crypto/openssh/umac.c

   1 /* $OpenBSD: umac.c,v 1.12 2017/05/31 08:09:45 markus Exp $ */
   2 /* -----------------------------------------------------------------------
   3  *
   4  * umac.c -- C Implementation UMAC Message Authentication
   5  *
   6  * Version 0.93b of rfc4418.txt -- 2006 July 18
   7  *
   8  * For a full description of UMAC message authentication see the UMAC
   9  * world-wide-web page at http://www.cs.ucdavis.edu/~rogaway/umac
  10  * Please report bugs and suggestions to the UMAC webpage.
  11  *
  12  * Copyright (c) 1999-2006 Ted Krovetz
  13  *
  14  * Permission to use, copy, modify, and distribute this software and
  15  * its documentation for any purpose and with or without fee, is hereby
  16  * granted provided that the above copyright notice appears in all copies
  17  * and in supporting documentation, and that the name of the copyright
  18  * holder not be used in advertising or publicity pertaining to
  19  * distribution of the software without specific, written prior permission.
  20  *
  21  * Comments should be directed to Ted Krovetz (tdk@acm.org)
  22  *
  23  * ---------------------------------------------------------------------- */
  24
  25  /* ////////////////////// IMPORTANT NOTES /////////////////////////////////
  26   *
  27   * 1) This version does not work properly on messages larger than 16MB
  28   *
  29   * 2) If you set the switch to use SSE2, then all data must be 16-byte
  30   *    aligned
  31   *
  32   * 3) When calling the function umac(), it is assumed that msg is in
  33   * a writable buffer of length divisible by 32 bytes. The message itself
  34   * does not have to fill the entire buffer, but bytes beyond msg may be
  35   * zeroed.
  36   *
  37   * 4) Three free AES implementations are supported by this implementation of
  38   * UMAC. Paulo Barreto's version is in the public domain and can be found
  39   * at http://www.esat.kuleuven.ac.be/~rijmen/rijndael/ (search for
  40   * "Barreto"). The only two files needed are rijndael-alg-fst.c and
  41   * rijndael-alg-fst.h. Brian Gladman's version is distributed with the GNU
  42   * Public lisence at http://fp.gladman.plus.com/AES/index.htm. It
  43   * includes a fast IA-32 assembly version. The OpenSSL crypo library is
  44   * the third.
  45   *
  46   * 5) With FORCE_C_ONLY flags set to 0, incorrect results are sometimes
  47   * produced under gcc with optimizations set -O3 or higher. Dunno why.
  48   *
  49   /////////////////////////////////////////////////////////////////////// */
  50
  51 /* ---------------------------------------------------------------------- */
  52 /* --- User Switches ---------------------------------------------------- */
  53 /* ---------------------------------------------------------------------- */
  54
  55 #ifndef UMAC_OUTPUT_LEN
  56 #define UMAC_OUTPUT_LEN     8  /* Alowable: 4, 8, 12, 16                  */
  57 #endif
  58
  59 #if UMAC_OUTPUT_LEN != 4 && UMAC_OUTPUT_LEN != 8 && \
  60     UMAC_OUTPUT_LEN != 12 && UMAC_OUTPUT_LEN != 16
  61 # error UMAC_OUTPUT_LEN must be defined to 4, 8, 12 or 16
  62 #endif
  63
  64 /* #define FORCE_C_ONLY        1  ANSI C and 64-bit integers req'd        */
  65 /* #define AES_IMPLEMENTAION   1  1 = OpenSSL, 2 = Barreto, 3 = Gladman   */
  66 /* #define SSE2                0  Is SSE2 is available?                   */
  67 /* #define RUN_TESTS           0  Run basic correctness/speed tests       */
  68 /* #define UMAC_AE_SUPPORT     0  Enable auhthenticated encrytion         */
  69
  70 /* ---------------------------------------------------------------------- */
  71 /* -- Global Includes --------------------------------------------------- */
  72 /* ---------------------------------------------------------------------- */
  73
  74 #include "includes.h"
  75 #include <sys/types.h>
  76 #include <string.h>
  77 #include <stdio.h>
  78 #include <stdlib.h>
  79 #include <stddef.h>
  80
  81 #include "xmalloc.h"
  82 #include "umac.h"
  83 #include "misc.h"
  84
  85 /* ---------------------------------------------------------------------- */
  86 /* --- Primitive Data Types ---                                           */
  87 /* ---------------------------------------------------------------------- */
  88
  89 /* The following assumptions may need change on your system */
  90 typedef u_int8_t        UINT8;  /* 1 byte   */
  91 typedef u_int16_t       UINT16; /* 2 byte   */
  92 typedef u_int32_t       UINT32; /* 4 byte   */
  93 typedef u_int64_t       UINT64; /* 8 bytes  */
  94 typedef unsigned int    UWORD;  /* Register */
  95
  96 /* ---------------------------------------------------------------------- */
  97 /* --- Constants -------------------------------------------------------- */
  98 /* ---------------------------------------------------------------------- */
  99
 100 #define UMAC_KEY_LEN           16  /* UMAC takes 16 bytes of external key */
 101
 102 /* Message "words" are read from memory in an endian-specific manner.     */
 103 /* For this implementation to behave correctly, __LITTLE_ENDIAN__ must    */
 104 /* be set true if the host computer is little-endian.                     */
 105
 106 #if BYTE_ORDER == LITTLE_ENDIAN
 107 #define __LITTLE_ENDIAN__ 1
 108 #else
 109 #define __LITTLE_ENDIAN__ 0
 110 #endif
 111
 112 /* ---------------------------------------------------------------------- */
 113 /* ---------------------------------------------------------------------- */
 114 /* ----- Architecture Specific ------------------------------------------ */
 115 /* ---------------------------------------------------------------------- */
 116 /* ---------------------------------------------------------------------- */
 117
 118
 119 /* ---------------------------------------------------------------------- */
 120 /* ---------------------------------------------------------------------- */
 121 /* ----- Primitive Routines --------------------------------------------- */
 122 /* ---------------------------------------------------------------------- */
 123 /* ---------------------------------------------------------------------- */
 124
 125
 126 /* ---------------------------------------------------------------------- */
 127 /* --- 32-bit by 32-bit to 64-bit Multiplication ------------------------ */
 128 /* ---------------------------------------------------------------------- */
 129
 130 #define MUL64(a,b) ((UINT64)((UINT64)(UINT32)(a) * (UINT64)(UINT32)(b)))
 131
 132 /* ---------------------------------------------------------------------- */
 133 /* --- Endian Conversion --- Forcing assembly on some platforms           */
 134 /* ---------------------------------------------------------------------- */
 135
 136 #if (__LITTLE_ENDIAN__)
 137 #define LOAD_UINT32_REVERSED(p)         get_u32(p)
 138 #define STORE_UINT32_REVERSED(p,v)      put_u32(p,v)
 139 #else
 140 #define LOAD_UINT32_REVERSED(p)         get_u32_le(p)
 141 #define STORE_UINT32_REVERSED(p,v)      put_u32_le(p,v)
 142 #endif
 143
 144 #define LOAD_UINT32_LITTLE(p)           (get_u32_le(p))
 145 #define STORE_UINT32_BIG(p,v)           put_u32(p, v)
 146
 147 /* ---------------------------------------------------------------------- */
 148 /* ---------------------------------------------------------------------- */
 149 /* ----- Begin KDF & PDF Section ---------------------------------------- */
 150 /* ---------------------------------------------------------------------- */
 151 /* ---------------------------------------------------------------------- */
 152
 153 /* UMAC uses AES with 16 byte block and key lengths */
 154 #define AES_BLOCK_LEN  16
 155
 156 /* OpenSSL's AES */
 157 #ifdef WITH_OPENSSL
 158 #include "openbsd-compat/openssl-compat.h"
 159 #ifndef USE_BUILTIN_RIJNDAEL
 160 # include <openssl/aes.h>
 161 #endif
 162 typedef AES_KEY aes_int_key[1];
 163 #define aes_encryption(in,out,int_key)                  \
 164   AES_encrypt((u_char *)(in),(u_char *)(out),(AES_KEY *)int_key)
 165 #define aes_key_setup(key,int_key)                      \
 166   AES_set_encrypt_key((const u_char *)(key),UMAC_KEY_LEN*8,int_key)
 167 #else
 168 #include "rijndael.h"
 169 #define AES_ROUNDS ((UMAC_KEY_LEN / 4) + 6)
 170 typedef UINT8 aes_int_key[AES_ROUNDS+1][4][4];  /* AES internal */
 171 #define aes_encryption(in,out,int_key) \
 172   rijndaelEncrypt((u32 *)(int_key), AES_ROUNDS, (u8 *)(in), (u8 *)(out))
 173 #define aes_key_setup(key,int_key) \
 174   rijndaelKeySetupEnc((u32 *)(int_key), (const unsigned char *)(key), \
 175   UMAC_KEY_LEN*8)
 176 #endif
 177
 178 /* The user-supplied UMAC key is stretched using AES in a counter
 179  * mode to supply all random bits needed by UMAC. The kdf function takes
 180  * an AES internal key representation 'key' and writes a stream of
 181  * 'nbytes' bytes to the memory pointed at by 'bufp'. Each distinct
 182  * 'ndx' causes a distinct byte stream.
