Home | History | Annotate | Line # | Download | only in raidframe
rf_states.c revision 1.37
      1 /*	$NetBSD: rf_states.c,v 1.37 2005/01/14 01:33:15 oster Exp $	*/
      2 /*
      3  * Copyright (c) 1995 Carnegie-Mellon University.
      4  * All rights reserved.
      5  *
      6  * Author: Mark Holland, William V. Courtright II, Robby Findler
      7  *
      8  * Permission to use, copy, modify and distribute this software and
      9  * its documentation is hereby granted, provided that both the copyright
     10  * notice and this permission notice appear in all copies of the
     11  * software, derivative works or modified versions, and any portions
     12  * thereof, and that both notices appear in supporting documentation.
     13  *
     14  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
     15  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
     16  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
     17  *
     18  * Carnegie Mellon requests users of this software to return to
     19  *
     20  *  Software Distribution Coordinator  or  Software.Distribution (at) CS.CMU.EDU
     21  *  School of Computer Science
     22  *  Carnegie Mellon University
     23  *  Pittsburgh PA 15213-3890
     24  *
     25  * any improvements or extensions that they make and grant Carnegie the
     26  * rights to redistribute these changes.
     27  */
     28 
     29 #include <sys/cdefs.h>
     30 __KERNEL_RCSID(0, "$NetBSD: rf_states.c,v 1.37 2005/01/14 01:33:15 oster Exp $");
     31 
     32 #include <sys/errno.h>
     33 
     34 #include "rf_archs.h"
     35 #include "rf_threadstuff.h"
     36 #include "rf_raid.h"
     37 #include "rf_dag.h"
     38 #include "rf_desc.h"
     39 #include "rf_aselect.h"
     40 #include "rf_general.h"
     41 #include "rf_states.h"
     42 #include "rf_dagutils.h"
     43 #include "rf_driver.h"
     44 #include "rf_engine.h"
     45 #include "rf_map.h"
     46 #include "rf_etimer.h"
     47 #include "rf_kintf.h"
     48 
     49 #ifndef RF_DEBUG_STATES
     50 #define RF_DEBUG_STATES 0
     51 #endif
     52 
     53 /* prototypes for some of the available states.
     54 
     55    States must:
     56 
     57      - not block.
     58 
     59      - either schedule rf_ContinueRaidAccess as a callback and return
     60        RF_TRUE, or complete all of their work and return RF_FALSE.
     61 
     62      - increment desc->state when they have finished their work.
     63 */
     64 
     65 #if RF_DEBUG_STATES
     66 static char *
     67 StateName(RF_AccessState_t state)
     68 {
     69 	switch (state) {
     70 		case rf_QuiesceState:return "QuiesceState";
     71 	case rf_MapState:
     72 		return "MapState";
     73 	case rf_LockState:
     74 		return "LockState";
     75 	case rf_CreateDAGState:
     76 		return "CreateDAGState";
     77 	case rf_ExecuteDAGState:
     78 		return "ExecuteDAGState";
     79 	case rf_ProcessDAGState:
     80 		return "ProcessDAGState";
     81 	case rf_CleanupState:
     82 		return "CleanupState";
     83 	case rf_LastState:
     84 		return "LastState";
     85 	case rf_IncrAccessesCountState:
     86 		return "IncrAccessesCountState";
     87 	case rf_DecrAccessesCountState:
     88 		return "DecrAccessesCountState";
     89 	default:
     90 		return "!!! UnnamedState !!!";
     91 	}
     92 }
     93 #endif
     94 
     95 void
     96 rf_ContinueRaidAccess(RF_RaidAccessDesc_t *desc)
     97 {
     98 	int     suspended = RF_FALSE;
     99 	int     current_state_index = desc->state;
    100 	RF_AccessState_t current_state = desc->states[current_state_index];
    101 #if RF_DEBUG_STATES
    102 	int     unit = desc->raidPtr->raidid;
    103 #endif
    104 
    105 	do {
    106 
    107 		current_state_index = desc->state;
    108 		current_state = desc->states[current_state_index];
    109 
