Linux Virtualization - Jun 2018 - [PATCH] net: vhost: improve performance when enable busyloop

This patch improves the guest receive performance from
host. On the handle_tx side, we poll the sock receive
queue at the same time. handle_rx do that in the same way.

we set the poll-us=100 us and use the iperf3 to test
its throughput. The iperf3 command is shown as below.

iperf3 -s -D
iperf3 -c 192.168.1.100 -i 1 -P 10 -t 10 -M 1400 --bandwidth 100000M

* With the patch:    21.1 Gbits/sec
* Without the patch: 12.7 Gbits/sec

Signed-off-by: Tonghao Zhang <zhangtonghao at didichuxing.com>
---
 drivers/vhost/net.c | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
index e7cf7d2..9364ede 100644
--- a/drivers/vhost/net.c
+++ b/drivers/vhost/net.c
@@ -429,22 +429,43 @@ static int vhost_net_enable_vq(struct vhost_net *n,
 	return vhost_poll_start(poll, sock->file);
 }
 
+static int sk_has_rx_data(struct sock *sk);
+
 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 				    struct vhost_virtqueue *vq,
 				    struct iovec iov[], unsigned int iov_size,
 				    unsigned int *out_num, unsigned int *in_num)
 {
 	unsigned long uninitialized_var(endtime);
+	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
+	struct vhost_virtqueue *rvq = &nvq->vq;
+	struct socket *sock = rvq->private_data;
+
 	int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 				  out_num, in_num, NULL, NULL);
 
 	if (r == vq->num && vq->busyloop_timeout) {
+		mutex_lock_nested(&rvq->mutex, 1);
+
+		vhost_disable_notify(&net->dev, rvq);
+
 		preempt_disable();
 		endtime = busy_clock() + vq->busyloop_timeout;
 		while (vhost_can_busy_poll(vq->dev, endtime) &&
+		       !(sock && sk_has_rx_data(sock->sk)) &&
 		       vhost_vq_avail_empty(vq->dev, vq))
 			cpu_relax();
 		preempt_enable();
+
+		if (sock && sk_has_rx_data(sock->sk))
+			vhost_poll_queue(&rvq->poll);
+		else if (unlikely(vhost_enable_notify(&net->dev, rvq))) {
+			vhost_disable_notify(&net->dev, rvq);
+			vhost_poll_queue(&rvq->poll);
+		}
+
+		mutex_unlock(&rvq->mutex);
+
 		r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 				      out_num, in_num, NULL, NULL);
 	}
-- 
1.8.3.1

On 2018?06?20? 21:28, Tonghao Zhang wrote:
> This patch improves the guest receive performance from
> host. On the handle_tx side, we poll the sock receive
> queue at the same time. handle_rx do that in the same way.
>
> we set the poll-us=100 us and use the iperf3 to test
> its throughput. The iperf3 command is shown as below.
>
> iperf3 -s -D
> iperf3 -c 192.168.1.100 -i 1 -P 10 -t 10 -M 1400 --bandwidth 100000M
>
> * With the patch:    21.1 Gbits/sec
> * Without the patch: 12.7 Gbits/sec
Thanks a lot for the patch. But looks like it needs some work to avoid 
e.g deadlock.

E.g in vhost_process_iotlb_msg() we call vhost_dev_lock_vqs() which did:

 ??? for (i = 0; i < d->nvqs; ++i)
 ??? ??? mutex_lock_nested(&d->vqs[i]->mutex, i);

I believe we need to change the code to lock the vq one by one like the 
attached (only compile test).
> Signed-off-by: Tonghao Zhang <zhangtonghao at didichuxing.com>
> ---
>   drivers/vhost/net.c | 21 +++++++++++++++++++++
>   1 file changed, 21 insertions(+)
>
> diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> index e7cf7d2..9364ede 100644
> --- a/drivers/vhost/net.c
> +++ b/drivers/vhost/net.c
> @@ -429,22 +429,43 @@ static int vhost_net_enable_vq(struct vhost_net *n,
>   	return vhost_poll_start(poll, sock->file);
>   }
>   
> +static int sk_has_rx_data(struct sock *sk);
> +
How about move sk_has_rx_data() here.
>   static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
>   				    struct vhost_virtqueue *vq,
>   				    struct iovec iov[], unsigned int iov_size,
>   				    unsigned int *out_num, unsigned int *in_num)
>   {
>   	unsigned long uninitialized_var(endtime);
> +	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
> +	struct vhost_virtqueue *rvq = &nvq->vq;
> +	struct socket *sock = rvq->private_data;
> +
>   	int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
>   				  out_num, in_num, NULL, NULL);
>   
>   	if (r == vq->num && vq->busyloop_timeout) {
> +		mutex_lock_nested(&rvq->mutex, 1);
> +
> +		vhost_disable_notify(&net->dev, rvq);
> +
>   		preempt_disable();
>   		endtime = busy_clock() + vq->busyloop_timeout;
>   		while (vhost_can_busy_poll(vq->dev, endtime) &&
> +		       !(sock && sk_has_rx_data(sock->sk)) &&
>   		       vhost_vq_avail_empty(vq->dev, vq))
>   			cpu_relax();
>   		preempt_enable();
> +
> +		if (sock && sk_has_rx_data(sock->sk))
> +			vhost_poll_queue(&rvq->poll);
> +		else if (unlikely(vhost_enable_notify(&net->dev, rvq))) {
> +			vhost_disable_notify(&net->dev, rvq);
> +			vhost_poll_queue(&rvq->poll);
> +		}
> +
> +		mutex_unlock(&rvq->mutex);
Some kinds of code duplication, can we try to unify them?

Btw, net-next is closed, so you need resubmit after it was open and use 
a "net-next" as the prefix of the patch.

Thanks
> +
>   		r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
>   				      out_num, in_num, NULL, NULL);
>   	}
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-vhost-lock-vqs-one-by-one.patch
Type: text/x-patch
Size: 2179 bytes
Desc: not available
URL:
<http://lists.linuxfoundation.org/pipermail/virtualization/attachments/20180621/ddd79f2a/attachment-0001.bin>