rf_decluster.c revision 1.5
1/* $NetBSD: rf_decluster.c,v 1.5 2000/03/07 01:54:29 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 61extern int rf_copyback_in_progress; /* debug only */ 62 63/* found in rf_kintf.c */ 64int rf_GetSpareTableFromDaemon(RF_SparetWait_t * req); 65 66/* configuration code */ 67 68int 69rf_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, 81 extraPUsPerDisk; 82 RF_StripeCount_t totSparePUsPerDisk; 83 RF_SectorNum_t diskOffsetOfLastFullTableInSUs; 84 RF_SectorCount_t SpareSpaceInSUs; 85 char *cfgBuf = (char *) (cfgPtr->layoutSpecific); 86 RF_StripeNum_t l, SUID; 87 88 SUID = l = 0; 89 numCompleteSpareRegionsPerDisk = 0; 90 91 /* 1. create layout specific structure */ 92 RF_MallocAndAdd(info, sizeof(RF_DeclusteredConfigInfo_t), (RF_DeclusteredConfigInfo_t *), raidPtr->cleanupList); 93 if (info == NULL) 94 return (ENOMEM); 95 layoutPtr->layoutSpecificInfo = (void *) info; 96 info->SpareTable = NULL; 97 98 /* 2. extract parameters from the config structure */ 99 if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) { 100 (void) bcopy(cfgBuf, info->sparemap_fname, RF_SPAREMAP_NAME_LEN); 101 } 102 cfgBuf += RF_SPAREMAP_NAME_LEN; 103 104 b = *((int *) cfgBuf); 105 cfgBuf += sizeof(int); 106 v = *((int *) cfgBuf); 107 cfgBuf += sizeof(int); 108 k = *((int *) cfgBuf); 109 cfgBuf += sizeof(int); 110 r = *((int *) cfgBuf); 111 cfgBuf += sizeof(int); 112 lambda = *((int *) cfgBuf); 113 cfgBuf += sizeof(int); 114 raidPtr->noRotate = *((int *) cfgBuf); 115 cfgBuf += sizeof(int); 116 117 /* the sparemaps are generated assuming that parity is rotated, so we 118 * issue a warning if both distributed sparing and no-rotate are on at 119 * the same time */ 120 if ((layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) && raidPtr->noRotate) { 121 RF_ERRORMSG("Warning: distributed sparing specified without parity rotation.\n"); 122 } 123 if (raidPtr->numCol != v) { 124 RF_ERRORMSG2("RAID: config error: table element count (%d) not equal to no. of cols (%d)\n", v, raidPtr->numCol); 125 return (EINVAL); 126 } 127 /* 3. set up the values used in the mapping code */ 128 info->BlocksPerTable = b; 129 info->Lambda = lambda; 130 info->NumParityReps = info->groupSize = k; 131 info->SUsPerTable = b * (k - 1) * layoutPtr->SUsPerPU; /* b blks, k-1 SUs each */ 132 info->SUsPerFullTable = k * info->SUsPerTable; /* rot k times */ 133 info->PUsPerBlock = k - 1; 134 info->SUsPerBlock = info->PUsPerBlock * layoutPtr->SUsPerPU; 135 info->TableDepthInPUs = (b * k) / v; 136 info->FullTableDepthInPUs = info->TableDepthInPUs * k; /* k repetitions */ 137 138 /* used only in distributed sparing case */ 139 info->FullTablesPerSpareRegion = (v - 1) / rf_gcd(r, v - 1); /* (v-1)/gcd fulltables */ 140 info->TablesPerSpareRegion = k * info->FullTablesPerSpareRegion; 141 info->SpareSpaceDepthPerRegionInSUs = (r * info->TablesPerSpareRegion / (v - 1)) * layoutPtr->SUsPerPU; 142 143 /* check to make sure the block design is sufficiently small */ 144 if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) { 145 if (info->FullTableDepthInPUs * layoutPtr->SUsPerPU + info->SpareSpaceDepthPerRegionInSUs > layoutPtr->stripeUnitsPerDisk) { 146 RF_ERRORMSG3("RAID: config error: Full Table depth (%d) + Spare Space (%d) larger than disk size (%d) (BD too big)\n", 147 (int) info->FullTableDepthInPUs, 148 (int) info->SpareSpaceDepthPerRegionInSUs, 149 (int) layoutPtr->stripeUnitsPerDisk); 