rf_dagdegwr.c revision 1.34
1/* $NetBSD: rf_dagdegwr.c,v 1.34 2019/02/09 03:34:00 christos 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 1.34 2019/02/09 03:34:00 christos 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#include "rf_map.h" 49 50 51/****************************************************************************** 52 * 53 * General comments on DAG creation: 54 * 55 * All DAGs in this file use roll-away error recovery. Each DAG has a single 56 * commit node, usually called "Cmt." If an error occurs before the Cmt node 57 * is reached, the execution engine will halt forward execution and work 58 * backward through the graph, executing the undo functions. Assuming that 59 * each node in the graph prior to the Cmt node are undoable and atomic - or - 60 * does not make changes to permanent state, the graph will fail atomically. 61 * If an error occurs after the Cmt node executes, the engine will roll-forward 62 * through the graph, blindly executing nodes until it reaches the end. 63 * If a graph reaches the end, it is assumed to have completed successfully. 64 * 65 * A graph has only 1 Cmt node. 66 * 67 */ 68 69 70/****************************************************************************** 71 * 72 * The following wrappers map the standard DAG creation interface to the 73 * DAG creation routines. Additionally, these wrappers enable experimentation 74 * with new DAG structures by providing an extra level of indirection, allowing 75 * the DAG creation routines to be replaced at this single point. 76 */ 77 78static 79RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG) 80{ 81 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, 82 flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE); 83} 84 85void 86rf_CreateDegradedWriteDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 87 RF_DagHeader_t *dag_h, void *bp, 88 RF_RaidAccessFlags_t flags, 89 RF_AllocListElem_t *allocList) 90{ 91 92 RF_ASSERT(asmap->numDataFailed == 1); 93 dag_h->creator = "DegradedWriteDAG"; 94 95 /* 96 * if the access writes only a portion of the failed unit, and also 97 * writes some portion of at least one surviving unit, we create two 98 * DAGs, one for the failed component and one for the non-failed 99 * component, and do them sequentially. Note that the fact that we're 100 * accessing only a portion of the failed unit indicates that the 101 * access either starts or ends in the failed unit, and hence we need 102 * create only two dags. This is inefficient in that the same data or 103 * parity can get read and written twice using this structure. I need 104 * to fix this to do the access all at once. 105 */ 106 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 && 107 asmap->failedPDAs[0]->numSector != 108 raidPtr->Layout.sectorsPerStripeUnit)); 109 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 110 allocList); 111} 112 113 114 115/****************************************************************************** 116 * 117 * DAG creation code begins here 118 */ 119#define BUF_ALLOC(num) \ 120 RF_MallocAndAdd(rf_RaidAddressToByte(raidPtr, num), allocList) 121 122 123 124/****************************************************************************** 125 * 126 * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode 127 * write, which is as follows 128 * 129 * / {Wnq} --\ 130 * hdr -> blockNode -> Rod -> Xor -> Cmt -> Wnp ----> unblock -> term 131 * \ {Rod} / \ Wnd ---/ 132 * \ {Wnd} -/ 133 * 134 * commit nodes: Xor, Wnd 135 * 136 * IMPORTANT: 137 * This DAG generator does not work for double-degraded archs since it does not 138 * generate Q 139 * 140 * This dag is essentially identical to the large-write dag, except that the 141 * write to the failed data unit is suppressed. 