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 (c) 2011, 2018 by Delphix. All rights reserved. 25 * Copyright 2017 Nexenta Systems, Inc. 26 * Copyright (c) 2014 Integros [integros.com] 27 * Copyright 2016 Toomas Soome <tsoome@me.com> 28 * Copyright 2017 Joyent, Inc. 29 */ 30 31 #include <sys/zfs_context.h> 32 #include <sys/fm/fs/zfs.h> 33 #include <sys/spa.h> 34 #include <sys/spa_impl.h> 35 #include <sys/bpobj.h> 36 #include <sys/dmu.h> 37 #include <sys/dmu_tx.h> 38 #include <sys/dsl_dir.h> 39 #include <sys/vdev_impl.h> 40 #include <sys/uberblock_impl.h> 41 #include <sys/metaslab.h> 42 #include <sys/metaslab_impl.h> 43 #include <sys/space_map.h> 44 #include <sys/space_reftree.h> 45 #include <sys/zio.h> 46 #include <sys/zap.h> 47 #include <sys/fs/zfs.h> 48 #include <sys/arc.h> 49 #include <sys/zil.h> 50 #include <sys/dsl_scan.h> 51 #include <sys/abd.h> 52 53 /* 54 * Virtual device management. 55 */ 56 57 static vdev_ops_t *vdev_ops_table[] = { 58 &vdev_root_ops, 59 &vdev_raidz_ops, 60 &vdev_mirror_ops, 61 &vdev_replacing_ops, 62 &vdev_spare_ops, 63 &vdev_disk_ops, 64 &vdev_file_ops, 65 &vdev_missing_ops, 66 &vdev_hole_ops, 67 &vdev_indirect_ops, 68 NULL 69 }; 70 71 /* maximum scrub/resilver I/O queue per leaf vdev */ 72 int zfs_scrub_limit = 10; 73 74 /* maximum number of metaslabs per top-level vdev */ 75 int vdev_max_ms_count = 200; 76 77 /* minimum amount of metaslabs per top-level vdev */ 78 int vdev_min_ms_count = 16; 79 80 /* see comment in vdev_metaslab_set_size() */ 81 int vdev_default_ms_shift = 29; 82 83 boolean_t vdev_validate_skip = B_FALSE; 84 85 /* 86 * Since the DTL space map of a vdev is not expected to have a lot of 87 * entries, we default its block size to 4K. 88 */ 89 int vdev_dtl_sm_blksz = (1 << 12); 90 91 /* 92 * vdev-wide space maps that have lots of entries written to them at 93 * the end of each transaction can benefit from a higher I/O bandwidth 94 * (e.g. vdev_obsolete_sm), thus we default their block size to 128K. 95 */ 96 int vdev_standard_sm_blksz = (1 << 17); 97 98 /*PRINTFLIKE2*/ 99 void 100 vdev_dbgmsg(vdev_t *vd, const char *fmt, ...) 101 { 102 va_list adx; 103 char buf[256]; 104 105 va_start(adx, fmt); 106 (void) vsnprintf(buf, sizeof (buf), fmt, adx); 107 va_end(adx); 108 109 if (vd->vdev_path != NULL) { 110 zfs_dbgmsg("%s vdev '%s': %s", vd->vdev_ops->vdev_op_type, 111 vd->vdev_path, buf); 112 } else { 113 zfs_dbgmsg("%s-%llu vdev (guid %llu): %s", 114 vd->vdev_ops->vdev_op_type, 115 (u_longlong_t)vd->vdev_id, 116 (u_longlong_t)vd->vdev_guid, buf); 117 } 118 } 119 120 void 121 vdev_dbgmsg_print_tree(vdev_t *vd, int indent) 122 { 123 char state[20]; 124 125 if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) { 126 zfs_dbgmsg("%*svdev %u: %s", indent, "", vd->vdev_id, 127 vd->vdev_ops->vdev_op_type); 128 return; 129 } 130 131 switch (vd->vdev_state) { 132 case VDEV_STATE_UNKNOWN: 133 (void) snprintf(state, sizeof (state), "unknown"); 134 break; 135 case VDEV_STATE_CLOSED: 136 (void) snprintf(state, sizeof (state), "closed"); 137 break; 138 case VDEV_STATE_OFFLINE: 139 (void) snprintf(state, sizeof (state), "offline"); 140 break; 141 case VDEV_STATE_REMOVED: 142 (void) snprintf(state, sizeof (state), "removed"); 143 break; 144 case VDEV_STATE_CANT_OPEN: 145 (void) snprintf(state, sizeof (state), "can't open"); 146 break; 147 case VDEV_STATE_FAULTED: 148 (void) snprintf(state, sizeof (state), "faulted"); 149 break; 150 case VDEV_STATE_DEGRADED: 151 (void) snprintf(state, sizeof (state), "degraded"); 152 break; 153 case VDEV_STATE_HEALTHY: 154 (void) snprintf(state, sizeof (state), "healthy"); 155 break; 156 default: 157 (void) snprintf(state, sizeof (state), "<state %u>", 158 (uint_t)vd->vdev_state); 159 } 160 161 zfs_dbgmsg("%*svdev %u: %s%s, guid: %llu, path: %s, %s", indent, 162 "", vd->vdev_id, vd->vdev_ops->vdev_op_type, 163 vd->vdev_islog ? " (log)" : "", 164 (u_longlong_t)vd->vdev_guid, 165 vd->vdev_path ? vd->vdev_path : "N/A", state); 166 167 for (uint64_t i = 0; i < vd->vdev_children; i++) 168 vdev_dbgmsg_print_tree(vd->vdev_child[i], indent + 2); 169 } 170 171 /* 172 * Given a vdev type, return the appropriate ops vector. 173 */ 174 static vdev_ops_t * 175 vdev_getops(const char *type) 176 { 177 vdev_ops_t *ops, **opspp; 178 179 for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++) 180 if (strcmp(ops->vdev_op_type, type) == 0) 181 break; 182 183 return (ops); 184 } 185 186 /* 187 * Default asize function: return the MAX of psize with the asize of 188 * all children. This is what's used by anything other than RAID-Z. 189 */ 190 uint64_t 191 vdev_default_asize(vdev_t *vd, uint64_t psize) 192 { 193 uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift); 194 uint64_t csize; 195 196 for (int c = 0; c < vd->vdev_children; c++) { 197 csize = vdev_psize_to_asize(vd->vdev_child[c], psize); 198 asize = MAX(asize, csize); 199 } 200 201 return (asize); 202 } 203 204 /* 205 * Get the minimum allocatable size. We define the allocatable size as 206 * the vdev's asize rounded to the nearest metaslab. This allows us to 207 * replace or attach devices which don't have the same physical size but 208 * can still satisfy the same number of allocations. 209 */ 210 uint64_t 211 vdev_get_min_asize(vdev_t *vd) 212 { 213 vdev_t *pvd = vd->vdev_parent; 214 215 /* 216 * If our parent is NULL (inactive spare or cache) or is the root, 217 * just return our own asize. 218 */ 219 if (pvd == NULL) 220 return (vd->vdev_asize); 221 222 /* 223 * The top-level vdev just returns the allocatable size rounded 224 * to the nearest metaslab. 225 */ 226 if (vd == vd->vdev_top) 227 return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift)); 228 229 /* 230 * The allocatable space for a raidz vdev is N * sizeof(smallest child), 231 * so each child must provide at least 1/Nth of its asize. 232 */ 233 if (pvd->vdev_ops == &vdev_raidz_ops) 234 return ((pvd->vdev_min_asize + pvd->vdev_children - 1) / 235 pvd->vdev_children); 236 237 return (pvd->vdev_min_asize); 238 } 239 240 void 241 vdev_set_min_asize(vdev_t *vd) 242 { 243 vd->vdev_min_asize = vdev_get_min_asize(vd); 244 245 for (int c = 0; c < vd->vdev_children; c++) 246 vdev_set_min_asize(vd->vdev_child[c]); 247 } 248 249 vdev_t * 250 vdev_lookup_top(spa_t *spa, uint64_t vdev) 251 { 252 vdev_t *rvd = spa->spa_root_vdev; 253 254 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0); 255 256 if (vdev < rvd->vdev_children) { 257 ASSERT(rvd->vdev_child[vdev] != NULL); 258 return (rvd->vdev_child[vdev]); 259 } 260 261 return (NULL); 262 } 263 264 vdev_t * 265 vdev_lookup_by_guid(vdev_t *vd, uint64_t guid) 266 { 267 vdev_t *mvd; 268 269 if (vd->vdev_guid == guid) 270 return (vd); 271 272 for (int c = 0; c < vd->vdev_children; c++) 273 if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) != 274 NULL) 275 return (mvd); 276 277 return (NULL); 278 } 279 280 static int 281 vdev_count_leaves_impl(vdev_t *vd) 282 { 283 int n = 0; 284 285 if (vd->vdev_ops->vdev_op_leaf) 286 return (1); 287 288 for (int c = 0; c < vd->vdev_children; c++) 289 n += vdev_count_leaves_impl(vd->vdev_child[c]); 290 291 return (n); 292 } 293 294 int 295 vdev_count_leaves(spa_t *spa) 296 { 297 return (vdev_count_leaves_impl(spa->spa_root_vdev)); 298 } 299 300 void 301 vdev_add_child(vdev_t *pvd, vdev_t *cvd) 302 { 303 size_t oldsize, newsize; 304 uint64_t id = cvd->vdev_id; 305 vdev_t **newchild; 306 spa_t *spa = cvd->vdev_spa; 307 308 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 309 ASSERT(cvd->vdev_parent == NULL); 310 311 cvd->vdev_parent = pvd; 312 313 if (pvd == NULL) 314 return; 315 316 ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL); 317 318 oldsize = pvd->vdev_children * sizeof (vdev_t *); 319 pvd->vdev_children = MAX(pvd->vdev_children, id + 1); 320 newsize = pvd->vdev_children * sizeof (vdev_t *); 321 322 newchild = kmem_zalloc(newsize, KM_SLEEP); 323 if (pvd->vdev_child != NULL) { 324 bcopy(pvd->vdev_child, newchild, oldsize); 325 kmem_free(pvd->vdev_child, oldsize); 326 } 327 328 pvd->vdev_child = newchild; 329 pvd->vdev_child[id] = cvd; 330 331 cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd); 332 ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL); 333 334 /* 335 * Walk up all ancestors to update guid sum. 336 */ 337 for (; pvd != NULL; pvd = pvd->vdev_parent) 338 pvd->vdev_guid_sum += cvd->vdev_guid_sum; 339 } 340 341 void 342 vdev_remove_child(vdev_t *pvd, vdev_t *cvd) 343 { 344 int c; 345 uint_t id = cvd->vdev_id; 346 347 ASSERT(cvd->vdev_parent == pvd); 348 349 if (pvd == NULL) 350 return; 351 352 ASSERT(id < pvd->vdev_children); 353 ASSERT(pvd->vdev_child[id] == cvd); 354 355 pvd->vdev_child[id] = NULL; 356 cvd->vdev_parent = NULL; 357 358 for (c = 0; c < pvd->vdev_children; c++) 359 if (pvd->vdev_child[c]) 360 break; 361 362 if (c == pvd->vdev_children) { 363 kmem_free(pvd->vdev_child, c * sizeof (vdev_t *)); 364 pvd->vdev_child = NULL; 365 pvd->vdev_children = 0; 366 } 367 368 /* 369 * Walk up all ancestors to update guid sum. 370 */ 371 for (; pvd != NULL; pvd = pvd->vdev_parent) 372 pvd->vdev_guid_sum -= cvd->vdev_guid_sum; 373 } 374 375 /* 376 * Remove any holes in the child array. 377 */ 378 void 379 vdev_compact_children(vdev_t *pvd) 380 { 381 vdev_t **newchild, *cvd; 382 int oldc = pvd->vdev_children; 383 int newc; 384 385 ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL); 386 387 for (int c = newc = 0; c < oldc; c++) 388 if (pvd->vdev_child[c]) 389 newc++; 390 391 newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP); 392 393 for (int c = newc = 0; c < oldc; c++) { 394 if ((cvd = pvd->vdev_child[c]) != NULL) { 395 newchild[newc] = cvd; 396 cvd->vdev_id = newc++; 397 } 398 } 399 400 kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *)); 401 pvd->vdev_child = newchild; 402 pvd->vdev_children = newc; 403 } 404 405 /* 406 * Allocate and minimally initialize a vdev_t. 407 */ 408 vdev_t * 409 vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops) 410 { 411 vdev_t *vd; 412 vdev_indirect_config_t *vic; 413 414 vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP); 415 vic = &vd->vdev_indirect_config; 416 417 if (spa->spa_root_vdev == NULL) { 418 ASSERT(ops == &vdev_root_ops); 419 spa->spa_root_vdev = vd; 420 spa->spa_load_guid = spa_generate_guid(NULL); 421 } 422 423 if (guid == 0 && ops != &vdev_hole_ops) { 424 if (spa->spa_root_vdev == vd) { 425 /* 426 * The root vdev's guid will also be the pool guid, 427 * which must be unique among all pools. 428 */ 429 guid = spa_generate_guid(NULL); 430 } else { 431 /* 432 * Any other vdev's guid must be unique within the pool. 433 */ 434 guid = spa_generate_guid(spa); 435 } 436 ASSERT(!spa_guid_exists(spa_guid(spa), guid)); 437 } 438 439 vd->vdev_spa = spa; 440 vd->vdev_id = id; 441 vd->vdev_guid = guid; 442 vd->vdev_guid_sum = guid; 443 vd->vdev_ops = ops; 444 vd->vdev_state = VDEV_STATE_CLOSED; 445 vd->vdev_ishole = (ops == &vdev_hole_ops); 446 vic->vic_prev_indirect_vdev = UINT64_MAX; 447 448 rw_init(&vd->vdev_indirect_rwlock, NULL, RW_DEFAULT, NULL); 449 mutex_init(&vd->vdev_obsolete_lock, NULL, MUTEX_DEFAULT, NULL); 450 vd->vdev_obsolete_segments = range_tree_create(NULL, NULL); 451 452 mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL); 453 mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL); 454 mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL); 455 mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL); 456 for (int t = 0; t < DTL_TYPES; t++) { 457 vd->vdev_dtl[t] = range_tree_create(NULL, NULL); 458 } 459 txg_list_create(&vd->vdev_ms_list, spa, 460 offsetof(struct metaslab, ms_txg_node)); 461 txg_list_create(&vd->vdev_dtl_list, spa, 462 offsetof(struct vdev, vdev_dtl_node)); 463 vd->vdev_stat.vs_timestamp = gethrtime(); 464 vdev_queue_init(vd); 465 vdev_cache_init(vd); 466 467 return (vd); 468 } 469 470 /* 471 * Allocate a new vdev. The 'alloctype' is used to control whether we are 472 * creating a new vdev or loading an existing one - the behavior is slightly 473 * different for each case. 474 */ 475 int 476 vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id, 477 int alloctype) 478 { 479 vdev_ops_t *ops; 480 char *type; 481 uint64_t guid = 0, islog, nparity; 482 vdev_t *vd; 483 vdev_indirect_config_t *vic; 484 485 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 486 487 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0) 488 return (SET_ERROR(EINVAL)); 489 490 if ((ops = vdev_getops(type)) == NULL) 491 return (SET_ERROR(EINVAL)); 492 493 /* 494 * If this is a load, get the vdev guid from the nvlist. 495 * Otherwise, vdev_alloc_common() will generate one for us. 496 */ 497 if (alloctype == VDEV_ALLOC_LOAD) { 498 uint64_t label_id; 499 500 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) || 501 label_id != id) 502 return (SET_ERROR(EINVAL)); 503 504 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0) 505 return (SET_ERROR(EINVAL)); 506 } else if (alloctype == VDEV_ALLOC_SPARE) { 507 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0) 508 return (SET_ERROR(EINVAL)); 509 } else if (alloctype == VDEV_ALLOC_L2CACHE) { 510 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0) 511 return (SET_ERROR(EINVAL)); 512 } else if (alloctype == VDEV_ALLOC_ROOTPOOL) { 513 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0) 514 return (SET_ERROR(EINVAL)); 515 } 516 517 /* 518 * The first allocated vdev must be of type 'root'. 519 */ 520 if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL) 521 return (SET_ERROR(EINVAL)); 522 523 /* 524 * Determine whether we're a log vdev. 525 */ 526 islog = 0; 527 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog); 528 if (islog && spa_version(spa) < SPA_VERSION_SLOGS) 529 return (SET_ERROR(ENOTSUP)); 530 531 if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES) 532 return (SET_ERROR(ENOTSUP)); 533 534 /* 535 * Set the nparity property for RAID-Z vdevs. 536 */ 537 nparity = -1ULL; 538 if (ops == &vdev_raidz_ops) { 539 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY, 540 &nparity) == 0) { 541 if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY) 542 return (SET_ERROR(EINVAL)); 543 /* 544 * Previous versions could only support 1 or 2 parity 545 * device. 546 */ 547 if (nparity > 1 && 548 spa_version(spa) < SPA_VERSION_RAIDZ2) 549 return (SET_ERROR(ENOTSUP)); 550 if (nparity > 2 && 551 spa_version(spa) < SPA_VERSION_RAIDZ3) 552 return (SET_ERROR(ENOTSUP)); 553 } else { 554 /* 555 * We require the parity to be specified for SPAs that 556 * support multiple parity levels. 557 */ 558 if (spa_version(spa) >= SPA_VERSION_RAIDZ2) 559 return (SET_ERROR(EINVAL)); 560 /* 561 * Otherwise, we default to 1 parity device for RAID-Z. 562 */ 563 nparity = 1; 564 } 565 } else { 566 nparity = 0; 567 } 568 ASSERT(nparity != -1ULL); 569 570 vd = vdev_alloc_common(spa, id, guid, ops); 571 vic = &vd->vdev_indirect_config; 572 573 vd->vdev_islog = islog; 574 vd->vdev_nparity = nparity; 575 576 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0) 577 vd->vdev_path = spa_strdup(vd->vdev_path); 578 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0) 579 vd->vdev_devid = spa_strdup(vd->vdev_devid); 580 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH, 581 &vd->vdev_physpath) == 0) 582 vd->vdev_physpath = spa_strdup(vd->vdev_physpath); 583 if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0) 584 vd->vdev_fru = spa_strdup(vd->vdev_fru); 585 586 /* 587 * Set the whole_disk property. If it's not specified, leave the value 588 * as -1. 