Andres Freund wrote:
> While validating my patch for
http://archives.postgresql.org/message-id/20160714060607.klwgq2qr7egt3zrr%40alap3.anarazel.de
> I noticed that the standby's infomask still had HEAP_XMAX_INVALID set
> after replaying a XLOG_HEAP_LOCK record.
>
> That's bad, but not really commonly fatal, because unless prepared
> transactions are used, locks don't need to be present / valid after
> crash-recovery. But it's clearly something we need to fix.
>
> Given that it's been this way for ages, it's not a blocker for
> committing the fix for above URL, but I'll try to get in something today
> for that as well.
Yes, it looks broken, and the consequences for prepared xacts should be
pretty obvious. Other than those, I think it's pretty innocuous.
> Looks like the minimal fix is just to add
> htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
> htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
> to heap_xlog_lock and heap_xlog_lock_updated.
Maybe we should change fix_infomask_from_infobits() to reset
HEAP_XMAX_BITS | HEAP_MOVED too (and HEAP_KEYS_UPDATED I suppose), to
avoid doing it in each callsite.
--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services