1 /* $NetBSD: rf_dagffrd.c,v 1.21 2019/10/10 03:43:59 christos 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_dagffrd.c 31 * 32 * code for creating fault-free read DAGs 33 * 34 */ 35 36 #include <sys/cdefs.h> 37 __KERNEL_RCSID(0, "$NetBSD: rf_dagffrd.c,v 1.21 2019/10/10 03:43:59 christos 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_dagffrd.h" 48 49 /****************************************************************************** 50 * 51 * General comments on DAG creation: 52 * 53 * All DAGs in this file use roll-away error recovery. Each DAG has a single 54 * commit node, usually called "Cmt." If an error occurs before the Cmt node 55 * is reached, the execution engine will halt forward execution and work 56 * backward through the graph, executing the undo functions. Assuming that 57 * each node in the graph prior to the Cmt node are undoable and atomic - or - 58 * does not make changes to permanent state, the graph will fail atomically. 59 * If an error occurs after the Cmt node executes, the engine will roll-forward 60 * through the graph, blindly executing nodes until it reaches the end. 61 * If a graph reaches the end, it is assumed to have completed successfully. 62 * 63 * A graph has only 1 Cmt node. 64 * 65 */ 66 67 68 /****************************************************************************** 69 * 70 * The following wrappers map the standard DAG creation interface to the 71 * DAG creation routines. Additionally, these wrappers enable experimentation 72 * with new DAG structures by providing an extra level of indirection, allowing 73 * the DAG creation routines to be replaced at this single point. 74 */ 75 76 void 77 rf_CreateFaultFreeReadDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 78 RF_DagHeader_t *dag_h, void *bp, 79 RF_RaidAccessFlags_t flags, 80 RF_AllocListElem_t *allocList) 81 { 82 rf_CreateNonredundantDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 83 RF_IO_TYPE_READ); 84 } 85 86 87 /****************************************************************************** 88 * 89 * DAG creation code begins here 90 */ 91 92 /****************************************************************************** 93 * 94 * creates a DAG to perform a nonredundant read or write of data within one 95 * stripe. 96 * For reads, this DAG is as follows: 97 * 98 * /---- read ----\ 99 * Header -- Block ---- read ---- Commit -- Terminate 100 * \---- read ----/ 101 * 102 * For writes, this DAG is as follows: 103 * 104 * /---- write ----\ 105 * Header -- Commit ---- write ---- Block -- Terminate 106 * \---- write ----/ 107 * 108 * There is one disk node per stripe unit accessed, and all disk nodes are in 109 * parallel. 110 * 111 * Tricky point here: The first disk node (read or write) is created 112 * normally. Subsequent disk nodes are created by copying the first one, 113 * and modifying a few params. The "succedents" and "antecedents" fields are 114 * _not_ re-created in each node, but rather left pointing to the same array 115 * that was malloc'd when the first node was created. Thus, it's essential 116 * that when this DAG is freed, the succedents and antecedents fields be freed 117 * in ONLY ONE of the read nodes. This does not apply to the "params" field 118 * because it is recreated for each READ node. 119 * 120 * Note that normal-priority accesses do not need to be tagged with their 121 * parity stripe ID, because they will never be promoted. Hence, I've 122 * commented-out the code to do this, and marked it with UNNEEDED. 