1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * POLYVAL library functions
4 *
5 * Copyright 2025 Google LLC
6 */
7
8 #include <crypto/polyval.h>
9 #include <linux/export.h>
10 #include <linux/module.h>
11 #include <linux/string.h>
12 #include <linux/unaligned.h>
13
14 /*
15 * POLYVAL is an almost-XOR-universal hash function. Similar to GHASH, POLYVAL
16 * interprets the message as the coefficients of a polynomial in GF(2^128) and
17 * evaluates that polynomial at a secret point. POLYVAL has a simple
18 * mathematical relationship with GHASH, but it uses a better field convention
19 * which makes it easier and faster to implement.
20 *
21 * POLYVAL is not a cryptographic hash function, and it should be used only by
22 * algorithms that are specifically designed to use it.
23 *
24 * POLYVAL is specified by "AES-GCM-SIV: Nonce Misuse-Resistant Authenticated
25 * Encryption" (https://datatracker.ietf.org/doc/html/rfc8452)
26 *
27 * POLYVAL is also used by HCTR2. See "Length-preserving encryption with HCTR2"
28 * (https://eprint.iacr.org/2021/1441.pdf).
29 *
30 * This file provides a library API for POLYVAL. This API can delegate to
31 * either a generic implementation or an architecture-optimized implementation.
32 *
33 * For the generic implementation, we don't use the traditional table approach
34 * to GF(2^128) multiplication. That approach is not constant-time and requires
35 * a lot of memory. Instead, we use a different approach which emulates
36 * carryless multiplication using standard multiplications by spreading the data
37 * bits apart using "holes". This allows the carries to spill harmlessly. This
38 * approach is borrowed from BoringSSL, which in turn credits BearSSL's
39 * documentation (https://bearssl.org/constanttime.html#ghash-for-gcm) for the
40 * "holes" trick and a presentation by Shay Gueron
41 * (https://crypto.stanford.edu/RealWorldCrypto/slides/gueron.pdf) for the
42 * 256-bit => 128-bit reduction algorithm.
43 */
44
45 #ifdef CONFIG_ARCH_SUPPORTS_INT128
46
47 /* Do a 64 x 64 => 128 bit carryless multiplication. */
clmul64(u64 a,u64 b,u64 * out_lo,u64 * out_hi)48 static void clmul64(u64 a, u64 b, u64 *out_lo, u64 *out_hi)
49 {
50 /*
51 * With 64-bit multiplicands and one term every 4 bits, there would be
52 * up to 64 / 4 = 16 one bits per column when each multiplication is
53 * written out as a series of additions in the schoolbook manner.
54 * Unfortunately, that doesn't work since the value 16 is 1 too large to
55 * fit in 4 bits. Carries would sometimes overflow into the next term.
56 *
57 * Using one term every 5 bits would work. However, that would cost
58 * 5 x 5 = 25 multiplications instead of 4 x 4 = 16.
59 *
60 * Instead, mask off 4 bits from one multiplicand, giving a max of 15
61 * one bits per column. Then handle those 4 bits separately.