589 */ 590 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK, 591 &vd->vdev_wholedisk) != 0) 592 vd->vdev_wholedisk = -1ULL; 593 594 ASSERT0(vic->vic_mapping_object); 595 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT, 596 &vic->vic_mapping_object); 597 ASSERT0(vic->vic_births_object); 598 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS, 599 &vic->vic_births_object); 600 ASSERT3U(vic->vic_prev_indirect_vdev, ==, UINT64_MAX); 601 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV, 602 &vic->vic_prev_indirect_vdev); 603 604 /* 605 * Look for the 'not present' flag. This will only be set if the device 606 * was not present at the time of import. 607 */ 608 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT, 609 &vd->vdev_not_present); 610 611 /* 612 * Get the alignment requirement. 613 */ 614 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift); 615 616 /* 617 * Retrieve the vdev creation time. 618 */ 619 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG, 620 &vd->vdev_crtxg); 621 622 /* 623 * If we're a top-level vdev, try to load the allocation parameters. 624 */ 625 if (parent && !parent->vdev_parent && 626 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) { 627 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY, 628 &vd->vdev_ms_array); 629 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT, 630 &vd->vdev_ms_shift); 631 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE, 632 &vd->vdev_asize); 633 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING, 634 &vd->vdev_removing); 635 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP, 636 &vd->vdev_top_zap); 637 } else { 638 ASSERT0(vd->vdev_top_zap); 639 } 640 641 if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) { 642 ASSERT(alloctype == VDEV_ALLOC_LOAD || 643 alloctype == VDEV_ALLOC_ADD || 644 alloctype == VDEV_ALLOC_SPLIT || 645 alloctype == VDEV_ALLOC_ROOTPOOL); 646 vd->vdev_mg = metaslab_group_create(islog ? 647 spa_log_class(spa) : spa_normal_class(spa), vd, 648 spa->spa_alloc_count); 649 } 650 651 if (vd->vdev_ops->vdev_op_leaf && 652 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) { 653 (void) nvlist_lookup_uint64(nv, 654 ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap); 655 } else { 656 ASSERT0(vd->vdev_leaf_zap); 657 } 658 659 /* 660 * If we're a leaf vdev, try to load the DTL object and other state. 661 */ 662 663 if (vd->vdev_ops->vdev_op_leaf && 664 (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE || 665 alloctype == VDEV_ALLOC_ROOTPOOL)) { 666 if (alloctype == VDEV_ALLOC_LOAD) { 667 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL, 668 &vd->vdev_dtl_object); 669 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE, 670 &vd->vdev_unspare); 671 } 672 673 if (alloctype == VDEV_ALLOC_ROOTPOOL) { 674 uint64_t spare = 0; 675 676 if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE, 677 &spare) == 0 && spare) 678 spa_spare_add(vd); 679 } 680 681 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE, 682 &vd->vdev_offline); 683 684 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG, 685 &vd->vdev_resilver_txg); 686 687 /* 688 * When importing a pool, we want to ignore the persistent fault 689 * state, as the diagnosis made on another system may not be 690 * valid in the current context. Local vdevs will 691 * remain in the faulted state. 692 */ 693 if (spa_load_state(spa) == SPA_LOAD_OPEN) { 694 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED, 695 &vd->vdev_faulted); 696 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED, 697 &vd->vdev_degraded); 698 (void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED, 699 &vd->vdev_removed); 700 701 if (vd->vdev_faulted || vd->vdev_degraded) { 702 char *aux; 703 704 vd->vdev_label_aux = 705 VDEV_AUX_ERR_EXCEEDED; 706 if (nvlist_lookup_string(nv, 707 ZPOOL_CONFIG_AUX_STATE, &aux) == 0 && 708 strcmp(aux, "external") == 0) 709 vd->vdev_label_aux = VDEV_AUX_EXTERNAL; 710 } 711 } 712 } 713 714 /* 715 * Add ourselves to the parent's list of children. 716 */ 717 vdev_add_child(parent, vd); 718 719 *vdp = vd; 720 721 return (0); 722 } 723 724 void 725 vdev_free(vdev_t *vd) 726 { 727 spa_t *spa = vd->vdev_spa; 728 729 /* 730 * vdev_free() implies closing the vdev first. This is simpler than 731 * trying to ensure complicated semantics for all callers. 732 */ 733 vdev_close(vd); 734 735 ASSERT(!list_link_active(&vd->vdev_config_dirty_node)); 736 ASSERT(!list_link_active(&vd->vdev_state_dirty_node)); 737 738 /* 739 * Free all children. 740 */ 741 for (int c = 0; c < vd->vdev_children; c++) 742 vdev_free(vd->vdev_child[c]); 743 744 ASSERT(vd->vdev_child == NULL); 745 ASSERT(vd->vdev_guid_sum == vd->vdev_guid); 746 747 /* 748 * Discard allocation state. 749 */ 750 if (vd->vdev_mg != NULL) { 751 vdev_metaslab_fini(vd); 752 metaslab_group_destroy(vd->vdev_mg); 753 } 754 755 ASSERT0(vd->vdev_stat.vs_space); 756 ASSERT0(vd->vdev_stat.vs_dspace); 757 ASSERT0(vd->vdev_stat.vs_alloc); 758 759 /* 760 * Remove this vdev from its parent's child list. 761 */ 762 vdev_remove_child(vd->vdev_parent, vd); 763 764 ASSERT(vd->vdev_parent == NULL); 765 766 /* 767 * Clean up vdev structure. 768 */ 769 vdev_queue_fini(vd); 770 vdev_cache_fini(vd); 771 772 if (vd->vdev_path) 773 spa_strfree(vd->vdev_path); 774 if (vd->vdev_devid) 775 spa_strfree(vd->vdev_devid); 776 if (vd->vdev_physpath) 777 spa_strfree(vd->vdev_physpath); 778 if (vd->vdev_fru) 779 spa_strfree(vd->vdev_fru); 780 781 if (vd->vdev_isspare) 782 spa_spare_remove(vd); 783 if (vd->vdev_isl2cache) 784 spa_l2cache_remove(vd); 785 786 txg_list_destroy(&vd->vdev_ms_list); 787 txg_list_destroy(&vd->vdev_dtl_list); 788 789 mutex_enter(&vd->vdev_dtl_lock); 790 space_map_close(vd->vdev_dtl_sm); 791 for (int t = 0; t < DTL_TYPES; t++) { 792 range_tree_vacate(vd->vdev_dtl[t], NULL, NULL); 793 range_tree_destroy(vd->vdev_dtl[t]); 794 } 795 mutex_exit(&vd->vdev_dtl_lock); 796 797 EQUIV(vd->vdev_indirect_births != NULL, 798 vd->vdev_indirect_mapping != NULL); 799 if (vd->vdev_indirect_births != NULL) { 800 vdev_indirect_mapping_close(vd->vdev_indirect_mapping); 801 vdev_indirect_births_close(vd->vdev_indirect_births); 802 } 803 804 if (vd->vdev_obsolete_sm != NULL) { 805 ASSERT(vd->vdev_removing || 806 vd->vdev_ops == &vdev_indirect_ops); 807 space_map_close(vd->vdev_obsolete_sm); 808 vd->vdev_obsolete_sm = NULL; 809 } 810 range_tree_destroy(vd->vdev_obsolete_segments); 811 rw_destroy(&vd->vdev_indirect_rwlock); 812 mutex_destroy(&vd->vdev_obsolete_lock); 813 814 mutex_destroy(&vd->vdev_queue_lock); 815 mutex_destroy(&vd->vdev_dtl_lock); 816 mutex_destroy(&vd->vdev_stat_lock); 817 mutex_destroy(&vd->vdev_probe_lock); 818 819 if (vd == spa->spa_root_vdev) 820 spa->spa_root_vdev = NULL; 821 822 kmem_free(vd, sizeof (vdev_t)); 823 } 824 825 /* 826 * Transfer top-level vdev state from svd to tvd. 827 */ 828 static void 829 vdev_top_transfer(vdev_t *svd, vdev_t *tvd) 830 { 831 spa_t *spa = svd->vdev_spa; 832 metaslab_t *msp; 833 vdev_t *vd; 834 int t; 835 836 ASSERT(tvd == tvd->vdev_top); 837 838 tvd->vdev_ms_array = svd->vdev_ms_array; 839 tvd->vdev_ms_shift = svd->vdev_ms_shift; 840 tvd->vdev_ms_count = svd->vdev_ms_count; 841 tvd->vdev_top_zap = svd->vdev_top_zap; 842 843 svd->vdev_ms_array = 0; 844 svd->vdev_ms_shift = 0; 845 svd->vdev_ms_count = 0; 846 svd->vdev_top_zap = 0; 847 848 if (tvd->vdev_mg) 849 ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg); 850 tvd->vdev_mg = svd->vdev_mg; 851 tvd->vdev_ms = svd->vdev_ms; 852 853 svd->vdev_mg = NULL; 854 svd->vdev_ms = NULL; 855 856 if (tvd->vdev_mg != NULL) 857 tvd->vdev_mg->mg_vd = tvd; 858 859 tvd->vdev_checkpoint_sm = svd->vdev_checkpoint_sm; 860 svd->vdev_checkpoint_sm = NULL; 861 862 tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc; 863 tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space; 864 tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace; 865 866 svd->vdev_stat.vs_alloc = 0; 867 svd->vdev_stat.vs_space = 0; 868 svd->vdev_stat.vs_dspace = 0; 869 870 /* 871 * State which may be set on a top-level vdev that's in the 872 * process of being removed. 873 */ 874 ASSERT0(tvd->vdev_indirect_config.vic_births_object); 875 ASSERT0(tvd->vdev_indirect_config.vic_mapping_object); 876 ASSERT3U(tvd->vdev_indirect_config.vic_prev_indirect_vdev, ==, -1ULL); 877 ASSERT3P(tvd->vdev_indirect_mapping, ==, NULL); 878 ASSERT3P(tvd->vdev_indirect_births, ==, NULL); 879 ASSERT3P(tvd->vdev_obsolete_sm, ==, NULL); 880 ASSERT0(tvd->vdev_removing); 881 tvd->vdev_removing = svd->vdev_removing; 882 tvd->vdev_indirect_config = svd->vdev_indirect_config; 883 tvd->vdev_indirect_mapping = svd->vdev_indirect_mapping; 884 tvd->vdev_indirect_births = svd->vdev_indirect_births; 885 range_tree_swap(&svd->vdev_obsolete_segments, 886 &tvd->vdev_obsolete_segments); 887 tvd->vdev_obsolete_sm = svd->vdev_obsolete_sm; 888 svd->vdev_indirect_config.vic_mapping_object = 0; 889 svd->vdev_indirect_config.vic_births_object = 0; 890 svd->vdev_indirect_config.vic_prev_indirect_vdev = -1ULL; 891 svd->vdev_indirect_mapping = NULL; 892 svd->vdev_indirect_births = NULL; 893 svd->vdev_obsolete_sm = NULL; 894 svd->vdev_removing = 0; 895 896 for (t = 0; t < TXG_SIZE; t++) { 897 while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL) 898 (void) txg_list_add(&tvd->vdev_ms_list, msp, t); 899 while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL) 900 (void) txg_list_add(&tvd->vdev_dtl_list, vd, t); 901 if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t)) 902 (void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t); 903 } 904 905 if (list_link_active(&svd->vdev_config_dirty_node)) { 906 vdev_config_clean(svd); 907 vdev_config_dirty(tvd); 908 } 909 910 if (list_link_active(&svd->vdev_state_dirty_node)) { 911 vdev_state_clean(svd); 912 vdev_state_dirty(tvd); 913 } 914 915 tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio; 916 svd->vdev_deflate_ratio = 0; 917 918 tvd->vdev_islog = svd->vdev_islog; 919 svd->vdev_islog = 0; 920 } 921 922 static void 923 vdev_top_update(vdev_t *tvd, vdev_t *vd) 924 { 925 if (vd == NULL) 926 return; 927 928 vd->vdev_top = tvd; 929 930 for (int c = 0; c < vd->vdev_children; c++) 931 vdev_top_update(tvd, vd->vdev_child[c]); 932 } 933 934 /* 935 * Add a mirror/replacing vdev above an existing vdev. 936 */ 937 vdev_t * 938 vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops) 939 { 940 spa_t *spa = cvd->vdev_spa; 941 vdev_t *pvd = cvd->vdev_parent; 942 vdev_t *mvd; 943 944 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 945 946 mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops); 947 948 mvd->vdev_asize = cvd->vdev_asize; 949 mvd->vdev_min_asize = cvd->vdev_min_asize; 950 mvd->vdev_max_asize = cvd->vdev_max_asize; 951 mvd->vdev_psize = cvd->vdev_psize; 952 mvd->vdev_ashift = cvd->vdev_ashift; 953 mvd->vdev_state = cvd->vdev_state; 954 mvd->vdev_crtxg = cvd->vdev_crtxg; 955 956 vdev_remove_child(pvd, cvd); 957 vdev_add_child(pvd, mvd); 958 cvd->vdev_id = mvd->vdev_children; 959 vdev_add_child(mvd, cvd); 960 vdev_top_update(cvd->vdev_top, cvd->vdev_top); 961 962 if (mvd == mvd->vdev_top) 963 vdev_top_transfer(cvd, mvd); 964 965 return (mvd); 966 } 967 968 /* 969 * Remove a 1-way mirror/replacing vdev from the tree. 970 */ 971 void 972 vdev_remove_parent(vdev_t *cvd) 973 { 974 vdev_t *mvd = cvd->vdev_parent; 975 vdev_t *pvd = mvd->vdev_parent; 976 977 ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL); 978 979 ASSERT(mvd->vdev_children == 1); 980 ASSERT(mvd->vdev_ops == &vdev_mirror_ops || 981 mvd->vdev_ops == &vdev_replacing_ops || 982 mvd->vdev_ops == &vdev_spare_ops); 983 cvd->vdev_ashift = mvd->vdev_ashift; 984 985 vdev_remove_child(mvd, cvd); 986 vdev_remove_child(pvd, mvd); 987 988 /* 989 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid. 990 * Otherwise, we could have detached an offline device, and when we 991 * go to import the pool we'll think we have two top-level vdevs, 992 * instead of a different version of the same top-level vdev. 993 */ 994 if (mvd->vdev_top == mvd) { 995 uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid; 996 cvd->vdev_orig_guid = cvd->vdev_guid; 997 cvd->vdev_guid += guid_delta; 998 cvd->vdev_guid_sum += guid_delta; 999 } 1000 cvd->vdev_id = mvd->vdev_id; 1001 vdev_add_child(pvd, cvd); 1002 vdev_top_update(cvd->vdev_top, cvd->vdev_top); 1003 1004 if (cvd == cvd->vdev_top) 1005 vdev_top_transfer(mvd, cvd); 1006 1007 ASSERT(mvd->vdev_children == 0); 1008 vdev_free(mvd); 1009 } 1010 1011 int 1012 vdev_metaslab_init(vdev_t *vd, uint64_t txg) 1013 { 1014 spa_t *spa = vd->vdev_spa; 1015 objset_t *mos = spa->spa_meta_objset; 1016 uint64_t m; 1017 uint64_t oldc = vd->vdev_ms_count; 1018 uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift; 1019 metaslab_t **mspp; 1020 int error; 1021 1022 ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER)); 1023 1024 /* 1025 * This vdev is not being allocated from yet or is a hole. 1026 */ 1027 if (vd->vdev_ms_shift == 0) 1028 return (0); 1029 1030 ASSERT(!vd->vdev_ishole); 1031 1032 ASSERT(oldc <= newc); 1033 1034 mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP); 1035 1036 if (oldc != 0) { 1037 bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp)); 1038 kmem_free(vd->vdev_ms, oldc * sizeof (*mspp)); 1039 } 1040 1041 vd->vdev_ms = mspp; 1042 vd->vdev_ms_count = newc; 1043 for (m = oldc; m < newc; m++) { 1044 uint64_t object = 0; 1045 1046 /* 1047 * vdev_ms_array may be 0 if we are creating the "fake" 1048 * metaslabs for an indirect vdev for zdb's leak detection. 1049 * See zdb_leak_init(). 1050 */ 1051 if (txg == 0 && vd->vdev_ms_array != 0) { 1052 error = dmu_read(mos, vd->vdev_ms_array, 1053 m * sizeof (uint64_t), sizeof (uint64_t), &object, 1054 DMU_READ_PREFETCH); 1055 if (error != 0) { 1056 vdev_dbgmsg(vd, "unable to read the metaslab " 1057 "array [error=%d]", error); 1058 return (error); 1059 } 1060 } 1061 1062 error = metaslab_init(vd->vdev_mg, m, object, txg, 1063 &(vd->vdev_ms[m])); 1064 if (error != 0) { 1065 vdev_dbgmsg(vd, "metaslab_init failed [error=%d]", 1066 error); 1067 return (error); 1068 } 1069 } 1070 1071 if (txg == 0) 1072 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER); 1073 1074 /* 1075 * If the vdev is being removed we don't activate 1076 * the metaslabs since we want to ensure that no new 1077 * allocations are performed on this device. 1078 */ 1079 if (oldc == 0 && !vd->vdev_removing) 1080 metaslab_group_activate(vd->vdev_mg); 1081 1082 if (txg == 0) 1083 spa_config_exit(spa, SCL_ALLOC, FTAG); 1084 1085 return (0); 1086 } 1087 1088 void 1089 vdev_metaslab_fini(vdev_t *vd) 1090 { 1091 if (vd->vdev_checkpoint_sm != NULL) { 1092 ASSERT(spa_feature_is_active(vd->vdev_spa, 1093 SPA_FEATURE_POOL_CHECKPOINT)); 1094 space_map_close(vd->vdev_checkpoint_sm); 1095 /* 1096 * Even though we close the space map, we need to set its 1097 * pointer to NULL. The reason is that vdev_metaslab_fini() 1098 * may be called multiple times for certain operations 1099 * (i.e. when destroying a pool) so we need to ensure that 1100 * this clause never executes twice. This logic is similar 1101 * to the one used for the vdev_ms clause below. 