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.
22 * The management thread for monitoring active md arrays.
23 * This thread does things which might block such as memory
27 * - Find out about new arrays in this container.
28 * Allocate the data structures and open the files.
30 * For this we watch /proc/mdstat and find new arrays with
31 * metadata type that confirms sharing. e.g. "md4"
32 * When we find a new array we slip it into the list of
33 * arrays and signal 'monitor' by writing to a pipe.
35 * - Respond to reshape requests by allocating new data structures
36 * and opening new files.
38 * These come as a change to raid_disks. We allocate a new
39 * version of the data structures and slip it into the list.
40 * 'monitor' will notice and release the old version.
41 * Changes to level, chunksize, layout.. do not need re-allocation.
42 * Reductions in raid_disks don't really either, but we handle
43 * them the same way for consistency.
45 * - When a device is added to the container, we add it to the metadata
48 * - Deal with degraded array
49 * We only do this when first noticing the array is degraded.
50 * This can be when we first see the array, when sync completes or
51 * when recovery completes.
53 * Check if number of failed devices suggests recovery is needed, and
55 * Ask metadata to allocate a spare device
56 * Add device as not in_sync and give a role
58 * Open sysfs files and pass to monitor.
59 * Make sure that monitor Starts recovery....
61 * - Pass on metadata updates from external programs such as
62 * mdadm creating a new array.
65 * It might involve adding a new array or changing the status of
66 * a spare, or any reconfig that the kernel doesn't get involved in.
68 * The required updates are received via a named pipe. There will
69 * be one named pipe for each container. Each message contains a
70 * sync marker: 0x5a5aa5a5, A byte count, and the message. This is
71 * passed to the metadata handler which will interpret and process it.
72 * For 'DDF' messages are internal data blocks with the leading
73 * 'magic number' signifying what sort of data it is.
78 * We select on /proc/mdstat and the named pipe.
79 * We create new arrays or updated version of arrays and slip
80 * them into the head of the list, then signal 'monitor' via a pipe write.
81 * 'monitor' will notice and place the old array on a return list.
82 * Metadata updates are placed on a queue just like they arrive
83 * from the named pipe.
85 * When new arrays are found based on correct metadata string, we
86 * need to identify them with an entry in the metadata. Maybe we require
87 * the metadata to be mdX/NN when NN is the index into an appropriate table.
93 * - Watch for spares to be added to the container, and write updated
95 * - Watch for new arrays using this container, confirm they match metadata
96 * and if so, start monitoring them
97 * - Watch for spares being added to monitored arrays. This shouldn't
98 * happen, as we should do all the adding. Just remove them.
99 * - Watch for change in raid-disks, chunk-size, etc. Update metadata and
107 #include <sys/syscall.h>
108 #include <sys/socket.h>
111 static void close_aa(struct active_array *aa)
115 for (d = aa->info.devs; d; d = d->next) {
116 close(d->recovery_fd);
122 if (aa->action_fd >= 0)
123 close(aa->action_fd);
124 if (aa->info.state_fd >= 0)
125 close(aa->info.state_fd);
126 if (aa->resync_start_fd >= 0)
127 close(aa->resync_start_fd);
128 if (aa->metadata_fd >= 0)
129 close(aa->metadata_fd);
130 if (aa->sync_completed_fd >= 0)
131 close(aa->sync_completed_fd);
134 static void free_aa(struct active_array *aa)
136 /* Note that this doesn't close fds if they are being used
137 * by a clone. ->container will be set for a clone
139 dprintf("sys_name: %s\n", aa->info.sys_name);
142 while (aa->info.devs) {
143 struct mdinfo *d = aa->info.devs;
144 aa->info.devs = d->next;
150 static struct active_array *duplicate_aa(struct active_array *aa)
152 struct active_array *newa = xmalloc(sizeof(*newa));
153 struct mdinfo **dp1, **dp2;
157 newa->replaces = NULL;
158 newa->info.next = NULL;
160 dp2 = &newa->info.devs;
162 for (dp1 = &aa->info.devs; *dp1; dp1 = &(*dp1)->next) {
164 if ((*dp1)->state_fd < 0)
167 d = xmalloc(sizeof(*d));
177 static void wakeup_monitor(void)
179 /* tgkill(getpid(), mon_tid, SIGUSR1); */
181 syscall(SYS_tgkill, pid, mon_tid, SIGUSR1);
184 static void remove_old(void)
187 discard_this->next = NULL;
188 free_aa(discard_this);
189 if (pending_discard == discard_this)
190 pending_discard = NULL;
196 static void replace_array(struct supertype *container,
197 struct active_array *old,
198 struct active_array *new)
200 /* To replace an array, we add it to the top of the list
201 * marked with ->replaces to point to the original.
