Re: FATAL: lock AccessShareLock on object 0/1260/0 is already held

On Wed, Sep 7, 2011 at 6:25 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes:
>> I thought about an error exit from client authentication, and that's a
>> somewhat appealing explanation, but I can't quite see why we wouldn't
>> clean up there the same as anywhere else.  The whole mechanism feels a
>> bit rickety to me - we don't actually release locks; we just abort the
>> transaction and *assume* that will cause locks to get released.
>
> Well, transaction abort will call LockReleaseAll, which is carefully
> coded to clean up the proclock lists regardless of what is in the
> locallocks table, so I'm not sure why you find that any more rickety
> than anything else.

Well, it's very clear that you CAN orphan locks if a backend holding a
session lock ever does CHECK_FOR_INTERRUPTS() outside of a
transaction. Try the attached patch.

rhaas=# vacuum full foo;
FATAL: terminating connection due to administrator command
FATAL: terminating connection due to administrator command
The connection to the server was lost. Attempting reset: Succeeded.
rhaas=# vacuum full foo;
ERROR: lock AccessExclusiveLock on object 16384/1431013/0 is already held

Now, I don't see any evidence of a live bug here (and on further
thought it can't be Dave's bug because he is orphaning
AccessShareLocks, not AccessExclusiveLocks), but I find this pretty
convincing as a demonstration of ricketiness. It is certainly not
obvious on its face that a misplaced CHECK_FOR_INTERRUPTS() can result
in a backend exiting without cleaning up its locks, and I'd argue its
a bad idea to leave it that way even if there's no user-visible
problem there today.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company