Home | History | Annotate | Line # | Download | only in raidframe
rf_states.c revision 1.39
      1 /*	$NetBSD: rf_states.c,v 1.39 2005/09/25 19:47:17 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 1.39 2005/09/25 19:47:17 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 	wakeup(&(desc->raidPtr->iodone));
    239 
    240 	/* printf("Calling biodone on 0x%x\n",desc->bp); */
    241 	biodone(desc->bp);	/* access came through ioctl */
    242 
    243 	if (callbackFunc)
    244 		callbackFunc(callbackArg);
    245 	rf_FreeRaidAccDesc(desc);
    246 
    247 	return RF_FALSE;
    248 }
    249 
    250 int
    251 rf_State_IncrAccessCount(RF_RaidAccessDesc_t *desc)
    252 {
    253 	RF_Raid_t *raidPtr;
    254 
    255 	raidPtr = desc->raidPtr;
    256 	/* Bummer. We have to do this to be 100% safe w.r.t. the increment
    257 	 * below */
    258 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    259 	raidPtr->accs_in_flight++;	/* used to detect quiescence */
    260 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    261 
    262 	desc->state++;
    263 	return RF_FALSE;
    264 }
    265 
    266 int
    267 rf_State_DecrAccessCount(RF_RaidAccessDesc_t *desc)
    268 {
    269 	RF_Raid_t *raidPtr;
    270 
    271 	raidPtr = desc->raidPtr;
    272 
    273 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    274 	raidPtr->accs_in_flight--;
    275 	if (raidPtr->accesses_suspended && raidPtr->accs_in_flight == 0) {
    276 		rf_SignalQuiescenceLock(raidPtr);
    277 	}
    278 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    279 
    280 	desc->state++;
    281 	return RF_FALSE;
    282 }
    283 
    284 int
    285 rf_State_Quiesce(RF_RaidAccessDesc_t *desc)
    286 {
    287 #if RF_ACC_TRACE > 0
    288 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    289 	RF_Etimer_t timer;
    290 #endif
    291 	RF_CallbackDesc_t *cb;
    292 	RF_Raid_t *raidPtr;
    293 	int     suspended = RF_FALSE;
    294 	int need_cb, used_cb;
    295 
    296 	raidPtr = desc->raidPtr;
    297 
    298 #if RF_ACC_TRACE > 0
    299 	RF_ETIMER_START(timer);
    300 	RF_ETIMER_START(desc->timer);
    301 #endif
    302 
    303 	need_cb = 0;
    304 	used_cb = 0;
    305 	cb = NULL;
    306 
    307 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    308 	/* Do an initial check to see if we might need a callback structure */
    309 	if (raidPtr->accesses_suspended) {
    310 		need_cb = 1;
    311 	}
    312 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    313 
    314 	if (need_cb) {
    315 		/* create a callback if we might need it...
    316 		   and we likely do. */
    317 		cb = rf_AllocCallbackDesc();
    318 	}
    319 
    320 	RF_LOCK_MUTEX(raidPtr->access_suspend_mutex);
    321 	if (raidPtr->accesses_suspended) {
    322 		cb->callbackFunc = (void (*) (RF_CBParam_t)) rf_ContinueRaidAccess;
    323 		cb->callbackArg.p = (void *) desc;
    324 		cb->next = raidPtr->quiesce_wait_list;
    325 		raidPtr->quiesce_wait_list = cb;
    326 		suspended = RF_TRUE;
    327 		used_cb = 1;
    328 	}
    329 	RF_UNLOCK_MUTEX(raidPtr->access_suspend_mutex);
    330 
    331 	if ((need_cb == 1) && (used_cb == 0)) {
    332 		rf_FreeCallbackDesc(cb);
    333 	}
    334 
    335 #if RF_ACC_TRACE > 0
    336 	RF_ETIMER_STOP(timer);
    337 	RF_ETIMER_EVAL(timer);
    338 	tracerec->specific.user.suspend_ovhd_us += RF_ETIMER_VAL_US(timer);
    339 #endif
    340 
    341 #if RF_DEBUG_QUIESCE
    342 	if (suspended && rf_quiesceDebug)
    343 		printf("Stalling access due to quiescence lock\n");
    344 #endif
    345 	desc->state++;
    346 	return suspended;
    347 }
    348 
    349 int
    350 rf_State_Map(RF_RaidAccessDesc_t *desc)
    351 {
