Linux Virtualization - May 2016 - [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

We don't stop rx polling socket during rx processing, this will lead
unnecessary wakeups from under layer net devices (E.g
sock_def_readable() form tun). Rx will be slowed down in this
way. This patch avoids this by stop polling socket during rx
processing. A small drawback is that this introduces some overheads in
light load case because of the extra start/stop polling, but single
netperf TCP_RR does not notice any change. In a super heavy load case,
e.g using pktgen to inject packet to guest, we get about ~8.8%
improvement on pps:

before: ~1240000 pkt/s
after:  ~1350000 pkt/s

Signed-off-by: Jason Wang <jasowang at redhat.com>
---
 drivers/vhost/net.c | 56 +++++++++++++++++++++++++++--------------------------
 1 file changed, 29 insertions(+), 27 deletions(-)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index 10ff494..e91603b 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c
@@ -301,6 +301,32 @@ static bool vhost_can_busy_poll(struct vhost_dev *dev,
 	       !vhost_has_work(dev);
 }
 
+static void vhost_net_disable_vq(struct vhost_net *n,
+				 struct vhost_virtqueue *vq)
+{
+	struct vhost_net_virtqueue *nvq +		container_of(vq, struct
vhost_net_virtqueue, vq);
+	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
+	if (!vq->private_data)
+		return;
+	vhost_poll_stop(poll);
+}
+
+static int vhost_net_enable_vq(struct vhost_net *n,
+				struct vhost_virtqueue *vq)
+{
+	struct vhost_net_virtqueue *nvq +		container_of(vq, struct
vhost_net_virtqueue, vq);
+	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
+	struct socket *sock;
+
+	sock = vq->private_data;
+	if (!sock)
+		return 0;
+
+	return vhost_poll_start(poll, sock->file);
+}
+
 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 				    struct vhost_virtqueue *vq,
 				    struct iovec iov[], unsigned int iov_size,
@@ -627,6 +653,7 @@ static void handle_rx(struct vhost_net *net)
 	if (!sock)
 		goto out;
 	vhost_disable_notify(&net->dev, vq);
+	vhost_net_disable_vq(net, vq);
 
 	vhost_hlen = nvq->vhost_hlen;
 	sock_hlen = nvq->sock_hlen;
@@ -715,9 +742,10 @@ static void handle_rx(struct vhost_net *net)
 		total_len += vhost_len;
 		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 			vhost_poll_queue(&vq->poll);
-			break;
+			goto out;
 		}
 	}
+	vhost_net_enable_vq(net, vq);
 out:
 	mutex_unlock(&vq->mutex);
 }
@@ -796,32 +824,6 @@ static int vhost_net_open(struct inode *inode, struct file
*f)
 	return 0;
 }
 
-static void vhost_net_disable_vq(struct vhost_net *n,
-				 struct vhost_virtqueue *vq)
-{
-	struct vhost_net_virtqueue *nvq -		container_of(vq, struct
vhost_net_virtqueue, vq);
-	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
-	if (!vq->private_data)
-		return;
-	vhost_poll_stop(poll);
-}
-
-static int vhost_net_enable_vq(struct vhost_net *n,
-				struct vhost_virtqueue *vq)
-{
-	struct vhost_net_virtqueue *nvq -		container_of(vq, struct
vhost_net_virtqueue, vq);
-	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
-	struct socket *sock;
-
-	sock = vq->private_data;
-	if (!sock)
-		return 0;
-
-	return vhost_poll_start(poll, sock->file);
-}
-
 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 					struct vhost_virtqueue *vq)
 {
-- 
1.8.3.1

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 at 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

