Unified · Split
| @@ -89,6 +89,7 @@ enum · 88 unmodified lines | |||
| 89 | ChachaBsize= 64, | 89 | ChachaBsize= 64, |
| 90 | ChachaKeylen= 256/8, | 90 | ChachaKeylen= 256/8, |
| 91 | ChachaIVlen= 96/8, | 91 | ChachaIVlen= 96/8, |
| 92 | XChachaIVlen= 192/8, | ||
| 92 | }; | 93 | }; |
| 93 | 94 | ||
| 94 | typedef struct Chachastate Chachastate; | 95 | typedef struct Chachastate Chachastate; |
| @@ -103,6 +104,7 @@ struct Chachastate · 8 unmodified lines | |||
| 103 | u32int iv[3]; | 104 | u32int iv[3]; |
| 104 | }; | 105 | }; |
| 105 | }; | 106 | }; |
| 107 | u32int xkey[8]; | ||
| 106 | int rounds; | 108 | int rounds; |
| 107 | int ivwords; | 109 | int ivwords; |
| 108 | }; | 110 | }; |
| @@ -113,6 +115,8 @@ void chacha_setblock(Chachastate*, u64int); · 4 unmodified lines | |||
| 113 | void chacha_encrypt(uchar*, ulong, Chachastate*); | 115 | void chacha_encrypt(uchar*, ulong, Chachastate*); |
| 114 | void chacha_encrypt2(uchar*, uchar*, ulong, Chachastate*); | 116 | void chacha_encrypt2(uchar*, uchar*, ulong, Chachastate*); |
| 115 | 117 | ||
| 118 | void hchacha(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds); | ||
| 119 | |||
| 116 | void ccpoly_encrypt(uchar *dat, ulong ndat, uchar *aad, ulong naad, uchar tag[16], Chachastate *cs); | 120 | void ccpoly_encrypt(uchar *dat, ulong ndat, uchar *aad, ulong naad, uchar tag[16], Chachastate *cs); |
| 117 | int ccpoly_decrypt(uchar *dat, ulong ndat, uchar *aad, ulong naad, uchar tag[16], Chachastate *cs); | 121 | int ccpoly_decrypt(uchar *dat, ulong ndat, uchar *aad, ulong naad, uchar tag[16], Chachastate *cs); |
| 118 | 122 | ||
| @@ -131,7 +135,7 @@ typedef struct Salsastate Salsastate; · 12 unmodified lines | |||
| 131 | struct Salsastate | 135 | struct Salsastate |
| 132 | { | 136 | { |
| 133 | u32int input[16]; | 137 | u32int input[16]; |
| 134 | u32int key[8]; | 138 | u32int xkey[8]; |
| 135 | int rounds; | 139 | int rounds; |
| 136 | int ivwords; | 140 | int ivwords; |
| 137 | }; | 141 | }; |
| @@ -142,6 +146,8 @@ void salsa_setblock(Salsastate*, u64int); · 4 unmodified lines | |||
| 142 | void salsa_encrypt(uchar*, ulong, Salsastate*); | 146 | void salsa_encrypt(uchar*, ulong, Salsastate*); |
| 143 | void salsa_encrypt2(uchar*, uchar*, ulong, Salsastate*); | 147 | void salsa_encrypt2(uchar*, uchar*, ulong, Salsastate*); |
| 144 | 148 | ||
| 149 | void salsa_core(u32int in[16], u32int out[16], int rounds); | ||
| 150 | |||
| 145 | void hsalsa(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds); | 151 | void hsalsa(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds); |
| 146 | 152 | ||
| 147 | /* | 153 | /* |
| @@ -485,7 +491,7 @@ typedef struct ECpoint{ · 337 unmodified lines | |||
| 485 | int inf; | 491 | int inf; |
| 486 | mpint *x; | 492 | mpint *x; |
| 487 | mpint *y; | 493 | mpint *y; |
| 488 | mpint *z; | 494 | mpint *z; /* nil when using affine coordinates */ |
| 489 | } ECpoint; | 495 | } ECpoint; |
| 490 | 496 | ||
| 491 | typedef ECpoint ECpub; | 497 | typedef ECpoint ECpub; |
| @@ -564,6 +570,11 @@ void curve25519_dh_finish(uchar x[32], uchar y[32], uchar z[32]); · 72 unmodified lines | |||
| 564 | void pbkdf2_x(uchar *p, ulong plen, uchar *s, ulong slen, ulong rounds, uchar *d, ulong dlen, | 570 | void pbkdf2_x(uchar *p, ulong plen, uchar *s, ulong slen, ulong rounds, uchar *d, ulong dlen, |
| 565 | DigestState* (*x)(uchar*, ulong, uchar*, ulong, uchar*, DigestState*), int xlen); | 571 | DigestState* (*x)(uchar*, ulong, uchar*, ulong, uchar*, DigestState*), int xlen); |
| 566 | 572 | ||
| 573 | /* scrypt password-based key derivation function */ | ||
| 574 | char* scrypt(uchar *p, ulong plen, uchar *s, ulong slen, | ||
| 575 | ulong N, ulong R, ulong P, | ||
| 576 | uchar *d, ulong dlen); | ||
| 577 | |||
| 567 | /* hmac-based key derivation function (rfc5869) */ | 578 | /* hmac-based key derivation function (rfc5869) */ |
