1 /* $NetBSD: rf_dagdegwr.c,v 1.11 2002/08/02 03:42:33 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 1.11 2002/08/02 03:42:33 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 77 static 78 RF_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 84 void 85 rf_CreateDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList) 86 RF_Raid_t *raidPtr; 87 RF_AccessStripeMap_t *asmap; 88 RF_DagHeader_t *dag_h; 89 void *bp; 90 RF_RaidAccessFlags_t flags; 91 RF_AllocListElem_t *allocList; 92 { 93 94 RF_ASSERT(asmap->numDataFailed == 1); 95 dag_h->creator = "DegradedWriteDAG"; 96 97 /* 98 * if the access writes only a portion of the failed unit, and also 99 * writes some portion of at least one surviving unit, we create two 100 * DAGs, one for the failed component and one for the non-failed 101 * component, and do them sequentially. Note that the fact that we're 102 * accessing only a portion of the failed unit indicates that the 103 * access either starts or ends in the failed unit, and hence we need 104 * create only two dags. This is inefficient in that the same data or 105 * parity can get read and written twice using this structure. I need 106 * to fix this to do the access all at once. 107 */ 108 RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 && 109 asmap->failedPDAs[0]->numSector != 110 raidPtr->Layout.sectorsPerStripeUnit)); 111 rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 112 allocList); 113 } 114 115 116 117 /****************************************************************************** 118 * 119 * DAG creation code begins here 120 */ 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 *****************************************************************************/ 155 156 void 157 rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp, flags, 158 allocList, nfaults, redFunc, allowBufferRecycle) 159 RF_Raid_t *raidPtr; 160 RF_AccessStripeMap_t *asmap; 161 RF_DagHeader_t *dag_h; 162 void *bp; 163 RF_RaidAccessFlags_t flags; 164 RF_AllocListElem_t *allocList; 165 int nfaults; 166 int (*redFunc) (RF_DagNode_t *); 167 int allowBufferRecycle; 168 { 169 int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum, 170 rdnodesFaked; 171 RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode; 172 RF_DagNode_t *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode; 173 RF_SectorCount_t sectorsPerSU; 174 RF_ReconUnitNum_t which_ru; 175 char *xorTargetBuf = NULL; /* the target buffer for the XOR 176 * operation */ 177 char *overlappingPDAs;/* a temporary array of flags */ 178 RF_AccessStripeMapHeader_t *new_asm_h[2]; 179 RF_PhysDiskAddr_t *pda, *parityPDA; 180 RF_StripeNum_t parityStripeID; 181 RF_PhysDiskAddr_t *failedPDA; 182 RF_RaidLayout_t *layoutPtr; 183 184 layoutPtr = &(raidPtr->Layout); 185 parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, 186 &which_ru); 187 sectorsPerSU = layoutPtr->sectorsPerStripeUnit; 188 /* failedPDA points to the pda within the asm that targets the failed 189 * disk */ 190 failedPDA = asmap->failedPDAs[0]; 191 192 if (rf_dagDebug) 193 printf("[Creating degraded-write DAG]\n"); 194 195 RF_ASSERT(asmap->numDataFailed == 1); 196 dag_h->creator = "SimpleDegradedWriteDAG"; 197 198 /* 199 * Generate two ASMs identifying the surviving data 200 * we need in order to recover the lost data. 201 */ 202 /* overlappingPDAs array must be zero'd */ 203 RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed, sizeof(char), (char *)); 204 rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, 205 &nXorBufs, NULL, overlappingPDAs, allocList); 206 207 /* create all the nodes at once */ 208 nWndNodes = asmap->numStripeUnitsAccessed - 1; /* no access is 209 * generated for the 210 * failed pda */ 211 212 nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) + 213 ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0); 214 /* 215 * XXX 216 * 217 * There's a bug with a complete stripe overwrite- that means 0 reads 218 * of old data, and the rest of the DAG generation code doesn't like 219 * that. A release is coming, and I don't wanna risk breaking a critical 220 * DAG generator, so here's what I'm gonna do- if there's no read nodes, 221 * I'm gonna fake there being a read node, and I'm gonna swap in a 222 * no-op node in its place (to make all the link-up code happy). 