 183  */
 184 static void kdf(void *bufp, aes_int_key key, UINT8 ndx, int nbytes)
 185 {
 186     UINT8 in_buf[AES_BLOCK_LEN] = {0};
 187     UINT8 out_buf[AES_BLOCK_LEN];
 188     UINT8 *dst_buf = (UINT8 *)bufp;
 189     int i;
 190
 191     /* Setup the initial value */
 192     in_buf[AES_BLOCK_LEN-9] = ndx;
 193     in_buf[AES_BLOCK_LEN-1] = i = 1;
 194
 195     while (nbytes >= AES_BLOCK_LEN) {
 196         aes_encryption(in_buf, out_buf, key);
 197         memcpy(dst_buf,out_buf,AES_BLOCK_LEN);
 198         in_buf[AES_BLOCK_LEN-1] = ++i;
 199         nbytes -= AES_BLOCK_LEN;
 200         dst_buf += AES_BLOCK_LEN;
 201     }
 202     if (nbytes) {
 203         aes_encryption(in_buf, out_buf, key);
 204         memcpy(dst_buf,out_buf,nbytes);
 205     }
 206     explicit_bzero(in_buf, sizeof(in_buf));
 207     explicit_bzero(out_buf, sizeof(out_buf));
 208 }
 209
 210 /* The final UHASH result is XOR'd with the output of a pseudorandom
 211  * function. Here, we use AES to generate random output and
 212  * xor the appropriate bytes depending on the last bits of nonce.
 213  * This scheme is optimized for sequential, increasing big-endian nonces.
 214  */
 215
 216 typedef struct {
 217     UINT8 cache[AES_BLOCK_LEN];  /* Previous AES output is saved      */
 218     UINT8 nonce[AES_BLOCK_LEN];  /* The AES input making above cache  */
 219     aes_int_key prf_key;         /* Expanded AES key for PDF          */
 220 } pdf_ctx;
 221
 222 static void pdf_init(pdf_ctx *pc, aes_int_key prf_key)
 223 {
 224     UINT8 buf[UMAC_KEY_LEN];
 225
 226     kdf(buf, prf_key, 0, UMAC_KEY_LEN);
 227     aes_key_setup(buf, pc->prf_key);
 228
 229     /* Initialize pdf and cache */
 230     memset(pc->nonce, 0, sizeof(pc->nonce));
 231     aes_encryption(pc->nonce, pc->cache, pc->prf_key);
 232     explicit_bzero(buf, sizeof(buf));
 233 }
 234
 235 static void pdf_gen_xor(pdf_ctx *pc, const UINT8 nonce[8], UINT8 buf[8])
 236 {
 237     /* 'ndx' indicates that we'll be using the 0th or 1st eight bytes
 238      * of the AES output. If last time around we returned the ndx-1st
 239      * element, then we may have the result in the cache already.
 240      */
 241
 242 #if (UMAC_OUTPUT_LEN == 4)
 243 #define LOW_BIT_MASK 3
 244 #elif (UMAC_OUTPUT_LEN == 8)
 245 #define LOW_BIT_MASK 1
 246 #elif (UMAC_OUTPUT_LEN > 8)
 247 #define LOW_BIT_MASK 0
 248 #endif
 249     union {
 250         UINT8 tmp_nonce_lo[4];
 251         UINT32 align;
 252     } t;
 253 #if LOW_BIT_MASK != 0
 254     int ndx = nonce[7] & LOW_BIT_MASK;
 255 #endif
 256     *(UINT32 *)t.tmp_nonce_lo = ((const UINT32 *)nonce)[1];
 257     t.tmp_nonce_lo[3] &= ~LOW_BIT_MASK; /* zero last bit */
 258
 259     if ( (((UINT32 *)t.tmp_nonce_lo)[0] != ((UINT32 *)pc->nonce)[1]) ||
 260          (((const UINT32 *)nonce)[0] != ((UINT32 *)pc->nonce)[0]) )
 261     {
 262         ((UINT32 *)pc->nonce)[0] = ((const UINT32 *)nonce)[0];
 263         ((UINT32 *)pc->nonce)[1] = ((UINT32 *)t.tmp_nonce_lo)[0];
 264         aes_encryption(pc->nonce, pc->cache, pc->prf_key);
 265     }
 266
 267 #if (UMAC_OUTPUT_LEN == 4)
 268     *((UINT32 *)buf) ^= ((UINT32 *)pc->cache)[ndx];
 269 #elif (UMAC_OUTPUT_LEN == 8)
 270     *((UINT64 *)buf) ^= ((UINT64 *)pc->cache)[ndx];
 271 #elif (UMAC_OUTPUT_LEN == 12)
 272     ((UINT64 *)buf)[0] ^= ((UINT64 *)pc->cache)[0];
 273     ((UINT32 *)buf)[2] ^= ((UINT32 *)pc->cache)[2];
 274 #elif (UMAC_OUTPUT_LEN == 16)
 275     ((UINT64 *)buf)[0] ^= ((UINT64 *)pc->cache)[0];
 276     ((UINT64 *)buf)[1] ^= ((UINT64 *)pc->cache)[1];
 277 #endif
 278 }
 279
 280 /* ---------------------------------------------------------------------- */
 281 /* ---------------------------------------------------------------------- */
 282 /* ----- Begin NH Hash Section ------------------------------------------ */
 283 /* ---------------------------------------------------------------------- */
 284 /* ---------------------------------------------------------------------- */
 285
 286 /* The NH-based hash functions used in UMAC are described in the UMAC paper
 287  * and specification, both of which can be found at the UMAC website.
 288  * The interface to this implementation has two
 289  * versions, one expects the entire message being hashed to be passed
 290  * in a single buffer and returns the hash result immediately. The second
 291  * allows the message to be passed in a sequence of buffers. In the
 292  * muliple-buffer interface, the client calls the routine nh_update() as
 293  * many times as necessary. When there is no more data to be fed to the
 294  * hash, the client calls nh_final() which calculates the hash output.
 295  * Before beginning another hash calculation the nh_reset() routine
 296  * must be called. The single-buffer routine, nh(), is equivalent to
 297  * the sequence of calls nh_update() and nh_final(); however it is
 298  * optimized and should be prefered whenever the multiple-buffer interface
 299  * is not necessary. When using either interface, it is the client's
 300  * responsability to pass no more than L1_KEY_LEN bytes per hash result.
