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)
48
static int read_attr(char *buf, int len, int fd)
57
n = read(fd, buf, len - 1);
69
static unsigned long long read_resync_start(int fd)
74
n = read_attr(buf, 30, fd);
77
if (strncmp(buf, "none", 4) == 0)
80
return strtoull(buf, NULL, 10);
83
static unsigned long long read_sync_completed(int fd)
85
unsigned long long val;
90
n = read_attr(buf, 50, fd);
95
val = strtoull(buf, &ep, 0);
96
if (ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))
101
static enum array_state read_state(int fd)
104
int n = read_attr(buf, 20, fd);
108
return (enum array_state) sysfs_match_word(buf, array_states);
111
static enum sync_action read_action( int fd)
114
int n = read_attr(buf, 20, fd);
118
return (enum sync_action) sysfs_match_word(buf, sync_actions);
121
int read_dev_state(int fd)
124
int n = read_attr(buf, 60, fd);
133
if (sysfs_attr_match(cp, "faulty"))
135
if (sysfs_attr_match(cp, "in_sync"))
137
if (sysfs_attr_match(cp, "write_mostly"))
138
rv |= DS_WRITE_MOSTLY;
139
if (sysfs_attr_match(cp, "spare"))
141
if (sysfs_attr_match(cp, "blocked"))
143
cp = strchr(cp, ',');
150
static void signal_manager(void)
152
/* tgkill(getpid(), mon_tid, SIGUSR1); */
154
syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
157
/* Monitor a set of active md arrays - all of which share the
158
* same metadata - and respond to events that require
161
* New arrays are detected by another thread which allocates
162
* required memory and attaches the data structure to our list.
166
* This is detected by array_state going to 'clear' or 'inactive'.
167
* while we thought it was active.
168
* Response is to mark metadata as clean and 'clear' the array(??)
170
* array_state if 'write-pending'
171
* We mark metadata as 'dirty' then set array to 'active'.
173
* Either ignore, or mark clean, then mark metadata as clean.
176
* detected by rd-N/state reporting "faulty"
177
* mark device as 'failed' in metadata, let the kernel release the
178
* device by writing '-blocked' to rd/state, and finally write 'remove' to
179
* rd/state. Before a disk can be replaced it must be failed and removed
180
* from all container members, this will be preemptive for the other
184
* sync_action was 'resync' and becomes 'idle' and resync_start becomes
186
* Notify metadata that sync is complete.
189
* sync_action changes from 'recover' to 'idle'
190
* Check each device state and mark metadata if 'faulty' or 'in_sync'.
193
* This only happens on finding a new array... mdadm will have set
194
* 'resync_start' to the correct value. If 'resync_start' indicates that an
195
* resync needs to occur set the array to the 'active' state rather than the
196
* initial read-auto state.
200
* We wait for a change (poll/select) on array_state, sync_action, and
201
* each rd-X/state file.
202
* When we get any change, we check everything. So read each state file,
203
* then decide what to do.
205
* The core action is to write new metadata to all devices in the array.
206
* This is done at most once on any wakeup.
207
* After that we might:
208
* - update the array_state
209
* - set the role of some devices.
210
* - request a sync_action
214
static int read_and_act(struct active_array *a)
216
unsigned long long sync_completed;
217
int check_degraded = 0;
222
a->next_state = bad_word;
223
a->next_action = bad_action;
225
a->curr_state = read_state(a->info.state_fd);
226
a->curr_action = read_action(a->action_fd);
227
a->info.resync_start = read_resync_start(a->resync_start_fd);
228
sync_completed = read_sync_completed(a->sync_completed_fd);
229
for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
231
if (mdi->state_fd >= 0) {
232
mdi->recovery_start = read_resync_start(mdi->recovery_fd);
233
mdi->curr_state = read_dev_state(mdi->state_fd);
237
if (a->curr_state <= inactive &&
238
a->prev_state > inactive) {
239
/* array has been stopped */
240
a->container->ss->set_array_state(a, 1);
241
a->next_state = clear;
244
if (a->curr_state == write_pending) {
245
a->container->ss->set_array_state(a, 0);
246
a->next_state = active;
249
if (a->curr_state == active_idle) {
250
/* Set array to 'clean' FIRST, then mark clean
253
a->next_state = clean;
256
if (a->curr_state == clean) {
257
a->container->ss->set_array_state(a, 1);
259
if (a->curr_state == active ||
260
a->curr_state == suspended ||
261
a->curr_state == bad_word)
263
if (a->curr_state == readonly) {
264
/* Well, I'm ready to handle things. If readonly
265
* wasn't requested, transition to read-auto.
