rf_parityloggingdags.c revision 1.16
1 /* $NetBSD: rf_parityloggingdags.c,v 1.16 2006/08/17 17:11:28 christos 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 (at) 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 DAGs specific to parity logging are created here 31 */ 32 33 #include <sys/cdefs.h> 34 __KERNEL_RCSID(0, "$NetBSD: rf_parityloggingdags.c,v 1.16 2006/08/17 17:11:28 christos Exp $"); 35 36 #include "rf_archs.h" 37 #include "opt_raid_diagnostic.h" 38 39 #if RF_INCLUDE_PARITYLOGGING > 0 40 41 #include <dev/raidframe/raidframevar.h> 42 43 #include "rf_raid.h" 44 #include "rf_dag.h" 45 #include "rf_dagutils.h" 46 #include "rf_dagfuncs.h" 47 #include "rf_debugMem.h" 48 #include "rf_paritylog.h" 49 #include "rf_general.h" 50 51 #include "rf_parityloggingdags.h" 52 53 /****************************************************************************** 54 * 55 * creates a DAG to perform a large-write operation: 56 * 57 * / Rod \ / Wnd \ 58 * H -- NIL- Rod - NIL - Wnd ------ NIL - T 59 * \ Rod / \ Xor - Lpo / 60 * 61 * The writes are not done until the reads complete because if they were done in 62 * parallel, a failure on one of the reads could leave the parity in an inconsistent 63 * state, so that the retry with a new DAG would produce erroneous parity. 64 * 65 * Note: this DAG has the nasty property that none of the buffers allocated for reading 66 * old data can be freed until the XOR node fires. Need to fix this. 67 * 68 * The last two arguments are the number of faults tolerated, and function for the 69 * redundancy calculation. The undo for the redundancy calc is assumed to be null 70 * 71 *****************************************************************************/ 72 73 void 74 rf_CommonCreateParityLoggingLargeWriteDAG( 75 RF_Raid_t * raidPtr, 76 RF_AccessStripeMap_t * asmap, 77 RF_DagHeader_t * dag_h, 78 void *bp, 79 RF_RaidAccessFlags_t flags, 80 RF_AllocListElem_t * allocList, 81 int nfaults, 82 int (*redFunc) (RF_DagNode_t *)) 83 { 84 RF_DagNode_t *nodes, *wndNodes, *rodNodes = NULL, *syncNode, *xorNode, 85 *lpoNode, *blockNode, *unblockNode, *termNode; 86 int nWndNodes, nRodNodes, i; 87 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 88 RF_AccessStripeMapHeader_t *new_asm_h[2]; 89 int nodeNum, asmNum; 90 RF_ReconUnitNum_t which_ru; 91 char *sosBuffer, *eosBuffer; 92 RF_PhysDiskAddr_t *pda; 93 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru); 94 95 if (rf_dagDebug) 96 printf("[Creating parity-logging large-write DAG]\n"); 97 RF_ASSERT(nfaults == 1);/* this arch only single fault tolerant */ 98 dag_h->creator = "ParityLoggingLargeWriteDAG"; 99 100 /* alloc the Wnd nodes, the xor node, and the Lpo node */ 101 nWndNodes = asmap->numStripeUnitsAccessed; 102 RF_MallocAndAdd(nodes, (nWndNodes + 6) * sizeof(RF_DagNode_t), 103 (RF_DagNode_t *), allocList); 104 i = 0; 105 wndNodes = &nodes[i]; 106 i += nWndNodes; 107 xorNode = &nodes[i]; 108 i += 1; 109 lpoNode = &nodes[i]; 110 i += 1; 111 blockNode = &nodes[i]; 112 i += 1; 113 syncNode = &nodes[i]; 114 i += 1; 115 unblockNode = &nodes[i]; 116 i += 1; 117 termNode = &nodes[i]; 118 i += 1; 119 120 dag_h->numCommitNodes = nWndNodes + 1; 121 dag_h->numCommits = 0; 122 dag_h->numSuccedents = 1; 123 124 rf_MapUnaccessedPortionOfStripe(raidPtr, layoutPtr, asmap, dag_h, new_asm_h, &nRodNodes, &sosBuffer, &eosBuffer, allocList); 125 if (nRodNodes > 0) 126 