 301  *
 302  * The routine nh_init() initializes the nh_ctx data structure and
 303  * must be called once, before any other PDF routine.
 304  */
 305
 306  /* The "nh_aux" routines do the actual NH hashing work. They
 307   * expect buffers to be multiples of L1_PAD_BOUNDARY. These routines
 308   * produce output for all STREAMS NH iterations in one call,
 309   * allowing the parallel implementation of the streams.
 310   */
 311
 312 #define STREAMS (UMAC_OUTPUT_LEN / 4) /* Number of times hash is applied  */
 313 #define L1_KEY_LEN         1024     /* Internal key bytes                 */
 314 #define L1_KEY_SHIFT         16     /* Toeplitz key shift between streams */
 315 #define L1_PAD_BOUNDARY      32     /* pad message to boundary multiple   */
 316 #define ALLOC_BOUNDARY       16     /* Keep buffers aligned to this       */
 317 #define HASH_BUF_BYTES       64     /* nh_aux_hb buffer multiple          */
 318
 319 typedef struct {
 320     UINT8  nh_key [L1_KEY_LEN + L1_KEY_SHIFT * (STREAMS - 1)]; /* NH Key */
 321     UINT8  data   [HASH_BUF_BYTES];    /* Incoming data buffer           */
 322     int next_data_empty;    /* Bookeeping variable for data buffer.       */
 323     int bytes_hashed;        /* Bytes (out of L1_KEY_LEN) incorperated.   */
 324     UINT64 state[STREAMS];               /* on-line state     */
 325 } nh_ctx;
 326
 327
 328 #if (UMAC_OUTPUT_LEN == 4)
 329
 330 static void nh_aux(void *kp, const void *dp, void *hp, UINT32 dlen)
 331 /* NH hashing primitive. Previous (partial) hash result is loaded and
 332 * then stored via hp pointer. The length of the data pointed at by "dp",
 333 * "dlen", is guaranteed to be divisible by L1_PAD_BOUNDARY (32).  Key
 334 * is expected to be endian compensated in memory at key setup.
 335 */
 336 {
 337     UINT64 h;
 338     UWORD c = dlen / 32;
 339     UINT32 *k = (UINT32 *)kp;
 340     const UINT32 *d = (const UINT32 *)dp;
 341     UINT32 d0,d1,d2,d3,d4,d5,d6,d7;
 342     UINT32 k0,k1,k2,k3,k4,k5,k6,k7;
 343
 344     h = *((UINT64 *)hp);
 345     do {
 346         d0 = LOAD_UINT32_LITTLE(d+0); d1 = LOAD_UINT32_LITTLE(d+1);
 347         d2 = LOAD_UINT32_LITTLE(d+2); d3 = LOAD_UINT32_LITTLE(d+3);
 348         d4 = LOAD_UINT32_LITTLE(d+4); d5 = LOAD_UINT32_LITTLE(d+5);
 349         d6 = LOAD_UINT32_LITTLE(d+6); d7 = LOAD_UINT32_LITTLE(d+7);
 350         k0 = *(k+0); k1 = *(k+1); k2 = *(k+2); k3 = *(k+3);
 351         k4 = *(k+4); k5 = *(k+5); k6 = *(k+6); k7 = *(k+7);
 352         h += MUL64((k0 + d0), (k4 + d4));
 353         h += MUL64((k1 + d1), (k5 + d5));
 354         h += MUL64((k2 + d2), (k6 + d6));
 355         h += MUL64((k3 + d3), (k7 + d7));
 356
 357         d += 8;
 358         k += 8;
 359     } while (--c);
 360   *((UINT64 *)hp) = h;
 361 }
 362
 363 #elif (UMAC_OUTPUT_LEN == 8)
 364
 365 static void nh_aux(void *kp, const void *dp, void *hp, UINT32 dlen)
 366 /* Same as previous nh_aux, but two streams are handled in one pass,
 367  * reading and writing 16 bytes of hash-state per call.
 368  */
 369 {
 370   UINT64 h1,h2;
 371   UWORD c = dlen / 32;
 372   UINT32 *k = (UINT32 *)kp;
 373   const UINT32 *d = (const UINT32 *)dp;
 374   UINT32 d0,d1,d2,d3,d4,d5,d6,d7;
 375   UINT32 k0,k1,k2,k3,k4,k5,k6,k7,
 376         k8,k9,k10,k11;
 377
 378   h1 = *((UINT64 *)hp);
 379   h2 = *((UINT64 *)hp + 1);
 380   k0 = *(k+0); k1 = *(k+1); k2 = *(k+2); k3 = *(k+3);
 381   do {
 382     d0 = LOAD_UINT32_LITTLE(d+0); d1 = LOAD_UINT32_LITTLE(d+1);
 383     d2 = LOAD_UINT32_LITTLE(d+2); d3 = LOAD_UINT32_LITTLE(d+3);
 384     d4 = LOAD_UINT32_LITTLE(d+4); d5 = LOAD_UINT32_LITTLE(d+5);
 385     d6 = LOAD_UINT32_LITTLE(d+6); d7 = LOAD_UINT32_LITTLE(d+7);
 386     k4 = *(k+4); k5 = *(k+5); k6 = *(k+6); k7 = *(k+7);
 387     k8 = *(k+8); k9 = *(k+9); k10 = *(k+10); k11 = *(k+11);
 388
 389     h1 += MUL64((k0 + d0), (k4 + d4));
 390     h2 += MUL64((k4 + d0), (k8 + d4));
 391
 392     h1 += MUL64((k1 + d1), (k5 + d5));
 393     h2 += MUL64((k5 + d1), (k9 + d5));
 394
 395     h1 += MUL64((k2 + d2), (k6 + d6));
 396     h2 += MUL64((k6 + d2), (k10 + d6));
 397
 398     h1 += MUL64((k3 + d3), (k7 + d7));
 399     h2 += MUL64((k7 + d3), (k11 + d7));
 400
 401     k0 = k8; k1 = k9; k2 = k10; k3 = k11;
 402
 403     d += 8;
 404     k += 8;
 405   } while (--c);
 406   ((UINT64 *)hp)[0] = h1;
 407   ((UINT64 *)hp)[1] = h2;
 408 }
 409
 410 #elif (UMAC_OUTPUT_LEN == 12)
 411
 412 static void nh_aux(void *kp, const void *dp, void *hp, UINT32 dlen)
 413 /* Same as previous nh_aux, but two streams are handled in one pass,
 414  * reading and writing 24 bytes of hash-state per call.
