2 * Copyright 1995-2022 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
10 #include "internal/cryptlib.h"
11 #include "internal/constant_time.h"
18 # define alloca _alloca
20 #elif defined(__GNUC__)
22 # define alloca(s) __builtin_alloca((s))
31 #if defined(OPENSSL_BN_ASM_MONT) && (defined(__sparc__) || defined(__sparc))
32 # include "sparc_arch.h"
33 extern unsigned int OPENSSL_sparcv9cap_P[];
34 # define SPARC_T4_MONT
37 /* maximum precomputation table size for *variable* sliding windows */
40 /* this one works - simple but works */
41 int BN_exp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, BN_CTX *ctx)
46 if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0
47 || BN_get_flags(a, BN_FLG_CONSTTIME) != 0) {
48 /* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
49 BNerr(BN_F_BN_EXP, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
54 rr = ((r == a) || (r == p)) ? BN_CTX_get(ctx) : r;
56 if (rr == NULL || v == NULL)
59 if (BN_copy(v, a) == NULL)
61 bits = BN_num_bits(p);
64 if (BN_copy(rr, a) == NULL)
71 for (i = 1; i < bits; i++) {
72 if (!BN_sqr(v, v, ctx))
74 if (BN_is_bit_set(p, i)) {
75 if (!BN_mul(rr, rr, v, ctx))
79 if (r != rr && BN_copy(r, rr) == NULL)
89 int BN_mod_exp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, const BIGNUM *m,
99 * For even modulus m = 2^k*m_odd, it might make sense to compute
100 * a^p mod m_odd and a^p mod 2^k separately (with Montgomery
101 * exponentiation for the odd part), using appropriate exponent
102 * reductions, and combine the results using the CRT.
104 * For now, we use Montgomery only if the modulus is odd; otherwise,
105 * exponentiation using the reciprocal-based quick remaindering
108 * (Timing obtained with expspeed.c [computations a^p mod m
109 * where a, p, m are of the same length: 256, 512, 1024, 2048,
110 * 4096, 8192 bits], compared to the running time of the
111 * standard algorithm:
113 * BN_mod_exp_mont 33 .. 40 % [AMD K6-2, Linux, debug configuration]
114 * 55 .. 77 % [UltraSparc processor, but
115 * debug-solaris-sparcv8-gcc conf.]
117 * BN_mod_exp_recp 50 .. 70 % [AMD K6-2, Linux, debug configuration]
118 * 62 .. 118 % [UltraSparc, debug-solaris-sparcv8-gcc]
120 * On the Sparc, BN_mod_exp_recp was faster than BN_mod_exp_mont
121 * at 2048 and more bits, but at 512 and 1024 bits, it was
122 * slower even than the standard algorithm!
124 * "Real" timings [linux-elf, solaris-sparcv9-gcc configurations]
125 * should be obtained when the new Montgomery reduction code
126 * has been integrated into OpenSSL.)
130 #define MONT_EXP_WORD
135 # ifdef MONT_EXP_WORD
136 if (a->top == 1 && !a->neg
137 && (BN_get_flags(p, BN_FLG_CONSTTIME) == 0)
138 && (BN_get_flags(a, BN_FLG_CONSTTIME) == 0)
139 && (BN_get_flags(m, BN_FLG_CONSTTIME) == 0)) {
140 BN_ULONG A = a->d[0];
141 ret = BN_mod_exp_mont_word(r, A, p, m, ctx, NULL);
144 ret = BN_mod_exp_mont(r, a, p, m, ctx, NULL);
149 ret = BN_mod_exp_recp(r, a, p, m, ctx);
153 ret = BN_mod_exp_simple(r, a, p, m, ctx);
161 int BN_mod_exp_recp(BIGNUM *r, const BIGNUM *a, const BIGNUM *p,
162 const BIGNUM *m, BN_CTX *ctx)
164 int i, j, bits, ret = 0, wstart, wend, window, wvalue;
167 /* Table of variables obtained from 'ctx' */
168 BIGNUM *val[TABLE_SIZE];
171 if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0
172 || BN_get_flags(a, BN_FLG_CONSTTIME) != 0
173 || BN_get_flags(m, BN_FLG_CONSTTIME) != 0) {
174 /* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
175 BNerr(BN_F_BN_MOD_EXP_RECP, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
179 bits = BN_num_bits(p);
181 /* x**0 mod 1, or x**0 mod -1 is still zero. */
182 if (BN_abs_is_word(m, 1)) {
191 BN_RECP_CTX_init(&recp);
194 aa = BN_CTX_get(ctx);
195 val[0] = BN_CTX_get(ctx);
200 /* ignore sign of 'm' */
204 if (BN_RECP_CTX_set(&recp, aa, ctx) <= 0)
207 if (BN_RECP_CTX_set(&recp, m, ctx) <= 0)
211 if (!BN_nnmod(val[0], a, m, ctx))
213 if (BN_is_zero(val[0])) {
