d6b1fbe1dda55ae84e7a8a4e36af2f6c1bbdd669
[thirdparty/mdadm.git] / monitor.c
1
2 #include "mdadm.h"
3 #include "mdmon.h"
4 #include <sys/syscall.h>
5 #include <sys/select.h>
6 #include <signal.h>
7
8 static char *array_states[] = {
9 "clear", "inactive", "suspended", "readonly", "read-auto",
10 "clean", "active", "write-pending", "active-idle", NULL };
11 static char *sync_actions[] = {
12 "idle", "reshape", "resync", "recover", "check", "repair", NULL
13 };
14
15 static int write_attr(char *attr, int fd)
16 {
17 return write(fd, attr, strlen(attr));
18 }
19
20 static 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
29 static 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
50
51 static int get_resync_start(struct active_array *a)
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 }
64
65 static 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
81 static 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
90 static 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
100 static 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
110 int 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) {
122 if (attr_match(cp, "faulty"))
123 rv |= DS_FAULTY;
124 if (attr_match(cp, "in_sync"))
125 rv |= DS_INSYNC;
126 if (attr_match(cp, "write_mostly"))
127 rv |= DS_WRITE_MOSTLY;
128 if (attr_match(cp, "spare"))
129 rv |= DS_SPARE;
130 if (attr_match(cp, "blocked"))
131 rv |= DS_BLOCKED;
132 cp = strchr(cp, ',');
133 if (cp)
134 cp++;
135 }
136 return rv;
137 }
138
139 static void signal_manager(void)
140 {
141 /* tgkill(getpid(), mon_tid, SIGUSR1); */
142 int pid = getpid();
143 syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
144 }
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"
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
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?
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.
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'.
180 *
181 * deal with resync
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.
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
203 static int read_and_act(struct active_array *a)
204 {
205 int check_degraded = 0;
206 int deactivate = 0;
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;
216 if (mdi->state_fd >= 0)
217 mdi->curr_state = read_dev_state(mdi->state_fd);
218 }
219
220 if (a->curr_state <= inactive &&
221 a->prev_state > inactive) {
222 /* array has been stopped */
223 a->container->ss->set_array_state(a, 1);
224 a->next_state = clear;
225 deactivate = 1;
226 }
227 if (a->curr_state == write_pending) {
228 get_resync_start(a);
229 a->container->ss->set_array_state(a, 0);
230 a->next_state = active;
231 }
232 if (a->curr_state == active_idle) {
233 /* Set array to 'clean' FIRST, then
234 * a->ss->mark_clean(a, ~0ULL);
235 * just ignore for now.
236 */
237 }
238
239 if (a->curr_state == readonly) {
240 /* Well, I'm ready to handle things, so
241 * read-auto is OK. FIXME what if we really want
242 * readonly ???
243 */
244 get_resync_start(a);
245 // printf("Found a readonly array at %llu\n", a->resync_start);
246 if (a->resync_start == ~0ULL)
247 a->next_state = read_auto; /* array is clean */
248 else {
249 a->container->ss->set_array_state(a, 0);
250 a->next_state = active;
251 }
252 }
253
254 if (!deactivate &&
255 a->curr_action == idle &&
256 a->prev_action == resync) {
257 /* A resync has finished. The endpoint is recorded in
258 * 'sync_start'. We don't update the metadata
259 * until the array goes inactive or readonly though.
260 * Just check if we need to fiddle spares.
261 */
262 get_resync_start(a);
263 a->container->ss->set_array_state(a, 0);
264 check_degraded = 1;
265 }
266
267 if (!deactivate &&
268 a->curr_action == idle &&
269 a->prev_action == recover) {
270 /* A recovery has finished. Some disks may be in sync now,
271 * and the array may no longer be degraded
272 */
273 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
274 a->container->ss->set_disk(a, mdi->disk.raid_disk,
275 mdi->curr_state);
276 if (! (mdi->curr_state & DS_INSYNC))
277 check_degraded = 1;
278 }
279 a->container->ss->set_array_state(a, 0);
280 }
281
282 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
283 if (mdi->curr_state & DS_FAULTY) {
284 a->container->ss->set_disk(a, mdi->disk.raid_disk,
285 mdi->curr_state);
286 check_degraded = 1;
287 mdi->next_state = DS_REMOVE;
288 }
289 }
290
291 a->container->ss->sync_metadata(a->container);
292 dprintf("%s: update[%d]: (", __func__, a->info.container_member);
293
294 /* Effect state changes in the array */
295 if (a->next_state != bad_word) {
296 dprintf(" state:%s", array_states[a->next_state]);
297 write_attr(array_states[a->next_state], a->info.state_fd);
298 }
299 if (a->next_action != bad_action) {
300 write_attr(sync_actions[a->next_action], a->action_fd);
301 dprintf(" action:%s", array_states[a->next_state]);
302 }
303 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
304 if (mdi->next_state == DS_REMOVE && mdi->state_fd >= 0) {
305 int remove_result;
306
307 write_attr("-blocked", mdi->state_fd);
308 /* the kernel may not be able to immediately remove the
309 * disk, we can simply wait until the next event to try
310 * again.