1102 */ 1103 vd->vdev_checkpoint_sm = NULL; 1104 } 1105 1106 if (vd->vdev_ms != NULL) { 1107 uint64_t count = vd->vdev_ms_count; 1108 1109 metaslab_group_passivate(vd->vdev_mg); 1110 for (uint64_t m = 0; m < count; m++) { 1111 metaslab_t *msp = vd->vdev_ms[m]; 1112 1113 if (msp != NULL) 1114 metaslab_fini(msp); 1115 } 1116 kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *)); 1117 vd->vdev_ms = NULL; 1118 1119 vd->vdev_ms_count = 0; 1120 } 1121 ASSERT0(vd->vdev_ms_count); 1122 } 1123 1124 typedef struct vdev_probe_stats { 1125 boolean_t vps_readable; 1126 boolean_t vps_writeable; 1127 int vps_flags; 1128 } vdev_probe_stats_t; 1129 1130 static void 1131 vdev_probe_done(zio_t *zio) 1132 { 1133 spa_t *spa = zio->io_spa; 1134 vdev_t *vd = zio->io_vd; 1135 vdev_probe_stats_t *vps = zio->io_private; 1136 1137 ASSERT(vd->vdev_probe_zio != NULL); 1138 1139 if (zio->io_type == ZIO_TYPE_READ) { 1140 if (zio->io_error == 0) 1141 vps->vps_readable = 1; 1142 if (zio->io_error == 0 && spa_writeable(spa)) { 1143 zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd, 1144 zio->io_offset, zio->io_size, zio->io_abd, 1145 ZIO_CHECKSUM_OFF, vdev_probe_done, vps, 1146 ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE)); 1147 } else { 1148 abd_free(zio->io_abd); 1149 } 1150 } else if (zio->io_type == ZIO_TYPE_WRITE) { 1151 if (zio->io_error == 0) 1152 vps->vps_writeable = 1; 1153 abd_free(zio->io_abd); 1154 } else if (zio->io_type == ZIO_TYPE_NULL) { 1155 zio_t *pio; 1156 1157 vd->vdev_cant_read |= !vps->vps_readable; 1158 vd->vdev_cant_write |= !vps->vps_writeable; 1159 1160 if (vdev_readable(vd) && 1161 (vdev_writeable(vd) || !spa_writeable(spa))) { 1162 zio->io_error = 0; 1163 } else { 1164 ASSERT(zio->io_error != 0); 1165 vdev_dbgmsg(vd, "failed probe"); 1166 zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE, 1167 spa, vd, NULL, 0, 0); 1168 zio->io_error = SET_ERROR(ENXIO); 1169 } 1170 1171 mutex_enter(&vd->vdev_probe_lock); 1172 ASSERT(vd->vdev_probe_zio == zio); 1173 vd->vdev_probe_zio = NULL; 1174 mutex_exit(&vd->vdev_probe_lock); 1175 1176 zio_link_t *zl = NULL; 1177 while ((pio = zio_walk_parents(zio, &zl)) != NULL) 1178 if (!vdev_accessible(vd, pio)) 1179 pio->io_error = SET_ERROR(ENXIO); 1180 1181 kmem_free(vps, sizeof (*vps)); 1182 } 1183 } 1184 1185 /* 1186 * Determine whether this device is accessible. 1187 * 1188 * Read and write to several known locations: the pad regions of each 1189 * vdev label but the first, which we leave alone in case it contains 1190 * a VTOC. 1191 */ 1192 zio_t * 1193 vdev_probe(vdev_t *vd, zio_t *zio) 1194 { 1195 spa_t *spa = vd->vdev_spa; 1196 vdev_probe_stats_t *vps = NULL; 1197 zio_t *pio; 1198 1199 ASSERT(vd->vdev_ops->vdev_op_leaf); 1200 1201 /* 1202 * Don't probe the probe. 1203 */ 1204 if (zio && (zio->io_flags & ZIO_FLAG_PROBE)) 1205 return (NULL); 1206 1207 /* 1208 * To prevent 'probe storms' when a device fails, we create 1209 * just one probe i/o at a time. All zios that want to probe 1210 * this vdev will become parents of the probe io. 1211 */ 1212 mutex_enter(&vd->vdev_probe_lock); 1213 1214 if ((pio = vd->vdev_probe_zio) == NULL) { 1215 vps = kmem_zalloc(sizeof (*vps), KM_SLEEP); 1216 1217 vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE | 1218 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE | 1219 ZIO_FLAG_TRYHARD; 1220 1221 if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) { 1222 /* 1223 * vdev_cant_read and vdev_cant_write can only 1224 * transition from TRUE to FALSE when we have the 1225 * SCL_ZIO lock as writer; otherwise they can only 1226 * transition from FALSE to TRUE. This ensures that 1227 * any zio looking at these values can assume that 1228 * failures persist for the life of the I/O. That's 1229 * important because when a device has intermittent 1230 * connectivity problems, we want to ensure that 1231 * they're ascribed to the device (ENXIO) and not 1232 * the zio (EIO). 1233 * 1234 * Since we hold SCL_ZIO as writer here, clear both 1235 * values so the probe can reevaluate from first 1236 * principles. 1237 */ 1238 vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER; 1239 vd->vdev_cant_read = B_FALSE; 1240 vd->vdev_cant_write = B_FALSE; 1241 } 1242 1243 vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd, 1244 vdev_probe_done, vps, 1245 vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE); 1246 1247 /* 1248 * We can't change the vdev state in this context, so we 1249 * kick off an async task to do it on our behalf. 1250 */ 1251 if (zio != NULL) { 1252 vd->vdev_probe_wanted = B_TRUE; 1253 spa_async_request(spa, SPA_ASYNC_PROBE); 1254 } 1255 } 1256 1257 if (zio != NULL) 1258 zio_add_child(zio, pio); 1259 1260 mutex_exit(&vd->vdev_probe_lock); 1261 1262 if (vps == NULL) { 1263 ASSERT(zio != NULL); 1264 return (NULL); 1265 } 1266 1267 for (int l = 1; l < VDEV_LABELS; l++) { 1268 zio_nowait(zio_read_phys(pio, vd, 1269 vdev_label_offset(vd->vdev_psize, l, 1270 offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE, 1271 abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE), 1272 ZIO_CHECKSUM_OFF, vdev_probe_done, vps, 1273 ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE)); 1274 } 1275 1276 if (zio == NULL) 1277 return (pio); 1278 1279 zio_nowait(pio); 1280 return (NULL); 1281 } 1282 1283 static void 1284 vdev_open_child(void *arg) 1285 { 1286 vdev_t *vd = arg; 1287 1288 vd->vdev_open_thread = curthread; 1289 vd->vdev_open_error = vdev_open(vd); 1290 vd->vdev_open_thread = NULL; 1291 } 1292 1293 boolean_t 1294 vdev_uses_zvols(vdev_t *vd) 1295 { 1296 if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR, 1297 strlen(ZVOL_DIR)) == 0) 1298 return (B_TRUE); 1299 for (int c = 0; c < vd->vdev_children; c++) 1300 if (vdev_uses_zvols(vd->vdev_child[c])) 1301 return (B_TRUE); 1302 return (B_FALSE); 1303 } 1304 1305 void 1306 vdev_open_children(vdev_t *vd) 1307 { 1308 taskq_t *tq; 1309 int children = vd->vdev_children; 1310 1311 /* 1312 * in order to handle pools on top of zvols, do the opens 1313 * in a single thread so that the same thread holds the 1314 * spa_namespace_lock 1315 */ 1316 if (vdev_uses_zvols(vd)) { 1317 for (int c = 0; c < children; c++) 1318 vd->vdev_child[c]->vdev_open_error = 1319 vdev_open(vd->vdev_child[c]); 1320 return; 1321 } 1322 tq = taskq_create("vdev_open", children, minclsyspri, 1323 children, children, TASKQ_PREPOPULATE); 1324 1325 for (int c = 0; c < children; c++) 1326 VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c], 1327 TQ_SLEEP) != NULL); 1328 1329 taskq_destroy(tq); 1330 } 1331 1332 /* 1333 * Compute the raidz-deflation ratio. Note, we hard-code 1334 * in 128k (1 << 17) because it is the "typical" blocksize. 1335 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change, 1336 * otherwise it would inconsistently account for existing bp's. 1337 */ 1338 static void 1339 vdev_set_deflate_ratio(vdev_t *vd) 1340 { 1341 if (vd == vd->vdev_top && !vd->vdev_ishole && vd->vdev_ashift != 0) { 1342 vd->vdev_deflate_ratio = (1 << 17) / 1343 (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT); 1344 } 1345 } 1346 1347 /* 1348 * Prepare a virtual device for access. 1349 */ 1350 int 1351 vdev_open(vdev_t *vd) 1352 { 1353 spa_t *spa = vd->vdev_spa; 1354 int error; 1355 uint64_t osize = 0; 1356 uint64_t max_osize = 0; 1357 uint64_t asize, max_asize, psize; 1358 uint64_t ashift = 0; 1359 1360 ASSERT(vd->vdev_open_thread == curthread || 1361 spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL); 1362 ASSERT(vd->vdev_state == VDEV_STATE_CLOSED || 1363 vd->vdev_state == VDEV_STATE_CANT_OPEN || 1364 vd->vdev_state == VDEV_STATE_OFFLINE); 1365 1366 vd->vdev_stat.vs_aux = VDEV_AUX_NONE; 1367 vd->vdev_cant_read = B_FALSE; 1368 vd->vdev_cant_write = B_FALSE; 1369 vd->vdev_min_asize = vdev_get_min_asize(vd); 1370 1371 /* 1372 * If this vdev is not removed, check its fault status. If it's 1373 * faulted, bail out of the open. 1374 */ 1375 if (!vd->vdev_removed && vd->vdev_faulted) { 1376 ASSERT(vd->vdev_children == 0); 1377 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED || 1378 vd->vdev_label_aux == VDEV_AUX_EXTERNAL); 1379 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED, 1380 vd->vdev_label_aux); 1381 return (SET_ERROR(ENXIO)); 1382 } else if (vd->vdev_offline) { 1383 ASSERT(vd->vdev_children == 0); 1384 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE); 1385 return (SET_ERROR(ENXIO)); 1386 } 1387 1388 error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift); 1389 1390 /* 1391 * Reset the vdev_reopening flag so that we actually close 1392 * the vdev on error. 1393 */ 1394 vd->vdev_reopening = B_FALSE; 1395 if (zio_injection_enabled && error == 0) 1396 error = zio_handle_device_injection(vd, NULL, ENXIO); 1397 1398 if (error) { 1399 if (vd->vdev_removed && 1400 vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED) 1401 vd->vdev_removed = B_FALSE; 1402 1403 if (vd->vdev_stat.vs_aux == VDEV_AUX_CHILDREN_OFFLINE) { 1404 vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, 1405 vd->vdev_stat.vs_aux); 1406 } else { 1407 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 1408 vd->vdev_stat.vs_aux); 1409 } 1410 return (error); 1411 } 1412 1413 vd->vdev_removed = B_FALSE; 1414 1415 /* 1416 * Recheck the faulted flag now that we have confirmed that 1417 * the vdev is accessible. If we're faulted, bail. 1418 */ 1419 if (vd->vdev_faulted) { 1420 ASSERT(vd->vdev_children == 0); 1421 ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED || 1422 vd->vdev_label_aux == VDEV_AUX_EXTERNAL); 1423 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED, 1424 vd->vdev_label_aux); 1425 return (SET_ERROR(ENXIO)); 1426 } 1427 1428 if (vd->vdev_degraded) { 1429 ASSERT(vd->vdev_children == 0); 1430 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED, 1431 VDEV_AUX_ERR_EXCEEDED); 1432 } else { 1433 vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0); 1434 } 1435 1436 /* 1437 * For hole or missing vdevs we just return success. 1438 */ 1439 if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) 1440 return (0); 1441 1442 for (int c = 0; c < vd->vdev_children; c++) { 1443 if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) { 1444 vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED, 1445 VDEV_AUX_NONE); 1446 break; 1447 } 1448 } 1449 1450 osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t)); 1451 max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t)); 1452 1453 if (vd->vdev_children == 0) { 1454 if (osize < SPA_MINDEVSIZE) { 1455 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 1456 VDEV_AUX_TOO_SMALL); 1457 return (SET_ERROR(EOVERFLOW)); 1458 } 1459 psize = osize; 1460 asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE); 1461 max_asize = max_osize - (VDEV_LABEL_START_SIZE + 1462 VDEV_LABEL_END_SIZE); 1463 } else { 1464 if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE - 1465 (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) { 1466 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 1467 VDEV_AUX_TOO_SMALL); 1468 return (SET_ERROR(EOVERFLOW)); 1469 } 1470 psize = 0; 1471 asize = osize; 1472 max_asize = max_osize; 1473 } 1474 1475 vd->vdev_psize = psize; 1476 1477 /* 1478 * Make sure the allocatable size hasn't shrunk too much. 1479 */ 1480 if (asize < vd->vdev_min_asize) { 1481 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 1482 VDEV_AUX_BAD_LABEL); 1483 return (SET_ERROR(EINVAL)); 1484 } 1485 1486 if (vd->vdev_asize == 0) { 1487 /* 1488 * This is the first-ever open, so use the computed values. 1489 * For testing purposes, a higher ashift can be requested. 1490 */ 1491 vd->vdev_asize = asize; 1492 vd->vdev_max_asize = max_asize; 1493 vd->vdev_ashift = MAX(ashift, vd->vdev_ashift); 1494 } else { 1495 /* 1496 * Detect if the alignment requirement has increased. 1497 * We don't want to make the pool unavailable, just 1498 * issue a warning instead. 1499 */ 1500 if (ashift > vd->vdev_top->vdev_ashift && 1501 vd->vdev_ops->vdev_op_leaf) { 1502 cmn_err(CE_WARN, 1503 "Disk, '%s', has a block alignment that is " 1504 "larger than the pool's alignment\n", 1505 vd->vdev_path); 1506 } 1507 vd->vdev_max_asize = max_asize; 1508 } 1509 1510 /* 1511 * If all children are healthy we update asize if either: 1512 * The asize has increased, due to a device expansion caused by dynamic 1513 * LUN growth or vdev replacement, and automatic expansion is enabled; 1514 * making the additional space available. 1515 * 1516 * The asize has decreased, due to a device shrink usually caused by a 1517 * vdev replace with a smaller device. This ensures that calculations 1518 * based of max_asize and asize e.g. esize are always valid. It's safe 1519 * to do this as we've already validated that asize is greater than 1520 * vdev_min_asize. 1521 */ 1522 if (vd->vdev_state == VDEV_STATE_HEALTHY && 1523 ((asize > vd->vdev_asize && 1524 (vd->vdev_expanding || spa->spa_autoexpand)) || 1525 (asize < vd->vdev_asize))) 1526 vd->vdev_asize = asize; 1527 1528 vdev_set_min_asize(vd); 1529 1530 /* 1531 * Ensure we can issue some IO before declaring the 1532 * vdev open for business. 1533 */ 1534 if (vd->vdev_ops->vdev_op_leaf && 1535 (error = zio_wait(vdev_probe(vd, NULL))) != 0) { 1536 vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED, 1537 VDEV_AUX_ERR_EXCEEDED); 1538 return (error); 1539 } 1540 1541 /* 1542 * Track the min and max ashift values for normal data devices. 1543 */ 1544 if (vd->vdev_top == vd && vd->vdev_ashift != 0 && 1545 !vd->vdev_islog && vd->vdev_aux == NULL) { 1546 if (vd->vdev_ashift > spa->spa_max_ashift) 1547 spa->spa_max_ashift = vd->vdev_ashift; 1548 if (vd->vdev_ashift < spa->spa_min_ashift) 1549 spa->spa_min_ashift = vd->vdev_ashift; 1550 } 1551 1552 /* 1553 * If a leaf vdev has a DTL, and seems healthy, then kick off a 1554 * resilver. But don't do this if we are doing a reopen for a scrub, 1555 * since this would just restart the scrub we are already doing. 1556 */ 1557 if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen && 1558 vdev_resilver_needed(vd, NULL, NULL)) 1559 spa_async_request(spa, SPA_ASYNC_RESILVER); 1560 1561 return (0); 1562 } 1563 1564 /* 1565 * Called once the vdevs are all opened, this routine validates the label 1566 * contents. This needs to be done before vdev_load() so that we don't 1567 * inadvertently do repair I/Os to the wrong device. 1568 * 1569 * This function will only return failure if one of the vdevs indicates that it 1570 * has since been destroyed or exported. This is only possible if 1571 * /etc/zfs/zpool.cache was readonly at the time. Otherwise, the vdev state 1572 * will be updated but the function will return 0. 1573 */ 1574 int 1575 vdev_validate(vdev_t *vd) 1576 { 1577 spa_t *spa = vd->vdev_spa; 1578 nvlist_t *label; 1579 uint64_t guid = 0, aux_guid = 0, top_guid; 1580 uint64_t state; 1581 nvlist_t *nvl; 1582 uint64_t txg; 1583 1584 if (vdev_validate_skip) 1585 return (0); 1586 1587 for (uint64_t c = 0; c < vd->vdev_children; c++) 1588 if (vdev_validate(vd->vdev_child[c]) != 0) 1589 return (SET_ERROR(EBADF)); 1590 1591 /* 1592 * If the device has already failed, or was marked offline, don't do 1593 * any further validation. Otherwise, label I/O will fail and we will 1594 * overwrite the previous state. 