lkml.org 
[lkml]   [2015]   [Feb]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH RFC v5 net-next 2/6] virtio_ring: try to disable event index callbacks in virtqueue_disable_cb()


On Tue, Feb 10, 2015 at 6:24 PM, Michael S. Tsirkin <mst@redhat.com>
wrote:
> On Mon, Feb 09, 2015 at 03:39:21AM -0500, Jason Wang wrote:
>> Currently, we do nothing to prevent the callbacks in
>> virtqueue_disable_cb() when event index is used. This may cause
>> spurious interrupts which may damage the performance. This patch
>> tries
>> to publish avail event as the used even to prevent the callbacks.
>>
>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>
> I'm surprised that this ever happens though.
> Normally we call this after getting an interrupt, and
> interrupts won't trigger again until the rings wraps around.

It was used to disable tx interrupt during start_xmit().
>
> When I tested this, touching an extra cache line was more
> expensive.
>
> Does this really reduce number of interrupts?
> Could you pls share some numbers with and without this patch?

Yes. It does reduce, see the following test results on multiple
sessions of TCP_RR

Datacopy:
/size/sessions/+-throughput%/+-cpu%/+-per_cpu_throughput%/+-tx_interrupts%/
/1/50/+0.1/+0.7/-0.6/-72.3/
/64/50/+1.7/+1.0/+0.7/-72.2/
/256/50/+0.0/-0.8/+0.9/-72.4/

Zerocopy:
/size/sessions/+-throughput%/+-cpu%/+-per_cpu_throughput%/+-tx_interrupts%/
/1/50/+3.2/-1.3/+4.6/-71.7/
/64/50/+1.9/-0.2/+2.1/-70.8/
/256/50/+4.4/+0.6/+3.9/-84.2/

No obvious changes for stream rx and tx.
>
>
>
>> ---
>> drivers/virtio/virtio_ring.c | 2 ++
>> 1 file changed, 2 insertions(+)
>>
>> diff --git a/drivers/virtio/virtio_ring.c
>> b/drivers/virtio/virtio_ring.c
>> index 545fed5..e9ffbfb 100644
>> --- a/drivers/virtio/virtio_ring.c
>> +++ b/drivers/virtio/virtio_ring.c
>> @@ -539,6 +539,8 @@ void virtqueue_disable_cb(struct virtqueue *_vq)
>> struct vring_virtqueue *vq = to_vvq(_vq);
>>
>> vq->vring.avail->flags |= cpu_to_virtio16(_vq->vdev,
>> VRING_AVAIL_F_NO_INTERRUPT);
>> + vring_used_event(&vq->vring) = cpu_to_virtio16(_vq->vdev,
>> + vq->vring.avail->idx);
>
> Hmm in fact, can't this actually cause an extra interrupt
> when avail->idx is completed?

We need to try best to avoid any interrupt after this and
virtqueue_disable_cb() were always need virtqueue_enable_cb() or
virtqueue_enable_cb_delayed() afterwards. Those two function will check
the pending used buffers and set a proper used event.
>
>
> I think that if you really can show disabling interrupts like this
> helps, you should
> set some invalid value like 0xfffff, or move it back to
> vq->vring.avail->idx - 1.
> No?

I tested avail->idx + vq.num but not obvious changes in the performance
compared to just use avail->idx here. So 0xffff probably won't help
much. Since we call virtqueue_enable_cb() or
virtqueue_disable_cb_delayed() afterwards, it doesn't matter that
whether avail->idx or avail->idx - 1 is used.
>
>
>
>
>> }
>> EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
>>
>> --
>> 1.8.3.1
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html



\
 
 \ /
  Last update: 2015-02-11 07:01    [W:0.050 / U:2.644 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site