rf_parityloggingdags.c revision 1.5
1/* $NetBSD: rf_parityloggingdags.c,v 1.5 2001/09/01 23:50:44 thorpej Exp $ */ 2/* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Author: 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#include "rf_archs.h" 30 31#if RF_INCLUDE_PARITYLOGGING > 0 32 33/* 34 DAGs specific to parity logging are created here 35 */ 36 37#include "rf_types.h" 38#include "rf_raid.h" 39#include "rf_dag.h" 40#include "rf_dagutils.h" 41#include "rf_dagfuncs.h" 42#include "rf_debugMem.h" 43#include "rf_paritylog.h" 44#include "rf_memchunk.h" 45#include "rf_general.h" 46 47#include "rf_parityloggingdags.h" 48 49/****************************************************************************** 50 * 51 * creates a DAG to perform a large-write operation: 52 * 53 * / Rod \ / Wnd \ 54 * H -- NIL- Rod - NIL - Wnd ------ NIL - T 55 * \ Rod / \ Xor - Lpo / 56 * 57 * The writes are not done until the reads complete because if they were done in 58 * parallel, a failure on one of the reads could leave the parity in an inconsistent 59 * state, so that the retry with a new DAG would produce erroneous parity. 60 * 61 * Note: this DAG has the nasty property that none of the buffers allocated for reading 62 * old data can be freed until the XOR node fires. Need to fix this. 63 * 64 * The last two arguments are the number of faults tolerated, and function for the 65 * redundancy calculation. The undo for the redundancy calc is assumed to be null 66 * 67 *****************************************************************************/ 68 69void 70rf_CommonCreateParityLoggingLargeWriteDAG( 71 RF_Raid_t * raidPtr, 72 RF_AccessStripeMap_t * asmap, 73 RF_DagHeader_t * dag_h, 74 void *bp, 75 RF_RaidAccessFlags_t flags, 76 RF_AllocListElem_t * allocList, 77 int nfaults, 78 int (*redFunc) (RF_DagNode_t *)) 79{ 80 RF_DagNode_t *nodes, *wndNodes, *rodNodes = NULL, *syncNode, *xorNode, 81 *lpoNode, *blockNode, *unblockNode, *termNode; 82 int nWndNodes, nRodNodes, i; 83 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 84 RF_AccessStripeMapHeader_t *new_asm_h[2]; 85 int nodeNum, asmNum; 86 RF_ReconUnitNum_t which_ru; 87 char *sosBuffer, *eosBuffer; 88 RF_PhysDiskAddr_t *pda; 89 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru); 90 91 if (rf_dagDebug) 92 printf("[Creating parity-logging large-write DAG]\n"); 93 RF_ASSERT(nfaults == 1);/* this arch only single fault tolerant */ 94 dag_h->creator = "ParityLoggingLargeWriteDAG"; 95 96 /* alloc the Wnd nodes, the xor node, and the Lpo node */ 97 nWndNodes = asmap->numStripeUnitsAccessed; 98 RF_CallocAndAdd(nodes, nWndNodes + 6, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 99 i = 0; 100 wndNodes = &nodes[i]; 101 i += nWndNodes; 102 xorNode = &nodes[i]; 103 i += 1; 104 lpoNode = &nodes[i]; 105 i += 1; 106 blockNode = &nodes[i]; 107 i += 1; 108 syncNode = &nodes[i]; 109 i += 1; 110 unblockNode = &nodes[i]; 111 i += 1; 112 termNode = &nodes[i]; 113 i += 1; 114 115 dag_h->numCommitNodes = nWndNodes + 1; 116 dag_h->numCommits = 0; 117 dag_h->numSuccedents = 1; 118 119 rf_MapUnaccessedPortionOfStripe(raidPtr, layoutPtr, asmap, dag_h, new_asm_h, &nRodNodes, &sosBuffer, &eosBuffer, allocList); 120 if (nRodNodes > 0) 121 RF_CallocAndAdd(rodNodes, nRodNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 122 123 /* begin node initialization */ 124 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nRodNodes + 1, 0, 0, 0, dag_h, "Nil", allocList); 125 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWndNodes + 1, 0, 0, dag_h, "Nil", allocList); 126 rf_InitNode(syncNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nWndNodes + 1, nRodNodes + 1, 0, 0, dag_h, "Nil", allocList); 127 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 128 129 /* initialize the Rod nodes */ 130 for (nodeNum = asmNum = 0; asmNum < 2; asmNum++) { 131 if (new_asm_h[asmNum]) { 132 pda = new_asm_h[asmNum]->stripeMap->physInfo; 133 while (pda) { 134 rf_InitNode(&rodNodes[nodeNum], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rod", allocList); 135 rodNodes[nodeNum].params[0].p = pda; 136 rodNodes[nodeNum].params[1].p = pda->bufPtr; 137 rodNodes[nodeNum].params[2].v = parityStripeID; 138 rodNodes[nodeNum].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 139 nodeNum++; 140 pda = pda->next; 141 } 142 } 143 } 144 RF_ASSERT(nodeNum == nRodNodes); 145 146 /* initialize the wnd nodes */ 147 pda = asmap->physInfo; 148 for (i = 0; i < nWndNodes; i++) { 149 rf_InitNode(&wndNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 150 RF_ASSERT(pda != NULL); 151 wndNodes[i].params[0].p = pda; 152 wndNodes[i].params[1].p = pda->bufPtr; 153 wndNodes[i].params[2].v = parityStripeID; 154 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 155 pda = pda->next; 156 } 157 158 /* initialize the redundancy node */ 159 rf_InitNode(xorNode, rf_wait, RF_TRUE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 1, 2 * (nWndNodes + nRodNodes) + 1, 1, dag_h, "Xr ", allocList); 160 xorNode->flags |= RF_DAGNODE_FLAG_YIELD; 161 for (i = 0; i < nWndNodes; i++) { 162 xorNode->params[2 * i + 0] = wndNodes[i].params[0]; /* pda */ 163 xorNode->params[2 * i + 1] = wndNodes[i].params[1]; /* buf ptr */ 164 } 165 for (i = 0; i < nRodNodes; i++) { 166 xorNode->params[2 * (nWndNodes + i) + 0] = rodNodes[i].params[0]; /* pda */ 167 xorNode->params[2 * (nWndNodes + i) + 1] = rodNodes[i].params[1]; /* buf ptr */ 168 } 169 xorNode->params[2 * (nWndNodes + nRodNodes)].p = raidPtr; /* xor node needs to get 170 * at RAID information */ 171 172 /* look for an Rod node that reads a complete SU. If none, alloc a 173 * buffer to receive the parity info. Note that we can't use a new 174 * data buffer because it will not have gotten written when the xor 175 * occurs. */ 176 for (i = 0; i < nRodNodes; i++) 177 if (((RF_PhysDiskAddr_t *) rodNodes[i].params[0].p)->numSector == raidPtr->Layout.sectorsPerStripeUnit) 178 break; 179 if (i == nRodNodes) { 180 RF_CallocAndAdd(xorNode->results[0], 1, rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit), (void *), allocList); 181 } else { 182 xorNode->results[0] = rodNodes[i].params[1].p; 183 } 184 185 /* initialize the Lpo node */ 186 rf_InitNode(lpoNode, rf_wait, RF_FALSE, rf_ParityLogOverwriteFunc, rf_ParityLogOverwriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpo", allocList); 187 188 lpoNode->params[0].p = asmap->parityInfo; 189 lpoNode->params[1].p = xorNode->results[0]; 190 RF_ASSERT(asmap->parityInfo->next == NULL); /* parityInfo must 191 * describe entire 192 * parity unit */ 193 194 /* connect nodes to form graph */ 195 196 /* connect dag header to block node */ 197 RF_ASSERT(dag_h->numSuccedents == 1); 198 RF_ASSERT(blockNode->numAntecedents == 0); 199 dag_h->succedents[0] = blockNode; 200 201 /* connect the block node to the Rod nodes */ 202 RF_ASSERT(blockNode->numSuccedents == nRodNodes + 1); 203 for (i = 0; i < nRodNodes; i++) { 204 RF_ASSERT(rodNodes[i].numAntecedents == 1); 205 blockNode->succedents[i] = &rodNodes[i]; 206 rodNodes[i].antecedents[0] = blockNode; 207 rodNodes[i].antType[0] = rf_control; 208 } 209 210 /* connect the block node to the sync node */ 211 /* necessary if nRodNodes == 0 */ 212 RF_ASSERT(syncNode->numAntecedents == nRodNodes + 1); 213 blockNode->succedents[nRodNodes] = syncNode; 214 syncNode->antecedents[0] = blockNode; 215 syncNode->antType[0] = rf_control; 216 217 /* connect the Rod nodes to the syncNode */ 218 for (i = 0; i < nRodNodes; i++) { 219 rodNodes[i].succedents[0] = syncNode; 220 syncNode->antecedents[1 + i] = &rodNodes[i]; 221 syncNode->antType[1 + i] = rf_control; 222 } 223 224 /* connect the sync node to the xor node */ 225 RF_ASSERT(syncNode->numSuccedents == nWndNodes + 1); 226 RF_ASSERT(xorNode->numAntecedents == 1); 227 syncNode->succedents[0] = xorNode; 228 xorNode->antecedents[0] = syncNode; 229 xorNode->antType[0] = rf_trueData; /* carry forward from sync */ 230 231 /* connect the sync node to the Wnd nodes */ 232 for (i = 0; i < nWndNodes; i++) { 233 RF_ASSERT(wndNodes->numAntecedents == 1); 234 syncNode->succedents[1 + i] = &wndNodes[i]; 235 wndNodes[i].antecedents[0] = syncNode; 236 wndNodes[i].antType[0] = rf_control; 237 } 238 239 /* connect the xor node to the Lpo node */ 240 RF_ASSERT(xorNode->numSuccedents == 1); 241 RF_ASSERT(lpoNode->numAntecedents == 1); 242 xorNode->succedents[0] = lpoNode; 243 lpoNode->antecedents[0] = xorNode; 244 lpoNode->antType[0] = rf_trueData; 245 246 /* connect the Wnd nodes to the unblock node */ 247 RF_ASSERT(unblockNode->numAntecedents == nWndNodes + 1); 248 for (i = 0; i < nWndNodes; i++) { 249 RF_ASSERT(wndNodes->numSuccedents == 1); 250 wndNodes[i].succedents[0] = unblockNode; 251 unblockNode->antecedents[i] = &wndNodes[i]; 252 unblockNode->antType[i] = rf_control; 253 } 254 255 /* connect the Lpo node to the unblock node */ 256 RF_ASSERT(lpoNode->numSuccedents == 1); 257 lpoNode->succedents[0] = unblockNode; 258 unblockNode->antecedents[nWndNodes] = lpoNode; 259 unblockNode->antType[nWndNodes] = rf_control; 260 261 /* connect unblock node to terminator */ 262 RF_ASSERT(unblockNode->numSuccedents == 1); 263 RF_ASSERT(termNode->numAntecedents == 1); 264 RF_ASSERT(termNode->numSuccedents == 0); 265 unblockNode->succedents[0] = termNode; 266 termNode->antecedents[0] = unblockNode; 267 termNode->antType[0] = rf_control; 268} 269 270 271 272 273/****************************************************************************** 274 * 275 * creates a DAG to perform a small-write operation (either raid 5 or pq), which is as follows: 276 * 277 * Header 278 * | 279 * Block 280 * / | ... \ \ 281 * / | \ \ 282 * Rod Rod Rod Rop 283 * | \ /| \ / | \/ | 284 * | | | /\ | 285 * Wnd Wnd Wnd X 286 * | \ / | 287 * | \ / | 288 * \ \ / Lpo 289 * \ \ / / 290 * +-> Unblock <-+ 291 * | 292 * T 293 * 294 * 295 * R = Read, W = Write, X = Xor, o = old, n = new, d = data, p = parity. 296 * When the access spans a stripe unit boundary and is less than one SU in size, there will 297 * be two Rop -- X -- Wnp branches. I call this the "double-XOR" case. 298 * The second output from each Rod node goes to the X node. In the double-XOR 299 * case, there are exactly 2 Rod nodes, and each sends one output to one X node. 300 * There is one Rod -- Wnd -- T branch for each stripe unit being updated. 301 * 302 * The block and unblock nodes are unused. See comment above CreateFaultFreeReadDAG. 303 * 304 * Note: this DAG ignores all the optimizations related to making the RMWs atomic. 