]>
git.ipfire.org Git - thirdparty/mdadm.git/blob - monitor.c
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("%s: Invalid fd %d\n", __func__
, fd
);
48 if (st
.st_nlink
== 0) {
49 dprintf("%s: fd %d was deleted\n", __func__
, 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("%s: Failed to read resync_start (%d)\n",
89 if (strncmp(buf
, "none", 4) == 0)
92 *v
= strtoull(buf
, NULL
, 10);
95 static unsigned long long read_sync_completed(int fd
)
97 unsigned long long val
;
102 n
= read_attr(buf
, 50, fd
);
107 val
= strtoull(buf
, &ep
, 0);
108 if (ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))
113 static enum array_state
read_state(int fd
)
116 int n
= read_attr(buf
, 20, fd
);
120 return (enum array_state
) sysfs_match_word(buf
, array_states
);
123 static enum sync_action
read_action( int fd
)
126 int n
= read_attr(buf
, 20, fd
);
130 return (enum sync_action
) sysfs_match_word(buf
, sync_actions
);
133 int read_dev_state(int fd
)
136 int n
= read_attr(buf
, 60, fd
);
145 if (sysfs_attr_match(cp
, "faulty"))
147 if (sysfs_attr_match(cp
, "in_sync"))
149 if (sysfs_attr_match(cp
, "write_mostly"))
150 rv
|= DS_WRITE_MOSTLY
;
151 if (sysfs_attr_match(cp
, "spare"))
153 if (sysfs_attr_match(cp
, "blocked"))
155 cp
= strchr(cp
, ',');
162 static void signal_manager(void)
164 /* tgkill(getpid(), mon_tid, SIGUSR1); */
166 syscall(SYS_tgkill
, pid
, mgr_tid
, SIGUSR1
);
169 /* Monitor a set of active md arrays - all of which share the
170 * same metadata - and respond to events that require
173 * New arrays are detected by another thread which allocates
174 * required memory and attaches the data structure to our list.
178 * This is detected by array_state going to 'clear' or 'inactive'.
179 * while we thought it was active.
180 * Response is to mark metadata as clean and 'clear' the array(??)
182 * array_state if 'write-pending'
183 * We mark metadata as 'dirty' then set array to 'active'.
185 * Either ignore, or mark clean, then mark metadata as clean.
188 * detected by rd-N/state reporting "faulty"
189 * mark device as 'failed' in metadata, let the kernel release the
190 * device by writing '-blocked' to rd/state, and finally write 'remove' to
191 * rd/state. Before a disk can be replaced it must be failed and removed
192 * from all container members, this will be preemptive for the other
196 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
198 * Notify metadata that sync is complete.
201 * sync_action changes from 'recover' to 'idle'
202 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
205 * This only happens on finding a new array... mdadm will have set
206 * 'resync_start' to the correct value. If 'resync_start' indicates that an
207 * resync needs to occur set the array to the 'active' state rather than the
208 * initial read-auto state.
212 * We wait for a change (poll/select) on array_state, sync_action, and
213 * each rd-X/state file.
214 * When we get any change, we check everything. So read each state file,
215 * then decide what to do.
217 * The core action is to write new metadata to all devices in the array.
218 * This is done at most once on any wakeup.
219 * After that we might:
220 * - update the array_state
221 * - set the role of some devices.
222 * - request a sync_action
226 #define ARRAY_DIRTY 1
228 static int read_and_act(struct active_array
*a
)
230 unsigned long long sync_completed
;
231 int check_degraded
= 0;
232 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 if (a
->curr_state
> inactive
&&
262 a
->prev_state
== inactive
) {
263 /* array has been started
264 * possible that container operation has to be completed
266 a
->container
->ss
->set_array_state(a
, 0);
268 if ((a
->curr_state
== bad_word
|| a
->curr_state
<= inactive
) &&
269 a
->prev_state
> inactive
) {
270 /* array has been stopped */
271 a
->container
->ss
->set_array_state(a
, 1);
272 a
->next_state
= clear
;
275 if (a
->curr_state
== write_pending
) {
276 a
->container
->ss
->set_array_state(a
, 0);
277 a
->next_state
= active
;
280 if (a
->curr_state
== active_idle
) {
281 /* Set array to 'clean' FIRST, then mark clean
284 a
->next_state
= clean
;
287 if (a
->curr_state
== clean
) {
288 a
->container
->ss
->set_array_state(a
, 1);
290 if (a
->curr_state
== active
||
291 a
->curr_state
== suspended
)
293 if (a
->curr_state
== readonly
) {
294 /* Well, I'm ready to handle things. If readonly
295 * wasn't requested, transition to read-auto.
298 read_attr(buf
, sizeof(buf
), a
->metadata_fd
);
299 if (strncmp(buf
, "external:-", 10) == 0) {
300 /* explicit request for readonly array. Leave it alone */
303 if (a
->container
->ss
->set_array_state(a
, 2))
304 a
->next_state
= read_auto
; /* array is clean */
306 a
->next_state
= active
; /* Now active for recovery etc */
313 a
->curr_action
== idle
&&
314 a
->prev_action
== resync
) {
315 /* A resync has finished. The endpoint is recorded in
316 * 'sync_start'. We don't update the metadata
317 * until the array goes inactive or readonly though.
