xref: /freebsd/crypto/openssl/crypto/ec/ec_mult.c (revision f25b8c9fb4f58cf61adb47d7570abe7caa6d385d)
1 /*
2  * Copyright 2001-2023 The OpenSSL Project Authors. All Rights Reserved.
3  * Copyright (c) 2002, Oracle and/or its affiliates. All rights reserved
4  *
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
9  */
10 
11 /*
12  * ECDSA low level APIs are deprecated for public use, but still ok for
13  * internal use.
14  */
15 #include "internal/deprecated.h"
16 
17 #include <string.h>
18 #include <openssl/err.h>
19 
20 #include "internal/cryptlib.h"
21 #include "crypto/bn.h"
22 #include "ec_local.h"
23 #include "internal/refcount.h"
24 
25 /*
26  * This file implements the wNAF-based interleaving multi-exponentiation method
27  * Formerly at:
28  *   http://www.informatik.tu-darmstadt.de/TI/Mitarbeiter/moeller.html#multiexp
29  * You might now find it here:
30  *   http://link.springer.com/chapter/10.1007%2F3-540-45537-X_13
31  *   http://www.bmoeller.de/pdf/TI-01-08.multiexp.pdf
32  * For multiplication with precomputation, we use wNAF splitting, formerly at:
33  *   http://www.informatik.tu-darmstadt.de/TI/Mitarbeiter/moeller.html#fastexp
34  */
35 
36 /* structure for precomputed multiples of the generator */
37 struct ec_pre_comp_st {
38     const EC_GROUP *group; /* parent EC_GROUP object */
39     size_t blocksize; /* block size for wNAF splitting */
40     size_t numblocks; /* max. number of blocks for which we have
41                        * precomputation */
42     size_t w; /* window size */
43     EC_POINT **points; /* array with pre-calculated multiples of
44                         * generator: 'num' pointers to EC_POINT
45                         * objects followed by a NULL */
46     size_t num; /* numblocks * 2^(w-1) */
47     CRYPTO_REF_COUNT references;
48 };
49 
ec_pre_comp_new(const EC_GROUP * group)50 static EC_PRE_COMP *ec_pre_comp_new(const EC_GROUP *group)
51 {
52     EC_PRE_COMP *ret = NULL;
53 
54     if (!group)
55         return NULL;
56 
57     ret = OPENSSL_zalloc(sizeof(*ret));
58     if (ret == NULL)
59         return ret;
60 
61     ret->group = group;
62     ret->blocksize = 8; /* default */
63     ret->w = 4; /* default */
64 
65     if (!CRYPTO_NEW_REF(&ret->references, 1)) {
66         OPENSSL_free(ret);
67         return NULL;
68     }
69     return ret;
70 }
71 
EC_ec_pre_comp_dup(EC_PRE_COMP * pre)72 EC_PRE_COMP *EC_ec_pre_comp_dup(EC_PRE_COMP *pre)
73 {
74     int i;
75     if (pre != NULL)
76         CRYPTO_UP_REF(&pre->references, &i);
77     return pre;
78 }
79 
EC_ec_pre_comp_free(EC_PRE_COMP * pre)80 void EC_ec_pre_comp_free(EC_PRE_COMP *pre)
81 {
82     int i;
83 
84     if (pre == NULL)
85         return;
86 
87     CRYPTO_DOWN_REF(&pre->references, &i);
88     REF_PRINT_COUNT("EC_ec", i, pre);
89     if (i > 0)
90         return;
91     REF_ASSERT_ISNT(i < 0);
92 
93     if (pre->points != NULL) {
94         EC_POINT **pts;
95 
96         for (pts = pre->points; *pts != NULL; pts++)
97             EC_POINT_free(*pts);
98         OPENSSL_free(pre->points);
99     }
100     CRYPTO_FREE_REF(&pre->references);
101     OPENSSL_free(pre);
102 }
103 
104 #define EC_POINT_BN_set_flags(P, flags) \
105     do {                                \
106         BN_set_flags((P)->X, (flags));  \
107         BN_set_flags((P)->Y, (flags));  \
108         BN_set_flags((P)->Z, (flags));  \
109     } while (0)
110 
111 /*-
112  * This functions computes a single point multiplication over the EC group,
113  * using, at a high level, a Montgomery ladder with conditional swaps, with
114  * various timing attack defenses.
115  *
116  * It performs either a fixed point multiplication
117  *          (scalar * generator)
118  * when point is NULL, or a variable point multiplication
119  *          (scalar * point)
120  * when point is not NULL.
121  *
122  * `scalar` cannot be NULL and should be in the range [0,n) otherwise all
123  * constant time bets are off (where n is the cardinality of the EC group).
