diff mbox series

[RFC] gpio: sysfs: make the sysfs export behavior consistent

Message ID 20240927074221.9985-1-brgl@bgdev.pl
State New
Headers show
Series [RFC] gpio: sysfs: make the sysfs export behavior consistent | expand

Commit Message

Bartosz Golaszewski Sept. 27, 2024, 7:42 a.m. UTC 
From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>

For drivers or board files that set gpio_chip->names, the links to the
GPIO attribute group created on sysfs export will be named after the
line's name set in that array. For lines that are named using device
properties, the names pointer of the gpio_chip struct is never assigned
so they are exported as if they're not named.

The ABI documentation does not mention the former behavior and given
that the majority of modern systems use device-tree, ACPI or other way
of passing GPIO names using device properties - bypassing gc->names -
it's better to make the behavior consistent by always exporting lines as
"gpioXYZ".

Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
---
Story time:

I decided to learn rust. I figured I'd best find me a project to work on
that would involve some proper coding but wouldn't have much impact on
anything important when I inevitably get it wrong the first few times.

I decided to write a sysfs-to-libgpiod compatibility layer based on
FUSE. Since Rust is hard, I started prototyping the thing in python
first to at least have the logic nailed down before I tackle the rust
part.

When working on the exporting part, I vagely recalled that when I used
to work with GPIO sysfs somewhere between 2009 and 2012 (still with
board-file based systems), named lines exported with sysfs would appear
under /sys/class/gpio as symbolic links named like the line and not the
usual "gpioXYZ". I realized that this is not the case now.

Quick glance at the sysfs code reveals that I didn't dream it up, but
that behavior is reserved to drivers setting gc->names. This has been
slowly going out of fashion with device-tree and device properties.

We could easily restore the behavior for everybody by taking the name
from the descriptor we already have access to or even just assign
gc->names from descriptors in gpiolib core but first: the sysfs ABI
document does not mention the named links at all and second: given how
this has naturally effectively been phased out over the years, it would
probably cause more harm than good when the exported names suddenly
change for existing users.

I'm proposing to just drop the named links alogether.

Let me know what you think.

 drivers/gpio/gpiolib-sysfs.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

Comments

Kent Gibson Sept. 30, 2024, 1:03 a.m. UTC | #1 
On Fri, Sep 27, 2024 at 09:42:21AM +0200, Bartosz Golaszewski wrote:
> From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
>
> For drivers or board files that set gpio_chip->names, the links to the
> GPIO attribute group created on sysfs export will be named after the
> line's name set in that array. For lines that are named using device
> properties, the names pointer of the gpio_chip struct is never assigned
> so they are exported as if they're not named.
>
> The ABI documentation does not mention the former behavior and given
> that the majority of modern systems use device-tree, ACPI or other way
> of passing GPIO names using device properties - bypassing gc->names -
> it's better to make the behavior consistent by always exporting lines as
> "gpioXYZ".
>

I have no opinions on sysfs changes.

> Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
> ---
> Story time:
>
> I decided to learn rust. I figured I'd best find me a project to work on
> that would involve some proper coding but wouldn't have much impact on
> anything important when I inevitably get it wrong the first few times.
>
> I decided to write a sysfs-to-libgpiod compatibility layer based on
> FUSE. Since Rust is hard, I started prototyping the thing in python
> first to at least have the logic nailed down before I tackle the rust
> part.
>

Something along these lines[1]?

Cheers,
Kent.

[1]https://dev.to/krjakbrjak/simulating-gpio-sysfs-interface-with-fuse-and-c-30ga
Bartosz Golaszewski Sept. 30, 2024, 7:20 a.m. UTC | #2 
On Mon, Sep 30, 2024 at 3:03 AM Kent Gibson <warthog618@gmail.com> wrote:
>
> > I decided to write a sysfs-to-libgpiod compatibility layer based on
> > FUSE. Since Rust is hard, I started prototyping the thing in python
> > first to at least have the logic nailed down before I tackle the rust
> > part.
> >
>
> Something along these lines[1]?
>
> Cheers,
> Kent.
>
> [1]https://dev.to/krjakbrjak/simulating-gpio-sysfs-interface-with-fuse-and-c-30ga
>

Well, this doesn't really do anything. I'm thinking about something
consuming the libgpiod rust bindings to actually be a useful
replacement for kernel sysfs.