142 * 143 * IMPORTANT: this dag does not work in the case where the access writes only 144 * a portion of the failed unit, and also writes some portion of at least one 145 * surviving SU. this case is handled in CreateDegradedWriteDAG above. 146 * 147 * The block & unblock nodes are leftovers from a previous version. They 148 * do nothing, but I haven't deleted them because it would be a tremendous 149 * effort to put them back in. 150 * 151 * This dag is used whenever a one of the data units in a write has failed. 152 * If it is the parity unit that failed, the nonredundant write dag (below) 153 * is used. 154 *****************************************************************************/ 155 156void 157rf_CommonCreateSimpleDegradedWriteDAG(RF_Raid_t *raidPtr, 158 RF_AccessStripeMap_t *asmap, 159 RF_DagHeader_t *dag_h, void *bp, 160 RF_RaidAccessFlags_t flags, 161 RF_AllocListElem_t *allocList, 162 int nfaults, 163 int (*redFunc) (RF_DagNode_t *), 164 int allowBufferRecycle) 165{ 166 int nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 167 rdnodesFaked; 168 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *termNode; 169#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 170 RF_DagNode_t *wnqNode; 171#endif 172 RF_DagNode_t *wndNodes, *rrdNodes, *xorNode, *commitNode; 173 RF_DagNode_t *tmpNode, *tmpwndNode, *tmprrdNode; 174 RF_SectorCount_t sectorsPerSU; 175 RF_ReconUnitNum_t which_ru; 176 char *xorTargetBuf = NULL; /* the target buffer for the XOR 177 * operation */ 178 char overlappingPDAs[RF_MAXCOL];/* a temporary array of flags */ 179 RF_AccessStripeMapHeader_t *new_asm_h[2]; 180 RF_PhysDiskAddr_t *pda, *parityPDA; 181 RF_StripeNum_t parityStripeID; 182 RF_PhysDiskAddr_t *failedPDA; 183 RF_RaidLayout_t *layoutPtr; 184 185 layoutPtr = &(raidPtr->Layout); 186 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 187 &which_ru); 188 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 189 /* failedPDA points to the pda within the asm that targets the failed 190 * disk */ 191 failedPDA = asmap->failedPDAs[0]; 192 193#if RF_DEBUG_DAG 194 if (rf_dagDebug) 195 printf("[Creating degraded-write DAG]\n"); 196#endif 197 198 RF_ASSERT(asmap->numDataFailed == 1); 199 dag_h->creator = "SimpleDegradedWriteDAG"; 200 201 /* 202 * Generate two ASMs identifying the surviving data 203 * we need in order to recover the lost data. 204 */ 205 /* overlappingPDAs array must be zero'd */ 206 memset(overlappingPDAs, 0, RF_MAXCOL); 207 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 208 &nXorBufs, NULL, overlappingPDAs, allocList); 209 210 /* create all the nodes at once */ 211 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 212 * generated for the 213 * failed pda */ 214 215 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 216 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 217 /* 218 * XXX 219 * 220 * There's a bug with a complete stripe overwrite- that means 0 reads 221 * of old data, and the rest of the DAG generation code doesn't like 222 * that. A release is coming, and I don't wanna risk breaking a critical 223 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 224 * I'm gonna fake there being a read node, and I'm gonna swap in a 225 * no-op node in its place (to make all the link-up code happy). 