On 05/05/2015 03:02 PM, Stas Sergeev wrote:
> 05.05.2015 11:22, Jacek Anaszewski пишет:
>> On 05/04/2015 07:20 PM, Stas Sergeev wrote:
>>> 04.05.2015 18:22, Jacek Anaszewski пишет:
>>>> On 05/04/2015 02:12 PM, Stas Sergeev wrote:
>>>>> Only under that condition:
>>>>> ---
>>>>> if (led_cdev->blink_delay_on || led_cdev->blink_delay_off) {
>>>>> led_cdev->delayed_set_value = brightness;
>>>>> schedule_work(&led_cdev->set_brightness_work);
>>>>> ---
>>>>>
>>>>> But the main condition is:
>>>>> ---
>>>>> if (led_cdev->flags & SET_BRIGHTNESS_ASYNC) {
>>>>> led_set_brightness_async(led_cdev, brightness);
>>>>> ---
>>>>>
>>>>> So I think it is actually unused.
>>>>> I don't see why schedule_work() above can't be just replaced
>>>>> with led_set_brightness_async(). Is there the reason not to do so?
>>>> set_brightness_work not only sets the brightness but also
>>>> stops software blinking, which was the primary reason
>>>> for adding this work queue I think. Here is the commit message:
>>> But led_trigger_set() does led_stop_software_blink(), which
>>> IMHO means led_set_brightness() will in most cases be called
>>> when sw blocking is already stopped. There seem to be just a
>>> few cases where this is not true: oneshot_trig_deactivate() and
>>> timer_trig_deactivate(), and I think I'll just change these two to
>>> led_stop_software_blink(). I am pretty sure the work-queue is
>>> not needed, but I'll have to test that with the patch it seems.
>> It is used e.g. in the following case:
>>
>> #echo "timer" > trigger
>> #echo 1 > brightness
> Indeed, thanks.
> I'll study that case next week when my board is back to me.
> Looking at sources, it seems in that case it would disable the
> software blinking (del_timer_sync()) without changing the
> trigger back to "none", which does not make sense to me.

Yes, this needs to be fixed.

>>
>>>
>>>> > Now your leds-aat1290 already asks for such a change,
>>>>> because it can sleep but does not use a work-queue the
>>>>> way other drivers do.
>>>> It doesn't need this change - it defines two ops: brightness_set
>>>> (the async one) and brightness_set_sync (the sync one). The
>>>> former is called from led_set_brightness_async and the latter
>>>> form led_set_brightness_sync.
>>>> led_set_brightness_async is called from led_set_brightness
>>>> for drivers that define SET_BRIGHTNESS_ASYNC flag and
>>>> led_set_brightness_sync for the drivers that define
>>>> SET_BRIGHTNESS_SYNC flags.
>>>>
>>>> led_timer_function calls always led_set_brightness_async.
>>> OK, I googled the patch:
>>> https://lkml.org/lkml/2015/3/4/960
>>> So the async one uses the work-queue, and the sync one
>>> does not. Since led_timer_function calls always
>>> led_set_brightness_async,
>>> it should always be using a work-queue.
>>> But then I fail to explain your diagnostic that with my patch and
>>> your driver, the hrtimer gives warning about a high interrupt
>>> latency. I thought this is because your driver does sleeps and
>>> does not use a work queue. Its not the case. Could you please
>>> clarify, what then caused the high interrupt latency warning in
>>> your testing?
>> An accurate explanation would require thorough investigation.
>> It can be related to the fact that the driver uses delays.
> Even if your driver just does schedule_work() and nothing
> more in an async method? Strange.

There can be indirect correlation.

>> In the first place we have to take into account that Linux is not
>> a real time operating system. The feature you're trying to implement
>> is realized by hardware with use of pwm. There might be narrow group
>> of drivers that could benefit from it in specific circumstances
>> (the system couldn't be too busy at the time when timer trigger is
>> running), but this is too weak argument in favour of supporting small
>> delay intervals.
> If you mean the drivers that don't have any sleeps, then the
> system load is irrelevant because the hrtimer callback is AFAIK
> running in an irq context. So for them it would be a clear win,
> not just in a specific circumstances. Of course I wonder if it is
> only leds-gpio, or anything else too. :) Though I could suspect
> that leds-gpio have a very wide usage, and it may worth the
> troubles even to improve just leds-gpio alone.

If you have a strong belief that it is possible to implement this
feature in a manner acceptable for everyone, feel free to experiment
with the implementation. If people will find it useful and reliable
then we will merge it.

--
Best Regards,
Jacek Anaszewski

2015-05-11 22:11:44

by Pavel Machek

[permalink] [raw]

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

> >>- You will set the delay in ms units. For example for 100us you'll
> >>write 0.1. IMHO it is counter-intuitive: people will make a mistake
> >>and try 0.0001 instead, wrongly assuming that this is in seconds.
> >>And nanoseconds should then better be removed, as writing
> >>nanosecond delay will just require too much zeros.
> >This is machine-to-machine interface. And users can handle this.
> OK. I think you mostly convinced me that this solution is not
> that bad. :) But I am going to add an explicit "Acked-by" then,
> to emphasize that it is your idea and not mine.

Feel free to note that it is my idea, but I'd have to see&like the
patch before you can add Acked-by.

Thanks,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html