rf_paritylogging.c revision 1.20
1 /* $NetBSD: rf_paritylogging.c,v 1.20 2003/12/29 05:36:19 oster Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: William V. Courtright II 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 (at) 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 parity logging configuration, dag selection, and mapping is implemented here 32 */ 33 34 #include <sys/cdefs.h> 35 __KERNEL_RCSID(0, "$NetBSD: rf_paritylogging.c,v 1.20 2003/12/29 05:36:19 oster Exp $"); 36 37 #include "rf_archs.h" 38 39 #if RF_INCLUDE_PARITYLOGGING > 0 40 41 #include <dev/raidframe/raidframevar.h> 42 43 #include "rf_raid.h" 44 #include "rf_dag.h" 45 #include "rf_dagutils.h" 46 #include "rf_dagfuncs.h" 47 #include "rf_dagffrd.h" 48 #include "rf_dagffwr.h" 49 #include "rf_dagdegrd.h" 50 #include "rf_dagdegwr.h" 51 #include "rf_paritylog.h" 52 #include "rf_paritylogDiskMgr.h" 53 #include "rf_paritylogging.h" 54 #include "rf_parityloggingdags.h" 55 #include "rf_general.h" 56 #include "rf_map.h" 57 #include "rf_utils.h" 58 #include "rf_shutdown.h" 59 60 typedef struct RF_ParityLoggingConfigInfo_s { 61 RF_RowCol_t **stripeIdentifier; /* filled in at config time & used by 62 * IdentifyStripe */ 63 } RF_ParityLoggingConfigInfo_t; 64 65 static void FreeRegionInfo(RF_Raid_t * raidPtr, RF_RegionId_t regionID); 66 static void rf_ShutdownParityLogging(RF_ThreadArg_t arg); 67 static void rf_ShutdownParityLoggingRegionInfo(RF_ThreadArg_t arg); 68 static void rf_ShutdownParityLoggingPool(RF_ThreadArg_t arg); 69 static void rf_ShutdownParityLoggingRegionBufferPool(RF_ThreadArg_t arg); 70 static void rf_ShutdownParityLoggingParityBufferPool(RF_ThreadArg_t arg); 71 static void rf_ShutdownParityLoggingDiskQueue(RF_ThreadArg_t arg); 72 73 int 74 rf_ConfigureParityLogging( 75 RF_ShutdownList_t ** listp, 76 RF_Raid_t * raidPtr, 77 RF_Config_t * cfgPtr) 78 { 79 int i, j, startdisk, rc; 80 RF_SectorCount_t totalLogCapacity, fragmentation, lastRegionCapacity; 81 RF_SectorCount_t parityBufferCapacity, maxRegionParityRange; 82 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 83 RF_ParityLoggingConfigInfo_t *info; 84 RF_ParityLog_t *l = NULL, *next; 85 caddr_t lHeapPtr; 86 87 if (rf_numParityRegions <= 0) 88 return(EINVAL); 89 90 /* 91 * We create multiple entries on the shutdown list here, since 92 * this configuration routine is fairly complicated in and of 93 * itself, and this makes backing out of a failed configuration 94 * much simpler. 95 */ 96 97 raidPtr->numSectorsPerLog = RF_DEFAULT_NUM_SECTORS_PER_LOG; 98 99 /* create a parity logging configuration structure */ 100 RF_MallocAndAdd(info, sizeof(RF_ParityLoggingConfigInfo_t), 101 (RF_ParityLoggingConfigInfo_t *), 102 raidPtr->cleanupList); 103 if (info == NULL) 104 return (ENOMEM); 105 layoutPtr->layoutSpecificInfo = (void *) info; 106 107 /* the stripe identifier must identify the disks in each stripe, IN 108 * THE ORDER THAT THEY APPEAR IN THE STRIPE. */ 109 info->stripeIdentifier = rf_make_2d_array((raidPtr->numCol), 110 (raidPtr->numCol), 111 raidPtr->cleanupList); 112 if (info->stripeIdentifier == NULL) 113 return (ENOMEM); 114 115 startdisk = 0; 116 for (i = 0; i < (raidPtr->numCol); i++) { 117 for (j = 0; j < (raidPtr->numCol); j++) { 118 info->stripeIdentifier[i][j] = (startdisk + j) % 119 (raidPtr->numCol - 1); 120 } 121 if ((--startdisk) < 0) 122 startdisk = raidPtr->numCol - 1 - 1; 123 } 124 125 /* fill in the remaining layout parameters */ 126 layoutPtr->numStripe = layoutPtr->stripeUnitsPerDisk; 127 layoutPtr->numParityCol = 1; 128 layoutPtr->numParityLogCol = 1; 129 layoutPtr->numDataCol = raidPtr->numCol - layoutPtr->numParityCol - 130 layoutPtr->numParityLogCol; 131 layoutPtr->dataSectorsPerStripe = layoutPtr->numDataCol * 132 layoutPtr->sectorsPerStripeUnit; 133 layoutPtr->dataStripeUnitsPerDisk = layoutPtr->stripeUnitsPerDisk; 134 raidPtr->sectorsPerDisk = layoutPtr->stripeUnitsPerDisk * 135 layoutPtr->sectorsPerStripeUnit; 136 137 raidPtr->totalSectors = layoutPtr->stripeUnitsPerDisk * 138 layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit; 139 140 /* configure parity log parameters 141 * 142 * parameter comment/constraints 143 * ------------------------------------------- 144 * numParityRegions* all regions (except possibly last) 145 * of equal size 146 * totalInCoreLogCapacity* amount of memory in bytes available 147 * for in-core logs (default 1 MB) 148 * numSectorsPerLog# capacity of an in-core log in sectors 149 * (1 * disk track) 150 * numParityLogs total number of in-core logs, 151 * should be at least numParityRegions 152 * regionLogCapacity size of a region log (except possibly 153 * last one) in sectors 154 * totalLogCapacity total amount of log space in sectors 155 * 156 * where '*' denotes a user settable parameter. 