X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=140cd19b98add54f276755e08b13bfad9e8c3d0c;hp=3132e59277bf782f9e2e03cb3b0f862981b80a2a;hb=27d65bc5db18f8563a9120924569acffb845ec4d;hpb=78ff941908d1c336106c06649c0892015cc98e18 diff --git a/urcu.c b/urcu.c index 3132e59..140cd19 100644 --- a/urcu.c +++ b/urcu.c @@ -23,6 +23,7 @@ * IBM's contributions to this file may be relicensed under LGPLv2 or later. */ +#define _BSD_SOURCE #include #include #include @@ -36,18 +37,27 @@ /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #include "urcu.h" -#ifndef RCU_MB +#ifdef RCU_MEMBARRIER static int init_done; +int has_sys_membarrier; void __attribute__((constructor)) rcu_init(void); -void __attribute__((destructor)) rcu_exit(void); -#else +#endif + +#ifdef RCU_MB void rcu_init(void) { } #endif -static pthread_mutex_t rcu_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; @@ -57,7 +67,7 @@ int gp_futex; * 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 rcu_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 @@ -72,18 +82,18 @@ unsigned int __thread rand_yield; static LIST_HEAD(registry); -static void internal_rcu_lock(void) +static void mutex_lock(pthread_mutex_t *mutex) { int ret; #ifndef DISTRUST_SIGNALS_EXTREME - ret = pthread_mutex_lock(&rcu_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(&rcu_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"); @@ -99,63 +109,35 @@ static void internal_rcu_lock(void) #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -static void internal_rcu_unlock(void) +static void mutex_unlock(pthread_mutex_t *mutex) { int ret; - ret = pthread_mutex_unlock(&rcu_mutex); + ret = pthread_mutex_unlock(mutex); if (ret) { perror("Error in pthread mutex unlock"); exit(-1); } } -/* - * called with rcu_mutex held. - */ -static void switch_next_rcu_qparity(void) +#ifdef RCU_MEMBARRIER +static void smp_mb_master(int group) { - STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); + if (likely(has_sys_membarrier)) + membarrier(MEMBARRIER_EXPEDITED); + else + smp_mb(); } +#endif #ifdef RCU_MB -#if 0 /* unused */ -static void force_mb_single_thread(struct rcu_reader *index) +static void smp_mb_master(int group) { smp_mb(); } -#endif //0 - -static void force_mb_all_readers(void) -{ - smp_mb(); -} -#else /* #ifdef RCU_MB */ -#if 0 /* unused */ -static void force_mb_single_thread(struct rcu_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, SIGRCU); - 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 +#ifdef RCU_SIGNAL static void force_mb_all_readers(void) { struct rcu_reader *index; @@ -198,7 +180,12 @@ static void force_mb_all_readers(void) } smp_mb(); /* read ->need_mb before ending the barrier */ } -#endif /* #else #ifdef RCU_MB */ + +static void smp_mb_master(int group) +{ + force_mb_all_readers(); +} +#endif /* #ifdef RCU_SIGNAL */ /* * synchronize_rcu() waiting. Single thread. @@ -206,20 +193,35 @@ static void force_mb_all_readers(void) static void wait_gp(void) { /* Read reader_gp before read futex */ - force_mb_all_readers(); + smp_mb_master(RCU_MB_GROUP); if (uatomic_read(&gp_futex) == -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 rcu_reader *index, *tmp; - if (list_empty(®istry)) - return; + /* Switch parity: 0 -> 1, 1 -> 0 */ + STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); + + /* + * Must commit qparity update to memory before waiting for other parity + * 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(); + /* * Wait for each thread rcu_reader.ctr count to become 0. */ @@ -228,7 +230,7 @@ void wait_for_quiescent_state(void) if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { uatomic_dec(&gp_futex); /* Write futex before read reader_gp */ - force_mb_all_readers(); + smp_mb_master(RCU_MB_GROUP); } list_for_each_entry_safe(index, tmp, ®istry, head) { @@ -240,7 +242,7 @@ void wait_for_quiescent_state(void) if (list_empty(®istry)) { if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ - force_mb_all_readers(); + smp_mb_master(RCU_MB_GROUP); uatomic_set(&gp_futex, 0); } break; @@ -258,7 +260,7 @@ void wait_for_quiescent_state(void) if (list_empty(®istry)) { if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ - force_mb_all_readers(); + smp_mb_master(RCU_MB_GROUP); uatomic_set(&gp_futex, 0); } break; @@ -268,7 +270,7 @@ void wait_for_quiescent_state(void) wait_gp(); break; /* only escape switch */ case KICK_READER_LOOPS: - force_mb_all_readers(); + smp_mb_master(RCU_MB_GROUP); wait_loops = 0; break; /* only escape switch */ default: @@ -283,35 +285,21 @@ void wait_for_quiescent_state(void) void synchronize_rcu(void) { - internal_rcu_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_rcu_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_readers(); - - switch_next_rcu_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 @@ -328,34 +316,17 @@ void synchronize_rcu(void) */ smp_mb(); - switch_next_rcu_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. - */ - - /* - * 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 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_rcu_lock because it iterates on - * reader threads. */ - force_mb_all_readers(); - - internal_rcu_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); } /* @@ -378,20 +349,31 @@ void rcu_register_thread(void) assert(rcu_reader.need_mb == 0); assert(rcu_reader.ctr == 0); - internal_rcu_lock(); + mutex_lock(&rcu_gp_lock); rcu_init(); /* In case gcc does not support constructor attribute */ list_add(&rcu_reader.head, ®istry); - internal_rcu_unlock(); + mutex_unlock(&rcu_gp_lock); } void rcu_unregister_thread(void) { - internal_rcu_lock(); + mutex_lock(&rcu_gp_lock); list_del(&rcu_reader.head); - internal_rcu_unlock(); + mutex_unlock(&rcu_gp_lock); } -#ifndef RCU_MB +#ifdef RCU_MEMBARRIER +void rcu_init(void) +{ + if (init_done) + return; + init_done = 1; + if (!membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_QUERY)) + has_sys_membarrier = 1; +} +#endif + +#ifdef RCU_SIGNAL static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) { /* @@ -408,9 +390,9 @@ static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) * 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_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. + * 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 rcu_init(void) { @@ -444,4 +426,4 @@ void rcu_exit(void) assert(act.sa_sigaction == sigrcu_handler); assert(list_empty(®istry)); } -#endif /* #ifndef RCU_MB */ +#endif /* #ifdef RCU_SIGNAL */