Re: Optimising latch signals

On Tue, Mar 9, 2021 at 12:20 PM Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
> On 2021-Mar-03, Thomas Munro wrote:
> > On Mon, Mar 1, 2021 at 2:29 PM Thomas Munro <thomas(dot)munro(at)gmail(dot)com> wrote:
> > > Time to watch the buildfarm to find out if my speculation about
> > > illumos is correct...
> >
> > I just heard that damselfly's host has been decommissioned with no
> > immediate plan for a replacement. That was the last of the
> > Solaris-family animals testing master. It may be some time before I
> > find out if my assumptions broke something on that OS...
>
> Hi, I don't know if you realized but we have two new Illumos members
> now (haddock and hake), and they're both failing initdb on signalfd()
> problems.

Ah, cool. I'd been discussing this with their owner, who saw my
message and wanted to provide replacements. Nice to see these start
up even though I don't love the colour of their first results. In
off-list emails, we got as far as determining that signalfd() fails on
illumos when running inside a zone (= container), because
/dev/signalfd is somehow not present. Apparently it works when
running on the main host. Tracing revealed that it's trying to open
that device and getting ENOENT here:

running bootstrap script ... FATAL: XX000: signalfd() failed
LOCATION: InitializeLatchSupport, latch.c:279

I'll wait a short time while he tries to see if that can be fixed (I
have no clue if it's a configuration problem in some kind of zone
creation scripts, or a bug, or what). If not, my fallback plan will
be to change it to default to WAIT_USE_POLL on that OS until it can be
fixed.