Re: [PATCH v3 3/3] linux/bits.h: Add compile time sanity check of GENMASK inputs

[Rikard Falkeborn]

On Tue, Oct 08, 2019 at 04:52:17PM +0900, Masahiro Yamada wrote:
> On Tue, Oct 8, 2019 at 4:44 PM Masahiro Yamada
> <yamada.masahiro@xxxxxxxxxxxxx> wrote:
> >
> > Hi Geert,
> >
> > On Tue, Oct 8, 2019 at 4:23 PM Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> > >
> > > Hi Rikard,
> > >
> > > On Sun, Aug 11, 2019 at 8:52 PM Rikard Falkeborn
> > > <rikard.falkeborn@xxxxxxxxx> wrote:
> > > > GENMASK() and GENMASK_ULL() are supposed to be called with the high bit
> > > > as the first argument and the low bit as the second argument. Mixing
> > > > them will return a mask with zero bits set.
> > > >
> > > > Recent commits show getting this wrong is not uncommon, see e.g.
> > > > commit aa4c0c9091b0 ("net: stmmac: Fix misuses of GENMASK macro") and
> > > > commit 9bdd7bb3a844 ("clocksource/drivers/npcm: Fix misuse of GENMASK
> > > > macro").
> > > >
> > > > To prevent such mistakes from appearing again, add compile time sanity
> > > > checking to the arguments of GENMASK() and GENMASK_ULL(). If both
> > > > arguments are known at compile time, and the low bit is higher than the
> > > > high bit, break the build to detect the mistake immediately.
> > > >
> > > > Since GENMASK() is used in declarations, BUILD_BUG_ON_ZERO() must be
> > > > used instead of BUILD_BUG_ON().
> > > >
> > > > __builtin_constant_p does not evaluate is argument, it only checks if it
> > > > is a constant or not at compile time, and __builtin_choose_expr does not
> > > > evaluate the expression that is not chosen. Therefore, GENMASK(x++, 0)
> > > > does only evaluate x++ once.
> > > >
> > > > Commit 95b980d62d52 ("linux/bits.h: make BIT(), GENMASK(), and friends
> > > > available in assembly") made the macros in linux/bits.h available in
> > > > assembly. Since BUILD_BUG_OR_ZERO() is not asm compatible, disable the
> > > > checks if the file is included in an asm file.
> > > >
> > > > Signed-off-by: Rikard Falkeborn <rikard.falkeborn@xxxxxxxxx>
> > > > ---
> > > > Changes in v3:
> > > >   - Changed back to shorter macro argument names
> > > >   - Remove casts and use 0 instead of UL(0) in GENMASK_INPUT_CHECK(),
> > > >     since all results in GENMASK_INPUT_CHECK() are now ints. Update
> > > >     commit message to reflect that.
> > > >
> > > > Changes in v2:
> > > >   - Add comment about why inputs are not checked when used in asm file
> > > >   - Use UL(0) instead of 0
> > > >   - Extract mask creation in a separate macro to improve readability
> > > >   - Use high and low instead of h and l (part of this was extracted to a
> > > >     separate patch)
> > > >   - Updated commit message
> > > >
> > > > Joe Perches sent a series to fix the existing misuses of GENMASK() that
> > > > needs to be merged before this to avoid build failures. Currently, 5 of
> > > > the patches are not in Linus tree, and 2 are not in linux-next. There is
> > > > also a patch pending by Nathan Chancellor that also needs to be merged
> > > > before this patch is merged to avoid build failures.
> > > >
> > > >  include/linux/bits.h | 21 +++++++++++++++++++--
> > > >  1 file changed, 19 insertions(+), 2 deletions(-)
> > > >
> > > > diff --git a/include/linux/bits.h b/include/linux/bits.h
> > > > index 669d69441a62..4ba0fb609239 100644
> > > > --- a/include/linux/bits.h
> > > > +++ b/include/linux/bits.h
