Home | History | Annotate | Line # | Download | only in raidframe
      1  1.39    andvar /*	$NetBSD: rf_reconmap.c,v 1.39 2022/04/08 10:27:04 andvar 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  1.39    andvar __KERNEL_RCSID(0, "$NetBSD: rf_reconmap.c,v 1.39 2022/04/08 10:27:04 andvar 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  1.26     perry  * code, the reconstructed sectors appeared aligned on stripe boundaries,
     58  1.26     perry  * 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  1.26     perry static void
     64  1.31     oster compact_stat_entry(RF_Raid_t *, RF_ReconMap_t *, int, int);
     65  1.31     oster static void crunch_list(RF_ReconMap_t *, RF_ReconMapListElem_t *);
     66   1.4     oster static RF_ReconMapListElem_t *
     67  1.31     oster MakeReconMapListElem(RF_ReconMap_t *, RF_SectorNum_t, RF_SectorNum_t,
     68  1.31     oster 		     RF_ReconMapListElem_t *);
     69  1.26     perry static void
     70  1.15     oster FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t * p);
     71   1.1     oster 
     72  1.12     oster /*---------------------------------------------------------------------------
     73   1.1     oster  *
     74   1.4     oster  * Creates and initializes new Reconstruction map
     75   1.1     oster  *
     76  1.22     oster  * ru_sectors   - size of reconstruction unit in sectors
     77  1.22     oster  * disk_sectors - size of disk in sectors
     78  1.22     oster  * spareUnitsPerDisk - zero unless distributed sparing
     79  1.12     oster  *-------------------------------------------------------------------------*/
     80   1.1     oster 
     81   1.4     oster RF_ReconMap_t *
     82  1.26     perry rf_MakeReconMap(RF_Raid_t *raidPtr, RF_SectorCount_t ru_sectors,
     83  1.22     oster 		RF_SectorCount_t disk_sectors,
     84  1.22     oster 		RF_ReconUnitCount_t spareUnitsPerDisk)
     85   1.4     oster {
     86   1.4     oster 	RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
     87   1.4     oster 	RF_ReconUnitCount_t num_rus = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerRU;
     88   1.4     oster 	RF_ReconMap_t *p;
     89   1.4     oster 
     90  1.37  christos 	p = RF_Malloc(sizeof(*p));
     91   1.4     oster 	p->sectorsPerReconUnit = ru_sectors;
     92   1.4     oster 	p->sectorsInDisk = disk_sectors;
     93   1.4     oster 
     94   1.4     oster 	p->totalRUs = num_rus;
     95   1.4     oster 	p->spareRUs = spareUnitsPerDisk;
     96   1.4     oster 	p->unitsLeft = num_rus - spareUnitsPerDisk;
     97  1.31     oster 	p->low_ru = 0;
     98  1.31     oster 	p->status_size = RF_RECONMAP_SIZE;
     99  1.31     oster 	p->high_ru = p->status_size - 1;
    100  1.31     oster 	p->head = 0;
    101   1.4     oster 
    102  1.37  christos 	p->status = RF_Malloc(p->status_size * sizeof(*p->status));
    103  1.33    plunky 	RF_ASSERT(p->status != NULL);
    104   1.4     oster 
    105  1.15     oster 	pool_init(&p->elem_pool, sizeof(RF_ReconMapListElem_t), 0,
    106  1.30        ad 	    0, 0, "raidreconpl", NULL, IPL_BIO);
    107  1.38       chs 	pool_prime(&p->elem_pool, RF_NUM_RECON_POOL_ELEM);
    108  1.15     oster 
    109  1.32       mrg 	rf_init_mutex2(p->mutex, IPL_VM);
    110  1.32       mrg 	rf_init_cond2(p->cv, "reconupdate");
    111  1.32       mrg 
    112   1.4     oster 	return (p);
    113   1.1     oster }
    114   1.1     oster 
    115   1.1     oster 
    116  1.12     oster /*---------------------------------------------------------------------------
    117   1.1     oster  *
    118  1.12     oster  * marks a new set of sectors as reconstructed.  All the possible
    119  1.12     oster  * mergings get complicated.  To simplify matters, the approach I take
    120  1.12     oster  * is to just dump something into the list, and then clean it up
    121  1.12     oster  * (i.e. merge elements and eliminate redundant ones) in a second pass
    122  1.12     oster  * over the list (compact_stat_entry()).  Not 100% efficient, since a
    123  1.12     oster  * structure can be allocated and then immediately freed, but it keeps
    124  1.12     oster  * this code from becoming (more of) a nightmare of special cases.
