*
* Userspace RCU QSBR library
*
- * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
+ * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
* Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
*
* This library is free software; you can redistribute it and/or
* IBM's contributions to this file may be relicensed under LGPLv2 or later.
*/
+#define _GNU_SOURCE
+#define _LGPL_SOURCE
#include <stdio.h>
#include <pthread.h>
#include <signal.h>
#include <assert.h>
#include <stdlib.h>
+#include <stdint.h>
#include <string.h>
#include <errno.h>
#include <poll.h>
+#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 urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER;
-int gp_futex;
+int32_t rcu_gp_futex;
/*
* Global grace period counter.
*/
-unsigned long urcu_gp_ctr = RCU_GP_ONLINE;
+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.
*/
-unsigned long __thread rcu_reader_qs_gp;
-
-/* Thread IDs of registered readers */
-#define INIT_NUM_THREADS 4
-
-struct reader_registry {
- pthread_t tid;
- unsigned long *rcu_reader_qs_gp;
-};
+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 struct reader_registry *registry;
-static int num_readers, alloc_readers;
+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 internal_urcu_lock(void)
+static void mutex_lock(pthread_mutex_t *mutex)
{
int ret;
#ifndef DISTRUST_SIGNALS_EXTREME
- ret = pthread_mutex_lock(&urcu_mutex);
- if (ret) {
- perror("Error in pthread mutex lock");
- exit(-1);
- }
+ ret = pthread_mutex_lock(mutex);
+ if (ret)
+ urcu_die(ret);
#else /* #ifndef DISTRUST_SIGNALS_EXTREME */
- while ((ret = pthread_mutex_trylock(&urcu_mutex)) != 0) {
- if (ret != EBUSY && ret != EINTR) {
- printf("ret = %d, errno = %d\n", ret, errno);
- perror("Error in pthread mutex lock");
- exit(-1);
- }
+ while ((ret = pthread_mutex_trylock(mutex)) != 0) {
+ if (ret != EBUSY && ret != EINTR)
+ urcu_die(ret);
poll(NULL,0,10);
}
#endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
}
-static void internal_urcu_unlock(void)
+static void mutex_unlock(pthread_mutex_t *mutex)
{
int ret;
- ret = pthread_mutex_unlock(&urcu_mutex);
- if (ret) {
- perror("Error in pthread mutex unlock");
- exit(-1);
- }
+ ret = pthread_mutex_unlock(mutex);
+ if (ret)
+ urcu_die(ret);
}
/*
* synchronize_rcu() waiting. Single thread.
*/
-static void wait_gp(struct reader_registry *index)
+static void wait_gp(void)
{
- uatomic_dec(&gp_futex);
- smp_mb(); /* Write futex before read reader_gp */
- if (!rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
- /* Read reader_gp before write futex */
- smp_mb();
- /* Callbacks are queued, don't wait. */
- uatomic_set(&gp_futex, 0);
- } else {
- /* Read reader_gp before read futex */
- smp_rmb();
- if (uatomic_read(&gp_futex) == -1)
- futex(&gp_futex, FUTEX_WAIT, -1,
- NULL, NULL, 0);
- }
+ /* Read reader_gp before read futex */
+ cmm_smp_rmb();
+ if (uatomic_read(&rcu_gp_futex) == -1)
+ futex_noasync(&rcu_gp_futex, FUTEX_WAIT, -1,
+ NULL, NULL, 0);
}
-static void wait_for_quiescent_state(void)
+static void wait_for_readers(struct cds_list_head *input_readers,
+ struct cds_list_head *cur_snap_readers,
+ struct cds_list_head *qsreaders)
{
- struct reader_registry *index;
+ int wait_loops = 0;
+ struct rcu_reader *index, *tmp;
- if (!registry)
- return;
/*
- * 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 (index = registry; index < registry + num_readers; index++) {
- int wait_loops = 0;
+ for (;;) {
+ wait_loops++;
+ 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 */
+ cmm_smp_mb();
+ }
+ 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;
+ }
+ }
- while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
- if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
- wait_gp(index);
+ if (cds_list_empty(input_readers)) {
+ if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
+ /* Read reader_gp before write futex */
+ cmm_smp_mb();
+ uatomic_set(&rcu_gp_futex, 0);
+ }
+ break;
+ } else {
+ 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 */
}
}
* long-size to ensure we do not encounter an overflow bug.
*/
-#if (BITS_PER_LONG < 64)
-/*
- * called with urcu_mutex held.
- */
-static void switch_next_urcu_qparity(void)
-{
- STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR);
-}
-
+#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_qs_gp;
+ was_online = URCU_TLS(rcu_reader).ctr;
/* 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_qs_gp, 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);
- internal_urcu_lock();
+ mutex_lock(&rcu_gp_lock);
- switch_next_urcu_qparity(); /* 0 -> 1 */
+ /*
+ * Move all waiters into our local queue.
+ */
+ urcu_move_waiters(&waiters, &gp_waiters);
+
+ if (cds_list_empty(®istry))
+ goto out;
/*
- * Must commit qparity update to memory before waiting for parity
- * 0 quiescent state. 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.
