1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 
  22 /*
  23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
  25  * Copyright (c) 2013 by Delphix. All rights reserved.
  26  */
  27 
  28 #include <sys/zfs_context.h>
  29 #include <sys/fm/fs/zfs.h>
  30 #include <sys/spa.h>
  31 #include <sys/spa_impl.h>
  32 #include <sys/dmu.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/vdev_impl.h>
  35 #include <sys/uberblock_impl.h>
  36 #include <sys/metaslab.h>
  37 #include <sys/metaslab_impl.h>
  38 #include <sys/space_map.h>
  39 #include <sys/zio.h>
  40 #include <sys/zap.h>
  41 #include <sys/fs/zfs.h>
  42 #include <sys/arc.h>
  43 #include <sys/zil.h>
  44 #include <sys/dsl_scan.h>
  45 
  46 /*
  47  * Virtual device management.
  48  */
  49 
  50 static vdev_ops_t *vdev_ops_table[] = {
  51         &vdev_root_ops,
  52         &vdev_raidz_ops,
  53         &vdev_mirror_ops,
  54         &vdev_replacing_ops,
  55         &vdev_spare_ops,
  56         &vdev_disk_ops,
  57         &vdev_file_ops,
  58         &vdev_missing_ops,
  59         &vdev_hole_ops,
  60         NULL
  61 };
  62 
  63 /* maximum scrub/resilver I/O queue per leaf vdev */
  64 int zfs_scrub_limit = 10;
  65 
  66 /*
  67  * Given a vdev type, return the appropriate ops vector.
  68  */
  69 static vdev_ops_t *
  70 vdev_getops(const char *type)
  71 {
  72         vdev_ops_t *ops, **opspp;
  73 
  74         for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
  75                 if (strcmp(ops->vdev_op_type, type) == 0)
  76                         break;
  77 
  78         return (ops);
  79 }
  80 
  81 /*
  82  * Default asize function: return the MAX of psize with the asize of
  83  * all children.  This is what's used by anything other than RAID-Z.
  84  */
  85 uint64_t
  86 vdev_default_asize(vdev_t *vd, uint64_t psize)
  87 {
  88         uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
  89         uint64_t csize;
  90 
  91         for (int c = 0; c < vd->vdev_children; c++) {
  92                 csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
  93                 asize = MAX(asize, csize);
  94         }
  95 
  96         return (asize);
  97 }
  98 
  99 /*
 100  * Get the minimum allocatable size. We define the allocatable size as
 101  * the vdev's asize rounded to the nearest metaslab. This allows us to
 102  * replace or attach devices which don't have the same physical size but
 103  * can still satisfy the same number of allocations.
 104  */
 105 uint64_t
 106 vdev_get_min_asize(vdev_t *vd)
 107 {
 108         vdev_t *pvd = vd->vdev_parent;
 109 
 110         /*
 111          * If our parent is NULL (inactive spare or cache) or is the root,
 112          * just return our own asize.
 113          */
 114         if (pvd == NULL)
 115                 return (vd->vdev_asize);
 116 
 117         /*
 118          * The top-level vdev just returns the allocatable size rounded
 119          * to the nearest metaslab.
 120          */
 121         if (vd == vd->vdev_top)
 122                 return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
 123 
 124         /*
 125          * The allocatable space for a raidz vdev is N * sizeof(smallest child),
 126          * so each child must provide at least 1/Nth of its asize.
 127          */
 128         if (pvd->vdev_ops == &vdev_raidz_ops)
 129                 return (pvd->vdev_min_asize / pvd->vdev_children);
 130 
 131         return (pvd->vdev_min_asize);
 132 }
 133 
 134 void
 135 vdev_set_min_asize(vdev_t *vd)
 136 {
 137         vd->vdev_min_asize = vdev_get_min_asize(vd);
 138 
 139         for (int c = 0; c < vd->vdev_children; c++)
 140                 vdev_set_min_asize(vd->vdev_child[c]);
 141 }
 142 
 143 vdev_t *
 144 vdev_lookup_top(spa_t *spa, uint64_t vdev)
 145 {
 146         vdev_t *rvd = spa->spa_root_vdev;
 147 
 148         ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
 149 
 150         if (vdev < rvd->vdev_children) {
 151                 ASSERT(rvd->vdev_child[vdev] != NULL);
 152                 return (rvd->vdev_child[vdev]);
 153         }
 154 
 155         return (NULL);
 156 }
 157 
 158 vdev_t *
 159 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
 160 {
 161         vdev_t *mvd;
 162 
 163         if (vd->vdev_guid == guid)
 164                 return (vd);
 165 
 166         for (int c = 0; c < vd->vdev_children; c++)
 167                 if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
 168                     NULL)
 169                         return (mvd);
 170 
 171         return (NULL);
 172 }
 173 
 174 void
 175 vdev_add_child(vdev_t *pvd, vdev_t *cvd)
 176 {
 177         size_t oldsize, newsize;
 178         uint64_t id = cvd->vdev_id;
 179         vdev_t **newchild;
 180 
 181         ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
 182         ASSERT(cvd->vdev_parent == NULL);
 183 
 184         cvd->vdev_parent = pvd;
 185 
 186         if (pvd == NULL)
 187                 return;
 188 
 189         ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
 190 
 191         oldsize = pvd->vdev_children * sizeof (vdev_t *);
 192         pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
 193         newsize = pvd->vdev_children * sizeof (vdev_t *);
 194 
 195         newchild = kmem_zalloc(newsize, KM_SLEEP);
 196         if (pvd->vdev_child != NULL) {
 197                 bcopy(pvd->vdev_child, newchild, oldsize);
 198                 kmem_free(pvd->vdev_child, oldsize);
 199         }
 200 
 201         pvd->vdev_child = newchild;
 202         pvd->vdev_child[id] = cvd;
 203 
 204         cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
 205         ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
 206 
 207         /*
 208          * Walk up all ancestors to update guid sum.
 209          */
 210         for (; pvd != NULL; pvd = pvd->vdev_parent)
 211                 pvd->vdev_guid_sum += cvd->vdev_guid_sum;
 212 }
 213 
 214 void
 215 vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
 216 {
 217         int c;
 218         uint_t id = cvd->vdev_id;
 219 
 220         ASSERT(cvd->vdev_parent == pvd);
 221 
 222         if (pvd == NULL)
 223                 return;
 224 
 225         ASSERT(id < pvd->vdev_children);
 226         ASSERT(pvd->vdev_child[id] == cvd);
 227 
 228         pvd->vdev_child[id] = NULL;
 229         cvd->vdev_parent = NULL;
 230 
 231         for (c = 0; c < pvd->vdev_children; c++)
 232                 if (pvd->vdev_child[c])
 233                         break;
 234 
 235         if (c == pvd->vdev_children) {
 236                 kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
 237                 pvd->vdev_child = NULL;
 238                 pvd->vdev_children = 0;
 239         }
 240 
 241         /*
 242          * Walk up all ancestors to update guid sum.
 243          */
 244         for (; pvd != NULL; pvd = pvd->vdev_parent)
 245                 pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
 246 }
 247 
 248 /*
 249  * Remove any holes in the child array.
 250  */
 251 void
 252 vdev_compact_children(vdev_t *pvd)
 253 {
 254         vdev_t **newchild, *cvd;
 255         int oldc = pvd->vdev_children;
 256         int newc;
 257 
 258         ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
 259 
 260         for (int c = newc = 0; c < oldc; c++)
 261                 if (pvd->vdev_child[c])
 262                         newc++;
 263 
 264         newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
 265 
 266         for (int c = newc = 0; c < oldc; c++) {
 267                 if ((cvd = pvd->vdev_child[c]) != NULL) {
 268                         newchild[newc] = cvd;
 269                         cvd->vdev_id = newc++;
 270                 }
 271         }
 272 
 273         kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
 274         pvd->vdev_child = newchild;
 275         pvd->vdev_children = newc;
 276 }
 277 
 278 /*
 279  * Allocate and minimally initialize a vdev_t.
 280  */
 281 vdev_t *
 282 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
 283 {
 284         vdev_t *vd;
 285 
 286         vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
 287 
 288         if (spa->spa_root_vdev == NULL) {
 289                 ASSERT(ops == &vdev_root_ops);
 290                 spa->spa_root_vdev = vd;
 291                 spa->spa_load_guid = spa_generate_guid(NULL);
 292         }
 293 
 294         if (guid == 0 && ops != &vdev_hole_ops) {
 295                 if (spa->spa_root_vdev == vd) {
 296                         /*
 297                          * The root vdev's guid will also be the pool guid,
 298                          * which must be unique among all pools.
 299                          */
 300                         guid = spa_generate_guid(NULL);
 301                 } else {
 302                         /*
 303                          * Any other vdev's guid must be unique within the pool.
 304                          */
 305                         guid = spa_generate_guid(spa);
 306                 }
 307                 ASSERT(!spa_guid_exists(spa_guid(spa), guid));
 308         }
 309 
 310         vd->vdev_spa = spa;
 311         vd->vdev_id = id;
 312         vd->vdev_guid = guid;
 313         vd->vdev_guid_sum = guid;
 314         vd->vdev_ops = ops;
 315         vd->vdev_state = VDEV_STATE_CLOSED;
 316         vd->vdev_ishole = (ops == &vdev_hole_ops);
 317 
 318         mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
 319         mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
 320         mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
 321         for (int t = 0; t < DTL_TYPES; t++) {
 322                 space_map_create(&vd->vdev_dtl[t], 0, -1ULL, 0,
 323                     &vd->vdev_dtl_lock);
 324         }
 325         txg_list_create(&vd->vdev_ms_list,
 326             offsetof(struct metaslab, ms_txg_node));
 327         txg_list_create(&vd->vdev_dtl_list,
 328             offsetof(struct vdev, vdev_dtl_node));
 329         vd->vdev_stat.vs_timestamp = gethrtime();
 330         vdev_queue_init(vd);
 331         vdev_cache_init(vd);
 332 
 333         return (vd);
 334 }
 335 
 336 /*
 337  * Allocate a new vdev.  The 'alloctype' is used to control whether we are
 338  * creating a new vdev or loading an existing one - the behavior is slightly
 339  * different for each case.
 340  */
 341 int
 342 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
 343     int alloctype)
 344 {
 345         vdev_ops_t *ops;
 346         char *type;
 347         uint64_t guid = 0, islog, nparity;
 348         vdev_t *vd;
 349 
 350         ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
 351 
 352         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
 353                 return (SET_ERROR(EINVAL));
 354 
 355         if ((ops = vdev_getops(type)) == NULL)
 356                 return (SET_ERROR(EINVAL));
 357 
 358         /*
 359          * If this is a load, get the vdev guid from the nvlist.
 360          * Otherwise, vdev_alloc_common() will generate one for us.
 361          */
 362         if (alloctype == VDEV_ALLOC_LOAD) {
 363                 uint64_t label_id;
 364 
 365                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
 366                     label_id != id)
 367                         return (SET_ERROR(EINVAL));
 368 
 369                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
 370                         return (SET_ERROR(EINVAL));
 371         } else if (alloctype == VDEV_ALLOC_SPARE) {
 372                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
 373                         return (SET_ERROR(EINVAL));
 374         } else if (alloctype == VDEV_ALLOC_L2CACHE) {
 375                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
 376                         return (SET_ERROR(EINVAL));
 377         } else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
 378                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
 379                         return (SET_ERROR(EINVAL));
 380         }
 381 
 382         /*
 383          * The first allocated vdev must be of type 'root'.
 384          */
 385         if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
 386                 return (SET_ERROR(EINVAL));
 387 
 388         /*
 389          * Determine whether we're a log vdev.
 390          */
 391         islog = 0;
 392         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
 393         if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
 394                 return (SET_ERROR(ENOTSUP));
 395 
 396         if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
 397                 return (SET_ERROR(ENOTSUP));
 398 
 399         /*
 400          * Set the nparity property for RAID-Z vdevs.
 401          */
 402         nparity = -1ULL;
 403         if (ops == &vdev_raidz_ops) {
 404                 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
 405                     &nparity) == 0) {
 406                         if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
 407                                 return (SET_ERROR(EINVAL));
 408                         /*
 409                          * Previous versions could only support 1 or 2 parity
 410                          * device.
 411                          */
 412                         if (nparity > 1 &&
 413                             spa_version(spa) < SPA_VERSION_RAIDZ2)
 414                                 return (SET_ERROR(ENOTSUP));
 415                         if (nparity > 2 &&
 416                             spa_version(spa) < SPA_VERSION_RAIDZ3)
 417                                 return (SET_ERROR(ENOTSUP));
 418                 } else {
 419                         /*
 420                          * We require the parity to be specified for SPAs that
 421                          * support multiple parity levels.
 422                          */
 423                         if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
 424                                 return (SET_ERROR(EINVAL));
 425                         /*
 426                          * Otherwise, we default to 1 parity device for RAID-Z.
 427                          */
 428                         nparity = 1;
 429                 }
 430         } else {
 431                 nparity = 0;
 432         }
 433         ASSERT(nparity != -1ULL);
 434 
 435         vd = vdev_alloc_common(spa, id, guid, ops);
 436 
 437         vd->vdev_islog = islog;
 438         vd->vdev_nparity = nparity;
 439 
 440         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
 441                 vd->vdev_path = spa_strdup(vd->vdev_path);
 442         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
 443                 vd->vdev_devid = spa_strdup(vd->vdev_devid);
 444         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
 445             &vd->vdev_physpath) == 0)
 446                 vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
 447         if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
 448                 vd->vdev_fru = spa_strdup(vd->vdev_fru);
 449 
 450         /*
 451          * Set the whole_disk property.  If it's not specified, leave the value
 452          * as -1.
