On Mon, Jul 01, 2024 at 04:28:03PM GMT, Luigi Leonardi via B4 Relay wrote:
From: Marco Pinna <marco.pinn95@xxxxxxxxx>
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@xxxxxxxxxxx>
Signed-off-by: Luigi Leonardi <luigi.leonardi@xxxxxxxxxxx>
Signed-off-by: Marco Pinna <marco.pinn95@xxxxxxxxx>
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
