]>
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;
239 a
->next_state
= bad_word
;
240 a
->next_action
= bad_action
;
242 a
->curr_state
= read_state(a
->info
.state_fd
);
243 a
->curr_action
= read_action(a
->action_fd
);
244 if (a
->curr_state
!= clear
)
246 * In "clear" state, resync_start may wrongly be set to "0"
247 * when the kernel called md_clean but didn't remove the
248 * sysfs attributes yet
250 read_resync_start(a
->resync_start_fd
, &a
->info
.resync_start
);
251 sync_completed
= read_sync_completed(a
->sync_completed_fd
);
252 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
255 if (mdi
->state_fd
>= 0) {
256 read_resync_start(mdi
->recovery_fd
,
257 &mdi
->recovery_start
);
258 mdi
->curr_state
= read_dev_state(mdi
->state_fd
);
262 gettimeofday(&tv
, NULL
);
263 dprintf("%s(%d): %ld.%06ld state:%s prev:%s action:%s prev: %s start:%llu\n",
264 __func__
, a
->info
.container_member
,
265 tv
.tv_sec
, tv
.tv_usec
,
266 array_states
[a
->curr_state
],
267 array_states
[a
->prev_state
],
268 sync_actions
[a
->curr_action
],
269 sync_actions
[a
->prev_action
],
273 if (a
->curr_state
> inactive
&&
274 a
->prev_state
== inactive
) {
275 /* array has been started
276 * possible that container operation has to be completed
278 a
->container
->ss
->set_array_state(a
, 0);
280 if ((a
->curr_state
== bad_word
|| a
->curr_state
<= inactive
) &&
281 a
->prev_state
> inactive
) {
282 /* array has been stopped */
283 a
->container
->ss
->set_array_state(a
, 1);
284 a
->next_state
= clear
;
287 if (a
->curr_state
== write_pending
) {
288 a
->container
->ss
->set_array_state(a
, 0);
289 a
->next_state
= active
;
292 if (a
->curr_state
== active_idle
) {
293 /* Set array to 'clean' FIRST, then mark clean
296 a
->next_state
= clean
;
299 if (a
->curr_state
== clean
) {
300 a
->container
->ss
->set_array_state(a
, 1);
302 if (a
->curr_state
== active
||
303 a
->curr_state
== suspended
)
305 if (a
->curr_state
== readonly
) {
306 /* Well, I'm ready to handle things. If readonly
307 * wasn't requested, transition to read-auto.
310 read_attr(buf
, sizeof(buf
), a
->metadata_fd
);
311 if (strncmp(buf
, "external:-", 10) == 0) {
312 /* explicit request for readonly array. Leave it alone */
315 if (a
->container
->ss
->set_array_state(a
, 2))
316 a
->next_state
= read_auto
; /* array is clean */
318 a
->next_state
= active
; /* Now active for recovery etc */
325 a
->curr_action
== idle
&&
326 a
->prev_action
== resync
) {
327 /* A resync has finished. The endpoint is recorded in
328 * 'sync_start'. We don't update the metadata
329 * until the array goes inactive or readonly though.
330 * Just check if we need to fiddle spares.
332 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
337 a
->curr_action
== idle
&&
338 a
->prev_action
== recover
) {
339 /* A recovery has finished. Some disks may be in sync now,
340 * and the array may no longer be degraded
342 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
343 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
345 if (! (mdi
->curr_state
& DS_INSYNC
))
349 if (count
!= a
->info
.array
.raid_disks
)
354 a
->curr_action
== reshape
&&
355 a
->prev_action
!= reshape
)
356 /* reshape was requested by mdadm. Need to see if
357 * new devices have been added. Manager does that
358 * when it sees check_reshape
362 /* Check for failures and if found:
363 * 1/ Record the failure in the metadata and unblock the device.
364 * FIXME update the kernel to stop notifying on failed drives when
365 * the array is readonly and we have cleared 'blocked'
366 * 2/ Try to remove the device if the array is writable, or can be
369 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
370 if (mdi
->curr_state
& DS_FAULTY
) {
371 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
374 if (mdi
->curr_state
& DS_BLOCKED
)
375 mdi
->next_state
|= DS_UNBLOCK
;
376 if (a
->curr_state
== read_auto
) {
377 a
->container
->ss
->set_array_state(a
, 0);
378 a
->next_state
= active
;
380 if (a
->curr_state
> readonly
)
381 mdi
->next_state
|= DS_REMOVE
;
385 /* Check for recovery checkpoint notifications. We need to be a
386 * minimum distance away from the last checkpoint to prevent
387 * over checkpointing. Note reshape checkpointing is handled
388 * in the second branch.
