Re: [PATCH v2] vmstate: fix subsection load name check

[Peter Xu <peterx@redhat.com>]

On Tue, Mar 10, 2026 at 01:39:04PM +0000, Peter Maydell wrote:
> On Tue, 10 Mar 2026 at 13:28, Alexandr Moshkov
> <dtalexundeer@yandex-team.ru> wrote:
> > Hi, this breaks migration for s390x and ppc64:
> >
> > qemu-system-s390x: Missing section footer for qemu-s390-flic
> >
> > Thanks for reply! It happens because "qemu-s390-flic" vmsd has "qemu-s390-flic-full" subsection. I'm not sure if we can change the names of the existing subsections. Peter, what do you think?
> >
> > It seems to be a bigger problem that the migration framework does not document the names for subsections in any way. At the same time, the code implies that at least the names should be a substring with its parent. There is even such a code comment `subsection name has to be "section_name/a"`, that seems to imply the presence of /.
> >
> > Also I tried to put together a list of devices whose separator in subsections is not /:
> 
> > or-irq
> 
> This one and probably some of the others are my mistake, I think.
> I think this is because:
>  * the migration code imposes a constraint on the subsection names
>  * the migration documentation does not mention this constraint
>  * the migration code does not effectively enforce this constraint
>    (e.g. by asserting when the vmstate with a bad name is registered)
> 
> My assumption when writing that code was very likely that the
> name of the subsection didn't have to have any relation to
> the name of its parent subsection (after all, the migration
> code knows it is a subsection, so if it needed to have the
> name on the wire be "parentname/subsectionname" it could construct
> that itself), and since it all just worked I never noticed the
> mistake...

Indeed, we can do better on the 2nd/3rd bullets.. so we should document it
in struct VMStateDescription definition, and enforce it when registering
new VMSDs.

One thing to double check with Alexandr on this one:

> They using `_`, `-` or `.` as separator

Does it mean we also can't simply whitelist all these characters (including
"/"), because it won't always work?

For example in your case, it's "virtio-blk" being the parent VMSD, "virtio"
being the child VMSD, followed with a subsection belongs to virtio-blk.
We want to make that subsection be recognizable as virtio-blk's.

Then if we treat "-" also to be a separator, then "virtio" will still
mis-recognize virtio-blk's as its own (because its name is "virtio-blk/..."
hence it also satisfies the "virtio" check)?

Thanks,

-- 
Peter Xu