]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * mdmon - monitor external metadata arrays | |
3 | * | |
4 | * Copyright (C) 2007-2008 Neil Brown <neilb@suse.de> | |
5 | * Copyright (C) 2007-2008 Intel Corporation | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify it | |
8 | * under the terms and conditions of the GNU General Public License, | |
9 | * version 2, as published by the Free Software Foundation. | |
10 | * | |
11 | * This program is distributed in the hope it will be useful, but WITHOUT | |
12 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
14 | * more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License along with | |
17 | * this program; if not, write to the Free Software Foundation, Inc., | |
18 | * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. | |
19 | */ | |
20 | ||
21 | #include "mdadm.h" | |
22 | #include "mdmon.h" | |
23 | #include <sys/syscall.h> | |
24 | #include <sys/select.h> | |
25 | #include <signal.h> | |
26 | ||
27 | static char *array_states[] = { | |
28 | "clear", "inactive", "suspended", "readonly", "read-auto", | |
29 | "clean", "active", "write-pending", "active-idle", NULL }; | |
30 | static char *sync_actions[] = { | |
31 | "idle", "reshape", "resync", "recover", "check", "repair", NULL | |
32 | }; | |
33 | ||
34 | static int write_attr(char *attr, int fd) | |
35 | { | |
36 | return write(fd, attr, strlen(attr)); | |
37 | } | |
38 | ||
39 | static void add_fd(fd_set *fds, int *maxfd, int fd) | |
40 | { | |
41 | if (fd < 0) | |
42 | return; | |
43 | if (fd > *maxfd) | |
44 | *maxfd = fd; | |
45 | FD_SET(fd, fds); | |
46 | } | |
47 | ||
48 | static int read_attr(char *buf, int len, int fd) | |
49 | { | |
50 | int n; | |
51 | ||
52 | if (fd < 0) { | |
53 | buf[0] = 0; | |
54 | return 0; | |
55 | } | |
56 | lseek(fd, 0, 0); | |
57 | n = read(fd, buf, len - 1); | |
58 | ||
59 | if (n <= 0) { | |
60 | buf[0] = 0; | |
61 | return 0; | |
62 | } | |
63 | buf[n] = 0; | |
64 | if (buf[n-1] == '\n') | |
65 | buf[n-1] = 0; | |
66 | return n; | |
67 | } | |
68 | ||
69 | int get_resync_start(struct active_array *a) | |
70 | { | |
71 | char buf[30]; | |
72 | int n; | |
73 | ||
74 | n = read_attr(buf, 30, a->resync_start_fd); | |
75 | if (n <= 0) | |
76 | return n; | |
77 | ||
78 | a->resync_start = strtoull(buf, NULL, 10); | |
79 | ||
80 | return 1; | |
81 | } | |
82 | ||
83 | ||
84 | static enum array_state read_state(int fd) | |
85 | { | |
86 | char buf[20]; | |
87 | int n = read_attr(buf, 20, fd); | |
88 | ||
89 | if (n <= 0) | |
90 | return bad_word; | |
91 | return (enum array_state) sysfs_match_word(buf, array_states); | |
92 | } | |
93 | ||
94 | static enum sync_action read_action( int fd) | |
95 | { | |
96 | char buf[20]; | |
97 | int n = read_attr(buf, 20, fd); | |
98 | ||
99 | if (n <= 0) | |
100 | return bad_action; | |
101 | return (enum sync_action) sysfs_match_word(buf, sync_actions); | |
102 | } | |
103 | ||
104 | int read_dev_state(int fd) | |
105 | { | |
106 | char buf[60]; | |
107 | int n = read_attr(buf, 60, fd); | |
108 | char *cp; | |
109 | int rv = 0; | |
110 | ||
111 | if (n <= 0) | |
112 | return 0; | |
113 | ||
114 | cp = buf; | |
115 | while (cp) { | |
116 | if (sysfs_attr_match(cp, "faulty")) | |
117 | rv |= DS_FAULTY; | |
118 | if (sysfs_attr_match(cp, "in_sync")) | |
119 | rv |= DS_INSYNC; | |
120 | if (sysfs_attr_match(cp, "write_mostly")) | |
121 | rv |= DS_WRITE_MOSTLY; | |
122 | if (sysfs_attr_match(cp, "spare")) | |
123 | rv |= DS_SPARE; | |
124 | if (sysfs_attr_match(cp, "blocked")) | |
125 | rv |= DS_BLOCKED; | |
126 | cp = strchr(cp, ','); | |
127 | if (cp) | |
128 | cp++; | |
129 | } | |
130 | return rv; | |
131 | } | |
132 | ||
133 | static void signal_manager(void) | |
134 | { | |
135 | /* tgkill(getpid(), mon_tid, SIGUSR1); */ | |
