]> git.ipfire.org Git - thirdparty/mdadm.git/blob - Assemble.c
tests/06name: adjust for homehost
[thirdparty/mdadm.git] / Assemble.c
1 /*
2 * mdadm - manage Linux "md" devices aka RAID arrays.
3 *
4 * Copyright (C) 2001-2006 Neil Brown <neilb@suse.de>
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 #include <ctype.h>
32
33 static int name_matches(char *found, char *required, char *homehost)
34 {
35 /* See if the name found matches the required name, possibly
36 * prefixed with 'homehost'
37 */
38 char fnd[33];
39
40 strncpy(fnd, found, 32);
41 fnd[32] = 0;
42 if (strcmp(found, required)==0)
43 return 1;
44 if (homehost) {
45 int l = strlen(homehost);
46 if (l < 32 && fnd[l] == ':' &&
47 strcmp(fnd+l+1, required)==0)
48 return 1;
49 }
50 return 0;
51 }
52
53 static int is_member_busy(char *metadata_version)
54 {
55 /* check if the given member array is active */
56 struct mdstat_ent *mdstat = mdstat_read(1, 0);
57 struct mdstat_ent *ent;
58 int busy = 0;
59
60 for (ent = mdstat; ent; ent = ent->next) {
61 if (ent->metadata_version == NULL)
62 continue;
63 if (strncmp(ent->metadata_version, "external:", 9) != 0)
64 continue;
65 if (!is_subarray(&ent->metadata_version[9]))
66 continue;
67 /* Skip first char - it can be '/' or '-' */
68 if (strcmp(&ent->metadata_version[10], metadata_version+1) == 0) {
69 busy = 1;
70 break;
71 }
72 }
73 free_mdstat(mdstat);
74
75 return busy;
76 }
77
78 int Assemble(struct supertype *st, char *mddev,
79 mddev_ident_t ident,
80 mddev_dev_t devlist, char *backup_file,
81 int readonly, int runstop,
82 char *update, char *homehost,
83 int verbose, int force)
84 {
85 /*
86 * The task of Assemble is to find a collection of
87 * devices that should (according to their superblocks)
88 * form an array, and to give this collection to the MD driver.
89 * In Linux-2.4 and later, this involves submitting a
90 * SET_ARRAY_INFO ioctl with no arg - to prepare
91 * the array - and then submit a number of
92 * ADD_NEW_DISK ioctls to add disks into
93 * the array. Finally RUN_ARRAY might
94 * be submitted to start the array.
95 *
96 * Much of the work of Assemble is in finding and/or
97 * checking the disks to make sure they look right.
98 *
99 * If mddev is not set, then scan must be set and we
100 * read through the config file for dev+uuid mapping
101 * We recurse, setting mddev, for each device that
102 * - isn't running
103 * - has a valid uuid (or any uuid if !uuidset)
104 *
105 * If mddev is set, we try to determine state of md.
106 * check version - must be at least 0.90.0
107 * check kernel version. must be at least 2.4.
108 * If not, we can possibly fall back on START_ARRAY
109 * Try to GET_ARRAY_INFO.
110 * If possible, give up
111 * If not, try to STOP_ARRAY just to make sure
112 *
113 * If !uuidset and scan, look in conf-file for uuid
114 * If not found, give up
115 * If !devlist and scan and uuidset, get list of devs from conf-file
116 *
117 * For each device:
118 * Check superblock - discard if bad
119 * Check uuid (set if we don't have one) - discard if no match
120 * Check superblock similarity if we have a superblock - discard if different
121 * Record events, devicenum
122 * This should give us a list of devices for the array
123 * We should collect the most recent event number
124 *
125 * Count disks with recent enough event count
126 * While force && !enough disks
127 * Choose newest rejected disks, update event count
128 * mark clean and rewrite superblock
129 * If recent kernel:
130 * SET_ARRAY_INFO
131 * foreach device with recent events : ADD_NEW_DISK
132 * if runstop == 1 || "enough" disks and runstop==0 -> RUN_ARRAY
133 * If old kernel:
134 * Check the device numbers in superblock are right
135 * update superblock if any changes
136 * START_ARRAY
137 *
138 */
139 int mdfd;
140 int clean;
141 int auto_assem = (mddev == NULL && !ident->uuid_set &&
142 ident->super_minor == UnSet && ident->name[0] == 0
143 && ident->container == NULL && ident->member == NULL);
144 int old_linux = 0;
145 int vers = vers; /* Keep gcc quite - it really is initialised */
146 struct {
147 char *devname;
148 int uptodate; /* set once we decide that this device is as
149 * recent as everything else in the array.
150 */
151 struct mdinfo i;
152 } *devices;
153 int *best = NULL; /* indexed by raid_disk */
154 unsigned int bestcnt = 0;
155 int devcnt = 0;
156 unsigned int okcnt, sparecnt;
157 unsigned int req_cnt;
158 unsigned int i;
159 int most_recent = 0;
160 int chosen_drive;
161 int change = 0;
162 int inargv = 0;
163 int report_missmatch;
164 int bitmap_done;
165 int start_partial_ok = (runstop >= 0) &&
166 (force || devlist==NULL || auto_assem);
167 unsigned int num_devs;
168 mddev_dev_t tmpdev;
169 struct mdinfo info;
170 struct mdinfo *content = NULL;
171 mdu_array_info_t tmp_inf;
172 char *avail;
173 int nextspare = 0;
174 char *name = NULL;
175 int trustworthy;
176 char chosen_name[1024];
177
178 if (get_linux_version() < 2004000)
179 old_linux = 1;
180
181 /*
182 * If any subdevs are listed, then any that don't
183 * match ident are discarded. Remainder must all match and
184 * become the array.
