mdmon: allow degraded arrays to be monitored
[thirdparty/mdadm.git] / monitor.c
CommitLineData
549e9569
NB
1
2#include "mdadm.h"
3#include "mdmon.h"
4d43913c 4#include <sys/syscall.h>
549e9569 5#include <sys/select.h>
1ed3f387 6#include <signal.h>
549e9569
NB
7
8static char *array_states[] = {
9 "clear", "inactive", "suspended", "readonly", "read-auto",
10 "clean", "active", "write-pending", "active-idle", NULL };
11static char *sync_actions[] = {
12 "idle", "reshape", "resync", "recover", "check", "repair", NULL
13};
14
15static int write_attr(char *attr, int fd)
16{
17 return write(fd, attr, strlen(attr));
18}
19
20static void add_fd(fd_set *fds, int *maxfd, int fd)
21{
22 if (fd < 0)
23 return;
24 if (fd > *maxfd)
25 *maxfd = fd;
26 FD_SET(fd, fds);
27}
28
29static int read_attr(char *buf, int len, int fd)
30{
31 int n;
32
33 if (fd < 0) {
34 buf[0] = 0;
35 return 0;
36 }
37 lseek(fd, 0, 0);
38 n = read(fd, buf, len - 1);
39
40 if (n <= 0) {
41 buf[0] = 0;
42 return 0;
43 }
44 buf[n] = 0;
45 if (buf[n-1] == '\n')
46 buf[n-1] = 0;
47 return n;
48}
49
549e9569 50
103f2410 51int get_resync_start(struct active_array *a)
c052ba30
DW
52{
53 char buf[30];
54 int n;
55
56 n = read_attr(buf, 30, a->resync_start_fd);
57 if (n <= 0)
58 return n;
59
60 a->resync_start = strtoull(buf, NULL, 10);
61
62 return 1;
63}
549e9569
NB
64
65static int attr_match(const char *attr, const char *str)
66{
67 /* See if attr, read from a sysfs file, matches
68 * str. They must either be the same, or attr can
69 * have a trailing newline or comma
70 */
71 while (*attr && *str && *attr == *str) {
72 attr++;
73 str++;
74 }
75
76 if (*str || (*attr && *attr != ',' && *attr != '\n'))
77 return 0;
78 return 1;
79}
80
81static int match_word(const char *word, char **list)
82{
83 int n;
84 for (n=0; list[n]; n++)
85 if (attr_match(word, list[n]))
86 break;
87 return n;
88}
89
90static enum array_state read_state(int fd)
91{
92 char buf[20];
93 int n = read_attr(buf, 20, fd);
94
95 if (n <= 0)
96 return bad_word;
97 return (enum array_state) match_word(buf, array_states);
98}
99
100static enum sync_action read_action( int fd)
101{
102 char buf[20];
103 int n = read_attr(buf, 20, fd);
104
105 if (n <= 0)
106 return bad_action;
107 return (enum sync_action) match_word(buf, sync_actions);
108}
109
549e9569
NB
110int read_dev_state(int fd)
111{
112 char buf[60];
113 int n = read_attr(buf, 60, fd);
114 char *cp;
115 int rv = 0;
116
117 if (n <= 0)
118 return 0;
119
120 cp = buf;
121 while (cp) {
8d45d196 122 if (attr_match(cp, "faulty"))
549e9569 123 rv |= DS_FAULTY;
8d45d196 124 if (attr_match(cp, "in_sync"))
549e9569 125 rv |= DS_INSYNC;
8d45d196 126 if (attr_match(cp, "write_mostly"))
549e9569 127 rv |= DS_WRITE_MOSTLY;
8d45d196 128 if (attr_match(cp, "spare"))
549e9569 129 rv |= DS_SPARE;
8d45d196
DW
130 if (attr_match(cp, "blocked"))
131 rv |= DS_BLOCKED;
549e9569
NB
132 cp = strchr(cp, ',');
133 if (cp)
134 cp++;
135 }
136 return rv;
137}
138
1ed3f387
NB
139static void signal_manager(void)
140{
4d43913c
NB
141 /* tgkill(getpid(), mon_tid, SIGUSR1); */
142 int pid = getpid();
143 syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
1ed3f387 144}
549e9569
NB
145
146/* Monitor a set of active md arrays - all of which share the
147 * same metadata - and respond to events that require
148 * metadata update.
149 *
150 * New arrays are detected by another thread which allocates
151 * required memory and attaches the data structure to our list.
152 *
153 * Events:
154 * Array stops.
155 * This is detected by array_state going to 'clear' or 'inactive'.
156 * while we thought it was active.