150 return (EINVAL); 151 } 152 } else { 153 if (info->TableDepthInPUs * layoutPtr->SUsPerPU > layoutPtr->stripeUnitsPerDisk) { 154 RF_ERRORMSG2("RAID: config error: Table depth (%d) larger than disk size (%d) (BD too big)\n", 155 (int) (info->TableDepthInPUs * layoutPtr->SUsPerPU), \ 156 (int) layoutPtr->stripeUnitsPerDisk); 157 return (EINVAL); 158 } 159 } 160 161 162 /* compute the size of each disk, and the number of tables in the last 163 * fulltable (which need not be complete) */ 164 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 165 166 PUsPerDisk = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU; 167 spareRegionDepthInPUs = (info->TablesPerSpareRegion * info->TableDepthInPUs + 168 (info->TablesPerSpareRegion * info->TableDepthInPUs) / (v - 1)); 169 info->SpareRegionDepthInSUs = spareRegionDepthInPUs * layoutPtr->SUsPerPU; 170 171 numCompleteSpareRegionsPerDisk = PUsPerDisk / spareRegionDepthInPUs; 172 info->NumCompleteSRs = numCompleteSpareRegionsPerDisk; 173 extraPUsPerDisk = PUsPerDisk % spareRegionDepthInPUs; 174 175 /* assume conservatively that we need the full amount of spare 176 * space in one region in order to provide spares for the 177 * partial spare region at the end of the array. We set "i" 178 * to the number of tables in the partial spare region. This 179 * may actually include some fulltables. */ 180 extraPUsPerDisk -= (info->SpareSpaceDepthPerRegionInSUs / layoutPtr->SUsPerPU); 181 if (extraPUsPerDisk <= 0) 182 i = 0; 183 else 184 i = extraPUsPerDisk / info->TableDepthInPUs; 185 186 complete_FT_count = raidPtr->numRow * (numCompleteSpareRegionsPerDisk * (info->TablesPerSpareRegion / k) + i / k); 187 info->FullTableLimitSUID = complete_FT_count * info->SUsPerFullTable; 188 info->ExtraTablesPerDisk = i % k; 189 190 /* note that in the last spare region, the spare space is 191 * complete even though data/parity space is not */ 192 totSparePUsPerDisk = (numCompleteSpareRegionsPerDisk + 1) * (info->SpareSpaceDepthPerRegionInSUs / layoutPtr->SUsPerPU); 193 info->TotSparePUsPerDisk = totSparePUsPerDisk; 194 195 layoutPtr->stripeUnitsPerDisk = 196 ((complete_FT_count / raidPtr->numRow) * info->FullTableDepthInPUs + /* data & parity space */ 197 info->ExtraTablesPerDisk * info->TableDepthInPUs + 198 totSparePUsPerDisk /* spare space */ 199 ) * layoutPtr->SUsPerPU; 200 layoutPtr->dataStripeUnitsPerDisk = 201 (complete_FT_count * info->FullTableDepthInPUs + info->ExtraTablesPerDisk * info->TableDepthInPUs) 202 * layoutPtr->SUsPerPU * (k - 1) / k; 203 204 } else { 205 /* non-dist spare case: force each disk to contain an 206 * integral number of tables */ 207 layoutPtr->stripeUnitsPerDisk /= (info->TableDepthInPUs * layoutPtr->SUsPerPU); 208 layoutPtr->stripeUnitsPerDisk *= (info->TableDepthInPUs * layoutPtr->SUsPerPU); 209 210 /* compute the number of tables in the last fulltable, which 211 * need not be complete */ 212 complete_FT_count = 213 ((layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU) / info->FullTableDepthInPUs) * raidPtr->numRow; 214 215 info->FullTableLimitSUID = complete_FT_count * info->SUsPerFullTable; 216 info->ExtraTablesPerDisk = 217 ((layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerPU) / info->TableDepthInPUs) % k; 218 } 219 220 raidPtr->sectorsPerDisk = layoutPtr->stripeUnitsPerDisk * layoutPtr->sectorsPerStripeUnit; 221 222 /* find the disk offset of the stripe unit where the last fulltable 223 * starts */ 224 numCompleteFullTablesPerDisk = complete_FT_count / raidPtr->numRow; 225 diskOffsetOfLastFullTableInSUs = numCompleteFullTablesPerDisk * info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 226 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 227 SpareSpaceInSUs = numCompleteSpareRegionsPerDisk * info->SpareSpaceDepthPerRegionInSUs; 228 diskOffsetOfLastFullTableInSUs += SpareSpaceInSUs; 229 info->DiskOffsetOfLastSpareSpaceChunkInSUs = 230 diskOffsetOfLastFullTableInSUs + info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU; 231 } 232 info->DiskOffsetOfLastFullTableInSUs = diskOffsetOfLastFullTableInSUs; 233 info->numCompleteFullTablesPerDisk = numCompleteFullTablesPerDisk; 234 235 /* 4. create and initialize the lookup tables */ 236 info->LayoutTable = rf_make_2d_array(b, k, raidPtr->cleanupList); 237 if (info->LayoutTable == NULL) 238 return (ENOMEM); 239 info->OffsetTable = rf_make_2d_array(b, k, raidPtr->cleanupList); 240 if (info->OffsetTable == NULL) 241 return (ENOMEM); 242 info->BlockTable = rf_make_2d_array(info->TableDepthInPUs * layoutPtr->SUsPerPU, raidPtr->numCol, raidPtr->cleanupList); 243 if (info->BlockTable == NULL) 244 return (ENOMEM); 245 246 first_avail_slot = rf_make_1d_array(v, NULL); 247 if (first_avail_slot == NULL) 248 return (ENOMEM); 249 250 for (i = 0; i < b; i++) 251 for (j = 0; j < k; j++) 252 info->LayoutTable[i][j] = *cfgBuf++; 253 254 /* initialize offset table */ 255 for (i = 0; i < b; i++) 256 for (j = 0; j < k; j++) { 257 info->OffsetTable[i][j] = first_avail_slot[info->LayoutTable[i][j]]; 258 first_avail_slot[info->LayoutTable[i][j]]++; 259 } 260 261 /* initialize block table */ 262 for (SUID = l = 0; l < layoutPtr->SUsPerPU; l++) { 263 for (i = 0; i < b; i++) { 264 for (j = 0; j < k; j++) { 265 info->BlockTable[(info->OffsetTable[i][j] * layoutPtr->SUsPerPU) + l] 266 [info->LayoutTable[i][j]] = SUID; 267 } 268 SUID++; 269 } 270 } 271 272 rf_free_1d_array(first_avail_slot, v); 273 274 /* 5. set up the remaining redundant-but-useful parameters */ 275 276 raidPtr->totalSectors = (k * complete_FT_count + raidPtr->numRow * info->ExtraTablesPerDisk) * 277 info->SUsPerTable * layoutPtr->sectorsPerStripeUnit; 278 layoutPtr->numStripe = (raidPtr->totalSectors / layoutPtr->sectorsPerStripeUnit) / (k - 1); 279 280 /* strange evaluation order below to try and minimize overflow 281 * problems */ 282 283 layoutPtr->dataSectorsPerStripe = (k - 1) * layoutPtr->sectorsPerStripeUnit; 284 layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector; 285 layoutPtr->numDataCol = k - 1; 286 layoutPtr->numParityCol = 1; 287 288 return (0); 289} 290/* declustering with distributed sparing */ 291static void rf_ShutdownDeclusteredDS(RF_ThreadArg_t); 292static void 293rf_ShutdownDeclusteredDS(arg) 294 RF_ThreadArg_t arg; 295{ 296 RF_DeclusteredConfigInfo_t *info; 297 RF_Raid_t *raidPtr; 298 299 raidPtr = (RF_Raid_t *) arg; 300 info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 301 if (info->SpareTable) 302 rf_FreeSpareTable(raidPtr); 303} 304 305int 306rf_ConfigureDeclusteredDS( 307 RF_ShutdownList_t ** listp, 308 RF_Raid_t * raidPtr, 309 RF_Config_t * cfgPtr) 310{ 311 int rc; 312 313 rc = rf_ConfigureDeclustered(listp, raidPtr, cfgPtr); 314 if (rc) 315 return (rc); 316 rc = rf_ShutdownCreate(listp, rf_ShutdownDeclusteredDS, raidPtr); 317 if (rc) { 318 RF_ERRORMSG1("Got %d adding shutdown event for DeclusteredDS\n", rc); 319 