Re: HeapTupleHeaderAdvanceLatestRemovedXid doing the wrong thing with multixacts

Excerpts from Bruce Momjian's message of jue ago 16 11:24:55 -0400 2012:
> On Mon, Oct 17, 2011 at 08:38:18PM +0100, Simon Riggs wrote:
> > On Mon, Oct 17, 2011 at 8:03 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> > > Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> writes:
> > >> I just noticed that HeapTupleHeaderAdvanceLatestRemovedXid is comparing Xmax as a TransactionId without verifying whether it is a multixact or not.  Since they advance separately, this could lead to bogus answers.  This probably needs to be fixed.  I didn't look into past releases to see if there's a live released bug here or not.
> > >
> > >> I think the fix is simply to ignore the Xmax if the HEAP_XMAX_IS_MULTI bit is set.
> > >
> > >> Additionally I think it should check HEAP_XMAX_INVALID before reading the Xmax at all.
> > >
> > > If it's failing to even check XMAX_INVALID, surely it's completely
> > > broken?  Perhaps it assumes its caller has checked all this?
> >
> > HeapTupleHeaderAdvanceLatestRemovedXid() is only ever called when
> > HeapTupleSatisfiesVacuum() returns HEAPTUPLE_DEAD, which only happens
> > when HEAP_XMAX_IS_MULTI is not set.
> >
> > I'll add an assert to check this and a comment to explain.
>
> Was this completed?

As far as I recall, there are changes related to this in my fklocks
patch. I am hoping to have some review happen on it during the upcoming
commitfest (which presumably means I need to do a merge to newer
sources.)

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