157 * Note that logs are fixed to be the size of a disk track, 158 * value #defined in rf_paritylog.h 159 * 160 */ 161 162 totalLogCapacity = layoutPtr->stripeUnitsPerDisk * layoutPtr->sectorsPerStripeUnit * layoutPtr->numParityLogCol; 163 raidPtr->regionLogCapacity = totalLogCapacity / rf_numParityRegions; 164 if (rf_parityLogDebug) 165 printf("bytes per sector %d\n", raidPtr->bytesPerSector); 166 167 /* reduce fragmentation within a disk region by adjusting the number 168 * of regions in an attempt to allow an integral number of logs to fit 169 * into a disk region */ 170 fragmentation = raidPtr->regionLogCapacity % raidPtr->numSectorsPerLog; 171 if (fragmentation > 0) 172 for (i = 1; i < (raidPtr->numSectorsPerLog / 2); i++) { 173 if (((totalLogCapacity / (rf_numParityRegions + i)) % 174 raidPtr->numSectorsPerLog) < fragmentation) { 175 rf_numParityRegions++; 176 raidPtr->regionLogCapacity = totalLogCapacity / 177 rf_numParityRegions; 178 fragmentation = raidPtr->regionLogCapacity % 179 raidPtr->numSectorsPerLog; 180 } 181 if (((totalLogCapacity / (rf_numParityRegions - i)) % 182 raidPtr->numSectorsPerLog) < fragmentation) { 183 rf_numParityRegions--; 184 raidPtr->regionLogCapacity = totalLogCapacity / 185 rf_numParityRegions; 186 fragmentation = raidPtr->regionLogCapacity % 187 raidPtr->numSectorsPerLog; 188 } 189 } 190 /* ensure integral number of regions per log */ 191 raidPtr->regionLogCapacity = (raidPtr->regionLogCapacity / 192 raidPtr->numSectorsPerLog) * 193 raidPtr->numSectorsPerLog; 194 195 raidPtr->numParityLogs = rf_totalInCoreLogCapacity / 196 (raidPtr->bytesPerSector * raidPtr->numSectorsPerLog); 197 /* to avoid deadlock, must ensure that enough logs exist for each 198 * region to have one simultaneously */ 199 if (raidPtr->numParityLogs < rf_numParityRegions) 200 raidPtr->numParityLogs = rf_numParityRegions; 201 202 /* create region information structs */ 203 printf("Allocating %d bytes for in-core parity region info\n", 204 (int) (rf_numParityRegions * sizeof(RF_RegionInfo_t))); 205 RF_Malloc(raidPtr->regionInfo, 206 (rf_numParityRegions * sizeof(RF_RegionInfo_t)), 207 (RF_RegionInfo_t *)); 208 if (raidPtr->regionInfo == NULL) 209 return (ENOMEM); 210 211 /* last region may not be full capacity */ 212 lastRegionCapacity = raidPtr->regionLogCapacity; 213 while ((rf_numParityRegions - 1) * raidPtr->regionLogCapacity + 214 lastRegionCapacity > totalLogCapacity) 215 lastRegionCapacity = lastRegionCapacity - 216 raidPtr->numSectorsPerLog; 217 218 raidPtr->regionParityRange = raidPtr->sectorsPerDisk / 219 rf_numParityRegions; 220 maxRegionParityRange = raidPtr->regionParityRange; 221 222 /* i can't remember why this line is in the code -wvcii 6/30/95 */ 223 /* if (raidPtr->sectorsPerDisk % rf_numParityRegions > 0) 224 regionParityRange++; */ 225 226 /* build pool of unused parity logs */ 227 printf("Allocating %d bytes for %d parity logs\n", 228 raidPtr->numParityLogs * raidPtr->numSectorsPerLog * 229 raidPtr->bytesPerSector, 230 raidPtr->numParityLogs); 231 RF_Malloc(raidPtr->parityLogBufferHeap, raidPtr->numParityLogs * 232 raidPtr->numSectorsPerLog * raidPtr->bytesPerSector, 233 (caddr_t)); 234 if (raidPtr->parityLogBufferHeap == NULL) 235 return (ENOMEM); 236 lHeapPtr = raidPtr->parityLogBufferHeap; 237 rc = rf_mutex_init(&raidPtr->parityLogPool.mutex); 238 if (rc) { 239 rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc); 240 RF_Free(raidPtr->parityLogBufferHeap, raidPtr->numParityLogs * 241 raidPtr->numSectorsPerLog * raidPtr->bytesPerSector); 242 return (ENOMEM); 243 } 244 for (i = 0; i < raidPtr->numParityLogs; i++) { 245 if (i == 0) { 246 RF_Malloc(raidPtr->parityLogPool.parityLogs, 247 sizeof(RF_ParityLog_t), (RF_ParityLog_t *)); 248 if (raidPtr->parityLogPool.parityLogs == NULL) { 249 RF_Free(raidPtr->parityLogBufferHeap, 