| 568 | void hkdf_x(uchar *salt, ulong nsalt, uchar *info, ulong ninfo, uchar *key, ulong nkey, uchar *d, ulong dlen, | 579 | void hkdf_x(uchar *salt, ulong nsalt, uchar *info, ulong ninfo, uchar *key, ulong nkey, uchar *d, ulong dlen, |
| 569 | DigestState* (*x)(uchar*, ulong, uchar*, ulong, uchar*, DigestState*), int xlen); | 580 | DigestState* (*x)(uchar*, ulong, uchar*, ulong, uchar*, DigestState*), int xlen); |
| @@ -58,6 +58,7 @@ OFILES=\ · 57 unmodified lines | |||
| 58 | rsagen.$O\ | 58 | rsagen.$O\ |
| 59 | rsaprivtopub.$O\ | 59 | rsaprivtopub.$O\ |
| 60 | salsa.$O\ | 60 | salsa.$O\ |
| 61 | scrypt.$O\ | ||
| 61 | secp256k1.$O\ | 62 | secp256k1.$O\ |
| 62 | secp256r1.$O\ | 63 | secp256r1.$O\ |
| 63 | secp384r1.$O\ | 64 | secp384r1.$O\ |
| @@ -15,13 +15,13 @@ enum{ · 14 unmodified lines | |||
| 15 | }; | 15 | }; |
| 16 | 16 | ||
| 17 | /* little-endian data order */ | 17 | /* little-endian data order */ |
| 18 | #define GET4(p) ((((((p)[3]<<8) | (p)[2])<<8) | (p)[1])<<8 | (p)[0]) | 18 | #define GET4(p) ((p)[0]|((p)[1]<<8)|((p)[2]<<16)|((p)[3]<<24)) |
| 19 | #define PUT4(p, v) (((p)[0]=v), (v>>=8), ((p)[1]=v), (v>>=8), ((p)[2]=v), (v>>=8), ((p)[3]=v)) | 19 | #define PUT4(p,v) (p)[0]=(v);(p)[1]=(v)>>8;(p)[2]=(v)>>16;(p)[3]=(v)>>24 |
| 20 | 20 | ||
| 21 | #define ROTATE(v,c) ((u32int)((v) << (c)) | ((v) >> (32 - (c)))) | 21 | #define ROTATE(v,c) ((u32int)((v) << (c)) | ((v) >> (32 - (c)))) |
| 22 | 22 | ||
| 23 | #define QUARTERROUND(ia,ib,ic,id) { \ | 23 | #define QUARTERROUND(ia,ib,ic,id) { \ |
| 24 | u32int a, b, c, d, t;\ | 24 | u32int a, b, c, d, t; \ |
| 25 | a = x[ia]; b = x[ib]; c = x[ic]; d = x[id]; \ | 25 | a = x[ia]; b = x[ib]; c = x[ic]; d = x[id]; \ |
| 26 | a += b; t = d^a; d = ROTATE(t,16); \ | 26 | a += b; t = d^a; d = ROTATE(t,16); \ |
| 27 | c += d; t = b^c; b = ROTATE(t,12); \ | 27 | c += d; t = b^c; b = ROTATE(t,12); \ |
| @@ -32,12 +32,9 @@ enum{ · 4 unmodified lines | |||
| 32 | 32 | ||
| 33 | #define ENCRYPT(s, x, y, d) {\ | 33 | #define ENCRYPT(s, x, y, d) {\ |
| 34 | u32int v; \ | 34 | u32int v; \ |
| 35 | uchar *sp, *dp; \ | 35 | v = GET4(s); \ |
| 36 | sp = (s); \ | ||
| 37 | v = GET4(sp); \ | ||
| 38 | v ^= (x)+(y); \ | 36 | v ^= (x)+(y); \ |
| 39 | dp = (d); \ | 37 | PUT4(d, v); \ |
| 40 | PUT4(dp, v); \ | ||
| 41 | } | 38 | } |
| 42 | 39 | ||
| 43 | static uchar sigma[16] = "expand 32-byte k"; | 40 | static uchar sigma[16] = "expand 32-byte k"; |
| @@ -57,7 +54,8 @@ setupChachastate(Chachastate *s, uchar *key, ulong keylen, uchar *iv, ulong ivle · 13 unmodified lines | |||
| 57 | { | 54 | { |
| 58 | if(keylen != 256/8 && keylen != 128/8) | 55 | if(keylen != 256/8 && keylen != 128/8) |
| 59 | sysfatal("invalid chacha key length"); | 56 | sysfatal("invalid chacha key length"); |
| 60 | if(ivlen != 96/8 && ivlen != 64/8) | 57 | if(ivlen != 64/8 && ivlen != 96/8 |
| 58 | && ivlen != 128/8 && ivlen != 192/8) /* hchacha, xchacha */ | ||
| 61 | sysfatal("invalid chacha iv length"); | 59 | sysfatal("invalid chacha iv length"); |
| 62 | if(rounds == 0) | 60 | if(rounds == 0) |
| 63 | rounds = 20; | 61 | rounds = 20; |
| @@ -70,7 +68,16 @@ setupChachastate(Chachastate *s, uchar *key, ulong keylen, uchar *iv, ulong ivle · 6 unmodified lines | |||
| 70 | load(&s->input[4], key, 4); | 68 | load(&s->input[4], key, 4); |
| 71 | load(&s->input[8], key, 4); | 69 | load(&s->input[8], key, 4); |
| 72 | } | 70 | } |
| 73 | s->ivwords = ivlen/sizeof(u32int); | 71 | s->xkey[0] = s->input[4]; |
| 72 | s->xkey[1] = s->input[5]; | ||
| 73 | s->xkey[2] = s->input[6]; | ||
| 74 | s->xkey[3] = s->input[7]; | ||
| 75 | s->xkey[4] = s->input[8]; | ||
| 76 | s->xkey[5] = s->input[9]; | ||
| 77 | s->xkey[6] = s->input[10]; | ||
| 78 | s->xkey[7] = s->input[11]; | ||
| 79 | |||
| 80 | s->ivwords = ivlen/4; | ||
