monitor.c

   1
   2 #include "mdadm.h"
   3 #include "mdmon.h"
   4 #include <sys/syscall.h>
   5 #include <sys/select.h>
   6 #include <signal.h>
   7
   8 static char *array_states[] = {
   9         "clear", "inactive", "suspended", "readonly", "read-auto",
  10         "clean", "active", "write-pending", "active-idle", NULL };
  11 static char *sync_actions[] = {
  12         "idle", "reshape", "resync", "recover", "check", "repair", NULL
  13 };
  14
  15 static int write_attr(char *attr, int fd)
  16 {
  17         return write(fd, attr, strlen(attr));
  18 }
  19
  20 static void add_fd(fd_set *fds, int *maxfd, int fd)
  21 {
  22         if (fd < 0)
  23                 return;
  24         if (fd > *maxfd)
  25                 *maxfd = fd;
  26         FD_SET(fd, fds);
  27 }
  28
  29 static int read_attr(char *buf, int len, int fd)
  30 {
  31         int n;
  32
  33         if (fd < 0) {
  34                 buf[0] = 0;
  35                 return 0;
  36         }
  37         lseek(fd, 0, 0);
  38         n = read(fd, buf, len - 1);
  39
  40         if (n <= 0) {
  41                 buf[0] = 0;
  42                 return 0;
  43         }
  44         buf[n] = 0;
  45         if (buf[n-1] == '\n')
  46                 buf[n-1] = 0;
  47         return n;
  48 }
  49
  50
  51 int get_resync_start(struct active_array *a)
  52 {
  53         char buf[30];
  54         int n;
  55
  56         n = read_attr(buf, 30, a->resync_start_fd);
  57         if (n <= 0)
  58                 return n;
  59
  60         a->resync_start = strtoull(buf, NULL, 10);
  61
  62         return 1;
  63 }
  64
  65 static int attr_match(const char *attr, const char *str)
  66 {
  67         /* See if attr, read from a sysfs file, matches
  68          * str.  They must either be the same, or attr can
  69          * have a trailing newline or comma
  70          */
  71         while (*attr && *str && *attr == *str) {
  72                 attr++;
  73                 str++;
  74         }
  75
  76         if (*str || (*attr && *attr != ',' && *attr != '\n'))
  77                 return 0;
  78         return 1;
  79 }
  80
  81 static int match_word(const char *word, char **list)
  82 {
  83         int n;
  84         for (n=0; list[n]; n++)
  85                 if (attr_match(word, list[n]))
  86                         break;
  87         return n;
  88 }
  89
  90 static enum array_state read_state(int fd)
  91 {
  92         char buf[20];
  93         int n = read_attr(buf, 20, fd);
  94
  95         if (n <= 0)
  96                 return bad_word;
  97         return (enum array_state) match_word(buf, array_states);
  98 }
  99
 100 static enum sync_action read_action( int fd)
 101 {
 102         char buf[20];
 103         int n = read_attr(buf, 20, fd);
 104
 105         if (n <= 0)
 106                 return bad_action;
 107         return (enum sync_action) match_word(buf, sync_actions);
 108 }
 109
 110 int read_dev_state(int fd)
 111 {
 112         char buf[60];
 113         int n = read_attr(buf, 60, fd);
 114         char *cp;
 115         int rv = 0;
 116
 117         if (n <= 0)
 118                 return 0;
 119
 120         cp = buf;
 121         while (cp) {
 122                 if (attr_match(cp, "faulty"))
 123                         rv |= DS_FAULTY;
 124                 if (attr_match(cp, "in_sync"))
 125                         rv |= DS_INSYNC;
 126                 if (attr_match(cp, "write_mostly"))
 127                         rv |= DS_WRITE_MOSTLY;
 128                 if (attr_match(cp, "spare"))
 129                         rv |= DS_SPARE;
 130                 if (attr_match(cp, "blocked"))
 131                         rv |= DS_BLOCKED;
 132                 cp = strchr(cp, ',');
 133                 if (cp)
 134                         cp++;
 135         }
 136         return rv;
 137 }
 138
 139 static void signal_manager(void)
 140 {
 141         /* tgkill(getpid(), mon_tid, SIGUSR1); */
 142         int pid = getpid();
 143         syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
 144 }
 145
 146 /* Monitor a set of active md arrays - all of which share the
 147  * same metadata - and respond to events that require
 148  * metadata update.
 149  *
 150  * New arrays are detected by another thread which allocates
 151  * required memory and attaches the data structure to our list.
 152  *
 153  * Events:
 154  *  Array stops.
 155  *    This is detected by array_state going to 'clear' or 'inactive'.
