[PATCH v1 0/2] mm: split underused anonymous mTHP folios

[PATCH v1 0/2] mm: split underused anonymous mTHP folios

[Joanne Koong]

PMD-sized THPs that are mostly zero-filled are reclaimed under memory pressure
by the deferred split shrinker, but this is not done for mTHP folios. At Meta,
we would like to deploy 2M THP=always on ARM (with 64k base pages), as 2M
provides contpte benefits and 512M is too big. However, 2M THP=always will
result in a memory regression unless the extra unused portions of THPs can be
broken down and reclaimed.

This series has two patches:
Patch 1 - makes thp_underused() work for non-PMD-sized folios by scaling
khugepaged_max_ptes_none proportionally to the folio size and using that
as the threshold for how many zero-filled pages a folio can contain before
it's considered underused. This is compatible with the mTHP collapse threshold
at both of the values mTHP collapse supports (0 and
KHUGEPAGED_MAX_PTES_LIMIT). If/when collapse supports intermediate values in
the future, the thresholds can be unified. There is no functional change for
PMD-sized folios.

Patch 2 - adds anonymous mTHP folios to the deferred split list from the
anonymous fault / collapse path so the shrinker can find and split them. This
is the same mechanism PMD THPs use, extended to mTHP. It places more folios
on the shrinker's list, but a folio that the shrinker determines is not
underused gets removed from the list and doesn't get rescanned in the future
unless the folio later becomes partially mapped. The logic to determine
whether a folio is underused or not (thp_underused()) is bounded work that
does a memcmp on at most nr_pages subpages, and runs only if shrink_underused
is on.

Testing:
This was sanity-checked by running a VM with 64K mTHP enabled (PMD THP
disabled, shrink_underused = 1, khugepaged/max_ptes_none=0) and having a
process fault in 256MB of mapped mostly zero-filled 64K mTHP inside a cgroup
with swap disabled, and then triggering reclaim (via memory.reclaim).
  - /sys/kernel/mm/transparent_hugepage/hugepages-64kB/stats/anon_fault_alloc
    showed that the mTHP were allocated during fault handling
  - /sys/kernel/mm/transparent_hugepage/hugepages-64kB/stats/split and
    /proc/vmstat's thp_underused_split_page each increased by 4096
  - the process's resident set size stat (/proc/<pid>/status) showed a drop
    from 257 MB to 17 MB

The scanning in patch 2 was checked on the same setup with 256 MB of fully
populated 64k mTHP folios. All 4096 folios were processed by the deferred
split shrinker only once and did not reappear on the list in future scans.

Joanne Koong (2):
  mm/huge_memory: extend thp_underused() to mTHP folios
  mm/memory: add anonymous mTHP folios to deferred split list

 mm/huge_memory.c | 12 ++++++++----
 mm/memory.c      |  2 ++
 2 files changed, 10 insertions(+), 4 deletions(-)

-- 
2.52.0

[PATCH v1 1/2] mm/huge_memory: extend thp_underused() to mTHP folios

[Joanne Koong]

thp_underused() decides whether a large folio on the deferred split
list is underused and should be split so that its zero subpages can be
reclaimed. However, the logic in it only accounts for PMD-sized folios.

As preparatory work for splitting underused mTHP folios, make the logic
in thp_underused() compatible with mTHP-sized folios. This uses the
existing khugepaged_max_ptes_none sysctl value and scales it
proportionally to the size of the folio as the threshold for how many
zero-filled pages a folio may contain before it's considered underused.

This introduces no functional changes for PMD-size folios.

Suggested-by: Usama Arif <usama.arif@linux.dev>
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
---
 mm/huge_memory.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index c0892cc533a9..ea5c350b4818 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -4413,24 +4413,28 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
 static bool thp_underused(struct folio *folio)
 {
 	int num_zero_pages = 0, num_filled_pages = 0;
+	int nr_pages = folio_nr_pages(folio);
+	unsigned int max_ptes_none;
 	int i;
 
-	if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1)
+	max_ptes_none = khugepaged_max_ptes_none * nr_pages / HPAGE_PMD_NR;
+
+	if (max_ptes_none >= nr_pages - 1)
 		return false;
 
 	if (folio_contain_hwpoisoned_page(folio))
 		return false;
 
