virtio: Recycle unused recv buffer pages for large skbs in net driver

If we hack the virtio_net driver to always allocate full-sized (64k+)
skbuffs, the driver slows down (lguest numbers):

  Time to receive 1GB (small buffers): 10.85 seconds
  Time to receive 1GB (64k+ buffers): 24.75 seconds

Of course, large buffers use up more space in the ring, so we increase
that from 128 to 2048:

  Time to receive 1GB (64k+ buffers, 2k ring): 16.61 seconds

If we recycle pages rather than using alloc_page/free_page:

  Time to receive 1GB (64k+ buffers, 2k ring, recycle pages): 10.81 seconds

This demonstrates that with efficient allocation, we don't need to
have a separate "small buffer" queue.

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 0886b8a2d92dc5dda7e75dfb70f41286df5c6140..0196a0df90210995ec6a5abe228eb23bb22d494c 100644 (file)
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -61,6 +61,9 @@ struct virtnet_info
        /* Receive & send queues. */
        struct sk_buff_head recv;
        struct sk_buff_head send;
+
+       /* Chain pages by the private ptr. */
+       struct page *pages;
 };
 
 static inline struct virtio_net_hdr *skb_vnet_hdr(struct sk_buff *skb)
@@ -73,6 +76,23 @@ static inline void vnet_hdr_to_sg(struct scatterlist *sg, struct sk_buff *skb)
        sg_init_one(sg, skb_vnet_hdr(skb), sizeof(struct virtio_net_hdr));
 }
 
+static void give_a_page(struct virtnet_info *vi, struct page *page)
+{
+       page->private = (unsigned long)vi->pages;
+       vi->pages = page;
+}
+
+static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
+{
+       struct page *p = vi->pages;
+
+       if (p)
+               vi->pages = (struct page *)p->private;
+       else
+               p = alloc_page(gfp_mask);
+       return p;
+}
+
 static void skb_xmit_done(struct virtqueue *svq)
 {
        struct virtnet_info *vi = svq->vdev->priv;
@@ -101,6 +121,15 @@ static void receive_skb(struct net_device *dev, struct sk_buff *skb,
        }
        len -= sizeof(struct virtio_net_hdr);
 
+       if (len <= MAX_PACKET_LEN) {
+               unsigned int i;
+
+               for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
+                       give_a_page(dev->priv, skb_shinfo(skb)->frags[i].page);
+               skb->data_len = 0;
+               skb_shinfo(skb)->nr_frags = 0;
+       }
+
        err = pskb_trim(skb, len);
        if (err) {
                pr_debug("%s: pskb_trim failed %i %d\n", dev->name, len, err);
@@ -183,7 +212,7 @@ static void try_fill_recv(struct virtnet_info *vi)
                if (vi->big_packets) {
                        for (i = 0; i < MAX_SKB_FRAGS; i++) {
                                skb_frag_t *f = &skb_shinfo(skb)->frags[i];
-                               f->page = alloc_page(GFP_ATOMIC);
+                               f->page = get_a_page(vi, GFP_ATOMIC);
                                if (!f->page)
                                        break;
 
@@ -506,6 +535,7 @@ static int virtnet_probe(struct virtio_device *vdev)
        vi->dev = dev;
        vi->vdev = vdev;
        vdev->priv = vi;
+       vi->pages = NULL;
 
        /* If they give us a callback when all buffers are done, we don't need
         * the timer. */
@@ -591,6 +621,10 @@ static void virtnet_remove(struct virtio_device *vdev)
        vdev->config->del_vq(vi->svq);
        vdev->config->del_vq(vi->rvq);
        unregister_netdev(vi->dev);
+
+       while (vi->pages)
+               __free_pages(get_a_page(vi, GFP_KERNEL), 0);
+
        free_netdev(vi->dev);
 }