rf_dagdegwr.c revision 1.16
1/* $NetBSD: rf_dagdegwr.c,v 1.16 2004/03/05 03:22:05 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 1.16 2004/03/05 03:22:05 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_DEBUG_DAG 187 if (rf_dagDebug) 188 printf("[Creating degraded-write DAG]\n"); 189#endif 190 191 RF_ASSERT(asmap->numDataFailed == 1); 192 dag_h->creator = "SimpleDegradedWriteDAG"; 193 194 /* 195 * Generate two ASMs identifying the surviving data 196 * we need in order to recover the lost data. 197 */ 198 /* overlappingPDAs array must be zero'd */ 199 RF_Malloc(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char), (char *)); 200 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 201 &nXorBufs, NULL, overlappingPDAs, allocList); 202 203 /* create all the nodes at once */ 204 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 205 * generated for the 206 * failed pda */ 207 208 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 209 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 210 /* 211 * XXX 212 * 213 * There's a bug with a complete stripe overwrite- that means 0 reads 214 * of old data, and the rest of the DAG generation code doesn't like 215 * that. A release is coming, and I don't wanna risk breaking a critical 216 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 217 * I'm gonna fake there being a read node, and I'm gonna swap in a 218 * no-op node in its place (to make all the link-up code happy). 219 * This should be fixed at some point. --jimz 220 */ 221 if (nRrdNodes == 0) { 222 nRrdNodes = 1; 223 rdnodesFaked = 1; 224 } else { 225 rdnodesFaked = 0; 226 } 227 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 228 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 229 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), 230 (RF_DagNode_t *), allocList); 231 i = 0; 232 blockNode = &nodes[i]; 233 i += 1; 234 commitNode = &nodes[i]; 235 i += 1; 236 unblockNode = &nodes[i]; 237 i += 1; 238 termNode = &nodes[i]; 239 i += 1; 240 xorNode = &nodes[i]; 241 i += 1; 242 wnpNode = &nodes[i]; 243 i += 1; 244 wndNodes = &nodes[i]; 245 i += nWndNodes; 246 rrdNodes = &nodes[i]; 247 i += nRrdNodes; 248 if (nfaults == 2) { 249 wnqNode = &nodes[i]; 250 i += 1; 251 } else { 252 wnqNode = NULL; 253 } 254 RF_ASSERT(i == nNodes); 255 256 /* this dag can not commit until all rrd and xor Nodes have completed */ 257 dag_h->numCommitNodes = 1; 258 dag_h->numCommits = 0; 259 dag_h->numSuccedents = 1; 260 261 RF_ASSERT(nRrdNodes > 0); 262 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 263 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 264 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 265 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 266 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 267 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 268 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 269 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 270 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 271 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 272 273 /* 274 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 275 * the failed buffer, save a pointer to it so we can use it as the target 276 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 277 * a buffer is the same size as the failed buffer, it must also be at the 278 * same alignment within the SU. 279 */ 280 i = 0; 281 if (new_asm_h[0]) { 282 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 283 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 284 i++, pda = pda->next) { 285 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 286 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 287 RF_ASSERT(pda); 288 rrdNodes[i].params[0].p = pda; 289 rrdNodes[i].params[1].p = pda->bufPtr; 290 rrdNodes[i].params[2].v = parityStripeID; 291 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 292 } 293 } 294 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 295 if (new_asm_h[1]) { 296 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 297 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 298 j++, pda = pda->next) { 299 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 300 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 301 RF_ASSERT(pda); 302 rrdNodes[i + j].params[0].p = pda; 303 rrdNodes[i + j].params[1].p = pda->bufPtr; 304 rrdNodes[i + j].params[2].v = parityStripeID; 305 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 306 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 307 xorTargetBuf = pda->bufPtr; 308 } 309 } 310 if (rdnodesFaked) { 311 /* 312 * This is where we'll init that fake noop read node 313 * (XXX should the wakeup func be different?) 