[PATCH 6/9] signal: Always call do_notify_parent_cldstop with siglock held

[Eric W. Biederman]

Oleg Nesterov <oleg at redhat.com> writes:

> On 04/27, Oleg Nesterov wrote:
>>
>> On 04/26, Eric W. Biederman wrote:
>> >
>> > @@ -2209,6 +2213,34 @@ static int ptrace_stop(int exit_code, int why, int clear_code,
>> >  		spin_lock_irq(&current->sighand->siglock);
>> >  	}
>> >
>> > +	/* Don't stop if current is not ptraced */
>> > +	if (unlikely(!current->ptrace))
>> > +		return (clear_code) ? 0 : exit_code;
>> > +
>> > +	/*
>> > +	 * If @why is CLD_STOPPED, we're trapping to participate in a group
>> > +	 * stop.  Do the bookkeeping.  Note that if SIGCONT was delievered
>> > +	 * across siglock relocks since INTERRUPT was scheduled, PENDING
>> > +	 * could be clear now.  We act as if SIGCONT is received after
>> > +	 * TASK_TRACED is entered - ignore it.
>> > +	 */
>> > +	if (why == CLD_STOPPED && (current->jobctl & JOBCTL_STOP_PENDING))
>> > +		gstop_done = task_participate_group_stop(current);
>> > +
>> > +	/*
>> > +	 * Notify parents of the stop.
>> > +	 *
>> > +	 * While ptraced, there are two parents - the ptracer and
>> > +	 * the real_parent of the group_leader.  The ptracer should
>> > +	 * know about every stop while the real parent is only
>> > +	 * interested in the completion of group stop.  The states
>> > +	 * for the two don't interact with each other.  Notify
>> > +	 * separately unless they're gonna be duplicates.
>> > +	 */
>> > +	do_notify_parent_cldstop(current, true, why);
>> > +	if (gstop_done && ptrace_reparented(current))
>> > +		do_notify_parent_cldstop(current, false, why);
>>
>> This doesn't look right too. The parent should be notified only after
>> we set __state = TASK_TRACED and ->exit code.
>>
>> Suppose that debugger sleeps in do_wait(). do_notify_parent_cldstop()
>> wakes it up, debugger calls wait_task_stopped() and then it will sleep
>> again, task_stopped_code() returns 0.
>>
>> This can be probably fixed if you remove the lockless (fast path)
>> task_stopped_code() check in wait_task_stopped(), but this is not
>> nice performance-wise...

Another detail I have overlooked.  Thank you.

Or we can change task_stopped_code look something like:

static int *task_stopped_code(struct task_struct *p, bool ptrace)
{
	if (ptrace) {
-		if (task_is_traced(p) && !(p->jobctl & JOBCTL_LISTENING))
+		if (p->ptrace && !(p->jobctl & JOBCTL_LISTENING))
			return &p->exit_code;
	} else {
		if (p->signal->flags & SIGNAL_STOP_STOPPED)
			return &p->signal->group_exit_code;
	}
	return NULL;
}

I probably need to do a little bit more to ensure that it isn't an
actual process exit_code in p->exit_code.  But the we don't have to
limit ourselves to being precisely in the task_is_traced stopped place
for the fast path.


> On the other hand, I don't understand why did you move the callsite
> of do_notify_parent_cldstop() up... just don't do this?

My goal and I still think it makes sense (if not my implementation)
is to move set_special_state as close as possible to schedule().

That way we can avoid sleeping spin_locks clobbering it and making
our life difficult.

My hope is we can just clean up ptrace_stop instead of making it more
complicated and harder to follow.  Not that I am fundamentally opposed
to the quiesce bit but the code is already very hard to follow because
of all it's nuance and complexity, and I would really like to reduce
that complexity if we can possibly figure out how.

Eric