]>
git.ipfire.org Git - thirdparty/mdadm.git/blob - monitor.c
5 #include <sys/select.h>
8 static char *array_states
[] = {
9 "clear", "inactive", "suspended", "readonly", "read-auto",
10 "clean", "active", "write-pending", "active-idle", NULL
};
11 static char *sync_actions
[] = {
12 "idle", "reshape", "resync", "recover", "check", "repair", NULL
15 static int write_attr(char *attr
, int fd
)
17 return write(fd
, attr
, strlen(attr
));
20 static void add_fd(fd_set
*fds
, int *maxfd
, int fd
)
29 static int read_attr(char *buf
, int len
, int fd
)
38 n
= read(fd
, buf
, len
- 1);
51 static int get_resync_start(struct active_array
*a
)
56 n
= read_attr(buf
, 30, a
->resync_start_fd
);
60 a
->resync_start
= strtoull(buf
, NULL
, 10);
65 static int attr_match(const char *attr
, const char *str
)
67 /* See if attr, read from a sysfs file, matches
68 * str. They must either be the same, or attr can
69 * have a trailing newline or comma
71 while (*attr
&& *str
&& *attr
== *str
) {
76 if (*str
|| (*attr
&& *attr
!= ',' && *attr
!= '\n'))
81 static int match_word(const char *word
, char **list
)
84 for (n
=0; list
[n
]; n
++)
85 if (attr_match(word
, list
[n
]))
90 static enum array_state
read_state(int fd
)
93 int n
= read_attr(buf
, 20, fd
);
97 return (enum array_state
) match_word(buf
, array_states
);
100 static enum sync_action
read_action( int fd
)
103 int n
= read_attr(buf
, 20, fd
);
107 return (enum sync_action
) match_word(buf
, sync_actions
);
110 int read_dev_state(int fd
)
113 int n
= read_attr(buf
, 60, fd
);
122 if (attr_match(cp
, "faulty"))
124 if (attr_match(cp
, "in_sync"))
126 if (attr_match(cp
, "write_mostly"))
127 rv
|= DS_WRITE_MOSTLY
;
128 if (attr_match(cp
, "spare"))
130 if (attr_match(cp
, "blocked"))
132 cp
= strchr(cp
, ',');
139 static void signal_manager(void)
141 kill(getpid(), SIGUSR1
);
144 /* Monitor a set of active md arrays - all of which share the
145 * same metadata - and respond to events that require
148 * New arrays are detected by another thread which allocates
149 * required memory and attaches the data structure to our list.
153 * This is detected by array_state going to 'clear' or 'inactive'.
154 * while we thought it was active.
155 * Response is to mark metadata as clean and 'clear' the array(??)
157 * array_state if 'write-pending'
158 * We mark metadata as 'dirty' then set array to 'active'.
160 * Either ignore, or mark clean, then mark metadata as clean.
163 * detected by rd-N/state reporting "faulty"
164 * mark device as 'failed' in metadata, let the kernel release the
165 * device by writing '-blocked' to rd/state, and finally write 'remove' to
166 * rd/state. Before a disk can be replaced it must be failed and removed
167 * from all container members, this will be preemptive for the other
171 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
173 * Notify metadata that sync is complete.
176 * sync_action changes from 'recover' to 'idle'
177 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
180 * This only happens on finding a new array... mdadm will have set
181 * 'resync_start' to the correct value. If 'resync_start' indicates that an
182 * resync needs to occur set the array to the 'active' state rather than the
183 * initial read-auto state.
187 * We wait for a change (poll/select) on array_state, sync_action, and
188 * each rd-X/state file.
189 * When we get any change, we check everything. So read each state file,
190 * then decide what to do.
192 * The core action is to write new metadata to all devices in the array.
193 * This is done at most once on any wakeup.
194 * After that we might:
195 * - update the array_state
196 * - set the role of some devices.
