From: Megha Dey Subject: [PATCH V8 0/5] crypto: AES CBC multibuffer implementation Date: Tue, 9 Jan 2018 16:09:03 -0800 Message-ID: <1515542948-24041-1-git-send-email-megha.dey@linux.intel.com> Cc: megha.dey@intel.com, Megha Dey To: linux-kernel@vger.kernel.org, linux-crypto@vger.kernel.org, davem@davemloft.net, herbert@gondor.apana.org.au Return-path: Received: from mga06.intel.com ([134.134.136.31]:37882 "EHLO mga06.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753115AbeAIXw6 (ORCPT ); Tue, 9 Jan 2018 18:52:58 -0500 Sender: linux-crypto-owner@vger.kernel.org List-ID: In this patch series, we introduce AES CBC encryption that is parallelized on x86_64 cpu with XMM registers. The multi-buffer technique encrypt 8 data streams in parallel with SIMD instructions. Decryption is handled as in the existing AESNI Intel CBC implementation which can already parallelize decryption even for a single data stream. Please see the multi-buffer whitepaper for details of the technique: http://www.intel.com/content/www/us/en/communications/communications-ia-multi-buffer-paper.html It is important that any driver uses this algorithm properly for scenarios where we have many data streams that can fill up the data lanes most of the time. It shouldn't be used when only a single data stream is expected mostly. Otherwise, we may incur extra delays when we have frequent gaps in data lanes, causing us to wait till data come in to fill the data lanes before initiating encryption. We may have to wait for flush operations to commence when no new data come in after some wait time. However, we keep this extra delay to a minimum by opportunistically flushing the unfinished jobs if crypto daemon is the only active task running on a cpu. By using this technique, we saw a throughput increase of up to 5.7x under optimal conditions when we have fully loaded encryption jobs filling up all the data lanes. Change Log: v8 1. Remove the notify_callback construct 2. Remove remaining irq_disabled check 3. Remove related tcrypt test as it is already merged v7 1. Add the CRYPTO_ALG_ASYNC flag to the internal algorithm 2. Remove the irq_disabled check v6 1. Move away from the compat naming scheme and update the names of the inner and outer algorithm 2. Move wrapper code around synchronous internal algorithm from simd.c to mcryptd.c v5 1. Use an async implementation of the inner algorithm instead of sync and use the latest skcipher interface instead of the older blkcipher interface. (we have picked up this work after a while) v4 1. Make the decrypt path also use ablkcpher walk. http://lkml.iu.edu/hypermail/linux/kernel/1512.0/01807.html v3 1. Use ablkcipher_walk helpers to walk the scatter gather list and eliminated needs to modify blkcipher_walk for multibuffer cipher v2 1. Update cpu feature check to make sure SSE is supported 2. Fix up unloading of aes-cbc-mb module to properly free memory Megha Dey (5): crypto: Multi-buffer encryption infrastructure support crypto: AES CBC multi-buffer data structures crypto: AES CBC multi-buffer scheduler crypto: AES CBC by8 encryption crypto: AES CBC multi-buffer glue code arch/x86/crypto/Makefile | 1 + arch/x86/crypto/aes-cbc-mb/Makefile | 22 + arch/x86/crypto/aes-cbc-mb/aes_cbc_enc_x8.S | 775 +++++++++++++++++++++ arch/x86/crypto/aes-cbc-mb/aes_cbc_mb.c | 698 +++++++++++++++++++ arch/x86/crypto/aes-cbc-mb/aes_cbc_mb_ctx.h | 97 +++ arch/x86/crypto/aes-cbc-mb/aes_cbc_mb_mgr.h | 132 ++++ arch/x86/crypto/aes-cbc-mb/aes_mb_mgr_init.c | 146 ++++ arch/x86/crypto/aes-cbc-mb/mb_mgr_datastruct.S | 271 +++++++ arch/x86/crypto/aes-cbc-mb/mb_mgr_inorder_x8_asm.S | 223 ++++++ arch/x86/crypto/aes-cbc-mb/mb_mgr_ooo_x8_asm.S | 417 +++++++++++ arch/x86/crypto/aes-cbc-mb/reg_sizes.S | 126 ++++ crypto/Kconfig | 15 + crypto/mcryptd.c | 475 +++++++++++++ include/crypto/mcryptd.h | 56 ++ 14 files changed, 3454 insertions(+) create mode 100644 arch/x86/crypto/aes-cbc-mb/Makefile create mode 100644 arch/x86/crypto/aes-cbc-mb/aes_cbc_enc_x8.S create mode 100644 arch/x86/crypto/aes-cbc-mb/aes_cbc_mb.c create mode 100644 arch/x86/crypto/aes-cbc-mb/aes_cbc_mb_ctx.h create mode 100644 arch/x86/crypto/aes-cbc-mb/aes_cbc_mb_mgr.h create mode 100644 arch/x86/crypto/aes-cbc-mb/aes_mb_mgr_init.c create mode 100644 arch/x86/crypto/aes-cbc-mb/mb_mgr_datastruct.S create mode 100644 arch/x86/crypto/aes-cbc-mb/mb_mgr_inorder_x8_asm.S create mode 100644 arch/x86/crypto/aes-cbc-mb/mb_mgr_ooo_x8_asm.S create mode 100644 arch/x86/crypto/aes-cbc-mb/reg_sizes.S -- 1.9.1