Message-ID: <20250306102540.7f0f6146@erd003.prtnl>
Date: Thu, 6 Mar 2025 10:25:40 +0100
From: David Jander <david@...tonic.nl>
To: Uwe Kleine-König <u.kleine-koenig@...libre.com>
Cc: linux-kernel@...r.kernel.org, linux-iio@...r.kernel.org, Jonathan Corbet
 <corbet@....net>, Rob Herring <robh@...nel.org>, Krzysztof Kozlowski
 <krzk+dt@...nel.org>, Conor Dooley <conor+dt@...nel.org>,
 devicetree@...r.kernel.org, linux-doc@...r.kernel.org, Nuno Sa
 <nuno.sa@...log.com>, Jonathan Cameron <jic23@...nel.org>, Oleksij Rempel
 <o.rempel@...gutronix.de>, Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Subject: Re: [RFC PATCH 1/7] drivers: Add motion control subsystem

On Thu, 6 Mar 2025 00:21:22 +0100
Uwe Kleine-König <u.kleine-koenig@...libre.com> wrote:

> Hello David,
> 
> On Wed, Mar 05, 2025 at 04:40:45PM +0100, David Jander wrote:
> > On Fri, 28 Feb 2025 17:44:27 +0100
> > Uwe Kleine-König <u.kleine-koenig@...libre.com> wrote:  
> > > On Thu, Feb 27, 2025 at 05:28:17PM +0100, David Jander wrote:
> > > [...]  
> > > > +static int motion_open(struct inode *inode, struct file *file)
> > > > +{
> > > > +	int minor = iminor(inode);
> > > > +	struct motion_device *mdev = NULL, *iter;
> > > > +	int err;
> > > > +
> > > > +	mutex_lock(&motion_mtx);    
> > > 
> > > If you use guard(), error handling gets a bit easier.  
> > 
> > This looks interesting. I didn't know about guard(). Thanks. I see the
> > benefits, but in some cases it also makes the locked region less clearly
> > visible. While I agree that guard() in this particular place is nice,
> > I'm hesitant to try and replace all mutex_lock()/_unlock() calls with guard().
> > Let me know if my assessment of the intended use of guard() is incorrect.  
> 
> I agree that guard() makes it harder for non-trivial functions to spot
> the critical section. In my eyes this is outweight by not having to
> unlock in all exit paths, but that might be subjective. Annother
> downside of guard is that sparse doesn't understand it and reports
> unbalanced locking.

What I was referring to, and what I want to know is, is it okay to mix guard()
with lock/unlock? I.e. Use guard() when there are multiple exit paths involved
and revert back to simple lock/unlock if it is just to encase a handful of
non-exiting operations?

> > > > +	list_for_each_entry(iter, &motion_list, list) {
> > > > +		if (iter->minor != minor)
> > > > +			continue;
> > > > +		mdev = iter;
> > > > +		break;
> > > > +	}    
> > > 
> > > This should be easier. If you use a cdev you can just do
> > > container_of(inode->i_cdev, ...);  
> > 
> > Hmm... I don't yet really understand what you mean. I will have to study the
> > involved code a bit more.  
> 
> The code that I'm convinced is correct is
> https://lore.kernel.org/linux-pwm/00c9f1181dc351e1e6041ba6e41e4c30b12b6a27.1725635013.git.u.kleine-koenig@baylibre.com/
> 
> This isn't in mainline because there is some feedback I still have to
> address, but I think it might serve as an example anyhow.

Thanks. I will study this example.

> > > > [...]
> > > > +
> > > > +static const struct class motion_class = {
> > > > +	.name		= "motion",
> > > > +	.devnode	= motion_devnode,    
> > > 
> > > IIRC it's recommended to not create new classes, but a bus.  
> > 
> > Interesting. I did some searching, and all I could find was that the chapter
> > in driver-api/driver-model about classes magically vanished between versions
> > 5.12 and 5.13. Does anyone know where I can find some information about this?
> > Sorry if I'm being blind...  
> 
> Half knowledge on my end at best. I would hope that Greg knows some
> details (which might even be "no, classes are fine"). I added him to Cc:
> 
> > > [...]  
> > > > +	devt = MKDEV(motion_major, mdev->minor);
> > > > +	mdev->dev = device_create_with_groups(&motion_class, mdev->parent,
> > > > +				devt, mdev, mdev->groups, "motion%d", mdev->minor);    
> > > 
> > > What makes sure that mdev doesn't go away while one of the attributes is
> > > accessed?  
> > 
> > Good question. I suppose you mean that since mdev is devres-managed and
> > device_create_with_groups() apparently isn't aware of that, so there is no
> > internal lock somewhere that prevents read() or ioctl() being called while the
> > devres code is freeing the memory of mdev?  
> 
> I'm not sure there is an issue, but when I developed the above mentioned
> patch it helped me to test these possible races. Just open the sysfs
> file, unbind the device (or unload the module) and only then start
> reading (or writing).

Will check this. Thanks.

