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
39 static int write_attr(char *attr, int fd)
41 return write(fd, attr, strlen(attr));
44 static void add_fd(fd_set *fds, int *maxfd, int fd)
49 if (fstat(fd, &st) == -1) {
50 dprintf("Invalid fd %d\n", fd);
53 if (st.st_nlink == 0) {
54 dprintf("fd %d was deleted\n", fd);
62 static int read_attr(char *buf, int len, int fd)
71 n = read(fd, buf, len - 1);
83 static void read_resync_start(int fd, unsigned long long *v)
88 n = read_attr(buf, 30, fd);
90 dprintf("Failed to read resync_start (%d)\n", fd);
93 if (strncmp(buf, "none", 4) == 0)
96 *v = strtoull(buf, NULL, 10);
99 static unsigned long long read_sync_completed(int fd)
101 unsigned long long val;
106 n = read_attr(buf, 50, fd);
111 val = strtoull(buf, &ep, 0);
112 if (ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))
117 static enum array_state read_state(int fd)
120 int n = read_attr(buf, 20, fd);
124 return (enum array_state) sysfs_match_word(buf, array_states);
127 static enum sync_action read_action( int fd)
130 int n = read_attr(buf, 20, fd);
134 return (enum sync_action) sysfs_match_word(buf, sync_actions);
137 int read_dev_state(int fd)
140 int n = read_attr(buf, sizeof(buf), fd);
149 if (sysfs_attr_match(cp, "faulty"))
151 if (sysfs_attr_match(cp, "in_sync"))
153 if (sysfs_attr_match(cp, "write_mostly"))
154 rv |= DS_WRITE_MOSTLY;
155 if (sysfs_attr_match(cp, "spare"))
157 if (sysfs_attr_match(cp, "blocked"))
159 cp = strchr(cp, ',');
166 int process_ubb(struct active_array *a, struct mdinfo *mdi, const unsigned long
167 long sector, const int length, const char *buf,
170 struct superswitch *ss = a->container->ss;
173 * record bad block in metadata first, then acknowledge it to the driver
176 if ((ss->record_bad_block(a, mdi->disk.raid_disk, sector, length)) &&
177 (write(mdi->bb_fd, buf, buf_len) == buf_len))
181 * failed to store or acknowledge bad block, switch of bad block support
182 * to get it out of blocked state
184 sysfs_set_str(&a->info, mdi, "state", "-external_bbl");
188 int compare_bb(struct active_array *a, struct mdinfo *mdi, const unsigned long
189 long sector, const unsigned int length, void *arg)
191 struct superswitch *ss = a->container->ss;
192 struct md_bb *bb = (struct md_bb *) arg;
196 for (i = 0; i < bb->count; i++) {
197 unsigned long long start = bb->entries[i].sector;
198 unsigned long long len = bb->entries[i].length;
201 * bad block in metadata exactly matches bad block in kernel
202 * list, just remove it from a list
204 if ((start == sector) && (len == length)) {
205 if (i < bb->count - 1)
206 bb->entries[i] = bb->entries[bb->count - 1];
212 * bad block in metadata spans bad block in kernel list,
213 * clear it and record new bad block
215 if ((sector >= start) && (sector + length <= start + len)) {
216 ss->clear_bad_block(a, mdi->disk.raid_disk, start, len);
221 /* record all bad blocks not in metadata list */
222 if (record && (ss->record_bad_block(a, mdi->disk.raid_disk, sector,
224 sysfs_set_str(&a->info, mdi, "state", "-external_bbl");
231 static int read_bb_file(int fd, struct active_array *a, struct mdinfo *mdi,
232 enum bb_action action, void *arg)
240 int preserve_pos = (action == RECORD_BB ? 0 : 1);
242 if (lseek(fd, 0, SEEK_SET) == (off_t) -1)
247 n = read(fd, buf + pos, sizeof(buf) - 1 - pos);
256 unsigned long long sector;
263 /* kernel sysfs file format: "sector length\n" */
264 matched = sscanf(buf + off, "%llu %d%c%n", §or,
265 &length, &newline, &consumed);
266 if ((matched != 3) && (off > 0)) {
267 /* truncated entry, read again */
269 pos = sizeof(buf) - off - 1;
270 memmove(buf, buf + off, pos);
272 if (lseek(fd, 0, SEEK_SET) ==
286 if (action == RECORD_BB)
287 rc = process_ubb(a, mdi, sector, length,
288 buf + off, consumed);
289 else if (action == COMPARE_BB)
290 rc = compare_bb(a, mdi, sector, length, arg);
