[Xenomai] imx28 rtcan flexcan system freezes

[Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Hi all,

I use two rtcan flexcan on imx28-evk board.

If both rtcan devices receive a data at the same time, this cause a freeze.
After investigation, i found that if I replace in flexcan_rx_interrupt 
these instructions :
-        *(__be32 *)(cf->data + 0) =
-            cpu_to_be32(flexcan_read(&mb->data[0]));
-        *(__be32 *)(cf->data + 4) =
-            cpu_to_be32(flexcan_read(&mb->data[1]));

by these

+        data0 = flexcan_read(&mb->data[0]);
+        data1 = flexcan_read(&mb->data[1]);
+
+        cf->data[3]=((data0 >> 0)  & 0xFF) ;
+        cf->data[2]=((data0 >> 8)  & 0xFF) ;
+        cf->data[1]=((data0 >> 16)  & 0xFF) ;
+        cf->data[0]=((data0 >> 24)  & 0xFF) ;
+        cf->data[7]=((data1 >> 0)  & 0xFF) ;
+        cf->data[6]=((data1 >> 8)  & 0xFF) ;
+        cf->data[5]=((data1 >> 16)  & 0xFF) ;
+        cf->data[4]=((data1 >> 24)  & 0xFF) ;

rtcan flexcan works perfectly.
is  anyone know what is the problem with first instructions ?

My config : xenomai 2.6.3,  linux 3.8.13,

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/22/2014 11:24 AM, Alexandre COFFIGNAL wrote:
> Hi all,
> 
> I use two rtcan flexcan on imx28-evk board.
> 
> If both rtcan devices receive a data at the same time, this cause a freeze.
> After investigation, i found that if I replace in flexcan_rx_interrupt 
> these instructions :
> -        *(__be32 *)(cf->data + 0) =
> -            cpu_to_be32(flexcan_read(&mb->data[0]));
> -        *(__be32 *)(cf->data + 4) =
> -            cpu_to_be32(flexcan_read(&mb->data[1]));
> 
> by these
> 
> +        data0 = flexcan_read(&mb->data[0]);
> +        data1 = flexcan_read(&mb->data[1]);
> +
> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
> 
> rtcan flexcan works perfectly.
> is  anyone know what is the problem with first instructions ?

Probably mb->data does not have the right alignment. Could you not
arrange to get it properly aligned? Failing that, you should use
put_unaligned instead of open coding it.


-- 
                                                                Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 22/01/2014 12:23, Gilles Chanteperdrix a écrit :
> On 01/22/2014 11:24 AM, Alexandre COFFIGNAL wrote:
>> Hi all,
>>
>> I use two rtcan flexcan on imx28-evk board.
>>
>> If both rtcan devices receive a data at the same time, this cause a freeze.
>> After investigation, i found that if I replace in flexcan_rx_interrupt
>> these instructions :
>> -        *(__be32 *)(cf->data + 0) =
>> -            cpu_to_be32(flexcan_read(&mb->data[0]));
>> -        *(__be32 *)(cf->data + 4) =
>> -            cpu_to_be32(flexcan_read(&mb->data[1]));
>>
>> by these
>>
>> +        data0 = flexcan_read(&mb->data[0]);
>> +        data1 = flexcan_read(&mb->data[1]);
>> +
>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>
>> rtcan flexcan works perfectly.
>> is  anyone know what is the problem with first instructions ?
> Probably mb->data does not have the right alignment. Could you not
> arrange to get it properly aligned? Failing that, you should use
> put_unaligned instead of open coding it.
>
>
here structures used in flexcan driver, it seem to be aligned

/* Structure of the message buffer */
struct flexcan_mb {
     u32 can_ctrl;
     u32 can_id;
     u32 data[2];
};

/* Structure of the hardware registers */
struct flexcan_regs {
     u32 mcr;        /* 0x00 */
     u32 ctrl;        /* 0x04 */
     u32 timer;        /* 0x08 */
     u32 _reserved1;        /* 0x0c */
     u32 rxgmask;        /* 0x10 */
     u32 rx14mask;        /* 0x14 */
     u32 rx15mask;        /* 0x18 */
     u32 ecr;        /* 0x1c */
     u32 esr;        /* 0x20 */
     u32 imask2;        /* 0x24 */
     u32 imask1;        /* 0x28 */
     u32 iflag2;        /* 0x2c */
     u32 iflag1;        /* 0x30 */
     u32 crl2;        /* 0x34 */
     u32 esr2;        /* 0x38 */
     u32 _reserved2[2];
     u32 crcr;        /* 0x44 */
     u32 rxfgmask;        /* 0x48 */
     u32 rxfir;        /* 0x4c */
     u32 _reserved3[12];
     struct flexcan_mb cantxfg[64];
};

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>
>>> rtcan flexcan works perfectly.
>>> is  anyone know what is the problem with first instructions ?
>> Probably mb->data does not have the right alignment. Could you not
>> arrange to get it properly aligned? Failing that, you should use
>> put_unaligned instead of open coding it.
>>
>>
> here structures used in flexcan driver, it seem to be aligned

I am talking about the alignment of cf->data, since obviously, that is 
the one which is causing problems.