    110 		switch (current_state) {
    111 
    112 		case rf_QuiesceState:
    113 			suspended = rf_State_Quiesce(desc);
    114 			break;
    115 		case rf_IncrAccessesCountState:
    116 			suspended = rf_State_IncrAccessCount(desc);
    117 			break;
    118 		case rf_MapState:
    119 			suspended = rf_State_Map(desc);
    120 			break;
    121 		case rf_LockState:
    122 			suspended = rf_State_Lock(desc);
    123 			break;
    124 		case rf_CreateDAGState:
    125 			suspended = rf_State_CreateDAG(desc);
    126 			break;
    127 		case rf_ExecuteDAGState:
    128 			suspended = rf_State_ExecuteDAG(desc);
    129 			break;
    130 		case rf_ProcessDAGState:
    131 			suspended = rf_State_ProcessDAG(desc);
    132 			break;
    133 		case rf_CleanupState:
    134 			suspended = rf_State_Cleanup(desc);
    135 			break;
    136 		case rf_DecrAccessesCountState:
    137 			suspended = rf_State_DecrAccessCount(desc);
    138 			break;
    139 		case rf_LastState:
    140 			suspended = rf_State_LastState(desc);
    141 			break;
    142 		}
    143 
    144 		/* after this point, we cannot dereference desc since
    145 		 * desc may have been freed. desc is only freed in
    146 		 * LastState, so if we renter this function or loop
    147 		 * back up, desc should be valid. */
    148 
    149 #if RF_DEBUG_STATES
    150 		if (rf_printStatesDebug) {
    151 			printf("raid%d: State: %-24s StateIndex: %3i desc: 0x%ld %s\n",
    152 			       unit, StateName(current_state),
    153 			       current_state_index, (long) desc,
    154 			       suspended ? "callback scheduled" : "looping");
    155 		}
    156 #endif
    157 	} while (!suspended && current_state != rf_LastState);
    158 
    159 	return;
    160 }
    161 
    162 
    163 void
    164 rf_ContinueDagAccess(RF_DagList_t *dagList)
    165 {
    166 #if RF_ACC_TRACE > 0
    167 	RF_AccTraceEntry_t *tracerec = &(dagList->desc->tracerec);
    168 	RF_Etimer_t timer;
    169 #endif
    170 	RF_RaidAccessDesc_t *desc;
    171 	RF_DagHeader_t *dag_h;
    172 	int     i;
    173 
    174 	desc = dagList->desc;
    175 
    176 #if RF_ACC_TRACE > 0
    177 	timer = tracerec->timer;
    178 	RF_ETIMER_STOP(timer);
    179 	RF_ETIMER_EVAL(timer);
    180 	tracerec->specific.user.exec_us = RF_ETIMER_VAL_US(timer);
    181 	RF_ETIMER_START(tracerec->timer);
    182 #endif
    183 
    184 	/* skip to dag which just finished */
    185 	dag_h = dagList->dags;
    186 	for (i = 0; i < dagList->numDagsDone; i++) {
    187 		dag_h = dag_h->next;
    188 	}
    189 
    190 	/* check to see if retry is required */
    191 	if (dag_h->status == rf_rollBackward) {
    192 		/* when a dag fails, mark desc status as bad and allow
    193 		 * all other dags in the desc to execute to
    194 		 * completion.  then, free all dags and start over */
    195 		desc->status = 1;	/* bad status */
    196 #if 0
    197 		printf("raid%d: DAG failure: %c addr 0x%lx "
    198 		       "(%ld) nblk 0x%x (%d) buf 0x%lx state %d\n",
    199 		       desc->raidPtr->raidid, desc->type,
    200 		       (long) desc->raidAddress,
    201 		       (long) desc->raidAddress, (int) desc->numBlocks,
    202 		       (int) desc->numBlocks,
    203 		       (unsigned long) (desc->bufPtr), desc->state);
    204 #endif
    205 	}
    206 	dagList->numDagsDone++;
    207 	rf_ContinueRaidAccess(desc);
    208 }
    209 
    210 int
    211 rf_State_LastState(RF_RaidAccessDesc_t *desc)
    212 {
    213 	void    (*callbackFunc) (RF_CBParam_t) = desc->callbackFunc;
    214 	RF_CBParam_t callbackArg;
    215 
    216 	callbackArg.p = desc->callbackArg;
    217 
    218 	/*
    219 	 * If this is not an async request, wake up the caller
    220 	 */
    221 	if (desc->async_flag == 0)
    222 		wakeup(desc->bp);
    223 
    224 	/*
    225 	 * That's all the IO for this one... unbusy the 'disk'.
    226 	 */
    227 
    228 	rf_disk_unbusy(desc);
    229 
    230 	/*
    231 	 * Wakeup any requests waiting to go.
    232 	 */
    233 
    234 	RF_LOCK_MUTEX(((RF_Raid_t *) desc->raidPtr)->mutex);
    235 	((RF_Raid_t *) desc->raidPtr)->openings++;
    236 	RF_UNLOCK_MUTEX(((RF_Raid_t *) desc->raidPtr)->mutex);
    237 
    238 	/* wake up any pending IO */
    239 	raidstart(((RF_Raid_t *) desc->raidPtr));
    240 
    241 	/* printf("Calling biodone on 0x%x\n",desc->bp); */
    242 	biodone(desc->bp);	/* access came through ioctl */
    243 
    244 	if (callbackFunc)
    245 		callbackFunc(callbackArg);
    246 	rf_FreeRaidAccDesc(desc);
    247 
    248 	return RF_FALSE;
    249 }
    250 
    251 int
    252 rf_State_IncrAccessCount(RF_RaidAccessDesc_t *desc)
    253 {
    254 	RF_Raid_t *raidPtr;
    255 
    256 	raidPtr = desc->raidPtr;
    257 	/* Bummer. We have to do this to be 100% safe w.r.t. the increment
    258 	 * below */
    259 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    260 	raidPtr->accs_in_flight++;	/* used to detect quiescence */
    261 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    262 
    263 	desc->state++;
    264 	return RF_FALSE;
    265 }
    266 
    267 int
    268 rf_State_DecrAccessCount(RF_RaidAccessDesc_t *desc)
    269 {
    270 	RF_Raid_t *raidPtr;
    271 
    272 	raidPtr = desc->raidPtr;
    273 
    274 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    275 	raidPtr->accs_in_flight--;
    276 	if (raidPtr->accesses_suspended && raidPtr->accs_in_flight == 0) {
    277 		rf_SignalQuiescenceLock(raidPtr);
    278 	}
    279 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    280 
    281 	desc->state++;
    282 	return RF_FALSE;
    283 }
    284 
    285 int
    286 rf_State_Quiesce(RF_RaidAccessDesc_t *desc)
    287 {
    288 #if RF_ACC_TRACE > 0
    289 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    290 	RF_Etimer_t timer;
    291 #endif
    292 	RF_CallbackDesc_t *cb;
    293 	RF_Raid_t *raidPtr;
    294 	int     suspended = RF_FALSE;
    295 	int need_cb, used_cb;
    296 
    297 	raidPtr = desc->raidPtr;
    298 
    299 #if RF_ACC_TRACE > 0
    300 	RF_ETIMER_START(timer);
    301 	RF_ETIMER_START(desc->timer);
    302 #endif
    303 
    304 	need_cb = 0;
    305 	used_cb = 0;
    306 	cb = NULL;
    307 
    308 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    309 	/* Do an initial check to see if we might need a callback structure */
    310 	if (raidPtr->accesses_suspended) {
    311 		need_cb = 1;
    312 	}
    313 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    314 
    315 	if (need_cb) {
    316 		/* create a callback if we might need it...