185 * If no subdevs, then we scan all devices in the config file, but
186 * there must be something in the identity
187 */
188
189 if (!devlist &&
190 ident->uuid_set == 0 &&
191 ident->super_minor < 0 &&
192 ident->devices == NULL) {
193 fprintf(stderr, Name ": No identity information available for %s - cannot assemble.\n",
194 mddev ? mddev : "further assembly");
195 return 1;
196 }
197
198 if (devlist == NULL)
199 devlist = conf_get_devs();
200 else if (mddev)
201 inargv = 1;
202
203 report_missmatch = ((inargv && verbose >= 0) || verbose > 0);
204 try_again:
205 /* We come back here when doing auto-assembly and attempting some
206 * set of devices failed. Those are now marked as ->used==2 and
207 * we ignore them and try again
208 */
209
210 tmpdev = devlist; num_devs = 0;
211 while (tmpdev) {
212 if (tmpdev->used)
213 tmpdev->used = 2;
214 else
215 num_devs++;
216 tmpdev = tmpdev->next;
217 }
218 devices = malloc(num_devs * sizeof(*devices));
219
220 if (!st && ident->st) st = ident->st;
221
222 if (verbose>0)
223 fprintf(stderr, Name ": looking for devices for %s\n",
224 mddev ? mddev : "further assembly");
225
226 /* first walk the list of devices to find a consistent set
227 * that match the criterea, if that is possible.
228 * We flag the ones we like with 'used'.
229 */
230 for (tmpdev = devlist;
231 tmpdev;
232 tmpdev = tmpdev->next) {
233 char *devname = tmpdev->devname;
234 int dfd;
235 struct stat stb;
236 struct supertype *tst = dup_super(st);
237
238 if (tmpdev->used > 1) continue;
239
240 if (ident->devices &&
241 !match_oneof(ident->devices, devname)) {
242 if (report_missmatch)
243 fprintf(stderr, Name ": %s is not one of %s\n", devname, ident->devices);
244 continue;
245 }
246
247 dfd = dev_open(devname, O_RDONLY|O_EXCL);
248 if (dfd < 0) {
249 if (report_missmatch)
250 fprintf(stderr, Name ": cannot open device %s: %s\n",
251 devname, strerror(errno));
252 tmpdev->used = 2;
253 } else if (fstat(dfd, &stb)< 0) {
254 /* Impossible! */
255 fprintf(stderr, Name ": fstat failed for %s: %s\n",
256 devname, strerror(errno));
257 tmpdev->used = 2;
258 } else if ((stb.st_mode & S_IFMT) != S_IFBLK) {
259 fprintf(stderr, Name ": %s is not a block device.\n",
260 devname);
261 tmpdev->used = 2;
262 } else if (!tst && (tst = guess_super(dfd)) == NULL) {
263 if (report_missmatch)
264 fprintf(stderr, Name ": no recogniseable superblock on %s\n",
265 devname);
266 tmpdev->used = 2;
267 } else if (tst->ss->load_super(tst,dfd, NULL)) {
268 if (report_missmatch)
269 fprintf( stderr, Name ": no RAID superblock on %s\n",
270 devname);
271 } else {
272 content = &info;
273 memset(content, 0, sizeof(*content));
274 tst->ss->getinfo_super(tst, content);
275 }
276 if (dfd >= 0) close(dfd);
277
278 if (tst && tst->sb && tst->ss->container_content
279 && tst->loaded_container) {
280 /* tmpdev is a container. We need to be either
281 * looking for a member, or auto-assembling
282 */
283 if (st) {
284 /* already found some components, this cannot
285 * be another one.
286 */
287 if (report_missmatch)
288 fprintf(stderr, Name ": %s is a container, but we are looking for components\n",
289 devname);
290 goto loop;
291 }
292
293 if (ident->container) {
294 if (ident->container[0] == '/' &&
295 !same_dev(ident->container, devname)) {
296 if (report_missmatch)
297 fprintf(stderr, Name ": %s is not the container required (%s)\n",
298 devname, ident->container);
299 goto loop;
300 }
301 if (ident->container[0] != '/') {
302 /* we have a uuid */
303 int uuid[4];
304 if (!parse_uuid(ident->container, uuid) ||
305 !same_uuid(content->uuid, uuid, tst->ss->swapuuid)) {
306 if (report_missmatch)
307 fprintf(stderr, Name ": %s has wrong UUID to be required container\n",
308 devname);
309 goto loop;
310 }
311 }
312 }
313 /* It is worth looking inside this container.