268
read_attr(buf, sizeof(buf), a->metadata_fd);
269
if (strncmp(buf, "external:-", 10) == 0) {
270
/* explicit request for readonly array. Leave it alone */
273
if (a->container->ss->set_array_state(a, 2))
274
a->next_state = read_auto; /* array is clean */
276
a->next_state = active; /* Now active for recovery etc */
283
a->curr_action == idle &&
284
a->prev_action == resync) {
285
/* A resync has finished. The endpoint is recorded in
286
* 'sync_start'. We don't update the metadata
287
* until the array goes inactive or readonly though.
288
* Just check if we need to fiddle spares.
290
a->container->ss->set_array_state(a, a->curr_state <= clean);
295
a->curr_action == idle &&
296
a->prev_action == recover) {
297
/* A recovery has finished. Some disks may be in sync now,
298
* and the array may no longer be degraded
300
for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
301
a->container->ss->set_disk(a, mdi->disk.raid_disk,
303
if (! (mdi->curr_state & DS_INSYNC))
308
/* Check for failures and if found:
309
* 1/ Record the failure in the metadata and unblock the device.
310
* FIXME update the kernel to stop notifying on failed drives when
311
* the array is readonly and we have cleared 'blocked'
312
* 2/ Try to remove the device if the array is writable, or can be
315
for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
316
if (mdi->curr_state & DS_FAULTY) {
317
a->container->ss->set_disk(a, mdi->disk.raid_disk,
320
mdi->next_state |= DS_UNBLOCK;
321
if (a->curr_state == read_auto) {
322
a->container->ss->set_array_state(a, 0);
323
a->next_state = active;
325
if (a->curr_state > readonly)
326
mdi->next_state |= DS_REMOVE;
330
/* Check for recovery checkpoint notifications. We need to be a
331
* minimum distance away from the last checkpoint to prevent
332
* over checkpointing. Note reshape checkpointing is not
335
if (sync_completed > a->last_checkpoint &&
336
sync_completed - a->last_checkpoint > a->info.component_size >> 4 &&
337
a->curr_action > reshape) {
338
/* A (non-reshape) sync_action has reached a checkpoint.
339
* Record the updated position in the metadata
341
a->last_checkpoint = sync_completed;
342
a->container->ss->set_array_state(a, a->curr_state <= clean);
343
} else if (sync_completed > a->last_checkpoint)
344
a->last_checkpoint = sync_completed;
346
a->container->ss->sync_metadata(a->container);
347
dprintf("%s(%d): state:%s action:%s next(", __func__, a->info.container_member,
348
array_states[a->curr_state], sync_actions[a->curr_action]);
350
/* Effect state changes in the array */
351
if (a->next_state != bad_word) {
352
dprintf(" state:%s", array_states[a->next_state]);
353
write_attr(array_states[a->next_state], a->info.state_fd);
355
if (a->next_action != bad_action) {
356
write_attr(sync_actions[a->next_action], a->action_fd);
357
dprintf(" action:%s", sync_actions[a->next_action]);
359
for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
360
if (mdi->next_state & DS_UNBLOCK) {
361
dprintf(" %d:-blocked", mdi->disk.raid_disk);
362
write_attr("-blocked", mdi->state_fd);
365
if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
368
/* the kernel may not be able to immediately remove the
369
* disk, we can simply wait until the next event to try
372
remove_result = write_attr("remove", mdi->state_fd);
373
if (remove_result > 0) {
374
dprintf(" %d:removed", mdi->disk.raid_disk);
375
close(mdi->state_fd);
379
if (mdi->next_state & DS_INSYNC) {
380
write_attr("+in_sync", mdi->state_fd);
381
dprintf(" %d:+in_sync", mdi->disk.raid_disk);
386
/* move curr_ to prev_ */
387
a->prev_state = a->curr_state;
389
a->prev_action = a->curr_action;
391
for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
392
