| Message ID | 20250305-fixed-type-genmasks-v4-3-1873dcdf6723@wanadoo.fr (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | bits: Fixed-type GENMASK()/BIT() | expand |
On 05/03/2025 at 23:30, Andy Shevchenko wrote: > On Wed, Mar 05, 2025 at 10:00:15PM +0900, Vincent Mailhol via B4 Relay wrote: >> From: Yury Norov <yury.norov@gmail.com> >> >> Add __GENMASK_t() which generalizes __GENMASK() to support different > > Is it with double underscore? I do not see it. This is my mistake. In an earlier draft, it was __GENMASK_t(), meanwhile, I dropped the __ prefix but forget to update the patch description. > _t is used for typedef simple types. It's unfortunate to have it > in such a macro. Ack. > Perhaps T or TYPE will suffice. Or perhaps we want > __GENMASK_Uxx() here? If no objection, I have a preference for GENMASK_TYPE(). >> types, and implement fixed-types versions of GENMASK() based on it. >> The fixed-type version allows more strict checks to the min/max values >> accepted, which is useful for defining registers like implemented by >> i915 and xe drivers with their REG_GENMASK*() macros. >> >> The strict checks rely on shift-count-overflow compiler check to fail >> the build if a number outside of the range allowed is passed. >> Example: >> >> #define FOO_MASK GENMASK_U32(33, 4) >> >> will generate a warning like: >> >> ../include/linux/bits.h:41:31: error: left shift count >= width of type [-Werror=shift-count-overflow] >> 41 | (((t)~0ULL - ((t)(1) << (l)) + 1) & \ >> | ^~ > > ... > >> + * __GENMASK_U*() depends on BITS_PER_TYPE() which would not work in the asm > > Where are the double underscore variants? I see it only for U128. Same as above. The description is incorrect. I will fix this in v5. >> + * code as BITS_PER_TYPE() relies on sizeof(), something not available in >> + * asm. Nethertheless, the concept of fixed width integers is a C thing which >> + * does not apply to assembly code. Yours sincerely, Vincent Mailhol
On Wed, Mar 05, 2025 at 10:00:15PM +0900, Vincent Mailhol via B4 Relay wrote: > From: Yury Norov <yury.norov@gmail.com> > > Add __GENMASK_t() which generalizes __GENMASK() to support different > types, and implement fixed-types versions of GENMASK() based on it. > The fixed-type version allows more strict checks to the min/max values > accepted, which is useful for defining registers like implemented by > i915 and xe drivers with their REG_GENMASK*() macros. > > The strict checks rely on shift-count-overflow compiler check to fail > the build if a number outside of the range allowed is passed. > Example: > > #define FOO_MASK GENMASK_U32(33, 4) > > will generate a warning like: > > ../include/linux/bits.h:41:31: error: left shift count >= width of type [-Werror=shift-count-overflow] > 41 | (((t)~0ULL - ((t)(1) << (l)) + 1) & \ > | ^~ > > Signed-off-by: Yury Norov <yury.norov@gmail.com> > Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com> > Acked-by: Jani Nikula <jani.nikula@intel.com> > Signed-off-by: Vincent Mailhol <mailhol.vincent@wanadoo.fr> Co-developed-by? > --- > Changelog: > > v3 -> v4: > > - The v3 is one year old. Meanwhile people started using > __GENMASK() directly. So instead of generalizing __GENMASK() to > support different types, add a new GENMASK_t(). > > - replace ~0ULL by ~_ULL(0). Otherwise, __GENMASK_t() would fail > in asm code. > > - Make GENMASK_U8() and GENMASK_U16() return an unsigned int. In > v3, due to the integer promotion rules, these were returning a > signed integer. By casting these to unsigned int, at least the This comment will disappear when I'll apply the patch. Can you comment it in the code instead? > signedness is kept. > --- > include/linux/bitops.h | 1 - > include/linux/bits.h | 33 +++++++++++++++++++++++++++++---- > 2 files changed, 29 insertions(+), 5 deletions(-) > > diff --git a/include/linux/bitops.h b/include/linux/bitops.h > index c1cb53cf2f0f8662ed3e324578f74330e63f935d..9be2d50da09a417966b3d11c84092bb2f4cd0bef 