 453          */
 454         if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
 455             &vd->vdev_wholedisk) != 0)
 456                 vd->vdev_wholedisk = -1ULL;
 457 
 458         /*
 459          * Look for the 'not present' flag.  This will only be set if the device
 460          * was not present at the time of import.
 461          */
 462         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
 463             &vd->vdev_not_present);
 464 
 465         /*
 466          * Get the alignment requirement.
 467          */
 468         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
 469 
 470         /*
 471          * Retrieve the vdev creation time.
 472          */
 473         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
 474             &vd->vdev_crtxg);
 475 
 476         /*
 477          * If we're a top-level vdev, try to load the allocation parameters.
 478          */
 479         if (parent && !parent->vdev_parent &&
 480             (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
 481                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
 482                     &vd->vdev_ms_array);
 483                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
 484                     &vd->vdev_ms_shift);
 485                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
 486                     &vd->vdev_asize);
 487                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
 488                     &vd->vdev_removing);
 489         }
 490 
 491         if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
 492                 ASSERT(alloctype == VDEV_ALLOC_LOAD ||
 493                     alloctype == VDEV_ALLOC_ADD ||
 494                     alloctype == VDEV_ALLOC_SPLIT ||
 495                     alloctype == VDEV_ALLOC_ROOTPOOL);
 496                 vd->vdev_mg = metaslab_group_create(islog ?
 497                     spa_log_class(spa) : spa_normal_class(spa), vd);
 498         }
 499 
 500         /*
 501          * If we're a leaf vdev, try to load the DTL object and other state.
 502          */
 503         if (vd->vdev_ops->vdev_op_leaf &&
 504             (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
 505             alloctype == VDEV_ALLOC_ROOTPOOL)) {
 506                 if (alloctype == VDEV_ALLOC_LOAD) {
 507                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
 508                             &vd->vdev_dtl_smo.smo_object);
 509                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
 510                             &vd->vdev_unspare);
 511                 }
 512 
 513                 if (alloctype == VDEV_ALLOC_ROOTPOOL) {
 514                         uint64_t spare = 0;
 515 
 516                         if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
 517                             &spare) == 0 && spare)
 518                                 spa_spare_add(vd);
 519                 }
 520 
 521                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
 522                     &vd->vdev_offline);
 523 
 524                 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
 525                     &vd->vdev_resilver_txg);
 526 
 527                 /*
 528                  * When importing a pool, we want to ignore the persistent fault
 529                  * state, as the diagnosis made on another system may not be
 530                  * valid in the current context.  Local vdevs will
 531                  * remain in the faulted state.
 532                  */
 533                 if (spa_load_state(spa) == SPA_LOAD_OPEN) {
 534                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
 535                             &vd->vdev_faulted);
 536                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
 537                             &vd->vdev_degraded);
 538                         (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
 539                             &vd->vdev_removed);
 540 
 541                         if (vd->vdev_faulted || vd->vdev_degraded) {
 542                                 char *aux;
 543 
 544                                 vd->vdev_label_aux =
 545                                     VDEV_AUX_ERR_EXCEEDED;
 546                                 if (nvlist_lookup_string(nv,
 547                                     ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
 548                                     strcmp(aux, "external") == 0)
 549                                         vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
 550                         }
 551                 }
 552         }
 553 
 554         /*
 555          * Add ourselves to the parent's list of children.
 556          */
 557         vdev_add_child(parent, vd);
 558 
 559         *vdp = vd;
 560 
 561         return (0);
 562 }
 563 
 564 void
 565 vdev_free(vdev_t *vd)
 566 {
 567         spa_t *spa = vd->vdev_spa;
 568 
 569         /*
 570          * vdev_free() implies closing the vdev first.  This is simpler than
 571          * trying to ensure complicated semantics for all callers.
 572          */
 573         vdev_close(vd);
 574 
 575         ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
 576         ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
 577 
 578         /*
 579          * Free all children.
 580          */
 581         for (int c = 0; c < vd->vdev_children; c++)
 582                 vdev_free(vd->vdev_child[c]);
 583 
 584         ASSERT(vd->vdev_child == NULL);
 585         ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
 586 
 587         /*
 588          * Discard allocation state.
 589          */
 590         if (vd->vdev_mg != NULL) {
 591                 vdev_metaslab_fini(vd);
 592                 metaslab_group_destroy(vd->vdev_mg);
 593         }
 594 
 595         ASSERT0(vd->vdev_stat.vs_space);
 596         ASSERT0(vd->vdev_stat.vs_dspace);
 597         ASSERT0(vd->vdev_stat.vs_alloc);
 598 
 599         /*
 600          * Remove this vdev from its parent's child list.
 601          */
 602         vdev_remove_child(vd->vdev_parent, vd);
 603 
 604         ASSERT(vd->vdev_parent == NULL);
 605 
 606         /*
 607          * Clean up vdev structure.
 608          */
 609         vdev_queue_fini(vd);
 610         vdev_cache_fini(vd);
 611 
 612         if (vd->vdev_path)
 613                 spa_strfree(vd->vdev_path);
 614         if (vd->vdev_devid)
 615                 spa_strfree(vd->vdev_devid);
 616         if (vd->vdev_physpath)
 617                 spa_strfree(vd->vdev_physpath);
 618         if (vd->vdev_fru)
 619                 spa_strfree(vd->vdev_fru);
 620 
 621         if (vd->vdev_isspare)
 622                 spa_spare_remove(vd);
 623         if (vd->vdev_isl2cache)
 624                 spa_l2cache_remove(vd);
 625 
 626         txg_list_destroy(&vd->vdev_ms_list);
 627         txg_list_destroy(&vd->vdev_dtl_list);
 628 
 629         mutex_enter(&vd->vdev_dtl_lock);
 630         for (int t = 0; t < DTL_TYPES; t++) {
 631                 space_map_unload(&vd->vdev_dtl[t]);
 632                 space_map_destroy(&vd->vdev_dtl[t]);
 633         }
 634         mutex_exit(&vd->vdev_dtl_lock);
 635 
 636         mutex_destroy(&vd->vdev_dtl_lock);
 637         mutex_destroy(&vd->vdev_stat_lock);
 638         mutex_destroy(&vd->vdev_probe_lock);
 639 
 640         if (vd == spa->spa_root_vdev)
 641                 spa->spa_root_vdev = NULL;
 642 
 643         kmem_free(vd, sizeof (vdev_t));
 644 }
 645 
 646 /*
 647  * Transfer top-level vdev state from svd to tvd.
 648  */
 649 static void
 650 vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
 651 {
 652         spa_t *spa = svd->vdev_spa;
 653         metaslab_t *msp;
 654         vdev_t *vd;
 655         int t;
 656 
 657         ASSERT(tvd == tvd->vdev_top);
 658 
 659         tvd->vdev_ms_array = svd->vdev_ms_array;
 660         tvd->vdev_ms_shift = svd->vdev_ms_shift;
 661         tvd->vdev_ms_count = svd->vdev_ms_count;
 662 
 663         svd->vdev_ms_array = 0;
 664         svd->vdev_ms_shift = 0;
 665         svd->vdev_ms_count = 0;
 666 
 667         if (tvd->vdev_mg)
 668                 ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
 669         tvd->vdev_mg = svd->vdev_mg;
 670         tvd->vdev_ms = svd->vdev_ms;
 671 
 672         svd->vdev_mg = NULL;
 673         svd->vdev_ms = NULL;
 674 
 675         if (tvd->vdev_mg != NULL)
 676                 tvd->vdev_mg->mg_vd = tvd;
 677 
 678         tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
 679         tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
 680         tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
 681 
 682         svd->vdev_stat.vs_alloc = 0;
 683         svd->vdev_stat.vs_space = 0;
 684         svd->vdev_stat.vs_dspace = 0;
 685 
 686         for (t = 0; t < TXG_SIZE; t++) {
 687                 while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
 688                         (void) txg_list_add(&tvd->vdev_ms_list, msp, t);
 689                 while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
 690                         (void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
 691                 if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
 692                         (void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
 693         }
 694 
 695         if (list_link_active(&svd->vdev_config_dirty_node)) {
 696                 vdev_config_clean(svd);
 697                 vdev_config_dirty(tvd);
 698         }
 699 
 700         if (list_link_active(&svd->vdev_state_dirty_node)) {
 701                 vdev_state_clean(svd);
 702                 vdev_state_dirty(tvd);
 703         }
 704 
 705         tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
 706         svd->vdev_deflate_ratio = 0;
 707 
 708         tvd->vdev_islog = svd->vdev_islog;
 709         svd->vdev_islog = 0;
 710 }
 711 
 712 static void
 713 vdev_top_update(vdev_t *tvd, vdev_t *vd)
 714 {
 715         if (vd == NULL)
 716                 return;
 717 
 718         vd->vdev_top = tvd;
 719 
 720         for (int c = 0; c < vd->vdev_children; c++)
 721                 vdev_top_update(tvd, vd->vdev_child[c]);
 722 }
 723 
 724 /*
 725  * Add a mirror/replacing vdev above an existing vdev.
 726  */
 727 vdev_t *
 728 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
 729 {
 730         spa_t *spa = cvd->vdev_spa;
 731         vdev_t *pvd = cvd->vdev_parent;
 732         vdev_t *mvd;
 733 
 734         ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
 735 
 736         mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
 737 
 738         mvd->vdev_asize = cvd->vdev_asize;
 739         mvd->vdev_min_asize = cvd->vdev_min_asize;
 740         mvd->vdev_max_asize = cvd->vdev_max_asize;
 741         mvd->vdev_ashift = cvd->vdev_ashift;
 742         mvd->vdev_state = cvd->vdev_state;
 743         mvd->vdev_crtxg = cvd->vdev_crtxg;
 744 
 745         vdev_remove_child(pvd, cvd);
 746         vdev_add_child(pvd, mvd);
 747         cvd->vdev_id = mvd->vdev_children;
 748         vdev_add_child(mvd, cvd);
 749         vdev_top_update(cvd->vdev_top, cvd->vdev_top);
 750 
 751         if (mvd == mvd->vdev_top)
 752                 vdev_top_transfer(cvd, mvd);
 753 
 754         return (mvd);
 755 }
 756 
 757 /*
 758  * Remove a 1-way mirror/replacing vdev from the tree.
 759  */
 760 void
 761 vdev_remove_parent(vdev_t *cvd)
 762 {
 763         vdev_t *mvd = cvd->vdev_parent;
 764         vdev_t *pvd = mvd->vdev_parent;
 765 
 766         ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
 767 
 768         ASSERT(mvd->vdev_children == 1);
 769         ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
 770             mvd->vdev_ops == &vdev_replacing_ops ||
 771             mvd->vdev_ops == &vdev_spare_ops);
 772         cvd->vdev_ashift = mvd->vdev_ashift;
 773 
 774         vdev_remove_child(mvd, cvd);
 775         vdev_remove_child(pvd, mvd);
 776 
 777         /*
 778          * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
 779          * Otherwise, we could have detached an offline device, and when we
 780          * go to import the pool we'll think we have two top-level vdevs,
 781          * instead of a different version of the same top-level vdev.
 782          */
 783         if (mvd->vdev_top == mvd) {
 784                 uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
 785                 cvd->vdev_orig_guid = cvd->vdev_guid;
 786                 cvd->vdev_guid += guid_delta;
 787                 cvd->vdev_guid_sum += guid_delta;
 788         }
 789         cvd->vdev_id = mvd->vdev_id;
 790         vdev_add_child(pvd, cvd);
 791         vdev_top_update(cvd->vdev_top, cvd->vdev_top);
 792 
 793         if (cvd == cvd->vdev_top)
 794                 vdev_top_transfer(mvd, cvd);
 795 
 796         ASSERT(mvd->vdev_children == 0);
 797         vdev_free(mvd);
 798 }
 799 
 800 int
 801 vdev_metaslab_init(vdev_t *vd, uint64_t txg)
 802 {
 803         spa_t *spa = vd->vdev_spa;
 804         objset_t *mos = spa->spa_meta_objset;
 805         uint64_t m;
 806         uint64_t oldc = vd->vdev_ms_count;
 807         uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
 808         metaslab_t **mspp;
 809         int error;
 810 
 811         ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
 812 
 813         /*
 814          * This vdev is not being allocated from yet or is a hole.
 815          */
 816         if (vd->vdev_ms_shift == 0)
 817                 return (0);
 818 
 819         ASSERT(!vd->vdev_ishole);
 820 
 821         /*
 822          * Compute the raidz-deflation ratio.  Note, we hard-code
 823          * in 128k (1 << 17) because it is the current "typical" blocksize.
 824          * Even if SPA_MAXBLOCKSIZE changes, this algorithm must never change,
 825          * or we will inconsistently account for existing bp's.
 826          */
 827         vd->vdev_deflate_ratio = (1 << 17) /
 828             (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
 829 
 830         ASSERT(oldc <= newc);
 831 
 832         mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
 833 
 834         if (oldc != 0) {
 835                 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
 836                 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
 837         }
 838 
 839         vd->vdev_ms = mspp;
 840         vd->vdev_ms_count = newc;
 841 
 842         for (m = oldc; m < newc; m++) {
 843                 space_map_obj_t smo = { 0, 0, 0 };
 844                 if (txg == 0) {
 845                         uint64_t object = 0;
 846                         error = dmu_read(mos, vd->vdev_ms_array,
 847                             m * sizeof (uint64_t), sizeof (uint64_t), &object,
 848                             DMU_READ_PREFETCH);
 849                         if (error)
 850                                 return (error);
 851                         if (object != 0) {
 852                                 dmu_buf_t *db;
 853                                 error = dmu_bonus_hold(mos, object, FTAG, &db);
 854                                 if (error)
 855                                         return (error);
 856                                 ASSERT3U(db->db_size, >=, sizeof (smo));
 857                                 bcopy(db->db_data, &smo, sizeof (smo));
 858                                 ASSERT3U(smo.smo_object, ==, object);
 859                                 dmu_buf_rele(db, FTAG);
 860                         }
 861                 }
 862                 vd->vdev_ms[m] = metaslab_init(vd->vdev_mg, &smo,
 863                     m << vd->vdev_ms_shift, 1ULL << vd->vdev_ms_shift, txg);
 864         }
 865 
 866         if (txg == 0)
 867                 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
 868 
 869         /*
 870          * If the vdev is being removed we don't activate
 871          * the metaslabs since we want to ensure that no new
 872          * allocations are performed on this device.
