]>
Commit | Line | Data |
---|---|---|
64c4757e | 1 | /* |
9a9dab36 | 2 | * mdadm - manage Linux "md" devices aka RAID arrays. |
64c4757e | 3 | * |
e736b623 | 4 | * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de> |
64c4757e NB |
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 | |
e736b623 | 22 | * Email: <neilb@suse.de> |
64c4757e NB |
23 | */ |
24 | ||
9a9dab36 | 25 | #include "mdadm.h" |
41a3b72a | 26 | #include <ctype.h> |
64c4757e | 27 | |
624920bb NB |
28 | static int name_matches(char *found, char *required, char *homehost) |
29 | { | |
30 | /* See if the name found matches the required name, possibly | |
31 | * prefixed with 'homehost' | |
32 | */ | |
33 | char fnd[33]; | |
34 | ||
35 | strncpy(fnd, found, 32); | |
36 | fnd[32] = 0; | |
37 | if (strcmp(found, required)==0) | |
38 | return 1; | |
39 | if (homehost) { | |
40 | int l = strlen(homehost); | |
41 | if (l < 32 && fnd[l] == ':' && | |
42 | strcmp(fnd+l+1, required)==0) | |
43 | return 1; | |
44 | } | |
45 | return 0; | |
46 | } | |
47 | ||
9008ed1c | 48 | static int is_member_busy(char *metadata_version) |
2de8884f DW |
49 | { |
50 | /* check if the given member array is active */ | |
51 | struct mdstat_ent *mdstat = mdstat_read(1, 0); | |
52 | struct mdstat_ent *ent; | |
53 | int busy = 0; | |
54 | ||
55 | for (ent = mdstat; ent; ent = ent->next) { | |
56 | if (ent->metadata_version == NULL) | |
57 | continue; | |
58 | if (strncmp(ent->metadata_version, "external:", 9) != 0) | |
59 | continue; | |
60 | if (!is_subarray(&ent->metadata_version[9])) | |
61 | continue; | |
62 | /* Skip first char - it can be '/' or '-' */ | |
63 | if (strcmp(&ent->metadata_version[10], metadata_version+1) == 0) { | |
64 | busy = 1; | |
65 | break; | |
66 | } | |
67 | } | |
68 | free_mdstat(mdstat); | |
69 | ||
70 | return busy; | |
71 | } | |
72 | ||
fa56eddb | 73 | static int ident_matches(struct mddev_ident *ident, |
08fb91a3 N |
74 | struct mdinfo *content, |
75 | struct supertype *tst, | |
76 | char *homehost, | |
77 | char *update, char *devname) | |
78 | { | |
79 | ||
80 | if (ident->uuid_set && (!update || strcmp(update, "uuid")!= 0) && | |
cbeeb0e5 AC |
81 | same_uuid(content->uuid, ident->uuid, tst->ss->swapuuid)==0 && |
82 | memcmp(content->uuid, uuid_zero, sizeof(int[4])) != 0) { | |
08fb91a3 | 83 | if (devname) |
e7b84f9d | 84 | pr_err("%s has wrong uuid.\n", devname); |
08fb91a3 N |
85 | return 0; |
86 | } | |
87 | if (ident->name[0] && (!update || strcmp(update, "name")!= 0) && | |
88 | name_matches(content->name, ident->name, homehost)==0) { | |
89 | if (devname) | |
e7b84f9d | 90 | pr_err("%s has wrong name.\n", devname); |
08fb91a3 N |
91 | return 0; |
92 | } | |
93 | if (ident->super_minor != UnSet && | |
94 | ident->super_minor != content->array.md_minor) { | |
95 | if (devname) | |
e7b84f9d N |
96 | pr_err("%s has wrong super-minor.\n", |
97 | devname); | |
08fb91a3 N |
98 | return 0; |
99 | } | |
100 | if (ident->level != UnSet && | |
101 | ident->level != content->array.level) { | |
102 | if (devname) | |
e7b84f9d N |
103 | pr_err("%s has wrong raid level.\n", |
104 | devname); | |
08fb91a3 N |
105 | return 0; |
106 | } | |
107 | if (ident->raid_disks != UnSet && | |
108 | ident->raid_disks!= content->array.raid_disks) { | |
109 | if (devname) | |
e7b84f9d N |
110 | pr_err("%s requires wrong number of drives.\n", |
111 | devname); | |
08fb91a3 N |
112 | return 0; |
113 | } | |
805d30b2 N |
114 | if (ident->member && ident->member[0]) { |
115 | /* content->text_version must match */ | |
116 | char *s = strchr(content->text_version+1, '/'); | |
117 | if (s == NULL) { | |
118 | if (devname) | |
e7b84f9d N |
119 | pr_err("%s is not a container and one is required.\n", |
120 | devname); | |
805d30b2 N |
121 | return 0; |
122 | } else if (strcmp(ident->member, s+1) != 0) { | |
123 | if (devname) | |
e7b84f9d N |
124 | pr_err("skipping wrong member %s is %s\n", |
125 | content->text_version, devname); | |
805d30b2 N |
126 | return 0; |
127 | } | |
128 | } | |
08fb91a3 N |
129 | return 1; |
130 | } | |
08fb91a3 | 131 | |
95425a89 N |
132 | static int select_devices(struct mddev_dev *devlist, |
133 | struct mddev_ident *ident, | |
134 | struct supertype **stp, | |
135 | struct mdinfo **contentp, | |
136 | struct context *c, | |
137 | int inargv, int auto_assem) | |
138 | ||
64c4757e | 139 | { |
a655e550 | 140 | struct mddev_dev *tmpdev; |
95425a89 N |
141 | int num_devs; |
142 | struct supertype *st = *stp; | |
98dbd966 | 143 | struct mdinfo *content = NULL; |
95425a89 | 144 | int report_missmatch = ((inargv && c->verbose >= 0) || c->verbose > 0); |
cbeeb0e5 | 145 | struct domainlist *domains = NULL; |
60e1bc1a | 146 | |
5787fa49 NB |
147 | tmpdev = devlist; num_devs = 0; |
148 | while (tmpdev) { | |
da6b5ca9 NB |
149 | if (tmpdev->used) |
150 | tmpdev->used = 2; | |
151 | else | |
152 | num_devs++; | |
0431869c | 153 | tmpdev->disposition = 0; |
5787fa49 NB |
154 | tmpdev = tmpdev->next; |
155 | } | |
5787fa49 | 156 | |
811e6cbe NB |
157 | /* first walk the list of devices to find a consistent set |
158 | * that match the criterea, if that is possible. | |
c30e5369 | 159 | * We flag the ones we like with 'used'. |
811e6cbe NB |
160 | */ |
161 | for (tmpdev = devlist; | |
162 | tmpdev; | |
5083d66b | 163 | tmpdev = tmpdev ? tmpdev->next : NULL) { |
811e6cbe | 164 | char *devname = tmpdev->devname; |
64c4757e NB |
165 | int dfd; |
166 | struct stat stb; | |
518a60f3 | 167 | struct supertype *tst; |
4e8d9f0a | 168 | struct dev_policy *pol = NULL; |
5083d66b | 169 | int found_container = 0; |
52826846 | 170 | |
95425a89 N |
171 | if (tmpdev->used > 1) |
172 | continue; | |
da6b5ca9 | 173 | |
52826846 | 174 | if (ident->devices && |
56eedc1a | 175 | !match_oneof(ident->devices, devname)) { |
52437b4f | 176 | if (report_missmatch) |
e7b84f9d | 177 | pr_err("%s is not one of %s\n", devname, ident->devices); |
52826846 | 178 | continue; |
56eedc1a | 179 | } |
4b1ac34b | 180 | |
518a60f3 JS |
181 | tst = dup_super(st); |
182 | ||
56d18859 | 183 | dfd = dev_open(devname, O_RDONLY); |
64c4757e | 184 | if (dfd < 0) { |
52437b4f | 185 | if (report_missmatch) |
e7b84f9d N |
186 | pr_err("cannot open device %s: %s\n", |
187 | devname, strerror(errno)); | |
da6b5ca9 | 188 | tmpdev->used = 2; |
52826846 NB |
189 | } else if (fstat(dfd, &stb)< 0) { |
190 | /* Impossible! */ | |
e7b84f9d N |
191 | pr_err("fstat failed for %s: %s\n", |
192 | devname, strerror(errno)); | |
da6b5ca9 | 193 | tmpdev->used = 2; |
cd29a5c8 | 194 | } else if ((stb.st_mode & S_IFMT) != S_IFBLK) { |
e7b84f9d N |
195 | pr_err("%s is not a block device.\n", |
196 | devname); | |
da6b5ca9 | 197 | tmpdev->used = 2; |
5083d66b N |
198 | } else if (must_be_container(dfd)) { |
199 | if (st) { | |
200 | /* already found some components, this cannot | |
201 | * be another one. | |
202 | */ | |
203 | if (report_missmatch) | |
e7b84f9d N |
204 | pr_err("%s is a container, but we are looking for components\n", |
205 | devname); | |
5083d66b | 206 | tmpdev->used = 2; |
71204a50 | 207 | #if !defined(MDASSEMBLE) || defined(MDASSEMBLE) && defined(MDASSEMBLE_AUTO) |
5083d66b N |
208 | } if (!tst && (tst = super_by_fd(dfd, NULL)) == NULL) { |
209 | if (report_missmatch) | |
e7b84f9d N |
210 | pr_err("not a recognisable container: %s\n", |
211 | devname); | |
5083d66b | 212 | tmpdev->used = 2; |
71204a50 | 213 | #endif |
417f346e HCP |
214 | } else if (!tst->ss->load_container |
215 | || tst->ss->load_container(tst, dfd, NULL)) { | |
5083d66b | 216 | if (report_missmatch) |
e7b84f9d N |
217 | pr_err("no correct container type: %s\n", |
218 | devname); | |
5083d66b N |
219 | tmpdev->used = 2; |
220 | } else if (auto_assem && | |
221 | !conf_test_metadata(tst->ss->name, (pol = devnum_policy(stb.st_rdev)), | |
4977146a | 222 | tst->ss->match_home(tst, c->homehost) == 1)) { |
5083d66b | 223 | if (report_missmatch) |
e7b84f9d N |
224 | pr_err("%s has metadata type %s for which " |
225 | "auto-assembly is disabled\n", | |
226 | devname, tst->ss->name); | |
5083d66b N |
227 | tmpdev->used = 2; |
228 | } else | |
229 | found_container = 1; | |
52826846 | 230 | } else { |
5083d66b N |
231 | if (!tst && (tst = guess_super(dfd)) == NULL) { |
232 | if (report_missmatch) | |
e7b84f9d N |
233 | pr_err("no recogniseable superblock on %s\n", |
234 | devname); | |
5083d66b N |
235 | tmpdev->used = 2; |
236 | } else if (tst->ss->load_super(tst,dfd, NULL)) { | |
237 | if (report_missmatch) | |
e7b84f9d N |
238 | pr_err("no RAID superblock on %s\n", |
239 | devname); | |
5083d66b N |
240 | tmpdev->used = 2; |
241 | } else if (tst->ss->compare_super == NULL) { | |
