Re: [libvirt] [PATCH 3/3] qemu_domain: use correct default USB controller on ppc64

On Tue, Mar 07, 2017 at 03:32:37PM +0100, Pavel Hrdina wrote: > On Tue, Mar 07, 2017 at 02:55:43PM +0100, Andrea Bolognani wrote: > > On Tue, 2017-03-07 at 09:19 +0100, Pavel Hrdina wrote: > > > > However, after migration is complete, the <controller> > > > > element has model='nec-xhci' instead of model='pci-ohci', > > > > which means that power cycling the guest results in > > > > breaking the guest ABI. > > >  > > > I'm not so sure that this is an ABI change.  The guest ABI is to ensure > > > that the same guest XML will always start the same QEMU guest.  However > > > the PERSISTENT migration can make ABI changes because it is the same as > > > virsh dumpxml $domain > $domain.xml && copy the XML onto remote host > > > and virsh define $domain.xml.  This would also change the *model*. > > >  > > > If this would be considered to be guest ABI stable it would mean that > > > other changes done by using this flag would be wrong because they also > > > modifies the persistent XML during migration. > > > > I assume there are very good reasons for persistent > > migration to behave differently, but as a user I find > > it extremely surprising. Do you have any insight on the > > rationale behind allowing ABI changes when performing > > persistent migration? > > Well it's a different kind of migration which is almost similar to the > virsh dumpxml --inactive && virsh define, with one more feature, migration > hook that can update the XML before it's defined on the destination. > > I agree that it may seem that we should handle this type of migration to > be guest ABI stable, on the other hand this basically defines new domain > on the destination where we allow to do the ABI updates. > > I'm adding Dan to CC to get his opinion about the guest ABI. So, there's no particular reason why things must be different for live vs persistent migration. The only fundamental requirement is that we must never change ABI underneath a running guest, which mandates the strong ABI guarantee for live migration & save/restore. For persistent/offline migration, there's no running guest, so from a technical POV we have more freedom in some sense. That said, a change in guest hardaware that can cause a guest to become unbootable is not a nice thing todo. For this reason, when you use 'defineXML' libvirt aims to immediately fill in (most) XML elements that would affect guest ABI. ie we assign addresses at define time, not start time, and we expand non-versioned machine types into versioned machine types, etc. So, yes, we should try to *not* change ABI across a dumpxml & define operational. When we first introduced PCI device addressing in libvirt, we had an old state where XML contained no addresses, and upon starting in new libvirt we have to assign addresses which might not be the same as those seen with previous libvirt before addressing was recorded. Once the addresses are assigned though, we must never change them.

So... - if we recorded 'pci-ohci' in the XML, we must never change that to 'nec-xhci'.

- if we didn't record 'pci-ochi' in the XML, we should fill in the model that is most likely to match what the previous default behaviour was, which would be 'pci-ochi' not 'nec-xhci' IIUC.