250 raidPtr->numParityLogs * 251 raidPtr->numSectorsPerLog * 252 raidPtr->bytesPerSector); 253 return (ENOMEM); 254 } 255 l = raidPtr->parityLogPool.parityLogs; 256 } else { 257 RF_Malloc(l->next, sizeof(RF_ParityLog_t), 258 (RF_ParityLog_t *)); 259 if (l->next == NULL) { 260 RF_Free(raidPtr->parityLogBufferHeap, 261 raidPtr->numParityLogs * 262 raidPtr->numSectorsPerLog * 263 raidPtr->bytesPerSector); 264 for (l = raidPtr->parityLogPool.parityLogs; 265 l; 266 l = next) { 267 next = l->next; 268 if (l->records) 269 RF_Free(l->records, (raidPtr->numSectorsPerLog * sizeof(RF_ParityLogRecord_t))); 270 RF_Free(l, sizeof(RF_ParityLog_t)); 271 } 272 return (ENOMEM); 273 } 274 l = l->next; 275 } 276 l->bufPtr = lHeapPtr; 277 lHeapPtr += raidPtr->numSectorsPerLog * 278 raidPtr->bytesPerSector; 279 RF_Malloc(l->records, (raidPtr->numSectorsPerLog * 280 sizeof(RF_ParityLogRecord_t)), 281 (RF_ParityLogRecord_t *)); 282 if (l->records == NULL) { 283 RF_Free(raidPtr->parityLogBufferHeap, 284 raidPtr->numParityLogs * 285 raidPtr->numSectorsPerLog * 286 raidPtr->bytesPerSector); 287 for (l = raidPtr->parityLogPool.parityLogs; 288 l; 289 l = next) { 290 next = l->next; 291 if (l->records) 292 RF_Free(l->records, 293 (raidPtr->numSectorsPerLog * 294 sizeof(RF_ParityLogRecord_t))); 295 RF_Free(l, sizeof(RF_ParityLog_t)); 296 } 297 return (ENOMEM); 298 } 299 } 300 rc = rf_ShutdownCreate(listp, rf_ShutdownParityLoggingPool, raidPtr); 301 if (rc) { 302 RF_ERRORMSG3("Unable to create shutdown entry file %s line %d rc=%d\n", __FILE__, 303 __LINE__, rc); 304 rf_ShutdownParityLoggingPool(raidPtr); 305 return (rc); 306 } 307 /* build pool of region buffers */ 308 rc = rf_mutex_init(&raidPtr->regionBufferPool.mutex); 309 if (rc) { 310 rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc); 311 return (ENOMEM); 312 } 313 raidPtr->regionBufferPool.cond = 0; 314 raidPtr->regionBufferPool.bufferSize = raidPtr->regionLogCapacity * 315 raidPtr->bytesPerSector; 316 printf("regionBufferPool.bufferSize %d\n", 317 raidPtr->regionBufferPool.bufferSize); 318 319 /* for now, only one region at a time may be reintegrated */ 320 raidPtr->regionBufferPool.totalBuffers = 1; 321 322 raidPtr->regionBufferPool.availableBuffers = 323 raidPtr->regionBufferPool.totalBuffers; 324 raidPtr->regionBufferPool.availBuffersIndex = 0; 325 raidPtr->regionBufferPool.emptyBuffersIndex = 0; 326 printf("Allocating %d bytes for regionBufferPool\n", 327 (int) (raidPtr->regionBufferPool.totalBuffers * 328 sizeof(caddr_t))); 329 RF_Malloc(raidPtr->regionBufferPool.buffers, 330 raidPtr->regionBufferPool.totalBuffers * sizeof(caddr_t), 331 (caddr_t *)); 332 if (raidPtr->regionBufferPool.buffers == NULL) { 333 return (ENOMEM); 334 } 335 for (i = 0; i < raidPtr->regionBufferPool.totalBuffers; i++) { 336 printf("Allocating %d bytes for regionBufferPool#%d\n", 337 (int) (raidPtr->regionBufferPool.bufferSize * 338 sizeof(char)), i); 339 RF_Malloc(raidPtr->regionBufferPool.buffers[i], 340 raidPtr->regionBufferPool.bufferSize * sizeof(char), 341 (caddr_t)); 342 if (raidPtr->regionBufferPool.buffers[i] == NULL) { 343 for (j = 0; j < i; j++) { 344 RF_Free(raidPtr->regionBufferPool.buffers[i], 345 raidPtr->regionBufferPool.bufferSize * 346 sizeof(char)); 347 } 348 RF_Free(raidPtr->regionBufferPool.buffers, 349 raidPtr->regionBufferPool.totalBuffers * 350 sizeof(caddr_t)); 351 return (ENOMEM); 352 } 353 printf("raidPtr->regionBufferPool.buffers[%d] = %lx\n", i, 354 (long) raidPtr->regionBufferPool.buffers[i]); 355 } 356 rc = rf_ShutdownCreate(listp, 357 rf_ShutdownParityLoggingRegionBufferPool, 358 raidPtr); 359 if (rc) { 360 RF_ERRORMSG3("Unable to create shutdown entry file %s line %d rc=%d\n", __FILE__, 361 __LINE__, rc); 362 rf_ShutdownParityLoggingRegionBufferPool(raidPtr); 363 return (rc); 364 } 365 /* build pool of parity buffers */ 366 parityBufferCapacity = maxRegionParityRange; 367 rc = rf_mutex_init(&raidPtr->parityBufferPool.mutex); 368 if (rc) { 369 rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc); 370 return (rc); 371 } 372 raidPtr->parityBufferPool.cond = 0; 373 raidPtr->parityBufferPool.bufferSize = parityBufferCapacity * 374 raidPtr->bytesPerSector; 375 printf("parityBufferPool.bufferSize %d\n", 376 raidPtr->parityBufferPool.bufferSize); 377 378 /* for now, only