gcm128.c: API modification and readability improvements,
[openssl.git] / crypto / modes / gcm128.c
index b21b6b3a1594718f1ab68d98689abe6abb4c2581..6f1eb6d484c300daaa8af6d9a60f354dc8a39511 100644 (file)
@@ -47,6 +47,7 @@
  * ====================================================================
  */
 
+#include "crypto.h"
 #include "modes_lcl.h"
 #include <string.h>
 
@@ -87,14 +88,34 @@ typedef struct { u64 hi,lo; } u128;
 /*
  * Even though permitted values for TABLE_BITS are 8, 4 and 1, it should
  * never be set to 8. 8 is effectively reserved for testing purposes.
- * Under ideal conditions "8-bit" version should be twice as fast as
- * "4-bit" one. For gcc-generated x86[_64] code, "8-bit" was observed to
- * run ~75% faster, closer to 100% for commercial compilers... But the
- * catch is that "8-bit" procedure consumes 16 times more memory, 4KB
- * per indivudual key + 1KB shared, and as access to these tables end up
- * on critical path, real-life execution time would be sensitive to
- * cache timing. It's not actually proven, but "4-bit" procedure is
- * believed to provide adequate all-round performance...
+ * TABLE_BITS>1 are lookup-table-driven implementations referred to as
+ * "Shoup's" in GCM specification. In other words OpenSSL does not cover
+ * whole spectrum of possible table driven implementations. Why? In
+ * non-"Shoup's" case memory access pattern is segmented in such manner,
+ * that it's trivial to see that cache timing information can reveal
+ * fair portion of intermediate hash value. Given that ciphertext is
+ * always available to attacker, it's possible for him to attempt to
+ * deduce secret parameter H and if successful, tamper with messages
+ * [which is nothing but trivial in CTR mode]. In "Shoup's" case it's
+ * not as trivial, but there is no reason to believe that it's resistant
+ * to cache-timing attack. And the thing about "8-bit" implementation is
+ * that it consumes 16 (sixteen) times more memory, 4KB per individual
+ * key + 1KB shared. Well, on pros side it should be twice as fast as
+ * "4-bit" version. And for gcc-generated x86[_64] code, "8-bit" version
+ * was observed to run ~75% faster, closer to 100% for commercial
+ * compilers... Yet "4-bit" procedure is preferred, because it's
+ * believed to provide better security-performance balance and adequate
+ * all-round performance. "All-round" refers to things like:
+ *
+ * - shorter setup time effectively improves overall timing for
+ *   handling short messages;
+ * - larger table allocation can become unbearable because of VM
+ *   subsystem penalties (for example on Windows large enough free
+ *   results in VM working set trimming, meaning that consequent
+ *   malloc would immediately incur working set expansion);
+ * - larger table has larger cache footprint, which can affect
+ *   performance of other code paths (not necessarily even from same
+ *   thread in Hyper-Threading world);
  */
 #define        TABLE_BITS 4
 
@@ -393,6 +414,7 @@ static void gcm_ghash_4bit(u64 Xi[2],const u128 Htable[16],
     size_t rem, nlo, nhi;
     const union { long one; char little; } is_endian = {1};
 
