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->container->ss->set_array_state(a, 2))
 247                         a->next_state = read_auto; /* array is clean */
 248                 else
 249                         a->next_state = active; /* Now active for recovery etc */
 250         }
 251
 252         if (!deactivate &&
 253             a->curr_action == idle &&
 254             a->prev_action == resync) {
 255                 /* A resync has finished.  The endpoint is recorded in
 256                  * 'sync_start'.  We don't update the metadata
 257                  * until the array goes inactive or readonly though.
 258                  * Just check if we need to fiddle spares.
 259                  */
 260                 get_resync_start(a);
 261                 a->container->ss->set_array_state(a, 0);
 262                 check_degraded = 1;
 263         }
 264
 265         if (!deactivate &&
 266             a->curr_action == idle &&
 267             a->prev_action == recover) {
 268                 /* A recovery has finished.  Some disks may be in sync now,
 269                  * and the array may no longer be degraded
 270                  */
 271                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 272                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 273                                                    mdi->curr_state);
 274                         if (! (mdi->curr_state & DS_INSYNC))
 275                                 check_degraded = 1;
 276                 }
 277         }
 278
 279         /* Check for failures and if found:
 280          * 1/ Record the failure in the metadata and unblock the device.
 281          *    FIXME update the kernel to stop notifying on failed drives when
 282          *    the array is readonly and we have cleared 'blocked'
 283          * 2/ Try to remove the device if the array is writable, or can be
 284          *    made writable.
 285          */
 286         for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 287                 if (mdi->curr_state & DS_FAULTY) {
 288                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 289                                                    mdi->curr_state);
 290                         check_degraded = 1;
 291                         mdi->next_state |= DS_UNBLOCK;
 292                         if (a->curr_state == read_auto) {
 293                                 a->container->ss->set_array_state(a, 0);
 294                                 a->next_state = active;
 295                         }
 296                         if (a->curr_state > readonly)
 297                                 mdi->next_state |= DS_REMOVE;
 298                 }
 299         }
 300
 301         a->container->ss->sync_metadata(a->container);
 302         dprintf("%s: update[%d]: (", __func__, a->info.container_member);
 303
 304         /* Effect state changes in the array */
 305         if (a->next_state != bad_word) {
 306                 dprintf(" state:%s", array_states[a->next_state]);
 307                 write_attr(array_states[a->next_state], a->info.state_fd);
 308         }
 309         if (a->next_action != bad_action) {
 310                 write_attr(sync_actions[a->next_action], a->action_fd);
 311                 dprintf(" action:%s", array_states[a->next_state]);
 312         }
 313         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 314                 if (mdi->next_state & DS_UNBLOCK) {
 315                         dprintf(" %d:-blocked", mdi->disk.raid_disk);
 316                         write_attr("-blocked", mdi->state_fd);
 317                 }
 318
 319                 if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
 320                         int remove_result;
 321
 322                         /* the kernel may not be able to immediately remove the
 323                          * disk, we can simply wait until the next event to try
 324                          * again.
 325                          */
 326                         remove_result = write_attr("remove", mdi->state_fd);
 327                         if (remove_result > 0) {
 328                                 dprintf(" %d:removed", mdi->disk.raid_disk);
 329                                 close(mdi->state_fd);
 330                                 mdi->state_fd = -1;
 331                         }
 332                 }
 333                 if (mdi->next_state & DS_INSYNC) {
 334                         write_attr("+in_sync", mdi->state_fd);
 335                         dprintf(" %d:+in_sync", mdi->disk.raid_disk);
 336                 }
 337         }
 338         dprintf(" )\n");
 339
 340         /* move curr_ to prev_ */
 341         a->prev_state = a->curr_state;
 342
 343         a->prev_action = a->curr_action;
 344
 345         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 346                 mdi->prev_state = mdi->curr_state;
 347                 mdi->next_state = 0;
 348         }
 349
 350         if (check_degraded) {
 351                 /* manager will do the actual check */
 352                 a->check_degraded = 1;
 353                 signal_manager();
 354         }
 355
 356         if (deactivate)
 357                 a->container = NULL;
 358
 359         return 1;
 360 }
 361
 362 static struct mdinfo *
 363 find_device(struct active_array *a, int major, int minor)
 364 {
 365         struct mdinfo *mdi;
 366
 367         for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 368                 if (mdi->disk.major == major && mdi->disk.minor == minor)
 369                         return mdi;
 370
 371         return NULL;
 372 }
 373
 374 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
 375 {
 376         struct active_array *a;
 377         struct mdinfo *victim;
 378
 379         for (a = aa; a; a = a->next) {
 380                 if (!a->container)
 381                         continue;
 382                 victim = find_device(a, failed->disk.major, failed->disk.minor);
