Re: [PATCH 2/5] sched,ptrace: Fix ptrace_check_attach() vs PREEMPT_RT

[Peter Zijlstra]

On Wed, Apr 13, 2022 at 09:20:53PM +0200, Peter Zijlstra wrote:
> On Wed, Apr 13, 2022 at 08:59:10PM +0200, Oleg Nesterov wrote:
> 
> 
> > +		// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
> > +		// wrong, needs siglock
> > +		current->jobctl &= ~JOBCTL_TRACED_XXX;
> > +		wake_up_bit(&current->jobctl, ~JOBCTL_TRACED_XXX_BIT);
> 		  __wake_up_common_lock()
> 		    spin_lock_irqsave()
> 		      current_save_and_set_rtlock_wait_state();
> 
> 			> +	if (wait_on_bit(&task->jobctl, JOBCTL_TRACED_XXX_BIT, TASK_KILLABLE))
> 			> +		return -EINTR;
> 			> +	// now that the tracee cleared JOBCTL_TRACED_XXX_BIT
> 			> +	// wait_task_inactive() should succeed or fail "really soon".
> 			> +	if (!wait_task_inactive(child, __TASK_TRACED))
> 			> +		return ret;
> 
> 
> 	*whoopsie* ?
> 
> >  		preempt_enable_no_resched();
> >  		freezable_schedule();
> >  		cgroup_leave_frozen(true);

Something that might work; since there's only the one ptracer, is to
instead do something like:

	current->jobctl &= ~JOBCTL_TRACED_XXX; // siglock
	if (current->ptrace)
		wake_up_process(current->parent);
	preempt_enable_no_resched();
	schedule();


vs

	for (;;) {
		set_current_state(TASK_UNINTERRUPTIBLE);
		if (!(p->jobctl & JOBCTL_TRACED_XXX))
			break;
		schedule();
	}
	__set_current_state(TASK_RUNNING);

	if (!wait_task_inactive(...))


ptrace_detach() needs some additional magic as well I think, but this
might just work.