100644 > --- a/include/linux/bitops.h > +++ b/include/linux/bitops.h > @@ -8,7 +8,6 @@ > > #include <uapi/linux/kernel.h> > > -#define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE) > #define BITS_TO_LONGS(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(long)) > #define BITS_TO_U64(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(u64)) > #define BITS_TO_U32(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(u32)) > diff --git a/include/linux/bits.h b/include/linux/bits.h > index 5f68980a1b98d771426872c74d7b5c0f79e5e802..f202e46d2f4b7899c16d975120f3fa3ae41556ae 100644 > --- a/include/linux/bits.h > +++ b/include/linux/bits.h > @@ -12,6 +12,7 @@ > #define BIT_ULL_MASK(nr) (ULL(1) << ((nr) % BITS_PER_LONG_LONG)) > #define BIT_ULL_WORD(nr) ((nr) / BITS_PER_LONG_LONG) > #define BITS_PER_BYTE 8 > +#define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE) > > /* > * Create a contiguous bitmask starting at bit position @l and ending at > @@ -25,14 +26,38 @@ > > #define GENMASK_INPUT_CHECK(h, l) BUILD_BUG_ON_ZERO(const_true((l) > (h))) > > -#define GENMASK(h, l) \ > - (GENMASK_INPUT_CHECK(h, l) + __GENMASK(h, l)) > -#define GENMASK_ULL(h, l) \ > - (GENMASK_INPUT_CHECK(h, l) + __GENMASK_ULL(h, l)) > +/* > + * Generate a mask for the specified type @t. Additional checks are made to > + * guarantee the value returned fits in that type, relying on > + * shift-count-overflow compiler check to detect incompatible arguments. > + * For example, all these create build errors or warnings: > + * > + * - GENMASK(15, 20): wrong argument order > + * - GENMASK(72, 15): doesn't fit unsigned long > + * - GENMASK_U32(33, 15): doesn't fit in a u32 > + */ > +#define GENMASK_t(t, h, l) \ Agree with Andy. This should be GENMASK_TYPE, or triple-underscored ___GENMASK() maybe. This _t thing looks misleading. > + (GENMASK_INPUT_CHECK(h, l) + \ > + (((t)~ULL(0) - ((t)1 << (l)) + 1) & \ > + ((t)~ULL(0) >> (BITS_PER_TYPE(t) - 1 - (h))))) Can you rebase it on top of -next? In this dev cycle I merge a patch that reverts the __GENMASK() back to: #define __GENMASK(h, l) (((~_UL(0)) << (l)) & (~_UL(0) >> (BITS_PER_LONG - 1 - (h)))) > +#define GENMASK(h, l) GENMASK_t(unsigned long, h, l) > +#define GENMASK_ULL(h, l) GENMASK_t(unsigned long long, h, l) This makes __GENMASK() and __GENMASK_ULL() unused in the kernel, other than in uapi. Or I misunderstand it? Having, in fact, different implementations of the same macro for kernel and userspace is a source of problems. Can we move GENMASK_TYPE() to uapi, and implement __GENMASK() on top of them? If not, I'd prefer to keep GENMASK and GENMASK_ULL untouched. Can you run bloat-o-meter and ensure there's no unwanted effects on code generation? > /* > * Missing asm support > * > + * __GENMASK_U*() depends on BITS_PER_TYPE() which would not work in the asm And there's no __GENMASK_U*(), right? > + * code as BITS_PER_TYPE() relies on sizeof(), something not available in > + * asm. Nethertheless, the concept of fixed width integers is a C thing which > + * does not apply to assembly code. > + */ > +#define GENMASK_U8(h, l) ((unsigned int)GENMASK_t(u8, h, l)) > +#define GENMASK_U16(h, l) ((unsigned int)GENMASK_t(u16, h, l)) Typecast to the type that user provides explicitly? And maybe do in GENMASK_TYPE() > +#define GENMASK_U32(h, l) GENMASK_t(u32, h, l) > +#define GENMASK_U64(h, l) GENMASK_t(u64, h, l) OK, this looks good. But GENMASK_U128() becomes a special case now. The 128-bit GENMASK is unsued, but it's exported in uapi. Is there any simple way to end up with a common implementation for all fixed-type GENMASKs? > + > +/* > * __GENMASK_U128() depends on _BIT128() which would not work > * in the asm code, as it shifts an 'unsigned __int128' data > * type instead of direct representation of 128 bit constants This comment is duplicated by the previous one. Maybe just join them? (Let's wait for a while for updates regarding GENMASK_U128 status before doing it.)
