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 <linux/kernel.h>
 10#include <asm/unaligned.h>
 11#include <crypto/internal/poly1305.h>
 12
 13void poly1305_core_setkey(struct poly1305_core_key *key,
 14			  const u8 raw_key[POLY1305_BLOCK_SIZE])
 15{
 16	u64 t0, t1;
 17
 18	/* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 19	t0 = get_unaligned_le64(&raw_key[0]);
 20	t1 = get_unaligned_le64(&raw_key[8]);
 21
 22	key->key.r64[0] = t0 & 0xffc0fffffffULL;
 23	key->key.r64[1] = ((t0 >> 44) | (t1 << 20)) & 0xfffffc0ffffULL;
 24	key->key.r64[2] = ((t1 >> 24)) & 0x00ffffffc0fULL;
 25
 26	/* s = 20*r */
 27	key->precomputed_s.r64[0] = key->key.r64[1] * 20;
 28	key->precomputed_s.r64[1] = key->key.r64[2] * 20;
 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	u64 hibit64;
 38	u64 r0, r1, r2;
 39	u64 s1, s2;
 40	u64 h0, h1, h2;
 41	u64 c;
 42	u128 d0, d1, d2, d;
 43
 44	if (!nblocks)
 45		return;
 46
 47	hibit64 = ((u64)hibit) << 40;
 48
 49	r0 = key->key.r64[0];
 50	r1 = key->key.r64[1];
 51	r2 = key->key.r64[2];
 52
 53	h0 = state->h64[0];
 54	h1 = state->h64[1];
 55	h2 = state->h64[2];
 56
 57	s1 = key->precomputed_s.r64[0];
 58	s2 = key->precomputed_s.r64[1];
 59
 60	do {
 61		u64 t0, t1;
 62
 63		/* h += m[i] */
 64		t0 = get_unaligned_le64(&input[0]);
 65		t1 = get_unaligned_le64(&input[8]);
 66
 67		h0 += t0 & 0xfffffffffffULL;
 68		h1 += ((t0 >> 44) | (t1 << 20)) & 0xfffffffffffULL;
 69		h2 += (((t1 >> 24)) & 0x3ffffffffffULL) | hibit64;
 70
 71		/* h *= r */
 72		d0 = (u128)h0 * r0;
 73		d = (u128)h1 * s2;
 74		d0 += d;
 75		d = (u128)h2 * s1;
 76		d0 += d;
 77		d1 = (u128)h0 * r1;
 78		d = (u128)h1 * r0;
 79		d1 += d;
 80		d = (u128)h2 * s2;
 81		d1 += d;
 82		d2 = (u128)h0 * r2;
 83		d = (u128)h1 * r1;
 84		d2 += d;
 85		d = (u128)h2 * r0;
 86		d2 += d;
 87
 88		/* (partial) h %= p */
 89		c = (u64)(d0 >> 44);
 90		h0 = (u64)d0 & 0xfffffffffffULL;
 91		d1 += c;
 92		c = (u64)(d1 >> 44);
 93		h1 = (u64)d1 & 0xfffffffffffULL;
 94		d2 += c;
 95		c = (u64)(d2 >> 42);
 96		h2 = (u64)d2 & 0x3ffffffffffULL;
 97		h0 += c * 5;
 98		c = h0 >> 44;
 99		h0 = h0 & 0xfffffffffffULL;
100		h1 += c;
101
102		input += POLY1305_BLOCK_SIZE;
103	} while (--nblocks);
104
105	state->h64[0] = h0;
106	state->h64[1] = h1;
107	state->h64[2] = h2;
108}
109EXPORT_SYMBOL(poly1305_core_blocks);
110
111void poly1305_core_emit(const struct poly1305_state *state, const u32 nonce[4],
112			void *dst)
113{
114	u8 *mac = dst;
115	u64 h0, h1, h2, c;
116	u64 g0, g1, g2;
117	u64 t0, t1;
118
119	/* fully carry h */
120	h0 = state->h64[0];
121	h1 = state->h64[1];
122	h2 = state->h64[2];
123
124	c = h1 >> 44;
125	h1 &= 0xfffffffffffULL;
126	h2 += c;
127	c = h2 >> 42;
128	h2 &= 0x3ffffffffffULL;
129	h0 += c * 5;
130	c = h0 >> 44;
131	h0 &= 0xfffffffffffULL;
132	h1 += c;
133	c = h1 >> 44;
134	h1 &= 0xfffffffffffULL;
135	h2 += c;
136	c = h2 >> 42;
137	h2 &= 0x3ffffffffffULL;
138	h0 += c * 5;
139	c = h0 >> 44;
140	h0 &= 0xfffffffffffULL;
141	h1 += c;
142
143	/* compute h + -p */
144	g0 = h0 + 5;
145	c  = g0 >> 44;
146	g0 &= 0xfffffffffffULL;
147	g1 = h1 + c;
148	c  = g1 >> 44;
149	g1 &= 0xfffffffffffULL;
150	g2 = h2 + c - (1ULL << 42);
151
152	/* select h if h < p, or h + -p if h >= p */
153	c = (g2 >> ((sizeof(u64) * 8) - 1)) - 1;
154	g0 &= c;
155	g1 &= c;
156	g2 &= c;
157	c  = ~c;
158	h0 = (h0 & c) | g0;
159	h1 = (h1 & c) | g1;
160	h2 = (h2 & c) | g2;
161
162	if (likely(nonce)) {
163		/* h = (h + nonce) */
164		t0 = ((u64)nonce[1] << 32) | nonce[0];
165		t1 = ((u64)nonce[3] << 32) | nonce[2];
166
167		h0 += t0 & 0xfffffffffffULL;
168		c = h0 >> 44;
169		h0 &= 0xfffffffffffULL;
170		h1 += (((t0 >> 44) | (t1 << 20)) & 0xfffffffffffULL) + c;
171		c = h1 >> 44;
172		h1 &= 0xfffffffffffULL;
173		h2 += (((t1 >> 24)) & 0x3ffffffffffULL) + c;
174		h2 &= 0x3ffffffffffULL;
175	}
176
177	/* mac = h % (2^128) */
178	h0 = h0 | (h1 << 44);
179	h1 = (h1 >> 20) | (h2 << 24);
180
181	put_unaligned_le64(h0, &mac[0]);
182	put_unaligned_le64(h1, &mac[8]);
183}
184EXPORT_SYMBOL(poly1305_core_emit);