Skip to contentSkip to conversation

virtio_net: add support for Byte Queue Limits

PATCH Superseded Jiri Pirko 4 participants 11 messages Jun 12, 2024 – Jun 18, 2024

Subsystems (4)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jiri Pirko

author
#0 lore raw

PATCH v2 1/1 virtio_net: add support for Byte Queue Limits

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +32 -18
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
 	u64 packets;
+	u64 xdp_packets;
 	u64 bytes;
+	u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			    bool in_napi, bool use_napi,
 			    struct virtnet_sq_free_stats *stats)
 {
 	unsigned int len;
 	void *ptr;
 
 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-		++stats->packets;
-
 		if (!is_xdp_frame(ptr)) {
 			struct sk_buff *skb = ptr;
 
 			pr_debug("Sent skb %p\n", skb);
 
+			stats->packets++;
 			stats->bytes += skb->len;
 			napi_consume_skb(skb, in_napi);
 		} else {
 			struct xdp_frame *frame = ptr_to_xdp(ptr);
 
-			stats->bytes += xdp_get_frame_len(frame);
+			stats->xdp_packets++;
+			stats->xdp_bytes += xdp_get_frame_len(frame);
 			xdp_return_frame(frame);
 		}
 	}
+	if (use_napi)
+		netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }
 
 /* Converting between virtqueue no. and kernel tx/rx queue no.
@@ -955,21 +961,22 @@ static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
 	virtnet_rq_free_buf(vi, rq, buf);
 }
 
-static void free_old_xmit(struct send_queue *sq, bool in_napi)
+static void free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			  bool in_napi, bool use_napi)
 {
 	struct virtnet_sq_free_stats stats = {0};
 
-	__free_old_xmit(sq, in_napi, &stats);
+	__free_old_xmit(sq, txq, in_napi, use_napi, &stats);
 
 	/* Avoid overhead when no packets have been processed
 	 * happens when called speculatively from start_xmit.
 	 */
-	if (!stats.packets)
+	if (!stats.packets && !stats.xdp_packets)
 		return;
 
 	u64_stats_update_begin(&sq->stats.syncp);
-	u64_stats_add(&sq->stats.bytes, stats.bytes);
-	u64_stats_add(&sq->stats.packets, stats.packets);
+	u64_stats_add(&sq->stats.bytes, stats.bytes + stats.xdp_bytes);
+	u64_stats_add(&sq->stats.packets, stats.packets + stats.xdp_packets);
 	u64_stats_update_end(&sq->stats.syncp);
 }
 
@@ -1003,7 +1010,9 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 	 * early means 16 slots are typically wasted.
 	 */
 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
-		netif_stop_subqueue(dev, qnum);
+		struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
+
+		netif_tx_stop_queue(txq);
 		u64_stats_update_begin(&sq->stats.syncp);
 		u64_stats_inc(&sq->stats.stop);
 		u64_stats_update_end(&sq->stats.syncp);
@@ -1012,7 +1021,7 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 				virtqueue_napi_schedule(&sq->napi, sq->vq);
 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 			/* More just got used, free them then recheck. */
-			free_old_xmit(sq, false);
+			free_old_xmit(sq, txq, false, use_napi);
 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 				netif_start_subqueue(dev, qnum);
 				u64_stats_update_begin(&sq->stats.syncp);
@@ -1138,7 +1147,8 @@ static int virtnet_xdp_xmit(struct net_device *dev,
 	}
 
 	/* Free up any pending old buffers before queueing new ones. */
-	__free_old_xmit(sq, false, &stats);
+	__free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq),
+			false, sq->napi.weight, &stats);
 
 	for (i = 0; i < n; i++) {
 		struct xdp_frame *xdpf = frames[i];
@@ -2313,7 +2323,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
 
 		do {
 			virtqueue_disable_cb(sq->vq);
-			free_old_xmit(sq, true);
+			free_old_xmit(sq, txq, true, true);
 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
@@ -2412,6 +2422,7 @@ static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
 		goto err_xdp_reg_mem_model;
 
 	virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
+	netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, qp_index));
 	virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
 
 	return 0;
@@ -2471,7 +2482,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
 	txq = netdev_get_tx_queue(vi->dev, index);
 	__netif_tx_lock(txq, raw_smp_processor_id());
 	virtqueue_disable_cb(sq->vq);
-	free_old_xmit(sq, true);
+	free_old_xmit(sq, txq, true, true);
 