On Mon, May 30, 2016 at 02:47:53AM -0400, Jason Wang
wrote:
> We don't stop rx polling socket during rx processing, this will lead
> unnecessary wakeups from under layer net devices (E.g
> sock_def_readable() form tun). Rx will be slowed down in this
> way. This patch avoids this by stop polling socket during rx
> processing. A small drawback is that this introduces some overheads in
> light load case because of the extra start/stop polling, but single
> netperf TCP_RR does not notice any change. In a super heavy load case,
> e.g using pktgen to inject packet to guest, we get about ~8.8%
> improvement on pps:
> 
> before: ~1240000 pkt/s
> after:  ~1350000 pkt/s
> 
> Signed-off-by: Jason Wang <jasowang at redhat.com>
> ---
>  drivers/vhost/net.c | 56
+++++++++++++++++++++++++++--------------------------
>  1 file changed, 29 insertions(+), 27 deletions(-)
> 
> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> index 10ff494..e91603b 100644
> --- a/drivers/vhost/net.c
> +++ b/drivers/vhost/net.c
> @@ -301,6 +301,32 @@ static bool vhost_can_busy_poll(struct vhost_dev *dev,
>  	       !vhost_has_work(dev);
>  }
>  
> +static void vhost_net_disable_vq(struct vhost_net *n,
> +				 struct vhost_virtqueue *vq)
> +{
> +	struct vhost_net_virtqueue *nvq > +		container_of(vq, struct
vhost_net_virtqueue, vq);
> +	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
> +	if (!vq->private_data)
> +		return;
> +	vhost_poll_stop(poll);
> +}
> +
> +static int vhost_net_enable_vq(struct vhost_net *n,
> +				struct vhost_virtqueue *vq)
> +{
> +	struct vhost_net_virtqueue *nvq > +		container_of(vq, struct
vhost_net_virtqueue, vq);
> +	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
> +	struct socket *sock;
> +
> +	sock = vq->private_data;
> +	if (!sock)
> +		return 0;
> +
> +	return vhost_poll_start(poll, sock->file);
> +}
> +
>  static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
>  				    struct vhost_virtqueue *vq,
>  				    struct iovec iov[], unsigned int iov_size,
BTW we might want to rename these functions, name no longer
reflects function ...

> @@ -627,6 +653,7 @@ static void handle_rx(struct vhost_net *net)
>  	if (!sock)
>  		goto out;
>  	vhost_disable_notify(&net->dev, vq);
> +	vhost_net_disable_vq(net, vq);
>  
>  	vhost_hlen = nvq->vhost_hlen;
>  	sock_hlen = nvq->sock_hlen;
> @@ -715,9 +742,10 @@ static void handle_rx(struct vhost_net *net)
>  		total_len += vhost_len;
>  		if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
>  			vhost_poll_queue(&vq->poll);
> -			break;
> +			goto out;
>  		}
>  	}
> +	vhost_net_enable_vq(net, vq);
OK so if sock is readable but RX VQ is empty, this will
immediately schedule another round of handle_rx and so ad
infinitum,

Looks like a bug.

>  out:
>  	mutex_unlock(&vq->mutex);
>  }
> @@ -796,32 +824,6 @@ static int vhost_net_open(struct inode *inode, struct
file *f)
>  	return 0;
>  }
>  
> -static void vhost_net_disable_vq(struct vhost_net *n,
> -				 struct vhost_virtqueue *vq)
> -{
> -	struct vhost_net_virtqueue *nvq > -		container_of(vq, struct
vhost_net_virtqueue, vq);
> -	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
> -	if (!vq->private_data)
> -		return;
> -	vhost_poll_stop(poll);
> -}
> -
> -static int vhost_net_enable_vq(struct vhost_net *n,
> -				struct vhost_virtqueue *vq)
> -{
> -	struct vhost_net_virtqueue *nvq > -		container_of(vq, struct
vhost_net_virtqueue, vq);
> -	struct vhost_poll *poll = n->poll + (nvq - n->vqs);
> -	struct socket *sock;
> -
> -	sock = vq->private_data;
> -	if (!sock)
> -		return 0;
> -
> -	return vhost_poll_start(poll, sock->file);
> -}
> -
>  static struct socket *vhost_net_stop_vq(struct vhost_net *n,
>  					struct vhost_virtqueue *vq)
>  {
> -- 
> 1.8.3.1

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 at 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