242 | if (report_missmatch) | |
e7b84f9d N |
243 | pr_err("Cannot assemble %s metadata on %s\n", |
244 | tst->ss->name, devname); | |
5083d66b N |
245 | tmpdev->used = 2; |
246 | } else if (auto_assem && st == NULL && | |
247 | !conf_test_metadata(tst->ss->name, (pol = devnum_policy(stb.st_rdev)), | |
4977146a | 248 | tst->ss->match_home(tst, c->homehost) == 1)) { |
5083d66b | 249 | if (report_missmatch) |
e7b84f9d N |
250 | pr_err("%s has metadata type %s for which " |
251 | "auto-assembly is disabled\n", | |
252 | devname, tst->ss->name); | |
5083d66b N |
253 | tmpdev->used = 2; |
254 | } | |
64c4757e | 255 | } |
c06487ce | 256 | if (dfd >= 0) close(dfd); |
d68ea4d7 | 257 | if (tmpdev->used == 2) { |
d4386799 | 258 | if (auto_assem || !inargv) |
d68ea4d7 N |
259 | /* Ignore unrecognised devices during auto-assembly */ |
260 | goto loop; | |
261 | if (ident->uuid_set || ident->name[0] || | |
262 | ident->super_minor != UnSet) | |
263 | /* Ignore unrecognised device if looking for | |
264 | * specific array */ | |
265 | goto loop; | |
d68ea4d7 | 266 | |
e7b84f9d | 267 | pr_err("%s has no superblock - assembly aborted\n", |
79f9f56d | 268 | devname); |
d68ea4d7 N |
269 | if (st) |
270 | st->ss->free_super(st); | |
271 | dev_policy_free(pol); | |
cbeeb0e5 | 272 | domain_free(domains); |
95425a89 | 273 | return -1; |
d68ea4d7 | 274 | } |
52826846 | 275 | |
5083d66b | 276 | if (found_container) { |
9008ed1c N |
277 | /* tmpdev is a container. We need to be either |
278 | * looking for a member, or auto-assembling | |
279 | */ | |
56d18859 N |
280 | /* should be safe to try an exclusive open now, we |
281 | * have rejected anything that some other mdadm might | |
282 | * be looking at | |
283 | */ | |
284 | dfd = dev_open(devname, O_RDONLY | O_EXCL); | |
285 | if (dfd < 0) { | |
286 | if (report_missmatch) | |
e7b84f9d | 287 | pr_err("%s is busy - skipping\n", devname); |
56d18859 N |
288 | goto loop; |
289 | } | |
290 | close(dfd); | |
9008ed1c N |
291 | |
292 | if (ident->container) { | |
293 | if (ident->container[0] == '/' && | |
294 | !same_dev(ident->container, devname)) { | |
295 | if (report_missmatch) | |
e7b84f9d | 296 | pr_err("%s is not the container required (%s)\n", |
79f9f56d | 297 | devname, ident->container); |
9008ed1c N |
298 | goto loop; |
299 | } | |
300 | if (ident->container[0] != '/') { | |
301 | /* we have a uuid */ | |
302 | int uuid[4]; | |
87477e6d | 303 | |
95425a89 | 304 | content = *contentp; |
87477e6d N |
305 | tst->ss->getinfo_super(tst, content, NULL); |
306 | ||
9008ed1c N |
307 | if (!parse_uuid(ident->container, uuid) || |
308 | !same_uuid(content->uuid, uuid, tst->ss->swapuuid)) { | |
309 | if (report_missmatch) | |
e7b84f9d | 310 | pr_err("%s has wrong UUID to be required container\n", |
79f9f56d | 311 | devname); |
9008ed1c N |
312 | goto loop; |
313 | } | |
314 | } | |
315 | } | |
316 | /* It is worth looking inside this container. | |
317 | */ | |
4977146a | 318 | if (c->verbose > 0) |
e7b84f9d | 319 | pr_err("looking in container %s\n", |
79f9f56d | 320 | devname); |
1415fe4b | 321 | |
88cef9b3 N |
322 | for (content = tst->ss->container_content(tst, NULL); |
323 | content; | |
324 | content = content->next) { | |
1415fe4b | 325 | |
88cef9b3 | 326 | if (!ident_matches(ident, content, tst, |
4977146a | 327 | c->homehost, c->update, |
88cef9b3 N |
328 | report_missmatch ? devname : NULL)) |
329 | /* message already printed */; | |
330 | else if (is_member_busy(content->text_version)) { | |
331 | if (report_missmatch) | |
e7b84f9d | 332 | pr_err("member %s in %s is already assembled\n", |
79f9f56d N |
333 | content->text_version, |
334 | devname); | |
81219e70 LM |
335 | } else if (content->array.state & (1<<MD_SB_BLOCK_VOLUME)) { |
336 | /* do not assemble arrays with unsupported configurations */ | |
e7b84f9d | 337 | pr_err("Cannot activate member %s in %s.\n", |
79f9f56d N |
338 | content->text_version, |
339 | devname); | |
88cef9b3 N |
340 | } else |
341 | break; | |
342 | } | |
343 | if (!content) { | |
f7ad3ccc | 344 | tmpdev->used = 2; |
88cef9b3 | 345 | goto loop; /* empty container */ |
fa031239 | 346 | } |
88cef9b3 | 347 | |
9008ed1c | 348 | st = tst; tst = NULL; |
4c1c3ad8 | 349 | if (!auto_assem && inargv && tmpdev->next != NULL) { |
e7b84f9d | 350 | pr_err("%s is a container, but is not " |
79f9f56d N |
351 | "only device given: confused and aborting\n", |
352 | devname); | |
9008ed1c | 353 | st->ss->free_super(st); |
4e8d9f0a | 354 | dev_policy_free(pol); |
cbeeb0e5 | 355 | domain_free(domains); |
95425a89 | 356 | return -1; |
9008ed1c | 357 | } |
4977146a | 358 | if (c->verbose > 0) |
e7b84f9d | 359 | pr_err("found match on member %s in %s\n", |
79f9f56d | 360 | content->text_version, devname); |
bac0d92e | 361 | |
5083d66b N |
362 | /* make sure we finished the loop */ |
363 | tmpdev = NULL; | |
bac0d92e | 364 | goto loop; |
5083d66b | 365 | } else { |
bac0d92e | 366 | |
95425a89 | 367 | content = *contentp; |
5083d66b N |
368 | tst->ss->getinfo_super(tst, content, NULL); |
369 | ||
370 | if (!ident_matches(ident, content, tst, | |
4977146a | 371 | c->homehost, c->update, |
5083d66b | 372 | report_missmatch ? devname : NULL)) |
df37ffc0 | 373 | goto loop; |
cbeeb0e5 | 374 | |
56d18859 N |
375 | /* should be safe to try an exclusive open now, we |
376 | * have rejected anything that some other mdadm might | |
377 | * be looking at | |
378 | */ | |
379 | dfd = dev_open(devname, O_RDONLY | O_EXCL); | |
380 | if (dfd < 0) { | |
381 | if (report_missmatch) | |
e7b84f9d | 382 | pr_err("%s is busy - skipping\n", devname); |
56d18859 N |
383 | goto loop; |
384 | } | |
385 | close(dfd); | |
386 | ||
5083d66b N |
387 | if (st == NULL) |
388 | st = dup_super(tst); | |
389 | if (st->minor_version == -1) | |
390 | st->minor_version = tst->minor_version; | |
ed7fc6b4 AC |
391 | |
392 | if (memcmp(content->uuid, uuid_zero, | |
393 | sizeof(int[4])) == 0) { | |
394 | /* this is a floating spare. It cannot define | |
395 | * an array unless there are no more arrays of | |
396 | * this type to be found. It can be included | |
397 | * in an array of this type though. | |
398 | */ | |
399 | tmpdev->used = 3; | |
400 | goto loop; | |
401 | } | |
402 | ||
5083d66b N |
403 | if (st->ss != tst->ss || |
404 | st->minor_version != tst->minor_version || | |
405 | st->ss->compare_super(st, tst) != 0) { | |
406 | /* Some mismatch. If exactly one array matches this host, | |
407 | * we can resolve on that one. | |
408 | * Or, if we are auto assembling, we just ignore the second | |
409 | * for now. | |
410 | */ | |
411 | if (auto_assem) | |
412 | goto loop; | |
4977146a N |
413 | if (c->homehost) { |
414 | int first = st->ss->match_home(st, c->homehost); | |
415 | int last = tst->ss->match_home(tst, c->homehost); | |
5083d66b N |
416 | if (first != last && |
417 | (first == 1 || last == 1)) { | |
418 | /* We can do something */ | |
419 | if (first) {/* just ignore this one */ | |
420 | if (report_missmatch) | |
e7b84f9d | 421 | pr_err("%s misses out due to wrong homehost\n", |
79f9f56d | 422 | devname); |
5083d66b N |
423 | goto loop; |
424 | } else { /* reject all those sofar */ | |
425 | struct mddev_dev *td; | |
426 | if (report_missmatch) | |
e7b84f9d | 427 | pr_err("%s overrides previous devices due to good homehost\n", |
79f9f56d | 428 | devname); |
5083d66b N |
429 | for (td=devlist; td != tmpdev; td=td->next) |
430 | if (td->used == 1) | |
431 | td->used = 0; | |
432 | tmpdev->used = 1; | |
433 | goto loop; | |
434 | } | |
83b6208e NB |
435 | } |
436 | } | |
e7b84f9d | 437 | pr_err("superblock on %s doesn't match others - assembly aborted\n", |
79f9f56d | 438 | devname); |
5083d66b N |
439 | tst->ss->free_super(tst); |
440 | st->ss->free_super(st); | |
441 | dev_policy_free(pol); | |
cbeeb0e5 | 442 | domain_free(domains); |
95425a89 | 443 | return -1; |
83b6208e | 444 | } |
5083d66b | 445 | tmpdev->used = 1; |
64c4757e | 446 | } |
df37ffc0 | 447 | loop: |
cbeeb0e5 | 448 | /* Collect domain information from members only */ |
26b05aea AC |
449 | if (tmpdev && tmpdev->used == 1) { |
450 | if (!pol) | |
451 | pol = devnum_policy(stb.st_rdev); | |
cbeeb0e5 | 452 | domain_merge(&domains, pol, tst?tst->ss->name:NULL); |
26b05aea | 453 | } |
4e8d9f0a N |
454 | dev_policy_free(pol); |
455 | pol = NULL; | |
3d2b16e7 NB |
456 | if (tst) |
457 | tst->ss->free_super(tst); | |
811e6cbe NB |
458 | } |
459 | ||
ed7fc6b4 | 460 | /* Check if we found some imsm spares but no members */ |
3c7b4a25 CA |
461 | if ((auto_assem || |
462 | (ident->uuid_set && | |
463 | memcmp(uuid_zero, ident->uuid,sizeof(uuid_zero)) == 0)) && | |
464 | (!st || !st->sb)) | |
ed7fc6b4 AC |
465 | for (tmpdev = devlist; tmpdev; tmpdev = tmpdev->next) { |
466 | if (tmpdev->used != 3) | |
467 | continue; | |
468 | tmpdev->used = 1; | |
95425a89 | 469 | content = *contentp; |
ed7fc6b4 AC |
470 | |
471 | if (!st->sb) { | |
472 | /* we need sb from one of the spares */ | |
473 | int dfd = dev_open(tmpdev->devname, O_RDONLY); | |
474 | if (dfd < 0 || | |
475 | st->ss->load_super(st, dfd, NULL)) | |
476 | tmpdev->used = 2; | |
477 | if (dfd > 0) | |
478 | close(dfd); | |
479 | } | |
480 | } | |
481 | ||
cbeeb0e5 AC |
482 | /* Now reject spares that don't match domains of identified members */ |
483 | for (tmpdev = devlist; tmpdev; tmpdev = tmpdev->next) { | |
484 | struct stat stb; | |
485 | if (tmpdev->used != 3) | |
486 | continue; | |
487 | if (stat(tmpdev->devname, &stb)< 0) { | |
e7b84f9d | 488 | pr_err("fstat failed for %s: %s\n", |
79f9f56d | 489 | tmpdev->devname, strerror(errno)); |
cbeeb0e5 AC |
490 | tmpdev->used = 2; |
491 | } else { | |
a5d10dce N |
492 | struct dev_policy *pol = devnum_policy(stb.st_rdev); |
493 | int dt = domain_test(domains, pol, NULL); | |
494 | if (inargv && dt != 0) | |
495 | /* take this spare as domains match | |
496 | * if there are any */ | |
497 | tmpdev->used = 1; | |
498 | else if (!inargv && dt == 1) | |
499 | /* device wasn't explicitly listed, so need | |
500 | * explicit domain match - which we have */ | |
cbeeb0e5 AC |
501 | tmpdev->used = 1; |
502 | else | |
503 | /* if domains don't match mark as unused */ | |
504 | tmpdev->used = 0; | |
505 | dev_policy_free(pol); | |
506 | } | |
507 | } | |
508 | domain_free(domains); | |
95425a89 N |
509 | *stp = st; |
510 | if (st && st->sb && content == *contentp) | |
511 | st->ss->getinfo_super(st, content, NULL); | |
512 | *contentp = content; | |
513 | ||
514 | return num_devs; | |
515 | } | |
516 | ||
2c355c22 N |
517 | struct devs { |
518 | char *devname; | |
519 | int uptodate; /* set once we decide that this device is as | |
520 | * recent as everything else in the array. | |
521 | */ | |
522 | int included; /* set if the device is already in the array | |
523 | * due to a previous '-I' | |
524 | */ | |
525 | struct mdinfo i; | |
526 | }; | |
527 | ||
528 | static int load_devices(struct devs *devices, char *devmap, | |
529 | struct mddev_ident *ident, struct supertype *st, | |
530 | struct mddev_dev *devlist, struct context *c, | |
531 | struct mdinfo *content, | |
532 | int mdfd, char *mddev, | |
533 | int *most_recentp, int *bestcntp, int **bestp, | |
534 | int inargv) | |
535 | { | |
536 | struct mddev_dev *tmpdev; | |
537 | int devcnt = 0; | |
538 | int nextspare = 0; | |
539 | #ifndef MDASSEMBLE | |
540 | int bitmap_done = 0; | |
541 | #endif | |
542 | int most_recent = 0; | |
543 | int bestcnt = 0; | |
544 | int *best = *bestp; | |
545 | ||
546 | for (tmpdev = devlist; tmpdev; tmpdev=tmpdev->next) { | |
547 | char *devname = tmpdev->devname; | |
548 | struct stat stb; | |
549 | int i; | |
550 | ||
551 | if (tmpdev->used != 1) | |
552 | continue; | |
553 | /* looks like a good enough match to update the super block if needed */ | |
554 | #ifndef MDASSEMBLE | |
555 | if (c->update) { | |
556 | int dfd; | |
557 | /* prepare useful information in info structures */ | |
558 | struct stat stb2; | |
559 | struct supertype *tst; | |
560 | int err; | |
561 | fstat(mdfd, &stb2); | |
562 | ||
563 | if (strcmp(c->update, "uuid")==0 && | |
564 | !ident->uuid_set) { | |
565 | int rfd; | |
566 | if ((rfd = open("/dev/urandom", O_RDONLY)) < 0 || | |
567 | read(rfd, ident->uuid, 16) != 16) { | |
568 | *(__u32*)(ident->uuid) = random(); | |
569 | *(__u32*)(ident->uuid+1) = random(); | |
570 | *(__u32*)(ident->uuid+2) = random(); | |
571 | *(__u32*)(ident->uuid+3) = random(); | |
572 | } | |
573 | if (rfd >= 0) close(rfd); | |
574 | } | |
575 | dfd = dev_open(devname, | |
576 | tmpdev->disposition == 'I' | |
577 | ? O_RDWR : (O_RDWR|O_EXCL)); | |
578 | ||
579 | tst = dup_super(st); | |
580 | if (dfd < 0 || tst->ss->load_super(tst, dfd, NULL) != 0) { | |
581 | pr_err("cannot re-read metadata from %s - aborting\n", | |
582 | devname); | |
583 | if (dfd >= 0) | |
584 | close(dfd); | |
585 | close(mdfd); | |
586 | free(devices); | |
587 | free(devmap); | |
588 | return -1; | |
589 | } | |
590 | tst->ss->getinfo_super(tst, content, devmap + devcnt * content->array.raid_disks); | |
591 | ||
592 | memcpy(content->uuid, ident->uuid, 16); | |
593 | strcpy(content->name, ident->name); | |
594 | content->array.md_minor = minor(stb2.st_rdev); | |
595 | ||
596 | if (strcmp(c->update, "byteorder") == 0) | |
597 | err = 0; | |
598 | else | |
599 | err = tst->ss->update_super(tst, content, c->update, | |
600 | devname, c->verbose, | |
601 | ident->uuid_set, | |
602 | c->homehost); | |
603 | if (err < 0) { | |
604 | pr_err("--update=%s not understood" | |
605 | " for %s metadata\n", | |
606 | c->update, tst->ss->name); | |
607 | tst->ss->free_super(tst); | |
608 | free(tst); | |
609 | close(mdfd); | |
610 | close(dfd); | |
611 | free(devices); | |
612 | free(devmap); | |
613 | return -1; | |
614 | } | |
615 | if (strcmp(c->update, "uuid")==0 && | |
616 | !ident->uuid_set) { | |
617 | ident->uuid_set = 1; | |
618 | memcpy(ident->uuid, content->uuid, 16); | |
619 | } | |
620 | if (tst->ss->store_super(tst, dfd)) | |
621 | pr_err("Could not re-write superblock on %s.\n", | |
622 | devname); | |
623 | close(dfd); | |
624 | ||
625 | if (strcmp(c->update, "uuid")==0 && | |
626 | ident->bitmap_fd >= 0 && !bitmap_done) { | |
627 | if (bitmap_update_uuid(ident->bitmap_fd, | |
628 | content->uuid, | |
629 | tst->ss->swapuuid) != 0) | |
630 | pr_err("Could not update uuid on external bitmap.\n"); | |
631 | else | |
632 | bitmap_done = 1; | |
633 | } | |
634 | tst->ss->free_super(tst); | |
635 | } else | |
636 | #endif | |
637 | { | |
638 | struct supertype *tst = dup_super(st); | |
639 | int dfd; | |
640 | dfd = dev_open(devname, | |
641 | tmpdev->disposition == 'I' | |
642 | ? O_RDWR : (O_RDWR|O_EXCL)); | |
643 | ||
644 | if (dfd < 0 || tst->ss->load_super(tst, dfd, NULL) != 0) { | |
645 | pr_err("cannot re-read metadata from %s - aborting\n", | |
646 | devname); | |
647 | if (dfd >= 0) | |
648 | close(dfd); | |
649 | close(mdfd); | |
650 | free(devices); | |
651 | free(devmap); | |
652 | return -1; | |
653 | } | |
654 | tst->ss->getinfo_super(tst, content, devmap + devcnt * content->array.raid_disks); | |
655 | tst->ss->free_super(tst); | |
656 | close(dfd); | |
657 | } | |
658 | ||
659 | stat(devname, &stb); | |
660 | ||
661 | if (c->verbose > 0) | |
662 | pr_err("%s is identified as a member of %s, slot %d.\n", | |
663 | devname, mddev, content->disk.raid_disk); | |
664 | devices[devcnt].devname = devname; | |
665 | devices[devcnt].uptodate = 0; | |
666 | devices[devcnt].included = (tmpdev->disposition == 'I'); | |
667 | devices[devcnt].i = *content; | |
668 | devices[devcnt].i.disk.major = major(stb.st_rdev); | |
669 | devices[devcnt].i.disk.minor = minor(stb.st_rdev); | |
670 | if (most_recent < devcnt) { | |
671 | if (devices[devcnt].i.events | |
672 | > devices[most_recent].i.events) | |
673 | most_recent = devcnt; | |
674 | } | |
675 | if (content->array.level == LEVEL_MULTIPATH) | |
676 | /* with multipath, the raid_disk from the superblock is meaningless */ | |
677 | i = devcnt; | |
678 | else | |
679 | i = devices[devcnt].i.disk.raid_disk; | |
680 | if (i+1 == 0) { | |
681 | if (nextspare < content->array.raid_disks) | |
682 | nextspare = content->array.raid_disks; | |
683 | i = nextspare++; | |
684 | } else { | |
685 | if (i >= content->array.raid_disks && | |
686 | i >= nextspare) | |
687 | nextspare = i+1; | |
688 | } | |
689 | if (i < 10000) { | |
690 | if (i >= bestcnt) { | |
691 | int newbestcnt = i+10; | |
692 | int *newbest = xmalloc(sizeof(int)*newbestcnt); | |
693 | int c; | |
694 | for (c=0; c < newbestcnt; c++) | |
695 | if (c < bestcnt) | |
696 | newbest[c] = best[c]; | |
697 | else | |
698 | newbest[c] = -1; | |
699 | if (best)free(best); | |
700 | best = newbest; | |
701 | bestcnt = newbestcnt; | |
702 | } | |
703 | if (best[i] >=0 && | |
704 | devices[best[i]].i.events | |
705 | == devices[devcnt].i.events | |
706 | && (devices[best[i]].i.disk.minor | |
707 | != devices[devcnt].i.disk.minor) | |
708 | && st->ss == &super0 | |
709 | && content->array.level != LEVEL_MULTIPATH) { | |
710 | /* two different devices with identical superblock. | |
711 | * Could be a mis-detection caused by overlapping | |
712 | * partitions. fail-safe. | |
713 | */ | |
714 | pr_err("WARNING %s and %s appear" | |
715 | " to have very similar superblocks.\n" | |
716 | " If they are really different, " | |
717 | "please --zero the superblock on one\n" | |
718 | " If they are the same or overlap," | |
719 | " please remove one from %s.\n", | |
720 | devices[best[i]].devname, devname, | |
721 | inargv ? "the list" : | |
722 | "the\n DEVICE list in mdadm.conf" | |
723 | ); | |
724 | close(mdfd); | |
725 | free(devices); | |
726 | free(devmap); | |
727 | return -1; | |
728 | } | |
729 | if (best[i] == -1 | |
730 | || (devices[best[i]].i.events | |
731 | < devices[devcnt].i.events)) | |
732 | best[i] = devcnt; | |
733 | } | |
734 | devcnt++; | |
735 | } | |
736 | *most_recentp = most_recent; | |
737 | *bestcntp = bestcnt; | |
738 | *bestp = best; | |
739 | return devcnt; | |
740 | } | |
741 | ||
9f5470ce N |
742 | static int force_array(struct mdinfo *content, |
743 | struct devs *devices, | |
744 | int *best, int bestcnt, char *avail, | |
745 | int most_recent, | |
746 | struct supertype *st, | |
747 | struct context *c) | |
748 | { | |
749 | int okcnt = 0; | |
750 | while (!enough(content->array.level, content->array.raid_disks, | |
751 | content->array.layout, 1, | |
752 | avail) | |
753 | || | |
754 | (content->reshape_active && content->delta_disks > 0 && | |
755 | !enough(content->array.level, (content->array.raid_disks | |
756 | - content->delta_disks), | |
757 | content->new_layout, 1, | |
758 | avail) | |
759 | )) { | |
760 | /* Choose the newest best drive which is | |
761 | * not up-to-date, update the superblock | |
762 | * and add it. | |
763 | */ | |
764 | int fd; | |
765 | struct supertype *tst; | |
766 | unsigned long long current_events; | |
767 | int chosen_drive = -1; | |
768 | int i; | |
769 | ||
770 | for (i = 0; i < content->array.raid_disks && i < bestcnt; i++) { | |
771 | int j = best[i]; | |
772 | if (j>=0 && | |
773 | !devices[j].uptodate && | |
774 | devices[j].i.recovery_start == MaxSector && | |
775 | (chosen_drive < 0 || | |
776 | devices[j].i.events | |
777 | > devices[chosen_drive].i.events)) | |
778 | chosen_drive = j; | |
779 | } | |
780 | if (chosen_drive < 0) | |
781 | break; | |
782 | current_events = devices[chosen_drive].i.events; | |
783 | add_another: | |
784 | if (c->verbose >= 0) | |
785 | pr_err("forcing event count in %s(%d) from %d upto %d\n", | |
79f9f56d N |
786 | devices[chosen_drive].devname, |
787 | devices[chosen_drive].i.disk.raid_disk, | |
788 | (int)(devices[chosen_drive].i.events), | |
789 | (int)(devices[most_recent].i.events)); | |
9f5470ce N |
790 | fd = dev_open(devices[chosen_drive].devname, |
791 | devices[chosen_drive].included ? O_RDWR | |
792 | : (O_RDWR|O_EXCL)); | |
793 | if (fd < 0) { | |
794 | pr_err("Couldn't open %s for write - not updating\n", | |
79f9f56d | 795 | devices[chosen_drive].devname); |
9f5470ce N |
796 | devices[chosen_drive].i.events = 0; |
797 | continue; | |
798 | } | |
799 | tst = dup_super(st); | |
800 | if (tst->ss->load_super(tst,fd, NULL)) { | |
801 | close(fd); | |
802 | pr_err("RAID superblock disappeared from %s - not updating.\n", | |
79f9f56d | 803 | devices[chosen_drive].devname); |
9f5470ce N |
804 | devices[chosen_drive].i.events = 0; |
805 | continue; | |
806 | } | |
807 | content->events = devices[most_recent].i.events; | |
808 | tst->ss->update_super(tst, content, "force-one", | |
79f9f56d N |
809 | devices[chosen_drive].devname, c->verbose, |
810 | 0, NULL); | |
9f5470ce N |
811 | |
812 | if (tst->ss->store_super(tst, fd)) { | |
813 | close(fd); | |
814 | pr_err("Could not re-write superblock on %s\n", | |
79f9f56d | 815 | devices[chosen_drive].devname); |
9f5470ce N |
816 | devices[chosen_drive].i.events = 0; |
817 | tst->ss->free_super(tst); | |
818 | continue; | |
819 | } | |
820 | close(fd); | |
821 | devices[chosen_drive].i.events = devices[most_recent].i.events; | |
822 | devices[chosen_drive].uptodate = 1; | |
823 | avail[chosen_drive] = 1; | |
824 | okcnt++; | |
825 | tst->ss->free_super(tst); | |
826 | ||
827 | /* If there are any other drives of the same vintage, | |
828 | * add them in as well. We can't lose and we might gain | |
829 | */ | |
830 | for (i = 0; i < content->array.raid_disks && i < bestcnt ; i++) { | |
831 | int j = best[i]; | |
832 | if (j >= 0 && | |
833 | !devices[j].uptodate && | |
834 | devices[j].i.recovery_start == MaxSector && | |
835 | devices[j].i.events == current_events) { | |
836 | chosen_drive = j; | |
837 | goto add_another; | |
838 | } | |
839 | } | |
840 | } | |
841 | return okcnt; | |
842 | } | |
843 | ||
ddc1b11f N |
844 | static int start_array(int mdfd, |
845 | char *mddev, | |
846 | struct mdinfo *content, | |
847 | struct supertype *st, | |
848 | struct mddev_ident *ident, | |
849 | int *best, int bestcnt, | |
850 | int chosen_drive, | |
851 | struct devs *devices, | |
852 | unsigned int okcnt, | |
853 | unsigned int sparecnt, | |
854 | unsigned int rebuilding_cnt, | |
855 | struct context *c, | |
856 | int clean, char *avail, | |
857 | int start_partial_ok | |
858 | ) | |
859 | { | |
860 | int rv; | |
861 | int i; | |
862 | unsigned int req_cnt; | |
863 | ||
864 | rv = set_array_info(mdfd, st, content); | |
865 | if (rv) { | |
866 | pr_err("failed to set array info for %s: %s\n", | |
867 | mddev, strerror(errno)); | |
868 | return 1; | |
869 | } | |
870 | if (ident->bitmap_fd >= 0) { | |
871 | if (ioctl(mdfd, SET_BITMAP_FILE, ident->bitmap_fd) != 0) { | |
872 | pr_err("SET_BITMAP_FILE failed.\n"); | |
873 | return 1; | |
874 | } | |
875 | } else if (ident->bitmap_file) { | |
876 | /* From config file */ | |
877 | int bmfd = open(ident->bitmap_file, O_RDWR); | |
878 | if (bmfd < 0) { | |
879 | pr_err("Could not open bitmap file %s\n", | |
880 | ident->bitmap_file); | |
881 | return 1; | |
882 | } | |
883 | if (ioctl(mdfd, SET_BITMAP_FILE, bmfd) != 0) { | |
884 | pr_err("Failed to set bitmapfile for %s\n", mddev); | |
885 | close(bmfd); | |
886 | return 1; | |
887 | } | |
888 | close(bmfd); | |
889 | } | |
890 | ||
891 | /* First, add the raid disks, but add the chosen one last */ | |
892 | for (i=0; i<= bestcnt; i++) { | |
893 | int j; | |
894 | if (i < bestcnt) { | |
895 | j = best[i]; | |
896 | if (j == chosen_drive) | |
897 | continue; | |
898 | } else | |
899 | j = chosen_drive; | |
900 | ||
901 | if (j >= 0 && !devices[j].included) { | |
902 | int dfd = dev_open(devices[j].devname, | |
903 | O_RDWR|O_EXCL); | |
904 | if (dfd >= 0) { | |
905 | remove_partitions(dfd); | |
906 | close(dfd); | |
907 | } | |
908 | rv = add_disk(mdfd, st, content, &devices[j].i); | |
909 | ||
910 | if (rv) { | |
911 | pr_err("failed to add " | |
912 | "%s to %s: %s\n", | |
913 | devices[j].devname, | |
914 | mddev, | |
915 | strerror(errno)); | |
916 | if (i < content->array.raid_disks | |
917 | || i == bestcnt) | |
918 | okcnt--; | |
919 | else | |
920 | sparecnt--; | |
921 | } else if (c->verbose > 0) | |
922 | pr_err("added %s to %s as %d%s\n", | |
923 | devices[j].devname, mddev, | |
924 | devices[j].i.disk.raid_disk, | |
925 | devices[j].uptodate?"": | |
926 | " (possibly out of date)"); | |
927 | } else if (j >= 0) { | |
928 | if (c->verbose > 0) | |
929 | pr_err("%s is already in %s as %d\n", | |
930 | devices[j].devname, mddev, | |
931 | devices[j].i.disk.raid_disk); | |
932 | } else if (c->verbose > 0 && i < content->array.raid_disks) | |
933 | pr_err("no uptodate device for slot %d of %s\n", | |
934 | i, mddev); | |
935 | } | |
936 | ||
937 | if (content->array.level == LEVEL_CONTAINER) { | |
938 | if (c->verbose >= 0) { | |
939 | pr_err("Container %s has been " | |
940 | "assembled with %d drive%s", | |
941 | mddev, okcnt+sparecnt, okcnt+sparecnt==1?"":"s"); | |
942 | if (okcnt < (unsigned)content->array.raid_disks) | |
943 | fprintf(stderr, " (out of %d)", | |
944 | content->array.raid_disks); | |
945 | fprintf(stderr, "\n"); | |
946 | } | |
947 | st->ss->free_super(st); | |
948 | sysfs_uevent(content, "change"); | |
949 | return 0; | |
950 | } | |
951 | ||
952 | /* Get number of in-sync devices according to the superblock. | |
953 | * We must have this number to start the array without -s or -R | |
954 | */ | |
955 | req_cnt = content->array.working_disks; | |
956 | ||
957 | if (c->runstop == 1 || | |
958 | (c->runstop <= 0 && | |
959 | ( enough(content->array.level, content->array.raid_disks, | |
960 | content->array.layout, clean, avail) && | |
961 | (okcnt + rebuilding_cnt >= req_cnt || start_partial_ok) | |
962 | ))) { | |
963 | /* This array is good-to-go. | |
964 | * If a reshape is in progress then we might need to | |
965 | * continue monitoring it. In that case we start | |
966 | * it read-only and let the grow code make it writable. | |
967 | */ | |
968 | int rv; | |
969 | #ifndef MDASSEMBLE | |
