rf_dagdegwr.c revision 1.4
1/* $NetBSD: rf_dagdegwr.c,v 1.4 1999/08/13 03:41:53 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 "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_threadid.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 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 92 RF_PhysDiskAddr_t *failedPDA = asmap->failedPDAs[0]; 93 94 RF_ASSERT(asmap->numDataFailed == 1); 95 dag_h->creator = "DegradedWriteDAG"; 96 97 /* if the access writes only a portion of the failed unit, and also 98 * writes some portion of at least one surviving unit, we create two 99 * DAGs, one for the failed component and one for the non-failed 100 * component, and do them sequentially. Note that the fact that we're 101 * accessing only a portion of the failed unit indicates that the 102 * access either starts or ends in the failed unit, and hence we need 103 * create only two dags. This is inefficient in that the same data or 104 * parity can get read and written twice using this structure. I need 105 * to fix this to do the access all at once. */ 106 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 && failedPDA->numSector != layoutPtr->sectorsPerStripeUnit)); 107 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList); 108} 109 110 111 112/****************************************************************************** 113 * 114 * DAG creation code begins here 115 */ 116 117 118 119/****************************************************************************** 120 * 121 * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode 122 * write, which is as follows 123 * 124 * / {Wnq} --\ 125 * hdr -> blockNode -> Rod -> Xor -> Cmt -> Wnp ----> unblock -> term 126 * \ {Rod} / \ Wnd ---/ 127 * \ {Wnd} -/ 128 * 129 * commit nodes: Xor, Wnd 130 * 131 * IMPORTANT: 132 * This DAG generator does not work for double-degraded archs since it does not 133 * generate Q 134 * 135 * This dag is essentially identical to the large-write dag, except that the 136 * write to the failed data unit is suppressed. 137 * 138 * IMPORTANT: this dag does not work in the case where the access writes only 139 * a portion of the failed unit, and also writes some portion of at least one 140 * surviving SU. this case is handled in CreateDegradedWriteDAG above. 141 * 142 * The block & unblock nodes are leftovers from a previous version. They 143 * do nothing, but I haven't deleted them because it would be a tremendous 144 * effort to put them back in. 145 * 146 * This dag is used whenever a one of the data units in a write has failed. 147 * If it is the parity unit that failed, the nonredundant write dag (below) 148 * is used. 149 *****************************************************************************/ 150 151void 152rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 153 allocList, nfaults, redFunc, allowBufferRecycle) 154 RF_Raid_t *raidPtr; 155 RF_AccessStripeMap_t *asmap; 156 RF_DagHeader_t *dag_h; 157 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 *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode; 168 RF_SectorCount_t sectorsPerSU; 169 RF_ReconUnitNum_t which_ru; 170 char *xorTargetBuf = NULL; /* the target buffer for the XOR 171 * operation */ 172 char *overlappingPDAs;/* a temporary array of flags */ 173 RF_AccessStripeMapHeader_t *new_asm_h[2]; 174 RF_PhysDiskAddr_t *pda, *parityPDA; 175 RF_StripeNum_t parityStripeID; 176 RF_PhysDiskAddr_t *failedPDA; 177 RF_RaidLayout_t *layoutPtr; 178 179 layoutPtr = &(raidPtr->Layout); 180 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 181 &which_ru); 182 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 183 /* failedPDA points to the pda within the asm that targets the failed 184 * disk */ 185 failedPDA = asmap->failedPDAs[0]; 186 187 if (rf_dagDebug) 188 printf("[Creating degraded-write DAG]\n"); 189 190 RF_ASSERT(asmap->numDataFailed == 1); 191 dag_h->creator = "SimpleDegradedWriteDAG"; 192 193 /* 194 * Generate two ASMs identifying the surviving data 195 * we need in order to recover the lost data. 196 */ 197 /* overlappingPDAs array must be zero'd */ 198 RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed, sizeof(char), (char *)); 199 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 200 &nXorBufs, NULL, overlappingPDAs, allocList); 201 202 /* create all the nodes at once */ 203 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 204 * generated for the 205 * failed pda */ 206 207 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 208 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 209 /* 210 * XXX 211 * 212 * There's a bug with a complete stripe overwrite- that means 0 reads 213 * of old data, and the rest of the DAG generation code doesn't like 214 * that. A release is coming, and I don't wanna risk breaking a critical 215 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 216 * I'm gonna fake there being a read node, and I'm gonna swap in a 217 * no-op node in its place (to make all the link-up code happy). 