On Fri 14 Feb 2025 at 17:33, Greg Kroah-Hartman [off-list ref] wrote:

On Tue, Feb 11, 2025 at 06:27:58PM +0100, Jerome Brunet wrote:

quoted

Add helper functions to create a device on the auxiliary bus.

This is meant for fairly simple usage of the auxiliary bus, to avoid having
the same code repeated in the different drivers.

Suggested-by: Stephen Boyd <sboyd@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jerome Brunet <jbrunet@baylibre.com>
---
 drivers/base/auxiliary.c      | 88 +++++++++++++++++++++++++++++++++++++++++++
 include/linux/auxiliary_bus.h | 10 +++++
 2 files changed, 98 insertions(+)

I like the idea, see much the same of what I recently did for the "faux"
bus here:
	https://lore.kernel.org/all/2025021023-sandstorm-precise-9f5d@gregkh/ (local)

Reading this, I'm getting the feeling that some (most?) simple auxiliary
driver might be better off migrating to "faux", instead of what I'm
proposing here ? Is this what you are suggesting ?

Few Q:
Is there some sort of 'platform_data' (sorry for the lack of a better
term, no provocation intended ;) ) ... it there a
simple way to pass an arbitrary struct to the created device with 'faux' ?

The difference between aux and faux I'm seeing it that aux seems to
decouple things a bit more. The only thing aux needs is a module name to
pop something up, while faux needs a reference to the ops instead.

I can see the appeal to use aux for maintainers trying to decouple
different subsystems.

Some review comments:

quoted

diff --git a/drivers/base/auxiliary.c b/drivers/base/auxiliary.c
index afa4df4c5a3f371b91d8dd8c4325495d32ad1291..0f697c9c243dc9a50498a52362806db594345faf 100644
--- a/drivers/base/auxiliary.c
+++ b/drivers/base/auxiliary.c

@@ -385,6 +385,94 @@ void auxiliary_driver_unregister(struct auxiliary_driver *auxdrv)
 }
 EXPORT_SYMBOL_GPL(auxiliary_driver_unregister);
 
+static void auxiliary_device_release(struct device *dev)
+{
+	struct auxiliary_device *auxdev = to_auxiliary_dev(dev);
+
+	kfree(auxdev);
+}
+
+static struct auxiliary_device *auxiliary_device_create(struct device *dev,
+							const char *modname,
+							const char *devname,
+							void *platform_data,

Can you have the caller set the platform_data if they need/want it after
the device is created?  Or do you need that in the probe callback?

My assumption was that it is needed in probe, but I guess that entirely
depends on the driver. If that was ever needed, it could be added later
I think.

And can't this be a global function too for those that don't want to
deal with devm stuff?

There was a note about that in the cover-letter of the v1 but I did not
repeat it after.

It can be exported but I had no use for it so I thought It was better not
export it until it was actually needed. I really do not have a strong
preference over this.

quoted

+							int id)
+{
+	struct auxiliary_device *auxdev;
+	int ret;
+
+	auxdev = kzalloc(sizeof(*auxdev), GFP_KERNEL);
+	if (!auxdev)
+		return ERR_PTR(-ENOMEM);

Ick, who cares what the error value really is?  Why not just do NULL or
a valid pointer?  That makes the caller much simpler to handle, right?

Sure why not

quoted

+
+	auxdev->id = id;
+	auxdev->name = devname;
+	auxdev->dev.parent = dev;
+	auxdev->dev.platform_data = platform_data;
+	auxdev->dev.release = auxiliary_device_release;
+	device_set_of_node_from_dev(&auxdev->dev, dev);
+
+	ret = auxiliary_device_init(auxdev);

Only way this will fail is if you forgot to set parent or a valid name.
So why not check for devname being non-NULL above this?

If auxiliary_device_init() ever changes it would be easy to forget to
update that and lead to something nasty to debug, don't you think ?

If you are OK with this, I could update in this direction.

quoted

+	if (ret) {
+		kfree(auxdev);
+		return ERR_PTR(ret);
+	}
+
+	ret = __auxiliary_device_add(auxdev, modname);
+	if (ret) {
+		/*
+		 * NOTE: It may look odd but auxdev should not be freed
+		 * here. auxiliary_device_uninit() calls device_put()
+		 * which call the device release function, freeing auxdev.
+		 */
+		auxiliary_device_uninit(auxdev);

Yes it is odd, are you SURE you should be calling device_del() on the
device if this fails?  auxiliary_device_uninit(), makes sense so why not
just call that here?

I'm confused ... I am call auxiliary_device_uninit() here. What do you
mean ? 

quoted

+		return ERR_PTR(ret);
+	}
+
+	return auxdev;
+}
+
+static void auxiliary_device_destroy(void *_auxdev)
+{
+	struct auxiliary_device *auxdev = _auxdev;
+
+	auxiliary_device_delete(auxdev);
+	auxiliary_device_uninit(auxdev);
+}
+
+/**
+ * __devm_auxiliary_device_create - create a device on the auxiliary bus
+ * @dev: parent device
+ * @modname: module name used to create the auxiliary driver name.
+ * @devname: auxiliary bus device name
+ * @platform_data: auxiliary bus device platform data
+ * @id: auxiliary bus device id
+ *
+ * Device managed helper to create an auxiliary bus device.
+ * The device create matches driver 'modname.devname' on the auxiliary bus.
+ */
+struct auxiliary_device *__devm_auxiliary_device_create(struct device *dev,
+							const char *modname,
+							const char *devname,
+							void *platform_data,
+							int id)
+{
+	struct auxiliary_device *auxdev;
+	int ret;
+
+	auxdev = auxiliary_device_create(dev, modname, devname, platform_data, id);
+	if (IS_ERR(auxdev))
+		return auxdev;
+
+	ret = devm_add_action_or_reset(dev, auxiliary_device_destroy,
+				       auxdev);

Oh this is going to be messy, but I trust that callers know what they
are doing here.  Good luck!  :)

thanks,

greg k-h

-- 
Jerome