The master plan is: provide a drop-in user-space replacement for
sysfs, make users convert to using it instead of the real thing,
eventually remove sysfs from the kernel and then some time after
remove the compatibility layer from existence forcing everybody to now
move to pure libgpiod. :)

Bart
Kent Gibson Sept. 30, 2024, 7:55 a.m. UTC | #3 
On Mon, Sep 30, 2024 at 09:20:48AM +0200, Bartosz Golaszewski wrote:
> On Mon, Sep 30, 2024 at 3:03 AM Kent Gibson <warthog618@gmail.com> wrote:
> >
> > > I decided to write a sysfs-to-libgpiod compatibility layer based on
> > > FUSE. Since Rust is hard, I started prototyping the thing in python
> > > first to at least have the logic nailed down before I tackle the rust
> > > part.
> > >
> >
> > Something along these lines[1]?
> >
> > Cheers,
> > Kent.
> >
> > [1]https://dev.to/krjakbrjak/simulating-gpio-sysfs-interface-with-fuse-and-c-30ga
> >
>
> Well, this doesn't really do anything. I'm thinking about something
> consuming the libgpiod rust bindings to actually be a useful
> replacement for kernel sysfs.
>

I haven't actually looked at that - but when you mentioned a sysfs fuse
interface that rang a bell.

> The master plan is: provide a drop-in user-space replacement for
> sysfs, make users convert to using it instead of the real thing,
> eventually remove sysfs from the kernel and then some time after
> remove the compatibility layer from existence forcing everybody to now
> move to pure libgpiod. :)
>

Ironically a lot of users would probably be happier with that than with
the D-bus daemon.  And for that reason I don't think that last step will
ever happen.  But that is ok too - as long as we can finally remove the
sysfs interface from the kernel.

Cheers,
Kent.
Linus Walleij Oct. 2, 2024, 1:12 p.m. UTC | #4 
On Fri, Sep 27, 2024 at 9:42 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:

> From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
>
> For drivers or board files that set gpio_chip->names, the links to the
> GPIO attribute group created on sysfs export will be named after the
> line's name set in that array. For lines that are named using device
> properties, the names pointer of the gpio_chip struct is never assigned
> so they are exported as if they're not named.
>
> The ABI documentation does not mention the former behavior and given
> that the majority of modern systems use device-tree, ACPI or other way
> of passing GPIO names using device properties - bypassing gc->names -
> it's better to make the behavior consistent by always exporting lines as
> "gpioXYZ".
>
> Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>

I'm in favor of this.
Reviewed-by: Linus Walleij <linus.walleij@linaro.org>

> Story time:

That's a good story :) Fun to see how you arrived at this.

Yours,
Linus Walleij
Bartosz Golaszewski Oct. 2, 2024, 2:31 p.m. UTC | #5 
On Wed, Oct 2, 2024 at 3:12 PM Linus Walleij <linus.walleij@linaro.org> wrote:
>
> On Fri, Sep 27, 2024 at 9:42 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
>
> > From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
> >
> > For drivers or board files that set gpio_chip->names, the links to the
> > GPIO attribute group created on sysfs export will be named after the
> > line's name set in that array. For lines that are named using device
> > properties, the names pointer of the gpio_chip struct is never assigned
> > so they are exported as if they're not named.
> >
> > The ABI documentation does not mention the former behavior and given
> > that the majority of modern systems use device-tree, ACPI or other way
> > of passing GPIO names using device properties - bypassing gc->names -
> > it's better to make the behavior consistent by always exporting lines as
> > "gpioXYZ".
> >
> > Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
>
> I'm in favor of this.
> Reviewed-by: Linus Walleij <linus.walleij@linaro.org>
>

I take it this also applies to v2 and am picking it up instead of this one.

Bart

> > Story time:
>
> That's a good story :) Fun to see how you arrived at this.
>
> Yours,
> Linus Walleij
diff mbox series

Patch

diff --git a/drivers/gpio/gpiolib-sysfs.c b/drivers/gpio/gpiolib-sysfs.c
index 17ed229412af..643620d261f5 100644
--- a/drivers/gpio/gpiolib-sysfs.c
+++ b/drivers/gpio/gpiolib-sysfs.c
@@ -577,7 +577,7 @@  int gpiod_export(struct gpio_desc *desc, bool direction_may_change)
 	struct gpio_device *gdev;
 	struct gpiod_data *data;
 	struct device *dev;
-	int status, offset;
+	int status;
 
 	/* can't export until sysfs is available ... */
 	if (!class_is_registered(&gpio_class)) {
@@ -626,10 +626,6 @@  int gpiod_export(struct gpio_desc *desc, bool direction_may_change)
 	else
 		data->direction_can_change = false;
 
-	offset = gpio_chip_hwgpio(desc);
-	if (guard.gc->names && guard.gc->names[offset])
-		ioname = guard.gc->names[offset];
-
 	dev = device_create_with_groups(&gpio_class, &gdev->dev,
 					MKDEV(0, 0), data, gpio_groups,
 					ioname ? ioname : "gpio%u",