X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=918d2742a0429683c11e42c60c77187bc908e862;hp=e70980f5b93e259f5506e3fb7e1b602708aa63b3;hb=eaf2c3f4b9041d2a6a85d7f9e56f6014603367d4;hpb=cfe78e252a63a8718e0d2f819d196cee5ca3f6c1 diff --git a/urcu.c b/urcu.c index e70980f..918d274 100644 --- a/urcu.c +++ b/urcu.c @@ -3,7 +3,7 @@ * * Userspace RCU library * - * Copyright (c) 2009 Mathieu Desnoyers + * Copyright (c) 2009 Mathieu Desnoyers * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. * * This library is free software; you can redistribute it and/or @@ -23,6 +23,7 @@ * IBM's contributions to this file may be relicensed under LGPLv2 or later. */ +#define _BSD_SOURCE #include #include #include @@ -36,34 +37,43 @@ /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #include "urcu.h" -#ifndef URCU_MB +#ifdef RCU_MEMBARRIER static int init_done; +int has_sys_membarrier; -void __attribute__((constructor)) urcu_init(void); -void __attribute__((destructor)) urcu_exit(void); -#else -void urcu_init(void) +void __attribute__((constructor)) rcu_init(void); +#endif + +#ifdef RCU_MB +void rcu_init(void) { } #endif -static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; +#ifdef RCU_SIGNAL +static int init_done; + +void __attribute__((constructor)) rcu_init(void); +void __attribute__((destructor)) rcu_exit(void); +#endif + +static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; int gp_futex; /* * Global grace period counter. - * Contains the current RCU_GP_CTR_BIT. + * Contains the current RCU_GP_CTR_PHASE. * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. * Written to only by writer with mutex taken. Read by both writer and readers. */ -long urcu_gp_ctr = RCU_GP_COUNT; +unsigned long rcu_gp_ctr = RCU_GP_COUNT; /* * Written to only by each individual reader. Read by both the reader and the * writers. */ -struct urcu_reader __thread urcu_reader; +struct rcu_reader __thread rcu_reader; #ifdef DEBUG_YIELD unsigned int yield_active; @@ -72,26 +82,26 @@ unsigned int __thread rand_yield; static LIST_HEAD(registry); -static void internal_urcu_lock(void) +static void mutex_lock(pthread_mutex_t *mutex) { int ret; #ifndef DISTRUST_SIGNALS_EXTREME - ret = pthread_mutex_lock(&urcu_mutex); + ret = pthread_mutex_lock(mutex); if (ret) { perror("Error in pthread mutex lock"); exit(-1); } #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ - while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) { + while ((ret = pthread_mutex_trylock(mutex)) != 0) { if (ret != EBUSY && ret != EINTR) { printf("ret = %d, errno = %d\n", ret, errno); perror("Error in pthread mutex lock"); exit(-1); } - if (urcu_reader.need_mb) { + if (LOAD_SHARED(rcu_reader.need_mb)) { smp_mb(); - urcu_reader.need_mb = 0; + _STORE_SHARED(rcu_reader.need_mb, 0); smp_mb(); } poll(NULL,0,10); @@ -99,66 +109,38 @@ static void internal_urcu_lock(void) #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -static void internal_urcu_unlock(void) +static void mutex_unlock(pthread_mutex_t *mutex) { int ret; - ret = pthread_mutex_unlock(&urcu_mutex); + ret = pthread_mutex_unlock(mutex); if (ret) { perror("Error in pthread mutex unlock"); exit(-1); } } -/* - * called with urcu_mutex held. - */ -static void switch_next_urcu_qparity(void) -{ - STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT); -} - -#ifdef URCU_MB -#if 0 /* unused */ -static void force_mb_single_thread(struct urcu_reader *index) +#ifdef RCU_MEMBARRIER +static void smp_mb_master(int group) { - smp_mb(); + if (likely(has_sys_membarrier)) + membarrier(MEMBARRIER_EXPEDITED); + else + smp_mb(); } -#endif //0 +#endif -static void force_mb_all_threads(void) +#ifdef RCU_MB +static void smp_mb_master(int group) { smp_mb(); } -#else /* #ifdef URCU_MB */ -#if 0 /* unused */ -static void force_mb_single_thread(struct urcu_reader *index) -{ - assert(!list_empty(®istry)); - /* - * pthread_kill has a smp_mb(). But beware, we assume it performs - * a cache flush on architectures with non-coherent cache. Let's play - * safe and don't assume anything : we use smp_mc() to make sure the - * cache flush is enforced. - */ - index->need_mb = 1; - smp_mc(); /* write ->need_mb before sending the signals */ - pthread_kill(index->tid, SIGURCU); - smp_mb(); - /* - * Wait for sighandler (and thus mb()) to execute on every thread. - * BUSY-LOOP. - */ - while (index->need_mb) { - poll(NULL, 0, 1); - } - smp_mb(); /* read ->need_mb before ending the barrier */ -} -#endif //0 +#endif -static void force_mb_all_threads(void) +#ifdef RCU_SIGNAL +static void force_mb_all_readers(void) { - struct urcu_reader *index; + struct rcu_reader *index; /* * Ask for each threads to execute a smp_mb() so we can consider the @@ -172,10 +154,9 @@ static void force_mb_all_threads(void) * safe and don't assume anything : we use smp_mc() to make sure the * cache flush is enforced. */ - list_for_each_entry(index, ®istry, head) { - index->need_mb = 1; - smp_mc(); /* write need_mb before sending the signal */ - pthread_kill(index->tid, SIGURCU); + list_for_each_entry(index, ®istry, node) { + STORE_SHARED(index->need_mb, 1); + pthread_kill(index->tid, SIGRCU); } /* * Wait for sighandler (and thus mb()) to execute on every thread. @@ -190,15 +171,20 @@ static void force_mb_all_threads(void) * relevant bug report. For Linux kernels, we recommend getting * the Linux Test Project (LTP). */ - list_for_each_entry(index, ®istry, head) { - while (index->need_mb) { - pthread_kill(index->tid, SIGURCU); + list_for_each_entry(index, ®istry, node) { + while (LOAD_SHARED(index->need_mb)) { + pthread_kill(index->tid, SIGRCU); poll(NULL, 0, 1); } } smp_mb(); /* read ->need_mb before ending the barrier */ } -#endif /* #else #ifdef URCU_MB */ + +static void smp_mb_master(int group) +{ + force_mb_all_readers(); +} +#endif /* #ifdef RCU_SIGNAL */ /* * synchronize_rcu() waiting. Single thread. @@ -206,41 +192,58 @@ static void force_mb_all_threads(void) static void wait_gp(void) { /* Read reader_gp before read futex */ - force_mb_all_threads(); + smp_mb_master(RCU_MB_GROUP); if (uatomic_read(&gp_futex) == -1) - futex(&gp_futex, FUTEX_WAIT, -1, + futex_async(&gp_futex, FUTEX_WAIT, -1, NULL, NULL, 0); } -void wait_for_quiescent_state(void) +void update_counter_and_wait(void) { LIST_HEAD(qsreaders); int wait_loops = 0; - struct urcu_reader *index; + struct rcu_reader *index, *tmp; + + /* Switch parity: 0 -> 1, 1 -> 0 */ + STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); - if (list_empty(®istry)) - return; /* - * Wait for each thread urcu_reader.ctr count to become 0. + * Must commit rcu_gp_ctr update to memory before waiting for quiescent + * state. Failure to do so could result in the writer waiting forever + * while new readers are always accessing data (no progress). Enforce + * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr. + */ + barrier(); + + /* + * + * Adding a smp_mb() which is _not_ formally required, but makes the + * model easier to understand. It does not have a big performance impact + * anyway, given this is the write-side. + */ + smp_mb(); + + /* + * Wait for each thread rcu_reader.ctr count to become 0. */ for (;;) { wait_loops++; if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { uatomic_dec(&gp_futex); /* Write futex before read reader_gp */ - force_mb_all_threads(); + smp_mb_master(RCU_MB_GROUP); } - list_for_each_entry(index, ®istry, head) { - if (!rcu_old_gp_ongoing(&index->ctr)) - list_move(&index->head, &qsreaders); + list_for_each_entry_safe(index, tmp, ®istry, node) { + if (!rcu_gp_ongoing(&index->ctr)) + list_move(&index->node, &qsreaders); } #ifndef HAS_INCOHERENT_CACHES if (list_empty(®istry)) { if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ - force_mb_all_threads(); + smp_mb_master(RCU_MB_GROUP); uatomic_set(&gp_futex, 0); } break; @@ -253,12 +256,12 @@ void wait_for_quiescent_state(void) #else /* #ifndef HAS_INCOHERENT_CACHES */ /* * BUSY-LOOP. Force the reader thread to commit its - * urcu_reader.ctr update to memory if we wait for too long. + * rcu_reader.ctr update to memory if we wait for too long. */ if (list_empty(®istry)) { if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ - force_mb_all_threads(); + smp_mb_master(RCU_MB_GROUP); uatomic_set(&gp_futex, 0); } break; @@ -268,7 +271,7 @@ void wait_for_quiescent_state(void) wait_gp(); break; /* only escape switch */ case KICK_READER_LOOPS: - force_mb_all_threads(); + smp_mb_master(RCU_MB_GROUP); wait_loops = 0; break; /* only escape switch */ default: @@ -278,65 +281,35 @@ void wait_for_quiescent_state(void) #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ } /* put back the reader list in the registry */ - list_move(&qsreaders, ®istry); + list_splice(&qsreaders, ®istry); } void synchronize_rcu(void) { - internal_urcu_lock(); + mutex_lock(&rcu_gp_lock); + + if (list_empty(®istry)) + goto out; /* All threads should read qparity before accessing data structure - * where new ptr points to. Must be done within internal_urcu_lock - * because it iterates on reader threads.*/ + * where new ptr points to. Must be done within rcu_gp_lock because it + * iterates on reader threads.*/ /* Write new ptr before changing the qparity */ - force_mb_all_threads(); - - switch_next_urcu_qparity(); /* 0 -> 1 */ - - /* - * Must commit qparity update to memory before waiting for parity - * 0 quiescent state. Failure to do so could result in the writer - * waiting forever while new readers are always accessing data (no - * progress). - * Ensured by STORE_SHARED and LOAD_SHARED. - */ - - /* - * Adding a smp_mb() which is _not_ formally required, but makes the - * model easier to understand. It does not have a big performance impact - * anyway, given this is the write-side. - */ - smp_mb(); + smp_mb_master(RCU_MB_GROUP); /* * Wait for previous parity to be empty of readers. */ - wait_for_quiescent_state(); /* Wait readers in parity 0 */ + update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ /* * Must finish waiting for quiescent state for parity 0 before - * committing qparity update to memory. Failure to do so could result in - * the writer waiting forever while new readers are always accessing - * data (no progress). - * Ensured by STORE_SHARED and LOAD_SHARED. - */ - - /* - * Adding a smp_mb() which is _not_ formally required, but makes the - * model easier to understand. It does not have a big performance impact - * anyway, given this is the write-side. - */ - smp_mb(); - - switch_next_urcu_qparity(); /* 1 -> 0 */ - - /* - * Must commit qparity update to memory before waiting for parity - * 1 quiescent state. Failure to do so could result in the writer - * waiting forever while new readers are always accessing data (no - * progress). - * Ensured by STORE_SHARED and LOAD_SHARED. + * committing next rcu_gp_ctr update to memory. Failure to do so could + * result in the writer waiting forever while new readers are always + * accessing data (no progress). Enforce compiler-order of load + * rcu_reader ctr before store to rcu_gp_ctr. */ + barrier(); /* * Adding a smp_mb() which is _not_ formally required, but makes the @@ -348,14 +321,14 @@ void synchronize_rcu(void) /* * Wait for previous parity to be empty of readers. */ - wait_for_quiescent_state(); /* Wait readers in parity 1 */ + update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ /* Finish waiting for reader threads before letting the old ptr being - * freed. Must be done within internal_urcu_lock because it iterates on - * reader threads. */ - force_mb_all_threads(); - - internal_urcu_unlock(); + * freed. Must be done within rcu_gp_lock because it iterates on reader + * threads. */ + smp_mb_master(RCU_MB_GROUP); +out: + mutex_unlock(&rcu_gp_lock); } /* @@ -372,59 +345,38 @@ void rcu_read_unlock(void) _rcu_read_unlock(); } -void *rcu_dereference(void *p) -{ - return _rcu_dereference(p); -} - -void *rcu_assign_pointer_sym(void **p, void *v) -{ - wmb(); - return STORE_SHARED(p, v); -} - -void *rcu_xchg_pointer_sym(void **p, void *v) -{ - wmb(); - return uatomic_xchg(p, v); -} - -void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new) -{ - wmb(); - return uatomic_cmpxchg(p, old, _new); -} - -void *rcu_publish_content_sym(void **p, void *v) +void rcu_register_thread(void) { - void *oldptr; - - oldptr = _rcu_xchg_pointer(p, v); - synchronize_rcu(); - return oldptr; + rcu_reader.tid = pthread_self(); + assert(rcu_reader.need_mb == 0); + assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK)); + + mutex_lock(&rcu_gp_lock); + rcu_init(); /* In case gcc does not support constructor attribute */ + list_add(&rcu_reader.node, ®istry); + mutex_unlock(&rcu_gp_lock); } -void rcu_register_thread(void) +void rcu_unregister_thread(void) { - urcu_reader.tid = pthread_self(); - assert(urcu_reader.need_mb == 0); - assert(urcu_reader.ctr == 0); - - internal_urcu_lock(); - urcu_init(); /* In case gcc does not support constructor attribute */ - list_add(&urcu_reader.head, ®istry); - internal_urcu_unlock(); + mutex_lock(&rcu_gp_lock); + list_del(&rcu_reader.node); + mutex_unlock(&rcu_gp_lock); } -void rcu_unregister_thread(void) +#ifdef RCU_MEMBARRIER +void rcu_init(void) { - internal_urcu_lock(); - list_del(&urcu_reader.head); - internal_urcu_unlock(); + if (init_done) + return; + init_done = 1; + if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY)) + has_sys_membarrier = 1; } +#endif -#ifndef URCU_MB -static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) +#ifdef RCU_SIGNAL +static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) { /* * Executing this smp_mb() is the only purpose of this signal handler. @@ -432,19 +384,19 @@ static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) * executed on. */ smp_mb(); - urcu_reader.need_mb = 0; + _STORE_SHARED(rcu_reader.need_mb, 0); smp_mb(); } /* - * urcu_init constructor. Called when the library is linked, but also when + * rcu_init constructor. Called when the library is linked, but also when * reader threads are calling rcu_register_thread(). * Should only be called by a single thread at a given time. This is ensured by - * holing the internal_urcu_lock() from rcu_register_thread() or by running at - * library load time, which should not be executed by multiple threads nor - * concurrently with rcu_register_thread() anyway. + * holing the rcu_gp_lock from rcu_register_thread() or by running at library + * load time, which should not be executed by multiple threads nor concurrently + * with rcu_register_thread() anyway. */ -void urcu_init(void) +void rcu_init(void) { struct sigaction act; int ret; @@ -453,27 +405,27 @@ void urcu_init(void) return; init_done = 1; - act.sa_sigaction = sigurcu_handler; + act.sa_sigaction = sigrcu_handler; act.sa_flags = SA_SIGINFO | SA_RESTART; sigemptyset(&act.sa_mask); - ret = sigaction(SIGURCU, &act, NULL); + ret = sigaction(SIGRCU, &act, NULL); if (ret) { perror("Error in sigaction"); exit(-1); } } -void urcu_exit(void) +void rcu_exit(void) { struct sigaction act; int ret; - ret = sigaction(SIGURCU, NULL, &act); + ret = sigaction(SIGRCU, NULL, &act); if (ret) { perror("Error in sigaction"); exit(-1); } - assert(act.sa_sigaction == sigurcu_handler); + assert(act.sa_sigaction == sigrcu_handler); assert(list_empty(®istry)); } -#endif /* #ifndef URCU_MB */ +#endif /* #ifdef RCU_SIGNAL */