one region at a time may be reintegrated */ 379 raidPtr->parityBufferPool.totalBuffers = 1; 380 381 raidPtr->parityBufferPool.availableBuffers = 382 raidPtr->parityBufferPool.totalBuffers; 383 raidPtr->parityBufferPool.availBuffersIndex = 0; 384 raidPtr->parityBufferPool.emptyBuffersIndex = 0; 385 printf("Allocating %d bytes for parityBufferPool of %d units\n", 386 (int) (raidPtr->parityBufferPool.totalBuffers * 387 sizeof(caddr_t)), 388 raidPtr->parityBufferPool.totalBuffers ); 389 RF_Malloc(raidPtr->parityBufferPool.buffers, 390 raidPtr->parityBufferPool.totalBuffers * sizeof(caddr_t), 391 (caddr_t *)); 392 if (raidPtr->parityBufferPool.buffers == NULL) { 393 return (ENOMEM); 394 } 395 for (i = 0; i < raidPtr->parityBufferPool.totalBuffers; i++) { 396 printf("Allocating %d bytes for parityBufferPool#%d\n", 397 (int) (raidPtr->parityBufferPool.bufferSize * 398 sizeof(char)),i); 399 RF_Malloc(raidPtr->parityBufferPool.buffers[i], 400 raidPtr->parityBufferPool.bufferSize * sizeof(char), 401 (caddr_t)); 402 if (raidPtr->parityBufferPool.buffers == NULL) { 403 for (j = 0; j < i; j++) { 404 RF_Free(raidPtr->parityBufferPool.buffers[i], 405 raidPtr->regionBufferPool.bufferSize * 406 sizeof(char)); 407 } 408 RF_Free(raidPtr->parityBufferPool.buffers, 409 raidPtr->regionBufferPool.totalBuffers * 410 sizeof(caddr_t)); 411 return (ENOMEM); 412 } 413 printf("parityBufferPool.buffers[%d] = %lx\n", i, 414 (long) raidPtr->parityBufferPool.buffers[i]); 415 } 416 rc = rf_ShutdownCreate(listp, 417 rf_ShutdownParityLoggingParityBufferPool, 418 raidPtr); 419 if (rc) { 420 RF_ERRORMSG3("Unable to create shutdown entry file %s line %d rc=%d\n", __FILE__, 421 __LINE__, rc); 422 rf_ShutdownParityLoggingParityBufferPool(raidPtr); 423 return (rc); 424 } 425 /* initialize parityLogDiskQueue */ 426 rf_mutex_init(&raidPtr->parityLogDiskQueue.mutex); 427 rc = rf_create_managed_cond(listp, &raidPtr->parityLogDiskQueue.cond); 428 if (rc) { 429 rf_print_unable_to_init_cond(__FILE__, __LINE__, rc); 430 return (rc); 431 } 432 raidPtr->parityLogDiskQueue.flushQueue = NULL; 433 raidPtr->parityLogDiskQueue.reintQueue = NULL; 434 raidPtr->parityLogDiskQueue.bufHead = NULL; 435 raidPtr->parityLogDiskQueue.bufTail = NULL; 436 raidPtr->parityLogDiskQueue.reintHead = NULL; 437 raidPtr->parityLogDiskQueue.reintTail = NULL; 438 raidPtr->parityLogDiskQueue.logBlockHead = NULL; 439 raidPtr->parityLogDiskQueue.logBlockTail = NULL; 440 raidPtr->parityLogDiskQueue.reintBlockHead = NULL; 441 raidPtr->parityLogDiskQueue.reintBlockTail = NULL; 442 raidPtr->parityLogDiskQueue.freeDataList = NULL; 443 raidPtr->parityLogDiskQueue.freeCommonList = NULL; 444 445 rc = rf_ShutdownCreate(listp, 446 rf_ShutdownParityLoggingDiskQueue, 447 raidPtr); 448 if (rc) { 449 RF_ERRORMSG3("Unable to create shutdown entry file %s line %d rc=%d\n", __FILE__, 450 __LINE__, rc); 451 return (rc); 452 } 453 for (i = 0; i < rf_numParityRegions; i++) { 454 rc = rf_mutex_init(&raidPtr->regionInfo[i].mutex); 455 if (rc) { 456 rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc); 457 for (j = 0; j < i; j++) 458 FreeRegionInfo(raidPtr, j); 459 RF_Free(raidPtr->regionInfo, 460 (rf_numParityRegions * 461 sizeof(RF_RegionInfo_t))); 462 return (ENOMEM); 463 } 464 rc = rf_mutex_init(&raidPtr->regionInfo[i].reintMutex); 465 if (rc) { 466 rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc); 467 for (j = 0; j < i; j++) 468 FreeRegionInfo(raidPtr, j); 469 RF_Free(raidPtr->regionInfo, 470 (rf_numParityRegions * 471 sizeof(RF_RegionInfo_t))); 472 return (ENOMEM); 473 } 474 raidPtr->regionInfo[i].reintInProgress = RF_FALSE; 475 raidPtr->regionInfo[i].regionStartAddr = 476 raidPtr->regionLogCapacity * i; 477 raidPtr->regionInfo[i].parityStartAddr = 478 raidPtr->regionParityRange * i; 479 if (i < rf_numParityRegions - 1) { 480 raidPtr->regionInfo[i].capacity = 481 raidPtr->regionLogCapacity; 482 raidPtr->regionInfo[i].numSectorsParity = 483 raidPtr->regionParityRange; 484 } else { 485 raidPtr->regionInfo[i].capacity = 486 lastRegionCapacity; 487 raidPtr->regionInfo[i].numSectorsParity = 488 raidPtr->sectorsPerDisk - 489 raidPtr->regionParityRange * i; 490 if (raidPtr->regionInfo[i].numSectorsParity > 491 maxRegionParityRange) 492 maxRegionParityRange = 493 raidPtr->regionInfo[i].numSectorsParity; 494 } 