diff --git a/include/linux/mm.h b/include/linux/mm.h
index 2dd73e4f3d8e..406ab9ea818f 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -641,8 +641,14 @@ static inline void vma_numab_state_free(struct vm_area_struct *vma) {}
  */
 static inline bool vma_start_read(struct vm_area_struct *vma)
 {
-	/* Check before locking. A race might cause false locked result. */
-	if (vma->vm_lock_seq == READ_ONCE(vma->vm_mm->mm_lock_seq))
+	/*
+	 * Check before locking. A race might cause false locked result.
+	 * We can use READ_ONCE() for the mm_lock_seq here, and don't need
+	 * ACQUIRE semantics, because this is just a lockless check whose result
+	 * we don't rely on for anything - the mm_lock_seq read against which we
+	 * need ordering is below.
+	 */
+	if (READ_ONCE(vma->vm_lock_seq) == READ_ONCE(vma->vm_mm->mm_lock_seq))
 		return false;
 
 	if (unlikely(down_read_trylock(&vma->vm_lock->lock) == 0))
@@ -653,8 +659,13 @@ static inline bool vma_start_read(struct vm_area_struct *vma)
 	 * False unlocked result is impossible because we modify and check
 	 * vma->vm_lock_seq under vma->vm_lock protection and mm->mm_lock_seq
 	 * modification invalidates all existing locks.
+	 *
+	 * We must use ACQUIRE semantics for the mm_lock_seq so that if we are
+	 * racing with vma_end_write_all(), we only start reading from the VMA
+	 * after it has been unlocked.
+	 * This pairs with RELEASE semantics in vma_end_write_all().
 	 */
-	if (unlikely(vma->vm_lock_seq == READ_ONCE(vma->vm_mm->mm_lock_seq))) {
+	if (unlikely(vma->vm_lock_seq == smp_load_acquire(&vma->vm_mm->mm_lock_seq))) {
 		up_read(&vma->vm_lock->lock);
 		return false;
 	}
@@ -676,7 +687,7 @@ static bool __is_vma_write_locked(struct vm_area_struct *vma, int *mm_lock_seq)
 	 * current task is holding mmap_write_lock, both vma->vm_lock_seq and
 	 * mm->mm_lock_seq can't be concurrently modified.
 	 */
-	*mm_lock_seq = READ_ONCE(vma->vm_mm->mm_lock_seq);
+	*mm_lock_seq = vma->vm_mm->mm_lock_seq;
 	return (vma->vm_lock_seq == *mm_lock_seq);
 }
 
@@ -688,7 +699,13 @@ static inline void vma_start_write(struct vm_area_struct *vma)
 		return;
 
 	down_write(&vma->vm_lock->lock);
-	vma->vm_lock_seq = mm_lock_seq;
+	/*
+	 * We should use WRITE_ONCE() here because we can have concurrent reads
+	 * from the early lockless pessimistic check in vma_start_read().
+	 * We don't really care about the correctness of that early check, but
+	 * we should use WRITE_ONCE() for cleanliness and to keep KCSAN happy.
+	 */
+	WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq);
 	up_write(&vma->vm_lock->lock);
 }
 
@@ -702,7 +719,7 @@ static inline bool vma_try_start_write(struct vm_area_struct *vma)
 	if (!down_write_trylock(&vma->vm_lock->lock))
 		return false;
 
-	vma->vm_lock_seq = mm_lock_seq;
+	WRITE_ONCE(vma->vm_lock_seq, mm_lock_seq);
 	up_write(&vma->vm_lock->lock);
 	return true;
 }
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index de10fc797c8e..5e74ce4a28cd 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -514,6 +514,20 @@ struct vm_area_struct {
 	};
 
 #ifdef CONFIG_PER_VMA_LOCK
+	/*
+	 * Can only be written (using WRITE_ONCE()) while holding both:
+	 *  - mmap_lock (in write mode)
+	 *  - vm_lock->lock (in write mode)
+	 * Can be read reliably while holding one of:
+	 *  - mmap_lock (in read or write mode)
+	 *  - vm_lock->lock (in read or write mode)
+	 * Can be read unreliably (using READ_ONCE()) for pessimistic bailout
+	 * while holding nothing (except RCU to keep the VMA struct allocated).
+	 *
+	 * This sequence counter is explicitly allowed to overflow; sequence
+	 * counter reuse can only lead to occasional unnecessary use of the
+	 * slowpath.
+	 */
 	int vm_lock_seq;
 	struct vma_lock *vm_lock;
 
@@ -679,6 +693,20 @@ struct mm_struct {
 					  * by mmlist_lock
 					  */
 #ifdef CONFIG_PER_VMA_LOCK
+		/*
+		 * This field has lock-like semantics, meaning it is sometimes
+		 * accessed with ACQUIRE/RELEASE semantics.
+		 * Roughly speaking, incrementing the sequence number is
+		 * equivalent to releasing locks on VMAs; reading the sequence
+		 * number can be part of taking a read lock on a VMA.
+		 *
+		 * Can be modified under write mmap_lock using RELEASE
+		 * semantics.
+		 * Can be read with no other protection when holding write
+		 * mmap_lock.
+		 * Can be read with ACQUIRE semantics if not holding write
+		 * mmap_lock.
+		 */
 		int mm_lock_seq;
 #endif
 
diff --git a/include/linux/mmap_lock.h b/include/linux/mmap_lock.h
index aab8f1b28d26..e05e167dbd16 100644
--- a/include/linux/mmap_lock.h
+++ b/include/linux/mmap_lock.h
@@ -76,8 +76,14 @@ static inline void mmap_assert_write_locked(struct mm_struct *mm)
 static inline void vma_end_write_all(struct mm_struct *mm)
 {
 	mmap_assert_write_locked(mm);
-	/* No races during update due to exclusive mmap_lock being held */
-	WRITE_ONCE(mm->mm_lock_seq, mm->mm_lock_seq + 1);
+	/*
+	 * Nobody can concurrently modify mm->mm_lock_seq due to exclusive
+	 * mmap_lock being held.
+	 * We need RELEASE semantics here to ensure that preceding stores into
+	 * the VMA take effect before we unlock it with this store.
+	 * Pairs with ACQUIRE semantics in vma_start_read().
+	 */
+	smp_store_release(&mm->mm_lock_seq, mm->mm_lock_seq + 1);
 }
 #else
 static inline void vma_end_write_all(struct mm_struct *mm) {}