 156  *    while we thought it was active.
 157  *    Response is to mark metadata as clean and 'clear' the array(??)
 158  *  write-pending
 159  *    array_state if 'write-pending'
 160  *    We mark metadata as 'dirty' then set array to 'active'.
 161  *  active_idle
 162  *    Either ignore, or mark clean, then mark metadata as clean.
 163  *
 164  *  device fails
 165  *    detected by rd-N/state reporting "faulty"
 166  *    mark device as 'failed' in metadata, let the kernel release the
 167  *    device by writing '-blocked' to rd/state, and finally write 'remove' to
 168  *    rd/state.  Before a disk can be replaced it must be failed and removed
 169  *    from all container members, this will be preemptive for the other
 170  *    arrays... safe?
 171  *
 172  *  sync completes
 173  *    sync_action was 'resync' and becomes 'idle' and resync_start becomes
 174  *    MaxSector
 175  *    Notify metadata that sync is complete.
 176  *
 177  *  recovery completes
 178  *    sync_action changes from 'recover' to 'idle'
 179  *    Check each device state and mark metadata if 'faulty' or 'in_sync'.
 180  *
 181  *  deal with resync
 182  *    This only happens on finding a new array... mdadm will have set
 183  *    'resync_start' to the correct value.  If 'resync_start' indicates that an
 184  *    resync needs to occur set the array to the 'active' state rather than the
 185  *    initial read-auto state.
 186  *
 187  *
 188  *
 189  * We wait for a change (poll/select) on array_state, sync_action, and
 190  * each rd-X/state file.
 191  * When we get any change, we check everything.  So read each state file,
 192  * then decide what to do.
 193  *
 194  * The core action is to write new metadata to all devices in the array.
 195  * This is done at most once on any wakeup.
 196  * After that we might:
 197  *   - update the array_state
 198  *   - set the role of some devices.
 199  *   - request a sync_action
 200  *
 201  */
 202
 203 static int read_and_act(struct active_array *a)
 204 {
 205         int check_degraded = 0;
 206         int deactivate = 0;
 207         struct mdinfo *mdi;
 208
 209         a->next_state = bad_word;
 210         a->next_action = bad_action;
 211
 212         a->curr_state = read_state(a->info.state_fd);
 213         a->curr_action = read_action(a->action_fd);
 214         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 215                 mdi->next_state = 0;
 216                 if (mdi->state_fd >= 0)
 217                         mdi->curr_state = read_dev_state(mdi->state_fd);
 218         }
 219
 220         if (a->curr_state <= inactive &&
 221             a->prev_state > inactive) {
 222                 /* array has been stopped */
 223                 get_resync_start(a);
 224                 a->container->ss->set_array_state(a, 1);
 225                 a->next_state = clear;
 226                 deactivate = 1;
 227         }
 228         if (a->curr_state == write_pending) {
 229                 get_resync_start(a);
 230                 a->container->ss->set_array_state(a, 0);
 231                 a->next_state = active;
 232         }
 233         if (a->curr_state == active_idle) {
 234                 /* Set array to 'clean' FIRST, then
 235                  * a->ss->mark_clean(a, ~0ULL);
 236                  * just ignore for now.
 237                  */
 238         }
 239
 240         if (a->curr_state == readonly) {
 241                 /* Well, I'm ready to handle things, so
 242                  * read-auto is OK. FIXME what if we really want
 243                  * readonly ???
 244                  */
 245                 get_resync_start(a);
 246                 if (a->resync_start == ~0ULL) {
 247                         a->next_state = read_auto; /* array is clean */
 248                         /* give the metadata a chance to force active if
 249                          * we have some recovery to do.  metadata sets
 250                          * resync_start to !MaxSector in this case
 251                          */
 252                         a->container->ss->set_array_state(a, 1);
 253                 }
 254                 if (a->resync_start != ~0ULL) {
 255                         a->container->ss->set_array_state(a, 0);
 256                         a->next_state = active;
 257                 }
 258         }
 259
 260         if (!deactivate &&
 261             a->curr_action == idle &&
 262             a->prev_action == resync) {
 263                 /* A resync has finished.  The endpoint is recorded in
 264                  * 'sync_start'.  We don't update the metadata
 265                  * until the array goes inactive or readonly though.
 266                  * Just check if we need to fiddle spares.
