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 */
99 memset(&rp[r->top], 0, sizeof(*rp) * i);
105 * Add multiples of |n| to |r| until R = 2^(nl * BN_BITS2) divides it. On
106 * input, we had |r| < |n| * R, so now |r| < 2 * |n| * R. Note that |r|
107 * includes |carry| which is stored separately.
109 for (carry = 0, i = 0; i < nl; i++, rp++) {
110 v = bn_mul_add_words(rp, np, nl, (rp[0] * n0) & BN_MASK2);
111 v = (v + carry + rp[nl]) & BN_MASK2;
112 carry |= (v != rp[nl]);
113 carry &= (v <= rp[nl]);
117 if (bn_wexpand(ret, nl) == NULL)
125 * Shift |nl| words to divide by R. We have |ap| < 2 * |n|. Note that |ap|
126 * includes |carry| which is stored separately.
131 * |v| is one if |ap| - |np| underflowed or zero if it did not. Note |v|
132 * cannot be -1. That would imply the subtraction did not fit in |nl| words,
133 * and we know at most one subtraction is needed.
135 v = bn_sub_words(rp, ap, np, nl) - carry;
137 for (i = 0; i < nl; i++) {
138 rp[i] = (v & ap[i]) | (~v & rp[i]);
147 #endif /* MONT_WORD */
149 int BN_from_montgomery(BIGNUM *ret, const BIGNUM *a, BN_MONT_CTX *mont,
157 if ((t = BN_CTX_get(ctx)) && BN_copy(t, a))
158 retn = BN_from_montgomery_word(ret, t, mont);
160 #else /* !MONT_WORD */
164 t1 = BN_CTX_get(ctx);
165 t2 = BN_CTX_get(ctx);
171 BN_mask_bits(t1, mont->ri);
173 if (!BN_mul(t2, t1, &mont->Ni, ctx))
175 BN_mask_bits(t2, mont->ri);
177 if (!BN_mul(t1, t2, &mont->N, ctx))
179 if (!BN_add(t2, a, t1))
181 if (!BN_rshift(ret, t2, mont->ri))
184 if (BN_ucmp(ret, &(mont->N)) >= 0) {
185 if (!BN_usub(ret, ret, &(mont->N)))
192 #endif /* MONT_WORD */
196 BN_MONT_CTX *BN_MONT_CTX_new(void)
200 if ((ret = OPENSSL_malloc(sizeof(*ret))) == NULL)
203 BN_MONT_CTX_init(ret);
204 ret->flags = BN_FLG_MALLOCED;
208 void BN_MONT_CTX_init(BN_MONT_CTX *ctx)
214 ctx->n0[0] = ctx->n0[1] = 0;
218 void BN_MONT_CTX_free(BN_MONT_CTX *mont)
223 BN_clear_free(&(mont->RR));
224 BN_clear_free(&(mont->N));
225 BN_clear_free(&(mont->Ni));
226 if (mont->flags & BN_FLG_MALLOCED)
230 int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx)
239 if ((Ri = BN_CTX_get(ctx)) == NULL)
241 R = &(mont->RR); /* grab RR as a temp */
242 if (!BN_copy(&(mont->N), mod))
243 goto err; /* Set N */
244 if (BN_get_flags(mod, BN_FLG_CONSTTIME) != 0)
245 BN_set_flags(&(mont->N), BN_FLG_CONSTTIME);
258 if (BN_get_flags(mod, BN_FLG_CONSTTIME) != 0)
259 BN_set_flags(&tmod, BN_FLG_CONSTTIME);
261 mont->ri = (BN_num_bits(mod) + (BN_BITS2 - 1)) / BN_BITS2 * BN_BITS2;
263 # if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
265 * Only certain BN_BITS2<=32 platforms actually make use of n0[1],
266 * and we could use the #else case (with a shorter R value) for the
267 * others. However, currently only the assembler files do know which
272 if (!(BN_set_bit(R, 2 * BN_BITS2)))
