1 /* $NetBSD: rf_dagdegwr.c,v 1.9 2001/11/13 07:11:13 lukem 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 1.9 2001/11/13 07:11:13 lukem 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_memchunk.h" 47 #include "rf_general.h" 48 #include "rf_dagdegwr.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 78 static 79 RF_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 85 void 86 rf_CreateDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList) 87 RF_Raid_t *raidPtr; 88 RF_AccessStripeMap_t *asmap; 89 RF_DagHeader_t *dag_h; 90 void *bp; 91 RF_RaidAccessFlags_t flags; 92 RF_AllocListElem_t *allocList; 93 { 94 95 RF_ASSERT(asmap->numDataFailed == 1); 96 dag_h->creator = "DegradedWriteDAG"; 97 98 /* 99 * if the access writes only a portion of the failed unit, and also 100 * writes some portion of at least one surviving unit, we create two 101 * DAGs, one for the failed component and one for the non-failed 102 * component, and do them sequentially. Note that the fact that we're 103 * accessing only a portion of the failed unit indicates that the 104 * access either starts or ends in the failed unit, and hence we need 105 * create only two dags. This is inefficient in that the same data or 106 * parity can get read and written twice using this structure. I need 107 * to fix this to do the access all at once. 108 */ 109 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 && 110 asmap->failedPDAs[0]->numSector != 111 raidPtr->Layout.sectorsPerStripeUnit)); 112 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 113 allocList); 114 } 115 116 117 118 /****************************************************************************** 119 * 120 * DAG creation code begins here 121 */ 122 123 124 125 /****************************************************************************** 126 * 127 * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode 128 * write, which is as follows 129 * 130 * / {Wnq} --\ 131 * hdr -> blockNode -> Rod -> Xor -> Cmt -> Wnp ----> unblock -> term 132 * \ {Rod} / \ Wnd ---/ 133 * \ {Wnd} -/ 134 * 135 * commit nodes: Xor, Wnd 136 * 137 * IMPORTANT: 138 * This DAG generator does not work for double-degraded archs since it does not 139 * generate Q 140 * 141 * This dag is essentially identical to the large-write dag, except that the 142 * write to the failed data unit is suppressed. 143 * 144 * IMPORTANT: this dag does not work in the case where the access writes only 145 * a portion of the failed unit, and also writes some portion of at least one 146 * surviving SU. this case is handled in CreateDegradedWriteDAG above. 147 * 148 * The block & unblock nodes are leftovers from a previous version. They 149 * do nothing, but I haven't deleted them because it would be a tremendous 150 * effort to put them back in. 151 * 152 * This dag is used whenever a one of the data units in a write has failed. 153 * If it is the parity unit that failed, the nonredundant write dag (below) 154 * is used. 155 *****************************************************************************/ 156 157 void 158 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 159 allocList, nfaults, redFunc, allowBufferRecycle) 160 RF_Raid_t *raidPtr; 161 RF_AccessStripeMap_t *asmap; 162 RF_DagHeader_t *dag_h; 163 void *bp; 164 RF_RaidAccessFlags_t flags; 165 RF_AllocListElem_t *allocList; 166 int nfaults; 167 int (*redFunc) (RF_DagNode_t *); 168 int allowBufferRecycle; 169 { 170 int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 171 rdnodesFaked; 172 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode; 173 RF_DagNode_t *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode; 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;/* 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_dagDebug) 194 printf("[Creating degraded-write DAG]\n"); 195 196 RF_ASSERT(asmap->numDataFailed == 1); 197 dag_h->creator = "SimpleDegradedWriteDAG"; 198 199 /* 200 * Generate two ASMs identifying the surviving data 201 * we need in order to recover the lost data. 202 */ 203 /* overlappingPDAs array must be zero'd */ 204 RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed, sizeof(char), (char *)); 205 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 206 &nXorBufs, NULL, overlappingPDAs, allocList); 207 208 /* create all the nodes at once */ 209 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 210 * generated for the 211 * failed pda */ 212 213 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 214 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 215 /* 216 * XXX 217 * 218 * There's a bug with a complete stripe overwrite- that means 0 reads 219 * of old data, and the rest of the DAG generation code doesn't like 220 * that. A release is coming, and I don't wanna risk breaking a critical 221 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 222 * I'm gonna fake there being a read node, and I'm gonna swap in a 223 * no-op node in its place (to make all the link-up code happy). 