# Re: \[RFC net-next\] net: veth: reduce page\_pool memory footprint using half page per-buffer

- From: Jesper Dangaard Brouer \<hidden\>
- Date: 2023-05-16 16:12:19 
- List: netdev
- Page: https://ratatoskr.run/netdev/2023/05/7848426
- Thread: https://ratatoskr.run/netdev/2023/05/7848299/t.md

```text

Cc. Alex Duyck + Eric, please criticize my idea below.

On 12/05/2023 15.08, Lorenzo Bianconi wrote:
> In order to reduce page_pool memory footprint, rely on
> page_pool_dev_alloc_frag routine and reduce buffer size
> (VETH_PAGE_POOL_FRAG_SIZE) to PAGE_SIZE / 2 in order to consume one page
> for two 1500B frames. Reduce VETH_XDP_PACKET_HEADROOM to 192 from 256
> (XDP_PACKET_HEADROOM) to fit max_head_size in VETH_PAGE_POOL_FRAG_SIZE.
> Please note, using default values (CONFIG_MAX_SKB_FRAGS=17), maximum
> supported MTU is now reduced to 36350B.
>
> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> ---
>   drivers/net/veth.c | 39 +++++++++++++++++++++++++--------------
>   1 file changed, 25 insertions(+), 14 deletions(-)
>
> diff --git a/drivers/net/veth.c b/drivers/net/veth.c
> index 614f3e3efab0..0e648703cccf 100644
> --- a/drivers/net/veth.c
> +++ b/drivers/net/veth.c
> @@ -31,9 +31,12 @@
>   #define DRV_NAME	"veth"
>   #define DRV_VERSION	"1.0"
>   
> -#define VETH_XDP_FLAG		BIT(0)
> -#define VETH_RING_SIZE		256
> -#define VETH_XDP_HEADROOM	(XDP_PACKET_HEADROOM + NET_IP_ALIGN)
> +#define VETH_XDP_FLAG			BIT(0)
> +#define VETH_RING_SIZE			256
> +#define VETH_XDP_PACKET_HEADROOM	192
> +#define VETH_XDP_HEADROOM		(VETH_XDP_PACKET_HEADROOM + \
> +					 NET_IP_ALIGN)
> +#define VETH_PAGE_POOL_FRAG_SIZE	2048
>   
>   #define VETH_XDP_TX_BULK_SIZE	16
>   #define VETH_XDP_BATCH		16
> @@ -736,7 +739,7 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
>   	if (skb_shared(skb) || skb_head_is_locked(skb) ||
>   	    skb_shinfo(skb)->nr_frags ||
>   	    skb_headroom(skb) < XDP_PACKET_HEADROOM) {
> -		u32 size, len, max_head_size, off;
> +		u32 size, len, max_head_size, off, pp_off;
>   		struct sk_buff *nskb;
>   		struct page *page;
>   		int i, head_off;
> @@ -747,17 +750,20 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
>   		 *
>   		 * Make sure we have enough space for linear and paged area
>   		 */
> -		max_head_size = SKB_WITH_OVERHEAD(PAGE_SIZE -
> +		max_head_size = SKB_WITH_OVERHEAD(VETH_PAGE_POOL_FRAG_SIZE -
>   						  VETH_XDP_HEADROOM);
> -		if (skb->len > PAGE_SIZE * MAX_SKB_FRAGS + max_head_size)
> +		if (skb->len >
> +		    VETH_PAGE_POOL_FRAG_SIZE * MAX_SKB_FRAGS + max_head_size)
>   			goto drop;
>   
>   		/* Allocate skb head */
> -		page = page_pool_dev_alloc_pages(rq->page_pool);

It seems wasteful to allocate a full page PAGE_SIZE.

> +		page = page_pool_dev_alloc_frag(rq->page_pool, &pp_off,
> +						VETH_PAGE_POOL_FRAG_SIZE);

Allocating PAGE_SIZE/2 isn't much better.

At this point we already know the skb->len (and skb_headlen).

Why don't we allocated the size that we need?

See page_frag_alloc() system invented by Eric and Duyck.


>   		if (!page)
>   			goto drop;
>   
> -		nskb = napi_build_skb(page_address(page), PAGE_SIZE);
> +		nskb = napi_build_skb(page_address(page) + pp_off,
> +				      VETH_PAGE_POOL_FRAG_SIZE);
>   		if (!nskb) {
>   			page_pool_put_full_page(rq->page_pool, page, true);
>   			goto drop;
> @@ -782,15 +788,18 @@ static int veth_convert_skb_to_xdp_buff(struct veth_rq *rq,
>   		len = skb->len - off;
>   
>   		for (i = 0; i < MAX_SKB_FRAGS && off < skb->len; i++) {
> -			page = page_pool_dev_alloc_pages(rq->page_pool);
> +			page = page_pool_dev_alloc_frag(rq->page_pool, &pp_off,
> +							VETH_PAGE_POOL_FRAG_SIZE);
>   			if (!page) {
>   				consume_skb(nskb);
>   				goto drop;
>   			}
>   
> -			size = min_t(u32, len, PAGE_SIZE);
> -			skb_add_rx_frag(nskb, i, page, 0, size, PAGE_SIZE);
> -			if (skb_copy_bits(skb, off, page_address(page),
> +			size = min_t(u32, len, VETH_PAGE_POOL_FRAG_SIZE);
> +			skb_add_rx_frag(nskb, i, page, pp_off, size,
> +					VETH_PAGE_POOL_FRAG_SIZE);
> +			if (skb_copy_bits(skb, off,
> +					  page_address(page) + pp_off,
>   					  size)) {
>   				consume_skb(nskb);
>   				goto drop;
[...]

```