rf_reconmap.c revision 1.9 1 1.9 oster /* $NetBSD: rf_reconmap.c,v 1.9 2002/09/14 17:53:58 oster 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 1.9 oster __KERNEL_RCSID(0, "$NetBSD: rf_reconmap.c,v 1.9 2002/09/14 17:53:58 oster 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 1.1 oster /* used to mark the end of the list */
55 1.1 oster #define RU_NIL ((RF_ReconMapListElem_t *) 0)
56 1.1 oster
57 1.1 oster
58 1.4 oster static void
59 1.4 oster compact_stat_entry(RF_Raid_t * raidPtr, RF_ReconMap_t * mapPtr,
60 1.4 oster int i);
61 1.4 oster static void crunch_list(RF_ReconMap_t * mapPtr, RF_ReconMapListElem_t * listPtr);
62 1.4 oster static RF_ReconMapListElem_t *
63 1.4 oster MakeReconMapListElem(RF_SectorNum_t startSector,
64 1.4 oster RF_SectorNum_t stopSector, RF_ReconMapListElem_t * next);
65 1.4 oster static void
66 1.4 oster FreeReconMapListElem(RF_ReconMap_t * mapPtr,
67 1.4 oster RF_ReconMapListElem_t * p);
68 1.4 oster static void update_size(RF_ReconMap_t * mapPtr, int size);
69 1.4 oster static void PrintList(RF_ReconMapListElem_t * listPtr);
70 1.1 oster
71 1.1 oster /*-----------------------------------------------------------------------------
72 1.1 oster *
73 1.4 oster * Creates and initializes new Reconstruction map
74 1.1 oster *
75 1.1 oster *-----------------------------------------------------------------------------*/
76 1.1 oster
77 1.4 oster RF_ReconMap_t *
78 1.4 oster rf_MakeReconMap(raidPtr, ru_sectors, disk_sectors, spareUnitsPerDisk)
79 1.4 oster RF_Raid_t *raidPtr;
80 1.4 oster RF_SectorCount_t ru_sectors; /* size of reconstruction unit in
81 1.4 oster * sectors */
82 1.4 oster RF_SectorCount_t disk_sectors; /* size of disk in sectors */
83 1.4 oster RF_ReconUnitCount_t spareUnitsPerDisk; /* zero unless distributed
84 1.4 oster * sparing */
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 int rc;
90 1.4 oster
91 1.4 oster RF_Malloc(p, sizeof(RF_ReconMap_t), (RF_ReconMap_t *));
92 1.4 oster p->sectorsPerReconUnit = ru_sectors;
93 1.4 oster p->sectorsInDisk = disk_sectors;
94 1.4 oster
95 1.4 oster p->totalRUs = num_rus;
96 1.4 oster p->spareRUs = spareUnitsPerDisk;
97 1.4 oster p->unitsLeft = num_rus - spareUnitsPerDisk;
98 1.4 oster
99 1.4 oster RF_Malloc(p->status, num_rus * sizeof(RF_ReconMapListElem_t *), (RF_ReconMapListElem_t **));
100 1.4 oster RF_ASSERT(p->status != (RF_ReconMapListElem_t **) NULL);
101 1.4 oster
102 1.7 thorpej (void) memset((char *) p->status, 0,
103 1.7 thorpej num_rus * sizeof(RF_ReconMapListElem_t *));
104 1.4 oster
105 1.4 oster p->size = sizeof(RF_ReconMap_t) + num_rus * sizeof(RF_ReconMapListElem_t *);
106 1.4 oster p->maxSize = p->size;
107 1.4 oster
108 1.4 oster rc = rf_mutex_init(&p->mutex);
109 1.4 oster if (rc) {
110 1.9 oster rf_print_unable_to_init_mutex(__FILE__, __LINE__, rc);
111 1.4 oster RF_Free(p->status, num_rus * sizeof(RF_ReconMapListElem_t *));
112 1.4 oster RF_Free(p, sizeof(RF_ReconMap_t));
113 1.4 oster return (NULL);
114 1.4 oster }
115 1.4 oster return (p);
116 1.1 oster }
117 1.1 oster
118 1.1 oster
119 1.1 oster /*-----------------------------------------------------------------------------
120 1.1 oster *
121 1.1 oster * marks a new set of sectors as reconstructed. All the possible mergings get
122 1.1 oster * complicated. To simplify matters, the approach I take is to just dump
123 1.1 oster * something into the list, and then clean it up (i.e. merge elements and
124 1.1 oster * eliminate redundant ones) in a second pass over the list (compact_stat_entry()).