-	for (i = 0; i < folio_nr_pages(folio); i++) {
+	for (i = 0; i < nr_pages; i++) {
 		if (pages_identical(folio_page(folio, i), ZERO_PAGE(0))) {
-			if (++num_zero_pages > khugepaged_max_ptes_none)
+			if (++num_zero_pages > max_ptes_none)
 				return true;
 		} else {
 			/*
 			 * Another path for early exit once the number
 			 * of non-zero filled pages exceeds threshold.
 			 */
-			if (++num_filled_pages >= HPAGE_PMD_NR - khugepaged_max_ptes_none)
+			if (++num_filled_pages >= nr_pages - max_ptes_none)
 				return false;
 		}
 	}
-- 
2.52.0

[PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[Joanne Koong]

Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
to the deferred split list at fault or collapse time. As a result, a
fully mapped mTHP folio that is mostly zero-filled doesn't get split by
the deferred split shrinker when the system is under memory pressure.

Add anonymous mTHP folios to the deferred split list so that if there's
memory pressure, a zero-filled mTHP can be split with its zero pages
remapped to the shared zero page and then reclaimed.

To minimize overhead on the common order-0 fault path, the
deferred_split_folio() call is guarded by an inline folio_test_large()
check.

Suggested-by: Usama Arif <usama.arif@linux.dev>
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
---
 mm/memory.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mm/memory.c b/mm/memory.c
index 6637c5b13c9b..441d918e3dc0 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
 	folio_add_lru_vma(folio, vma);
 	set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
 	update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
+	if (folio_test_large(folio))
+		deferred_split_folio(folio, false);
 }
 
 static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
-- 
2.52.0

Re: [PATCH v1 1/2] mm/huge_memory: extend thp_underused() to mTHP folios

[Joanne Koong]

On Tue, Jul 7, 2026 at 1:18 PM Joanne Koong <joannelkoong@gmail.com> wrote:
>
> thp_underused() decides whether a large folio on the deferred split
> list is underused and should be split so that its zero subpages can be
> reclaimed. However, the logic in it only accounts for PMD-sized folios.
>
> As preparatory work for splitting underused mTHP folios, make the logic
> in thp_underused() compatible with mTHP-sized folios. This uses the
> existing khugepaged_max_ptes_none sysctl value and scales it
> proportionally to the size of the folio as the threshold for how many
> zero-filled pages a folio may contain before it's considered underused.
>
> This introduces no functional changes for PMD-size folios.
>
> Suggested-by: Usama Arif <usama.arif@linux.dev>
> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
> ---
>  mm/huge_memory.c | 12 ++++++++----
>  1 file changed, 8 insertions(+), 4 deletions(-)
>
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index c0892cc533a9..ea5c350b4818 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -4413,24 +4413,28 @@ static unsigned long deferred_split_count(struct shrinker *shrink,
>  static bool thp_underused(struct folio *folio)
>  {
>         int num_zero_pages = 0, num_filled_pages = 0;
> +       int nr_pages = folio_nr_pages(folio);
> +       unsigned int max_ptes_none;
>         int i;
>
> -       if (khugepaged_max_ptes_none == HPAGE_PMD_NR - 1)
> +       max_ptes_none = khugepaged_max_ptes_none * nr_pages / HPAGE_PMD_NR;
> +
> +       if (max_ptes_none >= nr_pages - 1)
>                 return false;

Andrew kindly directed me to a comment by Sashiko [1]:

"Does this scale calculation unintentionally disable underused splitting for
small mTHPs with valid sysctl settings?
Due to integer division truncation, max_ptes_none >= nr_pages - 1 can match
configurations that aren't the sysctl sentinel (HPAGE_PMD_NR - 1).
...
Should the sentinel check remain against the raw
khugepaged_max_ptes_none value before scaling?"

Sashiko is right. I'll fix this up in v2 and keep the original
khugepaged_max_ptes_none == HPAGE_PMD_NR - 1 check on the raw
non-scaled value.

