monitor.c

   1
   2 #include "mdadm.h"
   3 #include "mdmon.h"
   4
   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 static 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         kill(getpid(), SIGUSR1);
 142 }
 143
 144 /* Monitor a set of active md arrays - all of which share the
 145  * same metadata - and respond to events that require
 146  * metadata update.
 147  *
 148  * New arrays are detected by another thread which allocates
 149  * required memory and attaches the data structure to our list.
 150  *
 151  * Events:
 152  *  Array stops.
 153  *    This is detected by array_state going to 'clear' or 'inactive'.
 154  *    while we thought it was active.
 155  *    Response is to mark metadata as clean and 'clear' the array(??)
 156  *  write-pending
 157  *    array_state if 'write-pending'
 158  *    We mark metadata as 'dirty' then set array to 'active'.
 159  *  active_idle
 160  *    Either ignore, or mark clean, then mark metadata as clean.
 161  *
 162  *  device fails
 163  *    detected by rd-N/state reporting "faulty"
 164  *    mark device as 'failed' in metadata, let the kernel release the
 165  *    device by writing '-blocked' to rd/state, and finally write 'remove' to
 166  *    rd/state.  Before a disk can be replaced it must be failed and removed
 167  *    from all container members, this will be preemptive for the other
 168  *    arrays... safe?
 169  *
 170  *  sync completes
 171  *    sync_action was 'resync' and becomes 'idle' and resync_start becomes
 172  *    MaxSector
 173  *    Notify metadata that sync is complete.
 174  *
 175  *  recovery completes
 176  *    sync_action changes from 'recover' to 'idle'
 177  *    Check each device state and mark metadata if 'faulty' or 'in_sync'.
 178  *
 179  *  deal with resync
 180  *    This only happens on finding a new array... mdadm will have set
 181  *    'resync_start' to the correct value.  If 'resync_start' indicates that an
 182  *    resync needs to occur set the array to the 'active' state rather than the
 183  *    initial read-auto state.
 184  *
 185  *
 186  *
 187  * We wait for a change (poll/select) on array_state, sync_action, and
 188  * each rd-X/state file.
 189  * When we get any change, we check everything.  So read each state file,
 190  * then decide what to do.
 191  *
 192  * The core action is to write new metadata to all devices in the array.
 193  * This is done at most once on any wakeup.
 194  * After that we might:
 195  *   - update the array_state
 196  *   - set the role of some devices.
 197  *   - request a sync_action
 198  *
 199  */
 200
 201 static int read_and_act(struct active_array *a)
 202 {
 203         int check_degraded = 0;
 204         int deactivate = 0;
 205         struct mdinfo *mdi;
 206
 207         a->next_state = bad_word;
 208         a->next_action = bad_action;
 209
 210         a->curr_state = read_state(a->info.state_fd);
 211         a->curr_action = read_action(a->action_fd);
 212         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 213                 mdi->next_state = 0;
 214                 if (mdi->state_fd >= 0)
 215                         mdi->curr_state = read_dev_state(mdi->state_fd);
 216         }
 217
 218         if (a->curr_state <= inactive &&
 219             a->prev_state > inactive) {
 220                 /* array has been stopped */
 221                 a->container->ss->set_array_state(a, 1);
 222                 a->next_state = clear;
 223                 deactivate = 1;
 224         }
 225         if (a->curr_state == write_pending) {
 226                 get_resync_start(a);
 227                 a->container->ss->set_array_state(a, 0);
 228                 a->next_state = active;
 229         }
 230         if (a->curr_state == active_idle) {
 231                 /* Set array to 'clean' FIRST, then
 232                  * a->ss->mark_clean(a, ~0ULL);
 233                  * just ignore for now.
 234                  */
 235         }
 236
 237         if (a->curr_state == readonly) {
 238                 /* Well, I'm ready to handle things, so
 239                  * read-auto is OK. FIXME what if we really want
 240                  * readonly ???
 241                  */
 242                 get_resync_start(a);
 243                 printf("Found a readonly array at %llu\n", a->resync_start);
 244                 if (a->resync_start == ~0ULL)
 245                         a->next_state = read_auto; /* array is clean */
 246                 else {
 247                         a->container->ss->set_array_state(a, 0);
 248                         a->next_state = active;
 249                 }
 250         }
 251
 252         if (a->curr_action == idle &&
 253             a->prev_action == resync) {
 254                 /* A resync has finished.  The endpoint is recorded in
 255                  * 'sync_start'.  We don't update the metadata
 256                  * until the array goes inactive or readonly though.
