Skip to contentSkip to conversation

don't clean packets in start_xmit in TX NAPI mode

RFC PATCH New Jason Wang 1 participant 3 messages Jan 22, 2025

Subsystems (4)
lore
Share & tools
lore

Jason Wang

author
#0 lore raw

cover don't clean packets in start_xmit in TX NAPI mode

Hi all:

This is an RFC to try to speed up TX by avoid cleaning transmitted packets in start_xmit() when TX NAPI mode is enabled. This increase 25%-62% PPS depending on the setups while keeping the TCP throughput.

More numbers could be seen in patch 2.

Thanks

Jason Wang (2):
virtio-net: factor out logic for stopping TX
virtio-net: free old xmit skbs only in NAPI

 drivers/net/virtio_net.c | 53 ++++++++++++++++++++++++++++++----------
 1 file changed, 40 insertions(+), 13 deletions(-)
Signature · 1 line
2.34.1
author
#1 lore raw

RFC PATCH 1/2 virtio-net: factor out logic for stopping TX

This patch factors out the logic of checking and stopping TX into a dedicated helper. This will be used by following patch.

---
 drivers/net/virtio_net.c | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)
drivers/net/virtio_net.c +22 -4
index 7646ddd9bef7..2f6c3dc68ba0 100644
@@ -1088,11 +1088,10 @@ static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
 		return false;
 }
 
-static void check_sq_full_and_disable(struct virtnet_info *vi,
-				      struct net_device *dev,
-				      struct send_queue *sq)
+static bool tx_may_stop(struct virtnet_info *vi,
+			struct net_device *dev,
+			struct send_queue *sq)
 {
-	bool use_napi = sq->napi.weight;
 	int qnum;
 
 	qnum = sq - vi->sq;
@@ -1114,6 +1113,25 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 		u64_stats_update_begin(&sq->stats.syncp);
 		u64_stats_inc(&sq->stats.stop);
 		u64_stats_update_end(&sq->stats.syncp);
+
+		return true;
+	}
+
+	return false;
+}
+
+static void check_sq_full_and_disable(struct virtnet_info *vi,
+				      struct net_device *dev,
+				      struct send_queue *sq)
+{
+	bool use_napi = sq->napi.weight;
+	int qnum;
+
+	qnum = sq - vi->sq;
+
+	if (tx_may_stop(vi, dev, sq)) {
+		struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
+
 		if (use_napi) {
 			if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
 				virtqueue_napi_schedule(&sq->napi, sq->vq);
Signature · 1 line
2.34.1
author
#2 lore raw

RFC PATCH 2/2 virtio-net: free old xmit skbs only in NAPI

In reply to Jason Wang

When NAPI mode is enabled, we try to free old transmited packets before sending a packet. This has several side effects:

- transmitted packets need to be freed before sending a packet, this introduces delay and increases the average packets transmit time.
- more time in hold the TX lock that causes more TX lock contention with the TX NAPI

This would be more noticeable when using a fast device like vhost-user/DPDK. So this patch tries to avoid those issues by not cleaning transmitted packets in start_xmit() when TX NAPI is enabled. Notification will be disabled at the beginning of the start_xmit() but we can't enable delayed notification after TX is stopped. Instead, the delayed notification needs to be enabled if we need to kick the virtqueue.

Performance numbers:

1) pktgen_sample03_burst_single_flow.sh (burst 256) + testpmd (rxonly) on the host:

- When pinning TX IRQ to pktgen VCPU: split virtqueue PPS were increased 62% from 6.45 Mpps to 10.5 Mpps; packed virtqueue PPS were increased 60% from 7.8 Mpps to 12.5 Mpps.
- When pinning TX IRQ to VCPU other than pktgen: split virtqueue PPS were increased 25% from 6.15 Mpps to 7.7 Mpps; packed virtqueue PPS were increased 50.6% from 8.3Mpps to 12.5 Mpps.

2) Netperf:

- Netperf in guest + vhost-net/TAP on the host doesn't show obvious differences.

---
 drivers/net/virtio_net.c | 27 ++++++++++++++++++---------
 1 file changed, 18 insertions(+), 9 deletions(-)
drivers/net/virtio_net.c +18 -9
index 2f6c3dc68ba0..3d5c44546dc1 100644
@@ -3271,15 +3271,10 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 	bool use_napi = sq->napi.weight;
 	bool kick;
 
-	/* Free up any pending old buffers before queueing new ones. */
-	do {
-		if (use_napi)
-			virtqueue_disable_cb(sq->vq);
-
+	if (!use_napi)
 		free_old_xmit(sq, txq, false);
-
-	} while (use_napi && !xmit_more &&
-	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
+	else
+		virtqueue_disable_cb(sq->vq);
 
 	/* timestamp packet in software */
 	skb_tx_timestamp(skb);
@@ -3305,7 +3300,18 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 		nf_reset_ct(skb);
 	}
 
-	check_sq_full_and_disable(vi, dev, sq);
+	if (tx_may_stop(vi, dev, sq) && !use_napi &&
+	    unlikely(virtqueue_enable_cb_delayed(sq->vq))) {
+		/* More just got used, free them then recheck. */
+		free_old_xmit(sq, txq, false);
+		if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
+			netif_start_subqueue(dev, qnum);
+			u64_stats_update_begin(&sq->stats.syncp);
+			u64_stats_inc(&sq->stats.wake);
+			u64_stats_update_end(&sq->stats.syncp);
+			virtqueue_disable_cb(sq->vq);
+		}
+	}
 
 	kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
 			  !xmit_more || netif_xmit_stopped(txq);
@@ -3317,6 +3323,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 		}
 	}
 
+	if (use_napi && kick && unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
+		virtqueue_napi_schedule(&sq->napi, sq->vq);
+
 	return NETDEV_TX_OK;
 }
 
Signature · 1 line
2.34.1