Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> Gilles Chanteperdrix wrote:
>>> Gilles Chanteperdrix wrote:
>>>> Jan Kiszka wrote:
>>>>> Gilles Chanteperdrix wrote:
>>>>>> I get this warning when compiling trunk:
>>>>>>
>>>>>> checking for __thread... rm: cannot remove `conftest1.dir': Is a directory
>>>>>> yes
>>>>>>
>>>>> That simple, it's a typo in this line:
>>>>>
>>>>> http://www.rts.uni-hannover.de/xenomai/lxr/source/configure.in?v=SVN-trunk#394
>>>>>
>>>>> Also, that rm should be done outside the if block to play safe.
>>>> Ok. But once this is fixed, I get a segmentation fault with __thread on
>>>> ARM in rt_task_trampoline. I am currently testing the SIGWINCH thing,
>>>> and if __thread is disabled, there is no segfault.
>>> If a #ifdef 0 xeno_set_current, there is no segfault.
>>>
>> Interesting. What about removing the initial-exec attributes? Is
>> __thread otherwise know to work on your platform?
> 
> Yes. This is a race condition, the real bug is, IMHO:
> 	__native_self = *iargs->task;
> When the segfault happens iargs->task is NULL. I suspect you are not
> supposed to use iargs after the __native_task_create syscall.

Don't see the race yet. iargs is provided by the trampoline creator, and
with that one we synchronize with the barrier AFTER this access - or
does the compiler some fancy reordering here?

Jan