Re: [Xenomai] iMX6q 1Hz tick

[Gilles Chanteperdrix <gilles.chanteperdrix@xenomai.org>]

On 12/10/2012 09:02 PM, Thierry Bultel wrote:

> Le 10/12/2012 19:56, Gilles Chanteperdrix a écrit :
>> On 12/10/2012 07:34 PM, Thierry Bultel wrote:
>>
>>> Le 10/12/2012 18:37, Gilles Chanteperdrix a écrit :
>>>> On 12/10/2012 05:31 PM, Thierry Bultel wrote:
>>>>> Le 10/12/2012 17:01, Gilles Chanteperdrix a écrit :
>>>>>> On 12/10/2012 03:53 PM, Thierry Bultel wrote:
>>>>>>> Le 10/12/2012 00:54, Gilles Chanteperdrix a écrit :
>>>>>>>> On 12/09/2012 09:06 PM, Thierry Bultel wrote:
>>>>>>>>
>>>>>>>>> Le 07/12/2012 19:34, Gilles Chanteperdrix a écrit :
>>>>>>>>>> On 12/07/2012 02:28 PM, Thierry Bultel wrote:
>>>>>>>>>>
>>>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> Hello Gilles,
>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> I have "successfully" patched a 3.0.35 kernel from Freescale with
>>>>>>>>>>> adeos-ipipe-3.0.36-1.18.11
>>>>>>>>>>>
>>>>>>>>>>> I have -not- applied the adeos-ipipe-3.0.36-arm-1.18-11-pre.patch
>>>>>>>>>>> that showed too much "revert patch detected" errors.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> The patch should apply cleanly, the question, if you get too many
>>>>>>>>>> rejections, it is because you do not apply it to the proper branch.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Of course it was not.
>>>>>>>>> I started with a kernel provided by the module manufacturer. I estimate
>>>>>>>>> that this kernel is "almost" a rel_imx_3.0.35_12.09.02, I mean, that
>>>>>>>>> I have not found a closer tag than this one.
>>>>>>>>>
>>>>>>>>> My first idea was to attempt to apply the patches to that kernel
>>>>>>>>> directly. I was not expecting it would have been simple, and it was
>>>>>>>>> actually not. But after a few hours I came to the initial result I said.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> BTW, I have understand that its purpose was mainly to upgrade to 3.0.36,
>>>>>>>>>>> but what is actually the role of the
>>>>>>>>>>> adeos-ipipe-3.0.36-arm-1.18-11-post.patch ? (I applied it).
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> I have explained this in a previous mail.
>>>>>>>>>
>>>>>>>>> Ok, I will seek in the history. I would be nice if the explanation
>>>>>>>>> was in the README.
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> I had to deal with some rejections but at least the kernel is booting
>>>>>>>>>>> and my application running.
>>>>>>>>>>>
>>>>>>>>>>> But the target is very slow, and /proc/xenomai/irq shows that the
>>>>>>>>>>> [timer] irq only comes about every 1 second.
>>>>>>>>>>>
>>>>>>>>>>> The /proc/interrupts shows an IRQ every 10ms, which was expected
>>>>>>>>>>> since I have CONFIG_HZ=100
>>>>>>>>>>>
>>>>>>>>>>> I am pretty sure that I messed up something in the patch process,
>>>>>>>>>>> however I do not know where to start to look for.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Apply the patches to the proper imx release, and everything should be fine.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> That is what I did, at last.
>>>>>>>>> To do so, I have extracted the BSP from the given 'almost'
>>>>>>>>> rel_imx_3.0.35_12.09.02 kernel,
>>>>>>>>> taken the rel_imx_3.0.15_12.03.00, applied the patches,
>>>>>>>>> and brought back the BSP to the patched kernel.
>>>>>>>>
>>>>>>>
>>>>>>> The issues I mentionned are not also present with CONFIG_IPIPE=no,
>>>>>>> so it is just that the 3.0.15 kernel does not have everything for my board.
>>>>>>>
>>>>>>>>
>>>>>>>> I would do what you want to do with git. Checkout the ipipe-3.0-imx6q
>>>>>>>> branch from the ipipe-gch.git, then merge it with the vendor branch you
>>>>>>>> want to use.
>>>>>>>>
>>>>>>>
>>>>>>> Sure it would work and would be fine with git.
>>>>>>> But unfortunalely I do not have access to the repository of the vendor,
>>>>>>> but only to a bz2 snapshot of their kernel.
>>>>>>>
>>>>>>> I wonder if the right method would be to use git to merge Ipipe to the
>>>>>>> 3.0.35_12_09.02, and then bring back the BSP that I have isolated as a patch
>>>>>>
>>>>>> Yes, if the vendor BSP is based on 3.0.35_12_09.02, you should make a
>>>>>> diff with that, and try and merge it with the ipipe-3.0-imx6q branch.
>>>>>>
>>>>>
>>>>>
>>>>> Done. And success. Thanks. I am running your 3.0.43 kernel with my BSP.
>>>>> no more network or reboot slowness issues.
>>>>>
>>>>> However, htop it saying that CPU0 is at 20% and CPU3 at 12%,
>>>>> no application is running. The only thing that I see running is the
>>>>> timer interrupt.
>>>>>
>>>>> Also, I still have
>>>>> 87:         13          0          0          0       GIC  i.MX Timer Tick
>>>>>
>>>>> .. the counter increases very slowly, about 1 every 10 minutes maybe.
>>>>>
>>>>> That does not happen with CONFIG_IPIPE=no
>>>>
>>>> Do you have CONFIG_LOCALTIMER (and CONFIG_SMP) enabled?
>>>>
>>>>
>>>
>>> CONFIG_SMP yes,
>>> You mean CONFIG_LOCAL_TIMERS ? Yes
>>
>>
>> Could you post the disassembly of the mxc_timer_interrupt function?
>>
> 
> Hope this helps:
> 
> 8006e93c <mxc_timer_interrupt>:
> 8006e93c:       e3031998        movw    r1, #14744      ; 0x3998
> 8006e940:       e92d4010        push    {r4, lr}
> 8006e944:       e34810c7        movt    r1, #32967      ; 0x80c7
> 8006e948:       e3020a00        movw    r0, #10752      ; 0x2a00
> 8006e94c:       e34800c3        movt    r0, #32963      ; 0x80c3
> 8006e950:       e3a04001        mov     r4, #1
> 8006e954:       e5912000        ldr     r2, [r1]
> 8006e958:       e5903000        ldr     r3, [r0]
> 8006e95c:       e5921008        ldr     r1, [r2, #8]
> 8006e960:       e5824008        str     r4, [r2, #8]
> 8006e964:       e12fff33        blx     r3
> 8006e968:       e1a00004        mov     r0, r4
> 8006e96c:       e8bd8010        pop     {r4, pc}
> 
> Tell me if you need more information
> 
> 

Ok, this looks compiled as the sources tell us. The thing I do not
understand is that since there is another clockevent running (the local
timers), the mxc_timer should not be ticking at all. What does cat
/proc/timer_list say?

Also, you can try adding the call to __ipipe_tsc_update in
mxc_timer_interrupt even in the SMP case. If you have a problem with the
tsc, the "tsc" test from Xenomai testsuite should fail after some time
when run with the "-w" argument, but normally the tsc should take a long
time to wrap, so, there should not be immediately visible effects. But
it is worth checking.

-- 
                                                                Gilles.