Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Common Twofish algorithm parts shared between the c and assembler
4 * implementations
5 *
6 * Originally Twofish for GPG
7 * By Matthew Skala <mskala@ansuz.sooke.bc.ca>, July 26, 1998
8 * 256-bit key length added March 20, 1999
9 * Some modifications to reduce the text size by Werner Koch, April, 1998
10 * Ported to the kerneli patch by Marc Mutz <Marc@Mutz.com>
11 * Ported to CryptoAPI by Colin Slater <hoho@tacomeat.net>
12 *
13 * The original author has disclaimed all copyright interest in this
14 * code and thus put it in the public domain. The subsequent authors
15 * have put this under the GNU General Public License.
16 *
17 * This code is a "clean room" implementation, written from the paper
18 * _Twofish: A 128-Bit Block Cipher_ by Bruce Schneier, John Kelsey,
19 * Doug Whiting, David Wagner, Chris Hall, and Niels Ferguson, available
20 * through http://www.counterpane.com/twofish.html
21 *
22 * For background information on multiplication in finite fields, used for
23 * the matrix operations in the key schedule, see the book _Contemporary
24 * Abstract Algebra_ by Joseph A. Gallian, especially chapter 22 in the
25 * Third Edition.
26 */
27
28#include <crypto/algapi.h>
29#include <crypto/twofish.h>
30#include <linux/bitops.h>
31#include <linux/errno.h>
32#include <linux/init.h>
33#include <linux/kernel.h>
34#include <linux/module.h>
35#include <linux/types.h>
36
37
38/* The large precomputed tables for the Twofish cipher (twofish.c)
39 * Taken from the same source as twofish.c
40 * Marc Mutz <Marc@Mutz.com>
41 */
42
43/* These two tables are the q0 and q1 permutations, exactly as described in
44 * the Twofish paper. */
45
46static const u8 q0[256] = {
47 0xA9, 0x67, 0xB3, 0xE8, 0x04, 0xFD, 0xA3, 0x76, 0x9A, 0x92, 0x80, 0x78,
48 0xE4, 0xDD, 0xD1, 0x38, 0x0D, 0xC6, 0x35, 0x98, 0x18, 0xF7, 0xEC, 0x6C,
49 0x43, 0x75, 0x37, 0x26, 0xFA, 0x13, 0x94, 0x48, 0xF2, 0xD0, 0x8B, 0x30,
50 0x84, 0x54, 0xDF, 0x23, 0x19, 0x5B, 0x3D, 0x59, 0xF3, 0xAE, 0xA2, 0x82,
51 0x63, 0x01, 0x83, 0x2E, 0xD9, 0x51, 0x9B, 0x7C, 0xA6, 0xEB, 0xA5, 0xBE,
52 0x16, 0x0C, 0xE3, 0x61, 0xC0, 0x8C, 0x3A, 0xF5, 0x73, 0x2C, 0x25, 0x0B,
53 0xBB, 0x4E, 0x89, 0x6B, 0x53, 0x6A, 0xB4, 0xF1, 0xE1, 0xE6, 0xBD, 0x45,
54 0xE2, 0xF4, 0xB6, 0x66, 0xCC, 0x95, 0x03, 0x56, 0xD4, 0x1C, 0x1E, 0xD7,
55 0xFB, 0xC3, 0x8E, 0xB5, 0xE9, 0xCF, 0xBF, 0xBA, 0xEA, 0x77, 0x39, 0xAF,
56 0x33, 0xC9, 0x62, 0x71, 0x81, 0x79, 0x09, 0xAD, 0x24, 0xCD, 0xF9, 0xD8,
57 0xE5, 0xC5, 0xB9, 0x4D, 0x44, 0x08, 0x86, 0xE7, 0xA1, 0x1D, 0xAA, 0xED,
58 0x06, 0x70, 0xB2, 0xD2, 0x41, 0x7B, 0xA0, 0x11, 0x31, 0xC2, 0x27, 0x90,
59 0x20, 0xF6, 0x60, 0xFF, 0x96, 0x5C, 0xB1, 0xAB, 0x9E, 0x9C, 0x52, 0x1B,
60 0x5F, 0x93, 0x0A, 0xEF, 0x91, 0x85, 0x49, 0xEE, 0x2D, 0x4F, 0x8F, 0x3B,
61 0x47, 0x87, 0x6D, 0x46, 0xD6, 0x3E, 0x69, 0x64, 0x2A, 0xCE, 0xCB, 0x2F,
62 0xFC, 0x97, 0x05, 0x7A, 0xAC, 0x7F, 0xD5, 0x1A, 0x4B, 0x0E, 0xA7, 0x5A,
63 0x28, 0x14, 0x3F, 0x29, 0x88, 0x3C, 0x4C, 0x02, 0xB8, 0xDA, 0xB0, 0x17,
64 0x55, 0x1F, 0x8A, 0x7D, 0x57, 0xC7, 0x8D, 0x74, 0xB7, 0xC4, 0x9F, 0x72,
65 0x7E, 0x15, 0x22, 0x12, 0x58, 0x07, 0x99, 0x34, 0x6E, 0x50, 0xDE, 0x68,
66 0x65, 0xBC, 0xDB, 0xF8, 0xC8, 0xA8, 0x2B, 0x40, 0xDC, 0xFE, 0x32, 0xA4,
67 0xCA, 0x10, 0x21, 0xF0, 0xD3, 0x5D, 0x0F, 0x00, 0x6F, 0x9D, 0x36, 0x42,
68 0x4A, 0x5E, 0xC1, 0xE0
69};
70
71static const u8 q1[256] = {
72 0x75, 0xF3, 0xC6, 0xF4, 0xDB, 0x7B, 0xFB, 0xC8, 0x4A, 0xD3, 0xE6, 0x6B,
73 0x45, 0x7D, 0xE8, 0x4B, 0xD6, 0x32, 0xD8, 0xFD, 0x37, 0x71, 0xF1, 0xE1,
74 0x30, 0x0F, 0xF8, 0x1B, 0x87, 0xFA, 0x06, 0x3F, 0x5E, 0xBA, 0xAE, 0x5B,
75 0x8A, 0x00, 0xBC, 0x9D, 0x6D, 0xC1, 0xB1, 0x0E, 0x80, 0x5D, 0xD2, 0xD5,
76 0xA0, 0x84, 0x07, 0x14, 0xB5, 0x90, 0x2C, 0xA3, 0xB2, 0x73, 0x4C, 0x54,
77 0x92, 0x74, 0x36, 0x51, 0x38, 0xB0, 0xBD, 0x5A, 0xFC, 0x60, 0x62, 0x96,
78 0x6C, 0x42, 0xF7, 0x10, 0x7C, 0x28, 0x27, 0x8C, 0x13, 0x95, 0x9C, 0xC7,
79 0x24, 0x46, 0x3B, 0x70, 0xCA, 0xE3, 0x85, 0xCB, 0x11, 0xD0, 0x93, 0xB8,
80 0xA6, 0x83, 0x20, 0xFF, 0x9F, 0x77, 0xC3, 0xCC, 0x03, 0x6F, 0x08, 0xBF,
81 0x40, 0xE7, 0x2B, 0xE2, 0x79, 0x0C, 0xAA, 0x82, 0x41, 0x3A, 0xEA, 0xB9,
82 0xE4, 0x9A, 0xA4, 0x97, 0x7E, 0xDA, 0x7A, 0x17, 0x66, 0x94, 0xA1, 0x1D,
83 0x3D, 0xF0, 0xDE, 0xB3, 0x0B, 0x72, 0xA7, 0x1C, 0xEF, 0xD1, 0x53, 0x3E,
84 0x8F, 0x33, 0x26, 0x5F, 0xEC, 0x76, 0x2A, 0x49, 0x81, 0x88, 0xEE, 0x21,
85 0xC4, 0x1A, 0xEB, 0xD9, 0xC5, 0x39, 0x99, 0xCD, 0xAD, 0x31, 0x8B, 0x01,
86 0x18, 0x23, 0xDD, 0x1F, 0x4E, 0x2D, 0xF9, 0x48, 0x4F, 0xF2, 0x65, 0x8E,
87 0x78, 0x5C, 0x58, 0x19, 0x8D, 0xE5, 0x98, 0x57, 0x67, 0x7F, 0x05, 0x64,
88 0xAF, 0x63, 0xB6, 0xFE, 0xF5, 0xB7, 0x3C, 0xA5, 0xCE, 0xE9, 0x68, 0x44,
89 0xE0, 0x4D, 0x43, 0x69, 0x29, 0x2E, 0xAC, 0x15, 0x59, 0xA8, 0x0A, 0x9E,
90 0x6E, 0x47, 0xDF, 0x34, 0x35, 0x6A, 0xCF, 0xDC, 0x22, 0xC9, 0xC0, 0x9B,
91 0x89, 0xD4, 0xED, 0xAB, 0x12, 0xA2, 0x0D, 0x52, 0xBB, 0x02, 0x2F, 0xA9,
92 0xD7, 0x61, 0x1E, 0xB4, 0x50, 0x04, 0xF6, 0xC2, 0x16, 0x25, 0x86, 0x56,
93 0x55, 0x09, 0xBE, 0x91
94};
95
96/* These MDS tables are actually tables of MDS composed with q0 and q1,
97 * because it is only ever used that way and we can save some time by
98 * precomputing. Of course the main saving comes from precomputing the
99 * GF(2^8) multiplication involved in the MDS matrix multiply; by looking
100 * things up in these tables we reduce the matrix multiply to four lookups
101 * and three XORs. Semi-formally, the definition of these tables is:
102 * mds[0][i] = MDS (q1[i] 0 0 0)^T mds[1][i] = MDS (0 q0[i] 0 0)^T
103 * mds[2][i] = MDS (0 0 q1[i] 0)^T mds[3][i] = MDS (0 0 0 q0[i])^T
104 * where ^T means "transpose", the matrix multiply is performed in GF(2^8)
105 * represented as GF(2)[x]/v(x) where v(x)=x^8+x^6+x^5+x^3+1 as described
106 * by Schneier et al, and I'm casually glossing over the byte/word
107 * conversion issues. */
108
109static const u32 mds[4][256] = {
110 {
111 0xBCBC3275, 0xECEC21F3, 0x202043C6, 0xB3B3C9F4, 0xDADA03DB, 0x02028B7B,
112 0xE2E22BFB, 0x9E9EFAC8, 0xC9C9EC4A, 0xD4D409D3, 0x18186BE6, 0x1E1E9F6B,
113 0x98980E45, 0xB2B2387D, 0xA6A6D2E8, 0x2626B74B, 0x3C3C57D6, 0x93938A32,
114 0x8282EED8, 0x525298FD, 0x7B7BD437, 0xBBBB3771, 0x5B5B97F1, 0x474783E1,
115 0x24243C30, 0x5151E20F, 0xBABAC6F8, 0x4A4AF31B, 0xBFBF4887, 0x0D0D70FA,
116 0xB0B0B306, 0x7575DE3F, 0xD2D2FD5E, 0x7D7D20BA, 0x666631AE, 0x3A3AA35B,
117 0x59591C8A, 0x00000000, 0xCDCD93BC, 0x1A1AE09D, 0xAEAE2C6D, 0x7F7FABC1,
118 0x2B2BC7B1, 0xBEBEB90E, 0xE0E0A080, 0x8A8A105D, 0x3B3B52D2, 0x6464BAD5,
119 0xD8D888A0, 0xE7E7A584, 0x5F5FE807, 0x1B1B1114, 0x2C2CC2B5, 0xFCFCB490,
120 0x3131272C, 0x808065A3, 0x73732AB2, 0x0C0C8173, 0x79795F4C, 0x6B6B4154,
121 0x4B4B0292, 0x53536974, 0x94948F36, 0x83831F51, 0x2A2A3638, 0xC4C49CB0,
122 0x2222C8BD, 0xD5D5F85A, 0xBDBDC3FC, 0x48487860, 0xFFFFCE62, 0x4C4C0796,
123 0x4141776C, 0xC7C7E642, 0xEBEB24F7, 0x1C1C1410, 0x5D5D637C, 0x36362228,
124 0x6767C027, 0xE9E9AF8C, 0x4444F913, 0x1414EA95, 0xF5F5BB9C, 0xCFCF18C7,
125 0x3F3F2D24, 0xC0C0E346, 0x7272DB3B, 0x54546C70, 0x29294CCA, 0xF0F035E3,
126 0x0808FE85, 0xC6C617CB, 0xF3F34F11, 0x8C8CE4D0, 0xA4A45993, 0xCACA96B8,
127 0x68683BA6, 0xB8B84D83, 0x38382820, 0xE5E52EFF, 0xADAD569F, 0x0B0B8477,
128 0xC8C81DC3, 0x9999FFCC, 0x5858ED03, 0x19199A6F, 0x0E0E0A08, 0x95957EBF,
129 0x70705040, 0xF7F730E7, 0x6E6ECF2B, 0x1F1F6EE2, 0xB5B53D79, 0x09090F0C,
130 0x616134AA, 0x57571682, 0x9F9F0B41, 0x9D9D803A, 0x111164EA, 0x2525CDB9,
131 0xAFAFDDE4, 0x4545089A, 0xDFDF8DA4, 0xA3A35C97, 0xEAEAD57E, 0x353558DA,
132 0xEDEDD07A, 0x4343FC17, 0xF8F8CB66, 0xFBFBB194, 0x3737D3A1, 0xFAFA401D,
133 0xC2C2683D, 0xB4B4CCF0, 0x32325DDE, 0x9C9C71B3, 0x5656E70B, 0xE3E3DA72,
134 0x878760A7, 0x15151B1C, 0xF9F93AEF, 0x6363BFD1, 0x3434A953, 0x9A9A853E,
135 0xB1B1428F, 0x7C7CD133, 0x88889B26, 0x3D3DA65F, 0xA1A1D7EC, 0xE4E4DF76,
136 0x8181942A, 0x91910149, 0x0F0FFB81, 0xEEEEAA88, 0x161661EE, 0xD7D77321,
137 0x9797F5C4, 0xA5A5A81A, 0xFEFE3FEB, 0x6D6DB5D9, 0x7878AEC5, 0xC5C56D39,
138 0x1D1DE599, 0x7676A4CD, 0x3E3EDCAD, 0xCBCB6731, 0xB6B6478B, 0xEFEF5B01,
139 0x12121E18, 0x6060C523, 0x6A6AB0DD, 0x4D4DF61F, 0xCECEE94E, 0xDEDE7C2D,
140 0x55559DF9, 0x7E7E5A48, 0x2121B24F, 0x03037AF2, 0xA0A02665, 0x5E5E198E,
141 0x5A5A6678, 0x65654B5C, 0x62624E58, 0xFDFD4519, 0x0606F48D, 0x404086E5,
142 0xF2F2BE98, 0x3333AC57, 0x17179067, 0x05058E7F, 0xE8E85E05, 0x4F4F7D64,
143 0x89896AAF, 0x10109563, 0x74742FB6, 0x0A0A75FE, 0x5C5C92F5, 0x9B9B74B7,
144 0x2D2D333C, 0x3030D6A5, 0x2E2E49CE, 0x494989E9, 0x46467268, 0x77775544,
145 0xA8A8D8E0, 0x9696044D, 0x2828BD43, 0xA9A92969, 0xD9D97929, 0x8686912E,
146 0xD1D187AC, 0xF4F44A15, 0x8D8D1559, 0xD6D682A8, 0xB9B9BC0A, 0x42420D9E,
147 0xF6F6C16E, 0x2F2FB847, 0xDDDD06DF, 0x23233934, 0xCCCC6235, 0xF1F1C46A,
148 0xC1C112CF, 0x8585EBDC, 0x8F8F9E22, 0x7171A1C9, 0x9090F0C0, 0xAAAA539B,
149 0x0101F189, 0x8B8BE1D4, 0x4E4E8CED, 0x8E8E6FAB, 0xABABA212, 0x6F6F3EA2,
150 0xE6E6540D, 0xDBDBF252, 0x92927BBB, 0xB7B7B602, 0x6969CA2F, 0x3939D9A9,
151 0xD3D30CD7, 0xA7A72361, 0xA2A2AD1E, 0xC3C399B4, 0x6C6C4450, 0x07070504,
152 0x04047FF6, 0x272746C2, 0xACACA716, 0xD0D07625, 0x50501386, 0xDCDCF756,
153 0x84841A55, 0xE1E15109, 0x7A7A25BE, 0x1313EF91},
154
155 {
156 0xA9D93939, 0x67901717, 0xB3719C9C, 0xE8D2A6A6, 0x04050707, 0xFD985252,
157 0xA3658080, 0x76DFE4E4, 0x9A084545, 0x92024B4B, 0x80A0E0E0, 0x78665A5A,
158 0xE4DDAFAF, 0xDDB06A6A, 0xD1BF6363, 0x38362A2A, 0x0D54E6E6, 0xC6432020,
159 0x3562CCCC, 0x98BEF2F2, 0x181E1212, 0xF724EBEB, 0xECD7A1A1, 0x6C774141,
160 0x43BD2828, 0x7532BCBC, 0x37D47B7B, 0x269B8888, 0xFA700D0D, 0x13F94444,
161 0x94B1FBFB, 0x485A7E7E, 0xF27A0303, 0xD0E48C8C, 0x8B47B6B6, 0x303C2424,
162 0x84A5E7E7, 0x54416B6B, 0xDF06DDDD, 0x23C56060, 0x1945FDFD, 0x5BA33A3A,
163 0x3D68C2C2, 0x59158D8D, 0xF321ECEC, 0xAE316666, 0xA23E6F6F, 0x82165757,
164 0x63951010, 0x015BEFEF, 0x834DB8B8, 0x2E918686, 0xD9B56D6D, 0x511F8383,
165 0x9B53AAAA, 0x7C635D5D, 0xA63B6868, 0xEB3FFEFE, 0xA5D63030, 0xBE257A7A,
166 0x16A7ACAC, 0x0C0F0909, 0xE335F0F0, 0x6123A7A7, 0xC0F09090, 0x8CAFE9E9,
167 0x3A809D9D, 0xF5925C5C, 0x73810C0C, 0x2C273131, 0x2576D0D0, 0x0BE75656,
168 0xBB7B9292, 0x4EE9CECE, 0x89F10101, 0x6B9F1E1E, 0x53A93434, 0x6AC4F1F1,
169 0xB499C3C3, 0xF1975B5B, 0xE1834747, 0xE66B1818, 0xBDC82222, 0x450E9898,
170 0xE26E1F1F, 0xF4C9B3B3, 0xB62F7474, 0x66CBF8F8, 0xCCFF9999, 0x95EA1414,
171 0x03ED5858, 0x56F7DCDC, 0xD4E18B8B, 0x1C1B1515, 0x1EADA2A2, 0xD70CD3D3,
172 0xFB2BE2E2, 0xC31DC8C8, 0x8E195E5E, 0xB5C22C2C, 0xE9894949, 0xCF12C1C1,
