Re: [PATCH v2 1/1] mm/madvise: enhance lazyfreeing with mTHP in madvise_free

From: Barry Song
Date: Fri Mar 08 2024 - 13:01:45 EST

On Fri, Mar 8, 2024 at 9:05 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>
> On 07/03/2024 18:54, Barry Song wrote:
> > On Fri, Mar 8, 2024 at 12:31 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>
> >> On 07/03/2024 12:01, Barry Song wrote:
> >>> On Thu, Mar 7, 2024 at 7:45 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >>>>
> >>>> On 07.03.24 12:42, Ryan Roberts wrote:
> >>>>> On 07/03/2024 11:31, David Hildenbrand wrote:
> >>>>>> On 07.03.24 12:26, Barry Song wrote:
> >>>>>>> On Thu, Mar 7, 2024 at 7:13 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>>>>>>>
> >>>>>>>> On 07/03/2024 10:54, David Hildenbrand wrote:
> >>>>>>>>> On 07.03.24 11:54, David Hildenbrand wrote:
> >>>>>>>>>> On 07.03.24 11:50, Ryan Roberts wrote:
> >>>>>>>>>>> On 07/03/2024 09:33, Barry Song wrote:
> >>>>>>>>>>>> On Thu, Mar 7, 2024 at 10:07 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> On 07/03/2024 08:10, Barry Song wrote:
> >>>>>>>>>>>>>> On Thu, Mar 7, 2024 at 9:00 PM Lance Yang <ioworker0@xxxxxxxxx> wrote:
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> Hey Barry,
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> Thanks for taking time to review!
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> On Thu, Mar 7, 2024 at 3:00 PM Barry Song <21cnbao@xxxxxxxxx> wrote:
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> On Thu, Mar 7, 2024 at 7:15 PM Lance Yang <ioworker0@xxxxxxxxx> wrote:
> >>>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> [...]
> >>>>>>>>>>>>>>>>> +static inline bool can_mark_large_folio_lazyfree(unsigned long addr,
> >>>>>>>>>>>>>>>>> + struct folio *folio,
> >>>>>>>>>>>>>>>>> pte_t *start_pte)
> >>>>>>>>>>>>>>>>> +{
> >>>>>>>>>>>>>>>>> + int nr_pages = folio_nr_pages(folio);
> >>>>>>>>>>>>>>>>> + fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
> >>>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>> + for (int i = 0; i < nr_pages; i++)
> >>>>>>>>>>>>>>>>> + if (page_mapcount(folio_page(folio, i)) != 1)
> >>>>>>>>>>>>>>>>> + return false;
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> we have moved to folio_estimated_sharers though it is not precise, so
> >>>>>>>>>>>>>>>> we don't do
> >>>>>>>>>>>>>>>> this check with lots of loops and depending on the subpage's mapcount.
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> If we don't check the subpage’s mapcount, and there is a cow folio
> >>>>>>>>>>>>>>> associated
> >>>>>>>>>>>>>>> with this folio and the cow folio has smaller size than this folio,
> >>>>>>>>>>>>>>> should we still
> >>>>>>>>>>>>>>> mark this folio as lazyfree?
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> I agree, this is true. However, we've somehow accepted the fact that
> >>>>>>>>>>>>>> folio_likely_mapped_shared
> >>>>>>>>>>>>>> can result in false negatives or false positives to balance the
> >>>>>>>>>>>>>> overhead. So I really don't know :-)
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>> Maybe David and Vishal can give some comments here.
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> BTW, do we need to rebase our work against David's changes[1]?
> >>>>>>>>>>>>>>>> [1]
> >>>>>>>>>>>>>>>> https://lore.kernel.org/linux-mm/20240227201548.857831-1-david@xxxxxxxxxx/
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> Yes, we should rebase our work against David’s changes.
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>> + return nr_pages == folio_pte_batch(folio, addr, start_pte,
> >>>>>>>>>>>>>>>>> + ptep_get(start_pte), nr_pages,
> >>>>>>>>>>>>>>>>> flags, NULL);
> >>>>>>>>>>>>>>>>> +}
> >>>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>> static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
> >>>>>>>>>>>>>>>>> unsigned long end, struct mm_walk
