multipathd/main.c

   1 /*
   2  * Copyright (c) 2004, 2005 Christophe Varoqui
   3  * Copyright (c) 2005 Kiyoshi Ueda, NEC
   4  * Copyright (c) 2005 Benjamin Marzinski, Redhat
   5  * Copyright (c) 2005 Edward Goggin, EMC
   6  */
   7 #include <unistd.h>
   8 #include <sys/stat.h>
   9 #include <libdevmapper.h>
  10 #include <wait.h>
  11 #include <sys/mman.h>
  12 #include <sys/types.h>
  13 #include <fcntl.h>
  14 #include <errno.h>
  15
  16 /*
  17  * libsysfs
  18  */
  19 #include <sysfs/libsysfs.h>
  20 #include <sysfs/dlist.h>
  21
  22 /*
  23  * libcheckers
  24  */
  25 #include <checkers.h>
  26
  27 /*
  28  * libmultipath
  29  */
  30 #include <parser.h>
  31 #include <vector.h>
  32 #include <memory.h>
  33 #include <config.h>
  34 #include <util.h>
  35 #include <hwtable.h>
  36 #include <defaults.h>
  37 #include <structs.h>
  38 #include <callout.h>
  39 #include <blacklist.h>
  40 #include <structs_vec.h>
  41 #include <dmparser.h>
  42 #include <devmapper.h>
  43 #include <dict.h>
  44 #include <discovery.h>
  45 #include <debug.h>
  46 #include <propsel.h>
  47 #include <uevent.h>
  48 #include <switchgroup.h>
  49 #include <print.h>
  50 #include <configure.h>
  51
  52 #include "main.h"
  53 #include "pidfile.h"
  54 #include "uxlsnr.h"
  55 #include "uxclnt.h"
  56 #include "cli.h"
  57 #include "cli_handlers.h"
  58 #include "lock.h"
  59 #include "waiter.h"
  60
  61 #define FILE_NAME_SIZE 256
  62 #define CMDSIZE 160
  63
  64 #define LOG_MSG(a,b) \
  65         if (strlen(b)) condlog(a, "%s: %s", pp->dev_t, b);
  66
  67 pthread_cond_t exit_cond = PTHREAD_COND_INITIALIZER;
  68 pthread_mutex_t exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  69
  70 /*
  71  * global copy of vecs for use in sig handlers
  72  */
  73 struct vectors * gvecs;
  74
  75 static int
  76 need_switch_pathgroup (struct multipath * mpp, int refresh)
  77 {
  78         struct pathgroup * pgp;
  79         struct path * pp;
  80         int i, j;
  81
  82         if (!mpp || mpp->pgfailback == -FAILBACK_MANUAL)
  83                 return 0;
  84
  85         /*
  86          * Refresh path priority values
  87          */
  88         if (refresh)
  89                 vector_foreach_slot (mpp->pg, pgp, i)
  90                         vector_foreach_slot (pgp->paths, pp, j)
  91                                 pathinfo(pp, conf->hwtable, DI_PRIO);
  92
  93         mpp->bestpg = select_path_group(mpp);
  94
  95         if (mpp->bestpg != mpp->nextpg)
  96                 return 1;
  97
  98         return 0;
  99 }
 100
 101 static void
 102 switch_pathgroup (struct multipath * mpp)
 103 {
 104         mpp->stat_switchgroup++;
 105         dm_switchgroup(mpp->alias, mpp->bestpg);
 106         condlog(2, "%s: switch to path group #%i",
 107                  mpp->alias, mpp->bestpg);
 108 }
 109
 110 static int
 111 coalesce_maps(struct vectors *vecs, vector nmpv)
 112 {
 113         struct multipath * ompp;
 114         vector ompv = vecs->mpvec;
 115         int i, j;
 116
 117         vector_foreach_slot (ompv, ompp, i) {
 118                 if (!find_mp_by_wwid(nmpv, ompp->wwid)) {
 119                         /*
 120                          * remove all current maps not allowed by the
 121                          * current configuration
 122                          */
 123                         if (dm_flush_map(ompp->alias, DEFAULT_TARGET)) {
 124                                 condlog(0, "%s: unable to flush devmap",
 125                                         ompp->alias);
 126                                 /*
 127                                  * may be just because the device is open
 128                                  */
 129                                 if (!vector_alloc_slot(nmpv))
 130                                         return 1;
 131
 132                                 vector_set_slot(nmpv, ompp);
 133                                 setup_multipath(vecs, ompp);
 134
 135                                 if ((j = find_slot(ompv, (void *)ompp)) != -1)
 136                                         vector_del_slot(ompv, j);
 137
 138                                 continue;
 139                         }
 140                         else {
 141                                 dm_lib_release();
 142                                 condlog(3, "%s devmap removed", ompp->alias);
 143                         }
 144                 }
 145         }
 146         return 0;
 147 }
 148
 149 static void
 150 sync_map_state(struct multipath *mpp)
 151 {
 152         int i, j;
 153         struct pathgroup *pgp;
 154         struct path *pp;
 155
 156         vector_foreach_slot (mpp->pg, pgp, i){
 157                 vector_foreach_slot (pgp->paths, pp, j){
 158                         if (pp->state <= PATH_UNCHECKED)
 159                                 continue;
 160                         if ((pp->dmstate == PSTATE_FAILED ||
 161                              pp->dmstate == PSTATE_UNDEF) &&
 162                             (pp->state == PATH_UP || pp->state == PATH_GHOST))
 163                                 dm_reinstate_path(mpp->alias, pp->dev_t);
 164                         else if ((pp->dmstate == PSTATE_ACTIVE ||
 165                                   pp->dmstate == PSTATE_UNDEF) &&
 166                                  (pp->state == PATH_DOWN ||
 167                                   pp->state == PATH_SHAKY))
 168                                 dm_fail_path(mpp->alias, pp->dev_t);
 169                 }
 170         }
 171 }
 172
 173 static void
 174 sync_maps_state(vector mpvec)
 175 {
 176         int i;
 177         struct multipath *mpp;
 178
 179         vector_foreach_slot (mpvec, mpp, i)
 180                 sync_map_state(mpp);
 181 }
 182
 183 static int
 184 flush_map(struct multipath * mpp, struct vectors * vecs)
 185 {
 186         /*
 187          * clear references to this map before flushing so we can ignore
 188          * the spurious uevent we may generate with the dm_flush_map call below
 189          */
 190         if (dm_flush_map(mpp->alias, DEFAULT_TARGET)) {
 191                 /*
 192                  * May not really be an error -- if the map was already flushed
 193                  * from the device mapper by dmsetup(8) for instance.
