Re: [PATCH bpf-next v4] virtio_net: add XDP meta data support

[Jason Wang <jasowang@redhat.com>]

On 2020/2/4 下午3:16, Yuya Kusakabe wrote:
> 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")
> Signed-off-by: Yuya Kusakabe <yuya.kusakabe@gmail.com>
> ---
>   drivers/net/virtio_net.c | 47 ++++++++++++++++++++++++++++------------
>   1 file changed, 33 insertions(+), 14 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 2fe7a3188282..5fdd6ea0e3f1 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -371,7 +371,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info *vi,
>   				   struct receive_queue *rq,
>   				   struct page *page, unsigned int offset,
>   				   unsigned int len, unsigned int truesize,
> -				   bool hdr_valid)
> +				   bool hdr_valid, unsigned int metasize)
>   {
>   	struct sk_buff *skb;
>   	struct virtio_net_hdr_mrg_rxbuf *hdr;
> @@ -393,7 +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?


>   		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);


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.


>   		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.


>   			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).


> -		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;


Any reason for doing this?

Thanks


>   			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))