X-Git-Url: https://git.lttng.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=94d1131e3dba6df7982a33def88f6ec3908c2bf2;hp=2d5c510aa6744418c5557d642b451e0afae415cd;hb=8999a9ee7abe868caf8f5327d8b2e8b8b44ee87d;hpb=c94886840694378921968b490ed1e09d1af3afb3 diff --git a/urcu.c b/urcu.c index 2d5c510..94d1131 100644 --- a/urcu.c +++ b/urcu.c @@ -26,6 +26,7 @@ #define _BSD_SOURCE #define _GNU_SOURCE #define _LGPL_SOURCE +#define _DEFAULT_SOURCE #include #include #include @@ -43,6 +44,7 @@ #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 @@ -52,15 +54,33 @@ /* * 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 a few loops waiting for it. + * per-se), kick it after 10 loops waiting for it. */ -#define KICK_READER_LOOPS 10000 +#define KICK_READER_LOOPS 10 /* * Active attempts to check for reader Q.S. before calling futex(). */ #define RCU_QS_ACTIVE_ATTEMPTS 100 +/* + * RCU_MEMBARRIER is only possibly available on Linux. + */ +#if defined(RCU_MEMBARRIER) && defined(__linux__) +#include +#endif + +/* If the headers do not support SYS_membarrier, fall back on RCU_MB */ +#ifdef SYS_membarrier +# define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__) +#else +# define membarrier(...) -ENOSYS +#endif + +#define MEMBARRIER_EXPEDITED (1 << 0) +#define MEMBARRIER_DELAYED (1 << 1) +#define MEMBARRIER_QUERY (1 << 16) + #ifdef RCU_MEMBARRIER static int init_done; int rcu_has_sys_membarrier; @@ -81,17 +101,22 @@ 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; - -int32_t rcu_gp_futex; - /* - * Global grace period counter. - * 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. + * 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_COUNT; +static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; +struct rcu_gp rcu_gp = { .ctr = RCU_GP_COUNT }; /* * Written to only by each individual reader. Read by both the reader and the @@ -99,13 +124,14 @@ unsigned long rcu_gp_ctr = RCU_GP_COUNT; */ DEFINE_URCU_TLS(struct rcu_reader, rcu_reader); -#ifdef DEBUG_YIELD -unsigned int rcu_yield_active; -DEFINE_URCU_TLS(unsigned int, rcu_rand_yield); -#endif - 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; @@ -123,7 +149,7 @@ static void mutex_lock(pthread_mutex_t *mutex) _CMM_STORE_SHARED(URCU_TLS(rcu_reader).need_mb, 0); cmm_smp_mb(); } - poll(NULL,0,10); + (void) poll(NULL, 0, 10); } #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ } @@ -141,7 +167,7 @@ static void mutex_unlock(pthread_mutex_t *mutex) static void smp_mb_master(int group) { if (caa_likely(rcu_has_sys_membarrier)) - membarrier(MEMBARRIER_EXPEDITED); + (void) membarrier(MEMBARRIER_EXPEDITED); else cmm_smp_mb(); } @@ -191,7 +217,7 @@ static void force_mb_all_readers(void) cds_list_for_each_entry(index, ®istry, node) { while (CMM_LOAD_SHARED(index->need_mb)) { pthread_kill(index->tid, SIGRCU); - poll(NULL, 0, 1); + (void) poll(NULL, 0, 1); } } cmm_smp_mb(); /* read ->need_mb before ending the barrier */ @@ -210,48 +236,79 @@ static void wait_gp(void) { /* Read reader_gp before read futex */ smp_mb_master(RCU_MB_GROUP); - if (uatomic_read(&rcu_gp_futex) == -1) - futex_async(&rcu_gp_futex, FUTEX_WAIT, -1, + if (uatomic_read(&rcu_gp.futex) == -1) + futex_async(&rcu_gp.futex, FUTEX_WAIT, -1, NULL, NULL, 0); } -static void wait_for_readers(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; +#ifdef HAS_INCOHERENT_CACHES + unsigned int wait_gp_loops = 0; +#endif /* HAS_INCOHERENT_CACHES */ /* * Wait for each thread URCU_TLS(rcu_reader).ctr to either * indicate quiescence (not nested), or observe the current - * rcu_gp_ctr value. + * rcu_gp.ctr value. */ for (;;) { - wait_loops++; - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) { - uatomic_dec(&rcu_gp_futex); + if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) + wait_loops++; + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { + uatomic_dec(&rcu_gp.futex); /* Write futex before read reader_gp */ smp_mb_master(RCU_MB_GROUP); } - 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; + } } #ifndef HAS_INCOHERENT_CACHES - if (cds_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_master(RCU_MB_GROUP); - uatomic_set(&rcu_gp_futex, 0); + uatomic_set(&rcu_gp.futex, 0); } break; } else { - if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) + /* Temporarily unlock the registry lock. */ + mutex_unlock(&rcu_registry_lock); + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) wait_gp(); else caa_cpu_relax(); + /* Re-lock the registry lock before the next loop. */ + mutex_lock(&rcu_registry_lock); } #else /* #ifndef HAS_INCOHERENT_CACHES */ /* @@ -259,56 +316,91 @@ static void wait_for_readers(void) * URCU_TLS(rcu_reader).ctr update to memory if we wait * for too long. */ - if (cds_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_master(RCU_MB_GROUP); - uatomic_set(&rcu_gp_futex, 0); + uatomic_set(&rcu_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: + wait_gp_loops = 0; + } + /* Temporarily unlock the registry lock. */ + mutex_unlock(&rcu_registry_lock); + if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) { + wait_gp(); + wait_gp_loops++; + } else { 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 */ - cds_list_splice(&qsreaders, ®istry); } void synchronize_rcu(void) { + CDS_LIST_HEAD(cur_snap_readers); + CDS_LIST_HEAD(qsreaders); + DEFINE_URCU_WAIT_NODE(wait, URCU_WAIT_WAITING); + struct urcu_waiters waiters; + + /* + * 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. + * The implicit memory barrier before urcu_wait_add() + * orders prior memory accesses of threads put into the wait + * queue before their insertion into the wait queue. + */ + if (urcu_wait_add(&gp_waiters, &wait) != 0) { + /* Not first in queue: will be awakened by another thread. */ + urcu_adaptative_busy_wait(&wait); + /* Order following memory accesses after grace period. */ + cmm_smp_mb(); + return; + } + /* 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; - /* 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 readers to observe original parity or be quiescent. + * wait_for_readers() can release and grab again rcu_registry_lock + * interally. */ - wait_for_readers(); + wait_for_readers(®istry, &cur_snap_readers, &qsreaders); /* * Must finish waiting for quiescent state for original parity before - * committing next rcu_gp_ctr update to memory. Failure to do so could + * 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. + * URCU_TLS(rcu_reader).ctr before store to rcu_gp.ctr. */ cmm_barrier(); @@ -320,13 +412,13 @@ 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_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 + * 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. + * compiler-order of store to rcu_gp.ctr before load rcu_reader ctr. */ cmm_barrier(); @@ -340,15 +432,32 @@ void synchronize_rcu(void) /* * 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(); + wait_for_readers(&cur_snap_readers, NULL, &qsreaders); - /* 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. */ + /* + * Put quiescent reader list back into registry. + */ + cds_list_splice(&qsreaders, ®istry); + + /* + * 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); + + /* + * Wakeup waiters only after we have completed the grace period + * and have ensured the memory barriers at the end of the grace + * period have been issued. + */ + urcu_wake_all_waiters(&waiters); } /* @@ -365,23 +474,28 @@ void rcu_read_unlock(void) _rcu_read_unlock(); } +int rcu_read_ongoing(void) +{ + return _rcu_read_ongoing(); +} + void rcu_register_thread(void) { 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 */ cds_list_add(&URCU_TLS(rcu_reader).node, ®istry); - mutex_unlock(&rcu_gp_lock); + mutex_unlock(&rcu_registry_lock); } void rcu_unregister_thread(void) { - mutex_lock(&rcu_gp_lock); + mutex_lock(&rcu_registry_lock); cds_list_del(&URCU_TLS(rcu_reader).node); - mutex_unlock(&rcu_gp_lock); + mutex_unlock(&rcu_registry_lock); } #ifdef RCU_MEMBARRIER @@ -412,9 +526,9 @@ static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context) * 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) { @@ -435,14 +549,14 @@ void rcu_init(void) void rcu_exit(void) { - struct sigaction act; - int ret; - - ret = sigaction(SIGRCU, NULL, &act); - if (ret) - urcu_die(errno); - assert(act.sa_sigaction == sigrcu_handler); - assert(cds_list_empty(®istry)); + /* + * Don't unregister the SIGRCU signal handler anymore, because + * call_rcu threads could still be using it shortly before the + * application exits. + * Assertion disabled because call_rcu threads are now rcu + * readers, and left running at exit. + * assert(cds_list_empty(®istry)); + */ } #endif /* #ifdef RCU_SIGNAL */