rf_decluster.c revision 1.2
1/* $NetBSD: rf_decluster.c,v 1.2 1999/01/26 02:33:55 oster Exp $ */ 2/* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland 7 * 8 * Permission to use, copy, modify and distribute this software and 9 * its documentation is hereby granted, provided that both the copyright 10 * notice and this permission notice appear in all copies of the 11 * software, derivative works or modified versions, and any portions 12 * thereof, and that both notices appear in supporting documentation. 13 * 14 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 15 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 16 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 17 * 18 * Carnegie Mellon requests users of this software to return to 19 * 20 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 21 * School of Computer Science 22 * Carnegie Mellon University 23 * Pittsburgh PA 15213-3890 24 * 25 * any improvements or extensions that they make and grant Carnegie the 26 * rights to redistribute these changes. 27 */ 28 29/*---------------------------------------------------------------------- 30 * 31 * rf_decluster.c -- code related to the declustered layout 32 * 33 * Created 10-21-92 (MCH) 34 * 35 * Nov 93: adding support for distributed sparing. This code is a little 36 * complex: the basic layout used is as follows: 37 * let F = (v-1)/GCD(r,v-1). The spare space for each set of 38 * F consecutive fulltables is grouped together and placed after 39 * that set of tables. 40 * +------------------------------+ 41 * | F fulltables | 42 * | Spare Space | 43 * | F fulltables | 44 * | Spare Space | 45 * | ... | 46 * +------------------------------+ 47 * 48 *--------------------------------------------------------------------*/ 49 50#include "rf_types.h" 51#include "rf_raid.h" 52#include "rf_raidframe.h" 53#include "rf_configure.h" 54#include "rf_decluster.h" 55#include "rf_debugMem.h" 56#include "rf_utils.h" 57#include "rf_alloclist.h" 58#include "rf_general.h" 59#include "rf_shutdown.h" 60#include "rf_sys.h" 61 62extern int rf_copyback_in_progress; /* debug only */ 63 64/* found in rf_kintf.c */ 65int rf_GetSpareTableFromDaemon(RF_SparetWait_t *req); 66 67/* configuration code */ 68 69int rf_ConfigureDeclustered( 70 RF_ShutdownList_t **listp, 71 RF_Raid_t *raidPtr, 72 RF_Config_t *cfgPtr) 73{ 74 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 75 int b, v, k, r, lambda; /* block design params */ 76 int i, j; 77 RF_RowCol_t *first_avail_slot; 78 RF_StripeCount_t complete_FT_count, numCompleteFullTablesPerDisk; 79 RF_DeclusteredConfigInfo_t *info; 80 RF_StripeCount_t PUsPerDisk, spareRegionDepthInPUs, numCompleteSpareRegionsPerDisk, extraPUsPerDisk; 81 RF_StripeCount_t totSparePUsPerDisk; 82 RF_SectorNum_t diskOffsetOfLastFullTableInSUs; 83 RF_SectorCount_t SpareSpaceInSUs; 84 char *cfgBuf = (char *) (cfgPtr->layoutSpecific); 85 RF_StripeNum_t l, SUID; 86 87 SUID = l = 0; 88 numCompleteSpareRegionsPerDisk = 0; 89 90 /* 1. create layout specific structure */ 91 RF_MallocAndAdd(info, sizeof(RF_DeclusteredConfigInfo_t), (RF_DeclusteredConfigInfo_t *), raidPtr->cleanupList); 92 if (info == NULL) 93 return(ENOMEM); 94 layoutPtr->layoutSpecificInfo = (void *) info; 95 info->SpareTable = NULL; 96 97 /* 2. extract parameters from the config structure */ 98 if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) { 99 (void) bcopy(cfgBuf, info->sparemap_fname, RF_SPAREMAP_NAME_LEN); 100 } 101 cfgBuf += RF_SPAREMAP_NAME_LEN; 102 103 b = *( (int *) cfgBuf); cfgBuf += sizeof(int); 104 v = *( (int *) cfgBuf); cfgBuf += sizeof(int); 105 k = *( (int *) cfgBuf); cfgBuf += sizeof(int); 106 r = *( (int *) cfgBuf); cfgBuf += sizeof(int); 107 lambda = *( (int *) cfgBuf); cfgBuf += sizeof(int); 108 raidPtr->noRotate = *( (int *) cfgBuf); cfgBuf += sizeof(int); 109 110 /* the sparemaps are generated assuming that parity is rotated, so we issue 111 * a warning if both distributed sparing and no-rotate are on at the same time 112 */ 113 if ((layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) && raidPtr->noRotate) { 114 RF_ERRORMSG("Warning: distributed sparing specified without parity rotation.