Re: [PATCH v5 10/11] mm: install PMD swap entries on swap-out

[Usama Arif]

On 23/07/2026 20:16, Matthew Wilcox wrote:
> On Wed, Jul 22, 2026 at 08:19:41AM -0700, Usama Arif wrote:
>> +++ b/mm/rmap.c
>> @@ -2282,6 +2282,25 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
>>  				goto walk_abort;
>>  			}
>>  
>> +#ifdef CONFIG_THP_SWAP
>> +			/*
>> +			 * If the folio is in the swap cache and we're not
>> +			 * asked to split, install a PMD-level swap entry.
>> +			 */
>> +			if (!(flags & TTU_SPLIT_HUGE_PMD) &&
>> +			    folio_test_anon(folio) &&
>> +			    folio_test_swapcache(folio)) {
>> +				if (set_pmd_swap_entry(&pvmw, folio))
>> +					goto walk_abort;
>> +
>> +				add_mm_counter(mm, MM_ANONPAGES,
>> +					       -HPAGE_PMD_NR);
>> +				add_mm_counter(mm, MM_SWAPENTS,
>> +					       HPAGE_PMD_NR);
>> +				goto walk_done;
>> +			}
>> +#endif
>> +
>>  			if (flags & TTU_SPLIT_HUGE_PMD) {
>>  				/*
>>  				 * We temporarily have to drop the PTL and
> 
> This makes me sad.  It feels like we're bolting more complexity onto
> try_to_unmap() instead of removing TTU_SPLIT_HUGE_PMD entirely.
> 
> Ideally we'd just cope with folios of whatever size (including
> PMD and PUD mapped folios) and only handle hugetlb weirdness when we
> absolutely have to.  Similarly, we could split PUDs to PMDs and
> PMDs to PTEs when we need to, without the caller specifying
> TTU_SPLIT_anything.
> 
> Is there a reason we can't do that?

Hi Matthew,

So I have been looking into this over the past few days, and I can't
come up with a way to do this without having another rmap walk, which
is expensive.

I tried to prototype this (code at the end of the series).
Removing the flag required an additional rmap walk (split_folio_pmd_mappings below),
MMU notifier handling and different fallback behavior on architectures without
PMD softleaf support (see #ifdef CONFIG_ARCH_HAS_PMD_SOFTLEAVES below).
It makes the code more complicated and expensive.

It also changes the behavior of every try_to_unmap() and try_to_migrate()
caller, while this series only needs reclaim to preserve a PMD when it can
install a PMD swap entry. I would therefore prefer to keep this series scoped
to PMD swap entries and handle removal of TTU_SPLIT_HUGE_PMD independently
with the broader rmap, migration, architecture, and THP-splitting review
and testing it requires.



Untested prototype below:


diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index a174758f7777..0e0172c0c852 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -93,7 +93,6 @@ struct anon_vma_chain {
 
 enum ttu_flags {
 	TTU_USE_SHARED_ZEROPAGE	= 0x2,	/* for unused pages of large folios */
-	TTU_SPLIT_HUGE_PMD	= 0x4,	/* split huge PMD if any */
 	TTU_IGNORE_MLOCK	= 0x8,	/* ignore mlock */
 	TTU_SYNC		= 0x10,	/* avoid racy checks with PVMW_SYNC */
 	TTU_HWPOISON		= 0x20,	/* do convert pte to hwpoison entry */
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 6f7acae5a5fd..1012cb625452 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -3905,6 +3905,38 @@ void vma_adjust_trans_huge(struct vm_area_struct *vma,
 		split_huge_pmd_if_needed(next, end);
 }
 
+static bool split_folio_pmd_mapping(struct folio *folio,
+				    struct vm_area_struct *vma,
+				    unsigned long address, void *arg)
+{
+	DEFINE_FOLIO_VMA_WALK(pvmw, folio, vma, address, PVMW_SYNC);
+	struct mmu_notifier_range range;
+
+	range.end = vma_address_end(&pvmw);
+	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma->vm_mm,
+				address, range.end);
+	mmu_notifier_invalidate_range_start(&range);
+
+	while (page_vma_mapped_walk(&pvmw)) {
+		if (pvmw.pte)
+			continue;
+		split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd, false);
+		page_vma_mapped_walk_restart(&pvmw);
+	}
+
+	mmu_notifier_invalidate_range_end(&range);
+	return true;
+}
+
+static void split_folio_pmd_mappings(struct folio *folio)
+{
+	struct rmap_walk_control rwc = {
+		.rmap_one = split_folio_pmd_mapping,
+	};
+
+	rmap_walk_locked(folio, &rwc);
+}
+
 static void unmap_folio(struct folio *folio)
 {
 	enum ttu_flags ttu_flags = TTU_RMAP_LOCKED | TTU_SYNC |
@@ -3912,18 +3944,23 @@ static void unmap_folio(struct folio *folio)
 
 	VM_BUG_ON_FOLIO(!folio_test_large(folio), folio);
 
-	if (folio_test_pmd_mappable(folio))
-		ttu_flags |= TTU_SPLIT_HUGE_PMD;
-
 	/*
 	 * Anon pages need migration entries to preserve them, but file
 	 * pages can simply be left unmapped, then faulted back on demand.
 	 * If that is ever changed (perhaps for mlock), update remap_page().
 	 */
-	if (folio_test_anon(folio))
+	if (folio_test_anon(folio)) {
+		if (folio_test_pmd_mappable(folio))
+			split_folio_pmd_mappings(folio);
 		try_to_migrate(folio, ttu_flags);
-	else
+	} else {
 		try_to_unmap(folio, ttu_flags | TTU_IGNORE_MLOCK);
+	}
 
 	try_to_unmap_flush();
 }
diff --git a/mm/rmap.c b/mm/rmap.c
index b7ead3e9f064..f77f380eb03c 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -2282,17 +2282,13 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 				goto walk_abort;
 			}
 
-			if (flags & TTU_SPLIT_HUGE_PMD) {
-				/*
-				 * We temporarily have to drop the PTL and
-				 * restart so we can process the PTE-mapped THP.
-				 */
-				split_huge_pmd_locked(vma, pvmw.address,
-						      pvmw.pmd, false);
-				flags &= ~TTU_SPLIT_HUGE_PMD;
-				page_vma_mapped_walk_restart(&pvmw);
-				continue;
-			}
+			/*
+			 * No PMD-level unmap operation handled this mapping, so
+			 * split it and retry at PTE granularity.
+			 */
+			split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd, false);
+			page_vma_mapped_walk_restart(&pvmw);
+			continue;
 		}
 
 		/* Unexpected PMD-mapped THP? */
@@ -2459,8 +2455,6 @@ void try_to_unmap(struct folio *folio, enum ttu_flags flags)
 /*
  * @arg: enum ttu_flags will be passed to this argument.
  *
- * If TTU_SPLIT_HUGE_PMD is specified any PMD mappings will be split into PTEs
- * containing migration entries.
  */
 static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
 		     unsigned long address, void *arg)
@@ -2509,24 +2503,11 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
 	mmu_notifier_invalidate_range_start(&range);
 
 	while (page_vma_mapped_walk(&pvmw)) {
-		/* PMD-mapped THP migration entry */
+		/* PMD-mapped THP */
 		if (!pvmw.pte) {
 			__maybe_unused unsigned long pfn;
 			__maybe_unused pmd_t pmdval;
 
-			if (flags & TTU_SPLIT_HUGE_PMD) {
-				/*
-				 * split_huge_pmd_locked() might leave the
-				 * folio mapped through PTEs. Retry the walk
-				 * so we can detect this scenario and properly
-				 * abort the walk.
-				 */
-				split_huge_pmd_locked(vma, pvmw.address,
-						      pvmw.pmd, true);
-				flags &= ~TTU_SPLIT_HUGE_PMD;
-				page_vma_mapped_walk_restart(&pvmw);
-				continue;
-			}
 #ifdef CONFIG_ARCH_HAS_PMD_SOFTLEAVES
 			pmdval = pmdp_get(pvmw.pmd);
 			if (likely(pmd_present(pmdval)))
@@ -2545,6 +2526,11 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
 				break;
 			}
 			continue;
+#else
+			/* Fall back to PTE migration entries on this architecture. */
+			split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd, true);
+			page_vma_mapped_walk_restart(&pvmw);
+			continue;
 #endif
 		}
 
@@ -2799,11 +2785,11 @@ void try_to_migrate(struct folio *folio, enum ttu_flags flags)
 	};
 
 	/*
-	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
-	 * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
+	 * Migration always ignores mlock and only supports TTU_RMAP_LOCKED,
+	 * TTU_SYNC, and TTU_BATCH_FLUSH flags.
 	 */
-	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
-					TTU_SYNC | TTU_BATCH_FLUSH)))
+	if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SYNC |
+					TTU_BATCH_FLUSH)))
 		return;
 
 	if (folio_is_zone_device(folio) &&
diff --git a/mm/truncate.c b/mm/truncate.c
index b58ba940be47..12e1fea5cf50 100644
--- a/mm/truncate.c
+++ b/mm/truncate.c
@@ -182,7 +182,6 @@ static int folio_split_or_unmap(struct folio *folio, struct page *split_at,
 {
 	enum ttu_flags ttu_flags =
 		TTU_SYNC |
-		TTU_SPLIT_HUGE_PMD |
 		TTU_IGNORE_MLOCK;
 	int ret;
 
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 9b8ee902f972..90751b384a5d 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -1333,8 +1333,6 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
 			enum ttu_flags flags = TTU_BATCH_FLUSH;
 			bool was_swapbacked = folio_test_swapbacked(folio);
 
-			if (folio_test_pmd_mappable(folio))
-				flags |= TTU_SPLIT_HUGE_PMD;
 			/*
 			 * Without TTU_SYNC, try_to_unmap will only begin to
 			 * hold PTL from the first present PTE within a large
-- 
2.53.0-Meta