223 * This should be fixed at some point. --jimz 224 */ 225 if (nRrdNodes == 0) { 226 nRrdNodes = 1; 227 rdnodesFaked = 1; 228 } else { 229 rdnodesFaked = 0; 230 } 231 /* lock, unlock, xor, Wnd, Rrd, W(nfaults) */ 232 nNodes = 5 + nfaults + nWndNodes + nRrdNodes; 233 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), 234 (RF_DagNode_t *), allocList); 235 i = 0; 236 blockNode = &nodes[i]; 237 i += 1; 238 commitNode = &nodes[i]; 239 i += 1; 240 unblockNode = &nodes[i]; 241 i += 1; 242 termNode = &nodes[i]; 243 i += 1; 244 xorNode = &nodes[i]; 245 i += 1; 246 wnpNode = &nodes[i]; 247 i += 1; 248 wndNodes = &nodes[i]; 249 i += nWndNodes; 250 rrdNodes = &nodes[i]; 251 i += nRrdNodes; 252 if (nfaults == 2) { 253 wnqNode = &nodes[i]; 254 i += 1; 255 } else { 256 wnqNode = NULL; 257 } 258 RF_ASSERT(i == nNodes); 259 260 /* this dag can not commit until all rrd and xor Nodes have completed */ 261 dag_h->numCommitNodes = 1; 262 dag_h->numCommits = 0; 263 dag_h->numSuccedents = 1; 264 265 RF_ASSERT(nRrdNodes > 0); 266 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 267 NULL, nRrdNodes, 0, 0, 0, dag_h, "Nil", allocList); 268 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 269 NULL, nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList); 270 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 271 NULL, 1, nWndNodes + nfaults, 0, 0, dag_h, "Nil", allocList); 272 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 273 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 274 rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 275 nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc", allocList); 276 277 /* 278 * Fill in the Rrd nodes. If any of the rrd buffers are the same size as 279 * the failed buffer, save a pointer to it so we can use it as the target 280 * of the XOR. The pdas in the rrd nodes have been range-restricted, so if 281 * a buffer is the same size as the failed buffer, it must also be at the 282 * same alignment within the SU. 283 */ 284 i = 0; 285 if (new_asm_h[0]) { 286 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo; 287 i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 288 i++, pda = pda->next) { 289 rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 290 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 291 RF_ASSERT(pda); 292 rrdNodes[i].params[0].p = pda; 293 rrdNodes[i].params[1].p = pda->bufPtr; 294 rrdNodes[i].params[2].v = parityStripeID; 295 rrdNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 296 } 297 } 298 /* i now equals the number of stripe units accessed in new_asm_h[0] */ 299 if (new_asm_h[1]) { 300 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo; 301 j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 302 j++, pda = pda->next) { 303 rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, 304 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rrd", allocList); 305 RF_ASSERT(pda); 306 rrdNodes[i + j].params[0].p = pda; 307 rrdNodes[i + j].params[1].p = pda->bufPtr; 308 rrdNodes[i + j].params[2].v = parityStripeID; 309 rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 310 if (allowBufferRecycle && (pda->numSector == failedPDA->numSector)) 311 xorTargetBuf = pda->bufPtr; 312 } 313 } 314 if (rdnodesFaked) { 315 /* 316 * This is where we'll init that fake noop read node 317 * (XXX should the wakeup func be different?) 