rf_dagdegwr.c revision 1.14
1/* $NetBSD: rf_dagdegwr.c,v 1.14 2003/12/30 21:59:03 oster Exp $ */ 2/* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: Mark Holland, Daniel Stodolsky, William V. Courtright II 7 * 8 * Permission to use, copy, modify and distribute this software and 9 * its documentation is hereby granted, provided that both the copyright 10 * notice and this permission notice appear in all copies of the 11 * software, derivative works or modified versions, and any portions 12 * thereof, and that both notices appear in supporting documentation. 13 * 14 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 15 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 16 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 17 * 18 * Carnegie Mellon requests users of this software to return to 19 * 20 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 21 * School of Computer Science 22 * Carnegie Mellon University 23 * Pittsburgh PA 15213-3890 24 * 25 * any improvements or extensions that they make and grant Carnegie the 26 * rights to redistribute these changes. 27 */ 28 29/* 30 * rf_dagdegwr.c 31 * 32 * code for creating degraded write DAGs 33 * 34 */ 35 36#include <sys/cdefs.h> 37__KERNEL_RCSID(0, "$NetBSD: rf_dagdegwr.c,v 1.14 2003/12/30 21:59:03 oster Exp $"); 38 39#include <dev/raidframe/raidframevar.h> 40 41#include "rf_raid.h" 42#include "rf_dag.h" 43#include "rf_dagutils.h" 44#include "rf_dagfuncs.h" 45#include "rf_debugMem.h" 46#include "rf_general.h" 47#include "rf_dagdegwr.h" 48 49 50/****************************************************************************** 51 * 52 * General comments on DAG creation: 53 * 54 * All DAGs in this file use roll-away error recovery. Each DAG has a single 55 * commit node, usually called "Cmt." If an error occurs before the Cmt node 56 * is reached, the execution engine will halt forward execution and work 57 * backward through the graph, executing the undo functions. Assuming that 58 * each node in the graph prior to the Cmt node are undoable and atomic - or - 59 * does not make changes to permanent state, the graph will fail atomically. 60 * If an error occurs after the Cmt node executes, the engine will roll-forward 61 * through the graph, blindly executing nodes until it reaches the end. 62 * If a graph reaches the end, it is assumed to have completed successfully. 63 * 64 * A graph has only 1 Cmt node. 65 * 66 */ 67 68 69/****************************************************************************** 70 * 71 * The following wrappers map the standard DAG creation interface to the 72 * DAG creation routines. Additionally, these wrappers enable experimentation 73 * with new DAG structures by providing an extra level of indirection, allowing 74 * the DAG creation routines to be replaced at this single point. 75 */ 76 77static 78RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG) 79{ 80 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, 81 flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE); 82} 83 84void 85rf_CreateDegradedWriteDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 86 RF_DagHeader_t *dag_h, 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(RF_Raid_t *raidPtr, 155 RF_AccessStripeMap_t *asmap, 156 RF_DagHeader_t *dag_h, void *bp, 157 RF_RaidAccessFlags_t flags, 158 RF_AllocListElem_t *allocList, 159 int nfaults, 160 int (*redFunc) (RF_DagNode_t *), 161 int allowBufferRecycle) 162{ 163 int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 164 rdnodesFaked; 165 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode; 166 RF_DagNode_t *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode; 167 RF_SectorCount_t sectorsPerSU; 168 RF_ReconUnitNum_t which_ru; 169 char *xorTargetBuf = NULL; /* the target buffer for the XOR 170 * operation */ 171 char *overlappingPDAs;/* a temporary array of flags */ 172 RF_AccessStripeMapHeader_t *new_asm_h[2]; 173 RF_PhysDiskAddr_t *pda, *parityPDA; 174 RF_StripeNum_t parityStripeID; 175 RF_PhysDiskAddr_t *failedPDA; 176 RF_RaidLayout_t *layoutPtr; 177 178 layoutPtr = &(raidPtr->Layout); 179 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 180 &which_ru); 181 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 182 /* failedPDA points to the pda within the asm that targets the failed 183 * disk */ 184 failedPDA = asmap->failedPDAs[0]; 185 186 if (rf_dagDebug) 187 printf("[Creating degraded-write DAG]\n"); 188 189 RF_ASSERT(asmap->numDataFailed == 1); 190 dag_h->creator = "SimpleDegradedWriteDAG"; 