crc32le-vx.S (7965B)
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 bitreflected CRC-32 checksums for IEEE 802.3 Ethernet 7 * and Castagnoli. 8 * 9 * This CRC-32 implementation algorithm is bitreflected and processes 10 * the least-significant bit first (Little-Endian). 11 * 12 * Copyright IBM Corp. 2015 13 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com> 14 */ 15 16#include <linux/linkage.h> 17#include <asm/nospec-insn.h> 18#include <asm/vx-insn.h> 19 20/* Vector register range containing CRC-32 constants */ 21#define CONST_PERM_LE2BE %v9 22#define CONST_R2R1 %v10 23#define CONST_R4R3 %v11 24#define CONST_R5 %v12 25#define CONST_RU_POLY %v13 26#define CONST_CRC_POLY %v14 27 28.data 29.align 8 30 31/* 32 * The CRC-32 constant block contains reduction constants to fold and 33 * process particular chunks of the input data stream in parallel. 34 * 35 * For the CRC-32 variants, the constants are precomputed according to 36 * these definitions: 37 * 38 * R1 = [(x4*128+32 mod P'(x) << 32)]' << 1 39 * R2 = [(x4*128-32 mod P'(x) << 32)]' << 1 40 * R3 = [(x128+32 mod P'(x) << 32)]' << 1 41 * R4 = [(x128-32 mod P'(x) << 32)]' << 1 42 * R5 = [(x64 mod P'(x) << 32)]' << 1 43 * R6 = [(x32 mod P'(x) << 32)]' << 1 44 * 45 * The bitreflected Barret reduction constant, u', is defined as 46 * the bit reversal of floor(x**64 / P(x)). 47 * 48 * where P(x) is the polynomial in the normal domain and the P'(x) is the 49 * polynomial in the reversed (bitreflected) domain. 50 * 51 * CRC-32 (IEEE 802.3 Ethernet, ...) polynomials: 52 * 53 * P(x) = 0x04C11DB7 54 * P'(x) = 0xEDB88320 55 * 56 * CRC-32C (Castagnoli) polynomials: 57 * 58 * P(x) = 0x1EDC6F41 59 * P'(x) = 0x82F63B78 60 */ 61 62.Lconstants_CRC_32_LE: 63 .octa 0x0F0E0D0C0B0A09080706050403020100 # BE->LE mask 64 .quad 0x1c6e41596, 0x154442bd4 # R2, R1 65 .quad 0x0ccaa009e, 0x1751997d0 # R4, R3 66 .octa 0x163cd6124 # R5 67 .octa 0x1F7011641 # u' 68 .octa 0x1DB710641 # P'(x) << 1 69 70.Lconstants_CRC_32C_LE: 71 .octa 0x0F0E0D0C0B0A09080706050403020100 # BE->LE mask 72 .quad 0x09e4addf8, 0x740eef02 # R2, R1 73 .quad 0x14cd00bd6, 0xf20c0dfe # R4, R3 74 .octa 0x0dd45aab8 # R5 75 .octa 0x0dea713f1 # u' 76 .octa 0x105ec76f0 # P'(x) << 1 77 78.previous 79 80 GEN_BR_THUNK %r14 81 82.text 83 84/* 85 * The CRC-32 functions use these calling conventions: 86 * 87 * Parameters: 88 * 89 * %r2: Initial CRC value, typically ~0; and final CRC (return) value. 90 * %r3: Input buffer pointer, performance might be improved if the 91 * buffer is on a doubleword boundary. 92 * %r4: Length of the buffer, must be 64 bytes or greater. 93 * 94 * Register usage: 95 * 96 * %r5: CRC-32 constant pool base pointer. 97 * V0: Initial CRC value and intermediate constants and results. 98 * V1..V4: Data for CRC computation. 