Re: [PATCH net-next 0/9] Split netmem from struct page

[Jesper Dangaard Brouer]

On 11/06/2025 22.48, Mina Almasry wrote:
> On Wed, Jun 11, 2025 at 7:24 AM Pavel Begunkov <asml.silence@xxxxxxxxx> wrote:
> > On 6/9/25 05:32, Byungchul Park wrote:
> > > Hi all,
> > >
> > > In this version, I'm posting non-controversial patches first.  I will
> > > post the rest more carefully later.  In this version, no update has been
> > > applied except excluding some patches from the previous version.  See
> > > the changes below.
> >
> > fwiw, I tried it with net_iov (zcrx), it didn't blow up during a
> > short test.
>
> FWIW, I ran my devmem TCP tests, and pp benchmark regression tests.
> Both look good to me. For the pp benchmark:

Thanks for verifying micro-benchmarks wasn't affected by this change.

> Before:
>
> Fast path results:
> no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.337 ns
>
> ptr_ring results:
> no-softirq-page_pool02 Per elem: 529 cycles(tsc) 196.073 ns

I'm surprised that ptr_ring return is this expensive on your system Mina.
I'll reply on the [v4] benchmark import patch, as this needs some more 
investigations.

[v4] 
https://lore.kernel.org/all/20250615205914.835368-1-almasrymina@xxxxxxxxxx/

> slow path results:
> no-softirq-page_pool03 Per elem: 554 cycles(tsc) 205.195 ns
>
> After:
>
> Fast path results:
> no-softirq-page_pool01 Per elem: 11 cycles(tsc) 4.401 ns
>
> ptr_ring results:
> no-softirq-page_pool02 Per elem: 530 cycles(tsc) 196.443 ns
>
> slow path results:
> no-softirq-page_pool03 Per elem: 551 cycles(tsc) 204.287 ns