Home | History | Annotate | Line # | Download | only in raidframe
rf_dagutils.c revision 1.6.6.9
      1  1.6.6.3  nathanw /*	$NetBSD: rf_dagutils.c,v 1.6.6.9 2002/12/11 06:38:32 thorpej Exp $	*/
      2      1.1    oster /*
      3      1.1    oster  * Copyright (c) 1995 Carnegie-Mellon University.
      4      1.1    oster  * All rights reserved.
      5      1.1    oster  *
      6      1.1    oster  * Authors: Mark Holland, William V. Courtright II, Jim Zelenka
      7      1.1    oster  *
      8      1.1    oster  * Permission to use, copy, modify and distribute this software and
      9      1.1    oster  * its documentation is hereby granted, provided that both the copyright
     10      1.1    oster  * notice and this permission notice appear in all copies of the
     11      1.1    oster  * software, derivative works or modified versions, and any portions
     12      1.1    oster  * thereof, and that both notices appear in supporting documentation.
     13      1.1    oster  *
     14      1.1    oster  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
     15      1.1    oster  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
     16      1.1    oster  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
     17      1.1    oster  *
     18      1.1    oster  * Carnegie Mellon requests users of this software to return to
     19      1.1    oster  *
     20      1.1    oster  *  Software Distribution Coordinator  or  Software.Distribution (at) CS.CMU.EDU
     21      1.1    oster  *  School of Computer Science
     22      1.1    oster  *  Carnegie Mellon University
     23      1.1    oster  *  Pittsburgh PA 15213-3890
     24      1.1    oster  *
     25      1.1    oster  * any improvements or extensions that they make and grant Carnegie the
     26      1.1    oster  * rights to redistribute these changes.
     27      1.1    oster  */
     28      1.1    oster 
     29      1.1    oster /******************************************************************************
     30      1.1    oster  *
     31      1.1    oster  * rf_dagutils.c -- utility routines for manipulating dags
     32      1.1    oster  *
     33      1.1    oster  *****************************************************************************/
     34  1.6.6.3  nathanw 
     35  1.6.6.3  nathanw #include <sys/cdefs.h>
     36  1.6.6.3  nathanw __KERNEL_RCSID(0, "$NetBSD: rf_dagutils.c,v 1.6.6.9 2002/12/11 06:38:32 thorpej Exp $");
     37      1.1    oster 
     38  1.6.6.2  nathanw #include <dev/raidframe/raidframevar.h>
     39  1.6.6.2  nathanw 
     40      1.1    oster #include "rf_archs.h"
     41      1.1    oster #include "rf_threadstuff.h"
     42      1.1    oster #include "rf_raid.h"
     43      1.1    oster #include "rf_dag.h"
     44      1.1    oster #include "rf_dagutils.h"
     45      1.1    oster #include "rf_dagfuncs.h"
     46      1.1    oster #include "rf_general.h"
     47      1.1    oster #include "rf_freelist.h"
     48      1.1    oster #include "rf_map.h"
     49      1.1    oster #include "rf_shutdown.h"
     50      1.1    oster 
     51      1.1    oster #define SNUM_DIFF(_a_,_b_) (((_a_)>(_b_))?((_a_)-(_b_)):((_b_)-(_a_)))
     52      1.1    oster 
     53      1.1    oster RF_RedFuncs_t rf_xorFuncs = {
     54      1.1    oster 	rf_RegularXorFunc, "Reg Xr",
     55      1.3    oster rf_SimpleXorFunc, "Simple Xr"};
     56      1.1    oster 
     57      1.1    oster RF_RedFuncs_t rf_xorRecoveryFuncs = {
     58      1.1    oster 	rf_RecoveryXorFunc, "Recovery Xr",
     59      1.3    oster rf_RecoveryXorFunc, "Recovery Xr"};
     60      1.1    oster 
     61  1.6.6.5  nathanw #if RF_DEBUG_VALIDATE_DAG
     62      1.1    oster static void rf_RecurPrintDAG(RF_DagNode_t *, int, int);
     63      1.1    oster static void rf_PrintDAG(RF_DagHeader_t *);
     64  1.6.6.5  nathanw static int rf_ValidateBranch(RF_DagNode_t *, int *, int *,
     65  1.6.6.5  nathanw 			     RF_DagNode_t **, int);
     66      1.1    oster static void rf_ValidateBranchVisitedBits(RF_DagNode_t *, int, int);
     67      1.1    oster static void rf_ValidateVisitedBits(RF_DagHeader_t *);
     68  1.6.6.5  nathanw #endif /* RF_DEBUG_VALIDATE_DAG */
     69      1.1    oster 
     70      1.1    oster /******************************************************************************
     71      1.1    oster  *
     72      1.1    oster  * InitNode - initialize a dag node
     73      1.1    oster  *
     74      1.1    oster  * the size of the propList array is always the same as that of the
     75      1.1    oster  * successors array.
     76      1.1    oster  *
     77      1.1    oster  *****************************************************************************/
     78      1.3    oster void
     79      1.3    oster rf_InitNode(
     80      1.3    oster     RF_DagNode_t * node,
     81      1.3    oster     RF_NodeStatus_t initstatus,
     82      1.3    oster     int commit,
     83      1.3    oster     int (*doFunc) (RF_DagNode_t * node),
     84      1.3    oster     int (*undoFunc) (RF_DagNode_t * node),
     85      1.3    oster     int (*wakeFunc) (RF_DagNode_t * node, int status),
     86      1.3    oster     int nSucc,
     87      1.3    oster     int nAnte,
     88      1.3    oster     int nParam,
     89      1.3    oster     int nResult,
     90      1.3    oster     RF_DagHeader_t * hdr,
     91      1.3    oster     char *name,
     92      1.3    oster     RF_AllocListElem_t * alist)
     93      1.3    oster {
     94      1.3    oster 	void  **ptrs;
     95      1.3    oster 	int     nptrs;
     96      1.3    oster 
     97      1.3    oster 	if (nAnte > RF_MAX_ANTECEDENTS)
     98      1.3    oster 		RF_PANIC();
     99      1.3    oster 	node->status = initstatus;
    100      1.3    oster 	node->commitNode = commit;
    101      1.3    oster 	node->doFunc = doFunc;
    102      1.3    oster 	node->undoFunc = undoFunc;
    103      1.3    oster 	node->wakeFunc = wakeFunc;
    104      1.3    oster 	node->numParams = nParam;
    105      1.3    oster 	node->numResults = nResult;
    106      1.3    oster 	node->numAntecedents = nAnte;
    107      1.3    oster 	node->numAntDone = 0;
    108      1.3    oster 	node->next = NULL;
    109      1.3    oster 	node->numSuccedents = nSucc;
    110      1.3    oster 	node->name = name;
    111      1.3    oster 	node->dagHdr = hdr;
    112      1.3    oster 	node->visited = 0;
    113      1.3    oster 
    114      1.3    oster 	/* allocate all the pointers with one call to malloc */
    115      1.3    oster 	nptrs = nSucc + nAnte + nResult + nSucc;
    116      1.3    oster 
    117      1.3    oster 	if (nptrs <= RF_DAG_PTRCACHESIZE) {
    118      1.3    oster 		/*
    119      1.3    oster 	         * The dag_ptrs field of the node is basically some scribble
    120      1.3    oster 	         * space to be used here. We could get rid of it, and always
    121      1.3    oster 	         * allocate the range of pointers, but that's expensive. So,
    122      1.3    oster 	         * we pick a "common case" size for the pointer cache. Hopefully,
    123      1.3    oster 	         * we'll find that:
    124      1.3    oster 	         * (1) Generally, nptrs doesn't exceed RF_DAG_PTRCACHESIZE by
    125      1.3    oster 	         *     only a little bit (least efficient case)
    126      1.3    oster 	         * (2) Generally, ntprs isn't a lot less than RF_DAG_PTRCACHESIZE
    127      1.3    oster 	         *     (wasted memory)
    128      1.3    oster 	         */
    129      1.3    oster 		ptrs = (void **) node->dag_ptrs;
    130      1.3    oster 	} else {
    131      1.3    oster 		RF_CallocAndAdd(ptrs, nptrs, sizeof(void *), (void **), alist);
    132      1.3    oster 	}
    133      1.3    oster 	node->succedents = (nSucc) ? (RF_DagNode_t **) ptrs : NULL;
    134      1.3    oster 	node->antecedents = (nAnte) ? (RF_DagNode_t **) (ptrs + nSucc) : NULL;
    135      1.3    oster 	node->results = (nResult) ? (void **) (ptrs + nSucc + nAnte) : NULL;
    136      1.3    oster 	node->propList = (nSucc) ? (RF_PropHeader_t **) (ptrs + nSucc + nAnte + nResult) : NULL;
    137      1.3    oster 
    138      1.3    oster 	if (nParam) {
    139      1.3    oster 		if (nParam <= RF_DAG_PARAMCACHESIZE) {
    140      1.3    oster 			node->params = (RF_DagParam_t *) node->dag_params;
    141      1.3    oster 		} else {
    142      1.3    oster 			RF_CallocAndAdd(node->params, nParam, sizeof(RF_DagParam_t), (RF_DagParam_t *), alist);
    143      1.3    oster 		}
    144      1.3    oster 	} else {
    145      1.3    oster 		node->params = NULL;
    146      1.3    oster 	}
    147      1.1    oster }
    148      1.1    oster 
    149      1.1    oster 
    150      1.1    oster 
    151      1.1    oster /******************************************************************************
    152      1.1    oster  *
    153      1.1    oster  * allocation and deallocation routines
    154      1.1    oster  *
    155      1.1    oster  *****************************************************************************/
    156      1.1    oster 
    157      1.3    oster void
    158      1.3    oster rf_FreeDAG(dag_h)
    159      1.3    oster 	RF_DagHeader_t *dag_h;
    160      1.3    oster {
    161      1.3    oster 	RF_AccessStripeMapHeader_t *asmap, *t_asmap;
    162      1.3    oster 	RF_DagHeader_t *nextDag;
