Heikki Linnakangas <heikki.linnakangas@enterprisedb.com> writes:
> A safer approach would be to just PGSemaphoreUnlock() in the signal
> handler, and do all the other processing outside it.
I don't see any particularly good reason to assume that
PGSemaphoreUnlock is safe either: you're still talking about nested
semop operations.
The pre-existing SIGALRM handler uses a self-signal (kill(MyProcPid,
SIGINT)) to kick the process off any wait it might be doing. I'd rather
do something like that.
Or maybe the work you're doing on latches would help ...
regards, tom lane