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, 2020 by Delphix. All rights reserved. 25 * Copyright (c) 2018, Nexenta Systems, Inc. All rights reserved. 26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 27 * Copyright 2013 Saso Kiselkov. All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 * Copyright 2016 Toomas Soome <tsoome@me.com> 30 * Copyright (c) 2016 Actifio, Inc. All rights reserved. 31 * Copyright 2018 Joyent, Inc. 32 * Copyright (c) 2017, 2019, Datto Inc. All rights reserved. 33 * Copyright 2017 Joyent, Inc. 34 * Copyright (c) 2017, Intel Corporation. 35 * Copyright (c) 2021, Colm Buckley <colm@tuatha.org> 36 */ 37 38 /* 39 * SPA: Storage Pool Allocator 40 * 41 * This file contains all the routines used when modifying on-disk SPA state. 42 * This includes opening, importing, destroying, exporting a pool, and syncing a 43 * pool. 44 */ 45 46 #include <sys/zfs_context.h> 47 #include <sys/fm/fs/zfs.h> 48 #include <sys/spa_impl.h> 49 #include <sys/zio.h> 50 #include <sys/zio_checksum.h> 51 #include <sys/dmu.h> 52 #include <sys/dmu_tx.h> 53 #include <sys/zap.h> 54 #include <sys/zil.h> 55 #include <sys/ddt.h> 56 #include <sys/vdev_impl.h> 57 #include <sys/vdev_removal.h> 58 #include <sys/vdev_indirect_mapping.h> 59 #include <sys/vdev_indirect_births.h> 60 #include <sys/vdev_initialize.h> 61 #include <sys/vdev_rebuild.h> 62 #include <sys/vdev_trim.h> 63 #include <sys/vdev_disk.h> 64 #include <sys/vdev_draid.h> 65 #include <sys/metaslab.h> 66 #include <sys/metaslab_impl.h> 67 #include <sys/mmp.h> 68 #include <sys/uberblock_impl.h> 69 #include <sys/txg.h> 70 #include <sys/avl.h> 71 #include <sys/bpobj.h> 72 #include <sys/dmu_traverse.h> 73 #include <sys/dmu_objset.h> 74 #include <sys/unique.h> 75 #include <sys/dsl_pool.h> 76 #include <sys/dsl_dataset.h> 77 #include <sys/dsl_dir.h> 78 #include <sys/dsl_prop.h> 79 #include <sys/dsl_synctask.h> 80 #include <sys/fs/zfs.h> 81 #include <sys/arc.h> 82 #include <sys/callb.h> 83 #include <sys/systeminfo.h> 84 #include <sys/spa_boot.h> 85 #include <sys/zfs_ioctl.h> 86 #include <sys/dsl_scan.h> 87 #include <sys/zfeature.h> 88 #include <sys/dsl_destroy.h> 89 #include <sys/zvol.h> 90 91 #ifdef _KERNEL 92 #include <sys/fm/protocol.h> 93 #include <sys/fm/util.h> 94 #include <sys/callb.h> 95 #include <sys/zone.h> 96 #include <sys/vmsystm.h> 97 #endif /* _KERNEL */ 98 99 #include "zfs_prop.h" 100 #include "zfs_comutil.h" 101 102 /* 103 * The interval, in seconds, at which failed configuration cache file writes 104 * should be retried. 105 */ 106 int zfs_ccw_retry_interval = 300; 107 108 typedef enum zti_modes { 109 ZTI_MODE_FIXED, /* value is # of threads (min 1) */ 110 ZTI_MODE_BATCH, /* cpu-intensive; value is ignored */ 111 ZTI_MODE_SCALE, /* Taskqs scale with CPUs. */ 112 ZTI_MODE_NULL, /* don't create a taskq */ 113 ZTI_NMODES 114 } zti_modes_t; 115 116 #define ZTI_P(n, q) { ZTI_MODE_FIXED, (n), (q) } 117 #define ZTI_PCT(n) { ZTI_MODE_ONLINE_PERCENT, (n), 1 } 118 #define ZTI_BATCH { ZTI_MODE_BATCH, 0, 1 } 119 #define ZTI_SCALE { ZTI_MODE_SCALE, 0, 1 } 120 #define ZTI_NULL { ZTI_MODE_NULL, 0, 0 } 121 122 #define ZTI_N(n) ZTI_P(n, 1) 123 #define ZTI_ONE ZTI_N(1) 124 125 typedef struct zio_taskq_info { 126 zti_modes_t zti_mode; 127 uint_t zti_value; 128 uint_t zti_count; 129 } zio_taskq_info_t; 130 131 static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = { 132 "iss", "iss_h", "int", "int_h" 133 }; 134 135 /* 136 * This table defines the taskq settings for each ZFS I/O type. When 137 * initializing a pool, we use this table to create an appropriately sized 138 * taskq. Some operations are low volume and therefore have a small, static 139 * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE 140 * macros. Other operations process a large amount of data; the ZTI_BATCH 141 * macro causes us to create a taskq oriented for throughput. Some operations 142 * are so high frequency and short-lived that the taskq itself can become a 143 * point of lock contention. The ZTI_P(#, #) macro indicates that we need an 144 * additional degree of parallelism specified by the number of threads per- 145 * taskq and the number of taskqs; when dispatching an event in this case, the 146 * particular taskq is chosen at random. ZTI_SCALE is similar to ZTI_BATCH, 147 * but with number of taskqs also scaling with number of CPUs. 148 * 149 * The different taskq priorities are to handle the different contexts (issue 150 * and interrupt) and then to reserve threads for ZIO_PRIORITY_NOW I/Os that 151 * need to be handled with minimum delay. 152 */ 153 static const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = { 154 /* ISSUE ISSUE_HIGH INTR INTR_HIGH */ 155 { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* NULL */ 156 { ZTI_N(8), ZTI_NULL, ZTI_SCALE, ZTI_NULL }, /* READ */ 157 { ZTI_BATCH, ZTI_N(5), ZTI_SCALE, ZTI_N(5) }, /* WRITE */ 158 { ZTI_SCALE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* FREE */ 159 { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* CLAIM */ 160 { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* IOCTL */ 161 { ZTI_N(4), ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* TRIM */ 162 }; 163 164 static void spa_sync_version(void *arg, dmu_tx_t *tx); 165 static void spa_sync_props(void *arg, dmu_tx_t *tx); 166 static boolean_t spa_has_active_shared_spare(spa_t *spa); 167 static int spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport); 168 static void spa_vdev_resilver_done(spa_t *spa); 169 170 static uint_t zio_taskq_batch_pct = 80; /* 1 thread per cpu in pset */ 171 static uint_t zio_taskq_batch_tpq; /* threads per taskq */ 172 static const boolean_t zio_taskq_sysdc = B_TRUE; /* use SDC scheduling class */ 173 static const uint_t zio_taskq_basedc = 80; /* base duty cycle */ 174 175 static const boolean_t spa_create_process = B_TRUE; /* no process => no sysdc */ 176 177 /* 178 * Report any spa_load_verify errors found, but do not fail spa_load. 179 * This is used by zdb to analyze non-idle pools. 180 */ 181 boolean_t spa_load_verify_dryrun = B_FALSE; 182 183 /* 184 * Allow read spacemaps in case of readonly import (spa_mode == SPA_MODE_READ). 185 * This is used by zdb for spacemaps verification. 186 */ 187 boolean_t spa_mode_readable_spacemaps = B_FALSE; 188 189 /* 190 * This (illegal) pool name is used when temporarily importing a spa_t in order 191 * to get the vdev stats associated with the imported devices. 192 */ 193 #define TRYIMPORT_NAME "$import" 194 195 /* 196 * For debugging purposes: print out vdev tree during pool import. 197 */ 198 static int spa_load_print_vdev_tree = B_FALSE; 199 200 /* 201 * A non-zero value for zfs_max_missing_tvds means that we allow importing 202 * pools with missing top-level vdevs. This is strictly intended for advanced 203 * pool recovery cases since missing data is almost inevitable. Pools with 204 * missing devices can only be imported read-only for safety reasons, and their 205 * fail-mode will be automatically set to "continue". 206 * 207 * With 1 missing vdev we should be able to import the pool and mount all 208 * datasets. User data that was not modified after the missing device has been 209 * added should be recoverable. This means that snapshots created prior to the 210 * addition of that device should be completely intact. 211 * 212 * With 2 missing vdevs, some datasets may fail to mount since there are 213 * dataset statistics that are stored as regular metadata. Some data might be 214 * recoverable if those vdevs were added recently. 215 * 216 * With 3 or more missing vdevs, the pool is severely damaged and MOS entries 217 * may be missing entirely. Chances of data recovery are very low. Note that 218 * there are also risks of performing an inadvertent rewind as we might be 219 * missing all the vdevs with the latest uberblocks. 220 */ 221 unsigned long zfs_max_missing_tvds = 0; 222 223 /* 224 * The parameters below are similar to zfs_max_missing_tvds but are only 225 * intended for a preliminary open of the pool with an untrusted config which 226 * might be incomplete or out-dated. 227 * 228 * We are more tolerant for pools opened from a cachefile since we could have 229 * an out-dated cachefile where a device removal was not registered. 230 * We could have set the limit arbitrarily high but in the case where devices 231 * are really missing we would want to return the proper error codes; we chose 232 * SPA_DVAS_PER_BP - 1 so that some copies of the MOS would still be available 233 * and we get a chance to retrieve the trusted config. 234 */ 235 uint64_t zfs_max_missing_tvds_cachefile = SPA_DVAS_PER_BP - 1; 236 237 /* 238 * In the case where config was assembled by scanning device paths (/dev/dsks 239 * by default) we are less tolerant since all the existing devices should have 240 * been detected and we want spa_load to return the right error codes. 241 */ 242 uint64_t zfs_max_missing_tvds_scan = 0; 243 244 /* 245 * Debugging aid that pauses spa_sync() towards the end. 246 */ 247 static const boolean_t zfs_pause_spa_sync = B_FALSE; 248 249 /* 250 * Variables to indicate the livelist condense zthr func should wait at certain 251 * points for the livelist to be removed - used to test condense/destroy races 252 */ 253 static int zfs_livelist_condense_zthr_pause = 0; 254 static int zfs_livelist_condense_sync_pause = 0; 255 256 /* 257 * Variables to track whether or not condense cancellation has been 258 * triggered in testing. 259 */ 260 static int zfs_livelist_condense_sync_cancel = 0; 261 static int zfs_livelist_condense_zthr_cancel = 0; 262 263 /* 264 * Variable to track whether or not extra ALLOC blkptrs were added to a 265 * livelist entry while it was being condensed (caused by the way we track 266 * remapped blkptrs in dbuf_remap_impl) 267 */ 268 static int zfs_livelist_condense_new_alloc = 0; 269 270 /* 271 * ========================================================================== 272 * SPA properties routines 273 * ========================================================================== 274 */ 275 276 /* 277 * Add a (source=src, propname=propval) list to an nvlist. 278 */ 279 static void 280 spa_prop_add_list(nvlist_t *nvl, zpool_prop_t prop, char *strval, 281 uint64_t intval, zprop_source_t src) 282 { 283 const char *propname = zpool_prop_to_name(prop); 284 nvlist_t *propval; 285 286 propval = fnvlist_alloc(); 287 fnvlist_add_uint64(propval, ZPROP_SOURCE, src); 288 289 if (strval != NULL) 290 fnvlist_add_string(propval, ZPROP_VALUE, strval); 291 else 292 fnvlist_add_uint64(propval, ZPROP_VALUE, intval); 293 294 fnvlist_add_nvlist(nvl, propname, propval); 295 nvlist_free(propval); 296 } 297 298 /* 299 * Get property values from the spa configuration. 300 */ 301 static void 302 spa_prop_get_config(spa_t *spa, nvlist_t **nvp) 303 { 304 vdev_t *rvd = spa->spa_root_vdev; 305 dsl_pool_t *pool = spa->spa_dsl_pool; 306 uint64_t size, alloc, cap, version; 307 const zprop_source_t src = ZPROP_SRC_NONE; 308 spa_config_dirent_t *dp; 309 metaslab_class_t *mc = spa_normal_class(spa); 310 311 ASSERT(MUTEX_HELD(&spa->spa_props_lock)); 312 313 if (rvd != NULL) { 314 alloc = metaslab_class_get_alloc(mc); 315 alloc += metaslab_class_get_alloc(spa_special_class(spa)); 316 alloc += metaslab_class_get_alloc(spa_dedup_class(spa)); 317 alloc += metaslab_class_get_alloc(spa_embedded_log_class(spa)); 318 319 size = metaslab_class_get_space(mc); 320 size += metaslab_class_get_space(spa_special_class(spa)); 321 size += metaslab_class_get_space(spa_dedup_class(spa)); 322 size += metaslab_class_get_space(spa_embedded_log_class(spa)); 323 324 spa_prop_add_list(*nvp, ZPOOL_PROP_NAME, spa_name(spa), 0, src); 325 spa_prop_add_list(*nvp, ZPOOL_PROP_SIZE, NULL, size, src); 326 spa_prop_add_list(*nvp, ZPOOL_PROP_ALLOCATED, NULL, alloc, src); 327 spa_prop_add_list(*nvp, ZPOOL_PROP_FREE, NULL, 328 size - alloc, src); 329 spa_prop_add_list(*nvp, ZPOOL_PROP_CHECKPOINT, NULL, 330 spa->spa_checkpoint_info.sci_dspace, src); 331 332 spa_prop_add_list(*nvp, ZPOOL_PROP_FRAGMENTATION, NULL, 333 metaslab_class_fragmentation(mc), src); 334 spa_prop_add_list(*nvp, ZPOOL_PROP_EXPANDSZ, NULL, 335 metaslab_class_expandable_space(mc), src); 336 spa_prop_add_list(*nvp, ZPOOL_PROP_READONLY, NULL, 337 (spa_mode(spa) == SPA_MODE_READ), src); 338 339 cap = (size == 0) ? 0 : (alloc * 100 / size); 340 spa_prop_add_list(*nvp, ZPOOL_PROP_CAPACITY, NULL, cap, src); 341 342 spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL, 343 ddt_get_pool_dedup_ratio(spa), src); 344 345 spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL, 346 rvd->vdev_state, src); 347 348 version = spa_version(spa); 349 if (version == zpool_prop_default_numeric(ZPOOL_PROP_VERSION)) { 350 spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, 351 version, ZPROP_SRC_DEFAULT); 352 } else { 353 spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, 354 version, ZPROP_SRC_LOCAL); 355 } 356 spa_prop_add_list(*nvp, ZPOOL_PROP_LOAD_GUID, 357 NULL, spa_load_guid(spa), src); 358 } 359 360 if (pool != NULL) { 361 /* 362 * The $FREE directory was introduced in SPA_VERSION_DEADLISTS, 363 * when opening pools before this version freedir will be NULL. 364 */ 365 if (pool->dp_free_dir != NULL) { 366 spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL, 367 dsl_dir_phys(pool->dp_free_dir)->dd_used_bytes, 368 src); 369 } else { 370 spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, 371 NULL, 0, src); 372 } 373 374 if (pool->dp_leak_dir != NULL) { 375 spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL, 376 dsl_dir_phys(pool->dp_leak_dir)->dd_used_bytes, 377 src); 378 } else { 379 spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, 380 NULL, 0, src); 381 } 382 } 383 384 spa_prop_add_list(*nvp, ZPOOL_PROP_GUID, NULL, spa_guid(spa), src); 385 386 if (spa->spa_comment != NULL) { 387 spa_prop_add_list(*nvp, ZPOOL_PROP_COMMENT, spa->spa_comment, 388 0, ZPROP_SRC_LOCAL); 389 } 390 391 if (spa->spa_compatibility != NULL) { 392 spa_prop_add_list(*nvp, ZPOOL_PROP_COMPATIBILITY, 393 spa->spa_compatibility, 0, ZPROP_SRC_LOCAL); 394 } 395 396 if (spa->spa_root != NULL) 397 spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root, 398 0, ZPROP_SRC_LOCAL); 399 400 if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) { 401 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL, 402 MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE); 403 } else { 404 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL, 405 SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE); 406 } 407 408 if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE)) { 409 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL, 410 DNODE_MAX_SIZE, ZPROP_SRC_NONE); 411 } else { 412 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL, 413 DNODE_MIN_SIZE, ZPROP_SRC_NONE); 414 } 415 416 if ((dp = list_head(&spa->spa_config_list)) != NULL) { 417 if (dp->scd_path == NULL) { 418 spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE, 419 "none", 0, ZPROP_SRC_LOCAL); 420 } else if (strcmp(dp->scd_path, spa_config_path) != 0) { 421 spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE, 422 dp->scd_path, 0, ZPROP_SRC_LOCAL); 423 } 424 } 425 } 426 427 /* 428 * Get zpool property values. 429 */ 430 int 431 spa_prop_get(spa_t *spa, nvlist_t **nvp) 432 { 433 objset_t *mos = spa->spa_meta_objset; 434 zap_cursor_t zc; 435 zap_attribute_t za; 436 dsl_pool_t *dp; 437 int err; 438 439 err = nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP); 440 if (err) 441 return (err); 442 443 dp = spa_get_dsl(spa); 444 dsl_pool_config_enter(dp, FTAG); 445 mutex_enter(&spa->spa_props_lock); 446 447 /* 448 * Get properties from the spa config. 449 */ 450 spa_prop_get_config(spa, nvp); 451 452 /* If no pool property object, no more prop to get. */ 453 if (mos == NULL || spa->spa_pool_props_object == 0) 454 goto out; 455 456 /* 457 * Get properties from the MOS pool property object. 458 */ 459 for (zap_cursor_init(&zc, mos, spa->spa_pool_props_object); 460 (err = zap_cursor_retrieve(&zc, &za)) == 0; 461 zap_cursor_advance(&zc)) { 462 uint64_t intval = 0; 463 char *strval = NULL; 464 zprop_source_t src = ZPROP_SRC_DEFAULT; 465 zpool_prop_t prop; 466 467 if ((prop = zpool_name_to_prop(za.za_name)) == ZPOOL_PROP_INVAL) 468 continue; 469 470 switch (za.za_integer_length) { 471 case 8: 472 /* integer property */ 473 if (za.za_first_integer != 474 zpool_prop_default_numeric(prop)) 475 src = ZPROP_SRC_LOCAL; 476 477 if (prop == ZPOOL_PROP_BOOTFS) { 478 dsl_dataset_t *ds = NULL; 479 480 err = dsl_dataset_hold_obj(dp, 481 za.za_first_integer, FTAG, &ds); 482 if (err != 0) 483 break; 484 485 strval = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, 486 KM_SLEEP); 487 dsl_dataset_name(ds, strval); 488 dsl_dataset_rele(ds, FTAG); 489 } else { 490 strval = NULL; 491 intval = za.za_first_integer; 492 } 493 494 spa_prop_add_list(*nvp, prop, strval, intval, src); 495 496 if (strval != NULL) 497 kmem_free(strval, ZFS_MAX_DATASET_NAME_LEN); 498 499 break; 500 501 case 1: 502 /* string property */ 503 strval = kmem_alloc(za.za_num_integers, KM_SLEEP); 504 err = zap_lookup(mos, spa->spa_pool_props_object, 505 za.za_name, 1, za.za_num_integers, strval); 506 if (err) { 507 kmem_free(strval, za.za_num_integers); 508 break; 509 } 510 spa_prop_add_list(*nvp, prop, strval, 0, src); 511 kmem_free(strval, za.za_num_integers); 512 break; 513 514 default: 515 break; 516 } 517 } 518 zap_cursor_fini(&zc); 519 out: 520 mutex_exit(&spa->spa_props_lock); 521 dsl_pool_config_exit(dp, FTAG); 522 if (err && err != ENOENT) { 523 nvlist_free(*nvp); 524 *nvp = NULL; 525 return (err); 526 } 527 528 return (0); 529 } 530 531 /* 532 * Validate the given pool properties nvlist and modify the list 533 * for the property values to be set. 534 */ 535 static int 536 spa_prop_validate(spa_t *spa, nvlist_t *props) 537 { 538 nvpair_t *elem; 539 int error = 0, reset_bootfs = 0; 540 uint64_t objnum = 0; 541 boolean_t has_feature = B_FALSE; 542 543 elem = NULL; 544 while ((elem = nvlist_next_nvpair(props, elem)) != NULL) { 545 uint64_t intval; 546 char *strval, *slash, *check, *fname; 547 const char *propname = nvpair_name(elem); 548 zpool_prop_t prop = zpool_name_to_prop(propname); 549 550 switch (prop) { 551 case ZPOOL_PROP_INVAL: 552 if (!zpool_prop_feature(propname)) { 553 error = SET_ERROR(EINVAL); 554 break; 555 } 556 557 /* 558 * Sanitize the input. 559 */ 560 if (nvpair_type(elem) != DATA_TYPE_UINT64) { 561 error = SET_ERROR(EINVAL); 562 break; 563 } 564 565 if (nvpair_value_uint64(elem, &intval) != 0) { 566 error = SET_ERROR(EINVAL); 567 break; 568 } 569 570 if (intval != 0) { 571 error = SET_ERROR(EINVAL); 572 break; 573 } 574 575 fname = strchr(propname, '@') + 1; 576 if (zfeature_lookup_name(fname, NULL) != 0) { 577 error = SET_ERROR(EINVAL); 578 break; 579 } 580 581 has_feature = B_TRUE; 582 break; 583 584 case ZPOOL_PROP_VERSION: 585 error = nvpair_value_uint64(elem, &intval); 586 if (!error && 587 (intval < spa_version(spa) || 588 intval > SPA_VERSION_BEFORE_FEATURES || 589 has_feature)) 590 error = SET_ERROR(EINVAL); 591 break; 592 593 case ZPOOL_PROP_DELEGATION: 594 case ZPOOL_PROP_AUTOREPLACE: 595 case ZPOOL_PROP_LISTSNAPS: 596 case ZPOOL_PROP_AUTOEXPAND: 597 case ZPOOL_PROP_AUTOTRIM: 598 error = nvpair_value_uint64(elem, &intval); 599 if (!error && intval > 1) 600 error = SET_ERROR(EINVAL); 601 break; 602 603 case ZPOOL_PROP_MULTIHOST: 604 error = nvpair_value_uint64(elem, &intval); 605 if (!error && intval > 1) 606 error = SET_ERROR(EINVAL); 607 608 if (!error) { 609 uint32_t hostid = zone_get_hostid(NULL); 610 if (hostid) 611 spa->spa_hostid = hostid; 612 else 613 error = SET_ERROR(ENOTSUP); 614 } 615 616 break; 617 618 case ZPOOL_PROP_BOOTFS: 619 /* 620 * If the pool version is less than SPA_VERSION_BOOTFS, 621 * or the pool is still being created (version == 0), 622 * the bootfs property cannot be set. 623 */ 624 if (spa_version(spa) < SPA_VERSION_BOOTFS) { 625 error = SET_ERROR(ENOTSUP); 626 break; 627 } 628 629 /* 630 * Make sure the vdev config is bootable 631 */ 632 if (!vdev_is_bootable(spa->spa_root_vdev)) { 633 error = SET_ERROR(ENOTSUP); 634 break; 635 } 636 637 reset_bootfs = 1; 638 639 error = nvpair_value_string(elem, &strval); 640 641 if (!error) { 642 objset_t *os; 643 644 if (strval == NULL || strval[0] == '\0') { 645 objnum = zpool_prop_default_numeric( 646 ZPOOL_PROP_BOOTFS); 647 break; 648 } 649 650 error = dmu_objset_hold(strval, FTAG, &os); 651 if (error != 0) 652 break; 653 654 /* Must be ZPL. */ 655 if (dmu_objset_type(os) != DMU_OST_ZFS) { 656 error = SET_ERROR(ENOTSUP); 657 } else { 658 objnum = dmu_objset_id(os); 659 } 660 dmu_objset_rele(os, FTAG); 661 } 662 break; 663 664 case ZPOOL_PROP_FAILUREMODE: 665 error = nvpair_value_uint64(elem, &intval); 666 if (!error && intval > ZIO_FAILURE_MODE_PANIC) 667 error = SET_ERROR(EINVAL); 668 669 /* 670 * This is a special case which only occurs when 671 * the pool has completely failed. This allows 672 * the user to change the in-core failmode property 673 * without syncing it out to disk (I/Os might 674 * currently be blocked). We do this by returning 675 * EIO to the caller (spa_prop_set) to trick it 676 * into thinking we encountered a property validation 677 * error. 678 */ 679 if (!error && spa_suspended(spa)) { 680 spa->spa_failmode = intval; 681 error = SET_ERROR(EIO); 682 } 683 break; 684 685 case ZPOOL_PROP_CACHEFILE: 686 if ((error = nvpair_value_string(elem, &strval)) != 0) 687 break; 688 689 if (strval[0] == '\0') 690 break; 691 692 if (strcmp(strval, "none") == 0) 693 break; 694 695 if (strval[0] != '/') { 696 error = SET_ERROR(EINVAL); 697 break; 698 } 699 700 slash = strrchr(strval, '/'); 701 ASSERT(slash != NULL); 702 703 if (slash[1] == '\0' || strcmp(slash, "/.") == 0 || 704 strcmp(slash, "/..") == 0) 705 error = SET_ERROR(EINVAL); 706 break; 707 708 case ZPOOL_PROP_COMMENT: 709 if ((error = nvpair_value_string(elem, &strval)) != 0) 710 break; 711 for (check = strval; *check != '\0'; check++) { 712 if (!isprint(*check)) { 713 error = SET_ERROR(EINVAL); 714 break; 715 } 716 } 717 if (strlen(strval) > ZPROP_MAX_COMMENT) 718 error = SET_ERROR(E2BIG); 719 break; 720 721 default: 722 break; 723 } 724 725 if (error) 726 break; 727 } 728 729 (void) nvlist_remove_all(props, 730 zpool_prop_to_name(ZPOOL_PROP_DEDUPDITTO)); 731 732 if (!error && reset_bootfs) { 733 error = nvlist_remove(props, 734 zpool_prop_to_name(ZPOOL_PROP_BOOTFS), DATA_TYPE_STRING); 735 736 if (!error) { 737 error = nvlist_add_uint64(props, 738 zpool_prop_to_name(ZPOOL_PROP_BOOTFS), objnum); 739 } 740 } 741 742 return (error); 743 } 744 745 void 746 spa_configfile_set(spa_t *spa, nvlist_t *nvp, boolean_t need_sync) 747 { 748 char *cachefile; 749 spa_config_dirent_t *dp; 750 751 if (nvlist_lookup_string(nvp, zpool_prop_to_name(ZPOOL_PROP_CACHEFILE), 752 &cachefile) != 0) 753 return; 754 755 dp = kmem_alloc(sizeof (spa_config_dirent_t), 756 KM_SLEEP); 757 758 if (cachefile[0] == '\0') 759 dp->scd_path = spa_strdup(spa_config_path); 760 else if (strcmp(cachefile, "none") == 0) 761 dp->scd_path = NULL; 762 else 763 dp->scd_path = spa_strdup(cachefile); 764 765 list_insert_head(&spa->spa_config_list, dp); 766 if (need_sync) 767 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 768 } 769 770 int 771 spa_prop_set(spa_t *spa, nvlist_t *nvp) 772 { 773 int error; 774 nvpair_t *elem = NULL; 775 boolean_t need_sync = B_FALSE; 776 777 if ((error = spa_prop_validate(spa, nvp)) != 0) 778 return (error); 779 780 while ((elem = nvlist_next_nvpair(nvp, elem)) != NULL) { 781 zpool_prop_t prop = zpool_name_to_prop(nvpair_name(elem)); 782 783 if (prop == ZPOOL_PROP_CACHEFILE || 784 prop == ZPOOL_PROP_ALTROOT || 785 prop == ZPOOL_PROP_READONLY) 786 continue; 787 788 if (prop == ZPOOL_PROP_VERSION || prop == ZPOOL_PROP_INVAL) { 789 uint64_t ver = 0; 790 791 if (prop == ZPOOL_PROP_VERSION) { 792 VERIFY(nvpair_value_uint64(elem, &ver) == 0); 793 } else { 794 ASSERT(zpool_prop_feature(nvpair_name(elem))); 795 ver = SPA_VERSION_FEATURES; 796 need_sync = B_TRUE; 797 } 798 799 /* Save time if the version is already set. */ 800 if (ver == spa_version(spa)) 801 continue; 802 803 /* 804 * In addition to the pool directory object, we might 805 * create the pool properties object, the features for 806 * read object, the features for write object, or the 807 * feature descriptions object. 