\n"); 115 } 116 117 if (raidPtr->numCol != v) { 118 RF_ERRORMSG2("RAID: config error: table element count (%d) not equal to no. of cols (%d)\n", v, raidPtr->numCol); 119 return(EINVAL); 120 } 121 122 /* 3. set up the values used in the mapping code */ 123 info->BlocksPerTable = b; 124 info->Lambda = lambda; 125 info->NumParityReps = info->groupSize = k; 126 info->SUsPerTable = b * (k-1) * layoutPtr->SUsPerPU;/* b blks, k-1 SUs each */ 127 info->SUsPerFullTable = k * info->SUsPerTable; /* rot k times */ 128 info->PUsPerBlock = k-1; 129 info->SUsPerBlock = info->PUsPerBlock * layoutPtr->SUsPerPU; 130 info->TableDepthInPUs = (b*k) / v; 131 info->FullTableDepthInPUs = info->TableDepthInPUs * k; /* k repetitions */ 132 133 /* used only in distributed sparing case */ 134 info->FullTablesPerSpareRegion = (v-1) / rf_gcd(r, v-1); /* (v-1)/gcd fulltables */ 135 info->TablesPerSpareRegion = k * info->FullTablesPerSpareRegion; 136 info->SpareSpaceDepthPerRegionInSUs = (r * info->TablesPerSpareRegion / (v-1)) * layoutPtr->SUsPerPU; 137 138 /* check to make sure the block design is sufficiently small */ 139 if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) { 140 if (info->FullTableDepthInPUs * layoutPtr->SUsPerPU + info->SpareSpaceDepthPerRegionInSUs > layoutPtr->stripeUnitsPerDisk) { 141 RF_ERRORMSG3("RAID: config error: Full Table depth (%d) + Spare Space (%d) larger than disk size (%d) (BD too big)\n", 142 (int)info->FullTableDepthInPUs, 143 (int)info->SpareSpaceDepthPerRegionInSUs, 144 (int)layoutPtr->stripeUnitsPerDisk); 145 return(EINVAL); 146 } 147 } else { 148 if (info->TableDepthInPUs * layoutPtr->SUsPerPU > layoutPtr->stripeUnitsPerDisk) { 149 RF_ERRORMSG2("RAID: config error: Table depth (%d) larger than disk size (%d) (BD too big)\n", 150 (int)(info->TableDepthInPUs * layoutPtr->SUsPerPU), \ 151 (int)layoutPtr->stripeUnitsPerDisk); 152 return(EINVAL); 153 } 154 } 155 156 157 /* compute the size of each disk, and the number of tables in the last fulltable (which 158 * need not be complete) 159 */ 160 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 161 162 PUsPerDisk = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU; 163 spareRegionDepthInPUs = (info->TablesPerSpareRegion * info->TableDepthInPUs + 164 (info->TablesPerSpareRegion * info->TableDepthInPUs) / (v-1)); 165 info->SpareRegionDepthInSUs = spareRegionDepthInPUs * layoutPtr->SUsPerPU; 166 167 numCompleteSpareRegionsPerDisk = PUsPerDisk / spareRegionDepthInPUs; 168 info->NumCompleteSRs = numCompleteSpareRegionsPerDisk; 169 extraPUsPerDisk = PUsPerDisk % spareRegionDepthInPUs; 170 171 /* assume conservatively that we need the full amount of spare space in one region in order 172 * to provide spares for the partial spare region at the end of the array. We set "i" to 173 * the number of tables in the partial spare region. This may actually include some fulltables. 