+#if 1
     do {
        cnt  = 15;
        nlo  = ((const u8 *)Xi)[15];
@@ -433,6 +455,99 @@ static void gcm_ghash_4bit(u64 Xi[2],const u128 Htable[16],
                Z.hi ^= Htable[nlo].hi;
                Z.lo ^= Htable[nlo].lo;
        }
+#else
+    /*
+     * Extra 256+16 bytes per-key plus 512 bytes shared tables
+     * [should] give ~50% improvement... One could have PACK()-ed
+     * the rem_8bit even here, but the priority is to minimize
+     * cache footprint...
+     */ 
+    u128 Hshr4[16];    /* Htable shifted right by 4 bits */
+    u8   Hshl4[16];    /* Htable shifted left  by 4 bits */
+    static const unsigned short rem_8bit[256] = {
+       0x0000, 0x01C2, 0x0384, 0x0246, 0x0708, 0x06CA, 0x048C, 0x054E,
+       0x0E10, 0x0FD2, 0x0D94, 0x0C56, 0x0918, 0x08DA, 0x0A9C, 0x0B5E,
+       0x1C20, 0x1DE2, 0x1FA4, 0x1E66, 0x1B28, 0x1AEA, 0x18AC, 0x196E,
+       0x1230, 0x13F2, 0x11B4, 0x1076, 0x1538, 0x14FA, 0x16BC, 0x177E,
+       0x3840, 0x3982, 0x3BC4, 0x3A06, 0x3F48, 0x3E8A, 0x3CCC, 0x3D0E,
+       0x3650, 0x3792, 0x35D4, 0x3416, 0x3158, 0x309A, 0x32DC, 0x331E,
+       0x2460, 0x25A2, 0x27E4, 0x2626, 0x2368, 0x22AA, 0x20EC, 0x212E,
+       0x2A70, 0x2BB2, 0x29F4, 0x2836, 0x2D78, 0x2CBA, 0x2EFC, 0x2F3E,
+       0x7080, 0x7142, 0x7304, 0x72C6, 0x7788, 0x764A, 0x740C, 0x75CE,
+       0x7E90, 0x7F52, 0x7D14, 0x7CD6, 0x7998, 0x785A, 0x7A1C, 0x7BDE,
+       0x6CA0, 0x6D62, 0x6F24, 0x6EE6, 0x6BA8, 0x6A6A, 0x682C, 0x69EE,
+       0x62B0, 0x6372, 0x6134, 0x60F6, 0x65B8, 0x647A, 0x663C, 0x67FE,
+       0x48C0, 0x4902, 0x4B44, 0x4A86, 0x4FC8, 0x4E0A, 0x4C4C, 0x4D8E,
+       0x46D0, 0x4712, 0x4554, 0x4496, 0x41D8, 0x401A, 0x425C, 0x439E,
+       0x54E0, 0x5522, 0x5764, 0x56A6, 0x53E8, 0x522A, 0x506C, 0x51AE,
+       0x5AF0, 0x5B32, 0x5974, 0x58B6, 0x5DF8, 0x5C3A, 0x5E7C, 0x5FBE,
+       0xE100, 0xE0C2, 0xE284, 0xE346, 0xE608, 0xE7CA, 0xE58C, 0xE44E,
+       0xEF10, 0xEED2, 0xEC94, 0xED56, 0xE818, 0xE9DA, 0xEB9C, 0xEA5E,
+       0xFD20, 0xFCE2, 0xFEA4, 0xFF66, 0xFA28, 0xFBEA, 0xF9AC, 0xF86E,
+       0xF330, 0xF2F2, 0xF0B4, 0xF176, 0xF438, 0xF5FA, 0xF7BC, 0xF67E,
+       0xD940, 0xD882, 0xDAC4, 0xDB06, 0xDE48, 0xDF8A, 0xDDCC, 0xDC0E,
+       0xD750, 0xD692, 0xD4D4, 0xD516, 0xD058, 0xD19A, 0xD3DC, 0xD21E,
+       0xC560, 0xC4A2, 0xC6E4, 0xC726, 0xC268, 0xC3AA, 0xC1EC, 0xC02E,
+       0xCB70, 0xCAB2, 0xC8F4, 0xC936, 0xCC78, 0xCDBA, 0xCFFC, 0xCE3E,
+       0x9180, 0x9042, 0x9204, 0x93C6, 0x9688, 0x974A, 0x950C, 0x94CE,
+       0x9F90, 0x9E52, 0x9C14, 0x9DD6, 0x9898, 0x995A, 0x9B1C, 0x9ADE,
+       0x8DA0, 0x8C62, 0x8E24, 0x8FE6, 0x8AA8, 0x8B6A, 0x892C, 0x88EE,
+       0x83B0, 0x8272, 0x8034, 0x81F6, 0x84B8, 0x857A, 0x873C, 0x86FE,
+       0xA9C0, 0xA802, 0xAA44, 0xAB86, 0xAEC8, 0xAF0A, 0xAD4C, 0xAC8E,
+       0xA7D0, 0xA612, 0xA454, 0xA596, 0xA0D8, 0xA11A, 0xA35C, 0xA29E,
+       0xB5E0, 0xB422, 0xB664, 0xB7A6, 0xB2E8, 0xB32A, 0xB16C, 0xB0AE,
+       0xBBF0, 0xBA32, 0xB874, 0xB9B6, 0xBCF8, 0xBD3A, 0xBF7C, 0xBEBE };
+    /*
+     * This pre-processing phase slows down procedure by approximately
+     * same time as it makes each loop spin faster. In other words
+     * single block performance is approximately same as straightforward
+     * "4-bit" implementation, and then it goes only faster...
+     */
+    for (cnt=0; cnt<16; ++cnt) {
+       Z.hi = Htable[cnt].hi;
+       Z.lo = Htable[cnt].lo;
+       Hshr4[cnt].lo = (Z.hi<<60)|(Z.lo>>4);
+       Hshr4[cnt].hi = (Z.hi>>4);
+       Hshl4[cnt]    = (u8)(Z.lo<<4);
+    }
+
+    do {
+       for (Z.lo=0, Z.hi=0, cnt=15; cnt; --cnt) {
+               nlo  = ((const u8 *)Xi)[cnt];
+               nlo ^= inp[cnt];
+               nhi  = nlo>>4;
+               nlo &= 0xf;
+
+               Z.hi ^= Htable[nlo].hi;
+               Z.lo ^= Htable[nlo].lo;
+
+               rem = (size_t)Z.lo&0xff;
+
+               Z.lo = (Z.hi<<56)|(Z.lo>>8);
+               Z.hi = (Z.hi>>8);
+
+               Z.hi ^= Hshr4[nhi].hi;
+               Z.lo ^= Hshr4[nhi].lo;
+               Z.hi ^= (u64)rem_8bit[rem^Hshl4[nhi]]<<48;
+       }
+
+       nlo  = ((const u8 *)Xi)[0];
+       nlo ^= inp[0];
+       nhi  = nlo>>4;
+       nlo &= 0xf;
+
+       Z.hi ^= Htable[nlo].hi;
+       Z.lo ^= Htable[nlo].lo;
+
+       rem = (size_t)Z.lo&0xf;
+
+       Z.lo = (Z.hi<<60)|(Z.lo>>4);
+       Z.hi = (Z.hi>>4);
+
+       Z.hi ^= Htable[nhi].hi;
+       Z.lo ^= Htable[nhi].lo;
+       Z.hi ^= ((u64)rem_8bit[rem<<4])<<48;
+#endif
 
