On Mon, Jan 20, 2025 at 09:00:08PM +0300, Vitaliy Makkoveev wrote:
> Is is such necessary? I suspect extra shifts on some architectures.
> 
> > On 20 Jan 2025, at 20:36, Alexander Bluhm <bluhm@openbsd.org> wrote:
> > 
> > -	u_int			 ph_ifidx;	/* rcv interface index */
> > +	u_int16_t		 ph_ifidx;	/* rcv interface index */

I think there are issues when making the mbuf header larger.  Like
TCP ACK packets not fitting in a single mbuf.  I tried to use
ph_cookie but that did not work.  It feels like something else might
use it in concurently.  So either add more paramters to if_input()
function pointers and pass the softnet down, or fixing the ph_cookie
approach, or making interface index 16 bit thoughout the stack, or
use pkthdr_pf pad[1] instead, or some other bright idea.

At least my diff shows that we can avoid lock flipping and that is
the solution for single stream performance.

bluhm