    352 	RF_Raid_t *raidPtr = desc->raidPtr;
    353 #if RF_ACC_TRACE > 0
    354 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    355 	RF_Etimer_t timer;
    356 
    357 	RF_ETIMER_START(timer);
    358 #endif
    359 
    360 	if (!(desc->asmap = rf_MapAccess(raidPtr, desc->raidAddress, desc->numBlocks,
    361 		    desc->bufPtr, RF_DONT_REMAP)))
    362 		RF_PANIC();
    363 
    364 #if RF_ACC_TRACE > 0
    365 	RF_ETIMER_STOP(timer);
    366 	RF_ETIMER_EVAL(timer);
    367 	tracerec->specific.user.map_us = RF_ETIMER_VAL_US(timer);
    368 #endif
    369 
    370 	desc->state++;
    371 	return RF_FALSE;
    372 }
    373 
    374 int
    375 rf_State_Lock(RF_RaidAccessDesc_t *desc)
    376 {
    377 #if RF_ACC_TRACE > 0
    378 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    379 	RF_Etimer_t timer;
    380 #endif
    381 	RF_Raid_t *raidPtr = desc->raidPtr;
    382 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    383 	RF_AccessStripeMap_t *asm_p;
    384 	RF_StripeNum_t lastStripeID = -1;
    385 	int     suspended = RF_FALSE;
    386 
    387 #if RF_ACC_TRACE > 0
    388 	RF_ETIMER_START(timer);
    389 #endif
    390 
    391 	/* acquire each lock that we don't already hold */
    392 	for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) {
    393 		RF_ASSERT(RF_IO_IS_R_OR_W(desc->type));
    394 		if (!rf_suppressLocksAndLargeWrites &&
    395 		    asm_p->parityInfo &&
    396 		    !(desc->flags & RF_DAG_SUPPRESS_LOCKS) &&
    397 		    !(asm_p->flags & RF_ASM_FLAGS_LOCK_TRIED)) {
    398 			asm_p->flags |= RF_ASM_FLAGS_LOCK_TRIED;
    399 				/* locks must be acquired hierarchically */
    400 			RF_ASSERT(asm_p->stripeID > lastStripeID);
    401 			lastStripeID = asm_p->stripeID;
    402 
    403 			RF_INIT_LOCK_REQ_DESC(asm_p->lockReqDesc, desc->type,
    404 					      (void (*) (struct buf *)) rf_ContinueRaidAccess, desc, asm_p,
    405 					      raidPtr->Layout.dataSectorsPerStripe);
    406 			if (rf_AcquireStripeLock(raidPtr->lockTable, asm_p->stripeID,
    407 						 &asm_p->lockReqDesc)) {
    408 				suspended = RF_TRUE;
    409 				break;
    410 			}
    411 		}
    412 		if (desc->type == RF_IO_TYPE_WRITE &&
    413 		    raidPtr->status == rf_rs_reconstructing) {
    414 			if (!(asm_p->flags & RF_ASM_FLAGS_FORCE_TRIED)) {
    415 				int     val;
    416 
    417 				asm_p->flags |= RF_ASM_FLAGS_FORCE_TRIED;
    418 				val = rf_ForceOrBlockRecon(raidPtr, asm_p,
    419 							   (void (*) (RF_Raid_t *, void *)) rf_ContinueRaidAccess, desc);
    420 				if (val == 0) {
    421 					asm_p->flags |= RF_ASM_FLAGS_RECON_BLOCKED;
    422 				} else {
    423 					suspended = RF_TRUE;
    424 					break;
    425 				}
    426 			} else {
    427 #if RF_DEBUG_PSS > 0
    428 				if (rf_pssDebug) {
    429 					printf("raid%d: skipping force/block because already done, psid %ld\n",
    430 					       desc->raidPtr->raidid,
    431 					       (long) asm_p->stripeID);
    432 				}
    433 #endif
    434 			}
    435 		} else {
    436 #if RF_DEBUG_PSS > 0
    437 			if (rf_pssDebug) {
    438 				printf("raid%d: skipping force/block because not write or not under recon, psid %ld\n",
    439 				       desc->raidPtr->raidid,
    440 				       (long) asm_p->stripeID);
    441 			}
    442 #endif
    443 		}
    444 	}
    445 #if RF_ACC_TRACE > 0
    446 	RF_ETIMER_STOP(timer);
    447 	RF_ETIMER_EVAL(timer);
    448 	tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer);
    449 #endif
    450 	if (suspended)
    451 		return (RF_TRUE);
    452 
    453 	desc->state++;
    454 	return (RF_FALSE);
    455 }
    456 /*
    457  * the following three states create, execute, and post-process dags
    458  * the error recovery unit is a single dag.