136 | int pid = getpid(); | |
137 | syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1); | |
138 | } | |
139 | ||
140 | /* Monitor a set of active md arrays - all of which share the | |
141 | * same metadata - and respond to events that require | |
142 | * metadata update. | |
143 | * | |
144 | * New arrays are detected by another thread which allocates | |
145 | * required memory and attaches the data structure to our list. | |
146 | * | |
147 | * Events: | |
148 | * Array stops. | |
149 | * This is detected by array_state going to 'clear' or 'inactive'. | |
150 | * while we thought it was active. | |
151 | * Response is to mark metadata as clean and 'clear' the array(??) | |
152 | * write-pending | |
153 | * array_state if 'write-pending' | |
154 | * We mark metadata as 'dirty' then set array to 'active'. | |
155 | * active_idle | |
156 | * Either ignore, or mark clean, then mark metadata as clean. | |
157 | * | |
158 | * device fails | |
159 | * detected by rd-N/state reporting "faulty" | |
160 | * mark device as 'failed' in metadata, let the kernel release the | |
161 | * device by writing '-blocked' to rd/state, and finally write 'remove' to | |
162 | * rd/state. Before a disk can be replaced it must be failed and removed | |
163 | * from all container members, this will be preemptive for the other | |
164 | * arrays... safe? | |
165 | * | |
166 | * sync completes | |
167 | * sync_action was 'resync' and becomes 'idle' and resync_start becomes | |
168 | * MaxSector | |
169 | * Notify metadata that sync is complete. | |
170 | * | |
171 | * recovery completes | |
172 | * sync_action changes from 'recover' to 'idle' | |
173 | * Check each device state and mark metadata if 'faulty' or 'in_sync'. | |
174 | * | |
175 | * deal with resync | |
176 | * This only happens on finding a new array... mdadm will have set | |
177 | * 'resync_start' to the correct value. If 'resync_start' indicates that an | |
178 | * resync needs to occur set the array to the 'active' state rather than the | |
179 | * initial read-auto state. | |
180 | * | |
181 | * | |
182 | * | |
183 | * We wait for a change (poll/select) on array_state, sync_action, and | |
184 | * each rd-X/state file. | |
185 | * When we get any change, we check everything. So read each state file, | |
186 | * then decide what to do. | |
187 | * | |
188 | * The core action is to write new metadata to all devices in the array. | |
189 | * This is done at most once on any wakeup. | |
190 | * After that we might: | |
191 | * - update the array_state | |
192 | * - set the role of some devices. | |
193 | * - request a sync_action | |
194 | * | |
195 | */ | |
196 | ||
197 | static int read_and_act(struct active_array *a) | |
198 | { | |
199 | int check_degraded = 0; | |
200 | int deactivate = 0; | |
201 | struct mdinfo *mdi; | |
202 | ||
203 | a->next_state = bad_word; | |
204 | a->next_action = bad_action; | |
205 | ||
206 | a->curr_state = read_state(a->info.state_fd); | |
207 | a->curr_action = read_action(a->action_fd); | |
208 | for (mdi = a->info.devs; mdi ; mdi = mdi->next) { | |
209 | mdi->next_state = 0; | |
210 | if (mdi->state_fd >= 0) | |
211 | mdi->curr_state = read_dev_state(mdi->state_fd); | |
212 | } | |
213 | ||
214 | if (a->curr_state <= inactive && | |
215 | a->prev_state > inactive) { | |
216 | /* array has been stopped */ | |
217 | get_resync_start(a); | |
218 | a->container->ss->set_array_state(a, 1); | |
219 | a->next_state = clear; | |
220 | deactivate = 1; | |
221 | } | |
222 | if (a->curr_state == write_pending) { | |
223 | get_resync_start(a); | |
224 | a->container->ss->set_array_state(a, 0); | |
225 | a->next_state = active; | |
226 | } | |
227 | if (a->curr_state == active_idle) { | |
228 | /* Set array to 'clean' FIRST, then mark clean | |
