2 * Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
4 * Permission is hereby granted, free of charge, to any person obtaining
5 * a copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sublicense, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
12 * The above copyright notice and this permission notice shall be
13 * included in all copies or substantial portions of the Software.
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
28 * Parameters for supported curves (field modulus, and 'b' equation
29 * parameter; both values use the 'i31' format, and 'b' is in Montgomery
33 static const uint32_t P256_P[] = {
35 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x00000007,
36 0x00000000, 0x00000000, 0x00000040, 0x7FFFFF80,
40 static const uint32_t P256_R2[] = {
42 0x00014000, 0x00018000, 0x00000000, 0x7FF40000,
43 0x7FEFFFFF, 0x7FF7FFFF, 0x7FAFFFFF, 0x005FFFFF,
47 static const uint32_t P256_B[] = {
49 0x6FEE1803, 0x6229C4BD, 0x21B139BE, 0x327150AA,
50 0x3567802E, 0x3F7212ED, 0x012E4355, 0x782DD38D,
54 static const uint32_t P384_P[] = {
56 0x7FFFFFFF, 0x00000001, 0x00000000, 0x7FFFFFF8,
57 0x7FFFFFEF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF,
58 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF,
62 static const uint32_t P384_R2[] = {
64 0x00000000, 0x00000080, 0x7FFFFE00, 0x000001FF,
65 0x00000800, 0x00000000, 0x7FFFE000, 0x00001FFF,
66 0x00008000, 0x00008000, 0x00000000, 0x00000000,
70 static const uint32_t P384_B[] = {
72 0x6E666840, 0x070D0392, 0x5D810231, 0x7651D50C,
73 0x17E218D6, 0x1B192002, 0x44EFE441, 0x3A524E2B,
74 0x2719BA5F, 0x41F02209, 0x36C5643E, 0x5813EFFE,
78 static const uint32_t P521_P[] = {
80 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF,
81 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF,
82 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF,
83 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF,
87 static const uint32_t P521_R2[] = {
89 0x00001000, 0x00000000, 0x00000000, 0x00000000,
90 0x00000000, 0x00000000, 0x00000000, 0x00000000,
91 0x00000000, 0x00000000, 0x00000000, 0x00000000,
92 0x00000000, 0x00000000, 0x00000000, 0x00000000,
96 static const uint32_t P521_B[] = {
98 0x540FC00A, 0x228FEA35, 0x2C34F1EF, 0x67BF107A,
99 0x46FC1CD5, 0x1605E9DD, 0x6937B165, 0x272A3D8F,
100 0x42785586, 0x44C8C778, 0x15F3B8B4, 0x64B73366,
101 0x03BA8B69, 0x0D05B42A, 0x21F929A2, 0x2C31C393,
112 static inline const curve_params *
113 id_to_curve(int curve)
115 static const curve_params pp[] = {
116 { P256_P, P256_B, P256_R2, 0x00000001 },
117 { P384_P, P384_B, P384_R2, 0x00000001 },
118 { P521_P, P521_B, P521_R2, 0x00000001 }
121 return &pp[curve - BR_EC_secp256r1];
124 #define I31_LEN ((BR_MAX_EC_SIZE + 61) / 31)
127 * Type for a point in Jacobian coordinates:
128 * -- three values, x, y and z, in Montgomery representation
129 * -- affine coordinates are X = x / z^2 and Y = y / z^3
130 * -- for the point at infinity, z = 0
133 uint32_t c[3][I31_LEN];
137 * We use a custom interpreter that uses a dozen registers, and
138 * only six operations:
139 * MSET(d, a) copy a into d
140 * MADD(d, a) d = d+a (modular)
141 * MSUB(d, a) d = d-a (modular)
142 * MMUL(d, a, b) d = a*b (Montgomery multiplication)
143 * MINV(d, a, b) invert d modulo p; a and b are used as scratch registers
144 * MTZ(d) clear return value if d = 0
145 * Destination of MMUL (d) must be distinct from operands (a and b).
146 * There is no such constraint for MSUB and MADD.
148 * Registers include the operand coordinates, and temporaries.
150 #define MSET(d, a) (0x0000 + ((d) << 8) + ((a) << 4))
151 #define MADD(d, a) (0x1000 + ((d) << 8) + ((a) << 4))
152 #define MSUB(d, a) (0x2000 + ((d) << 8) + ((a) << 4))
153 #define MMUL(d, a, b) (0x3000 + ((d) << 8) + ((a) << 4) + (b))
154 #define MINV(d, a, b) (0x4000 + ((d) << 8) + ((a) << 4) + (b))
155 #define MTZ(d) (0x5000 + ((d) << 8))
159 * Registers for the input operands.
