[Drbd-dev] DRBD8: nodes deadlock in PausedSync{ST]

[Philipp Reisner]

On Monday 29 October 2007 23:33:29 Montrose, Ernest wrote:
> Hi all,
> I have been struggling with a problem here where the nodes enter
> PausedSync[T|S] and stay there.
> This happens when one node come up from a fresh attach, connect
> sequence.  I think the issue happens this way. Say we have two volumes
> drbd5 and drbd16 and we attempt to connect both of them at roughly the
> same time.  Futhermore, drbd5 and 16 will require syncing say as sync
> target. What I observe is this:
> *	drbd16 is connecting and drbd5 is syncing. So 16 is paused isp=1
> *	drbd16 enters receive_state() but before acquiring the req_lock
> that thread loses the CPU to drbd5 that is finishing syncing.  After_isp
> is cleared on 16 giving drbd16 the green light to continue syncing. So
> far so good.
> *	Now drbd16 resumes with the old peer_isp=1
> *	So now we are paused forever.
>
> So I think receive_state() is just racy but I could be wrong. I am
> really not sure how to fix this but I include a patch here that may help
> to at least illustrate the problem. It seems close window for this
> particular race somewhat.
>

Hi Ernest,

You patch fixes the issue.

I spent an hour or so to understanding the exact timing, and drew 
diagrams of it... it is fixed with that change. No race left, I think.

I committed it to my git tree.

-Phil
-- 
: Dipl-Ing Philipp Reisner                      Tel +43-1-8178292-50 :
: LINBIT Information Technologies GmbH          Fax +43-1-8178292-82 :
: Vivenotgasse 48, 1120 Vienna, Austria        http://www.linbit.com :