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);
120 if (aa->action_fd >= 0)
121 close(aa->action_fd);
122 if (aa->info.state_fd >= 0)
123 close(aa->info.state_fd);
124 if (aa->resync_start_fd >= 0)
125 close(aa->resync_start_fd);
126 if (aa->metadata_fd >= 0)
127 close(aa->metadata_fd);
128 if (aa->sync_completed_fd >= 0)
129 close(aa->sync_completed_fd);
132 static void free_aa(struct active_array *aa)
134 /* Note that this doesn't close fds if they are being used
135 * by a clone. ->container will be set for a clone
137 dprintf("sys_name: %s\n", aa->info.sys_name);
140 while (aa->info.devs) {
141 struct mdinfo *d = aa->info.devs;
142 aa->info.devs = d->next;
148 static struct active_array *duplicate_aa(struct active_array *aa)
150 struct active_array *newa = xmalloc(sizeof(*newa));
151 struct mdinfo **dp1, **dp2;
155 newa->replaces = NULL;
156 newa->info.next = NULL;
158 dp2 = &newa->info.devs;
160 for (dp1 = &aa->info.devs; *dp1; dp1 = &(*dp1)->next) {
162 if ((*dp1)->state_fd < 0)
165 d = xmalloc(sizeof(*d));
175 static void wakeup_monitor(void)
177 /* tgkill(getpid(), mon_tid, SIGUSR1); */
179 syscall(SYS_tgkill, pid, mon_tid, SIGUSR1);
182 static void remove_old(void)
185 discard_this->next = NULL;
186 free_aa(discard_this);
187 if (pending_discard == discard_this)
188 pending_discard = NULL;
194 static void replace_array(struct supertype *container,
195 struct active_array *old,
196 struct active_array *new)
198 /* To replace an array, we add it to the top of the list
199 * marked with ->replaces to point to the original.
200 * 'monitor' will take the original out of the list
201 * and put it on 'discard_this'. We take it from there
205 while (pending_discard) {
206 while (discard_this == NULL)
210 pending_discard = old;
212 new->next = container->arrays;
213 container->arrays = new;
217 struct metadata_update *update_queue = NULL;
218 struct metadata_update *update_queue_handled = NULL;
219 struct metadata_update *update_queue_pending = NULL;
221 static void free_updates(struct metadata_update **update)
224 struct metadata_update *this = *update;
225 void **space_list = this->space_list;
227 *update = this->next;
231 void *space = space_list;
232 space_list = *space_list;
239 void check_update_queue(struct supertype *container)
241 free_updates(&update_queue_handled);
243 if (update_queue == NULL &&
244 update_queue_pending) {
245 update_queue = update_queue_pending;
246 update_queue_pending = NULL;
251 static void queue_metadata_update(struct metadata_update *mu)
253 struct metadata_update **qp;
255 qp = &update_queue_pending;
257 qp = & ((*qp)->next);
261 static void add_disk_to_container(struct supertype *st, struct mdinfo *sd)
265 struct supertype *st2;
266 struct metadata_update *update = NULL;
268 mdu_disk_info_t dk = {
270 .major = sd->disk.major,
271 .minor = sd->disk.minor,
276 dprintf("add %d:%d to container\n", sd->disk.major, sd->disk.minor);
281 sprintf(nm, "%d:%d", sd->disk.major, sd->disk.minor);
282 dfd = dev_open(nm, O_RDWR);
286 /* Check the metadata and see if it is already part of this
290 if (st2->ss->load_super(st2, dfd, NULL) == 0) {
291 st2->ss->getinfo_super(st2, &info, NULL);
292 if (st->ss->compare_super(st, st2) == 0 &&
293 info.disk.raid_disk >= 0) {
294 /* Looks like a good member of array.
