Home | History | Annotate | Line # | Download | only in raidframe
rf_raid5.c revision 1.3.8.1
      1  1.3.8.1  bouyer /*	$NetBSD: rf_raid5.c,v 1.3.8.1 2000/11/20 11:42:58 bouyer 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  * Author: Mark Holland
      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_raid5.c -- implements RAID Level 5
     32      1.1   oster  *
     33      1.1   oster  *****************************************************************************/
     34      1.1   oster 
     35      1.1   oster #include "rf_types.h"
     36      1.1   oster #include "rf_raid.h"
     37      1.1   oster #include "rf_raid5.h"
     38      1.1   oster #include "rf_dag.h"
     39      1.1   oster #include "rf_dagffrd.h"
     40      1.1   oster #include "rf_dagffwr.h"
     41      1.1   oster #include "rf_dagdegrd.h"
     42      1.1   oster #include "rf_dagdegwr.h"
     43      1.1   oster #include "rf_dagutils.h"
     44      1.1   oster #include "rf_general.h"
     45      1.1   oster #include "rf_map.h"
     46      1.1   oster #include "rf_utils.h"
     47      1.1   oster 
     48      1.1   oster typedef struct RF_Raid5ConfigInfo_s {
     49      1.3   oster 	RF_RowCol_t **stripeIdentifier;	/* filled in at config time and used
     50      1.3   oster 					 * by IdentifyStripe */
     51      1.3   oster }       RF_Raid5ConfigInfo_t;
     52      1.3   oster 
     53      1.3   oster int
     54      1.3   oster rf_ConfigureRAID5(
     55      1.3   oster     RF_ShutdownList_t ** listp,
     56      1.3   oster     RF_Raid_t * raidPtr,
     57      1.3   oster     RF_Config_t * cfgPtr)
     58      1.3   oster {
     59      1.3   oster 	RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
     60      1.3   oster 	RF_Raid5ConfigInfo_t *info;
     61      1.3   oster 	RF_RowCol_t i, j, startdisk;
     62      1.3   oster 
     63      1.3   oster 	/* create a RAID level 5 configuration structure */
     64      1.3   oster 	RF_MallocAndAdd(info, sizeof(RF_Raid5ConfigInfo_t), (RF_Raid5ConfigInfo_t *), raidPtr->cleanupList);
     65      1.3   oster 	if (info == NULL)
     66      1.3   oster 		return (ENOMEM);
     67      1.3   oster 	layoutPtr->layoutSpecificInfo = (void *) info;
     68      1.3   oster 
     69      1.3   oster 	RF_ASSERT(raidPtr->numRow == 1);
     70      1.3   oster 
     71      1.3   oster 	/* the stripe identifier must identify the disks in each stripe, IN
     72      1.3   oster 	 * THE ORDER THAT THEY APPEAR IN THE STRIPE. */
     73      1.3   oster 	info->stripeIdentifier = rf_make_2d_array(raidPtr->numCol, raidPtr->numCol, raidPtr->cleanupList);
     74      1.3   oster 	if (info->stripeIdentifier == NULL)
     75      1.3   oster 		return (ENOMEM);
     76      1.3   oster 	startdisk = 0;
     77      1.3   oster 	for (i = 0; i < raidPtr->numCol; i++) {
     78      1.3   oster 		for (j = 0; j < raidPtr->numCol; j++) {
     79      1.3   oster 			info->stripeIdentifier[i][j] = (startdisk + j) % raidPtr->numCol;
     80      1.3   oster 		}
     81      1.3   oster 		if ((--startdisk) < 0)
     82      1.3   oster 			startdisk = raidPtr->numCol - 1;
     83      1.3   oster 	}
     84      1.1   oster 
     85      1.3   oster 	/* fill in the remaining layout parameters */
     86      1.3   oster 	layoutPtr->numStripe = layoutPtr->stripeUnitsPerDisk;
     87      1.3   oster 	layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector;
     88      1.3   oster 	layoutPtr->numDataCol = raidPtr->numCol - 1;
     89      1.3   oster 	layoutPtr->dataSectorsPerStripe = layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
     90      1.3   oster 	layoutPtr->numParityCol = 1;
     91      1.3   oster 	layoutPtr->dataStripeUnitsPerDisk = layoutPtr->stripeUnitsPerDisk;
     92      1.1   oster 
     93      1.3   oster 	raidPtr->totalSectors = layoutPtr->stripeUnitsPerDisk * layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
     94      1.1   oster 
     95      1.3   oster 	return (0);
