crypto: arm64/crct10dif - port x86 SSE implementation to arm64
authorArd Biesheuvel <ard.biesheuvel@linaro.org>
Mon, 5 Dec 2016 18:42:25 +0000 (18:42 +0000)
committerHerbert Xu <herbert@gondor.apana.org.au>
Wed, 7 Dec 2016 12:01:17 +0000 (20:01 +0800)
This is a transliteration of the Intel algorithm implemented
using SSE and PCLMULQDQ instructions that resides in the file
arch/x86/crypto/crct10dif-pcl-asm_64.S, but simplified to only
operate on buffers that are 16 byte aligned (but of any size)

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
arch/arm64/crypto/Kconfig
arch/arm64/crypto/Makefile
arch/arm64/crypto/crct10dif-ce-core.S [new file with mode: 0644]
arch/arm64/crypto/crct10dif-ce-glue.c [new file with mode: 0644]

index c5ce39cd9b8f3206f879c6e373a0478ebdf89368..f1e6dd0fc174ab7445d185c7a6bedd7b2f5f81c5 100644 (file)
@@ -31,6 +31,11 @@ config CRYPTO_GHASH_ARM64_CE
        depends on ARM64 && KERNEL_MODE_NEON
        select CRYPTO_HASH
 
+config CRYPTO_CRCT10DIF_ARM64_CE
+       tristate "CRCT10DIF digest algorithm using PMULL instructions"
+       depends on KERNEL_MODE_NEON && CRC_T10DIF
+       select CRYPTO_HASH
+
 config CRYPTO_AES_ARM64_CE
        tristate "AES core cipher using ARMv8 Crypto Extensions"
        depends on ARM64 && KERNEL_MODE_NEON
index 7707867280823a8930e8f356cf868e2a48c691f5..d3f1ba6d4771ac5b96a937c040c9deccac83055c 100644 (file)
@@ -17,6 +17,9 @@ sha2-ce-y := sha2-ce-glue.o sha2-ce-core.o
 obj-$(CONFIG_CRYPTO_GHASH_ARM64_CE) += ghash-ce.o
 ghash-ce-y := ghash-ce-glue.o ghash-ce-core.o
 
+obj-$(CONFIG_CRYPTO_CRCT10DIF_ARM64_CE) += crct10dif-ce.o
+crct10dif-ce-y := crct10dif-ce-core.o crct10dif-ce-glue.o
+
 obj-$(CONFIG_CRYPTO_AES_ARM64_CE) += aes-ce-cipher.o
 CFLAGS_aes-ce-cipher.o += -march=armv8-a+crypto
 
diff --git a/arch/arm64/crypto/crct10dif-ce-core.S b/arch/arm64/crypto/crct10dif-ce-core.S
new file mode 100644 (file)
index 0000000..d5b5a8c
--- /dev/null
@@ -0,0 +1,392 @@
+//
+// Accelerated CRC-T10DIF using arm64 NEON and Crypto Extensions instructions
+//
+// Copyright (C) 2016 Linaro Ltd <ard.biesheuvel@linaro.org>
+//
+// This program is free software; you can redistribute it and/or modify
+// it under the terms of the GNU General Public License version 2 as
+// published by the Free Software Foundation.
+//
+
+//
+// Implement fast CRC-T10DIF computation with SSE and PCLMULQDQ instructions
+//
+// Copyright (c) 2013, Intel Corporation
+//
+// Authors:
+//     Erdinc Ozturk <erdinc.ozturk@intel.com>
+//     Vinodh Gopal <vinodh.gopal@intel.com>
+//     James Guilford <james.guilford@intel.com>
+//     Tim Chen <tim.c.chen@linux.intel.com>
+//
+// This software is available to you under a choice of one of two
+// licenses.  You may choose to be licensed under the terms of the GNU
+// General Public License (GPL) Version 2, available from the file
+// COPYING in the main directory of this source tree, or the
+// OpenIB.org BSD license below:
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+// * Redistributions of source code must retain the above copyright
+//   notice, this list of conditions and the following disclaimer.
+//
+// * Redistributions in binary form must reproduce the above copyright
+//   notice, this list of conditions and the following disclaimer in the
+//   documentation and/or other materials provided with the
+//   distribution.
+//
+// * Neither the name of the Intel Corporation nor the names of its
+//   contributors may be used to endorse or promote products derived from
+//   this software without specific prior written permission.
