VirtualBox

source: vbox/trunk/src/libs/openssl-3.0.7/crypto/bn/rsaz_exp_x2.c@ 97371

最後變更 在這個檔案從97371是 94404,由 vboxsync 提交於 3 年 前

libs/openssl: Update to 3.0.2 and switch to it, bugref:10128
檔案大小: 18.7 KB
行 
1/*
2 * Copyright 2020-2021 The OpenSSL Project Authors. All Rights Reserved.
3 * Copyright (c) 2020, Intel Corporation. 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 * Originally written by Ilya Albrekht, Sergey Kirillov and Andrey Matyukov
12 * Intel Corporation
13 *
14 */
15
16#include <openssl/opensslconf.h>
17#include <openssl/crypto.h>
18#include "rsaz_exp.h"
19
20#ifndef RSAZ_ENABLED
21NON_EMPTY_TRANSLATION_UNIT
22#else
23# include <assert.h>
24# include <string.h>
25
26# if defined(__GNUC__)
27# define ALIGN64 __attribute__((aligned(64)))
28# elif defined(_MSC_VER)
29# define ALIGN64 __declspec(align(64))
30# else
31# define ALIGN64
32# endif
33
34# define ALIGN_OF(ptr, boundary) \
35 ((unsigned char *)(ptr) + (boundary - (((size_t)(ptr)) & (boundary - 1))))
36
37/* Internal radix */
38# define DIGIT_SIZE (52)
39/* 52-bit mask */
40# define DIGIT_MASK ((uint64_t)0xFFFFFFFFFFFFF)
41
42# define BITS2WORD8_SIZE(x) (((x) + 7) >> 3)
43# define BITS2WORD64_SIZE(x) (((x) + 63) >> 6)
44
45static ossl_inline uint64_t get_digit52(const uint8_t *in, int in_len);
46static ossl_inline void put_digit52(uint8_t *out, int out_len, uint64_t digit);
47static void to_words52(BN_ULONG *out, int out_len, const BN_ULONG *in,
48 int in_bitsize);
49static void from_words52(BN_ULONG *bn_out, int out_bitsize, const BN_ULONG *in);
50static ossl_inline void set_bit(BN_ULONG *a, int idx);
51
52/* Number of |digit_size|-bit digits in |bitsize|-bit value */
53static ossl_inline int number_of_digits(int bitsize, int digit_size)
54{
55 return (bitsize + digit_size - 1) / digit_size;
56}
57
58typedef void (*AMM52)(BN_ULONG *res, const BN_ULONG *base,
59 const BN_ULONG *exp, const BN_ULONG *m, BN_ULONG k0);
60typedef void (*EXP52_x2)(BN_ULONG *res, const BN_ULONG *base,
61 const BN_ULONG *exp[2], const BN_ULONG *m,
62 const BN_ULONG *rr, const BN_ULONG k0[2]);
63
64/*
65 * For details of the methods declared below please refer to
66 * crypto/bn/asm/rsaz-avx512.pl
67 *
68 * Naming notes:
69 * amm = Almost Montgomery Multiplication
70 * ams = Almost Montgomery Squaring
71 * 52x20 - data represented as array of 20 digits in 52-bit radix
72 * _x1_/_x2_ - 1 or 2 independent inputs/outputs
73 * _256 suffix - uses 256-bit (AVX512VL) registers
74 */
75
76/*AMM = Almost Montgomery Multiplication. */
77void ossl_rsaz_amm52x20_x1_256(BN_ULONG *res, const BN_ULONG *base,
78 const BN_ULONG *exp, const BN_ULONG *m,
79 BN_ULONG k0);
80static void RSAZ_exp52x20_x2_256(BN_ULONG *res, const BN_ULONG *base,
81 const BN_ULONG *exp[2], const BN_ULONG *m,
82 const BN_ULONG *rr, const BN_ULONG k0[2]);
83void ossl_rsaz_amm52x20_x2_256(BN_ULONG *out, const BN_ULONG *a,
84 const BN_ULONG *b, const BN_ULONG *m,
85 const BN_ULONG k0[2]);
86void ossl_extract_multiplier_2x20_win5(BN_ULONG *red_Y,
87 const BN_ULONG *red_table,
88 int red_table_idx, int tbl_idx);
89
90/*
91 * Dual Montgomery modular exponentiation using prime moduli of the
92 * same bit size, optimized with AVX512 ISA.