229 | * in the metadata | |
230 | */ | |
231 | a->next_state = clean; | |
232 | } | |
233 | if (a->curr_state == clean) { | |
234 | get_resync_start(a); | |
235 | a->container->ss->set_array_state(a, 1); | |
236 | } | |
237 | ||
238 | if (a->curr_state == readonly) { | |
239 | /* Well, I'm ready to handle things. If readonly | |
240 | * wasn't requested, transition to read-auto. | |
241 | */ | |
242 | char buf[64]; | |
243 | read_attr(buf, sizeof(buf), a->metadata_fd); | |
244 | if (strncmp(buf, "external:-", 10) == 0) { | |
245 | /* explicit request for readonly array. Leave it alone */ | |
246 | ; | |
247 | } else { | |
248 | get_resync_start(a); | |
249 | if (a->container->ss->set_array_state(a, 2)) | |
250 | a->next_state = read_auto; /* array is clean */ | |
251 | else | |
252 | a->next_state = active; /* Now active for recovery etc */ | |
253 | } | |
254 | } | |
255 | ||
256 | if (!deactivate && | |
257 | a->curr_action == idle && | |
258 | a->prev_action == resync) { | |
259 | /* A resync has finished. The endpoint is recorded in | |
260 | * 'sync_start'. We don't update the metadata | |
261 | * until the array goes inactive or readonly though. | |
262 | * Just check if we need to fiddle spares. | |
263 | */ | |
264 | get_resync_start(a); | |
265 | a->container->ss->set_array_state(a, a->curr_state <= clean); | |
266 | check_degraded = 1; | |
267 | } | |
268 | ||
269 | if (!deactivate && | |
270 | a->curr_action == idle && | |
271 | a->prev_action == recover) { | |
272 | /* A recovery has finished. Some disks may be in sync now, | |
273 | * and the array may no longer be degraded | |
274 | */ | |
275 | for (mdi = a->info.devs ; mdi ; mdi = mdi->next) { | |
276 | a->container->ss->set_disk(a, mdi->disk.raid_disk, | |
277 | mdi->curr_state); | |
278 | if (! (mdi->curr_state & DS_INSYNC)) | |
279 | check_degraded = 1; | |
280 | } | |
281 | } | |
282 | ||
283 | /* Check for failures and if found: | |
284 | * 1/ Record the failure in the metadata and unblock the device. | |
285 | * FIXME update the kernel to stop notifying on failed drives when | |
286 | * the array is readonly and we have cleared 'blocked' | |
287 | * 2/ Try to remove the device if the array is writable, or can be | |
288 | * made writable. | |
289 | */ | |
290 | for (mdi = a->info.devs ; mdi ; mdi = mdi->next) { | |
291 | if (mdi->curr_state & DS_FAULTY) { | |
292 | a->container->ss->set_disk(a, mdi->disk.raid_disk, | |
293 | mdi->curr_state); | |
294 | check_degraded = 1; | |
295 | mdi->next_state |= DS_UNBLOCK; | |
296 | if (a->curr_state == read_auto) { | |
297 | a->container->ss->set_array_state(a, 0); | |
298 | a->next_state = active; | |
299 | } | |
300 | if (a->curr_state > readonly) | |
301 | mdi->next_state |= DS_REMOVE; | |
302 | } | |
303 | } | |
304 | ||
305 | a->container->ss->sync_metadata(a->container); | |
306 | dprintf("%s(%d): state:%s action:%s next(", __func__, a->info.container_member, | |
307 | array_states[a->curr_state], sync_actions[a->curr_action]); | |
308 | ||
309 | /* Effect state changes in the array */ | |
310 | if (a->next_state != bad_word) { | |
311 | dprintf(" state:%s", array_states[a->next_state]); | |
312 | write_attr(array_states[a->next_state], a->info.state_fd); | |
313 | } | |
314 | if (a->next_action != bad_action) { | |
315 | write_attr(sync_actions[a->next_action], a->action_fd); | |
316 | dprintf(" action:%s", sync_actions[a->next_action]); | |
317 | } | |
318 | for (mdi = a->info.devs; mdi ; mdi = mdi->next) { | |
319 | if (mdi->next_state & DS_UNBLOCK) { | |
320 | dprintf(" %d:-blocked", mdi->disk.raid_disk); | |
321 | write_attr("-blocked", mdi->state_fd); | |
322 | } | |
323 | ||
324 | if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) { | |
325 | int remove_result; | |
326 | ||
