Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1408803 > unrolled thread

[PATCH V2 0/2] vhost_net polling optimization

Back to article view | Back to linux.kernel

Contents

  [PATCH V2 0/2] vhost_net polling optimization Jason Wang <jasowang@redhat.com> - 2016-05-30 08:50 +0200
    [PATCH V2 2/2] vhost_net: conditionally enable tx polling Jason Wang <jasowang@redhat.com> - 2016-05-30 08:50 +0200
      Re: [PATCH V2 2/2] vhost_net: conditionally enable tx polling "Michael S. Tsirkin" <mst@redhat.com> - 2016-05-30 18:00 +0200
        Re: [PATCH V2 2/2] vhost_net: conditionally enable tx polling Jason Wang <jasowang@redhat.com> - 2016-05-31 05:30 +0200

#1408803 — [PATCH V2 0/2] vhost_net polling optimization

Hi:

This series tries to optimize vhost_net polling at two points:

- Stop rx polling for reduicng the unnecessary wakeups during
  handle_rx().
- Conditonally enable tx polling for reducing the unnecessary
  traversing and spinlock touching.

Test shows about 17% improvement on rx pps.

Please review

Changes from V1:
- use vhost_net_disable_vq()/vhost_net_enable_vq() instead of open
  coding.
- Add a new patch for conditionally enable tx polling.

Jason Wang (2):
  vhost_net: stop polling socket during rx processing
  vhost_net: conditionally enable tx polling

 drivers/vhost/net.c | 59 +++++++++++++++++++++++++++++------------------------
 1 file changed, 32 insertions(+), 27 deletions(-)

-- 
1.8.3.1

[toc] | [next] | [standalone]

#1408804 — [PATCH V2 2/2] vhost_net: conditionally enable tx polling

We always poll tx for socket, this is sub optimal since:

- it will be only used when we exceed the sndbuf of the socket.
- since we use two independent polls for tx and vq, this will slightly
  increase the waitqueue traversing time and more important, vhost
  could not benefit from commit
  9e641bdcfa4ef4d6e2fbaa59c1be0ad5d1551fd5 ("net-tun: restructure
  tun_do_read for better sleep/wakeup efficiency") even if we've
  stopped rx polling during handle_rx since tx poll were still left in
  the waitqueue.

Fix this by conditionally enable tx polling only when -EAGAIN were
met.

Test shows about 8% improvement on guest rx pps.

Before: ~1350000
After:  ~1460000

Signed-off-by: Jason Wang <jasowang@redhat.com>
---
 drivers/vhost/net.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index e91603b..5a05fa0 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c
@@ -378,6 +378,7 @@ static void handle_tx(struct vhost_net *net)
 		goto out;
 
 	vhost_disable_notify(&net->dev, vq);
+	vhost_net_disable_vq(net, vq);
 
 	hdr_size = nvq->vhost_hlen;
 	zcopy = nvq->ubufs;
@@ -459,6 +460,8 @@ static void handle_tx(struct vhost_net *net)
 					% UIO_MAXIOV;
 			}
 			vhost_discard_vq_desc(vq, 1);
+			if (err == -EAGAIN)
+				vhost_net_enable_vq(net, vq);
 			break;
 		}
 		if (err != len)
-- 
1.8.3.1

[toc] | [prev] | [next] | [standalone]

#1409251 — Re: [PATCH V2 2/2] vhost_net: conditionally enable tx polling

On Mon, May 30, 2016 at 02:47:54AM -0400, Jason Wang wrote:
> We always poll tx for socket, this is sub optimal since:
> 
> - it will be only used when we exceed the sndbuf of the socket.
> - since we use two independent polls for tx and vq, this will slightly
>   increase the waitqueue traversing time and more important, vhost
>   could not benefit from commit
>   9e641bdcfa4ef4d6e2fbaa59c1be0ad5d1551fd5 ("net-tun: restructure
>   tun_do_read for better sleep/wakeup efficiency") even if we've
>   stopped rx polling during handle_rx since tx poll were still left in
>   the waitqueue.

