1 /* $NetBSD: rf_dagutils.c,v 1.48 2006/01/09 01:33:27 oster Exp $ */ 2 /* 3 * Copyright (c) 1995 Carnegie-Mellon University. 4 * All rights reserved. 5 * 6 * Authors: Mark Holland, William V. Courtright II, Jim Zelenka 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 * rf_dagutils.c -- utility routines for manipulating dags 32 * 33 *****************************************************************************/ 34 35 #include <sys/cdefs.h> 36 __KERNEL_RCSID(0, "$NetBSD: rf_dagutils.c,v 1.48 2006/01/09 01:33:27 oster Exp $"); 37 38 #include <dev/raidframe/raidframevar.h> 39 40 #include "rf_archs.h" 41 #include "rf_threadstuff.h" 42 #include "rf_raid.h" 43 #include "rf_dag.h" 44 #include "rf_dagutils.h" 45 #include "rf_dagfuncs.h" 46 #include "rf_general.h" 47 #include "rf_map.h" 48 #include "rf_shutdown.h" 49 50 #define SNUM_DIFF(_a_,_b_) (((_a_)>(_b_))?((_a_)-(_b_)):((_b_)-(_a_))) 51 52 const RF_RedFuncs_t rf_xorFuncs = { 53 rf_RegularXorFunc, "Reg Xr", 54 rf_SimpleXorFunc, "Simple Xr"}; 55 56 const RF_RedFuncs_t rf_xorRecoveryFuncs = { 57 rf_RecoveryXorFunc, "Recovery Xr", 58 rf_RecoveryXorFunc, "Recovery Xr"}; 59 60 #if RF_DEBUG_VALIDATE_DAG 61 static void rf_RecurPrintDAG(RF_DagNode_t *, int, int); 62 static void rf_PrintDAG(RF_DagHeader_t *); 63 static int rf_ValidateBranch(RF_DagNode_t *, int *, int *, 64 RF_DagNode_t **, int); 65 static void rf_ValidateBranchVisitedBits(RF_DagNode_t *, int, int); 66 static void rf_ValidateVisitedBits(RF_DagHeader_t *); 67 #endif /* RF_DEBUG_VALIDATE_DAG */ 68 69 /* The maximum number of nodes in a DAG is bounded by 70 71 (2 * raidPtr->Layout->numDataCol) + (1 * layoutPtr->numParityCol) + 72 (1 * 2 * layoutPtr->numParityCol) + 3 73 74 which is: 2*RF_MAXCOL+1*2+1*2*2+3 75 76 For RF_MAXCOL of 40, this works out to 89. We use this value to provide an estimate 77 on the maximum size needed for RF_DAGPCACHE_SIZE. For RF_MAXCOL of 40, this structure 78 would be 534 bytes. Too much to have on-hand in a RF_DagNode_t, but should be ok to 79 have a few kicking around. 80 */ 81 #define RF_DAGPCACHE_SIZE ((2*RF_MAXCOL+1*2+1*2*2+3) *(RF_MAX(sizeof(RF_DagParam_t), sizeof(RF_DagNode_t *)))) 82 83 84 /****************************************************************************** 85 * 86 * InitNode - initialize a dag node 87 * 88 * the size of the propList array is always the same as that of the 89 * successors array. 90 * 91 *****************************************************************************/ 92 void 93 rf_InitNode(RF_DagNode_t *node, RF_NodeStatus_t initstatus, int commit, 94 int (*doFunc) (RF_DagNode_t *node), 95 int (*undoFunc) (RF_DagNode_t *node), 96 int (*wakeFunc) (RF_DagNode_t *node, int status), 97 int nSucc, int nAnte, int nParam, int nResult, 98 RF_DagHeader_t *hdr, const char *name, RF_AllocListElem_t *alist) 99 { 100 void **ptrs; 101 int nptrs; 102 103 if (nAnte > RF_MAX_ANTECEDENTS) 104 RF_PANIC(); 105 node->status = initstatus; 106 node->commitNode = commit; 107 node->doFunc = doFunc; 108 node->undoFunc = undoFunc; 109 node->wakeFunc = wakeFunc; 110 node->numParams = nParam; 111 node->numResults = nResult; 112 node->numAntecedents = nAnte; 113 node->numAntDone = 0; 114 node->next = NULL; 115 /* node->list_next = NULL */ /* Don't touch this here! 116 It may already be 117 in use by the caller! */ 118 node->numSuccedents = nSucc; 119 node->name = name; 120 node->dagHdr = hdr; 121 node->big_dag_ptrs = NULL; 122 node->big_dag_params = NULL; 123 node->visited = 0; 124 125 /* allocate all the pointers with one call to malloc */ 126 nptrs = nSucc + nAnte + nResult + nSucc; 127 128 if (nptrs <= RF_DAG_PTRCACHESIZE) { 129 /* 130 * The dag_ptrs field of the node is basically some scribble 131 * space to be used here. We could get rid of it, and always 132 * allocate the range of pointers, but that's expensive. So, 133 * we pick a "common case" size for the pointer cache. Hopefully, 134 * we'll find that: 135 * (1) Generally, nptrs doesn't exceed RF_DAG_PTRCACHESIZE by 136 * only a little bit (least efficient case) 137 * (2) Generally, ntprs isn't a lot less than RF_DAG_PTRCACHESIZE 138 * (wasted memory) 139 */ 140 ptrs = (void **) node->dag_ptrs; 141 } else if (nptrs <= (RF_DAGPCACHE_SIZE / sizeof(RF_DagNode_t *))) { 142 node->big_dag_ptrs = rf_AllocDAGPCache(); 143 ptrs = (void **) node->big_dag_ptrs; 144 } else { 145 RF_MallocAndAdd(ptrs, nptrs * sizeof(void *), 146 (void **), alist); 147 } 148 node->succedents = (nSucc) ? (RF_DagNode_t **) ptrs : NULL; 149 node->antecedents = (nAnte) ? (RF_DagNode_t **) (ptrs + nSucc) : NULL; 150 node->results = (nResult) ? (void **) (ptrs + nSucc + nAnte) : NULL; 151 node->propList = (nSucc) ? (RF_PropHeader_t **) (ptrs + nSucc + nAnte + nResult) : NULL; 152 153 if (nParam) { 154 if (nParam <= RF_DAG_PARAMCACHESIZE) { 155 node->params = (RF_DagParam_t *) node->dag_params; 156 } else if (nParam <= (RF_DAGPCACHE_SIZE / sizeof(RF_DagParam_t))) { 157 node->big_dag_params = rf_AllocDAGPCache(); 158 node->params = node->big_dag_params; 159 } else { 160 RF_MallocAndAdd(node->params, 161 nParam * sizeof(RF_DagParam_t), 162 (RF_DagParam_t *), alist); 163 } 164 } else { 165 node->params = NULL; 166 } 167 } 168 169 170 171 /****************************************************************************** 172 * 173 * allocation and deallocation routines 174 * 175 *****************************************************************************/ 176 177 void 178 rf_FreeDAG(RF_DagHeader_t *dag_h) 179 { 180 RF_AccessStripeMapHeader_t *asmap, *t_asmap; 181 RF_PhysDiskAddr_t *pda; 182 RF_DagNode_t *tmpnode; 183 RF_DagHeader_t *nextDag; 184 185 while (dag_h) { 186 nextDag = dag_h->next; 187 rf_FreeAllocList(dag_h->allocList); 188 for (asmap = dag_h->asmList; asmap;) { 189 t_asmap = asmap; 190 asmap = asmap->next; 191 rf_FreeAccessStripeMap(t_asmap); 