173 0xBF7E9595, 0xBA207D7D, 0xEA641111, 0x77840B0B, 0x396DC5C5, 0xAF6A8989,
174 0x33D17C7C, 0xC9A17171, 0x62CEFFFF, 0x7137BBBB, 0x81FB0F0F, 0x793DB5B5,
175 0x0951E1E1, 0xADDC3E3E, 0x242D3F3F, 0xCDA47676, 0xF99D5555, 0xD8EE8282,
176 0xE5864040, 0xC5AE7878, 0xB9CD2525, 0x4D049696, 0x44557777, 0x080A0E0E,
177 0x86135050, 0xE730F7F7, 0xA1D33737, 0x1D40FAFA, 0xAA346161, 0xED8C4E4E,
178 0x06B3B0B0, 0x706C5454, 0xB22A7373, 0xD2523B3B, 0x410B9F9F, 0x7B8B0202,
179 0xA088D8D8, 0x114FF3F3, 0x3167CBCB, 0xC2462727, 0x27C06767, 0x90B4FCFC,
180 0x20283838, 0xF67F0404, 0x60784848, 0xFF2EE5E5, 0x96074C4C, 0x5C4B6565,
181 0xB1C72B2B, 0xAB6F8E8E, 0x9E0D4242, 0x9CBBF5F5, 0x52F2DBDB, 0x1BF34A4A,
182 0x5FA63D3D, 0x9359A4A4, 0x0ABCB9B9, 0xEF3AF9F9, 0x91EF1313, 0x85FE0808,
183 0x49019191, 0xEE611616, 0x2D7CDEDE, 0x4FB22121, 0x8F42B1B1, 0x3BDB7272,
184 0x47B82F2F, 0x8748BFBF, 0x6D2CAEAE, 0x46E3C0C0, 0xD6573C3C, 0x3E859A9A,
185 0x6929A9A9, 0x647D4F4F, 0x2A948181, 0xCE492E2E, 0xCB17C6C6, 0x2FCA6969,
186 0xFCC3BDBD, 0x975CA3A3, 0x055EE8E8, 0x7AD0EDED, 0xAC87D1D1, 0x7F8E0505,
187 0xD5BA6464, 0x1AA8A5A5, 0x4BB72626, 0x0EB9BEBE, 0xA7608787, 0x5AF8D5D5,
188 0x28223636, 0x14111B1B, 0x3FDE7575, 0x2979D9D9, 0x88AAEEEE, 0x3C332D2D,
189 0x4C5F7979, 0x02B6B7B7, 0xB896CACA, 0xDA583535, 0xB09CC4C4, 0x17FC4343,
190 0x551A8484, 0x1FF64D4D, 0x8A1C5959, 0x7D38B2B2, 0x57AC3333, 0xC718CFCF,
191 0x8DF40606, 0x74695353, 0xB7749B9B, 0xC4F59797, 0x9F56ADAD, 0x72DAE3E3,
192 0x7ED5EAEA, 0x154AF4F4, 0x229E8F8F, 0x12A2ABAB, 0x584E6262, 0x07E85F5F,
193 0x99E51D1D, 0x34392323, 0x6EC1F6F6, 0x50446C6C, 0xDE5D3232, 0x68724646,
194 0x6526A0A0, 0xBC93CDCD, 0xDB03DADA, 0xF8C6BABA, 0xC8FA9E9E, 0xA882D6D6,
195 0x2BCF6E6E, 0x40507070, 0xDCEB8585, 0xFE750A0A, 0x328A9393, 0xA48DDFDF,
196 0xCA4C2929, 0x10141C1C, 0x2173D7D7, 0xF0CCB4B4, 0xD309D4D4, 0x5D108A8A,
197 0x0FE25151, 0x00000000, 0x6F9A1919, 0x9DE01A1A, 0x368F9494, 0x42E6C7C7,
198 0x4AECC9C9, 0x5EFDD2D2, 0xC1AB7F7F, 0xE0D8A8A8},
199
200 {
201 0xBC75BC32, 0xECF3EC21, 0x20C62043, 0xB3F4B3C9, 0xDADBDA03, 0x027B028B,
202 0xE2FBE22B, 0x9EC89EFA, 0xC94AC9EC, 0xD4D3D409, 0x18E6186B, 0x1E6B1E9F,
203 0x9845980E, 0xB27DB238, 0xA6E8A6D2, 0x264B26B7, 0x3CD63C57, 0x9332938A,
204 0x82D882EE, 0x52FD5298, 0x7B377BD4, 0xBB71BB37, 0x5BF15B97, 0x47E14783,
205 0x2430243C, 0x510F51E2, 0xBAF8BAC6, 0x4A1B4AF3, 0xBF87BF48, 0x0DFA0D70,
206 0xB006B0B3, 0x753F75DE, 0xD25ED2FD, 0x7DBA7D20, 0x66AE6631, 0x3A5B3AA3,
207 0x598A591C, 0x00000000, 0xCDBCCD93, 0x1A9D1AE0, 0xAE6DAE2C, 0x7FC17FAB,
208 0x2BB12BC7, 0xBE0EBEB9, 0xE080E0A0, 0x8A5D8A10, 0x3BD23B52, 0x64D564BA,
209 0xD8A0D888, 0xE784E7A5, 0x5F075FE8, 0x1B141B11, 0x2CB52CC2, 0xFC90FCB4,
210 0x312C3127, 0x80A38065, 0x73B2732A, 0x0C730C81, 0x794C795F, 0x6B546B41,
211 0x4B924B02, 0x53745369, 0x9436948F, 0x8351831F, 0x2A382A36, 0xC4B0C49C,
212 0x22BD22C8, 0xD55AD5F8, 0xBDFCBDC3, 0x48604878, 0xFF62FFCE, 0x4C964C07,
213 0x416C4177, 0xC742C7E6, 0xEBF7EB24, 0x1C101C14, 0x5D7C5D63, 0x36283622,
214 0x672767C0, 0xE98CE9AF, 0x441344F9, 0x149514EA, 0xF59CF5BB, 0xCFC7CF18,
215 0x3F243F2D, 0xC046C0E3, 0x723B72DB, 0x5470546C, 0x29CA294C, 0xF0E3F035,
216 0x088508FE, 0xC6CBC617, 0xF311F34F, 0x8CD08CE4, 0xA493A459, 0xCAB8CA96,
217 0x68A6683B, 0xB883B84D, 0x38203828, 0xE5FFE52E, 0xAD9FAD56, 0x0B770B84,
218 0xC8C3C81D, 0x99CC99FF, 0x580358ED, 0x196F199A, 0x0E080E0A, 0x95BF957E,
219 0x70407050, 0xF7E7F730, 0x6E2B6ECF, 0x1FE21F6E, 0xB579B53D, 0x090C090F,
220 0x61AA6134, 0x57825716, 0x9F419F0B, 0x9D3A9D80, 0x11EA1164, 0x25B925CD,
221 0xAFE4AFDD, 0x459A4508, 0xDFA4DF8D, 0xA397A35C, 0xEA7EEAD5, 0x35DA3558,
222 0xED7AEDD0, 0x431743FC, 0xF866F8CB, 0xFB94FBB1, 0x37A137D3, 0xFA1DFA40,
223 0xC23DC268, 0xB4F0B4CC, 0x32DE325D, 0x9CB39C71, 0x560B56E7, 0xE372E3DA,
224 0x87A78760, 0x151C151B, 0xF9EFF93A, 0x63D163BF, 0x345334A9, 0x9A3E9A85,
225 0xB18FB142, 0x7C337CD1, 0x8826889B, 0x3D5F3DA6, 0xA1ECA1D7, 0xE476E4DF,
226 0x812A8194, 0x91499101, 0x0F810FFB, 0xEE88EEAA, 0x16EE1661, 0xD721D773,
227 0x97C497F5, 0xA51AA5A8, 0xFEEBFE3F, 0x6DD96DB5, 0x78C578AE, 0xC539C56D,
228 0x1D991DE5, 0x76CD76A4, 0x3EAD3EDC, 0xCB31CB67, 0xB68BB647, 0xEF01EF5B,
229 0x1218121E, 0x602360C5, 0x6ADD6AB0, 0x4D1F4DF6, 0xCE4ECEE9, 0xDE2DDE7C,
230 0x55F9559D, 0x7E487E5A, 0x214F21B2, 0x03F2037A, 0xA065A026, 0x5E8E5E19,
231 0x5A785A66, 0x655C654B, 0x6258624E, 0xFD19FD45, 0x068D06F4, 0x40E54086,
232 0xF298F2BE, 0x335733AC, 0x17671790, 0x057F058E, 0xE805E85E, 0x4F644F7D,
233 0x89AF896A, 0x10631095, 0x74B6742F, 0x0AFE0A75, 0x5CF55C92, 0x9BB79B74,
234 0x2D3C2D33, 0x30A530D6, 0x2ECE2E49, 0x49E94989, 0x46684672, 0x77447755,
235 0xA8E0A8D8, 0x964D9604, 0x284328BD, 0xA969A929, 0xD929D979, 0x862E8691,
236 0xD1ACD187, 0xF415F44A, 0x8D598D15, 0xD6A8D682, 0xB90AB9BC, 0x429E420D,
237 0xF66EF6C1, 0x2F472FB8, 0xDDDFDD06, 0x23342339, 0xCC35CC62, 0xF16AF1C4,
238 0xC1CFC112, 0x85DC85EB, 0x8F228F9E, 0x71C971A1, 0x90C090F0, 0xAA9BAA53,
239 0x018901F1, 0x8BD48BE1, 0x4EED4E8C, 0x8EAB8E6F, 0xAB12ABA2, 0x6FA26F3E,
240 0xE60DE654, 0xDB52DBF2, 0x92BB927B, 0xB702B7B6, 0x692F69CA, 0x39A939D9,
241 0xD3D7D30C, 0xA761A723, 0xA21EA2AD, 0xC3B4C399, 0x6C506C44, 0x07040705,
242 0x04F6047F, 0x27C22746, 0xAC16ACA7, 0xD025D076, 0x50865013, 0xDC56DCF7,
243 0x8455841A, 0xE109E151, 0x7ABE7A25, 0x139113EF},
244
245 {
246 0xD939A9D9, 0x90176790, 0x719CB371, 0xD2A6E8D2, 0x05070405, 0x9852FD98,
247 0x6580A365, 0xDFE476DF, 0x08459A08, 0x024B9202, 0xA0E080A0, 0x665A7866,
248 0xDDAFE4DD, 0xB06ADDB0, 0xBF63D1BF, 0x362A3836, 0x54E60D54, 0x4320C643,
249 0x62CC3562, 0xBEF298BE, 0x1E12181E, 0x24EBF724, 0xD7A1ECD7, 0x77416C77,
250 0xBD2843BD, 0x32BC7532, 0xD47B37D4, 0x9B88269B, 0x700DFA70, 0xF94413F9,
251 0xB1FB94B1, 0x5A7E485A, 0x7A03F27A, 0xE48CD0E4, 0x47B68B47, 0x3C24303C,
252 0xA5E784A5, 0x416B5441, 0x06DDDF06, 0xC56023C5, 0x45FD1945, 0xA33A5BA3,
253 0x68C23D68, 0x158D5915, 0x21ECF321, 0x3166AE31, 0x3E6FA23E, 0x16578216,
254 0x95106395, 0x5BEF015B, 0x4DB8834D, 0x91862E91, 0xB56DD9B5, 0x1F83511F,
255 0x53AA9B53, 0x635D7C63, 0x3B68A63B, 0x3FFEEB3F, 0xD630A5D6, 0x257ABE25,
256 0xA7AC16A7, 0x0F090C0F, 0x35F0E335, 0x23A76123, 0xF090C0F0, 0xAFE98CAF,
257 0x809D3A80, 0x925CF592, 0x810C7381, 0x27312C27, 0x76D02576, 0xE7560BE7,
258 0x7B92BB7B, 0xE9CE4EE9, 0xF10189F1, 0x9F1E6B9F, 0xA93453A9, 0xC4F16AC4,
259 0x99C3B499, 0x975BF197, 0x8347E183, 0x6B18E66B, 0xC822BDC8, 0x0E98450E,
260 0x6E1FE26E, 0xC9B3F4C9, 0x2F74B62F, 0xCBF866CB, 0xFF99CCFF, 0xEA1495EA,
261 0xED5803ED, 0xF7DC56F7, 0xE18BD4E1, 0x1B151C1B, 0xADA21EAD, 0x0CD3D70C,
262 0x2BE2FB2B, 0x1DC8C31D, 0x195E8E19, 0xC22CB5C2, 0x8949E989, 0x12C1CF12,
263 0x7E95BF7E, 0x207DBA20, 0x6411EA64, 0x840B7784, 0x6DC5396D, 0x6A89AF6A,
264 0xD17C33D1, 0xA171C9A1, 0xCEFF62CE, 0x37BB7137, 0xFB0F81FB, 0x3DB5793D,
265 0x51E10951, 0xDC3EADDC, 0x2D3F242D, 0xA476CDA4, 0x9D55F99D, 0xEE82D8EE,
266 0x8640E586, 0xAE78C5AE, 0xCD25B9CD, 0x04964D04, 0x55774455, 0x0A0E080A,
267 0x13508613, 0x30F7E730, 0xD337A1D3, 0x40FA1D40, 0x3461AA34, 0x8C4EED8C,
268 0xB3B006B3, 0x6C54706C, 0x2A73B22A, 0x523BD252, 0x0B9F410B, 0x8B027B8B,
269 0x88D8A088, 0x4FF3114F, 0x67CB3167, 0x4627C246, 0xC06727C0, 0xB4FC90B4,
270 0x28382028, 0x7F04F67F, 0x78486078, 0x2EE5FF2E, 0x074C9607, 0x4B655C4B,
271 0xC72BB1C7, 0x6F8EAB6F, 0x0D429E0D, 0xBBF59CBB, 0xF2DB52F2, 0xF34A1BF3,
272 0xA63D5FA6, 0x59A49359, 0xBCB90ABC, 0x3AF9EF3A, 0xEF1391EF, 0xFE0885FE,
273 0x01914901, 0x6116EE61, 0x7CDE2D7C, 0xB2214FB2, 0x42B18F42, 0xDB723BDB,
274 0xB82F47B8, 0x48BF8748, 0x2CAE6D2C, 0xE3C046E3, 0x573CD657, 0x859A3E85,
275 0x29A96929, 0x7D4F647D, 0x94812A94, 0x492ECE49, 0x17C6CB17, 0xCA692FCA,
276 0xC3BDFCC3, 0x5CA3975C, 0x5EE8055E, 0xD0ED7AD0, 0x87D1AC87, 0x8E057F8E,
277 0xBA64D5BA, 0xA8A51AA8, 0xB7264BB7, 0xB9BE0EB9, 0x6087A760, 0xF8D55AF8,
278 0x22362822, 0x111B1411, 0xDE753FDE, 0x79D92979, 0xAAEE88AA, 0x332D3C33,
279 0x5F794C5F, 0xB6B702B6, 0x96CAB896, 0x5835DA58, 0x9CC4B09C, 0xFC4317FC,
280 0x1A84551A, 0xF64D1FF6, 0x1C598A1C, 0x38B27D38, 0xAC3357AC, 0x18CFC718,
281 0xF4068DF4, 0x69537469, 0x749BB774, 0xF597C4F5, 0x56AD9F56, 0xDAE372DA,
282 0xD5EA7ED5, 0x4AF4154A, 0x9E8F229E, 0xA2AB12A2, 0x4E62584E, 0xE85F07E8,
283 0xE51D99E5, 0x39233439, 0xC1F66EC1, 0x446C5044, 0x5D32DE5D, 0x72466872,
284 0x26A06526, 0x93CDBC93, 0x03DADB03, 0xC6BAF8C6, 0xFA9EC8FA, 0x82D6A882,
285 0xCF6E2BCF, 0x50704050, 0xEB85DCEB, 0x750AFE75, 0x8A93328A, 0x8DDFA48D,
286 0x4C29CA4C, 0x141C1014, 0x73D72173, 0xCCB4F0CC, 0x09D4D309, 0x108A5D10,
287 0xE2510FE2, 0x00000000, 0x9A196F9A, 0xE01A9DE0, 0x8F94368F, 0xE6C742E6,
288 0xECC94AEC, 0xFDD25EFD, 0xAB7FC1AB, 0xD8A8E0D8}
289};
290
291/* The exp_to_poly and poly_to_exp tables are used to perform efficient
292 * operations in GF(2^8) represented as GF(2)[x]/w(x) where
293 * w(x)=x^8+x^6+x^3+x^2+1. We care about doing that because it's part of the
294 * definition of the RS matrix in the key schedule. Elements of that field
295 * are polynomials of degree not greater than 7 and all coefficients 0 or 1,
296 * which can be represented naturally by bytes (just substitute x=2). In that
297 * form, GF(2^8) addition is the same as bitwise XOR, but GF(2^8)
298 * multiplication is inefficient without hardware support. To multiply
299 * faster, I make use of the fact x is a generator for the nonzero elements,
300 * so that every element p of GF(2)[x]/w(x) is either 0 or equal to (x)^n for
301 * some n in 0..254. Note that caret is exponentiation in GF(2^8),
302 * *not* polynomial notation. So if I want to compute pq where p and q are
303 * in GF(2^8), I can just say:
304 * 1. if p=0 or q=0 then pq=0
305 * 2. otherwise, find m and n such that p=x^m and q=x^n
306 * 3. pq=(x^m)(x^n)=x^(m+n), so add m and n and find pq
307 * The translations in steps 2 and 3 are looked up in the tables
308 * poly_to_exp (for step 2) and exp_to_poly (for step 3). To see this
309 * in action, look at the CALC_S macro. As additional wrinkles, note that
310 * one of my operands is always a constant, so the poly_to_exp lookup on it
311 * is done in advance; I included the original values in the comments so
312 * readers can have some chance of recognizing that this *is* the RS matrix
313 * from the Twofish paper. I've only included the table entries I actually
314 * need; I never do a lookup on a variable input of zero and the biggest
315 * exponents I'll ever see are 254 (variable) and 237 (constant), so they'll
316 * never sum to more than 491. I'm repeating part of the exp_to_poly table
317 * so that I don't have to do mod-255 reduction in the exponent arithmetic.