191 192 /* 193 * Generate two ASMs identifying the surviving data 194 * we need in order to recover the lost data. 195 */ 196 /* overlappingPDAs array must be zero'd */ 197 RF_Malloc(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char), (char *)); 198 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 199 &nXorBufs, NULL, overlappingPDAs, allocList); 200 201 /* create all the nodes at once */ 202 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 203 * generated for the 204 * failed pda */ 205 206 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 207 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 208 /* 209 * XXX 210 * 211 * There's a bug with a complete stripe overwrite- that means 0 reads 212 * of old data, and the rest of the DAG generation code doesn't like 213 * that. A release is coming, and I don't wanna risk breaking a critical 214 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 215 * I'm gonna fake there being a read node, and I'm gonna swap in a 216 * no-op node in its place (to make all the link-up code happy). 217 * This should be fixed at some point. --jimz 218 */ 219 if (nRrdNodes == 0) { 220 nRrdNodes = 1; 221 rdnodesFaked = 1; 222 } else { 223 rdnodesFaked = 0; 224 } 225 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 226 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 227 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), 228 (RF_DagNode_t *), allocList); 229 i = 0; 230 blockNode = &nodes[i]; 231 i += 1; 232 commitNode = &nodes[i]; 233 i += 1; 234 unblockNode = &nodes[i]; 235 i += 1; 236 termNode = &nodes[i]; 237 i += 1; 238 xorNode = &nodes[i]; 239 i += 1; 240 wnpNode = &nodes[i]; 241 i += 1; 242 wndNodes = &nodes[i]; 243 i += nWndNodes; 244 rrdNodes = &nodes[i]; 245 i += nRrdNodes; 246 if (nfaults == 2) { 247 wnqNode = &nodes[i]; 248 i += 1; 249 } else { 250 wnqNode = NULL; 251 } 252 RF_ASSERT(i == nNodes); 253 254 /* this dag can not commit until all rrd and xor Nodes have completed */ 255 dag_h->numCommitNodes = 1; 256 dag_h->numCommits = 0; 257 dag_h->numSuccedents = 1; 258 259 RF_ASSERT(nRrdNodes > 0); 260 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 261 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 262 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 263 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 264 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 265 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 266 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 267 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 268 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 269 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 270 271 /* 272 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 273 * the failed buffer, save a pointer to it so we can use it as the target 274 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 275 * a buffer is the same size as the failed buffer, it must also be at the 276 * same alignment within the SU. 277 */ 278 i = 0; 279 if (new_asm_h[0]) { 280 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 281 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 282 i++, pda = pda->next) { 283 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 284 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 285 RF_ASSERT(pda); 286 rrdNodes[i].params[0].p = pda; 287 rrdNodes[i].params[1].p = pda->bufPtr; 288 rrdNodes[i].params[2].v = parityStripeID; 289 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 290 } 291 } 292 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 293 if (new_asm_h[1]) { 294 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 295 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 296 j++, pda = pda->next) { 297 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 298 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 299 RF_ASSERT(pda); 300 rrdNodes[i + j].params[0].p = pda; 301 rrdNodes[i + j].params[1].p = pda->bufPtr; 302 rrdNodes[i + j].params[2].v = parityStripeID; 303 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 304 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 305 xorTargetBuf = pda->bufPtr; 306 } 307 } 308 if (rdnodesFaked) { 309 /* 310 * This is where we'll init that fake noop read node 311 * (XXX should the wakeup func be different?) 