]>
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>
26 static char *array_states
[] = {
27 "clear", "inactive", "suspended", "readonly", "read-auto",
28 "clean", "active", "write-pending", "active-idle", "broken", NULL
};
29 static char *sync_actions
[] = {
30 "idle", "reshape", "resync", "recover", "check", "repair", NULL
38 static int write_attr(char *attr
, int fd
)
40 return write(fd
, attr
, strlen(attr
));
43 static void add_fd(fd_set
*fds
, int *maxfd
, int fd
)
48 if (fstat(fd
, &st
) == -1) {
49 dprintf("Invalid fd %d\n", fd
);
52 if (st
.st_nlink
== 0) {
53 dprintf("fd %d was deleted\n", fd
);
61 static int read_attr(char *buf
, int len
, int fd
)
70 n
= read(fd
, buf
, len
- 1);
82 static void read_resync_start(int fd
, unsigned long long *v
)
84 char buf
[SYSFS_MAX_BUF_SIZE
];
87 n
= read_attr(buf
, sizeof(buf
), fd
);
89 dprintf("Failed to read resync_start (%d)\n", fd
);
92 if (str_is_none(buf
) == true)
95 *v
= strtoull(buf
, NULL
, 10);
98 static unsigned long long read_sync_completed(int fd
)
100 unsigned long long val
;
101 char buf
[SYSFS_MAX_BUF_SIZE
];
105 n
= read_attr(buf
, sizeof(buf
), fd
);
110 val
= strtoull(buf
, &ep
, 0);
111 if (ep
== buf
|| (*ep
!= 0 && *ep
!= '\n' && *ep
!= ' '))
116 static enum array_state
read_state(int fd
)
118 char buf
[SYSFS_MAX_BUF_SIZE
];
119 int n
= read_attr(buf
, sizeof(buf
), fd
);
123 return (enum array_state
) sysfs_match_word(buf
, array_states
);
126 static enum sync_action
read_action( int fd
)
128 char buf
[SYSFS_MAX_BUF_SIZE
];
129 int n
= read_attr(buf
, sizeof(buf
), fd
);
133 return (enum sync_action
) sysfs_match_word(buf
, sync_actions
);
136 int read_dev_state(int fd
)
138 char buf
[SYSFS_MAX_BUF_SIZE
];
139 int n
= read_attr(buf
, sizeof(buf
), fd
);
148 if (sysfs_attr_match(cp
, "faulty"))
150 if (sysfs_attr_match(cp
, "in_sync"))
152 if (sysfs_attr_match(cp
, "write_mostly"))
153 rv
|= DS_WRITE_MOSTLY
;
154 if (sysfs_attr_match(cp
, "spare"))
156 if (sysfs_attr_match(cp
, "blocked"))
158 cp
= strchr(cp
, ',');
165 int process_ubb(struct active_array
*a
, struct mdinfo
*mdi
, const unsigned long
166 long sector
, const int length
, const char *buf
,
169 struct superswitch
*ss
= a
->container
->ss
;
172 * record bad block in metadata first, then acknowledge it to the driver
175 if ((ss
->record_bad_block(a
, mdi
->disk
.raid_disk
, sector
, length
)) &&
176 (write(mdi
->bb_fd
, buf
, buf_len
) == buf_len
))
180 * failed to store or acknowledge bad block, switch of bad block support
181 * to get it out of blocked state
183 sysfs_set_str(&a
->info
, mdi
, "state", "-external_bbl");
187 int compare_bb(struct active_array
*a
, struct mdinfo
*mdi
, const unsigned long
188 long sector
, const unsigned int length
, void *arg
)
190 struct superswitch
*ss
= a
->container
->ss
;
191 struct md_bb
*bb
= (struct md_bb
*) arg
;
195 for (i
= 0; i
< bb
->count
; i
++) {
196 unsigned long long start
= bb
->entries
[i
].sector
;
197 unsigned long long len
= bb
->entries
[i
].length
;
200 * bad block in metadata exactly matches bad block in kernel
201 * list, just remove it from a list
203 if ((start
== sector
) && (len
== length
)) {
204 if (i
< bb
->count
- 1)
205 bb
->entries
[i
] = bb
->entries
[bb
->count
- 1];
211 * bad block in metadata spans bad block in kernel list,
212 * clear it and record new bad block
214 if ((sector
>= start
) && (sector
+ length
<= start
+ len
)) {
215 ss
->clear_bad_block(a
, mdi
->disk
.raid_disk
, start
, len
);
220 /* record all bad blocks not in metadata list */
221 if (record
&& (ss
->record_bad_block(a
, mdi
->disk
.raid_disk
, sector
,
223 sysfs_set_str(&a
->info
, mdi
, "state", "-external_bbl");