124  *
125  * This function expects `group->order` and `group->cardinality` to be well
126  * defined and non-zero: it fails with an error code otherwise.
127  *
128  * NB: This says nothing about the constant-timeness of the ladder step
129  * implementation (i.e., the default implementation is based on EC_POINT_add and
130  * EC_POINT_dbl, which of course are not constant time themselves) or the
131  * underlying multiprecision arithmetic.
132  *
133  * The product is stored in `r`.
134  *
135  * This is an internal function: callers are in charge of ensuring that the
136  * input parameters `group`, `r`, `scalar` and `ctx` are not NULL.
137  *
138  * Returns 1 on success, 0 otherwise.
139  */
ossl_ec_scalar_mul_ladder(const EC_GROUP * group,EC_POINT * r,const BIGNUM * scalar,const EC_POINT * point,BN_CTX * ctx)140 int ossl_ec_scalar_mul_ladder(const EC_GROUP *group, EC_POINT *r,
141     const BIGNUM *scalar, const EC_POINT *point,
142     BN_CTX *ctx)
143 {
144     int i, cardinality_bits, group_top, kbit, pbit, Z_is_one;
145     EC_POINT *p = NULL;
146     EC_POINT *s = NULL;
147     BIGNUM *k = NULL;
148     BIGNUM *lambda = NULL;
149     BIGNUM *cardinality = NULL;
150     int ret = 0;
151 
152     /* early exit if the input point is the point at infinity */
153     if (point != NULL && EC_POINT_is_at_infinity(group, point))
154         return EC_POINT_set_to_infinity(group, r);
155 
156     if (BN_is_zero(group->order)) {
157         ERR_raise(ERR_LIB_EC, EC_R_UNKNOWN_ORDER);
158         return 0;
159     }
160     if (BN_is_zero(group->cofactor)) {
161         ERR_raise(ERR_LIB_EC, EC_R_UNKNOWN_COFACTOR);
162         return 0;
163     }
164 
165     BN_CTX_start(ctx);
166 
167     if (((p = EC_POINT_new(group)) == NULL)
168         || ((s = EC_POINT_new(group)) == NULL)) {
169         ERR_raise(ERR_LIB_EC, ERR_R_EC_LIB);
170         goto err;
171     }
172 
173     if (point == NULL) {
174         if (!EC_POINT_copy(p, group->generator)) {
175             ERR_raise(ERR_LIB_EC, ERR_R_EC_LIB);
176             goto err;
177         }
178     } else {
179         if (!EC_POINT_copy(p, point)) {
180             ERR_raise(ERR_LIB_EC, ERR_R_EC_LIB);
181             goto err;
182         }
183     }
184 
185     EC_POINT_BN_set_flags(p, BN_FLG_CONSTTIME);
186     EC_POINT_BN_set_flags(r, BN_FLG_CONSTTIME);
187     EC_POINT_BN_set_flags(s, BN_FLG_CONSTTIME);
188 
189     cardinality = BN_CTX_get(ctx);
190     lambda = BN_CTX_get(ctx);
191     k = BN_CTX_get(ctx);
192     if (k == NULL) {
193         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
194         goto err;
195     }
196 
197     if (!BN_mul(cardinality, group->order, group->cofactor, ctx)) {
198         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
199         goto err;
200     }
201 
202     /*
203      * Group cardinalities are often on a word boundary.
204      * So when we pad the scalar, some timing diff might
205      * pop if it needs to be expanded due to carries.
206      * So expand ahead of time.
