rf_dagdegwr.c revision 1.8
1/* $NetBSD: rf_dagdegwr.c,v 1.8 2001/10/04 15:58:52 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 <dev/raidframe/raidframevar.h> 37 38#include "rf_raid.h" 39#include "rf_dag.h" 40#include "rf_dagutils.h" 41#include "rf_dagfuncs.h" 42#include "rf_debugMem.h" 43#include "rf_memchunk.h" 44#include "rf_general.h" 45#include "rf_dagdegwr.h" 46 47 48/****************************************************************************** 49 * 50 * General comments on DAG creation: 51 * 52 * All DAGs in this file use roll-away error recovery. Each DAG has a single 53 * commit node, usually called "Cmt." If an error occurs before the Cmt node 54 * is reached, the execution engine will halt forward execution and work 55 * backward through the graph, executing the undo functions. Assuming that 56 * each node in the graph prior to the Cmt node are undoable and atomic - or - 57 * does not make changes to permanent state, the graph will fail atomically. 58 * If an error occurs after the Cmt node executes, the engine will roll-forward 59 * through the graph, blindly executing nodes until it reaches the end. 60 * If a graph reaches the end, it is assumed to have completed successfully. 61 * 62 * A graph has only 1 Cmt node. 63 * 64 */ 65 66 67/****************************************************************************** 68 * 69 * The following wrappers map the standard DAG creation interface to the 70 * DAG creation routines. Additionally, these wrappers enable experimentation 71 * with new DAG structures by providing an extra level of indirection, allowing 72 * the DAG creation routines to be replaced at this single point. 73 */ 74 75static 76RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG) 77{ 78 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, 79 flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE); 80} 81 82void 83rf_CreateDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList) 84 RF_Raid_t *raidPtr; 85 RF_AccessStripeMap_t *asmap; 86 RF_DagHeader_t *dag_h; 87 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(raidPtr, asmap, dag_h, bp, flags, 156 allocList, nfaults, redFunc, allowBufferRecycle) 157 RF_Raid_t *raidPtr; 158 RF_AccessStripeMap_t *asmap; 159 RF_DagHeader_t *dag_h; 160 void *bp; 161 RF_RaidAccessFlags_t flags; 162 RF_AllocListElem_t *allocList; 163 int nfaults; 164 int (*redFunc) (RF_DagNode_t *); 165 int allowBufferRecycle; 166{ 167 int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 168 rdnodesFaked; 169 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode; 170 RF_DagNode_t *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode; 171 RF_SectorCount_t sectorsPerSU; 172 RF_ReconUnitNum_t which_ru; 173 char *xorTargetBuf = NULL; /* the target buffer for the XOR 174 * operation */ 175 char *overlappingPDAs;/* a temporary array of flags */ 176 RF_AccessStripeMapHeader_t *new_asm_h[2]; 177 RF_PhysDiskAddr_t *pda, *parityPDA; 178 RF_StripeNum_t parityStripeID; 179 RF_PhysDiskAddr_t *failedPDA; 180 RF_RaidLayout_t *layoutPtr; 181 182 layoutPtr = &(raidPtr->Layout); 183 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 184 &which_ru); 185 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 186 /* failedPDA points to the pda within the asm that targets the failed 187 * disk */ 188 failedPDA = asmap->failedPDAs[0]; 189 190 if (rf_dagDebug) 191 printf("[Creating degraded-write DAG]\n"); 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 RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed, sizeof(char), (char *)); 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 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), 232 (RF_DagNode_t *), allocList); 233 i = 0; 234 blockNode = &nodes[i]; 235 i += 1; 236 commitNode = &nodes[i]; 237 i += 1; 238 unblockNode = &nodes[i]; 239 i += 1; 240 termNode = &nodes[i]; 241 i += 1; 242 xorNode = &nodes[i]; 243 i += 1; 244 wnpNode = &nodes[i]; 245 i += 1; 246 wndNodes = &nodes[i]; 247 i += nWndNodes; 248 rrdNodes = &nodes[i]; 249 i += nRrdNodes; 250 if (nfaults == 2) { 251 wnqNode = &nodes[i]; 252 i += 1; 253 } else { 254 wnqNode = NULL; 255 } 256 RF_ASSERT(i == nNodes); 257 258 /* this dag can not commit until all rrd and xor Nodes have completed */ 259 dag_h->numCommitNodes = 1; 260 dag_h->numCommits = 0; 261 dag_h->numSuccedents = 1; 262 263 RF_ASSERT(nRrdNodes > 0); 264 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 265 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 266 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 267 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 268 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 269 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 270 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 271 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 272 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 273 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 274 275 /* 276 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 277 * the failed buffer, save a pointer to it so we can use it as the target 278 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 279 * a buffer is the same size as the failed buffer, it must also be at the 280 * same alignment within the SU. 281 */ 282 i = 0; 283 if (new_asm_h[0]) { 284 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 285 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 286 i++, pda = pda->next) { 287 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 288 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 289 RF_ASSERT(pda); 290 rrdNodes[i].params[0].p = pda; 291 rrdNodes[i].params[1].p = pda->bufPtr; 292 rrdNodes[i].params[2].v = parityStripeID; 293 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 294 } 295 } 296 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 297 if (new_asm_h[1]) { 298 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 299 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 300 j++, pda = pda->next) { 301 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 302 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 303 RF_ASSERT(pda); 304 rrdNodes[i + j].params[0].p = pda; 305 rrdNodes[i + j].params[1].p = pda->bufPtr; 306 rrdNodes[i + j].params[2].v = parityStripeID; 307 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 308 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 309 xorTargetBuf = pda->bufPtr; 310 } 311 } 312 if (rdnodesFaked) { 313 /* 314 * This is where we'll init that fake noop read node 315 * (XXX should the wakeup func be different?) 316 */ 317 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 318 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 319 } 320 /* 321 * Make a PDA for the parity unit. The parity PDA should start at 322 * the same offset into the SU as the failed PDA. 323 */ 324 /* Danner comment: I don't think this copy is really necessary. We are 325 * in one of two cases here. (1) The entire failed unit is written. 326 * Then asmap->parityInfo will describe the entire parity. (2) We are 327 * only writing a subset of the failed unit and nothing else. Then the 328 * asmap->parityInfo describes the failed unit and the copy can also 329 * be avoided. */ 330 331 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 332 parityPDA->row = asmap->parityInfo->row; 333 parityPDA->col = asmap->parityInfo->col; 334 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 335 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 336 parityPDA->numSector = failedPDA->numSector; 337 338 if (!xorTargetBuf) { 339 RF_CallocAndAdd(xorTargetBuf, 1, 340 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 341 } 342 /* init the Wnp node */ 343 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 344 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 345 wnpNode->params[0].p = parityPDA; 346 wnpNode->params[1].p = xorTargetBuf; 347 wnpNode->params[2].v = parityStripeID; 348 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 349 350 /* fill in the Wnq Node */ 351 if (nfaults == 2) { 352 { 353 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 354 (RF_PhysDiskAddr_t *), allocList); 355 parityPDA->row = asmap->qInfo->row; 356 parityPDA->col = asmap->qInfo->col; 357 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 358 