2 * mdmon - monitor external metadata arrays
4 * Copyright (C) 2007-2009 Neil Brown <neilb@suse.de>
5 * Copyright (C) 2007-2009 Intel Corporation
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.
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
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.
23 #include <sys/syscall.h>
24 #include <sys/select.h>
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
34 static int write_attr(char *attr, int fd)
36 return write(fd, attr, strlen(attr));
39 static void add_fd(fd_set *fds, int *maxfd, int fd)
44 if (fstat(fd, &st) == -1) {
45 dprintf("Invalid fd %d\n", fd);
48 if (st.st_nlink == 0) {
49 dprintf("fd %d was deleted\n", fd);
57 static int read_attr(char *buf, int len, int fd)
66 n = read(fd, buf, len - 1);
78 static void read_resync_start(int fd, unsigned long long *v)
83 n = read_attr(buf, 30, fd);
85 dprintf("Failed to read resync_start (%d)\n", fd);
88 if (strncmp(buf, "none", 4) == 0)
91 *v = strtoull(buf, NULL, 10);
94 static unsigned long long read_sync_completed(int fd)
96 unsigned long long val;
101 n = read_attr(buf, 50, fd);
106 val = strtoull(buf, &ep, 0);
107 if (ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))
112 static enum array_state read_state(int fd)
115 int n = read_attr(buf, 20, fd);
119 return (enum array_state) sysfs_match_word(buf, array_states);
122 static enum sync_action read_action( int fd)
125 int n = read_attr(buf, 20, fd);
129 return (enum sync_action) sysfs_match_word(buf, sync_actions);
132 int read_dev_state(int fd)
135 int n = read_attr(buf, sizeof(buf), fd);
144 if (sysfs_attr_match(cp, "faulty"))
146 if (sysfs_attr_match(cp, "in_sync"))
148 if (sysfs_attr_match(cp, "write_mostly"))
149 rv |= DS_WRITE_MOSTLY;
150 if (sysfs_attr_match(cp, "spare"))
152 if (sysfs_attr_match(cp, "blocked"))
154 cp = strchr(cp, ',');
161 static void signal_manager(void)
163 /* tgkill(getpid(), mon_tid, SIGUSR1); */
165 syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
168 /* Monitor a set of active md arrays - all of which share the
169 * same metadata - and respond to events that require
172 * New arrays are detected by another thread which allocates
173 * required memory and attaches the data structure to our list.
177 * This is detected by array_state going to 'clear' or 'inactive'.
178 * while we thought it was active.
179 * Response is to mark metadata as clean and 'clear' the array(??)
181 * array_state if 'write-pending'
182 * We mark metadata as 'dirty' then set array to 'active'.
184 * Either ignore, or mark clean, then mark metadata as clean.
187 * detected by rd-N/state reporting "faulty"
188 * mark device as 'failed' in metadata, let the kernel release the
189 * device by writing '-blocked' to rd/state, and finally write 'remove' to
190 * rd/state. Before a disk can be replaced it must be failed and removed
191 * from all container members, this will be preemptive for the other
195 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
197 * Notify metadata that sync is complete.
200 * sync_action changes from 'recover' to 'idle'
201 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
204 * This only happens on finding a new array... mdadm will have set
205 * 'resync_start' to the correct value. If 'resync_start' indicates that an
206 * resync needs to occur set the array to the 'active' state rather than the
207 * initial read-auto state.
211 * We wait for a change (poll/select) on array_state, sync_action, and
212 * each rd-X/state file.
213 * When we get any change, we check everything. So read each state file,
214 * then decide what to do.
216 * The core action is to write new metadata to all devices in the array.
217 * This is done at most once on any wakeup.
218 * After that we might:
219 * - update the array_state
220 * - set the role of some devices.
