1// Copyright (c) 2019, Google Inc.
2// Portions Copyright 2020 Brian Smith.
3//
4// Permission to use, copy, modify, and/or distribute this software for any
5// purpose with or without fee is hereby granted, provided that the above
6// copyright notice and this permission notice appear in all copies.
7//
8// THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
9// WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
10// MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY
11// SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
12// WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION
13// OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN
14// CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
15
16// This file is based on BoringSSL's gcm_nohw.c.
17
18// This file contains a constant-time implementation of GHASH based on the notes
19// in https://bearssl.org/constanttime.html#ghash-for-gcm and the reduction
20// algorithm described in
21// https://crypto.stanford.edu/RealWorldCrypto/slides/gueron.pdf.
22//
23// Unlike the BearSSL notes, we use u128 in the 64-bit implementation.
24
25use super::{Block, Xi, BLOCK_LEN};
26use crate::polyfill::ArraySplitMap;
27
28#[cfg(target_pointer_width = "64")]
29fn gcm_mul64_nohw(a: u64, b: u64) -> (u64, u64) {
30 #[allow(clippy::cast_possible_truncation)]
31 #[inline(always)]
32 fn lo(a: u128) -> u64 {
33 a as u64
34 }
35
36 #[inline(always)]
37 fn hi(a: u128) -> u64 {
38 lo(a >> 64)
39 }
40
41 #[inline(always)]
42 fn mul(a: u64, b: u64) -> u128 {
43 u128::from(a) * u128::from(b)
44 }
45
46 // One term every four bits means the largest term is 64/4 = 16, which barely
47 // overflows into the next term. Using one term every five bits would cost 25
48 // multiplications instead of 16. It is faster to mask off the bottom four
49 // bits of |a|, giving a largest term of 60/4 = 15, and apply the bottom bits
50 // separately.
51 let a0 = a & 0x1111111111111110;
52 let a1 = a & 0x2222222222222220;
53 let a2 = a & 0x4444444444444440;
54 let a3 = a & 0x8888888888888880;
55
56 let b0 = b & 0x1111111111111111;
57 let b1 = b & 0x2222222222222222;
58 let b2 = b & 0x4444444444444444;
59 let b3 = b & 0x8888888888888888;
60
61 let c0 = mul(a0, b0) ^ mul(a1, b3) ^ mul(a2, b2) ^ mul(a3, b1);
62 let c1 = mul(a0, b1) ^ mul(a1, b0) ^ mul(a2, b3) ^ mul(a3, b2);
63 let c2 = mul(a0, b2) ^ mul(a1, b1) ^ mul(a2, b0) ^ mul(a3, b3);
64 let c3 = mul(a0, b3) ^ mul(a1, b2) ^ mul(a2, b1) ^ mul(a3, b0);
65
66 // Multiply the bottom four bits of |a| with |b|.
67 let a0_mask = 0u64.wrapping_sub(a & 1);
68 let a1_mask = 0u64.wrapping_sub((a >> 1) & 1);
69 let a2_mask = 0u64.wrapping_sub((a >> 2) & 1);
70 let a3_mask = 0u64.wrapping_sub((a >> 3) & 1);
71 let extra = u128::from(a0_mask & b)
72 ^ (u128::from(a1_mask & b) << 1)
73 ^ (u128::from(a2_mask & b) << 2)
74 ^ (u128::from(a3_mask & b) << 3);
75
76 let lo = (lo(c0) & 0x1111111111111111)
77 ^ (lo(c1) & 0x2222222222222222)
78 ^ (lo(c2) & 0x4444444444444444)
79 ^ (lo(c3) & 0x8888888888888888)
80 ^ lo(extra);
81 let hi = (hi(c0) & 0x1111111111111111)
82 ^ (hi(c1) & 0x2222222222222222)
83 ^ (hi(c2) & 0x4444444444444444)
84 ^ (hi(c3) & 0x8888888888888888)
85 ^ hi(extra);
86 (lo, hi)
87}
88
89#[cfg(not(target_pointer_width = "64"))]
90fn gcm_mul32_nohw(a: u32, b: u32) -> u64 {
91 #[inline(always)]
92 fn mul(a: u32, b: u32) -> u64 {
93 u64::from(a) * u64::from(b)
94 }
95
96 // One term every four bits means the largest term is 32/4 = 8, which does not
97 // overflow into the next term.
98 let a0 = a & 0x11111111;
99 let a1 = a & 0x22222222;
100 let a2 = a & 0x44444444;
101 let a3 = a & 0x88888888;
102
103 let b0 = b & 0x11111111;
104 let b1 = b & 0x22222222;
105 let b2 = b & 0x44444444;
106 let b3 = b & 0x88888888;
107
108 let c0 = mul(a0, b0) ^ mul(a1, b3) ^ mul(a2, b2) ^ mul(a3, b1);
109 let c1 = mul(a0, b1) ^ mul(a1, b0) ^ mul(a2, b3) ^ mul(a3, b2);
110 let c2 = mul(a0, b2) ^ mul(a1, b1) ^ mul(a2, b0) ^ mul(a3, b3);
111 let c3 = mul(a0, b3) ^ mul(a1, b2) ^ mul(a2, b1) ^ mul(a3, b0);
112
113 (c0 & 0x1111111111111111)
114 | (c1 & 0x2222222222222222)
115 | (c2 & 0x4444444444444444)
116 | (c3 & 0x8888888888888888)
117}
118
119#[cfg(not(target_pointer_width = "64"))]
120fn gcm_mul64_nohw(a: u64, b: u64) -> (u64, u64) {
121 #[inline(always)]
122 fn lo(a: u64) -> u32 {
123 a as u32
124 }
125 #[inline(always)]
126 fn hi(a: u64) -> u32 {
127 lo(a >> 32)
128 }
129
130 let a0 = lo(a);
131 let a1 = hi(a);
132 let b0 = lo(b);
133 let b1 = hi(b);
134 // Karatsuba multiplication.
