On Thu, 2014-08-14 at 07:39 +0300, Sakari Ailus wrote:

Bryan and Richard,

Your opinion would be much appreciated to a question myself and Jacek were
pondering. Please see below.

On Thu, Aug 07, 2014 at 03:12:09PM +0200, Jacek Anaszewski wrote:

quoted

Hi Sakari,

On 08/04/2014 02:50 PM, Sakari Ailus wrote:

quoted

Hi Jacek,

Thank you for your continued efforts on this!

On Mon, Aug 04, 2014 at 02:35:26PM +0200, Jacek Anaszewski wrote:

quoted

On 07/16/2014 11:54 PM, Sakari Ailus wrote:

quoted

Hi Jacek,

Jacek Anaszewski wrote:
...

quoted

diff --git a/include/linux/leds.h b/include/linux/leds.h
index 1a130cc..9bea9e6 100644
--- a/include/linux/leds.h
+++ b/include/linux/leds.h

@@ -44,11 +44,21 @@ struct led_classdev {
 #define LED_BLINK_ONESHOT_STOP    (1 << 18)
 #define LED_BLINK_INVERT    (1 << 19)
 #define LED_SYSFS_LOCK        (1 << 20)
+#define LED_DEV_CAP_TORCH    (1 << 21)

     /* Set LED brightness level */
     /* Must not sleep, use a workqueue if needed */
     void        (*brightness_set)(struct led_classdev *led_cdev,
                       enum led_brightness brightness);
+    /*
+     * Set LED brightness immediately - it is required for flash led
+     * devices as they require setting torch brightness to have

immediate
+     * effect. brightness_set op cannot be used for this purpose because
+     * the led drivers schedule a work queue task in it to allow for
+     * being called from led-triggers, i.e. from the timer irq context.
+     */

Do we need to classify actual devices based on this? I think it's rather
a different API behaviour between the LED and the V4L2 APIs.

On devices that are slow to control, the behaviour should be asynchronous
over the LED API and synchronous when accessed through the V4L2 API. How
about implementing the work queue, as I have suggested, in the
framework, so
that individual drivers don't need to care about this and just implement
the
synchronous variant of this op? A flag could be added to distinguish
devices
that are fast so that the work queue isn't needed.

It'd be nice to avoid individual drivers having to implement multiple
ops to
do the same thing, just for differing user space interfacs.

It is not only the matter of a device controller speed. If a flash
device is to be made accessible from the LED subsystem, then it
should be also compatible with led-triggers. Some of led-triggers
call brightness_set op from the timer irq context and thus no
locking in the callback can occur. This requirement cannot be
met i.e. if i2c bus is to be used. This is probably the primary
reason for scheduling work queue tasks in brightness_set op.

Having the above in mind, setting a brightness in a work queue
task must be possible for all LED Class Flash drivers, regardless
whether related devices have fast or slow controller.

Let's recap the cost of possible solutions then:

1) Moving the work queues to the LED framework

  - it would probably require extending led_set_brightness and
    __led_set_brightness functions by a parameter indicating whether it
    should call brightness_set op in the work queue task or directly;
  - all existing triggers would have to be updated accordingly;
  - work queues would have to be removed from all the LED drivers;

2) adding led_set_torch_brightness API

  - no modifications in existing drivers and triggers would be required
  - instead, only the modifications from the discussed patch would
    be required

Solution 1 looks cleaner but requires much more modifications.

How about a combination of the two, i.e. option 1 with the old op remaining
there for compatibility with the old drivers (with a comment telling it's
deprecated)?

This way new drivers will benefit from having to implement this just once,
and modifications to the existing drivers could be left for later.

It's OK for me, but the opinion from the LED side guys is needed here
as well.

Ping.

I'm not a fan of forcing everything to the lowest common denominator. At
a basic level an LED is a binary on/off so the shear levels of
complexity we end up going through is kind of scary. Forcing everything
through a workqueue due to their being some hardware out there can can't
do it in interrupt context is kind of sad and wasn't where the API set
out from.

So personally I'd prefer not to see the API changed like that however I
appreciate I've been more distant from the code of late.

Cheers,

Richard