224 * This should be fixed at some point. --jimz 225 */ 226 if (nRrdNodes == 0) { 227 nRrdNodes = 1; 228 rdnodesFaked = 1; 229 } else { 230 rdnodesFaked = 0; 231 } 232 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 233 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 234 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), 235 (RF_DagNode_t *), allocList); 236 i = 0; 237 blockNode = &nodes[i]; 238 i += 1; 239 commitNode = &nodes[i]; 240 i += 1; 241 unblockNode = &nodes[i]; 242 i += 1; 243 termNode = &nodes[i]; 244 i += 1; 245 xorNode = &nodes[i]; 246 i += 1; 247 wnpNode = &nodes[i]; 248 i += 1; 249 wndNodes = &nodes[i]; 250 i += nWndNodes; 251 rrdNodes = &nodes[i]; 252 i += nRrdNodes; 253 if (nfaults == 2) { 254 wnqNode = &nodes[i]; 255 i += 1; 256 } else { 257 wnqNode = NULL; 258 } 259 RF_ASSERT(i == nNodes); 260 261 /* this dag can not commit until all rrd and xor Nodes have completed */ 262 dag_h->numCommitNodes = 1; 263 dag_h->numCommits = 0; 264 dag_h->numSuccedents = 1; 265 266 RF_ASSERT(nRrdNodes > 0); 267 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 268 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 269 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 270 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 271 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 272 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 273 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 274 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 275 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 276 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 277 278 /* 279 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 280 * the failed buffer, save a pointer to it so we can use it as the target 281 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 282 * a buffer is the same size as the failed buffer, it must also be at the 283 * same alignment within the SU. 284 */ 285 i = 0; 286 if (new_asm_h[0]) { 287 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 288 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 289 i++, pda = pda->next) { 290 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 291 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 292 RF_ASSERT(pda); 293 rrdNodes[i].params[0].p = pda; 294 rrdNodes[i].params[1].p = pda->bufPtr; 295 rrdNodes[i].params[2].v = parityStripeID; 296 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 297 } 298 } 299 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 300 if (new_asm_h[1]) { 301 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 302 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 303 j++, pda = pda->next) { 304 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 305 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 306 RF_ASSERT(pda); 307 rrdNodes[i + j].params[0].p = pda; 308 rrdNodes[i + j].params[1].p = pda->bufPtr; 309 rrdNodes[i + j].params[2].v = parityStripeID; 310 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 311 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 312 xorTargetBuf = pda->bufPtr; 313 } 314 } 315 if (rdnodesFaked) { 316 /* 317 * This is where we'll init that fake noop read node 318 * (XXX should the wakeup func be different?) 319 */ 320 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 321 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 322 } 323 /* 324 * Make a PDA for the parity unit. The parity PDA should start at 325 * the same offset into the SU as the failed PDA. 326 */ 327 /* Danner comment: I don't think this copy is really necessary. We are 328 * in one of two cases here. (1) The entire failed unit is written. 