This patch series adds 'canned' compression for Intel IAA and zlib. It also adds 'dynamic' compression for Intel IAA which is compatible with zlib deflate algorithms. The original IAA crypto submissions [1] included a 'canned' compression mode, but support for 'canned' compression was removed during the review because it didn't have an equivalent software implementation available [2]. Deflate compression can be done in a variety of modes. The dynamic mode uses Huffman tables that are generated/optimized for that particular input. This gives the best compression ratio but has the longest latency. The fixed mode uses Huffman tables that are defined by the Deflate standard. It generally gives the best latency but with a lower compression ratio. The 'canned' compression mode implements a compression scheme that uses a statically defined set of Huffman tables, but where the Deflate Block Header is implied rather than stored with the compressed data. The 'canned' mode results in lower compression/decompression latencies compared to using the dynamic mode, but it results in a better compression ratio than using the fixed mode. Below is a table showing the latency improvements with zlib, between zlib dynamic and zlib canned modes, and the compression ratio for each mode while using a set of 4300 4KB pages sampled from SPEC CPU17 workloads: _________________________________________________________ | Zlib Level | Canned Latency Gain | Comp Ratio | |------------|-----------------------|------------------| | | compress | decompress | dynamic | canned | |____________|__________|____________|_________|________| | 1 | 49% | 29% | 3.16 | 2.92 | |------------|----------|------------|---------|--------| | 6 | 27% | 28% | 3.35 | 3.09 | |------------|----------|------------|---------|--------| | 9 | 12% | 29% | 3.36 | 3.11 | |____________|__________|____________|_________|________| Using the same data set as for the above table, IAA fixed-mode compression results in a compression ratio of 2.69. Using IAA canned-mode compression results in a ratio of 2.88, which is a 7% improvement. This data shows that the canned mode results in better latencies than the dynamic mode, but a better ratio than the fixed mode. Thus, the canned mode would be preferred when compression and decompression latencies are valued more than the compression ratio. 'Dynamic' mode IAA compression is a HW accelerated dynamic Deflate that is fully compatible with software decompress implementations, including zlib. 'Dynamic' IAA compression allows users to perform hardware-accelerated compression that achieves a higher compression ratio than both 'canned' and 'fixed' compression modes. Thus, Dynamic IAA compression should be used when the goal is to produce the best compression ratio while minimizing latencies by using IAA hardware acceleration. IAA decompression is fully compatible with software compress implementations, when the uncompressed size is no more than 4KB. Below is a table showing the compression ratio seen when compressing a data set of 4300 4KB pages sampled from SPEC CPU17 workloads via various IAA compression modes: |---------------------------------------| | Intel IAA | |---------------------------------------| | compression mode | compression ratio | |---------------------------------------| | fixed | 2.69 | |---------------------------------------| | canned | 2.88 | |---------------------------------------| | dynamic | 3.14 | |---------------------------------------| Patch 1/4 adds a software implementation of “canned’ mode to the existing zlib software library and exposes it as “deflate-canned”. This was done instead of creating a new canned-mode library to avoid a lot of code duplication. Testing shows that this change has no performance impact to the existing zlib algorithms. Patch 2/4 adds IAA 'canned' support which is based on the original implementation [1] and will be exposed as 'deflate-iaa-canned' Patch 3/4 adds 'dynamic' mode IAA compression support and will be exposed as 'deflate-iaa-dynamic'. Patch 4/4 adds software compression stats to the optional debugfs statistics support for IAA. [1] https://lore.kernel.org/lkml/20230605201536.738396-1-tom.zanussi@xxxxxxxxxxxxxxx/ [2] https://lore.kernel.org/lkml/ZIw%2Fjtxdg6O1O0j3@xxxxxxxxxxxxxxxxxxx/ Andre Glover (4): crypto: Add 'canned' compression mode for zlib crypto: iaa - Add deflate-canned compression algorithm crypto: iaa - Add deflate-iaa-dynamic compression algorithm crypto: iaa - Add Software Compression stats to IAA Compression Accelerator stats .../driver-api/crypto/iaa/iaa-crypto.rst | 36 +- crypto/deflate.c | 72 +++- crypto/testmgr.c | 30 ++ crypto/testmgr.h | 220 +++++++++++ drivers/crypto/intel/iaa/Kconfig | 1 + drivers/crypto/intel/iaa/Makefile | 2 +- drivers/crypto/intel/iaa/iaa_crypto.h | 44 ++- .../crypto/intel/iaa/iaa_crypto_comp_canned.c | 116 ++++++ .../intel/iaa/iaa_crypto_comp_dynamic.c | 22 ++ .../crypto/intel/iaa/iaa_crypto_comp_fixed.c | 1 + drivers/crypto/intel/iaa/iaa_crypto_main.c | 361 ++++++++++++++++-- drivers/crypto/intel/iaa/iaa_crypto_stats.c | 8 + drivers/crypto/intel/iaa/iaa_crypto_stats.h | 2 + include/linux/zlib.h | 10 + lib/Kconfig | 9 + lib/zlib_deflate/defcanned.h | 118 ++++++ lib/zlib_deflate/deflate.c | 8 +- lib/zlib_deflate/deftree.c | 15 +- lib/zlib_inflate/infcanned.h | 191 +++++++++ lib/zlib_inflate/inflate.c | 15 +- lib/zlib_inflate/inflate.h | 5 +- lib/zlib_inflate/infutil.h | 16 + 22 files changed, 1255 insertions(+), 47 deletions(-) create mode 100644 drivers/crypto/intel/iaa/iaa_crypto_comp_canned.c create mode 100644 drivers/crypto/intel/iaa/iaa_crypto_comp_dynamic.c create mode 100644 lib/zlib_deflate/defcanned.h create mode 100644 lib/zlib_inflate/infcanned.h -- 2.27.0
