On Thu, Sep 22, 2005 at 10:36:41AM -0400, Alvaro Herrera wrote:
> On Fri, Sep 16, 2005 at 11:50:21PM -0700, Simon Riggs wrote:
> > > Alvaro Herrera wrote
> > > The only caller of both is
> > > repair_frag, whose only caller in turn is full_vacuum_rel.
> >
> > ...bgwriter still needs to access blocks. The WAL system relies on the
> > locking behaviour for recoverability, see comments in LockBuffer() and
> > SyncOneBuffer().
>
> Oh, certainly! In this case, may I point out that scan_heap() does not
> bother locking pages, mentioning that "we assume that holding exclusive
> lock on the relation will keep other backends from looking at the page".
> In particular, it calls PageRepairFragmentation which runs with the page
> unlocked AFAICT.
Looking again, PageRepairFragmentation is called on a copy of the page,
not on the page itself, so this is not a problem. The page is only
modified to exchange old Xids for FrozenTransactionId, or to set some
hint bits, so this really shouldn't be too much of a problem. I still
think it would be better to lock the page beforehand.
--
Alvaro Herrera Architect, http://www.EnterpriseDB.com
"Uno puede defenderse de los ataques; contra los elogios se esta indefenso"