296 * mdadm will incorporate any parts into
299 st2->ss->free_super(st2);
303 st2->ss->free_super(st2);
305 st->update_tail = &update;
306 st->ss->add_to_super(st, &dk, dfd, NULL, INVALID_SECTORS);
307 st->ss->write_init_super(st);
308 queue_metadata_update(update);
309 st->update_tail = NULL;
313 * Create and queue update structure about the removed disks.
314 * The update is prepared by super type handler and passed to the monitor
317 static void remove_disk_from_container(struct supertype *st, struct mdinfo *sd)
319 struct metadata_update *update = NULL;
320 mdu_disk_info_t dk = {
322 .major = sd->disk.major,
323 .minor = sd->disk.minor,
327 dprintf("remove %d:%d from container\n",
328 sd->disk.major, sd->disk.minor);
330 st->update_tail = &update;
331 st->ss->remove_from_super(st, &dk);
332 /* FIXME this write_init_super shouldn't be here.
333 * We have it after add_to_super to write to new device,
334 * but with 'remove' we don't ant to write to that device!
336 st->ss->write_init_super(st);
337 queue_metadata_update(update);
338 st->update_tail = NULL;
341 static void manage_container(struct mdstat_ent *mdstat,
342 struct supertype *container)
344 /* Of interest here are:
345 * - if a new device has been added to the container, we
346 * add it to the array ignoring any metadata on it.
347 * - if a device has been removed from the container, we
348 * remove it from the device list and update the metadata.
349 * FIXME should we look for compatible metadata and take hints
350 * about spare assignment.... probably not.
352 if (mdstat->devcnt != container->devcnt) {
353 struct mdinfo **cdp, *cd, *di, *mdi;
356 /* read /sys/block/NAME/md/dev-??/block/dev to find out
357 * what is there, and compare with container->info.devs
358 * To see what is removed and what is added.
359 * These need to be remove from, or added to, the array
361 mdi = sysfs_read(-1, mdstat->devnm, GET_DEVS);
363 /* invalidate the current count so we can try again */
364 container->devcnt = -1;
368 /* check for removals */
369 for (cdp = &container->devs; *cdp; ) {
371 for (di = mdi->devs; di; di = di->next)
372 if (di->disk.major == (*cdp)->disk.major &&
373 di->disk.minor == (*cdp)->disk.minor) {
380 remove_disk_from_container(container, cd);
386 /* check for additions */
387 for (di = mdi->devs; di; di = di->next) {
388 for (cd = container->devs; cd; cd = cd->next)
389 if (di->disk.major == cd->disk.major &&
390 di->disk.minor == cd->disk.minor)
393 struct mdinfo *newd = xmalloc(sizeof(*newd));
396 add_disk_to_container(container, newd);
400 container->devcnt = mdstat->devcnt;
404 static int sysfs_open2(char *devnum, char *name, char *attr)
406 int fd = sysfs_open(devnum, name, attr);
408 /* seq_file in the kernel allocates buffer space
409 * on the first read. Do that now so 'monitor'
413 if (read(fd, buf, sizeof(buf)) < 0)
414 /* pretend not to ignore return value */
420 static int disk_init_and_add(struct mdinfo *disk, struct mdinfo *clone,
421 struct active_array *aa)
427 disk->recovery_fd = sysfs_open2(aa->info.sys_name, disk->sys_name,
429 if (disk->recovery_fd < 0)
431 disk->state_fd = sysfs_open2(aa->info.sys_name, disk->sys_name, "state");
432 if (disk->state_fd < 0) {
433 close(disk->recovery_fd);
436 disk->prev_state = read_dev_state(disk->state_fd);
437 disk->curr_state = disk->prev_state;
438 disk->next = aa->info.devs;
439 aa->info.devs = disk;
444 static void manage_member(struct mdstat_ent *mdstat,
445 struct active_array *a)
447 /* Compare mdstat info with known state of member array.
448 * We do not need to look for device state changes here, that
449 * is dealt with by the monitor.
451 * If a reshape is being requested, monitor will have noticed
452 * that sync_action changed and will have set check_reshape.
