]>
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 unsigned long long read_resync_start(int fd
)
83 n
= read_attr(buf
, 30, fd
);
86 if (strncmp(buf
, "none", 4) == 0)
89 return strtoull(buf
, NULL
, 10);
92 static unsigned long long read_sync_completed(int fd
)
94 unsigned long long val
;
99 n
= read_attr(buf
, 50, fd
);
104 val
= strtoull(buf
, &ep
, 0);
105 if (ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))
110 static enum array_state
read_state(int fd
)
113 int n
= read_attr(buf
, 20, fd
);
117 return (enum array_state
) sysfs_match_word(buf
, array_states
);
120 static enum sync_action
read_action( int fd
)
123 int n
= read_attr(buf
, 20, fd
);
127 return (enum sync_action
) sysfs_match_word(buf
, sync_actions
);
130 int read_dev_state(int fd
)
133 int n
= read_attr(buf
, 60, fd
);
142 if (sysfs_attr_match(cp
, "faulty"))
144 if (sysfs_attr_match(cp
, "in_sync"))
146 if (sysfs_attr_match(cp
, "write_mostly"))
147 rv
|= DS_WRITE_MOSTLY
;
148 if (sysfs_attr_match(cp
, "spare"))
150 if (sysfs_attr_match(cp
, "blocked"))
152 cp
= strchr(cp
, ',');
159 static void signal_manager(void)
161 /* tgkill(getpid(), mon_tid, SIGUSR1); */
163 syscall(SYS_tgkill
, pid
, mgr_tid
, SIGUSR1
);
166 /* Monitor a set of active md arrays - all of which share the
167 * same metadata - and respond to events that require
170 * New arrays are detected by another thread which allocates
171 * required memory and attaches the data structure to our list.
175 * This is detected by array_state going to 'clear' or 'inactive'.
176 * while we thought it was active.
177 * Response is to mark metadata as clean and 'clear' the array(??)
179 * array_state if 'write-pending'
180 * We mark metadata as 'dirty' then set array to 'active'.
182 * Either ignore, or mark clean, then mark metadata as clean.
185 * detected by rd-N/state reporting "faulty"
186 * mark device as 'failed' in metadata, let the kernel release the
187 * device by writing '-blocked' to rd/state, and finally write 'remove' to
188 * rd/state. Before a disk can be replaced it must be failed and removed
189 * from all container members, this will be preemptive for the other
193 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
195 * Notify metadata that sync is complete.
198 * sync_action changes from 'recover' to 'idle'
199 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
202 * This only happens on finding a new array... mdadm will have set
203 * 'resync_start' to the correct value. If 'resync_start' indicates that an
204 * resync needs to occur set the array to the 'active' state rather than the
205 * initial read-auto state.
209 * We wait for a change (poll/select) on array_state, sync_action, and
210 * each rd-X/state file.
211 * When we get any change, we check everything. So read each state file,
212 * then decide what to do.
214 * The core action is to write new metadata to all devices in the array.
215 * This is done at most once on any wakeup.
216 * After that we might:
217 * - update the array_state
218 * - set the role of some devices.
219 * - request a sync_action
223 #define ARRAY_DIRTY 1
225 static int read_and_act(struct active_array
*a
)
227 unsigned long long sync_completed
;
228 int check_degraded
= 0;
229 int check_reshape
= 0;
235 a
->next_state
= bad_word
;
236 a
->next_action
= bad_action
;
238 a
->curr_state
= read_state(a
->info
.state_fd
);
239 a
->curr_action
= read_action(a
->action_fd
);
240 a
->info
.resync_start
= read_resync_start(a
->resync_start_fd
);
241 sync_completed
= read_sync_completed(a
->sync_completed_fd
);
242 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
245 if (mdi
->state_fd
>= 0) {
246 mdi
->recovery_start
= read_resync_start(mdi
->recovery_fd
);
247 mdi
->curr_state
= read_dev_state(mdi
->state_fd
);
251 if (a
->curr_state
> inactive
&&
252 a
->prev_state
== inactive
) {
253 /* array has been started
254 * possible that container operation has to be completed
256 a
->container
->ss
->set_array_state(a
, 0);
258 if (a
->curr_state
<= inactive
&&
259 a
->prev_state
> inactive
) {
260 /* array has been stopped */
261 a
->container
->ss
->set_array_state(a
, 1);
262 a
->next_state
= clear
;
265 if (a
->curr_state
== write_pending
) {
266 a
->container
->ss
->set_array_state(a
, 0);
267 a
->next_state
= active
;
270 if (a
->curr_state
== active_idle
) {
271 /* Set array to 'clean' FIRST, then mark clean
274 a
->next_state
= clean
;
277 if (a
->curr_state
== clean
) {
278 a
->container
->ss
->set_array_state(a
, 1);
280 if (a
->curr_state
== active
||
281 a
->curr_state
== suspended
||
282 a
->curr_state
== bad_word
)
284 if (a
->curr_state
== readonly
) {
285 /* Well, I'm ready to handle things. If readonly
286 * wasn't requested, transition to read-auto.