-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>
>>>> rtcan flexcan works perfectly.
>>>> is  anyone know what is the problem with first instructions ?
>>> Probably mb->data does not have the right alignment. Could you not
>>> arrange to get it properly aligned? Failing that, you should use
>>> put_unaligned instead of open coding it.
>>>
>>>
>> here structures used in flexcan driver, it seem to be aligned
>
> I am talking about the alignment of cf->data, since obviously, that is 
> the one which is causing problems.
>
>
Thank a lot, put_unaligned fix this issue if you want, i can send a path
-- 

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>
> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>
>>>>> rtcan flexcan works perfectly.
>>>>> is  anyone know what is the problem with first instructions ?
>>>> Probably mb->data does not have the right alignment. Could you not
>>>> arrange to get it properly aligned? Failing that, you should use
>>>> put_unaligned instead of open coding it.
>>>>
>>>>
>>> here structures used in flexcan driver, it seem to be aligned
>>
>> I am talking about the alignment of cf->data, since obviously, that is
>> the one which is causing problems.
>>
>>
> Thank a lot, put_unaligned fix this issue if you want, i can send a path
>

The other solution (getting cf->data to be properly aligned) would be 
more efficient, why is not it possible to get cf->data properly aligned?

-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>
>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>
>>>>>> rtcan flexcan works perfectly.
>>>>>> is  anyone know what is the problem with first instructions ?
>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>> put_unaligned instead of open coding it.
>>>>>
>>>>>
>>>> here structures used in flexcan driver, it seem to be aligned
>>>
>>> I am talking about the alignment of cf->data, since obviously, that is
>>> the one which is causing problems.
>>>
>>>
>> Thank a lot, put_unaligned fix this issue if you want, i can send a path
>>
>
> The other solution (getting cf->data to be properly aligned) would be 
> more efficient, why is not it possible to get cf->data properly aligned?
>
I think, i can't get cf->data to be properly aligned because "cf" is 
receive internal frame representation within the ring buffer
of a struct rtcan_socket and struct rtcan_rb_frame is a generic 
structure used in all rtcan drivers .

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>
> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>
>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>
>>>>>>> rtcan flexcan works perfectly.
>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>> put_unaligned instead of open coding it.
>>>>>>
>>>>>>
>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>
>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>> the one which is causing problems.
>>>>
>>>>
>>> Thank a lot, put_unaligned fix this issue if you want, i can send a path
>>>
>>
>> The other solution (getting cf->data to be properly aligned) would be
>> more efficient, why is not it possible to get cf->data properly aligned?
>>
> I think, i can't get cf->data to be properly aligned because "cf" is
> receive internal frame representation within the ring buffer
> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
> structure used in all rtcan drivers .

Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for 
all drivers, so that looks like a worthwile improvement...


-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>
>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>
>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>>
>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>> put_unaligned instead of open coding it.
>>>>>>>
>>>>>>>
>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>
>>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>>> the one which is causing problems.
>>>>>
>>>>>
>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a path
>>>>
>>>
>>> The other solution (getting cf->data to be properly aligned) would be
>>> more efficient, why is not it possible to get cf->data properly aligned?
>>>
>> I think, i can't get cf->data to be properly aligned because "cf" is
>> receive internal frame representation within the ring buffer
>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>> structure used in all rtcan drivers .
> 
> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for 
> all drivers, so that looks like a worthwile improvement...
> 
> 

Wolfgang, what do you think? Should we change the alignment of
rtcan_rb_frame::data, or use put_unaligned on ARM?

-- 
                                                                Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Wolfgang Grandegger]

Hi Gilles,



On Thu, 23 Jan 2014 11:04:44 +0100, Gilles Chanteperdrix

<gilles.chanteperdrix@xenomai.org> wrote:

> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:

>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:

>>>

>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :

>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:

>>>>>

>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :

>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:

>>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;

>>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;

>>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;

>>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;

>>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;

>>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;

>>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;

>>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;

>>>>>>>>>

>>>>>>>>> rtcan flexcan works perfectly.

>>>>>>>>> is  anyone know what is the problem with first instructions ?

>>>>>>>> Probably mb->data does not have the right alignment. Could you

not

>>>>>>>> arrange to get it properly aligned? Failing that, you should use

>>>>>>>> put_unaligned instead of open coding it.



How are unaligned accesses handled on your system? Alexandre,

what does "cat /proc/cpu/alignment" report? Aligned access

would be better, I agree.



>>>>>>>>

>>>>>>>>

>>>>>>> here structures used in flexcan driver, it seem to be aligned

>>>>>>

>>>>>> I am talking about the alignment of cf->data, since obviously, that

>>>>>> is

>>>>>> the one which is causing problems.

>>>>>>

>>>>>>

>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a

>>>>> path

>>>>>

>>>>

>>>> The other solution (getting cf->data to be properly aligned) would be

>>>> more efficient, why is not it possible to get cf->data properly

>>>> aligned?

>>>>

>>> I think, i can't get cf->data to be properly aligned because "cf" is

>>> receive internal frame representation within the ring buffer

>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic

>>> structure used in all rtcan drivers .

>> 

>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for 

>> all drivers, so that looks like a worthwile improvement...

>> 

>> 

> 

> Wolfgang, what do you think? Should we change the alignment of

> rtcan_rb_frame::data, or use put_unaligned on ARM?



I realized the mail and wonder why this problem does not show up with

vanilla

Linux. Maybe we need to port over some patch from there. I will have a

closer

look this even or over the weekend.



Wolfgang.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/23/2014 12:36 PM, Wolfgang Grandegger wrote:
> How are unaligned accesses handled on your system? Alexandre,
>
> what does "cat /proc/cpu/alignment" report? Aligned access
>
> would be better, I agree.

The imx28 can not support unaligned accesses in hardware. The unaligned 
access causes an exception, /proc/cpu/alignment thus only allow choosing 
how the exception is handled.


