]>
Commit | Line | Data |
---|---|---|
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 | * | |
29 | * - assist with activating spares by opening relevant sysfs file. | |
30 | * | |
31 | * - Pass on metadata updates from external programs such as | |
32 | * mdadm creating a new array. | |
33 | * | |
34 | * This is most-messy. | |
35 | * It might involve adding a new array or changing the status of | |
36 | * a spare, or any reconfig that the kernel doesn't get involved in. | |
37 | * | |
38 | * The required updates are received via a named pipe. There will | |
39 | * be one named pipe for each container. Each message contains a | |
40 | * sync marker: 0x5a5aa5a5, A byte count, and the message. This is | |
41 | * passed to the metadata handler which will interpret and process it. | |
42 | * For 'DDF' messages are internal data blocks with the leading | |
43 | * 'magic number' signifying what sort of data it is. | |
44 | * | |
45 | */ | |
46 | ||
47 | /* | |
48 | * We select on /proc/mdstat and the named pipe. | |
49 | * We create new arrays or updated version of arrays and slip | |
50 | * them into the head of the list, then signal 'monitor' via a pipe write. | |
51 | * 'monitor' will notice and place the old array on a return list. | |
52 | * Metadata updates are placed on a queue just like they arrive | |
53 | * from the named pipe. | |
54 | * | |
55 | * When new arrays are found based on correct metadata string, we | |
56 | * need to identify them with an entry in the metadata. Maybe we require | |
57 | * the metadata to be mdX/NN when NN is the index into an appropriate table. | |
58 | * | |
59 | */ | |
60 | ||
61 | /* | |
62 | * List of tasks: | |
63 | * - Watch for spares to be added to the container, and write updated | |
64 | * metadata to them. | |
65 | * - Watch for new arrays using this container, confirm they match metadata | |
66 | * and if so, start monitoring them | |
67 | * - Watch for spares being added to monitored arrays. This shouldn't | |
68 | * happen, as we should do all the adding. Just remove them. | |
69 | * - Watch for change in raid-disks, chunk-size, etc. Update metadata and | |
70 | * start a reshape. | |
71 | */ | |
72 | #ifndef _GNU_SOURCE | |
73 | #define _GNU_SOURCE | |
74 | #endif | |
75 | #include "mdadm.h" | |
76 | #include "mdmon.h" | |
77 | #include <sys/socket.h> | |
78 | #include <signal.h> | |
79 | ||
80 | static void close_aa(struct active_array *aa) | |
81 | { | |
82 | struct mdinfo *d; | |
83 | ||
84 | for (d = aa->info.devs; d; d = d->next) | |
85 | close(d->state_fd); | |
86 | ||
87 | close(aa->action_fd); | |
88 | close(aa->info.state_fd); | |
89 | close(aa->resync_start_fd); | |
90 | } | |
91 | ||
92 | static void free_aa(struct active_array *aa) | |
93 | { | |
94 | /* Note that this doesn't close fds if they are being used | |
95 | * by a clone. ->container will be set for a clone | |
96 | */ | |
97 | if (!aa->container) | |
98 | close_aa(aa); | |
99 | while (aa->info.devs) { | |
100 | struct mdinfo *d = aa->info.devs; | |
101 | aa->info.devs = d->next; | |
102 | free(d); | |
103 | } | |
104 | free(aa); | |
105 | } | |
106 | ||
107 | static void write_wakeup(struct supertype *c) | |
108 | { | |
109 | static struct md_generic_cmd cmd = { .action = md_action_ping_monitor }; | |
110 | int err; | |
111 | ||
112 | active_cmd = &cmd; | |
113 | ||
114 | /* send the monitor thread a pointer to the ping action */ | |
115 | write(c->mgr_pipe[1], &err, 1); | |
116 | read(c->mon_pipe[0], &err, 1); | |
117 | } | |
118 | ||
119 | static void remove_old(void) | |
120 | { | |
121 | if (discard_this) { | |
122 | discard_this->next = NULL; | |
123 | free_aa(discard_this); | |
124 | if (pending_discard == discard_this) | |
125 | pending_discard = NULL; | |
126 | discard_this = NULL; | |
127 | } | |
128 | } | |
129 | ||
130 | static void replace_array(struct supertype *container, | |
131 | struct active_array *old, | |
132 | struct active_array *new) | |
133 | { | |
134 | /* To replace an array, we add it to the top of the list | |
135 | * marked with ->replaces to point to the original. | |
136 | * 'monitor' will take the original out of the list | |
137 | * and put it on 'discard_this'. We take it from there | |
138 | * and discard it. | |
139 | */ | |
140 | remove_old(); | |
