On Mon, Jul 28, 2014 at 3:59 PM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Fri, Jul 25, 2014 at 3:25 PM, Noah Misch <noah@leadboat.com> wrote:
>> On a Windows or other EXEC_BACKEND build, the following eventually gets
>> failures because all, or all but one, max_connections slot is consumed:
>>
>> for run in `seq 1 100`; do make -C contrib/test_shm_mq installcheck; done
>>
>> When I use max_connections=40, it fails on the sixth iteration. Only the six
>> basic processes are actually running at that time.
>
> The tests start 7 workers each time, so that makes sense: 7 * 5 < 40
> but 7 * 6 > 40. What I'm not sure is why they are leaking connection
> slots, and why they're only doing it in EXEC_BACKEND mode. A quick
> code audit didn't uncover any obvious explanation, so I'll try to
> reproduce and debug.
OK, I think I see the problem. In EXEC_BACKEND mode,
SubPostmasterMain() calls InitProcess() before IsBackgroundWorker has
been set. InitProcess() therefore pulls the PGPROC for the worker
from freeProcs rather than bgworkerFreeProcs. By exit time,
IsBackgroundWorker has been set, so the PGPROC gets put back on the
bgworkerFreeProcs list. Eventually there are no regular PGPROCs left;
they've all been moved to the bgworkerFreeProcs list.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company