OSDN Git Service

crypto: arm64/aes-ce-cipher - move assembler code to .S file
authorArd Biesheuvel <ard.biesheuvel@linaro.org>
Tue, 21 Nov 2017 13:40:17 +0000 (13:40 +0000)
committerHerbert Xu <herbert@gondor.apana.org.au>
Wed, 29 Nov 2017 06:33:30 +0000 (17:33 +1100)
Most crypto drivers involving kernel mode NEON take care to put the code
that actually touches the NEON register file in a separate compilation
unit, to prevent the compiler from reordering code that preserves or
restores the NEON context with code that may corrupt it. This is
necessary because we currently have no way to express the restrictions
imposed upon use of the NEON in kernel mode in a way that the compiler
understands.

However, in the case of aes-ce-cipher, it did not seem unreasonable to
deviate from this rule, given how it does not seem possible for the
compiler to reorder cross object function calls with asm blocks whose
in- and output constraints reflect that it reads from and writes to
memory.

Now that LTO is being proposed for the arm64 kernel, it is time to
revisit this. The link time optimization may replace the function
calls to kernel_neon_begin() and kernel_neon_end() with instantiations
of the IR that make up its implementation, allowing further reordering
with the asm block.

So let's clean this up, and move the asm() blocks into a separate .S
file.

Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Reviewed-By: Nick Desaulniers <ndesaulniers@google.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
arch/arm64/crypto/Makefile
arch/arm64/crypto/aes-ce-core.S [new file with mode: 0644]
arch/arm64/crypto/aes-ce-glue.c [moved from arch/arm64/crypto/aes-ce-cipher.c with 62% similarity]

index b5edc59..f5e8295 100644 (file)
@@ -24,7 +24,7 @@ obj-$(CONFIG_CRYPTO_CRC32_ARM64_CE) += crc32-ce.o
 crc32-ce-y:= crc32-ce-core.o crc32-ce-glue.o
 
 obj-$(CONFIG_CRYPTO_AES_ARM64_CE) += aes-ce-cipher.o
-CFLAGS_aes-ce-cipher.o += -march=armv8-a+crypto
+aes-ce-cipher-y := aes-ce-core.o aes-ce-glue.o
 
 obj-$(CONFIG_CRYPTO_AES_ARM64_CE_CCM) += aes-ce-ccm.o
 aes-ce-ccm-y := aes-ce-ccm-glue.o aes-ce-ccm-core.o
diff --git a/arch/arm64/crypto/aes-ce-core.S b/arch/arm64/crypto/aes-ce-core.S
new file mode 100644 (file)
index 0000000..8efdfda
--- /dev/null
@@ -0,0 +1,87 @@
+/*
+ * Copyright (C) 2013 - 2017 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/linkage.h>
+#include <asm/assembler.h>
+
+       .arch           armv8-a+crypto
+
+ENTRY(__aes_ce_encrypt)
+       sub             w3, w3, #2
+       ld1             {v0.16b}, [x2]
+       ld1             {v1.4s}, [x0], #16
+       cmp             w3, #10
+       bmi             0f
+       bne             3f
+       mov             v3.16b, v1.16b
+       b               2f
+0:     mov             v2.16b, v1.16b
+       ld1             {v3.4s}, [x0], #16
+1:     aese            v0.16b, v2.16b
+       aesmc           v0.16b, v0.16b
+2:     ld1             {v1.4s}, [x0], #16
+       aese            v0.16b, v3.16b
+       aesmc           v0.16b, v0.16b
+3:     ld1             {v2.4s}, [x0], #16
+       subs            w3, w3, #3
+       aese            v0.16b, v1.16b
+       aesmc           v0.16b, v0.16b
+       ld1             {v3.4s}, [x0], #16
+       bpl             1b
+       aese            v0.16b, v2.16b
+       eor             v0.16b, v0.16b, v3.16b
+       st1             {v0.16b}, [x1]
+       ret
+ENDPROC(__aes_ce_encrypt)
+
+ENTRY(__aes_ce_decrypt)
+       sub             w3, w3, #2
+       ld1             {v0.16b}, [x2]
+       ld1             {v1.4s}, [x0], #16
+       cmp             w3, #10
+       bmi             0f
+       bne             3f
+       mov             v3.16b, v1.16b
+       b               2f
+0:     mov             v2.16b, v1.16b
+       ld1             {v3.4s}, [x0], #16
+1:     aesd            v0.16b, v2.16b
+       aesimc          v0.16b, v0.16b
+2:     ld1             {v1.4s}, [x0], #16
+       aesd            v0.16b, v3.16b
+       aesimc          v0.16b, v0.16b
+3:     ld1             {v2.4s}, [x0], #16
+       subs            w3, w3, #3
+       aesd            v0.16b, v1.16b
+       aesimc          v0.16b, v0.16b
+       ld1             {v3.4s}, [x0], #16
+       bpl             1b
+       aesd            v0.16b, v2.16b
+       eor             v0.16b, v0.16b, v3.16b
+       st1             {v0.16b}, [x1]
+       ret
+ENDPROC(__aes_ce_decrypt)
+
+/*
+ * __aes_ce_sub() - use the aese instruction to perform the AES sbox
+ *                  substitution on each byte in 'input'
+ */
+ENTRY(__aes_ce_sub)
+       dup             v1.4s, w0
+       movi            v0.16b, #0
+       aese            v0.16b, v1.16b
+       umov            w0, v0.s[0]
+       ret
+ENDPROC(__aes_ce_sub)
+
+ENTRY(__aes_ce_invert)
+       ld1             {v0.4s}, [x1]
+       aesimc          v1.16b, v0.16b
+       st1             {v1.4s}, [x0]
+       ret
+ENDPROC(__aes_ce_invert)
similarity index 62%
rename from arch/arm64/crypto/aes-ce-cipher.c
rename to arch/arm64/crypto/aes-ce-glue.c
index 6a75cd7..e6b3227 100644 (file)
@@ -29,6 +29,13 @@ struct aes_block {
        u8 b[AES_BLOCK_SIZE];
 };
 