329 * Then asmap->parityInfo will describe the entire parity. (2) We are 330 * only writing a subset of the failed unit and nothing else. Then the 331 * asmap->parityInfo describes the failed unit and the copy can also 332 * be avoided. */ 333 334 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 335 parityPDA->row = asmap->parityInfo->row; 336 parityPDA->col = asmap->parityInfo->col; 337 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 338 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 339 parityPDA->numSector = failedPDA->numSector; 340 341 if (!xorTargetBuf) { 342 RF_CallocAndAdd(xorTargetBuf, 1, 343 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 344 } 345 /* init the Wnp node */ 346 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 347 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 348 wnpNode->params[0].p = parityPDA; 349 wnpNode->params[1].p = xorTargetBuf; 350 wnpNode->params[2].v = parityStripeID; 351 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 352 353 /* fill in the Wnq Node */ 354 if (nfaults == 2) { 355 { 356 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 357 (RF_PhysDiskAddr_t *), allocList); 358 parityPDA->row = asmap->qInfo->row; 359 parityPDA->col = asmap->qInfo->col; 360 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 361 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 362 parityPDA->numSector = failedPDA->numSector; 363 364 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 365 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 366 wnqNode->params[0].p = parityPDA; 367 RF_CallocAndAdd(xorNode->results[1], 1, 368 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 369 wnqNode->params[1].p = xorNode->results[1]; 370 wnqNode->params[2].v = parityStripeID; 371 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 372 } 373 } 374 /* fill in the Wnd nodes */ 375 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 376 if (pda == failedPDA) { 377 i--; 378 continue; 379 } 380 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 381 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 382 RF_ASSERT(pda); 383 wndNodes[i].params[0].p = pda; 384 wndNodes[i].params[1].p = pda->bufPtr; 385 wndNodes[i].params[2].v = parityStripeID; 386 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 387 } 388 389 /* fill in the results of the xor node */ 390 xorNode->results[0] = xorTargetBuf; 391 392 /* fill in the params of the xor node */ 393 394 paramNum = 0; 395 if (rdnodesFaked == 0) { 396 for (i = 0; i < nRrdNodes; i++) { 397 /* all the Rrd nodes need to be xored together */ 398 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 399 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 400 } 401 } 402 for (i = 0; i < nWndNodes; i++) { 403 /* any Wnd nodes that overlap the failed access need to be 404 * xored in */ 405 if (overlappingPDAs[i]) { 406 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 407 bcopy((char *) wndNodes[i].params[0].p, (char *) pda, sizeof(RF_PhysDiskAddr_t)); 408 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 409 xorNode->params[paramNum++].p = pda; 410 xorNode->params[paramNum++].p = pda->bufPtr; 411 } 412 } 413 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 414 415 /* 416 * Install the failed PDA into the xor param list so that the 417 * new data gets xor'd in. 418 */ 419 xorNode->params[paramNum++].p = failedPDA; 420 xorNode->params[paramNum++].p = failedPDA->bufPtr; 421 422 /* 423 * The last 2 params to the recovery xor node are always the failed 424 * PDA and the raidPtr. install the failedPDA even though we have just 425 * done so above. This allows us to use the same XOR function for both 426 * degraded reads and degraded writes. 427 */ 428 xorNode->params[paramNum++].p = failedPDA; 429 xorNode->params[paramNum++].p = raidPtr; 430 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 431 432 /* 433 * Code to link nodes begins here 434 */ 435 436 /* link header to block node */ 437 RF_ASSERT(blockNode->numAntecedents == 0); 438 dag_h->succedents[0] = blockNode; 439 440 /* link block node to rd nodes */ 441 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 442 for (i = 0; i < nRrdNodes; i++) { 443 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 444 blockNode->succedents[i] = &rrdNodes[i]; 445 rrdNodes[i].antecedents[0] = blockNode; 446 rrdNodes[i].antType[0] = rf_control; 447 } 448 449 /* link read nodes to xor node */ 450 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 451 for (i = 0; i < nRrdNodes; i++) { 452 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 453 rrdNodes[i].succedents[0] = xorNode; 454 xorNode->antecedents[i] = &rrdNodes[i]; 