Re: [PATCH Resend] mm: Refine __{pgd,p4d,pud,pmd,pte}_alloc_one_*() about HIGHMEM

["Vishal Moola (Oracle)" <vishal.moola@gmail.com>]

On Fri, Nov 07, 2025 at 01:50:06PM +0100, Arnd Bergmann wrote:
> On Fri, Nov 7, 2025, at 12:44, Lance Yang wrote:
> > From: Lance Yang <lance.yang@linux.dev>
> > On Fri,  7 Nov 2025 17:59:22 +0800, Huacai Chen wrote:
> >> 
> >>   */
> >>  static inline pte_t *__pte_alloc_one_kernel_noprof(struct mm_struct *mm)
> >>  {
> >> -	struct ptdesc *ptdesc = pagetable_alloc_noprof(GFP_PGTABLE_KERNEL &
> >> -			~__GFP_HIGHMEM, 0);
> >> +	struct ptdesc *ptdesc = pagetable_alloc_noprof(GFP_PGTABLE_KERNEL, 0);
> >
> > I looked into the history and it seems you are right. This defensive pattern
> > was likely introduced by Vishal Moola in commit c787ae5[1].
> 
> Right, so not even so long ago, so we need to make sure we agree
> on a direction and don't send opposite patches in the name of
> cleanups.

I took a look, this patch is the direction we want to go :).

> > After this cleanup, would it make sense to add a BUILD_BUG_ON() somewhere
> > to check that __GFP_HIGHMEM is not present in GFP_PGTABLE_KERNEL and
> > GFP_PGTABLE_USER? This would prevent any future regression ;)
> >
> > Just a thought ...

In this case, I don't think the extra check is necessary. This is a
remnant of defensively converting callers to the ptdesc api from
get_free_pages() variants (which masks off GFP_HIGHMEM internally).
I doubt we'll ever be changing those macros to support highmem.

> I think we can go either way here, but I'd tend towards not
> adding more checks but instead removing any mention of __GFP_HIGHMEM
> that we can show is either pointless or can be avoided, with
> the goal of having only a small number of actual highmem
> allocations remaining in places we do care about (normal
> page cache, zram, possibly huge pages).

+1

I'm not familiar with which architectures use highmem or not, so
theres likely more cases like this patch that are remnants of the
ptdesc conversion.

git grep "pagetable_alloc.*GFP_HIGHMEM" shows at least 5 references
inline that can definitely be removed. I'll go ahead and clean those up,
but I'll leave the rest to people more familiar with the architectures.