 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
 		if (netif_tx_queue_stopped(txq)) {
@@ -2559,17 +2570,18 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct send_queue *sq = &vi->sq[qnum];
 	int err;
 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !netdev_xmit_more();
+	bool xmit_more = netdev_xmit_more();
 	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);
 
-		free_old_xmit(sq, false);
+		free_old_xmit(sq, txq, false, use_napi);
 
-	} while (use_napi && kick &&
+	} while (use_napi && !xmit_more &&
 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 	/* timestamp packet in software */
@@ -2598,7 +2610,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	check_sq_full_and_disable(vi, dev, sq);
 
-	if (kick || netif_xmit_stopped(txq)) {
+	kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
+			  !xmit_more && netif_xmit_stopped(txq);
+	if (kick) {
 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
 			u64_stats_update_begin(&sq->stats.syncp);
 			u64_stats_inc(&sq->stats.kicks);
Signature · 1 line
2.45.1
#1 lore raw
Quoted 21 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875
Quoted 1 line

I see you now support both napi and non-napi. Thanks a lot Jiri! Just coming out of a national holiday here, as usual with a backlog - pls allow until Monday to review. Thanks!

Quoted 184 lines

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +32 -18
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
 	u64 packets;
+	u64 xdp_packets;
 	u64 bytes;
+	u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			    bool in_napi, bool use_napi,
 			    struct virtnet_sq_free_stats *stats)
 {
 	unsigned int len;
 	void *ptr;
 
 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-		++stats->packets;
-
 		if (!is_xdp_frame(ptr)) {
 			struct sk_buff *skb = ptr;
 
 			pr_debug("Sent skb %p\n", skb);
 
+			stats->packets++;
 			stats->bytes += skb->len;
 			napi_consume_skb(skb, in_napi);
 		} else {
 			struct xdp_frame *frame = ptr_to_xdp(ptr);
 
-			stats->bytes += xdp_get_frame_len(frame);
+			stats->xdp_packets++;
+			stats->xdp_bytes += xdp_get_frame_len(frame);
 			xdp_return_frame(frame);
 		}
 	}
+	if (use_napi)
+		netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }
 
 /* Converting between virtqueue no. and kernel tx/rx queue no.
@@ -955,21 +961,22 @@ static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
 	virtnet_rq_free_buf(vi, rq, buf);
 }
 
-static void free_old_xmit(struct send_queue *sq, bool in_napi)
+static void free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			  bool in_napi, bool use_napi)
 {
 	struct virtnet_sq_free_stats stats = {0};
 
-	__free_old_xmit(sq, in_napi, &stats);
+	__free_old_xmit(sq, txq, in_napi, use_napi, &stats);
 
 	/* Avoid overhead when no packets have been processed
 	 * happens when called speculatively from start_xmit.
 	 */
-	if (!stats.packets)
+	if (!stats.packets && !stats.xdp_packets)
 		return;
 
 	u64_stats_update_begin(&sq->stats.syncp);
-	u64_stats_add(&sq->stats.bytes, stats.bytes);
-	u64_stats_add(&sq->stats.packets, stats.packets);
+	u64_stats_add(&sq->stats.bytes, stats.bytes + stats.xdp_bytes);
+	u64_stats_add(&sq->stats.packets, stats.packets + stats.xdp_packets);
 	u64_stats_update_end(&sq->stats.syncp);
 }
 
@@ -1003,7 +1010,9 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 	 * early means 16 slots are typically wasted.
 	 */
 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
-		netif_stop_subqueue(dev, qnum);
+		struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
+
+		netif_tx_stop_queue(txq);
 		u64_stats_update_begin(&sq->stats.syncp);
 		u64_stats_inc(&sq->stats.stop);
 		u64_stats_update_end(&sq->stats.syncp);
@@ -1012,7 +1021,7 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 				virtqueue_napi_schedule(&sq->napi, sq->vq);
 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 			/* More just got used, free them then recheck. */
-			free_old_xmit(sq, false);
+			free_old_xmit(sq, txq, false, use_napi);
 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 				netif_start_subqueue(dev, qnum);
 				u64_stats_update_begin(&sq->stats.syncp);
@@ -1138,7 +1147,8 @@ static int virtnet_xdp_xmit(struct net_device *dev,
 	}
 
