Skip to contentSkip to conversation

virtio_net: add XDP meta data support

PATCH Superseded Yuya Kusakabe 4 participants 28 messages Jun 27, 2019 – Feb 25, 2020

Subsystems (4)

v1 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 virtio_net: add XDP meta data support

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")
---
 drivers/net/virtio_net.c | 40 +++++++++++++++++++++++++++++-----------
 1 file changed, 29 insertions(+), 11 deletions(-)
drivers/net/virtio_net.c +29 -11
index 4f3de0ac8b0b..e787657fc568 100644
@@ -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);
+		__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;
 			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 */
 
+	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);
 				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))
Signature · 1 line
2.20.1
#1 lore raw
Quoted 32 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")
---
  drivers/net/virtio_net.c | 40 +++++++++++++++++++++++++++++-----------
  1 file changed, 29 insertions(+), 11 deletions(-)
drivers/net/virtio_net.c +2 -2
index 4f3de0ac8b0b..e787657fc568 100644
@@ -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;

drivers/net/virtio_net.c static struct sk_buff *page_to_skb(struct virtnet_info *vi, · line 408

Quoted 10 lines
drivers/net/virtio_net.c +4 -2
  	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.

Quoted 28 lines
drivers/net/virtio_net.c +8 -1
+		__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);

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 711

Quoted 8 lines
drivers/net/virtio_net.c +2 -0
@@ -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?

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 752

Quoted 12 lines
drivers/net/virtio_net.c +1 -1
  			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.

Quoted 38 lines
drivers/net/virtio_net.c +6 -2
+	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;

drivers/net/virtio_net.c static struct sk_buff *receive_mergeable(struct net_device *dev, · line 884

Quoted 10 lines
drivers/net/virtio_net.c +3 -2
+			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.

Thanks

Quoted 19 lines
drivers/net/virtio_net.c +3 -1
  				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))
#2 lore raw
Quoted 36 lines · Jason Wang
Quoted 34 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
  drivers/net/virtio_net.c | 40 +++++++++++++++++++++++++++++-----------

  1 file changed, 29 insertions(+), 11 deletions(-)
drivers/net/virtio_net.c +1 -1
index 4f3de0ac8b0b..e787657fc568 100644
@@ -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))

Quoted 8 lines
Quoted 5 lines

          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 29 lines
Quoted 29 lines

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

drivers/net/virtio_net.c +2 -1
@@ -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;
Quoted 8 lines
Quoted 5 lines
drivers/net/virtio_net.c +1 -0
+            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 16 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 752

Quoted 13 lines

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

drivers/net/virtio_net.c +1 -1
@@ -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 39 lines
Quoted 39 lines
drivers/net/virtio_net.c +2 -0
   +    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))

drivers/net/virtio_net.c +3 -1
@@ -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);
Quoted 8 lines

drivers/net/virtio_net.c static struct sk_buff *receive_mergeable(struct net_device *dev, · line 884

Quoted 5 lines
drivers/net/virtio_net.c +2 -2
                   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.

Quoted 21 lines

Thanks

Quoted 17 lines
drivers/net/virtio_net.c +0 -0
                   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;

drivers/net/virtio_net.c +2 -0
@@ -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))
#3 lore raw

PATCH v2 1/1 virtio_net: add XDP meta data support

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
 drivers/net/virtio_net.c | 39 +++++++++++++++++++++++++++------------
 1 file changed, 27 insertions(+), 12 deletions(-)
drivers/net/virtio_net.c +27 -12
index 4f3de0ac8b0b..2ebabb08b824 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 		copy = skb_tailroom(skb);
 	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
 	len -= copy;
 	offset += copy;
 
@@ -644,6 +649,7 @@ 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 +689,8 @@ 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 +701,11 @@ 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;
 			xdpf = convert_to_xdp_frame(&xdp);
 			if (unlikely(!xdpf))
 				goto err_xdp;
@@ -740,6 +748,9 @@ static struct sk_buff *receive_small(struct net_device *dev,
 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 	} /* keep zeroed vnet hdr since packet was changed by bpf */
 
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
 err:
 	return skb;
 
@@ -760,8 +771,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
 				   struct virtnet_rq_stats *stats)
 {
 	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
 	stats->bytes += len - vi->hdr_len;
 	if (unlikely(!skb))
@@ -793,6 +804,7 @@ 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 +851,8 @@ 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);
@@ -852,8 +864,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			 * adjustments. Note other cases do not build an
 			 * skb and avoid using offset
 			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
 			/* recalculate len if xdp.data or xdp.data_end were
 			 * adjusted
@@ -863,14 +876,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
 				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 +935,8 @@ 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))
Signature · 1 line
2.20.1
#4 lore raw
Quoted 79 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
  drivers/net/virtio_net.c | 39 +++++++++++++++++++++++++++------------
  1 file changed, 27 insertions(+), 12 deletions(-)
drivers/net/virtio_net.c +11 -3
index 4f3de0ac8b0b..2ebabb08b824 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
  		copy = skb_tailroom(skb);
  	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
  	len -= copy;
  	offset += copy;
 
@@ -644,6 +649,7 @@ 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 +689,8 @@ 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 +701,11 @@ 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;
  			xdpf = convert_to_xdp_frame(&xdp);
  			if (unlikely(!xdpf))
  				goto err_xdp;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 749

Quoted 2 lines
drivers/net/virtio_net.c +0 -0
@@ -740,6 +748,9 @@ static struct sk_buff *receive_small(struct net_device *dev,
  		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
  	} /* keep zeroed vnet hdr since packet was changed by bpf */

I wonder whether or not it's as simple as this. Consider bpf may adjust meta, it looks to me that the vnet header will be overwrite here? If yes, we probably need to have a device specific value then bpf can move the device metadata like vnet header for us?

Thanks

Quoted 77 lines
drivers/net/virtio_net.c +16 -9
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
  err:
  	return skb;
 
@@ -760,8 +771,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
  				   struct virtnet_rq_stats *stats)
  {
  	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
  	stats->bytes += len - vi->hdr_len;
  	if (unlikely(!skb))
@@ -793,6 +804,7 @@ 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 +851,8 @@ 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);
@@ -852,8 +864,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
  			 * adjustments. Note other cases do not build an
  			 * skb and avoid using offset
  			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
  			/* recalculate len if xdp.data or xdp.data_end were
  			 * adjusted
@@ -863,14 +876,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
  			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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
  				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 +935,8 @@ 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))
#5 lore raw
Quoted 73 lines · Jason Wang
Quoted 71 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.
---
  drivers/net/virtio_net.c | 39 +++++++++++++++++++++++++++------------

  1 file changed, 27 insertions(+), 12 deletions(-)
drivers/net/virtio_net.c +1 -1
index 4f3de0ac8b0b..2ebabb08b824 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;

drivers/net/virtio_net.c +4 -0
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
           copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -644,6 +649,7 @@ 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;

drivers/net/virtio_net.c +2 -1
@@ -683,8 +689,8 @@ 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 +701,11 @@ 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;
              xdpf = convert_to_xdp_frame(&xdp);
              if (unlikely(!xdpf))

Quoted 8 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 749

Quoted 4 lines

                  goto err_xdp;

drivers/net/virtio_net.c +0 -0
@@ -740,6 +748,9 @@ static struct sk_buff *receive_small(struct net_device *dev,
           memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */

I wonder whether or not it's as simple as this. Consider bpf may adjust meta, it looks to me that the vnet header will be overwrite here? If yes, we probably need to have a device specific value then bpf can move the device metadata like vnet header for us?> Thanks

Yes, the vnet header is overwrite by xdp metadata. I'll fix it.

Thanks.

