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

[David Hildenbrand <david@redhat.com>]

On 15.05.25 20:08, Lorenzo Stoakes wrote:
> On Thu, May 15, 2025 at 06:11:55PM +0200, David Hildenbrand wrote:
>>>>> So if you're not overriding VM_NOHUGEPAGE, the whole point of this exercise
>>>>> is to override global 'never'?
>>>>>
>>>>
>>>> Again, I am not overriding never.
>>>>
>>>> hugepage_global_always and hugepage_global_enabled will evaluate to false
>>>> and you will not get a hugepage.
>>>
>>> Yeah, again ack, but I kind of hate that we set VM_HUGEPAGE everywhere even
>>> if the policy is never.
>>
>> I think it should behave just as if someone does manually an madvise(). So
>> whatever we do here during an madvise, we should try to do the same thing
>> here.
> 
> Ack I agree with this.
> 
> It actually simplifies things a LOT to view it this way - we're saying 'by
> default apply madvise(...) to new VMAs'.
> 
> Hm I wonder if we could have a more generic version of this...
> 
> Note though that we're not _quite_ doing this.
> 
> So in hugepage_madvise():
> 
> int hugepage_madvise(struct vm_area_struct *vma,
> 		     unsigned long *vm_flags, int advice)
> {
> 	...
> 
> 	switch (advice) {
> 	case MADV_HUGEPAGE:
> 		*vm_flags &= ~VM_NOHUGEPAGE;
> 		*vm_flags |= VM_HUGEPAGE;
> 
> 		...
> 
> 		break;
> 
> 		...
> 	}
> 
> 	...
> }
> 
> So here we're actually clearing VM_NOHUGEPAGE and overriding it, but in the
> proposed code we're not.

Yeah, I think I suggested that, but probably we should just do exactly 
what madvise() does.

> 
> So we're back into confusing territory again :)
> 
> I wonder if we could...
> 
> 1. Add an MADV_xxx that mimics the desired behaviour here.
> 
> 2. Add a generic 'madvise() by default' thing at a process level?
> 
> Is this crazy?

I think that's what I had in mind, just a bit twisted.

What could work is

1) prctl to set the default

2) madvise() to adjust all existing VMAs


We might have to teach 2) to ignore non-compatible VMAs / holes. Maybe 
not, worth an investigation.


-- 
Cheers,

David / dhildenb