]> git.ipfire.org Git - thirdparty/mdadm.git/blob - managemon.c
Merge branch 'master' into devel-3.2
[thirdparty/mdadm.git] / managemon.c
1 /*
2 * mdmon - monitor external metadata arrays
3 *
4 * Copyright (C) 2007-2009 Neil Brown <neilb@suse.de>
5 * Copyright (C) 2007-2009 Intel Corporation
6 *
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.
10 *
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
14 * more details.
15 *
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.
19 */
20
21 /*
22 * The management thread for monitoring active md arrays.
23 * This thread does things which might block such as memory
24 * allocation.
25 * In particular:
26 *
27 * - Find out about new arrays in this container.
28 * Allocate the data structures and open the files.
29 *
30 * For this we watch /proc/mdstat and find new arrays with
31 * metadata type that confirms sharing. e.g. "md4"
32 * When we find a new array we slip it into the list of
33 * arrays and signal 'monitor' by writing to a pipe.
34 *
35 * - Respond to reshape requests by allocating new data structures
36 * and opening new files.
37 *
38 * These come as a change to raid_disks. We allocate a new
39 * version of the data structures and slip it into the list.
40 * 'monitor' will notice and release the old version.
41 * Changes to level, chunksize, layout.. do not need re-allocation.
42 * Reductions in raid_disks don't really either, but we handle
43 * them the same way for consistency.
44 *
45 * - When a device is added to the container, we add it to the metadata
46 * as a spare.
47 *
48 * - Deal with degraded array
49 * We only do this when first noticing the array is degraded.
50 * This can be when we first see the array, when sync completes or
51 * when recovery completes.
52 *
53 * Check if number of failed devices suggests recovery is needed, and
54 * skip if not.
55 * Ask metadata to allocate a spare device
56 * Add device as not in_sync and give a role
57 * Update metadata.
58 * Open sysfs files and pass to monitor.
59 * Make sure that monitor Starts recovery....
60 *
61 * - Pass on metadata updates from external programs such as
62 * mdadm creating a new array.
63 *
64 * This is most-messy.
65 * It might involve adding a new array or changing the status of
66 * a spare, or any reconfig that the kernel doesn't get involved in.
67 *
68 * The required updates are received via a named pipe. There will
69 * be one named pipe for each container. Each message contains a
70 * sync marker: 0x5a5aa5a5, A byte count, and the message. This is
71 * passed to the metadata handler which will interpret and process it.
72 * For 'DDF' messages are internal data blocks with the leading
73 * 'magic number' signifying what sort of data it is.
74 *
75 */
76
77 /*
78 * We select on /proc/mdstat and the named pipe.
79 * We create new arrays or updated version of arrays and slip
80 * them into the head of the list, then signal 'monitor' via a pipe write.
81 * 'monitor' will notice and place the old array on a return list.
82 * Metadata updates are placed on a queue just like they arrive
83 * from the named pipe.
84 *
85 * When new arrays are found based on correct metadata string, we
86 * need to identify them with an entry in the metadata. Maybe we require
87 * the metadata to be mdX/NN when NN is the index into an appropriate table.
88 *
89 */
90
91 /*
92 * List of tasks:
93 * - Watch for spares to be added to the container, and write updated
94 * metadata to them.
95 * - Watch for new arrays using this container, confirm they match metadata
96 * and if so, start monitoring them
97 * - Watch for spares being added to monitored arrays. This shouldn't
98 * happen, as we should do all the adding. Just remove them.
99 * - Watch for change in raid-disks, chunk-size, etc. Update metadata and
100 * start a reshape.
