[PATCH v2 03/18] lib/crc32: expose whether the lib is really optimized at runtime

[Eric Biggers]

On Fri, Oct 25, 2024 at 11:37:45PM +0200, Ard Biesheuvel wrote:
> On Fri, 25 Oct 2024 at 23:32, Eric Biggers <ebiggers at kernel.org> wrote:
> >
> > On Fri, Oct 25, 2024 at 10:32:14PM +0200, Ard Biesheuvel wrote:
> > > On Fri, 25 Oct 2024 at 21:15, Eric Biggers <ebiggers at kernel.org> wrote:
> > > >
> > > > From: Eric Biggers <ebiggers at google.com>
> > > >
> > > > Make the CRC32 library export some flags that indicate which CRC32
> > > > functions are actually executing optimized code at runtime.  Set these
> > > > correctly from the architectures that implement the CRC32 functions.
> > > >
> > > > This will be used to determine whether the crc32[c]-$arch shash
> > > > algorithms should be registered in the crypto API.  btrfs could also
> > > > start using these flags instead of the hack that it currently uses where
> > > > it parses the crypto_shash_driver_name.
> > > >
> > > > Signed-off-by: Eric Biggers <ebiggers at google.com>
> > > > ---
> > > >  arch/arm64/lib/crc32-glue.c  | 15 +++++++++++++++
> > > >  arch/riscv/lib/crc32-riscv.c | 15 +++++++++++++++
> > > >  include/linux/crc32.h        | 15 +++++++++++++++
> > > >  lib/crc32.c                  |  5 +++++
> > > >  4 files changed, 50 insertions(+)
> > > >
> > > ...
> > > > diff --git a/include/linux/crc32.h b/include/linux/crc32.h
> > > > index 58c632533b08..bf26d454b60d 100644
> > > > --- a/include/linux/crc32.h
> > > > +++ b/include/linux/crc32.h
> > > > @@ -35,10 +35,25 @@ static inline u32 __pure __crc32c_le(u32 crc, const u8 *p, size_t len)
> > > >         if (IS_ENABLED(CONFIG_CRC32_ARCH))
> > > >                 return crc32c_le_arch(crc, p, len);
> > > >         return crc32c_le_base(crc, p, len);
> > > >  }
> > > >
> > > > +/*
> > > > + * crc32_optimizations contains flags that indicate which CRC32 library
> > > > + * functions are using architecture-specific optimizations.  Unlike
> > > > + * IS_ENABLED(CONFIG_CRC32_ARCH) it takes into account the different CRC32
> > > > + * variants and also whether any needed CPU features are available at runtime.
> > > > + */
> > > > +#define CRC32_LE_OPTIMIZATION  BIT(0) /* crc32_le() is optimized */
> > > > +#define CRC32_BE_OPTIMIZATION  BIT(1) /* crc32_be() is optimized */
> > > > +#define CRC32C_OPTIMIZATION    BIT(2) /* __crc32c_le() is optimized */
> > > > +#if IS_ENABLED(CONFIG_CRC32_ARCH)
> > > > +extern u32 crc32_optimizations;
> > > > +#else
> > > > +#define crc32_optimizations 0
> > > > +#endif
> > > > +
> > >
> > > Wouldn't it be cleaner to add a new library function for this, instead
> > > of using a global variable?
> >
> > The architecture crc32 modules need to be able to write to this.  There could be
> > a setter function and a getter function, but just using a variable is simpler.
> >
> 
> If we just add 'u32 crc32_optimizations()', there is no need for those
> modules to have init/exit hooks, the only thing they need to export is
> this routine.
> 
> Or perhaps it could even be a static inline, with the right plumbing
> of header files. At least on arm64,
> 
> static inline u32 crc32_optimizations() {
>   if (!alternative_have_const_cap_likely(ARM64_HAS_CRC32))
>     return 0;
>   return CRC32_LE_OPTIMIZATION | CRC32_BE_OPTIMIZATION | CRC32C_OPTIMIZATION;
> }
> 
> should be all we need.

In 7 of the 9 affected arches, I already have a module_init function that checks
the CPU features in order to set up static keys.  (arm64 and riscv already used
alternatives patching, so I kept that.)  It's slightly convenient to set these
flags at the same time, but yes the above solution would work too.

- Eric