808 */ 809 error = dsl_sync_task(spa->spa_name, NULL, 810 spa_sync_version, &ver, 811 6, ZFS_SPACE_CHECK_RESERVED); 812 if (error) 813 return (error); 814 continue; 815 } 816 817 need_sync = B_TRUE; 818 break; 819 } 820 821 if (need_sync) { 822 return (dsl_sync_task(spa->spa_name, NULL, spa_sync_props, 823 nvp, 6, ZFS_SPACE_CHECK_RESERVED)); 824 } 825 826 return (0); 827 } 828 829 /* 830 * If the bootfs property value is dsobj, clear it. 831 */ 832 void 833 spa_prop_clear_bootfs(spa_t *spa, uint64_t dsobj, dmu_tx_t *tx) 834 { 835 if (spa->spa_bootfs == dsobj && spa->spa_pool_props_object != 0) { 836 VERIFY(zap_remove(spa->spa_meta_objset, 837 spa->spa_pool_props_object, 838 zpool_prop_to_name(ZPOOL_PROP_BOOTFS), tx) == 0); 839 spa->spa_bootfs = 0; 840 } 841 } 842 843 static int 844 spa_change_guid_check(void *arg, dmu_tx_t *tx) 845 { 846 uint64_t *newguid __maybe_unused = arg; 847 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 848 vdev_t *rvd = spa->spa_root_vdev; 849 uint64_t vdev_state; 850 851 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 852 int error = (spa_has_checkpoint(spa)) ? 853 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 854 return (SET_ERROR(error)); 855 } 856 857 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 858 vdev_state = rvd->vdev_state; 859 spa_config_exit(spa, SCL_STATE, FTAG); 860 861 if (vdev_state != VDEV_STATE_HEALTHY) 862 return (SET_ERROR(ENXIO)); 863 864 ASSERT3U(spa_guid(spa), !=, *newguid); 865 866 return (0); 867 } 868 869 static void 870 spa_change_guid_sync(void *arg, dmu_tx_t *tx) 871 { 872 uint64_t *newguid = arg; 873 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 874 uint64_t oldguid; 875 vdev_t *rvd = spa->spa_root_vdev; 876 877 oldguid = spa_guid(spa); 878 879 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 880 rvd->vdev_guid = *newguid; 881 rvd->vdev_guid_sum += (*newguid - oldguid); 882 vdev_config_dirty(rvd); 883 spa_config_exit(spa, SCL_STATE, FTAG); 884 885 spa_history_log_internal(spa, "guid change", tx, "old=%llu new=%llu", 886 (u_longlong_t)oldguid, (u_longlong_t)*newguid); 887 } 888 889 /* 890 * Change the GUID for the pool. This is done so that we can later 891 * re-import a pool built from a clone of our own vdevs. We will modify 892 * the root vdev's guid, our own pool guid, and then mark all of our 893 * vdevs dirty. Note that we must make sure that all our vdevs are 894 * online when we do this, or else any vdevs that weren't present 895 * would be orphaned from our pool. We are also going to issue a 896 * sysevent to update any watchers. 897 */ 898 int 899 spa_change_guid(spa_t *spa) 900 { 901 int error; 902 uint64_t guid; 903 904 mutex_enter(&spa->spa_vdev_top_lock); 905 mutex_enter(&spa_namespace_lock); 906 guid = spa_generate_guid(NULL); 907 908 error = dsl_sync_task(spa->spa_name, spa_change_guid_check, 909 spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED); 910 911 if (error == 0) { 912 spa_write_cachefile(spa, B_FALSE, B_TRUE); 913 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID); 914 } 915 916 mutex_exit(&spa_namespace_lock); 917 mutex_exit(&spa->spa_vdev_top_lock); 918 919 return (error); 920 } 921 922 /* 923 * ========================================================================== 924 * SPA state manipulation (open/create/destroy/import/export) 925 * ========================================================================== 926 */ 927 928 static int 929 spa_error_entry_compare(const void *a, const void *b) 930 { 931 const spa_error_entry_t *sa = (const spa_error_entry_t *)a; 932 const spa_error_entry_t *sb = (const spa_error_entry_t *)b; 933 int ret; 934 935 ret = memcmp(&sa->se_bookmark, &sb->se_bookmark, 936 sizeof (zbookmark_phys_t)); 937 938 return (TREE_ISIGN(ret)); 939 } 940 941 /* 942 * Utility function which retrieves copies of the current logs and 943 * re-initializes them in the process. 944 */ 945 void 946 spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub) 947 { 948 ASSERT(MUTEX_HELD(&spa->spa_errlist_lock)); 949 950 memcpy(last, &spa->spa_errlist_last, sizeof (avl_tree_t)); 951 memcpy(scrub, &spa->spa_errlist_scrub, sizeof (avl_tree_t)); 952 953 avl_create(&spa->spa_errlist_scrub, 954 spa_error_entry_compare, sizeof (spa_error_entry_t), 955 offsetof(spa_error_entry_t, se_avl)); 956 avl_create(&spa->spa_errlist_last, 957 spa_error_entry_compare, sizeof (spa_error_entry_t), 958 offsetof(spa_error_entry_t, se_avl)); 959 } 960 961 static void 962 spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q) 963 { 964 const zio_taskq_info_t *ztip = &zio_taskqs[t][q]; 965 enum zti_modes mode = ztip->zti_mode; 966 uint_t value = ztip->zti_value; 967 uint_t count = ztip->zti_count; 968 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 969 uint_t cpus, flags = TASKQ_DYNAMIC; 970 boolean_t batch = B_FALSE; 971 972 switch (mode) { 973 case ZTI_MODE_FIXED: 974 ASSERT3U(value, >, 0); 975 break; 976 977 case ZTI_MODE_BATCH: 978 batch = B_TRUE; 979 flags |= TASKQ_THREADS_CPU_PCT; 980 value = MIN(zio_taskq_batch_pct, 100); 981 break; 982 983 case ZTI_MODE_SCALE: 984 flags |= TASKQ_THREADS_CPU_PCT; 985 /* 986 * We want more taskqs to reduce lock contention, but we want 987 * less for better request ordering and CPU utilization. 988 */ 989 cpus = MAX(1, boot_ncpus * zio_taskq_batch_pct / 100); 990 if (zio_taskq_batch_tpq > 0) { 991 count = MAX(1, (cpus + zio_taskq_batch_tpq / 2) / 992 zio_taskq_batch_tpq); 993 } else { 994 /* 995 * Prefer 6 threads per taskq, but no more taskqs 996 * than threads in them on large systems. For 80%: 997 * 998 * taskq taskq total 999 * cpus taskqs percent threads threads 1000 * ------- ------- ------- ------- ------- 1001 * 1 1 80% 1 1 1002 * 2 1 80% 1 1 1003 * 4 1 80% 3 3 1004 * 8 2 40% 3 6 1005 * 16 3 27% 4 12 1006 * 32 5 16% 5 25 1007 * 64 7 11% 7 49 1008 * 128 10 8% 10 100 1009 * 256 14 6% 15 210 1010 */ 1011 count = 1 + cpus / 6; 1012 while (count * count > cpus) 1013 count--; 1014 } 1015 /* Limit each taskq within 100% to not trigger assertion. */ 1016 count = MAX(count, (zio_taskq_batch_pct + 99) / 100); 1017 value = (zio_taskq_batch_pct + count / 2) / count; 1018 break; 1019 1020 case ZTI_MODE_NULL: 1021 tqs->stqs_count = 0; 1022 tqs->stqs_taskq = NULL; 1023 return; 1024 1025 default: 1026 panic("unrecognized mode for %s_%s taskq (%u:%u) in " 1027 "spa_activate()", 1028 zio_type_name[t], zio_taskq_types[q], mode, value); 1029 break; 1030 } 1031 1032 ASSERT3U(count, >, 0); 1033 tqs->stqs_count = count; 1034 tqs->stqs_taskq = kmem_alloc(count * sizeof (taskq_t *), KM_SLEEP); 1035 1036 for (uint_t i = 0; i < count; i++) { 1037 taskq_t *tq; 1038 char name[32]; 1039 1040 if (count > 1) 1041 (void) snprintf(name, sizeof (name), "%s_%s_%u", 1042 zio_type_name[t], zio_taskq_types[q], i); 1043 else 1044 (void) snprintf(name, sizeof (name), "%s_%s", 1045 zio_type_name[t], zio_taskq_types[q]); 1046 1047 if (zio_taskq_sysdc && spa->spa_proc != &p0) { 1048 if (batch) 1049 flags |= TASKQ_DC_BATCH; 1050 1051 (void) zio_taskq_basedc; 1052 tq = taskq_create_sysdc(name, value, 50, INT_MAX, 1053 spa->spa_proc, zio_taskq_basedc, flags); 1054 } else { 1055 pri_t pri = maxclsyspri; 1056 /* 1057 * The write issue taskq can be extremely CPU 1058 * intensive. Run it at slightly less important 1059 * priority than the other taskqs. 1060 * 1061 * Under Linux and FreeBSD this means incrementing 1062 * the priority value as opposed to platforms like 1063 * illumos where it should be decremented. 1064 * 1065 * On FreeBSD, if priorities divided by four (RQ_PPQ) 1066 * are equal then a difference between them is 1067 * insignificant. 1068 */ 1069 if (t == ZIO_TYPE_WRITE && q == ZIO_TASKQ_ISSUE) { 1070 #if defined(__linux__) 1071 pri++; 1072 #elif defined(__FreeBSD__) 1073 pri += 4; 1074 #else 1075 #error "unknown OS" 1076 #endif 1077 } 1078 tq = taskq_create_proc(name, value, pri, 50, 1079 INT_MAX, spa->spa_proc, flags); 1080 } 1081 1082 tqs->stqs_taskq[i] = tq; 1083 } 1084 } 1085 1086 static void 1087 spa_taskqs_fini(spa_t *spa, zio_type_t t, zio_taskq_type_t q) 1088 { 1089 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1090 1091 if (tqs->stqs_taskq == NULL) { 1092 ASSERT3U(tqs->stqs_count, ==, 0); 1093 return; 1094 } 1095 1096 for (uint_t i = 0; i < tqs->stqs_count; i++) { 1097 ASSERT3P(tqs->stqs_taskq[i], !=, NULL); 1098 taskq_destroy(tqs->stqs_taskq[i]); 1099 } 1100 1101 kmem_free(tqs->stqs_taskq, tqs->stqs_count * sizeof (taskq_t *)); 1102 tqs->stqs_taskq = NULL; 1103 } 1104 1105 /* 1106 * Dispatch a task to the appropriate taskq for the ZFS I/O type and priority. 1107 * Note that a type may have multiple discrete taskqs to avoid lock contention 1108 * on the taskq itself. In that case we choose which taskq at random by using 1109 * the low bits of gethrtime(). 1110 */ 1111 void 1112 spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q, 1113 task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent) 1114 { 1115 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1116 taskq_t *tq; 1117 1118 ASSERT3P(tqs->stqs_taskq, !=, NULL); 1119 ASSERT3U(tqs->stqs_count, !=, 0); 1120 1121 if (tqs->stqs_count == 1) { 1122 tq = tqs->stqs_taskq[0]; 1123 } else { 1124 tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count]; 1125 } 1126 1127 taskq_dispatch_ent(tq, func, arg, flags, ent); 1128 } 1129 1130 /* 1131 * Same as spa_taskq_dispatch_ent() but block on the task until completion. 1132 */ 1133 void 1134 spa_taskq_dispatch_sync(spa_t *spa, zio_type_t t, zio_taskq_type_t q, 1135 task_func_t *func, void *arg, uint_t flags) 1136 { 1137 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1138 taskq_t *tq; 1139 taskqid_t id; 1140 1141 ASSERT3P(tqs->stqs_taskq, !=, NULL); 1142 ASSERT3U(tqs->stqs_count, !=, 0); 1143 1144 if (tqs->stqs_count == 1) { 1145 tq = tqs->stqs_taskq[0]; 1146 } else { 1147 tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count]; 1148 } 1149 1150 id = taskq_dispatch(tq, func, arg, flags); 1151 if (id) 1152 taskq_wait_id(tq, id); 1153 } 1154 1155 static void 1156 spa_create_zio_taskqs(spa_t *spa) 1157 { 1158 for (int t = 0; t < ZIO_TYPES; t++) { 1159 for (int q = 0; q < ZIO_TASKQ_TYPES; q++) { 1160 spa_taskqs_init(spa, t, q); 1161 } 1162 } 1163 } 1164 1165 /* 1166 * Disabled until spa_thread() can be adapted for Linux. 1167 */ 1168 #undef HAVE_SPA_THREAD 1169 1170 #if defined(_KERNEL) && defined(HAVE_SPA_THREAD) 1171 static void 1172 spa_thread(void *arg) 1173 { 1174 psetid_t zio_taskq_psrset_bind = PS_NONE; 1175 callb_cpr_t cprinfo; 1176 1177 spa_t *spa = arg; 1178 user_t *pu = PTOU(curproc); 1179 1180 CALLB_CPR_INIT(&cprinfo, &spa->spa_proc_lock, callb_generic_cpr, 1181 spa->spa_name); 1182 1183 ASSERT(curproc != &p0); 1184 (void) snprintf(pu->u_psargs, sizeof (pu->u_psargs), 1185 "zpool-%s", spa->spa_name); 1186 (void) strlcpy(pu->u_comm, pu->u_psargs, sizeof (pu->u_comm)); 1187 1188 /* bind this thread to the requested psrset */ 1189 if (zio_taskq_psrset_bind != PS_NONE) { 1190 pool_lock(); 1191 mutex_enter(&cpu_lock); 1192 mutex_enter(&pidlock); 1193 mutex_enter(&curproc->p_lock); 1194 1195 if (cpupart_bind_thread(curthread, zio_taskq_psrset_bind, 1196 0, NULL, NULL) == 0) { 1197 curthread->t_bind_pset = zio_taskq_psrset_bind; 1198 } else { 1199 cmn_err(CE_WARN, 1200 "Couldn't bind process for zfs pool \"%s\" to " 1201 "pset %d\n", spa->spa_name, zio_taskq_psrset_bind); 1202 } 1203 1204 mutex_exit(&curproc->p_lock); 1205 mutex_exit(&pidlock); 1206 mutex_exit(&cpu_lock); 1207 pool_unlock(); 1208 } 1209 1210 if (zio_taskq_sysdc) { 1211 sysdc_thread_enter(curthread, 100, 0); 1212 } 1213 1214 spa->spa_proc = curproc; 1215 spa->spa_did = curthread->t_did; 1216 1217 spa_create_zio_taskqs(spa); 1218 1219 mutex_enter(&spa->spa_proc_lock); 1220 ASSERT(spa->spa_proc_state == SPA_PROC_CREATED); 1221 1222 spa->spa_proc_state = SPA_PROC_ACTIVE; 1223 cv_broadcast(&spa->spa_proc_cv); 1224 1225 CALLB_CPR_SAFE_BEGIN(&cprinfo); 1226 while (spa->spa_proc_state == SPA_PROC_ACTIVE) 1227 cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); 1228 CALLB_CPR_SAFE_END(&cprinfo, &spa->spa_proc_lock); 1229 1230 ASSERT(spa->spa_proc_state == SPA_PROC_DEACTIVATE); 1231 spa->spa_proc_state = SPA_PROC_GONE; 1232 spa->spa_proc = &p0; 1233 cv_broadcast(&spa->spa_proc_cv); 1234 CALLB_CPR_EXIT(&cprinfo); /* drops spa_proc_lock */ 1235 1236 mutex_enter(&curproc->p_lock); 1237 lwp_exit(); 1238 } 1239 #endif 1240 1241 /* 1242 * Activate an uninitialized pool. 1243 */ 1244 static void 1245 spa_activate(spa_t *spa, spa_mode_t mode) 1246 { 1247 ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED); 1248 1249 spa->spa_state = POOL_STATE_ACTIVE; 1250 spa->spa_mode = mode; 1251 spa->spa_read_spacemaps = spa_mode_readable_spacemaps; 1252 1253 spa->spa_normal_class = metaslab_class_create(spa, &zfs_metaslab_ops); 1254 spa->spa_log_class = metaslab_class_create(spa, &zfs_metaslab_ops); 1255 spa->spa_embedded_log_class = 1256 metaslab_class_create(spa, &zfs_metaslab_ops); 1257 spa->spa_special_class = metaslab_class_create(spa, &zfs_metaslab_ops); 1258 spa->spa_dedup_class = metaslab_class_create(spa, &zfs_metaslab_ops); 1259 1260 /* Try to create a covering process */ 1261 mutex_enter(&spa->spa_proc_lock); 1262 ASSERT(spa->spa_proc_state == SPA_PROC_NONE); 1263 ASSERT(spa->spa_proc == &p0); 1264 spa->spa_did = 0; 1265 1266 (void) spa_create_process; 1267 #ifdef HAVE_SPA_THREAD 1268 /* Only create a process if we're going to be around a while. */ 1269 if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) { 1270 if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri, 1271 NULL, 0) == 0) { 1272 spa->spa_proc_state = SPA_PROC_CREATED; 1273 while (spa->spa_proc_state == SPA_PROC_CREATED) { 1274 cv_wait(&spa->spa_proc_cv, 1275 &spa->spa_proc_lock); 1276 } 1277 ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE); 1278 ASSERT(spa->spa_proc != &p0); 1279 ASSERT(spa->spa_did != 0); 1280 } else { 1281 #ifdef _KERNEL 1282 cmn_err(CE_WARN, 1283 "Couldn't create process for zfs pool \"%s\"\n", 1284 spa->spa_name); 1285 #endif 1286 } 1287 } 1288 #endif /* HAVE_SPA_THREAD */ 1289 mutex_exit(&spa->spa_proc_lock); 1290 1291 /* If we didn't create a process, we need to create our taskqs. */ 1292 if (spa->spa_proc == &p0) { 1293 spa_create_zio_taskqs(spa); 1294 } 1295 1296 for (size_t i = 0; i < TXG_SIZE; i++) { 1297 spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL, 1298 ZIO_FLAG_CANFAIL); 1299 } 1300 1301 list_create(&spa->spa_config_dirty_list, sizeof (vdev_t), 1302 offsetof(vdev_t, vdev_config_dirty_node)); 1303 list_create(&spa->spa_evicting_os_list, sizeof (objset_t), 1304 offsetof(objset_t, os_evicting_node)); 1305 list_create(&spa->spa_state_dirty_list, sizeof (vdev_t), 1306 offsetof(vdev_t, vdev_state_dirty_node)); 1307 1308 txg_list_create(&spa->spa_vdev_txg_list, spa, 1309 offsetof(struct vdev, vdev_txg_node)); 1310 1311 avl_create(&spa->spa_errlist_scrub, 1312 spa_error_entry_compare, sizeof (spa_error_entry_t), 1313 offsetof(spa_error_entry_t, se_avl)); 1314 avl_create(&spa->spa_errlist_last, 1315 spa_error_entry_compare, sizeof (spa_error_entry_t), 1316 offsetof(spa_error_entry_t, se_avl)); 1317 1318 spa_activate_os(spa); 1319 1320 spa_keystore_init(&spa->spa_keystore); 1321 1322 /* 1323 * This taskq is used to perform zvol-minor-related tasks 1324 * asynchronously. This has several advantages, including easy 1325 * resolution of various deadlocks. 1326 * 1327 * The taskq must be single threaded to ensure tasks are always 1328 * processed in the order in which they were dispatched. 1329 * 1330 * A taskq per pool allows one to keep the pools independent. 1331 * This way if one pool is suspended, it will not impact another. 1332 * 1333 * The preferred location to dispatch a zvol minor task is a sync 1334 * task. In this context, there is easy access to the spa_t and minimal 1335 * error handling is required because the sync task must succeed. 1336 */ 1337 spa->spa_zvol_taskq = taskq_create("z_zvol", 1, defclsyspri, 1338 1, INT_MAX, 0); 1339 1340 /* 1341 * Taskq dedicated to prefetcher threads: this is used to prevent the 1342 * pool traverse code from monopolizing the global (and limited) 1343 * system_taskq by inappropriately scheduling long running tasks on it. 1344 */ 1345 spa->spa_prefetch_taskq = taskq_create("z_prefetch", 100, 1346 defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 1347 1348 /* 1349 * The taskq to upgrade datasets in this pool. Currently used by 1350 * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA. 1351 */ 1352 spa->spa_upgrade_taskq = taskq_create("z_upgrade", 100, 1353 defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 1354 } 1355 1356 /* 1357 * Opposite of spa_activate(). 1358 */ 1359 static void 1360 spa_deactivate(spa_t *spa) 1361 { 1362 ASSERT(spa->spa_sync_on == B_FALSE); 1363 ASSERT(spa->spa_dsl_pool == NULL); 1364 ASSERT(spa->spa_root_vdev == NULL); 1365 ASSERT(spa->spa_async_zio_root == NULL); 1366 ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED); 1367 1368 spa_evicting_os_wait(spa); 1369 1370 if (spa->spa_zvol_taskq) { 1371 taskq_destroy(spa->spa_zvol_taskq); 1372 spa->spa_zvol_taskq = NULL; 1373 } 1374 1375 if (spa->spa_prefetch_taskq) { 1376 taskq_destroy(spa->spa_prefetch_taskq); 1377 spa->spa_prefetch_taskq = NULL; 1378 } 1379 1380 if (spa->spa_upgrade_taskq) { 1381 taskq_destroy(spa->spa_upgrade_taskq); 1382 spa->spa_upgrade_taskq = NULL; 1383 } 1384 1385 txg_list_destroy(&spa->spa_vdev_txg_list); 1386 1387 list_destroy(&spa->spa_config_dirty_list); 1388 list_destroy(&spa->spa_evicting_os_list); 1389 list_destroy(&spa->spa_state_dirty_list); 1390 1391 taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); 1392 1393 for (int t = 0; t < ZIO_TYPES; t++) { 1394 for (int q = 0; q < ZIO_TASKQ_TYPES; q++) { 1395 spa_taskqs_fini(spa, t, q); 1396 } 1397 } 1398 1399 for (size_t i = 0; i < TXG_SIZE; i++) { 1400 ASSERT3P(spa->spa_txg_zio[i], !=, NULL); 1401 VERIFY0(zio_wait(spa->spa_txg_zio[i])); 1402 spa->spa_txg_zio[i] = NULL; 1403 } 1404 1405 metaslab_class_destroy(spa->spa_normal_class); 1406 spa->spa_normal_class = NULL; 1407 1408 metaslab_class_destroy(spa->spa_log_class); 1409 spa->spa_log_class = NULL; 1410 1411 metaslab_class_destroy(spa->spa_embedded_log_class); 1412 spa->spa_embedded_log_class = NULL; 1413 1414 metaslab_class_destroy(spa->spa_special_class); 1415 spa->spa_special_class = NULL; 1416 1417 metaslab_class_destroy(spa->spa_dedup_class); 1418 spa->spa_dedup_class = NULL; 1419 1420 /* 1421 * If this was part of an import or the open otherwise failed, we may 1422 * still have errors left in the queues. Empty them just in case. 1423 */ 1424 spa_errlog_drain(spa); 1425 avl_destroy(&spa->spa_errlist_scrub); 1426 avl_destroy(&spa->spa_errlist_last); 1427 1428 spa_keystore_fini(&spa->spa_keystore); 1429 1430 spa->spa_state = POOL_STATE_UNINITIALIZED; 1431 1432 mutex_enter(&spa->spa_proc_lock); 1433 if (spa->spa_proc_state != SPA_PROC_NONE) { 1434 ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE); 1435 spa->spa_proc_state = SPA_PROC_DEACTIVATE; 1436 cv_broadcast(&spa->spa_proc_cv); 1437 while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) { 1438 ASSERT(spa->spa_proc != &p0); 1439 cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); 1440 } 1441 ASSERT(spa->spa_proc_state == SPA_PROC_GONE); 1442 spa->spa_proc_state = SPA_PROC_NONE; 1443 } 1444 ASSERT(spa->spa_proc == &p0); 1445 mutex_exit(&spa->spa_proc_lock); 1446 1447 /* 1448 * We want to make sure spa_thread() has actually exited the ZFS 1449 * module, so that the module can't be unloaded out from underneath 1450 * it. 1451 */ 1452 if (spa->spa_did != 0) { 1453 thread_join(spa->spa_did); 1454 spa->spa_did = 0; 1455 } 1456 1457 spa_deactivate_os(spa); 1458 1459 } 1460 1461 /* 1462 * Verify a pool configuration, and construct the vdev tree appropriately. This 1463 * will create all the necessary vdevs in the appropriate layout, with each vdev 1464 * in the CLOSED state. This will prep the pool before open/creation/import. 1465 * All vdev validation is done by the vdev_alloc() routine. 1466 */ 1467 int 1468 spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, 1469 uint_t id, int atype) 1470 { 1471 nvlist_t **child; 1472 uint_t children; 1473 int error; 1474 1475 if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0) 1476 return (error); 1477 1478 if ((*vdp)->vdev_ops->vdev_op_leaf) 1479 return (0); 1480 1481 error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN, 1482 &child, &children); 1483 1484 if (error == ENOENT) 1485 return (0); 1486 1487 if (error) { 1488 vdev_free(*vdp); 1489 *vdp = NULL; 1490 return (SET_ERROR(EINVAL)); 1491 } 1492 1493 for (int c = 0; c < children; c++) { 1494 vdev_t *vd; 1495 if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c, 1496 atype)) != 0) { 1497 vdev_free(*vdp); 1498 *vdp = NULL; 1499 return (error); 1500 } 1501 } 1502 1503 ASSERT(*vdp != NULL); 1504 1505 return (0); 1506 } 1507 1508 static boolean_t 1509 spa_should_flush_logs_on_unload(spa_t *spa) 1510 { 1511 if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) 1512 return (B_FALSE); 1513 1514 if (!spa_writeable(spa)) 1515 return (B_FALSE); 1516 1517 if (!spa->spa_sync_on) 1518 return (B_FALSE); 1519 1520 if (spa_state(spa) != POOL_STATE_EXPORTED) 1521 return (B_FALSE); 1522 1523 if (zfs_keep_log_spacemaps_at_export) 1524 return (B_FALSE); 1525 1526 return (B_TRUE); 1527 } 1528 1529 /* 1530 * Opens a transaction that will set the flag that will instruct 1531 * spa_sync to attempt to flush all the metaslabs for that txg. 1532 */ 1533 static void 1534 spa_unload_log_sm_flush_all(spa_t *spa) 1535 { 1536 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 1537 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 1538 1539 ASSERT3U(spa->spa_log_flushall_txg, ==, 0); 1540 spa->spa_log_flushall_txg = dmu_tx_get_txg(tx); 1541 1542 dmu_tx_commit(tx); 1543 txg_wait_synced(spa_get_dsl(spa), spa->spa_log_flushall_txg); 1544 } 1545 1546 static void 1547 spa_unload_log_sm_metadata(spa_t *spa) 1548 { 1549 void *cookie = NULL; 1550 spa_log_sm_t *sls; 1551 while ((sls = avl_destroy_nodes(&spa->spa_sm_logs_by_txg, 1552 &cookie)) != NULL) { 1553 VERIFY0(sls->sls_mscount); 1554 kmem_free(sls, sizeof (spa_log_sm_t)); 1555 } 1556 1557 for (log_summary_entry_t *e = list_head(&spa->spa_log_summary); 1558 e != NULL; e = list_head(&spa->spa_log_summary)) { 1559 VERIFY0(e->lse_mscount); 1560 list_remove(&spa->spa_log_summary, e); 1561 kmem_free(e, sizeof (log_summary_entry_t)); 1562 } 1563 1564 spa->spa_unflushed_stats.sus_nblocks = 0; 1565 spa->spa_unflushed_stats.sus_memused = 0; 1566 spa->spa_unflushed_stats.sus_blocklimit = 0; 1567 } 1568 1569 static void 1570 spa_destroy_aux_threads(spa_t *spa) 1571 { 1572 if (spa->spa_condense_zthr != NULL) { 1573 zthr_destroy(spa->spa_condense_zthr); 1574 spa->spa_condense_zthr = NULL; 1575 } 1576 if (spa->spa_checkpoint_discard_zthr != NULL) { 1577 zthr_destroy(spa->spa_checkpoint_discard_zthr); 1578 spa->spa_checkpoint_discard_zthr = NULL; 1579 } 1580 if (spa->spa_livelist_delete_zthr != NULL) { 1581 zthr_destroy(spa->spa_livelist_delete_zthr); 1582 spa->spa_livelist_delete_zthr = NULL; 1583 } 1584 if (spa->spa_livelist_condense_zthr != NULL) { 1585 zthr_destroy(spa->spa_livelist_condense_zthr); 1586 spa->spa_livelist_condense_zthr = NULL; 1587 } 1588 } 1589 1590 /* 1591 * Opposite of spa_load(). 1592 */ 1593 static void 1594 spa_unload(spa_t *spa) 1595 { 1596 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 1597 ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED); 1598 1599 spa_import_progress_remove(spa_guid(spa)); 1600 spa_load_note(spa, "UNLOADING"); 1601 1602 spa_wake_waiters(spa); 1603 1604 /* 1605 * If we have set the spa_final_txg, we have already performed the 1606 * tasks below in spa_export_common(). We should not redo it here since 1607 * we delay the final TXGs beyond what spa_final_txg is set at. 1608 */ 1609 if (spa->spa_final_txg == UINT64_MAX) { 1610 /* 1611 * If the log space map feature is enabled and the pool is 1612 * getting exported (but not destroyed), we want to spend some 1613 * time flushing as many metaslabs as we can in an attempt to 1614 * destroy log space maps and save import time. 1615 */ 1616 if (spa_should_flush_logs_on_unload(spa)) 1617 spa_unload_log_sm_flush_all(spa); 1618 1619 /* 1620 * Stop async tasks. 1621 */ 1622 spa_async_suspend(spa); 1623 1624 if (spa->spa_root_vdev) { 1625 vdev_t *root_vdev = spa->spa_root_vdev; 1626 vdev_initialize_stop_all(root_vdev, 1627 VDEV_INITIALIZE_ACTIVE); 1628 vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE); 1629 vdev_autotrim_stop_all(spa); 1630 vdev_rebuild_stop_all(spa); 1631 } 1632 } 1633 1634 /* 1635 * Stop syncing. 