Re: [PATCH v2] kpageflags: respect folio head-page flag placement

[Kefeng Wang]

On 2023/11/7 23:34, Gregory Price wrote:
> On Tue, Nov 07, 2023 at 09:03:53AM +0800, Kefeng Wang wrote:
> > On 2023/10/31 12:34, Gregory Price wrote:
> > > On Tue, Oct 31, 2023 at 09:13:44AM +0000, Matthew Wilcox wrote:
> > > > On Mon, Oct 30, 2023 at 07:41:23PM -0400, Gregory Price wrote:
> > > > > On Mon, Oct 30, 2023 at 11:22:18PM +0000, Matthew Wilcox wrote:
> > > > > > On Mon, Oct 30, 2023 at 02:00:05PM -0400, Gregory Price wrote:
> > > > > > > kpageflags reads page-flags directly from the page, even when the
> > > > > > > respective flag is only updated on the headpage of a folio.
> > > > > > >
> > > > > > > Update bitchecks to use PAGEFLAG() interfaces to check folio for the
> > > > > > > referenced, dirty, lru, active, and unevictable bits.
> > > > > >
> > > > > > But uptodate, writeback and reclaim (amongst others) are also defined
> > > > > > only on the head page.
> > > > >
> > > > > Ah yes i was only looking at the things defined w/ PAGEFLAG defines in
> > > > > page-flags.h. I'll give it full once over can collect them all, my bad.
> > > > >
> > > > > (also i forgot to update my commit message)
> > > > >
> > > > > Quick question here since i have your attention: any recommendation on
> > > > > what to do for ONLY_HEAD flags?  If the provided page is not the head,
> > > > > should the flag report 0... or whatever the head says?
> > > >
> > > > Thinking about it some more, really almost all flags are per-folio, not
> > > > per-page.  The only exceptions are HWPoison and AnonExclusive.  So
> > > > probably the right way to do this is to make k = folio->flags, and
> > > > then just change a few places rather than changing all the places that
> > > > test 'k'.
> > >
> > > Funny enough that's what i originally did but was confident it was
> > > correct so walked it back.  I'll take another crack at it.
> >
> > Hi Gregory, any update?
> > I changed stable_page_flags[1] when try to remove page idle wrapper,
> > Matthew pointed it will conflict with this, I could redo my patch
> > based on your new version:)
> >
> > [1]https://lore.kernel.org/linux-mm/20231103072906.2000381-5-wangkefeng.wang@xxxxxxxxxx/
> >
> > Thanks.
>
> looks like we were noticing the same thing.  I haven't done any further
> work, got caught up in another project.

Yes, I see your "Node Weights and Weighted Interleave", this is an 
interesting topic, we need some easy and efficient way to use tiered
memory.

> Matthew last pointed out:
>
> "probably the right way to do this is to make k = folio->flags, and then
> just change a few places rather than changing all the places that test
> 'k'."
>
> https://lore.kernel.org/linux-mm/ZUDFSEvpxxoGWmdG@xxxxxxxxxxxxxxxxxxxx/
>
> I took a quick look, and the only thing I'm not confident about is that
> some flags are stored in the head page, and some are stored on the
> second page.
>
> /* Which page is the flag stored in */
> #define FOLIO_PF_ANY            0
> #define FOLIO_PF_HEAD           0
> #define FOLIO_PF_ONLY_HEAD      0
> #define FOLIO_PF_NO_TAIL        0
> #define FOLIO_PF_NO_COMPOUND    0
> #define FOLIO_PF_SECOND         1
>
> There's only a handful, so yeah the best way is probably to go ahead and
> swap k = page->flags for k = *folio_flags(folio, 0) and then handle the
> couple of outliars.

If you don't mind, I maybe try to convert it in my changes.

Thanks.
> ~Gregory