312 */ 313 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 314 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 315 } 316 /* 317 * Make a PDA for the parity unit. The parity PDA should start at 318 * the same offset into the SU as the failed PDA. 319 */ 320 /* Danner comment: I don't think this copy is really necessary. We are 321 * in one of two cases here. (1) The entire failed unit is written. 322 * Then asmap->parityInfo will describe the entire parity. (2) We are 323 * only writing a subset of the failed unit and nothing else. Then the 324 * asmap->parityInfo describes the failed unit and the copy can also 325 * be avoided. */ 326 327 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 328 parityPDA->col = asmap->parityInfo->col; 329 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 330 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 331 parityPDA->numSector = failedPDA->numSector; 332 333 if (!xorTargetBuf) { 334 RF_MallocAndAdd(xorTargetBuf, 335 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 336 } 337 /* init the Wnp node */ 338 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 339 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 340 wnpNode->params[0].p = parityPDA; 341 wnpNode->params[1].p = xorTargetBuf; 342 wnpNode->params[2].v = parityStripeID; 343 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 344 345 /* fill in the Wnq Node */ 346 if (nfaults == 2) { 347 { 348 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 349 (RF_PhysDiskAddr_t *), allocList); 350 parityPDA->col = asmap->qInfo->col; 351 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 352 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 353 parityPDA->numSector = failedPDA->numSector; 354 355 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 356 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 357 wnqNode->params[0].p = parityPDA; 358 RF_MallocAndAdd(xorNode->results[1], 359 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 360 wnqNode->params[1].p = xorNode->results[1]; 361 wnqNode->params[2].v = parityStripeID; 362 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 363 } 364 } 365 /* fill in the Wnd nodes */ 366 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 367 if (pda == failedPDA) { 368 i--; 369 continue; 370 } 371 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 372 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 373 RF_ASSERT(pda); 374 wndNodes[i].params[0].p = pda; 375 wndNodes[i].params[1].p = pda->bufPtr; 376 wndNodes[i].params[2].v = parityStripeID; 377 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 378 } 379 380 /* fill in the results of the xor node */ 381 xorNode->results[0] = xorTargetBuf; 382 383 /* fill in the params of the xor node */ 384 385 paramNum = 0; 386 if (rdnodesFaked == 0) { 387 for (i = 0; i < nRrdNodes; i++) { 388 /* all the Rrd nodes need to be xored together */ 389 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 390 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 391 } 392 } 393 for (i = 0; i < nWndNodes; i++) { 394 /* any Wnd nodes that overlap the failed access need to be 395 * xored in */ 396 if (overlappingPDAs[i]) { 397 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 398 memcpy((char *) pda, (char *) wndNodes[i].params[0].p, sizeof(RF_PhysDiskAddr_t)); 399 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 400 xorNode->params[paramNum++].p = pda; 401 xorNode->params[paramNum++].p = pda->bufPtr; 402 } 403 } 404 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 405 406 /* 407 * Install the failed PDA into the xor param list so that the 408 * new data gets xor'd in. 409 */ 410 xorNode->params[paramNum++].p = failedPDA; 411 xorNode->params[paramNum++].p = failedPDA->bufPtr; 412 413 /* 414 * The last 2 params to the recovery xor node are always the failed 415 * PDA and the raidPtr. install the failedPDA even though we have just 416 * done so above. This allows us to use the same XOR function for both 417 * degraded reads and degraded writes. 