Re: NFS oops in 2.6.26rc4

["Chuck Lever" <chuck.lever@oracle.com>]

On Wed, Jun 4, 2008 at 2:20 PM, Dave Jones <davej@redhat.com> wrote:
> On Wed, Jun 04, 2008 at 02:13:08PM -0400, Chuck Lever wrote:
>  >
>  > On Jun 4, 2008, at 10:19 AM, Dave Jones wrote:
>  >
>  > > On Fri, May 30, 2008 at 03:37:01PM -0400, Chuck Lever wrote:
>  > >
>  > >>> Something else of note which I hadn't seen before, usually things
>  > >>> lock
>  > >>> up just after that first oops. For some reason, today it survived
>  > >>> a little longer, but things really went downhill fast.
>  > >>> It survived a 'dmesg ; scp dmesg davej@gelk', and then wedged solid.
>  > >>> So as well as the oops, it seems we're corrupting memory too.
>  > >>> For reference, this kernel has both SLUB_DEBUG and PAGEALLOC_DEBUG
>  > >>> enabled.
>  > >>
>  > >> I haven't seen this kind of problem here with .26, but yes, it does
>  > >> look like something is clobbering memory during an NFS mount.
>  > >>
>  > >> I introduced some NFS mount parsing changes in this commit range:
>  > >>
>  > >>   2d767432..82d101d5
>  > >>
>  > >> A quick bisect should show which, if any of these, is the guilty
>  > >> party.  If any of these are the problem, I suspect it's 3f8400d1.
>  > >
>  > > I didn't get time to try this out yet (hopefully tomorrow).
>  > > In the meantime, we've just gotten word of another user seeing memory
>  > > corruption with nfs - https://bugzilla.redhat.com/show_bug.cgi?id=449958
>  >
>  > 449958 could very well be the same problem.  The stack traceback is a
>  > lot cleaner than the one you originally sent, but there are a lot of
>  > similarities.  (I doubt this is related to symlinks, as the comment
>  > suggests).
>  >
>  > Is commit 86d61d863 applied to the current rawhide kernel?
>
> That kernel was .26rc4.git2, so unless it's only gone in in the last day
> or two, yes.  (Bandwidth impaired right now, and no local git repo to check)

Argh, I was afraid of that.  I expected that commit to improve things.
 Maybe it did, but this is a different problem?  You're going to force
me to actually think about this.  :-)

In any event, a bisect would be helpful here, when you can.  I will
also stare at the traceback in 449958 and see if anything new jumps
out.  It's certainly taken the heat off of the NFS client; it looks
like an rpcbind issue.

-- 
Chuck Lever
chuck[dot]lever[at]oracle[dot]com