162306a36Sopenharmony_ci// SPDX-License-Identifier: GPL-2.0 OR MIT
262306a36Sopenharmony_ci/*
362306a36Sopenharmony_ci * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
462306a36Sopenharmony_ci *
562306a36Sopenharmony_ci * This is based in part on Andrew Moon's poly1305-donna, which is in the
662306a36Sopenharmony_ci * public domain.
762306a36Sopenharmony_ci */
862306a36Sopenharmony_ci
962306a36Sopenharmony_ci#include <linux/kernel.h>
1062306a36Sopenharmony_ci#include <asm/unaligned.h>
1162306a36Sopenharmony_ci#include <crypto/internal/poly1305.h>
1262306a36Sopenharmony_ci
1362306a36Sopenharmony_civoid poly1305_core_setkey(struct poly1305_core_key *key,
1462306a36Sopenharmony_ci			  const u8 raw_key[POLY1305_BLOCK_SIZE])
1562306a36Sopenharmony_ci{
1662306a36Sopenharmony_ci	/* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
1762306a36Sopenharmony_ci	key->key.r[0] = (get_unaligned_le32(&raw_key[0])) & 0x3ffffff;
1862306a36Sopenharmony_ci	key->key.r[1] = (get_unaligned_le32(&raw_key[3]) >> 2) & 0x3ffff03;
1962306a36Sopenharmony_ci	key->key.r[2] = (get_unaligned_le32(&raw_key[6]) >> 4) & 0x3ffc0ff;
2062306a36Sopenharmony_ci	key->key.r[3] = (get_unaligned_le32(&raw_key[9]) >> 6) & 0x3f03fff;
2162306a36Sopenharmony_ci	key->key.r[4] = (get_unaligned_le32(&raw_key[12]) >> 8) & 0x00fffff;
2262306a36Sopenharmony_ci
2362306a36Sopenharmony_ci	/* s = 5*r */
2462306a36Sopenharmony_ci	key->precomputed_s.r[0] = key->key.r[1] * 5;
2562306a36Sopenharmony_ci	key->precomputed_s.r[1] = key->key.r[2] * 5;
2662306a36Sopenharmony_ci	key->precomputed_s.r[2] = key->key.r[3] * 5;
2762306a36Sopenharmony_ci	key->precomputed_s.r[3] = key->key.r[4] * 5;
2862306a36Sopenharmony_ci}
2962306a36Sopenharmony_ciEXPORT_SYMBOL(poly1305_core_setkey);
3062306a36Sopenharmony_ci
3162306a36Sopenharmony_civoid poly1305_core_blocks(struct poly1305_state *state,
3262306a36Sopenharmony_ci			  const struct poly1305_core_key *key, const void *src,
3362306a36Sopenharmony_ci			  unsigned int nblocks, u32 hibit)
3462306a36Sopenharmony_ci{
3562306a36Sopenharmony_ci	const u8 *input = src;
3662306a36Sopenharmony_ci	u32 r0, r1, r2, r3, r4;
3762306a36Sopenharmony_ci	u32 s1, s2, s3, s4;
3862306a36Sopenharmony_ci	u32 h0, h1, h2, h3, h4;
3962306a36Sopenharmony_ci	u64 d0, d1, d2, d3, d4;
4062306a36Sopenharmony_ci	u32 c;
4162306a36Sopenharmony_ci
4262306a36Sopenharmony_ci	if (!nblocks)
4362306a36Sopenharmony_ci		return;
4462306a36Sopenharmony_ci
4562306a36Sopenharmony_ci	hibit <<= 24;
4662306a36Sopenharmony_ci
