Index: head/sys/kern/kern_dtrace.c =================================================================== --- head/sys/kern/kern_dtrace.c (revision 323235) +++ head/sys/kern/kern_dtrace.c (revision 323236) @@ -1,126 +1,126 @@ /*- * Copyright (c) 2007-2008 John Birrell * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_kdb.h" #include #include #include #include #include #include #include #include #include #include #define KDTRACE_PROC_SIZE 64 #define KDTRACE_THREAD_SIZE 256 FEATURE(kdtrace_hooks, "Kernel DTrace hooks which are required to load DTrace kernel modules"); static MALLOC_DEFINE(M_KDTRACE, "kdtrace", "DTrace hooks"); /* Hooks used in the machine-dependent trap handlers. */ dtrace_trap_func_t dtrace_trap_func; dtrace_doubletrap_func_t dtrace_doubletrap_func; dtrace_pid_probe_ptr_t dtrace_pid_probe_ptr; dtrace_return_probe_ptr_t dtrace_return_probe_ptr; -systrace_probe_func_t systrace_probe_func; +systrace_probe_func_t __read_frequently systrace_probe_func; /* Return the DTrace process data size compiled in the kernel hooks. */ size_t kdtrace_proc_size() { return (KDTRACE_PROC_SIZE); } static void kdtrace_proc_ctor(void *arg __unused, struct proc *p) { p->p_dtrace = malloc(KDTRACE_PROC_SIZE, M_KDTRACE, M_WAITOK|M_ZERO); } static void kdtrace_proc_dtor(void *arg __unused, struct proc *p) { if (p->p_dtrace != NULL) { free(p->p_dtrace, M_KDTRACE); p->p_dtrace = NULL; } } /* Return the DTrace thread data size compiled in the kernel hooks. */ size_t kdtrace_thread_size() { return (KDTRACE_THREAD_SIZE); } static void kdtrace_thread_ctor(void *arg __unused, struct thread *td) { td->td_dtrace = malloc(KDTRACE_THREAD_SIZE, M_KDTRACE, M_WAITOK|M_ZERO); } static void kdtrace_thread_dtor(void *arg __unused, struct thread *td) { if (td->td_dtrace != NULL) { free(td->td_dtrace, M_KDTRACE); td->td_dtrace = NULL; } } /* * Initialise the kernel DTrace hooks. */ static void init_dtrace(void *dummy __unused) { EVENTHANDLER_REGISTER(process_ctor, kdtrace_proc_ctor, NULL, EVENTHANDLER_PRI_ANY); EVENTHANDLER_REGISTER(process_dtor, kdtrace_proc_dtor, NULL, EVENTHANDLER_PRI_ANY); EVENTHANDLER_REGISTER(thread_ctor, kdtrace_thread_ctor, NULL, EVENTHANDLER_PRI_ANY); EVENTHANDLER_REGISTER(thread_dtor, kdtrace_thread_dtor, NULL, EVENTHANDLER_PRI_ANY); } SYSINIT(kdtrace, SI_SUB_KDTRACE, SI_ORDER_FIRST, init_dtrace, NULL); Index: head/sys/kern/kern_lockstat.c =================================================================== --- head/sys/kern/kern_lockstat.c (revision 323235) +++ head/sys/kern/kern_lockstat.c (revision 323236) @@ -1,82 +1,82 @@ /*- * Copyright 2008-2009 Stacey Son * * 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 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 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 SDT_PROVIDER_DEFINE(lockstat); SDT_PROBE_DEFINE1(lockstat, , , adaptive__acquire, "struct mtx *"); SDT_PROBE_DEFINE1(lockstat, , , adaptive__release, "struct mtx *"); SDT_PROBE_DEFINE2(lockstat, , , adaptive__spin, "struct mtx *", "uint64_t"); SDT_PROBE_DEFINE2(lockstat, , , adaptive__block, "struct mtx *", "uint64_t"); SDT_PROBE_DEFINE1(lockstat, , , spin__acquire, "struct mtx *"); SDT_PROBE_DEFINE1(lockstat, , , spin__release, "struct mtx *"); SDT_PROBE_DEFINE2(lockstat, , , spin__spin, "struct mtx *", "uint64_t"); SDT_PROBE_DEFINE2(lockstat, , , rw__acquire, "struct rwlock *", "int"); SDT_PROBE_DEFINE2(lockstat, , , rw__release, "struct rwlock *", "int"); SDT_PROBE_DEFINE5(lockstat, , , rw__block, "struct rwlock *", "uint64_t", "int", "int", "int"); SDT_PROBE_DEFINE2(lockstat, , , rw__spin, "struct rwlock *", "uint64_t"); SDT_PROBE_DEFINE1(lockstat, , , rw__upgrade, "struct rwlock *"); SDT_PROBE_DEFINE1(lockstat, , , rw__downgrade, "struct rwlock *"); SDT_PROBE_DEFINE2(lockstat, , , sx__acquire, "struct sx *", "int"); SDT_PROBE_DEFINE2(lockstat, , , sx__release, "struct sx *", "int"); SDT_PROBE_DEFINE5(lockstat, , , sx__block, "struct sx *", "uint64_t", "int", "int", "int"); SDT_PROBE_DEFINE2(lockstat, , , sx__spin, "struct sx *", "uint64_t"); SDT_PROBE_DEFINE1(lockstat, , , sx__upgrade, "struct sx *"); SDT_PROBE_DEFINE1(lockstat, , , sx__downgrade, "struct sx *"); SDT_PROBE_DEFINE2(lockstat, , , thread__spin, "struct mtx *", "uint64_t"); -volatile int __read_mostly lockstat_enabled; +volatile int __read_frequently lockstat_enabled; uint64_t lockstat_nsecs(struct lock_object *lo) { struct bintime bt; uint64_t ns; if (!lockstat_enabled) return (0); if ((lo->lo_flags & LO_NOPROFILE) != 0) return (0); binuptime(&bt); ns = bt.sec * (uint64_t)1000000000; ns += ((uint64_t)1000000000 * (uint32_t)(bt.frac >> 32)) >> 32; return (ns); } Index: head/sys/kern/kern_mutex.c =================================================================== --- head/sys/kern/kern_mutex.c (revision 323235) +++ head/sys/kern/kern_mutex.c (revision 323236) @@ -1,1183 +1,1183 @@ /*- * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Berkeley Software Design Inc's name may not be used to endorse or * promote products derived from this software without specific prior * written permission. * * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ */ /* * Machine independent bits of mutex implementation. */ #include __FBSDID("$FreeBSD$"); #include "opt_adaptive_mutexes.h" #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(SMP) && !defined(NO_ADAPTIVE_MUTEXES) #define ADAPTIVE_MUTEXES #endif #ifdef HWPMC_HOOKS #include PMC_SOFT_DEFINE( , , lock, failed); #endif /* * Return the mutex address when the lock cookie address is provided. * This functionality assumes that struct mtx* have a member named mtx_lock. */ #define mtxlock2mtx(c) (__containerof(c, struct mtx, mtx_lock)) /* * Internal utility macros. */ #define mtx_unowned(m) ((m)->mtx_lock == MTX_UNOWNED) #define mtx_destroyed(m) ((m)->mtx_lock == MTX_DESTROYED) static void assert_mtx(const struct lock_object *lock, int what); #ifdef DDB static void db_show_mtx(const struct lock_object *lock); #endif static void lock_mtx(struct lock_object *lock, uintptr_t how); static void lock_spin(struct lock_object *lock, uintptr_t how); #ifdef KDTRACE_HOOKS static int owner_mtx(const struct lock_object *lock, struct thread **owner); #endif static uintptr_t unlock_mtx(struct lock_object *lock); static uintptr_t unlock_spin(struct lock_object *lock); /* * Lock classes for sleep and spin mutexes. */ struct lock_class lock_class_mtx_sleep = { .lc_name = "sleep mutex", .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE, .lc_assert = assert_mtx, #ifdef DDB .lc_ddb_show = db_show_mtx, #endif .lc_lock = lock_mtx, .lc_unlock = unlock_mtx, #ifdef KDTRACE_HOOKS .lc_owner = owner_mtx, #endif }; struct lock_class lock_class_mtx_spin = { .lc_name = "spin mutex", .lc_flags = LC_SPINLOCK | LC_RECURSABLE, .lc_assert = assert_mtx, #ifdef DDB .lc_ddb_show = db_show_mtx, #endif .lc_lock = lock_spin, .lc_unlock = unlock_spin, #ifdef KDTRACE_HOOKS .lc_owner = owner_mtx, #endif }; #ifdef ADAPTIVE_MUTEXES static SYSCTL_NODE(_debug, OID_AUTO, mtx, CTLFLAG_RD, NULL, "mtx debugging"); -static struct lock_delay_config __read_mostly mtx_delay; +static struct lock_delay_config __read_frequently mtx_delay; SYSCTL_INT(_debug_mtx, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_delay.base, 0, ""); SYSCTL_INT(_debug_mtx, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_delay.max, 0, ""); LOCK_DELAY_SYSINIT_DEFAULT(mtx_delay); #endif static SYSCTL_NODE(_debug, OID_AUTO, mtx_spin, CTLFLAG_RD, NULL, "mtx spin debugging"); -static struct lock_delay_config __read_mostly mtx_spin_delay; +static struct lock_delay_config __read_frequently mtx_spin_delay; SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_base, CTLFLAG_RW, &mtx_spin_delay.base, 0, ""); SYSCTL_INT(_debug_mtx_spin, OID_AUTO, delay_max, CTLFLAG_RW, &mtx_spin_delay.max, 0, ""); LOCK_DELAY_SYSINIT_DEFAULT(mtx_spin_delay); /* * System-wide mutexes */ struct mtx blocked_lock; struct mtx Giant; void assert_mtx(const struct lock_object *lock, int what) { mtx_assert((const struct mtx *)lock, what); } void lock_mtx(struct lock_object *lock, uintptr_t how) { mtx_lock((struct mtx *)lock); } void lock_spin(struct lock_object *lock, uintptr_t how) { panic("spin locks can only use msleep_spin"); } uintptr_t unlock_mtx(struct lock_object *lock) { struct mtx *m; m = (struct mtx *)lock; mtx_assert(m, MA_OWNED | MA_NOTRECURSED); mtx_unlock(m); return (0); } uintptr_t unlock_spin(struct lock_object *lock) { panic("spin locks can only use msleep_spin"); } #ifdef KDTRACE_HOOKS int owner_mtx(const struct lock_object *lock, struct thread **owner) { const struct mtx *m; uintptr_t x; m = (const struct mtx *)lock; x = m->mtx_lock; *owner = (struct thread *)(x & ~MTX_FLAGMASK); return (x != MTX_UNOWNED); } #endif /* * Function versions of the inlined __mtx_* macros. These are used by * modules and can also be called from assembly language if needed. */ void __mtx_lock_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; uintptr_t tid, v; m = mtxlock2mtx(c); KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() || !TD_IS_IDLETHREAD(curthread), ("mtx_lock() by idle thread %p on sleep mutex %s @ %s:%d", curthread, m->lock_object.lo_name, file, line)); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_lock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, ("mtx_lock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); WITNESS_CHECKORDER(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); tid = (uintptr_t)curthread; v = MTX_UNOWNED; if (!_mtx_obtain_lock_fetch(m, &v, tid)) _mtx_lock_sleep(m, v, tid, opts, file, line); else LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, 0, 0, file, line); LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, line); WITNESS_LOCK(&m->lock_object, (opts & ~MTX_RECURSE) | LOP_EXCLUSIVE, file, line); TD_LOCKS_INC(curthread); } void __mtx_unlock_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_unlock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, ("mtx_unlock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, line); mtx_assert(m, MA_OWNED); #ifdef LOCK_PROFILING __mtx_unlock_sleep(c, opts, file, line); #else __mtx_unlock(m, curthread, opts, file, line); #endif TD_LOCKS_DEC(curthread); } void __mtx_lock_spin_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_lock_spin() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("mtx_lock_spin() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); if (mtx_owned(m)) KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || (opts & MTX_RECURSE) != 0, ("mtx_lock_spin: recursed on non-recursive mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); opts &= ~MTX_RECURSE; WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); __mtx_lock_spin(m, curthread, opts, file, line); LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, line); WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); } int __mtx_trylock_spin_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return (1); m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_trylock_spin() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("mtx_trylock_spin() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); KASSERT((opts & MTX_RECURSE) == 0, ("mtx_trylock_spin: unsupp. opt MTX_RECURSE on mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); if (__mtx_trylock_spin(m, curthread, opts, file, line)) { LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 1, file, line); WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); return (1); } LOCK_LOG_TRY("LOCK", &m->lock_object, opts, 0, file, line); return (0); } void __mtx_unlock_spin_flags(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; if (SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_unlock_spin() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("mtx_unlock_spin() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); WITNESS_UNLOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("UNLOCK", &m->lock_object, opts, m->mtx_recurse, file, line); mtx_assert(m, MA_OWNED); __mtx_unlock_spin(m); } /* * The important part of mtx_trylock{,_flags}() * Tries to acquire lock `m.' If this function is called on a mutex that * is already owned, it will recursively acquire the lock. */ int _mtx_trylock_flags_(volatile uintptr_t *c, int opts, const char *file, int line) { struct mtx *m; struct thread *td; uintptr_t tid, v; #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif int rval; bool recursed; td = curthread; tid = (uintptr_t)td; if (SCHEDULER_STOPPED_TD(td)) return (1); m = mtxlock2mtx(c); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td), ("mtx_trylock() by idle thread %p on sleep mutex %s @ %s:%d", curthread, m->lock_object.lo_name, file, line)); KASSERT(m->mtx_lock != MTX_DESTROYED, ("mtx_trylock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_sleep, ("mtx_trylock() of spin mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); rval = 1; recursed = false; v = MTX_UNOWNED; for (;;) { if (_mtx_obtain_lock_fetch(m, &v, tid)) break; if (v == MTX_UNOWNED) continue; if (v == tid && ((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || (opts & MTX_RECURSE) != 0)) { m->mtx_recurse++; atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); recursed = true; break; } rval = 0; break; } opts &= ~MTX_RECURSE; LOCK_LOG_TRY("LOCK", &m->lock_object, opts, rval, file, line); if (rval) { WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE | LOP_TRYLOCK, file, line); TD_LOCKS_INC(curthread); if (!recursed) LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, waittime, file, line); } return (rval); } /* * __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock. * * We call this if the lock is either contested (i.e. we need to go to * sleep waiting for it), or if we need to recurse on it. */ #if LOCK_DEBUG > 0 void __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, uintptr_t tid, int opts, const char *file, int line) #else void __mtx_lock_sleep(volatile uintptr_t *c, uintptr_t v, uintptr_t tid) #endif { struct mtx *m; struct turnstile *ts; #ifdef ADAPTIVE_MUTEXES volatile struct thread *owner; #endif #ifdef KTR int cont_logged = 0; #endif #ifdef LOCK_PROFILING int contested = 0; uint64_t waittime = 0; #endif #if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS) struct lock_delay_arg lda; #endif #ifdef KDTRACE_HOOKS u_int sleep_cnt = 0; int64_t sleep_time = 0; int64_t all_time = 0; #endif #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) int