rf_dagdegwr.c revision 1.7
1/* $NetBSD: rf_dagdegwr.c,v 1.7 2001/09/01 23:50:44 thorpej 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 "rf_types.h" 37#include "rf_raid.h" 38#include "rf_dag.h" 39#include "rf_dagutils.h" 40#include "rf_dagfuncs.h" 41#include "rf_debugMem.h" 42#include "rf_memchunk.h" 43#include "rf_general.h" 44#include "rf_dagdegwr.h" 45 46 47/****************************************************************************** 48 * 49 * General comments on DAG creation: 50 * 51 * All DAGs in this file use roll-away error recovery. Each DAG has a single 52 * commit node, usually called "Cmt." If an error occurs before the Cmt node 53 * is reached, the execution engine will halt forward execution and work 54 * backward through the graph, executing the undo functions. Assuming that 55 * each node in the graph prior to the Cmt node are undoable and atomic - or - 56 * does not make changes to permanent state, the graph will fail atomically. 57 * If an error occurs after the Cmt node executes, the engine will roll-forward 58 * through the graph, blindly executing nodes until it reaches the end. 59 * If a graph reaches the end, it is assumed to have completed successfully. 60 * 61 * A graph has only 1 Cmt node. 62 * 63 */ 64 65 66/****************************************************************************** 67 * 68 * The following wrappers map the standard DAG creation interface to the 69 * DAG creation routines. Additionally, these wrappers enable experimentation 70 * with new DAG structures by providing an extra level of indirection, allowing 71 * the DAG creation routines to be replaced at this single point. 72 */ 73 74static 75RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG) 76{ 77 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, 78 flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE); 79} 80 81void 82rf_CreateDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList) 83 RF_Raid_t *raidPtr; 84 RF_AccessStripeMap_t *asmap; 85 RF_DagHeader_t *dag_h; 86 void *bp; 87 RF_RaidAccessFlags_t flags; 88 RF_AllocListElem_t *allocList; 89{ 90 91 RF_ASSERT(asmap->numDataFailed == 1); 92 dag_h->creator = "DegradedWriteDAG"; 93 94 /* 95 * if the access writes only a portion of the failed unit, and also 96 * writes some portion of at least one surviving unit, we create two 97 * DAGs, one for the failed component and one for the non-failed 98 * component, and do them sequentially. Note that the fact that we're 99 * accessing only a portion of the failed unit indicates that the 100 * access either starts or ends in the failed unit, and hence we need 101 * create only two dags. This is inefficient in that the same data or 102 * parity can get read and written twice using this structure. I need 103 * to fix this to do the access all at once. 104 */ 105 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 && 106 asmap->failedPDAs[0]->numSector != 107 raidPtr->Layout.sectorsPerStripeUnit)); 108 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 109 allocList); 110} 111 112 113 114/****************************************************************************** 115 * 116 * DAG creation code begins here 117 */ 118 119 120 121/****************************************************************************** 122 * 123 * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode 124 * write, which is as follows 125 * 126 * / {Wnq} --\ 127 * hdr -> blockNode -> Rod -> Xor -> Cmt -> Wnp ----> unblock -> term 128 * \ {Rod} / \ Wnd ---/ 129 * \ {Wnd} -/ 130 * 131 * commit nodes: Xor, Wnd 132 * 133 * IMPORTANT: 134 * This DAG generator does not work for double-degraded archs since it does not 135 * generate Q 136 * 137 * This dag is essentially identical to the large-write dag, except that the 138 * write to the failed data unit is suppressed. 139 * 140 * IMPORTANT: this dag does not work in the case where the access writes only 141 * a portion of the failed unit, and also writes some portion of at least one 142 * surviving SU. this case is handled in CreateDegradedWriteDAG above. 