Re: kernel BUG at fs/inode.c:LINE! (2)

[Al Viro]

On Fri, Aug 28, 2020 at 04:38:25PM +0100, Al Viro wrote:
> On Fri, Aug 28, 2020 at 06:18:17AM -0700, syzbot wrote:
> > Hello,
> > 
> > syzbot found the following issue on:
> > 
> > HEAD commit:    d012a719 Linux 5.9-rc2
> > git tree:       upstream
> > console output: https://syzkaller.appspot.com/x/log.txt?x=15aa650e900000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=891ca5711a9f1650
> > dashboard link: https://syzkaller.appspot.com/bug?extid=c92c93d1f1aaaacdb9db
> > compiler:       clang version 10.0.0 (https://github.com/llvm/llvm-project/ c2443155a0fb245c8f17f2c1c72b6ea391e86e81)
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=12ecb939900000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=140a19a9900000
> 
> Trying to reproduce...

OK, I think I see what's going on.  ep_loop_check_proc() runs into an already
doomed file that has already committed to getting killed (->f_count is already
at 0), but still hadn't gotten through its epitems removal (e.g. has its
eventpoll_release_file() sitting there trying to get epmutex).

Blindly bumping refcount here is worse than useless.  Try this, to verify that
this is what's going on; it's _not_ a proper fix, but it should at least tell
if we have something else going on.

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index e0decff22ae2..4590223b2de9 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -1995,7 +1995,7 @@ static int ep_loop_check_proc(void *priv, void *cookie, int call_nests)
 			 * during ep_insert().
 			 */
 			if (list_empty(&epi->ffd.file->f_tfile_llink)) {
-				get_file(epi->ffd.file);
+				if (get_file_rcu(epi->ffd.file))
 				list_add(&epi->ffd.file->f_tfile_llink,
 					 &tfile_check_list);
 			}