93 *
94 * Input and output parameters for each exponentiation are independent and
95 * denoted here by index |i|, i = 1..2.
96 *
97 * Input and output are all in regular 2^64 radix.
98 *
99 * Each moduli shall be |factor_size| bit size.
100 *
101 * NOTE: currently only 2x1024 case is supported.
102 *
103 * [out] res|i| - result of modular exponentiation: array of qword values
104 * in regular (2^64) radix. Size of array shall be enough
105 * to hold |factor_size| bits.
106 * [in] base|i| - base
107 * [in] exp|i| - exponent
108 * [in] m|i| - moduli
109 * [in] rr|i| - Montgomery parameter RR = R^2 mod m|i|
110 * [in] k0_|i| - Montgomery parameter k0 = -1/m|i| mod 2^64
111 * [in] factor_size - moduli bit size
112 *
113 * \return 0 in case of failure,
114 * 1 in case of success.
115 */
116int ossl_rsaz_mod_exp_avx512_x2(BN_ULONG *res1,
117 const BN_ULONG *base1,
118 const BN_ULONG *exp1,
119 const BN_ULONG *m1,
120 const BN_ULONG *rr1,
121 BN_ULONG k0_1,
122 BN_ULONG *res2,
123 const BN_ULONG *base2,
124 const BN_ULONG *exp2,
125 const BN_ULONG *m2,
126 const BN_ULONG *rr2,
127 BN_ULONG k0_2,
128 int factor_size)
129{
130 int ret = 0;
131
132 /*
133 * Number of word-size (BN_ULONG) digits to store exponent in redundant
134 * representation.
135 */
136 int exp_digits = number_of_digits(factor_size + 2, DIGIT_SIZE);
137 int coeff_pow = 4 * (DIGIT_SIZE * exp_digits - factor_size);
138 BN_ULONG *base1_red, *m1_red, *rr1_red;
139 BN_ULONG *base2_red, *m2_red, *rr2_red;
140 BN_ULONG *coeff_red;
141 BN_ULONG *storage = NULL;
142 BN_ULONG *storage_aligned = NULL;
143 BN_ULONG storage_len_bytes = 7 * exp_digits * sizeof(BN_ULONG);
144
145 /* AMM = Almost Montgomery Multiplication */
146 AMM52 amm = NULL;
147 /* Dual (2-exps in parallel) exponentiation */
148 EXP52_x2 exp_x2 = NULL;
149
150 const BN_ULONG *exp[2] = {0};
151 BN_ULONG k0[2] = {0};
152
153 /* Only 1024-bit factor size is supported now */
154 switch (factor_size) {
155 case 1024:
156 amm = ossl_rsaz_amm52x20_x1_256;
157 exp_x2 = RSAZ_exp52x20_x2_256;
158 break;
159 default:
160 goto err;
161 }
162
163 storage = (BN_ULONG *)OPENSSL_malloc(storage_len_bytes + 64);
164 if (storage == NULL)
165 goto err;
166 storage_aligned = (BN_ULONG *)ALIGN_OF(storage, 64);
167
168 /* Memory layout for red(undant) representations */
169 base1_red = storage_aligned;
170 base2_red = storage_aligned + 1 * exp_digits;
171 m1_red = storage_aligned + 2 * exp_digits;
172 m2_red = storage_aligned + 3 * exp_digits;
173 rr1_red = storage_aligned + 4 * exp_digits;
174 rr2_red = storage_aligned + 5 * exp_digits;
175 coeff_red = storage_aligned + 6 * exp_digits;
176
177 /* Convert base_i, m_i, rr_i, from regular to 52-bit radix */
178 to_words52(base1_red, exp_digits, base1, factor_size);
179 to_words52(base2_red, exp_digits, base2, factor_size);
180 to_words52(m1_red, exp_digits, m1, factor_size);
181 to_words52(m2_red, exp_digits, m2, factor_size);
182 to_words52(rr1_red, exp_digits, rr1, factor_size);
183 to_words52(rr2_red, exp_digits, rr2, factor_size);
184
185 /*
186 * Compute target domain Montgomery converters RR' for each modulus
187 * based on precomputed original domain's RR.