| 74 | s->input[12] = 0; | 81 | s->input[12] = 0; |
| 75 | s->input[13] = 0; | 82 | s->input[13] = 0; |
| 76 | if(iv == nil){ | 83 | if(iv == nil){ |
| @@ -80,9 +87,88 @@ setupChachastate(Chachastate *s, uchar *key, ulong keylen, uchar *iv, ulong ivle · 3 unmodified lines | |||
| 80 | chacha_setiv(s, iv); | 87 | chacha_setiv(s, iv); |
| 81 | } | 88 | } |
| 82 | 89 | ||
| 90 | static void | ||
| 91 | dorounds(u32int x[Blockwords], int rounds) | ||
| 92 | { | ||
| 93 | for(; rounds > 0; rounds -= 2) { | ||
| 94 | QUARTERROUND(0, 4, 8,12) | ||
| 95 | QUARTERROUND(1, 5, 9,13) | ||
| 96 | QUARTERROUND(2, 6,10,14) | ||
| 97 | QUARTERROUND(3, 7,11,15) | ||
| 98 | |||
| 99 | QUARTERROUND(0, 5,10,15) | ||
| 100 | QUARTERROUND(1, 6,11,12) | ||
| 101 | QUARTERROUND(2, 7, 8,13) | ||
| 102 | QUARTERROUND(3, 4, 9,14) | ||
| 103 | } | ||
| 104 | } | ||
| 105 | |||
| 106 | static void | ||
| 107 | hchachablock(uchar h[32], Chachastate *s) | ||
| 108 | { | ||
| 109 | u32int x[16]; | ||
| 110 | |||
| 111 | x[0] = s->input[0]; | ||
| 112 | x[1] = s->input[1]; | ||
| 113 | x[2] = s->input[2]; | ||
| 114 | x[3] = s->input[3]; | ||
| 115 | x[4] = s->input[4]; | ||
| 116 | x[5] = s->input[5]; | ||
| 117 | x[6] = s->input[6]; | ||
| 118 | x[7] = s->input[7]; | ||
| 119 | x[8] = s->input[8]; | ||
| 120 | x[9] = s->input[9]; | ||
| 121 | x[10] = s->input[10]; | ||
| 122 | x[11] = s->input[11]; | ||
| 123 | x[12] = s->input[12]; | ||
| 124 | x[13] = s->input[13]; | ||
| 125 | x[14] = s->input[14]; | ||
| 126 | x[15] = s->input[15]; | ||
| 127 | |||
| 128 | dorounds(x, s->rounds); | ||
| 129 | |||
| 130 | PUT4(h+0*4, x[0]); | ||
| 131 | PUT4(h+1*4, x[1]); | ||
| 132 | PUT4(h+2*4, x[2]); | ||
| 133 | PUT4(h+3*4, x[3]); | ||
| 134 | PUT4(h+4*4, x[12]); | ||
| 135 | PUT4(h+5*4, x[13]); | ||
| 136 | PUT4(h+6*4, x[14]); | ||
| 137 | PUT4(h+7*4, x[15]); | ||
| 138 | } | ||
| 139 | |||
| 83 | void | 140 | void |
| 84 | chacha_setiv(Chachastate *s, uchar *iv) | 141 | chacha_setiv(Chachastate *s, uchar *iv) |
| 85 | { | 142 | { |
| 143 | if(s->ivwords == 192/32){ | ||
| 144 | /* xchacha with 192-bit iv */ | ||
| 145 | u32int counter[2]; | ||
| 146 | uchar h[32]; | ||
| 147 | |||
| 148 | s->input[4] = s->xkey[0]; | ||
| 149 | s->input[5] = s->xkey[1]; | ||
| 150 | s->input[6] = s->xkey[2]; | ||
| 151 | s->input[7] = s->xkey[3]; | ||
| 152 | s->input[8] = s->xkey[4]; | ||
| 153 | s->input[9] = s->xkey[5]; | ||
| 154 | s->input[10] = s->xkey[6]; | ||
| 155 | s->input[11] = s->xkey[7]; | ||
| 156 | |||
| 157 | counter[0] = s->input[12]; | ||
| 158 | counter[1] = s->input[13]; | ||
| 159 | |||
| 160 | load(&s->input[12], iv, 4); | ||
| 161 | |||
| 162 | hchachablock(h, s); | ||
| 163 | load(&s->input[4], h, 8); | ||
| 164 | memset(h, 0, 32); | ||
| 165 | |||
| 166 | s->input[12] = counter[0]; | ||
| 167 | s->input[13] = counter[1]; | ||
| 168 | |||
| 169 | load(&s->input[14], iv+16, 2); | ||
| 170 | return; | ||
| 171 | } | ||
| 86 | load(&s->input[16 - s->ivwords], iv, s->ivwords); | 172 | load(&s->input[16 - s->ivwords], iv, s->ivwords); |
| 87 | } | 173 | } |
| 88 | 174 | ||
| @@ -90,7 +176,7 @@ void · 1 unmodified line | |||
| 90 | chacha_setblock(Chachastate *s, u64int blockno) | 176 | chacha_setblock(Chachastate *s, u64int blockno) |
| 91 | { | 177 | { |
| 92 | s->input[12] = blockno; | 178 | s->input[12] = blockno; |
| 93 | if(s->ivwords == 2) | 179 | if(s->ivwords != 3) |
| 94 | s->input[13] = blockno>>32; | 180 | s->input[13] = blockno>>32; |
| 95 | } | 181 | } |
| 96 | 182 | ||
| @@ -98,9 +184,8 @@ static void · 1 unmodified line | |||
| 98 | encryptblock(Chachastate *s, uchar *src, uchar *dst) | 184 | encryptblock(Chachastate *s, uchar *src, uchar *dst) |
| 99 | { | 185 | { |
| 100 | u32int x[Blockwords]; | 186 | u32int x[Blockwords]; |
| 101 | int i, rounds; | 187 | int i; |
| 102 | 188 | ||