rf_ShutdownDeclusteredDS(raidPtr); 320 return (rc); 321 } 322 return (0); 323} 324 325void 326rf_MapSectorDeclustered(raidPtr, raidSector, row, col, diskSector, remap) 327 RF_Raid_t *raidPtr; 328 RF_RaidAddr_t raidSector; 329 RF_RowCol_t *row; 330 RF_RowCol_t *col; 331 RF_SectorNum_t *diskSector; 332 int remap; 333{ 334 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 335 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 336 RF_StripeNum_t SUID = raidSector / layoutPtr->sectorsPerStripeUnit; 337 RF_StripeNum_t FullTableID, FullTableOffset, TableID, TableOffset; 338 RF_StripeNum_t BlockID, BlockOffset, RepIndex; 339 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 340 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 341 RF_StripeNum_t base_suid = 0, outSU, SpareRegion = 0, SpareSpace = 0; 342 343 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 344 345 FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array 346 * (across rows) */ 347 if (raidPtr->numRow == 1) 348 *row = 0; /* avoid a mod and a div in the common case */ 349 else { 350 *row = FullTableID % raidPtr->numRow; 351 FullTableID /= raidPtr->numRow; /* convert to fulltable ID on 352 * this disk */ 353 } 354 if (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) { 355 SpareRegion = FullTableID / info->FullTablesPerSpareRegion; 356 SpareSpace = SpareRegion * info->SpareSpaceDepthPerRegionInSUs; 357 } 358 FullTableOffset = SUID % sus_per_fulltable; 359 TableID = FullTableOffset / info->SUsPerTable; 360 TableOffset = FullTableOffset - TableID * info->SUsPerTable; 361 BlockID = TableOffset / info->PUsPerBlock; 362 BlockOffset = TableOffset - BlockID * info->PUsPerBlock; 363 BlockID %= info->BlocksPerTable; 364 RepIndex = info->PUsPerBlock - TableID; 365 if (!raidPtr->noRotate) 366 BlockOffset += ((BlockOffset >= RepIndex) ? 1 : 0); 367 *col = info->LayoutTable[BlockID][BlockOffset]; 368 369 /* remap to distributed spare space if indicated */ 370 if (remap) { 371 RF_ASSERT(raidPtr->Disks[*row][*col].status == rf_ds_reconstructing || raidPtr->Disks[*row][*col].status == rf_ds_dist_spared || 372 (rf_copyback_in_progress && raidPtr->Disks[*row][*col].status == rf_ds_optimal)); 373 rf_remap_to_spare_space(layoutPtr, info, *row, FullTableID, TableID, BlockID, (base_suid) ? 1 : 0, SpareRegion, col, &outSU); 374 } else { 375 376 outSU = base_suid; 377 outSU += FullTableID * fulltable_depth; /* offs to strt of FT */ 378 outSU += SpareSpace; /* skip rsvd spare space */ 379 outSU += TableID * info->TableDepthInPUs * layoutPtr->SUsPerPU; /* offs to strt of tble */ 380 outSU += info->OffsetTable[BlockID][BlockOffset] * layoutPtr->SUsPerPU; /* offs to the PU */ 381 } 382 outSU += TableOffset / (info->BlocksPerTable * info->PUsPerBlock); /* offs to the SU within 383 * a PU */ 384 385 /* convert SUs to sectors, and, if not aligned to SU boundary, add in 386 * offset to sector. */ 387 *diskSector = outSU * layoutPtr->sectorsPerStripeUnit + (raidSector % layoutPtr->sectorsPerStripeUnit); 388 389 RF_ASSERT(*col != -1); 390} 391 392 393/* prototyping this inexplicably causes the compile of the layout table (rf_layout.c) to fail */ 394void 395rf_MapParityDeclustered( 396 RF_Raid_t * raidPtr, 397 RF_RaidAddr_t raidSector, 398 RF_RowCol_t * row, 399 RF_RowCol_t * col, 400 RF_SectorNum_t * diskSector, 401 int remap) 402{ 403 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 404 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 