| 1 | /* $NetBSD: rf_engine.c,v 1.51 2016/01/03 08:17:24 mlelstv Exp $ */ |
| 2 | /* |
| 3 | * Copyright (c) 1995 Carnegie-Mellon University. |
| 4 | * All rights reserved. |
| 5 | * |
| 6 | * Author: William V. Courtright II, Mark Holland, Rachad Youssef |
| 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 | * * |
| 31 | * engine.c -- code for DAG execution engine * |
| 32 | * * |
| 33 | * Modified to work as follows (holland): * |
| 34 | * A user-thread calls into DispatchDAG, which fires off the nodes that * |
| 35 | * are direct successors to the header node. DispatchDAG then returns, * |
| 36 | * and the rest of the I/O continues asynchronously. As each node * |
| 37 | * completes, the node execution function calls FinishNode(). FinishNode * |
| 38 | * scans the list of successors to the node and increments the antecedent * |
| 39 | * counts. Each node that becomes enabled is placed on a central node * |
| 40 | * queue. A dedicated dag-execution thread grabs nodes off of this * |
| 41 | * queue and fires them. * |
| 42 | * * |
| 43 | * NULL nodes are never fired. * |
| 44 | * * |
| 45 | * Terminator nodes are never fired, but rather cause the callback * |
| 46 | * associated with the DAG to be invoked. * |
| 47 | * * |
| 48 | * If a node fails, the dag either rolls forward to the completion or * |
| 49 | * rolls back, undoing previously-completed nodes and fails atomically. * |
| 50 | * The direction of recovery is determined by the location of the failed * |
| 51 | * node in the graph. If the failure occurred before the commit node in * |
| 52 | * the graph, backward recovery is used. Otherwise, forward recovery is * |
| 53 | * used. * |
| 54 | * * |
| 55 | ****************************************************************************/ |
| 56 | |
| 57 | #include <sys/cdefs.h> |
| 58 | __KERNEL_RCSID(0, "$NetBSD: rf_engine.c,v 1.51 2016/01/03 08:17:24 mlelstv Exp $" ); |
| 59 | |
| 60 | #include <sys/errno.h> |
| 61 | |
| 62 | #include "rf_threadstuff.h" |
| 63 | #include "rf_dag.h" |
| 64 | #include "rf_engine.h" |
| 65 | #include "rf_etimer.h" |
| 66 | #include "rf_general.h" |
| 67 | #include "rf_dagutils.h" |
| 68 | #include "rf_shutdown.h" |
| 69 | #include "rf_raid.h" |
| 70 | #include "rf_kintf.h" |
| 71 | #include "rf_paritymap.h" |
| 72 | |
| 73 | static void rf_ShutdownEngine(void *); |
| 74 | static void DAGExecutionThread(RF_ThreadArg_t arg); |
| 75 | static void rf_RaidIOThread(RF_ThreadArg_t arg); |
| 76 | |
| 77 | /* synchronization primitives for this file. DO_WAIT should be enclosed in a while loop. */ |
| 78 | |
| 79 | #define DO_LOCK(_r_) \ |
| 80 | rf_lock_mutex2((_r_)->node_queue_mutex) |
| 81 | |
| 82 | #define DO_UNLOCK(_r_) \ |
| 83 | rf_unlock_mutex2((_r_)->node_queue_mutex) |
| 84 | |
| 85 | #define DO_WAIT(_r_) \ |
| 86 | rf_wait_cond2((_r_)->node_queue_cv, (_r_)->node_queue_mutex) |
| 87 | |
| 88 | #define DO_SIGNAL(_r_) \ |
| 89 | rf_broadcast_cond2((_r_)->node_queue_cv) /* XXX rf_signal_cond2? */ |
| 90 | |
| 91 | static void |
| 92 | rf_ShutdownEngine(void *arg) |
| 93 | { |
| 94 | RF_Raid_t *raidPtr; |
| 95 | |
| 96 | raidPtr = (RF_Raid_t *) arg; |
| 97 | |
| 98 | /* Tell the rf_RaidIOThread to shutdown */ |
| 99 | rf_lock_mutex2(raidPtr->iodone_lock); |
| 100 | |
| 101 | raidPtr->shutdown_raidio = 1; |
| 102 | rf_signal_cond2(raidPtr->iodone_cv); |
| 103 | |
| 104 | /* ...and wait for it to tell us it has finished */ |
| 105 | while (raidPtr->shutdown_raidio) |
| 106 | rf_wait_cond2(raidPtr->iodone_cv, raidPtr->iodone_lock); |
| 107 | |
| 108 | rf_unlock_mutex2(raidPtr->iodone_lock); |
| 109 | |
| 110 | /* Now shut down the DAG execution engine. */ |
| 111 | DO_LOCK(raidPtr); |
| 112 | raidPtr->shutdown_engine = 1; |
| 113 | DO_SIGNAL(raidPtr); |
| 114 | |
| 115 | /* ...and wait for it to tell us it has finished */ |