219 window = BN_window_bits_for_exponent_size(bits);
221 if (!BN_mod_mul_reciprocal(aa, val[0], val[0], &recp, ctx))
223 j = 1 << (window - 1);
224 for (i = 1; i < j; i++) {
225 if (((val[i] = BN_CTX_get(ctx)) == NULL) ||
226 !BN_mod_mul_reciprocal(val[i], val[i - 1], aa, &recp, ctx))
231 start = 1; /* This is used to avoid multiplication etc
232 * when there is only the value '1' in the
234 wvalue = 0; /* The 'value' of the window */
235 wstart = bits - 1; /* The top bit of the window */
236 wend = 0; /* The bottom bit of the window */
242 if (BN_is_bit_set(p, wstart) == 0) {
244 if (!BN_mod_mul_reciprocal(r, r, r, &recp, ctx))
252 * We now have wstart on a 'set' bit, we now need to work out how bit
253 * a window to do. To do this we need to scan forward until the last
254 * set bit before the end of the window
259 for (i = 1; i < window; i++) {
262 if (BN_is_bit_set(p, wstart - i)) {
263 wvalue <<= (i - wend);
269 /* wend is the size of the current window */
271 /* add the 'bytes above' */
273 for (i = 0; i < j; i++) {
274 if (!BN_mod_mul_reciprocal(r, r, r, &recp, ctx))
278 /* wvalue will be an odd number < 2^window */
279 if (!BN_mod_mul_reciprocal(r, r, val[wvalue >> 1], &recp, ctx))
282 /* move the 'window' down further */
292 BN_RECP_CTX_free(&recp);
297 int BN_mod_exp_mont(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
298 const BIGNUM *m, BN_CTX *ctx, BN_MONT_CTX *in_mont)
300 int i, j, bits, ret = 0, wstart, wend, window, wvalue;
304 /* Table of variables obtained from 'ctx' */
305 BIGNUM *val[TABLE_SIZE];
306 BN_MONT_CTX *mont = NULL;
308 if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0
309 || BN_get_flags(a, BN_FLG_CONSTTIME) != 0
310 || BN_get_flags(m, BN_FLG_CONSTTIME) != 0) {
311 return BN_mod_exp_mont_consttime(rr, a, p, m, ctx, in_mont);
319 BNerr(BN_F_BN_MOD_EXP_MONT, BN_R_CALLED_WITH_EVEN_MODULUS);
322 bits = BN_num_bits(p);
324 /* x**0 mod 1, or x**0 mod -1 is still zero. */
325 if (BN_abs_is_word(m, 1)) {
337 val[0] = BN_CTX_get(ctx);
342 * If this is not done, things will break in the montgomery part
348 if ((mont = BN_MONT_CTX_new()) == NULL)
350 if (!BN_MONT_CTX_set(mont, m, ctx))
354 if (a->neg || BN_ucmp(a, m) >= 0) {
355 if (!BN_nnmod(val[0], a, m, ctx))
360 if (!bn_to_mont_fixed_top(val[0], aa, mont, ctx))
363 window = BN_window_bits_for_exponent_size(bits);
365 if (!bn_mul_mont_fixed_top(d, val[0], val[0], mont, ctx))
367 j = 1 << (window - 1);
368 for (i = 1; i < j; i++) {
369 if (((val[i] = BN_CTX_get(ctx)) == NULL) ||
370 !bn_mul_mont_fixed_top(val[i], val[i - 1], d, mont, ctx))
375 start = 1; /* This is used to avoid multiplication etc
376 * when there is only the value '1' in the
378 wvalue = 0; /* The 'value' of the window */
379 wstart = bits - 1; /* The top bit of the window */
380 wend = 0; /* The bottom bit of the window */
382 #if 1 /* by Shay Gueron's suggestion */
383 j = m->top; /* borrow j */
384 if (m->d[j - 1] & (((BN_ULONG)1) << (BN_BITS2 - 1))) {
385 if (bn_wexpand(r, j) == NULL)
387 /* 2^(top*BN_BITS2) - m */
388 r->d[0] = (0 - m->d[0]) & BN_MASK2;
389 for (i = 1; i < j; i++)
390 r->d[i] = (~m->d[i]) & BN_MASK2;
392 r->flags |= BN_FLG_FIXED_TOP;
395 if (!bn_to_mont_fixed_top(r, BN_value_one(), mont, ctx))
398 if (BN_is_bit_set(p, wstart) == 0) {
400 if (!bn_mul_mont_fixed_top(r, r, r, mont, ctx))
409 * We now have wstart on a 'set' bit, we now need to work out how bit
410 * a window to do. To do this we need to scan forward until the last
411 * set bit before the end of the window
416 for (i = 1; i < window; i++) {
419 if (BN_is_bit_set(p, wstart - i)) {
420 wvalue <<= (i - wend);
426 /* wend is the size of the current window */
428 /* add the 'bytes above' */
430 for (i = 0; i < j; i++) {
431 if (!bn_mul_mont_fixed_top(r, r, r, mont, ctx))
435 /* wvalue will be an odd number < 2^window */
436 if (!bn_mul_mont_fixed_top(r, r, val[wvalue >> 1], mont, ctx))
439 /* move the 'window' down further */