218 * This should be fixed at some point. --jimz 219 */ 220 if (nRrdNodes == 0) { 221 nRrdNodes = 1; 222 rdnodesFaked = 1; 223 } else { 224 rdnodesFaked = 0; 225 } 226 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 227 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 228 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), 229 (RF_DagNode_t *), allocList); 230 i = 0; 231 blockNode = &nodes[i]; 232 i += 1; 233 commitNode = &nodes[i]; 234 i += 1; 235 unblockNode = &nodes[i]; 236 i += 1; 237 termNode = &nodes[i]; 238 i += 1; 239 xorNode = &nodes[i]; 240 i += 1; 241 wnpNode = &nodes[i]; 242 i += 1; 243 wndNodes = &nodes[i]; 244 i += nWndNodes; 245 rrdNodes = &nodes[i]; 246 i += nRrdNodes; 247 if (nfaults == 2) { 248 wnqNode = &nodes[i]; 249 i += 1; 250 } else { 251 wnqNode = NULL; 252 } 253 RF_ASSERT(i == nNodes); 254 255 /* this dag can not commit until all rrd and xor Nodes have completed */ 256 dag_h->numCommitNodes = 1; 257 dag_h->numCommits = 0; 258 dag_h->numSuccedents = 1; 259 260 RF_ASSERT(nRrdNodes > 0); 261 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 262 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 263 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 264 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 265 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 266 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 267 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 268 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 269 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 270 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 271 272 /* 273 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 274 * the failed buffer, save a pointer to it so we can use it as the target 275 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 276 * a buffer is the same size as the failed buffer, it must also be at the 277 * same alignment within the SU. 278 */ 279 i = 0; 280 if (new_asm_h[0]) { 281 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 282 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 283 i++, pda = pda->next) { 284 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 285 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 286 RF_ASSERT(pda); 287 rrdNodes[i].params[0].p = pda; 288 rrdNodes[i].params[1].p = pda->bufPtr; 289 rrdNodes[i].params[2].v = parityStripeID; 290 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 291 } 292 } 293 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 294 if (new_asm_h[1]) { 295 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 296 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 297 j++, pda = pda->next) { 298 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 299 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 300 RF_ASSERT(pda); 301 rrdNodes[i + j].params[0].p = pda; 302 rrdNodes[i + j].params[1].p = pda->bufPtr; 303 rrdNodes[i + j].params[2].v = parityStripeID; 304 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 305 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 306 xorTargetBuf = pda->bufPtr; 307 } 308 } 309 if (rdnodesFaked) { 310 /* 311 * This is where we'll init that fake noop read node 312 * (XXX should the wakeup func be different?) 313 */ 314 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 315 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 316 } 317 /* 318 * Make a PDA for the parity unit. The parity PDA should start at 319 * the same offset into the SU as the failed PDA. 320 */ 321 /* Danner comment: I don't think this copy is really necessary. We are 322 * in one of two cases here. (1) The entire failed unit is written. 