207      */
208     cardinality_bits = BN_num_bits(cardinality);
209     group_top = bn_get_top(cardinality);
210     if ((bn_wexpand(k, group_top + 2) == NULL)
211         || (bn_wexpand(lambda, group_top + 2) == NULL)) {
212         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
213         goto err;
214     }
215 
216     if (!BN_copy(k, scalar)) {
217         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
218         goto err;
219     }
220 
221     BN_set_flags(k, BN_FLG_CONSTTIME);
222 
223     if ((BN_num_bits(k) > cardinality_bits) || (BN_is_negative(k))) {
224         /*-
225          * this is an unusual input, and we don't guarantee
226          * constant-timeness
227          */
228         if (!BN_nnmod(k, k, cardinality, ctx)) {
229             ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
230             goto err;
231         }
232     }
233 
234     if (!BN_add(lambda, k, cardinality)) {
235         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
236         goto err;
237     }
238     BN_set_flags(lambda, BN_FLG_CONSTTIME);
239     if (!BN_add(k, lambda, cardinality)) {
240         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
241         goto err;
242     }
243     /*
244      * lambda := scalar + cardinality
245      * k := scalar + 2*cardinality
246      */
247     kbit = BN_is_bit_set(lambda, cardinality_bits);
248     BN_consttime_swap(kbit, k, lambda, group_top + 2);
249 
250     group_top = bn_get_top(group->field);
251     if ((bn_wexpand(s->X, group_top) == NULL)
252         || (bn_wexpand(s->Y, group_top) == NULL)
253         || (bn_wexpand(s->Z, group_top) == NULL)
254         || (bn_wexpand(r->X, group_top) == NULL)
255         || (bn_wexpand(r->Y, group_top) == NULL)
256         || (bn_wexpand(r->Z, group_top) == NULL)
257         || (bn_wexpand(p->X, group_top) == NULL)
258         || (bn_wexpand(p->Y, group_top) == NULL)
259         || (bn_wexpand(p->Z, group_top) == NULL)) {
260         ERR_raise(ERR_LIB_EC, ERR_R_BN_LIB);
261         goto err;
262     }
263 
264     /* ensure input point is in affine coords for ladder step efficiency */
265     if (!p->Z_is_one && (group->meth->make_affine == NULL || !group->meth->make_affine(group, p, ctx))) {
266         ERR_raise(ERR_LIB_EC, ERR_R_EC_LIB);
267         goto err;
268     }
269 
270     /* Initialize the Montgomery ladder */
271     if (!ec_point_ladder_pre(group, r, s, p, ctx)) {
272         ERR_raise(ERR_LIB_EC, EC_R_LADDER_PRE_FAILURE);
273         goto err;
274     }
275 
276     /* top bit is a 1, in a fixed pos */
277     pbit = 1;
278 
279 #define EC_POINT_CSWAP(c, a, b, w, t)              \
280     do {                                           \
281         BN_consttime_swap(c, (a)->X, (b)->X, w);   \
282         BN_consttime_swap(c, (a)->Y, (b)->Y, w);   \
283         BN_consttime_swap(c, (a)->Z, (b)->Z, w);   \
284         t = ((a)->Z_is_one ^ (b)->Z_is_one) & (c); \
285         (a)->Z_is_one ^= (t);                      \
286         (b)->Z_is_one ^= (t);                      \
287     } while (0)
288 
289     /*-
290      * The ladder step, with branches, is
291      *
292      * k[i] == 0: S = add(R, S), R = dbl(R)
293      * k[i] == 1: R = add(S, R), S = dbl(S)
294      *
295      * Swapping R, S conditionally on k[i] leaves you with state
296      *
297      * k[i] == 0: T, U = R, S
298      * k[i] == 1: T, U = S, R
299      *
300      * Then perform the ECC ops.
301      *
302      * U = add(T, U)
303      * T = dbl(T)
304      *
305      * Which leaves you with state
306      *
307      * k[i] == 0: U = add(R, S), T = dbl(R)
308      * k[i] == 1: U = add(S, R), T = dbl(S)
309      *
310      * Swapping T, U conditionally on k[i] leaves you with state
311      *
312      * k[i] == 0: R, S = T, U
313      * k[i] == 1: R, S = U, T
314      *
315      * Which leaves you with state
316      *
317      * k[i] == 0: S = add(R, S), R = dbl(R)
318      * k[i] == 1: R = add(S, R), S = dbl(S)
319      *
320      * So we get the same logic, but instead of a branch it's a
321      * conditional swap, followed by ECC ops, then another conditional swap.
322      *
323      * Optimization: The end of iteration i and start of i-1 looks like
324      *
325      * ...
326      * CSWAP(k[i], R, S)
327      * ECC
328      * CSWAP(k[i], R, S)
329      * (next iteration)
330      * CSWAP(k[i-1], R, S)
331      * ECC
332      * CSWAP(k[i-1], R, S)
333      * ...
334      *
335      * So instead of two contiguous swaps, you can merge the condition
336      * bits and do a single swap.
337      *
338      * k[i]   k[i-1]    Outcome
339      * 0      0         No Swap
340      * 0      1         Swap
341      * 1      0         Swap
342      * 1      1         No Swap
343      *
344      * This is XOR. pbit tracks the previous bit of k.
