2 * Copyright 2017-2020 The OpenSSL Project Authors. All Rights Reserved.
3 * Copyright 2015-2016 Cryptography Research, Inc.
5 * Licensed under the Apache License 2.0 (the "License"). You may not use
6 * this file except in compliance with the License. You can obtain a copy
7 * in the file LICENSE in the source distribution or at
8 * https://www.openssl.org/source/license.html
10 * Originally written by Mike Hamburg
12 #include <openssl/crypto.h>
16 #include "point_448.h"
18 #include "crypto/ecx.h"
19 #include "curve448_local.h"
23 #define C448_WNAF_FIXED_TABLE_BITS 5
24 #define C448_WNAF_VAR_TABLE_BITS 3
26 #define EDWARDS_D (-39081)
28 static const curve448_scalar_t precomputed_scalarmul_adjustment = {
31 SC_LIMB(0xc873d6d54a7bb0cfULL), SC_LIMB(0xe933d8d723a70aadULL),
32 SC_LIMB(0xbb124b65129c96fdULL), SC_LIMB(0x00000008335dc163ULL)
37 #define TWISTED_D (EDWARDS_D - 1)
39 #define WBITS C448_WORD_BITS /* NB this may be different from ARCH_WORD_BITS */
42 static void gf_invert(gf y, const gf x, int assert_nonzero)
47 gf_sqr(t1, x); /* o^2 */
48 ret = gf_isr(t2, t1); /* +-1/sqrt(o^2) = +-1/o */
53 gf_mul(t2, t1, x); /* not direct to y in case of alias. */
57 /** identity = (0,1) */
58 const curve448_point_t curve448_point_identity =
59 { {{{{0}}}, {{{1}}}, {{{1}}}, {{{0}}}} };
61 static void point_double_internal(curve448_point_t p, const curve448_point_t q,
68 gf_add_nr(d, c, a); /* 2+e */
69 gf_add_nr(p->t, q->y, q->x); /* 2+e */
71 gf_subx_nr(b, b, d, 3); /* 4+e */
72 gf_sub_nr(p->t, a, c); /* 3+e */
74 gf_add_nr(p->z, p->x, p->x); /* 2+e */
75 gf_subx_nr(a, p->z, p->t, 4); /* 6+e */
77 gf_weak_reduce(a); /* or 1+e */
79 gf_mul(p->z, p->t, a);
80 gf_mul(p->y, p->t, d);
85 void curve448_point_double(curve448_point_t p, const curve448_point_t q)
87 point_double_internal(p, q, 0);
90 /* Operations on [p]niels */
91 static ossl_inline void cond_neg_niels(niels_t n, mask_t neg)
93 gf_cond_swap(n->a, n->b, neg);
94 gf_cond_neg(n->c, neg);
97 static void pt_to_pniels(pniels_t b, const curve448_point_t a)
99 gf_sub(b->n->a, a->y, a->x);
100 gf_add(b->n->b, a->x, a->y);
101 gf_mulw(b->n->c, a->t, 2 * TWISTED_D);
102 gf_add(b->z, a->z, a->z);
105 static void pniels_to_pt(curve448_point_t e, const pniels_t d)
109 gf_add(eu, d->n->b, d->n->a);
110 gf_sub(e->y, d->n->b, d->n->a);
111 gf_mul(e->t, e->y, eu);
112 gf_mul(e->x, d->z, e->y);
113 gf_mul(e->y, d->z, eu);
117 static void niels_to_pt(curve448_point_t e, const niels_t n)
119 gf_add(e->y, n->b, n->a);
120 gf_sub(e->x, n->b, n->a);
121 gf_mul(e->t, e->y, e->x);
125 static void add_niels_to_pt(curve448_point_t d, const niels_t e,
130 gf_sub_nr(b, d->y, d->x); /* 3+e */
132 gf_add_nr(b, d->x, d->y); /* 2+e */
133 gf_mul(d->y, e->b, b);
134 gf_mul(d->x, e->c, d->t);
135 gf_add_nr(c, a, d->y); /* 2+e */
136 gf_sub_nr(b, d->y, a); /* 3+e */
