]>
Commit | Line | Data |
---|---|---|
a54d5262 DW |
1 | /* |
2 | * mdmon - monitor external metadata arrays | |
3 | * | |
e736b623 N |
4 | * Copyright (C) 2007-2009 Neil Brown <neilb@suse.de> |
5 | * Copyright (C) 2007-2009 Intel Corporation | |
a54d5262 DW |
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 | */ | |
549e9569 NB |
20 | |
21 | /* | |
22 | * md array manager. | |
23 | * When md arrays have user-space managed metadata, this is the program | |
24 | * that does the managing. | |
25 | * | |
26 | * Given one argument: the name of the array (e.g. /dev/md0) that is | |
27 | * the container. | |
28 | * We fork off a helper that runs high priority and mlocked. It responds to | |
29 | * device failures and other events that might stop writeout, or that are | |
30 | * trivial to deal with. | |
31 | * The main thread then watches for new arrays being created in the container | |
32 | * and starts monitoring them too ... along with a few other tasks. | |
33 | * | |
34 | * The main thread communicates with the priority thread by writing over | |
35 | * a pipe. | |
36 | * Separate programs can communicate with the main thread via Unix-domain | |
37 | * socket. | |
38 | * The two threads share address space and open file table. | |
39 | * | |
40 | */ | |
41 | ||
42 | #ifndef _GNU_SOURCE | |
43 | #define _GNU_SOURCE | |
44 | #endif | |
45 | ||
46 | #include <unistd.h> | |
47 | #include <stdlib.h> | |
4d43913c | 48 | #include <sys/types.h> |
549e9569 NB |
49 | #include <sys/stat.h> |
50 | #include <sys/socket.h> | |
51 | #include <sys/un.h> | |
52 | #include <sys/mman.h> | |
4d43913c | 53 | #include <sys/syscall.h> |
9fe32043 | 54 | #include <sys/wait.h> |
549e9569 NB |
55 | #include <stdio.h> |
56 | #include <errno.h> | |
57 | #include <string.h> | |
58 | #include <fcntl.h> | |
b109d928 | 59 | #include <signal.h> |
13047e4c | 60 | #include <dirent.h> |
549e9569 NB |
61 | |
62 | #include <sched.h> | |
63 | ||
64 | #include "mdadm.h" | |
65 | #include "mdmon.h" | |
66 | ||
549e9569 NB |
67 | struct active_array *discard_this; |
68 | struct active_array *pending_discard; | |
4d43913c NB |
69 | |
70 | int mon_tid, mgr_tid; | |
549e9569 | 71 | |
6144ed44 DW |
72 | int sigterm; |
73 | ||
549e9569 NB |
74 | int run_child(void *v) |
75 | { | |
76 | struct supertype *c = v; | |
1ed3f387 | 77 | |
549e9569 NB |
78 | do_monitor(c); |
79 | return 0; | |
80 | } | |
81 | ||
97f734fd N |
82 | #ifdef __ia64__ |
83 | int __clone2(int (*fn)(void *), | |
84 | void *child_stack_base, size_t stack_size, | |
85 | int flags, void *arg, ... | |
86 | /* pid_t *pid, struct user_desc *tls, pid_t *ctid */ ); | |
87 | #endif | |
88 | int clone_monitor(struct supertype *container) | |
549e9569 | 89 | { |
549e9569 | 90 | static char stack[4096]; |
549e9569 | 91 | |
97f734fd N |
92 | #ifdef __ia64__ |
93 | mon_tid = __clone2(run_child, stack, sizeof(stack), | |
94 | CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD, | |
95 | container); | |
96 | #else | |
2cc98f9e | 97 | mon_tid = clone(run_child, stack+4096-64, |
549e9569 NB |