 415 */
 416 {
 417     UINT64 h1,h2,h3;
 418     UWORD c = dlen / 32;
 419     UINT32 *k = (UINT32 *)kp;
 420     const UINT32 *d = (const UINT32 *)dp;
 421     UINT32 d0,d1,d2,d3,d4,d5,d6,d7;
 422     UINT32 k0,k1,k2,k3,k4,k5,k6,k7,
 423         k8,k9,k10,k11,k12,k13,k14,k15;
 424
 425     h1 = *((UINT64 *)hp);
 426     h2 = *((UINT64 *)hp + 1);
 427     h3 = *((UINT64 *)hp + 2);
 428     k0 = *(k+0); k1 = *(k+1); k2 = *(k+2); k3 = *(k+3);
 429     k4 = *(k+4); k5 = *(k+5); k6 = *(k+6); k7 = *(k+7);
 430     do {
 431         d0 = LOAD_UINT32_LITTLE(d+0); d1 = LOAD_UINT32_LITTLE(d+1);
 432         d2 = LOAD_UINT32_LITTLE(d+2); d3 = LOAD_UINT32_LITTLE(d+3);
 433         d4 = LOAD_UINT32_LITTLE(d+4); d5 = LOAD_UINT32_LITTLE(d+5);
 434         d6 = LOAD_UINT32_LITTLE(d+6); d7 = LOAD_UINT32_LITTLE(d+7);
 435         k8 = *(k+8); k9 = *(k+9); k10 = *(k+10); k11 = *(k+11);
 436         k12 = *(k+12); k13 = *(k+13); k14 = *(k+14); k15 = *(k+15);
 437
 438         h1 += MUL64((k0 + d0), (k4 + d4));
 439         h2 += MUL64((k4 + d0), (k8 + d4));
 440         h3 += MUL64((k8 + d0), (k12 + d4));
 441
 442         h1 += MUL64((k1 + d1), (k5 + d5));
 443         h2 += MUL64((k5 + d1), (k9 + d5));
 444         h3 += MUL64((k9 + d1), (k13 + d5));
 445
 446         h1 += MUL64((k2 + d2), (k6 + d6));
 447         h2 += MUL64((k6 + d2), (k10 + d6));
 448         h3 += MUL64((k10 + d2), (k14 + d6));
 449
 450         h1 += MUL64((k3 + d3), (k7 + d7));
 451         h2 += MUL64((k7 + d3), (k11 + d7));
 452         h3 += MUL64((k11 + d3), (k15 + d7));
 453
 454         k0 = k8; k1 = k9; k2 = k10; k3 = k11;
 455         k4 = k12; k5 = k13; k6 = k14; k7 = k15;
 456
 457         d += 8;
 458         k += 8;
 459     } while (--c);
 460     ((UINT64 *)hp)[0] = h1;
 461     ((UINT64 *)hp)[1] = h2;
 462     ((UINT64 *)hp)[2] = h3;
 463 }
 464
 465 #elif (UMAC_OUTPUT_LEN == 16)
 466
 467 static void nh_aux(void *kp, const void *dp, void *hp, UINT32 dlen)
 468 /* Same as previous nh_aux, but two streams are handled in one pass,
 469  * reading and writing 24 bytes of hash-state per call.
 470 */
 471 {
 472     UINT64 h1,h2,h3,h4;
 473     UWORD c = dlen / 32;
 474     UINT32 *k = (UINT32 *)kp;
 475     const UINT32 *d = (const UINT32 *)dp;
 476     UINT32 d0,d1,d2,d3,d4,d5,d6,d7;
 477     UINT32 k0,k1,k2,k3,k4,k5,k6,k7,
 478         k8,k9,k10,k11,k12,k13,k14,k15,
 479         k16,k17,k18,k19;
 480
 481     h1 = *((UINT64 *)hp);
 482     h2 = *((UINT64 *)hp + 1);
 483     h3 = *((UINT64 *)hp + 2);
 484     h4 = *((UINT64 *)hp + 3);
 485     k0 = *(k+0); k1 = *(k+1); k2 = *(k+2); k3 = *(k+3);
 486     k4 = *(k+4); k5 = *(k+5); k6 = *(k+6); k7 = *(k+7);
 487     do {
 488         d0 = LOAD_UINT32_LITTLE(d+0); d1 = LOAD_UINT32_LITTLE(d+1);
 489         d2 = LOAD_UINT32_LITTLE(d+2); d3 = LOAD_UINT32_LITTLE(d+3);
 490         d4 = LOAD_UINT32_LITTLE(d+4); d5 = LOAD_UINT32_LITTLE(d+5);
 491         d6 = LOAD_UINT32_LITTLE(d+6); d7 = LOAD_UINT32_LITTLE(d+7);
 492         k8 = *(k+8); k9 = *(k+9); k10 = *(k+10); k11 = *(k+11);
 493         k12 = *(k+12); k13 = *(k+13); k14 = *(k+14); k15 = *(k+15);
 494         k16 = *(k+16); k17 = *(k+17); k18 = *(k+18); k19 = *(k+19);
 495
 496         h1 += MUL64((k0 + d0), (k4 + d4));
 497         h2 += MUL64((k4 + d0), (k8 + d4));
 498         h3 += MUL64((k8 + d0), (k12 + d4));
 499         h4 += MUL64((k12 + d0), (k16 + d4));
 500
 501         h1 += MUL64((k1 + d1), (k5 + d5));
 502         h2 += MUL64((k5 + d1), (k9 + d5));
 503         h3 += MUL64((k9 + d1), (k13 + d5));
 504         h4 += MUL64((k13 + d1), (k17 + d5));
 505
 506         h1 += MUL64((k2 + d2), (k6 + d6));
 507         h2 += MUL64((k6 + d2), (k10 + d6));
 508         h3 += MUL64((k10 + d2), (k14 + d6));
 509         h4 += MUL64((k14 + d2), (k18 + d6));
 510
 511         h1 += MUL64((k3 + d3), (k7 + d7));
 512         h2 += MUL64((k7 + d3), (k11 + d7));
 513         h3 += MUL64((k11 + d3), (k15 + d7));
 514         h4 += MUL64((k15 + d3), (k19 + d7));
 515
 516         k0 = k8; k1 = k9; k2 = k10; k3 = k11;
 517         k4 = k12; k5 = k13; k6 = k14; k7 = k15;
 518         k8 = k16; k9 = k17; k10 = k18; k11 = k19;
 519
 520         d += 8;
 521         k += 8;
 522     } while (--c);
 523     ((UINT64 *)hp)[0] = h1;
 524     ((UINT64 *)hp)[1] = h2;
 525     ((UINT64 *)hp)[2] = h3;
 526     ((UINT64 *)hp)[3] = h4;
 527 }
 528
 529 /* ---------------------------------------------------------------------- */
 530 #endif  /* UMAC_OUTPUT_LENGTH */
 531 /* ---------------------------------------------------------------------- */
 532
 533
 534 /* ---------------------------------------------------------------------- */
 535
 536 static void nh_transform(nh_ctx *hc, const UINT8 *buf, UINT32 nbytes)
 537 /* This function is a wrapper for the primitive NH hash functions. It takes
 538  * as argument "hc" the current hash context and a buffer which must be a
 539  * multiple of L1_PAD_BOUNDARY. The key passed to nh_aux is offset
 540  * appropriately according to how much message has been hashed already.
