6.1-RELEASE, em0 high interrupt rate and nfsd eats lots of cpu

[Rong-en Fan]

On 5/15/06, Dmitriy Kirhlarov <dkirhlarov at oilspace.com> wrote:
> On Mon, May 15, 2006 at 02:15:08PM -0400, Rong-en Fan wrote:
> > Hi,
> >
> > After upgrading from 5.5-PRERELEASE to 6.1-RELEASE on one
> > nfs server today, I noticed that the load is very high, ranging from 4.x
> > to 30.x, depends how many nfsd I run. From mrtg traffic graph, I did
> > not notice there is high traffic. This box is 2 physical Xeon CPU w/
>
> I have same situation today on RC2.
> One client installing world from nfs share.
> nfsd eat 91% CPU, load average 6-8. Very small disk activitie.
> I don't look interrupt rate.
> I, also, have em0.

After some digging, I found the cpu-eater nfsd can be triggered by
running ``du'' on nfs client (both FreeBSD 6.1-R and Linux box).
The nfsd will eat lots of CPU. After the client's du is finished, the nfsd
still eat lots of CPU. The workaround is to

/etc/rc.d/nfsd restart

Everything will be just fine. Besides du, on FreeBSD 6.1-R client,
buildkernel over nfs will trigger the same behavior.

I just downgraded this box to 6.0-RELEASE and everything works
fine. Running du or buildkernel from nfs client do not trigger the
same behavior. I will try to do a binary search from 6.0-R to 6.1-R
see if I can find out related commits.

By the way, I have another nfs server running 6.1-RELEASE,
but it does not exhibit this behavior. Kernel conf and sysctl
are basically the same for both boxes.

Regards,
Rong-En Fan