Quoted 71 lines
Quoted 71 lines
drivers/net/virtio_net.c +2 -0
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +771,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +8 -6
@@ -793,6 +804,7 @@ 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 +851,8 @@ 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);
@@ -852,8 +864,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                * adjustments. Note other cases do not build an
                * skb and avoid using offset
                */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len;
+            metasize = xdp.data - xdp.data_meta;
+            offset = xdp.data - page_address(xdp_page) -
+                 vi->hdr_len - metasize;
                 /* recalculate len if xdp.data or xdp.data_end were
                * adjusted
@@ -863,14 +876,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
               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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset,
+                               len, PAGE_SIZE, false,
+                               metasize);
                   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;

drivers/net/virtio_net.c +2 -0
@@ -921,7 +935,8 @@ 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))
#6 lore raw

PATCH v3 1/1 virtio_net: add XDP meta data support

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
- fix preserve the vnet header in receive_small().
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
 drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
 1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +31 -14
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 		copy = skb_tailroom(skb);
 	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
 	len -= copy;
 	offset += copy;
 
@@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,13 @@ 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;
+		/* Copy the vnet header to the front of data_hard_start to avoid
+		 * overwriting by XDP meta data */
+		memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);
 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
 		stats->xdp_packets++;
 
@@ -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;
 			xdpf = convert_to_xdp_frame(&xdp);
 			if (unlikely(!xdpf))
 				goto err_xdp;
@@ -736,10 +747,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
 	skb_reserve(skb, headroom - delta);
 	skb_put(skb, len);
 	if (!delta) {
-		buf += header_offset;
-		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+		memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
 	} /* keep zeroed vnet hdr since packet was changed by bpf */
 
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
 err:
 	return skb;
 
@@ -760,8 +773,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
 				   struct virtnet_rq_stats *stats)
 {
 	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
 	stats->bytes += len - vi->hdr_len;
 	if (unlikely(!skb))
@@ -793,6 +806,7 @@ 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 +853,8 @@ 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);
@@ -852,8 +866,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			 * adjustments. Note other cases do not build an
 			 * skb and avoid using offset
 			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
 			/* recalculate len if xdp.data or xdp.data_end were
 			 * adjusted
@@ -863,14 +878,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
 				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 +937,8 @@ 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))
Signature · 1 line
2.20.1
#7 lore raw
Quoted 60 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
- fix preserve the vnet header in receive_small().
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
  drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
  1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +8 -2
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
  		copy = skb_tailroom(skb);
  	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
  	len -= copy;
  	offset += copy;
 
@@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,13 @@ static struct sk_buff *receive_small(struct net_device *dev,
 

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 698

Quoted 10 lines
drivers/net/virtio_net.c +4 -1
  		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;
+		/* Copy the vnet header to the front of data_hard_start to avoid
+		 * overwriting by XDP meta data */
+		memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

What happens if we have a large metadata that occupies all headroom here?

Thanks

Quoted 100 lines
drivers/net/virtio_net.c +19 -11
  		act = bpf_prog_run_xdp(xdp_prog, &xdp);
  		stats->xdp_packets++;
 
@@ -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;
  			xdpf = convert_to_xdp_frame(&xdp);
  			if (unlikely(!xdpf))
  				goto err_xdp;
@@ -736,10 +747,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
  	skb_reserve(skb, headroom - delta);
  	skb_put(skb, len);
  	if (!delta) {
-		buf += header_offset;
-		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+		memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
  	} /* keep zeroed vnet hdr since packet was changed by bpf */
 
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
  err:
  	return skb;
 
@@ -760,8 +773,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
  				   struct virtnet_rq_stats *stats)
  {
  	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
  	stats->bytes += len - vi->hdr_len;
  	if (unlikely(!skb))
@@ -793,6 +806,7 @@ 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 +853,8 @@ 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);
@@ -852,8 +866,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
  			 * adjustments. Note other cases do not build an
  			 * skb and avoid using offset
  			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
  			/* recalculate len if xdp.data or xdp.data_end were
  			 * adjusted
@@ -863,14 +878,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
  			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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
  				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 +937,8 @@ 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))
#8 lore raw
Quoted 70 lines · Jason Wang
Quoted 54 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.
---
  drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------

  1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +1 -1
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;

drivers/net/virtio_net.c +4 -0
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
           copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -644,6 +649,7 @@ 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;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 697

Quoted 10 lines
drivers/net/virtio_net.c +4 -1
@@ -683,10 +689,13 @@ 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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

What happens if we have a large metadata that occupies all headroom here?

Quoted 1 line

Thanks

Do you mean a large "XDP" metadata? If a large metadata is a large "XDP" metadata, I think we can not use a metadata that occupies all headroom. The size of metadata limited by bpf_xdp_adjust_meta() as below. bpf_xdp_adjust_meta() in net/core/filter.c:

	if (unlikely((metalen & (sizeof(__u32) - 1)) ||
		     (metalen > 32)))
		return -EACCES;

Thanks.

Quoted 93 lines
Quoted 93 lines
drivers/net/virtio_net.c +0 -0
           act = bpf_prog_run_xdp(xdp_prog, &xdp);
           stats->xdp_packets++;
   @@ -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;
              xdpf = convert_to_xdp_frame(&xdp);
              if (unlikely(!xdpf))
                  goto err_xdp;

drivers/net/virtio_net.c +3 -2
@@ -736,10 +747,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
       if (!delta) {
-        buf += header_offset;
-        memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+        memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +773,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +8 -6
@@ -793,6 +806,7 @@ 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 +853,8 @@ 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);
@@ -852,8 +866,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                * adjustments. Note other cases do not build an
                * skb and avoid using offset
                */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len;
+            metasize = xdp.data - xdp.data_meta;
+            offset = xdp.data - page_address(xdp_page) -
+                 vi->hdr_len - metasize;
                 /* recalculate len if xdp.data or xdp.data_end were
                * adjusted
@@ -863,14 +878,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
               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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset,
+                               len, PAGE_SIZE, false,
+                               metasize);
                   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;

drivers/net/virtio_net.c +2 -0
@@ -921,7 +937,8 @@ 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))
maintainer
#9 lore raw
Quoted 58 lines · Yuya Kusakabe
Quoted 57 lines · Jason Wang
Quoted 56 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.
---
  drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------

  1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +1 -1
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;

drivers/net/virtio_net.c +4 -0
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
           copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -644,6 +649,7 @@ 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;

drivers/net/virtio_net.c +0 -0
@@ -683,10 +689,13 @@ 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;
Quoted 8 lines
Quoted 8 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 698

Quoted 8 lines
drivers/net/virtio_net.c +4 -1
-        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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

I'm not fully sure if I'm following this one correctly, probably just missing something. Isn't the vnet header based on how we set up xdp.data_hard_start earlier already in front of it? Wouldn't we copy invalid data from xdp.data - vi->hdr_len into the vnet header at that point (given there can be up to 256 bytes of headroom between the two)? If it's relative to xdp.data and headroom is >0, then BPF prog could otherwise mangle this; something doesn't add up to me here. Could you clarify? Thx

Quoted 105 lines
Quoted 3 lines

What happens if we have a large metadata that occupies all headroom here?

Thanks

Do you mean a large "XDP" metadata? If a large metadata is a large "XDP" metadata, I think we can not use a metadata that occupies all headroom. The size of metadata limited by bpf_xdp_adjust_meta() as below. bpf_xdp_adjust_meta() in net/core/filter.c:

	if (unlikely((metalen & (sizeof(__u32) - 1)) ||
		     (metalen > 32)))
		return -EACCES;

Thanks.

