monitor.c

   1 /*
   2  * mdmon - monitor external metadata arrays
   3  *
   4  * Copyright (C) 2007-2009 Neil Brown <neilb@suse.de>
   5  * Copyright (C) 2007-2009 Intel Corporation
   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  */
  20
  21 #include "mdadm.h"
  22 #include "mdmon.h"
  23 #include <sys/syscall.h>
  24 #include <sys/select.h>
  25 #include <signal.h>
  26
  27 static char *array_states[] = {
  28         "clear", "inactive", "suspended", "readonly", "read-auto",
  29         "clean", "active", "write-pending", "active-idle", NULL };
  30 static char *sync_actions[] = {
  31         "idle", "reshape", "resync", "recover", "check", "repair", NULL
  32 };
  33
  34 static int write_attr(char *attr, int fd)
  35 {
  36         return write(fd, attr, strlen(attr));
  37 }
  38
  39 static void add_fd(fd_set *fds, int *maxfd, int fd)
  40 {
  41         if (fd < 0)
  42                 return;
  43         if (fd > *maxfd)
  44                 *maxfd = fd;
  45         FD_SET(fd, fds);
  46 }
  47
  48 static int read_attr(char *buf, int len, int fd)
  49 {
  50         int n;
  51
  52         if (fd < 0) {
  53                 buf[0] = 0;
  54                 return 0;
  55         }
  56         lseek(fd, 0, 0);
  57         n = read(fd, buf, len - 1);
  58
  59         if (n <= 0) {
  60                 buf[0] = 0;
  61                 return 0;
  62         }
  63         buf[n] = 0;
  64         if (buf[n-1] == '\n')
  65                 buf[n-1] = 0;
  66         return n;
  67 }
  68
  69 static unsigned long long read_resync_start(int fd)
  70 {
  71         char buf[30];
  72         int n;
  73
  74         n = read_attr(buf, 30, fd);
  75         if (n <= 0)
  76                 return 0;
  77         if (strncmp(buf, "none", 4) == 0)
  78                 return MaxSector;
  79         else
  80                 return strtoull(buf, NULL, 10);
  81 }
  82
  83 static unsigned long long read_sync_completed(int fd)
  84 {
  85         unsigned long long val;
  86         char buf[50];
  87         int n;
  88         char *ep;
  89
  90         n = read_attr(buf, 50, fd);
  91
  92         if (n <= 0)
  93                 return 0;
  94         buf[n] = 0;
  95         val = strtoull(buf, &ep, 0);
  96         if (ep == buf || (*ep != 0 && *ep != '\n' && *ep != ' '))
  97                 return 0;
  98         return val;
  99 }
 100
 101 static enum array_state read_state(int fd)
 102 {
 103         char buf[20];
 104         int n = read_attr(buf, 20, fd);
 105
 106         if (n <= 0)
 107                 return bad_word;
 108         return (enum array_state) sysfs_match_word(buf, array_states);
 109 }
 110
 111 static enum sync_action read_action( int fd)
 112 {
 113         char buf[20];
 114         int n = read_attr(buf, 20, fd);
 115
 116         if (n <= 0)
 117                 return bad_action;
 118         return (enum sync_action) sysfs_match_word(buf, sync_actions);
 119 }
 120
 121 int read_dev_state(int fd)
 122 {
 123         char buf[60];
 124         int n = read_attr(buf, 60, fd);
 125         char *cp;
 126         int rv = 0;
 127
 128         if (n <= 0)
 129                 return 0;
 130
 131         cp = buf;
 132         while (cp) {
 133                 if (sysfs_attr_match(cp, "faulty"))
 134                         rv |= DS_FAULTY;
 135                 if (sysfs_attr_match(cp, "in_sync"))
 136                         rv |= DS_INSYNC;
 137                 if (sysfs_attr_match(cp, "write_mostly"))
 138                         rv |= DS_WRITE_MOSTLY;
 139                 if (sysfs_attr_match(cp, "spare"))
 140                         rv |= DS_SPARE;
 141                 if (sysfs_attr_match(cp, "blocked"))
 142                         rv |= DS_BLOCKED;
 143                 cp = strchr(cp, ',');
 144                 if (cp)
 145                         cp++;
 146         }
 147         return rv;
 148 }
 149
 150 static void signal_manager(void)
 151 {
 152         /* tgkill(getpid(), mon_tid, SIGUSR1); */
 153         int pid = getpid();
 154         syscall(SYS_tgkill, pid, mgr_tid, SIGUSR1);
 155 }
 156
 157 /* Monitor a set of active md arrays - all of which share the
 158  * same metadata - and respond to events that require
 159  * metadata update.
