On Mon, 21 Oct 2024 at 02:29, Eric Biggers <ebiggers@xxxxxxxxxx> wrote: > > This patchset is also available in git via: > > git fetch https://git.kernel.org/pub/scm/linux/kernel/git/ebiggers/linux.git crc32-lib-v1 > > CRC32 is a family of common non-cryptographic integrity check algorithms > that are fairly fast with a portable C implementation and become far > faster still with the CRC32 or carryless multiplication instructions > that most CPUs have. 9 architectures already have optimized code for at > least some CRC32 variants; however, except for arm64 this optimized code > was only accessible through the crypto API, not the library functions. > > This patchset fixes that so that the CRC32 library functions use the > optimized code. This allows users to just use the library instead of > the crypto API. This is much simpler and also improves performance due > to eliminating the crypto API overhead including an indirect call. Some > examples of updating users are included at the end of the patchset. > > Note: crc32c() was a weird case. It was a library function layered on > top of the crypto API, which in turn is layered on top of the real > library functions. So while it was easy to use, it was still subject to > the crypto API overhead. This patchset provides CRC32C acceleration in > the real library functions directly. > > The updated CRC32 library design is: > > - Each arch's CRC32 code (all variants) is in arch/$ARCH/lib/crc32*. > This adopts what arm64 and riscv already did. Note, the crypto > directory is not used because CRC32 is not a cryptographic algorithm. > > - Weak symbols are no longer used. Instead there are crc32*_base() and > crc32*_arch(), and the appropriate ones are called based on the > kconfig. This is similar to how the ChaCha20 library code works. > > - Each arch's CRC32 code is enabled by default when CRC32 is enabled, > but it can now be disabled, controlled by the choice that previously > controlled the base implementation only. It can also now be built as > a module if CRC32 is a module too. > > - Instead of lots of pointless glue code that wires up each CRC32 > variant to the crypto API for each architecture, we now just rely on > the existing shash algorithms that use the library functions. > > - As before, the library functions don't provide access to off-CPU > crypto accelerators. But these appear to have very little, if any, > real-world relevance for CRC32 which is very fast on CPUs. > > Future work should apply a similar cleanup to crct10dif which is a > variant of CRC16. > > I tested all arches in QEMU using CONFIG_CRC32_SELFTEST and the crypto > self-tests, except for mips which I couldn't figure out how to do. > > This patchset has the following dependencies on recent patches: > > - "crypto - move crypto_simd_disabled_for_test to lib" > (https://lore.kernel.org/linux-crypto/20241018235343.425758-1-ebiggers@xxxxxxxxxx/) > - "crypto: x86/crc32c - jump table elimination and other cleanups" > (https://lore.kernel.org/linux-crypto/20241014042447.50197-1-ebiggers@xxxxxxxxxx/) > - "arm64: Speed up CRC-32 using PMULL instructions" > (https://lore.kernel.org/linux-crypto/20241018075347.2821102-5-ardb+git@xxxxxxxxxx/) > - "crypto: Enable fuzz testing for arch code" > (https://lore.kernel.org/linux-crypto/20241016185722.400643-4-ardb+git@xxxxxxxxxx/) > - "crypto: mips/crc32 - fix the CRC32C implementation" > (https://lore.kernel.org/linux-crypto/20241020180258.8060-1-ebiggers@xxxxxxxxxx/) > > Everything can be retrieved from git using the command given earlier. > > Since this patchset touches many areas, getting it merged may be > difficult. One option is a pull request with the whole patchset > directly to Linus. Another is to have at least patches 1-2 and the > above dependencies taken through the crypto tree in v6.13; then the arch > patches can land separately afterwards, followed by the rest. > > Eric Biggers (15): > lib/crc32: drop leading underscores from __crc32c_le_base > lib/crc32: improve support for arch-specific overrides > arm/crc32: expose CRC32 functions through lib > loongarch/crc32: expose CRC32 functions through lib > mips/crc32: expose CRC32 functions through lib > powerpc/crc32: expose CRC32 functions through lib > s390/crc32: expose CRC32 functions through lib > sparc/crc32: expose CRC32 functions through lib > x86/crc32: update prototype for crc_pcl() > x86/crc32: update prototype for crc32_pclmul_le_16() > x86/crc32: expose CRC32 functions through lib > lib/crc32: make crc32c() go directly to lib > ext4: switch to using the crc32c library > jbd2: switch to using the crc32c library > f2fs: switch to using the crc32 library > ... > 89 files changed, 1002 insertions(+), 2455 deletions(-) Very nice cleanup! For the series: Reviewed-by: Ard Biesheuvel <ardb@xxxxxxxxxx>
