poly1305-donna32.c

  1// SPDX-License-Identifier: GPL-2.0 OR MIT
  2/*
  3 * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
  4 *
  5 * This is based in part on Andrew Moon's poly1305-donna, which is in the
  6 * public domain.
  7 */
  8
  9#include <crypto/internal/poly1305.h>
 10#include <linux/export.h>
 11#include <linux/kernel.h>
 12#include <linux/unaligned.h>
 13
 14void poly1305_core_setkey(struct poly1305_core_key *key,
 15			  const u8 raw_key[POLY1305_BLOCK_SIZE])
 16{
 17	/* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 18	key->key.r[0] = (get_unaligned_le32(&raw_key[0])) & 0x3ffffff;
 19	key->key.r[1] = (get_unaligned_le32(&raw_key[3]) >> 2) & 0x3ffff03;
 20	key->key.r[2] = (get_unaligned_le32(&raw_key[6]) >> 4) & 0x3ffc0ff;
 21	key->key.r[3] = (get_unaligned_le32(&raw_key[9]) >> 6) & 0x3f03fff;
 22	key->key.r[4] = (get_unaligned_le32(&raw_key[12]) >> 8) & 0x00fffff;
 23
 24	/* s = 5*r */
 25	key->precomputed_s.r[0] = key->key.r[1] * 5;
 26	key->precomputed_s.r[1] = key->key.r[2] * 5;
 27	key->precomputed_s.r[2] = key->key.r[3] * 5;
 28	key->precomputed_s.r[3] = key->key.r[4] * 5;
 29}
 30EXPORT_SYMBOL(poly1305_core_setkey);
 31
 32void poly1305_core_blocks(struct poly1305_state *state,
 33			  const struct poly1305_core_key *key, const void *src,
 34			  unsigned int nblocks, u32 hibit)
 35{
 36	const u8 *input = src;
 37	u32 r0, r1, r2, r3, r4;
 38	u32 s1, s2, s3, s4;
 39	u32 h0, h1, h2, h3, h4;
 40	u64 d0, d1, d2, d3, d4;
 41	u32 c;
 42
 43	if (!nblocks)
 44		return;
 45
 46	hibit <<= 24;
 47
 48	r0 = key->key.r[0];
 49	r1 = key->key.r[1];
 50	r2 = key->key.r[2];
 51	r3 = key->key.r[3];
 52	r4 = key->key.r[4];
 53
 54	s1 = key->precomputed_s.r[0];
 55	s2 = key->precomputed_s.r[1];
 56	s3 = key->precomputed_s.r[2];
 57	s4 = key->precomputed_s.r[3];
 58
 59	h0 = state->h[0];
 60	h1 = state->h[1];
 61	h2 = state->h[2];
 62	h3 = state->h[3];
 63	h4 = state->h[4];
 64
 65	do {
 66		/* h += m[i] */
 67		h0 += (get_unaligned_le32(&input[0])) & 0x3ffffff;
 68		h1 += (get_unaligned_le32(&input[3]) >> 2) & 0x3ffffff;
 69		h2 += (get_unaligned_le32(&input[6]) >> 4) & 0x3ffffff;
 70		h3 += (get_unaligned_le32(&input[9]) >> 6) & 0x3ffffff;
 71		h4 += (get_unaligned_le32(&input[12]) >> 8) | hibit;
 72
 73		/* h *= r */
 74		d0 = ((u64)h0 * r0) + ((u64)h1 * s4) +
 75		     ((u64)h2 * s3) + ((u64)h3 * s2) +
 76		     ((u64)h4 * s1);
 77		d1 = ((u64)h0 * r1) + ((u64)h1 * r0) +
 78		     ((u64)h2 * s4) + ((u64)h3 * s3) +
 79		     ((u64)h4 * s2);
 80		d2 = ((u64)h0 * r2) + ((u64)h1 * r1) +
 81		     ((u64)h2 * r0) + ((u64)h3 * s4) +
 82		     ((u64)h4 * s3);
 83		d3 = ((u64)h0 * r3) + ((u64)h1 * r2) +
 84		     ((u64)h2 * r1) + ((u64)h3 * r0) +
 85		     ((u64)h4 * s4);
