Horton
Menu

Re: Implement waiting for wal lsn replay: reloaded

Andres Freund <andres@anarazel.de>

View thread
From: Andres Freund <andres@anarazel.de>
To: Xuneng Zhou <xunengzhou@gmail.com>
Cc: Tom Lane <tgl@sss.pgh.pa.us>, Alexander Korotkov <aekorotkov@gmail.com>, Heikki Linnakangas <hlinnaka@iki.fi>, Peter Eisentraut <peter@eisentraut.org>, Thomas Munro <thomas.munro@gmail.com>, Álvaro Herrera <alvherre@kurilemu.de>, Chao Li <li.evan.chao@gmail.com>, pgsql-hackers <pgsql-hackers@lists.postgresql.org>, Michael Paquier <michael@paquier.xyz>, jian he <jian.universality@gmail.com>, Tomas Vondra <tomas@vondra.me>, Yura Sokolov <y.sokolov@postgrespro.ru>
Date: 2026-04-07T03:07:45Z
Lists: pgsql-hackers 
Hi,

On 2026-04-07 10:28:52 +0800, Xuneng Zhou wrote:
> On Tue, Apr 7, 2026 at 10:08 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >
> > I wrote:
> > > I wondered why my buildfarm animals got noticeably slower today.
> > > There seem to be a couple of culprits, but one of them is that
> > > 7e8aeb9e4 (Use WAIT FOR LSN) has caused the runtime of pg_rewind's
> > > t/003_extrafiles.pl to go through the roof.  On indri's host, that
> > > TAP test took about 3 seconds immediately before that commit, and
> > > about 45 seconds immediately after.
> >
> > I'm wrong: there's only one culprit.  The other big change in runtime
> > today is that src/test/recovery's t/033_replay_tsp_drops.pl went from
> > about 4 seconds to about 46, and that jump also happened at 7e8aeb9e4.
> > So we still have a mystery, but it's "what do those two tests have in
> > common that is shared by no others?".
> >
> Thanks for reporting this. I think it could be related to the read of
> not-yet-updated writtenUpto position.  I'll look into this and propose
> a fix shortly.

Yes, it's not yet initialized and thus the WAIT FOR waits, even though the
position had already been reached.


But, leaving that aside, looking at this code I'm somewhat concerned - it
seems to not worry at all about memory ordering?


static void
XLogWalRcvWrite(char *buf, Size nbytes, XLogRecPtr recptr, TimeLineID tli)
...
	/* Update shared-memory status */
	pg_atomic_write_u64(&WalRcv->writtenUpto, LogstreamResult.Write);

	/*
	 * If we wrote an LSN that someone was waiting for, notify the waiters.
	 */
	if (waitLSNState &&
		(LogstreamResult.Write >=
		 pg_atomic_read_u64(&waitLSNState->minWaitedLSN[WAIT_LSN_TYPE_STANDBY_WRITE])))
		WaitLSNWakeup(WAIT_LSN_TYPE_STANDBY_WRITE, LogstreamResult.Write);

There are no memory barriers here, so the CPU would be entirely free to not
make the writtenUpto write visible to a waiter that's in the process of
registering and is checking whether it needs to wait in WaitForLSN().

And WaitForLSN()->GetCurrentLSNForWaitType()->GetWalRcvWriteRecPtr() also has
no barriers.  That MAYBE is ok, due addLSNWaiter() providing the barrier at
loop entry and maybe kinda you can think that WaitLatch() will somehow also
have barrier semantic.  But if so, that would need to be very carefully
documented.  And it seems completely unnecessary here, it's hard to believe
using a barrier (via pg_atomic_read_membarrier_u64() or such) would be a
performance issue

Greetings,

Andres Freund

Commits

Same data as JSON: GET /api/v1/messages/:b64id/commits the thread's linked commits as JSON, with link sources. API reference →

  1. Clean up 019_replslot_limit.pl comments

  2. Stabilize 019_replslot_limit.pl: wait on slot restart_lsn

  3. Fix memory ordering in WAIT FOR LSN wakeup mechanism

  4. Improve WAIT FOR LSN test coverage

  5. Remove redundant WAIT FOR LSN caller-side pre-checks

  6. Use barrier semantics when reading/writing writtenUpto

  7. Use replay position as floor for WAIT FOR LSN standby_(write|flush)

  8. Wake standby_write/standby_flush waiters from the WAL replay loop

  9. Minimal fix for WAIT FOR ... MODE 'standby_flush'

  10. Avoid syscache lookup while building a WAIT FOR tuple descriptor

  11. Document that WAIT FOR may be interrupted by recovery conflicts

  12. Use WAIT FOR LSN in PostgreSQL::Test::Cluster::wait_for_catchup()

  13. Wake LSN waiters before recovery target stop

  14. Remove redundant pg_unreachable() after elog(ERROR) from ExecWaitStmt()

  15. Revert "Use WAIT FOR LSN in PostgreSQL::Test::Cluster::wait_for_catchup()"

  16. Fix variable usage in wakeupWaiters()

  17. Add tab completion for the WAIT FOR LSN MODE option

  18. Add the MODE option to the WAIT FOR LSN command

  19. Extend xlogwait infrastructure with write and flush wait types

  20. Unify error messages

  21. Optimize shared memory usage for WaitLSNProcInfo

  22. Fix WaitLSNWakeup() fast-path check for InvalidXLogRecPtr

  23. Fix incorrect function name in comments

  24. Add infrastructure for efficient LSN waiting

  25. Add pairingheap_initialize() for shared memory usage

  26. Implement WAIT FOR command