virtio-net: put virtio net header inline with data
authorMichael S. Tsirkin <mst@redhat.com>
Thu, 25 Jul 2013 00:50:23 +0000 (10:20 +0930)
committerDavid S. Miller <davem@davemloft.net>
Sun, 28 Jul 2013 03:06:10 +0000 (20:06 -0700)
For small packets we can simplify xmit processing
by linearizing buffers with the header:
most packets seem to have enough head room
we can use for this purpose.
Since existing hypervisors require that header
is the first s/g element, we need a feature bit
for this.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/virtio_net.c
include/uapi/linux/virtio_net.h

index 3d2a90a..f216002 100644 (file)
@@ -106,6 +106,9 @@ struct virtnet_info {
        /* Has control virtqueue */
        bool has_cvq;
 
+       /* Host can handle any s/g split between our header and packet data */
+       bool any_header_sg;
+
        /* enable config space updates */
        bool config_enable;
 
@@ -669,12 +672,28 @@ static void free_old_xmit_skbs(struct send_queue *sq)
 
 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
 {
-       struct skb_vnet_hdr *hdr = skb_vnet_hdr(skb);
+       struct skb_vnet_hdr *hdr;
        const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
        struct virtnet_info *vi = sq->vq->vdev->priv;
        unsigned num_sg;
+       unsigned hdr_len;
+       bool can_push;
 
        pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
+       if (vi->mergeable_rx_bufs)
+               hdr_len = sizeof hdr->mhdr;
+       else
+               hdr_len = sizeof hdr->hdr;
+
+       can_push = vi->any_header_sg &&
+               !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
+               !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
+       /* Even if we can, don't push here yet as this would skew
+        * csum_start offset below. */
+       if (can_push)
+               hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
+       else
+               hdr = skb_vnet_hdr(skb);
 
        if (skb->ip_summed == CHECKSUM_PARTIAL) {
                hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
@@ -703,15 +722,18 @@ static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
                hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
        }
 
-       hdr->mhdr.num_buffers = 0;
-
-       /* Encode metadata header at front. */
        if (vi->mergeable_rx_bufs)
-               sg_set_buf(sq->sg, &hdr->mhdr, sizeof hdr->mhdr);
-       else
-               sg_set_buf(sq->sg, &hdr->hdr, sizeof hdr->hdr);
+               hdr->mhdr.num_buffers = 0;
 
-       num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
+       if (can_push) {
+               __skb_push(skb, hdr_len);
+               num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
+               /* Pull header back to avoid skew in tx bytes calculations. */
+               __skb_pull(skb, hdr_len);
+       } else {
+               sg_set_buf(sq->sg, hdr, hdr_len);
+               num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
+       }
        return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
 }
 
@@ -1552,6 +1574,9 @@ static int virtnet_probe(struct virtio_device *vdev)
        if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
                vi->mergeable_rx_bufs = true;
 
+       if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
+               vi->any_header_sg = true;
+
        if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
                vi->has_cvq = true;
 
@@ -1727,6 +1752,7 @@ static unsigned int features[] = {
        VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
        VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
        VIRTIO_NET_F_CTRL_MAC_ADDR,
+       VIRTIO_F_ANY_LAYOUT,
 };
 
 static struct virtio_driver virtio_net_driver = {
index c520203..227d4ce 100644 (file)
@@ -70,7 +70,9 @@ struct virtio_net_config {
        __u16 max_virtqueue_pairs;
 } __attribute__((packed));
 
-/* This is the first element of the scatter-gather list.  If you don't
+/* This header comes first in the scatter-gather list.
+ * If VIRTIO_F_ANY_LAYOUT is not negotiated, it must
+ * be the first element of the scatter-gather list.  If you don't
  * specify GSO or CSUM features, you can simply ignore the header. */
 struct virtio_net_hdr {
 #define VIRTIO_NET_HDR_F_NEEDS_CSUM    1       // Use csum_start, csum_offset