[PATCH v7 0/4] mm: avoid large folio splits when swap is unavailable

[PATCH v7 0/4] mm: avoid large folio splits when swap is unavailable

[Xueyuan Chen]

This is v7 of Barry's original RFC patch, "mm: Avoiding split large
folios if swap has no space":

https://lore.kernel.org/r/20260618221720.71768-1-baohua@kernel.org

Barry's RFC showed the no-swap case with MADV_PAGEOUT on 16KB mTHP: the
large-folio split counter increased by 1024 even though no swapout
progress was possible. Skipping the split in that case kept the counter
at 0.

This series makes folio_alloc_swap() classify failures according to
whether splitting a large folio might allow swapout to make progress.
Callers can then avoid destroying the large folio when neither global
swap availability nor the folio's memcg swap hierarchy has capacity for
even a smaller folio.

Patch #1 adds page_counter_margin(), a small helper that computes the
minimum remaining chargeable space across a page_counter hierarchy.

Patch #2 establishes the folio_alloc_swap() return-value contract:

  - -E2BIG: splitting may let smaller folios make progress
  - -ENOSPC: no global swap space is available
  - -ENOMEM: splitting is not expected to help, including when the
    folio's memcg swap hierarchy has no remaining capacity

Patch #3 makes vmscan split a large folio only when folio_alloc_swap()
returns -E2BIG. Other failures keep the existing activation path and
avoid destroying the large folio when no smaller part can be backed by
swap either.

Patch #4 applies the same contract to shmem_writeout(), which currently
splits a large folio on every folio_alloc_swap() failure. It now enters
the split fallback only on -E2BIG; other failures redirty and reactivate
the folio as before.

Testing:

With a 1GB anonymous mapping backed by 16KB mTHPs and
memory.swap.max=0, the patch reduced the median latency of 30
process_madvise(MADV_PAGEOUT) runs from 743.8 ms to 181.7 ms, while the
number of large-folio splits per run dropped from 65536 to 0. Neither
kernel swapped out any pages.

I also ran DaCapo h2 under swap pressure and found no statistically
significant change in wall time or CPU time. The overall benefit appears
minor and workload-dependent.

v6 -> v7:
- Add test results and clarify that the overall benefit is minor. No code
  changes.
- https://lore.kernel.org/r/20260813075025.1406585-1-xueyuan.chen21@gmail.com

RFC v5 -> v6:
- Simplify Patch #2's failure handling following Kairui Song's review,
  avoiding additional plumbing through the memcg charge path.
- https://lore.kernel.org/r/20260730122304.2496440-1-xueyuan.chen21@gmail.com

RFC v4 -> RFC v5:
- Fix Patch #3 to jump to activate_locked, not activate_locked_split, when
  folio_alloc_swap() fails with an error other than -E2BIG. The folio has
  not been split in that case, so activate_locked_split would incorrectly
  adjust nr_scanned and nr_pages as if tail pages had been split out.
- https://lore.kernel.org/r/20260730021630.2235914-1-xueyuan.chen21@gmail.com

RFC v3 -> RFC v4:
- Keep global swap availability and memcg swap margin separate, following
  feedback from Barry Song, Johannes Weiner, and Youngjun Park.
- Handle early rejections and memcg charge failures with the refined
  folio_alloc_swap() return-value contract.
- Apply the requested vmscan and shmem condition layout changes.
- https://lore.kernel.org/r/20260717122514.51514-1-xueyuan.chen21@gmail.com

RFC v2 -> RFC v3:
- Use Johannes Weiner's original page_counter_margin() patch and preserve
  his authorship. Move the mem_cgroup_get_nr_swap_pages() conversion into
  Patch #1 so the helper addition remains a pure refactoring.
- Add Patch #4 to make shmem_writeout() split large folios only on -E2BIG,
  as suggested by Baolin Wang.
- https://lore.kernel.org/r/20260709145124.764807-1-xueyuan.chen21@gmail.com

RFC v1 -> RFC v2:
- Split the RFC into helper, swap allocation, and vmscan patches.
- Add page_counter_margin() and use it for hierarchical memcg swap
  capacity checks.
- Make folio_alloc_swap() return -E2BIG only when a smaller folio may
  still be swapped out.
- Return -ENOSPC for no global swap space and -ENOMEM when splitting is
  not expected to help, including memcg swap exhaustion.
- Make vmscan split large folios only on -E2BIG from folio_alloc_swap().

