Re: [PATCH v2 01/34] mm: Add PAGE_TYPE_OP folio functions

On Thu, May 25, 2023 at 01:38:54PM -0700, Vishal Moola wrote:
> On Thu, May 25, 2023 at 1:20 PM Mike Rapoport <[email protected]> wrote:
> >
> > On Thu, May 25, 2023 at 10:00:23AM -0700, Vishal Moola wrote:
> > > On Thu, May 25, 2023 at 1:56 AM Mike Rapoport <[email protected]> wrote:
> > > >
> > > > Hi,
> > > >
> > > > On Mon, May 01, 2023 at 12:27:56PM -0700, Vishal Moola (Oracle) wrote:
> > > > > No folio equivalents for page type operations have been defined, so
> > > > > define them for later folio conversions.
> > > >
> > > > Can you please elaborate why would we need folios for page table descriptors?
> > >
> > > Thanks for the review!
> > >
> > > These macros are for callers that care about the page type, i.e. Table and
> > > Buddy. Aside from accounting for those cases, the page tables don't use folios.
> > > These are more for the cleanliness of those callers.
> >
> > But why using folio APIs for PageType will be cleaner than using page APIs?
> > Do you have an example?
> 
> Ah, for example in mm/memory-failure.c there are a couple uses of PageTable.
> Like the line :
> if (folio_test_slab(folio) || PageTable(&folio->page) ||
> folio_test_reserved(folio))
> where that PageTable(&folio->page) can now be written as folio_test_table(folio)
> instead.
> 
> Also there are numerous uses of PageBuddy in mm/compaction.c that will
> likely need to be converted to folios as well.

... and you can currently call PageTable() on the second/third/... page
of an allocation and it will return false, regardless of what the
first page is typed as.  For most architectures, this doesn't matter,
but /proc/kpageflags will underreport the amount of memory allocated
as page tables on architectures which use multi-page allocations for
their page tables as there's currently absolutely nothing to indicate
the size of the allocation.

To fix this, we need to use __GFP_COMP.