289 read_attr(buf
, sizeof(buf
), a
->metadata_fd
);
290 if (strncmp(buf
, "external:-", 10) == 0) {
291 /* explicit request for readonly array. Leave it alone */
294 if (a
->container
->ss
->set_array_state(a
, 2))
295 a
->next_state
= read_auto
; /* array is clean */
297 a
->next_state
= active
; /* Now active for recovery etc */
304 a
->curr_action
== idle
&&
305 a
->prev_action
== resync
) {
306 /* A resync has finished. The endpoint is recorded in
307 * 'sync_start'. We don't update the metadata
308 * until the array goes inactive or readonly though.
309 * Just check if we need to fiddle spares.
311 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
316 a
->curr_action
== idle
&&
317 a
->prev_action
== recover
) {
318 /* A recovery has finished. Some disks may be in sync now,
319 * and the array may no longer be degraded
321 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
322 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
324 if (! (mdi
->curr_state
& DS_INSYNC
))
328 if (count
!= a
->info
.array
.raid_disks
)
333 a
->curr_action
== reshape
&&
334 a
->prev_action
!= reshape
)
335 /* reshape was requested by mdadm. Need to see if
336 * new devices have been added. Manager does that
337 * when it sees check_reshape
341 /* Check for failures and if found:
342 * 1/ Record the failure in the metadata and unblock the device.
343 * FIXME update the kernel to stop notifying on failed drives when
344 * the array is readonly and we have cleared 'blocked'
345 * 2/ Try to remove the device if the array is writable, or can be
348 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
349 if (mdi
->curr_state
& DS_FAULTY
) {
350 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
353 if (mdi
->curr_state
& DS_BLOCKED
)
354 mdi
->next_state
|= DS_UNBLOCK
;
355 if (a
->curr_state
== read_auto
) {
356 a
->container
->ss
->set_array_state(a
, 0);
357 a
->next_state
= active
;
359 if (a
->curr_state
> readonly
)
360 mdi
->next_state
|= DS_REMOVE
;
364 /* Check for recovery checkpoint notifications. We need to be a
365 * minimum distance away from the last checkpoint to prevent
366 * over checkpointing. Note reshape checkpointing is handled
367 * in the second branch.
369 if (sync_completed
> a
->last_checkpoint
&&
370 sync_completed
- a
->last_checkpoint
> a
->info
.component_size
>> 4 &&
371 a
->curr_action
> reshape
) {
372 /* A (non-reshape) sync_action has reached a checkpoint.
373 * Record the updated position in the metadata
375 a
->last_checkpoint
= sync_completed
;
376 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
377 } else if ((a
->curr_action
== idle
&& a
->prev_action
== reshape
) ||
378 (a
->curr_action
== reshape
379 && sync_completed
> a
->last_checkpoint
) ) {
380 /* Reshape has progressed or completed so we need to
381 * update the array state - and possibly the array size
383 if (sync_completed
!= 0)
384 a
->last_checkpoint
= sync_completed
;
385 /* We might need to update last_checkpoint depending on
386 * the reason that reshape finished.
387 * if array reshape is really finished:
388 * set check point to the end, this allows
389 * set_array_state() to finalize reshape in metadata
390 * if reshape if broken: do not set checkpoint to the end
391 * this allows for reshape restart from checkpoint
393 if ((a
->curr_action
!= reshape
) &&
394 (a
->prev_action
== reshape
)) {
396 if ((sysfs_get_str(&a
->info
, NULL
,
399 sizeof(buf
)) >= 0) &&
400 strncmp(buf
, "none", 4) == 0)
401 a
->last_checkpoint
= a
->info
.component_size
;
403 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
404 a
->last_checkpoint
= sync_completed
;
407 if (sync_completed
> a
->last_checkpoint
)
408 a
->last_checkpoint
= sync_completed
;
410 a
->container
->ss
->sync_metadata(a
->container
);
411 dprintf("%s(%d): state:%s action:%s next(", __func__
, a
->info
.container_member
,
412 array_states
[a
->curr_state
], sync_actions
[a
->curr_action
]);
414 /* Effect state changes in the array */
415 if (a
->next_state
!= bad_word
) {
416 dprintf(" state:%s", array_states
[a
->next_state
]);
417 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
419 if (a
->next_action
!= bad_action
) {
420 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
421 dprintf(" action:%s", sync_actions
[a
->next_action
]);
423 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
424 if (mdi
->next_state
& DS_UNBLOCK
) {
425 dprintf(" %d:-blocked", mdi
->disk
.raid_disk
);
426 write_attr("-blocked", mdi
->state_fd
);
429 if ((mdi
->next_state
& DS_REMOVE
) && mdi
->state_fd
>= 0) {
432 /* The kernel may not be able to immediately remove the
433 * disk. In that case we wait a little while and
436 remove_result
= write_attr("remove", mdi
->state_fd
);
437 if (remove_result
> 0) {
438 dprintf(" %d:removed", mdi