mdi->prev_state = mdi->curr_state;
396
if (check_degraded) {
397
/* manager will do the actual check */
398
a->check_degraded = 1;
408
static struct mdinfo *
409
find_device(struct active_array *a, int major, int minor)
413
for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
414
if (mdi->disk.major == major && mdi->disk.minor == minor)
420
static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
422
struct active_array *a;
423
struct mdinfo *victim;
425
for (a = aa; a; a = a->next) {
428
victim = find_device(a, failed->disk.major, failed->disk.minor);
432
if (!(victim->curr_state & DS_FAULTY))
433
write_attr("faulty", victim->state_fd);
438
static void dprint_wake_reasons(fd_set *fds)
446
fprintf(stderr, "monitor: wake ( ");
447
for (i = 0; i < FD_SETSIZE; i++) {
448
if (FD_ISSET(i, fds)) {
449
sprintf(proc_path, "/proc/%d/fd/%d",
452
rv = readlink(proc_path, link, sizeof(link) - 1);
454
fprintf(stderr, "%d:unknown ", i);
458
basename = strrchr(link, '/');
459
fprintf(stderr, "%d:%s ",
460
i, basename ? ++basename : link);
463
fprintf(stderr, ")\n");
467
int monitor_loop_cnt;
469
static int wait_and_act(struct supertype *container, int nowait)
473
struct active_array **aap = &container->arrays;
474
struct active_array *a, **ap;
477
static unsigned int dirty_arrays = ~0; /* start at some non-zero value */
481
for (ap = aap ; *ap ;) {
483
/* once an array has been deactivated we want to
484
* ask the manager to discard it.
498
add_fd(&rfds, &maxfd, a->info.state_fd);
499
add_fd(&rfds, &maxfd, a->action_fd);
500
add_fd(&rfds, &maxfd, a->sync_completed_fd);
501
for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
502
add_fd(&rfds, &maxfd, mdi->state_fd);
507
if (manager_ready && (*aap == NULL || (sigterm && !dirty_arrays))) {
508
/* No interesting arrays, or we have been told to
509
* terminate and everything is clean. Lets see about
510
* exiting. Note that blocking at this point is not a
511
* problem as there are no active arrays, there is
512
* nothing that we need to be ready to do.
514
int fd = open_dev_excl(container->devnum);
515
if (fd >= 0 || errno != EBUSY) {
516
/* OK, we are safe to leave */
517
if (sigterm && !dirty_arrays)
518
dprintf("caught sigterm, all clean... exiting\n");
520
dprintf("no arrays to monitor... exiting\n");
522
/* On SIGTERM, someone (the take-over mdmon) will
525
remove_pidfile(container->devname);
534
sigprocmask(SIG_UNBLOCK, NULL, &set);
535
sigdelset(&set, SIGUSR1);
536
monitor_loop_cnt |= 1;
537
rv = pselect(maxfd+1, NULL, NULL, &rfds, NULL, &set);
538
monitor_loop_cnt += 1;
539
if (rv == -1 && errno == EINTR)
542
dprint_wake_reasons(&rfds);
548
struct metadata_update *this;
550
for (this = update_queue; this ; this = this->next)
551
container->ss->process_update(container, this);
553
update_queue_handled = update_queue;
556
container->ss->sync_metadata(container);
561
for (a = *aap; a ; a = a->next) {
564
if (a->replaces && !discard_this) {
565
struct active_array **ap;
566
for (ap = &a->next; *ap && *ap != a->replaces;
571
discard_this = a->replaces;
573
/* FIXME check if device->state_fd need to be cleared?*/
577
is_dirty = read_and_act(a);
579
dirty_arrays += is_dirty;
580
/* when terminating stop manipulating the array after it
581
* is clean, but make sure read_and_act() is given a
582
* chance to handle 'active_idle'
584
if (sigterm && !is_dirty)
585
a->container = NULL; /* stop touching this array */
589
/* propagate failures across container members */
590
for (a = *aap; a ; a = a->next) {
593
for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
594
if (mdi->curr_state & DS_FAULTY)
595
reconcile_failed(*aap, mdi);
601
void do_monitor(struct supertype *container)
606
rv = wait_and_act(container, first);