Bug 19421

> after upgrading to binutils 2.25.

this should be 2.25.90 ...

Hi Alan,

   Have you seen this bug report:

> https://sourceware.org/bugzilla/show_bug.cgi?id=19421

>             Summary: Missing weak symbols in kernel modules on ppc64el

> Quite a few symbols in kernel modules on ppc64el are dropped after upgrading to
> binutils 2.25.  This in turn completely breaks module loading on ppc64el
> systems, per this Debian bug report:
> https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=808246

It seems that the cause of this was the patch for PR 4317:

   https://sourceware.org/bugzilla/show_bug.cgi?id=4317

Perhaps we need to allow some targets to require that undefined locals 
are preserved ?

Cheers
   Nick

So far, I haven't been able to reproduce this problem using torvalds linux git kernel and current binutils.

There are lots of claims in the debian bug report:

1)
fat: no symbol version for TOC.
fat: Unknown symbol TOC. (err -22)
ext4: no symbol version for TOC.
ext4: Unknown symbol TOC. (err -22)
and others like this.

This can happen if there is no entry for "TOC." in the __versions section of a module, but I see for example
objdump -sj__versions fs/binfmt_misc.ko
...
 0e00 00000000 00000000 2e544f43 2e000000  .........TOC....
 0e10 00000000 00000000 00000000 00000000  ................
 0e20 00000000 00000000 00000000 00000000  ................
 0e30 00000000 00000000 00000000 00000000  ................
...
I picked on binfmt_misc.ko rather than the fat or ext4 modules since binfmt_misc.ko is built for "make defconfig", but I see the same with ext4.ko after flipping the appropriate .config lines.

2)
Changed section ordering for -z relro.  This shouldn't be a problem, but may be one of the triggers for (4) below.

3)
Normally a ppc64el kernel has the following undefined symbols:
| 44809: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND mach_powermac
| 52870: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND mach_chrp
| 62021: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND mach_cell
| 62383: 0000000000000000     0 NOTYPE  WEAK   DEFAULT  UND __crc_TOC.

Also not a problem if these symbols disappear when undefined, I think.  The mach_* symbols are used in the following macro
#define machine_is(name) \
        ({ \
                extern struct machdep_calls mach_##name \
                        __attribute__((weak));           \
                machine_id == &mach_##name; \
        })
A missing kernel symbol will be silently resolved to 0 if the reference is STB_WEAK as far as I can see from my reading kernel/module.c.  This of course is the same value if the symbol is found.  __crc_TOC. is similar (but I'm still investigating this one).

4)
the mcount symbol error

See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68759
If you're building with gcc-6, you will need a kernel patch.

I've reproduced lack of "TOC." in __versions.  (Lots of WARNING: ".TOC."  [*.ko] has no CRC! too.)

"make defconfig" followed by turning off CONFIG_CRASH_DUMP and CONFIG_RELOCATABLE in .config does the trick, at least with current gcc-6 and binutils.  I'll look into why tomorrow if someone else doesn't beat me to it.

Created attachment 8897 [details]
kernel fix

So the reason that CONFIG_RELOCATABLE (which is forced on by CONFIG_CRASH_DUMP) avoids the problem is that with relocatable output there is a relocation on "__crc_TOC." which results in the symbol being kept with newer binutils.  genksyms ought to be providing a value for "__crc_TOC." but it doesn't, which I think makes this a kernel bug.  Attached patch cures the problem.

The binutils-2_26-branch branch has been updated by Alan Modra <amodra@sourceware.org>:

https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=a82e3ef7ad90d9335c8f3801f06916ffed43327b

commit a82e3ef7ad90d9335c8f3801f06916ffed43327b
Author: Alan Modra <amodra@gmail.com>
Date:   Thu Jan 14 19:02:18 2016 +1030

    Revert "Strip undefined symbols from .symtab"
    
    This reverts commit d983c8c5503d680c6d4955ceb610a9beebc64460 and
    db95bb7c5946a109e1584ab2c43c052ff39e63bd
    
    The PR4317 patch introduced a ppc64le linux kernel build problem.
    
    	PR ld/19421
    	PR ld/4317
    bfd/
    	* elflink.c (elf_link_input_bfd): Revert 2015-02-19 changes.
    	(elf_link_output_extsym): Likewise.
    ld/
    	* ld-aarch64/gc-tls-relocs.d: Revert 2015-02-19 changes.
    	* ld-cris/locref2.d: Likewise.
    	* ld-elf/ehdr_start-weak.d: Likewise.
    	* ld-elf/group1.d: Likewise.
    	* ld-i386/compressed1.d: Likewise.
    	* ld-ia64/error1.d: Likewise.
    	* ld-ia64/error2.d: Likewise.
    	* ld-ia64/error3.d: Likewise.
    	* ld-mips-elf/pic-and-nonpic-1.nd: Likewise.
    	* ld-mmix/undef-3.d: Likewise.
    	* ld-powerpc/tlsexe.r: Likewise.
    	* ld-powerpc/tlsexetoc.r: Likewise.
    	* ld-powerpc/tlsso.r: Likewise.
    	* ld-powerpc/tlstocso.r: Likewise.
    	* ld-x86-64/compressed1.d: Likewise.
    	* ld-x86-64/pie1.d: Likewise.

Fixed for 2.26 by reverting the pr4317 patch.  Expect kernel patches too.

Alan, are you aware of https://git.kernel.org/cgit/linux/kernel/git/powerpc/linux.git/commit/?id=a61674bdfc7c2bf909c4010699 
and https://git.kernel.org/cgit/linux/kernel/git/powerpc/linux.git/commit/?id=2e50c4bef77511b42cc226865d ?