Barry Song (Xiaomi) (1):
  mm/vmscan: avoid pointless large folio splits without swap

Johannes Weiner (1):
  mm: add page_counter_margin()

Xueyuan Chen (2):
  mm: distinguish large folio swap allocation failures
  mm/shmem: split large folios only on -E2BIG

 include/linux/page_counter.h |  1 +
 include/linux/swap.h         |  6 ++++++
 mm/memcontrol.c              | 32 ++++++++++++++++++++++++++------
 mm/page_counter.c            | 20 ++++++++++++++++++++
 mm/shmem.c                   |  7 ++++---
 mm/swapfile.c                | 26 +++++++++++++++++++-------
 mm/vmscan.c                  |  7 ++++++-
 7 files changed, 82 insertions(+), 17 deletions(-)


base-commit: c73b725a57f276a3702ca213bde78fca029bc619
-- 
2.47.3

[PATCH v7 1/4] mm: add page_counter_margin()

[Xueyuan Chen]

From: Johannes Weiner <hannes@cmpxchg.org>

mem_cgroup_get_nr_swap_pages() open-codes the remaining capacity across
the memcg swap counter hierarchy.

Add page_counter_margin() to return the minimum usable space from a page
counter to the root, and use it in mem_cgroup_get_nr_swap_pages(). This
is a pure refactoring with no intended behavior change.

Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Signed-off-by: Xueyuan Chen <xueyuan.chen21@gmail.com>
Reviewed-by: David Hildenbrand (Arm) <david@kernel.org>
---
 include/linux/page_counter.h |  1 +
 mm/memcontrol.c              |  9 +++------
 mm/page_counter.c            | 20 ++++++++++++++++++++
 3 files changed, 24 insertions(+), 6 deletions(-)

diff --git a/include/linux/page_counter.h b/include/linux/page_counter.h
index d649b6bbbc87..07b7cb12249c 100644
--- a/include/linux/page_counter.h
+++ b/include/linux/page_counter.h
@@ -68,6 +68,7 @@ static inline unsigned long page_counter_read(struct page_counter *counter)
 	return atomic_long_read(&counter->usage);
 }
 
+long page_counter_margin(struct page_counter *counter);
 void page_counter_cancel(struct page_counter *counter, unsigned long nr_pages);
 void page_counter_charge(struct page_counter *counter, unsigned long nr_pages);
 bool page_counter_try_charge(struct page_counter *counter,
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 8319ad8c5c23..109c08be91cf 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -5670,12 +5670,9 @@ long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
 {
 	long nr_swap_pages = get_nr_swap_pages();
 
-	if (mem_cgroup_disabled() || do_memsw_account())
-		return nr_swap_pages;
-	for (; !mem_cgroup_is_root(memcg); memcg = parent_mem_cgroup(memcg))
-		nr_swap_pages = min_t(long, nr_swap_pages,
-				      READ_ONCE(memcg->swap.max) -
-				      page_counter_read(&memcg->swap));
+	if (!mem_cgroup_disabled() && !do_memsw_account())
+		nr_swap_pages = min(nr_swap_pages, page_counter_margin(&memcg->swap));
+
 	return nr_swap_pages;
 }
 
diff --git a/mm/page_counter.c b/mm/page_counter.c
index 661e0f2a5127..450543f4b318 100644
--- a/mm/page_counter.c
+++ b/mm/page_counter.c
@@ -46,6 +46,26 @@ static void propagate_protected_usage(struct page_counter *c,
 	}
 }
 
+/**
+ * page_counter_margin - remaining usable space within hierarchical limits
+ * @counter: counter
+ *
+ * Return: The minimum value of max minus usage across @counter and all of
+ * its ancestors. The value may be negative during a concurrent charge.
+ */
+long page_counter_margin(struct page_counter *counter)
+{
+	long margin = PAGE_COUNTER_MAX;
+
+	do {
+		long m = READ_ONCE(counter->max) - page_counter_read(counter);
+
+		margin = min(margin, m);
+	} while ((counter = counter->parent));
+
+	return margin;
+}
+
 /**
  * page_counter_cancel - take pages out of the local counter
  * @counter: counter

base-commit: c73b725a57f276a3702ca213bde78fca029bc619
-- 
2.47.3

[PATCH v7 2/4] mm: distinguish large folio swap allocation failures

[Xueyuan Chen]

folio_alloc_swap() reports most failures with generic negative error
codes. Reclaim callers consequently cannot tell whether splitting a large
folio could make progress, or whether no swap space is available for even
a single page.

