Re: [PATCH] pwm: sysfs: Set class on pwm devices

[Greg Kroah-Hartman]

On Mon, Oct 05, 2020 at 11:30:16AM +0200, Thierry Reding wrote:
> On Thu, Oct 01, 2020 at 01:24:49PM +0200, Greg Kroah-Hartman wrote:
> > On Thu, Oct 01, 2020 at 11:05:31AM +0200, Lars Poeschel wrote:
> > > On Wed, Sep 30, 2020 at 11:41:46AM +0200, Uwe Kleine-König wrote:
> > > > Hello,
> > > > 
> > > > I added Greg Kroah-Hartman who I discussed this with via irc a bit to
> > > > Cc:.
> > > > 
> > > > On Wed, Sep 30, 2020 at 11:20:56AM +0200, Lars Poeschel wrote:
> > > > > thank you for your review!
> > > > > 
> > > > > On Wed, Sep 30, 2020 at 08:57:26AM +0200, Uwe Kleine-König wrote:
> > > > > > On Tue, Sep 29, 2020 at 02:19:53PM +0200, poeschel@xxxxxxxxxxx wrote:
> > > > > > > From: Lars Poeschel <poeschel@xxxxxxxxxxx>
> > > > > > > 
> > > > > > > This adds a class to exported pwm devices.
> > > > > > > Exporting a pwm through sysfs did not yield udev events. The
> > > > > > 
> > > > > > I wonder what is your use-case here. This for sure also has a place to
> > > > > > be mentioned in the commit log. I suspect there is a better way to
> > > > > > accomplish you way.
> > > > > 
> > > > > Use-case is to be able to use a pwm from a non-root userspace process.
> > > > > I use udev rules to adjust permissions.
> > > > 
> > > > Hmm, how do you trigger the export? Without being aware of all the
> > > > details in the sysfs code I would expect that the exported stuff is
> > > > available instantly once the write used to export the PWM is completed.
> > > > So changing the permissions can be done directly after triggering the
> > > > export in the same process.
> > > 
> > > The export is triggered through the userspace process itself. Why can it
> > > do this ? Because there is another udev rule, that changes permissions
> > > when a pwmchip appears.
> > > Then I'd like to have the second udev rule, that changes permissions on
> > > the freshly exported pwm. The userspace process can't do this.
> > > You are right I could propably do everything from within udev: If a
> > > pwmchip appears, export certain pwms and right away change their
> > > permissions. It does not also not feel right. It'd require knowledge
> > > from the userspace application to be mapped to udev.
> > 
> > The way the kernel code is now, yes, you will not have any way to
> > trigger it by userspace as the kernel is creating a "raw" struct device
> > that isn't assigned to anything.  That is what needs to be fixed here.
> > 
> > > > Out of interest: What do you use the pwm for? Isn't there a suitable
> > > > kernel driver that can do the required stuff? Compared to the kernel-API
> > > > the sysfs interface isn't atomic. Is this an annoyance?
> > > 
> > > Use-case is generating a voltage from the pwm. This voltage is used to
> > > signal different states and does not change very often. This is
> > > absolutely not annoying that this is not atomic. We just change the duty
> > > cycle on the fly. Everything else is configured one time at startup.
> > > I'd call what I need pwm-dac. I could not find a ready to use driver.
> > > Maybe I could misuse some kernel driver for this. Maybe I could use
> > > pwm-led or pwm-brightness or pwm-fan. Propably pwm-regulator could work,
> > > there is even a userspace facing part for this, but this is not
> > > devicetree ready.
> > > ...and the worst, please don't blame me: The application is java, so
> > > ioctl is a problem.
> > 
> > I thought java could do ioctls, otherwise how would it ever be able to
> > talk to serial ports?
> > 
> > Anyway, this needs to be fixed in the kernel...
> 
> If atomicity was a problem, we could potentially add a mechanism to the
> sysfs interface to enable that. I don't see a good way of doing that in
> a single file, since that works against how sysfs is designed. But one
> thing I could imagine is adding a file ("lock", or whatever you want to
> call it) that you can use for atomic fencing:
> 
> 	$ echo 1 > lock # locks the hardware state
> 	$ echo 100 > period
> 	$ echo 50 > duty_cycle
> 	$ echo 0 > lock # flushes state to hardware
> 
> But it sounds like that's not even a big issue.

That is exactly what configfs was designed for :)

thanks,

greg k-h