Thanks,
Joanne

[1] https://sashiko.dev/#/patchset/20260707201735.4113107-1-joannelkoong@gmail.com

Re: [PATCH v1 0/2] mm: split underused anonymous mTHP folios

[David Hildenbrand (Arm)]

On 7/7/26 22:17, Joanne Koong wrote:
> PMD-sized THPs that are mostly zero-filled are reclaimed under memory pressure
> by the deferred split shrinker, but this is not done for mTHP folios. At Meta,
> we would like to deploy 2M THP=always on ARM (with 64k base pages), as 2M
> provides contpte benefits and 512M is too big. However, 2M THP=always will
> result in a memory regression unless the extra unused portions of THPs can be
> broken down and reclaimed.
> 
> This series has two patches:
> Patch 1 - makes thp_underused() work for non-PMD-sized folios by scaling
> khugepaged_max_ptes_none proportionally to the folio size and using that
> as the threshold for how many zero-filled pages a folio can contain before
> it's considered underused.
Did you look at the khugepaged side, and how we handle it there?

-- 
Cheers,

David

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[David Hildenbrand (Arm)]

On 7/7/26 22:17, Joanne Koong wrote:
> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
> to the deferred split list at fault or collapse time. As a result, a
> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
> the deferred split shrinker when the system is under memory pressure.
> 
> Add anonymous mTHP folios to the deferred split list so that if there's
> memory pressure, a zero-filled mTHP can be split with its zero pages
> remapped to the shared zero page and then reclaimed.
> 
> To minimize overhead on the common order-0 fault path, the
> deferred_split_folio() call is guarded by an inline folio_test_large()
> check.
> 
> Suggested-by: Usama Arif <usama.arif@linux.dev>
> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
> ---
>  mm/memory.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 6637c5b13c9b..441d918e3dc0 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
>  	folio_add_lru_vma(folio, vma);
>  	set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
>  	update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
> +	if (folio_test_large(folio))
> +		deferred_split_folio(folio, false);
>  }
>  
>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,

I had a session [1] at LSF/MM about having essentially all large anon folios
part of the the deferred split queue.

(1) I don't think this scales.

(2) I suspect the shrinker should make smarter decisions of what to scan/reclaim
first.

I think this needs more proper thought.

[1]
https://docs.google.com/presentation/d/1RfKWCY1AMVns-WLn-QdAWbI2a-rA7fbFyh7XD1Wn5BY/edit?usp=sharing

-- 
Cheers,

David

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[Usama Arif]

On 08/07/2026 08:56, David Hildenbrand (Arm) wrote:
> On 7/7/26 22:17, Joanne Koong wrote:
>> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
>> to the deferred split list at fault or collapse time. As a result, a
>> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
>> the deferred split shrinker when the system is under memory pressure.
>>
>> Add anonymous mTHP folios to the deferred split list so that if there's
>> memory pressure, a zero-filled mTHP can be split with its zero pages
>> remapped to the shared zero page and then reclaimed.
>>
>> To minimize overhead on the common order-0 fault path, the
>> deferred_split_folio() call is guarded by an inline folio_test_large()
>> check.
>>
>> Suggested-by: Usama Arif <usama.arif@linux.dev>
>> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
>> ---
>>  mm/memory.c | 2 ++
>>  1 file changed, 2 insertions(+)
>>
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 6637c5b13c9b..441d918e3dc0 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
>>  	folio_add_lru_vma(folio, vma);
>>  	set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
>>  	update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
>> +	if (folio_test_large(folio))
>> +		deferred_split_folio(folio, false);
>>  }
>>  
>>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
> 
> I had a session [1] at LSF/MM about having essentially all large anon folios
> part of the the deferred split queue.
> 
> (1) I don't think this scales.
> 

What if we only add the mTHPs for which the policy is set to always.

I think it doesn't scale if someone sets all orders to always, but if you only
set the 2M mTHP sysfs to always, it should be the same as what we have today
for PMD order?

The main motivation for this series is to try and bring the performance for
ARM on par with x86. One of the differences is TLB misses. I imagine the lower
churn in kernel by using larger page sizes will help as well. But we will
run into OOMs without the shrinker.
 

> (2) I suspect the shrinker should make smarter decisions of what to scan/reclaim
> first.

Yes definitely agree!

Alexandre is looking into this as well.

