On Wed, Jun 19, 2013 at 04:53:38PM +0800, Wenchao Xia wrote: > 于 2013-6-19 15:46, Stefan Hajnoczi 写道: > >On Wed, Jun 19, 2013 at 02:18:48PM +0800, Wenchao Xia wrote: > >>于 2013-6-18 22:20, Stefan Hajnoczi 写道: > >>>On Mon, Jun 17, 2013 at 11:25:26AM +0800, Wenchao Xia wrote: > >>>>于 2013-6-15 17:55, Eric Blake 写道: > >>>>>Should this command be made available via 'transaction'? That is, if I > >>>>>have a two-disk VM, and use 'transaction' to take a snapshot of both > >>>>>disks at once, shouldn't I also have a way to delete the snapshots of > >>>>>both at once, or gracefully fail without data loss if the second one has > >>>>>problems? > >>>> > >>>> I think adding it in transaction is not very useful but brings more > >>>>complexity. Transcation is used to guareentee all operations are taken > >>>>in one time point, for example, snapshot creation use it to make sure > >>>>all are consistent to VM. But for deletion, this requirement do not > >>>>exist. > >>> > >>>I guess the problem is: can we make internal snapshot deletion > >>>transactional? It's hard to do rollback for snapshot deletion. > >>> > >> Several deletion in transaction equals to several calls of > >>'blockdev-snapshot-delete-internal-sync', unlike creation, so I hope > >>not add it which have rollback issue. > >> > >> > >>>But batching is definitely useful for doing 'delvm' in QMP. I just > >>>don't think transactions help. We just need a 'delvm' equivalent in > >>>QMP. > >>> > >> Maybe the caller can encapsulate a batch interface at its level. > > > >'delvm' is a batch interface - it deletes internal snapshots that have > >the same name across multiple devices. > > > >It's not as flexible as: > >blkdev-internal-snapshot-delete drive0 drive2 drive4 > > > >Because that would allow you to specify specific drives. > > > Do you mean this interface should be changed as > blkdev-internal-snapshot-delete devices_array name *id? > > It seems not much difference with following methods: > > for device in device_list: > blkdev-internal-snapshot-delete device name
The ability to batch in a single QMP command feels a little nicer. There are less opportunities for the operation to stop half-way through. It would be usable as a QMP 'delvm' and in a more flexible way to delete multiple internal snapshots. Would be interesting to see what Eric Blake thinks from a libvirt perspective. Stefan
