]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Assemble.c
Fix assembling of array with spares when verion-1 superblock is used.
[thirdparty/mdadm.git] / Assemble.c
1 /*
2 * mdadm - manage Linux "md" devices aka RAID arrays.
3 *
4 * Copyright (C) 2001-2002 Neil Brown <neilb@cse.unsw.edu.au>
5 *
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License as published by
9 * the Free Software Foundation; either version 2 of the License, or
10 * (at your option) any later version.
11 *
12 * This program is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
16 *
17 * You should have received a copy of the GNU General Public License
18 * along with this program; if not, write to the Free Software
19 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 *
21 * Author: Neil Brown
22 * Email: <neilb@cse.unsw.edu.au>
23 * Paper: Neil Brown
24 * School of Computer Science and Engineering
25 * The University of New South Wales
26 * Sydney, 2052
27 * Australia
28 */
29
30 #include "mdadm.h"
31
32 int Assemble(struct supertype *st, char *mddev, int mdfd,
33 mddev_ident_t ident, char *conffile,
34 mddev_dev_t devlist,
35 int readonly, int runstop,
36 char *update,
37 int verbose, int force)
38 {
39 /*
40 * The task of Assemble is to find a collection of
41 * devices that should (according to their superblocks)
42 * form an array, and to give this collection to the MD driver.
43 * In Linux-2.4 and later, this involves submitting a
44 * SET_ARRAY_INFO ioctl with no arg - to prepare
45 * the array - and then submit a number of
46 * ADD_NEW_DISK ioctls to add disks into
47 * the array. Finally RUN_ARRAY might
48 * be submitted to start the array.
49 *
50 * Much of the work of Assemble is in finding and/or
51 * checking the disks to make sure they look right.
52 *
53 * If mddev is not set, then scan must be set and we
54 * read through the config file for dev+uuid mapping
55 * We recurse, setting mddev, for each device that
56 * - isn't running
57 * - has a valid uuid (or any uuid if !uuidset)
58 *
59 * If mddev is set, we try to determine state of md.
60 * check version - must be at least 0.90.0
61 * check kernel version. must be at least 2.4.
62 * If not, we can possibly fall back on START_ARRAY
63 * Try to GET_ARRAY_INFO.
64 * If possible, give up
65 * If not, try to STOP_ARRAY just to make sure
66 *
67 * If !uuidset and scan, look in conf-file for uuid
68 * If not found, give up
69 * If !devlist and scan and uuidset, get list of devs from conf-file
70 *
71 * For each device:
72 * Check superblock - discard if bad
73 * Check uuid (set if we don't have one) - discard if no match
74 * Check superblock similarity if we have a superblock - discard if different
75 * Record events, devicenum
76 * This should give us a list of devices for the array
77 * We should collect the most recent event number
78 *
79 * Count disks with recent enough event count
80 * While force && !enough disks
81 * Choose newest rejected disks, update event count
82 * mark clean and rewrite superblock
83 * If recent kernel:
84 * SET_ARRAY_INFO
85 * foreach device with recent events : ADD_NEW_DISK
86 * if runstop == 1 || "enough" disks and runstop==0 -> RUN_ARRAY
87 * If old kernel:
88 * Check the device numbers in superblock are right
89 * update superblock if any changes
90 * START_ARRAY
91 *
92 */
93 int old_linux = 0;
94 int vers;
95 void *first_super = NULL, *super = NULL;
96 struct {
97 char *devname;
98 unsigned int major, minor;
99 unsigned int oldmajor, oldminor;
100 long long events;
101 int uptodate;
102 int state;
103 int raid_disk;
104 int disk_nr;
105 } *devices;
106 int *best = NULL; /* indexed by raid_disk */
107 unsigned int bestcnt = 0;
108 int devcnt = 0;
109 unsigned int okcnt, sparecnt;
110 unsigned int req_cnt;
111 unsigned int i;
112 int most_recent = 0;
113 int chosen_drive;
114 int change = 0;
115 int inargv = 0;
116 int start_partial_ok = force || devlist==NULL;
117 unsigned int num_devs;
118 mddev_dev_t tmpdev;
119 struct mdinfo info;
120 struct mddev_ident_s ident2;
121 char *avail;
122 int nextspare = 0;
123
124 vers = md_get_version(mdfd);
125 if (vers <= 0) {
126 fprintf(stderr, Name ": %s appears not to be an md device.\n", mddev);
127 return 1;
128 }
129 if (vers < 9000) {
130 fprintf(stderr, Name ": Assemble requires driver version 0.90.0 or later.\n"
131 " Upgrade your kernel or try --build\n");
132 return 1;
133 }
134 if (get_linux_version() < 2004000)
135 old_linux = 1;
136
137 if (ioctl(mdfd, GET_ARRAY_INFO, &info.array)>=0) {
138 fprintf(stderr, Name ": device %s already active - cannot assemble it\n",
139 mddev);
140 return 1;
141 }
142 ioctl(mdfd, STOP_ARRAY, NULL); /* just incase it was started but has no content */
143
144 /*
145 * If any subdevs are listed, then any that don't
146 * match ident are discarded. Remainder must all match and
147 * become the array.
