Re: [net-next v2] virtio_net: replace netdev_alloc_skb_ip_align() with napi_alloc_skb()

[Michael S. Tsirkin]

On Fri, Mar 18, 2016 at 04:42:48PM +0100, Paolo Abeni wrote:
> This gives small but noticeable rx performance improvement (2-3%)
> and will allow exploiting future napi improvement.
> 
> Signed-off-by: Paolo Abeni 

I am not sure this is necessarily worth doing for this dumb hardware.
I queued v1 in vhost tree for now, let's see some performance
numbers before we start changing about other paths.

> --
>  v2: replace also netdev_alloc_skb_ip_align() invocation in
>   add_recvbuf_small(), suggested by Venkatesh Srinivas
> ---
>  drivers/net/virtio_net.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index fb0eae4..100e039 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -260,7 +260,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info 
> *vi,
>   p = page_address(page) + offset;
>  
>   /* copy small packet so we can reuse these pages for small data */
> - skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
> + skb = napi_alloc_skb(>napi, GOOD_COPY_LEN);
>   if (unlikely(!skb))
>   return NULL;
>  
> @@ -541,7 +541,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, 
> struct receive_queue *rq,
>   struct virtio_net_hdr_mrg_rxbuf *hdr;
>   int err;
>  
> - skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
> + skb = __napi_alloc_skb(>napi, GOOD_PACKET_LEN, gfp);
>   if (unlikely(!skb))
>   return -ENOMEM;
>  
> -- 
> 1.8.3.1

Re: [net-next v2] virtio_net: replace netdev_alloc_skb_ip_align() with napi_alloc_skb()

[Venkatesh Srinivas]

On Fri, Mar 18, 2016 at 8:42 AM, Paolo Abeni  wrote:
> This gives small but noticeable rx performance improvement (2-3%)
> and will allow exploiting future napi improvement.
>
> Signed-off-by: Paolo Abeni 
>
> --
>  v2: replace also netdev_alloc_skb_ip_align() invocation in
> add_recvbuf_small(), suggested by Venkatesh Srinivas
> ---
>  drivers/net/virtio_net.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index fb0eae4..100e039 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -260,7 +260,7 @@ static struct sk_buff *page_to_skb(struct virtnet_info 
> *vi,
> p = page_address(page) + offset;
>
> /* copy small packet so we can reuse these pages for small data */
> -   skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
> +   skb = napi_alloc_skb(>napi, GOOD_COPY_LEN);
> if (unlikely(!skb))
> return NULL;
>
> @@ -541,7 +541,7 @@ static int add_recvbuf_small(struct virtnet_info *vi, 
> struct receive_queue *rq,
> struct virtio_net_hdr_mrg_rxbuf *hdr;
> int err;
>
> -   skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
> +   skb = __napi_alloc_skb(>napi, GOOD_PACKET_LEN, gfp);
> if (unlikely(!skb))
> return -ENOMEM;

Sorry, I should've mentioned more in my first reply --
add_recvbuf_small is called from try_fill_recv in both
NAPI and non-NAPI context (NAPI: via virtnet_receive from
virtnet_poll; non-NAPI: via scheduled work,
via virtnet_probe, and via virtnet_restore).

I don't believe you can use __napi_alloc_skb() for the non-NAPI paths,
it'd be a bit more work to plumb
the context as an argument if you think its worth doing here.

Looking forward to bulk skb allocation either way.

HTH,
-- vs;