447 * Done with zero-padded intermediate BIGNUMs. Final BN_from_montgomery
448 * removes padding [if any] and makes return value suitable for public
451 #if defined(SPARC_T4_MONT)
452 if (OPENSSL_sparcv9cap_P[0] & (SPARCV9_VIS3 | SPARCV9_PREFER_FPU)) {
453 j = mont->N.top; /* borrow j */
454 val[0]->d[0] = 1; /* borrow val[0] */
455 for (i = 1; i < j; i++)
458 if (!BN_mod_mul_montgomery(rr, r, val[0], mont, ctx))
462 if (!BN_from_montgomery(rr, r, mont, ctx))
467 BN_MONT_CTX_free(mont);
473 static BN_ULONG bn_get_bits(const BIGNUM *a, int bitpos)
478 wordpos = bitpos / BN_BITS2;
480 if (wordpos >= 0 && wordpos < a->top) {
481 ret = a->d[wordpos] & BN_MASK2;
484 if (++wordpos < a->top)
485 ret |= a->d[wordpos] << (BN_BITS2 - bitpos);
489 return ret & BN_MASK2;
493 * BN_mod_exp_mont_consttime() stores the precomputed powers in a specific
494 * layout so that accessing any of these table values shows the same access
495 * pattern as far as cache lines are concerned. The following functions are
496 * used to transfer a BIGNUM from/to that table.
499 static int MOD_EXP_CTIME_COPY_TO_PREBUF(const BIGNUM *b, int top,
500 unsigned char *buf, int idx,
504 int width = 1 << window;
505 BN_ULONG *table = (BN_ULONG *)buf;
508 top = b->top; /* this works because 'buf' is explicitly
510 for (i = 0, j = idx; i < top; i++, j += width) {
517 static int MOD_EXP_CTIME_COPY_FROM_PREBUF(BIGNUM *b, int top,
518 unsigned char *buf, int idx,
522 int width = 1 << window;
524 * We declare table 'volatile' in order to discourage compiler
525 * from reordering loads from the table. Concern is that if
526 * reordered in specific manner loads might give away the
527 * information we are trying to conceal. Some would argue that
528 * compiler can reorder them anyway, but it can as well be
529 * argued that doing so would be violation of standard...
531 volatile BN_ULONG *table = (volatile BN_ULONG *)buf;
533 if (bn_wexpand(b, top) == NULL)
537 for (i = 0; i < top; i++, table += width) {
540 for (j = 0; j < width; j++) {
542 ((BN_ULONG)0 - (constant_time_eq_int(j,idx)&1));
548 int xstride = 1 << (window - 2);
549 BN_ULONG y0, y1, y2, y3;
551 i = idx >> (window - 2); /* equivalent of idx / xstride */
552 idx &= xstride - 1; /* equivalent of idx % xstride */
554 y0 = (BN_ULONG)0 - (constant_time_eq_int(i,0)&1);
555 y1 = (BN_ULONG)0 - (constant_time_eq_int(i,1)&1);
556 y2 = (BN_ULONG)0 - (constant_time_eq_int(i,2)&1);
557 y3 = (BN_ULONG)0 - (constant_time_eq_int(i,3)&1);
559 for (i = 0; i < top; i++, table += width) {
562 for (j = 0; j < xstride; j++) {
563 acc |= ( (table[j + 0 * xstride] & y0) |
564 (table[j + 1 * xstride] & y1) |
565 (table[j + 2 * xstride] & y2) |
566 (table[j + 3 * xstride] & y3) )
567 & ((BN_ULONG)0 - (constant_time_eq_int(j,idx)&1));
575 b->flags |= BN_FLG_FIXED_TOP;
580 * Given a pointer value, compute the next address that is a cache line
583 #define MOD_EXP_CTIME_ALIGN(x_) \
584 ((unsigned char*)(x_) + (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - (((size_t)(x_)) & (MOD_EXP_CTIME_MIN_CACHE_LINE_MASK))))
587 * This variant of BN_mod_exp_mont() uses fixed windows and the special
588 * precomputation memory layout to limit data-dependency to a minimum to
589 * protect secret exponents (cf. the hyper-threading timing attacks pointed
590 * out by Colin Percival,
591 * http://www.daemonology.net/hyperthreading-considered-harmful/)
593 int BN_mod_exp_mont_consttime(BIGNUM *rr, const BIGNUM *a, const BIGNUM *p,
594 const BIGNUM *m, BN_CTX *ctx,
595 BN_MONT_CTX *in_mont)
597 int i, bits, ret = 0, window, wvalue, wmask, window0;
599 BN_MONT_CTX *mont = NULL;
602 unsigned char *powerbufFree = NULL;
604 unsigned char *powerbuf = NULL;
606 #if defined(SPARC_T4_MONT)
615 BNerr(BN_F_BN_MOD_EXP_MONT_CONSTTIME, BN_R_CALLED_WITH_EVEN_MODULUS);
622 * Use all bits stored in |p|, rather than |BN_num_bits|, so we do not leak
623 * whether the top bits are zero.
