crypto/poly1305_generic.c

   1 /*
   2  * Poly1305 authenticator algorithm, RFC7539
   3  *
   4  * Copyright (C) 2015 Martin Willi
   5  *
   6  * Based on public domain code by Andrew Moon and Daniel J. Bernstein.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  */
  13
  14 #include <crypto/algapi.h>
  15 #include <crypto/internal/hash.h>
  16 #include <linux/crypto.h>
  17 #include <linux/kernel.h>
  18 #include <linux/module.h>
  19
  20 #define POLY1305_BLOCK_SIZE     16
  21 #define POLY1305_KEY_SIZE       32
  22 #define POLY1305_DIGEST_SIZE    16
  23
  24 struct poly1305_desc_ctx {
  25         /* key */
  26         u32 r[5];
  27         /* finalize key */
  28         u32 s[4];
  29         /* accumulator */
  30         u32 h[5];
  31         /* partial buffer */
  32         u8 buf[POLY1305_BLOCK_SIZE];
  33         /* bytes used in partial buffer */
  34         unsigned int buflen;
  35         /* r key has been set */
  36         bool rset;
  37         /* s key has been set */
  38         bool sset;
  39 };
  40
  41 static inline u64 mlt(u64 a, u64 b)
  42 {
  43         return a * b;
  44 }
  45
  46 static inline u32 sr(u64 v, u_char n)
  47 {
  48         return v >> n;
  49 }
  50
  51 static inline u32 and(u32 v, u32 mask)
  52 {
  53         return v & mask;
  54 }
  55
  56 static inline u32 le32_to_cpuvp(const void *p)
  57 {
  58         return le32_to_cpup(p);
  59 }
  60
  61 static int poly1305_init(struct shash_desc *desc)
  62 {
  63         struct poly1305_desc_ctx *dctx = shash_desc_ctx(desc);
  64
  65         memset(dctx->h, 0, sizeof(dctx->h));
  66         dctx->buflen = 0;
  67         dctx->rset = false;
  68         dctx->sset = false;
  69
  70         return 0;
  71 }
  72
  73 static int poly1305_setkey(struct crypto_shash *tfm,
  74                            const u8 *key, unsigned int keylen)
  75 {
  76         /* Poly1305 requires a unique key for each tag, which implies that
  77          * we can't set it on the tfm that gets accessed by multiple users
  78          * simultaneously. Instead we expect the key as the first 32 bytes in
  79          * the update() call. */
  80         return -ENOTSUPP;
  81 }
  82
  83 static void poly1305_setrkey(struct poly1305_desc_ctx *dctx, const u8 *key)
  84 {
  85         /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
  86         dctx->r[0] = (le32_to_cpuvp(key +  0) >> 0) & 0x3ffffff;
  87         dctx->r[1] = (le32_to_cpuvp(key +  3) >> 2) & 0x3ffff03;
  88         dctx->r[2] = (le32_to_cpuvp(key +  6) >> 4) & 0x3ffc0ff;
  89         dctx->r[3] = (le32_to_cpuvp(key +  9) >> 6) & 0x3f03fff;
  90         dctx->r[4] = (le32_to_cpuvp(key + 12) >> 8) & 0x00fffff;
  91 }
  92
  93 static void poly1305_setskey(struct poly1305_desc_ctx *dctx, const u8 *key)
  94 {
  95         dctx->s[0] = le32_to_cpuvp(key +  0);
  96         dctx->s[1] = le32_to_cpuvp(key +  4);
  97         dctx->s[2] = le32_to_cpuvp(key +  8);
  98         dctx->s[3] = le32_to_cpuvp(key + 12);
  99 }
 100
 101 static unsigned int poly1305_blocks(struct poly1305_desc_ctx *dctx,
 102                                     const u8 *src, unsigned int srclen,
 103                                     u32 hibit)
 104 {
 105         u32 r0, r1, r2, r3, r4;
 106         u32 s1, s2, s3, s4;
 107         u32 h0, h1, h2, h3, h4;
 108         u64 d0, d1, d2, d3, d4;
 109
 110         if (unlikely(!dctx->sset)) {
 111                 if (!dctx->rset && srclen >= POLY1305_BLOCK_SIZE) {
 112                         poly1305_setrkey(dctx, src);
 113                         src += POLY1305_BLOCK_SIZE;
 114                         srclen -= POLY1305_BLOCK_SIZE;
 115                         dctx->rset = true;
 116                 }
 117                 if (srclen >= POLY1305_BLOCK_SIZE) {
 118                         poly1305_setskey(dctx, src);