192 } 193 while (dag_h->pda_cleanup_list) { 194 pda = dag_h->pda_cleanup_list; 195 dag_h->pda_cleanup_list = dag_h->pda_cleanup_list->next; 196 rf_FreePhysDiskAddr(pda); 197 } 198 while (dag_h->nodes) { 199 tmpnode = dag_h->nodes; 200 dag_h->nodes = dag_h->nodes->list_next; 201 rf_FreeDAGNode(tmpnode); 202 } 203 rf_FreeDAGHeader(dag_h); 204 dag_h = nextDag; 205 } 206 } 207 208 #define RF_MAX_FREE_DAGH 128 209 #define RF_MIN_FREE_DAGH 32 210 211 #define RF_MAX_FREE_DAGNODE 512 /* XXX Tune this... */ 212 #define RF_MIN_FREE_DAGNODE 128 /* XXX Tune this... */ 213 214 #define RF_MAX_FREE_DAGLIST 128 215 #define RF_MIN_FREE_DAGLIST 32 216 217 #define RF_MAX_FREE_DAGPCACHE 128 218 #define RF_MIN_FREE_DAGPCACHE 8 219 220 #define RF_MAX_FREE_FUNCLIST 128 221 #define RF_MIN_FREE_FUNCLIST 32 222 223 #define RF_MAX_FREE_BUFFERS 128 224 #define RF_MIN_FREE_BUFFERS 32 225 226 static void rf_ShutdownDAGs(void *); 227 static void 228 rf_ShutdownDAGs(void *ignored) 229 { 230 pool_destroy(&rf_pools.dagh); 231 pool_destroy(&rf_pools.dagnode); 232 pool_destroy(&rf_pools.daglist); 233 pool_destroy(&rf_pools.dagpcache); 234 pool_destroy(&rf_pools.funclist); 235 } 236 237 int 238 rf_ConfigureDAGs(RF_ShutdownList_t **listp) 239 { 240 241 rf_pool_init(&rf_pools.dagnode, sizeof(RF_DagNode_t), 242 "rf_dagnode_pl", RF_MIN_FREE_DAGNODE, RF_MAX_FREE_DAGNODE); 243 rf_pool_init(&rf_pools.dagh, sizeof(RF_DagHeader_t), 244 "rf_dagh_pl", RF_MIN_FREE_DAGH, RF_MAX_FREE_DAGH); 245 rf_pool_init(&rf_pools.daglist, sizeof(RF_DagList_t), 246 "rf_daglist_pl", RF_MIN_FREE_DAGLIST, RF_MAX_FREE_DAGLIST); 247 rf_pool_init(&rf_pools.dagpcache, RF_DAGPCACHE_SIZE, 248 "rf_dagpcache_pl", RF_MIN_FREE_DAGPCACHE, RF_MAX_FREE_DAGPCACHE); 249 rf_pool_init(&rf_pools.funclist, sizeof(RF_FuncList_t), 250 "rf_funclist_pl", RF_MIN_FREE_FUNCLIST, RF_MAX_FREE_FUNCLIST); 251 rf_ShutdownCreate(listp, rf_ShutdownDAGs, NULL); 252 253 return (0); 254 } 255 256 RF_DagHeader_t * 257 rf_AllocDAGHeader() 258 { 259 RF_DagHeader_t *dh; 260 261 dh = pool_get(&rf_pools.dagh, PR_WAITOK); 262 memset((char *) dh, 0, sizeof(RF_DagHeader_t)); 263 return (dh); 264 } 265 266 void 267 rf_FreeDAGHeader(RF_DagHeader_t * dh) 268 { 269 pool_put(&rf_pools.dagh, dh); 270 } 271 272 RF_DagNode_t * 273 rf_AllocDAGNode() 274 { 275 RF_DagNode_t *node; 276 277 node = pool_get(&rf_pools.dagnode, PR_WAITOK); 278 memset(node, 0, sizeof(RF_DagNode_t)); 279 return (node); 280 } 281 282 void 283 rf_FreeDAGNode(RF_DagNode_t *node) 284 { 285 if (node->big_dag_ptrs) { 286 rf_FreeDAGPCache(node->big_dag_ptrs); 287 } 288 if (node->big_dag_params) { 289 rf_FreeDAGPCache(node->big_dag_params); 290 } 291 pool_put(&rf_pools.dagnode, node); 292 } 293 294 RF_DagList_t * 295 rf_AllocDAGList() 296 { 297 RF_DagList_t *dagList; 298 299 dagList = pool_get(&rf_pools.daglist, PR_WAITOK); 300 memset(dagList, 0, sizeof(RF_DagList_t)); 301 302 return (dagList); 303 } 304 305 void 306 rf_FreeDAGList(RF_DagList_t *dagList) 307 { 308 pool_put(&rf_pools.daglist, dagList); 309 } 310 311 void * 312 rf_AllocDAGPCache() 313 { 314 void *p; 315 p = pool_get(&rf_pools.dagpcache, PR_WAITOK); 316 memset(p, 0, RF_DAGPCACHE_SIZE); 317 318 return (p); 319 } 320 321 void 322 rf_FreeDAGPCache(void *p) 323 { 324 pool_put(&rf_pools.dagpcache, p); 325 } 326 327 RF_FuncList_t * 328 rf_AllocFuncList() 329 { 330 RF_FuncList_t *funcList; 331 332 funcList = pool_get(&rf_pools.funclist, PR_WAITOK); 333 memset(funcList, 0, sizeof(RF_FuncList_t)); 334 335 return (funcList); 336 } 337 338 void 339 rf_FreeFuncList(RF_FuncList_t *funcList) 340 { 341 pool_put(&rf_pools.funclist, funcList); 342 } 343 344 /* allocates a stripe buffer -- a buffer large enough to hold all the data 345 in an entire stripe. 346 */ 347 348 void * 349 rf_AllocStripeBuffer(RF_Raid_t *raidPtr, RF_DagHeader_t *dag_h, int size) 350 { 351 RF_VoidPointerListElem_t *vple; 352 void *p; 353 354 RF_ASSERT((size <= (raidPtr->numCol * (raidPtr->Layout.sectorsPerStripeUnit << 355 raidPtr->logBytesPerSector)))); 356 357 p = malloc( raidPtr->numCol * (raidPtr->Layout.sectorsPerStripeUnit << 358 raidPtr->logBytesPerSector), 359 M_RAIDFRAME, M_NOWAIT); 360 if (!p) { 361 RF_LOCK_MUTEX(raidPtr->mutex); 362 if (raidPtr->stripebuf_count > 0) { 363 vple = raidPtr->stripebuf; 364 raidPtr->stripebuf = vple->next; 365 p = vple->p; 366 rf_FreeVPListElem(vple); 367 raidPtr->stripebuf_count--; 368 } else { 369 #ifdef DIAGNOSTIC 370 printf("raid%d: Help! Out of emergency full-stripe buffers!\n", raidPtr->raidid); 371 #endif 372 } 373 RF_UNLOCK_MUTEX(raidPtr->mutex); 374 if (!p) { 375 /* We didn't get a buffer... not much we can do other than wait, 376 and hope that someone frees up memory for us.. */ 377 p = malloc( raidPtr->numCol * (raidPtr->Layout.sectorsPerStripeUnit << 378 raidPtr->logBytesPerSector), M_RAIDFRAME, M_WAITOK); 379 } 380 } 381 memset(p, 0, raidPtr->numCol * (raidPtr->Layout.sectorsPerStripeUnit << raidPtr->logBytesPerSector)); 382 383 vple = rf_AllocVPListElem(); 384 vple->p = p; 385 vple->next = dag_h->desc->stripebufs; 386 dag_h->desc->stripebufs = vple; 387 388 return (p); 389 } 390 391 392 void 393 rf_FreeStripeBuffer(RF_Raid_t *raidPtr, RF_VoidPointerListElem_t *vple) 394 { 395 RF_LOCK_MUTEX(raidPtr->mutex); 396 if (raidPtr->stripebuf_count < raidPtr->numEmergencyStripeBuffers) { 397 /* just tack it in */ 398 vple->next = raidPtr->stripebuf; 399 raidPtr->stripebuf = vple; 400 raidPtr->stripebuf_count++; 401 } else { 402 free(vple->p, M_RAIDFRAME); 403 rf_FreeVPListElem(vple); 404 } 405 RF_UNLOCK_MUTEX(raidPtr->mutex); 406 } 407 408 /* allocates a buffer big enough to hold the data described by the 409 caller (ie. the data of the associated PDA). Glue this buffer 410 into our dag_h cleanup structure. */ 411 412 void * 413 rf_AllocBuffer(RF_Raid_t *raidPtr, RF_DagHeader_t *dag_h, int size) 414 { 415 RF_VoidPointerListElem_t *vple; 416 void *p; 417 418 p = rf_AllocIOBuffer(raidPtr, size); 419 vple = rf_AllocVPListElem(); 420 vple->p = p; 421 vple->next = dag_h->desc->iobufs; 422 dag_h->desc->iobufs = vple; 423 424 return (p); 425 } 426 427 void * 428 rf_AllocIOBuffer(RF_Raid_t *raidPtr, int size) 429 { 430 RF_VoidPointerListElem_t *vple; 431 void *p; 432 433 RF_ASSERT((size <= (raidPtr->Layout.sectorsPerStripeUnit << 434 raidPtr->logBytesPerSector))); 435 436 p = malloc( raidPtr->Layout.sectorsPerStripeUnit << 437 raidPtr->logBytesPerSector, 438 M_RAIDFRAME, M_NOWAIT); 439 if (!p) { 440 RF_LOCK_MUTEX(raidPtr->mutex); 441 if (raidPtr->iobuf_count > 0) { 442 vple = raidPtr->iobuf; 443 raidPtr->iobuf = vple->next; 444 p = vple->p; 445 rf_FreeVPListElem(vple); 446 raidPtr->iobuf_count--; 447 } else { 448 #ifdef DIAGNOSTIC 449 printf("raid%d: Help! Out of emergency buffers!\n", raidPtr->raidid); 450 #endif 451 } 452 RF_UNLOCK_MUTEX(raidPtr->mutex); 453 if (!p) { 454 /* We didn't get a buffer... not much we can do other than wait, 455 and hope that someone frees up memory for us.. */ 456 p = malloc( raidPtr->Layout.sectorsPerStripeUnit << 457 raidPtr->logBytesPerSector, 458 M_RAIDFRAME, M_WAITOK); 459 } 460 } 461 memset(p, 0, raidPtr->Layout.sectorsPerStripeUnit << raidPtr->logBytesPerSector); 462 return (p); 463 } 464 465 void 466 rf_FreeIOBuffer(RF_Raid_t *raidPtr, RF_VoidPointerListElem_t *vple) 467 { 468 RF_LOCK_MUTEX(raidPtr->mutex); 469 if (raidPtr->iobuf_count < raidPtr->numEmergencyBuffers) { 470 /* just tack it in */ 471 vple->next = raidPtr->iobuf; 472 raidPtr->iobuf = vple; 473 raidPtr->iobuf_count++; 474 } else { 475 free(vple->p, M_RAIDFRAME); 476 rf_FreeVPListElem(vple); 477 } 478 RF_UNLOCK_MUTEX(raidPtr->mutex); 479 } 480 481 482 483 #if RF_DEBUG_VALIDATE_DAG 484 /****************************************************************************** 485 * 486 * debug routines 487 * 488 *****************************************************************************/ 489 490 char * 491 rf_NodeStatusString(RF_DagNode_t *node) 492 { 493 switch (node->status) { 494 case rf_wait: 495 return ("wait"); 496 case rf_fired: 497 return ("fired"); 498 case rf_good: 499 return ("good"); 500 case rf_bad: 501 return ("bad"); 502 default: 503 return ("?"); 504 } 505 } 506 507 void 508 rf_PrintNodeInfoString(RF_DagNode_t *node) 509 { 510 RF_PhysDiskAddr_t *pda; 511 int (*df) (RF_DagNode_t *) = node->doFunc; 512 int i, lk, unlk; 513 void *bufPtr; 514 515 if ((df == rf_DiskReadFunc) || (df == rf_DiskWriteFunc) 516 || (df == rf_DiskReadMirrorIdleFunc) 517 || (df == rf_DiskReadMirrorPartitionFunc)) { 518 pda = (RF_PhysDiskAddr_t *) node->params[0].p; 519 bufPtr = (void *) node->params[1].p; 520 lk = 0; 521 unlk = 0; 522 RF_ASSERT(!(lk && unlk)); 523 printf("c %d offs %ld nsect %d buf 0x%lx %s\n", pda->col, 524 (long) pda->startSector, (int) pda->numSector, (long) bufPtr, 525 (lk) ? "LOCK" : ((unlk) ? "UNLK" : " ")); 526 return; 527 } 528 if ((df == rf_SimpleXorFunc) || (df == rf_RegularXorFunc) 529 || (df == rf_RecoveryXorFunc)) { 530 printf("result buf 0x%lx\n", (long) node->results[0]); 531 for (i = 0; i < node->numParams - 1; i += 2) { 532 pda = (RF_PhysDiskAddr_t *) node->params[i].p; 533 bufPtr = (RF_PhysDiskAddr_t *) node->params[i + 1].p; 534 printf(" buf 0x%lx c%d offs %ld nsect %d\n", 535 (long) bufPtr, pda->col, 536 (long) pda->startSector, (int) pda->numSector); 537 } 538 return; 539 } 540 #if RF_INCLUDE_PARITYLOGGING > 0 541 if (df == rf_ParityLogOverwriteFunc || df == rf_ParityLogUpdateFunc) { 542 for (i = 0; i < node->numParams - 1; i += 2) { 543 pda = (RF_PhysDiskAddr_t *) node->params[i].p; 544 bufPtr = (RF_PhysDiskAddr_t *) node->params[i + 1].p; 545 printf(" c%d offs %ld nsect %d buf 0x%lx\n", 546 pda->col, (long) pda->startSector, 547 (int) pda->numSector, (long) bufPtr); 548 } 549 return; 550 } 551 #endif /* RF_INCLUDE_PARITYLOGGING > 0 */ 552 553 if ((df == rf_TerminateFunc) || (df == rf_NullNodeFunc)) { 554 printf("\n"); 555 return; 556 } 557 printf("?\n"); 558 } 559 #ifdef DEBUG 560 static void 561 rf_RecurPrintDAG(RF_DagNode_t *node, int depth, int unvisited) 562 { 563 char *anttype; 564 int i; 565 566 node->visited = (unvisited) ? 0 : 1; 567 printf("(%d) %d C%d %s: %s,s%d %d/%d,a%d/%d,p%d,r%d S{", depth, 568 node->nodeNum, node->commitNode, node->name, rf_NodeStatusString(node), 569 node->numSuccedents, node->numSuccFired, node->numSuccDone, 570 node->numAntecedents, node->numAntDone, node->numParams, node->numResults); 571 for (i = 0; i < node->numSuccedents; i++) { 572 printf("%d%s", node->succedents[i]->nodeNum, 573 ((i == node->numSuccedents - 1) ? "\0" : " ")); 574 } 575 printf("} A{"); 576 for (i = 0; i < node->numAntecedents; i++) { 577 switch (node->antType[i]) { 578 case rf_trueData: 579 anttype = "T"; 580 break; 581 case rf_antiData: 582 anttype = "A"; 583 break; 584 case rf_outputData: 585 anttype = "O"; 586 break; 587 case rf_control: 588 anttype = "C"; 589 break; 590 default: 591 anttype = "?"; 592 break; 593 } 594 printf("%d(%s)%s", node->antecedents[i]->nodeNum, anttype, (i == node->numAntecedents - 1) ? "\0" : " "); 595 } 596 printf("}; "); 597 rf_PrintNodeInfoString(node); 598 for (i = 0; i < node->numSuccedents; i++) { 599 if (node->succedents[i]->visited == unvisited) 600 rf_RecurPrintDAG(node->succedents[i], depth + 1, unvisited); 601 } 602 } 603 604 static void 605 rf_PrintDAG(RF_DagHeader_t *dag_h) 606 { 607 int unvisited, i; 608 char *status; 609 610 /* set dag status */ 611 switch (dag_h->status) { 612 case rf_enable: 613 status = "enable"; 614 break; 615 case rf_rollForward: 616 status = "rollForward"; 617 break; 618 case rf_rollBackward: 619 status = "rollBackward"; 620 break; 621 default: 622 status = "illegal!"; 623 break; 624 } 625 /* find out