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 359 parityPDA->numSector = failedPDA->numSector; 360 361 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 362 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 363 wnqNode->params[0].p = parityPDA; 364 RF_CallocAndAdd(xorNode->results[1], 1, 365 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 366 wnqNode->params[1].p = xorNode->results[1]; 367 wnqNode->params[2].v = parityStripeID; 368 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 369 } 370 } 371 /* fill in the Wnd nodes */ 372 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 373 if (pda == failedPDA) { 374 i--; 375 continue; 376 } 377 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 378 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 379 RF_ASSERT(pda); 380 wndNodes[i].params[0].p = pda; 381 wndNodes[i].params[1].p = pda->bufPtr; 382 wndNodes[i].params[2].v = parityStripeID; 383 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 384 } 385 386 /* fill in the results of the xor node */ 387 xorNode->results[0] = xorTargetBuf; 388 389 /* fill in the params of the xor node */ 390 391 paramNum = 0; 392 if (rdnodesFaked == 0) { 393 for (i = 0; i < nRrdNodes; i++) { 394 /* all the Rrd nodes need to be xored together */ 395 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 396 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 397 } 398 } 399 for (i = 0; i < nWndNodes; i++) { 400 /* any Wnd nodes that overlap the failed access need to be 401 * xored in */ 402 if (overlappingPDAs[i]) { 403 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 404 bcopy((char *) wndNodes[i].params[0].p, (char *) pda, sizeof(RF_PhysDiskAddr_t)); 405 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 406 xorNode->params[paramNum++].p = pda; 407 xorNode->params[paramNum++].p = pda->bufPtr; 408 } 409 } 410 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 411 412 /* 413 * Install the failed PDA into the xor param list so that the 414 * new data gets xor'd in. 415 */ 416 xorNode->params[paramNum++].p = failedPDA; 417 xorNode->params[paramNum++].p = failedPDA->bufPtr; 418 419 /* 420 * The last 2 params to the recovery xor node are always the failed 421 * PDA and the raidPtr. install the failedPDA even though we have just 422 * done so above. This allows us to use the same XOR function for both 423 * degraded reads and degraded writes. 424 */ 425 xorNode->params[paramNum++].p = failedPDA; 426 xorNode->params[paramNum++].p = raidPtr; 427 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 428 429 /* 430 * Code to link nodes begins here 431 */ 432 433 /* link header to block node */ 434 RF_ASSERT(blockNode->numAntecedents == 0); 435 dag_h->succedents[0] = blockNode; 436 437 /* link block node to rd nodes */ 438 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 439 for (i = 0; i < nRrdNodes; i++) { 440 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 441 blockNode->succedents[i] = &rrdNodes[i]; 442 rrdNodes[i].antecedents[0] = blockNode; 443 rrdNodes[i].antType[0] = rf_control; 444 } 445 446 /* link read nodes to xor node */ 447 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 448 for (i = 0; i < nRrdNodes; i++) { 449 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 450 rrdNodes[i].succedents[0] = xorNode; 451 xorNode->antecedents[i] = &rrdNodes[i]; 452 xorNode->antType[i] = rf_trueData; 453 } 454 455 /* link xor node to commit node */ 456 RF_ASSERT(xorNode->numSuccedents == 1); 457 RF_ASSERT(commitNode->numAntecedents == 1); 458 xorNode->succedents[0] = commitNode; 459 commitNode->antecedents[0] = xorNode; 460 commitNode->antType[0] = rf_control; 461 462 /* link commit node to wnd nodes */ 463 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 464 for (i = 0; i < nWndNodes; i++) { 465 RF_ASSERT(wndNodes[i].numAntecedents == 1); 466 commitNode->succedents[i] = &wndNodes[i]; 467 wndNodes[i].antecedents[0] = commitNode; 468 wndNodes[i].antType[0] = rf_control; 469 } 470 471 /* link the commit node to wnp, wnq nodes */ 472 RF_ASSERT(wnpNode->numAntecedents == 1); 473 commitNode->succedents[nWndNodes] = wnpNode; 474 wnpNode->antecedents[0] = commitNode; 