    163      1.3    oster 
    164      1.3    oster 	while (dag_h) {
    165      1.3    oster 		nextDag = dag_h->next;
    166      1.3    oster 		rf_FreeAllocList(dag_h->allocList);
    167      1.3    oster 		for (asmap = dag_h->asmList; asmap;) {
    168      1.3    oster 			t_asmap = asmap;
    169      1.3    oster 			asmap = asmap->next;
    170      1.3    oster 			rf_FreeAccessStripeMap(t_asmap);
    171      1.3    oster 		}
    172      1.3    oster 		rf_FreeDAGHeader(dag_h);
    173      1.3    oster 		dag_h = nextDag;
    174      1.3    oster 	}
    175      1.3    oster }
    176      1.3    oster 
    177      1.1    oster 
    178      1.1    oster static RF_FreeList_t *rf_dagh_freelist;
    179      1.1    oster 
    180      1.1    oster #define RF_MAX_FREE_DAGH 128
    181      1.1    oster #define RF_DAGH_INC       16
    182      1.1    oster #define RF_DAGH_INITIAL   32
    183      1.1    oster 
    184      1.1    oster static void rf_ShutdownDAGs(void *);
    185      1.3    oster static void
    186      1.3    oster rf_ShutdownDAGs(ignored)
    187      1.3    oster 	void   *ignored;
    188      1.1    oster {
    189      1.3    oster 	RF_FREELIST_DESTROY(rf_dagh_freelist, next, (RF_DagHeader_t *));
    190      1.1    oster }
    191      1.1    oster 
    192      1.3    oster int
    193      1.3    oster rf_ConfigureDAGs(listp)
    194      1.3    oster 	RF_ShutdownList_t **listp;
    195      1.1    oster {
    196      1.3    oster 	int     rc;
    197      1.1    oster 
    198      1.1    oster 	RF_FREELIST_CREATE(rf_dagh_freelist, RF_MAX_FREE_DAGH,
    199      1.3    oster 	    RF_DAGH_INC, sizeof(RF_DagHeader_t));
    200      1.1    oster 	if (rf_dagh_freelist == NULL)
    201      1.3    oster 		return (ENOMEM);
    202      1.1    oster 	rc = rf_ShutdownCreate(listp, rf_ShutdownDAGs, NULL);
    203      1.1    oster 	if (rc) {
    204  1.6.6.7  nathanw 		rf_print_unable_to_add_shutdown(__FILE__, __LINE__, rc);
    205      1.1    oster 		rf_ShutdownDAGs(NULL);
    206      1.3    oster 		return (rc);
    207      1.1    oster 	}
    208      1.3    oster 	RF_FREELIST_PRIME(rf_dagh_freelist, RF_DAGH_INITIAL, next,
    209      1.3    oster 	    (RF_DagHeader_t *));
    210      1.3    oster 	return (0);
    211      1.1    oster }
    212      1.1    oster 
    213      1.3    oster RF_DagHeader_t *
    214      1.3    oster rf_AllocDAGHeader()
    215      1.1    oster {
    216      1.1    oster 	RF_DagHeader_t *dh;
    217      1.1    oster 
    218      1.3    oster 	RF_FREELIST_GET(rf_dagh_freelist, dh, next, (RF_DagHeader_t *));
    219      1.1    oster 	if (dh) {
    220  1.6.6.1  nathanw 		memset((char *) dh, 0, sizeof(RF_DagHeader_t));
    221      1.1    oster 	}
    222      1.3    oster 	return (dh);
    223      1.1    oster }
    224      1.1    oster 
    225      1.3    oster void
    226      1.3    oster rf_FreeDAGHeader(RF_DagHeader_t * dh)
    227      1.1    oster {
    228      1.3    oster 	RF_FREELIST_FREE(rf_dagh_freelist, dh, next);
    229      1.1    oster }
    230      1.1    oster /* allocates a buffer big enough to hold the data described by pda */
    231      1.3    oster void   *
    232      1.3    oster rf_AllocBuffer(
    233      1.3    oster     RF_Raid_t * raidPtr,
    234      1.3    oster     RF_DagHeader_t * dag_h,
    235      1.3    oster     RF_PhysDiskAddr_t * pda,
    236      1.3    oster     RF_AllocListElem_t * allocList)
    237      1.3    oster {
    238      1.3    oster 	char   *p;
    239      1.3    oster 
    240      1.3    oster 	RF_MallocAndAdd(p, pda->numSector << raidPtr->logBytesPerSector,
    241      1.3    oster 	    (char *), allocList);
    242      1.3    oster 	return ((void *) p);
    243      1.1    oster }
    244  1.6.6.5  nathanw #if RF_DEBUG_VALIDATE_DAG
    245      1.1    oster /******************************************************************************
    246      1.1    oster  *
    247      1.1    oster  * debug routines
    248      1.1    oster  *
    249      1.1    oster  *****************************************************************************/
    250      1.1    oster 
    251      1.3    oster char   *
    252      1.3    oster rf_NodeStatusString(RF_DagNode_t * node)
    253      1.1    oster {
    254      1.3    oster 	switch (node->status) {
    255      1.3    oster 		case rf_wait:return ("wait");
    256      1.3    oster 	case rf_fired:
    257      1.3    oster 		return ("fired");
    258      1.3    oster 	case rf_good:
    259      1.3    oster 		return ("good");
    260      1.3    oster 	case rf_bad:
    261      1.3    oster 		return ("bad");
    262      1.3    oster 	default:
    263      1.3    oster 		return ("?");
    264      1.3    oster 	}
    265      1.3    oster }
    266      1.1    oster 
    267      1.3    oster void
    268      1.3    oster rf_PrintNodeInfoString(RF_DagNode_t * node)
    269      1.3    oster {
    270      1.3    oster 	RF_PhysDiskAddr_t *pda;
    271      1.3    oster 	int     (*df) (RF_DagNode_t *) = node->doFunc;
    272      1.3    oster 	int     i, lk, unlk;
    273      1.3    oster 	void   *bufPtr;
    274      1.3    oster 
    275      1.3    oster 	if ((df == rf_DiskReadFunc) || (df == rf_DiskWriteFunc)
    276      1.3    oster 	    || (df == rf_DiskReadMirrorIdleFunc)
    277      1.3    oster 	    || (df == rf_DiskReadMirrorPartitionFunc)) {
    278      1.3    oster 		pda = (RF_PhysDiskAddr_t *) node->params[0].p;
    279      1.3    oster 		bufPtr = (void *) node->params[1].p;
    280      1.3    oster 		lk = RF_EXTRACT_LOCK_FLAG(node->params[3].v);
    281      1.3    oster 		unlk = RF_EXTRACT_UNLOCK_FLAG(node->params[3].v);
    282      1.3    oster 		RF_ASSERT(!(lk && unlk));
    283      1.3    oster 		printf("r %d c %d offs %ld nsect %d buf 0x%lx %s\n", pda->row, pda->col,
    284      1.3    oster 		    (long) pda->startSector, (int) pda->numSector, (long) bufPtr,
    285      1.3    oster 		    (lk) ? "LOCK" : ((unlk) ? "UNLK" : " "));
    286      1.3    oster 		return;
    287      1.3    oster 	}
    288      1.3    oster 	if (df == rf_DiskUnlockFunc) {
    289      1.3    oster 		pda = (RF_PhysDiskAddr_t *) node->params[0].p;
    290      1.3    oster 		lk = RF_EXTRACT_LOCK_FLAG(node->params[3].v);
    291      1.3    oster 		unlk = RF_EXTRACT_UNLOCK_FLAG(node->params[3].v);
    292      1.3    oster 		RF_ASSERT(!(lk && unlk));
    293      1.3    oster 		printf("r %d c %d %s\n", pda->row, pda->col,
    294      1.3    oster 		    (lk) ? "LOCK" : ((unlk) ? "UNLK" : "nop"));
    295      1.3    oster 		return;
    296      1.3    oster 	}
    297      1.3    oster 	if ((df == rf_SimpleXorFunc) || (df == rf_RegularXorFunc)
    298      1.3    oster 	    || (df == rf_RecoveryXorFunc)) {
    299      1.3    oster 		printf("result buf 0x%lx\n", (long) node->results[0]);
    300      1.3    oster 		for (i = 0; i < node->numParams - 1; i += 2) {
    301      1.3    oster 			pda = (RF_PhysDiskAddr_t *) node->params[i].p;
    302      1.3    oster 			bufPtr = (RF_PhysDiskAddr_t *) node->params[i + 1].p;
    303      1.3    oster 			printf("    buf 0x%lx r%d c%d offs %ld nsect %d\n",
    304      1.3    oster 			    (long) bufPtr, pda->row, pda->col,
    305      1.3    oster 			    (long) pda->startSector, (int) pda->numSector);
    306      1.3    oster 		}
    307      1.3    oster 		return;
    308      1.3    oster 	}
    309      1.1    oster #if RF_INCLUDE_PARITYLOGGING > 0
    310      1.3    oster 	if (df == rf_ParityLogOverwriteFunc || df == rf_ParityLogUpdateFunc) {
    311      1.3    oster 		for (i = 0; i < node->numParams - 1; i += 2) {
    312      1.3    oster 			pda = (RF_PhysDiskAddr_t *) node->params[i].p;
    313      1.3    oster 			bufPtr = (RF_PhysDiskAddr_t *) node->params[i + 1].p;
    314      1.3    oster 			printf(" r%d c%d offs %ld nsect %d buf 0x%lx\n",
    315      1.3    oster 			    pda->row, pda->col, (long) pda->startSector,
    316      1.3    oster 			    (int) pda->numSector, (long) bufPtr);
    317      1.3    oster 		}
    318      1.3    oster 		return;
    319      1.3    oster 	}
    320      1.3    oster #endif				/* RF_INCLUDE_PARITYLOGGING > 0 */
    321      1.3    oster 
    322      1.3    oster 	if ((df == rf_TerminateFunc) || (df == rf_NullNodeFunc)) {
    323      1.3    oster 		printf("\n");
    324      1.3    oster 		return;
    325      1.3    oster 	}
    326      1.3    oster 	printf("?\n");
    327      1.3    oster }
    328  1.6.6.8  nathanw #ifdef DEBUG
    329      1.3    oster static void
    330      1.3    oster rf_RecurPrintDAG(node, depth, unvisited)
    331      1.3    oster 	RF_DagNode_t *node;
    332      1.3    oster 	int     depth;
    333      1.3    oster 	int     unvisited;
    334      1.3    oster {
    335      1.3    oster 	char   *anttype;
    336      1.3    oster 	int     i;
    337      1.3    oster 
    338      1.3    oster 	node->visited = (unvisited) ? 0 : 1;
    339      1.3    oster 	printf("(%d) %d C%d %s: %s,s%d %d/%d,a%d/%d,p%d,r%d S{", depth,
    340      1.3    oster 	    node->nodeNum, node->commitNode, node->name, rf_NodeStatusString(node),
    341      1.3    oster 	    node->numSuccedents, node->numSuccFired, node->numSuccDone,
    342      1.3    oster 	    node->numAntecedents, node->numAntDone, node->numParams, node->numResults);