     96      1.1   oster }
     97      1.1   oster 
     98      1.3   oster int
     99      1.3   oster rf_GetDefaultNumFloatingReconBuffersRAID5(RF_Raid_t * raidPtr)
    100      1.1   oster {
    101      1.3   oster 	return (20);
    102      1.1   oster }
    103      1.1   oster 
    104      1.3   oster RF_HeadSepLimit_t
    105      1.3   oster rf_GetDefaultHeadSepLimitRAID5(RF_Raid_t * raidPtr)
    106      1.1   oster {
    107      1.3   oster 	return (10);
    108      1.1   oster }
    109      1.1   oster #if !defined(__NetBSD__) && !defined(_KERNEL)
    110      1.1   oster /* not currently used */
    111      1.3   oster int
    112      1.3   oster rf_ShutdownRAID5(RF_Raid_t * raidPtr)
    113      1.1   oster {
    114      1.3   oster 	return (0);
    115      1.1   oster }
    116      1.1   oster #endif
    117      1.1   oster 
    118      1.3   oster void
    119      1.3   oster rf_MapSectorRAID5(
    120      1.3   oster     RF_Raid_t * raidPtr,
    121      1.3   oster     RF_RaidAddr_t raidSector,
    122      1.3   oster     RF_RowCol_t * row,
    123      1.3   oster     RF_RowCol_t * col,
    124      1.3   oster     RF_SectorNum_t * diskSector,
    125      1.3   oster     int remap)
    126      1.1   oster {
    127      1.3   oster 	RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
    128      1.3   oster 	*row = 0;
    129      1.3   oster 	*col = (SUID % raidPtr->numCol);
    130      1.3   oster 	*diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
    131      1.3   oster 	    (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
    132      1.1   oster }
    133      1.1   oster 
    134      1.3   oster void
    135      1.3   oster rf_MapParityRAID5(
    136      1.3   oster     RF_Raid_t * raidPtr,
    137      1.3   oster     RF_RaidAddr_t raidSector,
    138      1.3   oster     RF_RowCol_t * row,
    139      1.3   oster     RF_RowCol_t * col,
    140      1.3   oster     RF_SectorNum_t * diskSector,
    141      1.3   oster     int remap)
    142      1.1   oster {
    143      1.3   oster 	RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
    144      1.3   oster 
    145      1.3   oster 	*row = 0;
    146      1.3   oster 	*col = raidPtr->Layout.numDataCol - (SUID / raidPtr->Layout.numDataCol) % raidPtr->numCol;
    147      1.3   oster 	*diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
    148      1.3   oster 	    (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
    149      1.1   oster }
    150      1.1   oster 
    151      1.3   oster void
    152      1.3   oster rf_IdentifyStripeRAID5(
    153      1.3   oster     RF_Raid_t * raidPtr,
    154      1.3   oster     RF_RaidAddr_t addr,
    155      1.3   oster     RF_RowCol_t ** diskids,
    156      1.3   oster     RF_RowCol_t * outRow)
    157      1.1   oster {
    158      1.3   oster 	RF_StripeNum_t stripeID = rf_RaidAddressToStripeID(&raidPtr->Layout, addr);
    159      1.3   oster 	RF_Raid5ConfigInfo_t *info = (RF_Raid5ConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo;
    160      1.1   oster 
    161      1.3   oster 	*outRow = 0;
    162      1.3   oster 	*diskids = info->stripeIdentifier[stripeID % raidPtr->numCol];
    163      1.1   oster }
    164      1.1   oster 
    165      1.3   oster void
    166      1.3   oster rf_MapSIDToPSIDRAID5(
    167      1.3   oster     RF_RaidLayout_t * layoutPtr,
    168      1.3   oster     RF_StripeNum_t stripeID,
    169      1.3   oster     RF_StripeNum_t * psID,
    170      1.3   oster     RF_ReconUnitNum_t * which_ru)
    171      1.1   oster {
    172      1.3   oster 	*which_ru = 0;
    173      1.3   oster 	*psID = stripeID;
    174      1.1   oster }
    175      1.1   oster /* select an algorithm for performing an access.  Returns two pointers,
    176      1.1   oster  * one to a function that will return information about the DAG, and
    177      1.1   oster  * another to a function that will create the dag.
