Re: [PATCH 2/6] migration: introduce savevm, loadvm, delvm QMP commands

["Dr. David Alan Gilbert" <dgilbert@redhat.com>]

* Daniel P. Berrangé (berrange@redhat.com) wrote:
> On Fri, Jul 03, 2020 at 06:00:50PM +0100, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrangé (berrange@redhat.com) wrote:
> > > On Fri, Jul 03, 2020 at 05:22:46PM +0100, Dr. David Alan Gilbert wrote:
> > > > * Daniel P. Berrangé (berrange@redhat.com) wrote:
> > > > > On Fri, Jul 03, 2020 at 05:10:12PM +0100, Dr. David Alan Gilbert wrote:
> > > > > > * Daniel P. Berrangé (berrange@redhat.com) wrote:
> > > > > > > On Fri, Jul 03, 2020 at 04:49:33PM +0100, Dr. David Alan Gilbert wrote:
> > > > > > > > * Daniel P. Berrangé (berrange@redhat.com) wrote:
> > > > > > > > > On Thu, Jul 02, 2020 at 01:12:52PM -0500, Eric Blake wrote:
> > > > > > > > > > On 7/2/20 12:57 PM, Daniel P. Berrangé wrote:
> > > > > > > > > > > savevm, loadvm and delvm are some of the few commands that have never
> > > > > > > > > > > been converted to use QMP. The primary reason for this lack of
> > > > > > > > > > > conversion is that they block execution of the thread for as long as
> > > > > > > > > > > they run.
> > > > > > > > > > > 
> > > > > > > > > > > Despite this downside, however, libvirt and applications using libvirt
> > > > > > > > > > > has used these commands for as long as QMP has existed, via the
> > > > > > > > > > > "human-monitor-command" passthrough command. IOW, while it is clearly
> > > > > > > > > > > desirable to be able to fix the blocking problem, this is not an
> > > > > > > > > > > immediate obstacle to real world usage.
> > > > > > > > > > > 
> > > > > > > > > > > Meanwhile there is a need for other features which involve adding new
> > > > > > > > > > > parameters to the commands. This is possible with HMP passthrough, but
> > > > > > > > > > > it provides no reliable way for apps to introspect features, so using
> > > > > > > > > > > QAPI modelling is highly desirable.
> > > > > > > > > > > 
> > > > > > > > > > > This patch thus introduces trival savevm, loadvm, delvm commands
> > > > > > > > > > 
> > > > > > > > > > trivial
> > > > > > > > > > 
> > > > > > > > > > > to QMP that are functionally identical to the HMP counterpart, including
> > > > > > > > > > > the blocking problem.
> > > > > > > > > > 
> > > > > > > > > > Should we name them 'x-savevm', 'x-loadvm', 'x-delvm' to give ourselves room
> > > > > > > > > > to change them when we DO solve the blocking issue?  Or will the solution of
> > > > > > > > > > the blocking issue introduce new QMP commands, at which point we can add QMP
> > > > > > > > > > deprecation markers on these commands to eventually retire them?
> > > > > > > > > 
> > > > > > > > > I was in two minds about this, so I'm open to arguments either way.
> > > > > > > > > 
> > > > > > > > > The primary goal is for libvirt to consume the APIs as soon as possible,
> > > > > > > > > and generally libvirt doesn't want todo this is they are declared experimental
> > > > > > > > > via a "x-" prefix. So that pushes me away from "x-".
> > > > > > > > > 
> > > > > > > > > If we don't have an "x-" prefix and want to make changes, we can add extra
> > > > > > > > > parameters to trigger new behaviour in backwards compatible manner. Or we can
> > > > > > > > > simply deprecate these commands, deleting them 2 releases later, while adding
> > > > > > > > > completely new commands.
> > > > > > > > > 
> > > > > > > > > If we think the prposed design will definitely need incompatible changes in
> > > > > > > > > a very short time frame though, that would push towards "x-".
> > > > > > > > > 
> > > > > > > > > So IMHO the right answer largely depends on whether there is a credible
> > > > > > > > > strategy to implement the ideal non-blocking solution in a reasonable amount
> > > > > > > > > of time. I can't justify spending much time on this myself, but I'm willing
> > > > > > > > > to consider & try proposals for solving the blocking problem if they're not
> > > > > > > > > too complex / invasive.
> > > > > > > > 
> > > > > > > > Remind me, what was the problem with just making a block: migration
> > > > > > > > channel, and then we can migrate to it?
> > > > > > > 
> > > > > > > migration only does vmstate, not disks. The current blockdev commands
> > > > > > > are all related to external snapshots, nothing for internal snapshots
> > > > > > > AFAIK. So we still need commands to load/save internal snapshots of
> > > > > > > the disk data in the qcow2 files.
> > > > > > > 
> > > > > > > So we could look at loadvm/savevm conceptually as a syntax sugar above
> > > > > > > a migration transport that targets disk images, and blockdev QMP command
> > > > > > > that can do internal snapshots. Neither of these exist though and feel
> > > > > > > like a significantly larger amount of work than using existing functionality
> > > > > > > that is currently working.
> > > > > > 
> > > > > > I think that's what we should aim for; adding this wrapper isn't gaining
> > > > > > that much without moving a bit towards that; so I would stick with the
> > > > > > x- for now.
> > > > > 
> > > > > The question is how much work that approach will be and whether anyone can
> > > > > realistically commit to doing that ? It looks like a much larger piece of
> > > > > work in both QEMU and libvirt side to do that. I don't want to see us stuck
> > > > > with a x-savevm for years because no one has resource to work on the perfect
> > > > > solution. If we did get a perfect solution at a point in future, we can
> > > > > still deprecate and then remove any "savevm" command we add to QMP.
> > > > 
> > > > I'd at least like to understand that we've got a worklist for it though.
> > > > We've already got qemu_fopen_bdrv - what's actually wrong with that - is
> > > > that enough to do the migrate to a stream (given a tiny amount of
> > > > syntax) ?
> > > 
> > > It is close. The migration code works with the QEMUFile layer, but in terms
> > > of the monitor commands the current framework expects a QIOChannel based
> > > QEMUFile. It would be possible to add new helpers to work with the bdrv
> > > backed QEMUFile. The ideal would be to create a QIOChannel impl that is
> > > backed by a block device though. At that point there would only be a single
> > > QEMUFile impl based on QIOChannel. 
> > > 
> > > That would be another step closer to unlocking the ability to eliminate the
> > > QEMUFile wrapper entirely. QEMUFile does I/O buffering too for performance,
> > > that could be done in a QIOChannel layer too, as that's a concept useful
> > > for other QIOChannel users too.
> > 
> > There's some separate patches on list to do buffering in bdrv for
> > vmsave because apparently the qemufile ones don't play well with it.
> > 
> > > That's still only the vmstate part though, and a solution is needed for
> > > the internal snapshot handling.
> > 
> > Which bit is the bit that makes it block?
> 
> The entire  save_snapshot / load_snapshot methods really. They doing I/O
> operations throughout and this executes in context of the thread running
> the monitor, so their execution time is proportional to I/O time.

Is there any reason for the migrationy bit to run synchronously then?

Could those snapshotting things be done with any of the block copy
processes that run asynchronously?

Dave


> 
> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK