Re: 2.6.39-rc4 BUG: unable to handle kernel NULL pointer dereference at 0000000c IP: cfq_insert_request+0x1d/0x3f5

[Vivek Goyal <vgoyal@redhat.com>]

On Thu, Apr 21, 2011 at 11:04:45AM -0400, Jeff Moyer wrote:
> Jens Axboe <axboe@kernel.dk> writes:
> 
> > On 21/04/2011, at 09.16, Michal Hocko <mhocko@suse.cz> wrote:
> >
> >> On Wed 20-04-11 19:33:19, Jens Axboe wrote:
> >>> On 2011-04-20 15:29, Michal Hocko wrote:
> >>>> On Wed 20-04-11 15:13:15, Jens Axboe wrote:
> >>>>> On 2011-04-20 14:58, Michal Hocko wrote:
> >> [...]
> >>>>>> [   31.207888] Oops: 0000 [#1] PREEMPT [   31.208186] last sysfs file: /sys/devices/pci0000:00/0000:00:1f.2/host0/target0:0:0/0:0:0:0/block/sda/queue/scheduler
> >>>>> 
> >>>>> Ahh hang on, this may be a good clue. Does your boot scripts change the
> >>>>> IO scheduler?
> >>>> 
> >>>> Good one...
> >>>> Yes, I have:
> >>>> echo deadline > /sys/block/sda/queue/scheduler
> >>>> in /etc/rc.local
> >>>> 
> >>>> I am able to boot after I remove it. This is the first time I have seen
> >>>> "last sysfs file" being useful.
> >>>> Still want me to test the patch from the other email?
> >>> 
> >>> Is this a new addition to your system? IOW, how certain are you that
> >>> this is a regression that occured between rc3 and rc4?
> >> 
> >> No, I am setting the scheduler this way for quite some time. If I use it
> >> rc4 explodes while rc3 boots just fine. I am wondering, can this be a
> >> timing issue? I am able to set the scheduler after system settles down
> >> after boot and kde starts.
> >> 
> >> I am going to bisect, let's see if I can find anything.
> >
> > Thanks, that would be great!
> 
> OK, this is a long shot, but in a derivative kernel, I saw what may be
> the same issue.  Is this kernel built with CONFIG_BLK_CGROUP=n by
> chance?  The exact problem I saw was a panic on boot in
> cfq_insert_request+0x77, which mapped to this:
> 
> /usr/src/debug/kernel-2.6.32-135.el6/linux-2.6.32-135.el6.x86_64/block/cfq-iosched.c:1997
> ffffffff8125c390:       49 8b 84 24 a8 00 00    mov    0xa8(%r12),%rax
> ffffffff8125c397:       00                          <---------------------
> ffffffff8125c398:       83 80 ec 02 00 00 01    addl   $0x1,0x2ec(%rax)
> 
> cfq-iosched.c:1997 looks like this:
> 
>         (RQ_CFQG(rq))->dispatched++;
> 
> Enabling CONFIG_BLK_CGROUP made the problem go away.  Again, not sure
> it's the same thing, but I figured I'd speak up in case it helps.

Jeff, 

I think we had fixed this issue upstream with following commit.

commit 50eaeb323a170e231263ccb433bb2f99bd9e27ac
Author: Dmitry Monakhov <dmonakhov@openvz.org>
Date:   Wed Apr 28 19:50:33 2010 +0200

    cfq-iosched: fix broken cfq_ref_get_cfqf() for CONFIG_BLK_CGROUP=y &&
CFQ_GR
    
    We should return the cfq_group for this case, not NULL.
    
    Signed-off-by: Jens Axboe <jens.axboe@oracle.com>


I just booted my system with 39-rc4 with BLK_CGROUP=n. So I doubt that's the
issue here.

Thanks
Vivek