rf_dagdegwr.c revision 1.22
1/* $NetBSD: rf_dagdegwr.c,v 1.22 2004/03/21 03:22:08 oster 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 1.22 2004/03/21 03:22:08 oster 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 78static 79RF_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 85void 86rf_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 154void 155rf_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 RF_ASSERT(i == nNodes); 282 283 /* this dag can not commit until all rrd and xor Nodes have completed */ 284 dag_h->numCommitNodes = 1; 285 dag_h->numCommits = 0; 286 dag_h->numSuccedents = 1; 287 288 RF_ASSERT(nRrdNodes > 0); 289 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 290 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 291 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 292 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 293 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 294 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 295 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 296 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 297 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 298 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 299 300 /* 301 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 302 * the failed buffer, save a pointer to it so we can use it as the target 303 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 304 * a buffer is the same size as the failed buffer, it must also be at the 305 * same alignment within the SU. 306 */ 307 i = 0; 308 tmprrdNode = rrdNodes; 309 if (new_asm_h[0]) { 310 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 311 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 312 i++, pda = pda->next) { 313 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 314 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 315 RF_ASSERT(pda); 316 tmprrdNode->params[0].p = pda; 317 tmprrdNode->params[1].p = pda->bufPtr; 318 tmprrdNode->params[2].v = parityStripeID; 319 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 320 tmprrdNode = tmprrdNode->list_next; 321 } 322 } 323 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 324 /* Note that for tmprrdNode, this means a continuation from above, so no need to 325 assign it anything.. */ 326 if (new_asm_h[1]) { 327 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 328 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 329 j++, pda = pda->next) { 330 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 331 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 332 RF_ASSERT(pda); 333 tmprrdNode->params[0].p = pda; 334 tmprrdNode->params[1].p = pda->bufPtr; 335 tmprrdNode->params[2].v = parityStripeID; 336 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 337 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 338 xorTargetBuf = pda->bufPtr; 339 tmprrdNode = tmprrdNode->list_next; 340 } 341 } 342 if (rdnodesFaked) { 343 /* 344 * This is where we'll init that fake noop read node 345 * (XXX should the wakeup func be different?) 346 */ 347 /* node that rrdNodes will just be a single node... */ 348 rf_InitNode(rrdNodes, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 349 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 350 } 351 /* 352 * Make a PDA for the parity unit. The parity PDA should start at 353 * the same offset into the SU as the failed PDA. 354 */ 355 /* Danner comment: I don't think this copy is really necessary. We are 356 * in one of two cases here. (1) The entire failed unit is written. 357 * Then asmap->parityInfo will describe the entire parity. (2) We are 358 * only writing a subset of the failed unit and nothing else. Then the 359 * asmap->parityInfo describes the failed unit and the copy can also 360 * be avoided. */ 361 362 parityPDA = rf_AllocPhysDiskAddr(); 363 parityPDA->next = dag_h->pda_cleanup_list; 364 dag_h->pda_cleanup_list = parityPDA; 365 parityPDA->col = asmap->parityInfo->col; 366 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 367 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 368 parityPDA->numSector = failedPDA->numSector; 369 370 if (!xorTargetBuf) { 371 xorTargetBuf = rf_AllocBuffer(raidPtr, dag_h, 372 rf_RaidAddressToByte(raidPtr, failedPDA->numSector)); 373 } 374 /* init the Wnp node */ 375 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 376 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 