 194                  */
 195                 condlog(0, "%s: can't flush", mpp->alias);
 196                 return 1;
 197         }
 198         else {
 199                 dm_lib_release();
 200                 condlog(3, "%s: devmap removed", mpp->alias);
 201         }
 202
 203         orphan_paths(vecs->pathvec, mpp);
 204         remove_map(mpp, vecs, stop_waiter_thread, 1);
 205
 206         return 0;
 207 }
 208
 209 static int
 210 uev_add_map (char * devname, struct vectors * vecs)
 211 {
 212         condlog(2, "%s: add map (uevent)", devname);
 213         return ev_add_map(devname, vecs);
 214 }
 215
 216 int
 217 ev_add_map (char * devname, struct vectors * vecs)
 218 {
 219         int major, minor;
 220         char dev_t[BLK_DEV_SIZE];
 221         char * alias;
 222         char * refwwid;
 223         struct multipath * mpp;
 224         int map_present;
 225         int r = 1;
 226
 227         if (sscanf(devname, "dm-%d", &minor) == 1 &&
 228             !sysfs_get_dev(sysfs_path, devname, dev_t, BLK_DEV_SIZE) &&
 229             sscanf(dev_t, "%d:%d", &major, &minor) == 2)
 230                 alias = dm_mapname(major, minor);
 231         else
 232                 alias = STRDUP(devname);
 233
 234         if (!alias)
 235                 return 1;
 236
 237         map_present = dm_map_present(alias);
 238
 239         if (map_present && dm_type(alias, DEFAULT_TARGET) <= 0) {
 240                 condlog(4, "%s: not a multipath map", alias);
 241                 FREE(alias);
 242                 return 0;
 243         }
 244
 245         mpp = find_mp_by_alias(vecs->mpvec, alias);
 246
 247         if (mpp) {
 248                 /*
 249                  * Not really an error -- we generate our own uevent
 250                  * if we create a multipath mapped device as a result
 251                  * of uev_add_path
 252                  */
 253                 condlog(0, "%s: devmap already registered",
 254                         devname);
 255                 FREE(alias);
 256                 return 0;
 257         }
 258
 259         /*
 260          * now we can register the map
 261          */
 262         if (map_present && (mpp = add_map_without_path(vecs, minor, alias,
 263                                         start_waiter_thread))) {
 264                 sync_map_state(mpp);
 265                 condlog(3, "%s: devmap %s added", alias, devname);
 266                 return 0;
 267         }
 268         refwwid = get_refwwid(devname, DEV_DEVMAP, vecs->pathvec);
 269
 270         if (refwwid) {
 271                 r = coalesce_paths(vecs, NULL, refwwid);
 272                 dm_lib_release();
 273         }
 274
 275         if (!r)
 276                 condlog(3, "%s: devmap %s added", alias, devname);
 277         else
 278                 condlog(0, "%s: uev_add_map %s failed", alias, devname);
 279
 280         FREE(refwwid);
 281         FREE(alias);
 282         return r;
 283 }
 284
 285 static int
 286 uev_remove_map (char * devname, struct vectors * vecs)
 287 {
 288         condlog(2, "%s: remove map (uevent)", devname);
 289         return ev_remove_map(devname, vecs);
 290 }
 291
 292 int
 293 ev_remove_map (char * devname, struct vectors * vecs)
 294 {
 295         struct multipath * mpp;
 296
 297         mpp = find_mp_by_str(vecs->mpvec, devname);
 298
 299         if (!mpp) {
 300                 condlog(3, "%s: devmap not registered, can't remove",
 301                         devname);
 302                 return 0;
 303         }
 304         flush_map(mpp, vecs);
 305
 306         return 0;
 307 }
 308
 309 static int
 310 uev_umount_map (char * devname, struct vectors * vecs)
 311 {
 312         struct multipath * mpp;
 313
 314         condlog(2, "%s: umount map (uevent)", devname);
 315
 316         mpp = find_mp_by_str(vecs->mpvec, devname);
 317
 318         if (!mpp)
 319                 return 0;
 320
 321         update_mpp_paths(mpp, vecs->pathvec);
 322         verify_paths(mpp, vecs, NULL);
 323
 324         if (!VECTOR_SIZE(mpp->paths))
 325                 flush_map(mpp, vecs);
 326
 327         return 0;
 328 }
 329
 330 static int
 331 uev_add_path (char * devname, struct vectors * vecs)
 332 {
 333         condlog(2, "%s: add path (uevent)", devname);
 334         return (ev_add_path(devname, vecs) != 1)? 0 : 1;
 335 }
 336
 337
 338 /*
 339  * returns:
 340  * 0: added
 341  * 1: error
 342  * 2: blacklisted
 343  */
 344 int
 345 ev_add_path (char * devname, struct vectors * vecs)
 346 {
 347         struct multipath * mpp;
 348         struct path * pp;
 349         char empty_buff[WWID_SIZE] = {0};
 350
 351         pp = find_path_by_dev(vecs->pathvec, devname);
 352
 353         if (pp) {
 354                 condlog(0, "%s: spurious uevent, path already in pathvec",
 355                         devname);
 356                 if (pp->mpp)
 357                         return 0;
 358         }
 359         else {
 360                 /*
 361                  * get path vital state
 362                  */
 363                 if (!(pp = store_pathinfo(vecs->pathvec, conf->hwtable,
 364                       devname, DI_ALL))) {
 365                         condlog(0, "%s: failed to store path info", devname);
 366                         return 1;
 367                 }
 368                 pp->checkint = conf->checkint;
 369         }
 370
 371         /*
 372          * need path UID to go any further
 373          */
 374         if (memcmp(empty_buff, pp->wwid, WWID_SIZE) == 0) {
 375                 condlog(0, "%s: failed to get path uid", devname);
 376                 return 1; /* leave path added to pathvec */
 377         }
 378         if (blacklist_path(conf, pp)){
 379                 int i = find_slot(vecs->pathvec, (void *)pp);
 380                 if (i != -1)
