]>
Commit | Line | Data |
---|---|---|
549e9569 NB |
1 | |
2 | /* | |
3 | * The management thread for monitoring active md arrays. | |
4 | * This thread does things which might block such as memory | |
5 | * allocation. | |
6 | * In particular: | |
7 | * | |
8 | * - Find out about new arrays in this container. | |
9 | * Allocate the data structures and open the files. | |
10 | * | |
11 | * For this we watch /proc/mdstat and find new arrays with | |
12 | * metadata type that confirms sharing. e.g. "md4" | |
13 | * When we find a new array we slip it into the list of | |
14 | * arrays and signal 'monitor' by writing to a pipe. | |
15 | * | |
16 | * - Respond to reshape requests by allocating new data structures | |
17 | * and opening new files. | |
18 | * | |
19 | * These come as a change to raid_disks. We allocate a new | |
20 | * version of the data structures and slip it into the list. | |
21 | * 'monitor' will notice and release the old version. | |
22 | * Changes to level, chunksize, layout.. do not need re-allocation. | |
23 | * Reductions in raid_disks don't really either, but we handle | |
24 | * them the same way for consistency. | |
25 | * | |
26 | * - When a device is added to the container, we add it to the metadata | |
27 | * as a spare. | |
28 | * | |
6c3fb95c NB |
29 | * - Deal with degraded array |
30 | * We only do this when first noticing the array is degraded. | |
31 | * This can be when we first see the array, when sync completes or | |
32 | * when recovery completes. | |
33 | * | |
34 | * Check if number of failed devices suggests recovery is needed, and | |
35 | * skip if not. | |
36 | * Ask metadata to allocate a spare device | |
37 | * Add device as not in_sync and give a role | |
38 | * Update metadata. | |
39 | * Open sysfs files and pass to monitor. | |
40 | * Make sure that monitor Starts recovery.... | |
549e9569 NB |
41 | * |
42 | * - Pass on metadata updates from external programs such as | |
43 | * mdadm creating a new array. | |
44 | * | |
45 | * This is most-messy. | |
46 | * It might involve adding a new array or changing the status of | |
47 | * a spare, or any reconfig that the kernel doesn't get involved in. | |
48 | * | |
49 | * The required updates are received via a named pipe. There will | |
50 | * be one named pipe for each container. Each message contains a | |
51 | * sync marker: 0x5a5aa5a5, A byte count, and the message. This is | |
52 | * passed to the metadata handler which will interpret and process it. | |
53 | * For 'DDF' messages are internal data blocks with the leading | |
54 | * 'magic number' signifying what sort of data it is. | |
55 | * | |
56 | */ | |
57 | ||
58 | /* | |
59 | * We select on /proc/mdstat and the named pipe. | |
60 | * We create new arrays or updated version of arrays and slip | |
61 | * them into the head of the list, then signal 'monitor' via a pipe write. | |
62 | * 'monitor' will notice and place the old array on a return list. | |
63 | * Metadata updates are placed on a queue just like they arrive | |
64 | * from the named pipe. | |
65 | * | |
66 | * When new arrays are found based on correct metadata string, we | |
67 | * need to identify them with an entry in the metadata. Maybe we require | |
68 | * the metadata to be mdX/NN when NN is the index into an appropriate table. | |
69 | * | |
70 | */ | |
71 | ||
72 | /* | |
73 | * List of tasks: | |
74 | * - Watch for spares to be added to the container, and write updated | |
75 | * metadata to them. | |
76 | * - Watch for new arrays using this container, confirm they match metadata | |
77 | * and if so, start monitoring them | |
78 | * - Watch for spares being added to monitored arrays. This shouldn't | |
79 | * happen, as we should do all the adding. Just remove them. | |
80 | * - Watch for change in raid-disks, chunk-size, etc. Update metadata and | |