345      */
346 
347     for (i = cardinality_bits - 1; i >= 0; i--) {
348         kbit = BN_is_bit_set(k, i) ^ pbit;
349         EC_POINT_CSWAP(kbit, r, s, group_top, Z_is_one);
350 
351         /* Perform a single step of the Montgomery ladder */
352         if (!ec_point_ladder_step(group, r, s, p, ctx)) {
353             ERR_raise(ERR_LIB_EC, EC_R_LADDER_STEP_FAILURE);
354             goto err;
355         }
356         /*
357          * pbit logic merges this cswap with that of the
358          * next iteration
359          */
360         pbit ^= kbit;
361     }
362     /* one final cswap to move the right value into r */
363     EC_POINT_CSWAP(pbit, r, s, group_top, Z_is_one);
364 #undef EC_POINT_CSWAP
365 
366     /* Finalize ladder (and recover full point coordinates) */
367     if (!ec_point_ladder_post(group, r, s, p, ctx)) {
368         ERR_raise(ERR_LIB_EC, EC_R_LADDER_POST_FAILURE);
369         goto err;
370     }
371 
372     ret = 1;
373 
374 err:
375     EC_POINT_free(p);
376     EC_POINT_clear_free(s);
377     BN_CTX_end(ctx);
378 
379     return ret;
380 }
381 
382 #undef EC_POINT_BN_set_flags
383 
384 /*
385  * Table could be optimised for the wNAF-based implementation,
386  * sometimes smaller windows will give better performance (thus the
387  * boundaries should be increased)
388  */
389 #define EC_window_bits_for_scalar_size(b)      \
390     ((size_t)((b) >= 2000 ? 6 : (b) >= 800 ? 5 \
391             : (b) >= 300                   ? 4 \
392             : (b) >= 70                    ? 3 \
393             : (b) >= 20                    ? 2 \
394                                            : 1))
395 
396 /*-
397  * Compute
398  *      \sum scalars[i]*points[i],
399  * also including
400  *      scalar*generator
401  * in the addition if scalar != NULL
402  */
ossl_ec_wNAF_mul(const EC_GROUP * group,EC_POINT * r,const BIGNUM * scalar,size_t num,const EC_POINT * points[],const BIGNUM * scalars[],BN_CTX * ctx)403 int ossl_ec_wNAF_mul(const EC_GROUP *group, EC_POINT *r, const BIGNUM *scalar,
404     size_t num, const EC_POINT *points[],
405     const BIGNUM *scalars[], BN_CTX *ctx)
406 {
407     const EC_POINT *generator = NULL;
408     EC_POINT *tmp = NULL;
409     size_t totalnum;
410     size_t blocksize = 0, numblocks = 0; /* for wNAF splitting */
411     size_t pre_points_per_block = 0;
412     size_t i, j;
413     int k;
414     int r_is_inverted = 0;
415     int r_is_at_infinity = 1;
416     size_t *wsize = NULL; /* individual window sizes */
417     signed char **wNAF = NULL; /* individual wNAFs */
418     size_t *wNAF_len = NULL;
419     size_t max_len = 0;
420     size_t num_val;
421     EC_POINT **val = NULL; /* precomputation */
422     EC_POINT **v;
423     EC_POINT ***val_sub = NULL; /* pointers to sub-arrays of 'val' or
424                                  * 'pre_comp->points' */
425     const EC_PRE_COMP *pre_comp = NULL;
426     int num_scalar = 0; /* flag: will be set to 1 if 'scalar' must be
427                          * treated like other scalars, i.e.
428                          * precomputation is not available */
429     int ret = 0;
430 
431     if (!BN_is_zero(group->order) && !BN_is_zero(group->cofactor)) {
432         /*-
433          * Handle the common cases where the scalar is secret, enforcing a
434          * scalar multiplication implementation based on a Montgomery ladder,
435          * with various timing attack defenses.
436          */
437         if ((scalar != group->order) && (scalar != NULL) && (num == 0)) {
438             /*-
439              * In this case we want to compute scalar * GeneratorPoint: this
440              * codepath is reached most prominently by (ephemeral) key
441              * generation of EC cryptosystems (i.e. ECDSA keygen and sign setup,
442              * ECDH keygen/first half), where the scalar is always secret. This
443              * is why we ignore if BN_FLG_CONSTTIME is actually set and we
444              * always call the ladder version.
445              */
446             return ossl_ec_scalar_mul_ladder(group, r, scalar, NULL, ctx);
447         }
448         if ((scalar == NULL) && (num == 1) && (scalars[0] != group->order)) {
449             /*-
450              * In this case we want to compute scalar * VariablePoint: this
451              * codepath is reached most prominently by the second half of ECDH,
452              * where the secret scalar is multiplied by the peer's public point.
453              * To protect the secret scalar, we ignore if BN_FLG_CONSTTIME is
454              * actually set and we always call the ladder version.