314 */
315 next_member:
316 if (tmpdev->content)
317 content = tmpdev->content;
318 else
319 content = tst->ss->container_content(tst);
320
321 tmpdev->content = content->next;
322 if (tmpdev->content == NULL)
323 tmpdev->used = 1;
324
325 } else if (ident->container || ident->member) {
326 /* No chance of this matching if we don't have
327 * a container */
328 if (report_missmatch)
329 fprintf(stderr, Name "%s is not a container, and one is required.\n",
330 devname);
331 goto loop;
332 }
333
334 if (ident->uuid_set && (!update || strcmp(update, "uuid")!= 0) &&
335 (!tst || !tst->sb ||
336 same_uuid(content->uuid, ident->uuid, tst->ss->swapuuid)==0)) {
337 if (report_missmatch)
338 fprintf(stderr, Name ": %s has wrong uuid.\n",
339 devname);
340 goto loop;
341 }
342 if (ident->name[0] && (!update || strcmp(update, "name")!= 0) &&
343 (!tst || !tst->sb ||
344 name_matches(content->name, ident->name, homehost)==0)) {
345 if (report_missmatch)
346 fprintf(stderr, Name ": %s has wrong name.\n",
347 devname);
348 goto loop;
349 }
350 if (ident->super_minor != UnSet &&
351 (!tst || !tst->sb ||
352 ident->super_minor != content->array.md_minor)) {
353 if (report_missmatch)
354 fprintf(stderr, Name ": %s has wrong super-minor.\n",
355 devname);
356 goto loop;
357 }
358 if (ident->level != UnSet &&
359 (!tst || !tst->sb ||
360 ident->level != content->array.level)) {
361 if (report_missmatch)
362 fprintf(stderr, Name ": %s has wrong raid level.\n",
363 devname);
364 goto loop;
365 }
366 if (ident->raid_disks != UnSet &&
367 (!tst || !tst->sb ||
368 ident->raid_disks!= content->array.raid_disks)) {
369 if (report_missmatch)
370 fprintf(stderr, Name ": %s requires wrong number of drives.\n",
371 devname);
372 goto loop;
373 }
374 if (auto_assem) {
375 if (tst == NULL || tst->sb == NULL)
376 continue;
377 }
378 /* If we are this far, then we are nearly commited to this device.
379 * If the super_block doesn't exist, or doesn't match others,
380 * then we probably cannot continue
381 * However if one of the arrays is for the homehost, and
382 * the other isn't that can disambiguate.
383 */
384
385 if (!tst || !tst->sb) {
386 fprintf(stderr, Name ": %s has no superblock - assembly aborted\n",
387 devname);
388 if (st)
389 st->ss->free_super(st);
390 return 1;
391 }
392
393 if (tst && tst->sb && tst->ss->container_content
394 && tst->loaded_container) {
395 /* we have the one container we need, don't keep
396 * looking. If the chosen member is active, skip.
397 */
398 if (is_member_busy(content->text_version)) {
399 if (auto_assem)
400 goto loop;
401 fprintf(stderr, Name ": member %s in %s is already assembled\n",
402 content->text_version,
403 devname);
404 tst->ss->free_super(tst);
405 return 1;
406 }
407 st = tst; tst = NULL;
408 if (!auto_assem && tmpdev->next != NULL) {
409 fprintf(stderr, Name ": %s is a container, but is not "
410 "only device given: confused and aborting\n",
411 devname);
412 st->ss->free_super(st);
413 return 1;
414 }
415 break;
416 }
417 if (st == NULL)
418 st = dup_super(tst);
419 if (st->minor_version == -1)
420 st->minor_version = tst->minor_version;
421 if (st->ss != tst->ss ||
422 st->minor_version != tst->minor_version ||
423 st->ss->compare_super(st, tst) != 0) {
424 /* Some mismatch. If exactly one array matches this host,
425 * we can resolve on that one.
426 * Or, if we are auto assembling, we just ignore the second
427 * for now.
428 */
429 if (auto_assem)
430 goto loop;
431 if (homehost) {
432 int first = st->ss->match_home(st, homehost);
433 int last = tst->ss->match_home(tst, homehost);
434 if (first != last &&
435 (first == 1 || last == 1)) {
436 /* We can do something */
437 if (first) {/* just ignore this one */
438 if (report_missmatch)
439 fprintf(stderr, Name ": %s misses out due to wrong homehost\n",
440 devname);
441 goto loop;
442 } else { /* reject all those sofar */
443 mddev_dev_t td;
444 if (report_missmatch)
445 fprintf(stderr, Name ": %s overrides previous devices due to good homehost\n",
446 devname);
447 for (td=devlist; td != tmpdev; td=td->next)
448 if (td->used == 1)
449 td->used = 0;
450 tmpdev->used = 1;
451 goto loop;
452 }
453 }
454 }
455 fprintf(stderr, Name ": superblock on %s doesn't match others - assembly aborted\n",
456 devname);
457 tst->ss->free_super(tst);
458 st->ss->free_super(st);
459 return 1;
460 }
461
462 tmpdev->used = 1;
463
464 loop:
465 if (tmpdev->content)
466 goto next_member;
467 if (tst)
468 tst->ss->free_super(tst);
469 }
470
471 if (!st || !st->sb || !content)
472 return 2;
473
474 /* Now need to open array the device. Use create_mddev */
475 if (content == &info)
476 st->ss->getinfo_super(st, content);
477
478 trustworthy = FOREIGN;
479 switch (st->ss->match_home(st, homehost)) {
480 case 0:
481 trustworthy = FOREIGN;
482 name = content->name;
483 break;
484 case 1:
485 trustworthy = LOCAL;
486 name = strchr(content->name, ':');
487 if (name)
488 name++;
489 else
490 name = content->name;
491 break;
492 case -1:
493 trustworthy = FOREIGN;
494 break;
495 }
496 if (!auto_assem && trustworthy == FOREIGN)
497 /* If the array is listed in mdadm or on
498 * command line, then we trust the name
499 * even if the array doesn't look local
500 */
501 trustworthy = LOCAL;
502
503 if (content->name[0] == 0 &&