314 */ 315 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 316 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 317 } 318 /* 319 * Make a PDA for the parity unit. The parity PDA should start at 320 * the same offset into the SU as the failed PDA. 321 */ 322 /* Danner comment: I don't think this copy is really necessary. We are 323 * in one of two cases here. (1) The entire failed unit is written. 324 * Then asmap->parityInfo will describe the entire parity. (2) We are 325 * only writing a subset of the failed unit and nothing else. Then the 326 * asmap->parityInfo describes the failed unit and the copy can also 327 * be avoided. */ 328 329 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 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_MallocAndAdd(xorTargetBuf, 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, 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->col = asmap->qInfo->col; 353 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 354 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 355 parityPDA->numSector = failedPDA->numSector; 356 357 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 358 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 359 wnqNode->params[0].p = parityPDA; 360 RF_MallocAndAdd(xorNode->results[1], 361 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 362 wnqNode->params[1].p = xorNode->results[1]; 363 wnqNode->params[2].v = parityStripeID; 364 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 365 } 366 } 367 /* fill in the Wnd nodes */ 368 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 369 if (pda == failedPDA) { 370 i--; 371 continue; 372 } 373 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 374 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 375 RF_ASSERT(pda); 376 wndNodes[i].params[0].p = pda; 377 wndNodes[i].params[1].p = pda->bufPtr; 378 wndNodes[i].params[2].v = parityStripeID; 379 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 380 } 381 382 /* fill in the results of the xor node */ 383 xorNode->results[0] = xorTargetBuf; 384 385 /* fill in the params of the xor node */ 386 387 paramNum = 0; 388 if (rdnodesFaked == 0) { 389 for (i = 0; i < nRrdNodes; i++) { 390 /* all the Rrd nodes need to be xored together */ 391 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 392 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 393 } 394 } 395 for (i = 0; i < nWndNodes; i++) { 396 /* any Wnd nodes that overlap the failed access need to be 397 * xored in */ 398 if (overlappingPDAs[i]) { 399 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 400 memcpy((char *) pda, (char *) wndNodes[i].params[0].p, sizeof(RF_PhysDiskAddr_t)); 401 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 402 xorNode->params[paramNum++].p = pda; 403 xorNode->params[paramNum++].p = pda->bufPtr; 404 } 405 } 406 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 407 408 /* 409 * Install the failed PDA into the xor param list so that the 410 * new data gets xor'd in. 411 */ 412 xorNode->params[paramNum++].p = failedPDA; 413 xorNode->params[paramNum++].p = failedPDA->bufPtr; 414 415 /* 416 * The last 2 params to the recovery xor node are always the failed 417 * PDA and the raidPtr. install the failedPDA even though we have just 418 * done so above. This allows us to use the same XOR function for both 419 * degraded reads and degraded writes. 