Why is this an issue?
sock_def_write_space only wakes up when queue is half empty,
not on each packet.
        if ((atomic_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf)

I suspect the issue is with your previous patch,
it now pokes at the spinlock on data path
where it used not to.

Is that right?


> 
> Fix this by conditionally enable tx polling only when -EAGAIN were
> met.
> 
> Test shows about 8% improvement on guest rx pps.
> 
> Before: ~1350000
> After:  ~1460000
> 
> Signed-off-by: Jason Wang <jasowang@redhat.com>
> ---
>  drivers/vhost/net.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> index e91603b..5a05fa0 100644
> --- a/drivers/vhost/net.c
> +++ b/drivers/vhost/net.c
> @@ -378,6 +378,7 @@ static void handle_tx(struct vhost_net *net)
>  		goto out;
>  
>  	vhost_disable_notify(&net->dev, vq);
> +	vhost_net_disable_vq(net, vq);
>  
>  	hdr_size = nvq->vhost_hlen;
>  	zcopy = nvq->ubufs;
> @@ -459,6 +460,8 @@ static void handle_tx(struct vhost_net *net)
>  					% UIO_MAXIOV;
>  			}
>  			vhost_discard_vq_desc(vq, 1);
> +			if (err == -EAGAIN)
> +				vhost_net_enable_vq(net, vq);
>  			break;
>  		}
>  		if (err != len)
> -- 
> 1.8.3.1

[toc] | [prev] | [next] | [standalone]

#1409793 — Re: [PATCH V2 2/2] vhost_net: conditionally enable tx polling

On 2016年05月30日 23:55, Michael S. Tsirkin wrote:
> On Mon, May 30, 2016 at 02:47:54AM -0400, Jason Wang wrote:
>> We always poll tx for socket, this is sub optimal since:
>>
>> - it will be only used when we exceed the sndbuf of the socket.
>> - since we use two independent polls for tx and vq, this will slightly
>>    increase the waitqueue traversing time and more important, vhost
>>    could not benefit from commit
>>    9e641bdcfa4ef4d6e2fbaa59c1be0ad5d1551fd5 ("net-tun: restructure
>>    tun_do_read for better sleep/wakeup efficiency") even if we've
>>    stopped rx polling during handle_rx since tx poll were still left in
>>    the waitqueue.
> Why is this an issue?
> sock_def_write_space only wakes up when queue is half empty,
> not on each packet.
>          if ((atomic_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf)
>
> I suspect the issue is with your previous patch,
> it now pokes at the spinlock on data path
> where it used not to.
>
> Is that right?

The problem is not tx wake up but still rx wake up. Patch 1 removes rx 
poll, but still left tx poll. So in sock_def_readable(), 
skwq_has_sleeper() returns true, we still need to traverse waitqueue and 
touch spinlocks. With this patch, unless a heavy tx load, tx poll were 
disabled, sock_def_readable() can return finish very soon.

>
>
>> Fix this by conditionally enable tx polling only when -EAGAIN were
>> met.
>>
>> Test shows about 8% improvement on guest rx pps.
>>
>> Before: ~1350000
>> After:  ~1460000
>>
>> Signed-off-by: Jason Wang <jasowang@redhat.com>
>> ---
>>   drivers/vhost/net.c | 3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
>> index e91603b..5a05fa0 100644
>> --- a/drivers/vhost/net.c
>> +++ b/drivers/vhost/net.c
>> @@ -378,6 +378,7 @@ static void handle_tx(struct vhost_net *net)
>>   		goto out;
>>   
>>   	vhost_disable_notify(&net->dev, vq);
>> +	vhost_net_disable_vq(net, vq);
>>   
>>   	hdr_size = nvq->vhost_hlen;
>>   	zcopy = nvq->ubufs;
>> @@ -459,6 +460,8 @@ static void handle_tx(struct vhost_net *net)
>>   					% UIO_MAXIOV;
>>   			}
>>   			vhost_discard_vq_desc(vq, 1);
>> +			if (err == -EAGAIN)
>> +				vhost_net_enable_vq(net, vq);
>>   			break;
>>   		}
>>   		if (err != len)
>> -- 
>> 1.8.3.1

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web