    317 		   and we likely do. */
    318 		cb = rf_AllocCallbackDesc();
    319 	}
    320 
    321 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    322 	if (raidPtr->accesses_suspended) {
    323 		cb->callbackFunc = (void (*) (RF_CBParam_t)) rf_ContinueRaidAccess;
    324 		cb->callbackArg.p = (void *) desc;
    325 		cb->next = raidPtr->quiesce_wait_list;
    326 		raidPtr->quiesce_wait_list = cb;
    327 		suspended = RF_TRUE;
    328 		used_cb = 1;
    329 	}
    330 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    331 
    332 	if ((need_cb == 1) && (used_cb == 0)) {
    333 		rf_FreeCallbackDesc(cb);
    334 	}
    335 
    336 #if RF_ACC_TRACE > 0
    337 	RF_ETIMER_STOP(timer);
    338 	RF_ETIMER_EVAL(timer);
    339 	tracerec->specific.user.suspend_ovhd_us += RF_ETIMER_VAL_US(timer);
    340 #endif
    341 
    342 #if RF_DEBUG_QUIESCE
    343 	if (suspended && rf_quiesceDebug)
    344 		printf("Stalling access due to quiescence lock\n");
    345 #endif
    346 	desc->state++;
    347 	return suspended;
    348 }
    349 
    350 int
    351 rf_State_Map(RF_RaidAccessDesc_t *desc)
    352 {
    353 	RF_Raid_t *raidPtr = desc->raidPtr;
    354 #if RF_ACC_TRACE > 0
    355 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    356 	RF_Etimer_t timer;
    357 
    358 	RF_ETIMER_START(timer);
    359 #endif
    360 
    361 	if (!(desc->asmap = rf_MapAccess(raidPtr, desc->raidAddress, desc->numBlocks,
    362 		    desc->bufPtr, RF_DONT_REMAP)))
    363 		RF_PANIC();
    364 
    365 #if RF_ACC_TRACE > 0
    366 	RF_ETIMER_STOP(timer);
    367 	RF_ETIMER_EVAL(timer);
    368 	tracerec->specific.user.map_us = RF_ETIMER_VAL_US(timer);
    369 #endif
    370 
    371 	desc->state++;
    372 	return RF_FALSE;
    373 }
    374 
    375 int
    376 rf_State_Lock(RF_RaidAccessDesc_t *desc)
    377 {
    378 #if RF_ACC_TRACE > 0
    379 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    380 	RF_Etimer_t timer;
    381 #endif
    382 	RF_Raid_t *raidPtr = desc->raidPtr;
    383 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    384 	RF_AccessStripeMap_t *asm_p;
    385 	RF_StripeNum_t lastStripeID = -1;
    386 	int     suspended = RF_FALSE;
    387 
    388 #if RF_ACC_TRACE > 0
    389 	RF_ETIMER_START(timer);
    390 #endif
    391 
    392 	/* acquire each lock that we don't already hold */
    393 	for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) {
    394 		RF_ASSERT(RF_IO_IS_R_OR_W(desc->type));
    395 		if (!rf_suppressLocksAndLargeWrites &&
    396 		    asm_p->parityInfo &&
    397 		    !(desc->flags & RF_DAG_SUPPRESS_LOCKS) &&
    398 		    !(asm_p->flags & RF_ASM_FLAGS_LOCK_TRIED)) {
    399 			asm_p->flags |= RF_ASM_FLAGS_LOCK_TRIED;
    400 				/* locks must be acquired hierarchically */
    401 			RF_ASSERT(asm_p->stripeID > lastStripeID);
    402 			lastStripeID = asm_p->stripeID;
    403 
    404 			RF_INIT_LOCK_REQ_DESC(asm_p->lockReqDesc, desc->type,
    405 					      (void (*) (struct buf *)) rf_ContinueRaidAccess, desc, asm_p,
    406 					      raidPtr->Layout.dataSectorsPerStripe);
    407 			if (rf_AcquireStripeLock(raidPtr->lockTable, asm_p->stripeID,
    408 						 &asm_p->lockReqDesc)) {
    409 				suspended = RF_TRUE;
    410 				break;
    411 			}
    412 		}
    413 		if (desc->type == RF_IO_TYPE_WRITE &&
    414 		    raidPtr->status == rf_rs_reconstructing) {
    415 			if (!(asm_p->flags & RF_ASM_FLAGS_FORCE_TRIED)) {
    416 				int     val;
    417 
    418 				asm_p->flags |= RF_ASM_FLAGS_FORCE_TRIED;
    419 				val = rf_ForceOrBlockRecon(raidPtr, asm_p,
    420 							   (void (*) (RF_Raid_t *, void *)) rf_ContinueRaidAccess, desc);
    421 				if (val == 0) {
    422 					asm_p->flags |= RF_ASM_FLAGS_RECON_BLOCKED;
    423 				} else {
    424 					suspended = RF_TRUE;
    425 					break;
    426 				}
    427 			} else {
    428 #if RF_DEBUG_PSS > 0
    429 				if (rf_pssDebug) {
    430 					printf("raid%d: skipping force/block because already done, psid %ld\n",
    431 					       desc->raidPtr->raidid,
    432 					       (long) asm_p->stripeID);
    433 				}
    434 #endif
    435 			}
    436 		} else {
    437 #if RF_DEBUG_PSS > 0
    438 			if (rf_pssDebug) {
    439 				printf("raid%d: skipping force/block because not write or not under recon, psid %ld\n",
    440 				       desc->raidPtr->raidid,
    441 				       (long) asm_p->stripeID);
    442 			}
    443 #endif
    444 		}
    445 	}
    446 #if RF_ACC_TRACE > 0
    447 	RF_ETIMER_STOP(timer);
    448 	RF_ETIMER_EVAL(timer);
    449 	tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer);
    450 #endif
    451 	if (suspended)
    452 		return (RF_TRUE);
    453 
    454 	desc->state++;
    455 	return (RF_FALSE);
    456 }
    457 /*
    458  * the following three states create, execute, and post-process dags
    459  * the error recovery unit is a single dag.
