Home | History | Annotate | Line # | Download | only in raidframe
rf_states.c revision 1.49.32.1
      1 /*	$NetBSD: rf_states.c,v 1.49.32.1 2016/03/19 11:30:19 skrll 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 1.49.32.1 2016/03/19 11:30:19 skrll 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 #include "rf_paritymap.h"
     49 
     50 #ifndef RF_DEBUG_STATES
     51 #define RF_DEBUG_STATES 0
     52 #endif
     53 
     54 /* prototypes for some of the available states.
     55 
     56    States must:
     57 
     58      - not block.
     59 
     60      - either schedule rf_ContinueRaidAccess as a callback and return
     61        RF_TRUE, or complete all of their work and return RF_FALSE.
     62 
     63      - increment desc->state when they have finished their work.
     64 */
     65 
     66 #if RF_DEBUG_STATES
     67 static char *
     68 StateName(RF_AccessState_t state)
     69 {
     70 	switch (state) {
     71 		case rf_QuiesceState:return "QuiesceState";
     72 	case rf_MapState:
     73 		return "MapState";
     74 	case rf_LockState:
     75 		return "LockState";
     76 	case rf_CreateDAGState:
     77 		return "CreateDAGState";
     78 	case rf_ExecuteDAGState:
     79 		return "ExecuteDAGState";
     80 	case rf_ProcessDAGState:
     81 		return "ProcessDAGState";
     82 	case rf_CleanupState:
     83 		return "CleanupState";
     84 	case rf_LastState:
     85 		return "LastState";
     86 	case rf_IncrAccessesCountState:
     87 		return "IncrAccessesCountState";
     88 	case rf_DecrAccessesCountState:
     89 		return "DecrAccessesCountState";
     90 	default:
     91 		return "!!! UnnamedState !!!";
     92 	}
     93 }
     94 #endif
     95 
     96 void
     97 rf_ContinueRaidAccess(RF_RaidAccessDesc_t *desc)
     98 {
     99 	int     suspended = RF_FALSE;
    100 	int     current_state_index = desc->state;
    101 	RF_AccessState_t current_state = desc->states[current_state_index];
    102 #if RF_DEBUG_STATES
    103 	int     unit = desc->raidPtr->raidid;
    104 #endif
    105 
    106 	do {
    107 
    108 		current_state_index = desc->state;
    109 		current_state = desc->states[current_state_index];
    110 
    111 		switch (current_state) {
    112 
    113 		case rf_QuiesceState:
    114 			suspended = rf_State_Quiesce(desc);
    115 			break;
    116 		case rf_IncrAccessesCountState:
    117 			suspended = rf_State_IncrAccessCount(desc);
    118 			break;
    119 		case rf_MapState:
    120 			suspended = rf_State_Map(desc);
    121 			break;
    122 		case rf_LockState:
    123 			suspended = rf_State_Lock(desc);
    124 			break;
    125 		case rf_CreateDAGState:
    126 			suspended = rf_State_CreateDAG(desc);
    127 			break;
    128 		case rf_ExecuteDAGState:
    129 			suspended = rf_State_ExecuteDAG(desc);
    130 			break;
    131 		case rf_ProcessDAGState:
    132 			suspended = rf_State_ProcessDAG(desc);
    133 			break;
    134 		case rf_CleanupState:
    135 			suspended = rf_State_Cleanup(desc);
    136 			break;
    137 		case rf_DecrAccessesCountState:
    138 			suspended = rf_State_DecrAccessCount(desc);
    139 			break;
    140 		case rf_LastState:
    141 			suspended = rf_State_LastState(desc);
    142 			break;
    143 		}
    144 
    145 		/* after this point, we cannot dereference desc since
    146 		 * desc may have been freed. desc is only freed in