62 */
63 u64 a0 = a & 0x1111111111111110;
64 u64 a1 = a & 0x2222222222222220;
65 u64 a2 = a & 0x4444444444444440;
66 u64 a3 = a & 0x8888888888888880;
67
68 u64 b0 = b & 0x1111111111111111;
69 u64 b1 = b & 0x2222222222222222;
70 u64 b2 = b & 0x4444444444444444;
71 u64 b3 = b & 0x8888888888888888;
72
73 /* Multiply the high 60 bits of @a by @b. */
74 u128 c0 = (a0 * (u128)b0) ^ (a1 * (u128)b3) ^
75 (a2 * (u128)b2) ^ (a3 * (u128)b1);
76 u128 c1 = (a0 * (u128)b1) ^ (a1 * (u128)b0) ^
77 (a2 * (u128)b3) ^ (a3 * (u128)b2);
78 u128 c2 = (a0 * (u128)b2) ^ (a1 * (u128)b1) ^
79 (a2 * (u128)b0) ^ (a3 * (u128)b3);
80 u128 c3 = (a0 * (u128)b3) ^ (a1 * (u128)b2) ^
81 (a2 * (u128)b1) ^ (a3 * (u128)b0);
82
83 /* Multiply the low 4 bits of @a by @b. */
84 u64 e0 = -(a & 1) & b;
85 u64 e1 = -((a >> 1) & 1) & b;
86 u64 e2 = -((a >> 2) & 1) & b;
87 u64 e3 = -((a >> 3) & 1) & b;
88 u64 extra_lo = e0 ^ (e1 << 1) ^ (e2 << 2) ^ (e3 << 3);
89 u64 extra_hi = (e1 >> 63) ^ (e2 >> 62) ^ (e3 >> 61);
90
91 /* Add all the intermediate products together. */
92 *out_lo = (((u64)c0) & 0x1111111111111111) ^
93 (((u64)c1) & 0x2222222222222222) ^
94 (((u64)c2) & 0x4444444444444444) ^
95 (((u64)c3) & 0x8888888888888888) ^ extra_lo;
96 *out_hi = (((u64)(c0 >> 64)) & 0x1111111111111111) ^
97 (((u64)(c1 >> 64)) & 0x2222222222222222) ^
98 (((u64)(c2 >> 64)) & 0x4444444444444444) ^
99 (((u64)(c3 >> 64)) & 0x8888888888888888) ^ extra_hi;
100 }
101
102 #else /* CONFIG_ARCH_SUPPORTS_INT128 */
103
104 /* Do a 32 x 32 => 64 bit carryless multiplication. */
clmul32(u32 a,u32 b)105 static u64 clmul32(u32 a, u32 b)
106 {
107 /*
108 * With 32-bit multiplicands and one term every 4 bits, there are up to
109 * 32 / 4 = 8 one bits per column when each multiplication is written
110 * out as a series of additions in the schoolbook manner. The value 8
111 * fits in 4 bits, so the carries don't overflow into the next term.
112 */
113 u32 a0 = a & 0x11111111;
114 u32 a1 = a & 0x22222222;
115 u32 a2 = a & 0x44444444;
116 u32 a3 = a & 0x88888888;
117
118 u32 b0 = b & 0x11111111;
119 u32 b1 = b & 0x22222222;
120 u32 b2 = b & 0x44444444;
121 u32 b3 = b & 0x88888888;
122
123 u64 c0 = (a0 * (u64)b0) ^ (a1 * (u64)b3) ^
124 (a2 * (u64)b2) ^ (a3 * (u64)b1);
125 u64 c1 = (a0 * (u64)b1) ^ (a1 * (u64)b0) ^
126 (a2 * (u64)b3) ^ (a3 * (u64)b2);
127 u64 c2 = (a0 * (u64)b2) ^ (a1 * (u64)b1) ^
128 (a2 * (u64)b0) ^ (a3 * (u64)b3);
129 u64 c3 = (a0 * (u64)b3) ^ (a1 * (u64)b2) ^
130 (a2 * (u64)b1) ^ (a3 * (u64)b0);
131
132 /* Add all the intermediate products together. */
133 return (c0 & 0x1111111111111111) ^
134 (c1 & 0x2222222222222222) ^
135 (c2 & 0x4444444444444444) ^
136 (c3 & 0x8888888888888888);
137 }
138
139 /* Do a 64 x 64 => 128 bit carryless multiplication. */
clmul64(u64 a,u64 b,u64 * out_lo,u64 * out_hi)140 static void clmul64(u64 a, u64 b, u64 *out_lo, u64 *out_hi)
141 {
142 u32 a_lo = (u32)a;
143 u32 a_hi = a >> 32;
144 u32 b_lo = (u32)b;
145 u32 b_hi = b >> 32;
146
147 /* Karatsuba multiplication */
148 u64 lo = clmul32(a_lo, b_lo);
149 u64 hi = clmul32(a_hi, b_hi);
150 u64 mi = clmul32(a_lo ^ a_hi, b_lo ^ b_hi) ^ lo ^ hi;
151
152 *out_lo = lo ^ (mi << 32);
153 *out_hi = hi ^ (mi >> 32);
154 }
155 #endif /* !CONFIG_ARCH_SUPPORTS_INT128 */
156
157 /* Compute @a = @a * @b * x^-128 in the POLYVAL field. */
158 static void __maybe_unused
polyval_mul_generic(struct polyval_elem * a,const struct polyval_elem * b)159 polyval_mul_generic(struct polyval_elem *a, const struct polyval_elem *b)
160 {
161 u64 c0, c1, c2, c3, mi0, mi1;
162
163 /*
164 * Carryless-multiply @a by @b using Karatsuba multiplication. Store
165 * the 256-bit product in @c0 (low) through @c3 (high).