323 * Then asmap->parityInfo will describe the entire parity. (2) We are 324 * only writing a subset of the failed unit and nothing else. Then the 325 * asmap->parityInfo describes the failed unit and the copy can also 326 * be avoided. */ 327 328 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 329 parityPDA->row = asmap->parityInfo->row; 330 parityPDA->col = asmap->parityInfo->col; 331 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 332 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 333 parityPDA->numSector = failedPDA->numSector; 334 335 if (!xorTargetBuf) { 336 RF_CallocAndAdd(xorTargetBuf, 1, 337 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 338 } 339 /* init the Wnp node */ 340 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 341 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 342 wnpNode->params[0].p = parityPDA; 343 wnpNode->params[1].p = xorTargetBuf; 344 wnpNode->params[2].v = parityStripeID; 345 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 346 347 /* fill in the Wnq Node */ 348 if (nfaults == 2) { 349 { 350 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 351 (RF_PhysDiskAddr_t *), allocList); 352 parityPDA->row = asmap->qInfo->row; 353 parityPDA->col = asmap->qInfo->col; 354 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 355 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 356 parityPDA->numSector = failedPDA->numSector; 357 358 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 359 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 360 wnqNode->params[0].p = parityPDA; 361 RF_CallocAndAdd(xorNode->results[1], 1, 362 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 363 wnqNode->params[1].p = xorNode->results[1]; 364 wnqNode->params[2].v = parityStripeID; 365 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 366 } 367 } 368 /* fill in the Wnd nodes */ 369 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 370 if (pda == failedPDA) { 371 i--; 372 continue; 373 } 374 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 375 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 376 RF_ASSERT(pda); 377 wndNodes[i].params[0].p = pda; 378 wndNodes[i].params[1].p = pda->bufPtr; 379 wndNodes[i].params[2].v = parityStripeID; 380 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 381 } 382 383 /* fill in the results of the xor node */ 384 xorNode->results[0] = xorTargetBuf; 385 386 /* fill in the params of the xor node */ 387 388 paramNum = 0; 389 if (rdnodesFaked == 0) { 390 for (i = 0; i < nRrdNodes; i++) { 391 /* all the Rrd nodes need to be xored together */ 392 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 393 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 394 } 395 } 396 for (i = 0; i < nWndNodes; i++) { 397 /* any Wnd nodes that overlap the failed access need to be 398 * xored in */ 399 if (overlappingPDAs[i]) { 400 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 401 bcopy((char *) wndNodes[i].params[0].p, (char *) pda, sizeof(RF_PhysDiskAddr_t)); 402 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 403 xorNode->params[paramNum++].p = pda; 404 xorNode->params[paramNum++].p = pda->bufPtr; 405 } 406 } 407 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 408 409 /* 410 * Install the failed PDA into the xor param list so that the 411 * new data gets xor'd in. 412 */ 413 xorNode->params[paramNum++].p = failedPDA; 414 xorNode->params[paramNum++].p = failedPDA->bufPtr; 415 416 /* 417 * The last 2 params to the recovery xor node are always the failed 418 * PDA and the raidPtr. install the failedPDA even though we have just 419 * done so above. This allows us to use the same XOR function for both 420 * degraded reads and degraded writes. 421 */ 422 xorNode->params[paramNum++].p = failedPDA; 423 xorNode->params[paramNum++].p = raidPtr; 424 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 425 426 /* 427 * Code to link nodes begins here 428 */ 429 430 /* link header to block node */ 431 RF_ASSERT(blockNode->numAntecedents == 0); 432 dag_h->succedents[0] = blockNode; 433 434 /* link block node to rd nodes */ 435 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 436 for (i = 0; i < nRrdNodes; i++) { 437 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 438 blockNode->succedents[i] = &rrdNodes[i]; 439 rrdNodes[i].antecedents[0] = blockNode; 440 rrdNodes[i].antType[0] = rf_control; 441 } 442 443 /* link read nodes to xor node */ 444 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 445 for (i = 0; i < nRrdNodes; i++) { 446 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 