 	/* Free up any pending old buffers before queueing new ones. */
-	__free_old_xmit(sq, false, &stats);
+	__free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq),
+			false, sq->napi.weight, &stats);
 
 	for (i = 0; i < n; i++) {
 		struct xdp_frame *xdpf = frames[i];
@@ -2313,7 +2323,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
 
 		do {
 			virtqueue_disable_cb(sq->vq);
-			free_old_xmit(sq, true);
+			free_old_xmit(sq, txq, true, true);
 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
@@ -2412,6 +2422,7 @@ static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
 		goto err_xdp_reg_mem_model;
 
 	virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
+	netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, qp_index));
 	virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
 
 	return 0;
@@ -2471,7 +2482,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
 	txq = netdev_get_tx_queue(vi->dev, index);
 	__netif_tx_lock(txq, raw_smp_processor_id());
 	virtqueue_disable_cb(sq->vq);
-	free_old_xmit(sq, true);
+	free_old_xmit(sq, txq, true, true);
 
 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
 		if (netif_tx_queue_stopped(txq)) {
@@ -2559,17 +2570,18 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct send_queue *sq = &vi->sq[qnum];
 	int err;
 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !netdev_xmit_more();
+	bool xmit_more = netdev_xmit_more();
 	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);
 
-		free_old_xmit(sq, false);
+		free_old_xmit(sq, txq, false, use_napi);
 
-	} while (use_napi && kick &&
+	} while (use_napi && !xmit_more &&
 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 	/* timestamp packet in software */
@@ -2598,7 +2610,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	check_sq_full_and_disable(vi, dev, sq);
 
-	if (kick || netif_xmit_stopped(txq)) {
+	kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
+			  !xmit_more && netif_xmit_stopped(txq);
+	if (kick) {
 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
 			u64_stats_update_begin(&sq->stats.syncp);
 			u64_stats_inc(&sq->stats.kicks);
Signature · 1 line
2.45.1
#2 lore raw
In reply to Jiri Pirko

Hello Jiri,

Quoted 9 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

Quoted 12 lines
NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

I cannot get such a huge improvement when I was doing multiple tests between two VMs. I'm pretty sure the BQL feature is working, but the numbers look the same with/without BQL.

VM 1 (client):
16 cpus, x86_64, 4 queues, the latest net-next kernel with/without this patch, pfifo_fast, napi_tx=true, napi_weight=128

VM 2 (server):
16 cpus, aarch64, 4 queues, the latest net-next kernel without this patch, pfifo_fast

What the 'ping' command shows to me between two VMs is : rtt min/avg/max/mdev = 0.233/0.257/0.300/0.024 ms

I started 50 netperfs to communicate the other side with the following command:
#!/bin/bash

for i in $(seq 5000 5050);
do
netperf -p $i -H [ip addr] -l 60 -t TCP_RR -- -r 64,64 > /dev/null 2>&1 & done

The results are around 30423.62 txkB/s. If I remove '-r 64 64', they are still the same/similar.

Am I missing something?

Thanks,
Jason

author
#3 lore raw
Quoted 43 lines · Fri, Jun 14, 2024 at 11:54:04AM CEST, [email protected] wrote:

Hello Jiri,

Quoted 9 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

Quoted 12 lines
NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

I cannot get such a huge improvement when I was doing multiple tests between two VMs. I'm pretty sure the BQL feature is working, but the numbers look the same with/without BQL.

VM 1 (client):
16 cpus, x86_64, 4 queues, the latest net-next kernel with/without this patch, pfifo_fast, napi_tx=true, napi_weight=128

VM 2 (server):
16 cpus, aarch64, 4 queues, the latest net-next kernel without this patch, pfifo_fast

What the 'ping' command shows to me between two VMs is : rtt min/avg/max/mdev = 0.233/0.257/0.300/0.024 ms

I started 50 netperfs to communicate the other side with the following command:
#!/bin/bash

Quoted 7 lines

for i in $(seq 5000 5050);
do
netperf -p $i -H [ip addr] -l 60 -t TCP_RR -- -r 64,64 > /dev/null 2>&1 & done

The results are around 30423.62 txkB/s. If I remove '-r 64 64', they are still the same/similar.

You have to stress the line by parallel TCP_STREAM instances (50 in my case). For consistent results, use -p portnum,locport to specify the local port.

Then run TCP_RR, also use -p portnum,locport to specify the local port.

Also, double check the CONFIG_BQL=y is set.

Quoted 5 lines

Am I missing something?

