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)
48 static int read_attr(char *buf, int len, int fd)
57 n = read(fd, buf, len - 1);
69 static unsigned long long read_resync_start(int fd)
74 n = read_attr(buf, 30, fd);
77 if (strncmp(buf, "none", 4) == 0)
80 return strtoull(buf, NULL, 10);
83 static unsigned long long read_sync_completed(int fd)
85 unsigned long long val;
90 n = read_attr(buf, 50, fd);
95 val = strtoull(buf, &ep, 0);
96 if (ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))
101 static enum array_state read_state(int fd)
104 int n = read_attr(buf, 20, fd);
108 return (enum array_state) sysfs_match_word(buf, array_states);
111 static enum sync_action read_action( int fd)
114 int n = read_attr(buf, 20, fd);
118 return (enum sync_action) sysfs_match_word(buf, sync_actions);
121 int read_dev_state(int fd)
124 int n = read_attr(buf, 60, fd);
133 if (sysfs_attr_match(cp, "faulty"))
135 if (sysfs_attr_match(cp, "in_sync"))
137 if (sysfs_attr_match(cp, "write_mostly"))
138 rv |= DS_WRITE_MOSTLY;
139 if (sysfs_attr_match(cp, "spare"))
141 if (sysfs_attr_match(cp, "blocked"))
143 cp = strchr(cp, ',');
150 static void signal_manager(void)
152 /* tgkill(getpid(), mon_tid, SIGUSR1); */
154 syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
157 /* Monitor a set of active md arrays - all of which share the
158 * same metadata - and respond to events that require
161 * New arrays are detected by another thread which allocates
162 * required memory and attaches the data structure to our list.
166 * This is detected by array_state going to 'clear' or 'inactive'.
167 * while we thought it was active.
168 * Response is to mark metadata as clean and 'clear' the array(??)
170 * array_state if 'write-pending'
171 * We mark metadata as 'dirty' then set array to 'active'.
173 * Either ignore, or mark clean, then mark metadata as clean.
176 * detected by rd-N/state reporting "faulty"
177 * mark device as 'failed' in metadata, let the kernel release the
178 * device by writing '-blocked' to rd/state, and finally write 'remove' to
179 * rd/state. Before a disk can be replaced it must be failed and removed
180 * from all container members, this will be preemptive for the other
184 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
186 * Notify metadata that sync is complete.
189 * sync_action changes from 'recover' to 'idle'
190 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
193 * This only happens on finding a new array... mdadm will have set
194 * 'resync_start' to the correct value. If 'resync_start' indicates that an
195 * resync needs to occur set the array to the 'active' state rather than the
196 * initial read-auto state.
200 * We wait for a change (poll/select) on array_state, sync_action, and
201 * each rd-X/state file.
202 * When we get any change, we check everything. So read each state file,
203 * then decide what to do.
205 * The core action is to write new metadata to all devices in the array.
206 * This is done at most once on any wakeup.
207 * After that we might:
208 * - update the array_state
209 * - set the role of some devices.
210 * - request a sync_action
214 #define ARRAY_DIRTY 1
216 static int read_and_act(struct active_array *a)
218 unsigned long long sync_completed;
219 int check_degraded = 0;
220 int check_reshape = 0;
226 a->next_state = bad_word;
227 a->next_action = bad_action;
229 a->curr_state = read_state(a->info.state_fd);
230 a->curr_action = read_action(a->action_fd);
231 a->info.resync_start = read_resync_start(a->resync_start_fd);
232 sync_completed = read_sync_completed(a->sync_completed_fd);
233 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
236 if (mdi->state_fd >= 0) {
237 mdi->recovery_start = read_resync_start(mdi->recovery_fd);
238 mdi->curr_state = read_dev_state(mdi->state_fd);
242 if (a->curr_state > inactive &&
243 a->prev_state == inactive) {
244 /* array has been started
245 * possible that container operation has to be completed
247 a->container->ss->set_array_state(a, 0);
249 if (a->curr_state <= inactive &&
250 a->prev_state > inactive) {
251 /* array has been stopped */
252 a->container->ss->set_array_state(a, 1);
253 a->next_state = clear;
256 if (a->curr_state == write_pending) {
257 a->container->ss->set_array_state(a, 0);
258 a->next_state = active;
261 if (a->curr_state == active_idle) {
262 /* Set array to 'clean' FIRST, then mark clean
265 a->next_state = clean;
268 if (a->curr_state == clean) {
269 a->container->ss->set_array_state(a, 1);
271 if (a->curr_state == active ||
272 a->curr_state == suspended ||
273 a->curr_state == bad_word)
275 if (a->curr_state == readonly) {
276 /* Well, I'm ready to handle things. If readonly
277 * wasn't requested, transition to read-auto.
