rf_reconmap.c revision 1.30 1 1.30 ad /* $NetBSD: rf_reconmap.c,v 1.30 2007/03/12 18:18:31 ad 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 1.30 ad __KERNEL_RCSID(0, "$NetBSD: rf_reconmap.c,v 1.30 2007/03/12 18:18:31 ad 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 1.4 oster compact_stat_entry(RF_Raid_t * raidPtr, RF_ReconMap_t * mapPtr,
65 1.4 oster int i);
66 1.26 perry 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 1.26 perry 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 1.22 oster * ru_sectors - size of reconstruction unit in sectors
82 1.22 oster * disk_sectors - size of disk in sectors
83 1.22 oster * spareUnitsPerDisk - zero unless distributed sparing
84 1.12 oster *-------------------------------------------------------------------------*/
85 1.1 oster
86 1.4 oster RF_ReconMap_t *
87 1.26 perry rf_MakeReconMap(RF_Raid_t *raidPtr, RF_SectorCount_t ru_sectors,
88 1.22 oster RF_SectorCount_t disk_sectors,
89 1.22 oster RF_ReconUnitCount_t spareUnitsPerDisk)
90 1.4 oster {
91 1.4 oster RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
92 1.4 oster RF_ReconUnitCount_t num_rus = layoutPtr->stripeUnitsPerDisk / layoutPtr->SUsPerRU;
93 1.4 oster RF_ReconMap_t *p;
94 1.4 oster
95 1.4 oster RF_Malloc(p, sizeof(RF_ReconMap_t), (RF_ReconMap_t *));
96 1.4 oster p->sectorsPerReconUnit = ru_sectors;
97 1.4 oster p->sectorsInDisk = disk_sectors;
98 1.4 oster
99 1.4 oster p->totalRUs = num_rus;
100 1.4 oster p->spareRUs = spareUnitsPerDisk;
101 1.4 oster p->unitsLeft = num_rus - spareUnitsPerDisk;
102 1.4 oster
103 1.4 oster RF_Malloc(p->status, num_rus * sizeof(RF_ReconMapListElem_t *), (RF_ReconMapListElem_t **));
104 1.4 oster RF_ASSERT(p->status != (RF_ReconMapListElem_t **) NULL);
105 1.4 oster
106 1.7 thorpej (void) memset((char *) p->status, 0,
107 1.7 thorpej num_rus * sizeof(RF_ReconMapListElem_t *));
108 1.4 oster
109 1.26 perry
110 1.15 oster pool_init(&p->elem_pool, sizeof(RF_ReconMapListElem_t), 0,
111 1.30 ad 0, 0, "raidreconpl", NULL, IPL_BIO);
112 1.17 oster pool_prime(&p->elem_pool, RF_NUM_RECON_POOL_ELEM);
113 1.15 oster
114 1.20 oster rf_mutex_init(&p->mutex);
115 1.4 oster return (p);
116 1.1 oster }
117 1.1 oster
118 1.1 oster
119 1.12 oster /*---------------------------------------------------------------------------
120 1.1 oster *
121 1.12 oster * marks a new set of sectors as reconstructed. All the possible
122 1.12 oster * mergings get complicated. To simplify matters, the approach I take
123 1.12 oster * is to just dump something into the list, and then clean it up
124 1.12 oster * (i.e. merge elements and eliminate redundant ones) in a second pass
125 1.12 oster * over the list (compact_stat_entry()). Not 100% efficient, since a
126 1.12 oster * structure can be allocated and then immediately freed, but it keeps
127 1.12 oster * this code from becoming (more of) a nightmare of special cases.
128 1.12 oster * The only thing that compact_stat_entry() assumes is that the list
129 1.12 oster * is sorted by startSector, and so this is the only condition I
130 1.12 oster * maintain here. (MCH)
131 1.1 oster *
132 1.15 oster * This code now uses a pool instead of the previous malloc/free
133 1.15 oster * stuff.
