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 05:57:57PM +0200, David Hildenbrand wrote:
> On 15.05.25 17:45, Liam R. Howlett wrote:
> > * David Hildenbrand <david@redhat.com> [250515 10:44]:
> > > On 15.05.25 16:40, Lorenzo Stoakes wrote:
> > > > Overall I feel this series should _DEFINITELY_ be an RFC. This is pretty
> > > > outlandish stuff and needs discussion.
> > > >
> > > > You're basically making it so /sys/kernel/mm/transparent_hugepage/enabled =
> > > > never is completely ignored and overridden.
> > >
> > > I thought I made it very clear during earlier discussions that never means
> > > never.
> >
> > I also thought so, but the comments later made here [1] seem to
> > contradict that?
>
> It's ... complicated.
>
> >
> > It seems "never" means "default_no" and not actually "never"?
>
> We should consider these system toggles a single set of toggles that define a
> state, and not individual toggles that overwrite each other.
>
> If you say
> 	/sys/kernel/mm/transparent_hugepage/enabled = never
> and
> 	/sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled = always
>
> instead of the *default*
>
> 	/sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled = inherit
>
> the admin explicitly stats "I want the system behavior for 2048kB not to be configured using
> /sys/kernel/mm/transparent_hugepage/enabled". That's an admin decision, not a
> per-process overwrite or whatever.
>
>
> >
> > Maybe the global/system toggles need to affect the state of each other?
> > That is, if /sys/kernel/mm/transparent_hugepage/enabled is never and you
> > set /sys/kernel/mm/transparent_hugepage/hugepages-2048kB/enabled to
> > madvise, it should not leave /sys/kernel/mm/transparent_hugepage/enabled
> > as never.
>
> I recall we discussed that, but there was also a catch to that. :(
>
> --
> Cheers,
>
> David / dhildenb
>

Did we document all this? :)

It'd be good to be super explicit about these sorts of 'dependency chains'.