1636 */ 1637 if (spa->spa_sync_on) { 1638 txg_sync_stop(spa->spa_dsl_pool); 1639 spa->spa_sync_on = B_FALSE; 1640 } 1641 1642 /* 1643 * This ensures that there is no async metaslab prefetching 1644 * while we attempt to unload the spa. 1645 */ 1646 if (spa->spa_root_vdev != NULL) { 1647 for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++) { 1648 vdev_t *vc = spa->spa_root_vdev->vdev_child[c]; 1649 if (vc->vdev_mg != NULL) 1650 taskq_wait(vc->vdev_mg->mg_taskq); 1651 } 1652 } 1653 1654 if (spa->spa_mmp.mmp_thread) 1655 mmp_thread_stop(spa); 1656 1657 /* 1658 * Wait for any outstanding async I/O to complete. 1659 */ 1660 if (spa->spa_async_zio_root != NULL) { 1661 for (int i = 0; i < max_ncpus; i++) 1662 (void) zio_wait(spa->spa_async_zio_root[i]); 1663 kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *)); 1664 spa->spa_async_zio_root = NULL; 1665 } 1666 1667 if (spa->spa_vdev_removal != NULL) { 1668 spa_vdev_removal_destroy(spa->spa_vdev_removal); 1669 spa->spa_vdev_removal = NULL; 1670 } 1671 1672 spa_destroy_aux_threads(spa); 1673 1674 spa_condense_fini(spa); 1675 1676 bpobj_close(&spa->spa_deferred_bpobj); 1677 1678 spa_config_enter(spa, SCL_ALL, spa, RW_WRITER); 1679 1680 /* 1681 * Close all vdevs. 1682 */ 1683 if (spa->spa_root_vdev) 1684 vdev_free(spa->spa_root_vdev); 1685 ASSERT(spa->spa_root_vdev == NULL); 1686 1687 /* 1688 * Close the dsl pool. 1689 */ 1690 if (spa->spa_dsl_pool) { 1691 dsl_pool_close(spa->spa_dsl_pool); 1692 spa->spa_dsl_pool = NULL; 1693 spa->spa_meta_objset = NULL; 1694 } 1695 1696 ddt_unload(spa); 1697 spa_unload_log_sm_metadata(spa); 1698 1699 /* 1700 * Drop and purge level 2 cache 1701 */ 1702 spa_l2cache_drop(spa); 1703 1704 for (int i = 0; i < spa->spa_spares.sav_count; i++) 1705 vdev_free(spa->spa_spares.sav_vdevs[i]); 1706 if (spa->spa_spares.sav_vdevs) { 1707 kmem_free(spa->spa_spares.sav_vdevs, 1708 spa->spa_spares.sav_count * sizeof (void *)); 1709 spa->spa_spares.sav_vdevs = NULL; 1710 } 1711 if (spa->spa_spares.sav_config) { 1712 nvlist_free(spa->spa_spares.sav_config); 1713 spa->spa_spares.sav_config = NULL; 1714 } 1715 spa->spa_spares.sav_count = 0; 1716 1717 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) { 1718 vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]); 1719 vdev_free(spa->spa_l2cache.sav_vdevs[i]); 1720 } 1721 if (spa->spa_l2cache.sav_vdevs) { 1722 kmem_free(spa->spa_l2cache.sav_vdevs, 1723 spa->spa_l2cache.sav_count * sizeof (void *)); 1724 spa->spa_l2cache.sav_vdevs = NULL; 1725 } 1726 if (spa->spa_l2cache.sav_config) { 1727 nvlist_free(spa->spa_l2cache.sav_config); 1728 spa->spa_l2cache.sav_config = NULL; 1729 } 1730 spa->spa_l2cache.sav_count = 0; 1731 1732 spa->spa_async_suspended = 0; 1733 1734 spa->spa_indirect_vdevs_loaded = B_FALSE; 1735 1736 if (spa->spa_comment != NULL) { 1737 spa_strfree(spa->spa_comment); 1738 spa->spa_comment = NULL; 1739 } 1740 if (spa->spa_compatibility != NULL) { 1741 spa_strfree(spa->spa_compatibility); 1742 spa->spa_compatibility = NULL; 1743 } 1744 1745 spa_config_exit(spa, SCL_ALL, spa); 1746 } 1747 1748 /* 1749 * Load (or re-load) the current list of vdevs describing the active spares for 1750 * this pool. When this is called, we have some form of basic information in 1751 * 'spa_spares.sav_config'. We parse this into vdevs, try to open them, and 1752 * then re-generate a more complete list including status information. 1753 */ 1754 void 1755 spa_load_spares(spa_t *spa) 1756 { 1757 nvlist_t **spares; 1758 uint_t nspares; 1759 int i; 1760 vdev_t *vd, *tvd; 1761 1762 #ifndef _KERNEL 1763 /* 1764 * zdb opens both the current state of the pool and the 1765 * checkpointed state (if present), with a different spa_t. 1766 * 1767 * As spare vdevs are shared among open pools, we skip loading 1768 * them when we load the checkpointed state of the pool. 1769 */ 1770 if (!spa_writeable(spa)) 1771 return; 1772 #endif 1773 1774 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 1775 1776 /* 1777 * First, close and free any existing spare vdevs. 1778 */ 1779 for (i = 0; i < spa->spa_spares.sav_count; i++) { 1780 vd = spa->spa_spares.sav_vdevs[i]; 1781 1782 /* Undo the call to spa_activate() below */ 1783 if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid, 1784 B_FALSE)) != NULL && tvd->vdev_isspare) 1785 spa_spare_remove(tvd); 1786 vdev_close(vd); 1787 vdev_free(vd); 1788 } 1789 1790 if (spa->spa_spares.sav_vdevs) 1791 kmem_free(spa->spa_spares.sav_vdevs, 1792 spa->spa_spares.sav_count * sizeof (void *)); 1793 1794 if (spa->spa_spares.sav_config == NULL) 1795 nspares = 0; 1796 else 1797 VERIFY0(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, 1798 ZPOOL_CONFIG_SPARES, &spares, &nspares)); 1799 1800 spa->spa_spares.sav_count = (int)nspares; 1801 spa->spa_spares.sav_vdevs = NULL; 1802 1803 if (nspares == 0) 1804 return; 1805 1806 /* 1807 * Construct the array of vdevs, opening them to get status in the 1808 * process. For each spare, there is potentially two different vdev_t 1809 * structures associated with it: one in the list of spares (used only 1810 * for basic validation purposes) and one in the active vdev 1811 * configuration (if it's spared in). During this phase we open and 1812 * validate each vdev on the spare list. If the vdev also exists in the 1813 * active configuration, then we also mark this vdev as an active spare. 1814 */ 1815 spa->spa_spares.sav_vdevs = kmem_zalloc(nspares * sizeof (void *), 1816 KM_SLEEP); 1817 for (i = 0; i < spa->spa_spares.sav_count; i++) { 1818 VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0, 1819 VDEV_ALLOC_SPARE) == 0); 1820 ASSERT(vd != NULL); 1821 1822 spa->spa_spares.sav_vdevs[i] = vd; 1823 1824 if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid, 1825 B_FALSE)) != NULL) { 1826 if (!tvd->vdev_isspare) 1827 spa_spare_add(tvd); 1828 1829 /* 1830 * We only mark the spare active if we were successfully 1831 * able to load the vdev. Otherwise, importing a pool 1832 * with a bad active spare would result in strange 1833 * behavior, because multiple pool would think the spare 1834 * is actively in use. 1835 * 1836 * There is a vulnerability here to an equally bizarre 1837 * circumstance, where a dead active spare is later 1838 * brought back to life (onlined or otherwise). Given 1839 * the rarity of this scenario, and the extra complexity 1840 * it adds, we ignore the possibility. 1841 */ 1842 if (!vdev_is_dead(tvd)) 1843 spa_spare_activate(tvd); 1844 } 1845 1846 vd->vdev_top = vd; 1847 vd->vdev_aux = &spa->spa_spares; 1848 1849 if (vdev_open(vd) != 0) 1850 continue; 1851 1852 if (vdev_validate_aux(vd) == 0) 1853 spa_spare_add(vd); 1854 } 1855 1856 /* 1857 * Recompute the stashed list of spares, with status information 1858 * this time. 1859 */ 1860 fnvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES); 1861 1862 spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *), 1863 KM_SLEEP); 1864 for (i = 0; i < spa->spa_spares.sav_count; i++) 1865 spares[i] = vdev_config_generate(spa, 1866 spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE); 1867 fnvlist_add_nvlist_array(spa->spa_spares.sav_config, 1868 ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares, 1869 spa->spa_spares.sav_count); 1870 for (i = 0; i < spa->spa_spares.sav_count; i++) 1871 nvlist_free(spares[i]); 1872 kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *)); 1873 } 1874 1875 /* 1876 * Load (or re-load) the current list of vdevs describing the active l2cache for 1877 * this pool. When this is called, we have some form of basic information in 1878 * 'spa_l2cache.sav_config'. We parse this into vdevs, try to open them, and 1879 * then re-generate a more complete list including status information. 1880 * Devices which are already active have their details maintained, and are 1881 * not re-opened. 1882 */ 1883 void 1884 spa_load_l2cache(spa_t *spa) 1885 { 1886 nvlist_t **l2cache = NULL; 1887 uint_t nl2cache; 1888 int i, j, oldnvdevs; 1889 uint64_t guid; 1890 vdev_t *vd, **oldvdevs, **newvdevs; 1891 spa_aux_vdev_t *sav = &spa->spa_l2cache; 1892 1893 #ifndef _KERNEL 1894 /* 1895 * zdb opens both the current state of the pool and the 1896 * checkpointed state (if present), with a different spa_t. 1897 * 1898 * As L2 caches are part of the ARC which is shared among open 1899 * pools, we skip loading them when we load the checkpointed 1900 * state of the pool. 1901 */ 1902 if (!spa_writeable(spa)) 1903 return; 1904 #endif 1905 1906 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 1907 1908 oldvdevs = sav->sav_vdevs; 1909 oldnvdevs = sav->sav_count; 1910 sav->sav_vdevs = NULL; 1911 sav->sav_count = 0; 1912 1913 if (sav->sav_config == NULL) { 1914 nl2cache = 0; 1915 newvdevs = NULL; 1916 goto out; 1917 } 1918 1919 VERIFY0(nvlist_lookup_nvlist_array(sav->sav_config, 1920 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache)); 1921 newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP); 1922 1923 /* 1924 * Process new nvlist of vdevs. 1925 */ 1926 for (i = 0; i < nl2cache; i++) { 1927 guid = fnvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID); 1928 1929 newvdevs[i] = NULL; 1930 for (j = 0; j < oldnvdevs; j++) { 1931 vd = oldvdevs[j]; 1932 if (vd != NULL && guid == vd->vdev_guid) { 1933 /* 1934 * Retain previous vdev for add/remove ops. 1935 */ 1936 newvdevs[i] = vd; 1937 oldvdevs[j] = NULL; 1938 break; 1939 } 1940 } 1941 1942 if (newvdevs[i] == NULL) { 1943 /* 1944 * Create new vdev 1945 */ 1946 VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0, 1947 VDEV_ALLOC_L2CACHE) == 0); 1948 ASSERT(vd != NULL); 1949 newvdevs[i] = vd; 1950 1951 /* 1952 * Commit this vdev as an l2cache device, 1953 * even if it fails to open. 1954 */ 1955 spa_l2cache_add(vd); 1956 1957 vd->vdev_top = vd; 1958 vd->vdev_aux = sav; 1959 1960 spa_l2cache_activate(vd); 1961 1962 if (vdev_open(vd) != 0) 1963 continue; 1964 1965 (void) vdev_validate_aux(vd); 1966 1967 if (!vdev_is_dead(vd)) 1968 l2arc_add_vdev(spa, vd); 1969 1970 /* 1971 * Upon cache device addition to a pool or pool 1972 * creation with a cache device or if the header 1973 * of the device is invalid we issue an async 1974 * TRIM command for the whole device which will 1975 * execute if l2arc_trim_ahead > 0. 1976 */ 1977 spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM); 1978 } 1979 } 1980 1981 sav->sav_vdevs = newvdevs; 1982 sav->sav_count = (int)nl2cache; 1983 1984 /* 1985 * Recompute the stashed list of l2cache devices, with status 1986 * information this time. 1987 */ 1988 fnvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE); 1989 1990 if (sav->sav_count > 0) 1991 l2cache = kmem_alloc(sav->sav_count * sizeof (void *), 1992 KM_SLEEP); 1993 for (i = 0; i < sav->sav_count; i++) 1994 l2cache[i] = vdev_config_generate(spa, 1995 sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE); 1996 fnvlist_add_nvlist_array(sav->sav_config, ZPOOL_CONFIG_L2CACHE, 1997 (const nvlist_t * const *)l2cache, sav->sav_count); 1998 1999 out: 2000 /* 2001 * Purge vdevs that were dropped 2002 */ 2003 for (i = 0; i < oldnvdevs; i++) { 2004 uint64_t pool; 2005 2006 vd = oldvdevs[i]; 2007 if (vd != NULL) { 2008 ASSERT(vd->vdev_isl2cache); 2009 2010 if (spa_l2cache_exists(vd->vdev_guid, &pool) && 2011 pool != 0ULL && l2arc_vdev_present(vd)) 2012 l2arc_remove_vdev(vd); 2013 vdev_clear_stats(vd); 2014 vdev_free(vd); 2015 } 2016 } 2017 2018 if (oldvdevs) 2019 kmem_free(oldvdevs, oldnvdevs * sizeof (void *)); 2020 2021 for (i = 0; i < sav->sav_count; i++) 2022 nvlist_free(l2cache[i]); 2023 if (sav->sav_count) 2024 kmem_free(l2cache, sav->sav_count * sizeof (void *)); 2025 } 2026 2027 static int 2028 load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value) 2029 { 2030 dmu_buf_t *db; 2031 char *packed = NULL; 2032 size_t nvsize = 0; 2033 int error; 2034 *value = NULL; 2035 2036 error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db); 2037 if (error) 2038 return (error); 2039 2040 nvsize = *(uint64_t *)db->db_data; 2041 dmu_buf_rele(db, FTAG); 2042 2043 packed = vmem_alloc(nvsize, KM_SLEEP); 2044 error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed, 2045 DMU_READ_PREFETCH); 2046 if (error == 0) 2047 error = nvlist_unpack(packed, nvsize, value, 0); 2048 vmem_free(packed, nvsize); 2049 2050 return (error); 2051 } 2052 2053 /* 2054 * Concrete top-level vdevs that are not missing and are not logs. At every 2055 * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds. 2056 */ 2057 static uint64_t 2058 spa_healthy_core_tvds(spa_t *spa) 2059 { 2060 vdev_t *rvd = spa->spa_root_vdev; 2061 uint64_t tvds = 0; 2062 2063 for (uint64_t i = 0; i < rvd->vdev_children; i++) { 2064 vdev_t *vd = rvd->vdev_child[i]; 2065 if (vd->vdev_islog) 2066 continue; 2067 if (vdev_is_concrete(vd) && !vdev_is_dead(vd)) 2068 tvds++; 2069 } 2070 2071 return (tvds); 2072 } 2073 2074 /* 2075 * Checks to see if the given vdev could not be opened, in which case we post a 2076 * sysevent to notify the autoreplace code that the device has been removed. 2077 */ 2078 static void 2079 spa_check_removed(vdev_t *vd) 2080 { 2081 for (uint64_t c = 0; c < vd->vdev_children; c++) 2082 spa_check_removed(vd->vdev_child[c]); 2083 2084 if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) && 2085 vdev_is_concrete(vd)) { 2086 zfs_post_autoreplace(vd->vdev_spa, vd); 2087 spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK); 2088 } 2089 } 2090 2091 static int 2092 spa_check_for_missing_logs(spa_t *spa) 2093 { 2094 vdev_t *rvd = spa->spa_root_vdev; 2095 2096 /* 2097 * If we're doing a normal import, then build up any additional 2098 * diagnostic information about missing log devices. 2099 * We'll pass this up to the user for further processing. 2100 */ 2101 if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) { 2102 nvlist_t **child, *nv; 2103 uint64_t idx = 0; 2104 2105 child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t *), 2106 KM_SLEEP); 2107 nv = fnvlist_alloc(); 2108 2109 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 2110 vdev_t *tvd = rvd->vdev_child[c]; 2111 2112 /* 2113 * We consider a device as missing only if it failed 2114 * to open (i.e. offline or faulted is not considered 2115 * as missing). 2116 */ 2117 if (tvd->vdev_islog && 2118 tvd->vdev_state == VDEV_STATE_CANT_OPEN) { 2119 child[idx++] = vdev_config_generate(spa, tvd, 2120 B_FALSE, VDEV_CONFIG_MISSING); 2121 } 2122 } 2123 2124 if (idx > 0) { 2125 fnvlist_add_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN, 2126 (const nvlist_t * const *)child, idx); 2127 fnvlist_add_nvlist(spa->spa_load_info, 2128 ZPOOL_CONFIG_MISSING_DEVICES, nv); 2129 2130 for (uint64_t i = 0; i < idx; i++) 2131 nvlist_free(child[i]); 2132 } 2133 nvlist_free(nv); 2134 kmem_free(child, rvd->vdev_children * sizeof (char **)); 2135 2136 if (idx > 0) { 2137 spa_load_failed(spa, "some log devices are missing"); 2138 vdev_dbgmsg_print_tree(rvd, 2); 2139 return (SET_ERROR(ENXIO)); 2140 } 2141 } else { 2142 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 2143 vdev_t *tvd = rvd->vdev_child[c]; 2144 2145 if (tvd->vdev_islog && 2146 tvd->vdev_state == VDEV_STATE_CANT_OPEN) { 2147 spa_set_log_state(spa, SPA_LOG_CLEAR); 2148 spa_load_note(spa, "some log devices are " 2149 "missing, ZIL is dropped."); 2150 vdev_dbgmsg_print_tree(rvd, 2); 2151 break; 2152 } 2153 } 2154 } 2155 2156 return (0); 2157 } 2158 2159 /* 2160 * Check for missing log devices 2161 */ 2162 static boolean_t 2163 spa_check_logs(spa_t *spa) 2164 { 2165 boolean_t rv = B_FALSE; 2166 dsl_pool_t *dp = spa_get_dsl(spa); 2167 2168 switch (spa->spa_log_state) { 2169 default: 2170 break; 2171 case SPA_LOG_MISSING: 2172 /* need to recheck in case slog has been restored */ 2173 case SPA_LOG_UNKNOWN: 2174 rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 2175 zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0); 2176 if (rv) 2177 spa_set_log_state(spa, SPA_LOG_MISSING); 2178 break; 2179 } 2180 return (rv); 2181 } 2182 2183 /* 2184 * Passivate any log vdevs (note, does not apply to embedded log metaslabs). 2185 */ 2186 static boolean_t 2187 spa_passivate_log(spa_t *spa) 2188 { 2189 vdev_t *rvd = spa->spa_root_vdev; 2190 boolean_t slog_found = B_FALSE; 2191 2192 ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER)); 2193 2194 for (int c = 0; c < rvd->vdev_children; c++) { 2195 vdev_t *tvd = rvd->vdev_child[c]; 2196 2197 if (tvd->vdev_islog) { 2198 ASSERT3P(tvd->vdev_log_mg, ==, NULL); 2199 metaslab_group_passivate(tvd->vdev_mg); 2200 slog_found = B_TRUE; 2201 } 2202 } 2203 2204 return (slog_found); 2205 } 2206 2207 /* 2208 * Activate any log vdevs (note, does not apply to embedded log metaslabs). 2209 */ 2210 static void 2211 spa_activate_log(spa_t *spa) 2212 { 2213 vdev_t *rvd = spa->spa_root_vdev; 2214 2215 ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER)); 2216 2217 for (int c = 0; c < rvd->vdev_children; c++) { 2218 vdev_t *tvd = rvd->vdev_child[c]; 2219 2220 if (tvd->vdev_islog) { 2221 ASSERT3P(tvd->vdev_log_mg, ==, NULL); 2222 metaslab_group_activate(tvd->vdev_mg); 2223 } 2224 } 2225 } 2226 2227 int 2228 spa_reset_logs(spa_t *spa) 2229 { 2230 int error; 2231 2232 error = dmu_objset_find(spa_name(spa), zil_reset, 2233 NULL, DS_FIND_CHILDREN); 2234 if (error == 0) { 2235 /* 2236 * We successfully offlined the log device, sync out the 2237 * current txg so that the "stubby" block can be removed 2238 * by zil_sync(). 2239 */ 2240 txg_wait_synced(spa->spa_dsl_pool, 0); 2241 } 2242 return (error); 2243 } 2244 2245 static void 2246 spa_aux_check_removed(spa_aux_vdev_t *sav) 2247 { 2248 for (int i = 0; i < sav->sav_count; i++) 2249 spa_check_removed(sav->sav_vdevs[i]); 2250 } 2251 2252 void 2253 spa_claim_notify(zio_t *zio) 2254 { 2255 spa_t *spa = zio->io_spa; 2256 2257 if (zio->io_error) 2258 return; 2259 2260 mutex_enter(&spa->spa_props_lock); /* any mutex will do */ 2261 if (spa->spa_claim_max_txg < zio->io_bp->blk_birth) 2262 spa->spa_claim_max_txg = zio->io_bp->blk_birth; 2263 mutex_exit(&spa->spa_props_lock); 2264 } 2265 2266 typedef struct spa_load_error { 2267 boolean_t sle_verify_data; 2268 uint64_t sle_meta_count; 2269 uint64_t sle_data_count; 2270 } spa_load_error_t; 2271 2272 static void 2273 spa_load_verify_done(zio_t *zio) 2274 { 2275 blkptr_t *bp = zio->io_bp; 2276 spa_load_error_t *sle = zio->io_private; 2277 dmu_object_type_t type = BP_GET_TYPE(bp); 2278 int error = zio->io_error; 2279 spa_t *spa = zio->io_spa; 2280 2281 abd_free(zio->io_abd); 2282 if (error) { 2283 if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) && 2284 type != DMU_OT_INTENT_LOG) 2285 atomic_inc_64(&sle->sle_meta_count); 2286 else 2287 atomic_inc_64(&sle->sle_data_count); 2288 } 2289 2290 mutex_enter(&spa->spa_scrub_lock); 2291 spa->spa_load_verify_bytes -= BP_GET_PSIZE(bp); 2292 cv_broadcast(&spa->spa_scrub_io_cv); 2293 mutex_exit(&spa->spa_scrub_lock); 2294 } 2295 2296 /* 2297 * Maximum number of inflight bytes is the log2 fraction of the arc size. 2298 * By default, we set it to 1/16th of the arc. 2299 */ 2300 static int spa_load_verify_shift = 4; 2301 static int spa_load_verify_metadata = B_TRUE; 2302 static int spa_load_verify_data = B_TRUE; 2303 2304 static int 2305 spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 2306 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) 2307 { 2308 zio_t *rio = arg; 2309 spa_load_error_t *sle = rio->io_private; 2310 2311 (void) zilog, (void) dnp; 2312 2313 if (zb->zb_level == ZB_DNODE_LEVEL || BP_IS_HOLE(bp) || 2314 BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp)) 2315 return (0); 2316 /* 2317 * Note: normally this routine will not be called if 2318 * spa_load_verify_metadata is not set. However, it may be useful 2319 * to manually set the flag after the traversal has begun. 2320 */ 2321 if (!spa_load_verify_metadata) 2322 return (0); 2323 if (!BP_IS_METADATA(bp) && 2324 (!spa_load_verify_data || !sle->sle_verify_data)) 2325 return (0); 2326 2327 uint64_t maxinflight_bytes = 2328 arc_target_bytes() >> spa_load_verify_shift; 2329 size_t size = BP_GET_PSIZE(bp); 2330 2331 mutex_enter(&spa->spa_scrub_lock); 2332 while (spa->spa_load_verify_bytes >= maxinflight_bytes) 2333 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 2334 spa->spa_load_verify_bytes += size; 2335 mutex_exit(&spa->spa_scrub_lock); 2336 2337 zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size, 2338 spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB, 2339 ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL | 2340 ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb)); 2341 return (0); 2342 } 2343 2344 static int 2345 verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 2346 { 2347 (void) dp, (void) arg; 2348 2349 if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN) 2350 return (SET_ERROR(ENAMETOOLONG)); 2351 2352 return (0); 2353 } 2354 2355 static int 2356 spa_load_verify(spa_t *spa) 2357 { 2358 zio_t *rio; 2359 spa_load_error_t sle = { 0 }; 2360 zpool_load_policy_t policy; 2361 boolean_t verify_ok = B_FALSE; 2362 int error = 0; 2363 2364 zpool_get_load_policy(spa->spa_config, &policy); 2365 2366 if (policy.zlp_rewind & ZPOOL_NEVER_REWIND || 2367 policy.zlp_maxmeta == UINT64_MAX) 2368 return (0); 2369 2370 dsl_pool_config_enter(spa->spa_dsl_pool, FTAG); 2371 error = dmu_objset_find_dp(spa->spa_dsl_pool, 2372 spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL, 2373 DS_FIND_CHILDREN); 2374 dsl_pool_config_exit(spa->spa_dsl_pool, FTAG); 2375 if (error != 0) 2376 return (error); 2377 2378 /* 2379 * Verify data only if we are rewinding or error limit was set. 2380 * Otherwise nothing except dbgmsg care about it to waste time. 2381 */ 2382 sle.sle_verify_data = (policy.zlp_rewind & ZPOOL_REWIND_MASK) || 2383 (policy.zlp_maxdata < UINT64_MAX); 2384 2385 rio = zio_root(spa, NULL, &sle, 2386 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE); 2387 2388 if (spa_load_verify_metadata) { 2389 if (spa->spa_extreme_rewind) { 2390 spa_load_note(spa, "performing a complete scan of the " 2391 "pool since extreme rewind is on. This may take " 2392 "a very long time.\n (spa_load_verify_data=%u, " 2393 "spa_load_verify_metadata=%u)", 2394 spa_load_verify_data, spa_load_verify_metadata); 2395 } 2396 2397 error = traverse_pool(spa, spa->spa_verify_min_txg, 2398 TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA | 2399 TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio); 2400 } 2401 2402 (void) zio_wait(rio); 2403 ASSERT0(spa->spa_load_verify_bytes); 2404 2405 spa->spa_load_meta_errors = sle.sle_meta_count; 2406 spa->spa_load_data_errors = sle.sle_data_count; 2407 2408 if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) { 2409 spa_load_note(spa, "spa_load_verify found %llu metadata errors " 2410 "and %llu data errors", (u_longlong_t)sle.sle_meta_count, 2411 (u_longlong_t)sle.sle_data_count); 2412 } 2413 2414 if (spa_load_verify_dryrun || 2415 (!error && sle.sle_meta_count <= policy.zlp_maxmeta && 2416 sle.sle_data_count <= policy.zlp_maxdata)) { 2417 int64_t loss = 0; 2418 2419 verify_ok = B_TRUE; 2420 spa->spa_load_txg = spa->spa_uberblock.ub_txg; 2421 spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp; 2422 2423 loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts; 2424 fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_LOAD_TIME, 2425 spa->spa_load_txg_ts); 2426 fnvlist_add_int64(spa->spa_load_info, ZPOOL_CONFIG_REWIND_TIME, 2427 loss); 2428 fnvlist_add_uint64(spa->spa_load_info, 2429 ZPOOL_CONFIG_LOAD_META_ERRORS, sle.sle_meta_count); 2430 fnvlist_add_uint64(spa->spa_load_info, 2431 ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count); 2432 } else { 2433 spa->spa_load_max_txg = spa->spa_uberblock.ub_txg; 2434 } 2435 2436 if (spa_load_verify_dryrun) 2437 return (0); 2438 2439 if (error) { 2440 if (error != ENXIO && error != EIO) 2441 error = SET_ERROR(EIO); 2442 return (error); 2443 } 2444 2445 return (verify_ok ? 0 : EIO); 2446 } 2447 2448 /* 2449 * Find a value in the pool props object. 2450 */ 2451 static void 2452 spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val) 2453 { 2454 (void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object, 2455 zpool_prop_to_name(prop), sizeof (uint64_t), 1, val); 2456 } 2457 2458 /* 2459 * Find a value in the pool directory object. 