1595 */ 1596 if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd)) 1597 return (0); 1598 1599 /* 1600 * If we are performing an extreme rewind, we allow for a label that 1601 * was modified at a point after the current txg. 1602 * If config lock is not held do not check for the txg. spa_sync could 1603 * be updating the vdev's label before updating spa_last_synced_txg. 1604 */ 1605 if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0 || 1606 spa_config_held(spa, SCL_CONFIG, RW_WRITER) != SCL_CONFIG) 1607 txg = UINT64_MAX; 1608 else 1609 txg = spa_last_synced_txg(spa); 1610 1611 if ((label = vdev_label_read_config(vd, txg)) == NULL) { 1612 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 1613 VDEV_AUX_BAD_LABEL); 1614 vdev_dbgmsg(vd, "vdev_validate: failed reading config for " 1615 "txg %llu", (u_longlong_t)txg); 1616 return (0); 1617 } 1618 1619 /* 1620 * Determine if this vdev has been split off into another 1621 * pool. If so, then refuse to open it. 1622 */ 1623 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID, 1624 &aux_guid) == 0 && aux_guid == spa_guid(spa)) { 1625 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1626 VDEV_AUX_SPLIT_POOL); 1627 nvlist_free(label); 1628 vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool"); 1629 return (0); 1630 } 1631 1632 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) { 1633 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1634 VDEV_AUX_CORRUPT_DATA); 1635 nvlist_free(label); 1636 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label", 1637 ZPOOL_CONFIG_POOL_GUID); 1638 return (0); 1639 } 1640 1641 /* 1642 * If config is not trusted then ignore the spa guid check. This is 1643 * necessary because if the machine crashed during a re-guid the new 1644 * guid might have been written to all of the vdev labels, but not the 1645 * cached config. The check will be performed again once we have the 1646 * trusted config from the MOS. 1647 */ 1648 if (spa->spa_trust_config && guid != spa_guid(spa)) { 1649 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1650 VDEV_AUX_CORRUPT_DATA); 1651 nvlist_free(label); 1652 vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't " 1653 "match config (%llu != %llu)", (u_longlong_t)guid, 1654 (u_longlong_t)spa_guid(spa)); 1655 return (0); 1656 } 1657 1658 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl) 1659 != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID, 1660 &aux_guid) != 0) 1661 aux_guid = 0; 1662 1663 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) { 1664 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1665 VDEV_AUX_CORRUPT_DATA); 1666 nvlist_free(label); 1667 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label", 1668 ZPOOL_CONFIG_GUID); 1669 return (0); 1670 } 1671 1672 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid) 1673 != 0) { 1674 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1675 VDEV_AUX_CORRUPT_DATA); 1676 nvlist_free(label); 1677 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label", 1678 ZPOOL_CONFIG_TOP_GUID); 1679 return (0); 1680 } 1681 1682 /* 1683 * If this vdev just became a top-level vdev because its sibling was 1684 * detached, it will have adopted the parent's vdev guid -- but the 1685 * label may or may not be on disk yet. Fortunately, either version 1686 * of the label will have the same top guid, so if we're a top-level 1687 * vdev, we can safely compare to that instead. 1688 * However, if the config comes from a cachefile that failed to update 1689 * after the detach, a top-level vdev will appear as a non top-level 1690 * vdev in the config. Also relax the constraints if we perform an 1691 * extreme rewind. 1692 * 1693 * If we split this vdev off instead, then we also check the 1694 * original pool's guid. We don't want to consider the vdev 1695 * corrupt if it is partway through a split operation. 1696 */ 1697 if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) { 1698 boolean_t mismatch = B_FALSE; 1699 if (spa->spa_trust_config && !spa->spa_extreme_rewind) { 1700 if (vd != vd->vdev_top || vd->vdev_guid != top_guid) 1701 mismatch = B_TRUE; 1702 } else { 1703 if (vd->vdev_guid != top_guid && 1704 vd->vdev_top->vdev_guid != guid) 1705 mismatch = B_TRUE; 1706 } 1707 1708 if (mismatch) { 1709 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1710 VDEV_AUX_CORRUPT_DATA); 1711 nvlist_free(label); 1712 vdev_dbgmsg(vd, "vdev_validate: config guid " 1713 "doesn't match label guid"); 1714 vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu", 1715 (u_longlong_t)vd->vdev_guid, 1716 (u_longlong_t)vd->vdev_top->vdev_guid); 1717 vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, " 1718 "aux_guid %llu", (u_longlong_t)guid, 1719 (u_longlong_t)top_guid, (u_longlong_t)aux_guid); 1720 return (0); 1721 } 1722 } 1723 1724 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, 1725 &state) != 0) { 1726 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 1727 VDEV_AUX_CORRUPT_DATA); 1728 nvlist_free(label); 1729 vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label", 1730 ZPOOL_CONFIG_POOL_STATE); 1731 return (0); 1732 } 1733 1734 nvlist_free(label); 1735 1736 /* 1737 * If this is a verbatim import, no need to check the 1738 * state of the pool. 1739 */ 1740 if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) && 1741 spa_load_state(spa) == SPA_LOAD_OPEN && 1742 state != POOL_STATE_ACTIVE) { 1743 vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) " 1744 "for spa %s", (u_longlong_t)state, spa->spa_name); 1745 return (SET_ERROR(EBADF)); 1746 } 1747 1748 /* 1749 * If we were able to open and validate a vdev that was 1750 * previously marked permanently unavailable, clear that state 1751 * now. 1752 */ 1753 if (vd->vdev_not_present) 1754 vd->vdev_not_present = 0; 1755 1756 return (0); 1757 } 1758 1759 static void 1760 vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd) 1761 { 1762 if (svd->vdev_path != NULL && dvd->vdev_path != NULL) { 1763 if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) { 1764 zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed " 1765 "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid, 1766 dvd->vdev_path, svd->vdev_path); 1767 spa_strfree(dvd->vdev_path); 1768 dvd->vdev_path = spa_strdup(svd->vdev_path); 1769 } 1770 } else if (svd->vdev_path != NULL) { 1771 dvd->vdev_path = spa_strdup(svd->vdev_path); 1772 zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'", 1773 (u_longlong_t)dvd->vdev_guid, dvd->vdev_path); 1774 } 1775 } 1776 1777 /* 1778 * Recursively copy vdev paths from one vdev to another. Source and destination 1779 * vdev trees must have same geometry otherwise return error. Intended to copy 1780 * paths from userland config into MOS config. 1781 */ 1782 int 1783 vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd) 1784 { 1785 if ((svd->vdev_ops == &vdev_missing_ops) || 1786 (svd->vdev_ishole && dvd->vdev_ishole) || 1787 (dvd->vdev_ops == &vdev_indirect_ops)) 1788 return (0); 1789 1790 if (svd->vdev_ops != dvd->vdev_ops) { 1791 vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s", 1792 svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type); 1793 return (SET_ERROR(EINVAL)); 1794 } 1795 1796 if (svd->vdev_guid != dvd->vdev_guid) { 1797 vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != " 1798 "%llu)", (u_longlong_t)svd->vdev_guid, 1799 (u_longlong_t)dvd->vdev_guid); 1800 return (SET_ERROR(EINVAL)); 1801 } 1802 1803 if (svd->vdev_children != dvd->vdev_children) { 1804 vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: " 1805 "%llu != %llu", (u_longlong_t)svd->vdev_children, 1806 (u_longlong_t)dvd->vdev_children); 1807 return (SET_ERROR(EINVAL)); 1808 } 1809 1810 for (uint64_t i = 0; i < svd->vdev_children; i++) { 1811 int error = vdev_copy_path_strict(svd->vdev_child[i], 1812 dvd->vdev_child[i]); 1813 if (error != 0) 1814 return (error); 1815 } 1816 1817 if (svd->vdev_ops->vdev_op_leaf) 1818 vdev_copy_path_impl(svd, dvd); 1819 1820 return (0); 1821 } 1822 1823 static void 1824 vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd) 1825 { 1826 ASSERT(stvd->vdev_top == stvd); 1827 ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id); 1828 1829 for (uint64_t i = 0; i < dvd->vdev_children; i++) { 1830 vdev_copy_path_search(stvd, dvd->vdev_child[i]); 1831 } 1832 1833 if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd)) 1834 return; 1835 1836 /* 1837 * The idea here is that while a vdev can shift positions within 1838 * a top vdev (when replacing, attaching mirror, etc.) it cannot 1839 * step outside of it. 1840 */ 1841 vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid); 1842 1843 if (vd == NULL || vd->vdev_ops != dvd->vdev_ops) 1844 return; 1845 1846 ASSERT(vd->vdev_ops->vdev_op_leaf); 1847 1848 vdev_copy_path_impl(vd, dvd); 1849 } 1850 1851 /* 1852 * Recursively copy vdev paths from one root vdev to another. Source and 1853 * destination vdev trees may differ in geometry. For each destination leaf 1854 * vdev, search a vdev with the same guid and top vdev id in the source. 1855 * Intended to copy paths from userland config into MOS config. 1856 */ 1857 void 1858 vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd) 1859 { 1860 uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children); 1861 ASSERT(srvd->vdev_ops == &vdev_root_ops); 1862 ASSERT(drvd->vdev_ops == &vdev_root_ops); 1863 1864 for (uint64_t i = 0; i < children; i++) { 1865 vdev_copy_path_search(srvd->vdev_child[i], 1866 drvd->vdev_child[i]); 1867 } 1868 } 1869 1870 /* 1871 * Close a virtual device. 1872 */ 1873 void 1874 vdev_close(vdev_t *vd) 1875 { 1876 spa_t *spa = vd->vdev_spa; 1877 vdev_t *pvd = vd->vdev_parent; 1878 1879 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL); 1880 1881 /* 1882 * If our parent is reopening, then we are as well, unless we are 1883 * going offline. 1884 */ 1885 if (pvd != NULL && pvd->vdev_reopening) 1886 vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline); 1887 1888 vd->vdev_ops->vdev_op_close(vd); 1889 1890 vdev_cache_purge(vd); 1891 1892 /* 1893 * We record the previous state before we close it, so that if we are 1894 * doing a reopen(), we don't generate FMA ereports if we notice that 1895 * it's still faulted. 1896 */ 1897 vd->vdev_prevstate = vd->vdev_state; 1898 1899 if (vd->vdev_offline) 1900 vd->vdev_state = VDEV_STATE_OFFLINE; 1901 else 1902 vd->vdev_state = VDEV_STATE_CLOSED; 1903 vd->vdev_stat.vs_aux = VDEV_AUX_NONE; 1904 } 1905 1906 void 1907 vdev_hold(vdev_t *vd) 1908 { 1909 spa_t *spa = vd->vdev_spa; 1910 1911 ASSERT(spa_is_root(spa)); 1912 if (spa->spa_state == POOL_STATE_UNINITIALIZED) 1913 return; 1914 1915 for (int c = 0; c < vd->vdev_children; c++) 1916 vdev_hold(vd->vdev_child[c]); 1917 1918 if (vd->vdev_ops->vdev_op_leaf) 1919 vd->vdev_ops->vdev_op_hold(vd); 1920 } 1921 1922 void 1923 vdev_rele(vdev_t *vd) 1924 { 1925 spa_t *spa = vd->vdev_spa; 1926 1927 ASSERT(spa_is_root(spa)); 1928 for (int c = 0; c < vd->vdev_children; c++) 1929 vdev_rele(vd->vdev_child[c]); 1930 1931 if (vd->vdev_ops->vdev_op_leaf) 1932 vd->vdev_ops->vdev_op_rele(vd); 1933 } 1934 1935 /* 1936 * Reopen all interior vdevs and any unopened leaves. We don't actually 1937 * reopen leaf vdevs which had previously been opened as they might deadlock 1938 * on the spa_config_lock. Instead we only obtain the leaf's physical size. 1939 * If the leaf has never been opened then open it, as usual. 1940 */ 1941 void 1942 vdev_reopen(vdev_t *vd) 1943 { 1944 spa_t *spa = vd->vdev_spa; 1945 1946 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL); 1947 1948 /* set the reopening flag unless we're taking the vdev offline */ 1949 vd->vdev_reopening = !vd->vdev_offline; 1950 vdev_close(vd); 1951 (void) vdev_open(vd); 1952 1953 /* 1954 * Call vdev_validate() here to make sure we have the same device. 1955 * Otherwise, a device with an invalid label could be successfully 1956 * opened in response to vdev_reopen(). 1957 */ 1958 if (vd->vdev_aux) { 1959 (void) vdev_validate_aux(vd); 1960 if (vdev_readable(vd) && vdev_writeable(vd) && 1961 vd->vdev_aux == &spa->spa_l2cache && 1962 !l2arc_vdev_present(vd)) 1963 l2arc_add_vdev(spa, vd); 1964 } else { 1965 (void) vdev_validate(vd); 1966 } 1967 1968 /* 1969 * Reassess parent vdev's health. 1970 */ 1971 vdev_propagate_state(vd); 1972 } 1973 1974 int 1975 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing) 1976 { 1977 int error; 1978 1979 /* 1980 * Normally, partial opens (e.g. of a mirror) are allowed. 1981 * For a create, however, we want to fail the request if 1982 * there are any components we can't open. 1983 */ 1984 error = vdev_open(vd); 1985 1986 if (error || vd->vdev_state != VDEV_STATE_HEALTHY) { 1987 vdev_close(vd); 1988 return (error ? error : ENXIO); 1989 } 1990 1991 /* 1992 * Recursively load DTLs and initialize all labels. 1993 */ 1994 if ((error = vdev_dtl_load(vd)) != 0 || 1995 (error = vdev_label_init(vd, txg, isreplacing ? 1996 VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) { 1997 vdev_close(vd); 1998 return (error); 1999 } 2000 2001 return (0); 2002 } 2003 2004 void 2005 vdev_metaslab_set_size(vdev_t *vd) 2006 { 2007 uint64_t asize = vd->vdev_asize; 2008 uint64_t ms_shift = 0; 2009 2010 /* 2011 * For vdevs that are bigger than 8G the metaslab size varies in 2012 * a way that the number of metaslabs increases in powers of two, 2013 * linearly in terms of vdev_asize, starting from 16 metaslabs. 2014 * So for vdev_asize of 8G we get 16 metaslabs, for 16G, we get 32, 2015 * and so on, until we hit the maximum metaslab count limit 2016 * [vdev_max_ms_count] from which point the metaslab count stays 2017 * the same. 2018 */ 2019 ms_shift = vdev_default_ms_shift; 2020 2021 if ((asize >> ms_shift) < vdev_min_ms_count) { 2022 /* 2023 * For devices that are less than 8G we want to have 2024 * exactly 16 metaslabs. We don't want less as integer 2025 * division rounds down, so less metaslabs mean more 2026 * wasted space. We don't want more as these vdevs are 2027 * small and in the likely event that we are running 2028 * out of space, the SPA will have a hard time finding 2029 * space due to fragmentation. 2030 */ 2031 ms_shift = highbit64(asize / vdev_min_ms_count); 2032 ms_shift = MAX(ms_shift, SPA_MAXBLOCKSHIFT); 2033 2034 } else if ((asize >> ms_shift) > vdev_max_ms_count) { 2035 ms_shift = highbit64(asize / vdev_max_ms_count); 2036 } 2037 2038 vd->vdev_ms_shift = ms_shift; 2039 ASSERT3U(vd->vdev_ms_shift, >=, SPA_MAXBLOCKSHIFT); 2040 } 2041 2042 void 2043 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg) 2044 { 2045 ASSERT(vd == vd->vdev_top); 2046 /* indirect vdevs don't have metaslabs or dtls */ 2047 ASSERT(vdev_is_concrete(vd) || flags == 0); 2048 ASSERT(ISP2(flags)); 2049 ASSERT(spa_writeable(vd->vdev_spa)); 2050 2051 if (flags & VDD_METASLAB) 2052 (void) txg_list_add(&vd->vdev_ms_list, arg, txg); 2053 2054 if (flags & VDD_DTL) 2055 (void) txg_list_add(&vd->vdev_dtl_list, arg, txg); 2056 2057 (void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg); 2058 } 2059 2060 void 2061 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg) 2062 { 2063 for (int c = 0; c < vd->vdev_children; c++) 2064 vdev_dirty_leaves(vd->vdev_child[c], flags, txg); 2065 2066 if (vd->vdev_ops->vdev_op_leaf) 2067 vdev_dirty(vd->vdev_top, flags, vd, txg); 2068 } 2069 2070 /* 2071 * DTLs. 2072 * 2073 * A vdev's DTL (dirty time log) is the set of transaction groups for which 2074 * the vdev has less than perfect replication. There are four kinds of DTL: 2075 * 2076 * DTL_MISSING: txgs for which the vdev has no valid copies of the data 2077 * 2078 * DTL_PARTIAL: txgs for which data is available, but not fully replicated 2079 * 2080 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon 2081 * scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of 2082 * txgs that was scrubbed. 2083 * 2084 * DTL_OUTAGE: txgs which cannot currently be read, whether due to 2085 * persistent errors or just some device being offline. 2086 * Unlike the other three, the DTL_OUTAGE map is not generally 2087 * maintained; it's only computed when needed, typically to 2088 * determine whether a device can be detached. 2089 * 2090 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device 2091 * either has the data or it doesn't. 2092 * 2093 * For interior vdevs such as mirror and RAID-Z the picture is more complex. 