BUG? "Call fasync() functions without the BKL" is racy

[Oleg Nesterov <oleg@redhat.com>]

Hi.

Let's suppose we have the tasks T1, T2, T3 which share the same file,
all do sys_fcntl(file, F_SETFL) in parallel. file->f_flags == 0.

setfl(arg) does:

	if ((arg ^ filp->f_flags) & FASYNC)
// --- WINDOW_1 ---
		filp->f_op->fasync(fd, filp, (arg & FASYNC) != 0)
// --- WINDOW_2 ---
	filp->f_flags = arg;

T1 calls setfl(FASYNC), preempted in WINDOW_1.

T2 calls setfl(FASYNC), does all job and returns.

T3 calls setfl(0), sees ->f_flags & FASYNC, does ->fasync(on => 0),
preempted in WINDOW_2.

T1 resumes, does ->fasync(on => 1) again, update ->f_flags (it
already has FASYNC) and returns.

T3 resumes, and clears FASYNC from ->f_flags.



Now, this file was added into some "struct fasync_struct", but
->f_flags doesn't have FASYNC. This means __fput() will skip
->fasync(on => 0) and the next kill_fasync() can crash because
fa_file points to the freed/reused memory.

I think a238b790d5f99c7832f9b73ac8847025815b85f7 should be reverted.
Or do you see the better fix?

Unless I missed something of course.

Oleg.