2460 */ 2461 static int 2462 spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent) 2463 { 2464 int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 2465 name, sizeof (uint64_t), 1, val); 2466 2467 if (error != 0 && (error != ENOENT || log_enoent)) { 2468 spa_load_failed(spa, "couldn't get '%s' value in MOS directory " 2469 "[error=%d]", name, error); 2470 } 2471 2472 return (error); 2473 } 2474 2475 static int 2476 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err) 2477 { 2478 vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux); 2479 return (SET_ERROR(err)); 2480 } 2481 2482 boolean_t 2483 spa_livelist_delete_check(spa_t *spa) 2484 { 2485 return (spa->spa_livelists_to_delete != 0); 2486 } 2487 2488 static boolean_t 2489 spa_livelist_delete_cb_check(void *arg, zthr_t *z) 2490 { 2491 (void) z; 2492 spa_t *spa = arg; 2493 return (spa_livelist_delete_check(spa)); 2494 } 2495 2496 static int 2497 delete_blkptr_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 2498 { 2499 spa_t *spa = arg; 2500 zio_free(spa, tx->tx_txg, bp); 2501 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, 2502 -bp_get_dsize_sync(spa, bp), 2503 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); 2504 return (0); 2505 } 2506 2507 static int 2508 dsl_get_next_livelist_obj(objset_t *os, uint64_t zap_obj, uint64_t *llp) 2509 { 2510 int err; 2511 zap_cursor_t zc; 2512 zap_attribute_t za; 2513 zap_cursor_init(&zc, os, zap_obj); 2514 err = zap_cursor_retrieve(&zc, &za); 2515 zap_cursor_fini(&zc); 2516 if (err == 0) 2517 *llp = za.za_first_integer; 2518 return (err); 2519 } 2520 2521 /* 2522 * Components of livelist deletion that must be performed in syncing 2523 * context: freeing block pointers and updating the pool-wide data 2524 * structures to indicate how much work is left to do 2525 */ 2526 typedef struct sublist_delete_arg { 2527 spa_t *spa; 2528 dsl_deadlist_t *ll; 2529 uint64_t key; 2530 bplist_t *to_free; 2531 } sublist_delete_arg_t; 2532 2533 static void 2534 sublist_delete_sync(void *arg, dmu_tx_t *tx) 2535 { 2536 sublist_delete_arg_t *sda = arg; 2537 spa_t *spa = sda->spa; 2538 dsl_deadlist_t *ll = sda->ll; 2539 uint64_t key = sda->key; 2540 bplist_t *to_free = sda->to_free; 2541 2542 bplist_iterate(to_free, delete_blkptr_cb, spa, tx); 2543 dsl_deadlist_remove_entry(ll, key, tx); 2544 } 2545 2546 typedef struct livelist_delete_arg { 2547 spa_t *spa; 2548 uint64_t ll_obj; 2549 uint64_t zap_obj; 2550 } livelist_delete_arg_t; 2551 2552 static void 2553 livelist_delete_sync(void *arg, dmu_tx_t *tx) 2554 { 2555 livelist_delete_arg_t *lda = arg; 2556 spa_t *spa = lda->spa; 2557 uint64_t ll_obj = lda->ll_obj; 2558 uint64_t zap_obj = lda->zap_obj; 2559 objset_t *mos = spa->spa_meta_objset; 2560 uint64_t count; 2561 2562 /* free the livelist and decrement the feature count */ 2563 VERIFY0(zap_remove_int(mos, zap_obj, ll_obj, tx)); 2564 dsl_deadlist_free(mos, ll_obj, tx); 2565 spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx); 2566 VERIFY0(zap_count(mos, zap_obj, &count)); 2567 if (count == 0) { 2568 /* no more livelists to delete */ 2569 VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT, 2570 DMU_POOL_DELETED_CLONES, tx)); 2571 VERIFY0(zap_destroy(mos, zap_obj, tx)); 2572 spa->spa_livelists_to_delete = 0; 2573 spa_notify_waiters(spa); 2574 } 2575 } 2576 2577 /* 2578 * Load in the value for the livelist to be removed and open it. Then, 2579 * load its first sublist and determine which block pointers should actually 2580 * be freed. Then, call a synctask which performs the actual frees and updates 2581 * the pool-wide livelist data. 2582 */ 2583 static void 2584 spa_livelist_delete_cb(void *arg, zthr_t *z) 2585 { 2586 spa_t *spa = arg; 2587 uint64_t ll_obj = 0, count; 2588 objset_t *mos = spa->spa_meta_objset; 2589 uint64_t zap_obj = spa->spa_livelists_to_delete; 2590 /* 2591 * Determine the next livelist to delete. This function should only 2592 * be called if there is at least one deleted clone. 2593 */ 2594 VERIFY0(dsl_get_next_livelist_obj(mos, zap_obj, &ll_obj)); 2595 VERIFY0(zap_count(mos, ll_obj, &count)); 2596 if (count > 0) { 2597 dsl_deadlist_t *ll; 2598 dsl_deadlist_entry_t *dle; 2599 bplist_t to_free; 2600 ll = kmem_zalloc(sizeof (dsl_deadlist_t), KM_SLEEP); 2601 dsl_deadlist_open(ll, mos, ll_obj); 2602 dle = dsl_deadlist_first(ll); 2603 ASSERT3P(dle, !=, NULL); 2604 bplist_create(&to_free); 2605 int err = dsl_process_sub_livelist(&dle->dle_bpobj, &to_free, 2606 z, NULL); 2607 if (err == 0) { 2608 sublist_delete_arg_t sync_arg = { 2609 .spa = spa, 2610 .ll = ll, 2611 .key = dle->dle_mintxg, 2612 .to_free = &to_free 2613 }; 2614 zfs_dbgmsg("deleting sublist (id %llu) from" 2615 " livelist %llu, %lld remaining", 2616 (u_longlong_t)dle->dle_bpobj.bpo_object, 2617 (u_longlong_t)ll_obj, (longlong_t)count - 1); 2618 VERIFY0(dsl_sync_task(spa_name(spa), NULL, 2619 sublist_delete_sync, &sync_arg, 0, 2620 ZFS_SPACE_CHECK_DESTROY)); 2621 } else { 2622 VERIFY3U(err, ==, EINTR); 2623 } 2624 bplist_clear(&to_free); 2625 bplist_destroy(&to_free); 2626 dsl_deadlist_close(ll); 2627 kmem_free(ll, sizeof (dsl_deadlist_t)); 2628 } else { 2629 livelist_delete_arg_t sync_arg = { 2630 .spa = spa, 2631 .ll_obj = ll_obj, 2632 .zap_obj = zap_obj 2633 }; 2634 zfs_dbgmsg("deletion of livelist %llu completed", 2635 (u_longlong_t)ll_obj); 2636 VERIFY0(dsl_sync_task(spa_name(spa), NULL, livelist_delete_sync, 2637 &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY)); 2638 } 2639 } 2640 2641 static void 2642 spa_start_livelist_destroy_thread(spa_t *spa) 2643 { 2644 ASSERT3P(spa->spa_livelist_delete_zthr, ==, NULL); 2645 spa->spa_livelist_delete_zthr = 2646 zthr_create("z_livelist_destroy", 2647 spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa, 2648 minclsyspri); 2649 } 2650 2651 typedef struct livelist_new_arg { 2652 bplist_t *allocs; 2653 bplist_t *frees; 2654 } livelist_new_arg_t; 2655 2656 static int 2657 livelist_track_new_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 2658 dmu_tx_t *tx) 2659 { 2660 ASSERT(tx == NULL); 2661 livelist_new_arg_t *lna = arg; 2662 if (bp_freed) { 2663 bplist_append(lna->frees, bp); 2664 } else { 2665 bplist_append(lna->allocs, bp); 2666 zfs_livelist_condense_new_alloc++; 2667 } 2668 return (0); 2669 } 2670 2671 typedef struct livelist_condense_arg { 2672 spa_t *spa; 2673 bplist_t to_keep; 2674 uint64_t first_size; 2675 uint64_t next_size; 2676 } livelist_condense_arg_t; 2677 2678 static void 2679 spa_livelist_condense_sync(void *arg, dmu_tx_t *tx) 2680 { 2681 livelist_condense_arg_t *lca = arg; 2682 spa_t *spa = lca->spa; 2683 bplist_t new_frees; 2684 dsl_dataset_t *ds = spa->spa_to_condense.ds; 2685 2686 /* Have we been cancelled? */ 2687 if (spa->spa_to_condense.cancelled) { 2688 zfs_livelist_condense_sync_cancel++; 2689 goto out; 2690 } 2691 2692 dsl_deadlist_entry_t *first = spa->spa_to_condense.first; 2693 dsl_deadlist_entry_t *next = spa->spa_to_condense.next; 2694 dsl_deadlist_t *ll = &ds->ds_dir->dd_livelist; 2695 2696 /* 2697 * It's possible that the livelist was changed while the zthr was 2698 * running. Therefore, we need to check for new blkptrs in the two 2699 * entries being condensed and continue to track them in the livelist. 2700 * Because of the way we handle remapped blkptrs (see dbuf_remap_impl), 2701 * it's possible that the newly added blkptrs are FREEs or ALLOCs so 2702 * we need to sort them into two different bplists. 2703 */ 2704 uint64_t first_obj = first->dle_bpobj.bpo_object; 2705 uint64_t next_obj = next->dle_bpobj.bpo_object; 2706 uint64_t cur_first_size = first->dle_bpobj.bpo_phys->bpo_num_blkptrs; 2707 uint64_t cur_next_size = next->dle_bpobj.bpo_phys->bpo_num_blkptrs; 2708 2709 bplist_create(&new_frees); 2710 livelist_new_arg_t new_bps = { 2711 .allocs = &lca->to_keep, 2712 .frees = &new_frees, 2713 }; 2714 2715 if (cur_first_size > lca->first_size) { 2716 VERIFY0(livelist_bpobj_iterate_from_nofree(&first->dle_bpobj, 2717 livelist_track_new_cb, &new_bps, lca->first_size)); 2718 } 2719 if (cur_next_size > lca->next_size) { 2720 VERIFY0(livelist_bpobj_iterate_from_nofree(&next->dle_bpobj, 2721 livelist_track_new_cb, &new_bps, lca->next_size)); 2722 } 2723 2724 dsl_deadlist_clear_entry(first, ll, tx); 2725 ASSERT(bpobj_is_empty(&first->dle_bpobj)); 2726 dsl_deadlist_remove_entry(ll, next->dle_mintxg, tx); 2727 2728 bplist_iterate(&lca->to_keep, dsl_deadlist_insert_alloc_cb, ll, tx); 2729 bplist_iterate(&new_frees, dsl_deadlist_insert_free_cb, ll, tx); 2730 bplist_destroy(&new_frees); 2731 2732 char dsname[ZFS_MAX_DATASET_NAME_LEN]; 2733 dsl_dataset_name(ds, dsname); 2734 zfs_dbgmsg("txg %llu condensing livelist of %s (id %llu), bpobj %llu " 2735 "(%llu blkptrs) and bpobj %llu (%llu blkptrs) -> bpobj %llu " 2736 "(%llu blkptrs)", (u_longlong_t)tx->tx_txg, dsname, 2737 (u_longlong_t)ds->ds_object, (u_longlong_t)first_obj, 2738 (u_longlong_t)cur_first_size, (u_longlong_t)next_obj, 2739 (u_longlong_t)cur_next_size, 2740 (u_longlong_t)first->dle_bpobj.bpo_object, 2741 (u_longlong_t)first->dle_bpobj.bpo_phys->bpo_num_blkptrs); 2742 out: 2743 dmu_buf_rele(ds->ds_dbuf, spa); 2744 spa->spa_to_condense.ds = NULL; 2745 bplist_clear(&lca->to_keep); 2746 bplist_destroy(&lca->to_keep); 2747 kmem_free(lca, sizeof (livelist_condense_arg_t)); 2748 spa->spa_to_condense.syncing = B_FALSE; 2749 } 2750 2751 static void 2752 spa_livelist_condense_cb(void *arg, zthr_t *t) 2753 { 2754 while (zfs_livelist_condense_zthr_pause && 2755 !(zthr_has_waiters(t) || zthr_iscancelled(t))) 2756 delay(1); 2757 2758 spa_t *spa = arg; 2759 dsl_deadlist_entry_t *first = spa->spa_to_condense.first; 2760 dsl_deadlist_entry_t *next = spa->spa_to_condense.next; 2761 uint64_t first_size, next_size; 2762 2763 livelist_condense_arg_t *lca = 2764 kmem_alloc(sizeof (livelist_condense_arg_t), KM_SLEEP); 2765 bplist_create(&lca->to_keep); 2766 2767 /* 2768 * Process the livelists (matching FREEs and ALLOCs) in open context 2769 * so we have minimal work in syncing context to condense. 2770 * 2771 * We save bpobj sizes (first_size and next_size) to use later in 2772 * syncing context to determine if entries were added to these sublists 2773 * while in open context. This is possible because the clone is still 2774 * active and open for normal writes and we want to make sure the new, 2775 * unprocessed blockpointers are inserted into the livelist normally. 2776 * 2777 * Note that dsl_process_sub_livelist() both stores the size number of 2778 * blockpointers and iterates over them while the bpobj's lock held, so 2779 * the sizes returned to us are consistent which what was actually 2780 * processed. 2781 */ 2782 int err = dsl_process_sub_livelist(&first->dle_bpobj, &lca->to_keep, t, 2783 &first_size); 2784 if (err == 0) 2785 err = dsl_process_sub_livelist(&next->dle_bpobj, &lca->to_keep, 2786 t, &next_size); 2787 2788 if (err == 0) { 2789 while (zfs_livelist_condense_sync_pause && 2790 !(zthr_has_waiters(t) || zthr_iscancelled(t))) 2791 delay(1); 2792 2793 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 2794 dmu_tx_mark_netfree(tx); 2795 dmu_tx_hold_space(tx, 1); 2796 err = dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE); 2797 if (err == 0) { 2798 /* 2799 * Prevent the condense zthr restarting before 2800 * the synctask completes. 2801 */ 2802 spa->spa_to_condense.syncing = B_TRUE; 2803 lca->spa = spa; 2804 lca->first_size = first_size; 2805 lca->next_size = next_size; 2806 dsl_sync_task_nowait(spa_get_dsl(spa), 2807 spa_livelist_condense_sync, lca, tx); 2808 dmu_tx_commit(tx); 2809 return; 2810 } 2811 } 2812 /* 2813 * Condensing can not continue: either it was externally stopped or 2814 * we were unable to assign to a tx because the pool has run out of 2815 * space. In the second case, we'll just end up trying to condense 2816 * again in a later txg. 2817 */ 2818 ASSERT(err != 0); 2819 bplist_clear(&lca->to_keep); 2820 bplist_destroy(&lca->to_keep); 2821 kmem_free(lca, sizeof (livelist_condense_arg_t)); 2822 dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf, spa); 2823 spa->spa_to_condense.ds = NULL; 2824 if (err == EINTR) 2825 zfs_livelist_condense_zthr_cancel++; 2826 } 2827 2828 /* 2829 * Check that there is something to condense but that a condense is not 2830 * already in progress and that condensing has not been cancelled. 2831 */ 2832 static boolean_t 2833 spa_livelist_condense_cb_check(void *arg, zthr_t *z) 2834 { 2835 (void) z; 2836 spa_t *spa = arg; 2837 if ((spa->spa_to_condense.ds != NULL) && 2838 (spa->spa_to_condense.syncing == B_FALSE) && 2839 (spa->spa_to_condense.cancelled == B_FALSE)) { 2840 return (B_TRUE); 2841 } 2842 return (B_FALSE); 2843 } 2844 2845 static void 2846 spa_start_livelist_condensing_thread(spa_t *spa) 2847 { 2848 spa->spa_to_condense.ds = NULL; 2849 spa->spa_to_condense.first = NULL; 2850 spa->spa_to_condense.next = NULL; 2851 spa->spa_to_condense.syncing = B_FALSE; 2852 spa->spa_to_condense.cancelled = B_FALSE; 2853 2854 ASSERT3P(spa->spa_livelist_condense_zthr, ==, NULL); 2855 spa->spa_livelist_condense_zthr = 2856 zthr_create("z_livelist_condense", 2857 spa_livelist_condense_cb_check, 2858 spa_livelist_condense_cb, spa, minclsyspri); 2859 } 2860 2861 static void 2862 spa_spawn_aux_threads(spa_t *spa) 2863 { 2864 ASSERT(spa_writeable(spa)); 2865 2866 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 2867 2868 spa_start_indirect_condensing_thread(spa); 2869 spa_start_livelist_destroy_thread(spa); 2870 spa_start_livelist_condensing_thread(spa); 2871 2872 ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL); 2873 spa->spa_checkpoint_discard_zthr = 2874 zthr_create("z_checkpoint_discard", 2875 spa_checkpoint_discard_thread_check, 2876 spa_checkpoint_discard_thread, spa, minclsyspri); 2877 } 2878 2879 /* 2880 * Fix up config after a partly-completed split. This is done with the 2881 * ZPOOL_CONFIG_SPLIT nvlist. Both the splitting pool and the split-off 2882 * pool have that entry in their config, but only the splitting one contains 2883 * a list of all the guids of the vdevs that are being split off. 2884 * 2885 * This function determines what to do with that list: either rejoin 2886 * all the disks to the pool, or complete the splitting process. To attempt 2887 * the rejoin, each disk that is offlined is marked online again, and 2888 * we do a reopen() call. If the vdev label for every disk that was 2889 * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL) 2890 * then we call vdev_split() on each disk, and complete the split. 2891 * 2892 * Otherwise we leave the config alone, with all the vdevs in place in 2893 * the original pool. 2894 */ 2895 static void 2896 spa_try_repair(spa_t *spa, nvlist_t *config) 2897 { 2898 uint_t extracted; 2899 uint64_t *glist; 2900 uint_t i, gcount; 2901 nvlist_t *nvl; 2902 vdev_t **vd; 2903 boolean_t attempt_reopen; 2904 2905 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0) 2906 return; 2907 2908 /* check that the config is complete */ 2909 if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, 2910 &glist, &gcount) != 0) 2911 return; 2912 2913 vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP); 2914 2915 /* attempt to online all the vdevs & validate */ 2916 attempt_reopen = B_TRUE; 2917 for (i = 0; i < gcount; i++) { 2918 if (glist[i] == 0) /* vdev is hole */ 2919 continue; 2920 2921 vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE); 2922 if (vd[i] == NULL) { 2923 /* 2924 * Don't bother attempting to reopen the disks; 2925 * just do the split. 2926 */ 2927 attempt_reopen = B_FALSE; 2928 } else { 2929 /* attempt to re-online it */ 2930 vd[i]->vdev_offline = B_FALSE; 2931 } 2932 } 2933 2934 if (attempt_reopen) { 2935 vdev_reopen(spa->spa_root_vdev); 2936 2937 /* check each device to see what state it's in */ 2938 for (extracted = 0, i = 0; i < gcount; i++) { 2939 if (vd[i] != NULL && 2940 vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL) 2941 break; 2942 ++extracted; 2943 } 2944 } 2945 2946 /* 2947 * If every disk has been moved to the new pool, or if we never 2948 * even attempted to look at them, then we split them off for 2949 * good. 2950 */ 2951 if (!attempt_reopen || gcount == extracted) { 2952 for (i = 0; i < gcount; i++) 2953 if (vd[i] != NULL) 2954 vdev_split(vd[i]); 2955 vdev_reopen(spa->spa_root_vdev); 2956 } 2957 2958 kmem_free(vd, gcount * sizeof (vdev_t *)); 2959 } 2960 2961 static int 2962 spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type) 2963 { 2964 char *ereport = FM_EREPORT_ZFS_POOL; 2965 int error; 2966 2967 spa->spa_load_state = state; 2968 (void) spa_import_progress_set_state(spa_guid(spa), 2969 spa_load_state(spa)); 2970 2971 gethrestime(&spa->spa_loaded_ts); 2972 error = spa_load_impl(spa, type, &ereport); 2973 2974 /* 2975 * Don't count references from objsets that are already closed 2976 * and are making their way through the eviction process. 2977 */ 2978 spa_evicting_os_wait(spa); 2979 spa->spa_minref = zfs_refcount_count(&spa->spa_refcount); 2980 if (error) { 2981 if (error != EEXIST) { 2982 spa->spa_loaded_ts.tv_sec = 0; 2983 spa->spa_loaded_ts.tv_nsec = 0; 2984 } 2985 if (error != EBADF) { 2986 (void) zfs_ereport_post(ereport, spa, 2987 NULL, NULL, NULL, 0); 2988 } 2989 } 2990 spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE; 2991 spa->spa_ena = 0; 2992 2993 (void) spa_import_progress_set_state(spa_guid(spa), 2994 spa_load_state(spa)); 2995 2996 return (error); 2997 } 2998 2999 #ifdef ZFS_DEBUG 3000 /* 3001 * Count the number of per-vdev ZAPs associated with all of the vdevs in the 3002 * vdev tree rooted in the given vd, and ensure that each ZAP is present in the 3003 * spa's per-vdev ZAP list. 3004 */ 3005 static uint64_t 3006 vdev_count_verify_zaps(vdev_t *vd) 3007 { 3008 spa_t *spa = vd->vdev_spa; 3009 uint64_t total = 0; 3010 3011 if (vd->vdev_top_zap != 0) { 3012 total++; 3013 ASSERT0(zap_lookup_int(spa->spa_meta_objset, 3014 spa->spa_all_vdev_zaps, vd->vdev_top_zap)); 3015 } 3016 if (vd->vdev_leaf_zap != 0) { 3017 total++; 3018 ASSERT0(zap_lookup_int(spa->spa_meta_objset, 3019 spa->spa_all_vdev_zaps, vd->vdev_leaf_zap)); 3020 } 3021 3022 for (uint64_t i = 0; i < vd->vdev_children; i++) { 3023 total += vdev_count_verify_zaps(vd->vdev_child[i]); 3024 } 3025 3026 return (total); 3027 } 3028 #else 3029 #define vdev_count_verify_zaps(vd) ((void) sizeof (vd), 0) 3030 #endif 3031 3032 /* 3033 * Determine whether the activity check is required. 3034 */ 3035 static boolean_t 3036 spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label, 3037 nvlist_t *config) 3038 { 3039 uint64_t state = 0; 3040 uint64_t hostid = 0; 3041 uint64_t tryconfig_txg = 0; 3042 uint64_t tryconfig_timestamp = 0; 3043 uint16_t tryconfig_mmp_seq = 0; 3044 nvlist_t *nvinfo; 3045 3046 if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) { 3047 nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO); 3048 (void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG, 3049 &tryconfig_txg); 3050 (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP, 3051 &tryconfig_timestamp); 3052 (void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ, 3053 &tryconfig_mmp_seq); 3054 } 3055 3056 (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state); 3057 3058 /* 3059 * Disable the MMP activity check - This is used by zdb which 3060 * is intended to be used on potentially active pools. 3061 */ 3062 if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) 3063 return (B_FALSE); 3064 3065 /* 3066 * Skip the activity check when the MMP feature is disabled. 3067 */ 3068 if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0) 3069 return (B_FALSE); 3070 3071 /* 3072 * If the tryconfig_ values are nonzero, they are the results of an 3073 * earlier tryimport. If they all match the uberblock we just found, 3074 * then the pool has not changed and we return false so we do not test 3075 * a second time. 3076 */ 3077 if (tryconfig_txg && tryconfig_txg == ub->ub_txg && 3078 tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp && 3079 tryconfig_mmp_seq && tryconfig_mmp_seq == 3080 (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) 3081 return (B_FALSE); 3082 3083 /* 3084 * Allow the activity check to be skipped when importing the pool 3085 * on the same host which last imported it. Since the hostid from 3086 * configuration may be stale use the one read from the label. 3087 */ 3088 if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID)) 3089 hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID); 3090 3091 if (hostid == spa_get_hostid(spa)) 3092 return (B_FALSE); 3093 3094 /* 3095 * Skip the activity test when the pool was cleanly exported. 3096 */ 3097 if (state != POOL_STATE_ACTIVE) 3098 return (B_FALSE); 3099 3100 return (B_TRUE); 3101 } 3102 3103 /* 3104 * Nanoseconds the activity check must watch for changes on-disk. 3105 */ 3106 static uint64_t 3107 spa_activity_check_duration(spa_t *spa, uberblock_t *ub) 3108 { 3109 uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1); 3110 uint64_t multihost_interval = MSEC2NSEC( 3111 MMP_INTERVAL_OK(zfs_multihost_interval)); 3112 uint64_t import_delay = MAX(NANOSEC, import_intervals * 3113 multihost_interval); 3114 3115 /* 3116 * Local tunables determine a minimum duration except for the case 3117 * where we know when the remote host will suspend the pool if MMP 3118 * writes do not land. 3119 * 3120 * See Big Theory comment at the top of mmp.c for the reasoning behind 3121 * these cases and times. 3122 */ 3123 3124 ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100); 3125 3126 if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) && 3127 MMP_FAIL_INT(ub) > 0) { 3128 3129 /* MMP on remote host will suspend pool after failed writes */ 3130 import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) * 3131 MMP_IMPORT_SAFETY_FACTOR / 100; 3132 3133 zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp " 3134 "mmp_fails=%llu ub_mmp mmp_interval=%llu " 3135 "import_intervals=%llu", (u_longlong_t)import_delay, 3136 (u_longlong_t)MMP_FAIL_INT(ub), 3137 (u_longlong_t)MMP_INTERVAL(ub), 3138 (u_longlong_t)import_intervals); 3139 3140 } else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) && 3141 MMP_FAIL_INT(ub) == 0) { 3142 3143 /* MMP on remote host will never suspend pool */ 3144 import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) + 3145 ub->ub_mmp_delay) * import_intervals); 3146 3147 zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp " 3148 "mmp_interval=%llu ub_mmp_delay=%llu " 3149 "import_intervals=%llu", (u_longlong_t)import_delay, 3150 (u_longlong_t)MMP_INTERVAL(ub), 3151 (u_longlong_t)ub->ub_mmp_delay, 3152 (u_longlong_t)import_intervals); 3153 3154 } else if (MMP_VALID(ub)) { 3155 /* 3156 * zfs-0.7 compatibility case 3157 */ 3158 3159 import_delay = MAX(import_delay, (multihost_interval + 3160 ub->ub_mmp_delay) * import_intervals); 3161 3162 zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu " 3163 "import_intervals=%llu leaves=%u", 3164 (u_longlong_t)import_delay, 3165 (u_longlong_t)ub->ub_mmp_delay, 3166 (u_longlong_t)import_intervals, 3167 vdev_count_leaves(spa)); 3168 } else { 3169 /* Using local tunings is the only reasonable option */ 3170 zfs_dbgmsg("pool last imported on non-MMP aware " 3171 "host using import_delay=%llu multihost_interval=%llu " 3172 "import_intervals=%llu", (u_longlong_t)import_delay, 3173 (u_longlong_t)multihost_interval, 3174 (u_longlong_t)import_intervals); 3175 } 3176 3177 return (import_delay); 3178 } 3179 3180 /* 3181 * Perform the import activity check. If the user canceled the import or 3182 * we detected activity then fail. 3183 */ 3184 static int 3185 spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config) 3186 { 3187 uint64_t txg = ub->ub_txg; 3188 uint64_t timestamp = ub->ub_timestamp; 3189 uint64_t mmp_config = ub->ub_mmp_config; 3190 uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0; 3191 uint64_t import_delay; 3192 hrtime_t import_expire; 3193 nvlist_t *mmp_label = NULL; 3194 vdev_t *rvd = spa->spa_root_vdev; 3195 kcondvar_t cv; 3196 kmutex_t mtx; 3197 int error = 0; 3198 3199 cv_init(&cv, NULL, CV_DEFAULT, NULL); 3200 mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL); 3201 mutex_enter(&mtx); 3202 3203 /* 3204 * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed 3205 * during the earlier tryimport. If the txg recorded there is 0 then 3206 * the pool is known to be active on another host. 3207 * 3208 * Otherwise, the pool might be in use on another host. Check for 3209 * changes in the uberblocks on disk if necessary. 