134 1.12 oster *-------------------------------------------------------------------------*/
135 1.1 oster
136 1.26 perry void
137 1.22 oster rf_ReconMapUpdate(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr,
138 1.22 oster RF_SectorNum_t startSector, RF_SectorNum_t stopSector)
139 1.4 oster {
140 1.4 oster RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
141 1.4 oster RF_SectorNum_t i, first_in_RU, last_in_RU;
142 1.4 oster RF_ReconMapListElem_t *p, *pt;
143 1.4 oster
144 1.4 oster RF_LOCK_MUTEX(mapPtr->mutex);
145 1.24 oster while(mapPtr->lock) {
146 1.24 oster ltsleep(&mapPtr->lock, PRIBIO, "reconupdate", 0, &mapPtr->mutex);
147 1.24 oster }
148 1.24 oster mapPtr->lock = 1;
149 1.24 oster RF_UNLOCK_MUTEX(mapPtr->mutex);
150 1.26 perry RF_ASSERT(startSector >= 0 && stopSector < mapPtr->sectorsInDisk &&
151 1.12 oster stopSector >= startSector);
152 1.4 oster
153 1.4 oster while (startSector <= stopSector) {
154 1.4 oster i = startSector / mapPtr->sectorsPerReconUnit;
155 1.4 oster first_in_RU = i * sectorsPerReconUnit;
156 1.4 oster last_in_RU = first_in_RU + sectorsPerReconUnit - 1;
157 1.4 oster p = mapPtr->status[i];
158 1.4 oster if (p != RU_ALL) {
159 1.12 oster if (p == RU_NOTHING || p->startSector > startSector) {
160 1.12 oster /* insert at front of list */
161 1.4 oster
162 1.15 oster mapPtr->status[i] = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), (p == RU_NOTHING) ? NULL : p);
163 1.4 oster
164 1.4 oster } else {/* general case */
165 1.4 oster do { /* search for place to insert */
166 1.4 oster pt = p;
167 1.4 oster p = p->next;
168 1.4 oster } while (p && (p->startSector < startSector));
169 1.15 oster pt->next = MakeReconMapListElem(mapPtr,startSector, RF_MIN(stopSector, last_in_RU), p);
170 1.14 oster
171 1.4 oster }
172 1.4 oster compact_stat_entry(raidPtr, mapPtr, i);
173 1.4 oster }
174 1.4 oster startSector = RF_MIN(stopSector, last_in_RU) + 1;
175 1.4 oster }
176 1.24 oster RF_LOCK_MUTEX(mapPtr->mutex);
177 1.24 oster mapPtr->lock = 0;
178 1.24 oster wakeup(&mapPtr->lock);
179 1.4 oster RF_UNLOCK_MUTEX(mapPtr->mutex);
180 1.1 oster }
181 1.1 oster
182 1.1 oster
183 1.1 oster
184 1.12 oster /*---------------------------------------------------------------------------
185 1.1 oster *
186 1.1 oster * performs whatever list compactions can be done, and frees any space
187 1.1 oster * that is no longer necessary. Assumes only that the list is sorted
188 1.12 oster * by startSector. crunch_list() compacts a single list as much as
189 1.12 oster * possible, and the second block of code deletes the entire list if
190 1.12 oster * possible. crunch_list() is also called from
191 1.12 oster * MakeReconMapAccessList().
192 1.1 oster *
193 1.1 oster * When a recon unit is detected to be fully reconstructed, we set the
194 1.1 oster * corresponding bit in the parity stripe map so that the head follow
195 1.12 oster * code will not select this parity stripe again. This is redundant
196 1.12 oster * (but harmless) when compact_stat_entry is called from the
197 1.12 oster * reconstruction code, but necessary when called from the user-write
198 1.12 oster * code.
