mm: fix swapoff hang after page migration and fork
authorHugh Dickins <hughd@google.com>
Tue, 2 Dec 2014 23:59:39 +0000 (15:59 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 3 Dec 2014 17:36:03 +0000 (09:36 -0800)
I've been seeing swapoff hangs in recent testing: it's cycling around
trying unsuccessfully to find an mm for some remaining pages of swap.

I have been exercising swap and page migration more heavily recently,
and now notice a long-standing error in copy_one_pte(): it's trying to
add dst_mm to swapoff's mmlist when it finds a swap entry, but is doing
so even when it's a migration entry or an hwpoison entry.

Which wouldn't matter much, except it adds dst_mm next to src_mm,
assuming src_mm is already on the mmlist: which may not be so.  Then if
pages are later swapped out from dst_mm, swapoff won't be able to find
where to replace them.

There's already a !non_swap_entry() test for stats: move that up before
the swap_duplicate() and the addition to mmlist.

Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Kelley Nielsen <kelleynnn@gmail.com>
Cc: <stable@vger.kernel.org> [2.6.18+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/memory.c

index 3e50383..d5f2ae9 100644 (file)
@@ -815,20 +815,20 @@ copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
                if (!pte_file(pte)) {
                        swp_entry_t entry = pte_to_swp_entry(pte);
 
-                       if (swap_duplicate(entry) < 0)
-                               return entry.val;
-
-                       /* make sure dst_mm is on swapoff's mmlist. */
-                       if (unlikely(list_empty(&dst_mm->mmlist))) {
-                               spin_lock(&mmlist_lock);
-                               if (list_empty(&dst_mm->mmlist))
-                                       list_add(&dst_mm->mmlist,
-                                                &src_mm->mmlist);
-                               spin_unlock(&mmlist_lock);
-                       }
-                       if (likely(!non_swap_entry(entry)))
+                       if (likely(!non_swap_entry(entry))) {
+                               if (swap_duplicate(entry) < 0)
+                                       return entry.val;
+
+                               /* make sure dst_mm is on swapoff's mmlist. */
+                               if (unlikely(list_empty(&dst_mm->mmlist))) {
+                                       spin_lock(&mmlist_lock);
+                                       if (list_empty(&dst_mm->mmlist))
+                                               list_add(&dst_mm->mmlist,
+                                                        &src_mm->mmlist);
+                                       spin_unlock(&mmlist_lock);
+                               }
                                rss[MM_SWAPENTS]++;
-                       else if (is_migration_entry(entry)) {
+                       else if (is_migration_entry(entry)) {
                                page = migration_entry_to_page(entry);
 
                                if (PageAnon(page))