doing_lockprof; #endif if (SCHEDULER_STOPPED()) return; #if defined(ADAPTIVE_MUTEXES) lock_delay_arg_init(&lda, &mtx_delay); #elif defined(KDTRACE_HOOKS) lock_delay_arg_init(&lda, NULL); #endif m = mtxlock2mtx(c); if (__predict_false(v == MTX_UNOWNED)) v = MTX_READ_VALUE(m); if (__predict_false(lv_mtx_owner(v) == (struct thread *)tid)) { KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 || (opts & MTX_RECURSE) != 0, ("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); #if LOCK_DEBUG > 0 opts &= ~MTX_RECURSE; #endif m->mtx_recurse++; atomic_set_ptr(&m->mtx_lock, MTX_RECURSED); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m); return; } #if LOCK_DEBUG > 0 opts &= ~MTX_RECURSE; #endif #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR4(KTR_LOCK, "_mtx_lock_sleep: %s contested (lock=%p) at %s:%d", m->lock_object.lo_name, (void *)m->mtx_lock, file, line); #ifdef LOCK_PROFILING doing_lockprof = 1; #elif defined(KDTRACE_HOOKS) doing_lockprof = lockstat_enabled; if (__predict_false(doing_lockprof)) all_time -= lockstat_nsecs(&m->lock_object); #endif for (;;) { if (v == MTX_UNOWNED) { if (_mtx_obtain_lock_fetch(m, &v, tid)) break; continue; } #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif #ifdef ADAPTIVE_MUTEXES /* * If the owner is running on another CPU, spin until the * owner stops running or the state of the lock changes. */ owner = lv_mtx_owner(v); if (TD_IS_RUNNING(owner)) { if (LOCK_LOG_TEST(&m->lock_object, 0)) CTR3(KTR_LOCK, "%s: spinning on %p held by %p", __func__, m, owner); KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "spinning", "lockname:\"%s\"", m->lock_object.lo_name); do { lock_delay(&lda); v = MTX_READ_VALUE(m); owner = lv_mtx_owner(v); } while (v != MTX_UNOWNED && TD_IS_RUNNING(owner)); KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "running"); continue; } #endif ts = turnstile_trywait(&m->lock_object); v = MTX_READ_VALUE(m); /* * Check if the lock has been released while spinning for * the turnstile chain lock. */ if (v == MTX_UNOWNED) { turnstile_cancel(ts); continue; } #ifdef ADAPTIVE_MUTEXES /* * The current lock owner might have started executing * on another CPU (or the lock could have changed * owners) while we were waiting on the turnstile * chain lock. If so, drop the turnstile lock and try * again. */ owner = lv_mtx_owner(v); if (TD_IS_RUNNING(owner)) { turnstile_cancel(ts); continue; } #endif /* * If the mutex isn't already contested and a failure occurs * setting the contested bit, the mutex was either released * or the state of the MTX_RECURSED bit changed. */ if ((v & MTX_CONTESTED) == 0 && !atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) { turnstile_cancel(ts); v = MTX_READ_VALUE(m); continue; } /* * We definitely must sleep for this lock. */ mtx_assert(m, MA_NOTOWNED); #ifdef KTR if (!cont_logged) { CTR6(KTR_CONTENTION, "contention: %p at %s:%d wants %s, taken by %s:%d", (void *)tid, file, line, m->lock_object.lo_name, WITNESS_FILE(&m->lock_object), WITNESS_LINE(&m->lock_object)); cont_logged = 1; } #endif /* * Block on the turnstile. */ #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&m->lock_object); #endif turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&m->lock_object); sleep_cnt++; #endif v = MTX_READ_VALUE(m); } #ifdef KTR if (cont_logged) { CTR4(KTR_CONTENTION, "contention end: %s acquired by %p at %s:%d", m->lock_object.lo_name, (void *)tid, file, line); } #endif #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) if (__predict_true(!doing_lockprof)) return; #endif #ifdef KDTRACE_HOOKS all_time += lockstat_nsecs(&m->lock_object); #endif LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested, waittime, file, line); #ifdef KDTRACE_HOOKS if (sleep_time) LOCKSTAT_RECORD1(adaptive__block, m, sleep_time); /* * Only record the loops spinning and not sleeping. */ if (lda.spin_cnt > sleep_cnt) LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time); #endif } static void _mtx_lock_spin_failed(struct mtx *m) { struct thread *td; td = mtx_owner(m); /* If the mutex is unlocked, try again. */ if (td == NULL) return; printf( "spin lock %p (%s) held by %p (tid %d) too long\n", m, m->lock_object.lo_name, td, td->td_tid); #ifdef WITNESS witness_display_spinlock(&m->lock_object, td, printf); #endif panic("spin lock held too long"); } #ifdef SMP /* * _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock. * * This is only called if we need to actually spin for the lock. Recursion * is handled inline. */ void _mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v, uintptr_t tid, int opts, const char *file, int line) { struct mtx *m; struct lock_delay_arg lda; #ifdef LOCK_PROFILING int contested = 0; uint64_t waittime = 0; #endif #ifdef KDTRACE_HOOKS int64_t spin_time = 0; #endif #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) int doing_lockprof; #endif if (SCHEDULER_STOPPED()) return; lock_delay_arg_init(&lda, &mtx_spin_delay); m = mtxlock2mtx(c); if (__predict_false(v == MTX_UNOWNED)) v = MTX_READ_VALUE(m); if (__predict_false(v == tid)) { m->mtx_recurse++; return; } if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m); KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "spinning", "lockname:\"%s\"", m->lock_object.lo_name); #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); #ifdef LOCK_PROFILING doing_lockprof = 1; #elif defined(KDTRACE_HOOKS) doing_lockprof = lockstat_enabled; if (__predict_false(doing_lockprof)) spin_time -= lockstat_nsecs(&m->lock_object); #endif for (;;) { if (v == MTX_UNOWNED) { if (_mtx_obtain_lock_fetch(m, &v, tid)) break; continue; } /* Give interrupts a chance while we spin. */ spinlock_exit(); do { if (lda.spin_cnt < 10000000) { lock_delay(&lda); } else { lda.spin_cnt++; if (lda.spin_cnt < 60000000 || kdb_active || panicstr != NULL) DELAY(1); else _mtx_lock_spin_failed(m); cpu_spinwait(); } v = MTX_READ_VALUE(m); } while (v != MTX_UNOWNED); spinlock_enter(); } if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m); KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid), "running"); #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) if (__predict_true(!doing_lockprof)) return; #endif #ifdef KDTRACE_HOOKS spin_time += lockstat_nsecs(&m->lock_object); #endif LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, contested, waittime, file, line); #ifdef KDTRACE_HOOKS if (lda.spin_cnt != 0) LOCKSTAT_RECORD1(spin__spin, m, spin_time); #endif } #endif /* SMP */ void thread_lock_flags_(struct thread *td, int opts, const char *file, int line) { struct mtx *m; uintptr_t tid, v; struct lock_delay_arg lda; #ifdef LOCK_PROFILING int contested = 0; uint64_t waittime = 0; #endif #ifdef KDTRACE_HOOKS int64_t spin_time = 0; #endif #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) int doing_lockprof = 1; #endif tid = (uintptr_t)curthread; if (SCHEDULER_STOPPED()) { /* * Ensure that spinlock sections are balanced even when the * scheduler is stopped, since we may otherwise inadvertently * re-enable interrupts while dumping core. */ spinlock_enter(); return; } lock_delay_arg_init(&lda, &mtx_spin_delay); #ifdef LOCK_PROFILING doing_lockprof = 1; #elif defined(KDTRACE_HOOKS) doing_lockprof = lockstat_enabled; if (__predict_false(doing_lockprof)) spin_time -= lockstat_nsecs(&td->td_lock->lock_object); #endif for (;;) { retry: v = MTX_UNOWNED; spinlock_enter(); m = td->td_lock; KASSERT(m->mtx_lock != MTX_DESTROYED, ("thread_lock() of destroyed mutex @ %s:%d", file, line)); KASSERT(LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin, ("thread_lock() of sleep mutex %s @ %s:%d", m->lock_object.lo_name, file, line)); if (mtx_owned(m)) KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0, ("thread_lock: recursed on non-recursive mutex %s @ %s:%d\n", m->lock_object.lo_name, file, line)); WITNESS_CHECKORDER(&m->lock_object, opts | LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); for (;;) { if (_mtx_obtain_lock_fetch(m, &v, tid)) break; if (v == MTX_UNOWNED) continue; if (v == tid) { m->mtx_recurse++; break; } #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime); /* Give interrupts a chance while we spin. */ spinlock_exit(); do { if (lda.spin_cnt < 10000000) { lock_delay(&lda); } else { lda.spin_cnt++; if (lda.spin_cnt < 60000000 || kdb_active || panicstr != NULL) DELAY(1); else _mtx_lock_spin_failed(m); cpu_spinwait(); } if (m != td->td_lock) goto retry; v = MTX_READ_VALUE(m); } while (v != MTX_UNOWNED); spinlock_enter(); } if (m == td->td_lock) break; __mtx_unlock_spin(m); /* does spinlock_exit() */ } LOCK_LOG_LOCK("LOCK", &m->lock_object, opts, m->mtx_recurse, file, line); WITNESS_LOCK(&m->lock_object, opts | LOP_EXCLUSIVE, file, line); #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) if (__predict_true(!doing_lockprof)) return; #endif #ifdef KDTRACE_HOOKS spin_time += lockstat_nsecs(&m->lock_object); #endif if (m->mtx_recurse == 0) LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m, contested, waittime, file, line); #ifdef KDTRACE_HOOKS if (lda.spin_cnt != 0) LOCKSTAT_RECORD1(thread__spin, m, spin_time); #endif } struct mtx * thread_lock_block(struct thread *td) { struct mtx *lock; THREAD_LOCK_ASSERT(td, MA_OWNED); lock = td->td_lock; td->td_lock = &blocked_lock; mtx_unlock_spin(lock); return (lock); } void thread_lock_unblock(struct thread *td, struct mtx *new) { mtx_assert(new, MA_OWNED); MPASS(td->td_lock == &blocked_lock); atomic_store_rel_ptr((volatile void *)&td->td_lock, (uintptr_t)new); } void thread_lock_set(struct thread *td, struct mtx *new) { struct mtx *lock; mtx_assert(new, MA_OWNED); THREAD_LOCK_ASSERT(td, MA_OWNED); lock = td->td_lock; td->td_lock = new; mtx_unlock_spin(lock); } /* * __mtx_unlock_sleep: the tougher part of releasing an MTX_DEF lock. * * We are only called here if the lock is recursed, contested (i.e. we * need to wake up a blocked thread) or lockstat probe is active. */ #if LOCK_DEBUG > 0 void __mtx_unlock_sleep(volatile uintptr_t *c, int opts, const char *file, int line) #else void __mtx_unlock_sleep(volatile uintptr_t *c) #endif { struct mtx *m; struct turnstile *ts; uintptr_t tid, v; if (SCHEDULER_STOPPED()) return; tid = (uintptr_t)curthread; m = mtxlock2mtx(c); v = MTX_READ_VALUE(m); if (v & MTX_RECURSED) { if (--(m->mtx_recurse) == 0) atomic_clear_ptr(&m->mtx_lock, MTX_RECURSED); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p unrecurse", m); return; } LOCKSTAT_PROFILE_RELEASE_LOCK(adaptive__release, m); if (v == tid && _mtx_release_lock(m, tid)) return; /* * We have to lock the chain before the turnstile so this turnstile * can be removed from the hash list if it is empty. */ turnstile_chain_lock(&m->lock_object); ts = turnstile_lookup(&m->lock_object); if (LOCK_LOG_TEST(&m->lock_object, opts)) CTR1(KTR_LOCK, "_mtx_unlock_sleep: %p contested", m); MPASS(ts != NULL); turnstile_broadcast(ts, TS_EXCLUSIVE_QUEUE); _mtx_release_lock_quick(m); /* * This turnstile is now no longer associated with the mutex. We can * unlock the chain lock so a new turnstile may take it's place. */ turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); turnstile_chain_unlock(&m->lock_object); } /* * All the unlocking of MTX_SPIN locks is done inline. * See the __mtx_unlock_spin() macro for the details. */ /* * The backing function for the INVARIANTS-enabled mtx_assert() */ #ifdef INVARIANT_SUPPORT void __mtx_assert(const volatile uintptr_t *c, int what, const char *file, int line) { const struct mtx *m; if (panicstr != NULL || dumping || SCHEDULER_STOPPED()) return; m = mtxlock2mtx(c); switch (what) { case MA_OWNED: case MA_OWNED | MA_RECURSED: case MA_OWNED | MA_NOTRECURSED: if (!mtx_owned(m)) panic("mutex %s not owned at %s:%d", m->lock_object.lo_name, file, line); if (mtx_recursed(m)) { if ((what & MA_NOTRECURSED) != 0) panic("mutex %s recursed at %s:%d", m->lock_object.lo_name, file, line); } else if ((what & MA_RECURSED) != 0) { panic("mutex %s unrecursed at %s:%d", m->lock_object.lo_name, file, line); } break; case MA_NOTOWNED: if (mtx_owned(m)) panic("mutex %s owned at %s:%d", m->lock_object.lo_name, file, line); break; default: panic("unknown mtx_assert at %s:%d", file, line); } } #endif /* * General init routine used by the MTX_SYSINIT() macro. */ void mtx_sysinit(void *arg) { struct mtx_args *margs = arg; mtx_init((struct mtx *)margs->ma_mtx, margs->ma_desc, NULL, margs->ma_opts); } /* * Mutex initialization routine; initialize lock `m' of type contained in * `opts' with options contained in `opts' and name `name.' The optional * lock type `type' is used as a general lock category name for use with * witness. */ void _mtx_init(volatile uintptr_t *c, const char *name, const char *type, int opts) { struct mtx *m; struct lock_class *class; int flags; m = mtxlock2mtx(c); MPASS((opts & ~(MTX_SPIN | MTX_QUIET | MTX_RECURSE | MTX_NOWITNESS | MTX_DUPOK | MTX_NOPROFILE | MTX_NEW)) == 0); ASSERT_ATOMIC_LOAD_PTR(m->mtx_lock, ("%s: mtx_lock not aligned for %s: %p", __func__, name, &m->mtx_lock)); /* Determine lock class and lock flags. */ if (opts & MTX_SPIN) class = &lock_class_mtx_spin; else class = &lock_class_mtx_sleep; flags = 0; if (opts & MTX_QUIET) flags |= LO_QUIET; if (opts & MTX_RECURSE) flags |= LO_RECURSABLE; if ((opts & MTX_NOWITNESS) == 0) flags |= LO_WITNESS; if (opts & MTX_DUPOK) flags |= LO_DUPOK; if (opts & MTX_NOPROFILE) flags |= LO_NOPROFILE; if (opts & MTX_NEW) flags |= LO_NEW; /* Initialize mutex. */ lock_init(&m->lock_object, class, name, type, flags); m->mtx_lock = MTX_UNOWNED; m->mtx_recurse = 0; } /* * Remove lock `m' from all_mtx queue. We don't allow MTX_QUIET to be * passed in as a flag here because if the corresponding mtx_init() was * called with MTX_QUIET set, then it will already be set in the mutex's * flags. */ void _mtx_destroy(volatile uintptr_t *c) { struct mtx *m; m = mtxlock2mtx(c); if (!mtx_owned(m)) MPASS(mtx_unowned(m)); else { MPASS((m->mtx_lock & (MTX_RECURSED|MTX_CONTESTED)) == 0); /* Perform the non-mtx related part of mtx_unlock_spin(). */ if (LOCK_CLASS(&m->lock_object) == &lock_class_mtx_spin) spinlock_exit(); else TD_LOCKS_DEC(curthread); lock_profile_release_lock(&m->lock_object); /* Tell witness this isn't locked to make it happy. */ WITNESS_UNLOCK(&m->lock_object, LOP_EXCLUSIVE, __FILE__, __LINE__); } m->mtx_lock = MTX_DESTROYED; lock_destroy(&m->lock_object); } /* * Intialize the mutex code and system mutexes. This is called from the MD * startup code prior to mi_startup(). The per-CPU data space needs to be * setup before this is called. */ void mutex_init(void) { /* Setup turnstiles so that sleep mutexes work. */ init_turnstiles(); /* * Initialize mutexes. */ mtx_init(&Giant, "Giant", NULL, MTX_DEF | MTX_RECURSE); mtx_init(&blocked_lock, "blocked lock", NULL, MTX_SPIN); blocked_lock.mtx_lock = 0xdeadc0de; /* Always blocked. */ mtx_init(&proc0.p_mtx, "process lock", NULL, MTX_DEF | MTX_DUPOK); mtx_init(&proc0.p_slock, "process slock", NULL, MTX_SPIN); mtx_init(&proc0.p_statmtx, "pstatl", NULL, MTX_SPIN); mtx_init(&proc0.p_itimmtx, "pitiml", NULL, MTX_SPIN); mtx_init(&proc0.p_profmtx, "pprofl", NULL, MTX_SPIN); mtx_init(&devmtx, "cdev", NULL, MTX_DEF); mtx_lock(&Giant); } #ifdef DDB void db_show_mtx(const struct lock_object *lock) { struct thread *td; const struct mtx *m; m = (const struct mtx *)lock; db_printf(" flags: {"); if (LOCK_CLASS(lock) == &lock_class_mtx_spin) db_printf("SPIN"); else db_printf("DEF"); if (m->lock_object.lo_flags & LO_RECURSABLE) db_printf(", RECURSE"); if (m->lock_object.lo_flags & LO_DUPOK) db_printf(", DUPOK"); db_printf("}\n"); db_printf(" state: {"); if (mtx_unowned(m)) db_printf("UNOWNED"); else if (mtx_destroyed(m)) db_printf("DESTROYED"); else { db_printf("OWNED"); if (m->mtx_lock & MTX_CONTESTED) db_printf(", CONTESTED"); if (m->mtx_lock & MTX_RECURSED) db_printf(", RECURSED"); } db_printf("}\n"); if (!mtx_unowned(m) && !mtx_destroyed(m)) { td = mtx_owner(m); db_printf(" owner: %p (tid %d, pid %d, \"%s\")\n", td, td->td_tid, td->td_proc->p_pid, td->td_name); if (mtx_recursed(m)) db_printf(" recursed: %d\n", m->mtx_recurse); } } #endif Index: head/sys/kern/kern_rwlock.c =================================================================== --- head/sys/kern/kern_rwlock.c (revision 323235) +++ head/sys/kern/kern_rwlock.c (revision 323236) @@ -1,1396 +1,1396 @@ /*- * Copyright (c) 2006 John Baldwin * 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. */ /* * Machine independent bits of reader/writer lock implementation. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include "opt_no_adaptive_rwlocks.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) #define ADAPTIVE_RWLOCKS #endif #ifdef HWPMC_HOOKS #include PMC_SOFT_DECLARE( , , lock, failed); #endif /* * Return the rwlock address when the lock cookie address is provided. * This functionality assumes that struct rwlock* have a member named rw_lock. */ #define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock)) #ifdef DDB #include static void db_show_rwlock(const struct lock_object *lock); #endif static void assert_rw(const struct lock_object *lock, int what); static void lock_rw(struct lock_object *lock, uintptr_t how); #ifdef KDTRACE_HOOKS static int owner_rw(const struct lock_object *lock, struct thread **owner); #endif static uintptr_t unlock_rw(struct lock_object *lock); struct lock_class lock_class_rw = { .lc_name = "rw", .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, .lc_assert = assert_rw, #ifdef DDB .lc_ddb_show = db_show_rwlock, #endif .lc_lock = lock_rw, .lc_unlock = unlock_rw, #ifdef KDTRACE_HOOKS .lc_owner = owner_rw, #endif }; #ifdef ADAPTIVE_RWLOCKS -static int rowner_retries = 10; -static int rowner_loops = 10000; +static int __read_frequently rowner_retries = 10; +static int __read_frequently rowner_loops = 10000; static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, "rwlock debugging"); SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); -static struct lock_delay_config __read_mostly rw_delay; +static struct lock_delay_config __read_frequently rw_delay; SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base, 0, ""); SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max, 0, ""); LOCK_DELAY_SYSINIT_DEFAULT(rw_delay); #endif /* * Return a pointer to the owning thread if the lock is write-locked or * NULL if the lock is unlocked or read-locked. */ #define lv_rw_wowner(v) \ ((v) & RW_LOCK_READ ? NULL : \ (struct thread *)RW_OWNER((v))) #define rw_wowner(rw) lv_rw_wowner(RW_READ_VALUE(rw)) /* * Returns if a write owner is recursed. Write ownership is not assured * here and should be previously checked. */ #define rw_recursed(rw) ((rw)->rw_recurse != 0) /* * Return true if curthread helds the lock. */ #define rw_wlocked(rw) (rw_wowner((rw)) == curthread) /* * Return a pointer to the owning thread for this lock who should receive * any priority lent by threads that block on this lock. Currently this * is identical to rw_wowner(). */ #define rw_owner(rw) rw_wowner(rw) #ifndef INVARIANTS #define __rw_assert(c, what, file, line) #endif void assert_rw(const struct lock_object *lock, int what) { rw_assert((const struct rwlock *)lock, what); } void lock_rw(struct lock_object *lock, uintptr_t how) { struct rwlock *rw; rw = (struct rwlock *)lock; if (how) rw_rlock(rw); else rw_wlock(rw); } uintptr_t unlock_rw(struct lock_object *lock) { struct rwlock *rw; rw = (struct rwlock *)lock; rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); if (rw->rw_lock & RW_LOCK_READ) { rw_runlock(rw); return (1); } else { rw_wunlock(rw); return (0); } } #ifdef KDTRACE_HOOKS int owner_rw(const struct lock_object *lock, struct thread **owner) { const struct rwlock *rw = (const struct rwlock *)lock; uintptr_t x = rw->rw_lock; *owner = rw_wowner(rw); return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : (*owner != NULL)); } #endif void _rw_init_flags(volatile uintptr_t *c, const char *name, int opts) { struct rwlock *rw; int flags; rw = rwlock2rw(c); MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | RW_RECURSE | RW_NEW)) == 0); ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, ("%s: rw_lock not aligned for %s: %p", __func__, name, &rw->rw_lock)); flags = LO_UPGRADABLE; if (opts & RW_DUPOK) flags |= LO_DUPOK; if (opts & RW_NOPROFILE) flags |= LO_NOPROFILE; if (!(opts & RW_NOWITNESS)) flags |= LO_WITNESS; if (opts & RW_RECURSE) flags |= LO_RECURSABLE; if (opts & RW_QUIET) flags |= LO_QUIET; if (opts & RW_NEW) flags |= LO_NEW; lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); rw->rw_lock = RW_UNLOCKED; rw->rw_recurse = 0; } void _rw_destroy(volatile uintptr_t *c) { struct rwlock *rw; rw = rwlock2rw(c); KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw)); KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw)); rw->rw_lock = RW_DESTROYED; lock_destroy(&rw->lock_object); } void rw_sysinit(void *arg) { struct rw_args *args = arg; rw_init((struct rwlock *)args->ra_rw, args->ra_desc); } void rw_sysinit_flags(void *arg) { struct rw_args_flags *args = arg; rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc, args->ra_flags); } int _rw_wowned(const volatile uintptr_t *c) { return (rw_wowner(rwlock2rw(c)) == curthread); } void _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; uintptr_t tid, v; rw = rwlock2rw(c); KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() || !TD_IS_IDLETHREAD(curthread), ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d", curthread, rw->lock_object.lo_name, file, line)); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); tid = (uintptr_t)curthread; v = RW_UNLOCKED; if (!_rw_write_lock_fetch(rw, &v, tid)) _rw_wlock_hard(rw, v, tid, file, line); else LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, 0, 0, file, line, LOCKSTAT_WRITER); LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); TD_LOCKS_INC(curthread); } int __rw_try_wlock(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; struct thread *td; uintptr_t tid, v; int rval; bool recursed; td = curthread; tid = (uintptr_t)td; if (SCHEDULER_STOPPED_TD(td)) return (1); rw = rwlock2rw(c); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td), ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d", curthread, rw->lock_object.lo_name, file, line)); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); rval = 1; recursed = false; v = RW_UNLOCKED; for (;;) { if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid)) break; if (v == RW_UNLOCKED) continue; if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) { rw->rw_recurse++; atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED); break; } rval = 0; break; } LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); if (rval) { WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, file, line); if (!recursed) LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, 0, 0, file, line, LOCKSTAT_WRITER); TD_LOCKS_INC(curthread); } return (rval); } void _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; rw = rwlock2rw(c); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); __rw_assert(c, RA_WLOCKED, file, line); WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); #ifdef LOCK_PROFILING _rw_wunlock_hard(rw, (uintptr_t)curthread, file, line); #else __rw_wunlock(rw, curthread, file, line); #endif TD_LOCKS_DEC(curthread); } /* * Determines whether a new reader can acquire a lock. Succeeds if the * reader already owns a read lock and the lock is locked for read to * prevent deadlock from reader recursion. Also succeeds if the lock * is unlocked and has no writer waiters or spinners. Failing otherwise * prioritizes writers before readers. */ #define RW_CAN_READ(td, _rw) \ (((_rw) & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==\ RW_LOCK_READ || ((td)->td_rw_rlocks && (_rw) & RW_LOCK_READ)) static bool __always_inline __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp, const char *file, int line) { /* * Handle the easy case. If no other thread has a write * lock, then try to bump up the count of read locks. Note * that we have to preserve the current state of the * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a * read lock, then rw_lock must have changed, so restart * the loop. Note that this handles the case of a * completely unlocked rwlock since such a lock is encoded * as a read lock with no waiters. */ while (RW_CAN_READ(td, *vp)) { if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp, *vp + RW_ONE_READER)) { if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR4(KTR_LOCK, "%s: %p succeed %p -> %p", __func__, rw, (void *)*vp, (void *)(*vp + RW_ONE_READER)); td->td_rw_rlocks++; return (true); } } return (false); } static void __noinline __rw_rlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v, const char *file, int line) { struct rwlock *rw; struct turnstile *ts; #ifdef ADAPTIVE_RWLOCKS volatile struct thread *owner; int spintries = 0; int i; #endif #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS) struct lock_delay_arg lda; #endif #ifdef KDTRACE_HOOKS uintptr_t state; u_int sleep_cnt = 0; int64_t sleep_time = 0; int64_t all_time = 0; #endif if (SCHEDULER_STOPPED()) return; #if defined(ADAPTIVE_RWLOCKS) lock_delay_arg_init(&lda, &rw_delay); #elif defined(KDTRACE_HOOKS) lock_delay_arg_init(&lda, NULL); #endif rw = rwlock2rw(c); #ifdef KDTRACE_HOOKS all_time -= lockstat_nsecs(&rw->lock_object); #endif #ifdef KDTRACE_HOOKS state = v; #endif for (;;) { if (__rw_rlock_try(rw, td, &v, file, line)) break; #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&rw->lock_object, &contested, &waittime); #ifdef ADAPTIVE_RWLOCKS /* * If the owner is running on another CPU, spin until * the owner stops running or the state of the lock * changes. */ if ((v & RW_LOCK_READ) == 0) { owner = (struct thread *)RW_OWNER(v); if (TD_IS_RUNNING(owner)) { if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR3(KTR_LOCK, "%s: spinning on %p held by %p", __func__, rw, owner); KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", rw->lock_object.lo_name); do { lock_delay(&lda); v = RW_READ_VALUE(rw); owner = lv_rw_wowner(v); } while (owner != NULL && TD_IS_RUNNING(owner)); KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); continue; } } else if (spintries < rowner_retries) { spintries++; KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", rw->lock_object.lo_name); for (i = 0; i < rowner_loops; i++) { v = RW_READ_VALUE(rw); if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(td, v)) break; cpu_spinwait(); } v = RW_READ_VALUE(rw); #ifdef KDTRACE_HOOKS lda.spin_cnt += rowner_loops - i; #endif KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); if (i != rowner_loops) continue; } #endif /* * Okay, now it's the hard case. Some other thread already * has a write lock or there are write waiters present, * acquire the turnstile lock so we can begin the process * of blocking. */ ts = turnstile_trywait(&rw->lock_object); /* * The lock might have been released while we spun, so * recheck its state and restart the loop if needed. */ v = RW_READ_VALUE(rw); if (RW_CAN_READ(td, v)) { turnstile_cancel(ts); continue; } #ifdef ADAPTIVE_RWLOCKS /* * The current lock owner might have started executing * on another CPU (or the lock could have changed * owners) while we were waiting on the turnstile * chain lock. If so, drop the turnstile lock and try * again. */ if ((v & RW_LOCK_READ) == 0) { owner = (struct thread *)RW_OWNER(v); if (TD_IS_RUNNING(owner)) { turnstile_cancel(ts); continue; } } #endif /* * The lock is held in write mode or it already has waiters. */ MPASS(!RW_CAN_READ(td, v)); /* * If the RW_LOCK_READ_WAITERS flag is already set, then * we can go ahead and block. If it is not set then try * to set it. If we fail to set it drop the turnstile * lock and restart the loop. */ if (!(v & RW_LOCK_READ_WAITERS)) { if (!atomic_cmpset_ptr(&rw->rw_lock, v, v | RW_LOCK_READ_WAITERS)) { turnstile_cancel(ts); v = RW_READ_VALUE(rw); continue; } if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p set read waiters flag", __func__, rw); } /* * We were unable to acquire the lock and the read waiters * flag is set, so we must block on the turnstile. */ if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, rw); #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&rw->lock_object); #endif turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&rw->lock_object); sleep_cnt++; #endif if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p resuming from turnstile", __func__, rw); v = RW_READ_VALUE(rw); } #ifdef KDTRACE_HOOKS all_time += lockstat_nsecs(&rw->lock_object); if (sleep_time) LOCKSTAT_RECORD4(rw__block, rw, sleep_time, LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); /* Record only the loops spinning and not sleeping. */ if (lda.spin_cnt > sleep_cnt) LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); #endif /* * TODO: acquire "owner of record" here. Here be turnstile dragons * however. turnstiles don't like owners changing between calls to * turnstile_wait() currently. */ LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, waittime, file, line, LOCKSTAT_READER); } void __rw_rlock(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; struct thread *td; uintptr_t v; td = curthread; rw = rwlock2rw(c); KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) || !TD_IS_IDLETHREAD(td), ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d", td, rw->lock_object.lo_name, file, line)); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); KASSERT(rw_wowner(rw) != td, ("rw_rlock: wlock already held for %s @ %s:%d", rw->lock_object.lo_name, file, line)); WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); v = RW_READ_VALUE(rw); if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__acquire) || !