On Wed, May 31, 2023 at 12:36 AM David Ahern [off-list ref] wrote:

On 5/30/23 1:39 PM, Jakub Kicinski wrote:

quoted

On Tue, 30 May 2023 17:48:22 +0200 Eric Dumazet wrote:

quoted

quoted

tcp_gro_complete seems fairly trivial. Any reason not to make it an
inline and avoid another function call in the datapath?

Probably, although it is a regular function call, not an indirect one.

In the grand total of driver rx napi + GRO cost, saving a few cycles
per GRO completed packet is quite small.

IOW please make sure you include the performance analysis quantifying
the win, if you want to make this a static inline. Or let us know if
the patch is good as is, I'm keeping it in pw for now.

I am not suggesting holding up this patch; just constantly looking for
these little savings here and there to keep lowering the overhead.

100G, 1500 MTU, line rate is 8.3M pps so GRO wise that would be ~180k
fewer function calls.

Here with 4K MTU, this is called 67k per second

An __skb_put() instead of skb_put() in a driver (eg mlx5e_build_linear_skb())
would have 45x more impact, and would still be noise.