]>
git.ipfire.org Git - thirdparty/mdadm.git/blob - monitor.c
4 #include <sys/syscall.h>
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 /* tgkill(getpid(), mon_tid, SIGUSR1); */
143 syscall(SYS_tgkill
, pid
, mgr_tid
, SIGUSR1
);
146 /* Monitor a set of active md arrays - all of which share the
147 * same metadata - and respond to events that require
150 * New arrays are detected by another thread which allocates
151 * required memory and attaches the data structure to our list.
155 * This is detected by array_state going to 'clear' or 'inactive'.
156 * while we thought it was active.
157 * Response is to mark metadata as clean and 'clear' the array(??)
159 * array_state if 'write-pending'
160 * We mark metadata as 'dirty' then set array to 'active'.
162 * Either ignore, or mark clean, then mark metadata as clean.
165 * detected by rd-N/state reporting "faulty"
166 * mark device as 'failed' in metadata, let the kernel release the
167 * device by writing '-blocked' to rd/state, and finally write 'remove' to
168 * rd/state. Before a disk can be replaced it must be failed and removed
169 * from all container members, this will be preemptive for the other
173 * sync_action was 'resync' and becomes 'idle' and resync_start becomes
175 * Notify metadata that sync is complete.
178 * sync_action changes from 'recover' to 'idle'
179 * Check each device state and mark metadata if 'faulty' or 'in_sync'.
182 * This only happens on finding a new array... mdadm will have set
183 * 'resync_start' to the correct value. If 'resync_start' indicates that an
184 * resync needs to occur set the array to the 'active' state rather than the
185 * initial read-auto state.
189 * We wait for a change (poll/select) on array_state, sync_action, and
190 * each rd-X/state file.
191 * When we get any change, we check everything. So read each state file,
192 * then decide what to do.
194 * The core action is to write new metadata to all devices in the array.
195 * This is done at most once on any wakeup.
196 * After that we might:
197 * - update the array_state
198 * - set the role of some devices.
199 * - request a sync_action
203 static int read_and_act(struct active_array
*a
)
205 int check_degraded
= 0;
209 a
->next_state
= bad_word
;
210 a
->next_action
= bad_action
;
212 a
->curr_state
= read_state(a
->info
.state_fd
);
213 a
->curr_action
= read_action(a
->action_fd
);
214 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
216 if (mdi
->state_fd
>= 0)
217 mdi
->curr_state
= read_dev_state(mdi
->state_fd
);
220 if (a
->curr_state
<= inactive
&&
221 a
->prev_state
> inactive
) {
222 /* array has been stopped */
223 a
->container
->ss
->set_array_state(a
, 1);
224 a
->next_state
= clear
;
227 if (a
->curr_state
== write_pending
) {
229 a
->container
->ss
->set_array_state(a
, 0);
230 a
->next_state
= active
;
232 if (a
->curr_state
== active_idle
) {
233 /* Set array to 'clean' FIRST, then
234 * a->ss->mark_clean(a, ~0ULL);
235 * just ignore for now.
239 if (a
->curr_state
== readonly
) {
240 /* Well, I'm ready to handle things, so
241 * read-auto is OK. FIXME what if we really want
245 // printf("Found a readonly array at %llu\n", a->resync_start);
246 if (a
->resync_start
== ~0ULL)
247 a
->next_state
= read_auto
; /* array is clean */
249 a
->container
->ss
->set_array_state(a
, 0);
250 a
->next_state
= active
;
255 a
->curr_action
== idle
&&
256 a
->prev_action
== resync
) {
257 /* A resync has finished. The endpoint is recorded in
258 * 'sync_start'. We don't update the metadata
259 * until the array goes inactive or readonly though.
260 * Just check if we need to fiddle spares.
