[RFC PATCH 2/5] mm/mlock: fixup mlock_count during unmap

[Yosry Ahmed]

In the rare case where an mlocked order-0 folio is mapped 2^20 or more
times, the high mapcount can be interpreted mistakenly by munlock() as
an mlock_count, causing PG_mlocked to not be cleared, possibly leaving
the folio stranded as unevictable endlessly.

To fix this, add a hook during unmapping to check if the bits used for
mlock_count are 0s yet PG_mlocked is set. In this case, call make sure
to perform the missed munlock operation.

Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
---
 include/linux/mm.h |  4 ++++
 mm/mlock.c         | 18 +++++++++++++++++-
 mm/rmap.c          |  1 +
 3 files changed, 22 insertions(+), 1 deletion(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 3994580772b3..b341477a83e8 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1050,6 +1050,7 @@ unsigned long vmalloc_to_pfn(const void *addr);
 extern bool is_vmalloc_addr(const void *x);
 extern int is_vmalloc_or_module_addr(const void *x);
 extern int folio_mlocked_mapcount(struct folio *folio);
+extern void folio_mlock_unmap_check(struct folio *folio);
 #else
 static inline bool is_vmalloc_addr(const void *x)
 {
@@ -1063,6 +1064,9 @@ static inline int folio_mlocked_mapcount(struct folio *folio)
 {
 	return 0;
 }
+static inline void folio_mlock_unmap_check(struct folio *folio)
+{
+}
 #endif
 
 /*
diff --git a/mm/mlock.c b/mm/mlock.c
index 5c5462627391..8261df11d6a6 100644
--- a/mm/mlock.c
+++ b/mm/mlock.c
@@ -66,7 +66,8 @@ EXPORT_SYMBOL(can_do_mlock);
  * (1) The mapcount will be incorrect (underestimated). It will be correct again
  * once the number of mappings falls below MLOCK_COUNT_BIAS.
  * (2) munlock() can misinterpret the large number of mappings as an mlock_count
- * and leave PG_mlocked set.
+ * and leave PG_mlocked set. This will be fixed when the number of mappings
+ * falls below MLOCK_COUNT_BIAS by folio_mlock_unmap_check().
  */
 #define MLOCK_COUNT_SHIFT 20
 #define MLOCK_COUNT_BIAS (1U << MLOCK_COUNT_SHIFT)
@@ -139,6 +140,21 @@ static int folio_mlock_count_dec(struct folio *folio)
 	return mlock_count - 1;
 }
 
+/*
+ * Call after decrementing the mapcount. If the mapcount previously overflowed
+ * beyond the lower 20 bits for an order-0 mlocked folio, munlock() have
+ * mistakenly left the folio mlocked. Fix it here.
+ */
+void folio_mlock_unmap_check(struct folio *folio)
+{
+	int mapcount = atomic_read(&folio->_mapcount) + 1;
+	int mlock_count = mapcount >> MLOCK_COUNT_SHIFT;
+
+	if (unlikely(!folio_test_large(folio) && folio_test_mlocked(folio) &&
+		     mlock_count == 0))
+		munlock_folio(folio);
+}
+
 /*
  * Mlocked folios are marked with the PG_mlocked flag for efficient testing
  * in vmscan and, possibly, the fault path; and to support semi-accurate
diff --git a/mm/rmap.c b/mm/rmap.c
index 19392e090bec..02e558551f15 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1392,6 +1392,7 @@ void page_remove_rmap(struct page *page, struct vm_area_struct *vma,
 			nr = atomic_dec_return_relaxed(mapped);
 			nr = (nr < COMPOUND_MAPPED);
 		}
+		folio_mlock_unmap_check(folio);
 	} else if (folio_test_pmd_mappable(folio)) {
 		/* That test is redundant: it's for safety or to optimize out */
 
-- 
2.41.0.162.gfafddb0af9-goog