123 * 124 *****************************************************************************/ 125 126 void 127 rf_CreateNonredundantDAG(RF_Raid_t *raidPtr, 128 RF_AccessStripeMap_t *asmap, RF_DagHeader_t *dag_h, void *bp, 129 RF_RaidAccessFlags_t flags, RF_AllocListElem_t *allocList, 130 RF_IoType_t type) 131 { 132 RF_DagNode_t *diskNodes, *blockNode, *commitNode, *termNode; 133 RF_DagNode_t *tmpNode, *tmpdiskNode; 134 RF_PhysDiskAddr_t *pda = asmap->physInfo; 135 void (*doFunc) (RF_DagNode_t *), (*undoFunc) (RF_DagNode_t *); 136 int i, n; 137 const char *name; 138 139 n = asmap->numStripeUnitsAccessed; 140 dag_h->creator = "NonredundantDAG"; 141 142 doFunc = rf_NullNodeFunc; 143 undoFunc = rf_NullNodeUndoFunc; 144 name = NULL; 145 146 RF_ASSERT(RF_IO_IS_R_OR_W(type)); 147 switch (type) { 148 case RF_IO_TYPE_READ: 149 doFunc = rf_DiskReadFunc; 150 undoFunc = rf_DiskReadUndoFunc; 151 name = "R "; 152 #if RF_DEBUG_DAG 153 if (rf_dagDebug) 154 printf("[Creating non-redundant read DAG]\n"); 155 #endif 156 break; 157 case RF_IO_TYPE_WRITE: 158 doFunc = rf_DiskWriteFunc; 159 undoFunc = rf_DiskWriteUndoFunc; 160 name = "W "; 161 #if RF_DEBUG_DAG 162 if (rf_dagDebug) 163 printf("[Creating non-redundant write DAG]\n"); 164 #endif 165 break; 166 default: 167 RF_PANIC(); 168 } 169 170 /* 171 * For reads, the dag can not commit until the block node is reached. 172 * for writes, the dag commits immediately. 173 */ 174 dag_h->numCommitNodes = 1; 175 dag_h->numCommits = 0; 176 dag_h->numSuccedents = 1; 177 178 /* 179 * Node count: 180 * 1 block node 181 * n data reads (or writes) 182 * 1 commit node 183 * 1 terminator node 184 */ 185 RF_ASSERT(n > 0); 186 187 for (i = 0; i < n; i++) { 188 tmpNode = rf_AllocDAGNode(); 189 tmpNode->list_next = dag_h->nodes; 190 dag_h->nodes = tmpNode; 191 } 192 diskNodes = dag_h->nodes; 193 194 blockNode = rf_AllocDAGNode(); 195 blockNode->list_next = dag_h->nodes; 196 dag_h->nodes = blockNode; 197 198 commitNode = rf_AllocDAGNode(); 199 commitNode->list_next = dag_h->nodes; 200 dag_h->nodes = commitNode; 201 202 termNode = rf_AllocDAGNode(); 203 termNode->list_next = dag_h->nodes; 204 dag_h->nodes = termNode; 205 206 /* initialize nodes */ 207 switch (type) { 208 case RF_IO_TYPE_READ: 209 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 210 NULL, n, 0, 0, 0, dag_h, "Nil", allocList); 211 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 212 NULL, 1, n, 0, 0, dag_h, "Cmt", allocList); 213 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 214 NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 215 break; 216 case RF_IO_TYPE_WRITE: 217 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 218 NULL, 1, 0, 0, 0, dag_h, "Nil", allocList); 219 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, 220 NULL, n, 1, 0, 0, dag_h, "Cmt", allocList); 221 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, 222 NULL, 0, n, 0, 0, dag_h, "Trm", allocList); 223 break; 224 default: 225 RF_PANIC(); 226 } 227 228 tmpdiskNode = diskNodes; 229 for (i = 0; i < n; i++) { 230 RF_ASSERT(pda != NULL); 231 rf_InitNode(tmpdiskNode, rf_wait, RF_FALSE, doFunc, undoFunc, rf_GenericWakeupFunc, 232 1, 1, 4, 0, dag_h, name, allocList); 233 tmpdiskNode->params[0].p = pda; 234 tmpdiskNode->params[1].p = pda->bufPtr; 235 /* parity stripe id is not necessary */ 236 tmpdiskNode->params[2].v = 0; 237 tmpdiskNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0); 238 pda = pda->next; 239 tmpdiskNode = tmpdiskNode->list_next; 240 } 241 242 /* 243 * Connect nodes. 