[PATCH mmotm/next] fanotify: fix corruption preventing startup

[PATCH mmotm/next] fanotify: fix corruption preventing startup

[Hugh Dickins]

Two of my machines couldn't boot mmotm with fanotify enabled:
0x40000020 was being or'ed into a group's notification_waitq next
pointer, with sad results.  It comes from an over-simplification in
fanotify_merge(): test_event isn't NULL when the loop finds no merge.

Signed-off-by: Hugh Dickins <hughd@google.com>
---
Fixes fsnotify-do-not-share-events-between-notification-groups.patch

 fs/notify/fanotify/fanotify.c |    9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

--- mmotm/fs/notify/fanotify/fanotify.c	2014-01-10 18:25:01.700448941 -0800
+++ linux/fs/notify/fanotify/fanotify.c	2014-01-11 22:15:31.580961010 -0800
@@ -37,16 +37,19 @@ static bool should_merge(struct fsnotify
 static struct fsnotify_event *fanotify_merge(struct list_head *list,
 					     struct fsnotify_event *event)
 {
-	struct fsnotify_event *test_event = NULL;
+	struct fsnotify_event *test_event;
+	bool do_merge = false;
 
 	pr_debug("%s: list=%p event=%p\n", __func__, list, event);
 
 	list_for_each_entry_reverse(test_event, list, list) {
-		if (should_merge(test_event, event))
+		if (should_merge(test_event, event)) {
+			do_merge = true;
 			break;
+		}
 	}
 
-	if (!test_event)
+	if (!do_merge)
 		return NULL;
 
 	test_event->mask |= event->mask;

Re: [PATCH mmotm/next] fanotify: fix corruption preventing startup

[Jan Kara]

On Sat 11-01-14 22:54:15, Hugh Dickins wrote:
> Two of my machines couldn't boot mmotm with fanotify enabled:
> 0x40000020 was being or'ed into a group's notification_waitq next
> pointer, with sad results.  It comes from an over-simplification in
> fanotify_merge(): test_event isn't NULL when the loop finds no merge.
  That's a really dumb bug. Thanks for the fix! It looks good, you can add:
Reviewed-by: Jan Kara <jack@suse.cz>

  I was wondering why I didn't see this problem during my testing.  I was
always running my tests with various debug options enabled (in a hope to
catch more problems ;), in particular with spinlock debugging enabled which
shifted the waitqueue pointers further and the corruption landed in
spinlock debugging fields and never got noticed... Ew :-|

								Honza
> 
> Signed-off-by: Hugh Dickins <hughd@google.com>
> ---
> Fixes fsnotify-do-not-share-events-between-notification-groups.patch
> 
>  fs/notify/fanotify/fanotify.c |    9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> --- mmotm/fs/notify/fanotify/fanotify.c	2014-01-10 18:25:01.700448941 -0800
> +++ linux/fs/notify/fanotify/fanotify.c	2014-01-11 22:15:31.580961010 -0800
> @@ -37,16 +37,19 @@ static bool should_merge(struct fsnotify
>  static struct fsnotify_event *fanotify_merge(struct list_head *list,
>  					     struct fsnotify_event *event)
>  {
> -	struct fsnotify_event *test_event = NULL;
> +	struct fsnotify_event *test_event;
> +	bool do_merge = false;
>  
>  	pr_debug("%s: list=%p event=%p\n", __func__, list, event);
>  
>  	list_for_each_entry_reverse(test_event, list, list) {
> -		if (should_merge(test_event, event))
> +		if (should_merge(test_event, event)) {
> +			do_merge = true;
>  			break;
> +		}
>  	}
>  
> -	if (!test_event)
> +	if (!do_merge)
>  		return NULL;
>  
>  	test_event->mask |= event->mask;
-- 
Jan Kara <jack@suse.cz>
SUSE Labs, CR