xref: /linux/arch/s390/crypto/crc32be-vx.c (revision 79790b6818e96c58fe2bffee1b418c16e64e7b80)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Hardware-accelerated CRC-32 variants for Linux on z Systems
4  *
5  * Use the z/Architecture Vector Extension Facility to accelerate the
6  * computing of CRC-32 checksums.
7  *
8  * This CRC-32 implementation algorithm processes the most-significant
9  * bit first (BE).
10  *
11  * Copyright IBM Corp. 2015
12  * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
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 constants */
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 reduction constants to fold and
29  * process particular chunks of the input data stream in parallel.
30  *
31  * For the CRC-32 variants, the constants are precomputed according to
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 defined as floor(x**64 / P(x)).
42  *
43  *	where P(x) is the polynomial in the normal domain and the P'(x) is the
44  *	polynomial in the reversed (bitreflected) domain.
45  *
46  * Note that the constant definitions below are extended in order to compute
47  * intermediate results with a single VECTOR GALOIS FIELD MULTIPLY instruction.
48  * The rightmost doubleword can be 0 to prevent contribution to the result or
49  * can be multiplied by 1 to perform an XOR without the need for a separate
50  * VECTOR EXCLUSIVE OR instruction.
51  *
52  * CRC-32 (IEEE 802.3 Ethernet, ...) polynomials:
53  *
54  *	P(x)  = 0x04C11DB7
55  *	P'(x) = 0xEDB88320
56  */
57 
58 static unsigned long constants_CRC_32_BE[] = {
59 	0x08833794c, 0x0e6228b11,	/* R1, R2 */
60 	0x0c5b9cd4c, 0x0e8a45605,	/* R3, R4 */
61 	0x0f200aa66, 1UL << 32,		/* R5, x32 */
62 	0x0490d678d, 1,			/* R6, 1 */
63 	0x104d101df, 0,			/* u */
64 	0x104C11DB7, 0,			/* P(x) */
65 };
66 
67 /**
68  * crc32_be_vgfm_16 - Compute CRC-32 (BE variant) with vector registers
69  * @crc: Initial CRC value, typically ~0.
70  * @buf: Input buffer pointer, performance might be improved if the
71  *	  buffer is on a doubleword boundary.
72  * @size: Size of the buffer, must be 64 bytes or greater.
73  *
74  * Register usage:
75  *	V0:	Initial CRC value and intermediate constants and results.
76  *	V1..V4:	Data for CRC computation.
77  *	V5..V8:	Next data chunks that are fetched from the input buffer.
78  *	V9..V14: CRC-32 constants.
79  */
crc32_be_vgfm_16(u32 crc,unsigned char const * buf,size_t size)80 u32 crc32_be_vgfm_16(u32 crc, unsigned char const *buf, size_t size)
81 {
82 	/* Load CRC-32 constants */
83 	fpu_vlm(CONST_R1R2, CONST_CRC_POLY, &constants_CRC_32_BE);
84 	fpu_vzero(0);
85 
86 	/* Load the initial CRC value into the leftmost word of V0. */
87 	fpu_vlvgf(0, crc, 0);
88 
89 	/* Load a 64-byte data chunk and XOR with CRC */
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 chunk into V5 to V8 */
97 		fpu_vlm(5, 8, buf);
98 
99 		/*
100 		 * Perform a GF(2) multiplication of the doublewords in V1 with
101 		 * the reduction constants in V0.  The intermediate result is
102 		 * then folded (accumulated) with the next data chunk in V5 and
103 		 * stored in V1.  Repeat this step for the register contents
104 		 * in V2, V3, and V4 respectively.
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 value in V1 */
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 128-bit value into an 96-bit value
128 	 * that is XORed with the next 96-bit input data chunk.  To use a single
129 	 * VGFMG instruction, multiply the rightmost 64-bit with x^32 (1<<32) to
130 	 * form an intermediate 96-bit value (with appended zeros) which is then
131 	 * XORed with the intermediate reduction result.
132 	 */
133 	fpu_vgfmg(1, CONST_R5, 1);
134 
135 	/*
136 	 * Further reduce the remaining 96-bit value to a 64-bit value using a
137 	 * single VGFMG, the rightmost doubleword is multiplied with 0x1. The
138 	 * intermediate result is then XORed with the product of the leftmost
139 	 * doubleword with R6.	The result is a 64-bit value and is subject to
140 	 * the Barret reduction.
141 	 */
142 	fpu_vgfmg(1, CONST_R6, 1);
143 
144 	/*
145 	 * The input values to the Barret reduction are the degree-63 polynomial
146 	 * in V1 (R(x)), degree-32 generator polynomial, and the reduction
147 	 * constant u.	The Barret reduction result is the CRC value of R(x) mod
148 	 * P(x).
149 	 *
150 	 * The Barret reduction algorithm is defined as:
151 	 *
152 	 *    1. T1(x) = floor( R(x) / x^32 ) GF2MUL u
153 	 *    2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x)
154 	 *    3. C(x)  = R(x) XOR T2(x) mod x^32
155 	 *
156 	 * Note: To compensate the division by x^32, use the vector unpack
157 	 * instruction to move the leftmost word into the leftmost doubleword
158 	 * of the vector register.  The rightmost doubleword is multiplied
159 	 * with zero to not contribute to the intermediate results.
160 	 */
161 
162 	/* T1(x) = floor( R(x) / x^32 ) GF2MUL u */
163 	fpu_vupllf(2, 1);
164 	fpu_vgfmg(2, CONST_RU_POLY, 2);
165 
166 	/*
167 	 * Compute the GF(2) product of the CRC polynomial in VO with T1(x) in
168 	 * V2 and XOR the intermediate result, T2(x),  with the value in V1.
169 	 * The final result is in the rightmost word of V2.
170 	 */
171 	fpu_vupllf(2, 2);
172 	fpu_vgfmag(2, CONST_CRC_POLY, 2, 1);
173 	return fpu_vlgvf(2, 3);
174 }
175