]>
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 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 */
224 a
->container
->ss
->set_array_state(a
, 1);
225 a
->next_state
= clear
;
228 if (a
->curr_state
== write_pending
) {
230 a
->container
->ss
->set_array_state(a
, 0);
231 a
->next_state
= active
;
233 if (a
->curr_state
== active_idle
) {
234 /* Set array to 'clean' FIRST, then
235 * a->ss->mark_clean(a, ~0ULL);
236 * just ignore for now.
240 if (a
->curr_state
== readonly
) {
241 /* Well, I'm ready to handle things. If readonly
242 * wasn't requested, transition to read-auto.
245 read_attr(buf
, sizeof(buf
), a
->metadata_fd
);
246 if (strncmp(buf
, "external:-", 10) == 0) {
247 /* explicit request for readonly array. Leave it alone */
251 if (a
->container
->ss
->set_array_state(a
, 2))
252 a
->next_state
= read_auto
; /* array is clean */
254 a
->next_state
= active
; /* Now active for recovery etc */
259 a
->curr_action
== idle
&&
260 a
->prev_action
== resync
) {
261 /* A resync has finished. The endpoint is recorded in
262 * 'sync_start'. We don't update the metadata
263 * until the array goes inactive or readonly though.
264 * Just check if we need to fiddle spares.
267 a
->container
->ss
->set_array_state(a
, 0);
272 a
->curr_action
== idle
&&
273 a
->prev_action
== recover
) {
274 /* A recovery has finished. Some disks may be in sync now,
275 * and the array may no longer be degraded
277 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
278 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
280 if (! (mdi
->curr_state
& DS_INSYNC
))
285 /* Check for failures and if found:
286 * 1/ Record the failure in the metadata and unblock the device.
287 * FIXME update the kernel to stop notifying on failed drives when
288 * the array is readonly and we have cleared 'blocked'
289 * 2/ Try to remove the device if the array is writable, or can be
292 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
293 if (mdi
->curr_state
& DS_FAULTY
) {
294 a
->container
->ss
->set_disk(a
, mdi
->disk
.raid_disk
,
297 mdi
->next_state
|= DS_UNBLOCK
;
298 if (a
->curr_state
== read_auto
) {
299 a
->container
->ss
->set_array_state(a
, 0);
300 a
->next_state
= active
;
302 if (a
->curr_state
> readonly
)
303 mdi
->next_state
|= DS_REMOVE
;
307 a
->container
->ss
->sync_metadata(a
->container
);
308 dprintf("%s: update[%d]: (", __func__
, a
->info
.container_member
);
310 /* Effect state changes in the array */
311 if (a
->next_state
!= bad_word
) {
312 dprintf(" state:%s", array_states
[a
->next_state
]);
313 write_attr(array_states
[a
->next_state
], a
->info
.state_fd
);
315 if (a
->next_action
!= bad_action
) {
316 write_attr(sync_actions
[a
->next_action
], a
->action_fd
);
317 dprintf(" action:%s", array_states
[a
->next_state
]);
319 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
320 if (mdi
->next_state
& DS_UNBLOCK
) {
321 dprintf(" %d:-blocked", mdi
->disk
.raid_disk
);
322 write_attr("-blocked", mdi
->state_fd
);
325 if ((mdi
->next_state
& DS_REMOVE
) && mdi
->state_fd
>= 0) {
328 /* the kernel may not be able to immediately remove the
329 * disk, we can simply wait until the next event to try
332 remove_result
= write_attr("remove", mdi
->state_fd
);
333 if (remove_result
> 0) {
334 dprintf(" %d:removed", mdi
->disk
.raid_disk
);
335 close(mdi
->state_fd
);
339 if (mdi
->next_state
& DS_INSYNC
) {
340 write_attr("+in_sync", mdi
->state_fd
);
