X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=b4a5a7db948b8d57a03960747c5257d9631b1c7c;hp=ffdab3c2fc1eacbad2c4e73f5e89e83261fdb8d5;hb=06f22bdbb0c4c4d5db42a2e2dc35818aa61415be;hpb=32c15e4e6fa2c368f4649d904c91e4293a1193e3

diff --git a/urcu.c b/urcu.c
index ffdab3c..b4a5a7d 100644
--- a/urcu.c
+++ b/urcu.c
@@ -3,7 +3,7 @@
  *
  * Userspace RCU 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
@@ -67,7 +67,7 @@ int gp_futex;
  * 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 rcu_gp_ctr = RCU_GP_COUNT;
+unsigned long rcu_gp_ctr = RCU_GP_COUNT;
 
 /*
  * Written to only by each individual reader. Read by both the reader and the
@@ -99,10 +99,10 @@ static void mutex_lock(pthread_mutex_t *mutex)
 			perror("Error in pthread mutex lock");
 			exit(-1);
 		}
-		if (rcu_reader.need_mb) {
-			smp_mb();
-			rcu_reader.need_mb = 0;
-			smp_mb();
+		if (CAA_LOAD_SHARED(rcu_reader.need_mb)) {
+			cmm_smp_mb();
+			_CAA_STORE_SHARED(rcu_reader.need_mb, 0);
+			cmm_smp_mb();
 		}
 		poll(NULL,0,10);
 	}
@@ -121,19 +121,19 @@ static void mutex_unlock(pthread_mutex_t *mutex)
 }
 
 #ifdef RCU_MEMBARRIER
-static void smp_mb_heavy(void)
+static void smp_mb_master(int group)
 {
 	if (likely(has_sys_membarrier))
 		membarrier(MEMBARRIER_EXPEDITED);
 	else
-		smp_mb();
+		cmm_smp_mb();
 }
 #endif
 
 #ifdef RCU_MB
-static void smp_mb_heavy(void)
+static void smp_mb_master(int group)
 {
-	smp_mb();
+	cmm_smp_mb();
 }
 #endif
 
@@ -143,20 +143,19 @@ static void force_mb_all_readers(void)
 	struct rcu_reader *index;
 
 	/*
-	 * Ask for each threads to execute a smp_mb() so we can consider the
+	 * Ask for each threads to execute a cmm_smp_mb() so we can consider the
 	 * compiler barriers around rcu read lock as real memory barriers.
 	 */
 	if (list_empty(&registry))
 		return;
 	/*
-	 * pthread_kill has a smp_mb(). But beware, we assume it performs
+	 * pthread_kill has a cmm_smp_mb(). But beware, we assume it performs
 	 * a cache flush on architectures with non-coherent cache. Let's play
-	 * safe and don't assume anything : we use smp_mc() to make sure the
+	 * safe and don't assume anything : we use cmm_smp_mc() to make sure the
 	 * cache flush is enforced.
 	 */
-	list_for_each_entry(index, &registry, head) {
-		index->need_mb = 1;
-		smp_mc();	/* write need_mb before sending the signal */
+	list_for_each_entry(index, &registry, node) {
+		CAA_STORE_SHARED(index->need_mb, 1);
 		pthread_kill(index->tid, SIGRCU);
 	}
 	/*
@@ -172,16 +171,16 @@ static void force_mb_all_readers(void)
 	 * relevant bug report.  For Linux kernels, we recommend getting
 	 * the Linux Test Project (LTP).
 	 */
-	list_for_each_entry(index, &registry, head) {
-		while (index->need_mb) {
+	list_for_each_entry(index, &registry, node) {
+		while (CAA_LOAD_SHARED(index->need_mb)) {
 			pthread_kill(index->tid, SIGRCU);
 			poll(NULL, 0, 1);
 		}
 	}
-	smp_mb();	/* read ->need_mb before ending the barrier */
+	cmm_smp_mb();	/* read ->need_mb before ending the barrier */
 }
 
-static void smp_mb_heavy(void)
+static void smp_mb_master(int group)
 {
 	force_mb_all_readers();
 }
@@ -193,7 +192,7 @@ static void smp_mb_heavy(void)
 static void wait_gp(void)
 {
 	/* Read reader_gp before read futex */
-	smp_mb_heavy();
+	smp_mb_master(RCU_MB_GROUP);
 	if (uatomic_read(&gp_futex) == -1)
 		futex_async(&gp_futex, FUTEX_WAIT, -1,
 		      NULL, NULL, 0);
@@ -206,21 +205,23 @@ void update_counter_and_wait(void)
 	struct rcu_reader *index, *tmp;
 
 	/* Switch parity: 0 -> 1, 1 -> 0 */
-	STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
+	CAA_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE);
 
 	/*
-	 * Must commit qparity update to memory before waiting for other parity
-	 * 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 rcu_reader ctr.
 	 */
+	cmm_barrier();
 
 	/*
-	 * Adding a smp_mb() which is _not_ formally required, but makes the
+	 *
+	 * 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.
 	 */
-	smp_mb();
+	cmm_smp_mb();
 
