X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=53dfbd5286e4ea5a02044fe23a17425fe738cc4f;hp=32f74880492ce40527c8b5b40560249eac485b90;hb=9d7e3f89772f08cca26d727f47d44ecd47c94401;hpb=834a45ba9dc8fcb5c6c51e81674bd2db3bdc2df0 diff --git a/urcu.c b/urcu.c index 32f7488..53dfbd5 100644 --- a/urcu.c +++ b/urcu.c @@ -36,71 +36,62 @@ /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #include "urcu.h" -#ifndef URCU_MB +#ifndef RCU_MB static int init_done; -void __attribute__((constructor)) urcu_init(void); -void __attribute__((destructor)) urcu_exit(void); +void __attribute__((constructor)) rcu_init(void); +void __attribute__((destructor)) rcu_exit(void); #else -void urcu_init(void) +void rcu_init(void) { } #endif -static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; +static pthread_mutex_t rcu_mutex = 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; +long rcu_gp_ctr = RCU_GP_COUNT; /* * Written to only by each individual reader. Read by both the reader and the * writers. */ -long __thread urcu_active_readers; - -/* Thread IDs of registered readers */ -#define INIT_NUM_THREADS 4 - -struct reader_registry { - pthread_t tid; - long *urcu_active_readers; - char *need_mb; -}; +struct rcu_reader __thread rcu_reader; #ifdef DEBUG_YIELD unsigned int yield_active; unsigned int __thread rand_yield; #endif -static struct reader_registry *registry; -static char __thread need_mb; -static int num_readers, alloc_readers; +static LIST_HEAD(registry); -static void internal_urcu_lock(void) +static void internal_rcu_lock(void) { int ret; #ifndef DISTRUST_SIGNALS_EXTREME - ret = pthread_mutex_lock(&urcu_mutex); + ret = pthread_mutex_lock(&rcu_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(&rcu_mutex)) != 0) { if (ret != EBUSY && ret != EINTR) { printf("ret = %d, errno = %d\n", ret, errno); perror("Error in pthread mutex lock"); exit(-1); } - if (need_mb) { + if (rcu_reader.need_mb) { smp_mb(); - need_mb = 0; + rcu_reader.need_mb = 0; smp_mb(); } poll(NULL,0,10); @@ -108,11 +99,11 @@ static void internal_urcu_lock(void) #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -static void internal_urcu_unlock(void) +static void internal_rcu_unlock(void) { int ret; - ret = pthread_mutex_unlock(&urcu_mutex); + ret = pthread_mutex_unlock(&rcu_mutex); if (ret) { perror("Error in pthread mutex unlock"); exit(-1); @@ -120,59 +111,28 @@ static void internal_urcu_unlock(void) } /* - * called with urcu_mutex held. + * called with rcu_mutex held. */ -static void switch_next_urcu_qparity(void) +static void switch_next_rcu_qparity(void) { - STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT); + STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); } -#ifdef URCU_MB -#ifdef HAS_INCOHERENT_CACHES -static void force_mb_single_thread(struct reader_registry *index) +#ifdef RCU_MB +static void smp_mb_heavy() { smp_mb(); } -#endif /* #ifdef HAS_INCOHERENT_CACHES */ - -static void force_mb_all_threads(void) -{ - smp_mb(); -} -#else /* #ifdef URCU_MB */ -#ifdef HAS_INCOHERENT_CACHES -static void force_mb_single_thread(struct reader_registry *index) +#else +static void force_mb_all_readers(void) { - assert(registry); - /* - * 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 /* #ifdef HAS_INCOHERENT_CACHES */ + struct rcu_reader *index; -static void force_mb_all_threads(void) -{ - struct reader_registry *index; /* * Ask for each threads to execute a smp_mb() so we can consider the * compiler barriers around rcu read lock as real memory barriers. */ - if (!registry) + if (list_empty(®istry)) return; /* * pthread_kill has a smp_mb(). But beware, we assume it performs @@ -180,10 +140,10 @@ 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. */ - for (index = registry; index < registry + num_readers; index++) { - *index->need_mb = 1; + 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); + pthread_kill(index->tid, SIGRCU); } /* * Wait for sighandler (and thus mb()) to execute on every thread. @@ -198,58 +158,113 @@ static void force_mb_all_threads(void) * relevant bug report. For Linux kernels, we recommend getting * the Linux Test Project (LTP). */ - for (index = registry; index < registry + num_readers; index++) { - while (*index->need_mb) { - pthread_kill(index->tid, SIGURCU); + list_for_each_entry(index, ®istry, head) { + while (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_heavy() +{ + force_mb_all_readers(); +} +#endif /* #else #ifdef RCU_MB */ + +/* + * synchronize_rcu() waiting. Single thread. + */ +static void wait_gp(void) +{ + /* Read reader_gp before read futex */ + smp_mb_heavy(); + if (uatomic_read(&gp_futex) == -1) + futex_async(&gp_futex, FUTEX_WAIT, -1, + NULL, NULL, 0); +} void wait_for_quiescent_state(void) { - struct reader_registry *index; + LIST_HEAD(qsreaders); + int wait_loops = 0; + struct rcu_reader *index, *tmp; - if (!registry) + if (list_empty(®istry)) return; /* - * Wait for each thread urcu_active_readers count to become 0. + * Wait for each thread rcu_reader.ctr count to become 0. */ - for (index = registry; index < registry + num_readers; index++) { + for (;;) { + wait_loops++; + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + uatomic_dec(&gp_futex); + /* Write futex before read reader_gp */ + smp_mb_heavy(); + } + + list_for_each_entry_safe(index, tmp, ®istry, head) { + if (!rcu_old_gp_ongoing(&index->ctr)) + list_move(&index->head, &qsreaders); + } + #ifndef HAS_INCOHERENT_CACHES - while (rcu_old_gp_ongoing(index->urcu_active_readers)) - cpu_relax(); + if (list_empty(®istry)) { + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + /* Read reader_gp before write futex */ + smp_mb_heavy(); + uatomic_set(&gp_futex, 0); + } + break; + } else { + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) + wait_gp(); + else + cpu_relax(); + } #else /* #ifndef HAS_INCOHERENT_CACHES */ - int wait_loops = 0; /* * BUSY-LOOP. Force the reader thread to commit its - * urcu_active_readers update to memory if we wait for too long. + * rcu_reader.ctr update to memory if we wait for too long. */ - while (rcu_old_gp_ongoing(index->urcu_active_readers)) { - if (wait_loops++ == KICK_READER_LOOPS) { - force_mb_single_thread(index); + if (list_empty(®istry)) { + if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + /* Read reader_gp before write futex */ + smp_mb_heavy(); + uatomic_set(&gp_futex, 0); + } + break; + } else { + switch (wait_loops) { + case RCU_QS_ACTIVE_ATTEMPTS: + wait_gp(); + break; /* only escape switch */ + case KICK_READER_LOOPS: + smp_mb_heavy(); wait_loops = 0; - } else { + break; /* only escape switch */ + default: cpu_relax(); } } #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ } + /* put back the reader list in the registry */ + list_splice(&qsreaders, ®istry); } void synchronize_rcu(void) { - internal_urcu_lock(); + internal_rcu_lock(); /* All threads should read qparity before accessing data structure - * where new ptr points to. Must be done within internal_urcu_lock + * where new ptr points to. Must be done within internal_rcu_lock * because it iterates on reader threads.*/ /* Write new ptr before changing the qparity */ - force_mb_all_threads(); + smp_mb_heavy(); - switch_next_urcu_qparity(); /* 0 -> 1 */ + switch_next_rcu_qparity(); /* 0 -> 1 */ /* * Must commit qparity update to memory before waiting for parity @@ -286,7 +301,7 @@ void synchronize_rcu(void) */ smp_mb(); - switch_next_urcu_qparity(); /* 1 -> 0 */ + switch_next_rcu_qparity(); /* 1 -> 0 */ /* * Must commit qparity update to memory before waiting for parity @@ -309,11 +324,11 @@ void synchronize_rcu(void) wait_for_quiescent_state(); /* 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 + * freed. Must be done within internal_rcu_lock because it iterates on * reader threads. */ - force_mb_all_threads(); + smp_mb_heavy(); - internal_urcu_unlock(); + internal_rcu_unlock(); } /* @@ -330,104 +345,27 @@ 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 xchg(p, v); -} - -void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new) -{ - wmb(); - return cmpxchg(p, old, _new); -} - -void *rcu_publish_content_sym(void **p, void *v) -{ - void *oldptr; - - oldptr = _rcu_xchg_pointer(p, v); - synchronize_rcu(); - return oldptr; -} - -static void rcu_add_reader(pthread_t id) -{ - struct reader_registry *oldarray; - - if (!registry) { - alloc_readers = INIT_NUM_THREADS; - num_readers = 0; - registry = - malloc(sizeof(struct reader_registry) * alloc_readers); - } - if (alloc_readers < num_readers + 1) { - oldarray = registry; - registry = malloc(sizeof(struct reader_registry) - * (alloc_readers << 1)); - memcpy(registry, oldarray, - sizeof(struct reader_registry) * alloc_readers); - alloc_readers <<= 1; - free(oldarray); - } - registry[num_readers].tid = id; - /* reference to the TLS of _this_ reader thread. */ - registry[num_readers].urcu_active_readers = &urcu_active_readers; - registry[num_readers].need_mb = &need_mb; - num_readers++; -} - -/* - * Never shrink (implementation limitation). - * This is O(nb threads). Eventually use a hash table. - */ -static void rcu_remove_reader(pthread_t id) -{ - struct reader_registry *index; - - assert(registry != NULL); - for (index = registry; index < registry + num_readers; index++) { - if (pthread_equal(index->tid, id)) { - memcpy(index, ®istry[num_readers - 1], - sizeof(struct reader_registry)); - registry[num_readers - 1].tid = 0; - registry[num_readers - 1].urcu_active_readers = NULL; - num_readers--; - return; - } - } - /* Hrm not found, forgot to register ? */ - assert(0); -} - void rcu_register_thread(void) { - internal_urcu_lock(); - urcu_init(); /* In case gcc does not support constructor attribute */ - rcu_add_reader(pthread_self()); - internal_urcu_unlock(); + rcu_reader.tid = pthread_self(); + assert(rcu_reader.need_mb == 0); + assert(rcu_reader.ctr == 0); + + internal_rcu_lock(); + rcu_init(); /* In case gcc does not support constructor attribute */ + list_add(&rcu_reader.head, ®istry); + internal_rcu_unlock(); } void rcu_unregister_thread(void) { - internal_urcu_lock(); - rcu_remove_reader(pthread_self()); - internal_urcu_unlock(); + internal_rcu_lock(); + list_del(&rcu_reader.head); + internal_rcu_unlock(); } -#ifndef URCU_MB -static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) +#ifndef RCU_MB +static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) { /* * Executing this smp_mb() is the only purpose of this signal handler. @@ -435,19 +373,19 @@ static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) * executed on. */ smp_mb(); - need_mb = 0; + 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 + * holing the internal_rcu_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; @@ -456,27 +394,27 @@ 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); - free(registry); + assert(act.sa_sigaction == sigrcu_handler); + assert(list_empty(®istry)); } -#endif /* #ifndef URCU_MB */ +#endif /* #ifndef RCU_MB */