"do section mismatch check on full vmlinux" breaks powerpc build

Hello-

2.6.23-rc1 breaks the build for 64-bit powerpc for me (using
maple_defconfig):

  LD      vmlinux.o
powerpc64-unknown-linux-gnu-ld: dynreloc miscount for
kernel/built-in.o, section .opd
powerpc64-unknown-linux-gnu-ld: can not edit opd Bad value
make: *** [vmlinux.o] Error 1

This is on a i386 host with:
powerpc64-unknown-linux-gnu-gcc (GCC) 4.1.2
GNU ld version 2.16.1

Reverting the following commit fixes it:

commit 741f98fe298a73c9d47ed53703c1279a29718581
Author: Sam Ravnborg [off-list ref]
Date:   Tue Jul 17 10:54:06 2007 +0200

    kbuild: do section mismatch check on full vmlinux


However, I see a possibly related binutils patch:
http://article.gmane.org/gmane.comp.gnu.binutils/33650

Will there be a kbuild fix for this or should I update my binutils?

On Tue, Jul 24, 2007 at 05:41:05PM -0500, Nathan Lynch wrote:

Hello-

2.6.23-rc1 breaks the build for 64-bit powerpc for me (using
maple_defconfig):

  LD      vmlinux.o
powerpc64-unknown-linux-gnu-ld: dynreloc miscount for
kernel/built-in.o, section .opd
powerpc64-unknown-linux-gnu-ld: can not edit opd Bad value
make: *** [vmlinux.o] Error 1

This is on a i386 host with:
powerpc64-unknown-linux-gnu-gcc (GCC) 4.1.2
GNU ld version 2.16.1

Reverting the following commit fixes it:

commit 741f98fe298a73c9d47ed53703c1279a29718581
Author: Sam Ravnborg [off-list ref]
Date:   Tue Jul 17 10:54:06 2007 +0200

    kbuild: do section mismatch check on full vmlinux


However, I see a possibly related binutils patch:
http://article.gmane.org/gmane.comp.gnu.binutils/33650

Will there be a kbuild fix for this or should I update my binutils?

What kbuild does is a very basic link of a number of .o files.
Why this suddenly breaks I do not fully understand and I see no obvious
way to fix it in kbuild.

And browsing arch/ppc/kernel/vmlinux.lds.S I see no special handling of .opd.

Puzzeled?
I'm not sure how to deal with this one. Would prefer to say that this is
a binutils bug so fix it there but then I like to provide a workaround
for those with faulty toolchains.
I could put the link and the additional modpost check inside a
CONFIG_SOMETHING....

	Sam

On Tue, Jul 24, 2007 at 05:41:05PM -0500, Nathan Lynch wrote:

Hello-

2.6.23-rc1 breaks the build for 64-bit powerpc for me (using
maple_defconfig):

  LD      vmlinux.o
powerpc64-unknown-linux-gnu-ld: dynreloc miscount for
kernel/built-in.o, section .opd
powerpc64-unknown-linux-gnu-ld: can not edit opd Bad value
make: *** [vmlinux.o] Error 1

This is on a i386 host with:
powerpc64-unknown-linux-gnu-gcc (GCC) 4.1.2
GNU ld version 2.16.1

Reverting the following commit fixes it:

commit 741f98fe298a73c9d47ed53703c1279a29718581
Author: Sam Ravnborg [off-list ref]
Date:   Tue Jul 17 10:54:06 2007 +0200

    kbuild: do section mismatch check on full vmlinux


However, I see a possibly related binutils patch:
http://article.gmane.org/gmane.comp.gnu.binutils/33650

Will there be a kbuild fix for this or should I update my binutils?

Did a bit more testing.
Im am down to the following to reproduce the fault:
 /opt/crosstool/gcc-4.1.0-glibc-2.3.6/powerpc64-unknown-linux-gnu/bin/powerpc64-unknown-linux-gnu-ld -m elf64ppc  -r -o vmlinux.o kernel/signal.o arch/powerpc/kernel/vdso.o

Seems that vdso.o and signal.o does not like each other??