 160  *
 161  * New arrays are detected by another thread which allocates
 162  * required memory and attaches the data structure to our list.
 163  *
 164  * Events:
 165  *  Array stops.
 166  *    This is detected by array_state going to 'clear' or 'inactive'.
 167  *    while we thought it was active.
 168  *    Response is to mark metadata as clean and 'clear' the array(??)
 169  *  write-pending
 170  *    array_state if 'write-pending'
 171  *    We mark metadata as 'dirty' then set array to 'active'.
 172  *  active_idle
 173  *    Either ignore, or mark clean, then mark metadata as clean.
 174  *
 175  *  device fails
 176  *    detected by rd-N/state reporting "faulty"
 177  *    mark device as 'failed' in metadata, let the kernel release the
 178  *    device by writing '-blocked' to rd/state, and finally write 'remove' to
 179  *    rd/state.  Before a disk can be replaced it must be failed and removed
 180  *    from all container members, this will be preemptive for the other
 181  *    arrays... safe?
 182  *
 183  *  sync completes
 184  *    sync_action was 'resync' and becomes 'idle' and resync_start becomes
 185  *    MaxSector
 186  *    Notify metadata that sync is complete.
 187  *
 188  *  recovery completes
 189  *    sync_action changes from 'recover' to 'idle'
 190  *    Check each device state and mark metadata if 'faulty' or 'in_sync'.
 191  *
 192  *  deal with resync
 193  *    This only happens on finding a new array... mdadm will have set
 194  *    'resync_start' to the correct value.  If 'resync_start' indicates that an
 195  *    resync needs to occur set the array to the 'active' state rather than the
 196  *    initial read-auto state.
 197  *
 198  *
 199  *
 200  * We wait for a change (poll/select) on array_state, sync_action, and
 201  * each rd-X/state file.
 202  * When we get any change, we check everything.  So read each state file,
 203  * then decide what to do.
 204  *
 205  * The core action is to write new metadata to all devices in the array.
 206  * This is done at most once on any wakeup.
 207  * After that we might:
 208  *   - update the array_state
 209  *   - set the role of some devices.
 210  *   - request a sync_action
 211  *
 212  */
 213
 214 static int read_and_act(struct active_array *a)
 215 {
 216         unsigned long long sync_completed;
 217         int check_degraded = 0;
 218         int check_reshape = 0;
 219         int deactivate = 0;
 220         struct mdinfo *mdi;
 221         int dirty = 0;
 222         int count = 0;
 223
 224         a->next_state = bad_word;
 225         a->next_action = bad_action;
 226
 227         a->curr_state = read_state(a->info.state_fd);
 228         a->curr_action = read_action(a->action_fd);
 229         a->info.resync_start = read_resync_start(a->resync_start_fd);
 230         sync_completed = read_sync_completed(a->sync_completed_fd);
 231         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 232                 mdi->next_state = 0;
 233                 mdi->curr_state = 0;
 234                 if (mdi->state_fd >= 0) {
 235                         mdi->recovery_start = read_resync_start(mdi->recovery_fd);
 236                         mdi->curr_state = read_dev_state(mdi->state_fd);
 237                 }
 238         }
 239
 240         if (a->curr_state > inactive &&
 241             a->prev_state == inactive) {
 242                 /* array has been started
 243                  * possible that container operation has to be completed
 244                  */
 245                 a->container->ss->set_array_state(a, 0);
 246         }
 247         if (a->curr_state <= inactive &&
 248             a->prev_state > inactive) {
 249                 /* array has been stopped */
 250                 a->container->ss->set_array_state(a, 1);
 251                 a->next_state = clear;
 252                 deactivate = 1;
 253         }
 254         if (a->curr_state == write_pending) {
 255                 a->container->ss->set_array_state(a, 0);
 256                 a->next_state = active;
 257                 dirty = 1;
 258         }
 259         if (a->curr_state == active_idle) {
 260                 /* Set array to 'clean' FIRST, then mark clean
 261                  * in the metadata
 262                  */
 263                 a->next_state = clean;
 264                 dirty = 1;
 265         }
 266         if (a->curr_state == clean) {
 267                 a->container->ss->set_array_state(a, 1);
 268         }
 269         if (a->curr_state == active ||
 270             a->curr_state == suspended ||
 271             a->curr_state == bad_word)
 272                 dirty = 1;
 273         if (a->curr_state == readonly) {
 274                 /* Well, I'm ready to handle things.  If readonly
 275                  * wasn't requested, transition to read-auto.