475 wnpNode->antType[0] = rf_control; 476 if (nfaults == 2) { 477 RF_ASSERT(wnqNode->numAntecedents == 1); 478 commitNode->succedents[nWndNodes + 1] = wnqNode; 479 wnqNode->antecedents[0] = commitNode; 480 wnqNode->antType[0] = rf_control; 481 } 482 /* link write new data nodes to unblock node */ 483 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 484 for (i = 0; i < nWndNodes; i++) { 485 RF_ASSERT(wndNodes[i].numSuccedents == 1); 486 wndNodes[i].succedents[0] = unblockNode; 487 unblockNode->antecedents[i] = &wndNodes[i]; 488 unblockNode->antType[i] = rf_control; 489 } 490 491 /* link write new parity node to unblock node */ 492 RF_ASSERT(wnpNode->numSuccedents == 1); 493 wnpNode->succedents[0] = unblockNode; 494 unblockNode->antecedents[nWndNodes] = wnpNode; 495 unblockNode->antType[nWndNodes] = rf_control; 496 497 /* link write new q node to unblock node */ 498 if (nfaults == 2) { 499 RF_ASSERT(wnqNode->numSuccedents == 1); 500 wnqNode->succedents[0] = unblockNode; 501 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 502 unblockNode->antType[nWndNodes + 1] = rf_control; 503 } 504 /* link unblock node to term node */ 505 RF_ASSERT(unblockNode->numSuccedents == 1); 506 RF_ASSERT(termNode->numAntecedents == 1); 507 RF_ASSERT(termNode->numSuccedents == 0); 508 unblockNode->succedents[0] = termNode; 509 termNode->antecedents[0] = unblockNode; 510 termNode->antType[0] = rf_control; 511} 512#define CONS_PDA(if,start,num) \ 513 pda_p->row = asmap->if->row; pda_p->col = asmap->if->col; \ 514 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 515 pda_p->numSector = num; \ 516 pda_p->next = NULL; \ 517 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 518#if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 519void 520rf_WriteGenerateFailedAccessASMs( 521 RF_Raid_t * raidPtr, 522 RF_AccessStripeMap_t * asmap, 523 RF_PhysDiskAddr_t ** pdap, 524 int *nNodep, 525 RF_PhysDiskAddr_t ** pqpdap, 526 int *nPQNodep, 527 RF_AllocListElem_t * allocList) 528{ 529 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 530 int PDAPerDisk, i; 531 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 532 int numDataCol = layoutPtr->numDataCol; 533 int state; 534 unsigned napdas; 535 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 536 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 537 RF_PhysDiskAddr_t *pda_p; 538 RF_RaidAddr_t sosAddr; 539 540 /* determine how many pda's we will have to generate per unaccess 541 * stripe. If there is only one failed data unit, it is one; if two, 542 * possibly two, depending wether they overlap. */ 543 544 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 545 fone_end = fone_start + fone->numSector; 546 547 if (asmap->numDataFailed == 1) { 548 PDAPerDisk = 1; 549 state = 1; 550 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 551 pda_p = *pqpdap; 552 /* build p */ 553 CONS_PDA(parityInfo, fone_start, fone->numSector); 554 pda_p->type = RF_PDA_TYPE_PARITY; 555 pda_p++; 556 /* build q */ 557 CONS_PDA(qInfo, fone_start, fone->numSector); 558 pda_p->type = RF_PDA_TYPE_Q; 559 } else { 560 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 561 ftwo_end = ftwo_start + ftwo->numSector; 562 if (fone->numSector + ftwo->numSector > secPerSU) { 563 PDAPerDisk = 1; 564 state = 2; 565 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 566 pda_p = *pqpdap; 567 CONS_PDA(parityInfo, 0, secPerSU); 568 pda_p->type = RF_PDA_TYPE_PARITY; 569 pda_p++; 570 CONS_PDA(qInfo, 0, secPerSU); 571 pda_p->type = RF_PDA_TYPE_Q; 572 } else { 573 PDAPerDisk = 2; 574 state = 3; 575 /* four of them, fone, then ftwo */ 576 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 577 pda_p = *pqpdap; 578 CONS_PDA(parityInfo, fone_start, fone->numSector); 579 pda_p->type = RF_PDA_TYPE_PARITY; 580 pda_p++; 581 CONS_PDA(qInfo, fone_start, fone->numSector); 582 pda_p->type = RF_PDA_TYPE_Q; 583 pda_p++; 584 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 585 pda_p->type = RF_PDA_TYPE_PARITY; 586 pda_p++; 587 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 588 pda_p->type = RF_PDA_TYPE_Q; 589 } 590 } 591 /* figure out number of