arm64/crypto: improve performance of GHASH algorithm
authorArd Biesheuvel <ard.biesheuvel@linaro.org>
Mon, 16 Jun 2014 10:02:16 +0000 (11:02 +0100)
committerCatalin Marinas <catalin.marinas@arm.com>
Wed, 18 Jun 2014 11:40:54 +0000 (12:40 +0100)
This patches modifies the GHASH secure hash implementation to switch to a
faster, polynomial multiplication based reduction instead of one that uses
shifts and rotates.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
arch/arm64/crypto/ghash-ce-core.S
arch/arm64/crypto/ghash-ce-glue.c

index b9e6eaf41c9be14c5f5269477203e9eaf5565eda..dc457015884e04345ea814d0aa1d274793ad7bc4 100644 (file)
@@ -3,14 +3,6 @@
  *
  * Copyright (C) 2014 Linaro Ltd. <ard.biesheuvel@linaro.org>
  *
- * Based on arch/x86/crypto/ghash-pmullni-intel_asm.S
- *
- * Copyright (c) 2009 Intel Corp.
- *   Author: Huang Ying <ying.huang@intel.com>
- *           Vinodh Gopal
- *           Erdinc Ozturk
- *           Deniz Karakoyunlu
- *
  * 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/linkage.h>
 #include <asm/assembler.h>
 
-       DATA    .req    v0
-       SHASH   .req    v1
-       IN1     .req    v2
+       SHASH   .req    v0
+       SHASH2  .req    v1
        T1      .req    v2
        T2      .req    v3
-       T3      .req    v4
-       VZR     .req    v5
+       MASK    .req    v4
+       XL      .req    v5
+       XM      .req    v6
+       XH      .req    v7
+       IN1     .req    v7
 
        .text
        .arch           armv8-a+crypto
         *                         struct ghash_key const *k, const char *head)
         */
 ENTRY(pmull_ghash_update)
-       ld1             {DATA.16b}, [x1]
        ld1             {SHASH.16b}, [x3]
-       eor             VZR.16b, VZR.16b, VZR.16b
+       ld1             {XL.16b}, [x1]
+       movi            MASK.16b, #0xe1
+       ext             SHASH2.16b, SHASH.16b, SHASH.16b, #8
+       shl             MASK.2d, MASK.2d, #57
+       eor             SHASH2.16b, SHASH2.16b, SHASH.16b
 
        /* do the head block first, if supplied */
        cbz             x4, 0f
-       ld1             {IN1.2d}, [x4]
+       ld1             {T1.2d}, [x4]
        b               1f
 
-0:     ld1             {IN1.2d}, [x2], #16
+0:     ld1             {T1.2d}, [x2], #16
        sub             w0, w0, #1
-1:     ext             IN1.16b, IN1.16b, IN1.16b, #8
-CPU_LE(        rev64           IN1.16b, IN1.16b        )
-       eor             DATA.16b, DATA.16b, IN1.16b
 
-       /* multiply DATA by SHASH in GF(2^128) */
-       ext             T2.16b, DATA.16b, DATA.16b, #8
-       ext             T3.16b, SHASH.16b, SHASH.16b, #8
-       eor             T2.16b, T2.16b, DATA.16b
-       eor             T3.16b, T3.16b, SHASH.16b
+1:     /* multiply XL by SHASH in GF(2^128) */
+CPU_LE(        rev64           T1.16b, T1.16b  )
 
-       pmull2          T1.1q, SHASH.2d, DATA.2d        // a1 * b1
-       pmull           DATA.1q, SHASH.1d, DATA.1d      // a0 * b0
-       pmull           T2.1q, T2.1d, T3.1d             // (a1 + a0)(b1 + b0)
-       eor             T2.16b, T2.16b, T1.16b          // (a0 * b1) + (a1 * b0)
-       eor             T2.16b, T2.16b, DATA.16b
+       ext             T2.16b, XL.16b, XL.16b, #8
+       ext             IN1.16b, T1.16b, T1.16b, #8
+       eor             T1.16b, T1.16b, T2.16b
+       eor             XL.16b, XL.16b, IN1.16b
 
