %8s\n", "Zone", "Size", "Used", "Free", "Requests", "Sleeps", "Bucket"); LIST_FOREACH(kz, &uma_kegs, uk_link) { LIST_FOREACH(z, &kz->uk_zones, uz_link) { if (kz->uk_flags & UMA_ZFLAG_INTERNAL) { allocs = z->uz_allocs; frees = z->uz_frees; sleeps = z->uz_sleeps; cachefree = 0; } else uma_zone_sumstat(z, &cachefree, &allocs, &frees, &sleeps); if (!((z->uz_flags & UMA_ZONE_SECONDARY) && (LIST_FIRST(&kz->uk_zones) != z))) cachefree += kz->uk_free; LIST_FOREACH(bucket, &z->uz_buckets, ub_link) cachefree += bucket->ub_cnt; db_printf("%18s %8ju %8jd %8d %12ju %8ju %8u\n", z->uz_name, (uintmax_t)kz->uk_size, (intmax_t)(allocs - frees), cachefree, (uintmax_t)allocs, sleeps, z->uz_count); if (db_pager_quit) return; } } } DB_SHOW_COMMAND(umacache, db_show_umacache) { uint64_t allocs, frees; uma_bucket_t bucket; uma_zone_t z; int cachefree; db_printf("%18s %8s %8s %8s %12s %8s\n", "Zone", "Size", "Used", "Free", "Requests", "Bucket"); LIST_FOREACH(z, &uma_cachezones, uz_link) { uma_zone_sumstat(z, &cachefree, &allocs, &frees, NULL); LIST_FOREACH(bucket, &z->uz_buckets, ub_link) cachefree += bucket->ub_cnt; db_printf("%18s %8ju %8jd %8d %12ju %8u\n", z->uz_name, (uintmax_t)z->uz_size, (intmax_t)(allocs - frees), cachefree, (uintmax_t)allocs, z->uz_count); if (db_pager_quit) return; } } #endif /* DDB */