405 RF_StripeNum_t SUID = raidSector / layoutPtr->sectorsPerStripeUnit; 406 RF_StripeNum_t FullTableID, FullTableOffset, TableID, TableOffset; 407 RF_StripeNum_t BlockID, BlockOffset, RepIndex; 408 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 409 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 410 RF_StripeNum_t base_suid = 0, outSU, SpareRegion = 0, SpareSpace = 0; 411 412 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 413 414 /* compute row & (possibly) spare space exactly as before */ 415 FullTableID = SUID / sus_per_fulltable; 416 if (raidPtr->numRow == 1) 417 *row = 0; /* avoid a mod and a div in the common case */ 418 else { 419 *row = FullTableID % raidPtr->numRow; 420 FullTableID /= raidPtr->numRow; /* convert to fulltable ID on 421 * this disk */ 422 } 423 if ((raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE)) { 424 SpareRegion = FullTableID / info->FullTablesPerSpareRegion; 425 SpareSpace = SpareRegion * info->SpareSpaceDepthPerRegionInSUs; 426 } 427 /* compute BlockID and RepIndex exactly as before */ 428 FullTableOffset = SUID % sus_per_fulltable; 429 TableID = FullTableOffset / info->SUsPerTable; 430 TableOffset = FullTableOffset - TableID * info->SUsPerTable; 431 /* TableOffset = FullTableOffset % info->SUsPerTable; */ 432 /* BlockID = (TableOffset / info->PUsPerBlock) % 433 * info->BlocksPerTable; */ 434 BlockID = TableOffset / info->PUsPerBlock; 435 /* BlockOffset = TableOffset % info->PUsPerBlock; */ 436 BlockOffset = TableOffset - BlockID * info->PUsPerBlock; 437 BlockID %= info->BlocksPerTable; 438 439 /* the parity block is in the position indicated by RepIndex */ 440 RepIndex = (raidPtr->noRotate) ? info->PUsPerBlock : info->PUsPerBlock - TableID; 441 *col = info->LayoutTable[BlockID][RepIndex]; 442 443 if (remap) { 444 RF_ASSERT(raidPtr->Disks[*row][*col].status == rf_ds_reconstructing || raidPtr->Disks[*row][*col].status == rf_ds_dist_spared || 445 (rf_copyback_in_progress && raidPtr->Disks[*row][*col].status == rf_ds_optimal)); 446 rf_remap_to_spare_space(layoutPtr, info, *row, FullTableID, TableID, BlockID, (base_suid) ? 1 : 0, SpareRegion, col, &outSU); 447 } else { 448 449 /* compute sector as before, except use RepIndex instead of 450 * BlockOffset */ 451 outSU = base_suid; 452 outSU += FullTableID * fulltable_depth; 453 outSU += SpareSpace; /* skip rsvd spare space */ 454 outSU += TableID * info->TableDepthInPUs * layoutPtr->SUsPerPU; 455 outSU += info->OffsetTable[BlockID][RepIndex] * layoutPtr->SUsPerPU; 456 } 457 458 outSU += TableOffset / (info->BlocksPerTable * info->PUsPerBlock); 459 *diskSector = outSU * layoutPtr->sectorsPerStripeUnit + (raidSector % layoutPtr->sectorsPerStripeUnit); 460 461 RF_ASSERT(*col != -1); 462} 463/* returns an array of ints identifying the disks that comprise the stripe containing the indicated address. 464 * the caller must _never_ attempt to modify this array. 