202 * 'monitor' will take the original out of the list
203 * and put it on 'discard_this'. We take it from there
207 while (pending_discard) {
208 while (discard_this == NULL)
212 pending_discard = old;
214 new->next = container->arrays;
215 container->arrays = new;
219 struct metadata_update *update_queue = NULL;
220 struct metadata_update *update_queue_handled = NULL;
221 struct metadata_update *update_queue_pending = NULL;
223 static void free_updates(struct metadata_update **update)
226 struct metadata_update *this = *update;
227 void **space_list = this->space_list;
229 *update = this->next;
233 void *space = space_list;
234 space_list = *space_list;
241 void check_update_queue(struct supertype *container)
243 free_updates(&update_queue_handled);
245 if (update_queue == NULL &&
246 update_queue_pending) {
247 update_queue = update_queue_pending;
248 update_queue_pending = NULL;
253 static void queue_metadata_update(struct metadata_update *mu)
255 struct metadata_update **qp;
257 qp = &update_queue_pending;
259 qp = & ((*qp)->next);
263 static void add_disk_to_container(struct supertype *st, struct mdinfo *sd)
267 struct supertype *st2;
268 struct metadata_update *update = NULL;
270 mdu_disk_info_t dk = {
272 .major = sd->disk.major,
273 .minor = sd->disk.minor,
278 dprintf("add %d:%d to container\n", sd->disk.major, sd->disk.minor);
283 sprintf(nm, "%d:%d", sd->disk.major, sd->disk.minor);
284 dfd = dev_open(nm, O_RDWR);
288 /* Check the metadata and see if it is already part of this
292 if (st2->ss->load_super(st2, dfd, NULL) == 0) {
293 st2->ss->getinfo_super(st2, &info, NULL);
294 if (st->ss->compare_super(st, st2) == 0 &&
295 info.disk.raid_disk >= 0) {
296 /* Looks like a good member of array.
298 * mdadm will incorporate any parts into
301 st2->ss->free_super(st2);
305 st2->ss->free_super(st2);
307 st->update_tail = &update;
308 st->ss->add_to_super(st, &dk, dfd, NULL, INVALID_SECTORS);
309 st->ss->write_init_super(st);
310 queue_metadata_update(update);
311 st->update_tail = NULL;
315 * Create and queue update structure about the removed disks.
316 * The update is prepared by super type handler and passed to the monitor
319 static void remove_disk_from_container(struct supertype *st, struct mdinfo *sd)
321 struct metadata_update *update = NULL;
322 mdu_disk_info_t dk = {
324 .major = sd->disk.major,
325 .minor = sd->disk.minor,
329 dprintf("remove %d:%d from container\n",
330 sd->disk.major, sd->disk.minor);
332 st->update_tail = &update;
333 st->ss->remove_from_super(st, &dk);
334 /* FIXME this write_init_super shouldn't be here.
335 * We have it after add_to_super to write to new device,
336 * but with 'remove' we don't ant to write to that device!
338 st->ss->write_init_super(st);
339 queue_metadata_update(update);
340 st->update_tail = NULL;
343 static void manage_container(struct mdstat_ent *mdstat,
344 struct supertype *container)
346 /* Of interest here are:
347 * - if a new device has been added to the container, we
348 * add it to the array ignoring any metadata on it.
349 * - if a device has been removed from the container, we
350 * remove it from the device list and update the metadata.
351 * FIXME should we look for compatible metadata and take hints
352 * about spare assignment.... probably not.
354 if (mdstat->devcnt != container->devcnt) {
355 struct mdinfo **cdp, *cd, *di, *mdi;
358 /* read /sys/block/NAME/md/dev-??/block/dev to find out
359 * what is there, and compare with container->info.devs
360 * To see what is removed and what is added.
