Re: [PATCH 2/2] gpio: use a mutex to protect the list of GPIO devices

[Bartosz Golaszewski <brgl@xxxxxxxx>]

On Tue, Nov 28, 2023 at 3:21 PM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
>
> On Mon, Nov 27, 2023 at 8:37 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
>
> > From: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
> >
> > The global list of GPIO devices is never modified or accessed from
> > atomic context so it's fine to protect it using a mutex. Add a new
> > global lock dedicated to the gpio_devices list and use it whenever
> > accessing or modifying it.
> >
> > While at it: fold the sysfs registering of existing devices into
> > gpiolib.c and make gpio_devices static within its compilation unit.
> >
> > Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@xxxxxxxxxx>
>
> Nice! I might have found some snag:
>
> gpio_device_find() still does guard(spinlock_irqsave)(&gpio_lock);
> shouldn't that be switched to the mutex?
>

Good catch!

> On top of this I can update my patch to the delete the comment
> for gpio_lock to just rename that thing to gpio_descriptor_lock
> and document it as such.
>

No need, this will soon go away anyway. See below.

> But when I think about it: gpio[_decriptor]_lock can now (after this
> patch) be moved into struct gpio_chip as it is really just protecting
> the descriptors on the same chip from simultaneous modification,
> especially desc->flags. This is a BIG WIN because it makes it a local
> lock not a global one, do you wanna try it or should I? (On top of
> these two patches, then.)
>

I will have the series making locking in GPIOLIB more fine-grained
ready tomorrow or on Thursday. It will have separate locks for each
descriptor. We will use spinlock or mutex per descriptor depending on
the value of gc->can_sleep. I think it should work fine as a sleeping
chip can always use a mutex and a non-sleeping one cannot have
sleeping callbacks (correct me if I'm wrong).

We don't need to lock the GPIO device or chip separately - the
descriptor structs will stay alive as long as there's a live reference
to the GPIO device. GPIO device will have an SRCU cookie for
protecting API calls against removal of the chip.

To summarize: one mutex for the GPIO device list, one lock per GPIO
descriptor and SRCU protection of the GPIO device's chip.

Does it make sense?

Bart

> Yours,
> Linus Walleij