+asmlinkage void __aes_ce_encrypt(u32 *rk, u8 *out, const u8 *in, int rounds);
+asmlinkage void __aes_ce_decrypt(u32 *rk, u8 *out, const u8 *in, int rounds);
+
+asmlinkage u32 __aes_ce_sub(u32 l);
+asmlinkage void __aes_ce_invert(struct aes_block *out,
+                               const struct aes_block *in);
+
 static int num_rounds(struct crypto_aes_ctx *ctx)
 {
        /*
@@ -44,10 +51,6 @@ static int num_rounds(struct crypto_aes_ctx *ctx)
 static void aes_cipher_encrypt(struct crypto_tfm *tfm, u8 dst[], u8 const src[])
 {
        struct crypto_aes_ctx *ctx = crypto_tfm_ctx(tfm);
-       struct aes_block *out = (struct aes_block *)dst;
-       struct aes_block const *in = (struct aes_block *)src;
-       void *dummy0;
-       int dummy1;
 
        if (!may_use_simd()) {
                __aes_arm64_encrypt(ctx->key_enc, dst, src, num_rounds(ctx));
@@ -55,49 +58,13 @@ static void aes_cipher_encrypt(struct crypto_tfm *tfm, u8 dst[], u8 const src[])
        }
 
        kernel_neon_begin();
-
-       __asm__("       ld1     {v0.16b}, %[in]                 ;"
-               "       ld1     {v1.4s}, [%[key]], #16          ;"
-               "       cmp     %w[rounds], #10                 ;"
-               "       bmi     0f                              ;"
-               "       bne     3f                              ;"
-               "       mov     v3.16b, v1.16b                  ;"
-               "       b       2f                              ;"
-               "0:     mov     v2.16b, v1.16b                  ;"
-               "       ld1     {v3.4s}, [%[key]], #16          ;"
-               "1:     aese    v0.16b, v2.16b                  ;"
-               "       aesmc   v0.16b, v0.16b                  ;"
-               "2:     ld1     {v1.4s}, [%[key]], #16          ;"
-               "       aese    v0.16b, v3.16b                  ;"
-               "       aesmc   v0.16b, v0.16b                  ;"
-               "3:     ld1     {v2.4s}, [%[key]], #16          ;"
-               "       subs    %w[rounds], %w[rounds], #3      ;"
-               "       aese    v0.16b, v1.16b                  ;"
-               "       aesmc   v0.16b, v0.16b                  ;"
-               "       ld1     {v3.4s}, [%[key]], #16          ;"
-               "       bpl     1b                              ;"
-               "       aese    v0.16b, v2.16b                  ;"
-               "       eor     v0.16b, v0.16b, v3.16b          ;"
-               "       st1     {v0.16b}, %[out]                ;"
-
-       :       [out]           "=Q"(*out),
-               [key]           "=r"(dummy0),
-               [rounds]        "=r"(dummy1)
-       :       [in]            "Q"(*in),
-                               "1"(ctx->key_enc),
-                               "2"(num_rounds(ctx) - 2)
-       :       "cc");
-
+       __aes_ce_encrypt(ctx->key_enc, dst, src, num_rounds(ctx));
        kernel_neon_end();
 }
 
 static void aes_cipher_decrypt(struct crypto_tfm *tfm, u8 dst[], u8 const src[])
 {
        struct crypto_aes_ctx *ctx = crypto_tfm_ctx(tfm);
-       struct aes_block *out = (struct aes_block *)dst;
-       struct aes_block const *in = (struct aes_block *)src;
-       void *dummy0;
-       int dummy1;
 