504 content->array.level == LEVEL_CONTAINER) {
505 name = content->text_version;
506 trustworthy = METADATA;
507 }
508 mdfd = create_mddev(mddev, name, ident->autof, trustworthy,
509 chosen_name);
510 if (mdfd < 0) {
511 st->ss->free_super(st);
512 free(devices);
513 if (auto_assem)
514 goto try_again;
515 return 1;
516 }
517 mddev = chosen_name;
518 vers = md_get_version(mdfd);
519 if (vers < 9000) {
520 fprintf(stderr, Name ": Assemble requires driver version 0.90.0 or later.\n"
521 " Upgrade your kernel or try --build\n");
522 close(mdfd);
523 return 1;
524 }
525 if (ioctl(mdfd, GET_ARRAY_INFO, &tmp_inf)==0) {
526 fprintf(stderr, Name ": %s already active, cannot restart it!\n",
527 mddev);
528 for (tmpdev = devlist ;
529 tmpdev && tmpdev->used != 1;
530 tmpdev = tmpdev->next)
531 ;
532 if (tmpdev && auto_assem)
533 fprintf(stderr, Name ": %s needed for %s...\n",
534 mddev, tmpdev->devname);
535 close(mdfd);
536 mdfd = -3;
537 st->ss->free_super(st);
538 free(devices);
539 if (auto_assem)
540 goto try_again;
541 return 1;
542 }
543 ioctl(mdfd, STOP_ARRAY, NULL); /* just incase it was started but has no content */
544
545 #ifndef MDASSEMBLE
546 if (content != &info) {
547 /* This is a member of a container. Try starting the array. */
548 return assemble_container_content(st, mdfd, content, runstop,
549 chosen_name, verbose);
550 }
551 #endif
552 /* Ok, no bad inconsistancy, we can try updating etc */
553 bitmap_done = 0;
554 for (tmpdev = devlist; tmpdev; tmpdev=tmpdev->next) if (tmpdev->used == 1) {
555 char *devname = tmpdev->devname;
556 struct stat stb;
557 /* looks like a good enough match to update the super block if needed */
558 #ifndef MDASSEMBLE
559 if (update) {
560 int dfd;
561 /* prepare useful information in info structures */
562 struct stat stb2;
563 struct supertype *tst;
564 fstat(mdfd, &stb2);
565
566 if (strcmp(update, "uuid")==0 &&
567 !ident->uuid_set) {
568 int rfd;
569 if ((rfd = open("/dev/urandom", O_RDONLY)) < 0 ||
570 read(rfd, ident->uuid, 16) != 16) {
571 *(__u32*)(ident->uuid) = random();
572 *(__u32*)(ident->uuid+1) = random();
573 *(__u32*)(ident->uuid+2) = random();
574 *(__u32*)(ident->uuid+3) = random();
575 }
576 if (rfd >= 0) close(rfd);
577 }
578 dfd = dev_open(devname, O_RDWR|O_EXCL);
579
580 remove_partitions(dfd);
581
582 tst = dup_super(st);
583 tst->ss->load_super(tst, dfd, NULL);
584 tst->ss->getinfo_super(tst, content);
585
586 memcpy(content->uuid, ident->uuid, 16);
587 strcpy(content->name, ident->name);
588 content->array.md_minor = minor(stb2.st_rdev);
589
590 tst->ss->update_super(tst, content, update,
591 devname, verbose,
592 ident->uuid_set, homehost);
593 if (strcmp(update, "uuid")==0 &&
594 !ident->uuid_set) {
595 ident->uuid_set = 1;
596 memcpy(ident->uuid, content->uuid, 16);
597 }
598 if (dfd < 0)
599 fprintf(stderr, Name ": Cannot open %s for superblock update\n",
600 devname);
601 else if (tst->ss->store_super(tst, dfd))
602 fprintf(stderr, Name ": Could not re-write superblock on %s.\n",
603 devname);
604 if (dfd >= 0)
605 close(dfd);
606
607 if (strcmp(update, "uuid")==0 &&
608 ident->bitmap_fd >= 0 && !bitmap_done) {
609 if (bitmap_update_uuid(ident->bitmap_fd,
610 content->uuid,
611 tst->ss->swapuuid) != 0)
612 fprintf(stderr, Name ": Could not update uuid on external bitmap.\n");
613 else
614 bitmap_done = 1;
615 }
616 tst->ss->free_super(tst);
617 } else
618 #endif
619 {
620 struct supertype *tst = dup_super(st);
621 int dfd;
622 dfd = dev_open(devname, O_RDWR|O_EXCL);
623
624 remove_partitions(dfd);
625
626 tst->ss->load_super(tst, dfd, NULL);
627 tst->ss->getinfo_super(tst, content);
628 tst->ss->free_super(tst);
629 close(dfd);
630 }
631
632 stat(devname, &stb);
633
634 if (verbose > 0)
635 fprintf(stderr, Name ": %s is identified as a member of %s, slot %d.\n",
636 devname, mddev, content->disk.raid_disk);
637 devices[devcnt].devname = devname;
638 devices[devcnt].uptodate = 0;
639 devices[devcnt].i = *content;
640 devices[devcnt].i.disk.major = major(stb.st_rdev);
641 devices[devcnt].i.disk.minor = minor(stb.st_rdev);
642 if (most_recent < devcnt) {
643 if (devices[devcnt].i.events
644 > devices[most_recent].i.events)
645 most_recent = devcnt;
646 }
647 if (content->array.level == -4)
648 /* with multipath, the raid_disk from the superblock is meaningless */
649 i = devcnt;
650 else
651 i = devices[devcnt].i.disk.raid_disk;
652 if (i+1 == 0) {
653 if (nextspare < content->array.raid_disks)
654 nextspare = content->array.raid_disks;
655 i = nextspare++;
656 } else {
657 if (i >= content->array.raid_disks &&
658 i >= nextspare)
659 nextspare = i+1;
660 }
661 if (i < 10000) {
662 if (i >= bestcnt) {
663 unsigned int newbestcnt = i+10;
664 int *newbest = malloc(sizeof(int)*newbestcnt);
665 unsigned int c;
666 for (c=0; c < newbestcnt; c++)
667 if (c < bestcnt)
668 newbest[c] = best[c];
669 else
670 newbest[c] = -1;
671 if (best)free(best);
672 best = newbest;
673 bestcnt = newbestcnt;
674 }
675 if (best[i] >=0 &&
676 devices[best[i]].i.events
677 == devices[devcnt].i.events
678 && (devices[best[i]].i.disk.minor
679 != devices[devcnt].i.disk.minor)
680 && st->ss == &super0
681 && content->array.level != LEVEL_MULTIPATH) {
682 /* two different devices with identical superblock.