4762306a36Sopenharmony_ci	r0 = key->key.r[0];
4862306a36Sopenharmony_ci	r1 = key->key.r[1];
4962306a36Sopenharmony_ci	r2 = key->key.r[2];
5062306a36Sopenharmony_ci	r3 = key->key.r[3];
5162306a36Sopenharmony_ci	r4 = key->key.r[4];
5262306a36Sopenharmony_ci
5362306a36Sopenharmony_ci	s1 = key->precomputed_s.r[0];
5462306a36Sopenharmony_ci	s2 = key->precomputed_s.r[1];
5562306a36Sopenharmony_ci	s3 = key->precomputed_s.r[2];
5662306a36Sopenharmony_ci	s4 = key->precomputed_s.r[3];
5762306a36Sopenharmony_ci
5862306a36Sopenharmony_ci	h0 = state->h[0];
5962306a36Sopenharmony_ci	h1 = state->h[1];
6062306a36Sopenharmony_ci	h2 = state->h[2];
6162306a36Sopenharmony_ci	h3 = state->h[3];
6262306a36Sopenharmony_ci	h4 = state->h[4];
6362306a36Sopenharmony_ci
6462306a36Sopenharmony_ci	do {
6562306a36Sopenharmony_ci		/* h += m[i] */
6662306a36Sopenharmony_ci		h0 += (get_unaligned_le32(&input[0])) & 0x3ffffff;
6762306a36Sopenharmony_ci		h1 += (get_unaligned_le32(&input[3]) >> 2) & 0x3ffffff;
6862306a36Sopenharmony_ci		h2 += (get_unaligned_le32(&input[6]) >> 4) & 0x3ffffff;
6962306a36Sopenharmony_ci		h3 += (get_unaligned_le32(&input[9]) >> 6) & 0x3ffffff;
7062306a36Sopenharmony_ci		h4 += (get_unaligned_le32(&input[12]) >> 8) | hibit;
7162306a36Sopenharmony_ci
7262306a36Sopenharmony_ci		/* h *= r */
7362306a36Sopenharmony_ci		d0 = ((u64)h0 * r0) + ((u64)h1 * s4) +
7462306a36Sopenharmony_ci		     ((u64)h2 * s3) + ((u64)h3 * s2) +
7562306a36Sopenharmony_ci		     ((u64)h4 * s1);
7662306a36Sopenharmony_ci		d1 = ((u64)h0 * r1) + ((u64)h1 * r0) +
7762306a36Sopenharmony_ci		     ((u64)h2 * s4) + ((u64)h3 * s3) +
7862306a36Sopenharmony_ci		     ((u64)h4 * s2);
7962306a36Sopenharmony_ci		d2 = ((u64)h0 * r2) + ((u64)h1 * r1) +
8062306a36Sopenharmony_ci		     ((u64)h2 * r0) + ((u64)h3 * s4) +
8162306a36Sopenharmony_ci		     ((u64)h4 * s3);
8262306a36Sopenharmony_ci		d3 = ((u64)h0 * r3) + ((u64)h1 * r2) +
8362306a36Sopenharmony_ci		     ((u64)h2 * r1) + ((u64)h3 * r0) +
8462306a36Sopenharmony_ci		     ((u64)h4 * s4);
8562306a36Sopenharmony_ci		d4 = ((u64)h0 * r4) + ((u64)h1 * r3) +
8662306a36Sopenharmony_ci		     ((u64)h2 * r2) + ((u64)h3 * r1) +
8762306a36Sopenharmony_ci		     ((u64)h4 * r0);
8862306a36Sopenharmony_ci
8962306a36Sopenharmony_ci		/* (partial) h %= p */
9062306a36Sopenharmony_ci		c = (u32)(d0 >> 26);
9162306a36Sopenharmony_ci		h0 = (u32)d0 & 0x3ffffff;
9262306a36Sopenharmony_ci		d1 += c;
9362306a36Sopenharmony_ci		c = (u32)(d1 >> 26);
9462306a36Sopenharmony_ci		h1 = (u32)d1 & 0x3ffffff;