361 * These need to be remove from, or added to, the array
363 mdi = sysfs_read(-1, mdstat->devnm, GET_DEVS);
365 /* invalidate the current count so we can try again */
366 container->devcnt = -1;
370 /* check for removals */
371 for (cdp = &container->devs; *cdp; ) {
373 for (di = mdi->devs; di; di = di->next)
374 if (di->disk.major == (*cdp)->disk.major &&
375 di->disk.minor == (*cdp)->disk.minor) {
382 remove_disk_from_container(container, cd);
388 /* check for additions */
389 for (di = mdi->devs; di; di = di->next) {
390 for (cd = container->devs; cd; cd = cd->next)
391 if (di->disk.major == cd->disk.major &&
392 di->disk.minor == cd->disk.minor)
395 struct mdinfo *newd = xmalloc(sizeof(*newd));
398 add_disk_to_container(container, newd);
402 container->devcnt = mdstat->devcnt;
406 static int sysfs_open2(char *devnum, char *name, char *attr)
408 int fd = sysfs_open(devnum, name, attr);
410 /* seq_file in the kernel allocates buffer space
411 * on the first read. Do that now so 'monitor'
415 if (read(fd, buf, sizeof(buf)) < 0)
416 /* pretend not to ignore return value */
422 static int disk_init_and_add(struct mdinfo *disk, struct mdinfo *clone,
423 struct active_array *aa)
429 disk->recovery_fd = sysfs_open2(aa->info.sys_name, disk->sys_name,
431 if (disk->recovery_fd < 0)
433 disk->state_fd = sysfs_open2(aa->info.sys_name, disk->sys_name, "state");
434 if (disk->state_fd < 0) {
435 close(disk->recovery_fd);
438 disk->bb_fd = sysfs_open2(aa->info.sys_name, disk->sys_name,
440 if (disk->bb_fd < 0) {
441 close(disk->recovery_fd);
442 close(disk->state_fd);
445 disk->ubb_fd = sysfs_open2(aa->info.sys_name, disk->sys_name,
446 "unacknowledged_bad_blocks");
447 if (disk->ubb_fd < 0) {
448 close(disk->recovery_fd);
449 close(disk->state_fd);
453 disk->prev_state = read_dev_state(disk->state_fd);
454 disk->curr_state = disk->prev_state;
455 disk->next = aa->info.devs;
456 aa->info.devs = disk;
461 static void manage_member(struct mdstat_ent *mdstat,
462 struct active_array *a)
464 /* Compare mdstat info with known state of member array.
465 * We do not need to look for device state changes here, that
466 * is dealt with by the monitor.
468 * If a reshape is being requested, monitor will have noticed
469 * that sync_action changed and will have set check_reshape.
470 * We just need to see if new devices have appeared. All metadata
471 * updates will already have been processed.
473 * We also want to handle degraded arrays here by
474 * trying to find and assign a spare.
475 * We do that whenever the monitor tells us too.
479 struct supertype *container = a->container;
480 unsigned long long int component_size = 0;
482 if (container == NULL)
483 /* Raced with something */
486 if (mdstat->active) {
488 a->info.array.raid_disks = mdstat->raid_disks;
492 if (sysfs_get_ll(&a->info, NULL, "component_size", &component_size) >= 0)
493 a->info.component_size = component_size << 1;
496 if (sysfs_get_str(&a->info, NULL, "metadata_version", buf, sizeof(buf)) > 0)
497 frozen = buf[9] == '-';
499 frozen = 1; /* can't read metadata_version assume the worst */
501 /* If sync_action is not 'idle' then don't try recovery now */
503 && sysfs_get_str(&a->info, NULL, "sync_action", buf, sizeof(buf)) > 0
504 && strncmp(buf, "idle", 4) != 0)
508 int level = map_name(pers, mdstat->level);
509 if (level == 0 || level == LEVEL_LINEAR) {
514 else if (a->info.array.level != level && level > 0) {
515 struct active_array *newa = duplicate_aa(a);
517 newa->info.array.level = level;
518 replace_array(container, a, newa);
524 /* we are after monitor kick,
525 * so container field can be cleared - check it again
527 if (a->container == NULL)
530 if (sigterm && a->info.safe_mode_delay != 1) {
531 sysfs_set_safemode(&a->info, 1);
532 a->info.safe_mode_delay = 1;
535 /* We don't check the array while any update is pending, as it
536 * might container a change (such as a spare assignment) which
537 * could affect our decisions.
539 if (a->check_degraded && !frozen &&
540 update_queue == NULL && update_queue_pending == NULL) {
541 struct metadata_update *updates = NULL;
542 struct mdinfo *newdev = NULL;
543 struct active_array *newa;
546 a->check_degraded = 0;
548 /* The array may not be degraded, this is just a good time
551 newdev = container->ss->activate_spare(a, &updates);
555 newa = duplicate_aa(a);
558 /* prevent the kernel from activating the disk(s) before we
561 dprintf("freezing %s\n", a->info.sys_name);
562 sysfs_set_str(&a->info, NULL, "sync_action", "frozen");
564 /* Add device to array and set offset/size/slot.