    125  1.12     oster  * The only thing that compact_stat_entry() assumes is that the list
    126  1.12     oster  * is sorted by startSector, and so this is the only condition I
    127  1.12     oster  * maintain here.  (MCH)
    128   1.1     oster  *
    129  1.15     oster  * This code now uses a pool instead of the previous malloc/free
    130  1.15     oster  * stuff.
    131  1.12     oster  *-------------------------------------------------------------------------*/
    132   1.1     oster 
    133  1.26     perry void
    134  1.22     oster rf_ReconMapUpdate(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr,
    135  1.22     oster 		  RF_SectorNum_t startSector, RF_SectorNum_t stopSector)
    136   1.4     oster {
    137   1.4     oster 	RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
    138  1.31     oster 	RF_SectorNum_t i, first_in_RU, last_in_RU, ru;
    139   1.4     oster 	RF_ReconMapListElem_t *p, *pt;
    140   1.4     oster 
    141  1.32       mrg 	rf_lock_mutex2(mapPtr->mutex);
    142  1.24     oster 	while(mapPtr->lock) {
    143  1.32       mrg 		rf_wait_cond2(mapPtr->cv, mapPtr->mutex);
    144  1.24     oster 	}
    145  1.24     oster 	mapPtr->lock = 1;
    146  1.32       mrg 	rf_unlock_mutex2(mapPtr->mutex);
    147  1.26     perry 	RF_ASSERT(startSector >= 0 && stopSector < mapPtr->sectorsInDisk &&
    148  1.12     oster 		  stopSector >= startSector);
    149   1.4     oster 
    150   1.4     oster 	while (startSector <= stopSector) {
    151   1.4     oster 		i = startSector / mapPtr->sectorsPerReconUnit;
    152   1.4     oster 		first_in_RU = i * sectorsPerReconUnit;
    153   1.4     oster 		last_in_RU = first_in_RU + sectorsPerReconUnit - 1;
    154  1.31     oster 
    155  1.31     oster 		/* do we need to move the queue? */
    156  1.31     oster 		while (i > mapPtr->high_ru) {
    157  1.34     oster #if 0
    158  1.31     oster #ifdef DIAGNOSTIC
    159  1.34     oster 			/* XXX: The check below is not valid for
    160  1.34     oster 			 * RAID5_RS.  It is valid for RAID 1 and RAID 5.
    161  1.34     oster 			 * The issue is that we can easily have
    162  1.34     oster 			 * RU_NOTHING entries here too, and those are
    163  1.34     oster 			 * quite correct.