143 * 144 * The block & unblock nodes are leftovers from a previous version. They 145 * do nothing, but I haven't deleted them because it would be a tremendous 146 * effort to put them back in. 147 * 148 * This dag is used whenever a one of the data units in a write has failed. 149 * If it is the parity unit that failed, the nonredundant write dag (below) 150 * is used. 151 *****************************************************************************/ 152 153void 154rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 155 allocList, nfaults, redFunc, allowBufferRecycle) 156 RF_Raid_t *raidPtr; 157 RF_AccessStripeMap_t *asmap; 158 RF_DagHeader_t *dag_h; 159 void *bp; 160 RF_RaidAccessFlags_t flags; 161 RF_AllocListElem_t *allocList; 162 int nfaults; 163 int (*redFunc) (RF_DagNode_t *); 164 int allowBufferRecycle; 165{ 166 int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 167 rdnodesFaked; 168 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode; 169 RF_DagNode_t *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode; 170 RF_SectorCount_t sectorsPerSU; 171 RF_ReconUnitNum_t which_ru; 172 char *xorTargetBuf = NULL; /* the target buffer for the XOR 173 * operation */ 174 char *overlappingPDAs;/* a temporary array of flags */ 175 RF_AccessStripeMapHeader_t *new_asm_h[2]; 176 RF_PhysDiskAddr_t *pda, *parityPDA; 177 RF_StripeNum_t parityStripeID; 178 RF_PhysDiskAddr_t *failedPDA; 179 RF_RaidLayout_t *layoutPtr; 180 181 layoutPtr = &(raidPtr->Layout); 182 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 183 &which_ru); 184 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 185 /* failedPDA points to the pda within the asm that targets the failed 186 * disk */ 187 failedPDA = asmap->failedPDAs[0]; 188 189 if (rf_dagDebug) 190 printf("[Creating degraded-write DAG]\n"); 191 192 RF_ASSERT(asmap->numDataFailed == 1); 193 dag_h->creator = "SimpleDegradedWriteDAG"; 194 195 /* 196 * Generate two ASMs identifying the surviving data 197 * we need in order to recover the lost data. 198 */ 199 /* overlappingPDAs array must be zero'd */ 200 RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed, sizeof(char), (char *)); 201 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 202 &nXorBufs, NULL, overlappingPDAs, allocList); 203 204 /* create all the nodes at once */ 205 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 206 * generated for the 207 * failed pda */ 208 209 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 210 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 211 /* 212 * XXX 213 * 214 * There's a bug with a complete stripe overwrite- that means 0 reads 215 * of old data, and the rest of the DAG generation code doesn't like 216 * that. A release is coming, and I don't wanna risk breaking a critical 217 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 218 * I'm gonna fake there being a read node, and I'm gonna swap in a 219 * no-op node in its place (to make all the link-up code happy). 220 * This should be fixed at some point. --jimz 221 */ 222 if (nRrdNodes == 0) { 223 nRrdNodes = 1; 224 rdnodesFaked = 1; 225 } else { 226 rdnodesFaked = 0; 227 } 228 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 229 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 230 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), 231 (RF_DagNode_t *), allocList); 232 i = 0; 233 blockNode = &nodes[i]; 234 i += 1; 235 commitNode = &nodes[i]; 236 i += 1; 237 unblockNode = &nodes[i]; 238 i += 1; 239 termNode = &nodes[i]; 240 i += 1; 241 xorNode = &nodes[i]; 242 i += 1; 243 wnpNode = &nodes[i]; 244 i += 1; 245 wndNodes = &nodes[i]; 246 i += nWndNodes; 247 rrdNodes = &nodes[i]; 248 i += nRrdNodes; 249 if (nfaults == 2) { 250 wnqNode = &nodes[i]; 251 i += 1; 252 } else { 253 wnqNode = NULL; 254 } 255 RF_ASSERT(i == nNodes); 256 257 /* this dag can not commit until all rrd and xor Nodes have completed */ 258 dag_h->numCommitNodes = 1; 259 dag_h->numCommits = 0; 260 dag_h->numSuccedents = 1; 261 262 RF_ASSERT(nRrdNodes > 0); 263 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 264 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 