2 * Copyright 1995-2018 The OpenSSL Project Authors. All Rights Reserved.
4 * Licensed under the OpenSSL license (the "License"). You may not use
5 * this file except in compliance with the License. You can obtain a copy
6 * in the file LICENSE in the source distribution or at
7 * https://www.openssl.org/source/license.html
11 * Details about Montgomery multiplication algorithms can be found at
12 * http://security.ece.orst.edu/publications.html, e.g.
13 * http://security.ece.orst.edu/koc/papers/j37acmon.pdf and
14 * sections 3.8 and 4.2 in http://security.ece.orst.edu/koc/papers/r01rsasw.pdf
17 #include "internal/cryptlib.h"
20 #define MONT_WORD /* use the faster word-based algorithm */
23 static int BN_from_montgomery_word(BIGNUM *ret, BIGNUM *r, BN_MONT_CTX *mont);
26 int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
27 BN_MONT_CTX *mont, BN_CTX *ctx)
31 #if defined(OPENSSL_BN_ASM_MONT) && defined(MONT_WORD)
32 int num = mont->N.top;
34 if (num > 1 && a->top == num && b->top == num) {
35 if (bn_wexpand(r, num) == NULL)
37 if (bn_mul_mont(r->d, a->d, b->d, mont->N.d, mont->n0, num)) {
38 r->neg = a->neg ^ b->neg;
47 tmp = BN_CTX_get(ctx);
53 if (!BN_sqr(tmp, a, ctx))
56 if (!BN_mul(tmp, a, b, ctx))
59 /* reduce from aRR to aR */
61 if (!BN_from_montgomery_word(r, tmp, mont))
64 if (!BN_from_montgomery(r, tmp, mont, ctx))
75 static int BN_from_montgomery_word(BIGNUM *ret, BIGNUM *r, BN_MONT_CTX *mont)
78 BN_ULONG *ap, *np, *rp, n0, v, carry;
88 max = (2 * nl); /* carry is stored separately */
89 if (bn_wexpand(r, max) == NULL)
96 /* clear the top words of T */
101 memset(&rp[r->top], 0, sizeof(*rp) * i);
107 * Add multiples of |n| to |r| until R = 2^(nl * BN_BITS2) divides it. On
108 * input, we had |r| < |n| * R, so now |r| < 2 * |n| * R. Note that |r|
109 * includes |carry| which is stored separately.
111 for (carry = 0, i = 0; i < nl; i++, rp++) {
112 v = bn_mul_add_words(rp, np, nl, (rp[0] * n0) & BN_MASK2);
113 v = (v + carry + rp[nl]) & BN_MASK2;
114 carry |= (v != rp[nl]);
115 carry &= (v <= rp[nl]);
119 if (bn_wexpand(ret, nl) == NULL)
127 * Shift |nl| words to divide by R. We have |ap| < 2 * |n|. Note that |ap|
128 * includes |carry| which is stored separately.
133 * |v| is one if |ap| - |np| underflowed or zero if it did not. Note |v|
134 * cannot be -1. That would imply the subtraction did not fit in |nl| words,
135 * and we know at most one subtraction is needed.
137 v = bn_sub_words(rp, ap, np, nl) - carry;
139 for (i = 0; i < nl; i++) {
140 rp[i] = (v & ap[i]) | (~v & rp[i]);
149 #endif /* MONT_WORD */
151 int BN_from_montgomery(BIGNUM *ret, const BIGNUM *a, BN_MONT_CTX *mont,
159 if ((t = BN_CTX_get(ctx)) && BN_copy(t, a))
160 retn = BN_from_montgomery_word(ret, t, mont);
162 #else /* !MONT_WORD */
166 t1 = BN_CTX_get(ctx);
167 t2 = BN_CTX_get(ctx);
173 BN_mask_bits(t1, mont->ri);
175 if (!BN_mul(t2, t1, &mont->Ni, ctx))
177 BN_mask_bits(t2, mont->ri);
179 if (!BN_mul(t1, t2, &mont->N, ctx))
181 if (!BN_add(t2, a, t1))
183 if (!BN_rshift(ret, t2, mont->ri))
186 if (BN_ucmp(ret, &(mont->N)) >= 0) {
187 if (!BN_usub(ret, ret, &(mont->N)))
194 #endif /* MONT_WORD */
198 BN_MONT_CTX *BN_MONT_CTX_new(void)
202 if ((ret = OPENSSL_malloc(sizeof(*ret))) == NULL)
205 BN_MONT_CTX_init(ret);
206 ret->flags = BN_FLG_MALLOCED;
210 void BN_MONT_CTX_init(BN_MONT_CTX *ctx)
216 ctx->n0[0] = ctx->n0[1] = 0;
220 void BN_MONT_CTX_free(BN_MONT_CTX *mont)
224 BN_clear_free(&mont->RR);
225 BN_clear_free(&mont->N);
226 BN_clear_free(&mont->Ni);
227 if (mont->flags & BN_FLG_MALLOCED)