nonaccessed pda */ 592 napdas = PDAPerDisk * (numDataCol - 2); 593 *nPQNodep = PDAPerDisk; 594 595 *nNodep = napdas; 596 if (napdas == 0) 597 return; /* short circuit */ 598 599 /* allocate up our list of pda's */ 600 601 RF_CallocAndAdd(pda_p, napdas, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 602 *pdap = pda_p; 603 604 /* linkem together */ 605 for (i = 0; i < (napdas - 1); i++) 606 pda_p[i].next = pda_p + (i + 1); 607 608 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 609 for (i = 0; i < numDataCol; i++) { 610 if ((pda_p - (*pdap)) == napdas) 611 continue; 612 pda_p->type = RF_PDA_TYPE_DATA; 613 pda_p->raidAddress = sosAddr + (i * secPerSU); 614 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 615 /* skip over dead disks */ 616 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->row][pda_p->col].status)) 617 continue; 618 switch (state) { 619 case 1: /* fone */ 620 pda_p->numSector = fone->numSector; 621 pda_p->raidAddress += fone_start; 622 pda_p->startSector += fone_start; 623 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 624 break; 625 case 2: /* full stripe */ 626 pda_p->numSector = secPerSU; 627 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 628 break; 629 case 3: /* two slabs */ 630 pda_p->numSector = fone->numSector; 631 pda_p->raidAddress += fone_start; 632 pda_p->startSector += fone_start; 633 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 634 pda_p++; 635 pda_p->type = RF_PDA_TYPE_DATA; 636 pda_p->raidAddress = sosAddr + (i * secPerSU); 637 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 638 pda_p->numSector = ftwo->numSector; 639 pda_p->raidAddress += ftwo_start; 640 pda_p->startSector += ftwo_start; 641 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 642 break; 643 default: 644 RF_PANIC(); 645 } 646 pda_p++; 647 } 648 649 RF_ASSERT(pda_p - *pdap == napdas); 650 return; 651} 652#define DISK_NODE_PDA(node) ((node)->params[0].p) 653 654#define DISK_NODE_PARAMS(_node_,_p_) \ 655 (_node_).params[0].p = _p_ ; \ 656 (_node_).params[1].p = (_p_)->bufPtr; \ 657 (_node_).params[2].v = parityStripeID; \ 658 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru) 659 660void 661rf_DoubleDegSmallWrite( 662 RF_Raid_t * raidPtr, 663 RF_AccessStripeMap_t * asmap, 664 RF_DagHeader_t * dag_h, 665 void *bp, 666 RF_RaidAccessFlags_t flags, 667 RF_AllocListElem_t * allocList, 668 char *redundantReadNodeName, 669 char *redundantWriteNodeName, 670 char *recoveryNodeName, 671 int (*recovFunc) (RF_DagNode_t *)) 672{ 673 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 674 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 675 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 676 RF_PhysDiskAddr_t *pda, *pqPDAs; 677 RF_PhysDiskAddr_t *npdas; 678 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 679 RF_ReconUnitNum_t which_ru; 680 int nPQNodes; 681 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 682 683 /* simple small write case - First part looks like a reconstruct-read 684 * of the failed data units. Then a write of all data units not 685 * failed. */ 686 687 688 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 689 * / -------PQ----- / \ \ Wud Wp WQ \ | / 690 * --Unblock- | T 691 * 692 * Rrd = read recovery data (potentially none) Wud = write user data 693 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 694 * (could be two) 695 * 696 */ 697 698 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 699 700 RF_ASSERT(asmap->numDataFailed == 1); 701 702 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 703 nReadNodes = nRrdNodes + 2 * nPQNodes; 704 nWriteNodes = nWudNodes + 2 * nPQNodes; 705 nNodes = 4 + nReadNodes + nWriteNodes; 706 707 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 708 blockNode = nodes; 709 unblockNode = blockNode + 1; 710 termNode = unblockNode + 1; 711 recoveryNode = termNode + 1; 712 rrdNodes = recoveryNode + 1; 713 rpNodes = rrdNodes + nRrdNodes; 714 rqNodes = rpNodes + nPQNodes; 715 wudNodes = rqNodes + nPQNodes; 716 wpNodes = wudNodes + nWudNodes; 717 wqNodes = wpNodes + nPQNodes; 718 719 dag_h->creator = "PQ_DDSimpleSmallWrite"; 720 dag_h->numSuccedents = 1; 