+ * Wait for readers to observe original parity or be quiescent.
*/
+ wait_for_readers(®istry, &cur_snap_readers, &qsreaders);
/*
- * Wait for previous parity to be empty of readers.
+ * 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.
*/
- wait_for_quiescent_state(); /* Wait readers in parity 0 */
+ cmm_barrier();
/*
- * 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.
+ * 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_next_urcu_qparity(); /* 1 -> 0 */
+ /* Switch parity: 0 -> 1, 1 -> 0 */
+ CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR);
/*
- * Must commit qparity update to memory before waiting for parity
- * 1 quiescent state. 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 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();
/*
- * Wait for previous parity to be empty of readers.
+ * 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_for_quiescent_state(); /* Wait readers in parity 1 */
+ cmm_smp_mb();
- internal_urcu_unlock();
+ /*
+ * 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_qs_gp, LOAD_SHARED(urcu_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_qs_gp;
+ 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.
+ * our own quiescent state. This allows using synchronize_rcu()
+ * in threads registered as readers.
*/
- smp_mb();
if (was_online)
- STORE_SHARED(rcu_reader_qs_gp, 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);
+
+ /*
+ * Move all waiters into our local queue.
+ */
+ urcu_move_waiters(&waiters, &gp_waiters);
+
+ if (cds_list_empty(®istry))
+ goto out;
+
+ /* Increment current G.P. */
+ CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr + RCU_GP_CTR);
- internal_urcu_lock();
- STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR);
- wait_for_quiescent_state();
- internal_urcu_unlock();
+ /*
+ * 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_qs_gp, LOAD_SHARED(urcu_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.
_rcu_read_unlock();
}
-void *rcu_dereference(void *p)
-{
- return _rcu_dereference(p);
-}
-
-void *rcu_assign_pointer_sym(void **p, void *v)
-{
- wmb();
- return STORE_SHARED(p, v);
-}
-
-void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
-{
- wmb();
- return uatomic_cmpxchg(p, old, _new);
-}
-
-void *rcu_xchg_pointer_sym(void **p, void *v)
-{
- wmb();
- return uatomic_xchg(p, v);
-}
-
-void *rcu_publish_content_sym(void **p, void *v)
-{
- void *oldptr;
-
- oldptr = _rcu_xchg_pointer(p, v);
- synchronize_rcu();
- return oldptr;
-}
-
void rcu_quiescent_state(void)
{
_rcu_quiescent_state();
_rcu_thread_online();
}
-static void rcu_add_reader(pthread_t id)
-{
- struct reader_registry *oldarray;
-
- if (!registry) {
- alloc_readers = INIT_NUM_THREADS;
- num_readers = 0;
- registry =
- malloc(sizeof(struct reader_registry) * alloc_readers);
- }
- if (alloc_readers < num_readers + 1) {
- oldarray = registry;
- registry = malloc(sizeof(struct reader_registry)
- * (alloc_readers << 1));
- memcpy(registry, oldarray,
- sizeof(struct reader_registry) * alloc_readers);
- alloc_readers <<= 1;
- free(oldarray);
- }
- registry[num_readers].tid = id;
- /* reference to the TLS of _this_ reader thread. */
- registry[num_readers].rcu_reader_qs_gp = &rcu_reader_qs_gp;
- num_readers++;
-}
-
-/*
- * Never shrink (implementation limitation).
- * This is O(nb threads). Eventually use a hash table.
- */
-static void rcu_remove_reader(pthread_t id)
-{
- struct reader_registry *index;
-
- assert(registry != NULL);
- for (index = registry; index < registry + num_readers; index++) {
- if (pthread_equal(index->tid, id)) {
- memcpy(index, ®istry[num_readers - 1],
- sizeof(struct reader_registry));
- registry[num_readers - 1].tid = 0;
- registry[num_readers - 1].rcu_reader_qs_gp = NULL;
- num_readers--;
- return;
- }
- }
- /* Hrm not found, forgot to register ? */
- assert(0);
-}
-
void rcu_register_thread(void)
{
- internal_urcu_lock();
- rcu_add_reader(pthread_self());
- internal_urcu_unlock();
+ URCU_TLS(rcu_reader).tid = pthread_self();
+ assert(URCU_TLS(rcu_reader).ctr == 0);
+
+ mutex_lock(&rcu_gp_lock);
+ cds_list_add(&URCU_TLS(rcu_reader).node, ®istry);
+ mutex_unlock(&rcu_gp_lock);
_rcu_thread_online();
}
* with a waiting writer.
*/
_rcu_thread_offline();
- internal_urcu_lock();
- rcu_remove_reader(pthread_self());
- internal_urcu_unlock();
+ mutex_lock(&rcu_gp_lock);
+ cds_list_del(&URCU_TLS(rcu_reader).node);
+ mutex_unlock(&rcu_gp_lock);
}
+
+void rcu_exit(void)
+{
+ /*
+ * 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"