263 a
->container
->ss
->set_array_state(a
, 0);
268 a
->curr_action
== idle
&&
269 a
->prev_action
== recover
) {
270 /* A recovery has finished. Some disks may be in sync now,
271 * and the array may no longer be degraded
273 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
274 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
276 if (! (mdi
->curr_state
& DS_INSYNC
))
279 a
->container
->ss
->set_array_state(a
, 0);
282 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
283 if (mdi
->curr_state
& DS_FAULTY
) {
284 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
287 mdi
->next_state
= DS_REMOVE
;
291 a
->container
->ss
->sync_metadata(a
->container
);
292 dprintf("%s: update[%d]: (", __func__
, a
->info
.container_member
);
294 /* Effect state changes in the array */
295 if (a
->next_state
!= bad_word
) {
296 dprintf(" state:%s", array_states
[a
->next_state
]);
297 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
299 if (a
->next_action
!= bad_action
) {
300 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
301 dprintf(" action:%s", array_states
[a
->next_state
]);
303 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
304 if (mdi
->next_state
== DS_REMOVE
&& mdi
->state_fd
>= 0) {
307 write_attr("-blocked", mdi
->state_fd
);
308 /* the kernel may not be able to immediately remove the
309 * disk, we can simply wait until the next event to try
312 dprintf(" %d:-blocked", mdi
->disk
.raid_disk
);
313 remove_result
= write_attr("remove", mdi
->state_fd
);
314 if (remove_result
> 0) {
315 dprintf(" %d:removed", mdi
->disk
.raid_disk
);
316 close(mdi
->state_fd
);
320 if (mdi
->next_state
& DS_INSYNC
) {
321 write_attr("+in_sync", mdi
->state_fd
);
322 dprintf(" %d:+in_sync", mdi
->disk
.raid_disk
);
327 /* move curr_ to prev_ */
328 a
->prev_state
= a
->curr_state
;
330 a
->prev_action
= a
->curr_action
;
332 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
333 mdi
->prev_state
= mdi
->curr_state
;
337 if (check_degraded
) {
338 /* manager will do the actual check */
339 a
->check_degraded
= 1;
349 static struct mdinfo
*
350 find_device(struct active_array
*a
, int major
, int minor
)
354 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
355 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
361 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
363 struct active_array
*a
;
364 struct mdinfo
*victim
;
366 for (a
= aa
; a
; a
= a
->next
) {
369 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
373 if (!(victim
->curr_state
& DS_FAULTY
))
374 write_attr("faulty", victim
->state_fd
);
379 static void dprint_wake_reasons(fd_set
*fds
)
387 fprintf(stderr
, "monitor: wake ( ");
388 for (i
= 0; i
< FD_SETSIZE
; i
++) {
389 if (FD_ISSET(i
, fds
)) {
390 sprintf(proc_path
, "/proc/%d/fd/%d",
393 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
395 fprintf(stderr
, "%d:unknown ", i
);
399 basename
= strrchr(link
, '/');
400 fprintf(stderr
, "%d:%s ",
401 i
, basename
? ++basename
: link
);
404 fprintf(stderr
, ")\n");
408 static int wait_and_act(struct supertype
*container
, int nowait
)
412 struct active_array
**aap
= &container
->arrays
;
413 struct active_array
*a
, **ap
;
419 for (ap
= aap
; *ap
;) {
421 /* once an array has been deactivated we want to
422 * ask the manager to discard it.
436 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
437 add_fd(&rfds
, &maxfd
, a
->action_fd
);
438 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
439 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
444 if (manager_ready
&& *aap
== NULL
) {
445 /* No interesting arrays. Lets see about exiting.
446 * Note that blocking at this point is not a problem
447 * as there are no active arrays, there is nothing that
448 * we need to be ready to do.
450 int fd
= open(container
->device_name
, O_RDONLY
|O_EXCL
);
451 if (fd
>= 0 || errno
!= EBUSY
) {
452 /* OK, we are safe to leave */
453 dprintf("no arrays to monitor... exiting\n");
454 remove_pidfile(container
->devname
);
463 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
464 sigdelset(&set
, SIGUSR1
);
465 rv
= pselect(maxfd
+1, &rfds
, NULL
, NULL
, NULL
, &set
);
466 if (rv
== -1 && errno
== EINTR
)
469 dprint_wake_reasons(&rfds
);
475 struct metadata_update
*this;
477 for (this = update_queue
; this ; this = this->next
)
478 container
->ss
->process_update(container
, this);
480 update_queue_handled
= update_queue
;
483 container
->ss
->sync_metadata(container
);
486 for (a
= *aap
; a
; a
= a
->next
) {
487 if (a
->replaces
&& !discard_this
) {
488 struct active_array
**ap
;
489 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
494 discard_this
= a
->replaces
;
496 /* FIXME check if device->state_fd need to be cleared?*/
500 rv
+= read_and_act(a
);
503 /* propagate failures across container members */
504 for (a
= *aap
; a
; a
= a
->next
) {
507 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
508 if (mdi
->curr_state
& DS_FAULTY
)
509 reconcile_failed(*aap
, mdi
);
515 void do_monitor(struct supertype
*container
)
520 rv
= wait_and_act(container
, first
);