Thanks,
Jason

Tested-by
#4 lore raw
Quoted 50 lines · Jiri Pirko
Quoted 48 lines · Fri, Jun 14, 2024 at 11:54:04AM CEST, [email protected] wrote:

Hello Jiri,

Quoted 9 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

Quoted 12 lines
NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

I cannot get such a huge improvement when I was doing multiple tests between two VMs. I'm pretty sure the BQL feature is working, but the numbers look the same with/without BQL.

VM 1 (client):
16 cpus, x86_64, 4 queues, the latest net-next kernel with/without this patch, pfifo_fast, napi_tx=true, napi_weight=128

VM 2 (server):
16 cpus, aarch64, 4 queues, the latest net-next kernel without this patch, pfifo_fast

What the 'ping' command shows to me between two VMs is : rtt min/avg/max/mdev = 0.233/0.257/0.300/0.024 ms

I started 50 netperfs to communicate the other side with the following command:
#!/bin/bash

for i in $(seq 5000 5050);
do
netperf -p $i -H [ip addr] -l 60 -t TCP_RR -- -r 64,64 > /dev/null 2>&1 & done

Quoted 6 lines
Quoted 2 lines

The results are around 30423.62 txkB/s. If I remove '-r 64 64', they are still the same/similar.

You have to stress the line by parallel TCP_STREAM instances (50 in my case). For consistent results, use -p portnum,locport to specify the local port.

Thanks. Even though the results of TCP_RR mode vary sometimes, I can see a big improvement in the total value of those results under such circumstances.
With BQL, the throughput is 2159.17
Without BQL, it's 1099.33

Please feel free to add the tag:

Thanks,
Jason

#5 lore raw
In reply to Jiri Pirko
Quoted 76 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +5 -3
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
        u64 packets;
+       u64 xdp_packets;
        u64 bytes;
+       u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
        return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+                           bool in_napi, bool use_napi,
                            struct virtnet_sq_free_stats *stats)
 {
        unsigned int len;
        void *ptr;
 
        while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-               ++stats->packets;
-
                if (!is_xdp_frame(ptr)) {
                        struct sk_buff *skb = ptr;
 
                        pr_debug("Sent skb %p\n", skb);
 
+                       stats->packets++;
                        stats->bytes += skb->len;
                        napi_consume_skb(skb, in_napi);
                } else {
                        struct xdp_frame *frame = ptr_to_xdp(ptr);
 

drivers/net/virtio_net.c static struct xdp_frame *ptr_to_xdp(void *ptr) · line 538

Quoted 10 lines
drivers/net/virtio_net.c +5 -1
-                       stats->bytes += xdp_get_frame_len(frame);
+                       stats->xdp_packets++;
+                       stats->xdp_bytes += xdp_get_frame_len(frame);
                        xdp_return_frame(frame);
                }
        }
+       if (use_napi)
+               netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }

I wonder if this works correctly, for example NAPI could be enabled after queued but before sent. So __netdev_tx_sent_queue() is not called before.

Thanks

author
#6 lore raw
Quoted 77 lines · Mon, Jun 17, 2024 at 04:34:26AM CEST, [email protected] wrote:
Quoted 76 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +5 -3
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
        u64 packets;
+       u64 xdp_packets;
        u64 bytes;
+       u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
        return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+                           bool in_napi, bool use_napi,
                            struct virtnet_sq_free_stats *stats)
 {
        unsigned int len;
        void *ptr;
 
        while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-               ++stats->packets;
-
                if (!is_xdp_frame(ptr)) {
                        struct sk_buff *skb = ptr;
 
                        pr_debug("Sent skb %p\n", skb);
 
+                       stats->packets++;
                        stats->bytes += skb->len;
                        napi_consume_skb(skb, in_napi);
                } else {
                        struct xdp_frame *frame = ptr_to_xdp(ptr);
 
Quoted 14 lines

drivers/net/virtio_net.c static struct xdp_frame *ptr_to_xdp(void *ptr) · line 538

Quoted 10 lines
drivers/net/virtio_net.c +5 -1
-                       stats->bytes += xdp_get_frame_len(frame);
+                       stats->xdp_packets++;
+                       stats->xdp_bytes += xdp_get_frame_len(frame);
                        xdp_return_frame(frame);
                }
        }
+       if (use_napi)
+               netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }

I wonder if this works correctly, for example NAPI could be enabled after queued but before sent. So __netdev_tx_sent_queue() is not called before.

How is that possible? Napi weight can't change when link is up. Or am I missing something?