188 *
189 * RR -> RR' transformation steps:
190 * (1) coeff = 2^k
191 * (2) t = AMM(RR,RR) = RR^2 / R' mod m
192 * (3) RR' = AMM(t, coeff) = RR^2 * 2^k / R'^2 mod m
193 * where
194 * k = 4 * (52 * digits52 - modlen)
195 * R = 2^(64 * ceil(modlen/64)) mod m
196 * RR = R^2 mod M
197 * R' = 2^(52 * ceil(modlen/52)) mod m
198 *
199 * modlen = 1024: k = 64, RR = 2^2048 mod m, RR' = 2^2080 mod m
200 */
201 memset(coeff_red, 0, exp_digits * sizeof(BN_ULONG));
202 /* (1) in reduced domain representation */
203 set_bit(coeff_red, 64 * (int)(coeff_pow / 52) + coeff_pow % 52);
204
205 amm(rr1_red, rr1_red, rr1_red, m1_red, k0_1); /* (2) for m1 */
206 amm(rr1_red, rr1_red, coeff_red, m1_red, k0_1); /* (3) for m1 */
207
208 amm(rr2_red, rr2_red, rr2_red, m2_red, k0_2); /* (2) for m2 */
209 amm(rr2_red, rr2_red, coeff_red, m2_red, k0_2); /* (3) for m2 */
210
211 exp[0] = exp1;
212 exp[1] = exp2;
213
214 k0[0] = k0_1;
215 k0[1] = k0_2;
216
217 exp_x2(rr1_red, base1_red, exp, m1_red, rr1_red, k0);
218
219 /* Convert rr_i back to regular radix */
220 from_words52(res1, factor_size, rr1_red);
221 from_words52(res2, factor_size, rr2_red);
222
223 ret = 1;
224err:
225 if (storage != NULL) {
226 OPENSSL_cleanse(storage, storage_len_bytes);
227 OPENSSL_free(storage);
228 }
229 return ret;
230}
231
232/*
233 * Dual 1024-bit w-ary modular exponentiation using prime moduli of the same
234 * bit size using Almost Montgomery Multiplication, optimized with AVX512_IFMA
235 * ISA.
236 *
237 * The parameter w (window size) = 5.
238 *
239 * [out] res - result of modular exponentiation: 2x20 qword
240 * values in 2^52 radix.
241 * [in] base - base (2x20 qword values in 2^52 radix)
242 * [in] exp - array of 2 pointers to 16 qword values in 2^64 radix.
243 * Exponent is not converted to redundant representation.
244 * [in] m - moduli (2x20 qword values in 2^52 radix)
245 * [in] rr - Montgomery parameter for 2 moduli: RR = 2^2080 mod m.
246 * (2x20 qword values in 2^52 radix)
247 * [in] k0 - Montgomery parameter for 2 moduli: k0 = -1/m mod 2^64
248 *
249 * \return (void).
250 */
251static void RSAZ_exp52x20_x2_256(BN_ULONG *out, /* [2][20] */
252 const BN_ULONG *base, /* [2][20] */
253 const BN_ULONG *exp[2], /* 2x16 */
254 const BN_ULONG *m, /* [2][20] */
255 const BN_ULONG *rr, /* [2][20] */
256 const BN_ULONG k0[2])
257{
258# define BITSIZE_MODULUS (1024)
259# define EXP_WIN_SIZE (5)
260# define EXP_WIN_MASK ((1U << EXP_WIN_SIZE) - 1)
261/*
262 * Number of digits (64-bit words) in redundant representation to handle
263 * modulus bits
264 */
265# define RED_DIGITS (20)
266# define EXP_DIGITS (16)
267# define DAMM ossl_rsaz_amm52x20_x2_256
268/*
269 * Squaring is done using multiplication now. That can be a subject of
270 * optimization in future.
