From: Marco Pinna <marco.pinn95@gmail.com>

Introduce an optimization in virtio_transport_send_pkt:
when the work queue (send_pkt_queue) is empty the packet is
put directly in the virtqueue reducing latency.

In the following benchmark (pingpong mode) the host sends
a payload to the guest and waits for the same payload back.

All vCPUs pinned individually to pCPUs.
vhost process pinned to a pCPU
fio process pinned both inside the host and the guest system.

Host CPU: Intel i7-10700KF CPU @ 3.80GHz
Tool: Fio version 3.37-56
Env: Phys host + L1 Guest
Payload: 512
Runtime-per-test: 50s
Mode: pingpong (h-g-h)
Test runs: 50
Type: SOCK_STREAM

Before (Linux 6.8.11)
------
mean(1st percentile):    380.56 ns
mean(overall):           780.83 ns
mean(99th percentile):  8300.24 ns

After
------
mean(1st percentile):   370.59 ns
mean(overall):          720.66 ns
mean(99th percentile): 7600.27 ns

Same setup, using 4K payload:

Before (Linux 6.8.11)
------
mean(1st percentile):    458.84 ns
mean(overall):          1650.17 ns
mean(99th percentile): 42240.68 ns

After
------
mean(1st percentile):    450.12 ns
mean(overall):          1460.84 ns
mean(99th percentile): 37632.45 ns

virtqueue.

Throughput: iperf-vsock

Before (Linux 6.8.11)
G2H 28.7 Gb/s

After
G2H 40.8 Gb/s

The performance improvement is related to this optimization,
I checked that each packet was put directly on the vq
avoiding the work queue.

Co-developed-by: Luigi Leonardi <luigi.leonardi@outlook.com>
Signed-off-by: Luigi Leonardi <luigi.leonardi@outlook.com>
Signed-off-by: Marco Pinna <marco.pinn95@gmail.com>
---
 net/vmw_vsock/virtio_transport.c | 38 ++++++++++++++++++++++++++++++++++++--
 1 file changed, 36 insertions(+), 2 deletions(-)

diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
index a74083d28120..3815aa8d956b 100644
--- a/net/vmw_vsock/virtio_transport.c
+++ b/net/vmw_vsock/virtio_transport.c
@@ -213,6 +213,7 @@ virtio_transport_send_pkt(struct sk_buff *skb)
 {
 	struct virtio_vsock_hdr *hdr;
 	struct virtio_vsock *vsock;
+	bool use_worker = true;
 	int len = skb->len;
 
 	hdr = virtio_vsock_hdr(skb);
@@ -234,8 +235,41 @@ virtio_transport_send_pkt(struct sk_buff *skb)
 	if (virtio_vsock_skb_reply(skb))
 		atomic_inc(&vsock->queued_replies);
 
-	virtio_vsock_skb_queue_tail(&vsock->send_pkt_queue, skb);
-	queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
+	/* If the workqueue (send_pkt_queue) is empty there is no need to enqueue the packet.
+	 * Just put it on the virtqueue using virtio_transport_send_skb.
+	 */
+	if (skb_queue_empty_lockless(&vsock->send_pkt_queue)) {
+		bool restart_rx = false;
+		struct virtqueue *vq;
+		int ret;
+
+		/* Inside RCU, can't sleep! */
+		ret = mutex_trylock(&vsock->tx_lock);
+		if (unlikely(ret == 0))
+			goto out_worker;
+
+		/* Driver is being removed, no need to enqueue the packet */
+		if (!vsock->tx_run)
+			goto out_rcu;
+
+		vq = vsock->vqs[VSOCK_VQ_TX];
+
+		if (!virtio_transport_send_skb(skb, vq, vsock, &restart_rx)) {
+			use_worker = false;
+			virtqueue_kick(vq);
+		}
+
+		mutex_unlock(&vsock->tx_lock);
+
+		if (restart_rx)
+			queue_work(virtio_vsock_workqueue, &vsock->rx_work);
+	}
+
+out_worker:
+	if (use_worker) {
+		virtio_vsock_skb_queue_tail(&vsock->send_pkt_queue, skb);
+		queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
+	}
 
 out_rcu:
 	rcu_read_unlock();