230 static int read_bb_file(int fd
, struct active_array
*a
, struct mdinfo
*mdi
,
231 enum bb_action action
, void *arg
)
239 int preserve_pos
= (action
== RECORD_BB
? 0 : 1);
241 if (lseek(fd
, 0, SEEK_SET
) == (off_t
) -1)
246 n
= read(fd
, buf
+ pos
, sizeof(buf
) - 1 - pos
);
255 unsigned long long sector
;
262 /* kernel sysfs file format: "sector length\n" */
263 matched
= sscanf(buf
+ off
, "%llu %d%c%n", §or
,
264 &length
, &newline
, &consumed
);
265 if ((matched
!= 3) && (off
> 0)) {
266 /* truncated entry, read again */
268 pos
= sizeof(buf
) - off
- 1;
269 memmove(buf
, buf
+ off
, pos
);
271 if (lseek(fd
, 0, SEEK_SET
) ==
285 if (action
== RECORD_BB
)
286 rc
= process_ubb(a
, mdi
, sector
, length
,
287 buf
+ off
, consumed
);
288 else if (action
== COMPARE_BB
)
289 rc
= compare_bb(a
, mdi
, sector
, length
, arg
);
298 } while (read_again
);
303 static int process_dev_ubb(struct active_array
*a
, struct mdinfo
*mdi
)
305 return read_bb_file(mdi
->ubb_fd
, a
, mdi
, RECORD_BB
, NULL
);
308 static int check_for_cleared_bb(struct active_array
*a
, struct mdinfo
*mdi
)
310 struct superswitch
*ss
= a
->container
->ss
;
314 if (!ss
->get_bad_blocks
)
318 * Get a list of bad blocks for an array, then read list of
319 * acknowledged bad blocks from kernel and compare it against metadata
320 * list, clear all bad blocks remaining in metadata list
322 bb
= ss
->get_bad_blocks(a
, mdi
->disk
.raid_disk
);
326 if (read_bb_file(mdi
->bb_fd
, a
, mdi
, COMPARE_BB
, bb
) < 0)
329 for (i
= 0; i
< bb
->count
; i
++) {
330 unsigned long long sector
= bb
->entries
[i
].sector
;
331 int length
= bb
->entries
[i
].length
;
333 ss
->clear_bad_block(a
, mdi
->disk
.raid_disk
, sector
, length
);
339 static void signal_manager(void)
341 /* tgkill(getpid(), mon_tid, SIGUSR1); */
343 syscall(SYS_tgkill
, pid
, mgr_tid
, SIGUSR1
);
346 /* Monitor a set of active md arrays - all of which share the
347 * same metadata - and respond to events that require
350 * New arrays are detected by another thread which allocates
351 * required memory and attaches the data structure to our list.
355 * This is detected by array_state going to 'clear' or 'inactive'.
356 * while we thought it was active.
357 * Response is to mark metadata as clean and 'clear' the array(??)
359 * array_state if 'write-pending'
360 * We mark metadata as 'dirty' then set array to 'active'.
362 * Either ignore, or mark clean, then mark metadata as clean.
365 * detected by rd-N/state reporting "faulty"
366 * mark device as 'failed' in metadata, let the kernel release the
367 * device by writing '-blocked' to rd/state, and finally write 'remove' to
368 * rd/state. Before a disk can be replaced it must be failed and removed
369 * from all container members, this will be preemptive for the other
373 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
375 * Notify metadata that sync is complete.
378 * sync_action changes from 'recover' to 'idle'
379 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
382 * This only happens on finding a new array... mdadm will have set
383 * 'resync_start' to the correct value. If 'resync_start' indicates that an
384 * resync needs to occur set the array to the 'active' state rather than the
385 * initial read-auto state.
389 * We wait for a change (poll/select) on array_state, sync_action, and
390 * each rd-X/state file.
391 * When we get any change, we check everything. So read each state file,
392 * then decide what to do.
394 * The core action is to write new metadata to all devices in the array.
395 * This is done at most once on any wakeup.
396 * After that we might:
397 * - update the array_state
398 * - set the role of some devices.
