Re: pgsql: Fix a couple of bugs in MultiXactId freezing

Hi,

On 2013-12-12 18:39:43 -0300, Alvaro Herrera wrote:
> Alvaro Herrera wrote:
> > One last thing (I hope). It's not real easy to disable this check,
> > because it actually lives in GetNewMultiXactId. It would uglify the API
> > a lot if we were to pass a flag down two layers of routines; and moving
> > it to higher-level routines doesn't seem all that appropriate
> > either.

Unfortunately I find that too ugly. Adding a flag in the procarray
because of an Assert() in a lowlevel routine? That's overkill.
What's the problem with moving the check to MultiXactIdCreate() and
MultiXactIdExpand() instead? Since those are the ones where it's
required to have called SetOldest() before, I don't see why that would
be inappropriate?

> > I'm thinking we can have a new flag in MyPgXact->vacuumFlags, so
> > heap_prepare_freeze_tuple does this:
> >
> > PG_TRY();
> > {
> > /* set flag to let multixact.c know what we're doing */
> > MyPgXact->vacuumFlags |= PROC_FREEZING_MULTI;
> > newxmax = FreezeMultiXactId(xid, tuple->t_infomask,
> > cutoff_xid, cutoff_multi, &flags);
> > }
>
> Uhm, actually we don't need a PG_TRY block at all for this to work: we
> can rely on the flag being reset at transaction abort, if anything wrong
> happens and we lose control. So just set the flag, call
> FreezeMultiXactId, reset flag.

I don't think that'd be correct for a CLUSTER in a subtransaction? A
subtransaction's abort afaics doesn't call ProcArrayEndTransaction() and
thus don't clear vacuumFlags...

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services