447 rrdNodes[i].succedents[0] = xorNode; 448 xorNode->antecedents[i] = &rrdNodes[i]; 449 xorNode->antType[i] = rf_trueData; 450 } 451 452 /* link xor node to commit node */ 453 RF_ASSERT(xorNode->numSuccedents == 1); 454 RF_ASSERT(commitNode->numAntecedents == 1); 455 xorNode->succedents[0] = commitNode; 456 commitNode->antecedents[0] = xorNode; 457 commitNode->antType[0] = rf_control; 458 459 /* link commit node to wnd nodes */ 460 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 461 for (i = 0; i < nWndNodes; i++) { 462 RF_ASSERT(wndNodes[i].numAntecedents == 1); 463 commitNode->succedents[i] = &wndNodes[i]; 464 wndNodes[i].antecedents[0] = commitNode; 465 wndNodes[i].antType[0] = rf_control; 466 } 467 468 /* link the commit node to wnp, wnq nodes */ 469 RF_ASSERT(wnpNode->numAntecedents == 1); 470 commitNode->succedents[nWndNodes] = wnpNode; 471 wnpNode->antecedents[0] = commitNode; 472 wnpNode->antType[0] = rf_control; 473 if (nfaults == 2) { 474 RF_ASSERT(wnqNode->numAntecedents == 1); 475 commitNode->succedents[nWndNodes + 1] = wnqNode; 476 wnqNode->antecedents[0] = commitNode; 477 wnqNode->antType[0] = rf_control; 478 } 479 /* link write new data nodes to unblock node */ 480 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 481 for (i = 0; i < nWndNodes; i++) { 482 RF_ASSERT(wndNodes[i].numSuccedents == 1); 483 wndNodes[i].succedents[0] = unblockNode; 484 unblockNode->antecedents[i] = &wndNodes[i]; 485 unblockNode->antType[i] = rf_control; 486 } 487 488 /* link write new parity node to unblock node */ 489 RF_ASSERT(wnpNode->numSuccedents == 1); 490 wnpNode->succedents[0] = unblockNode; 491 unblockNode->antecedents[nWndNodes] = wnpNode; 492 unblockNode->antType[nWndNodes] = rf_control; 493 494 /* link write new q node to unblock node */ 495 if (nfaults == 2) { 496 RF_ASSERT(wnqNode->numSuccedents == 1); 497 wnqNode->succedents[0] = unblockNode; 498 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 499 unblockNode->antType[nWndNodes + 1] = rf_control; 500 } 501 /* link unblock node to term node */ 502 RF_ASSERT(unblockNode->numSuccedents == 1); 503 RF_ASSERT(termNode->numAntecedents == 1); 504 RF_ASSERT(termNode->numSuccedents == 0); 505 unblockNode->succedents[0] = termNode; 506 termNode->antecedents[0] = unblockNode; 507 termNode->antType[0] = rf_control; 508} 509#define CONS_PDA(if,start,num) \ 510 pda_p->row = asmap->if->row; pda_p->col = asmap->if->col; \ 511 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 512 pda_p->numSector = num; \ 513 pda_p->next = NULL; \ 514 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 515 516void 517rf_WriteGenerateFailedAccessASMs( 518 RF_Raid_t * raidPtr, 519 RF_AccessStripeMap_t * asmap, 520 RF_PhysDiskAddr_t ** pdap, 521 int *nNodep, 522 RF_PhysDiskAddr_t ** pqpdap, 523 int *nPQNodep, 524 RF_AllocListElem_t * allocList) 525{ 526 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 527 int PDAPerDisk, i; 528 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 529 int numDataCol = layoutPtr->numDataCol; 530 int state; 531 unsigned napdas; 532 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 533 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 534 RF_PhysDiskAddr_t *pda_p; 535 RF_RaidAddr_t sosAddr; 536 537 /* determine how many pda's we will have to generate per unaccess 538 * stripe. If there is only one failed data unit, it is one; if two, 539 * possibly two, depending wether they overlap. */ 540 541 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 542 fone_end = fone_start + fone->numSector; 543 544 if (asmap->numDataFailed == 1) { 545 PDAPerDisk = 1; 546 state = 1; 547 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 548 pda_p = *pqpdap; 549 /* build p */ 550 CONS_PDA(parityInfo, fone_start, fone->numSector); 551 pda_p->type = RF_PDA_TYPE_PARITY; 552 pda_p++; 553 /* build q */ 554 CONS_PDA(qInfo, fone_start, fone->numSector); 555 pda_p->type = RF_PDA_TYPE_Q; 556 } else { 557 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 558 ftwo_end = ftwo_start + ftwo->numSector; 559 if (fone->numSector + ftwo->numSector > secPerSU) { 560 PDAPerDisk = 1; 561 state = 2; 562 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 563 pda_p = *pqpdap; 564 CONS_PDA(parityInfo, 0, secPerSU); 565 pda_p->type = RF_PDA_TYPE_PARITY; 566 pda_p++; 567 CONS_PDA(qInfo, 0, secPerSU); 568 pda_p->type = RF_PDA_TYPE_Q; 569 } else { 570 PDAPerDisk = 2; 571 