    343      1.3    oster 	for (i = 0; i < node->numSuccedents; i++) {
    344      1.3    oster 		printf("%d%s", node->succedents[i]->nodeNum,
    345      1.3    oster 		    ((i == node->numSuccedents - 1) ? "\0" : " "));
    346      1.3    oster 	}
    347      1.3    oster 	printf("} A{");
    348      1.3    oster 	for (i = 0; i < node->numAntecedents; i++) {
    349      1.3    oster 		switch (node->antType[i]) {
    350      1.3    oster 		case rf_trueData:
    351      1.3    oster 			anttype = "T";
    352      1.3    oster 			break;
    353      1.3    oster 		case rf_antiData:
    354      1.3    oster 			anttype = "A";
    355      1.3    oster 			break;
    356      1.3    oster 		case rf_outputData:
    357      1.3    oster 			anttype = "O";
    358      1.3    oster 			break;
    359      1.3    oster 		case rf_control:
    360      1.3    oster 			anttype = "C";
    361      1.3    oster 			break;
    362      1.3    oster 		default:
    363      1.3    oster 			anttype = "?";
    364      1.3    oster 			break;
    365      1.3    oster 		}
    366      1.3    oster 		printf("%d(%s)%s", node->antecedents[i]->nodeNum, anttype, (i == node->numAntecedents - 1) ? "\0" : " ");
    367      1.3    oster 	}
    368      1.3    oster 	printf("}; ");
    369      1.3    oster 	rf_PrintNodeInfoString(node);
    370      1.3    oster 	for (i = 0; i < node->numSuccedents; i++) {
    371      1.3    oster 		if (node->succedents[i]->visited == unvisited)
    372      1.3    oster 			rf_RecurPrintDAG(node->succedents[i], depth + 1, unvisited);
    373      1.3    oster 	}
    374      1.1    oster }
    375      1.1    oster 
    376      1.3    oster static void
    377      1.3    oster rf_PrintDAG(dag_h)
    378      1.3    oster 	RF_DagHeader_t *dag_h;
    379      1.3    oster {
    380      1.3    oster 	int     unvisited, i;
    381      1.3    oster 	char   *status;
    382      1.3    oster 
    383      1.3    oster 	/* set dag status */
    384      1.3    oster 	switch (dag_h->status) {
    385      1.3    oster 	case rf_enable:
    386      1.3    oster 		status = "enable";
    387      1.3    oster 		break;
    388      1.3    oster 	case rf_rollForward:
    389      1.3    oster 		status = "rollForward";
    390      1.3    oster 		break;
    391      1.3    oster 	case rf_rollBackward:
    392      1.3    oster 		status = "rollBackward";
    393      1.3    oster 		break;
    394      1.3    oster 	default:
    395      1.3    oster 		status = "illegal!";
    396      1.3    oster 		break;
    397      1.3    oster 	}
    398      1.3    oster 	/* find out if visited bits are currently set or clear */
    399      1.3    oster 	unvisited = dag_h->succedents[0]->visited;
    400      1.3    oster 
    401      1.3    oster 	printf("DAG type:  %s\n", dag_h->creator);
    402      1.3    oster 	printf("format is (depth) num commit type: status,nSucc nSuccFired/nSuccDone,nAnte/nAnteDone,nParam,nResult S{x} A{x(type)};  info\n");
    403      1.3    oster 	printf("(0) %d Hdr: %s, s%d, (commit %d/%d) S{", dag_h->nodeNum,
    404      1.3    oster 	    status, dag_h->numSuccedents, dag_h->numCommitNodes, dag_h->numCommits);
    405      1.3    oster 	for (i = 0; i < dag_h->numSuccedents; i++) {
    406      1.3    oster 		printf("%d%s", dag_h->succedents[i]->nodeNum,
    407      1.3    oster 		    ((i == dag_h->numSuccedents - 1) ? "\0" : " "));
    408      1.3    oster 	}
    409      1.3    oster 	printf("};\n");
    410      1.3    oster 	for (i = 0; i < dag_h->numSuccedents; i++) {
    411      1.3    oster 		if (dag_h->succedents[i]->visited == unvisited)
    412      1.3    oster 			rf_RecurPrintDAG(dag_h->succedents[i], 1, unvisited);
    413      1.3    oster 	}
    414      1.3    oster }
    415  1.6.6.8  nathanw #endif
    416      1.1    oster /* assigns node numbers */
    417      1.3    oster int
    418      1.3    oster rf_AssignNodeNums(RF_DagHeader_t * dag_h)
    419      1.1    oster {
    420      1.3    oster 	int     unvisited, i, nnum;
    421      1.3    oster 	RF_DagNode_t *node;
    422      1.1    oster 
    423      1.3    oster 	nnum = 0;
    424      1.3    oster 	unvisited = dag_h->succedents[0]->visited;
    425      1.3    oster 
    426      1.3    oster 	dag_h->nodeNum = nnum++;
    427      1.3    oster 	for (i = 0; i < dag_h->numSuccedents; i++) {
    428      1.3    oster 		node = dag_h->succedents[i];
    429      1.3    oster 		if (node->visited == unvisited) {
    430      1.3    oster 			nnum = rf_RecurAssignNodeNums(dag_h->succedents[i], nnum, unvisited);
    431      1.3    oster 		}
    432      1.3    oster 	}
    433      1.3    oster 	return (nnum);
    434      1.1    oster }
    435      1.1    oster 
    436      1.3    oster int
    437      1.3    oster rf_RecurAssignNodeNums(node, num, unvisited)
    438      1.3    oster 	RF_DagNode_t *node;
    439      1.3    oster 	int     num;
    440      1.3    oster 	int     unvisited;
    441      1.3    oster {
    442      1.3    oster 	int     i;
    443      1.3    oster 
    444      1.3    oster 	node->visited = (unvisited) ? 0 : 1;
    445      1.3    oster 
    446      1.3    oster 	node->nodeNum = num++;
    447      1.3    oster 	for (i = 0; i < node->numSuccedents; i++) {
    448      1.3    oster 		if (node->succedents[i]->visited == unvisited) {
    449      1.3    oster 			num = rf_RecurAssignNodeNums(node->succedents[i], num, unvisited);
    450      1.3    oster 		}
    451      1.3    oster 	}
    452      1.3    oster 	return (num);
    453      1.3    oster }
    454      1.1    oster /* set the header pointers in each node to "newptr" */
    455      1.3    oster void
    456      1.3    oster rf_ResetDAGHeaderPointers(dag_h, newptr)
    457      1.3    oster 	RF_DagHeader_t *dag_h;
    458      1.3    oster 	RF_DagHeader_t *newptr;
    459      1.3    oster {
    460      1.3    oster 	int     i;
    461      1.3    oster 	for (i = 0; i < dag_h->numSuccedents; i++)
    462      1.3    oster 		if (dag_h->succedents[i]->dagHdr != newptr)
    463      1.3    oster 			rf_RecurResetDAGHeaderPointers(dag_h->succedents[i], newptr);
    464      1.1    oster }
    465      1.1    oster 
    466      1.3    oster void
    467      1.3    oster rf_RecurResetDAGHeaderPointers(node, newptr)
    468      1.3    oster 	RF_DagNode_t *node;
    469      1.3    oster 	RF_DagHeader_t *newptr;
    470      1.1    oster {
    471      1.3    oster 	int     i;
    472      1.3    oster 	node->dagHdr = newptr;
    473      1.3    oster 	for (i = 0; i < node->numSuccedents; i++)
    474      1.3    oster 		if (node->succedents[i]->dagHdr != newptr)
    475      1.3    oster 			rf_RecurResetDAGHeaderPointers(node->succedents[i], newptr);
    476      1.3    oster }
    477      1.1    oster 
    478      1.1    oster 
    479      1.3    oster void
    480      1.3    oster rf_PrintDAGList(RF_DagHeader_t * dag_h)
    481      1.3    oster {
    482      1.3    oster 	int     i = 0;
    483      1.3    oster 
    484      1.3    oster 	for (; dag_h; dag_h = dag_h->next) {
    485      1.3    oster 		rf_AssignNodeNums(dag_h);
    486      1.3    oster 		printf("\n\nDAG %d IN LIST:\n", i++);
    487      1.3    oster 		rf_PrintDAG(dag_h);
    488      1.3    oster 	}
    489      1.1    oster }
    490      1.1    oster 
    491      1.3    oster static int
    492      1.3    oster rf_ValidateBranch(node, scount, acount, nodes, unvisited)
    493      1.3    oster 	RF_DagNode_t *node;
    494      1.3    oster 	int    *scount;
    495      1.3    oster 	int    *acount;
    496      1.3    oster 	RF_DagNode_t **nodes;
    497      1.3    oster 	int     unvisited;
    498      1.3    oster {
    499      1.3    oster 	int     i, retcode = 0;
    500      1.3    oster 
    501      1.3    oster 	/* construct an array of node pointers indexed by node num */
    502      1.3    oster 	node->visited = (unvisited) ? 0 : 1;
    503      1.3    oster 	nodes[node->nodeNum] = node;
    504      1.3    oster 
    505      1.3    oster 	if (node->next != NULL) {
    506      1.3    oster 		printf("INVALID DAG: next pointer in node is not NULL\n");
    507      1.3    oster 		retcode = 1;
    508      1.3    oster 	}
    509      1.3    oster 	if (node->status != rf_wait) {
    510      1.3    oster 		printf("INVALID DAG: Node status is not wait\n");
    511      1.3    oster 		retcode = 1;
    512      1.3    oster 	}
    513      1.3    oster 	if (node->numAntDone != 0) {
    514      1.3    oster 		printf("INVALID DAG: numAntDone is not zero\n");
    515      1.3    oster 		retcode = 1;
    516      1.3    oster 	}
    517      1.3    oster 	if (node->doFunc == rf_TerminateFunc) {
    518      1.3    oster 		if (node->numSuccedents != 0) {
    519      1.3    oster 			printf("INVALID DAG: Terminator node has succedents\n");
    520      1.3    oster 			retcode = 1;
    521      1.3    oster 		}
    522      1.3    oster 	} else {
    523      1.3    oster 		if (node->numSuccedents == 0) {
    524      1.3    oster 			printf("INVALID DAG: Non-terminator node has no succedents\n");
    525      1.3    oster 			retcode = 1;
    526      1.3    oster 		}
    527      1.3    oster 	}
    528      1.3    oster 	for (i = 0; i < node->numSuccedents; i++) {
    529      1.3    oster 		if (!node->succedents[i]) {
    530      1.3    oster 			printf("INVALID DAG: succedent %d of node %s is NULL\n", i, node->name);