>> Wolfgang, what do you think? Should we change the alignment of
>
>> rtcan_rb_frame::data, or use put_unaligned on ARM?
>
>
>
> I realized the mail and wonder why this problem does not show up with
>
> vanilla

Vanilla handles the alignment trap by emulating the faulting 
instruction. Last time I checked, there were some problems related to 
running the emulation over xenomai domain, though looking at the code 
now, I see nothing which would prevent us from doing it.

-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Wolfgang Grandegger]

On Thu, 23 Jan 2014 13:05:40 +0100, Gilles Chanteperdrix

<gilles.chanteperdrix@xenomai.org> wrote:

> On 01/23/2014 12:36 PM, Wolfgang Grandegger wrote:

>> How are unaligned accesses handled on your system? Alexandre,

>>

>> what does "cat /proc/cpu/alignment" report? Aligned access

>>

>> would be better, I agree.

> 

> The imx28 can not support unaligned accesses in hardware. The unaligned 

> access causes an exception, /proc/cpu/alignment thus only allow choosing



> how the exception is handled.

> 

> 

>>> Wolfgang, what do you think? Should we change the alignment of

>>

>>> rtcan_rb_frame::data, or use put_unaligned on ARM?

>>

>>

>>

>> I realized the mail and wonder why this problem does not show up with

>>

>> vanilla

> 

> Vanilla handles the alignment trap by emulating the faulting 

> instruction. Last time I checked, there were some problems related to 

> running the emulation over xenomai domain, though looking at the code 

> now, I see nothing which would prevent us from doing it.



Yes, I know. But the handling of unaligned accesses via trap handler must

be

enabled by the user and I'm not aware that this is needed for vanilla

Linux.



Wolfgang.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/23/2014 01:21 PM, Wolfgang Grandegger wrote:
>> Vanilla handles the alignment trap by emulating the faulting
>
>> instruction. Last time I checked, there were some problems related to
>
>> running the emulation over xenomai domain, though looking at the code
>
>> now, I see nothing which would prevent us from doing it.
>
>
>
> Yes, I know. But the handling of unaligned accesses via trap handler must
>
> be
>
> enabled by the user and I'm not aware that this is needed for vanilla
>
> Linux.

I remember reading on the LAKML that the alignment trap is needed for 
kernel operations, notably for the IP stack.

-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/23/2014 01:05 PM, Gilles Chanteperdrix wrote:
> Vanilla handles the alignment trap by emulating the faulting
> instruction. Last time I checked, there were some problems related to
> running the emulation over xenomai domain, though looking at the code
> now, I see nothing which would prevent us from doing it.
>

Correction: the emulation code would require changing 
probe_kernel_address to be able to run over non-linux stacks.

-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 23/01/2014 12:36, Wolfgang Grandegger a écrit :
> Hi Gilles,
>
>
>
> On Thu, 23 Jan 2014 11:04:44 +0100, Gilles Chanteperdrix
>
> <gilles.chanteperdrix@xenomai.org> wrote:
>
>> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
>>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>>>> Probably mb->data does not have the right alignment. Could you
> not
>
>>>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>>>> put_unaligned instead of open coding it.
>
>
> How are unaligned accesses handled on your system? Alexandre,
>
> what does "cat /proc/cpu/alignment" report? Aligned access
>
> would be better, I agree.
>
>
>
>>>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>>> I am talking about the alignment of cf->data, since obviously, that
>>>>>>> is
>>>>>>> the one which is causing problems.
>>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a
>>>>>> path
>>>>> The other solution (getting cf->data to be properly aligned) would be
>>>>> more efficient, why is not it possible to get cf->data properly
>>>>> aligned?
>>>> I think, i can't get cf->data to be properly aligned because "cf" is
>>>> receive internal frame representation within the ring buffer
>>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>>>> structure used in all rtcan drivers .
>>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for
>>> all drivers, so that looks like a worthwile improvement...
>> Wolfgang, what do you think? Should we change the alignment of
>> rtcan_rb_frame::data, or use put_unaligned on ARM?
>
>
> I realized the mail and wonder why this problem does not show up with
>
> vanilla
>
> Linux. Maybe we need to port over some patch from there. I will have a
>
> closer
>
> look this even or over the weekend.
>
>
>
> Wolfgang.
>
>
>
> _______________________________________________
> Xenomai mailing list
> Xenomai@xenomai.org
> http://www.xenomai.org/mailman/listinfo/xenomai

Hi,

"cat /proc/cpu/alignment" report
User:           0
System:         0
Skipped:        0
Half:           0
Word:           0
DWord:          0
Multi:          0
User faults:    0 (ignored)

something wrong with that ?
  
