Re: [PATCH 0/2] abstract napi tracking strategy

[Olivier Langlois <olivier@trillion01.com>]

On Fri, 2024-08-16 at 15:26 +0100, Pavel Begunkov wrote:
> I pushed what I had (2 last patches), you can use it as a
> reference, but be aware that it's a completely untested
> draft with some obvious problems and ugly uapi.
> 
> https://github.com/isilence/linux.git manual-napi
> https://github.com/isilence/linux/commits/manual-napi/
> 
I did review what you have done...
I like your take on this feature idea... especially the autoremove
field.

the one aspect that I prefer in my version over yours, it is that my
implementation avoids totally to call __io_napi_add() which includes a
table lookup from io_poll_check_events() which is big chunk of the
dynamic tracking overhead.

also, when the napi devices are iterated in __io_napi_do_busy_loop(),
my version totally remove the conditional branching from the loop by
having 2 distinct busy loop functions.

but I guess the last point is arguably unimportant since the code is
busy polling...

Anyway, I had the time to create a v2 version of my implementation to
address the remarks made about v1 that I have completed testing. It has
been running on my system for the last 24h flawlessly...

I will post it here shortly. Please take a look at it and pick the best
features of both implementation.