125 1.1 oster * Not 100% efficient, since a structure can be allocated and then immediately
126 1.1 oster * freed, but it keeps this code from becoming (more of) a nightmare of
127 1.1 oster * special cases. The only thing that compact_stat_entry() assumes is that the
128 1.1 oster * list is sorted by startSector, and so this is the only condition I maintain
129 1.1 oster * here. (MCH)
130 1.1 oster *
131 1.1 oster *-----------------------------------------------------------------------------*/
132 1.1 oster
133 1.4 oster void
134 1.4 oster rf_ReconMapUpdate(raidPtr, mapPtr, startSector, stopSector)
135 1.4 oster RF_Raid_t *raidPtr;
136 1.4 oster RF_ReconMap_t *mapPtr;
137 1.4 oster RF_SectorNum_t startSector;
138 1.4 oster 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 1.6 oster RF_ASSERT(startSector >= 0 && stopSector < mapPtr->sectorsInDisk && stopSector >= startSector);
146 1.4 oster
147 1.4 oster while (startSector <= stopSector) {
148 1.4 oster i = startSector / mapPtr->sectorsPerReconUnit;
149 1.4 oster first_in_RU = i * sectorsPerReconUnit;
150 1.4 oster last_in_RU = first_in_RU + sectorsPerReconUnit - 1;
151 1.4 oster p = mapPtr->status[i];
152 1.4 oster if (p != RU_ALL) {
153 1.4 oster if (p == RU_NOTHING || p->startSector > startSector) { /* insert at front of
154 1.4 oster * list */
155 1.4 oster
156 1.4 oster mapPtr->status[i] = MakeReconMapListElem(startSector, RF_MIN(stopSector, last_in_RU), (p == RU_NOTHING) ? NULL : p);
157 1.4 oster update_size(mapPtr, sizeof(RF_ReconMapListElem_t));
158 1.4 oster
159 1.4 oster } else {/* general case */
160 1.4 oster do { /* search for place to insert */
161 1.4 oster pt = p;
162 1.4 oster p = p->next;
163 1.4 oster } while (p && (p->startSector < startSector));
164 1.4 oster pt->next = MakeReconMapListElem(startSector, RF_MIN(stopSector, last_in_RU), p);
165 1.4 oster update_size(mapPtr, sizeof(RF_ReconMapListElem_t));
166 1.4 oster }
167 1.4 oster compact_stat_entry(raidPtr, mapPtr, i);
168 1.4 oster }
169 1.4 oster startSector = RF_MIN(stopSector, last_in_RU) + 1;
170 1.4 oster }
171 1.4 oster RF_UNLOCK_MUTEX(mapPtr->mutex);
172 1.1 oster }
173 1.1 oster
174 1.1 oster
175 1.1 oster
176 1.1 oster /*-----------------------------------------------------------------------------
177 1.1 oster *
178 1.1 oster * performs whatever list compactions can be done, and frees any space
179 1.1 oster * that is no longer necessary. Assumes only that the list is sorted
180 1.1 oster * by startSector. crunch_list() compacts a single list as much as possible,
181 1.1 oster * and the second block of code deletes the entire list if possible.
182 1.1 oster * crunch_list() is also called from MakeReconMapAccessList().
183 1.1 oster *
184 1.1 oster * When a recon unit is detected to be fully reconstructed, we set the
185 1.1 oster * corresponding bit in the parity stripe map so that the head follow
186 1.1 oster * code will not select this parity stripe again. This is redundant (but
187 1.1 oster * harmless) when compact_stat_entry is called from the reconstruction code,
188 1.1 oster * but necessary when called from the user-write code.
