virtio-net: page_to_skb() use build_skb when there's sufficient tailroom
authorXuan Zhuo <xuanzhuo@linux.alibaba.com>
Fri, 16 Apr 2021 09:16:15 +0000 (17:16 +0800)
committerDavid S. Miller <davem@davemloft.net>
Fri, 16 Apr 2021 22:33:22 +0000 (15:33 -0700)
In page_to_skb(), if we have enough tailroom to save skb_shared_info, we
can use build_skb to create skb directly. No need to alloc for
additional space. And it can save a 'frags slot', which is very friendly
to GRO.

Here, if the payload of the received package is too small (less than
GOOD_COPY_LEN), we still choose to copy it directly to the space got by
napi_alloc_skb. So we can reuse these pages.

Testing Machine:
    The four queues of the network card are bound to the cpu1.

Test command:
    for ((i=0;i<5;++i)); do sockperf tp --ip 192.168.122.64 -m 1000 -t 150& done

The size of the udp package is 1000, so in the case of this patch, there
will always be enough tailroom to use build_skb. The sent udp packet
will be discarded because there is no port to receive it. The irqsoftd
of the machine is 100%, we observe the received quantity displayed by
sar -n DEV 1:

no build_skb:  956864.00 rxpck/s
build_skb:    1158465.00 rxpck/s

Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
Suggested-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/virtio_net.c

index 101659cd4b87f254652c2411a9f667e631535e93..8cd76037c72481200ea3e8429e9fdfec005dad85 100644 (file)
@@ -379,21 +379,17 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
                                   struct receive_queue *rq,
                                   struct page *page, unsigned int offset,
                                   unsigned int len, unsigned int truesize,
-                                  bool hdr_valid, unsigned int metasize)
+                                  bool hdr_valid, unsigned int metasize,
+                                  unsigned int headroom)
 {
        struct sk_buff *skb;
        struct virtio_net_hdr_mrg_rxbuf *hdr;
        unsigned int copy, hdr_len, hdr_padded_len;
-       char *p;
+       int tailroom, shinfo_size;
+       char *p, *hdr_p;
 
        p = page_address(page) + offset;
-
-       /* copy small packet so we can reuse these pages for small data */
-       skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
-       if (unlikely(!skb))
-               return NULL;
-
-       hdr = skb_vnet_hdr(skb);
+       hdr_p = p;
 
        hdr_len = vi->hdr_len;
        if (vi->mergeable_rx_bufs)
@@ -401,14 +397,38 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
        else
                hdr_padded_len = sizeof(struct padded_vnet_hdr);
 
-       /* hdr_valid means no XDP, so we can copy the vnet header */
-       if (hdr_valid)
-               memcpy(hdr, p, hdr_len);
+       /* If headroom is not 0, there is an offset between the beginning of the
+        * data and the allocated space, otherwise the data and the allocated
+        * space are aligned.
+        */
+       if (headroom) {
+               /* The actual allocated space size is PAGE_SIZE. */
+               truesize = PAGE_SIZE;
+               tailroom = truesize - len - offset;
+       } else {
+               tailroom = truesize - len;
+       }
 
        len -= hdr_len;
        offset += hdr_padded_len;
        p += hdr_padded_len;
 
+       shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
+
+       if (len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
+               skb = build_skb(p, truesize);
+               if (unlikely(!skb))
+                       return NULL;
+
+               skb_put(skb, len);
+               goto ok;
+       }
+
+       /* copy small packet so we can reuse these pages for small data */
+       skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
+       if (unlikely(!skb))
+               return NULL;
+
        /* Copy all frame if it fits skb->head, otherwise
         * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
         */
@@ -418,11 +438,6 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
                copy = ETH_HLEN + metasize;
        skb_put_data(skb, p, copy);
 
-       if (metasize) {
-               __skb_pull(skb, metasize);
-               skb_metadata_set(skb, metasize);
-       }
-
        len -= copy;
        offset += copy;
 
@@ -431,7 +446,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
                        skb_add_rx_frag(skb, 0, page, offset, len, truesize);
                else
                        put_page(page);
-               return skb;
+               goto ok;
        }
 
        /*
@@ -458,6 +473,18 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
        if (page)
                give_pages(rq, page);
 
+ok:
+       /* hdr_valid means no XDP, so we can copy the vnet header */
+       if (hdr_valid) {
+               hdr = skb_vnet_hdr(skb);
+               memcpy(hdr, hdr_p, hdr_len);
+       }
+
+       if (metasize) {
+               __skb_pull(skb, metasize);
+               skb_metadata_set(skb, metasize);
+       }
+
        return skb;
 }
 
@@ -808,7 +835,7 @@ static struct sk_buff *receive_big(struct net_device *dev,
 {
        struct page *page = buf;
        struct sk_buff *skb =
-               page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
+               page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
 
        stats->bytes += len - vi->hdr_len;
        if (unlikely(!skb))
@@ -922,7 +949,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                                put_page(page);
                                head_skb = page_to_skb(vi, rq, xdp_page, offset,
                                                       len, PAGE_SIZE, false,
-                                                      metasize);
+                                                      metasize, headroom);
                                return head_skb;
                        }
                        break;
@@ -980,7 +1007,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
        }
 
        head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
-                              metasize);
+                              metasize, headroom);
        curr_skb = head_skb;
 
        if (unlikely(!curr_skb))