X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=53dfbd5286e4ea5a02044fe23a17425fe738cc4f;hp=9542b264a6dbf53d36d2fb63e7578b4e7ff2b2f5;hb=9d7e3f89772f08cca26d727f47d44ecd47c94401;hpb=9598a4814c854780e9ca9bb2cfff8d77442c3db6 diff --git a/urcu.c b/urcu.c index 9542b26..53dfbd5 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 @@ -14,52 +29,81 @@ #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" -pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER; +#ifndef RCU_MB +static int init_done; + +void __attribute__((constructor)) rcu_init(void); +void __attribute__((destructor)) rcu_exit(void); +#else +void rcu_init(void) +{ +} +#endif + +static pthread_mutex_t rcu_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. + * 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 __thread urcu_active_readers; - -/* Thread IDs of registered readers */ -#define INIT_NUM_THREADS 4 +long rcu_gp_ctr = RCU_GP_COUNT; -struct reader_data { - pthread_t tid; - long *urcu_active_readers; -}; +/* + * Written to only by each individual reader. Read by both the reader and the + * writers. + */ +struct rcu_reader __thread rcu_reader; #ifdef DEBUG_YIELD unsigned int yield_active; unsigned int __thread rand_yield; #endif -static struct reader_data *reader_data; -static int num_readers, alloc_readers; -static int sig_done; +static LIST_HEAD(registry); -void internal_urcu_lock(void) +static void internal_rcu_lock(void) { int ret; - ret = pthread_mutex_lock(&urcu_mutex); + +#ifndef DISTRUST_SIGNALS_EXTREME + 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(&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 (rcu_reader.need_mb) { + smp_mb(); + rcu_reader.need_mb = 0; + smp_mb(); + } + poll(NULL,0,10); + } +#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } -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); @@ -67,223 +111,310 @@ 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) { - urcu_gp_ctr ^= RCU_GP_CTR_BIT; + STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); } -#ifdef DEBUG_FULL_MB -static void force_mb_all_threads(void) +#ifdef RCU_MB +static void smp_mb_heavy() { - mb(); + smp_mb(); } #else -static void force_mb_all_threads(void) +static void force_mb_all_readers(void) { - struct reader_data *index; + struct rcu_reader *index; + /* - * Ask for each threads to execute a mb() so we can consider the + * 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 (!reader_data) + if (list_empty(®istry)) return; - debug_yield_write(); - sig_done = 0; - debug_yield_write(); - mb(); /* write sig_done before sending the signals */ - debug_yield_write(); - for (index = reader_data; index < reader_data + num_readers; index++) { - pthread_kill(index->tid, SIGURCU); - debug_yield_write(); + /* + * 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. + */ + list_for_each_entry(index, ®istry, head) { + index->need_mb = 1; + smp_mc(); /* write need_mb before sending the signal */ + pthread_kill(index->tid, SIGRCU); } /* * 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 (sig_done < num_readers) - barrier(); - debug_yield_write(); - mb(); /* read sig_done before ending the barrier */ - debug_yield_write(); + 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 */ +} + +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); } -#endif void wait_for_quiescent_state(void) { - struct reader_data *index; + LIST_HEAD(qsreaders); + int wait_loops = 0; + struct rcu_reader *index, *tmp; - if (!reader_data) + 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 = reader_data; index < reader_data + 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 + 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 */ /* - * BUSY-LOOP. + * BUSY-LOOP. Force the reader thread to commit its + * rcu_reader.ctr update to memory if we wait for too long. */ - while (rcu_old_gp_ongoing(index->urcu_active_readers)) - barrier(); + 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; + 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_rcu_lock(); + /* All threads should read qparity before accessing data structure - * where new ptr points to. */ + * 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(); - debug_yield_write(); + smp_mb_heavy(); - internal_urcu_lock(); - debug_yield_write(); - - switch_next_urcu_qparity(); /* 0 -> 1 */ - debug_yield_write(); + 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. */ - mb(); + smp_mb(); /* * Wait for previous parity to be empty of readers. */ wait_for_quiescent_state(); /* Wait readers in parity 0 */ - debug_yield_write(); /* * Must finish waiting for quiescent state for parity 0 before * committing qparity update to memory. 