99 * V5..V8: Next data chunks that are fetched from the input buffer. 100 * V9: Constant for BE->LE conversion and shift operations 101 * 102 * V10..V14: CRC-32 constants. 103 */ 104 105ENTRY(crc32_le_vgfm_16) 106 larl %r5,.Lconstants_CRC_32_LE 107 j crc32_le_vgfm_generic 108ENDPROC(crc32_le_vgfm_16) 109 110ENTRY(crc32c_le_vgfm_16) 111 larl %r5,.Lconstants_CRC_32C_LE 112 j crc32_le_vgfm_generic 113ENDPROC(crc32c_le_vgfm_16) 114 115ENTRY(crc32_le_vgfm_generic) 116 /* Load CRC-32 constants */ 117 VLM CONST_PERM_LE2BE,CONST_CRC_POLY,0,%r5 118 119 /* 120 * Load the initial CRC value. 121 * 122 * The CRC value is loaded into the rightmost word of the 123 * vector register and is later XORed with the LSB portion 124 * of the loaded input data. 125 */ 126 VZERO %v0 /* Clear V0 */ 127 VLVGF %v0,%r2,3 /* Load CRC into rightmost word */ 128 129 /* Load a 64-byte data chunk and XOR with CRC */ 130 VLM %v1,%v4,0,%r3 /* 64-bytes into V1..V4 */ 131 VPERM %v1,%v1,%v1,CONST_PERM_LE2BE 132 VPERM %v2,%v2,%v2,CONST_PERM_LE2BE 133 VPERM %v3,%v3,%v3,CONST_PERM_LE2BE 134 VPERM %v4,%v4,%v4,CONST_PERM_LE2BE 135 136 VX %v1,%v0,%v1 /* V1 ^= CRC */ 137 aghi %r3,64 /* BUF = BUF + 64 */ 138 aghi %r4,-64 /* LEN = LEN - 64 */ 139 140 cghi %r4,64 141 jl .Lless_than_64bytes 142 143.Lfold_64bytes_loop: 144 /* Load the next 64-byte data chunk into V5 to V8 */ 145 VLM %v5,%v8,0,%r3 146 VPERM %v5,%v5,%v5,CONST_PERM_LE2BE 147 VPERM %v6,%v6,%v6,CONST_PERM_LE2BE 148 VPERM %v7,%v7,%v7,CONST_PERM_LE2BE 149 VPERM %v8,%v8,%v8,CONST_PERM_LE2BE 150 151 /* 152 * Perform a GF(2) multiplication of the doublewords in V1 with 153 * the R1 and R2 reduction constants in V0. The intermediate result 154 * is then folded (accumulated) with the next data chunk in V5 and 155 * stored in V1. Repeat this step for the register contents 156 * in V2, V3, and V4 respectively. 157 */ 158 VGFMAG %v1,CONST_R2R1,%v1,%v5 159 VGFMAG %v2,CONST_R2R1,%v2,%v6 160 VGFMAG %v3,CONST_R2R1,%v3,%v7 161 VGFMAG %v4,CONST_R2R1,%v4,%v8 162 163 aghi %r3,64 /* BUF = BUF + 64 */ 164 aghi %r4,-64 /* LEN = LEN - 64 */ 165 166 cghi %r4,64 167 jnl .Lfold_64bytes_loop 168 169.Lless_than_64bytes: 170 /* 171 * Fold V1 to V4 into a single 128-bit value in V1. Multiply V1 with R3 172 * and R4 and accumulating the next 128-bit chunk until a single 128-bit 173 * value remains. 174 */ 175 VGFMAG %v1,CONST_R4R3,%v1,%v2 176 VGFMAG %v1,CONST_R4R3,%v1,%v3 177 VGFMAG %v1,CONST_R4R3,%v1,%v4 178 179 cghi %r4,16 180 jl .Lfinal_fold 181 182.Lfold_16bytes_loop: 183 184 VL %v2,0,,%r3 /* Load next data chunk */ 185 VPERM %v2,%v2,%v2,CONST_PERM_LE2BE 186 VGFMAG %v1,CONST_R4R3,%v1,%v2 /* Fold next data chunk */ 187 188 aghi %r3,16 189 aghi %r4,-16 190 191 cghi %r4,16 192 jnl .Lfold_16bytes_loop 193 194.Lfinal_fold: 195 /* 196 * Set up a vector register for byte shifts. The shift value must 197 * be loaded in bits 1-4 in byte element 7 of a vector register. 