> >>>>>>>>>>>>>>>>> *walk)
> >>>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>> @@ -676,11 +690,45 @@ static int madvise_free_pte_range(pmd_t *pmd,
> >>>>>>>>>>>>>>>>> unsigned long addr,
> >>>>>>>>>>>>>>>>> */
> >>>>>>>>>>>>>>>>> if (folio_test_large(folio)) {
> >>>>>>>>>>>>>>>>> int err;
> >>>>>>>>>>>>>>>>> + unsigned long next_addr, align;
> >>>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>> - if (folio_estimated_sharers(folio) != 1)
> >>>>>>>>>>>>>>>>> - break;
> >>>>>>>>>>>>>>>>> - if (!folio_trylock(folio))
> >>>>>>>>>>>>>>>>> - break;
> >>>>>>>>>>>>>>>>> + if (folio_estimated_sharers(folio) != 1 ||
> >>>>>>>>>>>>>>>>> + !folio_trylock(folio))
> >>>>>>>>>>>>>>>>> + goto skip_large_folio;
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> I don't think we can skip all the PTEs for nr_pages, as some of them
> >>>>>>>>>>>>>>>> might be
> >>>>>>>>>>>>>>>> pointing to other folios.
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> for example, for a large folio with 16PTEs, you do MADV_DONTNEED(15-16),
> >>>>>>>>>>>>>>>> and write the memory of PTE15 and PTE16, you get page faults, thus PTE15
> >>>>>>>>>>>>>>>> and PTE16 will point to two different small folios. We can only skip
> >>>>>>>>>>>>>>>> when we
> >>>>>>>>>>>>>>>> are sure nr_pages == folio_pte_batch() is sure.
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>> Agreed. Thanks for pointing that out.
> >>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>> + align = folio_nr_pages(folio) * PAGE_SIZE;
> >>>>>>>>>>>>>>>>> + next_addr = ALIGN_DOWN(addr + align, align);
> >>>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>> + /*
> >>>>>>>>>>>>>>>>> + * If we mark only the subpages as lazyfree, or
> >>>>>>>>>>>>>>>>> + * cannot mark the entire large folio as
> >>>>>>>>>>>>>>>>> lazyfree,
> >>>>>>>>>>>>>>>>> + * then just split it.
> >>>>>>>>>>>>>>>>> + */
> >>>>>>>>>>>>>>>>> + if (next_addr > end || next_addr - addr !=
> >>>>>>>>>>>>>>>>> align ||
> >>>>>>>>>>>>>>>>> + !can_mark_large_folio_lazyfree(addr, folio,
> >>>>>>>>>>>>>>>>> pte))
> >>>>>>>>>>>>>>>>> + goto split_large_folio;
> >>>>>>>>>>>>>>>>> +
> >>>>>>>>>>>>>>>>> + /*
> >>>>>>>>>>>>>>>>> + * Avoid unnecessary folio splitting if the
> >>>>>>>>>>>>>>>>> large
> >>>>>>>>>>>>>>>>> + * folio is entirely within the given range.
> >>>>>>>>>>>>>>>>> + */
> >>>>>>>>>>>>>>>>> + folio_clear_dirty(folio);
> >>>>>>>>>>>>>>>>> + folio_unlock(folio);
> >>>>>>>>>>>>>>>>> + for (; addr != next_addr; pte++, addr +=
> >>>>>>>>>>>>>>>>> PAGE_SIZE) {
> >>>>>>>>>>>>>>>>> + ptent = ptep_get(pte);
> >>>>>>>>>>>>>>>>> + if (pte_young(ptent) ||
> >>>>>>>>>>>>>>>>> pte_dirty(ptent)) {
> >>>>>>>>>>>>>>>>> + ptent =
> >>>>>>>>>>>>>>>>> ptep_get_and_clear_full(
> >>>>>>>>>>>>>>>>> + mm, addr, pte,
> >>>>>>>>>>>>>>>>> tlb->fullmm);
> >>>>>>>>>>>>>>>>> + ptent = pte_mkold(ptent);
> >>>>>>>>>>>>>>>>> + ptent = pte_mkclean(ptent);
> >>>>>>>>>>>>>>>>> + set_pte_at(mm, addr, pte,
> >>>>>>>>>>>>>>>>> ptent);
> >>>>>>>>>>>>>>>>> + tlb_remove_tlb_entry(tlb, pte,
> >>>>>>>>>>>>>>>>> addr);
> >>>>>>>>>>>>>>>>> + }
> >>>>>>>>>>>>>>>>
> >>>>>>>>>>>>>>>> Can we do this in batches? for a CONT-PTE mapped large folio, you are
> >>>>>>>>>>>>>>>> unfolding
> >>>>>>>>>>>>>>>> and folding again. It seems quite expensive.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> I'm not convinced we should be doing this in batches. We want the initial
> >>>>>>>>>>>>> folio_pte_batch() to be as loose as possible regarding permissions so
> >>>>>>>>>>>>> that we
> >>>>>>>>>>>>> reduce our chances of splitting folios to the min. (e.g. ignore SW bits
> >>>>>>>>>>>>> like
> >>>>>>>>>>>>> soft dirty, etc). I think it might be possible that some PTEs are RO and
> >>>>>>>>>>>>> other
> >>>>>>>>>>>>> RW too (e.g. due to cow - although with the current cow impl, probably not.
> >>>>>>>>>>>>> But
