2 * Copyright 2017 The OpenSSL Project Authors. All Rights Reserved.
3 * Copyright 2015-2016 Cryptography Research, Inc.
5 * Licensed under the OpenSSL license (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 "curve448_lcl.h"
22 /* Comb config: number of combs, n, t, s. */
26 #define DECAF_WINDOW_BITS 5
27 #define DECAF_WNAF_FIXED_TABLE_BITS 5
28 #define DECAF_WNAF_VAR_TABLE_BITS 3
30 static const int EDWARDS_D = -39081;
31 static const curve448_scalar_t precomputed_scalarmul_adjustment = {{{
32 SC_LIMB(0xc873d6d54a7bb0cf), SC_LIMB(0xe933d8d723a70aad), SC_LIMB(0xbb124b65129c96fd), SC_LIMB(0x00000008335dc163)
35 const uint8_t decaf_x448_base_point[DECAF_X448_PUBLIC_BYTES] = { 0x05 };
38 #define TWISTED_D ((EDWARDS_D)-1)
40 #define EFF_D (-(TWISTED_D))
43 /* End of template stuff */
45 #define WBITS DECAF_WORD_BITS /* NB this may be different from ARCH_WORD_BITS */
47 /* Projective Niels coordinates */
48 typedef struct { gf a, b, c; } niels_s, niels_t[1];
49 typedef struct { niels_t n; gf z; } VECTOR_ALIGNED pniels_s, pniels_t[1];
51 /* Precomputed base */
52 struct curve448_precomputed_s { niels_t table [COMBS_N<<(COMBS_T-1)]; };
54 extern const gf curve448_precomputed_base_as_fe[];
55 const curve448_precomputed_s *curve448_precomputed_base =
56 (const curve448_precomputed_s *) &curve448_precomputed_base_as_fe;
60 gf_invert(gf y, const gf x, int assert_nonzero) {
64 gf_sqr(t1, x); /* o^2 */
65 ret = gf_isr(t2, t1); /* +-1/sqrt(o^2) = +-1/o */
67 if (assert_nonzero) assert(ret);
69 gf_mul(t2, t1, x); /* not direct to y in case of alias. */
73 /** identity = (0,1) */
74 const curve448_point_t curve448_point_identity = {{{{{0}}},{{{1}}},{{{1}}},{{{0}}}}};
77 point_double_internal (
79 const curve448_point_t q,
85 gf_add_nr ( d, c, a ); /* 2+e */
86 gf_add_nr ( p->t, q->y, q->x ); /* 2+e */
88 gf_subx_nr ( b, b, d, 3 ); /* 4+e */
89 gf_sub_nr ( p->t, a, c ); /* 3+e */
90 gf_sqr ( p->x, q->z );
91 gf_add_nr ( p->z, p->x, p->x ); /* 2+e */
92 gf_subx_nr ( a, p->z, p->t, 4 ); /* 6+e */
93 if (GF_HEADROOM == 5) gf_weak_reduce(a); /* or 1+e */
94 gf_mul ( p->x, a, b );
95 gf_mul ( p->z, p->t, a );
96 gf_mul ( p->y, p->t, d );
97 if (!before_double) gf_mul ( p->t, b, d );
100 void curve448_point_double(curve448_point_t p, const curve448_point_t q) {
101 point_double_internal(p,q,0);
104 /* Operations on [p]niels */
105 static ossl_inline void
110 gf_cond_swap(n->a, n->b, neg);
