mm/rmap: fix soft-dirty and uffd-wp bit loss when remapping zero-filled mTHP subpage to shared zeropage

From: Lance Yang <[email protected]>

When splitting an mTHP and replacing a zero-filled subpage with the shared zeropage, try_to_map_unused_to_zeropage() currently drops several important PTE bits.

For userspace tools like CRIU, which rely on the soft-dirty mechanism for incremental snapshots, losing the soft-dirty bit means modified pages are missed, leading to inconsistent memory state after restore.

As pointed out by David, the more critical uffd-wp bit is also dropped. This breaks the userfaultfd write-protection mechanism, causing writes to be silently missed by monitoring applications, which can lead to data corruption.

Preserve both the soft-dirty and uffd-wp bits from the old PTE when creating the new zeropage mapping to ensure they are correctly tracked.

Cc <[email protected]>

Fixes b1f202060afe ("mm: remap unused subpages to shared zeropage when splitting isolated thp")

Suggested-by David Hildenbrand <[email protected]>

Suggested-by Dev Jain <[email protected]>

Acked-by David Hildenbrand <[email protected]>

Signed-off-by Lance Yang <[email protected]>

---
v1 -> v2:
- Avoid calling ptep_get() multiple times (per Dev)
- Double-check the uffd-wp bit (per David)
- Collect Acked-by from David - thanks!
- https://lore.kernel.org/linux-mm/[email protected]/

  mm/migrate.c | 9 ++++++++-
  1 file changed, 8 insertions(+), 1 deletion(-)

mm/migrate.c +0 -0

index ce83c2c3c287..50aa91d9ab4e 100644

mm/migrate.c +7 -1

  	pte_t newpte;
 
  	if (PageCompound(page))
  		return false;
  	VM_BUG_ON_PAGE(!PageAnon(page), page);
  	VM_BUG_ON_PAGE(!PageLocked(page), page);
-	VM_BUG_ON_PAGE(pte_present(ptep_get(pvmw->pte)), page);
+	VM_BUG_ON_PAGE(pte_present(oldpte), page);
 
  	if (folio_test_mlocked(folio) || (pvmw->vma->vm_flags & VM_LOCKED) ||
  	    mm_forbids_zeropage(pvmw->vma->vm_mm))

@@ -322,6 +323,12 @@ static bool try_to_map_unused_to_zeropage(struct page_vma_mapped_walk *pvmw,
 
  	newpte = pte_mkspecial(pfn_pte(my_zero_pfn(pvmw->address),
  					pvmw->vma->vm_page_prot));
+
+	if (pte_swp_soft_dirty(oldpte))
+		newpte = pte_mksoft_dirty(newpte);
+	if (pte_swp_uffd_wp(oldpte))
+		newpte = pte_mkuffd_wp(newpte);
+
  	set_pte_at(pvmw->vma->vm_mm, pvmw->address, pvmw->pte, newpte);
 
  	dec_mm_counter(pvmw->vma->vm_mm, mm_counter(folio));

Quoted 34 lines · Lance Yang

From: Lance Yang <[email protected]>

When splitting an mTHP and replacing a zero-filled subpage with the shared zeropage, try_to_map_unused_to_zeropage() currently drops several important PTE bits.

For userspace tools like CRIU, which rely on the soft-dirty mechanism for incremental snapshots, losing the soft-dirty bit means modified pages are missed, leading to inconsistent memory state after restore.

As pointed out by David, the more critical uffd-wp bit is also dropped. This breaks the userfaultfd write-protection mechanism, causing writes to be silently missed by monitoring applications, which can lead to data corruption.

Preserve both the soft-dirty and uffd-wp bits from the old PTE when creating the new zeropage mapping to ensure they are correctly tracked.

