rf_decluster.c revision 1.3
1/* $NetBSD: rf_decluster.c,v 1.3 1999/02/05 00:06:08 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 70rf_ConfigureDeclustered( 71 RF_ShutdownList_t ** listp, 72 RF_Raid_t * raidPtr, 73 RF_Config_t * cfgPtr) 74{ 75 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 76 int b, v, k, r, lambda; /* block design params */ 77 int i, j; 78 RF_RowCol_t *first_avail_slot; 79 RF_StripeCount_t complete_FT_count, numCompleteFullTablesPerDisk; 80 RF_DeclusteredConfigInfo_t *info; 81 RF_StripeCount_t PUsPerDisk, spareRegionDepthInPUs, numCompleteSpareRegionsPerDisk, 82 extraPUsPerDisk; 83 RF_StripeCount_t totSparePUsPerDisk; 84 RF_SectorNum_t diskOffsetOfLastFullTableInSUs; 85 RF_SectorCount_t SpareSpaceInSUs; 86 char *cfgBuf = (char *) (cfgPtr->layoutSpecific); 87 RF_StripeNum_t l, SUID; 88 89 SUID = l = 0; 90 numCompleteSpareRegionsPerDisk = 0; 91 92 /* 1. create layout specific structure */ 93 RF_MallocAndAdd(info, sizeof(RF_DeclusteredConfigInfo_t), (RF_DeclusteredConfigInfo_t *), raidPtr->cleanupList); 94 if (info == NULL) 95 return (ENOMEM); 96 layoutPtr->layoutSpecificInfo = (void *) info; 97 info->SpareTable = NULL; 98 99 /* 2. extract parameters from the config structure */ 100 if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) { 101 (void) bcopy(cfgBuf, info->sparemap_fname, RF_SPAREMAP_NAME_LEN); 102 } 103 cfgBuf += RF_SPAREMAP_NAME_LEN; 104 105 b = *((int *) cfgBuf); 106 cfgBuf += sizeof(int); 107 v = *((int *) cfgBuf); 108 cfgBuf += sizeof(int); 109 k = *((int *) cfgBuf); 110 cfgBuf += sizeof(int); 111 r = *((int *) cfgBuf); 112 cfgBuf += sizeof(int); 113 lambda = *((int *) cfgBuf); 114 cfgBuf += sizeof(int); 115 raidPtr->noRotate = *((int *) cfgBuf); 116 cfgBuf += sizeof(int); 117 118 /* the sparemaps are generated assuming that parity is rotated, so we 119 * issue a warning if both distributed sparing and no-rotate are on at 120 * the same time */ 121 if ((layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) && raidPtr->noRotate) { 122 RF_ERRORMSG("Warning: distributed sparing specified without parity rotation.\n"); 123 } 124 if (raidPtr->numCol != v) { 125 RF_ERRORMSG2("RAID: config error: table element count (%d) not equal to no. of cols (%d)\n", v, raidPtr->numCol); 126 return (EINVAL); 127 } 128 /* 3. set up the values used in the mapping code */ 129 info->BlocksPerTable = b; 130 info->Lambda = lambda; 131 info->NumParityReps = info->groupSize = k; 132 info->SUsPerTable = b * (k - 1) * layoutPtr->SUsPerPU; /* b blks, k-1 SUs each */ 133 info->SUsPerFullTable = k * info->SUsPerTable; /* rot k times */ 134 info->PUsPerBlock = k - 1; 135 info->SUsPerBlock = info->PUsPerBlock * layoutPtr->SUsPerPU; 136 info->TableDepthInPUs = (b * k) / v; 137 info->FullTableDepthInPUs = info->TableDepthInPUs * k; /* k repetitions */ 138 139 /* used only in distributed sparing case */ 140 info->FullTablesPerSpareRegion = (v - 1) / rf_gcd(r, v - 1); /* (v-1)/gcd fulltables */ 141 info->TablesPerSpareRegion = k * info->FullTablesPerSpareRegion; 142 info->SpareSpaceDepthPerRegionInSUs = (r * info->TablesPerSpareRegion / (v - 1)) * layoutPtr->SUsPerPU; 143 144 /* check to make sure the block design is sufficiently small */ 145 if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) { 146 if (info->FullTableDepthInPUs * layoutPtr->SUsPerPU + info->SpareSpaceDepthPerRegionInSUs > layoutPtr->stripeUnitsPerDisk) { 147 RF_ERRORMSG3("RAID: config error: Full Table