 119                         src += POLY1305_BLOCK_SIZE;
 120                         srclen -= POLY1305_BLOCK_SIZE;
 121                         dctx->sset = true;
 122                 }
 123         }
 124
 125         r0 = dctx->r[0];
 126         r1 = dctx->r[1];
 127         r2 = dctx->r[2];
 128         r3 = dctx->r[3];
 129         r4 = dctx->r[4];
 130
 131         s1 = r1 * 5;
 132         s2 = r2 * 5;
 133         s3 = r3 * 5;
 134         s4 = r4 * 5;
 135
 136         h0 = dctx->h[0];
 137         h1 = dctx->h[1];
 138         h2 = dctx->h[2];
 139         h3 = dctx->h[3];
 140         h4 = dctx->h[4];
 141
 142         while (likely(srclen >= POLY1305_BLOCK_SIZE)) {
 143
 144                 /* h += m[i] */
 145                 h0 += (le32_to_cpuvp(src +  0) >> 0) & 0x3ffffff;
 146                 h1 += (le32_to_cpuvp(src +  3) >> 2) & 0x3ffffff;
 147                 h2 += (le32_to_cpuvp(src +  6) >> 4) & 0x3ffffff;
 148                 h3 += (le32_to_cpuvp(src +  9) >> 6) & 0x3ffffff;
 149                 h4 += (le32_to_cpuvp(src + 12) >> 8) | hibit;
 150
 151                 /* h *= r */
 152                 d0 = mlt(h0, r0) + mlt(h1, s4) + mlt(h2, s3) +
 153                      mlt(h3, s2) + mlt(h4, s1);
 154                 d1 = mlt(h0, r1) + mlt(h1, r0) + mlt(h2, s4) +
 155                      mlt(h3, s3) + mlt(h4, s2);
 156                 d2 = mlt(h0, r2) + mlt(h1, r1) + mlt(h2, r0) +
 157                      mlt(h3, s4) + mlt(h4, s3);
 158                 d3 = mlt(h0, r3) + mlt(h1, r2) + mlt(h2, r1) +
 159                      mlt(h3, r0) + mlt(h4, s4);
 160                 d4 = mlt(h0, r4) + mlt(h1, r3) + mlt(h2, r2) +
 161                      mlt(h3, r1) + mlt(h4, r0);
 162
 163                 /* (partial) h %= p */
 164                 d1 += sr(d0, 26);     h0 = and(d0, 0x3ffffff);
 165                 d2 += sr(d1, 26);     h1 = and(d1, 0x3ffffff);
 166                 d3 += sr(d2, 26);     h2 = and(d2, 0x3ffffff);
 167                 d4 += sr(d3, 26);     h3 = and(d3, 0x3ffffff);
 168                 h0 += sr(d4, 26) * 5; h4 = and(d4, 0x3ffffff);
 169                 h1 += h0 >> 26;       h0 = h0 & 0x3ffffff;
 170
 171                 src += POLY1305_BLOCK_SIZE;
 172                 srclen -= POLY1305_BLOCK_SIZE;
 173         }
 174
 175         dctx->h[0] = h0;
 176         dctx->h[1] = h1;
 177         dctx->h[2] = h2;
 178         dctx->h[3] = h3;
 179         dctx->h[4] = h4;
 180
 181         return srclen;
 182 }
 183
 184 static int poly1305_update(struct shash_desc *desc,
 185                            const u8 *src, unsigned int srclen)
 186 {
 187         struct poly1305_desc_ctx *dctx = shash_desc_ctx(desc);
 188         unsigned int bytes;
 189
 190         if (unlikely(dctx->buflen)) {
 191                 bytes = min(srclen, POLY1305_BLOCK_SIZE - dctx->buflen);
 192                 memcpy(dctx->buf + dctx->buflen, src, bytes);
 193                 src += bytes;
 194                 srclen -= bytes;
 195                 dctx->buflen += bytes;
 196
 197                 if (dctx->buflen == POLY1305_BLOCK_SIZE) {
 198                         poly1305_blocks(dctx, dctx->buf,
 199                                         POLY1305_BLOCK_SIZE, 1 << 24);
 200                         dctx->buflen = 0;
 201                 }
 202         }
 203
 204         if (likely(srclen >= POLY1305_BLOCK_SIZE)) {
 205                 bytes = poly1305_blocks(dctx, src, srclen, 1 << 24);
 206                 src += srclen - bytes;
 207                 srclen = bytes;
 208         }
 209
 210         if (unlikely(srclen)) {
 211                 dctx->buflen = srclen;
 212                 memcpy(dctx->buf, src, srclen);
 213         }
 214
 215         return 0;
 216 }
 217
 218 static int poly1305_final(struct shash_desc *desc, u8 *dst)
 219 {
