Home | History | Annotate | Line # | Download | only in raidframe
rf_reconmap.c revision 1.16
      1  1.16   simonb /*	$NetBSD: rf_reconmap.c,v 1.16 2003/12/21 07:53:59 simonb 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  * rf_reconmap.c
     31   1.1    oster  *
     32   1.1    oster  * code to maintain a map of what sectors have/have not been reconstructed
     33   1.1    oster  *
     34   1.1    oster  *************************************************************************/
     35   1.8    lukem 
     36   1.8    lukem #include <sys/cdefs.h>
     37  1.16   simonb __KERNEL_RCSID(0, "$NetBSD: rf_reconmap.c,v 1.16 2003/12/21 07:53:59 simonb Exp $");
     38   1.1    oster 
     39   1.1    oster #include "rf_raid.h"
     40   1.1    oster #include <sys/time.h>
     41   1.1    oster #include "rf_general.h"
     42   1.1    oster #include "rf_utils.h"
     43   1.1    oster 
     44   1.1    oster /* special pointer values indicating that a reconstruction unit
     45   1.1    oster  * has been either totally reconstructed or not at all.  Both
     46   1.1    oster  * are illegal pointer values, so you have to be careful not to
     47   1.1    oster  * dereference through them.  RU_NOTHING must be zero, since
     48   1.7  thorpej  * MakeReconMap uses memset to initialize the structure.  These are used
     49   1.1    oster  * only at the head of the list.
     50   1.1    oster  */
     51   1.1    oster #define RU_ALL      ((RF_ReconMapListElem_t *) -1)
     52   1.1    oster #define RU_NOTHING  ((RF_ReconMapListElem_t *) 0)
     53   1.1    oster 
     54  1.15    oster /* For most reconstructs we need at most 3 RF_ReconMapListElem_t's.
     55  1.15    oster  * Bounding the number we need is quite difficult, as it depends on how
     56  1.15    oster  * badly the sectors to be reconstructed get divided up.  In the current
     57  1.15    oster  * code, the reconstructed sectors appeared aligned on stripe boundaries,
     58  1.15    oster  * and are always presented in stripe width units, so we're probably
     59  1.15    oster  * allocating quite a bit more than we'll ever need.
     60  1.15    oster  */
     61  1.15    oster #define RF_NUM_RECON_POOL_ELEM 100
     62  1.15    oster 
     63   1.4    oster static void
     64   1.4    oster compact_stat_entry(RF_Raid_t * raidPtr, RF_ReconMap_t * mapPtr,
     65   1.4    oster     int i);
     66  1.15    oster static void crunch_list(RF_ReconMap_t *mapPtr,
     67  1.15    oster 			RF_ReconMapListElem_t * listPtr);
     68   1.4    oster static RF_ReconMapListElem_t *
     69  1.15    oster MakeReconMapListElem(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector,
     70  1.15    oster 		     RF_SectorNum_t stopSector, RF_ReconMapListElem_t * next);
     71   1.4    oster static void
     72  1.15    oster FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t * p);
     73  1.10    oster #if 0
     74   1.4    oster static void PrintList(RF_ReconMapListElem_t * listPtr);
     75  1.10    oster #endif
     76   1.1    oster 
     77  1.12    oster /*---------------------------------------------------------------------------
     78   1.1    oster  *
     79   1.4    oster  * Creates and initializes new Reconstruction map
     80   1.1    oster  *
     81  1.12    oster  *-------------------------------------------------------------------------*/
     82   1.1    oster 
     83   1.4    oster RF_ReconMap_t *
     84   1.4    oster rf_MakeReconMap(raidPtr, ru_sectors, disk_sectors, spareUnitsPerDisk)
     85   1.4    oster 	RF_Raid_t *raidPtr;
     86   1.4    oster 	RF_SectorCount_t ru_sectors;	/* size of reconstruction unit in
     87   1.4    oster 					 * sectors */
     88   1.4    oster 	RF_SectorCount_t disk_sectors;	/* size of disk in sectors */
     89   1.4    oster 	RF_ReconUnitCount_t spareUnitsPerDisk;	/* zero unless distributed
     90   1.4    oster 						 * sparing */
     91   1.4    oster {
