Hi Jan,
I think we need a memory barrier here between sync_in_process = 0 and totempg_trans_ack() or have to introduce a lock to protect them both. Otherwise, on some memory reordering multicore architectures, this code swaping seems meanless.
在 2013-3-21 下午5:11，"Jan Friesse" <jfriesse@xxxxxxxxxx>写道：
If IPC thread is waiting for allow_send and sync_in_process is changed
before trans_ack, message may be copied to incorrect queue. Solution is
to swap sync_in_process and trans_ack.
Signed-off-by: Jan Friesse <jfriesse@xxxxxxxxxx>
---
exec/main.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/exec/main.c b/exec/main.c
index 6c8ea35..45abc9f 100644
--- a/exec/main.c
+++ b/exec/main.c
@@ -272,11 +272,11 @@ static void corosync_sync_completed (void)
{
log_printf (LOGSYS_LEVEL_NOTICE,
"Completed service synchronization, ready to provide service.\n");
- sync_in_process = 0;
/*
* Inform totem to start using new message queue again
*/
totempg_trans_ack();
+ sync_in_process = 0;
}
static int corosync_sync_callbacks_retrieve (int sync_id,
--
1.7.1
_______________________________________________
discuss mailing list
discuss@xxxxxxxxxxxx
http://lists.corosync.org/mailman/listinfo/discuss
_______________________________________________ discuss mailing list discuss@xxxxxxxxxxxx http://lists.corosync.org/mailman/listinfo/discuss