Cc <[email protected]>

Fixes b1f202060afe ("mm: remap unused subpages to shared zeropage when splitting isolated thp")

Suggested-by David Hildenbrand <[email protected]>

Suggested-by Dev Jain <[email protected]>

Acked-by David Hildenbrand <[email protected]>

Signed-off-by Lance Yang <[email protected]>

---
v1 -> v2:
  - Avoid calling ptep_get() multiple times (per Dev)
  - Double-check the uffd-wp bit (per David)
  - Collect Acked-by from David - thanks!
  - https://lore.kernel.org/linux-mm/20250928044855.76359-1- [email protected]/

  mm/migrate.c | 9 ++++++++-

  1 file changed, 8 insertions(+), 1 deletion(-)

mm/migrate.c static bool try_to_map_unused_to_zeropage(struct · line 304

Quoted 10 lines

mm/migrate.c +1 -0

index ce83c2c3c287..50aa91d9ab4e 100644
@@ -300,13 +300,14 @@ static bool try_to_map_unused_to_zeropage(struct 
 page_vma_mapped_walk *pvmw,
                         unsigned long idx)
   {
       struct page *page = folio_page(folio, idx);
+    pte_t oldpte = ptep_get(pvmw->pte);

What I meant to say was, you can pass oldpte from remove_migration_pte to this function. Basically define old_pte = ptep_get(pvmw.pte) in the declarations of the start of the while block in remove_migration_pte and remove the existing one. That will ensure ptep_get() gets called only once per iteration.

Quoted 22 lines

mm/migrate.c +1 -1

       pte_t newpte;
       if (PageCompound(page))
           return false;
       VM_BUG_ON_PAGE(!PageAnon(page), page);
       VM_BUG_ON_PAGE(!PageLocked(page), page);
-    VM_BUG_ON_PAGE(pte_present(ptep_get(pvmw->pte)), page);
+    VM_BUG_ON_PAGE(pte_present(oldpte), page);

      if (folio_test_mlocked(folio) || (pvmw->vma->vm_flags & VM_LOCKED) || mm_forbids_zeropage(pvmw->vma->vm_mm))

mm/migrate.c +6 -0

@@ -322,6 +323,12 @@ static bool try_to_map_unused_to_zeropage(struct 
 page_vma_mapped_walk *pvmw,
       newpte = pte_mkspecial(pfn_pte(my_zero_pfn(pvmw->address),
                       pvmw->vma->vm_page_prot));
+
+    if (pte_swp_soft_dirty(oldpte))
+        newpte = pte_mksoft_dirty(newpte);
+    if (pte_swp_uffd_wp(oldpte))
+        newpte = pte_mkuffd_wp(newpte);
+
       set_pte_at(pvmw->vma->vm_mm, pvmw->address, pvmw->pte, newpte);
       dec_mm_counter(pvmw->vma->vm_mm, mm_counter(folio));

Quoted 36 lines · Dev Jain

Quoted 34 lines · Lance Yang

From: Lance Yang <[email protected]>

When splitting an mTHP and replacing a zero-filled subpage with the shared zeropage, try_to_map_unused_to_zeropage() currently drops several important PTE bits.

For userspace tools like CRIU, which rely on the soft-dirty mechanism for incremental snapshots, losing the soft-dirty bit means modified pages are missed, leading to inconsistent memory state after restore.

As pointed out by David, the more critical uffd-wp bit is also dropped. This breaks the userfaultfd write-protection mechanism, causing writes to be silently missed by monitoring applications, which can lead to data corruption.

Preserve both the soft-dirty and uffd-wp bits from the old PTE when creating the new zeropage mapping to ensure they are correctly tracked.

Cc <[email protected]>

Fixes b1f202060afe ("mm: remap unused subpages to shared zeropage when splitting isolated thp")

Suggested-by David Hildenbrand <[email protected]>

Suggested-by Dev Jain <[email protected]>

Acked-by David Hildenbrand <[email protected]>

Signed-off-by Lance Yang <[email protected]>

---
v1 -> v2:
  - Avoid calling ptep_get() multiple times (per Dev)
  - Double-check the uffd-wp bit (per David)
  - Collect Acked-by from David - thanks!
  - https://lore.kernel.org/linux-mm/20250928044855.76359-1- [email protected]/

  mm/migrate.c | 9 ++++++++-

  1 file changed, 8 insertions(+), 1 deletion(-)