depth (%d) + Spare Space (%d) larger than disk size (%d) (BD too big)\n", 148 (int) info->FullTableDepthInPUs, 149 (int) info->SpareSpaceDepthPerRegionInSUs, 150 (int) layoutPtr->stripeUnitsPerDisk); 151 return (EINVAL); 152 } 153 } else { 154 if (info->TableDepthInPUs * layoutPtr->SUsPerPU > layoutPtr->stripeUnitsPerDisk) { 155 RF_ERRORMSG2("RAID: config error: Table depth (%d) larger than disk size (%d) (BD too big)\n", 156 (int) (info->TableDepthInPUs * layoutPtr->SUsPerPU), \ 157 (int) layoutPtr->stripeUnitsPerDisk); 158 return (EINVAL); 159 } 160 } 161 162 163 /* compute the size of each disk, and the number of tables in the last 164 * fulltable (which need not be complete) */ 165 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 166 167 PUsPerDisk = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU; 168 spareRegionDepthInPUs = (info->TablesPerSpareRegion * info->TableDepthInPUs + 169 (info->TablesPerSpareRegion * info->TableDepthInPUs) / (v - 1)); 170 info->SpareRegionDepthInSUs = spareRegionDepthInPUs * layoutPtr->SUsPerPU; 171 172 numCompleteSpareRegionsPerDisk = PUsPerDisk / spareRegionDepthInPUs; 173 info->NumCompleteSRs = numCompleteSpareRegionsPerDisk; 174 extraPUsPerDisk = PUsPerDisk % spareRegionDepthInPUs; 175 176 /* assume conservatively that we need the full amount of spare 177 * space in one region in order to provide spares for the 178 * partial spare region at the end of the array. We set "i" 179 * to the number of tables in the partial spare region. This 180 * may actually include some fulltables. */ 181 extraPUsPerDisk -= (info->SpareSpaceDepthPerRegionInSUs / layoutPtr->SUsPerPU); 182 if (extraPUsPerDisk <= 0) 183 i = 0; 184 else 185 i = extraPUsPerDisk / info->TableDepthInPUs; 186 187 complete_FT_count = raidPtr->numRow * (numCompleteSpareRegionsPerDisk * (info->TablesPerSpareRegion / k) + i / k); 188 info->FullTableLimitSUID = complete_FT_count * info->SUsPerFullTable; 189 info->ExtraTablesPerDisk = i % k; 190 191 /* note that in the last spare region, the spare space is 192 * complete even though data/parity space is not */ 193 totSparePUsPerDisk = (numCompleteSpareRegionsPerDisk + 1) * (info->SpareSpaceDepthPerRegionInSUs / layoutPtr->SUsPerPU); 194 info->TotSparePUsPerDisk = totSparePUsPerDisk; 195 196 layoutPtr->stripeUnitsPerDisk = 197 ((complete_FT_count / raidPtr->numRow) * info->FullTableDepthInPUs + /* data & parity space */ 198 info->ExtraTablesPerDisk * info->TableDepthInPUs + 199 totSparePUsPerDisk /* spare space */ 200 ) * layoutPtr->SUsPerPU; 201 layoutPtr->dataStripeUnitsPerDisk = 202 (complete_FT_count * info->FullTableDepthInPUs + info->ExtraTablesPerDisk * info->TableDepthInPUs) 203 * layoutPtr->SUsPerPU * (k - 1) / k; 204 205 } else { 206 /* non-dist spare case: force each disk to contain an 207 * integral number of tables */ 208 layoutPtr->stripeUnitsPerDisk /= (info->TableDepthInPUs * layoutPtr->SUsPerPU); 209 layoutPtr->stripeUnitsPerDisk *= (info->TableDepthInPUs * layoutPtr->SUsPerPU); 210 211 /* compute the number of tables in the last fulltable, which 212 * need not be complete */ 213 complete_FT_count = 214 ((layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU) / info->FullTableDepthInPUs) * raidPtr->numRow; 215 216 info->FullTableLimitSUID = complete_FT_count * info->SUsPerFullTable; 217 info->ExtraTablesPerDisk = 218 ((layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU) / info->TableDepthInPUs) % k; 219 } 220 221 