625 bits = p->top * BN_BITS2;
627 /* x**0 mod 1, or x**0 mod -1 is still zero. */
628 if (BN_abs_is_word(m, 1)) {
640 * Allocate a montgomery context if it was not supplied by the caller. If
641 * this is not done, things will break in the montgomery part.
646 if ((mont = BN_MONT_CTX_new()) == NULL)
648 if (!BN_MONT_CTX_set(mont, m, ctx))
652 if (a->neg || BN_ucmp(a, m) >= 0) {
653 BIGNUM *reduced = BN_CTX_get(ctx);
655 || !BN_nnmod(reduced, a, m, ctx)) {
663 * If the size of the operands allow it, perform the optimized
664 * RSAZ exponentiation. For further information see
665 * crypto/bn/rsaz_exp.c and accompanying assembly modules.
667 if ((16 == a->top) && (16 == p->top) && (BN_num_bits(m) == 1024)
668 && rsaz_avx2_eligible()) {
669 if (NULL == bn_wexpand(rr, 16))
671 RSAZ_1024_mod_exp_avx2(rr->d, a->d, p->d, m->d, mont->RR.d,
678 } else if ((8 == a->top) && (8 == p->top) && (BN_num_bits(m) == 512)) {
679 if (NULL == bn_wexpand(rr, 8))
681 RSAZ_512_mod_exp(rr->d, a->d, p->d, m->d, mont->n0[0], mont->RR.d);
690 /* Get the window size to use with size of p. */
691 window = BN_window_bits_for_ctime_exponent_size(bits);
692 #if defined(SPARC_T4_MONT)
693 if (window >= 5 && (top & 15) == 0 && top <= 64 &&
694 (OPENSSL_sparcv9cap_P[1] & (CFR_MONTMUL | CFR_MONTSQR)) ==
695 (CFR_MONTMUL | CFR_MONTSQR) && (t4 = OPENSSL_sparcv9cap_P[0]))
699 #if defined(OPENSSL_BN_ASM_MONT5)
701 window = 5; /* ~5% improvement for RSA2048 sign, and even
703 /* reserve space for mont->N.d[] copy */
704 powerbufLen += top * sizeof(mont->N.d[0]);
710 * Allocate a buffer large enough to hold all of the pre-computed powers
711 * of am, am itself and tmp.
713 numPowers = 1 << window;
714 powerbufLen += sizeof(m->d[0]) * (top * numPowers +
716 numPowers ? (2 * top) : numPowers));
718 if (powerbufLen < 3072)
720 alloca(powerbufLen + MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH);
724 OPENSSL_malloc(powerbufLen + MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH))
728 powerbuf = MOD_EXP_CTIME_ALIGN(powerbufFree);
729 memset(powerbuf, 0, powerbufLen);
732 if (powerbufLen < 3072)
736 /* lay down tmp and am right after powers table */
737 tmp.d = (BN_ULONG *)(powerbuf + sizeof(m->d[0]) * top * numPowers);
739 tmp.top = am.top = 0;
740 tmp.dmax = am.dmax = top;
741 tmp.neg = am.neg = 0;
742 tmp.flags = am.flags = BN_FLG_STATIC_DATA;
744 /* prepare a^0 in Montgomery domain */
745 #if 1 /* by Shay Gueron's suggestion */
746 if (m->d[top - 1] & (((BN_ULONG)1) << (BN_BITS2 - 1))) {
747 /* 2^(top*BN_BITS2) - m */
748 tmp.d[0] = (0 - m->d[0]) & BN_MASK2;
749 for (i = 1; i < top; i++)
750 tmp.d[i] = (~m->d[i]) & BN_MASK2;
754 if (!bn_to_mont_fixed_top(&tmp, BN_value_one(), mont, ctx))
757 /* prepare a^1 in Montgomery domain */
758 if (!bn_to_mont_fixed_top(&am, a, mont, ctx))
761 #if defined(SPARC_T4_MONT)
763 typedef int (*bn_pwr5_mont_f) (BN_ULONG *tp, const BN_ULONG *np,
764 const BN_ULONG *n0, const void *table,
765 int power, int bits);
766 int bn_pwr5_mont_t4_8(BN_ULONG *tp, const BN_ULONG *np,
767 const BN_ULONG *n0, const void *table,
768 int power, int bits);
769 int bn_pwr5_mont_t4_16(BN_ULONG *tp, const BN_ULONG *np,
770 const BN_ULONG *n0, const void *table,
771 int power, int bits);
772 int bn_pwr5_mont_t4_24(BN_ULONG *tp, const BN_ULONG *np,
773 const BN_ULONG *n0, const void *table,
774 int power, int bits);
775 int bn_pwr5_mont_t4_32(BN_ULONG *tp, const BN_ULONG *np,
776 const BN_ULONG *n0, const void *table,
777 int power, int bits);
778 static const bn_pwr5_mont_f pwr5_funcs[4] = {
779 bn_pwr5_mont_t4_8, bn_pwr5_mont_t4_16,
780 bn_pwr5_mont_t4_24, bn_pwr5_mont_t4_32
782 bn_pwr5_mont_f pwr5_worker = pwr5_funcs[top / 16 - 1];
784 typedef int (*bn_mul_mont_f) (BN_ULONG *rp, const BN_ULONG *ap,