98 | CLONE_FS|CLONE_FILES|CLONE_VM|CLONE_SIGHAND|CLONE_THREAD, |
99 | container); | |
97f734fd | 100 | #endif |
3e70c845 | 101 | |
4d43913c | 102 | mgr_tid = syscall(SYS_gettid); |
2cc98f9e DW |
103 | |
104 | return mon_tid; | |
549e9569 NB |
105 | } |
106 | ||
ce744c97 DW |
107 | static struct superswitch *find_metadata_methods(char *vers) |
108 | { | |
109 | if (strcmp(vers, "ddf") == 0) | |
110 | return &super_ddf; | |
111 | if (strcmp(vers, "imsm") == 0) | |
112 | return &super_imsm; | |
113 | return NULL; | |
114 | } | |
115 | ||
96a8270d DW |
116 | static int test_pidfile(char *devname) |
117 | { | |
118 | char path[100]; | |
119 | struct stat st; | |
120 | ||
121 | sprintf(path, "/var/run/mdadm/%s.pid", devname); | |
122 | return stat(path, &st); | |
123 | } | |
549e9569 | 124 | |
295646b3 | 125 | int make_pidfile(char *devname, int o_excl) |
549e9569 NB |
126 | { |
127 | char path[100]; | |
128 | char pid[10]; | |
129 | int fd; | |
3d2c4fc7 DW |
130 | int n; |
131 | ||
6144ed44 DW |
132 | if (sigterm) |
133 | return -1; | |
134 | ||
549e9569 NB |
135 | sprintf(path, "/var/run/mdadm/%s.pid", devname); |
136 | ||
b109d928 | 137 | fd = open(path, O_RDWR|O_CREAT|o_excl, 0600); |
549e9569 | 138 | if (fd < 0) |
295646b3 | 139 | return -errno; |
549e9569 | 140 | sprintf(pid, "%d\n", getpid()); |
3d2c4fc7 | 141 | n = write(fd, pid, strlen(pid)); |
549e9569 | 142 | close(fd); |
3d2c4fc7 DW |
143 | if (n < 0) |
144 | return -errno; | |
549e9569 NB |
145 | return 0; |
146 | } | |
147 | ||
883a6142 DW |
148 | int is_container_member(struct mdstat_ent *mdstat, char *container) |
149 | { | |
150 | if (mdstat->metadata_version == NULL || | |
151 | strncmp(mdstat->metadata_version, "external:", 9) != 0 || | |
152 | !is_subarray(mdstat->metadata_version+9) || | |
153 | strncmp(mdstat->metadata_version+10, container, strlen(container)) != 0 || | |
154 | mdstat->metadata_version[10+strlen(container)] != '/') | |
155 | return 0; | |
156 | ||
157 | return 1; | |
158 | } | |
159 | ||
96a8270d | 160 | pid_t devname2mdmon(char *devname) |
b109d928 DW |
161 | { |
162 | char buf[100]; | |
96a8270d | 163 | pid_t pid = -1; |
b109d928 | 164 | int fd; |
b109d928 DW |
165 | |
166 | sprintf(buf, "/var/run/mdadm/%s.pid", devname); | |
96a8270d | 167 | fd = open(buf, O_RDONLY|O_NOATIME); |
b109d928 | 168 | if (fd < 0) |
96a8270d | 169 | return -1; |
b109d928 | 170 | |
96a8270d DW |
171 | if (read(fd, buf, sizeof(buf)) > 0) |
172 | sscanf(buf, "%d\n", &pid); | |
b109d928 | 173 | close(fd); |
96a8270d DW |
174 | |
175 | return pid; | |
176 | } | |
177 | ||
9f1da824 | 178 | static void try_kill_monitor(pid_t pid, char *devname, int sock) |
96a8270d DW |
179 | { |
180 | char buf[100]; | |
181 | int fd; | |
182 | struct mdstat_ent *mdstat; | |
b109d928 | 183 | |
8aae4219 DW |
184 | /* first rule of survival... don't off yourself */ |
185 | if (pid == getpid()) | |
186 | return; | |
187 | ||
b109d928 DW |
188 | /* kill this process if it is mdmon */ |
189 | sprintf(buf, "/proc/%lu/cmdline", (unsigned long) pid); | |
190 | fd = open(buf, O_RDONLY); | |
191 | if (fd < 0) | |