Classify failures using both the global free swap count and the remaining
capacity in the folio's memcg swap hierarchy. Return -ENOSPC when global
swap space is exhausted, -ENOMEM when splitting cannot overcome the
failure, and -E2BIG for a large folio when allocating or charging a
smaller folio might still succeed.

Use this classification for all folio_alloc_swap() failure paths,
including capability rejection, swap slot allocation failure, and memcg
swap charge failure. Callers are updated separately to split large folios
only on -E2BIG.

Suggested-by: Kairui Song <ryncsn@gmail.com>
Suggested-by: Barry Song <baohua@kernel.org>
Suggested-by: Youngjun Park <youngjun.park@lge.com>
Signed-off-by: Xueyuan Chen <xueyuan.chen21@gmail.com>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
---
 include/linux/swap.h |  6 ++++++
 mm/memcontrol.c      | 23 +++++++++++++++++++++++
 mm/swapfile.c        | 26 +++++++++++++++++++-------
 3 files changed, 48 insertions(+), 7 deletions(-)

diff --git a/include/linux/swap.h b/include/linux/swap.h
index 0544b2ec4c56..b23108d992aa 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -525,6 +525,7 @@ static inline void mem_cgroup_uncharge_swap(unsigned short id, unsigned int nr_p
 	__mem_cgroup_uncharge_swap(id, nr_pages);
 }
 
+long mem_cgroup_get_folio_swap_margin(struct folio *folio);
 extern long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg);
 extern bool mem_cgroup_swap_full(struct folio *folio);
 #else
@@ -538,6 +539,11 @@ static inline void mem_cgroup_uncharge_swap(unsigned short id,
 {
 }
 
+static inline long mem_cgroup_get_folio_swap_margin(struct folio *folio)
+{
+	return PAGE_COUNTER_MAX;
+}
+
 static inline long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
 {
 	return get_nr_swap_pages();
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 109c08be91cf..4b42f3fc6075 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -5676,6 +5676,29 @@ long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
 	return nr_swap_pages;
 }
 
+/**
+ * mem_cgroup_get_folio_swap_margin - get a folio's memcg swap margin
+ * @folio: folio whose memcg margin is queried
+ *
+ * Return: Remaining chargeable pages in the folio's memcg hierarchy.
+ */
+long mem_cgroup_get_folio_swap_margin(struct folio *folio)
+{
+	struct mem_cgroup *memcg;
+	long margin;
+
+	if (mem_cgroup_disabled() || do_memsw_account() ||
+	    !folio_memcg_charged(folio))
+		return PAGE_COUNTER_MAX;
+
+	rcu_read_lock();
+	memcg = folio_memcg(folio);
+	margin = page_counter_margin(&memcg->swap);
+	rcu_read_unlock();
+
+	return margin;
+}
+
 bool mem_cgroup_swap_full(struct folio *folio)
 {
 	struct mem_cgroup *memcg;
diff --git a/mm/swapfile.c b/mm/swapfile.c
index 70b90fa9c2a0..651682b1fe63 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -1735,7 +1735,9 @@ static int swap_dup_entries_cluster(struct swap_info_struct *si,
  * swap cache.
  *
  * Context: Caller needs to hold the folio lock.
- * Return: Whether the folio was added to the swap cache.
+ * Return: %0 on success, %-E2BIG if splitting the folio might allow swapout,
+ * %-ENOSPC if no global swap space is available, or %-ENOMEM if splitting
+ * would not help.
  */
 int folio_alloc_swap(struct folio *folio)
 {
@@ -1747,11 +1749,11 @@ int folio_alloc_swap(struct folio *folio)
 
 	if (order) {
 		/*
-		 * Reject large allocation when THP_SWAP is disabled,
-		 * the caller should split the folio and try again.
+		 * Reject large allocation when THP_SWAP is disabled. Check below
+		 * whether splitting and retrying can make progress.
 		 */
 		if (!IS_ENABLED(CONFIG_THP_SWAP))
-			return -EAGAIN;
+			goto failed;
 
 		/*
 		 * Allocation size should never exceed cluster size
@@ -1759,7 +1761,7 @@ int folio_alloc_swap(struct folio *folio)
 		 */
 		if (size > SWAPFILE_CLUSTER) {
 			VM_WARN_ON_ONCE(1);
-			return -EINVAL;
+			goto failed;
 		}
 	}
 