565 * and open files for each newdev */
566 for (d = newdev; d ; d = d->next) {
569 newd = xmalloc(sizeof(*newd));
570 if (sysfs_add_disk(&newa->info, d, 0) < 0) {
574 disk_init_and_add(newd, d, newa);
576 queue_metadata_update(updates);
578 while (update_queue_pending || update_queue) {
579 check_update_queue(container);
582 replace_array(container, a, newa);
583 if (sysfs_set_str(&a->info, NULL, "sync_action", "recover")
585 newa->prev_action = recover;
586 dprintf("recovery started on %s\n", a->info.sys_name);
593 free_updates(&updates);
596 if (a->check_reshape) {
597 /* mdadm might have added some devices to the array.
598 * We want to disk_init_and_add any such device to a
599 * duplicate_aa and replace a with that.
600 * mdstat doesn't have enough info so we sysfs_read
601 * and look for new stuff.
603 struct mdinfo *info, *d, *d2, *newd;
604 unsigned long long array_size;
605 struct active_array *newa = NULL;
606 a->check_reshape = 0;
607 info = sysfs_read(-1, mdstat->devnm,
608 GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE);
611 for (d = info->devs; d; d = d->next) {
612 if (d->disk.raid_disk < 0)
614 for (d2 = a->info.devs; d2; d2 = d2->next)
615 if (d2->disk.raid_disk ==
619 /* already have this one */
622 newa = duplicate_aa(a);
626 newd = xmalloc(sizeof(*newd));
627 disk_init_and_add(newd, d, newa);
629 if (sysfs_get_ll(info, NULL, "array_size", &array_size) == 0
630 && a->info.custom_array_size > array_size*2) {
631 sysfs_set_num(info, NULL, "array_size",
632 a->info.custom_array_size/2);
637 replace_array(container, a, newa);
641 static int aa_ready(struct active_array *aa)
644 int level = aa->info.array.level;
646 for (d = aa->info.devs; d; d = d->next)
650 if (aa->info.state_fd < 0)
653 if (level > 0 && (aa->action_fd < 0 || aa->resync_start_fd < 0))
662 static void manage_new(struct mdstat_ent *mdstat,
663 struct supertype *container,
664 struct active_array *victim)
666 /* A new array has appeared in this container.
667 * Hopefully it is already recorded in the metadata.
668 * Check, then create the new array to report it to
672 struct active_array *new;
673 struct mdinfo *mdi, *di;
679 /* check if array is ready to be monitored */
680 if (!mdstat->active || !mdstat->level)
682 if (strcmp(mdstat->level, "raid0") == 0 ||
683 strcmp(mdstat->level, "linear") == 0)
686 mdi = sysfs_read(-1, mdstat->devnm,
687 GET_LEVEL|GET_CHUNK|GET_DISKS|GET_COMPONENT|
688 GET_DEGRADED|GET_SAFEMODE|
689 GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE|GET_LAYOUT);
693 new = xcalloc(1, sizeof(*new));
695 strcpy(new->info.sys_name, mdstat->devnm);
697 new->prev_state = new->curr_state = new->next_state = inactive;
698 new->prev_action= new->curr_action= new->next_action= idle;
700 new->container = container;
702 inst = to_subarray(mdstat, container->devnm);
704 new->info.array = mdi->array;
705 new->info.component_size = mdi->component_size;
707 for (i = 0; i < new->info.array.raid_disks; i++) {
708 struct mdinfo *newd = xmalloc(sizeof(*newd));
710 for (di = mdi->devs; di; di = di->next)
711 if (i == di->disk.raid_disk)
714 if (disk_init_and_add(newd, di, new) != 0) {
719 if (failed > new->info.array.failed_disks) {
720 /* we cannot properly monitor without all working disks */
721 new->container = NULL;
727 new->action_fd = sysfs_open2(new->info.sys_name, NULL, "sync_action");
728 new->info.state_fd = sysfs_open2(new->info.sys_name, NULL, "array_state");
729 new->resync_start_fd = sysfs_open2(new->info.sys_name, NULL, "resync_start");
730 new->metadata_fd = sysfs_open2(new->info.sys_name, NULL, "metadata_version");
731 new->sync_completed_fd = sysfs_open2(new->info.sys_name, NULL, "sync_completed");
733 dprintf("inst: %s action: %d state: %d\n", inst,
734 new->action_fd, new->info.state_fd);
737 new->info.safe_mode_delay = 1;
738 else if (mdi->safe_mode_delay >= 50)
739 /* Normal start, mdadm set this. */
740 new->info.safe_mode_delay = mdi->safe_mode_delay;