390 if (sync_completed
> a
->last_checkpoint
&&
391 sync_completed
- a
->last_checkpoint
> a
->info
.component_size
>> 4 &&
392 a
->curr_action
> reshape
) {
393 /* A (non-reshape) sync_action has reached a checkpoint.
394 * Record the updated position in the metadata
396 a
->last_checkpoint
= sync_completed
;
397 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
398 } else if ((a
->curr_action
== idle
&& a
->prev_action
== reshape
) ||
399 (a
->curr_action
== reshape
400 && sync_completed
> a
->last_checkpoint
) ) {
401 /* Reshape has progressed or completed so we need to
402 * update the array state - and possibly the array size
404 if (sync_completed
!= 0)
405 a
->last_checkpoint
= sync_completed
;
406 /* We might need to update last_checkpoint depending on
407 * the reason that reshape finished.
408 * if array reshape is really finished:
409 * set check point to the end, this allows
410 * set_array_state() to finalize reshape in metadata
411 * if reshape if broken: do not set checkpoint to the end
412 * this allows for reshape restart from checkpoint
414 if ((a
->curr_action
!= reshape
) &&
415 (a
->prev_action
== reshape
)) {
417 if ((sysfs_get_str(&a
->info
, NULL
,
420 sizeof(buf
)) >= 0) &&
421 strncmp(buf
, "none", 4) == 0)
422 a
->last_checkpoint
= a
->info
.component_size
;
424 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
425 a
->last_checkpoint
= sync_completed
;
428 if (sync_completed
> a
->last_checkpoint
)
429 a
->last_checkpoint
= sync_completed
;
431 if (deactivate
|| a
->curr_state
>= clean
)
432 a
->container
->ss
->sync_metadata(a
->container
);
433 dprintf("%s(%d): state:%s action:%s next(", __func__
, a
->info
.container_member
,
434 array_states
[a
->curr_state
], sync_actions
[a
->curr_action
]);
436 /* Effect state changes in the array */
437 if (a
->next_state
!= bad_word
) {
438 dprintf(" state:%s", array_states
[a
->next_state
]);
439 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
441 if (a
->next_action
!= bad_action
) {
442 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
443 dprintf(" action:%s", sync_actions
[a
->next_action
]);
445 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
446 if (mdi
->next_state
& DS_UNBLOCK
) {
447 dprintf(" %d:-blocked", mdi
->disk
.raid_disk
);
448 write_attr("-blocked", mdi
->state_fd
);
451 if ((mdi
->next_state
& DS_REMOVE
) && mdi
->state_fd
>= 0) {
454 /* The kernel may not be able to immediately remove the
455 * disk. In that case we wait a little while and
458 remove_result
= write_attr("remove", mdi
->state_fd
);
459 if (remove_result
> 0) {
460 dprintf(" %d:removed", mdi
->disk
.raid_disk
);
461 close(mdi
->state_fd
);
462 close(mdi
->recovery_fd
);
467 if (mdi
->next_state
& DS_INSYNC
) {
468 write_attr("+in_sync", mdi
->state_fd
);
469 dprintf(" %d:+in_sync", mdi
->disk
.raid_disk
);
474 /* move curr_ to prev_ */
475 a
->prev_state
= a
->curr_state
;
477 a
->prev_action
= a
->curr_action
;
479 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
480 mdi
->prev_state
= mdi
->curr_state
;
484 if (check_degraded
|| check_reshape
) {
485 /* manager will do the actual check */
487 a
->check_degraded
= 1;
489 a
->check_reshape
= 1;
499 static struct mdinfo
*
500 find_device(struct active_array
*a
, int major
, int minor
)
504 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
505 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
511 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
513 struct active_array
*a
;
514 struct mdinfo
*victim
;
516 for (a
= aa
; a
; a
= a
->next
) {
517 if (!a
->container
|| a
->to_remove
)
519 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
523 if (!(victim
->curr_state
& DS_FAULTY
))
524 write_attr("faulty", victim
->state_fd
);
529 static void dprint_wake_reasons(fd_set
*fds
)
537 fprintf(stderr
, "monitor: wake ( ");
538 for (i
= 0; i
< FD_SETSIZE
; i
++) {
539 if (FD_ISSET(i
, fds
)) {
540 sprintf(proc_path
, "/proc/%d/fd/%d",
543 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
545 fprintf(stderr
, "%d:unknown ", i
);
549 basename
= strrchr(link
, '/');
550 fprintf(stderr
, "%d:%s ",
551 i
, basename
? ++basename
: link
);
554 fprintf(stderr
, ")\n");
558 int monitor_loop_cnt
;
560 static int wait_and_act(struct supertype
*container
, int nowait
)
564 struct active_array
**aap
= &container
->arrays
;
565 struct active_array
*a
, **ap
;
568 static unsigned int dirty_arrays
= ~0; /* start at some non-zero value */
572 for (ap
= aap
; *ap
;) {
574 /* once an array has been deactivated we want to
575 * ask the manager to discard it.
