Re: [PATCH v2] hw/display/virtio-gpu: Reject non-migratable large blobs

On 2026/07/28 16:04, Marc-André Lureau wrote:
> Hi
> 
> On Tue, Jul 28, 2026 at 10:26 AM Akihiko Odaki
> <[email protected]> wrote:
>>
>> On 2026/07/27 18:30, Daniel P. Berrangé wrote:
>>> On Mon, Jul 27, 2026 at 06:23:25PM +0900, Akihiko Odaki wrote:
>>>> The blob size is encoded as a 64-bit integer in the virtio protocol, but
>>>> it is encoded as a 32-bit integer in the migration stream, and
>>>> a large blob whose size cannot be expressed in a 32-bit integer will
>>>> be corrupted after migration. Deny creating such large blobs.
>>>
>>> Is there a valid use case for such large blobs ?   If so, rather than
>>> rejecting the blob, we could dynamically add a migration  blocker.
>>> That would allow the valid use cases, while preventing the migration
>>> problem.
>>
>> For the QEMU code path changed here, there should be no practical use
>> case for a blob larger than 4 GiB. The non-virgl implementation uses
>> blobs as scanout or cursor backing. Such a size is excessive for those uses.
>>
>> A dynamic migration blocker would also let the guest make an otherwise
>> migratable VM non-migratable at runtime merely by creating such a
>> resource. This could interfere with host-level VM management. So I think
>> rejecting sizes that the current migration format cannot represent is
>> better for now.
> 
> What's the difference? The dynamic blocker would simply happen earlier
> or be queryable, no?

The difference I had in mind was that the guest could make migration 
unavailable after the VM had started in an unexpected way for the user 
who manages the host.

But I now remember that, if QEMU is started with -only-migratable, 
migrate_add_blocker() fails, so resource creation can be rejected in 
that configuration. It also fails while a migration or snapshot is 
already in progress.

Now I feel safer with the dynamic blocker. I will change the patch to 
keep a migration blocker installed while any such blob exists.

Regards,
Akihiko Odaki

> 
>> If a concrete use case emerges, extending the migration format would be
>> preferable to making migratability depend on guest-controlled state.
>>
>> Regards,
>> Akihiko Odaki
>>
>>>
>>>>
>>>> Fixes: f66767f75c9c ("virtio-gpu: add virtio-gpu/blob vmstate subsection")
>>>> Fixes: 10b9ddbc83b9 ("Revert "virtio-gpu: block migration of VMs with blob=true"")
>>>> Signed-off-by: Akihiko Odaki <[email protected]>
>>>> ---
>>>> Changes in v2:
>>>> - Add the Fixes: tag for f66767f75c9c.
>>>> - Add a TODO to encode blob_size as 64 bits in virtio-gpu/blob
>>>>     VMState version 2.
>>>> - Link to v1: https://lore.kernel.org/qemu-devel/[email protected]
>>>> ---
>>>>    hw/display/virtio-gpu.c | 15 +++++++++++++++
>>>>    1 file changed, 15 insertions(+)
>>>>
>>>> diff --git a/hw/display/virtio-gpu.c b/hw/display/virtio-gpu.c
>>>> index 718ba3039290..987b1b3ea860 100644
>>>> --- a/hw/display/virtio-gpu.c
>>>> +++ b/hw/display/virtio-gpu.c
>>>> @@ -352,6 +352,14 @@ static void virtio_gpu_resource_create_blob(VirtIOGPU *g,
>>>>            return;
>>>>        }
>>>>
>>>> +    if (cblob.size > UINT32_MAX) {
>>>> +        qemu_log_mask(LOG_GUEST_ERROR,
>>>> +                      "%s: blob too large (%" PRIu64 "> %" PRIu32 ")\n",
>>>> +                      __func__, cblob.size, UINT32_MAX);
>>>> +        cmd->error = VIRTIO_GPU_RESP_ERR_OUT_OF_MEMORY;
>>>> +        return;
>>>> +    }
>>>> +
>>>>        if (virtio_gpu_find_resource(g, cblob.resource_id)) {
>>>>            qemu_log_mask(LOG_GUEST_ERROR, "%s: resource already exists %d\n",
>>>>                          __func__, cblob.resource_id);
>>>> @@ -1435,7 +1443,14 @@ static int virtio_gpu_blob_save(QEMUFile *f, void *opaque, size_t size,
>>>>            }
>>>>            assert(!res->image);
>>>>            qemu_put_be32(f, res->resource_id);
>>>> +
>>>> +        /*
>>>> +         * The migration stream encodes blob_size as a 32-bit integer for
>>>> +         * compatibility though virtio encodes it as a 64-bit integer. A newer
>>>> +         * version of the migration stream should encode it as a 64-bit integer.
>>>> +         */
>>>>            qemu_put_be32(f, res->blob_size);
>>>> +
>>>>            qemu_put_be32(f, res->iov_cnt);
>>>>            for (i = 0; i < res->iov_cnt; i++) {
>>>>                qemu_put_be64(f, res->addrs[i]);
>>>>
>>>> ---
>>>> base-commit: 006a22cb26998998385b104db1ff9466ef2f3153
>>>> change-id: 20260725-blob-8a6a1a85601a
>>>>
>>>> Best regards,
>>>> --
>>>> Akihiko Odaki <[email protected]>
>>>>
>>>>
>>>
>>> With regards,
>>> Daniel
>>
>>
> 
>