]>
git.ipfire.org Git - thirdparty/mdadm.git/blob - managemon.c
3 * The management thread for monitoring active md arrays.
4 * This thread does things which might block such as memory
8 * - Find out about new arrays in this container.
9 * Allocate the data structures and open the files.
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.
16 * - Respond to reshape requests by allocating new data structures
17 * and opening new files.
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.
26 * - When a device is added to the container, we add it to the metadata
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.
34 * Check if number of failed devices suggests recovery is needed, and
36 * Ask metadata to allocate a spare device
37 * Add device as not in_sync and give a role
39 * Open sysfs files and pass to monitor.
40 * Make sure that monitor Starts recovery....
42 * - Pass on metadata updates from external programs such as
43 * mdadm creating a new array.
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.
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.
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.
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.
74 * - Watch for spares to be added to the container, and write updated
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
88 #include <sys/socket.h>
91 static void close_aa(struct active_array
*aa
)
95 for (d
= aa
->info
.devs
; d
; d
= d
->next
)
99 close(aa
->info
.state_fd
);
100 close(aa
->resync_start_fd
);
103 static void free_aa(struct active_array
*aa
)
105 /* Note that this doesn't close fds if they are being used
106 * by a clone. ->container will be set for a clone
110 while (aa
->info
.devs
) {
111 struct mdinfo
*d
= aa
->info
.devs
;
112 aa
->info
.devs
= d
->next
;
118 static struct active_array
*duplicate_aa(struct active_array
*aa
)
120 struct active_array
*newa
= malloc(sizeof(*newa
));
121 struct mdinfo
**dp1
, **dp2
;
125 newa
->replaces
= NULL
;
126 newa
->info
.next
= NULL
;
128 dp2
= &newa
->info
.devs
;
130 for (dp1
= &aa
->info
.devs
; *dp1
; dp1
= &(*dp1
)->next
) {
132 if ((*dp1
)->state_fd
< 0)
135 d
= malloc(sizeof(*d
));
144 static void write_wakeup(struct supertype
*c
)
146 static struct md_generic_cmd cmd
= { .action
= md_action_ping_monitor
};
151 /* send the monitor thread a pointer to the ping action */
152 write(c
->mgr_pipe
[1], &err
, 1);
153 read(c
->mon_pipe
[0], &err
, 1);
156 static void remove_old(void)
159 discard_this
->next
= NULL
;
160 free_aa(discard_this
);
161 if (pending_discard
== discard_this
)
162 pending_discard
= NULL
;
167 static void replace_array(struct supertype
*container
,
168 struct active_array
*old
,
169 struct active_array
*new)
171 /* To replace an array, we add it to the top of the list
172 * marked with ->replaces to point to the original.
173 * 'monitor' will take the original out of the list
174 * and put it on 'discard_this'. We take it from there
178 while (pending_discard
) {
179 write_wakeup(container
);
180 while (discard_this
== NULL
)
184 pending_discard
= old
;
186 new->next
= container
->arrays
;
187 container
->arrays
= new;
188 write_wakeup(container
);
191 struct metadata_update
*update_queue
= NULL
;
192 struct metadata_update
*update_queue_handled
= NULL
;
193 struct metadata_update
*update_queue_pending
= NULL
;
195 void check_update_queue(struct supertype
*container
)
197 while (update_queue_handled
) {
198 struct metadata_update
*this = update_queue_handled
;
199 update_queue_handled
= this->next
;
203 if (update_queue
== NULL
&&
204 update_queue_pending
) {
205 update_queue
= update_queue_pending
;
206 update_queue_pending
= NULL
;
207 write_wakeup(container
);
211 static void queue_metadata_update(struct metadata_update
*mu
)
213 struct metadata_update
**qp
;
215 qp
= &update_queue_pending
;
217 qp
= & ((*qp
)->next
);
221 void wait_update_handled(void)
223 /* Wait for any pending update to be handled by monitor.
