Re: ERROR: modpost: "__delay" [drivers/net/mdio/mdio-cavium.ko] undefined!

[Randy Dunlap <rdunlap@infradead.org>]

On 5/31/21 8:45 AM, Geert Uytterhoeven wrote:
> Hi Randy,
> 
> On Mon, May 31, 2021 at 5:12 PM Randy Dunlap <rdunlap@infradead.org> wrote:
>> On 5/31/21 12:32 AM, Geert Uytterhoeven wrote:
>>> CC David (original author, asked by driver location change)
>>>
>>> On Mon, May 31, 2021 at 9:29 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
>>>>
>>>> On Mon, May 31, 2021 at 2:05 AM Randy Dunlap <rdunlap@infradead.org> wrote:
>>>>> On 5/29/21 4:25 PM, kernel test robot wrote:
>>>>>> First bad commit (maybe != root cause):
>>>>>>
>>>>>> tree:   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
>>>>>> head:   df8c66c4cfb91f2372d138b9b714f6df6f506966
>>>>>> commit: a9770eac511ad82390b9f4a3c1728e078c387ac7 net: mdio: Move MDIO drivers into a new subdirectory
>>>>>> date:   9 months ago
>>>>>> config: sh-allmodconfig (attached as .config)
>>>>>> compiler: sh4-linux-gcc (GCC) 9.3.0
>>>>>> reproduce (this is a W=1 build):
>>>>>>         wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
>>>>>>         chmod +x ~/bin/make.cross
>>>>>>         # https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a9770eac511ad82390b9f4a3c1728e078c387ac7
>>>>>>         git remote add linus https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
>>>>>>         git fetch --no-tags linus master
>>>>>>         git checkout a9770eac511ad82390b9f4a3c1728e078c387ac7
>>>>>>         # save the attached .config to linux build tree
>>>>>>         COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=sh
>>>>>>
>>>>>> If you fix the issue, kindly add following tag as appropriate
>>>>>> Reported-by: kernel test robot <lkp@intel.com>
>>>>>>
>>>>>> All errors (new ones prefixed by >>, old ones prefixed by <<):
>>>>>>
>>>>>>>> ERROR: modpost: "__delay" [drivers/net/mdio/mdio-cavium.ko] undefined!
>>>>>
>>>>> Just a comment here.  kernel test robot has reported this issue
>>>>> 5 times in 2021 that I know of -- and I could have missed some.
>>>>>
>>>>> I see that Geert recently (June 2020) reverted the
>>>>> EXPORT_SYMBOL(__delay) in arch/sh/lib/delay.c, with this comment:
>>>>>
>>>>>     __delay() is an internal implementation detail on several architectures.
>>>>>     Drivers should not call __delay() directly, as it has non-standardized
>>>>>     semantics, or may not even exist.
>>>>>     Hence there is no need to export __delay() to modules.
>>>>>
>>>>>     See also include/asm-generic/delay.h:
>>>>>
>>>>>         /* Undefined functions to get compile-time errors */
>>>>>         ...
>>>>>         extern void __delay(unsigned long loops);
>>>>>
>>>>> However, s/several architectures/all but one architecture: SH/.
>>>>> All architectures except for SH provide either an exported function,
>>>>> an inline function, or a macro for __delay(). Yeah, they probably
>>>>> don't all do the same delay.
>>>>
>>>> Hence it must not be used by drivers, as it might give the false assumption
>>>> of working everywhere.  While drivers/net/mdio/mdio-cavium is
>>>> platform-specific, code might be copied in a new driver, less restricted
>>>> to a specific platform.
>>
>> Geert, should all (15) of the other arch EXPORT_SYMBOL(__delay); exports
>> be removed?  (in theory?  I'm not planning to remove them.)
> 
> It depends.  If they are internal implementation details of an
> architecture's mdelay() or udelay() function (i.e. the latter are
> static inline functions that may call into out-of-line __delay()
> functions), then they should be kept.
> 
> I haven't checked all of them, but e.g. on arm64, mdelay() and udelay()
> don't call into __delay, so IMHO its export should be removed.
> 
> Generic drivers should not use __delay() with a hardcoded value, as
> its semantics are not defined (cfr. the undefined function comment
> in asm-generic).

Thanks, Geert. I get it.

Hopefully David Daney will jump in here with a patch at some point.

-- 
~Randy