On 02.07.25 19:38, SeongJae Park wrote:

On Wed, 2 Jul 2025 15:15:01 +0100 Usama Arif [off-list ref] wrote:

[...]

quoted

In terms of the approach of doing this, IMHO, I dont think the way to do this
is controversial. After the great feedback from Lorenzo on the prctl series, the
approach would be for userpsace to make a call that just does for_each_vma of the process,
madvises the VMAs,

One dirty hack that I can think off the top of my head for doing this without
new kernel changes is, unsurprisingly, using DAMOS.  Using DAMOS, users can do
madvise(MADV_HUGEPAGE) to virtual address ranges of specific access patterns.
It is aimed to be used for hot regions, while using similar one of
MADV_NOHUGEPAGE for cold regions.  An experiment with a prototype[1] showed it
eliminates about 80% of internal fragmentation caused memory overhead while
keeping 46% of performance improvement under a constrained situation.

If you set the access pattern as any pattern, hence, you can do
madvise(MADV_HUGEPAGE) for effectively entire virtual address space of the
process.  DAMON user-space tool supports periodically tracking childs and
applying same DAMOS scheme to those.  So, for example, below hack could be
tried.

     # damo start $(pidof XXX) --damos_action hugepage --include_child_tasks

IIRC, setting MADV_HUGEPAGE on arbitrary VMAs from arbitrary processes 
has the potential of breaking applications.

Just imagine them deliberately setting MADV_NOHUGEPAGE and are intending 
of using userfaultfd, where it is crucial that we don't over-allocate 
memory even before userfaultdf is actually registered.

(QEMU does that)

-- 
Cheers,

David / dhildenb