 381                         vector_del_slot(vecs->pathvec, i);
 382                 free_path(pp);
 383                 return 2;
 384         }
 385         mpp = pp->mpp = find_mp_by_wwid(vecs->mpvec, pp->wwid);
 386 rescan:
 387         if (mpp) {
 388                 if (adopt_paths(vecs->pathvec, mpp))
 389                         return 1; /* leave path added to pathvec */
 390
 391                 verify_paths(mpp, vecs, NULL);
 392                 mpp->action = ACT_RELOAD;
 393         }
 394         else {
 395                 if ((mpp = add_map_with_path(vecs, pp, 1)))
 396                         mpp->action = ACT_CREATE;
 397                 else
 398                         return 1; /* leave path added to pathvec */
 399         }
 400
 401         /*
 402          * push the map to the device-mapper
 403          */
 404         if (setup_map(mpp)) {
 405                 condlog(0, "%s: failed to setup map for addition of new "
 406                         "path %s", mpp->alias, devname);
 407                 goto out;
 408         }
 409         /*
 410          * reload the map for the multipath mapped device
 411          */
 412         if (domap(mpp) <= 0) {
 413                 condlog(0, "%s: failed in domap for addition of new "
 414                         "path %s", mpp->alias, devname);
 415                 /*
 416                  * deal with asynchronous uevents :((
 417                  */
 418                 if (mpp->action == ACT_RELOAD) {
 419                         condlog(0, "%s: uev_add_path sleep", mpp->alias);
 420                         sleep(1);
 421                         update_mpp_paths(mpp, vecs->pathvec);
 422                         goto rescan;
 423                 }
 424                 else
 425                         goto out;
 426         }
 427         dm_lib_release();
 428
 429         /*
 430          * update our state from kernel regardless of create or reload
 431          */
 432         if (setup_multipath(vecs, mpp))
 433                 goto out;
 434
 435         sync_map_state(mpp);
 436
 437         if (mpp->action == ACT_CREATE &&
 438             start_waiter_thread(mpp, vecs))
 439                         goto out;
 440
 441         condlog(3, "%s path added to devmap %s", devname, mpp->alias);
 442         return 0;
 443
 444 out:
 445         remove_map(mpp, vecs, NULL, 1);
 446         return 1;
 447 }
 448
 449 static int
 450 uev_remove_path (char * devname, struct vectors * vecs)
 451 {
 452         condlog(2, "%s: remove path (uevent)", devname);
 453         return ev_remove_path(devname, vecs);
 454 }
 455
 456 int
 457 ev_remove_path (char * devname, struct vectors * vecs)
 458 {
 459         struct multipath * mpp;
 460         struct path * pp;
 461         int i;
 462         int rm_path = 1;
 463
 464         pp = find_path_by_dev(vecs->pathvec, devname);
 465
 466         if (!pp) {
 467                 condlog(0, "%s: spurious uevent, path not in pathvec", devname);
 468                 return 1;
 469         }
 470
 471         /*
 472          * avoid referring to the map of an orphanned path
 473          */
 474         if ((mpp = pp->mpp)) {
 475
 476                 /*
 477                  * remove the map IFF removing the last path
 478                  */
 479                 if (pathcount(mpp, PATH_WILD) > 1) {
 480                         vector rpvec = vector_alloc();
 481
 482                         /*
 483                          * transform the mp->pg vector of vectors of paths
 484                          * into a mp->params string to feed the device-mapper
 485                          */
 486                         update_mpp_paths(mpp, vecs->pathvec);
 487                         if ((i = find_slot(mpp->paths, (void *)pp)) != -1)
 488                                 vector_del_slot(mpp->paths, i);
 489
 490                         if (VECTOR_SIZE(mpp->paths) == 0) {
 491                                 char alias[WWID_SIZE];
 492
 493                                 /*
 494                                  * flush_map will fail if the device is open
 495                                  */
 496                                 strncpy(alias, mpp->alias, WWID_SIZE);
 497                                 if (flush_map(mpp, vecs))
 498                                         rm_path = 0;
 499                                 else
 500                                         condlog(3, "%s: removed map after removing"
 501                                                 " multiple paths", alias);
 502                         }
 503                         else {
 504                                 if (setup_map(mpp)) {
 505                                         condlog(0, "%s: failed to setup map for"
 506                                                 " removal of path %s", mpp->alias, devname);
 507                                         free_pathvec(rpvec, KEEP_PATHS);
 508                                         goto out;
 509                                 }
 510                                 /*
 511                                  * reload the map
 512                                  */
 513                                 mpp->action = ACT_RELOAD;
 514                                 if (domap(mpp) <= 0) {
 515                                         condlog(0, "%s: failed in domap for "
 516                                                 "removal of path %s",
 517                                                 mpp->alias, devname);
 518                                         /*
 519                                          * Delete path from pathvec so that
 520                                          * update_mpp_paths wont find it later
 521                                          * when/if another path is removed.
