Linux Virtualization - Jan 2013 - [PATCH V8 1/3] virtio-net: fix the set affinity bug when CPU IDs are not consecutive

As Michael mentioned, set affinity and select queue will not work very
well when CPU IDs are not consecutive, this can happen with hot unplug.
Fix this bug by traversal the online CPUs, and create a per cpu variable
to find the mapping from CPU to the preferable virtual-queue.

Cc: Rusty Russell <rusty at rustcorp.com.au>
Cc: "Michael S. Tsirkin" <mst at redhat.com>
Cc: Jason Wang <jasowang at redhat.com>
Cc: Eric Dumazet <erdnetdev at gmail.com>
Cc: "David S. Miller" <davem at davemloft.net>
Cc: virtualization at lists.linux-foundation.org
Cc: netdev at vger.kernel.org
Signed-off-by: Wanlong Gao <gaowanlong at cn.fujitsu.com>
Acked-by: Michael S. Tsirkin <mst at redhat.com>
---
V7->V8:
	set vq index unconditionally, make bisect happy,
		make sure don't introduce bug. (Jason)
V6->V7:
	serialize virtnet_set_queues to avoid a race with cpu hotplug (Jason)
V5->V6:
	remove {get|put}_online_cpus from virtnet_del_vqs (Jason)
V4->V5:
	Add get/put_online_cpus to avoid CPUs go up and down during operations (Rusty)

V3->V4:
	move vq_index into virtnet_info	(Jason)
	change the mapping value when not setting affinity (Jason)
	address the comments about select_queue (Rusty)

 drivers/net/virtio_net.c | 67 ++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 54 insertions(+), 13 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index a6fcf15..fda214a 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -123,6 +123,9 @@ 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;
 };
 
 struct skb_vnet_hdr {
@@ -1016,6 +1019,7 @@ static int virtnet_vlan_rx_kill_vid(struct net_device
*dev, u16 vid)
 static void virtnet_set_affinity(struct virtnet_info *vi, bool set)
 {
 	int i;
+	int cpu;
 
 	/* In multiqueue mode, when the number of cpu is equal to the number of
 	 * queue pairs, we let the queue pairs to be private to one cpu by
@@ -1023,22 +1027,40 @@ static void virtnet_set_affinity(struct virtnet_info
*vi, bool set)
 	 */
 	if ((vi->curr_queue_pairs == 1 ||
 	     vi->max_queue_pairs != num_online_cpus()) && set) {
-		if (vi->affinity_hint_set)
+		if (vi->affinity_hint_set) {
 			set = false;
-		else
+		} else {
+			i = 0;
+			for_each_online_cpu(cpu)
+				*per_cpu_ptr(vi->vq_index, cpu) +					++i % vi->curr_queue_pairs;
 			return;
+		}
 	}
 
-	for (i = 0; i < vi->max_queue_pairs; i++) {
-		int cpu = set ? i : -1;
-		virtqueue_set_affinity(vi->rq[i].vq, cpu);
-		virtqueue_set_affinity(vi->sq[i].vq, cpu);
-	}
+	if (set) {
+		i = 0;
+		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;
+			i++;
+		}
 
-	if (set)
 		vi->affinity_hint_set = true;
-	else
+	} else {
+		for(i = 0; i < vi->max_queue_pairs; i++) {
+			virtqueue_set_affinity(vi->rq[i].vq, -1);
+			virtqueue_set_affinity(vi->sq[i].vq, -1);
+		}
+
+		i = 0;
+		for_each_online_cpu(cpu)
+			*per_cpu_ptr(vi->vq_index, cpu) +				++i % vi->curr_queue_pairs;
+
 		vi->affinity_hint_set = false;
+	}
 }
 
 static void virtnet_get_ringparam(struct net_device *dev,
@@ -1082,6 +1104,7 @@ static int virtnet_set_channels(struct net_device *dev,
 	if (queue_pairs > vi->max_queue_pairs)
 		return -EINVAL;
 
+	get_online_cpus();
 	err = virtnet_set_queues(vi, queue_pairs);
 	if (!err) {
 		netif_set_real_num_tx_queues(dev, queue_pairs);
@@ -1089,6 +1112,7 @@ static int virtnet_set_channels(struct net_device *dev,
 
 		virtnet_set_affinity(vi, true);
 	}
+	put_online_cpus();
 
 	return err;
 }
@@ -1127,12 +1151,19 @@ static int virtnet_change_mtu(struct net_device *dev,
int new_mtu)
 
 /* To avoid contending a lock hold by a vcpu who would exit to host, select the
  * txq based on the processor id.
- * TODO: handle cpu hotplug.
  */
 static u16 virtnet_select_queue(struct net_device *dev, struct sk_buff *skb)
 {
-	int txq = skb_rx_queue_recorded(skb) ? skb_get_rx_queue(skb) :
-		  smp_processor_id();
+	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;
@@ -1371,7 +1402,10 @@ static int init_vqs(struct virtnet_info *vi)
 	if (ret)
 		goto err_free;
 
+	get_online_cpus();
 	virtnet_set_affinity(vi, true);
+	put_online_cpus();
+
 	return 0;
 
 err_free:
@@ -1453,6 +1487,10 @@ 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);
@@ -1476,7 +1514,7 @@ 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_stats;
+		goto free_index;
 
 	netif_set_real_num_tx_queues(dev, 1);
 	netif_set_real_num_rx_queues(dev, 1);
@@ -1520,6 +1558,8 @@ 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:
@@ -1554,6 +1594,7 @@ 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);
 }
-- 
1.8.1

Split out the clean affinity function to virtnet_clean_affinity().

Cc: Rusty Russell <rusty at rustcorp.com.au>
Cc: "Michael S. Tsirkin" <mst at redhat.com>
Cc: Jason Wang <jasowang at redhat.com>
Cc: Eric Dumazet <erdnetdev at gmail.com>
Cc: "David S. Miller" <davem at davemloft.net>
Cc: virtualization at lists.linux-foundation.org
Cc: netdev at vger.kernel.org
Signed-off-by: Wanlong Gao <gaowanlong at cn.fujitsu.com>
Acked-by: Michael S. Tsirkin <mst at redhat.com>
---
V6->V7:
	always set vq_index (Jason)
V5->V6: NEW

 drivers/net/virtio_net.c | 76 +++++++++++++++++++++++++-----------------------
 1 file changed, 40 insertions(+), 36 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index fda214a..2f6fe9b 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -1016,51 +1016,55 @@ static int virtnet_vlan_rx_kill_vid(struct net_device
*dev, u16 vid)
 	return 0;
 }
 
-static void virtnet_set_affinity(struct virtnet_info *vi, bool set)
+static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
 {
 	int i;
 	int cpu;
 
-	/* In multiqueue mode, when the number of cpu is equal to the number of
-	 * queue pairs, we let the queue pairs to be private to one cpu by
-	 * setting the affinity hint to eliminate the contention.
-	 */
-	if ((vi->curr_queue_pairs == 1 ||
-	     vi->max_queue_pairs != num_online_cpus()) && set) {
-		if (vi->affinity_hint_set) {
-			set = false;
-		} else {
-			i = 0;
-			for_each_online_cpu(cpu)
-				*per_cpu_ptr(vi->vq_index, cpu) -					++i % vi->curr_queue_pairs;
-			return;
-		}
-	}
-
-	if (set) {
-		i = 0;
-		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;
-			i++;
-		}
-
-		vi->affinity_hint_set = true;
-	} else {
-		for(i = 0; i < vi->max_queue_pairs; i++) {
+	if (vi->affinity_hint_set) {
+		for (i = 0; i < vi->max_queue_pairs; i++) {
 			virtqueue_set_affinity(vi->rq[i].vq, -1);
 			virtqueue_set_affinity(vi->sq[i].vq, -1);
 		}
 
-		i = 0;
-		for_each_online_cpu(cpu)
+		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;
+		}
+	}
+}
 
