Re: [PATCH 1/2] pwm: sysfs: Set class on pwm devices

[Greg Kroah-Hartman]

On Fri, Oct 02, 2020 at 03:08:44PM +0200, Lars Poeschel wrote:
> On Fri, Oct 02, 2020 at 02:46:16PM +0200, Greg Kroah-Hartman wrote:
> > On Fri, Oct 02, 2020 at 02:30:47PM +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
> > > dev_uevent_filter function does filter-out devices without a bus or
> > > class.
> > > This was already addressed in commit
> > > commit 7e5d1fd75c3d ("pwm: Set class for exported channels in sysfs")
> > > but this did cause problems and the commit got reverted with
> > > commit c289d6625237 ("Revert "pwm: Set class for exported channels in
> > > sysfs"")
> > > Problem with the previous approach was, that there is a clash if we have
> > > multiple pwmchips:
> > > 	echo 0 > pwmchip0/export
> > > 	echo 0 > pwmchip1/export
> > > would both export /sys/class/pwm/pwm0 .
> > > 
> > > Now this patch changes the sysfs interface. We do include the pwmchip
> > > number into the pwm directory that gets exported.
> > > With the example above we get:
> > > 	/sys/class/pwm/pwm-0-0
> > > 	/sys/class/pwm/pwm-1-0
> > > We maintain ABI backward compatibility through symlinks.
> > > 	/sys/class/pwm/pwmchip0/pwm0
> > > 	/sys/class/pwm/pwmchip1/pwm0
> > > are now symbolic links to the new names.
> > > 
> > > Cc: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> > > Signed-off-by: Lars Poeschel <poeschel@xxxxxxxxxxx>
> > > ---
> > >  drivers/pwm/sysfs.c | 57 +++++++++++++++++++++++++++++++++++++--------
> > >  1 file changed, 47 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/drivers/pwm/sysfs.c b/drivers/pwm/sysfs.c
> > > index 449dbc0f49ed..c708da17a857 100644
> > > --- a/drivers/pwm/sysfs.c
> > > +++ b/drivers/pwm/sysfs.c
> > > @@ -240,8 +240,10 @@ static void pwm_export_release(struct device *child)
> > >  
> > >  static int pwm_export_child(struct device *parent, struct pwm_device *pwm)
> > >  {
> > > +	struct pwm_chip *chip = dev_get_drvdata(parent);
> > >  	struct pwm_export *export;
> > >  	char *pwm_prop[2];
> > > +	char *link_name;
> > >  	int ret;
> > >  
> > >  	if (test_and_set_bit(PWMF_EXPORTED, &pwm->flags))
> > > @@ -256,25 +258,39 @@ static int pwm_export_child(struct device *parent, struct pwm_device *pwm)
> > >  	export->pwm = pwm;
> > >  	mutex_init(&export->lock);
> > >  
> > > +	export->child.class = parent->class;
> > >  	export->child.release = pwm_export_release;
> > >  	export->child.parent = parent;
> > >  	export->child.devt = MKDEV(0, 0);
> > >  	export->child.groups = pwm_groups;
> > > -	dev_set_name(&export->child, "pwm%u", pwm->hwpwm);
> > > +	dev_set_name(&export->child, "pwm-%u-%u", chip->base, pwm->hwpwm);
> > >  
> > >  	ret = device_register(&export->child);
> > > -	if (ret) {
> > > -		clear_bit(PWMF_EXPORTED, &pwm->flags);
> > > -		put_device(&export->child);
> > > -		export = NULL;
> > > -		return ret;
> > > +	if (ret)
> > > +		goto error;
> > > +
> > > +	link_name = kasprintf(GFP_KERNEL, "pwm%u", pwm->hwpwm);
> > > +	if (link_name == NULL) {
> > > +		ret = -ENOMEM;
> > > +		goto dev_unregister;
> > >  	}
> > > -	pwm_prop[0] = kasprintf(GFP_KERNEL, "EXPORT=pwm%u", pwm->hwpwm);
> > > +
> > > +	pwm_prop[0] = kasprintf(GFP_KERNEL, "EXPORT=%s",
> > > +			export->child.kobj.name);
> > >  	pwm_prop[1] = NULL;
> > >  	kobject_uevent_env(&parent->kobj, KOBJ_CHANGE, pwm_prop);
> > 
> > Do you still need to do this by hand?  Why can't this uevent field
> > belong to the class and have it create this for you automatically when
> > the device is added?
> 
> I did not add this with my patch, it was there before and I wonder, what
> purpose it served, since the uevent was filtered because there was no
> class there.
> Now we have a class and now it works and this is what happens:
> 
> /sys/class/pwm# echo 0 > pwmchip1/export 
> KERNEL[2111.952725] add      /devices/platform/ocp/48302000.epwmss/48302200.pwm/pwm/pwmchip1/pwm-1-0 (pwm)
> ACTION=add
> DEVPATH=/devices/platform/ocp/48302000.epwmss/48302200.pwm/pwm/pwmchip1/pwm-1-0
> SEQNUM=1546
> SUBSYSTEM=pwm
> 
> KERNEL[2111.955155] change   /devices/platform/ocp/48302000.epwmss/48302200.pwm/pwm/pwmchip1 (pwm)
> ACTION=change
> DEVPATH=/devices/platform/ocp/48302000.epwmss/48302200.pwm/pwm/pwmchip1
> EXPORT=pwm-1-0
> SEQNUM=1547
> SUBSYSTEM=pwm
> 
> The first event is the event from device_register. It informs us that we
> now have a new pwm-1-0. Nice.
> The second is the event done here "by hand". It informs us, that
> pwmchip1 changed. It has a new export now. For me personally this is not
> needed, but also I don't think it is wrong.
> You decide!

If the uevent was being filtered out anyway, and never sent, then let's
just drop the thing as there is nothing to keep backwards compatible.

thanks,

greg k-h