    178      1.1   oster  */
    179      1.3   oster void
    180      1.3   oster rf_RaidFiveDagSelect(
    181      1.3   oster     RF_Raid_t * raidPtr,
    182      1.3   oster     RF_IoType_t type,
    183      1.3   oster     RF_AccessStripeMap_t * asmap,
    184      1.3   oster     RF_VoidFuncPtr * createFunc)
    185      1.1   oster {
    186      1.3   oster 	RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
    187      1.3   oster 	RF_PhysDiskAddr_t *failedPDA = NULL;
    188      1.3   oster 	RF_RowCol_t frow, fcol;
    189      1.3   oster 	RF_RowStatus_t rstat;
    190      1.3   oster 	int     prior_recon;
    191      1.3   oster 
    192      1.3   oster 	RF_ASSERT(RF_IO_IS_R_OR_W(type));
    193      1.3   oster 
    194      1.3   oster 	if (asmap->numDataFailed + asmap->numParityFailed > 1) {
    195      1.3   oster 		RF_ERRORMSG("Multiple disks failed in a single group!  Aborting I/O operation.\n");
    196      1.3   oster 		 /* *infoFunc = */ *createFunc = NULL;
    197      1.3   oster 		return;
    198      1.3   oster 	} else
    199      1.3   oster 		if (asmap->numDataFailed + asmap->numParityFailed == 1) {
    200      1.3   oster 
    201      1.3   oster 			/* if under recon & already reconstructed, redirect
    202      1.3   oster 			 * the access to the spare drive and eliminate the
    203      1.3   oster 			 * failure indication */
    204      1.3   oster 			failedPDA = asmap->failedPDAs[0];
    205      1.3   oster 			frow = failedPDA->row;
    206      1.3   oster 			fcol = failedPDA->col;
    207      1.3   oster 			rstat = raidPtr->status[failedPDA->row];
    208      1.3   oster 			prior_recon = (rstat == rf_rs_reconfigured) || (
    209      1.3   oster 			    (rstat == rf_rs_reconstructing) ?
    210      1.3   oster 			    rf_CheckRUReconstructed(raidPtr->reconControl[frow]->reconMap, failedPDA->startSector) : 0
    211      1.3   oster 			    );
    212      1.3   oster 			if (prior_recon) {
    213      1.3   oster 				RF_RowCol_t or = failedPDA->row, oc = failedPDA->col;
    214      1.3   oster 				RF_SectorNum_t oo = failedPDA->startSector;
    215      1.3   oster 
    216      1.3   oster 				if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) {	/* redirect to dist
    217      1.3   oster 											 * spare space */
    218      1.3   oster 
    219      1.3   oster 					if (failedPDA == asmap->parityInfo) {
    220      1.3   oster 
    221      1.3   oster 						/* parity has failed */
    222      1.3   oster 						(layoutPtr->map->MapParity) (raidPtr, failedPDA->raidAddress, &failedPDA->row,
    223      1.3   oster 						    &failedPDA->col, &failedPDA->startSector, RF_REMAP);
    224      1.3   oster 
    225      1.3   oster 						if (asmap->parityInfo->next) {	/* redir 2nd component,
    226      1.3   oster 										 * if any */
    227      1.3   oster 							RF_PhysDiskAddr_t *p = asmap->parityInfo->next;
    228      1.3   oster 							RF_SectorNum_t SUoffs = p->startSector % layoutPtr->sectorsPerStripeUnit;
    229      1.3   oster 							p->row = failedPDA->row;
    230      1.3   oster 							p->col = failedPDA->col;
    231      1.3   oster 							p->startSector = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->startSector) +
    232      1.3   oster 							    SUoffs;	/* cheating:
    233      1.3   oster 									 * startSector is not
    234      1.3   oster 									 * really a RAID address */
    235      1.3   oster 						}
    236      1.3   oster 					} else
    237      1.3   oster 						if (asmap->parityInfo->next && failedPDA == asmap->parityInfo->next) {
    238      1.3   oster 							RF_ASSERT(0);	/* should not ever
    239      1.3   oster 									 * happen */
    240      1.3   oster 						} else {
    241      1.3   oster 
    242      1.3   oster 							/* data has failed */
    243      1.3   oster 							(layoutPtr->map->MapSector) (raidPtr, failedPDA->raidAddress, &failedPDA->row,
    244      1.3   oster 							    &failedPDA->col, &failedPDA->startSector, RF_REMAP);
    245      1.3   oster 
    246      1.3   oster 						}
    247      1.3   oster 
    248      1.3   oster 				} else {	/* redirect to dedicated spare
    249      1.3   oster 						 * space */
    250      1.3   oster 
    251      1.3   oster 					failedPDA->row = raidPtr->Disks[frow][fcol].spareRow;
