Hi:
On 2021/1/28 8:09, Andrew Morton wrote:

On Wed, 27 Jan 2021 04:33:49 -0500 Miaohe Lin [off-list ref] wrote:

quoted

For PMD-mapped page (usually THP), pvmw->pte is NULL. For PTE-mapped THP,
pvmw->pte is mapped. But for HugeTLB pages, pvmw->pte is not mapped and set
to the relevant page table entry. So in page_vma_mapped_walk_done(), we may
do pte_unmap() for HugeTLB pte which is not mapped. Fix this by checking
pvmw->page against PageHuge before trying to do pte_unmap().

What are the runtime consequences of this?  Is there a workload which
is known to trigger it?

Not yet. This should not be backported. My bad. Sorry about it.

IOW, how do we justify a -stable backport of this fix?

quoted

quoted

--- a/include/linux/rmap.h
+++ b/include/linux/rmap.h

@@ -213,7 +213,8 @@ struct page_vma_mapped_walk {
 
 static inline void page_vma_mapped_walk_done(struct page_vma_mapped_walk *pvmw)
 {
-	if (pvmw->pte)
+	/* HugeTLB pte is set to the relevant page table entry without pte_mapped. */
+	if (pvmw->pte && !PageHuge(pvmw->page))
 		pte_unmap(pvmw->pte);
 	if (pvmw->ptl)
 		spin_unlock(pvmw->ptl);

-- 
2.19.1

.