Re: Pinning down a blocked task to extract diagnostics

["Paul E. McKenney" <paulmck@kernel.org>]

On Thu, Mar 05, 2020 at 09:28:45AM -0500, Steven Rostedt wrote:
> On Thu, 5 Mar 2020 06:22:45 -0800
> "Paul E. McKenney" <paulmck@kernel.org> wrote:
> 
> > On Thu, Mar 05, 2020 at 09:13:37AM +0100, Peter Zijlstra wrote:
> > > On Thu, Mar 05, 2020 at 09:07:55AM +0100, Peter Zijlstra wrote:  
> > > > On Wed, Mar 04, 2020 at 04:50:49PM -0800, Paul E. McKenney wrote:  
> > > > > Hello!
> > > > > 
> > > > > Suppose that I need to extract diagnostics information from a blocked
> > > > > task, but that I absolutely cannot tolerate this task awakening in the
> > > > > midst of this extraction process.  Is the following code the right way
> > > > > to make this work given a task "t"?
> > > > > 
> > > > > 	raw_spin_lock_irq(&t->pi_lock);
> > > > > 	if (t->on_rq) {
> > > > > 		/* Task no longer blocked, so ignore it. */
> > > > > 	} else {
> > > > > 		/* Extract consistent diagnostic information. */
> > > > > 	}
> > > > > 	raw_spin_unlock_irq(&t->pi_lock);
> > > > > 
> > > > > It looks like all the wakeup paths acquire ->pi_lock, but I figured I
> > > > > should actually ask...  
> > > > 
> > > > Close, the thing pi_lock actually guards is the t->state transition *to*
> > > > TASK_WAKING/TASK_RUNNING, so something like this:  
> > > 
> > > Almost, we must indeed also check ->on_rq, otherwise it might change the
> > > state back itself.
> > >   
> > > > 
> > > > 	raw_spin_lock_irq(&t->pi_lock);
> > > > 	switch (t->state) {
> > > > 	case TASK_RUNNING:
> > > > 	case TASK_WAKING:
> > > > 		/* ignore */
> > > > 		break;
> > > > 
> > > > 	default:  
> > > 		if (t->on_rq)
> > > 			break;
> > >   
> > > > 		/* Extract consistent diagnostic information. */
> > > > 		break;
> > > > 	}
> > > > 	raw_spin_unlock_irq(&t->pi_lock);
> > > > 
> > > > ought to work. But if you're going to do this, please add a reference to
> > > > that code in a comment on top of try_to_wake_up(), such that we can
> > > > later find all the code that relies on this.  
> > 
> > How about if I add something like this, located right by try_to_wake_up()?
> > 
> > 	bool try_to_keep_sleeping(struct task_struct *t)
> > 	{
> > 		raw_spin_lock_irq(&t->pi_lock);
> > 		switch (t->state) {
> > 		case TASK_RUNNING:
> > 		case TASK_WAKING:
> > 			raw_spin_unlock_irq(&t->pi_lock);
> > 			return false;
> > 
> > 		default:
> > 			if (t->on_rq) {
> 
> Somehow I think there still needs to be a read barrier before the test to
> on_rq.

This is nowhere near a fastpath, so if there is uncertainty it gets
the smp_rmb().  Or an smp_load_acquire() on t->state.

> > 				raw_spin_unlock_irq(&t->pi_lock);
> > 				return false;
> > 			}
> > 
> > 			/* OK to extract consistent diagnostic information. */
> > 			return true;
> > 		}
> > 		/* NOTREACHED */
> > 	}
> > 
> > Then a use might look like this:
> > 
> > 	if (try_to_keep_sleeping(t))
> > 		/* Extract consistent diagnostic information. */
> > 		raw_spin_unlock_irq(&t->pi_lock);
> 
> Perhaps we should have a allow_awake(t) to match it?
> 
> 		allow_awake(t);
> 
> Where we have:
> 
> static inline allow_awake(struct task_struct *t)
> {
> 	raw_spin_unlock_irq(&t->pi_lock);
> }

Makes sense to me!

							Thanx, Paul

> -- Steve?
> 
> > 	} else {
> > 		/* Woo-hoo!  It started running again!!! */
> > 	}
> > 
> > Is there a better way to approach this?
> > 
> > 							Thanx, Paul
>