169 * Alternate names for the first input operand.
187 * Extra scratch registers available when there is no second operand (e.g.
188 * for "double" and "affine").
195 * Doubling formulas are:
198 * m = 3*(x + z^2)*(x - z^2)
200 * y' = m*(s - x') - 8*y^4
203 * If y = 0 (P has order 2) then this yields infinity (z' = 0), as it
204 * should. This case should not happen anyway, because our curves have
205 * prime order, and thus do not contain any point of order 2.
207 * If P is infinity (z = 0), then again the formulas yield infinity,
208 * which is correct. Thus, this code works for all points.
210 * Cost: 8 multiplications
212 static const uint16_t code_double[] = {
214 * Compute z^2 (in t1).
219 * Compute x-z^2 (in t2) and then x+z^2 (in t1).
226 * Compute m = 3*(x+z^2)*(x-z^2) (in t1).
234 * Compute s = 4*x*y^2 (in t2) and 2*y^2 (in t3).
242 * Compute x' = m^2 - 2*s.
249 * Compute z' = 2*y*z.
256 * Compute y' = m*(s - x') - 8*y^4. Note that we already have
269 * Addtions formulas are:
277 * x3 = r^2 - h^3 - 2 * u1 * h^2
278 * y3 = r * (u1 * h^2 - x3) - s1 * h^3
281 * If both P1 and P2 are infinity, then z1 == 0 and z2 == 0, implying that
282 * z3 == 0, so the result is correct.
283 * If either of P1 or P2 is infinity, but not both, then z3 == 0, which is
285 * h == 0 only if u1 == u2; this happens in two cases:
286 * -- if s1 == s2 then P1 and/or P2 is infinity, or P1 == P2
287 * -- if s1 != s2 then P1 + P2 == infinity (but neither P1 or P2 is infinity)
289 * Thus, the following situations are not handled correctly:
290 * -- P1 = 0 and P2 != 0
291 * -- P1 != 0 and P2 = 0
293 * All other cases are properly computed. However, even in "incorrect"
294 * situations, the three coordinates still are properly formed field
297 * The returned flag is cleared if r == 0. This happens in the following
299 * -- Both points are on the same horizontal line (same Y coordinate).
300 * -- Both points are infinity.
301 * -- One point is infinity and the other is on line Y = 0.
302 * The third case cannot happen with our curves (there is no valid point
303 * on line Y = 0 since that would be a point of order 2). If the two
304 * source points are non-infinity, then remains only the case where the
305 * two points are on the same horizontal line.
307 * This allows us to detect the "P1 == P2" case, assuming that P1 != 0 and
309 * -- If the returned value is not the point at infinity, then it was properly
311 * -- Otherwise, if the returned flag is 1, then P1+P2 = 0, and the result
312 * is indeed the point at infinity.
313 * -- Otherwise (result is infinity, flag is 0), then P1 = P2 and we should
314 * use the 'double' code.
316 * Cost: 16 multiplications
318 static const uint16_t code_add[] = {
320 * Compute u1 = x1*z2^2 (in t1) and s1 = y1*z2^3 (in t3).
328 * Compute u2 = x2*z1^2 (in t2) and s2 = y2*z1^3 (in t4).
336 * Compute h = u2 - u1 (in t2) and r = s2 - s1 (in t4).
342 * Report cases where r = 0 through the returned flag.
347 * Compute u1*h^2 (in t6) and h^3 (in t5).
354 * Compute x3 = r^2 - h^3 - 2*u1*h^2.
355 * t1 and t7 can be used as scratch registers.
363 * Compute y3 = r*(u1*h^2 - x3) - s1*h^3.
371 * Compute z3 = h*z1*z2.
380 * Check that the point is on the curve. This code snippet assumes the
381 * following conventions:
382 * -- Coordinates x and y have been freshly decoded in P1 (but not
383 * converted to Montgomery coordinates yet).
384 * -- P2x, P2y and P2z are set to, respectively, R^2, b*R and 1.
386 static const uint16_t code_check[] = {
388 /* Convert x and y to Montgomery representation. */
394 /* Compute x^3 in t1. */
398 /* Subtract 3*x from t1. */
406 /* Compute y^2 in t2. */
409 /* Compare y^2 with x^3 - 3*x + b; they must match. */
413 /* Set z to 1 (in Montgomery representation). */
420 * Conversion back to affine coordinates. This code snippet assumes that
421 * the z coordinate of P2 is set to 1 (not in Montgomery representation).