495 raidPtr->regionInfo[i].diskCount = 0; 496 RF_ASSERT(raidPtr->regionInfo[i].capacity + 497 raidPtr->regionInfo[i].regionStartAddr <= 498 totalLogCapacity); 499 RF_ASSERT(raidPtr->regionInfo[i].parityStartAddr + 500 raidPtr->regionInfo[i].numSectorsParity <= 501 raidPtr->sectorsPerDisk); 502 printf("Allocating %d bytes for region %d\n", 503 (int) (raidPtr->regionInfo[i].capacity * 504 sizeof(RF_DiskMap_t)), i); 505 RF_Malloc(raidPtr->regionInfo[i].diskMap, 506 (raidPtr->regionInfo[i].capacity * 507 sizeof(RF_DiskMap_t)), 508 (RF_DiskMap_t *)); 509 if (raidPtr->regionInfo[i].diskMap == NULL) { 510 for (j = 0; j < i; j++) 511 FreeRegionInfo(raidPtr, j); 512 RF_Free(raidPtr->regionInfo, 513 (rf_numParityRegions * 514 sizeof(RF_RegionInfo_t))); 515 return (ENOMEM); 516 } 517 raidPtr->regionInfo[i].loggingEnabled = RF_FALSE; 518 raidPtr->regionInfo[i].coreLog = NULL; 519 } 520 rc = rf_ShutdownCreate(listp, 521 rf_ShutdownParityLoggingRegionInfo, 522 raidPtr); 523 if (rc) { 524 RF_ERRORMSG3("Unable to create shutdown entry file %s line %d rc=%d\n", __FILE__, 525 __LINE__, rc); 526 rf_ShutdownParityLoggingRegionInfo(raidPtr); 527 return (rc); 528 } 529 RF_ASSERT(raidPtr->parityLogDiskQueue.threadState == 0); 530 raidPtr->parityLogDiskQueue.threadState = RF_PLOG_CREATED; 531 rc = RF_CREATE_THREAD(raidPtr->pLogDiskThreadHandle, 532 rf_ParityLoggingDiskManager, raidPtr,"rf_log"); 533 if (rc) { 534 raidPtr->parityLogDiskQueue.threadState = 0; 535 RF_ERRORMSG3("Unable to create parity logging disk thread file %s line %d rc=%d\n", 536 __FILE__, __LINE__, rc); 537 return (ENOMEM); 538 } 539 /* wait for thread to start */ 540 RF_LOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex); 541 while (!(raidPtr->parityLogDiskQueue.threadState & RF_PLOG_RUNNING)) { 542 RF_WAIT_COND(raidPtr->parityLogDiskQueue.cond, 543 raidPtr->parityLogDiskQueue.mutex); 544 } 545 RF_UNLOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex); 546 547 rc = rf_ShutdownCreate(listp, rf_ShutdownParityLogging, raidPtr); 548 if (rc) { 549 RF_ERRORMSG1("Got rc=%d adding parity logging shutdown event\n", rc); 550 rf_ShutdownParityLogging(raidPtr); 551 return (rc); 552 } 553 if (rf_parityLogDebug) { 554 printf(" size of disk log in sectors: %d\n", 555 (int) totalLogCapacity); 556 printf(" total number of parity regions is %d\n", (int) rf_numParityRegions); 557 printf(" nominal sectors of log per parity region is %d\n", (int) raidPtr->regionLogCapacity); 558 printf(" nominal region fragmentation is %d sectors\n", (int) fragmentation); 559 printf(" total number of parity logs is %d\n", raidPtr->numParityLogs); 560 printf(" parity log size is %d sectors\n", raidPtr->numSectorsPerLog); 561 printf(" total in-core log space is %d bytes\n", (int) rf_totalInCoreLogCapacity); 562 } 563 rf_EnableParityLogging(raidPtr); 564 565 return (0); 566 } 567 568 static void 569 FreeRegionInfo( 570 RF_Raid_t * raidPtr, 571 RF_RegionId_t regionID) 572 { 573 RF_LOCK_MUTEX(raidPtr->regionInfo[regionID].mutex); 574 RF_Free(raidPtr->regionInfo[regionID].diskMap, 575 (raidPtr->regionInfo[regionID].capacity * 576 sizeof(RF_DiskMap_t))); 577 if (!rf_forceParityLogReint && raidPtr->regionInfo[regionID].coreLog) { 578 rf_ReleaseParityLogs(raidPtr, 579 raidPtr->regionInfo[regionID].coreLog); 580 raidPtr->regionInfo[regionID].coreLog = NULL; 581 } else { 582 RF_ASSERT(raidPtr->regionInfo[regionID].coreLog == NULL); 583 RF_ASSERT(raidPtr->regionInfo[regionID].diskCount == 0); 584 } 585 RF_UNLOCK_MUTEX(raidPtr->regionInfo[regionID].mutex); 586 } 587 588 589 static void 590 FreeParityLogQueue( 591 RF_Raid_t * raidPtr, 592 RF_ParityLogQueue_t * queue) 593 { 594 RF_ParityLog_t *l1, *l2; 595 596 RF_LOCK_MUTEX(queue->mutex); 597 l1 = queue->parityLogs; 598 while (l1) { 599 l2 = l1; 600 l1 = l2->next; 601 RF_Free(l2->records, (raidPtr->numSectorsPerLog * 602 sizeof(RF_ParityLogRecord_t))); 603 RF_Free(l2, sizeof(RF_ParityLog_t)); 604 } 605 RF_UNLOCK_MUTEX(queue->mutex); 606 } 607 608 609 static void 610 FreeRegionBufferQueue(RF_RegionBufferQueue_t * queue) 611 { 612 int i; 613 614 RF_LOCK_MUTEX(queue->mutex); 615 if (queue->availableBuffers != queue->totalBuffers) { 616 printf("Attempt to free region queue which is still in use!