Re: [PATCH] gpio: zynq: Implement irq_(request|release)_resources

[SÃren Brinkmann]

On Tue, 2015-10-27 at 05:37PM +0100, Lars-Peter Clausen wrote:
> On 10/27/2015 04:53 PM, Linus Walleij wrote:
> > On Fri, Oct 23, 2015 at 3:36 PM, Soren Brinkmann
> > <soren.brinkmann@xxxxxxxxxx> wrote:
> > 
> >> The driver uses runtime PM to leverage low power techniques. For
> >> use-cases using GPIO as interrupt the device needs to be in an
> >> appropriate state.
> >>
> >> Reported-by: John Linn <linnj@xxxxxxxxxx>
> >> Signed-off-by: Soren Brinkmann <soren.brinkmann@xxxxxxxxxx>
> >> Tested-by: John Linn <linnj@xxxxxxxxxx>
> > 
> > As pointed out by Grygorii in
> > commit aca82d1cbb49af34b69ecd4571a0fe48ad9247c1:
> > 
> >     The PM runtime API can't be used in atomic contex on -RT even if
> >     it's configured as irqsafe. As result, below error report can
> >     be seen when PM runtime API called from IRQ chip's callbacks
> >     irq_startup/irq_shutdown/irq_set_type, because they are
> >     protected by RAW spinlock:
> > (...)
> >     The IRQ chip interface defines only two callbacks which are executed in
> >     non-atomic contex - irq_bus_lock/irq_bus_sync_unlock, so lets move
> >     PM runtime calls there.
> > 
> > I.e. these calls are atomic context and it's just luck that it works
> > and this is fragile.
> > 
> > Can you please check if you can move it to
> > irq_bus_lock()/irq_sync_unlock()
> > like Grygorii does?
> 
> That only powers up the chip when the chip is accessed. For proper IRQ
> operation the chip needs to be powered up though as long as the IRQ is
> enabled. request_irq() and free_irq() must always be called from sleepable
> context. The thing is just that request_resource/release_resource are called
> from within a raw spinlock, which is necessary since otherwise you can't
> guarantee that they are only called once for shared interrupts.
> 
> It might make sense to add a separate set of callbacks to the irq_chip
> struct that are called from the sleepable sections of
> request_irq()/free_irq() which are meant for power management purposes and
> which wont have the guarantee that they are only called once for shared IRQs
> (but are still balanced).

Let me try to summarize what I've heard so far:
 - reqres/relres are called from atomic context, hence must not sleep
 - pm_runtime API must not be used from atomic context, even when the
   implementation of the callbacks does not sleep
 - when overriding regres/relres, a driver must re-implement the default
   behavior the core provides
 - bus lock/unlock is not sufficient for this case because it doesn't
   keep the device on as long as an IRQ is expected
 - a new pair of gpiochip ops might be helpful

Does that summarize the current situation correctly? If so, I'd tend to
agree with Lars that we might need another pair of callbacks in the
gpiochip struct.

	Thanks,
	SÃren
--
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/