224 * i.e. wait until update_queue is NULL
230 static void manage_container(struct mdstat_ent
*mdstat
,
231 struct supertype
*container
)
233 /* The only thing of interest here is if a new device
234 * has been added to the container. We add it to the
235 * array ignoring any metadata on it.
236 * FIXME should we look for compatible metadata and take hints
237 * about spare assignment.... probably not.
239 if (mdstat
->devcnt
!= container
->devcnt
) {
240 /* read /sys/block/NAME/md/dev-??/block/dev to find out
241 * what is there, and compare with container->info.devs
242 * To see what is removed and what is added.
243 * These need to be remove from, or added to, the array
246 container
->devcnt
= mdstat
->devcnt
;
250 static void manage_member(struct mdstat_ent
*mdstat
,
251 struct active_array
*a
)
253 /* Compare mdstat info with known state of member array.
254 * We do not need to look for device state changes here, that
255 * is dealt with by the monitor.
257 * We just look for changes which suggest that a reshape is
259 * Unfortunately decreases in raid_disks don't show up in
260 * mdstat until the reshape completes FIXME.
262 * Actually, we also want to handle degraded arrays here by
263 * trying to find and assign a spare.
264 * We do that whenever the monitor tells us too.
267 a
->info
.array
.raid_disks
= mdstat
->raid_disks
;
268 a
->info
.array
.chunk_size
= mdstat
->chunk_size
;
271 if (a
->check_degraded
) {
272 struct metadata_update
*updates
= NULL
;
273 struct mdinfo
*newdev
;
274 struct active_array
*newa
;
275 wait_update_handled();
276 a
->check_degraded
= 0;
278 /* The array may not be degraded, this is just a good time
281 newdev
= a
->container
->ss
->activate_spare(a
, &updates
);
284 /* Cool, we can add a device or several. */
285 newa
= duplicate_aa(a
);
286 /* suspend recovery - maybe not needed */
288 /* Add device to array and set offset/size/slot.
289 * and open files for each newdev */
290 for (d
= newdev
; d
; d
= d
->next
) {
292 if (sysfs_add_disk(&newa
->info
, d
))
294 newd
= newa
->info
.devs
;
295 newd
->state_fd
= sysfs_open(a
->devnum
,
299 = read_dev_state(newd
->state_fd
);
300 newd
->curr_state
= newd
->prev_state
;
302 queue_metadata_update(updates
);
303 replace_array(a
->container
, a
, newa
);
304 sysfs_set_str(&a
->info
, NULL
, "sync_action", "repair");
309 static void manage_new(struct mdstat_ent
*mdstat
,
310 struct supertype
*container
,
311 struct active_array
*victim
)
313 /* A new array has appeared in this container.
314 * Hopefully it is already recorded in the metadata.
315 * Check, then create the new array to report it to
319 struct active_array
*new;
320 struct mdinfo
*mdi
, *di
;
324 new = malloc(sizeof(*new));
326 memset(new, 0, sizeof(*new));
328 new->devnum
= mdstat
->devnum
;
330 new->prev_state
= new->curr_state
= new->next_state
= inactive
;
331 new->prev_action
= new->curr_action
= new->next_action
= idle
;
333 new->container
= container
;
335 inst
= &mdstat
->metadata_version
[10+strlen(container
->devname
)+1];
337 mdi
= sysfs_read(-1, new->devnum
,
338 GET_LEVEL
|GET_CHUNK
|GET_DISKS
|
339 GET_DEVS
|GET_OFFSET
|GET_SIZE
|GET_STATE
);
341 /* Eeek. Cannot monitor this array.
