"Tom Lane" <tgl@sss.pgh.pa.us> writes:
> Gregory Stark <stark@enterprisedb.com> writes:
>> "Tom Lane" <tgl@sss.pgh.pa.us> writes:
>>> How you figure that?
>
>> Well I'm not clear exactly what's going on with the semaphores here. If it's
>> possible for to be printing the messages only as a result of another backend
>> unlocking the semaphore then making the PGSemaphoreUnlock conditional on
>> log_lock_waits means you can't enable log_lock_waits after startup and get
>> deterministic behaviour because whether you get messages will depend on which
>> other backend happens to wake you up.
>
> I don't see how you arrive at that conclusion. The message is printed
> by the backend that is waiting for (or just obtained) a lock, dependent
> on its own local setting of log_lock_waits, and not dependent on who
> woke it up.
But in your version of the patch you're not calling PGSemaphoreUnlock() unless
log_lock_waits is set in the process doing the waking.
Hm, I suppose it'll wake up itself when its own deadlock timer runs out
anyways. So I guess the worst case is that it doesn't say anything after a
soft deadlock fixup.
> BTW, I just noticed that GUC allows deadlock_timeout to be set all the
> way down to zero. This seems bad --- surely the minimum value should at
> least be positive? As CVS HEAD stands, you're likely to get a lot of
> spurious/useless log messages if you have log_lock_waits = true and
> deadlock_timeout = 0. Do we care?
Does that actually work? I would expect setitimer to turn off the alarm in
that case.
--
Gregory Stark
EnterpriseDB http://www.enterprisedb.com