 873          */
 874         if (oldc == 0 && !vd->vdev_removing)
 875                 metaslab_group_activate(vd->vdev_mg);
 876 
 877         if (txg == 0)
 878                 spa_config_exit(spa, SCL_ALLOC, FTAG);
 879 
 880         return (0);
 881 }
 882 
 883 void
 884 vdev_metaslab_fini(vdev_t *vd)
 885 {
 886         uint64_t m;
 887         uint64_t count = vd->vdev_ms_count;
 888 
 889         if (vd->vdev_ms != NULL) {
 890                 metaslab_group_passivate(vd->vdev_mg);
 891                 for (m = 0; m < count; m++)
 892                         if (vd->vdev_ms[m] != NULL)
 893                                 metaslab_fini(vd->vdev_ms[m]);
 894                 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
 895                 vd->vdev_ms = NULL;
 896         }
 897 }
 898 
 899 typedef struct vdev_probe_stats {
 900         boolean_t       vps_readable;
 901         boolean_t       vps_writeable;
 902         int             vps_flags;
 903 } vdev_probe_stats_t;
 904 
 905 static void
 906 vdev_probe_done(zio_t *zio)
 907 {
 908         spa_t *spa = zio->io_spa;
 909         vdev_t *vd = zio->io_vd;
 910         vdev_probe_stats_t *vps = zio->io_private;
 911 
 912         ASSERT(vd->vdev_probe_zio != NULL);
 913 
 914         if (zio->io_type == ZIO_TYPE_READ) {
 915                 if (zio->io_error == 0)
 916                         vps->vps_readable = 1;
 917                 if (zio->io_error == 0 && spa_writeable(spa)) {
 918                         zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
 919                             zio->io_offset, zio->io_size, zio->io_data,
 920                             ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
 921                             ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
 922                 } else {
 923                         zio_buf_free(zio->io_data, zio->io_size);
 924                 }
 925         } else if (zio->io_type == ZIO_TYPE_WRITE) {
 926                 if (zio->io_error == 0)
 927                         vps->vps_writeable = 1;
 928                 zio_buf_free(zio->io_data, zio->io_size);
 929         } else if (zio->io_type == ZIO_TYPE_NULL) {
 930                 zio_t *pio;
 931 
 932                 vd->vdev_cant_read |= !vps->vps_readable;
 933                 vd->vdev_cant_write |= !vps->vps_writeable;
 934 
 935                 if (vdev_readable(vd) &&
 936                     (vdev_writeable(vd) || !spa_writeable(spa))) {
 937                         zio->io_error = 0;
 938                 } else {
 939                         ASSERT(zio->io_error != 0);
 940                         zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
 941                             spa, vd, NULL, 0, 0);
 942                         zio->io_error = SET_ERROR(ENXIO);
 943                 }
 944 
 945                 mutex_enter(&vd->vdev_probe_lock);
 946                 ASSERT(vd->vdev_probe_zio == zio);
 947                 vd->vdev_probe_zio = NULL;
 948                 mutex_exit(&vd->vdev_probe_lock);
 949 
 950                 while ((pio = zio_walk_parents(zio)) != NULL)
 951                         if (!vdev_accessible(vd, pio))
 952                                 pio->io_error = SET_ERROR(ENXIO);
 953 
 954                 kmem_free(vps, sizeof (*vps));
 955         }
 956 }
 957 
 958 /*
 959  * Determine whether this device is accessible.
 960  *
 961  * Read and write to several known locations: the pad regions of each
 962  * vdev label but the first, which we leave alone in case it contains
 963  * a VTOC.
 964  */
 965 zio_t *
 966 vdev_probe(vdev_t *vd, zio_t *zio)
 967 {
 968         spa_t *spa = vd->vdev_spa;
 969         vdev_probe_stats_t *vps = NULL;
 970         zio_t *pio;
 971 
 972         ASSERT(vd->vdev_ops->vdev_op_leaf);
 973 
 974         /*
 975          * Don't probe the probe.
 976          */
 977         if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
 978                 return (NULL);
 979 
 980         /*
 981          * To prevent 'probe storms' when a device fails, we create
 982          * just one probe i/o at a time.  All zios that want to probe
 983          * this vdev will become parents of the probe io.
 984          */
 985         mutex_enter(&vd->vdev_probe_lock);
 986 
 987         if ((pio = vd->vdev_probe_zio) == NULL) {
 988                 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
 989 
 990                 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
 991                     ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
 992                     ZIO_FLAG_TRYHARD;
 993 
 994                 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
 995                         /*
 996                          * vdev_cant_read and vdev_cant_write can only
 997                          * transition from TRUE to FALSE when we have the
 998                          * SCL_ZIO lock as writer; otherwise they can only
 999                          * transition from FALSE to TRUE.  This ensures that
1000                          * any zio looking at these values can assume that
1001                          * failures persist for the life of the I/O.  That's
1002                          * important because when a device has intermittent
1003                          * connectivity problems, we want to ensure that
1004                          * they're ascribed to the device (ENXIO) and not
1005                          * the zio (EIO).
1006                          *
1007                          * Since we hold SCL_ZIO as writer here, clear both
1008                          * values so the probe can reevaluate from first
1009                          * principles.
1010                          */
1011                         vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1012                         vd->vdev_cant_read = B_FALSE;
1013                         vd->vdev_cant_write = B_FALSE;
1014                 }
1015 
1016                 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1017                     vdev_probe_done, vps,
1018                     vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1019 
1020                 /*
1021                  * We can't change the vdev state in this context, so we
1022                  * kick off an async task to do it on our behalf.
1023                  */
1024                 if (zio != NULL) {
1025                         vd->vdev_probe_wanted = B_TRUE;
1026                         spa_async_request(spa, SPA_ASYNC_PROBE);
1027                 }
1028         }
1029 
1030         if (zio != NULL)
1031                 zio_add_child(zio, pio);
1032 
1033         mutex_exit(&vd->vdev_probe_lock);
1034 
1035         if (vps == NULL) {
1036                 ASSERT(zio != NULL);
1037                 return (NULL);
1038         }
1039 
1040         for (int l = 1; l < VDEV_LABELS; l++) {
1041                 zio_nowait(zio_read_phys(pio, vd,
1042                     vdev_label_offset(vd->vdev_psize, l,
1043                     offsetof(vdev_label_t, vl_pad2)),
1044                     VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1045                     ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1046                     ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1047         }
1048 
1049         if (zio == NULL)
1050                 return (pio);
1051 
1052         zio_nowait(pio);
1053         return (NULL);
1054 }
1055 
1056 static void
1057 vdev_open_child(void *arg)
1058 {
1059         vdev_t *vd = arg;
1060 
1061         vd->vdev_open_thread = curthread;
1062         vd->vdev_open_error = vdev_open(vd);
1063         vd->vdev_open_thread = NULL;
1064 }
1065 
1066 boolean_t
1067 vdev_uses_zvols(vdev_t *vd)
1068 {
1069         if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1070             strlen(ZVOL_DIR)) == 0)
1071                 return (B_TRUE);
1072         for (int c = 0; c < vd->vdev_children; c++)
1073                 if (vdev_uses_zvols(vd->vdev_child[c]))
1074                         return (B_TRUE);
1075         return (B_FALSE);
1076 }
1077 
1078 void
1079 vdev_open_children(vdev_t *vd)
1080 {
1081         taskq_t *tq;
1082         int children = vd->vdev_children;
1083 
1084         /*
1085          * in order to handle pools on top of zvols, do the opens
1086          * in a single thread so that the same thread holds the
1087          * spa_namespace_lock
1088          */
1089         if (vdev_uses_zvols(vd)) {
1090                 for (int c = 0; c < children; c++)
1091                         vd->vdev_child[c]->vdev_open_error =
1092                             vdev_open(vd->vdev_child[c]);
1093                 return;
1094         }
1095         tq = taskq_create("vdev_open", children, minclsyspri,
1096             children, children, TASKQ_PREPOPULATE);
1097 
1098         for (int c = 0; c < children; c++)
1099                 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1100                     TQ_SLEEP) != NULL);
1101 
1102         taskq_destroy(tq);
1103 }
1104 
1105 /*
1106  * Prepare a virtual device for access.
1107  */
1108 int
1109 vdev_open(vdev_t *vd)
1110 {
1111         spa_t *spa = vd->vdev_spa;
1112         int error;
1113         uint64_t osize = 0;
1114         uint64_t max_osize = 0;
1115         uint64_t asize, max_asize, psize;
1116         uint64_t ashift = 0;
1117 
1118         ASSERT(vd->vdev_open_thread == curthread ||
1119             spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1120         ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1121             vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1122             vd->vdev_state == VDEV_STATE_OFFLINE);
1123 
1124         vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1125         vd->vdev_cant_read = B_FALSE;
1126         vd->vdev_cant_write = B_FALSE;
1127         vd->vdev_min_asize = vdev_get_min_asize(vd);
1128 
1129         /*
1130          * If this vdev is not removed, check its fault status.  If it's
1131          * faulted, bail out of the open.
1132          */
1133         if (!vd->vdev_removed && vd->vdev_faulted) {
1134                 ASSERT(vd->vdev_children == 0);
1135                 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1136                     vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1137                 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1138                     vd->vdev_label_aux);
1139                 return (SET_ERROR(ENXIO));
1140         } else if (vd->vdev_offline) {
1141                 ASSERT(vd->vdev_children == 0);
1142                 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1143                 return (SET_ERROR(ENXIO));
1144         }
1145 
1146         error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1147 
1148         /*
1149          * Reset the vdev_reopening flag so that we actually close
1150          * the vdev on error.
1151          */
1152         vd->vdev_reopening = B_FALSE;
1153         if (zio_injection_enabled && error == 0)
1154                 error = zio_handle_device_injection(vd, NULL, ENXIO);
1155 
1156         if (error) {
1157                 if (vd->vdev_removed &&
1158                     vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1159                         vd->vdev_removed = B_FALSE;
1160 
1161                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1162                     vd->vdev_stat.vs_aux);
1163                 return (error);
1164         }
1165 
1166         vd->vdev_removed = B_FALSE;
1167 
1168         /*
1169          * Recheck the faulted flag now that we have confirmed that
1170          * the vdev is accessible.  If we're faulted, bail.
1171          */
1172         if (vd->vdev_faulted) {
1173                 ASSERT(vd->vdev_children == 0);
1174                 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1175                     vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1176                 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1177                     vd->vdev_label_aux);
1178                 return (SET_ERROR(ENXIO));
1179         }
1180 
1181         if (vd->vdev_degraded) {
1182                 ASSERT(vd->vdev_children == 0);
1183                 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1184                     VDEV_AUX_ERR_EXCEEDED);
1185         } else {
1186                 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1187         }
1188 
1189         /*
1190          * For hole or missing vdevs we just return success.
1191          */
1192         if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1193                 return (0);
1194 
1195         for (int c = 0; c < vd->vdev_children; c++) {
1196                 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1197                         vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1198                             VDEV_AUX_NONE);
1199                         break;
1200                 }
1201         }
1202 
1203         osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1204         max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1205 
1206         if (vd->vdev_children == 0) {
1207                 if (osize < SPA_MINDEVSIZE) {
1208                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1209                             VDEV_AUX_TOO_SMALL);
1210                         return (SET_ERROR(EOVERFLOW));
1211                 }
1212                 psize = osize;
1213                 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1214                 max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1215                     VDEV_LABEL_END_SIZE);
1216         } else {
1217                 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1218                     (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1219                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1220                             VDEV_AUX_TOO_SMALL);
1221                         return (SET_ERROR(EOVERFLOW));
1222                 }
1223                 psize = 0;
1224                 asize = osize;
1225                 max_asize = max_osize;
1226         }
1227 
1228         vd->vdev_psize = psize;
1229 
1230         /*
1231          * Make sure the allocatable size hasn't shrunk.
1232          */
1233         if (asize < vd->vdev_min_asize) {
1234                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1235                     VDEV_AUX_BAD_LABEL);
1236                 return (SET_ERROR(EINVAL));
1237         }
1238 
1239         if (vd->vdev_asize == 0) {
1240                 /*
1241                  * This is the first-ever open, so use the computed values.
1242                  * For testing purposes, a higher ashift can be requested.
1243                  */
1244                 vd->vdev_asize = asize;
1245                 vd->vdev_max_asize = max_asize;
1246                 vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1247         } else {
1248                 /*
1249                  * Detect if the alignment requirement has increased.
1250                  * We don't want to make the pool unavailable, just
1251                  * issue a warning instead.
1252                  */
1253                 if (ashift > vd->vdev_top->vdev_ashift &&
1254                     vd->vdev_ops->vdev_op_leaf) {
1255                         cmn_err(CE_WARN,
1256                             "Disk, '%s', has a block alignment that is "
1257                             "larger than the pool's alignment\n",
1258                             vd->vdev_path);
1259                 }
1260                 vd->vdev_max_asize = max_asize;
1261         }
1262 
1263         /*
1264          * If all children are healthy and the asize has increased,
1265          * then we've experienced dynamic LUN growth.  If automatic
1266          * expansion is enabled then use the additional space.