     92   1.4    oster 	RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
     93   1.4    oster 	RF_ReconUnitCount_t num_rus = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerRU;
     94   1.4    oster 	RF_ReconMap_t *p;
     95   1.4    oster 	int     rc;
     96   1.4    oster 
     97   1.4    oster 	RF_Malloc(p, sizeof(RF_ReconMap_t), (RF_ReconMap_t *));
     98   1.4    oster 	p->sectorsPerReconUnit = ru_sectors;
     99   1.4    oster 	p->sectorsInDisk = disk_sectors;
    100   1.4    oster 
    101   1.4    oster 	p->totalRUs = num_rus;
    102   1.4    oster 	p->spareRUs = spareUnitsPerDisk;
    103   1.4    oster 	p->unitsLeft = num_rus - spareUnitsPerDisk;
    104   1.4    oster 
    105   1.4    oster 	RF_Malloc(p->status, num_rus * sizeof(RF_ReconMapListElem_t *), (RF_ReconMapListElem_t **));
    106   1.4    oster 	RF_ASSERT(p->status != (RF_ReconMapListElem_t **) NULL);
    107   1.4    oster 
    108   1.7  thorpej 	(void) memset((char *) p->status, 0,
    109   1.7  thorpej 	    num_rus * sizeof(RF_ReconMapListElem_t *));
    110   1.4    oster 
    111  1.15    oster 
    112  1.15    oster 	pool_init(&p->elem_pool, sizeof(RF_ReconMapListElem_t), 0,
    113  1.16   simonb 	    0, 0, "raidreconpl", NULL);
    114  1.15    oster 
    115   1.4    oster 	rc = rf_mutex_init(&p->mutex);
    116   1.4    oster 	if (rc) {
    117   1.9    oster 		rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc);
    118   1.4    oster 		RF_Free(p->status, num_rus * sizeof(RF_ReconMapListElem_t *));
    119   1.4    oster 		RF_Free(p, sizeof(RF_ReconMap_t));
    120   1.4    oster 		return (NULL);
    121   1.4    oster 	}
    122   1.4    oster 	return (p);
    123   1.1    oster }
    124   1.1    oster 
    125   1.1    oster 
    126  1.12    oster /*---------------------------------------------------------------------------
    127   1.1    oster  *
    128  1.12    oster  * marks a new set of sectors as reconstructed.  All the possible
    129  1.12    oster  * mergings get complicated.  To simplify matters, the approach I take
    130  1.12    oster  * is to just dump something into the list, and then clean it up
    131  1.12    oster  * (i.e. merge elements and eliminate redundant ones) in a second pass
    132  1.12    oster  * over the list (compact_stat_entry()).  Not 100% efficient, since a
    133  1.12    oster  * structure can be allocated and then immediately freed, but it keeps
    134  1.12    oster  * this code from becoming (more of) a nightmare of special cases.
    135  1.12    oster  * The only thing that compact_stat_entry() assumes is that the list
    136  1.12    oster  * is sorted by startSector, and so this is the only condition I
    137  1.12    oster  * maintain here.  (MCH)
    138   1.1    oster  *
    139  1.15    oster  * This code now uses a pool instead of the previous malloc/free
    140  1.15    oster  * stuff.
    141  1.12    oster  *-------------------------------------------------------------------------*/
    142   1.1    oster 
    143   1.4    oster void
    144   1.4    oster rf_ReconMapUpdate(raidPtr, mapPtr, startSector, stopSector)
    145   1.4    oster 	RF_Raid_t *raidPtr;
    146   1.4    oster 	RF_ReconMap_t *mapPtr;
    147   1.4    oster 	RF_SectorNum_t startSector;
    148   1.4    oster 	RF_SectorNum_t stopSector;
    149   1.4    oster {
    150   1.4    oster 	RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
    151   1.4    oster 	RF_SectorNum_t i, first_in_RU, last_in_RU;
    152   1.4    oster 	RF_ReconMapListElem_t *p, *pt;
    153   1.4    oster 
    154   1.4    oster 	RF_LOCK_MUTEX(mapPtr->mutex);
    155  1.12    oster 	RF_ASSERT(startSector >= 0 && stopSector < mapPtr->sectorsInDisk &&
    156  1.12    oster 		  stopSector >= startSector);
    157   1.4    oster 
    158   1.4    oster 	while (startSector <= stopSector) {
    159   1.4    oster 		i = startSector / mapPtr->sectorsPerReconUnit;