683 * Could be a mis-detection caused by overlapping
684 * partitions. fail-safe.
685 */
686 fprintf(stderr, Name ": WARNING %s and %s appear"
687 " to have very similar superblocks.\n"
688 " If they are really different, "
689 "please --zero the superblock on one\n"
690 " If they are the same or overlap,"
691 " please remove one from %s.\n",
692 devices[best[i]].devname, devname,
693 inargv ? "the list" :
694 "the\n DEVICE list in mdadm.conf"
695 );
696 close(mdfd);
697 return 1;
698 }
699 if (best[i] == -1
700 || (devices[best[i]].i.events
701 < devices[devcnt].i.events))
702 best[i] = devcnt;
703 }
704 devcnt++;
705 }
706
707 if (devcnt == 0) {
708 fprintf(stderr, Name ": no devices found for %s\n",
709 mddev);
710 if (st)
711 st->ss->free_super(st);
712 close(mdfd);
713 return 1;
714 }
715
716 if (update && strcmp(update, "byteorder")==0)
717 st->minor_version = 90;
718
719 st->ss->getinfo_super(st, content);
720 clean = content->array.state & 1;
721
722 /* now we have some devices that might be suitable.
723 * I wonder how many
724 */
725 avail = malloc(content->array.raid_disks);
726 memset(avail, 0, content->array.raid_disks);
727 okcnt = 0;
728 sparecnt=0;
729 for (i=0; i< bestcnt ;i++) {
730 int j = best[i];
731 int event_margin = 1; /* always allow a difference of '1'
732 * like the kernel does
733 */
734 if (j < 0) continue;
735 /* note: we ignore error flags in multipath arrays
736 * as they don't make sense
737 */
738 if (content->array.level != -4)
739 if (!(devices[j].i.disk.state & (1<<MD_DISK_SYNC))) {
740 if (!(devices[j].i.disk.state
741 & (1<<MD_DISK_FAULTY)))
742 sparecnt++;
743 continue;
744 }
745 if (devices[j].i.events+event_margin >=
746 devices[most_recent].i.events) {
747 devices[j].uptodate = 1;
748 if (i < content->array.raid_disks) {
749 okcnt++;
750 avail[i]=1;
751 } else
752 sparecnt++;
753 }
754 }
755 while (force && !enough(content->array.level, content->array.raid_disks,
756 content->array.layout, 1,
757 avail, okcnt)) {
758 /* Choose the newest best drive which is
759 * not up-to-date, update the superblock
760 * and add it.
761 */
762 int fd;
763 struct supertype *tst;
764 long long current_events;
765 chosen_drive = -1;
766 for (i=0; i<content->array.raid_disks && i < bestcnt; i++) {
767 int j = best[i];
768 if (j>=0 &&
769 !devices[j].uptodate &&
770 devices[j].i.events > 0 &&
771 (chosen_drive < 0 ||
772 devices[j].i.events
773 > devices[chosen_drive].i.events))
774 chosen_drive = j;
775 }
776 if (chosen_drive < 0)
777 break;
778 current_events = devices[chosen_drive].i.events;
779 add_another:
780 if (verbose >= 0)
781 fprintf(stderr, Name ": forcing event count in %s(%d) from %d upto %d\n",
782 devices[chosen_drive].devname,
783 devices[chosen_drive].i.disk.raid_disk,
784 (int)(devices[chosen_drive].i.events),
785 (int)(devices[most_recent].i.events));
786 fd = dev_open(devices[chosen_drive].devname, O_RDWR|O_EXCL);
787 if (fd < 0) {
788 fprintf(stderr, Name ": Couldn't open %s for write - not updating\n",
789 devices[chosen_drive].devname);
790 devices[chosen_drive].i.events = 0;
791 continue;
792 }
793 tst = dup_super(st);
794 if (tst->ss->load_super(tst,fd, NULL)) {
795 close(fd);
796 fprintf(stderr, Name ": RAID superblock disappeared from %s - not updating.\n",
797 devices[chosen_drive].devname);
798 devices[chosen_drive].i.events = 0;
799 continue;
800 }
801 content->events = devices[most_recent].i.events;
802 tst->ss->update_super(tst, content, "force-one",
803 devices[chosen_drive].devname, verbose,
804 0, NULL);
805
806 if (tst->ss->store_super(tst, fd)) {
807 close(fd);
808 fprintf(stderr, Name ": Could not re-write superblock on %s\n",
809 devices[chosen_drive].devname);
810 devices[chosen_drive].i.events = 0;
811 tst->ss->free_super(tst);
812 continue;
813 }
814 close(fd);
815 devices[chosen_drive].i.events = devices[most_recent].i.events;
816 devices[chosen_drive].uptodate = 1;
817 avail[chosen_drive] = 1;
818 okcnt++;
819 tst->ss->free_super(tst);
820
821 /* If there are any other drives of the same vintage,
822 * add them in as well. We can't lose and we might gain
823 */
824 for (i=0; i<content->array.raid_disks && i < bestcnt ; i++) {
825 int j = best[i];
826 if (j >= 0 &&
827 !devices[j].uptodate &&
828 devices[j].i.events > 0 &&
829 devices[j].i.events == current_events) {
830 chosen_drive = j;
831 goto add_another;
832 }
833 }
834 }
835
836 /* Now we want to look at the superblock which the kernel will base things on
837 * and compare the devices that we think are working with the devices that the
838 * superblock thinks are working.