189 1.1 oster *
190 1.1 oster *-----------------------------------------------------------------------------*/
191 1.1 oster
192 1.4 oster static void
193 1.4 oster compact_stat_entry(raidPtr, mapPtr, i)
194 1.4 oster RF_Raid_t *raidPtr;
195 1.4 oster RF_ReconMap_t *mapPtr;
196 1.4 oster int i;
197 1.4 oster {
198 1.4 oster RF_SectorCount_t sectorsPerReconUnit = mapPtr->sectorsPerReconUnit;
199 1.4 oster RF_ReconMapListElem_t *p = mapPtr->status[i];
200 1.4 oster
201 1.4 oster crunch_list(mapPtr, p);
202 1.4 oster
203 1.4 oster if ((p->startSector == i * sectorsPerReconUnit) &&
204 1.4 oster (p->stopSector == i * sectorsPerReconUnit + sectorsPerReconUnit - 1)) {
205 1.4 oster mapPtr->status[i] = RU_ALL;
206 1.4 oster mapPtr->unitsLeft--;
207 1.4 oster FreeReconMapListElem(mapPtr, p);
208 1.4 oster }
209 1.4 oster }
210 1.4 oster
211 1.4 oster static void
212 1.4 oster crunch_list(mapPtr, listPtr)
213 1.4 oster RF_ReconMap_t *mapPtr;
214 1.4 oster RF_ReconMapListElem_t *listPtr;
215 1.4 oster {
216 1.4 oster RF_ReconMapListElem_t *pt, *p = listPtr;
217 1.4 oster
218 1.4 oster if (!p)
219 1.4 oster return;
220 1.4 oster pt = p;
221 1.4 oster p = p->next;
222 1.4 oster while (p) {
223 1.4 oster if (pt->stopSector >= p->startSector - 1) {
224 1.4 oster pt->stopSector = RF_MAX(pt->stopSector, p->stopSector);
225 1.4 oster pt->next = p->next;
226 1.4 oster FreeReconMapListElem(mapPtr, p);
227 1.4 oster p = pt->next;
228 1.4 oster } else {
229 1.4 oster pt = p;
230 1.4 oster p = p->next;
231 1.4 oster }
232 1.4 oster }
233 1.1 oster }
234 1.1 oster /*-----------------------------------------------------------------------------
235 1.4 oster *
236 1.1 oster * Allocate and fill a new list element
237 1.1 oster *
238 1.1 oster *-----------------------------------------------------------------------------*/
239 1.1 oster
240 1.4 oster static RF_ReconMapListElem_t *
241 1.4 oster MakeReconMapListElem(
242 1.4 oster RF_SectorNum_t startSector,
243 1.4 oster RF_SectorNum_t stopSector,
244 1.4 oster RF_ReconMapListElem_t * next)
245 1.4 oster {
246 1.4 oster RF_ReconMapListElem_t *p;
247 1.4 oster
248 1.4 oster RF_Malloc(p, sizeof(RF_ReconMapListElem_t), (RF_ReconMapListElem_t *));
249 1.4 oster if (p == NULL)
250 1.4 oster return (NULL);
251 1.4 oster p->startSector = startSector;
252 1.4 oster p->stopSector = stopSector;
253 1.4 oster p->next = next;
254 1.4 oster return (p);
255 1.1 oster }
256 1.1 oster /*-----------------------------------------------------------------------------
257 1.4 oster *
258 1.1 oster * Free a list element
259 1.1 oster *
260 1.1 oster *-----------------------------------------------------------------------------*/
261 1.1 oster
262 1.4 oster static void
263 1.4 oster FreeReconMapListElem(mapPtr, p)
264 1.4 oster RF_ReconMap_t *mapPtr;
265 1.4 oster RF_ReconMapListElem_t *p;
266 1.4 oster {
267 1.4 oster int delta;
268 1.4 oster
269 1.4 oster if (mapPtr) {
270 1.4 oster delta = 0 - (int) sizeof(RF_ReconMapListElem_t);
271 1.4 oster update_size(mapPtr, delta);
272 1.4 oster }
273 1.4 oster RF_Free(p, sizeof(*p));
274 1.1 oster }
275 1.1 oster /*-----------------------------------------------------------------------------
276 1.4 oster *
277 1.1 oster * Free an entire status structure. Inefficient, but can be called at any time.