Alexandre.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 23/01/2014 13:54, Alexandre COFFIGNAL a écrit :
>
> Le 23/01/2014 12:36, Wolfgang Grandegger a écrit :
>> Hi Gilles,
>>
>>
>>
>> On Thu, 23 Jan 2014 11:04:44 +0100, Gilles Chanteperdrix
>>
>> <gilles.chanteperdrix@xenomai.org> wrote:
>>
>>> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
>>>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>>>>> + cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[2]=((data0 >> 8) & 0xFF) ;
>>>>>>>>>>> +        cf->data[1]=((data0 >> 16) & 0xFF) ;
>>>>>>>>>>> +        cf->data[0]=((data0 >> 24) & 0xFF) ;
>>>>>>>>>>> +        cf->data[7]=((data1 >> 0) & 0xFF) ;
>>>>>>>>>>> +        cf->data[6]=((data1 >> 8) & 0xFF) ;
>>>>>>>>>>> +        cf->data[5]=((data1 >> 16) & 0xFF) ;
>>>>>>>>>>> +        cf->data[4]=((data1 >> 24) & 0xFF) ;
>>>>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>>>>> Probably mb->data does not have the right alignment. Could you
>> not
>>
>>>>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>>>>> put_unaligned instead of open coding it.
>>
>>
>> How are unaligned accesses handled on your system? Alexandre,
>>
>> what does "cat /proc/cpu/alignment" report? Aligned access
>>
>> would be better, I agree.
>>
>>
>>
>>>>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>>>> I am talking about the alignment of cf->data, since obviously, 
>>>>>>>> that
>>>>>>>> is
>>>>>>>> the one which is causing problems.
>>>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a
>>>>>>> path
>>>>>> The other solution (getting cf->data to be properly aligned) 
>>>>>> would be
>>>>>> more efficient, why is not it possible to get cf->data properly
>>>>>> aligned?
>>>>> I think, i can't get cf->data to be properly aligned because "cf" is
>>>>> receive internal frame representation within the ring buffer
>>>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>>>>> structure used in all rtcan drivers .
>>>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for
>>>> all drivers, so that looks like a worthwile improvement...
>>> Wolfgang, what do you think? Should we change the alignment of
>>> rtcan_rb_frame::data, or use put_unaligned on ARM?
>>
>>
>> I realized the mail and wonder why this problem does not show up with
>>
>> vanilla
>>
>> Linux. Maybe we need to port over some patch from there. I will have a
>>
>> closer
>>
>> look this even or over the weekend.
>>
>>
>>
>> Wolfgang.
>>
>>
>>
>> _______________________________________________
>> Xenomai mailing list
>> Xenomai@xenomai.org
>> http://www.xenomai.org/mailman/listinfo/xenomai
>
> Hi,
>
> "cat /proc/cpu/alignment" report
> User:           0
> System:         0
> Skipped:        0
> Half:           0
> Word:           0
> DWord:          0
> Multi:          0
> User faults:    0 (ignored)
>
> something wrong with that ?
>
> Alexandre.
>
>
> _______________________________________________
> Xenomai mailing list
> Xenomai@xenomai.org
> http://www.xenomai.org/mailman/listinfo/xenomai

For your information :
Actualy my kernel is linux 3.8.13 vanilla from 
https://www.kernel.org/pub/linux/kernel/v3.x/"
with Adeos/Xenomai Real-time patch 2.6.3 from
download.gna.org/xenomai/stable/xenomai-2.6.3.tar.bz2

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Wolfgang Grandegger]

On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>
>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>
>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>>
>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>> put_unaligned instead of open coding it.
>>>>>>>
>>>>>>>
>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>
>>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>>> the one which is causing problems.
>>>>>
>>>>>
>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a
>>>> path
>>>>
>>>
>>> The other solution (getting cf->data to be properly aligned) would be
>>> more efficient, why is not it possible to get cf->data properly aligned?
>>>
>> I think, i can't get cf->data to be properly aligned because "cf" is
>> receive internal frame representation within the ring buffer
>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>> structure used in all rtcan drivers .
> 
> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for
> all drivers, so that looks like a worthwile improvement...

I had a closer look. The CAN message (frame) queuing is optimized for
size to get as much as possible messages into the queue. Therefore the
CAN data are copied using memcpy, but only the real bytes. Directly
behind the real data is the 64 bit timestamp, if requested, which is
also not aligned. Therefore I suggest to use memcpy or byte accesses in
the flexcan driver as well. If this was a good decision is another question.

Wolfgang.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/26/2014 08:26 PM, Wolfgang Grandegger wrote:
> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>>
>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>>
>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>>>
>>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>>> put_unaligned instead of open coding it.
>>>>>>>>
>>>>>>>>
>>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>>
>>>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>>>> the one which is causing problems.
>>>>>>
>>>>>>
>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a
>>>>> path
>>>>>
>>>>
>>>> The other solution (getting cf->data to be properly aligned) would be
>>>> more efficient, why is not it possible to get cf->data properly aligned?
>>>>
>>> I think, i can't get cf->data to be properly aligned because "cf" is
>>> receive internal frame representation within the ring buffer
>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>>> structure used in all rtcan drivers .
>>
>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for
>> all drivers, so that looks like a worthwile improvement...
> 
> I had a closer look. The CAN message (frame) queuing is optimized for
> size to get as much as possible messages into the queue. Therefore the
> CAN data are copied using memcpy, but only the real bytes. Directly
> behind the real data is the 64 bit timestamp, if requested, which is
> also not aligned. Therefore I suggest to use memcpy or byte accesses in
> the flexcan driver as well. If this was a good decision is another question.

Aligning the data begining costs 2 bytes by struct rtcan_rb_frame. How
many such structures are there for a typical driver?


