xref: /linux/lib/crypto/mldsa.c (revision 64edccea594cf7cb1e2975fdf44531e3377b32db)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Support for verifying ML-DSA signatures
4  *
5  * Copyright 2025 Google LLC
6  */
7 
8 #include <crypto/mldsa.h>
9 #include <crypto/sha3.h>
10 #include <kunit/visibility.h>
11 #include <linux/export.h>
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/string.h>
15 #include <linux/unaligned.h>
16 
17 #define Q 8380417 /* The prime q = 2^23 - 2^13 + 1 */
18 #define QINV_MOD_2_32 58728449 /* Multiplicative inverse of q mod 2^32 */
19 #define N 256 /* Number of components per ring element */
20 #define D 13 /* Number of bits dropped from the public key vector t */
21 #define RHO_LEN 32 /* Length of the public random seed in bytes */
22 #define MAX_W1_ENCODED_LEN 192 /* Max encoded length of one element of w'_1 */
23 
24 /*
25  * The zetas array in Montgomery form, i.e. with extra factor of 2^32.
26  * Reference: FIPS 204 Section 7.5 "NTT and NTT^-1"
27  * Generated by the following Python code:
28  * q=8380417; [a%q - q*(a%q > q//2) for a in [1753**(int(f'{i:08b}'[::-1], 2)) << 32 for i in range(256)]]
29  */
30 static const s32 zetas_times_2_32[N] = {
31 	-4186625, 25847,    -2608894, -518909,	237124,	  -777960,  -876248,
32 	466468,	  1826347,  2353451,  -359251,	-2091905, 3119733,  -2884855,
33 	3111497,  2680103,  2725464,  1024112,	-1079900, 3585928,  -549488,
34 	-1119584, 2619752,  -2108549, -2118186, -3859737, -1399561, -3277672,
35 	1757237,  -19422,   4010497,  280005,	2706023,  95776,    3077325,
36 	3530437,  -1661693, -3592148, -2537516, 3915439,  -3861115, -3043716,
37 	3574422,  -2867647, 3539968,  -300467,	2348700,  -539299,  -1699267,
38 	-1643818, 3505694,  -3821735, 3507263,	-2140649, -1600420, 3699596,
39 	811944,	  531354,   954230,   3881043,	3900724,  -2556880, 2071892,
40 	-2797779, -3930395, -1528703, -3677745, -3041255, -1452451, 3475950,
41 	2176455,  -1585221, -1257611, 1939314,	-4083598, -1000202, -3190144,
42 	-3157330, -3632928, 126922,   3412210,	-983419,  2147896,  2715295,
43 	-2967645, -3693493, -411027,  -2477047, -671102,  -1228525, -22981,
44 	-1308169, -381987,  1349076,  1852771,	-1430430, -3343383, 264944,
45 	508951,	  3097992,  44288,    -1100098, 904516,	  3958618,  -3724342,
46 	-8578,	  1653064,  -3249728, 2389356,	-210977,  759969,   -1316856,
47 	189548,	  -3553272, 3159746,  -1851402, -2409325, -177440,  1315589,
48 	1341330,  1285669,  -1584928, -812732,	-1439742, -3019102, -3881060,
49 	-3628969, 3839961,  2091667,  3407706,	2316500,  3817976,  -3342478,
50 	2244091,  -2446433, -3562462, 266997,	2434439,  -1235728, 3513181,
51 	-3520352, -3759364, -1197226, -3193378, 900702,	  1859098,  909542,
52 	819034,	  495491,   -1613174, -43260,	-522500,  -655327,  -3122442,
53 	2031748,  3207046,  -3556995, -525098,	-768622,  -3595838, 342297,
54 	286988,	  -2437823, 4108315,  3437287,	-3342277, 1735879,  203044,
55 	2842341,  2691481,  -2590150, 1265009,	4055324,  1247620,  2486353,
56 	1595974,  -3767016, 1250494,  2635921,	-3548272, -2994039, 1869119,
57 	1903435,  -1050970, -1333058, 1237275,	-3318210, -1430225, -451100,
58 	1312455,  3306115,  -1962642, -1279661, 1917081,  -2546312, -1374803,
59 	1500165,  777191,   2235880,  3406031,	-542412,  -2831860, -1671176,