231 int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx)
240 if ((Ri = BN_CTX_get(ctx)) == NULL)
242 R = &(mont->RR); /* grab RR as a temp */
243 if (!BN_copy(&(mont->N), mod))
244 goto err; /* Set N */
245 if (BN_get_flags(mod, BN_FLG_CONSTTIME) != 0)
246 BN_set_flags(&(mont->N), BN_FLG_CONSTTIME);
259 if (BN_get_flags(mod, BN_FLG_CONSTTIME) != 0)
260 BN_set_flags(&tmod, BN_FLG_CONSTTIME);
262 mont->ri = (BN_num_bits(mod) + (BN_BITS2 - 1)) / BN_BITS2 * BN_BITS2;
264 # if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
266 * Only certain BN_BITS2<=32 platforms actually make use of n0[1],
267 * and we could use the #else case (with a shorter R value) for the
268 * others. However, currently only the assembler files do know which
273 if (!(BN_set_bit(R, 2 * BN_BITS2)))
277 if ((buf[0] = mod->d[0]))
279 if ((buf[1] = mod->top > 1 ? mod->d[1] : 0))
282 if ((BN_mod_inverse(Ri, R, &tmod, ctx)) == NULL)
284 if (!BN_lshift(Ri, Ri, 2 * BN_BITS2))
286 if (!BN_is_zero(Ri)) {
287 if (!BN_sub_word(Ri, 1))
289 } else { /* if N mod word size == 1 */
291 if (bn_expand(Ri, (int)sizeof(BN_ULONG) * 2) == NULL)
293 /* Ri-- (mod double word size) */
299 if (!BN_div(Ri, NULL, Ri, &tmod, ctx))
302 * Ni = (R*Ri-1)/N, keep only couple of least significant words:
304 mont->n0[0] = (Ri->top > 0) ? Ri->d[0] : 0;
305 mont->n0[1] = (Ri->top > 1) ? Ri->d[1] : 0;
308 if (!(BN_set_bit(R, BN_BITS2)))
311 buf[0] = mod->d[0]; /* tmod = N mod word size */
313 tmod.top = buf[0] != 0 ? 1 : 0;
314 /* Ri = R^-1 mod N */
315 if ((BN_mod_inverse(Ri, R, &tmod, ctx)) == NULL)
317 if (!BN_lshift(Ri, Ri, BN_BITS2))
319 if (!BN_is_zero(Ri)) {
320 if (!BN_sub_word(Ri, 1))
322 } else { /* if N mod word size == 1 */
324 if (!BN_set_word(Ri, BN_MASK2))
325 goto err; /* Ri-- (mod word size) */
327 if (!BN_div(Ri, NULL, Ri, &tmod, ctx))
330 * Ni = (R*Ri-1)/N, keep only least significant word:
332 mont->n0[0] = (Ri->top > 0) ? Ri->d[0] : 0;
336 #else /* !MONT_WORD */
337 { /* bignum version */
338 mont->ri = BN_num_bits(&mont->N);
340 if (!BN_set_bit(R, mont->ri))
341 goto err; /* R = 2^ri */
342 /* Ri = R^-1 mod N */
343 if ((BN_mod_inverse(Ri, R, &mont->N, ctx)) == NULL)
345 if (!BN_lshift(Ri, Ri, mont->ri))
347 if (!BN_sub_word(Ri, 1))
352 if (!BN_div(&(mont->Ni), NULL, Ri, &mont->N, ctx))
357 /* setup RR for conversions */
358 BN_zero(&(mont->RR));
359 if (!BN_set_bit(&(mont->RR), mont->ri * 2))
361 if (!BN_mod(&(mont->RR), &(mont->RR), &(mont->N), ctx))
370 BN_MONT_CTX *BN_MONT_CTX_copy(BN_MONT_CTX *to, BN_MONT_CTX *from)
375 if (!BN_copy(&(to->RR), &(from->RR)))
377 if (!BN_copy(&(to->N), &(from->N)))
379 if (!BN_copy(&(to->Ni), &(from->Ni)))
382 to->n0[0] = from->n0[0];
383 to->n0[1] = from->n0[1];
387 BN_MONT_CTX *BN_MONT_CTX_set_locked(BN_MONT_CTX **pmont, CRYPTO_RWLOCK *lock,
388 const BIGNUM *mod, BN_CTX *ctx)
392 CRYPTO_THREAD_read_lock(lock);
394 CRYPTO_THREAD_unlock(lock);
399 * We don't want to serialise globally while doing our lazy-init math in
400 * BN_MONT_CTX_set. That punishes threads that are doing independent
401 * things. Instead, punish the case where more than one thread tries to
402 * lazy-init the same 'pmont', by having each do the lazy-init math work
403 * independently and only use the one from the thread that wins the race
404 * (the losers throw away the work they've done).
406 ret = BN_MONT_CTX_new();
409 if (!BN_MONT_CTX_set(ret, mod, ctx)) {
410 BN_MONT_CTX_free(ret);
414 /* The locked compare-and-set, after the local work is done. */
415 CRYPTO_THREAD_write_lock(lock);
417 BN_MONT_CTX_free(ret);
421 CRYPTO_THREAD_unlock(lock);