    160   1.4    oster 		first_in_RU = i * sectorsPerReconUnit;
    161   1.4    oster 		last_in_RU = first_in_RU + sectorsPerReconUnit - 1;
    162   1.4    oster 		p = mapPtr->status[i];
    163   1.4    oster 		if (p != RU_ALL) {
    164  1.12    oster 			if (p == RU_NOTHING || p->startSector > startSector) {
    165  1.12    oster 				/* insert at front of list */
    166   1.4    oster 
    167  1.15    oster 				mapPtr->status[i] = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), (p == RU_NOTHING) ? NULL : p);
    168   1.4    oster 
    169   1.4    oster 			} else {/* general case */
    170   1.4    oster 				do {	/* search for place to insert */
    171   1.4    oster 					pt = p;
    172   1.4    oster 					p = p->next;
    173   1.4    oster 				} while (p && (p->startSector < startSector));
    174  1.15    oster 				pt->next = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), p);
    175  1.14    oster 
    176   1.4    oster 			}
    177   1.4    oster 			compact_stat_entry(raidPtr, mapPtr, i);
    178   1.4    oster 		}
    179   1.4    oster 		startSector = RF_MIN(stopSector, last_in_RU) + 1;
    180   1.4    oster 	}
    181   1.4    oster 	RF_UNLOCK_MUTEX(mapPtr->mutex);
    182   1.1    oster }
    183   1.1    oster 
    184   1.1    oster 
    185   1.1    oster 
    186  1.12    oster /*---------------------------------------------------------------------------
    187   1.1    oster  *
    188   1.1    oster  * performs whatever list compactions can be done, and frees any space
    189   1.1    oster  * that is no longer necessary.  Assumes only that the list is sorted
    190  1.12    oster  * by startSector.  crunch_list() compacts a single list as much as
    191  1.12    oster  * possible, and the second block of code deletes the entire list if
    192  1.12    oster  * possible.  crunch_list() is also called from
    193  1.12    oster  * MakeReconMapAccessList().
    194   1.1    oster  *
    195   1.1    oster  * When a recon unit is detected to be fully reconstructed, we set the
    196   1.1    oster  * corresponding bit in the parity stripe map so that the head follow
    197  1.12    oster  * code will not select this parity stripe again.  This is redundant
    198  1.12    oster  * (but harmless) when compact_stat_entry is called from the
    199  1.12    oster  * reconstruction code, but necessary when called from the user-write
    200  1.12    oster  * code.
    201   1.1    oster  *
    202  1.12    oster  *-------------------------------------------------------------------------*/
    203   1.1    oster 
    204   1.4    oster static void
    205   1.4    oster compact_stat_entry(raidPtr, mapPtr, i)
    206   1.4    oster 	RF_Raid_t *raidPtr;
    207   1.4    oster 	RF_ReconMap_t *mapPtr;
    208   1.4    oster 	int     i;
    209   1.4    oster {
    210   1.4    oster 	RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
    211   1.4    oster 	RF_ReconMapListElem_t *p = mapPtr->status[i];
    212   1.4    oster 
    213  1.15    oster 	crunch_list(mapPtr, p);
    214   1.4    oster 
    215   1.4    oster 	if ((p->startSector == i * sectorsPerReconUnit) &&
    216  1.12    oster 	    (p->stopSector == i * sectorsPerReconUnit +
    217  1.12    oster 			      sectorsPerReconUnit - 1)) {
    218   1.4    oster 		mapPtr->status[i] = RU_ALL;
    219   1.4    oster 		mapPtr->unitsLeft--;
    220  1.15    oster 		FreeReconMapListElem(mapPtr, p);
    221   1.4    oster 	}
    222   1.4    oster }
    223   1.4    oster 
    224   1.4    oster static void
    225  1.15    oster crunch_list(mapPtr, listPtr)
    226  1.15    oster 	RF_ReconMap_t *mapPtr;
    227   1.4    oster 	RF_ReconMapListElem_t *listPtr;
    228   1.4    oster {
    229   1.4    oster 	RF_ReconMapListElem_t *pt, *p = listPtr;
    230   1.4    oster 
    231   1.4    oster 	if (!p)
    232   1.4    oster 		return;
    233   1.4    oster 	pt = p;
    234   1.4    oster 	p = p->next;
    235   1.4    oster 	while (p) {
