Re: [PATCH V3 1/2] uapi: Define GENMASK_U128

[Anshuman Khandual <anshuman.khandual@arm.com>]

On 8/1/24 12:46, Anshuman Khandual wrote:
> This adds GENMASK_U128() and __GENMASK_U128() macros using __BITS_PER_U128
> and __int128 data types. These macros will be used in providing support for
> generating 128 bit masks.
> 
> Cc: Yury Norov <yury.norov@gmail.com>
> Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk>
> Cc: Arnd Bergmann <arnd@arndb.de>>
> Cc: linux-kernel@vger.kernel.org
> Cc: linux-arch@vger.kernel.org
> Reviewed-by: Arnd Bergmann <arnd@arndb.de>
> Signed-off-by: Anshuman Khandual <anshuman.khandual@arm.com>
> ---
>  include/linux/bits.h       | 13 +++++++++++++
>  include/uapi/linux/bits.h  |  3 +++
>  include/uapi/linux/const.h | 15 +++++++++++++++
>  3 files changed, 31 insertions(+)
> 
> diff --git a/include/linux/bits.h b/include/linux/bits.h
> index 0eb24d21aac2..bf99feb5570e 100644
> --- a/include/linux/bits.h
> +++ b/include/linux/bits.h
> @@ -36,4 +36,17 @@
>  #define GENMASK_ULL(h, l) \
>  	(GENMASK_INPUT_CHECK(h, l) + __GENMASK_ULL(h, l))
>  
> +/*
> + * Missing asm support
> + *
> + * __GENMASK_U128() depends on _BIT128() which would not work
> + * in the asm code, as it shifts an 'unsigned __init128' data
> + * type instead of direct representation of 128 bit constants
> + * such as long and unsigned long. The fundamental problem is
> + * that a 128 bit constant will get silently truncated by the
> + * gcc compiler.
> + */
> +#define GENMASK_U128(h, l) \
> +	(GENMASK_INPUT_CHECK(h, l) + __GENMASK_U128(h, l))
> +
>  #endif	/* __LINUX_BITS_H */
> diff --git a/include/uapi/linux/bits.h b/include/uapi/linux/bits.h
> index 3c2a101986a3..4d4b7b08003c 100644
> --- a/include/uapi/linux/bits.h
> +++ b/include/uapi/linux/bits.h
> @@ -12,4 +12,7 @@
>          (((~_ULL(0)) - (_ULL(1) << (l)) + 1) & \
>           (~_ULL(0) >> (__BITS_PER_LONG_LONG - 1 - (h))))
>  
> +#define __GENMASK_U128(h, l) \
> +	((_BIT128((h) + 1)) - (_BIT128(l)))
> +
>  #endif /* _UAPI_LINUX_BITS_H */
> diff --git a/include/uapi/linux/const.h b/include/uapi/linux/const.h
> index a429381e7ca5..5be12e8f8f9c 100644
> --- a/include/uapi/linux/const.h
> +++ b/include/uapi/linux/const.h
> @@ -28,6 +28,21 @@
>  #define _BITUL(x)	(_UL(1) << (x))
>  #define _BITULL(x)	(_ULL(1) << (x))
>  
> +/*
> + * Missing asm support
> + *
> + * __BIT128() would not work in the asm code, as it shifts an
> + * 'unsigned __init128' data type as direct representation of
> + * 128 bit constants is not supported in the gcc compiler, as
> + * they get silently truncated.
> + *
> + * TODO: Please revisit this implementation when gcc compiler
> + * starts representing 128 bit constants directly like long
> + * and unsigned long etc. Subsequently drop the comment for
> + * GENMASK_U128() which would then start supporting asm code.
> + */
> +#define _BIT128(x)	((unsigned __int128)(1) << (x))
> +
>  #define __ALIGN_KERNEL(x, a)		__ALIGN_KERNEL_MASK(x, (__typeof__(x))(a) - 1)
>  #define __ALIGN_KERNEL_MASK(x, mask)	(((x) + (mask)) & ~(mask))
>  

Hello Yuri/Arnd,

This proposed GENMASK_U128(h, l) warns during build when the higher end
bit is 127 (which in itself is a valid input).

./include/uapi/linux/const.h:45:44: warning: left shift count >= width of type [-Wshift-count-overflow]
   45 | #define _BIT128(x) ((unsigned __int128)(1) << (x))
      |                                            ^~
./include/asm-generic/bug.h:123:25: note: in definition of macro ‘WARN_ON’
  123 |  int __ret_warn_on = !!(condition);    \
      |                         ^~~~~~~~~
./include/uapi/linux/bits.h:16:4: note: in expansion of macro ‘_BIT128’
   16 |  ((_BIT128((h) + 1)) - (_BIT128(l)))
      |    ^~~~~~~
./include/linux/bits.h:51:31: note: in expansion of macro ‘__GENMASK_U128’
   51 |  (GENMASK_INPUT_CHECK(h, l) + __GENMASK_U128(h, l))
      |                               ^~~~~~~~~~~~~~

This is caused by ((unsigned __int128)(1) << (128)) which is generated
via (h + 1) element in __GENMASK_U128().

#define _BIT128(x)	((unsigned __int128)(1) << (x))
#define __GENMASK_U128(h, l) \
	((_BIT128((h) + 1)) - (_BIT128(l)))

Adding some extra tests in lib/test_bits.c exposes this build problem,
although it does not fail these new tests.

[    1.719221]     # Subtest: bits-test
[    1.719291]     # module: test_bits
[    1.720522]     ok 1 genmask_test
[    1.721570]     ok 2 genmask_ull_test
[    1.722668]     ok 3 genmask_u128_test
[    1.723760]     ok 4 genmask_input_check_test
[    1.723909] # bits-test: pass:4 fail:0 skip:0 total:4
[    1.724101] ok 1 bits-test

diff --git a/lib/test_bits.c b/lib/test_bits.c
index d3d858b24e02..7a972edc7122 100644
--- a/lib/test_bits.c
+++ b/lib/test_bits.c
@@ -49,6 +49,8 @@ static void genmask_u128_test(struct kunit *test)
        KUNIT_EXPECT_EQ(test, 0xffffffffffffffffULL, GENMASK_U128(63, 0));
        KUNIT_EXPECT_EQ(test, 0xffffffffffffffffULL, GENMASK_U128(64, 0) >> 1);
        KUNIT_EXPECT_EQ(test, 0x00000000ffffffffULL, GENMASK_U128(81, 50) >> 50);
+       KUNIT_EXPECT_EQ(test, 0x0000000000000003ULL, GENMASK_U128(127, 126) >> 126);
+       KUNIT_EXPECT_EQ(test, 0x0000000000000001ULL, GENMASK_U128(127, 127) >> 127);

The most significant bit in the generate mask can be added separately
, thus voiding that extra shift. The following patch solves the build
problem.

diff --git a/include/uapi/linux/bits.h b/include/uapi/linux/bits.h
index 4d4b7b08003c..4e50f635c6d9 100644
--- a/include/uapi/linux/bits.h
+++ b/include/uapi/linux/bits.h
@@ -13,6 +13,6 @@
          (~_ULL(0) >> (__BITS_PER_LONG_LONG - 1 - (h))))
 
 #define __GENMASK_U128(h, l) \
-       ((_BIT128((h) + 1)) - (_BIT128(l)))
+       (((_BIT128(h)) - (_BIT128(l))) | (_BIT128(h)))