Re: [RFC PATCH] mm: huge_memory: add folio_mark_accessed() when zapping file THP

[Barry Song]

On Fri, Apr 11, 2025 at 1:20 PM Baolin Wang
<baolin.wang@xxxxxxxxxxxxxxxxx> wrote:
>
>
>
> On 2025/4/11 05:56, Barry Song wrote:
> > On Fri, Apr 11, 2025 at 3:13 AM Zi Yan <ziy@xxxxxxxxxx> wrote:
> >>
> >> On 10 Apr 2025, at 6:29, Barry Song wrote:
> >>
> >>> On Thu, Apr 10, 2025 at 9:05 PM Baolin Wang
> >>> <baolin.wang@xxxxxxxxxxxxxxxxx> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 2025/4/10 16:14, Barry Song wrote:
> >>>>> On Wed, Apr 9, 2025 at 1:16 AM Baolin Wang
> >>>>> <baolin.wang@xxxxxxxxxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> When investigating performance issues during file folio unmap, I noticed some
> >>>>>> behavioral differences in handling non-PMD-sized folios and PMD-sized folios.
> >>>>>> For non-PMD-sized file folios, it will call folio_mark_accessed() to mark the
> >>>>>> folio as having seen activity, but this is not done for PMD-sized folios.
> >>>>>>
> >>>>>> This might not cause obvious issues, but a potential problem could be that,
> >>>>>> it might lead to more frequent refaults of PMD-sized file folios under memory
> >>>>>> pressure. Therefore, I am unsure whether the folio_mark_accessed() should be
> >>>>>> added for PMD-sized file folios?
> >>>>>>
> >>>>>> Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
> >>>>>> ---
> >>>>>>    mm/huge_memory.c | 4 ++++
> >>>>>>    1 file changed, 4 insertions(+)
> >>>>>>
> >>>>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >>>>>> index 6ac6d468af0d..b3ade7ac5bbf 100644
> >>>>>> --- a/mm/huge_memory.c
> >>>>>> +++ b/mm/huge_memory.c
> >>>>>> @@ -2262,6 +2262,10 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
> >>>>>>                                   zap_deposited_table(tlb->mm, pmd);
> >>>>>>                           add_mm_counter(tlb->mm, mm_counter_file(folio),
> >>>>>>                                          -HPAGE_PMD_NR);
> >>>>>> +
> >>>>>> +                       if (flush_needed && pmd_young(orig_pmd) &&
> >>>>>> +                           likely(vma_has_recency(vma)))
> >>>>>> +                               folio_mark_accessed(folio);
> >>>>>
> >>>>> Acked-by: Barry Song <baohua@xxxxxxxxxx>
> >>>>
> >>>> Thanks.
> >>>>
> >>>>> I also came across an interesting observation: on a memory-limited system,
> >>>>> demoting unmapped file folios in the LRU—specifically when their mapcount
> >>>>> drops from 1 to 0—can actually improve performance.
> >>>>
> >>>> These file folios are used only once? Can folio_set_dropbehind() be used
> >>>> to optimize it, which can avoid the LRU activity movement in
> >>>> folio_mark_accessed()?
> >>>
> >>> For instance, when a process, such as a game, just exits, it can be expected
> >>> that it won't be used again in the near future. As a result, demoting
> >>> its previously
> >>> unmapped file pages can improve performance.
> >>
> >> Is it possible to mark the dying VMAs either VM_SEQ_READ or VM_RAND_READ
> >> so that folio_mark_accessed() will be skipped? Or a new vm_flag?
> >> Will it work?
> >
> > Actually took a more aggressive approach and observed good performance
> > improvements on phones. After zap_pte_range() called remove_rmap(),
> > the following logic was added:
> >
> > if (file_folio && !folio_mapped())
> >      deactivate_file_folio();
> >
> > This helps file folios from exiting processes get reclaimed more quickly
> > during the MGLRU's min generation scan while the folios are probably
> > in max gen.
> >
> > I'm not entirely sure if this is universally applicable or worth submitting as
> > a patch.
>
> IMHO, I'm afraid this is not universally applicable. Although these file
> folios have been unmapped, it's not certain that they won't be accessed
> again. These file folios might be remapped and accessed again soon, or
> accessed through read()/write() operations using a file descriptor.

This might apply to interactive systems such as desktops and Android phones.
When an app exits, it's unlikely to be reopened very soon. For
example, Firefox’s
text and other file handles are of no use to LibreOffice. So, if we
can help reclaim
Firefox’s files promptly (rather than promoting them), we may be able to assist
LibreOffice in getting memory more efficiently.

Imagine a desktop system with limited memory that can only hold either Firefox
or LibreOffice at a time. When Firefox exits, its files are still
relatively "young"
in memory. If they’re marked as recently accessed, it becomes harder to reclaim
Firefox’s exclusive files.

Consider the current LRU list:

active -------------------------------------------------------------------->
inactive

firefox files - common .so file - firefox file - common .so file

If we demote Firefox’s files, the LRU could instead look like this:

active -------------------------------------------------------------------->
inactive

common .so files - common .so files - firefox files - firefox files

With this arrangement, when launching LibreOffice, the system can quickly
reclaim Firefox's files, rather than spending time evicting the commonly
used .so files that LibreOffice may also need.

>
> I agree with Zi's suggestion. Using some kind of madvise() hint to mark
> these file folios as those that won't be accessed after being unmapped,
> seems can work?

The issue is that userspace doesn’t know why or when it should call
madvise(). From its perspective, it’s simply the app exiting.

But I agree—there are always exceptions to the pattern I described above.
Just don't know how to tell the kernel the proper pattern.

Thanks
Barry