On Wed 24-08-16 11:55:39, Miklos Szeredi wrote: > List corruption was reported with a fanotify stress test. > > The bug turned out to be due to fsnotify_remove_event() being called on an > event on the fanotify_data.access_list and protected by > fanotify_data.access_lock instead of notification_mutex. This resulted in > list_del_init() being run concurrently on the same list entry. > > This was introduced by commit 09e5f14e57c7 ("fanotify: on group destroy > allow all waiters to bypass permission check") which made > fanotify_get_response() flush out events when bypass_perm was set. The > flush doesn't normally happen, since the wake_up() is called after the > access_list was cleaned in fsnotify_release(). But the two are not > synchronized, the fanotify_get_response() could still be processing a > previous wakeup by the time bypass_perm is true. This was seen in the > crashdumps in the report.
Thanks for the analysis and the patch! I agree there's a bug you describe, I just somewhat disagree with the solution. > This bug can be solved multiple ways, maybe the simplest is moving the > bypass_perm setting after the list has been processed. > > In theory there's also a memory ordering problem here. atomic_inc() in > itself doesn't imply a memory barrier, and spin_unlock() is a semi > permeable barrier, so we need an explicit memory barrier so that the > condition is precieved after the list is cleared. Well, the culprit of the problem seems to be that fanotify_get_response() does not use proper function to remove the event from the list in all the cases. The event can be in two states: 1) User didn't yet read the even at the time fanotify_release() is closed - note that a group can still receive new events queued while fanotify_release() is running until fsnotify_destroy_group() kills all the marks and that is the main reason why we have that bypass_perm thing to avoid blocking new processes in fanotify_get_response() because they could otherwise miss a wakeup and hang there indefinitely. In this state calling fsnotify_remove_event() is correct and that's what commit 5838d4442bd5 (fanotify: fix double free of pending permission events) had in mind. 2) User has read the event (thus the permission even was moved to access_list) but didn't write the response yet. In this state calling fsnotify_remove_event() from fanotify_get_response() is just wrong - as you noted it uses the wrong lock to protect the list removal but it also wrongly updates group->q_len. In this situation we should be calling dequeue_event(). The problem is that there's no easy way to distinguishing these two cases in fanotify_get_response(). We could flag this somehow inside the event structure but I think it's cleaner to remove all permission events in fanotify_release() (the same way we already handle the permission events on access_list). I'll send a patch shortly. > Similarly we need barriers for the case when event->response is set > (i.e. non zero): fsnotify_destroy_event() might destroy the event while > it's still on the access_list, since nothing guarantees that the storing > the response value in event->response will be preceived after the list > manipulation. So add the necessary barriers there as well. > > PS not sure why bypass_perm is an atomic_t, it could just as well be a > boolean flag. Agreed. I was just never bothered enough to fix this. > PPS all this subtlety could be removed if the waitq was per-event, which > would also allow better performance. I'm not sure how much this problem would be helped by a per-event waitqueue. Also it would add a significant memory cost to permission events and I prefer them small as there can be a lot of them queued... Honza -- Jan Kara <j...@suse.com> SUSE Labs, CR