    147 		 * LastState, so if we renter this function or loop
    148 		 * back up, desc should be valid. */
    149 
    150 #if RF_DEBUG_STATES
    151 		if (rf_printStatesDebug) {
    152 			printf("raid%d: State: %-24s StateIndex: %3i desc: 0x%ld %s\n",
    153 			       unit, StateName(current_state),
    154 			       current_state_index, (long) desc,
    155 			       suspended ? "callback scheduled" : "looping");
    156 		}
    157 #endif
    158 	} while (!suspended && current_state != rf_LastState);
    159 
    160 	return;
    161 }
    162 
    163 
    164 void
    165 rf_ContinueDagAccess(RF_DagList_t *dagList)
    166 {
    167 #if RF_ACC_TRACE > 0
    168 	RF_AccTraceEntry_t *tracerec = &(dagList->desc->tracerec);
    169 	RF_Etimer_t timer;
    170 #endif
    171 	RF_RaidAccessDesc_t *desc;
    172 	RF_DagHeader_t *dag_h;
    173 	int     i;
    174 
    175 	desc = dagList->desc;
    176 
    177 #if RF_ACC_TRACE > 0
    178 	timer = tracerec->timer;
    179 	RF_ETIMER_STOP(timer);
    180 	RF_ETIMER_EVAL(timer);
    181 	tracerec->specific.user.exec_us = RF_ETIMER_VAL_US(timer);
    182 	RF_ETIMER_START(tracerec->timer);
    183 #endif
    184 
    185 	/* skip to dag which just finished */
    186 	dag_h = dagList->dags;
    187 	for (i = 0; i < dagList->numDagsDone; i++) {
    188 		dag_h = dag_h->next;
    189 	}
    190 
    191 	/* check to see if retry is required */
    192 	if (dag_h->status == rf_rollBackward) {
    193 		/* when a dag fails, mark desc status as bad and allow
    194 		 * all other dags in the desc to execute to
    195 		 * completion.  then, free all dags and start over */
    196 		desc->status = 1;	/* bad status */
    197 #if 0
    198 		printf("raid%d: DAG failure: %c addr 0x%lx "
    199 		       "(%ld) nblk 0x%x (%d) buf 0x%lx state %d\n",
    200 		       desc->raidPtr->raidid, desc->type,
    201 		       (long) desc->raidAddress,
    202 		       (long) desc->raidAddress, (int) desc->numBlocks,
    203 		       (int) desc->numBlocks,
    204 		       (unsigned long) (desc->bufPtr), desc->state);
    205 #endif
    206 	}
    207 	dagList->numDagsDone++;
    208 	rf_ContinueRaidAccess(desc);
    209 }
    210 
    211 int
    212 rf_State_LastState(RF_RaidAccessDesc_t *desc)
    213 {
    214 	void    (*callbackFunc) (RF_CBParam_t) = desc->callbackFunc;
    215 	RF_CBParam_t callbackArg;
    216 
    217 	callbackArg.p = desc->callbackArg;
    218 
    219 	/*
    220 	 * We don't support non-async IO.
    221 	 */
    222 	KASSERT(desc->async_flag);
    223 
    224 	/*
    225 	 * The parity_map hook has to go here, because the iodone
    226 	 * callback goes straight into the kintf layer.
    227 	 */
    228 	if (desc->raidPtr->parity_map != NULL &&
    229 	    desc->type == RF_IO_TYPE_WRITE)
    230 		rf_paritymap_end(desc->raidPtr->parity_map,
    231 		    desc->raidAddress, desc->numBlocks);
    232 
    233 	/* printf("Calling raiddone on 0x%x\n",desc->bp); */
    234 	raiddone(desc->raidPtr, desc->bp); /* access came through ioctl */
    235 
    236 	if (callbackFunc)
    237 		callbackFunc(callbackArg);
    238 	rf_FreeRaidAccDesc(desc);
    239 
    240 	return RF_FALSE;
    241 }
    242 
    243 int
    244 rf_State_IncrAccessCount(RF_RaidAccessDesc_t *desc)
    245 {
    246 	RF_Raid_t *raidPtr;
    247 