 267                  */
 268                 get_resync_start(a);
 269                 a->container->ss->set_array_state(a, 0);
 270                 check_degraded = 1;
 271         }
 272
 273         if (!deactivate &&
 274             a->curr_action == idle &&
 275             a->prev_action == recover) {
 276                 /* A recovery has finished.  Some disks may be in sync now,
 277                  * and the array may no longer be degraded
 278                  */
 279                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 280                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 281                                                    mdi->curr_state);
 282                         if (! (mdi->curr_state & DS_INSYNC))
 283                                 check_degraded = 1;
 284                 }
 285         }
 286
 287         /* Check for failures and if found:
 288          * 1/ Record the failure in the metadata and unblock the device.
 289          *    FIXME update the kernel to stop notifying on failed drives when
 290          *    the array is readonly and we have cleared 'blocked'
 291          * 2/ Try to remove the device if the array is writable, or can be
 292          *    made writable.
 293          */
 294         for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 295                 if (mdi->curr_state & DS_FAULTY) {
 296                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 297                                                    mdi->curr_state);
 298                         check_degraded = 1;
 299                         mdi->next_state |= DS_UNBLOCK;
 300                         if (a->curr_state == read_auto) {
 301                                 a->container->ss->set_array_state(a, 0);
 302                                 a->next_state = active;
 303                         }
 304                         if (a->curr_state > readonly)
 305                                 mdi->next_state |= DS_REMOVE;
 306                 }
 307         }
 308
 309         a->container->ss->sync_metadata(a->container);
 310         dprintf("%s: update[%d]: (", __func__, a->info.container_member);
 311
 312         /* Effect state changes in the array */
 313         if (a->next_state != bad_word) {
 314                 dprintf(" state:%s", array_states[a->next_state]);
 315                 write_attr(array_states[a->next_state], a->info.state_fd);
 316         }
 317         if (a->next_action != bad_action) {
 318                 write_attr(sync_actions[a->next_action], a->action_fd);
 319                 dprintf(" action:%s", array_states[a->next_state]);
 320         }
 321         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 322                 if (mdi->next_state & DS_UNBLOCK) {
 323                         dprintf(" %d:-blocked", mdi->disk.raid_disk);
 324                         write_attr("-blocked", mdi->state_fd);
 325                 }
 326
 327                 if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
 328                         int remove_result;
 329
 330                         /* the kernel may not be able to immediately remove the
 331                          * disk, we can simply wait until the next event to try
 332                          * again.
 333                          */
 334                         remove_result = write_attr("remove", mdi->state_fd);
 335                         if (remove_result > 0) {
 336                                 dprintf(" %d:removed", mdi->disk.raid_disk);
 337                                 close(mdi->state_fd);
 338                                 mdi->state_fd = -1;
 339                         }
 340                 }
 341                 if (mdi->next_state & DS_INSYNC) {
 342                         write_attr("+in_sync", mdi->state_fd);
 343                         dprintf(" %d:+in_sync", mdi->disk.raid_disk);
 344                 }
 345         }
 346         dprintf(" )\n");
 347
 348         /* move curr_ to prev_ */
 349         a->prev_state = a->curr_state;
 350
 351         a->prev_action = a->curr_action;
 352
 353         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 354                 mdi->prev_state = mdi->curr_state;
 355                 mdi->next_state = 0;
 356         }
 357
 358         if (check_degraded) {
 359                 /* manager will do the actual check */
 360                 a->check_degraded = 1;
 361                 signal_manager();
 362         }
 363
 364         if (deactivate)
 365                 a->container = NULL;
 366
 367         return 1;
 368 }
 369
 370 static struct mdinfo *
 371 find_device(struct active_array *a, int major, int minor)
 372 {
 373         struct mdinfo *mdi;
 374
 375         for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 376                 if (mdi->disk.major == major && mdi->disk.minor == minor)
 377                         return mdi;
 378
 379         return NULL;
 380 }
 381
 382 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
 383 {
 384         struct active_array *a;
 385         struct mdinfo *victim;
 386
 387         for (a = aa; a; a = a->next) {
 388                 if (!a->container)
 389                         continue;
 390                 victim = find_device(a, failed->disk.major, failed->disk.minor);
 391                 if (!victim)
 392                         continue;
 393
 394                 if (!(victim->curr_state & DS_FAULTY))
 395                         write_attr("faulty", victim->state_fd);
 396         }
 397 }
 398
 399 #ifdef DEBUG
 400 static void dprint_wake_reasons(fd_set *fds)
 401 {
 402         int i;
 403         char proc_path[256];
 404         char link[256];
 405         char *basename;
 406         int rv;
 407
 408         fprintf(stderr, "monitor: wake ( ");
 409         for (i = 0; i < FD_SETSIZE; i++) {
 410                 if (FD_ISSET(i, fds)) {
 411                         sprintf(proc_path, "/proc/%d/fd/%d",
 412                                 (int) getpid(), i);
 413
 414                         rv = readlink(proc_path, link, sizeof(link) - 1);
 415                         if (rv < 0) {
 416                                 fprintf(stderr, "%d:unknown ", i);
 417                                 continue;
 418                         }
 419                         link[rv] = '\0';
 420                         basename = strrchr(link, '/');
 421                         fprintf(stderr, "%d:%s ",
 422                                 i, basename ? ++basename : link);
 423                 }
 424         }
 425         fprintf(stderr, ")\n");
 426 }
 427 #endif
 428
 429 int monitor_loop_cnt;
 430
 431 static int wait_and_act(struct supertype *container, int nowait)
 432 {
 433         fd_set rfds;
 434         int maxfd = 0;
 435         struct active_array **aap = &container->arrays;
 436         struct active_array *a, **ap;
 437         int rv;
 438         struct mdinfo *mdi;
 439
 440         FD_ZERO(&rfds);
 441
 442         for (ap = aap ; *ap ;) {
 443                 a = *ap;
 444                 /* once an array has been deactivated we want to
 445                  * ask the manager to discard it.