 276                  */
 277                 char buf[64];
 278                 read_attr(buf, sizeof(buf), a->metadata_fd);
 279                 if (strncmp(buf, "external:-", 10) == 0) {
 280                         /* explicit request for readonly array.  Leave it alone */
 281                         ;
 282                 } else {
 283                         if (a->container->ss->set_array_state(a, 2))
 284                                 a->next_state = read_auto; /* array is clean */
 285                         else {
 286                                 a->next_state = active; /* Now active for recovery etc */
 287                                 dirty = 1;
 288                         }
 289                 }
 290         }
 291
 292         if (!deactivate &&
 293             a->curr_action == idle &&
 294             a->prev_action == resync) {
 295                 /* A resync has finished.  The endpoint is recorded in
 296                  * 'sync_start'.  We don't update the metadata
 297                  * until the array goes inactive or readonly though.
 298                  * Just check if we need to fiddle spares.
 299                  */
 300                 a->container->ss->set_array_state(a, a->curr_state <= clean);
 301                 check_degraded = 1;
 302         }
 303
 304         if (!deactivate &&
 305             a->curr_action == idle &&
 306             a->prev_action == recover) {
 307                 /* A recovery has finished.  Some disks may be in sync now,
 308                  * and the array may no longer be degraded
 309                  */
 310                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 311                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 312                                                    mdi->curr_state);
 313                         if (! (mdi->curr_state & DS_INSYNC))
 314                                 check_degraded = 1;
 315                         count++;
 316                 }
 317                 if (count != a->info.array.raid_disks)
 318                         check_degraded = 1;
 319         }
 320
 321         if (!deactivate &&
 322             a->curr_action == reshape &&
 323             a->prev_action != reshape)
 324                 /* reshape was requested by mdadm.  Need to see if
 325                  * new devices have been added.  Manager does that
 326                  * when it sees check_reshape
 327                  */
 328                 check_reshape = 1;
 329
 330         /* Check for failures and if found:
 331          * 1/ Record the failure in the metadata and unblock the device.
 332          *    FIXME update the kernel to stop notifying on failed drives when
 333          *    the array is readonly and we have cleared 'blocked'
 334          * 2/ Try to remove the device if the array is writable, or can be
 335          *    made writable.
 336          */
 337         for (mdi = a->info.devs ; mdi ; mdi = mdi->next) {
 338                 if (mdi->curr_state & DS_FAULTY) {
 339                         a->container->ss->set_disk(a, mdi->disk.raid_disk,
 340                                                    mdi->curr_state);
 341                         check_degraded = 1;
 342                         mdi->next_state |= DS_UNBLOCK;
 343                         if (a->curr_state == read_auto) {
 344                                 a->container->ss->set_array_state(a, 0);
 345                                 a->next_state = active;
 346                         }
 347                         if (a->curr_state > readonly)
 348                                 mdi->next_state |= DS_REMOVE;
 349                 }
 350         }
 351
 352         /* Check for recovery checkpoint notifications.  We need to be a
 353          * minimum distance away from the last checkpoint to prevent
 354          * over checkpointing.  Note reshape checkpointing is handled
 355          * in the second branch.
