On Sat, Jun 20, 2015 at 12:07 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Michael Paquier <michael.paquier@gmail.com> writes:
>> Now if we look at RewindTest.pm, there is the following code:
>> if ($test_master_datadir)
>> {
>> system
>> "pg_ctl -D $test_master_datadir -s -m immediate stop
>> 2> /dev/null";
>> }
>> if ($test_standby_datadir)
>> {
>> system
>> "pg_ctl -D $test_standby_datadir -s -m immediate
>> stop 2> /dev/null";
>> }
>> And I think that the problem is triggered because we are missing a -w
>> switch here, meaning that we do not wait until the confirmation that
>> the server has stopped, and visibly if stop is slow enough the next
>> server to use cannot start because the port is already taken by the
>> server currently stopping.
>
> After I woke up a bit more, I remembered that -w is already the default
> for "pg_ctl stop", so your diagnosis here is incorrect.
Ah right. I forgot that. Perhaps I got just lucky in my runs.
> I suspect that the real problem is the arbitrary decision to use -m
> immediate. The postmaster would ordinarily wait for its children to
> die, but on a slow machine we could perhaps reach the end of that
> 5-second timeout, whereupon the postmaster would SIGKILL its children
> *and exit immediately*. I'm not sure how instantaneous SIGKILL is,
> but it seems possible that we could end up trying to start the new
> postmaster before all the children of the old one are dead. If the
> shmem interlock is working properly that ought to fail.
>
> I wonder whether it's such a good idea for the postmaster to give
> up waiting before all children are gone (postmaster.c:1722 in HEAD).
I don't think so as well.
--
Michael