318 * Just check if we need to fiddle spares.
320 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
325 a
->curr_action
== idle
&&
326 a
->prev_action
== recover
) {
327 /* A recovery has finished. Some disks may be in sync now,
328 * and the array may no longer be degraded
330 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
331 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
333 if (! (mdi
->curr_state
& DS_INSYNC
))
337 if (count
!= a
->info
.array
.raid_disks
)
342 a
->curr_action
== reshape
&&
343 a
->prev_action
!= reshape
)
344 /* reshape was requested by mdadm. Need to see if
345 * new devices have been added. Manager does that
346 * when it sees check_reshape
350 /* Check for failures and if found:
351 * 1/ Record the failure in the metadata and unblock the device.
352 * FIXME update the kernel to stop notifying on failed drives when
353 * the array is readonly and we have cleared 'blocked'
354 * 2/ Try to remove the device if the array is writable, or can be
357 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
358 if (mdi
->curr_state
& DS_FAULTY
) {
359 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
362 if (mdi
->curr_state
& DS_BLOCKED
)
363 mdi
->next_state
|= DS_UNBLOCK
;
364 if (a
->curr_state
== read_auto
) {
365 a
->container
->ss
->set_array_state(a
, 0);
366 a
->next_state
= active
;
368 if (a
->curr_state
> readonly
)
369 mdi
->next_state
|= DS_REMOVE
;
373 /* Check for recovery checkpoint notifications. We need to be a
374 * minimum distance away from the last checkpoint to prevent
375 * over checkpointing. Note reshape checkpointing is handled
376 * in the second branch.
378 if (sync_completed
> a
->last_checkpoint
&&
379 sync_completed
- a
->last_checkpoint
> a
->info
.component_size
>> 4 &&
380 a
->curr_action
> reshape
) {
381 /* A (non-reshape) sync_action has reached a checkpoint.
382 * Record the updated position in the metadata
384 a
->last_checkpoint
= sync_completed
;
385 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
386 } else if ((a
->curr_action
== idle
&& a
->prev_action
== reshape
) ||
387 (a
->curr_action
== reshape
388 && sync_completed
> a
->last_checkpoint
) ) {
389 /* Reshape has progressed or completed so we need to
390 * update the array state - and possibly the array size
392 if (sync_completed
!= 0)
393 a
->last_checkpoint
= sync_completed
;
394 /* We might need to update last_checkpoint depending on
395 * the reason that reshape finished.
396 * if array reshape is really finished:
397 * set check point to the end, this allows
398 * set_array_state() to finalize reshape in metadata
399 * if reshape if broken: do not set checkpoint to the end
400 * this allows for reshape restart from checkpoint
402 if ((a
->curr_action
!= reshape
) &&
403 (a
->prev_action
== reshape
)) {
405 if ((sysfs_get_str(&a
->info
, NULL
,
408 sizeof(buf
)) >= 0) &&
409 strncmp(buf
, "none", 4) == 0)
410 a
->last_checkpoint
= a
->info
.component_size
;
412 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
413 a
->last_checkpoint
= sync_completed
;
416 if (sync_completed
> a
->last_checkpoint
)
417 a
->last_checkpoint
= sync_completed
;
419 a
->container
->ss
->sync_metadata(a
->container
);
420 dprintf("%s(%d): state:%s action:%s next(", __func__
, a
->info
.container_member
,
421 array_states
[a
->curr_state
], sync_actions
[a
->curr_action
]);
423 /* Effect state changes in the array */
424 if (a
->next_state
!= bad_word
) {
425 dprintf(" state:%s", array_states
[a
->next_state
]);
426 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
428 if (a
->next_action
!= bad_action
) {
429 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
430 dprintf(" action:%s", sync_actions
[a
->next_action
]);
432 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
433 if (mdi
->next_state
& DS_UNBLOCK
) {
434 dprintf(" %d:-blocked", mdi
->disk
.raid_disk
);
435 write_attr("-blocked", mdi
->state_fd
);
438 if ((mdi
->next_state
& DS_REMOVE
) && mdi
->state_fd
>= 0) {
441 /* The kernel may not be able to immediately remove the
442 * disk. In that case we wait a little while and
445 remove_result
= write_attr("remove", mdi
->state_fd
);
446 if (remove_result
> 0) {
447 dprintf(" %d:removed", mdi
->disk
.raid_disk
);
448 close(mdi
->state_fd
);
449 close(mdi
->recovery_fd
);
454 if (mdi
->next_state
& DS_INSYNC
) {
455 write_attr("+in_sync", mdi
->state_fd
);
456 dprintf(" %d:+in_sync", mdi
->disk
.raid_disk
);
461 /* move curr_ to prev_ */
462 a
->prev_state
= a
->curr_state
;
464 a
->prev_action
= a
->curr_action
;
466 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
467 mdi
->prev_state
= mdi
->curr_state
;
471 if (check_degraded
|| check_reshape
) {
472 /* manager will do the actual check */
474 a
->check_degraded
= 1;
476 a
->check_reshape
= 1;
486 static struct mdinfo
*
487 find_device(struct active_array
*a
, int major
, int minor
)
491 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
492 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
498 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
500 struct active_array
*a
;
501 struct mdinfo
*victim
;
503 for (a
= aa
; a
; a
= a
->next
) {
504 if (!a
->container
|| a
->to_remove
)
506 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
510 if (!(victim
->curr_state
& DS_FAULTY
))
511 write_attr("faulty", victim
->state_fd
);
516 static void dprint_wake_reasons(fd_set
*fds
)
524 fprintf(stderr
, "monitor: wake ( ");
525 for (i
= 0; i
< FD_SETSIZE
; i
++) {
526 if (FD_ISSET(i
, fds
)) {
527 sprintf(proc_path
, "/proc/%d/fd/%d",
530 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
532 fprintf(stderr
, "%d:unknown ", i
);
536 basename
= strrchr(link
, '/');
537 fprintf(stderr
, "%d:%s ",
538 i
, basename
? ++basename
: link
);
541 fprintf(stderr
, ")\n");
545 int monitor_loop_cnt
;
547 static int wait_and_act(struct supertype
*container
, int nowait
)
551 struct active_array
**aap
= &container
->arrays
;
552 struct active_array
*a
, **ap
;
555 static unsigned int dirty_arrays
= ~0; /* start at some non-zero value */
559 for (ap
= aap
; *ap
;) {
561 /* once an array has been deactivated we want to
562 * ask the manager to discard it.
564 if (!a
->container
|| a
->to_remove
) {
576 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
577 add_fd(&rfds
, &maxfd
, a
->action_fd
);
578 add_fd(&rfds
, &maxfd
, a
->sync_completed_fd
);
579 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
580 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
585 if (manager_ready
&& (*aap
== NULL
|| (sigterm
&& !dirty_arrays
))) {
586 /* No interesting arrays, or we have been told to
587 * terminate and everything is clean. Lets see about
588 * exiting. Note that blocking at this point is not a
589 * problem as there are no active arrays, there is
590 * nothing that we need to be ready to do.
594 fd
= open_dev_excl(container
->devnm
);
596 fd
= open_dev_flags(container
->devnm
, O_RDONLY
|O_EXCL
);
597 if (fd
>= 0 || errno
!= EBUSY
) {
598 /* OK, we are safe to leave */
599 if (sigterm
&& !dirty_arrays
)
600 dprintf("caught sigterm, all clean... exiting\n");
602 dprintf("no arrays to monitor... exiting\n");
604 /* On SIGTERM, someone (the take-over mdmon) will
607 remove_pidfile(container
->devnm
);
620 if (*aap
== NULL
|| container
->retry_soon
) {
621 /* just waiting to get O_EXCL access */
623 ts
.tv_nsec
= 20000000ULL;
625 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
626 sigdelset(&set
, SIGUSR1
);
627 monitor_loop_cnt
|= 1;
628 rv
= pselect(maxfd
+1, NULL
, NULL
, &rfds
, &ts
, &set
);
629 monitor_loop_cnt
+= 1;
630 if (rv
== -1 && errno
== EINTR
)
633 dprint_wake_reasons(&rfds
);
635 container
->retry_soon
= 0;
639 struct metadata_update
*this;
641 for (this = update_queue
; this ; this = this->next
)
642 container
->ss
->process_update(container
, this);
644 update_queue_handled
= update_queue
;
647 container
->ss
->sync_metadata(container
);
652 for (a
= *aap
; a
; a
= a
->next
) {
654 if (a
->replaces
&& !discard_this
) {
655 struct active_array
**ap
;
656 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
661 discard_this
= a
->replaces
;
663 /* FIXME check if device->state_fd need to be cleared?*/
666 if (a
->container
&& !a
->to_remove
) {
667 int ret
= read_and_act(a
);
669 dirty_arrays
+= !!(ret
& ARRAY_DIRTY
);
670 /* when terminating stop manipulating the array after it
671 * is clean, but make sure read_and_act() is given a
672 * chance to handle 'active_idle'
674 if (sigterm
&& !(ret
& ARRAY_DIRTY
))
675 a
->container
= NULL
; /* stop touching this array */
676 if (ret
& ARRAY_BUSY
)
677 container
->retry_soon
= 1;
681 /* propagate failures across container members */
682 for (a
= *aap
; a
; a
= a
->next
) {
683 if (!a
->container
|| a
->to_remove
)
685 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
686 if (mdi
->curr_state
& DS_FAULTY
)
687 reconcile_failed(*aap
, mdi
);
693 void do_monitor(struct supertype
*container
)
698 rv
= wait_and_act(container
, first
);