111 gf_cond_neg(n->c, neg);
114 static void pt_to_pniels (
116 const curve448_point_t a
118 gf_sub ( b->n->a, a->y, a->x );
119 gf_add ( b->n->b, a->x, a->y );
120 gf_mulw ( b->n->c, a->t, 2*TWISTED_D );
121 gf_add ( b->z, a->z, a->z );
124 static void pniels_to_pt (
129 gf_add ( eu, d->n->b, d->n->a );
130 gf_sub ( e->y, d->n->b, d->n->a );
131 gf_mul ( e->t, e->y, eu);
132 gf_mul ( e->x, d->z, e->y );
133 gf_mul ( e->y, d->z, eu );
134 gf_sqr ( e->z, d->z );
142 gf_add ( e->y, n->b, n->a );
143 gf_sub ( e->x, n->b, n->a );
144 gf_mul ( e->t, e->y, e->x );
145 gf_copy ( e->z, ONE );
155 gf_sub_nr ( b, d->y, d->x ); /* 3+e */
156 gf_mul ( a, e->a, b );
157 gf_add_nr ( b, d->x, d->y ); /* 2+e */
158 gf_mul ( d->y, e->b, b );
159 gf_mul ( d->x, e->c, d->t );
160 gf_add_nr ( c, a, d->y ); /* 2+e */
161 gf_sub_nr ( b, d->y, a ); /* 3+e */
162 gf_sub_nr ( d->y, d->z, d->x ); /* 3+e */
163 gf_add_nr ( a, d->x, d->z ); /* 2+e */
164 gf_mul ( d->z, a, d->y );
165 gf_mul ( d->x, d->y, b );
166 gf_mul ( d->y, a, c );
167 if (!before_double) gf_mul ( d->t, b, c );
177 gf_sub_nr ( b, d->y, d->x ); /* 3+e */
178 gf_mul ( a, e->b, b );
179 gf_add_nr ( b, d->x, d->y ); /* 2+e */
180 gf_mul ( d->y, e->a, b );
181 gf_mul ( d->x, e->c, d->t );
182 gf_add_nr ( c, a, d->y ); /* 2+e */
183 gf_sub_nr ( b, d->y, a ); /* 3+e */
184 gf_add_nr ( d->y, d->z, d->x ); /* 2+e */
185 gf_sub_nr ( a, d->z, d->x ); /* 3+e */
186 gf_mul ( d->z, a, d->y );
187 gf_mul ( d->x, d->y, b );
188 gf_mul ( d->y, a, c );
189 if (!before_double) gf_mul ( d->t, b, c );
199 gf_mul ( L0, p->z, pn->z );
200 gf_copy ( p->z, L0 );
201 add_niels_to_pt( p, pn->n, before_double );
211 gf_mul ( L0, p->z, pn->z );
212 gf_copy ( p->z, L0 );
213 sub_niels_from_pt( p, pn->n, before_double );
216 decaf_bool_t curve448_point_eq ( const curve448_point_t p, const curve448_point_t q ) {
219 /* equality mod 2-torsion compares x/y */
221 gf_mul ( a, p->y, q->x );
222 gf_mul ( b, q->y, p->x );
225 return mask_to_bool(succ);
228 decaf_bool_t curve448_point_valid (
229 const curve448_point_t p
241 gf_mulw(c,b,TWISTED_D);
245 out &= ~gf_eq(p->z,ZERO);
246 return mask_to_bool(out);
249 static ossl_inline void
250 constant_time_lookup_niels (
251 niels_s *__restrict__ ni,
252 const niels_t *table,
256 constant_time_lookup(ni, table, sizeof(niels_s), nelts, idx);
259 void curve448_precomputed_scalarmul (
260 curve448_point_t out,
261 const curve448_precomputed_s *table,
