macvtap: switch to use skb array
authorJason Wang <jasowang@redhat.com>
Fri, 15 Jul 2016 07:46:31 +0000 (03:46 -0400)
committerDavid S. Miller <davem@davemloft.net>
Fri, 15 Jul 2016 21:40:39 +0000 (14:40 -0700)
This patch switch to use skb array instead of sk_receive_queue to
avoid spinlock contentions. Tests shows about 21% improvements for
guest rx pps:

Before: 1472731 pkts/s
After:  1786289 pkts/s

Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/macvtap.c

index 2476923..9204d19 100644 (file)
@@ -21,6 +21,7 @@
 #include <net/rtnetlink.h>
 #include <net/sock.h>
 #include <linux/virtio_net.h>
+#include <linux/skb_array.h>
 
 /*
  * A macvtap queue is the central object of this driver, it connects
@@ -43,6 +44,7 @@ struct macvtap_queue {
        u16 queue_index;
        bool enabled;
        struct list_head next;
+       struct skb_array skb_array;
 };
 
 #define MACVTAP_FEATURES (IFF_VNET_HDR | IFF_MULTI_QUEUE)
@@ -273,6 +275,7 @@ static void macvtap_put_queue(struct macvtap_queue *q)
        rtnl_unlock();
 
        synchronize_rcu();
+       skb_array_cleanup(&q->skb_array);
        sock_put(&q->sk);
 }
 
@@ -366,7 +369,7 @@ static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
        if (!q)
                return RX_HANDLER_PASS;
 
-       if (skb_queue_len(&q->sk.sk_receive_queue) >= dev->tx_queue_len)
+       if (__skb_array_full(&q->skb_array))
                goto drop;
 
        skb_push(skb, ETH_HLEN);
@@ -384,7 +387,8 @@ static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
                        goto drop;
 
                if (!segs) {
-                       skb_queue_tail(&q->sk.sk_receive_queue, skb);
+                       if (skb_array_produce(&q->skb_array, skb))
+                               goto drop;
                        goto wake_up;
                }
 
@@ -393,7 +397,11 @@ static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
                        struct sk_buff *nskb = segs->next;
 
                        segs->next = NULL;
-                       skb_queue_tail(&q->sk.sk_receive_queue, segs);
+                       if (skb_array_produce(&q->skb_array, segs)) {
+                               kfree_skb(segs);
+                               kfree_skb_list(nskb);
+                               break;
+                       }
                        segs = nskb;
                }
        } else {
@@ -406,7 +414,8 @@ static rx_handler_result_t macvtap_handle_frame(struct sk_buff **pskb)
                    !(features & NETIF_F_CSUM_MASK) &&
                    skb_checksum_help(skb))
                        goto drop;
-               skb_queue_tail(&q->sk.sk_receive_queue, skb);
+               if (skb_array_produce(&q->skb_array, skb))
+                       goto drop;
        }
 
 wake_up:
@@ -523,7 +532,11 @@ static void macvtap_sock_write_space(struct sock *sk)
 
 static void macvtap_sock_destruct(struct sock *sk)
 {
-       skb_queue_purge(&sk->sk_receive_queue);
+       struct macvtap_queue *q = container_of(sk, struct macvtap_queue, sk);
+       struct sk_buff *skb;
+
+       while ((skb = skb_array_consume(&q->skb_array)) != NULL)
+               kfree(skb);
 }
 
 static int macvtap_open(struct inode *inode, struct file *file)
@@ -536,13 +549,13 @@ static int macvtap_open(struct inode *inode, struct file *file)
        rtnl_lock();
        dev = dev_get_by_macvtap_minor(iminor(inode));
        if (!dev)
-               goto out;
+               goto err;
 
        err = -ENOMEM;
        q = (struct macvtap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
                                             &macvtap_proto, 0);
        if (!q)
-               goto out;
+               goto err;
 
