Don't reset latch in ConditionVariablePrepareToSleep().
authorThomas Munro <tmunro@postgresql.org>
Tue, 28 Jan 2020 02:28:36 +0000 (15:28 +1300)
committerThomas Munro <tmunro@postgresql.org>
Tue, 28 Jan 2020 02:28:36 +0000 (15:28 +1300)
commit78aaa0e82335527b8cc91601c98c9abbbc3ef3ee
treeb606f2b34a137a2110eaa06adc1a571ca51e58d6
parent05f18c6b6b6e4b44302ee20a042cedc664532aa2
Don't reset latch in ConditionVariablePrepareToSleep().

It's not OK to do that without calling CHECK_FOR_INTERRUPTS().
Let the next wait loop deal with it, following the usual pattern.

One consequence of this bug was that a SIGTERM delivered in a very
narrow timing window could leave a parallel worker process waiting
forever for a condition variable that will never be signaled, after
an error was raised in other process.

The code is a bit different in the stable branches due to commit
1321509f, making problems less likely there.  No back-patch for now,
but we may finish up deciding to make a similar change after more
discussion.

Author: Thomas Munro
Reviewed-by: Shawn Debnath
Reported-by: Tomas Vondra
Discussion: https://postgr.es/m/CA%2BhUKGJOm8zZHjVA8svoNT3tHY0XdqmaC_kHitmgXDQM49m1dA%40mail.gmail.com
src/backend/storage/lmgr/condition_variable.c