how fast can you run a hrtimer?

how fast can you run a hrtimer?

[Cliff Brake]

Hi,

I have an application where we need to tweak a PWM setting every 10uS.
My impression is this is beyond the capabilities of linux+rt, but want
to verify I'm not missing anything before we add an external
microcontroller to handle this.

I've run some experiments on a i.MX6UL with v4.1+rt, and observe
delays of ~50us when toggling a GPIO in a Linux driver using a
hrtimer.

Thanks,
Cliff

-- 
=================
http://bec-systems.com

Re: how fast can you run a hrtimer?

[Cliff Brake]

After a suggestion of using a NMI and a bit of research, it seems
something like the following article is the probably a reasonable
approach:

http://free-electrons.com/blog/fiq-handlers-in-the-arm-linux-kernel/

Cliff

On Fri, Aug 18, 2017 at 8:35 AM, Cliff Brake <cliff.brake@gmail.com> wrote:
> Hi,
>
> I have an application where we need to tweak a PWM setting every 10uS.
> My impression is this is beyond the capabilities of linux+rt, but want
> to verify I'm not missing anything before we add an external
> microcontroller to handle this.
>
> I've run some experiments on a i.MX6UL with v4.1+rt, and observe
> delays of ~50us when toggling a GPIO in a Linux driver using a
> hrtimer.
>
> Thanks,
> Cliff
>
> --
> =================
> http://bec-systems.com



-- 
=================
http://bec-systems.com

Re: how fast can you run a hrtimer?

[Julia Cartwright]

[-- Attachment #1: Type: text/plain, Size: 721 bytes --]

On Fri, Aug 18, 2017 at 10:34:19AM -0400, Cliff Brake wrote:
> After a suggestion of using a NMI and a bit of research, it seems
> something like the following article is the probably a reasonable
> approach:
> 
> http://free-electrons.com/blog/fiq-handlers-in-the-arm-linux-kernel/

The use of a FIQ probably gives you the best chance to achieve the
latencies you want; but keep in mind that the use of FIQ means you'll
need to get your hands pretty dirty.

The set of operations legal to perform from NMI context is very, very
small.  The usage of gpiolib (and therefore any existing i.MX gpio
driver) is probably not going to work from NMI context, so you'll have
to roll your own solution.

   Julia

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: how fast can you run a hrtimer?

[afzal mohammed]

Hi

> On Fri, Aug 18, 2017 at 8:35 AM, Cliff Brake <cliff.brake@gmail.com> wrote:

> > I have an application where we need to tweak a PWM setting every 10uS.
> > My impression is this is beyond the capabilities of linux+rt, but want
> > to verify I'm not missing anything before we add an external
> > microcontroller to handle this.
> >
> > I've run some experiments on a i.MX6UL with v4.1+rt, and observe
> > delays of ~50us when toggling a GPIO in a Linux driver using a
> > hrtimer.

On Fri, Aug 18, 2017 at 10:34:19AM -0400, Cliff Brake wrote:
> After a suggestion of using a NMI and a bit of research, it seems
> something like the following article is the probably a reasonable
> approach:
> 
> http://free-electrons.com/blog/fiq-handlers-in-the-arm-linux-kernel/

If a PWM IP is present on the chip, won't updating the PWM compare
registers within 10us (PWM @100KHz) achieve the required ?, toggling
delay & jitter shouldn't be a worry here, right ? (assuming PWM
compare registers can be updated every 10us using hrtimers)

afzal

Re: how fast can you run a hrtimer?

[Cliff Brake]

On Tue, Aug 22, 2017 at 8:45 AM, afzal mohammed <afzal.mohd.ma@gmail.com> wrote:

> If a PWM IP is present on the chip, won't updating the PWM compare
> registers within 10us (PWM @100KHz) achieve the required ?, toggling
> delay & jitter shouldn't be a worry here, right ? (assuming PWM
> compare registers can be updated every 10us using hrtimers)
>

There is timers on the cpu that can generate a PWM in hardware, but
what we need to do is update the PWM duty cycle every 10us (so the PWM
freq will be say 1MHz).  The idea is to ramp up/down the PWM value
over time (waveform generation).  I suppose if we could somehow feed
the capture/compare registers with DMA, that might work.  Gives me an
idea -- we should look into audio hardware and see if any of that can
generate a 100KHz waveform.

Cliff

-- 
=================
http://bec-systems.com