141 | while (pending_discard) { | |
142 | write_wakeup(container); | |
143 | while (discard_this == NULL) | |
144 | sleep(1); | |
145 | remove_old(); | |
146 | } | |
147 | pending_discard = old; | |
148 | new->replaces = old; | |
149 | new->next = container->arrays; | |
150 | container->arrays = new; | |
151 | write_wakeup(container); | |
152 | } | |
153 | ||
154 | static void manage_container(struct mdstat_ent *mdstat, | |
155 | struct supertype *container) | |
156 | { | |
157 | /* The only thing of interest here is if a new device | |
158 | * has been added to the container. We add it to the | |
159 | * array ignoring any metadata on it. | |
160 | * FIXME should we look for compatible metadata and take hints | |
161 | * about spare assignment.... probably not. | |
162 | * | |
163 | */ | |
164 | if (mdstat->devcnt != container->devcnt) { | |
165 | /* read /sys/block/NAME/md/dev-??/block/dev to find out | |
166 | * what is there, and compare with container->info.devs | |
167 | * To see what is removed and what is added. | |
168 | * These need to be remove from, or added to, the array | |
169 | */ | |
170 | // FIXME | |
171 | container->devcnt = mdstat->devcnt; | |
172 | } | |
173 | } | |
174 | ||
175 | static void manage_member(struct mdstat_ent *mdstat, | |
176 | struct active_array *a) | |
177 | { | |
178 | /* Compare mdstat info with known state of member array. | |
179 | * We do not need to look for device state changes here, that | |
180 | * is dealt with by the monitor. | |
181 | * | |
182 | * We just look for changes which suggest that a reshape is | |
183 | * being requested. | |
184 | * Unfortunately decreases in raid_disks don't show up in | |
185 | * mdstat until the reshape completes FIXME. | |
186 | */ | |
187 | // FIXME | |
188 | a->info.array.raid_disks = mdstat->raid_disks; | |
189 | a->info.array.chunk_size = mdstat->chunk_size; | |
190 | // MORE | |
191 | ||
192 | } | |
193 | ||
194 | static void manage_new(struct mdstat_ent *mdstat, | |
195 | struct supertype *container, | |
196 | struct active_array *victim) | |
197 | { | |
198 | /* A new array has appeared in this container. | |
199 | * Hopefully it is already recorded in the metadata. | |
200 | * Check, then create the new array to report it to | |
201 | * the monitor. | |
202 | */ | |
203 | ||
204 | struct active_array *new; | |
205 | struct mdinfo *mdi, *di; | |
206 | char *n; | |
207 | int inst; | |
208 | int i; | |
209 | ||
210 | new = malloc(sizeof(*new)); | |
211 | ||
212 | memset(new, 0, sizeof(*new)); | |
213 | ||
214 | new->devnum = mdstat->devnum; | |
215 | ||
216 | new->prev_state = new->curr_state = new->next_state = inactive; | |
217 | new->prev_action= new->curr_action= new->next_action= idle; | |
218 | ||
219 | new->container = container; | |
220 | ||
221 | n = &mdstat->metadata_version[10+strlen(container->devname)+1]; | |
222 | inst = atoi(n); | |
223 | if (inst < 0) | |
224 | abort();//FIXME | |
225 | ||
226 | mdi = sysfs_read(-1, new->devnum, | |
227 | GET_LEVEL|GET_CHUNK|GET_DISKS| | |
228 | GET_DEVS|GET_OFFSET|GET_SIZE|GET_STATE); | |
229 | if (!mdi) { | |
230 | /* Eeek. Cannot monitor this array. | |
231 | * Mark it to be ignored by setting container to NULL | |
232 | */ | |
233 | new->container = NULL; | |
234 | replace_array(container, victim, new); | |
235 | return; | |
236 | } | |
237 | ||
238 | new->info.array = mdi->array; | |
239 | ||
240 | for (i = 0; i < new->info.array.raid_disks; i++) { | |
241 | struct mdinfo *newd = malloc(sizeof(*newd)); | |
242 | ||
243 | for (di = mdi->devs; di; di = di->next) | |
244 | if (i == di->disk.raid_disk) | |
245 | break; | |
246 | ||
247 | if (di) { | |
248 | memcpy(newd, di, sizeof(*newd)); | |
249 | ||
250 | sprintf(newd->sys_name, "rd%d", i); | |
251 | ||
252 | newd->state_fd = sysfs_open(new->devnum, | |
253 | newd->sys_name, | |
254 | "state"); | |
255 | ||
256 | newd->prev_state = read_dev_state(newd->state_fd); | |
257 | newd->curr_state = newd->curr_state; | |
258 | } else { | |
259 | newd->state_fd = -1; | |
260 | } | |
261 | newd->next = new->info.devs; | |
262 | new->info.devs = newd; | |
263 | } | |
264 | new->action_fd = sysfs_open(new->devnum, NULL, "sync_action"); | |
265 | new->info.state_fd = sysfs_open(new->devnum, NULL, "array_state"); | |