 356          */
 357         if (sync_completed > a->last_checkpoint &&
 358             sync_completed - a->last_checkpoint > a->info.component_size >> 4 &&
 359             a->curr_action > reshape) {
 360                 /* A (non-reshape) sync_action has reached a checkpoint.
 361                  * Record the updated position in the metadata
 362                  */
 363                 a->last_checkpoint = sync_completed;
 364                 a->container->ss->set_array_state(a, a->curr_state <= clean);
 365         } else if ((a->curr_action == idle && a->prev_action == reshape) ||
 366                    (a->curr_action == reshape
 367                     && sync_completed > a->last_checkpoint) ) {
 368                 /* Reshape has progressed or completed so we need to
 369                  * update the array state - and possibly the array size
 370                  */
 371                 if (sync_completed != 0)
 372                         a->last_checkpoint = sync_completed;
 373                 /* We might need to update last_checkpoint depending on
 374                  * the reason that reshape finished.
 375                  * if array reshape is really finished:
 376                  *        set check point to the end, this allows
 377                  *        set_array_state() to finalize reshape in metadata
 378                  * if reshape if broken: do not set checkpoint to the end
 379                  *        this allows for reshape restart from checkpoint
 380                  */
 381                 if ((a->curr_action != reshape) &&
 382                     (a->prev_action == reshape)) {
 383                         char buf[40];
 384                         if ((sysfs_get_str(&a->info, NULL,
 385                                           "reshape_position",
 386                                           buf,
 387                                           sizeof(buf)) >= 0) &&
 388                              strncmp(buf, "none", 4) == 0)
 389                                 a->last_checkpoint = a->info.component_size;
 390                 }
 391                 a->container->ss->set_array_state(a, a->curr_state <= clean);
 392                 a->last_checkpoint = sync_completed;
 393         }
 394
 395         if (sync_completed > a->last_checkpoint)
 396                 a->last_checkpoint = sync_completed;
 397
 398         a->container->ss->sync_metadata(a->container);
 399         dprintf("%s(%d): state:%s action:%s next(", __func__, a->info.container_member,
 400                 array_states[a->curr_state], sync_actions[a->curr_action]);
 401
 402         /* Effect state changes in the array */
 403         if (a->next_state != bad_word) {
 404                 dprintf(" state:%s", array_states[a->next_state]);
 405                 write_attr(array_states[a->next_state], a->info.state_fd);
 406         }
 407         if (a->next_action != bad_action) {
 408                 write_attr(sync_actions[a->next_action], a->action_fd);
 409                 dprintf(" action:%s", sync_actions[a->next_action]);
 410         }
 411         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 412                 if (mdi->next_state & DS_UNBLOCK) {
 413                         dprintf(" %d:-blocked", mdi->disk.raid_disk);
 414                         write_attr("-blocked", mdi->state_fd);
 415                 }
 416
 417                 if ((mdi->next_state & DS_REMOVE) && mdi->state_fd >= 0) {
 418                         int remove_result;
 419
 420                         /* the kernel may not be able to immediately remove the
 421                          * disk, we can simply wait until the next event to try
 422                          * again.