101 */
102 #ifndef _GNU_SOURCE
103 #define _GNU_SOURCE
104 #endif
105 #include "mdadm.h"
106 #include "mdmon.h"
107 #include <sys/syscall.h>
108 #include <sys/socket.h>
109 #include <signal.h>
110
111 static void close_aa(struct active_array *aa)
112 {
113 struct mdinfo *d;
114
115 for (d = aa->info.devs; d; d = d->next) {
116 close(d->recovery_fd);
117 close(d->state_fd);
118 }
119
120 close(aa->action_fd);
121 close(aa->info.state_fd);
122 close(aa->resync_start_fd);
123 close(aa->metadata_fd);
124 close(aa->sync_completed_fd);
125 }
126
127 static void free_aa(struct active_array *aa)
128 {
129 /* Note that this doesn't close fds if they are being used
130 * by a clone. ->container will be set for a clone
131 */
132 dprintf("%s: devnum: %d\n", __func__, aa->devnum);
133 if (!aa->container)
134 close_aa(aa);
135 while (aa->info.devs) {
136 struct mdinfo *d = aa->info.devs;
137 aa->info.devs = d->next;
138 free(d);
139 }
140 free(aa);
141 }
142
143 static struct active_array *duplicate_aa(struct active_array *aa)
144 {
145 struct active_array *newa = malloc(sizeof(*newa));
146 struct mdinfo **dp1, **dp2;
147
148 *newa = *aa;
149 newa->next = NULL;
150 newa->replaces = NULL;
151 newa->info.next = NULL;
152
153 dp2 = &newa->info.devs;
154
155 for (dp1 = &aa->info.devs; *dp1; dp1 = &(*dp1)->next) {
156 struct mdinfo *d;
157 if ((*dp1)->state_fd < 0)
158 continue;
159
160 d = malloc(sizeof(*d));
161 *d = **dp1;
162 *dp2 = d;
163 dp2 = & d->next;
164 }
165 *dp2 = NULL;
166
167 return newa;
168 }
169
170 static void wakeup_monitor(void)
171 {
172 /* tgkill(getpid(), mon_tid, SIGUSR1); */
173 int pid = getpid();
174 syscall(SYS_tgkill, pid, mon_tid, SIGUSR1);
175 }
176
177 static void remove_old(void)
178 {
179 if (discard_this) {
180 discard_this->next = NULL;
181 free_aa(discard_this);
182 if (pending_discard == discard_this)
183 pending_discard = NULL;
184 discard_this = NULL;
185 wakeup_monitor();
186 }
187 }
188
189 static void replace_array(struct supertype *container,
190 struct active_array *old,
191 struct active_array *new)
192 {
193 /* To replace an array, we add it to the top of the list
194 * marked with ->replaces to point to the original.
195 * 'monitor' will take the original out of the list
196 * and put it on 'discard_this'. We take it from there
197 * and discard it.
198 */
199 remove_old();
200 while (pending_discard) {
201 while (discard_this == NULL)
202 sleep(1);
203 remove_old();
204 }
205 pending_discard = old;
206 new->replaces = old;
207 new->next = container->arrays;
208 container->arrays = new;
209 wakeup_monitor();
210 }
211
212 struct metadata_update *update_queue = NULL;
213 struct metadata_update *update_queue_handled = NULL;
214 struct metadata_update *update_queue_pending = NULL;
215
216 static void free_updates(struct metadata_update **update)
217 {
218 while (*update) {
219 struct metadata_update *this = *update;
220
221 *update = this->next;
222 free(this->buf);
223 free(this->space);
224 free(this);
225 }
226 }
227
228 void check_update_queue(struct supertype *container)
229 {
230 free_updates(&update_queue_handled);
231
232 if (update_queue == NULL &&
233 update_queue_pending) {
234 update_queue = update_queue_pending;
235 update_queue_pending = NULL;
236 wakeup_monitor();
237 }
238 }
239
240 static void queue_metadata_update(struct metadata_update *mu)
241 {
242 struct metadata_update **qp;
243
244 qp = &update_queue_pending;
245 while (*qp)
246 qp = & ((*qp)->next);
247 *qp = mu;
248 }
249
250 static void add_disk_to_container(struct supertype *st, struct mdinfo *sd)
251 {
252 int dfd;
253 char nm[20];
254 struct supertype *st2;
255 struct metadata_update *update = NULL;
256 struct mdinfo info;
257 mdu_disk_info_t dk = {
258 .number = -1,
259 .major = sd->disk.major,
260 .minor = sd->disk.minor,
261 .raid_disk = -1,
262 .state = 0,
263 };
264
265 dprintf("%s: add %d:%d to container\n",
266 __func__, sd->disk.major, sd->disk.minor);
267
268 sd->next = st->devs;
269 st->devs = sd;
270
271 sprintf(nm, "%d:%d", sd->disk.major, sd->disk.minor);
272 dfd = dev_open(nm, O_RDWR);
273 if (dfd < 0)
274 return;
275
276 /* Check the metadata and see if it is already part of this
277 * array
278 */
279 st2 = dup_super(st);
280 if (st2->ss->load_super(st2, dfd, NULL) == 0) {
281 st2->ss->getinfo_super(st, &info, NULL);
282 if (st->ss->compare_super(st, st2) == 0 &&
283 info.disk.raid_disk >= 0) {
284 /* Looks like a good member of array.
285 * Just accept it.