9562306a36Sopenharmony_ci		d2 += c;
9662306a36Sopenharmony_ci		c = (u32)(d2 >> 26);
9762306a36Sopenharmony_ci		h2 = (u32)d2 & 0x3ffffff;
9862306a36Sopenharmony_ci		d3 += c;
9962306a36Sopenharmony_ci		c = (u32)(d3 >> 26);
10062306a36Sopenharmony_ci		h3 = (u32)d3 & 0x3ffffff;
10162306a36Sopenharmony_ci		d4 += c;
10262306a36Sopenharmony_ci		c = (u32)(d4 >> 26);
10362306a36Sopenharmony_ci		h4 = (u32)d4 & 0x3ffffff;
10462306a36Sopenharmony_ci		h0 += c * 5;
10562306a36Sopenharmony_ci		c = (h0 >> 26);
10662306a36Sopenharmony_ci		h0 = h0 & 0x3ffffff;
10762306a36Sopenharmony_ci		h1 += c;
10862306a36Sopenharmony_ci
10962306a36Sopenharmony_ci		input += POLY1305_BLOCK_SIZE;
11062306a36Sopenharmony_ci	} while (--nblocks);
11162306a36Sopenharmony_ci
11262306a36Sopenharmony_ci	state->h[0] = h0;
11362306a36Sopenharmony_ci	state->h[1] = h1;
11462306a36Sopenharmony_ci	state->h[2] = h2;
11562306a36Sopenharmony_ci	state->h[3] = h3;
11662306a36Sopenharmony_ci	state->h[4] = h4;
11762306a36Sopenharmony_ci}
11862306a36Sopenharmony_ciEXPORT_SYMBOL(poly1305_core_blocks);
11962306a36Sopenharmony_ci
12062306a36Sopenharmony_civoid poly1305_core_emit(const struct poly1305_state *state, const u32 nonce[4],
12162306a36Sopenharmony_ci			void *dst)
12262306a36Sopenharmony_ci{
12362306a36Sopenharmony_ci	u8 *mac = dst;
12462306a36Sopenharmony_ci	u32 h0, h1, h2, h3, h4, c;
12562306a36Sopenharmony_ci	u32 g0, g1, g2, g3, g4;
12662306a36Sopenharmony_ci	u64 f;
12762306a36Sopenharmony_ci	u32 mask;
12862306a36Sopenharmony_ci
12962306a36Sopenharmony_ci	/* fully carry h */
13062306a36Sopenharmony_ci	h0 = state->h[0];
13162306a36Sopenharmony_ci	h1 = state->h[1];
13262306a36Sopenharmony_ci	h2 = state->h[2];
13362306a36Sopenharmony_ci	h3 = state->h[3];
13462306a36Sopenharmony_ci	h4 = state->h[4];
13562306a36Sopenharmony_ci
13662306a36Sopenharmony_ci	c = h1 >> 26;
13762306a36Sopenharmony_ci	h1 = h1 & 0x3ffffff;
13862306a36Sopenharmony_ci	h2 += c;
13962306a36Sopenharmony_ci	c = h2 >> 26;
14062306a36Sopenharmony_ci	h2 = h2 & 0x3ffffff;
14162306a36Sopenharmony_ci	h3 += c;
14262306a36Sopenharmony_ci	c = h3 >> 26;
14362306a36Sopenharmony_ci	h3 = h3 & 0x3ffffff;
14462306a36Sopenharmony_ci	h4 += c;
14562306a36Sopenharmony_ci	c = h4 >> 26;
14662306a36Sopenharmony_ci	h4 = h4 & 0x3ffffff;
14762306a36Sopenharmony_ci	h0 += c * 5;
14862306a36Sopenharmony_ci	c = h0 >> 26;
14962306a36Sopenharmony_ci	h0 = h0 & 0x3ffffff;
15062306a36Sopenharmony_ci	h1 += c;
15162306a36Sopenharmony_ci
15262306a36Sopenharmony_ci	/* compute h + -p */
