Home | History | Annotate | Line # | Download | only in raidframe
rf_raid5.c revision 1.6
      1 /*	$NetBSD: rf_raid5.c,v 1.6 2001/11/13 07:11:16 lukem Exp $	*/
      2 /*
      3  * Copyright (c) 1995 Carnegie-Mellon University.
      4  * All rights reserved.
      5  *
      6  * Author: Mark Holland
      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 /******************************************************************************
     30  *
     31  * rf_raid5.c -- implements RAID Level 5
     32  *
     33  *****************************************************************************/
     34 
     35 #include <sys/cdefs.h>
     36 __KERNEL_RCSID(0, "$NetBSD: rf_raid5.c,v 1.6 2001/11/13 07:11:16 lukem Exp $");
     37 
     38 #include <dev/raidframe/raidframevar.h>
     39 
     40 #include "rf_raid.h"
     41 #include "rf_raid5.h"
     42 #include "rf_dag.h"
     43 #include "rf_dagffrd.h"
     44 #include "rf_dagffwr.h"
     45 #include "rf_dagdegrd.h"
     46 #include "rf_dagdegwr.h"
     47 #include "rf_dagutils.h"
     48 #include "rf_general.h"
     49 #include "rf_map.h"
     50 #include "rf_utils.h"
     51 
     52 typedef struct RF_Raid5ConfigInfo_s {
     53 	RF_RowCol_t **stripeIdentifier;	/* filled in at config time and used
     54 					 * by IdentifyStripe */
     55 }       RF_Raid5ConfigInfo_t;
     56 
     57 int
     58 rf_ConfigureRAID5(
     59     RF_ShutdownList_t ** listp,
     60     RF_Raid_t * raidPtr,
     61     RF_Config_t * cfgPtr)
     62 {
     63 	RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
     64 	RF_Raid5ConfigInfo_t *info;
     65 	RF_RowCol_t i, j, startdisk;
     66 
     67 	/* create a RAID level 5 configuration structure */
     68 	RF_MallocAndAdd(info, sizeof(RF_Raid5ConfigInfo_t), (RF_Raid5ConfigInfo_t *), raidPtr->cleanupList);
     69 	if (info == NULL)
     70 		return (ENOMEM);
     71 	layoutPtr->layoutSpecificInfo = (void *) info;
     72 
     73 	RF_ASSERT(raidPtr->numRow == 1);
     74 
     75 	/* the stripe identifier must identify the disks in each stripe, IN
     76 	 * THE ORDER THAT THEY APPEAR IN THE STRIPE. */
     77 	info->stripeIdentifier = rf_make_2d_array(raidPtr->numCol, raidPtr->numCol, raidPtr->cleanupList);
     78 	if (info->stripeIdentifier == NULL)
     79 		return (ENOMEM);
     80 	startdisk = 0;
     81 	for (i = 0; i < raidPtr->numCol; i++) {
     82 		for (j = 0; j < raidPtr->numCol; j++) {
     83 			info->stripeIdentifier[i][j] = (startdisk + j) % raidPtr->numCol;
     84 		}
     85 		if ((--startdisk) < 0)
     86 			startdisk = raidPtr->numCol - 1;
     87 	}
     88 
     89 	/* fill in the remaining layout parameters */
     90 	layoutPtr->numStripe = layoutPtr->stripeUnitsPerDisk;
     91 	layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector;
     92 	layoutPtr->numDataCol = raidPtr->numCol - 1;
     93 	layoutPtr->dataSectorsPerStripe = layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
     94 	layoutPtr->numParityCol = 1;
     95 	layoutPtr->dataStripeUnitsPerDisk = layoutPtr->stripeUnitsPerDisk;
     96 
     97 	raidPtr->totalSectors = layoutPtr->stripeUnitsPerDisk * layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
     98 
     99 	return (0);
    100 }
    101 
    102 int
    103 rf_GetDefaultNumFloatingReconBuffersRAID5(RF_Raid_t * raidPtr)
    104 {
    105 	return (20);
    106 }
    107 
    108 RF_HeadSepLimit_t
    109 rf_GetDefaultHeadSepLimitRAID5(RF_Raid_t * raidPtr)
    110 {
    111 	return (10);
    112 }