455 xorNode->antType[i] = rf_trueData; 456 } 457 458 /* link xor node to commit node */ 459 RF_ASSERT(xorNode->numSuccedents == 1); 460 RF_ASSERT(commitNode->numAntecedents == 1); 461 xorNode->succedents[0] = commitNode; 462 commitNode->antecedents[0] = xorNode; 463 commitNode->antType[0] = rf_control; 464 465 /* link commit node to wnd nodes */ 466 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 467 for (i = 0; i < nWndNodes; i++) { 468 RF_ASSERT(wndNodes[i].numAntecedents == 1); 469 commitNode->succedents[i] = &wndNodes[i]; 470 wndNodes[i].antecedents[0] = commitNode; 471 wndNodes[i].antType[0] = rf_control; 472 } 473 474 /* link the commit node to wnp, wnq nodes */ 475 RF_ASSERT(wnpNode->numAntecedents == 1); 476 commitNode->succedents[nWndNodes] = wnpNode; 477 wnpNode->antecedents[0] = commitNode; 478 wnpNode->antType[0] = rf_control; 479 if (nfaults == 2) { 480 RF_ASSERT(wnqNode->numAntecedents == 1); 481 commitNode->succedents[nWndNodes + 1] = wnqNode; 482 wnqNode->antecedents[0] = commitNode; 483 wnqNode->antType[0] = rf_control; 484 } 485 /* link write new data nodes to unblock node */ 486 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 487 for (i = 0; i < nWndNodes; i++) { 488 RF_ASSERT(wndNodes[i].numSuccedents == 1); 489 wndNodes[i].succedents[0] = unblockNode; 490 unblockNode->antecedents[i] = &wndNodes[i]; 491 unblockNode->antType[i] = rf_control; 492 } 493 494 /* link write new parity node to unblock node */ 495 RF_ASSERT(wnpNode->numSuccedents == 1); 496 wnpNode->succedents[0] = unblockNode; 497 unblockNode->antecedents[nWndNodes] = wnpNode; 498 unblockNode->antType[nWndNodes] = rf_control; 499 500 /* link write new q node to unblock node */ 501 if (nfaults == 2) { 502 RF_ASSERT(wnqNode->numSuccedents == 1); 503 wnqNode->succedents[0] = unblockNode; 504 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 505 unblockNode->antType[nWndNodes + 1] = rf_control; 506 } 507 /* link unblock node to term node */ 508 RF_ASSERT(unblockNode->numSuccedents == 1); 509 RF_ASSERT(termNode->numAntecedents == 1); 510 RF_ASSERT(termNode->numSuccedents == 0); 511 unblockNode->succedents[0] = termNode; 512 termNode->antecedents[0] = unblockNode; 513 termNode->antType[0] = rf_control; 514 } 515 #define CONS_PDA(if,start,num) \ 516 pda_p->row = asmap->if->row; pda_p->col = asmap->if->col; \ 517 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 518 pda_p->numSector = num; \ 519 pda_p->next = NULL; \ 520 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 521 #if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 522 void 523 rf_WriteGenerateFailedAccessASMs( 524 RF_Raid_t * raidPtr, 525 RF_AccessStripeMap_t * asmap, 526 RF_PhysDiskAddr_t ** pdap, 527 int *nNodep, 528 RF_PhysDiskAddr_t ** pqpdap, 529 int *nPQNodep, 530 RF_AllocListElem_t * allocList) 531 { 532 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 533 int PDAPerDisk, i; 534 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 535 int numDataCol = layoutPtr->numDataCol; 536 int state; 537 unsigned napdas; 538 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 539 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 540 RF_PhysDiskAddr_t *pda_p; 541 RF_RaidAddr_t sosAddr; 542 543 /* determine how many pda's we will have to generate per unaccess 544 * stripe. If there is only one failed data unit, it is one; if two, 545 * possibly two, depending wether they overlap. */ 546 547 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 548 fone_end = fone_start + fone->numSector; 549 550 if (asmap->numDataFailed == 1) { 551 PDAPerDisk = 1; 552 state = 1; 553 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 554 pda_p = *pqpdap; 555 /* build p */ 556 CONS_PDA(parityInfo, fone_start, fone->numSector); 557 pda_p->type = RF_PDA_TYPE_PARITY; 558 pda_p++; 559 /* build q */ 560 CONS_PDA(qInfo, fone_start, fone->numSector); 561 pda_p->type = RF_PDA_TYPE_Q; 562 } else { 563 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 564 ftwo_end = ftwo_start + ftwo->numSector; 565 if (fone->numSector + ftwo->numSector > secPerSU) { 566 PDAPerDisk = 1; 567 state = 2; 568 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 569 pda_p = *pqpdap; 570 CONS_PDA(parityInfo, 0, secPerSU); 571 pda_p->type = RF_PDA_TYPE_PARITY; 572 pda_p++; 573 CONS_PDA(qInfo, 0, secPerSU); 574 pda_p->type = RF_PDA_TYPE_Q; 575 } else { 576 PDAPerDisk = 2; 577 state = 3; 578 /* four of