192 | return; | |
193 | ||
194 | if (read(fd, buf, sizeof(buf)) < 0) { | |
195 | close(fd); | |
196 | return; | |
197 | } | |
198 | ||
883a6142 DW |
199 | if (!strstr(buf, "mdmon")) |
200 | return; | |
201 | ||
202 | kill(pid, SIGTERM); | |
203 | ||
204 | mdstat = mdstat_read(0, 0); | |
205 | for ( ; mdstat; mdstat = mdstat->next) | |
206 | if (is_container_member(mdstat, devname)) { | |
207 | sprintf(buf, "/dev/%s", mdstat->dev); | |
9f1da824 | 208 | WaitClean(buf, sock, 0); |
883a6142 DW |
209 | } |
210 | free_mdstat(mdstat); | |
b109d928 DW |
211 | } |
212 | ||
e0d6609f NB |
213 | void remove_pidfile(char *devname) |
214 | { | |
215 | char buf[100]; | |
216 | ||
6144ed44 DW |
217 | if (sigterm) |
218 | return; | |
219 | ||
e0d6609f NB |
220 | sprintf(buf, "/var/run/mdadm/%s.pid", devname); |
221 | unlink(buf); | |
57752795 N |
222 | sprintf(buf, "/var/run/mdadm/%s.sock", devname); |
223 | unlink(buf); | |
e0d6609f NB |
224 | } |
225 | ||
295646b3 | 226 | int make_control_sock(char *devname) |
549e9569 NB |
227 | { |
228 | char path[100]; | |
229 | int sfd; | |
230 | long fl; | |
231 | struct sockaddr_un addr; | |
232 | ||
6144ed44 DW |
233 | if (sigterm) |
234 | return -1; | |
235 | ||
549e9569 NB |
236 | sprintf(path, "/var/run/mdadm/%s.sock", devname); |
237 | unlink(path); | |
238 | sfd = socket(PF_LOCAL, SOCK_STREAM, 0); | |
239 | if (sfd < 0) | |
240 | return -1; | |
241 | ||
242 | addr.sun_family = PF_LOCAL; | |
243 | strcpy(addr.sun_path, path); | |
244 | if (bind(sfd, &addr, sizeof(addr)) < 0) { | |
245 | close(sfd); | |
246 | return -1; | |
247 | } | |
248 | listen(sfd, 10); | |
249 | fl = fcntl(sfd, F_GETFL, 0); | |
250 | fl |= O_NONBLOCK; | |
251 | fcntl(sfd, F_SETFL, fl); | |
252 | return sfd; | |
253 | } | |
254 | ||
295646b3 DW |
255 | int socket_hup_requested; |
256 | static void hup(int sig) | |
257 | { | |
258 | socket_hup_requested = 1; | |
259 | } | |
260 | ||
6144ed44 DW |
261 | static void term(int sig) |
262 | { | |
263 | sigterm = 1; | |
264 | } | |
265 | ||
4d43913c NB |
266 | static void wake_me(int sig) |
267 | { | |
268 | ||
269 | } | |
270 | ||
16ddab0d DW |
271 | /* if we are debugging and starting mdmon by hand then don't fork */ |
272 | static int do_fork(void) | |
273 | { | |
274 | #ifdef DEBUG | |
40ebbb9c | 275 | if (check_env("MDADM_NO_MDMON")) |
16ddab0d DW |
276 | return 0; |
277 | #endif | |
278 | ||
279 | return 1; | |
280 | } | |
281 | ||
13047e4c DW |
282 | void usage(void) |
283 | { | |
5746141e | 284 | fprintf(stderr, "Usage: mdmon /device/name/for/container [target_dir]\n"); |
13047e4c DW |
285 | exit(2); |
286 | } | |
16ddab0d | 287 | |
1ffd2840 DW |
288 | int mdmon(char *devname, int devnum, int scan, char *switchroot); |
289 | ||
549e9569 NB |
290 | int main(int argc, char *argv[]) |
291 | { | |
13047e4c DW |
292 | char *container_name = NULL; |
293 | char *switchroot = NULL; | |
e8a70c89 N |
294 | int devnum; |
295 | char *devname; | |
1ffd2840 DW |
296 | int scan = 0; |
297 | int status = 0; | |
13047e4c DW |
298 | |
299 | switch (argc) { | |
5746141e DW |
300 | case 3: |