__rw_rlock_try(rw, td, &v, file, line))) __rw_rlock_hard(c, td, v, file, line); LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); WITNESS_LOCK(&rw->lock_object, 0, file, line); TD_LOCKS_INC(curthread); } int __rw_try_rlock(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; uintptr_t x; if (SCHEDULER_STOPPED()) return (1); rw = rwlock2rw(c); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d", curthread, rw->lock_object.lo_name, file, line)); x = rw->rw_lock; for (;;) { KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); if (!(x & RW_LOCK_READ)) break; if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) { LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, line); WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, 0, 0, file, line, LOCKSTAT_READER); TD_LOCKS_INC(curthread); curthread->td_rw_rlocks++; return (1); } } LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); return (0); } static bool __always_inline __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp) { for (;;) { /* * See if there is more than one read lock held. If so, * just drop one and return. */ if (RW_READERS(*vp) > 1) { if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp, *vp - RW_ONE_READER)) { if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR4(KTR_LOCK, "%s: %p succeeded %p -> %p", __func__, rw, (void *)*vp, (void *)(*vp - RW_ONE_READER)); td->td_rw_rlocks--; return (true); } continue; } /* * If there aren't any waiters for a write lock, then try * to drop it quickly. */ if (!(*vp & RW_LOCK_WAITERS)) { MPASS((*vp & ~RW_LOCK_WRITE_SPINNER) == RW_READERS_LOCK(1)); if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp, RW_UNLOCKED)) { if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p last succeeded", __func__, rw); td->td_rw_rlocks--; return (true); } continue; } break; } return (false); } static void __noinline __rw_runlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v, const char *file, int line) { struct rwlock *rw; struct turnstile *ts; uintptr_t x, queue; if (SCHEDULER_STOPPED()) return; rw = rwlock2rw(c); for (;;) { if (__rw_runlock_try(rw, td, &v)) break; /* * Ok, we know we have waiters and we think we are the * last reader, so grab the turnstile lock. */ turnstile_chain_lock(&rw->lock_object); v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); MPASS(v & RW_LOCK_WAITERS); /* * Try to drop our lock leaving the lock in a unlocked * state. * * If you wanted to do explicit lock handoff you'd have to * do it here. You'd also want to use turnstile_signal() * and you'd have to handle the race where a higher * priority thread blocks on the write lock before the * thread you wakeup actually runs and have the new thread * "steal" the lock. For now it's a lot simpler to just * wakeup all of the waiters. * * As above, if we fail, then another thread might have * acquired a read lock, so drop the turnstile lock and * restart. */ x = RW_UNLOCKED; if (v & RW_LOCK_WRITE_WAITERS) { queue = TS_EXCLUSIVE_QUEUE; x |= (v & RW_LOCK_READ_WAITERS); } else queue = TS_SHARED_QUEUE; if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v, x)) { turnstile_chain_unlock(&rw->lock_object); v = RW_READ_VALUE(rw); continue; } if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", __func__, rw); /* * Ok. The lock is released and all that's left is to * wake up the waiters. Note that the lock might not be * free anymore, but in that case the writers will just * block again if they run before the new lock holder(s) * release the lock. */ ts = turnstile_lookup(&rw->lock_object); MPASS(ts != NULL); turnstile_broadcast(ts, queue); turnstile_unpend(ts, TS_SHARED_LOCK); turnstile_chain_unlock(&rw->lock_object); td->td_rw_rlocks--; break; } LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER); } void _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; struct thread *td; uintptr_t v; rw = rwlock2rw(c); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); __rw_assert(c, RA_RLOCKED, file, line); WITNESS_UNLOCK(&rw->lock_object, 0, file, line); LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); td = curthread; v = RW_READ_VALUE(rw); if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__release) || !__rw_runlock_try(rw, td, &v))) __rw_runlock_hard(c, td, v, file, line); TD_LOCKS_DEC(curthread); } /* * This function is called when we are unable to obtain a write lock on the * first try. This means that at least one other thread holds either a * read or write lock. */ void __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v, uintptr_t tid, const char *file, int line) { struct rwlock *rw; struct turnstile *ts; #ifdef ADAPTIVE_RWLOCKS volatile struct thread *owner; int spintries = 0; int i; #endif uintptr_t x; #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS) struct lock_delay_arg lda; #endif #ifdef KDTRACE_HOOKS uintptr_t state; u_int sleep_cnt = 0; int64_t sleep_time = 0; int64_t all_time = 0; #endif if (SCHEDULER_STOPPED()) return; #if defined(ADAPTIVE_RWLOCKS) lock_delay_arg_init(&lda, &rw_delay); #elif defined(KDTRACE_HOOKS) lock_delay_arg_init(&lda, NULL); #endif rw = rwlock2rw(c); if (__predict_false(v == RW_UNLOCKED)) v = RW_READ_VALUE(rw); if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) { KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, ("%s: recursing but non-recursive rw %s @ %s:%d\n", __func__, rw->lock_object.lo_name, file, line)); rw->rw_recurse++; atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED); if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); return; } if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); #ifdef KDTRACE_HOOKS all_time -= lockstat_nsecs(&rw->lock_object); state = v; #endif for (;;) { if (v == RW_UNLOCKED) { if (_rw_write_lock_fetch(rw, &v, tid)) break; continue; } #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&rw->lock_object, &contested, &waittime); #ifdef ADAPTIVE_RWLOCKS /* * If the lock is write locked and the owner is * running on another CPU, spin until the owner stops * running or the state of the lock changes. */ owner = lv_rw_wowner(v); if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) { if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR3(KTR_LOCK, "%s: spinning on %p held by %p", __func__, rw, owner); KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", rw->lock_object.lo_name); do { lock_delay(&lda); v = RW_READ_VALUE(rw); owner = lv_rw_wowner(v); } while (owner != NULL && TD_IS_RUNNING(owner)); KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); continue; } if ((v & RW_LOCK_READ) && RW_READERS(v) && spintries < rowner_retries) { if (!(v & RW_LOCK_WRITE_SPINNER)) { if (!atomic_cmpset_ptr(&rw->rw_lock, v, v | RW_LOCK_WRITE_SPINNER)) { v = RW_READ_VALUE(rw); continue; } } spintries++; KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", rw->lock_object.lo_name); for (i = 0; i < rowner_loops; i++) { if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0) break; cpu_spinwait(); } KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); v = RW_READ_VALUE(rw); #ifdef KDTRACE_HOOKS lda.spin_cnt += rowner_loops - i; #endif if (i != rowner_loops) continue; } #endif ts = turnstile_trywait(&rw->lock_object); v = RW_READ_VALUE(rw); #ifdef ADAPTIVE_RWLOCKS /* * The current lock owner might have started executing * on another CPU (or the lock could have changed * owners) while we were waiting on the turnstile * chain lock. If so, drop the turnstile lock and try * again. */ if (!(v & RW_LOCK_READ)) { owner = (struct thread *)RW_OWNER(v); if (TD_IS_RUNNING(owner)) { turnstile_cancel(ts); continue; } } #endif /* * Check for the waiters flags about this rwlock. * If the lock was released, without maintain any pending * waiters queue, simply try to acquire it. * If a pending waiters queue is present, claim the lock * ownership and maintain the pending queue. */ x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); if ((v & ~x) == RW_UNLOCKED) { x &= ~RW_LOCK_WRITE_SPINNER; if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) { if (x) turnstile_claim(ts); else turnstile_cancel(ts); break; } turnstile_cancel(ts); v = RW_READ_VALUE(rw); continue; } /* * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to * set it. If we fail to set it, then loop back and try * again. */ if (!(v & RW_LOCK_WRITE_WAITERS)) { if (!atomic_cmpset_ptr(&rw->rw_lock, v, v | RW_LOCK_WRITE_WAITERS)) { turnstile_cancel(ts); v = RW_READ_VALUE(rw); continue; } if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p set write waiters flag", __func__, rw); } /* * We were unable to acquire the lock and the write waiters * flag is set, so we must block on the turnstile. */ if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, rw); #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&rw->lock_object); #endif turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&rw->lock_object); sleep_cnt++; #endif if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p resuming from turnstile", __func__, rw); #ifdef ADAPTIVE_RWLOCKS spintries = 0; #endif v = RW_READ_VALUE(rw); } #ifdef KDTRACE_HOOKS all_time += lockstat_nsecs(&rw->lock_object); if (sleep_time) LOCKSTAT_RECORD4(rw__block, rw, sleep_time, LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0, (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); /* Record only the loops spinning and not sleeping. */ if (lda.spin_cnt > sleep_cnt) LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0, (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); #endif LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, waittime, file, line, LOCKSTAT_WRITER); } /* * This function is called if lockstat is active or the first try at releasing * a write lock failed. The latter means that the lock is recursed or one of * the 2 waiter bits must be set indicating that at least one thread is waiting * on this lock. */ void __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file, int line) { struct rwlock *rw; struct turnstile *ts; uintptr_t v; int queue; if (SCHEDULER_STOPPED()) return; rw = rwlock2rw(c); v = RW_READ_VALUE(rw); if (v & RW_LOCK_WRITER_RECURSED) { if (--(rw->rw_recurse) == 0) atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED); if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); return; } LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER); if (v == tid && _rw_write_unlock(rw, tid)) return; KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), ("%s: neither of the waiter flags are set", __func__)); if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); turnstile_chain_lock(&rw->lock_object); ts = turnstile_lookup(&rw->lock_object); MPASS(ts != NULL); /* * Use the same algo as sx locks for now. Prefer waking up shared * waiters if we have any over writers. This is probably not ideal. * * 'v' is the value we are going to write back to rw_lock. If we * have waiters on both queues, we need to preserve the state of * the waiter flag for the queue we don't wake up. For now this is * hardcoded for the algorithm mentioned above. * * In the case of both readers and writers waiting we wakeup the * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a * new writer comes in before a reader it will claim the lock up * above. There is probably a potential priority inversion in * there that could be worked around either by waking both queues * of waiters or doing some complicated lock handoff gymnastics. */ v = RW_UNLOCKED; if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) { queue = TS_EXCLUSIVE_QUEUE; v |= (rw->rw_lock & RW_LOCK_READ_WAITERS); } else queue = TS_SHARED_QUEUE; /* Wake up all waiters for the specific queue. */ if (LOCK_LOG_TEST(&rw->lock_object, 0)) CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, queue == TS_SHARED_QUEUE ? "read" : "write"); turnstile_broadcast(ts, queue); atomic_store_rel_ptr(&rw->rw_lock, v); turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); turnstile_chain_unlock(&rw->lock_object); } /* * Attempt to do a non-blocking upgrade from a read lock to a write * lock. This will only succeed if this thread holds a single read * lock. Returns true if the upgrade succeeded and false otherwise. */ int __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; uintptr_t v, x, tid; struct turnstile *ts; int success; if (SCHEDULER_STOPPED()) return (1); rw = rwlock2rw(c); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); __rw_assert(c, RA_RLOCKED, file, line); /* * Attempt to switch from one reader to a writer. If there * are any write waiters, then we will have to lock the * turnstile first to prevent races with another writer * calling turnstile_wait() before we have claimed this * turnstile. So, do the simple case of no waiters first. */ tid = (uintptr_t)curthread; success = 0; for (;;) { v = rw->rw_lock; if (RW_READERS(v) > 1) break; if (!