state = 3; 572 /* four of them, fone, then ftwo */ 573 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 574 pda_p = *pqpdap; 575 CONS_PDA(parityInfo, fone_start, fone->numSector); 576 pda_p->type = RF_PDA_TYPE_PARITY; 577 pda_p++; 578 CONS_PDA(qInfo, fone_start, fone->numSector); 579 pda_p->type = RF_PDA_TYPE_Q; 580 pda_p++; 581 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 582 pda_p->type = RF_PDA_TYPE_PARITY; 583 pda_p++; 584 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 585 pda_p->type = RF_PDA_TYPE_Q; 586 } 587 } 588 /* figure out number of nonaccessed pda */ 589 napdas = PDAPerDisk * (numDataCol - 2); 590 *nPQNodep = PDAPerDisk; 591 592 *nNodep = napdas; 593 if (napdas == 0) 594 return; /* short circuit */ 595 596 /* allocate up our list of pda's */ 597 598 RF_CallocAndAdd(pda_p, napdas, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 599 *pdap = pda_p; 600 601 /* linkem together */ 602 for (i = 0; i < (napdas - 1); i++) 603 pda_p[i].next = pda_p + (i + 1); 604 605 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 606 for (i = 0; i < numDataCol; i++) { 607 if ((pda_p - (*pdap)) == napdas) 608 continue; 609 pda_p->type = RF_PDA_TYPE_DATA; 610 pda_p->raidAddress = sosAddr + (i * secPerSU); 611 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 612 /* skip over dead disks */ 613 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->row][pda_p->col].status)) 614 continue; 615 switch (state) { 616 case 1: /* fone */ 617 pda_p->numSector = fone->numSector; 618 pda_p->raidAddress += fone_start; 619 pda_p->startSector += fone_start; 620 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 621 break; 622 case 2: /* full stripe */ 623 pda_p->numSector = secPerSU; 624 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 625 break; 626 case 3: /* two slabs */ 627 pda_p->numSector = fone->numSector; 628 pda_p->raidAddress += fone_start; 629 pda_p->startSector += fone_start; 630 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 631 pda_p++; 632 pda_p->type = RF_PDA_TYPE_DATA; 633 pda_p->raidAddress = sosAddr + (i * secPerSU); 634 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 635 pda_p->numSector = ftwo->numSector; 636 pda_p->raidAddress += ftwo_start; 637 pda_p->startSector += ftwo_start; 638 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 639 break; 640 default: 641 RF_PANIC(); 642 } 643 pda_p++; 644 } 645 646 RF_ASSERT(pda_p - *pdap == napdas); 647 return; 648} 649#define DISK_NODE_PDA(node) ((node)->params[0].p) 650 651#define DISK_NODE_PARAMS(_node_,_p_) \ 652 (_node_).params[0].p = _p_ ; \ 653 (_node_).params[1].p = (_p_)->bufPtr; \ 654 (_node_).params[2].v = parityStripeID; \ 655 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru) 656 657void 658rf_DoubleDegSmallWrite( 659 RF_Raid_t * raidPtr, 660 RF_AccessStripeMap_t * asmap, 661 RF_DagHeader_t * dag_h, 662 void *bp, 663 RF_RaidAccessFlags_t flags, 664 RF_AllocListElem_t * allocList, 665 char *redundantReadNodeName, 666 char *redundantWriteNodeName, 667 char *recoveryNodeName, 668 int (*recovFunc) (RF_DagNode_t *)) 669{ 670 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 671 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 672 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 673 RF_PhysDiskAddr_t *pda, *pqPDAs; 674 RF_PhysDiskAddr_t *npdas; 675 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 676 RF_ReconUnitNum_t which_ru; 677 int nPQNodes; 678 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 679 680 /* simple small write case - First part looks like a reconstruct-read 681 * of the failed data units. Then a write of all data units not 682 * failed. */ 683 684 685 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 686 * / -------PQ----- / \ \ Wud Wp WQ \ | / 687 * --Unblock- | T 688 * 689 * Rrd = read recovery data (potentially none) Wud = write user data 690 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 691 * (could be two) 692 * 693 */ 694 695 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 696 697 RF_ASSERT(asmap->numDataFailed == 1); 698 699 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 700 nReadNodes = nRrdNodes + 2 * nPQNodes; 701 nWriteNodes = nWudNodes + 2 * nPQNodes; 702 nNodes = 4 + nReadNodes + nWriteNodes; 703 704 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 705 blockNode = nodes; 706 unblockNode = blockNode + 1; 707 termNode = unblockNode + 1; 708 