| 103 | rounds = s->rounds; | ||
| 104 | x[0] = s->input[0]; | 189 | x[0] = s->input[0]; |
| 105 | x[1] = s->input[1]; | 190 | x[1] = s->input[1]; |
| 106 | x[2] = s->input[2]; | 191 | x[2] = s->input[2]; |
| @@ -117,37 +202,8 @@ encryptblock(Chachastate *s, uchar *src, uchar *dst) · 10 unmodified lines | |||
| 117 | x[13] = s->input[13]; | 202 | x[13] = s->input[13]; |
| 118 | x[14] = s->input[14]; | 203 | x[14] = s->input[14]; |
| 119 | x[15] = s->input[15]; | 204 | x[15] = s->input[15]; |
| 205 | dorounds(x, s->rounds); | ||
| 120 | 206 | ||
| 121 | for(i = rounds; i > 0; i -= 2) { | ||
| 122 | QUARTERROUND(0, 4, 8,12) | ||
| 123 | QUARTERROUND(1, 5, 9,13) | ||
| 124 | QUARTERROUND(2, 6,10,14) | ||
| 125 | QUARTERROUND(3, 7,11,15) | ||
| 126 | |||
| 127 | QUARTERROUND(0, 5,10,15) | ||
| 128 | QUARTERROUND(1, 6,11,12) | ||
| 129 | QUARTERROUND(2, 7, 8,13) | ||
| 130 | QUARTERROUND(3, 4, 9,14) | ||
| 131 | } | ||
| 132 | |||
| 133 | #ifdef FULL_UNROLL | ||
| 134 | ENCRYPT(src+0*4, x[0], s->input[0], dst+0*4); | ||
| 135 | ENCRYPT(src+1*4, x[1], s->input[1], dst+1*4); | ||
| 136 | ENCRYPT(src+2*4, x[2], s->input[2], dst+2*4); | ||
| 137 | ENCRYPT(src+3*4, x[3], s->input[3], dst+3*4); | ||
| 138 | ENCRYPT(src+4*4, x[4], s->input[4], dst+4*4); | ||
| 139 | ENCRYPT(src+5*4, x[5], s->input[5], dst+5*4); | ||
| 140 | ENCRYPT(src+6*4, x[6], s->input[6], dst+6*4); | ||
| 141 | ENCRYPT(src+7*4, x[7], s->input[7], dst+7*4); | ||
| 142 | ENCRYPT(src+8*4, x[8], s->input[8], dst+8*4); | ||
| 143 | ENCRYPT(src+9*4, x[9], s->input[9], dst+9*4); | ||
| 144 | ENCRYPT(src+10*4, x[10], s->input[10], dst+10*4); | ||
| 145 | ENCRYPT(src+11*4, x[11], s->input[11], dst+11*4); | ||
| 146 | ENCRYPT(src+12*4, x[12], s->input[12], dst+12*4); | ||
| 147 | ENCRYPT(src+13*4, x[13], s->input[13], dst+13*4); | ||
| 148 | ENCRYPT(src+14*4, x[14], s->input[14], dst+14*4); | ||
| 149 | ENCRYPT(src+15*4, x[15], s->input[15], dst+15*4); | ||
| 150 | #else | ||
| 151 | for(i=0; i<nelem(x); i+=4){ | 207 | for(i=0; i<nelem(x); i+=4){ |
| 152 | ENCRYPT(src, x[i], s->input[i], dst); | 208 | ENCRYPT(src, x[i], s->input[i], dst); |
| 153 | ENCRYPT(src+4, x[i+1], s->input[i+1], dst+4); | 209 | ENCRYPT(src+4, x[i+1], s->input[i+1], dst+4); |
| @@ -156,9 +212,8 @@ encryptblock(Chachastate *s, uchar *src, uchar *dst) · 2 unmodified lines | |||
| 156 | src += 16; | 212 | src += 16; |
| 157 | dst += 16; | 213 | dst += 16; |
| 158 | } | 214 | } |
| 159 | #endif | ||
| 160 | 215 | ||
| 161 | if(++s->input[12] == 0 && s->ivwords == 2) | 216 | if(++s->input[12] == 0 && s->ivwords != 3) |
| 162 | s->input[13]++; | 217 | s->input[13]++; |
| 163 | } | 218 | } |
| 164 | 219 | ||
| @@ -184,3 +239,13 @@ chacha_encrypt(uchar *buf, ulong bytes, Chachastate *s) · 19 unmodified lines | |||
| 184 | { | 239 | { |
| 185 | chacha_encrypt2(buf, buf, bytes, s); | 240 | chacha_encrypt2(buf, buf, bytes, s); |
| 186 | } | 241 | } |
| 242 | |||
| 243 | void | ||
| 244 | hchacha(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds) | ||
| 245 | { | ||
| 246 | Chachastate s[1]; | ||
| 247 | |||
| 248 | setupChachastate(s, key, keylen, nonce, 16, rounds); | ||
| 249 | hchachablock(h, s); | ||
| 250 | memset(s, 0, sizeof(s)); | ||
| 251 | } | ||
| @@ -1,24 +1,17 @@ | |||
| 1 | #include "os.h" | 1 | #include "os.h" |
| 2 | #include <libsec.h> | 2 | #include <libsec.h> |
| 3 | 3 | ||
| 4 | enum{ | ||
| 5 | Blockwords= SalsaBsize/sizeof(u32int) | ||
| 6 | }; | ||
| 7 | |||
| 8 | /* little-endian data order */ | 4 | /* little-endian data order */ |
| 9 | #define GET4(p) ((((((p)[3]<<8) | (p)[2])<<8) | (p)[1])<<8 | (p)[0]) | 5 | #define GET4(p) ((p)[0]|((p)[1]<<8)|((p)[2]<<16)|((p)[3]<<24)) |
| 10 | #define PUT4(p, v) (((p)[0]=v), (v>>=8), ((p)[1]=v), (v>>=8), ((p)[2]=v), (v>>=8), ((p)[3]=v)) | 6 | #define PUT4(p,v) (p)[0]=(v);(p)[1]=(v)>>8;(p)[2]=(v)>>16;(p)[3]=(v)>>24 |