2094 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because 2095 * if any child is less than fully replicated, then so is its parent. 2096 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs, 2097 * comprising only those txgs which appear in 'maxfaults' or more children; 2098 * those are the txgs we don't have enough replication to read. For example, 2099 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2); 2100 * thus, its DTL_MISSING consists of the set of txgs that appear in more than 2101 * two child DTL_MISSING maps. 2102 * 2103 * It should be clear from the above that to compute the DTLs and outage maps 2104 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps. 2105 * Therefore, that is all we keep on disk. When loading the pool, or after 2106 * a configuration change, we generate all other DTLs from first principles. 2107 */ 2108 void 2109 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size) 2110 { 2111 range_tree_t *rt = vd->vdev_dtl[t]; 2112 2113 ASSERT(t < DTL_TYPES); 2114 ASSERT(vd != vd->vdev_spa->spa_root_vdev); 2115 ASSERT(spa_writeable(vd->vdev_spa)); 2116 2117 mutex_enter(&vd->vdev_dtl_lock); 2118 if (!range_tree_contains(rt, txg, size)) 2119 range_tree_add(rt, txg, size); 2120 mutex_exit(&vd->vdev_dtl_lock); 2121 } 2122 2123 boolean_t 2124 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size) 2125 { 2126 range_tree_t *rt = vd->vdev_dtl[t]; 2127 boolean_t dirty = B_FALSE; 2128 2129 ASSERT(t < DTL_TYPES); 2130 ASSERT(vd != vd->vdev_spa->spa_root_vdev); 2131 2132 /* 2133 * While we are loading the pool, the DTLs have not been loaded yet. 2134 * Ignore the DTLs and try all devices. This avoids a recursive 2135 * mutex enter on the vdev_dtl_lock, and also makes us try hard 2136 * when loading the pool (relying on the checksum to ensure that 2137 * we get the right data -- note that we while loading, we are 2138 * only reading the MOS, which is always checksummed). 2139 */ 2140 if (vd->vdev_spa->spa_load_state != SPA_LOAD_NONE) 2141 return (B_FALSE); 2142 2143 mutex_enter(&vd->vdev_dtl_lock); 2144 if (!range_tree_is_empty(rt)) 2145 dirty = range_tree_contains(rt, txg, size); 2146 mutex_exit(&vd->vdev_dtl_lock); 2147 2148 return (dirty); 2149 } 2150 2151 boolean_t 2152 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t) 2153 { 2154 range_tree_t *rt = vd->vdev_dtl[t]; 2155 boolean_t empty; 2156 2157 mutex_enter(&vd->vdev_dtl_lock); 2158 empty = range_tree_is_empty(rt); 2159 mutex_exit(&vd->vdev_dtl_lock); 2160 2161 return (empty); 2162 } 2163 2164 /* 2165 * Returns the lowest txg in the DTL range. 2166 */ 2167 static uint64_t 2168 vdev_dtl_min(vdev_t *vd) 2169 { 2170 range_seg_t *rs; 2171 2172 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock)); 2173 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0); 2174 ASSERT0(vd->vdev_children); 2175 2176 rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root); 2177 return (rs->rs_start - 1); 2178 } 2179 2180 /* 2181 * Returns the highest txg in the DTL. 2182 */ 2183 static uint64_t 2184 vdev_dtl_max(vdev_t *vd) 2185 { 2186 range_seg_t *rs; 2187 2188 ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock)); 2189 ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0); 2190 ASSERT0(vd->vdev_children); 2191 2192 rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root); 2193 return (rs->rs_end); 2194 } 2195 2196 /* 2197 * Determine if a resilvering vdev should remove any DTL entries from 2198 * its range. If the vdev was resilvering for the entire duration of the 2199 * scan then it should excise that range from its DTLs. Otherwise, this 2200 * vdev is considered partially resilvered and should leave its DTL 2201 * entries intact. The comment in vdev_dtl_reassess() describes how we 2202 * excise the DTLs. 2203 */ 2204 static boolean_t 2205 vdev_dtl_should_excise(vdev_t *vd) 2206 { 2207 spa_t *spa = vd->vdev_spa; 2208 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan; 2209 2210 ASSERT0(scn->scn_phys.scn_errors); 2211 ASSERT0(vd->vdev_children); 2212 2213 if (vd->vdev_state < VDEV_STATE_DEGRADED) 2214 return (B_FALSE); 2215 2216 if (vd->vdev_resilver_txg == 0 || 2217 range_tree_is_empty(vd->vdev_dtl[DTL_MISSING])) 2218 return (B_TRUE); 2219 2220 /* 2221 * When a resilver is initiated the scan will assign the scn_max_txg 2222 * value to the highest txg value that exists in all DTLs. If this 2223 * device's max DTL is not part of this scan (i.e. it is not in 2224 * the range (scn_min_txg, scn_max_txg] then it is not eligible 2225 * for excision. 2226 */ 2227 if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) { 2228 ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd)); 2229 ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg); 2230 ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg); 2231 return (B_TRUE); 2232 } 2233 return (B_FALSE); 2234 } 2235 2236 /* 2237 * Reassess DTLs after a config change or scrub completion. 2238 */ 2239 void 2240 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done) 2241 { 2242 spa_t *spa = vd->vdev_spa; 2243 avl_tree_t reftree; 2244 int minref; 2245 2246 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0); 2247 2248 for (int c = 0; c < vd->vdev_children; c++) 2249 vdev_dtl_reassess(vd->vdev_child[c], txg, 2250 scrub_txg, scrub_done); 2251 2252 if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux) 2253 return; 2254 2255 if (vd->vdev_ops->vdev_op_leaf) { 2256 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan; 2257 2258 mutex_enter(&vd->vdev_dtl_lock); 2259 2260 /* 2261 * If we've completed a scan cleanly then determine 2262 * if this vdev should remove any DTLs. We only want to 2263 * excise regions on vdevs that were available during 2264 * the entire duration of this scan. 2265 */ 2266 if (scrub_txg != 0 && 2267 (spa->spa_scrub_started || 2268 (scn != NULL && scn->scn_phys.scn_errors == 0)) && 2269 vdev_dtl_should_excise(vd)) { 2270 /* 2271 * We completed a scrub up to scrub_txg. If we 2272 * did it without rebooting, then the scrub dtl 2273 * will be valid, so excise the old region and 2274 * fold in the scrub dtl. Otherwise, leave the 2275 * dtl as-is if there was an error. 2276 * 2277 * There's little trick here: to excise the beginning 2278 * of the DTL_MISSING map, we put it into a reference 2279 * tree and then add a segment with refcnt -1 that 2280 * covers the range [0, scrub_txg). This means 2281 * that each txg in that range has refcnt -1 or 0. 2282 * We then add DTL_SCRUB with a refcnt of 2, so that 2283 * entries in the range [0, scrub_txg) will have a 2284 * positive refcnt -- either 1 or 2. We then convert 2285 * the reference tree into the new DTL_MISSING map. 2286 */ 2287 space_reftree_create(&reftree); 2288 space_reftree_add_map(&reftree, 2289 vd->vdev_dtl[DTL_MISSING], 1); 2290 space_reftree_add_seg(&reftree, 0, scrub_txg, -1); 2291 space_reftree_add_map(&reftree, 2292 vd->vdev_dtl[DTL_SCRUB], 2); 2293 space_reftree_generate_map(&reftree, 2294 vd->vdev_dtl[DTL_MISSING], 1); 2295 space_reftree_destroy(&reftree); 2296 } 2297 range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL); 2298 range_tree_walk(vd->vdev_dtl[DTL_MISSING], 2299 range_tree_add, vd->vdev_dtl[DTL_PARTIAL]); 2300 if (scrub_done) 2301 range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL); 2302 range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL); 2303 if (!vdev_readable(vd)) 2304 range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL); 2305 else 2306 range_tree_walk(vd->vdev_dtl[DTL_MISSING], 2307 range_tree_add, vd->vdev_dtl[DTL_OUTAGE]); 2308 2309 /* 2310 * If the vdev was resilvering and no longer has any 2311 * DTLs then reset its resilvering flag. 2312 */ 2313 if (vd->vdev_resilver_txg != 0 && 2314 range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) && 2315 range_tree_is_empty(vd->vdev_dtl[DTL_OUTAGE])) 2316 vd->vdev_resilver_txg = 0; 2317 2318 mutex_exit(&vd->vdev_dtl_lock); 2319 2320 if (txg != 0) 2321 vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg); 2322 return; 2323 } 2324 2325 mutex_enter(&vd->vdev_dtl_lock); 2326 for (int t = 0; t < DTL_TYPES; t++) { 2327 /* account for child's outage in parent's missing map */ 2328 int s = (t == DTL_MISSING) ? DTL_OUTAGE: t; 2329 if (t == DTL_SCRUB) 2330 continue; /* leaf vdevs only */ 2331 if (t == DTL_PARTIAL) 2332 minref = 1; /* i.e. non-zero */ 2333 else if (vd->vdev_nparity != 0) 2334 minref = vd->vdev_nparity + 1; /* RAID-Z */ 2335 else 2336 minref = vd->vdev_children; /* any kind of mirror */ 2337 space_reftree_create(&reftree); 2338 for (int c = 0; c < vd->vdev_children; c++) { 2339 vdev_t *cvd = vd->vdev_child[c]; 2340 mutex_enter(&cvd->vdev_dtl_lock); 2341 space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1); 2342 mutex_exit(&cvd->vdev_dtl_lock); 2343 } 2344 space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref); 2345 space_reftree_destroy(&reftree); 2346 } 2347 mutex_exit(&vd->vdev_dtl_lock); 2348 } 2349 2350 int 2351 vdev_dtl_load(vdev_t *vd) 2352 { 2353 spa_t *spa = vd->vdev_spa; 2354 objset_t *mos = spa->spa_meta_objset; 2355 int error = 0; 2356 2357 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) { 2358 ASSERT(vdev_is_concrete(vd)); 2359 2360 error = space_map_open(&vd->vdev_dtl_sm, mos, 2361 vd->vdev_dtl_object, 0, -1ULL, 0); 2362 if (error) 2363 return (error); 2364 ASSERT(vd->vdev_dtl_sm != NULL); 2365 2366 mutex_enter(&vd->vdev_dtl_lock); 2367 2368 /* 2369 * Now that we've opened the space_map we need to update 2370 * the in-core DTL. 2371 */ 2372 space_map_update(vd->vdev_dtl_sm); 2373 2374 error = space_map_load(vd->vdev_dtl_sm, 2375 vd->vdev_dtl[DTL_MISSING], SM_ALLOC); 2376 mutex_exit(&vd->vdev_dtl_lock); 2377 2378 return (error); 2379 } 2380 2381 for (int c = 0; c < vd->vdev_children; c++) { 2382 error = vdev_dtl_load(vd->vdev_child[c]); 2383 if (error != 0) 2384 break; 2385 } 2386 2387 return (error); 2388 } 2389 2390 void 2391 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx) 2392 { 2393 spa_t *spa = vd->vdev_spa; 2394 2395 VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx)); 2396 VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps, 2397 zapobj, tx)); 2398 } 2399 2400 uint64_t 2401 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx) 2402 { 2403 spa_t *spa = vd->vdev_spa; 2404 uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA, 2405 DMU_OT_NONE, 0, tx); 2406 2407 ASSERT(zap != 0); 2408 VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps, 2409 zap, tx)); 2410 2411 return (zap); 2412 } 2413 2414 void 2415 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx) 2416 { 2417 if (vd->vdev_ops != &vdev_hole_ops && 2418 vd->vdev_ops != &vdev_missing_ops && 2419 vd->vdev_ops != &vdev_root_ops && 2420 !vd->vdev_top->vdev_removing) { 2421 if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) { 2422 vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx); 2423 } 2424 if (vd == vd->vdev_top && vd->vdev_top_zap == 0) { 2425 vd->vdev_top_zap = vdev_create_link_zap(vd, tx); 2426 } 2427 } 2428 for (uint64_t i = 0; i < vd->vdev_children; i++) { 2429 vdev_construct_zaps(vd->vdev_child[i], tx); 2430 } 2431 } 2432 2433 void 2434 vdev_dtl_sync(vdev_t *vd, uint64_t txg) 2435 { 2436 spa_t *spa = vd->vdev_spa; 2437 range_tree_t *rt = vd->vdev_dtl[DTL_MISSING]; 2438 objset_t *mos = spa->spa_meta_objset; 2439 range_tree_t *rtsync; 2440 dmu_tx_t *tx; 2441 uint64_t object = space_map_object(vd->vdev_dtl_sm); 2442 2443 ASSERT(vdev_is_concrete(vd)); 2444 ASSERT(vd->vdev_ops->vdev_op_leaf); 2445 2446 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 2447 2448 if (vd->vdev_detached || vd->vdev_top->vdev_removing) { 2449 mutex_enter(&vd->vdev_dtl_lock); 2450 space_map_free(vd->vdev_dtl_sm, tx); 2451 space_map_close(vd->vdev_dtl_sm); 2452 vd->vdev_dtl_sm = NULL; 2453 mutex_exit(&vd->vdev_dtl_lock); 2454 2455 /* 2456 * We only destroy the leaf ZAP for detached leaves or for 2457 * removed log devices. Removed data devices handle leaf ZAP 2458 * cleanup later, once cancellation is no longer possible. 2459 */ 2460 if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached || 2461 vd->vdev_top->vdev_islog)) { 2462 vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx); 2463 vd->vdev_leaf_zap = 0; 2464 } 2465 2466 dmu_tx_commit(tx); 2467 return; 2468 } 2469 2470 if (vd->vdev_dtl_sm == NULL) { 2471 uint64_t new_object; 2472 2473 new_object = space_map_alloc(mos, vdev_dtl_sm_blksz, tx); 2474 VERIFY3U(new_object, !=, 0); 2475 2476 VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object, 2477 0, -1ULL, 0)); 2478 ASSERT(vd->vdev_dtl_sm != NULL); 2479 } 2480 2481 rtsync = range_tree_create(NULL, NULL); 2482 2483 mutex_enter(&vd->vdev_dtl_lock); 2484 range_tree_walk(rt, range_tree_add, rtsync); 2485 mutex_exit(&vd->vdev_dtl_lock); 2486 2487 space_map_truncate(vd->vdev_dtl_sm, vdev_dtl_sm_blksz, tx); 2488 space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, SM_NO_VDEVID, tx); 2489 range_tree_vacate(rtsync, NULL, NULL); 2490 2491 range_tree_destroy(rtsync); 2492 2493 /* 2494 * If the object for the space map has changed then dirty 2495 * the top level so that we update the config. 2496 */ 2497 if (object != space_map_object(vd->vdev_dtl_sm)) { 2498 vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, " 2499 "new object %llu", (u_longlong_t)txg, spa_name(spa), 2500 (u_longlong_t)object, 2501 (u_longlong_t)space_map_object(vd->vdev_dtl_sm)); 2502 vdev_config_dirty(vd->vdev_top); 2503 } 2504 2505 dmu_tx_commit(tx); 2506 2507 mutex_enter(&vd->vdev_dtl_lock); 2508 space_map_update(vd->vdev_dtl_sm); 2509 mutex_exit(&vd->vdev_dtl_lock); 2510 } 2511 2512 /* 2513 * Determine whether the specified vdev can be offlined/detached/removed 2514 * without losing data. 2515 */ 2516 boolean_t 2517 vdev_dtl_required(vdev_t *vd) 2518 { 2519 spa_t *spa = vd->vdev_spa; 2520 vdev_t *tvd = vd->vdev_top; 2521 uint8_t cant_read = vd->vdev_cant_read; 2522 boolean_t required; 2523 2524 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL); 2525 2526 if (vd == spa->spa_root_vdev || vd == tvd) 2527 return (B_TRUE); 2528 2529 /* 2530 * Temporarily mark the device as unreadable, and then determine 2531 * whether this results in any DTL outages in the top-level vdev. 2532 * If not, we can safely offline/detach/remove the device. 2533 */ 2534 vd->vdev_cant_read = B_TRUE; 2535 vdev_dtl_reassess(tvd, 0, 0, B_FALSE); 2536 required = !vdev_dtl_empty(tvd, DTL_OUTAGE); 2537 vd->vdev_cant_read = cant_read; 2538 vdev_dtl_reassess(tvd, 0, 0, B_FALSE); 2539 2540 if (!required && zio_injection_enabled) 2541 required = !!zio_handle_device_injection(vd, NULL, ECHILD); 2542 2543 return (required); 2544 } 2545 2546 /* 2547 * Determine if resilver is needed, and if so the txg range. 2548 */ 2549 boolean_t 2550 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp) 2551 { 2552 boolean_t needed = B_FALSE; 2553 uint64_t thismin = UINT64_MAX; 2554 uint64_t thismax = 0; 2555 2556 if (vd->vdev_children == 0) { 2557 mutex_enter(&vd->vdev_dtl_lock); 2558 if (!range_tree_is_empty(vd->vdev_dtl[DTL_MISSING]) && 2559 vdev_writeable(vd)) { 2560 2561 thismin = vdev_dtl_min(vd); 2562 thismax = vdev_dtl_max(vd); 2563 needed = B_TRUE; 2564 } 2565 mutex_exit(&vd->vdev_dtl_lock); 2566 } else { 2567 for (int c = 0; c < vd->vdev_children; c++) { 2568 vdev_t *cvd = vd->vdev_child[c]; 2569 uint64_t cmin, cmax; 2570 2571 if (vdev_resilver_needed(cvd, &cmin, &cmax)) { 2572 thismin = MIN(thismin, cmin); 2573 thismax = MAX(thismax, cmax); 2574 needed = B_TRUE; 2575 } 2576 } 2577 } 2578 2579 if (needed && minp) { 2580 *minp = thismin; 2581 *maxp = thismax; 2582 } 2583 return (needed); 2584 } 2585 2586 /* 2587 * Gets the checkpoint space map object from the vdev's ZAP. 2588 * Returns the spacemap object, or 0 if it wasn't in the ZAP 2589 * or the ZAP doesn't exist yet. 2590 */ 2591 int 2592 vdev_checkpoint_sm_object(vdev_t *vd) 2593 { 2594 ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER)); 2595 if (vd->vdev_top_zap == 0) { 2596 return (0); 2597 } 2598 2599 uint64_t sm_obj = 0; 2600 int err = zap_lookup(spa_meta_objset(vd->vdev_spa), vd->vdev_top_zap, 2601 VDEV_TOP_ZAP_POOL_CHECKPOINT_SM, sizeof (uint64_t), 1, &sm_obj); 2602 2603 ASSERT(err == 0 || err == ENOENT); 2604 2605 return (sm_obj); 2606 } 2607 2608 int 2609 vdev_load(vdev_t *vd) 2610 { 2611 int error = 0; 2612 /* 2613 * Recursively load all children. 