785 const void *bp, const BN_ULONG *np,
787 int bn_mul_mont_t4_8(BN_ULONG *rp, const BN_ULONG *ap, const void *bp,
788 const BN_ULONG *np, const BN_ULONG *n0);
789 int bn_mul_mont_t4_16(BN_ULONG *rp, const BN_ULONG *ap,
790 const void *bp, const BN_ULONG *np,
792 int bn_mul_mont_t4_24(BN_ULONG *rp, const BN_ULONG *ap,
793 const void *bp, const BN_ULONG *np,
795 int bn_mul_mont_t4_32(BN_ULONG *rp, const BN_ULONG *ap,
796 const void *bp, const BN_ULONG *np,
798 static const bn_mul_mont_f mul_funcs[4] = {
799 bn_mul_mont_t4_8, bn_mul_mont_t4_16,
800 bn_mul_mont_t4_24, bn_mul_mont_t4_32
802 bn_mul_mont_f mul_worker = mul_funcs[top / 16 - 1];
804 void bn_mul_mont_vis3(BN_ULONG *rp, const BN_ULONG *ap,
805 const void *bp, const BN_ULONG *np,
806 const BN_ULONG *n0, int num);
807 void bn_mul_mont_t4(BN_ULONG *rp, const BN_ULONG *ap,
808 const void *bp, const BN_ULONG *np,
809 const BN_ULONG *n0, int num);
810 void bn_mul_mont_gather5_t4(BN_ULONG *rp, const BN_ULONG *ap,
811 const void *table, const BN_ULONG *np,
812 const BN_ULONG *n0, int num, int power);
813 void bn_flip_n_scatter5_t4(const BN_ULONG *inp, size_t num,
814 void *table, size_t power);
815 void bn_gather5_t4(BN_ULONG *out, size_t num,
816 void *table, size_t power);
817 void bn_flip_t4(BN_ULONG *dst, BN_ULONG *src, size_t num);
819 BN_ULONG *np = mont->N.d, *n0 = mont->n0;
820 int stride = 5 * (6 - (top / 16 - 1)); /* multiple of 5, but less
824 * BN_to_montgomery can contaminate words above .top [in
825 * BN_DEBUG[_DEBUG] build]...
827 for (i = am.top; i < top; i++)
829 for (i = tmp.top; i < top; i++)
832 bn_flip_n_scatter5_t4(tmp.d, top, powerbuf, 0);
833 bn_flip_n_scatter5_t4(am.d, top, powerbuf, 1);
834 if (!(*mul_worker) (tmp.d, am.d, am.d, np, n0) &&
835 !(*mul_worker) (tmp.d, am.d, am.d, np, n0))
836 bn_mul_mont_vis3(tmp.d, am.d, am.d, np, n0, top);
837 bn_flip_n_scatter5_t4(tmp.d, top, powerbuf, 2);
839 for (i = 3; i < 32; i++) {
840 /* Calculate a^i = a^(i-1) * a */
841 if (!(*mul_worker) (tmp.d, tmp.d, am.d, np, n0) &&
842 !(*mul_worker) (tmp.d, tmp.d, am.d, np, n0))
843 bn_mul_mont_vis3(tmp.d, tmp.d, am.d, np, n0, top);
844 bn_flip_n_scatter5_t4(tmp.d, top, powerbuf, i);
847 /* switch to 64-bit domain */
848 np = alloca(top * sizeof(BN_ULONG));
850 bn_flip_t4(np, mont->N.d, top);
853 * The exponent may not have a whole number of fixed-size windows.
854 * To simplify the main loop, the initial window has between 1 and
855 * full-window-size bits such that what remains is always a whole
858 window0 = (bits - 1) % 5 + 1;
859 wmask = (1 << window0) - 1;
861 wvalue = bn_get_bits(p, bits) & wmask;
862 bn_gather5_t4(tmp.d, top, powerbuf, wvalue);
865 * Scan the exponent one window at a time starting from the most
872 wvalue = bn_get_bits(p, bits);
874 if ((*pwr5_worker) (tmp.d, np, n0, powerbuf, wvalue, stride))
876 /* retry once and fall back */
877 if ((*pwr5_worker) (tmp.d, np, n0, powerbuf, wvalue, stride))
881 wvalue >>= stride - 5;
883 bn_mul_mont_t4(tmp.d, tmp.d, tmp.d, np, n0, top);
884 bn_mul_mont_t4(tmp.d, tmp.d, tmp.d, np, n0, top);
885 bn_mul_mont_t4(tmp.d, tmp.d, tmp.d, np, n0, top);
886 bn_mul_mont_t4(tmp.d, tmp.d, tmp.d, np, n0, top);
887 bn_mul_mont_t4(tmp.d, tmp.d, tmp.d, np, n0, top);
888 bn_mul_mont_gather5_t4(tmp.d, tmp.d, powerbuf, np, n0, top,
892 bn_flip_t4(tmp.d, tmp.d, top);
894 /* back to 32-bit domain */
896 bn_correct_top(&tmp);
897 OPENSSL_cleanse(np, top * sizeof(BN_ULONG));
900 #if defined(OPENSSL_BN_ASM_MONT5)
901 if (window == 5 && top > 1) {
903 * This optimization uses ideas from http://eprint.iacr.org/2011/239,
904 * specifically optimization of cache-timing attack countermeasures
905 * and pre-computation optimization.
