Re: [PATCH 17/17] ptrace: implement PTRACE_LISTEN

[Tejun Heo <tj@kernel.org>]

Hello, Oleg.

On Thu, Jun 02, 2011 at 07:33:30PM +0200, Oleg Nesterov wrote:
> >  	p_code = task_stopped_code(p, ptrace);
> > -	if (unlikely(!p_code))
> > +	if (unlikely(!p_code) || p->jobctl & JOBCTL_LISTENING)
> >  		goto unlock_sig;
> 
> Up to you, but perhaps this JOBCTL_LISTENING check should go into
> task_stopped_code() ? Or do you think we can't check it without
> siglock?

So updated.  I don't think it's gonna introduce any new race
condition.

> > +		/*
> > +		 * If NOTIFY is set, it means event happened between start
> > +		 * of this trap and now.  Trigger re-trap immediately.
> > +		 */
> > +		if (child->jobctl & JOBCTL_TRAP_NOTIFY)
> > +			signal_wake_up(child, true);
> 
> Again, I won't insist if you prefer signal_wake_up(), but afaics
> wake_up_state(__TASK_TRACED) should be enough.

Re-trapping from attach/detach paths are already using
signal_wake_up() and I think it would be better to keep it consistent.

> > @@ -838,7 +840,7 @@ static void ptrace_trap_notify(struct task_struct *t)
> >  	assert_spin_locked(&t->sighand->siglock);
> >
> >  	task_set_jobctl_pending(t, JOBCTL_TRAP_NOTIFY);
> > -	signal_wake_up(t, 0);
> > +	signal_wake_up(t, t->jobctl & JOBCTL_LISTENING);
> >  }
> 
> OK. The only thing I can't understand is why prepare_signal(SIGCONT)
> calls ptrace_trap_notify() unconditionally. How about
> 
> 		if (likely(!(t->ptrace & PT_SEIZED)))
> 			wake_up_state(t, __TASK_STOPPED);
> 	-	else
> 	+	else if (why)
> 			ptrace_trap_notify(t);
> 
> ?

I'm having a Deja Vu.  Did I reply to this already?  Anyways, here are
my rationales.

* Tracer should be able to handle seemingly spurious notifications.
  e.g. rapid SIGSTOP/CONT sequence may generate seemingly spurious
  notifications even when it actually isn't spurious.

  SIGCONT always generating notification is correct and I don't see
  good reasons to optimize it.  Moreover, I think it doesn't hurt to
  have a way to reliably trigger spurious notification.

* If we're gonna optimize out SIGCONT processing if the target process
  doesn't need it, the proper way would be testing stopped state and
  exit before walking through the group list.  However, I think it's
  done the current way for a reason - always trying to wake up on
  SIGCONT is more robust in case something went out of sync &&
  optimizing spurious SIGCONT doesn't really help anyone.

So, I'd like to keep this one as it currently is.  It's more robust
and useful this way.

Thanks.

-- 
tejun