 541  */
 542 {
 543     UINT8 *key;
 544
 545     key = hc->nh_key + hc->bytes_hashed;
 546     nh_aux(key, buf, hc->state, nbytes);
 547 }
 548
 549 /* ---------------------------------------------------------------------- */
 550
 551 #if (__LITTLE_ENDIAN__)
 552 static void endian_convert(void *buf, UWORD bpw, UINT32 num_bytes)
 553 /* We endian convert the keys on little-endian computers to               */
 554 /* compensate for the lack of big-endian memory reads during hashing.     */
 555 {
 556     UWORD iters = num_bytes / bpw;
 557     if (bpw == 4) {
 558         UINT32 *p = (UINT32 *)buf;
 559         do {
 560             *p = LOAD_UINT32_REVERSED(p);
 561             p++;
 562         } while (--iters);
 563     } else if (bpw == 8) {
 564         UINT32 *p = (UINT32 *)buf;
 565         UINT32 t;
 566         do {
 567             t = LOAD_UINT32_REVERSED(p+1);
 568             p[1] = LOAD_UINT32_REVERSED(p);
 569             p[0] = t;
 570             p += 2;
 571         } while (--iters);
 572     }
 573 }
 574 #define endian_convert_if_le(x,y,z) endian_convert((x),(y),(z))
 575 #else
 576 #define endian_convert_if_le(x,y,z) do{}while(0)  /* Do nothing */
 577 #endif
 578
 579 /* ---------------------------------------------------------------------- */
 580
 581 static void nh_reset(nh_ctx *hc)
 582 /* Reset nh_ctx to ready for hashing of new data */
 583 {
 584     hc->bytes_hashed = 0;
 585     hc->next_data_empty = 0;
 586     hc->state[0] = 0;
 587 #if (UMAC_OUTPUT_LEN >= 8)
 588     hc->state[1] = 0;
 589 #endif
 590 #if (UMAC_OUTPUT_LEN >= 12)
 591     hc->state[2] = 0;
 592 #endif
 593 #if (UMAC_OUTPUT_LEN == 16)
 594     hc->state[3] = 0;
 595 #endif
 596
 597 }
 598
 599 /* ---------------------------------------------------------------------- */
 600
 601 static void nh_init(nh_ctx *hc, aes_int_key prf_key)
 602 /* Generate nh_key, endian convert and reset to be ready for hashing.   */
 603 {
 604     kdf(hc->nh_key, prf_key, 1, sizeof(hc->nh_key));
 605     endian_convert_if_le(hc->nh_key, 4, sizeof(hc->nh_key));
 606     nh_reset(hc);
 607 }
 608
 609 /* ---------------------------------------------------------------------- */
 610
 611 static void nh_update(nh_ctx *hc, const UINT8 *buf, UINT32 nbytes)
 612 /* Incorporate nbytes of data into a nh_ctx, buffer whatever is not an    */
 613 /* even multiple of HASH_BUF_BYTES.                                       */
 614 {
 615     UINT32 i,j;
 616
 617     j = hc->next_data_empty;
 618     if ((j + nbytes) >= HASH_BUF_BYTES) {
 619         if (j) {
 620             i = HASH_BUF_BYTES - j;
 621             memcpy(hc->data+j, buf, i);
 622             nh_transform(hc,hc->data,HASH_BUF_BYTES);
 623             nbytes -= i;
 624             buf += i;
 625             hc->bytes_hashed += HASH_BUF_BYTES;
 626         }
 627         if (nbytes >= HASH_BUF_BYTES) {
 628             i = nbytes & ~(HASH_BUF_BYTES - 1);
 629             nh_transform(hc, buf, i);
 630             nbytes -= i;
 631             buf += i;
 632             hc->bytes_hashed += i;
 633         }
 634         j = 0;
 635     }
 636     memcpy(hc->data + j, buf, nbytes);
 637     hc->next_data_empty = j + nbytes;
 638 }
 639
 640 /* ---------------------------------------------------------------------- */
 641
 642 static void zero_pad(UINT8 *p, int nbytes)
 643 {
 644 /* Write "nbytes" of zeroes, beginning at "p" */
 645     if (nbytes >= (int)sizeof(UWORD)) {
 646         while ((ptrdiff_t)p % sizeof(UWORD)) {
 647             *p = 0;
 648             nbytes--;
 649             p++;
 650         }
 651         while (nbytes >= (int)sizeof(UWORD)) {
 652             *(UWORD *)p = 0;
 653             nbytes -= sizeof(UWORD);
 654             p += sizeof(UWORD);
 655         }
 656     }
 657     while (nbytes) {
 658         *p = 0;
 659         nbytes--;
 660         p++;
 661     }
 662 }
 663
 664 /* ---------------------------------------------------------------------- */
 665
 666 static void nh_final(nh_ctx *hc, UINT8 *result)
 667 /* After passing some number of data buffers to nh_update() for integration
 668  * into an NH context, nh_final is called to produce a hash result. If any
 669  * bytes are in the buffer hc->data, incorporate them into the
 670  * NH context. Finally, add into the NH accumulation "state" the total number
 671  * of bits hashed. The resulting numbers are written to the buffer "result".
 672  * If nh_update was never called, L1_PAD_BOUNDARY zeroes are incorporated.
 673  */
 674 {
 675     int nh_len, nbits;
 676
 677     if (hc->next_data_empty != 0) {
 678         nh_len = ((hc->next_data_empty + (L1_PAD_BOUNDARY - 1)) &
 679                                                 ~(L1_PAD_BOUNDARY - 1));
 680         zero_pad(hc->data + hc->next_data_empty,
 681                                           nh_len - hc->next_data_empty);
 682         nh_transform(hc, hc->data, nh_len);
 683         hc->bytes_hashed += hc->next_data_empty;
 684     } else if (hc->bytes_hashed == 0) {
 685         nh_len = L1_PAD_BOUNDARY;
 686         zero_pad(hc->data, L1_PAD_BOUNDARY);
 687         nh_transform(hc, hc->data, nh_len);
 688     }
 689
 690     nbits = (hc->bytes_hashed << 3);
 691     ((UINT64 *)result)[0] = ((UINT64 *)hc->state)[0] + nbits;
 692 #if (UMAC_OUTPUT_LEN >= 8)
 693     ((UINT64 *)result)[1] = ((UINT64 *)hc->state)[1] + nbits;
 694 #endif
 695 #if (UMAC_OUTPUT_LEN >= 12)
 696     ((UINT64 *)result)[2] = ((UINT64 *)hc->state)[2] + nbits;
 697 #endif
 698 #if (UMAC_OUTPUT_LEN == 16)
 699     ((UINT64 *)result)[3] = ((UINT64 *)hc->state)[3] + nbits;
 700 #endif
 701     nh_reset(hc);
 702 }
 703
 704 /* ---------------------------------------------------------------------- */
 705
 706 static void nh(nh_ctx *hc, const UINT8 *buf, UINT32 padded_len,
 707                UINT32 unpadded_len, UINT8 *result)
 708 /* All-in-one nh_update() and nh_final() equivalent.
 709  * Assumes that padded_len is divisible by L1_PAD_BOUNDARY and result is
 710  * well aligned
 711  */
 712 {
 713     UINT32 nbits;
 714
 715     /* Initialize the hash state */
 716     nbits = (unpadded_len << 3);
 717
 718     ((UINT64 *)result)[0] = nbits;
 719 #if (UMAC_OUTPUT_LEN >= 8)
 720     ((UINT64 *)result)[1] = nbits;
 721 #endif
 722 #if (UMAC_OUTPUT_LEN >= 12)
 723     ((UINT64 *)result)[2] = nbits;
 724 #endif
 725 #if (UMAC_OUTPUT_LEN == 16)
 726     ((UINT64 *)result)[3] = nbits;
 727 #endif
 728
 729     nh_aux(hc->nh_key, buf, result, padded_len);
 730 }
 731
 732 /* ---------------------------------------------------------------------- */
 733 /* ---------------------------------------------------------------------- */
 734 /* ----- Begin UHASH Section -------------------------------------------- */
 735 /* ---------------------------------------------------------------------- */
 736 /* ---------------------------------------------------------------------- */
 737
 738 /* UHASH is a multi-layered algorithm. Data presented to UHASH is first
 739  * hashed by NH. The NH output is then hashed by a polynomial-hash layer
 740  * unless the initial data to be hashed is short. After the polynomial-
 741  * layer, an inner-product hash is used to produce the final UHASH output.
 742  *
 743  * UHASH provides two interfaces, one all-at-once and another where data
 744  * buffers are presented sequentially. In the sequential interface, the
 745  * UHASH client calls the routine uhash_update() as many times as necessary.
 746  * When there is no more data to be fed to UHASH, the client calls
 747  * uhash_final() which
 748  * calculates the UHASH output. Before beginning another UHASH calculation
 749  * the uhash_reset() routine must be called. The all-at-once UHASH routine,
 750  * uhash(), is equivalent to the sequence of calls uhash_update() and
 751  * uhash_final(); however it is optimized and should be
 752  * used whenever the sequential interface is not necessary.
 753  *
 754  * The routine uhash_init() initializes the uhash_ctx data structure and
 755  * must be called once, before any other UHASH routine.