909 * Dedicated window==4 case improves 512-bit RSA sign by ~15%, but as
910 * 512-bit RSA is hardly relevant, we omit it to spare size...
912 void bn_mul_mont_gather5(BN_ULONG *rp, const BN_ULONG *ap,
913 const void *table, const BN_ULONG *np,
914 const BN_ULONG *n0, int num, int power);
915 void bn_scatter5(const BN_ULONG *inp, size_t num,
916 void *table, size_t power);
917 void bn_gather5(BN_ULONG *out, size_t num, void *table, size_t power);
918 void bn_power5(BN_ULONG *rp, const BN_ULONG *ap,
919 const void *table, const BN_ULONG *np,
920 const BN_ULONG *n0, int num, int power);
921 int bn_get_bits5(const BN_ULONG *ap, int off);
922 int bn_from_montgomery(BN_ULONG *rp, const BN_ULONG *ap,
923 const BN_ULONG *not_used, const BN_ULONG *np,
924 const BN_ULONG *n0, int num);
926 BN_ULONG *n0 = mont->n0, *np;
929 * BN_to_montgomery can contaminate words above .top [in
930 * BN_DEBUG[_DEBUG] build]...
932 for (i = am.top; i < top; i++)
934 for (i = tmp.top; i < top; i++)
938 * copy mont->N.d[] to improve cache locality
940 for (np = am.d + top, i = 0; i < top; i++)
941 np[i] = mont->N.d[i];
943 bn_scatter5(tmp.d, top, powerbuf, 0);
944 bn_scatter5(am.d, am.top, powerbuf, 1);
945 bn_mul_mont(tmp.d, am.d, am.d, np, n0, top);
946 bn_scatter5(tmp.d, top, powerbuf, 2);
949 for (i = 3; i < 32; i++) {
950 /* Calculate a^i = a^(i-1) * a */
951 bn_mul_mont_gather5(tmp.d, am.d, powerbuf, np, n0, top, i - 1);
952 bn_scatter5(tmp.d, top, powerbuf, i);
955 /* same as above, but uses squaring for 1/2 of operations */
956 for (i = 4; i < 32; i *= 2) {
957 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
958 bn_scatter5(tmp.d, top, powerbuf, i);
960 for (i = 3; i < 8; i += 2) {
962 bn_mul_mont_gather5(tmp.d, am.d, powerbuf, np, n0, top, i - 1);
963 bn_scatter5(tmp.d, top, powerbuf, i);
964 for (j = 2 * i; j < 32; j *= 2) {
965 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
966 bn_scatter5(tmp.d, top, powerbuf, j);
969 for (; i < 16; i += 2) {
970 bn_mul_mont_gather5(tmp.d, am.d, powerbuf, np, n0, top, i - 1);
971 bn_scatter5(tmp.d, top, powerbuf, i);
972 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
973 bn_scatter5(tmp.d, top, powerbuf, 2 * i);
975 for (; i < 32; i += 2) {
976 bn_mul_mont_gather5(tmp.d, am.d, powerbuf, np, n0, top, i - 1);
977 bn_scatter5(tmp.d, top, powerbuf, i);
981 * The exponent may not have a whole number of fixed-size windows.
982 * To simplify the main loop, the initial window has between 1 and
983 * full-window-size bits such that what remains is always a whole
986 window0 = (bits - 1) % 5 + 1;
987 wmask = (1 << window0) - 1;
989 wvalue = bn_get_bits(p, bits) & wmask;
990 bn_gather5(tmp.d, top, powerbuf, wvalue);
993 * Scan the exponent one window at a time starting from the most
998 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
999 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
1000 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
1001 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
1002 bn_mul_mont(tmp.d, tmp.d, tmp.d, np, n0, top);
1003 bn_mul_mont_gather5(tmp.d, tmp.d, powerbuf, np, n0, top,
1004 bn_get_bits5(p->d, bits -= 5));
1008 bn_power5(tmp.d, tmp.d, powerbuf, np, n0, top,
1009 bn_get_bits5(p->d, bits -= 5));
1013 ret = bn_from_montgomery(tmp.d, tmp.d, NULL, np, n0, top);
1015 bn_correct_top(&tmp);
1017 if (!BN_copy(rr, &tmp))
1019 goto err; /* non-zero ret means it's not error */
1024 if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&tmp, top, powerbuf, 0, window))
1026 if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&am, top, powerbuf, 1, window))
1030 * If the window size is greater than 1, then calculate
1031 * val[i=2..2^winsize-1]. Powers are computed as a*a^(i-1) (even
1032 * powers could instead be computed as (a^(i/2))^2 to use the slight
1033 * performance advantage of sqr over mul).