318 */ 319 rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 320 NULL, 1, 1, 0, 0, dag_h, "RrN", allocList); 321 } 322 /* 323 * Make a PDA for the parity unit. The parity PDA should start at 324 * the same offset into the SU as the failed PDA. 325 */ 326 /* Danner comment: I don't think this copy is really necessary. We are 327 * in one of two cases here. (1) The entire failed unit is written. 328 * Then asmap->parityInfo will describe the entire parity. (2) We are 329 * only writing a subset of the failed unit and nothing else. Then the 330 * asmap->parityInfo describes the failed unit and the copy can also 331 * be avoided. */ 332 333 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 334 parityPDA->row = asmap->parityInfo->row; 335 parityPDA->col = asmap->parityInfo->col; 336 parityPDA->startSector = ((asmap->parityInfo->startSector / sectorsPerSU) 337 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 338 parityPDA->numSector = failedPDA->numSector; 339 340 if (!xorTargetBuf) { 341 RF_CallocAndAdd(xorTargetBuf, 1, 342 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 343 } 344 /* init the Wnp node */ 345 rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 346 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList); 347 wnpNode->params[0].p = parityPDA; 348 wnpNode->params[1].p = xorTargetBuf; 349 wnpNode->params[2].v = parityStripeID; 350 wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 351 352 /* fill in the Wnq Node */ 353 if (nfaults == 2) { 354 { 355 RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t), 356 (RF_PhysDiskAddr_t *), allocList); 357 parityPDA->row = asmap->qInfo->row; 358 parityPDA->col = asmap->qInfo->col; 359 parityPDA->startSector = ((asmap->qInfo->startSector / sectorsPerSU) 360 * sectorsPerSU) + (failedPDA->startSector % sectorsPerSU); 361 parityPDA->numSector = failedPDA->numSector; 362 363 rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 364 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList); 365 wnqNode->params[0].p = parityPDA; 366 RF_CallocAndAdd(xorNode->results[1], 1, 367 rf_RaidAddressToByte(raidPtr, failedPDA->numSector), (char *), allocList); 368 wnqNode->params[1].p = xorNode->results[1]; 369 wnqNode->params[2].v = parityStripeID; 370 wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 371 } 372 } 373 /* fill in the Wnd nodes */ 374 for (pda = asmap->physInfo, i = 0; i < nWndNodes; i++, pda = pda->next) { 375 if (pda == failedPDA) { 376 i--; 377 continue; 378 } 379 rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, 380 rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 381 RF_ASSERT(pda); 382 wndNodes[i].params[0].p = pda; 383 wndNodes[i].params[1].p = pda->bufPtr; 384 wndNodes[i].params[2].v = parityStripeID; 385 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 386 } 387 388 /* fill in the results of the xor node */ 389 xorNode->results[0] = xorTargetBuf; 390 391 /* fill in the params of the xor node */ 392 393 paramNum = 0; 394 if (rdnodesFaked == 0) { 395 for (i = 0; i < nRrdNodes; i++) { 396 /* all the Rrd nodes need to be xored together */ 397 xorNode->params[paramNum++] = rrdNodes[i].params[0]; 398 xorNode->params[paramNum++] = rrdNodes[i].params[1]; 399 } 400 } 401 for (i = 0; i < nWndNodes; i++) { 402 /* any Wnd nodes that overlap the failed access need to be 403 * xored in */ 404 if (overlappingPDAs[i]) { 405 RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 406 memcpy((char *) pda, (char *) wndNodes[i].params[0].p, sizeof(RF_PhysDiskAddr_t)); 407 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_DOBUFFER, 0); 408 xorNode->params[paramNum++].p = pda; 409 xorNode->params[paramNum++].p = pda->bufPtr; 410 } 411 } 412 RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char)); 413 414 /* 415 * Install the failed PDA into the xor param list so that the 416 * new data gets xor'd in. 417 */ 418 xorNode->params[paramNum++].p = failedPDA; 419 xorNode->params[paramNum++].p = failedPDA->bufPtr; 420 421 /* 422 * The last 2 params to the recovery xor node are always the failed 423 * PDA and the raidPtr. install the failedPDA even though we have just 424 * done so above. This allows us to use the same XOR function for both 425 * degraded reads and degraded writes. 