    531      1.3    oster 			retcode = 1;
    532      1.3    oster 		}
    533      1.3    oster 		scount[node->succedents[i]->nodeNum]++;
    534      1.3    oster 	}
    535      1.3    oster 	for (i = 0; i < node->numAntecedents; i++) {
    536      1.3    oster 		if (!node->antecedents[i]) {
    537      1.3    oster 			printf("INVALID DAG: antecedent %d of node %s is NULL\n", i, node->name);
    538      1.3    oster 			retcode = 1;
    539      1.3    oster 		}
    540      1.3    oster 		acount[node->antecedents[i]->nodeNum]++;
    541      1.3    oster 	}
    542      1.3    oster 	for (i = 0; i < node->numSuccedents; i++) {
    543      1.3    oster 		if (node->succedents[i]->visited == unvisited) {
    544      1.3    oster 			if (rf_ValidateBranch(node->succedents[i], scount,
    545      1.3    oster 				acount, nodes, unvisited)) {
    546      1.3    oster 				retcode = 1;
    547      1.3    oster 			}
    548      1.3    oster 		}
    549      1.3    oster 	}
    550      1.3    oster 	return (retcode);
    551      1.3    oster }
    552      1.3    oster 
    553      1.3    oster static void
    554      1.3    oster rf_ValidateBranchVisitedBits(node, unvisited, rl)
    555      1.3    oster 	RF_DagNode_t *node;
    556      1.3    oster 	int     unvisited;
    557      1.3    oster 	int     rl;
    558      1.3    oster {
    559      1.3    oster 	int     i;
    560      1.3    oster 
    561      1.3    oster 	RF_ASSERT(node->visited == unvisited);
    562      1.3    oster 	for (i = 0; i < node->numSuccedents; i++) {
    563      1.3    oster 		if (node->succedents[i] == NULL) {
    564      1.3    oster 			printf("node=%lx node->succedents[%d] is NULL\n", (long) node, i);
    565      1.3    oster 			RF_ASSERT(0);
    566      1.3    oster 		}
    567      1.3    oster 		rf_ValidateBranchVisitedBits(node->succedents[i], unvisited, rl + 1);
    568      1.3    oster 	}
    569      1.3    oster }
    570      1.3    oster /* NOTE:  never call this on a big dag, because it is exponential
    571      1.3    oster  * in execution time
    572      1.3    oster  */
    573      1.3    oster static void
    574      1.3    oster rf_ValidateVisitedBits(dag)
    575      1.3    oster 	RF_DagHeader_t *dag;
    576      1.3    oster {
    577      1.3    oster 	int     i, unvisited;
    578      1.3    oster 
    579      1.3    oster 	unvisited = dag->succedents[0]->visited;
    580      1.3    oster 
    581      1.3    oster 	for (i = 0; i < dag->numSuccedents; i++) {
    582      1.3    oster 		if (dag->succedents[i] == NULL) {
    583      1.3    oster 			printf("dag=%lx dag->succedents[%d] is NULL\n", (long) dag, i);
    584      1.3    oster 			RF_ASSERT(0);
    585      1.3    oster 		}
    586      1.3    oster 		rf_ValidateBranchVisitedBits(dag->succedents[i], unvisited, 0);
    587      1.3    oster 	}
    588      1.3    oster }
    589      1.1    oster /* validate a DAG.  _at entry_ verify that:
    590      1.1    oster  *   -- numNodesCompleted is zero
    591      1.1    oster  *   -- node queue is null
    592      1.1    oster  *   -- dag status is rf_enable
    593      1.1    oster  *   -- next pointer is null on every node
    594      1.1    oster  *   -- all nodes have status wait
    595      1.1    oster  *   -- numAntDone is zero in all nodes
    596      1.1    oster  *   -- terminator node has zero successors
    597      1.1    oster  *   -- no other node besides terminator has zero successors
    598      1.1    oster  *   -- no successor or antecedent pointer in a node is NULL
    599      1.1    oster  *   -- number of times that each node appears as a successor of another node
    600      1.1    oster  *      is equal to the antecedent count on that node
    601      1.1    oster  *   -- number of times that each node appears as an antecedent of another node
    602      1.1    oster  *      is equal to the succedent count on that node
    603      1.1    oster  *   -- what else?
    604      1.1    oster  */
    605      1.3    oster int
    606      1.3    oster rf_ValidateDAG(dag_h)
    607      1.3    oster 	RF_DagHeader_t *dag_h;
    608      1.3    oster {
    609      1.3    oster 	int     i, nodecount;
    610      1.3    oster 	int    *scount, *acount;/* per-node successor and antecedent counts */
    611      1.3    oster 	RF_DagNode_t **nodes;	/* array of ptrs to nodes in dag */
    612      1.3    oster 	int     retcode = 0;
    613      1.3    oster 	int     unvisited;
    614      1.3    oster 	int     commitNodeCount = 0;
    615      1.3    oster 
    616      1.3    oster 	if (rf_validateVisitedDebug)
    617      1.3    oster 		rf_ValidateVisitedBits(dag_h);
    618      1.3    oster 
    619      1.3    oster 	if (dag_h->numNodesCompleted != 0) {
    620      1.3    oster 		printf("INVALID DAG: num nodes completed is %d, should be 0\n", dag_h->numNodesCompleted);
    621      1.3    oster 		retcode = 1;
    622      1.3    oster 		goto validate_dag_bad;
    623      1.3    oster 	}
    624      1.3    oster 	if (dag_h->status != rf_enable) {
    625      1.3    oster 		printf("INVALID DAG: not enabled\n");
    626      1.3    oster 		retcode = 1;
    627      1.3    oster 		goto validate_dag_bad;
    628      1.3    oster 	}
    629      1.3    oster 	if (dag_h->numCommits != 0) {
    630      1.3    oster 		printf("INVALID DAG: numCommits != 0 (%d)\n", dag_h->numCommits);
    631      1.3    oster 		retcode = 1;
    632      1.3    oster 		goto validate_dag_bad;
    633      1.3    oster 	}
    634      1.3    oster 	if (dag_h->numSuccedents != 1) {
    635      1.3    oster 		/* currently, all dags must have only one succedent */
    636      1.3    oster 		printf("INVALID DAG: numSuccedents !1 (%d)\n", dag_h->numSuccedents);
    637      1.3    oster 		retcode = 1;
    638      1.3    oster 		goto validate_dag_bad;
    639      1.3    oster 	}
    640      1.3    oster 	nodecount = rf_AssignNodeNums(dag_h);
    641      1.3    oster 
    642      1.3    oster 	unvisited = dag_h->succedents[0]->visited;
    643      1.3    oster 
    644      1.3    oster 	RF_Calloc(scount, nodecount, sizeof(int), (int *));
    645      1.3    oster 	RF_Calloc(acount, nodecount, sizeof(int), (int *));
    646      1.3    oster 	RF_Calloc(nodes, nodecount, sizeof(RF_DagNode_t *), (RF_DagNode_t **));
    647      1.3    oster 	for (i = 0; i < dag_h->numSuccedents; i++) {
    648      1.3    oster 		if ((dag_h->succedents[i]->visited == unvisited)
    649      1.3    oster 		    && rf_ValidateBranch(dag_h->succedents[i], scount,
    650      1.3    oster 			acount, nodes, unvisited)) {
    651      1.3    oster 			retcode = 1;
    652      1.3    oster 		}
    653      1.3    oster 	}
    654      1.3    oster 	/* start at 1 to skip the header node */
    655      1.3    oster 	for (i = 1; i < nodecount; i++) {
    656      1.3    oster 		if (nodes[i]->commitNode)
    657      1.3    oster 			commitNodeCount++;
    658      1.3    oster 		if (nodes[i]->doFunc == NULL) {
    659      1.3    oster 			printf("INVALID DAG: node %s has an undefined doFunc\n", nodes[i]->name);
    660      1.3    oster 			retcode = 1;
    661      1.3    oster 			goto validate_dag_out;
    662      1.3    oster 		}
    663      1.3    oster 		if (nodes[i]->undoFunc == NULL) {
    664      1.3    oster 			printf("INVALID DAG: node %s has an undefined doFunc\n", nodes[i]->name);
    665      1.3    oster 			retcode = 1;
    666      1.3    oster 			goto validate_dag_out;
    667      1.3    oster 		}
    668      1.3    oster 		if (nodes[i]->numAntecedents != scount[nodes[i]->nodeNum]) {
    669      1.3    oster 			printf("INVALID DAG: node %s has %d antecedents but appears as a succedent %d times\n",
    670      1.3    oster 			    nodes[i]->name, nodes[i]->numAntecedents, scount[nodes[i]->nodeNum]);
    671      1.3    oster 			retcode = 1;
    672      1.3    oster 			goto validate_dag_out;
    673      1.3    oster 		}
    674      1.3    oster 		if (nodes[i]->numSuccedents != acount[nodes[i]->nodeNum]) {
    675      1.3    oster 			printf("INVALID DAG: node %s has %d succedents but appears as an antecedent %d times\n",
    676      1.3    oster 			    nodes[i]->name, nodes[i]->numSuccedents, acount[nodes[i]->nodeNum]);
    677      1.3    oster 			retcode = 1;
    678      1.3    oster 			goto validate_dag_out;
    679      1.3    oster 		}
    680      1.3    oster 	}
    681      1.1    oster 
    682      1.3    oster 	if (dag_h->numCommitNodes != commitNodeCount) {
    683      1.3    oster 		printf("INVALID DAG: incorrect commit node count.  hdr->numCommitNodes (%d) found (%d) commit nodes in graph\n",
    684      1.3    oster 		    dag_h->numCommitNodes, commitNodeCount);
    685      1.3    oster 		retcode = 1;
    686      1.3    oster 		goto validate_dag_out;
    687      1.3    oster 	}
    688      1.1    oster validate_dag_out:
    689      1.3    oster 	RF_Free(scount, nodecount * sizeof(int));
    690      1.3    oster 	RF_Free(acount, nodecount * sizeof(int));
    691      1.3    oster 	RF_Free(nodes, nodecount * sizeof(RF_DagNode_t *));
    692      1.3    oster 	if (retcode)
    693      1.3    oster 		rf_PrintDAGList(dag_h);
    694      1.3    oster 
    695      1.3    oster 	if (rf_validateVisitedDebug)
    696      1.3    oster 		rf_ValidateVisitedBits(dag_h);
    697      1.3    oster 