60 	-1846953, -2584293, -3724270, 594136,	-3776993, -2013608, 2432395,
61 	2454455,  -164721,  1957272,  3369112,	185531,	  -1207385, -3183426,
62 	162844,	  1616392,  3014001,  810149,	1652634,  -3694233, -1799107,
63 	-3038916, 3523897,  3866901,  269760,	2213111,  -975884,  1717735,
64 	472078,	  -426683,  1723600,  -1803090, 1910376,  -1667432, -1104333,
65 	-260646,  -3833893, -2939036, -2235985, -420899,  -2286327, 183443,
66 	-976891,  1612842,  -3545687, -554416,	3919660,  -48306,   -1362209,
67 	3937738,  1400424,  -846154,  1976782
68 };
69 
70 /* Reference: FIPS 204 Section 4 "Parameter Sets" */
71 static const struct mldsa_parameter_set {
72 	u8 k; /* num rows in the matrix A */
73 	u8 l; /* num columns in the matrix A */
74 	u8 ctilde_len; /* length of commitment hash ctilde in bytes; lambda/4 */
75 	u8 omega; /* max num of 1's in the hint vector h */
76 	u8 tau; /* num of +-1's in challenge c */
77 	u8 beta; /* tau times eta */
78 	u16 pk_len; /* length of public keys in bytes */
79 	u16 sig_len; /* length of signatures in bytes */
80 	s32 gamma1; /* coefficient range of y */
81 } mldsa_parameter_sets[] = {
82 	[MLDSA44] = {
83 		.k = 4,
84 		.l = 4,
85 		.ctilde_len = 32,
86 		.omega = 80,
87 		.tau = 39,
88 		.beta = 78,
89 		.pk_len = MLDSA44_PUBLIC_KEY_SIZE,
90 		.sig_len = MLDSA44_SIGNATURE_SIZE,
91 		.gamma1 = 1 << 17,
92 	},
93 	[MLDSA65] = {
94 		.k = 6,
95 		.l = 5,
96 		.ctilde_len = 48,
97 		.omega = 55,
98 		.tau = 49,
99 		.beta = 196,
100 		.pk_len = MLDSA65_PUBLIC_KEY_SIZE,
101 		.sig_len = MLDSA65_SIGNATURE_SIZE,
102 		.gamma1 = 1 << 19,
103 	},
104 	[MLDSA87] = {
105 		.k = 8,
106 		.l = 7,
107 		.ctilde_len = 64,
108 		.omega = 75,
109 		.tau = 60,
110 		.beta = 120,
111 		.pk_len = MLDSA87_PUBLIC_KEY_SIZE,
112 		.sig_len = MLDSA87_SIGNATURE_SIZE,
113 		.gamma1 = 1 << 19,
114 	},
115 };
116 
117 /*
118  * An element of the ring R_q (normal form) or the ring T_q (NTT form).  It
119  * consists of N integers mod q: either the polynomial coefficients of the R_q
120  * element or the components of the T_q element.  In either case, whether they
121  * are fully reduced to [0, q - 1] varies in the different parts of the code.
122  */
123 struct mldsa_ring_elem {
124 	s32 x[N];
125 };
126 
127 struct mldsa_verification_workspace {
128 	/* SHAKE context for computing c, mu, and ctildeprime */
129 	struct shake_ctx shake;
130 	/* The fields in this union are used in their order of declaration. */
131 	union {
132 		/* The hash of the public key */
133 		u8 tr[64];
134 		/* The message representative mu */
135 		u8 mu[64];
136 		/* Temporary space for rej_ntt_poly() */
137 		u8 block[SHAKE128_BLOCK_SIZE + 1];
138 		/* Encoded element of w'_1 */
139 		u8 w1_encoded[MAX_W1_ENCODED_LEN];
140 		/* The commitment hash.  Real length is params->ctilde_len */
141 		u8 ctildeprime[64];
142 	};
143 	/* SHAKE context for generating elements of the matrix A */
144 	struct shake_ctx a_shake;
145 	/*
146 	 * An element of the matrix A generated from the public seed, or an
147 	 * element of the vector t_1 decoded from the public key and pre-scaled
148 	 * by 2^d.  Both are in NTT form.  To reduce memory usage, we generate
149 	 * or decode these elements only as needed.
