[Qemu-devel] Re: [PATCH] rtl8139: IO memory is not part of vmstate

[Alex Williamson]

On Tue, 2010-12-14 at 14:32 +0200, Michael S. Tsirkin wrote:
> On Mon, Dec 13, 2010 at 10:00:48PM -0700, Alex Williamson wrote:
> > On Tue, 2010-12-14 at 06:43 +0200, Michael S. Tsirkin wrote:
> > > On Mon, Dec 13, 2010 at 12:15:08PM -0700, Alex Williamson wrote:
> > > > On Mon, 2010-12-13 at 21:06 +0200, Michael S. Tsirkin wrote:
> > > > > On Mon, Dec 13, 2010 at 11:59:16AM -0700, Alex Williamson wrote:
> > > > > > On Mon, 2010-12-13 at 20:54 +0200, Michael S. Tsirkin wrote:
> > > > > > > On Mon, Dec 13, 2010 at 11:00:44AM -0700, Alex Williamson wrote:
> > > > > > > > On Mon, 2010-12-13 at 19:50 +0200, Michael S. Tsirkin wrote:
> > > > > > > > > On Mon, Dec 13, 2010 at 10:43:22AM -0700, Alex Williamson 
> > > > > > > > > wrote:
> > > > > > > > > > So, unfortunately, I stand by my original patch.
> > > > > > > > > 
> > > > > > > > > What about the one that put -1 in saved index for a 
> > > > > > > > > hotplugged device?
> > > > > > > > 
> > > > > > > > There are still examples that don't work even without hotplug 
> > > > > > > > (example 2
> > > > > > > > and example 3 after the reboot).  That hack limits the damage, 
> > > > > > > > but still
> > > > > > > > leaves a latent bug for reboot and doesn't address the 
> > > > > > > > non-hotplug
> > > > > > > > scenarios.  So, I don't think it's worthwhile to pursue, and we
> > > > > > > > shouldn't pretend we can use it to avoid bumping the version_id.
> > > > > > > > Thanks,
> > > > > > > > 
> > > > > > > > Alex
> > > > > > > 
> > > > > > > I guess when we bump it we tell users: migration is completely
> > > > > > > borken to the old version, don't even try it.
> > > > > > > 
> > > > > > > Is there a way for libvirt to discover such incompatibilities
> > > > > > > and avoid the migration?
> > > > > > 
> > > > > > I don't know if libvirt has a way to query this in advance.  If a
> > > > > > migration is attempted, the target will report:
> > > > > > 
> > > > > > savevm: unsupported version 5 for '0000:00:03.0/rtl8139' v4
> > > > > > 
> > > > > > And the source will continue running.  We waste plenty of bits 
> > > > > > getting
> > > > > > to that point,
> > > > > 
> > > > > Yes, this happens after all of memory has been migrated.
> > > > 
> > > > Better late than never :^\
> > > 
> > > One other question: can we do the same by creating a new (empty)
> > > section? As was discussed in the past this is easier for
> > > downstreams to cherry-pick.
> > 
> > The only way I can think to do that would be to have a subsection that
> > is always included, but saves no data.  That would force a failure on
> > new->old migration, but I don't think it really matches the intended
> > purpose of subsections and feels like it's adding cruft for no gain.
> > Maybe I'm missing something.  Juan, is there any advantage to trapping
> > this in a subsection?  Thanks,
> > 
> > Alex
> 
> Maybe in this particular case the advantage is minimal.
> But it seems easier to stick to a rule of no more version
> bumps than argue about each case.

Do we have such a rule?  If we have a subsection who's needed function
is return 1, I think that's a good indication that it's not appropriate
for a subsection and the end result is equivalent to bumping the main
driver vmstate version.  It's convoluted to try to hide a one-way
upgrade in a subsection.  Thanks,

Alex