839 * If there are differences and --force is given, then update this chosen
840 * superblock.
841 */
842 chosen_drive = -1;
843 st->ss->free_super(st);
844 for (i=0; chosen_drive < 0 && i<bestcnt; i++) {
845 int j = best[i];
846 int fd;
847
848 if (j<0)
849 continue;
850 if (!devices[j].uptodate)
851 continue;
852 chosen_drive = j;
853 if ((fd=dev_open(devices[j].devname, O_RDONLY|O_EXCL))< 0) {
854 fprintf(stderr, Name ": Cannot open %s: %s\n",
855 devices[j].devname, strerror(errno));
856 close(mdfd);
857 return 1;
858 }
859 if (st->ss->load_super(st,fd, NULL)) {
860 close(fd);
861 fprintf(stderr, Name ": RAID superblock has disappeared from %s\n",
862 devices[j].devname);
863 close(mdfd);
864 return 1;
865 }
866 close(fd);
867 }
868 if (st->sb == NULL) {
869 fprintf(stderr, Name ": No suitable drives found for %s\n", mddev);
870 close(mdfd);
871 return 1;
872 }
873 st->ss->getinfo_super(st, content);
874 #ifndef MDASSEMBLE
875 sysfs_init(content, mdfd, 0);
876 #endif
877 for (i=0; i<bestcnt; i++) {
878 int j = best[i];
879 unsigned int desired_state;
880
881 if (i < content->array.raid_disks)
882 desired_state = (1<<MD_DISK_ACTIVE) | (1<<MD_DISK_SYNC);
883 else
884 desired_state = 0;
885
886 if (j<0)
887 continue;
888 if (!devices[j].uptodate)
889 continue;
890
891 devices[j].i.disk.state = desired_state;
892
893 if (st->ss->update_super(st, &devices[j].i, "assemble", NULL,
894 verbose, 0, NULL)) {
895 if (force) {
896 if (verbose >= 0)
897 fprintf(stderr, Name ": "
898 "clearing FAULTY flag for device %d in %s for %s\n",
899 j, mddev, devices[j].devname);
900 change = 1;
901 } else {
902 if (verbose >= -1)
903 fprintf(stderr, Name ": "
904 "device %d in %s has wrong state in superblock, but %s seems ok\n",
905 i, mddev, devices[j].devname);
906 }
907 }
908 #if 0
909 if (!(super.disks[i].i.disk.state & (1 << MD_DISK_FAULTY))) {
910 fprintf(stderr, Name ": devices %d of %s is not marked FAULTY in superblock, but cannot be found\n",
911 i, mddev);
912 }
913 #endif
914 }
915 if (force && !clean &&
916 !enough(content->array.level, content->array.raid_disks,
917 content->array.layout, clean,
918 avail, okcnt)) {
919 change += st->ss->update_super(st, content, "force-array",
920 devices[chosen_drive].devname, verbose,
921 0, NULL);
922 clean = 1;
923 }
924
925 if (change) {
926 int fd;
927 fd = dev_open(devices[chosen_drive].devname, O_RDWR|O_EXCL);
928 if (fd < 0) {
929 fprintf(stderr, Name ": Could not open %s for write - cannot Assemble array.\n",
930 devices[chosen_drive].devname);
931 close(mdfd);
932 return 1;
933 }
934 if (st->ss->store_super(st, fd)) {
935 close(fd);
936 fprintf(stderr, Name ": Could not re-write superblock on %s\n",
937 devices[chosen_drive].devname);
938 close(mdfd);
939 return 1;
940 }
941 close(fd);
942 }
943
944 /* If we are in the middle of a reshape we may need to restore saved data
945 * that was moved aside due to the reshape overwriting live data
946 * The code of doing this lives in Grow.c
947 */
948 #ifndef MDASSEMBLE
949 if (content->reshape_active) {
950 int err = 0;
951 int *fdlist = malloc(sizeof(int)* bestcnt);
952 for (i=0; i<bestcnt; i++) {
953 int j = best[i];
954 if (j >= 0) {
955 fdlist[i] = dev_open(devices[j].devname, O_RDWR|O_EXCL);
956 if (fdlist[i] < 0) {
957 fprintf(stderr, Name ": Could not open %s for write - cannot Assemble array.\n",
958 devices[j].devname);
959 err = 1;
960 break;
961 }
962 } else
963 fdlist[i] = -1;
964 }
965 if (!err)
966 err = Grow_restart(st, content, fdlist, bestcnt, backup_file);
967 while (i>0) {
968 i--;
969 if (fdlist[i]>=0) close(fdlist[i]);
970 }
971 if (err) {
972 fprintf(stderr, Name ": Failed to restore critical section for reshape, sorry.\n");
973 close(mdfd);
974 return err;
975 }
976 }
977 #endif
978 /* count number of in-sync devices according to the superblock.
