Re: Migration - Xen 4.3

[Jean-Yves Migeon <jeanyves.migeon%free.fr@localhost>]

Le 19/01/2015 13:09, Stephen Borrill a écrit :
> On Sun, 18 Jan 2015, Brian Marcotte wrote:
> > Has anyone tried Migration recently?
> >
> > I'm getting this on Xen 4.3, Linux dom0:
> >
> >  # xl migrate mydomU otherdom0host
> >  [...]
> >  libxl: error: libxl_dom.c:1063:libxl__domain_suspend_common_callback: 
> > \
> >  guest didn't acknowledge suspend, cancelling request

Curious, the ack is done when HYPERVISOR_suspend() gets called by domY 
(see link just below).

> > On the domU (netbsd) console, I see this:
> >
> >  xenbus_shutdown_handler: xenbus_rm 13
> >  Flushing disk caches: 21 done
> >
> > The domain is completely locked up and doesn't even respond to 
> > "+++++".

Before xenbus handlers get called, do you see any of the DPRINTK(...) 
messages found on this page:

http://nxr.netbsd.org/xref/src/sys/arch/xen/xen/xen_machdep.c#xen_prepare_suspend

> > It is similar with "xm migrate".

migration is merely a suspend => snapshot (done by hypervisor/xl/xend) 
=> resume on the other. So if any of "suspend" or "resume" does not 
work, so does migrate.

> I have never got migration to succeed, nor suspend/resume (which is
> fundamentally the same operation). Sometimes suspend appears to work,
> but resume fails.
>
> The usual symptoms for me are that the domain just continues
> regardless (i.e. it ignores the suspend request) rather than it locks.
> However, from a dom0 point of view, it waits for confirmation of
> suspension and so blocks any further operations on it.
>
> I understand that the problem is reasonably well understood by the
> relevant developers, but that debugging is hard.

Now that's a regression, suspend is supposed to finish correctly (e.g. 
the domain is descheduled and you get a core file that represents VM's 
state). Resuming is the culprit and I failed to get $SPARETIME to 
investigate the situation in the last few months; usual symptoms is a 
corrupted Xen I/O ring that completely thrashes xbdback(4) and leads to 
a paniced domU (or a very, very slow dom0 like an interrupt storm).

Do you try suspend/resume on UP or MP domU? Using xl or xm?

--
Jean-Yves Migeon