        if (is_endian.little) {
 #ifdef BSWAP8
@@ -601,7 +716,7 @@ void CRYPTO_gcm128_init(GCM128_CONTEXT *ctx,void *key,block128_f block)
 #if    TABLE_BITS==8
        gcm_init_8bit(ctx->Htable,ctx->H.u);
 #elif  TABLE_BITS==4
-# if   defined(GHASH_ASM_IAX)
+# if   defined(GHASH_ASM_IAX)                  /* both x86 and x86_64 */
        if (OPENSSL_ia32cap_P[1]&(1<<1)) {
                gcm_init_clmul(ctx->Htable,ctx->H.u);
                ctx->gmult = gcm_gmult_clmul;
@@ -609,7 +724,7 @@ void CRYPTO_gcm128_init(GCM128_CONTEXT *ctx,void *key,block128_f block)
                return;
        }
        gcm_init_4bit(ctx->Htable,ctx->H.u);
-#  if  defined(GHASH_ASM_X86)
+#  if  defined(GHASH_ASM_X86)                  /* x86 only */
        if (OPENSSL_ia32cap_P[0]&(1<<23)) {
                ctx->gmult = gcm_gmult_4bit_mmx;
                ctx->ghash = gcm_ghash_4bit_mmx;
@@ -977,7 +1092,8 @@ void CRYPTO_gcm128_decrypt(GCM128_CONTEXT *ctx,
        ctx->res = n;
 }
 