318 * Since I know my constant operands are never zero, I only have to worry
319 * about zero values in the variable operand, and I do it with a simple
320 * conditional branch. I know conditionals are expensive, but I couldn't
321 * see a non-horrible way of avoiding them, and I did manage to group the
322 * statements so that each if covers four group multiplications. */
323
324static const u8 poly_to_exp[255] = {
325 0x00, 0x01, 0x17, 0x02, 0x2E, 0x18, 0x53, 0x03, 0x6A, 0x2F, 0x93, 0x19,
326 0x34, 0x54, 0x45, 0x04, 0x5C, 0x6B, 0xB6, 0x30, 0xA6, 0x94, 0x4B, 0x1A,
327 0x8C, 0x35, 0x81, 0x55, 0xAA, 0x46, 0x0D, 0x05, 0x24, 0x5D, 0x87, 0x6C,
328 0x9B, 0xB7, 0xC1, 0x31, 0x2B, 0xA7, 0xA3, 0x95, 0x98, 0x4C, 0xCA, 0x1B,
329 0xE6, 0x8D, 0x73, 0x36, 0xCD, 0x82, 0x12, 0x56, 0x62, 0xAB, 0xF0, 0x47,
330 0x4F, 0x0E, 0xBD, 0x06, 0xD4, 0x25, 0xD2, 0x5E, 0x27, 0x88, 0x66, 0x6D,
331 0xD6, 0x9C, 0x79, 0xB8, 0x08, 0xC2, 0xDF, 0x32, 0x68, 0x2C, 0xFD, 0xA8,
332 0x8A, 0xA4, 0x5A, 0x96, 0x29, 0x99, 0x22, 0x4D, 0x60, 0xCB, 0xE4, 0x1C,
333 0x7B, 0xE7, 0x3B, 0x8E, 0x9E, 0x74, 0xF4, 0x37, 0xD8, 0xCE, 0xF9, 0x83,
334 0x6F, 0x13, 0xB2, 0x57, 0xE1, 0x63, 0xDC, 0xAC, 0xC4, 0xF1, 0xAF, 0x48,
335 0x0A, 0x50, 0x42, 0x0F, 0xBA, 0xBE, 0xC7, 0x07, 0xDE, 0xD5, 0x78, 0x26,
336 0x65, 0xD3, 0xD1, 0x5F, 0xE3, 0x28, 0x21, 0x89, 0x59, 0x67, 0xFC, 0x6E,
337 0xB1, 0xD7, 0xF8, 0x9D, 0xF3, 0x7A, 0x3A, 0xB9, 0xC6, 0x09, 0x41, 0xC3,
338 0xAE, 0xE0, 0xDB, 0x33, 0x44, 0x69, 0x92, 0x2D, 0x52, 0xFE, 0x16, 0xA9,
339 0x0C, 0x8B, 0x80, 0xA5, 0x4A, 0x5B, 0xB5, 0x97, 0xC9, 0x2A, 0xA2, 0x9A,
340 0xC0, 0x23, 0x86, 0x4E, 0xBC, 0x61, 0xEF, 0xCC, 0x11, 0xE5, 0x72, 0x1D,
341 0x3D, 0x7C, 0xEB, 0xE8, 0xE9, 0x3C, 0xEA, 0x8F, 0x7D, 0x9F, 0xEC, 0x75,
342 0x1E, 0xF5, 0x3E, 0x38, 0xF6, 0xD9, 0x3F, 0xCF, 0x76, 0xFA, 0x1F, 0x84,
343 0xA0, 0x70, 0xED, 0x14, 0x90, 0xB3, 0x7E, 0x58, 0xFB, 0xE2, 0x20, 0x64,
344 0xD0, 0xDD, 0x77, 0xAD, 0xDA, 0xC5, 0x40, 0xF2, 0x39, 0xB0, 0xF7, 0x49,
345 0xB4, 0x0B, 0x7F, 0x51, 0x15, 0x43, 0x91, 0x10, 0x71, 0xBB, 0xEE, 0xBF,
346 0x85, 0xC8, 0xA1
347};
348
349static const u8 exp_to_poly[492] = {
350 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x4D, 0x9A, 0x79, 0xF2,
351 0xA9, 0x1F, 0x3E, 0x7C, 0xF8, 0xBD, 0x37, 0x6E, 0xDC, 0xF5, 0xA7, 0x03,
352 0x06, 0x0C, 0x18, 0x30, 0x60, 0xC0, 0xCD, 0xD7, 0xE3, 0x8B, 0x5B, 0xB6,
353 0x21, 0x42, 0x84, 0x45, 0x8A, 0x59, 0xB2, 0x29, 0x52, 0xA4, 0x05, 0x0A,
354 0x14, 0x28, 0x50, 0xA0, 0x0D, 0x1A, 0x34, 0x68, 0xD0, 0xED, 0x97, 0x63,
355 0xC6, 0xC1, 0xCF, 0xD3, 0xEB, 0x9B, 0x7B, 0xF6, 0xA1, 0x0F, 0x1E, 0x3C,
356 0x78, 0xF0, 0xAD, 0x17, 0x2E, 0x5C, 0xB8, 0x3D, 0x7A, 0xF4, 0xA5, 0x07,
357 0x0E, 0x1C, 0x38, 0x70, 0xE0, 0x8D, 0x57, 0xAE, 0x11, 0x22, 0x44, 0x88,
358 0x5D, 0xBA, 0x39, 0x72, 0xE4, 0x85, 0x47, 0x8E, 0x51, 0xA2, 0x09, 0x12,
359 0x24, 0x48, 0x90, 0x6D, 0xDA, 0xF9, 0xBF, 0x33, 0x66, 0xCC, 0xD5, 0xE7,
360 0x83, 0x4B, 0x96, 0x61, 0xC2, 0xC9, 0xDF, 0xF3, 0xAB, 0x1B, 0x36, 0x6C,
361 0xD8, 0xFD, 0xB7, 0x23, 0x46, 0x8C, 0x55, 0xAA, 0x19, 0x32, 0x64, 0xC8,
362 0xDD, 0xF7, 0xA3, 0x0B, 0x16, 0x2C, 0x58, 0xB0, 0x2D, 0x5A, 0xB4, 0x25,
363 0x4A, 0x94, 0x65, 0xCA, 0xD9, 0xFF, 0xB3, 0x2B, 0x56, 0xAC, 0x15, 0x2A,
364 0x54, 0xA8, 0x1D, 0x3A, 0x74, 0xE8, 0x9D, 0x77, 0xEE, 0x91, 0x6F, 0xDE,
365 0xF1, 0xAF, 0x13, 0x26, 0x4C, 0x98, 0x7D, 0xFA, 0xB9, 0x3F, 0x7E, 0xFC,
366 0xB5, 0x27, 0x4E, 0x9C, 0x75, 0xEA, 0x99, 0x7F, 0xFE, 0xB1, 0x2F, 0x5E,
367 0xBC, 0x35, 0x6A, 0xD4, 0xE5, 0x87, 0x43, 0x86, 0x41, 0x82, 0x49, 0x92,
368 0x69, 0xD2, 0xE9, 0x9F, 0x73, 0xE6, 0x81, 0x4F, 0x9E, 0x71, 0xE2, 0x89,
369 0x5F, 0xBE, 0x31, 0x62, 0xC4, 0xC5, 0xC7, 0xC3, 0xCB, 0xDB, 0xFB, 0xBB,
370 0x3B, 0x76, 0xEC, 0x95, 0x67, 0xCE, 0xD1, 0xEF, 0x93, 0x6B, 0xD6, 0xE1,
371 0x8F, 0x53, 0xA6, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x4D,
372 0x9A, 0x79, 0xF2, 0xA9, 0x1F, 0x3E, 0x7C, 0xF8, 0xBD, 0x37, 0x6E, 0xDC,
373 0xF5, 0xA7, 0x03, 0x06, 0x0C, 0x18, 0x30, 0x60, 0xC0, 0xCD, 0xD7, 0xE3,
374 0x8B, 0x5B, 0xB6, 0x21, 0x42, 0x84, 0x45, 0x8A, 0x59, 0xB2, 0x29, 0x52,
375 0xA4, 0x05, 0x0A, 0x14, 0x28, 0x50, 0xA0, 0x0D, 0x1A, 0x34, 0x68, 0xD0,
376 0xED, 0x97, 0x63, 0xC6, 0xC1, 0xCF, 0xD3, 0xEB, 0x9B, 0x7B, 0xF6, 0xA1,
377 0x0F, 0x1E, 0x3C, 0x78, 0xF0, 0xAD, 0x17, 0x2E, 0x5C, 0xB8, 0x3D, 0x7A,
378 0xF4, 0xA5, 0x07, 0x0E, 0x1C, 0x38, 0x70, 0xE0, 0x8D, 0x57, 0xAE, 0x11,
379 0x22, 0x44, 0x88, 0x5D, 0xBA, 0x39, 0x72, 0xE4, 0x85, 0x47, 0x8E, 0x51,
380 0xA2, 0x09, 0x12, 0x24, 0x48, 0x90, 0x6D, 0xDA, 0xF9, 0xBF, 0x33, 0x66,
381 0xCC, 0xD5, 0xE7, 0x83, 0x4B, 0x96, 0x61, 0xC2, 0xC9, 0xDF, 0xF3, 0xAB,
382 0x1B, 0x36, 0x6C, 0xD8, 0xFD, 0xB7, 0x23, 0x46, 0x8C, 0x55, 0xAA, 0x19,
383 0x32, 0x64, 0xC8, 0xDD, 0xF7, 0xA3, 0x0B, 0x16, 0x2C, 0x58, 0xB0, 0x2D,
384 0x5A, 0xB4, 0x25, 0x4A, 0x94, 0x65, 0xCA, 0xD9, 0xFF, 0xB3, 0x2B, 0x56,
385 0xAC, 0x15, 0x2A, 0x54, 0xA8, 0x1D, 0x3A, 0x74, 0xE8, 0x9D, 0x77, 0xEE,
386 0x91, 0x6F, 0xDE, 0xF1, 0xAF, 0x13, 0x26, 0x4C, 0x98, 0x7D, 0xFA, 0xB9,
387 0x3F, 0x7E, 0xFC, 0xB5, 0x27, 0x4E, 0x9C, 0x75, 0xEA, 0x99, 0x7F, 0xFE,
388 0xB1, 0x2F, 0x5E, 0xBC, 0x35, 0x6A, 0xD4, 0xE5, 0x87, 0x43, 0x86, 0x41,
389 0x82, 0x49, 0x92, 0x69, 0xD2, 0xE9, 0x9F, 0x73, 0xE6, 0x81, 0x4F, 0x9E,
390 0x71, 0xE2, 0x89, 0x5F, 0xBE, 0x31, 0x62, 0xC4, 0xC5, 0xC7, 0xC3, 0xCB
391};
392
393
394/* The table constants are indices of
395 * S-box entries, preprocessed through q0 and q1. */
396static const u8 calc_sb_tbl[512] = {
397 0xA9, 0x75, 0x67, 0xF3, 0xB3, 0xC6, 0xE8, 0xF4,
398 0x04, 0xDB, 0xFD, 0x7B, 0xA3, 0xFB, 0x76, 0xC8,
399 0x9A, 0x4A, 0x92, 0xD3, 0x80, 0xE6, 0x78, 0x6B,
400 0xE4, 0x45, 0xDD, 0x7D, 0xD1, 0xE8, 0x38, 0x4B,
401 0x0D, 0xD6, 0xC6, 0x32, 0x35, 0xD8, 0x98, 0xFD,
402 0x18, 0x37, 0xF7, 0x71, 0xEC, 0xF1, 0x6C, 0xE1,
403 0x43, 0x30, 0x75, 0x0F, 0x37, 0xF8, 0x26, 0x1B,
404 0xFA, 0x87, 0x13, 0xFA, 0x94, 0x06, 0x48, 0x3F,
405 0xF2, 0x5E, 0xD0, 0xBA, 0x8B, 0xAE, 0x30, 0x5B,
406 0x84, 0x8A, 0x54, 0x00, 0xDF, 0xBC, 0x23, 0x9D,
407 0x19, 0x6D, 0x5B, 0xC1, 0x3D, 0xB1, 0x59, 0x0E,
408 0xF3, 0x80, 0xAE, 0x5D, 0xA2, 0xD2, 0x82, 0xD5,
409 0x63, 0xA0, 0x01, 0x84, 0x83, 0x07, 0x2E, 0x14,
410 0xD9, 0xB5, 0x51, 0x90, 0x9B, 0x2C, 0x7C, 0xA3,
411 0xA6, 0xB2, 0xEB, 0x73, 0xA5, 0x4C, 0xBE, 0x54,
412 0x16, 0x92, 0x0C, 0x74, 0xE3, 0x36, 0x61, 0x51,
413 0xC0, 0x38, 0x8C, 0xB0, 0x3A, 0xBD, 0xF5, 0x5A,
414 0x73, 0xFC, 0x2C, 0x60, 0x25, 0x62, 0x0B, 0x96,
415 0xBB, 0x6C, 0x4E, 0x42, 0x89, 0xF7, 0x6B, 0x10,
416 0x53, 0x7C, 0x6A, 0x28, 0xB4, 0x27, 0xF1, 0x8C,
417 0xE1, 0x13, 0xE6, 0x95, 0xBD, 0x9C, 0x45, 0xC7,
418 0xE2, 0x24, 0xF4, 0x46, 0xB6, 0x3B, 0x66, 0x70,
419 0xCC, 0xCA, 0x95, 0xE3, 0x03, 0x85, 0x56, 0xCB,
420 0xD4, 0x11, 0x1C, 0xD0, 0x1E, 0x93, 0xD7, 0xB8,
421 0xFB, 0xA6, 0xC3, 0x83, 0x8E, 0x20, 0xB5, 0xFF,
422 0xE9, 0x9F, 0xCF, 0x77, 0xBF, 0xC3, 0xBA, 0xCC,
423 0xEA, 0x03, 0x77, 0x6F, 0x39, 0x08, 0xAF, 0xBF,
424 0x33, 0x40, 0xC9, 0xE7, 0x62, 0x2B, 0x71, 0xE2,
425 0x81, 0x79, 0x79, 0x0C, 0x09, 0xAA, 0xAD, 0x82,
426 0x24, 0x41, 0xCD, 0x3A, 0xF9, 0xEA, 0xD8, 0xB9,
427 0xE5, 0xE4, 0xC5, 0x9A, 0xB9, 0xA4, 0x4D, 0x97,
428 0x44, 0x7E, 0x08, 0xDA, 0x86, 0x7A, 0xE7, 0x17,
429 0xA1, 0x66, 0x1D, 0x94, 0xAA, 0xA1, 0xED, 0x1D,
430 0x06, 0x3D, 0x70, 0xF0, 0xB2, 0xDE, 0xD2, 0xB3,
431 0x41, 0x0B, 0x7B, 0x72, 0xA0, 0xA7, 0x11, 0x1C,
432 0x31, 0xEF, 0xC2, 0xD1, 0x27, 0x53, 0x90, 0x3E,
433 0x20, 0x8F, 0xF6, 0x33, 0x60, 0x26, 0xFF, 0x5F,
434 0x96, 0xEC, 0x5C, 0x76, 0xB1, 0x2A, 0xAB, 0x49,
435 0x9E, 0x81, 0x9C, 0x88, 0x52, 0xEE, 0x1B, 0x21,
436 0x5F, 0xC4, 0x93, 0x1A, 0x0A, 0xEB, 0xEF, 0xD9,
437 0x91, 0xC5, 0x85, 0x39, 0x49, 0x99, 0xEE, 0xCD,
438 0x2D, 0xAD, 0x4F, 0x31, 0x8F, 0x8B, 0x3B, 0x01,
439 0x47, 0x18, 0x87, 0x23, 0x6D, 0xDD, 0x46, 0x1F,
440 0xD6, 0x4E, 0x3E, 0x2D, 0x69, 0xF9, 0x64, 0x48,
441 0x2A, 0x4F, 0xCE, 0xF2, 0xCB, 0x65, 0x2F, 0x8E,
442 0xFC, 0x78, 0x97, 0x5C, 0x05, 0x58, 0x7A, 0x19,
443 0xAC, 0x8D, 0x7F, 0xE5, 0xD5, 0x98, 0x1A, 0x57,
444 0x4B, 0x67, 0x0E, 0x7F, 0xA7, 0x05, 0x5A, 0x64,
445 0x28, 0xAF, 0x14, 0x63, 0x3F, 0xB6, 0x29, 0xFE,
446 0x88, 0xF5, 0x3C, 0xB7, 0x4C, 0x3C, 0x02, 0xA5,
447 0xB8, 0xCE, 0xDA, 0xE9, 0xB0, 0x68, 0x17, 0x44,
448 0x55, 0xE0, 0x1F, 0x4D, 0x8A, 0x43, 0x7D, 0x69,
449 0x57, 0x29, 0xC7, 0x2E, 0x8D, 0xAC, 0x74, 0x15,
450 0xB7, 0x59, 0xC4, 0xA8, 0x9F, 0x0A, 0x72, 0x9E,
451 0x7E, 0x6E, 0x15, 0x47, 0x22, 0xDF, 0x12, 0x34,
452 0x58, 0x35, 0x07, 0x6A, 0x99, 0xCF, 0x34, 0xDC,
453 0x6E, 0x22, 0x50, 0xC9, 0xDE, 0xC0, 0x68, 0x9B,
454 0x65, 0x89, 0xBC, 0xD4, 0xDB, 0xED, 0xF8, 0xAB,
455 0xC8, 0x12, 0xA8, 0xA2, 0x2B, 0x0D, 0x40, 0x52,
456 0xDC, 0xBB, 0xFE, 0x02, 0x32, 0x2F, 0xA4, 0xA9,
457 0xCA, 0xD7, 0x10, 0x61, 0x21, 0x1E, 0xF0, 0xB4,
458 0xD3, 0x50, 0x5D, 0x04, 0x0F, 0xF6, 0x00, 0xC2,
459 0x6F, 0x16, 0x9D, 0x25, 0x36, 0x86, 0x42, 0x56,
460 0x4A, 0x55, 0x5E, 0x09, 0xC1, 0xBE, 0xE0, 0x91
461};
462
463/* Macro to perform one column of the RS matrix multiplication. The
464 * parameters a, b, c, and d are the four bytes of output; i is the index
465 * of the key bytes, and w, x, y, and z, are the column of constants from
466 * the RS matrix, preprocessed through the poly_to_exp table. */
467
468#define CALC_S(a, b, c, d, i, w, x, y, z) \
469 if (key[i]) { \
470 tmp = poly_to_exp[key[i] - 1]; \
471 (a) ^= exp_to_poly[tmp + (w)]; \
472 (b) ^= exp_to_poly[tmp + (x)]; \
473 (c) ^= exp_to_poly[tmp + (y)]; \
474 (d) ^= exp_to_poly[tmp + (z)]; \
475 }
476
477/* Macros to calculate the key-dependent S-boxes for a 128-bit key using
478 * the S vector from CALC_S. CALC_SB_2 computes a single entry in all
479 * four S-boxes, where i is the index of the entry to compute, and a and b
480 * are the index numbers preprocessed through the q0 and q1 tables
481 * respectively. */
482
483#define CALC_SB_2(i, a, b) \
484 ctx->s[0][i] = mds[0][q0[(a) ^ sa] ^ se]; \
485 ctx->s[1][i] = mds[1][q0[(b) ^ sb] ^ sf]; \
486 ctx->s[2][i] = mds[2][q1[(a) ^ sc] ^ sg]; \
487 ctx->s[3][i] = mds[3][q1[(b) ^ sd] ^ sh]
488
489/* Macro exactly like CALC_SB_2, but for 192-bit keys. */
490
491#define CALC_SB192_2(i, a, b) \