453 * We just need to see if new devices have appeared. All metadata
454 * updates will already have been processed.
456 * We also want to handle degraded arrays here by
457 * trying to find and assign a spare.
458 * We do that whenever the monitor tells us too.
462 struct supertype *container = a->container;
463 unsigned long long int component_size = 0;
465 if (container == NULL)
466 /* Raced with something */
469 if (mdstat->active) {
471 a->info.array.raid_disks = mdstat->raid_disks;
475 if (sysfs_get_ll(&a->info, NULL, "component_size", &component_size) >= 0)
476 a->info.component_size = component_size << 1;
479 if (sysfs_get_str(&a->info, NULL, "metadata_version", buf, sizeof(buf)) > 0)
480 frozen = buf[9] == '-';
482 frozen = 1; /* can't read metadata_version assume the worst */
484 /* If sync_action is not 'idle' then don't try recovery now */
486 && sysfs_get_str(&a->info, NULL, "sync_action", buf, sizeof(buf)) > 0
487 && strncmp(buf, "idle", 4) != 0)
491 int level = map_name(pers, mdstat->level);
492 if (level == 0 || level == LEVEL_LINEAR) {
497 else if (a->info.array.level != level && level > 0) {
498 struct active_array *newa = duplicate_aa(a);
500 newa->info.array.level = level;
501 replace_array(container, a, newa);
507 /* we are after monitor kick,
508 * so container field can be cleared - check it again
510 if (a->container == NULL)
513 if (sigterm && a->info.safe_mode_delay != 1) {
514 sysfs_set_safemode(&a->info, 1);
515 a->info.safe_mode_delay = 1;
518 /* We don't check the array while any update is pending, as it
519 * might container a change (such as a spare assignment) which
520 * could affect our decisions.
522 if (a->check_degraded && !frozen &&
523 update_queue == NULL && update_queue_pending == NULL) {
524 struct metadata_update *updates = NULL;
525 struct mdinfo *newdev = NULL;
526 struct active_array *newa;
529 a->check_degraded = 0;
531 /* The array may not be degraded, this is just a good time
534 newdev = container->ss->activate_spare(a, &updates);
538 newa = duplicate_aa(a);
541 /* prevent the kernel from activating the disk(s) before we
544 dprintf("freezing %s\n", a->info.sys_name);
545 sysfs_set_str(&a->info, NULL, "sync_action", "frozen");
547 /* Add device to array and set offset/size/slot.
548 * and open files for each newdev */
549 for (d = newdev; d ; d = d->next) {
552 newd = xmalloc(sizeof(*newd));
553 if (sysfs_add_disk(&newa->info, d, 0) < 0) {
557 disk_init_and_add(newd, d, newa);
559 queue_metadata_update(updates);
561 while (update_queue_pending || update_queue) {
562 check_update_queue(container);
565 replace_array(container, a, newa);
566 if (sysfs_set_str(&a->info, NULL, "sync_action", "recover")
568 newa->prev_action = recover;
569 dprintf("recovery started on %s\n", a->info.sys_name);
576 free_updates(&updates);
579 if (a->check_reshape) {
580 /* mdadm might have added some devices to the array.
581 * We want to disk_init_and_add any such device to a
582 * duplicate_aa and replace a with that.
583 * mdstat doesn't have enough info so we sysfs_read
584 * and look for new stuff.
