Re: [PATCH net 6/9] virtio-net: make rx buf size estimation works for XDP

[John Fastabend <john.fastabend@gmail.com>]

On 16-12-23 06:37 AM, Jason Wang wrote:
> We don't update ewma rx buf size in the case of XDP. This will lead
> underestimation of rx buf size which causes host to produce more than
> one buffers. This will greatly increase the possibility of XDP page
> linearization.
> 
> Cc: John Fastabend <john.r.fastabend@intel.com>
> Signed-off-by: Jason Wang <jasowang@redhat.com>
> ---
>  drivers/net/virtio_net.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 0778dc8..77ae358 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -584,10 +584,12 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  				put_page(page);
>  				head_skb = page_to_skb(vi, rq, xdp_page,
>  						       0, len, PAGE_SIZE);
> +				ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
>  				return head_skb;
>  			}
>  			break;
>  		case XDP_TX:
> +			ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
>  			if (unlikely(xdp_page != page))
>  				goto err_xdp;
>  			rcu_read_unlock();
> @@ -596,6 +598,7 @@ static struct sk_buff *receive_mergeable(struct net_device *dev,
>  		default:
>  			if (unlikely(xdp_page != page))
>  				__free_pages(xdp_page, 0);
> +			ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
>  			goto err_xdp;
>  		}
>  	}
> 

Looks needed although I guess it will only be the case with
MTU > ETH_DATA_LEN because of the clamp in get_mergeable_buf_len().
Although XDP setup allows MTU up to page_size - hdr so certainly
will happen with ~MTU > 1500.

I need to add some various MTU tests to my setup.

Acked-by: John Fastabend <john.r.fastabend@intel.com>