81 | * start a reshape. | |
82 | */ | |
83 | #ifndef _GNU_SOURCE | |
84 | #define _GNU_SOURCE | |
85 | #endif | |
86 | #include "mdadm.h" | |
87 | #include "mdmon.h" | |
88 | #include <sys/socket.h> | |
1ed3f387 | 89 | #include <signal.h> |
549e9569 | 90 | |
2a0bb19e DW |
91 | static void close_aa(struct active_array *aa) |
92 | { | |
93 | struct mdinfo *d; | |
94 | ||
95 | for (d = aa->info.devs; d; d = d->next) | |
96 | close(d->state_fd); | |
97 | ||
98 | close(aa->action_fd); | |
99 | close(aa->info.state_fd); | |
100 | close(aa->resync_start_fd); | |
2a0bb19e DW |
101 | } |
102 | ||
549e9569 NB |
103 | static void free_aa(struct active_array *aa) |
104 | { | |
2a0bb19e DW |
105 | /* Note that this doesn't close fds if they are being used |
106 | * by a clone. ->container will be set for a clone | |
549e9569 | 107 | */ |
2a0bb19e DW |
108 | if (!aa->container) |
109 | close_aa(aa); | |
549e9569 NB |
110 | while (aa->info.devs) { |
111 | struct mdinfo *d = aa->info.devs; | |
112 | aa->info.devs = d->next; | |
113 | free(d); | |
114 | } | |
115 | free(aa); | |
116 | } | |
117 | ||
6c3fb95c NB |
118 | static struct active_array *duplicate_aa(struct active_array *aa) |
119 | { | |
120 | struct active_array *newa = malloc(sizeof(*newa)); | |
121 | struct mdinfo **dp1, **dp2; | |
122 | ||
123 | *newa = *aa; | |
124 | newa->next = NULL; | |
125 | newa->replaces = NULL; | |
126 | newa->info.next = NULL; | |
127 | ||
128 | dp2 = &newa->info.devs; | |
129 | ||
130 | for (dp1 = &aa->info.devs; *dp1; dp1 = &(*dp1)->next) { | |
131 | struct mdinfo *d; | |
132 | if ((*dp1)->state_fd < 0) | |
133 | continue; | |
134 | ||
135 | d = malloc(sizeof(*d)); | |
136 | *d = **dp1; | |
137 | *dp2 = d; | |
138 | dp2 = & d->next; | |
139 | } | |
7e1432fb | 140 | *dp2 = NULL; |
6c3fb95c NB |
141 | |
142 | return newa; | |
143 | } | |
144 | ||
2a0bb19e DW |
145 | static void write_wakeup(struct supertype *c) |
146 | { | |
3e70c845 DW |
147 | static struct md_generic_cmd cmd = { .action = md_action_ping_monitor }; |
148 | int err; | |
149 | ||
150 | active_cmd = &cmd; | |
151 | ||
152 | /* send the monitor thread a pointer to the ping action */ | |
153 | write(c->mgr_pipe[1], &err, 1); | |
154 | read(c->mon_pipe[0], &err, 1); | |
2a0bb19e DW |
155 | } |
156 | ||
1ed3f387 NB |
157 | static void remove_old(void) |
158 | { | |
159 | if (discard_this) { | |
160 | discard_this->next = NULL; | |
161 | free_aa(discard_this); | |
162 | if (pending_discard == discard_this) | |
163 | pending_discard = NULL; | |
164 | discard_this = NULL; | |
165 | } | |
166 | } | |
167 | ||
549e9569 NB |
168 | static void replace_array(struct supertype *container, |
169 | struct active_array *old, | |
170 | struct active_array *new) | |
171 | { | |
172 | /* To replace an array, we add it to the top of the list | |
173 | * marked with ->replaces to point to the original. | |
174 | * 'monitor' will take the original out of the list | |
175 | * and put it on 'discard_this'. We take it from there | |
176 | * and discard it. | |
177 | */ | |
1ed3f387 | 178 | remove_old(); |
549e9569 | 179 | while (pending_discard) { |
1ed3f387 | 180 | write_wakeup(container); |
549e9569 NB |
181 | while (discard_this == NULL) |
182 | sleep(1); | |
1ed3f387 | 183 | remove_old(); |
549e9569 NB |
184 | } |
185 | pending_discard = old; | |
186 | new->replaces = old; | |
187 | new->next = container->arrays; | |
188 | container->arrays = new; | |
2a0bb19e | 189 | write_wakeup(container); |
549e9569 NB |
190 | } |
191 | ||
2e735d19 NB |
192 | struct metadata_update *update_queue = NULL; |
193 | struct metadata_update *update_queue_handled = NULL; | |
194 | struct metadata_update *update_queue_pending = NULL; | |
195 | ||
196 | void check_update_queue(struct supertype *container) | |
197 | { | |
198 | while (update_queue_handled) { | |
