Re: use-after-free on log replay failure

[Dave Chinner <david@fromorbit.com>]

On Tue, Aug 12, 2014 at 03:39:02PM +0300, Alex Lyakas wrote:
> Hello Dave, Brian,
> I will describe a generic reproduction that you ask for.
> 
> It was performed on pristine XFS code from 3.8.13, taken from here:
> git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git
....
> I mounted XFS with the following options:
> rw,sync,noatime,wsync,attr2,inode64,noquota 0 0
> 
> I started a couple of processes writing files sequentially onto this
> mount point, and after few seconds crashed the VM.
> When the VM came up, I took the metadump file and placed it in:
> https://drive.google.com/file/d/0ByBy89zr3kJNa0ZpdmZFS242RVU/edit?usp=sharing
> 
> Then I set up the following Device Mapper target onto /dev/vde:
> dmsetup create VDE --table "0 41943040 linear-custom /dev/vde 0"
> I am attaching the code (and Makefile) of dm-linear-custom target.
> It is exact copy of dm-linear, except that it has a module
> parameter. With the parameter set to 0, this is an identity mapping
> onto /dev/vde. If the parameter is set to non-0, all WRITE bios are
> failed with ENOSPC. There is a workqueue to fail them in a different
> context (not sure if really needed, but that's what our "real"
> custom
> block device does).

Well, they you go. That explains it - an asynchronous dispatch error
happening fast enough to race with the synchronous XFS dispatch
processing.

dispatch thread			device workqueue
xfs_buf_hold();
atomic_set(b_io_remaining, 1)
atomic_inc(b_io_remaining)
submit_bio(bio)
queue_work(bio)
xfs_buf_ioend(bp, ....);
  atomic_dec(b_io_remaining)
xfs_buf_rele()
				bio error set to ENOSPC
				  bio->end_io()
				    xfs_buf_bio_endio()
				      bp->b_error = ENOSPC
				      _xfs_buf_ioend(bp, 1);
				        atomic_dec(b_io_remaining)
					  xfs_buf_ioend(bp, 1);
					    queue_work(bp)
xfs_buf_iowait()
 if (bp->b_error) return error;
if (error)
  xfs_buf_relse()
    xfs_buf_rele()
      xfs_buf_free()

And now we have a freed buffer that is queued on the io completion
queue. Basically, it requires the buffer error to be set
asynchronously *between* the dispatch decrementing it's I/O count
after dispatch, but before we wait on the IO.

Not sure what the right fix is yet - removing the bp->b_error check
from xfs_buf_iowait() doesn't solve the problem - it just prevents
this code path from being tripped over by the race condition.

But, just to validate this is the problem, you should be able to
reproduce this on a 3.16 kernel. Can you try that, Alex?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs