Skip to contentSkip to conversation

virtio-net: switch to use XPS to choose txq

PATCH In review Jason Wang 4 participants 5 messages Sep 30, 2013 – Oct 2, 2013

Subsystems (4)
lore
Share & tools
lore

Jason Wang

author
#0 lore raw

PATCH v2 1/1 virtio-net: switch to use XPS to choose txq

We used to use a percpu structure vq_index to record the cpu to queue mapping, this is suboptimal since it duplicates the work of XPS and loses all other XPS functionality such as allowing use to configure their own transmission steering strategy.

So this patch switches to use XPS and suggest a default mapping when the number of cpus is equal to the number of queues. With XPS support, there's no need for keeping per-cpu vq_index and .ndo_select_queue(), so they were removed also.

Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>

---
Changes from V1:
- use cpumask_of() instead of allocate dynamically

 drivers/net/virtio_net.c |   48 +--------------------------------------------
 1 files changed, 2 insertions(+), 46 deletions(-)
drivers/net/virtio_net.c +2 -46
index defec2b..4eca652 100644
@@ -127,9 +127,6 @@ struct virtnet_info {
 	/* Does the affinity hint is set for virtqueues? */
 	bool affinity_hint_set;
 
-	/* Per-cpu variable to show the mapping from CPU to virtqueue */
-	int __percpu *vq_index;
-
 	/* CPU hot plug notifier */
 	struct notifier_block nb;
 };
@@ -1063,7 +1060,6 @@ static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
 {
 	int i;
-	int cpu;
 
 	if (vi->affinity_hint_set) {
 		for (i = 0; i < vi->max_queue_pairs; i++) {
@@ -1073,16 +1069,6 @@ static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
 
 		vi->affinity_hint_set = false;
 	}
-
-	i = 0;
-	for_each_online_cpu(cpu) {
-		if (cpu == hcpu) {
-			*per_cpu_ptr(vi->vq_index, cpu) = -1;
-		} else {
-			*per_cpu_ptr(vi->vq_index, cpu) =
-				++i % vi->curr_queue_pairs;
-		}
-	}
 }
 
 static void virtnet_set_affinity(struct virtnet_info *vi)
@@ -1104,7 +1090,7 @@ static void virtnet_set_affinity(struct virtnet_info *vi)
 	for_each_online_cpu(cpu) {
 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
-		*per_cpu_ptr(vi->vq_index, cpu) = i;
+		netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
 		i++;
 	}
 
@@ -1217,28 +1203,6 @@ static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 	return 0;
 }
 
-/* To avoid contending a lock hold by a vcpu who would exit to host, select the
- * txq based on the processor id.
- */
-static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
-{
-	int txq;
-	struct virtnet_info *vi = netdev_priv(dev);
-
-	if (skb_rx_queue_recorded(skb)) {
-		txq = skb_get_rx_queue(skb);
-	} else {
-		txq = *__this_cpu_ptr(vi->vq_index);
-		if (txq == -1)
-			txq = 0;
-	}
-
-	while (unlikely(txq >= dev->real_num_tx_queues))
-		txq -= dev->real_num_tx_queues;
-
-	return txq;
-}
-
 static const struct net_device_ops virtnet_netdev = {
 	.ndo_open            = virtnet_open,
 	.ndo_stop   	     = virtnet_close,
@@ -1250,7 +1214,6 @@ static const struct net_device_ops virtnet_netdev = {
 	.ndo_get_stats64     = virtnet_stats,
 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
-	.ndo_select_queue     = virtnet_select_queue,
 #ifdef CONFIG_NET_POLL_CONTROLLER
 	.ndo_poll_controller = virtnet_netpoll,
 #endif
@@ -1559,10 +1522,6 @@ static int virtnet_probe(struct virtio_device *vdev)
 	if (vi->stats == NULL)
 		goto free;
 
-	vi->vq_index = alloc_percpu(int);
-	if (vi->vq_index == NULL)
-		goto free_stats;
-
 	mutex_init(&vi->config_lock);
 	vi->config_enable = true;
 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
@@ -1589,7 +1548,7 @@ static int virtnet_probe(struct virtio_device *vdev)
 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
 	err = init_vqs(vi);
 	if (err)
-		goto free_index;
+		goto free_stats;
 
 	netif_set_real_num_tx_queues(dev, 1);
 	netif_set_real_num_rx_queues(dev, 1);
@@ -1640,8 +1599,6 @@ free_recv_bufs:
 free_vqs:
 	cancel_delayed_work_sync(&vi->refill);
 	virtnet_del_vqs(vi);
-free_index:
-	free_percpu(vi->vq_index);
 free_stats:
 	free_percpu(vi->stats);
 free:
@@ -1678,7 +1635,6 @@ static void virtnet_remove(struct virtio_device *vdev)
 
 	flush_work(&vi->config_work);
 
-	free_percpu(vi->vq_index);
 	free_percpu(vi->stats);
 	free_netdev(vi->dev);
 }
Signature · 1 line
1.7.1
maintainer Acked-by on v2 1/1
#1 lore raw
Quoted 5 lines · Jason Wang

We used to use a percpu structure vq_index to record the cpu to queue mapping, this is suboptimal since it duplicates the work of XPS and loses all other XPS functionality such as allowing use to configure their own transmission steering strategy.

