* Stephan A. Rickauer  [2008-07-14 17:27]:

quoted text
> On Mon, 2008-07-14 at 14:22 +0200, Henning Brauer wrote:

> > perfect analysis!

> >

> > looks like the only sane thing to do in that case is to bail and not

> > send the icmp.

>

> I've compiled a new kernel with the patch. The machine is no longer

> crashing on pf_send_icmp(). However, I now see memory leaking until the

> machine locks up (it doesn't crash but its network becomes unusable).

> Unfortunately, it then also puts all CARP interfaces in MASTER state,

> though the other node works perfectly as master already. This will, of

> course, knock down our entire network until I manually put down the carp

> interfaces.

>

> I have increased kern.maxclusters to gain more time for debugging of the

> memory leak. However, all I could find out so far is that lots of mbufs

> are allocated while there is no significant traffic to be handled

> (remember the machine is the CARP backup). The machine crashes within 15

> minutes after reboot.

ok that is weird. icmp_error as called in pf_send_icmp does not m_free

anything but the passed mbuf, and we now just bail if tghe allocation

of it fails. so i have a hard time seeing this as related... might be

something completely different. and finding mbuf leaks tends to be

damn hard and following a lot of code...
--

Henning Brauer, hb@bsws.de, henning@openbsd.org

BS Web Services, http://bsws.de

Full-Service ISP - Secure Hosting, Mail and DNS Services

Dedicated Servers, Rootservers, Application Hosting - Hamburg & Amsterdam