Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[Andi Kleen]

"Robert Williamson" <robbiew@us.ibm.com> writes:
> 
> Obviously, we could add additional code to check for the running arch and
> define the syscall accordingly, however I'm not sure this is the correct
> way to go.  I'm copying our list, as well as the kernel mailing list about
> this, because I "think" the system calls should be defined to the same
> numbers across all architectures....but I'm not positive.  BTW,  I attached

No. Linux has traditionally used different syscall numbers for different
architectures. The original ports (alpha etc.) always used the syscall numbers
of the "native" Unix, so the numbering was often completely different.
Newer ports who weren't concered about such compatibility often did 
a renumbering too. For example x86-64 has a completely new 
"cache line optimized" ordering.

What should work on most architectures is 
(most = someone told me it doesn't work properly on IA64)

#include </path/to/kernel/source/include/asm-<arch>/unistd.h>
(you need the version in the kernel source because many glibc packagers
in their infinite wisdom use an old outdated copy of asm/ usually from
the last stable kernel only) 

_syscallN(...) 

-Andi

Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[Robert Williamson]

Hmmm...so I guess the only viable solution for a single test to cover as
many archs as possible, is to explicitly define the number for each arch.
Here's how I would do it Aniruddha:
------------------------------------------
#ifdef __i386__
#define __NR_timer_create 259
#endif

#ifdef __x86_64__
#define __NR_timer_create 222
#endif

#if defined(__ppc__) || defined(__ppc64__)
#define __NR_timer_create 240
#else /* Not defined on this architecture */

#include "test.h"
#include "usctest.h"

int TST_TOTAL = 0;      /* Total number of testcases */

int main()
{
  tst_resm(TCONF,"This system call is not defined for this architecture.");
  tst_exit();
  /* NOT REACHED */
  return(0);
}
#endif /* Not defined on this architecture */

<REST OF TEST HERE>
------------------------------------------

Any comments???

- Robbie

Robert V. Williamson <robbiew@us.ibm.com>
Linux Test Project
IBM Linux Technology Center
Phone: (512) 838-9295   T/L: 678-9295
Fax: (512) 838-4603
Web: http://ltp.sourceforge.net
IRC: #ltp on freenode.irc.net
====================
"Only two things are infinite, the universe and human stupidity, and I'm
not sure about the former." -Albert Einstein


                                                                                                                                      
                      Andi Kleen                                                                                                      
                      <ak@suse.de>             To:       Robert Williamson/Austin/IBM@IBMUS                                           
                      Sent by:                 cc:       linux-kernel@vger.kernel.org, aniruddha.marathe@wipro.com,                   
                      ak@suse.de                ltp-list@lists.sourceforge.net                                                        
                                               Subject:  Re: Same syscall is defined to different numbers on 3 different archs(was    
                                                Re: Makefile  issue)                                                                  
                      04/07/2003 10:54                                                                                                
                      AM                                                                                                              
                                                                                                                                      




"Robert Williamson" <robbiew@us.ibm.com> writes:
>
> Obviously, we could add additional code to check for the running arch and
> define the syscall accordingly, however I'm not sure this is the correct
> way to go.  I'm copying our list, as well as the kernel mailing list
about
> this, because I "think" the system calls should be defined to the same
> numbers across all architectures....but I'm not positive.  BTW,  I
attached

No. Linux has traditionally used different syscall numbers for different
architectures. The original ports (alpha etc.) always used the syscall
numbers
of the "native" Unix, so the numbering was often completely different.
Newer ports who weren't concered about such compatibility often did
a renumbering too. For example x86-64 has a completely new
"cache line optimized" ordering.

What should work on most architectures is
(most = someone told me it doesn't work properly on IA64)

#include </path/to/kernel/source/include/asm-<arch>/unistd.h>
(you need the version in the kernel source because many glibc packagers
in their infinite wisdom use an old outdated copy of asm/ usually from
the last stable kernel only)

_syscallN(...)

-Andi

Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[David Mosberger]

>>>>> On Mon, 7 Apr 2003 11:26:49 -0500, "Robert Williamson" <robbiew@us.ibm.com> said:

  Robert> Hmmm...so I guess the only viable solution for a single test
  Robert> to cover as many archs as possible, is to explicitly define
  Robert> the number for each arch.