301 | switchroot = argv[2]; | |
13047e4c DW |
302 | case 2: |
303 | container_name = argv[1]; | |
304 | break; | |
13047e4c DW |
305 | default: |
306 | usage(); | |
549e9569 | 307 | } |
13047e4c | 308 | |
1ffd2840 DW |
309 | if (strcmp(container_name, "/proc/mdstat") == 0) { |
310 | struct mdstat_ent *mdstat, *e; | |
311 | ||
312 | /* launch an mdmon instance for each container found */ | |
313 | scan = 1; | |
314 | mdstat = mdstat_read(0, 0); | |
315 | for (e = mdstat; e; e = e->next) { | |
316 | if (strncmp(e->metadata_version, "external:", 9) == 0 && | |
317 | !is_subarray(&e->metadata_version[9])) { | |
318 | devname = devnum2devname(e->devnum); | |
1b34f519 DW |
319 | /* update cmdline so this mdmon instance can be |
320 | * distinguished from others in a call to ps(1) | |
321 | */ | |
322 | if (strlen(devname) <= strlen(container_name)) { | |
323 | memset(container_name, 0, strlen(container_name)); | |
324 | sprintf(container_name, "%s", devname); | |
325 | } | |
1ffd2840 DW |
326 | status |= mdmon(devname, e->devnum, scan, |
327 | switchroot); | |
328 | } | |
329 | } | |
330 | free_mdstat(mdstat); | |
331 | ||
332 | return status; | |
333 | } else if (strncmp(container_name, "md", 2) == 0) { | |
6f4098a6 DW |
334 | devnum = devname2devnum(container_name); |
335 | devname = devnum2devname(devnum); | |
336 | if (strcmp(container_name, devname) != 0) | |
337 | devname = NULL; | |
338 | } else { | |
339 | struct stat st; | |
340 | ||
341 | devnum = NoMdDev; | |
342 | if (stat(container_name, &st) == 0) | |
343 | devnum = stat2devnum(&st); | |
344 | if (devnum == NoMdDev) | |
345 | devname = NULL; | |
346 | else | |
347 | devname = devnum2devname(devnum); | |
348 | } | |
349 | ||
350 | if (!devname) { | |
e8a70c89 N |
351 | fprintf(stderr, "mdmon: %s is not a valid md device name\n", |
352 | container_name); | |
353 | exit(1); | |
354 | } | |
1ffd2840 DW |
355 | return mdmon(devname, devnum, scan, switchroot); |
356 | } | |
357 | ||
358 | int mdmon(char *devname, int devnum, int scan, char *switchroot) | |
359 | { | |
360 | int mdfd; | |
361 | struct mdinfo *mdi, *di; | |
362 | struct supertype *container; | |
363 | sigset_t set; | |
364 | struct sigaction act; | |
365 | int pfd[2]; | |
366 | int status; | |
367 | int ignore; | |
96a8270d | 368 | pid_t victim = -1; |
9f1da824 | 369 | int victim_sock = -1; |
1ffd2840 | 370 | |
5746141e DW |
371 | dprintf("starting mdmon for %s in %s\n", |
372 | devname, switchroot ? : "/"); | |
b928b5a0 DW |
373 | |
374 | /* try to spawn mdmon instances from the target file system */ | |
375 | if (switchroot && strcmp(switchroot, "/") != 0) { | |
376 | char path[1024]; | |
377 | pid_t pid; | |
378 | ||
379 | sprintf(path, "%s/sbin/mdmon", switchroot); | |
380 | switch (fork()) { | |
381 | case 0: | |
382 | execl(path, "mdmon", devname, NULL); | |
383 | exit(1); | |
384 | case -1: | |
385 | return 1; | |
386 | default: | |
387 | pid = wait(&status); | |
388 | if (pid > -1 && WIFEXITED(status) && | |
389 | WEXITSTATUS(status) == 0) | |
390 | return 0; | |
391 | else | |
392 | return 1; | |
393 | } | |