| 116 | while (raidPtr->shutdown_engine) |
| 117 | DO_WAIT(raidPtr); |
| 118 | |
| 119 | DO_UNLOCK(raidPtr); |
| 120 | |
| 121 | rf_destroy_mutex2(raidPtr->node_queue_mutex); |
| 122 | rf_destroy_cond2(raidPtr->node_queue_cv); |
| 123 | |
| 124 | rf_destroy_mutex2(raidPtr->iodone_lock); |
| 125 | rf_destroy_cond2(raidPtr->iodone_cv); |
| 126 | } |
| 127 | |
| 128 | int |
| 129 | rf_ConfigureEngine(RF_ShutdownList_t **listp, RF_Raid_t *raidPtr, |
| 130 | RF_Config_t *cfgPtr) |
| 131 | { |
| 132 | |
| 133 | /* |
| 134 | * Initialise iodone for the IO thread. |
| 135 | */ |
| 136 | TAILQ_INIT(&(raidPtr->iodone)); |
| 137 | rf_init_mutex2(raidPtr->iodone_lock, IPL_VM); |
| 138 | rf_init_cond2(raidPtr->iodone_cv, "raidiow" ); |
| 139 | |
| 140 | rf_init_mutex2(raidPtr->node_queue_mutex, IPL_VM); |
| 141 | rf_init_cond2(raidPtr->node_queue_cv, "rfnodeq" ); |
| 142 | raidPtr->node_queue = NULL; |
| 143 | raidPtr->dags_in_flight = 0; |
| 144 | |
| 145 | /* we create the execution thread only once per system boot. no need |
| 146 | * to check return code b/c the kernel panics if it can't create the |
| 147 | * thread. */ |
| 148 | #if RF_DEBUG_ENGINE |
| 149 | if (rf_engineDebug) { |
| 150 | printf("raid%d: Creating engine thread\n" , raidPtr->raidid); |
| 151 | } |
| 152 | #endif |
| 153 | if (RF_CREATE_ENGINE_THREAD(raidPtr->engine_thread, |
| 154 | DAGExecutionThread, raidPtr, |
| 155 | "raid%d" , raidPtr->raidid)) { |
| 156 | printf("raid%d: Unable to create engine thread\n" , |
| 157 | raidPtr->raidid); |
| 158 | return (ENOMEM); |
| 159 | } |
| 160 | if (RF_CREATE_ENGINE_THREAD(raidPtr->engine_helper_thread, |
| 161 | rf_RaidIOThread, raidPtr, |
| 162 | "raidio%d" , raidPtr->raidid)) { |
| 163 | printf("raid%d: Unable to create raidio thread\n" , |
| 164 | raidPtr->raidid); |
| 165 | return (ENOMEM); |
| 166 | } |
| 167 | #if RF_DEBUG_ENGINE |
| 168 | if (rf_engineDebug) { |
| 169 | printf("raid%d: Created engine thread\n" , raidPtr->raidid); |
| 170 | } |
| 171 | #endif |
| 172 | |
| 173 | /* engine thread is now running and waiting for work */ |
| 174 | #if RF_DEBUG_ENGINE |
| 175 | if (rf_engineDebug) { |
| 176 | printf("raid%d: Engine thread running and waiting for events\n" , raidPtr->raidid); |
| 177 | } |
| 178 | #endif |
| 179 | rf_ShutdownCreate(listp, rf_ShutdownEngine, raidPtr); |
| 180 | |
| 181 | return (0); |
| 182 | } |
| 183 | |
| 184 | #if 0 |
| 185 | static int |
| 186 | BranchDone(RF_DagNode_t *node) |
| 187 | { |
| 188 | int i; |
| 189 | |
| 190 | /* return true if forward execution is completed for a node and its |
| 191 | * succedents */ |
| 192 | switch (node->status) { |
| 193 | case rf_wait: |
| 194 | /* should never be called in this state */ |
| 195 | RF_PANIC(); |
| 196 | break; |
| 197 | case rf_fired: |
| 198 | /* node is currently executing, so we're not done */ |
| 199 | return (RF_FALSE); |
| 200 | case rf_good: |
| 201 | /* for each succedent recursively check branch */ |
| 202 | for (i = 0; i < node->numSuccedents; i++) |
| 203 | if (!BranchDone(node->succedents[i])) |
| 204 | return RF_FALSE; |
| 205 | return RF_TRUE; /* node and all succedent branches aren't in |
| 206 | * fired state */ |
| 207 | case rf_bad: |
| 208 | /* succedents can't fire */ |
| 209 | return (RF_TRUE); |
| 210 | case rf_recover: |
| 211 | /* should never be called in this state */ |
| 212 | RF_PANIC(); |
| 213 | break; |
| 214 | case rf_undone: |
| 215 | case rf_panic: |
| 216 | /* XXX need to fix this case */ |
| 217 | /* for now, assume that we're done */ |
| 218 | return (RF_TRUE); |
| 219 | default: |
| 220 | /* illegal node status */ |
| 221 | RF_PANIC(); |
| 222 | break; |
| 223 | } |
| 224 | } |
| 225 | #endif |
| 226 | |
| 227 | static int |
| 228 | NodeReady(RF_DagNode_t *node) |
| 229 | { |
| 230 | int ready; |
| 231 | |
