feedc0de/wolfssl
1
0
Fork 0
forked from wolfSSL/wolfssl
Code Pull Requests Activity

1. Attempting to perform 8 AES-CBC decrypt operations simultaneously.

Browse Source 
2. Added code to test large AES-CBC decrypts.
This commit is contained in:
John Safranek
2016-04-08 11:53:40 -07:00
parent a0cd888fbf
commit b75dc127f3
3 changed files with 345 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
wolfcrypt/src/aes.c
View File

@@ -1094,10 +1094,17 @@ void AES_CBC_encrypt(const unsigned char* in, unsigned char* out,
XASM_LINK("AES_CBC_encrypt"); XASM_LINK("AES_CBC_encrypt");
#ifdef HAVE_AES_DECRYPT #ifdef HAVE_AES_DECRYPT
#ifndef HAVE_AES_DECRYPT_EX
void AES_CBC_decrypt(const unsigned char* in, unsigned char* out, void AES_CBC_decrypt(const unsigned char* in, unsigned char* out,
unsigned char* ivec, unsigned long length, unsigned char* ivec, unsigned long length,
const unsigned char* KS, int nr) const unsigned char* KS, int nr)
XASM_LINK("AES_CBC_decrypt"); XASM_LINK("AES_CBC_decrypt");
#else /* HAVE_AES_DECRYPT_EX */
void AES_CBC_decrypt_ex(const unsigned char* in, unsigned char* out,
unsigned char* ivec, unsigned long length,
const unsigned char* KS, int nr)
XASM_LINK("AES_CBC_decrypt_ex");
#endif /* HAVE_AES_DECRYPT_EX */
#endif /* HAVE_AES_DECRYPT */ #endif /* HAVE_AES_DECRYPT */
#endif /* HAVE_AES_CBC */ #endif /* HAVE_AES_CBC */
@@ -2549,8 +2556,13 @@ int wc_AesSetIV(Aes* aes, const byte* iv)
/* if input and output same will overwrite input iv */ /* if input and output same will overwrite input iv */
XMEMCPY(aes->tmp, in + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE); XMEMCPY(aes->tmp, in + sz - AES_BLOCK_SIZE, AES_BLOCK_SIZE);
#ifndef HAVE_AES_DECRYPT_EX
AES_CBC_decrypt(in, out, (byte*)aes->reg, sz, (byte*)aes->key, AES_CBC_decrypt(in, out, (byte*)aes->reg, sz, (byte*)aes->key,
aes->rounds); aes->rounds);
#else /* HAVE_AES_DECRYPT_EX */
AES_CBC_decrypt_ex(in, out, (byte*)aes->reg, sz, (byte*)aes->key,
aes->rounds);
#endif /* HAVE_AES_DECRYPT_EX */
/* store iv for next call */ /* store iv for next call */
XMEMCPY(aes->reg, aes->tmp, AES_BLOCK_SIZE); XMEMCPY(aes->reg, aes->tmp, AES_BLOCK_SIZE);
return 0; return 0;