3210 */ 3211 if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) { 3212 nvlist_t *nvinfo = fnvlist_lookup_nvlist(config, 3213 ZPOOL_CONFIG_LOAD_INFO); 3214 3215 if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) && 3216 fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) { 3217 vdev_uberblock_load(rvd, ub, &mmp_label); 3218 error = SET_ERROR(EREMOTEIO); 3219 goto out; 3220 } 3221 } 3222 3223 import_delay = spa_activity_check_duration(spa, ub); 3224 3225 /* Add a small random factor in case of simultaneous imports (0-25%) */ 3226 import_delay += import_delay * random_in_range(250) / 1000; 3227 3228 import_expire = gethrtime() + import_delay; 3229 3230 while (gethrtime() < import_expire) { 3231 (void) spa_import_progress_set_mmp_check(spa_guid(spa), 3232 NSEC2SEC(import_expire - gethrtime())); 3233 3234 vdev_uberblock_load(rvd, ub, &mmp_label); 3235 3236 if (txg != ub->ub_txg || timestamp != ub->ub_timestamp || 3237 mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) { 3238 zfs_dbgmsg("multihost activity detected " 3239 "txg %llu ub_txg %llu " 3240 "timestamp %llu ub_timestamp %llu " 3241 "mmp_config %#llx ub_mmp_config %#llx", 3242 (u_longlong_t)txg, (u_longlong_t)ub->ub_txg, 3243 (u_longlong_t)timestamp, 3244 (u_longlong_t)ub->ub_timestamp, 3245 (u_longlong_t)mmp_config, 3246 (u_longlong_t)ub->ub_mmp_config); 3247 3248 error = SET_ERROR(EREMOTEIO); 3249 break; 3250 } 3251 3252 if (mmp_label) { 3253 nvlist_free(mmp_label); 3254 mmp_label = NULL; 3255 } 3256 3257 error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz); 3258 if (error != -1) { 3259 error = SET_ERROR(EINTR); 3260 break; 3261 } 3262 error = 0; 3263 } 3264 3265 out: 3266 mutex_exit(&mtx); 3267 mutex_destroy(&mtx); 3268 cv_destroy(&cv); 3269 3270 /* 3271 * If the pool is determined to be active store the status in the 3272 * spa->spa_load_info nvlist. If the remote hostname or hostid are 3273 * available from configuration read from disk store them as well. 3274 * This allows 'zpool import' to generate a more useful message. 3275 * 3276 * ZPOOL_CONFIG_MMP_STATE - observed pool status (mandatory) 3277 * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool 3278 * ZPOOL_CONFIG_MMP_HOSTID - hostid from the active pool 3279 */ 3280 if (error == EREMOTEIO) { 3281 char *hostname = "<unknown>"; 3282 uint64_t hostid = 0; 3283 3284 if (mmp_label) { 3285 if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) { 3286 hostname = fnvlist_lookup_string(mmp_label, 3287 ZPOOL_CONFIG_HOSTNAME); 3288 fnvlist_add_string(spa->spa_load_info, 3289 ZPOOL_CONFIG_MMP_HOSTNAME, hostname); 3290 } 3291 3292 if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) { 3293 hostid = fnvlist_lookup_uint64(mmp_label, 3294 ZPOOL_CONFIG_HOSTID); 3295 fnvlist_add_uint64(spa->spa_load_info, 3296 ZPOOL_CONFIG_MMP_HOSTID, hostid); 3297 } 3298 } 3299 3300 fnvlist_add_uint64(spa->spa_load_info, 3301 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE); 3302 fnvlist_add_uint64(spa->spa_load_info, 3303 ZPOOL_CONFIG_MMP_TXG, 0); 3304 3305 error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO); 3306 } 3307 3308 if (mmp_label) 3309 nvlist_free(mmp_label); 3310 3311 return (error); 3312 } 3313 3314 static int 3315 spa_verify_host(spa_t *spa, nvlist_t *mos_config) 3316 { 3317 uint64_t hostid; 3318 char *hostname; 3319 uint64_t myhostid = 0; 3320 3321 if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config, 3322 ZPOOL_CONFIG_HOSTID, &hostid) == 0) { 3323 hostname = fnvlist_lookup_string(mos_config, 3324 ZPOOL_CONFIG_HOSTNAME); 3325 3326 myhostid = zone_get_hostid(NULL); 3327 3328 if (hostid != 0 && myhostid != 0 && hostid != myhostid) { 3329 cmn_err(CE_WARN, "pool '%s' could not be " 3330 "loaded as it was last accessed by " 3331 "another system (host: %s hostid: 0x%llx). " 3332 "See: https://openzfs.github.io/openzfs-docs/msg/" 3333 "ZFS-8000-EY", 3334 spa_name(spa), hostname, (u_longlong_t)hostid); 3335 spa_load_failed(spa, "hostid verification failed: pool " 3336 "last accessed by host: %s (hostid: 0x%llx)", 3337 hostname, (u_longlong_t)hostid); 3338 return (SET_ERROR(EBADF)); 3339 } 3340 } 3341 3342 return (0); 3343 } 3344 3345 static int 3346 spa_ld_parse_config(spa_t *spa, spa_import_type_t type) 3347 { 3348 int error = 0; 3349 nvlist_t *nvtree, *nvl, *config = spa->spa_config; 3350 int parse; 3351 vdev_t *rvd; 3352 uint64_t pool_guid; 3353 char *comment; 3354 char *compatibility; 3355 3356 /* 3357 * Versioning wasn't explicitly added to the label until later, so if 3358 * it's not present treat it as the initial version. 3359 */ 3360 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION, 3361 &spa->spa_ubsync.ub_version) != 0) 3362 spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL; 3363 3364 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) { 3365 spa_load_failed(spa, "invalid config provided: '%s' missing", 3366 ZPOOL_CONFIG_POOL_GUID); 3367 return (SET_ERROR(EINVAL)); 3368 } 3369 3370 /* 3371 * If we are doing an import, ensure that the pool is not already 3372 * imported by checking if its pool guid already exists in the 3373 * spa namespace. 3374 * 3375 * The only case that we allow an already imported pool to be 3376 * imported again, is when the pool is checkpointed and we want to 3377 * look at its checkpointed state from userland tools like zdb. 3378 */ 3379 #ifdef _KERNEL 3380 if ((spa->spa_load_state == SPA_LOAD_IMPORT || 3381 spa->spa_load_state == SPA_LOAD_TRYIMPORT) && 3382 spa_guid_exists(pool_guid, 0)) { 3383 #else 3384 if ((spa->spa_load_state == SPA_LOAD_IMPORT || 3385 spa->spa_load_state == SPA_LOAD_TRYIMPORT) && 3386 spa_guid_exists(pool_guid, 0) && 3387 !spa_importing_readonly_checkpoint(spa)) { 3388 #endif 3389 spa_load_failed(spa, "a pool with guid %llu is already open", 3390 (u_longlong_t)pool_guid); 3391 return (SET_ERROR(EEXIST)); 3392 } 3393 3394 spa->spa_config_guid = pool_guid; 3395 3396 nvlist_free(spa->spa_load_info); 3397 spa->spa_load_info = fnvlist_alloc(); 3398 3399 ASSERT(spa->spa_comment == NULL); 3400 if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0) 3401 spa->spa_comment = spa_strdup(comment); 3402 3403 ASSERT(spa->spa_compatibility == NULL); 3404 if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMPATIBILITY, 3405 &compatibility) == 0) 3406 spa->spa_compatibility = spa_strdup(compatibility); 3407 3408 (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, 3409 &spa->spa_config_txg); 3410 3411 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0) 3412 spa->spa_config_splitting = fnvlist_dup(nvl); 3413 3414 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) { 3415 spa_load_failed(spa, "invalid config provided: '%s' missing", 3416 ZPOOL_CONFIG_VDEV_TREE); 3417 return (SET_ERROR(EINVAL)); 3418 } 3419 3420 /* 3421 * Create "The Godfather" zio to hold all async IOs 3422 */ 3423 spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *), 3424 KM_SLEEP); 3425 for (int i = 0; i < max_ncpus; i++) { 3426 spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL, 3427 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 3428 ZIO_FLAG_GODFATHER); 3429 } 3430 3431 /* 3432 * Parse the configuration into a vdev tree. We explicitly set the 3433 * value that will be returned by spa_version() since parsing the 3434 * configuration requires knowing the version number. 3435 */ 3436 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3437 parse = (type == SPA_IMPORT_EXISTING ? 3438 VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT); 3439 error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse); 3440 spa_config_exit(spa, SCL_ALL, FTAG); 3441 3442 if (error != 0) { 3443 spa_load_failed(spa, "unable to parse config [error=%d]", 3444 error); 3445 return (error); 3446 } 3447 3448 ASSERT(spa->spa_root_vdev == rvd); 3449 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 3450 ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT); 3451 3452 if (type != SPA_IMPORT_ASSEMBLE) { 3453 ASSERT(spa_guid(spa) == pool_guid); 3454 } 3455 3456 return (0); 3457 } 3458 3459 /* 3460 * Recursively open all vdevs in the vdev tree. This function is called twice: 3461 * first with the untrusted config, then with the trusted config. 3462 */ 3463 static int 3464 spa_ld_open_vdevs(spa_t *spa) 3465 { 3466 int error = 0; 3467 3468 /* 3469 * spa_missing_tvds_allowed defines how many top-level vdevs can be 3470 * missing/unopenable for the root vdev to be still considered openable. 3471 */ 3472 if (spa->spa_trust_config) { 3473 spa->spa_missing_tvds_allowed = zfs_max_missing_tvds; 3474 } else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) { 3475 spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile; 3476 } else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) { 3477 spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan; 3478 } else { 3479 spa->spa_missing_tvds_allowed = 0; 3480 } 3481 3482 spa->spa_missing_tvds_allowed = 3483 MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed); 3484 3485 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3486 error = vdev_open(spa->spa_root_vdev); 3487 spa_config_exit(spa, SCL_ALL, FTAG); 3488 3489 if (spa->spa_missing_tvds != 0) { 3490 spa_load_note(spa, "vdev tree has %lld missing top-level " 3491 "vdevs.", (u_longlong_t)spa->spa_missing_tvds); 3492 if (spa->spa_trust_config && (spa->spa_mode & SPA_MODE_WRITE)) { 3493 /* 3494 * Although theoretically we could allow users to open 3495 * incomplete pools in RW mode, we'd need to add a lot 3496 * of extra logic (e.g. adjust pool space to account 3497 * for missing vdevs). 3498 * This limitation also prevents users from accidentally 3499 * opening the pool in RW mode during data recovery and 3500 * damaging it further. 3501 */ 3502 spa_load_note(spa, "pools with missing top-level " 3503 "vdevs can only be opened in read-only mode."); 3504 error = SET_ERROR(ENXIO); 3505 } else { 3506 spa_load_note(spa, "current settings allow for maximum " 3507 "%lld missing top-level vdevs at this stage.", 3508 (u_longlong_t)spa->spa_missing_tvds_allowed); 3509 } 3510 } 3511 if (error != 0) { 3512 spa_load_failed(spa, "unable to open vdev tree [error=%d]", 3513 error); 3514 } 3515 if (spa->spa_missing_tvds != 0 || error != 0) 3516 vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2); 3517 3518 return (error); 3519 } 3520 3521 /* 3522 * We need to validate the vdev labels against the configuration that 3523 * we have in hand. This function is called twice: first with an untrusted 3524 * config, then with a trusted config. The validation is more strict when the 3525 * config is trusted. 3526 */ 3527 static int 3528 spa_ld_validate_vdevs(spa_t *spa) 3529 { 3530 int error = 0; 3531 vdev_t *rvd = spa->spa_root_vdev; 3532 3533 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3534 error = vdev_validate(rvd); 3535 spa_config_exit(spa, SCL_ALL, FTAG); 3536 3537 if (error != 0) { 3538 spa_load_failed(spa, "vdev_validate failed [error=%d]", error); 3539 return (error); 3540 } 3541 3542 if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) { 3543 spa_load_failed(spa, "cannot open vdev tree after invalidating " 3544 "some vdevs"); 3545 vdev_dbgmsg_print_tree(rvd, 2); 3546 return (SET_ERROR(ENXIO)); 3547 } 3548 3549 return (0); 3550 } 3551 3552 static void 3553 spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub) 3554 { 3555 spa->spa_state = POOL_STATE_ACTIVE; 3556 spa->spa_ubsync = spa->spa_uberblock; 3557 spa->spa_verify_min_txg = spa->spa_extreme_rewind ? 3558 TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1; 3559 spa->spa_first_txg = spa->spa_last_ubsync_txg ? 3560 spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1; 3561 spa->spa_claim_max_txg = spa->spa_first_txg; 3562 spa->spa_prev_software_version = ub->ub_software_version; 3563 } 3564 3565 static int 3566 spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type) 3567 { 3568 vdev_t *rvd = spa->spa_root_vdev; 3569 nvlist_t *label; 3570 uberblock_t *ub = &spa->spa_uberblock; 3571 boolean_t activity_check = B_FALSE; 3572 3573 /* 3574 * If we are opening the checkpointed state of the pool by 3575 * rewinding to it, at this point we will have written the 3576 * checkpointed uberblock to the vdev labels, so searching 3577 * the labels will find the right uberblock. However, if 3578 * we are opening the checkpointed state read-only, we have 3579 * not modified the labels. Therefore, we must ignore the 3580 * labels and continue using the spa_uberblock that was set 3581 * by spa_ld_checkpoint_rewind. 3582 * 3583 * Note that it would be fine to ignore the labels when 3584 * rewinding (opening writeable) as well. However, if we 3585 * crash just after writing the labels, we will end up 3586 * searching the labels. Doing so in the common case means 3587 * that this code path gets exercised normally, rather than 3588 * just in the edge case. 3589 */ 3590 if (ub->ub_checkpoint_txg != 0 && 3591 spa_importing_readonly_checkpoint(spa)) { 3592 spa_ld_select_uberblock_done(spa, ub); 3593 return (0); 3594 } 3595 3596 /* 3597 * Find the best uberblock. 3598 */ 3599 vdev_uberblock_load(rvd, ub, &label); 3600 3601 /* 3602 * If we weren't able to find a single valid uberblock, return failure. 3603 */ 3604 if (ub->ub_txg == 0) { 3605 nvlist_free(label); 3606 spa_load_failed(spa, "no valid uberblock found"); 3607 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO)); 3608 } 3609 3610 if (spa->spa_load_max_txg != UINT64_MAX) { 3611 (void) spa_import_progress_set_max_txg(spa_guid(spa), 3612 (u_longlong_t)spa->spa_load_max_txg); 3613 } 3614 spa_load_note(spa, "using uberblock with txg=%llu", 3615 (u_longlong_t)ub->ub_txg); 3616 3617 3618 /* 3619 * For pools which have the multihost property on determine if the 3620 * pool is truly inactive and can be safely imported. Prevent 3621 * hosts which don't have a hostid set from importing the pool. 3622 */ 3623 activity_check = spa_activity_check_required(spa, ub, label, 3624 spa->spa_config); 3625 if (activity_check) { 3626 if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay && 3627 spa_get_hostid(spa) == 0) { 3628 nvlist_free(label); 3629 fnvlist_add_uint64(spa->spa_load_info, 3630 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID); 3631 return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO)); 3632 } 3633 3634 int error = spa_activity_check(spa, ub, spa->spa_config); 3635 if (error) { 3636 nvlist_free(label); 3637 return (error); 3638 } 3639 3640 fnvlist_add_uint64(spa->spa_load_info, 3641 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE); 3642 fnvlist_add_uint64(spa->spa_load_info, 3643 ZPOOL_CONFIG_MMP_TXG, ub->ub_txg); 3644 fnvlist_add_uint16(spa->spa_load_info, 3645 ZPOOL_CONFIG_MMP_SEQ, 3646 (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)); 3647 } 3648 3649 /* 3650 * If the pool has an unsupported version we can't open it. 3651 */ 3652 if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) { 3653 nvlist_free(label); 3654 spa_load_failed(spa, "version %llu is not supported", 3655 (u_longlong_t)ub->ub_version); 3656 return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP)); 3657 } 3658 3659 if (ub->ub_version >= SPA_VERSION_FEATURES) { 3660 nvlist_t *features; 3661 3662 /* 3663 * If we weren't able to find what's necessary for reading the 3664 * MOS in the label, return failure. 3665 */ 3666 if (label == NULL) { 3667 spa_load_failed(spa, "label config unavailable"); 3668 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, 3669 ENXIO)); 3670 } 3671 3672 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ, 3673 &features) != 0) { 3674 nvlist_free(label); 3675 spa_load_failed(spa, "invalid label: '%s' missing", 3676 ZPOOL_CONFIG_FEATURES_FOR_READ); 3677 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, 3678 ENXIO)); 3679 } 3680 3681 /* 3682 * Update our in-core representation with the definitive values 3683 * from the label. 3684 */ 3685 nvlist_free(spa->spa_label_features); 3686 spa->spa_label_features = fnvlist_dup(features); 3687 } 3688 3689 nvlist_free(label); 3690 3691 /* 3692 * Look through entries in the label nvlist's features_for_read. If 3693 * there is a feature listed there which we don't understand then we 3694 * cannot open a pool. 3695 */ 3696 if (ub->ub_version >= SPA_VERSION_FEATURES) { 3697 nvlist_t *unsup_feat; 3698 3699 unsup_feat = fnvlist_alloc(); 3700 3701 for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features, 3702 NULL); nvp != NULL; 3703 nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) { 3704 if (!zfeature_is_supported(nvpair_name(nvp))) { 3705 fnvlist_add_string(unsup_feat, 3706 nvpair_name(nvp), ""); 3707 } 3708 } 3709 3710 if (!nvlist_empty(unsup_feat)) { 3711 fnvlist_add_nvlist(spa->spa_load_info, 3712 ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat); 3713 nvlist_free(unsup_feat); 3714 spa_load_failed(spa, "some features are unsupported"); 3715 return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT, 3716 ENOTSUP)); 3717 } 3718 3719 nvlist_free(unsup_feat); 3720 } 3721 3722 if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) { 3723 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3724 spa_try_repair(spa, spa->spa_config); 3725 spa_config_exit(spa, SCL_ALL, FTAG); 3726 nvlist_free(spa->spa_config_splitting); 3727 spa->spa_config_splitting = NULL; 3728 } 3729 3730 /* 3731 * Initialize internal SPA structures. 3732 */ 3733 spa_ld_select_uberblock_done(spa, ub); 3734 3735 return (0); 3736 } 3737 3738 static int 3739 spa_ld_open_rootbp(spa_t *spa) 3740 { 3741 int error = 0; 3742 vdev_t *rvd = spa->spa_root_vdev; 3743 3744 error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool); 3745 if (error != 0) { 3746 spa_load_failed(spa, "unable to open rootbp in dsl_pool_init " 3747 "[error=%d]", error); 3748 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3749 } 3750 spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset; 3751 3752 return (0); 3753 } 3754 3755 static int 3756 spa_ld_trusted_config(spa_t *spa, spa_import_type_t type, 3757 boolean_t reloading) 3758 { 3759 vdev_t *mrvd, *rvd = spa->spa_root_vdev; 3760 nvlist_t *nv, *mos_config, *policy; 3761 int error = 0, copy_error; 3762 uint64_t healthy_tvds, healthy_tvds_mos; 3763 uint64_t mos_config_txg; 3764 3765 if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE) 3766 != 0) 3767 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3768 3769 /* 3770 * If we're assembling a pool from a split, the config provided is 3771 * already trusted so there is nothing to do. 3772 */ 3773 if (type == SPA_IMPORT_ASSEMBLE) 3774 return (0); 3775 3776 healthy_tvds = spa_healthy_core_tvds(spa); 3777 3778 if (load_nvlist(spa, spa->spa_config_object, &mos_config) 3779 != 0) { 3780 spa_load_failed(spa, "unable to retrieve MOS config"); 3781 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3782 } 3783 3784 /* 3785 * If we are doing an open, pool owner wasn't verified yet, thus do 3786 * the verification here. 3787 */ 3788 if (spa->spa_load_state == SPA_LOAD_OPEN) { 3789 error = spa_verify_host(spa, mos_config); 3790 if (error != 0) { 3791 nvlist_free(mos_config); 3792 return (error); 3793 } 3794 } 3795 3796 nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE); 3797 3798 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3799 3800 /* 3801 * Build a new vdev tree from the trusted config 3802 */ 3803 error = spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD); 3804 if (error != 0) { 3805 nvlist_free(mos_config); 3806 spa_config_exit(spa, SCL_ALL, FTAG); 3807 spa_load_failed(spa, "spa_config_parse failed [error=%d]", 3808 error); 3809 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 3810 } 3811 3812 /* 3813 * Vdev paths in the MOS may be obsolete. If the untrusted config was 3814 * obtained by scanning /dev/dsk, then it will have the right vdev 3815 * paths. We update the trusted MOS config with this information. 3816 * We first try to copy the paths with vdev_copy_path_strict, which 3817 * succeeds only when both configs have exactly the same vdev tree. 3818 * If that fails, we fall back to a more flexible method that has a 3819 * best effort policy. 3820 */ 3821 copy_error = vdev_copy_path_strict(rvd, mrvd); 3822 if (copy_error != 0 || spa_load_print_vdev_tree) { 3823 spa_load_note(spa, "provided vdev tree:"); 3824 vdev_dbgmsg_print_tree(rvd, 2); 3825 spa_load_note(spa, "MOS vdev tree:"); 3826 vdev_dbgmsg_print_tree(mrvd, 2); 3827 } 3828 if (copy_error != 0) { 3829 spa_load_note(spa, "vdev_copy_path_strict failed, falling " 3830 "back to vdev_copy_path_relaxed"); 3831 vdev_copy_path_relaxed(rvd, mrvd); 3832 } 3833 3834 vdev_close(rvd); 3835 vdev_free(rvd); 3836 spa->spa_root_vdev = mrvd; 3837 rvd = mrvd; 3838 spa_config_exit(spa, SCL_ALL, FTAG); 3839 3840 /* 3841 * We will use spa_config if we decide to reload the spa or if spa_load 3842 * fails and we rewind. We must thus regenerate the config using the 3843 * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to 3844 * pass settings on how to load the pool and is not stored in the MOS. 3845 * We copy it over to our new, trusted config. 3846 */ 3847 mos_config_txg = fnvlist_lookup_uint64(mos_config, 3848 ZPOOL_CONFIG_POOL_TXG); 3849 nvlist_free(mos_config); 3850 mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE); 3851 if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY, 3852 &policy) == 0) 3853 fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy); 3854 spa_config_set(spa, mos_config); 3855 spa->spa_config_source = SPA_CONFIG_SRC_MOS; 3856 3857 /* 3858 * Now that we got the config from the MOS, we should be more strict 3859 * in checking blkptrs and can make assumptions about the consistency 3860 * of the vdev tree. spa_trust_config must be set to true before opening 3861 * vdevs in order for them to be writeable. 3862 */ 3863 spa->spa_trust_config = B_TRUE; 3864 3865 /* 3866 * Open and validate the new vdev tree 3867 */ 3868 error = spa_ld_open_vdevs(spa); 3869 if (error != 0) 3870 return (error); 3871 3872 error = spa_ld_validate_vdevs(spa); 3873 if (error != 0) 3874 return (error); 3875 3876 if (copy_error != 0 || spa_load_print_vdev_tree) { 3877 spa_load_note(spa, "final vdev tree:"); 3878 vdev_dbgmsg_print_tree(rvd, 2); 3879 } 3880 3881 if (spa->spa_load_state != SPA_LOAD_TRYIMPORT && 3882 !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) { 3883 /* 3884 * Sanity check to make sure that we are indeed loading the 3885 * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds 3886 * in the config provided and they happened to be the only ones 3887 * to have the latest uberblock, we could involuntarily perform 3888 * an extreme rewind. 3889 */ 3890 healthy_tvds_mos = spa_healthy_core_tvds(spa); 3891 if (healthy_tvds_mos - healthy_tvds >= 3892 SPA_SYNC_MIN_VDEVS) { 3893 spa_load_note(spa, "config provided misses too many " 3894 "top-level vdevs compared to MOS (%lld vs %lld). ", 3895 (u_longlong_t)healthy_tvds, 3896 (u_longlong_t)healthy_tvds_mos); 3897 spa_load_note(spa, "vdev tree:"); 3898 vdev_dbgmsg_print_tree(rvd, 2); 3899 if (reloading) { 3900 spa_load_failed(spa, "config was already " 3901 "provided from MOS. Aborting."); 3902 return (spa_vdev_err(rvd, 3903 VDEV_AUX_CORRUPT_DATA, EIO)); 3904 } 3905 spa_load_note(spa, "spa must be reloaded using MOS " 3906 "config"); 3907 return (SET_ERROR(EAGAIN)); 3908 } 3909 } 3910 3911 error = spa_check_for_missing_logs(spa); 3912 if (error != 0) 3913 return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO)); 3914 3915 if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) { 3916 spa_load_failed(spa, "uberblock guid sum doesn't match MOS " 3917 "guid sum (%llu != %llu)", 3918 (u_longlong_t)spa->spa_uberblock.ub_guid_sum, 3919 (u_longlong_t)rvd->vdev_guid_sum); 3920 return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, 3921 ENXIO)); 3922 } 3923 3924 return (0); 3925 } 3926 3927 static int 3928 spa_ld_open_indirect_vdev_metadata(spa_t *spa) 3929 { 3930 int error = 0; 3931 vdev_t *rvd = spa->spa_root_vdev; 3932 3933 /* 3934 * Everything that we read before spa_remove_init() must be stored 3935 * on concreted vdevs. Therefore we do this as early as possible. 3936 */ 3937 error = spa_remove_init(spa); 3938 if (error != 0) { 3939 spa_load_failed(spa, "spa_remove_init failed [error=%d]", 3940 error); 3941 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3942 } 3943 3944 /* 3945 * Retrieve information needed to condense indirect vdev mappings. 