    252      1.3   oster 					failedPDA->col = raidPtr->Disks[frow][fcol].spareCol;
    253      1.3   oster 
    254      1.3   oster 					/* the parity may have two distinct
    255      1.3   oster 					 * components, both of which may need
    256      1.3   oster 					 * to be redirected */
    257      1.3   oster 					if (asmap->parityInfo->next) {
    258      1.3   oster 						if (failedPDA == asmap->parityInfo) {
    259      1.3   oster 							failedPDA->next->row = failedPDA->row;
    260      1.3   oster 							failedPDA->next->col = failedPDA->col;
    261      1.3   oster 						} else
    262      1.3   oster 							if (failedPDA == asmap->parityInfo->next) {	/* paranoid:  should
    263      1.3   oster 													 * never occur */
    264      1.3   oster 								asmap->parityInfo->row = failedPDA->row;
    265      1.3   oster 								asmap->parityInfo->col = failedPDA->col;
    266      1.3   oster 							}
    267      1.3   oster 					}
    268      1.3   oster 				}
    269      1.3   oster 
    270      1.3   oster 				RF_ASSERT(failedPDA->col != -1);
    271      1.3   oster 
    272      1.3   oster 				if (rf_dagDebug || rf_mapDebug) {
    273  1.3.8.1  bouyer 					printf("raid%d: Redirected type '%c' r %d c %d o %ld -> r %d c %d o %ld\n",
    274  1.3.8.1  bouyer 					       raidPtr->raidid, type, or, oc,
    275  1.3.8.1  bouyer 					       (long) oo, failedPDA->row,
    276  1.3.8.1  bouyer 					       failedPDA->col,
    277  1.3.8.1  bouyer 					       (long) failedPDA->startSector);
    278      1.3   oster 				}
    279      1.3   oster 				asmap->numDataFailed = asmap->numParityFailed = 0;
    280      1.3   oster 			}
    281      1.3   oster 		}
    282      1.3   oster 	/* all dags begin/end with block/unblock node therefore, hdrSucc &
    283      1.3   oster 	 * termAnt counts should always be 1 also, these counts should not be
    284      1.3   oster 	 * visible outside dag creation routines - manipulating the counts
    285      1.3   oster 	 * here should be removed */
    286      1.3   oster 	if (type == RF_IO_TYPE_READ) {
    287      1.3   oster 		if (asmap->numDataFailed == 0)
    288      1.3   oster 			*createFunc = (RF_VoidFuncPtr) rf_CreateFaultFreeReadDAG;
    289      1.3   oster 		else
    290      1.3   oster 			*createFunc = (RF_VoidFuncPtr) rf_CreateRaidFiveDegradedReadDAG;
    291      1.1   oster 	} else {
    292      1.1   oster 
    293      1.1   oster 
    294      1.3   oster 		/* if mirroring, always use large writes.  If the access
    295      1.3   oster 		 * requires two distinct parity updates, always do a small
    296      1.3   oster 		 * write.  If the stripe contains a failure but the access
    297      1.3   oster 		 * does not, do a small write. The first conditional
    298      1.3   oster 		 * (numStripeUnitsAccessed <= numDataCol/2) uses a
    299      1.3   oster 		 * less-than-or-equal rather than just a less-than because
    300      1.3   oster 		 * when G is 3 or 4, numDataCol/2 is 1, and I want
    301      1.3   oster 		 * single-stripe-unit updates to use just one disk. */
    302      1.3   oster 		if ((asmap->numDataFailed + asmap->numParityFailed) == 0) {
    303      1.3   oster 			if (rf_suppressLocksAndLargeWrites ||
    304      1.3   oster 			    (((asmap->numStripeUnitsAccessed <= (layoutPtr->numDataCol / 2)) && (layoutPtr->numDataCol != 1)) ||
    305      1.3   oster 				(asmap->parityInfo->next != NULL) || rf_CheckStripeForFailures(raidPtr, asmap))) {
    306      1.3   oster 				*createFunc = (RF_VoidFuncPtr) rf_CreateSmallWriteDAG;
    307      1.3   oster 			} else
    308      1.3   oster 				*createFunc = (RF_VoidFuncPtr) rf_CreateLargeWriteDAG;
    309      1.3   oster 		} else {
    310      1.3   oster 			if (asmap->numParityFailed == 1)
    311      1.3   oster 				*createFunc = (RF_VoidFuncPtr) rf_CreateNonRedundantWriteDAG;
    312      1.3   oster 			else
    313      1.3   oster 				if (asmap->numStripeUnitsAccessed != 1 && failedPDA->numSector != layoutPtr->sectorsPerStripeUnit)
    314      1.3   oster 					*createFunc = NULL;
    315      1.3   oster 				else
    316      1.3   oster 					*createFunc = (RF_VoidFuncPtr) rf_CreateDegradedWriteDAG;
    317      1.3   oster 		}
    318      1.1   oster 	}
    319      1.1   oster }
    320