198 * Shift by 8 bytes: 0x40 199 * Shift by 4 bytes: 0x20 200 */ 201 VLEIB %v9,0x40,7 202 203 /* 204 * Prepare V0 for the next GF(2) multiplication: shift V0 by 8 bytes 205 * to move R4 into the rightmost doubleword and set the leftmost 206 * doubleword to 0x1. 207 */ 208 VSRLB %v0,CONST_R4R3,%v9 209 VLEIG %v0,1,0 210 211 /* 212 * Compute GF(2) product of V1 and V0. The rightmost doubleword 213 * of V1 is multiplied with R4. The leftmost doubleword of V1 is 214 * multiplied by 0x1 and is then XORed with rightmost product. 215 * Implicitly, the intermediate leftmost product becomes padded 216 */ 217 VGFMG %v1,%v0,%v1 218 219 /* 220 * Now do the final 32-bit fold by multiplying the rightmost word 221 * in V1 with R5 and XOR the result with the remaining bits in V1. 222 * 223 * To achieve this by a single VGFMAG, right shift V1 by a word 224 * and store the result in V2 which is then accumulated. Use the 225 * vector unpack instruction to load the rightmost half of the 226 * doubleword into the rightmost doubleword element of V1; the other 227 * half is loaded in the leftmost doubleword. 228 * The vector register with CONST_R5 contains the R5 constant in the 229 * rightmost doubleword and the leftmost doubleword is zero to ignore 230 * the leftmost product of V1. 231 */ 232 VLEIB %v9,0x20,7 /* Shift by words */ 233 VSRLB %v2,%v1,%v9 /* Store remaining bits in V2 */ 234 VUPLLF %v1,%v1 /* Split rightmost doubleword */ 235 VGFMAG %v1,CONST_R5,%v1,%v2 /* V1 = (V1 * R5) XOR V2 */ 236 237 /* 238 * Apply a Barret reduction to compute the final 32-bit CRC value. 239 * 240 * The input values to the Barret reduction are the degree-63 polynomial 241 * in V1 (R(x)), degree-32 generator polynomial, and the reduction 242 * constant u. The Barret reduction result is the CRC value of R(x) mod 243 * P(x). 244 * 245 * The Barret reduction algorithm is defined as: 246 * 247 * 1. T1(x) = floor( R(x) / x^32 ) GF2MUL u 248 * 2. T2(x) = floor( T1(x) / x^32 ) GF2MUL P(x) 249 * 3. C(x) = R(x) XOR T2(x) mod x^32 250 * 251 * Note: The leftmost doubleword of vector register containing 252 * CONST_RU_POLY is zero and, thus, the intermediate GF(2) product 253 * is zero and does not contribute to the final result. 254 */ 255 256 /* T1(x) = floor( R(x) / x^32 ) GF2MUL u */ 257 VUPLLF %v2,%v1 258 VGFMG %v2,CONST_RU_POLY,%v2 259 260 /* 261 * Compute the GF(2) product of the CRC polynomial with T1(x) in 262 * V2 and XOR the intermediate result, T2(x), with the value in V1. 263 * The final result is stored in word element 2 of V2. 264 */ 265 VUPLLF %v2,%v2 266 VGFMAG %v2,CONST_CRC_POLY,%v2,%v1 267 268.Ldone: 269 VLGVF %r2,%v2,2 270 BR_EX %r14 271ENDPROC(crc32_le_vgfm_generic) 272 273.previous