Re: [PATCH 3/7] wait.[ch]: Introduce the simple waitqueue (swait) implementation

From: Paul Gortmaker <paul.gortmaker@windriver.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: <linux-rt-users@vger.kernel.org>, <linux-kernel@vger.kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
	"Paul E. McKenney" <paulmck@linux.vnet.ibm.com>,
	Steven Rostedt <rostedt@goodmis.org>
Subject: Re: [PATCH 3/7] wait.[ch]: Introduce the simple waitqueue (swait) implementation
Date: Mon, 20 Oct 2014 09:44:58 -0400	[thread overview]
Message-ID: <20141020134457.GE24595@windriver.com> (raw)
In-Reply-To: <20141018213417.GE23531@worktop.programming.kicks-ass.net>

[Re: [PATCH 3/7] wait.[ch]: Introduce the simple waitqueue (swait) implementation] On 18/10/2014 (Sat 23:34) Peter Zijlstra wrote:

> On Fri, Oct 17, 2014 at 08:22:58PM -0400, Paul Gortmaker wrote:
> > @@ -75,6 +123,32 @@ static void __cwake_up_common(struct cwait_head *q, unsigned int mode,
> >  	}
> >  }
> >  
> > +static void __swake_up_common(struct swait_head *q, unsigned int mode,
> > +			      int nr_exclusive)
> > +{
> > +	struct swait *curr, *next;
> > +	int woken = 0;
> > +
> > +	list_for_each_entry_safe(curr, next, &q->task_list, node) {
> > +		if (wake_up_state(curr->task, mode)) { /* <-- calls ttwu() */
> > +			__remove_swait(q, curr);
> > +			curr->task = NULL;
> > +			/*
> > +			 * The waiting task can free the waiter as
> > +			 * soon as curr->task = NULL is written,
> > +			 * without taking any locks. A memory barrier
> > +			 * is required here to prevent the following
> > +			 * store to curr->task from getting ahead of
> > +			 * the dequeue operation.
> > +			 */
> > +			smp_wmb();
> > +			if (++woken == nr_exclusive)
> > +				break;
> > +		}
> > +
> > +	}
> > +}
> > +
> >  /**
> >   * __cwake_up - wake up threads blocked on a waitqueue.
> >   * @q: the complex waitqueue
> > @@ -96,6 +170,19 @@ void __cwake_up(struct cwait_head *q, unsigned int mode, int nr_exclusive,
> >  }
> >  EXPORT_SYMBOL(__cwake_up);
> >  
> > +void __swake_up(struct swait_head *q, unsigned int mode, int nr_exclusive)
> > +{
> > +	unsigned long flags;
> > +
> > +	if (!swait_active(q))
> > +		return;
> > +
> > +	raw_spin_lock_irqsave(&q->lock, flags);
> > +	__swake_up_common(q, mode, nr_exclusive);
> > +	raw_spin_unlock_irqrestore(&q->lock, flags);
> > +}
> > +EXPORT_SYMBOL(__swake_up);
> 
> Same comment as before, that is an unbounded loop in a non preemptible
> section and therefore violates RT design principles.

Yep, I hadn't forgot about that ; see patch 6/7 -- which has your
tentative solution from before.  I didn't want to squish that into
here and lose sight of it ; same for the smp barriers - I wanted
to ensure we didn't lose visibility of things needing discussion.

> 
> We actually did talk about ways of fixing that.

I'll follow up to Steve's comment on what he described.

> 
> Also, I'm not entirely sure we want to do the cwait thing, it looks
> painful.

The simplewait vs. complex wait as a whole, or just the rework to
make it more aligned with the existing code?  FWIW, I'm not married
to this particular implementation; so if ideas have changed since,
and the plan is different than what v2 implements, that is no problem.

P.