1 | // Copyright (c) 2019, Google Inc. |
2 | // Portions Copyright 2020-2024 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 | |
25 | use super::{ffi::U128, KeyValue, UpdateBlock, UpdateBlocks, Xi, BLOCK_LEN}; |
26 | use crate::polyfill::{slice::AsChunks, ArraySplitMap as _}; |
27 | |
28 | #[derive (Clone)] |
29 | pub struct Key { |
30 | h: U128, |
31 | } |
32 | |
33 | impl Key { |
34 | pub(in super::super) fn new(value: KeyValue) -> Self { |
35 | Self { h: init(value) } |
36 | } |
37 | } |
38 | |
39 | impl UpdateBlock for Key { |
40 | fn update_block(&self, xi: &mut Xi, a: [u8; BLOCK_LEN]) { |
41 | xi.bitxor_assign(a); |
42 | gmult(xi, self.h); |
43 | } |
44 | } |
45 | |
46 | impl UpdateBlocks for Key { |
47 | fn update_blocks(&self, xi: &mut Xi, input: AsChunks<u8, BLOCK_LEN>) { |
48 | ghash(xi, self.h, input); |
49 | } |
50 | } |
51 | |
52 | #[cfg (target_pointer_width = "64" )] |
53 | fn gcm_mul64_nohw(a: u64, b: u64) -> (u64, u64) { |
54 | #[allow (clippy::cast_possible_truncation)] |
55 | #[inline (always)] |
56 | fn lo(a: u128) -> u64 { |
57 | a as u64 |
58 | } |
59 | |
60 | #[inline (always)] |
61 | fn hi(a: u128) -> u64 { |
62 | lo(a >> 64) |
63 | } |
64 | |
65 | #[inline (always)] |
66 | fn mul(a: u64, b: u64) -> u128 { |
67 | u128::from(a) * u128::from(b) |
68 | } |
69 | |
70 | // One term every four bits means the largest term is 64/4 = 16, which barely |
71 | // overflows into the next term. Using one term every five bits would cost 25 |
72 | // multiplications instead of 16. It is faster to mask off the bottom four |
73 | // bits of |a|, giving a largest term of 60/4 = 15, and apply the bottom bits |
74 | // separately. |
75 | let a0 = a & 0x1111111111111110; |
76 | let a1 = a & 0x2222222222222220; |
77 | let a2 = a & 0x4444444444444440; |
78 | let a3 = a & 0x8888888888888880; |
79 | |
80 | let b0 = b & 0x1111111111111111; |
81 | let b1 = b & 0x2222222222222222; |
82 | let b2 = b & 0x4444444444444444; |
83 | let b3 = b & 0x8888888888888888; |
84 | |
85 | let c0 = mul(a0, b0) ^ mul(a1, b3) ^ mul(a2, b2) ^ mul(a3, b1); |
86 | let c1 = mul(a0, b1) ^ mul(a1, b0) ^ mul(a2, b3) ^ mul(a3, b2); |
87 | let c2 = mul(a0, b2) ^ mul(a1, b1) ^ mul(a2, b0) ^ mul(a3, b3); |
88 | let c3 = mul(a0, b3) ^ mul(a1, b2) ^ mul(a2, b1) ^ mul(a3, b0); |
89 | |
90 | // Multiply the bottom four bits of |a| with |b|. |
91 | let a0_mask = 0u64.wrapping_sub(a & 1); |
92 | let a1_mask = 0u64.wrapping_sub((a >> 1) & 1); |
93 | let a2_mask = 0u64.wrapping_sub((a >> 2) & 1); |
94 | let a3_mask = 0u64.wrapping_sub((a >> 3) & 1); |
95 | let extra = u128::from(a0_mask & b) |
96 | ^ (u128::from(a1_mask & b) << 1) |
97 | ^ (u128::from(a2_mask & b) << 2) |
98 | ^ (u128::from(a3_mask & b) << 3); |
99 | |
100 | let lo = (lo(c0) & 0x1111111111111111) |
101 | ^ (lo(c1) & 0x2222222222222222) |
102 | ^ (lo(c2) & 0x4444444444444444) |
103 | ^ (lo(c3) & 0x8888888888888888) |
104 | ^ lo(extra); |