586 struct mdinfo *info, *d, *d2, *newd;
587 unsigned long long array_size;
588 struct active_array *newa = NULL;
589 a->check_reshape = 0;
590 info = sysfs_read(-1, mdstat->devnm,
591 GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE);
594 for (d = info->devs; d; d = d->next) {
595 if (d->disk.raid_disk < 0)
597 for (d2 = a->info.devs; d2; d2 = d2->next)
598 if (d2->disk.raid_disk ==
602 /* already have this one */
605 newa = duplicate_aa(a);
609 newd = xmalloc(sizeof(*newd));
610 disk_init_and_add(newd, d, newa);
612 if (sysfs_get_ll(info, NULL, "array_size", &array_size) == 0
613 && a->info.custom_array_size > array_size*2) {
614 sysfs_set_num(info, NULL, "array_size",
615 a->info.custom_array_size/2);
620 replace_array(container, a, newa);
624 static int aa_ready(struct active_array *aa)
627 int level = aa->info.array.level;
629 for (d = aa->info.devs; d; d = d->next)
633 if (aa->info.state_fd < 0)
636 if (level > 0 && (aa->action_fd < 0 || aa->resync_start_fd < 0))
645 static void manage_new(struct mdstat_ent *mdstat,
646 struct supertype *container,
647 struct active_array *victim)
649 /* A new array has appeared in this container.
650 * Hopefully it is already recorded in the metadata.
651 * Check, then create the new array to report it to
655 struct active_array *new;
656 struct mdinfo *mdi, *di;
662 /* check if array is ready to be monitored */
663 if (!mdstat->active || !mdstat->level)
665 if (strcmp(mdstat->level, "raid0") == 0 ||
666 strcmp(mdstat->level, "linear") == 0)
669 mdi = sysfs_read(-1, mdstat->devnm,
670 GET_LEVEL|GET_CHUNK|GET_DISKS|GET_COMPONENT|
671 GET_DEGRADED|GET_SAFEMODE|
672 GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE|GET_LAYOUT);
676 new = xcalloc(1, sizeof(*new));
678 strcpy(new->info.sys_name, mdstat->devnm);
680 new->prev_state = new->curr_state = new->next_state = inactive;
681 new->prev_action= new->curr_action= new->next_action= idle;
683 new->container = container;
685 inst = to_subarray(mdstat, container->devnm);
687 new->info.array = mdi->array;
688 new->info.component_size = mdi->component_size;
690 for (i = 0; i < new->info.array.raid_disks; i++) {
691 struct mdinfo *newd = xmalloc(sizeof(*newd));
693 for (di = mdi->devs; di; di = di->next)
694 if (i == di->disk.raid_disk)
697 if (disk_init_and_add(newd, di, new) != 0) {
702 if (failed > new->info.array.failed_disks) {
703 /* we cannot properly monitor without all working disks */
704 new->container = NULL;
710 new->action_fd = sysfs_open2(new->info.sys_name, NULL, "sync_action");
711 new->info.state_fd = sysfs_open2(new->info.sys_name, NULL, "array_state");
712 new->resync_start_fd = sysfs_open2(new->info.sys_name, NULL, "resync_start");
713 new->metadata_fd = sysfs_open2(new->info.sys_name, NULL, "metadata_version");
714 new->sync_completed_fd = sysfs_open2(new->info.sys_name, NULL, "sync_completed");
716 dprintf("inst: %s action: %d state: %d\n", inst,
717 new->action_fd, new->info.state_fd);
720 new->info.safe_mode_delay = 1;
721 else if (mdi->safe_mode_delay >= 50)
722 /* Normal start, mdadm set this. */
723 new->info.safe_mode_delay = mdi->safe_mode_delay;
725 /* Restart, just pick a number */
726 new->info.safe_mode_delay = 5000;
727 sysfs_set_safemode(&new->info, new->info.safe_mode_delay);
729 /* reshape_position is set by mdadm in sysfs
730 * read this information for new arrays only (empty victim)
732 if ((victim == NULL) &&
733 (sysfs_get_str(mdi, NULL, "sync_action", buf, 40) > 0) &&
734 (strncmp(buf, "reshape", 7) == 0)) {
735 if (sysfs_get_ll(mdi, NULL, "reshape_position",
736 &new->last_checkpoint) != 0)
737 new->last_checkpoint = 0;
739 int data_disks = mdi->array.raid_disks;
740 if (mdi->array.level == 4 || mdi->array.level == 5)
742 if (mdi->array.level == 6)
745 new->last_checkpoint /= data_disks;
747 dprintf("mdmon: New monitored array is under reshape.\n"
748 " Last checkpoint is: %llu\n",
749 new->last_checkpoint);
754 /* if everything checks out tell the metadata handler we want to
755 * manage this instance
757 if (!aa_ready(new) || container->ss->open_new(container, new, inst) < 0) {
758 pr_err("failed to monitor %s\n",
759 mdstat->metadata_version);
760 new->container = NULL;
763 replace_array(container, victim, new);
765 new->check_degraded = 1;
766 manage_member(mdstat, new);
771 void manage(struct mdstat_ent *mdstat, struct supertype *container)
773 /* We have just read mdstat and need to compare it with
774 * the known active arrays.