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. */ - mb(); - switch_next_urcu_qparity(); /* 1 -> 0 */ - debug_yield_write(); + /* + * 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_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. */ - mb(); + smp_mb(); /* * Wait for previous parity to be empty of readers. */ wait_for_quiescent_state(); /* Wait readers in parity 1 */ - debug_yield_write(); - internal_urcu_unlock(); - debug_yield_write(); - - /* All threads should finish using the data referred to by old ptr - * before decrementing their urcu_active_readers count */ /* Finish waiting for reader threads before letting the old ptr being - * freed. */ - force_mb_all_threads(); - debug_yield_write(); -} + * freed. Must be done within internal_rcu_lock because it iterates on + * reader threads. */ + smp_mb_heavy(); -void urcu_add_reader(pthread_t id) -{ - struct reader_data *oldarray; - - if (!reader_data) { - alloc_readers = INIT_NUM_THREADS; - num_readers = 0; - reader_data = - malloc(sizeof(struct reader_data) * alloc_readers); - } - if (alloc_readers < num_readers + 1) { - oldarray = reader_data; - reader_data = malloc(sizeof(struct reader_data) - * (alloc_readers << 1)); - memcpy(reader_data, oldarray, - sizeof(struct reader_data) * alloc_readers); - alloc_readers <<= 1; - free(oldarray); - } - reader_data[num_readers].tid = id; - /* reference to the TLS of _this_ reader thread. */ - reader_data[num_readers].urcu_active_readers = &urcu_active_readers; - num_readers++; + internal_rcu_unlock(); } /* - * Never shrink (implementation limitation). - * This is O(nb threads). Eventually use a hash table. + * library wrappers to be used by non-LGPL compatible source code. */ -void urcu_remove_reader(pthread_t id) + +void rcu_read_lock(void) { - struct reader_data *index; - - assert(reader_data != NULL); - for (index = reader_data; index < reader_data + num_readers; index++) { - if (pthread_equal(index->tid, id)) { - memcpy(index, &reader_data[num_readers - 1], - sizeof(struct reader_data)); - reader_data[num_readers - 1].tid = 0; - reader_data[num_readers - 1].urcu_active_readers = NULL; - num_readers--; - return; - } - } - /* Hrm not found, forgot to register ? */ - assert(0); + _rcu_read_lock(); } -void urcu_register_thread(void) +void rcu_read_unlock(void) { - internal_urcu_lock(); - urcu_add_reader(pthread_self()); - internal_urcu_unlock(); + _rcu_read_unlock(); } -void urcu_unregister_thread(void) +void rcu_register_thread(void) { - internal_urcu_lock(); - urcu_remove_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(); } -#ifndef DEBUG_FULL_MB -void sigurcu_handler(int signo, siginfo_t *siginfo, void *context) +void rcu_unregister_thread(void) { - mb(); - atomic_inc(&sig_done); + internal_rcu_lock(); + list_del(&rcu_reader.head); + internal_rcu_unlock(); } -void __attribute__((constructor)) urcu_init(void) +#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. + * It punctually promotes barrier() into smp_mb() on every thread it is + * executed on. + */ + smp_mb(); + rcu_reader.need_mb = 0; + smp_mb(); +} + +/* + * 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. + */ +void rcu_init(void) { struct sigaction act; int ret; - act.sa_sigaction = sigurcu_handler; - ret = sigaction(SIGURCU, &act, NULL); + if (init_done) + return; + init_done = 1; + + act.sa_sigaction = sigrcu_handler; + act.sa_flags = SA_SIGINFO | SA_RESTART; + sigemptyset(&act.sa_mask); + ret = sigaction(SIGRCU, &act, NULL); if (ret) { perror("Error in sigaction"); exit(-1); } } -void __attribute__((destructor)) 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(reader_data); + assert(act.sa_sigaction == sigrcu_handler); + assert(list_empty(®istry)); } -#endif +#endif /* #ifndef RCU_MB */