3946 */ 3947 error = spa_condense_init(spa); 3948 if (error != 0) { 3949 spa_load_failed(spa, "spa_condense_init failed [error=%d]", 3950 error); 3951 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 3952 } 3953 3954 return (0); 3955 } 3956 3957 static int 3958 spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep) 3959 { 3960 int error = 0; 3961 vdev_t *rvd = spa->spa_root_vdev; 3962 3963 if (spa_version(spa) >= SPA_VERSION_FEATURES) { 3964 boolean_t missing_feat_read = B_FALSE; 3965 nvlist_t *unsup_feat, *enabled_feat; 3966 3967 if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ, 3968 &spa->spa_feat_for_read_obj, B_TRUE) != 0) { 3969 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3970 } 3971 3972 if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE, 3973 &spa->spa_feat_for_write_obj, B_TRUE) != 0) { 3974 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3975 } 3976 3977 if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS, 3978 &spa->spa_feat_desc_obj, B_TRUE) != 0) { 3979 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3980 } 3981 3982 enabled_feat = fnvlist_alloc(); 3983 unsup_feat = fnvlist_alloc(); 3984 3985 if (!spa_features_check(spa, B_FALSE, 3986 unsup_feat, enabled_feat)) 3987 missing_feat_read = B_TRUE; 3988 3989 if (spa_writeable(spa) || 3990 spa->spa_load_state == SPA_LOAD_TRYIMPORT) { 3991 if (!spa_features_check(spa, B_TRUE, 3992 unsup_feat, enabled_feat)) { 3993 *missing_feat_writep = B_TRUE; 3994 } 3995 } 3996 3997 fnvlist_add_nvlist(spa->spa_load_info, 3998 ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat); 3999 4000 if (!nvlist_empty(unsup_feat)) { 4001 fnvlist_add_nvlist(spa->spa_load_info, 4002 ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat); 4003 } 4004 4005 fnvlist_free(enabled_feat); 4006 fnvlist_free(unsup_feat); 4007 4008 if (!missing_feat_read) { 4009 fnvlist_add_boolean(spa->spa_load_info, 4010 ZPOOL_CONFIG_CAN_RDONLY); 4011 } 4012 4013 /* 4014 * If the state is SPA_LOAD_TRYIMPORT, our objective is 4015 * twofold: to determine whether the pool is available for 4016 * import in read-write mode and (if it is not) whether the 4017 * pool is available for import in read-only mode. If the pool 4018 * is available for import in read-write mode, it is displayed 4019 * as available in userland; if it is not available for import 4020 * in read-only mode, it is displayed as unavailable in 4021 * userland. If the pool is available for import in read-only 4022 * mode but not read-write mode, it is displayed as unavailable 4023 * in userland with a special note that the pool is actually 4024 * available for open in read-only mode. 4025 * 4026 * As a result, if the state is SPA_LOAD_TRYIMPORT and we are 4027 * missing a feature for write, we must first determine whether 4028 * the pool can be opened read-only before returning to 4029 * userland in order to know whether to display the 4030 * abovementioned note. 4031 */ 4032 if (missing_feat_read || (*missing_feat_writep && 4033 spa_writeable(spa))) { 4034 spa_load_failed(spa, "pool uses unsupported features"); 4035 return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT, 4036 ENOTSUP)); 4037 } 4038 4039 /* 4040 * Load refcounts for ZFS features from disk into an in-memory 4041 * cache during SPA initialization. 4042 */ 4043 for (spa_feature_t i = 0; i < SPA_FEATURES; i++) { 4044 uint64_t refcount; 4045 4046 error = feature_get_refcount_from_disk(spa, 4047 &spa_feature_table[i], &refcount); 4048 if (error == 0) { 4049 spa->spa_feat_refcount_cache[i] = refcount; 4050 } else if (error == ENOTSUP) { 4051 spa->spa_feat_refcount_cache[i] = 4052 SPA_FEATURE_DISABLED; 4053 } else { 4054 spa_load_failed(spa, "error getting refcount " 4055 "for feature %s [error=%d]", 4056 spa_feature_table[i].fi_guid, error); 4057 return (spa_vdev_err(rvd, 4058 VDEV_AUX_CORRUPT_DATA, EIO)); 4059 } 4060 } 4061 } 4062 4063 if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) { 4064 if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG, 4065 &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0) 4066 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4067 } 4068 4069 /* 4070 * Encryption was added before bookmark_v2, even though bookmark_v2 4071 * is now a dependency. If this pool has encryption enabled without 4072 * bookmark_v2, trigger an errata message. 4073 */ 4074 if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) && 4075 !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) { 4076 spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION; 4077 } 4078 4079 return (0); 4080 } 4081 4082 static int 4083 spa_ld_load_special_directories(spa_t *spa) 4084 { 4085 int error = 0; 4086 vdev_t *rvd = spa->spa_root_vdev; 4087 4088 spa->spa_is_initializing = B_TRUE; 4089 error = dsl_pool_open(spa->spa_dsl_pool); 4090 spa->spa_is_initializing = B_FALSE; 4091 if (error != 0) { 4092 spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error); 4093 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4094 } 4095 4096 return (0); 4097 } 4098 4099 static int 4100 spa_ld_get_props(spa_t *spa) 4101 { 4102 int error = 0; 4103 uint64_t obj; 4104 vdev_t *rvd = spa->spa_root_vdev; 4105 4106 /* Grab the checksum salt from the MOS. */ 4107 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 4108 DMU_POOL_CHECKSUM_SALT, 1, 4109 sizeof (spa->spa_cksum_salt.zcs_bytes), 4110 spa->spa_cksum_salt.zcs_bytes); 4111 if (error == ENOENT) { 4112 /* Generate a new salt for subsequent use */ 4113 (void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes, 4114 sizeof (spa->spa_cksum_salt.zcs_bytes)); 4115 } else if (error != 0) { 4116 spa_load_failed(spa, "unable to retrieve checksum salt from " 4117 "MOS [error=%d]", error); 4118 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4119 } 4120 4121 if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0) 4122 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4123 error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj); 4124 if (error != 0) { 4125 spa_load_failed(spa, "error opening deferred-frees bpobj " 4126 "[error=%d]", error); 4127 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4128 } 4129 4130 /* 4131 * Load the bit that tells us to use the new accounting function 4132 * (raid-z deflation). If we have an older pool, this will not 4133 * be present. 4134 */ 4135 error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE); 4136 if (error != 0 && error != ENOENT) 4137 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4138 4139 error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION, 4140 &spa->spa_creation_version, B_FALSE); 4141 if (error != 0 && error != ENOENT) 4142 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4143 4144 /* 4145 * Load the persistent error log. If we have an older pool, this will 4146 * not be present. 4147 */ 4148 error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last, 4149 B_FALSE); 4150 if (error != 0 && error != ENOENT) 4151 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4152 4153 error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB, 4154 &spa->spa_errlog_scrub, B_FALSE); 4155 if (error != 0 && error != ENOENT) 4156 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4157 4158 /* 4159 * Load the livelist deletion field. If a livelist is queued for 4160 * deletion, indicate that in the spa 4161 */ 4162 error = spa_dir_prop(spa, DMU_POOL_DELETED_CLONES, 4163 &spa->spa_livelists_to_delete, B_FALSE); 4164 if (error != 0 && error != ENOENT) 4165 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4166 4167 /* 4168 * Load the history object. If we have an older pool, this 4169 * will not be present. 4170 */ 4171 error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE); 4172 if (error != 0 && error != ENOENT) 4173 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4174 4175 /* 4176 * Load the per-vdev ZAP map. If we have an older pool, this will not 4177 * be present; in this case, defer its creation to a later time to 4178 * avoid dirtying the MOS this early / out of sync context. See 4179 * spa_sync_config_object. 4180 */ 4181 4182 /* The sentinel is only available in the MOS config. */ 4183 nvlist_t *mos_config; 4184 if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) { 4185 spa_load_failed(spa, "unable to retrieve MOS config"); 4186 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4187 } 4188 4189 error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP, 4190 &spa->spa_all_vdev_zaps, B_FALSE); 4191 4192 if (error == ENOENT) { 4193 VERIFY(!nvlist_exists(mos_config, 4194 ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)); 4195 spa->spa_avz_action = AVZ_ACTION_INITIALIZE; 4196 ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev)); 4197 } else if (error != 0) { 4198 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4199 } else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) { 4200 /* 4201 * An older version of ZFS overwrote the sentinel value, so 4202 * we have orphaned per-vdev ZAPs in the MOS. Defer their 4203 * destruction to later; see spa_sync_config_object. 4204 */ 4205 spa->spa_avz_action = AVZ_ACTION_DESTROY; 4206 /* 4207 * We're assuming that no vdevs have had their ZAPs created 4208 * before this. Better be sure of it. 4209 */ 4210 ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev)); 4211 } 4212 nvlist_free(mos_config); 4213 4214 spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION); 4215 4216 error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object, 4217 B_FALSE); 4218 if (error && error != ENOENT) 4219 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4220 4221 if (error == 0) { 4222 uint64_t autoreplace = 0; 4223 4224 spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs); 4225 spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace); 4226 spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation); 4227 spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode); 4228 spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand); 4229 spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost); 4230 spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim); 4231 spa->spa_autoreplace = (autoreplace != 0); 4232 } 4233 4234 /* 4235 * If we are importing a pool with missing top-level vdevs, 4236 * we enforce that the pool doesn't panic or get suspended on 4237 * error since the likelihood of missing data is extremely high. 4238 */ 4239 if (spa->spa_missing_tvds > 0 && 4240 spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE && 4241 spa->spa_load_state != SPA_LOAD_TRYIMPORT) { 4242 spa_load_note(spa, "forcing failmode to 'continue' " 4243 "as some top level vdevs are missing"); 4244 spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE; 4245 } 4246 4247 return (0); 4248 } 4249 4250 static int 4251 spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type) 4252 { 4253 int error = 0; 4254 vdev_t *rvd = spa->spa_root_vdev; 4255 4256 /* 4257 * If we're assembling the pool from the split-off vdevs of 4258 * an existing pool, we don't want to attach the spares & cache 4259 * devices. 4260 */ 4261 4262 /* 4263 * Load any hot spares for this pool. 4264 */ 4265 error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object, 4266 B_FALSE); 4267 if (error != 0 && error != ENOENT) 4268 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4269 if (error == 0 && type != SPA_IMPORT_ASSEMBLE) { 4270 ASSERT(spa_version(spa) >= SPA_VERSION_SPARES); 4271 if (load_nvlist(spa, spa->spa_spares.sav_object, 4272 &spa->spa_spares.sav_config) != 0) { 4273 spa_load_failed(spa, "error loading spares nvlist"); 4274 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4275 } 4276 4277 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4278 spa_load_spares(spa); 4279 spa_config_exit(spa, SCL_ALL, FTAG); 4280 } else if (error == 0) { 4281 spa->spa_spares.sav_sync = B_TRUE; 4282 } 4283 4284 /* 4285 * Load any level 2 ARC devices for this pool. 4286 */ 4287 error = spa_dir_prop(spa, DMU_POOL_L2CACHE, 4288 &spa->spa_l2cache.sav_object, B_FALSE); 4289 if (error != 0 && error != ENOENT) 4290 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4291 if (error == 0 && type != SPA_IMPORT_ASSEMBLE) { 4292 ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE); 4293 if (load_nvlist(spa, spa->spa_l2cache.sav_object, 4294 &spa->spa_l2cache.sav_config) != 0) { 4295 spa_load_failed(spa, "error loading l2cache nvlist"); 4296 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4297 } 4298 4299 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4300 spa_load_l2cache(spa); 4301 spa_config_exit(spa, SCL_ALL, FTAG); 4302 } else if (error == 0) { 4303 spa->spa_l2cache.sav_sync = B_TRUE; 4304 } 4305 4306 return (0); 4307 } 4308 4309 static int 4310 spa_ld_load_vdev_metadata(spa_t *spa) 4311 { 4312 int error = 0; 4313 vdev_t *rvd = spa->spa_root_vdev; 4314 4315 /* 4316 * If the 'multihost' property is set, then never allow a pool to 4317 * be imported when the system hostid is zero. The exception to 4318 * this rule is zdb which is always allowed to access pools. 4319 */ 4320 if (spa_multihost(spa) && spa_get_hostid(spa) == 0 && 4321 (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) { 4322 fnvlist_add_uint64(spa->spa_load_info, 4323 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID); 4324 return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO)); 4325 } 4326 4327 /* 4328 * If the 'autoreplace' property is set, then post a resource notifying 4329 * the ZFS DE that it should not issue any faults for unopenable 4330 * devices. We also iterate over the vdevs, and post a sysevent for any 4331 * unopenable vdevs so that the normal autoreplace handler can take 4332 * over. 4333 */ 4334 if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) { 4335 spa_check_removed(spa->spa_root_vdev); 4336 /* 4337 * For the import case, this is done in spa_import(), because 4338 * at this point we're using the spare definitions from 4339 * the MOS config, not necessarily from the userland config. 4340 */ 4341 if (spa->spa_load_state != SPA_LOAD_IMPORT) { 4342 spa_aux_check_removed(&spa->spa_spares); 4343 spa_aux_check_removed(&spa->spa_l2cache); 4344 } 4345 } 4346 4347 /* 4348 * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc. 4349 */ 4350 error = vdev_load(rvd); 4351 if (error != 0) { 4352 spa_load_failed(spa, "vdev_load failed [error=%d]", error); 4353 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 4354 } 4355 4356 error = spa_ld_log_spacemaps(spa); 4357 if (error != 0) { 4358 spa_load_failed(spa, "spa_ld_log_sm_data failed [error=%d]", 4359 error); 4360 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 4361 } 4362 4363 /* 4364 * Propagate the leaf DTLs we just loaded all the way up the vdev tree. 4365 */ 4366 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4367 vdev_dtl_reassess(rvd, 0, 0, B_FALSE, B_FALSE); 4368 spa_config_exit(spa, SCL_ALL, FTAG); 4369 4370 return (0); 4371 } 4372 4373 static int 4374 spa_ld_load_dedup_tables(spa_t *spa) 4375 { 4376 int error = 0; 4377 vdev_t *rvd = spa->spa_root_vdev; 4378 4379 error = ddt_load(spa); 4380 if (error != 0) { 4381 spa_load_failed(spa, "ddt_load failed [error=%d]", error); 4382 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4383 } 4384 4385 return (0); 4386 } 4387 4388 static int 4389 spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, char **ereport) 4390 { 4391 vdev_t *rvd = spa->spa_root_vdev; 4392 4393 if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) { 4394 boolean_t missing = spa_check_logs(spa); 4395 if (missing) { 4396 if (spa->spa_missing_tvds != 0) { 4397 spa_load_note(spa, "spa_check_logs failed " 4398 "so dropping the logs"); 4399 } else { 4400 *ereport = FM_EREPORT_ZFS_LOG_REPLAY; 4401 spa_load_failed(spa, "spa_check_logs failed"); 4402 return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG, 4403 ENXIO)); 4404 } 4405 } 4406 } 4407 4408 return (0); 4409 } 4410 4411 static int 4412 spa_ld_verify_pool_data(spa_t *spa) 4413 { 4414 int error = 0; 4415 vdev_t *rvd = spa->spa_root_vdev; 4416 4417 /* 4418 * We've successfully opened the pool, verify that we're ready 4419 * to start pushing transactions. 4420 */ 4421 if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) { 4422 error = spa_load_verify(spa); 4423 if (error != 0) { 4424 spa_load_failed(spa, "spa_load_verify failed " 4425 "[error=%d]", error); 4426 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, 4427 error)); 4428 } 4429 } 4430 4431 return (0); 4432 } 4433 4434 static void 4435 spa_ld_claim_log_blocks(spa_t *spa) 4436 { 4437 dmu_tx_t *tx; 4438 dsl_pool_t *dp = spa_get_dsl(spa); 4439 4440 /* 4441 * Claim log blocks that haven't been committed yet. 4442 * This must all happen in a single txg. 4443 * Note: spa_claim_max_txg is updated by spa_claim_notify(), 4444 * invoked from zil_claim_log_block()'s i/o done callback. 4445 * Price of rollback is that we abandon the log. 4446 */ 4447 spa->spa_claiming = B_TRUE; 4448 4449 tx = dmu_tx_create_assigned(dp, spa_first_txg(spa)); 4450 (void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 4451 zil_claim, tx, DS_FIND_CHILDREN); 4452 dmu_tx_commit(tx); 4453 4454 spa->spa_claiming = B_FALSE; 4455 4456 spa_set_log_state(spa, SPA_LOG_GOOD); 4457 } 4458 4459 static void 4460 spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg, 4461 boolean_t update_config_cache) 4462 { 4463 vdev_t *rvd = spa->spa_root_vdev; 4464 int need_update = B_FALSE; 4465 4466 /* 4467 * If the config cache is stale, or we have uninitialized 4468 * metaslabs (see spa_vdev_add()), then update the config. 4469 * 4470 * If this is a verbatim import, trust the current 4471 * in-core spa_config and update the disk labels. 4472 */ 4473 if (update_config_cache || config_cache_txg != spa->spa_config_txg || 4474 spa->spa_load_state == SPA_LOAD_IMPORT || 4475 spa->spa_load_state == SPA_LOAD_RECOVER || 4476 (spa->spa_import_flags & ZFS_IMPORT_VERBATIM)) 4477 need_update = B_TRUE; 4478 4479 for (int c = 0; c < rvd->vdev_children; c++) 4480 if (rvd->vdev_child[c]->vdev_ms_array == 0) 4481 need_update = B_TRUE; 4482 4483 /* 4484 * Update the config cache asynchronously in case we're the 4485 * root pool, in which case the config cache isn't writable yet. 4486 */ 4487 if (need_update) 4488 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 4489 } 4490 4491 static void 4492 spa_ld_prepare_for_reload(spa_t *spa) 4493 { 4494 spa_mode_t mode = spa->spa_mode; 4495 int async_suspended = spa->spa_async_suspended; 4496 4497 spa_unload(spa); 4498 spa_deactivate(spa); 4499 spa_activate(spa, mode); 4500 4501 /* 4502 * We save the value of spa_async_suspended as it gets reset to 0 by 4503 * spa_unload(). We want to restore it back to the original value before 4504 * returning as we might be calling spa_async_resume() later. 4505 */ 4506 spa->spa_async_suspended = async_suspended; 4507 } 4508 4509 static int 4510 spa_ld_read_checkpoint_txg(spa_t *spa) 4511 { 4512 uberblock_t checkpoint; 4513 int error = 0; 4514 4515 ASSERT0(spa->spa_checkpoint_txg); 4516 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4517 4518 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 4519 DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t), 4520 sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint); 4521 4522 if (error == ENOENT) 4523 return (0); 4524 4525 if (error != 0) 4526 return (error); 4527 4528 ASSERT3U(checkpoint.ub_txg, !=, 0); 4529 ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0); 4530 ASSERT3U(checkpoint.ub_timestamp, !=, 0); 4531 spa->spa_checkpoint_txg = checkpoint.ub_txg; 4532 spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp; 4533 4534 return (0); 4535 } 4536 4537 static int 4538 spa_ld_mos_init(spa_t *spa, spa_import_type_t type) 4539 { 4540 int error = 0; 4541 4542 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4543 ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE); 4544 4545 /* 4546 * Never trust the config that is provided unless we are assembling 4547 * a pool following a split. 4548 * This means don't trust blkptrs and the vdev tree in general. This 4549 * also effectively puts the spa in read-only mode since 4550 * spa_writeable() checks for spa_trust_config to be true. 4551 * We will later load a trusted config from the MOS. 4552 */ 4553 if (type != SPA_IMPORT_ASSEMBLE) 4554 spa->spa_trust_config = B_FALSE; 4555 4556 /* 4557 * Parse the config provided to create a vdev tree. 4558 */ 4559 error = spa_ld_parse_config(spa, type); 4560 if (error != 0) 4561 return (error); 4562 4563 spa_import_progress_add(spa); 4564 4565 /* 4566 * Now that we have the vdev tree, try to open each vdev. This involves 4567 * opening the underlying physical device, retrieving its geometry and 4568 * probing the vdev with a dummy I/O. The state of each vdev will be set 4569 * based on the success of those operations. After this we'll be ready 4570 * to read from the vdevs. 4571 */ 4572 error = spa_ld_open_vdevs(spa); 4573 if (error != 0) 4574 return (error); 4575 4576 /* 4577 * Read the label of each vdev and make sure that the GUIDs stored 4578 * there match the GUIDs in the config provided. 4579 * If we're assembling a new pool that's been split off from an 4580 * existing pool, the labels haven't yet been updated so we skip 4581 * validation for now. 4582 */ 4583 if (type != SPA_IMPORT_ASSEMBLE) { 4584 error = spa_ld_validate_vdevs(spa); 4585 if (error != 0) 4586 return (error); 4587 } 4588 4589 /* 4590 * Read all vdev labels to find the best uberblock (i.e. latest, 4591 * unless spa_load_max_txg is set) and store it in spa_uberblock. We 4592 * get the list of features required to read blkptrs in the MOS from 4593 * the vdev label with the best uberblock and verify that our version 4594 * of zfs supports them all. 4595 */ 4596 error = spa_ld_select_uberblock(spa, type); 4597 if (error != 0) 4598 return (error); 4599 4600 /* 4601 * Pass that uberblock to the dsl_pool layer which will open the root 4602 * blkptr. This blkptr points to the latest version of the MOS and will 4603 * allow us to read its contents. 4604 */ 4605 error = spa_ld_open_rootbp(spa); 4606 if (error != 0) 4607 return (error); 4608 4609 return (0); 4610 } 4611 4612 static int 4613 spa_ld_checkpoint_rewind(spa_t *spa) 4614 { 4615 uberblock_t checkpoint; 4616 int error = 0; 4617 4618 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4619 ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); 4620 4621 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 4622 DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t), 4623 sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint); 4624 4625 if (error != 0) { 4626 spa_load_failed(spa, "unable to retrieve checkpointed " 4627 "uberblock from the MOS config [error=%d]", error); 4628 4629 if (error == ENOENT) 4630 error = ZFS_ERR_NO_CHECKPOINT; 4631 4632 return (error); 4633 } 4634 4635 ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg); 4636 ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg); 4637 4638 /* 4639 * We need to update the txg and timestamp of the checkpointed 4640 * uberblock to be higher than the latest one. This ensures that 4641 * the checkpointed uberblock is selected if we were to close and 4642 * reopen the pool right after we've written it in the vdev labels. 4643 * (also see block comment in vdev_uberblock_compare) 4644 */ 4645 checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1; 4646 checkpoint.ub_timestamp = gethrestime_sec(); 4647 4648 /* 4649 * Set current uberblock to be the checkpointed uberblock. 4650 */ 4651 spa->spa_uberblock = checkpoint; 4652 4653 /* 4654 * If we are doing a normal rewind, then the pool is open for 4655 * writing and we sync the "updated" checkpointed uberblock to 4656 * disk. Once this is done, we've basically rewound the whole 4657 * pool and there is no way back. 4658 * 4659 * There are cases when we don't want to attempt and sync the 4660 * checkpointed uberblock to disk because we are opening a 4661 * pool as read-only. Specifically, verifying the checkpointed 4662 * state with zdb, and importing the checkpointed state to get 4663 * a "preview" of its content. 