286 * mdadm will incorporate any parts into
287 * active arrays.
288 */
289 st2->ss->free_super(st2);
290 return;
291 }
292 }
293 st2->ss->free_super(st2);
294
295 st->update_tail = &update;
296 st->ss->add_to_super(st, &dk, dfd, NULL);
297 st->ss->write_init_super(st);
298 queue_metadata_update(update);
299 st->update_tail = NULL;
300 }
301
302 static void manage_container(struct mdstat_ent *mdstat,
303 struct supertype *container)
304 {
305 /* The only thing of interest here is if a new device
306 * has been added to the container. We add it to the
307 * array ignoring any metadata on it.
308 * FIXME should we look for compatible metadata and take hints
309 * about spare assignment.... probably not.
310 */
311 if (mdstat->devcnt != container->devcnt) {
312 struct mdinfo **cdp, *cd, *di, *mdi;
313 int found;
314
315 /* read /sys/block/NAME/md/dev-??/block/dev to find out
316 * what is there, and compare with container->info.devs
317 * To see what is removed and what is added.
318 * These need to be remove from, or added to, the array
319 */
320 mdi = sysfs_read(-1, mdstat->devnum, GET_DEVS);
321 if (!mdi) {
322 /* invalidate the current count so we can try again */
323 container->devcnt = -1;
324 return;
325 }
326
327 /* check for removals */
328 for (cdp = &container->devs; *cdp; ) {
329 found = 0;
330 for (di = mdi->devs; di; di = di->next)
331 if (di->disk.major == (*cdp)->disk.major &&
332 di->disk.minor == (*cdp)->disk.minor) {
333 found = 1;
334 break;
335 }
336 if (!found) {
337 cd = *cdp;
338 *cdp = (*cdp)->next;
339 free(cd);
340 } else
341 cdp = &(*cdp)->next;
342 }
343
344 /* check for additions */
345 for (di = mdi->devs; di; di = di->next) {
346 for (cd = container->devs; cd; cd = cd->next)
347 if (di->disk.major == cd->disk.major &&
348 di->disk.minor == cd->disk.minor)
349 break;
350 if (!cd) {
351 struct mdinfo *newd = malloc(sizeof(*newd));
352
353 if (!newd) {
354 container->devcnt = -1;
355 continue;
356 }
357 *newd = *di;
358 add_disk_to_container(container, newd);
359 }
360 }
361 sysfs_free(mdi);
362 container->devcnt = mdstat->devcnt;
363 }
364 }
365
366 static int disk_init_and_add(struct mdinfo *disk, struct mdinfo *clone,
367 struct active_array *aa)
368 {
369 if (!disk || !clone)
370 return -1;
371
372 *disk = *clone;
373 disk->recovery_fd = sysfs_open(aa->devnum, disk->sys_name, "recovery_start");
374 disk->state_fd = sysfs_open(aa->devnum, disk->sys_name, "state");
375 disk->prev_state = read_dev_state(disk->state_fd);
376 disk->curr_state = disk->prev_state;
377 disk->next = aa->info.devs;
378 aa->info.devs = disk;
379
380 return 0;
381 }
382
383 static void manage_member(struct mdstat_ent *mdstat,
384 struct active_array *a)
385 {
386 /* Compare mdstat info with known state of member array.
387 * We do not need to look for device state changes here, that
388 * is dealt with by the monitor.
389 *
390 * We just look for changes which suggest that a reshape is
391 * being requested.
392 * Unfortunately decreases in raid_disks don't show up in
393 * mdstat until the reshape completes FIXME.
394 *
395 * Actually, we also want to handle degraded arrays here by
396 * trying to find and assign a spare.
397 * We do that whenever the monitor tells us too.
398 */
399 char buf[64];
400 int frozen;
401
402 // FIXME
403 a->info.array.raid_disks = mdstat->raid_disks;
404 // MORE
405
406 /* honor 'frozen' */
407 if (sysfs_get_str(&a->info, NULL, "metadata_version", buf, sizeof(buf)) > 0)
408 frozen = buf[9] == '-';
409 else
410 frozen = 1; /* can't read metadata_version assume the worst */
411
412 if (a->check_degraded && !frozen) {
413 struct metadata_update *updates = NULL;
414 struct mdinfo *newdev = NULL;
415 struct active_array *newa;
416 struct mdinfo *d;
417
418 a->check_degraded = 0;
419
420 /* The array may not be degraded, this is just a good time
421 * to check.