221 * - request a sync_action
225 #define ARRAY_DIRTY 1
227 static int read_and_act(struct active_array *a)
229 unsigned long long sync_completed;
230 int check_degraded = 0;
231 int check_reshape = 0;
238 a->next_state = bad_word;
239 a->next_action = bad_action;
241 a->curr_state = read_state(a->info.state_fd);
242 a->curr_action = read_action(a->action_fd);
243 if (a->curr_state != clear)
245 * In "clear" state, resync_start may wrongly be set to "0"
246 * when the kernel called md_clean but didn't remove the
247 * sysfs attributes yet
249 read_resync_start(a->resync_start_fd, &a->info.resync_start);
250 sync_completed = read_sync_completed(a->sync_completed_fd);
251 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
254 if (mdi->state_fd >= 0) {
255 read_resync_start(mdi->recovery_fd,
256 &mdi->recovery_start);
257 mdi->curr_state = read_dev_state(mdi->state_fd);
261 gettimeofday(&tv, NULL);
262 dprintf("(%d): %ld.%06ld state:%s prev:%s action:%s prev: %s start:%llu\n",
263 a->info.container_member,
264 tv.tv_sec, tv.tv_usec,
265 array_states[a->curr_state],
266 array_states[a->prev_state],
267 sync_actions[a->curr_action],
268 sync_actions[a->prev_action],
272 if ((a->curr_state == bad_word || a->curr_state <= inactive) &&
273 a->prev_state > inactive) {
274 /* array has been stopped */
275 a->container->ss->set_array_state(a, 1);
276 a->next_state = clear;
279 if (a->curr_state == write_pending) {
280 a->container->ss->set_array_state(a, 0);
281 a->next_state = active;
284 if (a->curr_state == active_idle) {
285 /* Set array to 'clean' FIRST, then mark clean
288 a->next_state = clean;
291 if (a->curr_state == clean) {
292 a->container->ss->set_array_state(a, 1);
294 if (a->curr_state == active ||
295 a->curr_state == suspended)
297 if (a->curr_state == readonly) {
298 /* Well, I'm ready to handle things. If readonly
299 * wasn't requested, transition to read-auto.
302 read_attr(buf, sizeof(buf), a->metadata_fd);
303 if (strncmp(buf, "external:-", 10) == 0) {
304 /* explicit request for readonly array. Leave it alone */
307 if (a->container->ss->set_array_state(a, 2))
308 a->next_state = read_auto; /* array is clean */
310 a->next_state = active; /* Now active for recovery etc */
317 a->curr_action == idle &&
318 a->prev_action == resync) {
319 /* A resync has finished. The endpoint is recorded in
320 * 'sync_start'. We don't update the metadata
321 * until the array goes inactive or readonly though.
322 * Just check if we need to fiddle spares.
324 a->container->ss->set_array_state(a, a->curr_state <= clean);
329 a->curr_action == idle &&
330 a->prev_action == recover) {
331 /* A recovery has finished. Some disks may be in sync now,
332 * and the array may no longer be degraded
334 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
335 a->container->ss->set_disk(a, mdi->disk.raid_disk,
337 if (! (mdi->curr_state & DS_INSYNC))
341 if (count != a->info.array.raid_disks)
346 a->curr_action == reshape &&
347 a->prev_action != reshape)
348 /* reshape was requested by mdadm. Need to see if
349 * new devices have been added. Manager does that
350 * when it sees check_reshape
354 /* Check for failures and if found:
355 * 1/ Record the failure in the metadata and unblock the device.
356 * FIXME update the kernel to stop notifying on failed drives when
357 * the array is readonly and we have cleared 'blocked'
358 * 2/ Try to remove the device if the array is writable, or can be
361 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
362 if (mdi->curr_state & DS_FAULTY) {
363 a->container->ss->set_disk(a, mdi->disk.raid_disk,
366 if (mdi->curr_state & DS_BLOCKED)
367 mdi->next_state |= DS_UNBLOCK;
368 if (a->curr_state == read_auto) {
369 a->container->ss->set_array_state(a, 0);
370 a->next_state = active;
372 if (a->curr_state > readonly)
373 mdi->next_state |= DS_REMOVE;
377 /* Check for recovery checkpoint notifications. We need to be a
378 * minimum distance away from the last checkpoint to prevent
379 * over checkpointing. Note reshape checkpointing is handled
380 * in the second branch.