        if (!may_use_simd()) {
                __aes_arm64_decrypt(ctx->key_dec, dst, src, num_rounds(ctx));
@@ -105,62 +72,10 @@ static void aes_cipher_decrypt(struct crypto_tfm *tfm, u8 dst[], u8 const src[])
        }
 
        kernel_neon_begin();
-
-       __asm__("       ld1     {v0.16b}, %[in]                 ;"
-               "       ld1     {v1.4s}, [%[key]], #16          ;"
-               "       cmp     %w[rounds], #10                 ;"
-               "       bmi     0f                              ;"
-               "       bne     3f                              ;"
-               "       mov     v3.16b, v1.16b                  ;"
-               "       b       2f                              ;"
-               "0:     mov     v2.16b, v1.16b                  ;"
-               "       ld1     {v3.4s}, [%[key]], #16          ;"
-               "1:     aesd    v0.16b, v2.16b                  ;"
-               "       aesimc  v0.16b, v0.16b                  ;"
-               "2:     ld1     {v1.4s}, [%[key]], #16          ;"
-               "       aesd    v0.16b, v3.16b                  ;"
-               "       aesimc  v0.16b, v0.16b                  ;"
-               "3:     ld1     {v2.4s}, [%[key]], #16          ;"
-               "       subs    %w[rounds], %w[rounds], #3      ;"
-               "       aesd    v0.16b, v1.16b                  ;"
-               "       aesimc  v0.16b, v0.16b                  ;"
-               "       ld1     {v3.4s}, [%[key]], #16          ;"
-               "       bpl     1b                              ;"
-               "       aesd    v0.16b, v2.16b                  ;"
-               "       eor     v0.16b, v0.16b, v3.16b          ;"
-               "       st1     {v0.16b}, %[out]                ;"
-
-       :       [out]           "=Q"(*out),
-               [key]           "=r"(dummy0),
-               [rounds]        "=r"(dummy1)
-       :       [in]            "Q"(*in),
-                               "1"(ctx->key_dec),
-                               "2"(num_rounds(ctx) - 2)
-       :       "cc");
-
+       __aes_ce_decrypt(ctx->key_dec, dst, src, num_rounds(ctx));
        kernel_neon_end();
 }
 
-/*
- * aes_sub() - use the aese instruction to perform the AES sbox substitution
- *             on each byte in 'input'
- */
-static u32 aes_sub(u32 input)
-{
-       u32 ret;
-
-       __asm__("dup    v1.4s, %w[in]           ;"
-               "movi   v0.16b, #0              ;"
-               "aese   v0.16b, v1.16b          ;"
-               "umov   %w[out], v0.4s[0]       ;"
-
-       :       [out]   "=r"(ret)
-       :       [in]    "r"(input)
-       :               "v0","v1");
-
-       return ret;
-}
-
 int ce_aes_expandkey(struct crypto_aes_ctx *ctx, const u8 *in_key,
                     unsigned int key_len)
 {
@@ -189,7 +104,7 @@ int ce_aes_expandkey(struct crypto_aes_ctx *ctx, const u8 *in_key,
                u32 *rki = ctx->key_enc + (i * kwords);
                u32 *rko = rki + kwords;
 
-               rko[0] = ror32(aes_sub(rki[kwords - 1]), 8) ^ rcon[i] ^ rki[0];
+               rko[0] = ror32(__aes_ce_sub(rki[kwords - 1]), 8) ^ rcon[i] ^ rki[0];
                rko[1] = rko[0] ^ rki[1];
                rko[2] = rko[1] ^ rki[2];
                rko[3] = rko[2] ^ rki[3];
@@ -202,7 +117,7 @@ int ce_aes_expandkey(struct crypto_aes_ctx *ctx, const u8 *in_key,
                } else if (key_len == AES_KEYSIZE_256) {
                        if (i >= 6)
                                break;
-                       rko[4] = aes_sub(rko[3]) ^ rki[4];
+                       rko[4] = __aes_ce_sub(rko[3]) ^ rki[4];
                        rko[5] = rko[4] ^ rki[5];
                        rko[6] = rko[5] ^ rki[6];
                        rko[7] = rko[6] ^ rki[7];
@@ -221,13 +136,7 @@ int ce_aes_expandkey(struct crypto_aes_ctx *ctx, const u8 *in_key,
 
        key_dec[0] = key_enc[j];
        for (i = 1, j--; j > 0; i++, j--)
-               __asm__("ld1    {v0.4s}, %[in]          ;"
-                       "aesimc v1.16b, v0.16b          ;"
-                       "st1    {v1.4s}, %[out] ;"
-
-               :       [out]   "=Q"(key_dec[i])
-               :       [in]    "Q"(key_enc[j])
-               :               "v0","v1");
+               __aes_ce_invert(key_dec + i, key_enc + j);
        key_dec[i] = key_enc[0];
 
        kernel_neon_end();