Re: [PATCH bpf-next] virtio_net: add XDP meta data support

On 7/1/19 6:30 PM, Jason Wang wrote:

On 2019/6/27 下午4:06, Yuya Kusakabe wrote:

quoted

This adds XDP meta data support to both receive_small() and
receive_mergeable().

Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access")
Signed-off-by: Yuya Kusakabe <redacted>
---
  drivers/net/virtio_net.c | 40 +++++++++++++++++++++++++++++-----------
  1 file changed, 29 insertions(+), 11 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 4f3de0ac8b0b..e787657fc568 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -371,7 +371,7 @@ 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)
+                   bool hdr_valid, unsigned int metasize)
  {
      struct sk_buff *skb;
      struct virtio_net_hdr_mrg_rxbuf *hdr;

@@ -393,17 +393,25 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,

      else
          hdr_padded_len = sizeof(struct padded_vnet_hdr);
  -    if (hdr_valid)
+    if (hdr_valid && !metasize)
          memcpy(hdr, p, hdr_len);
        len -= hdr_len;
      offset += hdr_padded_len;
      p += hdr_padded_len;
  -    copy = len;
+    copy = len + metasize;
      if (copy > skb_tailroom(skb))
          copy = skb_tailroom(skb);
-    skb_put_data(skb, p, copy);
+
+    if (metasize) {
+        skb_put_data(skb, p - metasize, copy);

I would rather keep copy untouched above, and use copy + metasize here, then you can save the following decrement  as well. Or tweak the caller the count the meta in to offset, then we need only deal with skb_pull() and skb_metadata_set() here.

I think the latter is better, because copy + metasize must be smaller than skb tailroom size.

quoted

+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+        copy -= metasize;
+    } else {
+        skb_put_data(skb, p, copy);
+    }
        len -= copy;
      offset += copy;

@@ -644,6 +652,7 @@ static struct sk_buff *receive_small(struct net_device *dev,

      unsigned int delta = 0;
      struct page *xdp_page;
      int err;
+    unsigned int metasize = 0;
        len -= vi->hdr_len;
      stats->bytes += len;

@@ -683,8 +692,8 @@ static struct sk_buff *receive_small(struct net_device *dev,

            xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
          xdp.data = xdp.data_hard_start + xdp_headroom;
-        xdp_set_data_meta_invalid(&xdp);
          xdp.data_end = xdp.data + len;
+        xdp.data_meta = xdp.data;
          xdp.rxq = &rq->xdp_rxq;
          orig_data = xdp.data;
          act = bpf_prog_run_xdp(xdp_prog, &xdp);

@@ -695,9 +704,11 @@ static struct sk_buff *receive_small(struct net_device *dev,

              /* Recalculate length in case bpf program changed it */
              delta = orig_data - xdp.data;
              len = xdp.data_end - xdp.data;
+            metasize = xdp.data - xdp.data_meta;
              break;
          case XDP_TX:
              stats->xdp_tx++;
+            xdp.data_meta = xdp.data;

Why need this?

Because virtnet_xdp_xmit() doesn't support XDP meta data as below. And I suppose that we don't need to support XDP metadata for XDP_TX.

static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
				   struct send_queue *sq,
				   struct xdp_frame *xdpf)
{
	struct virtio_net_hdr_mrg_rxbuf *hdr;
	int err;

	/* virtqueue want to use data area in-front of packet */
	if (unlikely(xdpf->metasize > 0))
		return -EOPNOTSUPP;

quoted

              xdpf = convert_to_xdp_frame(&xdp);
              if (unlikely(!xdpf))
                  goto err_xdp;

@@ -735,11 +746,14 @@ static struct sk_buff *receive_small(struct net_device *dev,

      }
      skb_reserve(skb, headroom - delta);
      skb_put(skb, len);
-    if (!delta) {
+    if (!delta && !metasize) {
          buf += header_offset;
          memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
      } /* keep zeroed vnet hdr since packet was changed by bpf */

Is there any method to preserve the vnet header here? We probably don't want to lose it for XDP_PASS when packet is not modified.

I'll try to keep the vnet header with moving the meta data to the front of the vnet header.

quoted

  +    if (metasize)
+        skb_metadata_set(skb, metasize);
+
  err:
      return skb;
  @@ -761,7 +775,7 @@ 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);
+                      PAGE_SIZE, true, 0);
        stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

@@ -793,6 +807,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,

      unsigned int truesize;
      unsigned int headroom = mergeable_ctx_to_headroom(ctx);
      int err;
+    unsigned int metasize = 0;
        head_skb = NULL;
      stats->bytes += len - vi->hdr_len;

@@ -839,8 +854,8 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,

          data = page_address(xdp_page) + offset;
          xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
          xdp.data = data + vi->hdr_len;
-        xdp_set_data_meta_invalid(&xdp);
          xdp.data_end = xdp.data + (len - vi->hdr_len);
+        xdp.data_meta = xdp.data;
          xdp.rxq = &rq->xdp_rxq;
            act = bpf_prog_run_xdp(xdp_prog, &xdp);

@@ -859,18 +874,20 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,

               * adjusted
               */
              len = xdp.data_end - xdp.data + vi->hdr_len;
+            metasize = xdp.data - xdp.data_meta;
              /* We can only create skb based on xdp_page. */
              if (unlikely(xdp_page != page)) {
                  rcu_read_unlock();
                  put_page(page);
                  head_skb = page_to_skb(vi, rq, xdp_page,
-                               offset, len,
-                               PAGE_SIZE, false);
+                           offset, len,
+                           PAGE_SIZE, false, metasize);

Indentation is wired.

Sorry. I'll fix it.

Thanks

quoted

                  return head_skb;
              }
              break;
          case XDP_TX:
              stats->xdp_tx++;
+            xdp.data_meta = xdp.data;
              xdpf = convert_to_xdp_frame(&xdp);
              if (unlikely(!xdpf))
                  goto err_xdp;

@@ -921,7 +938,8 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,

          goto err_skb;
      }
  -    head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog);
+    head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
+                   metasize);
      curr_skb = head_skb;
        if (unlikely(!curr_skb))