Re: [PATCH v2 1/1] mm: pgtable: fix pte_swp_exclusive

[Al Viro <viro@zeniv.linux.org.uk>]

On Sat, Apr 05, 2025 at 06:09:11PM +0100, Sam James wrote:
> Sam James <sam@gentoo.org> writes:
> 
> > Lovely cleanup and a great suggestion from Al.
> >
> > Reviewed-by: Sam James <sam@gentoo.org>
> >
> > I'd suggest adding a:
> > Suggested-by: Al Viro <viro@zeniv.linux.org.uk>
> 
> Al, were you planning on taking this through your tree?

FWIW, I expected it to get sent to Linus as "please, run this
sed script before -rc1" kind of thing, script being something
like

sed -i -e 's/int pte_swp_exclusive/bool pte_swp_exclusive/' \
	`git grep -l 'int pte_swp_exclusive'`

with suggested commit message...  It's absolutely regular and
that kind of tree-wide change is easier handled that way.

	Oh, well...  To restore the context: Magnus had spotted a fun
bug on Alpha back in February - pte_swp_exclusive() there returned
pte_val(pte) & _PAGE_SWP_EXCLUSIVE as int.  The problem is that
_PAGE_SWP_EXCLUSIVE is 1UL<<39 there, with obvious results...

	I looked at the originally posted patch and suggested to
make pte_swp_exclusive() return bool instead of int.  All users
are in explicitly boolean contexts:

include/linux/swapops.h:        if (pte_swp_exclusive(pte))
mm/debug_vm_pgtable.c:  WARN_ON(pte_swp_exclusive(pte));
mm/debug_vm_pgtable.c:  WARN_ON(!pte_swp_exclusive(pte));
mm/debug_vm_pgtable.c:  WARN_ON(pte_swp_exclusive(pte));
mm/internal.h:  if (pte_swp_exclusive(pte))
mm/memory.c:            if (pte_swp_exclusive(orig_pte)) {
mm/memory.c:            exclusive = pte_swp_exclusive(vmf->orig_pte);
mm/swapfile.c:          if (pte_swp_exclusive(old_pte))
mm/userfaultfd.c:               if (!pte_swp_exclusive(orig_src_pte)) {

	Magnus posted patch of that form (see
https://lore.kernel.org/all/20250218175735.19882-2-linmag7@gmail.com/),
got no serious objections and then it went nowhere.

	Bug is real and fairly obvious, fix is entirely mechanical and
affects one line in each asm/pgtable.h out there.  Linus, could you
run that sed script just before -rc1?  Commit message from the patch refered
above looks sane:

mm: pgtable: fix pte_swp_exclusive

Make pte_swp_exclusive return bool instead of int. This will better reflect
how pte_swp_exclusive is actually used in the code. This fixes swap/swapoff
problems on Alpha due pte_swp_exclusive not returning correct values when
_PAGE_SWP_EXCLUSIVE bit resides in upper 32-bits of PTE (like on alpha).

Signed-off-by: Magnus Lindholm <linmag7@gmail.com>