    236   1.4    oster 		if (pt->stopSector >= p->startSector - 1) {
    237   1.4    oster 			pt->stopSector = RF_MAX(pt->stopSector, p->stopSector);
    238   1.4    oster 			pt->next = p->next;
    239  1.15    oster 			FreeReconMapListElem(mapPtr, p);
    240   1.4    oster 			p = pt->next;
    241   1.4    oster 		} else {
    242   1.4    oster 			pt = p;
    243   1.4    oster 			p = p->next;
    244   1.4    oster 		}
    245   1.4    oster 	}
    246   1.1    oster }
    247  1.12    oster /*---------------------------------------------------------------------------
    248   1.4    oster  *
    249   1.1    oster  * Allocate and fill a new list element
    250   1.1    oster  *
    251  1.12    oster  *-------------------------------------------------------------------------*/
    252   1.1    oster 
    253   1.4    oster static RF_ReconMapListElem_t *
    254   1.4    oster MakeReconMapListElem(
    255  1.15    oster     RF_ReconMap_t *mapPtr,
    256   1.4    oster     RF_SectorNum_t startSector,
    257   1.4    oster     RF_SectorNum_t stopSector,
    258   1.4    oster     RF_ReconMapListElem_t * next)
    259   1.4    oster {
    260   1.4    oster 	RF_ReconMapListElem_t *p;
    261   1.4    oster 
    262  1.15    oster 	p = pool_get(&mapPtr->elem_pool, PR_NOWAIT);
    263  1.15    oster 
    264   1.4    oster 	if (p == NULL)
    265   1.4    oster 		return (NULL);
    266  1.15    oster 
    267   1.4    oster 	p->startSector = startSector;
    268   1.4    oster 	p->stopSector = stopSector;
    269   1.4    oster 	p->next = next;
    270   1.4    oster 	return (p);
    271   1.1    oster }
    272  1.12    oster /*---------------------------------------------------------------------------
    273   1.4    oster  *
    274   1.1    oster  * Free a list element
    275   1.1    oster  *
    276  1.12    oster  *-------------------------------------------------------------------------*/
    277   1.1    oster 
    278   1.4    oster static void
    279  1.15    oster FreeReconMapListElem(mapPtr, p)
    280  1.15    oster 	RF_ReconMap_t *mapPtr;
    281   1.4    oster 	RF_ReconMapListElem_t *p;
    282   1.4    oster {
    283  1.15    oster 	pool_put(&mapPtr->elem_pool, p);
    284   1.1    oster }
    285  1.12    oster /*---------------------------------------------------------------------------
    286   1.4    oster  *
    287  1.12    oster  * Free an entire status structure.  Inefficient, but can be called at
    288  1.12    oster  * any time.
    289   1.1    oster  *
    290  1.12    oster  *-------------------------------------------------------------------------*/
    291   1.4    oster void
    292   1.4    oster rf_FreeReconMap(mapPtr)
    293   1.4    oster 	RF_ReconMap_t *mapPtr;
    294   1.4    oster {
    295   1.4    oster 	RF_ReconMapListElem_t *p, *q;
    296   1.4    oster 	RF_ReconUnitCount_t numRUs;
    297   1.4    oster 	RF_ReconUnitNum_t i;
    298   1.4    oster 
    299   1.4    oster 	numRUs = mapPtr->sectorsInDisk / mapPtr->sectorsPerReconUnit;
    300   1.4    oster 	if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
    301   1.4    oster 		numRUs++;
    302   1.4    oster 
    303   1.4    oster 	for (i = 0; i < numRUs; i++) {
    304   1.4    oster 		p = mapPtr->status[i];
    305   1.4    oster 		while (p != RU_NOTHING && p != RU_ALL) {
    306   1.4    oster 			q = p;
    307   1.4    oster 			p = p->next;
    308   1.4    oster 			RF_Free(q, sizeof(*q));
    309   1.4    oster 		}
    310   1.4    oster 	}
    311  1.15    oster 	pool_destroy(&mapPtr->elem_pool);
    312   1.4    oster 	rf_mutex_destroy(&mapPtr->mutex);
    313  1.12    oster 	RF_Free(mapPtr->status, mapPtr->totalRUs *
    314  1.12    oster 		sizeof(RF_ReconMapListElem_t *));
    315   1.4    oster 	RF_Free(mapPtr, sizeof(RF_ReconMap_t));
    316   1.1    oster }
    317  1.12    oster /*---------------------------------------------------------------------------