137 gf_sub_nr(d->y, d->z, d->x); /* 3+e */
138 gf_add_nr(a, d->x, d->z); /* 2+e */
139 gf_mul(d->z, a, d->y);
140 gf_mul(d->x, d->y, b);
146 static void sub_niels_from_pt(curve448_point_t d, const niels_t e,
151 gf_sub_nr(b, d->y, d->x); /* 3+e */
153 gf_add_nr(b, d->x, d->y); /* 2+e */
154 gf_mul(d->y, e->a, b);
155 gf_mul(d->x, e->c, d->t);
156 gf_add_nr(c, a, d->y); /* 2+e */
157 gf_sub_nr(b, d->y, a); /* 3+e */
158 gf_add_nr(d->y, d->z, d->x); /* 2+e */
159 gf_sub_nr(a, d->z, d->x); /* 3+e */
160 gf_mul(d->z, a, d->y);
161 gf_mul(d->x, d->y, b);
167 static void add_pniels_to_pt(curve448_point_t p, const pniels_t pn,
172 gf_mul(L0, p->z, pn->z);
174 add_niels_to_pt(p, pn->n, before_double);
177 static void sub_pniels_from_pt(curve448_point_t p, const pniels_t pn,
182 gf_mul(L0, p->z, pn->z);
184 sub_niels_from_pt(p, pn->n, before_double);
187 c448_bool_t curve448_point_eq(const curve448_point_t p,
188 const curve448_point_t q)
193 /* equality mod 2-torsion compares x/y */
194 gf_mul(a, p->y, q->x);
195 gf_mul(b, q->y, p->x);
198 return mask_to_bool(succ);
201 c448_bool_t curve448_point_valid(const curve448_point_t p)
206 gf_mul(a, p->x, p->y);
207 gf_mul(b, p->z, p->t);
213 gf_mulw(c, b, TWISTED_D);
217 out &= ~gf_eq(p->z, ZERO);
218 return mask_to_bool(out);
221 static ossl_inline void constant_time_lookup_niels(niels_s * RESTRICT ni,
222 const niels_t * table,
225 constant_time_lookup(ni, table, sizeof(niels_s), nelts, idx);
228 void curve448_precomputed_scalarmul(curve448_point_t out,
229 const curve448_precomputed_s * table,
230 const curve448_scalar_t scalar)
232 unsigned int i, j, k;
233 const unsigned int n = COMBS_N, t = COMBS_T, s = COMBS_S;
235 curve448_scalar_t scalar1x;
237 curve448_scalar_add(scalar1x, scalar, precomputed_scalarmul_adjustment);
238 curve448_scalar_halve(scalar1x, scalar1x);
240 for (i = s; i > 0; i--) {
242 point_double_internal(out, out, 0);
244 for (j = 0; j < n; j++) {
248 for (k = 0; k < t; k++) {
249 unsigned int bit = (i - 1) + s * (k + j * t);
251 if (bit < C448_SCALAR_BITS)
253 (scalar1x->limb[bit / WBITS] >> (bit % WBITS) & 1) << k;
256 invert = (tab >> (t - 1)) - 1;
258 tab &= (1 << (t - 1)) - 1;
260 constant_time_lookup_niels(ni, &table->table[j << (t - 1)],
263 cond_neg_niels(ni, invert);
264 if ((i != s) || j != 0)
265 add_niels_to_pt(out, ni, j == n - 1 && i != 1);
267 niels_to_pt(out, ni);
271 OPENSSL_cleanse(ni, sizeof(ni));
272 OPENSSL_cleanse(scalar1x, sizeof(scalar1x));
275 void curve448_point_mul_by_ratio_and_encode_like_eddsa(
276 uint8_t enc[EDDSA_448_PUBLIC_BYTES],
277 const curve448_point_t p)
282 /* The point is now on the twisted curve. Move it to untwisted. */
283 curve448_point_copy(q, p);
286 /* 4-isogeny: 2xy/(y^+x^2), (y^2-x^2)/(2z^2-y^2+x^2) */
292 gf_add(z, q->y, q->x);
302 OPENSSL_cleanse(u, sizeof(u));