> >>>>>>>>>>>>> its fragile to assume that). Anyway, if we do an initial batch that ignores
> >>>>>>>>>>>>> all
> >>>>>>>>>>>>
> >>>>>>>>>>>> You are correct. I believe this scenario could indeed occur. For instance,
> >>>>>>>>>>>> if process A forks process B and then unmaps itself, leaving B as the
> >>>>>>>>>>>> sole process owning the large folio. The current wp_page_reuse() function
> >>>>>>>>>>>> will reuse PTE one by one while the specific subpage is written.
> >>>>>>>>>>>
> >>>>>>>>>>> Hmm - I thought it would only reuse if the total mapcount for the folio
> >>>>>>>>>>> was 1.
> >>>>>>>>>>> And since it is a large folio with each page mapped once in proc B, I thought
> >>>>>>>>>>> every subpage write would cause a copy except the last one? I haven't
> >>>>>>>>>>> looked at
> >>>>>>>>>>> the code for a while. But I had it in my head that this is an area we need to
> >>>>>>>>>>> improve for mTHP.
> >>>>>>>
> >>>>>>> So sad I am wrong again 😢
> >>>>>>>
> >>>>>>>>>>
> >>>>>>>>>> wp_page_reuse() will currently reuse a PTE part of a large folio only if
> >>>>>>>>>> a single PTE remains mapped (refcount == 0).
> >>>>>>>>>
> >>>>>>>>> ^ == 1
> >>>>>>>
> >>>>>>> seems this needs improvement. it is a waste the last subpage can
> >>>>>>
> >>>>>> My take that is WIP:
> >>>>>>
> >>>>>> https://lore.kernel.org/all/20231124132626.235350-1-david@xxxxxxxxxx/T/#u
> >>>>>>
> >>>>>>> reuse the whole large folio. i was doing it in a quite different way,
> >>>>>>> if the large folio had only one subpage left, i would do copy and
> >>>>>>> released the large folio[1]. and if i could reuse the whole large folio
> >>>>>>> with CONT-PTE, i would reuse the whole large folio[2]. in mainline,
> >>>>>>> we don't have this cont-pte luxury exposed to mm, so i guess we can
> >>>>>>> not do [2] easily, but [1] seems to be an optimization.
> >>>>>>
> >>>>>> Yeah, I had essentially the same idea: just free up the large folio if most of
> >>>>>> the stuff is unmapped. But that's rather a corner-case optimization, so I did
> >>>>>> not proceed with that.
> >>>>>>
> >>>>>
> >>>>> I'm not sure it's a corner case, really? - process forks, then both parent and
> >>>>> child and write to all pages in what was previously a fully & contiguously
> >>>>> mapped large folio?
> >>>>
> >>>> Well, with 2 MiB my assumption was that while it can happen, it's rather
> >>>> rare. With smaller THP it might get more likely, agreed.
> >>>>
> >>>>>
> >>>>> Reggardless, why is it an optimization to do the copy for the last subpage and
> >>>>> syncrhonously free the large folio? It's already partially mapped so is on the
> >>>>> deferred split list and can be split if memory is tight.
> >>>
> >>> we don't want reclamation overhead later. and we want memories immediately
> >>> available to others.
> >>
> >> But by that logic, you also don't want to leave the large folio partially mapped
> >> all the way until the last subpage is CoWed. Surely you would want to reclaim it
> >> when you reach partial map status?
> >
> > To some extent, I agree. But then we will have two many copies. The last
> > subpage is small, and a safe place to copy instead.
> >
> > We actually had to tune userspace to decrease partial map as too much
> > partial map both unfolded CONT-PTE and wasted too much memory. if a
> > vma had too much partial map, we disabled mTHP on this VMA.
>
> I actually had a whacky idea around introducing selectable page size ABI
> per-process that might help here. I know Android is doing work to make the
> system 16K page compatible. You could run most of the system processes with 16K
> ABI on top of 4K kernel. Then those processes don't even have the ability to
> madvise/munmap/mprotect/mremap anything less than 16K alignment so that acts as
> an anti-fragmentation mechanism while allowing non-16K capable processes to run
> side-by-side. Just a passing thought...

