mm: fix race condition in MADV_FREE

From: Shaohua Li <[email protected]>

MADV_FREE clears pte dirty bit and then marks the page lazyfree (clear
SwapBacked). There is no lock to prevent the page is added to swap cache
between these two steps by page reclaim. If the page is added to swap cache, marking the page lazyfree will confuse page fault if the page is
reclaimed and refault.

Reported-and-tested-y: Artem Savkov <[email protected]>
Fix: 802a3a92ad7a(mm: reclaim MADV_FREE pages)

Signed-off-by Shaohua Li <[email protected]>

Cc [email protected]

Cc Johannes Weiner <[email protected]>

Cc Michal Hocko <[email protected]>

Cc Hillf Danton <[email protected]>

Cc Minchan Kim <[email protected]>

Cc Hugh Dickins <[email protected]>

Cc Rik van Riel <[email protected]>

Cc Mel Gorman <[email protected]>

Cc Andrew Morton <[email protected]>

From: Shaohua Li <[email protected]>

MADV_FREE clears pte dirty bit and then marks the page lazyfree (clear SwapBacked). There is no lock to prevent the page is added to swap cache between these two steps by page reclaim. If page reclaim finds such page, it will simply add the page to swap cache without pageout the page to swap because the page is marked as clean. Next time, page fault will read data from the swap slot which doesn't have the original data, so we have a data corruption. To fix issue, we mark the page dirty and pageout the page.

However, we shouldn't dirty all pages which is clean and in swap cache. swapin page is swap cache and clean too. So we only dirty page which is added into swap cache in page reclaim, which shouldn't be swapin page. Normal anonymous pages should be dirty already.

Reported-and-tested-y: Artem Savkov <[email protected]>
Fix: 802a3a92ad7a(mm: reclaim MADV_FREE pages)

Signed-off-by Shaohua Li <[email protected]>

Cc Johannes Weiner <[email protected]>

Cc Michal Hocko <[email protected]>

Cc Hillf Danton <[email protected]>

Cc Minchan Kim <[email protected]>

Cc Hugh Dickins <[email protected]>

Cc Rik van Riel <[email protected]>

Cc Mel Gorman <[email protected]>

Cc Andrew Morton <[email protected]>

---
 mm/vmscan.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

mm/vmscan.c +12 -0

index d811c81..820ee8d 100644
@@ -980,6 +980,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 		int may_enter_fs;
 		enum page_references references = PAGEREF_RECLAIM_CLEAN;
 		bool dirty, writeback;
+		bool new_swap_page = false;
 
 		cond_resched();
 

@@ -1165,6 +1166,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 
 				/* Adding to swap updated mapping */
 				mapping = page_mapping(page);
+				new_swap_page = true;
 			}
 		} else if (unlikely(PageTransHuge(page))) {
 			/* Split file THP */

@@ -1185,6 +1187,16 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 				nr_unmap_fail++;
 				goto activate_locked;
 			}
+
+			/*
+			 * MADV_FREE clear pte dirty bit, but not yet clear
+			 * SwapBacked for a page. We can't directly free the
+			 * page because we already set swap entry in pte. The
+			 * check guarantees this is such page and not a clean
+			 * swapin page
+			 */
+			if (!PageDirty(page) && new_swap_page)
+				set_page_dirty(page);
 		}
 
 		if (PageDirty(page)) {

Signature · 1 line

2.9.5

Quoted 71 lines · Shaohua Li

From: Shaohua Li <[email protected]>

MADV_FREE clears pte dirty bit and then marks the page lazyfree (clear SwapBacked). There is no lock to prevent the page is added to swap cache between these two steps by page reclaim. If page reclaim finds such page, it will simply add the page to swap cache without pageout the page to swap because the page is marked as clean. Next time, page fault will read data from the swap slot which doesn't have the original data, so we have a data corruption. To fix issue, we mark the page dirty and pageout the page.

However, we shouldn't dirty all pages which is clean and in swap cache. swapin page is swap cache and clean too. So we only dirty page which is added into swap cache in page reclaim, which shouldn't be swapin page. Normal anonymous pages should be dirty already.

Reported-and-tested-y: Artem Savkov <[email protected]>
Fix: 802a3a92ad7a(mm: reclaim MADV_FREE pages)

Signed-off-by Shaohua Li <[email protected]>

Cc Johannes Weiner <[email protected]>

Cc Michal Hocko <[email protected]>

Cc Hillf Danton <[email protected]>

Cc Minchan Kim <[email protected]>

Cc Hugh Dickins <[email protected]>

Cc Rik van Riel <[email protected]>

Cc Mel Gorman <[email protected]>

Cc Andrew Morton <[email protected]>

---
 mm/vmscan.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

mm/vmscan.c +12 -0

index d811c81..820ee8d 100644
@@ -980,6 +980,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 		int may_enter_fs;
 		enum page_references references = PAGEREF_RECLAIM_CLEAN;
 		bool dirty, writeback;
+		bool new_swap_page = false;
 
 		cond_resched();
 

@@ -1165,6 +1166,7 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 
 				/* Adding to swap updated mapping */
 				mapping = page_mapping(page);
+				new_swap_page = true;
 			}
 		} else if (unlikely(PageTransHuge(page))) {
 			/* Split file THP */

@@ -1185,6 +1187,16 @@ static unsigned long shrink_page_list(struct list_head *page_list,
 				nr_unmap_fail++;
 				goto activate_locked;
 			}
+
+			/*
+			 * MADV_FREE clear pte dirty bit, but not yet clear
+			 * SwapBacked for a page. We can't directly free the
+			 * page because we already set swap entry in pte. The
+			 * check guarantees this is such page and not a clean
+			 * swapin page
+			 */
+			if (!PageDirty(page) && new_swap_page)
+				set_page_dirty(page);
 		}
 
 		if (PageDirty(page)) {

Signature · 1 line

2.9.5

Couldn't we simple roll back to the logic before MADV_FREE's birth?

mm/swap_state.c +2 -1

index 71ce2d1ccbf7..548c19b5f78e 100644
@@ -231,7 +231,7 @@ int add_to_swap(struct page *page)
 	 * deadlock in the swap out path.
 	 */
 	/*
-	 * Add it to the swap cache.
+	 * Add it to the swap cache and mark it dirty
 	 */
 	err = add_to_swap_cache(page, entry,
 			__GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);

@@ -243,6 +243,7 @@ int add_to_swap(struct page *page)
 		 */
 		goto fail;
 
+	SetPageDirty(page);
 	return 1;
 
 fail:

To me, it would be more simple/readable rather than introducing a new branch in complicated shrink_page_list.

And I don't see why we cannot merge [1/2] and [2/2].