382 if (sync_completed > a->last_checkpoint &&
383 sync_completed - a->last_checkpoint > a->info.component_size >> 4 &&
384 a->curr_action > reshape) {
385 /* A (non-reshape) sync_action has reached a checkpoint.
386 * Record the updated position in the metadata
388 a->last_checkpoint = sync_completed;
389 a->container->ss->set_array_state(a, a->curr_state <= clean);
390 } else if ((a->curr_action == idle && a->prev_action == reshape) ||
391 (a->curr_action == reshape
392 && sync_completed > a->last_checkpoint) ) {
393 /* Reshape has progressed or completed so we need to
394 * update the array state - and possibly the array size
396 if (sync_completed != 0)
397 a->last_checkpoint = sync_completed;
398 /* We might need to update last_checkpoint depending on
399 * the reason that reshape finished.
400 * if array reshape is really finished:
401 * set check point to the end, this allows
402 * set_array_state() to finalize reshape in metadata
403 * if reshape if broken: do not set checkpoint to the end
404 * this allows for reshape restart from checkpoint
406 if ((a->curr_action != reshape) &&
407 (a->prev_action == reshape)) {
409 if ((sysfs_get_str(&a->info, NULL,
412 sizeof(buf)) >= 0) &&
413 strncmp(buf, "none", 4) == 0)
414 a->last_checkpoint = a->info.component_size;
416 a->container->ss->set_array_state(a, a->curr_state <= clean);
417 a->last_checkpoint = sync_completed;
420 if (sync_completed > a->last_checkpoint)
421 a->last_checkpoint = sync_completed;
423 if (sync_completed >= a->info.component_size)
424 a->last_checkpoint = 0;
426 a->container->ss->sync_metadata(a->container);
427 dprintf("(%d): state:%s action:%s next(", a->info.container_member,
428 array_states[a->curr_state], sync_actions[a->curr_action]);
430 /* Effect state changes in the array */
431 if (a->next_state != bad_word) {
432 dprintf_cont(" state:%s", array_states[a->next_state]);
433 write_attr(array_states[a->next_state], a->info.state_fd);
435 if (a->next_action != bad_action) {
436 write_attr(sync_actions[a->next_action], a->action_fd);
437 dprintf_cont(" action:%s", sync_actions[a->next_action]);
439 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
440 if (mdi->next_state & DS_UNBLOCK) {
441 dprintf_cont(" %d:-blocked", mdi->disk.raid_disk);
442 write_attr("-blocked", mdi->state_fd);
445 if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
448 /* The kernel may not be able to immediately remove the
449 * disk. In that case we wait a little while and
452 remove_result = write_attr("remove", mdi->state_fd);
453 if (remove_result > 0) {
454 dprintf_cont(" %d:removed", mdi->disk.raid_disk);
455 close(mdi->state_fd);
456 close(mdi->recovery_fd);
461 if (mdi->next_state & DS_INSYNC) {
462 write_attr("+in_sync", mdi->state_fd);
463 dprintf_cont(" %d:+in_sync", mdi->disk.raid_disk);
466 dprintf_cont(" )\n");
468 /* move curr_ to prev_ */
469 a->prev_state = a->curr_state;
471 a->prev_action = a->curr_action;
473 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
474 mdi->prev_state = mdi->curr_state;
478 if (check_degraded || check_reshape) {
479 /* manager will do the actual check */
481 a->check_degraded = 1;
483 a->check_reshape = 1;
493 static struct mdinfo *
494 find_device(struct active_array *a, int major, int minor)
498 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
499 if (mdi->disk.major == major && mdi->disk.minor == minor)
505 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
507 struct active_array *a;
508 struct mdinfo *victim;
510 for (a = aa; a; a = a->next) {
511 if (!a->container || a->to_remove)
513 victim = find_device(a, failed->disk.major, failed->disk.minor);
517 if (!(victim->curr_state & DS_FAULTY))
518 write_attr("faulty", victim->state_fd);
523 static void dprint_wake_reasons(fd_set *fds)
531 fprintf(stderr, "monitor: wake ( ");
532 for (i = 0; i < FD_SETSIZE; i++) {
533 if (FD_ISSET(i, fds)) {
534 sprintf(proc_path, "/proc/%d/fd/%d",
537 rv = readlink(proc_path, link, sizeof(link) - 1);
539 fprintf(stderr, "%d:unknown ", i);
543 basename = strrchr(link, '/');
544 fprintf(stderr, "%d:%s ",
545 i, basename ? ++basename : link);
548 fprintf(stderr, ")\n");
552 int monitor_loop_cnt;
554 static int wait_and_act(struct supertype *container, int nowait)
558 struct active_array **aap = &container->arrays;
559 struct active_array *a, **ap;
562 static unsigned int dirty_arrays = ~0; /* start at some non-zero value */
566 for (ap = aap ; *ap ;) {
568 /* once an array has been deactivated we want to
569 * ask the manager to discard it.
