Index: head/sys/kern/kern_lock.c =================================================================== --- head/sys/kern/kern_lock.c (revision 363414) +++ head/sys/kern/kern_lock.c (revision 363415) @@ -1,1760 +1,1853 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008 Attilio Rao * 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. */ #include "opt_ddb.h" #include "opt_hwpmc_hooks.h" #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #ifdef DEBUG_LOCKS #include #endif #include #include #include #ifdef DDB #include #endif #ifdef HWPMC_HOOKS #include PMC_SOFT_DECLARE( , , lock, failed); #endif CTASSERT(LK_UNLOCKED == (LK_UNLOCKED & ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS))); #define SQ_EXCLUSIVE_QUEUE 0 #define SQ_SHARED_QUEUE 1 #ifndef INVARIANTS #define _lockmgr_assert(lk, what, file, line) #endif #define TD_SLOCKS_INC(td) ((td)->td_lk_slocks++) #define TD_SLOCKS_DEC(td) ((td)->td_lk_slocks--) #ifndef DEBUG_LOCKS #define STACK_PRINT(lk) #define STACK_SAVE(lk) #define STACK_ZERO(lk) #else #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack) #define STACK_SAVE(lk) stack_save(&(lk)->lk_stack) #define STACK_ZERO(lk) stack_zero(&(lk)->lk_stack) #endif #define LOCK_LOG2(lk, string, arg1, arg2) \ if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \ CTR2(KTR_LOCK, (string), (arg1), (arg2)) #define LOCK_LOG3(lk, string, arg1, arg2, arg3) \ if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \ CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3)) #define GIANT_DECLARE \ int _i = 0; \ WITNESS_SAVE_DECL(Giant) #define GIANT_RESTORE() do { \ if (__predict_false(_i > 0)) { \ while (_i--) \ mtx_lock(&Giant); \ WITNESS_RESTORE(&Giant.lock_object, Giant); \ } \ } while (0) #define GIANT_SAVE() do { \ if (__predict_false(mtx_owned(&Giant))) { \ WITNESS_SAVE(&Giant.lock_object, Giant); \ while (mtx_owned(&Giant)) { \ _i++; \ mtx_unlock(&Giant); \ } \ } \ } while (0) static bool __always_inline LK_CAN_SHARE(uintptr_t x, int flags, bool fp) { if ((x & (LK_SHARE | LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == LK_SHARE) return (true); if (fp || (!(x & LK_SHARE))) return (false); if ((curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) || (curthread->td_pflags & TDP_DEADLKTREAT)) return (true); return (false); } #define LK_TRYOP(x) \ ((x) & LK_NOWAIT) #define LK_CAN_WITNESS(x) \ (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x)) #define LK_TRYWIT(x) \ (LK_TRYOP(x) ? LOP_TRYLOCK : 0) #define lockmgr_disowned(lk) \ (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC) #define lockmgr_xlocked_v(v) \ (((v) & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread) #define lockmgr_xlocked(lk) lockmgr_xlocked_v(lockmgr_read_value(lk)) static void assert_lockmgr(const struct lock_object *lock, int how); #ifdef DDB static void db_show_lockmgr(const struct lock_object *lock); #endif static void lock_lockmgr(struct lock_object *lock, uintptr_t how); #ifdef KDTRACE_HOOKS static int owner_lockmgr(const struct lock_object *lock, struct thread **owner); #endif static uintptr_t unlock_lockmgr(struct lock_object *lock); struct lock_class lock_class_lockmgr = { .lc_name = "lockmgr", .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE, .lc_assert = assert_lockmgr, #ifdef DDB .lc_ddb_show = db_show_lockmgr, #endif .lc_lock = lock_lockmgr, .lc_unlock = unlock_lockmgr, #ifdef KDTRACE_HOOKS .lc_owner = owner_lockmgr, #endif }; +static __read_mostly bool lk_adaptive = true; +static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging"); +SYSCTL_BOOL(_debug_lockmgr, OID_AUTO, adaptive_spinning, CTLFLAG_RW, &lk_adaptive, + 0, ""); +#define lockmgr_delay locks_delay + struct lockmgr_wait { const char *iwmesg; int ipri; int itimo; }; static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp); static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp); static void lockmgr_exit(u_int flags, struct lock_object *ilk, int wakeup_swapper) { struct lock_class *class; if (flags & LK_INTERLOCK) { class = LOCK_CLASS(ilk); class->lc_unlock(ilk); } if (__predict_false(wakeup_swapper)) kick_proc0(); } static void lockmgr_note_shared_acquire(struct lock *lk, int contested, uint64_t waittime, const char *file, int line, int flags) { LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested, waittime, file, line, LOCKSTAT_READER); LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line); WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line); TD_LOCKS_INC(curthread); TD_SLOCKS_INC(curthread); STACK_SAVE(lk); } static void lockmgr_note_shared_release(struct lock *lk, const char *file, int line) { WITNESS_UNLOCK(&lk->lock_object, 0, file, line); LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line); TD_LOCKS_DEC(curthread); TD_SLOCKS_DEC(curthread); } static void lockmgr_note_exclusive_acquire(struct lock *lk, int contested, uint64_t waittime, const char *file, int line, int flags) { LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested, waittime, file, line, LOCKSTAT_WRITER); LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line); WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file, line); TD_LOCKS_INC(curthread); STACK_SAVE(lk); } static void lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line) { if (LK_HOLDER(lockmgr_read_value(lk)) != LK_KERNPROC) { WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line); TD_LOCKS_DEC(curthread); } LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line); } static __inline struct thread * lockmgr_xholder(const struct lock *lk) { uintptr_t x; x = lockmgr_read_value(lk); return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x)); } /* * It assumes sleepq_lock held and returns with this one unheld. * It also assumes the generic interlock is sane and previously checked. * If LK_INTERLOCK is specified the interlock is not reacquired after the * sleep. */ static __inline int sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk, const char *wmesg, int pri, int timo, int queue) { GIANT_DECLARE; struct lock_class *class; int catch, error; class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL; catch = pri & PCATCH; pri &= PRIMASK; error = 0; LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk, (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared"); if (flags & LK_INTERLOCK) class->lc_unlock(ilk); if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0) lk->lk_exslpfail++; GIANT_SAVE(); sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ? SLEEPQ_INTERRUPTIBLE : 0), queue); if ((flags & LK_TIMELOCK) && timo) sleepq_set_timeout(&lk->lock_object, timo); /* * Decisional switch for real sleeping. */ if ((flags & LK_TIMELOCK) && timo && catch) error = sleepq_timedwait_sig(&lk->lock_object, pri); else if ((flags & LK_TIMELOCK) && timo) error = sleepq_timedwait(&lk->lock_object, pri); else if (catch) error = sleepq_wait_sig(&lk->lock_object, pri); else sleepq_wait(&lk->lock_object, pri); GIANT_RESTORE(); if ((flags & LK_SLEEPFAIL) && error == 0) error = ENOLCK; return (error); } static __inline int wakeupshlk(struct lock *lk, const char *file, int line) { uintptr_t v, x, orig_x; u_int realexslp; int queue, wakeup_swapper; wakeup_swapper = 0; for (;;) { x = lockmgr_read_value(lk); if (lockmgr_sunlock_try(lk, &x)) break; /* * We should have a sharer with waiters, so enter the hard * path in order to handle wakeups correctly. */ sleepq_lock(&lk->lock_object); orig_x = lockmgr_read_value(lk); retry_sleepq: x = orig_x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); v = LK_UNLOCKED; /* * If the lock has exclusive waiters, give them preference in * order to avoid deadlock with shared runners up. * If interruptible sleeps left the exclusive queue empty * avoid a starvation for the threads sleeping on the shared * queue by giving them precedence and cleaning up the * exclusive waiters bit anyway. * Please note that lk_exslpfail count may be lying about * the real number of waiters with the LK_SLEEPFAIL flag on * because they may be used in conjunction with interruptible * sleeps so lk_exslpfail might be considered an 'upper limit' * bound, including the edge cases. */ realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE); if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) { if (lk->lk_exslpfail < realexslp) { lk->lk_exslpfail = 0; queue = SQ_EXCLUSIVE_QUEUE; v |= (x & LK_SHARED_WAITERS); } else { lk->lk_exslpfail = 0; LOCK_LOG2(lk, "%s: %p has only LK_SLEEPFAIL sleepers", __func__, lk); LOCK_LOG2(lk, "%s: %p waking up threads on the exclusive queue", __func__, lk); wakeup_swapper = sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE); queue = SQ_SHARED_QUEUE; } } else { /* * Exclusive waiters sleeping with LK_SLEEPFAIL on * and using interruptible sleeps/timeout may have * left spourious lk_exslpfail counts on, so clean * it up anyway. */ lk->lk_exslpfail = 0; queue = SQ_SHARED_QUEUE; } if (lockmgr_sunlock_try(lk, &orig_x)) { sleepq_release(&lk->lock_object); break; } x |= LK_SHARERS_LOCK(1); if (!atomic_fcmpset_rel_ptr(&lk->lk_lock, &x, v)) { orig_x = x; goto retry_sleepq; } LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue", __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : "exclusive"); wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue); sleepq_release(&lk->lock_object); break; } LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER); return (wakeup_swapper); } static void assert_lockmgr(const struct lock_object *lock, int what) { panic("lockmgr locks do not support assertions"); } static void lock_lockmgr(struct lock_object *lock, uintptr_t how) { panic("lockmgr locks do not support sleep interlocking"); } static uintptr_t unlock_lockmgr(struct lock_object *lock) { panic("lockmgr locks do not support sleep interlocking"); } #ifdef KDTRACE_HOOKS static int owner_lockmgr(const struct lock_object *lock, struct thread **owner) { panic("lockmgr locks do not support owner inquiring"); } #endif void lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags) { int iflags; MPASS((flags & ~LK_INIT_MASK) == 0); ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock, ("%s: lockmgr not aligned for %s: %p", __func__, wmesg, &lk->lk_lock)); iflags = LO_SLEEPABLE | LO_UPGRADABLE; if (flags & LK_CANRECURSE) iflags |= LO_RECURSABLE; if ((flags & LK_NODUP) == 0) iflags |= LO_DUPOK; if (flags & LK_NOPROFILE) iflags |= LO_NOPROFILE; if ((flags & LK_NOWITNESS) == 0) iflags |= LO_WITNESS; if (flags & LK_QUIET) iflags |= LO_QUIET; if (flags & LK_IS_VNODE) iflags |= LO_IS_VNODE; if (flags & LK_NEW) iflags |= LO_NEW; iflags |= flags & LK_NOSHARE; lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags); lk->lk_lock = LK_UNLOCKED; lk->lk_recurse = 0; lk->lk_exslpfail = 0; lk->lk_timo = timo; lk->lk_pri = pri; STACK_ZERO(lk); } /* * XXX: Gross hacks to manipulate external lock flags after * initialization. Used for certain vnode and buf locks. */ void lockallowshare(struct lock *lk) { lockmgr_assert(lk, KA_XLOCKED); lk->lock_object.lo_flags &= ~LK_NOSHARE; } void lockdisableshare(struct lock *lk) { lockmgr_assert(lk, KA_XLOCKED); lk->lock_object.lo_flags |= LK_NOSHARE; } void lockallowrecurse(struct lock *lk) { lockmgr_assert(lk, KA_XLOCKED); lk->lock_object.lo_flags |= LO_RECURSABLE; } void lockdisablerecurse(struct lock *lk) { lockmgr_assert(lk, KA_XLOCKED); lk->lock_object.lo_flags &= ~LO_RECURSABLE; } void lockdestroy(struct lock *lk) { KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held")); KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed")); KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters")); lock_destroy(&lk->lock_object); } static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp) { /* * If no other thread has an exclusive lock, or * no exclusive waiter is present, bump the count of * sharers. Since we have to preserve the state of * waiters, if we fail to acquire the shared lock * loop back and retry. */ - *xp = lockmgr_read_value(lk); while (LK_CAN_SHARE(*xp, flags, fp)) { if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp, *xp + LK_ONE_SHARER)) { return (true); } } return (false); } static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp) { for (;;) { if (LK_SHARERS(*xp) > 1 || !