 86		d4 = ((u64)h0 * r4) + ((u64)h1 * r3) +
 87		     ((u64)h2 * r2) + ((u64)h3 * r1) +
 88		     ((u64)h4 * r0);
 89
 90		/* (partial) h %= p */
 91		c = (u32)(d0 >> 26);
 92		h0 = (u32)d0 & 0x3ffffff;
 93		d1 += c;
 94		c = (u32)(d1 >> 26);
 95		h1 = (u32)d1 & 0x3ffffff;
 96		d2 += c;
 97		c = (u32)(d2 >> 26);
 98		h2 = (u32)d2 & 0x3ffffff;
 99		d3 += c;
100		c = (u32)(d3 >> 26);
101		h3 = (u32)d3 & 0x3ffffff;
102		d4 += c;
103		c = (u32)(d4 >> 26);
104		h4 = (u32)d4 & 0x3ffffff;
105		h0 += c * 5;
106		c = (h0 >> 26);
107		h0 = h0 & 0x3ffffff;
108		h1 += c;
109
110		input += POLY1305_BLOCK_SIZE;
111	} while (--nblocks);
112
113	state->h[0] = h0;
114	state->h[1] = h1;
115	state->h[2] = h2;
116	state->h[3] = h3;
117	state->h[4] = h4;
118}
119EXPORT_SYMBOL(poly1305_core_blocks);
120
121void poly1305_core_emit(const struct poly1305_state *state, const u32 nonce[4],
122			void *dst)
123{
124	u8 *mac = dst;
125	u32 h0, h1, h2, h3, h4, c;
126	u32 g0, g1, g2, g3, g4;
127	u64 f;
128	u32 mask;
129
130	/* fully carry h */
131	h0 = state->h[0];
132	h1 = state->h[1];
133	h2 = state->h[2];
134	h3 = state->h[3];
135	h4 = state->h[4];
136
137	c = h1 >> 26;
138	h1 = h1 & 0x3ffffff;
139	h2 += c;
140	c = h2 >> 26;
141	h2 = h2 & 0x3ffffff;
142	h3 += c;
143	c = h3 >> 26;
144	h3 = h3 & 0x3ffffff;
145	h4 += c;
146	c = h4 >> 26;
147	h4 = h4 & 0x3ffffff;
148	h0 += c * 5;
149	c = h0 >> 26;
150	h0 = h0 & 0x3ffffff;
151	h1 += c;
152
153	/* compute h + -p */
154	g0 = h0 + 5;
155	c = g0 >> 26;
156	g0 &= 0x3ffffff;
157	g1 = h1 + c;
158	c = g1 >> 26;
159	g1 &= 0x3ffffff;
160	g2 = h2 + c;
161	c = g2 >> 26;
162	g2 &= 0x3ffffff;
163	g3 = h3 + c;
164	c = g3 >> 26;
165	g3 &= 0x3ffffff;
166	g4 = h4 + c - (1UL << 26);
167
168	/* select h if h < p, or h + -p if h >= p */
169	mask = (g4 >> ((sizeof(u32) * 8) - 1)) - 1;
170	g0 &= mask;
171	g1 &= mask;
172	g2 &= mask;
173	g3 &= mask;
174	g4 &= mask;
175	mask = ~mask;
176
177	h0 = (h0 & mask) | g0;
178	h1 = (h1 & mask) | g1;
179	h2 = (h2 & mask) | g2;
180	h3 = (h3 & mask) | g3;
181	h4 = (h4 & mask) | g4;
182
183	/* h = h % (2^128) */
184	h0 = ((h0) | (h1 << 26)) & 0xffffffff;
185	h1 = ((h1 >> 6) | (h2 << 20)) & 0xffffffff;
186	h2 = ((h2 >> 12) | (h3 << 14)) & 0xffffffff;
187	h3 = ((h3 >> 18) | (h4 << 8)) & 0xffffffff;
188
189	if (likely(nonce)) {
190		/* mac = (h + nonce) % (2^128) */
191		f = (u64)h0 + nonce[0];
192		h0 = (u32)f;
193		f = (u64)h1 + nonce[1] + (f >> 32);
194		h1 = (u32)f;
195		f = (u64)h2 + nonce[2] + (f >> 32);
196		h2 = (u32)f;
197		f = (u64)h3 + nonce[3] + (f >> 32);
198		h3 = (u32)f;
199	}
200
201	put_unaligned_le32(h0, &mac[0]);
202	put_unaligned_le32(h1, &mac[4]);
203	put_unaligned_le32(h2, &mac[8]);
204	put_unaligned_le32(h3, &mac[12]);
205}
206EXPORT_SYMBOL(poly1305_core_emit);