  Robert> Here's how I would do it Aniruddha:
  Robert> ------------------------------------------
  Robert> #ifdef __i386__
  Robert> #define __NR_timer_create 259
  Robert> #endif

  Robert> #ifdef __x86_64__
  Robert> #define __NR_timer_create 222
  Robert> #endif

  Robert> #if defined(__ppc__) || defined(__ppc64__)
  Robert> #define __NR_timer_create 240
  Robert> #else /* Not defined on this architecture */

  Robert> #include "test.h"
  Robert> #include "usctest.h"

  Robert> int TST_TOTAL = 0;      /* Total number of testcases */

  Robert> int main()
  Robert> {
  Robert> tst_resm(TCONF,"This system call is not defined for this architecture.");
  Robert> tst_exit();
  Robert> /* NOT REACHED */
  Robert> return(0);
  Robert> }
  Robert> #endif /* Not defined on this architecture */

  Robert> <REST OF TEST HERE>
  Robert> ------------------------------------------

  Robert> Any comments???

Why use such ugly, platform-dependent code when syscall(3) will do it
just fine?  (AFAIK, there is no man-page for syscall(3), but the glibc
info manual documents it in detail.)

	--david

Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[Robert Williamson]

The original note had the testcases in question, which uses
_syscall3......but since this was left off I'll summarize.  The author did
use the following:

_syscall3 (int, timer_create, clockid_t, which_clock, struct sigevent *,
      timer_event_spec, timer_t *, timer_id);

but compilation failed with:

timer_delete01.c: In function `timer_create':
timer_delete01.c:86: `__NR_timer_create' undeclared (first use in this
function)

The only way we were able to resolve this was to either to add the kernel
includes to the include path: "-I /usr/src/linux-2.5.66/include".
Obviously, this option will not work for the LTP and our users who
frequently change kernel levels and install locations, we have to use "such
ugly, platform-dependent code"

- Robbie

Robert V. Williamson <robbiew@us.ibm.com>
Linux Test Project
IBM Linux Technology Center
Phone: (512) 838-9295   T/L: 678-9295
Fax: (512) 838-4603
Web: http://ltp.sourceforge.net
IRC: #ltp on freenode.irc.net
====================
"Only two things are infinite, the universe and human stupidity, and I'm
not sure about the former." -Albert Einstein


                                                                                                                                      
                      David Mosberger                                                                                                 
                      <davidm@napali.hp        To:       Robert Williamson/Austin/IBM@IBMUS                                           
                      l.hp.com>                cc:       Andi Kleen <ak@suse.de>, aniruddha.marathe@wipro.com,                        
                                                linux-kernel@vger.kernel.org, ltp-list@lists.sourceforge.net                          
                      04/07/2003 01:08         Subject:  Re: Same syscall is defined to different numbers on 3 different archs(was    
                      PM                        Re: Makefile  issue)                                                                  
                      Please respond to                                                                                               
                      davidm                                                                                                          
                                                                                                                                      




>>>>> On Mon, 7 Apr 2003 11:26:49 -0500, "Robert Williamson"
<robbiew@us.ibm.com> said:

  Robert> Hmmm...so I guess the only viable solution for a single test
  Robert> to cover as many archs as possible, is to explicitly define
  Robert> the number for each arch.

  Robert> Here's how I would do it Aniruddha:
  Robert> ------------------------------------------
  Robert> #ifdef __i386__
  Robert> #define __NR_timer_create 259
  Robert> #endif

  Robert> #ifdef __x86_64__
  Robert> #define __NR_timer_create 222
  Robert> #endif

  Robert> #if defined(__ppc__) || defined(__ppc64__)
  Robert> #define __NR_timer_create 240
  Robert> #else /* Not defined on this architecture */

  Robert> #include "test.h"
  Robert> #include "usctest.h"

  Robert> int TST_TOTAL = 0;      /* Total number of testcases */

  Robert> int main()
  Robert> {
  Robert> tst_resm(TCONF,"This system call is not defined for this
architecture.");
  Robert> tst_exit();
  Robert> /* NOT REACHED */
  Robert> return(0);
  Robert> }
  Robert> #endif /* Not defined on this architecture */

  Robert> <REST OF TEST HERE>
  Robert> ------------------------------------------

  Robert> Any comments???