492 ctx->s[0][i] = mds[0][q0[q0[(b) ^ sa] ^ se] ^ si]; \
493 ctx->s[1][i] = mds[1][q0[q1[(b) ^ sb] ^ sf] ^ sj]; \
494 ctx->s[2][i] = mds[2][q1[q0[(a) ^ sc] ^ sg] ^ sk]; \
495 ctx->s[3][i] = mds[3][q1[q1[(a) ^ sd] ^ sh] ^ sl];
496
497/* Macro exactly like CALC_SB_2, but for 256-bit keys. */
498
499#define CALC_SB256_2(i, a, b) \
500 ctx->s[0][i] = mds[0][q0[q0[q1[(b) ^ sa] ^ se] ^ si] ^ sm]; \
501 ctx->s[1][i] = mds[1][q0[q1[q1[(a) ^ sb] ^ sf] ^ sj] ^ sn]; \
502 ctx->s[2][i] = mds[2][q1[q0[q0[(a) ^ sc] ^ sg] ^ sk] ^ so]; \
503 ctx->s[3][i] = mds[3][q1[q1[q0[(b) ^ sd] ^ sh] ^ sl] ^ sp];
504
505/* Macros to calculate the whitening and round subkeys. CALC_K_2 computes the
506 * last two stages of the h() function for a given index (either 2i or 2i+1).
507 * a, b, c, and d are the four bytes going into the last two stages. For
508 * 128-bit keys, this is the entire h() function and a and c are the index
509 * preprocessed through q0 and q1 respectively; for longer keys they are the
510 * output of previous stages. j is the index of the first key byte to use.
511 * CALC_K computes a pair of subkeys for 128-bit Twofish, by calling CALC_K_2
512 * twice, doing the Pseudo-Hadamard Transform, and doing the necessary
513 * rotations. Its parameters are: a, the array to write the results into,
514 * j, the index of the first output entry, k and l, the preprocessed indices
515 * for index 2i, and m and n, the preprocessed indices for index 2i+1.
516 * CALC_K192_2 expands CALC_K_2 to handle 192-bit keys, by doing an
517 * additional lookup-and-XOR stage. The parameters a, b, c and d are the
518 * four bytes going into the last three stages. For 192-bit keys, c = d
519 * are the index preprocessed through q0, and a = b are the index
520 * preprocessed through q1; j is the index of the first key byte to use.
521 * CALC_K192 is identical to CALC_K but for using the CALC_K192_2 macro
522 * instead of CALC_K_2.
523 * CALC_K256_2 expands CALC_K192_2 to handle 256-bit keys, by doing an
524 * additional lookup-and-XOR stage. The parameters a and b are the index
525 * preprocessed through q0 and q1 respectively; j is the index of the first
526 * key byte to use. CALC_K256 is identical to CALC_K but for using the
527 * CALC_K256_2 macro instead of CALC_K_2. */
528
529#define CALC_K_2(a, b, c, d, j) \
530 mds[0][q0[a ^ key[(j) + 8]] ^ key[j]] \
531 ^ mds[1][q0[b ^ key[(j) + 9]] ^ key[(j) + 1]] \
532 ^ mds[2][q1[c ^ key[(j) + 10]] ^ key[(j) + 2]] \
533 ^ mds[3][q1[d ^ key[(j) + 11]] ^ key[(j) + 3]]
534
535#define CALC_K(a, j, k, l, m, n) \
536 x = CALC_K_2 (k, l, k, l, 0); \
537 y = CALC_K_2 (m, n, m, n, 4); \
538 y = rol32(y, 8); \
539 x += y; y += x; ctx->a[j] = x; \
540 ctx->a[(j) + 1] = rol32(y, 9)
541
542#define CALC_K192_2(a, b, c, d, j) \
543 CALC_K_2 (q0[a ^ key[(j) + 16]], \
544 q1[b ^ key[(j) + 17]], \
545 q0[c ^ key[(j) + 18]], \
546 q1[d ^ key[(j) + 19]], j)
547
548#define CALC_K192(a, j, k, l, m, n) \
549 x = CALC_K192_2 (l, l, k, k, 0); \
550 y = CALC_K192_2 (n, n, m, m, 4); \
551 y = rol32(y, 8); \
552 x += y; y += x; ctx->a[j] = x; \
553 ctx->a[(j) + 1] = rol32(y, 9)
554
555#define CALC_K256_2(a, b, j) \
556 CALC_K192_2 (q1[b ^ key[(j) + 24]], \
557 q1[a ^ key[(j) + 25]], \
558 q0[a ^ key[(j) + 26]], \
559 q0[b ^ key[(j) + 27]], j)
560
561#define CALC_K256(a, j, k, l, m, n) \
562 x = CALC_K256_2 (k, l, 0); \
563 y = CALC_K256_2 (m, n, 4); \
564 y = rol32(y, 8); \
565 x += y; y += x; ctx->a[j] = x; \
566 ctx->a[(j) + 1] = rol32(y, 9)
567
568/* Perform the key setup. */
569int __twofish_setkey(struct twofish_ctx *ctx, const u8 *key,
570 unsigned int key_len)
571{
572 int i, j, k;
573
574 /* Temporaries for CALC_K. */
575 u32 x, y;
576
577 /* The S vector used to key the S-boxes, split up into individual bytes.
578 * 128-bit keys use only sa through sh; 256-bit use all of them. */
579 u8 sa = 0, sb = 0, sc = 0, sd = 0, se = 0, sf = 0, sg = 0, sh = 0;
580 u8 si = 0, sj = 0, sk = 0, sl = 0, sm = 0, sn = 0, so = 0, sp = 0;
581
582 /* Temporary for CALC_S. */
583 u8 tmp;
584
585 /* Check key length. */
586 if (key_len % 8)
587 return -EINVAL; /* unsupported key length */
588
589 /* Compute the first two words of the S vector. The magic numbers are
590 * the entries of the RS matrix, preprocessed through poly_to_exp. The
591 * numbers in the comments are the original (polynomial form) matrix
592 * entries. */
593 CALC_S (sa, sb, sc, sd, 0, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
594 CALC_S (sa, sb, sc, sd, 1, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
595 CALC_S (sa, sb, sc, sd, 2, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
596 CALC_S (sa, sb, sc, sd, 3, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
597 CALC_S (sa, sb, sc, sd, 4, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
598 CALC_S (sa, sb, sc, sd, 5, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
599 CALC_S (sa, sb, sc, sd, 6, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
600 CALC_S (sa, sb, sc, sd, 7, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
601 CALC_S (se, sf, sg, sh, 8, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
602 CALC_S (se, sf, sg, sh, 9, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
603 CALC_S (se, sf, sg, sh, 10, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
604 CALC_S (se, sf, sg, sh, 11, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
605 CALC_S (se, sf, sg, sh, 12, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
606 CALC_S (se, sf, sg, sh, 13, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
607 CALC_S (se, sf, sg, sh, 14, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
608 CALC_S (se, sf, sg, sh, 15, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
609
610 if (key_len == 24 || key_len == 32) { /* 192- or 256-bit key */
611 /* Calculate the third word of the S vector */
612 CALC_S (si, sj, sk, sl, 16, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
613 CALC_S (si, sj, sk, sl, 17, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
614 CALC_S (si, sj, sk, sl, 18, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
615 CALC_S (si, sj, sk, sl, 19, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
616 CALC_S (si, sj, sk, sl, 20, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
617 CALC_S (si, sj, sk, sl, 21, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
618 CALC_S (si, sj, sk, sl, 22, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
619 CALC_S (si, sj, sk, sl, 23, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
620 }
621
622 if (key_len == 32) { /* 256-bit key */
623 /* Calculate the fourth word of the S vector */
624 CALC_S (sm, sn, so, sp, 24, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
625 CALC_S (sm, sn, so, sp, 25, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
626 CALC_S (sm, sn, so, sp, 26, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
627 CALC_S (sm, sn, so, sp, 27, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
628 CALC_S (sm, sn, so, sp, 28, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
629 CALC_S (sm, sn, so, sp, 29, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
630 CALC_S (sm, sn, so, sp, 30, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
631 CALC_S (sm, sn, so, sp, 31, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
632
633 /* Compute the S-boxes. */
634 for ( i = j = 0, k = 1; i < 256; i++, j += 2, k += 2 ) {
635 CALC_SB256_2( i, calc_sb_tbl[j], calc_sb_tbl[k] );
636 }
637
638 /* CALC_K256/CALC_K192/CALC_K loops were unrolled.
639 * Unrolling produced x2.5 more code (+18k on i386),
640 * and speeded up key setup by 7%:
641 * unrolled: twofish_setkey/sec: 41128
642 * loop: twofish_setkey/sec: 38148
643 * CALC_K256: ~100 insns each
644 * CALC_K192: ~90 insns
645 * CALC_K: ~70 insns
646 */
647 /* Calculate whitening and round subkeys */
648 for ( i = 0; i < 8; i += 2 ) {
649 CALC_K256 (w, i, q0[i], q1[i], q0[i+1], q1[i+1]);
650 }
651 for ( i = 0; i < 32; i += 2 ) {
652 CALC_K256 (k, i, q0[i+8], q1[i+8], q0[i+9], q1[i+9]);
653 }
654 } else if (key_len == 24) { /* 192-bit key */
655 /* Compute the S-boxes. */
656 for ( i = j = 0, k = 1; i < 256; i++, j += 2, k += 2 ) {
657 CALC_SB192_2( i, calc_sb_tbl[j], calc_sb_tbl[k] );
658 }
659
660 /* Calculate whitening and round subkeys */
661 for ( i = 0; i < 8; i += 2 ) {
662 CALC_K192 (w, i, q0[i], q1[i], q0[i+1], q1[i+1]);
663 }
664 for ( i = 0; i < 32; i += 2 ) {
665 CALC_K192 (k, i, q0[i+8], q1[i+8], q0[i+9], q1[i+9]);
666 }
667 } else { /* 128-bit key */
668 /* Compute the S-boxes. */
669 for ( i = j = 0, k = 1; i < 256; i++, j += 2, k += 2 ) {
670 CALC_SB_2( i, calc_sb_tbl[j], calc_sb_tbl[k] );
671 }
672
673 /* Calculate whitening and round subkeys */
674 for ( i = 0; i < 8; i += 2 ) {
675 CALC_K (w, i, q0[i], q1[i], q0[i+1], q1[i+1]);
676 }
677 for ( i = 0; i < 32; i += 2 ) {
678 CALC_K (k, i, q0[i+8], q1[i+8], q0[i+9], q1[i+9]);
679 }
680 }
681
682 return 0;
683}
684EXPORT_SYMBOL_GPL(__twofish_setkey);
685
686int twofish_setkey(struct crypto_tfm *tfm, const u8 *key, unsigned int key_len)
687{
688 return __twofish_setkey(crypto_tfm_ctx(tfm), key, key_len);
689}
690EXPORT_SYMBOL_GPL(twofish_setkey);
691
692MODULE_LICENSE("GPL");
693MODULE_DESCRIPTION("Twofish cipher common functions");
1/*
2 * Common Twofish algorithm parts shared between the c and assembler
3 * implementations
4 *
5 * Originally Twofish for GPG
6 * By Matthew Skala <mskala@ansuz.sooke.bc.ca>, July 26, 1998
7 * 256-bit key length added March 20, 1999
8 * Some modifications to reduce the text size by Werner Koch, April, 1998
9 * Ported to the kerneli patch by Marc Mutz <Marc@Mutz.com>
10 * Ported to CryptoAPI by Colin Slater <hoho@tacomeat.net>
11 *
12 * The original author has disclaimed all copyright interest in this
13 * code and thus put it in the public domain. The subsequent authors
14 * have put this under the GNU General Public License.