RF_MallocAndAdd(rodNodes, nRodNodes * sizeof(RF_DagNode_t), 127 (RF_DagNode_t *), allocList); 128 129 /* begin node initialization */ 130 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nRodNodes + 1, 0, 0, 0, dag_h, "Nil", allocList); 131 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWndNodes + 1, 0, 0, dag_h, "Nil", allocList); 132 rf_InitNode(syncNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nWndNodes + 1, nRodNodes + 1, 0, 0, dag_h, "Nil", allocList); 133 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 134 135 /* initialize the Rod nodes */ 136 for (nodeNum = asmNum = 0; asmNum < 2; asmNum++) { 137 if (new_asm_h[asmNum]) { 138 pda = new_asm_h[asmNum]->stripeMap->physInfo; 139 while (pda) { 140 rf_InitNode(&rodNodes[nodeNum], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rod", allocList); 141 rodNodes[nodeNum].params[0].p = pda; 142 rodNodes[nodeNum].params[1].p = pda->bufPtr; 143 rodNodes[nodeNum].params[2].v = parityStripeID; 144 rodNodes[nodeNum].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 145 nodeNum++; 146 pda = pda->next; 147 } 148 } 149 } 150 RF_ASSERT(nodeNum == nRodNodes); 151 152 /* initialize the wnd nodes */ 153 pda = asmap->physInfo; 154 for (i = 0; i < nWndNodes; i++) { 155 rf_InitNode(&wndNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 156 RF_ASSERT(pda != NULL); 157 wndNodes[i].params[0].p = pda; 158 wndNodes[i].params[1].p = pda->bufPtr; 159 wndNodes[i].params[2].v = parityStripeID; 160 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 161 pda = pda->next; 162 } 163 164 /* initialize the redundancy node */ 165 rf_InitNode(xorNode, rf_wait, RF_TRUE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 1, 2 * (nWndNodes + nRodNodes) + 1, 1, dag_h, "Xr ", allocList); 166 xorNode->flags |= RF_DAGNODE_FLAG_YIELD; 167 for (i = 0; i < nWndNodes; i++) { 168 xorNode->params[2 * i + 0] = wndNodes[i].params[0]; /* pda */ 169 xorNode->params[2 * i + 1] = wndNodes[i].params[1]; /* buf ptr */ 170 } 171 for (i = 0; i < nRodNodes; i++) { 172 xorNode->params[2 * (nWndNodes + i) + 0] = rodNodes[i].params[0]; /* pda */ 173 xorNode->params[2 * (nWndNodes + i) + 1] = rodNodes[i].params[1]; /* buf ptr */ 174 } 175 xorNode->params[2 * (nWndNodes + nRodNodes)].p = raidPtr; /* xor node needs to get 176 * at RAID information */ 177 178 /* look for an Rod node that reads a complete SU. If none, alloc a 179 * buffer to receive the parity info. Note that we can't use a new 180 * data buffer because it will not have gotten written when the xor 181 * occurs. */ 182 for (i = 0; i < nRodNodes; i++) 183 if (((RF_PhysDiskAddr_t *) rodNodes[i].params[0].p)->numSector == raidPtr->Layout.sectorsPerStripeUnit) 184 break; 185 if (i == nRodNodes) { 186 RF_MallocAndAdd(xorNode->results[0], 187 rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit), (void *), allocList); 188 } else { 189 xorNode->results[0] = rodNodes[i].params[1].p; 190 } 191 192 /* initialize the Lpo node */ 193 rf_InitNode(lpoNode, rf_wait, RF_FALSE, rf_ParityLogOverwriteFunc, rf_ParityLogOverwriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpo", allocList); 194 195 lpoNode->params[0].p = asmap->parityInfo; 196 lpoNode->params[1].p = xorNode->results[0]; 197 RF_ASSERT(asmap->parityInfo->next == NULL); /* parityInfo must 198 * describe entire 199 * parity unit */ 200 201 /* connect nodes to form graph */ 202 203 /* connect