Quoted 93 lines
Quoted 93 lines
drivers/net/virtio_net.c +0 -0
           act = bpf_prog_run_xdp(xdp_prog, &xdp);
           stats->xdp_packets++;
   @@ -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;
              xdpf = convert_to_xdp_frame(&xdp);
              if (unlikely(!xdpf))
                  goto err_xdp;

drivers/net/virtio_net.c +3 -2
@@ -736,10 +747,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
       if (!delta) {
-        buf += header_offset;
-        memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+        memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +773,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +8 -6
@@ -793,6 +806,7 @@ 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 +853,8 @@ 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);
@@ -852,8 +866,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                * adjustments. Note other cases do not build an
                * skb and avoid using offset
                */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len;
+            metasize = xdp.data - xdp.data_meta;
+            offset = xdp.data - page_address(xdp_page) -
+                 vi->hdr_len - metasize;
                 /* recalculate len if xdp.data or xdp.data_end were
                * adjusted
@@ -863,14 +878,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
               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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset,
+                               len, PAGE_SIZE, false,
+                               metasize);
                   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;

drivers/net/virtio_net.c +2 -0
@@ -921,7 +937,8 @@ 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))
#10 lore raw
Quoted 66 lines · Daniel Borkmann
Quoted 65 lines · Yuya Kusakabe
Quoted 64 lines · Jason Wang
Quoted 63 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.

---
   drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
   1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +10 -2
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
           copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,13 @@ 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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
Quoted 8 lines
Quoted 1 line
Quoted 1 line

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 698

Quoted 1 line
drivers/net/virtio_net.c +1 -0
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

I'm not fully sure if I'm following this one correctly, probably just missing something. Isn't the vnet header based on how we set up xdp.data_hard_start earlier already in front of it? Wouldn't we copy invalid data from xdp.data - vi->hdr_len into the vnet header at that point (given there can be up to 256 bytes of headroom between the two)? If it's relative to xdp.data and headroom is >0, then BPF prog could otherwise mangle this; something doesn't add up to me here. Could you clarify? Thx

Vnet headr sits just in front of xdp.data not xdp.data_hard_start. So it could be overwrote by metadata, that's why we need a copy here.

Thanks

Quoted 104 lines
Quoted 104 lines
Quoted 4 lines

What happens if we have a large metadata that occupies all headroom here?

Thanks

Do you mean a large "XDP" metadata? If a large metadata is a large "XDP" metadata, I think we can not use a metadata that occupies all headroom. The size of metadata limited by bpf_xdp_adjust_meta() as below. bpf_xdp_adjust_meta() in net/core/filter.c:

	if (unlikely((metalen & (sizeof(__u32) - 1)) ||
		     (metalen > 32)))
		return -EACCES;

Thanks.

Quoted 92 lines
Quoted 92 lines
drivers/net/virtio_net.c +0 -0
           act = bpf_prog_run_xdp(xdp_prog, &xdp);
           stats->xdp_packets++;
   @@ -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;
              xdpf = convert_to_xdp_frame(&xdp);
              if (unlikely(!xdpf))
                  goto err_xdp;

drivers/net/virtio_net.c +16 -10
@@ -736,10 +747,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
       if (!delta) {
-        buf += header_offset;
-        memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+        memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+
   err:
       return skb;
   @@ -760,8 +773,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
                      struct virtnet_rq_stats *stats)
   {
       struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
         stats->bytes += len - vi->hdr_len;
       if (unlikely(!skb))
@@ -793,6 +806,7 @@ 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 +853,8 @@ 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);
@@ -852,8 +866,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                * adjustments. Note other cases do not build an
                * skb and avoid using offset
                */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len;
+            metasize = xdp.data - xdp.data_meta;
+            offset = xdp.data - page_address(xdp_page) -
+                 vi->hdr_len - metasize;
                 /* recalculate len if xdp.data or xdp.data_end were
                * adjusted
@@ -863,14 +878,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
               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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset,
+                               len, PAGE_SIZE, false,
+                               metasize);
                   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 +937,8 @@ 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))
maintainer
#11 lore raw
Quoted 76 lines · Jason Wang
Quoted 66 lines · Daniel Borkmann
Quoted 65 lines · Yuya Kusakabe
Quoted 64 lines · Jason Wang
Quoted 63 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
   - fix preserve the vnet header in receive_small().
v2:
   - keep copy untouched in page_to_skb().
   - preserve the vnet header in receive_small().
   - fix indentation.
---
   drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------

   1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +1 -1
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;

drivers/net/virtio_net.c +4 -0
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
            copy = skb_tailroom(skb);
        skb_put_data(skb, p, copy);
    +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
        len -= copy;
        offset += copy;
    @@ -644,6 +649,7 @@ 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;

drivers/net/virtio_net.c +3 -1
@@ -683,10 +689,13 @@ 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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
Quoted 8 lines
Quoted 1 line
Quoted 1 line

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 698

Quoted 1 line
drivers/net/virtio_net.c +1 -0
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

I'm not fully sure if I'm following this one correctly, probably just missing something. Isn't the vnet header based on how we set up xdp.data_hard_start earlier already in front of it? Wouldn't we copy invalid data from xdp.data - vi->hdr_len into the vnet header at that point (given there can be up to 256 bytes of headroom between the two)? If it's relative to xdp.data and headroom is >0, then BPF prog could otherwise mangle this; something doesn't add up to me here. Could you clarify? Thx

Quoted 1 line

Vnet headr sits just in front of xdp.data not xdp.data_hard_start. So it could be overwrote by metadata, that's why we need a copy here.

For the current code, you can adjust the xdp.data with a positive/negative offset already via bpf_xdp_adjust_head() helper. If vnet headr sits just in front of xdp.data, couldn't this be overridden today as well then? Anyway, just wondering how this is handled differently?

Thanks,
Daniel

#12 lore raw
Quoted 76 lines · Daniel Borkmann
Quoted 75 lines · Jason Wang
Quoted 74 lines · Daniel Borkmann
Quoted 58 lines · Yuya Kusakabe
Quoted 57 lines · Jason Wang
Quoted 56 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
   - fix preserve the vnet header in receive_small().
v2:
   - keep copy untouched in page_to_skb().
   - preserve the vnet header in receive_small().
   - fix indentation.

---
    drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
    1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +7 -1
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
            copy = skb_tailroom(skb);
        skb_put_data(skb, p, copy);
    +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
        len -= copy;
        offset += copy;
    @@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,13 @@ 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;
Quoted 8 lines
Quoted 8 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 698

Quoted 8 lines
drivers/net/virtio_net.c +4 -1
-        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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

I'm not fully sure if I'm following this one correctly, probably just missing something. Isn't the vnet header based on how we set up xdp.data_hard_start earlier already in front of it? Wouldn't we copy invalid data from xdp.data - vi->hdr_len into the vnet header at that point (given there can be up to 256 bytes of headroom between the two)? If it's relative to xdp.data and headroom is >0, then BPF prog could otherwise mangle this; something doesn't add up to me here. Could you clarify? Thx

Quoted 5 lines
Quoted 1 line

Vnet headr sits just in front of xdp.data not xdp.data_hard_start. So it could be overwrote by metadata, that's why we need a copy here.

For the current code, you can adjust the xdp.data with a positive/negative offset already via bpf_xdp_adjust_head() helper. If vnet headr sits just in front of xdp.data, couldn't this be overridden today as well then? Anyway, just wondering how this is handled differently?

We will invalidate the vnet header in this case. But for the case of metadata adjustment without header adjustment, we want to seek a way to preserve that.

Thanks

Quoted 3 lines

Thanks,
Daniel

#13 lore raw

I'm sorry for the late reply.

I saw the status of this patch is "Awaiting Upstream" on
https://patchwork.ozlabs.org/patch/1126046/. What is "Awaiting Upstream"? Is there anything that I can do?

