Re: [linux-next:master 13398/13940] drivers/net/ethernet/sfc/ef100_nic.c:610: undefined reference to `__umoddi3'

Re: [linux-next:master 13398/13940] drivers/net/ethernet/sfc/ef100_nic.c:610: undefined reference to `__umoddi3'

[Edward Cree]

On 06/08/2020 00:48, kernel test robot wrote:
> tree:   https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git master
> head:   d15fe4ec043588beee823781602ddb51d0bc84c8
> commit: adcfc3482ffff813fa2c34e5902005853f79c2aa [13398/13940] sfc_ef100: read Design Parameters at probe time
> config: microblaze-randconfig-r032-20200805 (attached as .config)
> compiler: microblaze-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
>         git checkout adcfc3482ffff813fa2c34e5902005853f79c2aa
>         # save the attached .config to linux build tree
>         COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=microblaze 
>
> 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 >>):
>
>    microblaze-linux-ld: drivers/net/ethernet/sfc/ef100_nic.o: in function `ef100_process_design_param':
>>> drivers/net/ethernet/sfc/ef100_nic.c:610: undefined reference to `__umoddi3'
>    605			/* Our TXQ and RXQ sizes are always power-of-two and thus divisible by
>    606			 * EFX_MIN_DMAQ_SIZE, so we just need to check that
>    607			 * EFX_MIN_DMAQ_SIZE is divisible by GRANULARITY.
>    608			 * This is very unlikely to fail.
>    609			 */
>  > 610			if (EFX_MIN_DMAQ_SIZE % reader->value) {
So, this is (unsigned long) % (u64), whichI guess doesn't go quite
 as smoothly 32-bit microcontrollers (though the thought of plugging
 a 100-gig smartNIC into a microblaze boggles the mind a little ;).
And none of the math64.h functions seem to have the shape we want —
 div_u64_rem() wants to write the remainder through a pointer, and
 do_div() wants to modify the dividend (which is a constant in this
 case).  So whatever I do, it's gonna be ugly :(

Maybe I should add a

static inline u32 mod_u64(u64 dividend, u32 divisor)
{
        return do_div(dividend, divisor);
}

to include/linux/math64.h?  At least that way the ugly is centralised
 in the header file.

Re: [linux-next:master 13398/13940] drivers/net/ethernet/sfc/ef100_nic.c:610: undefined reference to `__umoddi3'

[Guenter Roeck]

On Thu, Aug 06, 2020 at 07:17:43PM +0100, Edward Cree wrote:
> On 06/08/2020 00:48, kernel test robot wrote:
> > tree:   https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git master
> > head:   d15fe4ec043588beee823781602ddb51d0bc84c8
> > commit: adcfc3482ffff813fa2c34e5902005853f79c2aa [13398/13940] sfc_ef100: read Design Parameters at probe time
> > config: microblaze-randconfig-r032-20200805 (attached as .config)
> > compiler: microblaze-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
> >         git checkout adcfc3482ffff813fa2c34e5902005853f79c2aa
> >         # save the attached .config to linux build tree
> >         COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-9.3.0 make.cross ARCH=microblaze 
> >
> > 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 >>):
> >
> >    microblaze-linux-ld: drivers/net/ethernet/sfc/ef100_nic.o: in function `ef100_process_design_param':
> >>> drivers/net/ethernet/sfc/ef100_nic.c:610: undefined reference to `__umoddi3'
> >    605			/* Our TXQ and RXQ sizes are always power-of-two and thus divisible by
> >    606			 * EFX_MIN_DMAQ_SIZE, so we just need to check that
> >    607			 * EFX_MIN_DMAQ_SIZE is divisible by GRANULARITY.
> >    608			 * This is very unlikely to fail.
> >    609			 */
> >  > 610			if (EFX_MIN_DMAQ_SIZE % reader->value) {
> So, this is (unsigned long) % (u64), whichI guess doesn't go quite
>  as smoothly 32-bit microcontrollers (though the thought of plugging
>  a 100-gig smartNIC into a microblaze boggles the mind a little ;).
> And none of the math64.h functions seem to have the shape we want —
>  div_u64_rem() wants to write the remainder through a pointer, and
>  do_div() wants to modify the dividend (which is a constant in this
>  case).  So whatever I do, it's gonna be ugly :(
> 
> Maybe I should add a
> 
> static inline u32 mod_u64(u64 dividend, u32 divisor)
> {
>         return do_div(dividend, divisor);
> }

Your proposed function is an exact replicate of do_div() and thus doesn't
make much sense. Also, in your case, divisor is a 64-bit value, which is
causing the problem to start with. You could try something like

	if (reader->value > EFX_MIN_DMAQ_SIZE || EFX_MIN_DMAQ_SIZE % (u32)reader->value)

If EFX_MIN_DMAQ_SIZE is indeed known to be a power of 2, you could also use
the knowledge that a 2^n value can only be divided by a smaller 2^n value,
meaning that reader->value must have exactly one bit set. This would also
avoid divide-by-0 issues if reader->value can be 0.

	if (reader->value > EFX_MIN_DMAQ_SIZE || hweight64(reader->value) != 1)

Guenter

Re: [linux-next:master 13398/13940] drivers/net/ethernet/sfc/ef100_nic.c:610: undefined reference to `__umoddi3'

[Edward Cree]

On 10/08/2020 16:51, Guenter Roeck wrote:
> On Thu, Aug 06, 2020 at 07:17:43PM +0100, Edward Cree wrote:
>> Maybe I should add a
>>
>> static inline u32 mod_u64(u64 dividend, u32 divisor)
>> {
>>         return do_div(dividend, divisor);
>> }
> Your proposed function is an exact replicate of do_div()
No, because do_div() is a macro that modifies 'dividend', whereas by
 wrapping it in an inline function mod_u64() implicitly creates a
 local variable.  Thus do_div() cannot be used on a constant, whereas
 mod_u64() can.
> You could try something like
>
> 	if (reader->value > EFX_MIN_DMAQ_SIZE || EFX_MIN_DMAQ_SIZE % (u32)reader->value)
I considered that.  It's ugly, so while it will work I think it's
 worthlooking to see if there's a better way.
> If EFX_MIN_DMAQ_SIZE is indeed known to be a power of 2, you could also use
> the knowledge that a 2^n value can only be divided by a smaller 2^n value,
> meaning that reader->value must have exactly one bit set. This would also
> avoid divide-by-0 issues if reader->value can be 0.
>
> 	if (reader->value > EFX_MIN_DMAQ_SIZE || hweight64(reader->value) != 1)
This is also ugly and I don't like relying on the power-of-twoness —
 it just feels fragile.

But you're right to point out that there's a div/0 issue, and if I'm
 going to have to check for that, then ugliness is unavoidable.  So
 I think the least painful option available is probably

    if (!reader->value || reader->value > EFX_MIN_DMAQ_SIZE ||
        EFX_MIN_DMAQ_SIZE % (u32)reader->value)

 which only assumes EFX_MIN_DMAQ_SIZE <= U32_MAX, an assumption I'm
 comfortable with baking in.
I'll put together a formal patch with that.

Thanks for the help.

-ed