Re: [RFC PATCH 09/12] mm/huge_memory: refactor change_huge_pmd() non-present logic

[Lorenzo Stoakes]

On Fri, Oct 24, 2025 at 02:41:02PM -0400, Gregory Price wrote:
> On Fri, Oct 24, 2025 at 08:41:25AM +0100, Lorenzo Stoakes wrote:
> > Similar to copy_huge_pmd(), there is a large mass of open-coded logic for
> > the CONFIG_ARCH_ENABLE_THP_MIGRATION non-present entry case that does not
> > use thp_migration_supported() consistently.
> >
> > Resolve this by separating out this logic and introduce
> > change_non_present_huge_pmd().
> >
> > Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx>
> > ---
> --- >8
>
> > +	if (thp_migration_supported() && is_swap_pmd(*pmd)) {
> > +		change_non_present_huge_pmd(mm, addr, pmd, uffd_wp,
> > +					    uffd_wp_resolve);
>
> You point out the original code doesn't have thp_migration_supported()
>
> is this a bug? or is it benign and just leads to it failing (nicely)
> deeper in the stack?
>
> If it's a bug, maybe this patch should be pulled out ahead of the rest?

No it's not a bug, I mean it does:

#ifdef CONFIG_ARCH_ENBLE_THP_MIGRATION
if (is_swap_pmd(*pmd)) {
	...
}
#endif

Instead of the much nicer:

if (thp_migration_supported() && is_swap_pmd(*pmd)) {
}

Given:

static inline bool thp_migration_supported(void)
{
	return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
}

You can see it's equivalent except we rely on compiler removing dead code when
we use thp_migration_supported() obviously (which is fine)

>
> ~Gregory

Cheers, Lorenzo