377 wnpNode->params[0].p = parityPDA; 378 wnpNode->params[1].p = xorTargetBuf; 379 wnpNode->params[2].v = parityStripeID; 380 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 381 382#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 383 /* fill in the Wnq Node */ 384 if (nfaults == 2) { 385 { 386 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 387 (RF_PhysDiskAddr_t *), allocList); 388 parityPDA->col = asmap->qInfo->col; 389 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 390 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 391 parityPDA->numSector = failedPDA->numSector; 392 393 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 394 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 395 wnqNode->params[0].p = parityPDA; 396 RF_MallocAndAdd(xorNode->results[1], 397 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 398 wnqNode->params[1].p = xorNode->results[1]; 399 wnqNode->params[2].v = parityStripeID; 400 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 401 } 402 } 403#endif 404 /* fill in the Wnd nodes */ 405 tmpwndNode = wndNodes; 406 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 407 if (pda == failedPDA) { 408 i--; 409 continue; 410 } 411 rf_InitNode(tmpwndNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 412 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 413 RF_ASSERT(pda); 414 tmpwndNode->params[0].p = pda; 415 tmpwndNode->params[1].p = pda->bufPtr; 416 tmpwndNode->params[2].v = parityStripeID; 417 tmpwndNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 418 tmpwndNode = tmpwndNode->list_next; 419 } 420 421 /* fill in the results of the xor node */ 422 xorNode->results[0] = xorTargetBuf; 423 424 /* fill in the params of the xor node */ 425 426 paramNum = 0; 427 if (rdnodesFaked == 0) { 428 tmprrdNode = rrdNodes; 429 for (i = 0; i < nRrdNodes; i++) { 430 /* all the Rrd nodes need to be xored together */ 431 xorNode->params[paramNum++] = tmprrdNode->params[0]; 432 xorNode->params[paramNum++] = tmprrdNode->params[1]; 433 tmprrdNode = tmprrdNode->list_next; 434 } 435 } 436 tmpwndNode = wndNodes; 437 for (i = 0; i < nWndNodes; i++) { 438 /* any Wnd nodes that overlap the failed access need to be 439 * xored in */ 440 if (overlappingPDAs[i]) { 441 pda = rf_AllocPhysDiskAddr(); 442 memcpy((char *) pda, (char *) tmpwndNode->params[0].p, sizeof(RF_PhysDiskAddr_t)); 443 /* add it into the pda_cleanup_list *after* the copy, TYVM */ 444 pda->next = dag_h->pda_cleanup_list; 445 dag_h->pda_cleanup_list = pda; 446 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 447 xorNode->params[paramNum++].p = pda; 448 xorNode->params[paramNum++].p = pda->bufPtr; 449 } 450 tmpwndNode = tmpwndNode->list_next; 451 } 452 453 /* 454 * Install the failed PDA into the xor param list so that the 455 * new data gets xor'd in. 456 */ 457 xorNode->params[paramNum++].p = failedPDA; 458 xorNode->params[paramNum++].p = failedPDA->bufPtr; 459 460 /* 461 * The last 2 params to the recovery xor node are always the failed 462 * PDA and the raidPtr. install the failedPDA even though we have just 463 * done so above. This allows us to use the same XOR function for both 464 * degraded reads and degraded writes. 465 */ 466 xorNode->params[paramNum++].p = failedPDA; 467 xorNode->params[paramNum++].p = raidPtr; 468 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 469 470 /* 471 * Code to link nodes begins here 472 */ 473 474 /* link header to block node */ 475 RF_ASSERT(blockNode->numAntecedents == 0); 476 dag_h->succedents[0] = blockNode; 477 478 /* link block node to rd nodes */ 479 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 480 tmprrdNode = rrdNodes; 481 for (i = 0; i < nRrdNodes; i++) { 482 RF_ASSERT(tmprrdNode->numAntecedents == 1); 483 blockNode->succedents[i] = tmprrdNode; 484 tmprrdNode->antecedents[0] = blockNode; 485 tmprrdNode->antType[0] = rf_control; 486 tmprrdNode = tmprrdNode->list_next; 487 } 488 489 /* link read nodes to xor node */ 490 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 491 tmprrdNode = rrdNodes; 492 for (i = 0; i < nRrdNodes; i++) { 493 RF_ASSERT(tmprrdNode->numSuccedents == 1); 494 tmprrdNode->succedents[0] = xorNode; 495 xorNode->antecedents[i] = tmprrdNode; 496 xorNode->antType[i] = rf_trueData; 497 tmprrdNode = tmprrdNode->list_next; 498 } 499 500 /* link xor node to commit node */ 501 RF_ASSERT(xorNode->numSuccedents == 1); 502 RF_ASSERT(commitNode->numAntecedents == 1); 503 xorNode->succedents[0] = commitNode; 504 commitNode->antecedents[0] = xorNode; 505 commitNode->antType[0] = rf_control; 506 507 /* link commit node to wnd nodes */ 508 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 509 tmpwndNode = wndNodes; 510 for (i = 0; i < nWndNodes; i++) { 511 RF_ASSERT(tmpwndNode->numAntecedents == 1); 512 commitNode->succedents[i] = tmpwndNode; 513 tmpwndNode->antecedents[0] = commitNode; 514 tmpwndNode->antType[0] = rf_control; 515 tmpwndNode = tmpwndNode->list_next; 516 } 517 518 /* link the commit node to wnp, wnq nodes */ 519 RF_ASSERT(wnpNode->numAntecedents == 1); 520 commitNode->succedents[nWndNodes] = wnpNode; 521 wnpNode->antecedents[0] = commitNode; 522 wnpNode->antType[0] = rf_control; 523#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 524 if (nfaults == 2) { 525 RF_ASSERT(wnqNode->numAntecedents == 1); 526 commitNode->succedents[nWndNodes + 1] = wnqNode; 527 wnqNode->antecedents[0] = commitNode; 528 wnqNode->antType[0] = rf_control; 529 } 530#endif 531 /* link write new data nodes to unblock node */ 532 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 533 tmpwndNode = wndNodes; 534 for (i = 0; i < nWndNodes; i++) { 535 RF_ASSERT(tmpwndNode->numSuccedents == 1); 536 tmpwndNode->succedents[0] = unblockNode; 537 unblockNode->antecedents[i] = tmpwndNode; 538 unblockNode->antType[i] = rf_control; 539 tmpwndNode = tmpwndNode->list_next; 540 } 541 542 /* link write new parity node to unblock node */ 543 RF_ASSERT(wnpNode->numSuccedents == 1); 544 wnpNode->succedents[0] = unblockNode; 545 unblockNode->antecedents[nWndNodes] = wnpNode; 546 unblockNode->antType[nWndNodes] = rf_control; 547 548#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 549 /* link write new q node to unblock node */ 550 if (nfaults == 2) { 551 RF_ASSERT(wnqNode->numSuccedents == 1); 552 wnqNode->succedents[0] = unblockNode; 553 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 554 unblockNode->antType[nWndNodes + 1] = rf_control; 555 } 556#endif 557 /* link unblock node to term node */ 558 RF_ASSERT(unblockNode->numSuccedents == 1); 559 RF_ASSERT(termNode->numAntecedents == 1); 560 RF_ASSERT(termNode->numSuccedents == 0); 561 unblockNode->succedents[0] = termNode; 562 termNode->antecedents[0] = unblockNode; 563 termNode->antType[0] = rf_control; 564} 565#define CONS_PDA(if,start,num) \ 566 pda_p->col = asmap->if->col; \ 567 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 568 pda_p->numSector = num; \ 569 pda_p->next = NULL; \ 570 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 571#if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 572void 573rf_WriteGenerateFailedAccessASMs( 574 RF_Raid_t * raidPtr, 575 RF_AccessStripeMap_t * asmap, 576 RF_PhysDiskAddr_t ** pdap, 577 int *nNodep, 578 RF_PhysDiskAddr_t ** pqpdap, 579 int *nPQNodep, 580 RF_AllocListElem_t * allocList) 581{ 582 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 583 int PDAPerDisk, i; 584 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 585 int numDataCol = layoutPtr->numDataCol; 586 int state; 587 unsigned napdas; 588 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 589 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 590 RF_PhysDiskAddr_t *pda_p; 591 RF_RaidAddr_t sosAddr; 592 593 /* determine how many pda's we will have to generate per unaccess 594 * stripe. If there is only one failed data unit, it is one; if two, 595 * possibly two, depending wether they overlap. */ 596 597 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 598 fone_end = fone_start + fone->numSector; 599 600 if (asmap->numDataFailed == 1) { 601 PDAPerDisk = 1; 602 state = 1; 603 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 604 pda_p = *pqpdap; 605 /* build p */ 606 CONS_PDA(parityInfo, fone_start, fone->numSector); 607 pda_p->type = RF_PDA_TYPE_PARITY; 608 pda_p++; 609 /* build q */ 610 CONS_PDA(qInfo, fone_start, fone->numSector); 611 pda_p->type = RF_PDA_TYPE_Q; 612 } else { 613 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 614 ftwo_end = ftwo_start + ftwo->numSector; 615 if (fone->numSector + ftwo->numSector > secPerSU) { 616 PDAPerDisk = 1; 617 state = 2; 618 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 619 pda_p = *pqpdap; 620 CONS_PDA(parityInfo, 0, secPerSU); 621 pda_p->type = RF_PDA_TYPE_PARITY; 622 pda_p++; 623 CONS_PDA(qInfo, 0, secPerSU); 624 pda_p->type = RF_PDA_TYPE_Q; 625 } else { 626 PDAPerDisk = 2; 627 state = 3; 628 /* four of them, fone, then ftwo */ 629 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 630 pda_p = *pqpdap; 631 CONS_PDA(parityInfo, fone_start, fone->numSector); 632 pda_p->type = RF_PDA_TYPE_PARITY; 633 pda_p++; 634 CONS_PDA(qInfo, fone_start, fone->numSector); 635 pda_p->type = RF_PDA_TYPE_Q; 636 pda_p++; 637 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 638 pda_p->type = RF_PDA_TYPE_PARITY; 639 pda_p++; 640 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 641 pda_p->type = RF_PDA_TYPE_Q; 642 } 643 } 644 /* figure out number of nonaccessed pda */ 645 napdas = PDAPerDisk * (numDataCol - 2); 646 *nPQNodep = PDAPerDisk; 647 648 *nNodep = napdas; 649 if (napdas == 0) 650 return; /* short circuit */ 651 652 /* allocate up our list of pda's */ 653 654 RF_MallocAndAdd(pda_p, napdas * sizeof(RF_PhysDiskAddr_t), 655 (RF_PhysDiskAddr_t *), allocList); 656 *pdap = pda_p; 657 658 /* linkem together */ 659 for (i = 0; i < (napdas - 1); i++) 660 pda_p[i].next = pda_p + (i + 1); 661 662 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 663 for (i = 0; i < numDataCol; i++) { 664 if ((pda_p - (*pdap)) == napdas) 665 continue; 666 pda_p->type = RF_PDA_TYPE_DATA; 667 pda_p->raidAddress = sosAddr + (i * secPerSU); 668 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 669 /* skip over dead disks */ 670 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->col].status)) 671 continue; 672 switch (state) { 673 case 1: /* fone */ 674 pda_p->numSector = fone->numSector; 675 pda_p->raidAddress += fone_start; 676 pda_p->startSector += fone_start; 677 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 678 break; 679 case 2: /* full stripe */ 680 pda_p->numSector = secPerSU; 681 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 682 break; 683 case 3: /* two slabs */ 684 pda_p->numSector = fone->numSector; 685 pda_p->raidAddress += fone_start; 686 pda_p->startSector += fone_start; 687 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 688 pda_p++; 689 pda_p->type = RF_PDA_TYPE_DATA; 690 pda_p->raidAddress = sosAddr + (i * secPerSU); 691 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 692 pda_p->numSector = ftwo->numSector; 693 pda_p->raidAddress += ftwo_start; 694 pda_p->startSector += ftwo_start; 695 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 696 break; 697 default: 698 RF_PANIC(); 699 } 700 pda_p++; 701 } 702 703 RF_ASSERT(pda_p - *pdap == napdas); 704 return; 705} 706#define DISK_NODE_PDA(node) ((node)->params[0].p) 707 708#define DISK_NODE_PARAMS(_node_,_p_) \ 709 (_node_).params[0].p = _p_ ; \ 710 (_node_).params[1].p = (_p_)->bufPtr; \ 711 (_node_).params[2].v = parityStripeID; \ 712 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru) 713 714void 715rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 716 RF_DagHeader_t *dag_h, void *bp, 717 RF_RaidAccessFlags_t flags, 718 RF_AllocListElem_t *allocList, 719 char *redundantReadNodeName, 720 char *redundantWriteNodeName, 721 char *recoveryNodeName, 722 int (*recovFunc) (RF_DagNode_t *)) 723{ 724 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 725 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 726 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 727 RF_PhysDiskAddr_t *pda, *pqPDAs; 728 RF_PhysDiskAddr_t *npdas; 729 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 730 RF_ReconUnitNum_t which_ru; 731 int nPQNodes; 732 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 733 734 /* simple small write case - First part looks like a reconstruct-read 735 * of the failed data units. Then a write of all data units not 736 * failed. */ 737 738 739 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 740 * / -------PQ----- / \ \ Wud Wp WQ \ | / 741 * --Unblock- | T 742 * 743 * Rrd = read recovery data (potentially none) Wud = write user data 744 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 745 * (could be two) 746 * 747 */ 748 749 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 750 751 RF_ASSERT(asmap->numDataFailed == 1); 752 753 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 754 nReadNodes = nRrdNodes + 2 * nPQNodes; 755 nWriteNodes = nWudNodes + 2 * nPQNodes; 756 nNodes = 4 + nReadNodes + nWriteNodes; 757 758 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 759 blockNode = nodes; 760 unblockNode = blockNode + 1; 761 termNode = unblockNode + 1; 762 recoveryNode = termNode + 1; 763 rrdNodes = recoveryNode + 1; 764 rpNodes = rrdNodes + nRrdNodes; 765 rqNodes = rpNodes + nPQNodes; 766 wudNodes = rqNodes + nPQNodes; 767 wpNodes = wudNodes + nWudNodes; 768 wqNodes = wpNodes + nPQNodes; 769 770 dag_h->creator = "PQ_DDSimpleSmallWrite"; 771 dag_h->numSuccedents = 1; 772 dag_h->succedents[0] = blockNode; 773 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 774 termNode->antecedents[0] = unblockNode; 775 termNode->antType[0] = rf_control; 776 777 /* init the block and unblock nodes */ 778 /* The block node has all the read nodes as successors */ 779 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 780 for (i = 0; i < nReadNodes; i++) 781 blockNode->succedents[i] = rrdNodes + i; 782 783 /* The unblock node has all the writes as successors */ 784 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 785 for (i = 0; i < nWriteNodes; i++) { 786 unblockNode->antecedents[i] = wudNodes + i; 787 unblockNode->antType[i] = rf_control; 788 } 789 unblockNode->succedents[0] = termNode; 790 791#define INIT_READ_NODE(node,name) \ 792 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 793 (node)->succedents[0] = recoveryNode; \ 794 (node)->antecedents[0] = blockNode; \ 795 (node)->antType[0] = rf_control; 796 797 /* build the read nodes */ 798 pda = npdas; 799 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 800 INIT_READ_NODE(rrdNodes + i, "rrd"); 801 DISK_NODE_PARAMS(rrdNodes[i], pda); 802 } 803 804 /* read redundancy pdas */ 805 pda = pqPDAs; 806 INIT_READ_NODE(rpNodes, "Rp"); 807 RF_ASSERT(pda); 808 DISK_NODE_PARAMS(rpNodes[0], pda); 809 pda++; 810 INIT_READ_NODE(rqNodes, redundantReadNodeName); 811 RF_ASSERT(pda); 812 DISK_NODE_PARAMS(rqNodes[0], pda); 813 if (nPQNodes == 2) { 814 pda++; 815 INIT_READ_NODE(rpNodes + 1, "Rp"); 816 RF_ASSERT(pda); 817 DISK_NODE_PARAMS(rpNodes[1], pda); 818 pda++; 819 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 820 RF_ASSERT(pda); 821 DISK_NODE_PARAMS(rqNodes[1], pda); 822 } 823 /* the recovery node has all reads as precedessors and all writes as 824 * successors. It generates a result for every write P or write Q 825 * node. As parameters, it takes a pda per read and a pda per stripe 826 * of user data written. It also takes as the last params the raidPtr 827 * and asm. For results, it takes PDA for P & Q. */ 828 829 830 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 831 nWriteNodes, /* succesors */ 832 nReadNodes, /* preds */ 833 nReadNodes + nWudNodes + 3, /* params */ 834 2 * nPQNodes, /* results */ 835 dag_h, recoveryNodeName, allocList); 836 837 838 839 for (i = 0; i < nReadNodes; i++) { 840 recoveryNode->antecedents[i] = rrdNodes + i; 841 recoveryNode->antType[i] = rf_control; 842 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 843 } 844 for (i = 0; i < nWudNodes; i++) { 845 recoveryNode->succedents[i] = wudNodes + i; 846 } 847 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 848 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 849 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 850 851 for (; i < nWriteNodes; i++) 852 recoveryNode->succedents[i] = wudNodes + i; 853 854 pda = pqPDAs; 855 recoveryNode->results[0] = pda; 856 pda++; 857 recoveryNode->results[1] = pda; 858 if (nPQNodes == 2) { 859 pda++; 860 recoveryNode->results[2] = pda; 861 pda++; 862 recoveryNode->results[3] = pda; 863 } 864 /* fill writes */ 865#define INIT_WRITE_NODE(node,name) \ 866 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 867 (node)->succedents[0] = unblockNode; \ 868 (node)->antecedents[0] = recoveryNode; \ 869 (node)->antType[0] = rf_control; 870 871 pda = asmap->physInfo; 872 for (i = 0; i < nWudNodes; i++) { 873 INIT_WRITE_NODE(wudNodes + i, "Wd"); 874 DISK_NODE_PARAMS(wudNodes[i], pda); 875 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 876 pda = pda->next; 877 } 878 /* write redundancy pdas */ 879 pda = pqPDAs; 880 INIT_WRITE_NODE(wpNodes, "Wp"); 881 RF_ASSERT(pda); 882 DISK_NODE_PARAMS(wpNodes[0], pda); 883 pda++; 884 INIT_WRITE_NODE(wqNodes, "Wq"); 885 RF_ASSERT(pda); 886 DISK_NODE_PARAMS(wqNodes[0], pda); 887 if (nPQNodes == 2) { 888 pda++; 889 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 890 RF_ASSERT(pda); 891 DISK_NODE_PARAMS(wpNodes[1], pda); 892 pda++; 893 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 894 RF_ASSERT(pda); 895 DISK_NODE_PARAMS(wqNodes[1], pda); 896 } 897} 898#endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 899