-- 
                                                                Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Wolfgang Grandegger]

On 01/26/2014 08:29 PM, Gilles Chanteperdrix wrote:
> On 01/26/2014 08:26 PM, Wolfgang Grandegger wrote:
>> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
>>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>>>
>>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>>>
>>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>>>>
>>>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>>>> put_unaligned instead of open coding it.
>>>>>>>>>
>>>>>>>>>
>>>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>>>
>>>>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>>>>> the one which is causing problems.
>>>>>>>
>>>>>>>
>>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a
>>>>>> path
>>>>>>
>>>>>
>>>>> The other solution (getting cf->data to be properly aligned) would be
>>>>> more efficient, why is not it possible to get cf->data properly aligned?
>>>>>
>>>> I think, i can't get cf->data to be properly aligned because "cf" is
>>>> receive internal frame representation within the ring buffer
>>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>>>> structure used in all rtcan drivers .
>>>
>>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for
>>> all drivers, so that looks like a worthwile improvement...
>>
>> I had a closer look. The CAN message (frame) queuing is optimized for
>> size to get as much as possible messages into the queue. Therefore the
>> CAN data are copied using memcpy, but only the real bytes. Directly
>> behind the real data is the 64 bit timestamp, if requested, which is
>> also not aligned. Therefore I suggest to use memcpy or byte accesses in
>> the flexcan driver as well. If this was a good decision is another question.
> 
> Aligning the data begining costs 2 bytes by struct rtcan_rb_frame. How
> many such structures are there for a typical driver?

Any message in the queue will require two bytes more. I did not say that
optimizing for size is necessary, but that's how it was implemented long
time ago. And I do not see a need to break that just because of the
Flexcan driver.

Wolfgang.

> 
> 

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/26/2014 09:16 PM, Wolfgang Grandegger wrote:
> On 01/26/2014 08:29 PM, Gilles Chanteperdrix wrote:
>> On 01/26/2014 08:26 PM, Wolfgang Grandegger wrote:
>>> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:
>>>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
>>>>>
>>>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>>>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>>>>>
>>>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>>>>>
>>>>>>>>>>> rtcan flexcan works perfectly.
>>>>>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>>>>>> put_unaligned instead of open coding it.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>>>>>
>>>>>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>>>>>> the one which is causing problems.
>>>>>>>>
>>>>>>>>
>>>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send a
>>>>>>> path
>>>>>>>
>>>>>>
>>>>>> The other solution (getting cf->data to be properly aligned) would be
>>>>>> more efficient, why is not it possible to get cf->data properly aligned?
>>>>>>
>>>>> I think, i can't get cf->data to be properly aligned because "cf" is
>>>>> receive internal frame representation within the ring buffer
>>>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic
>>>>> structure used in all rtcan drivers .
>>>>
>>>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned for
>>>> all drivers, so that looks like a worthwile improvement...
>>>
>>> I had a closer look. The CAN message (frame) queuing is optimized for
>>> size to get as much as possible messages into the queue. Therefore the
>>> CAN data are copied using memcpy, but only the real bytes. Directly
>>> behind the real data is the 64 bit timestamp, if requested, which is
>>> also not aligned. Therefore I suggest to use memcpy or byte accesses in
>>> the flexcan driver as well. If this was a good decision is another question.
>>
>> Aligning the data begining costs 2 bytes by struct rtcan_rb_frame. How
>> many such structures are there for a typical driver?
> 
> Any message in the queue will require two bytes more. I did not say that
> optimizing for size is necessary, but that's how it was implemented long
> time ago. And I do not see a need to break that just because of the
> Flexcan driver.

IMO, it is not just for the flexcan driver, it will avoid unaligned
accesses for all drivers which access data 4 bytes at a time. And I do
not think it "breaks" anything. What I do not know is:
- how many struct rtcan_rb_frame are typically in use? If only a few,
then adding 2 bytes should not be a problem.
- do all drivers access data 4 bytes at a time? or only the flexcan driver?

Regards.

-- 
                                                                Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Wolfgang Grandegger]

On Sun, 26 Jan 2014 21:56:10 +0100, Gilles Chanteperdrix

<gilles.chanteperdrix@xenomai.org> wrote:

> On 01/26/2014 09:16 PM, Wolfgang Grandegger wrote:

>> On 01/26/2014 08:29 PM, Gilles Chanteperdrix wrote:

>>> On 01/26/2014 08:26 PM, Wolfgang Grandegger wrote:

>>>> On 01/22/2014 03:12 PM, Gilles Chanteperdrix wrote:

>>>>> On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:

>>>>>>

>>>>>> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :

>>>>>>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:

>>>>>>>>

>>>>>>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :

>>>>>>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:

>>>>>>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;

>>>>>>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;

>>>>>>>>>>>>

>>>>>>>>>>>> rtcan flexcan works perfectly.

>>>>>>>>>>>> is  anyone know what is the problem with first instructions ?

>>>>>>>>>>> Probably mb->data does not have the right alignment. Could you

>>>>>>>>>>> not

>>>>>>>>>>> arrange to get it properly aligned? Failing that, you should

use

>>>>>>>>>>> put_unaligned instead of open coding it.

>>>>>>>>>>>

>>>>>>>>>>>

>>>>>>>>>> here structures used in flexcan driver, it seem to be aligned

>>>>>>>>>

>>>>>>>>> I am talking about the alignment of cf->data, since obviously,

>>>>>>>>> that is

>>>>>>>>> the one which is causing problems.

>>>>>>>>>

>>>>>>>>>

>>>>>>>> Thank a lot, put_unaligned fix this issue if you want, i can send

a

>>>>>>>> path

>>>>>>>>

>>>>>>>

>>>>>>> The other solution (getting cf->data to be properly aligned) would

>>>>>>> be

>>>>>>> more efficient, why is not it possible to get cf->data properly

>>>>>>> aligned?

>>>>>>>

>>>>>> I think, i can't get cf->data to be properly aligned because "cf"

is

>>>>>> receive internal frame representation within the ring buffer

>>>>>> of a struct rtcan_socket and struct rtcan_rb_frame is a generic

>>>>>> structure used in all rtcan drivers .

