From: Darrick J. Wong <darrick.wong@xxxxxxxxxx> The crc32c code used in xfsprogs was copied directly from the Linux kernel. However, that code selects slice-by-4 by default, which isn't the fastest -- that's slice-by-8, which trades table size for speed. Fix some makefile dependency problems and explicitly select the algorithm we want. With this patch applied, I see about a 10% drop in CPU time running xfs_repair. Signed-off-by: Darrick J. Wong <darrick.wong@xxxxxxxxxx> --- libxfs/Makefile | 4 ++-- libxfs/crc32defs.h | 34 ++++++++++++++++++++++++++++++++++ 2 files changed, 36 insertions(+), 2 deletions(-) diff --git a/libxfs/Makefile b/libxfs/Makefile index baba02f..d248c1f 100644 --- a/libxfs/Makefile +++ b/libxfs/Makefile @@ -122,7 +122,7 @@ LDIRT = gen_crc32table crc32table.h crc32selftest default: crc32selftest ltdepend $(LTLIBRARY) -crc32table.h: gen_crc32table.c +crc32table.h: gen_crc32table.c crc32defs.h @echo " [CC] gen_crc32table" $(Q) $(BUILD_CC) $(BUILD_CFLAGS) -o gen_crc32table $< @echo " [GENERATE] $@" @@ -133,7 +133,7 @@ crc32table.h: gen_crc32table.c # systems/architectures. Hence we make sure that xfsprogs will never use a # busted CRC calculation at build time and hence avoid putting bad CRCs down on # disk. -crc32selftest: gen_crc32table.c crc32table.h crc32.c +crc32selftest: gen_crc32table.c crc32table.h crc32.c crc32defs.h @echo " [TEST] CRC32" $(Q) $(BUILD_CC) $(BUILD_CFLAGS) -D CRC32_SELFTEST=1 crc32.c -o $@ $(Q) ./$@ diff --git a/libxfs/crc32defs.h b/libxfs/crc32defs.h index 64cba2c..2999782 100644 --- a/libxfs/crc32defs.h +++ b/libxfs/crc32defs.h @@ -1,4 +1,38 @@ /* + * Use slice-by-8, which is the fastest variant. + * + * Calculate checksum 8 bytes at a time with a clever slicing algorithm. + * This is the fastest algorithm, but comes with a 8KiB lookup table. + * Most modern processors have enough cache to hold this table without + * thrashing the cache. + * + * The Linux kernel uses this as the default implementation "unless you + * have a good reason not to". The reason why Kconfig urges you to pick + * SLICEBY8 is because people challenged the assertion that we should + * always use slice by 8, so Darrick wrote a crc microbenchmark utility + * and ran it on as many machines as he could get his hands on to show + * that sb8 was the fastest. + * + * Every 64-bit machine (and most of the 32-bit ones too) saw the best + * results with sb8. Any machine with more than 4K of cache saw better + * results. The spreadsheet still exists today[1]; note that + * 'crc32-kern-le' corresponds to the slice by 4 algorithm which is the + * default unless CRC_LE_BITS is defined explicitly. + * + * FWIW, there are a handful of board defconfigs in the kernel that + * don't pick sliceby8. These are all embedded 32-bit mips/ppc systems + * with very small cache sizes which experience cache thrashing with the + * slice by 8 algorithm, and therefore chose to pick defaults that are + * saner for their particular board configuration. For nearly all of + * XFS' perceived userbase (which we assume are 32 and 64-bit machines + * with sufficiently large CPU cache and largeish storage devices) slice + * by 8 is the right choice. + * + * [1] https://goo.gl/0LSzsG ("crc32c_bench") + */ +#define CRC_LE_BITS 64 + +/* * There are multiple 16-bit CRC polynomials in common use, but this is * *the* standard CRC-32 polynomial, first popularized by Ethernet. * x^32+x^26+x^23+x^22+x^16+x^12+x^11+x^10+x^8+x^7+x^5+x^4+x^2+x^1+x^0 -- To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
