[Xenomai] iMX6 support

[Xenomai] iMX6 support

[Thierry Bultel]

Hi,

We are looking in a replacement solution for our obsolete 
pentium-Celeron calculator.
For many reasons, we would like to switch to the ARM architecture.
The aimed CPU is the iMX6 Quadcore.

Freescale provides a patched 2.6.35 and 2.6.38 kernel that do not have 
all the peripherals (the CAN seems to be missing)
and Emtrion, a CPU module manufacturer, say they provide the additional 
patches to have the complete BSP.
It is likely a 2.6.35 only, though.

 From my point of view, the remaining work would be to :

1) apply the Xenomai 2.6.0 patches on their kernel,
after having ported their diffs from 2.6.35 to 2.6.38

2) Once it is done, as we want CAN support, it remains the task to port 
the FlexCAN driver as an RTDM one,
that I estimate to 1 man/week work.

Do you experts see any possible issues or risks that I have not been 
thinking about ?
Any advice would be greatly appreciated.

Best Regards
Thierry Bultel

Re: [Xenomai] iMX6 support

[Gilles Chanteperdrix]

On 05/23/2012 11:56 AM, Thierry Bultel wrote:
> Hi,
> 
> We are looking in a replacement solution for our obsolete 
> pentium-Celeron calculator.
> For many reasons, we would like to switch to the ARM architecture.
> The aimed CPU is the iMX6 Quadcore.
> 
> Freescale provides a patched 2.6.35 and 2.6.38 kernel that do not have 
> all the peripherals (the CAN seems to be missing)
> and Emtrion, a CPU module manufacturer, say they provide the additional 
> patches to have the complete BSP.
> It is likely a 2.6.35 only, though.

Xenomai for mx6q does not exist for linux 2.6.35. It only exists for
freescale branch for 2.6.38, or 3.0, as a separate set of patches to
apply before and after the Adeos patches for the vanilla kernel. For
further details, see ksrc/arch/arm/patches/README in xenomai 2.6 sources.

Furthermore, we have some stability issues with Xenomai for mx6q with
3.0, the issues looking a lot like an issue with cache. We would ned to
recheck 2.6.38 to see if it had the same issues, but I would bet it does.

-- 
					    Gilles.

Re: [Xenomai] iMX6 support

[Thierry Bultel]

Le 23/05/2012 12:05, Gilles Chanteperdrix a écrit :
> On 05/23/2012 11:56 AM, Thierry Bultel wrote:
>> Hi,
>>
>> We are looking in a replacement solution for our obsolete
>> pentium-Celeron calculator.
>> For many reasons, we would like to switch to the ARM architecture.
>> The aimed CPU is the iMX6 Quadcore.
>>
>> Freescale provides a patched 2.6.35 and 2.6.38 kernel that do not have
>> all the peripherals (the CAN seems to be missing)
>> and Emtrion, a CPU module manufacturer, say they provide the additional
>> patches to have the complete BSP.
>> It is likely a 2.6.35 only, though.
> Xenomai for mx6q does not exist for linux 2.6.35. It only exists for
> freescale branch for 2.6.38, or 3.0, as a separate set of patches to
> apply before and after the Adeos patches for the vanilla kernel. For
> further details, see ksrc/arch/arm/patches/README in xenomai 2.6 sources.
Thanks, yes, as I said, I am aware that the very first work for me will 
be to get
the non-xenomai 2.6.38 kernel fully working
> Furthermore, we have some stability issues with Xenomai for mx6q with
> 3.0, the issues looking a lot like an issue with cache. We would ned to
> recheck 2.6.38 to see if it had the same issues, but I would bet it does.
>
Actually my concern was to know if there were currently developments
on that CPU. The fact that they are remaining bugs is not a concern,
as long as they would be likely be corrected in the coming months.
Do these bugs also exist on the mx6 single core ?

Re: [Xenomai] iMX6 support

[Gilles Chanteperdrix]

On 05/23/2012 12:14 PM, Thierry Bultel wrote:
> Le 23/05/2012 12:05, Gilles Chanteperdrix a écrit :
>> On 05/23/2012 11:56 AM, Thierry Bultel wrote:
>>> Hi,
>>>
>>> We are looking in a replacement solution for our obsolete
>>> pentium-Celeron calculator.
>>> For many reasons, we would like to switch to the ARM architecture.
>>> The aimed CPU is the iMX6 Quadcore.
>>>
>>> Freescale provides a patched 2.6.35 and 2.6.38 kernel that do not have
>>> all the peripherals (the CAN seems to be missing)
>>> and Emtrion, a CPU module manufacturer, say they provide the additional
>>> patches to have the complete BSP.
>>> It is likely a 2.6.35 only, though.
>> Xenomai for mx6q does not exist for linux 2.6.35. It only exists for
>> freescale branch for 2.6.38, or 3.0, as a separate set of patches to
>> apply before and after the Adeos patches for the vanilla kernel. For
>> further details, see ksrc/arch/arm/patches/README in xenomai 2.6 sources.
> Thanks, yes, as I said, I am aware that the very first work for me will 
> be to get
> the non-xenomai 2.6.38 kernel fully working
>> Furthermore, we have some stability issues with Xenomai for mx6q with
>> 3.0, the issues looking a lot like an issue with cache. We would ned to
>> recheck 2.6.38 to see if it had the same issues, but I would bet it does.
>>
> Actually my concern was to know if there were currently developments
> on that CPU. The fact that they are remaining bugs is not a concern,
> as long as they would be likely be corrected in the coming months.
> Do these bugs also exist on the mx6 single core ?