418 */ 419 xorNode->params[paramNum++].p = failedPDA; 420 xorNode->params[paramNum++].p = raidPtr; 421 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 422 423 /* 424 * Code to link nodes begins here 425 */ 426 427 /* link header to block node */ 428 RF_ASSERT(blockNode->numAntecedents == 0); 429 dag_h->succedents[0] = blockNode; 430 431 /* link block node to rd nodes */ 432 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 433 for (i = 0; i < nRrdNodes; i++) { 434 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 435 blockNode->succedents[i] = &rrdNodes[i]; 436 rrdNodes[i].antecedents[0] = blockNode; 437 rrdNodes[i].antType[0] = rf_control; 438 } 439 440 /* link read nodes to xor node */ 441 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 442 for (i = 0; i < nRrdNodes; i++) { 443 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 444 rrdNodes[i].succedents[0] = xorNode; 445 xorNode->antecedents[i] = &rrdNodes[i]; 446 xorNode->antType[i] = rf_trueData; 447 } 448 449 /* link xor node to commit node */ 450 RF_ASSERT(xorNode->numSuccedents == 1); 451 RF_ASSERT(commitNode->numAntecedents == 1); 452 xorNode->succedents[0] = commitNode; 453 commitNode->antecedents[0] = xorNode; 454 commitNode->antType[0] = rf_control; 455 456 /* link commit node to wnd nodes */ 457 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 458 for (i = 0; i < nWndNodes; i++) { 459 RF_ASSERT(wndNodes[i].numAntecedents == 1); 460 commitNode->succedents[i] = &wndNodes[i]; 461 wndNodes[i].antecedents[0] = commitNode; 462 wndNodes[i].antType[0] = rf_control; 463 } 464 465 /* link the commit node to wnp, wnq nodes */ 466 RF_ASSERT(wnpNode->numAntecedents == 1); 467 commitNode->succedents[nWndNodes] = wnpNode; 468 wnpNode->antecedents[0] = commitNode; 469 wnpNode->antType[0] = rf_control; 470 if (nfaults == 2) { 471 RF_ASSERT(wnqNode->numAntecedents == 1); 472 commitNode->succedents[nWndNodes + 1] = wnqNode; 473 wnqNode->antecedents[0] = commitNode; 474 wnqNode->antType[0] = rf_control; 475 } 476 /* link write new data nodes to unblock node */ 477 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 478 for (i = 0; i < nWndNodes; i++) { 479 RF_ASSERT(wndNodes[i].numSuccedents == 1); 480 wndNodes[i].succedents[0] = unblockNode; 481 unblockNode->antecedents[i] = &wndNodes[i]; 482 unblockNode->antType[i] = rf_control; 483 } 484 485 /* link write new parity node to unblock node */ 486 RF_ASSERT(wnpNode->numSuccedents == 1); 487 wnpNode->succedents[0] = unblockNode; 488 unblockNode->antecedents[nWndNodes] = wnpNode; 489 unblockNode->antType[nWndNodes] = rf_control; 490 491 /* link write new q node to unblock node */ 492 if (nfaults == 2) { 493 RF_ASSERT(wnqNode->numSuccedents == 1); 494 wnqNode->succedents[0] = unblockNode; 495 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 496 unblockNode->antType[nWndNodes + 1] = rf_control; 497 } 498 /* link unblock node to term node */ 499 RF_ASSERT(unblockNode->numSuccedents == 1); 500 RF_ASSERT(termNode->numAntecedents == 1); 501 RF_ASSERT(termNode->numSuccedents == 0); 502 unblockNode->succedents[0] = termNode; 503 termNode->antecedents[0] = unblockNode; 504 termNode->antType[0] = rf_control; 505} 506#define CONS_PDA(if,start,num) \ 507 pda_p->col = asmap->if->col; \ 508 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 509 pda_p->numSector = num; \ 510 pda_p->next = NULL; \ 511 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 512#if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 513void 514rf_WriteGenerateFailedAccessASMs( 515 RF_Raid_t * raidPtr, 516 RF_AccessStripeMap_t * asmap, 517 RF_PhysDiskAddr_t ** pdap, 518 int *nNodep, 519 RF_PhysDiskAddr_t ** pqpdap, 520 int *nPQNodep, 521 RF_AllocListElem_t * allocList) 522{ 523 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 524 int PDAPerDisk, i; 525 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 526 int numDataCol = layoutPtr->numDataCol; 527 int state; 528 unsigned napdas; 529 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 530 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 531 RF_PhysDiskAddr_t *pda_p; 532 RF_RaidAddr_t sosAddr; 533 534 /* determine how many pda's we will have to generate per unaccess 535 * stripe. If there is only one failed data unit, it is one; if two, 536 * possibly two, depending wether they overlap. */ 537 538 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 539 fone_end = fone_start + fone->numSector; 540 541 if (asmap->numDataFailed == 1) { 542 PDAPerDisk = 1; 543 state = 1; 544 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 545 pda_p = *pqpdap; 546 /* build p */ 547 CONS_PDA(parityInfo, fone_start, fone->numSector); 548 pda_p->type = RF_PDA_TYPE_PARITY; 549 pda_p++; 550 /* build q */ 551 CONS_PDA(qInfo, fone_start, fone->numSector); 552 pda_p->type = RF_PDA_TYPE_Q; 553 } else { 554 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 555 ftwo_end = ftwo_start + ftwo->numSector; 556 if (fone->numSector + ftwo->numSector > secPerSU) { 557 PDAPerDisk = 1; 558 state = 2; 559 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 560 pda_p = *pqpdap; 561 CONS_PDA(parityInfo, 0, secPerSU); 