From: | Andres Freund <andres(at)2ndquadrant(dot)com> |
---|---|
To: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
Cc: | Bruce Momjian <bruce(at)momjian(dot)us>, Robert Haas <robertmhaas(at)gmail(dot)com>, Noah Misch <noah(at)leadboat(dot)com>, Jeff Davis <pgsql(at)j-davis(dot)com>, "pgsql-hackers(at)postgresql(dot)org" <pgsql-hackers(at)postgresql(dot)org> |
Subject: | Re: mvcc catalo gsnapshots and TopTransactionContext |
Date: | 2014-02-02 22:50:39 |
Message-ID: | 20140202225039.GR5930@awork2.anarazel.de |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
On 2014-02-02 15:16:45 -0500, Tom Lane wrote:
> Andres Freund <andres(at)2ndquadrant(dot)com> writes:
> > On February 2, 2014 5:52:22 PM CET, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> >> More to the point, changing the Assert so it doesn't fire
> >> doesn't do one damn thing to ameliorate the fact that cache reload
> >> during transaction abort is wrong and unsafe.
>
> > And, as upthread, I still don't think that's correct. I don't have
> > sources available right now, but IIRC we already have aborted out of the
> > transaction. Released locks, the xid and everything.
>
> Nope ... the case exhibited in the example is dying in AtEOSubXact_Inval,
> which is in the very midst of subxact abort.
True. But we've done LWLockReleaseAll(), TransactionIdAbortTree(),
XidCacheRemoveRunningXids() and
ResourceOwnerRelease(RESOURCE_RELEASE_BEFORE_LOCKS), which is why we are
currently able to build correct entries, even though we are in an
aborted transaction.
> I've been thinking about this for the past little while, and I believe
> that it's probably okay to have RelationClearRelation leave the relcache
> entry un-rebuilt, but with rd_isvalid = false so it will be rebuilt when
> next opened. The rationale is explained in the comments in the attached
> patch. I've checked that this fixes Noah's test case and still passes
> the existing regression tests.
Hm, a bit scary, but I don't see an immediate problem.
The following comment now is violated for nailed relations
* We assume that at the time we are called, we have at least AccessShareLock
* on the target index. (Note: in the calls from RelationClearRelation,
* this is legitimate because we know the rel has positive refcount.)
but that should be easy to fix.
I wonder though, if we couldn't just stop doing the
RelationReloadIndexInfo() for nailed indexes. The corresponding comment
says:
* If it's a nailed index, then we need to re-read the pg_class row to see
* if its relfilenode changed. We do that immediately if we're inside a
* valid transaction. Otherwise just mark the entry as possibly invalid,
* and it'll be fixed when next opened.
*/
but any relfilenode change should have already been handled by
RelationInitPhysicalAddr()?
Do you plan to backpatch this? If so, even to 8.4?
Greetings,
Andres Freund
--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services
From | Date | Subject | |
---|---|---|---|
Next Message | David Fetter | 2014-02-02 22:52:42 | Re: GSOC13 proposal - extend RETURNING syntax |
Previous Message | Peter Eisentraut | 2014-02-02 21:57:28 | Re: pg_basebackup and pg_stat_tmp directory |