On Mon, Sep 11, 2017 at 06:19:54PM +0100, Dr. David Alan Gilbert wrote: > * Mark Cave-Ayland ([email protected]) wrote: > > On 11/09/17 11:48, David Gibson wrote: > > > > > On Mon, Sep 11, 2017 at 10:30:33AM +0100, Dr. David Alan Gilbert wrote: > > >> * Greg Kurz ([email protected]) wrote: > > >>> On Sun, 10 Sep 2017 15:37:33 +0100 > > >>> Mark Cave-Ayland <[email protected]> wrote: > > >>> > > >>>> Commit a90db15 "target-ppc: Convert ppc cpu savevm to > > >>>> VMStateDescription" > > >>>> appears to drop the internal CPU IRQ state from the migration stream. > > >>>> Whilst > > >>>> testing migration on g3beige/mac99 machines, test images would > > >>>> randomly fail to > > >>>> resume unless a key was pressed on the VGA console. > > >>>> > > >>>> Further investigation suggests that internal CPU IRQ state isn't being > > >>>> preserved and so interrupts asserted at the time of migration are > > >>>> lost. Adding > > >>>> the pending_interrupts and irq_input_state fields back into the > > >>>> migration > > >>>> stream appears to fix the problem here during local tests. > > >>>> > > >>>> As part of this commit we bump the vmstate_ppc version from 5 to 6 to > > >>>> handle > > >>>> the additional fields. > > >>>> > > >>> > > >>> And so this unconditionally breaks backward migration... what about > > >>> adding > > >>> a subsection for this ? > > >> > > >> and wiring it to a flag on the machine type so that older machine types > > >> don't send it. > > > > > > Right, a subsection is certainly necessary to avoid breaking backwards > > > migration. > > > > The suggestion of using the VMSTATE_*_V macros with an increased version > > number came from Alexey's original review of the patch many months ago > > which is why I did it that way. > > > > Out of curiosity though, what is the criteria for supporting backwards > > migration? Obviously forward migration is supported as-is in this > > manner, so what determines if a patch needs to be backwards compatible > > and how far? > > It's a bit fuzzy. Downstream we do backwards migration between various > versions - and those versions are pretty arbitrarily chosen. > Generally I prefer if we don't break that upstream either, although > it isn't tested much.
Right. In this case not breaking backwards migration upstream is
essentially a favour to downstream, since unbreaking it downstream
once its broken upstream is a real PITA.
> If it was in code that was specific to your g3beige I wouldn't mind;
> but for ppc in general then if it breaks the server migration it'll
> be a pain we'd have to then fix. Best to keep it working upstream.
Right.
> But it's fairly easy to put new fields in a subsection and tie it
> to a property; that makes it easy to switch it on/off in machine
> types.
In this case I'm not sure we even need a property - I think we could
migrate it only when it's non-zero. That shold only break it in cases
where actually it would already break.
> > > But apart from that I want to understand better exactly why this is
> > > necessary. What's the state that's being lost, and is it really not
> > > recoverable from anywhere else.
> >
> > The test case I have is installing Darwin PPC 6.02 with qemu-system-ppc
> > TCG and repeatedly pausing, executing "savevm foo", then quitting and
> > continuing with "-loadvm foo" during the installation phase. About 1 in
> > 10 times the installer hangs after the loadvm until I press a key, at
> > which point it carries on as normal.
> >
> > I then proceeded to going backwards through the git history until I
> > found out that it was the removal of the pending_interrupts,
> > irq_input_state and access_type fields during the conversion to
> > VMStateDescription commit a90db15 which seemed to cause the problem.
> >
> > > The other thing that concerns me is how we're encoding the
> > > information. These are essentially internal fields, not reflecting
> > > something with an architected encoding - adding those to the migration
> > > stream is often a bad idea - it inhibits our ability to rework
> > > internal encodings.
> >
> > I'm not sure how this should be managed, however there was a similar
> > issue with excp_prefix (which was also removed in a90db15) that was
> > fixed in 2360b6e by calling a helper in cpu_post_load(). I can't easily
> > see how could work with env.pending_interrupts and env.irq_input_state
> > though.
>
> Without knowing anything about this hardware... generally the migration
> stream should reflect the real state of the system rather than internal
> implementation detail, that way if you change the implementation you
> don't need to fudge the state. Having said that, there's generally
> some internal state that's perhaps not immediately obvious from specs
> until you try and implement it.
Right. I'm not really sure how to handle this yet. The CPU irq
numbers are pretty much arbitrarily assigned, I don't think much
thought has gone into them. And if its going to become part of the
migration ABI, some thought needs to be put into it.
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature
