1 /* $NetBSD: rf_parityloggingdags.c,v 1.7 2001/11/13 07:11:15 lukem 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 /* 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 1.7 2001/11/13 07:11:15 lukem Exp $"); 35 36 #include "rf_archs.h" 37 38 #if RF_INCLUDE_PARITYLOGGING > 0 39 40 #include <dev/raidframe/raidframevar.h> 41 42 #include "rf_raid.h" 43 #include "rf_dag.h" 44 #include "rf_dagutils.h" 45 #include "rf_dagfuncs.h" 46 #include "rf_debugMem.h" 47 #include "rf_paritylog.h" 48 #include "rf_memchunk.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_CallocAndAdd(nodes, nWndNodes + 6, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 103 i = 0; 104 wndNodes = &nodes[i]; 105 i += nWndNodes; 106 xorNode = &nodes[i]; 107 i += 1; 108 lpoNode = &nodes[i]; 109 i += 1; 110 blockNode = &nodes[i]; 111 i += 1; 112 syncNode = &nodes[i]; 113 i += 1; 114 unblockNode = &nodes[i]; 115 i += 1; 116 termNode = &nodes[i]; 117 i += 1; 118 119 dag_h->numCommitNodes = nWndNodes + 1; 120 dag_h->numCommits = 0; 121 dag_h->numSuccedents = 1; 122 123 rf_MapUnaccessedPortionOfStripe(raidPtr, layoutPtr, asmap, dag_h, new_asm_h, &nRodNodes, &sosBuffer, &eosBuffer, allocList); 124 if (nRodNodes > 0) 125 RF_CallocAndAdd(rodNodes, nRodNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 126 127 /* begin node initialization */ 128 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nRodNodes + 1, 0, 0, 0, dag_h, "Nil", allocList); 129 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nWndNodes + 1, 0, 0, dag_h, "Nil", allocList); 130 rf_InitNode(syncNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nWndNodes + 1, nRodNodes + 1, 0, 0, dag_h, "Nil", allocList); 131 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 132 133 /* initialize the Rod nodes */ 134 for (nodeNum = asmNum = 0; asmNum < 2; asmNum++) { 135 if (new_asm_h[asmNum]) { 136 pda = new_asm_h[asmNum]->stripeMap->physInfo; 137 while (pda) { 138 rf_InitNode(&rodNodes[nodeNum], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rod", allocList); 139 rodNodes[nodeNum].params[0].p = pda; 140 rodNodes[nodeNum].params[1].p = pda->bufPtr; 141 rodNodes[nodeNum].params[2].v = parityStripeID; 142 rodNodes[nodeNum].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 143 nodeNum++; 144 pda = pda->next; 145 } 146 } 147 } 148 RF_ASSERT(nodeNum == nRodNodes); 149 150 /* initialize the wnd nodes */ 151 pda = asmap->physInfo; 152 for (i = 0; i < nWndNodes; i++) { 153 rf_InitNode(&wndNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList); 154 RF_ASSERT(pda != NULL); 155 wndNodes[i].params[0].p = pda; 156 wndNodes[i].params[1].p = pda->bufPtr; 157 wndNodes[i].params[2].v = parityStripeID; 158 wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 159 pda = pda->next; 160 } 161 162 /* initialize the redundancy node */ 163 rf_InitNode(xorNode, rf_wait, RF_TRUE, redFunc, rf_NullNodeUndoFunc, NULL, 1, 1, 2 * (nWndNodes + nRodNodes) + 1, 1, dag_h, "Xr ", allocList); 164 xorNode->flags |= RF_DAGNODE_FLAG_YIELD; 165 for (i = 0; i < nWndNodes; i++) { 166 xorNode->params[2 * i + 0] = wndNodes[i].params[0]; /* pda */ 167 xorNode->params[2 * i + 1] = wndNodes[i].params[1]; /* buf ptr */ 168 } 169 for (i = 0; i < nRodNodes; i++) { 170 xorNode->params[2 * (nWndNodes + i) + 0] = rodNodes[i].params[0]; /* pda */ 171 xorNode->params[2 * (nWndNodes + i) + 1] = rodNodes[i].params[1]; /* buf ptr */ 172 } 173 xorNode->params[2 * (nWndNodes + nRodNodes)].p = raidPtr; /* xor node needs to get 174 * at RAID information */ 175 176 /* look for an Rod node that reads a complete SU. If none, alloc a 177 * buffer to receive the parity info. Note that we can't use a new 178 * data buffer because it will not have gotten written when the xor 179 * occurs. */ 180 for (i = 0; i < nRodNodes; i++) 181 if (((RF_PhysDiskAddr_t *) rodNodes[i].params[0].p)->numSector == raidPtr->Layout.sectorsPerStripeUnit) 182 break; 183 if (i == nRodNodes) { 184 RF_CallocAndAdd(xorNode->results[0], 1, rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit), (void *), allocList); 185 } else { 186 xorNode->results[0] = rodNodes[i].params[1].p; 187 } 188 189 /* initialize the Lpo node */ 190 rf_InitNode(lpoNode, rf_wait, RF_FALSE, rf_ParityLogOverwriteFunc, rf_ParityLogOverwriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpo", allocList); 191 192 lpoNode->params[0].p = asmap->parityInfo; 193 lpoNode->params[1].p = xorNode->results[0]; 194 RF_ASSERT(asmap->parityInfo->next == NULL); /* parityInfo must 195 * describe entire 196 * parity unit */ 197 198 /* connect nodes to form graph */ 199 200 /* connect dag header to block node */ 201 RF_ASSERT(dag_h->numSuccedents == 1); 202 RF_ASSERT(blockNode->numAntecedents == 0); 203 dag_h->succedents[0] = blockNode; 204 205 /* connect the block node to the Rod nodes */ 206 RF_ASSERT(blockNode->numSuccedents == nRodNodes + 1); 207 for (i = 0; i < nRodNodes; i++) { 208 RF_ASSERT(rodNodes[i].numAntecedents == 1); 209 blockNode->succedents[i] = &rodNodes[i]; 210 rodNodes[i].antecedents[0] = blockNode; 211 rodNodes[i].antType[0] = rf_control; 212 } 213 214 /* connect the block node to the sync node */ 215 /* necessary if nRodNodes == 0 */ 216 RF_ASSERT(syncNode->numAntecedents == nRodNodes + 1); 217 blockNode->succedents[nRodNodes] = syncNode; 218 syncNode->antecedents[0] = blockNode; 219 syncNode->antType[0] = rf_control; 220 221 /* connect the Rod nodes to the syncNode */ 222 for (i = 0; i < nRodNodes; i++) { 223 rodNodes[i].succedents[0] = syncNode; 224 syncNode->antecedents[1 + i] = &rodNodes[i]; 225 syncNode->antType[1 + i] = rf_control; 226 } 227 228 /* connect the sync node to the xor node */ 229 RF_ASSERT(syncNode->numSuccedents == nWndNodes + 1); 230 RF_ASSERT(xorNode->numAntecedents == 1); 231 syncNode->succedents[0] = xorNode; 232 xorNode->antecedents[0] = syncNode; 233 xorNode->antType[0] = rf_trueData; /* carry forward from sync */ 234 235 /* connect the sync node to the Wnd nodes */ 236 for (i = 0; i < nWndNodes; i++) { 237 RF_ASSERT(wndNodes->numAntecedents == 1); 238 syncNode->succedents[1 + i] = &wndNodes[i]; 239 wndNodes[i].antecedents[0] = syncNode; 240 wndNodes[i].antType[0] = rf_control; 241 } 242 243 /* connect the xor node to the Lpo node */ 244 RF_ASSERT(xorNode->numSuccedents == 1); 245 RF_ASSERT(lpoNode->numAntecedents == 1); 246 xorNode->succedents[0] = lpoNode; 247 lpoNode->antecedents[0] = xorNode; 248 lpoNode->antType[0] = rf_trueData; 249 250 /* connect the Wnd nodes to the unblock node */ 251 RF_ASSERT(unblockNode->numAntecedents == nWndNodes + 1); 252 for (i = 0; i < nWndNodes; i++) { 253 RF_ASSERT(wndNodes->numSuccedents == 1); 254 wndNodes[i].succedents[0] = unblockNode; 