>>>>>

>>>>> Well, if you fix rtcan_rb_frame to be aligned, it will be aligned

for

>>>>> all drivers, so that looks like a worthwile improvement...

>>>>

>>>> I had a closer look. The CAN message (frame) queuing is optimized for

>>>> size to get as much as possible messages into the queue. Therefore

the

>>>> CAN data are copied using memcpy, but only the real bytes. Directly

>>>> behind the real data is the 64 bit timestamp, if requested, which is

>>>> also not aligned. Therefore I suggest to use memcpy or byte accesses

in

>>>> the flexcan driver as well. If this was a good decision is another

>>>> question.

>>>

>>> Aligning the data begining costs 2 bytes by struct rtcan_rb_frame. How

>>> many such structures are there for a typical driver?

>> 

>> Any message in the queue will require two bytes more. I did not say

that

>> optimizing for size is necessary, but that's how it was implemented

long

>> time ago. And I do not see a need to break that just because of the

>> Flexcan driver.

> 

> IMO, it is not just for the flexcan driver, it will avoid unaligned

> accesses for all drivers which access data 4 bytes at a time. And I do

> not think it "breaks" anything. What I do not know is:



But it's just the Flexcan driver which is doing unaligned accesses.



> - how many struct rtcan_rb_frame are typically in use? If only a few,

> then adding 2 bytes should not be a problem.



This depends on the size of the queue. But if the message size changes

it will also change run-time behaviour because less messages will fit

into the queue. That's what I'm more concerned about than the extra

2 bytes.



> - do all drivers access data 4 bytes at a time? or only the flexcan

driver?



No, as far as I have seen only the Flexan driver. But I need to check.



Wolfgang.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/27/2014 08:56 AM, Wolfgang Grandegger wrote:
> On Sun, 26 Jan 2014 21:56:10 +0100, Gilles Chanteperdrix
>> IMO, it is not just for the flexcan driver, it will avoid unaligned
> 
>> accesses for all drivers which access data 4 bytes at a time. And I do
> 
>> not think it "breaks" anything. What I do not know is:
> 
> 
> 
> But it's just the Flexcan driver which is doing unaligned accesses.

Ok, can we agree on the following patch then?

diff --git a/ksrc/drivers/can/rtcan_flexcan.c b/ksrc/drivers/can/rtcan_flexcan.c
index 8f0e8d1..1674b73 100644
--- a/ksrc/drivers/can/rtcan_flexcan.c
+++ b/ksrc/drivers/can/rtcan_flexcan.c
@@ -38,6 +38,7 @@
 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,5,0)
 #include <linux/pinctrl/consumer.h>
 #endif
+#include <asm/unaligned.h>
 
 #include <rtdm/rtdm_driver.h>
 
@@ -497,11 +498,9 @@ static void flexcan_rx_interrupt(struct rtcan_device *dev,
 		cf->can_id |= CAN_RTR_FLAG;
 		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE;
 	} else {
-		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc ;
-		*(__be32 *)(cf->data + 0) =
-			cpu_to_be32(flexcan_read(&mb->data[0]));
-		*(__be32 *)(cf->data + 4) =
-			cpu_to_be32(flexcan_read(&mb->data[1]));
+		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc;
+		put_unaligned_be32(flexcan_read(&mb->data[0]), cf->data + 0);
+		put_unaligned_be32(flexcan_read(&mb->data[1]), cf->data + 4);
 	}
 
 	/* Store the interface index */



-- 
                                                                Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Wolfgang Grandegger]

On Tue, 28 Jan 2014 23:06:54 +0100, Gilles Chanteperdrix

<gilles.chanteperdrix@xenomai.org> wrote:

> On 01/27/2014 08:56 AM, Wolfgang Grandegger wrote:

>> On Sun, 26 Jan 2014 21:56:10 +0100, Gilles Chanteperdrix

>>> IMO, it is not just for the flexcan driver, it will avoid unaligned

>> 

>>> accesses for all drivers which access data 4 bytes at a time. And I do

>> 

>>> not think it "breaks" anything. What I do not know is:

>> 

>> 

>> 

>> But it's just the Flexcan driver which is doing unaligned accesses.

> 

> Ok, can we agree on the following patch then?

> 

> diff --git a/ksrc/drivers/can/rtcan_flexcan.c

> b/ksrc/drivers/can/rtcan_flexcan.c

> index 8f0e8d1..1674b73 100644

> --- a/ksrc/drivers/can/rtcan_flexcan.c

> +++ b/ksrc/drivers/can/rtcan_flexcan.c

> @@ -38,6 +38,7 @@

>  #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,5,0)

>  #include <linux/pinctrl/consumer.h>

>  #endif

> +#include <asm/unaligned.h>

>  

>  #include <rtdm/rtdm_driver.h>

>  

> @@ -497,11 +498,9 @@ static void flexcan_rx_interrupt(struct

rtcan_device

> *dev,

>  		cf->can_id |= CAN_RTR_FLAG;

>  		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE;

>  	} else {

> -		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc ;

> -		*(__be32 *)(cf->data + 0) =

> -			cpu_to_be32(flexcan_read(&mb->data[0]));

> -		*(__be32 *)(cf->data + 4) =

> -			cpu_to_be32(flexcan_read(&mb->data[1]));

> +		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc;

> +		put_unaligned_be32(flexcan_read(&mb->data[0]), cf->data + 0);

> +		put_unaligned_be32(flexcan_read(&mb->data[1]), cf->data + 4);

>  	}

>  

>  	/* Store the interface index */



D'accord, thanks.



