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/linkage.h>
16#include <asm/nospec-insn.h>
17#include <asm/vx-insn.h>
18
19/* Vector register range containing CRC-32 constants */
20#define CONST_R1R2		%v9
21#define CONST_R3R4		%v10
22#define CONST_R5		%v11
23#define CONST_R6		%v12
24#define CONST_RU_POLY		%v13
25#define CONST_CRC_POLY		%v14
26
27.data
28.align 8
29
30/*
31 * The CRC-32 constant block contains reduction constants to fold and
32 * process particular chunks of the input data stream in parallel.
33 *
34 * For the CRC-32 variants, the constants are precomputed according to
35 * these definitions:
36 *
37 *	R1 = x4*128+64 mod P(x)
38 *	R2 = x4*128    mod P(x)
39 *	R3 = x128+64   mod P(x)
40 *	R4 = x128      mod P(x)
41 *	R5 = x96       mod P(x)
42 *	R6 = x64       mod P(x)
43 *
44 *	Barret reduction constant, u, is defined as floor(x**64 / P(x)).
45 *
46 *	where P(x) is the polynomial in the normal domain and the P'(x) is the
47 *	polynomial in the reversed (bitreflected) domain.
48 *
49 * Note that the constant definitions below are extended in order to compute
50 * intermediate results with a single VECTOR GALOIS FIELD MULTIPLY instruction.
51 * The righmost doubleword can be 0 to prevent contribution to the result or
52 * can be multiplied by 1 to perform an XOR without the need for a separate
53 * VECTOR EXCLUSIVE OR instruction.
54 *
55 * CRC-32 (IEEE 802.3 Ethernet, ...) polynomials:
56 *
57 *	P(x)  = 0x04C11DB7
58 *	P'(x) = 0xEDB88320
59 */
60
61.Lconstants_CRC_32_BE:
62	.quad		0x08833794c, 0x0e6228b11	# R1, R2
63	.quad		0x0c5b9cd4c, 0x0e8a45605	# R3, R4
64	.quad		0x0f200aa66, 1 << 32		# R5, x32
65	.quad		0x0490d678d, 1			# R6, 1
66	.quad		0x104d101df, 0			# u
67	.quad		0x104C11DB7, 0			# P(x)
68
69.previous
70
71	GEN_BR_THUNK %r14
72
73.text
74/*
75 * The CRC-32 function(s) use these calling conventions:
76 *
77 * Parameters:
78 *
79 *	%r2:	Initial CRC value, typically ~0; and final CRC (return) value.
80 *	%r3:	Input buffer pointer, performance might be improved if the
81 *		buffer is on a doubleword boundary.
82 *	%r4:	Length of the buffer, must be 64 bytes or greater.
83 *
84 * Register usage:
85 *
86 *	%r5:	CRC-32 constant pool base pointer.
87 *	V0:	Initial CRC value and intermediate constants and results.
88 *	V1..V4:	Data for CRC computation.
89 *	V5..V8:	Next data chunks that are fetched from the input buffer.
90 *
91 *	V9..V14: CRC-32 constants.
92 */
93ENTRY(crc32_be_vgfm_16)
94	/* Load CRC-32 constants */
95	larl	%r5,.Lconstants_CRC_32_BE
96	VLM	CONST_R1R2,CONST_CRC_POLY,0,%r5
97
98	/* Load the initial CRC value into the leftmost word of V0. */
99	VZERO	%v0
100	VLVGF	%v0,%r2,0
101
102	/* Load a 64-byte data chunk and XOR with CRC */
103	VLM	%v1,%v4,0,%r3		/* 64-bytes into V1..V4 */
104	VX	%v1,%v0,%v1		/* V1 ^= CRC */
105	aghi	%r3,64			/* BUF = BUF + 64 */
106	aghi	%r4,-64			/* LEN = LEN - 64 */
107
108	/* Check remaining buffer size and jump to proper folding method */
109	cghi	%r4,64
110	jl	.Lless_than_64bytes
111
112.Lfold_64bytes_loop:
113	/* Load the next 64-byte data chunk into V5 to V8 */
114	VLM	%v5,%v8,0,%r3
115
116	/*
117	 * Perform a GF(2) multiplication of the doublewords in V1 with
118	 * the reduction constants in V0.  The intermediate result is
119	 * then folded (accumulated) with the next data chunk in V5 and
120	 * stored in V1.  Repeat this step for the register contents
121	 * in V2, V3, and V4 respectively.
