lib/crypto/poly1305-donna32.c

0001 // SPDX-License-Identifier: GPL-2.0 OR MIT
0002 /*
0003  * Copyright (C) 2015-2019 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved.
0004  *
0005  * This is based in part on Andrew Moon's poly1305-donna, which is in the
0006  * public domain.
0007  */
0008
0009 #include <linux/kernel.h>
0010 #include <asm/unaligned.h>
0011 #include <crypto/internal/poly1305.h>
0012
0013 void poly1305_core_setkey(struct poly1305_core_key *key,
0014               const u8 raw_key[POLY1305_BLOCK_SIZE])
0015 {
0016     /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
0017     key->key.r[0] = (get_unaligned_le32(&raw_key[0])) & 0x3ffffff;
0018     key->key.r[1] = (get_unaligned_le32(&raw_key[3]) >> 2) & 0x3ffff03;
0019     key->key.r[2] = (get_unaligned_le32(&raw_key[6]) >> 4) & 0x3ffc0ff;
0020     key->key.r[3] = (get_unaligned_le32(&raw_key[9]) >> 6) & 0x3f03fff;
0021     key->key.r[4] = (get_unaligned_le32(&raw_key[12]) >> 8) & 0x00fffff;
0022
0023     /* s = 5*r */
0024     key->precomputed_s.r[0] = key->key.r[1] * 5;
0025     key->precomputed_s.r[1] = key->key.r[2] * 5;
0026     key->precomputed_s.r[2] = key->key.r[3] * 5;
0027     key->precomputed_s.r[3] = key->key.r[4] * 5;
0028 }
0029 EXPORT_SYMBOL(poly1305_core_setkey);
0030
0031 void poly1305_core_blocks(struct poly1305_state *state,
0032               const struct poly1305_core_key *key, const void *src,
0033               unsigned int nblocks, u32 hibit)
0034 {
0035     const u8 *input = src;
0036     u32 r0, r1, r2, r3, r4;
0037     u32 s1, s2, s3, s4;
0038     u32 h0, h1, h2, h3, h4;
0039     u64 d0, d1, d2, d3, d4;
0040     u32 c;
0041
0042     if (!nblocks)
0043         return;
0044
0045     hibit <<= 24;
0046
0047     r0 = key->key.r[0];
0048     r1 = key->key.r[1];
0049     r2 = key->key.r[2];
0050     r3 = key->key.r[3];
0051     r4 = key->key.r[4];
0052
0053     s1 = key->precomputed_s.r[0];
0054     s2 = key->precomputed_s.r[1];
0055     s3 = key->precomputed_s.r[2];
0056     s4 = key->precomputed_s.r[3];
0057
0058     h0 = state->h[0];
0059     h1 = state->h[1];
0060     h2 = state->h[2];
0061     h3 = state->h[3];
0062     h4 = state->h[4];
0063
0064     do {
0065         /* h += m[i] */
0066         h0 += (get_unaligned_le32(&input[0])) & 0x3ffffff;
0067         h1 += (get_unaligned_le32(&input[3]) >> 2) & 0x3ffffff;
0068         h2 += (get_unaligned_le32(&input[6]) >> 4) & 0x3ffffff;
0069         h3 += (get_unaligned_le32(&input[9]) >> 6) & 0x3ffffff;
0070         h4 += (get_unaligned_le32(&input[12]) >> 8) | hibit;
0071
0072         /* h *= r */
0073         d0 = ((u64)h0 * r0) + ((u64)h1 * s4) +
0074              ((u64)h2 * s3) + ((u64)h3 * s2) +
0075              ((u64)h4 * s1);
0076         d1 = ((u64)h0 * r1) + ((u64)h1 * r0) +
0077              ((u64)h2 * s4) + ((u64)h3 * s3) +
0078              ((u64)h4 * s2);
0079         d2 = ((u64)h0 * r2) + ((u64)h1 * r1) +
0080              ((u64)h2 * r0) + ((u64)h3 * s4) +
0081              ((u64)h4 * s3);
0082         d3 = ((u64)h0 * r3) + ((u64)h1 * r2) +
0083              ((u64)h2 * r1) + ((u64)h3 * r0) +
0084              ((u64)h4 * s4);
0085         d4 = ((u64)h0 * r4) + ((u64)h1 * r3) +
0086              ((u64)h2 * r2) + ((u64)h3 * r1) +
0087              ((u64)h4 * r0);