250
wolfcrypt/src/aes_asm.s
View File

@@ -266,6 +266,256 @@ DEND_4:
ret ret
/*
AES_CBC_decrypt_ex (const unsigned char *in,
unsigned char *out,
unsigned char ivec[16],
unsigned long length,
const unsigned char *KS,
int nr)
*/
.globl AES_CBC_decrypt_ex
AES_CBC_decrypt_ex:
# parameter 1: %rdi - in
# parameter 2: %rsi - out
# parameter 3: %rdx - ivec
# parameter 4: %rcx - length
# parameter 5: %r8 - KS
# parameter 6: %r9d - nr
movq %rcx, %r10
shrq $4, %rcx
shlq $60, %r10
je ENO_PARTS_8
addq $1, %rcx
ENO_PARTS_8:
movq %rcx, %r10
shlq $61, %r10
shrq $61, %r10
shrq $3, %rcx
movdqu (%rdx), %xmm9
je EREMAINDER_8
subq $128, %rsi
ELOOP_8:
movdqu (%rdi), %xmm1
movdqu 16(%rdi), %xmm2
movdqu 32(%rdi), %xmm3
movdqu 48(%rdi), %xmm4
movdqu 64(%rdi), %xmm5
movdqu 80(%rdi), %xmm6
movdqu 96(%rdi), %xmm7
movdqu 112(%rdi), %xmm8
movdqa (%r8), %xmm10
movdqa 16(%r8), %xmm11
movdqa 32(%r8), %xmm12
movdqa 48(%r8), %xmm13
pxor %xmm10, %xmm1
pxor %xmm10, %xmm2
pxor %xmm10, %xmm3
pxor %xmm10, %xmm4
pxor %xmm10, %xmm5
pxor %xmm10, %xmm6
pxor %xmm10, %xmm7
pxor %xmm10, %xmm8
aesdec %xmm11, %xmm1
aesdec %xmm11, %xmm2
aesdec %xmm11, %xmm3
aesdec %xmm11, %xmm4
aesdec %xmm11, %xmm5
aesdec %xmm11, %xmm6
aesdec %xmm11, %xmm7
aesdec %xmm11, %xmm8
aesdec %xmm12, %xmm1
aesdec %xmm12, %xmm2
aesdec %xmm12, %xmm3
aesdec %xmm12, %xmm4
aesdec %xmm12, %xmm5
aesdec %xmm12, %xmm6
aesdec %xmm12, %xmm7
aesdec %xmm12, %xmm8
aesdec %xmm13, %xmm1
aesdec %xmm13, %xmm2
aesdec %xmm13, %xmm3
aesdec %xmm13, %xmm4
aesdec %xmm13, %xmm5
aesdec %xmm13, %xmm6
aesdec %xmm13, %xmm7
aesdec %xmm13, %xmm8
movdqa 64(%r8), %xmm10
movdqa 80(%r8), %xmm11
movdqa 96(%r8), %xmm12
movdqa 112(%r8), %xmm13
aesdec %xmm10, %xmm1
aesdec %xmm10, %xmm2
aesdec %xmm10, %xmm3
aesdec %xmm10, %xmm4
aesdec %xmm10, %xmm5
aesdec %xmm10, %xmm6
aesdec %xmm10, %xmm7
aesdec %xmm10, %xmm8
aesdec %xmm11, %xmm1
aesdec %xmm11, %xmm2
aesdec %xmm11, %xmm3
aesdec %xmm11, %xmm4
aesdec %xmm11, %xmm5
aesdec %xmm11, %xmm6
aesdec %xmm11, %xmm7
aesdec %xmm11, %xmm8
aesdec %xmm12, %xmm1
aesdec %xmm12, %xmm2
aesdec %xmm12, %xmm3
aesdec %xmm12, %xmm4
aesdec %xmm12, %xmm5
aesdec %xmm12, %xmm6
aesdec %xmm12, %xmm7
aesdec %xmm12, %xmm8
aesdec %xmm13, %xmm1
aesdec %xmm13, %xmm2
aesdec %xmm13, %xmm3
aesdec %xmm13, %xmm4
aesdec %xmm13, %xmm5
aesdec %xmm13, %xmm6
aesdec %xmm13, %xmm7
aesdec %xmm13, %xmm8
movdqa 128(%r8), %xmm10
movdqa 144(%r8), %xmm11
movdqa 160(%r8), %xmm12
cmpl $12, %r9d
aesdec %xmm10, %xmm1
aesdec %xmm10, %xmm2
aesdec %xmm10, %xmm3
aesdec %xmm10, %xmm4
aesdec %xmm10, %xmm5
aesdec %xmm10, %xmm6
aesdec %xmm10, %xmm7
aesdec %xmm10, %xmm8
aesdec %xmm11, %xmm1
