Re: [PATCH v1 3/8] sched/completion: convert completions to use simple wait queues

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Oct 12, 2015 at 12:03:06PM +0200, Daniel Wagner wrote:
> On 10/12/2015 11:17 AM, Daniel Wagner wrote:
> > On 09/09/2015 04:26 PM, Peter Zijlstra wrote:
> >> On Wed, Sep 09, 2015 at 02:05:29PM +0200, Daniel Wagner wrote:
> >>> @@ -50,10 +50,10 @@ void complete_all(struct completion *x)
> >>>  {
> >>>  	unsigned long flags;
> >>>  
> >>> -	spin_lock_irqsave(&x->wait.lock, flags);
> >>> +	raw_spin_lock_irqsave(&x->wait.lock, flags);
> >>>  	x->done += UINT_MAX/2;
> >>> -	__wake_up_locked(&x->wait, TASK_NORMAL, 0);
> >>> -	spin_unlock_irqrestore(&x->wait.lock, flags);
> >>> +	swake_up_locked(&x->wait);
> >>> +	raw_spin_unlock_irqrestore(&x->wait.lock, flags);
> >>>  }
> >>>  EXPORT_SYMBOL(complete_all);
> >>
> >> I don't think that's correct; __wake_up_locked(.nr=0) would wake all
> >> waiters, where swake_up_locked() will only wake one.
> > 
> > I read that x->done should be protected via wait.lock during the whole
> > operation. swake_up_all() will release and reacquire the lock while
> > processing the all waiters. So we need to get
> > 
> > Could we play a trick like setting the highest bit in done for
> > indicating the complete_all() operation. The UINT_MAX/2 update looks
> > like do this by setting a value which has the biggest offset from 0 (but
> > why adding instead of just going for assigning...).
> 
> 
> I had something like this here in mind:

I'm not exactly sure what problem you're trying to solve here.. The fact
that we cannot call swake_all() while holding &x->wait.lock, or the fact
that complete_all() is typically called from a context which cannot do
swake_all() either?

Note:

Documentation/scheduler/completion.txt:complete() and complete_all() can be called in hard-irq/atomic context safely.

Which is very much _NOT_ true of swake_all().