#define _BSD_SOURCE
#define _GNU_SOURCE
#define _LGPL_SOURCE
+#define _DEFAULT_SOURCE
#include <stdio.h>
#include <pthread.h>
#include <signal.h>
/*
* 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().
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;
+/*
+ * 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.
+ */
+static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER;
int32_t gp_futex;
* Written to only by writer with mutex taken. Read by both writer and readers.
*/
unsigned long rcu_gp_ctr = RCU_GP_COUNT;
-
/*
* Written to only by each individual reader. Read by both the reader and the
* writers.
*/
-DEFINE_URCU_TLS(struct rcu_reader, rcu_reader);
+__DEFINE_URCU_TLS_GLOBAL(struct rcu_reader, rcu_reader);
#ifdef DEBUG_YIELD
unsigned int yield_active;
-DEFINE_URCU_TLS(unsigned int, rand_yield);
+__DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield);
#endif
static CDS_LIST_HEAD(registry);
NULL, NULL, 0);
}
+/*
+ * Always called with rcu_registry lock held. Releases this lock between
+ * iterations and grabs it again. Holds the lock when it returns.
+ */
void update_counter_and_wait(void)
{
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 */
/* Switch parity: 0 -> 1, 1 -> 0 */
CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
* Wait for each thread URCU_TLS(rcu_reader).ctr count to become 0.
*/
for (;;) {
- wait_loops++;
- if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
+ wait_loops++;
+ if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
uatomic_dec(&gp_futex);
/* Write futex before read reader_gp */
smp_mb_master(RCU_MB_GROUP);
#ifndef HAS_INCOHERENT_CACHES
if (cds_list_empty(®istry)) {
- if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
/* Read reader_gp before write futex */
smp_mb_master(RCU_MB_GROUP);
uatomic_set(&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 */
/*
* for too long.
*/
if (cds_list_empty(®istry)) {
- if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
+ if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) {
/* Read reader_gp before write futex */
smp_mb_master(RCU_MB_GROUP);
uatomic_set(&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 */
}
void synchronize_rcu(void)
{
mutex_lock(&rcu_gp_lock);
+ 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 previous parity to be empty of readers.
+ * update_counter_and_wait() can release and grab again
+ * rcu_registry_lock interally.
*/
update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */
/*
* Wait for previous parity to be empty of readers.
+ * update_counter_and_wait() can release and grab again
+ * rcu_registry_lock interally.
*/
update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */
- /* 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. */
+ /*
+ * 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);
}
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
* 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)
{