X-Git-Url: https://git.lttng.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=c07df2e8cf21384a07da3bb47066cd2b2bb44084;hp=475b4d1c991a00832409ae7332e6317442b1f042;hb=c297c21c6eadc359a358d33e65f9f5419b55b586;hpb=0a52082baa5dd26455966c42a96f3185767590b7 diff --git a/urcu.c b/urcu.c index 475b4d1..c07df2e 100644 --- a/urcu.c +++ b/urcu.c @@ -3,9 +3,24 @@ * * 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. */ #include @@ -14,9 +29,24 @@ #include #include #include +#include +#include +#include "urcu-static.h" +/* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #include "urcu.h" +#ifndef DEBUG_FULL_MB +void __attribute__((constructor)) urcu_init(void); +void __attribute__((destructor)) urcu_exit(void); +#else +static inline urcu_init(void) +{ +} +#endif + +int init_done; + pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; /* @@ -39,6 +69,7 @@ long __thread urcu_active_readers; struct reader_registry { pthread_t tid; long *urcu_active_readers; + char *need_mb; }; #ifdef DEBUG_YIELD @@ -47,24 +78,34 @@ unsigned int __thread rand_yield; #endif static struct reader_registry *registry; +static char __thread need_mb; static int num_readers, alloc_readers; -#ifndef DEBUG_FULL_MB -static int sig_done; -#endif void internal_urcu_lock(void) { -#if 0 int ret; - /* Mutex sleeping does not play well with busy-waiting loop. */ + +#ifndef DISTRUST_SIGNALS_EXTREME ret = pthread_mutex_lock(&urcu_mutex); if (ret) { perror("Error in pthread mutex lock"); exit(-1); } -#endif - while (pthread_mutex_trylock(&urcu_mutex) != 0) - cpu_relax(); +#else /* #ifndef DISTRUST_SIGNALS_EXTREME */ + while ((ret = pthread_mutex_trylock(&urcu_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) { + smp_mb(); + need_mb = 0; + smp_mb(); + } + poll(NULL,0,10); + } +#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } void internal_urcu_unlock(void) @@ -87,38 +128,42 @@ static void switch_next_urcu_qparity(void) } #ifdef DEBUG_FULL_MB -static void force_mb_single_thread(pthread_t tid) +#ifdef HAS_INCOHERENT_CACHES +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 - -static void force_mb_single_thread(pthread_t tid) +#else /* #ifdef DEBUG_FULL_MB */ +#ifdef HAS_INCOHERENT_CACHES +static void force_mb_single_thread(struct reader_registry *index) { assert(registry); - sig_done = 0; /* * 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. - * smp_mb(); write sig_done before sending the signals */ - smp_mc(); /* write sig_done before sending the signals */ - pthread_kill(tid, SIGURCU); + *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 (LOAD_SHARED(sig_done) < 1) - cpu_relax(); - smp_mb(); /* read sig_done before ending the barrier */ + 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) { @@ -129,26 +174,39 @@ static void force_mb_all_threads(void) */ if (!registry) return; - sig_done = 0; /* * 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. - * smp_mb(); write sig_done before sending the signals */ - smp_mc(); /* write sig_done before sending the signals */ - for (index = registry; index < registry + num_readers; index++) + for (index = registry; index < registry + num_readers; index++) { + *index->need_mb = 1; + smp_mc(); /* write need_mb before sending the signal */ pthread_kill(index->tid, SIGURCU); + } /* * Wait for sighandler (and thus mb()) to execute on every thread. - * BUSY-LOOP. + * + * Note that the pthread_kill() will never be executed on systems + * that correctly deliver signals in a timely manner. However, it + * is not uncommon for kernels to have bugs that can result in + * lost or unduly delayed signals. + * + * If you are seeing the below pthread_kill() executing much at + * all, we suggest testing the underlying kernel and filing the + * relevant bug report. For Linux kernels, we recommend getting + * the Linux Test Project (LTP). */ - while (LOAD_SHARED(sig_done) < num_readers) - cpu_relax(); - smp_mb(); /* read sig_done before ending the barrier */ + for (index = registry; index < registry + num_readers; index++) { + while (*index->need_mb) { + pthread_kill(index->tid, SIGURCU); + poll(NULL, 0, 1); + } + } + smp_mb(); /* read ->need_mb before ending the barrier */ } -#endif +#endif /* #else #ifdef DEBUG_FULL_MB */ void wait_for_quiescent_state(void) { @@ -160,6 +218,10 @@ 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++) { +#ifndef HAS_INCOHERENT_CACHES + while (rcu_old_gp_ongoing(index->urcu_active_readers)) + cpu_relax(); +#else /* #ifndef HAS_INCOHERENT_CACHES */ int wait_loops = 0; /* * BUSY-LOOP. Force the reader thread to commit its @@ -167,12 +229,13 @@ void wait_for_quiescent_state(void) */ while (rcu_old_gp_ongoing(index->urcu_active_readers)) { if (wait_loops++ == KICK_READER_LOOPS) { - force_mb_single_thread(index->tid); + force_mb_single_thread(index); wait_loops = 0; } else { cpu_relax(); } } +#endif /* #else #ifndef HAS_INCOHERENT_CACHES */ } } @@ -196,6 +259,13 @@ 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. */ @@ -209,6 +279,13 @@ 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 */ /* @@ -219,6 +296,13 @@ 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. */ @@ -232,7 +316,47 @@ 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_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; @@ -254,6 +378,7 @@ void urcu_add_reader(pthread_t id) 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++; } @@ -261,7 +386,7 @@ 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; @@ -280,22 +405,23 @@ 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) +static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) { /* * Executing this smp_mb() is the only purpose of this signal handler. @@ -303,15 +429,30 @@ void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) * executed on. */ smp_mb(); - atomic_inc(&sig_done); + need_mb = 0; + 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; + sigemptyset(&act.sa_mask); ret = sigaction(SIGURCU, &act, NULL); if (ret) { perror("Error in sigaction"); @@ -319,7 +460,7 @@ void __attribute__((constructor)) urcu_init(void) } } -void __attribute__((destructor)) urcu_exit(void) +void urcu_exit(void) { struct sigaction act; int ret; @@ -332,4 +473,4 @@ void __attribute__((destructor)) urcu_exit(void) assert(act.sa_sigaction == sigurcu_handler); free(registry); } -#endif +#endif /* #ifndef DEBUG_FULL_MB */