15 *
16 * This program is free software; you can redistribute it and/or modify
17 * it under the terms of the GNU General Public License as published by
18 * the Free Software Foundation; either version 2 of the License, or
19 * (at your option) any later version.
20 *
21 * This program is distributed in the hope that it will be useful,
22 * but WITHOUT ANY WARRANTY; without even the implied warranty of
23 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * GNU General Public License for more details.
25 *
26 * You should have received a copy of the GNU General Public License
27 * along with this program; if not, write to the Free Software
28 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307
29 * USA
30 *
31 * This code is a "clean room" implementation, written from the paper
32 * _Twofish: A 128-Bit Block Cipher_ by Bruce Schneier, John Kelsey,
33 * Doug Whiting, David Wagner, Chris Hall, and Niels Ferguson, available
34 * through http://www.counterpane.com/twofish.html
35 *
36 * For background information on multiplication in finite fields, used for
37 * the matrix operations in the key schedule, see the book _Contemporary
38 * Abstract Algebra_ by Joseph A. Gallian, especially chapter 22 in the
39 * Third Edition.
40 */
41
42#include <crypto/twofish.h>
43#include <linux/bitops.h>
44#include <linux/crypto.h>
45#include <linux/errno.h>
46#include <linux/init.h>
47#include <linux/kernel.h>
48#include <linux/module.h>
49#include <linux/types.h>
50
51
52/* The large precomputed tables for the Twofish cipher (twofish.c)
53 * Taken from the same source as twofish.c
54 * Marc Mutz <Marc@Mutz.com>
55 */
56
57/* These two tables are the q0 and q1 permutations, exactly as described in
58 * the Twofish paper. */
59
60static const u8 q0[256] = {
61 0xA9, 0x67, 0xB3, 0xE8, 0x04, 0xFD, 0xA3, 0x76, 0x9A, 0x92, 0x80, 0x78,
62 0xE4, 0xDD, 0xD1, 0x38, 0x0D, 0xC6, 0x35, 0x98, 0x18, 0xF7, 0xEC, 0x6C,
63 0x43, 0x75, 0x37, 0x26, 0xFA, 0x13, 0x94, 0x48, 0xF2, 0xD0, 0x8B, 0x30,
64 0x84, 0x54, 0xDF, 0x23, 0x19, 0x5B, 0x3D, 0x59, 0xF3, 0xAE, 0xA2, 0x82,
65 0x63, 0x01, 0x83, 0x2E, 0xD9, 0x51, 0x9B, 0x7C, 0xA6, 0xEB, 0xA5, 0xBE,
66 0x16, 0x0C, 0xE3, 0x61, 0xC0, 0x8C, 0x3A, 0xF5, 0x73, 0x2C, 0x25, 0x0B,
67 0xBB, 0x4E, 0x89, 0x6B, 0x53, 0x6A, 0xB4, 0xF1, 0xE1, 0xE6, 0xBD, 0x45,
68 0xE2, 0xF4, 0xB6, 0x66, 0xCC, 0x95, 0x03, 0x56, 0xD4, 0x1C, 0x1E, 0xD7,
69 0xFB, 0xC3, 0x8E, 0xB5, 0xE9, 0xCF, 0xBF, 0xBA, 0xEA, 0x77, 0x39, 0xAF,
70 0x33, 0xC9, 0x62, 0x71, 0x81, 0x79, 0x09, 0xAD, 0x24, 0xCD, 0xF9, 0xD8,
71 0xE5, 0xC5, 0xB9, 0x4D, 0x44, 0x08, 0x86, 0xE7, 0xA1, 0x1D, 0xAA, 0xED,
72 0x06, 0x70, 0xB2, 0xD2, 0x41, 0x7B, 0xA0, 0x11, 0x31, 0xC2, 0x27, 0x90,
73 0x20, 0xF6, 0x60, 0xFF, 0x96, 0x5C, 0xB1, 0xAB, 0x9E, 0x9C, 0x52, 0x1B,
74 0x5F, 0x93, 0x0A, 0xEF, 0x91, 0x85, 0x49, 0xEE, 0x2D, 0x4F, 0x8F, 0x3B,
75 0x47, 0x87, 0x6D, 0x46, 0xD6, 0x3E, 0x69, 0x64, 0x2A, 0xCE, 0xCB, 0x2F,
76 0xFC, 0x97, 0x05, 0x7A, 0xAC, 0x7F, 0xD5, 0x1A, 0x4B, 0x0E, 0xA7, 0x5A,
77 0x28, 0x14, 0x3F, 0x29, 0x88, 0x3C, 0x4C, 0x02, 0xB8, 0xDA, 0xB0, 0x17,
78 0x55, 0x1F, 0x8A, 0x7D, 0x57, 0xC7, 0x8D, 0x74, 0xB7, 0xC4, 0x9F, 0x72,
79 0x7E, 0x15, 0x22, 0x12, 0x58, 0x07, 0x99, 0x34, 0x6E, 0x50, 0xDE, 0x68,
80 0x65, 0xBC, 0xDB, 0xF8, 0xC8, 0xA8, 0x2B, 0x40, 0xDC, 0xFE, 0x32, 0xA4,
81 0xCA, 0x10, 0x21, 0xF0, 0xD3, 0x5D, 0x0F, 0x00, 0x6F, 0x9D, 0x36, 0x42,
82 0x4A, 0x5E, 0xC1, 0xE0
83};
84
85static const u8 q1[256] = {
86 0x75, 0xF3, 0xC6, 0xF4, 0xDB, 0x7B, 0xFB, 0xC8, 0x4A, 0xD3, 0xE6, 0x6B,
87 0x45, 0x7D, 0xE8, 0x4B, 0xD6, 0x32, 0xD8, 0xFD, 0x37, 0x71, 0xF1, 0xE1,
88 0x30, 0x0F, 0xF8, 0x1B, 0x87, 0xFA, 0x06, 0x3F, 0x5E, 0xBA, 0xAE, 0x5B,
89 0x8A, 0x00, 0xBC, 0x9D, 0x6D, 0xC1, 0xB1, 0x0E, 0x80, 0x5D, 0xD2, 0xD5,
90 0xA0, 0x84, 0x07, 0x14, 0xB5, 0x90, 0x2C, 0xA3, 0xB2, 0x73, 0x4C, 0x54,
91 0x92, 0x74, 0x36, 0x51, 0x38, 0xB0, 0xBD, 0x5A, 0xFC, 0x60, 0x62, 0x96,
92 0x6C, 0x42, 0xF7, 0x10, 0x7C, 0x28, 0x27, 0x8C, 0x13, 0x95, 0x9C, 0xC7,
93 0x24, 0x46, 0x3B, 0x70, 0xCA, 0xE3, 0x85, 0xCB, 0x11, 0xD0, 0x93, 0xB8,
94 0xA6, 0x83, 0x20, 0xFF, 0x9F, 0x77, 0xC3, 0xCC, 0x03, 0x6F, 0x08, 0xBF,
95 0x40, 0xE7, 0x2B, 0xE2, 0x79, 0x0C, 0xAA, 0x82, 0x41, 0x3A, 0xEA, 0xB9,
96 0xE4, 0x9A, 0xA4, 0x97, 0x7E, 0xDA, 0x7A, 0x17, 0x66, 0x94, 0xA1, 0x1D,
97 0x3D, 0xF0, 0xDE, 0xB3, 0x0B, 0x72, 0xA7, 0x1C, 0xEF, 0xD1, 0x53, 0x3E,
98 0x8F, 0x33, 0x26, 0x5F, 0xEC, 0x76, 0x2A, 0x49, 0x81, 0x88, 0xEE, 0x21,
99 0xC4, 0x1A, 0xEB, 0xD9, 0xC5, 0x39, 0x99, 0xCD, 0xAD, 0x31, 0x8B, 0x01,
100 0x18, 0x23, 0xDD, 0x1F, 0x4E, 0x2D, 0xF9, 0x48, 0x4F, 0xF2, 0x65, 0x8E,
101 0x78, 0x5C, 0x58, 0x19, 0x8D, 0xE5, 0x98, 0x57, 0x67, 0x7F, 0x05, 0x64,
102 0xAF, 0x63, 0xB6, 0xFE, 0xF5, 0xB7, 0x3C, 0xA5, 0xCE, 0xE9, 0x68, 0x44,
103 0xE0, 0x4D, 0x43, 0x69, 0x29, 0x2E, 0xAC, 0x15, 0x59, 0xA8, 0x0A, 0x9E,
104 0x6E, 0x47, 0xDF, 0x34, 0x35, 0x6A, 0xCF, 0xDC, 0x22, 0xC9, 0xC0, 0x9B,
105 0x89, 0xD4, 0xED, 0xAB, 0x12, 0xA2, 0x0D, 0x52, 0xBB, 0x02, 0x2F, 0xA9,
106 0xD7, 0x61, 0x1E, 0xB4, 0x50, 0x04, 0xF6, 0xC2, 0x16, 0x25, 0x86, 0x56,
107 0x55, 0x09, 0xBE, 0x91
108};
109
110/* These MDS tables are actually tables of MDS composed with q0 and q1,
111 * because it is only ever used that way and we can save some time by
112 * precomputing. Of course the main saving comes from precomputing the
113 * GF(2^8) multiplication involved in the MDS matrix multiply; by looking
114 * things up in these tables we reduce the matrix multiply to four lookups
115 * and three XORs. Semi-formally, the definition of these tables is:
116 * mds[0][i] = MDS (q1[i] 0 0 0)^T mds[1][i] = MDS (0 q0[i] 0 0)^T
117 * mds[2][i] = MDS (0 0 q1[i] 0)^T mds[3][i] = MDS (0 0 0 q0[i])^T
118 * where ^T means "transpose", the matrix multiply is performed in GF(2^8)
119 * represented as GF(2)[x]/v(x) where v(x)=x^8+x^6+x^5+x^3+1 as described
120 * by Schneier et al, and I'm casually glossing over the byte/word
121 * conversion issues. */
122
123static const u32 mds[4][256] = {
124 {
125 0xBCBC3275, 0xECEC21F3, 0x202043C6, 0xB3B3C9F4, 0xDADA03DB, 0x02028B7B,
126 0xE2E22BFB, 0x9E9EFAC8, 0xC9C9EC4A, 0xD4D409D3, 0x18186BE6, 0x1E1E9F6B,
127 0x98980E45, 0xB2B2387D, 0xA6A6D2E8, 0x2626B74B, 0x3C3C57D6, 0x93938A32,
128 0x8282EED8, 0x525298FD, 0x7B7BD437, 0xBBBB3771, 0x5B5B97F1, 0x474783E1,
129 0x24243C30, 0x5151E20F, 0xBABAC6F8, 0x4A4AF31B, 0xBFBF4887, 0x0D0D70FA,
130 0xB0B0B306, 0x7575DE3F, 0xD2D2FD5E, 0x7D7D20BA, 0x666631AE, 0x3A3AA35B,
131 0x59591C8A, 0x00000000, 0xCDCD93BC, 0x1A1AE09D, 0xAEAE2C6D, 0x7F7FABC1,
132 0x2B2BC7B1, 0xBEBEB90E, 0xE0E0A080, 0x8A8A105D, 0x3B3B52D2, 0x6464BAD5,
133 0xD8D888A0, 0xE7E7A584, 0x5F5FE807, 0x1B1B1114, 0x2C2CC2B5, 0xFCFCB490,
134 0x3131272C, 0x808065A3, 0x73732AB2, 0x0C0C8173, 0x79795F4C, 0x6B6B4154,
135 0x4B4B0292, 0x53536974, 0x94948F36, 0x83831F51, 0x2A2A3638, 0xC4C49CB0,
136 0x2222C8BD, 0xD5D5F85A, 0xBDBDC3FC, 0x48487860, 0xFFFFCE62, 0x4C4C0796,
137 0x4141776C, 0xC7C7E642, 0xEBEB24F7, 0x1C1C1410, 0x5D5D637C, 0x36362228,
138 0x6767C027, 0xE9E9AF8C, 0x4444F913, 0x1414EA95, 0xF5F5BB9C, 0xCFCF18C7,
139 0x3F3F2D24, 0xC0C0E346, 0x7272DB3B, 0x54546C70, 0x29294CCA, 0xF0F035E3,
140 0x0808FE85, 0xC6C617CB, 0xF3F34F11, 0x8C8CE4D0, 0xA4A45993, 0xCACA96B8,
141 0x68683BA6, 0xB8B84D83, 0x38382820, 0xE5E52EFF, 0xADAD569F, 0x0B0B8477,
142 0xC8C81DC3, 0x9999FFCC, 0x5858ED03, 0x19199A6F, 0x0E0E0A08, 0x95957EBF,
143 0x70705040, 0xF7F730E7, 0x6E6ECF2B, 0x1F1F6EE2, 0xB5B53D79, 0x09090F0C,
144 0x616134AA, 0x57571682, 0x9F9F0B41, 0x9D9D803A, 0x111164EA, 0x2525CDB9,
145 0xAFAFDDE4, 0x4545089A, 0xDFDF8DA4, 0xA3A35C97, 0xEAEAD57E, 0x353558DA,
146 0xEDEDD07A, 0x4343FC17, 0xF8F8CB66, 0xFBFBB194, 0x3737D3A1, 0xFAFA401D,
147 0xC2C2683D, 0xB4B4CCF0, 0x32325DDE, 0x9C9C71B3, 0x5656E70B, 0xE3E3DA72,
148 0x878760A7, 0x15151B1C, 0xF9F93AEF, 0x6363BFD1, 0x3434A953, 0x9A9A853E,
149 0xB1B1428F, 0x7C7CD133, 0x88889B26, 0x3D3DA65F, 0xA1A1D7EC, 0xE4E4DF76,
150 0x8181942A, 0x91910149, 0x0F0FFB81, 0xEEEEAA88, 0x161661EE, 0xD7D77321,
151 0x9797F5C4, 0xA5A5A81A, 0xFEFE3FEB, 0x6D6DB5D9, 0x7878AEC5, 0xC5C56D39,
152 0x1D1DE599, 0x7676A4CD, 0x3E3EDCAD, 0xCBCB6731, 0xB6B6478B, 0xEFEF5B01,
153 0x12121E18, 0x6060C523, 0x6A6AB0DD, 0x4D4DF61F, 0xCECEE94E, 0xDEDE7C2D,
154 0x55559DF9, 0x7E7E5A48, 0x2121B24F, 0x03037AF2, 0xA0A02665, 0x5E5E198E,
155 0x5A5A6678, 0x65654B5C, 0x62624E58, 0xFDFD4519, 0x0606F48D, 0x404086E5,
156 0xF2F2BE98, 0x3333AC57, 0x17179067, 0x05058E7F, 0xE8E85E05, 0x4F4F7D64,
157 0x89896AAF, 0x10109563, 0x74742FB6, 0x0A0A75FE, 0x5C5C92F5, 0x9B9B74B7,
158 0x2D2D333C, 0x3030D6A5, 0x2E2E49CE, 0x494989E9, 0x46467268, 0x77775544,
159 0xA8A8D8E0, 0x9696044D, 0x2828BD43, 0xA9A92969, 0xD9D97929, 0x8686912E,
160 0xD1D187AC, 0xF4F44A15, 0x8D8D1559, 0xD6D682A8, 0xB9B9BC0A, 0x42420D9E,
161 0xF6F6C16E, 0x2F2FB847, 0xDDDD06DF, 0x23233934, 0xCCCC6235, 0xF1F1C46A,
162 0xC1C112CF, 0x8585EBDC, 0x8F8F9E22, 0x7171A1C9, 0x9090F0C0, 0xAAAA539B,
163 0x0101F189, 0x8B8BE1D4, 0x4E4E8CED, 0x8E8E6FAB, 0xABABA212, 0x6F6F3EA2,
164 0xE6E6540D, 0xDBDBF252, 0x92927BBB, 0xB7B7B602, 0x6969CA2F, 0x3939D9A9,
165 0xD3D30CD7, 0xA7A72361, 0xA2A2AD1E, 0xC3C399B4, 0x6C6C4450, 0x07070504,
166 0x04047FF6, 0x272746C2, 0xACACA716, 0xD0D07625, 0x50501386, 0xDCDCF756,
167 0x84841A55, 0xE1E15109, 0x7A7A25BE, 0x1313EF91},
168
169 {
170 0xA9D93939, 0x67901717, 0xB3719C9C, 0xE8D2A6A6, 0x04050707, 0xFD985252,
171 0xA3658080, 0x76DFE4E4, 0x9A084545, 0x92024B4B, 0x80A0E0E0, 0x78665A5A,
172 0xE4DDAFAF, 0xDDB06A6A, 0xD1BF6363, 0x38362A2A, 0x0D54E6E6, 0xC6432020,
173 0x3562CCCC, 0x98BEF2F2, 0x181E1212, 0xF724EBEB, 0xECD7A1A1, 0x6C774141,
174 0x43BD2828, 0x7532BCBC, 0x37D47B7B, 0x269B8888, 0xFA700D0D, 0x13F94444,
175 0x94B1FBFB, 0x485A7E7E, 0xF27A0303, 0xD0E48C8C, 0x8B47B6B6, 0x303C2424,
176 0x84A5E7E7, 0x54416B6B, 0xDF06DDDD, 0x23C56060, 0x1945FDFD, 0x5BA33A3A,
177 0x3D68C2C2, 0x59158D8D, 0xF321ECEC, 0xAE316666, 0xA23E6F6F, 0x82165757,
178 0x63951010, 0x015BEFEF, 0x834DB8B8, 0x2E918686, 0xD9B56D6D, 0x511F8383,
179 0x9B53AAAA, 0x7C635D5D, 0xA63B6868, 0xEB3FFEFE, 0xA5D63030, 0xBE257A7A,
180 0x16A7ACAC, 0x0C0F0909, 0xE335F0F0, 0x6123A7A7, 0xC0F09090, 0x8CAFE9E9,
181 0x3A809D9D, 0xF5925C5C, 0x73810C0C, 0x2C273131, 0x2576D0D0, 0x0BE75656,
182 0xBB7B9292, 0x4EE9CECE, 0x89F10101, 0x6B9F1E1E, 0x53A93434, 0x6AC4F1F1,
183 0xB499C3C3, 0xF1975B5B, 0xE1834747, 0xE66B1818, 0xBDC82222, 0x450E9898,
184 0xE26E1F1F, 0xF4C9B3B3, 0xB62F7474, 0x66CBF8F8, 0xCCFF9999, 0x95EA1414,
185 0x03ED5858, 0x56F7DCDC, 0xD4E18B8B, 0x1C1B1515, 0x1EADA2A2, 0xD70CD3D3,
186 0xFB2BE2E2, 0xC31DC8C8, 0x8E195E5E, 0xB5C22C2C, 0xE9894949, 0xCF12C1C1,