311 enc[EDDSA_448_PRIVATE_BYTES - 1] = 0;
312 gf_serialize(enc, x, 1);
313 enc[EDDSA_448_PRIVATE_BYTES - 1] |= 0x80 & gf_lobit(t);
315 OPENSSL_cleanse(x, sizeof(x));
316 OPENSSL_cleanse(y, sizeof(y));
317 OPENSSL_cleanse(z, sizeof(z));
318 OPENSSL_cleanse(t, sizeof(t));
319 curve448_point_destroy(q);
322 c448_error_t curve448_point_decode_like_eddsa_and_mul_by_ratio(
324 const uint8_t enc[EDDSA_448_PUBLIC_BYTES])
326 uint8_t enc2[EDDSA_448_PUBLIC_BYTES];
330 memcpy(enc2, enc, sizeof(enc2));
332 low = ~word_is_zero(enc2[EDDSA_448_PRIVATE_BYTES - 1] & 0x80);
333 enc2[EDDSA_448_PRIVATE_BYTES - 1] &= ~0x80;
335 succ = gf_deserialize(p->y, enc2, 1, 0);
336 succ &= word_is_zero(enc2[EDDSA_448_PRIVATE_BYTES - 1]);
339 gf_sub(p->z, ONE, p->x); /* num = 1-y^2 */
340 gf_mulw(p->t, p->x, EDWARDS_D); /* dy^2 */
341 gf_sub(p->t, ONE, p->t); /* denom = 1-dy^2 or 1-d + dy^2 */
343 gf_mul(p->x, p->z, p->t);
344 succ &= gf_isr(p->t, p->x); /* 1/sqrt(num * denom) */
346 gf_mul(p->x, p->t, p->z); /* sqrt(num / denom) */
347 gf_cond_neg(p->x, gf_lobit(p->x) ^ low);
353 /* 4-isogeny 2xy/(y^2-ax^2), (y^2+ax^2)/(2-y^2-ax^2) */
357 gf_add(p->t, p->y, p->x);
362 gf_add(p->z, p->x, p->x);
365 gf_mul(p->z, p->t, a);
366 gf_mul(p->y, p->t, d);
368 OPENSSL_cleanse(a, sizeof(a));
369 OPENSSL_cleanse(b, sizeof(b));
370 OPENSSL_cleanse(c, sizeof(c));
371 OPENSSL_cleanse(d, sizeof(d));
374 OPENSSL_cleanse(enc2, sizeof(enc2));
375 assert(curve448_point_valid(p) || ~succ);
377 return c448_succeed_if(mask_to_bool(succ));
380 c448_error_t x448_int(uint8_t out[X_PUBLIC_BYTES],
381 const uint8_t base[X_PUBLIC_BYTES],
382 const uint8_t scalar[X_PRIVATE_BYTES])
384 gf x1, x2, z2, x3, z3, t1, t2;
389 (void)gf_deserialize(x1, base, 1, 0);
395 for (t = X_PRIVATE_BITS - 1; t >= 0; t--) {
396 uint8_t sb = scalar[t / 8];
399 /* Scalar conditioning */
401 sb &= -(uint8_t)COFACTOR;
402 else if (t == X_PRIVATE_BITS - 1)
405 k_t = (sb >> (t % 8)) & 1;
406 k_t = 0 - k_t; /* set to all 0s or all 1s */
409 gf_cond_swap(x2, x3, swap);
410 gf_cond_swap(z2, z3, swap);
414 * The "_nr" below skips coefficient reduction. In the following
415 * comments, "2+e" is saying that the coefficients are at most 2+epsilon
416 * times the reduction limit.
418 gf_add_nr(t1, x2, z2); /* A = x2 + z2 */ /* 2+e */
419 gf_sub_nr(t2, x2, z2); /* B = x2 - z2 */ /* 3+e */
420 gf_sub_nr(z2, x3, z3); /* D = x3 - z3 */ /* 3+e */
421 gf_mul(x2, t1, z2); /* DA */
422 gf_add_nr(z2, z3, x3); /* C = x3 + z3 */ /* 2+e */
423 gf_mul(x3, t2, z2); /* CB */
424 gf_sub_nr(z3, x2, x3); /* DA-CB */ /* 3+e */
425 gf_sqr(z2, z3); /* (DA-CB)^2 */
426 gf_mul(z3, x1, z2); /* z3 = x1(DA-CB)^2 */
427 gf_add_nr(z2, x2, x3); /* (DA+CB) */ /* 2+e */
428 gf_sqr(x3, z2); /* x3 = (DA+CB)^2 */
430 gf_sqr(z2, t1); /* AA = A^2 */
431 gf_sqr(t1, t2); /* BB = B^2 */
432 gf_mul(x2, z2, t1); /* x2 = AA*BB */
