X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu.c;h=8cb32a117bb065405ca3f45e0224dfe8fbebbf5d;hp=162ce00fa2d4c5516ef6093722d4dfc7b17cda23;hb=8895e525a86e679e775b8267546799a0e7bc569f;hpb=b715b99eb0d52e0744bfea46c62849f260a5ec58

diff --git a/urcu.c b/urcu.c
index 162ce00..8cb32a1 100644
--- a/urcu.c
+++ b/urcu.c
@@ -77,11 +77,39 @@ static void switch_next_urcu_qparity(void)
 }
 
 #ifdef DEBUG_FULL_MB
+static void force_mb_single_thread(pthread_t tid)
+{
+	smp_mb();
+}
+
 static void force_mb_all_threads(void)
 {
 	smp_mb();
 }
 #else
+
+static void force_mb_single_thread(pthread_t tid)
+{
+	assert(reader_data);
+	sig_done = 0;
+	/*
+	 * 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 (LOAD_SHARED(sig_done) < 1)
+		cpu_relax();
+	smp_mb();	/* read sig_done before ending the barrier */
+}
+
 static void force_mb_all_threads(void)
 {
 	struct reader_data *index;
@@ -91,24 +119,24 @@ static void force_mb_all_threads(void)
 	 */
 	if (!reader_data)
 		return;
-	debug_yield_write();
 	sig_done = 0;
-	debug_yield_write();
-	smp_mb();	/* write sig_done before sending the signals */
-	debug_yield_write();
-	for (index = reader_data; index < reader_data + num_readers; index++) {
+	/*
+	 * 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);
-		debug_yield_write();
-	}
 	/*
 	 * Wait for sighandler (and thus mb()) to execute on every thread.
 	 * BUSY-LOOP.
 	 */
-	while (sig_done < num_readers)
-		barrier();
-	debug_yield_write();
+	while (LOAD_SHARED(sig_done) < num_readers)
+		cpu_relax();
 	smp_mb();	/* read sig_done before ending the barrier */
-	debug_yield_write();
 }
 #endif
 
@@ -118,30 +146,37 @@ void wait_for_quiescent_state(void)
 
 	if (!reader_data)
 		return;
-	/* Wait for each thread urcu_active_readers count to become 0.
+	/*
+	 * Wait for each thread urcu_active_readers count to become 0.
 	 */
 	for (index = reader_data; index < reader_data + num_readers; index++) {
+		int wait_loops = 0;
 		/*
-		 * BUSY-LOOP.
+		 * 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))
-			barrier();
+		while (rcu_old_gp_ongoing(index->urcu_active_readers)) {
+			if (wait_loops++ == KICK_READER_LOOPS) {
+				force_mb_single_thread(index->tid);
+				wait_loops = 0;
+			} else {
+				cpu_relax();
+			}
+		}
 	}
 }
 
 void synchronize_rcu(void)
 {
+	internal_urcu_lock();
+
 	/* All threads should read qparity before accessing data structure
-	 * where new ptr points to. */
+	 * where new ptr points to. Must be done within internal_urcu_lock
+	 * because it iterates on reader threads.*/
 	/* Write new ptr before changing the qparity */
 	force_mb_all_threads();
-	debug_yield_write();
-
-	internal_urcu_lock();
-	debug_yield_write();
 
 	switch_next_urcu_qparity();	/* 0 -> 1 */
-	debug_yield_write();
 
 	/*
 	 * Must commit qparity update to memory before waiting for parity
@@ -149,13 +184,12 @@ void synchronize_rcu(void)
 	 * waiting forever while new readers are always accessing data (no
 	 * progress).
 	 */
-	smp_mb();
+	smp_mc();
 
 	/*
 	 * Wait for previous parity to be empty of readers.
 	 */
 	wait_for_quiescent_state();	/* Wait readers in parity 0 */
-	debug_yield_write();
 
 	/*
 	 * Must finish waiting for quiescent state for parity 0 before
@@ -163,10 +197,9 @@ void synchronize_rcu(void)
 	 * the writer waiting forever while new readers are always accessing
 	 * data (no progress).
 	 */
-	smp_mb();
+	smp_mc();
 
 	switch_next_urcu_qparity();	/* 1 -> 0 */
-	debug_yield_write();
 
 	/*
 	 * Must commit qparity update to memory before waiting for parity
@@ -174,23 +207,19 @@ void synchronize_rcu(void)
 	 * waiting forever while new readers are always accessing data (no
 	 * progress).
 	 */
-	smp_mb();
+	smp_mc();
 
 	/*
 	 * Wait for previous parity to be empty of readers.
 	 */
 	wait_for_quiescent_state();	/* Wait readers in parity 1 */
-	debug_yield_write();
-
-	internal_urcu_unlock();
-	debug_yield_write();
 
-	/* All threads should finish using the data referred to by old ptr
-	 * before decrementing their urcu_active_readers count */
 	/* Finish waiting for reader threads before letting the old ptr being
-	 * freed. */
+	 * freed. Must be done within internal_urcu_lock because it iterates on
+	 * reader threads. */
 	force_mb_all_threads();
-	debug_yield_write();
+
+	internal_urcu_unlock();
 }
 
 void urcu_add_reader(pthread_t id)
@@ -258,6 +287,11 @@ void urcu_unregister_thread(void)
 #ifndef DEBUG_FULL_MB
 void sigurcu_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
+	 * executed on.
+	 */
 	smp_mb();
 	atomic_inc(&sig_done);
 }