150 	 */
151 	union {
152 		struct mldsa_ring_elem a;
153 		struct mldsa_ring_elem t1_scaled;
154 	};
155 	/* The challenge c, generated from ctilde */
156 	struct mldsa_ring_elem c;
157 	/* A temporary element used during calculations */
158 	struct mldsa_ring_elem tmp;
159 
160 	/* The following fields are variable-length: */
161 
162 	/* The signer's response vector */
163 	struct mldsa_ring_elem z[/* l */];
164 
165 	/* The signer's hint vector */
166 	/* u8 h[k * N]; */
167 };
168 
169 /*
170  * Compute a * b * 2^-32 mod q.  a * b must be in the range [-2^31 * q, 2^31 * q
171  * - 1] before reduction.  The return value is in the range [-q + 1, q - 1].
172  *
173  * To reduce mod q efficiently, this uses Montgomery reduction with R=2^32.
174  * That's where the factor of 2^-32 comes from.  The caller must include a
175  * factor of 2^32 at some point to compensate for that.
176  *
177  * To keep the input and output ranges very close to symmetric, this
178  * specifically does a "signed" Montgomery reduction.  That is, when computing
179  * d = c * q^-1 mod 2^32, this chooses a representative in [S32_MIN, S32_MAX]
180  * rather than [0, U32_MAX], i.e. s32 rather than u32.  This matters in the
181  * wider multiplication d * Q when d keeps its value via sign extension.
182  *
183  * Reference: FIPS 204 Appendix A "Montgomery Multiplication".  But, it doesn't
184  * explain it properly: it has an off-by-one error in the upper end of the input
185  * range, it doesn't clarify that the signed version should be used, and it
186  * gives an unnecessarily large output range.  A better citation is perhaps the
187  * Dilithium reference code, which functionally matches the below code and
188  * merely has the (benign) off-by-one error in its documentation.
189  */
190 static inline s32 Zq_mult(s32 a, s32 b)
191 {
192 	/* Compute the unreduced product c. */
193 	s64 c = (s64)a * b;
194 
195 	/*
196 	 * Compute d = c * q^-1 mod 2^32.  Generate a signed result, as
197 	 * explained above, but do the actual multiplication using an unsigned
198 	 * type to avoid signed integer overflow which is undefined behavior.
199 	 */
200 	s32 d = (u32)c * QINV_MOD_2_32;
201 
202 	/*
203 	 * Compute e = c - d * q.  This makes the low 32 bits zero, since
204 	 *   c - (c * q^-1) * q mod 2^32
205 	 * = c - c * (q^-1 * q) mod 2^32
206 	 * = c - c * 1 mod 2^32
207 	 * = c - c mod 2^32
208 	 * = 0 mod 2^32
209 	 */
210 	s64 e = c - (s64)d * Q;
211 
212 	/* Finally, return e * 2^-32. */
213 	return e >> 32;
214 }
215 
216 /*
217  * Convert @w to its number-theoretically-transformed representation in-place.
218  * Reference: FIPS 204 Algorithm 41, NTT
219  *
220  * To prevent intermediate overflows, all input coefficients must have absolute
221  * value < q.  All output components have absolute value < 9*q.
222  */
223 static void ntt(struct mldsa_ring_elem *w)
224 {
225 	int m = 0; /* index in zetas_times_2_32 */
226 
227 	for (int len = 128; len >= 1; len /= 2) {
228 		for (int start = 0; start < 256; start += 2 * len) {
229 			const s32 z = zetas_times_2_32[++m];
230 
231 			for (int j = start; j < start + len; j++) {
232 				s32 t = Zq_mult(z, w->x[j + len]);
233 
234 				w->x[j + len] = w->x[j] - t;
235 				w->x[j] += t;
236 			}
237 		}
238 	}
239 }
240 
241 /*
242  * Convert @w from its number-theoretically-transformed representation in-place.