426 */ 427 xorNode->params[paramNum++].p = failedPDA; 428 xorNode->params[paramNum++].p = raidPtr; 429 RF_ASSERT(paramNum == 2 * nXorBufs + 2); 430 431 /* 432 * Code to link nodes begins here 433 */ 434 435 /* link header to block node */ 436 RF_ASSERT(blockNode->numAntecedents == 0); 437 dag_h->succedents[0] = blockNode; 438 439 /* link block node to rd nodes */ 440 RF_ASSERT(blockNode->numSuccedents == nRrdNodes); 441 for (i = 0; i < nRrdNodes; i++) { 442 RF_ASSERT(rrdNodes[i].numAntecedents == 1); 443 blockNode->succedents[i] = &rrdNodes[i]; 444 rrdNodes[i].antecedents[0] = blockNode; 445 rrdNodes[i].antType[0] = rf_control; 446 } 447 448 /* link read nodes to xor node */ 449 RF_ASSERT(xorNode->numAntecedents == nRrdNodes); 450 for (i = 0; i < nRrdNodes; i++) { 451 RF_ASSERT(rrdNodes[i].numSuccedents == 1); 452 rrdNodes[i].succedents[0] = xorNode; 453 xorNode->antecedents[i] = &rrdNodes[i]; 454 xorNode->antType[i] = rf_trueData; 455 } 456 457 /* link xor node to commit node */ 458 RF_ASSERT(xorNode->numSuccedents == 1); 459 RF_ASSERT(commitNode->numAntecedents == 1); 460 xorNode->succedents[0] = commitNode; 461 commitNode->antecedents[0] = xorNode; 462 commitNode->antType[0] = rf_control; 463 464 /* link commit node to wnd nodes */ 465 RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes); 466 for (i = 0; i < nWndNodes; i++) { 467 RF_ASSERT(wndNodes[i].numAntecedents == 1); 468 commitNode->succedents[i] = &wndNodes[i]; 469 wndNodes[i].antecedents[0] = commitNode; 470 wndNodes[i].antType[0] = rf_control; 471 } 472 473 /* link the commit node to wnp, wnq nodes */ 474 RF_ASSERT(wnpNode->numAntecedents == 1); 475 commitNode->succedents[nWndNodes] = wnpNode; 476 wnpNode->antecedents[0] = commitNode; 477 wnpNode->antType[0] = rf_control; 478 if (nfaults == 2) { 479 RF_ASSERT(wnqNode->numAntecedents == 1); 480 commitNode->succedents[nWndNodes + 1] = wnqNode; 481 wnqNode->antecedents[0] = commitNode; 482 wnqNode->antType[0] = rf_control; 483 } 484 /* link write new data nodes to unblock node */ 485 RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults)); 486 for (i = 0; i < nWndNodes; i++) { 487 RF_ASSERT(wndNodes[i].numSuccedents == 1); 488 wndNodes[i].succedents[0] = unblockNode; 489 unblockNode->antecedents[i] = &wndNodes[i]; 490 unblockNode->antType[i] = rf_control; 491 } 492 493 /* link write new parity node to unblock node */ 494 RF_ASSERT(wnpNode->numSuccedents == 1); 495 wnpNode->succedents[0] = unblockNode; 496 unblockNode->antecedents[nWndNodes] = wnpNode; 497 unblockNode->antType[nWndNodes] = rf_control; 498 499 /* link write new q node to unblock node */ 500 if (nfaults == 2) { 501 RF_ASSERT(wnqNode->numSuccedents == 1); 502 wnqNode->succedents[0] = unblockNode; 503 unblockNode->antecedents[nWndNodes + 1] = wnqNode; 504 unblockNode->antType[nWndNodes + 1] = rf_control; 505 } 506 /* link unblock node to term node */ 507 RF_ASSERT(unblockNode->numSuccedents == 1); 508 RF_ASSERT(termNode->numAntecedents == 1); 509 RF_ASSERT(termNode->numSuccedents == 0); 510 unblockNode->succedents[0] = termNode; 511 termNode->antecedents[0] = unblockNode; 512 termNode->antType[0] = rf_control; 513 } 514 #define CONS_PDA(if,start,num) \ 515 pda_p->row = asmap->if->row; pda_p->col = asmap->if->col; \ 