    460  * by default, SelectAlgorithm creates an array of dags, one per parity stripe
    461  * in some tricky cases, multiple dags per stripe are created
    462  *   - dags within a parity stripe are executed sequentially (arbitrary order)
    463  *   - dags for distinct parity stripes are executed concurrently
    464  *
    465  * repeat until all dags complete successfully -or- dag selection fails
    466  *
    467  * while !done
    468  *   create dag(s) (SelectAlgorithm)
    469  *   if dag
    470  *     execute dag (DispatchDAG)
    471  *     if dag successful
    472  *       done (SUCCESS)
    473  *     else
    474  *       !done (RETRY - start over with new dags)
    475  *   else
    476  *     done (FAIL)
    477  */
    478 int
    479 rf_State_CreateDAG(RF_RaidAccessDesc_t *desc)
    480 {
    481 #if RF_ACC_TRACE > 0
    482 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    483 	RF_Etimer_t timer;
    484 #endif
    485 	RF_DagHeader_t *dag_h;
    486 	RF_DagList_t *dagList;
    487 	struct buf *bp;
    488 	int     i, selectStatus;
    489 
    490 	/* generate a dag for the access, and fire it off.  When the dag
    491 	 * completes, we'll get re-invoked in the next state. */
    492 #if RF_ACC_TRACE > 0
    493 	RF_ETIMER_START(timer);
    494 #endif
    495 	/* SelectAlgorithm returns one or more dags */
    496 	selectStatus = rf_SelectAlgorithm(desc, desc->flags | RF_DAG_SUPPRESS_LOCKS);
    497 #if RF_DEBUG_VALIDATE_DAG
    498 	if (rf_printDAGsDebug) {
    499 		dagList = desc->dagList;
    500 		for (i = 0; i < desc->numStripes; i++) {
    501 			rf_PrintDAGList(dagList.dags);
    502 			dagList = dagList->next;
    503 		}
    504 	}
    505 #endif /* RF_DEBUG_VALIDATE_DAG */
    506 #if RF_ACC_TRACE > 0
    507 	RF_ETIMER_STOP(timer);
    508 	RF_ETIMER_EVAL(timer);
    509 	/* update time to create all dags */
    510 	tracerec->specific.user.dag_create_us = RF_ETIMER_VAL_US(timer);
    511 #endif
    512 
    513 	desc->status = 0;	/* good status */
    514 
    515 	if (selectStatus || (desc->numRetries > RF_RETRY_THRESHOLD)) {
    516 		/* failed to create a dag */
    517 		/* this happens when there are too many faults or incomplete
    518 		 * dag libraries */
    519 		if (selectStatus) {
    520 			printf("raid%d: failed to create a dag. "
    521 			       "Too many component failures.\n",
    522 			       desc->raidPtr->raidid);
    523 		} else {
    524 			printf("raid%d: IO failed after %d retries.\n",
    525 			       desc->raidPtr->raidid, RF_RETRY_THRESHOLD);
    526 		}
    527 
    528 		desc->status = 1; /* bad status */
    529 		/* skip straight to rf_State_Cleanup() */
    530 		desc->state = rf_CleanupState;
    531 		bp = (struct buf *)desc->bp;
    532 		bp->b_flags |= B_ERROR;
    533 		bp->b_error = EIO;
    534 	} else {
    535 		/* bind dags to desc */
    536 		dagList = desc->dagList;
    537 		for (i = 0; i < desc->numStripes; i++) {
    538 			dag_h = dagList->dags;
    539 			while (dag_h) {
    540 				dag_h->bp = (struct buf *) desc->bp;
    541 #if RF_ACC_TRACE > 0
    542 				dag_h->tracerec = tracerec;
    543 #endif
    544 				dag_h = dag_h->next;
    545 			}
    546 			dagList = dagList->next;
    547 		}
    548 		desc->flags |= RF_DAG_DISPATCH_RETURNED;
    549 		desc->state++;	/* next state should be rf_State_ExecuteDAG */
    550 	}
    551 	return RF_FALSE;
    552 }
    553 
    554 
    555 
    556 /* the access has an list of dagLists, one dagList per parity stripe.