Right, this project faces a challenge in supporting legacy
4KiB-aligned applications.
but I don't find it will be an issue to run 16KiB-aligned applications
on a kernel whose
page size is 4KiB.

>
> >
> >>
> >>> reclamation will always cause latency and affect User
> >>> experience. split_folio is not cheap :-)
> >>
> >> But neither is memcpy(4K) I'd imagine. But I get your point.
> >
> > In a real product scenario, we need to consider the success rate of
> > allocating large folios.
> > Currently, it's only 7%, as reported here[1], with no method to keep
> > large folios intact in a
> > buddy system.
>
> Yes I saw that email - interesting. Is that 7% number for the Oppo
> implementation or upstream implementation? (I think Oppo?). Do you know how the
> other one compares (my guess is that upstream isn't complete enough yet to give
> viable numbers)? And I'm guessing you are running on a kernel/fs that doesn't
> support large folios in the page cache? What about large folio swap? My feeling
> is that once we have all these features, that number should significantly
> increase because you can create a system that essentially uses large quantities
> of a couple of sizes of page (e.g. 4K & (16K | 64K)) and fragmentation will be
> less of a problem. Perhaps that's wishful thinking though.

This is the number of OPPO's implementations which supports one kind of
large folio size - 64KiB only. Meanwhile, OPPO has a TAO-like
optimization by extending
migrate_type and marking some pageblocks dedicated for large folios(except some
corner cases , 3-order can also use them), it brings success rate to
around 50% in
do_anon_page and more than 90% in do_swap_page(we give this lower water as
we save large objects in zsmalloc/zram - compressing and decompressing 64KiB
as a whole instead of doing 16 * 4KiB).

The reported data is disabling the TAO-like optimization and just using buddy.

BTW, based on the previous observation, 16KiB allocation could still
be a problem
on phones, for example, kernel stacks allocation was a pain before it
was changed to
vmalloc.

>
> >
> > Yu's TAO[2] chose to release the large folio entirely after copying
> > the mapped parts
> > onto smaller folios in vmscan,
>
> Yes, TAO looks very interesting! It essentially partitions the memory IIUC?

kind of, adding two virtual zones to decrease compaction and keep
large folios intact/not being splitted.

>
> >
> > [1] https://lore.kernel.org/linux-mm/20240305083743.24950-1-21cnbao@xxxxxxxxx/
> > [2] https://lore.kernel.org/linux-mm/20240229183436.4110845-1-yuzhao@xxxxxxxxxx/
> >
> >>
> >>> if the number of this kind of
> >>> large folios
> >>> is huge, the waste can be huge for some while.
> >>>
> >>> it is not a corner case for large folio swap-in. while someone writes
> >>> one subpage, I swap-in a large folio, wp_reuse will immediately
> >>> be called. This can cause waste quite often. One outcome of this
> >>> discussion is that I realize I should investigate this issue immediately
> >>> in the swap-in series as my off-tree code has optimized reuse but
> >>> mainline hasn't.
> >>>
> >>>>
> >>>> At least for 2 MiB THP, it might make sense to make that large folio
> >>>> available immediately again, even without memory pressure. Even
> >>>> compaction would not compact it.
> >>>
> >>> It is also true for 64KiB. as we want other processes to allocate
> >>> 64KiB successfully as much as possible, and reduce the rate of
> >>> falling back to small folios. by releasing 64KiB directly to buddy
> >>> rather than splitting and returning 15*4KiB in shrinker, we reduce
> >>> buddy fragmentation too.
> >>>
> >>>>
> >>>> --
> >>>> Cheers,
> >>>>
> >>>> David / dhildenb
> >
Thanks
Barry
>