#include <openssl/objects.h>
#include <openssl/aes.h>
#include <openssl/sha.h>
+#include <openssl/rand.h>
+#include "modes_lcl.h"
#ifndef EVP_CIPH_FLAG_AEAD_CIPHER
#define EVP_CIPH_FLAG_AEAD_CIPHER 0x200000
#define EVP_CIPH_FLAG_DEFAULT_ASN1 0
#endif
+#if !defined(EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK)
+#define EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK 0
+#endif
+
#define TLS1_1_VERSION 0x0302
typedef struct
defined(_M_AMD64) || defined(_M_X64) || \
defined(__INTEL__) )
-#if defined(__GNUC__) && __GNUC__>=2 && !defined(PEDANTIC)
-# define BSWAP(x) ({ unsigned int r=(x); asm ("bswapl %0":"=r"(r):"0"(r)); r; })
-#endif
-
-extern unsigned int OPENSSL_ia32cap_P[2];
+extern unsigned int OPENSSL_ia32cap_P[3];
#define AESNI_CAPABLE (1<<(57-32))
int aesni_set_encrypt_key(const unsigned char *userKey, int bits,
#endif
#define SHA1_Update sha1_update
+#if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
+
+typedef struct { unsigned int A[8],B[8],C[8],D[8],E[8]; } SHA1_MB_CTX;
+typedef struct { const unsigned char *ptr; int blocks; } HASH_DESC;
+
+void sha1_multi_block(SHA1_MB_CTX *,const HASH_DESC *,int);
+
+typedef struct { const unsigned char *inp; unsigned char *out;
+ int blocks; u64 iv[2]; } CIPH_DESC;
+
+void aesni_multi_cbc_encrypt(CIPH_DESC *,void *,int);
+
+static size_t tls1_1_multi_block_encrypt(EVP_AES_HMAC_SHA1 *key,
+ unsigned char *out, const unsigned char *inp, size_t inp_len,
+ int n4x) /* n4x is 1 or 2 */
+{
+ HASH_DESC hash_d[8], edges[8];
+ CIPH_DESC ciph_d[8];
+ unsigned char storage[sizeof(SHA1_MB_CTX)+32];
+ union { u64 q[16];
+ u32 d[32];
+ u8 c[128]; } blocks[8];
+ SHA1_MB_CTX *ctx;
+ unsigned int frag, last, packlen, i, x4=4*n4x;
+ size_t ret = 0;
+ u8 *IVs;
+
+ ctx = (SHA1_MB_CTX *)(storage+32-((size_t)storage%32)); /* align */
+
+ frag = (unsigned int)inp_len>>(1+n4x);
+ last = (unsigned int)inp_len+frag-(frag<<(1+n4x));
+ if (last>frag && ((last+13+9)%64)<(x4-1)) {
+ frag++;
+ last -= x4-1;
+ }
+
+ hash_d[0].ptr = inp;
+ for (i=1;i<x4;i++) hash_d[i].ptr = hash_d[i-1].ptr+frag;
+
+ for (i=0;i<x4;i++) {
+ unsigned int len = (i==(x4-1)?last:frag);
+
+ ctx->A[i] = key->md.h0;
+ ctx->B[i] = key->md.h1;
+ ctx->C[i] = key->md.h2;
+ ctx->D[i] = key->md.h3;
+ ctx->E[i] = key->md.h4;
+
+ /* fix seqnum */
+#if defined(BSWAP8)
+ blocks[i].q[0] = BSWAP8(BSWAP8(*(u64*)key->md.data)+i);
+#else
+ blocks[i].c[7] += ((u8*)key->md.data)[7]+i;
+ if (blocks[i].c[7] < i) {
+ int j;
+
+ for (j=6;j>=0;j--) {
+ if (blocks[i].c[j]=((u8*)key->md.data)[j]+1) break;
+ }
+ }
+#endif
+ blocks[i].c[8] = ((u8*)key->md.data)[8];
+ blocks[i].c[9] = ((u8*)key->md.data)[9];
+ blocks[i].c[10] = ((u8*)key->md.data)[10];
+ /* fix length */
+ blocks[i].c[11] = (u8)(len>>8);
+ blocks[i].c[12] = (u8)(len);
+
+ memcpy(blocks[i].c+13,hash_d[i].ptr,64-13);
+ hash_d[i].ptr += 64-13;
+ hash_d[i].blocks = (len-(64-13))/64;
+
+ edges[i].ptr = blocks[i].c;
+ edges[i].blocks = 1;
+ }
+
+ /* hash 13-byte headers and first 64-13 bytes of inputs */