243  * Reference: FIPS 204 Algorithm 42, NTT^-1
244  *
245  * This also multiplies the coefficients by 2^32, undoing an extra factor of
246  * 2^-32 introduced earlier, and reduces the coefficients to [0, q - 1].
247  */
248 static void invntt_and_mul_2_32(struct mldsa_ring_elem *w)
249 {
250 	int m = 256; /* index in zetas_times_2_32 */
251 
252 	/* Prevent intermediate overflows. */
253 	for (int j = 0; j < 256; j++)
254 		w->x[j] %= Q;
255 
256 	for (int len = 1; len < 256; len *= 2) {
257 		for (int start = 0; start < 256; start += 2 * len) {
258 			const s32 z = -zetas_times_2_32[--m];
259 
260 			for (int j = start; j < start + len; j++) {
261 				s32 t = w->x[j];
262 
263 				w->x[j] = t + w->x[j + len];
264 				w->x[j + len] = Zq_mult(z, t - w->x[j + len]);
265 			}
266 		}
267 	}
268 	/*
269 	 * Multiply by 2^32 * 256^-1.  2^32 cancels the factor of 2^-32 from
270 	 * earlier Montgomery multiplications.  256^-1 is for NTT^-1.  This
271 	 * itself uses Montgomery multiplication, so *another* 2^32 is needed.
272 	 * Thus the actual multiplicand is 2^32 * 2^32 * 256^-1 mod q = 41978.
273 	 *
274 	 * Finally, also reduce from [-q + 1, q - 1] to [0, q - 1].
275 	 */
276 	for (int j = 0; j < 256; j++) {
277 		w->x[j] = Zq_mult(w->x[j], 41978);
278 		w->x[j] += (w->x[j] >> 31) & Q;
279 	}
280 }
281 
282 /*
283  * Decode an element of t_1, i.e. the high d bits of t = A*s_1 + s_2.
284  * Reference: FIPS 204 Algorithm 23, pkDecode.
285  * Also multiply it by 2^d and convert it to NTT form.
286  */
287 static const u8 *decode_t1_elem(struct mldsa_ring_elem *out,
288 				const u8 *t1_encoded)
289 {
290 	for (int j = 0; j < N; j += 4, t1_encoded += 5) {
291 		u32 v = get_unaligned_le32(t1_encoded);
292 
293 		out->x[j + 0] = ((v >> 0) & 0x3ff) << D;
294 		out->x[j + 1] = ((v >> 10) & 0x3ff) << D;
295 		out->x[j + 2] = ((v >> 20) & 0x3ff) << D;
296 		out->x[j + 3] = ((v >> 30) | (t1_encoded[4] << 2)) << D;
297 		static_assert(0x3ff << D < Q); /* All coefficients < q. */
298 	}
299 	ntt(out);
300 	return t1_encoded; /* Return updated pointer. */
301 }
302 
303 /*
304  * Decode the signer's response vector 'z' from the signature.
305  * Reference: FIPS 204 Algorithm 27, sigDecode.
306  *
307  * This also validates that the coefficients of z are in range, corresponding
308  * the infinity norm check at the end of Algorithm 8, ML-DSA.Verify_internal.
309  *
310  * Finally, this also converts z to NTT form.
