On 2012-11-30 18:09:49 +0530, Pavan Deolasee wrote:
> On Fri, Nov 30, 2012 at 5:28 PM, Andres Freund <andres@2ndquadrant.com>wrote:
>
> >
> > > >
> > > That seems to be safe to me. Anything thats been read above can't really
> > > change. The tuple is already locked, so a concurrent update/delete has to
> > > wait on us. We have a pin on the buffer, so VACUUM or HOT-prune can't
> > > happen either. I can't see any other operation that can really change
> > those
> > > fields.
> >
> > We only get the pin right there, I don't see any preexisting pin. Which
> > means we might have just opened a page thats in the process of being
> > pruned/vacuumed by another backend.
> >
>
> Hmm. Yeah, you're right. That is a possible risky scenario. Even though
> cleanup lock waits for all pins to go away, it will work only if every
> reader takes at least a SHARE lock unless it was continuously holding a pin
> on a buffer (in which case its OK to drop lock and read a tuple body
> without reacquiring it again). Otherwise, as you rightly pointed out, we
> could actually be reading a page which being actively cleaned up and tuples
> are being moved around.
Well, live (or recently dead) tuples can't be just moved arround unless
I miss something. That would cause problems with with HeapTuples
directly pointing into the page as youve noticed.
> > I think a concurrent heap_(insert|update)/PageAddItem might actually be
> > already dangerous because it might move line pointers around
> >
> >
> I don't we move the line pointers around ever because the indexes will be
> pointing to them.
The indexes point to a tid, not to a line pointer. So reshuffling the
linepointer array - while keeping the tids valid - is entirely
fine. Right?
Also, PageAddItem does that all the time, so I think we would have
noticed if not ;)
Greetings,
Andres Freund
--Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services