+ sha1_multi_block(ctx,edges,n4x);
+ /* hash bulk inputs */
+ sha1_multi_block(ctx,hash_d,n4x);
+
+ memset(blocks,0,sizeof(blocks));
+ for (i=0;i<x4;i++) {
+ unsigned int len = (i==(x4-1)?last:frag),
+ off = hash_d[i].blocks*64;
+ const unsigned char *ptr = hash_d[i].ptr+off;
+
+ off = len-(64-13)-off; /* remainder actually */
+ memcpy(blocks[i].c,ptr,off);
+ blocks[i].c[off]=0x80;
+ len += 64+13; /* 64 is HMAC header */
+ len *= 8; /* convert to bits */
+ if (off<(64-8)) {
+ blocks[i].d[15] = BSWAP4(len);
+ edges[i].blocks = 1;
+ } else {
+ blocks[i].d[31] = BSWAP4(len);
+ edges[i].blocks = 2;
+ }
+ edges[i].ptr = blocks[i].c;
+ }
+
+ /* hash input tails and finalize */
+ sha1_multi_block(ctx,edges,n4x);
+
+ memset(blocks,0,sizeof(blocks));
+ for (i=0;i<x4;i++) {
+ blocks[i].d[0] = BSWAP4(ctx->A[i]); ctx->A[i] = key->tail.h0;
+ blocks[i].d[1] = BSWAP4(ctx->B[i]); ctx->B[i] = key->tail.h1;
+ blocks[i].d[2] = BSWAP4(ctx->C[i]); ctx->C[i] = key->tail.h2;
+ blocks[i].d[3] = BSWAP4(ctx->D[i]); ctx->D[i] = key->tail.h3;
+ blocks[i].d[4] = BSWAP4(ctx->E[i]); ctx->E[i] = key->tail.h4;
+ blocks[i].c[20] = 0x80;
+ blocks[i].d[15] = BSWAP4((64+20)*8);
+ edges[i].ptr = blocks[i].c;
+ edges[i].blocks = 1;
+ }
+
+ /* finalize MACs */
+ sha1_multi_block(ctx,edges,n4x);
+
+ packlen = 5+16+((frag+20+16)&-16);
+
+ out += (packlen<<(1+n4x))-packlen;
+ inp += (frag<<(1+n4x))-frag;
+
+ RAND_bytes((IVs=blocks[0].c),16*x4); /* ask for IVs in bulk */
+
+ for (i=x4-1;;i--) {
+ unsigned int len = (i==(x4-1)?last:frag), pad, j;
+ unsigned char *out0 = out;
+
+ out += 5+16; /* place for header and explicit IV */
+ ciph_d[i].inp = out;
+ ciph_d[i].out = out;
+
+ memmove(out,inp,len);
+ out += len;
+
+ /* write MAC */
+ ((u32 *)out)[0] = BSWAP4(ctx->A[i]);
+ ((u32 *)out)[1] = BSWAP4(ctx->B[i]);
+ ((u32 *)out)[2] = BSWAP4(ctx->C[i]);
+ ((u32 *)out)[3] = BSWAP4(ctx->D[i]);
+ ((u32 *)out)[4] = BSWAP4(ctx->E[i]);
+ out += 20;
+ len += 20;
+
+ /* pad */
+ pad = 15-len%16;
+ for (j=0;j<=pad;j++) *(out++) = pad;
+ len += pad+1;
+
+ ciph_d[i].blocks = len/16;
+ len += 16; /* account for explicit iv */
+
+ /* arrange header */
+ out0[0] = ((u8*)key->md.data)[8];
+ out0[1] = ((u8*)key->md.data)[9];
+ out0[2] = ((u8*)key->md.data)[10];
+ out0[3] = (u8)(len>>8);
+ out0[4] = (u8)(len);
+
+ /* explicit iv */
+ memcpy(ciph_d[i].iv, IVs, 16);
+ memcpy(&out0[5], IVs, 16);
+
+ ret += len+5;
+
+ if (i==0) break;
+
+ out = out0-packlen;
+ inp -= frag;
+ IVs += 16;
+ }
+
+ aesni_multi_cbc_encrypt(ciph_d,&key->ks,n4x);
+
+ OPENSSL_cleanse(blocks,sizeof(blocks));
+ OPENSSL_cleanse(ctx,sizeof(*ctx));
+
+ return ret;
+}
+#endif
+
static int aesni_cbc_hmac_sha1_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
const unsigned char *in, size_t len)
{
}
} else {
union { unsigned int u[SHA_DIGEST_LENGTH/sizeof(unsigned int)];
- unsigned char c[SHA_DIGEST_LENGTH]; } mac;
+ unsigned char c[32+SHA_DIGEST_LENGTH]; } mac, *pmac;
+
+ /* arrange cache line alignment */