423 static const uint16_t code_affine[] = {
425 /* Save z*R in t1. */
428 /* Compute z^3 in t2. */
433 /* Invert to (1/z^3) in t2. */
440 /* Compute (1/z^2) in t3. */
451 run_code(jacobian *P1, const jacobian *P2,
452 const curve_params *cc, const uint16_t *code)
455 uint32_t t[13][I31_LEN];
461 * Copy the two operands in the dedicated registers.
463 memcpy(t[P1x], P1->c, 3 * I31_LEN * sizeof(uint32_t));
464 memcpy(t[P2x], P2->c, 3 * I31_LEN * sizeof(uint32_t));
470 unsigned op, d, a, b;
476 d = (op >> 8) & 0x0F;
477 a = (op >> 4) & 0x0F;
483 unsigned char tp[(BR_MAX_EC_SIZE + 7) >> 3];
486 memcpy(t[d], t[a], I31_LEN * sizeof(uint32_t));
489 ctl = br_i31_add(t[d], t[a], 1);
490 ctl |= NOT(br_i31_sub(t[d], cc->p, 0));
491 br_i31_sub(t[d], cc->p, ctl);
494 br_i31_add(t[d], cc->p, br_i31_sub(t[d], t[a], 1));
497 br_i31_montymul(t[d], t[a], t[b], cc->p, cc->p0i);
500 plen = (cc->p[0] - (cc->p[0] >> 5) + 7) >> 3;
501 br_i31_encode(tp, plen, cc->p);
503 br_i31_modpow(t[d], tp, plen,
504 cc->p, cc->p0i, t[a], t[b]);
507 r &= ~br_i31_iszero(t[d]);
515 memcpy(P1->c, t[P1x], 3 * I31_LEN * sizeof(uint32_t));
520 set_one(uint32_t *x, const uint32_t *p)
524 plen = (p[0] + 63) >> 5;
525 memset(x, 0, plen * sizeof *x);
531 point_zero(jacobian *P, const curve_params *cc)
533 memset(P, 0, sizeof *P);
534 P->c[0][0] = P->c[1][0] = P->c[2][0] = cc->p[0];
538 point_double(jacobian *P, const curve_params *cc)
540 run_code(P, P, cc, code_double);
543 static inline uint32_t
544 point_add(jacobian *P1, const jacobian *P2, const curve_params *cc)
546 return run_code(P1, P2, cc, code_add);
550 point_mul(jacobian *P, const unsigned char *x, size_t xlen,
551 const curve_params *cc)
554 * We do a simple double-and-add ladder with a 2-bit window
555 * to make only one add every two doublings. We thus first
556 * precompute 2P and 3P in some local buffers.
558 * We always perform two doublings and one addition; the
559 * addition is with P, 2P and 3P and is done in a temporary
562 * The addition code cannot handle cases where one of the
563 * operands is infinity, which is the case at the start of the
564 * ladder. We therefore need to maintain a flag that controls
568 jacobian P2, P3, Q, T, U;
570 memcpy(&P2, P, sizeof P2);
571 point_double(&P2, cc);
572 memcpy(&P3, P, sizeof P3);
573 point_add(&P3, &P2, cc);
577 while (xlen -- > 0) {
580 for (k = 6; k >= 0; k -= 2) {
584 point_double(&Q, cc);
585 point_double(&Q, cc);
586 memcpy(&T, P, sizeof T);
587 memcpy(&U, &Q, sizeof U);
588 bits = (*x >> k) & (uint32_t)3;
590 CCOPY(EQ(bits, 2), &T, &P2, sizeof T);
591 CCOPY(EQ(bits, 3), &T, &P3, sizeof T);
592 point_add(&U, &T, cc);
593 CCOPY(bnz & qz, &Q, &T, sizeof Q);
594 CCOPY(bnz & ~qz, &Q, &U, sizeof Q);
599 memcpy(P, &Q, sizeof Q);
603 * Decode point into Jacobian coordinates. This function does not support
604 * the point at infinity. If the point is invalid then this returns 0, but
605 * the coordinates are still set to properly formed field elements.
608 point_decode(jacobian *P, const void *src, size_t len, const curve_params *cc)
611 * Points must use uncompressed format:
612 * -- first byte is 0x04;
613 * -- coordinates X and Y use unsigned big-endian, with the same
614 * length as the field modulus.
616 * We don't support hybrid format (uncompressed, but first byte
617 * has value 0x06 or 0x07, depending on the least significant bit
618 * of Y) because it is rather useless, and explicitly forbidden
619 * by PKIX (RFC 5480, section 2.2).
621 * We don't support compressed format either, because it is not
622 * much used in practice (there are or were patent-related
623 * concerns about point compression, which explains the lack of
624 * generalised support). Also, point compression support would
625 * need a bit more code.