(v & RW_LOCK_WAITERS)) { success = atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid); if (!success) continue; break; } /* * Ok, we think we have waiters, so lock the turnstile. */ ts = turnstile_trywait(&rw->lock_object); v = rw->rw_lock; if (RW_READERS(v) > 1) { turnstile_cancel(ts); break; } /* * Try to switch from one reader to a writer again. This time * we honor the current state of the waiters flags. * If we obtain the lock with the flags set, then claim * ownership of the turnstile. */ x = rw->rw_lock & RW_LOCK_WAITERS; success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x); if (success) { if (x) turnstile_claim(ts); else turnstile_cancel(ts); break; } turnstile_cancel(ts); } LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); if (success) { curthread->td_rw_rlocks--; WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, file, line); LOCKSTAT_RECORD0(rw__upgrade, rw); } return (success); } /* * Downgrade a write lock into a single read lock. */ void __rw_downgrade(volatile uintptr_t *c, const char *file, int line) { struct rwlock *rw; struct turnstile *ts; uintptr_t tid, v; int rwait, wwait; if (SCHEDULER_STOPPED()) return; rw = rwlock2rw(c); KASSERT(rw->rw_lock != RW_DESTROYED, ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line); #ifndef INVARIANTS if (rw_recursed(rw)) panic("downgrade of a recursed lock"); #endif WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); /* * Convert from a writer to a single reader. First we handle * the easy case with no waiters. If there are any waiters, we * lock the turnstile and "disown" the lock. */ tid = (uintptr_t)curthread; if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) goto out; /* * Ok, we think we have waiters, so lock the turnstile so we can * read the waiter flags without any races. */ turnstile_chain_lock(&rw->lock_object); v = rw->rw_lock & RW_LOCK_WAITERS; rwait = v & RW_LOCK_READ_WAITERS; wwait = v & RW_LOCK_WRITE_WAITERS; MPASS(rwait | wwait); /* * Downgrade from a write lock while preserving waiters flag * and give up ownership of the turnstile. */ ts = turnstile_lookup(&rw->lock_object); MPASS(ts != NULL); if (!wwait) v &= ~RW_LOCK_READ_WAITERS; atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); /* * Wake other readers if there are no writers pending. Otherwise they * won't be able to acquire the lock anyway. */ if (rwait && !wwait) { turnstile_broadcast(ts, TS_SHARED_QUEUE); turnstile_unpend(ts, TS_EXCLUSIVE_LOCK); } else turnstile_disown(ts); turnstile_chain_unlock(&rw->lock_object); out: curthread->td_rw_rlocks++; LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); LOCKSTAT_RECORD0(rw__downgrade, rw); } #ifdef INVARIANT_SUPPORT #ifndef INVARIANTS #undef __rw_assert #endif /* * In the non-WITNESS case, rw_assert() can only detect that at least * *some* thread owns an rlock, but it cannot guarantee that *this* * thread owns an rlock. */ void __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line) { const struct rwlock *rw; if (panicstr != NULL) return; rw = rwlock2rw(c); switch (what) { case RA_LOCKED: case RA_LOCKED | RA_RECURSED: case RA_LOCKED | RA_NOTRECURSED: case RA_RLOCKED: case RA_RLOCKED | RA_RECURSED: case RA_RLOCKED | RA_NOTRECURSED: #ifdef WITNESS witness_assert(&rw->lock_object, what, file, line); #else /* * If some other thread has a write lock or we have one * and are asserting a read lock, fail. Also, if no one * has a lock at all, fail. */ if (rw->rw_lock == RW_UNLOCKED || (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED || rw_wowner(rw) != curthread))) panic("Lock %s not %slocked @ %s:%d\n", rw->lock_object.lo_name, (what & RA_RLOCKED) ? "read " : "", file, line); if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) { if (rw_recursed(rw)) { if (what & RA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", rw->lock_object.lo_name, file, line); } else if (what & RA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", rw->lock_object.lo_name, file, line); } #endif break; case RA_WLOCKED: case RA_WLOCKED | RA_RECURSED: case RA_WLOCKED | RA_NOTRECURSED: if (rw_wowner(rw) != curthread) panic("Lock %s not exclusively locked @ %s:%d\n", rw->lock_object.lo_name, file, line); if (rw_recursed(rw)) { if (what & RA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", rw->lock_object.lo_name, file, line); } else if (what & RA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", rw->lock_object.lo_name, file, line); break; case RA_UNLOCKED: #ifdef WITNESS witness_assert(&rw->lock_object, what, file, line); #else /* * If we hold a write lock fail. We can't reliably check * to see if we hold a read lock or not. */ if (rw_wowner(rw) == curthread) panic("Lock %s exclusively locked @ %s:%d\n", rw->lock_object.lo_name, file, line); #endif break; default: panic("Unknown rw lock assertion: %d @ %s:%d", what, file, line); } } #endif /* INVARIANT_SUPPORT */ #ifdef DDB void db_show_rwlock(const struct lock_object *lock) { const struct rwlock *rw; struct thread *td; rw = (const struct rwlock *)lock; db_printf(" state: "); if (rw->rw_lock == RW_UNLOCKED) db_printf("UNLOCKED\n"); else if (rw->rw_lock == RW_DESTROYED) { db_printf("DESTROYED\n"); return; } else if (rw->rw_lock & RW_LOCK_READ) db_printf("RLOCK: %ju locks\n", (uintmax_t)(RW_READERS(rw->rw_lock))); else { td = rw_wowner(rw); db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, td->td_tid, td->td_proc->p_pid, td->td_name); if (rw_recursed(rw)) db_printf(" recursed: %u\n", rw->rw_recurse); } db_printf(" waiters: "); switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { case RW_LOCK_READ_WAITERS: db_printf("readers\n"); break; case RW_LOCK_WRITE_WAITERS: db_printf("writers\n"); break; case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: db_printf("readers and writers\n"); break; default: db_printf("none\n"); break; } } #endif Index: head/sys/kern/kern_sx.c =================================================================== --- head/sys/kern/kern_sx.c (revision 323235) +++ head/sys/kern/kern_sx.c (revision 323236) @@ -1,1335 +1,1335 @@ /*- * Copyright (c) 2007 Attilio Rao * Copyright (c) 2001 Jason Evans * 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(s), this list of conditions and the following disclaimer as * the first lines of this file unmodified other than the possible * addition of one or more copyright notices. * 2. Redistributions in binary form must reproduce the above copyright * notice(s), this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``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 HOLDER(S) 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. */ /* * Shared/exclusive locks. This implementation attempts to ensure * deterministic lock granting behavior, so that slocks and xlocks are * interleaved. * * Priority propagation will not generally raise the priority of lock holders, * so should not be relied upon in combination with sx locks. */ #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include "opt_no_adaptive_sx.h" #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(SMP) && !defined(NO_ADAPTIVE_SX) #include #endif #ifdef DDB #include #endif #if defined(SMP) && !defined(NO_ADAPTIVE_SX) #define ADAPTIVE_SX #endif CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE); #ifdef HWPMC_HOOKS #include PMC_SOFT_DECLARE( , , lock, failed); #endif /* Handy macros for sleep queues. */ #define SQ_EXCLUSIVE_QUEUE 0 #define SQ_SHARED_QUEUE 1 /* * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file. We * drop Giant anytime we have to sleep or if we adaptively spin. */ #define GIANT_DECLARE \ int _giantcnt = 0; \ WITNESS_SAVE_DECL(Giant) \ #define GIANT_SAVE() do { \ if (mtx_owned(&Giant)) { \ WITNESS_SAVE(&Giant.lock_object, Giant); \ while (mtx_owned(&Giant)) { \ _giantcnt++; \ mtx_unlock(&Giant); \ } \ } \ } while (0) #define GIANT_RESTORE() do { \ if (_giantcnt > 0) { \ mtx_assert(&Giant, MA_NOTOWNED); \ while (_giantcnt--) \ mtx_lock(&Giant); \ WITNESS_RESTORE(&Giant.lock_object, Giant); \ } \ } while (0) /* * Returns true if an exclusive lock is recursed. It assumes * curthread currently has an exclusive lock. */ #define sx_recursed(sx) ((sx)->sx_recurse != 0) static void assert_sx(const struct lock_object *lock, int what); #ifdef DDB static void db_show_sx(const struct lock_object *lock); #endif static void lock_sx(struct lock_object *lock, uintptr_t how); #ifdef KDTRACE_HOOKS static int owner_sx(const struct lock_object *lock, struct thread **owner); #endif static uintptr_t unlock_sx(struct lock_object *lock); struct lock_class lock_class_sx = { .lc_name = "sx", .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE, .lc_assert = assert_sx, #ifdef DDB .lc_ddb_show = db_show_sx, #endif .lc_lock = lock_sx, .lc_unlock = unlock_sx, #ifdef KDTRACE_HOOKS .lc_owner = owner_sx, #endif }; #ifndef INVARIANTS #define _sx_assert(sx, what, file, line) #endif #ifdef ADAPTIVE_SX -static u_int asx_retries = 10; -static u_int asx_loops = 10000; +static __read_frequently u_int asx_retries = 10; +static __read_frequently u_int asx_loops = 10000; static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging"); SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, ""); SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, ""); -static struct lock_delay_config __read_mostly sx_delay; +static struct lock_delay_config __read_frequently sx_delay; SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base, 0, ""); SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max, 0, ""); LOCK_DELAY_SYSINIT_DEFAULT(sx_delay); #endif void assert_sx(const struct lock_object *lock, int what) { sx_assert((const struct sx *)lock, what); } void lock_sx(struct lock_object *lock, uintptr_t how) { struct sx *sx; sx = (struct sx *)lock; if (how) sx_slock(sx); else sx_xlock(sx); } uintptr_t unlock_sx(struct lock_object *lock) { struct sx *sx; sx = (struct sx *)lock; sx_assert(sx, SA_LOCKED | SA_NOTRECURSED); if (sx_xlocked(sx)) { sx_xunlock(sx); return (0); } else { sx_sunlock(sx); return (1); } } #ifdef KDTRACE_HOOKS int owner_sx(const struct lock_object *lock, struct thread **owner) { const struct sx *sx; uintptr_t x; sx = (const struct sx *)lock; x = sx->sx_lock; *owner = NULL; return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) : ((*owner = (struct thread *)SX_OWNER(x)) != NULL)); } #endif void sx_sysinit(void *arg) { struct sx_args *sargs = arg; sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags); } void sx_init_flags(struct sx *sx, const char *description, int opts) { int flags; MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK | SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0); ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock, ("%s: sx_lock not aligned for %s: %p", __func__, description, &sx->sx_lock)); flags = LO_SLEEPABLE | LO_UPGRADABLE; if (opts & SX_DUPOK) flags |= LO_DUPOK; if (opts & SX_NOPROFILE) flags |= LO_NOPROFILE; if (!(opts & SX_NOWITNESS)) flags |= LO_WITNESS; if (opts & SX_RECURSE) flags |= LO_RECURSABLE; if (opts & SX_QUIET) flags |= LO_QUIET; if (opts & SX_NEW) flags |= LO_NEW; flags |= opts & SX_NOADAPTIVE; lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags); sx->sx_lock = SX_LOCK_UNLOCKED; sx->sx_recurse = 0; } void sx_destroy(struct sx *sx) { KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held")); KASSERT(sx->sx_recurse == 0, ("sx lock still recursed")); sx->sx_lock = SX_LOCK_DESTROYED; lock_destroy(&sx->lock_object); } int sx_try_slock_(struct sx *sx, const char *file, int line) { uintptr_t x; if (SCHEDULER_STOPPED()) return (1); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("sx_try_slock() by idle thread %p on sx %s @ %s:%d", curthread, sx->lock_object.lo_name, file, line)); x = sx->sx_lock; for (;;) { KASSERT(x != SX_LOCK_DESTROYED, ("sx_try_slock() of destroyed sx @ %s:%d", file, line)); if (!(x & SX_LOCK_SHARED)) break; if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) { LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line); WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line); LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx, 0, 0, file, line, LOCKSTAT_READER); TD_LOCKS_INC(curthread); return (1); } } LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line); return (0); } int _sx_xlock(struct sx *sx, int opts, const char *file, int line) { uintptr_t tid, x; int error = 0; KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() || !TD_IS_IDLETHREAD(curthread), ("sx_xlock() by idle thread %p on sx %s @ %s:%d", curthread, sx->lock_object.lo_name, file, line)); KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_xlock() of destroyed sx @ %s:%d", file, line)); WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); tid = (uintptr_t)curthread; x = SX_LOCK_UNLOCKED; if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid)) error = _sx_xlock_hard(sx, x, tid, opts, file, line); else LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx, 0, 0, file, line, LOCKSTAT_WRITER); if (!error) { LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse, file, line); WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line); TD_LOCKS_INC(curthread); } return (error); } int sx_try_xlock_(struct sx *sx, const char *file, int line) { struct thread *td; uintptr_t tid, x; int rval; bool recursed; td = curthread; tid = (uintptr_t)td; if (SCHEDULER_STOPPED_TD(td)) return (1); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td), ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d", curthread, sx->lock_object.lo_name, file, line)); KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_try_xlock() of destroyed sx @ %s:%d", file, line)); rval = 1; recursed = false; x = SX_LOCK_UNLOCKED; for (;;) { if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid)) break; if (x == SX_LOCK_UNLOCKED) continue; if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) { sx->sx_recurse++; atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED); break; } rval = 0; break; } LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line); if (rval) { WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, file, line); if (!recursed) LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx, 0, 0, file, line, LOCKSTAT_WRITER); TD_LOCKS_INC(curthread); } return (rval); } void _sx_xunlock(struct sx *sx, const char *file, int line) { KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_xunlock() of destroyed sx @ %s:%d", file, line)); _sx_assert(sx, SA_XLOCKED, file, line); WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file, line); #if LOCK_DEBUG > 0 _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line); #else __sx_xunlock(sx, curthread, file, line); #endif TD_LOCKS_DEC(curthread); } /* * Try to do a non-blocking upgrade from a shared lock to an exclusive lock. * This will only succeed if this thread holds a single shared lock. * Return 1 if if the upgrade succeed, 0 otherwise. */ int sx_try_upgrade_(struct sx *sx, const char *file, int line) { uintptr_t x; int success; if (SCHEDULER_STOPPED()) return (1); KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line)); _sx_assert(sx, SA_SLOCKED, file, line); /* * Try to switch from one shared lock to an exclusive lock. We need * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that * we will wake up the exclusive waiters when we drop the lock. */ x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS; success = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x, (uintptr_t)curthread | x); LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line); if (success) { WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, file, line); LOCKSTAT_RECORD0(sx__upgrade, sx); } return (success); } /* * Downgrade an unrecursed exclusive lock into a single shared lock. */ void sx_downgrade_(struct sx *sx, const char *file, int line) { uintptr_t x; int wakeup_swapper; if (SCHEDULER_STOPPED()) return; KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_downgrade() of destroyed sx @ %s:%d", file, line)); _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line); #ifndef INVARIANTS if (sx_recursed(sx)) panic("downgrade of a recursed lock"); #endif WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line); /* * Try to switch from an exclusive lock with no shared waiters * to one sharer with no shared waiters. If there are * exclusive waiters, we don't need to lock the sleep queue so * long as we preserve the flag. We do one quick try and if * that fails we grab the sleepq lock to keep the flags from * changing and do it the slow way. * * We have to lock the sleep queue if there are shared waiters * so we can wake them up. */ x = sx->sx_lock; if (!