    164  1.34     oster 			 */
    165  1.31     oster 			if (mapPtr->status[mapPtr->head]!=RU_ALL) {
    166  1.31     oster 				printf("\nraid%d: reconmap incorrect -- working on i %" PRIu64 "\n",
    167  1.31     oster 				       raidPtr->raidid, i);
    168  1.31     oster 				printf("raid%d: ru %" PRIu64 " not completed!!!\n",
    169  1.31     oster 				       raidPtr->raidid, mapPtr->head);
    170  1.31     oster 
    171  1.31     oster 				printf("raid%d: low: %" PRIu64 " high: %" PRIu64 "\n",
    172  1.31     oster 				       raidPtr->raidid, mapPtr->low_ru, mapPtr->high_ru);
    173  1.31     oster 
    174  1.31     oster 				panic("reconmap incorrect");
    175  1.31     oster 			}
    176  1.31     oster #endif
    177  1.34     oster #endif
    178  1.31     oster 			mapPtr->low_ru++;
    179  1.31     oster 			mapPtr->high_ru++;
    180  1.31     oster 			/* initialize "highest" RU status entry, which
    181  1.39    andvar 			   will take over the current head position */
    182  1.31     oster 			mapPtr->status[mapPtr->head]=RU_NOTHING;
    183  1.31     oster 
    184  1.31     oster 			/* move head too */
    185  1.31     oster 			mapPtr->head++;
    186  1.31     oster 			if (mapPtr->head >= mapPtr->status_size)
    187  1.31     oster 				mapPtr->head = 0;
    188  1.31     oster 
    189  1.31     oster 		}
    190  1.31     oster 
    191  1.31     oster 		ru = i - mapPtr->low_ru + mapPtr->head;
    192  1.31     oster 		if (ru >= mapPtr->status_size)
    193  1.31     oster 			ru = ru - mapPtr->status_size;
    194  1.31     oster 
    195  1.31     oster 		if ((ru < 0) || (ru >= mapPtr->status_size)) {
    196  1.31     oster 			printf("raid%d: ru is bogus %" PRIu64 "%" PRIu64 "%" PRIu64 "%" PRIu64 "%" PRIu64 "\n",
    197  1.31     oster 			       raidPtr->raidid, i, ru, mapPtr->head, mapPtr->low_ru, mapPtr->high_ru);
    198  1.31     oster 			panic("bogus ru in reconmap");
    199  1.31     oster 		}
    200  1.31     oster 
    201  1.31     oster 		p = mapPtr->status[ru];
    202   1.4     oster 		if (p != RU_ALL) {
    203  1.12     oster 			if (p == RU_NOTHING || p->startSector > startSector) {
    204  1.12     oster 				/* insert at front of list */
    205   1.4     oster 
    206  1.31     oster 				mapPtr->status[ru] = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), (p == RU_NOTHING) ? NULL : p);
    207   1.4     oster 
    208   1.4     oster 			} else {/* general case */
    209   1.4     oster 				do {	/* search for place to insert */
    210   1.4     oster 					pt = p;
    211   1.4     oster 					p = p->next;
    212   1.4     oster 				} while (p && (p->startSector < startSector));
    213  1.15     oster 				pt->next = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), p);
    214  1.14     oster 
    215   1.4     oster 			}
    216  1.31     oster 			compact_stat_entry(raidPtr, mapPtr, i, ru);
    217   1.4     oster 		}
    218   1.4     oster 		startSector = RF_MIN(stopSector, last_in_RU) + 1;
    219   1.4     oster 	}
    220  1.32       mrg 	rf_lock_mutex2(mapPtr->mutex);
    221  1.24     oster 	mapPtr->lock = 0;
    222  1.32       mrg 	rf_broadcast_cond2(mapPtr->cv);
    223  1.32       mrg 	rf_unlock_mutex2(mapPtr->mutex);
    224   1.1     oster }
    225   1.1     oster 
    226   1.1     oster 
    227   1.1     oster 
    228  1.12     oster /*---------------------------------------------------------------------------
    229   1.1     oster  *
    230   1.1     oster  * performs whatever list compactions can be done, and frees any space
    231   1.1     oster  * that is no longer necessary.  Assumes only that the list is sorted
    232  1.12     oster  * by startSector.  crunch_list() compacts a single list as much as
    233  1.12     oster  * possible, and the second block of code deletes the entire list if
    234  1.12     oster  * possible.  crunch_list() is also called from
    235  1.12     oster  * MakeReconMapAccessList().
    236   1.1     oster  *
    237   1.1     oster  * When a recon unit is detected to be fully reconstructed, we set the
    238   1.1     oster  * corresponding bit in the parity stripe map so that the head follow
    239  1.12     oster  * code will not select this parity stripe again.  This is redundant
    240  1.12     oster  * (but harmless) when compact_stat_entry is called from the
    241  1.12     oster  * reconstruction code, but necessary when called from the user-write
    242  1.12     oster  * code.
