1 /* -*- linux-c -*- ------------------------------------------------------- * 2 * 3 * Copyright 2002 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 * raid6/sse2.c 15 * 16 * SSE-2 implementation of RAID-6 syndrome functions 17 * 18 */ 19 20 #include <linux/raid/pq.h> 21 #include "x86.h" 22 23 static const struct raid6_sse_constants { 24 u64 x1d[2]; 25 } raid6_sse_constants __attribute__((aligned(16))) = { 26 { 0x1d1d1d1d1d1d1d1dULL, 0x1d1d1d1d1d1d1d1dULL }, 27 }; 28 29 static int raid6_have_sse2(void) 30 { 31 /* Not really boot_cpu but "all_cpus" */ 32 return boot_cpu_has(X86_FEATURE_MMX) && 33 boot_cpu_has(X86_FEATURE_FXSR) && 34 boot_cpu_has(X86_FEATURE_XMM) && 35 boot_cpu_has(X86_FEATURE_XMM2); 36 } 37 38 /* 39 * Plain SSE2 implementation 40 */ 41 static void raid6_sse21_gen_syndrome(int disks, size_t bytes, void **ptrs) 42 { 43 u8 **dptr = (u8 **)ptrs; 44 u8 *p, *q; 45 int d, z, z0; 46 47 z0 = disks - 3; /* Highest data disk */ 48 p = dptr[z0+1]; /* XOR parity */ 49 q = dptr[z0+2]; /* RS syndrome */ 50 51 kernel_fpu_begin(); 52 53 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0])); 54 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */ 55 56 for ( d = 0 ; d < bytes ; d += 16 ) { 57 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d])); 58 asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */ 59 asm volatile("prefetchnta %0" : : "m" (dptr[z0-1][d])); 60 asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */ 61 asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z0-1][d])); 62 for ( z = z0-2 ; z >= 0 ; z-- ) { 63 asm volatile("prefetchnta %0" : : "m" (dptr[z][d])); 64 asm volatile("pcmpgtb %xmm4,%xmm5"); 65 asm volatile("paddb %xmm4,%xmm4"); 66 asm volatile("pand %xmm0,%xmm5"); 67 asm volatile("pxor %xmm5,%xmm4"); 68 asm volatile("pxor %xmm5,%xmm5"); 69 asm volatile("pxor %xmm6,%xmm2"); 70 asm volatile("pxor %xmm6,%xmm4"); 71 asm volatile("movdqa %0,%%xmm6" : : "m" (dptr[z][d])); 72 } 73 asm volatile("pcmpgtb %xmm4,%xmm5"); 74 asm volatile("paddb %xmm4,%xmm4"); 75 asm volatile("pand %xmm0,%xmm5"); 76 asm volatile("pxor %xmm5,%xmm4"); 77 asm volatile("pxor %xmm5,%xmm5"); 78 asm volatile("pxor %xmm6,%xmm2"); 79 asm volatile("pxor %xmm6,%xmm4"); 80 81 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d])); 82 asm volatile("pxor %xmm2,%xmm2"); 83 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d])); 84 asm volatile("pxor %xmm4,%xmm4"); 85 } 86 87 asm volatile("sfence" : : : "memory"); 88 kernel_fpu_end(); 89 } 90 91 const struct raid6_calls raid6_sse2x1 = { 92 raid6_sse21_gen_syndrome, 93 NULL, /* XOR not yet implemented */ 94 raid6_have_sse2, 95 "sse2x1", 96 1 /* Has cache hints */ 97 }; 98 99 /* 100 * Unrolled-by-2 SSE2 implementation 101 */ 102 static void raid6_sse22_gen_syndrome(int disks, size_t bytes, void **ptrs) 103 { 104 u8 **dptr = (u8 **)ptrs; 105 u8 *p, *q; 106 int d, z, z0; 107 108 z0 = disks - 