276 if ((buf[0] = mod->d[0]))
278 if ((buf[1] = mod->top > 1 ? mod->d[1] : 0))
281 if ((BN_mod_inverse(Ri, R, &tmod, ctx)) == NULL)
283 if (!BN_lshift(Ri, Ri, 2 * BN_BITS2))
285 if (!BN_is_zero(Ri)) {
286 if (!BN_sub_word(Ri, 1))
288 } else { /* if N mod word size == 1 */
290 if (bn_expand(Ri, (int)sizeof(BN_ULONG) * 2) == NULL)
292 /* Ri-- (mod double word size) */
298 if (!BN_div(Ri, NULL, Ri, &tmod, ctx))
301 * Ni = (R*Ri-1)/N, keep only couple of least significant words:
303 mont->n0[0] = (Ri->top > 0) ? Ri->d[0] : 0;
304 mont->n0[1] = (Ri->top > 1) ? Ri->d[1] : 0;
307 if (!(BN_set_bit(R, BN_BITS2)))
310 buf[0] = mod->d[0]; /* tmod = N mod word size */
312 tmod.top = buf[0] != 0 ? 1 : 0;
313 /* Ri = R^-1 mod N */
314 if ((BN_mod_inverse(Ri, R, &tmod, ctx)) == NULL)
316 if (!BN_lshift(Ri, Ri, BN_BITS2))
318 if (!BN_is_zero(Ri)) {
319 if (!BN_sub_word(Ri, 1))
321 } else { /* if N mod word size == 1 */
323 if (!BN_set_word(Ri, BN_MASK2))
324 goto err; /* Ri-- (mod word size) */
326 if (!BN_div(Ri, NULL, Ri, &tmod, ctx))
329 * Ni = (R*Ri-1)/N, keep only least significant word:
331 mont->n0[0] = (Ri->top > 0) ? Ri->d[0] : 0;
335 #else /* !MONT_WORD */
336 { /* bignum version */
337 mont->ri = BN_num_bits(&mont->N);
339 if (!BN_set_bit(R, mont->ri))
340 goto err; /* R = 2^ri */
341 /* Ri = R^-1 mod N */
342 if ((BN_mod_inverse(Ri, R, &mont->N, ctx)) == NULL)
344 if (!BN_lshift(Ri, Ri, mont->ri))
346 if (!BN_sub_word(Ri, 1))
351 if (!BN_div(&(mont->Ni), NULL, Ri, &mont->N, ctx))
356 /* setup RR for conversions */
357 BN_zero(&(mont->RR));
358 if (!BN_set_bit(&(mont->RR), mont->ri * 2))
360 if (!BN_mod(&(mont->RR), &(mont->RR), &(mont->N), ctx))
369 BN_MONT_CTX *BN_MONT_CTX_copy(BN_MONT_CTX *to, BN_MONT_CTX *from)
374 if (!BN_copy(&(to->RR), &(from->RR)))
376 if (!BN_copy(&(to->N), &(from->N)))
378 if (!BN_copy(&(to->Ni), &(from->Ni)))
381 to->n0[0] = from->n0[0];
382 to->n0[1] = from->n0[1];
386 BN_MONT_CTX *BN_MONT_CTX_set_locked(BN_MONT_CTX **pmont, CRYPTO_RWLOCK *lock,
387 const BIGNUM *mod, BN_CTX *ctx)
391 CRYPTO_THREAD_read_lock(lock);
393 CRYPTO_THREAD_unlock(lock);
398 * We don't want to serialise globally while doing our lazy-init math in
399 * BN_MONT_CTX_set. That punishes threads that are doing independent
400 * things. Instead, punish the case where more than one thread tries to
401 * lazy-init the same 'pmont', by having each do the lazy-init math work
402 * independently and only use the one from the thread that wins the race
403 * (the losers throw away the work they've done).
405 ret = BN_MONT_CTX_new();
408 if (!BN_MONT_CTX_set(ret, mod, ctx)) {
409 BN_MONT_CTX_free(ret);
413 /* The locked compare-and-set, after the local work is done. */
414 CRYPTO_THREAD_write_lock(lock);
416 BN_MONT_CTX_free(ret);
420 CRYPTO_THREAD_unlock(lock);