Re: [Qemu-devel] [PATCH 0/3] client_migrate_switch and auto_switch (RHBZ 725009)

[Daniel P. Berrange]

On Fri, Aug 26, 2011 at 01:00:29PM +0200, Gerd Hoffmann wrote:
>   Hi,
> 
> >>Ok, so I think we should be able to fix the race outlined above
> >>without adding new monitor commands, just by letting libvirt set the
> >>spice ticket in stage 2.
> >
> >Is that different then what I suggested in my reply to Daniel's 5 stage 
> >outline?
> 
> I think we have to care to not mix up switch-host and seamless spice
> client migration, your reply seems to refer top some seamless spice
> migration implementation ideas.
> 
> Today we seem to have this workflow:
> 
>    (1)  migration src->dst (stage 3)
>    (2a) libvirt sets spice ticket at dst (not sure which stage)
>    (2b) spice client switches connection to dst

No, that's not what we do currently.

> 2a and 2b are racing here.  I think we can fix the race by doing
> this instead:
> 
>   (1) libvirt sets spice ticket at dst (stage 2)
>   (2) migration src->dst (stage 3)
>   (3) spice client switches connection to dst

This is actually what we have already. The problem is that
the 'migration src->dst' here can take an arbitrary amount
of time. So by the time the spice client switchs to dst,
the ticket will likely have already expired.

The only way to fix this AFAICT is to ensure the SPICE
client connects to dst before migration starts, but delays
display switch until the end. eg

    (1) libvirt sets spice ticket at dst (libvirt stage 2)
    (2) spice client connects to dst (start of libvirt stage 3)
    (3) migration src->dst (middle libvirt stage 3)
    (4) spice client switches to dst (end of libvirt stage 3 or stage 5)


Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|