1 /* SPDX-License-Identifier: GPL-2.0 */ 1
2 /*
3 * Hardware-accelerated CRC-32 variants for Li
4 *
5 * Use the z/Architecture Vector Extension Fac
6 * computing of CRC-32 checksums.
7 *
8 * This CRC-32 implementation algorithm proces
9 * bit first (BE).
10 *
11 * Copyright IBM Corp. 2015
12 * Author(s): Hendrik Brueckner <brueckner@lin
13 */
14
15 #include <linux/types.h>
16 #include <asm/fpu.h>
17 #include "crc32-vx.h"
18
19 /* Vector register range containing CRC-32 con
20 #define CONST_R1R2 9
21 #define CONST_R3R4 10
22 #define CONST_R5 11
23 #define CONST_R6 12
24 #define CONST_RU_POLY 13
25 #define CONST_CRC_POLY 14
26
27 /*
28 * The CRC-32 constant block contains reductio
29 * process particular chunks of the input data
30 *
31 * For the CRC-32 variants, the constants are
32 * these definitions:
33 *
34 * R1 = x4*128+64 mod P(x)
35 * R2 = x4*128 mod P(x)
36 * R3 = x128+64 mod P(x)
37 * R4 = x128 mod P(x)
38 * R5 = x96 mod P(x)
39 * R6 = x64 mod P(x)
40 *
41 * Barret reduction constant, u, is defin
42 *
43 * where P(x) is the polynomial in the no
44 * polynomial in the reversed (bitreflect
45 *
46 * Note that the constant definitions below ar
47 * intermediate results with a single VECTOR G
48 * The rightmost doubleword can be 0 to preven
49 * can be multiplied by 1 to perform an XOR wi
50 * VECTOR EXCLUSIVE OR instruction.
51 *
52 * CRC-32 (IEEE 802.3 Ethernet, ...) polynomia
53 *
54 * P(x) = 0x04C11DB7
55 * P'(x) = 0xEDB88320
56 */
57
58 static unsigned long constants_CRC_32_BE[] = {
59 0x08833794c, 0x0e6228b11, /* R1,
60 0x0c5b9cd4c, 0x0e8a45605, /* R3,
61 0x0f200aa66, 1UL << 32, /* R5,
62 0x0490d678d, 1, /* R6,
63 0x104d101df, 0, /* u *
64 0x104C11DB7, 0, /* P(x
65 };
66
67 /**
68 * crc32_be_vgfm_16 - Compute CRC-32 (BE varia
69 * @crc: Initial CRC value, typically ~0.
70 * @buf: Input buffer pointer, performance mig
71 * buffer is on a doubleword boundary.
72 * @size: Size of the buffer, must be 64 bytes
73 *
74 * Register usage:
75 * V0: Initial CRC value and intermed
76 * V1..V4: Data for CRC computation.
77 * V5..V8: Next data chunks that are fetc
78 * V9..V14: CRC-32 constants.
79 */
80 u32 crc32_be_vgfm_16(u32 crc, unsigned char co
81 {
82 /* Load CRC-32 constants */
83 fpu_vlm(CONST_R1R2, CONST_CRC_POLY, &c
84 fpu_vzero(0);
85
86 /* Load the initial CRC value into the
87 fpu_vlvgf(0, crc, 0);
88
89 /* Load a 64-byte data chunk and XOR w
90 fpu_vlm(1, 4, buf);
91 fpu_vx(1, 0, 1);
92 buf += 64;
93 size -= 64;
94
95 while (size >= 64) {
96 /* Load the next 64-byte data
97 fpu_vlm(5, 8, buf);
98
99 /*
100 * Perform a GF(2) multiplicat
101 * the reduction constants in
102 * then folded (accumulated) w
103 * stored in V1. Repeat this
104 * in V2, V3, and V4 respectiv
105 */
106 fpu_vgfmag(1, CONST_R1R2, 1, 5
107 fpu_vgfmag(2, CONST_R1R2, 2, 6
108 fpu_vgfmag(3, CONST_R1R2, 3, 7
109 fpu_vgfmag(4, CONST_R1R2, 4, 8
110 buf += 64;
111 size -= 64;
112 }
113
114 /* Fold V1 to V4 into a single 128-bit
115 fpu_vgfmag(1, CONST_R3R4, 1, 2);
116 fpu_vgfmag(1, CONST_R3R4, 1, 3);
117 fpu_vgfmag(1, CONST_R3R4, 1, 4);
118
119 while (size >= 16) {
120 fpu_vl(2, buf);
121 fpu_vgfmag(1, CONST_R3R4, 1, 2
122 buf += 16;
123 size -= 16;
124 }
125
126 /*
127 * The R5 constant is used to fold a 1
128 * that is XORed with the next 96-bit
129 * VGFMG instruction, multiply the rig
130 * form an intermediate 96-bit value (
131 * XORed with the intermediate reducti
132 */
133 fpu_vgfmg(1, CONST_R5, 1);
134
135 /*
136 * Further reduce the remaining 96-bit
137 * single VGFMG, the rightmost doublew
138 * intermediate result is then XORed w
139 * doubleword with R6. The result is
140 * the Barret reduction.
141 */
142 fpu_vgfmg(1, CONST_R6, 1);
143
144 /*
145 * The input values to the Barret redu
146 * in V1 (R(x)), degree-32 generator p
147 * constant u. The Barret reduction r
148 * P(x).
149 *
150 * The Barret reduction algorithm is d
151 *
152 * 1. T1(x) = floor( R(x) / x^32 )
153 * 2. T2(x) = floor( T1(x) / x^32 )
154 * 3. C(x) = R(x) XOR T2(x) mod x^
155 *
156 * Note: To compensate the division by
157 * instruction to move the leftmost wo
158 * of the vector register. The rightm
159 * with zero to not contribute to the
160 */
161
162 /* T1(x) = floor( R(x) / x^32 ) GF2MUL
163 fpu_vupllf(2, 1);
164 fpu_vgfmg(2, CONST_RU_POLY, 2);
165
166 /*
167 * Compute the GF(2) product of the CR
168 * V2 and XOR the intermediate result,
169 * The final result is in the rightmos
170 */
171 fpu_vupllf(2, 2);
172 fpu_vgfmag(2, CONST_CRC_POLY, 2, 1);
173 return fpu_vlgvf(2, 3);
174 }
175
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.