    248 	raidPtr = desc->raidPtr;
    249 	/* Bummer. We have to do this to be 100% safe w.r.t. the increment
    250 	 * below */
    251 	rf_lock_mutex2(raidPtr->access_suspend_mutex);
    252 	raidPtr->accs_in_flight++;	/* used to detect quiescence */
    253 	rf_unlock_mutex2(raidPtr->access_suspend_mutex);
    254 
    255 	desc->state++;
    256 	return RF_FALSE;
    257 }
    258 
    259 int
    260 rf_State_DecrAccessCount(RF_RaidAccessDesc_t *desc)
    261 {
    262 	RF_Raid_t *raidPtr;
    263 
    264 	raidPtr = desc->raidPtr;
    265 
    266 	rf_lock_mutex2(raidPtr->access_suspend_mutex);
    267 	raidPtr->accs_in_flight--;
    268 	if (raidPtr->accesses_suspended && raidPtr->accs_in_flight == 0) {
    269 		rf_SignalQuiescenceLock(raidPtr);
    270 	}
    271 	rf_unlock_mutex2(raidPtr->access_suspend_mutex);
    272 
    273 	desc->state++;
    274 	return RF_FALSE;
    275 }
    276 
    277 int
    278 rf_State_Quiesce(RF_RaidAccessDesc_t *desc)
    279 {
    280 #if RF_ACC_TRACE > 0
    281 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    282 	RF_Etimer_t timer;
    283 #endif
    284 	RF_CallbackDesc_t *cb;
    285 	RF_Raid_t *raidPtr;
    286 	int     suspended = RF_FALSE;
    287 	int need_cb, used_cb;
    288 
    289 	raidPtr = desc->raidPtr;
    290 
    291 #if RF_ACC_TRACE > 0
    292 	RF_ETIMER_START(timer);
    293 	RF_ETIMER_START(desc->timer);
    294 #endif
    295 
    296 	need_cb = 0;
    297 	used_cb = 0;
    298 	cb = NULL;
    299 
    300 	rf_lock_mutex2(raidPtr->access_suspend_mutex);
    301 	/* Do an initial check to see if we might need a callback structure */
    302 	if (raidPtr->accesses_suspended) {
    303 		need_cb = 1;
    304 	}
    305 	rf_unlock_mutex2(raidPtr->access_suspend_mutex);
    306 
    307 	if (need_cb) {
    308 		/* create a callback if we might need it...
    309 		   and we likely do. */
    310 		cb = rf_AllocCallbackDesc();
    311 	}
    312 
    313 	rf_lock_mutex2(raidPtr->access_suspend_mutex);
    314 	if (raidPtr->accesses_suspended) {
    315 		cb->callbackFunc = (void (*) (RF_CBParam_t)) rf_ContinueRaidAccess;
    316 		cb->callbackArg.p = (void *) desc;
    317 		cb->next = raidPtr->quiesce_wait_list;
    318 		raidPtr->quiesce_wait_list = cb;
    319 		suspended = RF_TRUE;
    320 		used_cb = 1;
    321 	}
    322 	rf_unlock_mutex2(raidPtr->access_suspend_mutex);
    323 
    324 	if ((need_cb == 1) && (used_cb == 0)) {
    325 		rf_FreeCallbackDesc(cb);
    326 	}
    327 
    328 #if RF_ACC_TRACE > 0
    329 	RF_ETIMER_STOP(timer);
    330 	RF_ETIMER_EVAL(timer);
    331 	tracerec->specific.user.suspend_ovhd_us += RF_ETIMER_VAL_US(timer);
    332 #endif
    333 
    334 #if RF_DEBUG_QUIESCE
    335 	if (suspended && rf_quiesceDebug)
    336 		printf("Stalling access due to quiescence lock\n");
    337 #endif
    338 	desc->state++;
    339 	return suspended;
    340 }
    341 
    342 int
    343 rf_State_Map(RF_RaidAccessDesc_t *desc)
    344 {
    345 	RF_Raid_t *raidPtr = desc->raidPtr;
    346 #if RF_ACC_TRACE > 0
    347 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    348 	RF_Etimer_t timer;
    349 
    350 	RF_ETIMER_START(timer);
    351 #endif
    352 
    353 	if (!(desc->asmap = rf_MapAccess(raidPtr, desc->raidAddress, desc->numBlocks,
    354 		    desc->bufPtr, RF_DONT_REMAP)))
    355 		RF_PANIC();
    356 