> > > > @@ -18,12 +18,29 @@
> > > >   * position @h. For example
> > > >   * GENMASK_ULL(39, 21) gives us the 64bit vector 0x000000ffffe00000.
> > > >   */
> > > > -#define GENMASK(h, l) \
> > > > +#ifndef __ASSEMBLY__
> > > > +#include <linux/build_bug.h>
> > > > +#define GENMASK_INPUT_CHECK(h, l) \
> > > > +       (BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
> > > > +               __builtin_constant_p((l) > (h)), (l) > (h), 0)))
> > > > +#else
> > > > +/*
> > > > + * BUILD_BUG_ON_ZERO is not available in h files included from asm files,
> > > > + * disable the input check if that is the case.
> > > > + */
> > > > +#define GENMASK_INPUT_CHECK(h, l) 0
> > > > +#endif
> > > > +
> > > > +#define __GENMASK(h, l) \
> > > >         (((~UL(0)) - (UL(1) << (l)) + 1) & \
> > > >          (~UL(0) >> (BITS_PER_LONG - 1 - (h))))
> > > > +#define GENMASK(h, l) \
> > > > +       (GENMASK_INPUT_CHECK(h, l) + __GENMASK(h, l))
> > > >
> > > > -#define GENMASK_ULL(h, l) \
> > > > +#define __GENMASK_ULL(h, l) \
> > > >         (((~ULL(0)) - (ULL(1) << (l)) + 1) & \
> > > >          (~ULL(0) >> (BITS_PER_LONG_LONG - 1 - (h))))
> > > > +#define GENMASK_ULL(h, l) \
> > > > +       (GENMASK_INPUT_CHECK(h, l) + __GENMASK_ULL(h, l))
> > > >
> > > >  #endif /* __LINUX_BITS_H */
> > >
> > > This is now commit 0fd35cd30a2fece1 ("linux/bits.h: add compile time sanity
> > > check of GENMASK inputs") in next-20191008.
> > >
> > > <noreply@xxxxxxxxxxxxxx> reported the following failure in sun3_defconfig,
> > > which I managed to reproduce with gcc-4.6.3:
> >
> > Oh dear.
> >
> > I was able to reproduce this for gcc 4.7 or 4.8,
> > but I did not see any problem for gcc 4.9+
> >
> > Perhaps, is this due to broken __builtin_choose_expr or __builtin_constant_p
> > for old compilers?
> >
> >
> >
> >
> >
> >
> >
> > >
> > >     lib/842/842_compress.c: In function '__split_add_bits':
> > >     lib/842/842_compress.c:164:25: error: first argument to
> > > '__builtin_choose_expr' not a constant
> > >     lib/842/842_compress.c:164:25: error: bit-field '<anonymous>'
> > > width not an integer constant
> > >     scripts/Makefile.build:265: recipe for target
> > > 'lib/842/842_compress.o' failed
> > >
> > > __split_add_bits() calls GENMASK_ULL() with a non-constant.
> > > However __split_add_bits() itself is called with constants only.
> > > Apparently gcc fails to inline __split_add_bits().
> > > Adding inline or always_inline doesn't help.
> > >
> 
> 
> If this is broken for GCC < 4.9,
> we might be able to workaround it as follows:
> 
> 
> 
> 
> diff --git a/include/linux/bits.h b/include/linux/bits.h
> index 4ba0fb609239..f00417baf545 100644
> --- a/include/linux/bits.h
> +++ b/include/linux/bits.h
> @@ -18,7 +18,7 @@
>   * position @h. For example
>   * GENMASK_ULL(39, 21) gives us the 64bit vector 0x000000ffffe00000.
>   */
> -#ifndef __ASSEMBLY__
> +#if !defined(__ASSEMBLY__) && (!defined(CONFIG_CC_IS_GCC) ||
> CONFIG_GCC_VERSION >= 49000)
>  #include <linux/build_bug.h>
>  #define GENMASK_INPUT_CHECK(h, l) \
>         (BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
> 
> 
> -- 
> Best Regards
> Masahiro Yamada

Hi Masahiro, Geert,

It seems it is broken for GCC < 4.9, see [1]. I'll try disabling it for
too old compilers and resend.

In the meantime, I guess it should be dropped from linux-next?

Rikard

[1]: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=19449