174 */ 175 extraPUsPerDisk -= (info->SpareSpaceDepthPerRegionInSUs / layoutPtr->SUsPerPU); 176 if (extraPUsPerDisk <= 0) i = 0; 177 else i = extraPUsPerDisk/info->TableDepthInPUs; 178 179 complete_FT_count = raidPtr->numRow * (numCompleteSpareRegionsPerDisk * (info->TablesPerSpareRegion/k) + i/k); 180 info->FullTableLimitSUID = complete_FT_count * info->SUsPerFullTable; 181 info->ExtraTablesPerDisk = i % k; 182 183 /* note that in the last spare region, the spare space is complete even though data/parity space is not */ 184 totSparePUsPerDisk = (numCompleteSpareRegionsPerDisk+1) * (info->SpareSpaceDepthPerRegionInSUs / layoutPtr->SUsPerPU); 185 info->TotSparePUsPerDisk = totSparePUsPerDisk; 186 187 layoutPtr->stripeUnitsPerDisk = 188 ((complete_FT_count/raidPtr->numRow) * info->FullTableDepthInPUs + /* data & parity space */ 189 info->ExtraTablesPerDisk * info->TableDepthInPUs + 190 totSparePUsPerDisk /* spare space */ 191 ) * layoutPtr->SUsPerPU; 192 layoutPtr->dataStripeUnitsPerDisk = 193 (complete_FT_count * info->FullTableDepthInPUs + info->ExtraTablesPerDisk * info->TableDepthInPUs) 194 * layoutPtr->SUsPerPU * (k-1) / k; 195 196 } else { 197 /* non-dist spare case: force each disk to contain an integral number of tables */ 198 layoutPtr->stripeUnitsPerDisk /= (info->TableDepthInPUs * layoutPtr->SUsPerPU); 199 layoutPtr->stripeUnitsPerDisk *= (info->TableDepthInPUs * layoutPtr->SUsPerPU); 200 201 /* compute the number of tables in the last fulltable, which need not be complete */ 202 complete_FT_count = 203 ((layoutPtr->stripeUnitsPerDisk/layoutPtr->SUsPerPU) / info->FullTableDepthInPUs) * raidPtr->numRow; 204 205 info->FullTableLimitSUID = complete_FT_count * info->SUsPerFullTable; 206 info->ExtraTablesPerDisk = 207 ((layoutPtr->stripeUnitsPerDisk/layoutPtr->SUsPerPU) / info->TableDepthInPUs) % k; 208 } 209 210 raidPtr->sectorsPerDisk = layoutPtr->stripeUnitsPerDisk * layoutPtr->sectorsPerStripeUnit; 211 212 /* find the disk offset of the stripe unit where the last fulltable starts */ 213 numCompleteFullTablesPerDisk = complete_FT_count / raidPtr->numRow; 214 diskOffsetOfLastFullTableInSUs = numCompleteFullTablesPerDisk * info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 215 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 216 SpareSpaceInSUs = numCompleteSpareRegionsPerDisk * info->SpareSpaceDepthPerRegionInSUs; 217 diskOffsetOfLastFullTableInSUs += SpareSpaceInSUs; 218 info->DiskOffsetOfLastSpareSpaceChunkInSUs = 219 diskOffsetOfLastFullTableInSUs + info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU; 220 } 221 info->DiskOffsetOfLastFullTableInSUs = diskOffsetOfLastFullTableInSUs; 222 info->numCompleteFullTablesPerDisk = numCompleteFullTablesPerDisk; 223 224 /* 4. create and initialize the lookup tables */ 225 info->LayoutTable = rf_make_2d_array(b, k, raidPtr->cleanupList); 226 if (info->LayoutTable == NULL) 227 return(ENOMEM); 228 info->OffsetTable = rf_make_2d_array(b, k, raidPtr->cleanupList); 229 if (info->OffsetTable == NULL) 230 return(ENOMEM); 231 info->BlockTable = rf_make_2d_array(info->TableDepthInPUs*layoutPtr->SUsPerPU, raidPtr->numCol, raidPtr->cleanupList); 232 if (info->BlockTable == NULL) 233 return(ENOMEM); 234 235 first_avail_slot = rf_make_1d_array(v, NULL); 236 if (first_avail_slot == NULL) 237 return(ENOMEM); 238 239 for (i=0; i<b; i++) 240 for (j=0; j<k; j++) 241 info->LayoutTable[i][j] = *cfgBuf++; 242 243 /* initialize offset table */ 244 for (i=0; i<b; i++) for (j=0; j<k; j++) { 245 info->OffsetTable[i][j] = first_avail_slot[ info->LayoutTable[i][j] ]; 246 first_avail_slot[ info->LayoutTable[i][j] ]++; 247 } 248 249 /* initialize block table */ 250 for (SUID=l=0; l<layoutPtr->SUsPerPU; l++) { 251 for (i=0; i<b; i++) { 252 for (j=0; j<k; j++) { 253 info->BlockTable[ (info->OffsetTable[i][j] * layoutPtr->SUsPerPU) + l ] 254 [ info->LayoutTable[i][j] ] = SUID; 255 } 256 SUID++; 257 } 258 } 259 260 rf_free_1d_array(first_avail_slot, v); 261 262 /* 