xref: /freebsd/contrib/bearssl/src/symcipher/poly1305_ctmulq.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
1 /*
2  * Copyright (c) 2017 Thomas Pornin <pornin@bolet.org>
3  *
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:
11  *
12  * The above copyright notice and this permission notice shall be
13  * included in all copies or substantial portions of the Software.
14  *
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
22  * SOFTWARE.
23  */
24 
25 #include "inner.h"
26 
27 #if BR_INT128 || BR_UMUL128
28 
29 #if BR_INT128
30 
31 #define MUL128(hi, lo, x, y)   do { \
32 		unsigned __int128 mul128tmp; \
33 		mul128tmp = (unsigned __int128)(x) * (unsigned __int128)(y); \
34 		(hi) = (uint64_t)(mul128tmp >> 64); \
35 		(lo) = (uint64_t)mul128tmp; \
36 	} while (0)
37 
38 #elif BR_UMUL128
39 
40 #include <intrin.h>
41 
42 #define MUL128(hi, lo, x, y)   do { \
43 		(lo) = _umul128((x), (y), &(hi)); \
44 	} while (0)
45 
46 #endif
47 
48 #define MASK42   ((uint64_t)0x000003FFFFFFFFFF)
49 #define MASK44   ((uint64_t)0x00000FFFFFFFFFFF)
50 
51 /*
52  * The "accumulator" word is nominally a 130-bit value. We split it into
53  * words of 44 bits, each held in a 64-bit variable.
54  *
55  * If the current accumulator is a = a0 + a1*W + a2*W^2 (where W = 2^44)
56  * and r = r0 + r1*W + r2*W^2, then:
57  *
58  *   a*r = (a0*r0)
59  *       + (a0*r1 + a1*r0) * W
60  *       + (a0*r2 + a1*r1 + a2*r0) * W^2
61  *       + (a1*r2 + a2*r1) * W^3
62  *       + (a2*r2) * W^4
63  *
64  * We want to reduce that value modulo p = 2^130-5, so W^3 = 20 mod p,
65  * and W^4 = 20*W mod p. Thus, if we define u1 = 20*r1 and u2 = 20*r2,
66  * then the equations above become:
67  *
68  *  b0 = a0*r0 + a1*u2 + a2*u1
69  *  b1 = a0*r1 + a1*r0 + a2*u2
70  *  b2 = a0*r2 + a1*r1 + a2*r0
71  *
72  * In order to make u1 fit in 44 bits, we can change these equations
73  * into:
74  *
75  *  b0 = a0*r0 + a1*u2 + a2*t1
76  *  b1 = a0*r1 + a1*r0 + a2*t2
77  *  b2 = a0*r2 + a1*r1 + a2*r0
78  *
79  * Where t1 is u1 truncated to 44 bits, and t2 is u2 added to the extra
80  * bits of u1. Note that since r is clamped down to a 124-bit value, the
81  * values u2 and t2 fit on 44 bits too.
82  *
83  * The bx values are larger than 44 bits, so we may split them into a
84  * lower half (cx, 44 bits) and an upper half (dx). The new values for
85  * the accumulator are then:
86  *
87  *  e0 = c0 + 20*d2
88  *  e1 = c1 + d0
89  *  e2 = c2 + d1
90  *
91  * The equations allow for some room, i.e. the ax values may be larger
92  * than 44 bits. Similarly, the ex values will usually be larger than
93  * the ax. Thus, some sort of carry propagation must be done regularly,
94  * though not necessarily at each iteration. In particular, we do not
95  * need to compute the additions (for the bx values) over 128-bit
96  * quantities; we can stick to 64-bit computations.
97  *
98  *
99  * Since the 128-bit result of a 64x64 multiplication is actually
100  * represented over two 64-bit registers, it is cheaper to arrange for
101  * any split that happens between the "high" and "low" halves to be on
102  * that 64-bit boundary. This is done by left shifting the rx, ux and tx
103  * by 20 bits (since they all fit on 44 bits each, this shift is
104  * always possible).