raidPtr->sectorsPerDisk = layoutPtr->stripeUnitsPerDisk * layoutPtr->sectorsPerStripeUnit; 222 223 /* find the disk offset of the stripe unit where the last fulltable 224 * starts */ 225 numCompleteFullTablesPerDisk = complete_FT_count / raidPtr->numRow; 226 diskOffsetOfLastFullTableInSUs = numCompleteFullTablesPerDisk * info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 227 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 228 SpareSpaceInSUs = numCompleteSpareRegionsPerDisk * info->SpareSpaceDepthPerRegionInSUs; 229 diskOffsetOfLastFullTableInSUs += SpareSpaceInSUs; 230 info->DiskOffsetOfLastSpareSpaceChunkInSUs = 231 diskOffsetOfLastFullTableInSUs + info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU; 232 } 233 info->DiskOffsetOfLastFullTableInSUs = diskOffsetOfLastFullTableInSUs; 234 info->numCompleteFullTablesPerDisk = numCompleteFullTablesPerDisk; 235 236 /* 4. create and initialize the lookup tables */ 237 info->LayoutTable = rf_make_2d_array(b, k, raidPtr->cleanupList); 238 if (info->LayoutTable == NULL) 239 return (ENOMEM); 240 info->OffsetTable = rf_make_2d_array(b, k, raidPtr->cleanupList); 241 if (info->OffsetTable == NULL) 242 return (ENOMEM); 243 info->BlockTable = rf_make_2d_array(info->TableDepthInPUs * layoutPtr->SUsPerPU, raidPtr->numCol, raidPtr->cleanupList); 244 if (info->BlockTable == NULL) 245 return (ENOMEM); 246 247 first_avail_slot = rf_make_1d_array(v, NULL); 248 if (first_avail_slot == NULL) 249 return (ENOMEM); 250 251 for (i = 0; i < b; i++) 252 for (j = 0; j < k; j++) 253 info->LayoutTable[i][j] = *cfgBuf++; 254 255 /* initialize offset table */ 256 for (i = 0; i < b; i++) 257 for (j = 0; j < k; j++) { 258 info->OffsetTable[i][j] = first_avail_slot[info->LayoutTable[i][j]]; 259 first_avail_slot[info->LayoutTable[i][j]]++; 260 } 261 262 /* initialize block table */ 263 for (SUID = l = 0; l < layoutPtr->SUsPerPU; l++) { 264 for (i = 0; i < b; i++) { 265 for (j = 0; j < k; j++) { 266 info->BlockTable[(info->OffsetTable[i][j] * layoutPtr->SUsPerPU) + l] 267 [info->LayoutTable[i][j]] = SUID; 268 } 269 SUID++; 270 } 271 } 272 273 rf_free_1d_array(first_avail_slot, v); 274 275 /* 5. set up the remaining redundant-but-useful parameters */ 276 277 raidPtr->totalSectors = (k * complete_FT_count + raidPtr->numRow * info->ExtraTablesPerDisk) * 278 info->SUsPerTable * layoutPtr->sectorsPerStripeUnit; 279 layoutPtr->numStripe = (raidPtr->totalSectors / layoutPtr->sectorsPerStripeUnit) / (k - 1); 280 281 /* strange evaluation order below to try and minimize overflow 282 * problems */ 283 284 layoutPtr->dataSectorsPerStripe = (k - 1) * layoutPtr->sectorsPerStripeUnit; 285 layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector; 286 layoutPtr->numDataCol = k - 1; 287 layoutPtr->numParityCol = 1; 288 289 return (0); 290} 291/* declustering with distributed sparing */ 292static void rf_ShutdownDeclusteredDS(RF_ThreadArg_t); 293static void 294rf_ShutdownDeclusteredDS(arg) 295 RF_ThreadArg_t arg; 296{ 297 RF_DeclusteredConfigInfo_t *info; 298 RF_Raid_t *raidPtr; 299 300 raidPtr = (RF_Raid_t *) arg; 301 info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 302 if (info->SpareTable) 303 rf_FreeSpareTable(raidPtr); 304} 305 306int 307rf_ConfigureDeclusteredDS( 308 RF_ShutdownList_t ** listp, 309 RF_Raid_t * raidPtr, 310 RF_Config_t * cfgPtr) 311{ 312 int rc; 313 314 rc = rf_ConfigureDeclustered(listp, raidPtr, cfgPtr); 315 if (rc) 316 return (rc); 317 rc = rf_ShutdownCreate(listp, rf_ShutdownDeclusteredDS, raidPtr); 318 if (rc) { 319 RF_ERRORMSG1("Got %d adding shutdown event for DeclusteredDS\n", rc); 320 rf_ShutdownDeclusteredDS(raidPtr); 321 return (rc); 322 } 323 return (0); 324} 325 326void 327rf_MapSectorDeclustered(raidPtr, raidSector, row, col, diskSector, remap) 328 RF_Raid_t *raidPtr; 329 RF_RaidAddr_t raidSector; 330 RF_RowCol_t *row; 331 RF_RowCol_t *col; 332 RF_SectorNum_t *diskSector; 333 int remap; 334{ 335 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 336 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 337 RF_StripeNum_t SUID = raidSector / layoutPtr->sectorsPerStripeUnit; 338 RF_StripeNum_t FullTableID, FullTableOffset, TableID, TableOffset; 339 RF_StripeNum_t BlockID, BlockOffset, RepIndex; 340 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 341 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 342 RF_StripeNum_t base_suid = 0, outSU, SpareRegion = 0, SpareSpace = 0; 343 344 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 345 346 FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array 347 * (across rows) */ 348 if (raidPtr->numRow == 1) 349 *row = 0; /* avoid a mod and a div in the common case */ 350 else { 351 *row = FullTableID % raidPtr->numRow; 352 FullTableID /= raidPtr->numRow; /* convert to fulltable ID on 353 * this disk */ 354 } 355 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 356 SpareRegion = FullTableID / info->FullTablesPerSpareRegion; 357 SpareSpace = SpareRegion * info->SpareSpaceDepthPerRegionInSUs; 358 } 359 FullTableOffset = SUID % sus_per_fulltable; 360 TableID = FullTableOffset / info->SUsPerTable; 361 TableOffset = FullTableOffset - TableID * info->SUsPerTable; 362 BlockID = TableOffset / info->PUsPerBlock; 363 BlockOffset = TableOffset - BlockID * info->PUsPerBlock; 364 BlockID %= info->BlocksPerTable; 365 RepIndex = info->PUsPerBlock - TableID; 366 if (!raidPtr->noRotate) 367 BlockOffset += ((BlockOffset >= RepIndex) ? 1 : 0); 368 *col = info->LayoutTable[BlockID][BlockOffset]; 369 370 /* remap to distributed spare space if indicated */ 371 if (remap) { 372 RF_ASSERT(raidPtr->Disks[*row][*col].status == rf_ds_reconstructing || raidPtr->Disks[*row][*col].status == rf_ds_dist_spared || 373 (rf_copyback_in_progress && raidPtr->Disks[*row][*col].status == rf_ds_optimal)); 374 rf_remap_to_spare_space(layoutPtr, info, *row, FullTableID, TableID, BlockID, (base_suid) ? 1 : 0, SpareRegion, col, &outSU); 375 } else { 376 377 outSU = base_suid; 378 outSU += FullTableID * fulltable_depth; /* offs to strt of FT */ 379 outSU += SpareSpace; /* skip rsvd spare space */ 380 outSU += TableID * info->TableDepthInPUs * layoutPtr->SUsPerPU; /* offs to strt of tble */ 381 outSU += info->OffsetTable[BlockID][BlockOffset] * layoutPtr->SUsPerPU; /* offs to the PU */ 382 } 383 outSU += TableOffset / (info->BlocksPerTable * info->PUsPerBlock); /* offs to the SU within 384 * a PU */ 385 386 /* convert SUs to sectors, and, if not aligned to SU boundary, add in 387 * offset to sector. */ 388 *diskSector = outSU * layoutPtr->sectorsPerStripeUnit + (raidSector % layoutPtr->sectorsPerStripeUnit); 389 390 RF_ASSERT(*col != -1); 391} 392 393 394/* prototyping this inexplicably causes the compile of the layout table (rf_layout.c) to fail */ 395void 396rf_MapParityDeclustered( 397 RF_Raid_t * raidPtr, 398 RF_RaidAddr_t raidSector, 399 RF_RowCol_t * row, 400 RF_RowCol_t * col, 401 RF_SectorNum_t * diskSector, 402 int remap) 403{ 404 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 405 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 406 RF_StripeNum_t SUID = raidSector / layoutPtr->sectorsPerStripeUnit; 407 RF_StripeNum_t FullTableID, FullTableOffset, TableID, TableOffset; 408 RF_StripeNum_t BlockID, BlockOffset, RepIndex; 409 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 410 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 411 RF_StripeNum_t base_suid = 0, outSU, SpareRegion = 0, SpareSpace = 0; 412 413 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 414 415 /* compute row & (possibly) spare space exactly as before */ 416 FullTableID = SUID / sus_per_fulltable; 417 if (raidPtr->numRow == 1) 418 *row = 0; /* avoid a mod and a div in the common case */ 419 else { 420 *row = FullTableID % raidPtr->numRow; 421 FullTableID /= raidPtr->numRow; /* convert to fulltable ID on 422 * this disk */ 423 } 424 if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) { 425 SpareRegion = FullTableID / info->FullTablesPerSpareRegion; 426 SpareSpace = SpareRegion * info->SpareSpaceDepthPerRegionInSUs; 427 } 428 /* compute BlockID and RepIndex exactly as before */ 429 FullTableOffset = SUID % sus_per_fulltable; 430 TableID = FullTableOffset / info->SUsPerTable; 431 TableOffset = FullTableOffset - TableID * info->SUsPerTable; 432 /* TableOffset = FullTableOffset % info->SUsPerTable; */ 433 /* BlockID = (TableOffset / info->PUsPerBlock) % 434 * info->BlocksPerTable; */ 435 BlockID = TableOffset / info->PUsPerBlock; 436 /* BlockOffset = TableOffset % info->PUsPerBlock; */ 437 BlockOffset = TableOffset - BlockID * info->PUsPerBlock; 438 BlockID %= info->BlocksPerTable; 439 440 /* the parity block is in the position indicated by RepIndex */ 441 RepIndex = (raidPtr->noRotate) ? info->PUsPerBlock : info->PUsPerBlock - TableID; 442 *col = info->LayoutTable[BlockID][RepIndex]; 443 444 if (remap) { 445 RF_ASSERT(raidPtr->Disks[*row][*col].status == rf_ds_reconstructing || raidPtr->Disks[*row][*col].status == rf_ds_dist_spared || 446 (rf_copyback_in_progress && raidPtr->Disks[*row][*col].status == rf_ds_optimal)); 447 rf_remap_to_spare_space(layoutPtr, info, *row, FullTableID, TableID, BlockID, (base_suid) ? 1 : 0, SpareRegion, col, &outSU); 448 } else { 449 450 /* compute sector as before, except use RepIndex instead of 451 * BlockOffset */ 452 outSU = base_suid; 453 outSU += FullTableID * fulltable_depth; 454 outSU += SpareSpace; /* skip rsvd spare space */ 455 outSU += TableID * info->TableDepthInPUs * layoutPtr->SUsPerPU; 456 outSU += info->OffsetTable[BlockID][RepIndex] * layoutPtr->SUsPerPU; 457 } 458 459 outSU += TableOffset / (info->BlocksPerTable * info->PUsPerBlock); 460 *diskSector = outSU * layoutPtr->sectorsPerStripeUnit + (raidSector % layoutPtr->sectorsPerStripeUnit); 461 462 RF_ASSERT(*col != -1); 463} 464/* returns an array of ints identifying the disks that comprise the stripe containing the indicated address. 465 * the caller must _never_ attempt to modify this array. 