From: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
Date: Thu, 09 Feb 2023 21:58:07 +0100
> Alexander Lobakin <alexandr.lobakin@xxxxxxxxx> writes:
>
>> From: Alexander Lobakin <alexandr.lobakin@xxxxxxxxx>
>> Date: Thu, 9 Feb 2023 18:28:27 +0100
>>
>>> &xdp_buff and &xdp_frame are bound in a way that
>>>
>>> xdp_buff->data_hard_start == xdp_frame
>>
>> [...]
>>
>>> diff --git a/net/bpf/test_run.c b/net/bpf/test_run.c
>>> index 2723623429ac..c3cce7a8d47d 100644
>>> --- a/net/bpf/test_run.c
>>> +++ b/net/bpf/test_run.c
>>> @@ -97,8 +97,11 @@ static bool bpf_test_timer_continue(struct bpf_test_timer *t, int iterations,
>>> struct xdp_page_head {
>>> struct xdp_buff orig_ctx;
>>> struct xdp_buff ctx;
>>> - struct xdp_frame frm;
>>> - u8 data[];
>>> + union {
>>> + /* ::data_hard_start starts here */
>>> + DECLARE_FLEX_ARRAY(struct xdp_frame, frm);
>>> + DECLARE_FLEX_ARRAY(u8, data);
>>> + };
>>
>> BTW, xdp_frame here starts at 112 byte offset, i.e. in 16 bytes a
>> cacheline boundary is hit, so xdp_frame gets sliced into halves: 16
>> bytes in CL1 + 24 bytes in CL2. Maybe we'd better align this union to
>> %NET_SKB_PAD / %SMP_CACHE_BYTES / ... to avoid this?
>
> Hmm, IIRC my reasoning was that both those cache lines will be touched
> by the code in xdp_test_run_batch(), so it wouldn't matter? But if
> there's a performance benefit I don't mind adding an explicit alignment
> annotation, certainly!
Let me retest both ways and will see. I saw some huge CPU loads on
reading xdpf in ice_xdp_xmit(), so that was my first thought.
>
>> (but in bpf-next probably)
>
> Yeah...
>
> -Toke
>
Thanks,
Olek
