1 /* $NetBSD: rf_dagdegwr.c,v 1.30 2006/11/16 01:33:23 christos Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland, Daniel Stodolsky, 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@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 * rf_dagdegwr.c 31 * 32 * code for creating degraded write DAGs 33 * 34 */ 35 36 #include <sys/cdefs.h> 37 __KERNEL_RCSID(0, "$NetBSD: rf_dagdegwr.c,v 1.30 2006/11/16 01:33:23 christos Exp $"); 38 39 #include <dev/raidframe/raidframevar.h> 40 41 #include "rf_raid.h" 42 #include "rf_dag.h" 43 #include "rf_dagutils.h" 44 #include "rf_dagfuncs.h" 45 #include "rf_debugMem.h" 46 #include "rf_general.h" 47 #include "rf_dagdegwr.h" 48 #include "rf_map.h" 49 50 51 /****************************************************************************** 52 * 53 * General comments on DAG creation: 54 * 55 * All DAGs in this file use roll-away error recovery. Each DAG has a single 56 * commit node, usually called "Cmt." If an error occurs before the Cmt node 57 * is reached, the execution engine will halt forward execution and work 58 * backward through the graph, executing the undo functions. Assuming that 59 * each node in the graph prior to the Cmt node are undoable and atomic - or - 60 * does not make changes to permanent state, the graph will fail atomically. 61 * If an error occurs after the Cmt node executes, the engine will roll-forward 62 * through the graph, blindly executing nodes until it reaches the end. 63 * If a graph reaches the end, it is assumed to have completed successfully. 64 * 65 * A graph has only 1 Cmt node. 66 * 67 */ 68 69 70 /****************************************************************************** 71 * 72 * The following wrappers map the standard DAG creation interface to the 73 * DAG creation routines. Additionally, these wrappers enable experimentation 74 * with new DAG structures by providing an extra level of indirection, allowing 75 * the DAG creation routines to be replaced at this single point. 76 */ 77 78 static 79 RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG) 80 { 81 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, 82 flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE); 83 } 84 85 void 86 rf_CreateDegradedWriteDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 87 RF_DagHeader_t *dag_h, void *bp, 88 RF_RaidAccessFlags_t flags, 89 RF_AllocListElem_t *allocList) 90 { 91 92 RF_ASSERT(asmap->numDataFailed == 1); 93 dag_h->creator = "DegradedWriteDAG"; 94 95 /* 96 * if the access writes only a portion of the failed unit, and also 97 * writes some portion of at least one surviving unit, we create two 98 * DAGs, one for the failed component and one for the non-failed 99 * component, and do them sequentially. Note that the fact that we're 100 * accessing only a portion of the failed unit indicates that the 101 * access either starts or ends in the failed unit, and hence we need 102 * create only two dags. This is inefficient in that the same data or 103 * parity can get read and written twice using this structure. I need 104 * to fix this to do the access all at once. 105 */ 106 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 && 107 asmap->failedPDAs[0]->numSector != 108 raidPtr->Layout.sectorsPerStripeUnit)); 109 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 110 allocList); 111 } 112 113 114 115 /****************************************************************************** 116 * 117 * DAG creation code begins here 118 */ 119 120 121 122 /****************************************************************************** 123 * 124 * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode 125 * write, which is as follows 126 * 127 * / {Wnq} --\ 128 * hdr -> blockNode -> Rod -> Xor -> Cmt -> Wnp ----> unblock -> term 129 * \ {Rod} / \ Wnd ---/ 130 * \ {Wnd} -/ 131 * 132 * commit nodes: Xor, Wnd 133 * 134 * IMPORTANT: 135 * This DAG generator does not work for double-degraded archs since it does not 136 * generate Q 137 * 138 * This dag is essentially identical to the large-write dag, except that the 139 * write to the failed data unit is suppressed. 140 * 141 * IMPORTANT: this dag does not work in the case where the access writes only 142 * a portion of the failed unit, and also writes some portion of at least one 143 * surviving SU. this case is handled in CreateDegradedWriteDAG above. 144 * 145 * The block & unblock nodes are leftovers from a previous version. They 146 * do nothing, but I haven't deleted them because it would be a tremendous 147 * effort to put them back in. 148 * 149 * This dag is used whenever a one of the data units in a write has failed. 150 * If it is the parity unit that failed, the nonredundant write dag (below) 151 * is used. 152 *****************************************************************************/ 153 154 void 155 rf_CommonCreateSimpleDegradedWriteDAG(RF_Raid_t *raidPtr, 156 RF_AccessStripeMap_t *asmap, 157 RF_DagHeader_t *dag_h, void *bp, 158 RF_RaidAccessFlags_t flags, 159 RF_AllocListElem_t *allocList, 160 int nfaults, 161 int (*redFunc) (RF_DagNode_t *), 162 int allowBufferRecycle) 163 { 164 int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 165 rdnodesFaked; 166 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode; 167 RF_DagNode_t *wndNodes, *rrdNodes, *xorNode, *commitNode; 168 RF_DagNode_t *tmpNode, *tmpwndNode, *tmprrdNode; 169 RF_SectorCount_t sectorsPerSU; 170 RF_ReconUnitNum_t which_ru; 171 char *xorTargetBuf = NULL; /* the target buffer for the XOR 172 * operation */ 173 char overlappingPDAs[RF_MAXCOL];/* a temporary array of flags */ 174 RF_AccessStripeMapHeader_t *new_asm_h[2]; 175 RF_PhysDiskAddr_t *pda, *parityPDA; 176 RF_StripeNum_t parityStripeID; 177 RF_PhysDiskAddr_t *failedPDA; 178 RF_RaidLayout_t *layoutPtr; 179 180 layoutPtr = &(raidPtr->Layout); 181 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 182 &which_ru); 183 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 184 /* failedPDA points to the pda within the asm that targets the failed 185 * disk */ 186 failedPDA = asmap->failedPDAs[0]; 187 188 #if RF_DEBUG_DAG 189 if (rf_dagDebug) 190 printf("[Creating degraded-write DAG]\n"); 191 #endif 192 193 RF_ASSERT(asmap->numDataFailed == 1); 194 dag_h->creator = "SimpleDegradedWriteDAG"; 195 196 /* 197 * Generate two ASMs identifying the surviving data 198 * we need in order to recover the lost data. 199 */ 200 /* overlappingPDAs array must be zero'd */ 201 memset(overlappingPDAs, 0, RF_MAXCOL); 202 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 203 &nXorBufs, NULL, overlappingPDAs, allocList); 204 205 /* create all the nodes at once */ 206 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 207 * generated for the 208 * failed pda */ 209 210 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 211 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 212 /* 213 * XXX 214 * 215 * There's a bug with a complete stripe overwrite- that means 0 reads 216 * of old data, and the rest of the DAG generation code doesn't like 217 * that. A release is coming, and I don't wanna risk breaking a critical 218 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 219 * I'm gonna fake there being a read node, and I'm gonna swap in a 220 * no-op node in its place (to make all the link-up code happy). 221 * This should be fixed at some point. --jimz 222 */ 223 if (nRrdNodes == 0) { 224 nRrdNodes = 1; 225 rdnodesFaked = 1; 226 } else { 227 rdnodesFaked = 0; 228 } 229 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 230 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 231 232 blockNode = rf_AllocDAGNode(); 233 blockNode->list_next = dag_h->nodes; 234 dag_h->nodes = blockNode; 235 236 commitNode = rf_AllocDAGNode(); 237 commitNode->list_next = dag_h->nodes; 238 dag_h->nodes = commitNode; 239 240 unblockNode = rf_AllocDAGNode(); 241 unblockNode->list_next = dag_h->nodes; 242 dag_h->nodes = unblockNode; 243 244 termNode = rf_AllocDAGNode(); 245 termNode->list_next = dag_h->nodes; 246 dag_h->nodes = termNode; 247 248 xorNode = rf_AllocDAGNode(); 249 xorNode->list_next = dag_h->nodes; 250 dag_h->nodes = xorNode; 251 252 wnpNode = rf_AllocDAGNode(); 253 wnpNode->list_next = dag_h->nodes; 254 dag_h->nodes = wnpNode; 255 256 for (i = 0; i < nWndNodes; i++) { 257 tmpNode = rf_AllocDAGNode(); 258 tmpNode->list_next = dag_h->nodes; 259 dag_h->nodes = tmpNode; 260 } 261 wndNodes = dag_h->nodes; 262 263 for (i = 0; i < nRrdNodes; i++) { 264 tmpNode = rf_AllocDAGNode(); 265 tmpNode->list_next = dag_h->nodes; 266 dag_h->nodes = tmpNode; 267 } 268 rrdNodes = dag_h->nodes; 269 270 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 271 if (nfaults == 2) { 272 wnqNode = rf_AllocDAGNode(); 273 wnqNode->list_next = dag_h->nodes; 274 dag_h->nodes = wnqNode; 275 } else { 276 #endif 277 wnqNode = NULL; 278 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 279 } 280 #endif 281 282 /* this dag can not commit until all rrd and xor Nodes have completed */ 283 dag_h->numCommitNodes = 1; 284 dag_h->numCommits = 0; 285 dag_h->numSuccedents = 1; 286 287 RF_ASSERT(nRrdNodes > 0); 288 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 289 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 290 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 291 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 292 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 293 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 294 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 295 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 296 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 297 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 298 299 /* 300 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 301 * the failed buffer, save a pointer to it so we can use it as the target 302 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 303 * a buffer is the same size as the failed buffer, it must also be at the 304 * same alignment within the SU. 305 */ 306 i = 0; 307 tmprrdNode = rrdNodes; 308 if (new_asm_h[0]) { 309 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 310 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 311 i++, pda = pda->next) { 312 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 313 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 314 RF_ASSERT(pda); 315 tmprrdNode->params[0].p = pda; 316 tmprrdNode->params[1].p = pda->bufPtr; 317 tmprrdNode->params[2].v = parityStripeID; 318 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 319 tmprrdNode = tmprrdNode->list_next; 320 } 321 } 322 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 323 /* Note that for tmprrdNode, this means a continuation from above, so no need to 324 assign it anything.. */ 325 if (new_asm_h[1]) { 326 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 327 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 328 j++, pda = pda->next) { 329 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 330 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 331 RF_ASSERT(pda); 332 tmprrdNode->params[0].p = pda; 333 tmprrdNode->params[1].p = pda->bufPtr; 334 tmprrdNode->params[2].v = parityStripeID; 335 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 336 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 337 xorTargetBuf = pda->bufPtr; 338 tmprrdNode = tmprrdNode->list_next; 339 } 340 } 341 if (rdnodesFaked) { 342 /* 343 * This is where we'll init that fake noop read node 344 * (XXX should the wakeup func be different?) 