(*xp & LK_ALL_WAITERS)) { if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp, *xp - LK_ONE_SHARER)) return (true); continue; } break; } return (false); } +static bool +lockmgr_slock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp, + int flags) +{ + struct thread *owner; + uintptr_t x; + + x = *xp; + MPASS(x != LK_UNLOCKED); + owner = (struct thread *)LK_HOLDER(x); + for (;;) { + MPASS(owner != curthread); + if (owner == (struct thread *)LK_KERNPROC) + return (false); + if ((x & LK_SHARE) && LK_SHARERS(x) > 0) + return (false); + if (owner == NULL) + return (false); + if (!TD_IS_RUNNING(owner)) + return (false); + if ((x & LK_ALL_WAITERS) != 0) + return (false); + lock_delay(lda); + x = lockmgr_read_value(lk); + if (LK_CAN_SHARE(x, flags, false)) { + *xp = x; + return (true); + } + owner = (struct thread *)LK_HOLDER(x); + } +} + static __noinline int lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk, const char *file, int line, struct lockmgr_wait *lwa) { uintptr_t tid, x; int error = 0; const char *iwmesg; int ipri, itimo; #ifdef KDTRACE_HOOKS uint64_t sleep_time = 0; #endif #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif + struct lock_delay_arg lda; if (KERNEL_PANICKED()) goto out; tid = (uintptr_t)curthread; if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line, flags & LK_INTERLOCK ? ilk : NULL); + lock_delay_arg_init(&lda, &lockmgr_delay); + if (!lk_adaptive) + flags &= ~LK_ADAPTIVE; + x = lockmgr_read_value(lk); + /* + * The lock may already be locked exclusive by curthread, + * avoid deadlock. + */ + if (LK_HOLDER(x) == tid) { + LOCK_LOG2(lk, + "%s: %p already held in exclusive mode", + __func__, lk); + error = EDEADLK; + goto out; + } + for (;;) { if (lockmgr_slock_try(lk, &x, flags, false)) break; + + if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) { + if (lockmgr_slock_adaptive(&lda, lk, &x, flags)) + continue; + } + #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&lk->lock_object, &contested, &waittime); /* - * If the lock is already held by curthread in - * exclusive way avoid a deadlock. - */ - if (LK_HOLDER(x) == tid) { - LOCK_LOG2(lk, - "%s: %p already held in exclusive mode", - __func__, lk); - error = EDEADLK; - break; - } - - /* * If the lock is expected to not sleep just give up * and return. */ if (LK_TRYOP(flags)) { LOCK_LOG2(lk, "%s: %p fails the try operation", __func__, lk); error = EBUSY; break; } /* * Acquire the sleepqueue chain lock because we * probabilly will need to manipulate waiters flags. */ sleepq_lock(&lk->lock_object); x = lockmgr_read_value(lk); retry_sleepq: /* * if the lock can be acquired in shared mode, try * again. */ if (LK_CAN_SHARE(x, flags, false)) { sleepq_release(&lk->lock_object); continue; } /* * Try to set the LK_SHARED_WAITERS flag. If we fail, * loop back and retry. */ if ((x & LK_SHARED_WAITERS) == 0) { if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, x | LK_SHARED_WAITERS)) { goto retry_sleepq; } LOCK_LOG2(lk, "%s: %p set shared waiters flag", __func__, lk); } if (lwa == NULL) { iwmesg = lk->lock_object.lo_name; ipri = lk->lk_pri; itimo = lk->lk_timo; } else { iwmesg = lwa->iwmesg; ipri = lwa->ipri; itimo = lwa->itimo; } /* * As far as we have been unable to acquire the * shared lock and the shared waiters flag is set, * we will sleep. */ #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&lk->lock_object); #endif error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo, SQ_SHARED_QUEUE); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&lk->lock_object); #endif flags &= ~LK_INTERLOCK; if (error) { LOCK_LOG3(lk, "%s: interrupted sleep for %p with %d", __func__, lk, error); break; } LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", __func__, lk); + x = lockmgr_read_value(lk); } if (error == 0) { #ifdef KDTRACE_HOOKS if (sleep_time != 0) LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time, LOCKSTAT_READER, (x & LK_SHARE) == 0, (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x)); #endif #ifdef LOCK_PROFILING lockmgr_note_shared_acquire(lk, contested, waittime, file, line, flags); #else lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags); #endif } out: lockmgr_exit(flags, ilk, 0); return (error); } +static bool +lockmgr_xlock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp) +{ + struct thread *owner; + uintptr_t x; + + x = *xp; + MPASS(x != LK_UNLOCKED); + owner = (struct thread *)LK_HOLDER(x); + for (;;) { + MPASS(owner != curthread); + if (owner == NULL) + return (false); + if ((x & LK_SHARE) && LK_SHARERS(x) > 0) + return (false); + if (owner == (struct thread *)LK_KERNPROC) + return (false); + if (!TD_IS_RUNNING(owner)) + return (false); + if ((x & LK_ALL_WAITERS) != 0) + return (false); + lock_delay(lda); + x = lockmgr_read_value(lk); + if (x == LK_UNLOCKED) { + *xp = x; + return (true); + } + owner = (struct thread *)LK_HOLDER(x); + } +} + static __noinline int lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk, const char *file, int line, struct lockmgr_wait *lwa) { struct lock_class *class; uintptr_t tid, x, v; int error = 0; const char *iwmesg; int ipri, itimo; #ifdef KDTRACE_HOOKS uint64_t sleep_time = 0; #endif #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif + struct lock_delay_arg lda; if (KERNEL_PANICKED()) goto out; tid = (uintptr_t)curthread; if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? ilk : NULL); /* * If curthread already holds the lock and this one is * allowed to recurse, simply recurse on it. */ if (lockmgr_xlocked(lk)) { if ((flags & LK_CANRECURSE) == 0 && (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) { /* * If the lock is expected to not panic just * give up and return. */ if (LK_TRYOP(flags)) { LOCK_LOG2(lk, "%s: %p fails the try operation", __func__, lk); error = EBUSY; goto out; } if (flags & LK_INTERLOCK) { class = LOCK_CLASS(ilk); class->lc_unlock(ilk); } STACK_PRINT(lk); panic("%s: recursing on non recursive lockmgr %p " "@ %s:%d\n", __func__, lk, file, line); } atomic_set_ptr(&lk->lk_lock, LK_WRITER_RECURSED); lk->lk_recurse++; LOCK_LOG2(lk, "%s: %p recursing", __func__, lk); LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line); WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file, line); TD_LOCKS_INC(curthread); goto out; } + x = LK_UNLOCKED; + lock_delay_arg_init(&lda, &lockmgr_delay); + if (!lk_adaptive) + flags &= ~LK_ADAPTIVE; for (;;) { - if (lk->lk_lock == LK_UNLOCKED && - atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) - break; + if (x == LK_UNLOCKED) { + if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, tid)) + break; + } + if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) { + if (lockmgr_xlock_adaptive(&lda, lk, &x)) + continue; + } #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&lk->lock_object, &contested, &waittime); /* * If the lock is expected to not sleep just give up * and return. */ if (LK_TRYOP(flags)) { LOCK_LOG2(lk, "%s: %p fails the try operation", __func__, lk); error = EBUSY; break; } /* * Acquire the sleepqueue chain lock because we * probabilly will need to manipulate waiters flags. */ sleepq_lock(&lk->lock_object); x = lockmgr_read_value(lk); retry_sleepq: /* * if the lock has been released while we spun on * the sleepqueue chain lock just try again. */ if (x == LK_UNLOCKED) { sleepq_release(&lk->lock_object); continue; } /* * The lock can be in the state where there is a * pending queue of waiters, but still no owner. * This happens when the lock is contested and an * owner is going to claim the lock. * If curthread is the one successfully acquiring it * claim lock ownership and return, preserving waiters * flags. */ v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); if ((x & ~v) == LK_UNLOCKED) { v &= ~LK_EXCLUSIVE_SPINNERS; if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, tid | v)) { sleepq_release(&lk->lock_object); LOCK_LOG2(lk, "%s: %p claimed by a new writer", __func__, lk); break; } goto retry_sleepq; } /* * Try to set the LK_EXCLUSIVE_WAITERS flag. If we * fail, loop back and retry. */ if ((x & LK_EXCLUSIVE_WAITERS) == 0) { if (!atomic_fcmpset_ptr(&lk->lk_lock, &x, x | LK_EXCLUSIVE_WAITERS)) { goto retry_sleepq; } LOCK_LOG2(lk, "%s: %p set excl waiters flag", __func__, lk); } if (lwa == NULL) { iwmesg = lk->lock_object.lo_name; ipri = lk->lk_pri; itimo = lk->lk_timo; } else { iwmesg = lwa->iwmesg; ipri = lwa->ipri; itimo = lwa->itimo; } /* * As far as we have been unable to acquire the * exclusive lock and the exclusive waiters flag * is set, we will sleep. */ #ifdef KDTRACE_HOOKS sleep_time -= lockstat_nsecs(&lk->lock_object); #endif error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo, SQ_EXCLUSIVE_QUEUE); #ifdef KDTRACE_HOOKS sleep_time += lockstat_nsecs(&lk->lock_object); #endif flags &= ~LK_INTERLOCK; if (error) { LOCK_LOG3(lk, "%s: interrupted sleep for %p with %d", __func__, lk, error); break; } LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", __func__, lk); + x = lockmgr_read_value(lk); } if (error == 0) { #ifdef KDTRACE_HOOKS if (sleep_time != 0) LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time, LOCKSTAT_WRITER, (x & LK_SHARE) == 0, (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x)); #endif #ifdef LOCK_PROFILING lockmgr_note_exclusive_acquire(lk, contested, waittime, file, line, flags); #else lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, flags); #endif } out: lockmgr_exit(flags, ilk, 0); return (error); } static __noinline int lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk, const char *file, int line, struct lockmgr_wait *lwa) { uintptr_t tid, v, setv; int error = 0; int op; if (KERNEL_PANICKED()) goto out; tid = (uintptr_t)curthread; _lockmgr_assert(lk, KA_SLOCKED, file, line); op = flags & LK_TYPE_MASK; v = lockmgr_read_value(lk); for (;;) { if (LK_SHARERS_LOCK(v) > 1) { if (op == LK_TRYUPGRADE) { LOCK_LOG2(lk, "%s: %p failed the nowait upgrade", __func__, lk); error = EBUSY; goto out; } if (lockmgr_sunlock_try(lk, &v)) { lockmgr_note_shared_release(lk, file, line); goto out_xlock; } } MPASS((v & ~LK_ALL_WAITERS) == LK_SHARERS_LOCK(1)); setv = tid; setv |= (v & LK_ALL_WAITERS); /* * Try to switch from one shared lock to an exclusive one. * We need to preserve waiters flags during the operation. */ if (atomic_fcmpset_ptr(&lk->lk_lock, &v, setv)) { LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file, line); WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file, line); LOCKSTAT_RECORD0(lockmgr__upgrade, lk); TD_SLOCKS_DEC(curthread); goto out; } } out_xlock: error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa); flags &= ~LK_INTERLOCK; out: lockmgr_exit(flags, ilk, 0); return (error); } int lockmgr_lock_flags(struct lock *lk, u_int flags, struct lock_object *ilk, const char *file, int line) { struct lock_class *class; uintptr_t x, tid; u_int op; bool locked; if (KERNEL_PANICKED()) return (0); op = flags & LK_TYPE_MASK; locked = false; switch (op) { case LK_SHARED: if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line, flags & LK_INTERLOCK ? ilk : NULL); if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE)) break; + x = lockmgr_read_value(lk); if (lockmgr_slock_try(lk, &x, flags, true)) { lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags); locked = true; } else { return (lockmgr_slock_hard(lk, flags, ilk, file, line, NULL)); } break; case LK_EXCLUSIVE: if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? ilk : NULL); tid = (uintptr_t)curthread; if (lockmgr_read_value(lk) == LK_UNLOCKED && atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) { lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, flags); locked = true; } else { return (lockmgr_xlock_hard(lk, flags, ilk, file, line, NULL)); } break; case LK_UPGRADE: case LK_TRYUPGRADE: return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL)); default: break; } if (__predict_true(locked)) { if (__predict_false(flags & LK_INTERLOCK)) { class = LOCK_CLASS(ilk); class->lc_unlock(ilk); } return (0); } else { return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT, LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line)); } } static __noinline int lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk, const char *file, int line) { int wakeup_swapper = 0; if (KERNEL_PANICKED()) goto out; wakeup_swapper = wakeupshlk(lk, file, line); out: lockmgr_exit(flags, ilk, wakeup_swapper); return (0); } static __noinline int lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk, const char *file, int line) { uintptr_t tid, v; int wakeup_swapper = 0; u_int realexslp; int queue; if (KERNEL_PANICKED()) goto out; tid = (uintptr_t)curthread; /* * As first option, treact the lock as if it has not * any waiter. * Fix-up the tid var if the lock has been disowned. */ if (LK_HOLDER(x) == LK_KERNPROC) tid = LK_KERNPROC; /* * The lock is held in exclusive mode. * If the lock is recursed also, then unrecurse it. */ if (lockmgr_recursed_v(x)) { LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk); lk->lk_recurse--; if (lk->lk_recurse == 0) atomic_clear_ptr(&lk->lk_lock, LK_WRITER_RECURSED); goto out; } if (tid != LK_KERNPROC) LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_WRITER); if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) goto out; sleepq_lock(&lk->lock_object); x = lockmgr_read_value(lk); v = LK_UNLOCKED; /* * If the lock has exclusive waiters, give them * preference in order to avoid deadlock with * shared runners up. * If interruptible sleeps left the exclusive queue * empty avoid a starvation for the threads sleeping * on the shared queue by giving them precedence * and cleaning up the exclusive waiters bit anyway. * Please note that lk_exslpfail count may be lying * about the real number of waiters with the * LK_SLEEPFAIL flag on because they may be used in * conjunction with interruptible sleeps so * lk_exslpfail might be considered an 'upper limit' * bound, including the edge cases. */ MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE); if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) { if (lk->lk_exslpfail < realexslp) { lk->lk_exslpfail = 0; queue = SQ_EXCLUSIVE_QUEUE; v |= (x & LK_SHARED_WAITERS); } else { lk->lk_exslpfail = 0; LOCK_LOG2(lk, "%s: %p has only LK_SLEEPFAIL sleepers", __func__, lk); LOCK_LOG2(lk, "%s: %p waking up threads on the exclusive queue", __func__, lk); wakeup_swapper = sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE); queue = SQ_SHARED_QUEUE; } } else { /* * Exclusive waiters sleeping with LK_SLEEPFAIL * on and using interruptible sleeps/timeout * may have left spourious lk_exslpfail counts * on, so clean it up anyway. */ lk->lk_exslpfail = 0; queue = SQ_SHARED_QUEUE; } LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue", __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : "exclusive"); atomic_store_rel_ptr(&lk->lk_lock, v); wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue); sleepq_release(&lk->lock_object); out: lockmgr_exit(flags, ilk, wakeup_swapper); return (0); } /* * Lightweight entry points for common operations. * * Functionality is similar to sx locks, in that none of the additional lockmgr * features are supported. To be clear, these are NOT supported: * 1. shared locking disablement * 2. returning with an error after sleep * 3. unlocking the interlock * * If in doubt, use lockmgr_lock_flags. */ int lockmgr_slock(struct lock *lk, u_int flags, const char *file, int line) { uintptr_t x; MPASS((flags & LK_TYPE_MASK) == LK_SHARED); MPASS((flags & LK_INTERLOCK) == 0); MPASS((lk->lock_object.lo_flags & LK_NOSHARE) == 0); if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, file, line, NULL); + x = lockmgr_read_value(lk); if (__predict_true(lockmgr_slock_try(lk, &x, flags, true))) { lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags); return (0); } - return (lockmgr_slock_hard(lk, flags, NULL, file, line, NULL)); + return (lockmgr_slock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL)); } int lockmgr_xlock(struct lock *lk, u_int flags, const char *file, int line) { uintptr_t tid; MPASS((flags & LK_TYPE_MASK) == LK_EXCLUSIVE); MPASS((flags & LK_INTERLOCK) == 0); if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); tid = (uintptr_t)curthread; if (atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) { lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, flags); return (0); } - return (lockmgr_xlock_hard(lk, flags, NULL, file, line, NULL)); + return (lockmgr_xlock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL)); } int lockmgr_unlock(struct lock *lk) { uintptr_t x, tid; const char *file; int line; file = __FILE__; line = __LINE__; _lockmgr_assert(lk, KA_LOCKED, file, line); x = lockmgr_read_value(lk); if (__predict_true(x & LK_SHARE) != 0) { lockmgr_note_shared_release(lk, file, line); if (lockmgr_sunlock_try(lk, &x)) { LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER); } else { return (lockmgr_sunlock_hard(lk, x, LK_RELEASE, NULL, file, line)); } } else { tid = (uintptr_t)curthread; lockmgr_note_exclusive_release(lk, file, line); if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) { LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk,LOCKSTAT_WRITER); } else { return (lockmgr_xunlock_hard(lk, x, LK_RELEASE, NULL, file, line)); } } return (0); } int __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk, const char *wmesg, int pri, int timo, const char *file, int line) { GIANT_DECLARE; struct lockmgr_wait lwa; struct lock_class *class; const char *iwmesg; uintptr_t tid, v, x; u_int op, realexslp; int error, ipri, itimo, queue, wakeup_swapper; #ifdef LOCK_PROFILING uint64_t waittime = 0; int contested = 0; #endif if (KERNEL_PANICKED()) return (0); error = 0; tid = (uintptr_t)curthread; op = (flags & LK_TYPE_MASK); iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg; ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri; itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo; lwa.iwmesg = iwmesg; lwa.ipri = ipri; lwa.itimo = itimo; MPASS((flags & ~LK_TOTAL_MASK) == 0); KASSERT((op & (op - 1)) == 0, ("%s: Invalid requested operation @ %s:%d", __func__, file, line)); KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 || (op != LK_DOWNGRADE && op != LK_RELEASE), ("%s: Invalid flags in regard of the operation desired @ %s:%d", __func__, file, line)); KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL, ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d", __func__, file, line)); KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread, lk->lock_object.lo_name, file, line)); class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL; if (lk->lock_object.lo_flags & LK_NOSHARE) { switch (op) { case LK_SHARED: op = LK_EXCLUSIVE; break; case LK_UPGRADE: case LK_TRYUPGRADE: case LK_DOWNGRADE: _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, file, line); if (flags & LK_INTERLOCK) class->lc_unlock(ilk); return (0); } } wakeup_swapper = 0; switch (op) { case LK_SHARED: return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa)); break; case LK_UPGRADE: case LK_TRYUPGRADE: return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa)); break; case LK_EXCLUSIVE: return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa)); break; case LK_DOWNGRADE: _lockmgr_assert(lk, KA_XLOCKED, file, line); WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line); /* * Panic if the lock is recursed. */ if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) { if (flags & LK_INTERLOCK) class->lc_unlock(ilk); panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n", __func__, iwmesg, file, line); } TD_SLOCKS_INC(curthread); /* * In order to preserve waiters flags, just spin. */ for (;;) { x = lockmgr_read_value(lk); MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); x &= LK_ALL_WAITERS; if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x, LK_SHARERS_LOCK(1) | x)) break; cpu_spinwait(); } LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line); LOCKSTAT_RECORD0(lockmgr__downgrade, lk); break; case LK_RELEASE: _lockmgr_assert(lk, KA_LOCKED, file, line); x = lockmgr_read_value(lk); if (__predict_true(x & LK_SHARE) != 0) { lockmgr_note_shared_release(lk, file, line); return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line)); } else { lockmgr_note_exclusive_release(lk, file, line); return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line)); } break; case LK_DRAIN: if (LK_CAN_WITNESS(flags)) WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? ilk : NULL); /* * Trying to drain a lock we already own will result in a * deadlock. */ if (lockmgr_xlocked(lk)) { if (flags & LK_INTERLOCK) class->lc_unlock(ilk); panic("%s: draining %s with the lock held @ %s:%d\n", __func__, iwmesg, file, line); } for (;;) { if (lk->lk_lock == LK_UNLOCKED && atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) break; #ifdef HWPMC_HOOKS PMC_SOFT_CALL( , , lock, failed); #endif lock_profile_obtain_lock_failed(&lk->lock_object, &contested, &waittime); /* * If the lock is expected to not sleep just give up * and return. */ if (LK_TRYOP(flags)) { LOCK_LOG2(lk, "%s: %p fails the try operation", __func__, lk); error = EBUSY; break; } /* * Acquire the sleepqueue chain lock because we * probabilly will need to manipulate waiters flags. */ sleepq_lock(&lk->lock_object); x = lockmgr_read_value(lk); /* * if the lock has been released while we spun on * the sleepqueue chain lock just try again. */ if (x == LK_UNLOCKED) { sleepq_release(&lk->lock_object); continue; } v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); if ((x & ~v) == LK_UNLOCKED) { v = (x & ~LK_EXCLUSIVE_SPINNERS); /* * If interruptible sleeps left the exclusive * queue empty avoid a starvation for the * threads sleeping on the shared queue by * giving them precedence and cleaning up the * exclusive waiters bit anyway. * Please note that lk_exslpfail count may be * lying about the real number of waiters with * the LK_SLEEPFAIL flag on because they may * be used in conjunction with interruptible * sleeps so lk_exslpfail might be considered * an 'upper limit' bound, including the edge * cases. */ if (v & LK_EXCLUSIVE_WAITERS) { queue = SQ_EXCLUSIVE_QUEUE; v &= ~LK_EXCLUSIVE_WAITERS; } else { /* * Exclusive waiters sleeping with * LK_SLEEPFAIL on and using * interruptible sleeps/timeout may * have left spourious lk_exslpfail * counts on, so clean it up anyway. */ MPASS(v & LK_SHARED_WAITERS); lk->lk_exslpfail = 0; queue = SQ_SHARED_QUEUE; v &= ~LK_SHARED_WAITERS; } if (queue == SQ_EXCLUSIVE_QUEUE) { realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE); if (lk->lk_exslpfail >= realexslp) { lk->lk_exslpfail = 0; queue = SQ_SHARED_QUEUE; v &= ~LK_SHARED_WAITERS; if (realexslp != 0) { LOCK_LOG2(lk, "%s: %p has only LK_SLEEPFAIL sleepers", __func__, lk); LOCK_LOG2(lk, "%s: %p waking up threads on the exclusive queue", __func__, lk); wakeup_swapper = sleepq_broadcast( &lk->lock_object, SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE); } } else lk->lk_exslpfail = 0; } if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) { sleepq_release(&lk->lock_object); continue; } LOCK_LOG3(lk, "%s: %p waking up all threads on the %s queue", __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : "exclusive"); wakeup_swapper |= sleepq_broadcast( &lk->lock_object, SLEEPQ_LK, 0, queue); /* * If shared waiters have been woken up we need * to wait for one of them to acquire the lock * before to set the exclusive waiters in * order to avoid a deadlock. */ if (queue == SQ_SHARED_QUEUE) { for (v = lk->lk_lock; (v & LK_SHARE) && !LK_SHARERS(v); v = lk->lk_lock) cpu_spinwait(); } } /* * Try to set the LK_EXCLUSIVE_WAITERS flag. If we * fail, loop back and retry. */ if ((x & LK_EXCLUSIVE_WAITERS) == 0) { if (!atomic_cmpset_ptr(&lk->lk_lock, x, x | LK_EXCLUSIVE_WAITERS)) { sleepq_release(&lk->lock_object); continue; } LOCK_LOG2(lk, "%s: %p set drain waiters flag", __func__, lk); } /* * As far as we have been unable to acquire the * exclusive lock and the exclusive waiters flag * is set, we will sleep. */ if (flags & LK_INTERLOCK) { class->lc_unlock(ilk); flags &= ~LK_INTERLOCK; } GIANT_SAVE(); sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK, SQ_EXCLUSIVE_QUEUE); sleepq_wait(&lk->lock_object, ipri & PRIMASK); GIANT_RESTORE(); LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", __func__, lk); } if (error == 0) { lock_profile_obtain_lock_success(&lk->lock_object, contested, waittime, file, line); LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0, lk->lk_recurse, file, line); WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file, line); TD_LOCKS_INC(curthread); STACK_SAVE(lk); } break; default: if (flags & LK_INTERLOCK) class->lc_unlock(ilk); panic("%s: unknown lockmgr request 0x%x\n", __func__, op); } if (flags & LK_INTERLOCK) class->lc_unlock(ilk); if (wakeup_swapper) kick_proc0(); return (error); } void _lockmgr_disown(struct lock *lk, const char *file, int line) { uintptr_t tid, x; if (SCHEDULER_STOPPED()) return; tid = (uintptr_t)curthread; _lockmgr_assert(lk, KA_XLOCKED, file, line); /* * Panic if the lock is recursed. */ if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) panic("%s: disown a recursed lockmgr @ %s:%d\n", __func__, file, line); /* * If the owner is already LK_KERNPROC just skip the whole operation. */ if (LK_HOLDER(lk->lk_lock) != tid) return; lock_profile_release_lock(&lk->lock_object); LOCKSTAT_RECORD1(lockmgr__disown, lk, LOCKSTAT_WRITER); LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line); WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line); TD_LOCKS_DEC(curthread); STACK_SAVE(lk); /* * In order to preserve waiters flags, just spin. */ for (;;) { x = lockmgr_read_value(lk); MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); x &= LK_ALL_WAITERS; if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x, LK_KERNPROC | x)) return; cpu_spinwait(); } } void lockmgr_printinfo(const struct lock *lk) { struct thread *td; uintptr_t x; if (lk->lk_lock == LK_UNLOCKED) printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name); else if (lk->lk_lock & LK_SHARE) printf("lock type %s: SHARED (count %ju)\n", lk->lock_object.lo_name, (uintmax_t)LK_SHARERS(lk->lk_lock)); else { td = lockmgr_xholder(lk); if (td == (struct thread *)LK_KERNPROC) printf("lock type %s: EXCL by KERNPROC\n", lk->lock_object.lo_name); else printf("lock type %s: EXCL by thread %p " "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, td, td->td_proc->p_pid, td->td_proc->p_comm, td->td_tid); } x = lk->lk_lock; if (x & LK_EXCLUSIVE_WAITERS) printf(" with exclusive waiters pending\n"); if (x & LK_SHARED_WAITERS) printf(" with shared waiters pending\n"); if (x & LK_EXCLUSIVE_SPINNERS) printf(" with exclusive spinners pending\n"); STACK_PRINT(lk); } int lockstatus(const struct lock *lk) { uintptr_t v, x; int ret; ret = LK_SHARED; x = lockmgr_read_value(lk); v = LK_HOLDER(x); if ((x & LK_SHARE) == 0) { if (v == (uintptr_t)curthread || v == LK_KERNPROC) ret = LK_EXCLUSIVE; else ret = LK_EXCLOTHER; } else if (x == LK_UNLOCKED) ret = 0; return (ret); } #ifdef INVARIANT_SUPPORT FEATURE(invariant_support, "Support for modules compiled with INVARIANTS option"); #ifndef INVARIANTS #undef _lockmgr_assert #endif void _lockmgr_assert(const struct lock *lk, int what, const char *file, int line) { int slocked = 0; if (KERNEL_PANICKED()) return; switch (what) { case KA_SLOCKED: case KA_SLOCKED | KA_NOTRECURSED: case KA_SLOCKED | KA_RECURSED: slocked = 1; case KA_LOCKED: case KA_LOCKED | KA_NOTRECURSED: case KA_LOCKED | KA_RECURSED: #ifdef WITNESS /* * We cannot trust WITNESS if the lock is held in exclusive * mode and a call to lockmgr_disown() happened. * Workaround this skipping the check if the lock is held in * exclusive mode even for the KA_LOCKED case. */ if (slocked || (lk->lk_lock & LK_SHARE)) { witness_assert(&lk->lock_object, what, file, line); break; } #endif if (lk->lk_lock == LK_UNLOCKED || ((lk->lk_lock & LK_SHARE) == 0 && (slocked || (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))))) panic("Lock %s not %slocked @ %s:%d\n", lk->lock_object.lo_name, slocked ? "share" : "", file, line); if ((lk->lk_lock & LK_SHARE) == 0) { if (lockmgr_recursed(lk)) { if (what & KA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", lk->lock_object.lo_name, file, line); } else if (what & KA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", lk->lock_object.lo_name, file, line); } break; case KA_XLOCKED: case KA_XLOCKED | KA_NOTRECURSED: case KA_XLOCKED | KA_RECURSED: if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)) panic("Lock %s not exclusively locked @ %s:%d\n", lk->lock_object.lo_name, file, line); if (lockmgr_recursed(lk)) { if (what & KA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", lk->lock_object.lo_name, file, line); } else if (what & KA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", lk->lock_object.lo_name, file, line); break; case KA_UNLOCKED: if (lockmgr_xlocked(lk) || lockmgr_disowned(lk)) panic("Lock %s exclusively locked @ %s:%d\n", lk->lock_object.lo_name, file, line); break; default: panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file, line); } } #endif #ifdef DDB int lockmgr_chain(struct thread *td, struct thread **ownerp) { const struct lock *lk; lk = td->td_wchan; if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr) return (0); db_printf("blocked on lockmgr %s", lk->lock_object.lo_name); if (lk->lk_lock & LK_SHARE) db_printf("SHARED (count %ju)\n", (uintmax_t)LK_SHARERS(lk->lk_lock)); else db_printf("EXCL\n"); *ownerp = lockmgr_xholder(lk); return (1); } static void db_show_lockmgr(const struct lock_object *lock) { struct thread *td; const struct lock *lk; lk = (const struct lock *)lock; db_printf(" state: "); if (lk->lk_lock == LK_UNLOCKED) db_printf("UNLOCKED\n"); else if (lk->lk_lock & LK_SHARE) db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock)); else { td = lockmgr_xholder(lk); if (td == (struct thread *)LK_KERNPROC) db_printf("XLOCK: LK_KERNPROC\n"); else db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td, td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm); if (lockmgr_recursed(lk)) db_printf(" recursed: %d\n", lk->lk_recurse); } db_printf(" waiters: "); switch (lk->lk_lock & LK_ALL_WAITERS) { case LK_SHARED_WAITERS: db_printf("shared\n"); break; case LK_EXCLUSIVE_WAITERS: db_printf("exclusive\n"); break; case LK_ALL_WAITERS: db_printf("shared and exclusive\n"); break; default: db_printf("none\n"); } db_printf(" spinners: "); if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS) db_printf("exclusive\n"); else db_printf("none\n"); } #endif Index: head/sys/sys/lockmgr.h =================================================================== --- head/sys/sys/lockmgr.h (revision 363414) +++ head/sys/sys/lockmgr.h (revision 363415) @@ -1,210 +1,211 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008 Attilio Rao * 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. * * $FreeBSD$ */ #ifndef _SYS_LOCKMGR_H_ #define _SYS_LOCKMGR_H_ #include #include #include #include #define LK_SHARE 0x01 #define LK_SHARED_WAITERS 0x02 #define LK_EXCLUSIVE_WAITERS 0x04 #define LK_EXCLUSIVE_SPINNERS 0x08 #define LK_WRITER_RECURSED 0x10 #define LK_ALL_WAITERS \ (LK_SHARED_WAITERS | LK_EXCLUSIVE_WAITERS) #define LK_FLAGMASK \ (LK_SHARE | LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS | LK_WRITER_RECURSED) #define LK_HOLDER(x) ((x) & ~LK_FLAGMASK) #define LK_SHARERS_SHIFT 5 #define LK_SHARERS(x) (LK_HOLDER(x) >> LK_SHARERS_SHIFT) #define LK_SHARERS_LOCK(x) ((x) << LK_SHARERS_SHIFT | LK_SHARE) #define LK_ONE_SHARER (1 << LK_SHARERS_SHIFT) #define LK_UNLOCKED LK_SHARERS_LOCK(0) #define LK_KERNPROC ((uintptr_t)(-1) & ~LK_FLAGMASK) #ifdef _KERNEL #if !defined(LOCK_FILE) || !defined(LOCK_LINE) #error "LOCK_FILE and LOCK_LINE not defined, include before" #endif struct thread; #define lk_recurse lock_object.lo_data /* * Function prototipes. Routines that start with an underscore are not part * of the public interface and might be wrappered with a macro. */ int __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk, const char *wmesg, int prio, int timo, const char *file, int line); int lockmgr_lock_flags(struct lock *lk, u_int flags, struct lock_object *ilk, const char *file, int line); int lockmgr_slock(struct lock *lk, u_int flags, const char *file, int line); int lockmgr_xlock(struct lock *lk, u_int flags, const char *file, int line); int lockmgr_unlock(struct