X-Git-Url: http://git.liburcu.org/?p=userspace-rcu.git;a=blobdiff_plain;f=urcu-bp.c;h=3e5be5fb59e1baf0ed7a7a7b7c26f61603654a98;hp=71474bcbcb6fe85494635c004bfdeb0d5f7a3aae;hb=refs%2Fheads%2Fstable-0.7;hpb=e038e286679be66f3d8096ea9c7224a1f282a9c3

diff --git a/urcu-bp.c b/urcu-bp.c
index 71474bc..3e5be5f 100644
--- a/urcu-bp.c
+++ b/urcu-bp.c
@@ -79,8 +79,8 @@ void *mremap_wrapper(void *old_address, size_t old_size,
 }
 #endif
 
-/* Sleep delay in us */
-#define RCU_SLEEP_DELAY		1000
+/* Sleep delay in ms */
+#define RCU_SLEEP_DELAY_MS	10
 #define INIT_NR_THREADS		8
 #define ARENA_INIT_ALLOC		\
 	sizeof(struct registry_chunk)	\
@@ -91,12 +91,29 @@ void *mremap_wrapper(void *old_address, size_t old_size,
  */
 #define RCU_QS_ACTIVE_ATTEMPTS 100
 
+static
+int rcu_bp_refcount;
+
 static
 void __attribute__((constructor)) rcu_bp_init(void);
 static
-void __attribute__((destructor)) rcu_bp_exit(void);
+void __attribute__((destructor)) _rcu_bp_exit(void);
 
+/*
+ * 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;
@@ -105,7 +122,7 @@ static pthread_key_t urcu_bp_key;
 
 #ifdef DEBUG_YIELD
 unsigned int yield_active;
-DEFINE_URCU_TLS(unsigned int, rand_yield);
+__DEFINE_URCU_TLS_GLOBAL(unsigned int, rand_yield);
 #endif
 
 /*
@@ -120,7 +137,7 @@ long rcu_gp_ctr = RCU_GP_COUNT;
  * Pointer to registry elements. 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);
 
 static CDS_LIST_HEAD(registry);
 
@@ -168,10 +185,14 @@ static void mutex_unlock(pthread_mutex_t *mutex)
 		urcu_die(ret);
 }
 
+/*
+ * 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;
 
 	/* Switch parity: 0 -> 1, 1 -> 0 */
@@ -195,7 +216,9 @@ void update_counter_and_wait(void)
 	 * Wait for each thread rcu_reader.ctr count to become 0.
 	 */
 	for (;;) {
-		wait_loops++;
+		if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS)
+			wait_loops++;
+
 		cds_list_for_each_entry_safe(index, tmp, &registry, node) {
 			if (!rcu_old_gp_ongoing(&index->ctr))
 				cds_list_move(&index->node, &qsreaders);
@@ -204,10 +227,14 @@ void update_counter_and_wait(void)
 		if (cds_list_empty(&registry)) {
 			break;
 		} else {
-			if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
-				usleep(RCU_SLEEP_DELAY);
+			/* 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);
 		}
 	}
 	/* put back the reader list in the registry */
@@ -225,6 +252,7 @@ void synchronize_rcu(void)
 	assert(!ret);
 
 	mutex_lock(&rcu_gp_lock);
+	mutex_lock(&rcu_registry_lock);
 
 	if (cds_list_empty(&registry))
 		goto out;
@@ -236,6 +264,8 @@ void synchronize_rcu(void)
 
 	/*
 	 * 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 */
 
@@ -248,6 +278,8 @@ void synchronize_rcu(void)
 
 	/*
 	 * 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 */
 
@@ -257,6 +289,7 @@ void synchronize_rcu(void)
 	 */
 	cmm_smp_mb();
 out:
+	mutex_unlock(&rcu_registry_lock);
 	mutex_unlock(&rcu_gp_lock);
 	ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 	assert(!ret);
@@ -429,11 +462,8 @@ struct registry_chunk *find_chunk(struct rcu_reader *rcu_reader_reg)
 
 /* Called with signals off and mutex locked */
 static
-void remove_thread(void)
+void remove_thread(struct rcu_reader *rcu_reader_reg)
 {
-	struct rcu_reader *rcu_reader_reg;
-
-	rcu_reader_reg = URCU_TLS(rcu_reader);
 	cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg);
 	URCU_TLS(rcu_reader) = NULL;
 }
@@ -463,9 +493,9 @@ void rcu_bp_register(void)
 	 */
 	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)
@@ -474,7 +504,7 @@ end:
 
 /* Disable signals, take mutex, remove from registry */
 static
-void rcu_bp_unregister(void)
+void rcu_bp_unregister(struct rcu_reader *rcu_reader_reg)
 {
 	sigset_t newmask, oldmask;
 	int ret;
@@ -486,12 +516,13 @@ void rcu_bp_unregister(void)
 	if (ret)
 		abort();
 
-	mutex_lock(&rcu_gp_lock);
-	remove_thread();
-	mutex_unlock(&rcu_gp_lock);
+	mutex_lock(&rcu_registry_lock);
+	remove_thread(rcu_reader_reg);
+	mutex_unlock(&rcu_registry_lock);
 	ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 	if (ret)
 		abort();
+	_rcu_bp_exit();
 }
 
 /*
@@ -501,15 +532,14 @@ void rcu_bp_unregister(void)
 static
 void urcu_bp_thread_exit_notifier(void *rcu_key)
 {
-	assert(rcu_key == URCU_TLS(rcu_reader));
-	rcu_bp_unregister();
+	rcu_bp_unregister(rcu_key);
 }
 
 static
 void rcu_bp_init(void)
 {
 	mutex_lock(&init_lock);
-	if (!initialized) {
+	if (!rcu_bp_refcount++) {
 		int ret;
 
 		ret = pthread_key_create(&urcu_bp_key,
@@ -522,24 +552,39 @@ void rcu_bp_init(void)
 }
 
 static
-void rcu_bp_exit(void)
+void _rcu_bp_exit(void)
 {
-	struct registry_chunk *chunk, *tmp;
-	int ret;
+	mutex_lock(&init_lock);
+	if (!--rcu_bp_refcount) {
+		struct registry_chunk *chunk, *tmp;
+		int ret;
 
-	cds_list_for_each_entry_safe(chunk, tmp,
-			&registry_arena.chunk_list, node) {
-		munmap(chunk, chunk->data_len + sizeof(struct registry_chunk));
+		cds_list_for_each_entry_safe(chunk, tmp,
+				&registry_arena.chunk_list, node) {
+			munmap(chunk, chunk->data_len
+					+ sizeof(struct registry_chunk));
+		}
+		ret = pthread_key_delete(urcu_bp_key);
+		if (ret)
+			abort();
 	}
-	ret = pthread_key_delete(urcu_bp_key);
-	if (ret)
-		abort();
+	mutex_unlock(&init_lock);
+}
+
+/*
+ * Keep ABI compability within stable versions. This has never been
+ * exposed through a header, but needs to stay in the .so until the
+ * soname is bumped.
+ */
+void rcu_bp_exit(void)
+{
 }
 
 /*
- * 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)
 {
@@ -551,6 +596,7 @@ 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;
 }
 
@@ -560,6 +606,7 @@ void rcu_bp_after_fork_parent(void)
 	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);
@@ -567,7 +614,7 @@ void rcu_bp_after_fork_parent(void)
 
 /*
  * 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)
@@ -595,6 +642,7 @@ void rcu_bp_after_fork_child(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);