X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=b323711294ae25ffc7c65641ab1a73240aa03aaa;hp=efbb4b7890a71524fa8264877f3da63565291591;hb=63ff4873e046ba582d20e79ebff16ee7da2aa92f;hpb=67c2d80b2b808fe7a6d6d0660ca81c919a81010b

diff --git a/urcu.c b/urcu.c
index efbb4b7..b323711 100644
--- a/urcu.c
+++ b/urcu.c
@@ -36,12 +36,25 @@
 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 #include "urcu.h"
 
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+#ifndef URCU_MB
+static int init_done;
+
+void __attribute__((constructor)) urcu_init(void);
+void __attribute__((destructor)) urcu_exit(void);
+#else
+void urcu_init(void)
+{
+}
+#endif
+
+static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+int gp_futex;
 
 /*
  * Global grace period counter.
  * Contains the current RCU_GP_CTR_BIT.
- * Also has a RCU_GP_CTR_BIT of 1, to accelerate the reader fast path.
+ * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path.
  * Written to only by writer with mutex taken. Read by both writer and readers.
  */
 long urcu_gp_ctr = RCU_GP_COUNT;
@@ -70,7 +83,7 @@ static struct reader_registry *registry;
 static char __thread need_mb;
 static int num_readers, alloc_readers;
 
-void internal_urcu_lock(void)
+static void internal_urcu_lock(void)
 {
 	int ret;
 
@@ -97,7 +110,7 @@ void internal_urcu_lock(void)
 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 }
 
-void internal_urcu_unlock(void)
+static void internal_urcu_unlock(void)
 {
 	int ret;
 
@@ -116,20 +129,17 @@ static void switch_next_urcu_qparity(void)
 	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 }
 
-#ifdef DEBUG_FULL_MB
-#ifdef HAS_INCOHERENT_CACHES
+#ifdef URCU_MB
 static void force_mb_single_thread(struct reader_registry *index)
 {
 	smp_mb();
 }
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
 
 static void force_mb_all_threads(void)
 {
 	smp_mb();
 }
-#else /* #ifdef DEBUG_FULL_MB */
-#ifdef HAS_INCOHERENT_CACHES
+#else /* #ifdef URCU_MB */
 static void force_mb_single_thread(struct reader_registry *index)
 {
 	assert(registry);
@@ -152,7 +162,6 @@ static void force_mb_single_thread(struct reader_registry *index)
 	}
 	smp_mb();	/* read ->need_mb before ending the barrier */
 }
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
 
 static void force_mb_all_threads(void)
 {
@@ -195,7 +204,28 @@ static void force_mb_all_threads(void)
 	}
 	smp_mb();	/* read ->need_mb before ending the barrier */
 }
-#endif /* #else #ifdef DEBUG_FULL_MB */
+#endif /* #else #ifdef URCU_MB */
+
+/*
+ * synchronize_rcu() waiting. Single thread.
+ */
+static void wait_gp(struct reader_registry *index)
+{
+	uatomic_dec(&gp_futex);
+	force_mb_single_thread(index); /* Write futex before read reader_gp */
+	if (!rcu_old_gp_ongoing(index->urcu_active_readers)) {
+		/* Read reader_gp before write futex */
+		force_mb_single_thread(index);
+		/* Callbacks are queued, don't wait. */
+		uatomic_set(&gp_futex, 0);
+	} else {
+		/* Read reader_gp before read futex */
+		force_mb_single_thread(index);
+		if (uatomic_read(&gp_futex) == -1)
+			futex(&gp_futex, FUTEX_WAIT, -1,
+			      NULL, NULL, 0);
+	}
+}
 