5. set up the remaining redundant-but-useful parameters */ 263 264 raidPtr->totalSectors = (k*complete_FT_count + raidPtr->numRow*info->ExtraTablesPerDisk) * 265 info->SUsPerTable * layoutPtr->sectorsPerStripeUnit; 266 layoutPtr->numStripe = (raidPtr->totalSectors / layoutPtr->sectorsPerStripeUnit) / (k-1); 267 268 /* strange evaluation order below to try and minimize overflow problems */ 269 270 layoutPtr->dataSectorsPerStripe = (k-1) * layoutPtr->sectorsPerStripeUnit; 271 layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector; 272 layoutPtr->numDataCol = k-1; 273 layoutPtr->numParityCol = 1; 274 275 return(0); 276} 277 278/* declustering with distributed sparing */ 279static void rf_ShutdownDeclusteredDS(RF_ThreadArg_t); 280static void rf_ShutdownDeclusteredDS(arg) 281 RF_ThreadArg_t arg; 282{ 283 RF_DeclusteredConfigInfo_t *info; 284 RF_Raid_t *raidPtr; 285 286 raidPtr = (RF_Raid_t *)arg; 287 info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 288 if (info->SpareTable) 289 rf_FreeSpareTable(raidPtr); 290} 291 292int rf_ConfigureDeclusteredDS( 293 RF_ShutdownList_t **listp, 294 RF_Raid_t *raidPtr, 295 RF_Config_t *cfgPtr) 296{ 297 int rc; 298 299 rc = rf_ConfigureDeclustered(listp, raidPtr, cfgPtr); 300 if (rc) 301 return(rc); 302 rc = rf_ShutdownCreate(listp, rf_ShutdownDeclusteredDS, raidPtr); 303 if (rc) { 304 RF_ERRORMSG1("Got %d adding shutdown event for DeclusteredDS\n", rc); 305 rf_ShutdownDeclusteredDS(raidPtr); 306 return(rc); 307 } 308 return(0); 309} 310 311void rf_MapSectorDeclustered(raidPtr, raidSector, row, col, diskSector, remap) 312 RF_Raid_t *raidPtr; 313 RF_RaidAddr_t raidSector; 314 RF_RowCol_t *row; 315 RF_RowCol_t *col; 316 RF_SectorNum_t *diskSector; 317 int remap; 318{ 319 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 320 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 321 RF_StripeNum_t SUID = raidSector / layoutPtr->sectorsPerStripeUnit; 322 RF_StripeNum_t FullTableID, FullTableOffset, TableID, TableOffset; 323 RF_StripeNum_t BlockID, BlockOffset, RepIndex; 324 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 325 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 326 RF_StripeNum_t base_suid = 0, outSU, SpareRegion=0, SpareSpace=0; 327 328 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 329 330 FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array (across rows) */ 331 if (raidPtr->numRow == 1) *row = 0; /* avoid a mod and a div in the common case */ 332 else { 333 *row = FullTableID % raidPtr->numRow; 334 FullTableID /= raidPtr->numRow; /* convert to fulltable ID on this disk */ 335 } 336 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 337 SpareRegion = FullTableID / info->FullTablesPerSpareRegion; 338 SpareSpace = SpareRegion * info->SpareSpaceDepthPerRegionInSUs; 339 } 340 FullTableOffset = SUID % sus_per_fulltable; 341 TableID = FullTableOffset / info->SUsPerTable; 342 TableOffset = FullTableOffset - TableID * info->SUsPerTable; 343 BlockID = TableOffset / info->PUsPerBlock; 344 BlockOffset = TableOffset - BlockID * info->PUsPerBlock; 345 BlockID %= info->BlocksPerTable; 346 RepIndex = info->PUsPerBlock - TableID; 347 if (!raidPtr->noRotate) BlockOffset += ((BlockOffset >= RepIndex) ? 1 : 0); 348 *col = info->LayoutTable[BlockID][BlockOffset]; 349 350 /* remap to distributed spare space if indicated */ 351 if (remap) { 352 RF_ASSERT( raidPtr->Disks[*row][*col].status == rf_ds_reconstructing || raidPtr->Disks[*row][*col].status == rf_ds_dist_spared || 353 (rf_copyback_in_progress && raidPtr->Disks[*row][*col].status == rf_ds_optimal)); 354 rf_remap_to_spare_space(layoutPtr, info, *row, FullTableID, TableID, BlockID, (base_suid) ? 