394 | } | |
395 | ||
e8a70c89 | 396 | mdfd = open_dev(devnum); |
549e9569 | 397 | if (mdfd < 0) { |
1ffd2840 | 398 | fprintf(stderr, "mdmon: %s: %s\n", devname, |
549e9569 | 399 | strerror(errno)); |
1ffd2840 | 400 | return 1; |
549e9569 NB |
401 | } |
402 | if (md_get_version(mdfd) < 0) { | |
13047e4c | 403 | fprintf(stderr, "mdmon: %s: Not an md device\n", |
1ffd2840 DW |
404 | devname); |
405 | return 1; | |
549e9569 NB |
406 | } |
407 | ||
9fe32043 | 408 | /* Fork, and have the child tell us when they are ready */ |
1ffd2840 | 409 | if (do_fork() || scan) { |
3d2c4fc7 DW |
410 | if (pipe(pfd) != 0) { |
411 | fprintf(stderr, "mdmon: failed to create pipe\n"); | |
1ffd2840 | 412 | return 1; |
3d2c4fc7 | 413 | } |
16ddab0d DW |
414 | switch(fork()) { |
415 | case -1: | |
416 | fprintf(stderr, "mdmon: failed to fork: %s\n", | |
417 | strerror(errno)); | |
1ffd2840 | 418 | return 1; |
16ddab0d DW |
419 | case 0: /* child */ |
420 | close(pfd[0]); | |
421 | break; | |
422 | default: /* parent */ | |
423 | close(pfd[1]); | |
424 | if (read(pfd[0], &status, sizeof(status)) != sizeof(status)) { | |
425 | wait(&status); | |
426 | status = WEXITSTATUS(status); | |
427 | } | |
1ffd2840 | 428 | return status; |
9fe32043 | 429 | } |
16ddab0d DW |
430 | } else |
431 | pfd[0] = pfd[1] = -1; | |
549e9569 | 432 | |
f5df5d69 | 433 | container = calloc(1, sizeof(*container)); |
e8a70c89 N |
434 | container->devnum = devnum; |
435 | container->devname = devname; | |
13047e4c | 436 | container->arrays = NULL; |
c1363b40 | 437 | container->subarray[0] = 0; |
96a8270d | 438 | container->sock = -1; |
13047e4c DW |
439 | |
440 | if (!container->devname) { | |
441 | fprintf(stderr, "mdmon: failed to allocate container name string\n"); | |
442 | exit(3); | |
443 | } | |
444 | ||
445 | mdi = sysfs_read(mdfd, container->devnum, | |
7da80e6f | 446 | GET_VERSION|GET_LEVEL|GET_DEVS|SKIP_GONE_DEVS); |
13047e4c DW |
447 | |
448 | if (!mdi) { | |
449 | fprintf(stderr, "mdmon: failed to load sysfs info for %s\n", | |
450 | container->devname); | |
451 | exit(3); | |
452 | } | |
453 | if (mdi->array.level != UnSet) { | |
454 | fprintf(stderr, "mdmon: %s is not a container - cannot monitor\n", | |
1ffd2840 | 455 | devname); |
13047e4c DW |
456 | exit(3); |
457 | } | |
458 | if (mdi->array.major_version != -1 || | |
459 | mdi->array.minor_version != -2) { | |
460 | fprintf(stderr, "mdmon: %s does not use external metadata - cannot monitor\n", | |
1ffd2840 | 461 | devname); |
13047e4c DW |
462 | exit(3); |
463 | } | |
464 | ||
465 | container->ss = find_metadata_methods(mdi->text_version); | |
466 | if (container->ss == NULL) { | |
467 | fprintf(stderr, "mdmon: %s uses unknown metadata: %s\n", | |
1ffd2840 | 468 | devname, mdi->text_version); |
13047e4c DW |
469 | exit(3); |
470 | } | |
471 | ||
472 | container->devs = NULL; | |
473 | for (di = mdi->devs; di; di = di->next) { | |
474 | struct mdinfo *cd = malloc(sizeof(*cd)); | |
475 | *cd = *di; | |
476 | cd->next = container->devs; | |
477 | container->devs = cd; | |
478 | } | |
479 | sysfs_free(mdi); | |