+//
+//
+// THIS SOFTWARE IS PROVIDED BY INTEL CORPORATION ""AS IS"" AND ANY
+// EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+// PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL CORPORATION OR
+// CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+// EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+// PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+// PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+// LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+// SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+//       Function API:
+//       UINT16 crc_t10dif_pcl(
+//               UINT16 init_crc, //initial CRC value, 16 bits
+//               const unsigned char *buf, //buffer pointer to calculate CRC on
+//               UINT64 len //buffer length in bytes (64-bit data)
+//       );
+//
+//       Reference paper titled "Fast CRC Computation for Generic
+//     Polynomials Using PCLMULQDQ Instruction"
+//       URL: http://www.intel.com/content/dam/www/public/us/en/documents
+//  /white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
+//
+//
+
+#include <linux/linkage.h>
+#include <asm/assembler.h>
+
+       .text
+       .cpu            generic+crypto
+
+       arg1_low32      .req    w0
+       arg2            .req    x1
+       arg3            .req    x2
+
+       vzr             .req    v13
+
+ENTRY(crc_t10dif_pmull)
+       movi            vzr.16b, #0             // init zero register
+
+       // adjust the 16-bit initial_crc value, scale it to 32 bits
+       lsl             arg1_low32, arg1_low32, #16
+
+       // check if smaller than 256
+       cmp             arg3, #256
+
+       // for sizes less than 128, we can't fold 64B at a time...
+       b.lt            _less_than_128
+
+       // load the initial crc value
+       // crc value does not need to be byte-reflected, but it needs
+       // to be moved to the high part of the register.
+       // because data will be byte-reflected and will align with
+       // initial crc at correct place.
+       movi            v10.16b, #0
+       mov             v10.s[3], arg1_low32            // initial crc
+
+       // receive the initial 64B data, xor the initial crc value
+       ldp             q0, q1, [arg2]
+       ldp             q2, q3, [arg2, #0x20]
+       ldp             q4, q5, [arg2, #0x40]
+       ldp             q6, q7, [arg2, #0x60]
+       add             arg2, arg2, #0x80
+
+CPU_LE(        rev64           v0.16b, v0.16b                  )
+CPU_LE(        rev64           v1.16b, v1.16b                  )
+CPU_LE(        rev64           v2.16b, v2.16b                  )
+CPU_LE(        rev64           v3.16b, v3.16b                  )
+CPU_LE(        rev64           v4.16b, v4.16b                  )
+CPU_LE(        rev64           v5.16b, v5.16b                  )
+CPU_LE(        rev64           v6.16b, v6.16b                  )
+CPU_LE(        rev64           v7.16b, v7.16b                  )
+
+CPU_LE(        ext             v0.16b, v0.16b, v0.16b, #8      )
+CPU_LE(        ext             v1.16b, v1.16b, v1.16b, #8      )
+CPU_LE(        ext             v2.16b, v2.16b, v2.16b, #8      )
+CPU_LE(        ext             v3.16b, v3.16b, v3.16b, #8      )
+CPU_LE(        ext             v4.16b, v4.16b, v4.16b, #8      )
+CPU_LE(        ext             v5.16b, v5.16b, v5.16b, #8      )
+CPU_LE(        ext             v6.16b, v6.16b, v6.16b, #8      )
+CPU_LE(        ext             v7.16b, v7.16b, v7.16b, #8      )
+
+       // XOR the initial_crc value
+       eor             v0.16b, v0.16b, v10.16b
+
+       ldr             q10, rk3        // xmm10 has rk3 and rk4
+                                       // type of pmull instruction
+                                       // will determine which constant to use
+
+       //
+       // we subtract 256 instead of 128 to save one instruction from the loop
+       //
+       sub             arg3, arg3, #256
+
+       // at this section of the code, there is 64*x+y (0<=y<64) bytes of
+       // buffer. The _fold_64_B_loop will fold 64B at a time
+       // until we have 64+y Bytes of buffer
+
+