399 * - request a sync_action
403 #define ARRAY_DIRTY 1
405 static int read_and_act(struct active_array
*a
, fd_set
*fds
)
407 unsigned long long sync_completed
;
408 int check_degraded
= 0;
409 int check_reshape
= 0;
416 a
->next_state
= bad_word
;
417 a
->next_action
= bad_action
;
419 a
->curr_state
= read_state(a
->info
.state_fd
);
420 a
->curr_action
= read_action(a
->action_fd
);
421 if (a
->curr_state
!= clear
)
423 * In "clear" state, resync_start may wrongly be set to "0"
424 * when the kernel called md_clean but didn't remove the
425 * sysfs attributes yet
427 read_resync_start(a
->resync_start_fd
, &a
->info
.resync_start
);
428 sync_completed
= read_sync_completed(a
->sync_completed_fd
);
429 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
432 if (mdi
->state_fd
>= 0) {
433 read_resync_start(mdi
->recovery_fd
,
434 &mdi
->recovery_start
);
435 mdi
->curr_state
= read_dev_state(mdi
->state_fd
);
438 * If array is blocked and metadata handler is able to handle
439 * BB, check if you can acknowledge them to md driver. If
440 * successful, clear faulty state and unblock the array.
442 if ((mdi
->curr_state
& DS_BLOCKED
) &&
443 a
->container
->ss
->record_bad_block
&&
444 (process_dev_ubb(a
, mdi
) > 0)) {
445 mdi
->next_state
|= DS_UNBLOCK
;
447 if (FD_ISSET(mdi
->bb_fd
, fds
))
448 check_for_cleared_bb(a
, mdi
);
451 gettimeofday(&tv
, NULL
);
452 dprintf("(%d): %ld.%06ld state:%s prev:%s action:%s prev: %s start:%llu\n",
453 a
->info
.container_member
,
454 tv
.tv_sec
, tv
.tv_usec
,
455 array_states
[a
->curr_state
],
456 array_states
[a
->prev_state
],
457 sync_actions
[a
->curr_action
],
458 sync_actions
[a
->prev_action
],
462 if ((a
->curr_state
== bad_word
|| a
->curr_state
<= inactive
) &&
463 a
->prev_state
> inactive
) {
464 /* array has been stopped */
465 a
->container
->ss
->set_array_state(a
, 1);
466 a
->next_state
= clear
;
469 if (a
->curr_state
== write_pending
) {
470 a
->container
->ss
->set_array_state(a
, 0);
471 a
->next_state
= active
;
474 if (a
->curr_state
== active_idle
) {
475 /* Set array to 'clean' FIRST, then mark clean
478 a
->next_state
= clean
;
481 if ((a
->curr_state
== clean
) || (a
->curr_state
== broken
)) {
482 a
->container
->ss
->set_array_state(a
, 1);
484 if (a
->curr_state
== active
||
485 a
->curr_state
== suspended
)
487 if (a
->curr_state
== readonly
) {
488 /* Well, I'm ready to handle things. If readonly
489 * wasn't requested, transition to read-auto.
492 read_attr(buf
, sizeof(buf
), a
->metadata_fd
);
493 if (strncmp(buf
, "external:-", 10) == 0) {
494 /* explicit request for readonly array. Leave it alone */
497 if (a
->container
->ss
->set_array_state(a
, 2))
498 a
->next_state
= read_auto
; /* array is clean */
500 a
->next_state
= active
; /* Now active for recovery etc */
507 a
->curr_action
== idle
&&
508 a
->prev_action
== resync
) {
509 /* A resync has finished. The endpoint is recorded in
510 * 'sync_start'. We don't update the metadata
511 * until the array goes inactive or readonly though.
512 * Just check if we need to fiddle spares.
514 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
519 a
->curr_action
== idle
&&
520 a
->prev_action
== recover
) {
521 /* A recovery has finished. Some disks may be in sync now,
522 * and the array may no longer be degraded
524 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
525 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
527 if (! (mdi
->curr_state
& DS_INSYNC
))
531 if (count
!= a
->info
.array
.raid_disks
)
536 a
->curr_action
== reshape
&&
537 a
->prev_action
!= reshape
)
538 /* reshape was requested by mdadm. Need to see if
539 * new devices have been added. Manager does that
540 * when it sees check_reshape
544 /* Check for failures and if found:
545 * 1/ Record the failure in the metadata and unblock the device.
546 * FIXME update the kernel to stop notifying on failed drives when
547 * the array is readonly and we have cleared 'blocked'
548 * 2/ Try to remove the device if the array is writable, or can be
551 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
552 if (mdi
->curr_state
& DS_FAULTY
) {
553 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
556 if (mdi
->curr_state
& DS_BLOCKED
)
557 mdi
->next_state
|= DS_UNBLOCK
;
558 if (a
->curr_state
== read_auto
) {
559 a
->container
->ss
->set_array_state(a
, 0);
560 a
->next_state
= active
;
562 if (a
->curr_state
> readonly
)
563 mdi
->next_state
|= DS_REMOVE
;
567 /* Check for recovery checkpoint notifications. We need to be a
568 * minimum distance away from the last checkpoint to prevent
569 * over checkpointing. Note reshape checkpointing is handled
570 * in the second branch.