271 */
272# define DAMS(r,a,m,k0) \
273 ossl_rsaz_amm52x20_x2_256((r),(a),(a),(m),(k0))
274
275 /* Allocate stack for red(undant) result Y and multiplier X */
276 ALIGN64 BN_ULONG red_Y[2][RED_DIGITS];
277 ALIGN64 BN_ULONG red_X[2][RED_DIGITS];
278
279 /* Allocate expanded exponent */
280 ALIGN64 BN_ULONG expz[2][EXP_DIGITS + 1];
281
282 /* Pre-computed table of base powers */
283 ALIGN64 BN_ULONG red_table[1U << EXP_WIN_SIZE][2][RED_DIGITS];
284
285 int idx;
286
287 memset(red_Y, 0, sizeof(red_Y));
288 memset(red_table, 0, sizeof(red_table));
289 memset(red_X, 0, sizeof(red_X));
290
291 /*
292 * Compute table of powers base^i, i = 0, ..., (2^EXP_WIN_SIZE) - 1
293 * table[0] = mont(x^0) = mont(1)
294 * table[1] = mont(x^1) = mont(x)
295 */
296 red_X[0][0] = 1;
297 red_X[1][0] = 1;
298 DAMM(red_table[0][0], (const BN_ULONG*)red_X, rr, m, k0);
299 DAMM(red_table[1][0], base, rr, m, k0);
300
301 for (idx = 1; idx < (int)((1U << EXP_WIN_SIZE) / 2); idx++) {
302 DAMS(red_table[2 * idx + 0][0], red_table[1 * idx][0], m, k0);
303 DAMM(red_table[2 * idx + 1][0], red_table[2 * idx][0], red_table[1][0], m, k0);
304 }
305
306 /* Copy and expand exponents */
307 memcpy(expz[0], exp[0], EXP_DIGITS * sizeof(BN_ULONG));
308 expz[0][EXP_DIGITS] = 0;
309 memcpy(expz[1], exp[1], EXP_DIGITS * sizeof(BN_ULONG));
310 expz[1][EXP_DIGITS] = 0;
311
312 /* Exponentiation */
313 {
314 const int rem = BITSIZE_MODULUS % EXP_WIN_SIZE;
315 BN_ULONG table_idx_mask = EXP_WIN_MASK;
316
317 int exp_bit_no = BITSIZE_MODULUS - rem;
318 int exp_chunk_no = exp_bit_no / 64;
319 int exp_chunk_shift = exp_bit_no % 64;
320
321 BN_ULONG red_table_idx_0, red_table_idx_1;
322
323 /*
324 * If rem == 0, then
325 * exp_bit_no = modulus_bitsize - exp_win_size
326 * However, this isn't possible because rem is { 1024, 1536, 2048 } % 5
327 * which is { 4, 1, 3 } respectively.
328 *
329 * If this assertion ever fails the fix above is easy.
330 */
331 OPENSSL_assert(rem != 0);
332
333 /* Process 1-st exp window - just init result */
334 red_table_idx_0 = expz[0][exp_chunk_no];
335 red_table_idx_1 = expz[1][exp_chunk_no];
336 /*
337 * The function operates with fixed moduli sizes divisible by 64,
338 * thus table index here is always in supported range [0, EXP_WIN_SIZE).
