Andres Freund <andres@anarazel.de> writes:
> On 2017-03-16 13:00:54 +0100, Petr Jelinek wrote:
>> Looks like that didn't help either.
>>
>> I setup my own Windows machine and can reproduce the issue. I played
>> around a bit and could not really find a fix other than adding
>> WL_TIMEOUT and short timeout to WaitLatchOrSocket (it does wait a very
>> long time on the WaitLatchOrSocket otherwise before failing).
> Hm. Could you use process explorer or such to see the exact events
> happening? Seing that might help us to nail this down.
At this point it's hard to escape the conclusion that we're looking at
some sort of bug in the Windows version of the WaitEventSet infrastructure
--- or at least, the WaitEventSet-based waits for socket events are
evidently behaving differently from the implementation on the libpq side,
which is based on pqWaitTimed and thence pqSocketCheck and thence
pqSocketPoll.
Noticing that bowerbird builds with openssl, I'm a bit suspicious that
the issue might have something to do with the fact that pqSocketCheck
has a special case for SSL, which I don't think WaitEventSet knows about.
But I don't think SSL is actually active in the buildfarm run, so it's
hard to see how we get to that scenario exactly --- or even less why it
would only matter on Windows.
regards, tom lane