Re: [PATCH] Bluetooth: reorganize ioctls from hci_sock_bound_ioctl()

On 2021/07/22 3:17, Luiz Augusto von Dentz wrote:

I think it would have been cleaner if we have dedicated functions for
each command like I did in my patch:

https://patchwork.kernel.org/project/bluetooth/patch/20210717000731.3836303-1-luiz.dentz@gmail.com/

But your patch was proven to be unsafe. There is a use-after-unregister
race window which would require at least 1000 lines of modification and
a lot of careful review if we try to manage without my patch.
Such all-in-one-step change is too much for "sleep in atomic context"
regression fix which is preventing syzbot from testing Bluetooth module
and is preventing Linux distributors from fixing CVE-2021-3573.

As far as I can see, it is lock_sock() (not bh_lock_sock_nested() in your
patch) that is needed for protecting

	sk->sk_err = EPIPE;
	sk->sk_state = BT_OPEN;
	sk->sk_state_change(sk);

in hci_sock_dev_event(HCI_DEV_UNREG) from concurrent modification

	lock_sock(sk);

	if (sk->sk_state == BT_BOUND) {
		err = -EALREADY;
		goto done;
	}

	(...snipped...)

-		hci_pi(sk)->hdev = hdev;
+		if (hdev) {
+			hci_pi(sk)->dev = hdev->id;
+			hci_dev_put(hdev);
+		}

	(...snipped...)
	/* Race window is here. */
	(...snipped...)

	sk->sk_state = BT_BOUND;
done:
	release_sock(sk);

in hci_sock_bind().

That way it is simpler to protect the likes of
copy_from_user/copy_to_user, etc, even if we have to some checks
duplicated on each function we can have a helper function to checks
the flags, etc.

My patch calls copy_from_user()/copy_to_user() without lock_sock()
which works nicely with "[PATCH v3] Bluetooth: call lock_sock() outside
of spinlock section". I'd like to backport "[PATCH v2] Bluetooth:
reorganize ioctls from hci_sock_bound_ioctl()" together.