775 * Arrays with the wrong metadata are ignored.
778 for ( ; mdstat ; mdstat = mdstat->next) {
779 struct active_array *a;
780 if (strcmp(mdstat->devnm, container->devnm) == 0) {
781 manage_container(mdstat, container);
784 if (!is_container_member(mdstat, container->devnm))
785 /* Not for this array */
787 /* Looks like a member of this container */
788 for (a = container->arrays; a; a = a->next) {
789 if (strcmp(mdstat->devnm, a->info.sys_name) == 0) {
790 if (a->container && a->to_remove == 0)
791 manage_member(mdstat, a);
795 if (a == NULL || !a->container)
796 manage_new(mdstat, container, a);
800 static void handle_message(struct supertype *container, struct metadata_update *msg)
802 /* queue this metadata update through to the monitor */
804 struct metadata_update *mu;
807 while (update_queue_pending || update_queue) {
808 check_update_queue(container);
812 if (msg->len == 0) { /* ping_monitor */
815 cnt = monitor_loop_cnt;
817 cnt += 2; /* wait until next pselect */
819 cnt += 3; /* wait for 2 pselects */
822 while (monitor_loop_cnt - cnt < 0)
824 } else if (msg->len == -1) { /* ping_manager */
825 struct mdstat_ent *mdstat = mdstat_read(1, 0);
827 manage(mdstat, container);
829 } else if (!sigterm) {
830 mu = xmalloc(sizeof(*mu));
835 mu->space_list = NULL;
837 if (container->ss->prepare_update)
838 if (!container->ss->prepare_update(container, mu))
840 queue_metadata_update(mu);
844 void read_sock(struct supertype *container)
847 struct metadata_update msg;
850 int tmo = 3; /* 3 second timeout before hanging up the socket */
852 fd = accept(container->sock, NULL, NULL);
856 fl = fcntl(fd, F_GETFL, 0);
858 fcntl(fd, F_SETFL, fl);
863 /* read and validate the message */
864 if (receive_message(fd, &msg, tmo) == 0) {
865 handle_message(container, &msg);
867 /* ping reply with version */
869 msg.len = strlen(Version) + 1;
870 if (send_message(fd, &msg, tmo) < 0)
872 } else if (ack(fd, tmo) < 0)
877 } while (!terminate);
883 int manager_ready = 0;
884 void do_manager(struct supertype *container)
886 struct mdstat_ent *mdstat;
889 sigprocmask(SIG_UNBLOCK, NULL, &set);
890 sigdelset(&set, SIGUSR1);
891 sigdelset(&set, SIGTERM);
898 /* Can only 'manage' things if 'monitor' is not making
899 * structural changes to metadata, so need to check
902 if (update_queue == NULL) {
903 mdstat = mdstat_read(1, 0);
905 manage(mdstat, container);
907 read_sock(container);
913 check_update_queue(container);
920 if (update_queue == NULL)
921 mdstat_wait_fd(container->sock, &set);
923 /* If an update is happening, just wait for signal */
924 pselect(0, NULL, NULL, NULL, NULL, &set);