Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
lib/crc: simplify the kconfig options for CRC implementations
Make the following simplifications to the kconfig options for choosing CRC implementations for CRC32 and CRC_T10DIF: 1. Make the option to disable the arch-optimized code be visible only when CONFIG_EXPERT=y. 2. Make a single option control the inclusion of the arch-optimized code for all enabled CRC variants. 3. Make CRC32_SARWATE (a.k.a. slice-by-1 or byte-by-byte) be the only generic CRC32 implementation. The result is there is now just one option, CRC_OPTIMIZATIONS, which is default y and can be disabled only when CONFIG_EXPERT=y. Rationale: 1. Enabling the arch-optimized code is nearly always the right choice. However, people trying to build the tiniest kernel possible would find some use in disabling it. Anything we add to CRC32 is de facto unconditional, given that CRC32 gets selected by something in nearly all kernels. And unfortunately enabling the arch CRC code does not eliminate the need to build the generic CRC code into the kernel too, due to CPU feature dependencies. The size of the arch CRC code will also increase slightly over time as more CRC variants get added and more implementations targeting different instruction set extensions get added. Thus, it seems worthwhile to still provide an option to disable it, but it should be considered an expert-level tweak. 2. Considering the use case described in (1), there doesn't seem to be sufficient value in making the arch-optimized CRC code be independently configurable for different CRC variants. Note also that multiple variants were already grouped together, e.g. CONFIG_CRC32 actually enables three different variants of CRC32. 3. The bit-by-bit implementation is uselessly slow, whereas slice-by-n for n=4 and n=8 use tables that are inconveniently large: 4096 bytes and 8192 bytes respectively, compared to 1024 bytes for n=1. Higher n gives higher instruction-level parallelism, so higher n easily wins on traditional microbenchmarks on most CPUs. However, the larger tables, which are accessed randomly, can be harmful in real-world situations where the dcache may be cold or useful data may need be evicted from the dcache. Meanwhile, today most architectures have much faster CRC32 implementations using dedicated CRC32 instructions or carryless multiplication instructions anyway, which make the generic code obsolete in most cases especially on long messages. Another reason for going with n=1 is that this is already what is used by all the other CRC variants in the kernel. CRC32 was unique in having support for larger tables. But as per the above this can be considered an outdated optimization. The standardization on slice-by-1 a.k.a. CRC32_SARWATE makes much of the code in lib/crc32.c unused. A later patch will clean that up. Link: https://lore.kernel.org/r/20250123212904.118683-2-ebiggers@kernel.org Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com> Signed-off-by: Eric Biggers <ebiggers@google.com>
- Loading branch information