420 */ 421 xorNode->params[paramNum++].p = failedPDA; 422 xorNode->params[paramNum++].p = raidPtr; 423 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 424 425 /* 426 * Code to link nodes begins here 427 */ 428 429 /* link header to block node */ 430 RF_ASSERT(blockNode->numAntecedents == 0); 431 dag_h->succedents[0] = blockNode; 432 433 /* link block node to rd nodes */ 434 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 435 for (i = 0; i < nRrdNodes; i++) { 436 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 437 blockNode->succedents[i] = &rrdNodes[i]; 438 rrdNodes[i].antecedents[0] = blockNode; 439 rrdNodes[i].antType[0] = rf_control; 440 } 441 442 /* link read nodes to xor node */ 443 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 444 for (i = 0; i < nRrdNodes; i++) { 445 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 446 rrdNodes[i].succedents[0] = xorNode; 447 xorNode->antecedents[i] = &rrdNodes[i]; 448 xorNode->antType[i] = rf_trueData; 449 } 450 451 /* link xor node to commit node */ 452 RF_ASSERT(xorNode->numSuccedents == 1); 453 RF_ASSERT(commitNode->numAntecedents == 1); 454 xorNode->succedents[0] = commitNode; 455 commitNode->antecedents[0] = xorNode; 456 commitNode->antType[0] = rf_control; 457 458 /* link commit node to wnd nodes */ 459 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 460 for (i = 0; i < nWndNodes; i++) { 461 RF_ASSERT(wndNodes[i].numAntecedents == 1); 462 commitNode->succedents[i] = &wndNodes[i]; 463 wndNodes[i].antecedents[0] = commitNode; 464 wndNodes[i].antType[0] = rf_control; 465 } 466 467 /* link the commit node to wnp, wnq nodes */ 468 RF_ASSERT(wnpNode->numAntecedents == 1); 469 commitNode->succedents[nWndNodes] = wnpNode; 470 wnpNode->antecedents[0] = commitNode; 471 wnpNode->antType[0] = rf_control; 472 if (nfaults == 2) { 473 RF_ASSERT(wnqNode->numAntecedents == 1); 474 commitNode->succedents[nWndNodes + 1] = wnqNode; 475 wnqNode->antecedents[0] = commitNode; 476 wnqNode->antType[0] = rf_control; 477 } 478 /* link write new data nodes to unblock node */ 479 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 480 for (i = 0; i < nWndNodes; i++) { 481 RF_ASSERT(wndNodes[i].numSuccedents == 1); 482 wndNodes[i].succedents[0] = unblockNode; 483 unblockNode->antecedents[i] = &wndNodes[i]; 484 unblockNode->antType[i] = rf_control; 485 } 486 487 /* link write new parity node to unblock node */ 488 RF_ASSERT(wnpNode->numSuccedents == 1); 489 wnpNode->succedents[0] = unblockNode; 490 unblockNode->antecedents[nWndNodes] = wnpNode; 491 unblockNode->antType[nWndNodes] = rf_control; 492 493 /* link write new q node to unblock node */ 494 if (nfaults == 2) { 495 RF_ASSERT(wnqNode->numSuccedents == 1); 496 wnqNode->succedents[0] = unblockNode; 497 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 498 unblockNode->antType[nWndNodes + 1] = rf_control; 499 } 500 /* link unblock node to term node */ 501 RF_ASSERT(unblockNode->numSuccedents == 1); 502 RF_ASSERT(termNode->numAntecedents == 1); 503 RF_ASSERT(termNode->numSuccedents == 0); 504 unblockNode->succedents[0] = termNode; 505 termNode->antecedents[0] = unblockNode; 506 termNode->antType[0] = rf_control; 507} 508#define CONS_PDA(if,start,num) \ 509 pda_p->col = asmap->if->col; \ 510 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 511 pda_p->numSector = num; \ 512 pda_p->next = NULL; \ 513 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 514#if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 515void 516rf_WriteGenerateFailedAccessASMs( 517 RF_Raid_t * raidPtr, 518 RF_AccessStripeMap_t * asmap, 519 RF_PhysDiskAddr_t ** pdap, 520 int *nNodep, 521 RF_PhysDiskAddr_t ** pqpdap, 522 int *nPQNodep, 523 RF_AllocListElem_t * allocList) 524{ 525 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 526 int PDAPerDisk, i; 527 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 528 int numDataCol = layoutPtr->numDataCol; 529 int state; 530 unsigned napdas; 531 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 532 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 533 RF_PhysDiskAddr_t *pda_p; 534 RF_RaidAddr_t sosAddr; 535 536 /* determine how many pda's we will have to generate per unaccess 537 * stripe. If there is only one failed data unit, it is one; if two, 538 * possibly two, depending