516 pda_p->startSector = ((asmap->if->startSector / secPerSU) * secPerSU) + start; \ 517 pda_p->numSector = num; \ 518 pda_p->next = NULL; \ 519 RF_MallocAndAdd(pda_p->bufPtr,rf_RaidAddressToByte(raidPtr,num),(char *), allocList) 520 #if (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) 521 void 522 rf_WriteGenerateFailedAccessASMs( 523 RF_Raid_t * raidPtr, 524 RF_AccessStripeMap_t * asmap, 525 RF_PhysDiskAddr_t ** pdap, 526 int *nNodep, 527 RF_PhysDiskAddr_t ** pqpdap, 528 int *nPQNodep, 529 RF_AllocListElem_t * allocList) 530 { 531 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 532 int PDAPerDisk, i; 533 RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit; 534 int numDataCol = layoutPtr->numDataCol; 535 int state; 536 unsigned napdas; 537 RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end; 538 RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0], *ftwo = asmap->failedPDAs[1]; 539 RF_PhysDiskAddr_t *pda_p; 540 RF_RaidAddr_t sosAddr; 541 542 /* determine how many pda's we will have to generate per unaccess 543 * stripe. If there is only one failed data unit, it is one; if two, 544 * possibly two, depending wether they overlap. */ 545 546 fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector); 547 fone_end = fone_start + fone->numSector; 548 549 if (asmap->numDataFailed == 1) { 550 PDAPerDisk = 1; 551 state = 1; 552 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 553 pda_p = *pqpdap; 554 /* build p */ 555 CONS_PDA(parityInfo, fone_start, fone->numSector); 556 pda_p->type = RF_PDA_TYPE_PARITY; 557 pda_p++; 558 /* build q */ 559 CONS_PDA(qInfo, fone_start, fone->numSector); 560 pda_p->type = RF_PDA_TYPE_Q; 561 } else { 562 ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector); 563 ftwo_end = ftwo_start + ftwo->numSector; 564 if (fone->numSector + ftwo->numSector > secPerSU) { 565 PDAPerDisk = 1; 566 state = 2; 567 RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 568 pda_p = *pqpdap; 569 CONS_PDA(parityInfo, 0, secPerSU); 570 pda_p->type = RF_PDA_TYPE_PARITY; 571 pda_p++; 572 CONS_PDA(qInfo, 0, secPerSU); 573 pda_p->type = RF_PDA_TYPE_Q; 574 } else { 575 PDAPerDisk = 2; 576 state = 3; 577 /* four of them, fone, then ftwo */ 578 RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 579 pda_p = *pqpdap; 580 CONS_PDA(parityInfo, fone_start, fone->numSector); 581 pda_p->type = RF_PDA_TYPE_PARITY; 582 pda_p++; 583 CONS_PDA(qInfo, fone_start, fone->numSector); 584 pda_p->type = RF_PDA_TYPE_Q; 585 pda_p++; 586 CONS_PDA(parityInfo, ftwo_start, ftwo->numSector); 587 pda_p->type = RF_PDA_TYPE_PARITY; 588 pda_p++; 589 CONS_PDA(qInfo, ftwo_start, ftwo->numSector); 590 pda_p->type = RF_PDA_TYPE_Q; 591 } 592 } 593 /* figure out number of nonaccessed pda */ 594 napdas = PDAPerDisk * (numDataCol - 2); 595 *nPQNodep = PDAPerDisk; 596 597 *nNodep = napdas; 598 if (napdas == 0) 599 return; /* short circuit */ 600 601 /* allocate up our list of pda's */ 602 603 RF_CallocAndAdd(pda_p, napdas, sizeof(RF_PhysDiskAddr_t), (RF_PhysDiskAddr_t *), allocList); 604 *pdap = pda_p; 605 606 /* linkem together */ 607 for (i = 0; i < (napdas - 1); i++) 608 pda_p[i].next = pda_p + (i + 1); 609 610 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 611 for (i = 0; i < numDataCol; i++) { 612 if ((pda_p - (*pdap)) == napdas) 613 continue; 614 pda_p->type = RF_PDA_TYPE_DATA; 615 pda_p->raidAddress = sosAddr + (i * secPerSU); 616 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 617 /* skip over dead disks */ 618 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->row][pda_p->col].status)) 619 continue; 620 switch (state) { 621 case 1: /* fone */ 622 pda_p->numSector = fone->numSector; 623 pda_p->raidAddress += fone_start; 624 pda_p->startSector += fone_start; 625 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 