> 
> I think this needs more proper thought.
> 
> [1]
> https://docs.google.com/presentation/d/1RfKWCY1AMVns-WLn-QdAWbI2a-rA7fbFyh7XD1Wn5BY/edit?usp=sharing
> 

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[Joanne Koong]

On Wed, Jul 8, 2026 at 12:56 AM David Hildenbrand (Arm)
<david@kernel.org> wrote:
>
> On 7/7/26 22:17, Joanne Koong wrote:
> > Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
> > to the deferred split list at fault or collapse time. As a result, a
> > fully mapped mTHP folio that is mostly zero-filled doesn't get split by
> > the deferred split shrinker when the system is under memory pressure.
> >
> > Add anonymous mTHP folios to the deferred split list so that if there's
> > memory pressure, a zero-filled mTHP can be split with its zero pages
> > remapped to the shared zero page and then reclaimed.
> >
> > To minimize overhead on the common order-0 fault path, the
> > deferred_split_folio() call is guarded by an inline folio_test_large()
> > check.
> >
> > Suggested-by: Usama Arif <usama.arif@linux.dev>
> > Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
> > ---
> >  mm/memory.c | 2 ++
> >  1 file changed, 2 insertions(+)
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 6637c5b13c9b..441d918e3dc0 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
> >       folio_add_lru_vma(folio, vma);
> >       set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
> >       update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
> > +     if (folio_test_large(folio))
> > +             deferred_split_folio(folio, false);
> >  }
> >
> >  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
>
> I had a session [1] at LSF/MM about having essentially all large anon folios
> part of the the deferred split queue.
>
> (1) I don't think this scales.
>
> (2) I suspect the shrinker should make smarter decisions of what to scan/reclaim
> first.
>
> I think this needs more proper thought.
>
> [1]
> https://docs.google.com/presentation/d/1RfKWCY1AMVns-WLn-QdAWbI2a-rA7fbFyh7XD1Wn5BY/edit?usp=sharing

Thanks for the link to the slides! Was there a conclusion from the
LSF/MM discussion about the future path forward for deferred splitting
or is that still being determined?

Thanks,
Joanne

Re: [PATCH v1 0/2] mm: split underused anonymous mTHP folios

[Joanne Koong]

On Wed, Jul 8, 2026 at 12:47 AM David Hildenbrand (Arm)
<david@kernel.org> wrote:
>
> On 7/7/26 22:17, Joanne Koong wrote:
> > PMD-sized THPs that are mostly zero-filled are reclaimed under memory pressure
> > by the deferred split shrinker, but this is not done for mTHP folios. At Meta,
> > we would like to deploy 2M THP=always on ARM (with 64k base pages), as 2M
> > provides contpte benefits and 512M is too big. However, 2M THP=always will
> > result in a memory regression unless the extra unused portions of THPs can be
> > broken down and reclaimed.
> >
> > This series has two patches:
> > Patch 1 - makes thp_underused() work for non-PMD-sized folios by scaling
> > khugepaged_max_ptes_none proportionally to the folio size and using that
> > as the threshold for how many zero-filled pages a folio can contain before
> > it's considered underused.
> Did you look at the khugepaged side, and how we handle it there?

I'm not sure if I'm interpreting your suggestion correctly - is your
suggestion about using the khugepaged mTHP collapse threshold in
collapse_max_ptes_none()? I had looked at Nico's patchset in [1] and
my understanding of it was that the mTHP collapse threshold currently
only supports values of 0 and HPAGE_PMD_NR - 1. The patch 1 changes
use khugepaged_max_ptes_none's value as a proportional per-order
threshold for mTHP because 0 and HPAGE_PMD_NR - 1 are too extreme for
our use case where we're trying to find a middle ground that keeps
mostly-populated 2M folios whole but splits mostly zero-filled ones.

Is that what you had in mind or were you pointing at something else on
the khugepaged side?

Thanks,
Joanne

[1] https://lore.kernel.org/linux-mm/20260605161422.213817-1-npache@redhat.com/

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[David Hildenbrand (Arm)]

On 7/8/26 11:58, Usama Arif wrote:
> 
> 
> On 08/07/2026 08:56, David Hildenbrand (Arm) wrote:
>> On 7/7/26 22:17, Joanne Koong wrote:
>>> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
>>> to the deferred split list at fault or collapse time. As a result, a
>>> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
>>> the deferred split shrinker when the system is under memory pressure.
>>>
>>> Add anonymous mTHP folios to the deferred split list so that if there's
>>> memory pressure, a zero-filled mTHP can be split with its zero pages
>>> remapped to the shared zero page and then reclaimed.
>>>
>>> To minimize overhead on the common order-0 fault path, the
>>> deferred_split_folio() call is guarded by an inline folio_test_large()
>>> check.
>>>
>>> Suggested-by: Usama Arif <usama.arif@linux.dev>
>>> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
>>> ---
>>>  mm/memory.c | 2 ++
>>>  1 file changed, 2 insertions(+)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 6637c5b13c9b..441d918e3dc0 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
>>>  	folio_add_lru_vma(folio, vma);
>>>  	set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
>>>  	update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
>>> +	if (folio_test_large(folio))
>>> +		deferred_split_folio(folio, false);
>>>  }
>>>  
>>>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
>>
>> I had a session [1] at LSF/MM about having essentially all large anon folios
>> part of the the deferred split queue.
>>
>> (1) I don't think this scales.
>>
> 
> What if we only add the mTHPs for which the policy is set to always.