| 232 | switch (node->dagHdr->status) { |
| 233 | case rf_enable: |
| 234 | case rf_rollForward: |
| 235 | if ((node->status == rf_wait) && |
| 236 | (node->numAntecedents == node->numAntDone)) |
| 237 | ready = RF_TRUE; |
| 238 | else |
| 239 | ready = RF_FALSE; |
| 240 | break; |
| 241 | case rf_rollBackward: |
| 242 | RF_ASSERT(node->numSuccDone <= node->numSuccedents); |
| 243 | RF_ASSERT(node->numSuccFired <= node->numSuccedents); |
| 244 | RF_ASSERT(node->numSuccFired <= node->numSuccDone); |
| 245 | if ((node->status == rf_good) && |
| 246 | (node->numSuccDone == node->numSuccedents)) |
| 247 | ready = RF_TRUE; |
| 248 | else |
| 249 | ready = RF_FALSE; |
| 250 | break; |
| 251 | default: |
| 252 | printf("Execution engine found illegal DAG status in NodeReady\n" ); |
| 253 | RF_PANIC(); |
| 254 | break; |
| 255 | } |
| 256 | |
| 257 | return (ready); |
| 258 | } |
| 259 | |
| 260 | |
| 261 | |
| 262 | /* user context and dag-exec-thread context: Fire a node. The node's |
| 263 | * status field determines which function, do or undo, to be fired. |
| 264 | * This routine assumes that the node's status field has alread been |
| 265 | * set to "fired" or "recover" to indicate the direction of execution. |
| 266 | */ |
| 267 | static void |
| 268 | FireNode(RF_DagNode_t *node) |
| 269 | { |
| 270 | switch (node->status) { |
| 271 | case rf_fired: |
| 272 | /* fire the do function of a node */ |
| 273 | #if RF_DEBUG_ENGINE |
| 274 | if (rf_engineDebug) { |
| 275 | printf("raid%d: Firing node 0x%lx (%s)\n" , |
| 276 | node->dagHdr->raidPtr->raidid, |
| 277 | (unsigned long) node, node->name); |
| 278 | } |
| 279 | #endif |
| 280 | if (node->flags & RF_DAGNODE_FLAG_YIELD) { |
| 281 | #if defined(__NetBSD__) && defined(_KERNEL) |
| 282 | /* thread_block(); */ |
| 283 | /* printf("Need to block the thread here...\n"); */ |
| 284 | /* XXX thread_block is actually mentioned in |
| 285 | * /usr/include/vm/vm_extern.h */ |
| 286 | #else |
| 287 | thread_block(); |
| 288 | #endif |
| 289 | } |
| 290 | (*(node->doFunc)) (node); |
| 291 | break; |
| 292 | case rf_recover: |
| 293 | /* fire the undo function of a node */ |
| 294 | #if RF_DEBUG_ENGINE |
| 295 | if (rf_engineDebug) { |
| 296 | printf("raid%d: Firing (undo) node 0x%lx (%s)\n" , |
| 297 | node->dagHdr->raidPtr->raidid, |
| 298 | (unsigned long) node, node->name); |
| 299 | } |
| 300 | #endif |
| 301 | if (node->flags & RF_DAGNODE_FLAG_YIELD) |
| 302 | #if defined(__NetBSD__) && defined(_KERNEL) |
| 303 | /* thread_block(); */ |
| 304 | /* printf("Need to block the thread here...\n"); */ |
| 305 | /* XXX thread_block is actually mentioned in |
| 306 | * /usr/include/vm/vm_extern.h */ |
| 307 | #else |
| 308 | thread_block(); |
| 309 | #endif |
| 310 | (*(node->undoFunc)) (node); |
| 311 | break; |
| 312 | default: |
| 313 | RF_PANIC(); |
| 314 | break; |
| 315 | } |
| 316 | } |
| 317 | |
| 318 | |
| 319 | |
| 320 | /* user context: |
| 321 | * Attempt to fire each node in a linear array. |
| 322 | * The entire list is fired atomically. |
| 323 | */ |
| 324 | static void |
| 325 | FireNodeArray(int numNodes, RF_DagNode_t **nodeList) |
| 326 | { |
| 327 | RF_DagStatus_t dstat; |
| 328 | RF_DagNode_t *node; |
| 329 | int i, j; |
| 330 | |
| 331 | /* first, mark all nodes which are ready to be fired */ |
| 332 | for (i = 0; i < numNodes; i++) { |
| 333 | node = nodeList[i]; |
| 334 | dstat = node->dagHdr->status; |
| 335 | RF_ASSERT((node->status == rf_wait) || |
| 336 | (node->status == rf_good)); |
| 337 | if (NodeReady(node)) { |
| 338 | if ((dstat == rf_enable) || |
| 339 | (dstat == rf_rollForward)) { |
| 340 | RF_ASSERT(node->status == rf_wait); |
| 341 | if (node->commitNode) |
| 342 | node->dagHdr->numCommits++; |
| 343 | node->status = rf_fired; |
| 344 | for (j = 0; j < node->numAntecedents; j++) |
