1 /* $NetBSD: rf_dagdegwr.c,v 1.33 2014/03/23 03:42:39 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 1.33 2014/03/23 03:42:39 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 nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 165 rdnodesFaked; 166 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *termNode; 167 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 168 RF_DagNode_t *wnqNode; 169 #endif 170 RF_DagNode_t *wndNodes, *rrdNodes, *xorNode, *commitNode; 171 RF_DagNode_t *tmpNode, *tmpwndNode, *tmprrdNode; 172 RF_SectorCount_t sectorsPerSU; 173 RF_ReconUnitNum_t which_ru; 174 char *xorTargetBuf = NULL; /* the target buffer for the XOR 175 * operation */ 176 char overlappingPDAs[RF_MAXCOL];/* a temporary array of flags */ 177 RF_AccessStripeMapHeader_t *new_asm_h[2]; 178 RF_PhysDiskAddr_t *pda, *parityPDA; 179 RF_StripeNum_t parityStripeID; 180 RF_PhysDiskAddr_t *failedPDA; 181 RF_RaidLayout_t *layoutPtr; 182 183 layoutPtr = &(raidPtr->Layout); 184 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 185 &which_ru); 186 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 187 /* failedPDA points to the pda within the asm that targets the failed 188 * disk */ 189 failedPDA = asmap->failedPDAs[0]; 190 191 #if RF_DEBUG_DAG 192 if (rf_dagDebug) 193 printf("[Creating degraded-write DAG]\n"); 194 #endif 195 196 RF_ASSERT(asmap->numDataFailed == 1); 197 dag_h->creator = "SimpleDegradedWriteDAG"; 198 199 /* 200 * Generate two ASMs identifying the surviving data 201 * we need in order to recover the lost data. 202 */ 203 /* overlappingPDAs array must be zero'd */ 204 memset(overlappingPDAs, 0, RF_MAXCOL); 205 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 206 &nXorBufs, NULL, overlappingPDAs, allocList); 207 208 /* create all the nodes at once */ 209 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 210 * generated for the 211 * failed pda */ 212 213 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 214 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 215 /* 216 * XXX 217 * 218 * There's a bug with a complete stripe overwrite- that means 0 reads 219 * of old data, and the rest of the DAG generation code doesn't like 220 * that. A release is coming, and I don't wanna risk breaking a critical 221 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 222 * I'm gonna fake there being a read node, and I'm gonna swap in a 223 * no-op node in its place (to make all the link-up code happy). 224 * This should be fixed at some point. --jimz 225 */ 226 if (nRrdNodes == 0) { 227 nRrdNodes = 1; 228 rdnodesFaked = 1; 229 } else { 230 rdnodesFaked = 0; 231 } 232 233 blockNode = rf_AllocDAGNode(); 234 blockNode->list_next = dag_h->nodes; 235 dag_h->nodes = blockNode; 236 237 commitNode = rf_AllocDAGNode(); 238 commitNode->list_next = dag_h->nodes; 239 dag_h->nodes = commitNode; 240 241 unblockNode = rf_AllocDAGNode(); 242 unblockNode->list_next = dag_h->nodes; 243 dag_h->nodes = unblockNode; 244 245 termNode = rf_AllocDAGNode(); 246 termNode->list_next = dag_h->nodes; 247 dag_h->nodes = termNode; 248 249 xorNode = rf_AllocDAGNode(); 250 xorNode->list_next = dag_h->nodes; 251 dag_h->nodes = xorNode; 252 253 wnpNode = rf_AllocDAGNode(); 254 wnpNode->list_next = dag_h->nodes; 255 dag_h->nodes = wnpNode; 256 257 for (i = 0; i < nWndNodes; i++) { 258 tmpNode = rf_AllocDAGNode(); 259 tmpNode->list_next = dag_h->nodes; 260 dag_h->nodes = tmpNode; 261 } 262 wndNodes = dag_h->nodes; 263 264 for (i = 0; i < nRrdNodes; i++) { 265 tmpNode = rf_AllocDAGNode(); 266 tmpNode->list_next = dag_h->nodes; 267 dag_h->nodes = tmpNode; 268 } 269 rrdNodes = dag_h->nodes; 270 271 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 272 if (nfaults == 2) { 273 wnqNode = rf_AllocDAGNode(); 274 wnqNode->list_next = dag_h->nodes; 275 dag_h->nodes = wnqNode; 276 } else { 277 wnqNode = NULL; 278 } 279 #endif 280 281 /* this dag can not commit until all rrd and xor Nodes have completed */ 282 dag_h->numCommitNodes = 1; 283 dag_h->numCommits = 0; 284 dag_h->numSuccedents = 1; 285 286 RF_ASSERT(nRrdNodes > 0); 287 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 288 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 289 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 290 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 291 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 292 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 293 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 294 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 295 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 296 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 297 298 /* 299 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 300 * the failed buffer, save a pointer to it so we can use it as the target 301 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 302 * a buffer is the same size as the failed buffer, it must also be at the 303 * same alignment within the SU. 304 */ 305 i = 0; 306 tmprrdNode = rrdNodes; 307 if (new_asm_h[0]) { 308 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 309 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 310 i++, pda = pda->next) { 311 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 312 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 313 RF_ASSERT(pda); 314 tmprrdNode->params[0].p = pda; 315 tmprrdNode->params[1].p = pda->bufPtr; 316 tmprrdNode->params[2].v = parityStripeID; 317 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 318 tmprrdNode = tmprrdNode->list_next; 319 } 320 } 321 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 322 /* Note that for tmprrdNode, this means a continuation from above, so no need to 323 assign it anything.. */ 324 if (new_asm_h[1]) { 325 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 326 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 327 j++, pda = pda->next) { 328 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 329 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 330 RF_ASSERT(pda); 331 tmprrdNode->params[0].p = pda; 332 tmprrdNode->params[1].p = pda->bufPtr; 333 tmprrdNode->params[2].v = parityStripeID; 334 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 335 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 336 xorTargetBuf = pda->bufPtr; 337 tmprrdNode = tmprrdNode->list_next; 338 } 339 } 340 if (rdnodesFaked) { 341 /* 342 * This is where we'll init that fake noop read node 343 * (XXX should the wakeup func be different?) 344 */ 345 /* node that rrdNodes will just be a single node... */ 346 rf_InitNode(rrdNodes, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 347 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 348 } 349 /* 350 * Make a PDA for the parity unit. The parity PDA should start at 351 * the same offset into the SU as the failed PDA. 352 */ 353 /* Danner comment: I don't think this copy is really necessary. We are 354 * in one of two cases here. (1) The entire failed unit is written. 355 * Then asmap->parityInfo will describe the entire parity. (2) We are 356 * only writing a subset of the failed unit and nothing else. Then the 357 * asmap->parityInfo describes the failed unit and the copy can also 358 * be avoided. */ 359 360 parityPDA = rf_AllocPhysDiskAddr(); 361 parityPDA->next = dag_h->pda_cleanup_list; 362 dag_h->pda_cleanup_list = parityPDA; 363 parityPDA->col = asmap->parityInfo->col; 364 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 365 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 366 parityPDA->numSector = failedPDA->numSector; 367 368 if (!xorTargetBuf) { 369 xorTargetBuf = rf_AllocBuffer(raidPtr, dag_h, rf_RaidAddressToByte(raidPtr, failedPDA->numSector)); 370 } 371 /* init the Wnp node */ 372 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 373 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 374 wnpNode->params[0].p = parityPDA; 375 wnpNode->params[1].p = xorTargetBuf; 376 wnpNode->params[2].v = parityStripeID; 377 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 378 379 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 380 /* fill in the Wnq Node */ 381 if (nfaults == 2) { 382 { 383 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 384 (RF_PhysDiskAddr_t *), allocList); 385 parityPDA->col = asmap->qInfo->col; 386 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 387 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 388 parityPDA->numSector = failedPDA->numSector; 389 390 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 391 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 392 wnqNode->params[0].p = parityPDA; 393 RF_MallocAndAdd(xorNode->results[1], 394 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 395 wnqNode->params[1].p = xorNode->results[1]; 396 wnqNode->params[2].v = parityStripeID; 397 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 398 } 399 } 400 #endif 401 /* fill in the Wnd nodes */ 402 tmpwndNode = wndNodes; 403 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 404 if (pda == failedPDA) { 405 i--; 406 continue; 407 } 408 rf_InitNode(tmpwndNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 409 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 410 RF_ASSERT(pda); 411 tmpwndNode->params[0].p = pda; 412 tmpwndNode->params[1].p = pda->bufPtr; 413 tmpwndNode->params[2].v = parityStripeID; 414 tmpwndNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 415 tmpwndNode = tmpwndNode->list_next; 416 } 417 418 /* fill in the results of the xor node */ 419 xorNode->results[0] = xorTargetBuf; 420 421 /* fill in the params of the xor node */ 422 423 paramNum = 0; 424 if (rdnodesFaked == 0) { 425 tmprrdNode = rrdNodes; 426 for (i = 0; i < nRrdNodes; i++) { 427 /* all the Rrd nodes need to be xored together */ 428 xorNode->params[paramNum++] = tmprrdNode->params[0]; 429 xorNode->params[paramNum++] = tmprrdNode->params[1]; 430 tmprrdNode = tmprrdNode->list_next; 431 } 432 } 433 tmpwndNode = wndNodes; 434 for (i = 0; i < nWndNodes; i++) { 435 /* any Wnd nodes that overlap the failed access need to be 436 * xored in */ 437 if (overlappingPDAs[i]) { 438 pda = rf_AllocPhysDiskAddr(); 439 memcpy((char *) pda, (char *) tmpwndNode->params[0].p, sizeof(RF_PhysDiskAddr_t)); 440 /* add it into the pda_cleanup_list *after* the copy, TYVM */ 441 pda->next = dag_h->pda_cleanup_list; 442 dag_h->pda_cleanup_list = pda; 443 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 444 xorNode->params[paramNum++].p = pda; 445 xorNode->params[paramNum++].p = pda->bufPtr; 446 } 447 tmpwndNode = tmpwndNode->list_next; 448 } 449 450 /* 451 * Install the failed PDA into the xor param list so that the 452 * new data gets xor'd in. 453 */ 454 xorNode->params[paramNum++].p = failedPDA; 455 xorNode->params[paramNum++].p = failedPDA->bufPtr; 456 457 /* 458 * The last 2 params to the recovery xor node are always the failed 459 * PDA and the raidPtr. install the failedPDA even though we have just 460 * done so above. This allows us to use the same XOR function for both 461 * degraded reads and degraded writes. 462 */ 463 xorNode->params[paramNum++].p = failedPDA; 464 xorNode->params[paramNum++].p = raidPtr; 465 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 466 467 /* 468 * Code to link nodes begins here 469 */ 470 471 /* link header to block node */ 472 RF_ASSERT(blockNode->numAntecedents == 0); 473 dag_h->succedents[0] = blockNode; 474 475 /* link block node to rd nodes */ 476 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 477 tmprrdNode = rrdNodes; 478 for (i = 0; i < nRrdNodes; i++) { 479 RF_ASSERT(tmprrdNode->numAntecedents == 1); 480 blockNode->succedents[i] = tmprrdNode; 481 tmprrdNode->antecedents[0] = blockNode; 482 tmprrdNode->antType[0] = rf_control; 483 tmprrdNode = tmprrdNode->list_next; 484 } 485 486 /* link read nodes to xor node */ 487 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 488 tmprrdNode = rrdNodes; 489 for (i = 0; i < nRrdNodes; i++) { 490 RF_ASSERT(tmprrdNode->numSuccedents == 1); 491 tmprrdNode->succedents[0] = xorNode; 492 xorNode->antecedents[i] = tmprrdNode; 493 xorNode->antType[i] = rf_trueData; 494 tmprrdNode = tmprrdNode->list_next; 495 } 496 497 /* link xor node to commit node */ 498 RF_ASSERT(xorNode->numSuccedents == 1); 499 RF_ASSERT(commitNode->numAntecedents == 1); 500 xorNode->succedents[0] = commitNode; 501 commitNode->antecedents[0] = xorNode; 502 commitNode->antType[0] = rf_control; 503 504 /* link commit node to wnd nodes */ 505 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 506 tmpwndNode = wndNodes; 507 for (i = 0; i < nWndNodes; i++) { 508 RF_ASSERT(tmpwndNode->numAntecedents == 1); 509 commitNode->succedents[i] = tmpwndNode; 510 tmpwndNode->antecedents[0] = commitNode; 511 tmpwndNode->antType[0] = rf_control; 512 tmpwndNode = tmpwndNode->list_next; 513 } 514 515 /* link the commit node to wnp, wnq nodes */ 516 RF_ASSERT(wnpNode->numAntecedents == 1); 517 commitNode->succedents[nWndNodes] = wnpNode; 518 wnpNode->antecedents[0] = commitNode; 519 wnpNode->antType[0] = rf_control; 520 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 521 if (nfaults == 2) { 522 RF_ASSERT(wnqNode->numAntecedents == 1); 523 commitNode->succedents[nWndNodes + 1] = wnqNode; 524 wnqNode->antecedents[0] = commitNode; 525 wnqNode->antType[0] = rf_control; 526 } 527 #endif 528 /* link write new data nodes to unblock node */ 529 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 530 tmpwndNode = wndNodes; 531 for (i = 0; i < nWndNodes; i++) { 532 RF_ASSERT(tmpwndNode->numSuccedents == 1); 533 tmpwndNode->succedents[0] = unblockNode; 534 unblockNode->antecedents[i] = tmpwndNode; 535 unblockNode->antType[i] = rf_control; 536 tmpwndNode = tmpwndNode->list_next; 537 } 538 539 /* link write new parity node to unblock node */ 540 RF_ASSERT(wnpNode->numSuccedents == 1); 541 wnpNode->succedents[0] = unblockNode; 542 unblockNode->antecedents[nWndNodes] = wnpNode; 543 unblockNode->antType[nWndNodes] = rf_control; 544 545 #if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 546 /* link write new q node to unblock node */ 547 if (nfaults == 2) { 548 RF_ASSERT(wnqNode->numSuccedents == 1); 549 wnqNode->succedents[0] = unblockNode; 550 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 551 unblockNode->antType[nWndNodes + 1] = rf_control; 552 } 553 #endif 554 /* link unblock node to term node */ 555 RF_ASSERT(unblockNode->numSuccedents == 1); 556 RF_ASSERT(termNode->numAntecedents == 1); 557 RF_ASSERT(termNode->numSuccedents == 0); 558 unblockNode->succedents[0] = termNode; 559 termNode->antecedents[0] = unblockNode; 560 termNode->antType[0] = rf_control; 561 } 562 #define CONS_PDA(if,start,num) \ 563 pda_p->col = asmap->if->col; \ 564 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 565 pda_p->numSector = num; \ 566 pda_p->next = NULL; \ 567 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 568 #if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 569 void 570 rf_WriteGenerateFailedAccessASMs( 571 RF_Raid_t * raidPtr, 572 RF_AccessStripeMap_t * asmap, 573 RF_PhysDiskAddr_t ** pdap, 574 int *nNodep, 575 RF_PhysDiskAddr_t ** pqpdap, 576 int *nPQNodep, 577 RF_AllocListElem_t * allocList) 578 { 579 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 580 int PDAPerDisk, i; 581 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 582 int numDataCol = layoutPtr->numDataCol; 583 int state; 584 unsigned napdas; 585 RF_SectorNum_t fone_start, ftwo_start = 0; 586 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 587 RF_PhysDiskAddr_t *pda_p; 588 RF_RaidAddr_t sosAddr; 589 590 /* determine how many pda's we will have to generate per unaccess 591 * stripe. If there is only one failed data unit, it is one; if two, 592 * possibly two, depending whether they overlap. */ 593 594 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 595 596 if (asmap->numDataFailed == 1) { 597 PDAPerDisk = 1; 598 state = 1; 599 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 600 pda_p = *pqpdap; 601 /* build p */ 602 CONS_PDA(parityInfo, fone_start, fone->numSector); 603 pda_p->type = RF_PDA_TYPE_PARITY; 604 pda_p++; 605 /* build q */ 606 CONS_PDA(qInfo, fone_start, fone->numSector); 607 pda_p->type = RF_PDA_TYPE_Q; 608 } else { 609 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 610 if (fone->numSector + ftwo->numSector > secPerSU) { 611 PDAPerDisk = 1; 612 state = 2; 613 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 614 pda_p = *pqpdap; 615 CONS_PDA(parityInfo, 0, secPerSU); 616 pda_p->type = RF_PDA_TYPE_PARITY; 617 pda_p++; 618 CONS_PDA(qInfo, 0, secPerSU); 619 pda_p->type = RF_PDA_TYPE_Q; 620 } else { 621 PDAPerDisk = 2; 622 state = 3; 623 /* four of them, fone, then ftwo */ 624 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 625 pda_p = *pqpdap; 626 CONS_PDA(parityInfo, fone_start, fone->numSector); 627 pda_p->type = RF_PDA_TYPE_PARITY; 628 pda_p++; 629 CONS_PDA(qInfo, fone_start, fone->numSector); 630 pda_p->type = RF_PDA_TYPE_Q; 631 pda_p++; 632 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 