 522                                          */
 523                                         if ((i = find_slot(vecs->pathvec, (void *)pp)) != -1)
 524                                                 vector_del_slot(vecs->pathvec, i);
 525                                         free_path(pp);
 526                                         return 1;
 527                                 }
 528                                 /*
 529                                  * update our state from kernel
 530                                  */
 531                                 if (setup_multipath(vecs, mpp)) {
 532                                         free_pathvec(rpvec, KEEP_PATHS);
 533                                         goto out;
 534                                 }
 535                                 sync_map_state(mpp);
 536
 537                                 condlog(3, "%s path removed from devmap %s",
 538                                         devname, mpp->alias);
 539                         }
 540                         free_pathvec(rpvec, KEEP_PATHS);
 541                 }
 542                 else {
 543                         char alias[WWID_SIZE];
 544
 545                         /*
 546                          * flush_map will fail if the device is open
 547                          */
 548                         strncpy(alias, mpp->alias, WWID_SIZE);
 549                         if (flush_map(mpp, vecs))
 550                                 rm_path = 0;
 551                         else
 552                                 condlog(3, "%s: removed map", alias);
 553                 }
 554         }
 555
 556         if (rm_path) {
 557                 if ((i = find_slot(vecs->pathvec, (void *)pp)) != -1)
 558                         vector_del_slot(vecs->pathvec, i);
 559                 free_path(pp);
 560         }
 561
 562         return 0;
 563
 564 out:
 565         remove_map(mpp, vecs, stop_waiter_thread, 1);
 566         return 1;
 567 }
 568
 569 static int
 570 map_discovery (struct vectors * vecs)
 571 {
 572         int i;
 573         struct multipath * mpp;
 574
 575         if (dm_get_maps(vecs->mpvec, "multipath"))
 576                 return 1;
 577
 578         vector_foreach_slot (vecs->mpvec, mpp, i)
 579                 if (setup_multipath(vecs, mpp))
 580                         return 1;
 581
 582         return 0;
 583 }
 584
 585 int
 586 uxsock_trigger (char * str, char ** reply, int * len, void * trigger_data)
 587 {
 588         struct vectors * vecs;
 589         int r;
 590
 591         *reply = NULL;
 592         *len = 0;
 593         vecs = (struct vectors *)trigger_data;
 594
 595         pthread_cleanup_push(cleanup_lock, vecs->lock);
 596         lock(vecs->lock);
 597
 598         r = parse_cmd(str, reply, len, vecs);
 599
 600         if (r > 0) {
 601                 *reply = STRDUP("fail\n");
 602                 *len = strlen(*reply) + 1;
 603                 r = 1;
 604         }
 605         else if (!r && *len == 0) {
 606                 *reply = STRDUP("ok\n");
 607                 *len = strlen(*reply) + 1;
 608                 r = 0;
 609         }
 610         /* else if (r < 0) leave *reply alone */
 611
 612         lock_cleanup_pop(vecs->lock);
 613         return r;
 614 }
 615
 616 static int
 617 uev_discard(char * devpath)
 618 {
 619         char a[10], b[10];
 620
 621         /*
 622          * keep only block devices, discard partitions
 623          */
 624         if (sscanf(devpath, "/block/%10s", a) != 1 ||
 625             sscanf(devpath, "/block/%10[^/]/%10s", a, b) == 2) {
 626                 condlog(4, "discard event on %s", devpath);
 627                 return 1;
 628         }
 629         return 0;
 630 }
 631
 632 int
 633 uev_trigger (struct uevent * uev, void * trigger_data)
 634 {
 635         int r = 0;
 636         char devname[32];
 637         struct vectors * vecs;
 638
 639         vecs = (struct vectors *)trigger_data;
 640
 641         if (uev_discard(uev->devpath))
 642                 return 0;
 643
 644         basename(uev->devpath, devname);
 645         lock(vecs->lock);
 646
 647         /*
 648          * device map add/remove event
 649          */
 650         if (!strncmp(devname, "dm-", 3)) {
 651                 if (!strncmp(uev->action, "add", 3)) {
 652                         r = uev_add_map(devname, vecs);
 653                         goto out;
 654                 }
 655                 if (!strncmp(uev->action, "remove", 6)) {
 656                         r = uev_remove_map(devname, vecs);
 657                         goto out;
 658                 }
 659                 if (!strncmp(uev->action, "umount", 6)) {
 660                         r = uev_umount_map(devname, vecs);
 661                         goto out;
 662                 }
 663                 goto out;
 664         }
 665
 666         /*
 667          * path add/remove event
 668          */
 669         if (blacklist(conf->blist_devnode, devname))
 670                 goto out;
 671
 672         if (!strncmp(uev->action, "add", 3)) {
 673                 r = uev_add_path(devname, vecs);
 674                 goto out;
 675         }
 676         if (!strncmp(uev->action, "remove", 6)) {
 677                 r = uev_remove_path(devname, vecs);
 678                 goto out;
 679         }
 680
 681 out:
 682         unlock(vecs->lock);
 683         return r;
 684 }
 685
 686 static void *
 687 ueventloop (void * ap)
 688 {
 689         if (uevent_listen(&uev_trigger, ap))
 690                 fprintf(stderr, "error starting uevent listener");
 691
 692         return NULL;
 693 }
 694
 695 static void *
 696 uxlsnrloop (void * ap)
 697 {
 698         if (load_keys())
 699                 return NULL;
 700
 701         if (alloc_handlers())
 702                 return NULL;
 703
 704         add_handler(LIST+PATHS, cli_list_paths);
 705         add_handler(LIST+MAPS, cli_list_maps);
 706         add_handler(LIST+MAPS+STATUS, cli_list_maps_status);
 707         add_handler(LIST+MAPS+STATS, cli_list_maps_stats);
 708         add_handler(LIST+MAPS+TOPOLOGY, cli_list_maps_topology);
 709         add_handler(LIST+TOPOLOGY, cli_list_maps_topology);