105 | let hi = (hi(c0) & 0x1111111111111111) |
106 | ^ (hi(c1) & 0x2222222222222222) |
107 | ^ (hi(c2) & 0x4444444444444444) |
108 | ^ (hi(c3) & 0x8888888888888888) |
109 | ^ hi(extra); |
110 | (lo, hi) |
111 | } |
112 | |
113 | #[cfg (not(target_pointer_width = "64" ))] |
114 | fn gcm_mul32_nohw(a: u32, b: u32) -> u64 { |
115 | #[inline (always)] |
116 | fn mul(a: u32, b: u32) -> u64 { |
117 | u64::from(a) * u64::from(b) |
118 | } |
119 | |
120 | // One term every four bits means the largest term is 32/4 = 8, which does not |
121 | // overflow into the next term. |
122 | let a0 = a & 0x11111111; |
123 | let a1 = a & 0x22222222; |
124 | let a2 = a & 0x44444444; |
125 | let a3 = a & 0x88888888; |
126 | |
127 | let b0 = b & 0x11111111; |
128 | let b1 = b & 0x22222222; |
129 | let b2 = b & 0x44444444; |
130 | let b3 = b & 0x88888888; |
131 | |
132 | let c0 = mul(a0, b0) ^ mul(a1, b3) ^ mul(a2, b2) ^ mul(a3, b1); |
133 | let c1 = mul(a0, b1) ^ mul(a1, b0) ^ mul(a2, b3) ^ mul(a3, b2); |
134 | let c2 = mul(a0, b2) ^ mul(a1, b1) ^ mul(a2, b0) ^ mul(a3, b3); |
135 | let c3 = mul(a0, b3) ^ mul(a1, b2) ^ mul(a2, b1) ^ mul(a3, b0); |
136 | |
137 | (c0 & 0x1111111111111111) |
138 | | (c1 & 0x2222222222222222) |
139 | | (c2 & 0x4444444444444444) |
140 | | (c3 & 0x8888888888888888) |
141 | } |
142 | |
143 | #[cfg (not(target_pointer_width = "64" ))] |
144 | fn gcm_mul64_nohw(a: u64, b: u64) -> (u64, u64) { |
145 | #[inline (always)] |
146 | fn lo(a: u64) -> u32 { |
147 | a as u32 |
148 | } |
149 | #[inline (always)] |
150 | fn hi(a: u64) -> u32 { |
151 | lo(a >> 32) |
152 | } |
153 | |
154 | let a0 = lo(a); |
155 | let a1 = hi(a); |
156 | let b0 = lo(b); |
157 | let b1 = hi(b); |
158 | // Karatsuba multiplication. |
159 | let lo = gcm_mul32_nohw(a0, b0); |
160 | let hi = gcm_mul32_nohw(a1, b1); |
161 | let mid = gcm_mul32_nohw(a0 ^ a1, b0 ^ b1) ^ lo ^ hi; |
162 | (lo ^ (mid << 32), hi ^ (mid >> 32)) |
163 | } |
164 | |
165 | fn init(value: KeyValue) -> U128 { |
166 | let xi = value.into_inner(); |
167 | |
168 | // We implement GHASH in terms of POLYVAL, as described in RFC 8452. This |
169 | // avoids a shift by 1 in the multiplication, needed to account for bit |
170 | // reversal losing a bit after multiplication, that is, |
171 | // rev128(X) * rev128(Y) = rev255(X*Y). |
172 | // |
173 | // Per Appendix A, we run mulX_POLYVAL. Note this is the same transformation |
174 | // applied by |gcm_init_clmul|, etc. Note |Xi| has already been byteswapped. |
175 | // |
176 | // See also slide 16 of |
177 | // https://crypto.stanford.edu/RealWorldCrypto/slides/gueron.pdf |
178 | let mut lo = xi[1]; |
179 | let mut hi = xi[0]; |
180 | |
181 | let mut carry = hi >> 63; |
182 | carry = 0u64.wrapping_sub(carry); |
183 | |
184 | hi <<= 1; |
185 | hi |= lo >> 63; |
186 | lo <<= 1; |
187 | |
188 | // The irreducible polynomial is 1 + x^121 + x^126 + x^127 + x^128, so we |
189 | // conditionally add 0xc200...0001. |