166 */
167 clmul64(le64_to_cpu(a->lo), le64_to_cpu(b->lo), &c0, &c1);
168 clmul64(le64_to_cpu(a->hi), le64_to_cpu(b->hi), &c2, &c3);
169 clmul64(le64_to_cpu(a->lo ^ a->hi), le64_to_cpu(b->lo ^ b->hi),
170 &mi0, &mi1);
171 mi0 ^= c0 ^ c2;
172 mi1 ^= c1 ^ c3;
173 c1 ^= mi0;
174 c2 ^= mi1;
175
176 /*
177 * Cancel out the low 128 bits of the product by adding multiples of
178 * G(x) = x^128 + x^127 + x^126 + x^121 + 1. Do this in two steps, each
179 * of which cancels out 64 bits. Note that we break G(x) into three
180 * parts: 1, x^64 * (x^63 + x^62 + x^57), and x^128 * 1.
181 */
182
183 /*
184 * First, add G(x) times c0 as follows:
185 *
186 * (c0, c1, c2) = (0,
187 * c1 + (c0 * (x^63 + x^62 + x^57) mod x^64),
188 * c2 + c0 + floor((c0 * (x^63 + x^62 + x^57)) / x^64))
189 */
190 c1 ^= (c0 << 63) ^ (c0 << 62) ^ (c0 << 57);
191 c2 ^= c0 ^ (c0 >> 1) ^ (c0 >> 2) ^ (c0 >> 7);
192
193 /*
194 * Second, add G(x) times the new c1:
195 *
196 * (c1, c2, c3) = (0,
197 * c2 + (c1 * (x^63 + x^62 + x^57) mod x^64),
198 * c3 + c1 + floor((c1 * (x^63 + x^62 + x^57)) / x^64))
199 */
200 c2 ^= (c1 << 63) ^ (c1 << 62) ^ (c1 << 57);
201 c3 ^= c1 ^ (c1 >> 1) ^ (c1 >> 2) ^ (c1 >> 7);
202
203 /* Return (c2, c3). This implicitly multiplies by x^-128. */
204 a->lo = cpu_to_le64(c2);
205 a->hi = cpu_to_le64(c3);
206 }
207
208 static void __maybe_unused
polyval_blocks_generic(struct polyval_elem * acc,const struct polyval_elem * key,const u8 * data,size_t nblocks)209 polyval_blocks_generic(struct polyval_elem *acc, const struct polyval_elem *key,
210 const u8 *data, size_t nblocks)
211 {
212 do {
213 acc->lo ^= get_unaligned((__le64 *)data);
214 acc->hi ^= get_unaligned((__le64 *)(data + 8));
215 polyval_mul_generic(acc, key);
216 data += POLYVAL_BLOCK_SIZE;
217 } while (--nblocks);
218 }
219
220 /* Include the arch-optimized implementation of POLYVAL, if one is available. */
221 #ifdef CONFIG_CRYPTO_LIB_POLYVAL_ARCH
222 #include "polyval.h" /* $(SRCARCH)/polyval.h */
polyval_preparekey(struct polyval_key * key,const u8 raw_key[POLYVAL_BLOCK_SIZE])223 void polyval_preparekey(struct polyval_key *key,
224 const u8 raw_key[POLYVAL_BLOCK_SIZE])
225 {
226 polyval_preparekey_arch(key, raw_key);
227 }
228 EXPORT_SYMBOL_GPL(polyval_preparekey);
229 #endif /* Else, polyval_preparekey() is an inline function. */
230
231 /*
232 * polyval_mul_generic() and polyval_blocks_generic() take the key as a
233 * polyval_elem rather than a polyval_key, so that arch-optimized
234 * implementations with a different key format can use it as a fallback (if they
235 * have H^1 stored somewhere in their struct). Thus, the following dispatch
236 * code is needed to pass the appropriate key argument.