577 if (!a
->container
|| a
->to_remove
) {
589 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
590 add_fd(&rfds
, &maxfd
, a
->action_fd
);
591 add_fd(&rfds
, &maxfd
, a
->sync_completed_fd
);
592 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
593 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
598 if (manager_ready
&& (*aap
== NULL
|| (sigterm
&& !dirty_arrays
))) {
599 /* No interesting arrays, or we have been told to
600 * terminate and everything is clean. Lets see about
601 * exiting. Note that blocking at this point is not a
602 * problem as there are no active arrays, there is
603 * nothing that we need to be ready to do.
607 fd
= open_dev_excl(container
->devnm
);
609 fd
= open_dev_flags(container
->devnm
, O_RDONLY
|O_EXCL
);
610 if (fd
>= 0 || errno
!= EBUSY
) {
611 /* OK, we are safe to leave */
612 if (sigterm
&& !dirty_arrays
)
613 dprintf("caught sigterm, all clean... exiting\n");
615 dprintf("no arrays to monitor... exiting\n");
617 /* On SIGTERM, someone (the take-over mdmon) will
620 remove_pidfile(container
->devnm
);
633 if (*aap
== NULL
|| container
->retry_soon
) {
634 /* just waiting to get O_EXCL access */
636 ts
.tv_nsec
= 20000000ULL;
638 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
639 sigdelset(&set
, SIGUSR1
);
640 monitor_loop_cnt
|= 1;
641 rv
= pselect(maxfd
+1, NULL
, NULL
, &rfds
, &ts
, &set
);
642 monitor_loop_cnt
+= 1;
644 if (errno
== EINTR
) {
646 dprintf("monitor: caught signal\n");
648 dprintf("monitor: error %d in pselect\n",
653 dprint_wake_reasons(&rfds
);
655 container
->retry_soon
= 0;
659 struct metadata_update
*this;
661 for (this = update_queue
; this ; this = this->next
)
662 container
->ss
->process_update(container
, this);
664 update_queue_handled
= update_queue
;
667 container
->ss
->sync_metadata(container
);
672 for (a
= *aap
; a
; a
= a
->next
) {
674 if (a
->replaces
&& !discard_this
) {
675 struct active_array
**ap
;
676 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
681 discard_this
= a
->replaces
;
683 /* FIXME check if device->state_fd need to be cleared?*/
686 if (a
->container
&& !a
->to_remove
) {
687 int ret
= read_and_act(a
);
689 dirty_arrays
+= !!(ret
& ARRAY_DIRTY
);
690 /* when terminating stop manipulating the array after it
691 * is clean, but make sure read_and_act() is given a
692 * chance to handle 'active_idle'
694 if (sigterm
&& !(ret
& ARRAY_DIRTY
))
695 a
->container
= NULL
; /* stop touching this array */
696 if (ret
& ARRAY_BUSY
)
697 container
->retry_soon
= 1;
701 /* propagate failures across container members */
702 for (a
= *aap
; a
; a
= a
->next
) {
703 if (!a
->container
|| a
->to_remove
)
705 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
706 if (mdi
->curr_state
& DS_FAULTY
)
707 reconcile_failed(*aap
, mdi
);
713 void do_monitor(struct supertype
*container
)
718 rv
= wait_and_act(container
, first
);