190 | lo ^= carry & 1; |
191 | hi ^= carry & 0xc200000000000000; |
192 | |
193 | // This implementation does not use the rest of |Htable|. |
194 | U128 { hi, lo } |
195 | } |
196 | |
197 | fn gcm_polyval_nohw(xi: &mut [u64; 2], h: U128) { |
198 | // Karatsuba multiplication. The product of |Xi| and |H| is stored in |r0| |
199 | // through |r3|. Note there is no byte or bit reversal because we are |
200 | // evaluating POLYVAL. |
201 | let (r0, mut r1) = gcm_mul64_nohw(xi[0], h.lo); |
202 | let (mut r2, mut r3) = gcm_mul64_nohw(xi[1], h.hi); |
203 | let (mut mid0, mut mid1) = gcm_mul64_nohw(xi[0] ^ xi[1], h.hi ^ h.lo); |
204 | mid0 ^= r0 ^ r2; |
205 | mid1 ^= r1 ^ r3; |
206 | r2 ^= mid1; |
207 | r1 ^= mid0; |
208 | |
209 | // Now we multiply our 256-bit result by x^-128 and reduce. |r2| and |
210 | // |r3| shifts into position and we must multiply |r0| and |r1| by x^-128. We |
211 | // have: |
212 | // |
213 | // 1 = x^121 + x^126 + x^127 + x^128 |
214 | // x^-128 = x^-7 + x^-2 + x^-1 + 1 |
215 | // |
216 | // This is the GHASH reduction step, but with bits flowing in reverse. |
217 | |
218 | // The x^-7, x^-2, and x^-1 terms shift bits past x^0, which would require |
219 | // another reduction steps. Instead, we gather the excess bits, incorporate |
220 | // them into |r0| and |r1| and reduce once. See slides 17-19 |
221 | // of https://crypto.stanford.edu/RealWorldCrypto/slides/gueron.pdf. |
222 | r1 ^= (r0 << 63) ^ (r0 << 62) ^ (r0 << 57); |
223 | |
224 | // 1 |
225 | r2 ^= r0; |
226 | r3 ^= r1; |
227 | |
228 | // x^-1 |
229 | r2 ^= r0 >> 1; |
230 | r2 ^= r1 << 63; |
231 | r3 ^= r1 >> 1; |
232 | |
233 | // x^-2 |
234 | r2 ^= r0 >> 2; |
235 | r2 ^= r1 << 62; |
236 | r3 ^= r1 >> 2; |
237 | |
238 | // x^-7 |
239 | r2 ^= r0 >> 7; |
240 | r2 ^= r1 << 57; |
241 | r3 ^= r1 >> 7; |
242 | |
243 | *xi = [r2, r3]; |
244 | } |
245 | |
246 | fn gmult(xi: &mut Xi, h: U128) { |
247 | with_swapped_xi(xi, |swapped: &mut [u64; 2]| { |
248 | gcm_polyval_nohw(xi:swapped, h); |
249 | }) |
250 | } |
251 | |
252 | fn ghash(xi: &mut Xi, h: U128, input: AsChunks<u8, BLOCK_LEN>) { |
253 | with_swapped_xi(xi, |swapped: &mut [u64; 2]| { |
254 | input.into_iter().for_each(|&input: [u8; 16]| { |
255 | let input: [u64; _] = input.array_split_map(u64::from_be_bytes); |
256 | swapped[0] ^= input[1]; |
257 | swapped[1] ^= input[0]; |
258 | gcm_polyval_nohw(xi:swapped, h); |
259 | }); |
260 | }); |
261 | } |
262 | |
263 | #[inline ] |
264 | fn with_swapped_xi(Xi(xi: &mut [u8; 16]): &mut Xi, f: impl FnOnce(&mut [u64; 2])) { |
265 | let unswapped: [u64; 2] = xi.array_split_map(u64::from_be_bytes); |
266 | let mut swapped: [u64; 2] = [unswapped[1], unswapped[0]]; |
267 | f(&mut swapped); |
268 | let (xi_0: &mut [u8], xi_1: &mut [u8]) = xi.split_at_mut(BLOCK_LEN / 2); |
269 | xi_0.copy_from_slice(&u64::to_be_bytes(self:swapped[1])); |
270 | xi_1.copy_from_slice(&u64::to_be_bytes(self:swapped[0])); |
271 | } |
272 | |