dag header to block node */ 204 RF_ASSERT(dag_h->numSuccedents == 1); 205 RF_ASSERT(blockNode->numAntecedents == 0); 206 dag_h->succedents[0] = blockNode; 207 208 /* connect the block node to the Rod nodes */ 209 RF_ASSERT(blockNode->numSuccedents == nRodNodes + 1); 210 for (i = 0; i < nRodNodes; i++) { 211 RF_ASSERT(rodNodes[i].numAntecedents == 1); 212 blockNode->succedents[i] = &rodNodes[i]; 213 rodNodes[i].antecedents[0] = blockNode; 214 rodNodes[i].antType[0] = rf_control; 215 } 216 217 /* connect the block node to the sync node */ 218 /* necessary if nRodNodes == 0 */ 219 RF_ASSERT(syncNode->numAntecedents == nRodNodes + 1); 220 blockNode->succedents[nRodNodes] = syncNode; 221 syncNode->antecedents[0] = blockNode; 222 syncNode->antType[0] = rf_control; 223 224 /* connect the Rod nodes to the syncNode */ 225 for (i = 0; i < nRodNodes; i++) { 226 rodNodes[i].succedents[0] = syncNode; 227 syncNode->antecedents[1 + i] = &rodNodes[i]; 228 syncNode->antType[1 + i] = rf_control; 229 } 230 231 /* connect the sync node to the xor node */ 232 RF_ASSERT(syncNode->numSuccedents == nWndNodes + 1); 233 RF_ASSERT(xorNode->numAntecedents == 1); 234 syncNode->succedents[0] = xorNode; 235 xorNode->antecedents[0] = syncNode; 236 xorNode->antType[0] = rf_trueData; /* carry forward from sync */ 237 238 /* connect the sync node to the Wnd nodes */ 239 for (i = 0; i < nWndNodes; i++) { 240 RF_ASSERT(wndNodes->numAntecedents == 1); 241 syncNode->succedents[1 + i] = &wndNodes[i]; 242 wndNodes[i].antecedents[0] = syncNode; 243 wndNodes[i].antType[0] = rf_control; 244 } 245 246 /* connect the xor node to the Lpo node */ 247 RF_ASSERT(xorNode->numSuccedents == 1); 248 RF_ASSERT(lpoNode->numAntecedents == 1); 249 xorNode->succedents[0] = lpoNode; 250 lpoNode->antecedents[0] = xorNode; 251 lpoNode->antType[0] = rf_trueData; 252 253 /* connect the Wnd nodes to the unblock node */ 254 RF_ASSERT(unblockNode->numAntecedents == nWndNodes + 1); 255 for (i = 0; i < nWndNodes; i++) { 256 RF_ASSERT(wndNodes->numSuccedents == 1); 257 wndNodes[i].succedents[0] = unblockNode; 258 unblockNode->antecedents[i] = &wndNodes[i]; 259 unblockNode->antType[i] = rf_control; 260 } 261 262 /* connect the Lpo node to the unblock node */ 263 RF_ASSERT(lpoNode->numSuccedents == 1); 264 lpoNode->succedents[0] = unblockNode; 265 unblockNode->antecedents[nWndNodes] = lpoNode; 266 unblockNode->antType[nWndNodes] = rf_control; 267 268 /* connect unblock node to terminator */ 269 RF_ASSERT(unblockNode->numSuccedents == 1); 270 RF_ASSERT(termNode->numAntecedents == 1); 271 RF_ASSERT(termNode->numSuccedents == 0); 272 unblockNode->succedents[0] = termNode; 273 termNode->antecedents[0] = unblockNode; 274 termNode->antType[0] = rf_control; 275 } 276 277 278 279 280 /****************************************************************************** 281 * 282 * creates a DAG to perform a small-write operation (either raid 5 or pq), which is as follows: 283 * 284 * Header 285 * | 286 * Block 287 * / | ... \ \ 288 * / | \ \ 289 * Rod Rod Rod Rop 290 * | \ /| \ / | \/ | 291 * | | | /\ | 292 * Wnd Wnd Wnd X 293 * | \ / | 294 * | \ / | 295 * \ \ / Lpo 296 * \ \ / / 297 * +-> Unblock <-+ 298 * | 299 * T 300 * 301 * 302 * R = Read, W = Write, X = Xor, o = old, n = new, d = data, p = parity. 