On 9/30/26 9:06 AM, Akihiko Odaki wrote:
> On 2026/09/30 4:22, Peter Xu wrote:
>> On Tue, Sep 29, 2026 at 03:16:39PM -0400, Peter Xu wrote:
>>> On Tue, Sep 29, 2026 at 07:34:50PM +0100, David Woodhouse wrote:
>>>> On Tue, 2026-09-29 at 14:22 -0400, Peter Xu wrote:
>>>>> On Tue, Sep 29, 2026 at 08:23:23PM +0300, Andrey Drobyshev wrote:
>>>>>> The snapshot RAM-saving path does not support Xen's memory mappings.
>>>>>> Refuse 'savevm' and 'loadvm' when running as a Xen device model, the
>>>>>> same way they are refused for CPR migration modes.
>>>>>>
>>>>>> Suggested-by: Akihiko Odaki <[email protected]>
>>>>>> Signed-off-by: Andrey Drobyshev <[email protected]>
>>>>>
>>>>> Acked-by: Peter Xu <[email protected]>
>>>>>
>>>>> Cc: David Woodhouse <[email protected]>
>>>>> Cc: Paul Durrant <[email protected]>
>>>>
>>>> Hm, but surely if QEMU is running as a device model for a Xen domain,
>>>> all it should be saving and restoring is that device model. It's then
>>>> up to Xen and its toolstack to migrate the domain and its memory.
>>>>
>>>> We've had guest-transparent live migration working on Xen, as well as
>>>> in-place update of QEMU on the same host. Surely all of those require
>>>> save/load support?
>>>
>>> Looks like I was wise copying more people..
>>>
>>> I knew Xen had qmp_xen_save_devices_state(), I was thinking that is the
>>> save support wiring together with Xen's facilities.
>>>
>>> I confess I never know what is the load support, and I was taking it for
>>> granted RAM was managed elsewhere, so load perhaps the same.
>>>
>>> So I suppose this is still a question to the author, or maybe Akihiko?
>>
>> Ok, I just noticed xen also has @xen-load-devices-state, so maybe this is
>> the answer. I suppose Xen should be using those, rather than the default
>> savevm/loadvm.
>>
>> Note, that what this patch changes (migrate_can_snapshot()) only affects
>> savevm/loadvm command.
> That aligns with my understanding as well, so:
>
> Reviewed-by: Akihiko Odaki <[email protected]>
>
> Regards,
> Akihiko Odaki
I don't know much about Xen, but AFAICT Xen is indeed using
qmp_xen_{save,load}_devices_state for its needs, and therefore they
should not normally reach any patch calling migrate_can_snapshot(). We
just make it explicit by adding this guard. So I agree with Peter and
Akihiko here.
Let me know if this requires changes after all.
Thanks,
Andrey