466 */ 467void 468rf_IdentifyStripeDeclustered( 469 RF_Raid_t * raidPtr, 470 RF_RaidAddr_t addr, 471 RF_RowCol_t ** diskids, 472 RF_RowCol_t * outRow) 473{ 474 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 475 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 476 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 477 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 478 RF_StripeNum_t base_suid = 0; 479 RF_StripeNum_t SUID = rf_RaidAddressToStripeUnitID(layoutPtr, addr); 480 RF_StripeNum_t stripeID, FullTableID; 481 int tableOffset; 482 483 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 484 FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array 485 * (across rows) */ 486 *outRow = FullTableID % raidPtr->numRow; 487 stripeID = rf_StripeUnitIDToStripeID(layoutPtr, SUID); /* find stripe offset 488 * into array */ 489 tableOffset = (stripeID % info->BlocksPerTable); /* find offset into 490 * block design table */ 491 *diskids = info->LayoutTable[tableOffset]; 492} 493/* This returns the default head-separation limit, which is measured 494 * in "required units for reconstruction". Each time a disk fetches 495 * a unit, it bumps a counter. The head-sep code prohibits any disk 496 * from getting more than headSepLimit counter values ahead of any 497 * other. 498 * 499 * We assume here that the number of floating recon buffers is already 500 * set. There are r stripes to be reconstructed in each table, and so 501 * if we have a total of B buffers, we can have at most B/r tables 502 * under recon at any one time. In each table, lambda units are required 503 * from each disk, so given B buffers, the head sep limit has to be 504 * (lambda*B)/r units. We subtract one to avoid weird boundary cases. 505 * 506 * for example, suppose were given 50 buffers, r=19, and lambda=4 as in 507 * the 20.5 design. There are 19 stripes/table to be reconstructed, so 508 * we can have 50/19 tables concurrently under reconstruction, which means 509 * we can allow the fastest disk to get 50/19 tables ahead of the slower 510 * disk. There are lambda "required units" for each disk, so the fastest 511 * disk can get 4*50/19 = 10 counter values ahead of the slowest. 512 * 513 * If numBufsToAccumulate is not 1, we need to limit the head sep further 514 * because multiple bufs will be required for each stripe under recon. 515 */ 516RF_HeadSepLimit_t 517rf_GetDefaultHeadSepLimitDeclustered( 518 RF_Raid_t * raidPtr) 519{ 520 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 521 522 return (info->Lambda * raidPtr->numFloatingReconBufs / info->TableDepthInPUs / rf_numBufsToAccumulate); 523} 524/* returns the default number of recon buffers to use. The value 525 * is somewhat arbitrary...it's intended to be large enough to allow 526 * for a reasonably large head-sep limit, but small enough that you 527 * don't use up all your system memory with buffers. 528 */ 529int 530rf_GetDefaultNumFloatingReconBuffersDeclustered(RF_Raid_t * raidPtr) 531{ 532 return (100 * rf_numBufsToAccumulate); 533} 534/* sectors in the last fulltable of the array need to be handled 535 * specially since this fulltable can be incomplete. this function 536 * changes the values of certain params to handle this. 537 * 538 * the idea here is that MapSector et. al. figure out which disk the 539 * addressed unit lives on by computing the modulos of the unit number 540 * with the number of units per fulltable, table, etc. In the last 541 * fulltable, there are fewer units per fulltable, so we need to adjust 542 * the number of user data units per fulltable to reflect this. 543 * 544 * so, we (1) convert the fulltable size and depth parameters to 545 * the size of the partial fulltable at the end, (2) compute the 546 * disk sector offset where this fulltable starts, and (3) convert 547 * the users stripe unit number from an offset into the array to 548 * an offset into the last fulltable. 