Hi,
a significant number of WaitLatch's in the backend currently don't check
for postmaster death. That's imo wrong. E.g. SELECT pg_sleep(100); just
continues to run.
I think we should change most sites to error out in that case. I wonder
if we shouldn't add another WL_ flag that automatically makes the latch
code do so; instead of repeating the code at every callsite.
Places that I've noticed in a quick skim:
* pg_sleep()
* gather_getnext()
* shm_mq_send_bytes()?
* shm_mq_receive_bytes()?
* ProcSleep()?
* ProcWaitForSignal()
The only case where we don't necessarily want to react to postmaster
death is syslogger, which is supposed to finish logging before shutting
down.
Additionally I noticed that we're not always diligent about following
the correct pattern when using latches. For example check
gather_readnext(): /* Nothing to do except wait for developments. */ WaitLatch(MyLatch, WL_LATCH_SET, 0);
CHECK_FOR_INTERRUPTS(); ResetLatch(MyLatch);
we should reset the latch before checking for interrupts, not
after. This way an interrupt that arrives between the two will possibly
be ignored.
Greetings,
Andres Freund