Re: 2.6.36-rc3: EIP is at scsi_init_io+...

[James Bottomley <James.Bottomley@suse.de>]

On Thu, 2010-09-09 at 14:51 +0200, Jens Axboe wrote:
> On 2010-09-09 14:47, James Bottomley wrote:
> > On Wed, 2010-09-08 at 17:12 -0700, Linus Torvalds wrote:
> >> Hmm. No noise about this one.
> >>
> >> Jens, Fujita, James, any comments?
> >>
> >> On Mon, Aug 30, 2010 at 11:46 AM, Alexey Dobriyan <adobriyan@gmail.com> wrote:
> >>> Not much of a calltrace, it scrolled away because of hardlockup detector.
> >>> On the bright side, radeon KMS worked correctly and actually showed it.
> >>>
> >>> $ addr2line -e vmlinux ffffffff812d207b
> >>> drivers/scsi/scsi_lib.c:1015
> >>>
> >>>  1012  err_exit:
> >>>  1013          scsi_release_buffers(cmd);
> >>>  1014          scsi_put_command(cmd);
> >>>  1015  ===>    cmd->request->special = NULL; <===
> >>>  1016          return error;
> >>>  1017  }
> >>>  1018  EXPORT_SYMBOL(scsi_init_io);
> >>
> >> I do have to say that it looks rather wrong that it accesses "cmd"
> >> after it has done the "scsi_put_command(cmd)" on it.
> >>
> >> I also note that that was introduced pretty recently by commit
> >> 610a63498f7 ("scsi: fix discard page leak"), merged during this merge
> >> window. That does look suspicious to me.
> > 
> > It's a use after free:  The put actually frees the cmnd and then we use
> > it to get to the request.  Most of the time nothing notices, but if you
> > have poison on free enabled, we may see the problem.  The fix is just to
> > reverse the put and the set.
> 
> You are right, I misspoke in my original reply. It's clearing the
> request field, not the command field (which would be bogus of course).

Embarrassingly enough, Tejun spotted this a month ago, and I actually
created a patch and then forgot to apply it.  I've just pushed it out to
the rc-fixes branch:

http://marc.info/?l=linux-scsi&m=128197119525504

James