1 /* $NetBSD: rf_reconmap.c,v 1.25 2004/03/18 17:46:22 oster 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 1.25 2004/03/18 17:46:22 oster 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 * raidPtr, RF_ReconMap_t * mapPtr,
65 int i);
66 static void crunch_list(RF_ReconMap_t *mapPtr,
67 RF_ReconMapListElem_t * listPtr);
68 static RF_ReconMapListElem_t *
69 MakeReconMapListElem(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector,
70 RF_SectorNum_t stopSector, RF_ReconMapListElem_t * next);
71 static void
72 FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t * p);
73 #if 0
74 static void PrintList(RF_ReconMapListElem_t * listPtr);
75 #endif
76
77 /*---------------------------------------------------------------------------
78 *
79 * Creates and initializes new Reconstruction map
80 *
81 * ru_sectors - size of reconstruction unit in sectors
82 * disk_sectors - size of disk in sectors
83 * spareUnitsPerDisk - zero unless distributed sparing
84 *-------------------------------------------------------------------------*/
85
86 RF_ReconMap_t *
87 rf_MakeReconMap(RF_Raid_t *raidPtr, RF_SectorCount_t ru_sectors,
88 RF_SectorCount_t disk_sectors,
89 RF_ReconUnitCount_t spareUnitsPerDisk)
90 {
91 RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
92 RF_ReconUnitCount_t num_rus = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerRU;
93 RF_ReconMap_t *p;
94
95 RF_Malloc(p, sizeof(RF_ReconMap_t), (RF_ReconMap_t *));
96 p->sectorsPerReconUnit = ru_sectors;
97 p->sectorsInDisk = disk_sectors;
98
99 p->totalRUs = num_rus;
100 p->spareRUs = spareUnitsPerDisk;
101 p->unitsLeft = num_rus - spareUnitsPerDisk;
102
103 RF_Malloc(p->status, num_rus * sizeof(RF_ReconMapListElem_t *), (RF_ReconMapListElem_t **));
104 RF_ASSERT(p->status != (RF_ReconMapListElem_t **) NULL);
105
106 (void) memset((char *) p->status, 0,
107 num_rus * sizeof(RF_ReconMapListElem_t *));
108
109
110 pool_init(&p->elem_pool, sizeof(RF_ReconMapListElem_t), 0,
111 0, 0, "raidreconpl", NULL);
112 pool_prime(&p->elem_pool, RF_NUM_RECON_POOL_ELEM);
113
114 rf_mutex_init(&p->mutex);
115 return (p);
116 }
117
118
119 /*---------------------------------------------------------------------------
120 *
121 * marks a new set of sectors as reconstructed. All the possible
122 * mergings get complicated. To simplify matters, the approach I take
123 * is to just dump something into the list, and then clean it up
124 * (i.e. merge elements and eliminate redundant ones) in a second pass
125 * over the list (compact_stat_entry()). Not 100% efficient, since a
126 * structure can be allocated and then immediately freed, but it keeps
127 * this code from becoming (more of) a nightmare of special cases.
128 * The only thing that compact_stat_entry() assumes is that the list
129 * is sorted by startSector, and so this is the only condition I
130 * maintain here. (MCH)
131 *
132 * This code now uses a pool instead of the previous malloc/free
133 * stuff.