2614 */ 2615 for (int c = 0; c < vd->vdev_children; c++) { 2616 error = vdev_load(vd->vdev_child[c]); 2617 if (error != 0) { 2618 return (error); 2619 } 2620 } 2621 2622 vdev_set_deflate_ratio(vd); 2623 2624 /* 2625 * If this is a top-level vdev, initialize its metaslabs. 2626 */ 2627 if (vd == vd->vdev_top && vdev_is_concrete(vd)) { 2628 if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) { 2629 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 2630 VDEV_AUX_CORRUPT_DATA); 2631 vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, " 2632 "asize=%llu", (u_longlong_t)vd->vdev_ashift, 2633 (u_longlong_t)vd->vdev_asize); 2634 return (SET_ERROR(ENXIO)); 2635 } else if ((error = vdev_metaslab_init(vd, 0)) != 0) { 2636 vdev_dbgmsg(vd, "vdev_load: metaslab_init failed " 2637 "[error=%d]", error); 2638 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 2639 VDEV_AUX_CORRUPT_DATA); 2640 return (error); 2641 } 2642 2643 uint64_t checkpoint_sm_obj = vdev_checkpoint_sm_object(vd); 2644 if (checkpoint_sm_obj != 0) { 2645 objset_t *mos = spa_meta_objset(vd->vdev_spa); 2646 ASSERT(vd->vdev_asize != 0); 2647 ASSERT3P(vd->vdev_checkpoint_sm, ==, NULL); 2648 2649 if ((error = space_map_open(&vd->vdev_checkpoint_sm, 2650 mos, checkpoint_sm_obj, 0, vd->vdev_asize, 2651 vd->vdev_ashift))) { 2652 vdev_dbgmsg(vd, "vdev_load: space_map_open " 2653 "failed for checkpoint spacemap (obj %llu) " 2654 "[error=%d]", 2655 (u_longlong_t)checkpoint_sm_obj, error); 2656 return (error); 2657 } 2658 ASSERT3P(vd->vdev_checkpoint_sm, !=, NULL); 2659 space_map_update(vd->vdev_checkpoint_sm); 2660 2661 /* 2662 * Since the checkpoint_sm contains free entries 2663 * exclusively we can use sm_alloc to indicate the 2664 * culmulative checkpointed space that has been freed. 2665 */ 2666 vd->vdev_stat.vs_checkpoint_space = 2667 -vd->vdev_checkpoint_sm->sm_alloc; 2668 vd->vdev_spa->spa_checkpoint_info.sci_dspace += 2669 vd->vdev_stat.vs_checkpoint_space; 2670 } 2671 } 2672 2673 /* 2674 * If this is a leaf vdev, load its DTL. 2675 */ 2676 if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) { 2677 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 2678 VDEV_AUX_CORRUPT_DATA); 2679 vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed " 2680 "[error=%d]", error); 2681 return (error); 2682 } 2683 2684 uint64_t obsolete_sm_object = vdev_obsolete_sm_object(vd); 2685 if (obsolete_sm_object != 0) { 2686 objset_t *mos = vd->vdev_spa->spa_meta_objset; 2687 ASSERT(vd->vdev_asize != 0); 2688 ASSERT3P(vd->vdev_obsolete_sm, ==, NULL); 2689 2690 if ((error = space_map_open(&vd->vdev_obsolete_sm, mos, 2691 obsolete_sm_object, 0, vd->vdev_asize, 0))) { 2692 vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN, 2693 VDEV_AUX_CORRUPT_DATA); 2694 vdev_dbgmsg(vd, "vdev_load: space_map_open failed for " 2695 "obsolete spacemap (obj %llu) [error=%d]", 2696 (u_longlong_t)obsolete_sm_object, error); 2697 return (error); 2698 } 2699 space_map_update(vd->vdev_obsolete_sm); 2700 } 2701 2702 return (0); 2703 } 2704 2705 /* 2706 * The special vdev case is used for hot spares and l2cache devices. Its 2707 * sole purpose it to set the vdev state for the associated vdev. To do this, 2708 * we make sure that we can open the underlying device, then try to read the 2709 * label, and make sure that the label is sane and that it hasn't been 2710 * repurposed to another pool. 2711 */ 2712 int 2713 vdev_validate_aux(vdev_t *vd) 2714 { 2715 nvlist_t *label; 2716 uint64_t guid, version; 2717 uint64_t state; 2718 2719 if (!vdev_readable(vd)) 2720 return (0); 2721 2722 if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) { 2723 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 2724 VDEV_AUX_CORRUPT_DATA); 2725 return (-1); 2726 } 2727 2728 if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 || 2729 !SPA_VERSION_IS_SUPPORTED(version) || 2730 nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 || 2731 guid != vd->vdev_guid || 2732 nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) { 2733 vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN, 2734 VDEV_AUX_CORRUPT_DATA); 2735 nvlist_free(label); 2736 return (-1); 2737 } 2738 2739 /* 2740 * We don't actually check the pool state here. If it's in fact in 2741 * use by another pool, we update this fact on the fly when requested. 2742 */ 2743 nvlist_free(label); 2744 return (0); 2745 } 2746 2747 /* 2748 * Free the objects used to store this vdev's spacemaps, and the array 2749 * that points to them. 2750 */ 2751 void 2752 vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx) 2753 { 2754 if (vd->vdev_ms_array == 0) 2755 return; 2756 2757 objset_t *mos = vd->vdev_spa->spa_meta_objset; 2758 uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift; 2759 size_t array_bytes = array_count * sizeof (uint64_t); 2760 uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP); 2761 VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0, 2762 array_bytes, smobj_array, 0)); 2763 2764 for (uint64_t i = 0; i < array_count; i++) { 2765 uint64_t smobj = smobj_array[i]; 2766 if (smobj == 0) 2767 continue; 2768 2769 space_map_free_obj(mos, smobj, tx); 2770 } 2771 2772 kmem_free(smobj_array, array_bytes); 2773 VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx)); 2774 vd->vdev_ms_array = 0; 2775 } 2776 2777 static void 2778 vdev_remove_empty(vdev_t *vd, uint64_t txg) 2779 { 2780 spa_t *spa = vd->vdev_spa; 2781 dmu_tx_t *tx; 2782 2783 ASSERT(vd == vd->vdev_top); 2784 ASSERT3U(txg, ==, spa_syncing_txg(spa)); 2785 2786 if (vd->vdev_ms != NULL) { 2787 metaslab_group_t *mg = vd->vdev_mg; 2788 2789 metaslab_group_histogram_verify(mg); 2790 metaslab_class_histogram_verify(mg->mg_class); 2791 2792 for (int m = 0; m < vd->vdev_ms_count; m++) { 2793 metaslab_t *msp = vd->vdev_ms[m]; 2794 2795 if (msp == NULL || msp->ms_sm == NULL) 2796 continue; 2797 2798 mutex_enter(&msp->ms_lock); 2799 /* 2800 * If the metaslab was not loaded when the vdev 2801 * was removed then the histogram accounting may 2802 * not be accurate. Update the histogram information 2803 * here so that we ensure that the metaslab group 2804 * and metaslab class are up-to-date. 2805 */ 2806 metaslab_group_histogram_remove(mg, msp); 2807 2808 VERIFY0(space_map_allocated(msp->ms_sm)); 2809 space_map_close(msp->ms_sm); 2810 msp->ms_sm = NULL; 2811 mutex_exit(&msp->ms_lock); 2812 } 2813 2814 if (vd->vdev_checkpoint_sm != NULL) { 2815 ASSERT(spa_has_checkpoint(spa)); 2816 space_map_close(vd->vdev_checkpoint_sm); 2817 vd->vdev_checkpoint_sm = NULL; 2818 } 2819 2820 metaslab_group_histogram_verify(mg); 2821 metaslab_class_histogram_verify(mg->mg_class); 2822 for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++) 2823 ASSERT0(mg->mg_histogram[i]); 2824 } 2825 2826 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg); 2827 vdev_destroy_spacemaps(vd, tx); 2828 2829 if (vd->vdev_islog && vd->vdev_top_zap != 0) { 2830 vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx); 2831 vd->vdev_top_zap = 0; 2832 } 2833 dmu_tx_commit(tx); 2834 } 2835 2836 void 2837 vdev_sync_done(vdev_t *vd, uint64_t txg) 2838 { 2839 metaslab_t *msp; 2840 boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg)); 2841 2842 ASSERT(vdev_is_concrete(vd)); 2843 2844 while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg))) 2845 metaslab_sync_done(msp, txg); 2846 2847 if (reassess) 2848 metaslab_sync_reassess(vd->vdev_mg); 2849 } 2850 2851 void 2852 vdev_sync(vdev_t *vd, uint64_t txg) 2853 { 2854 spa_t *spa = vd->vdev_spa; 2855 vdev_t *lvd; 2856 metaslab_t *msp; 2857 dmu_tx_t *tx; 2858 2859 if (range_tree_space(vd->vdev_obsolete_segments) > 0) { 2860 dmu_tx_t *tx; 2861 2862 ASSERT(vd->vdev_removing || 2863 vd->vdev_ops == &vdev_indirect_ops); 2864 2865 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 2866 vdev_indirect_sync_obsolete(vd, tx); 2867 dmu_tx_commit(tx); 2868 2869 /* 2870 * If the vdev is indirect, it can't have dirty 2871 * metaslabs or DTLs. 2872 */ 2873 if (vd->vdev_ops == &vdev_indirect_ops) { 2874 ASSERT(txg_list_empty(&vd->vdev_ms_list, txg)); 2875 ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg)); 2876 return; 2877 } 2878 } 2879 2880 ASSERT(vdev_is_concrete(vd)); 2881 2882 if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 && 2883 !vd->vdev_removing) { 2884 ASSERT(vd == vd->vdev_top); 2885 ASSERT0(vd->vdev_indirect_config.vic_mapping_object); 2886 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 2887 vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset, 2888 DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx); 2889 ASSERT(vd->vdev_ms_array != 0); 2890 vdev_config_dirty(vd); 2891 dmu_tx_commit(tx); 2892 } 2893 2894 while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) { 2895 metaslab_sync(msp, txg); 2896 (void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg)); 2897 } 2898 2899 while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL) 2900 vdev_dtl_sync(lvd, txg); 2901 2902 /* 2903 * Remove the metadata associated with this vdev once it's empty. 2904 * Note that this is typically used for log/cache device removal; 2905 * we don't empty toplevel vdevs when removing them. But if 2906 * a toplevel happens to be emptied, this is not harmful. 2907 */ 2908 if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing) { 2909 vdev_remove_empty(vd, txg); 2910 } 2911 2912 (void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg)); 2913 } 2914 2915 uint64_t 2916 vdev_psize_to_asize(vdev_t *vd, uint64_t psize) 2917 { 2918 return (vd->vdev_ops->vdev_op_asize(vd, psize)); 2919 } 2920 2921 /* 2922 * Mark the given vdev faulted. A faulted vdev behaves as if the device could 2923 * not be opened, and no I/O is attempted. 2924 */ 2925 int 2926 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux) 2927 { 2928 vdev_t *vd, *tvd; 2929 2930 spa_vdev_state_enter(spa, SCL_NONE); 2931 2932 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) 2933 return (spa_vdev_state_exit(spa, NULL, ENODEV)); 2934 2935 if (!vd->vdev_ops->vdev_op_leaf) 2936 return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); 2937 2938 tvd = vd->vdev_top; 2939 2940 /* 2941 * We don't directly use the aux state here, but if we do a 2942 * vdev_reopen(), we need this value to be present to remember why we 2943 * were faulted. 2944 */ 2945 vd->vdev_label_aux = aux; 2946 2947 /* 2948 * Faulted state takes precedence over degraded. 2949 */ 2950 vd->vdev_delayed_close = B_FALSE; 2951 vd->vdev_faulted = 1ULL; 2952 vd->vdev_degraded = 0ULL; 2953 vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux); 2954 2955 /* 2956 * If this device has the only valid copy of the data, then 2957 * back off and simply mark the vdev as degraded instead. 2958 */ 2959 if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) { 2960 vd->vdev_degraded = 1ULL; 2961 vd->vdev_faulted = 0ULL; 2962 2963 /* 2964 * If we reopen the device and it's not dead, only then do we 2965 * mark it degraded. 2966 */ 2967 vdev_reopen(tvd); 2968 2969 if (vdev_readable(vd)) 2970 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux); 2971 } 2972 2973 return (spa_vdev_state_exit(spa, vd, 0)); 2974 } 2975 2976 /* 2977 * Mark the given vdev degraded. A degraded vdev is purely an indication to the 2978 * user that something is wrong. The vdev continues to operate as normal as far 2979 * as I/O is concerned. 2980 */ 2981 int 2982 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux) 2983 { 2984 vdev_t *vd; 2985 2986 spa_vdev_state_enter(spa, SCL_NONE); 2987 2988 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) 2989 return (spa_vdev_state_exit(spa, NULL, ENODEV)); 2990 2991 if (!vd->vdev_ops->vdev_op_leaf) 2992 return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); 2993 2994 /* 2995 * If the vdev is already faulted, then don't do anything. 2996 */ 2997 if (vd->vdev_faulted || vd->vdev_degraded) 2998 return (spa_vdev_state_exit(spa, NULL, 0)); 2999 3000 vd->vdev_degraded = 1ULL; 3001 if (!vdev_is_dead(vd)) 3002 vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, 3003 aux); 3004 3005 return (spa_vdev_state_exit(spa, vd, 0)); 3006 } 3007 3008 /* 3009 * Online the given vdev. 3010 * 3011 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things. First, any attached 3012 * spare device should be detached when the device finishes resilvering. 3013 * Second, the online should be treated like a 'test' online case, so no FMA 3014 * events are generated if the device fails to open. 3015 */ 3016 int 3017 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate) 3018 { 3019 vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev; 3020 boolean_t wasoffline; 3021 vdev_state_t oldstate; 3022 3023 spa_vdev_state_enter(spa, SCL_NONE); 3024 3025 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) 3026 return (spa_vdev_state_exit(spa, NULL, ENODEV)); 3027 3028 if (!vd->vdev_ops->vdev_op_leaf) 3029 return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); 3030 3031 wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline); 3032 oldstate = vd->vdev_state; 3033 3034 tvd = vd->vdev_top; 3035 vd->vdev_offline = B_FALSE; 3036 vd->vdev_tmpoffline = B_FALSE; 3037 vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE); 3038 vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT); 3039 3040 /* XXX - L2ARC 1.0 does not support expansion */ 3041 if (!vd->vdev_aux) { 3042 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent) 3043 pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND); 3044 } 3045 3046 vdev_reopen(tvd); 3047 vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE; 3048 3049 if (!vd->vdev_aux) { 3050 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent) 3051 pvd->vdev_expanding = B_FALSE; 3052 } 3053 3054 if (newstate) 3055 *newstate = vd->vdev_state; 3056 if ((flags & ZFS_ONLINE_UNSPARE) && 3057 !vdev_is_dead(vd) && vd->vdev_parent && 3058 vd->vdev_parent->vdev_ops == &vdev_spare_ops && 3059 vd->vdev_parent->vdev_child[0] == vd) 3060 vd->vdev_unspare = B_TRUE; 3061 3062 if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) { 3063 3064 /* XXX - L2ARC 1.0 does not support expansion */ 3065 if (vd->vdev_aux) 3066 return (spa_vdev_state_exit(spa, vd, ENOTSUP)); 3067 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 3068 } 3069 3070 if (wasoffline || 3071 (oldstate < VDEV_STATE_DEGRADED && 3072 vd->vdev_state >= VDEV_STATE_DEGRADED)) 3073 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE); 3074 3075 return (spa_vdev_state_exit(spa, vd, 0)); 3076 } 3077 3078 static int 3079 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags) 3080 { 3081 vdev_t *vd, *tvd; 3082 int error = 0; 3083 uint64_t generation; 3084 metaslab_group_t *mg; 3085 3086 top: 3087 spa_vdev_state_enter(spa, SCL_ALLOC); 3088 3089 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) 3090 return (spa_vdev_state_exit(spa, NULL, ENODEV)); 3091 3092 if (!vd->vdev_ops->vdev_op_leaf) 3093 return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); 3094 3095 tvd = vd->vdev_top; 3096 mg = tvd->vdev_mg; 3097 generation = spa->spa_config_generation + 1; 3098 3099 /* 3100 * If the device isn't already offline, try to offline it. 3101 */ 3102 if (!vd->vdev_offline) { 3103 /* 3104 * If this device has the only valid copy of some data, 3105 * don't allow it to be offlined. Log devices are always 3106 * expendable. 