237 */
238
polyval_mul(struct polyval_ctx * ctx)239 static void polyval_mul(struct polyval_ctx *ctx)
240 {
241 #ifdef CONFIG_CRYPTO_LIB_POLYVAL_ARCH
242 polyval_mul_arch(&ctx->acc, ctx->key);
243 #else
244 polyval_mul_generic(&ctx->acc, &ctx->key->h);
245 #endif
246 }
247
polyval_blocks(struct polyval_ctx * ctx,const u8 * data,size_t nblocks)248 static void polyval_blocks(struct polyval_ctx *ctx,
249 const u8 *data, size_t nblocks)
250 {
251 #ifdef CONFIG_CRYPTO_LIB_POLYVAL_ARCH
252 polyval_blocks_arch(&ctx->acc, ctx->key, data, nblocks);
253 #else
254 polyval_blocks_generic(&ctx->acc, &ctx->key->h, data, nblocks);
255 #endif
256 }
257
polyval_update(struct polyval_ctx * ctx,const u8 * data,size_t len)258 void polyval_update(struct polyval_ctx *ctx, const u8 *data, size_t len)
259 {
260 if (unlikely(ctx->partial)) {
261 size_t n = min(len, POLYVAL_BLOCK_SIZE - ctx->partial);
262
263 len -= n;
264 while (n--)
265 ctx->acc.bytes[ctx->partial++] ^= *data++;
266 if (ctx->partial < POLYVAL_BLOCK_SIZE)
267 return;
268 polyval_mul(ctx);
269 }
270 if (len >= POLYVAL_BLOCK_SIZE) {
271 size_t nblocks = len / POLYVAL_BLOCK_SIZE;
272
273 polyval_blocks(ctx, data, nblocks);
274 data += len & ~(POLYVAL_BLOCK_SIZE - 1);
275 len &= POLYVAL_BLOCK_SIZE - 1;
276 }
277 for (size_t i = 0; i < len; i++)
278 ctx->acc.bytes[i] ^= data[i];
279 ctx->partial = len;
280 }
281 EXPORT_SYMBOL_GPL(polyval_update);
282
polyval_final(struct polyval_ctx * ctx,u8 out[POLYVAL_BLOCK_SIZE])283 void polyval_final(struct polyval_ctx *ctx, u8 out[POLYVAL_BLOCK_SIZE])
284 {
285 if (unlikely(ctx->partial))
286 polyval_mul(ctx);
287 memcpy(out, &ctx->acc, POLYVAL_BLOCK_SIZE);
288 memzero_explicit(ctx, sizeof(*ctx));
289 }
290 EXPORT_SYMBOL_GPL(polyval_final);
291
292 #ifdef polyval_mod_init_arch
polyval_mod_init(void)293 static int __init polyval_mod_init(void)
294 {
295 polyval_mod_init_arch();
296 return 0;
297 }
298 subsys_initcall(polyval_mod_init);
299
polyval_mod_exit(void)300 static void __exit polyval_mod_exit(void)
301 {
302 }
303 module_exit(polyval_mod_exit);
304 #endif
305
306 MODULE_DESCRIPTION("POLYVAL almost-XOR-universal hash function");
307 MODULE_LICENSE("GPL");
308