| 345 | node->antecedents[j]->numSuccFired++; |
| 346 | } else { |
| 347 | RF_ASSERT(dstat == rf_rollBackward); |
| 348 | RF_ASSERT(node->status == rf_good); |
| 349 | /* only one commit node per graph */ |
| 350 | RF_ASSERT(node->commitNode == RF_FALSE); |
| 351 | node->status = rf_recover; |
| 352 | } |
| 353 | } |
| 354 | } |
| 355 | /* now, fire the nodes */ |
| 356 | for (i = 0; i < numNodes; i++) { |
| 357 | if ((nodeList[i]->status == rf_fired) || |
| 358 | (nodeList[i]->status == rf_recover)) |
| 359 | FireNode(nodeList[i]); |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | |
| 364 | /* user context: |
| 365 | * Attempt to fire each node in a linked list. |
| 366 | * The entire list is fired atomically. |
| 367 | */ |
| 368 | static void |
| 369 | FireNodeList(RF_DagNode_t *nodeList) |
| 370 | { |
| 371 | RF_DagNode_t *node, *next; |
| 372 | RF_DagStatus_t dstat; |
| 373 | int j; |
| 374 | |
| 375 | if (nodeList) { |
| 376 | /* first, mark all nodes which are ready to be fired */ |
| 377 | for (node = nodeList; node; node = next) { |
| 378 | next = node->next; |
| 379 | dstat = node->dagHdr->status; |
| 380 | RF_ASSERT((node->status == rf_wait) || |
| 381 | (node->status == rf_good)); |
| 382 | if (NodeReady(node)) { |
| 383 | if ((dstat == rf_enable) || |
| 384 | (dstat == rf_rollForward)) { |
| 385 | RF_ASSERT(node->status == rf_wait); |
| 386 | if (node->commitNode) |
| 387 | node->dagHdr->numCommits++; |
| 388 | node->status = rf_fired; |
| 389 | for (j = 0; j < node->numAntecedents; j++) |
| 390 | node->antecedents[j]->numSuccFired++; |
| 391 | } else { |
| 392 | RF_ASSERT(dstat == rf_rollBackward); |
| 393 | RF_ASSERT(node->status == rf_good); |
| 394 | /* only one commit node per graph */ |
| 395 | RF_ASSERT(node->commitNode == RF_FALSE); |
| 396 | node->status = rf_recover; |
| 397 | } |
| 398 | } |
| 399 | } |
| 400 | /* now, fire the nodes */ |
| 401 | for (node = nodeList; node; node = next) { |
| 402 | next = node->next; |
| 403 | if ((node->status == rf_fired) || |
| 404 | (node->status == rf_recover)) |
| 405 | FireNode(node); |
| 406 | } |
| 407 | } |
| 408 | } |
| 409 | /* interrupt context: |
| 410 | * for each succedent |
| 411 | * propagate required results from node to succedent |
| 412 | * increment succedent's numAntDone |
| 413 | * place newly-enable nodes on node queue for firing |
| 414 | * |
| 415 | * To save context switches, we don't place NIL nodes on the node queue, |
| 416 | * but rather just process them as if they had fired. Note that NIL nodes |
| 417 | * that are the direct successors of the header will actually get fired by |
| 418 | * DispatchDAG, which is fine because no context switches are involved. |
| 419 | * |
| 420 | * Important: when running at user level, this can be called by any |
| 421 | * disk thread, and so the increment and check of the antecedent count |
| 422 | * must be locked. I used the node queue mutex and locked down the |
| 423 | * entire function, but this is certainly overkill. |
| 424 | */ |
| 425 | static void |
| 426 | PropagateResults(RF_DagNode_t *node, int context) |
| 427 | { |
| 428 | RF_DagNode_t *s, *a; |
| 429 | RF_Raid_t *raidPtr; |
| 430 | int i; |
| 431 | RF_DagNode_t *finishlist = NULL; /* a list of NIL nodes to be |
| 432 | * finished */ |
| 433 | RF_DagNode_t *skiplist = NULL; /* list of nodes with failed truedata |
| 434 | * antecedents */ |
| 435 | RF_DagNode_t *firelist = NULL; /* a list of nodes to be fired */ |
| 436 | RF_DagNode_t *q = NULL, *qh = NULL, *next; |
| 437 | int j, skipNode; |
| 438 | |
| 439 | raidPtr = node->dagHdr->raidPtr; |
| 440 | |
| 441 | DO_LOCK(raidPtr); |
| 442 | |
| 443 | /* debug - validate fire counts */ |
| 444 | for (i = 0; i < node->numAntecedents; i++) { |
| 445 | a = *(node->antecedents + i); |
| 446 | RF_ASSERT(a->numSuccFired >= a->numSuccDone); |
| 447 | RF_ASSERT(a->numSuccFired <= a->numSuccedents); |
| 448 | a->numSuccDone++; |
| 449 | } |
| 450 | |
| 451 | switch (node->dagHdr->status) { |
| 452 | case rf_enable: |
| 453 | case rf_rollForward: |
| 454 | for (i = 0; i < node->numSuccedents; i++) { |
| 455 | s = *(node->succedents + i); |
| 456 | RF_ASSERT(s->status == rf_wait); |
| 457 | (s->numAntDone)++; |
| 458 | if (s->numAntDone == s->numAntecedents) { |
| 459 | /* look for NIL nodes */ |
| 460 | if (s->doFunc == rf_NullNodeFunc) { |
| 461 | /* don't fire NIL nodes, just process |
| 462 | * them */ |
| 463 | s->next = finishlist; |
| 464 | finishlist = s; |
| 465 | } else { |
| 466 | /* look to see if the node is to be |
| 467 | * skipped */ |
| 468 | skipNode = RF_FALSE; |
| 469 | for (j = 0; j < s->numAntecedents; j++) |
| 470 | if ((s->antType[j] == rf_trueData) && (s->antecedents[j]->status == rf_bad)) |
| 471 | skipNode = RF_TRUE; |
| 472 | if (skipNode) { |
| 473 | /* this node has one or more |
| 474 | * failed true data |
| 475 | * dependencies, so skip it */ |
| 476 | s->next = skiplist; |
| 477 | skiplist = s; |
| 478 | } else |
| 479 | /* add s to list of nodes (q) |
| 480 | * to execute */ |
| 481 | if (context != RF_INTR_CONTEXT) { |
| 482 | /* we only have to |
| 483 | * enqueue if we're at |
| 484 | * intr context */ |
| 485 | /* put node on |
| 486 | a list to |
| 487 | be fired |
| 488 | after we |
| 489 | unlock */ |
| 490 | s->next = firelist; |
| 491 | firelist = s; |
| 492 | } else { |
| 493 | /* enqueue the |
| 494 | node for |
| 495 | the dag |
| 496 | exec thread |
| 497 | to fire */ |
| 498 | RF_ASSERT(NodeReady(s)); |
| 499 | if (q) { |
| 500 | q->next = s; |
| 501 | q = s; |
| 502 | } else { |
| 503 | qh = q = s; |
| 504 | qh->next = NULL; |
| 505 | } |
| 506 | } |
| 507 | } |
| 508 | } |
| 509 | } |
| 510 | |
| 511 | if (q) { |
| 512 | /* xfer our local list of nodes to the node queue */ |
| 513 | q->next = raidPtr->node_queue; |
| 514 | raidPtr->node_queue = qh; |
| 515 | DO_SIGNAL(raidPtr); |
| 516 | } |
| 517 | DO_UNLOCK(raidPtr); |
| 518 | |
| 519 | for (; skiplist; skiplist = next) { |
| 520 | next = skiplist->next; |
| 521 | skiplist->status = rf_skipped; |
| 522 | for (i = 0; i < skiplist->numAntecedents; i++) { |
| 523 | skiplist->antecedents[i]->numSuccFired++; |
| 524 | } |
| 525 | if (skiplist->commitNode) { |
| 526 | skiplist->dagHdr->numCommits++; |
| 527 | } |
| 528 | rf_FinishNode(skiplist, context); |
| 529 | } |
| 530 | for (; finishlist; finishlist = next) { |
| 531 | /* NIL nodes: no need to fire them */ |
| 532 | next = finishlist->next; |
| 533 | finishlist->status = rf_good; |
| 534 | for (i = 0; i < finishlist->numAntecedents; i++) { |
| 535 | finishlist->antecedents[i]->numSuccFired++; |
| 536 | } |
| 537 | if (finishlist->commitNode) |
| 538 | finishlist->dagHdr->numCommits++; |
| 539 | /* |
| 540 | * Okay, here we're calling rf_FinishNode() on |
| 541 | * nodes that have the null function as their |
| 542 | * work proc. Such a node could be the |
| 543 | * terminal node in a DAG. If so, it will |
| 544 | * cause the DAG to complete, which will in |
| 545 | * turn free memory used by the DAG, which |
| 546 | * includes the node in question. Thus, we |
| 547 | * must avoid referencing the node at all |
| 548 | * after calling rf_FinishNode() on it. */ |
| 549 | rf_FinishNode(finishlist, context); /* recursive call */ |
| 550 | } |
| 551 | /* fire all nodes in firelist */ |
| 552 | FireNodeList(firelist); |
| 553 | break; |
| 554 | |
| 555 | case rf_rollBackward: |
| 556 | for (i = 0; i < node->numAntecedents; i++) { |
| 557 | a = *(node->antecedents + i); |
| 558 | RF_ASSERT(a->status == rf_good); |
| 559 | RF_ASSERT(a->numSuccDone <= a->numSuccedents); |
| 560 | RF_ASSERT(a->numSuccDone <= a->numSuccFired); |