303 * When the access spans a stripe unit boundary and is less than one SU in size, there will 304 * be two Rop -- X -- Wnp branches. I call this the "double-XOR" case. 305 * The second output from each Rod node goes to the X node. In the double-XOR 306 * case, there are exactly 2 Rod nodes, and each sends one output to one X node. 307 * There is one Rod -- Wnd -- T branch for each stripe unit being updated. 308 * 309 * The block and unblock nodes are unused. See comment above CreateFaultFreeReadDAG. 310 * 311 * Note: this DAG ignores all the optimizations related to making the RMWs atomic. 312 * it also has the nasty property that none of the buffers allocated for reading 313 * old data & parity can be freed until the XOR node fires. Need to fix this. 314 * 315 * A null qfuncs indicates single fault tolerant 316 *****************************************************************************/ 317 318 void 319 rf_CommonCreateParityLoggingSmallWriteDAG( 320 RF_Raid_t * raidPtr, 321 RF_AccessStripeMap_t * asmap, 322 RF_DagHeader_t * dag_h, 323 void *bp, 324 RF_RaidAccessFlags_t flags, 325 RF_AllocListElem_t * allocList, 326 const RF_RedFuncs_t * pfuncs, 327 const RF_RedFuncs_t * qfuncs) 328 { 329 RF_DagNode_t *xorNodes, *blockNode, *unblockNode, *nodes; 330 RF_DagNode_t *readDataNodes, *readParityNodes; 331 RF_DagNode_t *writeDataNodes, *lpuNodes; 332 RF_DagNode_t *termNode; 333 RF_PhysDiskAddr_t *pda = asmap->physInfo; 334 int numDataNodes = asmap->numStripeUnitsAccessed; 335 int numParityNodes = (asmap->parityInfo->next) ? 2 : 1; 336 int i, j, nNodes, totalNumNodes; 337 RF_ReconUnitNum_t which_ru; 338 int (*func) (RF_DagNode_t * node), (*undoFunc) (RF_DagNode_t * node); 339 int (*qfunc) (RF_DagNode_t * node); 340 const char *name, *qname; 341 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru); 342 #ifdef RAID_DIAGNOSTIC 343 long nfaults = qfuncs ? 2 : 1; 344 #endif /* RAID_DIAGNOSTIC */ 345 346 if (rf_dagDebug) 347 printf("[Creating parity-logging small-write DAG]\n"); 348 RF_ASSERT(numDataNodes > 0); 349 RF_ASSERT(nfaults == 1); 350 dag_h->creator = "ParityLoggingSmallWriteDAG"; 351 352 /* DAG creation occurs in three steps: 1. count the number of nodes in 353 * the DAG 2. create the nodes 3. initialize the nodes 4. connect the 354 * nodes */ 355 356 /* Step 1. compute number of nodes in the graph */ 357 358 /* number of nodes: a read and write for each data unit a redundancy 359 * computation node for each parity node a read and Lpu for each 360 * parity unit a block and unblock node (2) a terminator node if 361 * atomic RMW an unlock node for each data unit, redundancy unit */ 362 totalNumNodes = (2 * numDataNodes) + numParityNodes + (2 * numParityNodes) + 3; 363 364 nNodes = numDataNodes + numParityNodes; 365 366 dag_h->numCommitNodes = numDataNodes + numParityNodes; 367 dag_h->numCommits = 0; 368 dag_h->numSuccedents = 1; 369 370 /* Step 2. create the nodes */ 371 RF_MallocAndAdd(nodes, totalNumNodes * sizeof(RF_DagNode_t), 372 (RF_DagNode_t *), allocList); 373 i = 0; 374 blockNode = &nodes[i]; 375 i += 1; 376 unblockNode = &nodes[i]; 377 i += 1; 378 readDataNodes = &nodes[i]; 379 i += numDataNodes; 380 readParityNodes = &nodes[i]; 381 i += numParityNodes; 382 writeDataNodes = &nodes[i]; 383 i += numDataNodes; 384 lpuNodes = &nodes[i]; 385 i += numParityNodes; 386 xorNodes = &nodes[i]; 387 i += numParityNodes; 388 termNode = &nodes[i]; 389 i += 1; 390 391 RF_ASSERT(i == totalNumNodes); 392 393 /* Step 3. initialize the nodes */ 394 /* initialize block node (Nil) */ 395 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nNodes, 0, 0, 0, dag_h, "Nil", allocList); 396 397 /* initialize unblock node (Nil) */ 398 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nNodes, 0, 0, dag_h, "Nil", allocList); 399 400 /* initialize terminatory node (Trm) */ 401 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 402 403 /* initialize nodes which read old data (Rod) */ 404 for (i = 0; i < numDataNodes; i++) { 405 rf_InitNode(&readDataNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rod", allocList); 406 RF_ASSERT(pda != NULL); 407 readDataNodes[i].params[0].p = pda; /* physical disk addr 408 * desc */ 409 readDataNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda->numSector << raidPtr->logBytesPerSector); /* buffer to hold old data */ 410 readDataNodes[i].params[2].v = parityStripeID; 411 readDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, which_ru); 412 pda = pda->next; 413 readDataNodes[i].propList[0] = NULL; 414 readDataNodes[i].propList[1] = NULL; 415 } 416 417 /* initialize nodes which read old parity (Rop) */ 418 pda = asmap->parityInfo; 419 i = 0; 420 for (i = 0; i < numParityNodes; i++) { 421 RF_ASSERT(pda != NULL); 422 rf_InitNode(&readParityNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rop", allocList); 423 readParityNodes[i].params[0].p = pda; 424 readParityNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda->numSector << raidPtr->logBytesPerSector); /* buffer to hold old parity */ 425 readParityNodes[i].params[2].v = parityStripeID; 426 readParityNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 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, which_ru); 442 443 pda = pda->next; 444 } 445 446 447 /* initialize nodes which compute new parity */ 448 /* we use the simple XOR func in the double-XOR case, and when we're 449 * accessing only a portion of one stripe unit. the distinction 450 * between the two is that the regular XOR func assumes that the 451 * targbuf is a full SU in size, and examines the pda associated with 452 * the buffer to decide where within the buffer to XOR the data, 453 * whereas the simple XOR func just XORs the data into the start of 454 * the buffer. */ 455 if ((numParityNodes == 2) || ((numDataNodes == 1) && (asmap->totalSectorsAccessed < raidPtr->Layout.sectorsPerStripeUnit))) { 456 func = pfuncs->simple; 457 undoFunc = rf_NullNodeUndoFunc; 458 name = pfuncs->SimpleName; 459 if (qfuncs) { 460 qfunc = qfuncs->simple; 461 qname = qfuncs->SimpleName; 462 } 463 } else { 464 func = pfuncs->regular; 465 undoFunc = rf_NullNodeUndoFunc; 466 name = pfuncs->RegularName; 467 if (qfuncs) { 468 qfunc = qfuncs->regular; 469 qname = qfuncs->RegularName; 470 } 471 } 472 /* initialize the xor nodes: params are {pda,buf} from {Rod,Wnd,Rop} 473 * nodes, and raidPtr */ 474 if (numParityNodes == 2) { /* double-xor case */ 475 for (i = 0; i < numParityNodes; i++) { 476 rf_InitNode(&xorNodes[i], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, 7, 1, dag_h, name, allocList); /* no wakeup func for 477 * xor */ 478 xorNodes[i].flags |= RF_DAGNODE_FLAG_YIELD; 479 xorNodes[i].params[0] = readDataNodes[i].params[0]; 480 xorNodes[i].params[1] = readDataNodes[i].params[1]; 481 xorNodes[i].params[2] = readParityNodes[i].params[0]; 482 xorNodes[i].params[3] = readParityNodes[i].params[1]; 483 