 void wait_for_quiescent_state(void)
 {
@@ -207,20 +237,30 @@ void wait_for_quiescent_state(void)
 	 * Wait for each thread urcu_active_readers count to become 0.
 	 */
 	for (index = registry; index < registry + num_readers; index++) {
+		int wait_loops = 0;
 #ifndef HAS_INCOHERENT_CACHES
-		while (rcu_old_gp_ongoing(index->urcu_active_readers))
-			cpu_relax();
+		while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
+			if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
+				wait_gp(index);
+			} else {
+				cpu_relax();
+			}
+		}
 #else /* #ifndef HAS_INCOHERENT_CACHES */
-		int wait_loops = 0;
 		/*
 		 * BUSY-LOOP. Force the reader thread to commit its
 		 * urcu_active_readers update to memory if we wait for too long.
 		 */
 		while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
-			if (wait_loops++ == KICK_READER_LOOPS) {
+			switch (wait_loops++) {
+			case RCU_QS_ACTIVE_ATTEMPTS:
+				wait_gp(index);
+				break;
+			case KICK_READER_LOOPS:
 				force_mb_single_thread(index);
 				wait_loops = 0;
-			} else {
+				break;
+			default:
 				cpu_relax();
 			}
 		}
@@ -249,12 +289,11 @@ void synchronize_rcu(void)
 	 */
 
 	/*
-	 * Current RCU formal verification model assumes sequential execution of
-	 * the write-side. Add core synchronization instructions. Can be removed
-	 * if the formal model is extended to prove that reordering is still
-	 * correct.
+	 * Adding a 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.
 	 */
-	sync_core();	/* Formal model assumes serialized execution */
+	smp_mb();
 
 	/*
 	 * Wait for previous parity to be empty of readers.
@@ -269,7 +308,12 @@ void synchronize_rcu(void)
 	 * Ensured by STORE_SHARED and LOAD_SHARED.
 	 */
 
-	sync_core();	/* Formal model assumes serialized execution */
+	/*
+	 * Adding a 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.
+	 */
+	smp_mb();
 
 	switch_next_urcu_qparity();	/* 1 -> 0 */
 
@@ -281,7 +325,12 @@ void synchronize_rcu(void)
 	 * Ensured by STORE_SHARED and LOAD_SHARED.
 	 */
 
-	sync_core();	/* Formal model assumes serialized execution */
+	/*
+	 * Adding a 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.
+	 */
+	smp_mb();
 
 	/*
 	 * Wait for previous parity to be empty of readers.
@@ -324,7 +373,13 @@ void *rcu_assign_pointer_sym(void **p, void *v)
 void *rcu_xchg_pointer_sym(void **p, void *v)
 {
 	wmb();
-	return xchg(p, v);
+	return uatomic_xchg(p, v);
+}
+
+void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
+{
+	wmb();
+	return uatomic_cmpxchg(p, old, _new);
 }
 
 void *rcu_publish_content_sym(void **p, void *v)
@@ -388,6 +443,7 @@ static void rcu_remove_reader(pthread_t id)
 void rcu_register_thread(void)
 {
 	internal_urcu_lock();
+	urcu_init();	/* In case gcc does not support constructor attribute */
 	rcu_add_reader(pthread_self());
 	internal_urcu_unlock();
 }
@@ -399,7 +455,7 @@ void rcu_unregister_thread(void)
 	internal_urcu_unlock();
 }
 
-#ifndef DEBUG_FULL_MB
+#ifndef URCU_MB
 static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 {
 	/*
@@ -412,12 +468,26 @@ static void sigurcu_handler(int signo, siginfo_t *siginfo, void *context)
 	smp_mb();
 }
 
-void __attribute__((constructor)) urcu_init(void)
+/*
+ * urcu_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 internal_urcu_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 urcu_init(void)
 {
 	struct sigaction act;
 	int ret;
 
+	if (init_done)
+		return;
+	init_done = 1;
+
 	act.sa_sigaction = sigurcu_handler;
+	act.sa_flags = SA_SIGINFO | SA_RESTART;
+	sigemptyset(&act.sa_mask);
 	ret = sigaction(SIGURCU, &act, NULL);
 	if (ret) {
 		perror("Error in sigaction");
@@ -425,7 +495,7 @@ void __attribute__((constructor)) urcu_init(void)
 	}
 }
 
-void __attribute__((destructor)) urcu_exit(void)
+void urcu_exit(void)
 {
 	struct sigaction act;
 	int ret;
@@ -438,4 +508,4 @@ void __attribute__((destructor)) urcu_exit(void)
 	assert(act.sa_sigaction == sigurcu_handler);
 	free(registry);
 }
-#endif /* #ifndef DEBUG_FULL_MB */
+#endif /* #ifndef URCU_MB */