262 const curve448_scalar_t scalar
266 const unsigned int n = COMBS_N, t = COMBS_T, s = COMBS_S;
269 curve448_scalar_t scalar1x;
270 curve448_scalar_add(scalar1x, scalar, precomputed_scalarmul_adjustment);
271 curve448_scalar_halve(scalar1x,scalar1x);
273 for (i=s-1; i>=0; i--) {
274 if (i != (int)s-1) point_double_internal(out,out,0);
276 for (j=0; j<n; j++) {
280 for (k=0; k<t; k++) {
281 unsigned int bit = i + s*(k + j*t);
282 if (bit < DECAF_448_SCALAR_BITS) {
283 tab |= (scalar1x->limb[bit/WBITS] >> (bit%WBITS) & 1) << k;
287 invert = (tab>>(t-1))-1;
289 tab &= (1<<(t-1)) - 1;
291 constant_time_lookup_niels(ni, &table->table[j<<(t-1)], 1<<(t-1), tab);
293 cond_neg_niels(ni, invert);
294 if ((i!=(int)s-1)||j) {
295 add_niels_to_pt(out, ni, j==n-1 && i);
297 niels_to_pt(out, ni);
302 OPENSSL_cleanse(ni,sizeof(ni));
303 OPENSSL_cleanse(scalar1x,sizeof(scalar1x));
306 void curve448_point_mul_by_ratio_and_encode_like_eddsa (
307 uint8_t enc[DECAF_EDDSA_448_PUBLIC_BYTES],
308 const curve448_point_t p
311 /* The point is now on the twisted curve. Move it to untwisted. */
314 curve448_point_copy(q,p);
317 /* 4-isogeny: 2xy/(y^+x^2), (y^2-x^2)/(2z^2-y^2+x^2) */
322 gf_add( z, q->y, q->x );
332 OPENSSL_cleanse(u,sizeof(u));
341 enc[DECAF_EDDSA_448_PRIVATE_BYTES-1] = 0;
342 gf_serialize(enc, x, 1);
343 enc[DECAF_EDDSA_448_PRIVATE_BYTES-1] |= 0x80 & gf_lobit(t);
345 OPENSSL_cleanse(x,sizeof(x));
346 OPENSSL_cleanse(y,sizeof(y));
347 OPENSSL_cleanse(z,sizeof(z));
348 OPENSSL_cleanse(t,sizeof(t));
349 curve448_point_destroy(q);
353 decaf_error_t curve448_point_decode_like_eddsa_and_mul_by_ratio (
355 const uint8_t enc[DECAF_EDDSA_448_PUBLIC_BYTES]
357 uint8_t enc2[DECAF_EDDSA_448_PUBLIC_BYTES];
361 memcpy(enc2,enc,sizeof(enc2));
363 low = ~word_is_zero(enc2[DECAF_EDDSA_448_PRIVATE_BYTES-1] & 0x80);
364 enc2[DECAF_EDDSA_448_PRIVATE_BYTES-1] &= ~0x80;
366 succ = gf_deserialize(p->y, enc2, 1, 0);
368 succ &= word_is_zero(enc2[DECAF_EDDSA_448_PRIVATE_BYTES-1]);
372 gf_sub(p->z,ONE,p->x); /* num = 1-y^2 */
373 gf_mulw(p->t,p->x,EDWARDS_D); /* dy^2 */
374 gf_sub(p->t,ONE,p->t); /* denom = 1-dy^2 or 1-d + dy^2 */
376 gf_mul(p->x,p->z,p->t);
377 succ &= gf_isr(p->t,p->x); /* 1/sqrt(num * denom) */
379 gf_mul(p->x,p->t,p->z); /* sqrt(num / denom) */
380 gf_cond_neg(p->x,gf_lobit(p->x)^low);
384 /* 4-isogeny 2xy/(y^2-ax^2), (y^2+ax^2)/(2-y^2-ax^2) */
389 gf_add ( p->t, p->y, p->x );
392 gf_sub ( p->t, a, c );