244 */ 245 246 /* connect hdr to block node */ 247 RF_ASSERT(blockNode->numAntecedents == 0); 248 dag_h->succedents[0] = blockNode; 249 250 if (type == RF_IO_TYPE_READ) { 251 /* connecting a nonredundant read DAG */ 252 RF_ASSERT(blockNode->numSuccedents == n); 253 RF_ASSERT(commitNode->numAntecedents == n); 254 tmpdiskNode = diskNodes; 255 for (i = 0; i < n; i++) { 256 /* connect block node to each read node */ 257 RF_ASSERT(tmpdiskNode->numAntecedents == 1); 258 blockNode->succedents[i] = tmpdiskNode; 259 tmpdiskNode->antecedents[0] = blockNode; 260 tmpdiskNode->antType[0] = rf_control; 261 262 /* connect each read node to the commit node */ 263 RF_ASSERT(tmpdiskNode->numSuccedents == 1); 264 tmpdiskNode->succedents[0] = commitNode; 265 commitNode->antecedents[i] = tmpdiskNode; 266 commitNode->antType[i] = rf_control; 267 tmpdiskNode = tmpdiskNode->list_next; 268 } 269 /* connect the commit node to the term node */ 270 RF_ASSERT(commitNode->numSuccedents == 1); 271 RF_ASSERT(termNode->numAntecedents == 1); 272 RF_ASSERT(termNode->numSuccedents == 0); 273 commitNode->succedents[0] = termNode; 274 termNode->antecedents[0] = commitNode; 275 termNode->antType[0] = rf_control; 276 } else { 277 /* connecting a nonredundant write DAG */ 278 /* connect the block node to the commit node */ 279 RF_ASSERT(blockNode->numSuccedents == 1); 280 RF_ASSERT(commitNode->numAntecedents == 1); 281 blockNode->succedents[0] = commitNode; 282 commitNode->antecedents[0] = blockNode; 283 commitNode->antType[0] = rf_control; 284 285 RF_ASSERT(commitNode->numSuccedents == n); 286 RF_ASSERT(termNode->numAntecedents == n); 287 RF_ASSERT(termNode->numSuccedents == 0); 288 tmpdiskNode = diskNodes; 289 for (i = 0; i < n; i++) { 290 /* connect the commit node to each write node */ 291 RF_ASSERT(tmpdiskNode->numAntecedents == 1); 292 commitNode->succedents[i] = tmpdiskNode; 293 tmpdiskNode->antecedents[0] = commitNode; 294 tmpdiskNode->antType[0] = rf_control; 295 296 /* connect each write node to the term node */ 297 RF_ASSERT(tmpdiskNode->numSuccedents == 1); 298 tmpdiskNode->succedents[0] = termNode; 299 termNode->antecedents[i] = tmpdiskNode; 300 termNode->antType[i] = rf_control; 301 tmpdiskNode = tmpdiskNode->list_next; 302 } 303 } 304 } 305 /****************************************************************************** 306 * Create a fault-free read DAG for RAID level 1 307 * 308 * Hdr -> Nil -> Rmir -> Cmt -> Trm 309 * 310 * The "Rmir" node schedules a read from the disk in the mirror pair with the 311 * shortest disk queue. the proper queue is selected at Rmir execution. this 312 * deferred mapping is unlike other archs in RAIDframe which generally fix 313 * mapping at DAG creation time. 314 * 315 * Parameters: raidPtr - description of the physical array 316 * asmap - logical & physical addresses for this access 317 * bp - buffer ptr (for holding read data) 318 * flags - general flags (e.g. disk locking) 319 * allocList - list of memory allocated in DAG creation 320 *****************************************************************************/ 321 322 static void 323 CreateMirrorReadDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 324 RF_DagHeader_t *dag_h, void *bp, 325 RF_RaidAccessFlags_t flags, RF_AllocListElem_t *allocList, 326 void (*readfunc) (RF_DagNode_t * node)) 327 { 328 RF_DagNode_t *readNodes, *blockNode, *commitNode, *termNode; 329 RF_DagNode_t *tmpNode, *tmpreadNode; 330 RF_PhysDiskAddr_t *data_pda = asmap->physInfo; 331 RF_PhysDiskAddr_t *parity_pda = asmap->parityInfo; 332 int i, n; 333 334 n = asmap->numStripeUnitsAccessed; 335 dag_h->creator = "RaidOneReadDAG"; 336 #if RF_DEBUG_DAG 337 if (rf_dagDebug) { 338 printf("[Creating RAID level 1 read DAG]\n"); 339 } 340 #endif 341 /* 342 * This dag can not commit until the commit node is reached 343 * errors prior to the commit point imply the dag has failed. 