xorNodes[i].params[4] = writeDataNodes[i].params[0]; 484 xorNodes[i].params[5] = writeDataNodes[i].params[1]; 485 xorNodes[i].params[6].p = raidPtr; 486 xorNodes[i].results[0] = readParityNodes[i].params[1].p; /* use old parity buf as 487 * target buf */ 488 } 489 } else { 490 /* there is only one xor node in this case */ 491 rf_InitNode(&xorNodes[0], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, (2 * (numDataNodes + numDataNodes + 1) + 1), 1, dag_h, name, allocList); 492 xorNodes[0].flags |= RF_DAGNODE_FLAG_YIELD; 493 for (i = 0; i < numDataNodes + 1; i++) { 494 /* set up params related to Rod and Rop nodes */ 495 xorNodes[0].params[2 * i + 0] = readDataNodes[i].params[0]; /* pda */ 496 xorNodes[0].params[2 * i + 1] = readDataNodes[i].params[1]; /* buffer pointer */ 497 } 498 for (i = 0; i < numDataNodes; i++) { 499 /* set up params related to Wnd and Wnp nodes */ 500 xorNodes[0].params[2 * (numDataNodes + 1 + i) + 0] = writeDataNodes[i].params[0]; /* pda */ 501 xorNodes[0].params[2 * (numDataNodes + 1 + i) + 1] = writeDataNodes[i].params[1]; /* buffer pointer */ 502 } 503 xorNodes[0].params[2 * (numDataNodes + numDataNodes + 1)].p = raidPtr; /* xor node needs to get 504 * at RAID information */ 505 xorNodes[0].results[0] = readParityNodes[0].params[1].p; 506 } 507 508 /* initialize the log node(s) */ 509 pda = asmap->parityInfo; 510 for (i = 0; i < numParityNodes; i++) { 511 RF_ASSERT(pda); 512 rf_InitNode(&lpuNodes[i], rf_wait, RF_FALSE, rf_ParityLogUpdateFunc, rf_ParityLogUpdateUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpu", allocList); 513 lpuNodes[i].params[0].p = pda; /* PhysDiskAddr of parity */ 514 lpuNodes[i].params[1].p = xorNodes[i].results[0]; /* buffer pointer to 515 * parity */ 516 pda = pda->next; 517 } 518 519 520 /* Step 4. connect the nodes */ 521 522 /* connect header to block node */ 523 RF_ASSERT(dag_h->numSuccedents == 1); 524 RF_ASSERT(blockNode->numAntecedents == 0); 525 dag_h->succedents[0] = blockNode; 526 527 /* connect block node to read old data nodes */ 528 RF_ASSERT(blockNode->numSuccedents == (numDataNodes + numParityNodes)); 529 for (i = 0; i < numDataNodes; i++) { 530 blockNode->succedents[i] = &readDataNodes[i]; 531 RF_ASSERT(readDataNodes[i].numAntecedents == 1); 532 readDataNodes[i].antecedents[0] = blockNode; 533 readDataNodes[i].antType[0] = rf_control; 534 } 535 536 /* connect block node to read old parity nodes */ 537 for (i = 0; i < numParityNodes; i++) { 538 blockNode->succedents[numDataNodes + i] = &readParityNodes[i]; 539 RF_ASSERT(readParityNodes[i].numAntecedents == 1); 540 readParityNodes[i].antecedents[0] = blockNode; 541 readParityNodes[i].antType[0] = rf_control; 542 } 543 544 /* connect read old data nodes to write new data nodes */ 545 for (i = 0; i < numDataNodes; i++) { 546 RF_ASSERT(readDataNodes[i].numSuccedents == numDataNodes + numParityNodes); 547 for (j = 0; j < numDataNodes; j++) { 548 RF_ASSERT(writeDataNodes[j].numAntecedents == numDataNodes + numParityNodes); 549 readDataNodes[i].succedents[j] = &writeDataNodes[j]; 550 writeDataNodes[j].antecedents[i] = &readDataNodes[i]; 551 if (i == j) 552 writeDataNodes[j].antType[i] = rf_antiData; 553 else 554 writeDataNodes[j].antType[i] = rf_control; 555 } 556 } 557 558 /* connect read old data nodes to xor nodes */ 559 for (i = 0; i < numDataNodes; i++) 560 for (j = 0; j < numParityNodes; j++) { 561 