265 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 266 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 267 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 268 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 269 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 270 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 271 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 272 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 273 274 /* 275 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 276 * the failed buffer, save a pointer to it so we can use it as the target 277 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 278 * a buffer is the same size as the failed buffer, it must also be at the 279 * same alignment within the SU. 280 */ 281 i = 0; 282 if (new_asm_h[0]) { 283 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 284 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 285 i++, pda = pda->next) { 286 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 287 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 288 RF_ASSERT(pda); 289 rrdNodes[i].params[0].p = pda; 290 rrdNodes[i].params[1].p = pda->bufPtr; 291 rrdNodes[i].params[2].v = parityStripeID; 292 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 293 } 294 } 295 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 296 if (new_asm_h[1]) { 297 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 298 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 299 j++, pda = pda->next) { 300 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 301 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 302 RF_ASSERT(pda); 303 rrdNodes[i + j].params[0].p = pda; 304 rrdNodes[i + j].params[1].p = pda->bufPtr; 305 rrdNodes[i + j].params[2].v = parityStripeID; 306 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 307 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 308 xorTargetBuf = pda->bufPtr; 309 } 310 } 311 if (rdnodesFaked) { 312 /* 313 * This is where we'll init that fake noop read node 314 * (XXX should the wakeup func be different?) 315 */ 316 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 317 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 318 } 319 /* 320 * Make a PDA for the parity unit. The parity PDA should start at 321 * the same offset into the SU as the failed PDA. 322 */ 323 /* Danner comment: I don't think this copy is really necessary. We are 324 * in one of two cases here. (1) The entire failed unit is written. 325 * Then asmap->parityInfo will describe the entire parity. (2) We are 326 * only writing a subset of the failed unit and nothing else. Then the 327 * asmap->parityInfo describes the failed unit and the copy can also 328 * be avoided. */ 329 330 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 331 parityPDA->row = asmap->parityInfo->row; 332 parityPDA->col = asmap->parityInfo->col; 333 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 334 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 335 parityPDA->numSector = failedPDA->numSector; 336 337 if (!xorTargetBuf) { 338 RF_CallocAndAdd(xorTargetBuf, 1, 339 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 340 } 341 /* init the Wnp node */ 342 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 343 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 344 wnpNode->params[0].p = parityPDA; 345 wnpNode->params[1].p = xorTargetBuf; 346 wnpNode->params[2].v = parityStripeID; 347 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 348 349 /* fill in the Wnq Node */ 350 if (nfaults == 2) { 351 { 352 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 353 (RF_PhysDiskAddr_t *), allocList); 354 parityPDA->row = asmap->qInfo->row; 355 parityPDA->col = asmap->qInfo->col; 356 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 357 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 358 parityPDA->numSector = failedPDA->numSector; 359 360 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 361 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 362 wnqNode->params[0].p = parityPDA; 363 RF_CallocAndAdd(xorNode->results[1], 1, 364 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 365 wnqNode->params[1].p = xorNode->results[1]; 366 wnqNode->params[2].v = parityStripeID; 367 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 368 } 369 } 370 /* fill in the Wnd nodes */ 371 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 372 if (pda == failedPDA) { 373 i--; 374 continue; 375 } 376 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 377 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 378 RF_ASSERT(pda); 379 wndNodes[i].params[0].p = pda; 380 wndNodes[i].params[1].p = pda->bufPtr; 381 wndNodes[i].params[2].v = parityStripeID; 382 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 383 } 384 385 /* fill in the results of the xor node */ 386 xorNode->results[0] = xorTargetBuf; 387 388 /* fill in the params of the xor node */ 389 390 paramNum = 0; 391 if (rdnodesFaked == 0) { 392 for (i = 0; i < nRrdNodes; i++) { 393 /* all the Rrd nodes need to be xored together */ 394 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 395 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 396 } 397 } 398 for (i = 0; i < nWndNodes; i++) { 399 /* any Wnd nodes that overlap the failed access need to be 400 * xored in */ 401 if (overlappingPDAs[i]) { 402 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 403 bcopy((char *) wndNodes[i].params[0].p, (char *) pda, sizeof(RF_PhysDiskAddr_t)); 404 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 405 xorNode->params[paramNum++].p = pda; 406 xorNode->params[paramNum++].p = pda->bufPtr; 407 } 408 } 409 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 410 411 /* 412 * Install the failed PDA into the xor param list so that the 413 * new data gets xor'd in. 414 */ 415 xorNode->params[paramNum++].p = failedPDA; 416 xorNode->params[paramNum++].p = failedPDA->bufPtr; 417 418 /* 419 * The last 2 params to the recovery xor node are always the failed 420 * PDA and the raidPtr. install the failedPDA even though we have just 421 * done so above. This allows us to use the same XOR function for both 422 * degraded reads and degraded writes. 423 */ 424 xorNode->params[paramNum++].p = failedPDA; 425 xorNode->params[paramNum++].p = raidPtr; 426 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 427 428 /* 429 * Code to link nodes begins here 430 */ 431 432 /* link header to block node */ 433 RF_ASSERT(blockNode->numAntecedents == 0); 434 dag_h->succedents[0] = blockNode; 435 436 /* link block node to rd nodes */ 437 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 438 for (i = 0; i < nRrdNodes; i++) { 439 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 440 blockNode->succedents[i] = &rrdNodes[i]; 441 rrdNodes[i].antecedents[0] = blockNode; 442 rrdNodes[i].antType[0] = rf_control; 443 } 444 445 /* link read nodes to xor node */ 446 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 447 for (i = 0; i < nRrdNodes; i++) { 448 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 449 rrdNodes[i].succedents[0] = xorNode; 450 xorNode->antecedents[i] = &rrdNodes[i]; 451 xorNode->antType[i] = rf_trueData; 452 } 453 454 /* link xor node to commit node */ 455 RF_ASSERT(xorNode->numSuccedents == 1); 456 RF_ASSERT(commitNode->numAntecedents == 1); 457 xorNode->succedents[0] = commitNode; 458 commitNode->antecedents[0] = xorNode; 459 commitNode->antType[0] = rf_control; 460 461 /* link commit node to wnd nodes */ 462 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 463 for (i = 0; i < nWndNodes; i++) { 464 RF_ASSERT(wndNodes[i].numAntecedents == 1); 465 commitNode->succedents[i] = &wndNodes[i]; 466 wndNodes[i].antecedents[0] = commitNode; 467 wndNodes[i].antType[0] = rf_control; 468 } 469 470 /* link the commit node to wnp, wnq nodes */ 471 RF_ASSERT(wnpNode->numAntecedents == 1); 472 commitNode->succedents[nWndNodes] = wnpNode; 473 wnpNode->antecedents[0] = commitNode; 474 wnpNode->antType[0] = rf_control; 475 if (nfaults == 2) { 476 RF_ASSERT(wnqNode->numAntecedents == 1); 477 commitNode->succedents[nWndNodes + 1] = wnqNode; 478 wnqNode->antecedents[0] = commitNode; 479 wnqNode->antType[0] = rf_control; 480 } 481 /* link write new data nodes to unblock node */ 482 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 483 for (i = 0; i < nWndNodes; i++) { 484 RF_ASSERT(wndNodes[i].numSuccedents == 1); 485 wndNodes[i].succedents[0] = unblockNode; 486 unblockNode->antecedents[i] = &wndNodes[i]; 487 unblockNode->antType[i] = rf_control; 488 } 489 490 /* link write new parity node to unblock node */ 491 RF_ASSERT(wnpNode->numSuccedents == 1); 492 wnpNode->succedents[0] = unblockNode; 493 unblockNode->antecedents[nWndNodes] = wnpNode; 494 unblockNode->antType[nWndNodes] = rf_control; 495 496 /* link write new q node to unblock node */ 497 if (nfaults == 2) { 498 RF_ASSERT(wnqNode->numSuccedents == 1); 499 wnqNode->succedents[0] = unblockNode; 500 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 501 unblockNode->antType[nWndNodes + 1] = rf_control; 502 } 503 /* link unblock node to term node */ 504 RF_ASSERT(unblockNode->numSuccedents == 1); 505 RF_ASSERT(termNode->numAntecedents == 1); 506 RF_ASSERT(termNode->numSuccedents == 0); 507 unblockNode->succedents[0] = termNode; 508 termNode->antecedents[0] = unblockNode; 509 termNode->antType[0] = rf_control; 510} 511#define CONS_PDA(if,start,num) \ 512 pda_p->row = asmap->if->row; pda_p->col = asmap->if->col; \ 513 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 514 pda_p->numSector = num; \ 515 pda_p->next = NULL; \ 516 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 517#if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 518void 519rf_WriteGenerateFailedAccessASMs( 520 RF_Raid_t * raidPtr, 521 RF_AccessStripeMap_t * asmap, 522 RF_PhysDiskAddr_t ** pdap, 523 int *nNodep, 524 RF_PhysDiskAddr_t ** pqpdap, 525 int *nPQNodep, 526 RF_AllocListElem_t * allocList) 527{ 528 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 529 int PDAPerDisk, i; 530 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 531 int numDataCol = layoutPtr->numDataCol; 532 int state; 533 unsigned napdas; 534 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 535 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 536 RF_PhysDiskAddr_t *pda_p; 537 RF_RaidAddr_t sosAddr; 538 539 /* determine how many pda's we will have to generate per unaccess 540 * stripe. If there is only one failed data unit, it is one; if two, 541 * possibly two, depending wether they overlap. */ 542 543 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 544 fone_end = fone_start + fone->numSector; 545 546 if (asmap->numDataFailed == 1) { 547 PDAPerDisk = 1; 548 state = 1; 549 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 550 pda_p = *pqpdap; 551 /* build p */ 552 CONS_PDA(parityInfo, fone_start, fone->numSector); 553 pda_p->type = RF_PDA_TYPE_PARITY; 554 pda_p++; 555 /* build q */ 556 CONS_PDA(qInfo, fone_start, fone->numSector); 557 pda_p->type = RF_PDA_TYPE_Q; 558 } else { 559 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 560 ftwo_end = ftwo_start + ftwo->numSector; 561 if (fone->numSector + ftwo->numSector > secPerSU) { 562 PDAPerDisk = 1; 563 state = 2; 564 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 565 pda_p = *pqpdap; 566 CONS_PDA(parityInfo, 0, secPerSU); 567 pda_p->type = RF_PDA_TYPE_PARITY; 568 pda_p++; 569 CONS_PDA(qInfo, 0, secPerSU); 570 pda_p->type = RF_PDA_TYPE_Q; 571 } else { 572 PDAPerDisk = 2; 573 state = 3; 574 /* four of them, fone, then ftwo */ 575 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 576 pda_p = *pqpdap; 577 CONS_PDA(parityInfo, fone_start, fone->numSector); 578 pda_p->type = RF_PDA_TYPE_PARITY; 579 pda_p++; 580 CONS_PDA(qInfo, fone_start, fone->numSector); 581 pda_p->type = RF_PDA_TYPE_Q; 582 pda_p++; 583 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 584 pda_p->type = RF_PDA_TYPE_PARITY; 585 pda_p++; 586 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 587 pda_p->type = RF_PDA_TYPE_Q; 588 } 589 } 590 /* figure out number of nonaccessed pda */ 591 napdas = PDAPerDisk * (numDataCol - 2); 592 *nPQNodep = PDAPerDisk; 593 594 *nNodep = napdas; 595 if (napdas == 0) 596 return; /* short circuit */ 597 598 /* allocate up our list of pda's */ 599 600 RF_CallocAndAdd(pda_p, napdas, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 601 *pdap = pda_p; 602 603 /* linkem together */ 604 for (i = 0; i < (napdas - 1); i++) 605 pda_p[i].next = pda_p + (i + 1); 606 607 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 608 for (i = 0; i < numDataCol; i++) { 609 if ((pda_p - (*pdap)) == napdas) 610 continue; 611 pda_p->type = RF_PDA_TYPE_DATA; 612 pda_p->raidAddress = sosAddr + (i * secPerSU); 613 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 614 /* skip over dead disks */ 615 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->row][pda_p->col].status)) 616 continue; 617 switch (state) { 618 case 1: /* fone */ 619 pda_p->numSector = fone->numSector; 620 pda_p->raidAddress += fone_start; 621 pda_p->startSector += fone_start; 622 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 623 break; 624 case 2: /* full stripe */ 625 pda_p->numSector = secPerSU; 626 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 627 break; 628 case 3: /* two slabs */ 629 pda_p->numSector = fone->numSector; 630 pda_p->raidAddress += fone_start; 631 pda_p->startSector += fone_start; 632 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 633 pda_p++; 634 pda_p->type = RF_PDA_TYPE_DATA; 635 pda_p->raidAddress = sosAddr + (i * secPerSU); 636 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 637 pda_p->numSector = ftwo->numSector; 638 pda_p->raidAddress += ftwo_start; 639 pda_p->startSector += ftwo_start; 640 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 641 break; 642 default: 643 RF_PANIC(); 644 } 645 pda_p++; 646 } 647 648 RF_ASSERT(pda_p - *pdap == napdas); 649 return; 650} 651#define DISK_NODE_PDA(node) ((node)->params[0].p) 652 653#define DISK_NODE_PARAMS(_node_,_p_) \ 654 (_node_).params[0].p = _p_ ; \ 655 (_node_).params[1].p = (_p_)->bufPtr; \ 656 (_node_).params[2].v = parityStripeID; \ 657 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru) 658 659void 660rf_DoubleDegSmallWrite( 661 RF_Raid_t * raidPtr, 662 RF_AccessStripeMap_t * asmap, 663 RF_DagHeader_t * dag_h, 664 void *bp, 665 RF_RaidAccessFlags_t flags, 666 RF_AllocListElem_t * allocList, 667 char *redundantReadNodeName, 668 char *redundantWriteNodeName, 669 char *recoveryNodeName, 670 int (*recovFunc) (RF_DagNode_t *)) 671{ 672 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 673 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 674 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 675 RF_PhysDiskAddr_t *pda, *pqPDAs; 676 RF_PhysDiskAddr_t *npdas; 677 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 678 RF_ReconUnitNum_t which_ru; 679 int nPQNodes; 680 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 681 682 /* simple small write case - First part looks like a reconstruct-read 683 * of the failed data units. Then a write of all data units not 684 * failed. */ 685 686 687 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 688 * / -------PQ----- / \ \ Wud Wp WQ \ | / 689 * --Unblock- | T 690 * 691 * Rrd = read recovery data (potentially none) Wud = write user data 692 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 693 * (could be two) 694 * 695 */ 696 697 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 698 699 RF_ASSERT(asmap->numDataFailed == 1); 700 701 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 702 nReadNodes = nRrdNodes + 2 * nPQNodes; 703 nWriteNodes = nWudNodes + 2 * nPQNodes; 704 nNodes = 4 + nReadNodes + nWriteNodes; 705 706 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 707 blockNode = nodes; 708 unblockNode = blockNode + 1; 709 termNode = unblockNode + 1; 710 recoveryNode = termNode + 1; 711 rrdNodes = recoveryNode + 1; 712 rpNodes = rrdNodes + nRrdNodes; 713 rqNodes = rpNodes + nPQNodes; 714 wudNodes = rqNodes + nPQNodes; 715 wpNodes = wudNodes + nWudNodes; 716 wqNodes = wpNodes + nPQNodes; 717 718 dag_h->creator = "PQ_DDSimpleSmallWrite"; 719 dag_h->numSuccedents = 1; 720 dag_h->succedents[0] = blockNode; 721 