them, fone, then ftwo */ 579 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 580 pda_p = *pqpdap; 581 CONS_PDA(parityInfo, fone_start, fone->numSector); 582 pda_p->type = RF_PDA_TYPE_PARITY; 583 pda_p++; 584 CONS_PDA(qInfo, fone_start, fone->numSector); 585 pda_p->type = RF_PDA_TYPE_Q; 586 pda_p++; 587 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 588 pda_p->type = RF_PDA_TYPE_PARITY; 589 pda_p++; 590 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 591 pda_p->type = RF_PDA_TYPE_Q; 592 } 593 } 594 /* figure out number of nonaccessed pda */ 595 napdas = PDAPerDisk * (numDataCol - 2); 596 *nPQNodep = PDAPerDisk; 597 598 *nNodep = napdas; 599 if (napdas == 0) 600 return; /* short circuit */ 601 602 /* allocate up our list of pda's */ 603 604 RF_CallocAndAdd(pda_p, napdas, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 605 *pdap = pda_p; 606 607 /* linkem together */ 608 for (i = 0; i < (napdas - 1); i++) 609 pda_p[i].next = pda_p + (i + 1); 610 611 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 612 for (i = 0; i < numDataCol; i++) { 613 if ((pda_p - (*pdap)) == napdas) 614 continue; 615 pda_p->type = RF_PDA_TYPE_DATA; 616 pda_p->raidAddress = sosAddr + (i * secPerSU); 617 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 618 /* skip over dead disks */ 619 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->row][pda_p->col].status)) 620 continue; 621 switch (state) { 622 case 1: /* fone */ 623 pda_p->numSector = fone->numSector; 624 pda_p->raidAddress += fone_start; 625 pda_p->startSector += fone_start; 626 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 627 break; 628 case 2: /* full stripe */ 629 pda_p->numSector = secPerSU; 630 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 631 break; 632 case 3: /* two slabs */ 633 pda_p->numSector = fone->numSector; 634 pda_p->raidAddress += fone_start; 635 pda_p->startSector += fone_start; 636 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 637 pda_p++; 638 pda_p->type = RF_PDA_TYPE_DATA; 639 pda_p->raidAddress = sosAddr + (i * secPerSU); 640 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 641 pda_p->numSector = ftwo->numSector; 642 pda_p->raidAddress += ftwo_start; 643 pda_p->startSector += ftwo_start; 644 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 645 break; 646 default: 647 RF_PANIC(); 648 } 649 pda_p++; 650 } 651 652 RF_ASSERT(pda_p - *pdap == napdas); 653 return; 654 } 655 #define DISK_NODE_PDA(node) ((node)->params[0].p) 656 657 #define DISK_NODE_PARAMS(_node_,_p_) \ 658 (_node_).params[0].p = _p_ ; \ 659 (_node_).params[1].p = (_p_)->bufPtr; \ 660 (_node_).params[2].v = parityStripeID; \ 661 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru) 662 663 void 664 rf_DoubleDegSmallWrite( 665 RF_Raid_t * raidPtr, 666 RF_AccessStripeMap_t * asmap, 667 RF_DagHeader_t * dag_h, 668 void *bp, 669 RF_RaidAccessFlags_t flags, 670 RF_AllocListElem_t * allocList, 671 char *redundantReadNodeName, 672 char *redundantWriteNodeName, 673 char *recoveryNodeName, 674 int (*recovFunc) (RF_DagNode_t *)) 675 { 676 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 677 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 678 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 679 RF_PhysDiskAddr_t *pda, *pqPDAs; 680 RF_PhysDiskAddr_t *npdas; 681 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 682 RF_ReconUnitNum_t which_ru; 683 int nPQNodes; 684 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 685 686 /* simple small write case - First part looks like a reconstruct-read 687 * of the failed data units. Then a write of all data units not 688 * failed. */ 689 690 691 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 692 * / -------PQ----- / \ \ Wud Wp WQ \ | / 693 * --Unblock- | T 694 * 695 * Rrd = read recovery data (potentially none) Wud = write user data 696 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 697 * (could be two) 698 * 699 */ 700 701 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 702 703 RF_ASSERT(asmap->numDataFailed == 1); 704 705 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 706 nReadNodes = nRrdNodes + 2 * nPQNodes; 707 nWriteNodes = nWudNodes + 2 * nPQNodes; 708 nNodes = 4 + nReadNodes + nWriteNodes; 709 710 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 711 blockNode = nodes; 712 unblockNode = blockNode + 1; 713 termNode = unblockNode + 1; 714 recoveryNode = termNode + 1; 715 rrdNodes = recoveryNode + 1; 716 rpNodes = rrdNodes + nRrdNodes; 717 rqNodes = rpNodes + nPQNodes; 718 wudNodes = rqNodes + nPQNodes; 719 wpNodes = wudNodes + nWudNodes; 720 wqNodes = wpNodes + nPQNodes; 721 722 dag_h->creator = "PQ_DDSimpleSmallWrite"; 723 dag_h->numSuccedents = 1; 724 dag_h->succedents[0] = blockNode; 725 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 726 termNode->antecedents[0] = unblockNode; 727 termNode->antType[0] = rf_control; 728 729 /* init the block and unblock nodes */ 730 /* The block node has all the read nodes as successors */ 731 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 732 for (i = 0; i < nReadNodes; i++) 733 blockNode->succedents[i] = rrdNodes + i; 734 735 /* The unblock node has all the writes as successors */ 736 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 737 for (i = 0; i < nWriteNodes; i++) { 738 unblockNode->antecedents[i] = wudNodes + i; 739 unblockNode->antType[i] = rf_control; 740 } 741 unblockNode->succedents[0] = termNode; 742 743 #define INIT_READ_NODE(node,name) \ 744 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 745 (node)->succedents[0] = recoveryNode; \ 746 (node)->antecedents[0] = blockNode; \ 747 (node)->antType[0] = rf_control; 748 749 /* build the read nodes */ 750 pda = npdas; 751 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 752 INIT_READ_NODE(rrdNodes + i, "rrd"); 753 DISK_NODE_PARAMS(rrdNodes[i], pda); 754 } 755 756 /* read redundancy pdas */ 757 pda = pqPDAs; 758 INIT_READ_NODE(rpNodes, "Rp"); 759 RF_ASSERT(pda); 760 DISK_NODE_PARAMS(rpNodes[0], pda); 761 pda++; 762 INIT_READ_NODE(rqNodes, redundantReadNodeName); 763 RF_ASSERT(pda); 764 DISK_NODE_PARAMS(rqNodes[0], pda); 765 if (nPQNodes == 2) { 766 pda++; 767 INIT_READ_NODE(rpNodes + 1, "Rp"); 768 RF_ASSERT(pda); 769 DISK_NODE_PARAMS(rpNodes[1], pda); 770 pda++; 771 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 772 RF_ASSERT(pda); 773 DISK_NODE_PARAMS(rqNodes[1], pda); 774 } 775 /* the recovery node has all reads as precedessors and all writes as 776 * successors. It generates a result for every write P or write Q 777 * node. As parameters, it takes a pda per read and a pda per stripe 778 * of user data written. It also takes as the last params the raidPtr 779 * and asm. For results, it takes PDA for P & Q. */ 780 781 782 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 783 nWriteNodes, /* succesors */ 784 nReadNodes, /* preds */ 785 nReadNodes + nWudNodes + 3, /* params */ 786 2 * nPQNodes, /* results */ 787 dag_h, recoveryNodeName, allocList); 788 789 790 791 for (i = 0; i < nReadNodes; i++) { 792 recoveryNode->antecedents[i] = rrdNodes + i; 793 recoveryNode->antType[i] = rf_control; 794 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 795 } 796 for (i = 0; i < nWudNodes; i++) { 797 recoveryNode->succedents[i] = wudNodes + i; 798 } 799 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 800 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 801 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 802 803 for (; i < nWriteNodes; i++) 804 recoveryNode->succedents[i] = wudNodes + i; 805 806 pda = pqPDAs; 807 recoveryNode->results[0] = pda; 808 pda++; 809 recoveryNode->results[1] = pda; 810 if (nPQNodes == 2) { 811 pda++; 812 recoveryNode->results[2] = pda; 813 pda++; 814 recoveryNode->results[3] = pda; 815 } 816 /* fill writes */ 817 #define INIT_WRITE_NODE(node,name) \ 818 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 819 (node)->succedents[0] = unblockNode; \ 820 (node)->antecedents[0] = recoveryNode; \ 821 (node)->antType[0] = rf_control; 822 823 pda = asmap->physInfo; 824 for (i = 0; i < nWudNodes; i++) { 825 INIT_WRITE_NODE(wudNodes + i, "Wd"); 826 DISK_NODE_PARAMS(wudNodes[i], pda); 827 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 828 pda = pda->next; 829 } 830 /* write redundancy pdas */ 831 pda = pqPDAs; 832 INIT_WRITE_NODE(wpNodes, "Wp"); 833 RF_ASSERT(pda); 834 DISK_NODE_PARAMS(wpNodes[0], pda); 835 pda++; 836 INIT_WRITE_NODE(wqNodes, "Wq"); 837 RF_ASSERT(pda); 838 DISK_NODE_PARAMS(wqNodes[0], pda); 839 if (nPQNodes == 2) { 840 pda++; 841 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 842 RF_ASSERT(pda); 843 DISK_NODE_PARAMS(wpNodes[1], pda); 844 pda++; 845 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 846 RF_ASSERT(pda); 847 DISK_NODE_PARAMS(wqNodes[1], pda); 848 } 849 } 850 #endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 851