15362306a36Sopenharmony_ci	g0 = h0 + 5;
15462306a36Sopenharmony_ci	c = g0 >> 26;
15562306a36Sopenharmony_ci	g0 &= 0x3ffffff;
15662306a36Sopenharmony_ci	g1 = h1 + c;
15762306a36Sopenharmony_ci	c = g1 >> 26;
15862306a36Sopenharmony_ci	g1 &= 0x3ffffff;
15962306a36Sopenharmony_ci	g2 = h2 + c;
16062306a36Sopenharmony_ci	c = g2 >> 26;
16162306a36Sopenharmony_ci	g2 &= 0x3ffffff;
16262306a36Sopenharmony_ci	g3 = h3 + c;
16362306a36Sopenharmony_ci	c = g3 >> 26;
16462306a36Sopenharmony_ci	g3 &= 0x3ffffff;
16562306a36Sopenharmony_ci	g4 = h4 + c - (1UL << 26);
16662306a36Sopenharmony_ci
16762306a36Sopenharmony_ci	/* select h if h < p, or h + -p if h >= p */
16862306a36Sopenharmony_ci	mask = (g4 >> ((sizeof(u32) * 8) - 1)) - 1;
16962306a36Sopenharmony_ci	g0 &= mask;
17062306a36Sopenharmony_ci	g1 &= mask;
17162306a36Sopenharmony_ci	g2 &= mask;
17262306a36Sopenharmony_ci	g3 &= mask;
17362306a36Sopenharmony_ci	g4 &= mask;
17462306a36Sopenharmony_ci	mask = ~mask;
17562306a36Sopenharmony_ci
17662306a36Sopenharmony_ci	h0 = (h0 & mask) | g0;
17762306a36Sopenharmony_ci	h1 = (h1 & mask) | g1;
17862306a36Sopenharmony_ci	h2 = (h2 & mask) | g2;
17962306a36Sopenharmony_ci	h3 = (h3 & mask) | g3;
18062306a36Sopenharmony_ci	h4 = (h4 & mask) | g4;
18162306a36Sopenharmony_ci
18262306a36Sopenharmony_ci	/* h = h % (2^128) */
18362306a36Sopenharmony_ci	h0 = ((h0) | (h1 << 26)) & 0xffffffff;
18462306a36Sopenharmony_ci	h1 = ((h1 >> 6) | (h2 << 20)) & 0xffffffff;
18562306a36Sopenharmony_ci	h2 = ((h2 >> 12) | (h3 << 14)) & 0xffffffff;
18662306a36Sopenharmony_ci	h3 = ((h3 >> 18) | (h4 << 8)) & 0xffffffff;
18762306a36Sopenharmony_ci
18862306a36Sopenharmony_ci	if (likely(nonce)) {
18962306a36Sopenharmony_ci		/* mac = (h + nonce) % (2^128) */
19062306a36Sopenharmony_ci		f = (u64)h0 + nonce[0];
19162306a36Sopenharmony_ci		h0 = (u32)f;
19262306a36Sopenharmony_ci		f = (u64)h1 + nonce[1] + (f >> 32);
19362306a36Sopenharmony_ci		h1 = (u32)f;
19462306a36Sopenharmony_ci		f = (u64)h2 + nonce[2] + (f >> 32);
19562306a36Sopenharmony_ci		h2 = (u32)f;
19662306a36Sopenharmony_ci		f = (u64)h3 + nonce[3] + (f >> 32);
19762306a36Sopenharmony_ci		h3 = (u32)f;
19862306a36Sopenharmony_ci	}
19962306a36Sopenharmony_ci
20062306a36Sopenharmony_ci	put_unaligned_le32(h0, &mac[0]);
20162306a36Sopenharmony_ci	put_unaligned_le32(h1, &mac[4]);
20262306a36Sopenharmony_ci	put_unaligned_le32(h2, &mac[8]);
20362306a36Sopenharmony_ci	put_unaligned_le32(h3, &mac[12]);
20462306a36Sopenharmony_ci}
20562306a36Sopenharmony_ciEXPORT_SYMBOL(poly1305_core_emit);
206