433 gf_sub_nr(t2, z2, t1); /* E = AA-BB */ /* 3+e */
435 gf_mulw(t1, t2, -EDWARDS_D); /* E*-d = a24*E */
436 gf_add_nr(t1, t1, z2); /* AA + a24*E */ /* 2+e */
437 gf_mul(z2, t2, t1); /* z2 = E(AA+a24*E) */
441 gf_cond_swap(x2, x3, swap);
442 gf_cond_swap(z2, z3, swap);
443 gf_invert(z2, z2, 0);
445 gf_serialize(out, x1, 1);
446 nz = ~gf_eq(x1, ZERO);
448 OPENSSL_cleanse(x1, sizeof(x1));
449 OPENSSL_cleanse(x2, sizeof(x2));
450 OPENSSL_cleanse(z2, sizeof(z2));
451 OPENSSL_cleanse(x3, sizeof(x3));
452 OPENSSL_cleanse(z3, sizeof(z3));
453 OPENSSL_cleanse(t1, sizeof(t1));
454 OPENSSL_cleanse(t2, sizeof(t2));
456 return c448_succeed_if(mask_to_bool(nz));
459 void curve448_point_mul_by_ratio_and_encode_like_x448(uint8_t
461 const curve448_point_t p)
465 curve448_point_copy(q, p);
466 gf_invert(q->t, q->x, 0); /* 1/x */
467 gf_mul(q->z, q->t, q->y); /* y/x */
468 gf_sqr(q->y, q->z); /* (y/x)^2 */
469 gf_serialize(out, q->y, 1);
470 curve448_point_destroy(q);
473 void x448_derive_public_key(uint8_t out[X_PUBLIC_BYTES],
474 const uint8_t scalar[X_PRIVATE_BYTES])
476 /* Scalar conditioning */
477 uint8_t scalar2[X_PRIVATE_BYTES];
478 curve448_scalar_t the_scalar;
482 memcpy(scalar2, scalar, sizeof(scalar2));
483 scalar2[0] &= -(uint8_t)COFACTOR;
485 scalar2[X_PRIVATE_BYTES - 1] &= ~((0u - 1u) << ((X_PRIVATE_BITS + 7) % 8));
486 scalar2[X_PRIVATE_BYTES - 1] |= 1 << ((X_PRIVATE_BITS + 7) % 8);
488 curve448_scalar_decode_long(the_scalar, scalar2, sizeof(scalar2));
490 /* Compensate for the encoding ratio */
491 for (i = 1; i < X448_ENCODE_RATIO; i <<= 1)
492 curve448_scalar_halve(the_scalar, the_scalar);
494 curve448_precomputed_scalarmul(p, curve448_precomputed_base, the_scalar);
495 curve448_point_mul_by_ratio_and_encode_like_x448(out, p);
496 curve448_point_destroy(p);
499 /* Control for variable-time scalar multiply algorithms. */
500 struct smvt_control {
504 #if defined(__GNUC__) && (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ > 3))
505 # define NUMTRAILINGZEROS __builtin_ctz
507 # define NUMTRAILINGZEROS numtrailingzeros
508 static uint32_t numtrailingzeros(uint32_t i)
544 static int recode_wnaf(struct smvt_control *control,
545 /* [nbits/(table_bits + 1) + 3] */
546 const curve448_scalar_t scalar,
547 unsigned int table_bits)
549 unsigned int table_size = C448_SCALAR_BITS / (table_bits + 1) + 3;
550 int position = table_size - 1; /* at the end */
551 uint64_t current = scalar->limb[0] & 0xFFFF;
552 uint32_t mask = (1 << (table_bits + 1)) - 1;
554 const unsigned int B_OVER_16 = sizeof(scalar->limb[0]) / 2;
557 /* place the end marker */
558 control[position].power = -1;
559 control[position].addend = 0;
563 * PERF: Could negate scalar if it's large. But then would need more cases
564 * in the actual code that uses it, all for an expected reduction of like
565 * 1/5 op. Probably not worth it.