1267          */
1268         if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1269             (vd->vdev_expanding || spa->spa_autoexpand))
1270                 vd->vdev_asize = asize;
1271 
1272         vdev_set_min_asize(vd);
1273 
1274         /*
1275          * Ensure we can issue some IO before declaring the
1276          * vdev open for business.
1277          */
1278         if (vd->vdev_ops->vdev_op_leaf &&
1279             (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1280                 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1281                     VDEV_AUX_ERR_EXCEEDED);
1282                 return (error);
1283         }
1284 
1285         /*
1286          * If a leaf vdev has a DTL, and seems healthy, then kick off a
1287          * resilver.  But don't do this if we are doing a reopen for a scrub,
1288          * since this would just restart the scrub we are already doing.
1289          */
1290         if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1291             vdev_resilver_needed(vd, NULL, NULL))
1292                 spa_async_request(spa, SPA_ASYNC_RESILVER);
1293 
1294         return (0);
1295 }
1296 
1297 /*
1298  * Called once the vdevs are all opened, this routine validates the label
1299  * contents.  This needs to be done before vdev_load() so that we don't
1300  * inadvertently do repair I/Os to the wrong device.
1301  *
1302  * If 'strict' is false ignore the spa guid check. This is necessary because
1303  * if the machine crashed during a re-guid the new guid might have been written
1304  * to all of the vdev labels, but not the cached config. The strict check
1305  * will be performed when the pool is opened again using the mos config.
1306  *
1307  * This function will only return failure if one of the vdevs indicates that it
1308  * has since been destroyed or exported.  This is only possible if
1309  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1310  * will be updated but the function will return 0.
1311  */
1312 int
1313 vdev_validate(vdev_t *vd, boolean_t strict)
1314 {
1315         spa_t *spa = vd->vdev_spa;
1316         nvlist_t *label;
1317         uint64_t guid = 0, top_guid;
1318         uint64_t state;
1319 
1320         for (int c = 0; c < vd->vdev_children; c++)
1321                 if (vdev_validate(vd->vdev_child[c], strict) != 0)
1322                         return (SET_ERROR(EBADF));
1323 
1324         /*
1325          * If the device has already failed, or was marked offline, don't do
1326          * any further validation.  Otherwise, label I/O will fail and we will
1327          * overwrite the previous state.
1328          */
1329         if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1330                 uint64_t aux_guid = 0;
1331                 nvlist_t *nvl;
1332                 uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1333                     spa_last_synced_txg(spa) : -1ULL;
1334 
1335                 if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1336                         vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1337                             VDEV_AUX_BAD_LABEL);
1338                         return (0);
1339                 }
1340 
1341                 /*
1342                  * Determine if this vdev has been split off into another
1343                  * pool.  If so, then refuse to open it.
1344                  */
1345                 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1346                     &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1347                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1348                             VDEV_AUX_SPLIT_POOL);
1349                         nvlist_free(label);
1350                         return (0);
1351                 }
1352 
1353                 if (strict && (nvlist_lookup_uint64(label,
1354                     ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1355                     guid != spa_guid(spa))) {
1356                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1357                             VDEV_AUX_CORRUPT_DATA);
1358                         nvlist_free(label);
1359                         return (0);
1360                 }
1361 
1362                 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1363                     != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1364                     &aux_guid) != 0)
1365                         aux_guid = 0;
1366 
1367                 /*
1368                  * If this vdev just became a top-level vdev because its
1369                  * sibling was detached, it will have adopted the parent's
1370                  * vdev guid -- but the label may or may not be on disk yet.
1371                  * Fortunately, either version of the label will have the
1372                  * same top guid, so if we're a top-level vdev, we can
1373                  * safely compare to that instead.
1374                  *
1375                  * If we split this vdev off instead, then we also check the
1376                  * original pool's guid.  We don't want to consider the vdev
1377                  * corrupt if it is partway through a split operation.
1378                  */
1379                 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1380                     &guid) != 0 ||
1381                     nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1382                     &top_guid) != 0 ||
1383                     ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1384                     (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1385                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1386                             VDEV_AUX_CORRUPT_DATA);
1387                         nvlist_free(label);
1388                         return (0);
1389                 }
1390 
1391                 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1392                     &state) != 0) {
1393                         vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1394                             VDEV_AUX_CORRUPT_DATA);
1395                         nvlist_free(label);
1396                         return (0);
1397                 }
1398 
1399                 nvlist_free(label);
1400 
1401                 /*
1402                  * If this is a verbatim import, no need to check the
1403                  * state of the pool.
1404                  */
1405                 if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1406                     spa_load_state(spa) == SPA_LOAD_OPEN &&
1407                     state != POOL_STATE_ACTIVE)
1408                         return (SET_ERROR(EBADF));
1409 
1410                 /*
1411                  * If we were able to open and validate a vdev that was
1412                  * previously marked permanently unavailable, clear that state
1413                  * now.
1414                  */
1415                 if (vd->vdev_not_present)
1416                         vd->vdev_not_present = 0;
1417         }
1418 
1419         return (0);
1420 }
1421 
1422 /*
1423  * Close a virtual device.
1424  */
1425 void
1426 vdev_close(vdev_t *vd)
1427 {
1428         spa_t *spa = vd->vdev_spa;
1429         vdev_t *pvd = vd->vdev_parent;
1430 
1431         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1432 
1433         /*
1434          * If our parent is reopening, then we are as well, unless we are
1435          * going offline.
1436          */
1437         if (pvd != NULL && pvd->vdev_reopening)
1438                 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1439 
1440         vd->vdev_ops->vdev_op_close(vd);
1441 
1442         vdev_cache_purge(vd);
1443 
1444         /*
1445          * We record the previous state before we close it, so that if we are
1446          * doing a reopen(), we don't generate FMA ereports if we notice that
1447          * it's still faulted.
1448          */
1449         vd->vdev_prevstate = vd->vdev_state;
1450 
1451         if (vd->vdev_offline)
1452                 vd->vdev_state = VDEV_STATE_OFFLINE;
1453         else
1454                 vd->vdev_state = VDEV_STATE_CLOSED;
1455         vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1456 }
1457 
1458 void
1459 vdev_hold(vdev_t *vd)
1460 {
1461         spa_t *spa = vd->vdev_spa;
1462 
1463         ASSERT(spa_is_root(spa));
1464         if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1465                 return;
1466 
1467         for (int c = 0; c < vd->vdev_children; c++)
1468                 vdev_hold(vd->vdev_child[c]);
1469 
1470         if (vd->vdev_ops->vdev_op_leaf)
1471                 vd->vdev_ops->vdev_op_hold(vd);
1472 }
1473 
1474 void
1475 vdev_rele(vdev_t *vd)
1476 {
1477         spa_t *spa = vd->vdev_spa;
1478 
1479         ASSERT(spa_is_root(spa));
1480         for (int c = 0; c < vd->vdev_children; c++)
1481                 vdev_rele(vd->vdev_child[c]);
1482 
1483         if (vd->vdev_ops->vdev_op_leaf)
1484                 vd->vdev_ops->vdev_op_rele(vd);
1485 }
1486 
1487 /*
1488  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1489  * reopen leaf vdevs which had previously been opened as they might deadlock
1490  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1491  * If the leaf has never been opened then open it, as usual.
1492  */
1493 void
1494 vdev_reopen(vdev_t *vd)
1495 {
1496         spa_t *spa = vd->vdev_spa;
1497 
1498         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1499 
1500         /* set the reopening flag unless we're taking the vdev offline */
1501         vd->vdev_reopening = !vd->vdev_offline;
1502         vdev_close(vd);
1503         (void) vdev_open(vd);
1504 
1505         /*
1506          * Call vdev_validate() here to make sure we have the same device.
1507          * Otherwise, a device with an invalid label could be successfully
1508          * opened in response to vdev_reopen().
1509          */
1510         if (vd->vdev_aux) {
1511                 (void) vdev_validate_aux(vd);
1512                 if (vdev_readable(vd) && vdev_writeable(vd) &&
1513                     vd->vdev_aux == &spa->spa_l2cache &&
1514                     !l2arc_vdev_present(vd))
1515                         l2arc_add_vdev(spa, vd);
1516         } else {
1517                 (void) vdev_validate(vd, B_TRUE);
1518         }
1519 
1520         /*
1521          * Reassess parent vdev's health.
1522          */
1523         vdev_propagate_state(vd);
1524 }
1525 
1526 int
1527 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1528 {
1529         int error;
1530 
1531         /*
1532          * Normally, partial opens (e.g. of a mirror) are allowed.
1533          * For a create, however, we want to fail the request if
1534          * there are any components we can't open.
1535          */
1536         error = vdev_open(vd);
1537 
1538         if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1539                 vdev_close(vd);
1540                 return (error ? error : ENXIO);
1541         }
1542 
1543         /*
1544          * Recursively initialize all labels.
1545          */
1546         if ((error = vdev_label_init(vd, txg, isreplacing ?
1547             VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1548                 vdev_close(vd);
1549                 return (error);
1550         }
1551 
1552         return (0);
1553 }
1554 
1555 void
1556 vdev_metaslab_set_size(vdev_t *vd)
1557 {
1558         /*
1559          * Aim for roughly 200 metaslabs per vdev.
1560          */
1561         vd->vdev_ms_shift = highbit(vd->vdev_asize / 200);
1562         vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1563 }
1564 
1565 void
1566 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1567 {
1568         ASSERT(vd == vd->vdev_top);
1569         ASSERT(!vd->vdev_ishole);
1570         ASSERT(ISP2(flags));
1571         ASSERT(spa_writeable(vd->vdev_spa));
1572 
1573         if (flags & VDD_METASLAB)
1574                 (void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1575 
1576         if (flags & VDD_DTL)
1577                 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1578 
1579         (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1580 }
1581 
1582 /*
1583  * DTLs.
1584  *
1585  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1586  * the vdev has less than perfect replication.  There are four kinds of DTL:
1587  *
1588  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1589  *
1590  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1591  *
1592  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1593  *      scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1594  *      txgs that was scrubbed.
1595  *
1596  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1597  *      persistent errors or just some device being offline.
1598  *      Unlike the other three, the DTL_OUTAGE map is not generally
1599  *      maintained; it's only computed when needed, typically to
1600  *      determine whether a device can be detached.
1601  *
1602  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1603  * either has the data or it doesn't.
1604  *
1605  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1606  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1607  * if any child is less than fully replicated, then so is its parent.
1608  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1609  * comprising only those txgs which appear in 'maxfaults' or more children;
1610  * those are the txgs we don't have enough replication to read.  For example,
1611  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1612  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1613  * two child DTL_MISSING maps.
1614  *
1615  * It should be clear from the above that to compute the DTLs and outage maps
1616  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1617  * Therefore, that is all we keep on disk.  When loading the pool, or after
1618  * a configuration change, we generate all other DTLs from first principles.
1619  */
1620 void
1621 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1622 {
1623         space_map_t *sm = &vd->vdev_dtl[t];
1624 
1625         ASSERT(t < DTL_TYPES);
1626         ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1627         ASSERT(spa_writeable(vd->vdev_spa));
1628 
1629         mutex_enter(sm->sm_lock);
1630         if (!space_map_contains(sm, txg, size))
1631                 space_map_add(sm, txg, size);
1632         mutex_exit(sm->sm_lock);
1633 }
1634 
1635 boolean_t
1636 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1637 {
1638         space_map_t *sm = &vd->vdev_dtl[t];
1639         boolean_t dirty = B_FALSE;
1640 
1641         ASSERT(t < DTL_TYPES);
1642         ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1643 
1644         mutex_enter(sm->sm_lock);
1645         if (sm->sm_space != 0)
1646                 dirty = space_map_contains(sm, txg, size);
1647         mutex_exit(sm->sm_lock);
1648 
1649         return (dirty);
1650 }
1651 
1652 boolean_t
1653 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1654 {
1655         space_map_t *sm = &vd->vdev_dtl[t];
1656         boolean_t empty;
1657 
1658         mutex_enter(sm->sm_lock);
1659         empty = (sm->sm_space == 0);
1660         mutex_exit(sm->sm_lock);
1661 
1662         return (empty);
1663 }
1664 
1665 /*
1666  * Returns the lowest txg in the DTL range.
1667  */
1668 static uint64_t
1669 vdev_dtl_min(vdev_t *vd)
1670 {
1671         space_seg_t *ss;
1672 
1673         ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1674         ASSERT3U(vd->vdev_dtl[DTL_MISSING].sm_space, !=, 0);
1675         ASSERT0(vd->vdev_children);
1676 
1677         ss = avl_first(&vd->vdev_dtl[DTL_MISSING].sm_root);
1678         return (ss->ss_start - 1);
1679 }
1680 
1681 /*
1682  * Returns the highest txg in the DTL.
1683  */
1684 static uint64_t
1685 vdev_dtl_max(vdev_t *vd)
1686 {
1687         space_seg_t *ss;
1688 
1689         ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1690         ASSERT3U(vd->vdev_dtl[DTL_MISSING].sm_space, !=, 0);
1691         ASSERT0(vd->vdev_children);
1692 
1693         ss = avl_last(&vd->vdev_dtl[DTL_MISSING].sm_root);
1694         return (ss->ss_end);
1695 }
1696 
1697 /*
1698  * Determine if a resilvering vdev should remove any DTL entries from
1699  * its range. If the vdev was resilvering for the entire duration of the
1700  * scan then it should excise that range from its DTLs. Otherwise, this
1701  * vdev is considered partially resilvered and should leave its DTL
1702  * entries intact. The comment in vdev_dtl_reassess() describes how we
1703  * excise the DTLs.
