On Sat, Mar 8, 2025 at 6:31 AM Jacob Champion
<jacob.champion@enterprisedb.com> wrote:
> On Thu, Mar 6, 2025 at 9:13 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> > I don't see that behaviour on my Mac with a simple program, and that
> > seems like it couldn't possibly be intended.
>
> What version of macOS?
>
> Just to make sure I'm not chasing ghosts, I've attached my test
> program. Here are my CI results for running it:
Ah, right, yeah I see that here too. I thought you were saying that
kevent() could report an already triggered alarm even though we'd
replaced it (it doesn') but of course you meant poll(kq) as libpq
does.
I believe I know exactly why: kqueues are considered readable (by
poll/select/other kqueues) if there are any events queued[1]. Apple's
EVFILT_TIMER implementation is doing that trick[2] where it leaves
them queued, but filt_timerprocess() filters them out if its own
private _FIRED flag isn't set, so kevent() itself won't wake up or
return them. That trick doesn't survive nesting. I think I would
call that a bug. (I think I would keep the atomic CAS piece -- it
means you don't have to drain the timer callout synchronously when
reprogramming it which is a cool trick, but I think they overshot when
they left the knote queued.)
Maybe just do the delete-and-add in one call?
EV_SET(&ev[0], 1, EVFILT_TIMER, EV_DELETE, 0, 0, 0);
EV_SET(&ev[1], 1, EVFILT_TIMER, EV_ADD | EV_ONESHOT, 0, timeout, 0);
if (kevent(kq, &ev[0], 2, NULL, 0, NULL) < 0)
[1]
https://github.com/apple-oss-distributions/xnu/blob/8d741a5de7ff4191bf97d57b9f54c2f6d4a15585/bsd/kern/kern_event.c#L1078
[2]
https://github.com/apple-oss-distributions/xnu/blob/8d741a5de7ff4191bf97d57b9f54c2f6d4a15585/bsd/kern/kern_event.c#L1661
