Re: [PATCH] hw/char/pl011: Fix clock migration failure

[Andrew Jones]

On Wed, Mar 17, 2021 at 11:14:56AM +0000, Peter Maydell wrote:
> On Wed, 17 Mar 2021 at 10:59, Gavin Shan <gshan@redhat.com> wrote:
> >
> > Hi Peter,
> >
> > On 3/17/21 9:40 PM, Peter Maydell wrote:
> > > On Wed, 17 Mar 2021 at 10:37, Gavin Shan <gshan@redhat.com> wrote:
> > >> On 3/17/21 8:09 PM, Peter Maydell wrote:
> > >>> On Wed, 17 Mar 2021 at 04:44, Gavin Shan <gshan@redhat.com> wrote:
> > >>>>
> > >>>>    static const VMStateDescription vmstate_pl011 = {
> > >>>>        .name = "pl011",
> > >>>>        .version_id = 2,
> > >>>>        .minimum_version_id = 2,
> > >>>> +    .post_load = pl011_post_load,
> > >>>>        .fields = (VMStateField[]) {
> > >>>>            VMSTATE_UINT32(readbuff, PL011State),
> > >>>>            VMSTATE_UINT32(flags, PL011State),
> > >>>> @@ -355,10 +355,6 @@ static const VMStateDescription vmstate_pl011 = {
> > >>>>            VMSTATE_INT32(read_trigger, PL011State),
> > >>>>            VMSTATE_END_OF_LIST()
> > >>>>        },
> > >>>> -    .subsections = (const VMStateDescription * []) {
> > >>>> -        &vmstate_pl011_clock,
> > >>>> -        NULL
> > >>>> -    }
> > >>>>    };
> > >>>
> > >>> Doesn't dropping the subsection break migration compat ?
> > >>>
> > >>
> > >> It's why this patch needs to be backported to stable branches.
> > >> In that way, we won't have migration compatible issue.
> > >
> > > No, migration has to work from the existing already
> > > shipped 5.1, 5.2, etc releases to 6.0 (assuming you use
> > > the correct "virt-5.2" &c versioned machine type.)
> > >
> >
> > Commit aac63e0e6ea3 ("hw/char/pl011: add a clock input") is merged
> > to v5.2.0. The migration failure happens during migration from v6.0
> > to v5.1 with machine type as "virt-5.1", instead of migrating from
> > v5.1 to v6.0. One question is if we need support backwards migration?
> 
> Upstream doesn't care about backwards migration. AIUI
> RedHat as a downstream care about the backwards-migration
> case in some specific situations, but I don't know if that
> would include this one.

Right, we do prefer to be able to support "ping-pong" migrations. For
example, if we start a virt-5.1 machine on a 5.1 build of QEMU, and then
migrate it to a 5.2 build of QEMU, we'd like to also be able to go back
to the 5.1 build.

I agree this patch is not the right approach. I think the right approach
is to introduce a compat property and make the "new" section dependent
on it. And then update the hw_compat_* arrays. Gavin, please take a look
at "Connecting subsections to properties" of docs/devel/migration.rst.

I'm also curious what the state of mach-virt's machine types are for
migration. It'd be nice to exhaustively test both forward migration of
all machine types and ping-pong migrations of all machine types. We can
then consider each issue we find (the pessimist in me suggests we'll find
more than this pl011 issue) and how/if we want to resolve them.

Thanks,
drew