742 /* Restart, just pick a number */
743 new->info.safe_mode_delay = 5000;
744 sysfs_set_safemode(&new->info, new->info.safe_mode_delay);
746 /* reshape_position is set by mdadm in sysfs
747 * read this information for new arrays only (empty victim)
749 if ((victim == NULL) &&
750 (sysfs_get_str(mdi, NULL, "sync_action", buf, 40) > 0) &&
751 (strncmp(buf, "reshape", 7) == 0)) {
752 if (sysfs_get_ll(mdi, NULL, "reshape_position",
753 &new->last_checkpoint) != 0)
754 new->last_checkpoint = 0;
756 int data_disks = mdi->array.raid_disks;
757 if (mdi->array.level == 4 || mdi->array.level == 5)
759 if (mdi->array.level == 6)
762 new->last_checkpoint /= data_disks;
764 dprintf("mdmon: New monitored array is under reshape.\n"
765 " Last checkpoint is: %llu\n",
766 new->last_checkpoint);
771 /* if everything checks out tell the metadata handler we want to
772 * manage this instance
774 if (!aa_ready(new) || container->ss->open_new(container, new, inst) < 0) {
775 pr_err("failed to monitor %s\n",
776 mdstat->metadata_version);
777 new->container = NULL;
780 replace_array(container, victim, new);
782 new->check_degraded = 1;
783 manage_member(mdstat, new);
788 void manage(struct mdstat_ent *mdstat, struct supertype *container)
790 /* We have just read mdstat and need to compare it with
791 * the known active arrays.
792 * Arrays with the wrong metadata are ignored.
795 for ( ; mdstat ; mdstat = mdstat->next) {
796 struct active_array *a;
797 if (strcmp(mdstat->devnm, container->devnm) == 0) {
798 manage_container(mdstat, container);
801 if (!is_container_member(mdstat, container->devnm))
802 /* Not for this array */
804 /* Looks like a member of this container */
805 for (a = container->arrays; a; a = a->next) {
806 if (strcmp(mdstat->devnm, a->info.sys_name) == 0) {
807 if (a->container && a->to_remove == 0)
808 manage_member(mdstat, a);
812 if (a == NULL || !a->container)
813 manage_new(mdstat, container, a);
817 static void handle_message(struct supertype *container, struct metadata_update *msg)
819 /* queue this metadata update through to the monitor */
821 struct metadata_update *mu;
824 while (update_queue_pending || update_queue) {
825 check_update_queue(container);
829 if (msg->len == 0) { /* ping_monitor */
832 cnt = monitor_loop_cnt;
834 cnt += 2; /* wait until next pselect */
836 cnt += 3; /* wait for 2 pselects */
839 while (monitor_loop_cnt - cnt < 0)
841 } else if (msg->len == -1) { /* ping_manager */
842 struct mdstat_ent *mdstat = mdstat_read(1, 0);
844 manage(mdstat, container);
846 } else if (!sigterm) {
847 mu = xmalloc(sizeof(*mu));
852 mu->space_list = NULL;
854 if (container->ss->prepare_update)
855 if (!container->ss->prepare_update(container, mu))
857 queue_metadata_update(mu);
861 void read_sock(struct supertype *container)
864 struct metadata_update msg;
867 int tmo = 3; /* 3 second timeout before hanging up the socket */
869 fd = accept(container->sock, NULL, NULL);
873 fl = fcntl(fd, F_GETFL, 0);
875 fcntl(fd, F_SETFL, fl);
880 /* read and validate the message */
881 if (receive_message(fd, &msg, tmo) == 0) {
882 handle_message(container, &msg);
884 /* ping reply with version */
886 msg.len = strlen(Version) + 1;
887 if (send_message(fd, &msg, tmo) < 0)
889 } else if (ack(fd, tmo) < 0)
894 } while (!terminate);
900 int manager_ready = 0;
901 void do_manager(struct supertype *container)
903 struct mdstat_ent *mdstat;
906 sigprocmask(SIG_UNBLOCK, NULL, &set);
907 sigdelset(&set, SIGUSR1);
908 sigdelset(&set, SIGTERM);
915 /* Can only 'manage' things if 'monitor' is not making
916 * structural changes to metadata, so need to check
919 if (update_queue == NULL) {
920 mdstat = mdstat_read(1, 0);
922 manage(mdstat, container);
924 read_sock(container);
930 check_update_queue(container);
937 if (update_queue == NULL)
938 mdstat_wait_fd(container->sock, &set);
940 /* If an update is happening, just wait for signal */
941 pselect(0, NULL, NULL, NULL, NULL, &set);