1 : 0, SpareRegion, col, &outSU); 355 } else { 356 357 outSU = base_suid; 358 outSU += FullTableID * fulltable_depth; /* offs to strt of FT */ 359 outSU += SpareSpace; /* skip rsvd spare space */ 360 outSU += TableID * info->TableDepthInPUs * layoutPtr->SUsPerPU; /* offs to strt of tble */ 361 outSU += info->OffsetTable[BlockID][BlockOffset] * layoutPtr->SUsPerPU; /* offs to the PU */ 362 } 363 outSU += TableOffset / (info->BlocksPerTable * info->PUsPerBlock); /* offs to the SU within a PU */ 364 365 /* convert SUs to sectors, and, if not aligned to SU boundary, add in offset to sector. */ 366 *diskSector = outSU*layoutPtr->sectorsPerStripeUnit + (raidSector % layoutPtr->sectorsPerStripeUnit); 367 368 RF_ASSERT( *col != -1 ); 369} 370 371 372/* prototyping this inexplicably causes the compile of the layout table (rf_layout.c) to fail */ 373void rf_MapParityDeclustered( 374 RF_Raid_t *raidPtr, 375 RF_RaidAddr_t raidSector, 376 RF_RowCol_t *row, 377 RF_RowCol_t *col, 378 RF_SectorNum_t *diskSector, 379 int remap) 380{ 381 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 382 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 383 RF_StripeNum_t SUID = raidSector / layoutPtr->sectorsPerStripeUnit; 384 RF_StripeNum_t FullTableID, FullTableOffset, TableID, TableOffset; 385 RF_StripeNum_t BlockID, BlockOffset, RepIndex; 386 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 387 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 388 RF_StripeNum_t base_suid = 0, outSU, SpareRegion=0, SpareSpace=0; 389 390 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 391 392 /* compute row & (possibly) spare space exactly as before */ 393 FullTableID = SUID / sus_per_fulltable; 394 if (raidPtr->numRow == 1) *row = 0; /* avoid a mod and a div in the common case */ 395 else { 396 *row = FullTableID % raidPtr->numRow; 397 FullTableID /= raidPtr->numRow; /* convert to fulltable ID on this disk */ 398 } 399 if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) { 400 SpareRegion = FullTableID / info->FullTablesPerSpareRegion; 401 SpareSpace = SpareRegion * info->SpareSpaceDepthPerRegionInSUs; 402 } 403 404 /* compute BlockID and RepIndex exactly as before */ 405 FullTableOffset = SUID % sus_per_fulltable; 406 TableID = FullTableOffset / info->SUsPerTable; 407 TableOffset = FullTableOffset - TableID * info->SUsPerTable; 408 /*TableOffset = FullTableOffset % info->SUsPerTable;*/ 409 /*BlockID = (TableOffset / info->PUsPerBlock) % info->BlocksPerTable;*/ 410 BlockID = TableOffset / info->PUsPerBlock; 411 /*BlockOffset = TableOffset % info->PUsPerBlock;*/ 412 BlockOffset = TableOffset - BlockID * info->PUsPerBlock; 413 BlockID %= info->BlocksPerTable; 414 415 /* the parity block is in the position indicated by RepIndex */ 416 RepIndex = (raidPtr->noRotate) ? info->PUsPerBlock : info->PUsPerBlock - TableID; 417 *col = info->LayoutTable[BlockID][RepIndex]; 418 419 if (remap) { 420 RF_ASSERT( raidPtr->Disks[*row][*col].status == rf_ds_reconstructing || raidPtr->Disks[*row][*col].status == rf_ds_dist_spared || 421 (rf_copyback_in_progress && raidPtr->Disks[*row][*col].status == rf_ds_optimal)); 422 rf_remap_to_spare_space(layoutPtr, info, *row, FullTableID, TableID, BlockID, (base_suid) ? 1 : 0, SpareRegion, col, &outSU); 423 } else { 424 425 /* compute sector as before, except use RepIndex instead of BlockOffset */ 426 outSU = base_suid; 427 outSU += FullTableID * fulltable_depth; 428 outSU += SpareSpace; /* skip rsvd spare space */ 429 outSU += TableID * info->TableDepthInPUs * layoutPtr->SUsPerPU; 430 outSU += info->OffsetTable[BlockID][RepIndex] * layoutPtr->SUsPerPU; 431 } 432 433 outSU += TableOffset / (info->BlocksPerTable * info->PUsPerBlock); 434 *diskSector = outSU*layoutPtr->sectorsPerStripeUnit + (raidSector % layoutPtr->sectorsPerStripeUnit); 435 436 RF_ASSERT( *col != -1 ); 437} 438 439/* returns an array of ints identifying the disks that comprise the stripe containing the indicated address. 