    357 #if RF_ACC_TRACE > 0
    358 	RF_ETIMER_STOP(timer);
    359 	RF_ETIMER_EVAL(timer);
    360 	tracerec->specific.user.map_us = RF_ETIMER_VAL_US(timer);
    361 #endif
    362 
    363 	desc->state++;
    364 	return RF_FALSE;
    365 }
    366 
    367 int
    368 rf_State_Lock(RF_RaidAccessDesc_t *desc)
    369 {
    370 #if RF_ACC_TRACE > 0
    371 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    372 	RF_Etimer_t timer;
    373 #endif
    374 	RF_Raid_t *raidPtr = desc->raidPtr;
    375 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    376 	RF_AccessStripeMap_t *asm_p;
    377 	RF_StripeNum_t lastStripeID = -1;
    378 	int     suspended = RF_FALSE;
    379 
    380 #if RF_ACC_TRACE > 0
    381 	RF_ETIMER_START(timer);
    382 #endif
    383 
    384 	/* acquire each lock that we don't already hold */
    385 	for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) {
    386 		RF_ASSERT(RF_IO_IS_R_OR_W(desc->type));
    387 		if (!rf_suppressLocksAndLargeWrites &&
    388 		    asm_p->parityInfo &&
    389 		    !(desc->flags & RF_DAG_SUPPRESS_LOCKS) &&
    390 		    !(asm_p->flags & RF_ASM_FLAGS_LOCK_TRIED)) {
    391 			asm_p->flags |= RF_ASM_FLAGS_LOCK_TRIED;
    392 				/* locks must be acquired hierarchically */
    393 			RF_ASSERT(asm_p->stripeID > lastStripeID);
    394 			lastStripeID = asm_p->stripeID;
    395 
    396 			RF_INIT_LOCK_REQ_DESC(asm_p->lockReqDesc, desc->type,
    397 					      (void (*) (struct buf *)) rf_ContinueRaidAccess, desc, asm_p,
    398 					      raidPtr->Layout.dataSectorsPerStripe);
    399 			if (rf_AcquireStripeLock(raidPtr->lockTable, asm_p->stripeID,
    400 						 &asm_p->lockReqDesc)) {
    401 				suspended = RF_TRUE;
    402 				break;
    403 			}
    404 		}
    405 		if (desc->type == RF_IO_TYPE_WRITE &&
    406 		    raidPtr->status == rf_rs_reconstructing) {
    407 			if (!(asm_p->flags & RF_ASM_FLAGS_FORCE_TRIED)) {
    408 				int     val;
    409 
    410 				asm_p->flags |= RF_ASM_FLAGS_FORCE_TRIED;
    411 				val = rf_ForceOrBlockRecon(raidPtr, asm_p,
    412 							   (void (*) (RF_Raid_t *, void *)) rf_ContinueRaidAccess, desc);
    413 				if (val == 0) {
    414 					asm_p->flags |= RF_ASM_FLAGS_RECON_BLOCKED;
    415 				} else {
    416 					suspended = RF_TRUE;
    417 					break;
    418 				}
    419 			} else {
    420 #if RF_DEBUG_PSS > 0
    421 				if (rf_pssDebug) {
    422 					printf("raid%d: skipping force/block because already done, psid %ld\n",
    423 					       desc->raidPtr->raidid,
    424 					       (long) asm_p->stripeID);
    425 				}
    426 #endif
    427 			}
    428 		} else {
    429 #if RF_DEBUG_PSS > 0
    430 			if (rf_pssDebug) {
    431 				printf("raid%d: skipping force/block because not write or not under recon, psid %ld\n",
    432 				       desc->raidPtr->raidid,
    433 				       (long) asm_p->stripeID);
    434 			}
    435 #endif
    436 		}
    437 	}
    438 #if RF_ACC_TRACE > 0
    439 	RF_ETIMER_STOP(timer);
    440 	RF_ETIMER_EVAL(timer);
    441 	tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer);
    442 #endif
    443 	if (suspended)
    444 		return (RF_TRUE);
    445 
    446 	desc->state++;
    447 	return (RF_FALSE);
    448 }
    449 /*
    450  * the following three states create, execute, and post-process dags
    451  * the error recovery unit is a single dag.
