Re: [PATCH v2] n_tty: Protect tty->disc_data using refcount.

From: Tetsuo Handa
Date: Wed Jul 25 2018 - 08:35:33 EST


On 2018/07/25 17:06, Greg KH wrote:
> On Wed, Jul 25, 2018 at 12:22:16AM +0900, Tetsuo Handa wrote:
>> >From 118c64e86641a97d44dec39e313a95b12d9bc3b2 Mon Sep 17 00:00:00 2001
>> From: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>> Date: Wed, 25 Jul 2018 00:15:18 +0900
>> Subject: [PATCH v2] n_tty: Protect tty->disc_data using refcount.
>>
>> syzbot is reporting NULL pointer dereference at n_tty_set_termios() [1].
>> This is because ioctl(TIOCVHANGUP) versus ioctl(TCSETS) can race.
>>
>> Since we don't want to introduce new locking dependency, this patch
>> converts "struct n_tty_data *ldata = tty->disc_data;" in individual
>> function into a function argument which follows "struct tty *", and
>> holds tty->disc_data at each "struct tty_ldisc_ops" hook using refcount
>> in order to ensure that memory which contains "struct n_tty_data" will
>> not be released while processing individual function.
>>
>> [1] https://syzkaller.appspot.com/bug?id=1e850009fca0b64ce49dc16499bda4f7de0ab1a5
>>
>> Signed-off-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>> ---
>> drivers/tty/n_tty.c | 511 ++++++++++++++++++++++++++++++++--------------------
>> 1 file changed, 314 insertions(+), 197 deletions(-)
>
> What changed from v1? I haven't had the chance to review your first
> patch, sorry, it's still in my queue. I was hoping that someone else
> would help out with that as well :)
>

Just added a check in case I overlooked a subtle race condition that
n_tty_close() is _somehow_ called twice.

25,26c20,21
< 1 file changed, 308 insertions(+), 195 deletions(-)
> 1 file changed, 314 insertions(+), 197 deletions(-)
872,873c867,869
< struct n_tty_data *ldata = tty->disc_data;
> - struct n_tty_data *ldata = tty->disc_data;
> + struct n_tty_data *ldata = xchg(&tty->disc_data, NULL);
>
878,879c874,880
< tty->disc_data = NULL;
< + put_n_tty(ldata);
> - tty->disc_data = NULL;
> + /*
> + * The xchg() above and this NULL test are rather paranoid checks.
> + * Caller should not allow calling close() twice.
> + */
> + if (ldata)
> + put_n_tty(ldata);
1194c1195
< + if (!ldata) /* What value is most appropriate for this case? */
> + if (!ldata)

If n_tty_close() is _somehow_ called twice, v1 patch will trigger NULL
pointer dereference. If you are sure that n_tty_close() is never called
twice, you can ignore v2.