344 */ 345 dag_h->numCommitNodes = 1; 346 dag_h->numCommits = 0; 347 dag_h->numSuccedents = 1; 348 349 /* 350 * Node count: 351 * n data reads 352 * 1 block node 353 * 1 commit node 354 * 1 terminator node 355 */ 356 RF_ASSERT(n > 0); 357 358 for (i = 0; i < n; i++) { 359 tmpNode = rf_AllocDAGNode(); 360 tmpNode->list_next = dag_h->nodes; 361 dag_h->nodes = tmpNode; 362 } 363 readNodes = dag_h->nodes; 364 365 blockNode = rf_AllocDAGNode(); 366 blockNode->list_next = dag_h->nodes; 367 dag_h->nodes = blockNode; 368 369 commitNode = rf_AllocDAGNode(); 370 commitNode->list_next = dag_h->nodes; 371 dag_h->nodes = commitNode; 372 373 termNode = rf_AllocDAGNode(); 374 termNode->list_next = dag_h->nodes; 375 dag_h->nodes = termNode; 376 377 /* initialize nodes */ 378 rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, 379 rf_NullNodeUndoFunc, NULL, n, 0, 0, 0, dag_h, "Nil", allocList); 380 rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, 381 rf_NullNodeUndoFunc, NULL, 1, n, 0, 0, dag_h, "Cmt", allocList); 382 rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, 383 rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList); 384 385 tmpreadNode = readNodes; 386 for (i = 0; i < n; i++) { 387 RF_ASSERT(data_pda != NULL); 388 RF_ASSERT(parity_pda != NULL); 389 rf_InitNode(tmpreadNode, rf_wait, RF_FALSE, readfunc, 390 rf_DiskReadMirrorUndoFunc, rf_GenericWakeupFunc, 1, 1, 5, 0, dag_h, 391 "Rmir", allocList); 392 tmpreadNode->params[0].p = data_pda; 393 tmpreadNode->params[1].p = data_pda->bufPtr; 394 /* parity stripe id is not necessary */ 395 tmpreadNode->params[2].p = 0; 396 tmpreadNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0); 397 tmpreadNode->params[4].p = parity_pda; 398 data_pda = data_pda->next; 399 parity_pda = parity_pda->next; 400 tmpreadNode = tmpreadNode->list_next; 401 } 402 403 /* 404 * Connect nodes 405 */ 406 407 /* connect hdr to block node */ 408 RF_ASSERT(blockNode->numAntecedents == 0); 409 dag_h->succedents[0] = blockNode; 410 411 /* connect block node to read nodes */ 412 RF_ASSERT(blockNode->numSuccedents == n); 413 tmpreadNode = readNodes; 414 for (i = 0; i < n; i++) { 415 RF_ASSERT(tmpreadNode->numAntecedents == 1); 416 blockNode->succedents[i] = tmpreadNode; 417 tmpreadNode->antecedents[0] = blockNode; 418 tmpreadNode->antType[0] = rf_control; 419 tmpreadNode = tmpreadNode->list_next; 420 } 421 422 /* connect read nodes to commit node */ 423 RF_ASSERT(commitNode->numAntecedents == n); 424 tmpreadNode = readNodes; 425 for (i = 0; i < n; i++) { 426 RF_ASSERT(tmpreadNode->numSuccedents == 1); 427 tmpreadNode->succedents[0] = commitNode; 428 commitNode->antecedents[i] = tmpreadNode; 429 commitNode->antType[i] = rf_control; 430 tmpreadNode = tmpreadNode->list_next; 431 } 432 433 /* connect commit node to term node */ 434 RF_ASSERT(commitNode->numSuccedents == 1); 435 RF_ASSERT(termNode->numAntecedents == 1); 436 RF_ASSERT(termNode->numSuccedents == 0); 437 commitNode->succedents[0] = termNode; 438 termNode->antecedents[0] = commitNode; 439 termNode->antType[0] = rf_control; 440 } 441 442 void 443 rf_CreateMirrorIdleReadDAG( 444 RF_Raid_t * raidPtr, 445 RF_AccessStripeMap_t * asmap, 446 RF_DagHeader_t * dag_h, 447 void *bp, 448 RF_RaidAccessFlags_t flags, 449 RF_AllocListElem_t * allocList) 450 { 451 CreateMirrorReadDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 452 rf_DiskReadMirrorIdleFunc); 453 } 454 455 #if (RF_INCLUDE_CHAINDECLUSTER > 0) || (RF_INCLUDE_INTERDECLUSTER > 0) 456 457 void 458 rf_CreateMirrorPartitionReadDAG(RF_Raid_t *raidPtr, 459 RF_AccessStripeMap_t *asmap, 460 RF_DagHeader_t *dag_h, void *bp, 461 RF_RaidAccessFlags_t flags, 462 RF_AllocListElem_t *allocList) 463 { 464 CreateMirrorReadDAG(raidPtr, asmap, dag_h, bp, flags, allocList, 465 rf_DiskReadMirrorPartitionFunc); 466 } 467 #endif 468