Seq objects would be really hard to use if they did not provide
acquire-release semantics. Currently they do that via
ovs_mutex_lock()/ovs_mutex_unlock(), respectively. Document the
behavior so that it is safer to rely on that elsewhere.
Signed-off-by: Jarno Rajahalme <jrajaha...@nicira.com>
---
lib/ovs-thread.c | 8 +++++++-
lib/seq.h | 6 +++++-
2 files changed, 12 insertions(+), 2 deletions(-)
diff --git a/lib/ovs-thread.c b/lib/ovs-thread.c
index efbd60f..6d352ea 100644
--- a/lib/ovs-thread.c
+++ b/lib/ovs-thread.c
@@ -278,7 +278,11 @@ ovs_barrier_destroy(struct ovs_barrier *barrier)
}
/* Makes the calling thread block on the 'barrier' until all
- * 'barrier->size' threads hit the barrier. */
+ * 'barrier->size' threads hit the barrier.
+ * ovs_barrier provides the necessary acquire-release semantics to make
+ * the effects of prior memory accesses of all the participating threads
+ * visible on return and to prevent the following memory accesses to be
+ * reordered before the ovs_barrier_block(). */
void
ovs_barrier_block(struct ovs_barrier *barrier)
{
@@ -288,6 +292,8 @@ ovs_barrier_block(struct ovs_barrier *barrier)
atomic_add(&barrier->count, 1, &orig);
if (orig + 1 == barrier->size) {
atomic_store(&barrier->count, 0);
+ /* seq_change() serves as a release barrier against the other threads,
+ * so the zeroed count is visible to them as they continue. */
seq_change(barrier->seq);
}
diff --git a/lib/seq.h b/lib/seq.h
index 38c0e52..31097ff 100644
--- a/lib/seq.h
+++ b/lib/seq.h
@@ -107,7 +107,11 @@
* Thread-safety
* =============
*
- * Fully thread safe.
+ * Fully thread safe. seq_read() and seq_wait() synchronize with
+ * seq_change() on the same variable in acquire-release fashion. That
+ * is, all effects of the memory accesses performed by a thread prior
+ * to seq_change() are visible to the threads returning from
+ * seq_read() or seq_wait() observing that change.
*/
#include <stdint.h>
--
1.7.10.4
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