Quoted 3 lines

Thanks

#7 lore raw
Quoted 94 lines · Jiri Pirko
Quoted 77 lines · Mon, Jun 17, 2024 at 04:34:26AM CEST, [email protected] wrote:
Quoted 76 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +5 -3
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
        u64 packets;
+       u64 xdp_packets;
        u64 bytes;
+       u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
        return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+                           bool in_napi, bool use_napi,
                            struct virtnet_sq_free_stats *stats)
 {
        unsigned int len;
        void *ptr;
 
        while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-               ++stats->packets;
-
                if (!is_xdp_frame(ptr)) {
                        struct sk_buff *skb = ptr;
 
                        pr_debug("Sent skb %p\n", skb);
 
+                       stats->packets++;
                        stats->bytes += skb->len;
                        napi_consume_skb(skb, in_napi);
                } else {
                        struct xdp_frame *frame = ptr_to_xdp(ptr);
 
Quoted 14 lines

drivers/net/virtio_net.c static struct xdp_frame *ptr_to_xdp(void *ptr) · line 538

Quoted 10 lines
drivers/net/virtio_net.c +5 -1
-                       stats->bytes += xdp_get_frame_len(frame);
+                       stats->xdp_packets++;
+                       stats->xdp_bytes += xdp_get_frame_len(frame);
                        xdp_return_frame(frame);
                }
        }
+       if (use_napi)
+               netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }

I wonder if this works correctly, for example NAPI could be enabled after queued but before sent. So __netdev_tx_sent_queue() is not called before.

Quoted 2 lines

How is that possible? Napi weight can't change when link is up. Or am I missing something?

Something like this:

1) packet were queued
2) if down
3) enable NAPI
4) if up
5) packet were sent

?

Thanks

Quoted 5 lines
Quoted 4 lines

Thanks

author
#8 lore raw
Quoted 98 lines · Tue, Jun 18, 2024 at 02:53:42AM CEST, [email protected] wrote:
Quoted 94 lines · Jiri Pirko
Quoted 77 lines · Mon, Jun 17, 2024 at 04:34:26AM CEST, [email protected] wrote:
Quoted 76 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +5 -3
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
        u64 packets;
+       u64 xdp_packets;
        u64 bytes;
+       u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
        return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+                           bool in_napi, bool use_napi,
                            struct virtnet_sq_free_stats *stats)
 {
        unsigned int len;
        void *ptr;
 
        while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-               ++stats->packets;
-
                if (!is_xdp_frame(ptr)) {
                        struct sk_buff *skb = ptr;
 
                        pr_debug("Sent skb %p\n", skb);
 
+                       stats->packets++;
                        stats->bytes += skb->len;
                        napi_consume_skb(skb, in_napi);
                } else {
                        struct xdp_frame *frame = ptr_to_xdp(ptr);
 
Quoted 14 lines

drivers/net/virtio_net.c static struct xdp_frame *ptr_to_xdp(void *ptr) · line 538

Quoted 10 lines
drivers/net/virtio_net.c +5 -1
-                       stats->bytes += xdp_get_frame_len(frame);
+                       stats->xdp_packets++;
+                       stats->xdp_bytes += xdp_get_frame_len(frame);
                        xdp_return_frame(frame);
                }
        }
+       if (use_napi)
+               netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }

I wonder if this works correctly, for example NAPI could be enabled after queued but before sent. So __netdev_tx_sent_queue() is not called before.

Quoted 2 lines

How is that possible? Napi weight can't change when link is up. Or am I missing something?

Quoted 7 lines

Something like this:

1) packet were queued
2) if down
3) enable NAPI
4) if up
5) packet were sent

Gotcha, will try to fix. Thanks!

Quoted 11 lines

?

Thanks

Quoted 5 lines
Quoted 4 lines

Thanks

#9 lore raw
In reply to Jiri Pirko
Quoted 136 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +20 -11
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
 	u64 packets;
+	u64 xdp_packets;
 	u64 bytes;
+	u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			    bool in_napi, bool use_napi,
 			    struct virtnet_sq_free_stats *stats)
 {
 	unsigned int len;
 	void *ptr;
 
 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-		++stats->packets;
-
 		if (!is_xdp_frame(ptr)) {
 			struct sk_buff *skb = ptr;
 
 			pr_debug("Sent skb %p\n", skb);
 
+			stats->packets++;
 			stats->bytes += skb->len;
 			napi_consume_skb(skb, in_napi);
 		} else {
 			struct xdp_frame *frame = ptr_to_xdp(ptr);
 