 710         add_handler(LIST+MAP+TOPOLOGY, cli_list_map_topology);
 711         add_handler(LIST+CONFIG, cli_list_config);
 712         add_handler(ADD+PATH, cli_add_path);
 713         add_handler(DEL+PATH, cli_del_path);
 714         add_handler(ADD+MAP, cli_add_map);
 715         add_handler(DEL+MAP, cli_del_map);
 716         add_handler(SWITCH+MAP+GROUP, cli_switch_group);
 717         add_handler(RECONFIGURE, cli_reconfigure);
 718         add_handler(SUSPEND+MAP, cli_suspend);
 719         add_handler(RESUME+MAP, cli_resume);
 720         add_handler(REINSTATE+PATH, cli_reinstate);
 721         add_handler(FAIL+PATH, cli_fail);
 722
 723         uxsock_listen(&uxsock_trigger, ap);
 724
 725         return NULL;
 726 }
 727
 728 static int
 729 exit_daemon (int status)
 730 {
 731         if (status != 0)
 732                 fprintf(stderr, "bad exit status. see daemon.log\n");
 733
 734         condlog(3, "unlink pidfile");
 735         unlink(DEFAULT_PIDFILE);
 736
 737         lock(&exit_mutex);
 738         pthread_cond_signal(&exit_cond);
 739         unlock(&exit_mutex);
 740
 741         return status;
 742 }
 743
 744 static void
 745 fail_path (struct path * pp, int del_active)
 746 {
 747         if (!pp->mpp)
 748                 return;
 749
 750         condlog(2, "checker failed path %s in map %s",
 751                  pp->dev_t, pp->mpp->alias);
 752
 753         dm_fail_path(pp->mpp->alias, pp->dev_t);
 754         if (del_active)
 755                 update_queue_mode_del_path(pp->mpp);
 756 }
 757
 758 /*
 759  * caller must have locked the path list before calling that function
 760  */
 761 static void
 762 reinstate_path (struct path * pp, int add_active)
 763 {
 764         if (!pp->mpp)
 765                 return;
 766
 767         if (dm_reinstate_path(pp->mpp->alias, pp->dev_t))
 768                 condlog(0, "%s: reinstate failed", pp->dev_t);
 769         else {
 770                 condlog(2, "%s: reinstated", pp->dev_t);
 771                 if (add_active)
 772                         update_queue_mode_add_path(pp->mpp);
 773         }
 774 }
 775
 776 static void
 777 enable_group(struct path * pp)
 778 {
 779         struct pathgroup * pgp;
 780
 781         /*
 782          * if path is added through uev_add_path, pgindex can be unset.
 783          * next update_strings() will set it, upon map reload event.
 784          *
 785          * we can safely return here, because upon map reload, all
 786          * PG will be enabled.
 787          */
 788         if (!pp->mpp->pg || !pp->pgindex)
 789                 return;
 790
 791         pgp = VECTOR_SLOT(pp->mpp->pg, pp->pgindex - 1);
 792
 793         if (pgp->status == PGSTATE_DISABLED) {
 794                 condlog(2, "%s: enable group #%i", pp->mpp->alias, pp->pgindex);
 795                 dm_enablegroup(pp->mpp->alias, pp->pgindex);
 796         }
 797 }
 798
 799 static void
 800 mpvec_garbage_collector (struct vectors * vecs)
 801 {
 802         struct multipath * mpp;
 803         int i;
 804
 805         vector_foreach_slot (vecs->mpvec, mpp, i) {
 806                 if (mpp && mpp->alias && !dm_map_present(mpp->alias)) {
 807                         condlog(2, "%s: remove dead map", mpp->alias);
 808                         remove_map(mpp, vecs, stop_waiter_thread, 1);
 809                         i--;
 810                 }
 811         }
 812 }
 813
 814 static void
 815 defered_failback_tick (vector mpvec)
 816 {
 817         struct multipath * mpp;
 818         int i;
 819
 820         vector_foreach_slot (mpvec, mpp, i) {
 821                 /*
 822                  * defered failback getting sooner
 823                  */
 824                 if (mpp->pgfailback > 0 && mpp->failback_tick > 0) {
 825                         mpp->failback_tick--;
 826
 827                         if (!mpp->failback_tick && need_switch_pathgroup(mpp, 1))
 828                                 switch_pathgroup(mpp);
 829                 }
 830         }
 831 }
 832
 833 static void
 834 retry_count_tick(vector mpvec)
 835 {
 836         struct multipath *mpp;
 837         int i;
 838
 839         vector_foreach_slot (mpvec, mpp, i) {
 840                 if (mpp->retry_tick) {
 841                         mpp->stat_total_queueing_time++;
 842                         condlog(4, "%s: Retrying.. No active path", mpp->alias);
 843                         if(--mpp->retry_tick == 0) {
 844                                 dm_queue_if_no_path(mpp->alias, 0);
 845                                 condlog(2, "%s: Disable queueing", mpp->alias);
 846                         }
 847                 }
 848         }
 849 }
 850
 851 static void *
 852 checkerloop (void *ap)
 853 {
 854         struct vectors *vecs;
 855         struct path *pp;
 856         int i, count = 0;
 857         int newstate;
 858
 859         mlockall(MCL_CURRENT | MCL_FUTURE);
 860         vecs = (struct vectors *)ap;
 861         condlog(2, "path checkers start up");
 862
 863         /*
 864          * init the path check interval
 865          */
 866         vector_foreach_slot (vecs->pathvec, pp, i) {
 867                 pp->checkint = conf->checkint;
 868         }
 869
 870         while (1) {
 871                 pthread_cleanup_push(cleanup_lock, vecs->lock);
 872                 lock(vecs->lock);
 873                 condlog(4, "tick");
 874
 875                 vector_foreach_slot (vecs->pathvec, pp, i) {
 876                         if (!pp->mpp)
 877                                 continue;
 878
 879                         if (pp->tick && --pp->tick)
 880                                 continue; /* don't check this path yet */
 881
 882                         /*
 883                          * provision a next check soonest,
 884                          * in case we exit abnormaly from here
 885                          */
 886                         pp->tick = conf->checkint;
 887
 888                         if (!checker_selected(&pp->checker)) {
 889                                 pathinfo(pp, conf->hwtable, DI_SYSFS);
 890                                 select_checker(pp);
 891                         }