455              */
456             return ossl_ec_scalar_mul_ladder(group, r, scalars[0], points[0],
457                 ctx);
458         }
459     }
460 
461     if (scalar != NULL) {
462         generator = EC_GROUP_get0_generator(group);
463         if (generator == NULL) {
464             ERR_raise(ERR_LIB_EC, EC_R_UNDEFINED_GENERATOR);
465             goto err;
466         }
467 
468         /* look if we can use precomputed multiples of generator */
469 
470         pre_comp = group->pre_comp.ec;
471         if (pre_comp && pre_comp->numblocks
472             && (EC_POINT_cmp(group, generator, pre_comp->points[0], ctx) == 0)) {
473             blocksize = pre_comp->blocksize;
474 
475             /*
476              * determine maximum number of blocks that wNAF splitting may
477              * yield (NB: maximum wNAF length is bit length plus one)
478              */
479             numblocks = (BN_num_bits(scalar) / blocksize) + 1;
480 
481             /*
482              * we cannot use more blocks than we have precomputation for
483              */
484             if (numblocks > pre_comp->numblocks)
485                 numblocks = pre_comp->numblocks;
486 
487             pre_points_per_block = (size_t)1 << (pre_comp->w - 1);
488 
489             /* check that pre_comp looks sane */
490             if (pre_comp->num != (pre_comp->numblocks * pre_points_per_block)) {
491                 ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
492                 goto err;
493             }
494         } else {
495             /* can't use precomputation */
496             pre_comp = NULL;
497             numblocks = 1;
498             num_scalar = 1; /* treat 'scalar' like 'num'-th element of
499                              * 'scalars' */
500         }
501     }
502 
503     totalnum = num + numblocks;
504 
505     wsize = OPENSSL_malloc(totalnum * sizeof(wsize[0]));
506     wNAF_len = OPENSSL_malloc(totalnum * sizeof(wNAF_len[0]));
507     /* include space for pivot */
508     wNAF = OPENSSL_malloc((totalnum + 1) * sizeof(wNAF[0]));
509     val_sub = OPENSSL_malloc(totalnum * sizeof(val_sub[0]));
510 
511     /* Ensure wNAF is initialised in case we end up going to err */
512     if (wNAF != NULL)
513         wNAF[0] = NULL; /* preliminary pivot */
514 
515     if (wsize == NULL || wNAF_len == NULL || wNAF == NULL || val_sub == NULL)
516         goto err;
517 
518     /*
519      * num_val will be the total number of temporarily precomputed points
520      */
521     num_val = 0;
522 
523     for (i = 0; i < num + num_scalar; i++) {
524         size_t bits;
525 
526         bits = i < num ? BN_num_bits(scalars[i]) : BN_num_bits(scalar);
527         wsize[i] = EC_window_bits_for_scalar_size(bits);
528         num_val += (size_t)1 << (wsize[i] - 1);
529         wNAF[i + 1] = NULL; /* make sure we always have a pivot */
530         wNAF[i] = bn_compute_wNAF((i < num ? scalars[i] : scalar), wsize[i],
531             &wNAF_len[i]);
532         if (wNAF[i] == NULL)
533             goto err;
534         if (wNAF_len[i] > max_len)
535             max_len = wNAF_len[i];
536     }
537 
538     if (numblocks) {
539         /* we go here iff scalar != NULL */
540 
541         if (pre_comp == NULL) {
542             if (num_scalar != 1) {
543                 ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
544                 goto err;
545             }
546             /* we have already generated a wNAF for 'scalar' */
547         } else {
548             signed char *tmp_wNAF = NULL;
549             size_t tmp_len = 0;
550 
551             if (num_scalar != 0) {
552                 ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
553                 goto err;
554             }
555 
556             /*
557              * use the window size for which we have precomputation
558              */
559             wsize[num] = pre_comp->w;
560             tmp_wNAF = bn_compute_wNAF(scalar, wsize[num], &tmp_len);
561             if (!tmp_wNAF)
562                 goto err;
563 
564             if (tmp_len <= max_len) {
565                 /*
566                  * One of the other wNAFs is at least as long as the wNAF
567                  * belonging to the generator, so wNAF splitting will not buy
568                  * us anything.
