Re: [PATCH v2 0/5] fork: Page operation cleanups in the fork code

[Mateusz Guzik <mjguzik@gmail.com>]

On Fri, May 9, 2025 at 8:58 AM Linus Walleij <linus.walleij@linaro.org> wrote:
> And when the allocated stacks are reused, as you say the NUMA node
> is completely ignored.
>
> The rationale is in commit ac496bf48d97f2503eaa353996a4dd5e4383eaf0.
>
> I wonder what would be the best way to make this better? Improve
> the cache or assume that the MM/VMM itself should be able to sort this
> out in an optimized way by now and put it to the test by simply
> deleting the cache and see what happens?
>

The cache can be trivially touched up to reduce the problem:
scope_guard(preempt) around allocation and free from the cache and
check that the requested domain (including no domain) matches. There
is some already existing code to check which domains back pages on
free.

The real problem, expanding beyond just stack allocation, is that at
some point you are going to get a process initialized on one domain
but only ever running at another.

I was speculating this could be mostly fixed by asking the scheduler
in which domain does it think the process will run and then do all the
allocations with that domain as the requested target. Sorting this out
is way beyond the scope of the immediate problem though.

That aside the current caching mechanism is quite toy-ish, but I don't
have time right now to implement a RFC with something better.
-- 
Mateusz Guzik <mjguzik gmail.com>