    557  * fire the first dag in each parity stripe (dagList).
    558  * dags within a stripe (dagList) must be executed sequentially
    559  *  - this preserves atomic parity update
    560  * dags for independents parity groups (stripes) are fired concurrently */
    561 
    562 int
    563 rf_State_ExecuteDAG(RF_RaidAccessDesc_t *desc)
    564 {
    565 	int     i;
    566 	RF_DagHeader_t *dag_h;
    567 	RF_DagList_t *dagList;
    568 
    569 	/* next state is always rf_State_ProcessDAG important to do
    570 	 * this before firing the first dag (it may finish before we
    571 	 * leave this routine) */
    572 	desc->state++;
    573 
    574 	/* sweep dag array, a stripe at a time, firing the first dag
    575 	 * in each stripe */
    576 	dagList = desc->dagList;
    577 	for (i = 0; i < desc->numStripes; i++) {
    578 		RF_ASSERT(dagList->numDags > 0);
    579 		RF_ASSERT(dagList->numDagsDone == 0);
    580 		RF_ASSERT(dagList->numDagsFired == 0);
    581 #if RF_ACC_TRACE > 0
    582 		RF_ETIMER_START(dagList->tracerec.timer);
    583 #endif
    584 		/* fire first dag in this stripe */
    585 		dag_h = dagList->dags;
    586 		RF_ASSERT(dag_h);
    587 		dagList->numDagsFired++;
    588 		rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess, dagList);
    589 		dagList = dagList->next;
    590 	}
    591 
    592 	/* the DAG will always call the callback, even if there was no
    593 	 * blocking, so we are always suspended in this state */
    594 	return RF_TRUE;
    595 }
    596 
    597 
    598 
    599 /* rf_State_ProcessDAG is entered when a dag completes.
    600  * first, check to all dags in the access have completed
    601  * if not, fire as many dags as possible */
    602 
    603 int
    604 rf_State_ProcessDAG(RF_RaidAccessDesc_t *desc)
    605 {
    606 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    607 	RF_Raid_t *raidPtr = desc->raidPtr;
    608 	RF_DagHeader_t *dag_h;
    609 	int     i, j, done = RF_TRUE;
    610 	RF_DagList_t *dagList, *temp;
    611 
    612 	/* check to see if this is the last dag */
    613 	dagList = desc->dagList;
    614 	for (i = 0; i < desc->numStripes; i++) {
    615 		if (dagList->numDags != dagList->numDagsDone)
    616 			done = RF_FALSE;
    617 		dagList = dagList->next;
    618 	}
    619 
    620 	if (done) {
    621 		if (desc->status) {
    622 			/* a dag failed, retry */
    623 			/* free all dags */
    624 			dagList = desc->dagList;
    625 			for (i = 0; i < desc->numStripes; i++) {
    626 				rf_FreeDAG(dagList->dags);
    627 				temp = dagList;
    628 				dagList = dagList->next;
    629 				rf_FreeDAGList(temp);
    630 			}
    631 			desc->dagList = NULL;
    632 
    633 			rf_MarkFailuresInASMList(raidPtr, asmh);
    634 
    635 			/* note the retry so that we'll bail in
    636 			   rf_State_CreateDAG() once we've retired
    637 			   the IO RF_RETRY_THRESHOLD times */
    638 
    639 			desc->numRetries++;
    640 
    641 			/* back up to rf_State_CreateDAG */
    642 			desc->state = desc->state - 2;
    643 			return RF_FALSE;
    644 		} else {
    645 			/* move on to rf_State_Cleanup */
    646 			desc->state++;
    647 		}
    648 		return RF_FALSE;
    649 	} else {
    650 		/* more dags to execute */
    651 		/* see if any are ready to be fired.  if so, fire them */
    652 		/* don't fire the initial dag in a list, it's fired in
    653 		 * rf_State_ExecuteDAG */
    654 		dagList = desc->dagList;