    452  * by default, SelectAlgorithm creates an array of dags, one per parity stripe
    453  * in some tricky cases, multiple dags per stripe are created
    454  *   - dags within a parity stripe are executed sequentially (arbitrary order)
    455  *   - dags for distinct parity stripes are executed concurrently
    456  *
    457  * repeat until all dags complete successfully -or- dag selection fails
    458  *
    459  * while !done
    460  *   create dag(s) (SelectAlgorithm)
    461  *   if dag
    462  *     execute dag (DispatchDAG)
    463  *     if dag successful
    464  *       done (SUCCESS)
    465  *     else
    466  *       !done (RETRY - start over with new dags)
    467  *   else
    468  *     done (FAIL)
    469  */
    470 int
    471 rf_State_CreateDAG(RF_RaidAccessDesc_t *desc)
    472 {
    473 #if RF_ACC_TRACE > 0
    474 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    475 	RF_Etimer_t timer;
    476 #endif
    477 	RF_DagHeader_t *dag_h;
    478 	RF_DagList_t *dagList;
    479 	struct buf *bp;
    480 	int     i, selectStatus;
    481 
    482 	/* generate a dag for the access, and fire it off.  When the dag
    483 	 * completes, we'll get re-invoked in the next state. */
    484 #if RF_ACC_TRACE > 0
    485 	RF_ETIMER_START(timer);
    486 #endif
    487 	/* SelectAlgorithm returns one or more dags */
    488 	selectStatus = rf_SelectAlgorithm(desc, desc->flags | RF_DAG_SUPPRESS_LOCKS);
    489 #if RF_DEBUG_VALIDATE_DAG
    490 	if (rf_printDAGsDebug) {
    491 		dagList = desc->dagList;
    492 		for (i = 0; i < desc->numStripes; i++) {
    493 			rf_PrintDAGList(dagList->dags);
    494 			dagList = dagList->next;
    495 		}
    496 	}
    497 #endif /* RF_DEBUG_VALIDATE_DAG */
    498 #if RF_ACC_TRACE > 0
    499 	RF_ETIMER_STOP(timer);
    500 	RF_ETIMER_EVAL(timer);
    501 	/* update time to create all dags */
    502 	tracerec->specific.user.dag_create_us = RF_ETIMER_VAL_US(timer);
    503 #endif
    504 
    505 	desc->status = 0;	/* good status */
    506 
    507 	if (selectStatus || (desc->numRetries > RF_RETRY_THRESHOLD)) {
    508 		/* failed to create a dag */
    509 		/* this happens when there are too many faults or incomplete
    510 		 * dag libraries */
    511 		if (selectStatus) {
    512 			printf("raid%d: failed to create a dag. "
    513 			       "Too many component failures.\n",
    514 			       desc->raidPtr->raidid);
    515 		} else {
    516 			printf("raid%d: IO failed after %d retries.\n",
    517 			       desc->raidPtr->raidid, RF_RETRY_THRESHOLD);
    518 		}
    519 
    520 		desc->status = 1; /* bad status */
    521 		/* skip straight to rf_State_Cleanup() */
    522 		desc->state = rf_CleanupState;
    523 		bp = (struct buf *)desc->bp;
    524 		bp->b_error = EIO;
    525 		bp->b_resid = bp->b_bcount;
    526 	} else {
    527 		/* bind dags to desc */
    528 		dagList = desc->dagList;
    529 		for (i = 0; i < desc->numStripes; i++) {
    530 			dag_h = dagList->dags;
    531 			while (dag_h) {
    532 				dag_h->bp = (struct buf *) desc->bp;
    533 #if RF_ACC_TRACE > 0
    534 				dag_h->tracerec = tracerec;
    535 #endif
    536 				dag_h = dag_h->next;
    537 			}
    538 			dagList = dagList->next;
    539 		}
    540 		desc->flags |= RF_DAG_DISPATCH_RETURNED;
    541 		desc->state++;	/* next state should be rf_State_ExecuteDAG */
    542 	}
    543 	return RF_FALSE;
    544 }
    545 
    546 
    547 
    548 /* the access has an list of dagLists, one dagList per parity stripe.
    549  * fire the first dag in each parity stripe (dagList).