 257                  * Just check if we need to fiddle spares.
 258                  */
 259                 get_resync_start(a);
 260                 a->container->ss->set_array_state(a, 0);
 261                 check_degraded = 1;
 262         }
 263
 264         if (a->curr_action == idle &&
 265             a->prev_action == recover) {
 266                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 267                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 268                                                    mdi->curr_state);
 269                         if (! (mdi->curr_state & DS_INSYNC))
 270                                 check_degraded = 1;
 271                 }
 272         }
 273
 274
 275         for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 276                 if (mdi->curr_state & DS_FAULTY) {
 277                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 278                                                    mdi->curr_state);
 279                         check_degraded = 1;
 280                         mdi->next_state = DS_REMOVE;
 281                 }
 282         }
 283
 284         if (check_degraded) {
 285                 /* manager will do the actual check */
 286                 a->check_degraded = 1;
 287                 signal_manager();
 288         }
 289
 290         a->container->ss->sync_metadata(a->container);
 291
 292         /* Effect state changes in the array */
 293         if (a->next_state != bad_word)
 294                 write_attr(array_states[a->next_state], a->info.state_fd);
 295         if (a->next_action != bad_action)
 296                 write_attr(sync_actions[a->next_action], a->action_fd);
 297         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 298                 if (mdi->next_state == DS_REMOVE && mdi->state_fd >= 0) {
 299                         int remove_result;
 300
 301                         write_attr("-blocked", mdi->state_fd);
 302                         /* the kernel may not be able to immediately remove the
 303                          * disk, we can simply wait until the next event to try
 304                          * again.
 305                          */
 306                         remove_result = write_attr("remove", mdi->state_fd);
 307                         if (remove_result > 0) {
 308                                 close(mdi->state_fd);
 309                                 mdi->state_fd = -1;
 310                         }
 311                 }
 312                 if (mdi->next_state & DS_INSYNC)
 313                         write_attr("+in_sync", mdi->state_fd);
 314         }
 315
 316         /* move curr_ to prev_ */
 317         a->prev_state = a->curr_state;
 318
 319         a->prev_action = a->curr_action;
 320
 321         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 322                 mdi->prev_state = mdi->curr_state;
 323                 mdi->next_state = 0;
 324         }
 325
 326         if (deactivate)
 327                 a->container = NULL;
 328
 329         return 1;
 330 }
 331
 332 static struct mdinfo *
 333 find_device(struct active_array *a, int major, int minor)
 334 {
 335         struct mdinfo *mdi;
 336
 337         for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 338                 if (mdi->disk.major == major && mdi->disk.minor == minor)
 339                         return mdi;
 340
 341         return NULL;
 342 }
 343
 344 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
 345 {
 346         struct active_array *a;
 347         struct mdinfo *victim;
 348
 349         for (a = aa; a; a = a->next) {
 350                 if (!a->container)
 351                         continue;
 352                 victim = find_device(a, failed->disk.major, failed->disk.minor);
 353                 if (!victim)
 354                         continue;
 355
 356                 if (!(victim->curr_state & DS_FAULTY))
 357                         write_attr("faulty", victim->state_fd);
 358         }
 359 }
 360
 361 static int handle_remove_device(struct md_remove_device_cmd *cmd, struct active_array *aa)
 362 {
 363         struct active_array *a;
 364         struct mdinfo *victim;
 365         int rv;
 366
 367         /* scan all arrays for the given device, if ->state_fd is closed (-1)
 368          * in all cases then mark the disk as removed in the metadata.
 369          * Otherwise reply that it is busy.