1036 if (!bn_mul_mont_fixed_top(&tmp, &am, &am, mont, ctx))
1038 if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&tmp, top, powerbuf, 2,
1041 for (i = 3; i < numPowers; i++) {
1042 /* Calculate a^i = a^(i-1) * a */
1043 if (!bn_mul_mont_fixed_top(&tmp, &am, &tmp, mont, ctx))
1045 if (!MOD_EXP_CTIME_COPY_TO_PREBUF(&tmp, top, powerbuf, i,
1052 * The exponent may not have a whole number of fixed-size windows.
1053 * To simplify the main loop, the initial window has between 1 and
1054 * full-window-size bits such that what remains is always a whole
1057 window0 = (bits - 1) % window + 1;
1058 wmask = (1 << window0) - 1;
1060 wvalue = bn_get_bits(p, bits) & wmask;
1061 if (!MOD_EXP_CTIME_COPY_FROM_PREBUF(&tmp, top, powerbuf, wvalue,
1065 wmask = (1 << window) - 1;
1067 * Scan the exponent one window at a time starting from the most
1072 /* Square the result window-size times */
1073 for (i = 0; i < window; i++)
1074 if (!bn_mul_mont_fixed_top(&tmp, &tmp, &tmp, mont, ctx))
1078 * Get a window's worth of bits from the exponent
1079 * This avoids calling BN_is_bit_set for each bit, which
1080 * is not only slower but also makes each bit vulnerable to
1081 * EM (and likely other) side-channel attacks like One&Done
1082 * (for details see "One&Done: A Single-Decryption EM-Based
1083 * Attack on OpenSSL's Constant-Time Blinded RSA" by M. Alam,
1084 * H. Khan, M. Dey, N. Sinha, R. Callan, A. Zajic, and
1085 * M. Prvulovic, in USENIX Security'18)
1088 wvalue = bn_get_bits(p, bits) & wmask;
1090 * Fetch the appropriate pre-computed value from the pre-buf
1092 if (!MOD_EXP_CTIME_COPY_FROM_PREBUF(&am, top, powerbuf, wvalue,
1096 /* Multiply the result into the intermediate result */
1097 if (!bn_mul_mont_fixed_top(&tmp, &tmp, &am, mont, ctx))
1103 * Done with zero-padded intermediate BIGNUMs. Final BN_from_montgomery
1104 * removes padding [if any] and makes return value suitable for public
1107 #if defined(SPARC_T4_MONT)
1108 if (OPENSSL_sparcv9cap_P[0] & (SPARCV9_VIS3 | SPARCV9_PREFER_FPU)) {
1109 am.d[0] = 1; /* borrow am */
1110 for (i = 1; i < top; i++)
1112 if (!BN_mod_mul_montgomery(rr, &tmp, &am, mont, ctx))
1116 if (!BN_from_montgomery(rr, &tmp, mont, ctx))
1120 if (in_mont == NULL)
1121 BN_MONT_CTX_free(mont);
1122 if (powerbuf != NULL) {
1123 OPENSSL_cleanse(powerbuf, powerbufLen);
1124 OPENSSL_free(powerbufFree);
1130 int BN_mod_exp_mont_word(BIGNUM *rr, BN_ULONG a, const BIGNUM *p,
1131 const BIGNUM *m, BN_CTX *ctx, BN_MONT_CTX *in_mont)
1133 BN_MONT_CTX *mont = NULL;
1134 int b, bits, ret = 0;
1139 #define BN_MOD_MUL_WORD(r, w, m) \
1140 (BN_mul_word(r, (w)) && \
1141 (/* BN_ucmp(r, (m)) < 0 ? 1 :*/ \
1142 (BN_mod(t, r, m, ctx) && (swap_tmp = r, r = t, t = swap_tmp, 1))))
1144 * BN_MOD_MUL_WORD is only used with 'w' large, so the BN_ucmp test is
1145 * probably more overhead than always using BN_mod (which uses BN_copy if
1146 * a similar test returns true).