311  */
312 static bool decode_z(struct mldsa_ring_elem z[/* l */], int l, s32 gamma1,
313 		     int beta, const u8 **sig_ptr)
314 {
315 	const u8 *sig = *sig_ptr;
316 
317 	for (int i = 0; i < l; i++) {
318 		if (l == 4) { /* ML-DSA-44? */
319 			/* 18-bit coefficients: decode 4 from 9 bytes. */
320 			for (int j = 0; j < N; j += 4, sig += 9) {
321 				u64 v = get_unaligned_le64(sig);
322 
323 				z[i].x[j + 0] = (v >> 0) & 0x3ffff;
324 				z[i].x[j + 1] = (v >> 18) & 0x3ffff;
325 				z[i].x[j + 2] = (v >> 36) & 0x3ffff;
326 				z[i].x[j + 3] = (v >> 54) | (sig[8] << 10);
327 			}
328 		} else {
329 			/* 20-bit coefficients: decode 4 from 10 bytes. */
330 			for (int j = 0; j < N; j += 4, sig += 10) {
331 				u64 v = get_unaligned_le64(sig);
332 
333 				z[i].x[j + 0] = (v >> 0) & 0xfffff;
334 				z[i].x[j + 1] = (v >> 20) & 0xfffff;
335 				z[i].x[j + 2] = (v >> 40) & 0xfffff;
336 				z[i].x[j + 3] =
337 					(v >> 60) |
338 					(get_unaligned_le16(&sig[8]) << 4);
339 			}
340 		}
341 		for (int j = 0; j < N; j++) {
342 			z[i].x[j] = gamma1 - z[i].x[j];
343 			if (z[i].x[j] <= -(gamma1 - beta) ||
344 			    z[i].x[j] >= gamma1 - beta)
345 				return false;
346 		}
347 		ntt(&z[i]);
348 	}
349 	*sig_ptr = sig; /* Return updated pointer. */
350 	return true;
351 }
352 
353 /*
354  * Decode the signer's hint vector 'h' from the signature.
355  * Reference: FIPS 204 Algorithm 21, HintBitUnpack
356  *
357  * Note that there are several ways in which the hint vector can be malformed.
358  */
359 static bool decode_hint_vector(u8 h[/* k * N */], int k, int omega, const u8 *y)
360 {
361 	int index = 0;
362 
363 	memset(h, 0, k * N);
364 	for (int i = 0; i < k; i++) {
365 		int count = y[omega + i]; /* num 1's in elems 0 through i */
366 		int prev = -1;
367 
368 		/* Cumulative count mustn't decrease or exceed omega. */
369 		if (count < index || count > omega)
370 			return false;
371 		for (; index < count; index++) {
372 			if (prev >= y[index]) /* Coefficients out of order? */
373 				return false;
374 			prev = y[index];
375 			h[i * N + y[index]] = 1;
376 		}
377 	}
378 	return mem_is_zero(&y[index], omega - index);
379 }
380 
381 /*
382  * Expand @seed into an element of R_q @c with coefficients in {-1, 0, 1},
383  * exactly @tau of them nonzero.  Reference: FIPS 204 Algorithm 29, SampleInBall
384  */
385 static void sample_in_ball(struct mldsa_ring_elem *c, const u8 *seed,
386 			   size_t seed_len, int tau, struct shake_ctx *shake)
387 {
388 	u64 signs;
389 	u8 j;
390 
391 	shake256_init(shake);
392 	shake_update(shake, seed, seed_len);
393 	shake_squeeze(shake, (u8 *)&signs, sizeof(signs));
394 	le64_to_cpus(&signs);
395 	*c = (struct mldsa_ring_elem){};
396 	for (int i = N - tau; i < N; i++, signs >>= 1) {
397 		do {
398 			shake_squeeze(shake, &j, 1);
399 		} while (j > i);
400 		c->x[i] = c->x[j];
401 		c->x[j] = 1 - 2 * (s32)(signs & 1);
402 	}
403 }
404 
405 /*
406  * Expand the public seed @rho and @row_and_column into an element of T_q @out.
407  * Reference: FIPS 204 Algorithm 30, RejNTTPoly
408  *
409  * @shake and @block are temporary space used by the expansion.  @block has
410  * space for one SHAKE128 block, plus an extra byte to allow reading a u32 from
411  * the final 3-byte group without reading out-of-bounds.