That doesn't really help lock contention. And the lock contention is indeed a
thing, as we now also want to enable the LRU cache for smaller large folios,
which improves performance:

https://lore.kernel.org/r/20260709081536.82768-1-baohua@kernel.org

> 
> I think it doesn't scale if someone sets all orders to always, but if you only
> set the 2M mTHP sysfs to always, it should be the same as what we have today
> for PMD order?

I don't see a problem with large large folios, only with small large folios :)

> 
> The main motivation for this series is to try and bring the performance for
> ARM on par with x86. One of the differences is TLB misses. I imagine the lower
> churn in kernel by using larger page sizes will help as well. But we will
> run into OOMs without the shrinker.

Can you elaborate how this change helps here? I assume you want to enable mTHP
in production. Which sizes? large large ones? :)


-- 
Cheers,

David

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[David Hildenbrand (Arm)]

On 7/8/26 19:52, Joanne Koong wrote:
> On Wed, Jul 8, 2026 at 12:56 AM David Hildenbrand (Arm)
> <david@kernel.org> wrote:
>>
>> On 7/7/26 22:17, Joanne Koong wrote:
>>> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
>>> to the deferred split list at fault or collapse time. As a result, a
>>> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
>>> the deferred split shrinker when the system is under memory pressure.
>>>
>>> Add anonymous mTHP folios to the deferred split list so that if there's
>>> memory pressure, a zero-filled mTHP can be split with its zero pages
>>> remapped to the shared zero page and then reclaimed.
>>>
>>> To minimize overhead on the common order-0 fault path, the
>>> deferred_split_folio() call is guarded by an inline folio_test_large()
>>> check.
>>>
>>> Suggested-by: Usama Arif <usama.arif@linux.dev>
>>> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
>>> ---
>>>  mm/memory.c | 2 ++
>>>  1 file changed, 2 insertions(+)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 6637c5b13c9b..441d918e3dc0 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
>>>       folio_add_lru_vma(folio, vma);
>>>       set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
>>>       update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
>>> +     if (folio_test_large(folio))
>>> +             deferred_split_folio(folio, false);
>>>  }
>>>
>>>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
>>
>> I had a session [1] at LSF/MM about having essentially all large anon folios
>> part of the the deferred split queue.
>>
>> (1) I don't think this scales.
>>
>> (2) I suspect the shrinker should make smarter decisions of what to scan/reclaim
>> first.
>>
>> I think this needs more proper thought.
>>
>> [1]
>> https://docs.google.com/presentation/d/1RfKWCY1AMVns-WLn-QdAWbI2a-rA7fbFyh7XD1Wn5BY/edit?usp=sharing
> 
> Thanks for the link to the slides! Was there a conclusion from the
> LSF/MM discussion about the future path forward for deferred splitting
> or is that still being determined?

Sorry for the late reply. Unfortunately, it wasn't clear yet if we could
reporpuse the LRU, whereby we would just naturally benefit from the LRU cache
(soon) and only manage pages on a single list.

The concern was that we might end up scanning many items on the LRU to detect
splitting candidates.

I am not 100% sure if that is a real problem.

As raised during the last THP cabal, my gut feeling is that Johannes might have
an idea on how to improve things here.