if visited bits are currently set or clear */ 626 unvisited = dag_h->succedents[0]->visited; 627 628 printf("DAG type: %s\n", dag_h->creator); 629 printf("format is (depth) num commit type: status,nSucc nSuccFired/nSuccDone,nAnte/nAnteDone,nParam,nResult S{x} A{x(type)}; info\n"); 630 printf("(0) %d Hdr: %s, s%d, (commit %d/%d) S{", dag_h->nodeNum, 631 status, dag_h->numSuccedents, dag_h->numCommitNodes, dag_h->numCommits); 632 for (i = 0; i < dag_h->numSuccedents; i++) { 633 printf("%d%s", dag_h->succedents[i]->nodeNum, 634 ((i == dag_h->numSuccedents - 1) ? "\0" : " ")); 635 } 636 printf("};\n"); 637 for (i = 0; i < dag_h->numSuccedents; i++) { 638 if (dag_h->succedents[i]->visited == unvisited) 639 rf_RecurPrintDAG(dag_h->succedents[i], 1, unvisited); 640 } 641 } 642 #endif 643 /* assigns node numbers */ 644 int 645 rf_AssignNodeNums(RF_DagHeader_t * dag_h) 646 { 647 int unvisited, i, nnum; 648 RF_DagNode_t *node; 649 650 nnum = 0; 651 unvisited = dag_h->succedents[0]->visited; 652 653 dag_h->nodeNum = nnum++; 654 for (i = 0; i < dag_h->numSuccedents; i++) { 655 node = dag_h->succedents[i]; 656 if (node->visited == unvisited) { 657 nnum = rf_RecurAssignNodeNums(dag_h->succedents[i], nnum, unvisited); 658 } 659 } 660 return (nnum); 661 } 662 663 int 664 rf_RecurAssignNodeNums(RF_DagNode_t *node, int num, int unvisited) 665 { 666 int i; 667 668 node->visited = (unvisited) ? 0 : 1; 669 670 node->nodeNum = num++; 671 for (i = 0; i < node->numSuccedents; i++) { 672 if (node->succedents[i]->visited == unvisited) { 673 num = rf_RecurAssignNodeNums(node->succedents[i], num, unvisited); 674 } 675 } 676 return (num); 677 } 678 /* set the header pointers in each node to "newptr" */ 679 void 680 rf_ResetDAGHeaderPointers(RF_DagHeader_t *dag_h, RF_DagHeader_t *newptr) 681 { 682 int i; 683 for (i = 0; i < dag_h->numSuccedents; i++) 684 if (dag_h->succedents[i]->dagHdr != newptr) 685 rf_RecurResetDAGHeaderPointers(dag_h->succedents[i], newptr); 686 } 687 688 void 689 rf_RecurResetDAGHeaderPointers(RF_DagNode_t *node, RF_DagHeader_t *newptr) 690 { 691 int i; 692 node->dagHdr = newptr; 693 for (i = 0; i < node->numSuccedents; i++) 694 if (node->succedents[i]->dagHdr != newptr) 695 rf_RecurResetDAGHeaderPointers(node->succedents[i], newptr); 696 } 697 698 699 void 700 rf_PrintDAGList(RF_DagHeader_t * dag_h) 701 { 702 int i = 0; 703 704 for (; dag_h; dag_h = dag_h->next) { 705 rf_AssignNodeNums(dag_h); 706 printf("\n\nDAG %d IN LIST:\n", i++); 707 rf_PrintDAG(dag_h); 708 } 709 } 710 711 static int 712 rf_ValidateBranch(RF_DagNode_t *node, int *scount, int *acount, 713 RF_DagNode_t **nodes, int unvisited) 714 { 715 int i, retcode = 0; 716 717 /* construct an array of node pointers indexed by node num */ 718 node->visited = (unvisited) ? 0 : 1; 719 nodes[node->nodeNum] = node; 720 721 if (node->next != NULL) { 722 printf("INVALID DAG: next pointer in node is not NULL\n"); 723 retcode = 1; 724 } 725 if (node->status != rf_wait) { 726 printf("INVALID DAG: Node status is not wait\n"); 727 retcode = 1; 728 } 729 if (node->numAntDone != 0) { 730 printf("INVALID DAG: numAntDone is not zero\n"); 731 retcode = 1; 732 } 733 if (node->doFunc == rf_TerminateFunc) { 734 if (node->numSuccedents != 0) { 735 printf("INVALID DAG: Terminator node has succedents\n"); 736 retcode = 1; 737 } 738 } else { 739 if (node->numSuccedents == 0) { 740 printf("INVALID DAG: Non-terminator node has no succedents\n"); 741 retcode = 1; 742 } 743 } 744 for (i = 0; i < node->numSuccedents; i++) { 745 if (!node->succedents[i]) { 746 printf("INVALID DAG: succedent %d of node %s is NULL\n", i, node->name); 747 retcode = 1; 748 } 749 scount[node->succedents[i]->nodeNum]++; 750 } 751 for (i = 0; i < node->numAntecedents; i++) { 752 if (!node->antecedents[i]) { 753 printf("INVALID DAG: antecedent %d of node %s is NULL\n", i, node->name); 754 retcode = 1; 755 } 756 acount[node->antecedents[i]->nodeNum]++; 757 } 758 for (i = 0; i < node->numSuccedents; i++) { 759 if (node->succedents[i]->visited == unvisited) { 760 if (rf_ValidateBranch(node->succedents[i], scount, 761 acount, nodes, unvisited)) { 762 retcode = 1; 763 } 764 } 765 } 766 return (retcode); 767 } 768 769 static void 770 rf_ValidateBranchVisitedBits(RF_DagNode_t *node, int unvisited, int rl) 771 { 772 int i; 773 774 RF_ASSERT(node->visited == unvisited); 775 for (i = 0; i < node->numSuccedents; i++) { 776 if (node->succedents[i] == NULL) { 777 printf("node=%lx node->succedents[%d] is NULL\n", (long) node, i); 778 RF_ASSERT(0); 779 } 780 rf_ValidateBranchVisitedBits(node->succedents[i], unvisited, rl + 1); 781 } 782 } 783 /* NOTE: never call this on a big dag, because it is exponential 784 * in execution time 785 */ 786 static void 787 rf_ValidateVisitedBits(RF_DagHeader_t *dag) 788 { 789 int i, unvisited; 790 791 unvisited = dag->succedents[0]->visited; 792 793 for (i = 0; i < dag->numSuccedents; i++) { 794 if (dag->succedents[i] == NULL) { 795 printf("dag=%lx dag->succedents[%d] is NULL\n", (long) dag, i); 796 RF_ASSERT(0); 797 } 798 rf_ValidateBranchVisitedBits(dag->succedents[i], unvisited, 0); 799 } 800 } 801 /* validate a DAG. _at entry_ verify that: 802 * -- numNodesCompleted is zero 803 * -- node queue is null 804 * -- dag status is rf_enable 805 * -- next pointer is null on every node 806 * -- all nodes have status wait 807 * -- numAntDone is zero in all nodes 808 * -- terminator node has zero successors 809 * -- no other node besides terminator has zero successors 810 * -- no successor or antecedent pointer in a node is NULL 811 * -- number of times that each node appears as a successor of another node 812 * is equal to the antecedent count on that node 813 * -- number of times that each node appears as an antecedent of another node 814 * is equal to the succedent count on that node 815 * -- what else? 816 */ 817 int 818 rf_ValidateDAG(RF_DagHeader_t *dag_h) 819 { 820 int i, nodecount; 821 int *scount, *acount;/* per-node successor and antecedent counts */ 822 RF_DagNode_t **nodes; /* array of ptrs to nodes in dag */ 823 int retcode = 0; 824 int unvisited; 825 int commitNodeCount = 0; 826 827 if (rf_validateVisitedDebug) 828 rf_ValidateVisitedBits(dag_h); 829 830 if (dag_h->numNodesCompleted != 0) { 831 printf("INVALID DAG: num nodes completed is %d, should be 0\n", dag_h->numNodesCompleted); 832 retcode = 1; 833 goto validate_dag_bad; 834 } 835 if (dag_h->status != rf_enable) { 836 printf("INVALID DAG: not enabled\n"); 837 retcode = 1; 838 goto validate_dag_bad; 839 } 840 if (dag_h->numCommits != 0) { 841 printf("INVALID DAG: numCommits != 0 (%d)\n", dag_h->numCommits); 842 retcode = 1; 843 goto validate_dag_bad; 844 } 845 if (dag_h->numSuccedents != 1) { 846 /* currently, all dags must have only one succedent */ 847 printf("INVALID DAG: numSuccedents !1 (%d)\n", dag_h->numSuccedents); 848 retcode = 1; 849 goto validate_dag_bad; 850 } 851 nodecount = rf_AssignNodeNums(dag_h); 852 853 unvisited = dag_h->succedents[0]->visited; 854 855 RF_Malloc(scount, nodecount * sizeof(int), (int *)); 856 RF_Malloc(acount, nodecount * sizeof(int), (int *)); 857 RF_Malloc(nodes, nodecount * sizeof(RF_DagNode_t *), 858 (RF_DagNode_t **)); 859 for (i = 0; i < dag_h->numSuccedents; i++) { 860 if ((dag_h->succedents[i]->visited == unvisited) 861 && rf_ValidateBranch(dag_h->succedents[i], scount, 862 acount, nodes, unvisited)) { 863 retcode = 1; 864 } 865 } 866 /* start at 1 to skip the header node */ 867 for (i = 1; i < nodecount; i++) { 868 if (nodes[i]->commitNode) 869 commitNodeCount++; 870 if (nodes[i]->doFunc == NULL) { 871 printf("INVALID DAG: node %s has an undefined doFunc\n", nodes[i]->name); 872 retcode = 1; 873 goto validate_dag_out; 874 } 875 if (nodes[i]->undoFunc == NULL) { 876 printf("INVALID DAG: node %s has an undefined doFunc\n", nodes[i]->name); 877 retcode = 1; 878 goto validate_dag_out; 879 } 880 if (nodes[i]->numAntecedents != scount[nodes[i]->nodeNum]) { 881 printf("INVALID DAG: node %s has %d antecedents but appears as a succedent %d times\n", 882 nodes[i]->name, nodes[i]->numAntecedents, scount[nodes[i]->nodeNum]); 883 retcode = 1; 884 goto validate_dag_out; 885 } 886 if (nodes[i]->numSuccedents != acount[nodes[i]->nodeNum]) { 887 printf("INVALID DAG: node %s has %d succedents but appears as an antecedent %d times\n", 888 nodes[i]->name, nodes[i]->numSuccedents, acount[nodes[i]->nodeNum]); 889 retcode = 1; 890 goto validate_dag_out; 891 } 892 } 893 894 if (dag_h->numCommitNodes != commitNodeCount) { 895 printf("INVALID DAG: incorrect commit node count. hdr->numCommitNodes (%d) found (%d) commit nodes in graph\n", 896 dag_h->numCommitNodes, commitNodeCount); 897 retcode = 1; 898 goto validate_dag_out; 899 } 900 validate_dag_out: 901 RF_Free(scount, nodecount * sizeof(int)); 902 RF_Free(acount, nodecount * sizeof(int)); 903 RF_Free(nodes, nodecount * sizeof(RF_DagNode_t *)); 904 if (retcode) 905 rf_PrintDAGList(dag_h); 906 907 if (rf_validateVisitedDebug) 908 rf_ValidateVisitedBits(dag_h); 909 910 return (retcode); 911 912 validate_dag_bad: 913 rf_PrintDAGList(dag_h); 914 return (retcode); 915 } 916 917 #endif /* RF_DEBUG_VALIDATE_DAG */ 918 919 /****************************************************************************** 920 * 921 * misc construction routines 922 * 923 *****************************************************************************/ 924 925 void 926 rf_redirect_asm(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap) 927 { 928 int ds = (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) ? 1 : 0; 929 int fcol = raidPtr->reconControl->fcol; 930 int scol = raidPtr->reconControl->spareCol; 931 RF_PhysDiskAddr_t *pda; 932 933 RF_ASSERT(raidPtr->status == rf_rs_reconstructing); 934 for (pda = asmap->physInfo; pda; pda = pda->next) { 935 if (pda->col == fcol) { 936 #if RF_DEBUG_DAG 937 if (rf_dagDebug) { 938 if (!rf_CheckRUReconstructed(raidPtr->reconControl->reconMap, 939 pda->startSector)) { 940 RF_PANIC(); 941 } 942 } 943 #endif 944 /* printf("Remapped data for large write\n"); */ 945 if (ds) { 946 raidPtr->Layout.map->MapSector(raidPtr, pda->raidAddress, 947 &pda->col, &pda->startSector, RF_REMAP); 948 } else { 949 pda->col = scol; 950 } 951 } 952 } 953 for (pda = asmap->parityInfo; pda; pda = pda->next) { 954 if (pda->col == fcol) { 955 #if RF_DEBUG_DAG 956 if (rf_dagDebug) { 957 if (!rf_CheckRUReconstructed(raidPtr->reconControl->reconMap, pda->startSector)) { 958 RF_PANIC(); 959 } 960 } 961 #endif 962 } 963 if (ds) { 964 (raidPtr->Layout.map->MapParity) (raidPtr, pda->raidAddress, &pda->col, &pda->startSector, RF_REMAP); 965 } else { 966 pda->col = scol; 967 } 968 } 969 } 970 971 972 /* this routine allocates read buffers and generates stripe maps for the 973 * regions of the array from the start of the stripe to the start of the 974 * access, and from the end of the access to the end of the stripe. It also 975 * computes and returns the number of DAG nodes needed to read all this data. 976 * Note that this routine does the wrong thing if the access is fully 977 * contained within one stripe unit, so we RF_ASSERT against this case at the 978 * start. 