412  */
413 static void rej_ntt_poly(struct mldsa_ring_elem *out, const u8 rho[RHO_LEN],
414 			 __le16 row_and_column, struct shake_ctx *shake,
415 			 u8 block[SHAKE128_BLOCK_SIZE + 1])
416 {
417 	shake128_init(shake);
418 	shake_update(shake, rho, RHO_LEN);
419 	shake_update(shake, (u8 *)&row_and_column, sizeof(row_and_column));
420 	for (int i = 0; i < N;) {
421 		shake_squeeze(shake, block, SHAKE128_BLOCK_SIZE);
422 		block[SHAKE128_BLOCK_SIZE] = 0; /* for KMSAN */
423 		static_assert(SHAKE128_BLOCK_SIZE % 3 == 0);
424 		for (int j = 0; j < SHAKE128_BLOCK_SIZE && i < N; j += 3) {
425 			u32 x = get_unaligned_le32(&block[j]) & 0x7fffff;
426 
427 			if (x < Q) /* Ignore values >= q. */
428 				out->x[i++] = x;
429 		}
430 	}
431 }
432 
433 /*
434  * Return the HighBits of r adjusted according to hint h
435  * Reference: FIPS 204 Algorithm 40, UseHint
436  *
437  * This is needed because of the public key compression in ML-DSA.
438  *
439  * h is either 0 or 1, r is in [0, q - 1], and gamma2 is either (q - 1) / 88 or
440  * (q - 1) / 32.  Except when invoked via the unit test interface, gamma2 is a
441  * compile-time constant, so compilers will optimize the code accordingly.
442  */
443 static __always_inline s32 use_hint(u8 h, s32 r, const s32 gamma2)
444 {
445 	const s32 m = (Q - 1) / (2 * gamma2); /* 44 or 16, compile-time const */
446 	s32 r1;
447 
448 	/*
449 	 * Handle the special case where r - (r mod+- (2 * gamma2)) == q - 1,
450 	 * i.e. r >= q - gamma2.  This is also exactly where the computation of
451 	 * r1 below would produce 'm' and would need a correction.
452 	 */
453 	if (r >= Q - gamma2)
454 		return h == 0 ? 0 : m - 1;
455 
456 	/*
457 	 * Compute the (non-hint-adjusted) HighBits r1 as:
458 	 *
459 	 *  r1 = (r - (r mod+- (2 * gamma2))) / (2 * gamma2)
460 	 *     = floor((r + gamma2 - 1) / (2 * gamma2))
461 	 *
462 	 * Note that when '2 * gamma2' is a compile-time constant, compilers
463 	 * optimize the division to a reciprocal multiplication and shift.
464 	 */
465 	r1 = (u32)(r + gamma2 - 1) / (2 * gamma2);
466 
467 	/*
468 	 * Return the HighBits r1:
469 	 *	+ 0 if the hint is 0;
470 	 *	+ 1 (mod m) if the hint is 1 and the LowBits are positive;
471 	 *	- 1 (mod m) if the hint is 1 and the LowBits are negative or 0.
472 	 *
473 	 * r1 is in (and remains in) [0, m - 1].  Note that when 'm' is a
474 	 * compile-time constant, compilers optimize the '% m' accordingly.
475 	 */
476 	if (h == 0)
477 		return r1;
478 	if (r > r1 * (2 * gamma2))
479 		return (u32)(r1 + 1) % m;
480 	return (u32)(r1 + m - 1) % m;
481 }
482 
483 static __always_inline void use_hint_elem(struct mldsa_ring_elem *w,
484 					  const u8 h[N], const s32 gamma2)
485 {
486 	for (int j = 0; j < N; j++)
487 		w->x[j] = use_hint(h[j], w->x[j], gamma2);
488 }
489 
490 #if IS_ENABLED(CONFIG_CRYPTO_LIB_MLDSA_KUNIT_TEST)
491 /* Allow the __always_inline function use_hint() to be unit-tested. */
492 s32 mldsa_use_hint(u8 h, s32 r, s32 gamma2)
493 {
494 	return use_hint(h, r, gamma2);
495 }
496 EXPORT_SYMBOL_IF_KUNIT(mldsa_use_hint);
497 #endif
498 
499 /*
500  * Encode one element of the commitment vector w'_1 into a byte string.