341 dprintf(" %d:+in_sync", mdi
->disk
.raid_disk
);
346 /* move curr_ to prev_ */
347 a
->prev_state
= a
->curr_state
;
349 a
->prev_action
= a
->curr_action
;
351 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
) {
352 mdi
->prev_state
= mdi
->curr_state
;
356 if (check_degraded
) {
357 /* manager will do the actual check */
358 a
->check_degraded
= 1;
368 static struct mdinfo
*
369 find_device(struct active_array
*a
, int major
, int minor
)
373 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
374 if (mdi
->disk
.major
== major
&& mdi
->disk
.minor
== minor
)
380 static void reconcile_failed(struct active_array
*aa
, struct mdinfo
*failed
)
382 struct active_array
*a
;
383 struct mdinfo
*victim
;
385 for (a
= aa
; a
; a
= a
->next
) {
388 victim
= find_device(a
, failed
->disk
.major
, failed
->disk
.minor
);
392 if (!(victim
->curr_state
& DS_FAULTY
))
393 write_attr("faulty", victim
->state_fd
);
398 static void dprint_wake_reasons(fd_set
*fds
)
406 fprintf(stderr
, "monitor: wake ( ");
407 for (i
= 0; i
< FD_SETSIZE
; i
++) {
408 if (FD_ISSET(i
, fds
)) {
409 sprintf(proc_path
, "/proc/%d/fd/%d",
412 rv
= readlink(proc_path
, link
, sizeof(link
) - 1);
414 fprintf(stderr
, "%d:unknown ", i
);
418 basename
= strrchr(link
, '/');
419 fprintf(stderr
, "%d:%s ",
420 i
, basename
? ++basename
: link
);
423 fprintf(stderr
, ")\n");
427 int monitor_loop_cnt
;
429 static int wait_and_act(struct supertype
*container
, int nowait
)
433 struct active_array
**aap
= &container
->arrays
;
434 struct active_array
*a
, **ap
;
440 for (ap
= aap
; *ap
;) {
442 /* once an array has been deactivated we want to
443 * ask the manager to discard it.
457 add_fd(&rfds
, &maxfd
, a
->info
.state_fd
);
458 add_fd(&rfds
, &maxfd
, a
->action_fd
);
459 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
460 add_fd(&rfds
, &maxfd
, mdi
->state_fd
);
465 if (manager_ready
&& *aap
== NULL
) {
466 /* No interesting arrays. Lets see about exiting.
467 * Note that blocking at this point is not a problem
468 * as there are no active arrays, there is nothing that
469 * we need to be ready to do.
471 int fd
= open(container
->device_name
, O_RDONLY
|O_EXCL
);
472 if (fd
>= 0 || errno
!= EBUSY
) {
473 /* OK, we are safe to leave */
474 dprintf("no arrays to monitor... exiting\n");
475 remove_pidfile(container
->devname
);
484 sigprocmask(SIG_UNBLOCK
, NULL
, &set
);
485 sigdelset(&set
, SIGUSR1
);
486 monitor_loop_cnt
|= 1;
487 rv
= pselect(maxfd
+1, &rfds
, NULL
, NULL
, NULL
, &set
);
488 monitor_loop_cnt
+= 1;
489 if (rv
== -1 && errno
== EINTR
)
492 dprint_wake_reasons(&rfds
);
498 struct metadata_update
*this;
500 for (this = update_queue
; this ; this = this->next
)
501 container
->ss
->process_update(container
, this);
503 update_queue_handled
= update_queue
;
506 container
->ss
->sync_metadata(container
);
509 for (a
= *aap
; a
; a
= a
->next
) {
510 if (a
->replaces
&& !discard_this
) {
511 struct active_array
**ap
;
512 for (ap
= &a
->next
; *ap
&& *ap
!= a
->replaces
;
517 discard_this
= a
->replaces
;
519 /* FIXME check if device->state_fd need to be cleared?*/
523 rv
+= read_and_act(a
);
526 /* propagate failures across container members */
527 for (a
= *aap
; a
; a
= a
->next
) {
530 for (mdi
= a
->info
.devs
; mdi
; mdi
= mdi
->next
)
531 if (mdi
->curr_state
& DS_FAULTY
)
532 reconcile_failed(*aap
, mdi
);
538 void do_monitor(struct supertype
*container
)
543 rv
= wait_and_act(container
, first
);