1 /*        $NetBSD: rf_reconmap.c,v 1.39 2022/04/08 10:27:04 andvar 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@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  * rf_reconmap.c
31  *
32  * code to maintain a map of what sectors have/have not been reconstructed
33  *
34  *************************************************************************/
35 
36 #include <sys/cdefs.h>
37 __KERNEL_RCSID(0, "$NetBSD: rf_reconmap.c,v 1.39 2022/04/08 10:27:04 andvar Exp $");
38 
39 #include "rf_raid.h"
40 #include <sys/time.h>
41 #include "rf_general.h"
42 #include "rf_utils.h"
43 
44 /* special pointer values indicating that a reconstruction unit
45  * has been either totally reconstructed or not at all.  Both
46  * are illegal pointer values, so you have to be careful not to
47  * dereference through them.  RU_NOTHING must be zero, since
48  * MakeReconMap uses memset to initialize the structure.  These are used
49  * only at the head of the list.
50  */
51 #define RU_ALL      ((RF_ReconMapListElem_t *) -1)
52 #define RU_NOTHING  ((RF_ReconMapListElem_t *) 0)
53 
54 /* For most reconstructs we need at most 3 RF_ReconMapListElem_t's.
55  * Bounding the number we need is quite difficult, as it depends on how
56  * badly the sectors to be reconstructed get divided up.  In the current
57  * code, the reconstructed sectors appeared aligned on stripe boundaries,
58  * and are always presented in stripe width units, so we're probably
59  * allocating quite a bit more than we'll ever need.
60  */
61 #define RF_NUM_RECON_POOL_ELEM 100
62 
63 static void
64 compact_stat_entry(RF_Raid_t *, RF_ReconMap_t *, int, int);
65 static void crunch_list(RF_ReconMap_t *, RF_ReconMapListElem_t *);
66 static RF_ReconMapListElem_t *
67 MakeReconMapListElem(RF_ReconMap_t *, RF_SectorNum_t, RF_SectorNum_t,
68                          RF_ReconMapListElem_t *);
69 static void
70 FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t * p);
71 
72 /*---------------------------------------------------------------------------
73  *
74  * Creates and initializes new Reconstruction map
75  *
76  * ru_sectors   - size of reconstruction unit in sectors
77  * disk_sectors - size of disk in sectors
78  * spareUnitsPerDisk - zero unless distributed sparing
79  *-------------------------------------------------------------------------*/
80 
81 RF_ReconMap_t *
rf_MakeReconMap(RF_Raid_t * raidPtr,RF_SectorCount_t ru_sectors,RF_SectorCount_t disk_sectors,RF_ReconUnitCount_t spareUnitsPerDisk)82 rf_MakeReconMap(RF_Raid_t *raidPtr, RF_SectorCount_t ru_sectors,
83                     RF_SectorCount_t disk_sectors,
84                     RF_ReconUnitCount_t spareUnitsPerDisk)
85 {
86           RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
87           RF_ReconUnitCount_t num_rus = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerRU;
88           RF_ReconMap_t *p;
89 
90           p = RF_Malloc(sizeof(*p));
91           p->sectorsPerReconUnit = ru_sectors;
92           p->sectorsInDisk = disk_sectors;
93 
94           p->totalRUs = num_rus;
95           p->spareRUs = spareUnitsPerDisk;
96           p->unitsLeft = num_rus - spareUnitsPerDisk;
97           p->low_ru = 0;
98           p->status_size = RF_RECONMAP_SIZE;
99           p->high_ru = p->status_size - 1;
100           p->head = 0;
101 
102           p->status = RF_Malloc(p->status_size * sizeof(*p->status));
103           RF_ASSERT(p->status != NULL);
104 
105           pool_init(&p->elem_pool, sizeof(RF_ReconMapListElem_t), 0,
106               0, 0, "raidreconpl", NULL, IPL_BIO);
107           pool_prime(&p->elem_pool, RF_NUM_RECON_POOL_ELEM);
108 
109           rf_init_mutex2(p->mutex, IPL_VM);
110           rf_init_cond2(p->cv, "reconupdate");
111 
112           return (p);
113 }
114 
115 
116 /*---------------------------------------------------------------------------
117  *
118  * marks a new set of sectors as reconstructed.  All the possible
119  * mergings get complicated.  To simplify matters, the approach I take
120  * is to just dump something into the list, and then clean it up
121  * (i.e. merge elements and eliminate redundant ones) in a second pass
122  * over the list (compact_stat_entry()).  Not 100% efficient, since a
123  * structure can be allocated and then immediately freed, but it keeps
124  * this code from becoming (more of) a nightmare of special cases.