> > > > +	if (IS_ERR(mdev->dev)) {
> > > > +		dev_err(mdev->parent, "Error creating motion device %d\n",
> > > > +				mdev->minor);
> > > > +		mutex_unlock(&motion_mtx);
> > > > +		goto error_free_trig_info;
> > > > +	}
> > > > +	list_add_tail(&mdev->list, &motion_list);
> > > > +	mutex_unlock(&motion_mtx);
> > > > +
> > > > +	return 0;
> > > > +
> > > > +error_free_trig_info:
> > > > +	kfree(trig_info);
> > > > +error_free_trigger:
> > > > +	iio_trigger_free(mdev->iiotrig);
> > > > +error_free_minor:
> > > > +	motion_minor_free(mdev->minor);
> > > > +	dev_info(mdev->parent, "Registering motion device err=%d\n", err);
> > > > +	return err;
> > > > +}
> > > > +EXPORT_SYMBOL(motion_register_device);
> > > > [...]
> > > > +struct mot_capabilities {
> > > > +	__u32 features;
> > > > +	__u8 type;
> > > > +	__u8 num_channels;
> > > > +	__u8 num_int_triggers;
> > > > +	__u8 num_ext_triggers;
> > > > +	__u8 max_profiles;
> > > > +	__u8 max_vpoints;
> > > > +	__u8 max_apoints;
> > > > +	__u8 reserved1;
> > > > +	__u32 subdiv; /* Position unit sub-divisions, microsteps, etc... */
> > > > +	/*
> > > > +	 * Coefficients for converting to/from controller time <--> seconds.
> > > > +	 * Speed[1/s] = Speed[controller_units] * conv_mul / conv_div
> > > > +	 * Accel[1/s^2] = Accel[controller_units] * conv_mul / conv_div
> > > > +	 */
> > > > +	__u32 speed_conv_mul;
> > > > +	__u32 speed_conv_div;
> > > > +	__u32 accel_conv_mul;
> > > > +	__u32 accel_conv_div;
> > > > +	__u32 reserved2;
> > > > +};    
> > > 
> > > https://docs.kernel.org/gpu/imagination/uapi.html (which has some
> > > generic bits that apply here, too) has: "The overall struct must be
> > > padded to 64-bit alignment." If you drop reserved2 the struct is
> > > properly sized (or I counted wrongly).  
> > 
> > Oh, thanks for pointing that out... I wouldn't have searched for that
> > information in that particular place tbh. ;-)
> > 
> > I am tempted to add another __u32 reserved3 though instead. Better to have
> > some leeway if something needs to be added in a backwards-compatible way later.  
> 
> Note that you don't need reserved fields at the end because in the
> ioctl handler you know the size of the passed struct. So if the need to
> add members to the struct arise, you can do that by checking for the
> size. This is even more flexible because otherwise you can only add
> fields that must be 0 when the old behaviour is intended. Most of the
> time this is no problem. But only most.

You are right. Thanks.

> > > > +struct mot_speed_duration {
> > > > +	__u32 channel;
> > > > +	speed_raw_t speed;    
> > > 
> > > What is the unit here?  
> > 
> > Speed doesn't have a fixed unit in this case. Or rather, the unit is
> > device-dependent. For a motor it could be rotations per second, micro-steps per
> > second, etc... while for a linear actuator, it could be micrometers per second.
> > 
> > Why no fixed unit? That's because in practice many devices (controllers) have
> > their inherent base-unit, and it would get overly complicated if one needed to
> > convert back and forth between that and some universal unit just for the sake
> > of uniformity, and user-space most certainly expects the same unit as the
> > hardware device it was initially designed for. So in this case it is a design
> > decision to make user-space deal with unit-conversion if it is necessary to do
> > so.  
> 
> Sad, so a userspace process still has to know some internal things about
> the motor it drives. :-\

Unfortunately that is almost impossible to avoid entirely.
You can replace one stepper motor driver with another that might have
different micro-stepping subdivision, by looking at struct
mot_capabilities.subdiv, but a simple brushed DC motor just isn't able to
replace a stepper motor in all but the most trivial applications. I also think
that burdening the kernel with all sorts of complicated math to model the
mechanical conversion factors involved in anything that's connected to the
motor drive shaft is overkill. As well as trying to emulate all missing
capabilities from a motion device that is lacking that functionality natively.

So just like in IIO you cannot just replace one ADC with any other, in LMC you
also cannot replace any device with any other.

That's why there is struct mot_capabilities and MOT_IOCTL_GET_CAPA. It enables
user-space to optionally support different devices more easily. It is probably
best used in conjunction with a LMC user-space library, although I don't want
to rely on such a library for being able to use LMC. There is some middle
ground here I guess... just like in IIO.

One thing I could try to improve though, is to include some additional
information in struct mot_capabilities that tells something more about the
nature of the used units, just like the speed_conv and accel_conv constants do
for time conversion. Something that can be placed in the device tree (possibly
in a motor child-node connected to the motor-controller) that contains some
conversion constant for distance. That way, if one were to (for example)
replace a stepper motor with a BLDC motor + encoder in a new hardware
revision, this constant could be used to make the units backwards compatible.

As background information: A stepper motor controller counts distance in steps
and/or micro-steps. There are mot_capabilities.subdiv micro-steps in each
step. The amount of angle the actual motor shaft advances with each whole step
depends on the motor construction and is often 200 steps per revolution (1.8
degrees), but can vary from 4 to 400 steps per revolution depending on the
motor. So it is not only the controller that matters but also the type of
motor. This suggests the need of motor sub-nodes in the device-tree if one
wanted to extend the hardware knowledge further down from the motor driver.
But then there are gear boxes, pulleys, etc... it's basically conversion
factors all the way down. How many of them is sensible to bother the kernel
with?

Best regards,

-- 
David Jander

Powered by blists - more mailing lists