1704  */
1705 static boolean_t
1706 vdev_dtl_should_excise(vdev_t *vd)
1707 {
1708         spa_t *spa = vd->vdev_spa;
1709         dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1710 
1711         ASSERT0(scn->scn_phys.scn_errors);
1712         ASSERT0(vd->vdev_children);
1713 
1714         if (vd->vdev_resilver_txg == 0 ||
1715             vd->vdev_dtl[DTL_MISSING].sm_space == 0)
1716                 return (B_TRUE);
1717 
1718         /*
1719          * When a resilver is initiated the scan will assign the scn_max_txg
1720          * value to the highest txg value that exists in all DTLs. If this
1721          * device's max DTL is not part of this scan (i.e. it is not in
1722          * the range (scn_min_txg, scn_max_txg] then it is not eligible
1723          * for excision.
1724          */
1725         if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1726                 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1727                 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1728                 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1729                 return (B_TRUE);
1730         }
1731         return (B_FALSE);
1732 }
1733 
1734 /*
1735  * Reassess DTLs after a config change or scrub completion.
1736  */
1737 void
1738 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1739 {
1740         spa_t *spa = vd->vdev_spa;
1741         avl_tree_t reftree;
1742         int minref;
1743 
1744         ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1745 
1746         for (int c = 0; c < vd->vdev_children; c++)
1747                 vdev_dtl_reassess(vd->vdev_child[c], txg,
1748                     scrub_txg, scrub_done);
1749 
1750         if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1751                 return;
1752 
1753         if (vd->vdev_ops->vdev_op_leaf) {
1754                 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1755 
1756                 mutex_enter(&vd->vdev_dtl_lock);
1757 
1758                 /*
1759                  * If we've completed a scan cleanly then determine
1760                  * if this vdev should remove any DTLs. We only want to
1761                  * excise regions on vdevs that were available during
1762                  * the entire duration of this scan.
1763                  */
1764                 if (scrub_txg != 0 &&
1765                     (spa->spa_scrub_started ||
1766                     (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1767                     vdev_dtl_should_excise(vd)) {
1768                         /*
1769                          * We completed a scrub up to scrub_txg.  If we
1770                          * did it without rebooting, then the scrub dtl
1771                          * will be valid, so excise the old region and
1772                          * fold in the scrub dtl.  Otherwise, leave the
1773                          * dtl as-is if there was an error.
1774                          *
1775                          * There's little trick here: to excise the beginning
1776                          * of the DTL_MISSING map, we put it into a reference
1777                          * tree and then add a segment with refcnt -1 that
1778                          * covers the range [0, scrub_txg).  This means
1779                          * that each txg in that range has refcnt -1 or 0.
1780                          * We then add DTL_SCRUB with a refcnt of 2, so that
1781                          * entries in the range [0, scrub_txg) will have a
1782                          * positive refcnt -- either 1 or 2.  We then convert
1783                          * the reference tree into the new DTL_MISSING map.
1784                          */
1785                         space_map_ref_create(&reftree);
1786                         space_map_ref_add_map(&reftree,
1787                             &vd->vdev_dtl[DTL_MISSING], 1);
1788                         space_map_ref_add_seg(&reftree, 0, scrub_txg, -1);
1789                         space_map_ref_add_map(&reftree,
1790                             &vd->vdev_dtl[DTL_SCRUB], 2);
1791                         space_map_ref_generate_map(&reftree,
1792                             &vd->vdev_dtl[DTL_MISSING], 1);
1793                         space_map_ref_destroy(&reftree);
1794                 }
1795                 space_map_vacate(&vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1796                 space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1797                     space_map_add, &vd->vdev_dtl[DTL_PARTIAL]);
1798                 if (scrub_done)
1799                         space_map_vacate(&vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1800                 space_map_vacate(&vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1801                 if (!vdev_readable(vd))
1802                         space_map_add(&vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1803                 else
1804                         space_map_walk(&vd->vdev_dtl[DTL_MISSING],
1805                             space_map_add, &vd->vdev_dtl[DTL_OUTAGE]);
1806 
1807                 /*
1808                  * If the vdev was resilvering and no longer has any
1809                  * DTLs then reset its resilvering flag.
1810                  */
1811                 if (vd->vdev_resilver_txg != 0 &&
1812                     vd->vdev_dtl[DTL_MISSING].sm_space == 0 &&
1813                     vd->vdev_dtl[DTL_OUTAGE].sm_space == 0)
1814                         vd->vdev_resilver_txg = 0;
1815 
1816                 mutex_exit(&vd->vdev_dtl_lock);
1817 
1818                 if (txg != 0)
1819                         vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1820                 return;
1821         }
1822 
1823         mutex_enter(&vd->vdev_dtl_lock);
1824         for (int t = 0; t < DTL_TYPES; t++) {
1825                 /* account for child's outage in parent's missing map */
1826                 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1827                 if (t == DTL_SCRUB)
1828                         continue;                       /* leaf vdevs only */
1829                 if (t == DTL_PARTIAL)
1830                         minref = 1;                     /* i.e. non-zero */
1831                 else if (vd->vdev_nparity != 0)
1832                         minref = vd->vdev_nparity + 1;       /* RAID-Z */
1833                 else
1834                         minref = vd->vdev_children;  /* any kind of mirror */
1835                 space_map_ref_create(&reftree);
1836                 for (int c = 0; c < vd->vdev_children; c++) {
1837                         vdev_t *cvd = vd->vdev_child[c];
1838                         mutex_enter(&cvd->vdev_dtl_lock);
1839                         space_map_ref_add_map(&reftree, &cvd->vdev_dtl[s], 1);
1840                         mutex_exit(&cvd->vdev_dtl_lock);
1841                 }
1842                 space_map_ref_generate_map(&reftree, &vd->vdev_dtl[t], minref);
1843                 space_map_ref_destroy(&reftree);
1844         }
1845         mutex_exit(&vd->vdev_dtl_lock);
1846 }
1847 
1848 static int
1849 vdev_dtl_load(vdev_t *vd)
1850 {
1851         spa_t *spa = vd->vdev_spa;
1852         space_map_obj_t *smo = &vd->vdev_dtl_smo;
1853         objset_t *mos = spa->spa_meta_objset;
1854         dmu_buf_t *db;
1855         int error;
1856 
1857         ASSERT(vd->vdev_children == 0);
1858 
1859         if (smo->smo_object == 0)
1860                 return (0);
1861 
1862         ASSERT(!vd->vdev_ishole);
1863 
1864         if ((error = dmu_bonus_hold(mos, smo->smo_object, FTAG, &db)) != 0)
1865                 return (error);
1866 
1867         ASSERT3U(db->db_size, >=, sizeof (*smo));
1868         bcopy(db->db_data, smo, sizeof (*smo));
1869         dmu_buf_rele(db, FTAG);
1870 
1871         mutex_enter(&vd->vdev_dtl_lock);
1872         error = space_map_load(&vd->vdev_dtl[DTL_MISSING],
1873             NULL, SM_ALLOC, smo, mos);
1874         mutex_exit(&vd->vdev_dtl_lock);
1875 
1876         return (error);
1877 }
1878 
1879 void
1880 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1881 {
1882         spa_t *spa = vd->vdev_spa;
1883         space_map_obj_t *smo = &vd->vdev_dtl_smo;
1884         space_map_t *sm = &vd->vdev_dtl[DTL_MISSING];
1885         objset_t *mos = spa->spa_meta_objset;
1886         space_map_t smsync;
1887         kmutex_t smlock;
1888         dmu_buf_t *db;
1889         dmu_tx_t *tx;
1890 
1891         ASSERT(!vd->vdev_ishole);
1892 
1893         tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
1894 
1895         if (vd->vdev_detached) {
1896                 if (smo->smo_object != 0) {
1897                         int err = dmu_object_free(mos, smo->smo_object, tx);
1898                         ASSERT0(err);
1899                         smo->smo_object = 0;
1900                 }
1901                 dmu_tx_commit(tx);
1902                 return;
1903         }
1904 
1905         if (smo->smo_object == 0) {
1906                 ASSERT(smo->smo_objsize == 0);
1907                 ASSERT(smo->smo_alloc == 0);
1908                 smo->smo_object = dmu_object_alloc(mos,
1909                     DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
1910                     DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
1911                 ASSERT(smo->smo_object != 0);
1912                 vdev_config_dirty(vd->vdev_top);
1913         }
1914 
1915         mutex_init(&smlock, NULL, MUTEX_DEFAULT, NULL);
1916 
1917         space_map_create(&smsync, sm->sm_start, sm->sm_size, sm->sm_shift,
1918             &smlock);
1919 
1920         mutex_enter(&smlock);
1921 
1922         mutex_enter(&vd->vdev_dtl_lock);
1923         space_map_walk(sm, space_map_add, &smsync);
1924         mutex_exit(&vd->vdev_dtl_lock);
1925 
1926         space_map_truncate(smo, mos, tx);
1927         space_map_sync(&smsync, SM_ALLOC, smo, mos, tx);
1928         space_map_vacate(&smsync, NULL, NULL);
1929 
1930         space_map_destroy(&smsync);
1931 
1932         mutex_exit(&smlock);
1933         mutex_destroy(&smlock);
1934 
1935         VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
1936         dmu_buf_will_dirty(db, tx);
1937         ASSERT3U(db->db_size, >=, sizeof (*smo));
1938         bcopy(smo, db->db_data, sizeof (*smo));
1939         dmu_buf_rele(db, FTAG);
1940 
1941         dmu_tx_commit(tx);
1942 }
1943 
1944 /*
1945  * Determine whether the specified vdev can be offlined/detached/removed
1946  * without losing data.
1947  */
1948 boolean_t
1949 vdev_dtl_required(vdev_t *vd)
1950 {
1951         spa_t *spa = vd->vdev_spa;
1952         vdev_t *tvd = vd->vdev_top;
1953         uint8_t cant_read = vd->vdev_cant_read;
1954         boolean_t required;
1955 
1956         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1957 
1958         if (vd == spa->spa_root_vdev || vd == tvd)
1959                 return (B_TRUE);
1960 
1961         /*
1962          * Temporarily mark the device as unreadable, and then determine
1963          * whether this results in any DTL outages in the top-level vdev.
1964          * If not, we can safely offline/detach/remove the device.
1965          */
1966         vd->vdev_cant_read = B_TRUE;
1967         vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1968         required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
1969         vd->vdev_cant_read = cant_read;
1970         vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
1971 
1972         if (!required && zio_injection_enabled)
1973                 required = !!zio_handle_device_injection(vd, NULL, ECHILD);
1974 
1975         return (required);
1976 }
1977 
1978 /*
1979  * Determine if resilver is needed, and if so the txg range.
1980  */
1981 boolean_t
1982 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
1983 {
1984         boolean_t needed = B_FALSE;
1985         uint64_t thismin = UINT64_MAX;
1986         uint64_t thismax = 0;
1987 
1988         if (vd->vdev_children == 0) {
1989                 mutex_enter(&vd->vdev_dtl_lock);
1990                 if (vd->vdev_dtl[DTL_MISSING].sm_space != 0 &&
1991                     vdev_writeable(vd)) {
1992 
1993                         thismin = vdev_dtl_min(vd);
1994                         thismax = vdev_dtl_max(vd);
1995                         needed = B_TRUE;
1996                 }
1997                 mutex_exit(&vd->vdev_dtl_lock);
1998         } else {
1999                 for (int c = 0; c < vd->vdev_children; c++) {
2000                         vdev_t *cvd = vd->vdev_child[c];
2001                         uint64_t cmin, cmax;
2002 
2003                         if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2004                                 thismin = MIN(thismin, cmin);
2005                                 thismax = MAX(thismax, cmax);
2006                                 needed = B_TRUE;
2007                         }
2008                 }
2009         }
2010 
2011         if (needed && minp) {
2012                 *minp = thismin;
2013                 *maxp = thismax;
2014         }
2015         return (needed);
2016 }
2017 
2018 void
2019 vdev_load(vdev_t *vd)
2020 {
2021         /*
2022          * Recursively load all children.
2023          */
2024         for (int c = 0; c < vd->vdev_children; c++)
2025                 vdev_load(vd->vdev_child[c]);
2026 
2027         /*
2028          * If this is a top-level vdev, initialize its metaslabs.
2029          */
2030         if (vd == vd->vdev_top && !vd->vdev_ishole &&
2031             (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2032             vdev_metaslab_init(vd, 0) != 0))
2033                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2034                     VDEV_AUX_CORRUPT_DATA);
2035 
2036         /*
2037          * If this is a leaf vdev, load its DTL.
2038          */
2039         if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2040                 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2041                     VDEV_AUX_CORRUPT_DATA);
2042 }
2043 
2044 /*
2045  * The special vdev case is used for hot spares and l2cache devices.  Its
2046  * sole purpose it to set the vdev state for the associated vdev.  To do this,
2047  * we make sure that we can open the underlying device, then try to read the
2048  * label, and make sure that the label is sane and that it hasn't been
2049  * repurposed to another pool.
2050  */
2051 int
2052 vdev_validate_aux(vdev_t *vd)
2053 {
2054         nvlist_t *label;
2055         uint64_t guid, version;
2056         uint64_t state;
2057 
2058         if (!vdev_readable(vd))
2059                 return (0);
2060 
2061         if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2062                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2063                     VDEV_AUX_CORRUPT_DATA);
2064                 return (-1);
2065         }
2066 
2067         if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2068             !SPA_VERSION_IS_SUPPORTED(version) ||
2069             nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2070             guid != vd->vdev_guid ||
2071             nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2072                 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2073                     VDEV_AUX_CORRUPT_DATA);
2074                 nvlist_free(label);
2075                 return (-1);
2076         }
2077 
2078         /*
2079          * We don't actually check the pool state here.  If it's in fact in
2080          * use by another pool, we update this fact on the fly when requested.