327 | /* the kernel may not be able to immediately remove the | |
328 | * disk, we can simply wait until the next event to try | |
329 | * again. | |
330 | */ | |
331 | remove_result = write_attr("remove", mdi->state_fd); | |
332 | if (remove_result > 0) { | |
333 | dprintf(" %d:removed", mdi->disk.raid_disk); | |
334 | close(mdi->state_fd); | |
335 | mdi->state_fd = -1; | |
336 | } | |
337 | } | |
338 | if (mdi->next_state & DS_INSYNC) { | |
339 | write_attr("+in_sync", mdi->state_fd); | |
340 | dprintf(" %d:+in_sync", mdi->disk.raid_disk); | |
341 | } | |
342 | } | |
343 | dprintf(" )\n"); | |
344 | ||
345 | /* move curr_ to prev_ */ | |
346 | a->prev_state = a->curr_state; | |
347 | ||
348 | a->prev_action = a->curr_action; | |
349 | ||
350 | for (mdi = a->info.devs; mdi ; mdi = mdi->next) { | |
351 | mdi->prev_state = mdi->curr_state; | |
352 | mdi->next_state = 0; | |
353 | } | |
354 | ||
355 | if (check_degraded) { | |
356 | /* manager will do the actual check */ | |
357 | a->check_degraded = 1; | |
358 | signal_manager(); | |
359 | } | |
360 | ||
361 | if (deactivate) | |
362 | a->container = NULL; | |
363 | ||
364 | return 1; | |
365 | } | |
366 | ||
367 | static struct mdinfo * | |
368 | find_device(struct active_array *a, int major, int minor) | |
369 | { | |
370 | struct mdinfo *mdi; | |
371 | ||
372 | for (mdi = a->info.devs ; mdi ; mdi = mdi->next) | |
373 | if (mdi->disk.major == major && mdi->disk.minor == minor) | |
374 | return mdi; | |
375 | ||
376 | return NULL; | |
377 | } | |
378 | ||
379 | static void reconcile_failed(struct active_array *aa, struct mdinfo *failed) | |
380 | { | |
381 | struct active_array *a; | |
382 | struct mdinfo *victim; | |
383 | ||
384 | for (a = aa; a; a = a->next) { | |
385 | if (!a->container) | |
386 | continue; | |
387 | victim = find_device(a, failed->disk.major, failed->disk.minor); | |
388 | if (!victim) | |
389 | continue; | |
390 | ||
391 | if (!(victim->curr_state & DS_FAULTY)) | |
392 | write_attr("faulty", victim->state_fd); | |
393 | } | |
394 | } | |
395 | ||
396 | #ifdef DEBUG | |
397 | static void dprint_wake_reasons(fd_set *fds) | |
398 | { | |
399 | int i; | |
400 | char proc_path[256]; | |
401 | char link[256]; | |
402 | char *basename; | |
403 | int rv; | |
404 | ||
405 | fprintf(stderr, "monitor: wake ( "); | |
406 | for (i = 0; i < FD_SETSIZE; i++) { | |
407 | if (FD_ISSET(i, fds)) { | |
408 | sprintf(proc_path, "/proc/%d/fd/%d", | |
409 | (int) getpid(), i); | |
410 | ||
411 | rv = readlink(proc_path, link, sizeof(link) - 1); | |
412 | if (rv < 0) { | |
413 | fprintf(stderr, "%d:unknown ", i); | |
414 | continue; | |
415 | } | |
416 | link[rv] = '\0'; | |
417 | basename = strrchr(link, '/'); | |
418 | fprintf(stderr, "%d:%s ", | |
419 | i, basename ? ++basename : link); | |
420 | } | |
421 | } | |
422 | fprintf(stderr, ")\n"); | |
423 | } | |
424 | #endif | |
425 | ||
426 | int monitor_loop_cnt; | |
427 | ||
428 | static int wait_and_act(struct supertype *container, int nowait) | |
429 | { | |
430 | fd_set rfds; | |
431 | int maxfd = 0; | |
432 | struct active_array **aap = &container->arrays; | |
433 | struct active_array *a, **ap; | |
434 | int rv; | |
435 | struct mdinfo *mdi; | |
436 | static unsigned int dirty_arrays = ~0; /* start at some non-zero value */ | |
437 | ||
438 | FD_ZERO(&rfds); | |
439 | ||
440 | for (ap = aap ; *ap ;) { | |
441 | a = *ap; | |
442 | /* once an array has been deactivated we want to | |
443 | * ask the manager to discard it. | |
444 | */ | |
445 | if (!a->container) { | |
446 | if (discard_this) { | |
447 | ap = &(*ap)->next; | |
448 | continue; | |
449 | } | |
450 | *ap = a->next; | |
451 | a->next = NULL; | |
452 | discard_this = a; | |
453 | signal_manager(); | |
454 | continue; | |
455 | } | |
456 | ||
457 | add_fd(&rfds, &maxfd, a->info.state_fd); | |