148 * If no subdevs, then we scan all devices in the config file, but
149 * there must be something in the identity
150 */
151
152 if (!devlist &&
153 ident->uuid_set == 0 &&
154 ident->super_minor < 0 &&
155 ident->devices == NULL) {
156 fprintf(stderr, Name ": No identity information available for %s - cannot assemble.\n",
157 mddev);
158 return 1;
159 }
160 if (devlist == NULL)
161 devlist = conf_get_devs(conffile);
162 else inargv = 1;
163
164 tmpdev = devlist; num_devs = 0;
165 while (tmpdev) {
166 num_devs++;
167 tmpdev = tmpdev->next;
168 }
169 devices = malloc(num_devs * sizeof(*devices));
170
171 if (!st && ident->st) st = ident->st;
172
173 if (verbose>0)
174 fprintf(stderr, Name ": looking for devices for %s\n",
175 mddev);
176
177 while ( devlist) {
178 char *devname;
179 int dfd;
180 struct stat stb;
181 struct supertype *tst = st;
182
183 devname = devlist->devname;
184 devlist = devlist->next;
185
186 if (ident->devices &&
187 !match_oneof(ident->devices, devname)) {
188 if ((inargv && verbose>=0) || verbose > 0)
189 fprintf(stderr, Name ": %s is not one of %s\n", devname, ident->devices);
190 continue;
191 }
192
193 if (super) {
194 free(super);
195 super = NULL;
196 }
197
198 dfd = open(devname, O_RDONLY|O_EXCL, 0);
199 if (dfd < 0) {
200 if ((inargv && verbose >= 0) || verbose > 0)
201 fprintf(stderr, Name ": cannot open device %s: %s\n",
202 devname, strerror(errno));
203 } else if (fstat(dfd, &stb)< 0) {
204 /* Impossible! */
205 fprintf(stderr, Name ": fstat failed for %s: %s\n",
206 devname, strerror(errno));
207 close(dfd);
208 } else if ((stb.st_mode & S_IFMT) != S_IFBLK) {
209 fprintf(stderr, Name ": %s is not a block device.\n",
210 devname);
211 close(dfd);
212 } else if (!tst && (tst = guess_super(dfd)) == NULL) {
213 if ((inargv && verbose >= 0) || verbose > 0)
214 fprintf(stderr, Name ": no recogniseable superblock\n");
215 } else if (tst->ss->load_super(tst,dfd, &super, NULL)) {
216 if ((inargv && verbose >= 0) || verbose > 0)
217 fprintf( stderr, Name ": no RAID superblock on %s\n",
218 devname);
219 close(dfd);
220 } else {
221 tst->ss->getinfo_super(&info, &ident2, super);
222 close(dfd);
223 }
224
225 if (ident->uuid_set &&
226 (!super || same_uuid(info.uuid, ident->uuid, tst->ss->swapuuid)==0)) {
227 if ((inargv && verbose >= 0) || verbose > 0)
228 fprintf(stderr, Name ": %s has wrong uuid.\n",
229 devname);
230 continue;
231 }
232 if (ident->name[0] &&
233 (!super || strncmp(ident2.name, ident->name, 32)!=0)) {
234 if ((inargv && verbose >= 0) || verbose > 0)
235 fprintf(stderr, Name ": %s has wrong name.\n",
236 devname);
237 continue;
238 }
239 if (ident->super_minor != UnSet &&
240 (!super || ident->super_minor != info.array.md_minor)) {
241 if ((inargv && verbose >= 0) || verbose > 0)
242 fprintf(stderr, Name ": %s has wrong super-minor.\n",
243 devname);
244 continue;
245 }
246 if (ident->level != UnSet &&
247 (!super|| ident->level != info.array.level)) {
248 if ((inargv && verbose >= 0) || verbose > 0)
249 fprintf(stderr, Name ": %s has wrong raid level.\n",
250 devname);
251 continue;
252 }
253 if (ident->raid_disks != UnSet &&
254 (!super || ident->raid_disks!= info.array.raid_disks)) {
255 if ((inargv && verbose >= 0) || verbose > 0)
256 fprintf(stderr, Name ": %s requires wrong number of drives.\n",
257 devname);
258 continue;
259 }
260
261 /* If we are this far, then we are commited to this device.
