I wrote:
> Hmm. I think the problem is that poll_start() thinks it can just call
> start() a second time after a failure. If it wasn't a true failure
> but a timeout, then _pid is now set and the second call complains.
Oh, wait --- the case that is failing is after 017_shm.pl has
intentionally kill -9'd a postmaster, so that its pidfile is
left behind. The next attempted start fails on shmem id
conflict, but it doesn't remove the old pidfile, and then
the code I added to sub start erroneously picks that up
as a live postmaster PID.
Seems like we need to do 'kill 0' on the PID we get from
the file to verify that there's really a postmaster there.
(I wonder how well that works on Windows? perlport claims
it does, but ...)
I fear I still don't have the whole story though because
per this theory it should fail everywhere, yet it doesn't.
regards, tom lane