568 for (w = 1; w < (C448_SCALAR_BITS - 1) / 16 + 3; w++) {
569 if (w < (C448_SCALAR_BITS - 1) / 16 + 1) {
570 /* Refill the 16 high bits of current */
571 current += (uint32_t)((scalar->limb[w / B_OVER_16]
572 >> (16 * (w % B_OVER_16))) << 16);
575 while (current & 0xFFFF) {
576 uint32_t pos = NUMTRAILINGZEROS((uint32_t)current);
577 uint32_t odd = (uint32_t)current >> pos;
578 int32_t delta = odd & mask;
580 assert(position >= 0);
581 if (odd & (1 << (table_bits + 1)))
582 delta -= (1 << (table_bits + 1));
583 current -= delta * (1 << pos);
584 control[position].power = pos + 16 * (w - 1);
585 control[position].addend = delta;
590 assert(current == 0);
593 n = table_size - position;
594 for (i = 0; i < n; i++)
595 control[i] = control[i + position];
600 static void prepare_wnaf_table(pniels_t * output,
601 const curve448_point_t working,
604 curve448_point_t tmp;
608 pt_to_pniels(output[0], working);
613 curve448_point_double(tmp, working);
614 pt_to_pniels(twop, tmp);
616 add_pniels_to_pt(tmp, output[0], 0);
617 pt_to_pniels(output[1], tmp);
619 for (i = 2; i < 1 << tbits; i++) {
620 add_pniels_to_pt(tmp, twop, 0);
621 pt_to_pniels(output[i], tmp);
624 curve448_point_destroy(tmp);
625 OPENSSL_cleanse(twop, sizeof(twop));
628 void curve448_base_double_scalarmul_non_secret(curve448_point_t combo,
629 const curve448_scalar_t scalar1,
630 const curve448_point_t base2,
631 const curve448_scalar_t scalar2)
633 const int table_bits_var = C448_WNAF_VAR_TABLE_BITS;
634 const int table_bits_pre = C448_WNAF_FIXED_TABLE_BITS;
635 struct smvt_control control_var[C448_SCALAR_BITS /
636 (C448_WNAF_VAR_TABLE_BITS + 1) + 3];
637 struct smvt_control control_pre[C448_SCALAR_BITS /
638 (C448_WNAF_FIXED_TABLE_BITS + 1) + 3];
639 int ncb_pre = recode_wnaf(control_pre, scalar1, table_bits_pre);
640 int ncb_var = recode_wnaf(control_var, scalar2, table_bits_var);
641 pniels_t precmp_var[1 << C448_WNAF_VAR_TABLE_BITS];
642 int contp = 0, contv = 0, i;
644 prepare_wnaf_table(precmp_var, base2, table_bits_var);
645 i = control_var[0].power;
648 curve448_point_copy(combo, curve448_point_identity);
651 if (i > control_pre[0].power) {
652 pniels_to_pt(combo, precmp_var[control_var[0].addend >> 1]);
654 } else if (i == control_pre[0].power && i >= 0) {
655 pniels_to_pt(combo, precmp_var[control_var[0].addend >> 1]);
656 add_niels_to_pt(combo, curve448_wnaf_base[control_pre[0].addend >> 1],
661 i = control_pre[0].power;
662 niels_to_pt(combo, curve448_wnaf_base[control_pre[0].addend >> 1]);
666 for (i--; i >= 0; i--) {
667 int cv = (i == control_var[contv].power);
668 int cp = (i == control_pre[contp].power);
670 point_double_internal(combo, combo, i && !(cv || cp));
673 assert(control_var[contv].addend);
675 if (control_var[contv].addend > 0)
676 add_pniels_to_pt(combo,
677 precmp_var[control_var[contv].addend >> 1],
680 sub_pniels_from_pt(combo,
681 precmp_var[(-control_var[contv].addend)
687 assert(control_pre[contp].addend);
689 if (control_pre[contp].addend > 0)
690 add_niels_to_pt(combo,
691 curve448_wnaf_base[control_pre[contp].addend
694 sub_niels_from_pt(combo,
695 curve448_wnaf_base[(-control_pre
696 [contp].addend) >> 1], i);
701 /* This function is non-secret, but whatever this is cheap. */
702 OPENSSL_cleanse(control_var, sizeof(control_var));
703 OPENSSL_cleanse(control_pre, sizeof(control_pre));
704 OPENSSL_cleanse(precmp_var, sizeof(precmp_var));
706 assert(contv == ncb_var);
708 assert(contp == ncb_pre);
712 void curve448_point_destroy(curve448_point_t point)
714 OPENSSL_cleanse(point, sizeof(curve448_point_t));
717 int X448(uint8_t out_shared_key[56], const uint8_t private_key[56],
718 const uint8_t peer_public_value[56])
720 return x448_int(out_shared_key, peer_public_value, private_key)
724 void X448_public_from_private(uint8_t out_public_value[56],
725 const uint8_t private_key[56])
727 x448_derive_public_key(out_public_value, private_key);