549 */ 550void 551rf_decluster_adjust_params( 552 RF_RaidLayout_t * layoutPtr, 553 RF_StripeNum_t * SUID, 554 RF_StripeCount_t * sus_per_fulltable, 555 RF_StripeCount_t * fulltable_depth, 556 RF_StripeNum_t * base_suid) 557{ 558 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 559#if defined(__NetBSD__) && defined(_KERNEL) 560 /* Nothing! */ 561#else 562 char pc = layoutPtr->map->parityConfig; 563#endif 564 565 if (*SUID >= info->FullTableLimitSUID) { 566 /* new full table size is size of last full table on disk */ 567 *sus_per_fulltable = info->ExtraTablesPerDisk * info->SUsPerTable; 568 569 /* new full table depth is corresponding depth */ 570 *fulltable_depth = info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU; 571 572 /* set up the new base offset */ 573 *base_suid = info->DiskOffsetOfLastFullTableInSUs; 574 575 /* convert users array address to an offset into the last 576 * fulltable */ 577 *SUID -= info->FullTableLimitSUID; 578 } 579} 580/* 581 * map a stripe ID to a parity stripe ID. 582 * See comment above RaidAddressToParityStripeID in layout.c. 583 */ 584void 585rf_MapSIDToPSIDDeclustered( 586 RF_RaidLayout_t * layoutPtr, 587 RF_StripeNum_t stripeID, 588 RF_StripeNum_t * psID, 589 RF_ReconUnitNum_t * which_ru) 590{ 591 RF_DeclusteredConfigInfo_t *info; 592 593 info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 594 595 *psID = (stripeID / (layoutPtr->SUsPerPU * info->BlocksPerTable)) 596 * info->BlocksPerTable + (stripeID % info->BlocksPerTable); 597 *which_ru = (stripeID % (info->BlocksPerTable * layoutPtr->SUsPerPU)) 598 / info->BlocksPerTable; 599 RF_ASSERT((*which_ru) < layoutPtr->SUsPerPU / layoutPtr->SUsPerRU); 600} 601/* 602 * Called from MapSector and MapParity to retarget an access at the spare unit. 603 * Modifies the "col" and "outSU" parameters only. 604 */ 605void 606rf_remap_to_spare_space( 607 RF_RaidLayout_t * layoutPtr, 608 RF_DeclusteredConfigInfo_t * info, 609 RF_RowCol_t row, 610 RF_StripeNum_t FullTableID, 611 RF_StripeNum_t TableID, 612 RF_SectorNum_t BlockID, 613 RF_StripeNum_t base_suid, 614 RF_StripeNum_t SpareRegion, 615 RF_RowCol_t * outCol, 616 RF_StripeNum_t * outSU) 617{ 618 RF_StripeNum_t ftID, spareTableStartSU, TableInSpareRegion, lastSROffset, 619 which_ft; 620 621 /* 622 * note that FullTableID and hence SpareRegion may have gotten 623 * tweaked by rf_decluster_adjust_params. We detect this by 624 * noticing that base_suid is not 0. 625 */ 626 if (base_suid == 0) { 627 ftID = FullTableID; 628 } else { 629 /* 630 * There may be > 1.0 full tables in the last (i.e. partial) 631 * spare region. find out which of these we're in. 632 */ 633 lastSROffset = info->NumCompleteSRs * info->SpareRegionDepthInSUs; 634 which_ft = (info->DiskOffsetOfLastFullTableInSUs - lastSROffset) / (info->FullTableDepthInPUs * layoutPtr->SUsPerPU); 635 636 /* compute the actual full table ID */ 637 ftID = info->DiskOffsetOfLastFullTableInSUs / (info->FullTableDepthInPUs * layoutPtr->SUsPerPU) + which_ft; 638 SpareRegion = info->NumCompleteSRs; 639 } 640 TableInSpareRegion = (ftID * info->NumParityReps + TableID) % info->TablesPerSpareRegion; 641 642 *outCol = info->SpareTable[TableInSpareRegion][BlockID].spareDisk; 643 RF_ASSERT(*outCol != -1); 644 645 spareTableStartSU = (SpareRegion == info->NumCompleteSRs) ? 