+       // fold 64B at a time. This section of the code folds 4 vector
+       // registers in parallel
+_fold_64_B_loop:
+
+       .macro          fold64, reg1, reg2
+       ldp             q11, q12, [arg2], #0x20
+
+       pmull2          v8.1q, \reg1\().2d, v10.2d
+       pmull           \reg1\().1q, \reg1\().1d, v10.1d
+
+CPU_LE(        rev64           v11.16b, v11.16b                )
+CPU_LE(        rev64           v12.16b, v12.16b                )
+
+       pmull2          v9.1q, \reg2\().2d, v10.2d
+       pmull           \reg2\().1q, \reg2\().1d, v10.1d
+
+CPU_LE(        ext             v11.16b, v11.16b, v11.16b, #8   )
+CPU_LE(        ext             v12.16b, v12.16b, v12.16b, #8   )
+
+       eor             \reg1\().16b, \reg1\().16b, v8.16b
+       eor             \reg2\().16b, \reg2\().16b, v9.16b
+       eor             \reg1\().16b, \reg1\().16b, v11.16b
+       eor             \reg2\().16b, \reg2\().16b, v12.16b
+       .endm
+
+       fold64          v0, v1
+       fold64          v2, v3
+       fold64          v4, v5
+       fold64          v6, v7
+
+       subs            arg3, arg3, #128
+
+       // check if there is another 64B in the buffer to be able to fold
+       b.ge            _fold_64_B_loop
+
+       // at this point, the buffer pointer is pointing at the last y Bytes
+       // of the buffer the 64B of folded data is in 4 of the vector
+       // registers: v0, v1, v2, v3
+
+       // fold the 8 vector registers to 1 vector register with different
+       // constants
+
+       ldr             q10, rk9
+
+       .macro          fold16, reg, rk
+       pmull           v8.1q, \reg\().1d, v10.1d
+       pmull2          \reg\().1q, \reg\().2d, v10.2d
+       .ifnb           \rk
+       ldr             q10, \rk
+       .endif
+       eor             v7.16b, v7.16b, v8.16b
+       eor             v7.16b, v7.16b, \reg\().16b
+       .endm
+
+       fold16          v0, rk11
+       fold16          v1, rk13
+       fold16          v2, rk15
+       fold16          v3, rk17
+       fold16          v4, rk19
+       fold16          v5, rk1
+       fold16          v6
+
+       // instead of 64, we add 48 to the loop counter to save 1 instruction
+       // from the loop instead of a cmp instruction, we use the negative
+       // flag with the jl instruction
+       adds            arg3, arg3, #(128-16)
+       b.lt            _final_reduction_for_128
+
+       // now we have 16+y bytes left to reduce. 16 Bytes is in register v7
+       // and the rest is in memory. We can fold 16 bytes at a time if y>=16
+       // continue folding 16B at a time
+
+_16B_reduction_loop:
+       pmull           v8.1q, v7.1d, v10.1d
+       pmull2          v7.1q, v7.2d, v10.2d
+       eor             v7.16b, v7.16b, v8.16b
+
+       ldr             q0, [arg2], #16
+CPU_LE(        rev64           v0.16b, v0.16b                  )
+CPU_LE(        ext             v0.16b, v0.16b, v0.16b, #8      )
+       eor             v7.16b, v7.16b, v0.16b
+       subs            arg3, arg3, #16
+
+       // instead of a cmp instruction, we utilize the flags with the
+       // jge instruction equivalent of: cmp arg3, 16-16
+       // check if there is any more 16B in the buffer to be able to fold
+       b.ge            _16B_reduction_loop
+
+       // now we have 16+z bytes left to reduce, where 0<= z < 16.
+       // first, we reduce the data in the xmm7 register
+
+_final_reduction_for_128:
+       // check if any more data to fold. If not, compute the CRC of
+       // the final 128 bits
+       adds            arg3, arg3, #16
+       b.eq            _128_done
+
+       // here we are getting data that is less than 16 bytes.
+       // since we know that there was data before the pointer, we can
+       // offset the input pointer before the actual point, to receive
+       // exactly 16 bytes. after that the registers need to be adjusted.
+_get_last_two_regs:
+       add             arg2, arg2, arg3
+       ldr             q1, [arg2, #-16]
+CPU_LE(        rev64           v1.16b, v1.16b                  )
+CPU_LE(        ext             v1.16b, v1.16b, v1.16b, #8      )
+
+       // get rid of the extra data that was loaded before
