Linux Virtualization - Apr 2017 - [PATCH net-next 3/3] virtio-net: clean tx descriptors from rx napi

On Sun, Apr 2, 2017 at 10:47 PM, Michael S. Tsirkin <mst at redhat.com>
wrote:
> On Sun, Apr 02, 2017 at 04:10:12PM -0400, Willem de Bruijn wrote:
>> From: Willem de Bruijn <willemb at google.com>
>>
>> Amortize the cost of virtual interrupts by doing both rx and tx work
>> on reception of a receive interrupt if tx napi is enabled. With
>> VIRTIO_F_EVENT_IDX, this suppresses most explicit tx completion
>> interrupts for bidirectional workloads.
>>
>> Signed-off-by: Willem de Bruijn <willemb at google.com>
>> ---
>>  drivers/net/virtio_net.c | 22 ++++++++++++++++++++++
>>  1 file changed, 22 insertions(+)
>>
>> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
>> index 95d938e82080..af830eb212bf 100644
>> --- a/drivers/net/virtio_net.c
>> +++ b/drivers/net/virtio_net.c
>> @@ -1030,12 +1030,34 @@ static int virtnet_receive(struct receive_queue
*rq, int budget)
>>       return received;
>>  }
>>
>> +static void free_old_xmit_skbs(struct send_queue *sq);
>> +
>
> Could you pls re-arrange code to avoid forward declarations?
Okay. I'll do the move in a separate patch to simplify review.
>> +static void virtnet_poll_cleantx(struct receive_queue *rq)
>> +{
>> +     struct virtnet_info *vi = rq->vq->vdev->priv;
>> +     unsigned int index = vq2rxq(rq->vq);
>> +     struct send_queue *sq = &vi->sq[index];
>> +     struct netdev_queue *txq = netdev_get_tx_queue(vi->dev,
index);
>> +
>> +     if (!sq->napi.weight)
>> +             return;
>> +
>> +     __netif_tx_lock(txq, smp_processor_id());
>> +     free_old_xmit_skbs(sq);
>> +     __netif_tx_unlock(txq);
>> +
>> +     if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
>> +             netif_wake_subqueue(vi->dev, vq2txq(sq->vq));
>> +}
>> +
>
> Looks very similar to virtnet_poll_tx.
>
> I think this might be waking the tx queue too early, so
> it will tend to stay almost full for long periods of time.
> Why not defer wakeup until queue is at least half empty?
I'll test that. Delaying wake-up longer than necessary can cause
queue build up at the qdisc and higher tail latency, I imagine. But
it may reduce the number of __netif_schedule calls.
> I wonder whether it's worth it to handle very short queues
> correctly - they previously made very slow progress,
> not they are never woken up.
>
> I'm a bit concerned about the cost of these wakeups
> and locking. I note that this wake is called basically
> every time queue is not full.
>
> Would it make sense to limit the amount of tx polling?
> Maybe use trylock to reduce the conflict with xmit?
Yes, that sounds good. I did test that previously and saw no
difference then. But when multiple cpus contend for a single
txq it should help.

On Mon, Apr 03, 2017 at 01:02:13AM -0400, Willem de Bruijn
wrote:
> On Sun, Apr 2, 2017 at 10:47 PM, Michael S. Tsirkin <mst at
redhat.com> wrote:
> > On Sun, Apr 02, 2017 at 04:10:12PM -0400, Willem de Bruijn wrote:
> >> From: Willem de Bruijn <willemb at google.com>
> >>
> >> Amortize the cost of virtual interrupts by doing both rx and tx
work
> >> on reception of a receive interrupt if tx napi is enabled. With
> >> VIRTIO_F_EVENT_IDX, this suppresses most explicit tx completion
> >> interrupts for bidirectional workloads.
> >>
> >> Signed-off-by: Willem de Bruijn <willemb at google.com>
This is a popular approach, but I think this will only work well if tx
and rx interrupts are processed on the same CPU and if tx queue is per
cpu.  If they target different CPUs or if tx queue is used from multiple
CPUs they will conflict on the shared locks.

This can even change dynamically as CPUs/queues are reconfigured.
How about adding a flag and skipping the tx poll if there's no match?

-- 
MST

On Fri, Apr 7, 2017 at 3:28 PM, Michael S. Tsirkin <mst at redhat.com>
wrote:
> On Mon, Apr 03, 2017 at 01:02:13AM -0400, Willem de Bruijn wrote:
>> On Sun, Apr 2, 2017 at 10:47 PM, Michael S. Tsirkin <mst at
redhat.com> wrote:
>> > On Sun, Apr 02, 2017 at 04:10:12PM -0400, Willem de Bruijn wrote:
>> >> From: Willem de Bruijn <willemb at google.com>
>> >>
>> >> Amortize the cost of virtual interrupts by doing both rx and
tx work
>> >> on reception of a receive interrupt if tx napi is enabled.
With
>> >> VIRTIO_F_EVENT_IDX, this suppresses most explicit tx
completion
>> >> interrupts for bidirectional workloads.
>> >>
>> >> Signed-off-by: Willem de Bruijn <willemb at google.com>
>
> This is a popular approach, but I think this will only work well if tx
> and rx interrupts are processed on the same CPU and if tx queue is per
> cpu.  If they target different CPUs or if tx queue is used from multiple
> CPUs they will conflict on the shared locks.
Yes. As a result of this discussion I started running a few vcpu affinity tests.

The data is not complete. In particular, I don't have the data yet to
compare having tx and rx irq on the same cpu (0,0) vs on different
(0,2) for this patchset. Which is the relevant data to your point.

Initial results for unmodified upstream driver at {1, 10, 100}x
TCP_STREAM, for irq cpu affinity (rx,tx). Process is always pinned to cpu
1. This is a 4 vcpu system pinned by the host to 4 cores on the same
socket. The previously reported results were obtained with txq, rtx
and process on different vcpus (0,2). Running all on the same vcpu
lower cycle count considerably:

irq 0,0
1    throughput_Mbps=29767.14  391,488,924,526      cycles
10  throughput_Mbps=40808.64  424,530,251,896      cycles
100 throughput_Mbps=33475.13  414,622,071,167      cycles

irq 0,2
1   throughput_Mbps=30176.05  395,673,200,747      cycles
10 throughput_Mbps=40729.26  433,948,374,991      cycles
100 throughput_Mbps=33758.68 436,291,949,393      cycles

irq 1,1
1    throughput_Mbps=26635.20 269,071,002,844      cycles
10  throughput_Mbps=42385.05 299,945,944,516      cycles
100 throughput_Mbps=33580.98 283,272,895,507      cycles

With this patch set applied, cpu (1,1)

1     throughput_Mbps=34980.76  276,504,805,414      cycles
10   throughput_Mbps=42519.92 298,105,889,785      cycles
100 throughput_Mbps=35268.86 296,670,598,712      cycles

I will need to get data for (0,2) vs (0,0).
> This can even change dynamically as CPUs/queues are reconfigured.
> How about adding a flag and skipping the tx poll if there's no match?
I suspect that even with the cache invalidations this optimization
will be an improvement over handling all tx interrupts in the tx napi
handler. I will get the datapoint for that.

That said, we can make this conditional. What flag exactly do you
propose? Compare raw_smp_processor_id() in the rx softint with one
previously stored in the napi tx callback?