1149 * We can use BN_mod and do not need BN_nnmod because our accumulator is
1150 * never negative (the result of BN_mod does not depend on the sign of
1153 #define BN_TO_MONTGOMERY_WORD(r, w, mont) \
1154 (BN_set_word(r, (w)) && BN_to_montgomery(r, r, (mont), ctx))
1156 if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0
1157 || BN_get_flags(m, BN_FLG_CONSTTIME) != 0) {
1158 /* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
1159 BNerr(BN_F_BN_MOD_EXP_MONT_WORD, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
1166 if (!BN_is_odd(m)) {
1167 BNerr(BN_F_BN_MOD_EXP_MONT_WORD, BN_R_CALLED_WITH_EVEN_MODULUS);
1171 a %= m->d[0]; /* make sure that 'a' is reduced */
1173 bits = BN_num_bits(p);
1175 /* x**0 mod 1, or x**0 mod -1 is still zero. */
1176 if (BN_abs_is_word(m, 1)) {
1191 r = BN_CTX_get(ctx);
1192 t = BN_CTX_get(ctx);
1196 if (in_mont != NULL)
1199 if ((mont = BN_MONT_CTX_new()) == NULL)
1201 if (!BN_MONT_CTX_set(mont, m, ctx))
1205 r_is_one = 1; /* except for Montgomery factor */
1209 /* The result is accumulated in the product r*w. */
1210 w = a; /* bit 'bits-1' of 'p' is always set */
1211 for (b = bits - 2; b >= 0; b--) {
1212 /* First, square r*w. */
1214 if ((next_w / w) != w) { /* overflow */
1216 if (!BN_TO_MONTGOMERY_WORD(r, w, mont))
1220 if (!BN_MOD_MUL_WORD(r, w, m))
1227 if (!BN_mod_mul_montgomery(r, r, r, mont, ctx))
1231 /* Second, multiply r*w by 'a' if exponent bit is set. */
1232 if (BN_is_bit_set(p, b)) {
1234 if ((next_w / a) != w) { /* overflow */
1236 if (!BN_TO_MONTGOMERY_WORD(r, w, mont))
1240 if (!BN_MOD_MUL_WORD(r, w, m))
1249 /* Finally, set r:=r*w. */
1252 if (!BN_TO_MONTGOMERY_WORD(r, w, mont))
1256 if (!BN_MOD_MUL_WORD(r, w, m))
1261 if (r_is_one) { /* can happen only if a == 1 */
1265 if (!BN_from_montgomery(rr, r, mont, ctx))
1270 if (in_mont == NULL)
1271 BN_MONT_CTX_free(mont);
1277 /* The old fallback, simple version :-) */
1278 int BN_mod_exp_simple(BIGNUM *r, const BIGNUM *a, const BIGNUM *p,
1279 const BIGNUM *m, BN_CTX *ctx)
1281 int i, j, bits, ret = 0, wstart, wend, window, wvalue;
1284 /* Table of variables obtained from 'ctx' */
1285 BIGNUM *val[TABLE_SIZE];
1287 if (BN_get_flags(p, BN_FLG_CONSTTIME) != 0
1288 || BN_get_flags(a, BN_FLG_CONSTTIME) != 0
1289 || BN_get_flags(m, BN_FLG_CONSTTIME) != 0) {
1290 /* BN_FLG_CONSTTIME only supported by BN_mod_exp_mont() */
1291 BNerr(BN_F_BN_MOD_EXP_SIMPLE, ERR_R_SHOULD_NOT_HAVE_BEEN_CALLED);
1295 bits = BN_num_bits(p);
1297 /* x**0 mod 1, or x**0 mod -1 is still zero. */
1298 if (BN_abs_is_word(m, 1)) {
1308 d = BN_CTX_get(ctx);
1309 val[0] = BN_CTX_get(ctx);
1313 if (!BN_nnmod(val[0], a, m, ctx))
1315 if (BN_is_zero(val[0])) {
1321 window = BN_window_bits_for_exponent_size(bits);
1323 if (!BN_mod_mul(d, val[0], val[0], m, ctx))
1325 j = 1 << (window - 1);
1326 for (i = 1; i < j; i++) {
1327 if (((val[i] = BN_CTX_get(ctx)) == NULL) ||
1328 !BN_mod_mul(val[i], val[i - 1], d, m, ctx))
1333 start = 1; /* This is used to avoid multiplication etc
1334 * when there is only the value '1' in the
1336 wvalue = 0; /* The 'value' of the window */
1337 wstart = bits - 1; /* The top bit of the window */
1338 wend = 0; /* The bottom bit of the window */
1344 if (BN_is_bit_set(p, wstart) == 0) {
1346 if (!BN_mod_mul(r, r, r, m, ctx))
1354 * We now have wstart on a 'set' bit, we now need to work out how bit
1355 * a window to do. To do this we need to scan forward until the last
1356 * set bit before the end of the window
1361 for (i = 1; i < window; i++) {
1364 if (BN_is_bit_set(p, wstart - i)) {
1365 wvalue <<= (i - wend);
1371 /* wend is the size of the current window */
1373 /* add the 'bytes above' */
1375 for (i = 0; i < j; i++) {
1376 if (!BN_mod_mul(r, r, r, m, ctx))
1380 /* wvalue will be an odd number < 2^window */
1381 if (!BN_mod_mul(r, r, val[wvalue >> 1], m, ctx))
1384 /* move the 'window' down further */