    459  * by default, SelectAlgorithm creates an array of dags, one per parity stripe
    460  * in some tricky cases, multiple dags per stripe are created
    461  *   - dags within a parity stripe are executed sequentially (arbitrary order)
    462  *   - dags for distinct parity stripes are executed concurrently
    463  *
    464  * repeat until all dags complete successfully -or- dag selection fails
    465  *
    466  * while !done
    467  *   create dag(s) (SelectAlgorithm)
    468  *   if dag
    469  *     execute dag (DispatchDAG)
    470  *     if dag successful
    471  *       done (SUCCESS)
    472  *     else
    473  *       !done (RETRY - start over with new dags)
    474  *   else
    475  *     done (FAIL)
    476  */
    477 int
    478 rf_State_CreateDAG(RF_RaidAccessDesc_t *desc)
    479 {
    480 #if RF_ACC_TRACE > 0
    481 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    482 	RF_Etimer_t timer;
    483 #endif
    484 	RF_DagHeader_t *dag_h;
    485 	RF_DagList_t *dagList;
    486 	struct buf *bp;
    487 	int     i, selectStatus;
    488 
    489 	/* generate a dag for the access, and fire it off.  When the dag
    490 	 * completes, we'll get re-invoked in the next state. */
    491 #if RF_ACC_TRACE > 0
    492 	RF_ETIMER_START(timer);
    493 #endif
    494 	/* SelectAlgorithm returns one or more dags */
    495 	selectStatus = rf_SelectAlgorithm(desc, desc->flags | RF_DAG_SUPPRESS_LOCKS);
    496 #if RF_DEBUG_VALIDATE_DAG
    497 	if (rf_printDAGsDebug) {
    498 		dagList = desc->dagList;
    499 		for (i = 0; i < desc->numStripes; i++) {
    500 			rf_PrintDAGList(dagList.dags);
    501 			dagList = dagList->next;
    502 		}
    503 	}
    504 #endif /* RF_DEBUG_VALIDATE_DAG */
    505 #if RF_ACC_TRACE > 0
    506 	RF_ETIMER_STOP(timer);
    507 	RF_ETIMER_EVAL(timer);
    508 	/* update time to create all dags */
    509 	tracerec->specific.user.dag_create_us = RF_ETIMER_VAL_US(timer);
    510 #endif
    511 
    512 	desc->status = 0;	/* good status */
    513 
    514 	if (selectStatus || (desc->numRetries > RF_RETRY_THRESHOLD)) {
    515 		/* failed to create a dag */
    516 		/* this happens when there are too many faults or incomplete
    517 		 * dag libraries */
    518 		if (selectStatus) {
    519 			printf("raid%d: failed to create a dag. "
    520 			       "Too many component failures.\n",
    521 			       desc->raidPtr->raidid);
    522 		} else {
    523 			printf("raid%d: IO failed after %d retries.\n",
    524 			       desc->raidPtr->raidid, RF_RETRY_THRESHOLD);
    525 		}
    526 
    527 		desc->status = 1; /* bad status */
    528 		/* skip straight to rf_State_Cleanup() */
    529 		desc->state = rf_CleanupState;
    530 		bp = (struct buf *)desc->bp;
    531 		bp->b_flags |= B_ERROR;
    532 		bp->b_error = EIO;
    533 	} else {
    534 		/* bind dags to desc */
    535 		dagList = desc->dagList;
    536 		for (i = 0; i < desc->numStripes; i++) {
    537 			dag_h = dagList->dags;
    538 			while (dag_h) {
    539 				dag_h->bp = (struct buf *) desc->bp;
    540 #if RF_ACC_TRACE > 0
    541 				dag_h->tracerec = tracerec;
    542 #endif
    543 				dag_h = dag_h->next;
    544 			}
    545 			dagList = dagList->next;
    546 		}
    547 		desc->flags |= RF_DAG_DISPATCH_RETURNED;
    548 		desc->state++;	/* next state should be rf_State_ExecuteDAG */
    549 	}
    550 	return RF_FALSE;
    551 }
    552 
    553 
    554 
    555 /* the access has an list of dagLists, one dagList per parity stripe.
    556  * fire the first dag in each parity stripe (dagList).
    557  * dags within a stripe (dagList) must be executed sequentially
    558  *  - this preserves atomic parity update
    559  * dags for independents parity groups (stripes) are fired concurrently */
    560 
    561 int
    562 rf_State_ExecuteDAG(RF_RaidAccessDesc_t *desc)
    563 {
    564 	int     i;
    565 	RF_DagHeader_t *dag_h;
    566 	RF_DagList_t *dagList;
    567 
    568 	/* next state is always rf_State_ProcessDAG important to do
    569 	 * this before firing the first dag (it may finish before we
    570 	 * leave this routine) */
    571 	desc->state++;
    572 
    573 	/* sweep dag array, a stripe at a time, firing the first dag
    574 	 * in each stripe */
    575 	dagList = desc->dagList;
    576 	for (i = 0; i < desc->numStripes; i++) {
    577 		RF_ASSERT(dagList->numDags > 0);
    578 		RF_ASSERT(dagList->numDagsDone == 0);
    579 		RF_ASSERT(dagList->numDagsFired == 0);
    580 #if RF_ACC_TRACE > 0
    581 		RF_ETIMER_START(dagList->tracerec.timer);
    582 #endif
    583 		/* fire first dag in this stripe */
    584 		dag_h = dagList->dags;
    585 		RF_ASSERT(dag_h);
    586 		dagList->numDagsFired++;
    587 		rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess, dagList);
    588 		dagList = dagList->next;
    589 	}
    590 
    591 	/* the DAG will always call the callback, even if there was no
    592 	 * blocking, so we are always suspended in this state */
    593 	return RF_TRUE;
    594 }
    595 
    596 
    597 
    598 /* rf_State_ProcessDAG is entered when a dag completes.
    599  * first, check to all dags in the access have completed
    600  * if not, fire as many dags as possible */
    601 
    602 int
    603 rf_State_ProcessDAG(RF_RaidAccessDesc_t *desc)
    604 {
    605 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    606 	RF_Raid_t *raidPtr = desc->raidPtr;
    607 	RF_DagHeader_t *dag_h;
    608 	int     i, j, done = RF_TRUE;
    609 	RF_DagList_t *dagList, *temp;
    610 
    611 	/* check to see if this is the last dag */
    612 	dagList = desc->dagList;
    613 	for (i = 0; i < desc->numStripes; i++) {
    614 		if (dagList->numDags != dagList->numDagsDone)
    615 			done = RF_FALSE;
    616 		dagList = dagList->next;
    617 	}
    618 
    619 	if (done) {
    620 		if (desc->status) {
    621 			/* a dag failed, retry */
    622 			/* free all dags */
    623 			dagList = desc->dagList;
    624 			for (i = 0; i < desc->numStripes; i++) {
    625 				rf_FreeDAG(dagList->dags);
    626 				temp = dagList;
    627 				dagList = dagList->next;
    628 				rf_FreeDAGList(temp);
    629 			}
    630 			desc->dagList = NULL;
    631 
    632 			rf_MarkFailuresInASMList(raidPtr, asmh);
    633 
    634 			/* note the retry so that we'll bail in
    635 			   rf_State_CreateDAG() once we've retired
    636 			   the IO RF_RETRY_THRESHOLD times */
    637 
    638 			desc->numRetries++;
    639 
    640 			/* back up to rf_State_CreateDAG */
    641 			desc->state = desc->state - 2;
    642 			return RF_FALSE;
    643 		} else {
    644 			/* move on to rf_State_Cleanup */
    645 			desc->state++;
    646 		}
    647 		return RF_FALSE;
    648 	} else {
    649 		/* more dags to execute */
    650 		/* see if any are ready to be fired.  if so, fire them */
    651 		/* don't fire the initial dag in a list, it's fired in
    652 		 * rf_State_ExecuteDAG */
    653 		dagList = desc->dagList;