262 * If the super_block doesn't exist, or doesn't match others,
263 * then we cannot continue
264 */
265
266 if (!super) {
267 fprintf(stderr, Name ": %s has no superblock - assembly aborted\n",
268 devname);
269 free(first_super);
270 return 1;
271 }
272 st = tst; /* commit to this format, if haven't already */
273 if (st->ss->compare_super(&first_super, super)) {
274 fprintf(stderr, Name ": superblock on %s doesn't match others - assembly aborted\n",
275 devname);
276 free(super);
277 free(first_super);
278 return 1;
279 }
280
281 /* looks like a good enough match to update the super block if needed */
282 if (update) {
283 /* prepare useful information in info structures */
284 struct stat stb2;
285 fstat(mdfd, &stb2);
286 info.array.md_minor = minor(stb2.st_rdev);
287
288 st->ss->update_super(&info, super, update, devname, verbose);
289
290 dfd = open(devname, O_RDWR|O_EXCL, 0);
291 if (dfd < 0)
292 fprintf(stderr, Name ": Cannot open %s for superblock update\n",
293 devname);
294 else if (st->ss->store_super(st, dfd, super))
295 fprintf(stderr, Name ": Could not re-write superblock on %s.\n",
296 devname);
297 if (dfd >= 0)
298 close(dfd);
299 }
300
301 if (verbose > 0)
302 fprintf(stderr, Name ": %s is identified as a member of %s, slot %d.\n",
303 devname, mddev, info.disk.raid_disk);
304 devices[devcnt].devname = devname;
305 devices[devcnt].major = major(stb.st_rdev);
306 devices[devcnt].minor = minor(stb.st_rdev);
307 devices[devcnt].oldmajor = info.disk.major;
308 devices[devcnt].oldminor = info.disk.minor;
309 devices[devcnt].events = info.events;
310 devices[devcnt].raid_disk = info.disk.raid_disk;
311 devices[devcnt].disk_nr = info.disk.number;
312 devices[devcnt].uptodate = 0;
313 devices[devcnt].state = info.disk.state;
314 if (most_recent < devcnt) {
315 if (devices[devcnt].events
316 > devices[most_recent].events)
317 most_recent = devcnt;
318 }
319 if (info.array.level == -4)
320 /* with multipath, the raid_disk from the superblock is meaningless */
321 i = devcnt;
322 else
323 i = devices[devcnt].raid_disk;
324 if (i+1 == 0) {
325 if (nextspare < info.array.raid_disks)
326 nextspare = info.array.raid_disks;
327 i = nextspare++;
328 }
329 if (i < 10000) {
330 if (i >= bestcnt) {
331 unsigned int newbestcnt = i+10;
332 int *newbest = malloc(sizeof(int)*newbestcnt);
333 unsigned int c;
334 for (c=0; c < newbestcnt; c++)
335 if (c < bestcnt)
336 newbest[c] = best[c];
337 else
338 newbest[c] = -1;
339 if (best)free(best);
340 best = newbest;
341 bestcnt = newbestcnt;
342 }
343 if (best[i] == -1
344 || devices[best[i]].events < devices[devcnt].events)
345 best[i] = devcnt;
346 }
347 devcnt++;
348 }
349
350 if (super)
351 free(super);
352 super = NULL;
353
354 if (update && strcmp(update, "byteorder")==0)
355 st->minor_version = 90;
356
357 if (devcnt == 0) {
358 fprintf(stderr, Name ": no devices found for %s\n",
359 mddev);
360 free(first_super);
361 return 1;
362 }
363
364 st->ss->getinfo_super(&info, &ident2, first_super);
365
366 /* now we have some devices that might be suitable.
