Re: [PATCH 4/9] ptrace: relocate set_current_state(TASK_TRACED) in ptrace_stop()

[Oleg Nesterov <oleg@redhat.com>]

On 05/16, Tejun Heo wrote:
>
> Hey, Oleg.
>
> On Mon, May 16, 2011 at 01:57:11PM +0200, Oleg Nesterov wrote:
> > > and helps future updates to group stop participation.
> >
> > OK, so I assume we need this change.
>
> We don't necessarily need it but it makes things prettier later.
>
> > But the comment looks a bit confusing to me. This is fine, I almost never
> > read them ;) Just I'd like to ensure I din't miss something.
>
> Oleg, IIRC, those comments were taken from your email pointing out
> that set_current_state() needs to happen before clearing of TRAPPING,
> so, if you're confused, I'm confused too. :-)

So, we are both confused. Great!

> > > +	 * We're committing to trapping.  TRACED should be visible before
> > > +	 * TRAPPING is cleared
> >
> > This looks as if you explain the barrier in set_current_state(). And,
> > btw, why can't we use __set_current_state() here ?
> >
> > And. not only TRACED, at least ->exit_code should be visible as well.
>
> The racy part was task_is_stopped_or_traced() in task_stopped_code()
> and the value of exit_code doesn't matter at that point.

Why exit_code doesn't matter? task_stopped_code() needs
task_is_stopped_or_traced() && exit_code != 0. Both changes should be
visible.

> So, we need
> at least smp_wmb() between __set_current_state() and clearing
> TRAPPING.

I don't think so. Please see below,

> > IOW. It is not that TRACED should be visible before jobctl &= ~JOBCTL_TRAPPING,
> > we should correctly update the tracee before __wake_up_sync_key(), and I assume
> > this is what the comment says.
> >
> > Correct?
>
> All we need to update on the tracee is tracee->state and
> ~JOBCTL_TRAPPING and __wake_up_sync_key() can be considered single
> operation.

Yes! IOW, it safe to reorder the memory operations which change ->state,
->exit_code, and ->jobctl. This only important thing is that we should not
wake up the tracer before we change them.


And if I remember correctly this was the problem, the early patches did
something like

	task_clear_jobctl_trapping();
	set_current_state(TASK_TRACED);

Oleg.