[PATCH v8 0/6] elf: THP-aware load segment alignment

[WANG Rui]

On Fri, Apr 10, 2026 at 11:35 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Fri, Apr 10, 2026 at 10:57 AM WANG Rui <wangrui@loongson.cn> wrote:
> >
> > On Fri, Apr 10, 2026 at 9:26 AM H.J. Lu <hjl.tools@gmail.com> wrote:
> > >
> > > On Fri, Apr 10, 2026 at 7:33 AM Wilco Dijkstra <Wilco.Dijkstra@arm.com> wrote:
> > > >
> > > > Hi HJ,
> > > >
> > > > > Why a tunable at all?  Users may not set it.
> > > >
> > > > If it is already the default, users never need to set it. It makes benchmarking
> > > > simpler at a cost that is ~10000x lower than checking the huge page size.
> > > >
> > >
> > > Even if THP isn't enabled by default, like for 32-bit mode.  We may still
> > > want to enable THP on specific 32-bit applications.   A property bit provides
> > > such flexibility without requiring the user to set a tunable at run-time.
> >
> > Do we really need a flag in the executable file to hint the loader to
> > cooperate with THP? If we already have the ability to modify the
> > executable, why not just make the LOAD segmengts hugepage-aligned in
> > the first place?
> >
> > As you can see, this optimization is currently opt-in on most arches
> > rather than enabled by default. I see the tunable as more of a
> > transitional mechanism, it simply gives users a way to turn the opt on
> > and on arches where it is enbled by default, it can also be used to
> > turn it off.
> >
>
> This all or nothing approach is my main concern.  A programmer
> may want to enable THP on an application even if THP isn't enabled
> by default.    The overhead of checking a bit in the l_1_needed field
> is minimum.

In earlier discussions, there was mention of memory pressure caused by
THP, which is really a runtime system concern. When THP is not enabled
by default on the system, I'm a bit unsure whether it's a good idea
for a programer's decision to override the system default.

Ideally, the executable itself would take care of hugepage-friendly
alignment, the loader would make sure the mappings land on
hugepage-aligned virtual addresses, and the final decision on whether
to use huge pages would be left solely to THP. It has visibility into
the system's dynamic state, and users would only need to deal with
this single control point, making it relatively straightforward to
turn it on or off.

Thanks,
Rui