 370          */
 371
 372         /* pass1 check that it is not in use anywhere */
 373         /* note: we are safe from re-adds as long as the device exists in the
 374          * container
 375          */
 376         for (a = aa; a; a = a->next) {
 377                 if (!a->container)
 378                         continue;
 379                 victim = find_device(a, major(cmd->rdev), minor(cmd->rdev));
 380                 if (!victim)
 381                         continue;
 382                 if (victim->state_fd > 0)
 383                         return -EBUSY;
 384         }
 385
 386         /* pass2 schedule and process removal per array */
 387         for (a = aa; a; a = a->next) {
 388                 if (!a->container)
 389                         continue;
 390                 victim = find_device(a, major(cmd->rdev), minor(cmd->rdev));
 391                 if (!victim)
 392                         continue;
 393                 victim->curr_state |= DS_REMOVE;
 394                 rv = read_and_act(a);
 395                 if (rv < 0)
 396                         return rv;
 397         }
 398
 399         return 0;
 400 }
 401
 402 static int handle_pipe(struct md_generic_cmd *cmd, struct active_array *aa)
 403 {
 404         switch (cmd->action) {
 405         case md_action_ping_monitor:
 406                 return 0;
 407         case md_action_remove_device:
 408                 return handle_remove_device((void *) cmd, aa);
 409         }
 410
 411         return -1;
 412 }
 413
 414 static int wait_and_act(struct supertype *container, int pfd,
 415                         int monfd, int nowait)
 416 {
 417         fd_set rfds;
 418         int maxfd = 0;
 419         struct active_array **aap = &container->arrays;
 420         struct active_array *a, **ap;
 421         int rv;
 422         struct mdinfo *mdi;
 423
 424         FD_ZERO(&rfds);
 425
 426         add_fd(&rfds, &maxfd, pfd);
 427         for (ap = aap ; *ap ;) {
 428                 a = *ap;
 429                 /* once an array has been deactivated we want to
 430                  * ask the manager to discard it.
 431                  */
 432                 if (!a->container) {
 433                         if (discard_this) {
 434                                 ap = &(*ap)->next;
 435                                 continue;
 436                         }
 437                         *ap = a->next;
 438                         a->next = NULL;
 439                         discard_this = a;
 440                         signal_manager();
 441                         continue;
 442                 }
 443
 444                 add_fd(&rfds, &maxfd, a->info.state_fd);
 445                 add_fd(&rfds, &maxfd, a->action_fd);
 446                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 447                         add_fd(&rfds, &maxfd, mdi->state_fd);
 448
 449                 ap = &(*ap)->next;
 450         }
 451
 452         if (manager_ready && *aap == NULL) {
 453                 /* No interesting arrays. Lets see about exiting.
 454                  * Note that blocking at this point is not a problem
 455                  * as there are no active arrays, there is nothing that
 456                  * we need to be ready to do.
 457                  */
 458                 int fd = open(container->device_name, O_RDONLY|O_EXCL);
 459                 if (fd >= 0 || errno != EBUSY) {
 460                         /* OK, we are safe to leave */
 461                         exit_now = 1;
 462                         signal_manager();
 463                         remove_pidfile(container->devname);
 464                         exit(0);
 465                 }
 466         }
 467
 468         if (!nowait) {
 469                 rv = select(maxfd+1, &rfds, NULL, NULL, NULL);
 470
 471                 if (rv <= 0)
 472                         return rv;
 473
 474                 if (FD_ISSET(pfd, &rfds)) {
 475                         int err = -1;
 476
 477                         if (read(pfd, &err, 1) > 0)
 478                                 err = handle_pipe(active_cmd, *aap);
 479                         write(monfd, &err, 1);
 480                 }
 481         }
 482
 483         if (update_queue) {
 484                 struct metadata_update *this;
 485
 486                 for (this = update_queue; this ; this = this->next)
 487                         container->ss->process_update(container, this);
 488
 489                 update_queue_handled = update_queue;
 490                 update_queue = NULL;
 491                 signal_manager();
 492                 container->ss->sync_metadata(container);
 493         }
 494
 495         for (a = *aap; a ; a = a->next) {
 496                 if (a->replaces && !discard_this) {
 497                         struct active_array **ap;
 498                         for (ap = &a->next; *ap && *ap != a->replaces;
 499                              ap = & (*ap)->next)
 500                                 ;
 501                         if (*ap)
 502                                 *ap = (*ap)->next;
 503                         discard_this = a->replaces;
 504                         a->replaces = NULL;
 505                         /* FIXME check if device->state_fd need to be cleared?*/
 506                         signal_manager();
 507                 }
 508                 if (a->container)
 509                         rv += read_and_act(a);
 510         }
 511
 512         /* propagate failures across container members */
 513         for (a = *aap; a ; a = a->next) {
 514                 if (!a->container)
 515                         continue;
 516                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 517                         if (mdi->curr_state & DS_FAULTY)
 518                                 reconcile_failed(*aap, mdi);
 519         }
 520
 521         return rv;
 522 }
 523
 524 void do_monitor(struct supertype *container)
 525 {
 526         int rv;
 527         int first = 1;
 528         do {
 529                 rv = wait_and_act(container, container->mgr_pipe[0],
 530                                   container->mon_pipe[1], first);
 531                 first = 0;
 532         } while (rv >= 0);
 533 }