+ pmac = (void *)(((size_t)mac.c+31)&((size_t)0-32));
/* decrypt HMAC|padding at once */
aesni_cbc_encrypt(in,out,len,
/* but pretend as if we hashed padded payload */
bitlen = key->md.Nl+(inp_len<<3); /* at most 18 bits */
+#ifdef BSWAP4
+ bitlen = BSWAP4(bitlen);
+#else
mac.c[0] = 0;
mac.c[1] = (unsigned char)(bitlen>>16);
mac.c[2] = (unsigned char)(bitlen>>8);
mac.c[3] = (unsigned char)bitlen;
bitlen = mac.u[0];
+#endif
- mac.u[0]=0;
- mac.u[1]=0;
- mac.u[2]=0;
- mac.u[3]=0;
- mac.u[4]=0;
+ pmac->u[0]=0;
+ pmac->u[1]=0;
+ pmac->u[2]=0;
+ pmac->u[3]=0;
+ pmac->u[4]=0;
for (res=key->md.num, j=0;j<len;j++) {
size_t c = out[j];
if (res!=SHA_CBLOCK) continue;
- mask = 0-((inp_len+8-j)>>(sizeof(j)*8-1));
+ /* j is not incremented yet */
+ mask = 0-((inp_len+7-j)>>(sizeof(j)*8-1));
data->u[SHA_LBLOCK-1] |= bitlen&mask;
sha1_block_data_order(&key->md,data,1);
- mask &= 0-((j-inp_len-73)>>(sizeof(j)*8-1));
- mac.u[0] |= key->md.h0 & mask;
- mac.u[1] |= key->md.h1 & mask;
- mac.u[2] |= key->md.h2 & mask;
- mac.u[3] |= key->md.h3 & mask;
- mac.u[4] |= key->md.h4 & mask;
+ mask &= 0-((j-inp_len-72)>>(sizeof(j)*8-1));
+ pmac->u[0] |= key->md.h0 & mask;
+ pmac->u[1] |= key->md.h1 & mask;
+ pmac->u[2] |= key->md.h2 & mask;
+ pmac->u[3] |= key->md.h3 & mask;
+ pmac->u[4] |= key->md.h4 & mask;
res=0;
}
data->u[SHA_LBLOCK-1] |= bitlen&mask;
sha1_block_data_order(&key->md,data,1);
mask &= 0-((j-inp_len-73)>>(sizeof(j)*8-1));
- mac.u[0] |= key->md.h0 & mask;
- mac.u[1] |= key->md.h1 & mask;
- mac.u[2] |= key->md.h2 & mask;
- mac.u[3] |= key->md.h3 & mask;
- mac.u[4] |= key->md.h4 & mask;
+ pmac->u[0] |= key->md.h0 & mask;
+ pmac->u[1] |= key->md.h1 & mask;
+ pmac->u[2] |= key->md.h2 & mask;
+ pmac->u[3] |= key->md.h3 & mask;
+ pmac->u[4] |= key->md.h4 & mask;
memset(data,0,SHA_CBLOCK);
j+=64;
data->u[SHA_LBLOCK-1] = bitlen;
sha1_block_data_order(&key->md,data,1);
mask = 0-((j-inp_len-73)>>(sizeof(j)*8-1));
- mac.u[0] |= key->md.h0 & mask;
- mac.u[1] |= key->md.h1 & mask;
- mac.u[2] |= key->md.h2 & mask;
- mac.u[3] |= key->md.h3 & mask;
- mac.u[4] |= key->md.h4 & mask;
-
-#ifdef BSWAP
- mac.u[0] = BSWAP(mac.u[0]);
- mac.u[1] = BSWAP(mac.u[1]);
- mac.u[2] = BSWAP(mac.u[2]);
- mac.u[3] = BSWAP(mac.u[3]);
- mac.u[4] = BSWAP(mac.u[4]);
+ pmac->u[0] |= key->md.h0 & mask;
+ pmac->u[1] |= key->md.h1 & mask;
+ pmac->u[2] |= key->md.h2 & mask;
+ pmac->u[3] |= key->md.h3 & mask;
+ pmac->u[4] |= key->md.h4 & mask;
+
+#ifdef BSWAP4
+ pmac->u[0] = BSWAP4(pmac->u[0]);
+ pmac->u[1] = BSWAP4(pmac->u[1]);
+ pmac->u[2] = BSWAP4(pmac->u[2]);
+ pmac->u[3] = BSWAP4(pmac->u[3]);
+ pmac->u[4] = BSWAP4(pmac->u[4]);
#else
for (i=0;i<5;i++) {
- res = mac.u[i];
- mac.c[4*i+0]=(unsigned char)(res>>24);
- mac.c[4*i+1]=(unsigned char)(res>>16);
- mac.c[4*i+2]=(unsigned char)(res>>8);
- mac.c[4*i+3]=(unsigned char)res;
+ res = pmac->u[i];
+ pmac->c[4*i+0]=(unsigned char)(res>>24);
+ pmac->c[4*i+1]=(unsigned char)(res>>16);
+ pmac->c[4*i+2]=(unsigned char)(res>>8);
