Re: [PATCH] drivers: led: is31fl319x: 6/9-channel light effect led driver

[Jacek Anaszewski]

Hi Nikolaus,

On 07/06/2016 12:02 PM, H. Nikolaus Schaller wrote:
> Hi,
> finally, I found the time to update the driver according to the many comments
> received a while ago.
>
> Most of them have been worked in, including the regmap idea and
> brightness_set_blocking().
>
> The driver works on our system, so that I will mail [PATCH v2] as a followup.
>
> There is only one aspect of the new solution I am not sure if it is
> really better than our old proposal (see below).
>
>
> > Am 20.04.2016 um 23:04 schrieb Jacek Anaszewski <jacek.anaszewski@xxxxxxxxx>:
> >
> > On 04/19/2016 07:21 PM, H. Nikolaus Schaller wrote:
> > > > I believe the LEDS core now handles the workqueues generically for
> > > > blocking operations, so it's no longer needed in the individual drivers.
> > >
> > > We had a lot of trouble with locking and blocking especially if we
> > > want to indicate CPU or (root) disk activity.
> >
> > What kind of troubles you had? Could you share more details?
> > Does it mean that current LED core design doesn't fit for your
> > use cases?
>
> The system started to flicker the LEDs irregularily and sometimes
> the whole kernel stalled.
>
> > > So it is implemented in a way that changes can be requested faster
> > > than the I2C bus can write new values to the chip.
> > >
> > > Only after one sequence of I2C writes is done, another work function
> > > can be scheduled. And each group of writes updates as many LEDs
> > > in parallel if necessary.
> >
> > You can serialize the operations in brightness_set_blocking with
> > a mutex.
>
> Yes, that works fine in our (incomplete) test setup.
>
> But I think it assumes that the i2c bus is never congested by other i2c traffic.
>
> I have not found code that obviously takes care of the situation if led
> trigger events (e.g. mmc or cpu triggers) are coming in faster than the
> i2c (even using regmap) can write out over i2c.
>
> If I understand the led core code correctly, it will just do another schedule_work
> for every single change of led brightness.

Please look at schedule_work documentation in include/linux/workqueue.h:

/**
 * schedule_work - put work task in global workqueue
 * @work: job to be done
 *
 * Returns %false if @work was already on the kernel-global workqueue and
 * %true otherwise.
 *
 * This puts a job in the kernel-global workqueue if it was not already
 * queued and leaves it in the same position on the kernel-global
 * workqueue otherwise.
 */
static inline bool schedule_work(struct work_struct *work)
{
        return queue_work(system_wq, work);
}

> So I wonder if Is there some guarantee that this work queue will not fill
> up memory and is really processed faster than being filled? I.e. can the
> queue overflow?
>
> To reduce this risk, my original implementation strategy was different. The
> update speed was limited by i2c writing. A new register update batch job
> was only scheduled if the previous one is finished. If i2c was blocked/congested,
> the writing worker thread would come to a halt.
>
> All incoming led brightness changes were simply accumulated until a new
> batch job is started, because LEDs would anyways flicker invisibly fast.
>
> Tests with the new driver have shown that it seems not to run into this situation
> on our system but it might depend on factors we have not yet tested (slow i2c,
> other i2c traffic on the same bus, CPU speed, event types choosen).
>
> So I am a little in doubt about this risk. But I may have simply missed
> the reason why the standard approach works and can never overflow.
>
> BR,
> Nikolaus


--
Best regards,
Jacek Anaszewski