nfsclient process stucks in nfsaio

[Rong-En Fan]

Hi,

forget to mention all the clients/servers here are SMP kernel.
After some Googling, a post on current@ 2005/01/12
"NFS problems, locking up" is hightly related to my situation.
An workaround is to set debug.mpsafenet=0, just verified this
indeed works.

Now I'm turning on INVARIANTS, WITNESS to see if there
are some output. However, I'm afriad that I can not get a
serial console access to these machines (and thus no ddb
output :( ).

Thanks,
Rong-En Fan

On 3/10/06, Rong-En Fan <grafan at gmail.com> wrote:
> Hi,
>
> After upgrading several our nfs clients from 5.4-RELEASE to 6.0-RELEASE
> and some are now 6.1-PRERELEASE (a weeks ago). From time to time,
> we saw some processes stuck in nfsaio, and unkillable. These processes
> generate lots of traffic to nfs server (write to nfs, but nfs server's disk does
> not really in write. from netstat, client sends ~100Mbps, on nfs server, iostat
> does not show me ~12.5MB/s). The nfsd on the server side is either in RUN
> or in ufs state. Server is running 5.5-PRELEASE as of yesterday.
>
> Client mount options: rw,nosuid,bg,intr,nodev. Both client and server
> are running
> rpc.lockd, rpc.statd. I'm sure it's not related to any locking problems.
>
> I have another set of nfs server/client both running 6.0-RELEASE. And I can
> easily reproduce this situation on these two boxesnes, just by running
>
>   dd if=/dev/zero of=/nfs/ooo bs=1m
>
> If I do not add bs=1m, it works fine. Of all the boxes I mentioned above,
> I did not do anything special to kernel config, i.e., they are GENERIC w/o
> unnecessary devices and w/ firewal.  Basically, I can do anything on these
> two boxes (they are not in production mode). Any suggestion are welcome.
>
> Thanks,
> Rong-En Fan
>