970 | if (content->reshape_active && | |
971 | !(content->reshape_active & RESHAPE_NO_BACKUP) && | |
972 | content->delta_disks <= 0) { | |
973 | rv = sysfs_set_str(content, NULL, | |
974 | "array_state", "readonly"); | |
975 | if (rv == 0) | |
976 | rv = Grow_continue(mdfd, st, content, | |
977 | c->backup_file, | |
978 | c->freeze_reshape); | |
979 | } else if (c->readonly && | |
980 | sysfs_attribute_available( | |
981 | content, NULL, "array_state")) { | |
982 | rv = sysfs_set_str(content, NULL, | |
983 | "array_state", "readonly"); | |
984 | } else | |
985 | #endif | |
986 | rv = ioctl(mdfd, RUN_ARRAY, NULL); | |
987 | if (rv == 0) { | |
988 | if (c->verbose >= 0) { | |
989 | pr_err("%s has been started with %d drive%s", | |
990 | mddev, okcnt, okcnt==1?"":"s"); | |
991 | if (okcnt < (unsigned)content->array.raid_disks) | |
992 | fprintf(stderr, " (out of %d)", content->array.raid_disks); | |
993 | if (rebuilding_cnt) | |
994 | fprintf(stderr, "%s %d rebuilding", sparecnt?",":" and", rebuilding_cnt); | |
995 | if (sparecnt) | |
996 | fprintf(stderr, " and %d spare%s", sparecnt, sparecnt==1?"":"s"); | |
997 | fprintf(stderr, ".\n"); | |
998 | } | |
999 | if (content->reshape_active && | |
1000 | content->array.level >= 4 && | |
1001 | content->array.level <= 6) { | |
1002 | /* might need to increase the size | |
1003 | * of the stripe cache - default is 256 | |
1004 | */ | |
1005 | if (256 < 4 * (content->array.chunk_size/4096)) { | |
1006 | struct mdinfo *sra = sysfs_read(mdfd, 0, 0); | |
1007 | if (sra) | |
1008 | sysfs_set_num(sra, NULL, | |
1009 | "stripe_cache_size", | |
1010 | (4 * content->array.chunk_size / 4096) + 1); | |
1011 | sysfs_free(sra); | |
1012 | } | |
1013 | } | |
1014 | if (okcnt < (unsigned)content->array.raid_disks) { | |
1015 | /* If any devices did not get added | |
1016 | * because the kernel rejected them based | |
1017 | * on event count, try adding them | |
1018 | * again providing the action policy is | |
1019 | * 're-add' or greater. The bitmap | |
1020 | * might allow them to be included, or | |
1021 | * they will become spares. | |
1022 | */ | |
1023 | for (i = 0; i < bestcnt; i++) { | |
1024 | int j = best[i]; | |
1025 | if (j >= 0 && !devices[j].uptodate) { | |
1026 | if (!disk_action_allows(&devices[j].i, st->ss->name, act_re_add)) | |
1027 | continue; | |
1028 | rv = add_disk(mdfd, st, content, | |
1029 | &devices[j].i); | |
1030 | if (rv == 0 && c->verbose >= 0) | |
1031 | pr_err("%s has been re-added.\n", | |
1032 | devices[j].devname); | |
1033 | } | |
1034 | } | |
1035 | } | |
1036 | return 0; | |
1037 | } | |
1038 | pr_err("failed to RUN_ARRAY %s: %s\n", | |
1039 | mddev, strerror(errno)); | |
1040 | ||
1041 | if (!enough(content->array.level, content->array.raid_disks, | |
1042 | content->array.layout, 1, avail)) | |
1043 | pr_err("Not enough devices to " | |
1044 | "start the array.\n"); | |
1045 | else if (!enough(content->array.level, | |
1046 | content->array.raid_disks, | |
1047 | content->array.layout, clean, | |
1048 | avail)) | |
1049 | pr_err("Not enough devices to " | |
1050 | "start the array while not clean " | |
1051 | "- consider --force.\n"); | |
1052 | ||
1053 | return 1; | |
1054 | } | |
1055 | if (c->runstop == -1) { | |
1056 | pr_err("%s assembled from %d drive%s", | |
1057 | mddev, okcnt, okcnt==1?"":"s"); | |
1058 | if (okcnt != (unsigned)content->array.raid_disks) | |
1059 | fprintf(stderr, " (out of %d)", content->array.raid_disks); | |
1060 | fprintf(stderr, ", but not started.\n"); | |
1061 | return 2; | |
1062 | } | |
1063 | if (c->verbose >= -1) { | |
1064 | pr_err("%s assembled from %d drive%s", mddev, okcnt, okcnt==1?"":"s"); | |
1065 | if (rebuilding_cnt) | |
1066 | fprintf(stderr, "%s %d rebuilding", sparecnt?", ":" and ", rebuilding_cnt); | |
1067 | if (sparecnt) | |
1068 | fprintf(stderr, " and %d spare%s", sparecnt, sparecnt==1?"":"s"); | |
1069 | if (!enough(content->array.level, content->array.raid_disks, | |
1070 | content->array.layout, 1, avail)) | |
1071 | fprintf(stderr, " - not enough to start the array.\n"); | |
1072 | else if (!enough(content->array.level, | |
1073 | content->array.raid_disks, | |
1074 | content->array.layout, clean, | |
1075 | avail)) | |
1076 | fprintf(stderr, " - not enough to start the " | |
1077 | "array while not clean - consider " | |
1078 | "--force.\n"); | |
1079 | else { | |
1080 | if (req_cnt == (unsigned)content->array.raid_disks) | |
1081 | fprintf(stderr, " - need all %d to start it", req_cnt); | |
1082 | else | |
1083 | fprintf(stderr, " - need %d of %d to start", req_cnt, content->array.raid_disks); | |
1084 | fprintf(stderr, " (use --run to insist).\n"); | |
1085 | } | |
1086 | } | |
1087 | return 1; | |
1088 | } | |
1089 | ||
95425a89 N |
1090 | int Assemble(struct supertype *st, char *mddev, |
1091 | struct mddev_ident *ident, | |
1092 | struct mddev_dev *devlist, | |
1093 | struct context *c) | |
1094 | { | |
1095 | /* | |
1096 | * The task of Assemble is to find a collection of | |
1097 | * devices that should (according to their superblocks) | |
1098 | * form an array, and to give this collection to the MD driver. | |
1099 | * In Linux-2.4 and later, this involves submitting a | |
1100 | * SET_ARRAY_INFO ioctl with no arg - to prepare | |
1101 | * the array - and then submit a number of | |
1102 | * ADD_NEW_DISK ioctls to add disks into | |
1103 | * the array. Finally RUN_ARRAY might | |
1104 | * be submitted to start the array. | |
1105 | * | |
1106 | * Much of the work of Assemble is in finding and/or | |
1107 | * checking the disks to make sure they look right. | |
1108 | * | |
1109 | * If mddev is not set, then scan must be set and we | |
1110 | * read through the config file for dev+uuid mapping | |
1111 | * We recurse, setting mddev, for each device that | |
1112 | * - isn't running | |
1113 | * - has a valid uuid (or any uuid if !uuidset) | |
1114 | * | |
1115 | * If mddev is set, we try to determine state of md. | |
1116 | * check version - must be at least 0.90.0 | |
1117 | * check kernel version. must be at least 2.4. | |
1118 | * If not, we can possibly fall back on START_ARRAY | |
1119 | * Try to GET_ARRAY_INFO. | |
1120 | * If possible, give up | |
1121 | * If not, try to STOP_ARRAY just to make sure | |
1122 | * | |
1123 | * If !uuidset and scan, look in conf-file for uuid | |
1124 | * If not found, give up | |
1125 | * If !devlist and scan and uuidset, get list of devs from conf-file | |
1126 | * | |
1127 | * For each device: | |
1128 | * Check superblock - discard if bad | |
1129 | * Check uuid (set if we don't have one) - discard if no match | |
1130 | * Check superblock similarity if we have a superblock - discard if different | |
1131 | * Record events, devicenum | |
1132 | * This should give us a list of devices for the array | |
1133 | * We should collect the most recent event number | |
1134 | * | |
1135 | * Count disks with recent enough event count | |
1136 | * While force && !enough disks | |
1137 | * Choose newest rejected disks, update event count | |
1138 | * mark clean and rewrite superblock | |
1139 | * If recent kernel: | |
1140 | * SET_ARRAY_INFO | |
1141 | * foreach device with recent events : ADD_NEW_DISK | |
1142 | * if runstop == 1 || "enough" disks and runstop==0 -> RUN_ARRAY | |
1143 | * If old kernel: | |
1144 | * Check the device numbers in superblock are right | |
1145 | * update superblock if any changes | |
1146 | * START_ARRAY | |
1147 | * | |
1148 | */ | |
6f4dbdc4 | 1149 | int rv; |
95425a89 N |
1150 | int mdfd; |
1151 | int clean; | |
1152 | int auto_assem = (mddev == NULL && !ident->uuid_set && | |
1153 | ident->super_minor == UnSet && ident->name[0] == 0 | |
1154 | && (ident->container == NULL || ident->member == NULL)); | |
2c355c22 | 1155 | struct devs *devices; |
95425a89 N |
1156 | char *devmap; |
1157 | int *best = NULL; /* indexed by raid_disk */ | |
1158 | int bestcnt = 0; | |
2c355c22 | 1159 | int devcnt; |
95425a89 | 1160 | unsigned int okcnt, sparecnt, rebuilding_cnt; |
95425a89 N |
1161 | int i; |
1162 | int most_recent = 0; | |
1163 | int chosen_drive; | |
1164 | int change = 0; | |
1165 | int inargv = 0; | |
95425a89 N |
1166 | int start_partial_ok = (c->runstop >= 0) && |
1167 | (c->force || devlist==NULL || auto_assem); | |
1168 | int num_devs; | |
1169 | struct mddev_dev *tmpdev; | |
1170 | struct mdinfo info; | |
1171 | struct mdinfo *content = NULL; | |
1172 | struct mdinfo *pre_exist = NULL; | |
1173 | char *avail; | |
95425a89 N |
1174 | char *name = NULL; |
1175 | char chosen_name[1024]; | |
1176 | struct map_ent *map = NULL; | |
1177 | struct map_ent *mp; | |
1178 | ||
95425a89 N |