3; /* Highest data disk */ 109 p = dptr[z0+1]; /* XOR parity */ 110 q = dptr[z0+2]; /* RS syndrome */ 111 112 kernel_fpu_begin(); 113 114 asm volatile("movdqa %0,%%xmm0" : : "m" (raid6_sse_constants.x1d[0])); 115 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */ 116 asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */ 117 118 /* We uniformly assume a single prefetch covers at least 32 bytes */ 119 for ( d = 0 ; d < bytes ; d += 32 ) { 120 asm volatile("prefetchnta %0" : : "m" (dptr[z0][d])); 121 asm volatile("movdqa %0,%%xmm2" : : "m" (dptr[z0][d])); /* P[0] */ 122 asm volatile("movdqa %0,%%xmm3" : : "m" (dptr[z0][d+16])); /* P[1] */ 123 asm volatile("movdqa %xmm2,%xmm4"); /* Q[0] */ 124 asm volatile("movdqa %xmm3,%xmm6"); /* Q[1] */ 125 for ( z = z0-1 ; z >= 0 ; z-- ) { 126 asm volatile("prefetchnta %0" : : "m" (dptr[z][d])); 127 asm volatile("pcmpgtb %xmm4,%xmm5"); 128 asm volatile("pcmpgtb %xmm6,%xmm7"); 129 asm volatile("paddb %xmm4,%xmm4"); 130 asm volatile("paddb %xmm6,%xmm6"); 131 asm volatile("pand %xmm0,%xmm5"); 132 asm volatile("pand %xmm0,%xmm7"); 133 asm volatile("pxor %xmm5,%xmm4"); 134 asm volatile("pxor %xmm7,%xmm6"); 135 asm volatile("movdqa %0,%%xmm5" : : "m" (dptr[z][d])); 136 asm volatile("movdqa %0,%%xmm7" : : "m" (dptr[z][d+16])); 137 asm volatile("pxor %xmm5,%xmm2"); 138 asm volatile("pxor %xmm7,%xmm3"); 139 asm volatile("pxor %xmm5,%xmm4"); 140 asm volatile("pxor %xmm7,%xmm6"); 141 asm volatile("pxor %xmm5,%xmm5"); 142 asm volatile("pxor %xmm7,%xmm7"); 143 } 144 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d])); 145 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16])); 146 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d])); 147 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16])); 148 } 149 150 asm volatile("sfence" : : : "memory"); 151 kernel_fpu_end(); 152 } 153 154 const struct raid6_calls raid6_sse2x2 = { 155 raid6_sse22_gen_syndrome, 156 NULL, /* XOR not yet implemented */ 157 raid6_have_sse2, 158 "sse2x2", 159 1 /* Has cache hints */ 160 }; 161 162 #ifdef CONFIG_X86_64 163 164 /* 165 * Unrolled-by-4 SSE2 implementation 166 */ 167 static void raid6_sse24_gen_syndrome(int disks, size_t bytes, void **ptrs) 168 { 169 u8 **dptr = (u8 **)ptrs; 170 u8 *p, *q; 171 int d, z, z0; 172 173 z0 = disks - 3; /* Highest data disk */ 174 p = dptr[z0+1]; /* XOR parity */ 175 q = dptr[z0+2]; /* RS syndrome */ 176 177 kernel_fpu_begin(); 178 179 asm volatile("movdqa %0,%%xmm0" :: "m" (raid6_sse_constants.x1d[0])); 180 asm volatile("pxor %xmm2,%xmm2"); /* P[0] */ 181 asm volatile("pxor %xmm3,%xmm3"); /* P[1] */ 182 asm volatile("pxor %xmm4,%xmm4"); /* Q[0] */ 183 asm volatile("pxor %xmm5,%xmm5"); /* Zero temp */ 184 asm volatile("pxor %xmm6,%xmm6"); /* Q[1] */ 185 asm volatile("pxor %xmm7,%xmm7"); /* Zero temp */ 186 asm volatile("pxor %xmm10,%xmm10"); /* P[2] */ 187 asm volatile("pxor %xmm11,%xmm11"); /* P[3] */ 