I was unable to reproduce the cache issue on single core, but I have not
tested enough to be absolutely sure that we do not have it. The issue is
sufficiently rare that it may go unnoticed during one test, and appear
during the next.

-- 
					    Gilles.

Re: [Xenomai] iMX6 support

[Thierry Bultel]

Hi,
I have another question about the iMX6,
it is about the TSC emulation .

I have been told that the Pentium TSC could not be accurate enough
for a one of our application that performs high precision data acquisition.

Can we expect a better accuracy with the TSC emulation based
on the ARM timer or not ?

Is this based on the CCNT that is cycle-accurate ?

Thanks
Thierry

Le 23/05/2012 14:32, Gilles Chanteperdrix a écrit :
> On 05/23/2012 12:14 PM, Thierry Bultel wrote:
>> Le 23/05/2012 12:05, Gilles Chanteperdrix a écrit :
>>> On 05/23/2012 11:56 AM, Thierry Bultel wrote:
>>>> Hi,
>>>>
>>>> We are looking in a replacement solution for our obsolete
>>>> pentium-Celeron calculator.
>>>> For many reasons, we would like to switch to the ARM architecture.
>>>> The aimed CPU is the iMX6 Quadcore.
>>>>
>>>> Freescale provides a patched 2.6.35 and 2.6.38 kernel that do not have
>>>> all the peripherals (the CAN seems to be missing)
>>>> and Emtrion, a CPU module manufacturer, say they provide the additional
>>>> patches to have the complete BSP.
>>>> It is likely a 2.6.35 only, though.
>>> Xenomai for mx6q does not exist for linux 2.6.35. It only exists for
>>> freescale branch for 2.6.38, or 3.0, as a separate set of patches to
>>> apply before and after the Adeos patches for the vanilla kernel. For
>>> further details, see ksrc/arch/arm/patches/README in xenomai 2.6 sources.
>> Thanks, yes, as I said, I am aware that the very first work for me will
>> be to get
>> the non-xenomai 2.6.38 kernel fully working
>>> Furthermore, we have some stability issues with Xenomai for mx6q with
>>> 3.0, the issues looking a lot like an issue with cache. We would ned to
>>> recheck 2.6.38 to see if it had the same issues, but I would bet it does.
>>>
>> Actually my concern was to know if there were currently developments
>> on that CPU. The fact that they are remaining bugs is not a concern,
>> as long as they would be likely be corrected in the coming months.
>> Do these bugs also exist on the mx6 single core ?
> I was unable to reproduce the cache issue on single core, but I have not
> tested enough to be absolutely sure that we do not have it. The issue is
> sufficiently rare that it may go unnoticed during one test, and appear
> during the next.
>

Re: [Xenomai] iMX6 support

[Gilles Chanteperdrix]

On 06/07/2012 07:04 PM, Thierry Bultel wrote:
> Hi,
> I have another question about the iMX6,
> it is about the TSC emulation .
> 
> I have been told that the Pentium TSC could not be accurate enough
> for a one of our application that performs high precision data acquisition.
> 
> Can we expect a better accuracy with the TSC emulation based
> on the ARM timer or not ?

Xenomai tsc emulation is usually based on a hardware timer which has a
lower frequency than the core. On cortex A9 cores, we use the "global
timer". On imx6q, this timer frequency should be around 500MHz, from the
top of my head.

Anyway, if you require a precision better than the processor cycle, then
you should not use the processor to do the measurement. Because for
instance, simply reading the TSC has a latency. This latency is around
10 cycles on mx6q (I am talking about 500 MHz cycles, not processor
cycles). And at that level, you should also account for the processor
intrinsic lack of determinism, with such things as pipeline stalls. So,
you simply can not achieve a better precision than the TSC when
measuring with a processor.

> 
> Is this based on the CCNT that is cycle-accurate ?

No, it is based on the cortex a9 "global timer". I am sure we can read
this one from user-space, I am not sure we can read CCNT from user-space.

-- 
					    Gilles.

Re: [Xenomai] iMX6 support

[Gilles Chanteperdrix]

On 06/07/2012 07:15 PM, Gilles Chanteperdrix wrote:
> On 06/07/2012 07:04 PM, Thierry Bultel wrote:
>> Hi,
>> I have another question about the iMX6,
>> it is about the TSC emulation .
>>
>> I have been told that the Pentium TSC could not be accurate enough
>> for a one of our application that performs high precision data acquisition.
>>
>> Can we expect a better accuracy with the TSC emulation based
>> on the ARM timer or not ?
> 
> Xenomai tsc emulation is usually based on a hardware timer which has a
> lower frequency than the core.

ARM tsc emulation. Other architectures are different.

-- 
                                                                Gilles.