345 */ 346 /* node that rrdNodes will just be a single node... */ 347 rf_InitNode(rrdNodes, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 348 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 349 } 350 /* 351 * Make a PDA for the parity unit. The parity PDA should start at 352 * the same offset into the SU as the failed PDA. 353 */ 354 /* Danner comment: I don't think this copy is really necessary. We are 355 * in one of two cases here. (1) The entire failed unit is written. 356 * Then asmap->parityInfo will describe the entire parity. (2) We are 357 * only writing a subset of the failed unit and nothing else. Then the 358 * asmap->parityInfo describes the failed unit and the copy can also 359 * be avoided. */ 360 361 parityPDA = rf_AllocPhysDiskAddr(); 362 parityPDA->next = dag_h->pda_cleanup_list; 363 dag_h->pda_cleanup_list = parityPDA; 364 parityPDA->col = asmap->parityInfo->col; 365 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 366 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 367 parityPDA->numSector = failedPDA->numSector; 368 369 if (!xorTargetBuf) { 370 xorTargetBuf = rf_AllocBuffer(raidPtr, dag_h, rf_RaidAddressToByte(raidPtr, failedPDA->numSector)); 371 } 372 /* init the Wnp node */ 373 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 374 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 375 wnpNode->params[0].p = parityPDA; 376 wnpNode->params[1].p = xorTargetBuf; 377 wnpNode->params[2].v = parityStripeID; 378 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 379 380 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 381 /* fill in the Wnq Node */ 382 if (nfaults == 2) { 383 { 384 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 385 (RF_PhysDiskAddr_t *), allocList); 386 parityPDA->col = asmap->qInfo->col; 387 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 388 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 389 parityPDA->numSector = failedPDA->numSector; 390 391 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 392 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 393 wnqNode->params[0].p = parityPDA; 394 RF_MallocAndAdd(xorNode->results[1], 395 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 396 wnqNode->params[1].p = xorNode->results[1]; 397 wnqNode->params[2].v = parityStripeID; 398 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 399 } 400 } 401 #endif 402 /* fill in the Wnd nodes */ 403 tmpwndNode = wndNodes; 404 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 405 if (pda == failedPDA) { 406 i--; 407 continue; 408 } 409 rf_InitNode(tmpwndNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 410 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 411 RF_ASSERT(pda); 412 tmpwndNode->params[0].p = pda; 413 tmpwndNode->params[1].p = pda->bufPtr; 414 tmpwndNode->params[2].v = parityStripeID; 415 tmpwndNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 416 tmpwndNode = tmpwndNode->list_next; 417 } 418 419 /* fill in the results of the xor node */ 420 xorNode->results[0] = xorTargetBuf; 421 422 /* fill in the params of the xor node */ 423 424 paramNum = 0; 425 if (rdnodesFaked == 0) { 426 tmprrdNode = rrdNodes; 427 for (i = 0; i < nRrdNodes; i++) { 428 /* all the Rrd nodes need to be xored together */ 429 xorNode->params[paramNum++] = tmprrdNode->params[0]; 430 xorNode->params[paramNum++] = tmprrdNode->params[1]; 431 tmprrdNode = tmprrdNode->list_next; 432 } 433 } 434 tmpwndNode = wndNodes; 435 for (i = 0; i < nWndNodes; i++) { 436 /* any Wnd nodes that overlap the failed access need to be 437 * xored in */ 438 if (overlappingPDAs[i]) { 439 pda = rf_AllocPhysDiskAddr(); 440 memcpy((char *) pda, (char *) tmpwndNode->params[0].p, sizeof(RF_PhysDiskAddr_t)); 441 /* add it into the pda_cleanup_list *after* the copy, TYVM */ 442 