+       // load the shift constant
+       adr             x4, tbl_shf_table + 16
+       sub             x4, x4, arg3
+       ld1             {v0.16b}, [x4]
+
+       // shift v2 to the left by arg3 bytes
+       tbl             v2.16b, {v7.16b}, v0.16b
+
+       // shift v7 to the right by 16-arg3 bytes
+       movi            v9.16b, #0x80
+       eor             v0.16b, v0.16b, v9.16b
+       tbl             v7.16b, {v7.16b}, v0.16b
+
+       // blend
+       sshr            v0.16b, v0.16b, #7      // convert to 8-bit mask
+       bsl             v0.16b, v2.16b, v1.16b
+
+       // fold 16 Bytes
+       pmull           v8.1q, v7.1d, v10.1d
+       pmull2          v7.1q, v7.2d, v10.2d
+       eor             v7.16b, v7.16b, v8.16b
+       eor             v7.16b, v7.16b, v0.16b
+
+_128_done:
+       // compute crc of a 128-bit value
+       ldr             q10, rk5                // rk5 and rk6 in xmm10
+
+       // 64b fold
+       ext             v0.16b, vzr.16b, v7.16b, #8
+       mov             v7.d[0], v7.d[1]
+       pmull           v7.1q, v7.1d, v10.1d
+       eor             v7.16b, v7.16b, v0.16b
+
+       // 32b fold
+       ext             v0.16b, v7.16b, vzr.16b, #4
+       mov             v7.s[3], vzr.s[0]
+       pmull2          v0.1q, v0.2d, v10.2d
+       eor             v7.16b, v7.16b, v0.16b
+
+       // barrett reduction
+_barrett:
+       ldr             q10, rk7
+       mov             v0.d[0], v7.d[1]
+
+       pmull           v0.1q, v0.1d, v10.1d
+       ext             v0.16b, vzr.16b, v0.16b, #12
+       pmull2          v0.1q, v0.2d, v10.2d
+       ext             v0.16b, vzr.16b, v0.16b, #12
+       eor             v7.16b, v7.16b, v0.16b
+       mov             w0, v7.s[1]
+
+_cleanup:
+       // scale the result back to 16 bits
+       lsr             x0, x0, #16
+       ret
+
+_less_than_128:
+       cbz             arg3, _cleanup
+
+       movi            v0.16b, #0
+       mov             v0.s[3], arg1_low32     // get the initial crc value
+
+       ldr             q7, [arg2], #0x10
+CPU_LE(        rev64           v7.16b, v7.16b                  )
+CPU_LE(        ext             v7.16b, v7.16b, v7.16b, #8      )
+       eor             v7.16b, v7.16b, v0.16b  // xor the initial crc value
+
+       cmp             arg3, #16
+       b.eq            _128_done               // exactly 16 left
+       b.lt            _less_than_16_left
+
+       ldr             q10, rk1                // rk1 and rk2 in xmm10
+
+       // update the counter. subtract 32 instead of 16 to save one
+       // instruction from the loop
+       subs            arg3, arg3, #32
+       b.ge            _16B_reduction_loop
+
+       add             arg3, arg3, #16
+       b               _get_last_two_regs
+
+_less_than_16_left:
+       // shl r9, 4
+       adr             x0, tbl_shf_table + 16
+       sub             x0, x0, arg3
+       ld1             {v0.16b}, [x0]
+       movi            v9.16b, #0x80
+       eor             v0.16b, v0.16b, v9.16b
+       tbl             v7.16b, {v7.16b}, v0.16b
+       b               _128_done
+ENDPROC(crc_t10dif_pmull)
+
+// precomputed constants
+// these constants are precomputed from the poly:
+// 0x8bb70000 (0x8bb7 scaled to 32 bits)
+       .align          4
+// Q = 0x18BB70000
+// rk1 = 2^(32*3) mod Q << 32
+// rk2 = 2^(32*5) mod Q << 32
+// rk3 = 2^(32*15) mod Q << 32
+// rk4 = 2^(32*17) mod Q << 32
+// rk5 = 2^(32*3) mod Q << 32
+// rk6 = 2^(32*2) mod Q << 32
+// rk7 = floor(2^64/Q)
+// rk8 = Q
+
+rk1:   .octa           0x06df0000000000002d56000000000000
+rk3:   .octa           0x7cf50000000000009d9d000000000000
+rk5:   .octa           0x13680000000000002d56000000000000
+rk7:   .octa           0x000000018bb7000000000001f65a57f8
+rk9:   .octa           0xbfd6000000000000ceae000000000000
+rk11:  .octa           0x713c0000000000001e16000000000000
+rk13:  .octa           0x80a6000000000000f7f9000000000000
+rk15:  .octa           0xe658000000000000044c000000000000
+rk17:  .octa           0xa497000000000000ad18000000000000
+rk19:  .octa           0xe7b50000000000006ee3000000000000
+
+tbl_shf_table:
+// use these values for shift constants for the tbl/tbx instruction
+// different alignments result in values as shown:
+//     DDQ 0x008f8e8d8c8b8a898887868584838281 # shl 15 (16-1) / shr1
+//     DDQ 0x01008f8e8d8c8b8a8988878685848382 # shl 14 (16-3) / shr2
+//     DDQ 0x0201008f8e8d8c8b8a89888786858483 # shl 13 (16-4) / shr3
+//     DDQ 0x030201008f8e8d8c8b8a898887868584 # shl 12 (16-4) / shr4
+//     DDQ 0x04030201008f8e8d8c8b8a8988878685 # shl 11 (16-5) / shr5
+//     DDQ 0x0504030201008f8e8d8c8b8a89888786 # shl 10 (16-6) / shr6
+//     DDQ 0x060504030201008f8e8d8c8b8a898887 # shl 9  (16-7) / shr7
+//     DDQ 0x07060504030201008f8e8d8c8b8a8988 # shl 8  (16-8) / shr8
+//     DDQ 0x0807060504030201008f8e8d8c8b8a89 # shl 7  (16-9) / shr9
+//     DDQ 0x090807060504030201008f8e8d8c8b8a # shl 6  (16-10) / shr10
+//     DDQ 0x0a090807060504030201008f8e8d8c8b # shl 5  (16-11) / shr11
+//     DDQ 0x0b0a090807060504030201008f8e8d8c # shl 4  (16-12) / shr12
+//     DDQ 0x0c0b0a090807060504030201008f8e8d # shl 3  (16-13) / shr13
+//     DDQ 0x0d0c0b0a090807060504030201008f8e # shl 2  (16-14) / shr14
+//     DDQ 0x0e0d0c0b0a090807060504030201008f # shl 1  (16-15) / shr15
+
+       .byte            0x0, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87
+       .byte           0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f
+       .byte            0x0,  0x1,  0x2,  0x3,  0x4,  0x5,  0x6,  0x7
+       .byte            0x8,  0x9,  0xa,  0xb,  0xc,  0xd,  0xe , 0x0
diff --git a/arch/arm64/crypto/crct10dif-ce-glue.c b/arch/arm64/crypto/crct10dif-ce-glue.c
new file mode 100644 (file)
index 0000000..60cb590
--- /dev/null
@@ -0,0 +1,95 @@
+/*
+ * Accelerated CRC-T10DIF using arm64 NEON and Crypto Extensions instructions
+ *
+ * Copyright (C) 2016 Linaro Ltd <ard.biesheuvel@linaro.org>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+
+#include <linux/cpufeature.h>
+#include <linux/crc-t10dif.h>
+#include <linux/init.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/string.h>
+
+#include <crypto/internal/hash.h>
+
+#include <asm/neon.h>
+
+#define CRC_T10DIF_PMULL_CHUNK_SIZE    16U
+
+asmlinkage u16 crc_t10dif_pmull(u16 init_crc, const u8 buf[], u64 len);
+
+static int crct10dif_init(struct shash_desc *desc)
+{
+       u16 *crc = shash_desc_ctx(desc);
+
+       *crc = 0;
+       return 0;
+}
+
+static int crct10dif_update(struct shash_desc *desc, const u8 *data,
+                           unsigned int length)
+{
+       u16 *crc = shash_desc_ctx(desc);
+       unsigned int l;
+
+       if (unlikely((u64)data % CRC_T10DIF_PMULL_CHUNK_SIZE)) {
+               l = min_t(u32, length, CRC_T10DIF_PMULL_CHUNK_SIZE -
+                         ((u64)data % CRC_T10DIF_PMULL_CHUNK_SIZE));
+
+               *crc = crc_t10dif_generic(*crc, data, l);
+
+               length -= l;
+               data += l;
+       }
+
+       if (length > 0) {
+               kernel_neon_begin_partial(14);
+               *crc = crc_t10dif_pmull(*crc, data, length);
+               kernel_neon_end();
+       }
+
+       return 0;
+}
+
+static int crct10dif_final(struct shash_desc *desc, u8 *out)
+{
+       u16 *crc = shash_desc_ctx(desc);
+
+       *(u16 *)out = *crc;
+       return 0;
+}
+
+static struct shash_alg crc_t10dif_alg = {
+       .digestsize             = CRC_T10DIF_DIGEST_SIZE,
+       .init                   = crct10dif_init,
+       .update                 = crct10dif_update,
+       .final                  = crct10dif_final,
+       .descsize               = CRC_T10DIF_DIGEST_SIZE,
+
+       .base.cra_name          = "crct10dif",
+       .base.cra_driver_name   = "crct10dif-arm64-ce",
+       .base.cra_priority      = 200,
+       .base.cra_blocksize     = CRC_T10DIF_BLOCK_SIZE,
+       .base.cra_module        = THIS_MODULE,
+};
+
+static int __init crc_t10dif_mod_init(void)
+{
+       return crypto_register_shash(&crc_t10dif_alg);
+}
+
+static void __exit crc_t10dif_mod_exit(void)
+{
+       crypto_unregister_shash(&crc_t10dif_alg);
+}
+
+module_cpu_feature_match(PMULL, crc_t10dif_mod_init);
+module_exit(crc_t10dif_mod_exit);
+
+MODULE_AUTHOR("Ard Biesheuvel <ard.biesheuvel@linaro.org>");
+MODULE_LICENSE("GPL v2");