    243   1.1     oster  *
    244  1.12     oster  *-------------------------------------------------------------------------*/
    245   1.1     oster 
    246  1.26     perry static void
    247  1.31     oster compact_stat_entry(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr, int i, int j)
    248   1.4     oster {
    249   1.4     oster 	RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
    250  1.31     oster 	RF_ReconMapListElem_t *p = mapPtr->status[j];
    251   1.4     oster 
    252  1.15     oster 	crunch_list(mapPtr, p);
    253   1.4     oster 
    254   1.4     oster 	if ((p->startSector == i * sectorsPerReconUnit) &&
    255  1.26     perry 	    (p->stopSector == i * sectorsPerReconUnit +
    256  1.12     oster 			      sectorsPerReconUnit - 1)) {
    257  1.31     oster 		mapPtr->status[j] = RU_ALL;
    258   1.4     oster 		mapPtr->unitsLeft--;
    259  1.15     oster 		FreeReconMapListElem(mapPtr, p);
    260   1.4     oster 	}
    261   1.4     oster }
    262   1.4     oster 
    263  1.31     oster 
    264  1.26     perry static void
    265  1.22     oster crunch_list(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *listPtr)
    266   1.4     oster {
    267   1.4     oster 	RF_ReconMapListElem_t *pt, *p = listPtr;
    268   1.4     oster 
    269   1.4     oster 	if (!p)
    270   1.4     oster 		return;
    271   1.4     oster 	pt = p;
    272   1.4     oster 	p = p->next;
    273   1.4     oster 	while (p) {
    274   1.4     oster 		if (pt->stopSector >= p->startSector - 1) {
    275   1.4     oster 			pt->stopSector = RF_MAX(pt->stopSector, p->stopSector);
    276   1.4     oster 			pt->next = p->next;
    277  1.15     oster 			FreeReconMapListElem(mapPtr, p);
    278   1.4     oster 			p = pt->next;
    279   1.4     oster 		} else {
    280   1.4     oster 			pt = p;
    281   1.4     oster 			p = p->next;
    282   1.4     oster 		}
    283   1.4     oster 	}
    284   1.1     oster }
    285  1.12     oster /*---------------------------------------------------------------------------
    286   1.4     oster  *
    287   1.1     oster  * Allocate and fill a new list element
    288   1.1     oster  *
    289  1.12     oster  *-------------------------------------------------------------------------*/
    290   1.1     oster 
    291   1.4     oster static RF_ReconMapListElem_t *
    292  1.22     oster MakeReconMapListElem(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector,
    293  1.22     oster 		     RF_SectorNum_t stopSector, RF_ReconMapListElem_t *next)
    294   1.4     oster {
    295   1.4     oster 	RF_ReconMapListElem_t *p;
    296   1.4     oster 
    297  1.19     oster 	p = pool_get(&mapPtr->elem_pool, PR_WAITOK);
    298   1.4     oster 	p->startSector = startSector;
    299   1.4     oster 	p->stopSector = stopSector;
    300   1.4     oster 	p->next = next;
    301   1.4     oster 	return (p);
    302   1.1     oster }
    303  1.12     oster /*---------------------------------------------------------------------------
    304   1.4     oster  *
    305   1.1     oster  * Free a list element
    306   1.1     oster  *
    307  1.12     oster  *-------------------------------------------------------------------------*/
    308   1.1     oster 
    309  1.26     perry static void
    310  1.22     oster FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *p)
    311   1.4     oster {
    312  1.15     oster 	pool_put(&mapPtr->elem_pool, p);
    313   1.1     oster }
    314  1.12     oster /*---------------------------------------------------------------------------
    315   1.4     oster  *
    316  1.12     oster  * Free an entire status structure.  Inefficient, but can be called at
    317  1.12     oster  * any time.