157 * Response is to mark metadata as clean and 'clear' the array(??)
158 * write-pending
159 * array_state if 'write-pending'
160 * We mark metadata as 'dirty' then set array to 'active'.
161 * active_idle
162 * Either ignore, or mark clean, then mark metadata as clean.
163 *
164 * device fails
165 * detected by rd-N/state reporting "faulty"
8d45d196
DW
166 * mark device as 'failed' in metadata, let the kernel release the
167 * device by writing '-blocked' to rd/state, and finally write 'remove' to
0af73f61
DW
168 * rd/state. Before a disk can be replaced it must be failed and removed
169 * from all container members, this will be preemptive for the other
170 * arrays... safe?
549e9569
NB
171 *
172 * sync completes
173 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
174 * MaxSector
175 * Notify metadata that sync is complete.
549e9569
NB
176 *
177 * recovery completes
178 * sync_action changes from 'recover' to 'idle'
179 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
549e9569
NB
180 *
181 * deal with resync
c052ba30
DW
182 * This only happens on finding a new array... mdadm will have set
183 * 'resync_start' to the correct value. If 'resync_start' indicates that an
184 * resync needs to occur set the array to the 'active' state rather than the
185 * initial read-auto state.
549e9569
NB
186 *
187 *
188 *
189 * We wait for a change (poll/select) on array_state, sync_action, and
190 * each rd-X/state file.
191 * When we get any change, we check everything. So read each state file,
192 * then decide what to do.
193 *
194 * The core action is to write new metadata to all devices in the array.
195 * This is done at most once on any wakeup.
196 * After that we might:
197 * - update the array_state
198 * - set the role of some devices.
199 * - request a sync_action
200 *
201 */
202
203static int read_and_act(struct active_array *a)
204{
6c3fb95c 205 int check_degraded = 0;
2a0bb19e 206 int deactivate = 0;
549e9569
NB
207 struct mdinfo *mdi;
208
209 a->next_state = bad_word;
210 a->next_action = bad_action;
211
212 a->curr_state = read_state(a->info.state_fd);
213 a->curr_action = read_action(a->action_fd);
214 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
215 mdi->next_state = 0;
57632f4a 216 if (mdi->state_fd >= 0)
8d45d196 217 mdi->curr_state = read_dev_state(mdi->state_fd);
549e9569
NB
218 }
219
220 if (a->curr_state <= inactive &&
221 a->prev_state > inactive) {
222 /* array has been stopped */
33af8567 223 get_resync_start(a);
ed9d66aa 224 a->container->ss->set_array_state(a, 1);
549e9569 225 a->next_state = clear;
2a0bb19e 226 deactivate = 1;
549e9569
NB
227 }
228 if (a->curr_state == write_pending) {
ed9d66aa
NB
229 get_resync_start(a);
230 a->container->ss->set_array_state(a, 0);
549e9569
NB
231 a->next_state = active;
232 }
233 if (a->curr_state == active_idle) {
234 /* Set array to 'clean' FIRST, then
4e5528c6 235 * a->ss->mark_clean(a, ~0ULL);
549e9569
NB
236 * just ignore for now.
237 */
238 }
239
240 if (a->curr_state == readonly) {
241 /* Well, I'm ready to handle things, so
242 * read-auto is OK. FIXME what if we really want
243 * readonly ???
244 */
c052ba30 245 get_resync_start(a);
272906ef 246 if (a->resync_start == ~0ULL) {
c052ba30 247 a->next_state = read_auto; /* array is clean */
272906ef
DW
248 /* give the metadata a chance to force active if
249 * we have some recovery to do. metadata sets
250 * resync_start to !MaxSector in this case
251 */
252 a->container->ss->set_array_state(a, 1);
253 }
254 if (a->resync_start != ~0ULL) {
ed9d66aa 255 a->container->ss->set_array_state(a, 0);
c052ba30
DW
256 a->next_state = active;
257 }
549e9569
NB
258 }
259
00e02142
DW
260 if (!deactivate &&
261 a->curr_action == idle &&
549e9569 262 a->prev_action == resync) {
4e5528c6
NB
263 /* A resync has finished. The endpoint is recorded in
264 * 'sync_start'. We don't update the metadata
265 * until the array goes inactive or readonly though.
266 * Just check if we need to fiddle spares.