979 * We must have this number to start the array without -s or -R
980 */
981 req_cnt = content->array.working_disks;
982
983 /* Almost ready to actually *do* something */
984 if (!old_linux) {
985 int rv;
986
987 /* First, fill in the map, so that udev can find our name
988 * as soon as we become active.
989 */
990 map_update(NULL, fd2devnum(mdfd), content->text_version,
991 content->uuid, chosen_name);
992
993 rv = set_array_info(mdfd, st, content);
994 if (rv) {
995 fprintf(stderr, Name ": failed to set array info for %s: %s\n",
996 mddev, strerror(errno));
997 close(mdfd);
998 return 1;
999 }
1000 if (ident->bitmap_fd >= 0) {
1001 if (ioctl(mdfd, SET_BITMAP_FILE, ident->bitmap_fd) != 0) {
1002 fprintf(stderr, Name ": SET_BITMAP_FILE failed.\n");
1003 close(mdfd);
1004 return 1;
1005 }
1006 } else if (ident->bitmap_file) {
1007 /* From config file */
1008 int bmfd = open(ident->bitmap_file, O_RDWR);
1009 if (bmfd < 0) {
1010 fprintf(stderr, Name ": Could not open bitmap file %s\n",
1011 ident->bitmap_file);
1012 close(mdfd);
1013 return 1;
1014 }
1015 if (ioctl(mdfd, SET_BITMAP_FILE, bmfd) != 0) {
1016 fprintf(stderr, Name ": Failed to set bitmapfile for %s\n", mddev);
1017 close(bmfd);
1018 close(mdfd);
1019 return 1;
1020 }
1021 close(bmfd);
1022 }
1023
1024 /* First, add the raid disks, but add the chosen one last */
1025 for (i=0; i<= bestcnt; i++) {
1026 int j;
1027 if (i < bestcnt) {
1028 j = best[i];
1029 if (j == chosen_drive)
1030 continue;
1031 } else
1032 j = chosen_drive;
1033
1034 if (j >= 0 /* && devices[j].uptodate */) {
1035 rv = add_disk(mdfd, st, content, &devices[j].i);
1036
1037 if (rv) {
1038 fprintf(stderr, Name ": failed to add "
1039 "%s to %s: %s\n",
1040 devices[j].devname,
1041 mddev,
1042 strerror(errno));
1043 if (i < content->array.raid_disks
1044 || i == bestcnt)
1045 okcnt--;
1046 else
1047 sparecnt--;
1048 } else if (verbose > 0)
1049 fprintf(stderr, Name ": added %s "
1050 "to %s as %d\n",
1051 devices[j].devname, mddev,
1052 devices[j].i.disk.raid_disk);
1053 } else if (verbose > 0 && i < content->array.raid_disks)
1054 fprintf(stderr, Name ": no uptodate device for "
1055 "slot %d of %s\n",
1056 i, mddev);
1057 }
1058
1059 if (content->array.level == LEVEL_CONTAINER) {
1060 if (verbose >= 0) {
1061 fprintf(stderr, Name ": Container %s has been "
1062 "assembled with %d drive%s",
1063 mddev, okcnt+sparecnt, okcnt+sparecnt==1?"":"s");
1064 if (okcnt < content->array.raid_disks)
1065 fprintf(stderr, " (out of %d)",
1066 content->array.raid_disks);
1067 fprintf(stderr, "\n");
1068 }
1069 sysfs_uevent(content, "change");
1070 close(mdfd);
1071 return 0;
1072 }
1073
1074 if (runstop == 1 ||
1075 (runstop <= 0 &&
1076 ( enough(content->array.level, content->array.raid_disks,
1077 content->array.layout, clean, avail, okcnt) &&
1078 (okcnt >= req_cnt || start_partial_ok)
1079 ))) {
1080 if (ioctl(mdfd, RUN_ARRAY, NULL)==0) {
1081 if (verbose >= 0) {
1082 fprintf(stderr, Name ": %s has been started with %d drive%s",
1083 mddev, okcnt, okcnt==1?"":"s");
1084 if (okcnt < content->array.raid_disks)
1085 fprintf(stderr, " (out of %d)", content->array.raid_disks);
1086 if (sparecnt)
1087 fprintf(stderr, " and %d spare%s", sparecnt, sparecnt==1?"":"s");
1088 fprintf(stderr, ".\n");
1089 }
1090 close(mdfd);
1091 if (auto_assem) {
1092 int usecs = 1;
1093 /* There is a nasty race with 'mdadm --monitor'.
1094 * If it opens this device before we close it,
1095 * it gets an incomplete open on which IO
1096 * doesn't work and the capacity is
1097 * wrong.
1098 * If we reopen (to check for layered devices)
1099 * before --monitor closes, we loose.
1100 *
1101 * So: wait upto 1 second for there to be
1102 * a non-zero capacity.
