X-Git-Url: https://git.liburcu.org/?a=blobdiff_plain;f=urcu-qsbr.c;h=7f747ed5f9ce8ef133df3c2d026b8b76154e2e9e;hb=6362f68f024fd85fefe342b1f82d8787146c1ebb;hp=5e4348449ca9183b61bf1cbc6a041a549f784c2b;hpb=71c811bf8db75c9502b295edb1e190f978682b65;p=urcu.git diff --git a/urcu-qsbr.c b/urcu-qsbr.c index 5e43484..7f747ed 100644 --- a/urcu-qsbr.c +++ b/urcu-qsbr.c @@ -35,10 +35,15 @@ #include #include -#include "urcu/wfqueue.h" +#include "urcu/wfcqueue.h" +#include "urcu/wfstack.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" /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ #undef _LGPL_SOURCE @@ -49,7 +54,7 @@ void __attribute__((destructor)) rcu_exit(void); static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; -int32_t gp_futex; +int32_t rcu_gp_futex; /* * Global grace period counter. @@ -65,32 +70,56 @@ 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; +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 CDS_LIST_HEAD(registry); +/* + * Number of busy-loop attempts before waiting on futex for grace period + * batching. + */ +#define RCU_AWAKE_ATTEMPTS 1000 + +enum adapt_wakeup_state { + /* AWAKE_WAITING is compared directly (futex compares it). */ + AWAKE_WAITING = 0, + /* non-zero are used as masks. */ + AWAKE_WAKEUP = (1 << 0), + AWAKE_AWAKENED = (1 << 1), + AWAKE_TEARDOWN = (1 << 2), +}; + +struct gp_waiters_thread { + struct cds_wfs_node node; + int32_t wait_futex; +}; + +/* + * Stack keeping threads awaiting to wait for a grace period. Contains + * struct gp_waiters_thread objects. + */ +static struct cds_wfs_stack gp_waiters = { + .head = CDS_WFS_END, + .lock = PTHREAD_MUTEX_INITIALIZER, +}; + 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 */ @@ -101,10 +130,8 @@ 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); } /* @@ -114,69 +141,118 @@ 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) +/* + * Note: urcu_adaptative_wake_up needs "value" to stay allocated + * throughout its execution. In this scheme, the waiter owns the futex + * memory, and we only allow it to free this memory when it receives the + * AWAKE_TEARDOWN flag. + */ +static void urcu_adaptative_wake_up(int32_t *value) { - CDS_LIST_HEAD(qsreaders); - int wait_loops = 0; - struct rcu_reader *index, *tmp; + cmm_smp_mb(); + assert(uatomic_read(value) == AWAKE_WAITING); + uatomic_set(value, AWAKE_WAKEUP); + if (!(uatomic_read(value) & AWAKE_AWAKENED)) + futex_noasync(value, FUTEX_WAKE, 1, NULL, NULL, 0); + /* Allow teardown of "value" memory. */ + uatomic_or(value, AWAKE_TEARDOWN); +} -#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) */ +/* + * Caller must initialize "value" to AWAKE_WAITING before passing its + * memory to waker thread. + */ +static void urcu_adaptative_busy_wait(int32_t *value) +{ + unsigned int i; - /* - * 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(); + /* Load and test condition before read futex */ + cmm_smp_rmb(); + for (i = 0; i < RCU_AWAKE_ATTEMPTS; i++) { + if (uatomic_read(value) != AWAKE_WAITING) + goto skip_futex_wait; + caa_cpu_relax(); + } + futex_noasync(value, FUTEX_WAIT, AWAKE_WAITING, NULL, NULL, 0); +skip_futex_wait: + + /* Tell waker thread than we are awakened. */ + uatomic_or(value, AWAKE_AWAKENED); /* - * 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. + * Wait until waker thread lets us know it's ok to tear down + * memory allocated for value. */ - cmm_smp_mb(); + for (i = 0; i < RCU_AWAKE_ATTEMPTS; i++) { + if (uatomic_read(value) & AWAKE_TEARDOWN) + break; + caa_cpu_relax(); + } + while (!(uatomic_read(value) & AWAKE_TEARDOWN)) + poll(NULL, 0, 10); + assert(uatomic_read(value) & AWAKE_TEARDOWN); +} + +static void wait_for_readers(struct cds_list_head *input_readers, + struct cds_list_head *cur_snap_readers, + struct cds_list_head *qsreaders) +{ + int wait_loops = 0; + struct rcu_reader *index, *tmp; /* - * 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_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 { @@ -191,8 +267,6 @@ static void update_counter_and_wait(void) } } } - /* put back the reader list in the registry */ - cds_list_splice(&qsreaders, ®istry); } /* @@ -203,41 +277,82 @@ 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; + struct gp_waiters_thread gp_waiters_thread; + struct cds_wfs_head *gp_waiters_head; + struct cds_wfs_node *waiters_iter, *waiters_iter_n; - was_online = rcu_reader.ctr; + was_online = URCU_TLS(rcu_reader).ctr; /* All threads should read qparity before accessing data structure - * where new ptr points to. - */ - /* Write new ptr before changing the qparity */ - cmm_smp_mb(); - - /* + * 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. */ if (was_online) - CMM_STORE_SHARED(rcu_reader.ctr, 0); + rcu_thread_offline(); + else + cmm_smp_mb(); + + /* + * Add ourself to gp_waiters stack 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 stack. + */ + cds_wfs_node_init(&gp_waiters_thread.node); + gp_waiters_thread.wait_futex = AWAKE_WAITING; + if (cds_wfs_push(&gp_waiters, &gp_waiters_node) != 0) { + /* Not first in stack: will be awakened by another thread. */ + urcu_adaptative_busy_wait(&gp_waiters_thread.wait_futex); + goto gp_end; + } mutex_lock(&rcu_gp_lock); + /* + * Pop all waiters into our local stack head. + */ + gp_waiters_head = __cds_wfs_pop_all(&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 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(); + + /* + * 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(); + + /* Switch parity: 0 -> 1, 1 -> 0 */ + 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(); @@ -249,48 +364,131 @@ void synchronize_rcu(void) cmm_smp_mb(); /* - * Wait for previous parity to be empty of readers. + * 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. */ - update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ + cds_list_splice(&qsreaders, ®istry); out: mutex_unlock(&rcu_gp_lock); + /* Wake all waiters in our stack head, excluding ourself. */ + cds_wfs_for_each_blocking_safe(gp_waiters_head, waiters_iter, + waiters_iter_n) { + struct gp_waiters_thread *wt; + + wt = caa_container_of(waiters_iter, + struct gp_waiters_thread, node); + if (wt == &gp_waiters_thread) + continue; + urcu_adaptative_wake_up(&wt->wait_futex); + } + +gp_end: /* * Finish waiting for reader threads before letting the old ptr being * freed. */ if (was_online) - _CMM_STORE_SHARED(rcu_reader.ctr, - CMM_LOAD_SHARED(rcu_gp_ctr)); - cmm_smp_mb(); + rcu_thread_online(); + else + cmm_smp_mb(); } #else /* !(CAA_BITS_PER_LONG < 64) */ void synchronize_rcu(void) { + CDS_LIST_HEAD(qsreaders); unsigned long was_online; + struct gp_waiters_thread gp_waiters_thread; + struct cds_wfs_head *gp_waiters_head; + struct cds_wfs_node *waiters_iter, *waiters_iter_n; - was_online = rcu_reader.ctr; + was_online = URCU_TLS(rcu_reader).ctr; /* * 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. */ - cmm_smp_mb(); if (was_online) - CMM_STORE_SHARED(rcu_reader.ctr, 0); + rcu_thread_offline(); + else + cmm_smp_mb(); + + /* + * Add ourself to gp_waiters stack 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 stack. + */ + cds_wfs_node_init(&gp_waiters_thread.node); + gp_waiters_thread.wait_futex = AWAKE_WAITING; + if (cds_wfs_push(&gp_waiters, &gp_waiters_thread.node) != 0) { + /* Not first in stack: will be awakened by another thread. */ + urcu_adaptative_busy_wait(&gp_waiters_thread.wait_futex); + goto gp_end; + } mutex_lock(&rcu_gp_lock); + + /* + * Pop all waiters into our local stack head. + */ + gp_waiters_head = __cds_wfs_pop_all(&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); + /* Wake all waiters in our stack head, excluding ourself. */ + cds_wfs_for_each_blocking_safe(gp_waiters_head, waiters_iter, + waiters_iter_n) { + struct gp_waiters_thread *wt; + + wt = caa_container_of(waiters_iter, + struct gp_waiters_thread, node); + if (wt == &gp_waiters_thread) + continue; + urcu_adaptative_wake_up(&wt->wait_futex); + } + +gp_end: if (was_online) - _CMM_STORE_SHARED(rcu_reader.ctr, - CMM_LOAD_SHARED(rcu_gp_ctr)); - cmm_smp_mb(); + rcu_thread_online(); + else + cmm_smp_mb(); } #endif /* !(CAA_BITS_PER_LONG < 64) */ @@ -325,11 +523,11 @@ 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); - cds_list_add(&rcu_reader.node, ®istry); + cds_list_add(&URCU_TLS(rcu_reader).node, ®istry); mutex_unlock(&rcu_gp_lock); _rcu_thread_online(); } @@ -342,14 +540,20 @@ void rcu_unregister_thread(void) */ _rcu_thread_offline(); mutex_lock(&rcu_gp_lock); - cds_list_del(&rcu_reader.node); + cds_list_del(&URCU_TLS(rcu_reader).node); mutex_unlock(&rcu_gp_lock); } void rcu_exit(void) { - assert(cds_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"