299 } while (read_again);
304 static int process_dev_ubb(struct active_array *a, struct mdinfo *mdi)
306 return read_bb_file(mdi->ubb_fd, a, mdi, RECORD_BB, NULL);
309 static int check_for_cleared_bb(struct active_array *a, struct mdinfo *mdi)
311 struct superswitch *ss = a->container->ss;
316 * Get a list of bad blocks for an array, then read list of
317 * acknowledged bad blocks from kernel and compare it against metadata
318 * list, clear all bad blocks remaining in metadata list
320 bb = ss->get_bad_blocks(a, mdi->disk.raid_disk);
324 if (read_bb_file(mdi->bb_fd, a, mdi, COMPARE_BB, bb) < 0)
327 for (i = 0; i < bb->count; i++) {
328 unsigned long long sector = bb->entries[i].sector;
329 int length = bb->entries[i].length;
331 ss->clear_bad_block(a, mdi->disk.raid_disk, sector, length);
337 static void signal_manager(void)
339 /* tgkill(getpid(), mon_tid, SIGUSR1); */
341 syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
344 /* Monitor a set of active md arrays - all of which share the
345 * same metadata - and respond to events that require
348 * New arrays are detected by another thread which allocates
349 * required memory and attaches the data structure to our list.
353 * This is detected by array_state going to 'clear' or 'inactive'.
354 * while we thought it was active.
355 * Response is to mark metadata as clean and 'clear' the array(??)
357 * array_state if 'write-pending'
358 * We mark metadata as 'dirty' then set array to 'active'.
360 * Either ignore, or mark clean, then mark metadata as clean.
363 * detected by rd-N/state reporting "faulty"
364 * mark device as 'failed' in metadata, let the kernel release the
365 * device by writing '-blocked' to rd/state, and finally write 'remove' to
366 * rd/state. Before a disk can be replaced it must be failed and removed
367 * from all container members, this will be preemptive for the other
371 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
373 * Notify metadata that sync is complete.
376 * sync_action changes from 'recover' to 'idle'
377 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
380 * This only happens on finding a new array... mdadm will have set
381 * 'resync_start' to the correct value. If 'resync_start' indicates that an
382 * resync needs to occur set the array to the 'active' state rather than the
383 * initial read-auto state.
387 * We wait for a change (poll/select) on array_state, sync_action, and
388 * each rd-X/state file.
389 * When we get any change, we check everything. So read each state file,
390 * then decide what to do.
392 * The core action is to write new metadata to all devices in the array.
393 * This is done at most once on any wakeup.
394 * After that we might:
395 * - update the array_state
396 * - set the role of some devices.
397 * - request a sync_action
401 #define ARRAY_DIRTY 1
403 static int read_and_act(struct active_array *a, fd_set *fds)
405 unsigned long long sync_completed;
406 int check_degraded = 0;
407 int check_reshape = 0;
414 a->next_state = bad_word;
415 a->next_action = bad_action;
417 a->curr_state = read_state(a->info.state_fd);
418 a->curr_action = read_action(a->action_fd);
419 if (a->curr_state != clear)
421 * In "clear" state, resync_start may wrongly be set to "0"
422 * when the kernel called md_clean but didn't remove the
423 * sysfs attributes yet
425 read_resync_start(a->resync_start_fd, &a->info.resync_start);
426 sync_completed = read_sync_completed(a->sync_completed_fd);
427 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
430 if (mdi->state_fd >= 0) {
431 read_resync_start(mdi->recovery_fd,
432 &mdi->recovery_start);
433 mdi->curr_state = read_dev_state(mdi->state_fd);
436 * If array is blocked and metadata handler is able to handle
437 * BB, check if you can acknowledge them to md driver. If
438 * successful, clear faulty state and unblock the array.
