Re: [PATCH net-next] macvtap/macvlan: use IFF_NO_QUEUE

[Jason Wang <jasowang@redhat.com>]

On 08/28/2015 08:25 PM, Vlad Yasevich wrote:
> On 08/27/2015 10:42 PM, Jason Wang wrote:
>> > 
>> > 
>> > On 08/27/2015 06:43 PM, Michael S. Tsirkin wrote:
>>> >> On Wed, Aug 26, 2015 at 01:45:30PM +0800, Jason Wang wrote:
>>>> >>>
>>>> >>> On 08/26/2015 12:32 AM, Vlad Yasevich wrote:
>>>>> >>>> On 08/25/2015 07:30 AM, Jason Wang wrote:
>>>>>> >>>>> On 08/25/2015 06:17 PM, Michael S. Tsirkin wrote:
>>>>>>> >>>>>> On Mon, Aug 24, 2015 at 04:33:12PM +0800, Jason Wang wrote:
>>>>>>>>> >>>>>>>> For macvlan, switch to use IFF_NO_QUEUE instead of tx_queue_len = 0.
>>>>>>>>> >>>>>>>>
>>>>>>>>> >>>>>>>> For macvtap, after commit 6acf54f1cf0a6747bac9fea26f34cfc5a9029523
>>>>>>>>> >>>>>>>> ("macvtap: Add support of packet capture on macvtap
>>>>>>>>> >>>>>>>> device."). Multiqueue macvtap suffers from single qdisc lock
>>>>>>>>> >>>>>>>> contention. This is because macvtap claims a non zero tx_queue_len and
>>>>>>>>> >>>>>>>> it reuses this value as it socket receive queue size.Thanks to
>>>>>>>>> >>>>>>>> IFF_NO_QUEUE, we can remove the lock contention without breaking
>>>>>>>>> >>>>>>>> existing socket receive queue length logic.
>>>>>>>>> >>>>>>>>
>>>>>>>>> >>>>>>>> Cc: Patrick McHardy <kaber@trash.net>
>>>>>>>>> >>>>>>>> Cc: Vladislav Yasevich <vyasevic@redhat.com>
>>>>>>>>> >>>>>>>> Cc: Michael S. Tsirkin <mst@redhat.com>
>>>>>>>>> >>>>>>>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>>>>>>> >>>>>> Seems to make sense. Give me a day or two to get over the jet lag
>>>>>>> >>>>>> (and get out from under the pile of mail accumulated while I was traveling),
>>>>>>> >>>>>> I'll review properly and ack.
>>>>>>> >>>>>>
>>>>>> >>>>> A note on this patch: only default qdisc were removed but we don't lose
>>>>>> >>>>> the ability to attach a qdisc to macvtap (though it may cause lock
>>>>>> >>>>> contention on multiqueue case).
>>>>>> >>>>>
>>>>> >>>> Wouldn't that lock contention be solved if we really had multiple queues
>>>>> >>>> for multi-queue macvtaps?
>>>>> >>>>
>>>>> >>>> -vlad
>>>> >>> Yes, but this introduce another layer of txq locks contention?
>>> >> I don't follow - why does it? Could you clarify please?
>> > 
>> > I believe Vlad wants to remove NETIF_F_LLTX. If yes, core will do an
>> > extra tx lock at macvlan layer.
> No, I don't want to remove it.  In a sense, it would function similar to
> how it works when fwd_priv is populated.  I am still testing the code
> as it's showing some strange artifacts...  could be due to keeping LLTX.
>
> -vlad
>

I see. I'm ok to wait for your code. But if a patch of just two lines
works, probably no need to try complex method.

Thanks