Re: [Qemu-devel] Dropping the MONITOR_CMD_ASYNC

[Alon Levy]

On Mon, Dec 12, 2011 at 11:53:28AM -0600, Anthony Liguori wrote:
> On 12/12/2011 11:22 AM, Alon Levy wrote:
> >On Mon, Dec 12, 2011 at 10:24:53AM -0600, Anthony Liguori wrote:
> >>On 12/12/2011 10:00 AM, Alon Levy wrote:
> >>>On Mon, Dec 12, 2011 at 03:23:35PM +0000, Stefan Hajnoczi wrote:
> >>>>On Mon, Dec 12, 2011 at 12:39 PM, Gerd Hoffmann<address@hidden>   wrote:
> >>>>>On 12/12/11 13:10, Stefan Hajnoczi wrote:
> >>>>>>On Mon, Dec 12, 2011 at 11:29 AM, Gerd Hoffmann<address@hidden>   wrote:
> >>>>>>>On 12/12/11 11:18, Stefan Hajnoczi wrote:
> >>>>>>>>On Sun, Dec 11, 2011 at 10:29 AM, Alon Levy<address@hidden>   wrote:
> >>>>>>>>>On Thu, Dec 08, 2011 at 05:45:44PM -0200, Luiz Capitulino wrote:
> >>>>>>>>>>Hi there,
> >>>>>>>>>>
> >>>>>>>>>>I'm about to completely drop the MONITOR_CMD_ASYNC API, but it 
> >>>>>>>>>>turns out that
> >>>>>>>>>>the command client_migrate_info uses it. That's a legacy interface 
> >>>>>>>>>>and has to
> >>>>>>>>>>be dropped, no command should be using it...
> >>>>>>>>>>
> >>>>>>>>>>Something tells me that if I just drop it (and change the command 
> >>>>>>>>>>to use the
> >>>>>>>>>>regular interface), bad things will happen. Am I right? :)
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>The monitor command client_migrate_info needs to complete after 
> >>>>>>>>>getting
> >>>>>>>>>an ACK message from the currently connected spice client (this is the
> >>>>>>>>>only case where this is required - if there is no client then the
> >>>>>>>>>MONITOR_CMD_ASYNC API won't be used). This in turn requires the main
> >>>>>>>>>thread to perform select and call the callback that will process this
> >>>>>>>>>ACK. That's why the MONITOR_CMD_ASYNC API was used.
> >>>>>>>>>
> >>>>>>>>>I'm not aware of any other way to do this, I'll be glad for any help
> >>>>>>>>>here. Also, I understand this is not what is not true async, since 
> >>>>>>>>>one
> >>>>>>>>>would expect a true async interface to support multiple in flight
> >>>>>>>>>monitor commands. If there is any ETA or existing way to do this we
> >>>>>>>>>could change the implementation of client_migrate_info.
> >>>>>>>>
> >>>>>>>>Is it possible to use a QMP event to signal completion instead of a
> >>>>>>>>MONITOR_CMD_ASYNC command?
> >>>>>>>
> >>>>>>>Problem is this breaks the qemu<->   libvirt interface.
> >>>>>>
> >>>>>>I had the same issue with the block_job_cancel command, which Adam
> >>>>>>Litke and Eric Blake helped us fix and find a backward-compatible
> >>>>>>libvirt solution for:
> >>>>>>
> >>>>>>http://www.redhat.com/archives/libvir-list/2011-November/msg01351.html
> >>>>>
> >>>>>Isn't going to fly as waiting for completion isn't optional in that
> >>>>>case.  Workflow is this:
> >>>>>
> >>>>>(1) libvirt issues client_migrate_info command.
> >>>>>(2) qemu forwards it to spice-server, which in turn forwards it to
> >>>>>    the spice client (if connected).
> >>>>>(3) spice client connects to the migration target machine.
> >>>>>(4) spice client signals completion to spice-server, which in turn
> >>>>>    notifies qemu.
> >>>>>(5) qemu calls the monitor completion callback, libvirt gets
> >>>>>    client_migrate_info result.
> >>>>>(6) libvirt issues migrate command.
> >>>>>
> >>>>>The problem is that (3) must be finished before (6) because qemu on the
> >>>>>target side doesn't accept incoming tcp connections any more once the
> >>>>>migration started.
> >>>>>
> >>>>>I don't see a way to switch this to qmp events without breaking old
> >>>>>libvirt versions managing new qemu.
> >>>>
> >>>>I don't see a solution in this case either.  Looks like libvirt
> >>>>supports this command since 0.9.2 so it's not a good idea to just yank
> >>>>it.
> >>>>
> >>>>It might be possible for the QEMU client_migrate_info handler to run a
> >>>>nested event loop in the legacy libvirt case.  This would suck since
> >>>>the VM is unresponsive while waiting for spice migration to complete.
> >>>>New libvirt would call the async version of the command which is
> >>>>well-behaved and uses a QMP event to signal completion.
> >>>
> >>>I agree that a nested event loop would be a bad solution, the point is
> >>>to let the guest continue to work while waiting, otherwise you destroy
> >>>the live migration experience, might as well disconnect the client from
> >>>the source and have it connect to the target.
> >>>
> >>>I thought the whole point of MONITOR_CMD_ASYNC was to allow this
> >>>scenario. So iiuc QMP is the alternative but it would require a rewrite,
> >>>i.e. break existing users like libvirt. Hence my suggestion as a reply
> >>>to Luiz's initial email that we just deprecate MONITOR_CMD_ASYNC right
> >>>now instead of dropping it,
> >>
> >>It *has* to be dropped.  Any command using it is fundamentally broken.
> >>
> >
> >Care to explain what is fundamental about it?
> 
> The commands are broken as they sit today.  If you issue an async
> command and an error occurs, the error will not be associated with
> the right command.
> 
> >
> >>The command should have never been introduced in the first place to use 
> >>async.
> >
> >Gerd gave a very good explanation of the requirement, I don't think I
> >can add anything to it. What you suggest is to rewrite it, I just don't
> >understand why we can't leave the MONITOR_CMD_ASYNC around until there
> >is both an alternative and libvirt learns to use it.
> 
> It's broken and more importantly, the longer we leave it there the
> more likely people will use it and create more problems.
> 
> We aren't breaking an ABI here.  libvirt can detect that this
> command is gone and probe for the new command.

Obviously this will break current libvirt. So it means we need to make
sure we introduce the new command and libvirt's usage of it follows
ASAP. Still won't help with <=current libvirt.

How do you call the monitor interface btw? I thought it was a sort of
ABI.

> 
> This shouldn't have gone in in the first place and yes, that's my
> responsibility to enforce.
> 
> Regards,
> 
> Anthony Liguori
> 
> >
> >>
> >>Regards,
> >>
> >>Anthony Liguori
> >>
> >
> 
>