-		vi->affinity_hint_set = false;
+static void virtnet_set_affinity(struct virtnet_info *vi)
+{
+	int i;
+	int cpu;
+
+	/* In multiqueue mode, when the number of cpu is equal to the number of
+	 * queue pairs, we let the queue pairs to be private to one cpu by
+	 * setting the affinity hint to eliminate the contention.
+	 */
+	if (vi->curr_queue_pairs == 1 ||
+	    vi->max_queue_pairs != num_online_cpus()) {
+		virtnet_clean_affinity(vi, -1);
+		return;
 	}
+
+	i = 0;
+	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;
+		i++;
+	}
+
+	vi->affinity_hint_set = true;
 }
 
 static void virtnet_get_ringparam(struct net_device *dev,
@@ -1110,7 +1114,7 @@ static int virtnet_set_channels(struct net_device *dev,
 		netif_set_real_num_tx_queues(dev, queue_pairs);
 		netif_set_real_num_rx_queues(dev, queue_pairs);
 
-		virtnet_set_affinity(vi, true);
+		virtnet_set_affinity(vi);
 	}
 	put_online_cpus();
 
@@ -1279,7 +1283,7 @@ static void virtnet_del_vqs(struct virtnet_info *vi)
 {
 	struct virtio_device *vdev = vi->vdev;
 
-	virtnet_set_affinity(vi, false);
+	virtnet_clean_affinity(vi, -1);
 
 	vdev->config->del_vqs(vdev);
 
@@ -1403,7 +1407,7 @@ static int init_vqs(struct virtnet_info *vi)
 		goto err_free;
 
 	get_online_cpus();
-	virtnet_set_affinity(vi, true);
+	virtnet_set_affinity(vi);
 	put_online_cpus();
 
 	return 0;
-- 
1.8.1

Add a cpu notifier to virtio-net, so that we can reset the
virtqueue affinity if the cpu hotplug happens. It improve
the performance through enabling or disabling the virtqueue
affinity after doing cpu hotplug.

Cc: Rusty Russell <rusty at rustcorp.com.au>
Cc: "Michael S. Tsirkin" <mst at redhat.com>
Cc: Jason Wang <jasowang at redhat.com>
Cc: Eric Dumazet <erdnetdev at gmail.com>
Cc: "David S. Miller" <davem at davemloft.net>
Cc: virtualization at lists.linux-foundation.org
Cc: netdev at vger.kernel.org
Signed-off-by: Wanlong Gao <gaowanlong at cn.fujitsu.com>
Acked-by: Michael S. Tsirkin <mst at redhat.com>
---
V5->V6:
	deal with CPU_DOWN_PREPARE separately by just cleaning affinity
V4->V5:
	New method to deal with the cpu hotplug actions (Rusty)

 drivers/net/virtio_net.c | 33 +++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 2f6fe9b..35c00c5 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -26,6 +26,7 @@
 #include <linux/scatterlist.h>
 #include <linux/if_vlan.h>
 #include <linux/slab.h>
+#include <linux/cpu.h>
 
 static int napi_weight = 128;
 module_param(napi_weight, int, 0444);
@@ -126,6 +127,9 @@ struct virtnet_info {
 
 	/* Per-cpu variable to show the mapping from CPU to virtqueue */
 	int __percpu *vq_index;
+
+	/* CPU hot plug notifier */
+	struct notifier_block nb;
 };
 
 struct skb_vnet_hdr {
@@ -1067,6 +1071,26 @@ static void virtnet_set_affinity(struct virtnet_info *vi)
 	vi->affinity_hint_set = true;
 }
 
+static int virtnet_cpu_callback(struct notifier_block *nfb,
+			        unsigned long action, void *hcpu)
+{
+	struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
+
+	switch(action & ~CPU_TASKS_FROZEN) {
+	case CPU_ONLINE:
+	case CPU_DOWN_FAILED:
+	case CPU_DEAD:
+		virtnet_set_affinity(vi);
+		break;
+	case CPU_DOWN_PREPARE:
+		virtnet_clean_affinity(vi, (long)hcpu);
+		break;
+	default:
+		break;
+	}
+	return NOTIFY_OK;
+}
+
 static void virtnet_get_ringparam(struct net_device *dev,
 				struct ethtool_ringparam *ring)
 {
@@ -1541,6 +1565,13 @@ static int virtnet_probe(struct virtio_device *vdev)
 		}
 	}
 
