Re: [GIT PULL] TTY/Serial driver fixes for 4.11-rc4

[Dmitry Vyukov]

On Fri, Apr 14, 2017 at 2:30 PM, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Apr 14, 2017 at 11:41:26AM +0200, Vegard Nossum wrote:
>> On 13 April 2017 at 20:34, Greg KH <gregkh@xxxxxxxxxxxxxxxxxxx> wrote:
>> > On Thu, Apr 13, 2017 at 09:07:40AM -0700, Linus Torvalds wrote:
>> >> On Thu, Apr 13, 2017 at 3:50 AM, Vegard Nossum <vegard.nossum@xxxxxxxxx> wrote:
>> >> >
>> >> > I've bisected a syzkaller crash down to this commit
>> >> > (5362544bebe85071188dd9e479b5a5040841c895). The crash is:
>> >> >
>> >> > [   25.137552] BUG: unable to handle kernel paging request at 0000000000002280
>> >> > [   25.137579] IP: mutex_lock_interruptible+0xb/0x30
>> >>
>> >> It would seem to be the
>> >>
>> >>                 if (mutex_lock_interruptible(&ldata->atomic_read_lock))
>> >>
>> >> call in n_tty_read(), the offset is about right for a NULL 'ldata'
>> >> pointer (it's a big structure, it has a couple of character buffers of
>> >> size N_TTY_BUF_SIZE).
>> >>
>> >> I don't see the obvious fix, so I suspect at this point we should just
>> >> revert, as that commit seems to introduce worse problems that it is
>> >> supposed to fix. Greg?
>> >
>> > Unless Dmitry has a better idea, I will just revert it and send you the
>> > pull request in a day or so.
>>
>> I don't think we need to rush a revert, I'd hope there's a way to fix
>> it properly.
>
> For this late in the release cycle, for something as complex as tty
> ldisc handling, for an issue that has been present for over a decade,
> the safest thing right now is to go back to the old well-known code by
> applying a revert :)
>
>> So the original problem is that the vmalloc() in n_tty_open() can
>> fail, and that will panic in tty_set_ldisc()/tty_ldisc_restore()
>> because of its unwillingness to proceed if the tty doesn't have an
>> ldisc.
>>
>> Dmitry fixed this by allowing tty->ldisc == NULL in the case of memory
>> allocation failure as we can see from the comment in tty_set_ldisc().
>>
>> Unfortunately, it would appear that some other bits of code do not
>> like tty->ldisc == NULL (other than the crash in this thread, I saw
>> 2-3 similar crashes in other functions, e.g. poll()). I see two
>> possibilities:
>>
>> 1) make other code handle tty->ldisc == NULL.
>>
>> 2) don't close/free the old ldisc until the new one has been
>> successfully created/initialised/opened/attached to the tty, and
>> return an error to userspace if changing it failed.
>>
>> I'm leaning towards #2 as the more obviously correct fix, it makes
>> tty_set_ldisc() transactional, the fix seems limited in scope to
>> tty_set_ldisc() itself, and we don't need to make every other bit of
>> code that uses tty->ldisc handle the NULL case.
>
> That sounds reasonable to me, care to work on a patch for this?

Vegard, do you know how to do this?
That was first thing that I tried, but I did not manage to make it
work. disc is tied to tty, so it's not that one can create a fully
initialized disc on the side and then simply swap pointers. Looking at
the code now, there is at least TTY_LDISC_OPEN bit in tty. But as far
as I remember there were more fundamental problems. Or maybe I just
did not try too hard.