X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=07661a3cb448b34505f24a8552a421741bb282bd;hp=337f7640b44d208855277f6fc53d9322f212d638;hb=1e52eccf3cb657dccb197132f7cb47b683902538;hpb=54843abcc17c8e8b7600ed635e966c6970d8d20f diff --git a/urcu.c b/urcu.c index 337f764..07661a3 100644 --- a/urcu.c +++ b/urcu.c @@ -3,9 +3,22 @@ * * Userspace RCU library * - * Copyright February 2009 - Mathieu Desnoyers + * Copyright (c) 2009 Mathieu Desnoyers + * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. * - * Distributed under GPLv2 + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA * * IBM's contributions to this file may be relicensed under LGPLv2 or later. */ @@ -19,14 +32,29 @@ #include #include +#include "urcu-static.h" +/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #include "urcu.h" -pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; +#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 + +static 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; @@ -55,7 +83,7 @@ static struct reader_registry *registry; static char __thread need_mb; static int num_readers, alloc_readers; -void internal_urcu_lock(void) +static void internal_urcu_lock(void) { int ret; @@ -82,7 +110,7 @@ void internal_urcu_lock(void) #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -void internal_urcu_unlock(void) +static void internal_urcu_unlock(void) { int ret; @@ -101,20 +129,17 @@ 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 +#ifdef URCU_MB static void force_mb_single_thread(struct reader_registry *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 +#else /* #ifdef URCU_MB */ static void force_mb_single_thread(struct reader_registry *index) { assert(registry); @@ -137,7 +162,6 @@ static void force_mb_single_thread(struct reader_registry *index) } smp_mb(); /* read ->need_mb before ending the barrier */ } -#endif /* #ifdef HAS_INCOHERENT_CACHES */ static void force_mb_all_threads(void) { @@ -180,7 +204,28 @@ static void force_mb_all_threads(void) } 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 reader_registry *index) +{ + atomic_dec(&gp_futex); + force_mb_single_thread(index); /* Write futex before read reader_gp */ + if (!rcu_old_gp_ongoing(index->urcu_active_readers)) { + /* Read reader_gp before write futex */ + force_mb_single_thread(index); + /* Callbacks are queued, don't wait. */ + atomic_set(&gp_futex, 0); + } else { + /* Read reader_gp before read futex */ + force_mb_single_thread(index); + if (atomic_read(&gp_futex) == -1) + futex(&gp_futex, FUTEX_WAIT, -1, + NULL, NULL, 0); + } +} void wait_for_quiescent_state(void) { @@ -192,20 +237,30 @@ void wait_for_quiescent_state(void) * Wait for each thread urcu_active_readers count to become 0. */ for (index = registry; index < registry + num_readers; index++) { + 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->urcu_active_readers)) { + 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. */ while (rcu_old_gp_ongoing(index->urcu_active_readers)) { - if (wait_loops++ == KICK_READER_LOOPS) { + 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(); } } @@ -233,6 +288,13 @@ void synchronize_rcu(void) * 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. */ @@ -246,6 +308,13 @@ void synchronize_rcu(void) * 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 */ /* @@ -256,6 +325,13 @@ void synchronize_rcu(void) * 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. */ @@ -269,7 +345,53 @@ void synchronize_rcu(void) internal_urcu_unlock(); } -void urcu_add_reader(pthread_t id) +/* + * library wrappers to be used by non-LGPL compatible source code. + */ + +void rcu_read_lock(void) +{ + _rcu_read_lock(); +} + +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; @@ -299,7 +421,7 @@ void urcu_add_reader(pthread_t id) * Never shrink (implementation limitation). * This is O(nb threads). Eventually use a hash table. */ -void urcu_remove_reader(pthread_t id) +static void rcu_remove_reader(pthread_t id) { struct reader_registry *index; @@ -318,22 +440,23 @@ void urcu_remove_reader(pthread_t id) assert(0); } -void urcu_register_thread(void) +void rcu_register_thread(void) { internal_urcu_lock(); - urcu_add_reader(pthread_self()); + urcu_init(); /* In case gcc does not support constructor attribute */ + rcu_add_reader(pthread_self()); internal_urcu_unlock(); } -void urcu_unregister_thread(void) +void rcu_unregister_thread(void) { internal_urcu_lock(); - urcu_remove_reader(pthread_self()); + rcu_remove_reader(pthread_self()); internal_urcu_unlock(); } -#ifndef DEBUG_FULL_MB -void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) +#ifndef URCU_MB +static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) { /* * Executing this smp_mb() is the only purpose of this signal handler. @@ -345,12 +468,26 @@ void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) smp_mb(); } -void __attribute__((constructor)) urcu_init(void) +/* + * urcu_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. + */ +void urcu_init(void) { struct sigaction act; int ret; + if (init_done) + return; + 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"); @@ -358,7 +495,7 @@ void __attribute__((constructor)) urcu_init(void) } } -void __attribute__((destructor)) urcu_exit(void) +void urcu_exit(void) { struct sigaction act; int ret; @@ -371,4 +508,4 @@ void __attribute__((destructor)) urcu_exit(void) assert(act.sa_sigaction == sigurcu_handler); free(registry); } -#endif /* #ifndef DEBUG_FULL_MB */ +#endif /* #ifndef URCU_MB */