562 pda_p->type = RF_PDA_TYPE_PARITY; 563 pda_p++; 564 CONS_PDA(qInfo, 0, secPerSU); 565 pda_p->type = RF_PDA_TYPE_Q; 566 } else { 567 PDAPerDisk = 2; 568 state = 3; 569 /* four of them, fone, then ftwo */ 570 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 571 pda_p = *pqpdap; 572 CONS_PDA(parityInfo, fone_start, fone->numSector); 573 pda_p->type = RF_PDA_TYPE_PARITY; 574 pda_p++; 575 CONS_PDA(qInfo, fone_start, fone->numSector); 576 pda_p->type = RF_PDA_TYPE_Q; 577 pda_p++; 578 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 579 pda_p->type = RF_PDA_TYPE_PARITY; 580 pda_p++; 581 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 582 pda_p->type = RF_PDA_TYPE_Q; 583 } 584 } 585 /* figure out number of nonaccessed pda */ 586 napdas = PDAPerDisk * (numDataCol - 2); 587 *nPQNodep = PDAPerDisk; 588 589 *nNodep = napdas; 590 if (napdas == 0) 591 return; /* short circuit */ 592 593 /* allocate up our list of pda's */ 594 595 RF_MallocAndAdd(pda_p, napdas * sizeof(RF_PhysDiskAddr_t), 596 (RF_PhysDiskAddr_t *), allocList); 597 *pdap = pda_p; 598 599 /* linkem together */ 600 for (i = 0; i < (napdas - 1); i++) 601 pda_p[i].next = pda_p + (i + 1); 602 603 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 604 for (i = 0; i < numDataCol; i++) { 605 if ((pda_p - (*pdap)) == napdas) 606 continue; 607 pda_p->type = RF_PDA_TYPE_DATA; 608 pda_p->raidAddress = sosAddr + (i * secPerSU); 609 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 610 /* skip over dead disks */ 611 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->col].status)) 612 continue; 613 switch (state) { 614 case 1: /* fone */ 615 pda_p->numSector = fone->numSector; 616 pda_p->raidAddress += fone_start; 617 pda_p->startSector += fone_start; 618 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 619 break; 620 case 2: /* full stripe */ 621 pda_p->numSector = secPerSU; 622 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 623 break; 624 case 3: /* two slabs */ 625 pda_p->numSector = fone->numSector; 626 pda_p->raidAddress += fone_start; 627 pda_p->startSector += fone_start; 628 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 629 pda_p++; 630 pda_p->type = RF_PDA_TYPE_DATA; 631 pda_p->raidAddress = sosAddr + (i * secPerSU); 632 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 633 pda_p->numSector = ftwo->numSector; 634 pda_p->raidAddress += ftwo_start; 635 pda_p->startSector += ftwo_start; 636 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 637 break; 638 default: 639 RF_PANIC(); 640 } 641 pda_p++; 642 } 643 644 RF_ASSERT(pda_p - *pdap == napdas); 645 return; 646} 647#define DISK_NODE_PDA(node) ((node)->params[0].p) 648 649#define DISK_NODE_PARAMS(_node_,_p_) \ 650 (_node_).params[0].p = _p_ ; \ 651 (_node_).params[1].p = (_p_)->bufPtr; \ 652 (_node_).params[2].v = parityStripeID; \ 653 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru) 654 655void 656rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 657 RF_DagHeader_t *dag_h, void *bp, 658 RF_RaidAccessFlags_t flags, 659 RF_AllocListElem_t *allocList, 660 char *redundantReadNodeName, 661 char *redundantWriteNodeName, 662 char *recoveryNodeName, 663 int (*recovFunc) (RF_DagNode_t *)) 664{ 665 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 666 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 667 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 668 RF_PhysDiskAddr_t *pda, *pqPDAs; 669 RF_PhysDiskAddr_t *npdas; 670 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 671 RF_ReconUnitNum_t which_ru; 672 int nPQNodes; 673 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 674 675 /* simple small write case - First part looks like a reconstruct-read 676 * of the failed data units. Then a write of all data units not 677 * failed. */ 678 679 680 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 681 * / -------PQ----- / \ \ Wud Wp WQ \ | / 682 * --Unblock- | T 683 * 684 * Rrd = read recovery data (potentially none) Wud = write user data 685 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 686 * (could be two) 687 * 688 */ 689 690 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 691 692 RF_ASSERT(asmap->numDataFailed == 1); 693 694 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 695 nReadNodes = nRrdNodes + 2 * nPQNodes; 696 nWriteNodes = nWudNodes + 2 * nPQNodes; 697 nNodes = 4 + nReadNodes + nWriteNodes; 698 699 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 700 blockNode = nodes; 701 unblockNode = blockNode + 1; 702 