4664 */ 4665 if (spa_writeable(spa)) { 4666 vdev_t *rvd = spa->spa_root_vdev; 4667 4668 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4669 vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL }; 4670 int svdcount = 0; 4671 int children = rvd->vdev_children; 4672 int c0 = random_in_range(children); 4673 4674 for (int c = 0; c < children; c++) { 4675 vdev_t *vd = rvd->vdev_child[(c0 + c) % children]; 4676 4677 /* Stop when revisiting the first vdev */ 4678 if (c > 0 && svd[0] == vd) 4679 break; 4680 4681 if (vd->vdev_ms_array == 0 || vd->vdev_islog || 4682 !vdev_is_concrete(vd)) 4683 continue; 4684 4685 svd[svdcount++] = vd; 4686 if (svdcount == SPA_SYNC_MIN_VDEVS) 4687 break; 4688 } 4689 error = vdev_config_sync(svd, svdcount, spa->spa_first_txg); 4690 if (error == 0) 4691 spa->spa_last_synced_guid = rvd->vdev_guid; 4692 spa_config_exit(spa, SCL_ALL, FTAG); 4693 4694 if (error != 0) { 4695 spa_load_failed(spa, "failed to write checkpointed " 4696 "uberblock to the vdev labels [error=%d]", error); 4697 return (error); 4698 } 4699 } 4700 4701 return (0); 4702 } 4703 4704 static int 4705 spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type, 4706 boolean_t *update_config_cache) 4707 { 4708 int error; 4709 4710 /* 4711 * Parse the config for pool, open and validate vdevs, 4712 * select an uberblock, and use that uberblock to open 4713 * the MOS. 4714 */ 4715 error = spa_ld_mos_init(spa, type); 4716 if (error != 0) 4717 return (error); 4718 4719 /* 4720 * Retrieve the trusted config stored in the MOS and use it to create 4721 * a new, exact version of the vdev tree, then reopen all vdevs. 4722 */ 4723 error = spa_ld_trusted_config(spa, type, B_FALSE); 4724 if (error == EAGAIN) { 4725 if (update_config_cache != NULL) 4726 *update_config_cache = B_TRUE; 4727 4728 /* 4729 * Redo the loading process with the trusted config if it is 4730 * too different from the untrusted config. 4731 */ 4732 spa_ld_prepare_for_reload(spa); 4733 spa_load_note(spa, "RELOADING"); 4734 error = spa_ld_mos_init(spa, type); 4735 if (error != 0) 4736 return (error); 4737 4738 error = spa_ld_trusted_config(spa, type, B_TRUE); 4739 if (error != 0) 4740 return (error); 4741 4742 } else if (error != 0) { 4743 return (error); 4744 } 4745 4746 return (0); 4747 } 4748 4749 /* 4750 * Load an existing storage pool, using the config provided. This config 4751 * describes which vdevs are part of the pool and is later validated against 4752 * partial configs present in each vdev's label and an entire copy of the 4753 * config stored in the MOS. 4754 */ 4755 static int 4756 spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport) 4757 { 4758 int error = 0; 4759 boolean_t missing_feat_write = B_FALSE; 4760 boolean_t checkpoint_rewind = 4761 (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); 4762 boolean_t update_config_cache = B_FALSE; 4763 4764 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4765 ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE); 4766 4767 spa_load_note(spa, "LOADING"); 4768 4769 error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache); 4770 if (error != 0) 4771 return (error); 4772 4773 /* 4774 * If we are rewinding to the checkpoint then we need to repeat 4775 * everything we've done so far in this function but this time 4776 * selecting the checkpointed uberblock and using that to open 4777 * the MOS. 4778 */ 4779 if (checkpoint_rewind) { 4780 /* 4781 * If we are rewinding to the checkpoint update config cache 4782 * anyway. 4783 */ 4784 update_config_cache = B_TRUE; 4785 4786 /* 4787 * Extract the checkpointed uberblock from the current MOS 4788 * and use this as the pool's uberblock from now on. If the 4789 * pool is imported as writeable we also write the checkpoint 4790 * uberblock to the labels, making the rewind permanent. 4791 */ 4792 error = spa_ld_checkpoint_rewind(spa); 4793 if (error != 0) 4794 return (error); 4795 4796 /* 4797 * Redo the loading process again with the 4798 * checkpointed uberblock. 4799 */ 4800 spa_ld_prepare_for_reload(spa); 4801 spa_load_note(spa, "LOADING checkpointed uberblock"); 4802 error = spa_ld_mos_with_trusted_config(spa, type, NULL); 4803 if (error != 0) 4804 return (error); 4805 } 4806 4807 /* 4808 * Retrieve the checkpoint txg if the pool has a checkpoint. 4809 */ 4810 error = spa_ld_read_checkpoint_txg(spa); 4811 if (error != 0) 4812 return (error); 4813 4814 /* 4815 * Retrieve the mapping of indirect vdevs. Those vdevs were removed 4816 * from the pool and their contents were re-mapped to other vdevs. Note 4817 * that everything that we read before this step must have been 4818 * rewritten on concrete vdevs after the last device removal was 4819 * initiated. Otherwise we could be reading from indirect vdevs before 4820 * we have loaded their mappings. 4821 */ 4822 error = spa_ld_open_indirect_vdev_metadata(spa); 4823 if (error != 0) 4824 return (error); 4825 4826 /* 4827 * Retrieve the full list of active features from the MOS and check if 4828 * they are all supported. 4829 */ 4830 error = spa_ld_check_features(spa, &missing_feat_write); 4831 if (error != 0) 4832 return (error); 4833 4834 /* 4835 * Load several special directories from the MOS needed by the dsl_pool 4836 * layer. 4837 */ 4838 error = spa_ld_load_special_directories(spa); 4839 if (error != 0) 4840 return (error); 4841 4842 /* 4843 * Retrieve pool properties from the MOS. 4844 */ 4845 error = spa_ld_get_props(spa); 4846 if (error != 0) 4847 return (error); 4848 4849 /* 4850 * Retrieve the list of auxiliary devices - cache devices and spares - 4851 * and open them. 4852 */ 4853 error = spa_ld_open_aux_vdevs(spa, type); 4854 if (error != 0) 4855 return (error); 4856 4857 /* 4858 * Load the metadata for all vdevs. Also check if unopenable devices 4859 * should be autoreplaced. 4860 */ 4861 error = spa_ld_load_vdev_metadata(spa); 4862 if (error != 0) 4863 return (error); 4864 4865 error = spa_ld_load_dedup_tables(spa); 4866 if (error != 0) 4867 return (error); 4868 4869 /* 4870 * Verify the logs now to make sure we don't have any unexpected errors 4871 * when we claim log blocks later. 4872 */ 4873 error = spa_ld_verify_logs(spa, type, ereport); 4874 if (error != 0) 4875 return (error); 4876 4877 if (missing_feat_write) { 4878 ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT); 4879 4880 /* 4881 * At this point, we know that we can open the pool in 4882 * read-only mode but not read-write mode. We now have enough 4883 * information and can return to userland. 4884 */ 4885 return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT, 4886 ENOTSUP)); 4887 } 4888 4889 /* 4890 * Traverse the last txgs to make sure the pool was left off in a safe 4891 * state. When performing an extreme rewind, we verify the whole pool, 4892 * which can take a very long time. 4893 */ 4894 error = spa_ld_verify_pool_data(spa); 4895 if (error != 0) 4896 return (error); 4897 4898 /* 4899 * Calculate the deflated space for the pool. This must be done before 4900 * we write anything to the pool because we'd need to update the space 4901 * accounting using the deflated sizes. 4902 */ 4903 spa_update_dspace(spa); 4904 4905 /* 4906 * We have now retrieved all the information we needed to open the 4907 * pool. If we are importing the pool in read-write mode, a few 4908 * additional steps must be performed to finish the import. 4909 */ 4910 if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER || 4911 spa->spa_load_max_txg == UINT64_MAX)) { 4912 uint64_t config_cache_txg = spa->spa_config_txg; 4913 4914 ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT); 4915 4916 /* 4917 * In case of a checkpoint rewind, log the original txg 4918 * of the checkpointed uberblock. 4919 */ 4920 if (checkpoint_rewind) { 4921 spa_history_log_internal(spa, "checkpoint rewind", 4922 NULL, "rewound state to txg=%llu", 4923 (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg); 4924 } 4925 4926 /* 4927 * Traverse the ZIL and claim all blocks. 4928 */ 4929 spa_ld_claim_log_blocks(spa); 4930 4931 /* 4932 * Kick-off the syncing thread. 4933 */ 4934 spa->spa_sync_on = B_TRUE; 4935 txg_sync_start(spa->spa_dsl_pool); 4936 mmp_thread_start(spa); 4937 4938 /* 4939 * Wait for all claims to sync. We sync up to the highest 4940 * claimed log block birth time so that claimed log blocks 4941 * don't appear to be from the future. spa_claim_max_txg 4942 * will have been set for us by ZIL traversal operations 4943 * performed above. 4944 */ 4945 txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg); 4946 4947 /* 4948 * Check if we need to request an update of the config. On the 4949 * next sync, we would update the config stored in vdev labels 4950 * and the cachefile (by default /etc/zfs/zpool.cache). 4951 */ 4952 spa_ld_check_for_config_update(spa, config_cache_txg, 4953 update_config_cache); 4954 4955 /* 4956 * Check if a rebuild was in progress and if so resume it. 4957 * Then check all DTLs to see if anything needs resilvering. 4958 * The resilver will be deferred if a rebuild was started. 4959 */ 4960 if (vdev_rebuild_active(spa->spa_root_vdev)) { 4961 vdev_rebuild_restart(spa); 4962 } else if (!dsl_scan_resilvering(spa->spa_dsl_pool) && 4963 vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) { 4964 spa_async_request(spa, SPA_ASYNC_RESILVER); 4965 } 4966 4967 /* 4968 * Log the fact that we booted up (so that we can detect if 4969 * we rebooted in the middle of an operation). 4970 */ 4971 spa_history_log_version(spa, "open", NULL); 4972 4973 spa_restart_removal(spa); 4974 spa_spawn_aux_threads(spa); 4975 4976 /* 4977 * Delete any inconsistent datasets. 4978 * 4979 * Note: 4980 * Since we may be issuing deletes for clones here, 4981 * we make sure to do so after we've spawned all the 4982 * auxiliary threads above (from which the livelist 4983 * deletion zthr is part of). 4984 */ 4985 (void) dmu_objset_find(spa_name(spa), 4986 dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN); 4987 4988 /* 4989 * Clean up any stale temporary dataset userrefs. 4990 */ 4991 dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool); 4992 4993 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 4994 vdev_initialize_restart(spa->spa_root_vdev); 4995 vdev_trim_restart(spa->spa_root_vdev); 4996 vdev_autotrim_restart(spa); 4997 spa_config_exit(spa, SCL_CONFIG, FTAG); 4998 } 4999 5000 spa_import_progress_remove(spa_guid(spa)); 5001 spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD); 5002 5003 spa_load_note(spa, "LOADED"); 5004 5005 return (0); 5006 } 5007 5008 static int 5009 spa_load_retry(spa_t *spa, spa_load_state_t state) 5010 { 5011 spa_mode_t mode = spa->spa_mode; 5012 5013 spa_unload(spa); 5014 spa_deactivate(spa); 5015 5016 spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1; 5017 5018 spa_activate(spa, mode); 5019 spa_async_suspend(spa); 5020 5021 spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu", 5022 (u_longlong_t)spa->spa_load_max_txg); 5023 5024 return (spa_load(spa, state, SPA_IMPORT_EXISTING)); 5025 } 5026 5027 /* 5028 * If spa_load() fails this function will try loading prior txg's. If 5029 * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool 5030 * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this 5031 * function will not rewind the pool and will return the same error as 5032 * spa_load(). 5033 */ 5034 static int 5035 spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request, 5036 int rewind_flags) 5037 { 5038 nvlist_t *loadinfo = NULL; 5039 nvlist_t *config = NULL; 5040 int load_error, rewind_error; 5041 uint64_t safe_rewind_txg; 5042 uint64_t min_txg; 5043 5044 if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) { 5045 spa->spa_load_max_txg = spa->spa_load_txg; 5046 spa_set_log_state(spa, SPA_LOG_CLEAR); 5047 } else { 5048 spa->spa_load_max_txg = max_request; 5049 if (max_request != UINT64_MAX) 5050 spa->spa_extreme_rewind = B_TRUE; 5051 } 5052 5053 load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING); 5054 if (load_error == 0) 5055 return (0); 5056 if (load_error == ZFS_ERR_NO_CHECKPOINT) { 5057 /* 5058 * When attempting checkpoint-rewind on a pool with no 5059 * checkpoint, we should not attempt to load uberblocks 5060 * from previous txgs when spa_load fails. 5061 */ 5062 ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); 5063 spa_import_progress_remove(spa_guid(spa)); 5064 return (load_error); 5065 } 5066 5067 if (spa->spa_root_vdev != NULL) 5068 config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); 5069 5070 spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg; 5071 spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp; 5072 5073 if (rewind_flags & ZPOOL_NEVER_REWIND) { 5074 nvlist_free(config); 5075 spa_import_progress_remove(spa_guid(spa)); 5076 return (load_error); 5077 } 5078 5079 if (state == SPA_LOAD_RECOVER) { 5080 /* Price of rolling back is discarding txgs, including log */ 5081 spa_set_log_state(spa, SPA_LOG_CLEAR); 5082 } else { 5083 /* 5084 * If we aren't rolling back save the load info from our first 5085 * import attempt so that we can restore it after attempting 5086 * to rewind. 5087 */ 5088 loadinfo = spa->spa_load_info; 5089 spa->spa_load_info = fnvlist_alloc(); 5090 } 5091 5092 spa->spa_load_max_txg = spa->spa_last_ubsync_txg; 5093 safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE; 5094 min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ? 5095 TXG_INITIAL : safe_rewind_txg; 5096 5097 /* 5098 * Continue as long as we're finding errors, we're still within 5099 * the acceptable rewind range, and we're still finding uberblocks 5100 */ 5101 while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg && 5102 spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) { 5103 if (spa->spa_load_max_txg < safe_rewind_txg) 5104 spa->spa_extreme_rewind = B_TRUE; 5105 rewind_error = spa_load_retry(spa, state); 5106 } 5107 5108 spa->spa_extreme_rewind = B_FALSE; 5109 spa->spa_load_max_txg = UINT64_MAX; 5110 5111 if (config && (rewind_error || state != SPA_LOAD_RECOVER)) 5112 spa_config_set(spa, config); 5113 else 5114 nvlist_free(config); 5115 5116 if (state == SPA_LOAD_RECOVER) { 5117 ASSERT3P(loadinfo, ==, NULL); 5118 spa_import_progress_remove(spa_guid(spa)); 5119 return (rewind_error); 5120 } else { 5121 /* Store the rewind info as part of the initial load info */ 5122 fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO, 5123 spa->spa_load_info); 5124 5125 /* Restore the initial load info */ 5126 fnvlist_free(spa->spa_load_info); 5127 spa->spa_load_info = loadinfo; 5128 5129 spa_import_progress_remove(spa_guid(spa)); 5130 return (load_error); 5131 } 5132 } 5133 5134 /* 5135 * Pool Open/Import 5136 * 5137 * The import case is identical to an open except that the configuration is sent 5138 * down from userland, instead of grabbed from the configuration cache. For the 5139 * case of an open, the pool configuration will exist in the 5140 * POOL_STATE_UNINITIALIZED state. 5141 * 5142 * The stats information (gen/count/ustats) is used to gather vdev statistics at 5143 * the same time open the pool, without having to keep around the spa_t in some 5144 * ambiguous state. 5145 */ 5146 static int 5147 spa_open_common(const char *pool, spa_t **spapp, void *tag, nvlist_t *nvpolicy, 5148 nvlist_t **config) 5149 { 5150 spa_t *spa; 5151 spa_load_state_t state = SPA_LOAD_OPEN; 5152 int error; 5153 int locked = B_FALSE; 5154 int firstopen = B_FALSE; 5155 5156 *spapp = NULL; 5157 5158 /* 5159 * As disgusting as this is, we need to support recursive calls to this 5160 * function because dsl_dir_open() is called during spa_load(), and ends 5161 * up calling spa_open() again. The real fix is to figure out how to 5162 * avoid dsl_dir_open() calling this in the first place. 5163 */ 5164 if (MUTEX_NOT_HELD(&spa_namespace_lock)) { 5165 mutex_enter(&spa_namespace_lock); 5166 locked = B_TRUE; 5167 } 5168 5169 if ((spa = spa_lookup(pool)) == NULL) { 5170 if (locked) 5171 mutex_exit(&spa_namespace_lock); 5172 return (SET_ERROR(ENOENT)); 5173 } 5174 5175 if (spa->spa_state == POOL_STATE_UNINITIALIZED) { 5176 zpool_load_policy_t policy; 5177 5178 firstopen = B_TRUE; 5179 5180 zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config, 5181 &policy); 5182 if (policy.zlp_rewind & ZPOOL_DO_REWIND) 5183 state = SPA_LOAD_RECOVER; 5184 5185 spa_activate(spa, spa_mode_global); 5186 5187 if (state != SPA_LOAD_RECOVER) 5188 spa->spa_last_ubsync_txg = spa->spa_load_txg = 0; 5189 spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE; 5190 5191 zfs_dbgmsg("spa_open_common: opening %s", pool); 5192 error = spa_load_best(spa, state, policy.zlp_txg, 5193 policy.zlp_rewind); 5194 5195 if (error == EBADF) { 5196 /* 5197 * If vdev_validate() returns failure (indicated by 5198 * EBADF), it indicates that one of the vdevs indicates 5199 * that the pool has been exported or destroyed. If 5200 * this is the case, the config cache is out of sync and 5201 * we should remove the pool from the namespace. 5202 */ 5203 spa_unload(spa); 5204 spa_deactivate(spa); 5205 spa_write_cachefile(spa, B_TRUE, B_TRUE); 5206 spa_remove(spa); 5207 if (locked) 5208 mutex_exit(&spa_namespace_lock); 5209 return (SET_ERROR(ENOENT)); 5210 } 5211 5212 if (error) { 5213 /* 5214 * We can't open the pool, but we still have useful 5215 * information: the state of each vdev after the 5216 * attempted vdev_open(). Return this to the user. 5217 */ 5218 if (config != NULL && spa->spa_config) { 5219 *config = fnvlist_dup(spa->spa_config); 5220 fnvlist_add_nvlist(*config, 5221 ZPOOL_CONFIG_LOAD_INFO, 5222 spa->spa_load_info); 5223 } 5224 spa_unload(spa); 5225 spa_deactivate(spa); 5226 spa->spa_last_open_failed = error; 5227 if (locked) 5228 mutex_exit(&spa_namespace_lock); 5229 *spapp = NULL; 5230 return (error); 5231 } 5232 } 5233 5234 spa_open_ref(spa, tag); 5235 5236 if (config != NULL) 5237 *config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); 5238 5239 /* 5240 * If we've recovered the pool, pass back any information we 5241 * gathered while doing the load. 5242 */ 5243 if (state == SPA_LOAD_RECOVER) { 5244 fnvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO, 5245 spa->spa_load_info); 5246 } 5247 5248 if (locked) { 5249 spa->spa_last_open_failed = 0; 5250 spa->spa_last_ubsync_txg = 0; 5251 spa->spa_load_txg = 0; 5252 mutex_exit(&spa_namespace_lock); 5253 } 5254 5255 if (firstopen) 5256 zvol_create_minors_recursive(spa_name(spa)); 5257 5258 *spapp = spa; 5259 5260 return (0); 5261 } 5262 5263 int 5264 spa_open_rewind(const char *name, spa_t **spapp, void *tag, nvlist_t *policy, 5265 nvlist_t **config) 5266 { 5267 return (spa_open_common(name, spapp, tag, policy, config)); 5268 } 5269 5270 int 5271 spa_open(const char *name, spa_t **spapp, void *tag) 5272 { 5273 return (spa_open_common(name, spapp, tag, NULL, NULL)); 5274 } 5275 5276 /* 5277 * Lookup the given spa_t, incrementing the inject count in the process, 5278 * preventing it from being exported or destroyed. 5279 */ 5280 spa_t * 5281 spa_inject_addref(char *name) 5282 { 5283 spa_t *spa; 5284 5285 mutex_enter(&spa_namespace_lock); 5286 if ((spa = spa_lookup(name)) == NULL) { 5287 mutex_exit(&spa_namespace_lock); 5288 return (NULL); 5289 } 5290 spa->spa_inject_ref++; 5291 mutex_exit(&spa_namespace_lock); 5292 5293 return (spa); 5294 } 5295 5296 void 5297 spa_inject_delref(spa_t *spa) 5298 { 5299 mutex_enter(&spa_namespace_lock); 5300 spa->spa_inject_ref--; 5301 mutex_exit(&spa_namespace_lock); 5302 } 5303 5304 /* 5305 * Add spares device information to the nvlist. 5306 */ 5307 static void 5308 spa_add_spares(spa_t *spa, nvlist_t *config) 5309 { 5310 nvlist_t **spares; 5311 uint_t i, nspares; 5312 nvlist_t *nvroot; 5313 uint64_t guid; 5314 vdev_stat_t *vs; 5315 uint_t vsc; 5316 uint64_t pool; 5317 5318 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 5319 5320 if (spa->spa_spares.sav_count == 0) 5321 return; 5322 5323 nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE); 5324 VERIFY0(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, 5325 ZPOOL_CONFIG_SPARES, &spares, &nspares)); 5326 if (nspares != 0) { 5327 fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 5328 (const nvlist_t * const *)spares, nspares); 5329 VERIFY0(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 5330 &spares, &nspares)); 5331 5332 /* 5333 * Go through and find any spares which have since been 5334 * repurposed as an active spare. If this is the case, update 5335 * their status appropriately. 5336 */ 5337 for (i = 0; i < nspares; i++) { 5338 guid = fnvlist_lookup_uint64(spares[i], 5339 ZPOOL_CONFIG_GUID); 5340 if (spa_spare_exists(guid, &pool, NULL) && 5341 pool != 0ULL) { 5342 VERIFY0(nvlist_lookup_uint64_array(spares[i], 5343 ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, 5344 &vsc)); 5345 vs->vs_state = VDEV_STATE_CANT_OPEN; 5346 vs->vs_aux = VDEV_AUX_SPARED; 5347 } 5348 } 5349 } 5350 } 5351 5352 /* 5353 * Add l2cache device information to the nvlist, including vdev stats. 5354 */ 5355 static void 5356 spa_add_l2cache(spa_t *spa, nvlist_t *config) 5357 { 5358 nvlist_t **l2cache; 5359 uint_t i, j, nl2cache; 5360 nvlist_t *nvroot; 5361 uint64_t guid; 5362 vdev_t *vd; 5363 vdev_stat_t *vs; 5364 uint_t vsc; 5365 5366 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 5367 5368 if (spa->spa_l2cache.sav_count == 0) 5369 return; 5370 5371 nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE); 5372 VERIFY0(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config, 5373 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache)); 5374 if (nl2cache != 0) { 5375 fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, 5376 (const nvlist_t * const *)l2cache, nl2cache); 5377 VERIFY0(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, 5378 &l2cache, &nl2cache)); 5379 5380 /* 5381 * Update level 2 cache device stats. 5382 */ 5383 5384 for (i = 0; i < nl2cache; i++) { 5385 guid = fnvlist_lookup_uint64(l2cache[i], 5386 ZPOOL_CONFIG_GUID); 5387 5388 vd = NULL; 5389 for (j = 0; j < spa->spa_l2cache.sav_count; j++) { 5390 if (guid == 5391 spa->spa_l2cache.sav_vdevs[j]->vdev_guid) { 5392 vd = spa->spa_l2cache.sav_vdevs[j]; 5393 break; 5394 } 5395 } 5396 ASSERT(vd != NULL); 5397 5398 VERIFY0(nvlist_lookup_uint64_array(l2cache[i], 5399 ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc)); 5400 vdev_get_stats(vd, vs); 5401 vdev_config_generate_stats(vd, l2cache[i]); 5402 5403 } 5404 } 5405 } 5406 5407 static void 5408 spa_feature_stats_from_disk(spa_t *spa, nvlist_t *features) 5409 { 5410 zap_cursor_t zc; 5411 zap_attribute_t za; 5412 5413 if (spa->spa_feat_for_read_obj != 0) { 5414 for (zap_cursor_init(&zc, spa->spa_meta_objset, 5415 spa->spa_feat_for_read_obj); 5416 zap_cursor_retrieve(&zc, &za) == 0; 5417 zap_cursor_advance(&zc)) { 5418 ASSERT(za.za_integer_length == sizeof (uint64_t) && 5419 za.za_num_integers == 1); 5420 VERIFY0(nvlist_add_uint64(features, za.za_name, 5421 za.za_first_integer)); 5422 } 5423 zap_cursor_fini(&zc); 5424 } 5425 5426 if (spa->spa_feat_for_write_obj != 0) { 5427 for (zap_cursor_init(&zc, spa->spa_meta_objset, 5428 spa->spa_feat_for_write_obj); 5429 zap_cursor_retrieve(&zc, &za) == 0; 5430 zap_cursor_advance(&zc)) { 5431 ASSERT(za.za_integer_length == sizeof (uint64_t) && 5432 za.za_num_integers == 1); 5433 VERIFY0(nvlist_add_uint64(features, za.za_name, 5434 za.za_first_integer)); 5435 } 5436 zap_cursor_fini(&zc); 5437 } 5438 } 5439 5440 static void 5441 spa_feature_stats_from_cache(spa_t *spa, nvlist_t *features) 5442 { 5443 int i; 5444 5445 for (i = 0; i < SPA_FEATURES; i++) { 5446 zfeature_info_t feature = spa_feature_table[i]; 5447 uint64_t refcount; 5448 5449 if (feature_get_refcount(spa, &feature, &refcount) != 0) 5450 continue; 5451 5452 VERIFY0(nvlist_add_uint64(features, feature.fi_guid, refcount)); 5453 } 5454 } 5455 5456 /* 5457 * Store a list of pool features and their reference counts in the 5458 * config. 5459 * 5460 * The first time this is called on a spa, allocate a new nvlist, fetch 5461 * the pool features and reference counts from disk, then save the list 5462 * in the spa. In subsequent calls on the same spa use the saved nvlist 5463 * and refresh its values from the cached reference counts. This 5464 * ensures we don't block here on I/O on a suspended pool so 'zpool 5465 * clear' can resume the pool. 5466 */ 5467 static void 5468 spa_add_feature_stats(spa_t *spa, nvlist_t *config) 5469 { 5470 nvlist_t *features; 5471 5472 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 5473 5474 mutex_enter(&spa->spa_feat_stats_lock); 5475 features = spa->spa_feat_stats; 5476 5477 if (features != NULL) { 5478 spa_feature_stats_from_cache(spa, features); 5479 } else { 5480 VERIFY0(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP)); 5481 spa->spa_feat_stats = features; 5482 spa_feature_stats_from_disk(spa, features); 5483 } 5484 5485 VERIFY0(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS, 5486 features)); 5487 5488 mutex_exit(&spa->spa_feat_stats_lock); 5489 } 5490 5491 int 5492 spa_get_stats(const char *name, nvlist_t **config, 5493 char *altroot, size_t buflen) 5494 { 5495 int error; 5496 spa_t *spa; 5497 5498 *config = NULL; 5499 error = spa_open_common(name, &spa, FTAG, NULL, config); 5500 5501 if (spa != NULL) { 5502 /* 5503 * This still leaves a window of inconsistency where the spares 5504 * or l2cache devices could change and the config would be 5505 * self-inconsistent. 