105  */
106 
107 static void
108 poly1305_inner_big(uint64_t *acc, uint64_t *r, const void *data, size_t len)
109 {
110 
111 #define MX(hi, lo, m0, m1, m2)   do { \
112 		uint64_t mxhi, mxlo; \
113 		MUL128(mxhi, mxlo, a0, m0); \
114 		(hi) = mxhi; \
115 		(lo) = mxlo >> 20; \
116 		MUL128(mxhi, mxlo, a1, m1); \
117 		(hi) += mxhi; \
118 		(lo) += mxlo >> 20; \
119 		MUL128(mxhi, mxlo, a2, m2); \
120 		(hi) += mxhi; \
121 		(lo) += mxlo >> 20; \
122 	} while (0)
123 
124 	const unsigned char *buf;
125 	uint64_t a0, a1, a2;
126 	uint64_t r0, r1, r2, t1, t2, u2;
127 
128 	r0 = r[0];
129 	r1 = r[1];
130 	r2 = r[2];
131 	t1 = r[3];
132 	t2 = r[4];
133 	u2 = r[5];
134 	a0 = acc[0];
135 	a1 = acc[1];
136 	a2 = acc[2];
137 	buf = data;
138 
139 	while (len > 0) {
140 		uint64_t v0, v1, v2;
141 		uint64_t c0, c1, c2, d0, d1, d2;
142 
143 		v0 = br_dec64le(buf + 0);
144 		v1 = br_dec64le(buf + 8);
145 		v2 = v1 >> 24;
146 		v1 = ((v0 >> 44) | (v1 << 20)) & MASK44;
147 		v0 &= MASK44;
148 		a0 += v0;
149 		a1 += v1;
150 		a2 += v2 + ((uint64_t)1 << 40);
151 		MX(d0, c0, r0, u2, t1);
152 		MX(d1, c1, r1, r0, t2);
153 		MX(d2, c2, r2, r1, r0);
154 		a0 = c0 + 20 * d2;
155 		a1 = c1 + d0;
156 		a2 = c2 + d1;
157 
158 		v0 = br_dec64le(buf + 16);
159 		v1 = br_dec64le(buf + 24);
160 		v2 = v1 >> 24;
161 		v1 = ((v0 >> 44) | (v1 << 20)) & MASK44;
162 		v0 &= MASK44;
163 		a0 += v0;
164 		a1 += v1;
165 		a2 += v2 + ((uint64_t)1 << 40);
166 		MX(d0, c0, r0, u2, t1);
167 		MX(d1, c1, r1, r0, t2);
168 		MX(d2, c2, r2, r1, r0);
169 		a0 = c0 + 20 * d2;
170 		a1 = c1 + d0;
171 		a2 = c2 + d1;
172 
173 		v0 = br_dec64le(buf + 32);
174 		v1 = br_dec64le(buf + 40);
175 		v2 = v1 >> 24;
176 		v1 = ((v0 >> 44) | (v1 << 20)) & MASK44;
177 		v0 &= MASK44;
178 		a0 += v0;
179 		a1 += v1;
180 		a2 += v2 + ((uint64_t)1 << 40);
181 		MX(d0, c0, r0, u2, t1);
182 		MX(d1, c1, r1, r0, t2);
183 		MX(d2, c2, r2, r1, r0);
184 		a0 = c0 + 20 * d2;
185 		a1 = c1 + d0;
186 		a2 = c2 + d1;
187 
188 		v0 = br_dec64le(buf + 48);
189 		v1 = br_dec64le(buf + 56);
190 		v2 = v1 >> 24;
191 		v1 = ((v0 >> 44) | (v1 << 20)) & MASK44;
192 		v0 &= MASK44;
193 		a0 += v0;
194 		a1 += v1;
195 		a2 += v2 + ((uint64_t)1 << 40);
196 		MX(d0, c0, r0, u2, t1);
197 		MX(d1, c1, r1, r0, t2);
198 		MX(d2, c2, r2, r1, r0);
199 		a0 = c0 + 20 * d2;
200 		a1 = c1 + d0;
201 		a2 = c2 + d1;
202 
203 		a1 += a0 >> 44;
204 		a0 &= MASK44;
205 		a2 += a1 >> 44;
206 		a1 &= MASK44;
207 		a0 += 20 * (a2 >> 44);
208 		a2 &= MASK44;
209 
210 		buf += 64;
211 		len -= 64;
212 	}
213 	acc[0] = a0;
214 	acc[1] = a1;
215 	acc[2] = a2;
216 
217 #undef MX
218 }
219 
220 static void
221 poly1305_inner_small(uint64_t *acc, uint64_t *r, const void *data, size_t len)
222 {
223 	const unsigned char *buf;
224 	uint64_t a0, a1, a2;
225 	uint64_t r0, r1, r2, t1, t2, u2;
226 
227 	r0 = r[0];
228 	r1 = r[1];
229 	r2 = r[2];
230 	t1 = r[3];
231 	t2 = r[4];
232 	u2 = r[5];
233 	a0 = acc[0];
234 	a1 = acc[1];
235 	a2 = acc[2];
236 	buf = data;
237 
238 	while (len > 0) {
239 		uint64_t v0, v1, v2;
240 		uint64_t c0, c1, c2, d0, d1, d2;
241 		unsigned char tmp[16];
242 
243 		if (len < 16) {
244 			memcpy(tmp, buf, len);
245 			memset(tmp + len, 0, (sizeof tmp) - len);
