http://www.holm.cc/stress/log/cons141.html

[Gleb Smirnoff]

  Colleagues,

On Wed, Jul 13, 2005 at 10:51:55AM +0200, Peter Holm wrote:
P> > how often the panic in subject fires? I guess it is related to the problem
P> > discussed in this thread [1], and may be even the patch I've posted there
P> > will help.
P> >
P> 
P> The panic is almost instantanious if I run the stress test while doing
P> an "arp -d xxx".

The attached WIP patch helps me. The [1] script can't drop my box to panic
anymore, when running in 15 instances on 4-x CPU box.

The idea of the patch is to hold rtentry lock during all the manipulation
of the rtentry itself and its llinfo. To achieve this, I've changed 
rt_check() and arplookup() to return a locked rtentry.

Not everything is OK, yet. For example when I turn witness off via sysctl,
and all the scripts are running at this moment, I get a panic in
route_output() - "sleeping thread owns not-sleepable lock".

[1] ( while (true); do
	arp -d 81.19.64.111  >/dev/null 2>&1;
	ping -c 1 -t 1 81.19.64.111 >/dev/null 2>&1;
      done) &

-- 
Totus tuus, Glebius.
GLEBIUS-RIPN GLEB-RIPE