Quoted 8 lines

So this patch switches to use XPS and suggest a default mapping when the number of cpus is equal to the number of queues. With XPS support, there's no need for keeping per-cpu vq_index and .ndo_select_queue(), so they were removed also.

Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Quoted 131 lines

---
Changes from V1:
- use cpumask_of() instead of allocate dynamically

 drivers/net/virtio_net.c |   48 +--------------------------------------------
 1 files changed, 2 insertions(+), 46 deletions(-)
drivers/net/virtio_net.c +2 -46
index defec2b..4eca652 100644
@@ -127,9 +127,6 @@ struct virtnet_info {
 	/* Does the affinity hint is set for virtqueues? */
 	bool affinity_hint_set;
 
-	/* Per-cpu variable to show the mapping from CPU to virtqueue */
-	int __percpu *vq_index;
-
 	/* CPU hot plug notifier */
 	struct notifier_block nb;
 };
@@ -1063,7 +1060,6 @@ static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
 {
 	int i;
-	int cpu;
 
 	if (vi->affinity_hint_set) {
 		for (i = 0; i < vi->max_queue_pairs; i++) {
@@ -1073,16 +1069,6 @@ static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
 
 		vi->affinity_hint_set = false;
 	}
-
-	i = 0;
-	for_each_online_cpu(cpu) {
-		if (cpu == hcpu) {
-			*per_cpu_ptr(vi->vq_index, cpu) = -1;
-		} else {
-			*per_cpu_ptr(vi->vq_index, cpu) =
-				++i % vi->curr_queue_pairs;
-		}
-	}
 }
 
 static void virtnet_set_affinity(struct virtnet_info *vi)
@@ -1104,7 +1090,7 @@ static void virtnet_set_affinity(struct virtnet_info *vi)
 	for_each_online_cpu(cpu) {
 		virtqueue_set_affinity(vi->rq[i].vq, cpu);
 		virtqueue_set_affinity(vi->sq[i].vq, cpu);
-		*per_cpu_ptr(vi->vq_index, cpu) = i;
+		netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
 		i++;
 	}
 
@@ -1217,28 +1203,6 @@ static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
 	return 0;
 }
 
-/* To avoid contending a lock hold by a vcpu who would exit to host, select the
- * txq based on the processor id.
- */
-static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
-{
-	int txq;
-	struct virtnet_info *vi = netdev_priv(dev);
-
-	if (skb_rx_queue_recorded(skb)) {
-		txq = skb_get_rx_queue(skb);
-	} else {
-		txq = *__this_cpu_ptr(vi->vq_index);
-		if (txq == -1)
-			txq = 0;
-	}
-
-	while (unlikely(txq >= dev->real_num_tx_queues))
-		txq -= dev->real_num_tx_queues;
-
-	return txq;
-}
-
 static const struct net_device_ops virtnet_netdev = {
 	.ndo_open            = virtnet_open,
 	.ndo_stop   	     = virtnet_close,
@@ -1250,7 +1214,6 @@ static const struct net_device_ops virtnet_netdev = {
 	.ndo_get_stats64     = virtnet_stats,
 	.ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
 	.ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
-	.ndo_select_queue     = virtnet_select_queue,
 #ifdef CONFIG_NET_POLL_CONTROLLER
 	.ndo_poll_controller = virtnet_netpoll,
 #endif
@@ -1559,10 +1522,6 @@ static int virtnet_probe(struct virtio_device *vdev)
 	if (vi->stats == NULL)
 		goto free;
 