266 | new->resync_start_fd = sysfs_open(new->devnum, NULL, "resync_start"); | |
267 | new->resync_start = 0; | |
268 | ||
269 | sysfs_free(mdi); | |
270 | // finds and compares. | |
271 | if (container->ss->open_new(container, new, inst) < 0) { | |
272 | // FIXME close all those files | |
273 | new->container = NULL; | |
274 | replace_array(container, victim, new); | |
275 | return; | |
276 | } | |
277 | replace_array(container, victim, new); | |
278 | return; | |
279 | } | |
280 | ||
281 | void manage(struct mdstat_ent *mdstat, struct supertype *container) | |
282 | { | |
283 | /* We have just read mdstat and need to compare it with | |
284 | * the known active arrays. | |
285 | * Arrays with the wrong metadata are ignored. | |
286 | */ | |
287 | ||
288 | for ( ; mdstat ; mdstat = mdstat->next) { | |
289 | struct active_array *a; | |
290 | if (mdstat->devnum == container->devnum) { | |
291 | manage_container(mdstat, container); | |
292 | continue; | |
293 | } | |
294 | if (mdstat->metadata_version == NULL || | |
295 | strncmp(mdstat->metadata_version, "external:/", 10) != 0 || | |
296 | strncmp(mdstat->metadata_version+10, container->devname, | |
297 | strlen(container->devname)) != 0 || | |
298 | mdstat->metadata_version[10+strlen(container->devname)] | |
299 | != '/') | |
300 | /* Not for this array */ | |
301 | continue; | |
302 | /* Looks like a member of this container */ | |
303 | for (a = container->arrays; a; a = a->next) { | |
304 | if (mdstat->devnum == a->devnum) { | |
305 | if (a->container) | |
306 | manage_member(mdstat, a); | |
307 | break; | |
308 | } | |
309 | } | |
310 | if (a == NULL || !a->container) | |
311 | manage_new(mdstat, container, a); | |
312 | } | |
313 | } | |
314 | ||
315 | static int handle_message(struct supertype *container, struct md_message *msg) | |
316 | { | |
317 | int err; | |
318 | struct md_generic_cmd *cmd = msg->buf; | |
319 | ||
320 | if (!cmd) | |
321 | return 0; | |
322 | ||
323 | switch (cmd->action) { | |
324 | case md_action_remove_device: | |
325 | ||
326 | /* forward to the monitor */ | |
327 | active_cmd = cmd; | |
328 | write(container->mgr_pipe[1], &err, 1); | |
329 | read(container->mon_pipe[0], &err, 1); | |
330 | return err; | |
331 | ||
332 | default: | |
333 | return -1; | |
334 | } | |
335 | } | |
336 | ||
337 | void read_sock(struct supertype *container) | |
338 | { | |
339 | int fd; | |
340 | struct md_message msg; | |
341 | int terminate = 0; | |
342 | long fl; | |
343 | int tmo = 3; /* 3 second timeout before hanging up the socket */ | |
344 | ||
345 | fd = accept(container->sock, NULL, NULL); | |
346 | if (fd < 0) | |
347 | return; | |
348 | ||
349 | fl = fcntl(fd, F_GETFL, 0); | |
350 | fl |= O_NONBLOCK; | |
351 | fcntl(fd, F_SETFL, fl); | |
352 | ||
353 | do { | |
354 | int err; | |
355 | ||
356 | msg.buf = NULL; | |
357 | ||
358 | /* read and validate the message */ | |
359 | if (receive_message(fd, &msg, tmo) == 0) { | |
360 | err = handle_message(container, &msg); | |
361 | if (!err) | |
362 | ack(fd, msg.seq, tmo); | |
363 | else | |
364 | nack(fd, err, tmo); | |
365 | } else { | |
366 | terminate = 1; | |
367 | nack(fd, -1, tmo); | |
368 | } | |
369 | ||
370 | if (msg.buf) | |
371 | free(msg.buf); | |
372 | } while (!terminate); | |
373 | ||
374 | close(fd); | |
375 | } | |
376 | ||
377 | static int woke = 0; | |
378 | void wake_me(int sig) | |
379 | { | |
380 | woke = 1; | |
381 | } | |
382 | ||
383 | int exit_now = 0; | |
384 | int manager_ready = 0; | |
385 | void do_manager(struct supertype *container) | |
386 | { | |
387 | struct mdstat_ent *mdstat; | |
388 | sigset_t block, orig; | |
389 | ||
390 | sigemptyset(&block); | |
391 | sigaddset(&block, SIGUSR1); | |
392 | ||
393 | signal(SIGUSR1, wake_me); | |
394 | ||
395 | do { | |
396 | woke = 0; | |
397 | ||
398 | if (exit_now) | |
399 | exit(0); | |
400 | ||
401 | mdstat = mdstat_read(1, 0); | |
402 | ||
403 | manage(mdstat, container); | |
404 | ||
405 | read_sock(container); | |
406 | ||
407 | free_mdstat(mdstat); | |
408 | ||
409 | remove_old(); | |
410 | ||
411 | manager_ready = 1; | |
412 | sigprocmask(SIG_SETMASK, &block, &orig); | |
413 | if (woke == 0) | |
414 | mdstat_wait_fd(container->sock, &orig); | |
415 | sigprocmask(SIG_SETMASK, &orig, NULL); | |
416 | } while(1); | |
417 | } |