465 */ 466void 467rf_IdentifyStripeDeclustered( 468 RF_Raid_t * raidPtr, 469 RF_RaidAddr_t addr, 470 RF_RowCol_t ** diskids, 471 RF_RowCol_t * outRow) 472{ 473 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 474 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 475 RF_StripeCount_t sus_per_fulltable = info->SUsPerFullTable; 476 RF_StripeCount_t fulltable_depth = info->FullTableDepthInPUs * layoutPtr->SUsPerPU; 477 RF_StripeNum_t base_suid = 0; 478 RF_StripeNum_t SUID = rf_RaidAddressToStripeUnitID(layoutPtr, addr); 479 RF_StripeNum_t stripeID, FullTableID; 480 int tableOffset; 481 482 rf_decluster_adjust_params(layoutPtr, &SUID, &sus_per_fulltable, &fulltable_depth, &base_suid); 483 FullTableID = SUID / sus_per_fulltable; /* fulltable ID within array 484 * (across rows) */ 485 *outRow = FullTableID % raidPtr->numRow; 486 stripeID = rf_StripeUnitIDToStripeID(layoutPtr, SUID); /* find stripe offset 487 * into array */ 488 tableOffset = (stripeID % info->BlocksPerTable); /* find offset into 489 * block design table */ 490 *diskids = info->LayoutTable[tableOffset]; 491} 492/* This returns the default head-separation limit, which is measured 493 * in "required units for reconstruction". Each time a disk fetches 494 * a unit, it bumps a counter. The head-sep code prohibits any disk 495 * from getting more than headSepLimit counter values ahead of any 496 * other. 497 * 498 * We assume here that the number of floating recon buffers is already 499 * set. There are r stripes to be reconstructed in each table, and so 500 * if we have a total of B buffers, we can have at most B/r tables 501 * under recon at any one time. In each table, lambda units are required 502 * from each disk, so given B buffers, the head sep limit has to be 503 * (lambda*B)/r units. We subtract one to avoid weird boundary cases. 504 * 505 * for example, suppose were given 50 buffers, r=19, and lambda=4 as in 506 * the 20.5 design. There are 19 stripes/table to be reconstructed, so 507 * we can have 50/19 tables concurrently under reconstruction, which means 508 * we can allow the fastest disk to get 50/19 tables ahead of the slower 509 * disk. There are lambda "required units" for each disk, so the fastest 510 * disk can get 4*50/19 = 10 counter values ahead of the slowest. 511 * 512 * If numBufsToAccumulate is not 1, we need to limit the head sep further 513 * because multiple bufs will be required for each stripe under recon. 514 */ 515RF_HeadSepLimit_t 516rf_GetDefaultHeadSepLimitDeclustered( 517 RF_Raid_t * raidPtr) 518{ 519 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 520 521 return (info->Lambda * raidPtr->numFloatingReconBufs / info->TableDepthInPUs / rf_numBufsToAccumulate); 522} 523/* returns the default number of recon buffers to use. The value 524 * is somewhat arbitrary...it's intended to be large enough to allow 525 * for a reasonably large head-sep limit, but small enough that you 526 * don't use up all your system memory with buffers. 527 */ 528int 529rf_GetDefaultNumFloatingReconBuffersDeclustered(RF_Raid_t * raidPtr) 530{ 531 return (100 * rf_numBufsToAccumulate); 532} 533/* sectors in the last fulltable of the array need to be handled 534 * specially since this fulltable can be incomplete. this function 535 * changes the values of certain params to handle this. 536 * 537 * the idea here is that MapSector et. al. figure out which disk the 538 * addressed unit lives on by computing the modulos of the unit number 539 * with the number of units per fulltable, table, etc. In the last 540 * fulltable, there are fewer units per fulltable, so we need to adjust 541 * the number of user data units per fulltable to reflect this. 542 * 543 * so, we (1) convert the fulltable size and depth parameters to 544 * the size of the partial fulltable at the end, (2) compute the 545 * disk sector offset where this fulltable starts, and (3) convert 546 * the users stripe unit number from an offset into the array to 547 * an offset into the last fulltable. 