pda->next = dag_h->pda_cleanup_list; 443 dag_h->pda_cleanup_list = pda; 444 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 445 xorNode->params[paramNum++].p = pda; 446 xorNode->params[paramNum++].p = pda->bufPtr; 447 } 448 tmpwndNode = tmpwndNode->list_next; 449 } 450 451 /* 452 * Install the failed PDA into the xor param list so that the 453 * new data gets xor'd in. 454 */ 455 xorNode->params[paramNum++].p = failedPDA; 456 xorNode->params[paramNum++].p = failedPDA->bufPtr; 457 458 /* 459 * The last 2 params to the recovery xor node are always the failed 460 * PDA and the raidPtr. install the failedPDA even though we have just 461 * done so above. This allows us to use the same XOR function for both 462 * degraded reads and degraded writes. 463 */ 464 xorNode->params[paramNum++].p = failedPDA; 465 xorNode->params[paramNum++].p = raidPtr; 466 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 467 468 /* 469 * Code to link nodes begins here 470 */ 471 472 /* link header to block node */ 473 RF_ASSERT(blockNode->numAntecedents == 0); 474 dag_h->succedents[0] = blockNode; 475 476 /* link block node to rd nodes */ 477 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 478 tmprrdNode = rrdNodes; 479 for (i = 0; i < nRrdNodes; i++) { 480 RF_ASSERT(tmprrdNode->numAntecedents == 1); 481 blockNode->succedents[i] = tmprrdNode; 482 tmprrdNode->antecedents[0] = blockNode; 483 tmprrdNode->antType[0] = rf_control; 484 tmprrdNode = tmprrdNode->list_next; 485 } 486 487 /* link read nodes to xor node */ 488 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 489 tmprrdNode = rrdNodes; 490 for (i = 0; i < nRrdNodes; i++) { 491 RF_ASSERT(tmprrdNode->numSuccedents == 1); 492 tmprrdNode->succedents[0] = xorNode; 493 xorNode->antecedents[i] = tmprrdNode; 494 xorNode->antType[i] = rf_trueData; 495 tmprrdNode = tmprrdNode->list_next; 496 } 497 498 /* link xor node to commit node */ 499 RF_ASSERT(xorNode->numSuccedents == 1); 500 RF_ASSERT(commitNode->numAntecedents == 1); 501 xorNode->succedents[0] = commitNode; 502 commitNode->antecedents[0] = xorNode; 503 commitNode->antType[0] = rf_control; 504 505 /* link commit node to wnd nodes */ 506 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 507 tmpwndNode = wndNodes; 508 for (i = 0; i < nWndNodes; i++) { 509 RF_ASSERT(tmpwndNode->numAntecedents == 1); 510 commitNode->succedents[i] = tmpwndNode; 511 tmpwndNode->antecedents[0] = commitNode; 512 tmpwndNode->antType[0] = rf_control; 513 tmpwndNode = tmpwndNode->list_next; 514 } 515 516 /* link the commit node to wnp, wnq nodes */ 517 RF_ASSERT(wnpNode->numAntecedents == 1); 518 commitNode->succedents[nWndNodes] = wnpNode; 519 wnpNode->antecedents[0] = commitNode; 520 wnpNode->antType[0] = rf_control; 521 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 522 if (nfaults == 2) { 523 RF_ASSERT(wnqNode->numAntecedents == 1); 524 commitNode->succedents[nWndNodes + 1] = wnqNode; 525 wnqNode->antecedents[0] = commitNode; 526 wnqNode->antType[0] = rf_control; 527 } 528 #endif 529 /* link write new data nodes to unblock node */ 530 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 531 tmpwndNode = wndNodes; 532 for (i = 0; i < nWndNodes; i++) { 533 RF_ASSERT(tmpwndNode->numSuccedents == 1); 534 tmpwndNode->succedents[0] = unblockNode; 535 unblockNode->antecedents[i] = tmpwndNode; 536 unblockNode->antType[i] = rf_control; 537 tmpwndNode = tmpwndNode->list_next; 538 } 539 540 /* link write new parity node to unblock node */ 541 RF_ASSERT(wnpNode->numSuccedents == 1); 542 wnpNode->succedents[0] = unblockNode; 543 unblockNode->antecedents[nWndNodes] = wnpNode; 544 unblockNode->antType[nWndNodes] = rf_control; 545 546 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 547 /* link write new q node to unblock node */ 548 if (nfaults == 2) { 549 RF_ASSERT(wnqNode->numSuccedents == 1); 550 wnqNode->succedents[0] = unblockNode; 551 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 552 unblockNode->antType[nWndNodes + 1] = rf_control; 553 } 554 #endif 555 /* link unblock node to term node */ 556 RF_ASSERT(unblockNode->numSuccedents == 1); 557 RF_ASSERT(termNode->numAntecedents == 1); 558 RF_ASSERT(termNode->numSuccedents == 0); 559 unblockNode->succedents[0] = termNode; 560 termNode->antecedents[0] = unblockNode; 561 termNode->antType[0] = rf_control; 562 } 563 #define CONS_PDA(if,start,num) \ 564 pda_p->col = asmap->if->col; \ 565 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 566 pda_p->numSector = num; \ 567 pda_p->next = NULL; \ 568 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 569 #if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 570 void 571 rf_WriteGenerateFailedAccessASMs( 572 RF_Raid_t * raidPtr, 573 RF_AccessStripeMap_t * asmap, 574 RF_PhysDiskAddr_t ** pdap, 575 int *nNodep, 576 RF_PhysDiskAddr_t ** pqpdap, 577 int *nPQNodep, 578 RF_AllocListElem_t * allocList) 579 { 580 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 581 int PDAPerDisk, i; 582 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 583 int numDataCol = layoutPtr->numDataCol; 584 int state; 585 unsigned napdas; 586 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 587 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 588 RF_PhysDiskAddr_t *pda_p; 589 RF_RaidAddr_t sosAddr; 590 591 /* determine how many pda's we will have to generate per unaccess 592 * stripe. If there is only one failed data unit, it is one; if two, 593 * possibly two, depending wether they overlap. */ 594 595 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 596 fone_end = fone_start + fone->numSector; 597 598 if (asmap->numDataFailed == 1) { 599 PDAPerDisk = 1; 600 state = 1; 601 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 602 pda_p = *pqpdap; 603 /* build p */ 604 CONS_PDA(parityInfo, fone_start, fone->numSector); 605 pda_p->type = RF_PDA_TYPE_PARITY; 606 pda_p++; 607 /* build q */ 608 CONS_PDA(qInfo, fone_start, fone->numSector); 609 pda_p->type = RF_PDA_TYPE_Q; 610 } else { 611 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 612 ftwo_end = ftwo_start + ftwo->numSector; 613 if (fone->numSector + ftwo->numSector > secPerSU) { 614 PDAPerDisk = 1; 615 state = 2; 616 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 617 pda_p = *pqpdap; 618 CONS_PDA(parityInfo, 0, secPerSU); 619 pda_p->type = RF_PDA_TYPE_PARITY; 620 pda_p++; 621 CONS_PDA(qInfo, 0, secPerSU); 622 pda_p->type = RF_PDA_TYPE_Q; 623 } else { 624 PDAPerDisk = 2; 625 state = 3; 626 /* four of them, fone, then ftwo */ 627 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 628 pda_p = *pqpdap; 629 CONS_PDA(parityInfo, fone_start, fone->numSector); 630 pda_p->type = RF_PDA_TYPE_PARITY; 631 pda_p++; 632 CONS_PDA(qInfo, fone_start, fone->numSector); 633 pda_p->type = RF_PDA_TYPE_Q; 634 pda_p++; 635 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 636 pda_p->type = RF_PDA_TYPE_PARITY; 637 pda_p++; 638 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 639 pda_p->type = RF_PDA_TYPE_Q; 640 } 641 } 642 /* figure out number of nonaccessed pda */ 643 napdas = PDAPerDisk * (numDataCol - 2); 644 *nPQNodep = PDAPerDisk; 645 646 *nNodep = napdas; 647 if (napdas == 0) 648 return; /* short circuit */ 649 650 /* allocate up our list of pda's */ 651 652 RF_MallocAndAdd(pda_p, napdas * sizeof(RF_PhysDiskAddr_t), 653 (RF_PhysDiskAddr_t *), allocList); 654 *pdap = pda_p; 655 656 /* linkem together */ 657 for (i = 0; i < (napdas - 1); i++) 658 pda_p[i].next = pda_p + (i + 1); 659 660 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 661 for (i = 0; i < numDataCol; i++) { 662 if ((pda_p - (*pdap)) == napdas) 663 continue; 664 pda_p->type = RF_PDA_TYPE_DATA; 665 