305 * it also has the nasty property that none of the buffers allocated for reading 306 * old data & parity can be freed until the XOR node fires. Need to fix this. 307 * 308 * A null qfuncs indicates single fault tolerant 309 *****************************************************************************/ 310 311void 312rf_CommonCreateParityLoggingSmallWriteDAG( 313 RF_Raid_t * raidPtr, 314 RF_AccessStripeMap_t * asmap, 315 RF_DagHeader_t * dag_h, 316 void *bp, 317 RF_RaidAccessFlags_t flags, 318 RF_AllocListElem_t * allocList, 319 RF_RedFuncs_t * pfuncs, 320 RF_RedFuncs_t * qfuncs) 321{ 322 RF_DagNode_t *xorNodes, *blockNode, *unblockNode, *nodes; 323 RF_DagNode_t *readDataNodes, *readParityNodes; 324 RF_DagNode_t *writeDataNodes, *lpuNodes; 325 RF_DagNode_t *unlockDataNodes = NULL, *termNode; 326 RF_PhysDiskAddr_t *pda = asmap->physInfo; 327 int numDataNodes = asmap->numStripeUnitsAccessed; 328 int numParityNodes = (asmap->parityInfo->next) ? 2 : 1; 329 int i, j, nNodes, totalNumNodes; 330 RF_ReconUnitNum_t which_ru; 331 int (*func) (RF_DagNode_t * node), (*undoFunc) (RF_DagNode_t * node); 332 int (*qfunc) (RF_DagNode_t * node); 333 char *name, *qname; 334 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru); 335#ifdef RAID_DIAGNOSTIC 336 long nfaults = qfuncs ? 2 : 1; 337#endif /* RAID_DIAGNOSTIC */ 338 int lu_flag = (rf_enableAtomicRMW) ? 1 : 0; /* lock/unlock flag */ 339 340 if (rf_dagDebug) 341 printf("[Creating parity-logging small-write DAG]\n"); 342 RF_ASSERT(numDataNodes > 0); 343 RF_ASSERT(nfaults == 1); 344 dag_h->creator = "ParityLoggingSmallWriteDAG"; 345 346 /* DAG creation occurs in three steps: 1. count the number of nodes in 347 * the DAG 2. create the nodes 3. initialize the nodes 4. connect the 348 * nodes */ 349 350 /* Step 1. compute number of nodes in the graph */ 351 352 /* number of nodes: a read and write for each data unit a redundancy 353 * computation node for each parity node a read and Lpu for each 354 * parity unit a block and unblock node (2) a terminator node if 355 * atomic RMW an unlock node for each data unit, redundancy unit */ 356 totalNumNodes = (2 * numDataNodes) + numParityNodes + (2 * numParityNodes) + 3; 357 if (lu_flag) 358 totalNumNodes += numDataNodes; 359 360 nNodes = numDataNodes + numParityNodes; 361 362 dag_h->numCommitNodes = numDataNodes + numParityNodes; 363 dag_h->numCommits = 0; 364 dag_h->numSuccedents = 1; 365 366 /* Step 2. create the nodes */ 367 RF_CallocAndAdd(nodes, totalNumNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 368 i = 0; 369 blockNode = &nodes[i]; 370 i += 1; 371 unblockNode = &nodes[i]; 372 i += 1; 373 readDataNodes = &nodes[i]; 374 i += numDataNodes; 375 readParityNodes = &nodes[i]; 376 i += numParityNodes; 377 writeDataNodes = &nodes[i]; 378 i += numDataNodes; 379 lpuNodes = &nodes[i]; 380 i += numParityNodes; 381 xorNodes = &nodes[i]; 382 i += numParityNodes; 383 termNode = &nodes[i]; 384 i += 1; 385 if (lu_flag) { 386 unlockDataNodes = &nodes[i]; 387 i += numDataNodes; 388 } 389 RF_ASSERT(i == totalNumNodes); 390 391 /* Step 3. initialize the nodes */ 392 /* initialize block node (Nil) */ 393 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nNodes, 0, 0, 0, dag_h, "Nil", allocList); 394 395 /* initialize unblock node (Nil) */ 396 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nNodes, 0, 0, dag_h, "Nil", allocList); 397 398 /* initialize terminatory node (Trm) */ 399 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 400 401 /* initialize nodes which read old data (Rod) */ 402 for (i = 0; i < numDataNodes; i++) { 403 rf_InitNode(&readDataNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rod", allocList); 404 RF_ASSERT(pda != NULL); 405 readDataNodes[i].params[0].p = pda; /* physical disk addr 406 * desc */ 407 readDataNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda, allocList); /* buffer to hold old 408 * data */ 409 readDataNodes[i].params[2].v = parityStripeID; 410 readDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, lu_flag, 