571 if (!a->container || a->to_remove) {
583 add_fd(&rfds, &maxfd, a->info.state_fd);
584 add_fd(&rfds, &maxfd, a->action_fd);
585 add_fd(&rfds, &maxfd, a->sync_completed_fd);
586 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
587 add_fd(&rfds, &maxfd, mdi->state_fd);
592 if (manager_ready && (*aap == NULL || (sigterm && !dirty_arrays))) {
593 /* No interesting arrays, or we have been told to
594 * terminate and everything is clean. Lets see about
595 * exiting. Note that blocking at this point is not a
596 * problem as there are no active arrays, there is
597 * nothing that we need to be ready to do.
601 fd = open_dev_excl(container->devnm);
603 fd = open_dev_flags(container->devnm, O_RDONLY|O_EXCL);
604 if (fd >= 0 || errno != EBUSY) {
605 /* OK, we are safe to leave */
606 if (sigterm && !dirty_arrays)
607 dprintf("caught sigterm, all clean... exiting\n");
609 dprintf("no arrays to monitor... exiting\n");
611 /* On SIGTERM, someone (the take-over mdmon) will
614 remove_pidfile(container->devnm);
627 if (*aap == NULL || container->retry_soon) {
628 /* just waiting to get O_EXCL access */
630 ts.tv_nsec = 20000000ULL;
632 sigprocmask(SIG_UNBLOCK, NULL, &set);
633 sigdelset(&set, SIGUSR1);
634 monitor_loop_cnt |= 1;
635 rv = pselect(maxfd+1, NULL, NULL, &rfds, &ts, &set);
636 monitor_loop_cnt += 1;
638 if (errno == EINTR) {
640 dprintf("monitor: caught signal\n");
642 dprintf("monitor: error %d in pselect\n",
647 dprint_wake_reasons(&rfds);
649 container->retry_soon = 0;
653 struct metadata_update *this;
655 for (this = update_queue; this ; this = this->next)
656 container->ss->process_update(container, this);
658 update_queue_handled = update_queue;
661 container->ss->sync_metadata(container);
666 for (a = *aap; a ; a = a->next) {
668 if (a->replaces && !discard_this) {
669 struct active_array **ap;
670 for (ap = &a->next; *ap && *ap != a->replaces;
675 discard_this = a->replaces;
677 /* FIXME check if device->state_fd need to be cleared?*/
680 if (a->container && !a->to_remove) {
681 int ret = read_and_act(a);
683 dirty_arrays += !!(ret & ARRAY_DIRTY);
684 /* when terminating stop manipulating the array after it
685 * is clean, but make sure read_and_act() is given a
686 * chance to handle 'active_idle'
688 if (sigterm && !(ret & ARRAY_DIRTY))
689 a->container = NULL; /* stop touching this array */
690 if (ret & ARRAY_BUSY)
691 container->retry_soon = 1;
695 /* propagate failures across container members */
696 for (a = *aap; a ; a = a->next) {
697 if (!a->container || a->to_remove)
699 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
700 if (mdi->curr_state & DS_FAULTY)
701 reconcile_failed(*aap, mdi);
707 void do_monitor(struct supertype *container)
712 rv = wait_and_act(container, first);