199 | struct metadata_update *this = update_queue_handled; | |
200 | update_queue_handled = this->next; | |
7e1432fb | 201 | // free(this->buf); |
2e735d19 NB |
202 | free(this); |
203 | } | |
204 | if (update_queue == NULL && | |
205 | update_queue_pending) { | |
206 | update_queue = update_queue_pending; | |
207 | update_queue_pending = NULL; | |
208 | write_wakeup(container); | |
209 | } | |
210 | } | |
211 | ||
6c3fb95c | 212 | static void queue_metadata_update(struct metadata_update *mu) |
2e735d19 NB |
213 | { |
214 | struct metadata_update **qp; | |
215 | ||
216 | qp = &update_queue_pending; | |
217 | while (*qp) | |
218 | qp = & ((*qp)->next); | |
219 | *qp = mu; | |
220 | } | |
221 | ||
222 | void wait_update_handled(void) | |
223 | { | |
224 | /* Wait for any pending update to be handled by monitor. | |
225 | * i.e. wait until update_queue is NULL | |
226 | */ | |
227 | while (update_queue) | |
228 | usleep(100 * 1000); | |
229 | } | |
230 | ||
549e9569 NB |
231 | static void manage_container(struct mdstat_ent *mdstat, |
232 | struct supertype *container) | |
233 | { | |
234 | /* The only thing of interest here is if a new device | |
235 | * has been added to the container. We add it to the | |
236 | * array ignoring any metadata on it. | |
237 | * FIXME should we look for compatible metadata and take hints | |
238 | * about spare assignment.... probably not. | |
549e9569 NB |
239 | */ |
240 | if (mdstat->devcnt != container->devcnt) { | |
241 | /* read /sys/block/NAME/md/dev-??/block/dev to find out | |
242 | * what is there, and compare with container->info.devs | |
243 | * To see what is removed and what is added. | |
244 | * These need to be remove from, or added to, the array | |
245 | */ | |
246 | // FIXME | |
247 | container->devcnt = mdstat->devcnt; | |
248 | } | |
249 | } | |
250 | ||
251 | static void manage_member(struct mdstat_ent *mdstat, | |
252 | struct active_array *a) | |
253 | { | |
254 | /* Compare mdstat info with known state of member array. | |
255 | * We do not need to look for device state changes here, that | |
256 | * is dealt with by the monitor. | |
257 | * | |
258 | * We just look for changes which suggest that a reshape is | |
259 | * being requested. | |
260 | * Unfortunately decreases in raid_disks don't show up in | |
261 | * mdstat until the reshape completes FIXME. | |
6c3fb95c NB |
262 | * |
263 | * Actually, we also want to handle degraded arrays here by | |
264 | * trying to find and assign a spare. | |
265 | * We do that whenever the monitor tells us too. | |
549e9569 NB |
266 | */ |
267 | // FIXME | |
268 | a->info.array.raid_disks = mdstat->raid_disks; | |
269 | a->info.array.chunk_size = mdstat->chunk_size; | |
270 | // MORE | |
271 | ||
6c3fb95c NB |
272 | if (a->check_degraded) { |
273 | struct metadata_update *updates = NULL; | |
274 | struct mdinfo *newdev; | |
275 | struct active_array *newa; | |
276 | wait_update_handled(); | |
277 | a->check_degraded = 0; | |
278 | ||
279 | /* The array may not be degraded, this is just a good time | |
280 | * to check. | |
281 | */ | |
282 | newdev = a->container->ss->activate_spare(a, &updates); | |
283 | if (newdev) { | |
284 | struct mdinfo *d; | |
285 | /* Cool, we can add a device or several. */ | |
286 | newa = duplicate_aa(a); | |
287 | /* suspend recovery - maybe not needed */ | |
288 | ||
289 | /* Add device to array and set offset/size/slot. | |
290 | * and open files for each newdev */ | |
291 | for (d = newdev; d ; d = d->next) { | |
292 | struct mdinfo *newd; | |
293 | if (sysfs_add_disk(&newa->info, d)) | |
294 | continue; | |
295 | newd = newa->info.devs; | |
296 | newd->state_fd = sysfs_open(a->devnum, | |
297 | newd->sys_name, | |
298 | "state"); | |
299 | newd->prev_state | |
300 | = read_dev_state(newd->state_fd); | |
301 | newd->curr_state = newd->prev_state; | |
302 | } | |