548 */ 549void 550rf_decluster_adjust_params( 551 RF_RaidLayout_t * layoutPtr, 552 RF_StripeNum_t * SUID, 553 RF_StripeCount_t * sus_per_fulltable, 554 RF_StripeCount_t * fulltable_depth, 555 RF_StripeNum_t * base_suid) 556{ 557 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 558 559 if (*SUID >= info->FullTableLimitSUID) { 560 /* new full table size is size of last full table on disk */ 561 *sus_per_fulltable = info->ExtraTablesPerDisk * info->SUsPerTable; 562 563 /* new full table depth is corresponding depth */ 564 *fulltable_depth = info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU; 565 566 /* set up the new base offset */ 567 *base_suid = info->DiskOffsetOfLastFullTableInSUs; 568 569 /* convert users array address to an offset into the last 570 * fulltable */ 571 *SUID -= info->FullTableLimitSUID; 572 } 573} 574/* 575 * map a stripe ID to a parity stripe ID. 576 * See comment above RaidAddressToParityStripeID in layout.c. 577 */ 578void 579rf_MapSIDToPSIDDeclustered( 580 RF_RaidLayout_t * layoutPtr, 581 RF_StripeNum_t stripeID, 582 RF_StripeNum_t * psID, 583 RF_ReconUnitNum_t * which_ru) 584{ 585 RF_DeclusteredConfigInfo_t *info; 586 587 info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 588 589 *psID = (stripeID / (layoutPtr->SUsPerPU * info->BlocksPerTable)) 590 * info->BlocksPerTable + (stripeID % info->BlocksPerTable); 591 *which_ru = (stripeID % (info->BlocksPerTable * layoutPtr->SUsPerPU)) 592 / info->BlocksPerTable; 593 RF_ASSERT((*which_ru) < layoutPtr->SUsPerPU / layoutPtr->SUsPerRU); 594} 595/* 596 * Called from MapSector and MapParity to retarget an access at the spare unit. 597 * Modifies the "col" and "outSU" parameters only. 598 */ 599void 600rf_remap_to_spare_space( 601 RF_RaidLayout_t * layoutPtr, 602 RF_DeclusteredConfigInfo_t * info, 603 RF_RowCol_t row, 604 RF_StripeNum_t FullTableID, 605 RF_StripeNum_t TableID, 606 RF_SectorNum_t BlockID, 607 RF_StripeNum_t base_suid, 608 RF_StripeNum_t SpareRegion, 609 RF_RowCol_t * outCol, 610 RF_StripeNum_t * outSU) 611{ 612 RF_StripeNum_t ftID, spareTableStartSU, TableInSpareRegion, lastSROffset, 613 which_ft; 614 615 /* 616 * note that FullTableID and hence SpareRegion may have gotten 617 * tweaked by rf_decluster_adjust_params. We detect this by 618 * noticing that base_suid is not 0. 619 */ 620 if (base_suid == 0) { 621 ftID = FullTableID; 622 } else { 623 /* 624 * There may be > 1.0 full tables in the last (i.e. partial) 625 * spare region. find out which of these we're in. 626 */ 627 lastSROffset = info->NumCompleteSRs * info->SpareRegionDepthInSUs; 628 which_ft = (info->DiskOffsetOfLastFullTableInSUs - lastSROffset) / (info->FullTableDepthInPUs * layoutPtr->SUsPerPU); 629 630 /* compute the actual full table ID */ 631 ftID = info->DiskOffsetOfLastFullTableInSUs / (info->FullTableDepthInPUs * layoutPtr->SUsPerPU) + which_ft; 632 SpareRegion = info->NumCompleteSRs; 633 } 634 TableInSpareRegion = (ftID * info->NumParityReps + TableID) % info->TablesPerSpareRegion; 635 636 *outCol = info->SpareTable[TableInSpareRegion][BlockID].spareDisk; 637 RF_ASSERT(*outCol != -1); 638 639 spareTableStartSU = (SpareRegion == info->NumCompleteSRs) ? 