X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-qsbr.c;h=786a80d6b16b1936dc62b71b43db18850854dd41;hp=155505abc7843385fafef29de57509ceb096519c;hb=882f335739b978d1c55be2faeed077f315afe5d7;hpb=2dfb8b5e30bc2d268bcbc57eceb9c54d21b36d90 diff --git a/urcu-qsbr.c b/urcu-qsbr.c index 155505a..786a80d 100644 --- a/urcu-qsbr.c +++ b/urcu-qsbr.c @@ -3,7 +3,7 @@ * * Userspace RCU QSBR 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 @@ -23,43 +23,67 @@ * IBM's contributions to this file may be relicensed under LGPLv2 or later. */ +#define _GNU_SOURCE +#define _LGPL_SOURCE #include #include #include #include #include +#include #include #include #include +#include "urcu/wfcqueue.h" +#include "urcu/map/urcu-qsbr.h" #define BUILD_QSBR_LIB -#include "urcu-qsbr-static.h" +#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 #include "urcu-qsbr.h" +#define _LGPL_SOURCE void __attribute__((destructor)) rcu_exit(void); static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; -int gp_futex; +int32_t rcu_gp_futex; /* * Global grace period counter. */ unsigned long rcu_gp_ctr = RCU_GP_ONLINE; +/* + * Active attempts to check for reader Q.S. before calling futex(). + */ +#define RCU_QS_ACTIVE_ATTEMPTS 100 + /* * Written to only by each individual reader. Read by both the reader and the * writers. */ -struct rcu_reader __thread rcu_reader; +DEFINE_URCU_TLS(struct rcu_reader, rcu_reader); #ifdef DEBUG_YIELD -unsigned int yield_active; -unsigned int __thread rand_yield; +unsigned int rcu_yield_active; +DEFINE_URCU_TLS(unsigned int, rcu_rand_yield); #endif -static LIST_HEAD(registry); +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) { @@ -67,17 +91,12 @@ 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 (ret != EBUSY && ret != EINTR) + urcu_die(ret); poll(NULL,0,10); } #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ @@ -88,10 +107,8 @@ 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); } /* @@ -100,63 +117,81 @@ static void mutex_unlock(pthread_mutex_t *mutex) static void wait_gp(void) { /* Read reader_gp before read futex */ - smp_rmb(); - if (uatomic_read(&gp_futex) == -1) - futex_noasync(&gp_futex, FUTEX_WAIT, -1, + cmm_smp_rmb(); + 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) +static void wait_for_readers(struct cds_list_head *input_readers, + struct cds_list_head *cur_snap_readers, + struct cds_list_head *qsreaders) { - LIST_HEAD(qsreaders); int wait_loops = 0; struct rcu_reader *index, *tmp; -#if (BITS_PER_LONG < 64) - /* Switch parity: 1 -> 0, 0 -> 1 */ - STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR); -#else /* !(BITS_PER_LONG < 64) */ - /* Increment current G.P. */ - STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR); -#endif /* !(BITS_PER_LONG < 64) */ - /* - * 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) { - uatomic_dec(&gp_futex); + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { + 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, input_readers, node) { + _CMM_STORE_SHARED(index->waiting, 1); + } /* Write futex before read reader_gp */ - smp_mb(); + cmm_smp_mb(); } - - list_for_each_entry_safe(index, tmp, ®istry, head) { - if (!rcu_gp_ongoing(&index->ctr)) - list_move(&index->head, &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 (list_empty(®istry)) { - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + if (cds_list_empty(input_readers)) { + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { /* Read reader_gp before write futex */ - smp_mb(); - uatomic_set(&gp_futex, 0); + cmm_smp_mb(); + uatomic_set(&rcu_gp_futex, 0); } break; } else { - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { wait_gp(); } else { #ifndef HAS_INCOHERENT_CACHES - cpu_relax(); + caa_cpu_relax(); #else /* #ifndef HAS_INCOHERENT_CACHES */ - smp_mb(); + cmm_smp_mb(); #endif /* #else #ifndef HAS_INCOHERENT_CACHES */ } } } - /* put back the reader list in the