    113 #if !defined(__NetBSD__) && !defined(_KERNEL)
    114 /* not currently used */
    115 int
    116 rf_ShutdownRAID5(RF_Raid_t * raidPtr)
    117 {
    118 	return (0);
    119 }
    120 #endif
    121 
    122 void
    123 rf_MapSectorRAID5(
    124     RF_Raid_t * raidPtr,
    125     RF_RaidAddr_t raidSector,
    126     RF_RowCol_t * row,
    127     RF_RowCol_t * col,
    128     RF_SectorNum_t * diskSector,
    129     int remap)
    130 {
    131 	RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
    132 	*row = 0;
    133 	*col = (SUID % raidPtr->numCol);
    134 	*diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
    135 	    (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
    136 }
    137 
    138 void
    139 rf_MapParityRAID5(
    140     RF_Raid_t * raidPtr,
    141     RF_RaidAddr_t raidSector,
    142     RF_RowCol_t * row,
    143     RF_RowCol_t * col,
    144     RF_SectorNum_t * diskSector,
    145     int remap)
    146 {
    147 	RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
    148 
    149 	*row = 0;
    150 	*col = raidPtr->Layout.numDataCol - (SUID / raidPtr->Layout.numDataCol) % raidPtr->numCol;
    151 	*diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
    152 	    (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
    153 }
    154 
    155 void
    156 rf_IdentifyStripeRAID5(
    157     RF_Raid_t * raidPtr,
    158     RF_RaidAddr_t addr,
    159     RF_RowCol_t ** diskids,
    160     RF_RowCol_t * outRow)
    161 {
    162 	RF_StripeNum_t stripeID = rf_RaidAddressToStripeID(&raidPtr->Layout, addr);
    163 	RF_Raid5ConfigInfo_t *info = (RF_Raid5ConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo;
    164 
    165 	*outRow = 0;
    166 	*diskids = info->stripeIdentifier[stripeID % raidPtr->numCol];
    167 }
    168 
    169 void
    170 rf_MapSIDToPSIDRAID5(
    171     RF_RaidLayout_t * layoutPtr,
    172     RF_StripeNum_t stripeID,
    173     RF_StripeNum_t * psID,
    174     RF_ReconUnitNum_t * which_ru)
    175 {
    176 	*which_ru = 0;
    177 	*psID = stripeID;
    178 }
    179 /* select an algorithm for performing an access.  Returns two pointers,
    180  * one to a function that will return information about the DAG, and
    181  * another to a function that will create the dag.
    182  */
    183 void
    184 rf_RaidFiveDagSelect(
    185     RF_Raid_t * raidPtr,
    186     RF_IoType_t type,
    187     RF_AccessStripeMap_t * asmap,
    188     RF_VoidFuncPtr * createFunc)
    189 {
    190 	RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
    191 	RF_PhysDiskAddr_t *failedPDA = NULL;
    192 	RF_RowCol_t frow, fcol;
    193 	RF_RowStatus_t rstat;
    194 	int     prior_recon;
    195 
    196 	RF_ASSERT(RF_IO_IS_R_OR_W(type));
    197 
    198 	if (asmap->numDataFailed + asmap->numParityFailed > 1) {
    199 		RF_ERRORMSG("Multiple disks failed in a single group!  Aborting I/O operation.\n");
    200 		 /* *infoFunc = */ *createFunc = NULL;
    201 		return;
    202 	} else
    203 		if (asmap->numDataFailed + asmap->numParityFailed == 1) {
    204 
    205 			/* if under recon & already reconstructed, redirect
    206 			 * the access to the spare drive and eliminate the
    207 			 * failure indication */
    208 			failedPDA = asmap->failedPDAs[0];
    209 			frow = failedPDA->row;
    210 			fcol = failedPDA->col;
    211 			rstat = raidPtr->status[failedPDA->row];
    212 			prior_recon = (rstat == rf_rs_reconfigured) || (
    213 			    (rstat == rf_rs_reconstructing) ?