Thank you,
Yuya

Quoted 96 lines · Jason Wang
Quoted 76 lines · Daniel Borkmann
Quoted 75 lines · Jason Wang
Quoted 74 lines · Daniel Borkmann
Quoted 58 lines · Yuya Kusakabe
Quoted 57 lines · Jason Wang
Quoted 56 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
- fix preserve the vnet header in receive_small().
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
    drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
    1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +7 -1
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
            copy = skb_tailroom(skb);
        skb_put_data(skb, p, copy);
    +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
        len -= copy;
        offset += copy;
    @@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,13 @@ 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;
Quoted 8 lines
Quoted 8 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 697

Quoted 8 lines
drivers/net/virtio_net.c +4 -1
-        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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

I'm not fully sure if I'm following this one correctly, probably just missing something. Isn't the vnet header based on how we set up xdp.data_hard_start earlier already in front of it? Wouldn't we copy invalid data from xdp.data - vi->hdr_len into the vnet header at that point (given there can be up to 256 bytes of headroom between the two)? If it's relative to xdp.data and headroom is >0, then BPF prog could otherwise mangle this; something doesn't add up to me here. Could you clarify? Thx

Quoted 5 lines
Quoted 1 line

Vnet headr sits just in front of xdp.data not xdp.data_hard_start. So it could be overwrote by metadata, that's why we need a copy here.

For the current code, you can adjust the xdp.data with a positive/negative offset already via bpf_xdp_adjust_head() helper. If vnet headr sits just in front of xdp.data, couldn't this be overridden today as well then? Anyway, just wondering how this is handled differently?

We will invalidate the vnet header in this case. But for the case of metadata adjustment without header adjustment, we want to seek a way to preserve that.

Thanks

Quoted 3 lines

Thanks,
Daniel

Signature · 1 line
Yuya Kusakabe
#14 lore raw
Quoted 5 lines · Yuya Kusakabe

I'm sorry for the late reply.

I saw the status of this patch is "Awaiting Upstream" on
https://patchwork.ozlabs.org/patch/1126046/. What is "Awaiting Upstream"? Is there anything that I can do?

You can post a new version I think.

Thanks

Quoted 100 lines

Thank you,
Yuya

Quoted 93 lines · Jason Wang
Quoted 76 lines · Daniel Borkmann
Quoted 75 lines · Jason Wang
Quoted 74 lines · Daniel Borkmann
Quoted 58 lines · Yuya Kusakabe
Quoted 57 lines · Jason Wang
Quoted 56 lines · Yuya Kusakabe

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

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v3:
- fix preserve the vnet header in receive_small().
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
     drivers/net/virtio_net.c | 45 +++++++++++++++++++++++++++-------------
     1 file changed, 31 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +7 -1
index 4f3de0ac8b0b..03a1ae6fe267 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
             copy = skb_tailroom(skb);
         skb_put_data(skb, p, copy);
     +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
         len -= copy;
         offset += copy;
     @@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,13 @@ 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;
Quoted 8 lines
Quoted 8 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 697

Quoted 8 lines
drivers/net/virtio_net.c +4 -1
-        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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting by XDP meta data */
+        memcpy(xdp.data_hard_start - vi->hdr_len, xdp.data - vi->hdr_len, vi->hdr_len);

I'm not fully sure if I'm following this one correctly, probably just missing something. Isn't the vnet header based on how we set up xdp.data_hard_start earlier already in front of it? Wouldn't we copy invalid data from xdp.data - vi->hdr_len into the vnet header at that point (given there can be up to 256 bytes of headroom between the two)? If it's relative to xdp.data and headroom is >0, then BPF prog could otherwise mangle this; something doesn't add up to me here. Could you clarify? Thx

Quoted 5 lines
Quoted 1 line

Vnet headr sits just in front of xdp.data not xdp.data_hard_start. So it could be overwrote by metadata, that's why we need a copy here.

For the current code, you can adjust the xdp.data with a positive/negative offset already via bpf_xdp_adjust_head() helper. If vnet headr sits just in front of xdp.data, couldn't this be overridden today as well then? Anyway, just wondering how this is handled differently?

We will invalidate the vnet header in this case. But for the case of metadata adjustment without header adjustment, we want to seek a way to preserve that.

Thanks

Quoted 2 lines

Thanks,
Daniel

#15 lore raw

PATCH v4 1/1 virtio_net: add XDP meta data support

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data and copy vnet header to the front of xdp.data_hard_start to avoid overwriting it, where on program return with pass verdict, we call into skb_metadata_set().

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")
---
 drivers/net/virtio_net.c | 47 ++++++++++++++++++++++++++++------------
 1 file changed, 33 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +33 -14
index 2fe7a3188282..5fdd6ea0e3f1 100644
@@ -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;
@@ -393,7 +393,7 @@ 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;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 		copy = skb_tailroom(skb);
 	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
 	len -= copy;
 	offset += copy;
 
@@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,15 @@ 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;
+		/* Copy the vnet header to the front of data_hard_start to avoid
+		 * overwriting it by XDP meta data.
+		 */
+		memcpy(xdp.data_hard_start - vi->hdr_len,
+		       xdp.data - vi->hdr_len, vi->hdr_len);
 		act = bpf_prog_run_xdp(xdp_prog, &xdp);
 		stats->xdp_packets++;
 
@@ -695,9 +706,11 @@ 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;
 			xdpf = convert_to_xdp_frame(&xdp);
 			if (unlikely(!xdpf))
 				goto err_xdp;
@@ -736,10 +749,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
 	skb_reserve(skb, headroom - delta);
 	skb_put(skb, len);
 	if (!delta) {
-		buf += header_offset;
-		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+		memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
 	} /* keep zeroed vnet hdr since packet was changed by bpf */
 
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
 err:
 	return skb;
 
@@ -760,8 +775,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
 				   struct virtnet_rq_stats *stats)
 {
 	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
 	stats->bytes += len - vi->hdr_len;
 	if (unlikely(!skb))
@@ -793,6 +808,7 @@ 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 +855,8 @@ 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);
@@ -852,8 +868,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			 * adjustments. Note other cases do not build an
 			 * skb and avoid using offset
 			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
 			/* recalculate len if xdp.data or xdp.data_end were
 			 * adjusted
@@ -863,14 +880,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
 				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 +939,8 @@ 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))
Signature · 1 line
2.24.1
#16 lore raw
Quoted 26 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data and copy vnet header to the front of xdp.data_hard_start to avoid overwriting it, where on program return with pass verdict, we call into skb_metadata_set().

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")
---
  drivers/net/virtio_net.c | 47 ++++++++++++++++++++++++++++------------
  1 file changed, 33 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +1 -1
index 2fe7a3188282..5fdd6ea0e3f1 100644
@@ -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;

drivers/net/virtio_net.c static struct sk_buff *page_to_skb(struct virtnet_info *vi, · line 396

Quoted 5 lines
drivers/net/virtio_net.c +1 -1
@@ -393,7 +393,7 @@ 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)

hdr_valid means no XDP, so I think we can remove the check for metasize here and add a comment instead?

Quoted 27 lines
drivers/net/virtio_net.c +6 -0
  		memcpy(hdr, p, hdr_len);
 
  	len -= hdr_len;
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
  		copy = skb_tailroom(skb);
  	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
  	len -= copy;
  	offset += copy;
 
@@ -644,6 +649,7 @@ 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;
@@ -683,10 +689,15 @@ 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;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 700

Quoted 10 lines
drivers/net/virtio_net.c +6 -1
-		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;
+		/* Copy the vnet header to the front of data_hard_start to avoid
+		 * overwriting it by XDP meta data.
+		 */
+		memcpy(xdp.data_hard_start - vi->hdr_len,
+		       xdp.data - vi->hdr_len, vi->hdr_len);

I think we don't need this. And it looks to me there's a bug in the current code.