0, which_ru); 411 pda = pda->next; 412 readDataNodes[i].propList[0] = NULL; 413 readDataNodes[i].propList[1] = NULL; 414 } 415 416 /* initialize nodes which read old parity (Rop) */ 417 pda = asmap->parityInfo; 418 i = 0; 419 for (i = 0; i < numParityNodes; i++) { 420 RF_ASSERT(pda != NULL); 421 rf_InitNode(&readParityNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rop", allocList); 422 readParityNodes[i].params[0].p = pda; 423 readParityNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda, allocList); /* buffer to hold old 424 * parity */ 425 readParityNodes[i].params[2].v = parityStripeID; 426 readParityNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 427 readParityNodes[i].propList[0] = NULL; 428 pda = pda->next; 429 } 430 431 /* initialize nodes which write new data (Wnd) */ 432 pda = asmap->physInfo; 433 for (i = 0; i < numDataNodes; i++) { 434 RF_ASSERT(pda != NULL); 435 rf_InitNode(&writeDataNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, nNodes, 4, 0, dag_h, "Wnd", allocList); 436 writeDataNodes[i].params[0].p = pda; /* physical disk addr 437 * desc */ 438 writeDataNodes[i].params[1].p = pda->bufPtr; /* buffer holding new 439 * data to be written */ 440 writeDataNodes[i].params[2].v = parityStripeID; 441 writeDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 442 443 if (lu_flag) { 444 /* initialize node to unlock the disk queue */ 445 rf_InitNode(&unlockDataNodes[i], rf_wait, RF_FALSE, rf_DiskUnlockFunc, rf_DiskUnlockUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Und", allocList); 446 unlockDataNodes[i].params[0].p = pda; /* physical disk addr 447 * desc */ 448 unlockDataNodes[i].params[1].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, lu_flag, which_ru); 449 } 450 pda = pda->next; 451 } 452 453 454 /* initialize nodes which compute new parity */ 455 /* we use the simple XOR func in the double-XOR case, and when we're 456 * accessing only a portion of one stripe unit. the distinction 457 * between the two is that the regular XOR func assumes that the 458 * targbuf is a full SU in size, and examines the pda associated with 459 * the buffer to decide where within the buffer to XOR the data, 460 * whereas the simple XOR func just XORs the data into the start of 461 * the buffer. */ 462 if ((numParityNodes == 2) || ((numDataNodes == 1) && (asmap->totalSectorsAccessed < raidPtr->Layout.sectorsPerStripeUnit))) { 463 func = pfuncs->simple; 464 undoFunc = rf_NullNodeUndoFunc; 465 name = pfuncs->SimpleName; 466 if (qfuncs) { 467 qfunc = qfuncs->simple; 468 qname = qfuncs->SimpleName; 469 } 470 } else { 471 func = pfuncs->regular; 472 undoFunc = rf_NullNodeUndoFunc; 473 name = pfuncs->RegularName; 474 if (qfuncs) { 475 qfunc = qfuncs->regular; 476 qname = qfuncs->RegularName; 477 } 478 } 479 /* initialize the xor nodes: params are {pda,buf} from {Rod,Wnd,Rop} 480 * nodes, and raidPtr */ 481 if (numParityNodes == 2) { /* double-xor case */ 482 for (i = 0; i < numParityNodes; i++) { 483 rf_InitNode(&xorNodes[i], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, 7, 1, dag_h, name, allocList); /* no wakeup func for 484 * xor */ 485 xorNodes[i].flags |= RF_DAGNODE_FLAG_YIELD; 486 xorNodes[i].params[0] = readDataNodes[i].params[0]; 487 xorNodes[i].params[1] = readDataNodes[i].params[1]; 488 xorNodes[i].params[2] = readParityNodes[i].params[0]; 489 xorNodes[i].params[3] = readParityNodes[i].params[1]; 490 xorNodes[i].params[4] = writeDataNodes[i].params[0]; 491 xorNodes[i].params[5] = writeDataNodes[i].params[1]; 492 xorNodes[i].params[6].p = raidPtr; 493 