Re: [PATCH 1/6] prctl: introduce PR_THP_POLICY_DEFAULT_HUGE for the process

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Thu, May 15, 2025 at 09:17:28PM +0200, David Hildenbrand wrote:
> On 15.05.25 20:36, Lorenzo Stoakes wrote:

[snip]

> > But I don't think we should use vma_set_thp_policy() in these places, we
> > should just set the flag, to avoid trying to do a write lock etc. etc.,
> > plus we want to set the flag in a place that's not a VMA yet in both cases.
> >
> > So we'd need something like in do_mmap():
> >
> > +	vm_flags |= mm_implied_vma_flags(mm);
> > 	addr = mmap_region(file, addr, len, vm_flags, pgoff, uf);
> >
> > Where mm_implied_vma_flags() reads the MMF flags and sees if any imply VMA
> > flags.
> >
> > But we have something for that already don't we? mm->def_flags.
> >
> > Can't we use that actually? That should work for mmap too?
>
> Yeah, look at
>
> commit 1860033237d4be09c5d7382585f0c7229367a534
> Author: Michal Hocko <mhocko@suse.com>
> Date:   Mon Jul 10 15:48:02 2017 -0700
>
>     mm: make PR_SET_THP_DISABLE immediately active
>
>
> Where we moved away from that. As raised, I am not sure I like what we did
> with PR_SET_THP_DISABLE.
>
> And I don't want any new magical prctl like that that add new magical
> internal toggles.
>
> OTOH, I am completely fine with a prctl that just changes the default for
> new VMAs (just like applying madvise imemdiately afterwards). I'm also fine
> with a prtctl that changes all existing VMAs, but maybe just issuing a
> madvise() is the better solution, to cleanly separate it.
>
> All not too crazy and not too invasive -- piggybagging on VM_HUGEPAGE /
> VM_NOHUGEPAGE.
>
> I can life with that if it solves a use case.

I guess you're not suggesting using an MMF_ in this way which overrides
VMAs, I think the main reason Michael wanted to not use mm->def_flags here
is because doing so doesn't immediately change existing VMAs.

But we're doing things at the VMA level anyway, so we could just set:

1. mm->def_flags accordingly (no new MMF flag needed!)
2. update existing VMAs using (possibly improved) madvise() interface

And all should work.

The get policy stuff could then just check mm->def_flags & VM_HUGEPAGE or
VM_NOHUGEPAGE and use this as state.

This might be the least egregious way of doing this...

Maybe then I could hold my nose and possibly live with truly the most Evil
Interface in the History of Computing (TM), prctl() ;)

>
> --
> Cheers,
>
> David / dhildenb
>

Cheers, Lorenzo