Re: [PATCH v2 0/3] leds: blink resolution improvements

[Jacek Anaszewski <j.anaszewski@samsung.com>]

On 04/28/2015 03:26 PM, Stas Sergeev wrote:
> 28.04.2015 15:58, Jacek Anaszewski пишет:
>> On 04/28/2015 12:12 PM, Stas Sergeev wrote:
>>> 28.04.2015 11:57, Jacek Anaszewski пишет:
>>>> Hi Stas,
>>>>
>>>> Have you tested it?
>>> Of course I did.
>>> Works with gpio driver and provides up to 10usec precision on
>>> armada-xp board.
>>> This is 1000 times better than without my patch - the precision
>>> was 10ms (jiffy).
>> Please take into account that this could work reliably only for gpio
>> LEDs. For the LEDs driven though a bus (e.g. I2C) delays below 1ms
>> might be hard to achieve. The minimum available delay would vary from
>> driver to driver.
>>
>> We could think of adding the hr_timer mode to the led-class.
>> The mode could be turned on with use of a new led_set_high_res_timer
>> API. The API would be called by drivers/leds/leds-gpio.c driver when
>> a dedicated sysfs attribute was set adequately.
>> The other drivers could also set this mode if they controlled device
>> with a suitable LED switching rate. The minimum delay value could
>> be made configurable by the driver and readable through sysfs
>> when in hr_timer mode.
> Why such a complexity?
> Wouldn't it be enough if the driver can set the minimum delay
> value, yet to always use the hrtimer?

Hr timers would be advantageous only for leds-gpio AFAICS. There is no
reason for making them default for all LED class drivers. I assume that
they provide higher resolution at a cost of consuming more CPU/system
resources.

> Please note that 10ms (jiffy) is an inadequate minimum delay
> pretty much for any driver, I guess. At least 1ms should be possible,
> because, well, you can write 1ms to sysfs attribute even without
> my patch. So even the existing options need hrtimer to work right.

As I mentioned before even with hr timer it wouldn't be possible
to assure that 1ms is achievable due to locking and bus latency
reasons.

>>>> I tried it with Samsung M0 board and
>>>> my leds-aat1290 driver. It didn't work well. And for small delay
>>>> intervals it will not have a chance to work reliably with all drivers,
>>>> especially the ones which use mutex in their brightness_set op,
>>>> since mutex can sleep.
>>> OK, I can remove the nsec resolution.
>> usec also didn't work, please look at my use case and warning:
>>
>> echo "timer" > trigger
>> echo 1 > delay_on
>> echo 1 > delay_off
>> echo usec > delay_unit
>> [  178.584433] hrtimer: interrupt took 300747 ns
> I think I should try a tasklet then.
>
>> Only some time later I realized that for AAT1290 brightness is set
>> through ASCwire pulse protocol, which takes few ms.
>>
>> Please note that with this approach users would have to wonder why
>> they are getting the warnings and why they can't get their LEDs to work
>> with given settings.
> OK, so we need a tasklet and the ability for the driver to set the
> minimal delay. I think I can implement the former and hope someone
> else will later implement the later. :)
> Lets just solve the problem step-by-step. I can't solve all the
> problems at once, but you can't deny the fact that the problem
> exists and needs to be solved.
>

This seems not to be an urgent issue. Therefore it is better to
provide complex solution in one patch set.


-- 
Best Regards,
Jacek Anaszewski