Re: [PATCH v3 4/4] x86/folio_zero_user: multi-page clearing

[Mateusz Guzik]

On Wed, Apr 16, 2025 at 6:46 AM Ankur Arora <ankur.a.arora@xxxxxxxxxx> wrote:
> Mateusz Guzik <mjguzik@xxxxxxxxx> writes:
>
> > On Tue, Apr 15, 2025 at 11:46 PM Ankur Arora <ankur.a.arora@xxxxxxxxxx> wrote:
> >> Now, you could argue that the situation is less clear for 2MB pages.
> >>
> >
> > Well I was talking about 2MB. ;) I thought it is a foregone conclusion
> > that 1GB pages will be handled with non-temporal stores, but maybe I'm
> > crossing my wires.
>
> Or I am. I agree the situation is less clear for 2MB. I think for the
> immediate neighbourhood we clearly do want temporal.
>
> For the rest, L2 sizes are typically ~2MB or less. While there might be
> latency benefits for caching the rest of the page, there are clear
> negatives to blowing away state that was definitely useful in the
> near past.
>
> Would be good idea to quantify. Not sure what's a good benchmark to
> do that though.
>

L3 are not that great on many CPUs either, so you could argue you are
blown away the real stuff by zeroing a 2MB page.

I slept on it and I think proper evaluation is rather consuming and
tied to faulting more than one page at a time:
https://lore.kernel.org/linux-mm/CAGudoHGY012mwJqtGPUQ9mfQEVF1_otr9NSbbTYi_vazS09-CQ@xxxxxxxxxxxxxx/#t

I don't believe putting that effort in is a hard requirement at this
stage, albeit it would have been most welcome.

I think your patchset is a step in the right direction, even if the
specific policy is to change later.

All that is to say, I think this bit can be disregarded for now. I
definitely don't want to try to hold up the patchset.

cheers


--
Mateusz Guzik <mjguzik gmail.com>