125  * The only thing that compact_stat_entry() assumes is that the list
126  * is sorted by startSector, and so this is the only condition I
127  * maintain here.  (MCH)
128  *
129  * This code now uses a pool instead of the previous malloc/free
130  * stuff.
131  *-------------------------------------------------------------------------*/
132 
133 void
rf_ReconMapUpdate(RF_Raid_t * raidPtr,RF_ReconMap_t * mapPtr,RF_SectorNum_t startSector,RF_SectorNum_t stopSector)134 rf_ReconMapUpdate(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr,
135                       RF_SectorNum_t startSector, RF_SectorNum_t stopSector)
136 {
137           RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
138           RF_SectorNum_t i, first_in_RU, last_in_RU, ru;
139           RF_ReconMapListElem_t *p, *pt;
140 
141           rf_lock_mutex2(mapPtr->mutex);
142           while(mapPtr->lock) {
143                     rf_wait_cond2(mapPtr->cv, mapPtr->mutex);
144           }
145           mapPtr->lock = 1;
146           rf_unlock_mutex2(mapPtr->mutex);
147           RF_ASSERT(startSector >= 0 && stopSector < mapPtr->sectorsInDisk &&
148                       stopSector >= startSector);
149 
150           while (startSector <= stopSector) {
151                     i = startSector / mapPtr->sectorsPerReconUnit;
152                     first_in_RU = i * sectorsPerReconUnit;
153                     last_in_RU = first_in_RU + sectorsPerReconUnit - 1;
154 
155                     /* do we need to move the queue? */
156                     while (i > mapPtr->high_ru) {
157 #if 0
158 #ifdef DIAGNOSTIC
159                               /* XXX: The check below is not valid for
160                                * RAID5_RS.  It is valid for RAID 1 and RAID 5.
161                                * The issue is that we can easily have
162                                * RU_NOTHING entries here too, and those are
163                                * quite correct.
164                                */
165                               if (mapPtr->status[mapPtr->head]!=RU_ALL) {
166                                         printf("\nraid%d: reconmap incorrect -- working on i %" PRIu64 "\n",
167                                                raidPtr->raidid, i);
168                                         printf("raid%d: ru %" PRIu64 " not completed!!!\n",
169                                                raidPtr->raidid, mapPtr->head);
170 
171                                         printf("raid%d: low: %" PRIu64 " high: %" PRIu64 "\n",
172                                                raidPtr->raidid, mapPtr->low_ru, mapPtr->high_ru);
173 
174                                         panic("reconmap incorrect");
175                               }
176 #endif
177 #endif
178                               mapPtr->low_ru++;
179                               mapPtr->high_ru++;
180                               /* initialize "highest" RU status entry, which
181                                  will take over the current head position */
182                               mapPtr->status[mapPtr->head]=RU_NOTHING;
183 
184                               /* move head too */
185                               mapPtr->head++;
186                               if (mapPtr->head >= mapPtr->status_size)
187                                         mapPtr->head = 0;
188 
189                     }
190 
191                     ru = i - mapPtr->low_ru + mapPtr->head;
192                     if (ru >= mapPtr->status_size)
193                               ru = ru - mapPtr->status_size;
194 
195                     if ((ru < 0) || (ru >= mapPtr->status_size)) {
196                               printf("raid%d: ru is bogus %" PRIu64 "%" PRIu64 "%" PRIu64 "%" PRIu64 "%" PRIu64 "\n",
197                                      raidPtr->raidid, i, ru, mapPtr->head, mapPtr->low_ru, mapPtr->high_ru);
198                               panic("bogus ru in reconmap");
199                     }
200 
201                     p = mapPtr->status[ru];
202                     if (p != RU_ALL) {
203                               if (p == RU_NOTHING || p->startSector > startSector) {
204                                         /* insert at front of list */
205 
206                                         mapPtr->status[ru] = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), (p == RU_NOTHING) ? NULL : p);
207 
208                               } else {/* general case */
209                                         do {      /* search for place to insert */
210                                                   pt = p;