(x & SX_LOCK_SHARED_WAITERS) && atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) | (x & SX_LOCK_EXCLUSIVE_WAITERS))) { LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line); return; } /* * Lock the sleep queue so we can read the waiters bits * without any races and wakeup any shared waiters. */ sleepq_lock(&sx->lock_object); /* * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single * shared lock. If there are any shared waiters, wake them up. */ wakeup_swapper = 0; x = sx->sx_lock; atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | (x & SX_LOCK_EXCLUSIVE_WAITERS)); if (x & SX_LOCK_SHARED_WAITERS) wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, SQ_SHARED_QUEUE); sleepq_release(&sx->lock_object); LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line); LOCKSTAT_RECORD0(sx__downgrade, sx); if (wakeup_swapper) kick_proc0(); } /* * This function represents the so-called 'hard case' for sx_xlock * operation. All 'easy case' failures are redirected to this. Note * that ideally this would be a static function, but it needs to be * accessible from at least sx.h. */ int _sx_xlock_hard(struct sx *sx, uintptr_t x, uintptr_t tid, int opts, const char *file, int line) { GIANT_DECLARE; #ifdef ADAPTIVE_SX volatile struct thread *owner; u_int i, spintries = 0; #endif #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif int error = 0; #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS) struct lock_delay_arg lda; #endif #ifdef KDTRACE_HOOKS uintptr_t state; u_int sleep_cnt = 0; int64_t sleep_time = 0; int64_t all_time = 0; #endif if (SCHEDULER_STOPPED()) return (0); #if defined(ADAPTIVE_SX) lock_delay_arg_init(&lda, &sx_delay); #elif defined(KDTRACE_HOOKS) lock_delay_arg_init(&lda, NULL); #endif if (__predict_false(x == SX_LOCK_UNLOCKED)) x = SX_READ_VALUE(sx); /* If we already hold an exclusive lock, then recurse. */ if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) { KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0, ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n", sx->lock_object.lo_name, file, line)); sx->sx_recurse++; atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED); if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx); return (0); } if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, sx->lock_object.lo_name, (void *)sx->sx_lock, file, line); #ifdef KDTRACE_HOOKS all_time -= lockstat_nsecs(&sx->lock_object); state = x; #endif for (;;) { if (x == SX_LOCK_UNLOCKED) { if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid)) break; continue; } #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&sx->lock_object, &contested, &waittime); #ifdef ADAPTIVE_SX /* * If the lock is write locked and the owner is * running on another CPU, spin until the owner stops * running or the state of the lock changes. */ if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { if ((x & SX_LOCK_SHARED) == 0) { owner = lv_sx_owner(x); if (TD_IS_RUNNING(owner)) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR3(KTR_LOCK, "%s: spinning on %p held by %p", __func__, sx, owner); KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", sx->lock_object.lo_name); GIANT_SAVE(); do { lock_delay(&lda); x = SX_READ_VALUE(sx); owner = lv_sx_owner(x); } while (owner != NULL && TD_IS_RUNNING(owner)); KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); continue; } } else if (SX_SHARERS(x) && spintries < asx_retries) { KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", sx->lock_object.lo_name); GIANT_SAVE(); spintries++; for (i = 0; i < asx_loops; i++) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR4(KTR_LOCK, "%s: shared spinning on %p with %u and %u", __func__, sx, spintries, i); x = sx->sx_lock; if ((x & SX_LOCK_SHARED) == 0 || SX_SHARERS(x) == 0) break; cpu_spinwait(); #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif } KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); x = SX_READ_VALUE(sx); if (i != asx_loops) continue; } } #endif sleepq_lock(&sx->lock_object); x = SX_READ_VALUE(sx); /* * If the lock was released while spinning on the * sleep queue chain lock, try again. */ if (x == SX_LOCK_UNLOCKED) { sleepq_release(&sx->lock_object); continue; } #ifdef ADAPTIVE_SX /* * The current lock owner might have started executing * on another CPU (or the lock could have changed * owners) while we were waiting on the sleep queue * chain lock. If so, drop the sleep queue lock and try * again. */ if (!(x & SX_LOCK_SHARED) && (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { owner = (struct thread *)SX_OWNER(x); if (TD_IS_RUNNING(owner)) { sleepq_release(&sx->lock_object); continue; } } #endif /* * If an exclusive lock was released with both shared * and exclusive waiters and a shared waiter hasn't * woken up and acquired the lock yet, sx_lock will be * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS. * If we see that value, try to acquire it once. Note * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS * as there are other exclusive waiters still. If we * fail, restart the loop. */ if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) { if (atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS, tid | SX_LOCK_EXCLUSIVE_WAITERS)) { sleepq_release(&sx->lock_object); CTR2(KTR_LOCK, "%s: %p claimed by new writer", __func__, sx); break; } sleepq_release(&sx->lock_object); x = SX_READ_VALUE(sx); continue; } /* * Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail, * than loop back and retry. */ if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) { if (!atomic_cmpset_ptr(&sx->sx_lock, x, x | SX_LOCK_EXCLUSIVE_WAITERS)) { sleepq_release(&sx->lock_object); x = SX_READ_VALUE(sx); continue; } if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p set excl waiters flag", __func__, sx); } /* * Since we have been unable to acquire the exclusive * lock and the exclusive waiters flag is set, we have * to sleep. */ if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p blocking on sleep queue", __func__, sx); #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&sx->lock_object); #endif GIANT_SAVE(); sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name, SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ? SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE); if (!(opts & SX_INTERRUPTIBLE)) sleepq_wait(&sx->lock_object, 0); else error = sleepq_wait_sig(&sx->lock_object, 0); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&sx->lock_object); sleep_cnt++; #endif if (error) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: interruptible sleep by %p suspended by signal", __func__, sx); break; } if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p resuming from sleep queue", __func__, sx); x = SX_READ_VALUE(sx); } #ifdef KDTRACE_HOOKS all_time += lockstat_nsecs(&sx->lock_object); if (sleep_time) LOCKSTAT_RECORD4(sx__block, sx, sleep_time, LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0, (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state)); if (lda.spin_cnt > sleep_cnt) LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time, LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0, (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state)); #endif if (!error) LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx, contested, waittime, file, line, LOCKSTAT_WRITER); GIANT_RESTORE(); return (error); } /* * This function represents the so-called 'hard case' for sx_xunlock * operation. All 'easy case' failures are redirected to this. Note * that ideally this would be a static function, but it needs to be * accessible from at least sx.h. */ void _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line) { uintptr_t x; int queue, wakeup_swapper; if (SCHEDULER_STOPPED()) return; MPASS(!(sx->sx_lock & SX_LOCK_SHARED)); x = SX_READ_VALUE(sx); if (x & SX_LOCK_RECURSED) { /* The lock is recursed, unrecurse one level. */ if ((--sx->sx_recurse) == 0) atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED); if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx); return; } LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER); if (x == tid && atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED)) return; MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)); if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p contested", __func__, sx); sleepq_lock(&sx->lock_object); x = SX_LOCK_UNLOCKED; /* * The wake up algorithm here is quite simple and probably not * ideal. It gives precedence to shared waiters if they are * present. For this condition, we have to preserve the * state of the exclusive waiters flag. * If interruptible sleeps left the shared queue empty avoid a * starvation for the threads sleeping on the exclusive queue by giving * them precedence and cleaning up the shared waiters bit anyway. */ if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 && sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) { queue = SQ_SHARED_QUEUE; x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS); } else queue = SQ_EXCLUSIVE_QUEUE; /* Wake up all the waiters for the specific queue. */ if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue", __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" : "exclusive"); atomic_store_rel_ptr(&sx->sx_lock, x); wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, queue); sleepq_release(&sx->lock_object); if (wakeup_swapper) kick_proc0(); } static bool __always_inline __sx_slock_try(struct sx *sx, uintptr_t *xp, const char *file, int line) { /* * If no other thread has an exclusive lock then try to bump up * the count of sharers. Since we have to preserve the state * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the * shared lock loop back and retry. */ while (*xp & SX_LOCK_SHARED) { MPASS(!(*xp & SX_LOCK_SHARED_WAITERS)); if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp, *xp + SX_ONE_SHARER)) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR4(KTR_LOCK, "%s: %p succeed %p -> %p", __func__, sx, (void *)*xp, (void *)(*xp + SX_ONE_SHARER)); return (true); } } return (false); } static int __noinline _sx_slock_hard(struct sx *sx, int opts, const char *file, int line, uintptr_t x) { GIANT_DECLARE; #ifdef ADAPTIVE_SX volatile struct thread *owner; #endif #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif int error = 0; #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS) struct lock_delay_arg lda; #endif #ifdef KDTRACE_HOOKS uintptr_t state; u_int sleep_cnt = 0; int64_t sleep_time = 0; int64_t all_time = 0; #endif if (SCHEDULER_STOPPED()) return (0); #if defined(ADAPTIVE_SX) lock_delay_arg_init(&lda, &sx_delay); #elif defined(KDTRACE_HOOKS) lock_delay_arg_init(&lda, NULL); #endif #ifdef KDTRACE_HOOKS all_time -= lockstat_nsecs(&sx->lock_object); state = x; #endif /* * As with rwlocks, we don't make any attempt to try to block * shared locks once there is an exclusive waiter. */ for (;;) { if (__sx_slock_try(sx, &x, file, line)) break; #ifdef KDTRACE_HOOKS lda.spin_cnt++; #endif #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&sx->lock_object, &contested, &waittime); #ifdef ADAPTIVE_SX /* * If the owner is running on another CPU, spin until * the owner stops running or the state of the lock * changes. */ if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { owner = lv_sx_owner(x); if (TD_IS_RUNNING(owner)) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR3(KTR_LOCK, "%s: spinning on %p held by %p", __func__, sx, owner); KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), "spinning", "lockname:\"%s\"", sx->lock_object.lo_name); GIANT_SAVE(); do { lock_delay(&lda); x = SX_READ_VALUE(sx); owner = lv_sx_owner(x); } while (owner != NULL && TD_IS_RUNNING(owner)); KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), "running"); continue; } } #endif /* * Some other thread already has an exclusive lock, so * start the process of blocking. */ sleepq_lock(&sx->lock_object); x = SX_READ_VALUE(sx); /* * The lock could have been released while we spun. * In this case loop back and retry. */ if (x & SX_LOCK_SHARED) { sleepq_release(&sx->lock_object); continue; } #ifdef ADAPTIVE_SX /* * If the owner is running on another CPU, spin until * the owner stops running or the state of the lock * changes. */ if (!(x & SX_LOCK_SHARED) && (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) { owner = (struct thread *)SX_OWNER(x); if (TD_IS_RUNNING(owner)) { sleepq_release(&sx->lock_object); x = SX_READ_VALUE(sx); continue; } } #endif /* * Try to set the SX_LOCK_SHARED_WAITERS flag. If we * fail to set it drop the sleep queue lock and loop * back. */ if (!