Message-Id: <DBOJHHKW7OIB.2V01IOVQOGYXC@kernel.org>
Date: Tue, 29 Jul 2025 14:31:43 +0200
From: "Danilo Krummrich" <dakr@...nel.org>
To: "Leon Romanovsky" <leon@...nel.org>
Cc: "Sean Anderson" <sean.anderson@...ux.dev>, "Greg Kroah-Hartman"
 <gregkh@...uxfoundation.org>, "Ira Weiny" <ira.weiny@...el.com>,
 <linux-kernel@...r.kernel.org>, "Rafael J . Wysocki" <rafael@...nel.org>,
 "Dave Ertman" <david.m.ertman@...el.com>
Subject: Re: [PATCH] auxiliary: Automatically generate id

On Tue Jul 29, 2025 at 1:49 PM CEST, Leon Romanovsky wrote:
> On Tue, Jul 29, 2025 at 01:28:14PM +0200, Danilo Krummrich wrote:
>> On Tue Jul 29, 2025 at 1:11 PM CEST, Leon Romanovsky wrote:
>> > On Tue, Jul 29, 2025 at 12:51:42PM +0200, Danilo Krummrich wrote:
>> >> On Tue Jul 29, 2025 at 12:01 PM CEST, Leon Romanovsky wrote:
>> >> > On Tue, Jul 29, 2025 at 11:36:27AM +0200, Danilo Krummrich wrote:
>> >> >> On Mon Jul 28, 2025 at 11:10 PM CEST, Sean Anderson wrote:
>> >> >> > As it turns out, ids are not allowed to have semantic meaning. Their
>> >> >> > only purpose is to prevent sysfs collisions. To simplify things, just
>> >> >> > generate a unique id for each auxiliary device. Remove all references to
>> >> >> > filling in the id member of the device.
>> >> >> >
>> >> >> > Signed-off-by: Sean Anderson <sean.anderson@...ux.dev>
>> >> >> > ---
>> >> >> >
>> >> >> >  drivers/base/auxiliary.c      | 32 +++++++++++++++++++++++---------
>> >> >> >  include/linux/auxiliary_bus.h | 26 ++++++++------------------
>> >> >> >  2 files changed, 31 insertions(+), 27 deletions(-)
>> >> >> >
>> >> >> > diff --git a/drivers/base/auxiliary.c b/drivers/base/auxiliary.c
>> >> >> > index dba7c8e13a53..f66067df03ad 100644
>> >> >> > --- a/drivers/base/auxiliary.c
>> >> >> > +++ b/drivers/base/auxiliary.c
>> >> >> > @@ -264,6 +264,8 @@ static const struct bus_type auxiliary_bus_type = {
>> >> >> >  	.pm = &auxiliary_dev_pm_ops,
>> >> >> >  };
>> >> >> >  
>> >> >> > +static DEFINE_IDA(auxiliary_id);
>> >> >> 
>> >> >> I think this is the correct thing to do, even though the per device IDA drivers
>> >> >> typically went for so far produces IDs that are easier to handle when debugging
>> >> >> things.
>> >> >> 
>> >> >> > +
>> >> >> >  /**
>> >> >> >   * auxiliary_device_init - check auxiliary_device and initialize
>> >> >> >   * @auxdev: auxiliary device struct
>> >> >> > @@ -331,20 +333,37 @@ int __auxiliary_device_add(struct auxiliary_device *auxdev, const char *modname)
>> >> >> >  		return -EINVAL;
>> >> >> >  	}
>> >> >> >  
>> >> >> > +	ret = ida_alloc(&auxiliary_id, GFP_KERNEL);
>> >> >> > +	if (ret < 0) {
>> >> >> > +		dev_err(dev, "auxiliary device id_alloc fauiled: %d\n", ret);
>> >> >> > +		return ret;
>> >> >> > +	}
>> >> >> > +	auxdev->id = ret;
>> >> >> 
>> >> >> This overwrites the ID number set by various drivers that (still) use the
>> >> >> auxiliary_device_init() and auxiliary_device_add() pair.
>> >> >> 
>> >> >> While I agree with the general intent, I think it's a very bad idea to just
>> >> >> perform this change silently leaving drivers with their IDA instances not
>> >> >> knowing that the set ID numbers do not have an effect anymore.
>> >> >> 
>> >> >> I think this should be multiple steps:
>> >> >> 
>> >> >>   (1) Remove the id parameter and force an internal ID only for
>> >> >>       auxiliary_device_create().
>> >> >> 
>> >> >>   (2) Convert applicable drivers (and the Rust abstraction) to use
>> >> >>       auxiliary_device_create() rather than auxiliary_device_init() and
>> >> >>       auxiliary_device_add().
>> >> >> 
>> >> >>   (3) Treewide change to force an internal ID for all auxiliary devices
>> >> >>       considering this change in all affected drivers.
>> >> >
>> >> > I would suggest easier approach.
>> >> > 1. Add to the proposal patch, the sed generated line which removes auxdev->id
>> >> > assignment in the drivers.
>> >> > Something like this from mlx5:
>> >> >  - sf_dev->adev.id = id;
>> >> >
>> >> > 2. Add standalone patches to remove not used ida_alloc/ida_free calls
>> >> > from the drivers.
>> >> 
>> >> I assume you suggest this as an alternative to (3) above? If so, that's what I
>> >> meant in (3), I should have written "treewide series" instead of "treewide
>> >> change".
>> >
>> > I would say for all steps. Very important reason to use
>> > auxiliary_device_init() and not auxiliary_device_create() is to bind
>> > custom release callback, which is needed to release private data.
>> >
>> > In addition, complex devices embed struct auxiliary_device in their
>> > internal struct to rely on container_of to access the data.
>> > See mlx5_sf_dev_add() in drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
>> > as an example.
>> 
>> That's why I said "*applicable* drivers" everywhere. :)
>> 
>> The examples you mention don't fall under this category, but in general drivers
>> that *can* use auxiliary_device_create() should do it.
>
> Of course, the thing is that even drivers with auxiliary_device_init()
> shouldn't set "id" and because they need to be updated.