        RCU_INIT_POINTER(q->sock.wq, &q->wq);
        init_waitqueue_head(&q->wq.wait);
@@ -566,11 +579,24 @@ static int macvtap_open(struct inode *inode, struct file *file)
        if ((dev->features & NETIF_F_HIGHDMA) && (dev->features & NETIF_F_SG))
                sock_set_flag(&q->sk, SOCK_ZEROCOPY);
 
+       err = -ENOMEM;
+       if (skb_array_init(&q->skb_array, dev->tx_queue_len, GFP_KERNEL))
+               goto err_array;
+
        err = macvtap_set_queue(dev, file, q);
        if (err)
-               sock_put(&q->sk);
+               goto err_queue;
 
-out:
+       dev_put(dev);
+
+       rtnl_unlock();
+       return err;
+
+err_queue:
+       skb_array_cleanup(&q->skb_array);
+err_array:
+       sock_put(&q->sk);
+err:
        if (dev)
                dev_put(dev);
 
@@ -596,7 +622,7 @@ static unsigned int macvtap_poll(struct file *file, poll_table * wait)
        mask = 0;
        poll_wait(file, &q->wq.wait, wait);
 
-       if (!skb_queue_empty(&q->sk.sk_receive_queue))
+       if (!skb_array_empty(&q->skb_array))
                mask |= POLLIN | POLLRDNORM;
 
        if (sock_writeable(&q->sk) ||
@@ -856,7 +882,7 @@ static ssize_t macvtap_do_read(struct macvtap_queue *q,
                                        TASK_INTERRUPTIBLE);
 
                /* Read frames from the queue */
-               skb = skb_dequeue(&q->sk.sk_receive_queue);
+               skb = skb_array_consume(&q->skb_array);
                if (skb)
                        break;
                if (noblock) {
@@ -1180,10 +1206,18 @@ static int macvtap_recvmsg(struct socket *sock, struct msghdr *m,
        return ret;
 }
 
+static int macvtap_peek_len(struct socket *sock)
+{
+       struct macvtap_queue *q = container_of(sock, struct macvtap_queue,
+                                              sock);
+       return skb_array_peek_len(&q->skb_array);
+}
+
 /* Ops structure to mimic raw sockets with tun */
 static const struct proto_ops macvtap_socket_ops = {
        .sendmsg = macvtap_sendmsg,
        .recvmsg = macvtap_recvmsg,
+       .peek_len = macvtap_peek_len,
 };
 
 /* Get an underlying socket object from tun file.  Returns error unless file is
@@ -1202,6 +1236,28 @@ struct socket *macvtap_get_socket(struct file *file)
 }
 EXPORT_SYMBOL_GPL(macvtap_get_socket);
 
+static int macvtap_queue_resize(struct macvlan_dev *vlan)
+{
+       struct net_device *dev = vlan->dev;
+       struct macvtap_queue *q;
+       struct skb_array **arrays;
+       int n = vlan->numqueues;
+       int ret, i = 0;
+
+       arrays = kmalloc(sizeof *arrays * n, GFP_KERNEL);
+       if (!arrays)
+               return -ENOMEM;
+
+       list_for_each_entry(q, &vlan->queue_list, next)
+               arrays[i++] = &q->skb_array;
+
+       ret = skb_array_resize_multiple(arrays, n,
+                                       dev->tx_queue_len, GFP_KERNEL);
+
+       kfree(arrays);
+       return ret;
+}
+
 static int macvtap_device_event(struct notifier_block *unused,
                                unsigned long event, void *ptr)
 {
@@ -1249,6 +1305,10 @@ static int macvtap_device_event(struct notifier_block *unused,
                device_destroy(&macvtap_class, devt);
                macvtap_free_minor(vlan);
                break;
+       case NETDEV_CHANGE_TX_QUEUE_LEN:
+               if (macvtap_queue_resize(vlan))
+                       return NOTIFY_BAD;
+               break;
        }
 
        return NOTIFY_DONE;