303 | queue_metadata_update(updates); | |
304 | replace_array(a->container, a, newa); | |
305 | sysfs_set_str(&a->info, NULL, "sync_action", "repair"); | |
306 | } | |
307 | } | |
549e9569 NB |
308 | } |
309 | ||
549e9569 | 310 | static void manage_new(struct mdstat_ent *mdstat, |
2a0bb19e DW |
311 | struct supertype *container, |
312 | struct active_array *victim) | |
549e9569 NB |
313 | { |
314 | /* A new array has appeared in this container. | |
315 | * Hopefully it is already recorded in the metadata. | |
316 | * Check, then create the new array to report it to | |
317 | * the monitor. | |
318 | */ | |
319 | ||
320 | struct active_array *new; | |
321 | struct mdinfo *mdi, *di; | |
cba0191b | 322 | char *inst; |
549e9569 NB |
323 | int i; |
324 | ||
325 | new = malloc(sizeof(*new)); | |
326 | ||
d52690ac NB |
327 | memset(new, 0, sizeof(*new)); |
328 | ||
549e9569 | 329 | new->devnum = mdstat->devnum; |
7e1432fb | 330 | strcpy(new->info.sys_name, devnum2devname(new->devnum)); |
549e9569 NB |
331 | |
332 | new->prev_state = new->curr_state = new->next_state = inactive; | |
333 | new->prev_action= new->curr_action= new->next_action= idle; | |
334 | ||
335 | new->container = container; | |
336 | ||
cba0191b | 337 | inst = &mdstat->metadata_version[10+strlen(container->devname)+1]; |
549e9569 NB |
338 | |
339 | mdi = sysfs_read(-1, new->devnum, | |
340 | GET_LEVEL|GET_CHUNK|GET_DISKS| | |
341 | GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE); | |
342 | if (!mdi) { | |
343 | /* Eeek. Cannot monitor this array. | |
344 | * Mark it to be ignored by setting container to NULL | |
345 | */ | |
346 | new->container = NULL; | |
2a0bb19e | 347 | replace_array(container, victim, new); |
549e9569 NB |
348 | return; |
349 | } | |
350 | ||
351 | new->info.array = mdi->array; | |
352 | ||
353 | for (i = 0; i < new->info.array.raid_disks; i++) { | |
354 | struct mdinfo *newd = malloc(sizeof(*newd)); | |
355 | ||
356 | for (di = mdi->devs; di; di = di->next) | |
357 | if (i == di->disk.raid_disk) | |
358 | break; | |
359 | ||
360 | if (di) { | |
361 | memcpy(newd, di, sizeof(*newd)); | |
362 | ||
549e9569 NB |
363 | newd->state_fd = sysfs_open(new->devnum, |
364 | newd->sys_name, | |
365 | "state"); | |
366 | ||
367 | newd->prev_state = read_dev_state(newd->state_fd); | |
6c3fb95c | 368 | newd->curr_state = newd->prev_state; |
549e9569 NB |
369 | } else { |
370 | newd->state_fd = -1; | |
7e1432fb NB |
371 | newd->disk.raid_disk = i; |
372 | newd->prev_state = DS_REMOVE; | |
373 | newd->curr_state = DS_REMOVE; | |
549e9569 | 374 | } |
7e1432fb | 375 | sprintf(newd->sys_name, "rd%d", i); |
549e9569 NB |
376 | newd->next = new->info.devs; |
377 | new->info.devs = newd; | |
378 | } | |
379 | new->action_fd = sysfs_open(new->devnum, NULL, "sync_action"); | |
380 | new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state"); | |
c052ba30 | 381 | new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start"); |
77402e51 | 382 | new->resync_start = 0; |
549e9569 | 383 | |
4fa5aef9 | 384 | sysfs_free(mdi); |
549e9569 NB |
385 | // finds and compares. |
386 | if (container->ss->open_new(container, new, inst) < 0) { | |
387 | // FIXME close all those files | |
388 | new->container = NULL; | |
2a0bb19e | 389 | replace_array(container, victim, new); |
549e9569 NB |
390 | return; |
391 | } | |
2a0bb19e | 392 | replace_array(container, victim, new); |
549e9569 NB |
393 | return; |
394 | } | |
395 | ||
5d19760d | 396 | void manage(struct mdstat_ent *mdstat, struct supertype *container) |
549e9569 NB |
397 | { |
398 | /* We have just read mdstat and need to compare it with | |
399 | * the known active arrays. | |
400 | * Arrays with the wrong metadata are ignored. | |
401 | */ | |
402 | ||
403 | for ( ; mdstat ; mdstat = mdstat->next) { | |
404 | struct active_array *a; | |
405 | if (mdstat->devnum == container->devnum) { | |