197 * - request a sync_action
201 static int read_and_act(struct active_array
*a
)
203 int check_degraded
= 0;
207 a
->next_state
= bad_word
;
208 a
->next_action
= bad_action
;
210 a
->curr_state
= read_state(a
->info
.state_fd
);
211 a
->curr_action
= read_action(a
->action_fd
);
212 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
214 if (mdi
->state_fd
>= 0)
215 mdi
->curr_state
= read_dev_state(mdi
->state_fd
);
218 if (a
->curr_state
<= inactive
&&
219 a
->prev_state
> inactive
) {
220 /* array has been stopped */
221 a
->container
->ss
->set_array_state(a
, 1);
222 a
->next_state
= clear
;
225 if (a
->curr_state
== write_pending
) {
227 a
->container
->ss
->set_array_state(a
, 0);
228 a
->next_state
= active
;
230 if (a
->curr_state
== active_idle
) {
231 /* Set array to 'clean' FIRST, then
232 * a->ss->mark_clean(a, ~0ULL);
233 * just ignore for now.
237 if (a
->curr_state
== readonly
) {
238 /* Well, I'm ready to handle things, so
239 * read-auto is OK. FIXME what if we really want
243 printf("Found a readonly array at %llu\n", a
->resync_start
);
244 if (a
->resync_start
== ~0ULL)
245 a
->next_state
= read_auto
; /* array is clean */
247 a
->container
->ss
->set_array_state(a
, 0);
248 a
->next_state
= active
;
252 if (a
->curr_action
== idle
&&
253 a
->prev_action
== resync
) {
254 /* A resync has finished. The endpoint is recorded in
255 * 'sync_start'. We don't update the metadata
256 * until the array goes inactive or readonly though.
257 * Just check if we need to fiddle spares.
260 a
->container
->ss
->set_array_state(a
, 0);
264 if (a
->curr_action
== idle
&&
265 a
->prev_action
== recover
) {
266 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
267 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
269 if (! (mdi
->curr_state
& DS_INSYNC
))
274 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
275 if (mdi
->curr_state
& DS_FAULTY
) {
276 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
279 mdi
->next_state
= DS_REMOVE
;
283 a
->container
->ss
->sync_metadata(a
->container
);
284 dprintf("%s: update[%d]: (", __func__
, a
->info
.container_member
);
286 /* Effect state changes in the array */
287 if (a
->next_state
!= bad_word
) {
288 dprintf(" state:%s", array_states
[a
->next_state
]);
289 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
291 if (a
->next_action
!= bad_action
) {
292 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
293 dprintf(" action:%s", array_states
[a
->next_state
]);
295 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
296 if (mdi
->next_state
== DS_REMOVE
&& mdi
->state_fd
>= 0) {
299 write_attr("-blocked", mdi
->state_fd
);
300 /* the kernel may not be able to immediately remove the
301 * disk, we can simply wait until the next event to try
304 dprintf(" %d:-blocked", mdi
->disk
.raid_disk
);
305 remove_result
= write_attr("remove", mdi
->state_fd
);
306 if (remove_result
> 0) {
307 dprintf(" %d:removed", mdi
->disk
.raid_disk
);
308 close(mdi
->state_fd
);
312 if (mdi
->next_state
& DS_INSYNC
) {
313 write_attr("+in_sync", mdi
->state_fd
);
314 dprintf(" %d:+in_sync", mdi
->disk
.raid_disk
);
319 /* move curr_ to prev_ */
320 a
->prev_state
= a
->curr_state
;
322 a
->prev_action
= a
->curr_action
;
324 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
325 mdi
->prev_state
= mdi
->curr_state
;
329 if (check_degraded
) {
330 /* manager will do the actual check */
331 a
->check_degraded
= 1;
341 static struct mdinfo
*
342 find_device(struct active_array
*a
, int major
, int minor
)
346 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
347 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
353 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
355 struct active_array
*a
;
356 struct mdinfo
*victim
;
358 for (a
= aa
; a
; a
= a
->next
) {
361 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
365 if (!(victim
->curr_state
& DS_FAULTY
))
366 write_attr("faulty", victim
->state_fd
);
370 static int handle_remove_device(struct md_remove_device_cmd
*cmd
, struct active_array
*aa
)
372 struct active_array
*a
;
373 struct mdinfo
*victim
;
376 /* scan all arrays for the given device, if ->state_fd is closed (-1)
377 * in all cases then mark the disk as removed in the metadata.