342 * Mark it to be ignored by setting container to NULL
344 new->container
= NULL
;
345 replace_array(container
, victim
, new);
349 new->info
.array
= mdi
->array
;
351 for (i
= 0; i
< new->info
.array
.raid_disks
; i
++) {
352 struct mdinfo
*newd
= malloc(sizeof(*newd
));
354 for (di
= mdi
->devs
; di
; di
= di
->next
)
355 if (i
== di
->disk
.raid_disk
)
359 memcpy(newd
, di
, sizeof(*newd
));
361 sprintf(newd
->sys_name
, "rd%d", i
);
363 newd
->state_fd
= sysfs_open(new->devnum
,
367 newd
->prev_state
= read_dev_state(newd
->state_fd
);
368 newd
->curr_state
= newd
->prev_state
;
372 newd
->next
= new->info
.devs
;
373 new->info
.devs
= newd
;
375 new->action_fd
= sysfs_open(new->devnum
, NULL
, "sync_action");
376 new->info
.state_fd
= sysfs_open(new->devnum
, NULL
, "array_state");
377 new->resync_start_fd
= sysfs_open(new->devnum
, NULL
, "resync_start");
378 new->resync_start
= 0;
381 // finds and compares.
382 if (container
->ss
->open_new(container
, new, inst
) < 0) {
383 // FIXME close all those files
384 new->container
= NULL
;
385 replace_array(container
, victim
, new);
388 replace_array(container
, victim
, new);
392 void manage(struct mdstat_ent
*mdstat
, struct supertype
*container
)
394 /* We have just read mdstat and need to compare it with
395 * the known active arrays.
396 * Arrays with the wrong metadata are ignored.
399 for ( ; mdstat
; mdstat
= mdstat
->next
) {
400 struct active_array
*a
;
401 if (mdstat
->devnum
== container
->devnum
) {
402 manage_container(mdstat
, container
);
405 if (mdstat
->metadata_version
== NULL
||
406 strncmp(mdstat
->metadata_version
, "external:/", 10) != 0 ||
407 strncmp(mdstat
->metadata_version
+10, container
->devname
,
408 strlen(container
->devname
)) != 0 ||
409 mdstat
->metadata_version
[10+strlen(container
->devname
)]
411 /* Not for this array */
413 /* Looks like a member of this container */
414 for (a
= container
->arrays
; a
; a
= a
->next
) {
415 if (mdstat
->devnum
== a
->devnum
) {
417 manage_member(mdstat
, a
);
421 if (a
== NULL
|| !a
->container
)
422 manage_new(mdstat
, container
, a
);
426 static int handle_message(struct supertype
*container
, struct md_message
*msg
)
429 struct md_generic_cmd
*cmd
= msg
->buf
;
434 switch (cmd
->action
) {
435 case md_action_remove_device
:
437 /* forward to the monitor */
439 write(container
->mgr_pipe
[1], &err
, 1);
440 read(container
->mon_pipe
[0], &err
, 1);
448 void read_sock(struct supertype
*container
)
451 struct md_message msg
;
454 int tmo
= 3; /* 3 second timeout before hanging up the socket */
456 fd
= accept(container
->sock
, NULL
, NULL
);
460 fl
= fcntl(fd
, F_GETFL
, 0);
462 fcntl(fd
, F_SETFL
, fl
);
469 /* read and validate the message */
470 if (receive_message(fd
, &msg
, tmo
) == 0) {
471 err
= handle_message(container
, &msg
);
473 ack(fd
, msg
.seq
, tmo
);
483 } while (!terminate
);
489 void wake_me(int sig
)
495 int manager_ready
= 0;
496 void do_manager(struct supertype
*container
)
498 struct mdstat_ent
*mdstat
;
499 sigset_t block
, orig
;
502 sigaddset(&block
, SIGUSR1
);
504 signal(SIGUSR1
, wake_me
);
512 mdstat
= mdstat_read(1, 0);
514 manage(mdstat
, container
);
516 read_sock(container
);
522 check_update_queue(container
);
525 sigprocmask(SIG_SETMASK
, &block
, &orig
);
527 mdstat_wait_fd(container
->sock
, &orig
);
528 sigprocmask(SIG_SETMASK
, &orig
, NULL
);