339 */
340 red_table_idx_0 >>= exp_chunk_shift;
341 red_table_idx_1 >>= exp_chunk_shift;
342
343 ossl_extract_multiplier_2x20_win5(red_Y[0], (const BN_ULONG*)red_table,
344 (int)red_table_idx_0, 0);
345 ossl_extract_multiplier_2x20_win5(red_Y[1], (const BN_ULONG*)red_table,
346 (int)red_table_idx_1, 1);
347
348 /* Process other exp windows */
349 for (exp_bit_no -= EXP_WIN_SIZE; exp_bit_no >= 0; exp_bit_no -= EXP_WIN_SIZE) {
350 /* Extract pre-computed multiplier from the table */
351 {
352 BN_ULONG T;
353
354 exp_chunk_no = exp_bit_no / 64;
355 exp_chunk_shift = exp_bit_no % 64;
356 {
357 red_table_idx_0 = expz[0][exp_chunk_no];
358 T = expz[0][exp_chunk_no + 1];
359
360 red_table_idx_0 >>= exp_chunk_shift;
361 /*
362 * Get additional bits from then next quadword
363 * when 64-bit boundaries are crossed.
364 */
365 if (exp_chunk_shift > 64 - EXP_WIN_SIZE) {
366 T <<= (64 - exp_chunk_shift);
367 red_table_idx_0 ^= T;
368 }
369 red_table_idx_0 &= table_idx_mask;
370
371 ossl_extract_multiplier_2x20_win5(red_X[0],
372 (const BN_ULONG*)red_table,
373 (int)red_table_idx_0, 0);
374 }
375 {
376 red_table_idx_1 = expz[1][exp_chunk_no];
377 T = expz[1][exp_chunk_no + 1];
378
379 red_table_idx_1 >>= exp_chunk_shift;
380 /*
381 * Get additional bits from then next quadword
382 * when 64-bit boundaries are crossed.
383 */
384 if (exp_chunk_shift > 64 - EXP_WIN_SIZE) {
385 T <<= (64 - exp_chunk_shift);
386 red_table_idx_1 ^= T;
387 }
388 red_table_idx_1 &= table_idx_mask;
389
390 ossl_extract_multiplier_2x20_win5(red_X[1],
391 (const BN_ULONG*)red_table,
392 (int)red_table_idx_1, 1);
393 }
394 }
395
396 /* Series of squaring */
397 DAMS((BN_ULONG*)red_Y, (const BN_ULONG*)red_Y, m, k0);
398 DAMS((BN_ULONG*)red_Y, (const BN_ULONG*)red_Y, m, k0);
399 DAMS((BN_ULONG*)red_Y, (const BN_ULONG*)red_Y, m, k0);
400 DAMS((BN_ULONG*)red_Y, (const BN_ULONG*)red_Y, m, k0);
401 DAMS((BN_ULONG*)red_Y, (const BN_ULONG*)red_Y, m, k0);
402
403 DAMM((BN_ULONG*)red_Y, (const BN_ULONG*)red_Y, (const BN_ULONG*)red_X, m, k0);
404 }
405 }
406
407 /*
408 *
409 * NB: After the last AMM of exponentiation in Montgomery domain, the result
410 * may be 1025-bit, but the conversion out of Montgomery domain performs an
411 * AMM(x,1) which guarantees that the final result is less than |m|, so no
412 * conditional subtraction is needed here. See "Efficient Software
413 * Implementations of Modular Exponentiation" (by Shay Gueron) paper for details.
414 */
415
416 /* Convert result back in regular 2^52 domain */
417 memset(red_X, 0, sizeof(red_X));
418 red_X[0][0] = 1;
419 red_X[1][0] = 1;
420 DAMM(out, (const BN_ULONG*)red_Y, (const BN_ULONG*)red_X, m, k0);
421
422 /* Clear exponents */
423 OPENSSL_cleanse(expz, sizeof(expz));
424 OPENSSL_cleanse(red_Y, sizeof(red_Y));
425
426# undef DAMS
427# undef DAMM
428# undef EXP_DIGITS
429# undef RED_DIGITS
430# undef EXP_WIN_MASK
431# undef EXP_WIN_SIZE
432# undef BITSIZE_MODULUS
433}
434
435static ossl_inline uint64_t get_digit52(const uint8_t *in, int in_len)
436{
437 uint64_t digit = 0;
438
439 assert(in != NULL);
440
441 for (; in_len > 0; in_len--) {
442 digit <<= 8;
443 digit += (uint64_t)(in[in_len - 1]);
444 }
445 return digit;
446}
447
448/*
449 * Convert array of words in regular (base=2^64) representation to array of
450 * words in redundant (base=2^52) one.