 756  */
 757
 758 /* ---------------------------------------------------------------------- */
 759 /* ----- Constants and uhash_ctx ---------------------------------------- */
 760 /* ---------------------------------------------------------------------- */
 761
 762 /* ---------------------------------------------------------------------- */
 763 /* ----- Poly hash and Inner-Product hash Constants --------------------- */
 764 /* ---------------------------------------------------------------------- */
 765
 766 /* Primes and masks */
 767 #define p36    ((UINT64)0x0000000FFFFFFFFBull)              /* 2^36 -  5 */
 768 #define p64    ((UINT64)0xFFFFFFFFFFFFFFC5ull)              /* 2^64 - 59 */
 769 #define m36    ((UINT64)0x0000000FFFFFFFFFull)  /* The low 36 of 64 bits */
 770
 771
 772 /* ---------------------------------------------------------------------- */
 773
 774 typedef struct uhash_ctx {
 775     nh_ctx hash;                          /* Hash context for L1 NH hash  */
 776     UINT64 poly_key_8[STREAMS];           /* p64 poly keys                */
 777     UINT64 poly_accum[STREAMS];           /* poly hash result             */
 778     UINT64 ip_keys[STREAMS*4];            /* Inner-product keys           */
 779     UINT32 ip_trans[STREAMS];             /* Inner-product translation    */
 780     UINT32 msg_len;                       /* Total length of data passed  */
 781                                           /* to uhash */
 782 } uhash_ctx;
 783 typedef struct uhash_ctx *uhash_ctx_t;
 784
 785 /* ---------------------------------------------------------------------- */
 786
 787
 788 /* The polynomial hashes use Horner's rule to evaluate a polynomial one
 789  * word at a time. As described in the specification, poly32 and poly64
 790  * require keys from special domains. The following implementations exploit
 791  * the special domains to avoid overflow. The results are not guaranteed to
 792  * be within Z_p32 and Z_p64, but the Inner-Product hash implementation
 793  * patches any errant values.
 794  */
 795
 796 static UINT64 poly64(UINT64 cur, UINT64 key, UINT64 data)
 797 {
 798     UINT32 key_hi = (UINT32)(key >> 32),
 799            key_lo = (UINT32)key,
 800            cur_hi = (UINT32)(cur >> 32),
 801            cur_lo = (UINT32)cur,
 802            x_lo,
 803            x_hi;
 804     UINT64 X,T,res;
 805
 806     X =  MUL64(key_hi, cur_lo) + MUL64(cur_hi, key_lo);
 807     x_lo = (UINT32)X;
 808     x_hi = (UINT32)(X >> 32);
 809
 810     res = (MUL64(key_hi, cur_hi) + x_hi) * 59 + MUL64(key_lo, cur_lo);
 811
 812     T = ((UINT64)x_lo << 32);
 813     res += T;
 814     if (res < T)
 815         res += 59;
 816
 817     res += data;
 818     if (res < data)
 819         res += 59;
 820
 821     return res;
 822 }
 823
 824
 825 /* Although UMAC is specified to use a ramped polynomial hash scheme, this
 826  * implementation does not handle all ramp levels. Because we don't handle
 827  * the ramp up to p128 modulus in this implementation, we are limited to
 828  * 2^14 poly_hash() invocations per stream (for a total capacity of 2^24
 829  * bytes input to UMAC per tag, ie. 16MB).
 830  */
 831 static void poly_hash(uhash_ctx_t hc, UINT32 data_in[])
 832 {
 833     int i;
 834     UINT64 *data=(UINT64*)data_in;
 835
 836     for (i = 0; i < STREAMS; i++) {
 837         if ((UINT32)(data[i] >> 32) == 0xfffffffful) {
 838             hc->poly_accum[i] = poly64(hc->poly_accum[i],
 839                                        hc->poly_key_8[i], p64 - 1);
 840             hc->poly_accum[i] = poly64(hc->poly_accum[i],
 841                                        hc->poly_key_8[i], (data[i] - 59));
 842         } else {
 843             hc->poly_accum[i] = poly64(hc->poly_accum[i],
 844                                        hc->poly_key_8[i], data[i]);
 845         }
 846     }
 847 }
 848
 849
 850 /* ---------------------------------------------------------------------- */
 851
 852
 853 /* The final step in UHASH is an inner-product hash. The poly hash
 854  * produces a result not neccesarily WORD_LEN bytes long. The inner-
 855  * product hash breaks the polyhash output into 16-bit chunks and
 856  * multiplies each with a 36 bit key.
 857  */
 858
 859 static UINT64 ip_aux(UINT64 t, UINT64 *ipkp, UINT64 data)
 860 {
 861     t = t + ipkp[0] * (UINT64)(UINT16)(data >> 48);
 862     t = t + ipkp[1] * (UINT64)(UINT16)(data >> 32);
 863     t = t + ipkp[2] * (UINT64)(UINT16)(data >> 16);
 864     t = t + ipkp[3] * (UINT64)(UINT16)(data);
 865
 866     return t;
 867 }
 868
 869 static UINT32 ip_reduce_p36(UINT64 t)
 870 {
 871 /* Divisionless modular reduction */
 872     UINT64 ret;
 873
 874     ret = (t & m36) + 5 * (t >> 36);
 875     if (ret >= p36)
 876         ret -= p36;
 877
 878     /* return least significant 32 bits */
 879     return (UINT32)(ret);
 880 }
 881
 882
 883 /* If the data being hashed by UHASH is no longer than L1_KEY_LEN, then
 884  * the polyhash stage is skipped and ip_short is applied directly to the
 885  * NH output.
 886  */
 887 static void ip_short(uhash_ctx_t ahc, UINT8 *nh_res, u_char *res)
 888 {
 889     UINT64 t;
 890     UINT64 *nhp = (UINT64 *)nh_res;
 891
 892     t  = ip_aux(0,ahc->ip_keys, nhp[0]);
 893     STORE_UINT32_BIG((UINT32 *)res+0, ip_reduce_p36(t) ^ ahc->ip_trans[0]);
 894 #if (UMAC_OUTPUT_LEN >= 8)
 895     t  = ip_aux(0,ahc->ip_keys+4, nhp[1]);
 896     STORE_UINT32_BIG((UINT32 *)res+1, ip_reduce_p36(t) ^ ahc->ip_trans[1]);
 897 #endif
 898 #if (UMAC_OUTPUT_LEN >= 12)
 899     t  = ip_aux(0,ahc->ip_keys+8, nhp[2]);
 900     STORE_UINT32_BIG((UINT32 *)res+2, ip_reduce_p36(t) ^ ahc->ip_trans[2]);
 901 #endif
 902 #if (UMAC_OUTPUT_LEN == 16)
 903     t  = ip_aux(0,ahc->ip_keys+12, nhp[3]);
 904     STORE_UINT32_BIG((UINT32 *)res+3, ip_reduce_p36(t) ^ ahc->ip_trans[3]);
 905 #endif
 906 }
 907
 908 /* If the data being hashed by UHASH is longer than L1_KEY_LEN, then
 909  * the polyhash stage is not skipped and ip_long is applied to the
 910  * polyhash output.
