Re: [PATCH] x86_32: Fix module version table mismatch.

["H. Peter Anvin" <hpa@zytor.com>]

On 04/23/2013 05:40 AM, Tetsuo Handa wrote:
> Commit a4b6a77b "module: fix symbol versioning with symbol prefixes" broke
> loading of net/ipv6/ipv6.ko built with CONFIG_MODVERSIONS=y for x86_32.
> 
>   # modprobe ipv6
>   FATAL: Error inserting ipv6 (/lib/modules/3.9.0-rc8-next-20130422/kernel/net/ipv6/ipv6.ko): Invalid argument
>   # dmesg
>   ipv6: no symbol version for memcmp
>   ipv6: Unknown symbol memcmp (err -22)
> 
> The reason for breakage is that check_version() in kernel/module.c tries to
> find symname == "memcmp" but versions[i].name == "__builtin_memcmp".
> 
> The reason for versions[i].name == "__builtin_memcmp" is that
> memcmp() for x86_32 is defined as
> 
>   #define memcmp __builtin_memcmp
> 
> in arch/x86/include/asm/string_32.h while memcmp() for x86_64 is defined as
> 
>   int memcmp(const void *cs, const void *ct, size_t count);
> 
> in arch/x86/include/asm/string_64.h.
> 
> Since __builtin_memcmp is a gcc's built-in function which might emit a call to
> memcmp, __builtin_memcmp should not be used for versions[i].name field.
> 
> In order to make sure that versions[i].name == "memcmp", make the definition of
> memcmp() for x86_32 identical with that of x86_64.
> 

I'm still confused by all of this.  VMLINUX_SYMBOL_STR() ought to be a
noop on x86, so how on Earth could a4b6a77b break anything?

Secondly, although memcmp is a macro, it is #undef'd before the
definition in lib/string.c, which is the one that is exported.

I'm wondering if the real culprit isn't b92021b0, but I'm looking into
it now.

	-hpa