367 * I wonder how many
368 */
369 avail = malloc(info.array.raid_disks);
370 memset(avail, 0, info.array.raid_disks);
371 okcnt = 0;
372 sparecnt=0;
373 for (i=0; i< bestcnt ;i++) {
374 int j = best[i];
375 int event_margin = !force;
376 if (j < 0) continue;
377 /* note: we ignore error flags in multipath arrays
378 * as they don't make sense
379 */
380 if (info.array.level != -4)
381 if (!(devices[j].state & (1<<MD_DISK_SYNC))) {
382 if (!(devices[j].state & (1<<MD_DISK_FAULTY)))
383 sparecnt++;
384 continue;
385 }
386 if (devices[j].events+event_margin >=
387 devices[most_recent].events) {
388 devices[j].uptodate = 1;
389 if (i < info.array.raid_disks) {
390 okcnt++;
391 avail[i]=1;
392 } else
393 sparecnt++;
394 }
395 }
396 while (force && !enough(info.array.level, info.array.raid_disks,
397 info.array.layout,
398 avail, okcnt)) {
399 /* Choose the newest best drive which is
400 * not up-to-date, update the superblock
401 * and add it.
402 */
403 int fd;
404 chosen_drive = -1;
405 for (i=0; i<info.array.raid_disks && i < bestcnt; i++) {
406 int j = best[i];
407 if (j>=0 &&
408 !devices[j].uptodate &&
409 devices[j].events > 0 &&
410 (chosen_drive < 0 ||
411 devices[j].events > devices[chosen_drive].events))
412 chosen_drive = j;
413 }
414 if (chosen_drive < 0)
415 break;
416 if (verbose >= 0)
417 fprintf(stderr, Name ": forcing event count in %s(%d) from %d upto %d\n",
418 devices[chosen_drive].devname, devices[chosen_drive].raid_disk,
419 (int)(devices[chosen_drive].events),
420 (int)(devices[most_recent].events));
421 fd = open(devices[chosen_drive].devname, O_RDWR|O_EXCL);
422 if (fd < 0) {
423 fprintf(stderr, Name ": Couldn't open %s for write - not updating\n",
424 devices[chosen_drive].devname);
425 devices[chosen_drive].events = 0;
426 continue;
427 }
428 if (st->ss->load_super(st,fd, &super, NULL)) {
429 close(fd);
430 fprintf(stderr, Name ": RAID superblock disappeared from %s - not updating.\n",
431 devices[chosen_drive].devname);
432 devices[chosen_drive].events = 0;
433 continue;
434 }
435 info.events = devices[most_recent].events;
436 st->ss->update_super(&info, super, "force", devices[chosen_drive].devname, verbose);
437
438 if (st->ss->store_super(st, fd, super)) {
439 close(fd);
440 fprintf(stderr, Name ": Could not re-write superblock on %s\n",
441 devices[chosen_drive].devname);
442 devices[chosen_drive].events = 0;
443 free(super);
444 continue;
445 }
446 close(fd);
447 devices[chosen_drive].events = devices[most_recent].events;
448 devices[chosen_drive].uptodate = 1;
449 avail[chosen_drive] = 1;
450 okcnt++;
451 free(super);
452 }
453
454 /* Now we want to look at the superblock which the kernel will base things on
455 * and compare the devices that we think are working with the devices that the
456 * superblock thinks are working.
