Re: [BUG] 4.11.0-rc3 xterm hung in D state on exit, wchan is tty_release_struct

[Rob Herring]

+Peter

On Thu, Mar 23, 2017 at 4:22 PM,  <lkml@xxxxxxxxxxx> wrote:
> On Thu, Mar 23, 2017 at 11:57:22AM -0500, Rob Herring wrote:
>> On Thu, Mar 23, 2017 at 08:46:03AM -0500, Rob Herring wrote:
>> > On Thu, Mar 23, 2017 at 12:30:18AM -0700, lkml@xxxxxxxxxxx wrote:
>> > > On Wed, Mar 22, 2017 at 11:44:18PM -0700, lkml@xxxxxxxxxxx wrote:
>> > > > On Wed, Mar 22, 2017 at 07:08:46PM -0700, lkml@xxxxxxxxxxx wrote:
>> > > > > Hello list,
>> > > > >
>> > > > > After approximately one day day of running 4.11.0-rc3 with 7e54d9d reverted to
>> > > > > enable regular use, this happened upon destroying an xterm:
>> > > > >
>>
>> [...]
>>
>> > > >
>> > > > Added Rob Herring, author of c3485ee to CC list.
>> > > >
>> > >
>> > > I suspect this part was a mistake:
>> > >
>> > >  -       tty = READ_ONCE(port->itty);
>> > >  -       if (tty == NULL)
>> > >  -               return;
>> > >
>> > > Note release_tty() tty->port->itty is assigned NULL before calling
>> > > tty_buffer_cancel_work():
>> >
>> > The READ_ONCE should still handle that.
>> >
>> > Anyway, the changes were purely to try to remove the need for a ldisc in
>> > the serdev case and avoid referencing it. In fact we still have an
>> > ldisc, it's just not used. So we can restore the original ordering.
>> >
>> > Can you try this patch:
>> >
>>
>> Please try this one instead. It passes the tty struct around instead of
>> the ldisc.
>>
>
> Happy to test the fix, except reproducing the bug without changing anything
> at all has proven elusive.  AFAIK we just have my single experience
> described in the report, I'm unconfident in my ability to validate any
> fixes for this specific bug.
>
> If you think you understand the root cause, maybe you can conceive of a
> more reliable reproducer than me just using my machine?  It appears I may
> have just been very (un)lucky.

Honestly, I don't understand why my change caused a problem. The fix
just changes things back to how things we ordered before.

Looks like the need for READ_ONCE was originally found with ktsan, so
maybe it could help here. But I've never used it.

Rob