    698      1.3    oster 	return (retcode);
    699      1.1    oster 
    700      1.1    oster validate_dag_bad:
    701      1.3    oster 	rf_PrintDAGList(dag_h);
    702      1.3    oster 	return (retcode);
    703      1.1    oster }
    704      1.1    oster 
    705  1.6.6.5  nathanw #endif /* RF_DEBUG_VALIDATE_DAG */
    706      1.1    oster 
    707      1.1    oster /******************************************************************************
    708      1.1    oster  *
    709      1.1    oster  * misc construction routines
    710      1.1    oster  *
    711      1.1    oster  *****************************************************************************/
    712      1.1    oster 
    713      1.3    oster void
    714      1.3    oster rf_redirect_asm(
    715      1.3    oster     RF_Raid_t * raidPtr,
    716      1.3    oster     RF_AccessStripeMap_t * asmap)
    717      1.3    oster {
    718      1.3    oster 	int     ds = (raidPtr->Layout.map->flags & RF_DISTRIBUTE_SPARE) ? 1 : 0;
    719      1.3    oster 	int     row = asmap->physInfo->row;
    720      1.3    oster 	int     fcol = raidPtr->reconControl[row]->fcol;
    721      1.3    oster 	int     srow = raidPtr->reconControl[row]->spareRow;
    722      1.3    oster 	int     scol = raidPtr->reconControl[row]->spareCol;
    723      1.3    oster 	RF_PhysDiskAddr_t *pda;
    724      1.3    oster 
    725      1.3    oster 	RF_ASSERT(raidPtr->status[row] == rf_rs_reconstructing);
    726      1.3    oster 	for (pda = asmap->physInfo; pda; pda = pda->next) {
    727      1.3    oster 		if (pda->col == fcol) {
    728      1.3    oster 			if (rf_dagDebug) {
    729      1.3    oster 				if (!rf_CheckRUReconstructed(raidPtr->reconControl[row]->reconMap,
    730      1.3    oster 					pda->startSector)) {
    731      1.3    oster 					RF_PANIC();
    732      1.3    oster 				}
    733      1.3    oster 			}
    734      1.3    oster 			/* printf("Remapped data for large write\n"); */
    735      1.3    oster 			if (ds) {
    736      1.3    oster 				raidPtr->Layout.map->MapSector(raidPtr, pda->raidAddress,
    737      1.3    oster 				    &pda->row, &pda->col, &pda->startSector, RF_REMAP);
    738      1.3    oster 			} else {
    739      1.3    oster 				pda->row = srow;
    740      1.3    oster 				pda->col = scol;
    741      1.3    oster 			}
    742      1.3    oster 		}
    743      1.3    oster 	}
    744      1.3    oster 	for (pda = asmap->parityInfo; pda; pda = pda->next) {
    745      1.3    oster 		if (pda->col == fcol) {
    746      1.3    oster 			if (rf_dagDebug) {
    747      1.3    oster 				if (!rf_CheckRUReconstructed(raidPtr->reconControl[row]->reconMap, pda->startSector)) {
    748      1.3    oster 					RF_PANIC();
    749      1.3    oster 				}
    750      1.3    oster 			}
    751      1.3    oster 		}
    752      1.3    oster 		if (ds) {
    753      1.3    oster 			(raidPtr->Layout.map->MapParity) (raidPtr, pda->raidAddress, &pda->row, &pda->col, &pda->startSector, RF_REMAP);
    754      1.3    oster 		} else {
    755      1.3    oster 			pda->row = srow;
    756      1.3    oster 			pda->col = scol;
    757      1.3    oster 		}
    758      1.3    oster 	}
    759      1.1    oster }
    760      1.1    oster 
    761      1.1    oster 
    762      1.1    oster /* this routine allocates read buffers and generates stripe maps for the
    763      1.1    oster  * regions of the array from the start of the stripe to the start of the
    764      1.1    oster  * access, and from the end of the access to the end of the stripe.  It also
    765      1.1    oster  * computes and returns the number of DAG nodes needed to read all this data.
    766      1.1    oster  * Note that this routine does the wrong thing if the access is fully
    767      1.1    oster  * contained within one stripe unit, so we RF_ASSERT against this case at the
    768      1.1    oster  * start.
    769      1.1    oster  */
    770      1.3    oster void
    771      1.3    oster rf_MapUnaccessedPortionOfStripe(
    772      1.3    oster     RF_Raid_t * raidPtr,
    773      1.3    oster     RF_RaidLayout_t * layoutPtr,/* in: layout information */
    774      1.3    oster     RF_AccessStripeMap_t * asmap,	/* in: access stripe map */
    775      1.3    oster     RF_DagHeader_t * dag_h,	/* in: header of the dag to create */
    776      1.3    oster     RF_AccessStripeMapHeader_t ** new_asm_h,	/* in: ptr to array of 2
    777      1.3    oster 						 * headers, to be filled in */
    778      1.3    oster     int *nRodNodes,		/* out: num nodes to be generated to read
    779      1.3    oster 				 * unaccessed data */
    780      1.3    oster     char **sosBuffer,		/* out: pointers to newly allocated buffer */
    781      1.3    oster     char **eosBuffer,
    782      1.3    oster     RF_AllocListElem_t * allocList)
    783      1.3    oster {
    784      1.3    oster 	RF_RaidAddr_t sosRaidAddress, eosRaidAddress;
    785      1.3    oster 	RF_SectorNum_t sosNumSector, eosNumSector;
    786      1.3    oster 
    787      1.3    oster 	RF_ASSERT(asmap->numStripeUnitsAccessed > (layoutPtr->numDataCol / 2));
    788      1.3    oster 	/* generate an access map for the region of the array from start of
    789      1.3    oster 	 * stripe to start of access */
    790      1.3    oster 	new_asm_h[0] = new_asm_h[1] = NULL;
    791      1.3    oster 	*nRodNodes = 0;
    792      1.3    oster 	if (!rf_RaidAddressStripeAligned(layoutPtr, asmap->raidAddress)) {
    793      1.3    oster 		sosRaidAddress = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress);
    794      1.3    oster 		sosNumSector = asmap->raidAddress - sosRaidAddress;
    795      1.3    oster 		RF_MallocAndAdd(*sosBuffer, rf_RaidAddressToByte(raidPtr, sosNumSector), (char *), allocList);
    796      1.3    oster 		new_asm_h[0] = rf_MapAccess(raidPtr, sosRaidAddress, sosNumSector, *sosBuffer, RF_DONT_REMAP);
    797      1.3    oster 		new_asm_h[0]->next = dag_h->asmList;
    798      1.3    oster 		dag_h->asmList = new_asm_h[0];
    799      1.3    oster 		*nRodNodes += new_asm_h[0]->stripeMap->numStripeUnitsAccessed;
    800      1.3    oster 
    801      1.3    oster 		RF_ASSERT(new_asm_h[0]->stripeMap->next == NULL);
    802      1.3    oster 		/* we're totally within one stripe here */
    803      1.3    oster 		if (asmap->flags & RF_ASM_REDIR_LARGE_WRITE)
    804      1.3    oster 			rf_redirect_asm(raidPtr, new_asm_h[0]->stripeMap);
    805      1.3    oster 	}
    806      1.3    oster 	/* generate an access map for the region of the array from end of
    807      1.3    oster 	 * access to end of stripe */
    808      1.3    oster 	if (!rf_RaidAddressStripeAligned(layoutPtr, asmap->endRaidAddress)) {
    809      1.3    oster 		eosRaidAddress = asmap->endRaidAddress;
    810      1.3    oster 		eosNumSector = rf_RaidAddressOfNextStripeBoundary(layoutPtr, eosRaidAddress) - eosRaidAddress;
    811      1.3    oster 		RF_MallocAndAdd(*eosBuffer, rf_RaidAddressToByte(raidPtr, eosNumSector), (char *), allocList);
    812      1.3    oster 		new_asm_h[1] = rf_MapAccess(raidPtr, eosRaidAddress, eosNumSector, *eosBuffer, RF_DONT_REMAP);
    813      1.3    oster 		new_asm_h[1]->next = dag_h->asmList;
    814      1.3    oster 		dag_h->asmList = new_asm_h[1];
    815      1.3    oster 		*nRodNodes += new_asm_h[1]->stripeMap->numStripeUnitsAccessed;
    816      1.3    oster 
    817      1.3    oster 		RF_ASSERT(new_asm_h[1]->stripeMap->next == NULL);
    818      1.3    oster 		/* we're totally within one stripe here */
    819      1.3    oster 		if (asmap->flags & RF_ASM_REDIR_LARGE_WRITE)
    820      1.3    oster 			rf_redirect_asm(raidPtr, new_asm_h[1]->stripeMap);
    821      1.3    oster 	}
    822      1.1    oster }
    823      1.1    oster 
    824      1.1    oster 
    825      1.1    oster 
    826      1.1    oster /* returns non-zero if the indicated ranges of stripe unit offsets overlap */
    827      1.3    oster int
    828      1.3    oster rf_PDAOverlap(
    829      1.3    oster     RF_RaidLayout_t * layoutPtr,
    830      1.3    oster     RF_PhysDiskAddr_t * src,
    831      1.3    oster     RF_PhysDiskAddr_t * dest)
    832      1.3    oster {
    833      1.3    oster 	RF_SectorNum_t soffs = rf_StripeUnitOffset(layoutPtr, src->startSector);
    834      1.3    oster 	RF_SectorNum_t doffs = rf_StripeUnitOffset(layoutPtr, dest->startSector);
    835      1.3    oster 	/* use -1 to be sure we stay within SU */
    836      1.3    oster 	RF_SectorNum_t send = rf_StripeUnitOffset(layoutPtr, src->startSector + src->numSector - 1);
    837      1.3    oster 	RF_SectorNum_t dend = rf_StripeUnitOffset(layoutPtr, dest->startSector + dest->numSector - 1);
    838      1.3    oster 	return ((RF_MAX(soffs, doffs) <= RF_MIN(send, dend)) ? 1 : 0);
    839      1.1    oster }
    840      1.1    oster 
    841      1.1    oster 
    842      1.1    oster /* GenerateFailedAccessASMs
    843      1.1    oster  *
    844      1.1    oster  * this routine figures out what portion of the stripe needs to be read
    845      1.1    oster  * to effect the degraded read or write operation.  It's primary function
    846      1.1    oster  * is to identify everything required to recover the data, and then
    847      1.1    oster  * eliminate anything that is already being accessed by the user.
