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