    550  * dags within a stripe (dagList) must be executed sequentially
    551  *  - this preserves atomic parity update
    552  * dags for independents parity groups (stripes) are fired concurrently */
    553 
    554 int
    555 rf_State_ExecuteDAG(RF_RaidAccessDesc_t *desc)
    556 {
    557 	int     i;
    558 	RF_DagHeader_t *dag_h;
    559 	RF_DagList_t *dagList;
    560 
    561 	/* next state is always rf_State_ProcessDAG important to do
    562 	 * this before firing the first dag (it may finish before we
    563 	 * leave this routine) */
    564 	desc->state++;
    565 
    566 	/* sweep dag array, a stripe at a time, firing the first dag
    567 	 * in each stripe */
    568 	dagList = desc->dagList;
    569 	for (i = 0; i < desc->numStripes; i++) {
    570 		RF_ASSERT(dagList->numDags > 0);
    571 		RF_ASSERT(dagList->numDagsDone == 0);
    572 		RF_ASSERT(dagList->numDagsFired == 0);
    573 #if RF_ACC_TRACE > 0
    574 		RF_ETIMER_START(dagList->tracerec.timer);
    575 #endif
    576 		/* fire first dag in this stripe */
    577 		dag_h = dagList->dags;
    578 		RF_ASSERT(dag_h);
    579 		dagList->numDagsFired++;
    580 		rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess, dagList);
    581 		dagList = dagList->next;
    582 	}
    583 
    584 	/* the DAG will always call the callback, even if there was no
    585 	 * blocking, so we are always suspended in this state */
    586 	return RF_TRUE;
    587 }
    588 
    589 
    590 
    591 /* rf_State_ProcessDAG is entered when a dag completes.
    592  * first, check to all dags in the access have completed
    593  * if not, fire as many dags as possible */
    594 
    595 int
    596 rf_State_ProcessDAG(RF_RaidAccessDesc_t *desc)
    597 {
    598 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    599 	RF_Raid_t *raidPtr = desc->raidPtr;
    600 	RF_DagHeader_t *dag_h;
    601 	int     i, j, done = RF_TRUE;
    602 	RF_DagList_t *dagList, *temp;
    603 
    604 	/* check to see if this is the last dag */
    605 	dagList = desc->dagList;
    606 	for (i = 0; i < desc->numStripes; i++) {
    607 		if (dagList->numDags != dagList->numDagsDone)
    608 			done = RF_FALSE;
    609 		dagList = dagList->next;
    610 	}
    611 
    612 	if (done) {
    613 		if (desc->status) {
    614 			/* a dag failed, retry */
    615 			/* free all dags */
    616 			dagList = desc->dagList;
    617 			for (i = 0; i < desc->numStripes; i++) {
    618 				rf_FreeDAG(dagList->dags);
    619 				temp = dagList;
    620 				dagList = dagList->next;
    621 				rf_FreeDAGList(temp);
    622 			}
    623 			desc->dagList = NULL;
    624 
    625 			rf_MarkFailuresInASMList(raidPtr, asmh);
    626 
    627 			/* note the retry so that we'll bail in
    628 			   rf_State_CreateDAG() once we've retired
    629 			   the IO RF_RETRY_THRESHOLD times */
    630 
    631 			desc->numRetries++;
    632 
    633 			/* back up to rf_State_CreateDAG */
    634 			desc->state = desc->state - 2;
    635 			return RF_FALSE;
    636 		} else {
    637 			/* move on to rf_State_Cleanup */
    638 			desc->state++;
    639 		}
    640 		return RF_FALSE;
    641 	} else {
    642 		/* more dags to execute */
    643 		/* see if any are ready to be fired.  if so, fire them */
    644 		/* don't fire the initial dag in a list, it's fired in
    645 		 * rf_State_ExecuteDAG */
    646 		dagList = desc->dagList;
