Re: [PATCH v2] mm: use memory barriers for non PMD-sized folios to sync

From: Baolin Wang

Date: Wed Nov 18 2026 - 01:17:34 EST




On 11/17/26 9:53 PM, Lorenzo Stoakes (ARM) wrote:
On Tue, Nov 17, 2026 at 03:00:30PM +0800, Baolin Wang wrote:
Commit 243cf1f67b36 ("mm: memory: fix truncation removing mapped folios")
uses PMD lock via sync_with_folio_pmd_zap() to synchronize when a cleared
PMD is found while unmapping a folio.

However, for non PMD-sized folios, the only PMD clear that can race with
this unmap is PTE table reclaim, which clears the PMD only after the rmaps
for the table have been removed. This is plain message passing, and taking
the PMD lock is more than needed, like I did in the previous patch[1].

But it doesn't automatically mean that you _have_ to use fragile memory
barriers?

Correctness and robustness in the future are important right? You need to
justify this properly.

Sorry, I wasn't clear. The point of this patch is to avoid a possible performance regression from taking the PMD lock to sync for non PMD-sized folios. So using memory barriers can avoid this.

The description below explains why memory barriers are enough here.

Without barriers, the walker can observe the cleared PMD but a stale
mapcount on weakly ordered architectures:

You mean without the PMD lock?

Yes, will make it clear.

CPU 0 (unmap folio) CPU 1 (MADV_DONTNEED, full PMD)
================================ ================================
folio_lock()
lock(ptl)
clear PTEs
remove rmaps /* W mapcount */
[mapcount load satisfied early,
folio still mapped]
pmd_clear() /* W pmd */
unlock(ptl)
pmdp_get_lockless() -> none /* R pmd */
skip the PMD
folio_mapped() -> true /* R mapcount, stale */
unmapping assumed to have failed

Pairing an smp_wmb() before the PMD clear with an smp_rmb() after the
PMD read forbids this outcome:

CPU 0 (unmap folio) CPU 1 (MADV_DONTNEED, full PMD)
================================ ================================
remove rmaps /* W mapcount */
smp_wmb()
pmd_clear() /* W pmd */
pmdp_get_lockless() -> none /* R pmd */
smp_rmb()
folio_mapped() /* R mapcount */

If R pmd observes W pmd, R mapcount must observe W mapcount, so
folio_mapped() no longer reports the removed mappings.

This is insufficient IMO, you need to explain firstly why acquire/release isn't
feasible (presumably because you can't synchronise the clear properly that way).

Sure. Using the PMD lock (with its implicit acquire/release semantics) would also work. However, as I mentioned above, the concern is the potential performance regression.

Pass a 'folio' parameter to sync_with_folio_pmd_zap(), allowing us to
skip those pfn walkers: pfn_mkclean_range and mapping_wrprotect_range,
which don't need synchronization.

Similarly, skip sync for page_mapped_in_vma() since the subsequent
operations don't check the folio's mapcount.

No changes for PMD-sized folios unmapping.

[1] https://lore.kernel.org/all/542c97482eae84f86704bcee56482633d1b32321.1791616645.git.baolin.wang@xxxxxxxxxxxxxxxxx/
Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>

This replaces an easy-to-reason about lock with a very-hard-to-reason about
memory barriers that are fragile and could easily break resulting in subtle data
race bugs.

And you don't seem to have provided any perf numbers or anything to justify it
unless I'm missing something?

Yes. Performance measument for this race path may be tricky, but I'll try to see whether I can show a measurable benefit with memory barriers.

I'm a little uneasy about it honestly.

---
Changes from v1:
- Pass in a folio for sync_with_folio_pmd_zap(), per David.
---
include/linux/rmap.h | 2 ++
mm/internal.h | 26 ++++++++++++++++++++++----
mm/memory.c | 12 ++++++++++--
mm/page_vma_mapped.c | 4 ++--
4 files changed, 36 insertions(+), 8 deletions(-)

diff --git a/include/linux/rmap.h b/include/linux/rmap.h
index 74cca0e3c726..62e7341de179 100644
--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h
@@ -862,6 +862,7 @@ struct page *make_device_exclusive(struct mm_struct *mm, unsigned long addr,
#define PVMW_PGTABLE_CROSSED (1 << 16)

struct page_vma_mapped_walk {
+ struct folio *folio;
unsigned long pfn;
unsigned long nr_pages;
pgoff_t pgoff; /* Only meaningful if nr_pages > 1 and not a KSM walk */
@@ -876,6 +877,7 @@ struct page_vma_mapped_walk {

#define DEFINE_FOLIO_VMA_WALK(name, _folio, _vma, _address, _flags) \
struct page_vma_mapped_walk name = { \
+ .folio = _folio, \
.pfn = folio_pfn(_folio), \
.nr_pages = folio_nr_pages(_folio), \
.pgoff = folio_pgoff(_folio), \
diff --git a/mm/internal.h b/mm/internal.h
index cd3675c044e0..ad7e051cb198 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -592,6 +592,7 @@ void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte);
* sync_with_folio_pmd_zap - sync with concurrent zapping of folios
* @mm: The mm_struct.
* @pmdp: Pointer to the pmd that was found to be pmd_none().
+ * @folio: The folio being zapped.
*
* When we find a pmd_none() while unmapping a folio without holding the PTL,
* zap_huge_pmd() may have cleared the PMD but not yet updated the folio
@@ -600,13 +601,30 @@ void pmd_install(struct mm_struct *mm, pmd_t *pmd, pgtable_t *pte);
* against later mapcount reads. Skipping the PMD without synchronization
* could make folio unmapping code assume that unmapping failed.
*
- * Wait for concurrent zapping to complete by grabbing the PTL.
+ * Wait for concurrent zapping to complete by grabbing the PTL for PMD-sized
+ * folios, otherwise smp_rmb() is sufficient.
*/
-static inline void sync_with_folio_pmd_zap(struct mm_struct *mm, pmd_t *pmdp)
+static inline void sync_with_folio_pmd_zap(struct mm_struct *mm,
+ pmd_t *pmdp, struct folio *folio)
{
- spinlock_t *ptl = pmd_lock(mm, pmdp);
+ if (folio_test_pmd_mappable(folio)) {
+ spinlock_t *ptl = pmd_lock(mm, pmdp);

- spin_unlock(ptl);
+ spin_unlock(ptl);
+ return;
+ }
+
+ /*
+ * For non PMD-sized folios zapping, concurrent zap_empty_pte_table()
+ * may clear the PMD after the rmaps for the table have been removed,
+ * but later mapcount reads could still be observed before the PMD
+ * read, making the folio appear to be still mapped after unmapping.
+ * Order the PMD read before later mapcount reads, so that observing
+ * the cleared PMD implies observing the rmap removals.
+ *

This also makes me very uneasy, complicated reasoning like this is asking for it
to break.

Also could this be smaller and clearer?

OK. Let me try.

Thanks.