187 0xBF7E9595, 0xBA207D7D, 0xEA641111, 0x77840B0B, 0x396DC5C5, 0xAF6A8989,
188 0x33D17C7C, 0xC9A17171, 0x62CEFFFF, 0x7137BBBB, 0x81FB0F0F, 0x793DB5B5,
189 0x0951E1E1, 0xADDC3E3E, 0x242D3F3F, 0xCDA47676, 0xF99D5555, 0xD8EE8282,
190 0xE5864040, 0xC5AE7878, 0xB9CD2525, 0x4D049696, 0x44557777, 0x080A0E0E,
191 0x86135050, 0xE730F7F7, 0xA1D33737, 0x1D40FAFA, 0xAA346161, 0xED8C4E4E,
192 0x06B3B0B0, 0x706C5454, 0xB22A7373, 0xD2523B3B, 0x410B9F9F, 0x7B8B0202,
193 0xA088D8D8, 0x114FF3F3, 0x3167CBCB, 0xC2462727, 0x27C06767, 0x90B4FCFC,
194 0x20283838, 0xF67F0404, 0x60784848, 0xFF2EE5E5, 0x96074C4C, 0x5C4B6565,
195 0xB1C72B2B, 0xAB6F8E8E, 0x9E0D4242, 0x9CBBF5F5, 0x52F2DBDB, 0x1BF34A4A,
196 0x5FA63D3D, 0x9359A4A4, 0x0ABCB9B9, 0xEF3AF9F9, 0x91EF1313, 0x85FE0808,
197 0x49019191, 0xEE611616, 0x2D7CDEDE, 0x4FB22121, 0x8F42B1B1, 0x3BDB7272,
198 0x47B82F2F, 0x8748BFBF, 0x6D2CAEAE, 0x46E3C0C0, 0xD6573C3C, 0x3E859A9A,
199 0x6929A9A9, 0x647D4F4F, 0x2A948181, 0xCE492E2E, 0xCB17C6C6, 0x2FCA6969,
200 0xFCC3BDBD, 0x975CA3A3, 0x055EE8E8, 0x7AD0EDED, 0xAC87D1D1, 0x7F8E0505,
201 0xD5BA6464, 0x1AA8A5A5, 0x4BB72626, 0x0EB9BEBE, 0xA7608787, 0x5AF8D5D5,
202 0x28223636, 0x14111B1B, 0x3FDE7575, 0x2979D9D9, 0x88AAEEEE, 0x3C332D2D,
203 0x4C5F7979, 0x02B6B7B7, 0xB896CACA, 0xDA583535, 0xB09CC4C4, 0x17FC4343,
204 0x551A8484, 0x1FF64D4D, 0x8A1C5959, 0x7D38B2B2, 0x57AC3333, 0xC718CFCF,
205 0x8DF40606, 0x74695353, 0xB7749B9B, 0xC4F59797, 0x9F56ADAD, 0x72DAE3E3,
206 0x7ED5EAEA, 0x154AF4F4, 0x229E8F8F, 0x12A2ABAB, 0x584E6262, 0x07E85F5F,
207 0x99E51D1D, 0x34392323, 0x6EC1F6F6, 0x50446C6C, 0xDE5D3232, 0x68724646,
208 0x6526A0A0, 0xBC93CDCD, 0xDB03DADA, 0xF8C6BABA, 0xC8FA9E9E, 0xA882D6D6,
209 0x2BCF6E6E, 0x40507070, 0xDCEB8585, 0xFE750A0A, 0x328A9393, 0xA48DDFDF,
210 0xCA4C2929, 0x10141C1C, 0x2173D7D7, 0xF0CCB4B4, 0xD309D4D4, 0x5D108A8A,
211 0x0FE25151, 0x00000000, 0x6F9A1919, 0x9DE01A1A, 0x368F9494, 0x42E6C7C7,
212 0x4AECC9C9, 0x5EFDD2D2, 0xC1AB7F7F, 0xE0D8A8A8},
213
214 {
215 0xBC75BC32, 0xECF3EC21, 0x20C62043, 0xB3F4B3C9, 0xDADBDA03, 0x027B028B,
216 0xE2FBE22B, 0x9EC89EFA, 0xC94AC9EC, 0xD4D3D409, 0x18E6186B, 0x1E6B1E9F,
217 0x9845980E, 0xB27DB238, 0xA6E8A6D2, 0x264B26B7, 0x3CD63C57, 0x9332938A,
218 0x82D882EE, 0x52FD5298, 0x7B377BD4, 0xBB71BB37, 0x5BF15B97, 0x47E14783,
219 0x2430243C, 0x510F51E2, 0xBAF8BAC6, 0x4A1B4AF3, 0xBF87BF48, 0x0DFA0D70,
220 0xB006B0B3, 0x753F75DE, 0xD25ED2FD, 0x7DBA7D20, 0x66AE6631, 0x3A5B3AA3,
221 0x598A591C, 0x00000000, 0xCDBCCD93, 0x1A9D1AE0, 0xAE6DAE2C, 0x7FC17FAB,
222 0x2BB12BC7, 0xBE0EBEB9, 0xE080E0A0, 0x8A5D8A10, 0x3BD23B52, 0x64D564BA,
223 0xD8A0D888, 0xE784E7A5, 0x5F075FE8, 0x1B141B11, 0x2CB52CC2, 0xFC90FCB4,
224 0x312C3127, 0x80A38065, 0x73B2732A, 0x0C730C81, 0x794C795F, 0x6B546B41,
225 0x4B924B02, 0x53745369, 0x9436948F, 0x8351831F, 0x2A382A36, 0xC4B0C49C,
226 0x22BD22C8, 0xD55AD5F8, 0xBDFCBDC3, 0x48604878, 0xFF62FFCE, 0x4C964C07,
227 0x416C4177, 0xC742C7E6, 0xEBF7EB24, 0x1C101C14, 0x5D7C5D63, 0x36283622,
228 0x672767C0, 0xE98CE9AF, 0x441344F9, 0x149514EA, 0xF59CF5BB, 0xCFC7CF18,
229 0x3F243F2D, 0xC046C0E3, 0x723B72DB, 0x5470546C, 0x29CA294C, 0xF0E3F035,
230 0x088508FE, 0xC6CBC617, 0xF311F34F, 0x8CD08CE4, 0xA493A459, 0xCAB8CA96,
231 0x68A6683B, 0xB883B84D, 0x38203828, 0xE5FFE52E, 0xAD9FAD56, 0x0B770B84,
232 0xC8C3C81D, 0x99CC99FF, 0x580358ED, 0x196F199A, 0x0E080E0A, 0x95BF957E,
233 0x70407050, 0xF7E7F730, 0x6E2B6ECF, 0x1FE21F6E, 0xB579B53D, 0x090C090F,
234 0x61AA6134, 0x57825716, 0x9F419F0B, 0x9D3A9D80, 0x11EA1164, 0x25B925CD,
235 0xAFE4AFDD, 0x459A4508, 0xDFA4DF8D, 0xA397A35C, 0xEA7EEAD5, 0x35DA3558,
236 0xED7AEDD0, 0x431743FC, 0xF866F8CB, 0xFB94FBB1, 0x37A137D3, 0xFA1DFA40,
237 0xC23DC268, 0xB4F0B4CC, 0x32DE325D, 0x9CB39C71, 0x560B56E7, 0xE372E3DA,
238 0x87A78760, 0x151C151B, 0xF9EFF93A, 0x63D163BF, 0x345334A9, 0x9A3E9A85,
239 0xB18FB142, 0x7C337CD1, 0x8826889B, 0x3D5F3DA6, 0xA1ECA1D7, 0xE476E4DF,
240 0x812A8194, 0x91499101, 0x0F810FFB, 0xEE88EEAA, 0x16EE1661, 0xD721D773,
241 0x97C497F5, 0xA51AA5A8, 0xFEEBFE3F, 0x6DD96DB5, 0x78C578AE, 0xC539C56D,
242 0x1D991DE5, 0x76CD76A4, 0x3EAD3EDC, 0xCB31CB67, 0xB68BB647, 0xEF01EF5B,
243 0x1218121E, 0x602360C5, 0x6ADD6AB0, 0x4D1F4DF6, 0xCE4ECEE9, 0xDE2DDE7C,
244 0x55F9559D, 0x7E487E5A, 0x214F21B2, 0x03F2037A, 0xA065A026, 0x5E8E5E19,
245 0x5A785A66, 0x655C654B, 0x6258624E, 0xFD19FD45, 0x068D06F4, 0x40E54086,
246 0xF298F2BE, 0x335733AC, 0x17671790, 0x057F058E, 0xE805E85E, 0x4F644F7D,
247 0x89AF896A, 0x10631095, 0x74B6742F, 0x0AFE0A75, 0x5CF55C92, 0x9BB79B74,
248 0x2D3C2D33, 0x30A530D6, 0x2ECE2E49, 0x49E94989, 0x46684672, 0x77447755,
249 0xA8E0A8D8, 0x964D9604, 0x284328BD, 0xA969A929, 0xD929D979, 0x862E8691,
250 0xD1ACD187, 0xF415F44A, 0x8D598D15, 0xD6A8D682, 0xB90AB9BC, 0x429E420D,
251 0xF66EF6C1, 0x2F472FB8, 0xDDDFDD06, 0x23342339, 0xCC35CC62, 0xF16AF1C4,
252 0xC1CFC112, 0x85DC85EB, 0x8F228F9E, 0x71C971A1, 0x90C090F0, 0xAA9BAA53,
253 0x018901F1, 0x8BD48BE1, 0x4EED4E8C, 0x8EAB8E6F, 0xAB12ABA2, 0x6FA26F3E,
254 0xE60DE654, 0xDB52DBF2, 0x92BB927B, 0xB702B7B6, 0x692F69CA, 0x39A939D9,
255 0xD3D7D30C, 0xA761A723, 0xA21EA2AD, 0xC3B4C399, 0x6C506C44, 0x07040705,
256 0x04F6047F, 0x27C22746, 0xAC16ACA7, 0xD025D076, 0x50865013, 0xDC56DCF7,
257 0x8455841A, 0xE109E151, 0x7ABE7A25, 0x139113EF},
258
259 {
260 0xD939A9D9, 0x90176790, 0x719CB371, 0xD2A6E8D2, 0x05070405, 0x9852FD98,
261 0x6580A365, 0xDFE476DF, 0x08459A08, 0x024B9202, 0xA0E080A0, 0x665A7866,
262 0xDDAFE4DD, 0xB06ADDB0, 0xBF63D1BF, 0x362A3836, 0x54E60D54, 0x4320C643,
263 0x62CC3562, 0xBEF298BE, 0x1E12181E, 0x24EBF724, 0xD7A1ECD7, 0x77416C77,
264 0xBD2843BD, 0x32BC7532, 0xD47B37D4, 0x9B88269B, 0x700DFA70, 0xF94413F9,
265 0xB1FB94B1, 0x5A7E485A, 0x7A03F27A, 0xE48CD0E4, 0x47B68B47, 0x3C24303C,
266 0xA5E784A5, 0x416B5441, 0x06DDDF06, 0xC56023C5, 0x45FD1945, 0xA33A5BA3,
267 0x68C23D68, 0x158D5915, 0x21ECF321, 0x3166AE31, 0x3E6FA23E, 0x16578216,
268 0x95106395, 0x5BEF015B, 0x4DB8834D, 0x91862E91, 0xB56DD9B5, 0x1F83511F,
269 0x53AA9B53, 0x635D7C63, 0x3B68A63B, 0x3FFEEB3F, 0xD630A5D6, 0x257ABE25,
270 0xA7AC16A7, 0x0F090C0F, 0x35F0E335, 0x23A76123, 0xF090C0F0, 0xAFE98CAF,
271 0x809D3A80, 0x925CF592, 0x810C7381, 0x27312C27, 0x76D02576, 0xE7560BE7,
272 0x7B92BB7B, 0xE9CE4EE9, 0xF10189F1, 0x9F1E6B9F, 0xA93453A9, 0xC4F16AC4,
273 0x99C3B499, 0x975BF197, 0x8347E183, 0x6B18E66B, 0xC822BDC8, 0x0E98450E,
274 0x6E1FE26E, 0xC9B3F4C9, 0x2F74B62F, 0xCBF866CB, 0xFF99CCFF, 0xEA1495EA,
275 0xED5803ED, 0xF7DC56F7, 0xE18BD4E1, 0x1B151C1B, 0xADA21EAD, 0x0CD3D70C,
276 0x2BE2FB2B, 0x1DC8C31D, 0x195E8E19, 0xC22CB5C2, 0x8949E989, 0x12C1CF12,
277 0x7E95BF7E, 0x207DBA20, 0x6411EA64, 0x840B7784, 0x6DC5396D, 0x6A89AF6A,
278 0xD17C33D1, 0xA171C9A1, 0xCEFF62CE, 0x37BB7137, 0xFB0F81FB, 0x3DB5793D,
279 0x51E10951, 0xDC3EADDC, 0x2D3F242D, 0xA476CDA4, 0x9D55F99D, 0xEE82D8EE,
280 0x8640E586, 0xAE78C5AE, 0xCD25B9CD, 0x04964D04, 0x55774455, 0x0A0E080A,
281 0x13508613, 0x30F7E730, 0xD337A1D3, 0x40FA1D40, 0x3461AA34, 0x8C4EED8C,
282 0xB3B006B3, 0x6C54706C, 0x2A73B22A, 0x523BD252, 0x0B9F410B, 0x8B027B8B,
283 0x88D8A088, 0x4FF3114F, 0x67CB3167, 0x4627C246, 0xC06727C0, 0xB4FC90B4,
284 0x28382028, 0x7F04F67F, 0x78486078, 0x2EE5FF2E, 0x074C9607, 0x4B655C4B,
285 0xC72BB1C7, 0x6F8EAB6F, 0x0D429E0D, 0xBBF59CBB, 0xF2DB52F2, 0xF34A1BF3,
286 0xA63D5FA6, 0x59A49359, 0xBCB90ABC, 0x3AF9EF3A, 0xEF1391EF, 0xFE0885FE,
287 0x01914901, 0x6116EE61, 0x7CDE2D7C, 0xB2214FB2, 0x42B18F42, 0xDB723BDB,
288 0xB82F47B8, 0x48BF8748, 0x2CAE6D2C, 0xE3C046E3, 0x573CD657, 0x859A3E85,
289 0x29A96929, 0x7D4F647D, 0x94812A94, 0x492ECE49, 0x17C6CB17, 0xCA692FCA,
290 0xC3BDFCC3, 0x5CA3975C, 0x5EE8055E, 0xD0ED7AD0, 0x87D1AC87, 0x8E057F8E,
291 0xBA64D5BA, 0xA8A51AA8, 0xB7264BB7, 0xB9BE0EB9, 0x6087A760, 0xF8D55AF8,
292 0x22362822, 0x111B1411, 0xDE753FDE, 0x79D92979, 0xAAEE88AA, 0x332D3C33,
293 0x5F794C5F, 0xB6B702B6, 0x96CAB896, 0x5835DA58, 0x9CC4B09C, 0xFC4317FC,
294 0x1A84551A, 0xF64D1FF6, 0x1C598A1C, 0x38B27D38, 0xAC3357AC, 0x18CFC718,
295 0xF4068DF4, 0x69537469, 0x749BB774, 0xF597C4F5, 0x56AD9F56, 0xDAE372DA,
296 0xD5EA7ED5, 0x4AF4154A, 0x9E8F229E, 0xA2AB12A2, 0x4E62584E, 0xE85F07E8,
297 0xE51D99E5, 0x39233439, 0xC1F66EC1, 0x446C5044, 0x5D32DE5D, 0x72466872,
298 0x26A06526, 0x93CDBC93, 0x03DADB03, 0xC6BAF8C6, 0xFA9EC8FA, 0x82D6A882,
299 0xCF6E2BCF, 0x50704050, 0xEB85DCEB, 0x750AFE75, 0x8A93328A, 0x8DDFA48D,
300 0x4C29CA4C, 0x141C1014, 0x73D72173, 0xCCB4F0CC, 0x09D4D309, 0x108A5D10,
301 0xE2510FE2, 0x00000000, 0x9A196F9A, 0xE01A9DE0, 0x8F94368F, 0xE6C742E6,
302 0xECC94AEC, 0xFDD25EFD, 0xAB7FC1AB, 0xD8A8E0D8}
303};
304
305/* The exp_to_poly and poly_to_exp tables are used to perform efficient
306 * operations in GF(2^8) represented as GF(2)[x]/w(x) where
307 * w(x)=x^8+x^6+x^3+x^2+1. We care about doing that because it's part of the
308 * definition of the RS matrix in the key schedule. Elements of that field
309 * are polynomials of degree not greater than 7 and all coefficients 0 or 1,
310 * which can be represented naturally by bytes (just substitute x=2). In that
311 * form, GF(2^8) addition is the same as bitwise XOR, but GF(2^8)
312 * multiplication is inefficient without hardware support. To multiply
313 * faster, I make use of the fact x is a generator for the nonzero elements,
314 * so that every element p of GF(2)[x]/w(x) is either 0 or equal to (x)^n for
315 * some n in 0..254. Note that that caret is exponentiation in GF(2^8),
316 * *not* polynomial notation. So if I want to compute pq where p and q are
317 * in GF(2^8), I can just say:
318 * 1. if p=0 or q=0 then pq=0
319 * 2. otherwise, find m and n such that p=x^m and q=x^n
320 * 3. pq=(x^m)(x^n)=x^(m+n), so add m and n and find pq
321 * The translations in steps 2 and 3 are looked up in the tables
322 * poly_to_exp (for step 2) and exp_to_poly (for step 3). To see this
323 * in action, look at the CALC_S macro. As additional wrinkles, note that
324 * one of my operands is always a constant, so the poly_to_exp lookup on it
325 * is done in advance; I included the original values in the comments so
326 * readers can have some chance of recognizing that this *is* the RS matrix
327 * from the Twofish paper. I've only included the table entries I actually
328 * need; I never do a lookup on a variable input of zero and the biggest
329 * exponents I'll ever see are 254 (variable) and 237 (constant), so they'll
330 * never sum to more than 491. I'm repeating part of the exp_to_poly table
331 * so that I don't have to do mod-255 reduction in the exponent arithmetic.
