Re: [PATCH v4 3/3] gpio: ws16c48: Migrate to the regmap API

[William Breathitt Gray]

On Wed, Mar 08, 2023 at 03:06:18PM +0200, Andy Shevchenko wrote:
> On Tue, Mar 07, 2023 at 09:51:26PM -0500, William Breathitt Gray wrote:
> > On Mon, Mar 06, 2023 at 04:20:03PM +0200, Andy Shevchenko wrote:
> > > On Mon, Mar 06, 2023 at 07:59:53AM -0500, William Breathitt Gray wrote:
> 
> ...
> 
> > > > -	raw_spinlock_t lock;
> > > > +	spinlock_t lock;
> > > 
> > > This is a regression.
> > > That said, do we need a support of raw spin locks in the regmap IRQ?
> > 
> > So this code has a similar need as the gpio-pcie-idio-24 patch: guard
> > registers between handle_mask_sync() and set_type_config(); however, now
> > we also need to protect registers in regmap_irq_thread(). We can't use a
> > mutex here because regmap_irq_thread() is executed in an interrupt
> > context so we cannot sleep.
> > 
> > This might be a mistake in my understanding: I chose spinlock_t here
> > because I believed it to map out to a raw_spinlock_t anyway underneath,
> > whereas on RT kernels it would map out to whatever the equivalent is. I
> > suspect this is not actually the case. Would using raw_spinlock_t
> > explicitly be the correct way to go for this particular case?
> 
> You may read the commit message of the 27d9098cff6e ("pinctrl: intel:
> Use raw_spinlock for locking"). TL;DR: this is only affects IRQ chips,
> so if your GPIO controller is _not_ an IRQ chip, you are fine.
> 
> WRT the other driver, can_sleep may reduce scope of the use of GPIOs
> and even make a regression if any consumer don't want that behaviour
> and currently works.

Looking through kernel/irq/manage.c, I see the raw_spinlock desc->lock
is taken in __setup_irq() before potentially calling __irq_set_trigger()
which ultimate calls the chip->irq_set_type() callback. So it seems
unsafe to sleep within at least this callback which is utilized by both
drivers, so both gpio-pcie-idio-24 and gpio-ws16c48 will need the
raw_spinlock lock type afterall.

I'll make the necessary changes and release a v5 of this patchset.

As an aside, I wonder if locking is not needed if we only utilize the
set_type_config() callback, because the desc->lock taken by the irq
subsystem will be enough to guard between regmap_irq_set_type() and
regmap_irq_thread(). It's not valid for our particular case here because
we also utilize a handle_mask_sync() callback (chip_bus_lock() is not
protected by desc->chip) but it's something to think about.

William Breathitt Gray

Attachment: signature.asc
Description: PGP signature