SEO

I'm trying to track down the source of a memory buffer leak in my
OpenBSD 3.8 box.

It started around the time I switched from four seperate NICs to a
Soekris lan1651, an affordable 4-port NIC. For a while I thought it
was my pf config, but lots of people use pf and nobody else seems to
be having the problem.

The behavior is that it works for a while, but eventually (sometimes
18 minutes, sometimes 4 days) networking comes to a grinding halt, and
I must reboot the machine from the console.

I have tried increasing mbufs in the kernel config, and this seems to
help a bit, maybe.

Killing processes does not help. It is not a leak of userland memory;
it is a leak of some kind of kernel pool. If I try to ping something
from the console, it says "no memory buffers available".

I cannot tell if this is a OpenBSD driver problem, a pf problem, or
a lan1641 hardware problem, or something else.

"vmstat -m" when it is hung shows nothing unusual; the types of memory
allocated indicate that most of the memory, maybe 2MB, is used by
"devbuf"s.

I was wondering if there were any other reports. One hypothesis is
that if two of the NICs interrupt nearly simultaneously, that there's
some kind of race condition that leads to a kernel memory leak.

This is very hard to debug and any advice would be appreciated,
because going net-dead is not good for a firewall, and very annoying.

I can provide any debugging information that you might find helpful.

I have tried running tcpdump on my interfaces, so that I can see the last
few packets being passed before it hangs, and I haven't found anything
unusual in the traffic itself.
--
"Curiousity killed the cat, but for a while I was a suspect" -- Steven Wright
Security Guru for Hire http://www.lightconsulting.com/~travis/ -><-
GPG fingerprint: 9D3F 395A DAC5 5CCC 9066 151D 0A6B 4098 0C55 1484