 423                          */
 424                         remove_result = write_attr("remove", mdi->state_fd);
 425                         if (remove_result > 0) {
 426                                 dprintf(" %d:removed", mdi->disk.raid_disk);
 427                                 close(mdi->state_fd);
 428                                 close(mdi->recovery_fd);
 429                                 mdi->state_fd = -1;
 430                         }
 431                 }
 432                 if (mdi->next_state & DS_INSYNC) {
 433                         write_attr("+in_sync", mdi->state_fd);
 434                         dprintf(" %d:+in_sync", mdi->disk.raid_disk);
 435                 }
 436         }
 437         dprintf(" )\n");
 438
 439         /* move curr_ to prev_ */
 440         a->prev_state = a->curr_state;
 441
 442         a->prev_action = a->curr_action;
 443
 444         for (mdi = a->info.devs; mdi ; mdi = mdi->next) {
 445                 mdi->prev_state = mdi->curr_state;
 446                 mdi->next_state = 0;
 447         }
 448
 449         if (check_degraded || check_reshape) {
 450                 /* manager will do the actual check */
 451                 if (check_degraded)
 452                         a->check_degraded = 1;
 453                 if (check_reshape)
 454                         a->check_reshape = 1;
 455                 signal_manager();
 456         }
 457
 458         if (deactivate)
 459                 a->container = NULL;
 460
 461         return dirty;
 462 }
 463
 464 static struct mdinfo *
 465 find_device(struct active_array *a, int major, int minor)
 466 {
 467         struct mdinfo *mdi;
 468
 469         for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 470                 if (mdi->disk.major == major && mdi->disk.minor == minor)
 471                         return mdi;
 472
 473         return NULL;
 474 }
 475
 476 static void reconcile_failed(struct active_array *aa, struct mdinfo *failed)
 477 {
 478         struct active_array *a;
 479         struct mdinfo *victim;
 480
 481         for (a = aa; a; a = a->next) {
 482                 if (!a->container)
 483                         continue;
 484                 victim = find_device(a, failed->disk.major, failed->disk.minor);
 485                 if (!victim)
 486                         continue;
 487
 488                 if (!(victim->curr_state & DS_FAULTY))
 489                         write_attr("faulty", victim->state_fd);
 490         }
 491 }
 492
 493 #ifdef DEBUG
 494 static void dprint_wake_reasons(fd_set *fds)
 495 {
 496         int i;
 497         char proc_path[256];
 498         char link[256];
 499         char *basename;
 500         int rv;
 501
 502         fprintf(stderr, "monitor: wake ( ");
 503         for (i = 0; i < FD_SETSIZE; i++) {
 504                 if (FD_ISSET(i, fds)) {
 505                         sprintf(proc_path, "/proc/%d/fd/%d",
 506                                 (int) getpid(), i);
 507
 508                         rv = readlink(proc_path, link, sizeof(link) - 1);
 509                         if (rv < 0) {
 510                                 fprintf(stderr, "%d:unknown ", i);
 511                                 continue;
 512                         }
 513                         link[rv] = '\0';
 514                         basename = strrchr(link, '/');
 515                         fprintf(stderr, "%d:%s ",
 516                                 i, basename ? ++basename : link);
 517                 }
 518         }
 519         fprintf(stderr, ")\n");
 520 }
 521 #endif
 522
 523 int monitor_loop_cnt;
 524
 525 static int wait_and_act(struct supertype *container, int nowait)
 526 {
 527         fd_set rfds;
 528         int maxfd = 0;
 529         struct active_array **aap = &container->arrays;
 530         struct active_array *a, **ap;
 531         int rv;
 532         struct mdinfo *mdi;
 533         static unsigned int dirty_arrays = ~0; /* start at some non-zero value */
 534
 535         FD_ZERO(&rfds);
 536
 537         for (ap = aap ; *ap ;) {
 538                 a = *ap;
 539                 /* once an array has been deactivated we want to
 540                  * ask the manager to discard it.
 541                  */
 542                 if (!a->container) {
 543                         if (discard_this) {
 544                                 ap = &(*ap)->next;
 545                                 continue;
 546                         }
 547                         *ap = a->next;
 548                         a->next = NULL;
 549                         discard_this = a;
 550                         signal_manager();
 551                         continue;
 552                 }
 553
 554                 add_fd(&rfds, &maxfd, a->info.state_fd);
 555                 add_fd(&rfds, &maxfd, a->action_fd);
 556                 add_fd(&rfds, &maxfd, a->sync_completed_fd);
 557                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 558                         add_fd(&rfds, &maxfd, mdi->state_fd);
 559
 560                 ap = &(*ap)->next;
 561         }
 562
 563         if (manager_ready && (*aap == NULL || (sigterm && !dirty_arrays))) {
 564                 /* No interesting arrays, or we have been told to
 565                  * terminate and everything is clean.  Lets see about
 566                  * exiting.  Note that blocking at this point is not a
 567                  * problem as there are no active arrays, there is
 568                  * nothing that we need to be ready to do.