211                                                   p = p->next;
212                                         } while (p && (p->startSector < startSector));
213                                         pt->next = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), p);
214 
215                               }
216                               compact_stat_entry(raidPtr, mapPtr, i, ru);
217                     }
218                     startSector = RF_MIN(stopSector, last_in_RU) + 1;
219           }
220           rf_lock_mutex2(mapPtr->mutex);
221           mapPtr->lock = 0;
222           rf_broadcast_cond2(mapPtr->cv);
223           rf_unlock_mutex2(mapPtr->mutex);
224 }
225 
226 
227 
228 /*---------------------------------------------------------------------------
229  *
230  * performs whatever list compactions can be done, and frees any space
231  * that is no longer necessary.  Assumes only that the list is sorted
232  * by startSector.  crunch_list() compacts a single list as much as
233  * possible, and the second block of code deletes the entire list if
234  * possible.  crunch_list() is also called from
235  * MakeReconMapAccessList().
236  *
237  * When a recon unit is detected to be fully reconstructed, we set the
238  * corresponding bit in the parity stripe map so that the head follow
239  * code will not select this parity stripe again.  This is redundant
240  * (but harmless) when compact_stat_entry is called from the
241  * reconstruction code, but necessary when called from the user-write
242  * code.
243  *
244  *-------------------------------------------------------------------------*/
245 
246 static void
compact_stat_entry(RF_Raid_t * raidPtr,RF_ReconMap_t * mapPtr,int i,int j)247 compact_stat_entry(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr, int i, int j)
248 {
249           RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
250           RF_ReconMapListElem_t *p = mapPtr->status[j];
251 
252           crunch_list(mapPtr, p);
253 
254           if ((p->startSector == i * sectorsPerReconUnit) &&
255               (p->stopSector == i * sectorsPerReconUnit +
256                                     sectorsPerReconUnit - 1)) {
257                     mapPtr->status[j] = RU_ALL;
258                     mapPtr->unitsLeft--;
259                     FreeReconMapListElem(mapPtr, p);
260           }
261 }
262 
263 
264 static void
crunch_list(RF_ReconMap_t * mapPtr,RF_ReconMapListElem_t * listPtr)265 crunch_list(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *listPtr)
266 {
267           RF_ReconMapListElem_t *pt, *p = listPtr;
268 
269           if (!p)
270                     return;
271           pt = p;
272           p = p->next;
273           while (p) {
274                     if (pt->stopSector >= p->startSector - 1) {
275                               pt->stopSector = RF_MAX(pt->stopSector, p->stopSector);
276                               pt->next = p->next;
277                               FreeReconMapListElem(mapPtr, p);
278                               p = pt->next;
279                     } else {
280                               pt = p;
281                               p = p->next;
282                     }
283           }
284 }
285 /*---------------------------------------------------------------------------
286  *
287  * Allocate and fill a new list element
288  *
289  *-------------------------------------------------------------------------*/
290 
291 static RF_ReconMapListElem_t *
MakeReconMapListElem(RF_ReconMap_t * mapPtr,RF_SectorNum_t startSector,RF_SectorNum_t stopSector,RF_ReconMapListElem_t * next)292 MakeReconMapListElem(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector,
293                          RF_SectorNum_t stopSector, RF_ReconMapListElem_t *next)
294 {
295           RF_ReconMapListElem_t *p;
296 
297           p = pool_get(&mapPtr->elem_pool, PR_WAITOK);
298           p->startSector = startSector;
299           p->stopSector = stopSector;
300           p->next = next;
301           return (p);
302 }
303 /*---------------------------------------------------------------------------
304  *
305  * Free a list element
306  *
307  *-------------------------------------------------------------------------*/
308 
309 static void
FreeReconMapListElem(RF_ReconMap_t * mapPtr,RF_ReconMapListElem_t * p)310 FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *p)
311 {
312           pool_put(&mapPtr->elem_pool, p);
313 }
314 /*---------------------------------------------------------------------------
315  *
316  * Free an entire status structure.  Inefficient, but can be called at
317  * any time.