979 * 980 * layoutPtr - in: layout information 981 * asmap - in: access stripe map 982 * dag_h - in: header of the dag to create 983 * new_asm_h - in: ptr to array of 2 headers. to be filled in 984 * nRodNodes - out: num nodes to be generated to read unaccessed data 985 * sosBuffer, eosBuffer - out: pointers to newly allocated buffer 986 */ 987 void 988 rf_MapUnaccessedPortionOfStripe(RF_Raid_t *raidPtr, 989 RF_RaidLayout_t *layoutPtr, 990 RF_AccessStripeMap_t *asmap, 991 RF_DagHeader_t *dag_h, 992 RF_AccessStripeMapHeader_t **new_asm_h, 993 int *nRodNodes, 994 char **sosBuffer, char **eosBuffer, 995 RF_AllocListElem_t *allocList) 996 { 997 RF_RaidAddr_t sosRaidAddress, eosRaidAddress; 998 RF_SectorNum_t sosNumSector, eosNumSector; 999 1000 RF_ASSERT(asmap->numStripeUnitsAccessed > (layoutPtr->numDataCol / 2)); 1001 /* generate an access map for the region of the array from start of 1002 * stripe to start of access */ 1003 new_asm_h[0] = new_asm_h[1] = NULL; 1004 *nRodNodes = 0; 1005 if (!rf_RaidAddressStripeAligned(layoutPtr, asmap->raidAddress)) { 1006 sosRaidAddress = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 1007 sosNumSector = asmap->raidAddress - sosRaidAddress; 1008 *sosBuffer = rf_AllocStripeBuffer(raidPtr, dag_h, rf_RaidAddressToByte(raidPtr, sosNumSector)); 1009 new_asm_h[0] = rf_MapAccess(raidPtr, sosRaidAddress, sosNumSector, *sosBuffer, RF_DONT_REMAP); 1010 new_asm_h[0]->next = dag_h->asmList; 1011 dag_h->asmList = new_asm_h[0]; 1012 *nRodNodes += new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 1013 1014 RF_ASSERT(new_asm_h[0]->stripeMap->next == NULL); 1015 /* we're totally within one stripe here */ 1016 if (asmap->flags & RF_ASM_REDIR_LARGE_WRITE) 1017 rf_redirect_asm(raidPtr, new_asm_h[0]->stripeMap); 1018 } 1019 /* generate an access map for the region of the array from end of 1020 * access to end of stripe */ 1021 if (!rf_RaidAddressStripeAligned(layoutPtr, asmap->endRaidAddress)) { 1022 eosRaidAddress = asmap->endRaidAddress; 1023 eosNumSector = rf_RaidAddressOfNextStripeBoundary(layoutPtr, eosRaidAddress) - eosRaidAddress; 1024 *eosBuffer = rf_AllocStripeBuffer(raidPtr, dag_h, rf_RaidAddressToByte(raidPtr, eosNumSector)); 1025 new_asm_h[1] = rf_MapAccess(raidPtr, eosRaidAddress, eosNumSector, *eosBuffer, RF_DONT_REMAP); 1026 new_asm_h[1]->next = dag_h->asmList; 1027 dag_h->asmList = new_asm_h[1]; 1028 *nRodNodes += new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 1029 1030 RF_ASSERT(new_asm_h[1]->stripeMap->next == NULL); 1031 /* we're totally within one stripe here */ 1032 if (asmap->flags & RF_ASM_REDIR_LARGE_WRITE) 1033 rf_redirect_asm(raidPtr, new_asm_h[1]->stripeMap); 1034 } 1035 } 1036 1037 1038 1039 /* returns non-zero if the indicated ranges of stripe unit offsets overlap */ 1040 int 1041 rf_PDAOverlap(RF_RaidLayout_t *layoutPtr, 1042 RF_PhysDiskAddr_t *src, RF_PhysDiskAddr_t *dest) 1043 { 1044 RF_SectorNum_t soffs = rf_StripeUnitOffset(layoutPtr, src->startSector); 1045 RF_SectorNum_t doffs = rf_StripeUnitOffset(layoutPtr, dest->startSector); 1046 /* use -1 to be sure we stay within SU */ 1047 RF_SectorNum_t send = rf_StripeUnitOffset(layoutPtr, src->startSector + src->numSector - 1); 1048 RF_SectorNum_t dend = rf_StripeUnitOffset(layoutPtr, dest->startSector + dest->numSector - 1); 1049 return ((RF_MAX(soffs, doffs) <= RF_MIN(send, dend)) ? 1 : 0); 1050 } 1051 1052 1053 /* GenerateFailedAccessASMs 1054 * 1055 * this routine figures out what portion of the stripe needs to be read 1056 * to effect the degraded read or write operation. It's primary function 1057 * is to identify everything required to recover the data, and then 1058 * eliminate anything that is already being accessed by the user. 1059 * 1060 * The main result is two new ASMs, one for the region from the start of the 1061 * stripe to the start of the access, and one for the region from the end of 1062 * the access to the end of the stripe. These ASMs describe everything that 1063 * needs to be read to effect the degraded access. Other results are: 1064 * nXorBufs -- the total number of buffers that need to be XORed together to 1065 * recover the lost data, 1066 * rpBufPtr -- ptr to a newly-allocated buffer to hold the parity. If NULL 1067 * at entry, not allocated. 1068 * overlappingPDAs -- 1069 * describes which of the non-failed PDAs in the user access 1070 * overlap data that needs to be read to effect recovery. 1071 * overlappingPDAs[i]==1 if and only if, neglecting the failed 1072 * PDA, the ith pda in the input asm overlaps data that needs 1073 * to be read for recovery. 1074 */ 1075 /* in: asm - ASM for the actual access, one stripe only */ 1076 /* in: failedPDA - which component of the access has failed */ 1077 /* in: dag_h - header of the DAG we're going to create */ 1078 /* out: new_asm_h - the two new ASMs */ 1079 /* out: nXorBufs - the total number of xor bufs required */ 1080 /* out: rpBufPtr - a buffer for the parity read */ 1081 void 1082 rf_GenerateFailedAccessASMs(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap, 1083 RF_PhysDiskAddr_t *failedPDA, 1084 RF_DagHeader_t *dag_h, 1085 RF_AccessStripeMapHeader_t **new_asm_h, 1086 int *nXorBufs, char **rpBufPtr, 1087 char *overlappingPDAs, 1088 RF_AllocListElem_t *allocList) 1089 { 1090 RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout); 1091 1092 /* s=start, e=end, s=stripe, a=access, f=failed, su=stripe unit */ 1093 RF_RaidAddr_t sosAddr, sosEndAddr, eosStartAddr, eosAddr; 1094 RF_PhysDiskAddr_t *pda; 1095 int foundit, i; 1096 1097 foundit = 0; 1098 /* first compute the following raid addresses: start of stripe, 1099 * (sosAddr) MIN(start of access, start of failed SU), (sosEndAddr) 1100 * MAX(end of access, end of failed SU), (eosStartAddr) end of 1101 * stripe (i.e. start of next stripe) (eosAddr) */ 1102 sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress); 1103 sosEndAddr = RF_MIN(asmap->raidAddress, rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->raidAddress)); 1104 eosStartAddr = RF_MAX(asmap->endRaidAddress, rf_RaidAddressOfNextStripeUnitBoundary(layoutPtr, failedPDA->raidAddress)); 1105 eosAddr = rf_RaidAddressOfNextStripeBoundary(layoutPtr, asmap->raidAddress); 1106 1107 /* now generate access stripe maps for each of the above regions of 1108 * the stripe. Use a dummy (NULL) buf ptr