termNode = unblockNode + 1; 703 recoveryNode = termNode + 1; 704 rrdNodes = recoveryNode + 1; 705 rpNodes = rrdNodes + nRrdNodes; 706 rqNodes = rpNodes + nPQNodes; 707 wudNodes = rqNodes + nPQNodes; 708 wpNodes = wudNodes + nWudNodes; 709 wqNodes = wpNodes + nPQNodes; 710 711 dag_h->creator = "PQ_DDSimpleSmallWrite"; 712 dag_h->numSuccedents = 1; 713 dag_h->succedents[0] = blockNode; 714 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 715 termNode->antecedents[0] = unblockNode; 716 termNode->antType[0] = rf_control; 717 718 /* init the block and unblock nodes */ 719 /* The block node has all the read nodes as successors */ 720 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 721 for (i = 0; i < nReadNodes; i++) 722 blockNode->succedents[i] = rrdNodes + i; 723 724 /* The unblock node has all the writes as successors */ 725 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 726 for (i = 0; i < nWriteNodes; i++) { 727 unblockNode->antecedents[i] = wudNodes + i; 728 unblockNode->antType[i] = rf_control; 729 } 730 unblockNode->succedents[0] = termNode; 731 732#define INIT_READ_NODE(node,name) \ 733 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 734 (node)->succedents[0] = recoveryNode; \ 735 (node)->antecedents[0] = blockNode; \ 736 (node)->antType[0] = rf_control; 737 738 /* build the read nodes */ 739 pda = npdas; 740 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 741 INIT_READ_NODE(rrdNodes + i, "rrd"); 742 DISK_NODE_PARAMS(rrdNodes[i], pda); 743 } 744 745 /* read redundancy pdas */ 746 pda = pqPDAs; 747 INIT_READ_NODE(rpNodes, "Rp"); 748 RF_ASSERT(pda); 749 DISK_NODE_PARAMS(rpNodes[0], pda); 750 pda++; 751 INIT_READ_NODE(rqNodes, redundantReadNodeName); 752 RF_ASSERT(pda); 753 DISK_NODE_PARAMS(rqNodes[0], pda); 754 if (nPQNodes == 2) { 755 pda++; 756 INIT_READ_NODE(rpNodes + 1, "Rp"); 757 RF_ASSERT(pda); 758 DISK_NODE_PARAMS(rpNodes[1], pda); 759 pda++; 760 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 761 RF_ASSERT(pda); 762 DISK_NODE_PARAMS(rqNodes[1], pda); 763 } 764 /* the recovery node has all reads as precedessors and all writes as 765 * successors. It generates a result for every write P or write Q 766 * node. As parameters, it takes a pda per read and a pda per stripe 767 * of user data written. It also takes as the last params the raidPtr 768 * and asm. For results, it takes PDA for P & Q. */ 769 770 771 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 772 nWriteNodes, /* succesors */ 773 nReadNodes, /* preds */ 774 nReadNodes + nWudNodes + 3, /* params */ 775 2 * nPQNodes, /* results */ 776 dag_h, recoveryNodeName, allocList); 777 778 779 780 for (i = 0; i < nReadNodes; i++) { 781 recoveryNode->antecedents[i] = rrdNodes + i; 782 recoveryNode->antType[i] = rf_control; 783 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 784 } 785 for (i = 0; i < nWudNodes; i++) { 786 recoveryNode->succedents[i] = wudNodes + i; 787 } 788 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 789 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 790 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 791 792 for (; i < nWriteNodes; i++) 793 recoveryNode->succedents[i] = wudNodes + i; 794 795 pda = pqPDAs; 796 recoveryNode->results[0] = pda; 797 pda++; 798 recoveryNode->results[1] = pda; 799 if (nPQNodes == 2) { 800 pda++; 801 recoveryNode->results[2] = pda; 802 pda++; 803 recoveryNode->results[3] = pda; 804 } 805 /* fill writes */ 806#define INIT_WRITE_NODE(node,name) \ 807 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 808 (node)->succedents[0] = unblockNode; \ 809 (node)->antecedents[0] = recoveryNode; \ 810 (node)->antType[0] = rf_control; 811 812 pda = asmap->physInfo; 813 for (i = 0; i < nWudNodes; i++) { 814 INIT_WRITE_NODE(wudNodes + i, "Wd"); 815 DISK_NODE_PARAMS(wudNodes[i], pda); 816 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 817 pda = pda->next; 818 } 819 /* write redundancy pdas */ 820 pda = pqPDAs; 821 INIT_WRITE_NODE(wpNodes, "Wp"); 822 RF_ASSERT(pda); 823 DISK_NODE_PARAMS(wpNodes[0], pda); 824 pda++; 825 INIT_WRITE_NODE(wqNodes, "Wq"); 826 RF_ASSERT(pda); 827 DISK_NODE_PARAMS(wqNodes[0], pda); 828 if (nPQNodes == 2) { 829 pda++; 830 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 831 RF_ASSERT(pda); 832 DISK_NODE_PARAMS(wpNodes[1], pda); 833 pda++; 834 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 835 RF_ASSERT(pda); 836 DISK_NODE_PARAMS(wqNodes[1], pda); 837 } 838} 839#endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 840