(x & SX_LOCK_SHARED_WAITERS)) { if (!atomic_cmpset_ptr(&sx->sx_lock, x, x | SX_LOCK_SHARED_WAITERS)) { sleepq_release(&sx->lock_object); x = SX_READ_VALUE(sx); continue; } if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p set shared waiters flag", __func__, sx); } /* * Since we have been unable to acquire the shared lock, * we have to sleep. */ if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p blocking on sleep queue", __func__, sx); #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&sx->lock_object); #endif GIANT_SAVE(); sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name, SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ? SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE); if (!(opts & SX_INTERRUPTIBLE)) sleepq_wait(&sx->lock_object, 0); else error = sleepq_wait_sig(&sx->lock_object, 0); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&sx->lock_object); sleep_cnt++; #endif if (error) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: interruptible sleep by %p suspended by signal", __func__, sx); break; } if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p resuming from sleep queue", __func__, sx); x = SX_READ_VALUE(sx); } #ifdef KDTRACE_HOOKS all_time += lockstat_nsecs(&sx->lock_object); if (sleep_time) LOCKSTAT_RECORD4(sx__block, sx, sleep_time, LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0, (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state)); if (lda.spin_cnt > sleep_cnt) LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time, LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0, (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state)); #endif if (error == 0) { LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx, contested, waittime, file, line, LOCKSTAT_READER); } GIANT_RESTORE(); return (error); } int _sx_slock(struct sx *sx, int opts, const char *file, int line) { uintptr_t x; int error; KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() || !TD_IS_IDLETHREAD(curthread), ("sx_slock() by idle thread %p on sx %s @ %s:%d", curthread, sx->lock_object.lo_name, file, line)); KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_slock() of destroyed sx @ %s:%d", file, line)); WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL); error = 0; x = SX_READ_VALUE(sx); if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) || !__sx_slock_try(sx, &x, file, line))) error = _sx_slock_hard(sx, opts, file, line, x); if (error == 0) { LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line); WITNESS_LOCK(&sx->lock_object, 0, file, line); TD_LOCKS_INC(curthread); } return (error); } static bool __always_inline _sx_sunlock_try(struct sx *sx, uintptr_t *xp) { for (;;) { /* * We should never have sharers while at least one thread * holds a shared lock. */ KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS), ("%s: waiting sharers", __func__)); /* * See if there is more than one shared lock held. If * so, just drop one and return. */ if (SX_SHARERS(*xp) > 1) { if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp, *xp - SX_ONE_SHARER)) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR4(KTR_LOCK, "%s: %p succeeded %p -> %p", __func__, sx, (void *)*xp, (void *)(*xp - SX_ONE_SHARER)); return (true); } continue; } /* * If there aren't any waiters for an exclusive lock, * then try to drop it quickly. */ if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) { MPASS(*xp == SX_SHARERS_LOCK(1)); *xp = SX_SHARERS_LOCK(1); if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp, SX_LOCK_UNLOCKED)) { if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p last succeeded", __func__, sx); return (true); } continue; } break; } return (false); } static void __noinline _sx_sunlock_hard(struct sx *sx, uintptr_t x, const char *file, int line) { int wakeup_swapper; if (SCHEDULER_STOPPED()) return; LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER); for (;;) { if (_sx_sunlock_try(sx, &x)) break; /* * At this point, there should just be one sharer with * exclusive waiters. */ MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS)); sleepq_lock(&sx->lock_object); /* * Wake up semantic here is quite simple: * Just wake up all the exclusive waiters. * Note that the state of the lock could have changed, * so if it fails loop back and retry. */ if (!atomic_cmpset_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS, SX_LOCK_UNLOCKED)) { sleepq_release(&sx->lock_object); x = SX_READ_VALUE(sx); continue; } if (LOCK_LOG_TEST(&sx->lock_object, 0)) CTR2(KTR_LOCK, "%s: %p waking up all thread on" "exclusive queue", __func__, sx); wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0, SQ_EXCLUSIVE_QUEUE); sleepq_release(&sx->lock_object); if (wakeup_swapper) kick_proc0(); break; } } void _sx_sunlock(struct sx *sx, const char *file, int line) { uintptr_t x; KASSERT(sx->sx_lock != SX_LOCK_DESTROYED, ("sx_sunlock() of destroyed sx @ %s:%d", file, line)); _sx_assert(sx, SA_SLOCKED, file, line); WITNESS_UNLOCK(&sx->lock_object, 0, file, line); LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line); x = SX_READ_VALUE(sx); if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) || !_sx_sunlock_try(sx, &x))) _sx_sunlock_hard(sx, x, file, line); TD_LOCKS_DEC(curthread); } #ifdef INVARIANT_SUPPORT #ifndef INVARIANTS #undef _sx_assert #endif /* * In the non-WITNESS case, sx_assert() can only detect that at least * *some* thread owns an slock, but it cannot guarantee that *this* * thread owns an slock. */ void _sx_assert(const struct sx *sx, int what, const char *file, int line) { #ifndef WITNESS int slocked = 0; #endif if (panicstr != NULL) return; switch (what) { case SA_SLOCKED: case SA_SLOCKED | SA_NOTRECURSED: case SA_SLOCKED | SA_RECURSED: #ifndef WITNESS slocked = 1; /* FALLTHROUGH */ #endif case SA_LOCKED: case SA_LOCKED | SA_NOTRECURSED: case SA_LOCKED | SA_RECURSED: #ifdef WITNESS witness_assert(&sx->lock_object, what, file, line); #else /* * If some other thread has an exclusive lock or we * have one and are asserting a shared lock, fail. * Also, if no one has a lock at all, fail. */ if (sx->sx_lock == SX_LOCK_UNLOCKED || (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked || sx_xholder(sx) != curthread))) panic("Lock %s not %slocked @ %s:%d\n", sx->lock_object.lo_name, slocked ? "share " : "", file, line); if (!(sx->sx_lock & SX_LOCK_SHARED)) { if (sx_recursed(sx)) { if (what & SA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", sx->lock_object.lo_name, file, line); } else if (what & SA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", sx->lock_object.lo_name, file, line); } #endif break; case SA_XLOCKED: case SA_XLOCKED | SA_NOTRECURSED: case SA_XLOCKED | SA_RECURSED: if (sx_xholder(sx) != curthread) panic("Lock %s not exclusively locked @ %s:%d\n", sx->lock_object.lo_name, file, line); if (sx_recursed(sx)) { if (what & SA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", sx->lock_object.lo_name, file, line); } else if (what & SA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", sx->lock_object.lo_name, file, line); break; case SA_UNLOCKED: #ifdef WITNESS witness_assert(&sx->lock_object, what, file, line); #else /* * If we hold an exclusve lock fail. We can't * reliably check to see if we hold a shared lock or * not. */ if (sx_xholder(sx) == curthread) panic("Lock %s exclusively locked @ %s:%d\n", sx->lock_object.lo_name, file, line); #endif break; default: panic("Unknown sx lock assertion: %d @ %s:%d", what, file, line); } } #endif /* INVARIANT_SUPPORT */ #ifdef DDB static void db_show_sx(const struct lock_object *lock) { struct thread *td; const struct sx *sx; sx = (const struct sx *)lock; db_printf(" state: "); if (sx->sx_lock == SX_LOCK_UNLOCKED) db_printf("UNLOCKED\n"); else if (sx->sx_lock == SX_LOCK_DESTROYED) { db_printf("DESTROYED\n"); return; } else if (sx->sx_lock & SX_LOCK_SHARED) db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock)); else { td = sx_xholder(sx); db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td, td->td_tid, td->td_proc->p_pid, td->td_name); if (sx_recursed(sx)) db_printf(" recursed: %d\n", sx->sx_recurse); } db_printf(" waiters: "); switch(sx->sx_lock & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) { case SX_LOCK_SHARED_WAITERS: db_printf("shared\n"); break; case SX_LOCK_EXCLUSIVE_WAITERS: db_printf("exclusive\n"); break; case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS: db_printf("exclusive and shared\n"); break; default: db_printf("none\n"); } } /* * Check to see if a thread that is blocked on a sleep queue is actually * blocked on an sx lock. If so, output some details and return true. * If the lock has an exclusive owner, return that in *ownerp. */ int sx_chain(struct thread *td, struct thread **ownerp) { struct sx *sx; /* * Check to see if this thread is blocked on an sx lock. * First, we check the lock class. If that is ok, then we * compare the lock name against the wait message. */ sx = td->td_wchan; if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx || sx->lock_object.lo_name != td->td_wmesg) return (0); /* We think we have an sx lock, so output some details. */ db_printf("blocked on sx \"%s\" ", td->td_wmesg); *ownerp = sx_xholder(sx); if (sx->sx_lock & SX_LOCK_SHARED) db_printf("SLOCK (count %ju)\n", (uintmax_t)SX_SHARERS(sx->sx_lock)); else db_printf("XLOCK\n"); return (1); } #endif Index: head/sys/security/audit/audit.c =================================================================== --- head/sys/security/audit/audit.c (revision 323235) +++ head/sys/security/audit/audit.c (revision 323236) @@ -1,809 +1,809 @@ /*- * Copyright (c) 1999-2005 Apple Inc. * Copyright (c) 2006-2007, 2016-2017 Robert N. M. Watson * All rights reserved. * * Portions of this software were developed by BAE Systems, the University of * Cambridge Computer Laboratory, and Memorial University under DARPA/AFRL * contract FA8650-15-C-7558 ("CADETS"), as part of the DARPA Transparent * Computing (TC) research program. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of Apple Inc. ("Apple") 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 APPLE AND ITS 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 APPLE OR ITS 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 #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include FEATURE(audit, "BSM audit support"); static uma_zone_t audit_record_zone; static MALLOC_DEFINE(M_AUDITCRED, "audit_cred", "Audit cred storage"); MALLOC_DEFINE(M_AUDITDATA, "audit_data", "Audit data storage"); MALLOC_DEFINE(M_AUDITPATH, "audit_path", "Audit path storage"); MALLOC_DEFINE(M_AUDITTEXT, "audit_text", "Audit text storage"); MALLOC_DEFINE(M_AUDITGIDSET, "audit_gidset", "Audit GID set storage"); static SYSCTL_NODE(_security, OID_AUTO, audit, CTLFLAG_RW, 0, "TrustedBSD audit controls"); /* * Audit control settings that are set/read by system calls and are hence * non-static. * * Define the audit control flags. */ -int audit_enabled; +int __read_frequently audit_enabled; int audit_suspended; /* * Flags controlling behavior in low storage situations. Should we panic if * a write fails? Should we fail stop if we're out of disk space? */ int audit_panic_on_write_fail; int audit_fail_stop; int audit_argv; int audit_arge; /* * Are we currently "failing stop" due to out of disk space? */ int audit_in_failure; /* * Global audit statistics. */ struct audit_fstat audit_fstat; /* * Preselection mask for non-attributable events. */ struct au_mask audit_nae_mask; /* * Mutex to protect global variables shared between various threads and * processes. */ struct mtx audit_mtx; /* * Queue of audit records ready for delivery to disk. We insert new records * at the tail, and remove records from the head. Also, a count of the * number of records used for checking queue depth. In addition, a counter * of records that we have allocated but are not yet in the queue, which is * needed to estimate the total size of the combined set of records * outstanding in the system. */ struct kaudit_queue audit_q; int audit_q_len; int audit_pre_q_len; /* * Audit queue control settings (minimum free, low/high water marks, etc.) */ struct au_qctrl audit_qctrl; /* * Condition variable to signal to the worker that it has work to do: either * new records are in the queue, or a log replacement is taking place. */ struct cv audit_worker_cv; /* * Condition variable to flag when crossing the low watermark, meaning that * threads blocked due to hitting the high watermark can wake up and continue * to commit records. */ struct cv audit_watermark_cv; /* * Condition variable for auditing threads wait on when in fail-stop mode. * Threads wait on this CV forever (and ever), never seeing the light of day * again. */ static struct cv audit_fail_cv; /* * Optional DTrace audit provider support: function pointers for preselection * and commit events. */ #ifdef KDTRACE_HOOKS void *(*dtaudit_hook_preselect)(au_id_t auid, au_event_t event, au_class_t class); int (*dtaudit_hook_commit)(struct kaudit_record *kar, au_id_t auid, au_event_t event, au_class_t class, int sorf); void (*dtaudit_hook_bsm)(struct kaudit_record *kar, au_id_t auid, au_event_t event, au_class_t class, int sorf, void *bsm_data, size_t bsm_lenlen); #endif /* * Kernel audit information. This will store the current audit address * or host information that the kernel will use when it's generating * audit records. This data is modified by the A_GET{SET}KAUDIT auditon(2) * command. */ static struct auditinfo_addr audit_kinfo; static struct rwlock audit_kinfo_lock; #define KINFO_LOCK_INIT() rw_init(&audit_kinfo_lock, \ "audit_kinfo_lock") #define KINFO_RLOCK() rw_rlock(&audit_kinfo_lock) #define KINFO_WLOCK() rw_wlock(&audit_kinfo_lock) #define KINFO_RUNLOCK() rw_runlock(&audit_kinfo_lock) #define KINFO_WUNLOCK() rw_wunlock(&audit_kinfo_lock) void audit_set_kinfo(struct auditinfo_addr *ak) { KASSERT(ak->ai_termid.at_type == AU_IPv4 || ak->ai_termid.at_type == AU_IPv6, ("audit_set_kinfo: invalid address type")); KINFO_WLOCK(); audit_kinfo = *ak; KINFO_WUNLOCK(); } void audit_get_kinfo(struct auditinfo_addr *ak) { KASSERT(audit_kinfo.ai_termid.at_type == AU_IPv4 || audit_kinfo.ai_termid.at_type == AU_IPv6, ("audit_set_kinfo: invalid address type")); KINFO_RLOCK(); *ak = audit_kinfo; KINFO_RUNLOCK(); } /* * Construct an audit record for the passed thread. */ static int audit_record_ctor(void *mem, int size, void *arg, int flags) { struct kaudit_record *ar; struct thread *td; struct ucred *cred; struct prison *pr; KASSERT(sizeof(*ar) == size, ("audit_record_ctor: wrong size")); td = arg; ar = mem; bzero(ar, sizeof(*ar)); ar->k_ar.ar_magic = AUDIT_RECORD_MAGIC; nanotime(&ar->k_ar.ar_starttime); /* * Export the subject credential. */ cred = td->td_ucred; cru2x(cred, &ar->k_ar.ar_subj_cred); ar->k_ar.ar_subj_ruid = cred->cr_ruid; ar->k_ar.ar_subj_rgid = cred->cr_rgid; ar->k_ar.ar_subj_egid = cred->cr_groups[0]; ar->k_ar.ar_subj_auid = cred->cr_audit.ai_auid; ar->k_ar.ar_subj_asid = cred->cr_audit.ai_asid; ar->k_ar.ar_subj_pid = td->td_proc->p_pid; ar->k_ar.ar_subj_amask = cred->cr_audit.ai_mask; ar->k_ar.ar_subj_term_addr = cred->cr_audit.ai_termid; /* * If this process is jailed, make sure we capture the name of the * jail so we can use it to generate a zonename token when we covert * this record to BSM. */ if (jailed(cred)) { pr = cred->cr_prison; (void) strlcpy(ar->k_ar.ar_jailname, pr->pr_name, sizeof(ar->k_ar.ar_jailname)); } else ar->k_ar.ar_jailname[0] = '\0'; return (0); } static void audit_record_dtor(void *mem, int size, void *arg) { struct kaudit_record *ar; KASSERT(sizeof(*ar) == size, ("audit_record_dtor: wrong size")); ar = mem; if (ar->k_ar.ar_arg_upath1 != NULL) free(ar->k_ar.ar_arg_upath1, M_AUDITPATH); if (ar->k_ar.ar_arg_upath2 != NULL) free(ar->k_ar.ar_arg_upath2, M_AUDITPATH); if (ar->k_ar.ar_arg_text != NULL) free(ar->k_ar.ar_arg_text, M_AUDITTEXT); if (ar->k_udata != NULL) free(ar->k_udata, M_AUDITDATA); if (ar->k_ar.ar_arg_argv != NULL) free(ar->k_ar.ar_arg_argv, M_AUDITTEXT); if (ar->k_ar.ar_arg_envv != NULL) free(ar->k_ar.ar_arg_envv, M_AUDITTEXT); if (ar->k_ar.ar_arg_groups.gidset != NULL) free(ar->k_ar.ar_arg_groups.gidset, M_AUDITGIDSET); } /* * Initialize the Audit subsystem: configuration state, work queue, * synchronization primitives, worker thread, and trigger device node. Also * call into the BSM assembly code to initialize it. */ static void audit_init(void) { audit_enabled = 0; audit_suspended = 0; audit_panic_on_write_fail = 0; audit_fail_stop = 0; audit_in_failure = 0; audit_argv = 0; audit_arge = 0; audit_fstat.af_filesz = 0; /* '0' means unset, unbounded. */ audit_fstat.af_currsz = 0; audit_nae_mask.am_success = 0; audit_nae_mask.am_failure = 0; TAILQ_INIT(&audit_q); audit_q_len = 0; audit_pre_q_len = 0; audit_qctrl.aq_hiwater = AQ_HIWATER; audit_qctrl.aq_lowater = AQ_LOWATER; audit_qctrl.aq_bufsz = AQ_BUFSZ; audit_qctrl.aq_minfree = AU_FS_MINFREE; audit_kinfo.ai_termid.at_type = AU_IPv4; audit_kinfo.ai_termid.at_addr[0] = INADDR_ANY; mtx_init(&audit_mtx, "audit_mtx", NULL, MTX_DEF); KINFO_LOCK_INIT(); cv_init(&audit_worker_cv, "audit_worker_cv"); cv_init(&audit_watermark_cv, "audit_watermark_cv"); cv_init(&audit_fail_cv, "audit_fail_cv"); audit_record_zone = uma_zcreate("audit_record", sizeof(struct kaudit_record), audit_record_ctor, audit_record_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); /* Initialize the BSM audit subsystem. */ kau_init(); audit_trigger_init(); /* Register shutdown handler. */ EVENTHANDLER_REGISTER(shutdown_pre_sync, audit_shutdown, NULL, SHUTDOWN_PRI_FIRST); /* Start audit worker thread. */ audit_worker_init(); } SYSINIT(audit_init, SI_SUB_AUDIT, SI_ORDER_FIRST, audit_init, NULL); /* * Drain the audit queue and close the log at shutdown. Note that this can * be called both from the system shutdown path and also from audit * configuration syscalls, so 'arg' and 'howto' are ignored. * * XXXRW: In FreeBSD 7.x and 8.x, this fails to wait for the record queue to * drain before returning, which could lead to lost records on shutdown. */ void audit_shutdown(void *arg, int howto) { audit_rotate_vnode(NULL, NULL); } /* * Return the current thread's audit record, if any. */ struct kaudit_record * currecord(void) { return (curthread->td_ar); } /* * XXXAUDIT: There are a number of races present in the code below due to * release and re-grab of the mutex. The code should be revised to become * slightly less racy. * * XXXAUDIT: Shouldn't there be logic here to sleep waiting on available * pre_q space, suspending the system call until there is room? */ struct kaudit_record * audit_new(int event, struct thread *td) { struct kaudit_record *ar; int no_record; mtx_lock(&audit_mtx); no_record = (audit_suspended || !audit_enabled); mtx_unlock(&audit_mtx); if (no_record) return (NULL); /* * Note: the number of outstanding uncommitted audit records is * limited to the number of concurrent threads servicing system calls * in the kernel. */ ar = uma_zalloc_arg(audit_record_zone, td, M_WAITOK); ar->k_ar.ar_event = event; mtx_lock(&audit_mtx); audit_pre_q_len++; mtx_unlock(&audit_mtx); return (ar); } void audit_free(struct kaudit_record *ar) { uma_zfree(audit_record_zone, ar); } void audit_commit(struct kaudit_record *ar, int error, int retval) { au_event_t event; au_class_t class; au_id_t auid; int sorf; struct au_mask *aumask; if (ar == NULL) return; ar->k_ar.ar_errno = error; ar->k_ar.ar_retval = retval; nanotime(&ar->k_ar.ar_endtime); /* * Decide whether to commit the audit record by checking the error * value from the system call and using the appropriate audit mask. */ if (ar->k_ar.ar_subj_auid == AU_DEFAUDITID) aumask = &audit_nae_mask; else aumask = &ar->k_ar.ar_subj_amask; if (error) sorf = AU_PRS_FAILURE; else sorf = AU_PRS_SUCCESS; /* * syscalls.master sometimes contains a prototype event number, which * we will transform into a more specific event number now that we * have more complete information gathered during the system call. */ switch(ar->k_ar.ar_event) { case AUE_OPEN_RWTC: ar->k_ar.ar_event = audit_flags_and_error_to_openevent( ar->k_ar.ar_arg_fflags, error); break; case AUE_OPENAT_RWTC: ar->k_ar.ar_event = audit_flags_and_error_to_openatevent( ar->k_ar.ar_arg_fflags, error); break; case AUE_SYSCTL: ar->k_ar.ar_event = audit_ctlname_to_sysctlevent( ar->k_ar.ar_arg_ctlname, ar->k_ar.ar_valid_arg); break; case AUE_AUDITON: /* Convert the auditon() command to an event. */ ar->k_ar.ar_event = auditon_command_event(ar->k_ar.ar_arg_cmd); break; case AUE_MSGSYS: if (ARG_IS_VALID(ar, ARG_SVIPC_WHICH)) ar->k_ar.ar_event = audit_msgsys_to_event(ar->k_ar.ar_arg_svipc_which); break; case AUE_SEMSYS: if (ARG_IS_VALID(ar, ARG_SVIPC_WHICH)) ar->k_ar.ar_event = audit_semsys_to_event(ar->k_ar.ar_arg_svipc_which); break; case AUE_SHMSYS: if (ARG_IS_VALID(ar, ARG_SVIPC_WHICH)) ar->k_ar.ar_event = audit_shmsys_to_event(ar->k_ar.ar_arg_svipc_which); break; } auid = ar->k_ar.ar_subj_auid; event = ar->k_ar.ar_event; class = au_event_class(event); ar->k_ar_commit |= AR_COMMIT_KERNEL; if (au_preselect(event, class, aumask, sorf) != 0) ar->k_ar_commit |= AR_PRESELECT_TRAIL; if (audit_pipe_preselect(auid, event, class, sorf, ar->k_ar_commit & AR_PRESELECT_TRAIL) != 0) ar->k_ar_commit |= AR_PRESELECT_PIPE; #ifdef KDTRACE_HOOKS /* * Expose the audit record to DTrace, both to allow the "commit" probe * to fire if it's desirable, and also to allow a decision to be made * about later firing with BSM in the audit worker. */ if (dtaudit_hook_commit != NULL) { if (dtaudit_hook_commit(ar, auid, event, class, sorf) != 0) ar->k_ar_commit |= AR_PRESELECT_DTRACE; } #endif if ((ar->k_ar_commit & (AR_PRESELECT_TRAIL | AR_PRESELECT_PIPE | AR_PRESELECT_USER_TRAIL | AR_PRESELECT_USER_PIPE | AR_PRESELECT_DTRACE)) == 0) { mtx_lock(&audit_mtx); audit_pre_q_len--; mtx_unlock(&audit_mtx); audit_free(ar); return; } /* * Note: it could be that some records initiated while audit was * enabled should still be committed? */ mtx_lock(&audit_mtx); if (audit_suspended || !audit_enabled) { audit_pre_q_len--; mtx_unlock(&audit_mtx); audit_free(ar); return; } /* * Constrain the number of committed audit records based on the * configurable parameter. */ while (audit_q_len >= audit_qctrl.aq_hiwater) cv_wait(&audit_watermark_cv, &audit_mtx); TAILQ_INSERT_TAIL(&audit_q, ar, k_q); audit_q_len++; audit_pre_q_len--; cv_signal(&audit_worker_cv); mtx_unlock(&audit_mtx); } /* * audit_syscall_enter() is called on entry to each system call. It is * responsible for deciding whether or not to audit the call (preselection), * and if so, allocating a per-thread audit record. audit_new() will fill in * basic thread/credential properties. */ void audit_syscall_enter(unsigned short code, struct thread *td) { struct au_mask *aumask; #ifdef KDTRACE_HOOKS void *dtaudit_state; #endif au_class_t class; au_event_t event; au_id_t auid; int record_needed; KASSERT(td->td_ar == NULL, ("audit_syscall_enter: td->td_ar != NULL")); KASSERT((td->td_pflags & TDP_AUDITREC) == 0, ("audit_syscall_enter: TDP_AUDITREC set")); /* * In FreeBSD, each ABI has its own system call table, and hence * mapping of system call codes to audit events. Convert the code to * an audit event identifier using the process system call table * reference. In Darwin, there's only one, so we use the global * symbol for the system call table. No audit record is generated * for bad system calls, as no operation has been performed. */ if (code >= td->td_proc->p_sysent->sv_size) return; event = td->td_proc->p_sysent->sv_table[code].sy_auevent; if (event == AUE_NULL) return; /* * Check which audit mask to use; either the kernel non-attributable * event mask or the process audit mask. */ auid = td->td_ucred->cr_audit.ai_auid; if (auid == AU_DEFAUDITID) aumask = &audit_nae_mask; else aumask = &td->td_ucred->cr_audit.ai_mask; /* * Determine whether trail or pipe preselection would like an audit * record allocated for this system call. */ class = au_event_class(event); if (au_preselect(event, class, aumask, AU_PRS_BOTH)) { /* * If we're out of space and need to suspend unprivileged * processes, do that here rather than trying to allocate * another audit record. * * Note: we might wish to be able to continue here in the * future, if the system recovers. That should be possible * by means of checking the condition in a loop around * cv_wait(). It might be desirable to reevaluate whether an * audit record is still required for this event by * re-calling au_preselect(). */ if (audit_in_failure && priv_check(td, PRIV_AUDIT_FAILSTOP) != 0) { cv_wait(&audit_fail_cv, &audit_mtx); panic("audit_failing_stop: thread continued"); } record_needed = 1; } else if (audit_pipe_preselect(auid, event, class, AU_PRS_BOTH, 0)) { record_needed = 1; } else { record_needed = 0; } /* * After audit trails and pipes have made their policy choices, DTrace * may request that records be generated as well. This is a slightly * complex affair, as the DTrace audit provider needs the audit * framework to maintain some state on the audit record, which has not * been allocated at the point where the decision has to be made. * This hook must run even if we are not changing the decision, as * DTrace may want to stick event state onto a record we were going to * produce due to the trail or pipes. The event state returned by the * DTrace provider must be safe without locks held between here and * below -- i.e., dtaudit_state must must refer to stable memory. */ #ifdef KDTRACE_HOOKS dtaudit_state = NULL; if (dtaudit_hook_preselect != NULL) { dtaudit_state = dtaudit_hook_preselect(auid, event, class); if (dtaudit_state != NULL) record_needed = 1; } #endif /* * If a record is required, allocate it and attach it to the thread * for use throughout the system call. Also attach DTrace state if * required. * * XXXRW: If we decide to reference count the evname_elem underlying * dtaudit_state, we will need to free here if no record is allocated * or allocatable. */ if (record_needed) { td->td_ar = audit_new(event, td); if (td->td_ar != NULL) { td->td_pflags |= TDP_AUDITREC; #ifdef KDTRACE_HOOKS td->td_ar->k_dtaudit_state = dtaudit_state; #endif } } else td->td_ar = NULL; } /* * audit_syscall_exit() is called from the return of every system call, or in * the event of exit1(), during the execution of exit1(). It is responsible * for committing the audit record, if any, along with return condition. */ void audit_syscall_exit(int error, struct thread *td) { int retval; /* * Commit the audit record as desired; once we pass the record into * audit_commit(), the memory is owned by the audit subsystem. The * return value from the system call is stored on the user thread. * If there was an error, the return value is set to -1, imitating * the behavior of the cerror routine. */ if (error) retval = -1; else retval = td->td_retval[0]; audit_commit(td->td_ar, error, retval); td->td_ar = NULL; td->td_pflags &= ~TDP_AUDITREC; } void audit_cred_copy(struct ucred *src, struct ucred *dest) { bcopy(&src->cr_audit, &dest->cr_audit, sizeof(dest->cr_audit)); } void audit_cred_destroy(struct ucred *cred) { } void audit_cred_init(struct ucred *cred) { bzero(&cred->cr_audit, sizeof(cred->cr_audit)); } /* * Initialize audit information for the first kernel process (proc 0) and for * the first user process (init). */ void audit_cred_kproc0(struct ucred *cred) { cred->cr_audit.ai_auid = AU_DEFAUDITID; cred->cr_audit.ai_termid.at_type = AU_IPv4; } void audit_cred_proc1(struct ucred *cred) { cred->cr_audit.ai_auid = AU_DEFAUDITID; cred->cr_audit.ai_termid.at_type = AU_IPv4; } void audit_thread_alloc(struct thread *td) { td->td_ar = NULL; } void audit_thread_free(struct thread *td) { KASSERT(td->td_ar == NULL, ("audit_thread_free: td_ar != NULL")); KASSERT((td->td_pflags & TDP_AUDITREC) == 0, ("audit_thread_free: TDP_AUDITREC set")); } void audit_proc_coredump(struct thread *td, char *path, int errcode) { struct kaudit_record *ar; struct au_mask *aumask; struct ucred *cred; au_class_t class; int ret, sorf; char **pathp; au_id_t auid; ret = 0; /* * Make sure we are using the correct preselection mask. */ cred = td->td_ucred; auid = cred->cr_audit.ai_auid; if (auid == AU_DEFAUDITID) aumask = &audit_nae_mask; else aumask = &cred->cr_audit.ai_mask; /* * It's possible for coredump(9) generation to fail. Make sure that * we handle this case correctly for preselection. */ if (errcode != 0) sorf = AU_PRS_FAILURE; else sorf = AU_PRS_SUCCESS; class = au_event_class(AUE_CORE); if (au_preselect(AUE_CORE, class, aumask, sorf) == 0 && audit_pipe_preselect(auid, AUE_CORE, class, sorf, 0) == 0) return; /* * If we are interested in seeing this audit record, allocate it. * Where possible coredump records should contain a pathname and arg32 * (signal) tokens. */ ar = audit_new(AUE_CORE, td); if (ar == NULL) return; if (path != NULL) { pathp = &ar->k_ar.ar_arg_upath1; *pathp = malloc(MAXPATHLEN, M_AUDITPATH, M_WAITOK); audit_canon_path(td, AT_FDCWD, path, *pathp); ARG_SET_VALID(ar, ARG_UPATH1); } ar->k_ar.ar_arg_signum = td->td_proc->p_sig; ARG_SET_VALID(ar, ARG_SIGNUM); if (errcode != 0) ret = 1; audit_commit(ar, errcode, ret); }