572 if (sync_completed
> a
->last_checkpoint
&&
573 sync_completed
- a
->last_checkpoint
> a
->info
.component_size
>> 4 &&
574 a
->curr_action
> reshape
) {
575 /* A (non-reshape) sync_action has reached a checkpoint.
576 * Record the updated position in the metadata
578 a
->last_checkpoint
= sync_completed
;
579 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
580 } else if ((a
->curr_action
== idle
&& a
->prev_action
== reshape
) ||
581 (a
->curr_action
== reshape
&&
582 sync_completed
> a
->last_checkpoint
)) {
583 /* Reshape has progressed or completed so we need to
584 * update the array state - and possibly the array size
586 if (sync_completed
!= 0)
587 a
->last_checkpoint
= sync_completed
;
588 /* We might need to update last_checkpoint depending on
589 * the reason that reshape finished.
590 * if array reshape is really finished:
591 * set check point to the end, this allows
592 * set_array_state() to finalize reshape in metadata
593 * if reshape if broken: do not set checkpoint to the end
594 * this allows for reshape restart from checkpoint
596 if ((a
->curr_action
!= reshape
) &&
597 (a
->prev_action
== reshape
)) {
598 char buf
[SYSFS_MAX_BUF_SIZE
];
599 if ((sysfs_get_str(&a
->info
, NULL
,
602 sizeof(buf
)) >= 0) &&
603 str_is_none(buf
) == true)
604 a
->last_checkpoint
= a
->info
.component_size
;
606 a
->container
->ss
->set_array_state(a
, a
->curr_state
<= clean
);
607 a
->last_checkpoint
= sync_completed
;
610 if (sync_completed
> a
->last_checkpoint
)
611 a
->last_checkpoint
= sync_completed
;
613 if (sync_completed
>= a
->info
.component_size
)
614 a
->last_checkpoint
= 0;
616 a
->container
->ss
->sync_metadata(a
->container
);
617 dprintf("(%d): state:%s action:%s next(", a
->info
.container_member
,
618 array_states
[a
->curr_state
], sync_actions
[a
->curr_action
]);
620 /* Effect state changes in the array */
621 if (a
->next_state
!= bad_word
) {
622 dprintf_cont(" state:%s", array_states
[a
->next_state
]);
623 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
625 if (a
->next_action
!= bad_action
) {
626 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
627 dprintf_cont(" action:%s", sync_actions
[a
->next_action
]);
629 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
630 if (mdi
->next_state
& DS_UNBLOCK
) {
631 dprintf_cont(" %d:-blocked", mdi
->disk
.raid_disk
);
632 write_attr("-blocked", mdi
->state_fd
);
635 if ((mdi
->next_state
& DS_REMOVE
) && mdi
->state_fd
>= 0) {
638 /* The kernel may not be able to immediately remove the
639 * disk. In that case we wait a little while and
642 remove_result
= write_attr("remove", mdi
->state_fd
);
643 if (remove_result
> 0) {
644 dprintf_cont(" %d:removed", mdi
->disk
.raid_disk
);
645 close(mdi
->state_fd
);
646 close(mdi
->recovery_fd
);
653 if (mdi
->next_state
& DS_INSYNC
) {
654 write_attr("+in_sync", mdi
->state_fd
);
655 dprintf_cont(" %d:+in_sync", mdi
->disk
.raid_disk
);
658 dprintf_cont(" )\n");
660 /* move curr_ to prev_ */
661 a
->prev_state
= a
->curr_state
;
663 a
->prev_action
= a
->curr_action
;
665 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
666 mdi
->prev_state
= mdi
->curr_state
;
670 if (check_degraded
|| check_reshape
) {
671 /* manager will do the actual check */
673 a
->check_degraded
= 1;
675 a
->check_reshape
= 1;
685 static struct mdinfo
*
686 find_device(struct active_array
*a
, int major
, int minor
)
690 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
691 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
697 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
699 struct active_array
*a
;
700 struct mdinfo
*victim
;
702 for (a
= aa
; a
; a
= a
->next
) {
703 if (!a
->container
|| a
->to_remove
)
705 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
709 if (!(victim
->curr_state
& DS_FAULTY
))
710 write_attr("faulty", victim
->state_fd
);
715 static void dprint_wake_reasons(fd_set
*fds
)
723 fprintf(stderr
, "monitor: wake ( ");
724 for (i
= 0; i
< FD_SETSIZE
; i
++) {
725 if (FD_ISSET(i
, fds
)) {
726 sprintf(proc_path
, "/proc/%d/fd/%d",
729 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
731 fprintf(stderr
, "%d:unknown ", i
);
735 basename
= strrchr(link
, '/');
736 fprintf(stderr
, "%d:%s ",
737 i
, basename
? ++basename
: link
);
740 fprintf(stderr
, ")\n");
744 int monitor_loop_cnt
;
746 static int wait_and_act(struct supertype
*container
, int nowait
)
750 struct active_array
**aap
= &container
->arrays
;
751 struct active_array
*a
, **ap
;
754 static unsigned int dirty_arrays
= ~0; /* start at some non-zero value */
758 for (ap
= aap
; *ap
;) {
760 /* once an array has been deactivated we want to
761 * ask the manager to discard it.