134 *-------------------------------------------------------------------------*/
135
136 void
137 rf_ReconMapUpdate(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr,
138 RF_SectorNum_t startSector, RF_SectorNum_t stopSector)
139 {
140 RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
141 RF_SectorNum_t i, first_in_RU, last_in_RU;
142 RF_ReconMapListElem_t *p, *pt;
143
144 RF_LOCK_MUTEX(mapPtr->mutex);
145 while(mapPtr->lock) {
146 ltsleep(&mapPtr->lock, PRIBIO, "reconupdate", 0, &mapPtr->mutex);
147 }
148 mapPtr->lock = 1;
149 RF_UNLOCK_MUTEX(mapPtr->mutex);
150 RF_ASSERT(startSector >= 0 && stopSector < mapPtr->sectorsInDisk &&
151 stopSector >= startSector);
152
153 while (startSector <= stopSector) {
154 i = startSector / mapPtr->sectorsPerReconUnit;
155 first_in_RU = i * sectorsPerReconUnit;
156 last_in_RU = first_in_RU + sectorsPerReconUnit - 1;
157 p = mapPtr->status[i];
158 if (p != RU_ALL) {
159 if (p == RU_NOTHING || p->startSector > startSector) {
160 /* insert at front of list */
161
162 mapPtr->status[i] = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), (p == RU_NOTHING) ? NULL : p);
163
164 } else {/* general case */
165 do { /* search for place to insert */
166 pt = p;
167 p = p->next;
168 } while (p && (p->startSector < startSector));
169 pt->next = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), p);
170
171 }
172 compact_stat_entry(raidPtr, mapPtr, i);
173 }
174 startSector = RF_MIN(stopSector, last_in_RU) + 1;
175 }
176 RF_LOCK_MUTEX(mapPtr->mutex);
177 mapPtr->lock = 0;
178 wakeup(&mapPtr->lock);
179 RF_UNLOCK_MUTEX(mapPtr->mutex);
180 }
181
182
183
184 /*---------------------------------------------------------------------------
185 *
186 * performs whatever list compactions can be done, and frees any space
187 * that is no longer necessary. Assumes only that the list is sorted
188 * by startSector. crunch_list() compacts a single list as much as
189 * possible, and the second block of code deletes the entire list if
190 * possible. crunch_list() is also called from
191 * MakeReconMapAccessList().
192 *
193 * When a recon unit is detected to be fully reconstructed, we set the
194 * corresponding bit in the parity stripe map so that the head follow
195 * code will not select this parity stripe again. This is redundant
196 * (but harmless) when compact_stat_entry is called from the
197 * reconstruction code, but necessary when called from the user-write
198 * code.
199 *
200 *-------------------------------------------------------------------------*/
201
202 static void
203 compact_stat_entry(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr, int i)
204 {
205 RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
206 RF_ReconMapListElem_t *p = mapPtr->status[i];
207
208 crunch_list(mapPtr, p);
209
210 if ((p->startSector == i * sectorsPerReconUnit) &&
211 (p->stopSector == i * sectorsPerReconUnit +
212 sectorsPerReconUnit - 1)) {
213 mapPtr->status[i] = RU_ALL;
214 mapPtr->unitsLeft--;
215 FreeReconMapListElem(mapPtr, p);
216 }
217 }
218
219 static void
220 crunch_list(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *listPtr)
221 {
222 RF_ReconMapListElem_t *pt, *p = listPtr;
223
224 if (!p)
225 return;
226 pt = p;
227 p = p->next;
228 while (p) {
229 if (pt->stopSector >= p->startSector - 1) {
230 pt->stopSector = RF_MAX(pt->stopSector, p->stopSector);
231 pt->next = p->next;
232 FreeReconMapListElem(mapPtr, p);
233 p = pt->next;
234 } else {
235 pt = p;
236 p = p->next;
237 }
238 }
239 }
240 /*---------------------------------------------------------------------------
241 *
242 * Allocate and fill a new list element
243 *
244 *-------------------------------------------------------------------------*/
245
246 static RF_ReconMapListElem_t *
247 MakeReconMapListElem(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector,
248 RF_SectorNum_t stopSector, RF_ReconMapListElem_t *next)
249 {
250 RF_ReconMapListElem_t *p;
251
252 p = pool_get(&mapPtr->elem_pool, PR_WAITOK);
253 p->startSector = startSector;
254 p->stopSector = stopSector;
255 p->next = next;
256 return (p);
257 }
258 /*---------------------------------------------------------------------------
259 *
260 * Free a list element
261 *
262 *-------------------------------------------------------------------------*/
263
264 static void
265 FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *p)
266 {
267 pool_put(&mapPtr->elem_pool, p);
268 }
269 /*---------------------------------------------------------------------------
270 *
271 * Free an entire status structure. Inefficient, but can be called at
272 * any time.