| 561 | |
| 562 | if (a->numSuccDone == a->numSuccFired) { |
| 563 | if (a->undoFunc == rf_NullNodeFunc) { |
| 564 | /* don't fire NIL nodes, just process |
| 565 | * them */ |
| 566 | a->next = finishlist; |
| 567 | finishlist = a; |
| 568 | } else { |
| 569 | if (context != RF_INTR_CONTEXT) { |
| 570 | /* we only have to enqueue if |
| 571 | * we're at intr context */ |
| 572 | /* put node on a list to be |
| 573 | fired after we unlock */ |
| 574 | a->next = firelist; |
| 575 | |
| 576 | firelist = a; |
| 577 | } else { |
| 578 | /* enqueue the node for the |
| 579 | dag exec thread to fire */ |
| 580 | RF_ASSERT(NodeReady(a)); |
| 581 | if (q) { |
| 582 | q->next = a; |
| 583 | q = a; |
| 584 | } else { |
| 585 | qh = q = a; |
| 586 | qh->next = NULL; |
| 587 | } |
| 588 | } |
| 589 | } |
| 590 | } |
| 591 | } |
| 592 | if (q) { |
| 593 | /* xfer our local list of nodes to the node queue */ |
| 594 | q->next = raidPtr->node_queue; |
| 595 | raidPtr->node_queue = qh; |
| 596 | DO_SIGNAL(raidPtr); |
| 597 | } |
| 598 | DO_UNLOCK(raidPtr); |
| 599 | for (; finishlist; finishlist = next) { |
| 600 | /* NIL nodes: no need to fire them */ |
| 601 | next = finishlist->next; |
| 602 | finishlist->status = rf_good; |
| 603 | /* |
| 604 | * Okay, here we're calling rf_FinishNode() on |
| 605 | * nodes that have the null function as their |
| 606 | * work proc. Such a node could be the first |
| 607 | * node in a DAG. If so, it will cause the DAG |
| 608 | * to complete, which will in turn free memory |
| 609 | * used by the DAG, which includes the node in |
| 610 | * question. Thus, we must avoid referencing |
| 611 | * the node at all after calling |
| 612 | * rf_FinishNode() on it. */ |
| 613 | rf_FinishNode(finishlist, context); /* recursive call */ |
| 614 | } |
| 615 | /* fire all nodes in firelist */ |
| 616 | FireNodeList(firelist); |
| 617 | |
| 618 | break; |
| 619 | default: |
| 620 | printf("Engine found illegal DAG status in PropagateResults()\n" ); |
| 621 | RF_PANIC(); |
| 622 | break; |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | |
| 627 | |
| 628 | /* |
| 629 | * Process a fired node which has completed |
| 630 | */ |
| 631 | static void |
| 632 | ProcessNode(RF_DagNode_t *node, int context) |
| 633 | { |
| 634 | #if RF_DEBUG_ENGINE |
| 635 | RF_Raid_t *raidPtr; |
| 636 | |
| 637 | raidPtr = node->dagHdr->raidPtr; |
| 638 | #endif |
| 639 | |
| 640 | switch (node->status) { |
| 641 | case rf_good: |
| 642 | /* normal case, don't need to do anything */ |
| 643 | break; |
| 644 | case rf_bad: |
| 645 | if ((node->dagHdr->numCommits > 0) || |
| 646 | (node->dagHdr->numCommitNodes == 0)) { |
| 647 | /* crossed commit barrier */ |
| 648 | node->dagHdr->status = rf_rollForward; |
| 649 | #if RF_DEBUG_ENGINE |
| 650 | if (rf_engineDebug) { |
| 651 | printf("raid%d: node (%s) returned fail, rolling forward\n" , raidPtr->raidid, node->name); |
| 652 | } |
| 653 | #endif |
| 654 | } else { |
| 655 | /* never reached commit barrier */ |
| 656 | node->dagHdr->status = rf_rollBackward; |
| 657 | #if RF_DEBUG_ENGINE |
| 658 | if (rf_engineDebug) { |
| 659 | printf("raid%d: node (%s) returned fail, rolling backward\n" , raidPtr->raidid, node->name); |
| 660 | } |
| 661 | #endif |
| 662 | } |
| 663 | break; |
| 664 | case rf_undone: |
| 665 | /* normal rollBackward case, don't need to do anything */ |
| 666 | break; |
| 667 | case rf_panic: |
| 668 | /* an undo node failed!!! */ |
| 669 | printf("UNDO of a node failed!!!\n" ); |
| 670 | break; |
| 671 | default: |
| 672 | printf("node finished execution with an illegal status!!!\n" ); |
| 673 | RF_PANIC(); |
| 674 | break; |
| 675 | } |
| 676 | |
| 677 | /* enqueue node's succedents (antecedents if rollBackward) for |
| 678 | * execution */ |
| 679 | PropagateResults(node, context); |