    647 		for (i = 0; i < desc->numStripes; i++) {
    648 			if ((dagList->numDagsDone < dagList->numDags)
    649 			    && (dagList->numDagsDone == dagList->numDagsFired)
    650 			    && (dagList->numDagsFired > 0)) {
    651 #if RF_ACC_TRACE > 0
    652 				RF_ETIMER_START(dagList->tracerec.timer);
    653 #endif
    654 				/* fire next dag in this stripe */
    655 				/* first, skip to next dag awaiting execution */
    656 				dag_h = dagList->dags;
    657 				for (j = 0; j < dagList->numDagsDone; j++)
    658 					dag_h = dag_h->next;
    659 				dagList->numDagsFired++;
    660 				rf_DispatchDAG(dag_h, (void (*) (void *)) rf_ContinueDagAccess,
    661 				    dagList);
    662 			}
    663 			dagList = dagList->next;
    664 		}
    665 		return RF_TRUE;
    666 	}
    667 }
    668 /* only make it this far if all dags complete successfully */
    669 int
    670 rf_State_Cleanup(RF_RaidAccessDesc_t *desc)
    671 {
    672 #if RF_ACC_TRACE > 0
    673 	RF_AccTraceEntry_t *tracerec = &desc->tracerec;
    674 	RF_Etimer_t timer;
    675 #endif
    676 	RF_AccessStripeMapHeader_t *asmh = desc->asmap;
    677 	RF_Raid_t *raidPtr = desc->raidPtr;
    678 	RF_AccessStripeMap_t *asm_p;
    679 	RF_DagList_t *dagList;
    680 	int i;
    681 
    682 	desc->state++;
    683 
    684 #if RF_ACC_TRACE > 0
    685 	timer = tracerec->timer;
    686 	RF_ETIMER_STOP(timer);
    687 	RF_ETIMER_EVAL(timer);
    688 	tracerec->specific.user.dag_retry_us = RF_ETIMER_VAL_US(timer);
    689 
    690 	/* the RAID I/O is complete.  Clean up. */
    691 	tracerec->specific.user.dag_retry_us = 0;
    692 
    693 	RF_ETIMER_START(timer);
    694 #endif
    695 	/* free all dags */
    696 	dagList = desc->dagList;
    697 	for (i = 0; i < desc->numStripes; i++) {
    698 		rf_FreeDAG(dagList->dags);
    699 		dagList = dagList->next;
    700 	}
    701 #if RF_ACC_TRACE > 0
    702 	RF_ETIMER_STOP(timer);
    703 	RF_ETIMER_EVAL(timer);
    704 	tracerec->specific.user.cleanup_us = RF_ETIMER_VAL_US(timer);
    705 
    706 	RF_ETIMER_START(timer);
    707 #endif
    708 	for (asm_p = asmh->stripeMap; asm_p; asm_p = asm_p->next) {
    709 		if (!rf_suppressLocksAndLargeWrites &&
    710 		    asm_p->parityInfo &&
    711 		    !(desc->flags & RF_DAG_SUPPRESS_LOCKS)) {
    712 			RF_ASSERT_VALID_LOCKREQ(&asm_p->lockReqDesc);
    713 			rf_ReleaseStripeLock(raidPtr->lockTable,
    714 					     asm_p->stripeID,
    715 					     &asm_p->lockReqDesc);
    716 		}
    717 		if (asm_p->flags & RF_ASM_FLAGS_RECON_BLOCKED) {
    718 			rf_UnblockRecon(raidPtr, asm_p);
    719 		}
    720 	}
    721 #if RF_ACC_TRACE > 0
    722 	RF_ETIMER_STOP(timer);
    723 	RF_ETIMER_EVAL(timer);
    724 	tracerec->specific.user.lock_us += RF_ETIMER_VAL_US(timer);
    725 
    726 	RF_ETIMER_START(timer);
    727 #endif
    728 	rf_FreeAccessStripeMap(asmh);
    729 #if RF_ACC_TRACE > 0
    730 	RF_ETIMER_STOP(timer);
    731 	RF_ETIMER_EVAL(timer);
    732 	tracerec->specific.user.cleanup_us += RF_ETIMER_VAL_US(timer);
    733 
    734 	RF_ETIMER_STOP(desc->timer);
    735 	RF_ETIMER_EVAL(desc->timer);
    736 
    737 	timer = desc->tracerec.tot_timer;
    738 	RF_ETIMER_STOP(timer);
    739 	RF_ETIMER_EVAL(timer);
    740 	desc->tracerec.total_us = RF_ETIMER_VAL_US(timer);
    741 
    742 	rf_LogTraceRec(raidPtr, tracerec);
    743 #endif
    744 	desc->flags |= RF_DAG_ACCESS_COMPLETE;
    745 
    746 	return RF_FALSE;
    747 }
    748