422 */
423 newdev = a->container->ss->activate_spare(a, &updates);
424 if (!newdev)
425 return;
426
427 newa = duplicate_aa(a);
428 if (!newa)
429 goto out;
430 /* Cool, we can add a device or several. */
431
432 /* Add device to array and set offset/size/slot.
433 * and open files for each newdev */
434 for (d = newdev; d ; d = d->next) {
435 struct mdinfo *newd;
436
437 newd = malloc(sizeof(*newd));
438 if (!newd)
439 continue;
440 if (sysfs_add_disk(&newa->info, d, 0) < 0) {
441 free(newd);
442 continue;
443 }
444 disk_init_and_add(newd, d, newa);
445 }
446 queue_metadata_update(updates);
447 updates = NULL;
448 replace_array(a->container, a, newa);
449 sysfs_set_str(&a->info, NULL, "sync_action", "recover");
450 out:
451 while (newdev) {
452 d = newdev->next;
453 free(newdev);
454 newdev = d;
455 }
456 free_updates(&updates);
457 }
458 }
459
460 static int aa_ready(struct active_array *aa)
461 {
462 struct mdinfo *d;
463 int level = aa->info.array.level;
464
465 for (d = aa->info.devs; d; d = d->next)
466 if (d->state_fd < 0)
467 return 0;
468
469 if (aa->info.state_fd < 0)
470 return 0;
471
472 if (level > 0 && (aa->action_fd < 0 || aa->resync_start_fd < 0))
473 return 0;
474
475 if (!aa->container)
476 return 0;
477
478 return 1;
479 }
480
481 static void manage_new(struct mdstat_ent *mdstat,
482 struct supertype *container,
483 struct active_array *victim)
484 {
485 /* A new array has appeared in this container.
486 * Hopefully it is already recorded in the metadata.
487 * Check, then create the new array to report it to
488 * the monitor.
489 */
490
491 struct active_array *new;
492 struct mdinfo *mdi, *di;
493 char *inst;
494 int i;
495 int failed = 0;
496
497 /* check if array is ready to be monitored */
498 if (!mdstat->active)
499 return;
500
501 mdi = sysfs_read(-1, mdstat->devnum,
502 GET_LEVEL|GET_CHUNK|GET_DISKS|GET_COMPONENT|
503 GET_DEGRADED|GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE);
504
505 new = malloc(sizeof(*new));
506
507 if (!new || !mdi) {
508 if (mdi)
509 sysfs_free(mdi);
510 if (new)
511 free(new);
512 return;
513 }
514 memset(new, 0, sizeof(*new));
515
516 new->devnum = mdstat->devnum;
517 strcpy(new->info.sys_name, devnum2devname(new->devnum));
518
519 new->prev_state = new->curr_state = new->next_state = inactive;
520 new->prev_action= new->curr_action= new->next_action= idle;
521
522 new->container = container;
523
524 inst = to_subarray(mdstat, container->devname);
525
526 new->info.array = mdi->array;
527 new->info.component_size = mdi->component_size;
528
529 for (i = 0; i < new->info.array.raid_disks; i++) {
530 struct mdinfo *newd = malloc(sizeof(*newd));
531
532 for (di = mdi->devs; di; di = di->next)
533 if (i == di->disk.raid_disk)
534 break;
535
536 if (disk_init_and_add(newd, di, new) != 0) {
537 if (newd)
538 free(newd);
539
540 failed++;
541 if (failed > new->info.array.failed_disks) {
542 /* we cannot properly monitor without all working disks */
543 new->container = NULL;
544 break;
545 }
546 }
547 }
548
549 new->action_fd = sysfs_open(new->devnum, NULL, "sync_action");
550 new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state");
551 new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start");
552 new->metadata_fd = sysfs_open(new->devnum, NULL, "metadata_version");
553 new->sync_completed_fd = sysfs_open(new->devnum, NULL, "sync_completed");
554 dprintf("%s: inst: %d action: %d state: %d\n", __func__, atoi(inst),
555 new->action_fd, new->info.state_fd);
556
557 sysfs_free(mdi);
558
559 /* if everything checks out tell the metadata handler we want to
560 * manage this instance
561 */
562 if (!aa_ready(new) || container->ss->open_new(container, new, inst) < 0) {
563 fprintf(stderr, "mdmon: failed to monitor %s\n",
564 mdstat->metadata_version);
565 new->container = NULL;
566 free_aa(new);
567 } else {
568 replace_array(container, victim, new);
569 if (failed) {
570 new->check_degraded = 1;
571 manage_member(mdstat, new);
572 }
573 }
574 }
575
576 void manage(struct mdstat_ent *mdstat, struct supertype *container)
577 {
578 /* We have just read mdstat and need to compare it with
579 * the known active arrays.