recoveryNode = termNode + 1; 709 rrdNodes = recoveryNode + 1; 710 rpNodes = rrdNodes + nRrdNodes; 711 rqNodes = rpNodes + nPQNodes; 712 wudNodes = rqNodes + nPQNodes; 713 wpNodes = wudNodes + nWudNodes; 714 wqNodes = wpNodes + nPQNodes; 715 716 dag_h->creator = "PQ_DDSimpleSmallWrite"; 717 dag_h->numSuccedents = 1; 718 dag_h->succedents[0] = blockNode; 719 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 720 termNode->antecedents[0] = unblockNode; 721 termNode->antType[0] = rf_control; 722 723 /* init the block and unblock nodes */ 724 /* The block node has all the read nodes as successors */ 725 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 726 for (i = 0; i < nReadNodes; i++) 727 blockNode->succedents[i] = rrdNodes + i; 728 729 /* The unblock node has all the writes as successors */ 730 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 731 for (i = 0; i < nWriteNodes; i++) { 732 unblockNode->antecedents[i] = wudNodes + i; 733 unblockNode->antType[i] = rf_control; 734 } 735 unblockNode->succedents[0] = termNode; 736 737#define INIT_READ_NODE(node,name) \ 738 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 739 (node)->succedents[0] = recoveryNode; \ 740 (node)->antecedents[0] = blockNode; \ 741 (node)->antType[0] = rf_control; 742 743 /* build the read nodes */ 744 pda = npdas; 745 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 746 INIT_READ_NODE(rrdNodes + i, "rrd"); 747 DISK_NODE_PARAMS(rrdNodes[i], pda); 748 } 749 750 /* read redundancy pdas */ 751 pda = pqPDAs; 752 INIT_READ_NODE(rpNodes, "Rp"); 753 RF_ASSERT(pda); 754 DISK_NODE_PARAMS(rpNodes[0], pda); 755 pda++; 756 INIT_READ_NODE(rqNodes, redundantReadNodeName); 757 RF_ASSERT(pda); 758 DISK_NODE_PARAMS(rqNodes[0], pda); 759 if (nPQNodes == 2) { 760 pda++; 761 INIT_READ_NODE(rpNodes + 1, "Rp"); 762 RF_ASSERT(pda); 763 DISK_NODE_PARAMS(rpNodes[1], pda); 764 pda++; 765 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 766 RF_ASSERT(pda); 767 DISK_NODE_PARAMS(rqNodes[1], pda); 768 } 769 /* the recovery node has all reads as precedessors and all writes as 770 * successors. It generates a result for every write P or write Q 771 * node. As parameters, it takes a pda per read and a pda per stripe 772 * of user data written. It also takes as the last params the raidPtr 773 * and asm. For results, it takes PDA for P & Q. */ 774 775 776 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 777 nWriteNodes, /* succesors */ 778 nReadNodes, /* preds */ 779 nReadNodes + nWudNodes + 3, /* params */ 780 2 * nPQNodes, /* results */ 781 dag_h, recoveryNodeName, allocList); 782 783 784 785 for (i = 0; i < nReadNodes; i++) { 786 recoveryNode->antecedents[i] = rrdNodes + i; 787 recoveryNode->antType[i] = rf_control; 788 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 789 } 790 for (i = 0; i < nWudNodes; i++) { 791 recoveryNode->succedents[i] = wudNodes + i; 792 } 793 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 794 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 795 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 796 797 for (; i < nWriteNodes; i++) 798 recoveryNode->succedents[i] = wudNodes + i; 799 800 pda = pqPDAs; 801 recoveryNode->results[0] = pda; 802 pda++; 803 recoveryNode->results[1] = pda; 804 if (nPQNodes == 2) { 805 pda++; 806 recoveryNode->results[2] = pda; 807 pda++; 808 recoveryNode->results[3] = pda; 809 } 810 /* fill writes */ 811#define INIT_WRITE_NODE(node,name) \ 812 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 813 (node)->succedents[0] = unblockNode; \ 814 (node)->antecedents[0] = recoveryNode; \ 815 (node)->antType[0] = rf_control; 816 817 pda = asmap->physInfo; 818 for (i = 0; i < nWudNodes; i++) { 819 INIT_WRITE_NODE(wudNodes + i, "Wd"); 820 DISK_NODE_PARAMS(wudNodes[i], pda); 821 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 822 pda = pda->next; 823 } 824 /* write redundancy pdas */ 825 pda = pqPDAs; 826 INIT_WRITE_NODE(wpNodes, "Wp"); 827 RF_ASSERT(pda); 828 DISK_NODE_PARAMS(wpNodes[0], pda); 829 pda++; 830 INIT_WRITE_NODE(wqNodes, "Wq"); 831 RF_ASSERT(pda); 832 DISK_NODE_PARAMS(wqNodes[0], pda); 833 if (nPQNodes == 2) { 834 pda++; 835 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 836 RF_ASSERT(pda); 837 DISK_NODE_PARAMS(wpNodes[1], pda); 838 pda++; 839 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 840 RF_ASSERT(pda); 841 DISK_NODE_PARAMS(wqNodes[1], pda); 842 } 843} 844