On Fri, 2008-11-21 at 15:21 +0530, Pavan Deolasee wrote:
>
> + /*
> + * Release locks, if any.
> + */
> + RelationReleaseRecoveryLocks(xlrec->slotId);
>
>
> If I am reading the patch correctly, AccessExclusiveLocks acquired by
> a transaction will be released when the transaction is committed or
> aborted. If the transaction errors out with FATAL, the locks will be
> released when the next transaction occupying the same slot is
> committed/aborted.
>
> I smell some sort of deadlock condition here. What if the following
> events happen:
>
> - transaction A (slot 1) starts and acquires AEL lock on relation
> - transaction A errors out with FATAL error
> - transaction B (slot 1) starts and requests AEL lock on the same
> relation
>
> Won't B deadlock with A ? Since B hasn't yet committed/aborted, the
> lock held by A is not released and
> relation_redo_lock() would indefinitely wait for the lock.
This won't happen because the lock is held by the startup process on
behalf of slot 1. Explained in comments in inval.c code.
-- Simon Riggs www.2ndQuadrant.comPostgreSQL Training, Services and Support