[RFC PATCH] Rework gpio cansleep (was Re: gpiolib and sleeping gpios)

[jamie@shareable.org (Jamie Lokier)]

Ryan Mallon wrote:
> On 06/24/2010 10:53 AM, Jamie Lokier wrote:
> > Ryan Mallon wrote:
> >> On 06/23/2010 04:37 PM, David Brownell wrote:
> >> I'm not. Some gpios, such as those on io expanders, may sleep in their
> >> implementations of the gpio_(set/get) functions.
> > 
> > I'm having a hard time figuring out where some GPIOs I'm using fit
> > into this picture.
> > 
> > I have some hardware that is currently using a 2.4.26 kernel, but I
> > look from time to time at forward-porting all the drivers to 2.6.recent.
> > 
> > It has an I2C driven GPIO expander, with a watchdog reset chip hanging
> > off the expander.
> > 
> > The watchdog is kept alive off the back end of a timer BH, which means
> > the I2C GPIO routines are written to be safe in BH context (which
> > isn't sleepable), but they can't be used in IRQ context because the
> > necessary spin_lock_irqsave() would turn off interrupts for too long
> > for other subsystems to function properly.
> 
> Do the implementations of the get/set calls for the io expander gpios
> sleep at all?

No, because sleeping isn't allowed in BH context.  (Note that this is
2.4.26 code - things have changed a bit for 2.6, but the hardware is
the same, and still needs the I2C watchdog to be driven from a BH-like
context).

> > How should I flag those GPIO routines in your scheme?  They're safe to
> > use in some non-sleeping contexts, but not safe in irq context.
> 
> The idea in my proposal is to use gpio_request in a driver if the
> requested gpio can never sleep (ie because of the context it is used
> in), and gpio_request_cansleep if the gpio is never used from non-sleep
> safe context in a driver. I suggested stripping back the patch to just
> add the gpio_request_cansleep function.
> 
> In the current code, if a driver ever calls gpio_(set/get)_value on a
> gpio then you cannot pass a sleeping gpio to that driver. The request
> will succeed, but you will get warnings with the get/get calls are made.
> My idea is basically to move the denotation of whether a gpio will be
> used in non-sleep safe context to the gpio request.

The reason I'm asking about my scenario is because the GPIO routines
can't sleep and are used from a non-sleep safe context - but they are
not safe to call in irq contexts.

So my watchdog driver would have to call gpio_request (not _cansleep)
- that's fine.  But if I connected other GPIOs from the same GPIO
driver (other lines on the same I/O expander chip) to another
GPIO-using driver which happens to use them from irq context, then
your changes won't detect the problem - the code will just break at
runtime.

Of course if I did that, it would be my fault and my problem.  I get
to keep both pieces etc.  But it's a scenario which your proposal
would fail to catch at compile time, that's why I bring it up.

-- Jamie