RF_ASSERT(xorNodes[j].numAntecedents == numDataNodes + numParityNodes); 562 readDataNodes[i].succedents[numDataNodes + j] = &xorNodes[j]; 563 xorNodes[j].antecedents[i] = &readDataNodes[i]; 564 xorNodes[j].antType[i] = rf_trueData; 565 } 566 567 /* connect read old parity nodes to write new data nodes */ 568 for (i = 0; i < numParityNodes; i++) { 569 RF_ASSERT(readParityNodes[i].numSuccedents == numDataNodes + numParityNodes); 570 for (j = 0; j < numDataNodes; j++) { 571 readParityNodes[i].succedents[j] = &writeDataNodes[j]; 572 writeDataNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i]; 573 writeDataNodes[j].antType[numDataNodes + i] = rf_control; 574 } 575 } 576 577 /* connect read old parity nodes to xor nodes */ 578 for (i = 0; i < numParityNodes; i++) 579 for (j = 0; j < numParityNodes; j++) { 580 readParityNodes[i].succedents[numDataNodes + j] = &xorNodes[j]; 581 xorNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i]; 582 xorNodes[j].antType[numDataNodes + i] = rf_trueData; 583 } 584 585 /* connect xor nodes to write new parity nodes */ 586 for (i = 0; i < numParityNodes; i++) { 587 RF_ASSERT(xorNodes[i].numSuccedents == 1); 588 RF_ASSERT(lpuNodes[i].numAntecedents == 1); 589 xorNodes[i].succedents[0] = &lpuNodes[i]; 590 lpuNodes[i].antecedents[0] = &xorNodes[i]; 591 lpuNodes[i].antType[0] = rf_trueData; 592 } 593 594 for (i = 0; i < numDataNodes; i++) { 595 /* connect write new data nodes to unblock node */ 596 RF_ASSERT(writeDataNodes[i].numSuccedents == 1); 597 RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes))); 598 writeDataNodes[i].succedents[0] = unblockNode; 599 unblockNode->antecedents[i] = &writeDataNodes[i]; 600 unblockNode->antType[i] = rf_control; 601 } 602 603 /* connect write new parity nodes to unblock node */ 604 for (i = 0; i < numParityNodes; i++) { 605 RF_ASSERT(lpuNodes[i].numSuccedents == 1); 606 lpuNodes[i].succedents[0] = unblockNode; 607 unblockNode->antecedents[numDataNodes + i] = &lpuNodes[i]; 608 unblockNode->antType[numDataNodes + i] = rf_control; 609 } 610 611 /* connect unblock node to terminator */ 612 RF_ASSERT(unblockNode->numSuccedents == 1); 613 RF_ASSERT(termNode->numAntecedents == 1); 614 RF_ASSERT(termNode->numSuccedents == 0); 615 unblockNode->succedents[0] = termNode; 616 termNode->antecedents[0] = unblockNode; 617 termNode->antType[0] = rf_control; 618 } 619 620 621 void 622 rf_CreateParityLoggingSmallWriteDAG( 623 RF_Raid_t * raidPtr, 624 RF_AccessStripeMap_t * asmap, 625 RF_DagHeader_t * dag_h, 626 void *bp, 627 RF_RaidAccessFlags_t flags, 628 RF_AllocListElem_t * allocList, 629 const RF_RedFuncs_t * pfuncs, 630 const RF_RedFuncs_t * qfuncs) 631 { 632 dag_h->creator = "ParityLoggingSmallWriteDAG"; 633 rf_CommonCreateParityLoggingSmallWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, &rf_xorFuncs, NULL); 634 } 635 636 637 void 638 rf_CreateParityLoggingLargeWriteDAG( 639 RF_Raid_t * raidPtr, 640 RF_AccessStripeMap_t * asmap, 641 RF_DagHeader_t * dag_h, 642 void *bp, 643 RF_RaidAccessFlags_t flags, 644 RF_AllocListElem_t * allocList, 645 int nfaults, 646 int (*redFunc) (RF_DagNode_t *)) 647 { 648 dag_h->creator = "ParityLoggingSmallWriteDAG"; 649 rf_CommonCreateParityLoggingLargeWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 1, rf_RegularXorFunc); 650 } 651 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 652
Indexes created Thu Sep 25 02:09:54 GMT 2025