Re: c_can: (newbie) high system load when frame not acked?

[Andri Yngvason <andri.yngvason@marel.com>]

Quoting Viktor Babrian (2015-01-13 15:10:29)
> > Maybe you can turn the error interrupts off and still have it function. Maybe 
> > it won't (may be required for bus off recovery).
> 
> Thanks for the detailed answer. I read the pages and took a look into the 
> code.
> 
> Regarding state-change behavior, DCAN controller is similar to the one in 
> i.MX6. (It has explicit interrupt source for Act->Warn and Pass->BOFF 
> transitions).
> Although as I understand CAN, you have to have a successful RX or TX event 
> to have the counters decrease. That  means that you will have an immediate 
> report (i.e. interrupt) on Warn->Active, Pass->Warn and BOFF->Active. As 
> far as I see, this latter is explicitly handled by upper layers so driver 
> will have to know it anyway.
Correct.
> This means that if I turn off bus error monitoring, the only event I will 
> not get immediate report of is Warn->Pass. (Pass->Warn can be reported on 
> successful message events if the driver is implemented accordingly. 
Also correct.
> Currently it isn't). However it could be reported later upon the first 
> successful message event. I beleive missing an accurate report of 
> active->passive state change is far better than having a non-responding 
> system in the field in scenarios that do happen.
I agree.

Note: The FlexCAN driver does actually respect berr-reporting for i.MX6.
However, it does not respect it for some older chips.

> 
> Anyway I modified the code so that state change interrupt is not 
> enabled when user sets berr-reporting off (as you kindly suggested). It 
> solves the load issue. Bus-off recovery behavior did not change (I made a 
> few tests only). Btw the driver disables all interrupts in bus off so I 
> guess it does not matter what exact flags were set.
Perhaps, you would like to post your patch? Turning off the interrupts when
berr-reporting is off is definitely correct behaviour. Not turning them off
seems rather pointless to me. AFAIK the whole point of this flag is to suppress
ack-error flood.
> 
> > The consequence (depending on the CPU and Linux build options) is 10%-100% 
> > CPU utilisation when there's nothing to ACK the packet, and a higher figure 
> > again when the bus isn't terminated.
> >
> > CAN is meant to be used in cars or factories. Running with only one node or a 
> > disconnected bus is not a "normal" condition, so the hardware and drivers 
> > don't handle those conditions well.
> Well error states are defined by the spec so they ARE normal. What is not 
> normal is to bring down system performace when they happen.
> 
> 
> I alose added a line to put the controller in init mode in 
> c_can_stop() so that bringing the bus down will actually cause the 
> controller stop transmitting. I hope it's the way to do this, it would be 
> nice if some expert could verify this.
> 
Put this in a separate patch if you post this change.

Best regards,
Andri Yngvason