The second version of this patch with Linus' comments addressed.  I've
changed the code for a few architectures to use a new name (<arch>_pipe
instead of sys_pipe) as per Linus' suggestion.  I haven't touched MIPS
and Sparc, they use too much magic in this area.  Should be easy enough
for the arch maintainer to clean up.  sys_pipe is now unconditionally
created in fs/pipe.c and they will lead to link errors in case the
arch code isn't cleaned up.

Some arch maintainers might want to take a look at their remaining
special versions.  The cris version, for instance, only differs from
the generic version in that it takes the BKL.  Is this still needed
these days?


 arch/alpha/kernel/entry.S         |    8 ++++----
 arch/alpha/kernel/systbls.S       |    2 +-
 arch/arm/kernel/sys_arm.c         |   17 -----------------
 arch/avr32/kernel/sys_avr32.c     |   13 -------------
 arch/blackfin/kernel/sys_bfin.c   |   17 -----------------
 arch/cris/arch-v10/kernel/entry.S |    2 +-
 arch/cris/arch-v32/kernel/entry.S |    2 +-
 arch/cris/kernel/sys_cris.c       |    4 ++--
 arch/frv/kernel/sys_frv.c         |   17 -----------------
 arch/h8300/kernel/sys_h8300.c     |   17 -----------------
 arch/ia64/kernel/entry.S          |    2 +-
 arch/ia64/kernel/sys_ia64.c       |    2 +-
 arch/m32r/kernel/sys_m32r.c       |    4 ++--
 arch/m32r/kernel/syscall_table.S  |    2 +-
 arch/m68k/kernel/sys_m68k.c       |   17 -----------------
 arch/m68knommu/kernel/sys_m68k.c  |   17 -----------------
 arch/mn10300/kernel/sys_mn10300.c |   17 -----------------
 arch/parisc/kernel/sys_parisc.c   |   13 -------------
 arch/powerpc/kernel/syscalls.c    |   17 -----------------
 arch/s390/kernel/entry.h          |    1 -
 arch/s390/kernel/sys_s390.c       |   17 -----------------
 arch/sh/kernel/sys_sh32.c         |    4 ++--
 arch/sh/kernel/sys_sh64.c         |   17 -----------------
 arch/sh/kernel/syscalls_32.S      |    2 +-
 arch/um/kernel/syscall.c          |   17 -----------------
 arch/v850...
[ message continues ]

On Sat, May 3, 2008 at 8:01 PM, Ulrich Drepper <drepper@redhat.com> wrote:
[...]

I realize this code is old, but wouldn't file descriptors leak if
copy_to_user fails?

BR,
dm.n9107
--

The MIPS implementation doesn't have this problem; it returns the
file descriptors in the result registers $v0 and $v1.

But an interesting catch after so many years!

  Ralf
--

Yes, it does -- I just tested this.

-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Found a bug? http://www.kernel.org/doc/man-pages/reporting_bugs.html
--

No, definitely not.

That said, I think that in order to not break other architectures, and to 
make it even easier to do this transformation, how about we just mark the 
generic version with __weak?

That way, odd architectures can just continue to call their own versions 
"sys_pipe()", and we don't break them unnecessarily.

		Linus
--