 892
 893                         if (!checker_selected(&pp->checker)) {
 894                                 condlog(0, "%s: checker is not set", pp->dev);
 895                                 continue;
 896                         }
 897                         newstate = checker_check(&pp->checker);
 898
 899                         if (newstate < 0) {
 900                                 condlog(2, "%s: unusable path", pp->dev);
 901                                 pathinfo(pp, conf->hwtable, 0);
 902                                 continue;
 903                         }
 904
 905                         if (newstate != pp->state) {
 906                                 int oldstate = pp->state;
 907                                 pp->state = newstate;
 908                                 LOG_MSG(1, checker_message(&pp->checker));
 909
 910                                 /*
 911                                  * upon state change, reset the checkint
 912                                  * to the shortest delay
 913                                  */
 914                                 pp->checkint = conf->checkint;
 915
 916                                 if (newstate == PATH_DOWN ||
 917                                     newstate == PATH_SHAKY ||
 918                                     update_multipath_strings(pp->mpp,
 919                                                              vecs->pathvec)) {
 920                                         /*
 921                                          * proactively fail path in the DM
 922                                          */
 923                                         if (oldstate == PATH_UP ||
 924                                             oldstate == PATH_GHOST)
 925                                                 fail_path(pp, 1);
 926                                         else
 927                                                 fail_path(pp, 0);
 928
 929                                         /*
 930                                          * cancel scheduled failback
 931                                          */
 932                                         pp->mpp->failback_tick = 0;
 933
 934                                         pp->mpp->stat_path_failures++;
 935                                         continue;
 936                                 }
 937
 938                                 /*
 939                                  * reinstate this path
 940                                  */
 941                                 if (oldstate != PATH_UP &&
 942                                     oldstate != PATH_GHOST)
 943                                         reinstate_path(pp, 1);
 944                                 else
 945                                         reinstate_path(pp, 0);
 946
 947                                 /*
 948                                  * schedule [defered] failback
 949                                  */
 950                                 if (pp->mpp->pgfailback > 0)
 951                                         pp->mpp->failback_tick =
 952                                                 pp->mpp->pgfailback + 1;
 953                                 else if (pp->mpp->pgfailback == -FAILBACK_IMMEDIATE &&
 954                                     need_switch_pathgroup(pp->mpp, 1))
 955                                         switch_pathgroup(pp->mpp);
 956
 957                                 /*
 958                                  * if at least one path is up in a group, and
 959                                  * the group is disabled, re-enable it
 960                                  */
 961                                 if (newstate == PATH_UP)
 962                                         enable_group(pp);
 963                         }
 964                         else if (newstate == PATH_UP || newstate == PATH_GHOST) {
 965                                 LOG_MSG(4, checker_message(&pp->checker));
 966                                 /*
 967                                  * double the next check delay.
 968                                  * max at conf->max_checkint
 969                                  */
 970                                 if (pp->checkint < (conf->max_checkint / 2))
 971                                         pp->checkint = 2 * pp->checkint;
 972                                 else
 973                                         pp->checkint = conf->max_checkint;
 974
 975                                 pp->tick = pp->checkint;
 976                                 condlog(4, "%s: delay next check %is",
 977                                                 pp->dev_t, pp->tick);
 978                         }
 979                         else if (newstate == PATH_DOWN)
 980                                 LOG_MSG(2, checker_message(&pp->checker));
 981
 982                         pp->state = newstate;
 983
 984                         /*
 985                          * path prio refreshing
 986                          */
 987                         condlog(4, "path prio refresh");
 988                         pathinfo(pp, conf->hwtable, DI_PRIO);
 989
 990                         if (need_switch_pathgroup(pp->mpp, 0)) {
 991                                 if (pp->mpp->pgfailback > 0 &&
 992                                     pp->mpp->failback_tick <= 0)
 993                                         pp->mpp->failback_tick =
 994                                                 pp->mpp->pgfailback + 1;
 995                                 else if (pp->mpp->pgfailback ==
 996                                                 -FAILBACK_IMMEDIATE)
 997                                         switch_pathgroup(pp->mpp);
 998                         }
 999                 }
1000                 defered_failback_tick(vecs->mpvec);
1001                 retry_count_tick(vecs->mpvec);
1002
1003                 if (count)
1004                         count--;
1005                 else {
1006                         condlog(4, "map garbage collection");
1007                         mpvec_garbage_collector(vecs);
1008                         count = MAPGCINT;