226 * This should be fixed at some point. --jimz 227 */ 228 if (nRrdNodes == 0) { 229 nRrdNodes = 1; 230 rdnodesFaked = 1; 231 } else { 232 rdnodesFaked = 0; 233 } 234 235 blockNode = rf_AllocDAGNode(); 236 blockNode->list_next = dag_h->nodes; 237 dag_h->nodes = blockNode; 238 239 commitNode = rf_AllocDAGNode(); 240 commitNode->list_next = dag_h->nodes; 241 dag_h->nodes = commitNode; 242 243 unblockNode = rf_AllocDAGNode(); 244 unblockNode->list_next = dag_h->nodes; 245 dag_h->nodes = unblockNode; 246 247 termNode = rf_AllocDAGNode(); 248 termNode->list_next = dag_h->nodes; 249 dag_h->nodes = termNode; 250 251 xorNode = rf_AllocDAGNode(); 252 xorNode->list_next = dag_h->nodes; 253 dag_h->nodes = xorNode; 254 255 wnpNode = rf_AllocDAGNode(); 256 wnpNode->list_next = dag_h->nodes; 257 dag_h->nodes = wnpNode; 258 259 for (i = 0; i < nWndNodes; i++) { 260 tmpNode = rf_AllocDAGNode(); 261 tmpNode->list_next = dag_h->nodes; 262 dag_h->nodes = tmpNode; 263 } 264 wndNodes = dag_h->nodes; 265 266 for (i = 0; i < nRrdNodes; i++) { 267 tmpNode = rf_AllocDAGNode(); 268 tmpNode->list_next = dag_h->nodes; 269 dag_h->nodes = tmpNode; 270 } 271 rrdNodes = dag_h->nodes; 272 273#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 274 if (nfaults == 2) { 275 wnqNode = rf_AllocDAGNode(); 276 wnqNode->list_next = dag_h->nodes; 277 dag_h->nodes = wnqNode; 278 } else { 279 wnqNode = NULL; 280 } 281#endif 282 283 /* this dag can not commit until all rrd and xor Nodes have completed */ 284 dag_h->numCommitNodes = 1; 285 dag_h->numCommits = 0; 286 dag_h->numSuccedents = 1; 287 288 RF_ASSERT(nRrdNodes > 0); 289 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 290 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 291 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 292 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 293 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 294 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 295 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 296 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 297 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 298 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 299 300 /* 301 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 302 * the failed buffer, save a pointer to it so we can use it as the target 303 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 304 * a buffer is the same size as the failed buffer, it must also be at the 305 * same alignment within the SU. 306 */ 307 i = 0; 308 tmprrdNode = rrdNodes; 309 if (new_asm_h[0]) { 310 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 311 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 312 i++, pda = pda->next) { 313 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 314 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 315 RF_ASSERT(pda); 316 tmprrdNode->params[0].p = pda; 317 tmprrdNode->params[1].p = pda->bufPtr; 318 tmprrdNode->params[2].v = parityStripeID; 319 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 320 tmprrdNode = tmprrdNode->list_next; 321 } 322 } 323 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 324 /* Note that for tmprrdNode, this means a continuation from above, so no need to 325 assign it anything.. */ 326 if (new_asm_h[1]) { 327 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 328 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 329 j++, pda = pda->next) { 330 rf_InitNode(tmprrdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 331 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 332 RF_ASSERT(pda); 333 tmprrdNode->params[0].p = pda; 334 tmprrdNode->params[1].p = pda->bufPtr; 335 tmprrdNode->params[2].v = parityStripeID; 336 tmprrdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 337 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 338 xorTargetBuf = pda->bufPtr; 339 tmprrdNode = tmprrdNode->list_next; 340 } 341 } 342 if (rdnodesFaked) { 343 /* 344 * This is where we'll init that fake noop read node 345 * (XXX should the wakeup func be different?) 