Mike Mascari <mascarim@yahoo.com> writes:
> Thanks for the response, Tom. When looking at the system log,
> the kernel was logging messages regarding IPX network name collisions
> which apprently can happen when there are autoconfigured Win95 boxes
> on the same subnet. These messages were flooding the log at a rate of
> one every second or two...Even though #2 seems improbable, and just
> glancing at the IPX kernel code didn't point to how that may have
> caused a continual consumption of file descriptors, I'm willing to
> blame the kernel on this (and me for using autoprimary and autointerface
> options).
That doesn't strike me as a bulletproof explanation. fd.c has a tight
loop that close()s an FD and then tries to open() the file it wants,
repeat until success or an error other than ENFILE/EMFILE. If the
scenario really is that it got ENFILE every time until it was down to
zero FDs, there'd have to be something sucking up each freed FD within
microseconds of its being freed. Repeatedly. Forty or fifty (or more)
times in a row. I don't think a once-a-second Win95 lossage will do
that. And if you were down to zero free FDs system-wide, Postgres
wouldn't be the only thing having troubles!
I take it you don't use Postgres password authentication at all? If you
do, the other theory looks a lot more viable to me... I haven't had time
to try to reproduce a crash yet, but I'm pretty sure there's one there.
regards, tom lane