Wolfgang.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 29/01/2014 08:16, Wolfgang Grandegger a écrit :
> On Tue, 28 Jan 2014 23:06:54 +0100, Gilles Chanteperdrix
> <gilles.chanteperdrix@xenomai.org> wrote:
>> On 01/27/2014 08:56 AM, Wolfgang Grandegger wrote:
>>> On Sun, 26 Jan 2014 21:56:10 +0100, Gilles Chanteperdrix
>>>> IMO, it is not just for the flexcan driver, it will avoid unaligned
>>>> accesses for all drivers which access data 4 bytes at a time. And I do
>>>> not think it "breaks" anything. What I do not know is:
>>>
>>>
>>> But it's just the Flexcan driver which is doing unaligned accesses.
>> Ok, can we agree on the following patch then?
>>
>> diff --git a/ksrc/drivers/can/rtcan_flexcan.c
>> b/ksrc/drivers/can/rtcan_flexcan.c
>> index 8f0e8d1..1674b73 100644
>> --- a/ksrc/drivers/can/rtcan_flexcan.c
>> +++ b/ksrc/drivers/can/rtcan_flexcan.c
>> @@ -38,6 +38,7 @@
>>   #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,5,0)
>>   #include <linux/pinctrl/consumer.h>
>>   #endif
>> +#include <asm/unaligned.h>
>>   
>>   #include <rtdm/rtdm_driver.h>
>>   
>> @@ -497,11 +498,9 @@ static void flexcan_rx_interrupt(struct
> rtcan_device
>> *dev,
>>   		cf->can_id |= CAN_RTR_FLAG;
>>   		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE;
>>   	} else {
>> -		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc ;
>> -		*(__be32 *)(cf->data + 0) =
>> -			cpu_to_be32(flexcan_read(&mb->data[0]));
>> -		*(__be32 *)(cf->data + 4) =
>> -			cpu_to_be32(flexcan_read(&mb->data[1]));
>> +		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc;
>> +		put_unaligned_be32(flexcan_read(&mb->data[0]), cf->data + 0);
>> +		put_unaligned_be32(flexcan_read(&mb->data[1]), cf->data + 4);
>>   	}
>>   
>>   	/* Store the interface index */
> D'accord, thanks.
>
> Wolfgang.
>
Thank you all for your replies

Alexandre.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/29/2014 04:50 PM, Alexandre COFFIGNAL wrote:
>
> Le 29/01/2014 08:16, Wolfgang Grandegger a écrit :
>> On Tue, 28 Jan 2014 23:06:54 +0100, Gilles Chanteperdrix
>> <gilles.chanteperdrix@xenomai.org> wrote:
>>> On 01/27/2014 08:56 AM, Wolfgang Grandegger wrote:
>>>> On Sun, 26 Jan 2014 21:56:10 +0100, Gilles Chanteperdrix
>>>>> IMO, it is not just for the flexcan driver, it will avoid unaligned
>>>>> accesses for all drivers which access data 4 bytes at a time. And I do
>>>>> not think it "breaks" anything. What I do not know is:
>>>>
>>>>
>>>> But it's just the Flexcan driver which is doing unaligned accesses.
>>> Ok, can we agree on the following patch then?
>>>
>>> diff --git a/ksrc/drivers/can/rtcan_flexcan.c
>>> b/ksrc/drivers/can/rtcan_flexcan.c
>>> index 8f0e8d1..1674b73 100644
>>> --- a/ksrc/drivers/can/rtcan_flexcan.c
>>> +++ b/ksrc/drivers/can/rtcan_flexcan.c
>>> @@ -38,6 +38,7 @@
>>>    #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,5,0)
>>>    #include <linux/pinctrl/consumer.h>
>>>    #endif
>>> +#include <asm/unaligned.h>
>>>
>>>    #include <rtdm/rtdm_driver.h>
>>>
>>> @@ -497,11 +498,9 @@ static void flexcan_rx_interrupt(struct
>> rtcan_device
>>> *dev,
>>>    		cf->can_id |= CAN_RTR_FLAG;
>>>    		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE;
>>>    	} else {
>>> -		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc ;
>>> -		*(__be32 *)(cf->data + 0) =
>>> -			cpu_to_be32(flexcan_read(&mb->data[0]));
>>> -		*(__be32 *)(cf->data + 4) =
>>> -			cpu_to_be32(flexcan_read(&mb->data[1]));
>>> +		skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc;
>>> +		put_unaligned_be32(flexcan_read(&mb->data[0]), cf->data + 0);
>>> +		put_unaligned_be32(flexcan_read(&mb->data[1]), cf->data + 4);
>>>    	}
>>>
>>>    	/* Store the interface index */
>> D'accord, thanks.
>>
>> Wolfgang.
>>
> Thank you all for your replies

Hi Alexandre,

Did you check the last patch? Does it compile and work for you?

Regards.