    214 			    rf_CheckRUReconstructed(raidPtr->reconControl[frow]->reconMap, failedPDA->startSector) : 0
    215 			    );
    216 			if (prior_recon) {
    217 				RF_RowCol_t or = failedPDA->row, oc = failedPDA->col;
    218 				RF_SectorNum_t oo = failedPDA->startSector;
    219 
    220 				if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) {	/* redirect to dist
    221 											 * spare space */
    222 
    223 					if (failedPDA == asmap->parityInfo) {
    224 
    225 						/* parity has failed */
    226 						(layoutPtr->map->MapParity) (raidPtr, failedPDA->raidAddress, &failedPDA->row,
    227 						    &failedPDA->col, &failedPDA->startSector, RF_REMAP);
    228 
    229 						if (asmap->parityInfo->next) {	/* redir 2nd component,
    230 										 * if any */
    231 							RF_PhysDiskAddr_t *p = asmap->parityInfo->next;
    232 							RF_SectorNum_t SUoffs = p->startSector % layoutPtr->sectorsPerStripeUnit;
    233 							p->row = failedPDA->row;
    234 							p->col = failedPDA->col;
    235 							p->startSector = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->startSector) +
    236 							    SUoffs;	/* cheating:
    237 									 * startSector is not
    238 									 * really a RAID address */
    239 						}
    240 					} else
    241 						if (asmap->parityInfo->next && failedPDA == asmap->parityInfo->next) {
    242 							RF_ASSERT(0);	/* should not ever
    243 									 * happen */
    244 						} else {
    245 
    246 							/* data has failed */
    247 							(layoutPtr->map->MapSector) (raidPtr, failedPDA->raidAddress, &failedPDA->row,
    248 							    &failedPDA->col, &failedPDA->startSector, RF_REMAP);
    249 
    250 						}
    251 
    252 				} else {	/* redirect to dedicated spare
    253 						 * space */
    254 
    255 					failedPDA->row = raidPtr->Disks[frow][fcol].spareRow;
    256 					failedPDA->col = raidPtr->Disks[frow][fcol].spareCol;
    257 
    258 					/* the parity may have two distinct
    259 					 * components, both of which may need
    260 					 * to be redirected */
    261 					if (asmap->parityInfo->next) {
    262 						if (failedPDA == asmap->parityInfo) {
    263 							failedPDA->next->row = failedPDA->row;
    264 							failedPDA->next->col = failedPDA->col;
    265 						} else
    266 							if (failedPDA == asmap->parityInfo->next) {	/* paranoid:  should
    267 													 * never occur */
    268 								asmap->parityInfo->row = failedPDA->row;
    269 								asmap->parityInfo->col = failedPDA->col;
    270 							}
    271 					}
    272 				}
    273 
    274 				RF_ASSERT(failedPDA->col != -1);
    275 
    276 				if (rf_dagDebug || rf_mapDebug) {
    277 					printf("raid%d: Redirected type '%c' r %d c %d o %ld -> r %d c %d o %ld\n",
    278 					       raidPtr->raidid, type, or, oc,
    279 					       (long) oo, failedPDA->row,
    280 					       failedPDA->col,
    281 					       (long) failedPDA->startSector);
    282 				}
    283 				asmap->numDataFailed = asmap->numParityFailed = 0;
    284 			}
    285 		}
    286 	/* all dags begin/end with block/unblock node therefore, hdrSucc &
    287 	 * termAnt counts should always be 1 also, these counts should not be
    288 	 * visible outside dag creation routines - manipulating the counts
    289 	 * here should be removed */
    290 	if (type == RF_IO_TYPE_READ) {
    291 		if (asmap->numDataFailed == 0)
    292 			*createFunc = (RF_VoidFuncPtr) rf_CreateFaultFreeReadDAG;
    293 		else
    294 			*createFunc = (RF_VoidFuncPtr) rf_CreateRaidFiveDegradedReadDAG;
    295 	} else {
    296 
    297 
    298 		/* if mirroring, always use large writes.  If the access
    299 		 * requires two distinct parity updates, always do a small
    300 		 * write.  If the stripe contains a failure but the access
    301 		 * does not, do a small write. The first conditional
    302 		 * (numStripeUnitsAccessed <= numDataCol/2) uses a
    303 		 * less-than-or-equal rather than just a less-than because
    304 		 * when G is 3 or 4, numDataCol/2 is 1, and I want
    305 		 * single-stripe-unit updates to use just one disk. */
    306 		if ((asmap->numDataFailed + asmap->numParityFailed) == 0) {
    307 			if (rf_suppressLocksAndLargeWrites ||
    308 			    (((asmap->numStripeUnitsAccessed <= (layoutPtr->numDataCol / 2)) && (layoutPtr->numDataCol != 1)) ||
    309 				(asmap->parityInfo->next != NULL) || rf_CheckStripeForFailures(raidPtr, asmap))) {
    310 				*createFunc = (RF_VoidFuncPtr) rf_CreateSmallWriteDAG;
    311 			} else
    312 				*createFunc = (RF_VoidFuncPtr) rf_CreateLargeWriteDAG;
    313 		} else {
    314 			if (asmap->numParityFailed == 1)
    315 				*createFunc = (RF_VoidFuncPtr) rf_CreateNonRedundantWriteDAG;
    316 			else
    317 				if (asmap->numStripeUnitsAccessed != 1 && failedPDA->numSector != layoutPtr->sectorsPerStripeUnit)
    318 					*createFunc = NULL;
    319 				else
    320 					*createFunc = (RF_VoidFuncPtr) rf_CreateDegradedWriteDAG;
    321 		}
    322 	}
    323 }
    324