246 			buf = tmp;
247 			len = 16;
248 		}
249 		v0 = br_dec64le(buf + 0);
250 		v1 = br_dec64le(buf + 8);
251 
252 		v2 = v1 >> 24;
253 		v1 = ((v0 >> 44) | (v1 << 20)) & MASK44;
254 		v0 &= MASK44;
255 
256 		a0 += v0;
257 		a1 += v1;
258 		a2 += v2 + ((uint64_t)1 << 40);
259 
260 #define MX(hi, lo, m0, m1, m2)   do { \
261 		uint64_t mxhi, mxlo; \
262 		MUL128(mxhi, mxlo, a0, m0); \
263 		(hi) = mxhi; \
264 		(lo) = mxlo >> 20; \
265 		MUL128(mxhi, mxlo, a1, m1); \
266 		(hi) += mxhi; \
267 		(lo) += mxlo >> 20; \
268 		MUL128(mxhi, mxlo, a2, m2); \
269 		(hi) += mxhi; \
270 		(lo) += mxlo >> 20; \
271 	} while (0)
272 
273 		MX(d0, c0, r0, u2, t1);
274 		MX(d1, c1, r1, r0, t2);
275 		MX(d2, c2, r2, r1, r0);
276 
277 #undef MX
278 
279 		a0 = c0 + 20 * d2;
280 		a1 = c1 + d0;
281 		a2 = c2 + d1;
282 
283 		a1 += a0 >> 44;
284 		a0 &= MASK44;
285 		a2 += a1 >> 44;
286 		a1 &= MASK44;
287 		a0 += 20 * (a2 >> 44);
288 		a2 &= MASK44;
289 
290 		buf += 16;
291 		len -= 16;
292 	}
293 	acc[0] = a0;
294 	acc[1] = a1;
295 	acc[2] = a2;
296 }
297 
298 static inline void
299 poly1305_inner(uint64_t *acc, uint64_t *r, const void *data, size_t len)
300 {
301 	if (len >= 64) {
302 		size_t len2;
303 
304 		len2 = len & ~(size_t)63;
305 		poly1305_inner_big(acc, r, data, len2);
306 		data = (const unsigned char *)data + len2;
307 		len -= len2;
308 	}
309 	if (len > 0) {
310 		poly1305_inner_small(acc, r, data, len);
311 	}
312 }
313 
314 /* see bearssl_block.h */
315 void
316 br_poly1305_ctmulq_run(const void *key, const void *iv,
317 	void *data, size_t len, const void *aad, size_t aad_len,
318 	void *tag, br_chacha20_run ichacha, int encrypt)
319 {
320 	unsigned char pkey[32], foot[16];
321 	uint64_t r[6], acc[3], r0, r1;
322 	uint32_t v0, v1, v2, v3, v4;
323 	uint64_t w0, w1, w2, w3;
324 	uint32_t ctl;
325 
326 	/*
327 	 * Compute the MAC key. The 'r' value is the first 16 bytes of
328 	 * pkey[].
329 	 */
330 	memset(pkey, 0, sizeof pkey);
331 	ichacha(key, iv, 0, pkey, sizeof pkey);
332 
333 	/*
334 	 * If encrypting, ChaCha20 must run first, followed by Poly1305.
335 	 * When decrypting, the operations are reversed.
336 	 */
337 	if (encrypt) {
338 		ichacha(key, iv, 1, data, len);
339 	}
340 
341 	/*
342 	 * Run Poly1305. We must process the AAD, then ciphertext, then
343 	 * the footer (with the lengths). Note that the AAD and ciphertext
344 	 * are meant to be padded with zeros up to the next multiple of 16,
345 	 * and the length of the footer is 16 bytes as well.
346 	 */
347 
348 	/*
349 	 * Apply the "clamping" on r.
350 	 */
351 	pkey[ 3] &= 0x0F;
352 	pkey[ 4] &= 0xFC;
353 	pkey[ 7] &= 0x0F;
354 	pkey[ 8] &= 0xFC;
355 	pkey[11] &= 0x0F;
356 	pkey[12] &= 0xFC;
357 	pkey[15] &= 0x0F;
358 
359 	/*
360 	 * Decode the 'r' value into 44-bit words, left-shifted by 20 bits.
361 	 * Also compute the u1 and u2 values.
362 	 */
363 	r0 = br_dec64le(pkey +  0);
364 	r1 = br_dec64le(pkey +  8);
365 	r[0] = r0 << 20;
366 	r[1] = ((r0 >> 24) | (r1 << 40)) & ~(uint64_t)0xFFFFF;
367 	r[2] = (r1 >> 4) & ~(uint64_t)0xFFFFF;
368 	r1 = 20 * (r[1] >> 20);
369 	r[3] = r1 << 20;
370 	r[5] = 20 * r[2];
371 	r[4] = (r[5] + (r1 >> 24)) & ~(uint64_t)0xFFFFF;
372 
373 	/*
374 	 * Accumulator is 0.