-- 
Cheers,

David

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[Usama Arif]

On 29/07/2026 13:36, David Hildenbrand (Arm) wrote:
> On 7/8/26 11:58, Usama Arif wrote:
>>
>>
>> On 08/07/2026 08:56, David Hildenbrand (Arm) wrote:
>>> On 7/7/26 22:17, Joanne Koong wrote:
>>>> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
>>>> to the deferred split list at fault or collapse time. As a result, a
>>>> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
>>>> the deferred split shrinker when the system is under memory pressure.
>>>>
>>>> Add anonymous mTHP folios to the deferred split list so that if there's
>>>> memory pressure, a zero-filled mTHP can be split with its zero pages
>>>> remapped to the shared zero page and then reclaimed.
>>>>
>>>> To minimize overhead on the common order-0 fault path, the
>>>> deferred_split_folio() call is guarded by an inline folio_test_large()
>>>> check.
>>>>
>>>> Suggested-by: Usama Arif <usama.arif@linux.dev>
>>>> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
>>>> ---
>>>>  mm/memory.c | 2 ++
>>>>  1 file changed, 2 insertions(+)
>>>>
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 6637c5b13c9b..441d918e3dc0 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
>>>>  	folio_add_lru_vma(folio, vma);
>>>>  	set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
>>>>  	update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
>>>> +	if (folio_test_large(folio))
>>>> +		deferred_split_folio(folio, false);
>>>>  }
>>>>  
>>>>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
>>>
>>> I had a session [1] at LSF/MM about having essentially all large anon folios
>>> part of the the deferred split queue.
>>>
>>> (1) I don't think this scales.
>>>
>>
>> What if we only add the mTHPs for which the policy is set to always.
> 
> That doesn't really help lock contention. And the lock contention is indeed a
> thing, as we now also want to enable the LRU cache for smaller large folios,
> which improves performance:
> 
> https://lore.kernel.org/r/20260709081536.82768-1-baohua@kernel.org
> 
>>
>> I think it doesn't scale if someone sets all orders to always, but if you only
>> set the 2M mTHP sysfs to always, it should be the same as what we have today
>> for PMD order?
> 
> I don't see a problem with large large folios, only with small large folios :)
> 
>>
>> The main motivation for this series is to try and bring the performance for
>> ARM on par with x86. One of the differences is TLB misses. I imagine the lower
>> churn in kernel by using larger page sizes will help as well. But we will
>> run into OOMs without the shrinker.
> 
> Can you elaborate how this change helps here? I assume you want to enable mTHP
> in production. Which sizes? large large ones? :)
> 

One of the differences in our production between x86 and ARM hosts which we think
should perform at a similar level (but currently aren't) is TLB misses.

On x86, we have 2M THP set to always.
On ARM, we have base page size of 64K. 2M mTHPs should help bridge the gap due to
TLB coalescing. We can't really set 2M mTHP to always as without the mTHP shrinker
we will end up with severe memory pressure and a significant increase in OOMs.

We don't plan to enable other mTHP sizes on ARM, only 2M.

The series by Barry is for folios above costly order, so won't impact this.

David, Joanne, Johannes: what do you think about only adding folios that are larger
than COSTLY_ORDER to deferred split list?

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[Usama Arif]

