From: Jakub Kicinski <kuba@xxxxxxxxxx>
Date: Mon, 13 Jan 2025 13:01:04 -0800
> On Mon, 13 Jan 2025 14:50:02 +0100 Alexander Lobakin wrote:
>> From: Paolo Abeni <pabeni@xxxxxxxxxx>
>> Date: Thu, 9 Jan 2025 15:24:16 +0100
>>
>>> On 1/7/25 4:29 PM, Alexander Lobakin wrote:
>>>> @@ -623,21 +622,21 @@ static gro_result_t napi_skb_finish(struct napi_struct *napi,
>>>> return ret;
>>>> }
>>>>
>>>> -gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
>>>> +gro_result_t gro_receive_skb(struct gro_node *gro, struct sk_buff *skb)
>>>> {
>>>> gro_result_t ret;
>>>>
>>>> - skb_mark_napi_id(skb, napi);
>>>> + __skb_mark_napi_id(skb, gro->napi_id);
>>>
>>> Is this the only place where gro->napi_id is needed? If so, what about
>>> moving skb_mark_napi_id() in napi_gro_receive() and remove such field?
>>
>> Yes, only here. I thought of this, too. But this will increase the
>> object code of each napi_gro_receive() caller as it's now inline. So I
>> stopped on this one.
>> What do you think?
>
> What if we make napi_gro_receive() a real function (not inline)
> and tail call gro_receive_skb()? Is the compiler not clever
> enough too optimize that?
Worth trying. I'll be glad to do it that way if perf doesn't regress.
>
> Very nice work in general, the napi_id is gro sticks out..
Thanks,
Olek
