1/* -*- linux-c -*- ------------------------------------------------------- * 2 * 3 * Copyright 2002-2004 H. Peter Anvin - All Rights Reserved 4 * 5 * This program is free software; you can redistribute it and/or modify 6 * it under the terms of the GNU General Public License as published by 7 * the Free Software Foundation, Inc., 53 Temple Place Ste 330, 8 * Boston MA 02111-1307, USA; either version 2 of the License, or 9 * (at your option) any later version; incorporated herein by reference. 10 * 11 * ----------------------------------------------------------------------- */ 12 13/* 14 * int$#.c 15 * 16 * $#-way unrolled portable integer math RAID-6 instruction set 17 * 18 * This file is postprocessed using unroll.awk 19 */ 20 21#include <linux/raid/pq.h> 22 23/* 24 * This is the C data type to use 25 */ 26 27/* Change this from BITS_PER_LONG if there is something better... */ 28#if BITS_PER_LONG == 64 29# define NBYTES(x) ((x) * 0x0101010101010101UL) 30# define NSIZE 8 31# define NSHIFT 3 32# define NSTRING "64" 33typedef u64 unative_t; 34#else 35# define NBYTES(x) ((x) * 0x01010101U) 36# define NSIZE 4 37# define NSHIFT 2 38# define NSTRING "32" 39typedef u32 unative_t; 40#endif 41 42 43 44/* 45 * IA-64 wants insane amounts of unrolling. On other architectures that 46 * is just a waste of space. 47 */ 48#if ($# <= 8) || defined(__ia64__) 49 50 51/* 52 * These sub-operations are separate inlines since they can sometimes be 53 * specially optimized using architecture-specific hacks. 54 */ 55 56/* 57 * The SHLBYTE() operation shifts each byte left by 1, *not* 58 * rolling over into the next byte 59 */ 60static inline __attribute_const__ unative_t SHLBYTE(unative_t v) 61{ 62 unative_t vv; 63 64 vv = (v << 1) & NBYTES(0xfe); 65 return vv; 66} 67 68/* 69 * The MASK() operation returns 0xFF in any byte for which the high 70 * bit is 1, 0x00 for any byte for which the high bit is 0. 71 */ 72static inline __attribute_const__ unative_t MASK(unative_t v) 73{ 74 unative_t vv; 75 76 vv = v & NBYTES(0x80); 77 vv = (vv << 1) - (vv >> 7); /* Overflow on the top bit is OK */ 78 return vv; 79} 80 81 82static void raid6_int$#_gen_syndrome(int disks, size_t bytes, void **ptrs) 83{ 84 u8 **dptr = (u8 **)ptrs; 85 u8 *p, *q; 86 int d, z, z0; 87 88 unative_t wd$$, wq$$, wp$$, w1$$, w2$$; 89 90 z0 = disks - 3; /* Highest data disk */ 91 p = dptr[z0+1]; /* XOR parity */ 92 q = dptr[z0+2]; /* RS syndrome */ 93 94 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) { 95 wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE]; 96 for ( z = z0-1 ; z >= 0 ; z-- ) { 97 wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE]; 98 wp$$ ^= wd$$; 99 w2$$ = MASK(wq$$); 100 w1$$ = SHLBYTE(wq$$); 101 w2$$ &= NBYTES(0x1d); 102 w1$$ ^= w2$$; 103 wq$$ = w1$$ ^ wd$$; 104 } 105 *(unative_t *)&p[d+NSIZE*$$] = wp$$; 106 *(unative_t *)&q[d+NSIZE*$$] = wq$$; 107 } 108} 109 110static void raid6_int$#_xor_syndrome(int disks, int start, int stop, 111 size_t bytes, void **ptrs) 112{ 113 u8 **dptr = (u8 **)ptrs; 114 u8 *p, *q; 115 int d, z, z0; 116 117 unative_t wd$$, wq$$, wp$$, w1$$, w2$$; 118 119 z0 = stop; /* P/Q right side optimization */ 120 p = dptr[disks-2]; /* XOR parity */ 121 q = dptr[disks-1]; /* RS syndrome */ 122 123 for ( d = 0 ; d < bytes ; d += NSIZE*$# ) { 124 /* P/Q data pages */ 125 wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE]; 126 for ( z = z0-1 ; z >= start ; z-- ) { 127 wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE]; 128 wp$$ ^= wd$$; 129 w2$$ = MASK(wq$$); 130 w1$$ = SHLBYTE(wq$$); 131 w2$$ &= NBYTES(0x1d); 132 w1$$ ^= w2$$; 133 wq$$ = w1$$ ^ wd$$; 134 } 135 /* P/Q left side optimization */ 136 for ( z = start-1 ; z >= 0 ; z-- ) { 137 w2$$ = MASK(wq$$); 138 w1$$ = SHLBYTE(wq$$); 139 w2$$ &= NBYTES(0x1d); 140 wq$$ = w1$$ ^ w2$$; 141 } 142 *(unative_t *)&p[d+NSIZE*$$] ^= wp$$; 143 *(unative_t *)&q[d+NSIZE*$$] ^= wq$$; 144 } 145 146} 147 148const struct raid6_calls raid6_intx$# = { 149 raid6_int$#_gen_syndrome, 150 raid6_int$#_xor_syndrome, 151 NULL, /* always valid */ 152 "int" NSTRING "x$#", 153 0 154}; 155 156#endif 157