On 06/03/2025 at 00:47, Yury Norov wrote: > On Wed, Mar 05, 2025 at 10:00:15PM +0900, Vincent Mailhol via B4 Relay wrote: >> From: Yury Norov <yury.norov@gmail.com> >> >> Add __GENMASK_t() which generalizes __GENMASK() to support different >> types, and implement fixed-types versions of GENMASK() based on it. >> The fixed-type version allows more strict checks to the min/max values >> accepted, which is useful for defining registers like implemented by >> i915 and xe drivers with their REG_GENMASK*() macros. >> >> The strict checks rely on shift-count-overflow compiler check to fail >> the build if a number outside of the range allowed is passed. >> Example: >> >> #define FOO_MASK GENMASK_U32(33, 4) >> >> will generate a warning like: >> >> ../include/linux/bits.h:41:31: error: left shift count >= width of type [-Werror=shift-count-overflow] >> 41 | (((t)~0ULL - ((t)(1) << (l)) + 1) & \ >> | ^~ >> >> Signed-off-by: Yury Norov <yury.norov@gmail.com> >> Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com> >> Acked-by: Jani Nikula <jani.nikula@intel.com> >> Signed-off-by: Vincent Mailhol <mailhol.vincent@wanadoo.fr> > > Co-developed-by? OK. I will keep you as the main author and credit me as Co-developer. >> --- >> Changelog: >> >> v3 -> v4: >> >> - The v3 is one year old. Meanwhile people started using >> __GENMASK() directly. So instead of generalizing __GENMASK() to >> support different types, add a new GENMASK_t(). >> >> - replace ~0ULL by ~_ULL(0). Otherwise, __GENMASK_t() would fail >> in asm code. >> >> - Make GENMASK_U8() and GENMASK_U16() return an unsigned int. In >> v3, due to the integer promotion rules, these were returning a >> signed integer. By casting these to unsigned int, at least the > > This comment will disappear when I'll apply the patch. Can you comment > it in the code instead? Ack. I will add below comment in the code: /* * Because of the C integer promotion rules, the U8 and the U16 * variants would immediately become signed integers when used in * expressions. Cast them to unsigned int so that, at least, the * signedness is preserved. */ (unless if you prefer to go back to the u8 and u16 casts, c.f. below). >> signedness is kept. >> --- >> include/linux/bitops.h | 1 - >> include/linux/bits.h | 33 +++++++++++++++++++++++++++++---- >> 2 files changed, 29 insertions(+), 5 deletions(-) >> >> diff --git a/include/linux/bitops.h b/include/linux/bitops.h >> index c1cb53cf2f0f8662ed3e324578f74330e63f935d..9be2d50da09a417966b3d11c84092bb2f4cd0bef 100644 >> --- a/include/linux/bitops.h >> +++ b/include/linux/bitops.h >> @@ -8,7 +8,6 @@ >> >> #include <uapi/linux/kernel.h> >> >> -#define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE) >> #define BITS_TO_LONGS(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(long)) >> #define BITS_TO_U64(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(u64)) >> #define BITS_TO_U32(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(u32)) >> diff --git a/include/linux/bits.h b/include/linux/bits.h >> index 5f68980a1b98d771426872c74d7b5c0f79e5e802..f202e46d2f4b7899c16d975120f3fa3ae41556ae 100644 >> --- a/include/linux/bits.h >> +++ b/include/linux/bits.h >> @@ -12,6 +12,7 @@ >> #define BIT_ULL_MASK(nr) (ULL(1) << ((nr) % BITS_PER_LONG_LONG)) >> #define BIT_ULL_WORD(nr) ((nr) / BITS_PER_LONG_LONG) >> #define BITS_PER_BYTE 8 >> +#define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE) >> >> /* >> * Create a contiguous bitmask starting at bit position @l and ending at >> @@ -25,14 +26,38 @@ >> >> #define GENMASK_INPUT_CHECK(h, l) BUILD_BUG_ON_ZERO(const_true((l) > (h))) >> >> -#define GENMASK(h, l) \ >> - (GENMASK_INPUT_CHECK(h, l) + __GENMASK(h, l)) >> -#define GENMASK_ULL(h, l) \ >> - (GENMASK_INPUT_CHECK(h, l) + __GENMASK_ULL(h, l)) >> +/* >> + * Generate a mask for the specified type @t. Additional checks are made to >> + * guarantee the value returned fits in that type, relying on >> + * shift-count-overflow compiler check to detect incompatible arguments. >> + * For example, all these create build errors or warnings: >> + * >> + * - GENMASK(15, 20): wrong argument order >> + * - GENMASK(72, 15): doesn't fit