    848      1.1    oster  *
    849      1.1    oster  * The main result is two new ASMs, one for the region from the start of the
    850      1.1    oster  * stripe to the start of the access, and one for the region from the end of
    851      1.1    oster  * the access to the end of the stripe.  These ASMs describe everything that
    852      1.1    oster  * needs to be read to effect the degraded access.  Other results are:
    853      1.1    oster  *    nXorBufs -- the total number of buffers that need to be XORed together to
    854      1.1    oster  *                recover the lost data,
    855      1.1    oster  *    rpBufPtr -- ptr to a newly-allocated buffer to hold the parity.  If NULL
    856      1.1    oster  *                at entry, not allocated.
    857      1.1    oster  *    overlappingPDAs --
    858      1.1    oster  *                describes which of the non-failed PDAs in the user access
    859      1.1    oster  *                overlap data that needs to be read to effect recovery.
    860      1.1    oster  *                overlappingPDAs[i]==1 if and only if, neglecting the failed
    861      1.1    oster  *                PDA, the ith pda in the input asm overlaps data that needs
    862      1.1    oster  *                to be read for recovery.
    863      1.1    oster  */
    864      1.1    oster  /* in: asm - ASM for the actual access, one stripe only */
    865  1.6.6.4  nathanw  /* in: failedPDA - which component of the access has failed */
    866      1.1    oster  /* in: dag_h - header of the DAG we're going to create */
    867      1.1    oster  /* out: new_asm_h - the two new ASMs */
    868      1.1    oster  /* out: nXorBufs - the total number of xor bufs required */
    869      1.1    oster  /* out: rpBufPtr - a buffer for the parity read */
    870      1.3    oster void
    871      1.3    oster rf_GenerateFailedAccessASMs(
    872      1.3    oster     RF_Raid_t * raidPtr,
    873      1.3    oster     RF_AccessStripeMap_t * asmap,
    874      1.3    oster     RF_PhysDiskAddr_t * failedPDA,
    875      1.3    oster     RF_DagHeader_t * dag_h,
    876      1.3    oster     RF_AccessStripeMapHeader_t ** new_asm_h,
    877      1.3    oster     int *nXorBufs,
    878      1.3    oster     char **rpBufPtr,
    879      1.3    oster     char *overlappingPDAs,
    880      1.3    oster     RF_AllocListElem_t * allocList)
    881      1.3    oster {
    882      1.3    oster 	RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
    883      1.3    oster 
    884      1.3    oster 	/* s=start, e=end, s=stripe, a=access, f=failed, su=stripe unit */
    885      1.3    oster 	RF_RaidAddr_t sosAddr, sosEndAddr, eosStartAddr, eosAddr;
    886      1.3    oster 
    887      1.3    oster 	RF_SectorCount_t numSect[2], numParitySect;
    888      1.3    oster 	RF_PhysDiskAddr_t *pda;
    889      1.3    oster 	char   *rdBuf, *bufP;
    890      1.3    oster 	int     foundit, i;
    891      1.3    oster 
    892      1.3    oster 	bufP = NULL;
    893      1.3    oster 	foundit = 0;
    894      1.3    oster 	/* first compute the following raid addresses: start of stripe,
    895      1.3    oster 	 * (sosAddr) MIN(start of access, start of failed SU),   (sosEndAddr)
    896      1.3    oster 	 * MAX(end of access, end of failed SU),       (eosStartAddr) end of
    897      1.3    oster 	 * stripe (i.e. start of next stripe)   (eosAddr) */
    898      1.3    oster 	sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr, asmap->raidAddress);
    899      1.3    oster 	sosEndAddr = RF_MIN(asmap->raidAddress, rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->raidAddress));
    900      1.3    oster 	eosStartAddr = RF_MAX(asmap->endRaidAddress, rf_RaidAddressOfNextStripeUnitBoundary(layoutPtr, failedPDA->raidAddress));
    901      1.3    oster 	eosAddr = rf_RaidAddressOfNextStripeBoundary(layoutPtr, asmap->raidAddress);
    902      1.3    oster 
    903      1.3    oster 	/* now generate access stripe maps for each of the above regions of
    904      1.3    oster 	 * the stripe.  Use a dummy (NULL) buf ptr for now */
    905      1.3    oster 
    906      1.3    oster 	new_asm_h[0] = (sosAddr != sosEndAddr) ? rf_MapAccess(raidPtr, sosAddr, sosEndAddr - sosAddr, NULL, RF_DONT_REMAP) : NULL;
    907      1.3    oster 	new_asm_h[1] = (eosStartAddr != eosAddr) ? rf_MapAccess(raidPtr, eosStartAddr, eosAddr - eosStartAddr, NULL, RF_DONT_REMAP) : NULL;
    908      1.3    oster 
    909      1.3    oster 	/* walk through the PDAs and range-restrict each SU to the region of
    910      1.3    oster 	 * the SU touched on the failed PDA.  also compute total data buffer
    911      1.3    oster 	 * space requirements in this step.  Ignore the parity for now. */
    912      1.3    oster 
    913      1.3    oster 	numSect[0] = numSect[1] = 0;
    914      1.3    oster 	if (new_asm_h[0]) {
    915      1.3    oster 		new_asm_h[0]->next = dag_h->asmList;
    916      1.3    oster 		dag_h->asmList = new_asm_h[0];
    917      1.3    oster 		for (pda = new_asm_h[0]->stripeMap->physInfo; pda; pda = pda->next) {
    918      1.3    oster 			rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_NOBUFFER, 0);
    919      1.3    oster 			numSect[0] += pda->numSector;
    920      1.3    oster 		}
    921      1.3    oster 	}
    922      1.3    oster 	if (new_asm_h[1]) {
    923      1.3    oster 		new_asm_h[1]->next = dag_h->asmList;
    924      1.3    oster 		dag_h->asmList = new_asm_h[1];
    925      1.3    oster 		for (pda = new_asm_h[1]->stripeMap->physInfo; pda; pda = pda->next) {
    926      1.3    oster 			rf_RangeRestrictPDA(raidPtr, failedPDA, pda, RF_RESTRICT_NOBUFFER, 0);
    927      1.3    oster 			numSect[1] += pda->numSector;
    928      1.3    oster 		}
    929      1.3    oster 	}
    930      1.3    oster 	numParitySect = failedPDA->numSector;
    931      1.3    oster 
    932      1.3    oster 	/* allocate buffer space for the data & parity we have to read to
    933      1.3    oster 	 * recover from the failure */
    934      1.3    oster 
    935      1.3    oster 	if (numSect[0] + numSect[1] + ((rpBufPtr) ? numParitySect : 0)) {	/* don't allocate parity
    936      1.3    oster 										 * buf if not needed */
    937      1.3    oster 		RF_MallocAndAdd(rdBuf, rf_RaidAddressToByte(raidPtr, numSect[0] + numSect[1] + numParitySect), (char *), allocList);
    938      1.3    oster 		bufP = rdBuf;
    939      1.3    oster 		if (rf_degDagDebug)
    940      1.3    oster 			printf("Newly allocated buffer (%d bytes) is 0x%lx\n",
    941      1.3    oster 			    (int) rf_RaidAddressToByte(raidPtr, numSect[0] + numSect[1] + numParitySect), (unsigned long) bufP);
    942      1.3    oster 	}
    943      1.3    oster 	/* now walk through the pdas one last time and assign buffer pointers
    944      1.3    oster 	 * (ugh!).  Again, ignore the parity.  also, count nodes to find out
    945      1.3    oster 	 * how many bufs need to be xored together */
    946      1.3    oster 	(*nXorBufs) = 1;	/* in read case, 1 is for parity.  In write
    947      1.3    oster 				 * case, 1 is for failed data */
    948      1.3    oster 	if (new_asm_h[0]) {
    949      1.3    oster 		for (pda = new_asm_h[0]->stripeMap->physInfo; pda; pda = pda->next) {
    950      1.3    oster 			pda->bufPtr = bufP;
    951      1.3    oster 			bufP += rf_RaidAddressToByte(raidPtr, pda->numSector);
    952      1.3    oster 		}
    953      1.3    oster 		*nXorBufs += new_asm_h[0]->stripeMap->numStripeUnitsAccessed;
    954      1.3    oster 	}
    955      1.3    oster 	if (new_asm_h[1]) {
    956      1.3    oster 		for (pda = new_asm_h[1]->stripeMap->physInfo; pda; pda = pda->next) {
    957      1.3    oster 			pda->bufPtr = bufP;
    958      1.3    oster 			bufP += rf_RaidAddressToByte(raidPtr, pda->numSector);
    959      1.3    oster 		}
    960      1.3    oster 		(*nXorBufs) += new_asm_h[1]->stripeMap->numStripeUnitsAccessed;
    961      1.3    oster 	}
    962      1.3    oster 	if (rpBufPtr)
    963      1.3    oster 		*rpBufPtr = bufP;	/* the rest of the buffer is for
    964      1.3    oster 					 * parity */
    965      1.3    oster 
    966      1.3    oster 	/* the last step is to figure out how many more distinct buffers need
    967      1.3    oster 	 * to get xor'd to produce the missing unit.  there's one for each
    968      1.3    oster 	 * user-data read node that overlaps the portion of the failed unit
    969      1.3    oster 	 * being accessed */
    970      1.3    oster 
    971      1.3    oster 	for (foundit = i = 0, pda = asmap->physInfo; pda; i++, pda = pda->next) {
    972      1.3    oster 		if (pda == failedPDA) {
    973      1.3    oster 			i--;
    974      1.3    oster 			foundit = 1;
    975      1.3    oster 			continue;
    976      1.3    oster 		}
    977      1.3    oster 		if (rf_PDAOverlap(layoutPtr, pda, failedPDA)) {
    978      1.3    oster 			overlappingPDAs[i] = 1;
    979      1.3    oster 			(*nXorBufs)++;
    980      1.3    oster 		}
    981      1.3    oster 	}
    982      1.3    oster 	if (!foundit) {
    983      1.3    oster 		RF_ERRORMSG("GenerateFailedAccessASMs: did not find failedPDA in asm list\n");
    984      1.3    oster 		RF_ASSERT(0);
    985      1.3    oster 	}
    986      1.3    oster 	if (rf_degDagDebug) {
    987      1.3    oster 		if (new_asm_h[0]) {
    988      1.3    oster 			printf("First asm:\n");
    989      1.3    oster 			rf_PrintFullAccessStripeMap(new_asm_h[0], 1);
    990      1.3    oster 		}
    991      1.3    oster 		if (new_asm_h[1]) {
    992      1.3    oster 			printf("Second asm:\n");
    993      1.3    oster 			rf_PrintFullAccessStripeMap(new_asm_h[1], 1);
    994      1.3    oster 		}
    995      1.3    oster 	}
    996      1.1    oster }
    997      1.1    oster 
    998      1.1    oster 
    999      1.1    oster /* adjusts the offset and number of sectors in the destination pda so that
   1000      1.1    oster  * it covers at most the region of the SU covered by the source PDA.  This
   1001      1.1    oster  * is exclusively a restriction:  the number of sectors indicated by the
   1002      1.1    oster  * target PDA can only shrink.
