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
 13typedef __uint128_t u128;
 14
 15void poly1305_core_setkey(struct poly1305_core_key *key, const u8 raw_key[16])
 16{
 17	u64 t0, t1;
 18
 19	/* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
 20	t0 = get_unaligned_le64(&raw_key[0]);
 21	t1 = get_unaligned_le64(&raw_key[8]);
 22
 23	key->key.r64[0] = t0 & 0xffc0fffffffULL;
 24	key->key.r64[1] = ((t0 >> 44) | (t1 << 20)) & 0xfffffc0ffffULL;
 25	key->key.r64[2] = ((t1 >> 24)) & 0x00ffffffc0fULL;
 26
 27	/* s = 20*r */
 28	key->precomputed_s.r64[0] = key->key.r64[1] * 20;
 29	key->precomputed_s.r64[1] = key->key.r64[2] * 20;
 30}
 31EXPORT_SYMBOL(poly1305_core_setkey);
 32
 33void poly1305_core_blocks(struct poly1305_state *state,
 34			  const struct poly1305_core_key *key, const void *src,
 35			  unsigned int nblocks, u32 hibit)
 36{
 37	const u8 *input = src;
 38	u64 hibit64;
 39	u64 r0, r1, r2;
 40	u64 s1, s2;
 41	u64 h0, h1, h2;
 42	u64 c;
 43	u128 d0, d1, d2, d;
 44
 45	if (!nblocks)
 46		return;
 47
 48	hibit64 = ((u64)hibit) << 40;
 49
 50	r0 = key->key.r64[0];
 51	r1 = key->key.r64[1];
 52	r2 = key->key.r64[2];
 53
 54	h0 = state->h64[0];
 55	h1 = state->h64[1];
 56	h2 = state->h64[2];
 57
 58	s1 = key->precomputed_s.r64[0];
 59	s2 = key->precomputed_s.r64[1];
 60
 61	do {
 62		u64 t0, t1;
 63
 64		/* h += m[i] */
 65		t0 = get_unaligned_le64(&input[0]);
 66		t1 = get_unaligned_le64(&input[8]);
 67
 68		h0 += t0 & 0xfffffffffffULL;
 69		h1 += ((t0 >> 44) | (t1 << 20)) & 0xfffffffffffULL;
 70		h2 += (((t1 >> 24)) & 0x3ffffffffffULL) | hibit64;
 71
 72		/* h *= r */
 73		d0 = (u128)h0 * r0;
 74		d = (u128)h1 * s2;
 75		d0 += d;
 76		d = (u128)h2 * s1;
 77		d0 += d;
 78		d1 = (u128)h0 * r1;
 79		d = (u128)h1 * r0;
 80		d1 += d;
 81		d = (u128)h2 * s2;
 82		d1 += d;
 83		d2 = (u128)h0 * r2;
 84		d = (u128)h1 * r1;
 85		d2 += d;
 86		d = (u128)h2 * r0;
 87		d2 += d;
 88
 89		/* (partial) h %= p */
 90		c = (u64)(d0 >> 44);
 91		h0 = (u64)d0 & 0xfffffffffffULL;
 92		d1 += c;
 93		c = (u64)(d1 >> 44);
 94		h1 = (u64)d1 & 0xfffffffffffULL;
 95		d2 += c;
 96		c = (u64)(d2 >> 42);
 97		h2 = (u64)d2 & 0x3ffffffffffULL;
 98		h0 += c * 5;
 99		c = h0 >> 44;
100		h0 = h0 & 0xfffffffffffULL;
101		h1 += c;
102
103		input += POLY1305_BLOCK_SIZE;
104	} while (--nblocks);
105
106	state->h64[0] = h0;
107	state->h64[1] = h1;
108	state->h64[2] = h2;
109}
110EXPORT_SYMBOL(poly1305_core_blocks);
111
112void poly1305_core_emit(const struct poly1305_state *state, const u32 nonce[4],
113			void *dst)
114{
115	u8 *mac = dst;
116	u64 h0, h1, h2, c;
117	u64 g0, g1, g2;
118	u64 t0, t1;
119
120	/* fully carry h */
121	h0 = state->h64[0];
122	h1 = state->h64[1];
123	h2 = state->h64[2];
124
125	c = h1 >> 44;
126	h1 &= 0xfffffffffffULL;
127	h2 += c;
128	c = h2 >> 42;
129	h2 &= 0x3ffffffffffULL;
130	h0 += c * 5;
131	c = h0 >> 44;
132	h0 &= 0xfffffffffffULL;
133	h1 += c;
134	c = h1 >> 44;
135	h1 &= 0xfffffffffffULL;
136	h2 += c;
137	c = h2 >> 42;
138	h2 &= 0x3ffffffffffULL;
139	h0 += c * 5;
140	c = h0 >> 44;
141	h0 &= 0xfffffffffffULL;
142	h1 += c;
143
144	/* compute h + -p */
145	g0 = h0 + 5;
146	c  = g0 >> 44;
147	g0 &= 0xfffffffffffULL;
148	g1 = h1 + c;
149	c  = g1 >> 44;
150	g1 &= 0xfffffffffffULL;
151	g2 = h2 + c - (1ULL << 42);
152
153	/* select h if h < p, or h + -p if h >= p */
154	c = (g2 >> ((sizeof(u64) * 8) - 1)) - 1;
155	g0 &= c;
156	g1 &= c;
157	g2 &= c;
158	c  = ~c;
159	h0 = (h0 & c) | g0;
160	h1 = (h1 & c) | g1;
161	h2 = (h2 & c) | g2;
162
163	if (likely(nonce)) {
164		/* h = (h + nonce) */
165		t0 = ((u64)nonce[1] << 32) | nonce[0];
166		t1 = ((u64)nonce[3] << 32) | nonce[2];
167
168		h0 += t0 & 0xfffffffffffULL;
169		c = h0 >> 44;
170		h0 &= 0xfffffffffffULL;
171		h1 += (((t0 >> 44) | (t1 << 20)) & 0xfffffffffffULL) + c;
172		c = h1 >> 44;
173		h1 &= 0xfffffffffffULL;
174		h2 += (((t1 >> 24)) & 0x3ffffffffffULL) + c;
175		h2 &= 0x3ffffffffffULL;
176	}
177
178	/* mac = h % (2^128) */
179	h0 = h0 | (h1 << 44);
180	h1 = (h1 >> 20) | (h2 << 24);
181
182	put_unaligned_le64(h0, &mac[0]);
183	put_unaligned_le64(h1, &mac[8]);
184}
185EXPORT_SYMBOL(poly1305_core_emit);