-- 
2.45.2

On Mon, Jul 01, 2024 at 04:28:03PM GMT, Luigi Leonardi via B4 Relay wrote:
>From: Marco Pinna <marco.pinn95@gmail.com>
>
>Introduce an optimization in virtio_transport_send_pkt:
>when the work queue (send_pkt_queue) is empty the packet is
>put directly in the virtqueue reducing latency.
>
>In the following benchmark (pingpong mode) the host sends
>a payload to the guest and waits for the same payload back.
>
>All vCPUs pinned individually to pCPUs.
>vhost process pinned to a pCPU
>fio process pinned both inside the host and the guest system.
>
>Host CPU: Intel i7-10700KF CPU @ 3.80GHz
>Tool: Fio version 3.37-56
>Env: Phys host + L1 Guest
>Payload: 512
>Runtime-per-test: 50s
>Mode: pingpong (h-g-h)
>Test runs: 50
>Type: SOCK_STREAM
>
>Before (Linux 6.8.11)
>------
>mean(1st percentile):    380.56 ns
>mean(overall):           780.83 ns
>mean(99th percentile):  8300.24 ns
>
>After
>------
>mean(1st percentile):   370.59 ns
>mean(overall):          720.66 ns
>mean(99th percentile): 7600.27 ns
>
>Same setup, using 4K payload:
>
>Before (Linux 6.8.11)
>------
>mean(1st percentile):    458.84 ns
>mean(overall):          1650.17 ns
>mean(99th percentile): 42240.68 ns
>
>After
>------
>mean(1st percentile):    450.12 ns
>mean(overall):          1460.84 ns
>mean(99th percentile): 37632.45 ns
>
>virtqueue.
>
>Throughput: iperf-vsock
>
>Before (Linux 6.8.11)
>G2H 28.7 Gb/s
>
>After
>G2H 40.8 Gb/s

Cool!

I'd suggest to add the length of buffer (-l param) used, and also
check more lenghts, like at least 4k, 64k, 128k.

>
>The performance improvement is related to this optimization,
>I checked that each packet was put directly on the vq
>avoiding the work queue.

How?

>
>Co-developed-by: Luigi Leonardi <luigi.leonardi@outlook.com>
>Signed-off-by: Luigi Leonardi <luigi.leonardi@outlook.com>
>Signed-off-by: Marco Pinna <marco.pinn95@gmail.com>

I think you might want to change the author of this patch, since it's 
changed a lot from Marco's original one. Obviously if you both agree on 
this.

Thanks,
Stefano

>---
> net/vmw_vsock/virtio_transport.c | 38 ++++++++++++++++++++++++++++++++++++--
> 1 file changed, 36 insertions(+), 2 deletions(-)
>
>diff --git a/net/vmw_vsock/virtio_transport.c b/net/vmw_vsock/virtio_transport.c
>index a74083d28120..3815aa8d956b 100644
>--- a/net/vmw_vsock/virtio_transport.c
>+++ b/net/vmw_vsock/virtio_transport.c
>@@ -213,6 +213,7 @@ virtio_transport_send_pkt(struct sk_buff *skb)
> {
> 	struct virtio_vsock_hdr *hdr;
> 	struct virtio_vsock *vsock;
>+	bool use_worker = true;
> 	int len = skb->len;
>
> 	hdr = virtio_vsock_hdr(skb);
>@@ -234,8 +235,41 @@ virtio_transport_send_pkt(struct sk_buff *skb)
> 	if (virtio_vsock_skb_reply(skb))
> 		atomic_inc(&vsock->queued_replies);
>
>-	virtio_vsock_skb_queue_tail(&vsock->send_pkt_queue, skb);
>-	queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
>+	/* If the workqueue (send_pkt_queue) is empty there is no need to enqueue the packet.
>+	 * Just put it on the virtqueue using virtio_transport_send_skb.
>+	 */
>+	if (skb_queue_empty_lockless(&vsock->send_pkt_queue)) {
>+		bool restart_rx = false;
>+		struct virtqueue *vq;
>+		int ret;
>+
>+		/* Inside RCU, can't sleep! */
>+		ret = mutex_trylock(&vsock->tx_lock);
>+		if (unlikely(ret == 0))
>+			goto out_worker;
>+
>+		/* Driver is being removed, no need to enqueue the packet */
>+		if (!vsock->tx_run)
>+			goto out_rcu;
>+
>+		vq = vsock->vqs[VSOCK_VQ_TX];
>+
>+		if (!virtio_transport_send_skb(skb, vq, vsock, &restart_rx)) {
>+			use_worker = false;
>+			virtqueue_kick(vq);
>+		}
>+
>+		mutex_unlock(&vsock->tx_lock);
>+
>+		if (restart_rx)
>+			queue_work(virtio_vsock_workqueue, &vsock->rx_work);
>+	}
>+
>+out_worker:
>+	if (use_worker) {
>+		virtio_vsock_skb_queue_tail(&vsock->send_pkt_queue, skb);
>+		queue_work(virtio_vsock_workqueue, &vsock->send_pkt_work);
>+	}
>
> out_rcu:
> 	rcu_read_unlock();
>
>-- 2.45.2
>
>