 911  */
 912 static void ip_long(uhash_ctx_t ahc, u_char *res)
 913 {
 914     int i;
 915     UINT64 t;
 916
 917     for (i = 0; i < STREAMS; i++) {
 918         /* fix polyhash output not in Z_p64 */
 919         if (ahc->poly_accum[i] >= p64)
 920             ahc->poly_accum[i] -= p64;
 921         t  = ip_aux(0,ahc->ip_keys+(i*4), ahc->poly_accum[i]);
 922         STORE_UINT32_BIG((UINT32 *)res+i,
 923                          ip_reduce_p36(t) ^ ahc->ip_trans[i]);
 924     }
 925 }
 926
 927
 928 /* ---------------------------------------------------------------------- */
 929
 930 /* ---------------------------------------------------------------------- */
 931
 932 /* Reset uhash context for next hash session */
 933 static int uhash_reset(uhash_ctx_t pc)
 934 {
 935     nh_reset(&pc->hash);
 936     pc->msg_len = 0;
 937     pc->poly_accum[0] = 1;
 938 #if (UMAC_OUTPUT_LEN >= 8)
 939     pc->poly_accum[1] = 1;
 940 #endif
 941 #if (UMAC_OUTPUT_LEN >= 12)
 942     pc->poly_accum[2] = 1;
 943 #endif
 944 #if (UMAC_OUTPUT_LEN == 16)
 945     pc->poly_accum[3] = 1;
 946 #endif
 947     return 1;
 948 }
 949
 950 /* ---------------------------------------------------------------------- */
 951
 952 /* Given a pointer to the internal key needed by kdf() and a uhash context,
 953  * initialize the NH context and generate keys needed for poly and inner-
 954  * product hashing. All keys are endian adjusted in memory so that native
 955  * loads cause correct keys to be in registers during calculation.
 956  */
 957 static void uhash_init(uhash_ctx_t ahc, aes_int_key prf_key)
 958 {
 959     int i;
 960     UINT8 buf[(8*STREAMS+4)*sizeof(UINT64)];
 961
 962     /* Zero the entire uhash context */
 963     memset(ahc, 0, sizeof(uhash_ctx));
 964
 965     /* Initialize the L1 hash */
 966     nh_init(&ahc->hash, prf_key);
 967
 968     /* Setup L2 hash variables */
 969     kdf(buf, prf_key, 2, sizeof(buf));    /* Fill buffer with index 1 key */
 970     for (i = 0; i < STREAMS; i++) {
 971         /* Fill keys from the buffer, skipping bytes in the buffer not
 972          * used by this implementation. Endian reverse the keys if on a
 973          * little-endian computer.
 974          */
 975         memcpy(ahc->poly_key_8+i, buf+24*i, 8);
 976         endian_convert_if_le(ahc->poly_key_8+i, 8, 8);
 977         /* Mask the 64-bit keys to their special domain */
 978         ahc->poly_key_8[i] &= ((UINT64)0x01ffffffu << 32) + 0x01ffffffu;
 979         ahc->poly_accum[i] = 1;  /* Our polyhash prepends a non-zero word */
 980     }
 981
 982     /* Setup L3-1 hash variables */
 983     kdf(buf, prf_key, 3, sizeof(buf)); /* Fill buffer with index 2 key */
 984     for (i = 0; i < STREAMS; i++)
 985           memcpy(ahc->ip_keys+4*i, buf+(8*i+4)*sizeof(UINT64),
 986                                                  4*sizeof(UINT64));
 987     endian_convert_if_le(ahc->ip_keys, sizeof(UINT64),
 988                                                   sizeof(ahc->ip_keys));
 989     for (i = 0; i < STREAMS*4; i++)
 990         ahc->ip_keys[i] %= p36;  /* Bring into Z_p36 */
 991
 992     /* Setup L3-2 hash variables    */
 993     /* Fill buffer with index 4 key */
 994     kdf(ahc->ip_trans, prf_key, 4, STREAMS * sizeof(UINT32));
 995     endian_convert_if_le(ahc->ip_trans, sizeof(UINT32),
 996                          STREAMS * sizeof(UINT32));
 997     explicit_bzero(buf, sizeof(buf));
 998 }
 999
1000 /* ---------------------------------------------------------------------- */
1001
1002 #if 0
1003 static uhash_ctx_t uhash_alloc(u_char key[])
1004 {
1005 /* Allocate memory and force to a 16-byte boundary. */
1006     uhash_ctx_t ctx;
1007     u_char bytes_to_add;
1008     aes_int_key prf_key;
1009
1010     ctx = (uhash_ctx_t)malloc(sizeof(uhash_ctx)+ALLOC_BOUNDARY);
1011     if (ctx) {
1012         if (ALLOC_BOUNDARY) {
1013             bytes_to_add = ALLOC_BOUNDARY -
1014                               ((ptrdiff_t)ctx & (ALLOC_BOUNDARY -1));
1015             ctx = (uhash_ctx_t)((u_char *)ctx + bytes_to_add);
1016             *((u_char *)ctx - 1) = bytes_to_add;
1017         }
1018         aes_key_setup(key,prf_key);
1019         uhash_init(ctx, prf_key);
1020     }
1021     return (ctx);
1022 }
1023 #endif
1024
1025 /* ---------------------------------------------------------------------- */
1026
1027 #if 0
1028 static int uhash_free(uhash_ctx_t ctx)
1029 {
1030 /* Free memory allocated by uhash_alloc */
1031     u_char bytes_to_sub;
1032
1033     if (ctx) {
1034         if (ALLOC_BOUNDARY) {
1035             bytes_to_sub = *((u_char *)ctx - 1);
1036             ctx = (uhash_ctx_t)((u_char *)ctx - bytes_to_sub);
1037         }
1038         free(ctx);
1039     }
1040     return (1);
1041 }
1042 #endif
1043 /* ---------------------------------------------------------------------- */
1044
1045 static int uhash_update(uhash_ctx_t ctx, const u_char *input, long len)
1046 /* Given len bytes of data, we parse it into L1_KEY_LEN chunks and
1047  * hash each one with NH, calling the polyhash on each NH output.
1048  */
1049 {
1050     UWORD bytes_hashed, bytes_remaining;
1051     UINT64 result_buf[STREAMS];
1052     UINT8 *nh_result = (UINT8 *)&result_buf;
1053
1054     if (ctx->msg_len + len <= L1_KEY_LEN) {
1055         nh_update(&ctx->hash, (const UINT8 *)input, len);
1056         ctx->msg_len += len;
1057     } else {
1058
1059          bytes_hashed = ctx->msg_len % L1_KEY_LEN;
1060          if (ctx->msg_len == L1_KEY_LEN)
1061              bytes_hashed = L1_KEY_LEN;
1062
1063          if (bytes_hashed + len >= L1_KEY_LEN) {
1064
1065              /* If some bytes have been passed to the hash function      */
1066              /* then we want to pass at most (L1_KEY_LEN - bytes_hashed) */
1067              /* bytes to complete the current nh_block.                  */
1068              if (bytes_hashed) {
1069                  bytes_remaining = (L1_KEY_LEN - bytes_hashed);
1070                  nh_update(&ctx->hash, (const UINT8 *)input, bytes_remaining);
1071                  nh_final(&ctx->hash, nh_result);
1072                  ctx->msg_len += bytes_remaining;
1073                  poly_hash(ctx,(UINT32 *)nh_result);
1074                  len -= bytes_remaining;
1075                  input += bytes_remaining;
1076              }
1077
1078              /* Hash directly from input stream if enough bytes */
1079              while (len >= L1_KEY_LEN) {
1080                  nh(&ctx->hash, (const UINT8 *)input, L1_KEY_LEN,
1081                                    L1_KEY_LEN, nh_result);
1082                  ctx->msg_len += L1_KEY_LEN;
1083                  len -= L1_KEY_LEN;
1084                  input += L1_KEY_LEN;
1085                  poly_hash(ctx,(UINT32 *)nh_result);
1086              }
1087          }
1088
1089          /* pass remaining < L1_KEY_LEN bytes of input data to NH */
1090          if (len) {
1091              nh_update(&ctx->hash, (const UINT8 *)input, len);
1092              ctx->msg_len += len;
1093          }
1094      }
1095
1096     return (1);
1097 }
1098
1099 /* ---------------------------------------------------------------------- */
1100
1101 static int uhash_final(uhash_ctx_t ctx, u_char *res)
1102 /* Incorporate any pending data, pad, and generate tag */
1103 {
1104     UINT64 result_buf[STREAMS];
1105     UINT8 *nh_result = (UINT8 *)&result_buf;
1106
1107     if (ctx->msg_len > L1_KEY_LEN) {
1108         if (ctx->msg_len % L1_KEY_LEN) {
1109             nh_final(&ctx->hash, nh_result);
1110             poly_hash(ctx,(UINT32 *)nh_result);
1111         }
1112         ip_long(ctx, res);
1113     } else {
1114         nh_final(&ctx->hash, nh_result);
1115         ip_short(ctx,nh_result, res);
1116     }
1117     uhash_reset(ctx);
1118     return (1);
1119 }
1120
1121 /* ---------------------------------------------------------------------- */
1122
1123 #if 0
1124 static int uhash(uhash_ctx_t ahc, u_char *msg, long len, u_char *res)
1125 /* assumes that msg is in a writable buffer of length divisible by */
1126 /* L1_PAD_BOUNDARY. Bytes beyond msg[len] may be zeroed.           */
1127 {
1128     UINT8 nh_result[STREAMS*sizeof(UINT64)];
1129     UINT32 nh_len;
1130     int extra_zeroes_needed;
1131
1132     /* If the message to be hashed is no longer than L1_HASH_LEN, we skip
1133      * the polyhash.