1009                 }
1010
1011                 lock_cleanup_pop(vecs->lock);
1012                 sleep(1);
1013         }
1014         return NULL;
1015 }
1016
1017 int
1018 configure (struct vectors * vecs, int start_waiters)
1019 {
1020         struct multipath * mpp;
1021         struct path * pp;
1022         vector mpvec;
1023         int i;
1024
1025         if (!vecs->pathvec && !(vecs->pathvec = vector_alloc()))
1026                 return 1;
1027
1028         if (!vecs->mpvec && !(vecs->mpvec = vector_alloc()))
1029                 return 1;
1030
1031         if (!(mpvec = vector_alloc()))
1032                 return 1;
1033
1034         /*
1035          * probe for current path (from sysfs) and map (from dm) sets
1036          */
1037         path_discovery(vecs->pathvec, conf, DI_ALL);
1038
1039         vector_foreach_slot (vecs->pathvec, pp, i){
1040                 if (blacklist_path(conf, pp)){
1041                         vector_del_slot(vecs->pathvec, i);
1042                         free_path(pp);
1043                         i--;
1044                 }
1045                 else
1046                         pp->checkint = conf->checkint;
1047         }
1048         if (map_discovery(vecs))
1049                 return 1;
1050
1051         /*
1052          * create new set of maps & push changed ones into dm
1053          */
1054         if (coalesce_paths(vecs, mpvec, NULL))
1055                 return 1;
1056
1057         /*
1058          * may need to remove some maps which are no longer relevant
1059          * e.g., due to blacklist changes in conf file
1060          */
1061         if (coalesce_maps(vecs, mpvec))
1062                 return 1;
1063
1064         dm_lib_release();
1065
1066         sync_maps_state(mpvec);
1067
1068         if (conf->verbosity > 2)
1069                 vector_foreach_slot(mpvec, mpp, i)
1070                         print_map(mpp);
1071
1072         /*
1073          * purge dm of old maps
1074          */
1075         remove_maps(vecs, NULL);
1076
1077         /*
1078          * save new set of maps formed by considering current path state
1079          */
1080         vector_free(vecs->mpvec);
1081         vecs->mpvec = mpvec;
1082
1083         /*
1084          * start dm event waiter threads for these new maps
1085          */
1086         vector_foreach_slot(vecs->mpvec, mpp, i) {
1087                 if (setup_multipath(vecs, mpp))
1088                         return 1;
1089                 if (start_waiters)
1090                         if (start_waiter_thread(mpp, vecs))
1091                                 return 1;
1092         }
1093         return 0;
1094 }
1095
1096 int
1097 reconfigure (struct vectors * vecs)
1098 {
1099         struct config * old = conf;
1100
1101         /*
1102          * free old map and path vectors ... they use old conf state
1103          */
1104         if (VECTOR_SIZE(vecs->mpvec))
1105                 remove_maps(vecs, stop_waiter_thread);
1106
1107         if (VECTOR_SIZE(vecs->pathvec))
1108                 free_pathvec(vecs->pathvec, FREE_PATHS);
1109
1110         vecs->pathvec = NULL;
1111         conf = NULL;
1112
1113         if (load_config(DEFAULT_CONFIGFILE))
1114                 return 1;
1115
1116         conf->verbosity = old->verbosity;
1117
1118         if (!conf->checkint) {
1119                 conf->checkint = DEFAULT_CHECKINT;
1120                 conf->max_checkint = MAX_CHECKINT(conf->checkint);
1121         }
1122         configure(vecs, 1);
1123         free_config(old);
1124         return 0;
1125 }
1126
1127 static struct vectors *
1128 init_vecs (void)
1129 {
1130         struct vectors * vecs;
1131
1132         vecs = (struct vectors *)MALLOC(sizeof(struct vectors));
1133
1134         if (!vecs)
1135                 return NULL;
1136
1137         vecs->lock =
1138                 (pthread_mutex_t *)MALLOC(sizeof(pthread_mutex_t));
1139
1140         if (!vecs->lock)
1141                 goto out;
1142
1143         pthread_mutex_init(vecs->lock, NULL);
1144
1145         return vecs;
1146
1147 out:
1148         FREE(vecs);
1149         condlog(0, "failed to init paths");
1150         return NULL;
1151 }
1152
1153 static void *
1154 signal_set(int signo, void (*func) (int))
1155 {
1156         int r;
1157         struct sigaction sig;
1158         struct sigaction osig;
1159
1160         sig.sa_handler = func;
1161         sigemptyset(&sig.sa_mask);
1162         sig.sa_flags = 0;
1163
1164         r = sigaction(signo, &sig, &osig);
1165
1166         if (r < 0)
1167                 return (SIG_ERR);
1168         else
1169                 return (osig.sa_handler);
1170 }
1171
1172 static void
1173 sighup (int sig)
1174 {
1175         condlog(2, "reconfigure (SIGHUP)");
1176
1177         lock(gvecs->lock);
1178         reconfigure(gvecs);
1179         unlock(gvecs->lock);
1180
1181 #ifdef _DEBUG_
1182         dbg_free_final(NULL);
1183 #endif
1184 }
1185
1186 static void
1187 sigend (int sig)
1188 {
1189         exit_daemon(0);
1190 }
1191
1192 static void
1193 sigusr1 (int sig)
1194 {
1195         condlog(3, "SIGUSR1 received");
1196 }
1197
1198 static void
1199 signal_init(void)
1200 {
1201         signal_set(SIGHUP, sighup);
1202         signal_set(SIGUSR1, sigusr1);
1203         signal_set(SIGINT, sigend);
1204         signal_set(SIGTERM, sigend);
1205         signal(SIGPIPE, SIG_IGN);
1206 }
1207
1208 static void
1209 setscheduler (void)
1210 {
1211         int res;
1212         static struct sched_param sched_param = {
1213                 sched_priority: 99
1214         };
1215
1216         res = sched_setscheduler (0, SCHED_RR, &sched_param);
1217
1218         if (res == -1)
1219                 condlog(LOG_WARNING, "Could not set SCHED_RR at priority 99");
1220         return;
1221 }
1222
1223 static void
1224 set_oom_adj (int val)
1225 {
1226         FILE *fp;
1227
1228         fp = fopen("/proc/self/oom_adj", "w");
1229
1230         if (!fp)
1231                 return;
1232
1233         fprintf(fp, "%i", val);
1234         fclose(fp);
1235 }
1236
1237 static int
1238 child (void * param)
1239 {
1240         pthread_t check_thr, uevent_thr, uxlsnr_thr;
1241         pthread_attr_t attr;