pda_p->raidAddress = sosAddr + (i * secPerSU); 666 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 667 /* skip over dead disks */ 668 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->col].status)) 669 continue; 670 switch (state) { 671 case 1: /* fone */ 672 pda_p->numSector = fone->numSector; 673 pda_p->raidAddress += fone_start; 674 pda_p->startSector += fone_start; 675 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 676 break; 677 case 2: /* full stripe */ 678 pda_p->numSector = secPerSU; 679 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 680 break; 681 case 3: /* two slabs */ 682 pda_p->numSector = fone->numSector; 683 pda_p->raidAddress += fone_start; 684 pda_p->startSector += fone_start; 685 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 686 pda_p++; 687 pda_p->type = RF_PDA_TYPE_DATA; 688 pda_p->raidAddress = sosAddr + (i * secPerSU); 689 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 690 pda_p->numSector = ftwo->numSector; 691 pda_p->raidAddress += ftwo_start; 692 pda_p->startSector += ftwo_start; 693 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 694 break; 695 default: 696 RF_PANIC(); 697 } 698 pda_p++; 699 } 700 701 RF_ASSERT(pda_p - *pdap == napdas); 702 return; 703 } 704 #define DISK_NODE_PDA(node) ((node)->params[0].p) 705 706 #define DISK_NODE_PARAMS(_node_,_p_) \ 707 (_node_).params[0].p = _p_ ; \ 708 (_node_).params[1].p = (_p_)->bufPtr; \ 709 (_node_).params[2].v = parityStripeID; \ 710 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru) 711 712 void 713 rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 714 RF_DagHeader_t *dag_h, void *bp, 715 RF_RaidAccessFlags_t flags, 716 RF_AllocListElem_t *allocList, 717 const char *redundantReadNodeName, 718 const char *redundantWriteNodeName, 719 const char *recoveryNodeName, 720 int (*recovFunc) (RF_DagNode_t *)) 721 { 722 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 723 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 724 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 725 RF_PhysDiskAddr_t *pda, *pqPDAs; 726 RF_PhysDiskAddr_t *npdas; 727 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 728 RF_ReconUnitNum_t which_ru; 729 int nPQNodes; 730 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 731 732 /* simple small write case - First part looks like a reconstruct-read 733 * of the failed data units. Then a write of all data units not 734 * failed. */ 735 736 737 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 738 * / -------PQ----- / \ \ Wud Wp WQ \ | / 739 * --Unblock- | T 740 * 741 * Rrd = read recovery data (potentially none) Wud = write user data 742 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 743 * (could be two) 744 * 745 */ 746 747 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 748 749 RF_ASSERT(asmap->numDataFailed == 1); 750 751 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 752 nReadNodes = nRrdNodes + 2 * nPQNodes; 753 nWriteNodes = nWudNodes + 2 * nPQNodes; 754 nNodes = 4 + nReadNodes + nWriteNodes; 755 756 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 757 blockNode = nodes; 758 unblockNode = blockNode + 1; 759 termNode = unblockNode + 1; 760 recoveryNode = termNode + 1; 761 rrdNodes = recoveryNode + 1; 762 rpNodes = rrdNodes + nRrdNodes; 763 rqNodes = rpNodes + nPQNodes; 764 wudNodes = rqNodes + nPQNodes; 765 wpNodes = wudNodes + nWudNodes; 766 wqNodes = wpNodes + nPQNodes; 767 768 dag_h->creator = "PQ_DDSimpleSmallWrite"; 769 dag_h->numSuccedents = 1; 770 dag_h->succedents[0] = blockNode; 771 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 772 termNode->antecedents[0] = unblockNode; 773 termNode->antType[0] = rf_control; 774 775 /* init the block and unblock nodes */ 776 /* The block node has all the read nodes as successors */ 777 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 