Re: [Qemu-devel] [PATCH] net/virtio: fix multi-queue negotiation

[Jason Wang <jasowang@redhat.com>]

On 06/26/2015 12:54 AM, Marcel Apfelbaum wrote:
> On 06/24/2015 11:00 AM, Jason Wang wrote:
>>
>>
>> On 06/19/2015 02:05 AM, Marcel Apfelbaum wrote:
>>> Clear host multi-queue related features if the peer
>>> doesn't support it.
>>>
>>> Signed-off-by: Marcel Apfelbaum <marcel@redhat.com>
>>> ---
>>> Notes:
>>>   This fixes a guest CPU soft lock, however the virtio-net
>>>   device will not work correctly. It seems that is
>>>   peer's "fault", not knowing how to handle the situation.
>>>   However, I submit this patch since it corrects the negotiation
>>>   and saves us from a guest crash (!).
>>
>> Could you please describe how to reproduce this issue?
> Hi Jason,
> Sorry for the late reply.
> I was hoping that a vhost-user multi-queue "insider" will ask
> questions :)
>
> This happens when we have OVS as backend without multi-queue support
> and qemu/guest with multi-queue-support.
>
>
>>
>>>
>>>   Any ideas from the virtio/multi-queue developers on how to debug
>>>   this further are welcomed.
>>>
>>>   hw/net/virtio-net.c | 13 ++++++++++++-
>>>   1 file changed, 12 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
>>> index 9281aa1..63e59e8 100644
>>> --- a/hw/net/virtio-net.c
>>> +++ b/hw/net/virtio-net.c
>>> @@ -367,6 +367,11 @@ static int peer_has_ufo(VirtIONet *n)
>>>       return n->has_ufo;
>>>   }
>>>
>>> +static int peer_has_multiqueue(VirtIONet *n)
>>> +{
>>> +    return n->multiqueue;
>>> +}
>>
>> The name is confusing, this is in fact whether or not guest support
>> multiqueue. To check peer's ability, you need check
>> n->nic_conf.peers.queues instead.
> I just wanted to mimic similar code, I have no issue against
> this approach.
>
>>
>>> +
>>>   static void virtio_net_set_mrg_rx_bufs(VirtIONet *n, int
>>> mergeable_rx_bufs,
>>>                                          int version_1)
>>>   {
>>> @@ -469,6 +474,13 @@ static uint64_t
>>> virtio_net_get_features(VirtIODevice *vdev, uint64_t features)
>>>           virtio_clear_feature(&features, VIRTIO_NET_F_HOST_UFO);
>>>       }
>>>
>>> +    if (!peer_has_multiqueue(n)) {
>>> +        virtio_clear_feature(&features, VIRTIO_NET_F_MQ);
>>
>> I'm not quite understand this, VIRTIO_NET_F_MQ should work if peer has
>> only 1 queue.
> Please explain, I thought  VIRTIO_NET_F_MQ flag *enables*
> multi-queue, and in our case we pass queue=2 to vhost.
>
>>
>>> +        virtio_clear_feature(&features, VIRTIO_NET_F_GUEST_ANNOUNCE);
>>> +        virtio_clear_feature(&features, VIRTIO_NET_F_CTRL_VQ);
>>> +        virtio_clear_feature(&features, VIRTIO_NET_F_CTRL_RX);
>>> +    }
>>> +
>>
>> Those features don't depend on multiqueue, why clear them?
> Once VIRTIO_NET_F_MQ is cleared, the virtio driver from guest is
> complaining
> about those other flags too.
>
> Any explanations about how *it should* work are welcomed.
> Again, the scenario is: backend doesn't support multi-queue,
> QEMU/guest do,
> and queues=2 is passed on command line.
>

In this case, backend should just fail the initialization like what
we've done for tap. I believe vhost-user should check whether or not its
backend can support up to 2 queues and fail if not.

Thanks

>
> Thanks,
> Marcel
>>
>>>       if (!get_vhost_net(nc->peer)) {
>>>           virtio_add_feature(&features, VIRTIO_F_VERSION_1);
>>>           return features;
>>> @@ -1314,7 +1326,6 @@ static void virtio_net_tx_bh(void *opaque)
>>>   static void virtio_net_set_multiqueue(VirtIONet *n, int multiqueue)
>>>   {
>>>       n->multiqueue = multiqueue;
>>> -
>>>       virtio_net_set_queues(n);
>>>   }
>>>
>>
>