627 const unsigned char *buf;
634 plen = (cc->p[0] - (cc->p[0] >> 5) + 7) >> 3;
635 if (len != 1 + (plen << 1)) {
638 r = br_i31_decode_mod(P->c[0], buf + 1, plen, cc->p);
639 r &= br_i31_decode_mod(P->c[1], buf + 1 + plen, plen, cc->p);
644 r &= EQ(buf[0], 0x04);
646 r &= EQ(buf[0], 0x04) | (EQ(buf[0] & 0xFE, 0x06)
647 & ~(uint32_t)(buf[0] ^ buf[plen << 1]));
651 * Convert coordinates and check that the point is valid.
653 zlen = ((cc->p[0] + 63) >> 5) * sizeof(uint32_t);
654 memcpy(Q.c[0], cc->R2, zlen);
655 memcpy(Q.c[1], cc->b, zlen);
656 set_one(Q.c[2], cc->p);
657 r &= ~run_code(P, &Q, cc, code_check);
662 * Encode a point. This method assumes that the point is correct and is
663 * not the point at infinity. Encoded size is always 1+2*plen, where
664 * plen is the field modulus length, in bytes.
667 point_encode(void *dst, const jacobian *P, const curve_params *cc)
677 plen = (xbl + 7) >> 3;
679 memcpy(&Q, P, sizeof *P);
680 set_one(T.c[2], cc->p);
681 run_code(&Q, &T, cc, code_affine);
682 br_i31_encode(buf + 1, plen, Q.c[0]);
683 br_i31_encode(buf + 1 + plen, plen, Q.c[1]);
686 static const br_ec_curve_def *
687 id_to_curve_def(int curve)
690 case BR_EC_secp256r1:
691 return &br_secp256r1;
692 case BR_EC_secp384r1:
693 return &br_secp384r1;
694 case BR_EC_secp521r1:
695 return &br_secp521r1;
700 static const unsigned char *
701 api_generator(int curve, size_t *len)
703 const br_ec_curve_def *cd;
705 cd = id_to_curve_def(curve);
706 *len = cd->generator_len;
707 return cd->generator;
710 static const unsigned char *
711 api_order(int curve, size_t *len)
713 const br_ec_curve_def *cd;
715 cd = id_to_curve_def(curve);
716 *len = cd->order_len;
721 api_xoff(int curve, size_t *len)
723 api_generator(curve, len);
729 api_mul(unsigned char *G, size_t Glen,
730 const unsigned char *x, size_t xlen, int curve)
733 const curve_params *cc;
736 cc = id_to_curve(curve);
737 r = point_decode(&P, G, Glen, cc);
738 point_mul(&P, x, xlen, cc);
739 point_encode(G, &P, cc);
744 api_mulgen(unsigned char *R,
745 const unsigned char *x, size_t xlen, int curve)
747 const unsigned char *G;
750 G = api_generator(curve, &Glen);
752 api_mul(R, Glen, x, xlen, curve);
757 api_muladd(unsigned char *A, const unsigned char *B, size_t len,
758 const unsigned char *x, size_t xlen,
759 const unsigned char *y, size_t ylen, int curve)
762 const curve_params *cc;
766 * TODO: see about merging the two ladders. Right now, we do
767 * two independent point multiplications, which is a bit
768 * wasteful of CPU resources (but yields short code).
771 cc = id_to_curve(curve);
772 r = point_decode(&P, A, len, cc);
776 B = api_generator(curve, &Glen);
778 r &= point_decode(&Q, B, len, cc);
779 point_mul(&P, x, xlen, cc);
780 point_mul(&Q, y, ylen, cc);
783 * We want to compute P+Q. Since the base points A and B are distinct
784 * from infinity, and the multipliers are non-zero and lower than the
785 * curve order, then we know that P and Q are non-infinity. This
786 * leaves two special situations to test for:
787 * -- If P = Q then we must use point_double().
788 * -- If P+Q = 0 then we must report an error.
790 t = point_add(&P, &Q, cc);
791 point_double(&Q, cc);
792 z = br_i31_iszero(P.c[2]);
795 * If z is 1 then either P+Q = 0 (t = 1) or P = Q (t = 0). So we
796 * have the following:
798 * z = 0, t = 0 return P (normal addition)
799 * z = 0, t = 1 return P (normal addition)
800 * z = 1, t = 0 return Q (a 'double' case)
801 * z = 1, t = 1 report an error (P+Q = 0)
803 CCOPY(z & ~t, &P, &Q, sizeof Q);
804 point_encode(A, &P, cc);
810 /* see bearssl_ec.h */
811 const br_ec_impl br_ec_prime_i31 = {
812 (uint32_t)0x03800000,