Re: [PATCH v8 3/5] amba: Don't unprepare the clocks if device driver wants IRQ safe runtime PM

[Krzysztof Kozlowski]

On sob, 2014-11-01 at 01:01 +0000, Russell King - ARM Linux wrote:
> On Sat, Nov 01, 2014 at 12:55:14AM +0000, Russell King - ARM Linux wrote:
> > On Sat, Nov 01, 2014 at 01:45:47AM +0100, Rafael J. Wysocki wrote:
> > > On Monday, October 20, 2014 11:04:46 AM Krzysztof Kozlowski wrote:
> > > > @@ -198,8 +217,10 @@ static int amba_probe(struct device *dev)
> > > >  		pm_runtime_enable(dev);
> > > >  
> > > >  		ret = pcdrv->probe(pcdev, id);
> > > > -		if (ret == 0)
> > > > +		if (ret == 0) {
> > > > +			pcdev->irq_safe = pm_runtime_is_irq_safe(dev);
> > > 
> > > This looks racy.
> > > 
> > > Is it guaranteed that runtime PM callbacks won't be run for the device
> > > after pcdrv->probe() has returned and before setting pcdev->irq_safe?
> > > If not, inconsistent behavior may ensue.
> > 
> > You are absolutely correct.  So that knocks that idea on its head.
> 
> Actually, I think we shouldn't give up hope here.  Currently, we do this:
> 
>                 pm_runtime_get_noresume(dev);
>                 pm_runtime_set_active(dev);
>                 pm_runtime_enable(dev);
> 
>                 ret = pcdrv->probe(pcdev, id);
> 
> What we could do is:
> 
> 		pm_runtime_get_noresume(dev);
>                 pm_runtime_get_noresume(dev);
>                 pm_runtime_set_active(dev);
>                 pm_runtime_enable(dev);
> 
>                 ret = pcdrv->probe(pcdev, id);
> 		if (ret == 0) {
> 			pcdev->irq_safe = pm_runtime_is_irq_safe(dev);
> 			pm_runtime_put(dev);
> 			break;
> 		}
> 
>                 pm_runtime_disable(dev);
>                 pm_runtime_set_suspended(dev);
>                 pm_runtime_put_noidle(dev);
>                 pm_runtime_put_noidle(dev);
> 
> which would ensure that we hold a usecount until after the probe function
> has returned.  Would that work?
> 
> I'll give you that it's pretty horrid.

> Would another possible solution be to remember the irq-safeness in the
> suspend handler, and use that in the resume handler?  Resume should
> /always/ undo what the suspend handler previously did wrt clk API stuff.

I think the second solution could be more readable. The WARN_ON wouldn't
be needed. However this won't solve the two dual nature of runtime
callbacks.

I wondered also about removing runtime PM callbacks from amba/bus.c
completely and moving this to child drivers. This way runtime PM would
be obvious in each driver case.

Best regards,
Krzysztof


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/