580 * Arrays with the wrong metadata are ignored.
581 */
582
583 for ( ; mdstat ; mdstat = mdstat->next) {
584 struct active_array *a;
585 if (mdstat->devnum == container->devnum) {
586 manage_container(mdstat, container);
587 continue;
588 }
589 if (!is_container_member(mdstat, container->devname))
590 /* Not for this array */
591 continue;
592 /* Looks like a member of this container */
593 for (a = container->arrays; a; a = a->next) {
594 if (mdstat->devnum == a->devnum) {
595 if (a->container)
596 manage_member(mdstat, a);
597 break;
598 }
599 }
600 if (a == NULL || !a->container)
601 manage_new(mdstat, container, a);
602 }
603 }
604
605 static void handle_message(struct supertype *container, struct metadata_update *msg)
606 {
607 /* queue this metadata update through to the monitor */
608
609 struct metadata_update *mu;
610
611 if (msg->len <= 0)
612 while (update_queue_pending || update_queue) {
613 check_update_queue(container);
614 usleep(15*1000);
615 }
616
617 if (msg->len == 0) { /* ping_monitor */
618 int cnt;
619
620 cnt = monitor_loop_cnt;
621 if (cnt & 1)
622 cnt += 2; /* wait until next pselect */
623 else
624 cnt += 3; /* wait for 2 pselects */
625 wakeup_monitor();
626
627 while (monitor_loop_cnt - cnt < 0)
628 usleep(10 * 1000);
629 } else if (msg->len == -1) { /* ping_manager */
630 struct mdstat_ent *mdstat = mdstat_read(1, 0);
631
632 manage(mdstat, container);
633 free_mdstat(mdstat);
634 } else if (!sigterm) {
635 mu = malloc(sizeof(*mu));
636 mu->len = msg->len;
637 mu->buf = msg->buf;
638 msg->buf = NULL;
639 mu->space = NULL;
640 mu->next = NULL;
641 if (container->ss->prepare_update)
642 container->ss->prepare_update(container, mu);
643 queue_metadata_update(mu);
644 }
645 }
646
647 void read_sock(struct supertype *container)
648 {
649 int fd;
650 struct metadata_update msg;
651 int terminate = 0;
652 long fl;
653 int tmo = 3; /* 3 second timeout before hanging up the socket */
654
655 fd = accept(container->sock, NULL, NULL);
656 if (fd < 0)
657 return;
658
659 fl = fcntl(fd, F_GETFL, 0);
660 fl |= O_NONBLOCK;
661 fcntl(fd, F_SETFL, fl);
662
663 do {
664 msg.buf = NULL;
665
666 /* read and validate the message */
667 if (receive_message(fd, &msg, tmo) == 0) {
668 handle_message(container, &msg);
669 if (msg.len == 0) {
670 /* ping reply with version */
671 msg.buf = Version;
672 msg.len = strlen(Version) + 1;
673 if (send_message(fd, &msg, tmo) < 0)
674 terminate = 1;
675 } else if (ack(fd, tmo) < 0)
676 terminate = 1;
677 } else
678 terminate = 1;
679
680 } while (!terminate);
681
682 close(fd);
683 }
684
685 int exit_now = 0;
686 int manager_ready = 0;
687 void do_manager(struct supertype *container)
688 {
689 struct mdstat_ent *mdstat;
690 sigset_t set;
691
692 sigprocmask(SIG_UNBLOCK, NULL, &set);
693 sigdelset(&set, SIGUSR1);
694 sigdelset(&set, SIGTERM);
695
696 do {
697
698 if (exit_now)
699 exit(0);
700
701 /* Can only 'manage' things if 'monitor' is not making
702 * structural changes to metadata, so need to check
703 * update_queue
704 */
705 if (update_queue == NULL) {
706 mdstat = mdstat_read(1, 0);
707
708 manage(mdstat, container);
709
710 read_sock(container);
711
712 free_mdstat(mdstat);
713 }
714 remove_old();
715
716 check_update_queue(container);
717
718 manager_ready = 1;
719
720 if (sigterm)
721 wakeup_monitor();
722
723 if (update_queue == NULL)
724 mdstat_wait_fd(container->sock, &set);
725 else
726 /* If an update is happening, just wait for signal */
727 pselect(0, NULL, NULL, NULL, NULL, &set);
728 } while(1);
729 }