2081          */
2082         nvlist_free(label);
2083         return (0);
2084 }
2085 
2086 void
2087 vdev_remove(vdev_t *vd, uint64_t txg)
2088 {
2089         spa_t *spa = vd->vdev_spa;
2090         objset_t *mos = spa->spa_meta_objset;
2091         dmu_tx_t *tx;
2092 
2093         tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2094 
2095         if (vd->vdev_dtl_smo.smo_object) {
2096                 ASSERT0(vd->vdev_dtl_smo.smo_alloc);
2097                 (void) dmu_object_free(mos, vd->vdev_dtl_smo.smo_object, tx);
2098                 vd->vdev_dtl_smo.smo_object = 0;
2099         }
2100 
2101         if (vd->vdev_ms != NULL) {
2102                 for (int m = 0; m < vd->vdev_ms_count; m++) {
2103                         metaslab_t *msp = vd->vdev_ms[m];
2104 
2105                         if (msp == NULL || msp->ms_smo.smo_object == 0)
2106                                 continue;
2107 
2108                         ASSERT0(msp->ms_smo.smo_alloc);
2109                         (void) dmu_object_free(mos, msp->ms_smo.smo_object, tx);
2110                         msp->ms_smo.smo_object = 0;
2111                 }
2112         }
2113 
2114         if (vd->vdev_ms_array) {
2115                 (void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2116                 vd->vdev_ms_array = 0;
2117                 vd->vdev_ms_shift = 0;
2118         }
2119         dmu_tx_commit(tx);
2120 }
2121 
2122 void
2123 vdev_sync_done(vdev_t *vd, uint64_t txg)
2124 {
2125         metaslab_t *msp;
2126         boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2127 
2128         ASSERT(!vd->vdev_ishole);
2129 
2130         while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2131                 metaslab_sync_done(msp, txg);
2132 
2133         if (reassess)
2134                 metaslab_sync_reassess(vd->vdev_mg);
2135 }
2136 
2137 void
2138 vdev_sync(vdev_t *vd, uint64_t txg)
2139 {
2140         spa_t *spa = vd->vdev_spa;
2141         vdev_t *lvd;
2142         metaslab_t *msp;
2143         dmu_tx_t *tx;
2144 
2145         ASSERT(!vd->vdev_ishole);
2146 
2147         if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2148                 ASSERT(vd == vd->vdev_top);
2149                 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2150                 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2151                     DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2152                 ASSERT(vd->vdev_ms_array != 0);
2153                 vdev_config_dirty(vd);
2154                 dmu_tx_commit(tx);
2155         }
2156 
2157         /*
2158          * Remove the metadata associated with this vdev once it's empty.
2159          */
2160         if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2161                 vdev_remove(vd, txg);
2162 
2163         while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2164                 metaslab_sync(msp, txg);
2165                 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2166         }
2167 
2168         while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2169                 vdev_dtl_sync(lvd, txg);
2170 
2171         (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2172 }
2173 
2174 uint64_t
2175 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2176 {
2177         return (vd->vdev_ops->vdev_op_asize(vd, psize));
2178 }
2179 
2180 /*
2181  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2182  * not be opened, and no I/O is attempted.
2183  */
2184 int
2185 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2186 {
2187         vdev_t *vd, *tvd;
2188 
2189         spa_vdev_state_enter(spa, SCL_NONE);
2190 
2191         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2192                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2193 
2194         if (!vd->vdev_ops->vdev_op_leaf)
2195                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2196 
2197         tvd = vd->vdev_top;
2198 
2199         /*
2200          * We don't directly use the aux state here, but if we do a
2201          * vdev_reopen(), we need this value to be present to remember why we
2202          * were faulted.
2203          */
2204         vd->vdev_label_aux = aux;
2205 
2206         /*
2207          * Faulted state takes precedence over degraded.
2208          */
2209         vd->vdev_delayed_close = B_FALSE;
2210         vd->vdev_faulted = 1ULL;
2211         vd->vdev_degraded = 0ULL;
2212         vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2213 
2214         /*
2215          * If this device has the only valid copy of the data, then
2216          * back off and simply mark the vdev as degraded instead.
2217          */
2218         if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2219                 vd->vdev_degraded = 1ULL;
2220                 vd->vdev_faulted = 0ULL;
2221 
2222                 /*
2223                  * If we reopen the device and it's not dead, only then do we
2224                  * mark it degraded.
2225                  */
2226                 vdev_reopen(tvd);
2227 
2228                 if (vdev_readable(vd))
2229                         vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2230         }
2231 
2232         return (spa_vdev_state_exit(spa, vd, 0));
2233 }
2234 
2235 /*
2236  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2237  * user that something is wrong.  The vdev continues to operate as normal as far
2238  * as I/O is concerned.
2239  */
2240 int
2241 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2242 {
2243         vdev_t *vd;
2244 
2245         spa_vdev_state_enter(spa, SCL_NONE);
2246 
2247         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2248                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2249 
2250         if (!vd->vdev_ops->vdev_op_leaf)
2251                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2252 
2253         /*
2254          * If the vdev is already faulted, then don't do anything.
2255          */
2256         if (vd->vdev_faulted || vd->vdev_degraded)
2257                 return (spa_vdev_state_exit(spa, NULL, 0));
2258 
2259         vd->vdev_degraded = 1ULL;
2260         if (!vdev_is_dead(vd))
2261                 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2262                     aux);
2263 
2264         return (spa_vdev_state_exit(spa, vd, 0));
2265 }
2266 
2267 /*
2268  * Online the given vdev.
2269  *
2270  * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2271  * spare device should be detached when the device finishes resilvering.
2272  * Second, the online should be treated like a 'test' online case, so no FMA
2273  * events are generated if the device fails to open.
2274  */
2275 int
2276 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2277 {
2278         vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2279 
2280         spa_vdev_state_enter(spa, SCL_NONE);
2281 
2282         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2283                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2284 
2285         if (!vd->vdev_ops->vdev_op_leaf)
2286                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2287 
2288         tvd = vd->vdev_top;
2289         vd->vdev_offline = B_FALSE;
2290         vd->vdev_tmpoffline = B_FALSE;
2291         vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2292         vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2293 
2294         /* XXX - L2ARC 1.0 does not support expansion */
2295         if (!vd->vdev_aux) {
2296                 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2297                         pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2298         }
2299 
2300         vdev_reopen(tvd);
2301         vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2302 
2303         if (!vd->vdev_aux) {
2304                 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2305                         pvd->vdev_expanding = B_FALSE;
2306         }
2307 
2308         if (newstate)
2309                 *newstate = vd->vdev_state;
2310         if ((flags & ZFS_ONLINE_UNSPARE) &&
2311             !vdev_is_dead(vd) && vd->vdev_parent &&
2312             vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2313             vd->vdev_parent->vdev_child[0] == vd)
2314                 vd->vdev_unspare = B_TRUE;
2315 
2316         if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2317 
2318                 /* XXX - L2ARC 1.0 does not support expansion */
2319                 if (vd->vdev_aux)
2320                         return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2321                 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2322         }
2323         return (spa_vdev_state_exit(spa, vd, 0));
2324 }
2325 
2326 static int
2327 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2328 {
2329         vdev_t *vd, *tvd;
2330         int error = 0;
2331         uint64_t generation;
2332         metaslab_group_t *mg;
2333 
2334 top:
2335         spa_vdev_state_enter(spa, SCL_ALLOC);
2336 
2337         if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2338                 return (spa_vdev_state_exit(spa, NULL, ENODEV));
2339 
2340         if (!vd->vdev_ops->vdev_op_leaf)
2341                 return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2342 
2343         tvd = vd->vdev_top;
2344         mg = tvd->vdev_mg;
2345         generation = spa->spa_config_generation + 1;
2346 
2347         /*
2348          * If the device isn't already offline, try to offline it.
2349          */
2350         if (!vd->vdev_offline) {
2351                 /*
2352                  * If this device has the only valid copy of some data,
2353                  * don't allow it to be offlined. Log devices are always
2354                  * expendable.
2355                  */
2356                 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2357                     vdev_dtl_required(vd))
2358                         return (spa_vdev_state_exit(spa, NULL, EBUSY));
2359 
2360                 /*
2361                  * If the top-level is a slog and it has had allocations
2362                  * then proceed.  We check that the vdev's metaslab group
2363                  * is not NULL since it's possible that we may have just
2364                  * added this vdev but not yet initialized its metaslabs.
2365                  */
2366                 if (tvd->vdev_islog && mg != NULL) {
2367                         /*
2368                          * Prevent any future allocations.
2369                          */
2370                         metaslab_group_passivate(mg);
2371                         (void) spa_vdev_state_exit(spa, vd, 0);
2372 
2373                         error = spa_offline_log(spa);
2374 
2375                         spa_vdev_state_enter(spa, SCL_ALLOC);
2376 
2377                         /*
2378                          * Check to see if the config has changed.
2379                          */
2380                         if (error || generation != spa->spa_config_generation) {
2381                                 metaslab_group_activate(mg);
2382                                 if (error)
2383                                         return (spa_vdev_state_exit(spa,
2384                                             vd, error));
2385                                 (void) spa_vdev_state_exit(spa, vd, 0);
2386                                 goto top;
2387                         }
2388                         ASSERT0(tvd->vdev_stat.vs_alloc);
2389                 }
2390 
2391                 /*
2392                  * Offline this device and reopen its top-level vdev.
2393                  * If the top-level vdev is a log device then just offline
2394                  * it. Otherwise, if this action results in the top-level
2395                  * vdev becoming unusable, undo it and fail the request.
2396                  */
2397                 vd->vdev_offline = B_TRUE;
2398                 vdev_reopen(tvd);
2399 
2400                 if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2401                     vdev_is_dead(tvd)) {
2402                         vd->vdev_offline = B_FALSE;
2403                         vdev_reopen(tvd);
2404                         return (spa_vdev_state_exit(spa, NULL, EBUSY));
2405                 }
2406 
2407                 /*
2408                  * Add the device back into the metaslab rotor so that
2409                  * once we online the device it's open for business.
2410                  */
2411                 if (tvd->vdev_islog && mg != NULL)
2412                         metaslab_group_activate(mg);
2413         }
2414 
2415         vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2416 
2417         return (spa_vdev_state_exit(spa, vd, 0));
2418 }
2419 
2420 int
2421 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2422 {
2423         int error;
2424 
2425         mutex_enter(&spa->spa_vdev_top_lock);
2426         error = vdev_offline_locked(spa, guid, flags);
2427         mutex_exit(&spa->spa_vdev_top_lock);
2428 
2429         return (error);
2430 }
2431 
2432 /*
2433  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2434  * vdev_offline(), we assume the spa config is locked.  We also clear all
2435  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2436  */
2437 void
2438 vdev_clear(spa_t *spa, vdev_t *vd)
2439 {
2440         vdev_t *rvd = spa->spa_root_vdev;
2441 
2442         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2443 
2444         if (vd == NULL)
2445                 vd = rvd;
2446 
2447         vd->vdev_stat.vs_read_errors = 0;
2448         vd->vdev_stat.vs_write_errors = 0;
2449         vd->vdev_stat.vs_checksum_errors = 0;
2450 
2451         for (int c = 0; c < vd->vdev_children; c++)
2452                 vdev_clear(spa, vd->vdev_child[c]);
2453 
2454         /*
2455          * If we're in the FAULTED state or have experienced failed I/O, then
2456          * clear the persistent state and attempt to reopen the device.  We
2457          * also mark the vdev config dirty, so that the new faulted state is
2458          * written out to disk.
2459          */
2460         if (vd->vdev_faulted || vd->vdev_degraded ||
2461             !vdev_readable(vd) || !vdev_writeable(vd)) {
2462 
2463                 /*
2464                  * When reopening in reponse to a clear event, it may be due to
2465                  * a fmadm repair request.  In this case, if the device is
2466                  * still broken, we want to still post the ereport again.
2467                  */
2468                 vd->vdev_forcefault = B_TRUE;
2469 
2470                 vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2471                 vd->vdev_cant_read = B_FALSE;
2472                 vd->vdev_cant_write = B_FALSE;
2473 
2474                 vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2475 
2476                 vd->vdev_forcefault = B_FALSE;
2477 
2478                 if (vd != rvd && vdev_writeable(vd->vdev_top))
2479                         vdev_state_dirty(vd->vdev_top);
2480 
2481                 if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2482                         spa_async_request(spa, SPA_ASYNC_RESILVER);
2483 
2484                 spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2485         }
2486 
2487         /*
2488          * When clearing a FMA-diagnosed fault, we always want to
2489          * unspare the device, as we assume that the original spare was
2490          * done in response to the FMA fault.
2491          */
2492         if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2493             vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2494             vd->vdev_parent->vdev_child[0] == vd)
2495                 vd->vdev_unspare = B_TRUE;
2496 }
2497 
2498 boolean_t
2499 vdev_is_dead(vdev_t *vd)
2500 {
2501         /*
2502          * Holes and missing devices are always considered "dead".
2503          * This simplifies the code since we don't have to check for
2504          * these types of devices in the various code paths.
2505          * Instead we rely on the fact that we skip over dead devices
2506          * before issuing I/O to them.