549e9569 | 480 | |
883a6142 DW |
481 | /* SIGUSR is sent between parent and child. So both block it |
482 | * and enable it only with pselect. | |
483 | */ | |
484 | sigemptyset(&set); | |
485 | sigaddset(&set, SIGUSR1); | |
486 | sigaddset(&set, SIGHUP); | |
487 | sigaddset(&set, SIGALRM); | |
488 | sigaddset(&set, SIGTERM); | |
489 | sigprocmask(SIG_BLOCK, &set, NULL); | |
490 | act.sa_handler = wake_me; | |
491 | act.sa_flags = 0; | |
492 | sigaction(SIGUSR1, &act, NULL); | |
493 | sigaction(SIGALRM, &act, NULL); | |
494 | act.sa_handler = hup; | |
495 | sigaction(SIGHUP, &act, NULL); | |
496 | act.sa_handler = term; | |
497 | sigaction(SIGTERM, &act, NULL); | |
498 | act.sa_handler = SIG_IGN; | |
499 | sigaction(SIGPIPE, &act, NULL); | |
500 | ||
13047e4c DW |
501 | if (switchroot) { |
502 | /* we assume we assume that /sys /proc /dev are available in | |
96a8270d | 503 | * the new root |
13047e4c | 504 | */ |
96a8270d | 505 | victim = devname2mdmon(container->devname); |
9f1da824 | 506 | victim_sock = connect_monitor(container->devname); |
13047e4c DW |
507 | if (chroot(switchroot) != 0) { |
508 | fprintf(stderr, "mdmon: failed to chroot to '%s': %s\n", | |
509 | switchroot, strerror(errno)); | |
510 | exit(4); | |
511 | } | |
512 | } | |
513 | ||
13047e4c | 514 | ignore = chdir("/"); |
96a8270d | 515 | if (victim < 0 && test_pidfile(container->devname) == 0) { |
b109d928 DW |
516 | if (ping_monitor(container->devname) == 0) { |
517 | fprintf(stderr, "mdmon: %s already managed\n", | |
518 | container->devname); | |
519 | exit(3); | |
96a8270d DW |
520 | } else if (victim < 0) |
521 | victim = devname2mdmon(container->devname); | |
549e9569 | 522 | } |
1ffd2840 | 523 | if (container->ss->load_super(container, mdfd, devname)) { |
549e9569 | 524 | fprintf(stderr, "mdmon: Cannot load metadata for %s\n", |
1ffd2840 | 525 | devname); |
549e9569 NB |
526 | exit(3); |
527 | } | |
e8a70c89 | 528 | close(mdfd); |
549e9569 | 529 | |
9fe32043 N |
530 | /* Ok, this is close enough. We can say goodbye to our parent now. |
531 | */ | |
532 | status = 0; | |
3d2c4fc7 DW |
533 | if (write(pfd[1], &status, sizeof(status)) < 0) |
534 | fprintf(stderr, "mdmon: failed to notify our parent: %d\n", | |
535 | getppid()); | |
9fe32043 N |
536 | close(pfd[1]); |
537 | ||
9fe32043 N |
538 | setsid(); |
539 | close(0); | |
540 | open("/dev/null", O_RDWR); | |
541 | close(1); | |
3d2c4fc7 | 542 | ignore = dup(0); |
9fe32043 N |
543 | #ifndef DEBUG |
544 | close(2); | |
3d2c4fc7 | 545 | ignore = dup(0); |
9fe32043 N |
546 | #endif |
547 | ||
1373b07d | 548 | mlockall(MCL_CURRENT | MCL_FUTURE); |
549e9569 | 549 | |
3e70c845 | 550 | if (clone_monitor(container) < 0) { |
295646b3 | 551 | fprintf(stderr, "mdmon: failed to start monitor process: %s\n", |
549e9569 NB |
552 | strerror(errno)); |
553 | exit(2); | |
554 | } | |
555 | ||
9f1da824 DW |
556 | if (victim > -1) { |
557 | try_kill_monitor(victim, container->devname, victim_sock); | |
558 | close(victim_sock); | |
559 | } | |
549e9569 NB |
560 | do_manager(container); |
561 | ||
562 | exit(0); | |
563 | } |