X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=d441355e8a85b85a814c778f52acba7edb47ea2b;hp=b71e162b3fca1e29f83e02b6cd405e68b154147e;hb=e3b0cef0b290dcc757347989a6288c9952efc54a;hpb=5dba80f9bed985c23090ec74bb4173b31b28513a diff --git a/urcu.c b/urcu.c index b71e162..d441355 100644 --- a/urcu.c +++ b/urcu.c @@ -36,17 +36,25 @@ /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #include "urcu.h" +#ifndef URCU_MB +static int init_done; + void __attribute__((constructor)) urcu_init(void); void __attribute__((destructor)) urcu_exit(void); +#else +void urcu_init(void) +{ +} +#endif -int init_done; +static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; -pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; +int gp_futex; /* * Global grace period counter. * Contains the current RCU_GP_CTR_BIT. - * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path. + * 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; @@ -55,27 +63,19 @@ long urcu_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; +struct urcu_reader __thread urcu_reader; /* Thread IDs of registered readers */ #define INIT_NUM_THREADS 4 -struct reader_registry { - pthread_t tid; - long *urcu_active_readers; - char *need_mb; -}; - #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); -void internal_urcu_lock(void) +static void internal_urcu_lock(void) { int ret; @@ -92,9 +92,9 @@ void internal_urcu_lock(void) perror("Error in pthread mutex lock"); exit(-1); } - if (need_mb) { + if (urcu_reader.need_mb) { smp_mb(); - need_mb = 0; + urcu_reader.need_mb = 0; smp_mb(); } poll(NULL,0,10); @@ -102,7 +102,7 @@ void internal_urcu_lock(void) #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -void internal_urcu_unlock(void) +static void internal_urcu_unlock(void) { int ret; @@ -121,30 +121,27 @@ static void switch_next_urcu_qparity(void) STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT); } -#ifdef DEBUG_FULL_MB -#ifdef HAS_INCOHERENT_CACHES -static void force_mb_single_thread(struct reader_registry *index) +#ifdef URCU_MB +static void force_mb_single_thread(struct urcu_reader *index) { smp_mb(); } -#endif /* #ifdef HAS_INCOHERENT_CACHES */ static void force_mb_all_threads(void) { smp_mb(); } -#else /* #ifdef DEBUG_FULL_MB */ -#ifdef HAS_INCOHERENT_CACHES -static void force_mb_single_thread(struct reader_registry *index) +#else /* #ifdef URCU_MB */ +static void force_mb_single_thread(struct urcu_reader *index) { - assert(registry); + 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; + index->need_mb = 1; smp_mc(); /* write ->need_mb before sending the signals */ pthread_kill(index->tid, SIGURCU); smp_mb(); @@ -152,21 +149,21 @@ static void force_mb_single_thread(struct reader_registry *index) * Wait for sighandler (and thus mb()) to execute on every thread. * BUSY-LOOP. */ - while (*index->need_mb) { + while (index->need_mb) { poll(NULL, 0, 1); } smp_mb(); /* read ->need_mb before ending the barrier */ } -#endif /* #ifdef HAS_INCOHERENT_CACHES */ static void force_mb_all_threads(void) { - struct reader_registry *index; + struct urcu_reader *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 @@ -174,8 +171,8 @@ 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); } @@ -192,40 +189,71 @@ 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) { + list_for_each_entry(index, ®istry, head) { + while (index->need_mb) { pthread_kill(index->tid, SIGURCU); poll(NULL, 0, 1); } } smp_mb(); /* read ->need_mb before ending the barrier */ } -#endif /* #else #ifdef DEBUG_FULL_MB */ +#endif /* #else #ifdef URCU_MB */ + +/* + * synchronize_rcu() waiting. Single thread. + */ +static void wait_gp(struct urcu_reader *index) +{ + uatomic_dec(&gp_futex); + force_mb_single_thread(index); /* Write futex before read reader_gp */ + if (!rcu_old_gp_ongoing(&index->ctr)) { + /* Read reader_gp before write futex */ + force_mb_single_thread(index); + /* Callbacks are queued, don't wait. */ + uatomic_set(&gp_futex, 0); + } else { + /* Read reader_gp before read futex */ + force_mb_single_thread(index); + if (uatomic_read(&gp_futex) == -1) + futex(&gp_futex, FUTEX_WAIT, -1, + NULL, NULL, 0); + } +} void wait_for_quiescent_state(void) { - struct reader_registry *index; + struct urcu_reader *index; - if (!registry) + if (list_empty(®istry)) return; /* - * Wait for each thread urcu_active_readers count to become 0. + * Wait for each thread urcu_reader.ctr count to become 0. */ - for (index = registry; index < registry + num_readers; index++) { + list_for_each_entry(index, ®istry, head) { + int wait_loops = 0; #ifndef HAS_INCOHERENT_CACHES - while (rcu_old_gp_ongoing(index->urcu_active_readers)) - cpu_relax(); + while (rcu_old_gp_ongoing(&index->ctr)) { + if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) { + wait_gp(index); + } 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. + * urcu_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) { + while (rcu_old_gp_ongoing(&index->ctr)) { + switch (wait_loops++) { + case RCU_QS_ACTIVE_ATTEMPTS: + wait_gp(index); + break; + case KICK_READER_LOOPS: force_mb_single_thread(index); wait_loops = 0; - } else { + break; + default: cpu_relax(); } } @@ -338,7 +366,13 @@ void *rcu_assign_pointer_sym(void **p, void *v) void *rcu_xchg_pointer_sym(void **p, void *v) { wmb(); - return xchg(p, v); + 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) @@ -350,71 +384,25 @@ void *rcu_publish_content_sym(void **p, void *v) 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()); + urcu_reader.tid = pthread_self(); + assert(urcu_reader.need_mb == 0); + assert(urcu_reader.ctr == 0); + list_add(&urcu_reader.head, ®istry); internal_urcu_unlock(); } void rcu_unregister_thread(void) { internal_urcu_lock(); - rcu_remove_reader(pthread_self()); + list_del(&urcu_reader.head); internal_urcu_unlock(); } -#ifndef DEBUG_FULL_MB +#ifndef URCU_MB static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) { /* @@ -423,7 +411,7 @@ static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) * executed on. */ smp_mb(); - need_mb = 0; + urcu_reader.need_mb = 0; smp_mb(); } @@ -445,6 +433,8 @@ void urcu_init(void) init_done = 1; act.sa_sigaction = sigurcu_handler; + act.sa_flags = SA_SIGINFO | SA_RESTART; + sigemptyset(&act.sa_mask); ret = sigaction(SIGURCU, &act, NULL); if (ret) { perror("Error in sigaction"); @@ -463,6 +453,6 @@ void urcu_exit(void) exit(-1); } assert(act.sa_sigaction == sigurcu_handler); - free(registry); + assert(list_empty(®istry)); } -#endif /* #ifndef DEBUG_FULL_MB */ +#endif /* #ifndef URCU_MB */