\n"); 617 RF_ASSERT(0); 618 } 619 for (i = 0; i < queue->totalBuffers; i++) 620 RF_Free(queue->buffers[i], queue->bufferSize); 621 RF_Free(queue->buffers, queue->totalBuffers * sizeof(caddr_t)); 622 RF_UNLOCK_MUTEX(queue->mutex); 623 } 624 625 static void 626 rf_ShutdownParityLoggingRegionInfo(RF_ThreadArg_t arg) 627 { 628 RF_Raid_t *raidPtr; 629 RF_RegionId_t i; 630 631 raidPtr = (RF_Raid_t *) arg; 632 if (rf_parityLogDebug) { 633 printf("raid%d: ShutdownParityLoggingRegionInfo\n", 634 raidPtr->raidid); 635 } 636 /* free region information structs */ 637 for (i = 0; i < rf_numParityRegions; i++) 638 FreeRegionInfo(raidPtr, i); 639 RF_Free(raidPtr->regionInfo, (rf_numParityRegions * 640 sizeof(raidPtr->regionInfo))); 641 raidPtr->regionInfo = NULL; 642 } 643 644 static void 645 rf_ShutdownParityLoggingPool(RF_ThreadArg_t arg) 646 { 647 RF_Raid_t *raidPtr; 648 649 raidPtr = (RF_Raid_t *) arg; 650 if (rf_parityLogDebug) { 651 printf("raid%d: ShutdownParityLoggingPool\n", raidPtr->raidid); 652 } 653 /* free contents of parityLogPool */ 654 FreeParityLogQueue(raidPtr, &raidPtr->parityLogPool); 655 RF_Free(raidPtr->parityLogBufferHeap, raidPtr->numParityLogs * 656 raidPtr->numSectorsPerLog * raidPtr->bytesPerSector); 657 } 658 659 static void 660 rf_ShutdownParityLoggingRegionBufferPool(RF_ThreadArg_t arg) 661 { 662 RF_Raid_t *raidPtr; 663 664 raidPtr = (RF_Raid_t *) arg; 665 if (rf_parityLogDebug) { 666 printf("raid%d: ShutdownParityLoggingRegionBufferPool\n", 667 raidPtr->raidid); 668 } 669 FreeRegionBufferQueue(&raidPtr->regionBufferPool); 670 } 671 672 static void 673 rf_ShutdownParityLoggingParityBufferPool(RF_ThreadArg_t arg) 674 { 675 RF_Raid_t *raidPtr; 676 677 raidPtr = (RF_Raid_t *) arg; 678 if (rf_parityLogDebug) { 679 printf("raid%d: ShutdownParityLoggingParityBufferPool\n", 680 raidPtr->raidid); 681 } 682 FreeRegionBufferQueue(&raidPtr->parityBufferPool); 683 } 684 685 static void 686 rf_ShutdownParityLoggingDiskQueue(RF_ThreadArg_t arg) 687 { 688 RF_ParityLogData_t *d; 689 RF_CommonLogData_t *c; 690 RF_Raid_t *raidPtr; 691 692 raidPtr = (RF_Raid_t *) arg; 693 if (rf_parityLogDebug) { 694 printf("raid%d: ShutdownParityLoggingDiskQueue\n", 695 raidPtr->raidid); 696 } 697 /* free disk manager stuff */ 698 RF_ASSERT(raidPtr->parityLogDiskQueue.bufHead == NULL); 699 RF_ASSERT(raidPtr->parityLogDiskQueue.bufTail == NULL); 700 RF_ASSERT(raidPtr->parityLogDiskQueue.reintHead == NULL); 701 RF_ASSERT(raidPtr->parityLogDiskQueue.reintTail == NULL); 702 while (raidPtr->parityLogDiskQueue.freeDataList) { 703 d = raidPtr->parityLogDiskQueue.freeDataList; 704 raidPtr->parityLogDiskQueue.freeDataList = 705 raidPtr->parityLogDiskQueue.freeDataList->next; 706 RF_Free(d, sizeof(RF_ParityLogData_t)); 707 } 708 while (raidPtr->parityLogDiskQueue.freeCommonList) { 709 c = raidPtr->parityLogDiskQueue.freeCommonList; 710 raidPtr->parityLogDiskQueue.freeCommonList = 711 raidPtr->parityLogDiskQueue.freeCommonList->next; 712 RF_Free(c, sizeof(RF_CommonLogData_t)); 713 } 714 } 715 716 static void 717 rf_ShutdownParityLogging(RF_ThreadArg_t arg) 718 { 719 RF_Raid_t *raidPtr; 720 721 raidPtr = (RF_Raid_t *) arg; 722 if (rf_parityLogDebug) { 723 printf("raid%d: ShutdownParityLogging\n", raidPtr->raidid); 724 } 725 /* shutdown disk thread */ 726 /* This has the desirable side-effect of forcing all regions to be 727 * reintegrated. This is necessary since all parity log maps are 728 * currently held in volatile memory. */ 729 730 RF_LOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex); 731 raidPtr->parityLogDiskQueue.threadState |= RF_PLOG_TERMINATE; 732 RF_UNLOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex); 733 RF_SIGNAL_COND(raidPtr->parityLogDiskQueue.cond); 734 /* 735 * pLogDiskThread will now terminate when queues are cleared 736 * now wait for it to be done 737 */ 738 RF_LOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex); 739 while (!