Re: [PATCH 41/75] hexagon: Add pmd_pfn()

[Mike Rapoport]

On Sun, Feb 06, 2022 at 10:05:35PM +0000, Matthew Wilcox wrote:
> On Sun, Feb 06, 2022 at 11:33:40PM +0200, Mike Rapoport wrote:
> > On Sun, Feb 06, 2022 at 08:46:45PM +0000, Matthew Wilcox wrote:
> > > On Sun, Feb 06, 2022 at 08:13:11PM +0200, Mike Rapoport wrote:
> > > > On Fri, Feb 04, 2022 at 07:58:18PM +0000, Matthew Wilcox (Oracle) wrote:
> > > > > I need to use this function in common code, so define it for hexagon.
> > > > > 
> > > > > Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > > > > ---
> > > > >  arch/hexagon/include/asm/pgtable.h | 3 ++-
> > > > 
> > > > Why hexagon out of all architectures?
> > > > What about m68k, nios2, nds32 etc?
> > 
> > > Presumably they don't support CONFIG_TRANSPARENT_HUGEPAGE.
> > > This code isn't compiled when THP are disabled; at least I haven't
> > > had a buildbot complaint for any other architectures.
> > 
> > m68k defconfig fails:
> > 
> >   CC      mm/page_vma_mapped.o
> > mm/page_vma_mapped.c: In function 'page_vma_mapped_walk':
> > mm/page_vma_mapped.c:219:20: error: implicit declaration of function 'pmd_pfn'; did you mean 'pmd_off'? [-Werror=implicit-function-declaration]
> >   219 |     if (!check_pmd(pmd_pfn(pmde), pvmw))
> >       |                    ^~~~~~~
> >       |                    pmd_off
> 
> Ok, so this whole thing is protected by "if (pmd_trans_huge(pmd))"
> which is defined to be 0 if !THP.  But obviously, gcc is trying to parse
> it.  So I think we need something like:
> 
> #ifndef CONFIG_TRANSPARENT_HUGEPAGE
> static inline unsigned long pmd_pfn(pmd_t pmd)
> {
>         BUG();
>         return 0UL;
> }
> #endif
> 
> but that can have its own problem if architectures define pmd_pfn()
> when they don't have CONFIG_TRANSPARENT_HUGEPAGE.  Which of course, they
> can do.

I think the safest would be to define pmd_pfn where it's missing. Then
the whole pmd_pfn(), pmd_page() and pmd_page_vaddr() can be cleaned up on
top, but this won't block folio work.

I've baked a patch that defines pmd_pfn() based on pmd_page_vaddr() or pmd_page()
definition. With pmd_pfn() defined we won't need special care for THP or
!THP cases and PMD is always there even if it's called PGD sometimes.