On 29/07/2026 14:43, Usama Arif wrote:
> 
> 
> On 29/07/2026 13:36, David Hildenbrand (Arm) wrote:
>> On 7/8/26 11:58, Usama Arif wrote:
>>>
>>>
>>> On 08/07/2026 08:56, David Hildenbrand (Arm) wrote:
>>>> On 7/7/26 22:17, Joanne Koong wrote:
>>>>> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
>>>>> to the deferred split list at fault or collapse time. As a result, a
>>>>> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
>>>>> the deferred split shrinker when the system is under memory pressure.
>>>>>
>>>>> Add anonymous mTHP folios to the deferred split list so that if there's
>>>>> memory pressure, a zero-filled mTHP can be split with its zero pages
>>>>> remapped to the shared zero page and then reclaimed.
>>>>>
>>>>> To minimize overhead on the common order-0 fault path, the
>>>>> deferred_split_folio() call is guarded by an inline folio_test_large()
>>>>> check.
>>>>>
>>>>> Suggested-by: Usama Arif <usama.arif@linux.dev>
>>>>> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
>>>>> ---
>>>>>  mm/memory.c | 2 ++
>>>>>  1 file changed, 2 insertions(+)
>>>>>
>>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>>> index 6637c5b13c9b..441d918e3dc0 100644
>>>>> --- a/mm/memory.c
>>>>> +++ b/mm/memory.c
>>>>> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
>>>>>  	folio_add_lru_vma(folio, vma);
>>>>>  	set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
>>>>>  	update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
>>>>> +	if (folio_test_large(folio))
>>>>> +		deferred_split_folio(folio, false);
>>>>>  }
>>>>>  
>>>>>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
>>>>
>>>> I had a session [1] at LSF/MM about having essentially all large anon folios
>>>> part of the the deferred split queue.
>>>>
>>>> (1) I don't think this scales.
>>>>
>>>
>>> What if we only add the mTHPs for which the policy is set to always.
>>
>> That doesn't really help lock contention. And the lock contention is indeed a
>> thing, as we now also want to enable the LRU cache for smaller large folios,
>> which improves performance:
>>
>> https://lore.kernel.org/r/20260709081536.82768-1-baohua@kernel.org
>>
>>>
>>> I think it doesn't scale if someone sets all orders to always, but if you only
>>> set the 2M mTHP sysfs to always, it should be the same as what we have today
>>> for PMD order?
>>
>> I don't see a problem with large large folios, only with small large folios :)
>>
>>>
>>> The main motivation for this series is to try and bring the performance for
>>> ARM on par with x86. One of the differences is TLB misses. I imagine the lower
>>> churn in kernel by using larger page sizes will help as well. But we will
>>> run into OOMs without the shrinker.
>>
>> Can you elaborate how this change helps here? I assume you want to enable mTHP
>> in production. Which sizes? large large ones? :)
>>
> 
> One of the differences in our production between x86 and ARM hosts which we think
> should perform at a similar level (but currently aren't) is TLB misses.
> 
> On x86, we have 2M THP set to always.
> On ARM, we have base page size of 64K. 2M mTHPs should help bridge the gap due to
> TLB coalescing. We can't really set 2M mTHP to always as without the mTHP shrinker
> we will end up with severe memory pressure and a significant increase in OOMs.
> 
> We don't plan to enable other mTHP sizes on ARM, only 2M.
> 
> The series by Barry is for folios above costly order, so won't impact this.

s/above/below/
> 
> David, Joanne, Johannes: what do you think about only adding folios that are larger
> than COSTLY_ORDER to deferred split list?

Re: [PATCH v1 2/2] mm/memory: add anonymous mTHP folios to deferred split list

[Johannes Weiner]

On Wed, Jul 29, 2026 at 02:38:53PM +0200, David Hildenbrand (Arm) wrote:
> On 7/8/26 19:52, Joanne Koong wrote:
> > On Wed, Jul 8, 2026 at 12:56 AM David Hildenbrand (Arm)
> > <david@kernel.org> wrote:
> >>
> >> On 7/7/26 22:17, Joanne Koong wrote:
> >>> Unlike for PMD-sized folios, an anonymous mTHP folio doesn't get added
> >>> to the deferred split list at fault or collapse time. As a result, a
> >>> fully mapped mTHP folio that is mostly zero-filled doesn't get split by
> >>> the deferred split shrinker when the system is under memory pressure.
> >>>
> >>> Add anonymous mTHP folios to the deferred split list so that if there's
> >>> memory pressure, a zero-filled mTHP can be split with its zero pages
> >>> remapped to the shared zero page and then reclaimed.
> >>>
> >>> To minimize overhead on the common order-0 fault path, the
> >>> deferred_split_folio() call is guarded by an inline folio_test_large()
> >>> check.
> >>>
> >>> Suggested-by: Usama Arif <usama.arif@linux.dev>
> >>> Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
> >>> ---
> >>>  mm/memory.c | 2 ++
> >>>  1 file changed, 2 insertions(+)
> >>>
> >>> diff --git a/mm/memory.c b/mm/memory.c
> >>> index 6637c5b13c9b..441d918e3dc0 100644
> >>> --- a/mm/memory.c
> >>> +++ b/mm/memory.c
> >>> @@ -5259,6 +5259,8 @@ void map_anon_folio_pte_nopf(struct folio *folio, pte_t *pte,
> >>>       folio_add_lru_vma(folio, vma);
> >>>       set_ptes(vma->vm_mm, addr, pte, entry, nr_pages);
> >>>       update_mmu_cache_range(NULL, vma, addr, pte, nr_pages);
> >>> +     if (folio_test_large(folio))
> >>> +             deferred_split_folio(folio, false);
> >>>  }
> >>>
> >>>  static void map_anon_folio_pte_pf(struct folio *folio, pte_t *pte,
> >>
> >> I had a session [1] at LSF/MM about having essentially all large anon folios
> >> part of the the deferred split queue.
> >>
> >> (1) I don't think this scales.
> >>
> >> (2) I suspect the shrinker should make smarter decisions of what to scan/reclaim
> >> first.
> >>
> >> I think this needs more proper thought.
> >>
> >> [1]
> >> https://docs.google.com/presentation/d/1RfKWCY1AMVns-WLn-QdAWbI2a-rA7fbFyh7XD1Wn5BY/edit?usp=sharing
> > 
> > Thanks for the link to the slides! Was there a conclusion from the
> > LSF/MM discussion about the future path forward for deferred splitting
> > or is that still being determined?
> 
> Sorry for the late reply. Unfortunately, it wasn't clear yet if we could
> reporpuse the LRU, whereby we would just naturally benefit from the LRU cache
> (soon) and only manage pages on a single list.
> 
> The concern was that we might end up scanning many items on the LRU to detect
> splitting candidates.
> 
> I am not 100% sure if that is a real problem.
> 
> As raised during the last THP cabal, my gut feeling is that Johannes might have
> an idea on how to improve things here.

