Re: [RFC PATCH 5/5] can: m_can: Add hrtimer to generate software interrupt

[Mendez, Judith]

Hello Marc,

On 4/17/2023 2:26 PM, Marc Kleine-Budde wrote:
> On 17.04.2023 19:34:03, Oliver Hartkopp wrote:
> > On 17.04.23 09:26, Marc Kleine-Budde wrote:
> > > On 16.04.2023 21:46:40, Oliver Hartkopp wrote:
> > > > > I had the 5ms that are actually used in the code in mind. But this is a
> > > > > good calculation.
> > > >
> > > > @Judith: Can you acknowledge the value calculation?
> > > >
> > > > > > The "shortest" 11 bit CAN ID CAN frame is a Classical CAN frame with DLC = 0
> > > > > > and 1 Mbit/s (arbitration) bitrate. This should be 48 bits @1Mbit => ~50
> > > > > > usecs
> > > > > >
> > > > > > So it should be something about
> > > > > >
> > > > > >        50 usecs * (FIFO queue len - 2)
> > > > >
> > > > > Where does the "2" come from?
> > > >
> > > > I thought about handling the FIFO earlier than it gets completely "full".
> > > >
> > > > The fetching routine would need some time too and the hrtimer could also
> > > > jitter to some extend.
> > >
> > > I was assuming something like this.
> > >
> > > I would argue that the polling time should be:
> > >
> > >       50 µs * FIFO length - IRQ overhead.
> > >
> > > The max IRQ overhead depends on your SoC and kernel configuration.
> >
> > I just tried an educated guess to prevent the FIFO to be filled up
> > completely. How can you estimate the "IRQ overhead"? And how do you catch
> > the CAN frames that are received while the IRQ is handled?
>
> We're talking about polling, better call it "overhead" or "latency from
> timer expiration until FIFO has at least one frame room". This value
> depends on your system.
>
> It depends on many, many factors, SoC, Kernel configuration (preempt RT,
> powersaving, frequency scaling, system load. In your example it's 100
> µs. I wanted to say there's an overhead (or latency) and we need enough
> space in the FIFO, to cover it.

I am not sure how to estimate IRQ overhead, but FIFO length should be 64
elements.

50 us * 62 is about 3.1 ms and we are using 1 ms timer polling interval.

Running a few benchmarks showed that using 0.5 ms timer polling interval 
starts to take a toll on CPU load, that is why I chose 1 ms polling 
interval.

regards,
Judith