I think we agree here. :) It's only about the ordering between "convert to
auxiliary_device_create()" and "use internal IDA".

I think it'd be a good synergy to convert applicable drivers to
auxiliary_device_create() first, but I'm not insisting on it.

> The auxiliary_device_create() relies on auxiliary_device_init() under the hood,
> so most likely the change should be there.
>
>> 
>> >> Technically (2) is orthogonal, yet I think it's a bit better to do the desired
>> >> change right away. Otherwise we end up converting all applicable drivers to
>> >> implement the auxiliary device release callback (which we need for a common
>> >> ida_free()) first, just to remove it later on when we convert to
>> >> auxiliary_device_create().
>> >
>> > My expectation is to see extension of driver/base/core.c. Something like that:
>> >
>> > diff --git a/drivers/base/core.c b/drivers/base/core.c
>> > index cbc0099d8ef24..63847c84dbdc0 100644
>> > --- a/drivers/base/core.c
>> > +++ b/drivers/base/core.c
>> > @@ -2560,8 +2560,10 @@ static void device_release(struct kobject *kobj)
>> >
>> >         kfree(dev->dma_range_map);
>> >
>> > -       if (dev->release)
>> > +       if (dev->release) {
>> > +               dev->bus_specific_cleanup(dev);
>> >                 dev->release(dev);
>> > +       }
>> >         else if (dev->type && dev->type->release)
>> >                 dev->type->release(dev);
>> >         else if (dev->class && dev->class->dev_release)
>> 
>> The common pattern is to have custom release callbacks for class or bus specific
>> device types.
>> 
>> In this case drivers would set struct auxiliary_device::release. And the
>> auxiliary bus would implement the underlying struct device::release to call the
>> driver provided struct auxiliary_device::release plus the additional cleanup.
>> 
>> What you propose works as well, but it moves bus or class device specifics into
>> the generic struct device, where the normal inheritance pattern already solves
>> this.
>
> It was just a sketch, everything that allows to set custom release
> callback is fine by me.

Yeah, what I meant is that we shouldn't add an additional release callback to
struct device for such things in general.

device::release() is for the user of the struct device, in this case this is
struct auxiliary_device.

auxiliary_device::release() will be for the user of struct auxiliary_device,
which could be some generic driver specific device structure, let's say
struct foo_device.

Now, struct foo_device may have another release callback for its specific
cleanup.

So, the callchain would look like this:

	device::release {
		auxiliary_device::release {
			foo_device::release {
				// clean up struct foo_device
			}

			// clean up struct auxiliary_device
		}

		// clean up struct device
	}

Having additional release callbacks on struct device does not scale.

Powered by blists - more mailing lists