@@ -1775,13 +1777,23 @@ int folio_alloc_swap(struct folio *folio)
 	}
 
 	/* Need to call this even if allocation failed, for MEMCG_SWAP_FAIL. */
-	if (unlikely(mem_cgroup_try_charge_swap(folio)))
+	if (unlikely(mem_cgroup_try_charge_swap(folio))) {
 		swap_cache_del_folio(folio);
+		goto failed;
+	}
 
 	if (unlikely(!folio_test_swapcache(folio)))
-		return -ENOMEM;
+		goto failed;
 
 	return 0;
+
+failed:
+	if (get_nr_swap_pages() <= 0)
+		return -ENOSPC;
+	if (mem_cgroup_get_folio_swap_margin(folio) <= 0)
+		return -ENOMEM;
+
+	return order ? -E2BIG : -ENOMEM;
 }
 
 /**
-- 
2.47.3

[PATCH v7 3/4] mm/vmscan: avoid pointless large folio splits without swap

[Xueyuan Chen]

From: "Barry Song (Xiaomi)" <baohua@kernel.org>

When swap is disabled, exhausted, or unavailable due to memcg swap
limits, splitting a large anonymous folio cannot make swapout progress.
The fallback only destroys the large folio and inflates split statistics.

Use -E2BIG from folio_alloc_swap() as the explicit signal that splitting
the folio might allow swapout of smaller pieces. For other allocation
failures, keep the existing activation path and avoid the split.

This preserves the split fallback for fragmented or partially available
swap, while avoiding it when there is no backing space for any part of the
folio.

Reported-by: Nanzhe Zhao <zhaonanzhe@xiaomi.com>
Signed-off-by: Barry Song (Xiaomi) <baohua@kernel.org>
Signed-off-by: Xueyuan Chen <xueyuan.chen21@gmail.com>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
---
 mm/vmscan.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 26df81e773ff..457049e130df 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1263,6 +1263,8 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
 		 */
 		if (folio_test_anon(folio) && folio_test_swapbacked(folio) &&
 				!folio_test_swapcache(folio)) {
+			int ret;
+
 			if (!(sc->gfp_mask & __GFP_IO))
 				goto keep_locked;
 			if (folio_maybe_dma_pinned(folio))
@@ -1281,11 +1283,14 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
 				    split_folio_to_list(folio, folio_list))
 					goto activate_locked;
 			}
-			if (folio_alloc_swap(folio)) {
+			ret = folio_alloc_swap(folio);
+			if (ret) {
 				int __maybe_unused order = folio_order(folio);
 
 				if (!folio_test_large(folio))
 					goto activate_locked_split;
+				if (ret != -E2BIG)
+					goto activate_locked;
 				/* Fallback to swap normal pages */
 				if (split_folio_to_list(folio, folio_list))
 					goto activate_locked;
-- 
2.47.3

[PATCH v7 4/4] mm/shmem: split large folios only on -E2BIG

[Xueyuan Chen]

shmem_writeout() currently splits a large folio on every
folio_alloc_swap() failure. With the refined return-value contract, only
-E2BIG indicates that splitting might allow smaller folios to be swapped
out.

Enter the split fallback only for -E2BIG. For -ENOSPC and -ENOMEM,
redirty and reactivate the folio as before.

Suggested-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Signed-off-by: Xueyuan Chen <xueyuan.chen21@gmail.com>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Reviewed-by: Barry Song <baohua@kernel.org>
---
 mm/shmem.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/mm/shmem.c b/mm/shmem.c
index 6641823bed16..eff38bb33d8a 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1608,7 +1608,7 @@ int shmem_writeout(struct swap_io_ctx *ctx, struct folio *folio,
 	struct shmem_inode_info *info = SHMEM_I(inode);
 	struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
 	pgoff_t index;
-	int nr_pages;
+	int nr_pages, ret;
 	bool split = false;
 
 	if ((info->flags & SHMEM_F_LOCKED) || sbinfo->noswap)
@@ -1689,7 +1689,8 @@ int shmem_writeout(struct swap_io_ctx *ctx, struct folio *folio,
 		folio_mark_uptodate(folio);
 	}
 
-	if (!folio_alloc_swap(folio)) {
+	ret = folio_alloc_swap(folio);
+	if (!ret) {
 		bool first_swapped = shmem_recalc_inode(inode, 0, nr_pages);
 		int error;
 