2507          */
2508         return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2509             vd->vdev_ops == &vdev_missing_ops);
2510 }
2511 
2512 boolean_t
2513 vdev_readable(vdev_t *vd)
2514 {
2515         return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2516 }
2517 
2518 boolean_t
2519 vdev_writeable(vdev_t *vd)
2520 {
2521         return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2522 }
2523 
2524 boolean_t
2525 vdev_allocatable(vdev_t *vd)
2526 {
2527         uint64_t state = vd->vdev_state;
2528 
2529         /*
2530          * We currently allow allocations from vdevs which may be in the
2531          * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2532          * fails to reopen then we'll catch it later when we're holding
2533          * the proper locks.  Note that we have to get the vdev state
2534          * in a local variable because although it changes atomically,
2535          * we're asking two separate questions about it.
2536          */
2537         return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2538             !vd->vdev_cant_write && !vd->vdev_ishole);
2539 }
2540 
2541 boolean_t
2542 vdev_accessible(vdev_t *vd, zio_t *zio)
2543 {
2544         ASSERT(zio->io_vd == vd);
2545 
2546         if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2547                 return (B_FALSE);
2548 
2549         if (zio->io_type == ZIO_TYPE_READ)
2550                 return (!vd->vdev_cant_read);
2551 
2552         if (zio->io_type == ZIO_TYPE_WRITE)
2553                 return (!vd->vdev_cant_write);
2554 
2555         return (B_TRUE);
2556 }
2557 
2558 /*
2559  * Get statistics for the given vdev.
2560  */
2561 void
2562 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2563 {
2564         vdev_t *rvd = vd->vdev_spa->spa_root_vdev;
2565 
2566         mutex_enter(&vd->vdev_stat_lock);
2567         bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2568         vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2569         vs->vs_state = vd->vdev_state;
2570         vs->vs_rsize = vdev_get_min_asize(vd);
2571         if (vd->vdev_ops->vdev_op_leaf)
2572                 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2573         vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2574         mutex_exit(&vd->vdev_stat_lock);
2575 
2576         /*
2577          * If we're getting stats on the root vdev, aggregate the I/O counts
2578          * over all top-level vdevs (i.e. the direct children of the root).
2579          */
2580         if (vd == rvd) {
2581                 for (int c = 0; c < rvd->vdev_children; c++) {
2582                         vdev_t *cvd = rvd->vdev_child[c];
2583                         vdev_stat_t *cvs = &cvd->vdev_stat;
2584 
2585                         mutex_enter(&vd->vdev_stat_lock);
2586                         for (int t = 0; t < ZIO_TYPES; t++) {
2587                                 vs->vs_ops[t] += cvs->vs_ops[t];
2588                                 vs->vs_bytes[t] += cvs->vs_bytes[t];
2589                         }
2590                         cvs->vs_scan_removing = cvd->vdev_removing;
2591                         mutex_exit(&vd->vdev_stat_lock);
2592                 }
2593         }
2594 }
2595 
2596 void
2597 vdev_clear_stats(vdev_t *vd)
2598 {
2599         mutex_enter(&vd->vdev_stat_lock);
2600         vd->vdev_stat.vs_space = 0;
2601         vd->vdev_stat.vs_dspace = 0;
2602         vd->vdev_stat.vs_alloc = 0;
2603         mutex_exit(&vd->vdev_stat_lock);
2604 }
2605 
2606 void
2607 vdev_scan_stat_init(vdev_t *vd)
2608 {
2609         vdev_stat_t *vs = &vd->vdev_stat;
2610 
2611         for (int c = 0; c < vd->vdev_children; c++)
2612                 vdev_scan_stat_init(vd->vdev_child[c]);
2613 
2614         mutex_enter(&vd->vdev_stat_lock);
2615         vs->vs_scan_processed = 0;
2616         mutex_exit(&vd->vdev_stat_lock);
2617 }
2618 
2619 void
2620 vdev_stat_update(zio_t *zio, uint64_t psize)
2621 {
2622         spa_t *spa = zio->io_spa;
2623         vdev_t *rvd = spa->spa_root_vdev;
2624         vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2625         vdev_t *pvd;
2626         uint64_t txg = zio->io_txg;
2627         vdev_stat_t *vs = &vd->vdev_stat;
2628         zio_type_t type = zio->io_type;
2629         int flags = zio->io_flags;
2630 
2631         /*
2632          * If this i/o is a gang leader, it didn't do any actual work.
2633          */
2634         if (zio->io_gang_tree)
2635                 return;
2636 
2637         if (zio->io_error == 0) {
2638                 /*
2639                  * If this is a root i/o, don't count it -- we've already
2640                  * counted the top-level vdevs, and vdev_get_stats() will
2641                  * aggregate them when asked.  This reduces contention on
2642                  * the root vdev_stat_lock and implicitly handles blocks
2643                  * that compress away to holes, for which there is no i/o.
2644                  * (Holes never create vdev children, so all the counters
2645                  * remain zero, which is what we want.)
2646                  *
2647                  * Note: this only applies to successful i/o (io_error == 0)
2648                  * because unlike i/o counts, errors are not additive.
2649                  * When reading a ditto block, for example, failure of
2650                  * one top-level vdev does not imply a root-level error.
2651                  */
2652                 if (vd == rvd)
2653                         return;
2654 
2655                 ASSERT(vd == zio->io_vd);
2656 
2657                 if (flags & ZIO_FLAG_IO_BYPASS)
2658                         return;
2659 
2660                 mutex_enter(&vd->vdev_stat_lock);
2661 
2662                 if (flags & ZIO_FLAG_IO_REPAIR) {
2663                         if (flags & ZIO_FLAG_SCAN_THREAD) {
2664                                 dsl_scan_phys_t *scn_phys =
2665                                     &spa->spa_dsl_pool->dp_scan->scn_phys;
2666                                 uint64_t *processed = &scn_phys->scn_processed;
2667 
2668                                 /* XXX cleanup? */
2669                                 if (vd->vdev_ops->vdev_op_leaf)
2670                                         atomic_add_64(processed, psize);
2671                                 vs->vs_scan_processed += psize;
2672                         }
2673 
2674                         if (flags & ZIO_FLAG_SELF_HEAL)
2675                                 vs->vs_self_healed += psize;
2676                 }
2677 
2678                 vs->vs_ops[type]++;
2679                 vs->vs_bytes[type] += psize;
2680 
2681                 mutex_exit(&vd->vdev_stat_lock);
2682                 return;
2683         }
2684 
2685         if (flags & ZIO_FLAG_SPECULATIVE)
2686                 return;
2687 
2688         /*
2689          * If this is an I/O error that is going to be retried, then ignore the
2690          * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2691          * hard errors, when in reality they can happen for any number of
2692          * innocuous reasons (bus resets, MPxIO link failure, etc).
2693          */
2694         if (zio->io_error == EIO &&
2695             !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2696                 return;
2697 
2698         /*
2699          * Intent logs writes won't propagate their error to the root
2700          * I/O so don't mark these types of failures as pool-level
2701          * errors.
2702          */
2703         if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2704                 return;
2705 
2706         mutex_enter(&vd->vdev_stat_lock);
2707         if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2708                 if (zio->io_error == ECKSUM)
2709                         vs->vs_checksum_errors++;
2710                 else
2711                         vs->vs_read_errors++;
2712         }
2713         if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2714                 vs->vs_write_errors++;
2715         mutex_exit(&vd->vdev_stat_lock);
2716 
2717         if (type == ZIO_TYPE_WRITE && txg != 0 &&
2718             (!(flags & ZIO_FLAG_IO_REPAIR) ||
2719             (flags & ZIO_FLAG_SCAN_THREAD) ||
2720             spa->spa_claiming)) {
2721                 /*
2722                  * This is either a normal write (not a repair), or it's
2723                  * a repair induced by the scrub thread, or it's a repair
2724                  * made by zil_claim() during spa_load() in the first txg.
2725                  * In the normal case, we commit the DTL change in the same
2726                  * txg as the block was born.  In the scrub-induced repair
2727                  * case, we know that scrubs run in first-pass syncing context,
2728                  * so we commit the DTL change in spa_syncing_txg(spa).
2729                  * In the zil_claim() case, we commit in spa_first_txg(spa).
2730                  *
2731                  * We currently do not make DTL entries for failed spontaneous
2732                  * self-healing writes triggered by normal (non-scrubbing)
2733                  * reads, because we have no transactional context in which to
2734                  * do so -- and it's not clear that it'd be desirable anyway.
2735                  */
2736                 if (vd->vdev_ops->vdev_op_leaf) {
2737                         uint64_t commit_txg = txg;
2738                         if (flags & ZIO_FLAG_SCAN_THREAD) {
2739                                 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2740                                 ASSERT(spa_sync_pass(spa) == 1);
2741                                 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2742                                 commit_txg = spa_syncing_txg(spa);
2743                         } else if (spa->spa_claiming) {
2744                                 ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2745                                 commit_txg = spa_first_txg(spa);
2746                         }
2747                         ASSERT(commit_txg >= spa_syncing_txg(spa));
2748                         if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2749                                 return;
2750                         for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2751                                 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2752                         vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2753                 }
2754                 if (vd != rvd)
2755                         vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2756         }
2757 }
2758 
2759 /*
2760  * Update the in-core space usage stats for this vdev, its metaslab class,
2761  * and the root vdev.
2762  */
2763 void
2764 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2765     int64_t space_delta)
2766 {
2767         int64_t dspace_delta = space_delta;
2768         spa_t *spa = vd->vdev_spa;
2769         vdev_t *rvd = spa->spa_root_vdev;
2770         metaslab_group_t *mg = vd->vdev_mg;
2771         metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2772 
2773         ASSERT(vd == vd->vdev_top);
2774 
2775         /*
2776          * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2777          * factor.  We must calculate this here and not at the root vdev
2778          * because the root vdev's psize-to-asize is simply the max of its
2779          * childrens', thus not accurate enough for us.
2780          */
2781         ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2782         ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2783         dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2784             vd->vdev_deflate_ratio;
2785 
2786         mutex_enter(&vd->vdev_stat_lock);
2787         vd->vdev_stat.vs_alloc += alloc_delta;
2788         vd->vdev_stat.vs_space += space_delta;
2789         vd->vdev_stat.vs_dspace += dspace_delta;
2790         mutex_exit(&vd->vdev_stat_lock);
2791 
2792         if (mc == spa_normal_class(spa)) {
2793                 mutex_enter(&rvd->vdev_stat_lock);
2794                 rvd->vdev_stat.vs_alloc += alloc_delta;
2795                 rvd->vdev_stat.vs_space += space_delta;
2796                 rvd->vdev_stat.vs_dspace += dspace_delta;
2797                 mutex_exit(&rvd->vdev_stat_lock);
2798         }
2799 
2800         if (mc != NULL) {
2801                 ASSERT(rvd == vd->vdev_parent);
2802                 ASSERT(vd->vdev_ms_count != 0);
2803 
2804                 metaslab_class_space_update(mc,
2805                     alloc_delta, defer_delta, space_delta, dspace_delta);
2806         }
2807 }
2808 
2809 /*
2810  * Mark a top-level vdev's config as dirty, placing it on the dirty list
2811  * so that it will be written out next time the vdev configuration is synced.
2812  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2813  */
2814 void
2815 vdev_config_dirty(vdev_t *vd)
2816 {
2817         spa_t *spa = vd->vdev_spa;
2818         vdev_t *rvd = spa->spa_root_vdev;
2819         int c;
2820 
2821         ASSERT(spa_writeable(spa));
2822 
2823         /*
2824          * If this is an aux vdev (as with l2cache and spare devices), then we
2825          * update the vdev config manually and set the sync flag.
2826          */
2827         if (vd->vdev_aux != NULL) {
2828                 spa_aux_vdev_t *sav = vd->vdev_aux;
2829                 nvlist_t **aux;
2830                 uint_t naux;
2831 
2832                 for (c = 0; c < sav->sav_count; c++) {
2833                         if (sav->sav_vdevs[c] == vd)
2834                                 break;
2835                 }
2836 
2837                 if (c == sav->sav_count) {
2838                         /*
2839                          * We're being removed.  There's nothing more to do.
2840                          */
2841                         ASSERT(sav->sav_sync == B_TRUE);
2842                         return;
2843                 }
2844 
2845                 sav->sav_sync = B_TRUE;
2846 
2847                 if (nvlist_lookup_nvlist_array(sav->sav_config,
2848                     ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
2849                         VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
2850                             ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
2851                 }
2852 
2853                 ASSERT(c < naux);
2854 
2855                 /*
2856                  * Setting the nvlist in the middle if the array is a little
2857                  * sketchy, but it will work.
2858                  */
2859                 nvlist_free(aux[c]);
2860                 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
2861 
2862                 return;
2863         }
2864 
2865         /*
2866          * The dirty list is protected by the SCL_CONFIG lock.  The caller
2867          * must either hold SCL_CONFIG as writer, or must be the sync thread
2868          * (which holds SCL_CONFIG as reader).  There's only one sync thread,
2869          * so this is sufficient to ensure mutual exclusion.
2870          */
2871         ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2872             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2873             spa_config_held(spa, SCL_CONFIG, RW_READER)));
2874 
2875         if (vd == rvd) {
2876                 for (c = 0; c < rvd->vdev_children; c++)
2877                         vdev_config_dirty(rvd->vdev_child[c]);
2878         } else {
2879                 ASSERT(vd == vd->vdev_top);
2880 
2881                 if (!list_link_active(&vd->vdev_config_dirty_node) &&
2882                     !vd->vdev_ishole)
2883                         list_insert_head(&spa->spa_config_dirty_list, vd);
2884         }
2885 }
2886 
2887 void
2888 vdev_config_clean(vdev_t *vd)
2889 {
2890         spa_t *spa = vd->vdev_spa;
2891 
2892         ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
2893             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2894             spa_config_held(spa, SCL_CONFIG, RW_READER)));
2895 
2896         ASSERT(list_link_active(&vd->vdev_config_dirty_node));
2897         list_remove(&spa->spa_config_dirty_list, vd);
2898 }
2899 
2900 /*
2901  * Mark a top-level vdev's state as dirty, so that the next pass of
2902  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
2903  * the state changes from larger config changes because they require
2904  * much less locking, and are often needed for administrative actions.