451 */
452static void to_words52(BN_ULONG *out, int out_len,
453 const BN_ULONG *in, int in_bitsize)
454{
455 uint8_t *in_str = NULL;
456
457 assert(out != NULL);
458 assert(in != NULL);
459 /* Check destination buffer capacity */
460 assert(out_len >= number_of_digits(in_bitsize, DIGIT_SIZE));
461
462 in_str = (uint8_t *)in;
463
464 for (; in_bitsize >= (2 * DIGIT_SIZE); in_bitsize -= (2 * DIGIT_SIZE), out += 2) {
465 out[0] = (*(uint64_t *)in_str) & DIGIT_MASK;
466 in_str += 6;
467 out[1] = ((*(uint64_t *)in_str) >> 4) & DIGIT_MASK;
468 in_str += 7;
469 out_len -= 2;
470 }
471
472 if (in_bitsize > DIGIT_SIZE) {
473 uint64_t digit = get_digit52(in_str, 7);
474
475 out[0] = digit & DIGIT_MASK;
476 in_str += 6;
477 in_bitsize -= DIGIT_SIZE;
478 digit = get_digit52(in_str, BITS2WORD8_SIZE(in_bitsize));
479 out[1] = digit >> 4;
480 out += 2;
481 out_len -= 2;
482 } else if (in_bitsize > 0) {
483 out[0] = get_digit52(in_str, BITS2WORD8_SIZE(in_bitsize));
484 out++;
485 out_len--;
486 }
487
488 while (out_len > 0) {
489 *out = 0;
490 out_len--;
491 out++;
492 }
493}
494
495static ossl_inline void put_digit52(uint8_t *pStr, int strLen, uint64_t digit)
496{
497 assert(pStr != NULL);
498
499 for (; strLen > 0; strLen--) {
500 *pStr++ = (uint8_t)(digit & 0xFF);
501 digit >>= 8;
502 }
503}
504
505/*
506 * Convert array of words in redundant (base=2^52) representation to array of
507 * words in regular (base=2^64) one.
508 */
509static void from_words52(BN_ULONG *out, int out_bitsize, const BN_ULONG *in)
510{
511 int i;
512 int out_len = BITS2WORD64_SIZE(out_bitsize);
513
514 assert(out != NULL);
515 assert(in != NULL);
516
517 for (i = 0; i < out_len; i++)
518 out[i] = 0;
519
520 {
521 uint8_t *out_str = (uint8_t *)out;
522
523 for (; out_bitsize >= (2 * DIGIT_SIZE); out_bitsize -= (2 * DIGIT_SIZE), in += 2) {
524 (*(uint64_t *)out_str) = in[0];
525 out_str += 6;
526 (*(uint64_t *)out_str) ^= in[1] << 4;
527 out_str += 7;
528 }
529
530 if (out_bitsize > DIGIT_SIZE) {
531 put_digit52(out_str, 7, in[0]);
532 out_str += 6;
533 out_bitsize -= DIGIT_SIZE;
534 put_digit52(out_str, BITS2WORD8_SIZE(out_bitsize),
535 (in[1] << 4 | in[0] >> 48));
536 } else if (out_bitsize) {
537 put_digit52(out_str, BITS2WORD8_SIZE(out_bitsize), in[0]);
538 }
539 }
540}
541
542/*
543 * Set bit at index |idx| in the words array |a|.
544 * It does not do any boundaries checks, make sure the index is valid before
545 * calling the function.
546 */
547static ossl_inline void set_bit(BN_ULONG *a, int idx)
548{
549 assert(a != NULL);
550
551 {
552 int i, j;
553
554 i = idx / BN_BITS2;
555 j = idx % BN_BITS2;
556 a[i] |= (((BN_ULONG)1) << j);
557 }
558}
559
560#endif
注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

© 2025 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette