Re: flexcan napi poll and error frames

[Wolfgang Grandegger <wg@grandegger.com>]

On 10/24/2014 06:04 PM, Andri Yngvason wrote:
> 
> On fös 24.okt 2014 14:39, Andri Yngvason wrote:
>> On fös 24.okt 2014 12:33, Wolfgang Grandegger wrote:
>>> On Fri, 24 Oct 2014 10:55:48 +0000, Andri Yngvason
>>> <andri.yngvason@marel.com> wrote:
>>>> On fös 24.okt 2014 10:43, Wolfgang Grandegger wrote:
>>>>> On Fri, 24 Oct 2014 10:26:11 +0000, Andri Yngvason
>>>>> <andri.yngvason@marel.com> wrote:
>>>>>> Hi,
>>>>>>
>>>>>> I was running some tests on my patches when I noticed the following:
>>>>>> If I have 2 flexcan devices on the bus, each sending to the bus using
>>>>>> cangen,and then I disconnect the cable to one of them, that device
>>>>>> will enter"error-warning" state, but it will not continue on to
>>>>>> "error-passive" as itshould.
>>>>>>
>>>>>> However, when I reconnect the cable, I get the "error-passive" message
>>>>>> followed by an "error-warning" and eventually "back-to-error-active".
>>>>> Yes, I think I observed that behaviour as well as you can see here:
>>>>>
>>> https://gitorious.org/linux-can/wg-linux-can-next/commit/bd3acb12dbb9551541d28ae8766c154d3cf6ed57.patch
>>>> Good to know.
>>>>>
> ...
>>>>>
>>>>> I suspect that the problem is that the driver doesn't receive any
>>>>> interruptsother than the one for "error-passive" and so things
>>>>> won't "weigh" enoughfor napi. There seems to be some truth in this
>>>>> conjecture, because when Itried setting the napi weight to 1, the
>>>>> message got through.
>>>>> Hm, why should it depend on NAPI. It does not delay messages for
>>>>> a long time. I think the problem is that the state change is not
>>>>> signalled my an interrupt but some time later when another event
>>>>> (message) occurs.
>>>>>  
>>>> Perhaps, but how do you explain that the message got through when I
>>>> set the weight to 1?
>>> If it's really true it would be a bug in the NAPI handling. Could you
>>> please elaborate a bit more by adding some printouts in the interrupt
>>> handler. I will have a closer look tomorrow.
>> I wasn't lying about it. Perhaps by changing the weight it got through with
>> something else. I don't know; I'm not an expert on the inner workings of napi.
>>
>> But let's just forget about the weight thing. I found out by looking in the
>> i.mx6 reference manual that there is no interrupt for this transition. I
>> found that quite incredible so I searched through it a few times. Anyway,
>> there are only interrupts for active->tx-warning, active->rx-warning and
>> active->bus-off.
>>
>>>>>> Another thing that I found peculiar was that I had to be sending on
>>>>>> both devices for the error states to change to anything other than
>>>>>> "error-warning".
>>>>> Well, the error reporting on the SJA1000 is perfect... on all other
>>>>> CAN controllers it's more or less worse.
>>>>>
>>>> Should we just ignore this problem then? I'd rather like to figure
>>>> out if this is problem with the controller or not. Do you remember
>>>> if you've had this problem with flexcan?
>>> We can do little if the CAN controller does not notify the Software
>>> via interrupt.
>> Yes, that's why I wanted to figure out if it's a controller problem or not.
>> Turns out it's a controller problem, but perhaps we can work around it?
>> E.g. if we check esr for state changes every time someone transmits a
>> frame, both of these problems would go away. Would it be unacceptable
>> overhead to do so?
>>
> I've just confirmed that this "fix" works, but only if berr-reporting is
> enabled.

Ah, oh, this reminds me that there is a related bug in some versions of
the FLEXCAN core. If you look to "flexcan.c" you will find:

#define FLEXCAN_HAS_BROKEN_ERR_STATE	BIT(2) /* [TR]WRN INT not connected */

On these cores bus-error reporting is required to realize these state
changes.

Wolfgang.