Commit 436c9453a1ac0 ("virtio-net: keep vnet header zeroed after processing XDP") leave the a corner case for receive_small() which still use:

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

Which seems wrong, we need check xdp_prog instead of delta.

With this fixed, there's no need to care about the vnet header here since we don't know whether or not packet is modified by XDP.

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 713

Quoted 12 lines
drivers/net/virtio_net.c +2 -0
  		act = bpf_prog_run_xdp(xdp_prog, &xdp);
  		stats->xdp_packets++;
 
@@ -695,9 +706,11 @@ 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;

I think we should remove the xdp_set_data_meta_invalid() at least? And move this initialization just after xdp.data is initialized.

Testing receive_small() requires to disable mrg_rxbuf, guest_tso4, guest_tso6 and guest_ufo from qemu command line.

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 751

Quoted 7 lines
drivers/net/virtio_net.c +0 -0
  			xdpf = convert_to_xdp_frame(&xdp);
  			if (unlikely(!xdpf))
  				goto err_xdp;
@@ -736,10 +749,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
  	skb_reserve(skb, headroom - delta);
  	skb_put(skb, len);
  	if (!delta) {

Need to check xdp_prog (need another patch).

Quoted 58 lines
drivers/net/virtio_net.c +11 -9
-		buf += header_offset;
-		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+		memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
  	} /* keep zeroed vnet hdr since packet was changed by bpf */
 
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
  err:
  	return skb;
 
@@ -760,8 +775,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
  				   struct virtnet_rq_stats *stats)
  {
  	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
  	stats->bytes += len - vi->hdr_len;
  	if (unlikely(!skb))
@@ -793,6 +808,7 @@ 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 +855,8 @@ 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);
@@ -852,8 +868,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
  			 * adjustments. Note other cases do not build an
  			 * skb and avoid using offset
  			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
  			/* recalculate len if xdp.data or xdp.data_end were
  			 * adjusted
@@ -863,14 +880,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
  			if (unlikely(xdp_page != page)) {
  				rcu_read_unlock();
  				put_page(page);
-				head_skb = page_to_skb(vi, rq, xdp_page,
-						       offset, len,

drivers/net/virtio_net.c static struct sk_buff *receive_mergeable(struct net_device *dev, · line 891

Quoted 10 lines
drivers/net/virtio_net.c +4 -1
-						       PAGE_SIZE, false);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
  				return head_skb;
  			}
  			break;
  		case XDP_TX:
  			stats->xdp_tx++;
+			xdp.data_meta = xdp.data;

Any reason for doing this?

Thanks

Quoted 13 lines
drivers/net/virtio_net.c +2 -1
  			xdpf = convert_to_xdp_frame(&xdp);
  			if (unlikely(!xdpf))
  				goto err_xdp;
@@ -921,7 +939,8 @@ 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))
#17 lore raw
Quoted 27 lines · Jason Wang
Quoted 25 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data and copy vnet header to the front of xdp.data_hard_start to avoid overwriting it, where on program return with pass verdict, we call into skb_metadata_set().

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
  drivers/net/virtio_net.c | 47 ++++++++++++++++++++++++++++------------

  1 file changed, 33 insertions(+), 14 deletions(-)
drivers/net/virtio_net.c +1 -1
index 2fe7a3188282..5fdd6ea0e3f1 100644
@@ -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;
Quoted 8 lines
Quoted 5 lines
drivers/net/virtio_net.c +0 -0
@@ -393,7 +393,7 @@ 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)

hdr_valid means no XDP, so I think we can remove the check for metasize here and add a comment instead?

I will fix it on next patch.

Quoted 39 lines
Quoted 23 lines

          memcpy(hdr, p, hdr_len);
        len -= hdr_len;

drivers/net/virtio_net.c +4 -0
@@ -405,6 +405,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
           copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -644,6 +649,7 @@ 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;

drivers/net/virtio_net.c +0 -0
@@ -683,10 +689,15 @@ 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;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 699

Quoted 10 lines
drivers/net/virtio_net.c +6 -1
-        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;
+        /* Copy the vnet header to the front of data_hard_start to avoid
+         * overwriting it by XDP meta data.
+         */
+        memcpy(xdp.data_hard_start - vi->hdr_len,
+               xdp.data - vi->hdr_len, vi->hdr_len);

I think we don't need this. And it looks to me there's a bug in the current code.

Commit 436c9453a1ac0 ("virtio-net: keep vnet header zeroed after processing XDP") leave the a corner case for receive_small() which still use:

Quoted 8 lines

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

Which seems wrong, we need check xdp_prog instead of delta.

With this fixed, there's no need to care about the vnet header here since we don't know whether or not packet is modified by XDP.

I missed this commit. I understand this is the reason for "Awaiting Upstream".

Quoted 19 lines
Quoted 5 lines
drivers/net/virtio_net.c +0 -0
           act = bpf_prog_run_xdp(xdp_prog, &xdp);
           stats->xdp_packets++;
   @@ -695,9 +706,11 @@ static struct sk_buff *receive_small(struct net_device *dev,
               /* Recalculate length in case bpf program changed it */
Quoted 7 lines

              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;

I think we should remove the xdp_set_data_meta_invalid() at least? And move this initialization just after xdp.data is initialized.

Testing receive_small() requires to disable mrg_rxbuf, guest_tso4, guest_tso6 and guest_ufo from qemu command line.

Quoted 10 lines

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 751

Quoted 7 lines

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

drivers/net/virtio_net.c +0 -0
@@ -736,10 +749,12 @@ static struct sk_buff *receive_small(struct net_device *dev,
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
       if (!delta) {

Need to check xdp_prog (need another patch).

I will fix it on next patch.

Quoted 59 lines
Quoted 59 lines
drivers/net/virtio_net.c +3 -2
-        buf += header_offset;
-        memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
+        memcpy(skb_vnet_hdr(skb), buf + VIRTNET_RX_PAD, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +775,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +8 -6
@@ -793,6 +808,7 @@ 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 +855,8 @@ 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);
@@ -852,8 +868,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
                * adjustments. Note other cases do not build an
                * skb and avoid using offset
                */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len;
+            metasize = xdp.data - xdp.data_meta;
+            offset = xdp.data - page_address(xdp_page) -
+                 vi->hdr_len - metasize;
                 /* recalculate len if xdp.data or xdp.data_end were
                * adjusted
@@ -863,14 +880,15 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
               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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset,
+                               len, PAGE_SIZE, false,
+                               metasize);
                   return head_skb;
Quoted 8 lines
Quoted 5 lines
drivers/net/virtio_net.c +0 -0
               }
               break;

          case XDP_TX:
              stats->xdp_tx++;
+            xdp.data_meta = xdp.data;

Any reason for doing this?

XDP_TX can not support metadata for now, because if metasize > 0, __virtnet_xdp_xmit_one() returns EOPNOTSUPP.

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 4 lines

Thanks

Quoted 12 lines

drivers/net/virtio_net.c static struct sk_buff *receive_mergeable(struct net_device *dev, · line 945

Quoted 11 lines

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

drivers/net/virtio_net.c +2 -0
@@ -921,7 +939,8 @@ 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))

Thank you for your kind review.

#18 lore raw
Quoted 9 lines · Yuya Kusakabe
Quoted 4 lines
Quoted 3 lines

          case XDP_TX:
              stats->xdp_tx++;
+            xdp.data_meta = xdp.data;

Any reason for doing this?

XDP_TX can not support metadata for now, because if metasize > 0, __virtnet_xdp_xmit_one() returns EOPNOTSUPP.

static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
struct send_queue *sq, struct xdp_frame *xdpf)

Quoted 9 lines
{
	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;

I see.

Then I think it's better to fix __virtnet_xdp_xmit_one() instead.