393 gf_sqr ( p->x, p->z );
394 gf_add ( p->z, p->x, p->x );
395 gf_sub ( a, p->z, d );
396 gf_mul ( p->x, a, b );
397 gf_mul ( p->z, p->t, a );
398 gf_mul ( p->y, p->t, d );
399 gf_mul ( p->t, b, d );
400 OPENSSL_cleanse(a,sizeof(a));
401 OPENSSL_cleanse(b,sizeof(b));
402 OPENSSL_cleanse(c,sizeof(c));
403 OPENSSL_cleanse(d,sizeof(d));
406 OPENSSL_cleanse(enc2,sizeof(enc2));
407 assert(curve448_point_valid(p) || ~succ);
409 return decaf_succeed_if(mask_to_bool(succ));
412 decaf_error_t decaf_x448 (
413 uint8_t out[X_PUBLIC_BYTES],
414 const uint8_t base[X_PUBLIC_BYTES],
415 const uint8_t scalar[X_PRIVATE_BYTES]
417 gf x1, x2, z2, x3, z3, t1, t2;
422 ignore_result(gf_deserialize(x1,base,1,0));
428 for (t = X_PRIVATE_BITS-1; t>=0; t--) {
429 uint8_t sb = scalar[t/8];
432 /* Scalar conditioning */
433 if (t/8==0) sb &= -(uint8_t)COFACTOR;
434 else if (t == X_PRIVATE_BITS-1) sb = -1;
436 k_t = (sb>>(t%8)) & 1;
437 k_t = -k_t; /* set to all 0s or all 1s */
440 gf_cond_swap(x2,x3,swap);
441 gf_cond_swap(z2,z3,swap);
444 gf_add_nr(t1,x2,z2); /* A = x2 + z2 */ /* 2+e */
445 gf_sub_nr(t2,x2,z2); /* B = x2 - z2 */ /* 3+e */
446 gf_sub_nr(z2,x3,z3); /* D = x3 - z3 */ /* 3+e */
447 gf_mul(x2,t1,z2); /* DA */
448 gf_add_nr(z2,z3,x3); /* C = x3 + z3 */ /* 2+e */
449 gf_mul(x3,t2,z2); /* CB */
450 gf_sub_nr(z3,x2,x3); /* DA-CB */ /* 3+e */
451 gf_sqr(z2,z3); /* (DA-CB)^2 */
452 gf_mul(z3,x1,z2); /* z3 = x1(DA-CB)^2 */
453 gf_add_nr(z2,x2,x3); /* (DA+CB) */ /* 2+e */
454 gf_sqr(x3,z2); /* x3 = (DA+CB)^2 */
456 gf_sqr(z2,t1); /* AA = A^2 */
457 gf_sqr(t1,t2); /* BB = B^2 */
458 gf_mul(x2,z2,t1); /* x2 = AA*BB */
459 gf_sub_nr(t2,z2,t1); /* E = AA-BB */ /* 3+e */
461 gf_mulw(t1,t2,-EDWARDS_D); /* E*-d = a24*E */
462 gf_add_nr(t1,t1,z2); /* AA + a24*E */ /* 2+e */
463 gf_mul(z2,t2,t1); /* z2 = E(AA+a24*E) */
467 gf_cond_swap(x2,x3,swap);
468 gf_cond_swap(z2,z3,swap);
471 gf_serialize(out,x1,1);
472 nz = ~gf_eq(x1,ZERO);
474 OPENSSL_cleanse(x1,sizeof(x1));
475 OPENSSL_cleanse(x2,sizeof(x2));
476 OPENSSL_cleanse(z2,sizeof(z2));
477 OPENSSL_cleanse(x3,sizeof(x3));
478 OPENSSL_cleanse(z3,sizeof(z3));
479 OPENSSL_cleanse(t1,sizeof(t1));
480 OPENSSL_cleanse(t2,sizeof(t2));
482 return decaf_succeed_if(mask_to_bool(nz));
485 /* Thanks Johan Pascal */
486 void decaf_ed448_convert_public_key_to_x448 (
487 uint8_t x[DECAF_X448_PUBLIC_BYTES],
488 const uint8_t ed[DECAF_EDDSA_448_PUBLIC_BYTES]