406 | manage_container(mdstat, container); | |
407 | continue; | |
408 | } | |
409 | if (mdstat->metadata_version == NULL || | |
410 | strncmp(mdstat->metadata_version, "external:/", 10) != 0 || | |
411 | strncmp(mdstat->metadata_version+10, container->devname, | |
412 | strlen(container->devname)) != 0 || | |
413 | mdstat->metadata_version[10+strlen(container->devname)] | |
414 | != '/') | |
415 | /* Not for this array */ | |
416 | continue; | |
417 | /* Looks like a member of this container */ | |
5d19760d | 418 | for (a = container->arrays; a; a = a->next) { |
549e9569 NB |
419 | if (mdstat->devnum == a->devnum) { |
420 | if (a->container) | |
421 | manage_member(mdstat, a); | |
422 | break; | |
423 | } | |
424 | } | |
2a0bb19e DW |
425 | if (a == NULL || !a->container) |
426 | manage_new(mdstat, container, a); | |
549e9569 NB |
427 | } |
428 | } | |
429 | ||
3e70c845 DW |
430 | static int handle_message(struct supertype *container, struct md_message *msg) |
431 | { | |
432 | int err; | |
433 | struct md_generic_cmd *cmd = msg->buf; | |
434 | ||
435 | if (!cmd) | |
436 | return 0; | |
437 | ||
438 | switch (cmd->action) { | |
439 | case md_action_remove_device: | |
440 | ||
441 | /* forward to the monitor */ | |
442 | active_cmd = cmd; | |
443 | write(container->mgr_pipe[1], &err, 1); | |
444 | read(container->mon_pipe[0], &err, 1); | |
445 | return err; | |
446 | ||
447 | default: | |
448 | return -1; | |
449 | } | |
450 | } | |
451 | ||
452 | void read_sock(struct supertype *container) | |
549e9569 NB |
453 | { |
454 | int fd; | |
b109d928 DW |
455 | struct md_message msg; |
456 | int terminate = 0; | |
457 | long fl; | |
458 | int tmo = 3; /* 3 second timeout before hanging up the socket */ | |
549e9569 | 459 | |
3e70c845 | 460 | fd = accept(container->sock, NULL, NULL); |
549e9569 NB |
461 | if (fd < 0) |
462 | return; | |
b109d928 DW |
463 | |
464 | fl = fcntl(fd, F_GETFL, 0); | |
465 | fl |= O_NONBLOCK; | |
466 | fcntl(fd, F_SETFL, fl); | |
467 | ||
468 | do { | |
3e70c845 DW |
469 | int err; |
470 | ||
b109d928 DW |
471 | msg.buf = NULL; |
472 | ||
473 | /* read and validate the message */ | |
474 | if (receive_message(fd, &msg, tmo) == 0) { | |
3e70c845 DW |
475 | err = handle_message(container, &msg); |
476 | if (!err) | |
477 | ack(fd, msg.seq, tmo); | |
478 | else | |
479 | nack(fd, err, tmo); | |
b109d928 DW |
480 | } else { |
481 | terminate = 1; | |
482 | nack(fd, -1, tmo); | |
483 | } | |
484 | ||
485 | if (msg.buf) | |
486 | free(msg.buf); | |
487 | } while (!terminate); | |
488 | ||
549e9569 NB |
489 | close(fd); |
490 | } | |
1ed3f387 NB |
491 | |
492 | static int woke = 0; | |
493 | void wake_me(int sig) | |
494 | { | |
495 | woke = 1; | |
496 | } | |
497 | ||
e0d6609f NB |
498 | int exit_now = 0; |
499 | int manager_ready = 0; | |
549e9569 NB |
500 | void do_manager(struct supertype *container) |
501 | { | |
502 | struct mdstat_ent *mdstat; | |
1ed3f387 NB |
503 | sigset_t block, orig; |
504 | ||
505 | sigemptyset(&block); | |
506 | sigaddset(&block, SIGUSR1); | |
507 | ||
508 | signal(SIGUSR1, wake_me); | |
549e9569 NB |
509 | |
510 | do { | |
1ed3f387 NB |
511 | woke = 0; |
512 | ||
e0d6609f NB |
513 | if (exit_now) |
514 | exit(0); | |
515 | ||
549e9569 NB |
516 | mdstat = mdstat_read(1, 0); |
517 | ||
5d19760d | 518 | manage(mdstat, container); |
549e9569 | 519 | |
3e70c845 | 520 | read_sock(container); |
549e9569 | 521 | |
4fa5aef9 DW |
522 | free_mdstat(mdstat); |
523 | ||
1ed3f387 NB |
524 | remove_old(); |
525 | ||
2e735d19 NB |
526 | check_update_queue(container); |
527 | ||
e0d6609f | 528 | manager_ready = 1; |
1ed3f387 NB |
529 | sigprocmask(SIG_SETMASK, &block, &orig); |
530 | if (woke == 0) | |
531 | mdstat_wait_fd(container->sock, &orig); | |
532 | sigprocmask(SIG_SETMASK, &orig, NULL); | |
549e9569 NB |
533 | } while(1); |
534 | } |