Re: vhost + multiqueue + RSS question.

[Jason Wang <jasowang@redhat.com>]

On 11/17/2014 07:58 PM, Michael S. Tsirkin wrote:
> On Mon, Nov 17, 2014 at 01:22:07PM +0200, Gleb Natapov wrote:
>> > On Mon, Nov 17, 2014 at 12:38:16PM +0200, Michael S. Tsirkin wrote:
>>> > > On Mon, Nov 17, 2014 at 09:44:23AM +0200, Gleb Natapov wrote:
>>>> > > > On Sun, Nov 16, 2014 at 08:56:04PM +0200, Michael S. Tsirkin wrote:
>>>>> > > > > On Sun, Nov 16, 2014 at 06:18:18PM +0200, Gleb Natapov wrote:
>>>>>> > > > > > Hi Michael,
>>>>>> > > > > > 
>>>>>> > > > > >  I am playing with vhost multiqueue capability and have a question about
>>>>>> > > > > > vhost multiqueue and RSS (receive side steering). My setup has Mellanox
>>>>>> > > > > > ConnectX-3 NIC which supports multiqueue and RSS. Network related
>>>>>> > > > > > parameters for qemu are:
>>>>>> > > > > > 
>>>>>> > > > > >    -netdev tap,id=hn0,script=qemu-ifup.sh,vhost=on,queues=4
>>>>>> > > > > >    -device virtio-net-pci,netdev=hn0,id=nic1,mq=on,vectors=10
>>>>>> > > > > > 
>>>>>> > > > > > In a guest I ran "ethtool -L eth0 combined 4" to enable multiqueue.
>>>>>> > > > > > 
>>>>>> > > > > > I am running one tcp stream into the guest using iperf. Since there is
>>>>>> > > > > > only one tcp stream I expect it to be handled by one queue only but
>>>>>> > > > > > this seams to be not the case. ethtool -S on a host shows that the
>>>>>> > > > > > stream is handled by one queue in the NIC, just like I would expect,
>>>>>> > > > > > but in a guest all 4 virtio-input interrupt are incremented. Am I
>>>>>> > > > > > missing any configuration?
>>>>> > > > > 
>>>>> > > > > I don't see anything obviously wrong with what you describe.
>>>>> > > > > Maybe, somehow, same irqfd got bound to multiple MSI vectors?
>>>> > > > It does not look like this is what is happening judging by the way
>>>> > > > interrupts are distributed between queues. They are not distributed
>>>> > > > uniformly and often I see one queue gets most interrupt and others get
>>>> > > > much less and then it changes.
>>> > > 
>>> > > Weird. It would happen if you transmitted from multiple CPUs.
>>> > > You did pin iperf to a single CPU within guest, did you not?
>>> > > 
>> > No, I didn't because I didn't expect it to matter for input interrupts.
>> > When I run iperf on a host rx queue that receives all packets depends
>> > only on a connection itself, not on a cpu iperf is running on (I tested
>> > that).
> This really depends on the type of networking card you have
> on the host, and how it's configured.
>
> I think you will get something more closely resembling this
> behaviour if you enable RFS in host.
>
>> > When I pin iperf in a guest I do indeed see that all interrupts
>> > are arriving to the same irq vector. Is a number after virtio-input
>> > in /proc/interrupt any indication of a queue a packet arrived to (on
>> > a host I can use ethtool -S to check what queue receives packets, but
>> > unfortunately this does not work for virtio nic in a guest)?
> I think it is.
>
>> > Because if
>> > it is the way RSS works in virtio is not how it works on a host and not
>> > what I would expect after reading about RSS. The queue a packets arrives
>> > to should be calculated by hashing fields from a packet header only.
> Yes, what virtio has is not RSS - it's an accelerated RFS really.

Strictly speaking, not aRFS. aRFS requires a programmable filter and
needs driver to fill the filter on demand. For virtio-net, this is done
automatically in host side (tun/tap). There's no guest involvement.

>
> The point is to try and take application locality into account.
>

Yes, the locality was done through (consider a N vcpu guest with N queue):

- virtio-net driver will provide a default 1:1 mapping between vcpu and
txq through XPS
- virtio-net driver will suggest a default irq affinity hint also for a
1:1 mapping bettwen vcpu and txq/rxq

With all these, each vcpu get its private txq/rxq paris. And host side
implementation (tun/tap) will make sure if the packets of a flow were
received from queue N, if will also use queue N to transmit the packets
of this flow to guest.