| 11 | 7 | ||
| 12 | #define ROTATE(v,c) (t = v, (u32int)(t << (c)) | (t >> (32 - (c)))) | 8 | #define ROTATE(v,c) (t = v, (u32int)(t << (c)) | (t >> (32 - (c)))) |
| 13 | 9 | ||
| 14 | #define ENCRYPT(s, x, y, d) {\ | 10 | #define ENCRYPT(s, x, y, d) {\ |
| 15 | u32int v; \ | 11 | u32int v; \ |
| 16 | uchar *sp, *dp; \ | 12 | v = GET4(s); \ |
| 17 | sp = (s); \ | ||
| 18 | v = GET4(sp); \ | ||
| 19 | v ^= (x)+(y); \ | 13 | v ^= (x)+(y); \ |
| 20 | dp = (d); \ | 14 | PUT4(d, v); \ |
| 21 | PUT4(dp, v); \ | ||
| 22 | } | 15 | } |
| 23 | 16 | ||
| 24 | static uchar sigma[16] = "expand 32-byte k"; | 17 | static uchar sigma[16] = "expand 32-byte k"; |
| @@ -38,7 +31,8 @@ setupSalsastate(Salsastate *s, uchar *key, ulong keylen, uchar *iv, ulong ivlen, · 13 unmodified lines | |||
| 38 | { | 31 | { |
| 39 | if(keylen != 256/8 && keylen != 128/8) | 32 | if(keylen != 256/8 && keylen != 128/8) |
| 40 | sysfatal("invalid salsa key length"); | 33 | sysfatal("invalid salsa key length"); |
| 41 | if(ivlen != 64/8 && ivlen != 128/8 && ivlen != 192/8) | 34 | if(ivlen != 64/8 |
| 35 | && ivlen != 128/8 && ivlen != 192/8) /* hsalsa, xsalsa */ | ||
| 42 | sysfatal("invalid salsa iv length"); | 36 | sysfatal("invalid salsa iv length"); |
| 43 | if(rounds == 0) | 37 | if(rounds == 0) |
| 44 | rounds = 20; | 38 | rounds = 20; |
| @@ -58,14 +52,14 @@ setupSalsastate(Salsastate *s, uchar *key, ulong keylen, uchar *iv, ulong ivlen, · 13 unmodified lines | |||
| 58 | load(&s->input[11], key, 4); | 52 | load(&s->input[11], key, 4); |
| 59 | load(&s->input[15], tau +4*3, 1); | 53 | load(&s->input[15], tau +4*3, 1); |
| 60 | } | 54 | } |
| 61 | s->key[0] = s->input[1]; | 55 | s->xkey[0] = s->input[1]; |
| 62 | s->key[1] = s->input[2]; | 56 | s->xkey[1] = s->input[2]; |
| 63 | s->key[2] = s->input[3]; | 57 | s->xkey[2] = s->input[3]; |
| 64 | s->key[3] = s->input[4]; | 58 | s->xkey[3] = s->input[4]; |
| 65 | s->key[4] = s->input[11]; | 59 | s->xkey[4] = s->input[11]; |
| 66 | s->key[5] = s->input[12]; | 60 | s->xkey[5] = s->input[12]; |
| 67 | s->key[6] = s->input[13]; | 61 | s->xkey[6] = s->input[13]; |
| 68 | s->key[7] = s->input[14]; | 62 | s->xkey[7] = s->input[14]; |
| 69 | 63 | ||
| 70 | s->ivwords = ivlen/4; | 64 | s->ivwords = ivlen/4; |
| 71 | s->input[8] = 0; | 65 | s->input[8] = 0; |
| @@ -78,30 +72,11 @@ setupSalsastate(Salsastate *s, uchar *key, ulong keylen, uchar *iv, ulong ivlen, · 6 unmodified lines | |||
| 78 | } | 72 | } |
| 79 | 73 | ||
| 80 | static void | 74 | static void |
| 81 | hsalsablock(uchar h[32], Salsastate *s) | 75 | dorounds(u32int x[16], int rounds) |
| 82 | { | 76 | { |
| 83 | u32int x[Blockwords], t; | 77 | u32int t; |
| 84 | int i, rounds; | ||
| 85 | 78 | ||
| 86 | rounds = s->rounds; | 79 | for(; rounds > 0; rounds -= 2) { |
| 87 | x[0] = s->input[0]; | ||
| 88 | x[1] = s->input[1]; | ||
| 89 | x[2] = s->input[2]; | ||
| 90 | x[3] = s->input[3]; | ||
| 91 | x[4] = s->input[4]; | ||
| 92 | x[5] = s->input[5]; | ||
| 93 | x[6] = s->input[6]; | ||
| 94 | x[7] = s->input[7]; | ||
| 95 | x[8] = s->input[8]; | ||
| 96 | x[9] = s->input[9]; | ||
| 97 | x[10] = s->input[10]; | ||
| 98 | x[11] = s->input[11]; | ||
| 99 | x[12] = s->input[12]; | ||
| 100 | x[13] = s->input[13]; | ||
| 101 | x[14] = s->input[14]; | ||
| 102 | x[15] = s->input[15]; | ||
| 103 | |||
| 104 | for(i = rounds; i > 0; i -= 2) { | ||
| 105 | x[4] ^= ROTATE( x[0]+x[12], 7); | 80 | x[4] ^= ROTATE( x[0]+x[12], 7); |
| 106 | x[8] ^= ROTATE( x[4]+ x[0], 9); | 81 | x[8] ^= ROTATE( x[4]+ x[0], 9); |
| 107 | x[12] ^= ROTATE( x[8]+ x[4],13); | 82 | x[12] ^= ROTATE( x[8]+ x[4],13); |
| @@ -135,6 +110,31 @@ hsalsablock(uchar h[32], Salsastate *s) · 27 unmodified lines | |||
| 135 | x[14] ^= ROTATE(x[13]+x[12],13); | 110 | x[14] ^= ROTATE(x[13]+x[12],13); |