+	vi->nb.notifier_call = &virtnet_cpu_callback;
+	err = register_hotcpu_notifier(&vi->nb);
+	if (err) {
+		pr_debug("virtio_net: registering cpu notifier failed\n");
+		goto free_recv_bufs;
+	}
+
 	/* Assume link up if device can't report link status,
 	   otherwise get link status from config. */
 	if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
@@ -1587,6 +1618,8 @@ static void virtnet_remove(struct virtio_device *vdev)
 {
 	struct virtnet_info *vi = vdev->priv;
 
+	unregister_hotcpu_notifier(&vi->nb);
+
 	/* Prevent config work handler from accessing the device. */
 	mutex_lock(&vi->config_lock);
 	vi->config_enable = false;
-- 
1.8.1

From: Wanlong Gao <gaowanlong at cn.fujitsu.com>
Date: Fri, 25 Jan 2013 17:51:29 +0800
> As Michael mentioned, set affinity and select queue will not work very
> well when CPU IDs are not consecutive, this can happen with hot unplug.
> Fix this bug by traversal the online CPUs, and create a per cpu variable
> to find the mapping from CPU to the preferable virtual-queue.
> 
> Cc: Rusty Russell <rusty at rustcorp.com.au>
> Cc: "Michael S. Tsirkin" <mst at redhat.com>
> Cc: Jason Wang <jasowang at redhat.com>
> Cc: Eric Dumazet <erdnetdev at gmail.com>
> Cc: "David S. Miller" <davem at davemloft.net>
> Cc: virtualization at lists.linux-foundation.org
> Cc: netdev at vger.kernel.org
> Signed-off-by: Wanlong Gao <gaowanlong at cn.fujitsu.com>
> Acked-by: Michael S. Tsirkin <mst at redhat.com>
Applied.

From: Wanlong Gao <gaowanlong at cn.fujitsu.com>
Date: Fri, 25 Jan 2013 17:51:30 +0800
> Split out the clean affinity function to virtnet_clean_affinity().
> 
> Cc: Rusty Russell <rusty at rustcorp.com.au>
> Cc: "Michael S. Tsirkin" <mst at redhat.com>
> Cc: Jason Wang <jasowang at redhat.com>
> Cc: Eric Dumazet <erdnetdev at gmail.com>
> Cc: "David S. Miller" <davem at davemloft.net>
> Cc: virtualization at lists.linux-foundation.org
> Cc: netdev at vger.kernel.org
> Signed-off-by: Wanlong Gao <gaowanlong at cn.fujitsu.com>
> Acked-by: Michael S. Tsirkin <mst at redhat.com>
Applied.

From: Wanlong Gao <gaowanlong at cn.fujitsu.com>
Date: Fri, 25 Jan 2013 17:51:31 +0800
> Add a cpu notifier to virtio-net, so that we can reset the
> virtqueue affinity if the cpu hotplug happens. It improve
> the performance through enabling or disabling the virtqueue
> affinity after doing cpu hotplug.
> 
> Cc: Rusty Russell <rusty at rustcorp.com.au>
> Cc: "Michael S. Tsirkin" <mst at redhat.com>
> Cc: Jason Wang <jasowang at redhat.com>
> Cc: Eric Dumazet <erdnetdev at gmail.com>
> Cc: "David S. Miller" <davem at davemloft.net>
> Cc: virtualization at lists.linux-foundation.org
> Cc: netdev at vger.kernel.org
> Signed-off-by: Wanlong Gao <gaowanlong at cn.fujitsu.com>
> Acked-by: Michael S. Tsirkin <mst at redhat.com>
Applied.