 383                 if (!victim)
 384                         continue;
 385
 386                 if (!(victim->curr_state & DS_FAULTY))
 387                         write_attr("faulty", victim->state_fd);
 388         }
 389 }
 390
 391 #ifdef DEBUG
 392 static void dprint_wake_reasons(fd_set *fds)
 393 {
 394         int i;
 395         char proc_path[256];
 396         char link[256];
 397         char *basename;
 398         int rv;
 399
 400         fprintf(stderr, "monitor: wake ( ");
 401         for (i = 0; i < FD_SETSIZE; i++) {
 402                 if (FD_ISSET(i, fds)) {
 403                         sprintf(proc_path, "/proc/%d/fd/%d",
 404                                 (int) getpid(), i);
 405
 406                         rv = readlink(proc_path, link, sizeof(link) - 1);
 407                         if (rv < 0) {
 408                                 fprintf(stderr, "%d:unknown ", i);
 409                                 continue;
 410                         }
 411                         link[rv] = '\0';
 412                         basename = strrchr(link, '/');
 413                         fprintf(stderr, "%d:%s ",
 414                                 i, basename ? ++basename : link);
 415                 }
 416         }
 417         fprintf(stderr, ")\n");
 418 }
 419 #endif
 420
 421 int monitor_loop_cnt;
 422
 423 static int wait_and_act(struct supertype *container, int nowait)
 424 {
 425         fd_set rfds;
 426         int maxfd = 0;
 427         struct active_array **aap = &container->arrays;
 428         struct active_array *a, **ap;
 429         int rv;
 430         struct mdinfo *mdi;
 431
 432         FD_ZERO(&rfds);
 433
 434         for (ap = aap ; *ap ;) {
 435                 a = *ap;
 436                 /* once an array has been deactivated we want to
 437                  * ask the manager to discard it.
 438                  */
 439                 if (!a->container) {
 440                         if (discard_this) {
 441                                 ap = &(*ap)->next;
 442                                 continue;
 443                         }
 444                         *ap = a->next;
 445                         a->next = NULL;
 446                         discard_this = a;
 447                         signal_manager();
 448                         continue;
 449                 }
 450
 451                 add_fd(&rfds, &maxfd, a->info.state_fd);
 452                 add_fd(&rfds, &maxfd, a->action_fd);
 453                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 454                         add_fd(&rfds, &maxfd, mdi->state_fd);
 455
 456                 ap = &(*ap)->next;
 457         }
 458
 459         if (manager_ready && *aap == NULL) {
 460                 /* No interesting arrays. Lets see about exiting.
 461                  * Note that blocking at this point is not a problem
 462                  * as there are no active arrays, there is nothing that
 463                  * we need to be ready to do.
 464                  */
 465                 int fd = open(container->device_name, O_RDONLY|O_EXCL);
 466                 if (fd >= 0 || errno != EBUSY) {
 467                         /* OK, we are safe to leave */
 468                         dprintf("no arrays to monitor... exiting\n");
 469                         remove_pidfile(container->devname);
 470                         exit_now = 1;
 471                         signal_manager();
 472                         exit(0);
 473                 }
 474         }
 475
 476         if (!nowait) {
 477                 sigset_t set;
 478                 sigprocmask(SIG_UNBLOCK, NULL, &set);
 479                 sigdelset(&set, SIGUSR1);
 480                 monitor_loop_cnt |= 1;
 481                 rv = pselect(maxfd+1, &rfds, NULL, NULL, NULL, &set);
 482                 monitor_loop_cnt += 1;
 483                 if (rv == -1 && errno == EINTR)
 484                         rv = 0;
 485                 #ifdef DEBUG
 486                 dprint_wake_reasons(&rfds);
 487                 #endif
 488
 489         }
 490
 491         if (update_queue) {
 492                 struct metadata_update *this;
 493
 494                 for (this = update_queue; this ; this = this->next)
 495                         container->ss->process_update(container, this);
 496
 497                 update_queue_handled = update_queue;
 498                 update_queue = NULL;
 499                 signal_manager();
 500                 container->ss->sync_metadata(container);
 501         }
 502
 503         for (a = *aap; a ; a = a->next) {
 504                 if (a->replaces && !discard_this) {
 505                         struct active_array **ap;
 506                         for (ap = &a->next; *ap && *ap != a->replaces;
 507                              ap = & (*ap)->next)
 508                                 ;
 509                         if (*ap)
 510                                 *ap = (*ap)->next;
 511                         discard_this = a->replaces;
 512                         a->replaces = NULL;
 513                         /* FIXME check if device->state_fd need to be cleared?*/
 514                         signal_manager();
 515                 }
 516                 if (a->container)
 517                         rv += read_and_act(a);
 518         }
 519
 520         /* propagate failures across container members */
 521         for (a = *aap; a ; a = a->next) {
 522                 if (!a->container)
 523                         continue;
 524                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 525                         if (mdi->curr_state & DS_FAULTY)
 526                                 reconcile_failed(*aap, mdi);
 527         }
 528
 529         return rv;
 530 }
 531
 532 void do_monitor(struct supertype *container)
 533 {
 534         int rv;
 535         int first = 1;
 536         do {
 537                 rv = wait_and_act(container, first);
 538                 first = 0;
 539         } while (rv >= 0);
 540 }