| 136 | x[15] ^= ROTATE(x[14]+x[13],18); | 111 | x[15] ^= ROTATE(x[14]+x[13],18); |
| 137 | } | 112 | } |
| 113 | } | ||
| 114 | |||
| 115 | static void | ||
| 116 | hsalsablock(uchar h[32], Salsastate *s) | ||
| 117 | { | ||
| 118 | u32int x[16]; | ||
| 119 | |||
| 120 | x[0] = s->input[0]; | ||
| 121 | x[1] = s->input[1]; | ||
| 122 | x[2] = s->input[2]; | ||
| 123 | x[3] = s->input[3]; | ||
| 124 | x[4] = s->input[4]; | ||
| 125 | x[5] = s->input[5]; | ||
| 126 | x[6] = s->input[6]; | ||
| 127 | x[7] = s->input[7]; | ||
| 128 | x[8] = s->input[8]; | ||
| 129 | x[9] = s->input[9]; | ||
| 130 | x[10] = s->input[10]; | ||
| 131 | x[11] = s->input[11]; | ||
| 132 | x[12] = s->input[12]; | ||
| 133 | x[13] = s->input[13]; | ||
| 134 | x[14] = s->input[14]; | ||
| 135 | x[15] = s->input[15]; | ||
| 136 | |||
| 137 | dorounds(x, s->rounds); | ||
| 138 | 138 | ||
| 139 | PUT4(h+0*4, x[0]); | 139 | PUT4(h+0*4, x[0]); |
| 140 | PUT4(h+1*4, x[5]); | 140 | PUT4(h+1*4, x[5]); |
| @@ -150,7 +150,7 @@ void · 9 unmodified lines | |||
| 150 | salsa_setiv(Salsastate *s, uchar *iv) | 150 | salsa_setiv(Salsastate *s, uchar *iv) |
| 151 | { | 151 | { |
| 152 | if(s->ivwords == 128/32){ | 152 | if(s->ivwords == 128/32){ |
| 153 | /* hsalsa 128-bit iv */ | 153 | /* hsalsa with 128-bit iv */ |
| 154 | load(&s->input[6], iv, 4); | 154 | load(&s->input[6], iv, 4); |
| 155 | return; | 155 | return; |
| 156 | } | 156 | } |
| @@ -162,14 +162,14 @@ salsa_setiv(Salsastate *s, uchar *iv) · 5 unmodified lines | |||
| 162 | counter[0] = s->input[8]; | 162 | counter[0] = s->input[8]; |
| 163 | counter[1] = s->input[9]; | 163 | counter[1] = s->input[9]; |
| 164 | 164 | ||
| 165 | s->input[1] = s->key[0]; | 165 | s->input[1] = s->xkey[0]; |
| 166 | s->input[2] = s->key[1]; | 166 | s->input[2] = s->xkey[1]; |
| 167 | s->input[3] = s->key[2]; | 167 | s->input[3] = s->xkey[2]; |
| 168 | s->input[4] = s->key[3]; | 168 | s->input[4] = s->xkey[3]; |
| 169 | s->input[11] = s->key[4]; | 169 | s->input[11] = s->xkey[4]; |
| 170 | s->input[12] = s->key[5]; | 170 | s->input[12] = s->xkey[5]; |
| 171 | s->input[13] = s->key[6]; | 171 | s->input[13] = s->xkey[6]; |
| 172 | s->input[14] = s->key[7]; | 172 | s->input[14] = s->xkey[7]; |
| 173 | 173 | ||
| 174 | load(&s->input[6], iv, 4); | 174 | load(&s->input[6], iv, 4); |
| 175 | 175 | ||
| @@ -197,10 +197,9 @@ salsa_setblock(Salsastate *s, u64int blockno) · 21 unmodified lines | |||
| 197 | static void | 197 | static void |
| 198 | encryptblock(Salsastate *s, uchar *src, uchar *dst) | 198 | encryptblock(Salsastate *s, uchar *src, uchar *dst) |
| 199 | { | 199 | { |
| 200 | u32int x[Blockwords], t; | 200 | u32int x[16]; |
| 201 | int i, rounds; | 201 | int i; |
| 202 | 202 | ||
| 203 | rounds = s->rounds; | ||
| 204 | x[0] = s->input[0]; | 203 | x[0] = s->input[0]; |
| 205 | x[1] = s->input[1]; | 204 | x[1] = s->input[1]; |
| 206 | x[2] = s->input[2]; | 205 | x[2] = s->input[2]; |
| @@ -218,59 +217,8 @@ encryptblock(Salsastate *s, uchar *src, uchar *dst) · 11 unmodified lines | |||
| 218 | x[14] = s->input[14]; | 217 | x[14] = s->input[14]; |
| 219 | x[15] = s->input[15]; | 218 | x[15] = s->input[15]; |
| 220 | 219 | ||
| 221 | for(i = rounds; i > 0; i -= 2) { | 220 | dorounds(x, s->rounds); |
| 222 | x[4] ^= ROTATE( x[0]+x[12], 7); | ||
| 223 | x[8] ^= ROTATE( x[4]+ x[0], 9); | ||
| 224 | x[12] ^= ROTATE( x[8]+ x[4],13); | ||
| 225 | x[0] ^= ROTATE(x[12]+ x[8],18); | ||
| 226 | x[9] ^= ROTATE( x[5]+ x[1], 7); | ||
| 227 | x[13] ^= ROTATE( x[9]+ x[5], 9); | ||
| 228 | x[1] ^= ROTATE(x[13]+ x[9],13); | ||
| 229 | x[5] ^= ROTATE( x[1]+x[13],18); | ||
| 230 | x[14] ^= ROTATE(x[10]+ x[6], 7); | ||
| 231 | x[2] ^= ROTATE(x[14]+x[10], 9); | ||
| 232 | x[6] ^= ROTATE( x[2]+x[14],13); | ||
| 233 | x[10] ^= ROTATE( x[6]+ x[2],18); | ||
| 234 | x[3] ^= ROTATE(x[15]+x[11], 7); | ||
