[Drbd-dev] DRBD-8: Patch to fix incorrect split-brain occurrence.

[Drbd-dev] DRBD-8: Patch to fix incorrect split-brain occurrence.

[Graham, Simon]

[-- Attachment #1: Type: text/plain, Size: 1068 bytes --]

In testing, we have run into a case where we unexpectedly get a
split-brain situation - after some investigation I think I have found
the reason - at the end of syncing, the SyncTarget side finishes up by
setting it's UUID info to the same as the Source (in
drbd_resync_finished) - the code that does this checks for connection
state SyncTarget before doing this. 

Now, I have seen that sometimes the resync completes when in PausedSyncT
state - for example, if the last set of blocks is already under way when
we decide to go into the PausedSyncT state - thus it's entirely possible
to enter drbd_resync_finished with connection state PausedSyncT in which
case we do finish up the resync but do not copy the UUID info - this
means that the next time we need to resync, the code determines that we
have a split-brain situation.

In other places in the code, checks are made for either SyncTarget or
PausedSyncT so I propose fixing this (patch attached) the same way -
have drbd_resync_finished check for either SyncTarget OR PausedSyncT.

Simon


[-- Attachment #2: drbd-resync.patch --]
[-- Type: application/octet-stream, Size: 755 bytes --]

Index: src/drbd/drbd_worker.c
===================================================================
--- src/drbd/drbd_worker.c	(revision 7932)
+++ src/drbd/drbd_worker.c	(working copy)
@@ -402,7 +402,8 @@
 	if (mdev->rs_failed) {
 		INFO("            %lu failed blocks\n",mdev->rs_failed);
 
-		if (mdev->state.conn == SyncTarget) {
+		if (mdev->state.conn == SyncTarget || 
+		    mdev->state.conn == PausedSyncT) {
 			dstate = Inconsistent;
 			pdstate = UpToDate;
 		} else {
@@ -412,7 +413,8 @@
 	} else {
 		dstate = pdstate = UpToDate;
 
-		if ( mdev->state.conn == SyncTarget) {
+		if ( mdev->state.conn == SyncTarget || 
+		     mdev->state.conn == PausedSyncT) {
 			if( mdev->p_uuid ) {
 				int i;
 				for ( i=Bitmap ; i<=History_end ; i++ ) {

Re: [Drbd-dev] DRBD-8: Patch to fix incorrect split-brain occurrence.

[Lars Ellenberg]

/ 2006-12-19 10:45:04 -0500
\ Graham, Simon:
> In testing, we have run into a case where we unexpectedly get a
> split-brain situation - after some investigation I think I have found
> the reason - at the end of syncing, the SyncTarget side finishes up by
> setting it's UUID info to the same as the Source (in
> drbd_resync_finished) - the code that does this checks for connection
> state SyncTarget before doing this. 
> 
> Now, I have seen that sometimes the resync completes when in PausedSyncT
> state - for example, if the last set of blocks is already under way when
> we decide to go into the PausedSyncT state 

right, or "resync" "just happens" by application writes.

> - thus it's entirely possible
> to enter drbd_resync_finished with connection state PausedSyncT in which
> case we do finish up the resync but do not copy the UUID info - this
> means that the next time we need to resync, the code determines that we
> have a split-brain situation.
> 
> In other places in the code, checks are made for either SyncTarget or
> PausedSyncT so I propose fixing this (patch attached) the same way -
> have drbd_resync_finished check for either SyncTarget OR PausedSyncT.

right.

-- 
: Lars Ellenberg                            Tel +43-1-8178292-55 :
: LINBIT Information Technologies GmbH      Fax +43-1-8178292-82 :
: Vivenotgasse 48, A-1120 Vienna/Europe    http://www.linbit.com :