-       ext             T3.16b, VZR.16b, T2.16b, #8
-       ext             T2.16b, T2.16b, VZR.16b, #8
-       eor             DATA.16b, DATA.16b, T3.16b
-       eor             T1.16b, T1.16b, T2.16b  // <T1:DATA> is result of
-                                               // carry-less multiplication
+       pmull2          XH.1q, SHASH.2d, XL.2d          // a1 * b1
+       eor             T1.16b, T1.16b, XL.16b
+       pmull           XL.1q, SHASH.1d, XL.1d          // a0 * b0
+       pmull           XM.1q, SHASH2.1d, T1.1d         // (a1 + a0)(b1 + b0)
 
-       /* first phase of the reduction */
-       shl             T3.2d, DATA.2d, #1
-       eor             T3.16b, T3.16b, DATA.16b
-       shl             T3.2d, T3.2d, #5
-       eor             T3.16b, T3.16b, DATA.16b
-       shl             T3.2d, T3.2d, #57
-       ext             T2.16b, VZR.16b, T3.16b, #8
-       ext             T3.16b, T3.16b, VZR.16b, #8
-       eor             DATA.16b, DATA.16b, T2.16b
-       eor             T1.16b, T1.16b, T3.16b
+       ext             T1.16b, XL.16b, XH.16b, #8
+       eor             T2.16b, XL.16b, XH.16b
+       eor             XM.16b, XM.16b, T1.16b
+       eor             XM.16b, XM.16b, T2.16b
+       pmull           T2.1q, XL.1d, MASK.1d
 
-       /* second phase of the reduction */
-       ushr            T2.2d, DATA.2d, #5
-       eor             T2.16b, T2.16b, DATA.16b
-       ushr            T2.2d, T2.2d, #1
-       eor             T2.16b, T2.16b, DATA.16b
-       ushr            T2.2d, T2.2d, #1
-       eor             T1.16b, T1.16b, T2.16b
-       eor             DATA.16b, DATA.16b, T1.16b
+       mov             XH.d[0], XM.d[1]
+       mov             XM.d[1], XL.d[0]
+
+       eor             XL.16b, XM.16b, T2.16b
+       ext             T2.16b, XL.16b, XL.16b, #8
+       pmull           XL.1q, XL.1d, MASK.1d
+       eor             T2.16b, T2.16b, XH.16b
+       eor             XL.16b, XL.16b, T2.16b
 
        cbnz            w0, 0b
 
-       st1             {DATA.16b}, [x1]
+       st1             {XL.16b}, [x1]
        ret
 ENDPROC(pmull_ghash_update)
index ef6aa69c4e0c9d398e1c809aea55682b543cb306..833ec1e3f3e9b7491cc26da24fc0ba386b73de81 100644 (file)
@@ -67,7 +67,7 @@ static int ghash_update(struct shash_desc *desc, const u8 *src,
                blocks = len / GHASH_BLOCK_SIZE;
                len %= GHASH_BLOCK_SIZE;
 
-               kernel_neon_begin_partial(6);
+               kernel_neon_begin_partial(8);
                pmull_ghash_update(blocks, ctx->digest, src, key,
                                   partial ? ctx->buf : NULL);
                kernel_neon_end();
@@ -89,7 +89,7 @@ static int ghash_final(struct shash_desc *desc, u8 *dst)
 
                memset(ctx->buf + partial, 0, GHASH_BLOCK_SIZE - partial);
 
-               kernel_neon_begin_partial(6);
+               kernel_neon_begin_partial(8);
                pmull_ghash_update(1, ctx->digest, ctx->buf, key, NULL);
                kernel_neon_end();
        }