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 722 termNode->antecedents[0] = unblockNode; 723 termNode->antType[0] = rf_control; 724 725 /* init the block and unblock nodes */ 726 /* The block node has all the read nodes as successors */ 727 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 728 for (i = 0; i < nReadNodes; i++) 729 blockNode->succedents[i] = rrdNodes + i; 730 731 /* The unblock node has all the writes as successors */ 732 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 733 for (i = 0; i < nWriteNodes; i++) { 734 unblockNode->antecedents[i] = wudNodes + i; 735 unblockNode->antType[i] = rf_control; 736 } 737 unblockNode->succedents[0] = termNode; 738 739#define INIT_READ_NODE(node,name) \ 740 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 741 (node)->succedents[0] = recoveryNode; \ 742 (node)->antecedents[0] = blockNode; \ 743 (node)->antType[0] = rf_control; 744 745 /* build the read nodes */ 746 pda = npdas; 747 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 748 INIT_READ_NODE(rrdNodes + i, "rrd"); 749 DISK_NODE_PARAMS(rrdNodes[i], pda); 750 } 751 752 /* read redundancy pdas */ 753 pda = pqPDAs; 754 INIT_READ_NODE(rpNodes, "Rp"); 755 RF_ASSERT(pda); 756 DISK_NODE_PARAMS(rpNodes[0], pda); 757 pda++; 758 INIT_READ_NODE(rqNodes, redundantReadNodeName); 759 RF_ASSERT(pda); 760 DISK_NODE_PARAMS(rqNodes[0], pda); 761 if (nPQNodes == 2) { 762 pda++; 763 INIT_READ_NODE(rpNodes + 1, "Rp"); 764 RF_ASSERT(pda); 765 DISK_NODE_PARAMS(rpNodes[1], pda); 766 pda++; 767 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 768 RF_ASSERT(pda); 769 DISK_NODE_PARAMS(rqNodes[1], pda); 770 } 771 /* the recovery node has all reads as precedessors and all writes as 772 * successors. It generates a result for every write P or write Q 773 * node. As parameters, it takes a pda per read and a pda per stripe 774 * of user data written. It also takes as the last params the raidPtr 775 * and asm. For results, it takes PDA for P & Q. */ 776 777 778 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 779 nWriteNodes, /* succesors */ 780 nReadNodes, /* preds */ 781 nReadNodes + nWudNodes + 3, /* params */ 782 2 * nPQNodes, /* results */ 783 dag_h, recoveryNodeName, allocList); 784 785 786 787 for (i = 0; i < nReadNodes; i++) { 788 recoveryNode->antecedents[i] = rrdNodes + i; 789 recoveryNode->antType[i] = rf_control; 790 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 791 } 792 for (i = 0; i < nWudNodes; i++) { 793 recoveryNode->succedents[i] = wudNodes + i; 794 } 795 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 796 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 797 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 798 799 for (; i < nWriteNodes; i++) 800 recoveryNode->succedents[i] = wudNodes + i; 801 802 pda = pqPDAs; 803 recoveryNode->results[0] = pda; 804 pda++; 805 recoveryNode->results[1] = pda; 806 if (nPQNodes == 2) { 807 pda++; 808 recoveryNode->results[2] = pda; 809 pda++; 810 recoveryNode->results[3] = pda; 811 } 812 /* fill writes */ 813#define INIT_WRITE_NODE(node,name) \ 814 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 815 (node)->succedents[0] = unblockNode; \ 816 (node)->antecedents[0] = recoveryNode; \ 817 (node)->antType[0] = rf_control; 818 819 pda = asmap->physInfo; 820 for (i = 0; i < nWudNodes; i++) { 821 INIT_WRITE_NODE(wudNodes + i, "Wd"); 822 DISK_NODE_PARAMS(wudNodes[i], pda); 823 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 824 pda = pda->next; 825 } 826 /* write redundancy pdas */ 827 pda = pqPDAs; 828 INIT_WRITE_NODE(wpNodes, "Wp"); 829 RF_ASSERT(pda); 830 DISK_NODE_PARAMS(wpNodes[0], pda); 831 pda++; 832 INIT_WRITE_NODE(wqNodes, "Wq"); 833 RF_ASSERT(pda); 834 DISK_NODE_PARAMS(wqNodes[0], pda); 835 if (nPQNodes == 2) { 836 pda++; 837 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 838 RF_ASSERT(pda); 839 DISK_NODE_PARAMS(wpNodes[1], pda); 840 pda++; 841 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 842 RF_ASSERT(pda); 843 DISK_NODE_PARAMS(wqNodes[1], pda); 844 } 845} 846#endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 847