569                  */
570 
571                 numblocks = 1;
572                 totalnum = num + 1; /* don't use wNAF splitting */
573                 wNAF[num] = tmp_wNAF;
574                 wNAF[num + 1] = NULL;
575                 wNAF_len[num] = tmp_len;
576                 /*
577                  * pre_comp->points starts with the points that we need here:
578                  */
579                 val_sub[num] = pre_comp->points;
580             } else {
581                 /*
582                  * don't include tmp_wNAF directly into wNAF array - use wNAF
583                  * splitting and include the blocks
584                  */
585 
586                 signed char *pp;
587                 EC_POINT **tmp_points;
588 
589                 if (tmp_len < numblocks * blocksize) {
590                     /*
591                      * possibly we can do with fewer blocks than estimated
592                      */
593                     numblocks = (tmp_len + blocksize - 1) / blocksize;
594                     if (numblocks > pre_comp->numblocks) {
595                         ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
596                         OPENSSL_free(tmp_wNAF);
597                         goto err;
598                     }
599                     totalnum = num + numblocks;
600                 }
601 
602                 /* split wNAF in 'numblocks' parts */
603                 pp = tmp_wNAF;
604                 tmp_points = pre_comp->points;
605 
606                 for (i = num; i < totalnum; i++) {
607                     if (i < totalnum - 1) {
608                         wNAF_len[i] = blocksize;
609                         if (tmp_len < blocksize) {
610                             ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
611                             OPENSSL_free(tmp_wNAF);
612                             goto err;
613                         }
614                         tmp_len -= blocksize;
615                     } else
616                         /*
617                          * last block gets whatever is left (this could be
618                          * more or less than 'blocksize'!)
619                          */
620                         wNAF_len[i] = tmp_len;
621 
622                     wNAF[i + 1] = NULL;
623                     wNAF[i] = OPENSSL_malloc(wNAF_len[i]);
624                     if (wNAF[i] == NULL) {
625                         OPENSSL_free(tmp_wNAF);
626                         goto err;
627                     }
628                     memcpy(wNAF[i], pp, wNAF_len[i]);
629                     if (wNAF_len[i] > max_len)
630                         max_len = wNAF_len[i];
631 
632                     if (*tmp_points == NULL) {
633                         ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
634                         OPENSSL_free(tmp_wNAF);
635                         goto err;
636                     }
637                     val_sub[i] = tmp_points;
638                     tmp_points += pre_points_per_block;
639                     pp += blocksize;
640                 }
641                 OPENSSL_free(tmp_wNAF);
642             }
643         }
644     }
645 
646     /*
647      * All points we precompute now go into a single array 'val'.
648      * 'val_sub[i]' is a pointer to the subarray for the i-th point, or to a
649      * subarray of 'pre_comp->points' if we already have precomputation.
650      */
651     val = OPENSSL_malloc((num_val + 1) * sizeof(val[0]));
652     if (val == NULL)
653         goto err;
654     val[num_val] = NULL; /* pivot element */
655 
656     /* allocate points for precomputation */
657     v = val;
658     for (i = 0; i < num + num_scalar; i++) {
659         val_sub[i] = v;
660         for (j = 0; j < ((size_t)1 << (wsize[i] - 1)); j++) {
661             *v = EC_POINT_new(group);
662             if (*v == NULL)
663                 goto err;
664             v++;
665         }
666     }
667     if (!(v == val + num_val)) {
668         ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
669         goto err;
670     }
671 
672     if ((tmp = EC_POINT_new(group)) == NULL)
673         goto err;
674 
675     /*-
676      * prepare precomputed values:
677      *    val_sub[i][0] :=     points[i]
678      *    val_sub[i][1] := 3 * points[i]
679      *    val_sub[i][2] := 5 * points[i]
680      *    ...
681      */
682     for (i = 0; i < num + num_scalar; i++) {
683         if (i < num) {
684             if (!EC_POINT_copy(val_sub[i][0], points[i]))
685                 goto err;
686         } else {
687             if (!EC_POINT_copy(val_sub[i][0], generator))
688                 goto err;
689         }
690 
691         if (wsize[i] > 1) {
692             if (!EC_POINT_dbl(group, tmp, val_sub[i][0], ctx))
693                 goto err;
694             for (j = 1; j < ((size_t)1 << (wsize[i] - 1)); j++) {
695                 if (!EC_POINT_add(group, val_sub[i][j], val_sub[i][j - 1], tmp, ctx))
696                     goto err;
697             }
698         }
699     }
700 
701     if (group->meth->points_make_affine == NULL
702         || !group->meth->points_make_affine(group, num_val, val, ctx))
703         goto err;
704 
705     r_is_at_infinity = 1;
706 
707     for (k = max_len - 1; k >= 0; k--) {
708         if (!r_is_at_infinity) {
709             if (!EC_POINT_dbl(group, r, r, ctx))
710                 goto err;
711         }
712 
713         for (i = 0; i < totalnum; i++) {
714             if (wNAF_len[i] > (size_t)k) {
715                 int digit = wNAF[i][k];
716                 int is_neg;
717 
718                 if (digit) {
719                     is_neg = digit < 0;
720 
721                     if (is_neg)
722                         digit = -digit;
723 
724                     if (is_neg != r_is_inverted) {
725                         if (!r_is_at_infinity) {
726                             if (!EC_POINT_invert(group, r, ctx))
727                                 goto err;
728                         }
729                         r_is_inverted = !r_is_inverted;
730                     }
731 
732                     /* digit > 0 */
733 
734                     if (r_is_at_infinity) {
735                         if (!EC_POINT_copy(r, val_sub[i][digit >> 1]))
736                             goto err;
737 
738                         /*-
739                          * Apply coordinate blinding for EC_POINT.