(raidPtr->parityLogDiskQueue.threadState & RF_PLOG_SHUTDOWN)) { 740 RF_WAIT_COND(raidPtr->parityLogDiskQueue.cond, 741 raidPtr->parityLogDiskQueue.mutex); 742 } 743 RF_UNLOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex); 744 if (rf_parityLogDebug) { 745 printf("raid%d: ShutdownParityLogging done (thread completed)\n", raidPtr->raidid); 746 } 747 } 748 749 int 750 rf_GetDefaultNumFloatingReconBuffersParityLogging(RF_Raid_t * raidPtr) 751 { 752 return (20); 753 } 754 755 RF_HeadSepLimit_t 756 rf_GetDefaultHeadSepLimitParityLogging(RF_Raid_t * raidPtr) 757 { 758 return (10); 759 } 760 /* return the region ID for a given RAID address */ 761 RF_RegionId_t 762 rf_MapRegionIDParityLogging( 763 RF_Raid_t * raidPtr, 764 RF_SectorNum_t address) 765 { 766 RF_RegionId_t regionID; 767 768 /* regionID = address / (raidPtr->regionParityRange * raidPtr->Layout.numDataCol); */ 769 regionID = address / raidPtr->regionParityRange; 770 if (regionID == rf_numParityRegions) { 771 /* last region may be larger than other regions */ 772 regionID--; 773 } 774 RF_ASSERT(address >= raidPtr->regionInfo[regionID].parityStartAddr); 775 RF_ASSERT(address < raidPtr->regionInfo[regionID].parityStartAddr + 776 raidPtr->regionInfo[regionID].numSectorsParity); 777 RF_ASSERT(regionID < rf_numParityRegions); 778 return (regionID); 779 } 780 781 782 /* given a logical RAID sector, determine physical disk address of data */ 783 void 784 rf_MapSectorParityLogging( 785 RF_Raid_t * raidPtr, 786 RF_RaidAddr_t raidSector, 787 RF_RowCol_t * col, 788 RF_SectorNum_t * diskSector, 789 int remap) 790 { 791 RF_StripeNum_t SUID = raidSector / 792 raidPtr->Layout.sectorsPerStripeUnit; 793 /* *col = (SUID % (raidPtr->numCol - 794 * raidPtr->Layout.numParityLogCol)); */ 795 *col = SUID % raidPtr->Layout.numDataCol; 796 *diskSector = (SUID / (raidPtr->Layout.numDataCol)) * 797 raidPtr->Layout.sectorsPerStripeUnit + 798 (raidSector % raidPtr->Layout.sectorsPerStripeUnit); 799 } 800 801 802 /* given a logical RAID sector, determine physical disk address of parity */ 803 void 804 rf_MapParityParityLogging( 805 RF_Raid_t * raidPtr, 806 RF_RaidAddr_t raidSector, 807 RF_RowCol_t * col, 808 RF_SectorNum_t * diskSector, 809 int remap) 810 { 811 RF_StripeNum_t SUID = raidSector / 812 raidPtr->Layout.sectorsPerStripeUnit; 813 814 /* *col = 815 * raidPtr->Layout.numDataCol-(SUID/raidPtr->Layout.numDataCol)%(raidPt 816 * r->numCol - raidPtr->Layout.numParityLogCol); */ 817 *col = raidPtr->Layout.numDataCol; 818 *diskSector = (SUID / (raidPtr->Layout.numDataCol)) * 819 raidPtr->Layout.sectorsPerStripeUnit + 820 (raidSector % raidPtr->Layout.sectorsPerStripeUnit); 821 } 822 823 824 /* given a regionID and sector offset, determine the physical disk address of the parity log */ 825 void 826 rf_MapLogParityLogging( 827 RF_Raid_t * raidPtr, 828 RF_RegionId_t regionID, 829 RF_SectorNum_t regionOffset, 830 RF_RowCol_t * col, 831 RF_SectorNum_t * startSector) 832 { 833 *col = raidPtr->numCol - 1; 834 *startSector = raidPtr->regionInfo[regionID].regionStartAddr + regionOffset; 835 } 836 837 838 /* given a regionID, determine the physical disk address of the logged 839 parity for that region */ 840 void 841 rf_MapRegionParity( 842 RF_Raid_t * raidPtr, 843 RF_RegionId_t regionID, 844 RF_RowCol_t * col, 845 RF_SectorNum_t * startSector, 846 RF_SectorCount_t * numSector) 847 { 848 *col = raidPtr->numCol - 2; 849 *startSector = raidPtr->regionInfo[regionID].parityStartAddr; 850 *numSector = raidPtr->regionInfo[regionID].numSectorsParity; 851 } 852 853 854 /* given a logical RAID address, determine the participating disks in 855 the stripe */ 856 void 857 rf_IdentifyStripeParityLogging( 858 RF_Raid_t * raidPtr, 859 RF_RaidAddr_t addr, 860 RF_RowCol_t ** diskids) 861 { 862 RF_StripeNum_t stripeID = rf_RaidAddressToStripeID(&raidPtr->Layout, 863 addr); 864 RF_ParityLoggingConfigInfo_t *info = (RF_ParityLoggingConfigInfo_t *) 865 raidPtr->Layout.layoutSpecificInfo; 866 *diskids = info->stripeIdentifier[stripeID % raidPtr->numCol]; 867 } 868 869 870 void 871 rf_MapSIDToPSIDParityLogging( 872 RF_RaidLayout_t * layoutPtr, 873 RF_StripeNum_t stripeID, 874 RF_StripeNum_t * psID, 875 RF_ReconUnitNum_t * which_ru) 876 { 877 *which_ru = 0; 878 *psID = stripeID; 879 } 880 881 882 /* select an algorithm for performing an access. Returns two pointers, 883 * one to a function that will return information about the DAG, and 884 * another to a function that will create the dag. 885 */ 886 void 887 rf_ParityLoggingDagSelect( 888 RF_Raid_t * raidPtr, 889 RF_IoType_t type, 890 RF_AccessStripeMap_t * asmp, 891 RF_VoidFuncPtr * createFunc) 892 { 893 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 894 RF_PhysDiskAddr_t *failedPDA = NULL; 895 RF_RowCol_t fcol; 896 RF_RowStatus_t rstat; 897 int prior_recon; 898 899 RF_ASSERT(RF_IO_IS_R_OR_W(type)); 900 901 if (asmp->numDataFailed + asmp->numParityFailed > 1) { 902 RF_ERRORMSG("Multiple disks failed in a single group! Aborting I/O operation.