633 pda_p->type = RF_PDA_TYPE_PARITY; 634 pda_p++; 635 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 636 pda_p->type = RF_PDA_TYPE_Q; 637 } 638 } 639 /* figure out number of nonaccessed pda */ 640 napdas = PDAPerDisk * (numDataCol - 2); 641 *nPQNodep = PDAPerDisk; 642 643 *nNodep = napdas; 644 if (napdas == 0) 645 return; /* short circuit */ 646 647 /* allocate up our list of pda's */ 648 649 RF_MallocAndAdd(pda_p, napdas * sizeof(RF_PhysDiskAddr_t), 650 (RF_PhysDiskAddr_t *), allocList); 651 *pdap = pda_p; 652 653 /* linkem together */ 654 for (i = 0; i < (napdas - 1); i++) 655 pda_p[i].next = pda_p + (i + 1); 656 657 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 658 for (i = 0; i < numDataCol; i++) { 659 if ((pda_p - (*pdap)) == napdas) 660 continue; 661 pda_p->type = RF_PDA_TYPE_DATA; 662 pda_p->raidAddress = sosAddr + (i * secPerSU); 663 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 664 /* skip over dead disks */ 665 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->col].status)) 666 continue; 667 switch (state) { 668 case 1: /* fone */ 669 pda_p->numSector = fone->numSector; 670 pda_p->raidAddress += fone_start; 671 pda_p->startSector += fone_start; 672 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 673 break; 674 case 2: /* full stripe */ 675 pda_p->numSector = secPerSU; 676 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 677 break; 678 case 3: /* two slabs */ 679 pda_p->numSector = fone->numSector; 680 pda_p->raidAddress += fone_start; 681 pda_p->startSector += fone_start; 682 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 683 pda_p++; 684 pda_p->type = RF_PDA_TYPE_DATA; 685 pda_p->raidAddress = sosAddr + (i * secPerSU); 686 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 687 pda_p->numSector = ftwo->numSector; 688 pda_p->raidAddress += ftwo_start; 689 pda_p->startSector += ftwo_start; 690 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 691 break; 692 default: 693 RF_PANIC(); 694 } 695 pda_p++; 696 } 697 698 RF_ASSERT(pda_p - *pdap == napdas); 699 return; 700 } 701 #define DISK_NODE_PDA(node) ((node)->params[0].p) 702 703 #define DISK_NODE_PARAMS(_node_,_p_) \ 704 (_node_).params[0].p = _p_ ; \ 705 (_node_).params[1].p = (_p_)->bufPtr; \ 706 (_node_).params[2].v = parityStripeID; \ 707 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru) 708 709 void 710 rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 711 RF_DagHeader_t *dag_h, void *bp, 712 RF_RaidAccessFlags_t flags, 713 RF_AllocListElem_t *allocList, 714 const char *redundantReadNodeName, 715 const char *redundantWriteNodeName, 716 const char *recoveryNodeName, 717 int (*recovFunc) (RF_DagNode_t *)) 718 { 719 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 720 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 721 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 722 RF_PhysDiskAddr_t *pda, *pqPDAs; 723 RF_PhysDiskAddr_t *npdas; 724 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 725 RF_ReconUnitNum_t which_ru; 726 int nPQNodes; 727 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 728 729 /* simple small write case - First part looks like a reconstruct-read 730 * of the failed data units. Then a write of all data units not 731 * failed. */ 732 733 734 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 735 * / -------PQ----- / \ \ Wud Wp WQ \ | / 736 * --Unblock- | T 737 * 738 * Rrd = read recovery data (potentially none) Wud = write user data 739 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 740 * (could be two) 741 * 742 */ 743 744 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 745 746 RF_ASSERT(asmap->numDataFailed == 1); 747 748 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 749 nReadNodes = nRrdNodes + 2 * nPQNodes; 750 nWriteNodes = nWudNodes + 2 * nPQNodes; 751 nNodes = 4 + nReadNodes + nWriteNodes; 752 753 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 754 blockNode = nodes; 755 unblockNode = blockNode + 1; 756 termNode = unblockNode + 1; 757 recoveryNode = termNode + 1; 758 rrdNodes = recoveryNode + 1; 759 rpNodes = rrdNodes + nRrdNodes; 760 rqNodes = rpNodes + nPQNodes; 761 wudNodes = rqNodes + nPQNodes; 762 wpNodes = wudNodes + nWudNodes; 763 wqNodes = wpNodes + nPQNodes; 764 765 dag_h->creator = "PQ_DDSimpleSmallWrite"; 766 dag_h->numSuccedents = 1; 767 dag_h->succedents[0] = blockNode; 768 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 769 termNode->antecedents[0] = unblockNode; 770 termNode->antType[0] = rf_control; 771 772 /* init the block and unblock nodes */ 773 /* The block node has all the read nodes as successors */ 774 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 775 for (i = 0; i < nReadNodes; i++) 776 blockNode->succedents[i] = rrdNodes + i; 777 778 /* The unblock node has all the writes as successors */ 779 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 780 for (i = 0; i < nWriteNodes; i++) { 781 unblockNode->antecedents[i] = wudNodes + i; 782 unblockNode->antType[i] = rf_control; 783 } 784 unblockNode->succedents[0] = termNode; 785 786 #define INIT_READ_NODE(node,name) \ 787 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 788 (node)->succedents[0] = recoveryNode; \ 789 (node)->antecedents[0] = blockNode; \ 790 (node)->antType[0] = rf_control; 791 792 /* build the read nodes */ 793 pda = npdas; 794 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 795 INIT_READ_NODE(rrdNodes + i, "rrd"); 796 DISK_NODE_PARAMS(rrdNodes[i], pda); 797 } 798 799 /* read redundancy pdas */ 800 pda = pqPDAs; 801 INIT_READ_NODE(rpNodes, "Rp"); 802 RF_ASSERT(pda); 803 DISK_NODE_PARAMS(rpNodes[0], pda); 804 pda++; 805 INIT_READ_NODE(rqNodes, redundantReadNodeName); 806 RF_ASSERT(pda); 807 DISK_NODE_PARAMS(rqNodes[0], pda); 808 if (nPQNodes == 2) { 809 pda++; 810 INIT_READ_NODE(rpNodes + 1, "Rp"); 811 RF_ASSERT(pda); 812 DISK_NODE_PARAMS(rpNodes[1], pda); 813 pda++; 814 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 815 RF_ASSERT(pda); 816 DISK_NODE_PARAMS(rqNodes[1], pda); 817 } 818 /* the recovery node has all reads as precedessors and all writes as 819 * successors. It generates a result for every write P or write Q 820 * node. As parameters, it takes a pda per read and a pda per stripe 821 * of user data written. It also takes as the last params the raidPtr 822 * and asm. For results, it takes PDA for P & Q. */ 823 824 825 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 826 nWriteNodes, /* succesors */ 827 nReadNodes, /* preds */ 828 nReadNodes + nWudNodes + 3, /* params */ 829 2 * nPQNodes, /* results */ 830 dag_h, recoveryNodeName, allocList); 831 832 833 834 for (i = 0; i < nReadNodes; i++) { 835 recoveryNode->antecedents[i] = rrdNodes + i; 836 recoveryNode->antType[i] = rf_control; 837 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 838 } 839 for (i = 0; i < nWudNodes; i++) { 840 recoveryNode->succedents[i] = wudNodes + i; 841 } 842 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 843 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 844 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 845 846 for (; i < nWriteNodes; i++) 847 recoveryNode->succedents[i] = wudNodes + i; 848 849 pda = pqPDAs; 850 recoveryNode->results[0] = pda; 851 pda++; 852 recoveryNode->results[1] = pda; 853 if (nPQNodes == 2) { 854 pda++; 855 recoveryNode->results[2] = pda; 856 pda++; 857 recoveryNode->results[3] = pda; 858 } 859 /* fill writes */ 860 #define INIT_WRITE_NODE(node,name) \ 861 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 862 (node)->succedents[0] = unblockNode; \ 863 (node)->antecedents[0] = recoveryNode; \ 864 (node)->antType[0] = rf_control; 865 866 pda = asmap->physInfo; 867 for (i = 0; i < nWudNodes; i++) { 868 INIT_WRITE_NODE(wudNodes + i, "Wd"); 869 DISK_NODE_PARAMS(wudNodes[i], pda); 870 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 871 pda = pda->next; 872 } 873 /* write redundancy pdas */ 874 pda = pqPDAs; 875 INIT_WRITE_NODE(wpNodes, "Wp"); 876 RF_ASSERT(pda); 877 DISK_NODE_PARAMS(wpNodes[0], pda); 878 pda++; 879 INIT_WRITE_NODE(wqNodes, "Wq"); 880 RF_ASSERT(pda); 881 DISK_NODE_PARAMS(wqNodes[0], pda); 882 if (nPQNodes == 2) { 883 pda++; 884 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 885 RF_ASSERT(pda); 886 DISK_NODE_PARAMS(wpNodes[1], pda); 887 pda++; 888 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 889 RF_ASSERT(pda); 890 DISK_NODE_PARAMS(wqNodes[1], pda); 891 } 892 } 893 #endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 894