267 */
77402e51 268 get_resync_start(a);
ed9d66aa 269 a->container->ss->set_array_state(a, 0);
549e9569
NB
270 check_degraded = 1;
271 }
272
00e02142
DW
273 if (!deactivate &&
274 a->curr_action == idle &&
549e9569 275 a->prev_action == recover) {
0a6bdbee
DW
276 /* A recovery has finished. Some disks may be in sync now,
277 * and the array may no longer be degraded
278 */
549e9569 279 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
8d45d196
DW
280 a->container->ss->set_disk(a, mdi->disk.raid_disk,
281 mdi->curr_state);
549e9569
NB
282 if (! (mdi->curr_state & DS_INSYNC))
283 check_degraded = 1;
284 }
285 }
286
549e9569
NB
287 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
288 if (mdi->curr_state & DS_FAULTY) {
8d45d196
DW
289 a->container->ss->set_disk(a, mdi->disk.raid_disk,
290 mdi->curr_state);
549e9569
NB
291 check_degraded = 1;
292 mdi->next_state = DS_REMOVE;
293 }
294 }
295
2e735d19 296 a->container->ss->sync_metadata(a->container);
4e6e574a 297 dprintf("%s: update[%d]: (", __func__, a->info.container_member);
549e9569
NB
298
299 /* Effect state changes in the array */
4e6e574a
DW
300 if (a->next_state != bad_word) {
301 dprintf(" state:%s", array_states[a->next_state]);
549e9569 302 write_attr(array_states[a->next_state], a->info.state_fd);
4e6e574a
DW
303 }
304 if (a->next_action != bad_action) {
549e9569 305 write_attr(sync_actions[a->next_action], a->action_fd);
4e6e574a
DW
306 dprintf(" action:%s", array_states[a->next_state]);
307 }
549e9569 308 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
57632f4a
NB
309 if (mdi->next_state == DS_REMOVE && mdi->state_fd >= 0) {
310 int remove_result;
8d45d196
DW
311
312 write_attr("-blocked", mdi->state_fd);
313 /* the kernel may not be able to immediately remove the
314 * disk, we can simply wait until the next event to try
315 * again.
316 */
4e6e574a 317 dprintf(" %d:-blocked", mdi->disk.raid_disk);
57632f4a
NB
318 remove_result = write_attr("remove", mdi->state_fd);
319 if (remove_result > 0) {
4e6e574a 320 dprintf(" %d:removed", mdi->disk.raid_disk);
8d45d196
DW
321 close(mdi->state_fd);
322 mdi->state_fd = -1;
323 }
324 }
4e6e574a 325 if (mdi->next_state & DS_INSYNC) {
549e9569 326 write_attr("+in_sync", mdi->state_fd);
4e6e574a
DW
327 dprintf(" %d:+in_sync", mdi->disk.raid_disk);
328 }
549e9569 329 }
4e6e574a 330 dprintf(" )\n");
549e9569
NB
331
332 /* move curr_ to prev_ */
333 a->prev_state = a->curr_state;
334
335 a->prev_action = a->curr_action;
336
337 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
338 mdi->prev_state = mdi->curr_state;
339 mdi->next_state = 0;
340 }
341
7e1432fb
NB
342 if (check_degraded) {
343 /* manager will do the actual check */
344 a->check_degraded = 1;
345 signal_manager();
346 }
347
2a0bb19e
DW
348 if (deactivate)
349 a->container = NULL;
350
549e9569
NB
351 return 1;
352}
353
0af73f61
DW
354static struct mdinfo *
355find_device(struct active_array *a, int major, int minor)
356{
357 struct mdinfo *mdi;
358
359 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
360 if (mdi->disk.major == major && mdi->disk.minor == minor)
361 return mdi;
362
363 return NULL;
364}
365
366static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
367{
368 struct active_array *a;
369 struct mdinfo *victim;
370
371 for (a = aa; a; a = a->next) {
372 if (!a->container)
373 continue;
374 victim = find_device(a, failed->disk.major, failed->disk.minor);
375 if (!victim)
376 continue;
377
378 if (!(victim->curr_state & DS_FAULTY))
379 write_attr("faulty", victim->state_fd);
380 }
381}
382
4e6e574a
DW
383#ifdef DEBUG
384static void dprint_wake_reasons(fd_set *fds)
385{
386 int i;
387 char proc_path[256];
388 char link[256];
389 char *basename;
390 int rv;
391
392 fprintf(stderr, "monitor: wake ( ");
393 for (i = 0; i < FD_SETSIZE; i++) {
394 if (FD_ISSET(i, fds)) {
395 sprintf(proc_path, "/proc/%d/fd/%d",
396 (int) getpid(), i);
397
398 rv = readlink(proc_path, link, sizeof(link) - 1);
399 if (rv < 0) {
400 fprintf(stderr, "%d:unknown ", i);
401 continue;
402 }
403 link[rv] = '\0';
404 basename = strrchr(link, '/');
405 fprintf(stderr, "%d:%s ",
406 i, basename ? ++basename : link);
407 }
408 }
409 fprintf(stderr, ")\n");
410}
411#endif
412
1eb252b8
N
413int monitor_loop_cnt;
414
4d43913c 415static int wait_and_act(struct supertype *container, int nowait)
549e9569
NB
416{
417 fd_set rfds;
418 int maxfd = 0;
e0d6609f 419 struct active_array **aap = &container->arrays;
1ed3f387 420 struct active_array *a, **ap;
549e9569 421 int rv;
0af73f61 422 struct mdinfo *mdi;
549e9569
NB
423
424 FD_ZERO(&rfds);
425
1ed3f387
NB
426 for (ap = aap ; *ap ;) {
427 a = *ap;
428 /* once an array has been deactivated we want to
429 * ask the manager to discard it.