318  *
319  *-------------------------------------------------------------------------*/
320 void
rf_FreeReconMap(RF_ReconMap_t * mapPtr)321 rf_FreeReconMap(RF_ReconMap_t *mapPtr)
322 {
323           RF_ReconMapListElem_t *p, *q;
324           RF_ReconUnitNum_t i;
325 
326           for (i = 0; i < mapPtr->status_size; i++) {
327                     p = mapPtr->status[i];
328                     while (p != RU_NOTHING && p != RU_ALL) {
329                               q = p;
330                               p = p->next;
331                               RF_Free(q, sizeof(*q));
332                     }
333           }
334 
335           rf_destroy_mutex2(mapPtr->mutex);
336           rf_destroy_cond2(mapPtr->cv);
337 
338           pool_destroy(&mapPtr->elem_pool);
339           RF_Free(mapPtr->status, mapPtr->status_size *
340                     sizeof(RF_ReconMapListElem_t *));
341           RF_Free(mapPtr, sizeof(RF_ReconMap_t));
342 }
343 /*---------------------------------------------------------------------------
344  *
345  * returns nonzero if the indicated RU has been reconstructed already
346  *
347  *-------------------------------------------------------------------------*/
348 
349 int
rf_CheckRUReconstructed(RF_ReconMap_t * mapPtr,RF_SectorNum_t startSector)350 rf_CheckRUReconstructed(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector)
351 {
352           RF_ReconUnitNum_t i;
353           int rv;
354 
355           i = startSector / mapPtr->sectorsPerReconUnit;
356 
357           if (i < mapPtr->low_ru)
358                     rv = 1;
359           else if (i > mapPtr->high_ru)
360                     rv = 0;
361           else {
362                     i = i - mapPtr->low_ru + mapPtr->head;
363                     if (i >= mapPtr->status_size)
364                               i = i - mapPtr->status_size;
365                     if (mapPtr->status[i] == RU_ALL)
366                               rv = 1;
367                     else
368                               rv = 0;
369           }
370 
371           return rv;
372 }
373 
374 RF_ReconUnitCount_t
rf_UnitsLeftToReconstruct(RF_ReconMap_t * mapPtr)375 rf_UnitsLeftToReconstruct(RF_ReconMap_t *mapPtr)
376 {
377           RF_ASSERT(mapPtr != NULL);
378           return (mapPtr->unitsLeft);
379 }
380 
381 #if RF_DEBUG_RECON
382 void
rf_PrintReconSchedule(RF_ReconMap_t * mapPtr,struct timeval * starttime)383 rf_PrintReconSchedule(RF_ReconMap_t *mapPtr, struct timeval *starttime)
384 {
385           static int old_pctg = -1;
386           struct timeval tv, diff;
387           int     new_pctg;
388 
389           new_pctg = 100 - (rf_UnitsLeftToReconstruct(mapPtr) *
390                                 100 / mapPtr->totalRUs);
391           if (new_pctg != old_pctg) {
392                     RF_GETTIME(tv);
393                     RF_TIMEVAL_DIFF(starttime, &tv, &diff);
394                     printf("%d %d.%06d\n", (int) new_pctg, (int) diff.tv_sec,
395                            (int) diff.tv_usec);
396                     old_pctg = new_pctg;
397           }
398 }
399 #endif
400