-void CRYPTO_gcm128_finish(GCM128_CONTEXT *ctx)
+int CRYPTO_gcm128_finish(GCM128_CONTEXT *ctx,const unsigned char *tag,
+                       size_t len)
 {
        const union { long one; char little; } is_endian = {1};
        u64 alen = ctx->len.u[0]<<3;
@@ -1007,6 +1123,29 @@ void CRYPTO_gcm128_finish(GCM128_CONTEXT *ctx)
 
        ctx->Xi.u[0] ^= ctx->EK0.u[0];
        ctx->Xi.u[1] ^= ctx->EK0.u[1];
+
+       if (tag && len<=sizeof(ctx->Xi))
+               return memcmp(ctx->Xi.c,tag,len);
+       else
+               return -1;
+}
+
+GCM128_CONTEXT *CRYPTO_gcm128_new(void *key, block128_f block)
+{
+       GCM128_CONTEXT *ret;
+
+       if ((ret = (GCM128_CONTEXT *)OPENSSL_malloc(sizeof(GCM128_CONTEXT))))
+               CRYPTO_gcm128_init(ret,key,block);
+
+       return ret;
+}
+
+void CRYPTO_gcm128_release(GCM128_CONTEXT *ctx)
+{
+       if (ctx) {
+               OPENSSL_cleanse(ctx,sizeof(*ctx));
+               OPENSSL_free(ctx);
+       }
 }
 
 #if defined(SELFTEST)
@@ -1041,7 +1180,7 @@ static const u8   K3[]=  {0xfe,0xff,0xe9,0x92,0x86,0x65,0x73,0x1c,0x6d,0x6a,0x8f,0
                        0xe3,0xaa,0x21,0x2f,0x2c,0x02,0xa4,0xe0,0x35,0xc1,0x7e,0x23,0x29,0xac,0xa1,0x2e,
                        0x21,0xd5,0x14,0xb2,0x54,0x66,0x93,0x1c,0x7d,0x8f,0x6a,0x5a,0xac,0x84,0xaa,0x05,
                        0x1b,0xa3,0x0b,0x39,0x6a,0x0a,0xac,0x97,0x3d,0x58,0xe0,0x91,0x47,0x3f,0x59,0x85},
-               T3[]=  {0x4d,0x5c,0x2a,0xf3,0x27,0xcd,0x64,0xa6,0x2c,0xf3,0x5a,0xbd,0x2b,0xa6,0xfa,0xb4,};
+               T3[]=  {0x4d,0x5c,0x2a,0xf3,0x27,0xcd,0x64,0xa6,0x2c,0xf3,0x5a,0xbd,0x2b,0xa6,0xfa,0xb4};
 
 /* Test Case 4 */
 #define K4 K3
@@ -1233,15 +1372,15 @@ static const u8 IV18[]={0x93,0x13,0x22,0x5d,0xf8,0x84,0x06,0xe5,0x55,0x90,0x9c,0
        CRYPTO_gcm128_setiv(&ctx,IV##n,sizeof(IV##n));          \
        if (A##n) CRYPTO_gcm128_aad(&ctx,A##n,sizeof(A##n));    \
        if (P##n) CRYPTO_gcm128_encrypt(&ctx,P##n,out,sizeof(out));     \
-       CRYPTO_gcm128_finish(&ctx);                             \
-       if (memcmp(ctx.Xi.c,T##n,16) || (C##n && memcmp(out,C##n,sizeof(out)))) \
+       if (CRYPTO_gcm128_finish(&ctx,T##n,16) ||               \
+           (C##n && memcmp(out,C##n,sizeof(out))))             \
                ret++, printf ("encrypt test#%d failed.\n",n);\
        CRYPTO_gcm128_setiv(&ctx,IV##n,sizeof(IV##n));          \
        if (A##n) CRYPTO_gcm128_aad(&ctx,A##n,sizeof(A##n));    \
        if (C##n) CRYPTO_gcm128_decrypt(&ctx,C##n,out,sizeof(out));     \
-       CRYPTO_gcm128_finish(&ctx);                             \
-       if (memcmp(ctx.Xi.c,T##n,16) || (P##n && memcmp(out,P##n,sizeof(out)))) \
-               ret++, printf ("decrypt test#%d failed.\n",n);\
+       if (CRYPTO_gcm128_finish(&ctx,T##n,16) ||               \
+           (P##n && memcmp(out,P##n,sizeof(out))))             \
+               ret++, printf ("decrypt test#%d failed.\n",n);  \
        } while(0)
 
 int main()