1242         struct vectors * vecs;
1243
1244         mlockall(MCL_CURRENT | MCL_FUTURE);
1245
1246         if (logsink)
1247                 log_thread_start();
1248
1249         condlog(2, "--------start up--------");
1250         condlog(2, "read " DEFAULT_CONFIGFILE);
1251
1252         if (load_config(DEFAULT_CONFIGFILE))
1253                 exit(1);
1254
1255         setlogmask(LOG_UPTO(conf->verbosity + 3));
1256
1257         /*
1258          * fill the voids left in the config file
1259          */
1260         if (!conf->checkint) {
1261                 conf->checkint = DEFAULT_CHECKINT;
1262                 conf->max_checkint = MAX_CHECKINT(conf->checkint);
1263         }
1264
1265         if (pidfile_create(DEFAULT_PIDFILE, getpid())) {
1266                 if (logsink)
1267                         log_thread_stop();
1268
1269                 exit(1);
1270         }
1271         signal_init();
1272         setscheduler();
1273         set_oom_adj(-16);
1274         vecs = gvecs = init_vecs();
1275
1276         if (!vecs)
1277                 exit(1);
1278
1279         if (sysfs_get_mnt_path(sysfs_path, FILE_NAME_SIZE)) {
1280                 condlog(0, "can not find sysfs mount point");
1281                 exit(1);
1282         }
1283
1284         /*
1285          * fetch and configure both paths and multipaths
1286          */
1287         if (configure(vecs, 1)) {
1288                 condlog(0, "failure during configuration");
1289                 exit(1);
1290         }
1291
1292         /*
1293          * start threads
1294          */
1295         pthread_attr_init(&attr);
1296         pthread_attr_setstacksize(&attr, 64 * 1024);
1297         pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
1298
1299         pthread_create(&check_thr, &attr, checkerloop, vecs);
1300         pthread_create(&uevent_thr, &attr, ueventloop, vecs);
1301         pthread_create(&uxlsnr_thr, &attr, uxlsnrloop, vecs);
1302
1303         pthread_cond_wait(&exit_cond, &exit_mutex);
1304
1305         /*
1306          * exit path
1307          */
1308         lock(vecs->lock);
1309         remove_maps(vecs, stop_waiter_thread);
1310         free_pathvec(vecs->pathvec, FREE_PATHS);
1311
1312         pthread_cancel(check_thr);
1313         pthread_cancel(uevent_thr);
1314         pthread_cancel(uxlsnr_thr);
1315
1316         free_keys(keys);
1317         keys = NULL;
1318         free_handlers(handlers);
1319         handlers = NULL;
1320         free_polls();
1321
1322         unlock(vecs->lock);
1323         pthread_mutex_destroy(vecs->lock);
1324         FREE(vecs->lock);
1325         vecs->lock = NULL;
1326         FREE(vecs);
1327         vecs = NULL;
1328         free_config(conf);
1329         conf = NULL;
1330
1331         condlog(2, "--------shut down-------");
1332
1333         if (logsink)
1334                 log_thread_stop();
1335
1336         dm_lib_release();
1337         dm_lib_exit();
1338
1339 #ifdef _DEBUG_
1340         dbg_free_final(NULL);
1341 #endif
1342
1343         exit(0);
1344 }
1345
1346 static int
1347 daemonize(void)
1348 {
1349         int pid;
1350         int in_fd, out_fd;
1351
1352         if( (pid = fork()) < 0){
1353                 fprintf(stderr, "Failed first fork : %s\n", strerror(errno));
1354                 return -1;
1355         }
1356         else if (pid != 0)
1357                 return pid;
1358
1359         setsid();
1360
1361         if ( (pid = fork()) < 0)
1362                 fprintf(stderr, "Failed second fork : %s\n", strerror(errno));
1363         else if (pid != 0)
1364                 _exit(0);
1365
1366         in_fd = open("/dev/null", O_RDONLY);
1367         if (in_fd < 0){
1368                 fprintf(stderr, "cannot open /dev/null for input : %s\n",
1369                         strerror(errno));
1370                 _exit(0);
1371         }
1372         out_fd = open("/dev/console", O_WRONLY);
1373         if (out_fd < 0){
1374                 fprintf(stderr, "cannot open /dev/console for output : %s\n",
1375                         strerror(errno));
1376                 _exit(0);
1377         }
1378
1379         close(STDIN_FILENO);
1380         dup(in_fd);
1381         close(STDOUT_FILENO);
1382         dup(out_fd);
1383         close(STDERR_FILENO);
1384         dup(out_fd);
1385
1386         close(in_fd);
1387         close(out_fd);
1388         chdir("/");
1389         umask(0);
1390         return 0;
1391 }
1392
1393 int
1394 main (int argc, char *argv[])
1395 {
1396         extern char *optarg;
1397         extern int optind;
1398         int arg;
1399         int err;
1400
1401         logsink = 1;
1402
1403         if (getuid() != 0) {
1404                 fprintf(stderr, "need to be root\n");
1405                 exit(1);
1406         }
1407
1408         /* make sure we don't lock any path */
1409         chdir("/");
1410         umask(umask(077) | 022);
1411
1412         conf = alloc_config();
1413
1414         if (!conf)
1415                 exit(1);
1416
1417         while ((arg = getopt(argc, argv, ":dv:k::")) != EOF ) {
1418         switch(arg) {
1419                 case 'd':
1420                         logsink = 0;
1421                         //debug=1; /* ### comment me out ### */
1422                         break;
1423                 case 'v':
1424                         if (sizeof(optarg) > sizeof(char *) ||
1425                             !isdigit(optarg[0]))
1426                                 exit(1);
1427
1428                         conf->verbosity = atoi(optarg);
1429                         break;
1430                 case 'k':
1431                         uxclnt(optarg);
1432                         exit(0);
1433                 default:
1434                         ;
1435                 }
1436         }
1437
1438         if (!logsink)
1439                 err = 0;
1440         else
1441                 err = daemonize();
1442
1443         if (err < 0)
1444                 /* error */
1445                 exit(1);
1446         else if (err > 0)
1447                 /* parent dies */
1448                 exit(0);
1449         else
1450                 /* child lives */
1451                 return (child(NULL));
1452 }
1453