280 read_attr(buf, sizeof(buf), a->metadata_fd);
281 if (strncmp(buf, "external:-", 10) == 0) {
282 /* explicit request for readonly array. Leave it alone */
285 if (a->container->ss->set_array_state(a, 2))
286 a->next_state = read_auto; /* array is clean */
288 a->next_state = active; /* Now active for recovery etc */
295 a->curr_action == idle &&
296 a->prev_action == resync) {
297 /* A resync has finished. The endpoint is recorded in
298 * 'sync_start'. We don't update the metadata
299 * until the array goes inactive or readonly though.
300 * Just check if we need to fiddle spares.
302 a->container->ss->set_array_state(a, a->curr_state <= clean);
307 a->curr_action == idle &&
308 a->prev_action == recover) {
309 /* A recovery has finished. Some disks may be in sync now,
310 * and the array may no longer be degraded
312 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
313 a->container->ss->set_disk(a, mdi->disk.raid_disk,
315 if (! (mdi->curr_state & DS_INSYNC))
319 if (count != a->info.array.raid_disks)
324 a->curr_action == reshape &&
325 a->prev_action != reshape)
326 /* reshape was requested by mdadm. Need to see if
327 * new devices have been added. Manager does that
328 * when it sees check_reshape
332 /* Check for failures and if found:
333 * 1/ Record the failure in the metadata and unblock the device.
334 * FIXME update the kernel to stop notifying on failed drives when
335 * the array is readonly and we have cleared 'blocked'
336 * 2/ Try to remove the device if the array is writable, or can be
339 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
340 if (mdi->curr_state & DS_FAULTY) {
341 a->container->ss->set_disk(a, mdi->disk.raid_disk,
344 if (mdi->curr_state & DS_BLOCKED)
345 mdi->next_state |= DS_UNBLOCK;
346 if (a->curr_state == read_auto) {
347 a->container->ss->set_array_state(a, 0);
348 a->next_state = active;
350 if (a->curr_state > readonly)
351 mdi->next_state |= DS_REMOVE;
355 /* Check for recovery checkpoint notifications. We need to be a
356 * minimum distance away from the last checkpoint to prevent
357 * over checkpointing. Note reshape checkpointing is handled
358 * in the second branch.
360 if (sync_completed > a->last_checkpoint &&
361 sync_completed - a->last_checkpoint > a->info.component_size >> 4 &&
362 a->curr_action > reshape) {
363 /* A (non-reshape) sync_action has reached a checkpoint.
364 * Record the updated position in the metadata
366 a->last_checkpoint = sync_completed;
367 a->container->ss->set_array_state(a, a->curr_state <= clean);
368 } else if ((a->curr_action == idle && a->prev_action == reshape) ||
369 (a->curr_action == reshape
370 && sync_completed > a->last_checkpoint) ) {
371 /* Reshape has progressed or completed so we need to
372 * update the array state - and possibly the array size
374 if (sync_completed != 0)
375 a->last_checkpoint = sync_completed;
376 /* We might need to update last_checkpoint depending on
377 * the reason that reshape finished.
378 * if array reshape is really finished:
379 * set check point to the end, this allows
380 * set_array_state() to finalize reshape in metadata
381 * if reshape if broken: do not set checkpoint to the end
382 * this allows for reshape restart from checkpoint
384 if ((a->curr_action != reshape) &&
385 (a->prev_action == reshape)) {
387 if ((sysfs_get_str(&a->info, NULL,
390 sizeof(buf)) >= 0) &&
391 strncmp(buf, "none", 4) == 0)
392 a->last_checkpoint = a->info.component_size;
394 a->container->ss->set_array_state(a, a->curr_state <= clean);
395 a->last_checkpoint = sync_completed;
398 if (sync_completed > a->last_checkpoint)
399 a->last_checkpoint = sync_completed;
401 a->container->ss->sync_metadata(a->container);
402 dprintf("%s(%d): state:%s action:%s next(", __func__, a->info.container_member,
403 array_states[a->curr_state], sync_actions[a->curr_action]);
405 /* Effect state changes in the array */
406 if (a->next_state != bad_word) {
407 dprintf(" state:%s", array_states[a->next_state]);
408 write_attr(array_states[a->next_state], a->info.state_fd);
410 if (a->next_action != bad_action) {
411 write_attr(sync_actions[a->next_action], a->action_fd);
412 dprintf(" action:%s", sync_actions[a->next_action]);
414 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
415 if (mdi->next_state & DS_UNBLOCK) {
416 dprintf(" %d:-blocked", mdi->disk.raid_disk);
417 write_attr("-blocked", mdi->state_fd);
420 if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
423 /* The kernel may not be able to immediately remove the
424 * disk. In that case we wait a little while and
427 remove_result = write_attr("remove", mdi->state_fd);
