On Thu, Oct 12, 2023 at 7:09 PM Michael Paquier <michael@paquier.xyz> wrote:
> On Thu, Oct 12, 2023 at 01:32:29PM -0400, Robert Haas wrote:
> > IPC seems right to me. Yeah, a timeout is being used, but as you say,
> > that's an implementation detail.
> >
> > +1 for the idea, too.
>
> Agreed that timeout makes little sense in this context, and IPC looks
> correct.
>
> + pgstat_report_wait_start(WAIT_EVENT_CHECKPOINT_DELAY_START);
> do
> {
> pg_usleep(10000L); /* wait for 10 msec */
> } while (HaveVirtualXIDsDelayingChkpt(vxids, nvxids,
> DELAY_CHKPT_START));
> + pgstat_report_wait_end();
>
> HaveVirtualXIDsDelayingChkpt() does immediately a LWLockAcquire()
> which would itself report a wait event for ProcArrayLock, overwriting
> this new one, no?
Ah, right: the wait event should be set and cleared around pg_usleep,
not the whole loop.
--
Robert Haas
EDB: http://www.enterprisedb.com