1179 | /* |
1180 | * If any subdevs are listed, then any that don't | |
1181 | * match ident are discarded. Remainder must all match and | |
1182 | * become the array. | |
1183 | * If no subdevs, then we scan all devices in the config file, but | |
1184 | * there must be something in the identity | |
1185 | */ | |
1186 | ||
1187 | if (!devlist && | |
1188 | ident->uuid_set == 0 && | |
1189 | (ident->super_minor < 0 || ident->super_minor == UnSet) && | |
1190 | ident->name[0] == 0 && | |
1191 | (ident->container == NULL || ident->member == NULL) && | |
1192 | ident->devices == NULL) { | |
1193 | pr_err("No identity information available for %s - cannot assemble.\n", | |
1194 | mddev ? mddev : "further assembly"); | |
1195 | return 1; | |
1196 | } | |
1197 | ||
1198 | if (devlist == NULL) | |
1199 | devlist = conf_get_devs(); | |
1200 | else if (mddev) | |
1201 | inargv = 1; | |
1202 | ||
79f9f56d | 1203 | try_again: |
95425a89 N |
1204 | /* We come back here when doing auto-assembly and attempting some |
1205 | * set of devices failed. Those are now marked as ->used==2 and | |
1206 | * we ignore them and try again | |
1207 | */ | |
1208 | if (!st && ident->st) | |
1209 | st = ident->st; | |
1210 | if (c->verbose>0) | |
1211 | pr_err("looking for devices for %s\n", | |
1212 | mddev ? mddev : "further assembly"); | |
1213 | ||
1214 | content = &info; | |
1215 | num_devs = select_devices(devlist, ident, &st, &content, c, | |
1216 | inargv, auto_assem); | |
1217 | if (num_devs < 0) | |
1218 | return 1; | |
cbeeb0e5 | 1219 | |
98dbd966 | 1220 | if (!st || !st->sb || !content) |
c30e5369 N |
1221 | return 2; |
1222 | ||
0431869c N |
1223 | /* We have a full set of devices - we now need to find the |
1224 | * array device. | |
1225 | * However there is a risk that we are racing with "mdadm -I" | |
1226 | * and the array is already partially assembled - we will have | |
1227 | * rejected any devices already in this address. | |
1228 | * So we take a lock on the map file - to prevent further races - | |
1229 | * and look for the uuid in there. If found and the array is | |
1230 | * active, we abort. If found and the array is not active | |
1231 | * we commit to that md device and add all the contained devices | |
1232 | * to our list. We flag them so that we don't try to re-add, | |
1233 | * but can remove if they turn out to not be wanted. | |
1234 | */ | |
1235 | if (map_lock(&map)) | |
1236 | pr_err("failed to get exclusive lock on mapfile - continue anyway...\n"); | |
1237 | mp = map_by_uuid(&map, content->uuid); | |
1238 | if (mp) { | |
1239 | struct mdinfo *dv; | |
1240 | /* array already exists. */ | |
1241 | pre_exist = sysfs_read(-1, mp->devnum, GET_LEVEL|GET_DEVS); | |
1242 | if (pre_exist->array.level != UnSet) { | |
1243 | pr_err("Found some drive for an array that is already active: %s\n", | |
1244 | mp->path); | |
1245 | pr_err("giving up.\n"); | |
1246 | return 1; | |
1247 | } | |
1248 | for (dv = pre_exist->devs; dv; dv = dv->next) { | |
1249 | /* We want to add this device to our list, | |
1250 | * but it could already be there if "mdadm -I" | |
1251 | * started *after* we checked for O_EXCL. | |
1252 | * If we add it to the top of the list | |
1253 | * it will be preferred over later copies. | |
1254 | */ | |
1255 | struct mddev_dev *newdev; | |
1256 | char *devname = map_dev(dv->disk.major, | |
1257 | dv->disk.minor, | |
1258 | 0); | |
1259 | if (!devname) | |
1260 | continue; | |
1261 | newdev = xmalloc(sizeof(*newdev)); | |
1262 | newdev->devname = devname; | |
1263 | newdev->disposition = 'I'; | |
1264 | newdev->used = 1; | |
1265 | newdev->next = devlist; | |
1266 | devlist = newdev; | |
1267 | num_devs++; | |
1268 | } | |
1269 | strcpy(chosen_name, mp->path); | |
1270 | if (c->verbose > 0 || mddev == NULL || | |
1271 | strcmp(mddev, chosen_name) != 0) | |
1272 | pr_err("Merging with already-assembled %s\n", | |
1273 | chosen_name); | |
1274 | mdfd = open_dev_excl(mp->devnum); | |
1275 | } else { | |
1276 | int trustworthy = FOREIGN; | |
1277 | name = content->name; | |
1278 | switch (st->ss->match_home(st, c->homehost) | |
1279 | ?: st->ss->match_home(st, "any")) { | |
1280 | case 1: | |
1281 | trustworthy = LOCAL; | |
1282 | name = strchr(content->name, ':'); | |
1283 | if (name) | |
1284 | name++; | |
1285 | else | |
1286 | name = content->name; | |
1287 | break; | |
1288 | } | |
1289 | if (!auto_assem) | |
1290 | /* If the array is listed in mdadm.conf or on | |
1291 | * command line, then we trust the name | |
1292 | * even if the array doesn't look local | |
1293 | */ | |
1294 | trustworthy = LOCAL; | |
ac2ecf55 | 1295 | |
0431869c N |
1296 | if (name[0] == 0 && |
1297 | content->array.level == LEVEL_CONTAINER) { | |
1298 | name = content->text_version; | |
1299 | trustworthy = METADATA; | |
1300 | } | |
0ac91628 | 1301 | |
0431869c N |
1302 | if (name[0] && trustworthy != LOCAL && |
1303 | ! c->require_homehost && | |
1304 | conf_name_is_free(name)) | |
1305 | trustworthy = LOCAL; | |
0ac91628 | 1306 | |
0431869c N |
1307 | if (trustworthy == LOCAL && |
1308 | strchr(name, ':')) | |
1309 | /* Ignore 'host:' prefix of name */ | |
1310 | name = strchr(name, ':')+1; | |
7cdc0872 | 1311 | |
0431869c N |
1312 | mdfd = create_mddev(mddev, name, ident->autof, trustworthy, |
1313 | chosen_name); | |
1314 | } | |
c30e5369 N |
1315 | if (mdfd < 0) { |
1316 | st->ss->free_super(st); | |
c30e5369 | 1317 | if (auto_assem) |
60e1bc1a | 1318 | goto try_again; |
c30e5369 N |
1319 | return 1; |
1320 | } | |
a04d5763 | 1321 | mddev = chosen_name; |
6f4dbdc4 N |
1322 | if (get_linux_version() < 2004000 || |
1323 | md_get_version(mdfd) < 9000) { | |
1324 | pr_err("Assemble requires Linux 2.4 or later, and\n" | |
79f9f56d N |
1325 | " md driver version 0.90.0 or later.\n" |
1326 | " Upgrade your kernel or try --build\n"); | |
c30e5369 N |
1327 | close(mdfd); |
1328 | return 1; | |
1329 | } | |
0431869c N |
1330 | if (pre_exist == NULL) { |
1331 | if (mddev_busy(fd2devnum(mdfd))) { | |
1332 | pr_err("%s already active, cannot restart it!\n", | |
1333 | mddev); | |
1334 | for (tmpdev = devlist ; | |
1335 | tmpdev && tmpdev->used != 1; | |
1336 | tmpdev = tmpdev->next) | |
1337 | ; | |
1338 | if (tmpdev && auto_assem) | |
1339 | pr_err("%s needed for %s...\n", | |
1340 | mddev, tmpdev->devname); | |
1341 | close(mdfd); | |
1342 | mdfd = -3; | |
1343 | st->ss->free_super(st); | |
1344 | if (auto_assem) | |
1345 | goto try_again; | |
1346 | return 1; | |
1347 | } | |
1348 | /* just incase it was started but has no content */ | |
1349 | ioctl(mdfd, STOP_ARRAY, NULL); | |
8a46fe84 NB |
1350 | } |
1351 | ||
9008ed1c N |
1352 | #ifndef MDASSEMBLE |
1353 | if (content != &info) { | |
1354 | /* This is a member of a container. Try starting the array. */ | |
588bebfc | 1355 | int err; |
11b6d91d N |
1356 | err = assemble_container_content(st, mdfd, content, c, |
1357 | chosen_name); | |
588bebfc | 1358 | close(mdfd); |
588bebfc | 1359 | return err; |
9008ed1c N |
1360 | } |
1361 | #endif | |
811e6cbe | 1362 | /* Ok, no bad inconsistancy, we can try updating etc */ |
503975b9 N |
1363 | devices = xcalloc(num_devs, sizeof(*devices)); |
1364 | devmap = xcalloc(num_devs, content->array.raid_disks); | |
2c355c22 N |
1365 | devcnt = load_devices(devices, devmap, ident, st, devlist, |
1366 | c, content, mdfd, mddev, | |
1367 | &most_recent, &bestcnt, &best, inargv); | |
1368 | if (devcnt < 0) | |
1369 | return 1; | |
4b1ac34b | 1370 | |
64c4757e | 1371 | if (devcnt == 0) { |
e7b84f9d | 1372 | pr_err("no devices found for %s\n", |
79f9f56d | 1373 | mddev); |
3d2b16e7 NB |
1374 | if (st) |
1375 | st->ss->free_super(st); | |
7f91af49 | 1376 | close(mdfd); |
d7f7ebb7 | 1377 | free(devices); |
02e7c5b7 | 1378 | free(devmap); |
64c4757e NB |
1379 | return 1; |
1380 | } | |
4b1ac34b | 1381 | |
4977146a | 1382 | if (c->update && strcmp(c->update, "byteorder")==0) |
3d2b16e7 NB |
1383 | st->minor_version = 90; |
1384 | ||
a5d85af7 | 1385 | st->ss->getinfo_super(st, content, NULL); |
98dbd966 | 1386 | clean = content->array.state & 1; |
4b1ac34b | 1387 | |
64c4757e NB |
1388 | /* now we have some devices that might be suitable. |
1389 | * I wonder how many | |
1390 | */ | |
503975b9 | 1391 | avail = xcalloc(content->array.raid_disks, 1); |
64c4757e | 1392 | okcnt = 0; |
52826846 | 1393 | sparecnt=0; |
1ff98339 | 1394 | rebuilding_cnt=0; |
f21e18ca | 1395 | for (i=0; i< bestcnt; i++) { |
64c4757e | 1396 | int j = best[i]; |
ee04451c NB |
1397 | int event_margin = 1; /* always allow a difference of '1' |
1398 | * like the kernel does | |
1399 | */ | |
64c4757e | 1400 | if (j < 0) continue; |
d013a55e NB |