3107 */ 3108 if (!tvd->vdev_islog && vd->vdev_aux == NULL && 3109 vdev_dtl_required(vd)) 3110 return (spa_vdev_state_exit(spa, NULL, EBUSY)); 3111 3112 /* 3113 * If the top-level is a slog and it has had allocations 3114 * then proceed. We check that the vdev's metaslab group 3115 * is not NULL since it's possible that we may have just 3116 * added this vdev but not yet initialized its metaslabs. 3117 */ 3118 if (tvd->vdev_islog && mg != NULL) { 3119 /* 3120 * Prevent any future allocations. 3121 */ 3122 metaslab_group_passivate(mg); 3123 (void) spa_vdev_state_exit(spa, vd, 0); 3124 3125 error = spa_reset_logs(spa); 3126 3127 /* 3128 * If the log device was successfully reset but has 3129 * checkpointed data, do not offline it. 3130 */ 3131 if (error == 0 && 3132 tvd->vdev_checkpoint_sm != NULL) { 3133 ASSERT3U(tvd->vdev_checkpoint_sm->sm_alloc, 3134 !=, 0); 3135 error = ZFS_ERR_CHECKPOINT_EXISTS; 3136 } 3137 3138 spa_vdev_state_enter(spa, SCL_ALLOC); 3139 3140 /* 3141 * Check to see if the config has changed. 3142 */ 3143 if (error || generation != spa->spa_config_generation) { 3144 metaslab_group_activate(mg); 3145 if (error) 3146 return (spa_vdev_state_exit(spa, 3147 vd, error)); 3148 (void) spa_vdev_state_exit(spa, vd, 0); 3149 goto top; 3150 } 3151 ASSERT0(tvd->vdev_stat.vs_alloc); 3152 } 3153 3154 /* 3155 * Offline this device and reopen its top-level vdev. 3156 * If the top-level vdev is a log device then just offline 3157 * it. Otherwise, if this action results in the top-level 3158 * vdev becoming unusable, undo it and fail the request. 3159 */ 3160 vd->vdev_offline = B_TRUE; 3161 vdev_reopen(tvd); 3162 3163 if (!tvd->vdev_islog && vd->vdev_aux == NULL && 3164 vdev_is_dead(tvd)) { 3165 vd->vdev_offline = B_FALSE; 3166 vdev_reopen(tvd); 3167 return (spa_vdev_state_exit(spa, NULL, EBUSY)); 3168 } 3169 3170 /* 3171 * Add the device back into the metaslab rotor so that 3172 * once we online the device it's open for business. 3173 */ 3174 if (tvd->vdev_islog && mg != NULL) 3175 metaslab_group_activate(mg); 3176 } 3177 3178 vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY); 3179 3180 return (spa_vdev_state_exit(spa, vd, 0)); 3181 } 3182 3183 int 3184 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags) 3185 { 3186 int error; 3187 3188 mutex_enter(&spa->spa_vdev_top_lock); 3189 error = vdev_offline_locked(spa, guid, flags); 3190 mutex_exit(&spa->spa_vdev_top_lock); 3191 3192 return (error); 3193 } 3194 3195 /* 3196 * Clear the error counts associated with this vdev. Unlike vdev_online() and 3197 * vdev_offline(), we assume the spa config is locked. We also clear all 3198 * children. If 'vd' is NULL, then the user wants to clear all vdevs. 3199 */ 3200 void 3201 vdev_clear(spa_t *spa, vdev_t *vd) 3202 { 3203 vdev_t *rvd = spa->spa_root_vdev; 3204 3205 ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL); 3206 3207 if (vd == NULL) 3208 vd = rvd; 3209 3210 vd->vdev_stat.vs_read_errors = 0; 3211 vd->vdev_stat.vs_write_errors = 0; 3212 vd->vdev_stat.vs_checksum_errors = 0; 3213 3214 for (int c = 0; c < vd->vdev_children; c++) 3215 vdev_clear(spa, vd->vdev_child[c]); 3216 3217 /* 3218 * It makes no sense to "clear" an indirect vdev. 3219 */ 3220 if (!vdev_is_concrete(vd)) 3221 return; 3222 3223 /* 3224 * If we're in the FAULTED state or have experienced failed I/O, then 3225 * clear the persistent state and attempt to reopen the device. We 3226 * also mark the vdev config dirty, so that the new faulted state is 3227 * written out to disk. 3228 */ 3229 if (vd->vdev_faulted || vd->vdev_degraded || 3230 !vdev_readable(vd) || !vdev_writeable(vd)) { 3231 3232 /* 3233 * When reopening in reponse to a clear event, it may be due to 3234 * a fmadm repair request. In this case, if the device is 3235 * still broken, we want to still post the ereport again. 3236 */ 3237 vd->vdev_forcefault = B_TRUE; 3238 3239 vd->vdev_faulted = vd->vdev_degraded = 0ULL; 3240 vd->vdev_cant_read = B_FALSE; 3241 vd->vdev_cant_write = B_FALSE; 3242 3243 vdev_reopen(vd == rvd ? rvd : vd->vdev_top); 3244 3245 vd->vdev_forcefault = B_FALSE; 3246 3247 if (vd != rvd && vdev_writeable(vd->vdev_top)) 3248 vdev_state_dirty(vd->vdev_top); 3249 3250 if (vd->vdev_aux == NULL && !vdev_is_dead(vd)) 3251 spa_async_request(spa, SPA_ASYNC_RESILVER); 3252 3253 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR); 3254 } 3255 3256 /* 3257 * When clearing a FMA-diagnosed fault, we always want to 3258 * unspare the device, as we assume that the original spare was 3259 * done in response to the FMA fault. 3260 */ 3261 if (!vdev_is_dead(vd) && vd->vdev_parent != NULL && 3262 vd->vdev_parent->vdev_ops == &vdev_spare_ops && 3263 vd->vdev_parent->vdev_child[0] == vd) 3264 vd->vdev_unspare = B_TRUE; 3265 } 3266 3267 boolean_t 3268 vdev_is_dead(vdev_t *vd) 3269 { 3270 /* 3271 * Holes and missing devices are always considered "dead". 3272 * This simplifies the code since we don't have to check for 3273 * these types of devices in the various code paths. 3274 * Instead we rely on the fact that we skip over dead devices 3275 * before issuing I/O to them. 3276 */ 3277 return (vd->vdev_state < VDEV_STATE_DEGRADED || 3278 vd->vdev_ops == &vdev_hole_ops || 3279 vd->vdev_ops == &vdev_missing_ops); 3280 } 3281 3282 boolean_t 3283 vdev_readable(vdev_t *vd) 3284 { 3285 return (!vdev_is_dead(vd) && !vd->vdev_cant_read); 3286 } 3287 3288 boolean_t 3289 vdev_writeable(vdev_t *vd) 3290 { 3291 return (!vdev_is_dead(vd) && !vd->vdev_cant_write && 3292 vdev_is_concrete(vd)); 3293 } 3294 3295 boolean_t 3296 vdev_allocatable(vdev_t *vd) 3297 { 3298 uint64_t state = vd->vdev_state; 3299 3300 /* 3301 * We currently allow allocations from vdevs which may be in the 3302 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device 3303 * fails to reopen then we'll catch it later when we're holding 3304 * the proper locks. Note that we have to get the vdev state 3305 * in a local variable because although it changes atomically, 3306 * we're asking two separate questions about it. 3307 */ 3308 return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) && 3309 !vd->vdev_cant_write && vdev_is_concrete(vd) && 3310 vd->vdev_mg->mg_initialized); 3311 } 3312 3313 boolean_t 3314 vdev_accessible(vdev_t *vd, zio_t *zio) 3315 { 3316 ASSERT(zio->io_vd == vd); 3317 3318 if (vdev_is_dead(vd) || vd->vdev_remove_wanted) 3319 return (B_FALSE); 3320 3321 if (zio->io_type == ZIO_TYPE_READ) 3322 return (!vd->vdev_cant_read); 3323 3324 if (zio->io_type == ZIO_TYPE_WRITE) 3325 return (!vd->vdev_cant_write); 3326 3327 return (B_TRUE); 3328 } 3329 3330 boolean_t 3331 vdev_is_spacemap_addressable(vdev_t *vd) 3332 { 3333 /* 3334 * Assuming 47 bits of the space map entry dedicated for the entry's 3335 * offset (see description in space_map.h), we calculate the maximum 3336 * address that can be described by a space map entry for the given 3337 * device. 3338 */ 3339 uint64_t shift = vd->vdev_ashift + 47; 3340 3341 if (shift >= 63) /* detect potential overflow */ 3342 return (B_TRUE); 3343 3344 return (vd->vdev_asize < (1ULL << shift)); 3345 } 3346 3347 /* 3348 * Get statistics for the given vdev. 3349 */ 3350 void 3351 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs) 3352 { 3353 spa_t *spa = vd->vdev_spa; 3354 vdev_t *rvd = spa->spa_root_vdev; 3355 vdev_t *tvd = vd->vdev_top; 3356 3357 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0); 3358 3359 mutex_enter(&vd->vdev_stat_lock); 3360 bcopy(&vd->vdev_stat, vs, sizeof (*vs)); 3361 vs->vs_timestamp = gethrtime() - vs->vs_timestamp; 3362 vs->vs_state = vd->vdev_state; 3363 vs->vs_rsize = vdev_get_min_asize(vd); 3364 if (vd->vdev_ops->vdev_op_leaf) 3365 vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE; 3366 /* 3367 * Report expandable space on top-level, non-auxillary devices only. 3368 * The expandable space is reported in terms of metaslab sized units 3369 * since that determines how much space the pool can expand. 3370 */ 3371 if (vd->vdev_aux == NULL && tvd != NULL) { 3372 vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize - 3373 spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift); 3374 } 3375 if (vd->vdev_aux == NULL && vd == vd->vdev_top && 3376 vdev_is_concrete(vd)) { 3377 vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation; 3378 } 3379 3380 /* 3381 * If we're getting stats on the root vdev, aggregate the I/O counts 3382 * over all top-level vdevs (i.e. the direct children of the root). 3383 */ 3384 if (vd == rvd) { 3385 for (int c = 0; c < rvd->vdev_children; c++) { 3386 vdev_t *cvd = rvd->vdev_child[c]; 3387 vdev_stat_t *cvs = &cvd->vdev_stat; 3388 3389 for (int t = 0; t < ZIO_TYPES; t++) { 3390 vs->vs_ops[t] += cvs->vs_ops[t]; 3391 vs->vs_bytes[t] += cvs->vs_bytes[t]; 3392 } 3393 cvs->vs_scan_removing = cvd->vdev_removing; 3394 } 3395 } 3396 mutex_exit(&vd->vdev_stat_lock); 3397 } 3398 3399 void 3400 vdev_clear_stats(vdev_t *vd) 3401 { 3402 mutex_enter(&vd->vdev_stat_lock); 3403 vd->vdev_stat.vs_space = 0; 3404 vd->vdev_stat.vs_dspace = 0; 3405 vd->vdev_stat.vs_alloc = 0; 3406 mutex_exit(&vd->vdev_stat_lock); 3407 } 3408 3409 void 3410 vdev_scan_stat_init(vdev_t *vd) 3411 { 3412 vdev_stat_t *vs = &vd->vdev_stat; 3413 3414 for (int c = 0; c < vd->vdev_children; c++) 3415 vdev_scan_stat_init(vd->vdev_child[c]); 3416 3417 mutex_enter(&vd->vdev_stat_lock); 3418 vs->vs_scan_processed = 0; 3419 mutex_exit(&vd->vdev_stat_lock); 3420 } 3421 3422 void 3423 vdev_stat_update(zio_t *zio, uint64_t psize) 3424 { 3425 spa_t *spa = zio->io_spa; 3426 vdev_t *rvd = spa->spa_root_vdev; 3427 vdev_t *vd = zio->io_vd ? zio->io_vd : rvd; 3428 vdev_t *pvd; 3429 uint64_t txg = zio->io_txg; 3430 vdev_stat_t *vs = &vd->vdev_stat; 3431 zio_type_t type = zio->io_type; 3432 int flags = zio->io_flags; 3433 3434 /* 3435 * If this i/o is a gang leader, it didn't do any actual work. 3436 */ 3437 if (zio->io_gang_tree) 3438 return; 3439 3440 if (zio->io_error == 0) { 3441 /* 3442 * If this is a root i/o, don't count it -- we've already 3443 * counted the top-level vdevs, and vdev_get_stats() will 3444 * aggregate them when asked. This reduces contention on 3445 * the root vdev_stat_lock and implicitly handles blocks 3446 * that compress away to holes, for which there is no i/o. 3447 * (Holes never create vdev children, so all the counters 3448 * remain zero, which is what we want.) 3449 * 3450 * Note: this only applies to successful i/o (io_error == 0) 3451 * because unlike i/o counts, errors are not additive. 3452 * When reading a ditto block, for example, failure of 3453 * one top-level vdev does not imply a root-level error. 3454 */ 3455 if (vd == rvd) 3456 return; 3457 3458 ASSERT(vd == zio->io_vd); 3459 3460 if (flags & ZIO_FLAG_IO_BYPASS) 3461 return; 3462 3463 mutex_enter(&vd->vdev_stat_lock); 3464 3465 if (flags & ZIO_FLAG_IO_REPAIR) { 3466 if (flags & ZIO_FLAG_SCAN_THREAD) { 3467 dsl_scan_phys_t *scn_phys = 3468 &spa->spa_dsl_pool->dp_scan->scn_phys; 3469 uint64_t *processed = &scn_phys->scn_processed; 3470 3471 /* XXX cleanup? */ 3472 if (vd->vdev_ops->vdev_op_leaf) 3473 atomic_add_64(processed, psize); 3474 vs->vs_scan_processed += psize; 3475 } 3476 3477 if (flags & ZIO_FLAG_SELF_HEAL) 3478 vs->vs_self_healed += psize; 3479 } 3480 3481 vs->vs_ops[type]++; 3482 vs->vs_bytes[type] += psize; 3483 3484 mutex_exit(&vd->vdev_stat_lock); 3485 return; 3486 } 3487 3488 if (flags & ZIO_FLAG_SPECULATIVE) 3489 return; 3490 3491 /* 3492 * If this is an I/O error that is going to be retried, then ignore the 3493 * error. Otherwise, the user may interpret B_FAILFAST I/O errors as 3494 * hard errors, when in reality they can happen for any number of 3495 * innocuous reasons (bus resets, MPxIO link failure, etc). 3496 */ 3497 if (zio->io_error == EIO && 3498 !(zio->io_flags & ZIO_FLAG_IO_RETRY)) 3499 return; 3500 3501 /* 3502 * Intent logs writes won't propagate their error to the root 3503 * I/O so don't mark these types of failures as pool-level 3504 * errors. 3505 */ 3506 if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) 3507 return; 3508 3509 mutex_enter(&vd->vdev_stat_lock); 3510 if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) { 3511 if (zio->io_error == ECKSUM) 3512 vs->vs_checksum_errors++; 3513 else 3514 vs->vs_read_errors++; 3515 } 3516 if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd)) 3517 vs->vs_write_errors++; 3518 mutex_exit(&vd->vdev_stat_lock); 3519 3520 if (spa->spa_load_state == SPA_LOAD_NONE && 3521 type == ZIO_TYPE_WRITE && txg != 0 && 3522 (!(flags & ZIO_FLAG_IO_REPAIR) || 3523 (flags & ZIO_FLAG_SCAN_THREAD) || 3524 spa->spa_claiming)) { 3525 /* 3526 * This is either a normal write (not a repair), or it's 3527 * a repair induced by the scrub thread, or it's a repair 3528 * made by zil_claim() during spa_load() in the first txg. 3529 * In the normal case, we commit the DTL change in the same 3530 * txg as the block was born. In the scrub-induced repair 3531 * case, we know that scrubs run in first-pass syncing context, 3532 * so we commit the DTL change in spa_syncing_txg(spa). 3533 * In the zil_claim() case, we commit in spa_first_txg(spa). 3534 * 3535 * We currently do not make DTL entries for failed spontaneous 3536 * self-healing writes triggered by normal (non-scrubbing) 3537 * reads, because we have no transactional context in which to 3538 * do so -- and it's not clear that it'd be desirable anyway. 3539 */ 3540 if (vd->vdev_ops->vdev_op_leaf) { 3541 uint64_t commit_txg = txg; 3542 if (flags & ZIO_FLAG_SCAN_THREAD) { 3543 ASSERT(flags & ZIO_FLAG_IO_REPAIR); 3544 ASSERT(spa_sync_pass(spa) == 1); 3545 vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1); 3546 commit_txg = spa_syncing_txg(spa); 3547 } else if (spa->spa_claiming) { 3548 ASSERT(flags & ZIO_FLAG_IO_REPAIR); 3549 commit_txg = spa_first_txg(spa); 3550 } 3551 ASSERT(commit_txg >= spa_syncing_txg(spa)); 3552 if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1)) 3553 return; 3554 for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent) 3555 vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1); 3556 vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg); 3557 } 3558 if (vd != rvd) 3559 vdev_dtl_dirty(vd, DTL_MISSING, txg, 1); 3560 } 3561 } 3562 3563 /* 3564 * Update the in-core space usage stats for this vdev, its metaslab class, 3565 * and the root vdev. 3566 */ 3567 void 3568 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta, 3569 int64_t space_delta) 3570 { 3571 int64_t dspace_delta = space_delta; 3572 spa_t *spa = vd->vdev_spa; 3573 vdev_t *rvd = spa->spa_root_vdev; 3574 metaslab_group_t *mg = vd->vdev_mg; 3575 metaslab_class_t *mc = mg ? mg->mg_class : NULL; 3576 3577 ASSERT(vd == vd->vdev_top); 3578 3579 /* 3580 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion 3581 * factor. We must calculate this here and not at the root vdev 3582 * because the root vdev's psize-to-asize is simply the max of its 3583 * childrens', thus not accurate enough for us. 3584 */ 3585 ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0); 3586 ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache); 3587 dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) * 3588 vd->vdev_deflate_ratio; 3589 3590 mutex_enter(&vd->vdev_stat_lock); 3591 vd->vdev_stat.vs_alloc += alloc_delta; 3592 vd->vdev_stat.vs_space += space_delta; 3593 vd->vdev_stat.vs_dspace += dspace_delta; 3594 mutex_exit(&vd->vdev_stat_lock); 3595 3596 if (mc == spa_normal_class(spa)) { 3597 mutex_enter(&rvd->vdev_stat_lock); 3598 rvd->vdev_stat.vs_alloc += alloc_delta; 3599 rvd->vdev_stat.vs_space += space_delta; 3600 rvd->vdev_stat.vs_dspace += dspace_delta; 3601 mutex_exit(&rvd->vdev_stat_lock); 3602 } 3603 3604 if (mc != NULL) { 3605 ASSERT(rvd == vd->vdev_parent); 3606 ASSERT(vd->vdev_ms_count != 0); 3607 3608 metaslab_class_space_update(mc, 3609 alloc_delta, defer_delta, space_delta, dspace_delta); 3610 } 3611 } 3612 3613 /* 3614 * Mark a top-level vdev's config as dirty, placing it on the dirty list 3615 * so that it will be written out next time the vdev configuration is synced. 