Re: [Xenomai-help] compiling xenomai on x86_64

[Jan Kiszka <kiszka@domain.hid>]

[-- Attachment #1: Type: text/plain, Size: 2272 bytes --]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>  > I just analysed this further: the problem disappears when I manually
>  > remove any "-I<path-to-kernel-includes>" from the makefiles. Seems to be
>  > the well-known user space include issue again... ;)
> 
> I tried to compile every kernel arch/i386 has a patch for with gcc 2.95,
> 3.3 and 3.4, and I saw no issue with inclusions of kernel headers
> from user-space. At all.

I think you were lucky. As far as I remember, it is generally no good
idea to include kernel header for userspace compilation as long as you
do not _exactly_ know (or test) what they bring with them. Some
structures or defines locate in different headers under /usr/include
compared to some kernel tree. So, if you are not lucky, you drag those
definitions in twice - see below.

> 
> So, there is definitely something you are doing differently from me. 
> If this issue is different from the '-m32' flag issue, could you please
> fill a bug report, and this time, attach the exact procedure you are
> following to compile (whether in-tree, or out-of-tree, whether using
> configure or Kconfig, everything that would allow someone to ), the
> .config and .xeno_config files you used, in order for me to be able to
> reproduce the problem you observed ?
> 

As I'm not sure if you can reproduce it easily without my environment
(that damn x86_64 gcc4 on SuSE10), I digged a little bit deeper, and
here is a new theory:

The compilation fails in skins/native/libs/task.c (there exist similar
files in other skins as well). The error: structures in
<linux>/include/asm/sigcontext.h are being redefined. Unfortunately, gcc
does not tell me where the first definition was. But here is my finding:
commenting out #include <signal.h> in libs/task.c makes that error
disappear. Following the include chain of /usr/include/signal.h
(userspace headers...), you get to /usr/include/bits/sigcontext.h.
Bingo. On the x86_64 box all structs are defined here according to the
target address width. And as /usr/include/bits/sigcontext.h and
<linux>/include/asm/sigcontext.h do not exclude each other, we get
redefines.

To sum up, I still think that kernel headers need to be avoided in
userspace. Or we have to work around such issues case-by-case...

Jan

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 254 bytes --]