763 if (!a
->container
|| a
->to_remove
) {
775 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
776 add_fd(&rfds
, &maxfd
, a
->action_fd
);
777 add_fd(&rfds
, &maxfd
, a
->sync_completed_fd
);
778 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
779 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
780 add_fd(&rfds
, &maxfd
, mdi
->bb_fd
);
781 add_fd(&rfds
, &maxfd
, mdi
->ubb_fd
);
787 if (manager_ready
&& (*aap
== NULL
|| (sigterm
&& !dirty_arrays
))) {
788 /* No interesting arrays, or we have been told to
789 * terminate and everything is clean. Lets see about
790 * exiting. Note that blocking at this point is not a
791 * problem as there are no active arrays, there is
792 * nothing that we need to be ready to do.
796 fd
= open_dev_excl(container
->devnm
);
798 fd
= open_dev_flags(container
->devnm
, O_RDONLY
|O_EXCL
);
799 if (fd
>= 0 || errno
!= EBUSY
) {
800 /* OK, we are safe to leave */
801 if (sigterm
&& !dirty_arrays
)
802 dprintf("caught sigterm, all clean... exiting\n");
804 dprintf("no arrays to monitor... exiting\n");
806 /* On SIGTERM, someone (the take-over mdmon) will
809 remove_pidfile(container
->devnm
);
822 if (*aap
== NULL
|| container
->retry_soon
) {
823 /* just waiting to get O_EXCL access */
825 ts
.tv_nsec
= 20000000ULL;
827 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
828 sigdelset(&set
, SIGUSR1
);
829 monitor_loop_cnt
|= 1;
830 rv
= pselect(maxfd
+1, NULL
, NULL
, &rfds
, &ts
, &set
);
831 monitor_loop_cnt
+= 1;
833 if (errno
== EINTR
) {
836 dprintf("monitor: caught signal\n");
838 dprintf("monitor: error %d in pselect\n",
843 dprint_wake_reasons(&rfds
);
845 container
->retry_soon
= 0;
849 struct metadata_update
*this;
851 for (this = update_queue
; this ; this = this->next
)
852 container
->ss
->process_update(container
, this);
854 update_queue_handled
= update_queue
;
857 container
->ss
->sync_metadata(container
);
862 for (a
= *aap
; a
; a
= a
->next
) {
864 if (a
->replaces
&& !discard_this
) {
865 struct active_array
**ap
;
866 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
871 discard_this
= a
->replaces
;
873 /* FIXME check if device->state_fd need to be cleared?*/
876 if (a
->container
&& !a
->to_remove
) {
877 int ret
= read_and_act(a
, &rfds
);
879 dirty_arrays
+= !!(ret
& ARRAY_DIRTY
);
880 /* when terminating stop manipulating the array after it
881 * is clean, but make sure read_and_act() is given a
882 * chance to handle 'active_idle'
884 if (sigterm
&& !(ret
& ARRAY_DIRTY
))
885 a
->container
= NULL
; /* stop touching this array */
886 if (ret
& ARRAY_BUSY
)
887 container
->retry_soon
= 1;
891 /* propagate failures across container members */
892 for (a
= *aap
; a
; a
= a
->next
) {
893 if (!a
->container
|| a
->to_remove
)
895 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
896 if (mdi
->curr_state
& DS_FAULTY
)
897 reconcile_failed(*aap
, mdi
);
903 void do_monitor(struct supertype
*container
)
908 rv
= wait_and_act(container
, first
);