Bug 175830

If you read from the snapshot at the same time as the origin is being written
to, the data you get back may be contaminated by what is being written to the
origin.

Neither the origin nor snapshot are actually corrupted on disk by this bug.
Rather it is read() that can return the wrong blocks of data, so anything
reading the snapshot device (such as a backup program) may see corruption.

The correct read path is fast compared to the time taken by the code in the path
that can race against it - I think this explains why occurrences of the problem
are fortunately rare and hard to reproduce.  (If you re-read the corrupt data,
it'll always be correct - the race can't happen twice on the same data.)

This might be partly responsible for bug 174742 and others.



Explanation
-----------
dm-snap.c:

                /*
                 * FIXME: this read path scares me because we
                 * always use the origin when we have a pending
                 * exception.  However I can't think of a
                 * situation where this is wrong - ejt.
                 */
                                                                                
                /* Do reads */
                down_read(&s->lock);
                                                                                
                /* See if it it has been remapped */
                e = lookup_exception(&s->complete, chunk);
                if (e)
                        remap_exception(s, e, bio);
                else
                        bio->bi_bdev = s->origin->bdev;
                                                                                
                up_read(&s->lock);
                                                                                

As the FIXME suggests, there are situation when                                
              what it does is wrong.

Consider:                                         
  A write to origin creates a pe.
  A read from snapshot gets mapped to the origin.
  The pe processing is completed and the write to the origin completes.
  The snapshot read happens but sees what was just written to the origin.


A new dm target, rwsplit, was written to make it straightforward to reproduce
this scenario.  

http://www.kernel.org/pub/linux/kernel/people/agk/patches/2.6/editing/dm-rwsplit.patch

This target sends reads to one device and writes to another.  The test script
attached uses this to suspend reads to a device whilst allowing writes to
proceed.  The script takes 3 copies of the snapshot device.  They should all be
identical as nothing is writing to the snapshot.  But the middle one gets
contaminated with the data written to the origin.  This happens every time on my
test machine.


A fix will have to add code to track the I/O through the snapshot target -
either directly (as elsewhere in dm targets) or possibly by extending the
pending exception code.