X-Git-Url: https://git.lttng.org/?p=urcu.git;a=blobdiff_plain;f=urcu-qsbr.c;h=685efb5297dca06632697d810f5f078bc0aa835d;hp=745676e209d3ff44dc828507e43ca8c09f023e9c;hb=731ccb963c80afd067e20acee2f9bd7cb4875ffb;hpb=aee1715b914e5f4ba5003079d5d954ef2209f04e diff --git a/urcu-qsbr.c b/urcu-qsbr.c index 745676e..685efb5 100644 --- a/urcu-qsbr.c +++ b/urcu-qsbr.c @@ -35,11 +35,15 @@ #include #include -#include "urcu/wfqueue.h" +#include "urcu/wfcqueue.h" #include "urcu/map/urcu-qsbr.h" #define BUILD_QSBR_LIB #include "urcu/static/urcu-qsbr.h" #include "urcu-pointer.h" +#include "urcu/tls-compat.h" + +#include "urcu-die.h" +#include "urcu-wait.h" /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #undef _LGPL_SOURCE @@ -48,14 +52,22 @@ void __attribute__((destructor)) rcu_exit(void); +/* + * rcu_gp_lock ensures mutual exclusion between threads calling + * synchronize_rcu(). + */ static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; - -int32_t gp_futex; - /* - * Global grace period counter. + * 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. */ -unsigned long rcu_gp_ctr = RCU_GP_ONLINE; +static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; +struct rcu_gp rcu_gp = { .ctr = RCU_GP_ONLINE }; /* * Active attempts to check for reader Q.S. before calling futex(). @@ -66,32 +78,28 @@ unsigned long rcu_gp_ctr = RCU_GP_ONLINE; * 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 +DEFINE_URCU_TLS(struct rcu_reader, rcu_reader); static CDS_LIST_HEAD(registry); +/* + * Queue keeping threads awaiting to wait for a grace period. Contains + * struct gp_waiters_thread objects. + */ +static DEFINE_URCU_WAIT_QUEUE(gp_waiters); + static void mutex_lock(pthread_mutex_t *mutex) { int ret; #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 (ret != EBUSY && ret != EINTR) + urcu_die(ret); poll(NULL,0,10); } #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ @@ -102,10 +110,8 @@ 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); } /* @@ -115,72 +121,76 @@ static void wait_gp(void) { /* Read reader_gp before read futex */ cmm_smp_rmb(); - if (uatomic_read(&gp_futex) == -1) - futex_noasync(&gp_futex, FUTEX_WAIT, -1, + if (uatomic_read(&rcu_gp.futex) == -1) + futex_noasync(&rcu_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0); } -static void update_counter_and_wait(void) +/* + * Always called with rcu_registry lock held. Releases this lock between + * iterations and grabs it again. Holds the lock when it returns. + */ +static void wait_for_readers(struct cds_list_head *input_readers, + struct cds_list_head *cur_snap_readers, + struct cds_list_head *qsreaders) { - CDS_LIST_HEAD(qsreaders); - int wait_loops = 0; + unsigned int wait_loops = 0; struct rcu_reader *index, *tmp; -#if (CAA_BITS_PER_LONG < 64) - /* Switch parity: 0 -> 1, 1 -> 0 */ - CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR); -#else /* !(CAA_BITS_PER_LONG < 64) */ - /* Increment current G.P. */ - CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR); -#endif /* !(CAA_BITS_PER_LONG < 64) */ - - /* - * Must commit rcu_gp_ctr update to memory before waiting for - * quiescent state. Failure to do so could result in the writer - * waiting forever while new readers are always accessing data - * (no progress). Enforce compiler-order of store to rcu_gp_ctr - * before load rcu_reader ctr. - */ - cmm_barrier(); - - /* - * 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. - */ - cmm_smp_mb(); - /* - * Wait for each thread rcu_reader_qs_gp count to become 0. + * Wait for each thread URCU_TLS(rcu_reader).ctr to either + * indicate quiescence (offline), or for them to observe the + * current rcu_gp.ctr value. */ for (;;) { - wait_loops++; + if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) + wait_loops++; if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { - uatomic_set(&gp_futex, -1); + uatomic_set(&rcu_gp.futex, -1); /* * Write futex before write waiting (the other side * reads them in the opposite order). */ cmm_smp_wmb(); - cds_list_for_each_entry(index, ®istry, node) { + cds_list_for_each_entry(index, input_readers, node) { _CMM_STORE_SHARED(index->waiting, 1); } /* Write futex before read reader_gp */ cmm_smp_mb(); } - cds_list_for_each_entry_safe(index, tmp, ®istry, node) { - if (!rcu_gp_ongoing(&index->ctr)) - cds_list_move(&index->node, &qsreaders); + cds_list_for_each_entry_safe(index, tmp, input_readers, node) { + switch (rcu_reader_state(&index->ctr)) { + case RCU_READER_ACTIVE_CURRENT: + if (cur_snap_readers) { + cds_list_move(&index->node, + cur_snap_readers); + break; + } + /* Fall-through */ + case RCU_READER_INACTIVE: + cds_list_move(&index->node, qsreaders); + break; + case RCU_READER_ACTIVE_OLD: + /* + * Old snapshot. Leaving node in + * input_readers will make us busy-loop + * until the snapshot becomes current or + * the reader becomes inactive. + */ + break; + } } - if (cds_list_empty(®istry)) { + if (cds_list_empty(input_readers)) { if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ cmm_smp_mb(); - uatomic_set(&gp_futex, 0); + uatomic_set(&rcu_gp.futex, 0); } break; } else { + /* Temporarily unlock the registry lock. */ + mutex_unlock(&rcu_registry_lock); if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { wait_gp(); } else { @@ -190,10 +200,10 @@ static void update_counter_and_wait(void) cmm_smp_mb(); #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ } + /* Re-lock the registry lock before the next loop. */ + mutex_lock(&rcu_registry_lock); } } - /* put back the reader list in the registry */ - cds_list_splice(&qsreaders, ®istry); } /* @@ -204,9 +214,13 @@ static void update_counter_and_wait(void) #if (CAA_BITS_PER_LONG < 64) void synchronize_rcu(void) { + CDS_LIST_HEAD(cur_snap_readers); + CDS_LIST_HEAD(qsreaders); unsigned long was_online; + DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING); + struct urcu_waiters waiters; - was_online = rcu_reader.ctr; + was_online = rcu_read_ongoing(); /* All threads should read qparity before accessing data structure * where new ptr points to. In the "then" case, rcu_thread_offline @@ -221,23 +235,45 @@ void synchronize_rcu(void) else cmm_smp_mb(); + /* + * Add ourself to gp_waiters queue of threads awaiting to wait + * for a grace period. Proceed to perform the grace period only + * if we are the first thread added into the queue. + */ + if (urcu_wait_add(&gp_waiters, &wait) != 0) { + /* Not first in queue: will be awakened by another thread. */ + urcu_adaptative_busy_wait(&wait); + goto gp_end; + } + /* We won't need to wake ourself up */ + urcu_wait_set_state(&wait, URCU_WAIT_RUNNING); + mutex_lock(&rcu_gp_lock); + /* + * Move all waiters into our local queue. + */ + urcu_move_waiters(&waiters, &gp_waiters); + + mutex_lock(&rcu_registry_lock); + if (cds_list_empty(®istry)) goto out; /* - * Wait for previous parity to be empty of readers. + * Wait for readers to observe original parity or be quiescent. + * wait_for_readers() can release and grab again rcu_registry_lock + * interally. */ - update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ + wait_for_readers(®istry, &cur_snap_readers, &qsreaders); /* - * Must finish waiting for quiescent state for parity 0 before - * committing next rcu_gp_ctr update to memory. Failure to - * do so could result in the writer waiting forever while new + * Must finish waiting for quiescent state for original parity + * before 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. + * compiler-order of load URCU_TLS(rcu_reader).ctr before store + * to rcu_gp.ctr. */ cmm_barrier(); @@ -248,13 +284,41 @@ void synchronize_rcu(void) */ cmm_smp_mb(); + /* Switch parity: 0 -> 1, 1 -> 0 */ + CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr ^ RCU_GP_CTR); + /* - * Wait for previous parity to be empty of readers. + * Must commit rcu_gp.ctr update to memory