Thanks

#19 lore raw

PATCH v5 1/1 virtio_net: add XDP meta data support

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

---
v5:
- page_to_skb(): copy vnet header if hdr_valid without checking metasize.
- receive_small(): do not copy vnet header if xdp_prog is availavle.
- __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
- improve comments.
v4:
- improve commit message
v3:
- fix preserve the vnet header in receive_small().
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
 drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------
 1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +33 -21
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
 
@@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
 		copy = skb_tailroom(skb);
 	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
 	len -= copy;
 	offset += copy;
 
@@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
 	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;
-
 	if (unlikely(xdpf->headroom < vi->hdr_len))
 		return -EOVERFLOW;
 
@@ -644,6 +646,7 @@ 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 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
 	}
 	skb_reserve(skb, headroom - delta);
 	skb_put(skb, len);
-	if (!delta) {
+	if (!xdp_prog) {
 		buf += header_offset;
 		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
 	} /* keep zeroed vnet hdr since packet was changed by bpf */
 
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
 err:
 	return skb;
 
@@ -760,8 +767,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
 				   struct virtnet_rq_stats *stats)
 {
 	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
 	stats->bytes += len - vi->hdr_len;
 	if (unlikely(!skb))
@@ -793,6 +800,7 @@ 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 +847,8 @@ 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);
@@ -848,24 +856,27 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 
 		switch (act) {
 		case XDP_PASS:
+			metasize = xdp.data - xdp.data_meta;
+
 			/* recalculate offset to account for any header
-			 * adjustments. Note other cases do not build an
-			 * skb and avoid using offset
+			 * adjustments and minus the metasize to copy the
+			 * metadata in page_to_skb(). Note other cases do not
+			 * build an skb and avoid using offset
 			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
-			/* recalculate len if xdp.data or xdp.data_end were
-			 * adjusted
+			/* recalculate len if xdp.data, xdp.data_end or
+			 * xdp.data_meta were adjusted
 			 */
-			len = xdp.data_end - xdp.data + vi->hdr_len;
+			len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
 			/* 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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
 				return head_skb;
 			}
 			break;
@@ -921,7 +932,8 @@ 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))
Signature · 1 line
2.24.1
#20 lore raw
Quoted 9 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

I'm not sure this is correct since virtio-net claims to not support metadata by calling xdp_set_data_meta_invalid()?

Quoted 90 lines

---
v5:
- page_to_skb(): copy vnet header if hdr_valid without checking metasize.
- receive_small(): do not copy vnet header if xdp_prog is availavle.
- __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
- improve comments.
v4:
- improve commit message
v3:
- fix preserve the vnet header in receive_small().
v2:
- keep copy untouched in page_to_skb().
- preserve the vnet header in receive_small().
- fix indentation.

---
  drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------
  1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +10 -6
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
 
@@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
  		copy = skb_tailroom(skb);
  	skb_put_data(skb, p, copy);
 
+	if (metasize) {
+		__skb_pull(skb, metasize);
+		skb_metadata_set(skb, metasize);
+	}
+
  	len -= copy;
  	offset += copy;
 
@@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
  	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;
-
  	if (unlikely(xdpf->headroom < vi->hdr_len))
  		return -EOVERFLOW;
 
@@ -644,6 +646,7 @@ 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 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 745

Quoted 8 lines
drivers/net/virtio_net.c +1 -1
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
  	}
  	skb_reserve(skb, headroom - delta);
  	skb_put(skb, len);
-	if (!delta) {
+	if (!xdp_prog) {
  		buf += header_offset;
  		memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
  	} /* keep zeroed vnet hdr since packet was changed by bpf */

I prefer to make this an independent patch and cc stable.

Other looks good.

Thanks

Quoted 84 lines
drivers/net/virtio_net.c +22 -14
+	if (metasize)
+		skb_metadata_set(skb, metasize);
+
  err:
  	return skb;
 
@@ -760,8 +767,8 @@ static struct sk_buff *receive_big(struct net_device *dev,
  				   struct virtnet_rq_stats *stats)
  {
  	struct page *page = buf;
-	struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-					  PAGE_SIZE, true);
+	struct sk_buff *skb =
+		page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
 
  	stats->bytes += len - vi->hdr_len;
  	if (unlikely(!skb))
@@ -793,6 +800,7 @@ 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 +847,8 @@ 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);
@@ -848,24 +856,27 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 
  		switch (act) {
  		case XDP_PASS:
+			metasize = xdp.data - xdp.data_meta;
+
  			/* recalculate offset to account for any header
-			 * adjustments. Note other cases do not build an
-			 * skb and avoid using offset
+			 * adjustments and minus the metasize to copy the
+			 * metadata in page_to_skb(). Note other cases do not
+			 * build an skb and avoid using offset
  			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
-			/* recalculate len if xdp.data or xdp.data_end were
-			 * adjusted
+			/* recalculate len if xdp.data, xdp.data_end or
+			 * xdp.data_meta were adjusted
  			 */
-			len = xdp.data_end - xdp.data + vi->hdr_len;
+			len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
  			/* 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);
+				head_skb = page_to_skb(vi, rq, xdp_page, offset,
+						       len, PAGE_SIZE, false,
+						       metasize);
  				return head_skb;
  			}
  			break;
@@ -921,7 +932,8 @@ 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))
#21 lore raw
Quoted 7 lines · Jason Wang
Quoted 5 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Quoted 7 lines
Quoted 4 lines

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

I'm not sure this is correct since virtio-net claims to not support metadata by calling xdp_set_data_meta_invalid()?

virtio_net doesn't support by calling xdp_set_data_meta_invalid() for now.

https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n686
https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n842

And xdp_set_data_meta_invalid() are added by de8f3a83b0a0.

$ git blame ./drivers/net/virtio_net.c | grep xdp_set_data_meta_invalid
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  686)                xdp_set_data_meta_invalid(&xdp);
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  842)                xdp_set_data_meta_invalid(&xdp);

So I added Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access") to the comment.

Quoted 96 lines
Quoted 84 lines

---
v5:
  - page_to_skb(): copy vnet header if hdr_valid without checking metasize.
  - receive_small(): do not copy vnet header if xdp_prog is availavle.
  - __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
  - improve comments.
v4:
  - improve commit message
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.
---
  drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------

  1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +1 -1
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
  @@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi, copy = skb_tailroom(skb);
      skb_put_data(skb, p, copy);
  +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
      len -= copy;
      offset += copy;
  @@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 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;
-
      if (unlikely(xdpf->headroom < vi->hdr_len))
          return -EOVERFLOW;
  @@ -644,6 +646,7 @@ 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;

drivers/net/virtio_net.c +2 -1
@@ -683,8 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 745

Quoted 8 lines
drivers/net/virtio_net.c +1 -1
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
       }
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
-    if (!delta) {
+    if (!xdp_prog) {
           buf += header_offset;
           memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */

I prefer to make this an independent patch and cc stable.

Quoted 3 lines

Other looks good.

Thanks

I see. So I need to revert to delta from xdp_prog?

Thank you.