registry */ - list_splice(&qsreaders, ®istry); } /* @@ -164,95 +199,197 @@ static void update_counter_and_wait(void) * long-size to ensure we do not encounter an overflow bug. */ -#if (BITS_PER_LONG < 64) +#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. + * where new ptr points to. In the "then" case, rcu_thread_offline + * includes a memory barrier. + * + * Mark the writer thread offline to make sure we don't wait for + * our own quiescent state. This allows using synchronize_rcu() + * in threads registered as readers. */ - /* Write new ptr before changing the qparity */ - smp_mb(); + if (was_online) + rcu_thread_offline(); + else + cmm_smp_mb(); /* - * Mark the writer thread offline to make sure we don't wait for - * our own quiescent state. This allows using synchronize_rcu() in - * threads registered as readers. + * 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 (was_online) - STORE_SHARED(rcu_reader.ctr, 0); + 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); - if (list_empty(®istry)) + /* + * Move all waiters into our local queue. + */ + urcu_move_waiters(&waiters, &gp_waiters); + + 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. */ - 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 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. + * 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 URCU_TLS(rcu_reader).ctr before store + * to rcu_gp_ctr. */ + 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(); + + /* 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(&cur_snap_readers, NULL, &qsreaders); + + /* + * Put quiescent reader list back into registry. + */ + cds_list_splice(&qsreaders, ®istry); out: mutex_unlock(&rcu_gp_lock); - + urcu_wake_all_waiters(&waiters); +gp_end: /* * Finish waiting for reader threads before letting the old ptr being * freed. */ if (was_online) - _STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr)); - smp_mb(); + rcu_thread_online(); + else + cmm_smp_mb(); } -#else /* !(BITS_PER_LONG < 64) */ +#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 - * our own quiescent state. This allows using synchronize_rcu() in - * threads registered as readers. + * our own quiescent state. This allows using synchronize_rcu() + * in threads registered as readers. */ - smp_mb(); if (was_online) - STORE_SHARED(rcu_reader.ctr, 0); + rcu_thread_offline(); + 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); - if (list_empty(®istry)) + + /* + * Move all waiters into our local queue. + */ + urcu_move_waiters(&waiters, &gp_waiters); + + 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(®istry, NULL, &qsreaders); + + /* + * Put quiescent reader list back into registry. + */ + cds_list_splice(&qsreaders, ®istry); out: mutex_unlock(&rcu_gp_lock); - + urcu_wake_all_waiters(&waiters); +gp_end: if (was_online) - _STORE_SHARED(rcu_reader.ctr, LOAD_SHARED(rcu_gp_ctr)); - smp_mb(); + rcu_thread_online(); + else + cmm_smp_mb(); } -#endif /* !(BITS_PER_LONG < 64) */ +#endif /* !(CAA_BITS_PER_LONG < 64) */ /* * library wrappers to be used by non-LGPL compatible source code. @@ -268,6 +405,11 @@ 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(); @@ -285,11 +427,11 @@ 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); - list_add(&rcu_reader.head, ®istry); + cds_list_add(&URCU_TLS(rcu_reader).node, ®istry); mutex_unlock(&rcu_gp_lock); _rcu_thread_online(); } @@ -302,11 +444,20 @@ void rcu_unregister_thread(void) */ _rcu_thread_offline(); mutex_lock(&rcu_gp_lock); - list_del(&rcu_reader.head); + cds_list_del(&URCU_TLS(rcu_reader).node); mutex_unlock(&rcu_gp_lock); } void rcu_exit(void) { - assert(list_empty(®istry)); + /* + * Assertion disabled because call_rcu threads are now rcu + * readers, and left running at exit. + * assert(cds_list_empty(®istry)); + */ } + +DEFINE_RCU_FLAVOR(rcu_flavor); + +#include "urcu-call-rcu-impl.h" +#include "urcu-defer-impl.h"