346 */ 347 /* node that rrdNodes will just be a single node... */ 348 rf_InitNode(rrdNodes, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 349 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 350 } 351 /* 352 * Make a PDA for the parity unit. The parity PDA should start at 353 * the same offset into the SU as the failed PDA. 354 */ 355 /* Danner comment: I don't think this copy is really necessary. We are 356 * in one of two cases here. (1) The entire failed unit is written. 357 * Then asmap->parityInfo will describe the entire parity. (2) We are 358 * only writing a subset of the failed unit and nothing else. Then the 359 * asmap->parityInfo describes the failed unit and the copy can also 360 * be avoided. */ 361 362 parityPDA = rf_AllocPhysDiskAddr(); 363 parityPDA->next = dag_h->pda_cleanup_list; 364 dag_h->pda_cleanup_list = parityPDA; 365 parityPDA->col = asmap->parityInfo->col; 366 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 367 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 368 parityPDA->numSector = failedPDA->numSector; 369 370 if (!xorTargetBuf) { 371 xorTargetBuf = rf_AllocBuffer(raidPtr, dag_h, rf_RaidAddressToByte(raidPtr, failedPDA->numSector)); 372 } 373 /* init the Wnp node */ 374 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 375 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 376 wnpNode->params[0].p = parityPDA; 377 wnpNode->params[1].p = xorTargetBuf; 378 wnpNode->params[2].v = parityStripeID; 379 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 380 381#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 382 /* fill in the Wnq Node */ 383 if (nfaults == 2) { 384 { 385 parityPA = RF_MallocAndAdd(sizeof(*parityPA), 386 allocList); 387 parityPDA->col = asmap->qInfo->col; 388 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 389 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 390 parityPDA->numSector = failedPDA->numSector; 391 392 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 393 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 394 wnqNode->params[0].p = parityPDA; 395 xorNode->results[1] = BUF_ALLOC(failedPDA->numSector); 396 wnqNode->params[1].p = xorNode->results[1]; 397 wnqNode->params[2].v = parityStripeID; 398 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 399 } 400 } 401#endif 402 /* fill in the Wnd nodes */ 403 tmpwndNode = wndNodes; 404 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 405 if (pda == failedPDA) { 406 i--; 407 continue; 408 } 409 rf_InitNode(tmpwndNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 410 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 411 RF_ASSERT(pda); 412 tmpwndNode->params[0].p = pda; 413 tmpwndNode->params[1].p = pda->bufPtr; 414 tmpwndNode->params[2].v = parityStripeID; 415 tmpwndNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 416 tmpwndNode = tmpwndNode->list_next; 417 } 418 419 /* fill in the results of the xor node */ 420 xorNode->results[0] = xorTargetBuf; 421 422 /* fill in the params of the xor node */ 423 424 paramNum = 0; 425 if (rdnodesFaked == 0) { 426 tmprrdNode = rrdNodes; 427 for (i = 0; i < nRrdNodes; i++) { 428 /* all the Rrd nodes need to be xored together */ 429 xorNode->params[paramNum++] = tmprrdNode->params[0]; 430 xorNode->params[paramNum++] = tmprrdNode->params[1]; 431 tmprrdNode = tmprrdNode->list_next; 432 } 433 } 434 tmpwndNode = wndNodes; 435 for (i = 0; i < nWndNodes; i++) { 436 /* any Wnd nodes that overlap the failed access need to be 437 * xored in */ 438 if (overlappingPDAs[i]) { 439 pda = rf_AllocPhysDiskAddr(); 440 memcpy((char *) pda, (char *) tmpwndNode->params[0].p, sizeof(RF_PhysDiskAddr_t)); 441 /* add it into the pda_cleanup_list *after* the copy, TYVM */ 442 pda->next = dag_h->pda_cleanup_list; 443 dag_h->pda_cleanup_list = pda; 444 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 445 xorNode->params[paramNum++].p = pda; 