0088
0089         /* (partial) h %= p */
0090         c = (u32)(d0 >> 26);
0091         h0 = (u32)d0 & 0x3ffffff;
0092         d1 += c;
0093         c = (u32)(d1 >> 26);
0094         h1 = (u32)d1 & 0x3ffffff;
0095         d2 += c;
0096         c = (u32)(d2 >> 26);
0097         h2 = (u32)d2 & 0x3ffffff;
0098         d3 += c;
0099         c = (u32)(d3 >> 26);
0100         h3 = (u32)d3 & 0x3ffffff;
0101         d4 += c;
0102         c = (u32)(d4 >> 26);
0103         h4 = (u32)d4 & 0x3ffffff;
0104         h0 += c * 5;
0105         c = (h0 >> 26);
0106         h0 = h0 & 0x3ffffff;
0107         h1 += c;
0108
0109         input += POLY1305_BLOCK_SIZE;
0110     } while (--nblocks);
0111
0112     state->h[0] = h0;
0113     state->h[1] = h1;
0114     state->h[2] = h2;
0115     state->h[3] = h3;
0116     state->h[4] = h4;
0117 }
0118 EXPORT_SYMBOL(poly1305_core_blocks);
0119
0120 void poly1305_core_emit(const struct poly1305_state *state, const u32 nonce[4],
0121             void *dst)
0122 {
0123     u8 *mac = dst;
0124     u32 h0, h1, h2, h3, h4, c;
0125     u32 g0, g1, g2, g3, g4;
0126     u64 f;
0127     u32 mask;
0128
0129     /* fully carry h */
0130     h0 = state->h[0];
0131     h1 = state->h[1];
0132     h2 = state->h[2];
0133     h3 = state->h[3];
0134     h4 = state->h[4];
0135
0136     c = h1 >> 26;
0137     h1 = h1 & 0x3ffffff;
0138     h2 += c;
0139     c = h2 >> 26;
0140     h2 = h2 & 0x3ffffff;
0141     h3 += c;
0142     c = h3 >> 26;
0143     h3 = h3 & 0x3ffffff;
0144     h4 += c;
0145     c = h4 >> 26;
0146     h4 = h4 & 0x3ffffff;
0147     h0 += c * 5;
0148     c = h0 >> 26;
0149     h0 = h0 & 0x3ffffff;
0150     h1 += c;
0151
0152     /* compute h + -p */
0153     g0 = h0 + 5;
0154     c = g0 >> 26;
0155     g0 &= 0x3ffffff;
0156     g1 = h1 + c;
0157     c = g1 >> 26;
0158     g1 &= 0x3ffffff;
0159     g2 = h2 + c;
0160     c = g2 >> 26;
0161     g2 &= 0x3ffffff;
0162     g3 = h3 + c;
0163     c = g3 >> 26;
0164     g3 &= 0x3ffffff;
0165     g4 = h4 + c - (1UL << 26);
0166
0167     /* select h if h < p, or h + -p if h >= p */
0168     mask = (g4 >> ((sizeof(u32) * 8) - 1)) - 1;
0169     g0 &= mask;
0170     g1 &= mask;
0171     g2 &= mask;
0172     g3 &= mask;
0173     g4 &= mask;
0174     mask = ~mask;
0175
0176     h0 = (h0 & mask) | g0;
0177     h1 = (h1 & mask) | g1;
0178     h2 = (h2 & mask) | g2;
0179     h3 = (h3 & mask) | g3;
0180     h4 = (h4 & mask) | g4;
0181
0182     /* h = h % (2^128) */
0183     h0 = ((h0) | (h1 << 26)) & 0xffffffff;
0184     h1 = ((h1 >> 6) | (h2 << 20)) & 0xffffffff;
0185     h2 = ((h2 >> 12) | (h3 << 14)) & 0xffffffff;
0186     h3 = ((h3 >> 18) | (h4 << 8)) & 0xffffffff;
0187
0188     if (likely(nonce)) {
0189         /* mac = (h + nonce) % (2^128) */
0190         f = (u64)h0 + nonce[0];
0191         h0 = (u32)f;
0192         f = (u64)h1 + nonce[1] + (f >> 32);
0193         h1 = (u32)f;
0194         f = (u64)h2 + nonce[2] + (f >> 32);
0195         h2 = (u32)f;
0196         f = (u64)h3 + nonce[3] + (f >> 32);
0197         h3 = (u32)f;
0198     }
0199
0200     put_unaligned_le32(h0, &mac[0]);
0201     put_unaligned_le32(h1, &mac[4]);
0202     put_unaligned_le32(h2, &mac[8]);
0203     put_unaligned_le32(h3, &mac[12]);
0204 }
0205 EXPORT_SYMBOL(poly1305_core_emit);