Re: [Qemu-devel] qemu and qemu.git -> Migration + disk stress introduces qcow2 corruptions

[Juan Quintela]

Anthony Liguori <address@hidden> wrote:
> On 11/09/2011 11:02 AM, Avi Kivity wrote:
>> On 11/09/2011 06:39 PM, Anthony Liguori wrote:
>>>
>>> Migration with qcow2 is not a supported feature for 1.0.  Migration is
>>> only supported with raw images using coherent shared storage[1].
>>>
>>> [1] NFS is only coherent with close-to-open which right now is not
>>> good enough for migration.
>>
>> Say what?
>
> Due to block format probing, we read at least the first sector of the
> disk during start up.
>
> Strictly going by what NFS guarantees, since we don't open on the
> destination *after* as close on the source, we aren't guaranteed to
> see what's written by the source.
>
> In practice, because of block format probing, unless we're using
> cache=none, the first sector can be out of sync with the source on the
> destination.  If you use cache=none on a Linux client with at least a
> Linux NFS server, you should be relatively safe.

You are not :-(

If you are using a format that "caches" data, like qcow2 with the L1/L2
cache, you are not safe.  You need to reopen (or discard metadata +
re-read it).  Notice that raw nowadays also has metadata (we can resize
the image on the flight, and we need to reopen to find that).

About the coherence problem, I just sent the patches that we had on RHEL
to the list.  With cache=none, both NFS & iSCSI & Fiberchannel are ok
(module the previous problem of metadata).  If you look at the second
patch that I sent, it "tries" to flush the read cache for a block
device.  Problem with the patch are:
- BLKFLSBUF is linux specific
- BLKFLSBUF only works for "some block devices"
- Christoph just Nacked it due to previous reasons.

In resume:
- If we use raw, we don't resize images, and we use a clustered
  filesystem, qemu.git migration works.

- If we change metadata (qcow2, raw resize, ...) we need to re-read
  metadata (we just close +open on RHEL).

- If we use NFS: we need to use cache=none, or close+open consistency

- if we use iSCSI: we need to use cache=none. close+open is not enough
  for consistency.  The ioctl patch that I sent happens to work on
  linux, but it is not even guaranteed to work there.  And if our block
  layer gurus told us not to use the ioctl() I think that we need to do
  just that.

Later, Juan.