Re: [Xenomai] Sharing memory between rt-posix and standard-posix apps

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

On 06/20/2012 09:33 AM, Wolfgang Grandegger wrote:
> On 06/19/2012 11:38 AM, Gilles Chanteperdrix wrote:
>> On 06/19/2012 11:29 AM, Wolfgang Grandegger wrote:
>>> On 06/19/2012 11:04 AM, Gilles Chanteperdrix wrote:
>>>> On 06/19/2012 10:52 AM, Wolfgang Grandegger wrote:
>>>>> On 06/19/2012 10:50 AM, Gilles Chanteperdrix wrote:
>>>>>> On 06/19/2012 10:49 AM, Wolfgang Grandegger wrote:
>>>>>>> On 06/19/2012 10:33 AM, Gilles Chanteperdrix wrote:
>>>>>>>> On 06/19/2012 10:25 AM, Wolfgang Grandegger wrote:
>>>>>>>>> Hello,
>>>>>>>>>
>>>>>>>>> the shared memory support of Xenomai's POSIX skin currently only support
>>>>>>>>> shared memory between rt processes.
>>>>>>>>
>>>>>>>> These services are for sharing memory between user-space and
>>>>>>>> kernel-space. In order to share memory between user-space processes, you
>>>>>>>> can use plain linux shared memories.
>>>>>>>
>>>>>>> But then I will get mode switches due to page faults, right?
>>>>>>
>>>>>> No, there should not be any page fault.
>>>>>
>>>>> But that's what we observe on a i.MX53 and i.MX6Q board. IIRC, the first
>>>>> access usually results in a page fault.
>>>>
>>>> And you do not get the same behaviour with posix skin shared memories ?
>>>
>>> Right, and that's because do_pagefault() is called somewhere in the
>>> initialization, IIRC. But I will redo my tests now.
>>
>> Yes, I forgot about that, we pre-fault shared memory on ARM, in order to
>> avoid the page fault on first access.
>>
>>>
>>> Anyway, if I use the POSIX skin the shared memory functions in there
>>> will be used automatically. But I could use  __real_shm_open and
>>> __real_mmap, right?
>>
>> Or you can used the wrapped version, but without enabling support for
>> xenomai posix shared memories in the kernel configuration.
>>
>>
>>  ... and access the page once per page in the
>>> initialization phase. Then no more page faults should happen.
>>
>> Yes. It is a bit ugly, but should work.
> 
> I forgot. The problem was not the read but the first write to a page
> triggering the page fault. Therefore it would make sense to wrap just
> mmap and doing the pre-faulting if XENO_OPT_POSIX_SHM is not set.

Yes, when we prefault in kernel-space, we prefault write faults. We can
do that, yes.

-- 
					    Gilles.