| 680 | } |
| 681 | |
| 682 | |
| 683 | |
| 684 | /* user context or dag-exec-thread context: |
| 685 | * This is the first step in post-processing a newly-completed node. |
| 686 | * This routine is called by each node execution function to mark the node |
| 687 | * as complete and fire off any successors that have been enabled. |
| 688 | */ |
| 689 | int |
| 690 | rf_FinishNode(RF_DagNode_t *node, int context) |
| 691 | { |
| 692 | int retcode = RF_FALSE; |
| 693 | node->dagHdr->numNodesCompleted++; |
| 694 | ProcessNode(node, context); |
| 695 | |
| 696 | return (retcode); |
| 697 | } |
| 698 | |
| 699 | |
| 700 | /* user context: submit dag for execution, return non-zero if we have |
| 701 | * to wait for completion. if and only if we return non-zero, we'll |
| 702 | * cause cbFunc to get invoked with cbArg when the DAG has completed. |
| 703 | * |
| 704 | * for now we always return 1. If the DAG does not cause any I/O, |
| 705 | * then the callback may get invoked before DispatchDAG returns. |
| 706 | * There's code in state 5 of ContinueRaidAccess to handle this. |
| 707 | * |
| 708 | * All we do here is fire the direct successors of the header node. |
| 709 | * The DAG execution thread does the rest of the dag processing. */ |
| 710 | int |
| 711 | rf_DispatchDAG(RF_DagHeader_t *dag, void (*cbFunc) (void *), |
| 712 | void *cbArg) |
| 713 | { |
| 714 | RF_Raid_t *raidPtr; |
| 715 | |
| 716 | raidPtr = dag->raidPtr; |
| 717 | #if RF_ACC_TRACE > 0 |
| 718 | if (dag->tracerec) { |
| 719 | RF_ETIMER_START(dag->tracerec->timer); |
| 720 | } |
| 721 | #endif |
| 722 | #if DEBUG |
| 723 | #if RF_DEBUG_VALIDATE_DAG |
| 724 | if (rf_engineDebug || rf_validateDAGDebug) { |
| 725 | if (rf_ValidateDAG(dag)) |
| 726 | RF_PANIC(); |
| 727 | } |
| 728 | #endif |
| 729 | #endif |
| 730 | #if RF_DEBUG_ENGINE |
| 731 | if (rf_engineDebug) { |
| 732 | printf("raid%d: Entering DispatchDAG\n" , raidPtr->raidid); |
| 733 | } |
| 734 | #endif |
| 735 | raidPtr->dags_in_flight++; /* debug only: blow off proper |
| 736 | * locking */ |
| 737 | dag->cbFunc = cbFunc; |
| 738 | dag->cbArg = cbArg; |
| 739 | dag->numNodesCompleted = 0; |
| 740 | dag->status = rf_enable; |
| 741 | FireNodeArray(dag->numSuccedents, dag->succedents); |
| 742 | return (1); |
| 743 | } |
| 744 | /* dedicated kernel thread: the thread that handles all DAG node |
| 745 | * firing. To minimize locking and unlocking, we grab a copy of the |
| 746 | * entire node queue and then set the node queue to NULL before doing |
| 747 | * any firing of nodes. This way we only have to release the lock |
| 748 | * once. Of course, it's probably rare that there's more than one |
| 749 | * node in the queue at any one time, but it sometimes happens. |
| 750 | */ |
| 751 | |
| 752 | static void |
| 753 | DAGExecutionThread(RF_ThreadArg_t arg) |
| 754 | { |
| 755 | RF_DagNode_t *nd, *local_nq, *term_nq, *fire_nq; |
| 756 | RF_Raid_t *raidPtr; |
| 757 | |
| 758 | raidPtr = (RF_Raid_t *) arg; |
| 759 | |
| 760 | #if RF_DEBUG_ENGINE |
| 761 | if (rf_engineDebug) { |
| 762 | printf("raid%d: Engine thread is running\n" , raidPtr->raidid); |
| 763 | } |
| 764 | #endif |
| 765 | |
| 766 | DO_LOCK(raidPtr); |
| 767 | while (!raidPtr->shutdown_engine) { |
| 768 | |
| 769 | while (raidPtr->node_queue != NULL) { |
| 770 | local_nq = raidPtr->node_queue; |
| 771 | fire_nq = NULL; |
| 772 | term_nq = NULL; |
| 773 | raidPtr->node_queue = NULL; |
| 774 | DO_UNLOCK(raidPtr); |
| 775 | |
| 776 | /* first, strip out the terminal nodes */ |
| 777 | while (local_nq) { |
| 778 | nd = local_nq; |
| 779 | local_nq = local_nq->next; |
| 780 | switch (nd->dagHdr->status) { |
| 781 | case rf_enable: |
| 782 | case rf_rollForward: |
| 783 | if (nd->numSuccedents == 0) { |
| 784 | /* end of the dag, add to |
| 785 | * callback list */ |