aesdec %xmm11, %xmm2
aesdec %xmm11, %xmm3
aesdec %xmm11, %xmm4
aesdec %xmm11, %xmm5
aesdec %xmm11, %xmm6
aesdec %xmm11, %xmm7
aesdec %xmm11, %xmm8
jb ELAST_8
movdqa 160(%r8), %xmm10
movdqa 176(%r8), %xmm11
movdqa 192(%r8), %xmm12
cmpl $14, %r9d
aesdec %xmm10, %xmm1
aesdec %xmm10, %xmm2
aesdec %xmm10, %xmm3
aesdec %xmm10, %xmm4
aesdec %xmm10, %xmm5
aesdec %xmm10, %xmm6
aesdec %xmm10, %xmm7
aesdec %xmm10, %xmm8
aesdec %xmm11, %xmm1
aesdec %xmm11, %xmm2
aesdec %xmm11, %xmm3
aesdec %xmm11, %xmm4
aesdec %xmm11, %xmm5
aesdec %xmm11, %xmm6
aesdec %xmm11, %xmm7
aesdec %xmm11, %xmm8
jb ELAST_8
movdqa 192(%r8), %xmm10
movdqa 208(%r8), %xmm11
movdqa 224(%r8), %xmm12
aesdec %xmm10, %xmm1
aesdec %xmm10, %xmm2
aesdec %xmm10, %xmm3
aesdec %xmm10, %xmm4
aesdec %xmm10, %xmm5
aesdec %xmm10, %xmm6
aesdec %xmm10, %xmm7
aesdec %xmm10, %xmm8
aesdec %xmm11, %xmm1
aesdec %xmm11, %xmm2
aesdec %xmm11, %xmm3
aesdec %xmm11, %xmm4
aesdec %xmm11, %xmm5
aesdec %xmm11, %xmm6
aesdec %xmm11, %xmm7
aesdec %xmm11, %xmm8
ELAST_8:
addq $128, %rsi
aesdeclast %xmm12, %xmm1
aesdeclast %xmm12, %xmm2
aesdeclast %xmm12, %xmm3
aesdeclast %xmm12, %xmm4
aesdeclast %xmm12, %xmm5
aesdeclast %xmm12, %xmm6
aesdeclast %xmm12, %xmm7
aesdeclast %xmm12, %xmm8
movdqu (%rdi), %xmm10
movdqu 16(%rdi), %xmm11
movdqu 32(%rdi), %xmm12
movdqu 48(%rdi), %xmm13
pxor %xmm9, %xmm1
pxor %xmm10, %xmm2
pxor %xmm11, %xmm3
pxor %xmm12, %xmm4
pxor %xmm13, %xmm5
movdqu 64(%rdi), %xmm10
movdqu 80(%rdi), %xmm11
movdqu 96(%rdi), %xmm12
movdqu 112(%rdi), %xmm9
pxor %xmm10, %xmm6
pxor %xmm11, %xmm7
pxor %xmm12, %xmm8
movdqu %xmm1, (%rsi)
movdqu %xmm2, 16(%rsi)
movdqu %xmm3, 32(%rsi)
movdqu %xmm4, 48(%rsi)
movdqu %xmm5, 64(%rsi)
movdqu %xmm6, 80(%rsi)
movdqu %xmm7, 96(%rsi)
movdqu %xmm8, 112(%rsi)
addq $128, %rdi
decq %rcx
jne ELOOP_8
addq $128, %rsi
EREMAINDER_8:
cmpq $0, %r10
je EEND_8
ELOOP_8_2:
movdqu (%rdi), %xmm1
movdqa %xmm1 ,%xmm10
addq $16, %rdi
pxor (%r8), %xmm1
movdqu 160(%r8), %xmm2
cmpl $12, %r9d
aesdec 16(%r8), %xmm1
aesdec 32(%r8), %xmm1
aesdec 48(%r8), %xmm1
aesdec 64(%r8), %xmm1
aesdec 80(%r8), %xmm1
aesdec 96(%r8), %xmm1
aesdec 112(%r8), %xmm1
aesdec 128(%r8), %xmm1
aesdec 144(%r8), %xmm1
jb ELAST_8_2
movdqu 192(%r8), %xmm2
cmpl $14, %r9d
aesdec 160(%r8), %xmm1
aesdec 176(%r8), %xmm1
jb ELAST_8_2
movdqu 224(%r8), %xmm2
aesdec 192(%r8), %xmm1
aesdec 208(%r8), %xmm1
ELAST_8_2:
aesdeclast %xmm2, %xmm1
pxor %xmm9, %xmm1
movdqa %xmm10, %xmm9
movdqu %xmm1, (%rsi)
addq $16, %rsi
decq %r10
jne ELOOP_8_2
EEND_8:
ret
/* /*
AES_ECB_encrypt (const unsigned char *in, AES_ECB_encrypt (const unsigned char *in,
unsigned char *out, unsigned char *out,