332 * Since I know my constant operands are never zero, I only have to worry
333 * about zero values in the variable operand, and I do it with a simple
334 * conditional branch. I know conditionals are expensive, but I couldn't
335 * see a non-horrible way of avoiding them, and I did manage to group the
336 * statements so that each if covers four group multiplications. */
337
338static const u8 poly_to_exp[255] = {
339 0x00, 0x01, 0x17, 0x02, 0x2E, 0x18, 0x53, 0x03, 0x6A, 0x2F, 0x93, 0x19,
340 0x34, 0x54, 0x45, 0x04, 0x5C, 0x6B, 0xB6, 0x30, 0xA6, 0x94, 0x4B, 0x1A,
341 0x8C, 0x35, 0x81, 0x55, 0xAA, 0x46, 0x0D, 0x05, 0x24, 0x5D, 0x87, 0x6C,
342 0x9B, 0xB7, 0xC1, 0x31, 0x2B, 0xA7, 0xA3, 0x95, 0x98, 0x4C, 0xCA, 0x1B,
343 0xE6, 0x8D, 0x73, 0x36, 0xCD, 0x82, 0x12, 0x56, 0x62, 0xAB, 0xF0, 0x47,
344 0x4F, 0x0E, 0xBD, 0x06, 0xD4, 0x25, 0xD2, 0x5E, 0x27, 0x88, 0x66, 0x6D,
345 0xD6, 0x9C, 0x79, 0xB8, 0x08, 0xC2, 0xDF, 0x32, 0x68, 0x2C, 0xFD, 0xA8,
346 0x8A, 0xA4, 0x5A, 0x96, 0x29, 0x99, 0x22, 0x4D, 0x60, 0xCB, 0xE4, 0x1C,
347 0x7B, 0xE7, 0x3B, 0x8E, 0x9E, 0x74, 0xF4, 0x37, 0xD8, 0xCE, 0xF9, 0x83,
348 0x6F, 0x13, 0xB2, 0x57, 0xE1, 0x63, 0xDC, 0xAC, 0xC4, 0xF1, 0xAF, 0x48,
349 0x0A, 0x50, 0x42, 0x0F, 0xBA, 0xBE, 0xC7, 0x07, 0xDE, 0xD5, 0x78, 0x26,
350 0x65, 0xD3, 0xD1, 0x5F, 0xE3, 0x28, 0x21, 0x89, 0x59, 0x67, 0xFC, 0x6E,
351 0xB1, 0xD7, 0xF8, 0x9D, 0xF3, 0x7A, 0x3A, 0xB9, 0xC6, 0x09, 0x41, 0xC3,
352 0xAE, 0xE0, 0xDB, 0x33, 0x44, 0x69, 0x92, 0x2D, 0x52, 0xFE, 0x16, 0xA9,
353 0x0C, 0x8B, 0x80, 0xA5, 0x4A, 0x5B, 0xB5, 0x97, 0xC9, 0x2A, 0xA2, 0x9A,
354 0xC0, 0x23, 0x86, 0x4E, 0xBC, 0x61, 0xEF, 0xCC, 0x11, 0xE5, 0x72, 0x1D,
355 0x3D, 0x7C, 0xEB, 0xE8, 0xE9, 0x3C, 0xEA, 0x8F, 0x7D, 0x9F, 0xEC, 0x75,
356 0x1E, 0xF5, 0x3E, 0x38, 0xF6, 0xD9, 0x3F, 0xCF, 0x76, 0xFA, 0x1F, 0x84,
357 0xA0, 0x70, 0xED, 0x14, 0x90, 0xB3, 0x7E, 0x58, 0xFB, 0xE2, 0x20, 0x64,
358 0xD0, 0xDD, 0x77, 0xAD, 0xDA, 0xC5, 0x40, 0xF2, 0x39, 0xB0, 0xF7, 0x49,
359 0xB4, 0x0B, 0x7F, 0x51, 0x15, 0x43, 0x91, 0x10, 0x71, 0xBB, 0xEE, 0xBF,
360 0x85, 0xC8, 0xA1
361};
362
363static const u8 exp_to_poly[492] = {
364 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x4D, 0x9A, 0x79, 0xF2,
365 0xA9, 0x1F, 0x3E, 0x7C, 0xF8, 0xBD, 0x37, 0x6E, 0xDC, 0xF5, 0xA7, 0x03,
366 0x06, 0x0C, 0x18, 0x30, 0x60, 0xC0, 0xCD, 0xD7, 0xE3, 0x8B, 0x5B, 0xB6,
367 0x21, 0x42, 0x84, 0x45, 0x8A, 0x59, 0xB2, 0x29, 0x52, 0xA4, 0x05, 0x0A,
368 0x14, 0x28, 0x50, 0xA0, 0x0D, 0x1A, 0x34, 0x68, 0xD0, 0xED, 0x97, 0x63,
369 0xC6, 0xC1, 0xCF, 0xD3, 0xEB, 0x9B, 0x7B, 0xF6, 0xA1, 0x0F, 0x1E, 0x3C,
370 0x78, 0xF0, 0xAD, 0x17, 0x2E, 0x5C, 0xB8, 0x3D, 0x7A, 0xF4, 0xA5, 0x07,
371 0x0E, 0x1C, 0x38, 0x70, 0xE0, 0x8D, 0x57, 0xAE, 0x11, 0x22, 0x44, 0x88,
372 0x5D, 0xBA, 0x39, 0x72, 0xE4, 0x85, 0x47, 0x8E, 0x51, 0xA2, 0x09, 0x12,
373 0x24, 0x48, 0x90, 0x6D, 0xDA, 0xF9, 0xBF, 0x33, 0x66, 0xCC, 0xD5, 0xE7,
374 0x83, 0x4B, 0x96, 0x61, 0xC2, 0xC9, 0xDF, 0xF3, 0xAB, 0x1B, 0x36, 0x6C,
375 0xD8, 0xFD, 0xB7, 0x23, 0x46, 0x8C, 0x55, 0xAA, 0x19, 0x32, 0x64, 0xC8,
376 0xDD, 0xF7, 0xA3, 0x0B, 0x16, 0x2C, 0x58, 0xB0, 0x2D, 0x5A, 0xB4, 0x25,
377 0x4A, 0x94, 0x65, 0xCA, 0xD9, 0xFF, 0xB3, 0x2B, 0x56, 0xAC, 0x15, 0x2A,
378 0x54, 0xA8, 0x1D, 0x3A, 0x74, 0xE8, 0x9D, 0x77, 0xEE, 0x91, 0x6F, 0xDE,
379 0xF1, 0xAF, 0x13, 0x26, 0x4C, 0x98, 0x7D, 0xFA, 0xB9, 0x3F, 0x7E, 0xFC,
380 0xB5, 0x27, 0x4E, 0x9C, 0x75, 0xEA, 0x99, 0x7F, 0xFE, 0xB1, 0x2F, 0x5E,
381 0xBC, 0x35, 0x6A, 0xD4, 0xE5, 0x87, 0x43, 0x86, 0x41, 0x82, 0x49, 0x92,
382 0x69, 0xD2, 0xE9, 0x9F, 0x73, 0xE6, 0x81, 0x4F, 0x9E, 0x71, 0xE2, 0x89,
383 0x5F, 0xBE, 0x31, 0x62, 0xC4, 0xC5, 0xC7, 0xC3, 0xCB, 0xDB, 0xFB, 0xBB,
384 0x3B, 0x76, 0xEC, 0x95, 0x67, 0xCE, 0xD1, 0xEF, 0x93, 0x6B, 0xD6, 0xE1,
385 0x8F, 0x53, 0xA6, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x4D,
386 0x9A, 0x79, 0xF2, 0xA9, 0x1F, 0x3E, 0x7C, 0xF8, 0xBD, 0x37, 0x6E, 0xDC,
387 0xF5, 0xA7, 0x03, 0x06, 0x0C, 0x18, 0x30, 0x60, 0xC0, 0xCD, 0xD7, 0xE3,
388 0x8B, 0x5B, 0xB6, 0x21, 0x42, 0x84, 0x45, 0x8A, 0x59, 0xB2, 0x29, 0x52,
389 0xA4, 0x05, 0x0A, 0x14, 0x28, 0x50, 0xA0, 0x0D, 0x1A, 0x34, 0x68, 0xD0,
390 0xED, 0x97, 0x63, 0xC6, 0xC1, 0xCF, 0xD3, 0xEB, 0x9B, 0x7B, 0xF6, 0xA1,
391 0x0F, 0x1E, 0x3C, 0x78, 0xF0, 0xAD, 0x17, 0x2E, 0x5C, 0xB8, 0x3D, 0x7A,
392 0xF4, 0xA5, 0x07, 0x0E, 0x1C, 0x38, 0x70, 0xE0, 0x8D, 0x57, 0xAE, 0x11,
393 0x22, 0x44, 0x88, 0x5D, 0xBA, 0x39, 0x72, 0xE4, 0x85, 0x47, 0x8E, 0x51,
394 0xA2, 0x09, 0x12, 0x24, 0x48, 0x90, 0x6D, 0xDA, 0xF9, 0xBF, 0x33, 0x66,
395 0xCC, 0xD5, 0xE7, 0x83, 0x4B, 0x96, 0x61, 0xC2, 0xC9, 0xDF, 0xF3, 0xAB,
396 0x1B, 0x36, 0x6C, 0xD8, 0xFD, 0xB7, 0x23, 0x46, 0x8C, 0x55, 0xAA, 0x19,
397 0x32, 0x64, 0xC8, 0xDD, 0xF7, 0xA3, 0x0B, 0x16, 0x2C, 0x58, 0xB0, 0x2D,
398 0x5A, 0xB4, 0x25, 0x4A, 0x94, 0x65, 0xCA, 0xD9, 0xFF, 0xB3, 0x2B, 0x56,
399 0xAC, 0x15, 0x2A, 0x54, 0xA8, 0x1D, 0x3A, 0x74, 0xE8, 0x9D, 0x77, 0xEE,
400 0x91, 0x6F, 0xDE, 0xF1, 0xAF, 0x13, 0x26, 0x4C, 0x98, 0x7D, 0xFA, 0xB9,
401 0x3F, 0x7E, 0xFC, 0xB5, 0x27, 0x4E, 0x9C, 0x75, 0xEA, 0x99, 0x7F, 0xFE,
402 0xB1, 0x2F, 0x5E, 0xBC, 0x35, 0x6A, 0xD4, 0xE5, 0x87, 0x43, 0x86, 0x41,
403 0x82, 0x49, 0x92, 0x69, 0xD2, 0xE9, 0x9F, 0x73, 0xE6, 0x81, 0x4F, 0x9E,
404 0x71, 0xE2, 0x89, 0x5F, 0xBE, 0x31, 0x62, 0xC4, 0xC5, 0xC7, 0xC3, 0xCB
405};
406
407
408/* The table constants are indices of
409 * S-box entries, preprocessed through q0 and q1. */
410static const u8 calc_sb_tbl[512] = {
411 0xA9, 0x75, 0x67, 0xF3, 0xB3, 0xC6, 0xE8, 0xF4,
412 0x04, 0xDB, 0xFD, 0x7B, 0xA3, 0xFB, 0x76, 0xC8,
413 0x9A, 0x4A, 0x92, 0xD3, 0x80, 0xE6, 0x78, 0x6B,
414 0xE4, 0x45, 0xDD, 0x7D, 0xD1, 0xE8, 0x38, 0x4B,
415 0x0D, 0xD6, 0xC6, 0x32, 0x35, 0xD8, 0x98, 0xFD,
416 0x18, 0x37, 0xF7, 0x71, 0xEC, 0xF1, 0x6C, 0xE1,
417 0x43, 0x30, 0x75, 0x0F, 0x37, 0xF8, 0x26, 0x1B,
418 0xFA, 0x87, 0x13, 0xFA, 0x94, 0x06, 0x48, 0x3F,
419 0xF2, 0x5E, 0xD0, 0xBA, 0x8B, 0xAE, 0x30, 0x5B,
420 0x84, 0x8A, 0x54, 0x00, 0xDF, 0xBC, 0x23, 0x9D,
421 0x19, 0x6D, 0x5B, 0xC1, 0x3D, 0xB1, 0x59, 0x0E,
422 0xF3, 0x80, 0xAE, 0x5D, 0xA2, 0xD2, 0x82, 0xD5,
423 0x63, 0xA0, 0x01, 0x84, 0x83, 0x07, 0x2E, 0x14,
424 0xD9, 0xB5, 0x51, 0x90, 0x9B, 0x2C, 0x7C, 0xA3,
425 0xA6, 0xB2, 0xEB, 0x73, 0xA5, 0x4C, 0xBE, 0x54,
426 0x16, 0x92, 0x0C, 0x74, 0xE3, 0x36, 0x61, 0x51,
427 0xC0, 0x38, 0x8C, 0xB0, 0x3A, 0xBD, 0xF5, 0x5A,
428 0x73, 0xFC, 0x2C, 0x60, 0x25, 0x62, 0x0B, 0x96,
429 0xBB, 0x6C, 0x4E, 0x42, 0x89, 0xF7, 0x6B, 0x10,
430 0x53, 0x7C, 0x6A, 0x28, 0xB4, 0x27, 0xF1, 0x8C,
431 0xE1, 0x13, 0xE6, 0x95, 0xBD, 0x9C, 0x45, 0xC7,
432 0xE2, 0x24, 0xF4, 0x46, 0xB6, 0x3B, 0x66, 0x70,
433 0xCC, 0xCA, 0x95, 0xE3, 0x03, 0x85, 0x56, 0xCB,
434 0xD4, 0x11, 0x1C, 0xD0, 0x1E, 0x93, 0xD7, 0xB8,
435 0xFB, 0xA6, 0xC3, 0x83, 0x8E, 0x20, 0xB5, 0xFF,
436 0xE9, 0x9F, 0xCF, 0x77, 0xBF, 0xC3, 0xBA, 0xCC,
437 0xEA, 0x03, 0x77, 0x6F, 0x39, 0x08, 0xAF, 0xBF,
438 0x33, 0x40, 0xC9, 0xE7, 0x62, 0x2B, 0x71, 0xE2,
439 0x81, 0x79, 0x79, 0x0C, 0x09, 0xAA, 0xAD, 0x82,
440 0x24, 0x41, 0xCD, 0x3A, 0xF9, 0xEA, 0xD8, 0xB9,
441 0xE5, 0xE4, 0xC5, 0x9A, 0xB9, 0xA4, 0x4D, 0x97,
442 0x44, 0x7E, 0x08, 0xDA, 0x86, 0x7A, 0xE7, 0x17,
443 0xA1, 0x66, 0x1D, 0x94, 0xAA, 0xA1, 0xED, 0x1D,
444 0x06, 0x3D, 0x70, 0xF0, 0xB2, 0xDE, 0xD2, 0xB3,
445 0x41, 0x0B, 0x7B, 0x72, 0xA0, 0xA7, 0x11, 0x1C,
446 0x31, 0xEF, 0xC2, 0xD1, 0x27, 0x53, 0x90, 0x3E,
447 0x20, 0x8F, 0xF6, 0x33, 0x60, 0x26, 0xFF, 0x5F,
448 0x96, 0xEC, 0x5C, 0x76, 0xB1, 0x2A, 0xAB, 0x49,
449 0x9E, 0x81, 0x9C, 0x88, 0x52, 0xEE, 0x1B, 0x21,
450 0x5F, 0xC4, 0x93, 0x1A, 0x0A, 0xEB, 0xEF, 0xD9,
451 0x91, 0xC5, 0x85, 0x39, 0x49, 0x99, 0xEE, 0xCD,
452 0x2D, 0xAD, 0x4F, 0x31, 0x8F, 0x8B, 0x3B, 0x01,
453 0x47, 0x18, 0x87, 0x23, 0x6D, 0xDD, 0x46, 0x1F,
454 0xD6, 0x4E, 0x3E, 0x2D, 0x69, 0xF9, 0x64, 0x48,
455 0x2A, 0x4F, 0xCE, 0xF2, 0xCB, 0x65, 0x2F, 0x8E,
456 0xFC, 0x78, 0x97, 0x5C, 0x05, 0x58, 0x7A, 0x19,
457 0xAC, 0x8D, 0x7F, 0xE5, 0xD5, 0x98, 0x1A, 0x57,
458 0x4B, 0x67, 0x0E, 0x7F, 0xA7, 0x05, 0x5A, 0x64,
459 0x28, 0xAF, 0x14, 0x63, 0x3F, 0xB6, 0x29, 0xFE,
460 0x88, 0xF5, 0x3C, 0xB7, 0x4C, 0x3C, 0x02, 0xA5,
461 0xB8, 0xCE, 0xDA, 0xE9, 0xB0, 0x68, 0x17, 0x44,
462 0x55, 0xE0, 0x1F, 0x4D, 0x8A, 0x43, 0x7D, 0x69,
463 0x57, 0x29, 0xC7, 0x2E, 0x8D, 0xAC, 0x74, 0x15,
464 0xB7, 0x59, 0xC4, 0xA8, 0x9F, 0x0A, 0x72, 0x9E,
465 0x7E, 0x6E, 0x15, 0x47, 0x22, 0xDF, 0x12, 0x34,
466 0x58, 0x35, 0x07, 0x6A, 0x99, 0xCF, 0x34, 0xDC,
467 0x6E, 0x22, 0x50, 0xC9, 0xDE, 0xC0, 0x68, 0x9B,
468 0x65, 0x89, 0xBC, 0xD4, 0xDB, 0xED, 0xF8, 0xAB,
469 0xC8, 0x12, 0xA8, 0xA2, 0x2B, 0x0D, 0x40, 0x52,
470 0xDC, 0xBB, 0xFE, 0x02, 0x32, 0x2F, 0xA4, 0xA9,
471 0xCA, 0xD7, 0x10, 0x61, 0x21, 0x1E, 0xF0, 0xB4,
472 0xD3, 0x50, 0x5D, 0x04, 0x0F, 0xF6, 0x00, 0xC2,
473 0x6F, 0x16, 0x9D, 0x25, 0x36, 0x86, 0x42, 0x56,
474 0x4A, 0x55, 0x5E, 0x09, 0xC1, 0xBE, 0xE0, 0x91
475};
476
477/* Macro to perform one column of the RS matrix multiplication. The
478 * parameters a, b, c, and d are the four bytes of output; i is the index
479 * of the key bytes, and w, x, y, and z, are the column of constants from
480 * the RS matrix, preprocessed through the poly_to_exp table. */
481
482#define CALC_S(a, b, c, d, i, w, x, y, z) \
483 if (key[i]) { \
484 tmp = poly_to_exp[key[i] - 1]; \
485 (a) ^= exp_to_poly[tmp + (w)]; \
486 (b) ^= exp_to_poly[tmp + (x)]; \
487 (c) ^= exp_to_poly[tmp + (y)]; \
488 (d) ^= exp_to_poly[tmp + (z)]; \
489 }
490
491/* Macros to calculate the key-dependent S-boxes for a 128-bit key using
492 * the S vector from CALC_S. CALC_SB_2 computes a single entry in all
493 * four S-boxes, where i is the index of the entry to compute, and a and b
494 * are the index numbers preprocessed through the q0 and q1 tables
495 * respectively. */
496