188 asm volatile("pxor %xmm12,%xmm12"); /* Q[2] */ 189 asm volatile("pxor %xmm13,%xmm13"); /* Zero temp */ 190 asm volatile("pxor %xmm14,%xmm14"); /* Q[3] */ 191 asm volatile("pxor %xmm15,%xmm15"); /* Zero temp */ 192 193 for ( d = 0 ; d < bytes ; d += 64 ) { 194 for ( z = z0 ; z >= 0 ; z-- ) { 195 /* The second prefetch seems to improve performance... */ 196 asm volatile("prefetchnta %0" :: "m" (dptr[z][d])); 197 asm volatile("prefetchnta %0" :: "m" (dptr[z][d+32])); 198 asm volatile("pcmpgtb %xmm4,%xmm5"); 199 asm volatile("pcmpgtb %xmm6,%xmm7"); 200 asm volatile("pcmpgtb %xmm12,%xmm13"); 201 asm volatile("pcmpgtb %xmm14,%xmm15"); 202 asm volatile("paddb %xmm4,%xmm4"); 203 asm volatile("paddb %xmm6,%xmm6"); 204 asm volatile("paddb %xmm12,%xmm12"); 205 asm volatile("paddb %xmm14,%xmm14"); 206 asm volatile("pand %xmm0,%xmm5"); 207 asm volatile("pand %xmm0,%xmm7"); 208 asm volatile("pand %xmm0,%xmm13"); 209 asm volatile("pand %xmm0,%xmm15"); 210 asm volatile("pxor %xmm5,%xmm4"); 211 asm volatile("pxor %xmm7,%xmm6"); 212 asm volatile("pxor %xmm13,%xmm12"); 213 asm volatile("pxor %xmm15,%xmm14"); 214 asm volatile("movdqa %0,%%xmm5" :: "m" (dptr[z][d])); 215 asm volatile("movdqa %0,%%xmm7" :: "m" (dptr[z][d+16])); 216 asm volatile("movdqa %0,%%xmm13" :: "m" (dptr[z][d+32])); 217 asm volatile("movdqa %0,%%xmm15" :: "m" (dptr[z][d+48])); 218 asm volatile("pxor %xmm5,%xmm2"); 219 asm volatile("pxor %xmm7,%xmm3"); 220 asm volatile("pxor %xmm13,%xmm10"); 221 asm volatile("pxor %xmm15,%xmm11"); 222 asm volatile("pxor %xmm5,%xmm4"); 223 asm volatile("pxor %xmm7,%xmm6"); 224 asm volatile("pxor %xmm13,%xmm12"); 225 asm volatile("pxor %xmm15,%xmm14"); 226 asm volatile("pxor %xmm5,%xmm5"); 227 asm volatile("pxor %xmm7,%xmm7"); 228 asm volatile("pxor %xmm13,%xmm13"); 229 asm volatile("pxor %xmm15,%xmm15"); 230 } 231 asm volatile("movntdq %%xmm2,%0" : "=m" (p[d])); 232 asm volatile("pxor %xmm2,%xmm2"); 233 asm volatile("movntdq %%xmm3,%0" : "=m" (p[d+16])); 234 asm volatile("pxor %xmm3,%xmm3"); 235 asm volatile("movntdq %%xmm10,%0" : "=m" (p[d+32])); 236 asm volatile("pxor %xmm10,%xmm10"); 237 asm volatile("movntdq %%xmm11,%0" : "=m" (p[d+48])); 238 asm volatile("pxor %xmm11,%xmm11"); 239 asm volatile("movntdq %%xmm4,%0" : "=m" (q[d])); 240 asm volatile("pxor %xmm4,%xmm4"); 241 asm volatile("movntdq %%xmm6,%0" : "=m" (q[d+16])); 242 asm volatile("pxor %xmm6,%xmm6"); 243 asm volatile("movntdq %%xmm12,%0" : "=m" (q[d+32])); 244 asm volatile("pxor %xmm12,%xmm12"); 245 asm volatile("movntdq %%xmm14,%0" : "=m" (q[d+48])); 246 asm volatile("pxor %xmm14,%xmm14"); 247 } 248 249 asm volatile("sfence" : : : "memory"); 250 kernel_fpu_end(); 251 } 252 253 const struct raid6_calls raid6_sse2x4 = { 254 raid6_sse24_gen_syndrome, 255 NULL, /* XOR not yet implemented */ 256 raid6_have_sse2, 257 "sse2x4", 258 1 /* Has cache hints */ 259 }; 260 261 #endif /* CONFIG_X86_64 */ 262