-			stats->bytes += xdp_get_frame_len(frame);
+			stats->xdp_packets++;
+			stats->xdp_bytes += xdp_get_frame_len(frame);
 			xdp_return_frame(frame);
 		}
 	}
+	if (use_napi)
+		netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }
 
 /* Converting between virtqueue no. and kernel tx/rx queue no.
@@ -955,21 +961,22 @@ static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
 	virtnet_rq_free_buf(vi, rq, buf);
 }
 
-static void free_old_xmit(struct send_queue *sq, bool in_napi)
+static void free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			  bool in_napi, bool use_napi)
 {
 	struct virtnet_sq_free_stats stats = {0};
 
-	__free_old_xmit(sq, in_napi, &stats);
+	__free_old_xmit(sq, txq, in_napi, use_napi, &stats);
 
 	/* Avoid overhead when no packets have been processed
 	 * happens when called speculatively from start_xmit.
 	 */
-	if (!stats.packets)
+	if (!stats.packets && !stats.xdp_packets)
 		return;
 
 	u64_stats_update_begin(&sq->stats.syncp);
-	u64_stats_add(&sq->stats.bytes, stats.bytes);
-	u64_stats_add(&sq->stats.packets, stats.packets);
+	u64_stats_add(&sq->stats.bytes, stats.bytes + stats.xdp_bytes);
+	u64_stats_add(&sq->stats.packets, stats.packets + stats.xdp_packets);
 	u64_stats_update_end(&sq->stats.syncp);
 }
 
@@ -1003,7 +1010,9 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 	 * early means 16 slots are typically wasted.
 	 */
 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
-		netif_stop_subqueue(dev, qnum);
+		struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
+
+		netif_tx_stop_queue(txq);
 		u64_stats_update_begin(&sq->stats.syncp);
 		u64_stats_inc(&sq->stats.stop);
 		u64_stats_update_end(&sq->stats.syncp);
@@ -1012,7 +1021,7 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 				virtqueue_napi_schedule(&sq->napi, sq->vq);
 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 			/* More just got used, free them then recheck. */
-			free_old_xmit(sq, false);
+			free_old_xmit(sq, txq, false, use_napi);
 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 				netif_start_subqueue(dev, qnum);
 				u64_stats_update_begin(&sq->stats.syncp);

drivers/net/virtio_net.c static int virtnet_xdp_xmit(struct net_device *dev, · line 1151

Quoted 6 lines
drivers/net/virtio_net.c +2 -1
@@ -1138,7 +1147,8 @@ static int virtnet_xdp_xmit(struct net_device *dev,
 	}
 
 	/* Free up any pending old buffers before queueing new ones. */
-	__free_old_xmit(sq, false, &stats);
+	__free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq),
+			false, sq->napi.weight, &stats);

So what happens if sq->napi.weight changes while packet is in the queue? What prevents that?

Quoted 64 lines
drivers/net/virtio_net.c +10 -6
 	for (i = 0; i < n; i++) {
 		struct xdp_frame *xdpf = frames[i];
@@ -2313,7 +2323,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
 
 		do {
 			virtqueue_disable_cb(sq->vq);
-			free_old_xmit(sq, true);
+			free_old_xmit(sq, txq, true, true);
 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
@@ -2412,6 +2422,7 @@ static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
 		goto err_xdp_reg_mem_model;
 
 	virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
+	netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, qp_index));
 	virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
 
 	return 0;
@@ -2471,7 +2482,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
 	txq = netdev_get_tx_queue(vi->dev, index);
 	__netif_tx_lock(txq, raw_smp_processor_id());
 	virtqueue_disable_cb(sq->vq);
-	free_old_xmit(sq, true);
+	free_old_xmit(sq, txq, true, true);
 
 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
 		if (netif_tx_queue_stopped(txq)) {
@@ -2559,17 +2570,18 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct send_queue *sq = &vi->sq[qnum];
 	int err;
 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !netdev_xmit_more();
+	bool xmit_more = netdev_xmit_more();
 	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);
 
-		free_old_xmit(sq, false);
+		free_old_xmit(sq, txq, false, use_napi);
 
-	} while (use_napi && kick &&
+	} while (use_napi && !xmit_more &&
 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 	/* timestamp packet in software */
@@ -2598,7 +2610,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	check_sq_full_and_disable(vi, dev, sq);
 
-	if (kick || netif_xmit_stopped(txq)) {
+	kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
+			  !xmit_more && netif_xmit_stopped(txq);
+	if (kick) {
 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
 			u64_stats_update_begin(&sq->stats.syncp);
 			u64_stats_inc(&sq->stats.kicks);
Signature · 1 line
2.45.1
author
#10 lore raw
Quoted 139 lines · Mon, Jun 17, 2024 at 06:14:11PM CEST, [email protected] wrote:
Quoted 138 lines · Jiri Pirko

From: Jiri Pirko <[email protected]>

Add support for Byte Queue Limits (BQL).