for now */ 1109 1110 new_asm_h[0] = (sosAddr != sosEndAddr) ? rf_MapAccess(raidPtr, sosAddr, sosEndAddr - sosAddr, NULL, RF_DONT_REMAP) : NULL; 1111 new_asm_h[1] = (eosStartAddr != eosAddr) ? rf_MapAccess(raidPtr, eosStartAddr, eosAddr - eosStartAddr, NULL, RF_DONT_REMAP) : NULL; 1112 1113 /* walk through the PDAs and range-restrict each SU to the region of 1114 * the SU touched on the failed PDA. also compute total data buffer 1115 * space requirements in this step. Ignore the parity for now. */ 1116 /* Also count nodes to find out how many bufs need to be xored together */ 1117 (*nXorBufs) = 1; /* in read case, 1 is for parity. In write 1118 * case, 1 is for failed data */ 1119 1120 if (new_asm_h[0]) { 1121 new_asm_h[0]->next = dag_h->asmList; 1122 dag_h->asmList = new_asm_h[0]; 1123 for (pda = new_asm_h[0]->stripeMap->physInfo; pda; pda = pda->next) { 1124 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_NOBUFFER, 0); 1125 pda->bufPtr = rf_AllocBuffer(raidPtr, dag_h, pda->numSector << raidPtr->logBytesPerSector); 1126 } 1127 (*nXorBufs) += new_asm_h[0]->stripeMap->numStripeUnitsAccessed; 1128 } 1129 if (new_asm_h[1]) { 1130 new_asm_h[1]->next = dag_h->asmList; 1131 dag_h->asmList = new_asm_h[1]; 1132 for (pda = new_asm_h[1]->stripeMap->physInfo; pda; pda = pda->next) { 1133 rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_NOBUFFER, 0); 1134 pda->bufPtr = rf_AllocBuffer(raidPtr, dag_h, pda->numSector << raidPtr->logBytesPerSector); 1135 } 1136 (*nXorBufs) += new_asm_h[1]->stripeMap->numStripeUnitsAccessed; 1137 } 1138 1139 /* allocate a buffer for parity */ 1140 if (rpBufPtr) 1141 *rpBufPtr = rf_AllocBuffer(raidPtr, dag_h, failedPDA->numSector << raidPtr->logBytesPerSector); 1142 1143 /* the last step is to figure out how many more distinct buffers need 1144 * to get xor'd to produce the missing unit. there's one for each 1145 * user-data read node that overlaps the portion of the failed unit 1146 * being accessed */ 1147 1148 for (foundit = i = 0, pda = asmap->physInfo; pda; i++, pda = pda->next) { 1149 if (pda == failedPDA) { 1150 i--; 1151 foundit = 1; 1152 continue; 1153 } 1154 if (rf_PDAOverlap(layoutPtr, pda, failedPDA)) { 1155 overlappingPDAs[i] = 1; 1156 (*nXorBufs)++; 1157 } 1158 } 1159 if (!foundit) { 1160 RF_ERRORMSG("GenerateFailedAccessASMs: did not find failedPDA in asm list\n"); 1161 RF_ASSERT(0); 1162 } 1163 #if RF_DEBUG_DAG 1164 if (rf_degDagDebug) { 1165 if (new_asm_h[0]) { 1166 printf("First asm:\n"); 1167 rf_PrintFullAccessStripeMap(new_asm_h[0], 1); 1168 } 1169 if (new_asm_h[1]) { 1170 printf("Second asm:\n"); 1171 rf_PrintFullAccessStripeMap(new_asm_h[1], 1); 1172 } 1173 } 1174 #endif 1175 } 1176 1177 1178 /* adjusts the offset and number of sectors in the destination pda so that 1179 * it covers at most the region of the SU covered by the source PDA. This 1180 * is exclusively a restriction: the number of sectors indicated by the 1181 * target PDA can only shrink. 1182 * 1183 * For example: s = sectors within SU indicated by source PDA 1184 * d = sectors within SU indicated by dest PDA 1185 * r = results, stored in dest PDA 1186 * 1187 * |--------------- one stripe unit ---------------------| 1188 * | sssssssssssssssssssssssssssssssss | 1189 * | ddddddddddddddddddddddddddddddddddddddddddddd | 1190 * | rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr | 1191 * 1192 * Another example: 1193 * 1194 * |--------------- one stripe unit ---------------------| 1195 * | sssssssssssssssssssssssssssssssss | 1196 * | ddddddddddddddddddddddd | 1197 * | rrrrrrrrrrrrrrrr | 1198 * 1199 */ 1200 void 1201 rf_RangeRestrictPDA(RF_Raid_t *raidPtr, RF_PhysDiskAddr_t *src, 1202 RF_PhysDiskAddr_t *dest, int dobuffer, int doraidaddr) 1203 { 1204 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout; 1205 RF_SectorNum_t soffs = rf_StripeUnitOffset(layoutPtr, src->startSector); 1206 RF_SectorNum_t doffs = rf_StripeUnitOffset(layoutPtr, dest->startSector); 1207 RF_SectorNum_t send = rf_StripeUnitOffset(layoutPtr, src->startSector + src->numSector - 1); /* use -1 to be sure we 1208 * stay within SU */ 1209 RF_SectorNum_t dend = rf_StripeUnitOffset(layoutPtr, dest->startSector + dest->numSector - 1); 1210 RF_SectorNum_t subAddr = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, dest->startSector); /* stripe unit boundary */ 1211 1212 dest->startSector = subAddr + RF_MAX(soffs, doffs); 1213 dest->numSector = subAddr + RF_MIN(send, dend) + 1 - dest->startSector; 1214 1215 if (dobuffer) 1216 dest->bufPtr += (soffs > doffs) ? rf_RaidAddressToByte(raidPtr, soffs - doffs) : 0; 1217 if (doraidaddr) { 1218 dest->raidAddress = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, dest->raidAddress) + 1219 rf_StripeUnitOffset(layoutPtr, dest->startSector); 1220 } 1221 } 1222 1223 #if (RF_INCLUDE_CHAINDECLUSTER > 0) 1224 1225 /* 1226 * Want the highest of these primes to be the largest one 1227 * less than the max expected number of columns (won't hurt 1228 * to be too small or too large, but won't be optimal, either) 1229 * --jimz 1230 */ 1231 #define NLOWPRIMES 8 1232 static int lowprimes[NLOWPRIMES] = {2, 3, 5, 7, 11, 13, 17, 19}; 1233 /***************************************************************************** 1234 * compute the workload shift factor. (chained declustering) 1235 * 1236 * return nonzero if access should shift to secondary, otherwise, 1237 * access is to primary 1238 *****************************************************************************/ 1239 int 1240 rf_compute_workload_shift(RF_Raid_t *raidPtr, RF_PhysDiskAddr_t *pda) 1241 { 1242 /* 1243 * variables: 1244 * d = column of disk containing primary 1245 * f = column of failed disk 1246 * n = number of disks in array 1247 * sd = "shift distance" (number of columns that d is to the right of f) 1248 * v = numerator of redirection ratio 1249 * k = denominator of redirection ratio 1250 */ 1251 RF_RowCol_t d, f, sd, n; 1252 int k, v, ret, i; 1253 1254 n = raidPtr->numCol; 1255 1256 /* assign column of primary copy to d */ 1257 d = pda->col; 1258 1259 /* assign column of dead disk to f */ 1260 for (f = 0; ((!RF_DEAD_DISK(raidPtr->Disks[f].status)) && (f < n)); f++); 1261 1262 RF_ASSERT(f < n); 1263 RF_ASSERT(f != d); 1264 1265 sd = (f > d) ? (n + d - f) : (d - f); 1266 RF_ASSERT(sd < n); 1267 1268 /* 1269 * v of every k accesses should be redirected 1270 * 1271 * v/k := (n-1-sd)/(n-1) 1272 */ 1273 v = (n - 1 - sd); 1274 k = (n - 1); 1275 1276 #if 1 1277 /* 1278 * XXX 1279 * Is this worth it? 1280 * 1281 * Now reduce the fraction, by repeatedly factoring 1282 * out primes (just like they teach in elementary school!) 