lock *lk); #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) void _lockmgr_assert(const struct lock *lk, int what, const char *file, int line); #endif void _lockmgr_disown(struct lock *lk, const char *file, int line); void lockallowrecurse(struct lock *lk); void lockallowshare(struct lock *lk); void lockdestroy(struct lock *lk); void lockdisablerecurse(struct lock *lk); void lockdisableshare(struct lock *lk); void lockinit(struct lock *lk, int prio, const char *wmesg, int timo, int flags); #ifdef DDB int lockmgr_chain(struct thread *td, struct thread **ownerp); #endif void lockmgr_printinfo(const struct lock *lk); int lockstatus(const struct lock *lk); /* * As far as the ilk can be a static NULL pointer these functions need a * strict prototype in order to safely use the lock_object member. */ static __inline int _lockmgr_args(struct lock *lk, u_int flags, struct mtx *ilk, const char *wmesg, int prio, int timo, const char *file, int line) { return (__lockmgr_args(lk, flags, (ilk != NULL) ? &ilk->lock_object : NULL, wmesg, prio, timo, file, line)); } static __inline int _lockmgr_args_rw(struct lock *lk, u_int flags, struct rwlock *ilk, const char *wmesg, int prio, int timo, const char *file, int line) { return (__lockmgr_args(lk, flags, (ilk != NULL) ? &ilk->lock_object : NULL, wmesg, prio, timo, file, line)); } /* * Define aliases in order to complete lockmgr KPI. */ #define lockmgr_read_value(lk) ((lk)->lk_lock) #define lockmgr(lk, flags, ilk) \ _lockmgr_args((lk), (flags), (ilk), LK_WMESG_DEFAULT, \ LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, LOCK_FILE, LOCK_LINE) #define lockmgr_args(lk, flags, ilk, wmesg, prio, timo) \ _lockmgr_args((lk), (flags), (ilk), (wmesg), (prio), (timo), \ LOCK_FILE, LOCK_LINE) #define lockmgr_args_rw(lk, flags, ilk, wmesg, prio, timo) \ _lockmgr_args_rw((lk), (flags), (ilk), (wmesg), (prio), (timo), \ LOCK_FILE, LOCK_LINE) #define lockmgr_disown(lk) \ _lockmgr_disown((lk), LOCK_FILE, LOCK_LINE) #define lockmgr_recursed_v(v) \ (v & LK_WRITER_RECURSED) #define lockmgr_recursed(lk) \ lockmgr_recursed_v((lk)->lk_lock) #define lockmgr_rw(lk, flags, ilk) \ _lockmgr_args_rw((lk), (flags), (ilk), LK_WMESG_DEFAULT, \ LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, LOCK_FILE, LOCK_LINE) #ifdef INVARIANTS #define lockmgr_assert(lk, what) \ _lockmgr_assert((lk), (what), LOCK_FILE, LOCK_LINE) #else #define lockmgr_assert(lk, what) #endif /* * Flags for lockinit(). */ #define LK_INIT_MASK 0x0001FF #define LK_CANRECURSE 0x000001 #define LK_NODUP 0x000002 #define LK_NOPROFILE 0x000004 #define LK_NOSHARE 0x000008 #define LK_NOWITNESS 0x000010 #define LK_QUIET 0x000020 #define LK_UNUSED0 0x000040 /* Was LK_ADAPTIVE */ #define LK_IS_VNODE 0x000080 /* Tell WITNESS about a VNODE lock */ #define LK_NEW 0x000100 /* * Additional attributes to be used in lockmgr(). */ #define LK_EATTR_MASK 0x00FF00 #define LK_INTERLOCK 0x000100 #define LK_NOWAIT 0x000200 #define LK_RETRY 0x000400 #define LK_SLEEPFAIL 0x000800 #define LK_TIMELOCK 0x001000 #define LK_NODDLKTREAT 0x002000 +#define LK_ADAPTIVE 0x004000 /* * Operations for lockmgr(). */ #define LK_TYPE_MASK 0xFF0000 #define LK_DOWNGRADE 0x010000 #define LK_DRAIN 0x020000 #define LK_EXCLOTHER 0x040000 #define LK_EXCLUSIVE 0x080000 #define LK_RELEASE 0x100000 #define LK_SHARED 0x200000 #define LK_UPGRADE 0x400000 #define LK_TRYUPGRADE 0x800000 #define LK_TOTAL_MASK (LK_INIT_MASK | LK_EATTR_MASK | LK_TYPE_MASK) /* * Default values for lockmgr_args(). */ #define LK_WMESG_DEFAULT (NULL) #define LK_PRIO_DEFAULT (0) #define LK_TIMO_DEFAULT (0) /* * Assertion flags. */ #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT) #define KA_LOCKED LA_LOCKED #define KA_SLOCKED LA_SLOCKED #define KA_XLOCKED LA_XLOCKED #define KA_UNLOCKED LA_UNLOCKED #define KA_RECURSED LA_RECURSED #define KA_NOTRECURSED LA_NOTRECURSED #endif #endif /* _KERNEL */ #endif /* !_SYS_LOCKMGR_H_ */ Index: head/sys/ufs/ffs/ffs_vnops.c =================================================================== --- head/sys/ufs/ffs/ffs_vnops.c (revision 363414) +++ head/sys/ufs/ffs/ffs_vnops.c (revision 363415) @@ -1,1823 +1,1825 @@ /*- * SPDX-License-Identifier: (BSD-2-Clause-FreeBSD AND BSD-3-Clause) * * Copyright (c) 2002, 2003 Networks Associates Technology, Inc. * All rights reserved. * * This software was developed for the FreeBSD Project by Marshall * Kirk McKusick and Network Associates Laboratories, the Security * Research Division of Network Associates, Inc. under DARPA/SPAWAR * contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA CHATS * research program * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)ufs_readwrite.c 8.11 (Berkeley) 5/8/95 * from: $FreeBSD: .../ufs/ufs_readwrite.c,v 1.96 2002/08/12 09:22:11 phk ... * @(#)ffs_vnops.c 8.15 (Berkeley) 5/14/95 */ #include __FBSDID("$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 "opt_directio.h" #include "opt_ffs.h" #define ALIGNED_TO(ptr, s) \ (((uintptr_t)(ptr) & (_Alignof(s) - 1)) == 0) #ifdef DIRECTIO extern int ffs_rawread(struct vnode *vp, struct uio *uio, int *workdone); #endif static vop_fdatasync_t ffs_fdatasync; static vop_fsync_t ffs_fsync; static vop_getpages_t ffs_getpages; static vop_getpages_async_t ffs_getpages_async; static vop_lock1_t ffs_lock; #ifdef INVARIANTS static vop_unlock_t ffs_unlock_debug; #endif static vop_read_t ffs_read; static vop_write_t ffs_write; static int ffs_extread(struct vnode *vp, struct uio *uio, int ioflag); static int ffs_extwrite(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *cred); static vop_strategy_t ffsext_strategy; static vop_closeextattr_t ffs_closeextattr; static vop_deleteextattr_t ffs_deleteextattr; static vop_getextattr_t ffs_getextattr; static vop_listextattr_t ffs_listextattr; static vop_openextattr_t ffs_openextattr; static vop_setextattr_t ffs_setextattr; static vop_vptofh_t ffs_vptofh; /* Global vfs data structures for ufs. */ struct vop_vector ffs_vnodeops1 = { .vop_default = &ufs_vnodeops, .vop_fsync = ffs_fsync, .vop_fdatasync = ffs_fdatasync, .vop_getpages = ffs_getpages, .vop_getpages_async = ffs_getpages_async, .vop_lock1 = ffs_lock, #ifdef INVARIANTS .vop_unlock = ffs_unlock_debug, #endif .vop_read = ffs_read, .vop_reallocblks = ffs_reallocblks, .vop_write = ffs_write, .vop_vptofh = ffs_vptofh, }; VFS_VOP_VECTOR_REGISTER(ffs_vnodeops1); struct vop_vector ffs_fifoops1 = { .vop_default = &ufs_fifoops, .vop_fsync = ffs_fsync, .vop_fdatasync = ffs_fdatasync, .vop_lock1 = ffs_lock, #ifdef INVARIANTS .vop_unlock = ffs_unlock_debug, #endif .vop_vptofh = ffs_vptofh, }; VFS_VOP_VECTOR_REGISTER(ffs_fifoops1); /* Global vfs data structures for ufs. */ struct vop_vector ffs_vnodeops2 = { .vop_default = &ufs_vnodeops, .vop_fsync = ffs_fsync, .vop_fdatasync = ffs_fdatasync, .vop_getpages = ffs_getpages, .vop_getpages_async = ffs_getpages_async, .vop_lock1 = ffs_lock, #ifdef INVARIANTS .vop_unlock = ffs_unlock_debug, #endif .vop_read = ffs_read, .vop_reallocblks = ffs_reallocblks, .vop_write = ffs_write, .vop_closeextattr = ffs_closeextattr, .vop_deleteextattr = ffs_deleteextattr, .vop_getextattr = ffs_getextattr, .vop_listextattr = ffs_listextattr, .vop_openextattr = ffs_openextattr, .vop_setextattr = ffs_setextattr, .vop_vptofh = ffs_vptofh, }; VFS_VOP_VECTOR_REGISTER(ffs_vnodeops2); struct vop_vector ffs_fifoops2 = { .vop_default = &ufs_fifoops, .vop_fsync = ffs_fsync, .vop_fdatasync = ffs_fdatasync, .vop_lock1 = ffs_lock, #ifdef INVARIANTS .vop_unlock = ffs_unlock_debug, #endif .vop_reallocblks = ffs_reallocblks, .vop_strategy = ffsext_strategy, .vop_closeextattr = ffs_closeextattr, .vop_deleteextattr = ffs_deleteextattr, .vop_getextattr = ffs_getextattr, .vop_listextattr = ffs_listextattr, .vop_openextattr = ffs_openextattr, .vop_setextattr = ffs_setextattr, .vop_vptofh = ffs_vptofh, }; VFS_VOP_VECTOR_REGISTER(ffs_fifoops2); /* * Synch an open file. */ /* ARGSUSED */ static int ffs_fsync(struct vop_fsync_args *ap) { struct vnode *vp; struct bufobj *bo; int error; vp = ap->a_vp; bo = &vp->v_bufobj; retry: error = ffs_syncvnode(vp, ap->a_waitfor, 0); if (error) return (error); if (ap->a_waitfor == MNT_WAIT && DOINGSOFTDEP(vp)) { error = softdep_fsync(vp); if (error) return (error); /* * The softdep_fsync() function may drop vp lock, * allowing for dirty buffers to reappear on the * bo_dirty list. Recheck and resync as needed. */ BO_LOCK(bo); if ((vp->v_type == VREG || vp->v_type == VDIR) && (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)) { BO_UNLOCK(bo); goto retry; } BO_UNLOCK(bo); } if (ffs_fsfail_cleanup(VFSTOUFS(vp->v_mount), 0)) return (ENXIO); return (0); } int ffs_syncvnode(struct vnode *vp, int waitfor, int flags) { struct inode *ip; struct bufobj *bo; struct ufsmount *ump; struct buf *bp, *nbp; ufs_lbn_t lbn; int error, passes; bool still_dirty, wait; ip = VTOI(vp); ip->i_flag &= ~IN_NEEDSYNC; bo = &vp->v_bufobj; ump = VFSTOUFS(vp->v_mount); /* * When doing MNT_WAIT we must first flush all dependencies * on the inode. */ if (DOINGSOFTDEP(vp) && waitfor == MNT_WAIT && (error = softdep_sync_metadata(vp)) != 0) { if (ffs_fsfail_cleanup(ump, error)) error = 0; return (error); } /* * Flush all dirty buffers associated with a vnode. */ error = 0; passes = 0; wait = false; /* Always do an async pass first. */ lbn = lblkno(ITOFS(ip), (ip->i_size + ITOFS(ip)->fs_bsize - 1)); BO_LOCK(bo); loop: TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) bp->b_vflags &= ~BV_SCANNED; TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) { /* * Reasons to skip this buffer: it has already been considered * on this pass, the buffer has dependencies that will cause * it to be redirtied and it has not already been deferred, * or it is already being written. */ if ((bp->b_vflags & BV_SCANNED) != 0) continue; bp->b_vflags |= BV_SCANNED; /* * Flush indirects in order, if requested. * * Note that if only datasync is requested, we can * skip indirect blocks when softupdates are not * active. Otherwise we must flush them with data, * since dependencies prevent data block writes. */ if (waitfor == MNT_WAIT && bp->b_lblkno <= -UFS_NDADDR && (lbn_level(bp->b_lblkno) >= passes || ((flags & DATA_ONLY) != 0 && !DOINGSOFTDEP(vp)))) continue; if (bp->b_lblkno > lbn) panic("ffs_syncvnode: syncing truncated data."); if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) { BO_UNLOCK(bo); } else if (wait) { if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo)) != 0) { bp->b_vflags &= ~BV_SCANNED; goto next; } } else continue; if ((bp->b_flags & B_DELWRI) == 0) panic("ffs_fsync: not dirty"); /* * Check for dependencies and potentially complete them. */ if (!LIST_EMPTY(&bp->b_dep) && (error = softdep_sync_buf(vp, bp, wait ? MNT_WAIT : MNT_NOWAIT)) != 0) { /* I/O error. */ if (error != EBUSY) { BUF_UNLOCK(bp); return (error); } /* If we deferred once, don't defer again. */ if ((bp->b_flags & B_DEFERRED) == 0) { bp->b_flags |= B_DEFERRED; BUF_UNLOCK(bp); goto next; } } if (wait) { bremfree(bp); error = bwrite(bp); if (ffs_fsfail_cleanup(ump, error)) error = 0; if (error != 0) return (error); } else if ((bp->b_flags & B_CLUSTEROK)) { (void) vfs_bio_awrite(bp); } else { bremfree(bp); (void) bawrite(bp); } next: /* * Since we may have slept during the I/O, we need * to start from a known point. */ BO_LOCK(bo); nbp = TAILQ_FIRST(&bo->bo_dirty.bv_hd); } if (waitfor != MNT_WAIT) { BO_UNLOCK(bo); if ((flags & NO_INO_UPDT) != 0) return (0); else return (ffs_update(vp, 0)); } /* Drain IO to see if we're done. */ bufobj_wwait(bo, 0, 0); /* * Block devices associated with filesystems may have new I/O * requests posted for them even if the vnode is locked, so no * amount of trying will get them clean. We make several passes * as a best effort. * * Regular files may need multiple passes to flush all dependency * work as it is possible that we must write once per indirect * level, once for the leaf, and once for the inode and each of * these will be done with one sync and one async pass. */ if (bo->bo_dirty.bv_cnt > 0) { if ((flags & DATA_ONLY) == 0) { still_dirty = true; } else { /* * For data-only sync, dirty indirect buffers * are ignored. */ still_dirty = false; TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) { if (bp->b_lblkno > -UFS_NDADDR) { still_dirty = true; break; } } } if (still_dirty) { /* Write the inode after sync passes to flush deps. */ if (wait && DOINGSOFTDEP(vp) && (flags & NO_INO_UPDT) == 0) { BO_UNLOCK(bo); ffs_update(vp, 1); BO_LOCK(bo); } /* switch between sync/async. */ wait = !wait; if (wait || ++passes < UFS_NIADDR + 2) goto loop; } } BO_UNLOCK(bo); error = 0; if ((flags & DATA_ONLY) == 0) { if ((flags & NO_INO_UPDT) == 0) error = ffs_update(vp, 1); if (DOINGSUJ(vp)) softdep_journal_fsync(VTOI(vp)); } else if ((ip->i_flags & (IN_SIZEMOD | IN_IBLKDATA)) != 0) { error = ffs_update(vp, 1); } return (error); } static int ffs_fdatasync(struct vop_fdatasync_args *ap) { return (ffs_syncvnode(ap->a_vp, MNT_WAIT, DATA_ONLY)); } static int ffs_lock(ap) struct vop_lock1_args /* { struct vnode *a_vp; int a_flags; struct thread *a_td; char *file; int line; } */ *ap; { #ifndef NO_FFS_SNAPSHOT struct vnode *vp; int flags; struct lock *lkp; int result; + ap->a_flags |= LK_ADAPTIVE; switch (ap->a_flags & LK_TYPE_MASK) { case LK_SHARED: case LK_UPGRADE: case LK_EXCLUSIVE: vp = ap->a_vp; flags = ap->a_flags; for (;;) { #ifdef DEBUG_VFS_LOCKS VNPASS(vp->v_holdcnt != 0, vp); #endif lkp = vp->v_vnlock; result = lockmgr_lock_flags(lkp, flags, &VI_MTX(vp)->lock_object, ap->a_file, ap->a_line); if (lkp == vp->v_vnlock || result != 0) break; /* * Apparent success, except that the vnode * mutated between snapshot file vnode and * regular file vnode while this process * slept. The lock currently held is not the * right lock. Release it, and try to get the * new lock. */ lockmgr_unlock(lkp); if ((flags & (LK_INTERLOCK | LK_NOWAIT)) == (LK_INTERLOCK | LK_NOWAIT)) return (EBUSY); if ((flags & LK_TYPE_MASK) == LK_UPGRADE) flags = (flags & ~LK_TYPE_MASK) | LK_EXCLUSIVE; flags &= ~LK_INTERLOCK; } break; default: result = VOP_LOCK1_APV(&ufs_vnodeops, ap); } return (result); #else + ap->a_flags |= LK_ADAPTIVE; return (VOP_LOCK1_APV(&ufs_vnodeops, ap)); #endif } #ifdef INVARIANTS static int ffs_unlock_debug(struct vop_unlock_args *ap) { struct vnode *vp = ap->a_vp; struct inode *ip = VTOI(vp); if (ip->i_flag & UFS_INODE_FLAG_LAZY_MASK_ASSERTABLE) { if ((vp->v_mflag & VMP_LAZYLIST) == 0) { VI_LOCK(vp); VNASSERT((vp->v_mflag & VMP_LAZYLIST), vp, ("%s: modified vnode (%x) not on lazy list", __func__, ip->i_flag)); VI_UNLOCK(vp); } } return (VOP_UNLOCK_APV(&ufs_vnodeops, ap)); } #endif static int ffs_read_hole(struct uio *uio, long xfersize, long *size) { ssize_t saved_resid, tlen; int error; while (xfersize > 0) { tlen = min(xfersize, ZERO_REGION_SIZE); saved_resid = uio->uio_resid; error = vn_io_fault_uiomove(__DECONST(void *, zero_region), tlen, uio); if (error != 0) return (error); tlen = saved_resid - uio->uio_resid; xfersize -= tlen; *size -= tlen; } return (0); } /* * Vnode op for reading. */ static int ffs_read(ap) struct vop_read_args /* { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; } */ *ap; { struct vnode *vp; struct inode *ip; struct uio *uio; struct fs *fs; struct buf *bp; ufs_lbn_t lbn, nextlbn; off_t bytesinfile; long size, xfersize, blkoffset; ssize_t orig_resid; int bflag, error, ioflag, seqcount; vp = ap->a_vp; uio = ap->a_uio; ioflag = ap->a_ioflag; if (ap->a_ioflag & IO_EXT) #ifdef notyet return (ffs_extread(vp, uio, ioflag)); #else panic("ffs_read+IO_EXT"); #endif #ifdef DIRECTIO if ((ioflag & IO_DIRECT) != 0) { int workdone; error = ffs_rawread(vp, uio, &workdone); if (error != 0 || workdone != 0) return error; } #endif seqcount = ap->a_ioflag >> IO_SEQSHIFT; ip = VTOI(vp); #ifdef INVARIANTS if (uio->uio_rw != UIO_READ) panic("ffs_read: mode"); if (vp->v_type == VLNK) { if ((int)ip->i_size < vp->v_mount->mnt_maxsymlinklen) panic("ffs_read: short symlink"); } else if (vp->v_type != VREG && vp->v_type != VDIR) panic("ffs_read: type %d", vp->v_type); #endif orig_resid = uio->uio_resid; KASSERT(orig_resid >= 0, ("ffs_read: uio->uio_resid < 0")); if (orig_resid == 0) return (0); KASSERT(uio->uio_offset >= 0, ("ffs_read: uio->uio_offset < 0")); fs = ITOFS(ip); if (uio->uio_offset < ip->i_size && uio->uio_offset >= fs->fs_maxfilesize) return (EOVERFLOW); bflag = GB_UNMAPPED | (uio->uio_segflg == UIO_NOCOPY ? 0 : GB_NOSPARSE); for (error = 0, bp = NULL; uio->uio_resid > 0; bp = NULL) { if ((bytesinfile = ip->i_size - uio->uio_offset) <= 0) break; lbn = lblkno(fs, uio->uio_offset); nextlbn = lbn + 1; /* * size of buffer. The buffer representing the * end of the file is rounded up to the size of * the block type ( fragment or full block, * depending ). */ size = blksize(fs, ip, lbn); blkoffset = blkoff(fs, uio->uio_offset); /* * The amount we want to transfer in this iteration is * one FS block less the amount of the data before * our startpoint (duh!) */ xfersize = fs->fs_bsize - blkoffset; /* * But if we actually want less than the block, * or the file doesn't have a whole block more of data, * then use the lesser number. */ if (uio->uio_resid < xfersize) xfersize = uio->uio_resid; if (bytesinfile < xfersize) xfersize = bytesinfile; if (lblktosize(fs, nextlbn) >= ip->i_size) { /* * Don't do readahead if this is the end of the file. */ error = bread_gb(vp, lbn, size, NOCRED, bflag, &bp); } else if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERR) == 0) { /* * Otherwise if we are allowed to cluster, * grab as much as we can. * * XXX This may not be a win if we are not * doing sequential access. */ error = cluster_read(vp, ip->i_size, lbn, size, NOCRED, blkoffset + uio->uio_resid, seqcount, bflag, &bp); } else if (seqcount > 1) { /* * If we are NOT allowed to cluster, then * if we appear to be acting sequentially, * fire off a request for a readahead * as well as a read. Note that the 4th and 5th * arguments point to arrays of the size specified in * the 6th argument. */ u_int nextsize = blksize(fs, ip, nextlbn); error = breadn_flags(vp, lbn, lbn, size, &nextlbn, &nextsize, 1, NOCRED, bflag, NULL, &bp); } else { /* * Failing all of the above, just read what the * user asked for. Interestingly, the same as * the first option above. */ error = bread_gb(vp, lbn, size, NOCRED, bflag, &bp); } if (error == EJUSTRETURN) { error = ffs_read_hole(uio, xfersize, &size); if (error == 0) continue; } if (error != 0) { brelse(bp); bp = NULL; break; } /* * We should only get non-zero b_resid when an I/O error * has occurred, which should cause us to break above. * However, if the short read did not cause an error, * then we want to ensure that we do not uiomove bad * or uninitialized data. */ size -= bp->b_resid; if (size < xfersize) { if (size == 0) break; xfersize = size; } if (buf_mapped(bp)) { error = vn_io_fault_uiomove((char *)bp->b_data + blkoffset, (int)xfersize, uio); } else { error = vn_io_fault_pgmove(bp->b_pages, blkoffset, (int)xfersize, uio); } if (error) break; vfs_bio_brelse(bp, ioflag); } /* * This can only happen in the case of an error * because the loop above resets bp to NULL on each iteration * and on normal completion has not set a new value into it. * so it must have come from a 'break' statement */ if (bp != NULL) vfs_bio_brelse(bp, ioflag); if ((error == 0 || uio->uio_resid != orig_resid) && (vp->v_mount->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0) UFS_INODE_SET_FLAG_SHARED(ip, IN_ACCESS); return (error); } /* * Vnode op for writing. */ static int ffs_write(ap) struct vop_write_args /* { struct vnode *a_vp; struct uio *a_uio; int a_ioflag; struct ucred *a_cred; } */ *ap; { struct vnode *vp; struct uio *uio; struct inode *ip; struct fs *fs; struct buf *bp; ufs_lbn_t lbn; off_t osize; ssize_t resid; int seqcount; int blkoffset, error, flags, ioflag, size, xfersize; vp = ap->a_vp; uio = ap->a_uio; ioflag = ap->a_ioflag; if (ap->a_ioflag & IO_EXT) #ifdef notyet return (ffs_extwrite(vp, uio, ioflag, ap->a_cred)); #else panic("ffs_write+IO_EXT"); #endif seqcount = ap->a_ioflag >> IO_SEQSHIFT; ip = VTOI(vp); #ifdef INVARIANTS if (uio->uio_rw != UIO_WRITE) panic("ffs_write: mode"); #endif switch (vp->v_type) { case VREG: if (ioflag & IO_APPEND) uio->uio_offset = ip->i_size; if ((ip->i_flags & APPEND) && uio->uio_offset != ip->i_size) return (EPERM); /* FALLTHROUGH */ case VLNK: break; case VDIR: panic("ffs_write: dir write"); break; default: panic("ffs_write: type %p %d (%d,%d)", vp, (int)vp->v_type, (int)uio->uio_offset, (int)uio->uio_resid ); } KASSERT(uio->uio_resid >= 0, ("ffs_write: uio->uio_resid < 0")); KASSERT(uio->uio_offset >= 0, ("ffs_write: uio->uio_offset < 0")); fs = ITOFS(ip); if ((uoff_t)uio->uio_offset + uio->uio_resid > fs->fs_maxfilesize) return (EFBIG); /* * Maybe this should be above the vnode op call, but so long as * file servers have no limits, I don't think it matters. */ if (vn_rlimit_fsize(vp, uio, uio->uio_td)) return (EFBIG); resid = uio->uio_resid; osize = ip->i_size; if (seqcount > BA_SEQMAX) flags = BA_SEQMAX << BA_SEQSHIFT; else flags = seqcount << BA_SEQSHIFT; if (ioflag & IO_SYNC) flags |= IO_SYNC; flags |= BA_UNMAPPED; for (error = 0; uio->uio_resid > 0;) { lbn = lblkno(fs, uio->uio_offset); blkoffset = blkoff(fs, uio->uio_offset); xfersize = fs->fs_bsize - blkoffset; if (uio->uio_resid < xfersize) xfersize = uio->uio_resid; if (uio->uio_offset + xfersize > ip->i_size) vnode_pager_setsize(vp, uio->uio_offset + xfersize); /* * We must perform a read-before-write if the transfer size * does not cover the entire buffer. */ if (fs->fs_bsize > xfersize) flags |= BA_CLRBUF; else flags &= ~BA_CLRBUF; /* XXX is uio->uio_offset the right thing here? */ error = UFS_BALLOC(vp, uio->uio_offset, xfersize, ap->a_cred, flags, &bp); if (error != 0) { vnode_pager_setsize(vp, ip->i_size); break; } if ((ioflag & (IO_SYNC|IO_INVAL)) == (IO_SYNC|IO_INVAL)) bp->b_flags |= B_NOCACHE; if (uio->uio_offset + xfersize > ip->i_size) { ip->i_size = uio->uio_offset + xfersize; DIP_SET(ip, i_size, ip->i_size); UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE); } size = blksize(fs, ip, lbn) - bp->b_resid; if (size < xfersize) xfersize = size; if (buf_mapped(bp)) { error = vn_io_fault_uiomove((char *)bp->b_data + blkoffset, (int)xfersize, uio); } else { error = vn_io_fault_pgmove(bp->b_pages, blkoffset, (int)xfersize, uio); } /* * If the buffer is not already filled and we encounter an * error while trying to fill it, we have to clear out any * garbage data from the pages instantiated for the buffer. * If we do not, a failed uiomove() during a write can leave * the prior contents of the pages exposed to a userland mmap. * * Note that we need only clear buffers with a transfer size * equal to the block size because buffers with a shorter * transfer size were cleared above by the call to UFS_BALLOC() * with the BA_CLRBUF flag set. * * If the source region for uiomove identically mmaps the * buffer, uiomove() performed the NOP copy, and the buffer * content remains valid because the page fault handler * validated the pages. */ if (error != 0 && (bp->b_flags & B_CACHE) == 0 && fs->fs_bsize == xfersize) vfs_bio_clrbuf(bp); vfs_bio_set_flags(bp, ioflag); /* * If IO_SYNC each buffer is written synchronously. Otherwise * if we have a severe page deficiency write the buffer * asynchronously. Otherwise try to cluster, and if that * doesn't do it then either do an async write (if O_DIRECT), * or a delayed write (if not). */ if (ioflag & IO_SYNC) { (void)bwrite(bp); } else if (vm_page_count_severe() || buf_dirty_count_severe() || (ioflag & IO_ASYNC)) { bp->b_flags |= B_CLUSTEROK; bawrite(bp); } else if (xfersize + blkoffset == fs->fs_bsize) { if ((vp->v_mount->mnt_flag & MNT_NOCLUSTERW) == 0) { bp->b_flags |= B_CLUSTEROK; cluster_write(vp, bp, ip->i_size, seqcount, GB_UNMAPPED); } else { bawrite(bp); } } else if (ioflag & IO_DIRECT) { bp->b_flags |= B_CLUSTEROK; bawrite(bp); } else { bp->b_flags |= B_CLUSTEROK; bdwrite(bp); } if (error || xfersize == 0) break; UFS_INODE_SET_FLAG(ip, IN_CHANGE | IN_UPDATE); } /* * If we successfully wrote any data, and we are not the superuser * we clear the setuid and setgid bits as a precaution against * tampering. */ if ((ip->i_mode & (ISUID | ISGID)) && resid > uio->uio_resid && ap->a_cred) { if (priv_check_cred(ap->a_cred, PRIV_VFS_RETAINSUGID)) { ip->i_mode &= ~(ISUID | ISGID); DIP_SET(ip, i_mode, ip->i_mode); } } if (error) { if (ioflag & IO_UNIT) { (void)ffs_truncate(vp, osize, IO_NORMAL | (ioflag & IO_SYNC), ap->a_cred); uio->uio_offset -= resid - uio->uio_resid; uio->uio_resid = resid; } } else if (resid > uio->uio_resid && (ioflag & IO_SYNC)) { error = ffs_update(vp, 1); if (ffs_fsfail_cleanup(VFSTOUFS(vp->v_mount), error)) error = ENXIO; } return (error); } /* * Extended attribute area reading. */ static int ffs_extread(struct vnode *vp, struct uio *uio, int ioflag) { struct inode *ip; struct ufs2_dinode *dp; struct fs *fs; struct buf *bp; ufs_lbn_t lbn, nextlbn; off_t bytesinfile; long size, xfersize, blkoffset; ssize_t orig_resid; int error; ip = VTOI(vp); fs = ITOFS(ip); dp = ip->i_din2; #ifdef INVARIANTS if (uio->uio_rw != UIO_READ || fs->fs_magic != FS_UFS2_MAGIC) panic("ffs_extread: mode"); #endif orig_resid = uio->uio_resid; KASSERT(orig_resid >= 0, ("ffs_extread: uio->uio_resid < 0")); if (orig_resid == 0) return (0); KASSERT(uio->uio_offset >= 0, ("ffs_extread: uio->uio_offset < 0")); for (error = 0, bp = NULL; uio->uio_resid > 0; bp = NULL) { if ((bytesinfile = dp->di_extsize - uio->uio_offset) <= 0) break; lbn = lblkno(fs, uio->uio_offset); nextlbn = lbn + 1; /* * size of buffer. The buffer representing the * end of the file is rounded up to the size of * the block type ( fragment or full block, * depending ). */ size = sblksize(fs, dp->di_extsize, lbn); blkoffset = blkoff(fs, uio->uio_offset); /* * The amount we want to transfer in this iteration is * one FS block less the amount of the data before * our startpoint (duh!) */ xfersize = fs->fs_bsize - blkoffset; /* * But if we actually want less than the block, * or the file doesn't have a whole block more of data, * then use the lesser number. */ if (uio->uio_resid < xfersize) xfersize = uio->uio_resid; if (bytesinfile < xfersize) xfersize = bytesinfile; if (lblktosize(fs, nextlbn) >= dp->di_extsize) { /* * Don't do readahead if this is the end of the info. */ error = bread(vp, -1 - lbn, size, NOCRED, &bp); } else { /* * If we have a second block, then * fire off a request for a readahead * as well as a read. Note that the 4th and 5th * arguments point to arrays of the size specified in * the 6th argument. */ u_int nextsize = sblksize(fs, dp->di_extsize, nextlbn); nextlbn = -1 - nextlbn; error = breadn(vp, -1 - lbn, size, &nextlbn, &nextsize, 1, NOCRED, &bp); } if (error) { brelse(bp); bp = NULL; break; } /* * We should only get non-zero b_resid when an I/O error * has occurred, which should cause us to break above. * However, if the short read did not cause an error, * then we want to ensure that we do not uiomove bad * or uninitialized data. */ size -= bp->b_resid; if (size < xfersize) { if (size == 0) break; xfersize = size; } error = uiomove((char *)bp->b_data + blkoffset, (int)xfersize, uio); if (error) break; vfs_bio_brelse(bp, ioflag); } /* * This can only happen in the case of an error * because the loop above resets bp to NULL on each iteration * and on normal completion has not set a new value into it. * so it must have come from a 'break' statement */ if (bp != NULL) vfs_bio_brelse(bp, ioflag); return (error); } /* * Extended attribute area writing. */ static int ffs_extwrite(struct vnode *vp, struct uio *uio, int ioflag, struct ucred *ucred) { struct inode *ip; struct ufs2_dinode *dp; struct fs *fs; struct buf *bp; ufs_lbn_t lbn; off_t osize; ssize_t resid; int blkoffset, error, flags, size, xfersize; ip = VTOI(vp); fs = ITOFS(ip); dp = ip->i_din2; #ifdef INVARIANTS if (uio->uio_rw != UIO_WRITE || fs->fs_magic != FS_UFS2_MAGIC) panic("ffs_extwrite: mode"); #endif if (ioflag & IO_APPEND) uio->uio_offset = dp->di_extsize; KASSERT(uio->uio_offset >= 0, ("ffs_extwrite: uio->uio_offset < 0")); KASSERT(uio->uio_resid >= 0, ("ffs_extwrite: uio->uio_resid < 0")); if ((uoff_t)uio->uio_offset + uio->uio_resid > UFS_NXADDR * fs->fs_bsize) return (EFBIG); resid = uio->uio_resid; osize = dp->di_extsize; flags = IO_EXT; if (ioflag & IO_SYNC) flags |= IO_SYNC; for (error = 0; uio->uio_resid > 0;) { lbn = lblkno(fs, uio->uio_offset); blkoffset = blkoff(fs, uio->uio_offset); xfersize = fs->fs_bsize - blkoffset; if (uio->uio_resid < xfersize) xfersize = uio->uio_resid; /* * We must perform a read-before-write if the transfer size * does not cover the entire buffer. */ if (fs->fs_bsize > xfersize) flags |= BA_CLRBUF; else flags &= ~BA_CLRBUF; error = UFS_BALLOC(vp, uio->uio_offset, xfersize, ucred, flags, &bp); if (error != 0) break; /* * If the buffer is not valid we have to clear out any * garbage data from the pages instantiated for the buffer. * If we do not, a failed uiomove() during a write can leave * the prior contents of the pages exposed to a userland * mmap(). XXX deal with uiomove() errors a better way. */ if ((bp->b_flags & B_CACHE) == 0 && fs->fs_bsize <= xfersize) vfs_bio_clrbuf(bp); if (uio->uio_offset + xfersize > dp->di_extsize) { dp->di_extsize = uio->uio_offset + xfersize; UFS_INODE_SET_FLAG(ip, IN_SIZEMOD | IN_CHANGE); } size = sblksize(fs, dp->di_extsize, lbn) - bp->b_resid; if (size < xfersize) xfersize = size; error = uiomove((char *)bp->b_data + blkoffset, (int)xfersize, uio); vfs_bio_set_flags(bp, ioflag); /* * If IO_SYNC each buffer is written synchronously. Otherwise * if we have a severe page deficiency write the buffer * asynchronously. Otherwise try to cluster, and if that * doesn't do it then either do an async write (if O_DIRECT), * or a delayed write (if not). */ if (ioflag & IO_SYNC) { (void)bwrite(bp); } else if (vm_page_count_severe() || buf_dirty_count_severe() || xfersize + blkoffset == fs->fs_bsize || (ioflag & (IO_ASYNC | IO_DIRECT))) bawrite(bp); else bdwrite(bp); if (error || xfersize == 0) break; UFS_INODE_SET_FLAG(ip, IN_CHANGE); } /* * If we successfully wrote any data, and we are not the superuser * we clear the setuid and setgid bits as a precaution against * tampering. */ if ((ip->i_mode & (ISUID | ISGID)) && resid > uio->uio_resid && ucred) { if (priv_check_cred(ucred, PRIV_VFS_RETAINSUGID)) { ip->i_mode &= ~(ISUID | ISGID); dp->di_mode = ip->i_mode; } } if (error) { if (ioflag & IO_UNIT) { (void)ffs_truncate(vp, osize, IO_EXT | (ioflag&IO_SYNC), ucred); uio->uio_offset -= resid - uio->uio_resid; uio->uio_resid = resid; } } else if (resid > uio->uio_resid && (ioflag & IO_SYNC)) error = ffs_update(vp, 1); return (error); } /* * Vnode operating to retrieve a named extended attribute. * * Locate a particular EA (nspace:name) in the area (ptr:length), and return * the length of the EA, and possibly the pointer to the entry and to the data. */ static int ffs_findextattr(u_char *ptr, u_int length, int nspace, const char *name, struct extattr **eapp, u_char **eac) { struct extattr *eap, *eaend; size_t nlen; nlen = strlen(name); KASSERT(ALIGNED_TO(ptr, struct extattr), ("unaligned")); eap = (struct extattr *)ptr; eaend = (struct extattr *)(ptr + length); for (; eap < eaend; eap = EXTATTR_NEXT(eap)) { /* make sure this entry is complete */ if (EXTATTR_NEXT(eap) > eaend) break; if (eap->ea_namespace != nspace || eap->ea_namelength != nlen || memcmp(eap->ea_name, name, nlen) != 0) continue; if (eapp != NULL) *eapp = eap; if (eac != NULL) *eac = EXTATTR_CONTENT(eap); return (EXTATTR_CONTENT_SIZE(eap)); } return (-1); } static int ffs_rdextattr(u_char **p, struct vnode *vp, struct thread *td, int extra) { struct inode *ip; struct ufs2_dinode *dp; struct fs *fs; struct uio luio; struct iovec liovec; u_int easize; int error; u_char *eae; ip = VTOI(vp); fs = ITOFS(ip); dp = ip->i_din2; easize = dp->di_extsize; if ((uoff_t)easize + extra > UFS_NXADDR * fs->fs_bsize) return (EFBIG); eae = malloc(easize + extra, M_TEMP, M_WAITOK); liovec.iov_base = eae; liovec.iov_len = easize; luio.uio_iov = &liovec; luio.uio_iovcnt = 1; luio.uio_offset = 0; luio.uio_resid = easize; luio.uio_segflg = UIO_SYSSPACE; luio.uio_rw = UIO_READ; luio.uio_td = td; error = ffs_extread(vp, &luio, IO_EXT | IO_SYNC); if (error) { free(eae, M_TEMP); return(error); } *p = eae; return (0); } static void ffs_lock_ea(struct vnode *vp) { struct inode *ip; ip = VTOI(vp); VI_LOCK(vp); while (ip->i_flag & IN_EA_LOCKED) { UFS_INODE_SET_FLAG(ip, IN_EA_LOCKWAIT); msleep(&ip->i_ea_refs, &vp->v_interlock, PINOD + 2, "ufs_ea", 0); } UFS_INODE_SET_FLAG(ip, IN_EA_LOCKED); VI_UNLOCK(vp); } static void ffs_unlock_ea(struct vnode *vp) { struct inode *ip; ip = VTOI(vp); VI_LOCK(vp); if (ip->i_flag & IN_EA_LOCKWAIT) wakeup(&ip->i_ea_refs); ip->i_flag &= ~(IN_EA_LOCKED | IN_EA_LOCKWAIT); VI_UNLOCK(vp); } static int ffs_open_ea(struct vnode *vp, struct ucred *cred, struct thread *td) { struct inode *ip; struct ufs2_dinode *dp; int error; ip = VTOI(vp); ffs_lock_ea(vp); if (ip->i_ea_area != NULL) { ip->i_ea_refs++; ffs_unlock_ea(vp); return (0); } dp = ip->i_din2; error = ffs_rdextattr(&ip->i_ea_area, vp, td, 0); if (error) { ffs_unlock_ea(vp); return (error); } ip->i_ea_len = dp->di_extsize; ip->i_ea_error = 0; ip->i_ea_refs++; ffs_unlock_ea(vp); return (0); } /* * Vnode extattr transaction commit/abort */ static int ffs_close_ea(struct vnode *vp, int