\n"); 903 *createFunc = NULL; 904 return; 905 } else 906 if (asmp->numDataFailed + asmp->numParityFailed == 1) { 907 908 /* if under recon & already reconstructed, redirect 909 * the access to the spare drive and eliminate the 910 * failure indication */ 911 failedPDA = asmp->failedPDAs[0]; 912 fcol = failedPDA->col; 913 rstat = raidPtr->status; 914 prior_recon = (rstat == rf_rs_reconfigured) || ( 915 (rstat == rf_rs_reconstructing) ? 916 rf_CheckRUReconstructed(raidPtr->reconControl->reconMap, failedPDA->startSector) : 0 917 ); 918 if (prior_recon) { 919 RF_RowCol_t oc = failedPDA->col; 920 RF_SectorNum_t oo = failedPDA->startSector; 921 if (layoutPtr->map->flags & 922 RF_DISTRIBUTE_SPARE) { 923 /* redirect to dist spare space */ 924 925 if (failedPDA == asmp->parityInfo) { 926 927 /* parity has failed */ 928 (layoutPtr->map->MapParity) (raidPtr, failedPDA->raidAddress, 929 &failedPDA->col, &failedPDA->startSector, RF_REMAP); 930 931 if (asmp->parityInfo->next) { /* redir 2nd component, 932 * if any */ 933 RF_PhysDiskAddr_t *p = asmp->parityInfo->next; 934 RF_SectorNum_t SUoffs = p->startSector % layoutPtr->sectorsPerStripeUnit; 935 p->col = failedPDA->col; 936 p->startSector = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->startSector) + 937 SUoffs; /* cheating: 938 * startSector is not 939 * really a RAID address */ 940 } 941 } else 942 if (asmp->parityInfo->next && failedPDA == asmp->parityInfo->next) { 943 RF_ASSERT(0); /* should not ever 944 * happen */ 945 } else { 946 947 /* data has failed */ 948 (layoutPtr->map->MapSector) (raidPtr, failedPDA->raidAddress, 949 &failedPDA->col, &failedPDA->startSector, RF_REMAP); 950 951 } 952 953 } else { 954 /* redirect to dedicated spare space */ 955 956 failedPDA->col = raidPtr->Disks[fcol].spareCol; 957 958 /* the parity may have two distinct 959 * components, both of which may need 960 * to be redirected */ 961 if (asmp->parityInfo->next) { 962 if (failedPDA == asmp->parityInfo) { 963 failedPDA->next->col = failedPDA->col; 964 } else 965 if (failedPDA == asmp->parityInfo->next) { /* paranoid: should never occur */ 966 asmp->parityInfo->col = failedPDA->col; 967 } 968 } 969 } 970 971 RF_ASSERT(failedPDA->col != -1); 972 973 if (rf_dagDebug || rf_mapDebug) { 974 printf("raid%d: Redirected type '%c' c %d o %ld -> c %d o %ld\n", 975 raidPtr->raidid, type, oc, (long) oo, failedPDA->col, (long) failedPDA->startSector); 976 } 977 asmp->numDataFailed = asmp->numParityFailed = 0; 978 } 979 } 980 if (type == RF_IO_TYPE_READ) { 981 982 if (asmp->numDataFailed == 0) 983 *createFunc = (RF_VoidFuncPtr) rf_CreateFaultFreeReadDAG; 984 else 985 *createFunc = (RF_VoidFuncPtr) rf_CreateRaidFiveDegradedReadDAG; 986 987 } else { 988 989 990 /* if mirroring, always use large writes. If the access 991 * requires two distinct parity updates, always do a small 992 * write. If the stripe contains a failure but the access 993 * does not, do a small write. The first conditional 994 * (numStripeUnitsAccessed <= numDataCol/2) uses a 995 * less-than-or-equal rather than just a less-than because 996 * when G is 3 or 4, numDataCol/2 is 1, and I want 997 * single-stripe-unit updates to use just one disk. */ 998 if ((asmp->numDataFailed + asmp->numParityFailed) == 0) { 999 if (((asmp->numStripeUnitsAccessed <= 1000 (layoutPtr->numDataCol / 2)) && 1001 (layoutPtr->numDataCol != 1)) || 1002 (asmp->parityInfo->next != NULL) || 1003 rf_CheckStripeForFailures(raidPtr, asmp)) { 1004 *createFunc = (RF_VoidFuncPtr) rf_CreateParityLoggingSmallWriteDAG; 1005 } else 1006 *createFunc = (RF_VoidFuncPtr) rf_CreateParityLoggingLargeWriteDAG; 1007 } else 1008 if (asmp->numParityFailed == 1) 1009 *createFunc = (RF_VoidFuncPtr) rf_CreateNonRedundantWriteDAG; 1010 else 1011 if (asmp->numStripeUnitsAccessed != 1 && failedPDA->numSector != layoutPtr->sectorsPerStripeUnit) 1012 *createFunc = NULL; 1013 else 1014 *createFunc = (RF_VoidFuncPtr) rf_CreateDegradedWriteDAG; 1015 } 1016 } 1017 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 1018
Indexes created Mon Sep 22 21:09:42 GMT 2025