440 * the caller must _never_ attempt to modify this array. 441 */ 442void rf_IdentifyStripeDeclustered( 443 RF_Raid_t *raidPtr, 444 RF_RaidAddr_t addr, 445 RF_RowCol_t **diskids, 446 RF_RowCol_t *outRow) 447{ 448 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 449 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 450 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 451 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 452 RF_StripeNum_t base_suid = 0; 453 RF_StripeNum_t SUID = rf_RaidAddressToStripeUnitID(layoutPtr, addr); 454 RF_StripeNum_t stripeID, FullTableID; 455 int tableOffset; 456 457 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 458 FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array (across rows) */ 459 *outRow = FullTableID % raidPtr->numRow; 460 stripeID = rf_StripeUnitIDToStripeID(layoutPtr, SUID); /* find stripe offset into array */ 461 tableOffset = (stripeID % info->BlocksPerTable); /* find offset into block design table */ 462 *diskids = info->LayoutTable[tableOffset]; 463} 464 465/* This returns the default head-separation limit, which is measured 466 * in "required units for reconstruction". Each time a disk fetches 467 * a unit, it bumps a counter. The head-sep code prohibits any disk 468 * from getting more than headSepLimit counter values ahead of any 469 * other. 470 * 471 * We assume here that the number of floating recon buffers is already 472 * set. There are r stripes to be reconstructed in each table, and so 473 * if we have a total of B buffers, we can have at most B/r tables 474 * under recon at any one time. In each table, lambda units are required 475 * from each disk, so given B buffers, the head sep limit has to be 476 * (lambda*B)/r units. We subtract one to avoid weird boundary cases. 477 * 478 * for example, suppose were given 50 buffers, r=19, and lambda=4 as in 479 * the 20.5 design. There are 19 stripes/table to be reconstructed, so 480 * we can have 50/19 tables concurrently under reconstruction, which means 481 * we can allow the fastest disk to get 50/19 tables ahead of the slower 482 * disk. There are lambda "required units" for each disk, so the fastest 483 * disk can get 4*50/19 = 10 counter values ahead of the slowest. 484 * 485 * If numBufsToAccumulate is not 1, we need to limit the head sep further 486 * because multiple bufs will be required for each stripe under recon. 487 */ 488RF_HeadSepLimit_t rf_GetDefaultHeadSepLimitDeclustered( 489 RF_Raid_t *raidPtr) 490{ 491 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 492 493 return(info->Lambda * raidPtr->numFloatingReconBufs / info->TableDepthInPUs / rf_numBufsToAccumulate); 494} 495 496/* returns the default number of recon buffers to use. The value 497 * is somewhat arbitrary...it's intended to be large enough to allow 498 * for a reasonably large head-sep limit, but small enough that you 499 * don't use up all your system memory with buffers. 