Re: [PATCH 3/4] mm: khugepaged: move mm to list tail when MADV_COLD/MADV_FREE

[Vernon Yang <vernon2gm@gmail.com>]

On Tue, Dec 23, 2025 at 10:59:29AM +0100, David Hildenbrand (Red Hat) wrote:
> On 12/21/25 13:34, Vernon Yang wrote:
> > On Sun, Dec 21, 2025 at 10:24:11AM +0100, David Hildenbrand (Red Hat) wrote:
> > > On 12/21/25 05:25, Vernon Yang wrote:
> > > > On Sun, Dec 21, 2025 at 02:10:44AM +0000, Wei Yang wrote:
> > > > > On Fri, Dec 19, 2025 at 09:58:17AM +0100, David Hildenbrand (Red Hat) wrote:
> > > > > > On 12/19/25 06:29, Vernon Yang wrote:
> > > > > > > On Thu, Dec 18, 2025 at 10:31:58AM +0100, David Hildenbrand (Red Hat) wrote:
> > > > > > > > On 12/15/25 10:04, Vernon Yang wrote:
> > > > > > > > > For example, create three task: hot1 -> cold -> hot2. After all three
> > > > > > > > > task are created, each allocate memory 128MB. the hot1/hot2 task
> > > > > > > > > continuously access 128 MB memory, while the cold task only accesses
> > > > > > > > > its memory briefly andthen call madvise(MADV_COLD). However, khugepaged
> > > > > > > > > still prioritizes scanning the cold task and only scans the hot2 task
> > > > > > > > > after completing the scan of the cold task.
> > > > > > > > >
> > > > > > > > > So if the user has explicitly informed us via MADV_COLD/FREE that this
> > > > > > > > > memory is cold or will be freed, it is appropriate for khugepaged to
> > > > > > > > > scan it only at the latest possible moment, thereby avoiding unnecessary
> > > > > > > > > scan and collapse operations to reducing CPU wastage.
> > > > > > > > >
> > > > > > > > > Here are the performance test results:
> > > > > > > > > (Throughput bigger is better, other smaller is better)
> > > > > > > > >
> > > > > > > > > Testing on x86_64 machine:
> > > > > > > > >
> > > > > > > > > | task hot2           | without patch | with patch    |  delta  |
> > > > > > > > > |---------------------|---------------|---------------|---------|
> > > > > > > > > | total accesses time |  3.14 sec     |  2.92 sec     | -7.01%  |
> > > > > > > > > | cycles per access   |  4.91         |  2.07         | -57.84% |
> > > > > > > > > | Throughput          |  104.38 M/sec |  112.12 M/sec | +7.42%  |
> > > > > > > > > | dTLB-load-misses    |  288966432    |  1292908      | -99.55% |
> > > > > > > > >
> > > > > > > > > Testing on qemu-system-x86_64 -enable-kvm:
> > > > > > > > >
> > > > > > > > > | task hot2           | without patch | with patch    |  delta  |
> > > > > > > > > |---------------------|---------------|---------------|---------|
> > > > > > > > > | total accesses time |  3.35 sec     |  2.96 sec     | -11.64% |
> > > > > > > > > | cycles per access   |  7.23         |  2.12         | -70.68% |
> > > > > > > > > | Throughput          |  97.88 M/sec  |  110.76 M/sec | +13.16% |
> > > > > > > > > | dTLB-load-misses    |  237406497    |  3189194      | -98.66% |
> > > > > > > >
> > > > > > > > Again, I also don't like that because you make assumptions on a full process
> > > > > > > > based on some part of it's address space.
> > > > > > > >
> > > > > > > > E.g., if a library issues a MADV_COLD on some part of the memory the library
> > > > > > > > manages, why should the remaining part of the process suffer as well?
> > > > > > >
> > > > > > > Yes, you make a good point, thanks!
> > > > > > >
> > > > > > > > This seems to be an heuristic focused on some specific workloads, no?
> > > > > > >
> > > > > > > Right.
> > > > > > >
> > > > > > > Could we use the VM_NOHUGEPAGE flag to indicate that this region should
> > > > > > > not be collapsed, so that khugepaged can simply skip this VMA during
> > > > > > > scanning? This way, it won't affect the remaining part of the task's
> > > > > > > memory regions.
> > > > > >
> > > > > > I thought we would skip these regions already properly in khugeapged, or
> > > > > > maybe I misunderstood your question.
> > > > > >
> > > > >
> > > > > I think we should, but seems we didn't do this for anonymous memory during
> > > > > khugepaged.
> > > > >
> > > > > We check the vma with thp_vma_allowable_order() during scan.
> > > > >
> > > > >     * For anonymous memory during khugepaged, if we always enable 2M collapse,
> > > > >       we will scan this vma. Even VM_NOHUGEPAGE is set.
> > > > >
> > > > >     * For other cases, it looks good since __thp_vma_allowable_order() will skip
> > > > >       this vma with vma_thp_disabled().
> > > >
> > > > Hi David, Wei,
> > > >
> > > > The khugepaged has already checked the VM_NOHUGEPAGE flag for anonymous
> > > > memory during scan, as below:
> > > >
> > > > khugepaged_scan_mm_slot()
> > > >       thp_vma_allowable_order()
> > > >           thp_vma_allowable_orders()
> > > >               __thp_vma_allowable_orders()
> > > >                   vma_thp_disabled() {
> > > >                        if (vm_flags & VM_NOHUGEPAGE)
> > > >                            return true;
> > > >                   }
> > > >
> > > > REAL ISSUE: when madvise(MADV_COLD)，not set VM_NOHUGEPAGE flag to vma,
> > > > so the khugepaged will continue scan this vma.
> > > >
> > > > I set VM_NOHUGEPAGE flag to vma when madvise(MADV_COLD), the test has
> > > > been successful. I will send it in the next version.
> > >
> > > No we must not do that. That's a user-space visible change. :/
> >
> > David, what good ideas do you have to achieve this goal? let me know
> > please, thank!
>
> Your idea would be to skip a VMA when we issues madvise(MADV_COLD).
>
> That sounds like yet another heuristic that can easily be wrong? :/
>
> In particular, imagine if the VMA is much larger than the madvise'd region
> (other parts used for something else) or if the previously cold memory area
> is used for something that is now hot.
>
> With memory allocators that manage most of the memory in a single large VMA,
> it's rather easy to see how such a heuristic would be bad, no?

Thanks for your explain, but I current approach is as follows, the large
VMA will split at this case.

madvise_vma_behavior
    madvise_cold
    madvise_update_vma

Maybe I'll send v2 first, and we'll discuss it more clearly :)

--
Merry Christmas,
Vernon