I've been trying to reconstruct all the details we talked about at
LSFMM ;)

Going over this again, I still have to conclude using the anon LRU for
splitting is not a good idea. Let me try to lay it out:

(1) The anon LRU isn't scanned at all when there is no swap. This is
    fixable, but requires some re-architecting of the vmscan stack.

(2) With a mix of basepages and THPs, there could indeed be a lot of
    basepages ahead of underused THPs. That means swapping before
    getting to space that is much cheaper to reclaim.

    The current setup isn't perfect in that regard, as the shrinker
    runs simultaneously as the LRU. But it's making guaranteed forward
    progress through the THPs, even as the first LRU pages are scanned.

(3) The anon LRU has folio lifetime, but the splitqueue is one-shot:
    we scan each THP once, and then it's either split and dropped, or
    found full and dropped. That THP never needs to be revisited. The
    queue actually empties as the workload establishes itself.

    The anon LRU ~ splitqueue argument is only true around startup.

    If we used the anon LRU, we'd need per-page state to avoid repeat
    underused checks. And we need external state to not scan the anon
    LRU at all if there are no new THPs (and no swap). And if that's
    just a counter for "new, not yet scanned THPs", a single fault
    will cause you to walk the entire anon LRU before you get to it.

(4) The anon LRU is driven based on the cost of swap and observed
    refaults. These metrics are inherently bad modulators for scanning
    underused THP space.

    Using the anon LRU for splits means that if anon scanning slows
    down and we lean more on the file cache, we'd also slow down the
    search for unused THP space. This is undesirable. File cache is
    still more valuable than uninitialized anonymous memory:

                    | anon | file | uTHP |
    ----------------+--------------------+
    cost to reclaim |    1 |    0 |    0 |
    ----------------+--------------------+
    cost to refault |    1 |    1 |    0 |

    We're thinking anon LRU because those splittable, potentially
    underused THPs happen to be anon. But anon user data that needs to
    be swapped out is an inherently different class of reclaim targets
    than the uninitialized space *between* such anon user data.

    We really want uTHP -> clean cache -> swap reclaim ordering.

    Classic LRU takes this even further. Because of how the page cache
    grows endlessly compared to heap memory, classic will scan *only*
    the file LRU until those pages start refaulting. Using the anon
    LRU would get us a clean cache -> swap / uTHP ordering.

So I think reusing the anon LRU is flawed. It's fundamentally
different needles in fundamentally different haystacks.

If we can agree on that, then the lock contention problem has a
different scope as well: it's a simple optimization issue, not a
fundamental data structure arrangement issue.

If I understand you correctly, the concern is that people will enable
all manner of mTHP orders, and 99% of the anon faults, including all
the order-3, order-4 pagelets, will go through the list_lru lock on
fault, with no batching.

I do think that's valid, but how concrete is that right now? Is anyone
actually doing that? I would have some concerns purely from a
servability POV: the page allocator, watermarking, compaction etc. are
still a bottle neck for high rates of lower orders, as we've been
noticing with the page cache and the optimistic vmalloc higher orders.

The concrete proposal I've seen from several places was much simpler:
I have ARM 64k basepages and I want 2M THPs.

But in that case, the splitqueue looks no different than on x86 today.

My take is that we should add mTHPs to the split queue as-is. Deal
with the locking/batching concern when real usecases say we should.