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

[Vishal Moola (Oracle)]

+Cc: Mike

On Fri, Nov 07, 2025 at 12:21:38PM +0100, Arnd Bergmann wrote:
> On Fri, Nov 7, 2025, at 10:59, Huacai Chen wrote:
> > __{pgd,p4d,pud,pmd,pte}_alloc_one_*() always allocate pages with GFP
> > flag GFP_PGTABLE_KERNEL/GFP_PGTABLE_USER. These two macros are defined
> > as follows:
> >
> >  #define GFP_PGTABLE_KERNEL	(GFP_KERNEL | __GFP_ZERO)
> >  #define GFP_PGTABLE_USER	(GFP_PGTABLE_KERNEL | __GFP_ACCOUNT)
> >
> > There is no __GFP_HIGHMEM in them, so we needn't to clear __GFP_HIGHMEM
> > explicitly.
> >
> > Signed-off-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
> > ---
> > Resend because the lines begin with # was eaten by git.
> 
> Thanks for your patch, this is an area I've also started
> looking at, with the intention to reduce the references
> to __GFO_HIGHMEM to the minimum we need for supporting the
> remaining platforms that need to use highmem somewhere.

Yay! Thanks for doing that, I like less highmem :)

> I'm not sure what the reason is for your patch, I assume
> this is meant purely as a cleanup, correct? Are you looking
> at a wider set of related cleanups, or did you just notice
> this one instance?
> 
> Note that for the moment, the 32-bit arm __pte_alloc_one() function
> still passes __GFP_HIGHMEM when CONFIG_HIGHPTE is set, though
> I would like to remove that code path. Unless we remove
> that at the same time, this should probably be explained in your
> patch description.

Skimming the functions, __pte_alloc_one_kernel() doesn't get passed in
a gfp, while __pte_alloc_one() does. IOW I __pte_alloc_one_kernel()
cares about architecture gfp, while the latter does care - so they are
2 very different cases.

Might be helpful to explain, although I don't think it matters much.

I've cc-ed Mike, he might have more useful opinions these functions.