Re: [Qemu-devel] [PATCH V5 0/8] add internal snapshot support at block device level

[Stefan Hajnoczi <stefanha@redhat.com>]

On Thu, Jul 18, 2013 at 02:34:52PM +0200, Kevin Wolf wrote:
> Am 11.07.2013 um 07:46 hat Wenchao Xia geschrieben:
> >   This series brings internal snapshot support at block devices level, now we
> > have two three methods to do block snapshot lively: 1) backing chain,
> > 2) internal one and 3) drive-back up approach.
> > 
> > Comparation:
> >              Advantages:                            Disadvantages:
> > 1)    delta data, taken fast, export, size        performance, delete slow.
> > 2)  taken fast, delete fast, performance, size       delta data, format
> > 3)      performance, export, format               taken slow, delta data, size
> > 
> >   I think in most case, saving vmstate in an standalone file is better than
> > saving it inside qcow2, So suggest treat internal snapshot as block level
> > methods and not encourage user to savevm in qcow2 any more.
> > 
> > Implemention details:
> >   To avoid trouble, this serial have hide ID in create interfaces, this make
> > sure no chaos of ID and name will be introduced by these interfaces.
> >   There is one patch may be common to Pavel's savvm transaction, patch 1/11,
> > others are not quite related. Patch 1/11 will not set errp when no snapshot
> > find, since patch 3/11 need to distinguish real error case.
> > 
> > Next steps to better full VM snapshot:
> >   Improve internal snapshot's export capability.
> >   Better vmstate saving.
> > 
> >   Thanks Kevin to give advisement about how add it in qmp_transaction, oldest
> > version comes drom Dietmar Maurer.
> > 
> > v3:
> >   General:
> >   Rebased after Stenfan's driver-backup patch V6.
> > 
> >   Address Eric's comments:
> >   4/9: grammar fix and better doc.
> >   5/9: parameter name is mandatory now. grammar fix.
> >   6/9: redesiged interface: take both id and name as optional parameter, return
> > the deleted snapshot's info.
> > 
> >   Address Stefan's comments:
> >   4/9: add '' around %s in message. drop code comments about vm_clock.
> >   9/9: better doc, refined the code and add more test case.
> > 
> > v4:
> >   Address Stefan's comments:
> >   4/9: use error_setg_errno() to show error reason for bdrv_snapshot_create(),
> > spell fix and better doc.
> >   5/9: better doc.
> >   6/9: remove spurious ';' in code, spell fix and better doc.
> > 
> > v5:
> >   Address Kevin's comments:
> >   3/8, 4/8, 8/8: remove the limit of numeric snapshot name.
> >   General change:
> >   4/8: use existing type as parameter in qapi schema.
> > 
> > Wenchao Xia (8):
> >   1 snapshot: new function bdrv_snapshot_find_by_id_and_name()
> >   2 snapshot: distinguish id and name in snapshot delete
> >   3 qmp: add internal snapshot support in qmp_transaction
> >   4 qmp: add interface blockdev-snapshot-internal-sync
> >   5 qmp: add interface blockdev-snapshot-delete-internal-sync
> >   6 hmp: add interface hmp_snapshot_blkdev_internal
> >   7 hmp: add interface hmp_snapshot_delete_blkdev_internal
> >   8 qemu-iotests: add 056 internal snapshot for block device test case
> 
> Okay, reviewed the whole series. I had some comments on two or three
> patches.
> 
> One thing that I want to add is that in order to provide transactional
> behaviour (i.e. snapshotting multiple device in one transaction should
> result in a consistent snapshot), I think we may in fact need to stop
> the VM while taking the snapshots.

If we are doing guest memory snapshots we definitely need to stop vcpus.

For disk-only snapshots I do not think it's necessary to stop the guest
because:

1. Guest I/O can only be processed from the QEMU global mutex.  But we
   currently hold it so no one else can make progress.

2. In the dataplane case the drain callback that I added stops the
   dataplane thread temporarily.  It cannot be restarted until the main
   loop iterates again (see #1).

Therefore I think stopping vcpus is not necessary for disk-only
snapshots.

Stefan