2a0bb19e 430 */
1ed3f387
NB
431 if (!a->container) {
432 if (discard_this) {
433 ap = &(*ap)->next;
434 continue;
435 }
436 *ap = a->next;
437 a->next = NULL;
438 discard_this = a;
439 signal_manager();
2a0bb19e 440 continue;
1ed3f387 441 }
2a0bb19e 442
549e9569
NB
443 add_fd(&rfds, &maxfd, a->info.state_fd);
444 add_fd(&rfds, &maxfd, a->action_fd);
445 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
446 add_fd(&rfds, &maxfd, mdi->state_fd);
1ed3f387
NB
447
448 ap = &(*ap)->next;
549e9569
NB
449 }
450
e0d6609f
NB
451 if (manager_ready && *aap == NULL) {
452 /* No interesting arrays. Lets see about exiting.
453 * Note that blocking at this point is not a problem
454 * as there are no active arrays, there is nothing that
455 * we need to be ready to do.
456 */
457 int fd = open(container->device_name, O_RDONLY|O_EXCL);
458 if (fd >= 0 || errno != EBUSY) {
459 /* OK, we are safe to leave */
4e6e574a 460 dprintf("no arrays to monitor... exiting\n");
207aac36 461 remove_pidfile(container->devname);
e0d6609f
NB
462 exit_now = 1;
463 signal_manager();
e0d6609f
NB
464 exit(0);
465 }
466 }
467
549e9569 468 if (!nowait) {
4d43913c
NB
469 sigset_t set;
470 sigprocmask(SIG_UNBLOCK, NULL, &set);
471 sigdelset(&set, SIGUSR1);
1eb252b8 472 monitor_loop_cnt |= 1;
4d43913c 473 rv = pselect(maxfd+1, &rfds, NULL, NULL, NULL, &set);
1eb252b8 474 monitor_loop_cnt += 1;
bfa44e2e
NB
475 if (rv == -1 && errno == EINTR)
476 rv = 0;
4e6e574a
DW
477 #ifdef DEBUG
478 dprint_wake_reasons(&rfds);
479 #endif
480
549e9569
NB
481 }
482
2e735d19
NB
483 if (update_queue) {
484 struct metadata_update *this;
485
486 for (this = update_queue; this ; this = this->next)
487 container->ss->process_update(container, this);
488
489 update_queue_handled = update_queue;
490 update_queue = NULL;
491 signal_manager();
492 container->ss->sync_metadata(container);
493 }
494
1ed3f387 495 for (a = *aap; a ; a = a->next) {
2a0bb19e 496 if (a->replaces && !discard_this) {
549e9569
NB
497 struct active_array **ap;
498 for (ap = &a->next; *ap && *ap != a->replaces;
499 ap = & (*ap)->next)
500 ;
501 if (*ap)
502 *ap = (*ap)->next;
503 discard_this = a->replaces;
504 a->replaces = NULL;
6c3fb95c 505 /* FIXME check if device->state_fd need to be cleared?*/
1ed3f387 506 signal_manager();
549e9569 507 }
2a0bb19e
DW
508 if (a->container)
509 rv += read_and_act(a);
549e9569 510 }
0af73f61
DW
511
512 /* propagate failures across container members */
1ed3f387 513 for (a = *aap; a ; a = a->next) {
0af73f61
DW
514 if (!a->container)
515 continue;
516 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
517 if (mdi->curr_state & DS_FAULTY)
1ed3f387 518 reconcile_failed(*aap, mdi);
0af73f61
DW
519 }
520
549e9569
NB
521 return rv;
522}
523
524void do_monitor(struct supertype *container)
525{
526 int rv;
527 int first = 1;
528 do {
4d43913c 529 rv = wait_and_act(container, first);
549e9569
NB
530 first = 0;
531 } while (rv >= 0);
532}