3616 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs. 3617 */ 3618 void 3619 vdev_config_dirty(vdev_t *vd) 3620 { 3621 spa_t *spa = vd->vdev_spa; 3622 vdev_t *rvd = spa->spa_root_vdev; 3623 int c; 3624 3625 ASSERT(spa_writeable(spa)); 3626 3627 /* 3628 * If this is an aux vdev (as with l2cache and spare devices), then we 3629 * update the vdev config manually and set the sync flag. 3630 */ 3631 if (vd->vdev_aux != NULL) { 3632 spa_aux_vdev_t *sav = vd->vdev_aux; 3633 nvlist_t **aux; 3634 uint_t naux; 3635 3636 for (c = 0; c < sav->sav_count; c++) { 3637 if (sav->sav_vdevs[c] == vd) 3638 break; 3639 } 3640 3641 if (c == sav->sav_count) { 3642 /* 3643 * We're being removed. There's nothing more to do. 3644 */ 3645 ASSERT(sav->sav_sync == B_TRUE); 3646 return; 3647 } 3648 3649 sav->sav_sync = B_TRUE; 3650 3651 if (nvlist_lookup_nvlist_array(sav->sav_config, 3652 ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) { 3653 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, 3654 ZPOOL_CONFIG_SPARES, &aux, &naux) == 0); 3655 } 3656 3657 ASSERT(c < naux); 3658 3659 /* 3660 * Setting the nvlist in the middle if the array is a little 3661 * sketchy, but it will work. 3662 */ 3663 nvlist_free(aux[c]); 3664 aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0); 3665 3666 return; 3667 } 3668 3669 /* 3670 * The dirty list is protected by the SCL_CONFIG lock. The caller 3671 * must either hold SCL_CONFIG as writer, or must be the sync thread 3672 * (which holds SCL_CONFIG as reader). There's only one sync thread, 3673 * so this is sufficient to ensure mutual exclusion. 3674 */ 3675 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) || 3676 (dsl_pool_sync_context(spa_get_dsl(spa)) && 3677 spa_config_held(spa, SCL_CONFIG, RW_READER))); 3678 3679 if (vd == rvd) { 3680 for (c = 0; c < rvd->vdev_children; c++) 3681 vdev_config_dirty(rvd->vdev_child[c]); 3682 } else { 3683 ASSERT(vd == vd->vdev_top); 3684 3685 if (!list_link_active(&vd->vdev_config_dirty_node) && 3686 vdev_is_concrete(vd)) { 3687 list_insert_head(&spa->spa_config_dirty_list, vd); 3688 } 3689 } 3690 } 3691 3692 void 3693 vdev_config_clean(vdev_t *vd) 3694 { 3695 spa_t *spa = vd->vdev_spa; 3696 3697 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) || 3698 (dsl_pool_sync_context(spa_get_dsl(spa)) && 3699 spa_config_held(spa, SCL_CONFIG, RW_READER))); 3700 3701 ASSERT(list_link_active(&vd->vdev_config_dirty_node)); 3702 list_remove(&spa->spa_config_dirty_list, vd); 3703 } 3704 3705 /* 3706 * Mark a top-level vdev's state as dirty, so that the next pass of 3707 * spa_sync() can convert this into vdev_config_dirty(). We distinguish 3708 * the state changes from larger config changes because they require 3709 * much less locking, and are often needed for administrative actions. 3710 */ 3711 void 3712 vdev_state_dirty(vdev_t *vd) 3713 { 3714 spa_t *spa = vd->vdev_spa; 3715 3716 ASSERT(spa_writeable(spa)); 3717 ASSERT(vd == vd->vdev_top); 3718 3719 /* 3720 * The state list is protected by the SCL_STATE lock. The caller 3721 * must either hold SCL_STATE as writer, or must be the sync thread 3722 * (which holds SCL_STATE as reader). There's only one sync thread, 3723 * so this is sufficient to ensure mutual exclusion. 3724 */ 3725 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) || 3726 (dsl_pool_sync_context(spa_get_dsl(spa)) && 3727 spa_config_held(spa, SCL_STATE, RW_READER))); 3728 3729 if (!list_link_active(&vd->vdev_state_dirty_node) && 3730 vdev_is_concrete(vd)) 3731 list_insert_head(&spa->spa_state_dirty_list, vd); 3732 } 3733 3734 void 3735 vdev_state_clean(vdev_t *vd) 3736 { 3737 spa_t *spa = vd->vdev_spa; 3738 3739 ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) || 3740 (dsl_pool_sync_context(spa_get_dsl(spa)) && 3741 spa_config_held(spa, SCL_STATE, RW_READER))); 3742 3743 ASSERT(list_link_active(&vd->vdev_state_dirty_node)); 3744 list_remove(&spa->spa_state_dirty_list, vd); 3745 } 3746 3747 /* 3748 * Propagate vdev state up from children to parent. 3749 */ 3750 void 3751 vdev_propagate_state(vdev_t *vd) 3752 { 3753 spa_t *spa = vd->vdev_spa; 3754 vdev_t *rvd = spa->spa_root_vdev; 3755 int degraded = 0, faulted = 0; 3756 int corrupted = 0; 3757 vdev_t *child; 3758 3759 if (vd->vdev_children > 0) { 3760 for (int c = 0; c < vd->vdev_children; c++) { 3761 child = vd->vdev_child[c]; 3762 3763 /* 3764 * Don't factor holes or indirect vdevs into the 3765 * decision. 3766 */ 3767 if (!vdev_is_concrete(child)) 3768 continue; 3769 3770 if (!vdev_readable(child) || 3771 (!vdev_writeable(child) && spa_writeable(spa))) { 3772 /* 3773 * Root special: if there is a top-level log 3774 * device, treat the root vdev as if it were 3775 * degraded. 3776 */ 3777 if (child->vdev_islog && vd == rvd) 3778 degraded++; 3779 else 3780 faulted++; 3781 } else if (child->vdev_state <= VDEV_STATE_DEGRADED) { 3782 degraded++; 3783 } 3784 3785 if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA) 3786 corrupted++; 3787 } 3788 3789 vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded); 3790 3791 /* 3792 * Root special: if there is a top-level vdev that cannot be 3793 * opened due to corrupted metadata, then propagate the root 3794 * vdev's aux state as 'corrupt' rather than 'insufficient 3795 * replicas'. 3796 */ 3797 if (corrupted && vd == rvd && 3798 rvd->vdev_state == VDEV_STATE_CANT_OPEN) 3799 vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN, 3800 VDEV_AUX_CORRUPT_DATA); 3801 } 3802 3803 if (vd->vdev_parent) 3804 vdev_propagate_state(vd->vdev_parent); 3805 } 3806 3807 /* 3808 * Set a vdev's state. If this is during an open, we don't update the parent 3809 * state, because we're in the process of opening children depth-first. 3810 * Otherwise, we propagate the change to the parent. 3811 * 3812 * If this routine places a device in a faulted state, an appropriate ereport is 3813 * generated. 3814 */ 3815 void 3816 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux) 3817 { 3818 uint64_t save_state; 3819 spa_t *spa = vd->vdev_spa; 3820 3821 if (state == vd->vdev_state) { 3822 vd->vdev_stat.vs_aux = aux; 3823 return; 3824 } 3825 3826 save_state = vd->vdev_state; 3827 3828 vd->vdev_state = state; 3829 vd->vdev_stat.vs_aux = aux; 3830 3831 /* 3832 * If we are setting the vdev state to anything but an open state, then 3833 * always close the underlying device unless the device has requested 3834 * a delayed close (i.e. we're about to remove or fault the device). 3835 * Otherwise, we keep accessible but invalid devices open forever. 3836 * We don't call vdev_close() itself, because that implies some extra 3837 * checks (offline, etc) that we don't want here. This is limited to 3838 * leaf devices, because otherwise closing the device will affect other 3839 * children. 3840 */ 3841 if (!vd->vdev_delayed_close && vdev_is_dead(vd) && 3842 vd->vdev_ops->vdev_op_leaf) 3843 vd->vdev_ops->vdev_op_close(vd); 3844 3845 /* 3846 * If we have brought this vdev back into service, we need 3847 * to notify fmd so that it can gracefully repair any outstanding 3848 * cases due to a missing device. We do this in all cases, even those 3849 * that probably don't correlate to a repaired fault. This is sure to 3850 * catch all cases, and we let the zfs-retire agent sort it out. If 3851 * this is a transient state it's OK, as the retire agent will 3852 * double-check the state of the vdev before repairing it. 3853 */ 3854 if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf && 3855 vd->vdev_prevstate != state) 3856 zfs_post_state_change(spa, vd); 3857 3858 if (vd->vdev_removed && 3859 state == VDEV_STATE_CANT_OPEN && 3860 (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) { 3861 /* 3862 * If the previous state is set to VDEV_STATE_REMOVED, then this 3863 * device was previously marked removed and someone attempted to 3864 * reopen it. If this failed due to a nonexistent device, then 3865 * keep the device in the REMOVED state. We also let this be if 3866 * it is one of our special test online cases, which is only 3867 * attempting to online the device and shouldn't generate an FMA 3868 * fault. 3869 */ 3870 vd->vdev_state = VDEV_STATE_REMOVED; 3871 vd->vdev_stat.vs_aux = VDEV_AUX_NONE; 3872 } else if (state == VDEV_STATE_REMOVED) { 3873 vd->vdev_removed = B_TRUE; 3874 } else if (state == VDEV_STATE_CANT_OPEN) { 3875 /* 3876 * If we fail to open a vdev during an import or recovery, we 3877 * mark it as "not available", which signifies that it was 3878 * never there to begin with. Failure to open such a device 3879 * is not considered an error. 3880 */ 3881 if ((spa_load_state(spa) == SPA_LOAD_IMPORT || 3882 spa_load_state(spa) == SPA_LOAD_RECOVER) && 3883 vd->vdev_ops->vdev_op_leaf) 3884 vd->vdev_not_present = 1; 3885 3886 /* 3887 * Post the appropriate ereport. If the 'prevstate' field is 3888 * set to something other than VDEV_STATE_UNKNOWN, it indicates 3889 * that this is part of a vdev_reopen(). In this case, we don't 3890 * want to post the ereport if the device was already in the 3891 * CANT_OPEN state beforehand. 3892 * 3893 * If the 'checkremove' flag is set, then this is an attempt to 3894 * online the device in response to an insertion event. If we 3895 * hit this case, then we have detected an insertion event for a 3896 * faulted or offline device that wasn't in the removed state. 3897 * In this scenario, we don't post an ereport because we are 3898 * about to replace the device, or attempt an online with 3899 * vdev_forcefault, which will generate the fault for us. 3900 */ 3901 if ((vd->vdev_prevstate != state || vd->vdev_forcefault) && 3902 !vd->vdev_not_present && !vd->vdev_checkremove && 3903 vd != spa->spa_root_vdev) { 3904 const char *class; 3905 3906 switch (aux) { 3907 case VDEV_AUX_OPEN_FAILED: 3908 class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED; 3909 break; 3910 case VDEV_AUX_CORRUPT_DATA: 3911 class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA; 3912 break; 3913 case VDEV_AUX_NO_REPLICAS: 3914 class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS; 3915 break; 3916 case VDEV_AUX_BAD_GUID_SUM: 3917 class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM; 3918 break; 3919 case VDEV_AUX_TOO_SMALL: 3920 class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL; 3921 break; 3922 case VDEV_AUX_BAD_LABEL: 3923 class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL; 3924 break; 3925 default: 3926 class = FM_EREPORT_ZFS_DEVICE_UNKNOWN; 3927 } 3928 3929 zfs_ereport_post(class, spa, vd, NULL, save_state, 0); 3930 } 3931 3932 /* Erase any notion of persistent removed state */ 3933 vd->vdev_removed = B_FALSE; 3934 } else { 3935 vd->vdev_removed = B_FALSE; 3936 } 3937 3938 if (!isopen && vd->vdev_parent) 3939 vdev_propagate_state(vd->vdev_parent); 3940 } 3941 3942 boolean_t 3943 vdev_children_are_offline(vdev_t *vd) 3944 { 3945 ASSERT(!vd->vdev_ops->vdev_op_leaf); 3946 3947 for (uint64_t i = 0; i < vd->vdev_children; i++) { 3948 if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE) 3949 return (B_FALSE); 3950 } 3951 3952 return (B_TRUE); 3953 } 3954 3955 /* 3956 * Check the vdev configuration to ensure that it's capable of supporting 3957 * a root pool. We do not support partial configuration. 3958 * In addition, only a single top-level vdev is allowed. 3959 */ 3960 boolean_t 3961 vdev_is_bootable(vdev_t *vd) 3962 { 3963 if (!vd->vdev_ops->vdev_op_leaf) { 3964 char *vdev_type = vd->vdev_ops->vdev_op_type; 3965 3966 if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 && 3967 vd->vdev_children > 1) { 3968 return (B_FALSE); 3969 } else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0 || 3970 strcmp(vdev_type, VDEV_TYPE_INDIRECT) == 0) { 3971 return (B_FALSE); 3972 } 3973 } 3974 3975 for (int c = 0; c < vd->vdev_children; c++) { 3976 if (!vdev_is_bootable(vd->vdev_child[c])) 3977 return (B_FALSE); 3978 } 3979 return (B_TRUE); 3980 } 3981 3982 boolean_t 3983 vdev_is_concrete(vdev_t *vd) 3984 { 3985 vdev_ops_t *ops = vd->vdev_ops; 3986 if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops || 3987 ops == &vdev_missing_ops || ops == &vdev_root_ops) { 3988 return (B_FALSE); 3989 } else { 3990 return (B_TRUE); 3991 } 3992 } 3993 3994 /* 3995 * Determine if a log device has valid content. If the vdev was 3996 * removed or faulted in the MOS config then we know that 3997 * the content on the log device has already been written to the pool. 3998 */ 3999 boolean_t 4000 vdev_log_state_valid(vdev_t *vd) 4001 { 4002 if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted && 4003 !vd->vdev_removed) 4004 return (B_TRUE); 4005 4006 for (int c = 0; c < vd->vdev_children; c++) 4007 if (vdev_log_state_valid(vd->vdev_child[c])) 4008 return (B_TRUE); 4009 4010 return (B_FALSE); 4011 } 4012 4013 /* 4014 * Expand a vdev if possible. 4015 */ 4016 void 4017 vdev_expand(vdev_t *vd, uint64_t txg) 4018 { 4019 ASSERT(vd->vdev_top == vd); 4020 ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL); 4021 4022 vdev_set_deflate_ratio(vd); 4023 4024 if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count && 4025 vdev_is_concrete(vd)) { 4026 VERIFY(vdev_metaslab_init(vd, txg) == 0); 4027 vdev_config_dirty(vd); 4028 } 4029 } 4030 4031 /* 4032 * Split a vdev. 4033 */ 4034 void 4035 vdev_split(vdev_t *vd) 4036 { 4037 vdev_t *cvd, *pvd = vd->vdev_parent; 4038 4039 vdev_remove_child(pvd, vd); 4040 vdev_compact_children(pvd); 4041 4042 cvd = pvd->vdev_child[0]; 4043 if (pvd->vdev_children == 1) { 4044 vdev_remove_parent(cvd); 4045 cvd->vdev_splitting = B_TRUE; 4046 } 4047 vdev_propagate_state(cvd); 4048 } 4049 4050 void 4051 vdev_deadman(vdev_t *vd) 4052 { 4053 for (int c = 0; c < vd->vdev_children; c++) { 4054 vdev_t *cvd = vd->vdev_child[c]; 4055 4056 vdev_deadman(cvd); 4057 } 4058 4059 if (vd->vdev_ops->vdev_op_leaf) { 4060 vdev_queue_t *vq = &vd->vdev_queue; 4061 4062 mutex_enter(&vq->vq_lock); 4063 if (avl_numnodes(&vq->vq_active_tree) > 0) { 4064 spa_t *spa = vd->vdev_spa; 4065 zio_t *fio; 4066 uint64_t delta; 4067 4068 /* 4069 * Look at the head of all the pending queues, 4070 * if any I/O has been outstanding for longer than 4071 * the spa_deadman_synctime we panic the system. 4072 */ 4073 fio = avl_first(&vq->vq_active_tree); 4074 delta = gethrtime() - fio->io_timestamp; 4075 if (delta > spa_deadman_synctime(spa)) { 4076 vdev_dbgmsg(vd, "SLOW IO: zio timestamp " 4077 "%lluns, delta %lluns, last io %lluns", 4078 fio->io_timestamp, (u_longlong_t)delta, 4079 vq->vq_io_complete_ts); 4080 fm_panic("I/O to pool '%s' appears to be " 4081 "hung.", spa_name(spa)); 4082 } 4083 } 4084 mutex_exit(&vq->vq_lock); 4085 } 4086 } 4087