 	/*
 	 * Wait for each thread rcu_reader.ctr count to become 0.
@@ -230,19 +231,19 @@ void update_counter_and_wait(void)
 		if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 			uatomic_dec(&gp_futex);
 			/* Write futex before read reader_gp */
-			smp_mb_heavy();
+			smp_mb_master(RCU_MB_GROUP);
 		}
 
-		list_for_each_entry_safe(index, tmp, &registry, head) {
-			if (!rcu_old_gp_ongoing(&index->ctr))
-				list_move(&index->head, &qsreaders);
+		list_for_each_entry_safe(index, tmp, &registry, node) {
+			if (!rcu_gp_ongoing(&index->ctr))
+				list_move(&index->node, &qsreaders);
 		}
 
 #ifndef HAS_INCOHERENT_CACHES
 		if (list_empty(&registry)) {
 			if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 				/* Read reader_gp before write futex */
-				smp_mb_heavy();
+				smp_mb_master(RCU_MB_GROUP);
 				uatomic_set(&gp_futex, 0);
 			}
 			break;
@@ -250,7 +251,7 @@ void update_counter_and_wait(void)
 			if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS)
 				wait_gp();
 			else
-				cpu_relax();
+				caa_cpu_relax();
 		}
 #else /* #ifndef HAS_INCOHERENT_CACHES */
 		/*
@@ -260,7 +261,7 @@ void update_counter_and_wait(void)
 		if (list_empty(&registry)) {
 			if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) {
 				/* Read reader_gp before write futex */
-				smp_mb_heavy();
+				smp_mb_master(RCU_MB_GROUP);
 				uatomic_set(&gp_futex, 0);
 			}
 			break;
@@ -270,11 +271,11 @@ void update_counter_and_wait(void)
 				wait_gp();
 				break; /* only escape switch */
 			case KICK_READER_LOOPS:
-				smp_mb_heavy();
+				smp_mb_master(RCU_MB_GROUP);
 				wait_loops = 0;
 				break; /* only escape switch */
 			default:
-				cpu_relax();
+				caa_cpu_relax();
 			}
 		}
 #endif /* #else #ifndef HAS_INCOHERENT_CACHES */
@@ -294,7 +295,7 @@ void synchronize_rcu(void)
 	 * where new ptr points to. Must be done within rcu_gp_lock because it
 	 * iterates on reader threads.*/
 	/* Write new ptr before changing the qparity */
-	smp_mb_heavy();
+	smp_mb_master(RCU_MB_GROUP);
 
 	/*
 	 * Wait for previous parity to be empty of readers.
@@ -303,18 +304,19 @@ void synchronize_rcu(void)
 
 	/*
 	 * 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.
+	 * 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
+	 * rcu_reader ctr before store to rcu_gp_ctr.
 	 */
+	cmm_barrier();
 
 	/*
-	 * Adding a smp_mb() which is _not_ formally required, but makes the
+	 * 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.
 	 */
-	smp_mb();
+	cmm_smp_mb();
 
 	/*
 	 * Wait for previous parity to be empty of readers.
@@ -324,7 +326,7 @@ void synchronize_rcu(void)
 	/* 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. */
-	smp_mb_heavy();
+	smp_mb_master(RCU_MB_GROUP);
 out:
 	mutex_unlock(&rcu_gp_lock);
 }
@@ -347,18 +349,18 @@ void rcu_register_thread(void)
 {
 	rcu_reader.tid = pthread_self();
 	assert(rcu_reader.need_mb == 0);
-	assert(rcu_reader.ctr == 0);
+	assert(!(rcu_reader.ctr & RCU_GP_CTR_NEST_MASK));
 
 	mutex_lock(&rcu_gp_lock);
 	rcu_init();	/* In case gcc does not support constructor attribute */
-	list_add(&rcu_reader.head, &registry);
+	list_add(&rcu_reader.node, &registry);
 	mutex_unlock(&rcu_gp_lock);
 }
 
 void rcu_unregister_thread(void)
 {
 	mutex_lock(&rcu_gp_lock);
-	list_del(&rcu_reader.head);
+	list_del(&rcu_reader.node);
 	mutex_unlock(&rcu_gp_lock);
 }
 
@@ -377,13 +379,13 @@ void rcu_init(void)
 static void sigrcu_handler(int signo, siginfo_t *siginfo, void *context)
 {
 	/*
-	 * Executing this smp_mb() is the only purpose of this signal handler.
-	 * It punctually promotes barrier() into smp_mb() on every thread it is
+	 * Executing this cmm_smp_mb() is the only purpose of this signal handler.
+	 * It punctually promotes cmm_barrier() into cmm_smp_mb() on every thread it is
 	 * executed on.
 	 */
-	smp_mb();
-	rcu_reader.need_mb = 0;
-	smp_mb();
+	cmm_smp_mb();
+	_CAA_STORE_SHARED(rcu_reader.need_mb, 0);
+	cmm_smp_mb();
 }
 
 /*