	Sam

On Tue, Jul 24, 2007 at 05:41:05PM -0500, Nathan Lynch wrote:

Hello-

2.6.23-rc1 breaks the build for 64-bit powerpc for me (using
maple_defconfig):

  LD      vmlinux.o
powerpc64-unknown-linux-gnu-ld: dynreloc miscount for
kernel/built-in.o, section .opd
powerpc64-unknown-linux-gnu-ld: can not edit opd Bad value
make: *** [vmlinux.o] Error 1

This is on a i386 host with:
powerpc64-unknown-linux-gnu-gcc (GCC) 4.1.2
GNU ld version 2.16.1

Reverting the following commit fixes it:

commit 741f98fe298a73c9d47ed53703c1279a29718581
Author: Sam Ravnborg [off-list ref]
Date:   Tue Jul 17 10:54:06 2007 +0200

    kbuild: do section mismatch check on full vmlinux


However, I see a possibly related binutils patch:
http://article.gmane.org/gmane.comp.gnu.binutils/33650

Will there be a kbuild fix for this or should I update my binutils?

I tried searching a bit more to find the culprint.
I narrowed it down to the following change to avoid the breakage:

diff --git a/include/linux/mm.h b/include/linux/mm.h
index c456c3a..2ea222f 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h

@@ -1246,7 +1246,7 @@ void drop_slab(void);
 extern int randomize_va_space;
 #endif
 
-__attribute__((weak)) const char *arch_vma_name(struct vm_area_struct *vma);
+//__attribute__((weak)) const char *arch_vma_name(struct vm_area_struct *vma);
 
 #endif /* __KERNEL__ */
 #endif /* _LINUX_MM_H */

So seems that something goes a bit fishy when using weak symbols and this trigges 
a binutils bug.

The above line was introdused in the following commit:

commit f269fdd1829acc5e53bf57b145003e5733133f2b
Author: David Howells [off-list ref]
Date:   Wed Sep 27 01:50:23 2006 -0700

    [PATCH] NOMMU: move the fallback arch_vma_name() to a sensible place
    
    Move the fallback arch_vma_name() to a sensible place (kernel/signal.c).
    
    Currently it's in fs/proc/task_mmu.c, a file that is dependent on both
    CONFIG_PROC_FS and CONFIG_MMU being enabled, but it's used from
    kernel/signal.c from where it is called unconditionally.
    
    [akpm@osdl.org: build fix]
    Signed-off-by: David Howells [off-list ref]
    Signed-off-by: Andrew Morton [off-list ref]
    Signed-off-by: Linus Torvalds [off-list ref]


	Sam

Hi Sam-

Sam Ravnborg wrote:

Thanks for looking into this.  Removing the "__attribute__((weak))"
from arch_vma_name's declaration in linux/mm.h unbreaks the build for
me.

Maybe it shouldn't matter, but it seems unusual to have the weak
attribute specified at the function's declaration.  I wasn't able to
find any examples of that for other weak functions in the kernel
(e.g. sched_clock).

On Wed, Jul 25, 2007 at 04:16:10PM -0500, Nathan Lynch wrote:

Hi Sam-

Sam Ravnborg wrote:

quoted

On Tue, Jul 24, 2007 at 05:41:05PM -0500, Nathan Lynch wrote:

quoted

2.6.23-rc1 breaks the build for 64-bit powerpc for me (using
maple_defconfig):

  LD      vmlinux.o
powerpc64-unknown-linux-gnu-ld: dynreloc miscount for
kernel/built-in.o, section .opd
powerpc64-unknown-linux-gnu-ld: can not edit opd Bad value
make: *** [vmlinux.o] Error 1

I narrowed it down to the following change to avoid the breakage:

diff --git a/include/linux/mm.h b/include/linux/mm.h
index c456c3a..2ea222f 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h

@@ -1246,7 +1246,7 @@ void drop_slab(void);
 extern int randomize_va_space;
 #endif
 
-__attribute__((weak)) const char *arch_vma_name(struct vm_area_struct *vma);
+//__attribute__((weak)) const char *arch_vma_name(struct vm_area_struct *vma);
 
 #endif /* __KERNEL__ */
 #endif /* _LINUX_MM_H */

So seems that something goes a bit fishy when using weak symbols and this trigges 
a binutils bug.

The above line was introdused in the following commit:

commit f269fdd1829acc5e53bf57b145003e5733133f2b
Author: David Howells [off-list ref]
Date:   Wed Sep 27 01:50:23 2006 -0700

    [PATCH] NOMMU: move the fallback arch_vma_name() to a sensible place

Thanks for looking into this.  Removing the "__attribute__((weak))"
from arch_vma_name's declaration in linux/mm.h unbreaks the build for
me.

Maybe it shouldn't matter, but it seems unusual to have the weak
attribute specified at the function's declaration.  I wasn't able to
find any examples of that for other weak functions in the kernel
(e.g. sched_clock).

Unfortunately removing the weak attribute uncovered that x86_64 has
two functions with the same name => link error.
Needs to have that sorted first out but ball is rolling.
(Se lkml for details)

	Sam