before waiting for + * quiescent state. Failure to do so could result in the writer + * waiting forever while new readers are always accessing data + * (no progress). Enforce compiler-order of store to rcu_gp.ctr + * before load URCU_TLS(rcu_reader).ctr. */ - update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ + cmm_barrier(); + + /* + * 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. + */ + cmm_smp_mb(); + + /* + * Wait for readers to observe new parity or be quiescent. + * wait_for_readers() can release and grab again rcu_registry_lock + * interally. + */ + wait_for_readers(&cur_snap_readers, NULL, &qsreaders); + + /* + * Put quiescent reader list back into registry. + */ + cds_list_splice(&qsreaders, ®istry); out: + mutex_unlock(&rcu_registry_lock); mutex_unlock(&rcu_gp_lock); - + urcu_wake_all_waiters(&waiters); +gp_end: /* * Finish waiting for reader threads before letting the old ptr being * freed. @@ -267,9 +331,12 @@ out: #else /* !(CAA_BITS_PER_LONG < 64) */ void synchronize_rcu(void) { + CDS_LIST_HEAD(qsreaders); unsigned long was_online; + DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING); + struct urcu_waiters waiters; - was_online = rcu_reader.ctr; + was_online = rcu_read_ongoing(); /* * Mark the writer thread offline to make sure we don't wait for @@ -281,13 +348,66 @@ void synchronize_rcu(void) else cmm_smp_mb(); + /* + * Add ourself to gp_waiters queue of threads awaiting to wait + * for a grace period. Proceed to perform the grace period only + * if we are the first thread added into the queue. + */ + if (urcu_wait_add(&gp_waiters, &wait) != 0) { + /* Not first in queue: will be awakened by another thread. */ + urcu_adaptative_busy_wait(&wait); + goto gp_end; + } + /* We won't need to wake ourself up */ + urcu_wait_set_state(&wait, URCU_WAIT_RUNNING); + mutex_lock(&rcu_gp_lock); + + /* + * Move all waiters into our local queue. + */ + urcu_move_waiters(&waiters, &gp_waiters); + + mutex_lock(&rcu_registry_lock); + if (cds_list_empty(®istry)) goto out; - update_counter_and_wait(); + + /* Increment current G.P. */ + CMM_STORE_SHARED(rcu_gp.ctr, rcu_gp.ctr + RCU_GP_CTR); + + /* + * Must commit rcu_gp.ctr update to memory before waiting for + * quiescent state. Failure to do so could result in the writer + * waiting forever while new readers are always accessing data + * (no progress). Enforce compiler-order of store to rcu_gp.ctr + * before load URCU_TLS(rcu_reader).ctr. + */ + cmm_barrier(); + + /* + * 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. + */ + cmm_smp_mb(); + + /* + * Wait for readers to observe new count of be quiescent. + * wait_for_readers() can release and grab again rcu_registry_lock + * interally. + */ + wait_for_readers(®istry, NULL, &qsreaders); + + /* + * Put quiescent reader list back into registry. + */ + cds_list_splice(&qsreaders, ®istry); out: + mutex_unlock(&rcu_registry_lock); mutex_unlock(&rcu_gp_lock); - + urcu_wake_all_waiters(&waiters); +gp_end: if (was_online) rcu_thread_online(); else @@ -309,6 +429,11 @@ void rcu_read_unlock(void) _rcu_read_unlock(); } +int rcu_read_ongoing(void) +{ + return _rcu_read_ongoing(); +} + void rcu_quiescent_state(void) { _rcu_quiescent_state(); @@ -326,12 +451,12 @@ void rcu_thread_online(void) void rcu_register_thread(void) { - rcu_reader.tid = pthread_self(); - assert(rcu_reader.ctr == 0); + URCU_TLS(rcu_reader).tid = pthread_self(); + assert(URCU_TLS(rcu_reader).ctr == 0); - mutex_lock(&rcu_gp_lock); - cds_list_add(&rcu_reader.node, ®istry); - mutex_unlock(&rcu_gp_lock); + mutex_lock(&rcu_registry_lock); + cds_list_add(&URCU_TLS(rcu_reader).node, ®istry); + mutex_unlock(&rcu_registry_lock); _rcu_thread_online(); } @@ -342,9 +467,9 @@ void rcu_unregister_thread(void) * with a waiting writer. */ _rcu_thread_offline(); - mutex_lock(&rcu_gp_lock); - cds_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); } void rcu_exit(void)