501  * Reference: FIPS 204 Algorithm 28, w1Encode.
502  * Return the number of bytes used: 192 for ML-DSA-44 and 128 for the others.
503  */
504 static size_t encode_w1(u8 out[MAX_W1_ENCODED_LEN],
505 			const struct mldsa_ring_elem *w1, int k)
506 {
507 	size_t pos = 0;
508 
509 	static_assert(N * 6 / 8 == MAX_W1_ENCODED_LEN);
510 	if (k == 4) { /* ML-DSA-44? */
511 		/* 6 bits per coefficient.  Pack 4 at a time. */
512 		for (int j = 0; j < N; j += 4) {
513 			u32 v = (w1->x[j + 0] << 0) | (w1->x[j + 1] << 6) |
514 				(w1->x[j + 2] << 12) | (w1->x[j + 3] << 18);
515 			out[pos++] = v >> 0;
516 			out[pos++] = v >> 8;
517 			out[pos++] = v >> 16;
518 		}
519 	} else {
520 		/* 4 bits per coefficient.  Pack 2 at a time. */
521 		for (int j = 0; j < N; j += 2)
522 			out[pos++] = w1->x[j] | (w1->x[j + 1] << 4);
523 	}
524 	return pos;
525 }
526 
527 /* Reference: FIPS 204 Section 6.3 "ML-DSA Verifying (Internal)" */
528 int mldsa_verify(enum mldsa_alg alg, const u8 *sig, size_t sig_len,
529 		 const u8 *msg, size_t msg_len, const u8 *pk, size_t pk_len)
530 {
531 	const struct mldsa_parameter_set *params = &mldsa_parameter_sets[alg];
532 	const int k = params->k, l = params->l;
533 	/* For now this just does pure ML-DSA with an empty context string. */
534 	static const u8 msg_prefix[2] = { /* dom_sep= */ 0, /* ctx_len= */ 0 };
535 	const u8 *ctilde; /* The signer's commitment hash */
536 	const u8 *t1_encoded = &pk[RHO_LEN]; /* Next encoded element of t_1 */
537 	u8 *h; /* The signer's hint vector, length k * N */
538 	size_t w1_enc_len;
539 
540 	/* Validate the public key and signature lengths. */
541 	if (pk_len != params->pk_len || sig_len != params->sig_len)
542 		return -EBADMSG;
543 
544 	/*
545 	 * Allocate the workspace, including variable-length fields.  Its size
546 	 * depends only on the ML-DSA parameter set, not the other inputs.
547 	 *
548 	 * For freeing it, use kfree_sensitive() rather than kfree().  This is
549 	 * mainly to comply with FIPS 204 Section 3.6.3 "Intermediate Values".
550 	 * In reality it's a bit gratuitous, as this is a public key operation.
551 	 */
552 	struct mldsa_verification_workspace *ws __free(kfree_sensitive) =
553 		kmalloc(sizeof(*ws) + (l * sizeof(ws->z[0])) + (k * N),
554 			GFP_KERNEL);
555 	if (!ws)
556 		return -ENOMEM;
557 	h = (u8 *)&ws->z[l];
558 
559 	/* Decode the signature.  Reference: FIPS 204 Algorithm 27, sigDecode */
560 	ctilde = sig;
561 	sig += params->ctilde_len;
562 	if (!decode_z(ws->z, l, params->gamma1, params->beta, &sig))
563 		return -EBADMSG;
564 	if (!decode_hint_vector(h, k, params->omega, sig))
565 		return -EBADMSG;
566 
567 	/* Recreate the challenge c from the signer's commitment hash. */
568 	sample_in_ball(&ws->c, ctilde, params->ctilde_len, params->tau,
569 		       &ws->shake);
570 	ntt(&ws->c);
571 
572 	/* Compute the message representative mu. */
573 	shake256(pk, pk_len, ws->tr, sizeof(ws->tr));
574 	shake256_init(&ws->shake);
575 	shake_update(&ws->shake, ws->tr, sizeof(ws->tr));
576 	shake_update(&ws->shake, msg_prefix, sizeof(msg_prefix));
577 	shake_update(&ws->shake, msg, msg_len);
578 	shake_squeeze(&ws->shake, ws->mu, sizeof(ws->mu));
579 
580 	/* Start computing ctildeprime = H(mu || w1Encode(w'_1)). */
581 	shake256_init(&ws->shake);
582 	shake_update(&ws->shake, ws->mu, sizeof(ws->mu));
583 
584 	/*
585 	 * Compute the commitment w'_1 from A, z, c, t_1, and h.