2905  */
2906 void
2907 vdev_state_dirty(vdev_t *vd)
2908 {
2909         spa_t *spa = vd->vdev_spa;
2910 
2911         ASSERT(spa_writeable(spa));
2912         ASSERT(vd == vd->vdev_top);
2913 
2914         /*
2915          * The state list is protected by the SCL_STATE lock.  The caller
2916          * must either hold SCL_STATE as writer, or must be the sync thread
2917          * (which holds SCL_STATE as reader).  There's only one sync thread,
2918          * so this is sufficient to ensure mutual exclusion.
2919          */
2920         ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2921             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2922             spa_config_held(spa, SCL_STATE, RW_READER)));
2923 
2924         if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
2925                 list_insert_head(&spa->spa_state_dirty_list, vd);
2926 }
2927 
2928 void
2929 vdev_state_clean(vdev_t *vd)
2930 {
2931         spa_t *spa = vd->vdev_spa;
2932 
2933         ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
2934             (dsl_pool_sync_context(spa_get_dsl(spa)) &&
2935             spa_config_held(spa, SCL_STATE, RW_READER)));
2936 
2937         ASSERT(list_link_active(&vd->vdev_state_dirty_node));
2938         list_remove(&spa->spa_state_dirty_list, vd);
2939 }
2940 
2941 /*
2942  * Propagate vdev state up from children to parent.
2943  */
2944 void
2945 vdev_propagate_state(vdev_t *vd)
2946 {
2947         spa_t *spa = vd->vdev_spa;
2948         vdev_t *rvd = spa->spa_root_vdev;
2949         int degraded = 0, faulted = 0;
2950         int corrupted = 0;
2951         vdev_t *child;
2952 
2953         if (vd->vdev_children > 0) {
2954                 for (int c = 0; c < vd->vdev_children; c++) {
2955                         child = vd->vdev_child[c];
2956 
2957                         /*
2958                          * Don't factor holes into the decision.
2959                          */
2960                         if (child->vdev_ishole)
2961                                 continue;
2962 
2963                         if (!vdev_readable(child) ||
2964                             (!vdev_writeable(child) && spa_writeable(spa))) {
2965                                 /*
2966                                  * Root special: if there is a top-level log
2967                                  * device, treat the root vdev as if it were
2968                                  * degraded.
2969                                  */
2970                                 if (child->vdev_islog && vd == rvd)
2971                                         degraded++;
2972                                 else
2973                                         faulted++;
2974                         } else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
2975                                 degraded++;
2976                         }
2977 
2978                         if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
2979                                 corrupted++;
2980                 }
2981 
2982                 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
2983 
2984                 /*
2985                  * Root special: if there is a top-level vdev that cannot be
2986                  * opened due to corrupted metadata, then propagate the root
2987                  * vdev's aux state as 'corrupt' rather than 'insufficient
2988                  * replicas'.
2989                  */
2990                 if (corrupted && vd == rvd &&
2991                     rvd->vdev_state == VDEV_STATE_CANT_OPEN)
2992                         vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
2993                             VDEV_AUX_CORRUPT_DATA);
2994         }
2995 
2996         if (vd->vdev_parent)
2997                 vdev_propagate_state(vd->vdev_parent);
2998 }
2999 
3000 /*
3001  * Set a vdev's state.  If this is during an open, we don't update the parent
3002  * state, because we're in the process of opening children depth-first.
3003  * Otherwise, we propagate the change to the parent.
3004  *
3005  * If this routine places a device in a faulted state, an appropriate ereport is
3006  * generated.
3007  */
3008 void
3009 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3010 {
3011         uint64_t save_state;
3012         spa_t *spa = vd->vdev_spa;
3013 
3014         if (state == vd->vdev_state) {
3015                 vd->vdev_stat.vs_aux = aux;
3016                 return;
3017         }
3018 
3019         save_state = vd->vdev_state;
3020 
3021         vd->vdev_state = state;
3022         vd->vdev_stat.vs_aux = aux;
3023 
3024         /*
3025          * If we are setting the vdev state to anything but an open state, then
3026          * always close the underlying device unless the device has requested
3027          * a delayed close (i.e. we're about to remove or fault the device).
3028          * Otherwise, we keep accessible but invalid devices open forever.
3029          * We don't call vdev_close() itself, because that implies some extra
3030          * checks (offline, etc) that we don't want here.  This is limited to
3031          * leaf devices, because otherwise closing the device will affect other
3032          * children.
3033          */
3034         if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3035             vd->vdev_ops->vdev_op_leaf)
3036                 vd->vdev_ops->vdev_op_close(vd);
3037 
3038         /*
3039          * If we have brought this vdev back into service, we need
3040          * to notify fmd so that it can gracefully repair any outstanding
3041          * cases due to a missing device.  We do this in all cases, even those
3042          * that probably don't correlate to a repaired fault.  This is sure to
3043          * catch all cases, and we let the zfs-retire agent sort it out.  If
3044          * this is a transient state it's OK, as the retire agent will
3045          * double-check the state of the vdev before repairing it.
3046          */
3047         if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3048             vd->vdev_prevstate != state)
3049                 zfs_post_state_change(spa, vd);
3050 
3051         if (vd->vdev_removed &&
3052             state == VDEV_STATE_CANT_OPEN &&
3053             (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3054                 /*
3055                  * If the previous state is set to VDEV_STATE_REMOVED, then this
3056                  * device was previously marked removed and someone attempted to
3057                  * reopen it.  If this failed due to a nonexistent device, then
3058                  * keep the device in the REMOVED state.  We also let this be if
3059                  * it is one of our special test online cases, which is only
3060                  * attempting to online the device and shouldn't generate an FMA
3061                  * fault.
3062                  */
3063                 vd->vdev_state = VDEV_STATE_REMOVED;
3064                 vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3065         } else if (state == VDEV_STATE_REMOVED) {
3066                 vd->vdev_removed = B_TRUE;
3067         } else if (state == VDEV_STATE_CANT_OPEN) {
3068                 /*
3069                  * If we fail to open a vdev during an import or recovery, we
3070                  * mark it as "not available", which signifies that it was
3071                  * never there to begin with.  Failure to open such a device
3072                  * is not considered an error.
3073                  */
3074                 if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3075                     spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3076                     vd->vdev_ops->vdev_op_leaf)
3077                         vd->vdev_not_present = 1;
3078 
3079                 /*
3080                  * Post the appropriate ereport.  If the 'prevstate' field is
3081                  * set to something other than VDEV_STATE_UNKNOWN, it indicates
3082                  * that this is part of a vdev_reopen().  In this case, we don't
3083                  * want to post the ereport if the device was already in the
3084                  * CANT_OPEN state beforehand.
3085                  *
3086                  * If the 'checkremove' flag is set, then this is an attempt to
3087                  * online the device in response to an insertion event.  If we
3088                  * hit this case, then we have detected an insertion event for a
3089                  * faulted or offline device that wasn't in the removed state.
3090                  * In this scenario, we don't post an ereport because we are
3091                  * about to replace the device, or attempt an online with
3092                  * vdev_forcefault, which will generate the fault for us.
3093                  */
3094                 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3095                     !vd->vdev_not_present && !vd->vdev_checkremove &&
3096                     vd != spa->spa_root_vdev) {
3097                         const char *class;
3098 
3099                         switch (aux) {
3100                         case VDEV_AUX_OPEN_FAILED:
3101                                 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3102                                 break;
3103                         case VDEV_AUX_CORRUPT_DATA:
3104                                 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3105                                 break;
3106                         case VDEV_AUX_NO_REPLICAS:
3107                                 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3108                                 break;
3109                         case VDEV_AUX_BAD_GUID_SUM:
3110                                 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3111                                 break;
3112                         case VDEV_AUX_TOO_SMALL:
3113                                 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3114                                 break;
3115                         case VDEV_AUX_BAD_LABEL:
3116                                 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3117                                 break;
3118                         default:
3119                                 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3120                         }
3121 
3122                         zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3123                 }
3124 
3125                 /* Erase any notion of persistent removed state */
3126                 vd->vdev_removed = B_FALSE;
3127         } else {
3128                 vd->vdev_removed = B_FALSE;
3129         }
3130 
3131         if (!isopen && vd->vdev_parent)
3132                 vdev_propagate_state(vd->vdev_parent);
3133 }
3134 
3135 /*
3136  * Check the vdev configuration to ensure that it's capable of supporting
3137  * a root pool. Currently, we do not support RAID-Z or partial configuration.
3138  * In addition, only a single top-level vdev is allowed and none of the leaves
3139  * can be wholedisks.
3140  */
3141 boolean_t
3142 vdev_is_bootable(vdev_t *vd)
3143 {
3144         if (!vd->vdev_ops->vdev_op_leaf) {
3145                 char *vdev_type = vd->vdev_ops->vdev_op_type;
3146 
3147                 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3148                     vd->vdev_children > 1) {
3149                         return (B_FALSE);
3150                 } else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3151                     strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3152                         return (B_FALSE);
3153                 }
3154         } else if (vd->vdev_wholedisk == 1) {
3155                 return (B_FALSE);
3156         }
3157 
3158         for (int c = 0; c < vd->vdev_children; c++) {
3159                 if (!vdev_is_bootable(vd->vdev_child[c]))
3160                         return (B_FALSE);
3161         }
3162         return (B_TRUE);
3163 }
3164 
3165 /*
3166  * Load the state from the original vdev tree (ovd) which
3167  * we've retrieved from the MOS config object. If the original
3168  * vdev was offline or faulted then we transfer that state to the
3169  * device in the current vdev tree (nvd).
3170  */
3171 void
3172 vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3173 {
3174         spa_t *spa = nvd->vdev_spa;
3175 
3176         ASSERT(nvd->vdev_top->vdev_islog);
3177         ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3178         ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3179 
3180         for (int c = 0; c < nvd->vdev_children; c++)
3181                 vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3182 
3183         if (nvd->vdev_ops->vdev_op_leaf) {
3184                 /*
3185                  * Restore the persistent vdev state
3186                  */
3187                 nvd->vdev_offline = ovd->vdev_offline;
3188                 nvd->vdev_faulted = ovd->vdev_faulted;
3189                 nvd->vdev_degraded = ovd->vdev_degraded;
3190                 nvd->vdev_removed = ovd->vdev_removed;
3191         }
3192 }
3193 
3194 /*
3195  * Determine if a log device has valid content.  If the vdev was
3196  * removed or faulted in the MOS config then we know that
3197  * the content on the log device has already been written to the pool.
3198  */
3199 boolean_t
3200 vdev_log_state_valid(vdev_t *vd)
3201 {
3202         if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3203             !vd->vdev_removed)
3204                 return (B_TRUE);
3205 
3206         for (int c = 0; c < vd->vdev_children; c++)
3207                 if (vdev_log_state_valid(vd->vdev_child[c]))
3208                         return (B_TRUE);
3209 
3210         return (B_FALSE);
3211 }
3212 
3213 /*
3214  * Expand a vdev if possible.
3215  */
3216 void
3217 vdev_expand(vdev_t *vd, uint64_t txg)
3218 {
3219         ASSERT(vd->vdev_top == vd);
3220         ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3221 
3222         if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3223                 VERIFY(vdev_metaslab_init(vd, txg) == 0);
3224                 vdev_config_dirty(vd);
3225         }
3226 }
3227 
3228 /*
3229  * Split a vdev.
3230  */
3231 void
3232 vdev_split(vdev_t *vd)
3233 {
3234         vdev_t *cvd, *pvd = vd->vdev_parent;
3235 
3236         vdev_remove_child(pvd, vd);
3237         vdev_compact_children(pvd);
3238 
3239         cvd = pvd->vdev_child[0];
3240         if (pvd->vdev_children == 1) {
3241                 vdev_remove_parent(cvd);
3242                 cvd->vdev_splitting = B_TRUE;
3243         }
3244         vdev_propagate_state(cvd);
3245 }
3246 
3247 void
3248 vdev_deadman(vdev_t *vd)
3249 {
3250         for (int c = 0; c < vd->vdev_children; c++) {
3251                 vdev_t *cvd = vd->vdev_child[c];
3252 
3253                 vdev_deadman(cvd);
3254         }
3255 
3256         if (vd->vdev_ops->vdev_op_leaf) {
3257                 vdev_queue_t *vq = &vd->vdev_queue;
3258 
3259                 mutex_enter(&vq->vq_lock);
3260                 if (avl_numnodes(&vq->vq_pending_tree) > 0) {
3261                         spa_t *spa = vd->vdev_spa;
3262                         zio_t *fio;
3263                         uint64_t delta;
3264 
3265                         /*
3266                          * Look at the head of all the pending queues,
3267                          * if any I/O has been outstanding for longer than
3268                          * the spa_deadman_synctime we panic the system.
3269                          */
3270                         fio = avl_first(&vq->vq_pending_tree);
3271                         delta = gethrtime() - fio->io_timestamp;
3272                         if (delta > spa_deadman_synctime(spa)) {
3273                                 zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3274                                     "delta %lluns, last io %lluns",
3275                                     fio->io_timestamp, delta,
3276                                     vq->vq_io_complete_ts);
3277                                 fm_panic("I/O to pool '%s' appears to be "
3278                                     "hung.", spa_name(spa));
3279                         }
3280                 }
3281                 mutex_exit(&vq->vq_lock);
3282         }
3283 }