Tested on qemu emulated virtio_net device with 1, 2 and 4 queues. Tested with fq_codel and pfifo_fast. Super netperf with 50 threads is running in background. Netperf TCP_RR results:

NOBQL FQC 1q:  159.56  159.33  158.50  154.31    agv: 157.925
NOBQL FQC 2q:  184.64  184.96  174.73  174.15    agv: 179.62
NOBQL FQC 4q:  994.46  441.96  416.50  499.56    agv: 588.12
NOBQL PFF 1q:  148.68  148.92  145.95  149.48    agv: 148.2575
NOBQL PFF 2q:  171.86  171.20  170.42  169.42    agv: 170.725
NOBQL PFF 4q: 1505.23 1137.23 2488.70 3507.99    agv: 2159.7875
  BQL FQC 1q: 1332.80 1297.97 1351.41 1147.57    agv: 1282.4375
  BQL FQC 2q:  768.30  817.72  864.43  974.40    agv: 856.2125
  BQL FQC 4q:  945.66  942.68  878.51  822.82    agv: 897.4175
  BQL PFF 1q:  149.69  151.49  149.40  147.47    agv: 149.5125
  BQL PFF 2q: 2059.32  798.74 1844.12  381.80    agv: 1270.995
  BQL PFF 4q: 1871.98 4420.02 4916.59 13268.16   agv: 6119.1875

---
v1->v2:
- moved netdev_tx_completed_queue() call into __free_old_xmit(), propagate use_napi flag to __free_old_xmit() and only call netdev_tx_completed_queue() in case it is true
- added forgotten call to netdev_tx_reset_queue()
- fixed stats for xdp packets
- fixed bql accounting when __free_old_xmit() is called from xdp path
- handle the !use_napi case in start_xmit() kick section

---
 drivers/net/virtio_net.c | 50 +++++++++++++++++++++++++---------------
 1 file changed, 32 insertions(+), 18 deletions(-)
drivers/net/virtio_net.c +20 -11
index 61a57d134544..5863c663ccab 100644
@@ -84,7 +84,9 @@ struct virtnet_stat_desc {
 
 struct virtnet_sq_free_stats {
 	u64 packets;
+	u64 xdp_packets;
 	u64 bytes;
+	u64 xdp_bytes;
 };
 
 struct virtnet_sq_stats {
@@ -506,29 +508,33 @@ static struct xdp_frame *ptr_to_xdp(void *ptr)
 	return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
 }
 
-static void __free_old_xmit(struct send_queue *sq, bool in_napi,
+static void __free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			    bool in_napi, bool use_napi,
 			    struct virtnet_sq_free_stats *stats)
 {
 	unsigned int len;
 	void *ptr;
 
 	while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
-		++stats->packets;
-
 		if (!is_xdp_frame(ptr)) {
 			struct sk_buff *skb = ptr;
 
 			pr_debug("Sent skb %p\n", skb);
 
+			stats->packets++;
 			stats->bytes += skb->len;
 			napi_consume_skb(skb, in_napi);
 		} else {
 			struct xdp_frame *frame = ptr_to_xdp(ptr);
 
-			stats->bytes += xdp_get_frame_len(frame);
+			stats->xdp_packets++;
+			stats->xdp_bytes += xdp_get_frame_len(frame);
 			xdp_return_frame(frame);
 		}
 	}
+	if (use_napi)
+		netdev_tx_completed_queue(txq, stats->packets, stats->bytes);
+
 }
 
 /* Converting between virtqueue no. and kernel tx/rx queue no.
@@ -955,21 +961,22 @@ static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
 	virtnet_rq_free_buf(vi, rq, buf);
 }
 
-static void free_old_xmit(struct send_queue *sq, bool in_napi)
+static void free_old_xmit(struct send_queue *sq, struct netdev_queue *txq,
+			  bool in_napi, bool use_napi)
 {
 	struct virtnet_sq_free_stats stats = {0};
 
-	__free_old_xmit(sq, in_napi, &stats);
+	__free_old_xmit(sq, txq, in_napi, use_napi, &stats);
 