778 for (i = 0; i < nReadNodes; i++) 779 blockNode->succedents[i] = rrdNodes + i; 780 781 /* The unblock node has all the writes as successors */ 782 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 783 for (i = 0; i < nWriteNodes; i++) { 784 unblockNode->antecedents[i] = wudNodes + i; 785 unblockNode->antType[i] = rf_control; 786 } 787 unblockNode->succedents[0] = termNode; 788 789 #define INIT_READ_NODE(node,name) \ 790 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 791 (node)->succedents[0] = recoveryNode; \ 792 (node)->antecedents[0] = blockNode; \ 793 (node)->antType[0] = rf_control; 794 795 /* build the read nodes */ 796 pda = npdas; 797 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 798 INIT_READ_NODE(rrdNodes + i, "rrd"); 799 DISK_NODE_PARAMS(rrdNodes[i], pda); 800 } 801 802 /* read redundancy pdas */ 803 pda = pqPDAs; 804 INIT_READ_NODE(rpNodes, "Rp"); 805 RF_ASSERT(pda); 806 DISK_NODE_PARAMS(rpNodes[0], pda); 807 pda++; 808 INIT_READ_NODE(rqNodes, redundantReadNodeName); 809 RF_ASSERT(pda); 810 DISK_NODE_PARAMS(rqNodes[0], pda); 811 if (nPQNodes == 2) { 812 pda++; 813 INIT_READ_NODE(rpNodes + 1, "Rp"); 814 RF_ASSERT(pda); 815 DISK_NODE_PARAMS(rpNodes[1], pda); 816 pda++; 817 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 818 RF_ASSERT(pda); 819 DISK_NODE_PARAMS(rqNodes[1], pda); 820 } 821 /* the recovery node has all reads as precedessors and all writes as 822 * successors. It generates a result for every write P or write Q 823 * node. As parameters, it takes a pda per read and a pda per stripe 824 * of user data written. It also takes as the last params the raidPtr 825 * and asm. For results, it takes PDA for P & Q. */ 826 827 828 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 829 nWriteNodes, /* succesors */ 830 nReadNodes, /* preds */ 831 nReadNodes + nWudNodes + 3, /* params */ 832 2 * nPQNodes, /* results */ 833 dag_h, recoveryNodeName, allocList); 834 835 836 837 for (i = 0; i < nReadNodes; i++) { 838 recoveryNode->antecedents[i] = rrdNodes + i; 839 recoveryNode->antType[i] = rf_control; 840 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 841 } 842 for (i = 0; i < nWudNodes; i++) { 843 recoveryNode->succedents[i] = wudNodes + i; 844 } 845 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 846 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 847 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 848 849 for (; i < nWriteNodes; i++) 850 recoveryNode->succedents[i] = wudNodes + i; 851 852 pda = pqPDAs; 853 recoveryNode->results[0] = pda; 854 pda++; 855 recoveryNode->results[1] = pda; 856 if (nPQNodes == 2) { 857 pda++; 858 recoveryNode->results[2] = pda; 859 pda++; 860 recoveryNode->results[3] = pda; 861 } 862 /* fill writes */ 863 #define INIT_WRITE_NODE(node,name) \ 864 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 865 (node)->succedents[0] = unblockNode; \ 866 (node)->antecedents[0] = recoveryNode; \ 867 (node)->antType[0] = rf_control; 868 869 pda = asmap->physInfo; 870 for (i = 0; i < nWudNodes; i++) { 871 INIT_WRITE_NODE(wudNodes + i, "Wd"); 872 DISK_NODE_PARAMS(wudNodes[i], pda); 873 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 874 pda = pda->next; 875 } 876 /* write redundancy pdas */ 877 pda = pqPDAs; 878 INIT_WRITE_NODE(wpNodes, "Wp"); 879 RF_ASSERT(pda); 880 DISK_NODE_PARAMS(wpNodes[0], pda); 881 pda++; 882 INIT_WRITE_NODE(wqNodes, "Wq"); 883 RF_ASSERT(pda); 884 DISK_NODE_PARAMS(wqNodes[0], pda); 885 if (nPQNodes == 2) { 886 pda++; 887 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 888 RF_ASSERT(pda); 889 DISK_NODE_PARAMS(wpNodes[1], pda); 890 pda++; 891 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 892 RF_ASSERT(pda); 893 DISK_NODE_PARAMS(wqNodes[1], pda); 894 } 895 } 896 #endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 897