440 if ((mdi->curr_state & DS_BLOCKED) &&
441 a->container->ss->record_bad_block &&
442 (process_dev_ubb(a, mdi) > 0)) {
443 mdi->next_state |= DS_UNBLOCK;
445 if (FD_ISSET(mdi->bb_fd, fds))
446 check_for_cleared_bb(a, mdi);
449 gettimeofday(&tv, NULL);
450 dprintf("(%d): %ld.%06ld state:%s prev:%s action:%s prev: %s start:%llu\n",
451 a->info.container_member,
452 tv.tv_sec, tv.tv_usec,
453 array_states[a->curr_state],
454 array_states[a->prev_state],
455 sync_actions[a->curr_action],
456 sync_actions[a->prev_action],
460 if ((a->curr_state == bad_word || a->curr_state <= inactive) &&
461 a->prev_state > inactive) {
462 /* array has been stopped */
463 a->container->ss->set_array_state(a, 1);
464 a->next_state = clear;
467 if (a->curr_state == write_pending) {
468 a->container->ss->set_array_state(a, 0);
469 a->next_state = active;
472 if (a->curr_state == active_idle) {
473 /* Set array to 'clean' FIRST, then mark clean
476 a->next_state = clean;
479 if (a->curr_state == clean) {
480 a->container->ss->set_array_state(a, 1);
482 if (a->curr_state == active ||
483 a->curr_state == suspended)
485 if (a->curr_state == readonly) {
486 /* Well, I'm ready to handle things. If readonly
487 * wasn't requested, transition to read-auto.
490 read_attr(buf, sizeof(buf), a->metadata_fd);
491 if (strncmp(buf, "external:-", 10) == 0) {
492 /* explicit request for readonly array. Leave it alone */
495 if (a->container->ss->set_array_state(a, 2))
496 a->next_state = read_auto; /* array is clean */
498 a->next_state = active; /* Now active for recovery etc */
505 a->curr_action == idle &&
506 a->prev_action == resync) {
507 /* A resync has finished. The endpoint is recorded in
508 * 'sync_start'. We don't update the metadata
509 * until the array goes inactive or readonly though.
510 * Just check if we need to fiddle spares.
512 a->container->ss->set_array_state(a, a->curr_state <= clean);
517 a->curr_action == idle &&
518 a->prev_action == recover) {
519 /* A recovery has finished. Some disks may be in sync now,
520 * and the array may no longer be degraded
522 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
523 a->container->ss->set_disk(a, mdi->disk.raid_disk,
525 if (! (mdi->curr_state & DS_INSYNC))
529 if (count != a->info.array.raid_disks)
534 a->curr_action == reshape &&
535 a->prev_action != reshape)
536 /* reshape was requested by mdadm. Need to see if
537 * new devices have been added. Manager does that
538 * when it sees check_reshape
542 /* Check for failures and if found:
543 * 1/ Record the failure in the metadata and unblock the device.
544 * FIXME update the kernel to stop notifying on failed drives when
545 * the array is readonly and we have cleared 'blocked'
546 * 2/ Try to remove the device if the array is writable, or can be
549 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
550 if (mdi->curr_state & DS_FAULTY) {
551 a->container->ss->set_disk(a, mdi->disk.raid_disk,
554 if (mdi->curr_state & DS_BLOCKED)
555 mdi->next_state |= DS_UNBLOCK;
556 if (a->curr_state == read_auto) {
557 a->container->ss->set_array_state(a, 0);
558 a->next_state = active;
560 if (a->curr_state > readonly)
561 mdi->next_state |= DS_REMOVE;
565 /* Check for recovery checkpoint notifications. We need to be a
566 * minimum distance away from the last checkpoint to prevent
567 * over checkpointing. Note reshape checkpointing is handled
568 * in the second branch.
570 if (sync_completed > a->last_checkpoint &&
571 sync_completed - a->last_checkpoint > a->info.component_size >> 4 &&
572 a->curr_action > reshape) {
573 /* A (non-reshape) sync_action has reached a checkpoint.
574 * Record the updated position in the metadata
576 a->last_checkpoint = sync_completed;
577 a->container->ss->set_array_state(a, a->curr_state <= clean);
578 } else if ((a->curr_action == idle && a->prev_action == reshape) ||
579 (a->curr_action == reshape
580 && sync_completed > a->last_checkpoint) ) {
581 /* Reshape has progressed or completed so we need to
582 * update the array state - and possibly the array size
584 if (sync_completed != 0)
585 a->last_checkpoint = sync_completed;
586 /* We might need to update last_checkpoint depending on
587 * the reason that reshape finished.
588 * if array reshape is really finished:
589 * set check point to the end, this allows
590 * set_array_state() to finalize reshape in metadata
591 * if reshape if broken: do not set checkpoint to the end
592 * this allows for reshape restart from checkpoint
594 if ((a->curr_action != reshape) &&
595 (a->prev_action == reshape)) {
597 if ((sysfs_get_str(&a->info, NULL,
600 sizeof(buf)) >= 0) &&
601 strncmp(buf, "none", 4) == 0)
602 a->last_checkpoint = a->info.component_size;
604 a->container->ss->set_array_state(a, a->curr_state <= clean);
605 a->last_checkpoint = sync_completed;
608 if (sync_completed > a->last_checkpoint)
609 a->last_checkpoint = sync_completed;
611 if (sync_completed >= a->info.component_size)
612 a->last_checkpoint = 0;
614 a->container->ss->sync_metadata(a->container);
615 dprintf("(%d): state:%s action:%s next(", a->info.container_member,
616 array_states[a->curr_state], sync_actions[a->curr_action]);
618 /* Effect state changes in the array */
619 if (a->next_state != bad_word) {
620 dprintf_cont(" state:%s", array_states[a->next_state]);
621 write_attr(array_states[a->next_state], a->info.state_fd);
623 if (a->next_action != bad_action) {
624 write_attr(sync_actions[a->next_action], a->action_fd);
625 dprintf_cont(" action:%s", sync_actions[a->next_action]);
627 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
628 if (mdi->next_state & DS_UNBLOCK) {
629 dprintf_cont(" %d:-blocked", mdi->disk.raid_disk);
630 write_attr("-blocked", mdi->state_fd);
633 if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
636 /* The kernel may not be able to immediately remove the
637 * disk. In that case we wait a little while and
640 remove_result = write_attr("remove", mdi->state_fd);
641 if (remove_result > 0) {
642 dprintf_cont(" %d:removed", mdi->disk.raid_disk);
643 close(mdi->state_fd);
644 close(mdi->recovery_fd);
651 if (mdi->next_state & DS_INSYNC) {
652 write_attr("+in_sync", mdi->state_fd);
653 dprintf_cont(" %d:+in_sync", mdi->disk.raid_disk);
656 dprintf_cont(" )\n");
658 /* move curr_ to prev_ */
659 a->prev_state = a->curr_state;
661 a->prev_action = a->curr_action;
663 for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
664 mdi->prev_state = mdi->curr_state;
668 if (check_degraded || check_reshape) {
669 /* manager will do the actual check */
671 a->check_degraded = 1;
673 a->check_reshape = 1;
683 static struct mdinfo *
684 find_device(struct active_array *a, int major, int minor)
688 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
689 if (mdi->disk.major == major && mdi->disk.minor == minor)
695 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
697 struct active_array *a;
698 struct mdinfo *victim;
700 for (a = aa; a; a = a->next) {
701 if (!a->container || a->to_remove)
703 victim = find_device(a, failed->disk.major, failed->disk.minor);
707 if (!(victim->curr_state & DS_FAULTY))
708 write_attr("faulty", victim->state_fd);
713 static void dprint_wake_reasons(fd_set *fds)
721 fprintf(stderr, "monitor: wake ( ");
722 for (i = 0; i < FD_SETSIZE; i++) {
723 if (FD_ISSET(i, fds)) {
724 sprintf(proc_path, "/proc/%d/fd/%d",
727 rv = readlink(proc_path, link, sizeof(link) - 1);
729 fprintf(stderr, "%d:unknown ", i);
733 basename = strrchr(link, '/');
734 fprintf(stderr, "%d:%s ",
735 i, basename ? ++basename : link);
738 fprintf(stderr, ")\n");
742 int monitor_loop_cnt;
744 static int wait_and_act(struct supertype *container, int nowait)
748 struct active_array **aap = &container->arrays;
749 struct active_array *a, **ap;
752 static unsigned int dirty_arrays = ~0; /* start at some non-zero value */
756 for (ap = aap ; *ap ;) {
758 /* once an array has been deactivated we want to
759 * ask the manager to discard it.