   1003      1.1    oster  *
   1004      1.1    oster  * For example:  s = sectors within SU indicated by source PDA
   1005      1.1    oster  *               d = sectors within SU indicated by dest PDA
   1006      1.1    oster  *               r = results, stored in dest PDA
   1007      1.1    oster  *
   1008      1.1    oster  * |--------------- one stripe unit ---------------------|
   1009      1.1    oster  * |           sssssssssssssssssssssssssssssssss         |
   1010      1.1    oster  * |    ddddddddddddddddddddddddddddddddddddddddddddd    |
   1011      1.1    oster  * |           rrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr         |
   1012      1.1    oster  *
   1013      1.1    oster  * Another example:
   1014      1.1    oster  *
   1015      1.1    oster  * |--------------- one stripe unit ---------------------|
   1016      1.1    oster  * |           sssssssssssssssssssssssssssssssss         |
   1017      1.1    oster  * |    ddddddddddddddddddddddd                          |
   1018      1.1    oster  * |           rrrrrrrrrrrrrrrr                          |
   1019      1.1    oster  *
   1020      1.1    oster  */
   1021      1.3    oster void
   1022      1.3    oster rf_RangeRestrictPDA(
   1023      1.3    oster     RF_Raid_t * raidPtr,
   1024      1.3    oster     RF_PhysDiskAddr_t * src,
   1025      1.3    oster     RF_PhysDiskAddr_t * dest,
   1026      1.3    oster     int dobuffer,
   1027      1.3    oster     int doraidaddr)
   1028      1.3    oster {
   1029      1.3    oster 	RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
   1030      1.3    oster 	RF_SectorNum_t soffs = rf_StripeUnitOffset(layoutPtr, src->startSector);
   1031      1.3    oster 	RF_SectorNum_t doffs = rf_StripeUnitOffset(layoutPtr, dest->startSector);
   1032      1.3    oster 	RF_SectorNum_t send = rf_StripeUnitOffset(layoutPtr, src->startSector + src->numSector - 1);	/* use -1 to be sure we
   1033      1.3    oster 													 * stay within SU */
   1034      1.3    oster 	RF_SectorNum_t dend = rf_StripeUnitOffset(layoutPtr, dest->startSector + dest->numSector - 1);
   1035      1.3    oster 	RF_SectorNum_t subAddr = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, dest->startSector);	/* stripe unit boundary */
   1036      1.3    oster 
   1037      1.3    oster 	dest->startSector = subAddr + RF_MAX(soffs, doffs);
   1038      1.3    oster 	dest->numSector = subAddr + RF_MIN(send, dend) + 1 - dest->startSector;
   1039      1.3    oster 
   1040      1.3    oster 	if (dobuffer)
   1041      1.3    oster 		dest->bufPtr += (soffs > doffs) ? rf_RaidAddressToByte(raidPtr, soffs - doffs) : 0;
   1042      1.3    oster 	if (doraidaddr) {
   1043      1.3    oster 		dest->raidAddress = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, dest->raidAddress) +
   1044      1.3    oster 		    rf_StripeUnitOffset(layoutPtr, dest->startSector);
   1045      1.3    oster 	}
   1046      1.1    oster }
   1047  1.6.6.5  nathanw 
   1048  1.6.6.5  nathanw #if (RF_INCLUDE_CHAINDECLUSTER > 0)
   1049  1.6.6.5  nathanw 
   1050      1.1    oster /*
   1051      1.1    oster  * Want the highest of these primes to be the largest one
   1052      1.1    oster  * less than the max expected number of columns (won't hurt
   1053      1.1    oster  * to be too small or too large, but won't be optimal, either)
   1054      1.1    oster  * --jimz
   1055      1.1    oster  */
   1056      1.1    oster #define NLOWPRIMES 8
   1057      1.3    oster static int lowprimes[NLOWPRIMES] = {2, 3, 5, 7, 11, 13, 17, 19};
   1058      1.1    oster /*****************************************************************************
   1059      1.1    oster  * compute the workload shift factor.  (chained declustering)
   1060      1.1    oster  *
   1061      1.1    oster  * return nonzero if access should shift to secondary, otherwise,
   1062      1.1    oster  * access is to primary
   1063      1.1    oster  *****************************************************************************/
   1064      1.3    oster int
   1065      1.3    oster rf_compute_workload_shift(
   1066      1.3    oster     RF_Raid_t * raidPtr,
   1067      1.3    oster     RF_PhysDiskAddr_t * pda)
   1068      1.3    oster {
   1069      1.3    oster 	/*
   1070      1.3    oster          * variables:
   1071      1.3    oster          *  d   = column of disk containing primary
   1072      1.3    oster          *  f   = column of failed disk
   1073      1.3    oster          *  n   = number of disks in array
   1074      1.3    oster          *  sd  = "shift distance" (number of columns that d is to the right of f)
   1075      1.3    oster          *  row = row of array the access is in
   1076      1.3    oster          *  v   = numerator of redirection ratio
   1077      1.3    oster          *  k   = denominator of redirection ratio
   1078      1.3    oster          */
   1079      1.3    oster 	RF_RowCol_t d, f, sd, row, n;
   1080      1.3    oster 	int     k, v, ret, i;
   1081      1.3    oster 
   1082      1.3    oster 	row = pda->row;
   1083      1.3    oster 	n = raidPtr->numCol;
   1084      1.3    oster 
   1085      1.3    oster 	/* assign column of primary copy to d */
   1086      1.3    oster 	d = pda->col;
   1087      1.3    oster 
   1088      1.3    oster 	/* assign column of dead disk to f */
   1089      1.3    oster 	for (f = 0; ((!RF_DEAD_DISK(raidPtr->Disks[row][f].status)) && (f < n)); f++);
   1090      1.3    oster 
   1091      1.3    oster 	RF_ASSERT(f < n);
   1092      1.3    oster 	RF_ASSERT(f != d);
   1093      1.3    oster 
   1094      1.3    oster 	sd = (f > d) ? (n + d - f) : (d - f);
   1095      1.3    oster 	RF_ASSERT(sd < n);
   1096      1.3    oster 
   1097      1.3    oster 	/*
   1098      1.3    oster          * v of every k accesses should be redirected
   1099      1.3    oster          *
   1100      1.3    oster          * v/k := (n-1-sd)/(n-1)
   1101      1.3    oster          */
   1102      1.3    oster 	v = (n - 1 - sd);
   1103      1.3    oster 	k = (n - 1);
   1104      1.1    oster 
   1105      1.1    oster #if 1
   1106      1.3    oster 	/*
   1107      1.3    oster          * XXX
   1108      1.3    oster          * Is this worth it?
   1109      1.3    oster          *
   1110      1.3    oster          * Now reduce the fraction, by repeatedly factoring
   1111      1.3    oster          * out primes (just like they teach in elementary school!)