-	vi->vq_index = alloc_percpu(int);
-	if (vi->vq_index == NULL)
-		goto free_stats;
-
 	mutex_init(&vi->config_lock);
 	vi->config_enable = true;
 	INIT_WORK(&vi->config_work, virtnet_config_changed_work);
@@ -1589,7 +1548,7 @@ static int virtnet_probe(struct virtio_device *vdev)
 	/* Allocate/initialize the rx/tx queues, and invoke find_vqs */
 	err = init_vqs(vi);
 	if (err)
-		goto free_index;
+		goto free_stats;
 
 	netif_set_real_num_tx_queues(dev, 1);
 	netif_set_real_num_rx_queues(dev, 1);
@@ -1640,8 +1599,6 @@ free_recv_bufs:
 free_vqs:
 	cancel_delayed_work_sync(&vi->refill);
 	virtnet_del_vqs(vi);
-free_index:
-	free_percpu(vi->vq_index);
 free_stats:
 	free_percpu(vi->stats);
 free:
@@ -1678,7 +1635,6 @@ static void virtnet_remove(struct virtio_device *vdev)
 
 	flush_work(&vi->config_work);
 
-	free_percpu(vi->vq_index);
 	free_percpu(vi->stats);
 	free_netdev(vi->dev);
 }
Signature · 1 line
1.7.1
Acked-by
#2 lore raw
Quoted 11 lines · Michael S. Tsirkin
Quoted 10 lines · Jason Wang

We used to use a percpu structure vq_index to record the cpu to queue mapping, this is suboptimal since it duplicates the work of XPS and loses all other XPS functionality such as allowing use to configure their own transmission steering strategy.

So this patch switches to use XPS and suggest a default mapping when the number of cpus is equal to the number of queues. With XPS support, there's no need for keeping per-cpu vq_index and .ndo_select_queue(), so they were removed also.

Quoted 5 lines
Quoted 3 lines
Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>

Dave, please apply.

Cheers,
Rusty.

maintainer
#3 lore raw
In reply to Jason Wang

From: Jason Wang <[email protected]>
Date: Mon, 30 Sep 2013 15:37:17 +0800

Quoted 10 lines

We used to use a percpu structure vq_index to record the cpu to queue mapping, this is suboptimal since it duplicates the work of XPS and loses all other XPS functionality such as allowing use to configure their own transmission steering strategy.

So this patch switches to use XPS and suggest a default mapping when the number of cpus is equal to the number of queues. With XPS support, there's no need for keeping per-cpu vq_index and .ndo_select_queue(), so they were removed also.

Quoted 6 lines
Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>

---
Changes from V1:
- use cpumask_of() instead of allocate dynamically

This generates build warnings:

drivers/net/virtio_net.c: In function ‘virtnet_set_affinity’:
drivers/net/virtio_net.c:1093:3: warning: passing argument 2 of ‘netif_set_xps_queue’ discards ‘const’ qualifier from pointer target type [enabled by default] In file included from drivers/net/virtio_net.c:20:0:
include/linux/netdevice.h:2275:5: note: expected ‘struct cpumask *’ but argument is of type ‘const struct cpumask *’ �τθΊ{.nΗ+‰·Ÿ�‰­†+%ŠΛ�±ιέΆ�Šw�Ί{.nΗ+‰·�Š{±ώG«ι�Š{ayΊΚ‡Ϊ™λ,j­’f£’·hšοκ�‘κηz_θ�(­ιšŽŠέ’j"ϊΆm§��Ύ«ώG«ι�’Έ?™¨θ­Ϊ&£ψ§~αΆiO•ζ¬z·švΨ^Άm§��Γ �Άμ�’Έ?–I�

#4 lore raw
Quoted 20 lines · David Miller

From: Jason Wang <[email protected]>
Date: Mon, 30 Sep 2013 15:37:17 +0800

Quoted 10 lines

We used to use a percpu structure vq_index to record the cpu to queue mapping, this is suboptimal since it duplicates the work of XPS and loses all other XPS functionality such as allowing use to configure their own transmission steering strategy.

So this patch switches to use XPS and suggest a default mapping when the number of cpus is equal to the number of queues. With XPS support, there's no need for keeping per-cpu vq_index and .ndo_select_queue(), so they were removed also.

Quoted 6 lines
Cc Rusty Russell <[email protected]>
Cc Michael S. Tsirkin <[email protected]>

---
Changes from V1:
- use cpumask_of() instead of allocate dynamically

Quoted 6 lines

This generates build warnings:

drivers/net/virtio_net.c: In function ‘virtnet_set_affinity’:
drivers/net/virtio_net.c:1093:3: warning: passing argument 2 of ‘netif_set_xps_queue’ discards ‘const’ qualifier from pointer target type [enabled by default] In file included from drivers/net/virtio_net.c:20:0:
include/linux/netdevice.h:2275:5: note: expected ‘struct cpumask *’ but argument is of type ‘const struct cpumask *’

Looks like we just need to tweak the signature there to const struct cpumask *: it's never changed.

I just sent a patch
[PATCH] netif_set_xps_queue: make cpu mask const
that does exactly that.

Signature · 1 line
MST