428 if (remove_result > 0) {
429 dprintf(" %d:removed", mdi->disk.raid_disk);
430 close(mdi->state_fd);
431 close(mdi->recovery_fd);
436 if (mdi->next_state & DS_INSYNC) {
437 write_attr("+in_sync", mdi->state_fd);
438 dprintf(" %d:+in_sync", mdi->disk.raid_disk);
443 /* move curr_ to prev_ */
444 a->prev_state = a->curr_state;
446 a->prev_action = a->curr_action;
448 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
449 mdi->prev_state = mdi->curr_state;
453 if (check_degraded || check_reshape) {
454 /* manager will do the actual check */
456 a->check_degraded = 1;
458 a->check_reshape = 1;
468 static struct mdinfo *
469 find_device(struct active_array *a, int major, int minor)
473 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
474 if (mdi->disk.major == major && mdi->disk.minor == minor)
480 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
482 struct active_array *a;
483 struct mdinfo *victim;
485 for (a = aa; a; a = a->next) {
486 if (!a->container || a->to_remove)
488 victim = find_device(a, failed->disk.major, failed->disk.minor);
492 if (!(victim->curr_state & DS_FAULTY))
493 write_attr("faulty", victim->state_fd);
498 static void dprint_wake_reasons(fd_set *fds)
506 fprintf(stderr, "monitor: wake ( ");
507 for (i = 0; i < FD_SETSIZE; i++) {
508 if (FD_ISSET(i, fds)) {
509 sprintf(proc_path, "/proc/%d/fd/%d",
512 rv = readlink(proc_path, link, sizeof(link) - 1);
514 fprintf(stderr, "%d:unknown ", i);
518 basename = strrchr(link, '/');
519 fprintf(stderr, "%d:%s ",
520 i, basename ? ++basename : link);
523 fprintf(stderr, ")\n");
527 int monitor_loop_cnt;
529 static int wait_and_act(struct supertype *container, int nowait)
533 struct active_array **aap = &container->arrays;
534 struct active_array *a, **ap;
537 static unsigned int dirty_arrays = ~0; /* start at some non-zero value */
541 for (ap = aap ; *ap ;) {
543 /* once an array has been deactivated we want to
544 * ask the manager to discard it.
546 if (!a->container || a->to_remove) {
558 add_fd(&rfds, &maxfd, a->info.state_fd);
559 add_fd(&rfds, &maxfd, a->action_fd);
560 add_fd(&rfds, &maxfd, a->sync_completed_fd);
561 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
562 add_fd(&rfds, &maxfd, mdi->state_fd);
567 if (manager_ready && (*aap == NULL || (sigterm && !dirty_arrays))) {
568 /* No interesting arrays, or we have been told to
569 * terminate and everything is clean. Lets see about
570 * exiting. Note that blocking at this point is not a
571 * problem as there are no active arrays, there is
572 * nothing that we need to be ready to do.
576 fd = open_dev_excl(container->devnm);
578 fd = open_dev_flags(container->devnm, O_RDONLY|O_EXCL);
579 if (fd >= 0 || errno != EBUSY) {
580 /* OK, we are safe to leave */
581 if (sigterm && !dirty_arrays)
582 dprintf("caught sigterm, all clean... exiting\n");
584 dprintf("no arrays to monitor... exiting\n");
586 /* On SIGTERM, someone (the take-over mdmon) will
589 remove_pidfile(container->devnm);
602 if (*aap == NULL || container->retry_soon) {
603 /* just waiting to get O_EXCL access */
605 ts.tv_nsec = 20000000ULL;
607 sigprocmask(SIG_UNBLOCK, NULL, &set);
608 sigdelset(&set, SIGUSR1);
609 monitor_loop_cnt |= 1;
610 rv = pselect(maxfd+1, NULL, NULL, &rfds, &ts, &set);
611 monitor_loop_cnt += 1;
612 if (rv == -1 && errno == EINTR)
615 dprint_wake_reasons(&rfds);
617 container->retry_soon = 0;
621 struct metadata_update *this;
623 for (this = update_queue; this ; this = this->next)
624 container->ss->process_update(container, this);
626 update_queue_handled = update_queue;
629 container->ss->sync_metadata(container);
634 for (a = *aap; a ; a = a->next) {
636 if (a->replaces && !discard_this) {
637 struct active_array **ap;
638 for (ap = &a->next; *ap && *ap != a->replaces;
643 discard_this = a->replaces;
645 /* FIXME check if device->state_fd need to be cleared?*/
648 if (a->container && !a->to_remove) {
649 int ret = read_and_act(a);
651 dirty_arrays += !!(ret & ARRAY_DIRTY);
652 /* when terminating stop manipulating the array after it
653 * is clean, but make sure read_and_act() is given a
654 * chance to handle 'active_idle'
656 if (sigterm && !(ret & ARRAY_DIRTY))
657 a->container = NULL; /* stop touching this array */
658 if (ret & ARRAY_BUSY)
659 container->retry_soon = 1;
663 /* propagate failures across container members */
664 for (a = *aap; a ; a = a->next) {
665 if (!a->container || a->to_remove)
667 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
668 if (mdi->curr_state & DS_FAULTY)
669 reconcile_failed(*aap, mdi);
675 void do_monitor(struct supertype *container)
680 rv = wait_and_act(container, first);