640 info->DiskOffsetOfLastFullTableInSUs + info->ExtraTablesPerDisk * info->TableDepthInPUs * layoutPtr->SUsPerPU : 641 (SpareRegion + 1) * info->SpareRegionDepthInSUs - info->SpareSpaceDepthPerRegionInSUs; 642 *outSU = spareTableStartSU + info->SpareTable[TableInSpareRegion][BlockID].spareBlockOffsetInSUs; 643 if (*outSU >= layoutPtr->stripeUnitsPerDisk) { 644 printf("rf_remap_to_spare_space: invalid remapped disk SU offset %ld\n", (long) *outSU); 645 } 646} 647 648int 649rf_InstallSpareTable( 650 RF_Raid_t * raidPtr, 651 RF_RowCol_t frow, 652 RF_RowCol_t fcol) 653{ 654 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 655 RF_SparetWait_t *req; 656 int retcode; 657 658 RF_Malloc(req, sizeof(*req), (RF_SparetWait_t *)); 659 req->C = raidPtr->numCol; 660 req->G = raidPtr->Layout.numDataCol + raidPtr->Layout.numParityCol; 661 req->fcol = fcol; 662 req->SUsPerPU = raidPtr->Layout.SUsPerPU; 663 req->TablesPerSpareRegion = info->TablesPerSpareRegion; 664 req->BlocksPerTable = info->BlocksPerTable; 665 req->TableDepthInPUs = info->TableDepthInPUs; 666 req->SpareSpaceDepthPerRegionInSUs = info->SpareSpaceDepthPerRegionInSUs; 667 668 retcode = rf_GetSpareTableFromDaemon(req); 669 RF_ASSERT(!retcode); /* XXX -- fix this to recover gracefully -- 670 * XXX */ 671 return (retcode); 672} 673/* 674 * Invoked via ioctl to install a spare table in the kernel. 675 */ 676int 677rf_SetSpareTable(raidPtr, data) 678 RF_Raid_t *raidPtr; 679 void *data; 680{ 681 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo; 682 RF_SpareTableEntry_t **ptrs; 683 int i, retcode; 684 685 /* what we need to copyin is a 2-d array, so first copyin the user 686 * pointers to the rows in the table */ 687 RF_Malloc(ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *), (RF_SpareTableEntry_t **)); 688 retcode = copyin((caddr_t) data, (caddr_t) ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 689 690 if (retcode) 691 return (retcode); 692 693 /* now allocate kernel space for the row pointers */ 694 RF_Malloc(info->SpareTable, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *), (RF_SpareTableEntry_t **)); 695 696 /* now allocate kernel space for each row in the table, and copy it in 697 * from user space */ 698 for (i = 0; i < info->TablesPerSpareRegion; i++) { 699 RF_Malloc(info->SpareTable[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t), (RF_SpareTableEntry_t *)); 700 retcode = copyin(ptrs[i], info->SpareTable[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t)); 701 if (retcode) { 702 info->SpareTable = NULL; /* blow off the memory 703 * we've allocated */ 704 return (retcode); 705 } 706 } 707 708 /* free up the temporary array we used */ 709 RF_Free(ptrs, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 710 711 return (0); 712} 713 714RF_ReconUnitCount_t 715rf_GetNumSpareRUsDeclustered(raidPtr) 716 RF_Raid_t *raidPtr; 717{ 718 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 719 720 return (((RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo)->TotSparePUsPerDisk); 721} 722 723 724void 725rf_FreeSpareTable(raidPtr) 726 RF_Raid_t *raidPtr; 727{ 728 long i; 729 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 730 RF_DeclusteredConfigInfo_t *info = (RF_DeclusteredConfigInfo_t *) layoutPtr->layoutSpecificInfo; 731 RF_SpareTableEntry_t **table = info->SpareTable; 732 733 for (i = 0; i < info->TablesPerSpareRegion; i++) { 734 RF_Free(table[i], info->BlocksPerTable * sizeof(RF_SpareTableEntry_t)); 735 } 736 RF_Free(table, info->TablesPerSpareRegion * sizeof(RF_SpareTableEntry_t *)); 737 info->SpareTable = (RF_SpareTableEntry_t **) NULL; 738} 739