 569                  */
 570                 int fd;
 571                 if (sigterm)
 572                         fd = open_dev_excl(container->devnum);
 573                 else
 574                         fd = open_dev_flags(container->devnum, O_RDONLY|O_EXCL);
 575                 if (fd >= 0 || errno != EBUSY) {
 576                         /* OK, we are safe to leave */
 577                         if (sigterm && !dirty_arrays)
 578                                 dprintf("caught sigterm, all clean... exiting\n");
 579                         else
 580                                 dprintf("no arrays to monitor... exiting\n");
 581                         if (!sigterm)
 582                                 /* On SIGTERM, someone (the take-over mdmon) will
 583                                  * clean up
 584                                  */
 585                                 remove_pidfile(container->devname);
 586                         exit_now = 1;
 587                         signal_manager();
 588                         close(fd);
 589                         exit(0);
 590                 }
 591         }
 592
 593         if (!nowait) {
 594                 sigset_t set;
 595                 struct timespec ts;
 596                 ts.tv_sec = 24*3600;
 597                 ts.tv_nsec = 0;
 598                 if (*aap == NULL) {
 599                         /* just waiting to get O_EXCL access */
 600                         ts.tv_sec = 0;
 601                         ts.tv_nsec = 20000000ULL;
 602                 }
 603                 sigprocmask(SIG_UNBLOCK, NULL, &set);
 604                 sigdelset(&set, SIGUSR1);
 605                 monitor_loop_cnt |= 1;
 606                 rv = pselect(maxfd+1, NULL, NULL, &rfds, &ts, &set);
 607                 monitor_loop_cnt += 1;
 608                 if (rv == -1 && errno == EINTR)
 609                         rv = 0;
 610                 #ifdef DEBUG
 611                 dprint_wake_reasons(&rfds);
 612                 #endif
 613
 614         }
 615
 616         if (update_queue) {
 617                 struct metadata_update *this;
 618
 619                 for (this = update_queue; this ; this = this->next)
 620                         container->ss->process_update(container, this);
 621
 622                 update_queue_handled = update_queue;
 623                 update_queue = NULL;
 624                 signal_manager();
 625                 container->ss->sync_metadata(container);
 626         }
 627
 628         rv = 0;
 629         dirty_arrays = 0;
 630         for (a = *aap; a ; a = a->next) {
 631                 int is_dirty;
 632
 633                 if (a->replaces && !discard_this) {
 634                         struct active_array **ap;
 635                         for (ap = &a->next; *ap && *ap != a->replaces;
 636                              ap = & (*ap)->next)
 637                                 ;
 638                         if (*ap)
 639                                 *ap = (*ap)->next;
 640                         discard_this = a->replaces;
 641                         a->replaces = NULL;
 642                         /* FIXME check if device->state_fd need to be cleared?*/
 643                         signal_manager();
 644                 }
 645                 if (a->container) {
 646                         is_dirty = read_and_act(a);
 647                         rv |= 1;
 648                         dirty_arrays += is_dirty;
 649                         /* when terminating stop manipulating the array after it
 650                          * is clean, but make sure read_and_act() is given a
 651                          * chance to handle 'active_idle'
 652                          */
 653                         if (sigterm && !is_dirty)
 654                                 a->container = NULL; /* stop touching this array */
 655                 }
 656         }
 657
 658         /* propagate failures across container members */
 659         for (a = *aap; a ; a = a->next) {
 660                 if (!a->container)
 661                         continue;
 662                 for (mdi = a->info.devs ; mdi ; mdi = mdi->next)
 663                         if (mdi->curr_state & DS_FAULTY)
 664                                 reconcile_failed(*aap, mdi);
 665         }
 666
 667         return rv;
 668 }
 669
 670 void do_monitor(struct supertype *container)
 671 {
 672         int rv;
 673         int first = 1;
 674         do {
 675                 rv = wait_and_act(container, first);
 676                 first = 0;
 677         } while (rv >= 0);
 678 }