 	/* Avoid overhead when no packets have been processed
 	 * happens when called speculatively from start_xmit.
 	 */
-	if (!stats.packets)
+	if (!stats.packets && !stats.xdp_packets)
 		return;
 
 	u64_stats_update_begin(&sq->stats.syncp);
-	u64_stats_add(&sq->stats.bytes, stats.bytes);
-	u64_stats_add(&sq->stats.packets, stats.packets);
+	u64_stats_add(&sq->stats.bytes, stats.bytes + stats.xdp_bytes);
+	u64_stats_add(&sq->stats.packets, stats.packets + stats.xdp_packets);
 	u64_stats_update_end(&sq->stats.syncp);
 }
 
@@ -1003,7 +1010,9 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 	 * early means 16 slots are typically wasted.
 	 */
 	if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
-		netif_stop_subqueue(dev, qnum);
+		struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
+
+		netif_tx_stop_queue(txq);
 		u64_stats_update_begin(&sq->stats.syncp);
 		u64_stats_inc(&sq->stats.stop);
 		u64_stats_update_end(&sq->stats.syncp);
@@ -1012,7 +1021,7 @@ static void check_sq_full_and_disable(struct virtnet_info *vi,
 				virtqueue_napi_schedule(&sq->napi, sq->vq);
 		} else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
 			/* More just got used, free them then recheck. */
-			free_old_xmit(sq, false);
+			free_old_xmit(sq, txq, false, use_napi);
 			if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
 				netif_start_subqueue(dev, qnum);
 				u64_stats_update_begin(&sq->stats.syncp);
@@ -1138,7 +1147,8 @@ static int virtnet_xdp_xmit(struct net_device *dev,
 	}
 
Quoted 8 lines

drivers/net/virtio_net.c static int virtnet_xdp_xmit(struct net_device *dev, · line 1151

Quoted 4 lines
drivers/net/virtio_net.c +2 -1
 	/* Free up any pending old buffers before queueing new ones. */
-	__free_old_xmit(sq, false, &stats);
+	__free_old_xmit(sq, netdev_get_tx_queue(dev, sq - vi->sq),
+			false, sq->napi.weight, &stats);

So what happens if sq->napi.weight changes while packet is in the queue? What prevents that?

Sigh. You are correct. That is what Jason points out as well.

Quoted 68 lines
Quoted 67 lines
drivers/net/virtio_net.c +10 -6
 	for (i = 0; i < n; i++) {
 		struct xdp_frame *xdpf = frames[i];
@@ -2313,7 +2323,7 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
 
 		do {
 			virtqueue_disable_cb(sq->vq);
-			free_old_xmit(sq, true);
+			free_old_xmit(sq, txq, true, true);
 		} while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 		if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
@@ -2412,6 +2422,7 @@ static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
 		goto err_xdp_reg_mem_model;
 
 	virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
+	netdev_tx_reset_queue(netdev_get_tx_queue(vi->dev, qp_index));
 	virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
 
 	return 0;
@@ -2471,7 +2482,7 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
 	txq = netdev_get_tx_queue(vi->dev, index);
 	__netif_tx_lock(txq, raw_smp_processor_id());
 	virtqueue_disable_cb(sq->vq);
-	free_old_xmit(sq, true);
+	free_old_xmit(sq, txq, true, true);
 
 	if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS) {
 		if (netif_tx_queue_stopped(txq)) {
@@ -2559,17 +2570,18 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 	struct send_queue *sq = &vi->sq[qnum];
 	int err;
 	struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
-	bool kick = !netdev_xmit_more();
+	bool xmit_more = netdev_xmit_more();
 	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);
 
-		free_old_xmit(sq, false);
+		free_old_xmit(sq, txq, false, use_napi);
 
-	} while (use_napi && kick &&
+	} while (use_napi && !xmit_more &&
 	       unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
 
 	/* timestamp packet in software */
@@ -2598,7 +2610,9 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	check_sq_full_and_disable(vi, dev, sq);
 
-	if (kick || netif_xmit_stopped(txq)) {
+	kick = use_napi ? __netdev_tx_sent_queue(txq, skb->len, xmit_more) :
+			  !xmit_more && netif_xmit_stopped(txq);
+	if (kick) {
 		if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
 			u64_stats_update_begin(&sq->stats.syncp);
 			u64_stats_inc(&sq->stats.kicks);
Signature · 1 line
2.45.1