 220         struct poly1305_desc_ctx *dctx = shash_desc_ctx(desc);
 221         __le32 *mac = (__le32 *)dst;
 222         u32 h0, h1, h2, h3, h4;
 223         u32 g0, g1, g2, g3, g4;
 224         u32 mask;
 225         u64 f = 0;
 226
 227         if (unlikely(!dctx->sset))
 228                 return -ENOKEY;
 229
 230         if (unlikely(dctx->buflen)) {
 231                 dctx->buf[dctx->buflen++] = 1;
 232                 memset(dctx->buf + dctx->buflen, 0,
 233                        POLY1305_BLOCK_SIZE - dctx->buflen);
 234                 poly1305_blocks(dctx, dctx->buf, POLY1305_BLOCK_SIZE, 0);
 235         }
 236
 237         /* fully carry h */
 238         h0 = dctx->h[0];
 239         h1 = dctx->h[1];
 240         h2 = dctx->h[2];
 241         h3 = dctx->h[3];
 242         h4 = dctx->h[4];
 243
 244         h2 += (h1 >> 26);     h1 = h1 & 0x3ffffff;
 245         h3 += (h2 >> 26);     h2 = h2 & 0x3ffffff;
 246         h4 += (h3 >> 26);     h3 = h3 & 0x3ffffff;
 247         h0 += (h4 >> 26) * 5; h4 = h4 & 0x3ffffff;
 248         h1 += (h0 >> 26);     h0 = h0 & 0x3ffffff;
 249
 250         /* compute h + -p */
 251         g0 = h0 + 5;
 252         g1 = h1 + (g0 >> 26);             g0 &= 0x3ffffff;
 253         g2 = h2 + (g1 >> 26);             g1 &= 0x3ffffff;
 254         g3 = h3 + (g2 >> 26);             g2 &= 0x3ffffff;
 255         g4 = h4 + (g3 >> 26) - (1 << 26); g3 &= 0x3ffffff;
 256
 257         /* select h if h < p, or h + -p if h >= p */
 258         mask = (g4 >> ((sizeof(u32) * 8) - 1)) - 1;
 259         g0 &= mask;
 260         g1 &= mask;
 261         g2 &= mask;
 262         g3 &= mask;
 263         g4 &= mask;
 264         mask = ~mask;
 265         h0 = (h0 & mask) | g0;
 266         h1 = (h1 & mask) | g1;
 267         h2 = (h2 & mask) | g2;
 268         h3 = (h3 & mask) | g3;
 269         h4 = (h4 & mask) | g4;
 270
 271         /* h = h % (2^128) */
 272         h0 = (h0 >>  0) | (h1 << 26);
 273         h1 = (h1 >>  6) | (h2 << 20);
 274         h2 = (h2 >> 12) | (h3 << 14);
 275         h3 = (h3 >> 18) | (h4 <<  8);
 276
 277         /* mac = (h + s) % (2^128) */
 278         f = (f >> 32) + h0 + dctx->s[0]; mac[0] = cpu_to_le32(f);
 279         f = (f >> 32) + h1 + dctx->s[1]; mac[1] = cpu_to_le32(f);
 280         f = (f >> 32) + h2 + dctx->s[2]; mac[2] = cpu_to_le32(f);
 281         f = (f >> 32) + h3 + dctx->s[3]; mac[3] = cpu_to_le32(f);
 282
 283         return 0;
 284 }
 285
 286 static struct shash_alg poly1305_alg = {
 287         .digestsize     = POLY1305_DIGEST_SIZE,
 288         .init           = poly1305_init,
 289         .update         = poly1305_update,
 290         .final          = poly1305_final,
 291         .setkey         = poly1305_setkey,
 292         .descsize       = sizeof(struct poly1305_desc_ctx),
 293         .base           = {
 294                 .cra_name               = "poly1305",
 295                 .cra_driver_name        = "poly1305-generic",
 296                 .cra_priority           = 100,
 297                 .cra_flags              = CRYPTO_ALG_TYPE_SHASH,
 298                 .cra_alignmask          = sizeof(u32) - 1,
 299                 .cra_blocksize          = POLY1305_BLOCK_SIZE,
 300                 .cra_module             = THIS_MODULE,
 301         },
 302 };
 303
 304 static int __init poly1305_mod_init(void)
 305 {
 306         return crypto_register_shash(&poly1305_alg);
 307 }
 308
 309 static void __exit poly1305_mod_exit(void)
 310 {
 311         crypto_unregister_shash(&poly1305_alg);
 312 }
 313
 314 module_init(poly1305_mod_init);
 315 module_exit(poly1305_mod_exit);
 316
 317 MODULE_LICENSE("GPL");
 318 MODULE_AUTHOR("Martin Willi <martin@strongswan.org>");
 319 MODULE_DESCRIPTION("Poly1305 authenticator");
 320 MODULE_ALIAS_CRYPTO("poly1305");
 321 MODULE_ALIAS_CRYPTO("poly1305-generic");