 446                  */
 447                 if (!a->container) {
 448                         if (discard_this) {
 449                                 ap = &(*ap)->next;
 450                                 continue;
 451                         }
 452                         *ap = a->next;
 453                         a->next = NULL;
 454                         discard_this = a;
 455                         signal_manager();
 456                         continue;
 457                 }
 458
 459                 add_fd(&rfds, &maxfd, a->info.state_fd);
 460                 add_fd(&rfds, &maxfd, a->action_fd);
 461                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 462                         add_fd(&rfds, &maxfd, mdi->state_fd);
 463
 464                 ap = &(*ap)->next;
 465         }
 466
 467         if (manager_ready && *aap == NULL) {
 468                 /* No interesting arrays. Lets see about exiting.
 469                  * Note that blocking at this point is not a problem
 470                  * as there are no active arrays, there is nothing that
 471                  * we need to be ready to do.
 472                  */
 473                 int fd = open(container->device_name, O_RDONLY|O_EXCL);
 474                 if (fd >= 0 || errno != EBUSY) {
 475                         /* OK, we are safe to leave */
 476                         dprintf("no arrays to monitor... exiting\n");
 477                         remove_pidfile(container->devname);
 478                         exit_now = 1;
 479                         signal_manager();
 480                         exit(0);
 481                 }
 482         }
 483
 484         if (!nowait) {
 485                 sigset_t set;
 486                 sigprocmask(SIG_UNBLOCK, NULL, &set);
 487                 sigdelset(&set, SIGUSR1);
 488                 monitor_loop_cnt |= 1;
 489                 rv = pselect(maxfd+1, &rfds, NULL, NULL, NULL, &set);
 490                 monitor_loop_cnt += 1;
 491                 if (rv == -1 && errno == EINTR)
 492                         rv = 0;
 493                 #ifdef DEBUG
 494                 dprint_wake_reasons(&rfds);
 495                 #endif
 496
 497         }
 498
 499         if (update_queue) {
 500                 struct metadata_update *this;
 501
 502                 for (this = update_queue; this ; this = this->next)
 503                         container->ss->process_update(container, this);
 504
 505                 update_queue_handled = update_queue;
 506                 update_queue = NULL;
 507                 signal_manager();
 508                 container->ss->sync_metadata(container);
 509         }
 510
 511         for (a = *aap; a ; a = a->next) {
 512                 if (a->replaces && !discard_this) {
 513                         struct active_array **ap;
 514                         for (ap = &a->next; *ap && *ap != a->replaces;
 515                              ap = & (*ap)->next)
 516                                 ;
 517                         if (*ap)
 518                                 *ap = (*ap)->next;
 519                         discard_this = a->replaces;
 520                         a->replaces = NULL;
 521                         /* FIXME check if device->state_fd need to be cleared?*/
 522                         signal_manager();
 523                 }
 524                 if (a->container)
 525                         rv += read_and_act(a);
 526         }
 527
 528         /* propagate failures across container members */
 529         for (a = *aap; a ; a = a->next) {
 530                 if (!a->container)
 531                         continue;
 532                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 533                         if (mdi->curr_state & DS_FAULTY)
 534                                 reconcile_failed(*aap, mdi);
 535         }
 536
 537         return rv;
 538 }
 539
 540 void do_monitor(struct supertype *container)
 541 {
 542         int rv;
 543         int first = 1;
 544         do {
 545                 rv = wait_and_act(container, first);
 546                 first = 0;
 547         } while (rv >= 0);
 548 }