199 1.1 oster *
200 1.12 oster *-------------------------------------------------------------------------*/
201 1.1 oster
202 1.26 perry static void
203 1.29 christos compact_stat_entry(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr, int i)
204 1.4 oster {
205 1.4 oster RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
206 1.4 oster RF_ReconMapListElem_t *p = mapPtr->status[i];
207 1.4 oster
208 1.15 oster crunch_list(mapPtr, p);
209 1.4 oster
210 1.4 oster if ((p->startSector == i * sectorsPerReconUnit) &&
211 1.26 perry (p->stopSector == i * sectorsPerReconUnit +
212 1.12 oster sectorsPerReconUnit - 1)) {
213 1.4 oster mapPtr->status[i] = RU_ALL;
214 1.4 oster mapPtr->unitsLeft--;
215 1.15 oster FreeReconMapListElem(mapPtr, p);
216 1.4 oster }
217 1.4 oster }
218 1.4 oster
219 1.26 perry static void
220 1.22 oster crunch_list(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *listPtr)
221 1.4 oster {
222 1.4 oster RF_ReconMapListElem_t *pt, *p = listPtr;
223 1.4 oster
224 1.4 oster if (!p)
225 1.4 oster return;
226 1.4 oster pt = p;
227 1.4 oster p = p->next;
228 1.4 oster while (p) {
229 1.4 oster if (pt->stopSector >= p->startSector - 1) {
230 1.4 oster pt->stopSector = RF_MAX(pt->stopSector, p->stopSector);
231 1.4 oster pt->next = p->next;
232 1.15 oster FreeReconMapListElem(mapPtr, p);
233 1.4 oster p = pt->next;
234 1.4 oster } else {
235 1.4 oster pt = p;
236 1.4 oster p = p->next;
237 1.4 oster }
238 1.4 oster }
239 1.1 oster }
240 1.12 oster /*---------------------------------------------------------------------------
241 1.4 oster *
242 1.1 oster * Allocate and fill a new list element
243 1.1 oster *
244 1.12 oster *-------------------------------------------------------------------------*/
245 1.1 oster
246 1.4 oster static RF_ReconMapListElem_t *
247 1.22 oster MakeReconMapListElem(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector,
248 1.22 oster RF_SectorNum_t stopSector, RF_ReconMapListElem_t *next)
249 1.4 oster {
250 1.4 oster RF_ReconMapListElem_t *p;
251 1.4 oster
252 1.19 oster p = pool_get(&mapPtr->elem_pool, PR_WAITOK);
253 1.4 oster p->startSector = startSector;
254 1.4 oster p->stopSector = stopSector;
255 1.4 oster p->next = next;
256 1.4 oster return (p);
257 1.1 oster }
258 1.12 oster /*---------------------------------------------------------------------------
259 1.4 oster *
260 1.1 oster * Free a list element
261 1.1 oster *
262 1.12 oster *-------------------------------------------------------------------------*/
263 1.1 oster
264 1.26 perry static void
265 1.22 oster FreeReconMapListElem(RF_ReconMap_t *mapPtr, RF_ReconMapListElem_t *p)
266 1.4 oster {
267 1.15 oster pool_put(&mapPtr->elem_pool, p);
268 1.1 oster }
269 1.12 oster /*---------------------------------------------------------------------------
270 1.4 oster *
271 1.12 oster * Free an entire status structure. Inefficient, but can be called at
272 1.12 oster * any time.
273 1.1 oster *
274 1.12 oster *-------------------------------------------------------------------------*/
275 1.26 perry void
276 1.22 oster rf_FreeReconMap(RF_ReconMap_t *mapPtr)
277 1.4 oster {
278 1.4 oster RF_ReconMapListElem_t *p, *q;
279 1.4 oster RF_ReconUnitCount_t numRUs;
280 1.4 oster RF_ReconUnitNum_t i;
281 1.4 oster
282 1.4 oster numRUs = mapPtr->sectorsInDisk / mapPtr->sectorsPerReconUnit;
283 1.4 oster if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
