In the caller of map_pte(), we may modify the pvmw->pte after acquiring
the pvmw->ptl, so convert it to using pte_offset_map_rw_nolock(). At
this time, the pte_same() check is not performed after the pvmw->ptl held,
so we should get pmdval and do pmd_same() check to ensure the stability of
pvmw->pmd.

Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
---
 mm/page_vma_mapped.c | 24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)

diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
index ae5cc42aa2087..ab1671e71cb2d 100644
--- a/mm/page_vma_mapped.c
+++ b/mm/page_vma_mapped.c
@@ -13,7 +13,8 @@ static inline bool not_found(struct page_vma_mapped_walk *pvmw)
 	return false;
 }
 
-static bool map_pte(struct page_vma_mapped_walk *pvmw, spinlock_t **ptlp)
+static bool map_pte(struct page_vma_mapped_walk *pvmw, pmd_t *pmdvalp,
+		    spinlock_t **ptlp)
 {
 	pte_t ptent;
 
@@ -25,6 +26,7 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, spinlock_t **ptlp)
 		return !!pvmw->pte;
 	}
 
+again:
 	/*
 	 * It is important to return the ptl corresponding to pte,
 	 * in case *pvmw->pmd changes underneath us; so we need to
@@ -32,8 +34,8 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, spinlock_t **ptlp)
 	 * proceeds to loop over next ptes, and finds a match later.
 	 * Though, in most cases, page lock already protects this.
 	 */
-	pvmw->pte = pte_offset_map_nolock(pvmw->vma->vm_mm, pvmw->pmd,
-					  pvmw->address, ptlp);
+	pvmw->pte = pte_offset_map_rw_nolock(pvmw->vma->vm_mm, pvmw->pmd,
+					     pvmw->address, pmdvalp, ptlp);
 	if (!pvmw->pte)
 		return false;
 
@@ -67,8 +69,13 @@ static bool map_pte(struct page_vma_mapped_walk *pvmw, spinlock_t **ptlp)
 	} else if (!pte_present(ptent)) {
 		return false;
 	}
+	spin_lock(*ptlp);
+	if (unlikely(!pmd_same(*pmdvalp, pmdp_get_lockless(pvmw->pmd)))) {
+		pte_unmap_unlock(pvmw->pte, *ptlp);
+		goto again;
+	}
 	pvmw->ptl = *ptlp;
-	spin_lock(pvmw->ptl);
+
 	return true;
 }
 
@@ -278,7 +285,7 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 			step_forward(pvmw, PMD_SIZE);
 			continue;
 		}
-		if (!map_pte(pvmw, &ptl)) {
+		if (!map_pte(pvmw, &pmde, &ptl)) {
 			if (!pvmw->pte)
 				goto restart;
 			goto next_pte;
@@ -305,8 +312,13 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
 		} while (pte_none(ptep_get(pvmw->pte)));
 
 		if (!pvmw->ptl) {
+			spin_lock(ptl);
+			if (unlikely(!pmd_same(pmde, pmdp_get_lockless(pvmw->pmd)))) {
+				pte_unmap_unlock(pvmw->pte, ptl);
+				pvmw->pte = NULL;
+				goto restart;
+			}
 			pvmw->ptl = ptl;
-			spin_lock(pvmw->ptl);
 		}
 		goto this_pte;
 	} while (pvmw->address < end);
-- 
2.20.1

> On Sep 26, 2024, at 14:46, Qi Zheng <zhengqi.arch@bytedance.com> wrote:
> 
> In the caller of map_pte(), we may modify the pvmw->pte after acquiring
> the pvmw->ptl, so convert it to using pte_offset_map_rw_nolock(). At
> this time, the pte_same() check is not performed after the pvmw->ptl held,
> so we should get pmdval and do pmd_same() check to ensure the stability of
> pvmw->pmd.
> 
> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>

Reviewed-by: Muchun Song <muchun.song@linux.dev>

Thanks.