457 * If there are differences and --force is given, then update this chosen
458 * superblock.
459 */
460 chosen_drive = -1;
461 super = NULL;
462 for (i=0; chosen_drive < 0 && i<bestcnt; i++) {
463 int j = best[i];
464 int fd;
465
466 if (j<0)
467 continue;
468 if (!devices[j].uptodate)
469 continue;
470 chosen_drive = j;
471 if ((fd=open(devices[j].devname, O_RDONLY|O_EXCL))< 0) {
472 fprintf(stderr, Name ": Cannot open %s: %s\n",
473 devices[j].devname, strerror(errno));
474 return 1;
475 }
476 if (st->ss->load_super(st,fd, &super, NULL)) {
477 close(fd);
478 fprintf(stderr, Name ": RAID superblock has disappeared from %s\n",
479 devices[j].devname);
480 return 1;
481 }
482 close(fd);
483 }
484 if (super == NULL) {
485 fprintf(stderr, Name ": No suitable drives found for %s\n", mddev);
486 return 1;
487 }
488 st->ss->getinfo_super(&info, &ident2, super);
489 for (i=0; i<bestcnt; i++) {
490 int j = best[i];
491 unsigned int desired_state;
492
493 if (i < info.array.raid_disks)
494 desired_state = (1<<MD_DISK_ACTIVE) | (1<<MD_DISK_SYNC);
495 else
496 desired_state = 0;
497
498 if (j<0)
499 continue;
500 if (!devices[j].uptodate)
501 continue;
502 info.disk.number = devices[j].disk_nr;
503 info.disk.raid_disk = i;
504 info.disk.state = desired_state;
505
506 if (devices[j].uptodate &&
507 st->ss->update_super(&info, super, "assemble", NULL, verbose)) {
508 if (force) {
509 if (verbose >= 0)
510 fprintf(stderr, Name ": "
511 "clearing FAULTY flag for device %d in %s for %s\n",
512 j, mddev, devices[j].devname);
513 change = 1;
514 } else {
515 if (verbose >= -1)
516 fprintf(stderr, Name ": "
517 "device %d in %s has wrong state in superblock, but %s seems ok\n",
518 i, mddev, devices[j].devname);
519 }
520 }
521 #if 0
522 if (!devices[j].uptodate &&
523 !(super.disks[i].state & (1 << MD_DISK_FAULTY))) {
524 fprintf(stderr, Name ": devices %d of %s is not marked FAULTY in superblock, but cannot be found\n",
525 i, mddev);
526 }
527 #endif
528 }
529 if (force && okcnt == info.array.raid_disks-1) {
530 /* FIXME check event count */
531 change += st->ss->update_super(&info, super, "force",
532 devices[chosen_drive].devname, verbose);
533 }
534
535 if (change) {
536 int fd;
537 fd = open(devices[chosen_drive].devname, O_RDWR|O_EXCL);
538 if (fd < 0) {
539 fprintf(stderr, Name ": Could open %s for write - cannot Assemble array.\n",
540 devices[chosen_drive].devname);
541 return 1;
542 }
543 if (st->ss->store_super(st, fd, super)) {
544 close(fd);
545 fprintf(stderr, Name ": Could not re-write superblock on %s\n",
546 devices[chosen_drive].devname);
547 return 1;
548 }
549 close(fd);
550 }
551
552 /* count number of in-sync devices according to the superblock.