497#define CALC_SB_2(i, a, b) \
498 ctx->s[0][i] = mds[0][q0[(a) ^ sa] ^ se]; \
499 ctx->s[1][i] = mds[1][q0[(b) ^ sb] ^ sf]; \
500 ctx->s[2][i] = mds[2][q1[(a) ^ sc] ^ sg]; \
501 ctx->s[3][i] = mds[3][q1[(b) ^ sd] ^ sh]
502
503/* Macro exactly like CALC_SB_2, but for 192-bit keys. */
504
505#define CALC_SB192_2(i, a, b) \
506 ctx->s[0][i] = mds[0][q0[q0[(b) ^ sa] ^ se] ^ si]; \
507 ctx->s[1][i] = mds[1][q0[q1[(b) ^ sb] ^ sf] ^ sj]; \
508 ctx->s[2][i] = mds[2][q1[q0[(a) ^ sc] ^ sg] ^ sk]; \
509 ctx->s[3][i] = mds[3][q1[q1[(a) ^ sd] ^ sh] ^ sl];
510
511/* Macro exactly like CALC_SB_2, but for 256-bit keys. */
512
513#define CALC_SB256_2(i, a, b) \
514 ctx->s[0][i] = mds[0][q0[q0[q1[(b) ^ sa] ^ se] ^ si] ^ sm]; \
515 ctx->s[1][i] = mds[1][q0[q1[q1[(a) ^ sb] ^ sf] ^ sj] ^ sn]; \
516 ctx->s[2][i] = mds[2][q1[q0[q0[(a) ^ sc] ^ sg] ^ sk] ^ so]; \
517 ctx->s[3][i] = mds[3][q1[q1[q0[(b) ^ sd] ^ sh] ^ sl] ^ sp];
518
519/* Macros to calculate the whitening and round subkeys. CALC_K_2 computes the
520 * last two stages of the h() function for a given index (either 2i or 2i+1).
521 * a, b, c, and d are the four bytes going into the last two stages. For
522 * 128-bit keys, this is the entire h() function and a and c are the index
523 * preprocessed through q0 and q1 respectively; for longer keys they are the
524 * output of previous stages. j is the index of the first key byte to use.
525 * CALC_K computes a pair of subkeys for 128-bit Twofish, by calling CALC_K_2
526 * twice, doing the Pseudo-Hadamard Transform, and doing the necessary
527 * rotations. Its parameters are: a, the array to write the results into,
528 * j, the index of the first output entry, k and l, the preprocessed indices
529 * for index 2i, and m and n, the preprocessed indices for index 2i+1.
530 * CALC_K192_2 expands CALC_K_2 to handle 192-bit keys, by doing an
531 * additional lookup-and-XOR stage. The parameters a, b, c and d are the
532 * four bytes going into the last three stages. For 192-bit keys, c = d
533 * are the index preprocessed through q0, and a = b are the index
534 * preprocessed through q1; j is the index of the first key byte to use.
535 * CALC_K192 is identical to CALC_K but for using the CALC_K192_2 macro
536 * instead of CALC_K_2.
537 * CALC_K256_2 expands CALC_K192_2 to handle 256-bit keys, by doing an
538 * additional lookup-and-XOR stage. The parameters a and b are the index
539 * preprocessed through q0 and q1 respectively; j is the index of the first
540 * key byte to use. CALC_K256 is identical to CALC_K but for using the
541 * CALC_K256_2 macro instead of CALC_K_2. */
542
543#define CALC_K_2(a, b, c, d, j) \
544 mds[0][q0[a ^ key[(j) + 8]] ^ key[j]] \
545 ^ mds[1][q0[b ^ key[(j) + 9]] ^ key[(j) + 1]] \
546 ^ mds[2][q1[c ^ key[(j) + 10]] ^ key[(j) + 2]] \
547 ^ mds[3][q1[d ^ key[(j) + 11]] ^ key[(j) + 3]]
548
549#define CALC_K(a, j, k, l, m, n) \
550 x = CALC_K_2 (k, l, k, l, 0); \
551 y = CALC_K_2 (m, n, m, n, 4); \
552 y = rol32(y, 8); \
553 x += y; y += x; ctx->a[j] = x; \
554 ctx->a[(j) + 1] = rol32(y, 9)
555
556#define CALC_K192_2(a, b, c, d, j) \
557 CALC_K_2 (q0[a ^ key[(j) + 16]], \
558 q1[b ^ key[(j) + 17]], \
559 q0[c ^ key[(j) + 18]], \
560 q1[d ^ key[(j) + 19]], j)
561
562#define CALC_K192(a, j, k, l, m, n) \
563 x = CALC_K192_2 (l, l, k, k, 0); \
564 y = CALC_K192_2 (n, n, m, m, 4); \
565 y = rol32(y, 8); \
566 x += y; y += x; ctx->a[j] = x; \
567 ctx->a[(j) + 1] = rol32(y, 9)
568
569#define CALC_K256_2(a, b, j) \
570 CALC_K192_2 (q1[b ^ key[(j) + 24]], \
571 q1[a ^ key[(j) + 25]], \
572 q0[a ^ key[(j) + 26]], \
573 q0[b ^ key[(j) + 27]], j)
574
575#define CALC_K256(a, j, k, l, m, n) \
576 x = CALC_K256_2 (k, l, 0); \
577 y = CALC_K256_2 (m, n, 4); \
578 y = rol32(y, 8); \
579 x += y; y += x; ctx->a[j] = x; \
580 ctx->a[(j) + 1] = rol32(y, 9)
581
582/* Perform the key setup. */
583int __twofish_setkey(struct twofish_ctx *ctx, const u8 *key,
584 unsigned int key_len, u32 *flags)
585{
586 int i, j, k;
587
588 /* Temporaries for CALC_K. */
589 u32 x, y;
590
591 /* The S vector used to key the S-boxes, split up into individual bytes.
592 * 128-bit keys use only sa through sh; 256-bit use all of them. */
593 u8 sa = 0, sb = 0, sc = 0, sd = 0, se = 0, sf = 0, sg = 0, sh = 0;
594 u8 si = 0, sj = 0, sk = 0, sl = 0, sm = 0, sn = 0, so = 0, sp = 0;
595
596 /* Temporary for CALC_S. */
597 u8 tmp;
598
599 /* Check key length. */
600 if (key_len % 8)
601 {
602 *flags |= CRYPTO_TFM_RES_BAD_KEY_LEN;
603 return -EINVAL; /* unsupported key length */
604 }
605
606 /* Compute the first two words of the S vector. The magic numbers are
607 * the entries of the RS matrix, preprocessed through poly_to_exp. The
608 * numbers in the comments are the original (polynomial form) matrix
609 * entries. */
610 CALC_S (sa, sb, sc, sd, 0, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
611 CALC_S (sa, sb, sc, sd, 1, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
612 CALC_S (sa, sb, sc, sd, 2, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
613 CALC_S (sa, sb, sc, sd, 3, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
614 CALC_S (sa, sb, sc, sd, 4, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
615 CALC_S (sa, sb, sc, sd, 5, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
616 CALC_S (sa, sb, sc, sd, 6, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
617 CALC_S (sa, sb, sc, sd, 7, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
618 CALC_S (se, sf, sg, sh, 8, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
619 CALC_S (se, sf, sg, sh, 9, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
620 CALC_S (se, sf, sg, sh, 10, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
621 CALC_S (se, sf, sg, sh, 11, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
622 CALC_S (se, sf, sg, sh, 12, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
623 CALC_S (se, sf, sg, sh, 13, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
624 CALC_S (se, sf, sg, sh, 14, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
625 CALC_S (se, sf, sg, sh, 15, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
626
627 if (key_len == 24 || key_len == 32) { /* 192- or 256-bit key */
628 /* Calculate the third word of the S vector */
629 CALC_S (si, sj, sk, sl, 16, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
630 CALC_S (si, sj, sk, sl, 17, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
631 CALC_S (si, sj, sk, sl, 18, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
632 CALC_S (si, sj, sk, sl, 19, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
633 CALC_S (si, sj, sk, sl, 20, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
634 CALC_S (si, sj, sk, sl, 21, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
635 CALC_S (si, sj, sk, sl, 22, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
636 CALC_S (si, sj, sk, sl, 23, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
637 }
638
639 if (key_len == 32) { /* 256-bit key */
640 /* Calculate the fourth word of the S vector */
641 CALC_S (sm, sn, so, sp, 24, 0x00, 0x2D, 0x01, 0x2D); /* 01 A4 02 A4 */
642 CALC_S (sm, sn, so, sp, 25, 0x2D, 0xA4, 0x44, 0x8A); /* A4 56 A1 55 */
643 CALC_S (sm, sn, so, sp, 26, 0x8A, 0xD5, 0xBF, 0xD1); /* 55 82 FC 87 */
644 CALC_S (sm, sn, so, sp, 27, 0xD1, 0x7F, 0x3D, 0x99); /* 87 F3 C1 5A */
645 CALC_S (sm, sn, so, sp, 28, 0x99, 0x46, 0x66, 0x96); /* 5A 1E 47 58 */
646 CALC_S (sm, sn, so, sp, 29, 0x96, 0x3C, 0x5B, 0xED); /* 58 C6 AE DB */
647 CALC_S (sm, sn, so, sp, 30, 0xED, 0x37, 0x4F, 0xE0); /* DB 68 3D 9E */
648 CALC_S (sm, sn, so, sp, 31, 0xE0, 0xD0, 0x8C, 0x17); /* 9E E5 19 03 */
649
650 /* Compute the S-boxes. */
651 for ( i = j = 0, k = 1; i < 256; i++, j += 2, k += 2 ) {
652 CALC_SB256_2( i, calc_sb_tbl[j], calc_sb_tbl[k] );
653 }
654
655 /* CALC_K256/CALC_K192/CALC_K loops were unrolled.
656 * Unrolling produced x2.5 more code (+18k on i386),
657 * and speeded up key setup by 7%:
658 * unrolled: twofish_setkey/sec: 41128
659 * loop: twofish_setkey/sec: 38148
660 * CALC_K256: ~100 insns each
661 * CALC_K192: ~90 insns
662 * CALC_K: ~70 insns
663 */
664 /* Calculate whitening and round subkeys */
665 for ( i = 0; i < 8; i += 2 ) {
666 CALC_K256 (w, i, q0[i], q1[i], q0[i+1], q1[i+1]);
667 }
668 for ( i = 0; i < 32; i += 2 ) {
669 CALC_K256 (k, i, q0[i+8], q1[i+8], q0[i+9], q1[i+9]);
670 }
671 } else if (key_len == 24) { /* 192-bit key */
672 /* Compute the S-boxes. */
673 for ( i = j = 0, k = 1; i < 256; i++, j += 2, k += 2 ) {
674 CALC_SB192_2( i, calc_sb_tbl[j], calc_sb_tbl[k] );
675 }
676
677 /* Calculate whitening and round subkeys */
678 for ( i = 0; i < 8; i += 2 ) {
679 CALC_K192 (w, i, q0[i], q1[i], q0[i+1], q1[i+1]);
680 }
681 for ( i = 0; i < 32; i += 2 ) {
682 CALC_K192 (k, i, q0[i+8], q1[i+8], q0[i+9], q1[i+9]);
683 }
684 } else { /* 128-bit key */
685 /* Compute the S-boxes. */
686 for ( i = j = 0, k = 1; i < 256; i++, j += 2, k += 2 ) {
687 CALC_SB_2( i, calc_sb_tbl[j], calc_sb_tbl[k] );
688 }
689
690 /* Calculate whitening and round subkeys */
691 for ( i = 0; i < 8; i += 2 ) {
692 CALC_K (w, i, q0[i], q1[i], q0[i+1], q1[i+1]);
693 }
694 for ( i = 0; i < 32; i += 2 ) {
695 CALC_K (k, i, q0[i+8], q1[i+8], q0[i+9], q1[i+9]);
696 }
697 }
698
699 return 0;
700}
701EXPORT_SYMBOL_GPL(__twofish_setkey);
702
703int twofish_setkey(struct crypto_tfm *tfm, const u8 *key, unsigned int key_len)
704{
705 return __twofish_setkey(crypto_tfm_ctx(tfm), key, key_len,
706 &tfm->crt_flags);
707}
708EXPORT_SYMBOL_GPL(twofish_setkey);
709
710MODULE_LICENSE("GPL");
711MODULE_DESCRIPTION("Twofish cipher common functions");