1283 */ 1284 for (i = 0; i < NLOWPRIMES; i++) { 1285 if (lowprimes[i] > v) 1286 break; 1287 while (((v % lowprimes[i]) == 0) && ((k % lowprimes[i]) == 0)) { 1288 v /= lowprimes[i]; 1289 k /= lowprimes[i]; 1290 } 1291 } 1292 #endif 1293 1294 raidPtr->hist_diskreq[d]++; 1295 if (raidPtr->hist_diskreq[d] > v) { 1296 ret = 0; /* do not redirect */ 1297 } else { 1298 ret = 1; /* redirect */ 1299 } 1300 1301 #if 0 1302 printf("d=%d f=%d sd=%d v=%d k=%d ret=%d h=%d\n", d, f, sd, v, k, ret, 1303 raidPtr->hist_diskreq[d]); 1304 #endif 1305 1306 if (raidPtr->hist_diskreq[d] >= k) { 1307 /* reset counter */ 1308 raidPtr->hist_diskreq[d] = 0; 1309 } 1310 return (ret); 1311 } 1312 #endif /* (RF_INCLUDE_CHAINDECLUSTER > 0) */ 1313 1314 /* 1315 * Disk selection routines 1316 */ 1317 1318 /* 1319 * Selects the disk with the shortest queue from a mirror pair. 1320 * Both the disk I/Os queued in RAIDframe as well as those at the physical 1321 * disk are counted as members of the "queue" 1322 */ 1323 void 1324 rf_SelectMirrorDiskIdle(RF_DagNode_t * node) 1325 { 1326 RF_Raid_t *raidPtr = (RF_Raid_t *) node->dagHdr->raidPtr; 1327 RF_RowCol_t colData, colMirror; 1328 int dataQueueLength, mirrorQueueLength, usemirror; 1329 RF_PhysDiskAddr_t *data_pda = (RF_PhysDiskAddr_t *) node->params[0].p; 1330 RF_PhysDiskAddr_t *mirror_pda = (RF_PhysDiskAddr_t *) node->params[4].p; 1331 RF_PhysDiskAddr_t *tmp_pda; 1332 RF_RaidDisk_t *disks = raidPtr->Disks; 1333 RF_DiskQueue_t *dqs = raidPtr->Queues, *dataQueue, *mirrorQueue; 1334 1335 /* return the [row col] of the disk with the shortest queue */ 1336 colData = data_pda->col; 1337 colMirror = mirror_pda->col; 1338 dataQueue = &(dqs[colData]); 1339 mirrorQueue = &(dqs[colMirror]); 1340 1341 #ifdef RF_LOCK_QUEUES_TO_READ_LEN 1342 RF_LOCK_QUEUE_MUTEX(dataQueue, "SelectMirrorDiskIdle"); 1343 #endif /* RF_LOCK_QUEUES_TO_READ_LEN */ 1344 dataQueueLength = dataQueue->queueLength + dataQueue->numOutstanding; 1345 #ifdef RF_LOCK_QUEUES_TO_READ_LEN 1346 RF_UNLOCK_QUEUE_MUTEX(dataQueue, "SelectMirrorDiskIdle"); 1347 RF_LOCK_QUEUE_MUTEX(mirrorQueue, "SelectMirrorDiskIdle"); 1348 #endif /* RF_LOCK_QUEUES_TO_READ_LEN */ 1349 mirrorQueueLength = mirrorQueue->queueLength + mirrorQueue->numOutstanding; 1350 #ifdef RF_LOCK_QUEUES_TO_READ_LEN 1351 RF_UNLOCK_QUEUE_MUTEX(mirrorQueue, "SelectMirrorDiskIdle"); 1352 #endif /* RF_LOCK_QUEUES_TO_READ_LEN */ 1353 1354 usemirror = 0; 1355 if (RF_DEAD_DISK(disks[colMirror].status)) { 1356 usemirror = 0; 1357 } else 1358 if (RF_DEAD_DISK(disks[colData].status)) { 1359 usemirror = 1; 1360 } else 1361 if (raidPtr->parity_good == RF_RAID_DIRTY) { 1362 /* Trust only the main disk */ 1363 usemirror = 0; 1364 } else 1365 if (dataQueueLength < mirrorQueueLength) { 1366 usemirror = 0; 1367 } else 1368 if (mirrorQueueLength < dataQueueLength) { 1369 usemirror = 1; 1370 } else { 1371 /* queues are equal length. attempt 1372 * cleverness. */ 1373 if (SNUM_DIFF(dataQueue->last_deq_sector, data_pda->startSector) 1374 <= SNUM_DIFF(mirrorQueue->last_deq_sector, mirror_pda->startSector)) { 1375 usemirror = 0; 1376 } else { 1377 usemirror = 1; 1378 } 1379 } 1380 1381 if (usemirror) { 1382 /* use mirror (parity) disk, swap params 0 & 4 */ 1383 tmp_pda = data_pda; 1384 node->params[0].p = mirror_pda; 1385 node->params[4].p = tmp_pda; 1386 } else { 1387 /* use data disk, leave param 0 unchanged */ 1388 } 1389 /* printf("dataQueueLength %d, mirrorQueueLength 1390 * %d\n",dataQueueLength, mirrorQueueLength); */ 1391 } 1392 #if (RF_INCLUDE_CHAINDECLUSTER > 0) || (RF_INCLUDE_INTERDECLUSTER > 0) || (RF_DEBUG_VALIDATE_DAG > 0) 1393 /* 1394 * Do simple partitioning. This assumes that 1395 * the data and parity disks are laid out identically. 1396 */ 1397 void 1398 rf_SelectMirrorDiskPartition(RF_DagNode_t * node) 1399 { 1400 RF_Raid_t *raidPtr = (RF_Raid_t *) node->dagHdr->raidPtr; 1401 RF_RowCol_t colData, colMirror; 1402 RF_PhysDiskAddr_t *data_pda = (RF_PhysDiskAddr_t *) node->params[0].p; 1403 RF_PhysDiskAddr_t *mirror_pda = (RF_PhysDiskAddr_t *) node->params[4].p; 1404 RF_PhysDiskAddr_t *tmp_pda; 1405 RF_RaidDisk_t *disks = raidPtr->Disks; 1406 int usemirror; 1407 1408 /* return the [row col] of the disk with the shortest queue */ 1409 colData = data_pda->col; 1410 colMirror = mirror_pda->col; 1411 1412 usemirror = 0; 1413 if (RF_DEAD_DISK(disks[colMirror].status)) { 1414 usemirror = 0; 1415 } else 1416 if (RF_DEAD_DISK(disks[colData].status)) { 1417 usemirror = 1; 1418 } else 1419 if (raidPtr->parity_good == RF_RAID_DIRTY) { 1420 /* Trust only the main disk */ 1421 usemirror = 0; 1422 } else 1423 if (data_pda->startSector < 1424 (disks[colData].numBlocks / 2)) { 1425 usemirror = 0; 1426 } else { 1427 usemirror = 1; 1428 } 1429 1430 if (usemirror) { 1431 /* use mirror (parity) disk, swap params 0 & 4 */ 1432 tmp_pda = data_pda; 1433 node->params[0].p = mirror_pda; 1434 node->params[4].p = tmp_pda; 1435 } else { 1436 /* use data disk, leave param 0 unchanged */ 1437 } 1438 } 1439 #endif 1440