Why use such ugly, platform-dependent code when syscall(3) will do it
just fine?  (AFAIK, there is no man-page for syscall(3), but the glibc
info manual documents it in detail.)

             --david

Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[David Mosberger]

>>>>> On Mon, 7 Apr 2003 13:16:21 -0500, "Robert Williamson" <robbiew@us.ibm.com> said:

  Robert> The original note had the testcases in question, which uses
  Robert> _syscall3......but since this was left off I'll summarize.
  Robert> The author did use the following:

  Robert> _syscall3 (int, timer_create, clockid_t, which_clock, struct sigevent *,
  Robert> timer_event_spec, timer_t *, timer_id);

  Robert> but compilation failed with:

  Robert> timer_delete01.c: In function `timer_create':
  Robert> timer_delete01.c:86: `__NR_timer_create' undeclared (first use in this
  Robert> function)

  Robert> The only way we were able to resolve this was to either to
  Robert> add the kernel includes to the include path: "-I
  Robert> /usr/src/linux-2.5.66/include".  Obviously, this option will
  Robert> not work for the LTP and our users who frequently change
  Robert> kernel levels and install locations, we have to use "such
  Robert> ugly, platform-dependent code"

But you still can use syscall() instead of the non-portable syscallN()
macros.  Also, it should go something like this:

#include <sys/syscall.h>

#ifndef SYS_timer_create
# if defined(__i386)
#  define SYS_timer_create	259
# elif defined(...)
   ...
# endif
#endif

	--david

Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[Werner Almesberger]

Andi Kleen wrote:
> #include </path/to/kernel/source/include/asm-<arch>/unistd.h>
> (you need the version in the kernel source because many glibc packagers
> in their infinite wisdom use an old outdated copy of asm/ usually from
> the last stable kernel only) 

Do we need a /proc/syscalls ? :-)

- Werner

-- 
  _________________________________________________________________________
 / Werner Almesberger, Buenos Aires, Argentina         wa@almesberger.net /
/_http://www.almesberger.net/____________________________________________/

Re: [LTP] Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[Paul Larson]

[-- Attachment #1: Type: text/plain, Size: 691 bytes --]

On Mon, 2003-04-07 at 21:23, Werner Almesberger wrote:
> Andi Kleen wrote:
> > #include </path/to/kernel/source/include/asm-<arch>/unistd.h>
> > (you need the version in the kernel source because many glibc packagers
> > in their infinite wisdom use an old outdated copy of asm/ usually from
> > the last stable kernel only) 
> 
> Do we need a /proc/syscalls ? :-)
I don't think so.  Apps should be accessing things through libraries
anyway.  The only reason we don't in some cases like this is that it's
new and not in libs on any distro yet.  Besides, I don't think having a
/proc/syscalls would be any better than having to do ifdefs or use
kernel headers.

-Paul Larson


[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 240 bytes --]

Re: [LTP] Re: Same syscall is defined to different numbers on 3 different archs(was Re: Makefile issue)

[Werner Almesberger]

Paul Larson wrote:
> I don't think so.  Apps should be accessing things through libraries
> anyway.  The only reason we don't in some cases like this is that it's
> new and not in libs on any distro yet.

Well yes, if the syscall is correctly implemented in the library,
there's no problem.

> Besides, I don't think having a
> /proc/syscalls would be any better than having to do ifdefs or use
> kernel headers.

#ifdef may be hard pressed to identify a specific kernel with a
specific patch. Kernel headers are obviously a solution (although
this violates the "user space must never include kernel headers"
pseudo-rule), but add the problem of needing to identify the
"current" source tree, which is a configuration step almost always
requiring manual intervention.

My suggestion for /proc/syscalls isn't entirely serious, but it 
would be nice if we could eventually solve this problem ...

- Werner

-- 
  _________________________________________________________________________
 / Werner Almesberger, Buenos Aires, Argentina         wa@almesberger.net /
/_http://www.almesberger.net/____________________________________________/