500 */ 501int rf_GetDefaultNumFloatingReconBuffersDeclustered(RF_Raid_t *raidPtr) 502{ 503 return(100 * rf_numBufsToAccumulate); 504} 505 506/* sectors in the last fulltable of the array need to be handled 507 * specially since this fulltable can be incomplete. this function 508 * changes the values of certain params to handle this. 509 * 510 * the idea here is that MapSector et. al. figure out which disk the 511 * addressed unit lives on by computing the modulos of the unit number 512 * with the number of units per fulltable, table, etc. In the last 513 * fulltable, there are fewer units per fulltable, so we need to adjust 514 * the number of user data units per fulltable to reflect this. 515 * 516 * so, we (1) convert the fulltable size and depth parameters to 517 * the size of the partial fulltable at the end, (2) compute the 518 * disk sector offset where this fulltable starts, and (3) convert 519 * the users stripe unit number from an offset into the array to 520 * an offset into the last fulltable. 521 */ 522void rf_decluster_adjust_params( 523 RF_RaidLayout_t *layoutPtr, 524 RF_StripeNum_t *SUID, 525 RF_StripeCount_t *sus_per_fulltable, 526 RF_StripeCount_t *fulltable_depth, 527 RF_StripeNum_t *base_suid) 528{ 529 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 530#if defined(__NetBSD__) && defined(_KERNEL) 531 /* Nothing! */ 532#else 533 char pc = layoutPtr->map->parityConfig; 534#endif 535 536 if (*SUID >= info->FullTableLimitSUID) { 537 /* new full table size is size of last full table on disk */ 538 *sus_per_fulltable = info->ExtraTablesPerDisk * info->SUsPerTable; 539 540 /* new full table depth is corresponding depth */ 541 *fulltable_depth = info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU; 542 543 /* set up the new base offset */ 544 *base_suid = info->DiskOffsetOfLastFullTableInSUs; 545 546 /* convert users array address to an offset into the last fulltable */ 547 *SUID -= info->FullTableLimitSUID; 548 } 549} 550 551/* 552 * map a stripe ID to a parity stripe ID. 553 * See comment above RaidAddressToParityStripeID in layout.c. 554 */ 555void rf_MapSIDToPSIDDeclustered( 556 RF_RaidLayout_t *layoutPtr, 557 RF_StripeNum_t stripeID, 558 RF_StripeNum_t *psID, 559 RF_ReconUnitNum_t *which_ru) 560{ 561 RF_DeclusteredConfigInfo_t *info; 562 563 info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 564 565 *psID = (stripeID / (layoutPtr->SUsPerPU * info->BlocksPerTable)) 566 * info->BlocksPerTable + (stripeID % info->BlocksPerTable); 567 *which_ru = (stripeID % (info->BlocksPerTable * layoutPtr->SUsPerPU)) 568 / info->BlocksPerTable; 569 RF_ASSERT( (*which_ru) < layoutPtr->SUsPerPU/layoutPtr->SUsPerRU); 570} 571 572/* 573 * Called from MapSector and MapParity to retarget an access at the spare unit. 574 * Modifies the "col" and "outSU" parameters only. 575 */ 576void rf_remap_to_spare_space( 577 RF_RaidLayout_t *layoutPtr, 578 RF_DeclusteredConfigInfo_t *info, 579 RF_RowCol_t row, 580 RF_StripeNum_t FullTableID, 581 RF_StripeNum_t TableID, 582 RF_SectorNum_t BlockID, 583 RF_StripeNum_t base_suid, 584 RF_StripeNum_t SpareRegion, 585 RF_RowCol_t *outCol, 586 RF_StripeNum_t *outSU) 587{ 588 RF_StripeNum_t ftID, spareTableStartSU, TableInSpareRegion, lastSROffset, which_ft; 589 590 /* 591 * note that FullTableID and hence SpareRegion may have gotten 592 * tweaked by rf_decluster_adjust_params. We detect this by 593 * noticing that base_suid is not 0. 594 */ 595 if (base_suid == 0) { 596 ftID = FullTableID; 597 } 598 else { 599 /* 600 * There may be > 1.0 full tables in the last (i.e. partial) 601 * spare region. find out which of these we're in. 602 */ 603 lastSROffset = info->NumCompleteSRs * info->SpareRegionDepthInSUs; 604 which_ft = (info->DiskOffsetOfLastFullTableInSUs - lastSROffset) / (info->FullTableDepthInPUs * layoutPtr->SUsPerPU); 605 606 /* compute the actual full table ID */ 607 ftID = info->DiskOffsetOfLastFullTableInSUs / (info->FullTableDepthInPUs * layoutPtr->SUsPerPU) + which_ft; 608 SpareRegion = info->NumCompleteSRs; 609 } 610 TableInSpareRegion = (ftID * info->NumParityReps + TableID) % info->TablesPerSpareRegion; 611 612 *outCol = info->SpareTable[TableInSpareRegion][BlockID].spareDisk; 613 RF_ASSERT( *outCol != -1); 614 615 spareTableStartSU = (SpareRegion == info->NumCompleteSRs) ? 