lp50xx: LED banking appears to be broken

["Jan Kundrát" <jan.kundrat@cesnet.cz>]

Hi there,
I'm trying to use the LP5009 chip with the following HW setup:

- channels 1-6 drive a big, 20mm LED module which internally consists of 
six independent LEDs
- channels 7, 8 and 9 drive a RGB LED as usual.

I thought that a DT bindings like this will work:

	led-controller@0c {
		compatible = "ti,lp5009";
		reg = <0x0c>;
		#address-cells = <1>;
		#size-cells = <0>;
		status = "okay";

		multi-led@1 {
			#address-cells = <1>;
			#size-cells = <0>;
			reg = <1>;
			label = "tally:1";
			led {
				color = <LED_COLOR_ID_RED>;
			};
		};
		multi-led@2 {
			#address-cells = <1>;
			#size-cells = <0>;
			reg = <2>;
			label = "tally:2";
			led {
				color = <LED_COLOR_ID_RED>;
			};
		};
		multi-led@3 {
			#address-cells = <1>;
			#size-cells = <0>;
			reg = <3>;
			label = "tally:3";
			led {
				color = <LED_COLOR_ID_RED>;
			};
		};
		multi-led@4 {
			#address-cells = <1>;
			#size-cells = <0>;
			reg = <4>;
			label = "tally:4";
			led {
				color = <LED_COLOR_ID_RED>;
			};
		};
		multi-led@5 {
			#address-cells = <1>;
			#size-cells = <0>;
			reg = <5>;
			label = "tally:5";
			led {
				color = <LED_COLOR_ID_RED>;
			};
		};
		multi-led@6 {
			#address-cells = <1>;
			#size-cells = <0>;
			reg = <6>;
			label = "tally:6";
			led {
				color = <LED_COLOR_ID_RED>;
			};
		};

		multi-led@7 {
			#address-cells = <1>;
			#size-cells = <2>;
			reg = <7 8 9>;
			color = <LED_COLOR_ID_RGB>;
			label = "preview";

			led@7 {
				color = <LED_COLOR_ID_RED>;
			};
			led@8 {
				color = <LED_COLOR_ID_GREEN>;
			};
			led@9 {
				color = <LED_COLOR_ID_BLUE>;
			};
		};
	};

This has drawbacks:

- I get a multicolor sysfs entry for each of the six red sub-LEDs, which 
probably doesn't make much sense. I cannot do a one "multicolor" LED with 
six channels because there appears to be a limit of 3 channels, and because 
the order of channels is documented to be non-deterministic, so that would 
require me to come up with fake names or something. Also, driving this from 
userspace means two writes for each sub-LED.

- The sysfs entries do not appear to drive correct LEDs. For example, a 
write to tally:5 or tally:6 results in an error:

 lp50xx 1-000c: Cannot write intensity value -5
 leds tally:6: Setting an LED's brightness failed (-5)

I tried to simplify this, and kept just the one RGB LED (that is, the 
multi-led@7 and led@7, led@8 and led@9 stanzas). This resulted in the 
following regmap entries after init:

# cat /sys/kernel/debug/regmap/1-000c/registers 
00: 40
01: 3c
02: 80
03: ff
04: 0f
05: 0f
06: 0f
07: 0f
08: ff
09: ff
0a: ff
0b: 0f
0c: 00
0d: 00
0e: 00
0f: 00
10: 00
11: 00
12: 00
13: 00
14: 00
15: 00
16: 00
17: ff

Clearly, that's wrong because it sets register's 0x02 reserved bits to 
non-zero. It looks as if the LED's channel number gets translated to the 
bank number, which is wrong. There are nine LEDs on LP5009, but only three 
individual banks.

Also, I don't think that the concept of "banks" as defined in LP50xx chips 
should be used in the Linux driver. The datasheet is not terribly specific 
on details, but it looks to me that the "banks" are for a use case where 
multiple physical LEDs are to, e.g., "breathe together". The chip indeed 
imposes some limitations when banking is enabled:

- LED0 will always be on channels 1-3. That's incompatible with the current 
code which uses the `reg` DT property and allows arbitrary assigning of 
channels to a LED's color inputs. I can have a Linux RGB LED which uses 
channels 1, 5 and 9 just fine, but I cannot use banking for that.

- Bank A always drives the first color of all LEDs that have banking 
enabled. Bank B is always for the second color, and bank C always applies 
to the third color.

As far as I can tell, there's no support for cross-LED control in Linux, so 
I think that we can just rip support for banking from this driver. The main 
motivation appears to be saving some I2C bandwidth and MCU cycles. If the 
driver was serious about this, it would use register auto-increment as a 
first step I suppose, but the regmap subsystem as-is updates all registers 
independently.

Before I send a patch which implements all that, I wanted to ask if I 
understood everything right, and to check whether these suggestions make 
sense to the maintainer and to the original author of the driver (and, 
hopefully, to the users as well).

With kind regards,
Jan