255 unblockNode->antecedents[i] = &wndNodes[i]; 256 unblockNode->antType[i] = rf_control; 257 } 258 259 /* connect the Lpo node to the unblock node */ 260 RF_ASSERT(lpoNode->numSuccedents == 1); 261 lpoNode->succedents[0] = unblockNode; 262 unblockNode->antecedents[nWndNodes] = lpoNode; 263 unblockNode->antType[nWndNodes] = rf_control; 264 265 /* connect unblock node to terminator */ 266 RF_ASSERT(unblockNode->numSuccedents == 1); 267 RF_ASSERT(termNode->numAntecedents == 1); 268 RF_ASSERT(termNode->numSuccedents == 0); 269 unblockNode->succedents[0] = termNode; 270 termNode->antecedents[0] = unblockNode; 271 termNode->antType[0] = rf_control; 272 } 273 274 275 276 277 /****************************************************************************** 278 * 279 * creates a DAG to perform a small-write operation (either raid 5 or pq), which is as follows: 280 * 281 * Header 282 * | 283 * Block 284 * / | ... \ \ 285 * / | \ \ 286 * Rod Rod Rod Rop 287 * | \ /| \ / | \/ | 288 * | | | /\ | 289 * Wnd Wnd Wnd X 290 * | \ / | 291 * | \ / | 292 * \ \ / Lpo 293 * \ \ / / 294 * +-> Unblock <-+ 295 * | 296 * T 297 * 298 * 299 * R = Read, W = Write, X = Xor, o = old, n = new, d = data, p = parity. 300 * When the access spans a stripe unit boundary and is less than one SU in size, there will 301 * be two Rop -- X -- Wnp branches. I call this the "double-XOR" case. 302 * The second output from each Rod node goes to the X node. In the double-XOR 303 * case, there are exactly 2 Rod nodes, and each sends one output to one X node. 304 * There is one Rod -- Wnd -- T branch for each stripe unit being updated. 305 * 306 * The block and unblock nodes are unused. See comment above CreateFaultFreeReadDAG. 307 * 308 * Note: this DAG ignores all the optimizations related to making the RMWs atomic. 309 * it also has the nasty property that none of the buffers allocated for reading 310 * old data & parity can be freed until the XOR node fires. Need to fix this. 311 * 312 * A null qfuncs indicates single fault tolerant 313 *****************************************************************************/ 314 315 void 316 rf_CommonCreateParityLoggingSmallWriteDAG( 317 RF_Raid_t * raidPtr, 318 RF_AccessStripeMap_t * asmap, 319 RF_DagHeader_t * dag_h, 320 void *bp, 321 RF_RaidAccessFlags_t flags, 322 RF_AllocListElem_t * allocList, 323 RF_RedFuncs_t * pfuncs, 324 RF_RedFuncs_t * qfuncs) 325 { 326 RF_DagNode_t *xorNodes, *blockNode, *unblockNode, *nodes; 327 RF_DagNode_t *readDataNodes, *readParityNodes; 328 RF_DagNode_t *writeDataNodes, *lpuNodes; 329 RF_DagNode_t *unlockDataNodes = NULL, *termNode; 330 RF_PhysDiskAddr_t *pda = asmap->physInfo; 331 int numDataNodes = asmap->numStripeUnitsAccessed; 332 int numParityNodes = (asmap->parityInfo->next) ? 2 : 1; 333 int i, j, nNodes, totalNumNodes; 334 RF_ReconUnitNum_t which_ru; 335 int (*func) (RF_DagNode_t * node), (*undoFunc) (RF_DagNode_t * node); 336 int (*qfunc) (RF_DagNode_t * node); 337 char *name, *qname; 338 RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout), asmap->raidAddress, &which_ru); 339 #ifdef RAID_DIAGNOSTIC 340 long nfaults = qfuncs ? 2 : 1; 341 #endif /* RAID_DIAGNOSTIC */ 342 int lu_flag = (rf_enableAtomicRMW) ? 