375 	 */
376 	acc[0] = 0;
377 	acc[1] = 0;
378 	acc[2] = 0;
379 
380 	/*
381 	 * Process the additional authenticated data, ciphertext, and
382 	 * footer in due order.
383 	 */
384 	br_enc64le(foot, (uint64_t)aad_len);
385 	br_enc64le(foot + 8, (uint64_t)len);
386 	poly1305_inner(acc, r, aad, aad_len);
387 	poly1305_inner(acc, r, data, len);
388 	poly1305_inner_small(acc, r, foot, sizeof foot);
389 
390 	/*
391 	 * Finalise modular reduction. At that point, the value consists
392 	 * in three 44-bit values (the lowest one might be slightly above
393 	 * 2^44). Two loops shall be sufficient.
394 	 */
395 	acc[1] += (acc[0] >> 44);
396 	acc[0] &= MASK44;
397 	acc[2] += (acc[1] >> 44);
398 	acc[1] &= MASK44;
399 	acc[0] += 5 * (acc[2] >> 42);
400 	acc[2] &= MASK42;
401 	acc[1] += (acc[0] >> 44);
402 	acc[0] &= MASK44;
403 	acc[2] += (acc[1] >> 44);
404 	acc[1] &= MASK44;
405 	acc[0] += 5 * (acc[2] >> 42);
406 	acc[2] &= MASK42;
407 
408 	/*
409 	 * The value may still fall in the 2^130-5..2^130-1 range, in
410 	 * which case we must reduce it again. The code below selects,
411 	 * in constant-time, between 'acc' and 'acc-p'. We encode the
412 	 * value over four 32-bit integers to finish the operation.
413 	 */
414 	v0 = (uint32_t)acc[0];
415 	v1 = (uint32_t)(acc[0] >> 32) | ((uint32_t)acc[1] << 12);
416 	v2 = (uint32_t)(acc[1] >> 20) | ((uint32_t)acc[2] << 24);
417 	v3 = (uint32_t)(acc[2] >> 8);
418 	v4 = (uint32_t)(acc[2] >> 40);
419 
420 	ctl = GT(v0, 0xFFFFFFFA);
421 	ctl &= EQ(v1, 0xFFFFFFFF);
422 	ctl &= EQ(v2, 0xFFFFFFFF);
423 	ctl &= EQ(v3, 0xFFFFFFFF);
424 	ctl &= EQ(v4, 0x00000003);
425 	v0 = MUX(ctl, v0 + 5, v0);
426 	v1 = MUX(ctl, 0, v1);
427 	v2 = MUX(ctl, 0, v2);
428 	v3 = MUX(ctl, 0, v3);
429 
430 	/*
431 	 * Add the "s" value. This is done modulo 2^128. Don't forget
432 	 * carry propagation...
433 	 */
434 	w0 = (uint64_t)v0 + (uint64_t)br_dec32le(pkey + 16);
435 	w1 = (uint64_t)v1 + (uint64_t)br_dec32le(pkey + 20) + (w0 >> 32);
436 	w2 = (uint64_t)v2 + (uint64_t)br_dec32le(pkey + 24) + (w1 >> 32);
437 	w3 = (uint64_t)v3 + (uint64_t)br_dec32le(pkey + 28) + (w2 >> 32);
438 	v0 = (uint32_t)w0;
439 	v1 = (uint32_t)w1;
440 	v2 = (uint32_t)w2;
441 	v3 = (uint32_t)w3;
442 
443 	/*
444 	 * Encode the tag.
445 	 */
446 	br_enc32le((unsigned char *)tag +  0, v0);
447 	br_enc32le((unsigned char *)tag +  4, v1);
448 	br_enc32le((unsigned char *)tag +  8, v2);
449 	br_enc32le((unsigned char *)tag + 12, v3);
450 
451 	/*
452 	 * If decrypting, then ChaCha20 runs _after_ Poly1305.
453 	 */
454 	if (!encrypt) {
455 		ichacha(key, iv, 1, data, len);
456 	}
457 }
458 
459 /* see bearssl_block.h */
460 br_poly1305_run
461 br_poly1305_ctmulq_get(void)
462 {
463 	return &br_poly1305_ctmulq_run;
464 }
465 
466 #else
467 
468 /* see bearssl_block.h */
469 br_poly1305_run
470 br_poly1305_ctmulq_get(void)
471 {
472 	return 0;
473 }
474 
475 #endif
476