311 */
312 dprintf(" %d:-blocked", mdi->disk.raid_disk);
313 remove_result = write_attr("remove", mdi->state_fd);
314 if (remove_result > 0) {
315 dprintf(" %d:removed", mdi->disk.raid_disk);
316 close(mdi->state_fd);
317 mdi->state_fd = -1;
318 }
319 }
320 if (mdi->next_state & DS_INSYNC) {
321 write_attr("+in_sync", mdi->state_fd);
322 dprintf(" %d:+in_sync", mdi->disk.raid_disk);
323 }
324 }
325 dprintf(" )\n");
326
327 /* move curr_ to prev_ */
328 a->prev_state = a->curr_state;
329
330 a->prev_action = a->curr_action;
331
332 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
333 mdi->prev_state = mdi->curr_state;
334 mdi->next_state = 0;
335 }
336
337 if (check_degraded) {
338 /* manager will do the actual check */
339 a->check_degraded = 1;
340 signal_manager();
341 }
342
343 if (deactivate)
344 a->container = NULL;
345
346 return 1;
347 }
348
349 static struct mdinfo *
350 find_device(struct active_array *a, int major, int minor)
351 {
352 struct mdinfo *mdi;
353
354 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
355 if (mdi->disk.major == major && mdi->disk.minor == minor)
356 return mdi;
357
358 return NULL;
359 }
360
361 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
362 {
363 struct active_array *a;
364 struct mdinfo *victim;
365
366 for (a = aa; a; a = a->next) {
367 if (!a->container)
368 continue;
369 victim = find_device(a, failed->disk.major, failed->disk.minor);
370 if (!victim)
371 continue;
372
373 if (!(victim->curr_state & DS_FAULTY))
374 write_attr("faulty", victim->state_fd);
375 }
376 }
377
378 #ifdef DEBUG
379 static void dprint_wake_reasons(fd_set *fds)
380 {
381 int i;
382 char proc_path[256];
383 char link[256];
384 char *basename;
385 int rv;
386
387 fprintf(stderr, "monitor: wake ( ");
388 for (i = 0; i < FD_SETSIZE; i++) {
389 if (FD_ISSET(i, fds)) {
390 sprintf(proc_path, "/proc/%d/fd/%d",
391 (int) getpid(), i);
392
393 rv = readlink(proc_path, link, sizeof(link) - 1);
394 if (rv < 0) {
395 fprintf(stderr, "%d:unknown ", i);
396 continue;
397 }
398 link[rv] = '\0';
399 basename = strrchr(link, '/');
400 fprintf(stderr, "%d:%s ",
401 i, basename ? ++basename : link);
402 }
403 }
404 fprintf(stderr, ")\n");
405 }
406 #endif
407
408 static int wait_and_act(struct supertype *container, int nowait)
409 {
410 fd_set rfds;
411 int maxfd = 0;
412 struct active_array **aap = &container->arrays;
413 struct active_array *a, **ap;
414 int rv;
415 struct mdinfo *mdi;
416
417 FD_ZERO(&rfds);
418
419 for (ap = aap ; *ap ;) {
420 a = *ap;
421 /* once an array has been deactivated we want to
422 * ask the manager to discard it.
423 */
424 if (!a->container) {
425 if (discard_this) {
426 ap = &(*ap)->next;
427 continue;
428 }
429 *ap = a->next;
430 a->next = NULL;
431 discard_this = a;
432 signal_manager();
433 continue;
434 }
435
436 add_fd(&rfds, &maxfd, a->info.state_fd);
437 add_fd(&rfds, &maxfd, a->action_fd);
438 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
439 add_fd(&rfds, &maxfd, mdi->state_fd);
440
441 ap = &(*ap)->next;
442 }
443
444 if (manager_ready && *aap == NULL) {
445 /* No interesting arrays. Lets see about exiting.
446 * Note that blocking at this point is not a problem
447 * as there are no active arrays, there is nothing that
448 * we need to be ready to do.
449 */
450 int fd = open(container->device_name, O_RDONLY|O_EXCL);
451 if (fd >= 0 || errno != EBUSY) {
452 /* OK, we are safe to leave */
453 dprintf("no arrays to monitor... exiting\n");
454 remove_pidfile(container->devname);
455 exit_now = 1;
456 signal_manager();
457 exit(0);
458 }
459 }
460
461 if (!nowait) {
462 sigset_t set;
463 sigprocmask(SIG_UNBLOCK, NULL, &set);
464 sigdelset(&set, SIGUSR1);
465 rv = pselect(maxfd+1, &rfds, NULL, NULL, NULL, &set);
466 if (rv == -1 && errno == EINTR)
467 rv = 0;
468 #ifdef DEBUG
469 dprint_wake_reasons(&rfds);
470 #endif
471
472 }
473
474 if (update_queue) {
475 struct metadata_update *this;
476
477 for (this = update_queue; this ; this = this->next)
478 container->ss->process_update(container, this);
479
480 update_queue_handled = update_queue;
481 update_queue = NULL;
482 signal_manager();
483 container->ss->sync_metadata(container);
484 }
485
486 for (a = *aap; a ; a = a->next) {
487 if (a->replaces && !discard_this) {
488 struct active_array **ap;
489 for (ap = &a->next; *ap && *ap != a->replaces;
490 ap = & (*ap)->next)
491 ;
492 if (*ap)
493 *ap = (*ap)->next;
494 discard_this = a->replaces;
495 a->replaces = NULL;
496 /* FIXME check if device->state_fd need to be cleared?*/
497 signal_manager();
498 }
499 if (a->container)
500 rv += read_and_act(a);
501 }
502
503 /* propagate failures across container members */
504 for (a = *aap; a ; a = a->next) {
505 if (!a->container)
506 continue;
507 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
508 if (mdi->curr_state & DS_FAULTY)
509 reconcile_failed(*aap, mdi);
510 }
511
512 return rv;
513 }
514
515 void do_monitor(struct supertype *container)
516 {
517 int rv;
518 int first = 1;
519 do {
520 rv = wait_and_act(container, first);
521 first = 0;
522 } while (rv >= 0);
523 }