Hi Jakub,
El Wed, Oct 04, 2017 at 03:22:03PM -0700 Jakub Kicinski ha dit:
> On Wed, 4 Oct 2017 11:49:57 -0700, Matthias Kaehlcke wrote:
> > Hi Joe,
> >
> > El Wed, Oct 04, 2017 at 11:07:19AM -0700 Joe Perches ha dit:
> >
> > > On Tue, 2017-10-03 at 13:05 -0700, Matthias Kaehlcke wrote:
> > > > nfp_eth_set_bit_config() is marked as __always_inline to allow gcc to
> > > > identify the 'mask' parameter as known to be constant at compile time,
> > > > which is required to use the FIELD_GET() macro.
> > > >
> > > > The forced inlining does the trick for gcc, but for kernel builds with
> > > > clang it results in undefined symbols:
> > >
> > > Can't you use local different FIELD_PREP/FIELD_GET macros
> > > with a different name without the BUILD_BUG tests?
> > >
> > > i.e.:
> > >
> > > #define NFP_FIELD_PREP(_mask, _val) \
> > > ({ \
> > > ((typeof(_mask))(_val) << __bf_shf(_mask)) & (_mask); \
> > > })
> > >
> > > #define NFP_FIELD_GET(_mask, _reg) \
> > > ({ \
> > > (typeof(_mask))(((_reg) & (_mask)) >> __bf_shf(_mask)); \
> > > })
> > >
> > > Then the __always_inline can be removed from
> > > nfp_eth_set_bit_config too.
> >
> > Thanks for the suggestion. This seems a viable alternative if David
> > and the NFP owners can live without the extra checking provided by
> > __BF_FIELD_CHECK.
>
> The reason the __BF_FIELD_CHECK refuses to compile non-constant masks
> is that it will require runtime ffs on the mask, which is potentially
> costly. I would also feel quite stupid adding those macros to the nfp
> driver, given that I specifically created the bitfield.h header to not
> have to reimplement these in every driver I write/maintain.
That make sense, thanks for providing more context.
> Can you please test the patch I provided in the other reply?
With this patch there are no errors when building the kernel with
clang.
Thanks!
Matthias