740                          *
741                          * The underlying EC_METHOD can optionally implement this function:
742                          * ossl_ec_point_blind_coordinates() returns 0 in case of errors or 1 on
743                          * success or if coordinate blinding is not implemented for this
744                          * group.
745                          */
746                         if (!ossl_ec_point_blind_coordinates(group, r, ctx)) {
747                             ERR_raise(ERR_LIB_EC, EC_R_POINT_COORDINATES_BLIND_FAILURE);
748                             goto err;
749                         }
750 
751                         r_is_at_infinity = 0;
752                     } else {
753                         if (!EC_POINT_add(group, r, r, val_sub[i][digit >> 1], ctx))
754                             goto err;
755                     }
756                 }
757             }
758         }
759     }
760 
761     if (r_is_at_infinity) {
762         if (!EC_POINT_set_to_infinity(group, r))
763             goto err;
764     } else {
765         if (r_is_inverted)
766             if (!EC_POINT_invert(group, r, ctx))
767                 goto err;
768     }
769 
770     ret = 1;
771 
772 err:
773     EC_POINT_free(tmp);
774     OPENSSL_free(wsize);
775     OPENSSL_free(wNAF_len);
776     if (wNAF != NULL) {
777         signed char **w;
778 
779         for (w = wNAF; *w != NULL; w++)
780             OPENSSL_free(*w);
781 
782         OPENSSL_free(wNAF);
783     }
784     if (val != NULL) {
785         for (v = val; *v != NULL; v++)
786             EC_POINT_clear_free(*v);
787 
788         OPENSSL_free(val);
789     }
790     OPENSSL_free(val_sub);
791     return ret;
792 }
793 
794 /*-
795  * ossl_ec_wNAF_precompute_mult()
796  * creates an EC_PRE_COMP object with preprecomputed multiples of the generator
797  * for use with wNAF splitting as implemented in ossl_ec_wNAF_mul().
798  *
799  * 'pre_comp->points' is an array of multiples of the generator
800  * of the following form:
801  * points[0] =     generator;
802  * points[1] = 3 * generator;
803  * ...
804  * points[2^(w-1)-1] =     (2^(w-1)-1) * generator;
805  * points[2^(w-1)]   =     2^blocksize * generator;
806  * points[2^(w-1)+1] = 3 * 2^blocksize * generator;
807  * ...
808  * points[2^(w-1)*(numblocks-1)-1] = (2^(w-1)) *  2^(blocksize*(numblocks-2)) * generator
809  * points[2^(w-1)*(numblocks-1)]   =              2^(blocksize*(numblocks-1)) * generator
810  * ...
811  * points[2^(w-1)*numblocks-1]     = (2^(w-1)) *  2^(blocksize*(numblocks-1)) * generator
812  * points[2^(w-1)*numblocks]       = NULL
813  */
ossl_ec_wNAF_precompute_mult(EC_GROUP * group,BN_CTX * ctx)814 int ossl_ec_wNAF_precompute_mult(EC_GROUP *group, BN_CTX *ctx)
815 {
816     const EC_POINT *generator;
817     EC_POINT *tmp_point = NULL, *base = NULL, **var;
818     const BIGNUM *order;
819     size_t i, bits, w, pre_points_per_block, blocksize, numblocks, num;
820     EC_POINT **points = NULL;
821     EC_PRE_COMP *pre_comp;
822     int ret = 0;
823     int used_ctx = 0;
824 #ifndef FIPS_MODULE
825     BN_CTX *new_ctx = NULL;
826 #endif
827 
828     /* if there is an old EC_PRE_COMP object, throw it away */
829     EC_pre_comp_free(group);
830     if ((pre_comp = ec_pre_comp_new(group)) == NULL)
831         return 0;
832 
833     generator = EC_GROUP_get0_generator(group);
834     if (generator == NULL) {
835         ERR_raise(ERR_LIB_EC, EC_R_UNDEFINED_GENERATOR);
836         goto err;
837     }
838 
839 #ifndef FIPS_MODULE
840     if (ctx == NULL)
841         ctx = new_ctx = BN_CTX_new();
842 #endif
843     if (ctx == NULL)
844         goto err;
845 
846     BN_CTX_start(ctx);
847     used_ctx = 1;
848 
849     order = EC_GROUP_get0_order(group);
850     if (order == NULL)
851         goto err;
852     if (BN_is_zero(order)) {
853         ERR_raise(ERR_LIB_EC, EC_R_UNKNOWN_ORDER);
854         goto err;
855     }
856 
857     bits = BN_num_bits(order);
858     /*
859      * The following parameters mean we precompute (approximately) one point
860      * per bit. TBD: The combination 8, 4 is perfect for 160 bits; for other
861      * bit lengths, other parameter combinations might provide better
862      * efficiency.