commit, struct ucred *cred, struct thread *td) { struct inode *ip; struct uio luio; struct iovec liovec; int error; struct ufs2_dinode *dp; ip = VTOI(vp); ffs_lock_ea(vp); if (ip->i_ea_area == NULL) { ffs_unlock_ea(vp); return (EINVAL); } dp = ip->i_din2; error = ip->i_ea_error; if (commit && error == 0) { ASSERT_VOP_ELOCKED(vp, "ffs_close_ea commit"); if (cred == NOCRED) cred = vp->v_mount->mnt_cred; liovec.iov_base = ip->i_ea_area; liovec.iov_len = ip->i_ea_len; luio.uio_iov = &liovec; luio.uio_iovcnt = 1; luio.uio_offset = 0; luio.uio_resid = ip->i_ea_len; luio.uio_segflg = UIO_SYSSPACE; luio.uio_rw = UIO_WRITE; luio.uio_td = td; /* XXX: I'm not happy about truncating to zero size */ if (ip->i_ea_len < dp->di_extsize) error = ffs_truncate(vp, 0, IO_EXT, cred); error = ffs_extwrite(vp, &luio, IO_EXT | IO_SYNC, cred); } if (--ip->i_ea_refs == 0) { free(ip->i_ea_area, M_TEMP); ip->i_ea_area = NULL; ip->i_ea_len = 0; ip->i_ea_error = 0; } ffs_unlock_ea(vp); return (error); } /* * Vnode extattr strategy routine for fifos. * * We need to check for a read or write of the external attributes. * Otherwise we just fall through and do the usual thing. */ static int ffsext_strategy(struct vop_strategy_args *ap) /* struct vop_strategy_args { struct vnodeop_desc *a_desc; struct vnode *a_vp; struct buf *a_bp; }; */ { struct vnode *vp; daddr_t lbn; vp = ap->a_vp; lbn = ap->a_bp->b_lblkno; if (I_IS_UFS2(VTOI(vp)) && lbn < 0 && lbn >= -UFS_NXADDR) return (VOP_STRATEGY_APV(&ufs_vnodeops, ap)); if (vp->v_type == VFIFO) return (VOP_STRATEGY_APV(&ufs_fifoops, ap)); panic("spec nodes went here"); } /* * Vnode extattr transaction commit/abort */ static int ffs_openextattr(struct vop_openextattr_args *ap) /* struct vop_openextattr_args { struct vnodeop_desc *a_desc; struct vnode *a_vp; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { if (ap->a_vp->v_type == VCHR || ap->a_vp->v_type == VBLK) return (EOPNOTSUPP); return (ffs_open_ea(ap->a_vp, ap->a_cred, ap->a_td)); } /* * Vnode extattr transaction commit/abort */ static int ffs_closeextattr(struct vop_closeextattr_args *ap) /* struct vop_closeextattr_args { struct vnodeop_desc *a_desc; struct vnode *a_vp; int a_commit; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { if (ap->a_vp->v_type == VCHR || ap->a_vp->v_type == VBLK) return (EOPNOTSUPP); if (ap->a_commit && (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY)) return (EROFS); return (ffs_close_ea(ap->a_vp, ap->a_commit, ap->a_cred, ap->a_td)); } /* * Vnode operation to remove a named attribute. */ static int ffs_deleteextattr(struct vop_deleteextattr_args *ap) /* vop_deleteextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { struct inode *ip; struct extattr *eap; uint32_t ul; int olen, error, i, easize; u_char *eae; void *tmp; ip = VTOI(ap->a_vp); if (ap->a_vp->v_type == VCHR || ap->a_vp->v_type == VBLK) return (EOPNOTSUPP); if (strlen(ap->a_name) == 0) return (EINVAL); if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY) return (EROFS); error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VWRITE); if (error) { /* * ffs_lock_ea is not needed there, because the vnode * must be exclusively locked. */ if (ip->i_ea_area != NULL && ip->i_ea_error == 0) ip->i_ea_error = error; return (error); } error = ffs_open_ea(ap->a_vp, ap->a_cred, ap->a_td); if (error) return (error); /* CEM: delete could be done in-place instead */ eae = malloc(ip->i_ea_len, M_TEMP, M_WAITOK); bcopy(ip->i_ea_area, eae, ip->i_ea_len); easize = ip->i_ea_len; olen = ffs_findextattr(eae, easize, ap->a_attrnamespace, ap->a_name, &eap, NULL); if (olen == -1) { /* delete but nonexistent */ free(eae, M_TEMP); ffs_close_ea(ap->a_vp, 0, ap->a_cred, ap->a_td); return (ENOATTR); } ul = eap->ea_length; i = (u_char *)EXTATTR_NEXT(eap) - eae; bcopy(EXTATTR_NEXT(eap), eap, easize - i); easize -= ul; tmp = ip->i_ea_area; ip->i_ea_area = eae; ip->i_ea_len = easize; free(tmp, M_TEMP); error = ffs_close_ea(ap->a_vp, 1, ap->a_cred, ap->a_td); return (error); } /* * Vnode operation to retrieve a named extended attribute. */ static int ffs_getextattr(struct vop_getextattr_args *ap) /* vop_getextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; INOUT struct uio *a_uio; OUT size_t *a_size; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { struct inode *ip; u_char *eae, *p; unsigned easize; int error, ealen; ip = VTOI(ap->a_vp); if (ap->a_vp->v_type == VCHR || ap->a_vp->v_type == VBLK) return (EOPNOTSUPP); error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VREAD); if (error) return (error); error = ffs_open_ea(ap->a_vp, ap->a_cred, ap->a_td); if (error) return (error); eae = ip->i_ea_area; easize = ip->i_ea_len; ealen = ffs_findextattr(eae, easize, ap->a_attrnamespace, ap->a_name, NULL, &p); if (ealen >= 0) { error = 0; if (ap->a_size != NULL) *ap->a_size = ealen; else if (ap->a_uio != NULL) error = uiomove(p, ealen, ap->a_uio); } else error = ENOATTR; ffs_close_ea(ap->a_vp, 0, ap->a_cred, ap->a_td); return (error); } /* * Vnode operation to retrieve extended attributes on a vnode. */ static int ffs_listextattr(struct vop_listextattr_args *ap) /* vop_listextattr { IN struct vnode *a_vp; IN int a_attrnamespace; INOUT struct uio *a_uio; OUT size_t *a_size; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { struct inode *ip; struct extattr *eap, *eaend; int error, ealen; ip = VTOI(ap->a_vp); if (ap->a_vp->v_type == VCHR || ap->a_vp->v_type == VBLK) return (EOPNOTSUPP); error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VREAD); if (error) return (error); error = ffs_open_ea(ap->a_vp, ap->a_cred, ap->a_td); if (error) return (error); error = 0; if (ap->a_size != NULL) *ap->a_size = 0; KASSERT(ALIGNED_TO(ip->i_ea_area, struct extattr), ("unaligned")); eap = (struct extattr *)ip->i_ea_area; eaend = (struct extattr *)(ip->i_ea_area + ip->i_ea_len); for (; error == 0 && eap < eaend; eap = EXTATTR_NEXT(eap)) { /* make sure this entry is complete */ if (EXTATTR_NEXT(eap) > eaend) break; if (eap->ea_namespace != ap->a_attrnamespace) continue; ealen = eap->ea_namelength; if (ap->a_size != NULL) *ap->a_size += ealen + 1; else if (ap->a_uio != NULL) error = uiomove(&eap->ea_namelength, ealen + 1, ap->a_uio); } ffs_close_ea(ap->a_vp, 0, ap->a_cred, ap->a_td); return (error); } /* * Vnode operation to set a named attribute. */ static int ffs_setextattr(struct vop_setextattr_args *ap) /* vop_setextattr { IN struct vnode *a_vp; IN int a_attrnamespace; IN const char *a_name; INOUT struct uio *a_uio; IN struct ucred *a_cred; IN struct thread *a_td; }; */ { struct inode *ip; struct fs *fs; struct extattr *eap; uint32_t ealength, ul; ssize_t ealen; int olen, eapad1, eapad2, error, i, easize; u_char *eae; void *tmp; ip = VTOI(ap->a_vp); fs = ITOFS(ip); if (ap->a_vp->v_type == VCHR || ap->a_vp->v_type == VBLK) return (EOPNOTSUPP); if (strlen(ap->a_name) == 0) return (EINVAL); /* XXX Now unsupported API to delete EAs using NULL uio. */ if (ap->a_uio == NULL) return (EOPNOTSUPP); if (ap->a_vp->v_mount->mnt_flag & MNT_RDONLY) return (EROFS); ealen = ap->a_uio->uio_resid; if (ealen < 0 || ealen > lblktosize(fs, UFS_NXADDR)) return (EINVAL); error = extattr_check_cred(ap->a_vp, ap->a_attrnamespace, ap->a_cred, ap->a_td, VWRITE); if (error) { /* * ffs_lock_ea is not needed there, because the vnode * must be exclusively locked. */ if (ip->i_ea_area != NULL && ip->i_ea_error == 0) ip->i_ea_error = error; return (error); } error = ffs_open_ea(ap->a_vp, ap->a_cred, ap->a_td); if (error) return (error); ealength = sizeof(uint32_t) + 3 + strlen(ap->a_name); eapad1 = roundup2(ealength, 8) - ealength; eapad2 = roundup2(ealen, 8) - ealen; ealength += eapad1 + ealen + eapad2; /* * CEM: rewrites of the same size or smaller could be done in-place * instead. (We don't acquire any fine-grained locks in here either, * so we could also do bigger writes in-place.) */ eae = malloc(ip->i_ea_len + ealength, M_TEMP, M_WAITOK); bcopy(ip->i_ea_area, eae, ip->i_ea_len); easize = ip->i_ea_len; olen = ffs_findextattr(eae, easize, ap->a_attrnamespace, ap->a_name, &eap, NULL); if (olen == -1) { /* new, append at end */ KASSERT(ALIGNED_TO(eae + easize, struct extattr), ("unaligned")); eap = (struct extattr *)(eae + easize); easize += ealength; } else { ul = eap->ea_length; i = (u_char *)EXTATTR_NEXT(eap) - eae; if (ul != ealength) { bcopy(EXTATTR_NEXT(eap), (u_char *)eap + ealength, easize - i); easize += (ealength - ul); } } if (easize > lblktosize(fs, UFS_NXADDR)) { free(eae, M_TEMP); ffs_close_ea(ap->a_vp, 0, ap->a_cred, ap->a_td); if (ip->i_ea_area != NULL && ip->i_ea_error == 0) ip->i_ea_error = ENOSPC; return (ENOSPC); } eap->ea_length = ealength; eap->ea_namespace = ap->a_attrnamespace; eap->ea_contentpadlen = eapad2; eap->ea_namelength = strlen(ap->a_name); memcpy(eap->ea_name, ap->a_name, strlen(ap->a_name)); bzero(&eap->ea_name[strlen(ap->a_name)], eapad1); error = uiomove(EXTATTR_CONTENT(eap), ealen, ap->a_uio); if (error) { free(eae, M_TEMP); ffs_close_ea(ap->a_vp, 0, ap->a_cred, ap->a_td); if (ip->i_ea_area != NULL && ip->i_ea_error == 0) ip->i_ea_error = error; return (error); } bzero((u_char *)EXTATTR_CONTENT(eap) + ealen, eapad2); tmp = ip->i_ea_area; ip->i_ea_area = eae; ip->i_ea_len = easize; free(tmp, M_TEMP); error = ffs_close_ea(ap->a_vp, 1, ap->a_cred, ap->a_td); return (error); } /* * Vnode pointer to File handle */ static int ffs_vptofh(struct vop_vptofh_args *ap) /* vop_vptofh { IN struct vnode *a_vp; IN struct fid *a_fhp; }; */ { struct inode *ip; struct ufid *ufhp; ip = VTOI(ap->a_vp); ufhp = (struct ufid *)ap->a_fhp; ufhp->ufid_len = sizeof(struct ufid); ufhp->ufid_ino = ip->i_number; ufhp->ufid_gen = ip->i_gen; return (0); } SYSCTL_DECL(_vfs_ffs); static int use_buf_pager = 1; SYSCTL_INT(_vfs_ffs, OID_AUTO, use_buf_pager, CTLFLAG_RWTUN, &use_buf_pager, 0, "Always use buffer pager instead of bmap"); static daddr_t ffs_gbp_getblkno(struct vnode *vp, vm_ooffset_t off) { return (lblkno(VFSTOUFS(vp->v_mount)->um_fs, off)); } static int ffs_gbp_getblksz(struct vnode *vp, daddr_t lbn) { return (blksize(VFSTOUFS(vp->v_mount)->um_fs, VTOI(vp), lbn)); } static int ffs_getpages(struct vop_getpages_args *ap) { struct vnode *vp; struct ufsmount *um; vp = ap->a_vp; um = VFSTOUFS(vp->v_mount); if (!use_buf_pager && um->um_devvp->v_bufobj.bo_bsize <= PAGE_SIZE) return (vnode_pager_generic_getpages(vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, NULL, NULL)); return (vfs_bio_getpages(vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, ffs_gbp_getblkno, ffs_gbp_getblksz)); } static int ffs_getpages_async(struct vop_getpages_async_args *ap) { struct vnode *vp; struct ufsmount *um; bool do_iodone; int error; vp = ap->a_vp; um = VFSTOUFS(vp->v_mount); do_iodone = true; if (um->um_devvp->v_bufobj.bo_bsize <= PAGE_SIZE) { error = vnode_pager_generic_getpages(vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, ap->a_iodone, ap->a_arg); if (error == 0) do_iodone = false; } else { error = vfs_bio_getpages(vp, ap->a_m, ap->a_count, ap->a_rbehind, ap->a_rahead, ffs_gbp_getblkno, ffs_gbp_getblksz); } if (do_iodone && ap->a_iodone != NULL) ap->a_iodone(ap->a_arg, ap->a_m, ap->a_count, error); return (error); }