unsigned long >> + * - GENMASK_U32(33, 15): doesn't fit in a u32 >> + */ >> +#define GENMASK_t(t, h, l) \ > > Agree with Andy. This should be GENMASK_TYPE, or triple-underscored > ___GENMASK() maybe. This _t thing looks misleading. My preference goes to GENMASK_TYPE(). >> + (GENMASK_INPUT_CHECK(h, l) + \ >> + (((t)~ULL(0) - ((t)1 << (l)) + 1) & \ >> + ((t)~ULL(0) >> (BITS_PER_TYPE(t) - 1 - (h))))) > > Can you rebase it on top of -next? In this dev cycle I merge a patch > that reverts the __GENMASK() back to: Oh, I did not realize that. Do you mean a rebase on top of: https://github.com/norov/linux/tree/bitmap-for-next ? I will do so. > #define __GENMASK(h, l) (((~_UL(0)) << (l)) & (~_UL(0) >> (BITS_PER_LONG - 1 - (h)))) > >> +#define GENMASK(h, l) GENMASK_t(unsigned long, h, l) >> +#define GENMASK_ULL(h, l) GENMASK_t(unsigned long long, h, l) > > This makes __GENMASK() and __GENMASK_ULL() unused in the kernel, other > than in uapi. Or I misunderstand it? Correct. > Having, in fact, different implementations of the same macro for kernel > and userspace is a source of problems. Can we move GENMASK_TYPE() to uapi, > and implement __GENMASK() on top of them? If not, I'd prefer to keep > GENMASK and GENMASK_ULL untouched. This is something which I tried to explain in the cover letter. I am not confident to declare GENMASK_TYPE() in the uapi and expose it to the userland. If we do so, any future change in the parameters would be a user breaking change. __GENMASK_U128() looks already too much to me for the uapi, I am not keen to bloat it even more with GENMASK_TYPE(). This plus the fact that if we use GENMASK_TYPE() to generate the asm variant, then we can not rely on sizeof() in the definition which makes everything over complicated. I acknowledge that not having a common denominator is not best, but I see this as an acceptable tradeoff. > Can you run bloat-o-meter and ensure there's no unwanted effects on > code generation? Ack, but that will be tomorrow :) >> /* >> * Missing asm support >> * >> + * __GENMASK_U*() depends on BITS_PER_TYPE() which would not work in the asm > > And there's no __GENMASK_U*(), right? Yes, silly typo, sorry. Will fix in v5. >> + * code as BITS_PER_TYPE() relies on sizeof(), something not available in >> + * asm. Nethertheless, the concept of fixed width integers is a C thing which >> + * does not apply to assembly code. >> + */ >> +#define GENMASK_U8(h, l) ((unsigned int)GENMASK_t(u8, h, l)) >> +#define GENMASK_U16(h, l) ((unsigned int)GENMASK_t(u16, h, l)) > > Typecast to the type that user provides explicitly? And maybe do > in GENMASK_TYPE() I have a slight preference for the cast to unsigned int for the reason explained above. But that is not a deal breaker. If you believe that the u8/u16 casts are better, let me know, I will be happy to change it :) >> +#define GENMASK_U32(h, l) GENMASK_t(u32, h, l) >> +#define GENMASK_U64(h, l) GENMASK_t(u64, h, l) > > OK, this looks good. But GENMASK_U128() becomes a special case now. > The 128-bit GENMASK is unsued, but it's exported in uapi. Is there any > simple way to end up with a common implementation for all fixed-type > GENMASKs? What bothers me is that the 128 bit types are not something available on all architectures, c.f. the CONFIG_ARCH_SUPPORTS_INT128. So, I would need a U128() equivalent to the ULL() but which does not break on architectures which do not support 128 bits integers. This is where I am stuck. If someone can guide me on how to write a robust U128() macro, then I think the common implementation could be feasible. >> + >> +/* >> * __GENMASK_U128() depends on _BIT128() which would not work >> * in the asm code, as it shifts an 'unsigned __int128' data >> * type instead of direct representation of 128 bit constants > > This comment is duplicated by the previous one. Maybe just join them? > (Let's wait for a while for updates regarding GENMASK_U128 status before > doing it.) OK. I will wait for this one. I will probably send the v5 before we get the answer but I do not this this is an issue if we have two parallel streams. Yours sincerely, Vincent Mailhol