378 * Otherwise reply that it is busy.
381 /* pass1 check that it is not in use anywhere */
382 /* note: we are safe from re-adds as long as the device exists in the
385 for (a
= aa
; a
; a
= a
->next
) {
388 victim
= find_device(a
, major(cmd
->rdev
), minor(cmd
->rdev
));
391 if (victim
->state_fd
> 0)
395 /* pass2 schedule and process removal per array */
396 for (a
= aa
; a
; a
= a
->next
) {
399 victim
= find_device(a
, major(cmd
->rdev
), minor(cmd
->rdev
));
402 victim
->curr_state
|= DS_REMOVE
;
403 rv
= read_and_act(a
);
411 static int handle_pipe(struct md_generic_cmd
*cmd
, struct active_array
*aa
)
413 switch (cmd
->action
) {
414 case md_action_ping_monitor
:
416 case md_action_remove_device
:
417 return handle_remove_device((void *) cmd
, aa
);
424 static void dprint_wake_reasons(fd_set
*fds
)
432 fprintf(stderr
, "monitor: wake ( ");
433 for (i
= 0; i
< FD_SETSIZE
; i
++) {
434 if (FD_ISSET(i
, fds
)) {
435 sprintf(proc_path
, "/proc/%d/fd/%d",
438 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
440 fprintf(stderr
, "%d:unknown ", i
);
444 basename
= strrchr(link
, '/');
445 fprintf(stderr
, "%d:%s ",
446 i
, basename
? ++basename
: link
);
449 fprintf(stderr
, ")\n");
453 static int wait_and_act(struct supertype
*container
, int pfd
,
454 int monfd
, int nowait
)
458 struct active_array
**aap
= &container
->arrays
;
459 struct active_array
*a
, **ap
;
465 add_fd(&rfds
, &maxfd
, pfd
);
466 for (ap
= aap
; *ap
;) {
468 /* once an array has been deactivated we want to
469 * ask the manager to discard it.
483 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
484 add_fd(&rfds
, &maxfd
, a
->action_fd
);
485 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
486 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
491 if (manager_ready
&& *aap
== NULL
) {
492 /* No interesting arrays. Lets see about exiting.
493 * Note that blocking at this point is not a problem
494 * as there are no active arrays, there is nothing that
495 * we need to be ready to do.
497 int fd
= open(container
->device_name
, O_RDONLY
|O_EXCL
);
498 if (fd
>= 0 || errno
!= EBUSY
) {
499 /* OK, we are safe to leave */
500 dprintf("no arrays to monitor... exiting\n");
503 remove_pidfile(container
->devname
);
509 rv
= select(maxfd
+1, &rfds
, NULL
, NULL
, NULL
);
512 dprint_wake_reasons(&rfds
);
518 if (FD_ISSET(pfd
, &rfds
)) {
521 if (read(pfd
, &err
, 1) > 0)
522 err
= handle_pipe(active_cmd
, *aap
);
523 write(monfd
, &err
, 1);
528 struct metadata_update
*this;
530 for (this = update_queue
; this ; this = this->next
)
531 container
->ss
->process_update(container
, this);
533 update_queue_handled
= update_queue
;
536 container
->ss
->sync_metadata(container
);
539 for (a
= *aap
; a
; a
= a
->next
) {
540 if (a
->replaces
&& !discard_this
) {
541 struct active_array
**ap
;
542 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
547 discard_this
= a
->replaces
;
549 /* FIXME check if device->state_fd need to be cleared?*/
553 rv
+= read_and_act(a
);
556 /* propagate failures across container members */
557 for (a
= *aap
; a
; a
= a
->next
) {
560 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
561 if (mdi
->curr_state
& DS_FAULTY
)
562 reconcile_failed(*aap
, mdi
);
568 void do_monitor(struct supertype
*container
)
573 rv
= wait_and_act(container
, container
->mgr_pipe
[0],
574 container
->mon_pipe
[1], first
);