| 235 | x[7] ^= ROTATE( x[3]+x[15], 9); | ||
| 236 | x[11] ^= ROTATE( x[7]+ x[3],13); | ||
| 237 | x[15] ^= ROTATE(x[11]+ x[7],18); | ||
| 238 | x[1] ^= ROTATE( x[0]+ x[3], 7); | ||
| 239 | x[2] ^= ROTATE( x[1]+ x[0], 9); | ||
| 240 | x[3] ^= ROTATE( x[2]+ x[1],13); | ||
| 241 | x[0] ^= ROTATE( x[3]+ x[2],18); | ||
| 242 | x[6] ^= ROTATE( x[5]+ x[4], 7); | ||
| 243 | x[7] ^= ROTATE( x[6]+ x[5], 9); | ||
| 244 | x[4] ^= ROTATE( x[7]+ x[6],13); | ||
| 245 | x[5] ^= ROTATE( x[4]+ x[7],18); | ||
| 246 | x[11] ^= ROTATE(x[10]+ x[9], 7); | ||
| 247 | x[8] ^= ROTATE(x[11]+x[10], 9); | ||
| 248 | x[9] ^= ROTATE( x[8]+x[11],13); | ||
| 249 | x[10] ^= ROTATE( x[9]+ x[8],18); | ||
| 250 | x[12] ^= ROTATE(x[15]+x[14], 7); | ||
| 251 | x[13] ^= ROTATE(x[12]+x[15], 9); | ||
| 252 | x[14] ^= ROTATE(x[13]+x[12],13); | ||
| 253 | x[15] ^= ROTATE(x[14]+x[13],18); | ||
| 254 | } | ||
| 255 | 221 | ||
| 256 | #ifdef FULL_UNROLL | ||
| 257 | ENCRYPT(src+0*4, x[0], s->input[0], dst+0*4); | ||
| 258 | ENCRYPT(src+1*4, x[1], s->input[1], dst+1*4); | ||
| 259 | ENCRYPT(src+2*4, x[2], s->input[2], dst+2*4); | ||
| 260 | ENCRYPT(src+3*4, x[3], s->input[3], dst+3*4); | ||
| 261 | ENCRYPT(src+4*4, x[4], s->input[4], dst+4*4); | ||
| 262 | ENCRYPT(src+5*4, x[5], s->input[5], dst+5*4); | ||
| 263 | ENCRYPT(src+6*4, x[6], s->input[6], dst+6*4); | ||
| 264 | ENCRYPT(src+7*4, x[7], s->input[7], dst+7*4); | ||
| 265 | ENCRYPT(src+8*4, x[8], s->input[8], dst+8*4); | ||
| 266 | ENCRYPT(src+9*4, x[9], s->input[9], dst+9*4); | ||
| 267 | ENCRYPT(src+10*4, x[10], s->input[10], dst+10*4); | ||
| 268 | ENCRYPT(src+11*4, x[11], s->input[11], dst+11*4); | ||
| 269 | ENCRYPT(src+12*4, x[12], s->input[12], dst+12*4); | ||
| 270 | ENCRYPT(src+13*4, x[13], s->input[13], dst+13*4); | ||
| 271 | ENCRYPT(src+14*4, x[14], s->input[14], dst+14*4); | ||
| 272 | ENCRYPT(src+15*4, x[15], s->input[15], dst+15*4); | ||
| 273 | #else | ||
| 274 | for(i=0; i<nelem(x); i+=4){ | 222 | for(i=0; i<nelem(x); i+=4){ |
| 275 | ENCRYPT(src, x[i], s->input[i], dst); | 223 | ENCRYPT(src, x[i], s->input[i], dst); |
| 276 | ENCRYPT(src+4, x[i+1], s->input[i+1], dst+4); | 224 | ENCRYPT(src+4, x[i+1], s->input[i+1], dst+4); |
| @@ -279,7 +227,6 @@ encryptblock(Salsastate *s, uchar *src, uchar *dst) · 2 unmodified lines | |||
| 279 | src += 16; | 227 | src += 16; |
| 280 | dst += 16; | 228 | dst += 16; |
| 281 | } | 229 | } |
| 282 | #endif | ||
| 283 | 230 | ||
| 284 | if(++s->input[8] == 0) | 231 | if(++s->input[8] == 0) |
| 285 | s->input[9]++; | 232 | s->input[9]++; |
| @@ -308,6 +255,48 @@ salsa_encrypt(uchar *buf, ulong bytes, Salsastate *s) · 22 unmodified lines | |||
| 308 | salsa_encrypt2(buf, buf, bytes, s); | 255 | salsa_encrypt2(buf, buf, bytes, s); |
| 309 | } | 256 | } |
| 310 | 257 | ||
| 258 | void | ||
| 259 | salsa_core(u32int in[16], u32int out[16], int rounds) | ||
| 260 | { | ||
| 261 | u32int x[16]; | ||
| 262 | |||
| 263 | x[0] = in[0]; | ||
| 264 | x[1] = in[1]; | ||
| 265 | x[2] = in[2]; | ||
| 266 | x[3] = in[3]; | ||
| 267 | x[4] = in[4]; | ||
| 268 | x[5] = in[5]; | ||
| 269 | x[6] = in[6]; | ||
| 270 | x[7] = in[7]; | ||
| 271 | x[8] = in[8]; | ||
| 272 | x[9] = in[9]; | ||
| 273 | x[10] = in[10]; | ||
| 274 | x[11] = in[11]; | ||
| 275 | x[12] = in[12]; | ||
| 276 | x[13] = in[13]; | ||
| 277 | x[14] = in[14]; | ||
| 278 | x[15] = in[15]; | ||
| 279 | |||
| 280 | dorounds(x, rounds); | ||
| 281 | |||
| 282 | out[0] = x[0] + in[0]; | ||
| 283 | out[1] = x[1] + in[1]; | ||
| 284 | out[2] = x[2] + in[2]; | ||
| 285 | out[3] = x[3] + in[3]; | ||
| 286 | out[4] = x[4] + in[4]; | ||
| 287 | out[5] = x[5] + in[5]; | ||
| 288 | out[6] = x[6] + in[6]; | ||
| 289 | out[7] = x[7] + in[7]; | ||
| 290 | out[8] = x[8] + in[8]; | ||
| 291 | out[9] = x[9] + in[9]; | ||
| 292 | out[10] = x[10] + in[10]; | ||
| 293 | out[11] = x[11] + in[11]; | ||