284 1.4 oster numRUs++;
285 1.4 oster
286 1.4 oster for (i = 0; i < numRUs; i++) {
287 1.4 oster p = mapPtr->status[i];
288 1.4 oster while (p != RU_NOTHING && p != RU_ALL) {
289 1.4 oster q = p;
290 1.4 oster p = p->next;
291 1.4 oster RF_Free(q, sizeof(*q));
292 1.4 oster }
293 1.4 oster }
294 1.15 oster pool_destroy(&mapPtr->elem_pool);
295 1.26 perry RF_Free(mapPtr->status, mapPtr->totalRUs *
296 1.12 oster sizeof(RF_ReconMapListElem_t *));
297 1.4 oster RF_Free(mapPtr, sizeof(RF_ReconMap_t));
298 1.1 oster }
299 1.12 oster /*---------------------------------------------------------------------------
300 1.1 oster *
301 1.1 oster * returns nonzero if the indicated RU has been reconstructed already
302 1.1 oster *
303 1.12 oster *-------------------------------------------------------------------------*/
304 1.1 oster
305 1.26 perry int
306 1.22 oster rf_CheckRUReconstructed(RF_ReconMap_t *mapPtr, RF_SectorNum_t startSector)
307 1.1 oster {
308 1.4 oster RF_ReconMapListElem_t *l; /* used for searching */
309 1.4 oster RF_ReconUnitNum_t i;
310 1.1 oster
311 1.4 oster i = startSector / mapPtr->sectorsPerReconUnit;
312 1.4 oster l = mapPtr->status[i];
313 1.4 oster return ((l == RU_ALL) ? 1 : 0);
314 1.1 oster }
315 1.1 oster
316 1.26 perry RF_ReconUnitCount_t
317 1.22 oster rf_UnitsLeftToReconstruct(RF_ReconMap_t *mapPtr)
318 1.1 oster {
319 1.4 oster RF_ASSERT(mapPtr != NULL);
320 1.4 oster return (mapPtr->unitsLeft);
321 1.4 oster }
322 1.4 oster
323 1.10 oster #if 0
324 1.26 perry static void
325 1.22 oster PrintList(RF_ReconMapListElem_t *listPtr)
326 1.4 oster {
327 1.4 oster while (listPtr) {
328 1.26 perry printf("%d,%d -> ", (int) listPtr->startSector,
329 1.12 oster (int) listPtr->stopSector);
330 1.4 oster listPtr = listPtr->next;
331 1.4 oster }
332 1.4 oster printf("\n");
333 1.4 oster }
334 1.4 oster
335 1.26 perry void
336 1.22 oster rf_PrintReconMap(RF_Raid_t *raidPtr, RF_ReconMap_t *mapPtr, RF_RowCol_t fcol)
337 1.4 oster {
338 1.4 oster RF_ReconUnitCount_t numRUs;
339 1.4 oster RF_ReconMapListElem_t *p;
340 1.4 oster RF_ReconUnitNum_t i;
341 1.4 oster
342 1.4 oster numRUs = mapPtr->totalRUs;
343 1.4 oster if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
344 1.4 oster numRUs++;
345 1.4 oster
346 1.4 oster for (i = 0; i < numRUs; i++) {
347 1.4 oster p = mapPtr->status[i];
348 1.4 oster if (p == RU_ALL)/* printf("[%d] ALL\n",i) */
349 1.4 oster ;
350 1.4 oster else
351 1.4 oster if (p == RU_NOTHING) {
352 1.4 oster printf("%d: Unreconstructed\n", i);
353 1.4 oster } else {
354 1.4 oster printf("%d: ", i);
355 1.4 oster PrintList(p);
356 1.4 oster }
357 1.4 oster }
358 1.4 oster }
359 1.10 oster #endif
360 1.4 oster
361 1.11 oster #if RF_DEBUG_RECON
362 1.26 perry void
363 1.22 oster rf_PrintReconSchedule(RF_ReconMap_t *mapPtr, struct timeval *starttime)
364 1.4 oster {
365 1.4 oster static int old_pctg = -1;
366 1.4 oster struct timeval tv, diff;
367 1.4 oster int new_pctg;
368 1.4 oster
369 1.26 perry new_pctg = 100 - (rf_UnitsLeftToReconstruct(mapPtr) *
370 1.12 oster 100 / mapPtr->totalRUs);
371 1.4 oster if (new_pctg != old_pctg) {
372 1.4 oster RF_GETTIME(tv);
373 1.4 oster RF_TIMEVAL_DIFF(starttime, &tv, &diff);
374 1.26 perry printf("%d %d.%06d\n", (int) new_pctg, (int) diff.tv_sec,
375 1.12 oster (int) diff.tv_usec);
376 1.4 oster old_pctg = new_pctg;
377 1.4 oster }
378 1.1 oster }
379 1.11 oster #endif
380 1.11 oster
381