491 const uint8_t mask = (uint8_t)(0xFE<<(7));
492 ignore_result(gf_deserialize(y, ed, 1, mask));
497 /* u = y^2 * (1-dy^2) / (1-y^2) */
498 gf_sqr(n,y); /* y^2*/
499 gf_sub(d,ONE,n); /* 1-y^2*/
500 gf_invert(d,d,0); /* 1/(1-y^2)*/
501 gf_mul(y,n,d); /* y^2 / (1-y^2) */
502 gf_mulw(d,n,EDWARDS_D); /* dy^2*/
503 gf_sub(d, ONE, d); /* 1-dy^2*/
504 gf_mul(n, y, d); /* y^2 * (1-dy^2) / (1-y^2) */
507 OPENSSL_cleanse(y,sizeof(y));
508 OPENSSL_cleanse(n,sizeof(n));
509 OPENSSL_cleanse(d,sizeof(d));
513 void curve448_point_mul_by_ratio_and_encode_like_x448 (
514 uint8_t out[X_PUBLIC_BYTES],
515 const curve448_point_t p
518 curve448_point_copy(q,p);
519 gf_invert(q->t,q->x,0); /* 1/x */
520 gf_mul(q->z,q->t,q->y); /* y/x */
521 gf_sqr(q->y,q->z); /* (y/x)^2 */
522 gf_serialize(out,q->y,1);
523 curve448_point_destroy(q);
526 void decaf_x448_derive_public_key (
527 uint8_t out[X_PUBLIC_BYTES],
528 const uint8_t scalar[X_PRIVATE_BYTES]
530 /* Scalar conditioning */
531 uint8_t scalar2[X_PRIVATE_BYTES];
532 curve448_scalar_t the_scalar;
536 memcpy(scalar2,scalar,sizeof(scalar2));
537 scalar2[0] &= -(uint8_t)COFACTOR;
539 scalar2[X_PRIVATE_BYTES-1] &= ~(-1u<<((X_PRIVATE_BITS+7)%8));
540 scalar2[X_PRIVATE_BYTES-1] |= 1<<((X_PRIVATE_BITS+7)%8);
542 curve448_scalar_decode_long(the_scalar,scalar2,sizeof(scalar2));
544 /* Compensate for the encoding ratio */
545 for (i=1; i<DECAF_X448_ENCODE_RATIO; i<<=1) {
546 curve448_scalar_halve(the_scalar,the_scalar);
548 curve448_precomputed_scalarmul(p,curve448_precomputed_base,the_scalar);
549 curve448_point_mul_by_ratio_and_encode_like_x448(out,p);
550 curve448_point_destroy(p);
555 * Control for variable-time scalar multiply algorithms.
557 struct smvt_control {
561 static int recode_wnaf (
562 struct smvt_control *control, /* [nbits/(table_bits+1) + 3] */
563 const curve448_scalar_t scalar,
564 unsigned int table_bits
566 unsigned int table_size = DECAF_448_SCALAR_BITS/(table_bits+1) + 3;
567 int position = table_size - 1; /* at the end */
568 uint64_t current = scalar->limb[0] & 0xFFFF;
569 uint32_t mask = (1<<(table_bits+1))-1;
571 const unsigned int B_OVER_16 = sizeof(scalar->limb[0]) / 2;
574 /* place the end marker */
575 control[position].power = -1;
576 control[position].addend = 0;
579 /* PERF: Could negate scalar if it's large. But then would need more cases
580 * in the actual code that uses it, all for an expected reduction of like 1/5 op.
581 * Probably not worth it.