1 : 0; /* lock/unlock flag */ 343 344 if (rf_dagDebug) 345 printf("[Creating parity-logging small-write DAG]\n"); 346 RF_ASSERT(numDataNodes > 0); 347 RF_ASSERT(nfaults == 1); 348 dag_h->creator = "ParityLoggingSmallWriteDAG"; 349 350 /* DAG creation occurs in three steps: 1. count the number of nodes in 351 * the DAG 2. create the nodes 3. initialize the nodes 4. connect the 352 * nodes */ 353 354 /* Step 1. compute number of nodes in the graph */ 355 356 /* number of nodes: a read and write for each data unit a redundancy 357 * computation node for each parity node a read and Lpu for each 358 * parity unit a block and unblock node (2) a terminator node if 359 * atomic RMW an unlock node for each data unit, redundancy unit */ 360 totalNumNodes = (2 * numDataNodes) + numParityNodes + (2 * numParityNodes) + 3; 361 if (lu_flag) 362 totalNumNodes += numDataNodes; 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_CallocAndAdd(nodes, totalNumNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList); 372 i = 0; 373 blockNode = &nodes[i]; 374 i += 1; 375 unblockNode = &nodes[i]; 376 i += 1; 377 readDataNodes = &nodes[i]; 378 i += numDataNodes; 379 readParityNodes = &nodes[i]; 380 i += numParityNodes; 381 writeDataNodes = &nodes[i]; 382 i += numDataNodes; 383 lpuNodes = &nodes[i]; 384 i += numParityNodes; 385 xorNodes = &nodes[i]; 386 i += numParityNodes; 387 termNode = &nodes[i]; 388 i += 1; 389 if (lu_flag) { 390 unlockDataNodes = &nodes[i]; 391 i += numDataNodes; 392 } 393 RF_ASSERT(i == totalNumNodes); 394 395 /* Step 3. initialize the nodes */ 396 /* initialize block node (Nil) */ 397 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, nNodes, 0, 0, 0, dag_h, "Nil", allocList); 398 399 /* initialize unblock node (Nil) */ 400 rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL, 1, nNodes, 0, 0, dag_h, "Nil", allocList); 401 402 /* initialize terminatory node (Trm) */ 403 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 404 405 /* initialize nodes which read old data (Rod) */ 406 for (i = 0; i < numDataNodes; i++) { 407 rf_InitNode(&readDataNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rod", allocList); 408 RF_ASSERT(pda != NULL); 409 readDataNodes[i].params[0].p = pda; /* physical disk addr 410 * desc */ 411 readDataNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda, allocList); /* buffer to hold old 412 * data */ 413 readDataNodes[i].params[2].v = parityStripeID; 414 readDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, lu_flag, 0, which_ru); 415 pda = pda->next; 416 readDataNodes[i].propList[0] = NULL; 417 readDataNodes[i].propList[1] = NULL; 418 } 419 420 /* initialize nodes which read old parity (Rop) */ 421 pda = asmap->parityInfo; 422 i = 0; 423 for (i = 0; i < numParityNodes; i++) { 424 RF_ASSERT(pda != NULL); 425 rf_InitNode(&readParityNodes[i], rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc, rf_GenericWakeupFunc, nNodes, 1, 4, 0, dag_h, "Rop", allocList); 426 readParityNodes[i].params[0].p = pda; 427 readParityNodes[i].params[1].p = rf_AllocBuffer(raidPtr, dag_h, pda, allocList); /* buffer to hold old 428 * parity */ 429 readParityNodes[i].params[2].v = parityStripeID; 430 readParityNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 431 readParityNodes[i].propList[0] = NULL; 432 pda = pda->next; 433 } 434 435 /* initialize nodes which write new data (Wnd) */ 436 pda = asmap->physInfo; 437 for (i = 0; i < numDataNodes; i++) { 438 RF_ASSERT(pda != NULL); 439 rf_InitNode(&writeDataNodes[i], rf_wait, RF_TRUE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, nNodes, 4, 0, dag_h, "Wnd", allocList); 440 writeDataNodes[i].params[0].p = pda; /* physical disk addr 441 * desc */ 442 writeDataNodes[i].params[1].p = pda->bufPtr; /* buffer holding new 443 * data to be written */ 444 writeDataNodes[i].params[2].v = parityStripeID; 445 writeDataNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru); 446 447 if (lu_flag) { 448 /* initialize node to unlock the disk queue */ 449 rf_InitNode(&unlockDataNodes[i], rf_wait, RF_FALSE, rf_DiskUnlockFunc, rf_DiskUnlockUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Und", allocList); 450 unlockDataNodes[i].params[0].p = pda; /* physical disk addr 451 * desc */ 452 unlockDataNodes[i].params[1].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, lu_flag, which_ru); 453 } 454 pda = pda->next; 455 } 456 457 458 /* initialize nodes which compute new parity */ 459 /* we use the simple XOR func in the double-XOR case, and when we're 460 * accessing only a portion of one stripe unit. the distinction 461 * between the two is that the regular XOR func assumes that the 462 * targbuf is a full SU in size, and examines the pda associated with 463 * the buffer to decide where within the buffer to XOR the data, 464 * whereas the simple XOR func just XORs the data into the start of 465 * the buffer. */ 466 if ((numParityNodes == 2) || ((numDataNodes == 1) && (asmap->totalSectorsAccessed < raidPtr->Layout.sectorsPerStripeUnit))) { 467 func = pfuncs->simple; 468 undoFunc = rf_NullNodeUndoFunc; 469 name = pfuncs->SimpleName; 470 if (qfuncs) { 471 qfunc = qfuncs->simple; 472 qname = qfuncs->SimpleName; 473 } 474 } else { 475 func = pfuncs->regular; 476 undoFunc = rf_NullNodeUndoFunc; 477 name = pfuncs->RegularName; 478 if (qfuncs) { 479 qfunc = qfuncs->regular; 480 qname = qfuncs->RegularName; 481 } 482 } 483 /* initialize the xor nodes: params are {pda,buf} from {Rod,Wnd,Rop} 484 * nodes, and raidPtr */ 485 if (numParityNodes == 2) { /* double-xor case */ 486 for (i = 0; i < numParityNodes; i++) { 487 rf_InitNode(&xorNodes[i], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, 7, 1, dag_h, name, allocList); /* no wakeup func for 488 * xor */ 489 xorNodes[i].flags |= RF_DAGNODE_FLAG_YIELD; 490 xorNodes[i].params[0] = readDataNodes[i].params[0]; 491 xorNodes[i].params[1] = readDataNodes[i].params[1]; 492 xorNodes[i].params[2] = readParityNodes[i].params[0]; 493 xorNodes[i].params[3] = readParityNodes[i].params[1]; 494 xorNodes[i].params[4] = writeDataNodes[i].params[0]; 495 xorNodes[i].params[5] = writeDataNodes[i].params[1]; 496 xorNodes[i].params[6].p = raidPtr; 497 xorNodes[i].results[0] = readParityNodes[i].params[1].p; /* use old parity buf as 498 * target buf */ 499 } 500 } else { 501 /* there is only one xor node in this case */ 502 rf_InitNode(&xorNodes[0], rf_wait, RF_TRUE, func, undoFunc, NULL, 1, nNodes, (2 * (numDataNodes + numDataNodes + 1) + 1), 1, dag_h, name, allocList); 503 xorNodes[0].flags |= RF_DAGNODE_FLAG_YIELD; 504 for (i = 0; i < numDataNodes + 1; i++) { 505 /* set up params related to Rod and Rop nodes */ 506 xorNodes[0].params[2 * i + 0] = readDataNodes[i].params[0]; /* pda */ 507 xorNodes[0].params[2 * i + 1] = readDataNodes[i].params[1]; /* buffer pointer */ 508 } 509 for (i = 0; i < numDataNodes; i++) { 510 /* set up params related to Wnd and Wnp nodes */ 511 xorNodes[0].params[2 * (numDataNodes + 1 + i) + 0] = writeDataNodes[i].params[0]; /* pda */ 512 xorNodes[0].params[2 * (numDataNodes + 1 + i) + 1] = writeDataNodes[i].params[1]; /* buffer pointer */ 513 } 514 xorNodes[0].params[2 * (numDataNodes + numDataNodes + 1)].p = raidPtr; /* xor node needs to get 515 * at RAID information */ 516 xorNodes[0].results[0] = readParityNodes[0].params[1].p; 517 } 518 519 /* initialize the log node(s) */ 520 pda = asmap->parityInfo; 521 for (i = 0; i < numParityNodes; i++) { 522 RF_ASSERT(pda); 523 rf_InitNode(&lpuNodes[i], rf_wait, RF_FALSE, rf_ParityLogUpdateFunc, rf_ParityLogUpdateUndoFunc, rf_GenericWakeupFunc, 1, 1, 2, 0, dag_h, "Lpu", allocList); 524 lpuNodes[i].params[0].p = pda; /* PhysDiskAddr of parity */ 525 lpuNodes[i].params[1].p = xorNodes[i].results[0]; /* buffer pointer to 526 * parity */ 527 pda = pda->next; 528 } 529 530 531 /* Step 4. connect the nodes */ 532 533 /* connect header to block node */ 534 RF_ASSERT(dag_h->numSuccedents == 1); 535 RF_ASSERT(blockNode->numAntecedents == 0); 536 dag_h->succedents[0] = blockNode; 537 538 /* connect block node to read old data nodes */ 539 RF_ASSERT(blockNode->numSuccedents == (numDataNodes + numParityNodes)); 540 for (i = 0; i < numDataNodes; i++) { 541 blockNode->succedents[i] = &readDataNodes[i]; 542 RF_ASSERT(readDataNodes[i].numAntecedents == 1); 543 readDataNodes[i].antecedents[0] = blockNode; 544 readDataNodes[i].antType[0] = rf_control; 545 } 546 547 /* connect block node to read old parity nodes */ 548 for (i = 0; i < numParityNodes; i++) { 549 blockNode->succedents[numDataNodes + i] = &readParityNodes[i]; 550 RF_ASSERT(readParityNodes[i].numAntecedents == 1); 551 readParityNodes[i].antecedents[0] = blockNode; 552 readParityNodes[i].antType[0] = rf_control; 553 } 554 555 /* connect read old data nodes to write new data nodes */ 556 for (i = 0; i < numDataNodes; i++) { 557 RF_ASSERT(readDataNodes[i].numSuccedents == numDataNodes + numParityNodes); 558 for (j = 0; j < numDataNodes; j++) { 559 RF_ASSERT(writeDataNodes[j].numAntecedents == numDataNodes + numParityNodes); 560 readDataNodes[i].succedents[j] = &writeDataNodes[j]; 561 writeDataNodes[j].antecedents[i] = &readDataNodes[i]; 562 if (i == j) 563 writeDataNodes[j].antType[i] = rf_antiData; 564 else 565 writeDataNodes[j].antType[i] = rf_control; 566 } 567 } 568 569 /* connect read old data nodes to xor nodes */ 570 for (i = 0; i < numDataNodes; i++) 571 for (j = 0; j < numParityNodes; j++) { 572 RF_ASSERT(xorNodes[j].numAntecedents == numDataNodes + numParityNodes); 573 readDataNodes[i].succedents[numDataNodes + j] = &xorNodes[j]; 574 xorNodes[j].antecedents[i] = &readDataNodes[i]; 575 xorNodes[j].antType[i] = rf_trueData; 576 } 577 578 /* connect read old parity nodes to write new data nodes */ 579 for (i = 0; i < numParityNodes; i++) { 580 RF_ASSERT(readParityNodes[i].numSuccedents == numDataNodes + numParityNodes); 581 for (j = 0; j < numDataNodes; j++) { 582 readParityNodes[i].succedents[j] = &writeDataNodes[j]; 583 writeDataNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i]; 584 writeDataNodes[j].antType[numDataNodes + i] = rf_control; 585 } 586 } 587 588 /* connect read old parity nodes to xor nodes */ 589 for (i = 0; i < numParityNodes; i++) 590 for (j = 0; j < numParityNodes; j++) { 591 readParityNodes[i].succedents[numDataNodes + j] = &xorNodes[j]; 592 xorNodes[j].antecedents[numDataNodes + i] = &readParityNodes[i]; 593 xorNodes[j].antType[numDataNodes + i] = rf_trueData; 594 } 595 596 /* connect xor nodes to write new parity nodes */ 597 for (i = 0; i < numParityNodes; i++) { 598 RF_ASSERT(xorNodes[i].numSuccedents == 1); 599 RF_ASSERT(lpuNodes[i].numAntecedents == 1); 600 xorNodes[i].succedents[0] = &lpuNodes[i]; 601 lpuNodes[i].antecedents[0] = &xorNodes[i]; 602 lpuNodes[i].antType[0] = rf_trueData; 603 } 604 605 for (i = 0; i < numDataNodes; i++) { 606 if (lu_flag) { 607 /* connect write new data nodes to unlock nodes */ 608 RF_ASSERT(writeDataNodes[i].numSuccedents == 1); 609 RF_ASSERT(unlockDataNodes[i].numAntecedents == 1); 610 writeDataNodes[i].succedents[0] = &unlockDataNodes[i]; 611 unlockDataNodes[i].antecedents[0] = &writeDataNodes[i]; 612 unlockDataNodes[i].antType[0] = rf_control; 613 614 /* connect unlock nodes to unblock node */ 615 RF_ASSERT(unlockDataNodes[i].numSuccedents == 1); 616 RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes))); 617 unlockDataNodes[i].succedents[0] = unblockNode; 618 unblockNode->antecedents[i] = &unlockDataNodes[i]; 619 unblockNode->antType[i] = rf_control; 620 } else { 621 /* connect write new data nodes to unblock node */ 622 RF_ASSERT(writeDataNodes[i].numSuccedents == 1); 623 RF_ASSERT(unblockNode->numAntecedents == (numDataNodes + (nfaults * numParityNodes))); 624 writeDataNodes[i].succedents[0] = unblockNode; 625 unblockNode->antecedents[i] = &writeDataNodes[i]; 626 unblockNode->antType[i] = rf_control; 627 } 628 } 629 630 /* connect write new parity nodes to unblock node */ 631 for (i = 0; i < numParityNodes; i++) { 632 RF_ASSERT(lpuNodes[i].numSuccedents == 1); 633 lpuNodes[i].succedents[0] = unblockNode; 634 unblockNode->antecedents[numDataNodes + i] = &lpuNodes[i]; 635 unblockNode->antType[numDataNodes + i] = rf_control; 636 } 637 638 /* connect unblock node to terminator */ 639 RF_ASSERT(unblockNode->numSuccedents == 1); 640 RF_ASSERT(termNode->numAntecedents == 1); 641 RF_ASSERT(termNode->numSuccedents == 0); 642 unblockNode->succedents[0] = termNode; 643 termNode->antecedents[0] = unblockNode; 644 termNode->antType[0] = rf_control; 645 } 646 647 648 void 649 rf_CreateParityLoggingSmallWriteDAG( 650 RF_Raid_t * raidPtr, 651 RF_AccessStripeMap_t * asmap, 652 RF_DagHeader_t * dag_h, 653 void *bp, 654 RF_RaidAccessFlags_t flags, 655 RF_AllocListElem_t * allocList, 656 RF_RedFuncs_t * pfuncs, 657 RF_RedFuncs_t * qfuncs) 658 { 659 dag_h->creator = "ParityLoggingSmallWriteDAG"; 660 rf_CommonCreateParityLoggingSmallWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, &rf_xorFuncs, NULL); 661 } 662 663 664 void 665 rf_CreateParityLoggingLargeWriteDAG( 666 RF_Raid_t * raidPtr, 667 RF_AccessStripeMap_t * asmap, 668 RF_DagHeader_t * dag_h, 669 void *bp, 670 RF_RaidAccessFlags_t flags, 671 RF_AllocListElem_t * allocList, 672 int nfaults, 673 int (*redFunc) (RF_DagNode_t *)) 674 { 675 dag_h->creator = "ParityLoggingSmallWriteDAG"; 676 rf_CommonCreateParityLoggingLargeWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 1, rf_RegularXorFunc); 677 } 678 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 679