X-Git-Url: http://git.liburcu.org/?a=blobdiff_plain;f=urcu.c;h=5e2d2afff5641c57a80548c1f78469c52ccf7b09;hb=06d6106d9482f3a2805a2512ef0cda203088f2f5;hp=f2aae3433eeb310c16b6ed82a45aacafea76135a;hpb=135530fd6a47925c76a3e828f2fc7d6751bc9ff7;p=urcu.git

diff --git a/urcu.c b/urcu.c
index f2aae34..5e2d2af 100644
--- a/urcu.c
+++ b/urcu.c
@@ -23,9 +23,14 @@ pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
  * 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.
+ * Written to only by writer with mutex taken. Read by both writer and readers.
  */
 long urcu_gp_ctr = RCU_GP_COUNT;
 
+/*
+ * Written to only by each individual reader. Read by both the reader and the
+ * writers.
+ */
 long __thread urcu_active_readers;
 
 /* Thread IDs of registered readers */
@@ -73,7 +78,7 @@ void internal_urcu_unlock(void)
  */
 static void switch_next_urcu_qparity(void)
 {
-	urcu_gp_ctr ^= RCU_GP_CTR_BIT;
+	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR_BIT);
 }
 
 #ifdef DEBUG_FULL_MB
@@ -92,14 +97,21 @@ static void force_mb_single_thread(pthread_t tid)
 {
 	assert(reader_data);
 	sig_done = 0;
-	smp_mb();	/* write sig_done before sending the signals */
+	/*
+	 * pthread_kill has a 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
+	 * cache flush is enforced.
+	 * smp_mb();    write sig_done before sending the signals
+	 */
+	smp_mc();	/* write sig_done before sending the signals */
 	pthread_kill(tid, SIGURCU);
 	/*
 	 * Wait for sighandler (and thus mb()) to execute on every thread.
 	 * BUSY-LOOP.
 	 */
-	while (sig_done < 1)
-		smp_rmb();	/* ensure we re-read sig-done */
+	while (LOAD_SHARED(sig_done) < 1)
+		cpu_relax();
 	smp_mb();	/* read sig_done before ending the barrier */
 }
 
@@ -113,15 +125,22 @@ static void force_mb_all_threads(void)
 	if (!reader_data)
 		return;
 	sig_done = 0;
-	smp_mb();	/* write sig_done before sending the signals */
+	/*
+	 * pthread_kill has a 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
+	 * cache flush is enforced.
+	 * smp_mb();    write sig_done before sending the signals
+	 */
+	smp_mc();	/* write sig_done before sending the signals */
 	for (index = reader_data; index < reader_data + num_readers; index++)
 		pthread_kill(index->tid, SIGURCU);
 	/*
 	 * Wait for sighandler (and thus mb()) to execute on every thread.
 	 * BUSY-LOOP.
 	 */
-	while (sig_done < num_readers)
-		smp_rmb();	/* ensure we re-read sig-done */
+	while (LOAD_SHARED(sig_done) < num_readers)
+		cpu_relax();
 	smp_mb();	/* read sig_done before ending the barrier */
 }
 #endif
@@ -145,6 +164,8 @@ void wait_for_quiescent_state(void)
 			if (wait_loops++ == KICK_READER_LOOPS) {
 				force_mb_single_thread(index->tid);
 				wait_loops = 0;
+			} else {
+				cpu_relax();
 			}
 		}
 	}
@@ -167,8 +188,8 @@ void synchronize_rcu(void)
 	 * 0 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.
 	 */
-	smp_mb();
 
 	/*
 	 * Wait for previous parity to be empty of readers.
@@ -180,8 +201,8 @@ void synchronize_rcu(void)
 	 * 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.
 	 */
-	smp_mb();
 
 	switch_next_urcu_qparity();	/* 1 -> 0 */
 
@@ -190,8 +211,8 @@ void synchronize_rcu(void)
 	 * 1 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.
 	 */
-	smp_mb();
 
 	/*
 	 * Wait for previous parity to be empty of readers.