/* * Copyright (c) 2002, Jeffrey Roberson * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD: src/sys/kern/kern_umtx.c,v 1.18 2004/11/30 12:18:53 davidxu Exp $"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct umtx_key { vm_object_t uk_object; vm_ooffset_t uk_offset; }; struct umtx_q { LIST_ENTRY(umtx_q) uq_next; /* Linked list for the hash. */ struct umtx_key uq_key; /* Umtx key */ struct thread *uq_thread; /* The thread waits on */ TAILQ_ENTRY(umtx_q) uq_rqnext; /* Linked list for requeuing. */ vm_offset_t uq_addr; }; LIST_HEAD(umtx_head, umtx_q); struct umtxq_chain { struct mtx uc_lock; /* Lock for this chain. */ struct umtx_head uc_queue; /* List of sleep queues. */ }; #define GOLDEN_RATIO_PRIME 2654404609U #define UMTX_CHAINS 128 #define UMTX_SHIFTS (__WORD_BIT - 7) static struct umtxq_chain umtxq_chains[UMTX_CHAINS]; static MALLOC_DEFINE(M_UMTX, "umtx", "UMTX queue memory"); #define UMTX_CONTESTED LONG_MIN static void umtx_init_chains(void *); static int umtxq_hash(struct umtx_key *); static struct mtx *umtxq_mtx(int chain); static void umtxq_lock(struct umtx_key *); static void umtxq_unlock(struct umtx_key *); static void umtxq_insert(struct umtx_q *); static int umtxq_sleep(struct thread *td, struct umtx_key *key, int priority, const char *wmesg, int timo); static int umtxq_count(struct umtx_key *); static void umtxq_signal(struct umtx_key *); static void fork_handler(void *arg, struct proc *p1, struct proc *p2, int flags); SYSINIT(umtx, SI_SUB_EVENTHANDLER+1, SI_ORDER_MIDDLE, umtx_init_chains, NULL); static void umtx_init_chains(void *arg __unused) { int i; for (i = 0; i < UMTX_CHAINS; ++i) { mtx_init(&umtxq_chains[i].uc_lock, "umtxq_lock", NULL, MTX_DEF | MTX_DUPOK); LIST_INIT(&umtxq_chains[i].uc_queue); } EVENTHANDLER_REGISTER(process_fork, fork_handler, 0, 10000); } static void fork_handler(void *arg, struct proc *p1, struct proc *p2, int flags) { struct thread *td; PROC_LOCK(p1); FOREACH_THREAD_IN_PROC(p1, td) { if (td->td_flags & TDF_UMTXQ) wakeup(td); } PROC_UNLOCK(p1); } static inline int umtxq_hash(struct umtx_key *key) { unsigned n = (uintptr_t)key->uk_object + key->uk_offset; return (((n * GOLDEN_RATIO_PRIME) >> UMTX_SHIFTS) % UMTX_CHAINS); } static inline struct mtx * umtxq_mtx(int chain) { return (&umtxq_chains[chain].uc_lock); } static inline void umtxq_lock(struct umtx_key *key) { int chain = umtxq_hash(key); mtx_lock(umtxq_mtx(chain)); } static inline void umtxq_unlock(struct umtx_key *key) { int chain = umtxq_hash(key); mtx_unlock(umtxq_mtx(chain)); } /* * Insert a thread onto the umtx queue. */ static inline void umtxq_insert(struct umtx_q *uq) { struct umtx_head *head; int chain = umtxq_hash(&uq->uq_key); head = &umtxq_chains[chain].uc_queue; LIST_INSERT_HEAD(head, uq, uq_next); mtx_lock_spin(&sched_lock); uq->uq_thread->td_flags |= TDF_UMTXQ; mtx_unlock_spin(&sched_lock); } /* * Remove thread from umtx queue released. */ static inline void umtxq_remove(struct umtx_q *uq) { if (uq->uq_thread->td_flags & TDF_UMTXQ) { LIST_REMOVE(uq, uq_next); mtx_lock_spin(&sched_lock); uq->uq_thread->td_flags &= ~TDF_UMTXQ; mtx_unlock_spin(&sched_lock); } } static int umtxq_count(struct umtx_key *key) { struct umtx_q *uq; struct umtx_head *head; int chain, count = 0; chain = umtxq_hash(key); umtxq_lock(key); head = &umtxq_chains[chain].uc_queue; LIST_FOREACH(uq, head, uq_next) { if (uq->uq_key.uk_object == key->uk_object && uq->uq_key.uk_offset == key->uk_offset) { if (++count > 1) break; } } umtxq_unlock(key); return (count); } static void umtxq_signal(struct umtx_key *key) { struct umtx_q *uq; struct umtx_head *head; struct thread *blocked = NULL; int chain; chain = umtxq_hash(key); umtxq_lock(key); head = &umtxq_chains[chain].uc_queue; LIST_FOREACH(uq, head, uq_next) { if (uq->uq_key.uk_object == key->uk_object && uq->uq_key.uk_offset == key->uk_offset) { blocked = uq->uq_thread; umtxq_remove(uq); break; } } umtxq_unlock(key); if (blocked != NULL) wakeup(blocked); } static inline int umtxq_sleep(struct thread *td, struct umtx_key *key, int priority, const char *wmesg, int timo) { int error; int chain = umtxq_hash(key); error = msleep(td, umtxq_mtx(chain), td->td_priority | PCATCH | PDROP, "umtx", timo); return (error); } static int queue_me(struct thread *td, struct umtx *umtx, struct umtx_q *uq) { vm_map_t map; vm_map_entry_t entry; vm_object_t object; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; map = &td->td_proc->p_vmspace->vm_map; if (vm_map_lookup(&map, (vm_offset_t)umtx, VM_PROT_WRITE, &entry, &object, &pindex, &prot, &wired) != KERN_SUCCESS) { return EFAULT; } uq->uq_key.uk_object = object; uq->uq_key.uk_offset = entry->offset + entry->start - (vm_offset_t)umtx; /* done accessing vm object and map entry */ vm_map_lookup_done(map, entry); uq->uq_addr = (vm_offset_t)umtx; uq->uq_thread = td; umtxq_lock(&uq->uq_key); umtxq_insert(uq); umtxq_unlock(&uq->uq_key); return (0); } static int _do_lock(struct thread *td, struct umtx *umtx, long id, int timo) { struct umtx_q uq; intptr_t owner; intptr_t old; int page_off; int error = 0; /* * Care must be exercised when dealing with umtx structure. It * can fault on any access. */ page_off = ((unsigned long)umtx) % PAGE_SIZE; /* Must not on page boundary. */ if (page_off + sizeof(void *) > PAGE_SIZE) return (EINVAL); for (;;) { /* * Try the uncontested case. This should be done in userland. */ owner = casuptr((intptr_t *)&umtx->u_owner, UMTX_UNOWNED, id); /* The acquire succeeded. */ if (owner == UMTX_UNOWNED) return (0); /* The address was invalid. */ if (owner == -1) return (EFAULT); /* If no one owns it but it is contested try to acquire it. */ if (owner == UMTX_CONTESTED) { owner = casuptr((intptr_t *)&umtx->u_owner, UMTX_CONTESTED, id | UMTX_CONTESTED); if (owner == UMTX_CONTESTED) return (0); /* The address was invalid. */ if (owner == -1) return (EFAULT); /* If this failed the lock has changed, restart. */ continue; } /* * If we caught a signal, we have retried and now * exit immediately. */ if (error || (error = queue_me(td, umtx, &uq)) != 0) return (error); /* * Set the contested bit so that a release in user space * knows to use the system call for unlock. If this fails * either some one else has acquired the lock or it has been * released. */ old = casuptr((intptr_t *)&umtx->u_owner, owner, owner | UMTX_CONTESTED); /* The address was invalid. */ if (old == -1) { umtxq_lock(&uq.uq_key); umtxq_remove(&uq); umtxq_unlock(&uq.uq_key); return (EFAULT); } /* * We set the contested bit, sleep. Otherwise the lock changed * and we need to retry or we lost a race to the thread * unlocking the umtx. */ umtxq_lock(&uq.uq_key); if (old == owner && (td->td_flags & TDF_UMTXQ)) { error = umtxq_sleep(td, &uq.uq_key, td->td_priority | PCATCH | PDROP, "umtx", timo); if (td->td_flags & TDF_UMTXQ) { umtxq_lock(&uq.uq_key); umtxq_remove(&uq); umtxq_unlock(&uq.uq_key); } } else { umtxq_remove(&uq); umtxq_unlock(&uq.uq_key); error = 0; } } return (0); } static int do_lock(struct thread *td, struct umtx *umtx, long id, struct timespec *abstime) { struct timespec ts1, ts2; struct timeval tv; int timo, error; if (abstime == NULL) { error = _do_lock(td, umtx, id, 0); } else { for (;;) { ts1 = *abstime; getnanotime(&ts2); timespecsub(&ts1, &ts2); TIMESPEC_TO_TIMEVAL(&tv, &ts1); if (tv.tv_sec < 0) { error = ETIMEDOUT; break; } timo = tvtohz(&tv); error = _do_lock(td, umtx, id, timo); if (error != ETIMEDOUT) break; } } return (error); } static int do_unlock(struct thread *td, struct umtx *umtx, long id) { struct umtx_key key; vm_map_t map; vm_map_entry_t entry; vm_pindex_t pindex; vm_prot_t prot; boolean_t wired; intptr_t owner; intptr_t old; int count; /* * Make sure we own this mtx. * * XXX Need a {fu,su}ptr this is not correct on arch where * sizeof(intptr_t) != sizeof(long). */ if ((owner = fuword(&umtx->u_owner)) == -1) return (EFAULT); if ((owner & ~UMTX_CONTESTED) != id) return (EPERM); /* We should only ever be in here for contested locks */ if ((owner & UMTX_CONTESTED) == 0) return (EINVAL); /* * When unlocking the umtx, it must be marked as unowned if * there is zero or one thread only waiting for it. * Otherwise, it must be marked as contested. */ old = casuptr((intptr_t *)&umtx->u_owner, owner, UMTX_UNOWNED); if (old == -1) return (EFAULT); if (old != owner) return (EINVAL); map = &td->td_proc->p_vmspace->vm_map; if (vm_map_lookup(&map, (vm_offset_t)umtx, VM_PROT_WRITE, &entry, &key.uk_object, &pindex, &prot, &wired) != KERN_SUCCESS) { return EFAULT; } key.uk_offset = entry->offset + entry->start - (vm_offset_t)umtx; vm_map_lookup_done(map, entry); /* * At the point, a new thread can lock the umtx before we * reach here, so contested bit will not be set, if there * are two or more threads on wait queue, we should set * contensted bit for them. */ count = umtxq_count(&key); if (count <= 0) return (0); /* * If there is second thread waiting on umtx, set contested bit, * if they are resumed before we reach here, it is harmless, * just a bit unefficient. */ if (count > 1) { owner = UMTX_UNOWNED; for (;;) { old = casuptr((intptr_t *)&umtx->u_owner, owner, owner | UMTX_CONTESTED); if (old == owner) break; if (old == -1) return (EFAULT); owner = old; } /* * Another thread locked the umtx before us, so don't bother * to wake more threads, that thread will do it when it unlocks * the umtx. */ if ((owner & ~UMTX_CONTESTED) != 0) return (0); } /* Wake blocked thread. */ umtxq_signal(&key); return (0); } int _umtx_lock(struct thread *td, struct _umtx_lock_args *uap) /* struct umtx *umtx */ { return _do_lock(td, uap->umtx, td->td_tid, 0); } int _umtx_unlock(struct thread *td, struct _umtx_unlock_args *uap) /* struct umtx *umtx */ { return do_unlock(td, uap->umtx, td->td_tid); } int _umtx_op(struct thread *td, struct _umtx_op_args *uap) { struct timespec abstime; struct timespec *ts; int error; switch(uap->op) { case UMTX_LOCK: /* Allow a null timespec (wait forever). */ if (uap->abstime == NULL) ts = NULL; else { error = copyin(uap->abstime, &abstime, sizeof(abstime)); if (error != 0) return (error); if (abstime.tv_nsec >= 1000000000 || abstime.tv_nsec < 0) return (EINVAL); ts = &abstime; } return do_lock(td, uap->umtx, uap->id, ts); case UMTX_UNLOCK: return do_unlock(td, uap->umtx, uap->id); default: return (EINVAL); } }