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

[Laurent Vivier]

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?

> 
> 
>>
>>> 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.

Thanks,
Laurent