+ pmac->c[4*i+3]=(unsigned char)res;
}
#endif
len += SHA_DIGEST_LENGTH;
#else
SHA1_Update(&key->md,out,inp_len);
res = key->md.num;
- SHA1_Final(mac.c,&key->md);
+ SHA1_Final(pmac->c,&key->md);
{
unsigned int inp_blocks, pad_blocks;
}
#endif
key->md = key->tail;
- SHA1_Update(&key->md,mac.c,SHA_DIGEST_LENGTH);
- SHA1_Final(mac.c,&key->md);
+ SHA1_Update(&key->md,pmac->c,SHA_DIGEST_LENGTH);
+ SHA1_Final(pmac->c,&key->md);
/* verify HMAC */
out += inp_len;
cmask = ((int)(j-off-SHA_DIGEST_LENGTH))>>(sizeof(int)*8-1);
res |= (c^pad)&~cmask; /* ... and padding */
cmask &= ((int)(off-1-j))>>(sizeof(int)*8-1);
- res |= (c^mac.c[i])&cmask;
+ res |= (c^pmac->c[i])&cmask;
i += 1&cmask;
}
maxpad -= SHA_DIGEST_LENGTH;
}
#else
for (res=0,i=0;i<SHA_DIGEST_LENGTH;i++)
- res |= out[i]^mac.c[i];
+ res |= out[i]^pmac->c[i];
res = 0-((0-res)>>(sizeof(res)*8-1));
ret &= (int)~res;
return SHA_DIGEST_LENGTH;
}
}
+#if !defined(OPENSSL_NO_MULTIBLOCK) && EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK
+ case EVP_CTRL_TLS1_1_MULTIBLOCK_AAD:
+ {
+ EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
+ (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
+ unsigned int n4x=1, x4;
+ unsigned int frag, last, packlen, inp_len;
+
+ if (arg<sizeof(EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM)) return -1;
+
+ inp_len = param->inp[11]<<8|param->inp[12];
+
+ if (ctx->encrypt)
+ {
+ if ((param->inp[9]<<8|param->inp[10]) < TLS1_1_VERSION)
+ return -1;
+
+ if (inp_len<4096) return 0; /* too short */
+
+ if (inp_len>=8192 && OPENSSL_ia32cap_P[2]&(1<<5))
+ n4x=2; /* AVX2 */
+
+ key->md = key->head;
+ SHA1_Update(&key->md,param->inp,13);
+
+ x4 = 4*n4x; n4x += 1;
+
+ frag = inp_len>>n4x;
+ last = inp_len+frag-(frag<<n4x);
+ if (last>frag && ((last+13+9)%64<(x4-1))) {
+ frag++;
+ last -= x4-1;
+ }
+
+ packlen = 5+16+((frag+20+16)&-16);
+ packlen = (packlen<<n4x)-packlen;
+ packlen += 5+16+((last+20+16)&-16);
+
+ param->interleave = x4;
+
+ return (int)packlen;
+ }
+ else
+ return -1; /* not yet */
+ }
+ case EVP_CTRL_TLS1_1_MULTIBLOCK_ENCRYPT:
+ {
+ EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *param =
+ (EVP_CTRL_TLS1_1_MULTIBLOCK_PARAM *)ptr;
+
+ return (int)tls1_1_multi_block_encrypt(key,param->out,param->inp,
+ param->len,param->interleave/4);
+ }
+ case EVP_CTRL_TLS1_1_MULTIBLOCK_DECRYPT:
+#endif
default:
return -1;
}
NID_undef,
#endif
16,16,16,
- EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|EVP_CIPH_FLAG_AEAD_CIPHER,
+ EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
+ EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
aesni_cbc_hmac_sha1_init_key,
aesni_cbc_hmac_sha1_cipher,
NULL,
NID_undef,
#endif
16,32,16,
- EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|EVP_CIPH_FLAG_AEAD_CIPHER,
+ EVP_CIPH_CBC_MODE|EVP_CIPH_FLAG_DEFAULT_ASN1|
+ EVP_CIPH_FLAG_AEAD_CIPHER|EVP_CIPH_FLAG_TLS1_1_MULTIBLOCK,
aesni_cbc_hmac_sha1_init_key,
aesni_cbc_hmac_sha1_cipher,
NULL,
projects / openssl.git / blobdiff