1134      */
1135     if (len <= L1_KEY_LEN) {
1136         if (len == 0)                  /* If zero length messages will not */
1137                 nh_len = L1_PAD_BOUNDARY;  /* be seen, comment out this case   */
1138         else
1139                 nh_len = ((len + (L1_PAD_BOUNDARY - 1)) & ~(L1_PAD_BOUNDARY - 1));
1140         extra_zeroes_needed = nh_len - len;
1141         zero_pad((UINT8 *)msg + len, extra_zeroes_needed);
1142         nh(&ahc->hash, (UINT8 *)msg, nh_len, len, nh_result);
1143         ip_short(ahc,nh_result, res);
1144     } else {
1145         /* Otherwise, we hash each L1_KEY_LEN chunk with NH, passing the NH
1146          * output to poly_hash().
1147          */
1148         do {
1149             nh(&ahc->hash, (UINT8 *)msg, L1_KEY_LEN, L1_KEY_LEN, nh_result);
1150             poly_hash(ahc,(UINT32 *)nh_result);
1151             len -= L1_KEY_LEN;
1152             msg += L1_KEY_LEN;
1153         } while (len >= L1_KEY_LEN);
1154         if (len) {
1155             nh_len = ((len + (L1_PAD_BOUNDARY - 1)) & ~(L1_PAD_BOUNDARY - 1));
1156             extra_zeroes_needed = nh_len - len;
1157             zero_pad((UINT8 *)msg + len, extra_zeroes_needed);
1158             nh(&ahc->hash, (UINT8 *)msg, nh_len, len, nh_result);
1159             poly_hash(ahc,(UINT32 *)nh_result);
1160         }
1161
1162         ip_long(ahc, res);
1163     }
1164
1165     uhash_reset(ahc);
1166     return 1;
1167 }
1168 #endif
1169
1170 /* ---------------------------------------------------------------------- */
1171 /* ---------------------------------------------------------------------- */
1172 /* ----- Begin UMAC Section --------------------------------------------- */
1173 /* ---------------------------------------------------------------------- */
1174 /* ---------------------------------------------------------------------- */
1175
1176 /* The UMAC interface has two interfaces, an all-at-once interface where
1177  * the entire message to be authenticated is passed to UMAC in one buffer,
1178  * and a sequential interface where the message is presented a little at a
1179  * time. The all-at-once is more optimaized than the sequential version and
1180  * should be preferred when the sequential interface is not required.
1181  */
1182 struct umac_ctx {
1183     uhash_ctx hash;          /* Hash function for message compression    */
1184     pdf_ctx pdf;             /* PDF for hashed output                    */
1185     void *free_ptr;          /* Address to free this struct via          */
1186 } umac_ctx;
1187
1188 /* ---------------------------------------------------------------------- */
1189
1190 #if 0
1191 int umac_reset(struct umac_ctx *ctx)
1192 /* Reset the hash function to begin a new authentication.        */
1193 {
1194     uhash_reset(&ctx->hash);
1195     return (1);
1196 }
1197 #endif
1198
1199 /* ---------------------------------------------------------------------- */
1200
1201 int umac_delete(struct umac_ctx *ctx)
1202 /* Deallocate the ctx structure */
1203 {
1204     if (ctx) {
1205         if (ALLOC_BOUNDARY)
1206             ctx = (struct umac_ctx *)ctx->free_ptr;
1207         explicit_bzero(ctx, sizeof(*ctx) + ALLOC_BOUNDARY);
1208         free(ctx);
1209     }
1210     return (1);
1211 }
1212
1213 /* ---------------------------------------------------------------------- */
1214
1215 struct umac_ctx *umac_new(const u_char key[])
1216 /* Dynamically allocate a umac_ctx struct, initialize variables,
1217  * generate subkeys from key. Align to 16-byte boundary.
1218  */
1219 {
1220     struct umac_ctx *ctx, *octx;
1221     size_t bytes_to_add;
1222     aes_int_key prf_key;
1223
1224     octx = ctx = xcalloc(1, sizeof(*ctx) + ALLOC_BOUNDARY);
1225     if (ctx) {
1226         if (ALLOC_BOUNDARY) {
1227             bytes_to_add = ALLOC_BOUNDARY -
1228                               ((ptrdiff_t)ctx & (ALLOC_BOUNDARY - 1));
1229             ctx = (struct umac_ctx *)((u_char *)ctx + bytes_to_add);
1230         }
1231         ctx->free_ptr = octx;
1232         aes_key_setup(key, prf_key);
1233         pdf_init(&ctx->pdf, prf_key);
1234         uhash_init(&ctx->hash, prf_key);
1235         explicit_bzero(prf_key, sizeof(prf_key));
1236     }
1237
1238     return (ctx);
1239 }
1240
1241 /* ---------------------------------------------------------------------- */
1242
1243 int umac_final(struct umac_ctx *ctx, u_char tag[], const u_char nonce[8])
1244 /* Incorporate any pending data, pad, and generate tag */
1245 {
1246     uhash_final(&ctx->hash, (u_char *)tag);
1247     pdf_gen_xor(&ctx->pdf, (const UINT8 *)nonce, (UINT8 *)tag);
1248
1249     return (1);
1250 }
1251
1252 /* ---------------------------------------------------------------------- */
1253
1254 int umac_update(struct umac_ctx *ctx, const u_char *input, long len)
1255 /* Given len bytes of data, we parse it into L1_KEY_LEN chunks and   */
1256 /* hash each one, calling the PDF on the hashed output whenever the hash- */
1257 /* output buffer is full.                                                 */
1258 {
1259     uhash_update(&ctx->hash, input, len);
1260     return (1);
1261 }
1262
1263 /* ---------------------------------------------------------------------- */
1264
1265 #if 0
1266 int umac(struct umac_ctx *ctx, u_char *input,
1267          long len, u_char tag[],
1268          u_char nonce[8])
1269 /* All-in-one version simply calls umac_update() and umac_final().        */
1270 {
1271     uhash(&ctx->hash, input, len, (u_char *)tag);
1272     pdf_gen_xor(&ctx->pdf, (UINT8 *)nonce, (UINT8 *)tag);
1273
1274     return (1);
1275 }
1276 #endif
1277
1278 /* ---------------------------------------------------------------------- */
1279 /* ---------------------------------------------------------------------- */
1280 /* ----- End UMAC Section ----------------------------------------------- */
1281 /* ---------------------------------------------------------------------- */
1282 /* ---------------------------------------------------------------------- */