Quoted 77 lines
Quoted 76 lines
drivers/net/virtio_net.c +2 -0
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +767,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +2 -1
@@ -793,6 +800,7 @@ 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 +847,8 @@ 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);
@@ -848,24 +856,27 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
             switch (act) {

          case XDP_PASS:
+            metasize = xdp.data - xdp.data_meta;
+
              /* recalculate offset to account for any header -             * adjustments. Note other cases do not build an -             * skb and avoid using offset
+             * adjustments and minus the metasize to copy the +             * metadata in page_to_skb(). Note other cases do not +             * build an skb and avoid using offset
               */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len; +            offset = xdp.data - page_address(xdp_page) - +                 vi->hdr_len - metasize;
  -            /* recalculate len if xdp.data or xdp.data_end were -             * adjusted
+            /* recalculate len if xdp.data, xdp.data_end or +             * xdp.data_meta were adjusted
               */
-            len = xdp.data_end - xdp.data + vi->hdr_len; +            len = xdp.data_end - xdp.data + vi->hdr_len + metasize; /* 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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset, +                               len, PAGE_SIZE, false,
+                               metasize);
                  return head_skb;
              }
              break;

drivers/net/virtio_net.c +2 -0
@@ -921,7 +932,8 @@ 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))
#22 lore raw
Quoted 21 lines · Yuya Kusakabe
Quoted 7 lines · Jason Wang
Quoted 5 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Quoted 7 lines
Quoted 4 lines

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

I'm not sure this is correct since virtio-net claims to not support metadata by calling xdp_set_data_meta_invalid()?

virtio_net doesn't support by calling xdp_set_data_meta_invalid() for now.

https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n686
https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n842

Quoted 7 lines

And xdp_set_data_meta_invalid() are added by de8f3a83b0a0.

$ git blame ./drivers/net/virtio_net.c | grep xdp_set_data_meta_invalid
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  686)                xdp_set_data_meta_invalid(&xdp);
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  842)                xdp_set_data_meta_invalid(&xdp);

So I added Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access") to the comment.

Fixes basically means "must be backported to any kernel that has de8f3a83b0a0 in order to fix a bug". This looks more like a feature than a bug though, so I'm not sure Fixes
is approproate. Correct me if I'm wrong.

Quoted 181 lines
Quoted 96 lines
Quoted 84 lines

---
v5:
  - page_to_skb(): copy vnet header if hdr_valid without checking metasize.
  - receive_small(): do not copy vnet header if xdp_prog is availavle.
  - __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
  - improve comments.
v4:
  - improve commit message
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.
---
  drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------

  1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +1 -1
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
  @@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi, copy = skb_tailroom(skb);
      skb_put_data(skb, p, copy);
  +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
      len -= copy;
      offset += copy;
  @@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 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;
-
      if (unlikely(xdpf->headroom < vi->hdr_len))
          return -EOVERFLOW;
  @@ -644,6 +646,7 @@ 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;

drivers/net/virtio_net.c +2 -1
@@ -683,8 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 745

Quoted 8 lines
drivers/net/virtio_net.c +1 -1
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
       }
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
-    if (!delta) {
+    if (!xdp_prog) {
           buf += header_offset;
           memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */

I prefer to make this an independent patch and cc stable.

Quoted 3 lines

Other looks good.

Thanks

I see. So I need to revert to delta from xdp_prog?

Thank you.

Quoted 77 lines
Quoted 76 lines
drivers/net/virtio_net.c +2 -0
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +767,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +2 -1
@@ -793,6 +800,7 @@ 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 +847,8 @@ 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);
@@ -848,24 +856,27 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
             switch (act) {

          case XDP_PASS:
+            metasize = xdp.data - xdp.data_meta;
+
              /* recalculate offset to account for any header -             * adjustments. Note other cases do not build an -             * skb and avoid using offset
+             * adjustments and minus the metasize to copy the +             * metadata in page_to_skb(). Note other cases do not +             * build an skb and avoid using offset
               */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len; +            offset = xdp.data - page_address(xdp_page) - +                 vi->hdr_len - metasize;
  -            /* recalculate len if xdp.data or xdp.data_end were -             * adjusted
+            /* recalculate len if xdp.data, xdp.data_end or +             * xdp.data_meta were adjusted
               */
-            len = xdp.data_end - xdp.data + vi->hdr_len; +            len = xdp.data_end - xdp.data + vi->hdr_len + metasize; /* 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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset, +                               len, PAGE_SIZE, false,
+                               metasize);
                  return head_skb;
              }
              break;

drivers/net/virtio_net.c +2 -0
@@ -921,7 +932,8 @@ 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))
#23 lore raw
In reply to Yuya Kusakabe
Quoted 125 lines · Yuya Kusakabe
Quoted 7 lines · Jason Wang
Quoted 5 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Quoted 7 lines
Quoted 4 lines

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

I'm not sure this is correct since virtio-net claims to not support metadata by calling xdp_set_data_meta_invalid()?

virtio_net doesn't support by calling xdp_set_data_meta_invalid() for now.

https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n686
https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n842

And xdp_set_data_meta_invalid() are added by de8f3a83b0a0.

$ git blame ./drivers/net/virtio_net.c | grep xdp_set_data_meta_invalid
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  686)                xdp_set_data_meta_invalid(&xdp);
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  842)                xdp_set_data_meta_invalid(&xdp);

So I added Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access") to the comment.

Quoted 96 lines
Quoted 84 lines

---
v5:
  - page_to_skb(): copy vnet header if hdr_valid without checking metasize.
  - receive_small(): do not copy vnet header if xdp_prog is availavle.
  - __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
  - improve comments.
v4:
  - improve commit message
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.
---
  drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------

  1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +1 -1
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
  @@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi, copy = skb_tailroom(skb);
      skb_put_data(skb, p, copy);
  +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
      len -= copy;
      offset += copy;
  @@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 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;
-
      if (unlikely(xdpf->headroom < vi->hdr_len))
          return -EOVERFLOW;
  @@ -644,6 +646,7 @@ 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;

drivers/net/virtio_net.c +2 -1
@@ -683,8 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 745

Quoted 8 lines
drivers/net/virtio_net.c +1 -1
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
       }
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
-    if (!delta) {
+    if (!xdp_prog) {
           buf += header_offset;
           memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */

I prefer to make this an independent patch and cc stable.

Quoted 7 lines
Quoted 3 lines

Other looks good.

Thanks

I see. So I need to revert to delta from xdp_prog?

Thank you.

So maybe send a 2 patch series: 1/2 is this chunk with the appropriate description. Actually for netdev David prefers that people do not cc stable directly, just include Fixes tag and mention in the commit log it's also needed for stable. Patch 2/2 is the rest handling metadata.

Quoted 77 lines
Quoted 77 lines
Quoted 76 lines
drivers/net/virtio_net.c +2 -0
   +    if (metasize)
+        skb_metadata_set(skb, metasize);
+

  err:
      return skb;
  @@ -760,8 +767,8 @@ static struct sk_buff *receive_big(struct net_device *dev, struct virtnet_rq_stats *stats)
  {
      struct page *page = buf;
-    struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
-                      PAGE_SIZE, true);
+    struct sk_buff *skb =
+        page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0); stats->bytes += len - vi->hdr_len;
      if (unlikely(!skb))

drivers/net/virtio_net.c +2 -1
@@ -793,6 +800,7 @@ 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 +847,8 @@ 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);
@@ -848,24 +856,27 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
             switch (act) {

          case XDP_PASS:
+            metasize = xdp.data - xdp.data_meta;
+
              /* recalculate offset to account for any header -             * adjustments. Note other cases do not build an -             * skb and avoid using offset
+             * adjustments and minus the metasize to copy the +             * metadata in page_to_skb(). Note other cases do not +             * build an skb and avoid using offset
               */
-            offset = xdp.data -
-                    page_address(xdp_page) - vi->hdr_len; +            offset = xdp.data - page_address(xdp_page) - +                 vi->hdr_len - metasize;
  -            /* recalculate len if xdp.data or xdp.data_end were -             * adjusted
+            /* recalculate len if xdp.data, xdp.data_end or +             * xdp.data_meta were adjusted
               */
-            len = xdp.data_end - xdp.data + vi->hdr_len; +            len = xdp.data_end - xdp.data + vi->hdr_len + metasize; /* 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);
+                head_skb = page_to_skb(vi, rq, xdp_page, offset, +                               len, PAGE_SIZE, false,
+                               metasize);
                  return head_skb;
              }
              break;

drivers/net/virtio_net.c +2 -0
@@ -921,7 +932,8 @@ 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))
#24 lore raw
Quoted 121 lines · Michael S. Tsirkin
Quoted 120 lines · Yuya Kusakabe
Quoted 11 lines · Jason Wang
Quoted 9 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

I'm not sure this is correct since virtio-net claims to not support metadata by calling xdp_set_data_meta_invalid()?

virtio_net doesn't support by calling xdp_set_data_meta_invalid() for now.

https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n686
https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n842

And xdp_set_data_meta_invalid() are added by de8f3a83b0a0.