1103 */
1104 while (usecs < 1000) {
1105 mdfd = open(mddev, O_RDONLY);
1106 if (mdfd >= 0) {
1107 unsigned long long size;
1108 if (get_dev_size(mdfd, NULL, &size) &&
1109 size > 0)
1110 break;
1111 close(mdfd);
1112 }
1113 usleep(usecs);
1114 usecs <<= 1;
1115 }
1116 }
1117 return 0;
1118 }
1119 fprintf(stderr, Name ": failed to RUN_ARRAY %s: %s\n",
1120 mddev, strerror(errno));
1121
1122 if (!enough(content->array.level, content->array.raid_disks,
1123 content->array.layout, 1, avail, okcnt))
1124 fprintf(stderr, Name ": Not enough devices to "
1125 "start the array.\n");
1126 else if (!enough(content->array.level,
1127 content->array.raid_disks,
1128 content->array.layout, clean,
1129 avail, okcnt))
1130 fprintf(stderr, Name ": Not enough devices to "
1131 "start the array while not clean "
1132 "- consider --force.\n");
1133
1134 if (auto_assem)
1135 ioctl(mdfd, STOP_ARRAY, NULL);
1136 close(mdfd);
1137 return 1;
1138 }
1139 if (runstop == -1) {
1140 fprintf(stderr, Name ": %s assembled from %d drive%s",
1141 mddev, okcnt, okcnt==1?"":"s");
1142 if (okcnt != content->array.raid_disks)
1143 fprintf(stderr, " (out of %d)", content->array.raid_disks);
1144 fprintf(stderr, ", but not started.\n");
1145 close(mdfd);
1146 return 0;
1147 }
1148 if (verbose >= -1) {
1149 fprintf(stderr, Name ": %s assembled from %d drive%s", mddev, okcnt, okcnt==1?"":"s");
1150 if (sparecnt)
1151 fprintf(stderr, " and %d spare%s", sparecnt, sparecnt==1?"":"s");
1152 if (!enough(content->array.level, content->array.raid_disks,
1153 content->array.layout, 1, avail, okcnt))
1154 fprintf(stderr, " - not enough to start the array.\n");
1155 else if (!enough(content->array.level,
1156 content->array.raid_disks,
1157 content->array.layout, clean,
1158 avail, okcnt))
1159 fprintf(stderr, " - not enough to start the "
1160 "array while not clean - consider "
1161 "--force.\n");
1162 else {
1163 if (req_cnt == content->array.raid_disks)
1164 fprintf(stderr, " - need all %d to start it", req_cnt);
1165 else
1166 fprintf(stderr, " - need %d of %d to start", req_cnt, content->array.raid_disks);
1167 fprintf(stderr, " (use --run to insist).\n");
1168 }
1169 }
1170 if (auto_assem)
1171 ioctl(mdfd, STOP_ARRAY, NULL);
1172 close(mdfd);
1173 return 1;
1174 } else {
1175 /* The "chosen_drive" is a good choice, and if necessary, the superblock has
1176 * been updated to point to the current locations of devices.
1177 * so we can just start the array
1178 */
1179 unsigned long dev;
1180 dev = makedev(devices[chosen_drive].i.disk.major,
1181 devices[chosen_drive].i.disk.minor);
1182 if (ioctl(mdfd, START_ARRAY, dev)) {
1183 fprintf(stderr, Name ": Cannot start array: %s\n",
1184 strerror(errno));
1185 }
1186
1187 }
1188 close(mdfd);
1189 return 0;
1190 }
1191
1192 #ifndef MDASSEMBLE
1193 int assemble_container_content(struct supertype *st, int mdfd,
1194 struct mdinfo *content, int runstop,
1195 char *chosen_name, int verbose)
1196 {
1197 struct mdinfo *dev, *sra;
1198 int working = 0, preexist = 0;
1199 struct map_ent *map = NULL;
1200
1201 sysfs_init(content, mdfd, 0);
1202
1203 sra = sysfs_read(mdfd, 0, GET_VERSION);
1204 if (sra == NULL || strcmp(sra->text_version, content->text_version) != 0)
1205 if (sysfs_set_array(content, md_get_version(mdfd)) != 0)
1206 return 1;
1207 if (sra)
1208 sysfs_free(sra);
1209
1210 for (dev = content->devs; dev; dev = dev->next)
1211 if (sysfs_add_disk(content, dev) == 0)
1212 working++;
1213 else if (errno == EEXIST)
1214 preexist++;
1215 if (working == 0)
1216 /* Nothing new, don't try to start */ ;
1217 else if (runstop > 0 ||
1218 (working + preexist) >= content->array.working_disks) {
1219 switch(content->array.level) {
1220 case LEVEL_LINEAR:
1221 case LEVEL_MULTIPATH:
1222 case 0:
1223 sysfs_set_str(content, NULL, "array_state",
1224 "active");
1225 break;
1226 default:
1227 sysfs_set_str(content, NULL, "array_state",
1228 "readonly");
1229 /* start mdmon if needed. */
1230 if (!mdmon_running(st->container_dev))
1231 start_mdmon(st->container_dev);
1232 ping_monitor(devnum2devname(st->container_dev));
1233 break;
1234 }
1235 sysfs_set_safemode(content, content->safe_mode_delay);
1236 if (verbose >= 0) {
1237 fprintf(stderr, Name
1238 ": Started %s with %d devices",
1239 chosen_name, working + preexist);
1240 if (preexist)
1241 fprintf(stderr, " (%d new)", working);
1242 fprintf(stderr, "\n");
1243 }
1244 /* FIXME should have an O_EXCL and wait for read-auto */
1245 } else
1246 if (verbose >= 0)
1247 fprintf(stderr, Name
1248 ": %s assembled with %d devices but "
1249 "not started\n",
1250 chosen_name, working);
1251 map_update(&map, fd2devnum(mdfd),
1252 content->text_version,
1253 content->uuid, chosen_name);
1254
1255 return 0;
1256 }
1257 #endif
1258