446 xorNode->params[paramNum++].p = pda->bufPtr; 447 } 448 tmpwndNode = tmpwndNode->list_next; 449 } 450 451 /* 452 * Install the failed PDA into the xor param list so that the 453 * new data gets xor'd in. 454 */ 455 xorNode->params[paramNum++].p = failedPDA; 456 xorNode->params[paramNum++].p = failedPDA->bufPtr; 457 458 /* 459 * The last 2 params to the recovery xor node are always the failed 460 * PDA and the raidPtr. install the failedPDA even though we have just 461 * done so above. This allows us to use the same XOR function for both 462 * degraded reads and degraded writes. 463 */ 464 xorNode->params[paramNum++].p = failedPDA; 465 xorNode->params[paramNum++].p = raidPtr; 466 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 467 468 /* 469 * Code to link nodes begins here 470 */ 471 472 /* link header to block node */ 473 RF_ASSERT(blockNode->numAntecedents == 0); 474 dag_h->succedents[0] = blockNode; 475 476 /* link block node to rd nodes */ 477 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 478 tmprrdNode = rrdNodes; 479 for (i = 0; i < nRrdNodes; i++) { 480 RF_ASSERT(tmprrdNode->numAntecedents == 1); 481 blockNode->succedents[i] = tmprrdNode; 482 tmprrdNode->antecedents[0] = blockNode; 483 tmprrdNode->antType[0] = rf_control; 484 tmprrdNode = tmprrdNode->list_next; 485 } 486 487 /* link read nodes to xor node */ 488 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 489 tmprrdNode = rrdNodes; 490 for (i = 0; i < nRrdNodes; i++) { 491 RF_ASSERT(tmprrdNode->numSuccedents == 1); 492 tmprrdNode->succedents[0] = xorNode; 493 xorNode->antecedents[i] = tmprrdNode; 494 xorNode->antType[i] = rf_trueData; 495 tmprrdNode = tmprrdNode->list_next; 496 } 497 498 /* link xor node to commit node */ 499 RF_ASSERT(xorNode->numSuccedents == 1); 500 RF_ASSERT(commitNode->numAntecedents == 1); 501 xorNode->succedents[0] = commitNode; 502 commitNode->antecedents[0] = xorNode; 503 commitNode->antType[0] = rf_control; 504 505 /* link commit node to wnd nodes */ 506 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 507 tmpwndNode = wndNodes; 508 for (i = 0; i < nWndNodes; i++) { 509 RF_ASSERT(tmpwndNode->numAntecedents == 1); 510 commitNode->succedents[i] = tmpwndNode; 511 tmpwndNode->antecedents[0] = commitNode; 512 tmpwndNode->antType[0] = rf_control; 513 tmpwndNode = tmpwndNode->list_next; 514 } 515 516 /* link the commit node to wnp, wnq nodes */ 517 RF_ASSERT(wnpNode->numAntecedents == 1); 518 commitNode->succedents[nWndNodes] = wnpNode; 519 wnpNode->antecedents[0] = commitNode; 520 wnpNode->antType[0] = rf_control; 521#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 522 if (nfaults == 2) { 523 RF_ASSERT(wnqNode->numAntecedents == 1); 524 commitNode->succedents[nWndNodes + 1] = wnqNode; 525 wnqNode->antecedents[0] = commitNode; 526 wnqNode->antType[0] = rf_control; 527 } 528#endif 529 /* link write new data nodes to unblock node */ 530 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 531 tmpwndNode = wndNodes; 532 for (i = 0; i < nWndNodes; i++) { 533 RF_ASSERT(tmpwndNode->numSuccedents == 1); 534 tmpwndNode->succedents[0] = unblockNode; 535 unblockNode->antecedents[i] = tmpwndNode; 536 unblockNode->antType[i] = rf_control; 537 tmpwndNode = tmpwndNode->list_next; 538 } 539 540 /* link write new parity node to unblock node */ 541 RF_ASSERT(wnpNode->numSuccedents == 1); 542 wnpNode->succedents[0] = unblockNode; 543 unblockNode->antecedents[nWndNodes] = wnpNode; 544 unblockNode->antType[nWndNodes] = rf_control; 545 546#if (RF_INCLUDE_DECL_PQ > 0) || (RF_INCLUDE_RAID6 > 0) 547 /* link write new q node to unblock node */ 548 if (nfaults == 2) { 549 RF_ASSERT(wnqNode->numSuccedents == 1); 550 wnqNode->succedents[0] = unblockNode; 551 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 