wether they overlap. */ 539 540 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 541 fone_end = fone_start + fone->numSector; 542 543 if (asmap->numDataFailed == 1) { 544 PDAPerDisk = 1; 545 state = 1; 546 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 547 pda_p = *pqpdap; 548 /* build p */ 549 CONS_PDA(parityInfo, fone_start, fone->numSector); 550 pda_p->type = RF_PDA_TYPE_PARITY; 551 pda_p++; 552 /* build q */ 553 CONS_PDA(qInfo, fone_start, fone->numSector); 554 pda_p->type = RF_PDA_TYPE_Q; 555 } else { 556 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 557 ftwo_end = ftwo_start + ftwo->numSector; 558 if (fone->numSector + ftwo->numSector > secPerSU) { 559 PDAPerDisk = 1; 560 state = 2; 561 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 562 pda_p = *pqpdap; 563 CONS_PDA(parityInfo, 0, secPerSU); 564 pda_p->type = RF_PDA_TYPE_PARITY; 565 pda_p++; 566 CONS_PDA(qInfo, 0, secPerSU); 567 pda_p->type = RF_PDA_TYPE_Q; 568 } else { 569 PDAPerDisk = 2; 570 state = 3; 571 /* four of them, fone, then ftwo */ 572 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 573 pda_p = *pqpdap; 574 CONS_PDA(parityInfo, fone_start, fone->numSector); 575 pda_p->type = RF_PDA_TYPE_PARITY; 576 pda_p++; 577 CONS_PDA(qInfo, fone_start, fone->numSector); 578 pda_p->type = RF_PDA_TYPE_Q; 579 pda_p++; 580 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 581 pda_p->type = RF_PDA_TYPE_PARITY; 582 pda_p++; 583 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 584 pda_p->type = RF_PDA_TYPE_Q; 585 } 586 } 587 /* figure out number of nonaccessed pda */ 588 napdas = PDAPerDisk * (numDataCol - 2); 589 *nPQNodep = PDAPerDisk; 590 591 *nNodep = napdas; 592 if (napdas == 0) 593 return; /* short circuit */ 594 595 /* allocate up our list of pda's */ 596 597 RF_MallocAndAdd(pda_p, napdas * sizeof(RF_PhysDiskAddr_t), 598 (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->col), &(pda_p->startSector), 0); 612 /* skip over dead disks */ 613 if (RF_DEAD_DISK(raidPtr->Disks[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->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, which_ru) 656 657void 658rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 659 RF_DagHeader_t *dag_h, void *bp, 660 RF_RaidAccessFlags_t flags, 661 RF_AllocListElem_t *allocList, 662 char *redundantReadNodeName, 663 char *redundantWriteNodeName, 664 char *recoveryNodeName, 665 int (*recovFunc) (RF_DagNode_t *)) 666{ 667 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 668 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 669 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 670 RF_PhysDiskAddr_t *pda, *pqPDAs; 671 RF_PhysDiskAddr_t *npdas; 672 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 673 RF_ReconUnitNum_t which_ru; 674 int nPQNodes; 675 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 676 677 /* simple small write case - First part looks like a reconstruct-read 678 * of the failed data units. Then a write of all data units not 679 * failed. */ 680 681 682 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 683 * / -------PQ----- / \ \ Wud Wp WQ \ | / 684 * --Unblock- | T 685 * 686 * Rrd = read recovery data (potentially none) Wud = write user data 687 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 688 * (could be two) 689 * 690 */ 691 692 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 693 694 RF_ASSERT(asmap->numDataFailed == 1); 695 696 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 697 nReadNodes = nRrdNodes + 2 * nPQNodes; 698 nWriteNodes = nWudNodes + 2 * nPQNodes; 699 nNodes = 4 + nReadNodes + nWriteNodes; 700 701 RF_MallocAndAdd(nodes, nNodes * sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 702 blockNode = nodes; 703 unblockNode = blockNode + 1; 704 termNode = unblockNode + 1; 705 recoveryNode = termNode + 1; 706 rrdNodes = recoveryNode + 1; 707 rpNodes = rrdNodes + nRrdNodes; 708 rqNodes = rpNodes + nPQNodes; 709 wudNodes = rqNodes + nPQNodes; 710 wpNodes = wudNodes + nWudNodes; 711 wqNodes = wpNodes + nPQNodes; 712 713 dag_h->creator = "PQ_DDSimpleSmallWrite"; 714 dag_h->numSuccedents = 1; 715 