X-Git-Url: http://git.liburcu.org/?p=userspace-rcu.git;a=blobdiff_plain;f=urcu.c;h=e9df01bc7b56cdb25b6e225e6ba46d2182ab28be;hp=a3df976e3cea8fd8269d7cb3158b6e38c18fab01;hb=refs%2Fheads%2Fstable-0.7;hpb=18c3778a6e548f0926cada96db00f8da7cc81454 diff --git a/urcu.c b/urcu.c index a3df976..e9df01b 100644 --- a/urcu.c +++ b/urcu.c @@ -3,7 +3,7 @@ * * Userspace RCU library * - * Copyright (c) 2009 Mathieu Desnoyers + * Copyright (c) 2009 Mathieu Desnoyers * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. * * This library is free software; you can redistribute it and/or @@ -24,18 +24,43 @@ */ #define _BSD_SOURCE +#define _GNU_SOURCE +#define _LGPL_SOURCE +#define _DEFAULT_SOURCE #include #include #include #include #include +#include #include #include #include -#include "urcu-static.h" +#include "urcu/wfqueue.h" +#include "urcu/map/urcu.h" +#include "urcu/static/urcu.h" +#include "urcu-pointer.h" +#include "urcu/tls-compat.h" + +#include "urcu-die.h" + /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ +#undef _LGPL_SOURCE #include "urcu.h" +#define _LGPL_SOURCE + +/* + * If a reader is really non-cooperative and refuses to commit its + * rcu_active_readers count to memory (there is no barrier in the reader + * per-se), kick it after 10 loops waiting for it. + */ +#define KICK_READER_LOOPS 10 + +/* + * Active attempts to check for reader Q.S. before calling futex(). + */ +#define RCU_QS_ACTIVE_ATTEMPTS 100 #ifdef RCU_MEMBARRIER static int init_done; @@ -57,9 +82,23 @@ void __attribute__((constructor)) rcu_init(void); void __attribute__((destructor)) rcu_exit(void); #endif +/* + * rcu_gp_lock ensures mutual exclusion between threads calling + * synchronize_rcu(). + */ static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; +/* + * rcu_registry_lock ensures mutual exclusion between threads + * registering and unregistering themselves to/from the registry, and + * with threads reading that registry from synchronize_rcu(). However, + * this lock is not held all the way through the completion of awaiting + * for the grace period. It is sporadically released between iterations + * on the registry. + * rcu_registry_lock may nest inside rcu_gp_lock. + */ +static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; -int gp_futex; +int32_t gp_futex; /* * Global grace period counter. @@ -68,19 +107,18 @@ int gp_futex; * Written to only by writer with mutex taken. Read by both writer and readers. */ unsigned long rcu_gp_ctr = RCU_GP_COUNT; - /* * Written to only by each individual reader. Read by both the reader and the * writers. */ -struct rcu_reader __thread rcu_reader; +__DEFINE_URCU_TLS_GLOBAL(struct rcu_reader, rcu_reader); #ifdef DEBUG_YIELD unsigned int yield_active; -unsigned int __thread rand_yield; +__DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield); #endif -static LIST_HEAD(registry); +static CDS_LIST_HEAD(registry); static void mutex_lock(pthread_mutex_t *mutex) { @@ -88,21 +126,16 @@ static void mutex_lock(pthread_mutex_t *mutex) #ifndef DISTRUST_SIGNALS_EXTREME ret = pthread_mutex_lock(mutex); - if (ret) { - perror("Error in pthread mutex lock"); - exit(-1); - } + if (ret) + urcu_die(ret); #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ 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"); - exit(-1); - } - if (LOAD_SHARED(rcu_reader.need_mb)) { - smp_mb(); - _STORE_SHARED(rcu_reader.need_mb, 0); - smp_mb(); + if (ret != EBUSY && ret != EINTR) + urcu_die(ret); + if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader).need_mb)) { + cmm_smp_mb(); + _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0); + cmm_smp_mb(); } poll(NULL,0,10); } @@ -114,26 +147,24 @@ static void mutex_unlock(pthread_mutex_t *mutex) int ret; ret = pthread_mutex_unlock(mutex); - if (ret) { - perror("Error in pthread mutex unlock"); - exit(-1); - } + if (ret) + urcu_die(ret); } #ifdef RCU_MEMBARRIER static void smp_mb_master(int group) { - if (likely(has_sys_membarrier)) + if (caa_likely(has_sys_membarrier)) membarrier(MEMBARRIER_EXPEDITED); else - smp_mb(); + cmm_smp_mb(); } #endif #ifdef RCU_MB static void smp_mb_master(int group) { - smp_mb(); + cmm_smp_mb(); } #endif @@ -143,19 +174,19 @@ static void force_mb_all_readers(void) struct rcu_reader *index; /* - * Ask for each threads to execute a smp_mb() so we can consider the + * Ask for each threads to execute a cmm_smp_mb() so we can consider the * compiler barriers around rcu read lock as real memory barriers. */ - if (list_empty(®istry)) + if (cds_list_empty(®istry)) return; /* - * pthread_kill has a smp_mb(). But beware, we assume it performs + * pthread_kill has a cmm_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 + * safe and don't assume anything : we use cmm_smp_mc() to