xorNodes[i].results[0] = readParityNodes[i].params[1].p; /* use old parity buf as 494 * target buf */ 495 } 496 } else { 497 /* there is only one xor node in this case */ 498 rf_InitNode(&xorNodes[0], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, (2 * (numDataNodes + numDataNodes + 1) + 1), 1, dag_h, name, allocList); 499 xorNodes[0].flags |= RF_DAGNODE_FLAG_YIELD; 500 for (i = 0; i < numDataNodes + 1; i++) { 501 /* set up params related to Rod and Rop nodes */ 502 xorNodes[0].params[2 * i + 0] = readDataNodes[i].params[0]; /* pda */ 503 xorNodes[0].params[2 * i + 1] = readDataNodes[i].params[1]; /* buffer pointer */ 504 } 505 for (i = 0; i < numDataNodes; i++) { 506 /* set up params related to Wnd and Wnp nodes */ 507 xorNodes[0].params[2 * (numDataNodes + 1 + i) + 0] = writeDataNodes[i].params[0]; /* pda */ 508 xorNodes[0].params[2 * (numDataNodes + 1 + i) + 1] = writeDataNodes[i].params[1]; /* buffer pointer */ 509 } 510 xorNodes[0].params[2 * (numDataNodes + numDataNodes + 1)].p = raidPtr; /* xor node needs to get 511 * at RAID information */ 512 xorNodes[0].results[0] = readParityNodes[0].params[1].p; 513 } 514 515 /* initialize the log node(s) */ 516 pda = asmap->parityInfo; 517 for (i = 0; i < numParityNodes; i++) { 518 RF_ASSERT(pda); 519 rf_InitNode(&lpuNodes[i], rf_wait, RF_FALSE, rf_ParityLogUpdateFunc, rf_ParityLogUpdateUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpu", allocList); 520 lpuNodes[i].params[0].p = pda; /* PhysDiskAddr of parity */ 521 lpuNodes[i].params[1].p = xorNodes[i].results[0]; /* buffer pointer to 522 * parity */ 523 pda = pda->next; 524 } 525 526 527 /* Step 4. connect the nodes */ 528 529 /* connect header to block node */ 530 RF_ASSERT(dag_h->numSuccedents == 1); 531 RF_ASSERT(blockNode->numAntecedents == 0); 532 dag_h->succedents[0] = blockNode; 533 534 /* connect block node to read old data nodes */ 535 RF_ASSERT(blockNode->numSuccedents == (numDataNodes + numParityNodes)); 536 for (i = 0; i < numDataNodes; i++) { 537 blockNode->succedents[i] = &readDataNodes[i]; 538 RF_ASSERT(readDataNodes[i].numAntecedents == 1); 539 readDataNodes[i].antecedents[0] = blockNode; 540 readDataNodes[i].antType[0] = rf_control; 541 } 542 543 /* connect block node to read old parity nodes */ 544 for (i = 0; i < numParityNodes; i++) { 545 blockNode->succedents[numDataNodes + i] = &readParityNodes[i]; 546 RF_ASSERT(readParityNodes[i].numAntecedents == 1); 547 readParityNodes[i].antecedents[0] = blockNode; 548 readParityNodes[i].antType[0] = rf_control; 549 } 550 551 /* connect read old data nodes to write new data nodes */ 552 for (i = 0; i < numDataNodes; i++) { 553 RF_ASSERT(readDataNodes[i].numSuccedents == numDataNodes + numParityNodes); 554 for (j = 0; j < numDataNodes; j++) { 555 RF_ASSERT(writeDataNodes[j].numAntecedents == numDataNodes + numParityNodes); 556 readDataNodes[i].succedents[j] = &writeDataNodes[j]; 557 writeDataNodes[j].antecedents[i] = &readDataNodes[i]; 558 if (i == j) 559 writeDataNodes[j].antType[i] = rf_antiData; 560 else 561 writeDataNodes[j].antType[i] = rf_control; 562 } 563 } 564 565 /* connect read old data nodes to xor nodes */ 566 for (i = 0; i < numDataNodes; i++) 567 for (j = 0; j < numParityNodes; j++) { 568 RF_ASSERT(xorNodes[j].numAntecedents == numDataNodes + numParityNodes); 569 readDataNodes[i].succedents[numDataNodes + j] = &xorNodes[j]; 570 xorNodes[j].antecedents[i] = &readDataNodes[i]; 571 xorNodes[j].antType[i] = rf_trueData; 572 } 573 574 /* connect read old parity nodes to write new data nodes */ 575 for (i = 0; i < numParityNodes; i++) { 576 RF_ASSERT(readParityNodes[i].numSuccedents == numDataNodes + numParityNodes); 577 for (j = 0; j < numDataNodes; j++) { 578 readParityNodes[i].succedents[j] = &writeDataNodes[j]; 579 writeDataNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i]; 580 writeDataNodes[j].antType[numDataNodes + i] = rf_control; 581 } 582 } 583 584 /* connect read old parity nodes to xor nodes */ 585 for (i = 0; i < numParityNodes; i++) 586 for (j = 0; j < numParityNodes; j++) { 587 readParityNodes[i].succedents[numDataNodes + j] = &xorNodes[j]; 588 xorNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i]; 589 xorNodes[j].antType[numDataNodes + i] = rf_trueData; 590 } 591 592 /* connect xor nodes to write new parity nodes */ 593 for (i = 0; i < numParityNodes; i++) { 594 RF_ASSERT(xorNodes[i].numSuccedents == 1); 595 RF_ASSERT(lpuNodes[i].numAntecedents == 1); 596 xorNodes[i].succedents[0] = &lpuNodes[i]; 597 lpuNodes[i].antecedents[0] = &xorNodes[i]; 598 lpuNodes[i].antType[0] = rf_trueData; 599 } 600 601 for (i = 0; i < numDataNodes; i++) { 602 if (lu_flag) { 603 /* connect write new data nodes to unlock nodes */ 604 RF_ASSERT(writeDataNodes[i].numSuccedents == 1); 605 RF_ASSERT(unlockDataNodes[i].numAntecedents == 1); 606 writeDataNodes[i].succedents[0] = &unlockDataNodes[i]; 607 unlockDataNodes[i].antecedents[0] = &writeDataNodes[i]; 608 unlockDataNodes[i].antType[0] = rf_control; 609 610 /* connect unlock nodes to unblock node */ 611 RF_ASSERT(unlockDataNodes[i].numSuccedents == 1); 612 RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes))); 613 unlockDataNodes[i].succedents[0] = unblockNode; 614 unblockNode->antecedents[i] = &unlockDataNodes[i]; 615 unblockNode->antType[i] = rf_control; 616 } else { 617 /* connect write new data nodes to unblock node */ 618 RF_ASSERT(writeDataNodes[i].numSuccedents == 1); 619 RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes))); 620 writeDataNodes[i].succedents[0] = unblockNode; 621 unblockNode->antecedents[i] = &writeDataNodes[i]; 622 unblockNode->antType[i] = rf_control; 623 } 624 } 625 626 /* connect write new parity nodes to unblock node */ 627 for (i = 0; i < numParityNodes; i++) { 628 RF_ASSERT(lpuNodes[i].numSuccedents == 1); 629 lpuNodes[i].succedents[0] = unblockNode; 630 unblockNode->antecedents[numDataNodes + i] = &lpuNodes[i]; 631 unblockNode->antType[numDataNodes + i] = rf_control; 632 } 633 634 /* connect unblock node to terminator */ 635 RF_ASSERT(unblockNode->numSuccedents == 1); 636 RF_ASSERT(termNode->numAntecedents == 1); 637 RF_ASSERT(termNode->numSuccedents == 0); 638 unblockNode->succedents[0] = termNode; 639 termNode->antecedents[0] = unblockNode; 640 termNode->antType[0] = rf_control; 641} 642 643 644void 645rf_CreateParityLoggingSmallWriteDAG( 646 RF_Raid_t * raidPtr, 647 RF_AccessStripeMap_t * asmap, 648 RF_DagHeader_t * dag_h, 649 void *bp, 650 RF_RaidAccessFlags_t flags, 651 RF_AllocListElem_t * allocList, 652 RF_RedFuncs_t * pfuncs, 653 RF_RedFuncs_t * qfuncs) 654{ 655 dag_h->creator = "ParityLoggingSmallWriteDAG"; 656 rf_CommonCreateParityLoggingSmallWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, &rf_xorFuncs, NULL); 657} 658 659 660void 661rf_CreateParityLoggingLargeWriteDAG( 662 RF_Raid_t * raidPtr, 663 RF_AccessStripeMap_t * asmap, 664 RF_DagHeader_t * dag_h, 665 void *bp, 666 RF_RaidAccessFlags_t flags, 667 RF_AllocListElem_t * allocList, 668 int nfaults, 669 int (*redFunc) (RF_DagNode_t *)) 670{ 671 dag_h->creator = "ParityLoggingSmallWriteDAG"; 672 rf_CommonCreateParityLoggingLargeWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 1, rf_RegularXorFunc); 673} 674#endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 675