Re: 001_rep_changes.pl stalls

At Sat, 18 Apr 2020 00:01:42 -0700, Noah Misch <noah(at)leadboat(dot)com> wrote in
> On Fri, Apr 17, 2020 at 05:06:29PM +0900, Kyotaro Horiguchi wrote:
> > At Fri, 17 Apr 2020 17:00:15 +0900 (JST), Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote in
> > > By the way, if latch is consumed in WalSndLoop, succeeding call to
> > > WalSndWaitForWal cannot be woke-up by the latch-set. Doesn't that
> > > cause missing wakeups? (in other words, overlooking of wakeup latch).
> >
> > - Since the only source other than timeout of walsender wakeup is latch,
> > - we should avoid wasteful consuming of latch. (It is the same issue
> > - with [1]).
> >
> > + Since walsender is wokeup by LSN advancement via latch, we should
> > + avoid wasteful consuming of latch. (It is the same issue with [1]).
> >
> >
> > > If wakeup signal is not remembered on walsender (like
> > > InterruptPending), WalSndPhysical cannot enter a sleep with
> > > confidence.
>
> No; per latch.h, "What must be avoided is placing any checks for asynchronous
> events after WaitLatch and before ResetLatch, as that creates a race
> condition." In other words, the thing to avoid is calling ResetLatch()
> without next examining all pending work that a latch would signal. Each
> walsender.c WaitLatch call does follow the rules.

I didn't meant that, of course. I thought of more or less the same
with moving the trigger from latch to signal then the handler sets a
flag and SetLatch(). If we use bare latch, we should avoid false
entering to sleep, which also makes thinks compolex.

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center