273 *
274 *-------------------------------------------------------------------------*/
275 void
276 rf_FreeReconMap(RF_ReconMap_t *mapPtr)
277 {
278 RF_ReconMapListElem_t *p, *q;
279 RF_ReconUnitCount_t numRUs;
280 RF_ReconUnitNum_t i;
281
282 numRUs = mapPtr->sectorsInDisk / mapPtr->sectorsPerReconUnit;
283 if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
284 numRUs++;
285
286 for (i = 0; i < numRUs; i++) {
287 p = mapPtr->status[i];
288 while (p != RU_NOTHING && p != RU_ALL) {
289 q = p;
290 p = p->next;
291 RF_Free(q, sizeof(*q));
292 }
293 }
294 pool_destroy(&mapPtr->elem_pool);
295 RF_Free(mapPtr->status, mapPtr->totalRUs *
296 sizeof(RF_ReconMapListElem_t *));
297 RF_Free(mapPtr, sizeof(RF_ReconMap_t));
298 }
299 /*---------------------------------------------------------------------------
300 *
301 * returns nonzero if the indicated RU has been reconstructed already
302 *
303 *-------------------------------------------------------------------------*/
304
305 int
306 rf_CheckRUReconstructed(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector)
307 {
308 RF_ReconMapListElem_t *l; /* used for searching */
309 RF_ReconUnitNum_t i;
310
311 i = startSector / mapPtr->sectorsPerReconUnit;
312 l = mapPtr->status[i];
313 return ((l == RU_ALL) ? 1 : 0);
314 }
315
316 RF_ReconUnitCount_t
317 rf_UnitsLeftToReconstruct(RF_ReconMap_t *mapPtr)
318 {
319 RF_ASSERT(mapPtr != NULL);
320 return (mapPtr->unitsLeft);
321 }
322
323 #if 0
324 static void
325 PrintList(RF_ReconMapListElem_t *listPtr)
326 {
327 while (listPtr) {
328 printf("%d,%d -> ", (int) listPtr->startSector,
329 (int) listPtr->stopSector);
330 listPtr = listPtr->next;
331 }
332 printf("\n");
333 }
334
335 void
336 rf_PrintReconMap(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr, RF_RowCol_t fcol)
337 {
338 RF_ReconUnitCount_t numRUs;
339 RF_ReconMapListElem_t *p;
340 RF_ReconUnitNum_t i;
341
342 numRUs = mapPtr->totalRUs;
343 if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
344 numRUs++;
345
346 for (i = 0; i < numRUs; i++) {
347 p = mapPtr->status[i];
348 if (p == RU_ALL)/* printf("[%d] ALL\n",i) */
349 ;
350 else
351 if (p == RU_NOTHING) {
352 printf("%d: Unreconstructed\n", i);
353 } else {
354 printf("%d: ", i);
355 PrintList(p);
356 }
357 }
358 }
359 #endif
360
361 #if RF_DEBUG_RECON
362 void
363 rf_PrintReconSchedule(RF_ReconMap_t *mapPtr, struct timeval *starttime)
364 {
365 static int old_pctg = -1;
366 struct timeval tv, diff;
367 int new_pctg;
368
369 new_pctg = 100 - (rf_UnitsLeftToReconstruct(mapPtr) *
370 100 / mapPtr->totalRUs);
371 if (new_pctg != old_pctg) {
372 RF_GETTIME(tv);
373 RF_TIMEVAL_DIFF(starttime, &tv, &diff);
374 printf("%d %d.%06d\n", (int) new_pctg, (int) diff.tv_sec,
375 (int) diff.tv_usec);
376 old_pctg = new_pctg;
377 }
378 }
379 #endif
380
Cache object: 2898475dc75332b57f19f0144a7a785e
|