122	 */
123	VGFMAG	%v1,CONST_R1R2,%v1,%v5
124	VGFMAG	%v2,CONST_R1R2,%v2,%v6
125	VGFMAG	%v3,CONST_R1R2,%v3,%v7
126	VGFMAG	%v4,CONST_R1R2,%v4,%v8
127
128	/* Adjust buffer pointer and length for next loop */
129	aghi	%r3,64			/* BUF = BUF + 64 */
130	aghi	%r4,-64			/* LEN = LEN - 64 */
131
132	cghi	%r4,64
133	jnl	.Lfold_64bytes_loop
134
135.Lless_than_64bytes:
136	/* Fold V1 to V4 into a single 128-bit value in V1 */
137	VGFMAG	%v1,CONST_R3R4,%v1,%v2
138	VGFMAG	%v1,CONST_R3R4,%v1,%v3
139	VGFMAG	%v1,CONST_R3R4,%v1,%v4
140
141	/* Check whether to continue with 64-bit folding */
142	cghi	%r4,16
143	jl	.Lfinal_fold
144
145.Lfold_16bytes_loop:
146
147	VL	%v2,0,,%r3		/* Load next data chunk */
148	VGFMAG	%v1,CONST_R3R4,%v1,%v2	/* Fold next data chunk */
149
150	/* Adjust buffer pointer and size for folding next data chunk */
151	aghi	%r3,16
152	aghi	%r4,-16
153
154	/* Process remaining data chunks */
155	cghi	%r4,16
156	jnl	.Lfold_16bytes_loop
157
158.Lfinal_fold:
159	/*
160	 * The R5 constant is used to fold a 128-bit value into an 96-bit value
161	 * that is XORed with the next 96-bit input data chunk.  To use a single
162	 * VGFMG instruction, multiply the rightmost 64-bit with x^32 (1<<32) to
163	 * form an intermediate 96-bit value (with appended zeros) which is then
164	 * XORed with the intermediate reduction result.
165	 */
166	VGFMG	%v1,CONST_R5,%v1
167
168	/*
169	 * Further reduce the remaining 96-bit value to a 64-bit value using a
170	 * single VGFMG, the rightmost doubleword is multiplied with 0x1. The
171	 * intermediate result is then XORed with the product of the leftmost
172	 * doubleword with R6.	The result is a 64-bit value and is subject to
173	 * the Barret reduction.
174	 */
175	VGFMG	%v1,CONST_R6,%v1
176
177	/*
178	 * The input values to the Barret reduction are the degree-63 polynomial
179	 * in V1 (R(x)), degree-32 generator polynomial, and the reduction
180	 * constant u.	The Barret reduction result is the CRC value of R(x) mod
181	 * P(x).
182	 *
183	 * The Barret reduction algorithm is defined as:
184	 *
185	 *    1. T1(x) = floor( R(x) / x^32 ) GF2MUL u
186	 *    2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x)
187	 *    3. C(x)  = R(x) XOR T2(x) mod x^32
188	 *
189	 * Note: To compensate the division by x^32, use the vector unpack
190	 * instruction to move the leftmost word into the leftmost doubleword
191	 * of the vector register.  The rightmost doubleword is multiplied
192	 * with zero to not contribute to the intermediate results.
193	 */
194
195	/* T1(x) = floor( R(x) / x^32 ) GF2MUL u */
196	VUPLLF	%v2,%v1
197	VGFMG	%v2,CONST_RU_POLY,%v2
198
199	/*
200	 * Compute the GF(2) product of the CRC polynomial in VO with T1(x) in
201	 * V2 and XOR the intermediate result, T2(x),  with the value in V1.
202	 * The final result is in the rightmost word of V2.
203	 */
204	VUPLLF	%v2,%v2
205	VGFMAG	%v2,CONST_CRC_POLY,%v2,%v1
206
207.Ldone:
208	VLGVF	%r2,%v2,3
209	BR_EX	%r14
210ENDPROC(crc32_be_vgfm_16)
211
212.previous
213