Re: [bug fix] "pg_ctl stop" times out when it should respond quickly

Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> On Mon, Feb 17, 2014 at 11:29 AM, Alvaro Herrera
> <alvherre(at)2ndquadrant(dot)com> wrote:
>> The pg_regress part is ugly. However, pg_regress is doing something
>> unusual when starting postmaster itself, so the ugly coding to stop it
>> seems to match. If we wanted to avoid the ugliness here, the right fix
>> would be to use pg_ctl to start postmaster as well as to stop it.

> I wonder if this would change the behavior in cases where we hit ^C
> during the regression tests. Right now I think that kills the
> postmaster as well as pg_regress, but if we used pg_ctl, it might not,
> because pg_regress uses fork()+exec(), but pg_ctl uses system() to
> launch a shell which is in turn instructed to background the
> postmaster.

After re-reading this thread, I am convinced that this patch is failing
to fix the real problem, which is that the postmaster isn't defending
against early arrival of a signal; see
http://www.postgresql.org/message-id/30805.1386110129@sss.pgh.pa.us

Hacking pg_ctl is a band-aid solution, and given these concerns,
not a particularly safe band-aid.

I experimented with simply moving postmaster.c's initialization of signals
further up, and that seems to fix the complained-of problem just fine.
It could be an issue if anything in postmaster initialization is expecting
to run with signals unblocked, but I don't know what that would be.
So I suggest we just do the attached.

regards, tom lane