458 | add_fd(&rfds, &maxfd, a->action_fd); | |
459 | for (mdi = a->info.devs ; mdi ; mdi = mdi->next) | |
460 | add_fd(&rfds, &maxfd, mdi->state_fd); | |
461 | ||
462 | ap = &(*ap)->next; | |
463 | } | |
464 | ||
465 | if (manager_ready && (*aap == NULL || (sigterm && !dirty_arrays))) { | |
466 | /* No interesting arrays, or we have been told to | |
467 | * terminate and everything is clean. Lets see about | |
468 | * exiting. Note that blocking at this point is not a | |
469 | * problem as there are no active arrays, there is | |
470 | * nothing that we need to be ready to do. | |
471 | */ | |
472 | int fd = open_dev_excl(container->devnum); | |
473 | if (fd >= 0 || errno != EBUSY) { | |
474 | /* OK, we are safe to leave */ | |
475 | if (sigterm && !dirty_arrays) | |
476 | dprintf("caught sigterm, all clean... exiting\n"); | |
477 | else | |
478 | dprintf("no arrays to monitor... exiting\n"); | |
479 | remove_pidfile(container->devname); | |
480 | exit_now = 1; | |
481 | signal_manager(); | |
482 | exit(0); | |
483 | } | |
484 | } | |
485 | ||
486 | if (!nowait) { | |
487 | sigset_t set; | |
488 | sigprocmask(SIG_UNBLOCK, NULL, &set); | |
489 | sigdelset(&set, SIGUSR1); | |
490 | monitor_loop_cnt |= 1; | |
491 | rv = pselect(maxfd+1, &rfds, NULL, NULL, NULL, &set); | |
492 | monitor_loop_cnt += 1; | |
493 | if (rv == -1 && errno == EINTR) | |
494 | rv = 0; | |
495 | #ifdef DEBUG | |
496 | dprint_wake_reasons(&rfds); | |
497 | #endif | |
498 | ||
499 | } | |
500 | ||
501 | if (update_queue) { | |
502 | struct metadata_update *this; | |
503 | ||
504 | for (this = update_queue; this ; this = this->next) | |
505 | container->ss->process_update(container, this); | |
506 | ||
507 | update_queue_handled = update_queue; | |
508 | update_queue = NULL; | |
509 | signal_manager(); | |
510 | container->ss->sync_metadata(container); | |
511 | } | |
512 | ||
513 | rv = 0; | |
514 | dirty_arrays = 0; | |
515 | for (a = *aap; a ; a = a->next) { | |
516 | int is_dirty; | |
517 | ||
518 | if (a->replaces && !discard_this) { | |
519 | struct active_array **ap; | |
520 | for (ap = &a->next; *ap && *ap != a->replaces; | |
521 | ap = & (*ap)->next) | |
522 | ; | |
523 | if (*ap) | |
524 | *ap = (*ap)->next; | |
525 | discard_this = a->replaces; | |
526 | a->replaces = NULL; | |
527 | /* FIXME check if device->state_fd need to be cleared?*/ | |
528 | signal_manager(); | |
529 | } | |
530 | if (a->container) | |
531 | rv += read_and_act(a); | |
532 | else | |
533 | continue; | |
534 | ||
535 | /* when terminating stop manipulating the array after it is | |
536 | * clean, but make sure read_and_act() is given a chance to | |
537 | * handle 'active_idle' | |
538 | */ | |
539 | switch (read_state(a->info.state_fd)) { | |
540 | case active: | |
541 | case active_idle: | |
542 | case suspended: | |
543 | case bad_word: | |
544 | is_dirty = 1; | |
545 | break; | |
546 | default: | |
547 | if (a->curr_state == active_idle) | |
548 | is_dirty = 1; | |
549 | else | |
550 | is_dirty = 0; | |
551 | break; | |
552 | } | |
553 | dirty_arrays += is_dirty; | |
554 | if (sigterm && !is_dirty) | |
555 | a->container = NULL; /* stop touching this array */ | |
556 | } | |
557 | ||
558 | /* propagate failures across container members */ | |
559 | for (a = *aap; a ; a = a->next) { | |
560 | if (!a->container) | |
561 | continue; | |
562 | for (mdi = a->info.devs ; mdi ; mdi = mdi->next) | |
563 | if (mdi->curr_state & DS_FAULTY) | |
564 | reconcile_failed(*aap, mdi); | |
565 | } | |
566 | ||
567 | return rv; | |
568 | } | |
569 | ||
570 | void do_monitor(struct supertype *container) | |
571 | { | |
572 | int rv; | |
573 | int first = 1; | |
574 | do { | |
575 | rv = wait_and_act(container, first); | |
576 | first = 0; | |
577 | } while (rv >= 0); | |
578 | } |