721 dag_h->succedents[0] = blockNode; 722 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 723 termNode->antecedents[0] = unblockNode; 724 termNode->antType[0] = rf_control; 725 726 /* init the block and unblock nodes */ 727 /* The block node has all the read nodes as successors */ 728 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 729 for (i = 0; i < nReadNodes; i++) 730 blockNode->succedents[i] = rrdNodes + i; 731 732 /* The unblock node has all the writes as successors */ 733 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 734 for (i = 0; i < nWriteNodes; i++) { 735 unblockNode->antecedents[i] = wudNodes + i; 736 unblockNode->antType[i] = rf_control; 737 } 738 unblockNode->succedents[0] = termNode; 739 740#define INIT_READ_NODE(node,name) \ 741 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 742 (node)->succedents[0] = recoveryNode; \ 743 (node)->antecedents[0] = blockNode; \ 744 (node)->antType[0] = rf_control; 745 746 /* build the read nodes */ 747 pda = npdas; 748 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 749 INIT_READ_NODE(rrdNodes + i, "rrd"); 750 DISK_NODE_PARAMS(rrdNodes[i], pda); 751 } 752 753 /* read redundancy pdas */ 754 pda = pqPDAs; 755 INIT_READ_NODE(rpNodes, "Rp"); 756 RF_ASSERT(pda); 757 DISK_NODE_PARAMS(rpNodes[0], pda); 758 pda++; 759 INIT_READ_NODE(rqNodes, redundantReadNodeName); 760 RF_ASSERT(pda); 761 DISK_NODE_PARAMS(rqNodes[0], pda); 762 if (nPQNodes == 2) { 763 pda++; 764 INIT_READ_NODE(rpNodes + 1, "Rp"); 765 RF_ASSERT(pda); 766 DISK_NODE_PARAMS(rpNodes[1], pda); 767 pda++; 768 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 769 RF_ASSERT(pda); 770 DISK_NODE_PARAMS(rqNodes[1], pda); 771 } 772 /* the recovery node has all reads as precedessors and all writes as 773 * successors. It generates a result for every write P or write Q 774 * node. As parameters, it takes a pda per read and a pda per stripe 775 * of user data written. It also takes as the last params the raidPtr 776 * and asm. For results, it takes PDA for P & Q. */ 777 778 779 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 780 nWriteNodes, /* succesors */ 781 nReadNodes, /* preds */ 782 nReadNodes + nWudNodes + 3, /* params */ 783 2 * nPQNodes, /* results */ 784 dag_h, recoveryNodeName, allocList); 785 786 787 788 for (i = 0; i < nReadNodes; i++) { 789 recoveryNode->antecedents[i] = rrdNodes + i; 790 recoveryNode->antType[i] = rf_control; 791 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 792 } 793 for (i = 0; i < nWudNodes; i++) { 794 recoveryNode->succedents[i] = wudNodes + i; 795 } 796 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 797 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 798 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 799 800 for (; i < nWriteNodes; i++) 801 recoveryNode->succedents[i] = wudNodes + i; 802 803 pda = pqPDAs; 804 recoveryNode->results[0] = pda; 805 pda++; 806 recoveryNode->results[1] = pda; 807 if (nPQNodes == 2) { 808 pda++; 809 recoveryNode->results[2] = pda; 810 pda++; 811 recoveryNode->results[3] = pda; 812 } 813 /* fill writes */ 814#define INIT_WRITE_NODE(node,name) \ 815 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 816 (node)->succedents[0] = unblockNode; \ 817 (node)->antecedents[0] = recoveryNode; \ 818 (node)->antType[0] = rf_control; 819 820 pda = asmap->physInfo; 821 for (i = 0; i < nWudNodes; i++) { 822 INIT_WRITE_NODE(wudNodes + i, "Wd"); 823 DISK_NODE_PARAMS(wudNodes[i], pda); 824 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 825 pda = pda->next; 826 } 827 /* write redundancy pdas */ 828 pda = pqPDAs; 829 INIT_WRITE_NODE(wpNodes, "Wp"); 830 RF_ASSERT(pda); 831 DISK_NODE_PARAMS(wpNodes[0], pda); 832 pda++; 833 INIT_WRITE_NODE(wqNodes, "Wq"); 834 RF_ASSERT(pda); 835 DISK_NODE_PARAMS(wqNodes[0], pda); 836 if (nPQNodes == 2) { 837 pda++; 838 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 839 RF_ASSERT(pda); 840 DISK_NODE_PARAMS(wpNodes[1], pda); 841 pda++; 842 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 843 RF_ASSERT(pda); 844 DISK_NODE_PARAMS(wqNodes[1], pda); 845 } 846} 847#endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 848