While debugging some crashes related to virtio-balloon deflation that
happened under the old balloon migration code, I stumbled over a possible
race I think.
What we experienced:
drivers/virtio/virtio_balloon.c:release_pages_balloon():
- WARNING: CPU: 13 PID: 6586 at lib/list_debug.c:59 __list_del_entry+0xa1/0xd0
- list_del corruption. prev->next should be ffffe253961090a0, but was dead000000000100
Turns out after having added the page to a local list when dequeuing,
the page would suddenly be moved to an LRU list before we would free it
via the local list, corrupting both lists. So a page we own and that is
!LRU was moved to an LRU list.
For us, this was triggered by backporting 195a8c43e93d8 ("virtio-balloon:
deflate via a page list") onto old balloon compaction code, but I think
this only made the BUG become visible and the race still exists in
new !LRU migraton code.
My theory:
In __unmap_and_move(), we lock the old and newpage and perform the
migration. In case of vitio-balloon, the new page will become
movable, the old page will no longer be movable.
However, after unlocking newpage, I think there is nothing stopping
the newpage from getting dequeued and freed by virtio-balloon. This
will result in the newpage
1. No longer having PageMovable()
2. Getting moved to the local list before finally freeing it (using
page->lru)
Back in the migration thread in __unmap_and_move(), we would after
unlocking the newpage suddenly no longer have PageMovable(newpage) and
will therefore call putback_lru_page(newpage), modifying page->lru
although that list is still in use by virtio-balloon.
To summarize, we have a race between migrating the newpage and checking
for PageMovable(newpage). Instead of checking __PageMovable(newpage), we
can simply rely on is_lru. Because if the old page was not an LRU page, the
new page also shouldn't be. If there is a problem with that, we could check
for __PageMovable(newpage) before unlocking the new page.
As I am not yet sure if this race actually exists, especially also
upstream, I am sending this as RFC.
Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
Cc: "Kirill A. Shutemov" <kirill.shutemov@xxxxxxxxxxxxxxx>
Cc: Michal Hocko <mhocko@xxxxxxxx>
Cc: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx>
Cc: Jan Kara <jack@xxxxxxx>
Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
Cc: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Cc: Vratislav Bendel <vbendel@xxxxxxxxxx>
Cc: Rafael Aquini <aquini@xxxxxxxxxx>
Reported-by: Vratislav Bendel <vbendel@xxxxxxxxxx>
Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
---
mm/migrate.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/mm/migrate.c b/mm/migrate.c
index 4512afab46ac..31e002270b05 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1135,10 +1135,12 @@ static int __unmap_and_move(struct page *page, struct page *newpage,
* If migration is successful, decrease refcount of the newpage
* which will not free the page because new page owner increased
* refcounter. As well, if it is LRU page, add the page to LRU
- * list in here.
+ * list in here. Don't rely on PageMovable(newpage), as that could
+ * already have changed after unlocking newpage (e.g.
+ * virtio-balloon deflation).
*/
if (rc == MIGRATEPAGE_SUCCESS) {
- if (unlikely(__PageMovable(newpage)))
+ if (unlikely(!is_lru))
put_page(newpage);
else
putback_lru_page(newpage);
--
2.17.2