552 unblockNode->antType[nWndNodes + 1] = rf_control; 553 } 554#endif 555 /* link unblock node to term node */ 556 RF_ASSERT(unblockNode->numSuccedents == 1); 557 RF_ASSERT(termNode->numAntecedents == 1); 558 RF_ASSERT(termNode->numSuccedents == 0); 559 unblockNode->succedents[0] = termNode; 560 termNode->antecedents[0] = unblockNode; 561 termNode->antType[0] = rf_control; 562} 563#define CONS_PDA(if,start,num) \ 564 pda_p->col = asmap->if->col; \ 565 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 566 pda_p->numSector = num; \ 567 pda_p->next = NULL; \ 568 pda_p->bufPtr = BUF_ALLOC(num) 569#if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 570void 571rf_WriteGenerateFailedAccessASMs( 572 RF_Raid_t * raidPtr, 573 RF_AccessStripeMap_t * asmap, 574 RF_PhysDiskAddr_t ** pdap, 575 int *nNodep, 576 RF_PhysDiskAddr_t ** pqpdap, 577 int *nPQNodep, 578 RF_AllocListElem_t * allocList) 579{ 580 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 581 int PDAPerDisk, i; 582 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 583 int numDataCol = layoutPtr->numDataCol; 584 int state; 585 unsigned napdas; 586 RF_SectorNum_t fone_start, ftwo_start = 0; 587 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 588 RF_PhysDiskAddr_t *pda_p; 589 RF_RaidAddr_t sosAddr; 590 591 /* determine how many pda's we will have to generate per unaccess 592 * stripe. If there is only one failed data unit, it is one; if two, 593 * possibly two, depending whether they overlap. */ 594 595 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 596 597 if (asmap->numDataFailed == 1) { 598 PDAPerDisk = 1; 599 state = 1; 600 *pqpdap = RF_MallocAndAdd(2 * sizeof(**pqpdap), allocList); 601 pda_p = *pqpdap; 602 /* build p */ 603 CONS_PDA(parityInfo, fone_start, fone->numSector); 604 pda_p->type = RF_PDA_TYPE_PARITY; 605 pda_p++; 606 /* build q */ 607 CONS_PDA(qInfo, fone_start, fone->numSector); 608 pda_p->type = RF_PDA_TYPE_Q; 609 } else { 610 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 611 if (fone->numSector + ftwo->numSector > secPerSU) { 612 PDAPerDisk = 1; 613 state = 2; 614 *pqpdap = RF_MallocAndAdd(2 * sizeof(**pqpdap), 615 allocList); 616 pda_p = *pqpdap; 617 CONS_PDA(parityInfo, 0, secPerSU); 618 pda_p->type = RF_PDA_TYPE_PARITY; 619 pda_p++; 620 CONS_PDA(qInfo, 0, secPerSU); 621 pda_p->type = RF_PDA_TYPE_Q; 622 } else { 623 PDAPerDisk = 2; 624 state = 3; 625 /* four of them, fone, then ftwo */ 626 *pqpdap = RF_MallocAndAdd(4 * sizeof(*pqpdap), 627 allocList); 628 pda_p = *pqpdap; 629 CONS_PDA(parityInfo, fone_start, fone->numSector); 630 pda_p->type = RF_PDA_TYPE_PARITY; 631 pda_p++; 632 CONS_PDA(qInfo, fone_start, fone->numSector); 633 pda_p->type = RF_PDA_TYPE_Q; 634 pda_p++; 635 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 636 pda_p->type = RF_PDA_TYPE_PARITY; 637 pda_p++; 638 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 639 pda_p->type = RF_PDA_TYPE_Q; 640 } 641 } 642 /* figure out number of nonaccessed pda */ 643 napdas = PDAPerDisk * (numDataCol - 2); 644 *nPQNodep = PDAPerDisk; 645 646 *nNodep = napdas; 647 if (napdas == 0) 648 return; /* short circuit */ 649 650 /* allocate up our list of pda's */ 651 652 pda_p = RF_MallocAndAdd(napdas * sizeof(*pda_p), allocList); 653 *pdap = pda_p; 654 655 /* linkem together */ 656 for (i = 0; i < (napdas - 1); i++) 657 pda_p[i].next = pda_p + (i + 1); 658 659 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 660 for (i = 0; i < numDataCol; i++) { 661 if ((pda_p - (*pdap)) == napdas) 662 continue; 663 pda_p->type = RF_PDA_TYPE_DATA; 664 pda_p->raidAddress = sosAddr + (i * secPerSU); 665 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 666 /* skip over dead disks */ 667 