$ git blame ./drivers/net/virtio_net.c | grep xdp_set_data_meta_invalid
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  686)                xdp_set_data_meta_invalid(&xdp);
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  842)                xdp_set_data_meta_invalid(&xdp);

So I added Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access") to the comment.

Quoted 95 lines
Quoted 82 lines

---
v5:
  - page_to_skb(): copy vnet header if hdr_valid without checking metasize.
  - receive_small(): do not copy vnet header if xdp_prog is availavle.
  - __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
  - improve comments.
v4:
  - improve commit message
v3:
  - fix preserve the vnet header in receive_small().
v2:
  - keep copy untouched in page_to_skb().
  - preserve the vnet header in receive_small().
  - fix indentation.

---
   drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------
   1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +8 -5
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
   @@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
           copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
       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;
-
       if (unlikely(xdpf->headroom < vi->hdr_len))
           return -EOVERFLOW;
   @@ -644,6 +646,7 @@ 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 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 745

Quoted 8 lines
drivers/net/virtio_net.c +1 -1
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
       }
       skb_reserve(skb, headroom - delta);
       skb_put(skb, len);
-    if (!delta) {
+    if (!xdp_prog) {
           buf += header_offset;
           memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
       } /* keep zeroed vnet hdr since packet was changed by bpf */

I prefer to make this an independent patch and cc stable.

Other looks good.

Thanks

Quoted 8 lines
Quoted 3 lines

I see. So I need to revert to delta from xdp_prog?

Thank you.

So maybe send a 2 patch series: 1/2 is this chunk with the appropriate description. Actually for netdev David prefers that people do not cc stable directly, just include Fixes tag and mention in the commit log it's also needed for stable. Patch 2/2 is the rest handling metadata.

+1

Thanks

#25 lore raw
Quoted 133 lines · Jason Wang
Quoted 121 lines · Michael S. Tsirkin
Quoted 120 lines · Yuya Kusakabe
Quoted 11 lines · Jason Wang
Quoted 9 lines · Yuya Kusakabe

Implement support for transferring XDP meta data into skb for virtio_net driver; before calling into the program, xdp.data_meta points to xdp.data, where on program return with pass verdict, we call into skb_metadata_set().

Tested with the script at
https://github.com/higebu/virtio_net-xdp-metadata-test.

Fixes de8f3a83b0a0 ("bpf: add meta pointer for direct access")

I'm not sure this is correct since virtio-net claims to not support metadata by calling xdp_set_data_meta_invalid()?

virtio_net doesn't support by calling xdp_set_data_meta_invalid() for now.

https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n686
https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/drivers/net/virtio_net.c?id=e42da4c62abb547d9c9138e0e7fcd1f36057b5e8#n842

And xdp_set_data_meta_invalid() are added by de8f3a83b0a0.

$ git blame ./drivers/net/virtio_net.c | grep xdp_set_data_meta_invalid
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  686)                xdp_set_data_meta_invalid(&xdp);
de8f3a83b0a0f (Daniel Borkmann           2017-09-25 02:25:51 +0200  842)                xdp_set_data_meta_invalid(&xdp);

So I added Fixes: de8f3a83b0a0 ("bpf: add meta pointer for direct access") to the comment.

Quoted 95 lines
Quoted 82 lines

---
v5:
   - page_to_skb(): copy vnet header if hdr_valid without checking metasize.
   - receive_small(): do not copy vnet header if xdp_prog is availavle.
   - __virtnet_xdp_xmit_one(): remove the xdp_set_data_meta_invalid().
   - improve comments.
v4:
   - improve commit message
v3:
   - fix preserve the vnet header in receive_small().
v2:
   - keep copy untouched in page_to_skb().
   - preserve the vnet header in receive_small().
   - fix indentation.
---
   drivers/net/virtio_net.c | 54 ++++++++++++++++++++++++----------------

   1 file changed, 33 insertions(+), 21 deletions(-)
drivers/net/virtio_net.c +1 -1
index 2fe7a3188282..4ea0ae60c000 100644
@@ -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;
@@ -393,6 +393,7 @@ 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);
   @@ -405,6 +406,11 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi, copy = skb_tailroom(skb);
       skb_put_data(skb, p, copy);
   +    if (metasize) {
+        __skb_pull(skb, metasize);
+        skb_metadata_set(skb, metasize);
+    }
+
       len -= copy;
       offset += copy;
   @@ -450,10 +456,6 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi, 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;
-
       if (unlikely(xdpf->headroom < vi->hdr_len))
           return -EOVERFLOW;
   @@ -644,6 +646,7 @@ 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;

drivers/net/virtio_net.c +2 -1
@@ -683,8 +686,8 @@ 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);
@@ -695,6 +698,7 @@ 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++;

drivers/net/virtio_net.c static struct sk_buff *receive_small(struct net_device *dev, · line 745

Quoted 8 lines
drivers/net/virtio_net.c +1 -1
@@ -735,11 +739,14 @@ static struct sk_buff *receive_small(struct net_device *dev,
        }
        skb_reserve(skb, headroom - delta);
        skb_put(skb, len);
-    if (!delta) {
+    if (!xdp_prog) {
            buf += header_offset;
            memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
        } /* keep zeroed vnet hdr since packet was changed by bpf */

I prefer to make this an independent patch and cc stable.

Other looks good.

Thanks

Quoted 8 lines
Quoted 3 lines

I see. So I need to revert to delta from xdp_prog?

Thank you.

So maybe send a 2 patch series: 1/2 is this chunk with the appropriate description. Actually for netdev David prefers that people do not cc stable directly, just include Fixes tag and mention in the commit log it's also needed for stable. Patch 2/2 is the rest handling metadata.

Quoted 5 lines

+1

Thanks

Thank you for the detailed explanation. I will make a 2 patch series.

#26 lore raw
In reply to Yuya Kusakabe

drivers/net/virtio_net.c static struct sk_buff *receive_mergeable(struct net_device *dev, · line 876

Quoted 12 lines · Yuya Kusakabe
drivers/net/virtio_net.c +3 -2
@@ -852,8 +868,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			 * adjustments. Note other cases do not build an
 			 * skb and avoid using offset
 			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
 			/* recalculate len if xdp.data or xdp.data_end were
 			 * adjusted

Tricky to get one's head around.
Can you pls update the comment above to document the new math?

Signature · 1 line
MST
#27 lore raw
Quoted 11 lines · Michael S. Tsirkin
Quoted 10 lines · Yuya Kusakabe
drivers/net/virtio_net.c +3 -2
@@ -852,8 +868,9 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
 			 * adjustments. Note other cases do not build an
 			 * skb and avoid using offset
 			 */
-			offset = xdp.data -
-					page_address(xdp_page) - vi->hdr_len;
+			metasize = xdp.data - xdp.data_meta;
+			offset = xdp.data - page_address(xdp_page) -
+				 vi->hdr_len - metasize;
 
Quoted 6 lines

drivers/net/virtio_net.c static struct sk_buff *receive_mergeable(struct net_device *dev, · line 876

Quoted 2 lines
drivers/net/virtio_net.c +0 -0
 			/* recalculate len if xdp.data or xdp.data_end were
 			 * adjusted

Tricky to get one's head around.
Can you pls update the comment above to document the new math?

Thank you for your review.

I will update the comment on next patch.