[PATCH 9/9] ptrace: Don't change __state

[Oleg Nesterov]

On 04/27, Eric W. Biederman wrote:
>
> "Eric W. Biederman" <ebiederm at xmission.com> writes:
>
> > diff --git a/include/linux/sched/signal.h b/include/linux/sched/signal.h
> > index 3c8b34876744..1947c85aa9d9 100644
> > --- a/include/linux/sched/signal.h
> > +++ b/include/linux/sched/signal.h
> > @@ -437,7 +437,8 @@ extern void signal_wake_up_state(struct task_struct *t, unsigned int state);
> >
> >  static inline void signal_wake_up(struct task_struct *t, bool resume)
> >  {
> > -	signal_wake_up_state(t, resume ? TASK_WAKEKILL : 0);
> > +	bool wakekill = resume && !(t->jobctl & JOBCTL_DELAY_WAKEKILL);
> > +	signal_wake_up_state(t, wakekill ? TASK_WAKEKILL : 0);
> >  }
> >  static inline void ptrace_signal_wake_up(struct task_struct *t, bool resume)
> >  {
>
> Grrr.  While looking through everything today I have realized that there
> is a bug.
>
> Suppose we have 3 processes: TRACER, TRACEE, KILLER.
>
> Meanwhile TRACEE is in the middle of ptrace_stop, just after siglock has
> been dropped.
>
> The TRACER process has performed ptrace_attach on TRACEE and is in the
> middle of a ptrace operation and has just set JOBCTL_DELAY_WAKEKILL.
>
> Then comes in the KILLER process and sends the TRACEE a SIGKILL.
> The TRACEE __state remains TASK_TRACED, as designed.
>
> The bug appears when the TRACEE makes it to schedule().  Inside
> schedule there is a call to signal_pending_state() which notices
> a SIGKILL is pending and refuses to sleep.

And I think this is fine. This doesn't really differ from the case
when the tracee was killed before it takes siglock.

The only problem (afaics) is that, once we introduce JOBCTL_TRACED,
ptrace_stop() can leak this flag. That is why I suggested to clear
it along with LISTENING/DELAY_WAKEKILL before return, exactly because
schedule() won't block if fatal_signal_pending() is true.

But may be I misunderstood you concern?

Oleg.