dag_h->succedents[0] = blockNode; 716 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 717 termNode->antecedents[0] = unblockNode; 718 termNode->antType[0] = rf_control; 719 720 /* init the block and unblock nodes */ 721 /* The block node has all the read nodes as successors */ 722 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 723 for (i = 0; i < nReadNodes; i++) 724 blockNode->succedents[i] = rrdNodes + i; 725 726 /* The unblock node has all the writes as successors */ 727 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 728 for (i = 0; i < nWriteNodes; i++) { 729 unblockNode->antecedents[i] = wudNodes + i; 730 unblockNode->antType[i] = rf_control; 731 } 732 unblockNode->succedents[0] = termNode; 733 734#define INIT_READ_NODE(node,name) \ 735 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 736 (node)->succedents[0] = recoveryNode; \ 737 (node)->antecedents[0] = blockNode; \ 738 (node)->antType[0] = rf_control; 739 740 /* build the read nodes */ 741 pda = npdas; 742 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 743 INIT_READ_NODE(rrdNodes + i, "rrd"); 744 DISK_NODE_PARAMS(rrdNodes[i], pda); 745 } 746 747 /* read redundancy pdas */ 748 pda = pqPDAs; 749 INIT_READ_NODE(rpNodes, "Rp"); 750 RF_ASSERT(pda); 751 DISK_NODE_PARAMS(rpNodes[0], pda); 752 pda++; 753 INIT_READ_NODE(rqNodes, redundantReadNodeName); 754 RF_ASSERT(pda); 755 DISK_NODE_PARAMS(rqNodes[0], pda); 756 if (nPQNodes == 2) { 757 pda++; 758 INIT_READ_NODE(rpNodes + 1, "Rp"); 759 RF_ASSERT(pda); 760 DISK_NODE_PARAMS(rpNodes[1], pda); 761 pda++; 762 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 763 RF_ASSERT(pda); 764 DISK_NODE_PARAMS(rqNodes[1], pda); 765 } 766 /* the recovery node has all reads as precedessors and all writes as 767 * successors. It generates a result for every write P or write Q 768 * node. As parameters, it takes a pda per read and a pda per stripe 769 * of user data written. It also takes as the last params the raidPtr 770 * and asm. For results, it takes PDA for P & Q. */ 771 772 773 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 774 nWriteNodes, /* succesors */ 775 nReadNodes, /* preds */ 776 nReadNodes + nWudNodes + 3, /* params */ 777 2 * nPQNodes, /* results */ 778 dag_h, recoveryNodeName, allocList); 779 780 781 782 for (i = 0; i < nReadNodes; i++) { 783 recoveryNode->antecedents[i] = rrdNodes + i; 784 recoveryNode->antType[i] = rf_control; 785 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 786 } 787 for (i = 0; i < nWudNodes; i++) { 788 recoveryNode->succedents[i] = wudNodes + i; 789 } 790 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 791 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 792 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 793 794 for (; i < nWriteNodes; i++) 795 recoveryNode->succedents[i] = wudNodes + i; 796 797 pda = pqPDAs; 798 recoveryNode->results[0] = pda; 799 pda++; 800 recoveryNode->results[1] = pda; 801 if (nPQNodes == 2) { 802 pda++; 803 recoveryNode->results[2] = pda; 804 pda++; 805 recoveryNode->results[3] = pda; 806 } 807 /* fill writes */ 808#define INIT_WRITE_NODE(node,name) \ 809 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 810 (node)->succedents[0] = unblockNode; \ 811 (node)->antecedents[0] = recoveryNode; \ 812 (node)->antType[0] = rf_control; 813 814 pda = asmap->physInfo; 815 for (i = 0; i < nWudNodes; i++) { 816 INIT_WRITE_NODE(wudNodes + i, "Wd"); 817 DISK_NODE_PARAMS(wudNodes[i], pda); 818 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 819 pda = pda->next; 820 } 821 /* write redundancy pdas */ 822 pda = pqPDAs; 823 INIT_WRITE_NODE(wpNodes, "Wp"); 824 RF_ASSERT(pda); 825 DISK_NODE_PARAMS(wpNodes[0], pda); 826 pda++; 827 INIT_WRITE_NODE(wqNodes, "Wq"); 828 RF_ASSERT(pda); 829 DISK_NODE_PARAMS(wqNodes[0], pda); 830 if (nPQNodes == 2) { 831 pda++; 832 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 833 RF_ASSERT(pda); 834 DISK_NODE_PARAMS(wpNodes[1], pda); 835 pda++; 836 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 837 RF_ASSERT(pda); 838 DISK_NODE_PARAMS(wqNodes[1], pda); 839 } 840} 841#endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 842