On Tue 23-04-19 21:05:22, Amir Goldstein wrote: > On Tue, Apr 23, 2019 at 7:27 PM Jan Kara <jack@xxxxxxx> wrote: > > > > On Fri 19-04-19 12:33:02, Amir Goldstein wrote: > > > On Thu, Apr 18, 2019 at 8:14 PM syzbot > > > <syzbot+15927486a4f1bfcbaf91@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote: > > > > > > > > syzbot has bisected this bug to: > > > > > > > > commit 77115225acc67d9ac4b15f04dd138006b9cd1ef2 > > > > Author: Amir Goldstein <amir73il@xxxxxxxxx> > > > > Date: Thu Jan 10 17:04:37 2019 +0000 > > > > > > > > fanotify: cache fsid in fsnotify_mark_connector > > > > > > > > bisection log: https://syzkaller.appspot.com/x/bisect.txt?x=1627632d200000 > > > > start commit: 3f018f4a Add linux-next specific files for 20190418 > > > > git tree: linux-next > > > > final crash: https://syzkaller.appspot.com/x/report.txt?x=1527632d200000 > > > > console output: https://syzkaller.appspot.com/x/log.txt?x=1127632d200000 > > > > kernel config: https://syzkaller.appspot.com/x/.config?x=faa7bdc352fc157e > > > > dashboard link: https://syzkaller.appspot.com/bug?extid=15927486a4f1bfcbaf91 > > > > syz repro: https://syzkaller.appspot.com/x/repro.syz?x=155543d3200000 > > > > > > > > Reported-by: syzbot+15927486a4f1bfcbaf91@xxxxxxxxxxxxxxxxxxxxxxxxx > > > > Fixes: 77115225acc6 ("fanotify: cache fsid in fsnotify_mark_connector") > > > > > > > > [...] > > > > > I'd prefer if we could make only fully initialized marks visible on > > connector's list. Just to make things simple in the fast path of handling > > events. So I'd just set mark->connector before adding mark to connector's > > list and having smp_wmb() there to force ordering in > > fsnotify_add_mark_list(). And we should use READ_ONCE() as a counter-part > > to this in fanotify_get_fsid(). It is somewhat unfortunate that there are > > three different ways how fsnotify_add_mark_list() can add mark to a list so > > we may need to either repeat this in three different places or just use > > some macro magic like: > > > > #define fanotify_attach_obj_list(where, mark, conn, head) \ > > do { \ > > mark->connector = conn; \ > > smp_wmb(); \ > > hlist_add_##where##_rcu(&mark->obj_list, head); \ > > } while (0) > > > > And I would not really worry about fsnotify_put_mark() - if it ever sees > > mark->connector set, mark really is on the connector's list and > > fsnotify_put_mark() does the right thing (i.e. locks connector->lock if > > needed). > > > > Sounds good to me. > > I was worried about fsnotify_put_mark() NOT seeing mark->connector set > even though it is on the list (add mark; get lockless ref; remove > mark; put lockless ref). > Its subtle, not sure if possible. Not possible for the cases it matters. fsnotify_put_mark() just decrements the refcount. If someone is adding mark to the list, he is by definition holding one reference so racing fsnotify_put_mark() cannot be dropping the last one and so the result of the check doesn't really matter. > I am assuming you will be sending the above patch to syzbot for testing. > If you want me do help with anything please let me know. OK, I'll write it and send it. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