->disk
.raid_disk
);
439 close(mdi
->state_fd
);
440 close(mdi
->recovery_fd
);
445 if (mdi
->next_state
& DS_INSYNC
) {
446 write_attr("+in_sync", mdi
->state_fd
);
447 dprintf(" %d:+in_sync", mdi
->disk
.raid_disk
);
452 /* move curr_ to prev_ */
453 a
->prev_state
= a
->curr_state
;
455 a
->prev_action
= a
->curr_action
;
457 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
458 mdi
->prev_state
= mdi
->curr_state
;
462 if (check_degraded
|| check_reshape
) {
463 /* manager will do the actual check */
465 a
->check_degraded
= 1;
467 a
->check_reshape
= 1;
477 static struct mdinfo
*
478 find_device(struct active_array
*a
, int major
, int minor
)
482 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
483 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
489 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
491 struct active_array
*a
;
492 struct mdinfo
*victim
;
494 for (a
= aa
; a
; a
= a
->next
) {
495 if (!a
->container
|| a
->to_remove
)
497 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
501 if (!(victim
->curr_state
& DS_FAULTY
))
502 write_attr("faulty", victim
->state_fd
);
507 static void dprint_wake_reasons(fd_set
*fds
)
515 fprintf(stderr
, "monitor: wake ( ");
516 for (i
= 0; i
< FD_SETSIZE
; i
++) {
517 if (FD_ISSET(i
, fds
)) {
518 sprintf(proc_path
, "/proc/%d/fd/%d",
521 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
523 fprintf(stderr
, "%d:unknown ", i
);
527 basename
= strrchr(link
, '/');
528 fprintf(stderr
, "%d:%s ",
529 i
, basename
? ++basename
: link
);
532 fprintf(stderr
, ")\n");
536 int monitor_loop_cnt
;
538 static int wait_and_act(struct supertype
*container
, int nowait
)
542 struct active_array
**aap
= &container
->arrays
;
543 struct active_array
*a
, **ap
;
546 static unsigned int dirty_arrays
= ~0; /* start at some non-zero value */
550 for (ap
= aap
; *ap
;) {
552 /* once an array has been deactivated we want to
553 * ask the manager to discard it.
555 if (!a
->container
|| a
->to_remove
) {
567 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
568 add_fd(&rfds
, &maxfd
, a
->action_fd
);
569 add_fd(&rfds
, &maxfd
, a
->sync_completed_fd
);
570 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
571 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
576 if (manager_ready
&& (*aap
== NULL
|| (sigterm
&& !dirty_arrays
))) {
577 /* No interesting arrays, or we have been told to
578 * terminate and everything is clean. Lets see about
579 * exiting. Note that blocking at this point is not a
580 * problem as there are no active arrays, there is
581 * nothing that we need to be ready to do.
585 fd
= open_dev_excl(container
->devnm
);
587 fd
= open_dev_flags(container
->devnm
, O_RDONLY
|O_EXCL
);
588 if (fd
>= 0 || errno
!= EBUSY
) {
589 /* OK, we are safe to leave */
590 if (sigterm
&& !dirty_arrays
)
591 dprintf("caught sigterm, all clean... exiting\n");
593 dprintf("no arrays to monitor... exiting\n");
595 /* On SIGTERM, someone (the take-over mdmon) will
598 remove_pidfile(container
->devnm
);
611 if (*aap
== NULL
|| container
->retry_soon
) {
612 /* just waiting to get O_EXCL access */
614 ts
.tv_nsec
= 20000000ULL;
616 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
617 sigdelset(&set
, SIGUSR1
);
618 monitor_loop_cnt
|= 1;
619 rv
= pselect(maxfd
+1, NULL
, NULL
, &rfds
, &ts
, &set
);
620 monitor_loop_cnt
+= 1;
621 if (rv
== -1 && errno
== EINTR
)
624 dprint_wake_reasons(&rfds
);
626 container
->retry_soon
= 0;
630 struct metadata_update
*this;
632 for (this = update_queue
; this ; this = this->next
)
633 container
->ss
->process_update(container
, this);
635 update_queue_handled
= update_queue
;
638 container
->ss
->sync_metadata(container
);
643 for (a
= *aap
; a
; a
= a
->next
) {
645 if (a
->replaces
&& !discard_this
) {
646 struct active_array
**ap
;
647 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
652 discard_this
= a
->replaces
;
654 /* FIXME check if device->state_fd need to be cleared?*/
657 if (a
->container
&& !a
->to_remove
) {
658 int ret
= read_and_act(a
);
660 dirty_arrays
+= !!(ret
& ARRAY_DIRTY
);
661 /* when terminating stop manipulating the array after it
662 * is clean, but make sure read_and_act() is given a
663 * chance to handle 'active_idle'
665 if (sigterm
&& !(ret
& ARRAY_DIRTY
))
666 a
->container
= NULL
; /* stop touching this array */
667 if (ret
& ARRAY_BUSY
)
668 container
->retry_soon
= 1;
672 /* propagate failures across container members */
673 for (a
= *aap
; a
; a
= a
->next
) {
674 if (!a
->container
|| a
->to_remove
)
676 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
677 if (mdi
->curr_state
& DS_FAULTY
)
678 reconcile_failed(*aap
, mdi
);
684 void do_monitor(struct supertype
*container
)
689 rv
= wait_and_act(container
, first
);