Re: [PATCH net] net: avoid 32 x truesize under-estimation for tiny skbs

[Eric Dumazet <edumazet@google.com>]

On Fri, Apr 2, 2021 at 12:35 PM Xuan Zhuo <xuanzhuo@linux.alibaba.com> wrote:
>
>
> Great!!
>
> test cmd:
>         sockperf tp --tcp --ip <ip> -m 8192 -t 300
>
>                 server softirq cpu | server sockperf cpu | rxpck/s | rxkB/s
> page_frag_alloc 100%.                86.1%                 2211785   3270145
> kmalloc         100%.                100%                  2226950   3292571
> copy ETH        100%                 69.3%                 2140552   3163408
>
> Compared with page_frag_alloc, the performance has also increased by 19.4%. The
> reason for such a big improvement I think is that although merge is turned on in
> my test environment, there is only one buffer for each packet received.
>
> So in the page_frag_alloc scenario, when performing gro merge, one skb will
> occupy two frags, one is used to store the data in the linear space, and the
> other is a frag in the skb frags. Now, the linear space only has the header, so
> one skb only occupies one frag space.
>
> In the case of using page_frag_alloc, an skb can only contain the original 8
> packets. Now a skb merges the original 17 packets

Exactly. Never put payload in skb->head, unless you are sure that :
1) All payload can be put there. (or it is filling a full page,
followed by remaining payload in further frags for big MTU)
2) skb->head is backed by a page fraf as well (skb->head_frag == 1)

>
> It's really exciting.
>
> I fixed the retval check of pskb_may_pull and added a pskb_may_pull call.
> Here is the patch I used for testing.
>
> Thanks very much.

Excellent, I am going to submit this patch officially, since it is not
risky for stable kernels.
I will give you credits as Reporter, and co-developer.

Thanks for your help !