X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-qsbr.c;h=dac664935d7b898bd3dcce494845c7471c09749d;hp=c0e643dab54f460b59e3646ade4af850fd9d05fb;hb=ba59a0c7b244a0939a2298fc76a9002436ef9674;hpb=f0f7dbdd9ba9bff3860d8aea63a3e01b158351de

diff --git a/urcu-qsbr.c b/urcu-qsbr.c
index c0e643d..dac6649 100644
--- a/urcu-qsbr.c
+++ b/urcu-qsbr.c
@@ -37,12 +37,14 @@
 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 #include "urcu-qsbr.h"
 
-pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+static pthread_mutex_t urcu_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+int gp_futex;
 
 /*
  * Global grace period counter.
  */
-unsigned long urcu_gp_ctr = 1;
+unsigned long urcu_gp_ctr = RCU_GP_ONLINE;
 
 /*
  * Written to only by each individual reader. Read by both the reader and the
@@ -99,12 +101,26 @@ static void internal_urcu_unlock(void)
 	}
 }
 
-#ifdef HAS_INCOHERENT_CACHES
-static void force_mb_single_thread(struct reader_registry *index)
+/*
+ * synchronize_rcu() waiting. Single thread.
+ */
+static void wait_gp(struct reader_registry *index)
 {
-	smp_mb();
+	atomic_dec(&gp_futex);
+	smp_mb(); /* Write futex before read reader_gp */
+	if (!rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
+		/* Read reader_gp before write futex */
+		smp_mb();
+		/* Callbacks are queued, don't wait. */
+		atomic_set(&gp_futex, 0);
+	} else {
+		/* Read reader_gp before read futex */
+		smp_rmb();
+		if (atomic_read(&gp_futex) == -1)
+			futex(&gp_futex, FUTEX_WAIT, -1,
+			      NULL, NULL, 0);
+	}
 }
-#endif /* #ifdef HAS_INCOHERENT_CACHES */
 
 static void wait_for_quiescent_state(void)
 {
@@ -116,27 +132,107 @@ static void wait_for_quiescent_state(void)
 	 * Wait for each thread rcu_reader_qs_gp count to become 0.
 	 */
 	for (index = registry; index < registry + num_readers; index++) {
-#ifndef HAS_INCOHERENT_CACHES
-		while (rcu_gp_ongoing(index->rcu_reader_qs_gp))
-			cpu_relax();
-#else /* #ifndef HAS_INCOHERENT_CACHES */
 		int wait_loops = 0;
-		/*
-		 * BUSY-LOOP. Force the reader thread to commit its
-		 * rcu_reader_qs_gp update to memory if we wait for too long.
-		 */
+
 		while (rcu_gp_ongoing(index->rcu_reader_qs_gp)) {
-			if (wait_loops++ == KICK_READER_LOOPS) {
-				force_mb_single_thread(index);
-				wait_loops = 0;
+			if (wait_loops++ == RCU_QS_ACTIVE_ATTEMPTS) {
+				wait_gp(index);
 			} else {
+#ifndef HAS_INCOHERENT_CACHES
 				cpu_relax();
+#else /* #ifndef HAS_INCOHERENT_CACHES */
+				smp_mb();
+#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 			}
 		}
-#endif /* #else #ifndef HAS_INCOHERENT_CACHES */
 	}
 }
 
+/*
+ * Using a two-subphases algorithm for architectures with smaller than 64-bit
+ * long-size to ensure we do not encounter an overflow bug.
+ */
+
+#if (BITS_PER_LONG < 64)
+/*
+ * called with urcu_mutex held.
+ */
+static void switch_next_urcu_qparity(void)
+{
+	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr ^ RCU_GP_CTR);
+}
+
+void synchronize_rcu(void)
+{
+	unsigned long was_online;
+
+	was_online = rcu_reader_qs_gp;
+
+	/* All threads should read qparity before accessing data structure
+	 * where new ptr points to.
+	 */
+	/* Write new ptr before changing the qparity */
+	smp_mb();
+
+	/*
+	 * Mark the writer thread offline to make sure we don't wait for
+	 * our own quiescent state. This allows using synchronize_rcu() in
+	 * threads registered as readers.
+	 */
+	if (was_online)
+		STORE_SHARED(rcu_reader_qs_gp, 0);
+
+	internal_urcu_lock();
+
+	switch_next_urcu_qparity();	/* 0 -> 1 */
+
+	/*
+	 * Must commit qparity update to memory before waiting for parity
+	 * 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.
+	 */
+
+	/*
+	 * Wait for previous parity to be empty of readers.
+	 */
+	wait_for_quiescent_state();	/* Wait readers in parity 0 */
+
+	/*
+	 * 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.
+	 */
+
+	switch_next_urcu_qparity();	/* 1 -> 0 */
+
+	/*
+	 * Must commit qparity update to memory before waiting for parity
+	 * 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.
+	 */
+
+	/*
+	 * Wait for previous parity to be empty of readers.
+	 */
+	wait_for_quiescent_state();	/* Wait readers in parity 1 */
+
+	internal_urcu_unlock();
+
+	/*
+	 * Finish waiting for reader threads before letting the old ptr being
+	 * freed.
+	 */
+	if (was_online)
+		_STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
+	smp_mb();
+}
+#else /* !(BITS_PER_LONG < 64) */
 void synchronize_rcu(void)
 {
 	unsigned long was_online;
@@ -153,7 +249,7 @@ void synchronize_rcu(void)
 		STORE_SHARED(rcu_reader_qs_gp, 0);
 
 	internal_urcu_lock();
-	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + 2);
+	STORE_SHARED(urcu_gp_ctr, urcu_gp_ctr + RCU_GP_CTR);
 	wait_for_quiescent_state();
 	internal_urcu_unlock();
 
@@ -161,6 +257,7 @@ void synchronize_rcu(void)
 		_STORE_SHARED(rcu_reader_qs_gp, LOAD_SHARED(urcu_gp_ctr));
 	smp_mb();
 }
+#endif  /* !(BITS_PER_LONG < 64) */
 
 /*
  * library wrappers to be used by non-LGPL compatible source code.
@@ -187,6 +284,12 @@ void *rcu_assign_pointer_sym(void **p, void *v)
 	return STORE_SHARED(p, v);
 }
 
+void *rcu_cmpxchg_pointer_sym(void **p, void *old, void *_new)
+{
+	wmb();
+	return cmpxchg(p, old, _new);
+}
+
 void *rcu_xchg_pointer_sym(void **p, void *v)
 {
 	wmb();