584 for (w = 1; w<(DECAF_448_SCALAR_BITS-1)/16+3; w++) {
585 if (w < (DECAF_448_SCALAR_BITS-1)/16+1) {
586 /* Refill the 16 high bits of current */
587 current += (uint32_t)((scalar->limb[w/B_OVER_16]>>(16*(w%B_OVER_16)))<<16);
590 while (current & 0xFFFF) {
591 uint32_t pos = __builtin_ctz((uint32_t)current), odd = (uint32_t)current >> pos;
592 int32_t delta = odd & mask;
594 assert(position >= 0);
595 if (odd & 1<<(table_bits+1)) delta -= (1<<(table_bits+1));
596 current -= delta << pos;
597 control[position].power = pos + 16*(w-1);
598 control[position].addend = delta;
606 n = table_size - position;
607 for (i=0; i<n; i++) {
608 control[i] = control[i+position];
616 const curve448_point_t working,
619 curve448_point_t tmp;
623 pt_to_pniels(output[0], working);
625 if (tbits == 0) return;
627 curve448_point_double(tmp,working);
628 pt_to_pniels(twop, tmp);
630 add_pniels_to_pt(tmp, output[0],0);
631 pt_to_pniels(output[1], tmp);
633 for (i=2; i < 1<<tbits; i++) {
634 add_pniels_to_pt(tmp, twop,0);
635 pt_to_pniels(output[i], tmp);
638 curve448_point_destroy(tmp);
639 OPENSSL_cleanse(twop,sizeof(twop));
642 extern const gf curve448_precomputed_wnaf_as_fe[];
643 static const niels_t *curve448_wnaf_base = (const niels_t *)curve448_precomputed_wnaf_as_fe;
645 void curve448_base_double_scalarmul_non_secret (
646 curve448_point_t combo,
647 const curve448_scalar_t scalar1,
648 const curve448_point_t base2,
649 const curve448_scalar_t scalar2
651 const int table_bits_var = DECAF_WNAF_VAR_TABLE_BITS,
652 table_bits_pre = DECAF_WNAF_FIXED_TABLE_BITS;
653 struct smvt_control control_var[DECAF_448_SCALAR_BITS/(DECAF_WNAF_VAR_TABLE_BITS+1)+3];
654 struct smvt_control control_pre[DECAF_448_SCALAR_BITS/(DECAF_WNAF_FIXED_TABLE_BITS+1)+3];
655 int ncb_pre = recode_wnaf(control_pre, scalar1, table_bits_pre);
656 int ncb_var = recode_wnaf(control_var, scalar2, table_bits_var);
657 pniels_t precmp_var[1<<DECAF_WNAF_VAR_TABLE_BITS];
658 int contp=0, contv=0, i;
660 prepare_wnaf_table(precmp_var, base2, table_bits_var);
661 i = control_var[0].power;
664 curve448_point_copy(combo, curve448_point_identity);
666 } else if (i > control_pre[0].power) {
667 pniels_to_pt(combo, precmp_var[control_var[0].addend >> 1]);
669 } else if (i == control_pre[0].power && i >=0 ) {
670 pniels_to_pt(combo, precmp_var[control_var[0].addend >> 1]);
671 add_niels_to_pt(combo, curve448_wnaf_base[control_pre[0].addend >> 1], i);
674 i = control_pre[0].power;
675 niels_to_pt(combo, curve448_wnaf_base[control_pre[0].addend >> 1]);
679 for (i--; i >= 0; i--) {
680 int cv = (i==control_var[contv].power), cp = (i==control_pre[contp].power);
681 point_double_internal(combo,combo,i && !(cv||cp));
684 assert(control_var[contv].addend);
686 if (control_var[contv].addend > 0) {
687 add_pniels_to_pt(combo, precmp_var[control_var[contv].addend >> 1], i&&!cp);
689 sub_pniels_from_pt(combo, precmp_var[(-control_var[contv].addend) >> 1], i&&!cp);
695 assert(control_pre[contp].addend);
697 if (control_pre[contp].addend > 0) {
698 add_niels_to_pt(combo, curve448_wnaf_base[control_pre[contp].addend >> 1], i);
700 sub_niels_from_pt(combo, curve448_wnaf_base[(-control_pre[contp].addend) >> 1], i);
706 /* This function is non-secret, but whatever this is cheap. */
707 OPENSSL_cleanse(control_var,sizeof(control_var));
708 OPENSSL_cleanse(control_pre,sizeof(control_pre));
709 OPENSSL_cleanse(precmp_var,sizeof(precmp_var));
711 assert(contv == ncb_var); (void)ncb_var;
712 assert(contp == ncb_pre); (void)ncb_pre;
715 void curve448_point_destroy (
716 curve448_point_t point
718 OPENSSL_cleanse(point, sizeof(curve448_point_t));
721 int X448(uint8_t out_shared_key[56], const uint8_t private_key[56],
722 const uint8_t peer_public_value[56])
724 return decaf_x448(out_shared_key, peer_public_value, private_key)
728 void X448_public_from_private(uint8_t out_public_value[56],
729 const uint8_t private_key[56])
731 decaf_x448_derive_public_key(out_public_value, private_key);