646 info->DiskOffsetOfLastFullTableInSUs + info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU : 647 (SpareRegion + 1) * info->SpareRegionDepthInSUs - info->SpareSpaceDepthPerRegionInSUs; 648 *outSU = spareTableStartSU + info->SpareTable[TableInSpareRegion][BlockID].spareBlockOffsetInSUs; 649 if (*outSU >= layoutPtr->stripeUnitsPerDisk) { 650 printf("rf_remap_to_spare_space: invalid remapped disk SU offset %ld\n", (long) *outSU); 651 } 652} 653 654int 655rf_InstallSpareTable( 656 RF_Raid_t * raidPtr, 657 RF_RowCol_t frow, 658 RF_RowCol_t fcol) 659{ 660 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 661 RF_SparetWait_t *req; 662 int retcode; 663 664 RF_Malloc(req, sizeof(*req), (RF_SparetWait_t *)); 665 req->C = raidPtr->numCol; 666 req->G = raidPtr->Layout.numDataCol + raidPtr->Layout.numParityCol; 667 req->fcol = fcol; 668 req->SUsPerPU = raidPtr->Layout.SUsPerPU; 669 req->TablesPerSpareRegion = info->TablesPerSpareRegion; 670 req->BlocksPerTable = info->BlocksPerTable; 671 req->TableDepthInPUs = info->TableDepthInPUs; 672 req->SpareSpaceDepthPerRegionInSUs = info->SpareSpaceDepthPerRegionInSUs; 673 674 retcode = rf_GetSpareTableFromDaemon(req); 675 RF_ASSERT(!retcode); /* XXX -- fix this to recover gracefully -- 676 * XXX */ 677 return (retcode); 678} 679/* 680 * Invoked via ioctl to install a spare table in the kernel. 681 */ 682int 683rf_SetSpareTable(raidPtr, data) 684 RF_Raid_t *raidPtr; 685 void *data; 686{ 687 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 688 RF_SpareTableEntry_t **ptrs; 689 int i, retcode; 690 691 /* what we need to copyin is a 2-d array, so first copyin the user 692 * pointers to the rows in the table */ 693 RF_Malloc(ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *), (RF_SpareTableEntry_t **)); 694 retcode = copyin((caddr_t) data, (caddr_t) ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 695 696 if (retcode) 697 return (retcode); 698 699 /* now allocate kernel space for the row pointers */ 700 RF_Malloc(info->SpareTable, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *), (RF_SpareTableEntry_t **)); 701 702 /* now allocate kernel space for each row in the table, and copy it in 703 * from user space */ 704 for (i = 0; i < info->TablesPerSpareRegion; i++) { 705 RF_Malloc(info->SpareTable[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t), (RF_SpareTableEntry_t *)); 706 retcode = copyin(ptrs[i], info->SpareTable[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t)); 707 if (retcode) { 708 info->SpareTable = NULL; /* blow off the memory 709 * we've allocated */ 710 return (retcode); 711 } 712 } 713 714 /* free up the temporary array we used */ 715 RF_Free(ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 716 717 return (0); 718} 719 720RF_ReconUnitCount_t 721rf_GetNumSpareRUsDeclustered(raidPtr) 722 RF_Raid_t *raidPtr; 723{ 724 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 725 726 return (((RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo)->TotSparePUsPerDisk); 727} 728 729 730void 731rf_FreeSpareTable(raidPtr) 732 RF_Raid_t *raidPtr; 733{ 734 long i; 735 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 736 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 737 RF_SpareTableEntry_t **table = info->SpareTable; 738 739 for (i = 0; i < info->TablesPerSpareRegion; i++) { 740 RF_Free(table[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t)); 741 } 742 RF_Free(table, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 743 info->SpareTable = (RF_SpareTableEntry_t **) NULL; 744} 745