626 break; 627 case 2: /* full stripe */ 628 pda_p->numSector = secPerSU; 629 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, secPerSU), (char *), allocList); 630 break; 631 case 3: /* two slabs */ 632 pda_p->numSector = fone->numSector; 633 pda_p->raidAddress += fone_start; 634 pda_p->startSector += fone_start; 635 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 636 pda_p++; 637 pda_p->type = RF_PDA_TYPE_DATA; 638 pda_p->raidAddress = sosAddr + (i * secPerSU); 639 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress, &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0); 640 pda_p->numSector = ftwo->numSector; 641 pda_p->raidAddress += ftwo_start; 642 pda_p->startSector += ftwo_start; 643 RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(raidPtr, pda_p->numSector), (char *), allocList); 644 break; 645 default: 646 RF_PANIC(); 647 } 648 pda_p++; 649 } 650 651 RF_ASSERT(pda_p - *pdap == napdas); 652 return; 653 } 654 #define DISK_NODE_PDA(node) ((node)->params[0].p) 655 656 #define DISK_NODE_PARAMS(_node_,_p_) \ 657 (_node_).params[0].p = _p_ ; \ 658 (_node_).params[1].p = (_p_)->bufPtr; \ 659 (_node_).params[2].v = parityStripeID; \ 660 (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru) 661 662 void 663 rf_DoubleDegSmallWrite( 664 RF_Raid_t * raidPtr, 665 RF_AccessStripeMap_t * asmap, 666 RF_DagHeader_t * dag_h, 667 void *bp, 668 RF_RaidAccessFlags_t flags, 669 RF_AllocListElem_t * allocList, 670 char *redundantReadNodeName, 671 char *redundantWriteNodeName, 672 char *recoveryNodeName, 673 int (*recovFunc) (RF_DagNode_t *)) 674 { 675 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 676 RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode, 677 *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode; 678 RF_PhysDiskAddr_t *pda, *pqPDAs; 679 RF_PhysDiskAddr_t *npdas; 680 int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i; 681 RF_ReconUnitNum_t which_ru; 682 int nPQNodes; 683 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress, &which_ru); 684 685 /* simple small write case - First part looks like a reconstruct-read 686 * of the failed data units. Then a write of all data units not 687 * failed. */ 688 689 690 /* Hdr | ------Block- / / \ Rrd Rrd ... Rrd Rp Rq \ \ 691 * / -------PQ----- / \ \ Wud Wp WQ \ | / 692 * --Unblock- | T 693 * 694 * Rrd = read recovery data (potentially none) Wud = write user data 695 * (not incl. failed disks) Wp = Write P (could be two) Wq = Write Q 696 * (could be two) 697 * 698 */ 699 700 rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes, &pqPDAs, &nPQNodes, allocList); 701 702 RF_ASSERT(asmap->numDataFailed == 1); 703 704 nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed); 705 nReadNodes = nRrdNodes + 2 * nPQNodes; 706 nWriteNodes = nWudNodes + 2 * nPQNodes; 707 nNodes = 4 + nReadNodes + nWriteNodes; 708 709 RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 710 blockNode = nodes; 711 unblockNode = blockNode + 1; 712 termNode = unblockNode + 1; 713 recoveryNode = termNode + 1; 714 rrdNodes = recoveryNode + 1; 715 rpNodes = rrdNodes + nRrdNodes; 716 rqNodes = rpNodes + nPQNodes; 717 wudNodes = rqNodes + nPQNodes; 718 wpNodes = wudNodes + nWudNodes; 719 wqNodes = wpNodes + nPQNodes; 720 721 dag_h->creator = "PQ_DDSimpleSmallWrite"; 722 dag_h->numSuccedents = 1; 723 dag_h->succedents[0] = blockNode; 724 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 725 termNode->antecedents[0] = unblockNode; 726 termNode->antType[0] = rf_control; 727 728 /* init the block and unblock nodes */ 729 /* The block node has all the read nodes as successors */ 730 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h, "Nil", allocList); 731 for (i = 0; i < nReadNodes; i++) 732 blockNode->succedents[i] = rrdNodes + i; 733 734 /* The unblock node has all the writes as successors */ 735 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h, "Nil", allocList); 736 for (i = 0; i < nWriteNodes; i++) { 737 unblockNode->antecedents[i] = wudNodes + i; 738 unblockNode->antType[i] = rf_control; 739 } 740 unblockNode->succedents[0] = termNode; 741 742 #define INIT_READ_NODE(node,name) \ 743 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 744 (node)->succedents[0] = recoveryNode; \ 745 (node)->antecedents[0] = blockNode; \ 746 (node)->antType[0] = rf_control; 747 748 /* build the read nodes */ 749 pda = npdas; 750 for (i = 0; i < nRrdNodes; i++, pda = pda->next) { 751 INIT_READ_NODE(rrdNodes + i, "rrd"); 752 DISK_NODE_PARAMS(rrdNodes[i], pda); 753 } 754 755 /* read redundancy pdas */ 756 pda = pqPDAs; 757 INIT_READ_NODE(rpNodes, "Rp"); 758 RF_ASSERT(pda); 759 DISK_NODE_PARAMS(rpNodes[0], pda); 760 pda++; 761 INIT_READ_NODE(rqNodes, redundantReadNodeName); 762 RF_ASSERT(pda); 763 DISK_NODE_PARAMS(rqNodes[0], pda); 764 if (nPQNodes == 2) { 765 pda++; 766 INIT_READ_NODE(rpNodes + 1, "Rp"); 767 RF_ASSERT(pda); 768 DISK_NODE_PARAMS(rpNodes[1], pda); 769 pda++; 770 INIT_READ_NODE(rqNodes + 1, redundantReadNodeName); 771 RF_ASSERT(pda); 772 DISK_NODE_PARAMS(rqNodes[1], pda); 773 } 774 /* the recovery node has all reads as precedessors and all writes as 775 * successors. It generates a result for every write P or write Q 776 * node. As parameters, it takes a pda per read and a pda per stripe 777 * of user data written. It also takes as the last params the raidPtr 778 * and asm. For results, it takes PDA for P & Q. */ 779 780 781 rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc, rf_NullNodeUndoFunc, NULL, 782 nWriteNodes, /* succesors */ 783 nReadNodes, /* preds */ 784 nReadNodes + nWudNodes + 3, /* params */ 785 2 * nPQNodes, /* results */ 786 dag_h, recoveryNodeName, allocList); 787 788 789 790 for (i = 0; i < nReadNodes; i++) { 791 recoveryNode->antecedents[i] = rrdNodes + i; 792 recoveryNode->antType[i] = rf_control; 793 recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i); 794 } 795 for (i = 0; i < nWudNodes; i++) { 796 recoveryNode->succedents[i] = wudNodes + i; 797 } 798 recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0]; 799 recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr; 800 recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap; 801 802 for (; i < nWriteNodes; i++) 803 recoveryNode->succedents[i] = wudNodes + i; 804 805 pda = pqPDAs; 806 recoveryNode->results[0] = pda; 807 pda++; 808 recoveryNode->results[1] = pda; 809 if (nPQNodes == 2) { 810 pda++; 811 recoveryNode->results[2] = pda; 812 pda++; 813 recoveryNode->results[3] = pda; 814 } 815 /* fill writes */ 816 #define INIT_WRITE_NODE(node,name) \ 817 rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, name, allocList); \ 818 (node)->succedents[0] = unblockNode; \ 819 (node)->antecedents[0] = recoveryNode; \ 820 (node)->antType[0] = rf_control; 821 822 pda = asmap->physInfo; 823 for (i = 0; i < nWudNodes; i++) { 824 INIT_WRITE_NODE(wudNodes + i, "Wd"); 825 DISK_NODE_PARAMS(wudNodes[i], pda); 826 recoveryNode->params[nReadNodes + i].p = DISK_NODE_PDA(wudNodes + i); 827 pda = pda->next; 828 } 829 /* write redundancy pdas */ 830 pda = pqPDAs; 831 INIT_WRITE_NODE(wpNodes, "Wp"); 832 RF_ASSERT(pda); 833 DISK_NODE_PARAMS(wpNodes[0], pda); 834 pda++; 835 INIT_WRITE_NODE(wqNodes, "Wq"); 836 RF_ASSERT(pda); 837 DISK_NODE_PARAMS(wqNodes[0], pda); 838 if (nPQNodes == 2) { 839 pda++; 840 INIT_WRITE_NODE(wpNodes + 1, "Wp"); 841 RF_ASSERT(pda); 842 DISK_NODE_PARAMS(wpNodes[1], pda); 843 pda++; 844 INIT_WRITE_NODE(wqNodes + 1, "Wq"); 845 RF_ASSERT(pda); 846 DISK_NODE_PARAMS(wqNodes[1], pda); 847 } 848 } 849 #endif /* (RF_INCLUDE_PQ > 0) || (RF_INCLUDE_EVENODD > 0) */ 850