On 06/03/2025 at 04:45, Andy Shevchenko wrote: >>> But GENMASK_U128() becomes a special case now. >>> The 128-bit GENMASK is unsued, but it's exported in uapi. Is there any >>> simple way to end up with a common implementation for all fixed-type >>> GENMASKs? >> >> What bothers me is that the 128 bit types are not something available on >> all architectures, c.f. the CONFIG_ARCH_SUPPORTS_INT128. So, I would >> need a U128() equivalent to the ULL() but which does not break on >> architectures which do not support 128 bits integers. >> >> This is where I am stuck. If someone can guide me on how to write a >> robust U128() macro, then I think the common implementation could be >> feasible. > > I think we may leave that U128 stuff alone for now. I found the solution! The trick is to use type_max() from overflow.h. With this, GENMASK_TYPE() becomes: #define GENMASK_TYPE(t, h, l) \ ((t)(GENMASK_INPUT_CHECK(h, l) + \ (type_max(t) << (l) & \ type_max(t) >> (BITS_PER_TYPE(t) - 1 - (h))))) and works with all the GENMASK variants, including the U128 one! The unit tests under lib/test_bits.c are all green. Of course, this does *not* work in assembly. But as explained before, GENMASK_TYPE() is guarded by a #if !defined(__ASSEMBLY__), so all good! The question raised by Yury on whether or not we should keep __GENMASK_U128() in the uapi still remains. And in full honesty, I will not touch that one. This is not in the scope of this series. Yours sincerely, Vincent Mailhol
diff --git a/include/linux/bitops.h b/include/linux/bitops.h index c1cb53cf2f0f8662ed3e324578f74330e63f935d..9be2d50da09a417966b3d11c84092bb2f4cd0bef 100644 --- a/include/linux/bitops.h +++ b/include/linux/bitops.h @@ -8,7 +8,6 @@ #include <uapi/linux/kernel.h> -#define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE) #define BITS_TO_LONGS(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(long)) #define BITS_TO_U64(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(u64)) #define BITS_TO_U32(nr) __KERNEL_DIV_ROUND_UP(nr, BITS_PER_TYPE(u32)) diff --git a/include/linux/bits.h b/include/linux/bits.h index 5f68980a1b98d771426872c74d7b5c0f79e5e802..f202e46d2f4b7899c16d975120f3fa3ae41556ae 100644 --- a/include/linux/bits.h +++ b/include/linux/bits.h @@ -12,6 +12,7 @@ #define BIT_ULL_MASK(nr) (ULL(1) << ((nr) % BITS_PER_LONG_LONG)) #define BIT_ULL_WORD(nr) ((nr) / BITS_PER_LONG_LONG) #define BITS_PER_BYTE 8 +#define BITS_PER_TYPE(type) (sizeof(type) * BITS_PER_BYTE) /* * Create a contiguous bitmask starting at bit position @l and ending at @@ -25,14 +26,38 @@ #define GENMASK_INPUT_CHECK(h, l) BUILD_BUG_ON_ZERO(const_true((l) > (h))) -#define GENMASK(h, l) \ - (GENMASK_INPUT_CHECK(h, l) + __GENMASK(h, l)) -#define GENMASK_ULL(h, l) \ - (GENMASK_INPUT_CHECK(h, l) + __GENMASK_ULL(h, l)) +/* + * Generate a mask for the specified type @t. Additional checks are made to + * guarantee the value returned fits in that type, relying on + * shift-count-overflow compiler check to detect incompatible arguments. + * For example, all these create build errors or warnings: + * + * - GENMASK(15, 20): wrong argument order + * - GENMASK(72, 15): doesn't fit unsigned long + * - GENMASK_U32(33, 15): doesn't fit in a u32 + */ +#define GENMASK_t(t, h, l) \ + (GENMASK_INPUT_CHECK(h, l) + \ + (((t)~ULL(0) - ((t)1 << (l)) + 1) & \ + ((t)~ULL(0) >> (BITS_PER_TYPE(t) - 1 - (h))))) + +#define GENMASK(h, l) GENMASK_t(unsigned long, h, l) +#define GENMASK_ULL(h, l) GENMASK_t(unsigned long long, h, l) /* * Missing asm support * + * __GENMASK_U*() depends on BITS_PER_TYPE() which would not work in the asm + * code as BITS_PER_TYPE() relies on sizeof(), something not available in + * asm. Nethertheless, the concept of fixed width integers is a C thing which + * does not apply to assembly code. + */ +#define GENMASK_U8(h, l) ((unsigned int)GENMASK_t(u8, h, l)) +#define GENMASK_U16(h, l) ((unsigned int)GENMASK_t(u16, h, l)) +#define GENMASK_U32(h, l) GENMASK_t(u32, h, l) +#define GENMASK_U64(h, l) GENMASK_t(u64, h, l) + +/* * __GENMASK_U128() depends on _BIT128() which would not work * in the asm code, as it shifts an 'unsigned __int128' data * type instead of direct representation of 128 bit constants