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->col].status)) 668 continue; 669 switch (state) { 670 case 1: /* fone */ 671 pda_p->numSector = fone->numSector; 672 pda_p->raidAddress += fone_start; 673 pda_p->startSector += fone_start; 674 pda_p->bufPtr = BUF_ALLOC(pda_p->numSector); 675 break; 676 case 2: /* full stripe */ 677 pda_p->numSector = secPerSU; 678 pda_p->bufPtr = BUF_ALLOC(secPerSU); 679 break; 680 case 3: /* two slabs */ 681 pda_p->numSector = fone->numSector; 682 pda_p->raidAddress += fone_start; 683 pda_p->startSector += fone_start; 684 pda_p->bufPtr = BUF_ALLOC(pda_p->numSector); 685 pda_p++; 686 pda_p->type = RF_PDA_TYPE_DATA; 687 pda_p->raidAddress = sosAddr + (i * secPerSU); 688 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->col), &(pda_p->startSector), 0); 689 pda_p->numSector = ftwo->numSector; 690 pda_p->raidAddress += ftwo_start; 691 pda_p->startSector += ftwo_start; 692 pda_p->bufPtr = BUF_ALLOC(pda_p->numSector); 693 break; 694 default: 695 RF_PANIC(); 696 } 697 pda_p++; 698 } 699 700 RF_ASSERT(pda_p - *pdap == napdas); 701 return; 702} 703#define DISK_NODE_PDA(node) ((node)->params[0].p) 704 705#define DISK_NODE_PARAMS(_node_,_p_) \ 706 (_node_).params[0].p = _p_ ; \ 707 (_node_).params[1].p = (_p_)->bufPtr; \ 708 (_node_).params[2].v = parityStripeID; \ 709 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru) 710 711void 712rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 713 RF_DagHeader_t *dag_h, void *bp, 714 RF_RaidAccessFlags_t flags, 715 RF_AllocListElem_t *allocList, 716 const char *redundantReadNodeName, 717 const char *redundantWriteNodeName, 718 const char *recoveryNodeName, 719 int (*recovFunc) (RF_DagNode_t *)) 720{ 721 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 722 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 723 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 724 RF_PhysDiskAddr_t *pda, *pqPDAs; 725 RF_PhysDiskAddr_t *npdas; 726 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 727 RF_ReconUnitNum_t which_ru; 728 int nPQNodes; 729 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 730 731 /* simple small write case - First part looks like a reconstruct-read 732 * of the failed data units. Then a write of all data units not 733 * failed. */ 734 735 736 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 737 * / -------PQ----- / \ \ Wud Wp WQ \ | / 738 * --Unblock- | T 739 * 740 * Rrd = read recovery data (potentially none) Wud = write user data 741 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 742 * (could be two) 743 * 744 */ 745 746 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 747 748 RF_ASSERT(asmap->numDataFailed == 1); 749 750 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 751 nReadNodes = nRrdNodes + 2 * nPQNodes; 752 nWriteNodes = nWudNodes + 2 * nPQNodes; 753 nNodes = 4 + nReadNodes + nWriteNodes; 754 755 nodes = RF_MallocAndAdd(nNodes * sizeof(*nodes), allocList); 756 blockNode = nodes; 757 unblockNode = blockNode + 1; 758 termNode = unblockNode + 1; 759 recoveryNode = termNode + 1; 760 rrdNodes = recoveryNode + 1; 761 rpNodes = rrdNodes + nRrdNodes; 762 rqNodes = rpNodes + nPQNodes; 763 wudNodes = rqNodes + nPQNodes; 764 wpNodes = wudNodes + nWudNodes; 765 wqNodes = wpNodes + nPQNodes; 766 767 dag_h->creator = "PQ_DDSimpleSmallWrite"; 768 dag_h->numSuccedents = 1; 769 dag_h->succedents[0] = blockNode; 770 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 771 termNode->antecedents[0] = unblockNode; 772 termNode->antType[0] = rf_control; 773 774 /* init the block and unblock nodes */ 775 /* The block node has all the read nodes as successors */ 776 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 777 for (i = 0; i < nReadNodes; i++) 778 blockNode->succedents[i] = rrdNodes + i; 779 780 /* The unblock node has all the writes as successors */ 781 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 782 for (i = 0; i < nWriteNodes; i++) { 783 unblockNode->antecedents[i] = wudNodes + i; 784 unblockNode->antType[i] = rf_control; 785 } 786 unblockNode->succedents[0] = termNode; 787 788#define INIT_READ_NODE(node,name) \ 789 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 790 (node)->succedents[0] = recoveryNode; \ 791 (node)->antecedents[0] = blockNode; \ 792 (node)->antType[0] = rf_control; 793 794 /* build the read nodes */ 795 pda = npdas; 796 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 797 INIT_READ_NODE(rrdNodes + i, "rrd"); 798 DISK_NODE_PARAMS(rrdNodes[i], pda); 799 } 800 801 /* read redundancy pdas */ 802 pda = pqPDAs; 803 INIT_READ_NODE(rpNodes, "Rp"); 804 RF_ASSERT(pda); 805 DISK_NODE_PARAMS(rpNodes[0], pda); 806 pda++; 807 INIT_READ_NODE(rqNodes, redundantReadNodeName); 808 RF_ASSERT(pda); 809 DISK_NODE_PARAMS(rqNodes[0], pda); 810 if (nPQNodes == 2) { 811 pda++; 812 INIT_READ_NODE(rpNodes + 1, "Rp"); 813 RF_ASSERT(pda); 814 DISK_NODE_PARAMS(rpNodes[1], pda); 815 pda++; 816 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 817 RF_ASSERT(pda); 818 DISK_NODE_PARAMS(rqNodes[1], pda); 819 } 820 /* the recovery node has all reads as precedessors and all writes as 821 * successors. It generates a result for every write P or write Q 822 * node. As parameters, it takes a pda per read and a pda per stripe 823 * of user data written. It also takes as the last params the raidPtr 824 * and asm. For results, it takes PDA for P & Q. */ 825 826 827 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 828 nWriteNodes, /* succesors */ 829 nReadNodes, /* preds */ 830 nReadNodes + nWudNodes + 3, /* params */ 831 2 * nPQNodes, /* results */ 832 dag_h, recoveryNodeName, allocList); 833 834 835 836 for (i = 0; i < nReadNodes; i++) { 837 recoveryNode->antecedents[i] = rrdNodes + i; 838 recoveryNode->antType[i] = rf_control; 839 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 840 } 841 for (i = 0; i < nWudNodes; i++) { 842 recoveryNode->succedents[i] = wudNodes + i; 843 } 844 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 845 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 846 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 847 848 for (; i < nWriteNodes; i++) 849 recoveryNode->succedents[i] = wudNodes + i; 850 851 pda = pqPDAs; 852 recoveryNode->results[0] = pda; 853 pda++; 854 recoveryNode->results[1] = pda; 855 if (nPQNodes == 2) { 856 pda++; 857 recoveryNode->results[2] = pda; 858 pda++; 859 recoveryNode->results[3] = pda; 860 } 861 /* fill writes */ 862#define INIT_WRITE_NODE(node,name) \ 863 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 864 (node)->succedents[0] = unblockNode; \ 865 (node)->antecedents[0] = recoveryNode; \ 866 (node)->antType[0] = rf_control; 867 868 pda = asmap->physInfo; 869 for (i = 0; i < nWudNodes; i++) { 870 INIT_WRITE_NODE(wudNodes + i, "Wd"); 871 DISK_NODE_PARAMS(wudNodes[i], pda); 872 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 873 pda = pda->next; 874 } 875 /* write redundancy pdas */ 876 pda = pqPDAs; 877 INIT_WRITE_NODE(wpNodes, "Wp"); 878 RF_ASSERT(pda); 879 DISK_NODE_PARAMS(wpNodes[0], pda); 880 pda++; 881 INIT_WRITE_NODE(wqNodes, "Wq"); 882 RF_ASSERT(pda); 883 DISK_NODE_PARAMS(wqNodes[0], pda); 884 if (nPQNodes == 2) { 885 pda++; 886 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 887 RF_ASSERT(pda); 888 DISK_NODE_PARAMS(wpNodes[1], pda); 889 pda++; 890 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 891 RF_ASSERT(pda); 892 DISK_NODE_PARAMS(wqNodes[1], pda); 893 } 894} 895#endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 896