761 if (!a->container || a->to_remove) {
773 add_fd(&rfds, &maxfd, a->info.state_fd);
774 add_fd(&rfds, &maxfd, a->action_fd);
775 add_fd(&rfds, &maxfd, a->sync_completed_fd);
776 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
777 add_fd(&rfds, &maxfd, mdi->state_fd);
778 add_fd(&rfds, &maxfd, mdi->bb_fd);
779 add_fd(&rfds, &maxfd, mdi->ubb_fd);
785 if (manager_ready && (*aap == NULL || (sigterm && !dirty_arrays))) {
786 /* No interesting arrays, or we have been told to
787 * terminate and everything is clean. Lets see about
788 * exiting. Note that blocking at this point is not a
789 * problem as there are no active arrays, there is
790 * nothing that we need to be ready to do.
794 fd = open_dev_excl(container->devnm);
796 fd = open_dev_flags(container->devnm, O_RDONLY|O_EXCL);
797 if (fd >= 0 || errno != EBUSY) {
798 /* OK, we are safe to leave */
799 if (sigterm && !dirty_arrays)
800 dprintf("caught sigterm, all clean... exiting\n");
802 dprintf("no arrays to monitor... exiting\n");
804 /* On SIGTERM, someone (the take-over mdmon) will
807 remove_pidfile(container->devnm);
820 if (*aap == NULL || container->retry_soon) {
821 /* just waiting to get O_EXCL access */
823 ts.tv_nsec = 20000000ULL;
825 sigprocmask(SIG_UNBLOCK, NULL, &set);
826 sigdelset(&set, SIGUSR1);
827 monitor_loop_cnt |= 1;
828 rv = pselect(maxfd+1, NULL, NULL, &rfds, &ts, &set);
829 monitor_loop_cnt += 1;
831 if (errno == EINTR) {
834 dprintf("monitor: caught signal\n");
836 dprintf("monitor: error %d in pselect\n",
841 dprint_wake_reasons(&rfds);
843 container->retry_soon = 0;
847 struct metadata_update *this;
849 for (this = update_queue; this ; this = this->next)
850 container->ss->process_update(container, this);
852 update_queue_handled = update_queue;
855 container->ss->sync_metadata(container);
860 for (a = *aap; a ; a = a->next) {
862 if (a->replaces && !discard_this) {
863 struct active_array **ap;
864 for (ap = &a->next; *ap && *ap != a->replaces;
869 discard_this = a->replaces;
871 /* FIXME check if device->state_fd need to be cleared?*/
874 if (a->container && !a->to_remove) {
875 int ret = read_and_act(a, &rfds);
877 dirty_arrays += !!(ret & ARRAY_DIRTY);
878 /* when terminating stop manipulating the array after it
879 * is clean, but make sure read_and_act() is given a
880 * chance to handle 'active_idle'
882 if (sigterm && !(ret & ARRAY_DIRTY))
883 a->container = NULL; /* stop touching this array */
884 if (ret & ARRAY_BUSY)
885 container->retry_soon = 1;
889 /* propagate failures across container members */
890 for (a = *aap; a ; a = a->next) {
891 if (!a->container || a->to_remove)
893 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
894 if (mdi->curr_state & DS_FAULTY)
895 reconcile_failed(*aap, mdi);
901 void do_monitor(struct supertype *container)
906 rv = wait_and_act(container, first);