| 294 | out[12] = x[12] + in[12]; | ||
| 295 | out[13] = x[13] + in[13]; | ||
| 296 | out[14] = x[14] + in[14]; | ||
| 297 | out[15] = x[15] + in[15]; | ||
| 298 | } | ||
| 299 | |||
| 311 | void | 300 | void |
| 312 | hsalsa(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds) | 301 | hsalsa(uchar h[32], uchar *key, ulong keylen, uchar nonce[16], int rounds) |
| 313 | { | 302 | { |
| @@ -0,0 +1,119 @@ | |||
| 1 | #include "os.h" | ||
| 2 | #include <libsec.h> | ||
| 3 | |||
| 4 | #define movw(w, S, D) memmove(D, S, (w)*4) | ||
| 5 | |||
| 6 | static void | ||
| 7 | xorw(ulong w, u32int *S, u32int *D) | ||
| 8 | { | ||
| 9 | for(w /= 8; w; w--, D += 8, S += 8){ | ||
| 10 | D[0] ^= S[0]; | ||
| 11 | D[1] ^= S[1]; | ||
| 12 | D[2] ^= S[2]; | ||
| 13 | D[3] ^= S[3]; | ||
| 14 | D[4] ^= S[4]; | ||
| 15 | D[5] ^= S[5]; | ||
| 16 | D[6] ^= S[6]; | ||
| 17 | D[7] ^= S[7]; | ||
| 18 | } | ||
| 19 | } | ||
| 20 | |||
| 21 | static void | ||
| 22 | scryptBlockMix(ulong R, u32int *B, u32int *Y) | ||
| 23 | { | ||
| 24 | u32int X[16]; | ||
| 25 | ulong i; | ||
| 26 | |||
| 27 | R *= 2; | ||
| 28 | movw(16, &B[(R-1)*16], X); | ||
| 29 | for(i = 0; i < R; i += 2){ | ||
| 30 | xorw(16, &B[i*16], X); | ||
| 31 | salsa_core(X, X, 8); | ||
| 32 | movw(16, X, &Y[i*8]); | ||
| 33 | |||
| 34 | xorw(16, &B[(i+1)*16], X); | ||
| 35 | salsa_core(X, X, 8); | ||
| 36 | movw(16, X, &Y[i*8 + R*8]); | ||
| 37 | } | ||
| 38 | } | ||
| 39 | |||
| 40 | static void | ||
| 41 | scryptROMix(ulong R, ulong N, u32int *V, u32int *X, uchar *B) | ||
| 42 | { | ||
| 43 | ulong w, i, d; | ||
| 44 | u32int *Y; | ||
| 45 | |||
| 46 | w = R*32; | ||
| 47 | for(i=0; i<w; i++, B+=4) | ||
| 48 | X[i] = B[0] | (B[1]<<8) | (B[2]<<16) | (B[3]<<24); | ||
| 49 | |||
| 50 | Y = &X[w]; | ||
| 51 | for(i=0; i<N; i += 2){ | ||
| 52 | movw(w, X, &V[i*w]); | ||
| 53 | scryptBlockMix(R, X, Y); | ||
| 54 | |||
| 55 | movw(w, Y, &V[(i+1)*w]); | ||
| 56 | scryptBlockMix(R, Y, X); | ||
| 57 | } | ||
| 58 | for(i=0; i<N; i += 2){ | ||
| 59 | xorw(w, &V[(X[w-16] & (N-1))*w], X); | ||
| 60 | scryptBlockMix(R, X, Y); | ||
| 61 | |||
| 62 | xorw(w, &V[(Y[w-16] & (N-1))*w], Y); | ||
| 63 | scryptBlockMix(R, Y, X); | ||
| 64 | } | ||
| 65 | |||
| 66 | B -= w*4; | ||
| 67 | for(i=0; i<w; i++, B+=4) | ||
| 68 | d = X[i], B[0]=d, B[1]=d>>8, B[2]=d>>16, B[3]=d>>24; | ||
| 69 | } | ||
| 70 | |||
| 71 | char* | ||
| 72 | scrypt(p, plen, s, slen, N, R, P, d, dlen) | ||
| 73 | ulong plen, slen, dlen, N, R, P; | ||
| 74 | uchar *p, *s, *d; | ||
| 75 | { | ||
| 76 | static char oom[] = "out of memory"; | ||
| 77 | |||
| 78 | ulong rb, i; | ||
| 79 | u32int *V, *X; | ||
| 80 | uchar *B; | ||
| 81 | |||
| 82 | if(P < 1) | ||
| 83 | return "invalid parallelization parameter P"; | ||
| 84 | if(R < 1 || R >= (1UL<<(31-7))/P) | ||
| 85 | return "invalid block size parameter R"; | ||
| 86 | if(N < 2 || (N & (N-1)) != 0 || N >= (1UL<<(31-7))/R) | ||
| 87 | return "invalid cpu/memory cost parameter N"; | ||
| 88 | |||
| 89 | rb = R<<7; | ||
| 90 | if((B = malloc(P*rb)) == nil) | ||
| 91 | return oom; | ||
| 92 | if((V = malloc(N*rb)) == nil){ | ||
| 93 | free(B); | ||
| 94 | return oom; | ||
| 95 | } | ||
| 96 | if((X = malloc(2*rb)) == nil){ | ||
| 97 | free(V); | ||
| 98 | free(B); | ||
| 99 | return oom; | ||
| 100 | } | ||
| 101 | |||
| 102 | pbkdf2_x(p, plen, s, slen, 1, B, P*rb, hmac_sha2_256, SHA2_256dlen); | ||
| 103 | |||
| 104 | for(i=0; i<P; i++) | ||
| 105 | scryptROMix(R, N, V, X, &B[i*rb]); | ||
| 106 | |||
| 107 | memset(X, 0, 2*rb); | ||
| 108 | free(X); | ||
| 109 | |||
| 110 | memset(V, 0, N*rb); | ||
| 111 | free(V); | ||
| 112 | |||
| 113 | pbkdf2_x(p, plen, B, P*rb, 1, d, dlen, hmac_sha2_256, SHA2_256dlen); | ||
| 114 | |||
| 115 | memset(B, 0, P*rb); | ||
| 116 | free(B); | ||
| 117 | |||
| 118 | return nil; | ||
| 119 | } | ||