135 let lo = gcm_mul32_nohw(a0, b0);
136 let hi = gcm_mul32_nohw(a1, b1);
137 let mid = gcm_mul32_nohw(a0 ^ a1, b0 ^ b1) ^ lo ^ hi;
138 (lo ^ (mid << 32), hi ^ (mid >> 32))
139}
140
141pub(super) fn init(xi: [u64; 2]) -> super::u128 {
142 // We implement GHASH in terms of POLYVAL, as described in RFC 8452. This
143 // avoids a shift by 1 in the multiplication, needed to account for bit
144 // reversal losing a bit after multiplication, that is,
145 // rev128(X) * rev128(Y) = rev255(X*Y).
146 //
147 // Per Appendix A, we run mulX_POLYVAL. Note this is the same transformation
148 // applied by |gcm_init_clmul|, etc. Note |Xi| has already been byteswapped.
149 //
150 // See also slide 16 of
151 // https://crypto.stanford.edu/RealWorldCrypto/slides/gueron.pdf
152 let mut lo = xi[1];
153 let mut hi = xi[0];
154
155 let mut carry = hi >> 63;
156 carry = 0u64.wrapping_sub(carry);
157
158 hi <<= 1;
159 hi |= lo >> 63;
160 lo <<= 1;
161
162 // The irreducible polynomial is 1 + x^121 + x^126 + x^127 + x^128, so we
163 // conditionally add 0xc200...0001.
164 lo ^= carry & 1;
165 hi ^= carry & 0xc200000000000000;
166
167 // This implementation does not use the rest of |Htable|.
168 super::u128 { hi, lo }
169}
170
171fn gcm_polyval_nohw(xi: &mut [u64; 2], h: super::u128) {
172 // Karatsuba multiplication. The product of |Xi| and |H| is stored in |r0|
173 // through |r3|. Note there is no byte or bit reversal because we are
174 // evaluating POLYVAL.
175 let (r0, mut r1) = gcm_mul64_nohw(xi[0], h.lo);
176 let (mut r2, mut r3) = gcm_mul64_nohw(xi[1], h.hi);
177 let (mut mid0, mut mid1) = gcm_mul64_nohw(xi[0] ^ xi[1], h.hi ^ h.lo);
178 mid0 ^= r0 ^ r2;
179 mid1 ^= r1 ^ r3;
180 r2 ^= mid1;
181 r1 ^= mid0;
182
183 // Now we multiply our 256-bit result by x^-128 and reduce. |r2| and
184 // |r3| shifts into position and we must multiply |r0| and |r1| by x^-128. We
185 // have:
186 //
187 // 1 = x^121 + x^126 + x^127 + x^128
188 // x^-128 = x^-7 + x^-2 + x^-1 + 1
189 //
190 // This is the GHASH reduction step, but with bits flowing in reverse.
191
192 // The x^-7, x^-2, and x^-1 terms shift bits past x^0, which would require
193 // another reduction steps. Instead, we gather the excess bits, incorporate
194 // them into |r0| and |r1| and reduce once. See slides 17-19
195 // of https://crypto.stanford.edu/RealWorldCrypto/slides/gueron.pdf.
196 r1 ^= (r0 << 63) ^ (r0 << 62) ^ (r0 << 57);
197
198 // 1
199 r2 ^= r0;
200 r3 ^= r1;
201
202 // x^-1
203 r2 ^= r0 >> 1;
204 r2 ^= r1 << 63;
205 r3 ^= r1 >> 1;
206
207 // x^-2
208 r2 ^= r0 >> 2;
209 r2 ^= r1 << 62;
210 r3 ^= r1 >> 2;
211
212 // x^-7
213 r2 ^= r0 >> 7;
214 r2 ^= r1 << 57;
215 r3 ^= r1 >> 7;
216
217 *xi = [r2, r3];
218}
219
220pub(super) fn gmult(xi: &mut Xi, h: super::u128) {
221 with_swapped_xi(xi, |swapped: &mut [u64; 2]| {
222 gcm_polyval_nohw(xi:swapped, h);
223 })
224}
225
226pub(super) fn ghash(xi: &mut Xi, h: super::u128, input: &[[u8; BLOCK_LEN]]) {
227 with_swapped_xi(xi, |swapped: &mut [u64; 2]| {
228 input.iter().for_each(|&input: [u8; 16]| {
229 let input: [u64; 2] = input.array_split_map(u64::from_be_bytes);
230 swapped[0] ^= input[1];
231 swapped[1] ^= input[0];
232 gcm_polyval_nohw(xi:swapped, h);
233 });
234 });
235}
236
237#[inline]
238fn with_swapped_xi(Xi(xi: &mut Block): &mut Xi, f: impl FnOnce(&mut [u64; 2])) {
239 let unswapped: [u64; 2] = xi.as_ref().array_split_map(u64::from_be_bytes);
240 let mut swapped: [u64; 2] = [unswapped[1], unswapped[0]];
241 f(&mut swapped);
242 let reswapped: [u64; 2] = [swapped[1], swapped[0]];
243 *xi = Block::from(reswapped.map(u64::to_be_bytes))
244}
245