Quoted 15 lines

mm/migrate.c static bool · line 304

Quoted 10 lines

mm/migrate.c +1 -0

index ce83c2c3c287..50aa91d9ab4e 100644
@@ -300,13 +300,14 @@ static bool 
 try_to_map_unused_to_zeropage(struct page_vma_mapped_walk *pvmw,
                         unsigned long idx)
   {
       struct page *page = folio_page(folio, idx);
+    pte_t oldpte = ptep_get(pvmw->pte);

What I meant to say was, you can pass oldpte from remove_migration_pte to this function. Basically define old_pte = ptep_get(pvmw.pte) in the declarations of the start of the while block in remove_migration_pte and remove the existing one. That will ensure ptep_get() gets called only once per iteration.

Ah, got it. Thanks for the clarification!

IIUC, you mean something like this:

```

mm/migrate.c +1 -0

index ce83c2c3c287..bafd8cb3bebe 100644
@@ -297,6 +297,7 @@ bool isolate_folio_to_list(struct folio *folio, 
 struct list_head *list)
 
  static bool try_to_map_unused_to_zeropage(struct page_vma_mapped_walk *pvmw,
                        struct folio *folio,
+                      pte_t old_pte,
                        unsigned long idx)
  {

     struct page *page = folio_page(folio, idx);

mm/migrate.c +1 -1

@@ -306,7 +307,7 @@ static bool try_to_map_unused_to_zeropage(struct 
 page_vma_mapped_walk *pvmw,
          return false;
      VM_BUG_ON_PAGE(!PageAnon(page), page);
      VM_BUG_ON_PAGE(!PageLocked(page), page);
-    VM_BUG_ON_PAGE(pte_present(ptep_get(pvmw->pte)), page);
+    VM_BUG_ON_PAGE(pte_present(old_pte), page);
 

     if (folio_test_mlocked(folio) || (pvmw->vma->vm_flags & VM_LOCKED) || mm_forbids_zeropage(pvmw->vma->vm_mm))

mm/migrate.c +6 -0

@@ -322,6 +323,12 @@ static bool try_to_map_unused_to_zeropage(struct 
 page_vma_mapped_walk *pvmw,
 
      newpte = pte_mkspecial(pfn_pte(my_zero_pfn(pvmw->address),
                      pvmw->vma->vm_page_prot));
+
+    if (pte_swp_soft_dirty(old_pte))
+        newpte = pte_mksoft_dirty(newpte);
+    if (pte_swp_uffd_wp(old_pte))
+        newpte = pte_mkuffd_wp(newpte);
+
      set_pte_at(pvmw->vma->vm_mm, pvmw->address, pvmw->pte, newpte);
 

     dec_mm_counter(pvmw->vma->vm_mm, mm_counter(folio));

mm/migrate.c +1 -1

@@ -344,7 +351,7 @@ static bool remove_migration_pte(struct folio *folio,
 
      while (page_vma_mapped_walk(&pvmw)) {
          rmap_t rmap_flags = RMAP_NONE;
-        pte_t old_pte;
+        pte_t old_pte = ptep_get(pvmw.pte);
          pte_t pte;
          swp_entry_t entry;
          struct page *new;

@@ -365,12 +372,11 @@ static bool remove_migration_pte(struct folio 
 *folio,
          }
  #endif

         if (rmap_walk_arg->map_unused_to_zeropage &&
-            try_to_map_unused_to_zeropage(&pvmw, folio, idx)) +            try_to_map_unused_to_zeropage(&pvmw, folio, old_pte, idx)) continue;

         folio_get(folio);
         pte = mk_pte(new, READ_ONCE(vma->vm_page_prot)); -        old_pte = ptep_get(pvmw.pte);

         entry = pte_to_swp_entry(old_pte);
         if (!is_migration_entry_young(entry))
```

ptep_get() gets called only once per iteration, right?