Re: [PATCH] mm: madvise: fix uneven accounting of psi

[Suren Baghdasaryan <surenb@google.com>]

On Mon, Jun 26, 2023 at 7:31 AM Charan Teja Kalla
<quic_charante@quicinc.com> wrote:
>
> Hi Suren,
>
> On 6/10/2023 4:43 AM, Suren Baghdasaryan wrote:
> >>> I can see use for that from userspace to detect incorrect madvise()
> >>> and adjust its aggressiveness. I think the API might get a bit complex
> >>> because of the need to associate refaults with specific madvise()/VMAs
> >>> to understand which hint was incorrect and adjust the behavior.
> >>> Instead what is the opinion about giving an MADV_PAGEOUT_INACTIVE
> >> interface which does operate on a page only If it is on the inactive
> >> list and !PageWorkingset ?
> > IOW you want a less aggressive mechanism which can be used by the
> > userspace to tell the kernel "I think these pages won't be used but
> > I'm not 100% sure, so drop them only if they are inactive"?
> >  I don't know how much that will help when the madvise() ends up being
> > wrong but maybe you can quickly experiment and tell us if the
> > difference is substantial?
>
> We did some extensive testing on Android and this ask is not helping us
> much. I am really not sure if there is some other usecase that can
> benefit from this. So, for now I just stick to your suggestion of making
> the pages on the Active list as the Workingset at the time of pageout.
> >

Thanks for checking that. Your plan SGTM.

>
> Thanks.
>