553 * We must have this number to start the array without -s or -R
554 */
555 req_cnt = info.array.working_disks;
556
557 /* Almost ready to actually *do* something */
558 if (!old_linux) {
559 int rv;
560 if ((vers % 100) >= 1) { /* can use different versions */
561 mdu_array_info_t inf;
562 memset(&inf, 0, sizeof(inf));
563 inf.major_version = st->ss->major;
564 inf.minor_version = st->minor_version;
565 rv = ioctl(mdfd, SET_ARRAY_INFO, &inf);
566 } else
567 rv = ioctl(mdfd, SET_ARRAY_INFO, NULL);
568
569 if (rv) {
570 fprintf(stderr, Name ": SET_ARRAY_INFO failed for %s: %s\n",
571 mddev, strerror(errno));
572 return 1;
573 }
574 if (ident->bitmap_fd >= 0) {
575 if (ioctl(mdfd, SET_BITMAP_FILE, ident->bitmap_fd) != 0) {
576 fprintf(stderr, Name ": SET_BITMAP_FILE failed.\n");
577 return 1;
578 }
579 }
580
581 /* First, add the raid disks, but add the chosen one last */
582 for (i=0; i<= bestcnt; i++) {
583 int j;
584 if (i < bestcnt) {
585 j = best[i];
586 if (j == chosen_drive)
587 continue;
588 } else
589 j = chosen_drive;
590
591 if (j >= 0 /* && devices[j].uptodate */) {
592 mdu_disk_info_t disk;
593 memset(&disk, 0, sizeof(disk));
594 disk.major = devices[j].major;
595 disk.minor = devices[j].minor;
596 if (ioctl(mdfd, ADD_NEW_DISK, &disk)!=0) {
597 fprintf(stderr, Name ": failed to add %s to %s: %s\n",
598 devices[j].devname,
599 mddev,
600 strerror(errno));
601 if (i < info.array.raid_disks || i == bestcnt)
602 okcnt--;
603 else
604 sparecnt--;
605 } else if (verbose > 0)
606 fprintf(stderr, Name ": added %s to %s as %d\n",
607 devices[j].devname, mddev, devices[j].raid_disk);
608 } else if (verbose > 0 && i < info.array.raid_disks)
609 fprintf(stderr, Name ": no uptodate device for slot %d of %s\n",
610 i, mddev);
611 }
612
613 if (runstop == 1 ||
614 (runstop == 0 &&
615 ( enough(info.array.level, info.array.raid_disks, info.array.layout, avail, okcnt) &&
616 (okcnt >= req_cnt || start_partial_ok)
617 ))) {
618 if (ioctl(mdfd, RUN_ARRAY, NULL)==0) {
619 if (verbose >= 0) {
620 fprintf(stderr, Name ": %s has been started with %d drive%s",
621 mddev, okcnt, okcnt==1?"":"s");
622 if (okcnt < info.array.raid_disks)
623 fprintf(stderr, " (out of %d)", info.array.raid_disks);
624 if (sparecnt)
625 fprintf(stderr, " and %d spare%s", sparecnt, sparecnt==1?"":"s");
626 fprintf(stderr, ".\n");
627 }
628 return 0;
629 }
630 fprintf(stderr, Name ": failed to RUN_ARRAY %s: %s\n",
631 mddev, strerror(errno));
632 return 1;
633 }
634 if (runstop == -1) {
635 fprintf(stderr, Name ": %s assembled from %d drive%s, but not started.\n",
636 mddev, okcnt, okcnt==1?"":"s");
637 return 0;
638 }
639 if (verbose >= 0) {
640 fprintf(stderr, Name ": %s assembled from %d drive%s", mddev, okcnt, okcnt==1?"":"s");
641 if (sparecnt)
642 fprintf(stderr, " and %d spare%s", sparecnt, sparecnt==1?"":"s");
643 if (!enough(info.array.level, info.array.raid_disks, info.array.layout, avail, okcnt))
644 fprintf(stderr, " - not enough to start the array.\n");
645 else {
646 if (req_cnt == info.array.raid_disks)
647 fprintf(stderr, " - need all %d to start it", req_cnt);
648 else
649 fprintf(stderr, " - need %d of %d to start", req_cnt, info.array.raid_disks);
650 fprintf(stderr, " (use --run to insist).\n");
651 }
652 }
653 return 1;
654 } else {
655 /* The "chosen_drive" is a good choice, and if necessary, the superblock has
656 * been updated to point to the current locations of devices.
657 * so we can just start the array
658 */
659 unsigned long dev;
660 dev = makedev(devices[chosen_drive].major,
661 devices[chosen_drive].minor);
662 if (ioctl(mdfd, START_ARRAY, dev)) {
663 fprintf(stderr, Name ": Cannot start array: %s\n",
664 strerror(errno));
665 }
666
667 }
668 return 0;
669 }