1401 | /* note: we ignore error flags in multipath arrays |
1402 | * as they don't make sense | |
1403 | */ | |
df0d4ea0 | 1404 | if (content->array.level != LEVEL_MULTIPATH) |
f22385f9 | 1405 | if (!(devices[j].i.disk.state & (1<<MD_DISK_ACTIVE))) { |
213ee40b | 1406 | if (!(devices[j].i.disk.state |
ed7fc6b4 AC |
1407 | & (1<<MD_DISK_FAULTY))) { |
1408 | devices[j].uptodate = 1; | |
aa88f531 | 1409 | sparecnt++; |
ed7fc6b4 | 1410 | } |
d013a55e | 1411 | continue; |
aa88f531 | 1412 | } |
de5a472e | 1413 | /* If this device thinks that 'most_recent' has failed, then |
02e7c5b7 N |
1414 | * we must reject this device. |
1415 | */ | |
1416 | if (j != most_recent && | |
1417 | content->array.raid_disks > 0 && | |
1418 | devices[most_recent].i.disk.raid_disk >= 0 && | |
1419 | devmap[j * content->array.raid_disks + devices[most_recent].i.disk.raid_disk] == 0) { | |
4977146a | 1420 | if (c->verbose > -1) |
e7b84f9d | 1421 | pr_err("ignoring %s as it reports %s as failed\n", |
79f9f56d | 1422 | devices[j].devname, devices[most_recent].devname); |
02e7c5b7 N |
1423 | best[i] = -1; |
1424 | continue; | |
1425 | } | |
213ee40b NB |
1426 | if (devices[j].i.events+event_margin >= |
1427 | devices[most_recent].i.events) { | |
64c4757e | 1428 | devices[j].uptodate = 1; |
1ff98339 | 1429 | if (i < content->array.raid_disks) { |
dcc4210f DW |
1430 | if (devices[j].i.recovery_start == MaxSector || |
1431 | (content->reshape_active && | |
b720636a N |
1432 | ((i >= content->array.raid_disks - content->delta_disks) || |
1433 | (i >= content->array.raid_disks - content->delta_disks - 1 | |
1434 | && content->array.level == 4)))) { | |
1ff98339 N |
1435 | okcnt++; |
1436 | avail[i]=1; | |
1437 | } else | |
1438 | rebuilding_cnt++; | |
265e0f17 | 1439 | } else |
52826846 | 1440 | sparecnt++; |
64c4757e NB |
1441 | } |
1442 | } | |
02e7c5b7 | 1443 | free(devmap); |
9f5470ce N |
1444 | if (c->force) |
1445 | okcnt += force_array(content, devices, best, bestcnt, | |
1446 | avail, most_recent, st, c); | |
52826846 NB |
1447 | |
1448 | /* Now we want to look at the superblock which the kernel will base things on | |
1449 | * and compare the devices that we think are working with the devices that the | |
1450 | * superblock thinks are working. | |
1451 | * If there are differences and --force is given, then update this chosen | |
1452 | * superblock. | |
1453 | */ | |
cd29a5c8 | 1454 | chosen_drive = -1; |
3da92f27 | 1455 | st->ss->free_super(st); |
56eedc1a | 1456 | for (i=0; chosen_drive < 0 && i<bestcnt; i++) { |
52826846 NB |
1457 | int j = best[i]; |
1458 | int fd; | |
4b1ac34b | 1459 | |
52826846 NB |
1460 | if (j<0) |
1461 | continue; | |
1462 | if (!devices[j].uptodate) | |
1463 | continue; | |
a28232b8 N |
1464 | if (devices[j].i.events < devices[most_recent].i.events) |
1465 | continue; | |
52826846 | 1466 | chosen_drive = j; |
0431869c N |
1467 | if ((fd=dev_open(devices[j].devname, |
1468 | devices[j].included ? O_RDONLY | |
1469 | : (O_RDONLY|O_EXCL)))< 0) { | |
e7b84f9d | 1470 | pr_err("Cannot open %s: %s\n", |
79f9f56d | 1471 | devices[j].devname, strerror(errno)); |
7f91af49 | 1472 | close(mdfd); |
d7f7ebb7 | 1473 | free(devices); |
52826846 NB |
1474 | return 1; |
1475 | } | |
3da92f27 | 1476 | if (st->ss->load_super(st,fd, NULL)) { |
52826846 | 1477 | close(fd); |
e7b84f9d | 1478 | pr_err("RAID superblock has disappeared from %s\n", |
79f9f56d | 1479 | devices[j].devname); |
7f91af49 | 1480 | close(mdfd); |
d7f7ebb7 | 1481 | free(devices); |
52826846 NB |
1482 | return 1; |
1483 | } | |
1484 | close(fd); | |
1485 | } | |
3da92f27 | 1486 | if (st->sb == NULL) { |
e7b84f9d | 1487 | pr_err("No suitable drives found for %s\n", mddev); |
7f91af49 | 1488 | close(mdfd); |
d7f7ebb7 | 1489 | free(devices); |
4b1ac34b NB |
1490 | return 1; |
1491 | } | |
a5d85af7 | 1492 | st->ss->getinfo_super(st, content, NULL); |
f35f2525 | 1493 | #ifndef MDASSEMBLE |
98dbd966 | 1494 | sysfs_init(content, mdfd, 0); |
f35f2525 | 1495 | #endif |
56eedc1a | 1496 | for (i=0; i<bestcnt; i++) { |
52826846 | 1497 | int j = best[i]; |
98c6faba | 1498 | unsigned int desired_state; |
11a3e71d | 1499 | |
98dbd966 | 1500 | if (i < content->array.raid_disks) |
11a3e71d NB |
1501 | desired_state = (1<<MD_DISK_ACTIVE) | (1<<MD_DISK_SYNC); |
1502 | else | |
1503 | desired_state = 0; | |
1504 | ||
52826846 NB |
1505 | if (j<0) |
1506 | continue; | |
1507 | if (!devices[j].uptodate) | |
1508 | continue; | |
4b1ac34b | 1509 | |
213ee40b | 1510 | devices[j].i.disk.state = desired_state; |
4e9a6ff7 N |
1511 | if (!(devices[j].i.array.state & 1)) |
1512 | clean = 0; | |
213ee40b NB |
1513 | |
1514 | if (st->ss->update_super(st, &devices[j].i, "assemble", NULL, | |
4977146a N |
1515 | c->verbose, 0, NULL)) { |
1516 | if (c->force) { | |
1517 | if (c->verbose >= 0) | |
e7b84f9d | 1518 | pr_err("clearing FAULTY flag for device %d in %s for %s\n", |
79f9f56d | 1519 | j, mddev, devices[j].devname); |
4b1ac34b | 1520 | change = 1; |
52826846 | 1521 | } else { |
4977146a | 1522 | if (c->verbose >= -1) |
e7b84f9d | 1523 | pr_err("device %d in %s has wrong state in superblock, but %s seems ok\n", |
79f9f56d | 1524 | i, mddev, devices[j].devname); |
52826846 NB |
1525 | } |
1526 | } | |
4b1ac34b | 1527 | #if 0 |
213ee40b | 1528 | if (!(super.disks[i].i.disk.state & (1 << MD_DISK_FAULTY))) { |
e7b84f9d | 1529 | pr_err("devices %d of %s is not marked FAULTY in superblock, but cannot be found\n", |
79f9f56d | 1530 | i, mddev); |
52826846 | 1531 | } |
4b1ac34b | 1532 | #endif |
52826846 | 1533 | } |
4977146a | 1534 | if (c->force && !clean && |
98dbd966 DW |
1535 | !enough(content->array.level, content->array.raid_disks, |
1536 | content->array.layout, clean, | |
de5a472e | 1537 | avail)) { |
98dbd966 | 1538 | change += st->ss->update_super(st, content, "force-array", |
6f4dbdc4 | 1539 | devices[chosen_drive].devname, c->verbose, |
67a8c82d | 1540 | 0, NULL); |
583315d9 | 1541 | clean = 1; |
d013a55e | 1542 | } |
52826846 | 1543 | |
4b1ac34b | 1544 | if (change) { |
52826846 | 1545 | int fd; |
0431869c N |
1546 | fd = dev_open(devices[chosen_drive].devname, |
1547 | devices[chosen_drive].included ? | |
1548 | O_RDWR : (O_RDWR|O_EXCL)); | |
52826846 | 1549 | if (fd < 0) { |
e7b84f9d | 1550 | pr_err("Could not open %s for write - cannot Assemble array.\n", |
79f9f56d | 1551 | devices[chosen_drive].devname); |
7f91af49 | 1552 | close(mdfd); |
d7f7ebb7 | 1553 | free(devices); |
52826846 NB |
1554 | return 1; |
1555 | } | |
3da92f27 | 1556 | if (st->ss->store_super(st, fd)) { |
52826846 | 1557 | close(fd); |
e7b84f9d | 1558 | pr_err("Could not re-write superblock on %s\n", |
79f9f56d | 1559 | devices[chosen_drive].devname); |
7f91af49 | 1560 | close(mdfd); |
d7f7ebb7 | 1561 | free(devices); |
52826846 NB |
1562 | return 1; |
1563 | } | |
4977146a | 1564 | if (c->verbose >= 0) |
e7b84f9d | 1565 | pr_err("Marking array %s as 'clean'\n", |
79f9f56d | 1566 | mddev); |
52826846 | 1567 | close(fd); |
52826846 NB |
1568 | } |
1569 | ||
353632d9 NB |
1570 | /* If we are in the middle of a reshape we may need to restore saved data |
1571 | * that was moved aside due to the reshape overwriting live data | |
1572 | * The code of doing this lives in Grow.c | |
1573 | */ | |
39c5a909 | 1574 | #ifndef MDASSEMBLE |
5e88ab2e N |
1575 | if (content->reshape_active && |
1576 | !(content->reshape_active & RESHAPE_NO_BACKUP)) { | |
353632d9 | 1577 | int err = 0; |
503975b9 | 1578 | int *fdlist = xmalloc(sizeof(int)* bestcnt); |
4977146a | 1579 | if (c->verbose > 0) |
e7b84f9d | 1580 | pr_err(":%s has an active reshape - checking " |
79f9f56d N |
1581 | "if critical section needs to be restored\n", |
1582 | chosen_name); | |
353632d9 NB |
1583 | for (i=0; i<bestcnt; i++) { |
1584 | int j = best[i]; | |
1585 | if (j >= 0) { | |
0431869c N |
1586 | fdlist[i] = dev_open(devices[j].devname, |
1587 | devices[j].included | |
1588 | ? O_RDWR : (O_RDWR|O_EXCL)); | |
353632d9 | 1589 | if (fdlist[i] < 0) { |
e7b84f9d | 1590 | pr_err("Could not open %s for write - cannot Assemble array.\n", |
79f9f56d | 1591 | devices[j].devname); |
353632d9 NB |
1592 | err = 1; |
1593 | break; | |
1594 | } | |
1595 | } else | |
1596 | fdlist[i] = -1; | |
1597 | } | |
87f26d14 | 1598 | if (!err) { |
ba53ea59 AK |
1599 | if (st->ss->external && st->ss->recover_backup) |
1600 | err = st->ss->recover_backup(st, content); | |
1601 | else | |
1602 | err = Grow_restart(st, content, fdlist, bestcnt, | |
4977146a N |
1603 | c->backup_file, c->verbose > 0); |
1604 | if (err && c->invalid_backup) { | |
1605 | if (c->verbose > 0) | |
e7b84f9d | 1606 | pr_err("continuing" |
79f9f56d | 1607 | " without restoring backup\n"); |
87f26d14 N |
1608 | err = 0; |
1609 | } | |
1610 | } | |
353632d9 NB |
1611 | while (i>0) { |
1612 | i--; | |
1613 | if (fdlist[i]>=0) close(fdlist[i]); | |
1614 | } | |
56dcaa6b | 1615 | free(fdlist); |
353632d9 | 1616 | if (err) { |
e7b84f9d | 1617 | pr_err("Failed to restore critical section for reshape, sorry.\n"); |
4977146a | 1618 | if (c->backup_file == NULL) |
e7b84f9d | 1619 | cont_err("Possibly you needed to specify the --backup-file\n"); |
7f91af49 | 1620 | close(mdfd); |
d7f7ebb7 | 1621 | free(devices); |
353632d9 NB |
1622 | return err; |
1623 | } | |
1624 | } | |
39c5a909 | 1625 | #endif |
aa88f531 | 1626 | |
64c4757e | 1627 | /* Almost ready to actually *do* something */ |
6f4dbdc4 N |
1628 | /* First, fill in the map, so that udev can find our name |
1629 | * as soon as we become active. | |
1630 | */ | |
1631 | map_update(&map, fd2devnum(mdfd), content->text_version, | |
1632 | content->uuid, chosen_name); | |
1633 | ||
1634 | rv = start_array(mdfd, mddev, content, | |
1635 | st, ident, best, bestcnt, | |
1636 | chosen_drive, devices, okcnt, sparecnt, | |
1637 | rebuilding_cnt, | |
1638 | c, | |
1639 | clean, avail, start_partial_ok); | |
1640 | if (rv == 1 && !pre_exist) | |
1641 | ioctl(mdfd, STOP_ARRAY, NULL); | |
1642 | close(mdfd); | |
1643 | free(devices); | |
1644 | map_unlock(&map); | |
1645 | if (rv == 0) { | |
1646 | wait_for(chosen_name, mdfd); | |
1647 | if (auto_assem) { | |
1648 | int usecs = 1; | |
1649 | /* There is a nasty race with 'mdadm --monitor'. | |
1650 | * If it opens this device before we close it, | |
1651 | * it gets an incomplete open on which IO | |
1652 | * doesn't work and the capacity is | |
1653 | * wrong. | |
1654 | * If we reopen (to check for layered devices) | |
1655 | * before --monitor closes, we loose. | |
1656 | * | |
1657 | * So: wait upto 1 second for there to be | |
1658 | * a non-zero capacity. | |
1659 | */ | |
1660 | while (usecs < 1000) { | |
1661 | mdfd = open(mddev, O_RDONLY); | |
1662 | if (mdfd >= 0) { | |
1663 | unsigned long long size; | |
1664 | if (get_dev_size(mdfd, NULL, &size) && | |
1665 | size > 0) | |
1666 | break; | |
1667 | close(mdfd); | |
589395d6 | 1668 | } |
6f4dbdc4 N |
1669 | usleep(usecs); |
1670 | usecs <<= 1; | |
dab6685f | 1671 | } |
aa88f531 | 1672 | } |
64c4757e | 1673 | } |
6f4dbdc4 N |
1674 | /* '2' means 'OK, but not started yet' */ |
1675 | return rv == 2 ? 0 : rv; | |
64c4757e | 1676 | } |
6234c63c DW |
1677 | |
1678 | #ifndef MDASSEMBLE | |
1679 | int assemble_container_content(struct supertype *st, int mdfd, | |
11b6d91d N |
1680 | struct mdinfo *content, struct context *c, |
1681 | char *chosen_name) | |
6234c63c DW |
1682 | { |
1683 | struct mdinfo *dev, *sra; | |
1684 | int working = 0, preexist = 0; | |
882029c8 | 1685 | int expansion = 0; |
6234c63c | 1686 | struct map_ent *map = NULL; |
7af03341 | 1687 | int old_raid_disks; |
4aecb54a | 1688 | int start_reshape; |
6234c63c DW |
1689 | |
1690 | sysfs_init(content, mdfd, 0); | |
1691 | ||
1692 | sra = sysfs_read(mdfd, 0, GET_VERSION); | |
0ea8f5b1 N |
1693 | if (sra == NULL || strcmp(sra->text_version, content->text_version) != 0) { |
1694 | if (content->array.major_version == -1 && | |
1695 | content->array.minor_version == -2 && | |
11b6d91d | 1696 | c->readonly && |
0ea8f5b1 N |
1697 | content->text_version[0] == '/') |
1698 | content->text_version[0] = '-'; | |
22472ee1 JS |
1699 | if (sysfs_set_array(content, md_get_version(mdfd)) != 0) { |
1700 | if (sra) | |
1701 | sysfs_free(sra); | |
6234c63c | 1702 | return 1; |
22472ee1 | 1703 | } |
0ea8f5b1 | 1704 | } |
588bebfc | 1705 | |
4aecb54a AK |
1706 | /* There are two types of reshape: container wide or sub-array specific |
1707 | * Check if metadata requests blocking container wide reshapes | |
1708 | */ | |
1709 | start_reshape = (content->reshape_active && | |
79f9f56d N |
1710 | !((content->reshape_active == CONTAINER_RESHAPE) && |
1711 | (content->array.state & (1<<MD_SB_BLOCK_CONTAINER_RESHAPE)))); | |
4aecb54a AK |
1712 | |
1713 | /* Block subarray here if it is under reshape now | |
1714 | * Do not allow for any changes in this array | |
1715 | */ | |
1716 | if (st->ss->external && content->recovery_blocked && start_reshape) | |
b8063f07 AK |
1717 | block_subarray(content); |
1718 | ||
6234c63c DW |
1719 | if (sra) |
1720 | sysfs_free(sra); | |
7af03341 | 1721 | old_raid_disks = content->array.raid_disks - content->delta_disks; |
6234c63c | 1722 | for (dev = content->devs; dev; dev = dev->next) |
14032016 | 1723 | if (sysfs_add_disk(content, dev, 1) == 0) { |
7af03341 | 1724 | if (dev->disk.raid_disk >= old_raid_disks && |
14032016 AK |
1725 | content->reshape_active) |
1726 | expansion++; | |
1727 | else | |
1728 | working++; | |
1729 | } else if (errno == EEXIST) | |
6234c63c | 1730 | preexist++; |
111e9fda | 1731 | if (working + expansion == 0) |
7cb2aa33 | 1732 | return 1;/* Nothing new, don't try to start */ |
588bebfc | 1733 | |
8b4e5ea9 N |
1734 | map_update(&map, fd2devnum(mdfd), |
1735 | content->text_version, | |
1736 | content->uuid, chosen_name); | |
1737 | ||
11b6d91d | 1738 | if (c->runstop > 0 || |
79f9f56d N |
1739 | (working + preexist + expansion) >= |
1740 | content->array.working_disks) { | |
bb50e5d3 | 1741 | int err; |
a714580e | 1742 | |
81219e70 | 1743 | if (start_reshape) { |
49680258 | 1744 | int spare = content->array.raid_disks + expansion; |
3f54bd62 AK |
1745 | if (restore_backup(st, content, |
1746 | working, | |
11b6d91d | 1747 | spare, c->backup_file, c->verbose) == 1) |
49680258 | 1748 | return 1; |
49680258 | 1749 | |
3bd58dc6 AK |
1750 | err = sysfs_set_str(content, NULL, |
1751 | "array_state", "readonly"); | |
1752 | if (err) | |
1753 | return 1; | |
1754 | ||
a93ada3b AK |
1755 | if (st->ss->external) { |
1756 | if (!mdmon_running(st->container_dev)) | |
1757 | start_mdmon(st->container_dev); | |
1758 | ping_monitor_by_id(st->container_dev); | |
7728e1c6 | 1759 | if (mdmon_running(st->container_dev) && |
79f9f56d | 1760 | st->update_tail == NULL) |
7728e1c6 | 1761 | st->update_tail = &st->updates; |
a93ada3b AK |
1762 | } |
1763 | ||
11b6d91d N |
1764 | err = Grow_continue(mdfd, st, content, c->backup_file, |
1765 | c->freeze_reshape); | |
49680258 | 1766 | } else switch(content->array.level) { |
79f9f56d N |
1767 | case LEVEL_LINEAR: |
1768 | case LEVEL_MULTIPATH: | |
1769 | case 0: | |
1770 | err = sysfs_set_str(content, NULL, "array_state", | |
1771 | c->readonly ? "readonly" : "active"); | |
1772 | break; | |
1773 | default: | |
1774 | err = sysfs_set_str(content, NULL, "array_state", | |
1775 | "readonly"); | |
1776 | /* start mdmon if needed. */ | |
1777 | if (!err) { | |
1778 | if (!mdmon_running(st->container_dev)) | |
1779 | start_mdmon(st->container_dev); | |
1780 | ping_monitor_by_id(st->container_dev); | |
1781 | } | |
1782 | break; | |
bb50e5d3 | 1783 | } |
bb50e5d3 N |
1784 | if (!err) |
1785 | sysfs_set_safemode(content, content->safe_mode_delay); | |
4aecb54a AK |
1786 | |
1787 | /* Block subarray here if it is not reshaped now | |
1788 | * It has be blocked a little later to allow mdmon to switch in | |
1789 | * in to R/W state | |
1790 | */ | |
1791 | if (st->ss->external && content->recovery_blocked && | |
1792 | !start_reshape) | |
1793 | block_subarray(content); | |
1794 | ||
11b6d91d | 1795 | if (c->verbose >= 0) { |
bb50e5d3 | 1796 | if (err) |
e7b84f9d N |
1797 | pr_err("array %s now has %d device%s", |
1798 | chosen_name, working + preexist, | |
1799 | working + preexist == 1 ? "":"s"); | |
bb50e5d3 | 1800 | else |
e7b84f9d N |
1801 | pr_err("Started %s with %d device%s", |
1802 | chosen_name, working + preexist, | |
1803 | working + preexist == 1 ? "":"s"); | |
6234c63c DW |
1804 | if (preexist) |
1805 | fprintf(stderr, " (%d new)", working); | |
882029c8 AK |
1806 | if (expansion) |
1807 | fprintf(stderr, " ( + %d for expansion)", | |
1808 | expansion); | |
6234c63c DW |
1809 | fprintf(stderr, "\n"); |
1810 | } | |
bb50e5d3 | 1811 | if (!err) |
a7c6e3fb | 1812 | wait_for(chosen_name, mdfd); |
588bebfc | 1813 | return err; |
6234c63c | 1814 | /* FIXME should have an O_EXCL and wait for read-auto */ |
7cb2aa33 | 1815 | } else { |
11b6d91d | 1816 | if (c->verbose >= 0) { |
e7b84f9d N |
1817 | pr_err("%s assembled with %d device%s", |
1818 | chosen_name, preexist + working, | |
1819 | preexist + working == 1 ? "":"s"); | |
88716263 N |
1820 | if (preexist) |
1821 | fprintf(stderr, " (%d new)", working); | |
1822 | fprintf(stderr, " but not started\n"); | |
1823 | } | |
7cb2aa33 N |
1824 | return 1; |
1825 | } | |
6234c63c DW |
1826 | } |
1827 | #endif | |
1828 |