   1112      1.3    oster          */
   1113      1.3    oster 	for (i = 0; i < NLOWPRIMES; i++) {
   1114      1.3    oster 		if (lowprimes[i] > v)
   1115      1.3    oster 			break;
   1116      1.3    oster 		while (((v % lowprimes[i]) == 0) && ((k % lowprimes[i]) == 0)) {
   1117      1.3    oster 			v /= lowprimes[i];
   1118      1.3    oster 			k /= lowprimes[i];
   1119      1.3    oster 		}
   1120      1.3    oster 	}
   1121      1.1    oster #endif
   1122      1.1    oster 
   1123      1.3    oster 	raidPtr->hist_diskreq[row][d]++;
   1124      1.3    oster 	if (raidPtr->hist_diskreq[row][d] > v) {
   1125      1.3    oster 		ret = 0;	/* do not redirect */
   1126      1.3    oster 	} else {
   1127      1.3    oster 		ret = 1;	/* redirect */
   1128      1.3    oster 	}
   1129      1.1    oster 
   1130      1.1    oster #if 0
   1131      1.3    oster 	printf("d=%d f=%d sd=%d v=%d k=%d ret=%d h=%d\n", d, f, sd, v, k, ret,
   1132      1.3    oster 	    raidPtr->hist_diskreq[row][d]);
   1133      1.1    oster #endif
   1134      1.1    oster 
   1135      1.3    oster 	if (raidPtr->hist_diskreq[row][d] >= k) {
   1136      1.3    oster 		/* reset counter */
   1137      1.3    oster 		raidPtr->hist_diskreq[row][d] = 0;
   1138      1.3    oster 	}
   1139      1.3    oster 	return (ret);
   1140      1.1    oster }
   1141  1.6.6.5  nathanw #endif /* (RF_INCLUDE_CHAINDECLUSTER > 0) */
   1142  1.6.6.5  nathanw 
   1143      1.1    oster /*
   1144      1.1    oster  * Disk selection routines
   1145      1.1    oster  */
   1146      1.1    oster 
   1147      1.1    oster /*
   1148      1.1    oster  * Selects the disk with the shortest queue from a mirror pair.
   1149      1.1    oster  * Both the disk I/Os queued in RAIDframe as well as those at the physical
   1150      1.1    oster  * disk are counted as members of the "queue"
   1151      1.1    oster  */
   1152      1.3    oster void
   1153      1.3    oster rf_SelectMirrorDiskIdle(RF_DagNode_t * node)
   1154      1.1    oster {
   1155      1.3    oster 	RF_Raid_t *raidPtr = (RF_Raid_t *) node->dagHdr->raidPtr;
   1156      1.3    oster 	RF_RowCol_t rowData, colData, rowMirror, colMirror;
   1157      1.3    oster 	int     dataQueueLength, mirrorQueueLength, usemirror;
   1158      1.3    oster 	RF_PhysDiskAddr_t *data_pda = (RF_PhysDiskAddr_t *) node->params[0].p;
   1159      1.3    oster 	RF_PhysDiskAddr_t *mirror_pda = (RF_PhysDiskAddr_t *) node->params[4].p;
   1160      1.3    oster 	RF_PhysDiskAddr_t *tmp_pda;
   1161      1.3    oster 	RF_RaidDisk_t **disks = raidPtr->Disks;
   1162      1.3    oster 	RF_DiskQueue_t **dqs = raidPtr->Queues, *dataQueue, *mirrorQueue;
   1163      1.3    oster 
   1164      1.3    oster 	/* return the [row col] of the disk with the shortest queue */
   1165      1.3    oster 	rowData = data_pda->row;
   1166      1.3    oster 	colData = data_pda->col;
   1167      1.3    oster 	rowMirror = mirror_pda->row;
   1168      1.3    oster 	colMirror = mirror_pda->col;
   1169      1.3    oster 	dataQueue = &(dqs[rowData][colData]);
   1170      1.3    oster 	mirrorQueue = &(dqs[rowMirror][colMirror]);
   1171      1.1    oster 
   1172      1.1    oster #ifdef RF_LOCK_QUEUES_TO_READ_LEN
   1173      1.3    oster 	RF_LOCK_QUEUE_MUTEX(dataQueue, "SelectMirrorDiskIdle");
   1174      1.3    oster #endif				/* RF_LOCK_QUEUES_TO_READ_LEN */
   1175      1.3    oster 	dataQueueLength = dataQueue->queueLength + dataQueue->numOutstanding;
   1176      1.1    oster #ifdef RF_LOCK_QUEUES_TO_READ_LEN
   1177      1.3    oster 	RF_UNLOCK_QUEUE_MUTEX(dataQueue, "SelectMirrorDiskIdle");
   1178      1.3    oster 	RF_LOCK_QUEUE_MUTEX(mirrorQueue, "SelectMirrorDiskIdle");
   1179      1.3    oster #endif				/* RF_LOCK_QUEUES_TO_READ_LEN */
   1180      1.3    oster 	mirrorQueueLength = mirrorQueue->queueLength + mirrorQueue->numOutstanding;
   1181      1.1    oster #ifdef RF_LOCK_QUEUES_TO_READ_LEN
   1182      1.3    oster 	RF_UNLOCK_QUEUE_MUTEX(mirrorQueue, "SelectMirrorDiskIdle");
   1183      1.3    oster #endif				/* RF_LOCK_QUEUES_TO_READ_LEN */
   1184      1.1    oster 
   1185      1.3    oster 	usemirror = 0;
   1186      1.3    oster 	if (RF_DEAD_DISK(disks[rowMirror][colMirror].status)) {
   1187      1.3    oster 		usemirror = 0;
   1188      1.3    oster 	} else
   1189      1.3    oster 		if (RF_DEAD_DISK(disks[rowData][colData].status)) {
   1190      1.3    oster 			usemirror = 1;
   1191      1.3    oster 		} else
   1192      1.5    oster 			if (raidPtr->parity_good == RF_RAID_DIRTY) {
   1193      1.5    oster 				/* Trust only the main disk */
   1194      1.3    oster 				usemirror = 0;
   1195      1.3    oster 			} else
   1196      1.5    oster 				if (dataQueueLength < mirrorQueueLength) {
   1197      1.5    oster 					usemirror = 0;
   1198      1.5    oster 				} else
   1199      1.5    oster 					if (mirrorQueueLength < dataQueueLength) {
   1200      1.5    oster 						usemirror = 1;
   1201      1.3    oster 					} else {
   1202      1.5    oster 						/* queues are equal length. attempt
   1203      1.5    oster 						 * cleverness. */
   1204      1.5    oster 						if (SNUM_DIFF(dataQueue->last_deq_sector, data_pda->startSector)
   1205      1.5    oster 						    <= SNUM_DIFF(mirrorQueue->last_deq_sector, mirror_pda->startSector)) {
   1206      1.5    oster 							usemirror = 0;
   1207      1.5    oster 						} else {
   1208      1.5    oster 							usemirror = 1;
   1209      1.5    oster 						}
   1210      1.3    oster 					}
   1211      1.3    oster 
   1212      1.3    oster 	if (usemirror) {
   1213      1.3    oster 		/* use mirror (parity) disk, swap params 0 & 4 */
   1214      1.3    oster 		tmp_pda = data_pda;
   1215      1.3    oster 		node->params[0].p = mirror_pda;
   1216      1.3    oster 		node->params[4].p = tmp_pda;
   1217      1.3    oster 	} else {
   1218      1.3    oster 		/* use data disk, leave param 0 unchanged */
   1219      1.3    oster 	}
   1220      1.3    oster 	/* printf("dataQueueLength %d, mirrorQueueLength
   1221      1.3    oster 	 * %d\n",dataQueueLength, mirrorQueueLength); */
   1222      1.1    oster }
   1223  1.6.6.9  thorpej #if (RF_INCLUDE_CHAINDECLUSTER > 0) || (RF_INCLUDE_INTERDECLUSTER > 0) || (RF_DEBUG_VALIDATE_DAG > 0)
   1224      1.1    oster /*
   1225      1.1    oster  * Do simple partitioning. This assumes that
   1226      1.1    oster  * the data and parity disks are laid out identically.
   1227      1.1    oster  */
   1228      1.3    oster void
   1229      1.3    oster rf_SelectMirrorDiskPartition(RF_DagNode_t * node)
   1230      1.1    oster {
   1231      1.3    oster 	RF_Raid_t *raidPtr = (RF_Raid_t *) node->dagHdr->raidPtr;
   1232      1.3    oster 	RF_RowCol_t rowData, colData, rowMirror, colMirror;
   1233      1.3    oster 	RF_PhysDiskAddr_t *data_pda = (RF_PhysDiskAddr_t *) node->params[0].p;
   1234      1.3    oster 	RF_PhysDiskAddr_t *mirror_pda = (RF_PhysDiskAddr_t *) node->params[4].p;
   1235      1.3    oster 	RF_PhysDiskAddr_t *tmp_pda;
   1236      1.3    oster 	RF_RaidDisk_t **disks = raidPtr->Disks;
   1237      1.3    oster 	int     usemirror;
   1238      1.3    oster 
   1239      1.3    oster 	/* return the [row col] of the disk with the shortest queue */
   1240      1.3    oster 	rowData = data_pda->row;
   1241      1.3    oster 	colData = data_pda->col;
   1242      1.3    oster 	rowMirror = mirror_pda->row;
   1243      1.3    oster 	colMirror = mirror_pda->col;
   1244      1.3    oster 
   1245      1.3    oster 	usemirror = 0;
   1246      1.3    oster 	if (RF_DEAD_DISK(disks[rowMirror][colMirror].status)) {
   1247      1.3    oster 		usemirror = 0;
   1248      1.3    oster 	} else
   1249      1.3    oster 		if (RF_DEAD_DISK(disks[rowData][colData].status)) {
   1250      1.3    oster 			usemirror = 1;
   1251      1.6    oster 		} else
   1252      1.6    oster 			if (raidPtr->parity_good == RF_RAID_DIRTY) {
   1253      1.6    oster 				/* Trust only the main disk */
   1254      1.3    oster 				usemirror = 0;
   1255      1.6    oster 			} else
   1256      1.6    oster 				if (data_pda->startSector <
   1257      1.6    oster 				    (disks[rowData][colData].numBlocks / 2)) {
   1258      1.6    oster 					usemirror = 0;
   1259      1.6    oster 				} else {
   1260      1.6    oster 					usemirror = 1;
   1261      1.6    oster 				}
   1262      1.3    oster 
   1263      1.3    oster 	if (usemirror) {
   1264      1.3    oster 		/* use mirror (parity) disk, swap params 0 & 4 */
   1265      1.3    oster 		tmp_pda = data_pda;
   1266      1.3    oster 		node->params[0].p = mirror_pda;
   1267      1.3    oster 		node->params[4].p = tmp_pda;
   1268      1.3    oster 	} else {
   1269      1.3    oster 		/* use data disk, leave param 0 unchanged */
   1270      1.3    oster 	}
   1271      1.1    oster }
   1272  1.6.6.9  thorpej #endif
   1273