586 	 *
587 	 * The computation is the same for each of the k rows.  Just do each row
588 	 * before moving on to the next, resulting in only one loop over k.
589 	 */
590 	for (int i = 0; i < k; i++) {
591 		/*
592 		 * tmp = NTT(A) * NTT(z) * 2^-32
593 		 * To reduce memory use, generate each element of NTT(A)
594 		 * on-demand.  Note that each element is used only once.
595 		 */
596 		ws->tmp = (struct mldsa_ring_elem){};
597 		for (int j = 0; j < l; j++) {
598 			rej_ntt_poly(&ws->a, pk /* rho is first field of pk */,
599 				     cpu_to_le16((i << 8) | j), &ws->a_shake,
600 				     ws->block);
601 			for (int n = 0; n < N; n++)
602 				ws->tmp.x[n] +=
603 					Zq_mult(ws->a.x[n], ws->z[j].x[n]);
604 		}
605 		/* All components of tmp now have abs value < l*q. */
606 
607 		/* Decode the next element of t_1. */
608 		t1_encoded = decode_t1_elem(&ws->t1_scaled, t1_encoded);
609 
610 		/*
611 		 * tmp -= NTT(c) * NTT(t_1 * 2^d) * 2^-32
612 		 *
613 		 * Taking a conservative bound for the output of ntt(), the
614 		 * multiplicands can have absolute value up to 9*q.  That
615 		 * corresponds to a product with absolute value 81*q^2.  That is
616 		 * within the limits of Zq_mult() which needs < ~256*q^2.
617 		 */
618 		for (int j = 0; j < N; j++)
619 			ws->tmp.x[j] -= Zq_mult(ws->c.x[j], ws->t1_scaled.x[j]);
620 		/* All components of tmp now have abs value < (l+1)*q. */
621 
622 		/* tmp = w'_Approx = NTT^-1(tmp) * 2^32 */
623 		invntt_and_mul_2_32(&ws->tmp);
624 		/* All coefficients of tmp are now in [0, q - 1]. */
625 
626 		/*
627 		 * tmp = w'_1 = UseHint(h, w'_Approx)
628 		 * For efficiency, set gamma2 to a compile-time constant.
629 		 */
630 		if (k == 4)
631 			use_hint_elem(&ws->tmp, &h[i * N], (Q - 1) / 88);
632 		else
633 			use_hint_elem(&ws->tmp, &h[i * N], (Q - 1) / 32);
634 
635 		/* Encode and hash the next element of w'_1. */
636 		w1_enc_len = encode_w1(ws->w1_encoded, &ws->tmp, k);
637 		shake_update(&ws->shake, ws->w1_encoded, w1_enc_len);
638 	}
639 
640 	/* Finish computing ctildeprime. */
641 	shake_squeeze(&ws->shake, ws->ctildeprime, params->ctilde_len);
642 
643 	/* Verify that ctilde == ctildeprime. */
644 	if (memcmp(ws->ctildeprime, ctilde, params->ctilde_len) != 0)
645 		return -EKEYREJECTED;
646 	/* ||z||_infinity < gamma1 - beta was already checked in decode_z(). */
647 	return 0;
648 }
649 EXPORT_SYMBOL_GPL(mldsa_verify);
650 
651 MODULE_DESCRIPTION("ML-DSA signature verification");
652 MODULE_LICENSE("GPL");
653