    318   1.1    oster  *
    319   1.1    oster  * returns nonzero if the indicated RU has been reconstructed already
    320   1.1    oster  *
    321  1.12    oster  *-------------------------------------------------------------------------*/
    322   1.1    oster 
    323   1.4    oster int
    324   1.4    oster rf_CheckRUReconstructed(mapPtr, startSector)
    325   1.4    oster 	RF_ReconMap_t *mapPtr;
    326   1.4    oster 	RF_SectorNum_t startSector;
    327   1.1    oster {
    328   1.4    oster 	RF_ReconMapListElem_t *l;	/* used for searching */
    329   1.4    oster 	RF_ReconUnitNum_t i;
    330   1.1    oster 
    331   1.4    oster 	i = startSector / mapPtr->sectorsPerReconUnit;
    332   1.4    oster 	l = mapPtr->status[i];
    333   1.4    oster 	return ((l == RU_ALL) ? 1 : 0);
    334   1.1    oster }
    335   1.1    oster 
    336   1.4    oster RF_ReconUnitCount_t
    337   1.4    oster rf_UnitsLeftToReconstruct(mapPtr)
    338   1.4    oster 	RF_ReconMap_t *mapPtr;
    339   1.1    oster {
    340   1.4    oster 	RF_ASSERT(mapPtr != NULL);
    341   1.4    oster 	return (mapPtr->unitsLeft);
    342   1.4    oster }
    343   1.4    oster 
    344  1.10    oster #if 0
    345   1.4    oster static void
    346   1.4    oster PrintList(listPtr)
    347   1.4    oster 	RF_ReconMapListElem_t *listPtr;
    348   1.4    oster {
    349   1.4    oster 	while (listPtr) {
    350  1.12    oster 		printf("%d,%d -> ", (int) listPtr->startSector,
    351  1.12    oster 		       (int) listPtr->stopSector);
    352   1.4    oster 		listPtr = listPtr->next;
    353   1.4    oster 	}
    354   1.4    oster 	printf("\n");
    355   1.4    oster }
    356   1.4    oster 
    357   1.4    oster void
    358   1.4    oster rf_PrintReconMap(raidPtr, mapPtr, frow, fcol)
    359   1.4    oster 	RF_Raid_t *raidPtr;
    360   1.4    oster 	RF_ReconMap_t *mapPtr;
    361   1.4    oster 	RF_RowCol_t frow;
    362   1.4    oster 	RF_RowCol_t fcol;
    363   1.4    oster {
    364   1.4    oster 	RF_ReconUnitCount_t numRUs;
    365   1.4    oster 	RF_ReconMapListElem_t *p;
    366   1.4    oster 	RF_ReconUnitNum_t i;
    367   1.4    oster 
    368   1.4    oster 	numRUs = mapPtr->totalRUs;
    369   1.4    oster 	if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
    370   1.4    oster 		numRUs++;
    371   1.4    oster 
    372   1.4    oster 	for (i = 0; i < numRUs; i++) {
    373   1.4    oster 		p = mapPtr->status[i];
    374   1.4    oster 		if (p == RU_ALL)/* printf("[%d] ALL\n",i) */
    375   1.4    oster 			;
    376   1.4    oster 		else
    377   1.4    oster 			if (p == RU_NOTHING) {
    378   1.4    oster 				printf("%d: Unreconstructed\n", i);
    379   1.4    oster 			} else {
    380   1.4    oster 				printf("%d: ", i);
    381   1.4    oster 				PrintList(p);
    382   1.4    oster 			}
    383   1.4    oster 	}
    384   1.4    oster }
    385  1.10    oster #endif
    386   1.4    oster 
    387  1.11    oster #if RF_DEBUG_RECON
    388   1.4    oster void
    389   1.4    oster rf_PrintReconSchedule(mapPtr, starttime)
    390   1.4    oster 	RF_ReconMap_t *mapPtr;
    391   1.4    oster 	struct timeval *starttime;
    392   1.4    oster {
    393   1.4    oster 	static int old_pctg = -1;
    394   1.4    oster 	struct timeval tv, diff;
    395   1.4    oster 	int     new_pctg;
    396   1.4    oster 
    397  1.12    oster 	new_pctg = 100 - (rf_UnitsLeftToReconstruct(mapPtr) *
    398  1.12    oster 			  100 / mapPtr->totalRUs);
    399   1.4    oster 	if (new_pctg != old_pctg) {
    400   1.4    oster 		RF_GETTIME(tv);
    401   1.4    oster 		RF_TIMEVAL_DIFF(starttime, &tv, &diff);
    402  1.12    oster 		printf("%d %d.%06d\n", (int) new_pctg, (int) diff.tv_sec,
    403  1.12    oster 		       (int) diff.tv_usec);
    404   1.4    oster 		old_pctg = new_pctg;
    405   1.4    oster 	}
    406   1.1    oster }
    407  1.11    oster #endif
    408  1.11    oster 
    409