make sure the * cache flush is enforced. */ - list_for_each_entry(index, ®istry, node) { - STORE_SHARED(index->need_mb, 1); + cds_list_for_each_entry(index, ®istry, node) { + CMM_STORE_SHARED(index->need_mb, 1); pthread_kill(index->tid, SIGRCU); } /* @@ -171,13 +202,13 @@ static void force_mb_all_readers(void) * relevant bug report. For Linux kernels, we recommend getting * the Linux Test Project (LTP). */ - list_for_each_entry(index, ®istry, node) { - while (LOAD_SHARED(index->need_mb)) { + cds_list_for_each_entry(index, ®istry, node) { + while (CMM_LOAD_SHARED(index->need_mb)) { pthread_kill(index->tid, SIGRCU); poll(NULL, 0, 1); } } - smp_mb(); /* read ->need_mb before ending the barrier */ + cmm_smp_mb(); /* read ->need_mb before ending the barrier */ } static void smp_mb_master(int group) @@ -188,24 +219,57 @@ static void smp_mb_master(int group) /* * synchronize_rcu() waiting. Single thread. + * Always called with rcu_registry lock held. Releases this lock and + * grabs it again. Holds the lock when it returns. */ static void wait_gp(void) { - /* Read reader_gp before read futex */ + /* + * Read reader_gp before read futex. smp_mb_master() needs to + * be called with the rcu registry lock held in RCU_SIGNAL + * flavor. + */ smp_mb_master(RCU_MB_GROUP); - if (uatomic_read(&gp_futex) == -1) - futex_async(&gp_futex, FUTEX_WAIT, -1, - NULL, NULL, 0); + /* Temporarily unlock the registry lock. */ + mutex_unlock(&rcu_registry_lock); + if (uatomic_read(&gp_futex) != -1) + goto end; + while (futex_async(&gp_futex, FUTEX_WAIT, -1, + NULL, NULL, 0)) { + switch (errno) { + case EWOULDBLOCK: + /* Value already changed. */ + goto end; + case EINTR: + /* Retry if interrupted by signal. */ + break; /* Get out of switch. */ + default: + /* Unexpected error. */ + urcu_die(errno); + } + } +end: + /* + * Re-lock the registry lock before the next loop. + */ + mutex_lock(&rcu_registry_lock); } +/* + * Always called with rcu_registry lock held. Releases this lock between + * iterations and grabs it again. Holds the lock when it returns. + */ void update_counter_and_wait(void) { - LIST_HEAD(qsreaders); - int wait_loops = 0; + CDS_LIST_HEAD(qsreaders); + unsigned int wait_loops = 0; struct rcu_reader *index, *tmp; +#ifdef HAS_INCOHERENT_CACHES + unsigned int wait_gp_loops = 0; +#endif /* HAS_INCOHERENT_CACHES */ /* Switch parity: 0 -> 1, 1 -> 0 */ - STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); + CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); /* * Must commit rcu_gp_ctr update to memory before waiting for quiescent @@ -213,92 +277,115 @@ void update_counter_and_wait(void) * while new readers are always accessing data (no progress). Enforce * compiler-order of store to rcu_gp_ctr before load rcu_reader ctr. */ - barrier(); + cmm_barrier(); /* * - * Adding a smp_mb() which is _not_ formally required, but makes the + * Adding a cmm_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(); + cmm_smp_mb(); /* - * Wait for each thread rcu_reader.ctr count to become 0. + * Wait for each thread URCU_TLS(rcu_reader).ctr count to become 0. */ for (;;) { - wait_loops++; - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) + wait_loops++; + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { uatomic_dec(&gp_futex); /* Write futex before read reader_gp */ smp_mb_master(RCU_MB_GROUP); } - list_for_each_entry_safe(index, tmp, ®istry, node) { + cds_list_for_each_entry_safe(index, tmp, ®istry, node) { if (!rcu_gp_ongoing(&index->ctr)) - list_move(&index->node, &qsreaders); + cds_list_move(&index->node, &qsreaders); } #ifndef HAS_INCOHERENT_CACHES - if (list_empty(®istry)) { - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + if (cds_list_empty(®istry)) { + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ smp_mb_master(RCU_MB_GROUP); uatomic_set(&gp_futex, 0); } break; } else { - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { + /* wait_gp unlocks/locks registry lock. */ wait_gp(); - else - cpu_relax(); + } else { + /* Temporarily unlock the registry lock. */ + mutex_unlock(&rcu_registry_lock); + caa_cpu_relax(); + /* + * Re-lock the registry lock before the + * next loop. + */ + mutex_lock(&rcu_registry_lock); + } } #else /* #ifndef HAS_INCOHERENT_CACHES */ /* * BUSY-LOOP. Force the reader thread to commit its - * rcu_reader.ctr update to memory if we wait for too long. + * URCU_TLS(rcu_reader).ctr update to memory if we wait + * for too long. */ - if (list_empty(®istry)) { - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + if (cds_list_empty(®istry)) { + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ smp_mb_master(RCU_MB_GROUP); 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: + if (wait_gp_loops == KICK_READER_LOOPS) { smp_mb_master(RCU_MB_GROUP); - wait_loops = 0; - break; /* only escape switch */ - default: - cpu_relax(); + wait_gp_loops = 0; + } + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { + /* wait_gp unlocks/locks registry lock. */ + wait_gp(); + wait_gp_loops++; + } else { + /* Temporarily unlock the registry lock. */ + mutex_unlock(&rcu_registry_lock); + caa_cpu_relax(); + /* + * Re-lock the registry lock before the + * next loop. + */ + mutex_lock(&rcu_registry_lock); } } #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ } /* put back the reader list in the registry */ - list_splice(&qsreaders, ®istry); + cds_list_splice(&qsreaders, ®istry); } void synchronize_rcu(void) { mutex_lock(&rcu_gp_lock); + mutex_lock(&rcu_registry_lock); - if (list_empty(®istry)) + if (cds_list_empty(®istry)) goto out; - /* All threads should read qparity before accessing data structure - * where new ptr points to. Must be done within rcu_gp_lock because it - * iterates on reader threads.*/ + /* + * All threads should read qparity before accessing data structure + * where new ptr points to. Must be done within rcu_registry_lock + * because it iterates on reader threads. + */ /* Write new ptr before changing the qparity */ smp_mb_master(RCU_MB_GROUP); /* * Wait for previous parity to be empty of readers. + * update_counter_and_wait() can release and grab again + * rcu_registry_lock interally. */ update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ @@ -307,27 +394,32 @@ void synchronize_rcu(void) * committing next rcu_gp_ctr update to memory. Failure to do so could * result in the writer waiting forever while new readers are always * accessing data (no progress). Enforce compiler-order of load - * rcu_reader ctr before store to rcu_gp_ctr. + * URCU_TLS(rcu_reader).ctr before store to rcu_gp_ctr. */ - barrier(); + cmm_barrier(); /* - * Adding a smp_mb() which is _not_ formally required, but makes the + * Adding a cmm_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(); + cmm_smp_mb(); /* * Wait for previous parity to be empty of readers. + * update_counter_and_wait() can release and grab again + * rcu_registry_lock interally. */ 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 rcu_gp_lock because it iterates on reader - * threads. */ + /* + * Finish waiting for reader threads before letting the old ptr + * being freed. Must be done within rcu_registry_lock because it + * iterates on reader threads. + */ smp_mb_master(RCU_MB_GROUP); out: + mutex_unlock(&rcu_registry_lock); mutex_unlock(&rcu_gp_lock); } @@ -347,21 +439,21 @@ void rcu_read_unlock(void) void rcu_register_thread(void) { - rcu_reader.tid = pthread_self(); - assert(rcu_reader.need_mb == 0); - assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK)); + URCU_TLS(rcu_reader).tid = pthread_self(); + assert(URCU_TLS(rcu_reader).need_mb == 0); + assert(!(URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK)); - mutex_lock(&rcu_gp_lock); + mutex_lock(&rcu_registry_lock); rcu_init(); /* In case gcc does not support constructor attribute */ - list_add(&rcu_reader.node, ®istry); - mutex_unlock(&rcu_gp_lock); + cds_list_add(&URCU_TLS(rcu_reader).node, ®istry); + mutex_unlock(&rcu_registry_lock); } void rcu_unregister_thread(void) { - mutex_lock(&rcu_gp_lock); - list_del(&rcu_reader.node); - mutex_unlock(&rcu_gp_lock); + mutex_lock(&rcu_registry_lock); + cds_list_del(&URCU_TLS(rcu_reader).node); + mutex_unlock(&rcu_registry_lock); } #ifdef RCU_MEMBARRIER @@ -379,22 +471,22 @@ void rcu_init(void) 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 + * Executing this cmm_smp_mb() is the only purpose of this signal handler. + * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is * executed on. */ - smp_mb(); - _STORE_SHARED(rcu_reader.need_mb, 0); - smp_mb(); + cmm_smp_mb(); + _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0); + cmm_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 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. + * holing the rcu_registry_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) { @@ -409,10 +501,8 @@ void rcu_init(void) act.sa_flags = SA_SIGINFO | SA_RESTART; sigemptyset(&act.sa_mask); ret = sigaction(SIGRCU, &act, NULL); - if (ret) { - perror("Error in sigaction"); - exit(-1); - } + if (ret) + urcu_die(errno); } void rcu_exit(void) @@ -421,11 +511,15 @@ void rcu_exit(void) int ret; ret = sigaction(SIGRCU, NULL, &act); - if (ret) { - perror("Error in sigaction"); - exit(-1); - } + if (ret) + urcu_die(errno); assert(act.sa_sigaction == sigrcu_handler); - assert(list_empty(®istry)); + assert(cds_list_empty(®istry)); } + #endif /* #ifdef RCU_SIGNAL */ + +DEFINE_RCU_FLAVOR(rcu_flavor); + +#include "urcu-call-rcu-impl.h" +#include "urcu-defer-impl.h"