@@ -1742,7 +1743,7 @@ int shmem_writeout(struct swap_io_ctx *ctx, struct folio *folio,
 		swap_cache_del_folio(folio);
 		goto redirty;
 	}
-	if (nr_pages > 1)
+	if (nr_pages > 1 && ret == -E2BIG)
 		goto try_split;
 redirty:
 	folio_mark_dirty(folio);
-- 
2.47.3

Re: [PATCH v7 1/4] mm: add page_counter_margin()

[Barry Song]

On Sun, Aug 30, 2026 at 12:29 PM Xueyuan Chen <xueyuan.chen21@gmail.com> wrote:
>
> From: Johannes Weiner <hannes@cmpxchg.org>
>
> mem_cgroup_get_nr_swap_pages() open-codes the remaining capacity across
> the memcg swap counter hierarchy.
>
> Add page_counter_margin() to return the minimum usable space from a page
> counter to the root, and use it in mem_cgroup_get_nr_swap_pages(). This
> is a pure refactoring with no intended behavior change.
>
> Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
> Signed-off-by: Xueyuan Chen <xueyuan.chen21@gmail.com>
> Reviewed-by: David Hildenbrand (Arm) <david@kernel.org>
> ---

Thanks for working over the weekend. I've been waiting for this
resending for a while. :-).  LGTM,

Reviewed-by: Barry Song <baohua@kernel.org>

Re: [PATCH v7 0/4] mm: avoid large folio splits when swap is unavailable

[Andrew Morton]

On Sun, 30 Aug 2026 12:29:16 +0800 Xueyuan Chen <xueyuan.chen21@gmail.com> wrote:

> This is v7 of Barry's original RFC patch, "mm: Avoiding split large
> folios if swap has no space":
> 
> https://lore.kernel.org/r/20260618221720.71768-1-baohua@kernel.org
> 
> Barry's RFC showed the no-swap case with MADV_PAGEOUT on 16KB mTHP: the
> large-folio split counter increased by 1024 even though no swapout
> progress was possible. Skipping the split in that case kept the counter
> at 0.
> 
> This series makes folio_alloc_swap() classify failures according to
> whether splitting a large folio might allow swapout to make progress.
> Callers can then avoid destroying the large folio when neither global
> swap availability nor the folio's memcg swap hierarchy has capacity for
> even a smaller folio.

Thanks.  Sashiko had a question which didn't occur against v6:

	https://sashiko.dev/#/patchset/20260830-folio_swap_entry-v1-0-7786b52acbc8@columbia.edu

Re: [PATCH v7 0/4] mm: avoid large folio splits when swap is unavailable

[Xueyuan Chen]

[-- Attachment #1: Type: text/plain, Size: 1434 bytes --]

On Tue, Sep 1, 2026 at 11:39 AM Andrew Morton <akpm@linux-foundation.org>
wrote:

> On Sun, 30 Aug 2026 12:29:16 +0800 Xueyuan Chen <xueyuan.chen21@gmail.com>
> wrote:
>
> > This is v7 of Barry's original RFC patch, "mm: Avoiding split large
> > folios if swap has no space":
> >
> > https://lore.kernel.org/r/20260618221720.71768-1-baohua@kernel.org
> >
> > Barry's RFC showed the no-swap case with MADV_PAGEOUT on 16KB mTHP: the
> > large-folio split counter increased by 1024 even though no swapout
> > progress was possible. Skipping the split in that case kept the counter
> > at 0.
> >
> > This series makes folio_alloc_swap() classify failures according to
> > whether splitting a large folio might allow swapout to make progress.
> > Callers can then avoid destroying the large folio when neither global
> > swap availability nor the folio's memcg swap hierarchy has capacity for
> > even a smaller folio.
>
> Thanks.  Sashiko had a question which didn't occur against v6:
>
>
> https://sashiko.dev/#/patchset/20260830-folio_swap_entry-v1-0-7786b52acbc8@columbia.edu


Hi Andrew,

I think the Sashiko link points to a different series:

[PATCH 0/8] mm: remove page_swap_entry()

The Sashiko page for this v7 reports "No issues found":

https://sashiko.dev/#/patchset/20260830042920.2280454-1-xueyuan.chen21@gmail.com


Did you intend to link a different question?

Thanks,
Xueyuan

[-- Attachment #2: Type: text/html, Size: 2444 bytes --]