    318   1.1     oster  *
    319  1.12     oster  *-------------------------------------------------------------------------*/
    320  1.26     perry void
    321  1.22     oster rf_FreeReconMap(RF_ReconMap_t *mapPtr)
    322   1.4     oster {
    323   1.4     oster 	RF_ReconMapListElem_t *p, *q;
    324   1.4     oster 	RF_ReconUnitNum_t i;
    325   1.4     oster 
    326  1.31     oster 	for (i = 0; i < mapPtr->status_size; i++) {
    327   1.4     oster 		p = mapPtr->status[i];
    328   1.4     oster 		while (p != RU_NOTHING && p != RU_ALL) {
    329   1.4     oster 			q = p;
    330   1.4     oster 			p = p->next;
    331   1.4     oster 			RF_Free(q, sizeof(*q));
    332   1.4     oster 		}
    333   1.4     oster 	}
    334  1.31     oster 
    335  1.32       mrg 	rf_destroy_mutex2(mapPtr->mutex);
    336  1.32       mrg 	rf_destroy_cond2(mapPtr->cv);
    337  1.32       mrg 
    338  1.15     oster 	pool_destroy(&mapPtr->elem_pool);
    339  1.31     oster 	RF_Free(mapPtr->status, mapPtr->status_size *
    340  1.12     oster 		sizeof(RF_ReconMapListElem_t *));
    341   1.4     oster 	RF_Free(mapPtr, sizeof(RF_ReconMap_t));
    342   1.1     oster }
    343  1.12     oster /*---------------------------------------------------------------------------
    344   1.1     oster  *
    345   1.1     oster  * returns nonzero if the indicated RU has been reconstructed already
    346   1.1     oster  *
    347  1.12     oster  *-------------------------------------------------------------------------*/
    348   1.1     oster 
    349  1.26     perry int
    350  1.22     oster rf_CheckRUReconstructed(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector)
    351   1.1     oster {
    352   1.4     oster 	RF_ReconUnitNum_t i;
    353  1.31     oster 	int rv;
    354   1.1     oster 
    355   1.4     oster 	i = startSector / mapPtr->sectorsPerReconUnit;
    356  1.31     oster 
    357  1.31     oster 	if (i < mapPtr->low_ru)
    358  1.31     oster 		rv = 1;
    359  1.31     oster 	else if (i > mapPtr->high_ru)
    360  1.31     oster 		rv = 0;
    361  1.31     oster 	else {
    362  1.31     oster 		i = i - mapPtr->low_ru + mapPtr->head;
    363  1.31     oster 		if (i >= mapPtr->status_size)
    364  1.31     oster 			i = i - mapPtr->status_size;
    365  1.31     oster 		if (mapPtr->status[i] == RU_ALL)
    366  1.31     oster 			rv = 1;
    367  1.31     oster 		else
    368  1.31     oster 			rv = 0;
    369  1.31     oster 	}
    370  1.31     oster 
    371  1.31     oster 	return rv;
    372   1.1     oster }
    373   1.1     oster 
    374  1.26     perry RF_ReconUnitCount_t
    375  1.22     oster rf_UnitsLeftToReconstruct(RF_ReconMap_t *mapPtr)
    376   1.1     oster {
    377   1.4     oster 	RF_ASSERT(mapPtr != NULL);
    378   1.4     oster 	return (mapPtr->unitsLeft);
    379   1.4     oster }
    380   1.4     oster 
    381  1.11     oster #if RF_DEBUG_RECON
    382  1.26     perry void
    383  1.22     oster rf_PrintReconSchedule(RF_ReconMap_t *mapPtr, struct timeval *starttime)
    384   1.4     oster {
    385   1.4     oster 	static int old_pctg = -1;
    386   1.4     oster 	struct timeval tv, diff;
    387   1.4     oster 	int     new_pctg;
    388   1.4     oster 
    389  1.26     perry 	new_pctg = 100 - (rf_UnitsLeftToReconstruct(mapPtr) *
    390  1.12     oster 			  100 / mapPtr->totalRUs);
    391   1.4     oster 	if (new_pctg != old_pctg) {
    392   1.4     oster 		RF_GETTIME(tv);
    393   1.4     oster 		RF_TIMEVAL_DIFF(starttime, &tv, &diff);
    394  1.26     perry 		printf("%d %d.%06d\n", (int) new_pctg, (int) diff.tv_sec,
    395  1.12     oster 		       (int) diff.tv_usec);
    396   1.4     oster 		old_pctg = new_pctg;
    397   1.4     oster 	}
    398   1.1     oster }
    399  1.11     oster #endif
    400