Re: [PATCH 1/2] virtio: add a way to disable a queue

[Laurent Vivier <lvivier@redhat.com>]

On 09/08/2021 05:01, Jason Wang wrote:
> 
> 在 2021/8/6 下午3:27, Laurent Vivier 写道:
>> On 06/08/2021 08:25, Jason Wang wrote:
>>> 在 2021/8/2 下午4:42, Laurent Vivier 写道:
>>>> On 02/08/2021 06:50, Jason Wang wrote:
>>>>> 在 2021/7/30 上午3:19, Laurent Vivier 写道:
>>>>>> Add virtio_queue_disable()/virtio_queue_enable() to disable/enable a queue
>>>>>> by setting vring.num to 0 (or num_default).
>>>>>> This is needed to be able to disable a guest driver from the host side
>>>>> I suspect this won't work correclty for vhost.
>>>> With my test it seems to work with vhost too.
>>>
>>> So setting 0 will lead -EINVAL to be returned during VHOST_SET_VRING_NUM. I think qemu
>>> will warn the failure in this case.
>> I didn't see any error when I tried. I will check the code.
>>
>>> What's more important, it's not guaranteed to work for the case of vhost-user or
>>> vhost-vDPA.
>> Perhaps we can target only the vhost host case, as this is used for failover and usually
>> the virtio-net device is backed by a bridge on same network as the VFIO device?
> 
> 
> Probably not, it should be a general feature that can work for all types of virtio/vhost
> backends.
> 
> 
>>
>>>
>>>>> And I believe we should only do this after the per queue enabling/disabling is supported
>>>>> by the spec.
>>>>>
>>>>> (only MMIO support that AFAIK)
>>>> I don't want to modify the spec.
>>>>
>>>> I need something that works without modifying existing (old) drivers.
>>>>
>>>> The idea is to be able to disable the virtio-net kernel driver from QEMU if the driver is
>>>> too old (i.e. it doesn't support STANDBY feature).
>>>>
>>>> Setting vring.num to 0 forces the kernel driver to exit on error in the probe function.
>>>> It's what I want: the device is present but disabled (the driver is not loaded).
>>>>
>>>> Any other suggestion?
>>>
>>> I think we should probably disable the device instead of doing it per virtqueue.
>>>
>> I tried to use virtio_set_disabled() but it doesn't work.
>> Perhaps it's too late when I call the function (I need to do that in
>> virtio_net_set_features()). What I want is to prevent the load of the driver in the guest
>> kernel to hide the virtio-net device. Setting vring.num to 0 triggers an error in the
>> driver probe function and prevents the load of the driver.
> 
> 
> How about fail the validate_features() in this case?

It's a good suggestion and it seems to work.

I'm going to send an updated patch.

Thanks,
Laurent