On 1/30/19 3:32 AM, Johan Hovold wrote:

On Mon, Jan 28, 2019 at 02:29:22PM -0700, shuah wrote:

quoted

On 1/25/19 9:14 PM, Al Viro wrote:

quoted

On Fri, Jan 25, 2019 at 04:29:05PM -0700, Shuah Khan wrote:

quoted

tty_set_termios() has the following WARMN_ON which can be triggered with a
syscall to invoke TIOCGETD __NR_ioctl.

You meant TIOCSETD here, and in fact its the call which sets the uart
protocol that triggers the warning.

Right. It is a TIOCSETD.

quoted

quoted

quoted

WARN_ON(tty->driver->type == TTY_DRIVER_TYPE_PTY &&
                  tty->driver->subtype == PTY_TYPE_MASTER);
Reference: https://syzkaller.appspot.com/bug?id=2410d22f1d8e5984217329dd0884b01d99e3e48d

A simple change would have been to print error message instead of WARN_ON.
However, the callers assume that tty_set_termios() always returns 0 and
don't check return value. The complete solution is fixing all the callers
to check error and bail out to fix the WARN_ON.

This fix changes tty_set_termios() to return error and all the callers
to check error and bail out. The reproducer is used to reproduce the
problem and verify the fix.

quoted

--- a/drivers/bluetooth/hci_ldisc.c
+++ b/drivers/bluetooth/hci_ldisc.c

@@ -321,6 +321,8 @@ void hci_uart_set_flow_control(struct hci_uart *hu, bool enable)
   		status = tty_set_termios(tty, &ktermios);
   		BT_DBG("Disabling hardware flow control: %s",
   		       status ? "failed" : "success");
+		if (status)
+			return;

Can that ldisc end up set on pty master?  And does it make any sense there?

The initial objective of the patch is to prevent the WARN_ON by making
the change to return error instead of WARN_ON. However, without changes
to places that don't check the return and keep making progress, there
will be secondary problems.

Without this change to return here, instead of WARN_ON, it will fail
with the following NULL pointer dereference at the next thing
hci_uart_set_flow_control() attempts.

status = tty->driver->ops->tiocmget(tty);

kernel: [10140.649783] BUG: unable to handle kernel NULL pointer

That's a separate issue, which is being fixed:

	https://lkml.kernel.org/r/20190130095938.GP3691@localhost

Ah good to know.

quoted

quoted

IOW, I don't believe that this patch makes any sense.  If anything,
we need to prevent unconditional tty_set_termios() on the path
that *does* lead to calling it for pty.

I don't think preventing unconditional tty_set_termios() is enough to
prevent secondary problems such as the one above.

For example, the following call chain leads to the WARN_ON that was
reported. Even if void hci_uart_set_baudrate() prevents the very first
tty_set_termios() call, its caller hci_uart_setup() continues with
more tty setup. It goes ahead to call driver setup callback. The
driver callback goes on to do more setup calling tty_set_termios().

WARN_ON call path:
   hci_uart_set_baudrate+0x1cc/0x250 drivers/bluetooth/hci_ldisc.c:378
   hci_uart_setup+0xa2/0x490 drivers/bluetooth/hci_ldisc.c:401
   hci_dev_do_open+0x6b1/0x1920 net/bluetooth/hci_core.c:1423

Once this WARN_ON is changed to return error, the following
happens, when hci_uart_setup() does driver setup callback.

kernel: [10140.649836]  mrvl_setup+0x17/0x80 [hci_uart]
kernel: [10140.649840]  hci_uart_setup+0x56/0x160 [hci_uart]
kernel: [10140.649850]  hci_dev_do_open+0xe6/0x630 [bluetooth]
kernel: [10140.649860]  hci_power_on+0x52/0x220 [bluetooth]

I think continuing to catch the invalid condition in tty_set_termios()
and preventing progress by checking return value is a straight forward
change to avoid secondary problems, and it might be difficult to catch
all the cases where it could fail.

I agree with Al that this change doesn't make much sense. The WARN_ON
is there to catch any bugs leading to the termios being changed for a
master side pty. Those should bugs should be fixed, and not worked
around in order to silence a WARN_ON.

The problem started with 7721383f4199 ("Bluetooth: hci_uart: Support
operational speed during setup") which introduced a new way for how
tty_set_termios() could end up being called for a master pty.

Ah. Thanks for the context.

As Al hinted at, setting these ldiscs for a master pty really makes no
sense and perhaps that is what we should prevent unless simply making
sure they do not call tty_set_termios() is sufficient for the time
being.

I will take a look to see if not calling tty_set_termios() is enough.

Finally, note that serdev never operates on a pty, and that this is only
an issue for (the three) line disciplines.

Thanks for the detailed message explaining the evolution. Now it makes
sense.

-- Shuah