    655 		for (i = 0; i < desc->numStripes; i++) {
    656 			if ((dagList->numDagsDone < dagList->numDags)
    657 			    && (dagList->numDagsDone == dagList->numDagsFired)
    658 			    && (dagList->numDagsFired > 0)) {
    659 #if RF_ACC_TRACE > 0
    660 				RF_ETIMER_START(dagList->tracerec.timer);
    661 #endif
    662 				/* fire next dag in this stripe */
    663 				/* first, skip to next dag awaiting execution */
    664 				dag_h = dagList->dags;
    665 				for (j = 0; j < dagList->numDagsDone; j++)
    666 					dag_h = dag_h->next;
    667 				dagList->numDagsFired++;
    668 				rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess,
    669 				    dagList);
    670 			}
    671 			dagList = dagList->next;
    672 		}
    673 		return RF_TRUE;
    674 	}
    675 }
    676 /* only make it this far if all dags complete successfully */
    677 int
    678 rf_State_Cleanup(RF_RaidAccessDesc_t *desc)
    679 {
    680 #if RF_ACC_TRACE > 0
    681 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    682 	RF_Etimer_t timer;
    683 #endif
    684 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    685 	RF_Raid_t *raidPtr = desc->raidPtr;
    686 	RF_AccessStripeMap_t *asm_p;
    687 	RF_DagList_t *dagList;
    688 	int i;
    689 
    690 	desc->state++;
    691 
    692 #if RF_ACC_TRACE > 0
    693 	timer = tracerec->timer;
    694 	RF_ETIMER_STOP(timer);
    695 	RF_ETIMER_EVAL(timer);
    696 	tracerec->specific.user.dag_retry_us = RF_ETIMER_VAL_US(timer);
    697 
    698 	/* the RAID I/O is complete.  Clean up. */
    699 	tracerec->specific.user.dag_retry_us = 0;
    700 
    701 	RF_ETIMER_START(timer);
    702 #endif
    703 	/* free all dags */
    704 	dagList = desc->dagList;
    705 	for (i = 0; i < desc->numStripes; i++) {
    706 		rf_FreeDAG(dagList->dags);
    707 		dagList = dagList->next;
    708 	}
    709 #if RF_ACC_TRACE > 0
    710 	RF_ETIMER_STOP(timer);
    711 	RF_ETIMER_EVAL(timer);
    712 	tracerec->specific.user.cleanup_us = RF_ETIMER_VAL_US(timer);
    713 
    714 	RF_ETIMER_START(timer);
    715 #endif
    716 	for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) {
    717 		if (!rf_suppressLocksAndLargeWrites &&
    718 		    asm_p->parityInfo &&
    719 		    !(desc->flags & RF_DAG_SUPPRESS_LOCKS)) {
    720 			RF_ASSERT_VALID_LOCKREQ(&asm_p->lockReqDesc);
    721 			rf_ReleaseStripeLock(raidPtr->lockTable,
    722 					     asm_p->stripeID,
    723 					     &asm_p->lockReqDesc);
    724 		}
    725 		if (asm_p->flags & RF_ASM_FLAGS_RECON_BLOCKED) {
    726 			rf_UnblockRecon(raidPtr, asm_p);
    727 		}
    728 	}
    729 #if RF_ACC_TRACE > 0
    730 	RF_ETIMER_STOP(timer);
    731 	RF_ETIMER_EVAL(timer);
    732 	tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer);
    733 
    734 	RF_ETIMER_START(timer);
    735 #endif
    736 	rf_FreeAccessStripeMap(asmh);
    737 #if RF_ACC_TRACE > 0
    738 	RF_ETIMER_STOP(timer);
    739 	RF_ETIMER_EVAL(timer);
    740 	tracerec->specific.user.cleanup_us += RF_ETIMER_VAL_US(timer);
    741 
    742 	RF_ETIMER_STOP(desc->timer);
    743 	RF_ETIMER_EVAL(desc->timer);
    744 
    745 	timer = desc->tracerec.tot_timer;
    746 	RF_ETIMER_STOP(timer);
    747 	RF_ETIMER_EVAL(timer);
    748 	desc->tracerec.total_us = RF_ETIMER_VAL_US(timer);
    749 
    750 	rf_LogTraceRec(raidPtr, tracerec);
    751 #endif
    752 	desc->flags |= RF_DAG_ACCESS_COMPLETE;
    753 
    754 	return RF_FALSE;
    755 }
    756