83
wolfcrypt/test/test.c
View File

@@ -2669,6 +2669,89 @@ int aes_test(void)
if (memcmp(cipher, verify, AES_BLOCK_SIZE)) if (memcmp(cipher, verify, AES_BLOCK_SIZE))
return -61; return -61;
#if defined(WOLFSSL_AESNI) && \
defined(HAVE_AES_DECRYPT) && defined(HAVE_AES_DECRYPT_EX)
{
const byte bigMsg[] = {
/* "All work and no play makes Jack a dull boy. " */
0x41,0x6c,0x6c,0x20,0x77,0x6f,0x72,0x6b,
0x20,0x61,0x6e,0x64,0x20,0x6e,0x6f,0x20,
0x70,0x6c,0x61,0x79,0x20,0x6d,0x61,0x6b,
0x65,0x73,0x20,0x4a,0x61,0x63,0x6b,0x20,
0x61,0x20,0x64,0x75,0x6c,0x6c,0x20,0x62,
0x6f,0x79,0x2e,0x20,0x41,0x6c,0x6c,0x20,
0x77,0x6f,0x72,0x6b,0x20,0x61,0x6e,0x64,
0x20,0x6e,0x6f,0x20,0x70,0x6c,0x61,0x79,
0x20,0x6d,0x61,0x6b,0x65,0x73,0x20,0x4a,
0x61,0x63,0x6b,0x20,0x61,0x20,0x64,0x75,
0x6c,0x6c,0x20,0x62,0x6f,0x79,0x2e,0x20,
0x41,0x6c,0x6c,0x20,0x77,0x6f,0x72,0x6b,
0x20,0x61,0x6e,0x64,0x20,0x6e,0x6f,0x20,
0x70,0x6c,0x61,0x79,0x20,0x6d,0x61,0x6b,
0x65,0x73,0x20,0x4a,0x61,0x63,0x6b,0x20,
0x61,0x20,0x64,0x75,0x6c,0x6c,0x20,0x62,
0x6f,0x79,0x2e,0x20,0x41,0x6c,0x6c,0x20,
0x77,0x6f,0x72,0x6b,0x20,0x61,0x6e,0x64,
0x20,0x6e,0x6f,0x20,0x70,0x6c,0x61,0x79,
0x20,0x6d,0x61,0x6b,0x65,0x73,0x20,0x4a,
0x61,0x63,0x6b,0x20,0x61,0x20,0x64,0x75,
0x6c,0x6c,0x20,0x62,0x6f,0x79,0x2e,0x20,
0x41,0x6c,0x6c,0x20,0x77,0x6f,0x72,0x6b,
0x20,0x61,0x6e,0x64,0x20,0x6e,0x6f,0x20,
0x70,0x6c,0x61,0x79,0x20,0x6d,0x61,0x6b,
0x65,0x73,0x20,0x4a,0x61,0x63,0x6b,0x20,
0x61,0x20,0x64,0x75,0x6c,0x6c,0x20,0x62,
0x6f,0x79,0x2e,0x20,0x41,0x6c,0x6c,0x20,
0x77,0x6f,0x72,0x6b,0x20,0x61,0x6e,0x64,
0x20,0x6e,0x6f,0x20,0x70,0x6c,0x61,0x79,
0x20,0x6d,0x61,0x6b,0x65,0x73,0x20,0x4a,
0x61,0x63,0x6b,0x20,0x61,0x20,0x64,0x75,
0x6c,0x6c,0x20,0x62,0x6f,0x79,0x2e,0x20,
0x41,0x6c,0x6c,0x20,0x77,0x6f,0x72,0x6b,
0x20,0x61,0x6e,0x64,0x20,0x6e,0x6f,0x20,
0x70,0x6c,0x61,0x79,0x20,0x6d,0x61,0x6b,
0x65,0x73,0x20,0x4a,0x61,0x63,0x6b,0x20,
0x61,0x20,0x64,0x75,0x6c,0x6c,0x20,0x62,
0x6f,0x79,0x2e,0x20,0x41,0x6c,0x6c,0x20,
0x77,0x6f,0x72,0x6b,0x20,0x61,0x6e,0x64,
0x20,0x6e,0x6f,0x20,0x70,0x6c,0x61,0x79,
0x20,0x6d,0x61,0x6b,0x65,0x73,0x20,0x4a,
0x61,0x63,0x6b,0x20,0x61,0x20,0x64,0x75,
0x6c,0x6c,0x20,0x62,0x6f,0x79,0x2e,0x20,
0x41,0x6c,0x6c,0x20,0x77,0x6f,0x72,0x6b,
0x20,0x61,0x6e,0x64,0x20,0x6e,0x6f,0x20,
0x70,0x6c,0x61,0x79,0x20,0x6d,0x61,0x6b,
0x65,0x73,0x20,0x4a,0x61,0x63,0x6b,0x20
};
byte bigCipher[sizeof(bigMsg)];
byte bigPlain[sizeof(bigMsg)];
ret = wc_AesSetKey(&enc, key, AES_BLOCK_SIZE, iv, AES_ENCRYPTION);
if (ret != 0)
return -1030;
ret = wc_AesSetKey(&dec, key, AES_BLOCK_SIZE, iv, AES_DECRYPTION);
if (ret != 0)
return -1031;
#define AESNI_DECRYPT_SIZE (AES_BLOCK_SIZE*24)
if ((sizeof(bigMsg) < AESNI_DECRYPT_SIZE) ||
(AESNI_DECRYPT_SIZE == 0) ||
(AESNI_DECRYPT_SIZE % AES_BLOCK_SIZE != 0))
return -1032;
ret = wc_AesCbcEncrypt(&enc, bigCipher, bigMsg, AESNI_DECRYPT_SIZE);
if (ret != 0)
return -1033;
ret = wc_AesCbcDecrypt(&dec, bigPlain, bigCipher, AESNI_DECRYPT_SIZE);
if (ret != 0)
return -1034;
if (memcmp(bigPlain, bigMsg, AESNI_DECRYPT_SIZE))
return -1035;
}
#endif /* WOLFSSL_AESNI HAVE_AES_DECRYPT HAVE_AES_DECRYPT_EX */
#ifdef HAVE_CAVIUM #ifdef HAVE_CAVIUM
wc_AesFreeCavium(&enc); wc_AesFreeCavium(&enc);
wc_AesFreeCavium(&dec); wc_AesFreeCavium(&dec);