863      */
864     blocksize = 8;
865     w = 4;
866     if (EC_window_bits_for_scalar_size(bits) > w) {
867         /* let's not make the window too small ... */
868         w = EC_window_bits_for_scalar_size(bits);
869     }
870 
871     numblocks = (bits + blocksize - 1) / blocksize; /* max. number of blocks
872                                                      * to use for wNAF
873                                                      * splitting */
874 
875     pre_points_per_block = (size_t)1 << (w - 1);
876     num = pre_points_per_block * numblocks; /* number of points to compute
877                                              * and store */
878 
879     points = OPENSSL_malloc(sizeof(*points) * (num + 1));
880     if (points == NULL)
881         goto err;
882 
883     var = points;
884     var[num] = NULL; /* pivot */
885     for (i = 0; i < num; i++) {
886         if ((var[i] = EC_POINT_new(group)) == NULL) {
887             ERR_raise(ERR_LIB_EC, ERR_R_EC_LIB);
888             goto err;
889         }
890     }
891 
892     if ((tmp_point = EC_POINT_new(group)) == NULL
893         || (base = EC_POINT_new(group)) == NULL) {
894         ERR_raise(ERR_LIB_EC, ERR_R_EC_LIB);
895         goto err;
896     }
897 
898     if (!EC_POINT_copy(base, generator))
899         goto err;
900 
901     /* do the precomputation */
902     for (i = 0; i < numblocks; i++) {
903         size_t j;
904 
905         if (!EC_POINT_dbl(group, tmp_point, base, ctx))
906             goto err;
907 
908         if (!EC_POINT_copy(*var++, base))
909             goto err;
910 
911         for (j = 1; j < pre_points_per_block; j++, var++) {
912             /*
913              * calculate odd multiples of the current base point
914              */
915             if (!EC_POINT_add(group, *var, tmp_point, *(var - 1), ctx))
916                 goto err;
917         }
918 
919         if (i < numblocks - 1) {
920             /*
921              * get the next base (multiply current one by 2^blocksize)
922              */
923             size_t k;
924 
925             if (blocksize <= 2) {
926                 ERR_raise(ERR_LIB_EC, ERR_R_INTERNAL_ERROR);
927                 goto err;
928             }
929 
930             if (!EC_POINT_dbl(group, base, tmp_point, ctx))
931                 goto err;
932             for (k = 2; k < blocksize; k++) {
933                 if (!EC_POINT_dbl(group, base, base, ctx))
934                     goto err;
935             }
936         }
937     }
938 
939     if (group->meth->points_make_affine == NULL
940         || !group->meth->points_make_affine(group, num, points, ctx))
941         goto err;
942 
943     pre_comp->group = group;
944     pre_comp->blocksize = blocksize;
945     pre_comp->numblocks = numblocks;
946     pre_comp->w = w;
947     pre_comp->points = points;
948     points = NULL;
949     pre_comp->num = num;
950     SETPRECOMP(group, ec, pre_comp);
951     pre_comp = NULL;
952     ret = 1;
953 
954 err:
955     if (used_ctx)
956         BN_CTX_end(ctx);
957 #ifndef FIPS_MODULE
958     BN_CTX_free(new_ctx);
959 #endif
960     EC_ec_pre_comp_free(pre_comp);
961     if (points) {
962         EC_POINT **p;
963 
964         for (p = points; *p != NULL; p++)
965             EC_POINT_free(*p);
966         OPENSSL_free(points);
967     }
968     EC_POINT_free(tmp_point);
969     EC_POINT_free(base);
970     return ret;
971 }
972 
ossl_ec_wNAF_have_precompute_mult(const EC_GROUP * group)973 int ossl_ec_wNAF_have_precompute_mult(const EC_GROUP *group)
974 {
975     return HAVEPRECOMP(group, ec);
976 }
977