278 1.1 oster *
279 1.1 oster *-----------------------------------------------------------------------------*/
280 1.4 oster void
281 1.4 oster rf_FreeReconMap(mapPtr)
282 1.4 oster RF_ReconMap_t *mapPtr;
283 1.4 oster {
284 1.4 oster RF_ReconMapListElem_t *p, *q;
285 1.4 oster RF_ReconUnitCount_t numRUs;
286 1.4 oster RF_ReconUnitNum_t i;
287 1.4 oster
288 1.4 oster numRUs = mapPtr->sectorsInDisk / mapPtr->sectorsPerReconUnit;
289 1.4 oster if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
290 1.4 oster numRUs++;
291 1.4 oster
292 1.4 oster for (i = 0; i < numRUs; i++) {
293 1.4 oster p = mapPtr->status[i];
294 1.4 oster while (p != RU_NOTHING && p != RU_ALL) {
295 1.4 oster q = p;
296 1.4 oster p = p->next;
297 1.4 oster RF_Free(q, sizeof(*q));
298 1.4 oster }
299 1.4 oster }
300 1.4 oster rf_mutex_destroy(&mapPtr->mutex);
301 1.4 oster RF_Free(mapPtr->status, mapPtr->totalRUs * sizeof(RF_ReconMapListElem_t *));
302 1.4 oster RF_Free(mapPtr, sizeof(RF_ReconMap_t));
303 1.1 oster }
304 1.1 oster /*-----------------------------------------------------------------------------
305 1.1 oster *
306 1.1 oster * returns nonzero if the indicated RU has been reconstructed already
307 1.1 oster *
308 1.1 oster *---------------------------------------------------------------------------*/
309 1.1 oster
310 1.4 oster int
311 1.4 oster rf_CheckRUReconstructed(mapPtr, startSector)
312 1.4 oster RF_ReconMap_t *mapPtr;
313 1.4 oster RF_SectorNum_t startSector;
314 1.1 oster {
315 1.4 oster RF_ReconMapListElem_t *l; /* used for searching */
316 1.4 oster RF_ReconUnitNum_t i;
317 1.1 oster
318 1.4 oster i = startSector / mapPtr->sectorsPerReconUnit;
319 1.4 oster l = mapPtr->status[i];
320 1.4 oster return ((l == RU_ALL) ? 1 : 0);
321 1.1 oster }
322 1.1 oster
323 1.4 oster RF_ReconUnitCount_t
324 1.4 oster rf_UnitsLeftToReconstruct(mapPtr)
325 1.4 oster RF_ReconMap_t *mapPtr;
326 1.1 oster {
327 1.4 oster RF_ASSERT(mapPtr != NULL);
328 1.4 oster return (mapPtr->unitsLeft);
329 1.1 oster }
330 1.1 oster /* updates the size fields of a status descriptor */
331 1.4 oster static void
332 1.4 oster update_size(mapPtr, size)
333 1.4 oster RF_ReconMap_t *mapPtr;
334 1.4 oster int size;
335 1.4 oster {
336 1.4 oster mapPtr->size += size;
337 1.4 oster mapPtr->maxSize = RF_MAX(mapPtr->size, mapPtr->maxSize);
338 1.4 oster }
339 1.4 oster
340 1.4 oster static void
341 1.4 oster PrintList(listPtr)
342 1.4 oster RF_ReconMapListElem_t *listPtr;
343 1.4 oster {
344 1.4 oster while (listPtr) {
345 1.4 oster printf("%d,%d -> ", (int) listPtr->startSector, (int) listPtr->stopSector);
346 1.4 oster listPtr = listPtr->next;
347 1.4 oster }
348 1.4 oster printf("\n");
349 1.4 oster }
350 1.4 oster
351 1.4 oster void
352 1.4 oster rf_PrintReconMap(raidPtr, mapPtr, frow, fcol)
353 1.4 oster RF_Raid_t *raidPtr;
354 1.4 oster RF_ReconMap_t *mapPtr;
355 1.4 oster RF_RowCol_t frow;
356 1.4 oster RF_RowCol_t fcol;
357 1.4 oster {
358 1.4 oster RF_ReconUnitCount_t numRUs;
359 1.4 oster RF_ReconMapListElem_t *p;
360 1.4 oster RF_ReconUnitNum_t i;
361 1.4 oster
362 1.4 oster numRUs = mapPtr->totalRUs;
363 1.4 oster if (mapPtr->sectorsInDisk % mapPtr->sectorsPerReconUnit)
364 1.4 oster numRUs++;
365 1.4 oster
366 1.4 oster for (i = 0; i < numRUs; i++) {
367 1.4 oster p = mapPtr->status[i];
368 1.4 oster if (p == RU_ALL)/* printf("[%d] ALL\n",i) */
369 1.4 oster ;
370 1.4 oster else
371 1.4 oster if (p == RU_NOTHING) {
372 1.4 oster printf("%d: Unreconstructed\n", i);
373 1.4 oster } else {
374 1.4 oster printf("%d: ", i);
375 1.4 oster PrintList(p);
376 1.4 oster }
377 1.4 oster }
378 1.4 oster }
379 1.4 oster
380 1.4 oster void
381 1.4 oster rf_PrintReconSchedule(mapPtr, starttime)
382 1.4 oster RF_ReconMap_t *mapPtr;
383 1.4 oster 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 1.4 oster new_pctg = 100 - (rf_UnitsLeftToReconstruct(mapPtr) * 100 / mapPtr->totalRUs);
390 1.4 oster if (new_pctg != old_pctg) {
391 1.4 oster RF_GETTIME(tv);
392 1.4 oster RF_TIMEVAL_DIFF(starttime, &tv, &diff);
393 1.4 oster printf("%d %d.%06d\n", (int) new_pctg, (int) diff.tv_sec, (int) diff.tv_usec);
394 1.4 oster old_pctg = new_pctg;
395 1.4 oster }
396 1.1 oster }
397