    654 		for (i = 0; i < desc->numStripes; i++) {
    655 			if ((dagList->numDagsDone < dagList->numDags)
    656 			    && (dagList->numDagsDone == dagList->numDagsFired)
    657 			    && (dagList->numDagsFired > 0)) {
    658 #if RF_ACC_TRACE > 0
    659 				RF_ETIMER_START(dagList->tracerec.timer);
    660 #endif
    661 				/* fire next dag in this stripe */
    662 				/* first, skip to next dag awaiting execution */
    663 				dag_h = dagList->dags;
    664 				for (j = 0; j < dagList->numDagsDone; j++)
    665 					dag_h = dag_h->next;
    666 				dagList->numDagsFired++;
    667 				rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess,
    668 				    dagList);
    669 			}
    670 			dagList = dagList->next;
    671 		}
    672 		return RF_TRUE;
    673 	}
    674 }
    675 /* only make it this far if all dags complete successfully */
    676 int
    677 rf_State_Cleanup(RF_RaidAccessDesc_t *desc)
    678 {
    679 #if RF_ACC_TRACE > 0
    680 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    681 	RF_Etimer_t timer;
    682 #endif
    683 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    684 	RF_Raid_t *raidPtr = desc->raidPtr;
    685 	RF_AccessStripeMap_t *asm_p;
    686 	RF_DagList_t *dagList;
    687 	int i;
    688 
    689 	desc->state++;
    690 
    691 #if RF_ACC_TRACE > 0
    692 	timer = tracerec->timer;
    693 	RF_ETIMER_STOP(timer);
    694 	RF_ETIMER_EVAL(timer);
    695 	tracerec->specific.user.dag_retry_us = RF_ETIMER_VAL_US(timer);
    696 
    697 	/* the RAID I/O is complete.  Clean up. */
    698 	tracerec->specific.user.dag_retry_us = 0;
    699 
    700 	RF_ETIMER_START(timer);
    701 #endif
    702 	/* free all dags */
    703 	dagList = desc->dagList;
    704 	for (i = 0; i < desc->numStripes; i++) {
    705 		rf_FreeDAG(dagList->dags);
    706 		dagList = dagList->next;
    707 	}
    708 #if RF_ACC_TRACE > 0
    709 	RF_ETIMER_STOP(timer);
    710 	RF_ETIMER_EVAL(timer);
    711 	tracerec->specific.user.cleanup_us = RF_ETIMER_VAL_US(timer);
    712 
    713 	RF_ETIMER_START(timer);
    714 #endif
    715 	for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) {
    716 		if (!rf_suppressLocksAndLargeWrites &&
    717 		    asm_p->parityInfo &&
    718 		    !(desc->flags & RF_DAG_SUPPRESS_LOCKS)) {
    719 			RF_ASSERT_VALID_LOCKREQ(&asm_p->lockReqDesc);
    720 			rf_ReleaseStripeLock(raidPtr->lockTable,
    721 					     asm_p->stripeID,
    722 					     &asm_p->lockReqDesc);
    723 		}
    724 		if (asm_p->flags & RF_ASM_FLAGS_RECON_BLOCKED) {
    725 			rf_UnblockRecon(raidPtr, asm_p);
    726 		}
    727 	}
    728 #if RF_ACC_TRACE > 0
    729 	RF_ETIMER_STOP(timer);
    730 	RF_ETIMER_EVAL(timer);
    731 	tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer);
    732 
    733 	RF_ETIMER_START(timer);
    734 #endif
    735 	rf_FreeAccessStripeMap(asmh);
    736 #if RF_ACC_TRACE > 0
    737 	RF_ETIMER_STOP(timer);
    738 	RF_ETIMER_EVAL(timer);
    739 	tracerec->specific.user.cleanup_us += RF_ETIMER_VAL_US(timer);
    740 
    741 	RF_ETIMER_STOP(desc->timer);
    742 	RF_ETIMER_EVAL(desc->timer);
    743 
    744 	timer = desc->tracerec.tot_timer;
    745 	RF_ETIMER_STOP(timer);
    746 	RF_ETIMER_EVAL(timer);
    747 	desc->tracerec.total_us = RF_ETIMER_VAL_US(timer);
    748 
    749 	rf_LogTraceRec(raidPtr, tracerec);
    750 #endif
    751 	desc->flags |= RF_DAG_ACCESS_COMPLETE;
    752 
    753 	return RF_FALSE;
    754 }
    755