#include <unistd.h>
#include <sys/mman.h>
+#include "urcu/arch.h"
#include "urcu/wfcqueue.h"
#include "urcu/map/urcu-bp.h"
#include "urcu/static/urcu-bp.h"
static
int rcu_bp_refcount;
+/* If the headers do not support membarrier system call, fall back smp_mb. */
+#ifdef __NR_membarrier
+# define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__)
+#else
+# define membarrier(...) -ENOSYS
+#endif
+
+enum membarrier_cmd {
+ MEMBARRIER_CMD_QUERY = 0,
+ MEMBARRIER_CMD_SHARED = (1 << 0),
+};
+
static
void __attribute__((constructor)) rcu_bp_init(void);
static
void __attribute__((destructor)) rcu_bp_exit(void);
+int urcu_bp_has_sys_membarrier;
+
+/*
+ * 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;
static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER;
static int initialized;
urcu_die(ret);
}
+static void smp_mb_master(void)
+{
+ if (caa_likely(urcu_bp_has_sys_membarrier))
+ (void) membarrier(MEMBARRIER_CMD_SHARED, 0);
+ else
+ cmm_smp_mb();
+}
+
+/*
+ * 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)
if (cds_list_empty(input_readers)) {
break;
} else {
+ /* Temporarily unlock the registry lock. */
+ mutex_unlock(&rcu_registry_lock);
if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS)
(void) poll(NULL, 0, RCU_SLEEP_DELAY_MS);
else
caa_cpu_relax();
+ /* Re-lock the registry lock before the next loop. */
+ mutex_lock(&rcu_registry_lock);
}
}
}
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. */
/* Write new ptr before changing the qparity */
- cmm_smp_mb();
+ smp_mb_master();
/*
* 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(®istry, &cur_snap_readers, &qsreaders);
/*
* 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(&cur_snap_readers, NULL, &qsreaders);
* Finish waiting for reader threads before letting the old ptr being
* freed.
*/
- cmm_smp_mb();
+ smp_mb_master();
out:
+ mutex_unlock(&rcu_registry_lock);
mutex_unlock(&rcu_gp_lock);
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
assert(!ret);
-1, 0);
if (new_chunk == MAP_FAILED)
abort();
- bzero(new_chunk, new_chunk_len);
+ memset(new_chunk, 0, new_chunk_len);
new_chunk->data_len =
new_chunk_len - sizeof(struct registry_chunk);
cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
if (new_chunk != MAP_FAILED) {
/* Should not have moved. */
assert(new_chunk == last_chunk);
- bzero((char *) last_chunk + old_chunk_len,
+ memset((char *) last_chunk + old_chunk_len, 0,
new_chunk_len - old_chunk_len);
last_chunk->data_len =
new_chunk_len - sizeof(struct registry_chunk);
-1, 0);
if (new_chunk == MAP_FAILED)
abort();
- bzero(new_chunk, new_chunk_len);
+ memset(new_chunk, 0, new_chunk_len);
new_chunk->data_len =
new_chunk_len - sizeof(struct registry_chunk);
cds_list_add_tail(&new_chunk->node, &arena->chunk_list);
*/
rcu_bp_init();
- mutex_lock(&rcu_gp_lock);
+ mutex_lock(&rcu_registry_lock);
add_thread();
- mutex_unlock(&rcu_gp_lock);
+ mutex_unlock(&rcu_registry_lock);
end:
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
if (ret)
if (ret)
abort();
- mutex_lock(&rcu_gp_lock);
+ mutex_lock(&rcu_registry_lock);
remove_thread(rcu_reader_reg);
- mutex_unlock(&rcu_gp_lock);
+ mutex_unlock(&rcu_registry_lock);
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
if (ret)
abort();
urcu_bp_thread_exit_notifier);
if (ret)
abort();
+ ret = membarrier(MEMBARRIER_CMD_QUERY, 0);
+ if (ret >= 0 && (ret & MEMBARRIER_CMD_SHARED)) {
+ urcu_bp_has_sys_membarrier = 1;
+ }
initialized = 1;
}
mutex_unlock(&init_lock);
}
/*
- * Holding the rcu_gp_lock across fork will make sure we fork() don't race with
- * a concurrent thread executing with this same lock held. This ensures that the
- * registry is in a coherent state in the child.
+ * Holding the rcu_gp_lock and rcu_registry_lock across fork will make
+ * sure we fork() don't race with a concurrent thread executing with
+ * any of those locks held. This ensures that the registry and data
+ * protected by rcu_gp_lock are in a coherent state in the child.
*/
void rcu_bp_before_fork(void)
{
ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
assert(!ret);
mutex_lock(&rcu_gp_lock);
+ mutex_lock(&rcu_registry_lock);
saved_fork_signal_mask = oldmask;
}
int ret;
oldmask = saved_fork_signal_mask;
+ mutex_unlock(&rcu_registry_lock);
mutex_unlock(&rcu_gp_lock);
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
assert(!ret);
/*
* Prune all entries from registry except our own thread. Fits the Linux
- * fork behavior. Called with rcu_gp_lock held.
+ * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held.
*/
static
void urcu_bp_prune_registry(void)
urcu_bp_prune_registry();
oldmask = saved_fork_signal_mask;
+ mutex_unlock(&rcu_registry_lock);
mutex_unlock(&rcu_gp_lock);
ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
assert(!ret);