-- 
					    Gilles.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Alexandre COFFIGNAL]

Le 29/01/2014 18:58, Gilles Chanteperdrix a écrit :
> On 01/29/2014 04:50 PM, Alexandre COFFIGNAL wrote:
>>
>> Le 29/01/2014 08:16, Wolfgang Grandegger a écrit :
>>> On Tue, 28 Jan 2014 23:06:54 +0100, Gilles Chanteperdrix
>>> <gilles.chanteperdrix@xenomai.org> wrote:
>>>> On 01/27/2014 08:56 AM, Wolfgang Grandegger wrote:
>>>>> On Sun, 26 Jan 2014 21:56:10 +0100, Gilles Chanteperdrix
>>>>>> IMO, it is not just for the flexcan driver, it will avoid unaligned
>>>>>> accesses for all drivers which access data 4 bytes at a time. And 
>>>>>> I do
>>>>>> not think it "breaks" anything. What I do not know is:
>>>>>
>>>>>
>>>>> But it's just the Flexcan driver which is doing unaligned accesses.
>>>> Ok, can we agree on the following patch then?
>>>>
>>>> diff --git a/ksrc/drivers/can/rtcan_flexcan.c
>>>> b/ksrc/drivers/can/rtcan_flexcan.c
>>>> index 8f0e8d1..1674b73 100644
>>>> --- a/ksrc/drivers/can/rtcan_flexcan.c
>>>> +++ b/ksrc/drivers/can/rtcan_flexcan.c
>>>> @@ -38,6 +38,7 @@
>>>>    #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,5,0)
>>>>    #include <linux/pinctrl/consumer.h>
>>>>    #endif
>>>> +#include <asm/unaligned.h>
>>>>
>>>>    #include <rtdm/rtdm_driver.h>
>>>>
>>>> @@ -497,11 +498,9 @@ static void flexcan_rx_interrupt(struct
>>> rtcan_device
>>>> *dev,
>>>>            cf->can_id |= CAN_RTR_FLAG;
>>>>            skb->rb_frame_size = EMPTY_RB_FRAME_SIZE;
>>>>        } else {
>>>> -        skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc ;
>>>> -        *(__be32 *)(cf->data + 0) =
>>>> -            cpu_to_be32(flexcan_read(&mb->data[0]));
>>>> -        *(__be32 *)(cf->data + 4) =
>>>> -            cpu_to_be32(flexcan_read(&mb->data[1]));
>>>> +        skb->rb_frame_size = EMPTY_RB_FRAME_SIZE + cf->can_dlc;
>>>> + put_unaligned_be32(flexcan_read(&mb->data[0]), cf->data + 0);
>>>> + put_unaligned_be32(flexcan_read(&mb->data[1]), cf->data + 4);
>>>>        }
>>>>
>>>>        /* Store the interface index */
>>> D'accord, thanks.
>>>
>>> Wolfgang.
>>>
>> Thank you all for your replies
>
> Hi Alexandre,
>
> Did you check the last patch? Does it compile and work for you?
>
> Regards.
>

I tested this patch, it works perfectly on i.MX28. Thank you very much.


Regards.

Re: [Xenomai] imx28 rtcan flexcan system freezes

[Gilles Chanteperdrix]

On 01/22/2014 03:04 PM, Alexandre COFFIGNAL wrote:
> 
> Le 22/01/2014 14:30, Gilles Chanteperdrix a écrit :
>> On 01/22/2014 02:27 PM, Alexandre COFFIGNAL wrote:
>>>
>>> Le 22/01/2014 13:00, Gilles Chanteperdrix a écrit :
>>>> On 01/22/2014 12:36 PM, Alexandre COFFIGNAL wrote:
>>>>>>> +        cf->data[3]=((data0 >> 0)  & 0xFF) ;
>>>>>>> +        cf->data[2]=((data0 >> 8)  & 0xFF) ;
>>>>>>> +        cf->data[1]=((data0 >> 16)  & 0xFF) ;
>>>>>>> +        cf->data[0]=((data0 >> 24)  & 0xFF) ;
>>>>>>> +        cf->data[7]=((data1 >> 0)  & 0xFF) ;
>>>>>>> +        cf->data[6]=((data1 >> 8)  & 0xFF) ;
>>>>>>> +        cf->data[5]=((data1 >> 16)  & 0xFF) ;
>>>>>>> +        cf->data[4]=((data1 >> 24)  & 0xFF) ;
>>>>>>>
>>>>>>> rtcan flexcan works perfectly.
>>>>>>> is  anyone know what is the problem with first instructions ?
>>>>>> Probably mb->data does not have the right alignment. Could you not
>>>>>> arrange to get it properly aligned? Failing that, you should use
>>>>>> put_unaligned instead of open coding it.
>>>>>>
>>>>>>
>>>>> here structures used in flexcan driver, it seem to be aligned
>>>>
>>>> I am talking about the alignment of cf->data, since obviously, that is
>>>> the one which is causing problems.
>>>>
>>>>
>>> Thank a lot, put_unaligned fix this issue if you want, i can send a path
>>>
>>
>> The other solution (getting cf->data to be properly aligned) would be 
>> more efficient, why is not it possible to get cf->data properly aligned?
>>
> I think, i can't get cf->data to be properly aligned because "cf" is 
> receive internal frame representation within the ring buffer
> of a struct rtcan_socket and struct rtcan_rb_frame is a generic 
> structure used in all rtcan drivers .

Hi Alexandr,

Could you try the following patch?

diff --git a/ksrc/drivers/can/rtcan_socket.h
b/ksrc/drivers/can/rtcan_socket.h
index e0a395d..8984121 100644
--- a/ksrc/drivers/can/rtcan_socket.h
+++ b/ksrc/drivers/can/rtcan_socket.h
@@ -77,6 +77,7 @@ struct rtcan_rb_frame {
      * existence of a timestamp is indicated by the RTCAN_HAS_TIMESTAMP
      * bit. */
     unsigned char       can_dlc;
+    unsigned short	pad;

     /* Data bytes */
     uint8_t             data[8];

Regards.

-- 
                                                                Gilles.