Re: [PATCH] mm: move PG_slab flag to page_type

From: Hyeonggon Yoo
Date: Sat Oct 08 2022 - 00:21:45 EST


On Fri, Oct 07, 2022 at 07:02:35PM +0100, Matthew Wilcox wrote:
> On Fri, Oct 07, 2022 at 10:36:56PM +0900, Hyeonggon Yoo wrote:
> > > First, you say that folio_mapped() returns false for slab pages. That's
> > > only true for order-0 slab pages. For larger pages,
> > >
> > > if (!folio_test_large(folio))
> > > return atomic_read(&folio->_mapcount) >= 0;
> > > if (atomic_read(folio_mapcount_ptr(folio)) >= 0)
> > > return true;
> > >
> > > so that's going to depend what folio_mapcount_ptr() aliases with.
> >
> > IIUC it's true for order > 0 slab too.
> >
> > As slab pages are not mapped to userspace at all,
> > entire compound page nor base pages are not mapped to userspace.
> >
> > AFAIK followings are true for order > 0 slab:
> > - (first tail page)->compound_mapcount is -1
>
> That's the part I wasn't sure of. I think we do, in
> prep_compound_head().

Right, exactly!

>
> > - _mapcount of base pages are -1
> >
> > So:
> > folio_mapped() and page_mapped() (if applied to head page)
> > returns false for larger pages with this patch.
> >
> > I wrote simple testcase and did check that folio_mapped() and page_mapped()
> > returns false for both order-0 page and larger pages. (and SLAB
> > returned true for them before)

FYI, This is still true even after fixing my mistaken test case (see below)

> >
> > > Second, this patch changes the behaviour of PageSlab() when applied to
> > > tail pages.
> >
> > Altough it changes the way it checks the flag,
> >
> > it does not change behavior when applied to tail pages - PageSlab() on tail
> > page returns false with or without this patch.
>
> Really? It seems to me that it returns true at the moment. Look:
>
> __PAGEFLAG(Slab, slab, PF_NO_TAIL)
>
> #define PF_NO_TAIL(page, enforce) ({ \
> VM_BUG_ON_PGFLAGS(enforce && PageTail(page), page); \
> PF_POISONED_CHECK(compound_head(page)); })
>
> so AFAICS, PageSlab checks the Slab bit on the head page, not the
> tail page.

You are right. I misunderstood it due to my mistakenly written test case
(without passing __GFP_COMP... how silly of me :D)

Hmm okay, then I will implement PF_NO_TAIL policy that works on page_type.

>
> > If PageSlab() need to return true for tail pages too,
> > we may make it check page_type at head page.
> >
> > But I'm not sure when it the behavior is needed.
> > Can you please share your insight on this?
>
> There are tools like tools/vm/page-types.c which expect PageSlab to
> return true for tail pages.
>
> > > Which raises the further question of what PageBuddy(),
> > > PageTable(), PageGuard() and PageIsolated() should do for multi-page
> > > folios, if that is even possible.
> >
> > For users that uses real compound page like slab, we can make it check
> > page_type of head page. (if needed)
> >
> > But for cases David described, there isn't much thing we can do
> > except making them to use real compound pages.
> >
> > > Third, can we do this without that awkward __u16 thing? Perhaps
> > >
> > > -#define PG_buddy 0x00000080
> > > -#define PG_offline 0x00000100
> > > -#define PG_table 0x00000200
> > > -#define PG_guard 0x00000400
> > > +#define PG_buddy 0x00010000
> > > +#define PG_offline 0x00020000
> > > +#define PG_table 0x00040000
> > > +#define PG_guard 0x00080000
> > > +#define PG_slab 0x00100000
> > >
> > > ... and then use wrappers in slab.c to access the bottom 16 bits?
> >
> > Definitely! I prefer that way and will adjust in RFC v2.
> >
> > Thank you for precious feedback.
>
> No problem. I suggested (in an off-list email) that you consider counting
> 'active' by subtraction rather than addition because I have a feeling that
>
> int active(struct slab *slab)
> {
> return ~(slab->page_type | PG_slab);
> }
>
> would be better than
>
> int active(struct slab *slab)
> {
> return slab->page_type & 0xffff;
> }
>
> at least in part because you don't have to clear the bottom 16 bits of
> page_type when you clear PG_slab, and you don't have to re-set them
> when you set PG_slab.

Yeah, I was wondering what is the benefit of the that approach.
After implementing both approach, your suggestion seems better to me too.

Many thanks, Matthew!

--
Hyeonggon