| 786 | nd->next = term_nq; |
| 787 | term_nq = nd; |
| 788 | } else { |
| 789 | /* not the end, add to the |
| 790 | * fire queue */ |
| 791 | nd->next = fire_nq; |
| 792 | fire_nq = nd; |
| 793 | } |
| 794 | break; |
| 795 | case rf_rollBackward: |
| 796 | if (nd->numAntecedents == 0) { |
| 797 | /* end of the dag, add to the |
| 798 | * callback list */ |
| 799 | nd->next = term_nq; |
| 800 | term_nq = nd; |
| 801 | } else { |
| 802 | /* not the end, add to the |
| 803 | * fire queue */ |
| 804 | nd->next = fire_nq; |
| 805 | fire_nq = nd; |
| 806 | } |
| 807 | break; |
| 808 | default: |
| 809 | RF_PANIC(); |
| 810 | break; |
| 811 | } |
| 812 | } |
| 813 | |
| 814 | /* execute callback of dags which have reached the |
| 815 | * terminal node */ |
| 816 | while (term_nq) { |
| 817 | nd = term_nq; |
| 818 | term_nq = term_nq->next; |
| 819 | nd->next = NULL; |
| 820 | (nd->dagHdr->cbFunc) (nd->dagHdr->cbArg); |
| 821 | raidPtr->dags_in_flight--; /* debug only */ |
| 822 | } |
| 823 | |
| 824 | /* fire remaining nodes */ |
| 825 | FireNodeList(fire_nq); |
| 826 | |
| 827 | DO_LOCK(raidPtr); |
| 828 | } |
| 829 | while (!raidPtr->shutdown_engine && |
| 830 | raidPtr->node_queue == NULL) { |
| 831 | DO_WAIT(raidPtr); |
| 832 | } |
| 833 | } |
| 834 | |
| 835 | /* Let rf_ShutdownEngine know that we're done... */ |
| 836 | raidPtr->shutdown_engine = 0; |
| 837 | DO_SIGNAL(raidPtr); |
| 838 | |
| 839 | DO_UNLOCK(raidPtr); |
| 840 | |
| 841 | kthread_exit(0); |
| 842 | } |
| 843 | |
| 844 | /* |
| 845 | * rf_RaidIOThread() -- When I/O to a component begins, raidstrategy() |
| 846 | * puts the I/O on a buffer queue, and then signals raidPtr->iodone. If |
| 847 | * necessary, this function calls raidstart() to initiate the I/O. |
| 848 | * When I/O to a component completes, KernelWakeupFunc() puts the |
| 849 | * completed request onto raidPtr->iodone TAILQ. This function looks |
| 850 | * after requests on that queue by calling rf_DiskIOComplete() for the |
| 851 | * request, and by calling any required CompleteFunc for the request. |
| 852 | */ |
| 853 | |
| 854 | static void |
| 855 | rf_RaidIOThread(RF_ThreadArg_t arg) |
| 856 | { |
| 857 | RF_Raid_t *raidPtr; |
| 858 | RF_DiskQueueData_t *req; |
| 859 | |
| 860 | raidPtr = (RF_Raid_t *) arg; |
| 861 | |
| 862 | rf_lock_mutex2(raidPtr->iodone_lock); |
| 863 | |
| 864 | while (!raidPtr->shutdown_raidio) { |
| 865 | /* if there is nothing to do, then snooze. */ |
| 866 | if (TAILQ_EMPTY(&(raidPtr->iodone)) && |
| 867 | rf_buf_queue_check(raidPtr)) { |
| 868 | rf_wait_cond2(raidPtr->iodone_cv, raidPtr->iodone_lock); |
| 869 | } |
| 870 | |
| 871 | /* Check for deferred parity-map-related work. */ |
| 872 | if (raidPtr->parity_map != NULL) { |
| 873 | rf_unlock_mutex2(raidPtr->iodone_lock); |
| 874 | rf_paritymap_checkwork(raidPtr->parity_map); |
| 875 | rf_lock_mutex2(raidPtr->iodone_lock); |
| 876 | } |
| 877 | |
| 878 | /* See what I/Os, if any, have arrived */ |
| 879 | while ((req = TAILQ_FIRST(&(raidPtr->iodone))) != NULL) { |
| 880 | TAILQ_REMOVE(&(raidPtr->iodone), req, iodone_entries); |
| 881 | rf_unlock_mutex2(raidPtr->iodone_lock); |
| 882 | rf_DiskIOComplete(req->queue, req, req->error); |
| 883 | (req->CompleteFunc) (req->argument, req->error); |
| 884 | rf_lock_mutex2(raidPtr->iodone_lock); |
| 885 | } |
| 886 | |
| 887 | /* process any pending outgoing IO */ |
| 888 | rf_unlock_mutex2(raidPtr->iodone_lock); |
| 889 | raidstart(raidPtr); |
| 890 | rf_lock_mutex2(raidPtr->iodone_lock); |
| 891 | |
| 892 | } |
| 893 | |
| 894 | /* Let rf_ShutdownEngine know that we're done... */ |
| 895 | raidPtr->shutdown_raidio = 0; |
| 896 | rf_signal_cond2(raidPtr->iodone_cv); |
| 897 | |
| 898 | rf_unlock_mutex2(raidPtr->iodone_lock); |
| 899 | |
| 900 | kthread_exit(0); |
| 901 | } |
| 902 | |