616 info->DiskOffsetOfLastFullTableInSUs + info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU : 617 (SpareRegion+1) * info->SpareRegionDepthInSUs - info->SpareSpaceDepthPerRegionInSUs; 618 *outSU = spareTableStartSU + info->SpareTable[TableInSpareRegion][BlockID].spareBlockOffsetInSUs; 619 if (*outSU >= layoutPtr->stripeUnitsPerDisk) { 620 printf("rf_remap_to_spare_space: invalid remapped disk SU offset %ld\n",(long)*outSU); 621 } 622} 623 624int rf_InstallSpareTable( 625 RF_Raid_t *raidPtr, 626 RF_RowCol_t frow, 627 RF_RowCol_t fcol) 628{ 629 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 630 RF_SparetWait_t *req; 631 int retcode; 632 633 RF_Malloc(req, sizeof(*req), (RF_SparetWait_t *)); 634 req->C = raidPtr->numCol; 635 req->G = raidPtr->Layout.numDataCol + raidPtr->Layout.numParityCol; 636 req->fcol = fcol; 637 req->SUsPerPU = raidPtr->Layout.SUsPerPU; 638 req->TablesPerSpareRegion = info->TablesPerSpareRegion; 639 req->BlocksPerTable = info->BlocksPerTable; 640 req->TableDepthInPUs = info->TableDepthInPUs; 641 req->SpareSpaceDepthPerRegionInSUs = info->SpareSpaceDepthPerRegionInSUs; 642 643 retcode = rf_GetSpareTableFromDaemon(req); 644 RF_ASSERT(!retcode); /* XXX -- fix this to recover gracefully -- XXX */ 645 return(retcode); 646} 647 648/* 649 * Invoked via ioctl to install a spare table in the kernel. 650 */ 651int rf_SetSpareTable(raidPtr, data) 652 RF_Raid_t *raidPtr; 653 void *data; 654{ 655 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 656 RF_SpareTableEntry_t **ptrs; 657 int i, retcode; 658 659 /* what we need to copyin is a 2-d array, so first copyin the user pointers to the rows in the table */ 660 RF_Malloc(ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *), (RF_SpareTableEntry_t **)); 661 retcode = copyin((caddr_t) data, (caddr_t) ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 662 663 if (retcode) return(retcode); 664 665 /* now allocate kernel space for the row pointers */ 666 RF_Malloc(info->SpareTable, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *), (RF_SpareTableEntry_t **)); 667 668 /* now allocate kernel space for each row in the table, and copy it in from user space */ 669 for (i=0; i<info->TablesPerSpareRegion; i++) { 670 RF_Malloc(info->SpareTable[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t), (RF_SpareTableEntry_t *)); 671 retcode = copyin(ptrs[i], info->SpareTable[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t)); 672 if (retcode) { 673 info->SpareTable = NULL; /* blow off the memory we've allocated */ 674 return(retcode); 675 } 676 } 677 678 /* free up the temporary array we used */ 679 RF_Free(ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 680 681 return(0); 682} 683 684RF_ReconUnitCount_t rf_GetNumSpareRUsDeclustered(raidPtr) 685 RF_Raid_t *raidPtr; 686{ 687 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 688 689 return( ((RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo)->TotSparePUsPerDisk ); 690} 691 692 693void rf_FreeSpareTable(raidPtr) 694 RF_Raid_t *raidPtr; 695{ 696 long i; 697 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 698 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 699 RF_SpareTableEntry_t **table = info->SpareTable; 700 701 for (i=0; i<info->TablesPerSpareRegion; i++) {RF_Free(table[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t));} 702 RF_Free(table, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 703 info->SpareTable = (RF_SpareTableEntry_t **) NULL; 704} 705