5506 */ 5507 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 5508 5509 if (*config != NULL) { 5510 uint64_t loadtimes[2]; 5511 5512 loadtimes[0] = spa->spa_loaded_ts.tv_sec; 5513 loadtimes[1] = spa->spa_loaded_ts.tv_nsec; 5514 fnvlist_add_uint64_array(*config, 5515 ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2); 5516 5517 fnvlist_add_uint64(*config, 5518 ZPOOL_CONFIG_ERRCOUNT, 5519 spa_get_errlog_size(spa)); 5520 5521 if (spa_suspended(spa)) { 5522 fnvlist_add_uint64(*config, 5523 ZPOOL_CONFIG_SUSPENDED, 5524 spa->spa_failmode); 5525 fnvlist_add_uint64(*config, 5526 ZPOOL_CONFIG_SUSPENDED_REASON, 5527 spa->spa_suspended); 5528 } 5529 5530 spa_add_spares(spa, *config); 5531 spa_add_l2cache(spa, *config); 5532 spa_add_feature_stats(spa, *config); 5533 } 5534 } 5535 5536 /* 5537 * We want to get the alternate root even for faulted pools, so we cheat 5538 * and call spa_lookup() directly. 5539 */ 5540 if (altroot) { 5541 if (spa == NULL) { 5542 mutex_enter(&spa_namespace_lock); 5543 spa = spa_lookup(name); 5544 if (spa) 5545 spa_altroot(spa, altroot, buflen); 5546 else 5547 altroot[0] = '\0'; 5548 spa = NULL; 5549 mutex_exit(&spa_namespace_lock); 5550 } else { 5551 spa_altroot(spa, altroot, buflen); 5552 } 5553 } 5554 5555 if (spa != NULL) { 5556 spa_config_exit(spa, SCL_CONFIG, FTAG); 5557 spa_close(spa, FTAG); 5558 } 5559 5560 return (error); 5561 } 5562 5563 /* 5564 * Validate that the auxiliary device array is well formed. We must have an 5565 * array of nvlists, each which describes a valid leaf vdev. If this is an 5566 * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be 5567 * specified, as long as they are well-formed. 5568 */ 5569 static int 5570 spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode, 5571 spa_aux_vdev_t *sav, const char *config, uint64_t version, 5572 vdev_labeltype_t label) 5573 { 5574 nvlist_t **dev; 5575 uint_t i, ndev; 5576 vdev_t *vd; 5577 int error; 5578 5579 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 5580 5581 /* 5582 * It's acceptable to have no devs specified. 5583 */ 5584 if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0) 5585 return (0); 5586 5587 if (ndev == 0) 5588 return (SET_ERROR(EINVAL)); 5589 5590 /* 5591 * Make sure the pool is formatted with a version that supports this 5592 * device type. 5593 */ 5594 if (spa_version(spa) < version) 5595 return (SET_ERROR(ENOTSUP)); 5596 5597 /* 5598 * Set the pending device list so we correctly handle device in-use 5599 * checking. 5600 */ 5601 sav->sav_pending = dev; 5602 sav->sav_npending = ndev; 5603 5604 for (i = 0; i < ndev; i++) { 5605 if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0, 5606 mode)) != 0) 5607 goto out; 5608 5609 if (!vd->vdev_ops->vdev_op_leaf) { 5610 vdev_free(vd); 5611 error = SET_ERROR(EINVAL); 5612 goto out; 5613 } 5614 5615 vd->vdev_top = vd; 5616 5617 if ((error = vdev_open(vd)) == 0 && 5618 (error = vdev_label_init(vd, crtxg, label)) == 0) { 5619 fnvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID, 5620 vd->vdev_guid); 5621 } 5622 5623 vdev_free(vd); 5624 5625 if (error && 5626 (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE)) 5627 goto out; 5628 else 5629 error = 0; 5630 } 5631 5632 out: 5633 sav->sav_pending = NULL; 5634 sav->sav_npending = 0; 5635 return (error); 5636 } 5637 5638 static int 5639 spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode) 5640 { 5641 int error; 5642 5643 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 5644 5645 if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode, 5646 &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES, 5647 VDEV_LABEL_SPARE)) != 0) { 5648 return (error); 5649 } 5650 5651 return (spa_validate_aux_devs(spa, nvroot, crtxg, mode, 5652 &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE, 5653 VDEV_LABEL_L2CACHE)); 5654 } 5655 5656 static void 5657 spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs, 5658 const char *config) 5659 { 5660 int i; 5661 5662 if (sav->sav_config != NULL) { 5663 nvlist_t **olddevs; 5664 uint_t oldndevs; 5665 nvlist_t **newdevs; 5666 5667 /* 5668 * Generate new dev list by concatenating with the 5669 * current dev list. 5670 */ 5671 VERIFY0(nvlist_lookup_nvlist_array(sav->sav_config, config, 5672 &olddevs, &oldndevs)); 5673 5674 newdevs = kmem_alloc(sizeof (void *) * 5675 (ndevs + oldndevs), KM_SLEEP); 5676 for (i = 0; i < oldndevs; i++) 5677 newdevs[i] = fnvlist_dup(olddevs[i]); 5678 for (i = 0; i < ndevs; i++) 5679 newdevs[i + oldndevs] = fnvlist_dup(devs[i]); 5680 5681 fnvlist_remove(sav->sav_config, config); 5682 5683 fnvlist_add_nvlist_array(sav->sav_config, config, 5684 (const nvlist_t * const *)newdevs, ndevs + oldndevs); 5685 for (i = 0; i < oldndevs + ndevs; i++) 5686 nvlist_free(newdevs[i]); 5687 kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *)); 5688 } else { 5689 /* 5690 * Generate a new dev list. 5691 */ 5692 sav->sav_config = fnvlist_alloc(); 5693 fnvlist_add_nvlist_array(sav->sav_config, config, 5694 (const nvlist_t * const *)devs, ndevs); 5695 } 5696 } 5697 5698 /* 5699 * Stop and drop level 2 ARC devices 5700 */ 5701 void 5702 spa_l2cache_drop(spa_t *spa) 5703 { 5704 vdev_t *vd; 5705 int i; 5706 spa_aux_vdev_t *sav = &spa->spa_l2cache; 5707 5708 for (i = 0; i < sav->sav_count; i++) { 5709 uint64_t pool; 5710 5711 vd = sav->sav_vdevs[i]; 5712 ASSERT(vd != NULL); 5713 5714 if (spa_l2cache_exists(vd->vdev_guid, &pool) && 5715 pool != 0ULL && l2arc_vdev_present(vd)) 5716 l2arc_remove_vdev(vd); 5717 } 5718 } 5719 5720 /* 5721 * Verify encryption parameters for spa creation. If we are encrypting, we must 5722 * have the encryption feature flag enabled. 5723 */ 5724 static int 5725 spa_create_check_encryption_params(dsl_crypto_params_t *dcp, 5726 boolean_t has_encryption) 5727 { 5728 if (dcp->cp_crypt != ZIO_CRYPT_OFF && 5729 dcp->cp_crypt != ZIO_CRYPT_INHERIT && 5730 !has_encryption) 5731 return (SET_ERROR(ENOTSUP)); 5732 5733 return (dmu_objset_create_crypt_check(NULL, dcp, NULL)); 5734 } 5735 5736 /* 5737 * Pool Creation 5738 */ 5739 int 5740 spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, 5741 nvlist_t *zplprops, dsl_crypto_params_t *dcp) 5742 { 5743 spa_t *spa; 5744 char *altroot = NULL; 5745 vdev_t *rvd; 5746 dsl_pool_t *dp; 5747 dmu_tx_t *tx; 5748 int error = 0; 5749 uint64_t txg = TXG_INITIAL; 5750 nvlist_t **spares, **l2cache; 5751 uint_t nspares, nl2cache; 5752 uint64_t version, obj, ndraid = 0; 5753 boolean_t has_features; 5754 boolean_t has_encryption; 5755 boolean_t has_allocclass; 5756 spa_feature_t feat; 5757 char *feat_name; 5758 char *poolname; 5759 nvlist_t *nvl; 5760 5761 if (props == NULL || 5762 nvlist_lookup_string(props, "tname", &poolname) != 0) 5763 poolname = (char *)pool; 5764 5765 /* 5766 * If this pool already exists, return failure. 5767 */ 5768 mutex_enter(&spa_namespace_lock); 5769 if (spa_lookup(poolname) != NULL) { 5770 mutex_exit(&spa_namespace_lock); 5771 return (SET_ERROR(EEXIST)); 5772 } 5773 5774 /* 5775 * Allocate a new spa_t structure. 5776 */ 5777 nvl = fnvlist_alloc(); 5778 fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool); 5779 (void) nvlist_lookup_string(props, 5780 zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); 5781 spa = spa_add(poolname, nvl, altroot); 5782 fnvlist_free(nvl); 5783 spa_activate(spa, spa_mode_global); 5784 5785 if (props && (error = spa_prop_validate(spa, props))) { 5786 spa_deactivate(spa); 5787 spa_remove(spa); 5788 mutex_exit(&spa_namespace_lock); 5789 return (error); 5790 } 5791 5792 /* 5793 * Temporary pool names should never be written to disk. 5794 */ 5795 if (poolname != pool) 5796 spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME; 5797 5798 has_features = B_FALSE; 5799 has_encryption = B_FALSE; 5800 has_allocclass = B_FALSE; 5801 for (nvpair_t *elem = nvlist_next_nvpair(props, NULL); 5802 elem != NULL; elem = nvlist_next_nvpair(props, elem)) { 5803 if (zpool_prop_feature(nvpair_name(elem))) { 5804 has_features = B_TRUE; 5805 5806 feat_name = strchr(nvpair_name(elem), '@') + 1; 5807 VERIFY0(zfeature_lookup_name(feat_name, &feat)); 5808 if (feat == SPA_FEATURE_ENCRYPTION) 5809 has_encryption = B_TRUE; 5810 if (feat == SPA_FEATURE_ALLOCATION_CLASSES) 5811 has_allocclass = B_TRUE; 5812 } 5813 } 5814 5815 /* verify encryption params, if they were provided */ 5816 if (dcp != NULL) { 5817 error = spa_create_check_encryption_params(dcp, has_encryption); 5818 if (error != 0) { 5819 spa_deactivate(spa); 5820 spa_remove(spa); 5821 mutex_exit(&spa_namespace_lock); 5822 return (error); 5823 } 5824 } 5825 if (!has_allocclass && zfs_special_devs(nvroot, NULL)) { 5826 spa_deactivate(spa); 5827 spa_remove(spa); 5828 mutex_exit(&spa_namespace_lock); 5829 return (ENOTSUP); 5830 } 5831 5832 if (has_features || nvlist_lookup_uint64(props, 5833 zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) { 5834 version = SPA_VERSION; 5835 } 5836 ASSERT(SPA_VERSION_IS_SUPPORTED(version)); 5837 5838 spa->spa_first_txg = txg; 5839 spa->spa_uberblock.ub_txg = txg - 1; 5840 spa->spa_uberblock.ub_version = version; 5841 spa->spa_ubsync = spa->spa_uberblock; 5842 spa->spa_load_state = SPA_LOAD_CREATE; 5843 spa->spa_removing_phys.sr_state = DSS_NONE; 5844 spa->spa_removing_phys.sr_removing_vdev = -1; 5845 spa->spa_removing_phys.sr_prev_indirect_vdev = -1; 5846 spa->spa_indirect_vdevs_loaded = B_TRUE; 5847 5848 /* 5849 * Create "The Godfather" zio to hold all async IOs 5850 */ 5851 spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *), 5852 KM_SLEEP); 5853 for (int i = 0; i < max_ncpus; i++) { 5854 spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL, 5855 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 5856 ZIO_FLAG_GODFATHER); 5857 } 5858 5859 /* 5860 * Create the root vdev. 5861 */ 5862 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 5863 5864 error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD); 5865 5866 ASSERT(error != 0 || rvd != NULL); 5867 ASSERT(error != 0 || spa->spa_root_vdev == rvd); 5868 5869 if (error == 0 && !zfs_allocatable_devs(nvroot)) 5870 error = SET_ERROR(EINVAL); 5871 5872 if (error == 0 && 5873 (error = vdev_create(rvd, txg, B_FALSE)) == 0 && 5874 (error = vdev_draid_spare_create(nvroot, rvd, &ndraid, 0)) == 0 && 5875 (error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) == 0) { 5876 /* 5877 * instantiate the metaslab groups (this will dirty the vdevs) 5878 * we can no longer error exit past this point 5879 */ 5880 for (int c = 0; error == 0 && c < rvd->vdev_children; c++) { 5881 vdev_t *vd = rvd->vdev_child[c]; 5882 5883 vdev_metaslab_set_size(vd); 5884 vdev_expand(vd, txg); 5885 } 5886 } 5887 5888 spa_config_exit(spa, SCL_ALL, FTAG); 5889 5890 if (error != 0) { 5891 spa_unload(spa); 5892 spa_deactivate(spa); 5893 spa_remove(spa); 5894 mutex_exit(&spa_namespace_lock); 5895 return (error); 5896 } 5897 5898 /* 5899 * Get the list of spares, if specified. 5900 */ 5901 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 5902 &spares, &nspares) == 0) { 5903 spa->spa_spares.sav_config = fnvlist_alloc(); 5904 fnvlist_add_nvlist_array(spa->spa_spares.sav_config, 5905 ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares, 5906 nspares); 5907 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 5908 spa_load_spares(spa); 5909 spa_config_exit(spa, SCL_ALL, FTAG); 5910 spa->spa_spares.sav_sync = B_TRUE; 5911 } 5912 5913 /* 5914 * Get the list of level 2 cache devices, if specified. 5915 */ 5916 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, 5917 &l2cache, &nl2cache) == 0) { 5918 VERIFY0(nvlist_alloc(&spa->spa_l2cache.sav_config, 5919 NV_UNIQUE_NAME, KM_SLEEP)); 5920 fnvlist_add_nvlist_array(spa->spa_l2cache.sav_config, 5921 ZPOOL_CONFIG_L2CACHE, (const nvlist_t * const *)l2cache, 5922 nl2cache); 5923 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 5924 spa_load_l2cache(spa); 5925 spa_config_exit(spa, SCL_ALL, FTAG); 5926 spa->spa_l2cache.sav_sync = B_TRUE; 5927 } 5928 5929 spa->spa_is_initializing = B_TRUE; 5930 spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg); 5931 spa->spa_is_initializing = B_FALSE; 5932 5933 /* 5934 * Create DDTs (dedup tables). 5935 */ 5936 ddt_create(spa); 5937 5938 spa_update_dspace(spa); 5939 5940 tx = dmu_tx_create_assigned(dp, txg); 5941 5942 /* 5943 * Create the pool's history object. 5944 */ 5945 if (version >= SPA_VERSION_ZPOOL_HISTORY && !spa->spa_history) 5946 spa_history_create_obj(spa, tx); 5947 5948 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE); 5949 spa_history_log_version(spa, "create", tx); 5950 5951 /* 5952 * Create the pool config object. 5953 */ 5954 spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset, 5955 DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE, 5956 DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx); 5957 5958 if (zap_add(spa->spa_meta_objset, 5959 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG, 5960 sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) { 5961 cmn_err(CE_PANIC, "failed to add pool config"); 5962 } 5963 5964 if (zap_add(spa->spa_meta_objset, 5965 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION, 5966 sizeof (uint64_t), 1, &version, tx) != 0) { 5967 cmn_err(CE_PANIC, "failed to add pool version"); 5968 } 5969 5970 /* Newly created pools with the right version are always deflated. */ 5971 if (version >= SPA_VERSION_RAIDZ_DEFLATE) { 5972 spa->spa_deflate = TRUE; 5973 if (zap_add(spa->spa_meta_objset, 5974 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE, 5975 sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) { 5976 cmn_err(CE_PANIC, "failed to add deflate"); 5977 } 5978 } 5979 5980 /* 5981 * Create the deferred-free bpobj. Turn off compression 5982 * because sync-to-convergence takes longer if the blocksize 5983 * keeps changing. 5984 */ 5985 obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx); 5986 dmu_object_set_compress(spa->spa_meta_objset, obj, 5987 ZIO_COMPRESS_OFF, tx); 5988 if (zap_add(spa->spa_meta_objset, 5989 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ, 5990 sizeof (uint64_t), 1, &obj, tx) != 0) { 5991 cmn_err(CE_PANIC, "failed to add bpobj"); 5992 } 5993 VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj, 5994 spa->spa_meta_objset, obj)); 5995 5996 /* 5997 * Generate some random noise for salted checksums to operate on. 5998 */ 5999 (void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes, 6000 sizeof (spa->spa_cksum_salt.zcs_bytes)); 6001 6002 /* 6003 * Set pool properties. 6004 */ 6005 spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS); 6006 spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION); 6007 spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE); 6008 spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND); 6009 spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST); 6010 spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM); 6011 6012 if (props != NULL) { 6013 spa_configfile_set(spa, props, B_FALSE); 6014 spa_sync_props(props, tx); 6015 } 6016 6017 for (int i = 0; i < ndraid; i++) 6018 spa_feature_incr(spa, SPA_FEATURE_DRAID, tx); 6019 6020 dmu_tx_commit(tx); 6021 6022 spa->spa_sync_on = B_TRUE; 6023 txg_sync_start(dp); 6024 mmp_thread_start(spa); 6025 txg_wait_synced(dp, txg); 6026 6027 spa_spawn_aux_threads(spa); 6028 6029 spa_write_cachefile(spa, B_FALSE, B_TRUE); 6030 6031 /* 6032 * Don't count references from objsets that are already closed 6033 * and are making their way through the eviction process. 6034 */ 6035 spa_evicting_os_wait(spa); 6036 spa->spa_minref = zfs_refcount_count(&spa->spa_refcount); 6037 spa->spa_load_state = SPA_LOAD_NONE; 6038 6039 spa_import_os(spa); 6040 6041 mutex_exit(&spa_namespace_lock); 6042 6043 return (0); 6044 } 6045 6046 /* 6047 * Import a non-root pool into the system. 6048 */ 6049 int 6050 spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags) 6051 { 6052 spa_t *spa; 6053 char *altroot = NULL; 6054 spa_load_state_t state = SPA_LOAD_IMPORT; 6055 zpool_load_policy_t policy; 6056 spa_mode_t mode = spa_mode_global; 6057 uint64_t readonly = B_FALSE; 6058 int error; 6059 nvlist_t *nvroot; 6060 nvlist_t **spares, **l2cache; 6061 uint_t nspares, nl2cache; 6062 6063 /* 6064 * If a pool with this name exists, return failure. 6065 */ 6066 mutex_enter(&spa_namespace_lock); 6067 if (spa_lookup(pool) != NULL) { 6068 mutex_exit(&spa_namespace_lock); 6069 return (SET_ERROR(EEXIST)); 6070 } 6071 6072 /* 6073 * Create and initialize the spa structure. 6074 */ 6075 (void) nvlist_lookup_string(props, 6076 zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); 6077 (void) nvlist_lookup_uint64(props, 6078 zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly); 6079 if (readonly) 6080 mode = SPA_MODE_READ; 6081 spa = spa_add(pool, config, altroot); 6082 spa->spa_import_flags = flags; 6083 6084 /* 6085 * Verbatim import - Take a pool and insert it into the namespace 6086 * as if it had been loaded at boot. 6087 */ 6088 if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) { 6089 if (props != NULL) 6090 spa_configfile_set(spa, props, B_FALSE); 6091 6092 spa_write_cachefile(spa, B_FALSE, B_TRUE); 6093 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT); 6094 zfs_dbgmsg("spa_import: verbatim import of %s", pool); 6095 mutex_exit(&spa_namespace_lock); 6096 return (0); 6097 } 6098 6099 spa_activate(spa, mode); 6100 6101 /* 6102 * Don't start async tasks until we know everything is healthy. 6103 */ 6104 spa_async_suspend(spa); 6105 6106 zpool_get_load_policy(config, &policy); 6107 if (policy.zlp_rewind & ZPOOL_DO_REWIND) 6108 state = SPA_LOAD_RECOVER; 6109 6110 spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT; 6111 6112 if (state != SPA_LOAD_RECOVER) { 6113 spa->spa_last_ubsync_txg = spa->spa_load_txg = 0; 6114 zfs_dbgmsg("spa_import: importing %s", pool); 6115 } else { 6116 zfs_dbgmsg("spa_import: importing %s, max_txg=%lld " 6117 "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg); 6118 } 6119 error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind); 6120 6121 /* 6122 * Propagate anything learned while loading the pool and pass it 6123 * back to caller (i.e. rewind info, missing devices, etc). 6124 */ 6125 fnvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info); 6126 6127 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6128 /* 6129 * Toss any existing sparelist, as it doesn't have any validity 6130 * anymore, and conflicts with spa_has_spare(). 6131 */ 6132 if (spa->spa_spares.sav_config) { 6133 nvlist_free(spa->spa_spares.sav_config); 6134 spa->spa_spares.sav_config = NULL; 6135 spa_load_spares(spa); 6136 } 6137 if (spa->spa_l2cache.sav_config) { 6138 nvlist_free(spa->spa_l2cache.sav_config); 6139 spa->spa_l2cache.sav_config = NULL; 6140 spa_load_l2cache(spa); 6141 } 6142 6143 nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE); 6144 spa_config_exit(spa, SCL_ALL, FTAG); 6145 6146 if (props != NULL) 6147 spa_configfile_set(spa, props, B_FALSE); 6148 6149 if (error != 0 || (props && spa_writeable(spa) && 6150 (error = spa_prop_set(spa, props)))) { 6151 spa_unload(spa); 6152 spa_deactivate(spa); 6153 spa_remove(spa); 6154 mutex_exit(&spa_namespace_lock); 6155 return (error); 6156 } 6157 6158 spa_async_resume(spa); 6159 6160 /* 6161 * Override any spares and level 2 cache devices as specified by 6162 * the user, as these may have correct device names/devids, etc. 6163 */ 6164 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 6165 &spares, &nspares) == 0) { 6166 if (spa->spa_spares.sav_config) 6167 fnvlist_remove(spa->spa_spares.sav_config, 6168 ZPOOL_CONFIG_SPARES); 6169 else 6170 spa->spa_spares.sav_config = fnvlist_alloc(); 6171 fnvlist_add_nvlist_array(spa->spa_spares.sav_config, 6172 ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares, 6173 nspares); 6174 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6175 spa_load_spares(spa); 6176 spa_config_exit(spa, SCL_ALL, FTAG); 6177 spa->spa_spares.sav_sync = B_TRUE; 6178 } 6179 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, 6180 &l2cache, &nl2cache) == 0) { 6181 if (spa->spa_l2cache.sav_config) 6182 fnvlist_remove(spa->spa_l2cache.sav_config, 6183 ZPOOL_CONFIG_L2CACHE); 6184 else 6185 spa->spa_l2cache.sav_config = fnvlist_alloc(); 6186 fnvlist_add_nvlist_array(spa->spa_l2cache.sav_config, 6187 ZPOOL_CONFIG_L2CACHE, (const nvlist_t * const *)l2cache, 6188 nl2cache); 6189 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6190 spa_load_l2cache(spa); 6191 spa_config_exit(spa, SCL_ALL, FTAG); 6192 spa->spa_l2cache.sav_sync = B_TRUE; 6193 } 6194 6195 /* 6196 * Check for any removed devices. 6197 */ 6198 if (spa->spa_autoreplace) { 6199 spa_aux_check_removed(&spa->spa_spares); 6200 spa_aux_check_removed(&spa->spa_l2cache); 6201 } 6202 6203 if (spa_writeable(spa)) { 6204 /* 6205 * Update the config cache to include the newly-imported pool. 6206 */ 6207 spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); 6208 } 6209 6210 /* 6211 * It's possible that the pool was expanded while it was exported. 6212 * We kick off an async task to handle this for us. 6213 */ 6214 spa_async_request(spa, SPA_ASYNC_AUTOEXPAND); 6215 6216 spa_history_log_version(spa, "import", NULL); 6217 6218 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT); 6219 6220 mutex_exit(&spa_namespace_lock); 6221 6222 zvol_create_minors_recursive(pool); 6223 6224 spa_import_os(spa); 6225 6226 return (0); 6227 } 6228 6229 nvlist_t * 6230 spa_tryimport(nvlist_t *tryconfig) 6231 { 6232 nvlist_t *config = NULL; 6233 char *poolname, *cachefile; 6234 spa_t *spa; 6235 uint64_t state; 6236 int error; 6237 zpool_load_policy_t policy; 6238 6239 if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname)) 6240 return (NULL); 6241 6242 if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state)) 6243 return (NULL); 6244 6245 /* 6246 * Create and initialize the spa structure. 6247 */ 6248 mutex_enter(&spa_namespace_lock); 6249 spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL); 6250 spa_activate(spa, SPA_MODE_READ); 6251 6252 /* 6253 * Rewind pool if a max txg was provided. 6254 */ 6255 zpool_get_load_policy(spa->spa_config, &policy); 6256 if (policy.zlp_txg != UINT64_MAX) { 6257 spa->spa_load_max_txg = policy.zlp_txg; 6258 spa->spa_extreme_rewind = B_TRUE; 6259 zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld", 6260 poolname, (longlong_t)policy.zlp_txg); 6261 } else { 6262 zfs_dbgmsg("spa_tryimport: importing %s", poolname); 6263 } 6264 6265 if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile) 6266 == 0) { 6267 zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile); 6268 spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE; 6269 } else { 6270 spa->spa_config_source = SPA_CONFIG_SRC_SCAN; 6271 } 6272 6273 error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING); 6274 6275 /* 6276 * If 'tryconfig' was at least parsable, return the current config. 6277 */ 6278 if (spa->spa_root_vdev != NULL) { 6279 config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); 6280 fnvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, poolname); 6281 fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, state); 6282 fnvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP, 6283 spa->spa_uberblock.ub_timestamp); 6284 fnvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, 6285 spa->spa_load_info); 6286 fnvlist_add_uint64(config, ZPOOL_CONFIG_ERRATA, 6287 spa->spa_errata); 6288 6289 /* 6290 * If the bootfs property exists on this pool then we 6291 * copy it out so that external consumers can tell which 6292 * pools are bootable. 6293 */ 6294 if ((!error || error == EEXIST) && spa->spa_bootfs) { 6295 char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP); 6296 6297 /* 6298 * We have to play games with the name since the 6299 * pool was opened as TRYIMPORT_NAME. 6300 */ 6301 if (dsl_dsobj_to_dsname(spa_name(spa), 6302 spa->spa_bootfs, tmpname) == 0) { 6303 char *cp; 6304 char *dsname; 6305 6306 dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP); 6307 6308 cp = strchr(tmpname, '/'); 6309 if (cp == NULL) { 6310 (void) strlcpy(dsname, tmpname, 6311 MAXPATHLEN); 6312 } else { 6313 (void) snprintf(dsname, MAXPATHLEN, 6314 "%s/%s", poolname, ++cp); 6315 } 6316 fnvlist_add_string(config, ZPOOL_CONFIG_BOOTFS, 6317 dsname); 6318 kmem_free(dsname, MAXPATHLEN); 6319 } 6320 kmem_free(tmpname, MAXPATHLEN); 6321 } 6322 6323 /* 6324 * Add the list of hot spares and level 2 cache devices. 6325 */ 6326 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 6327 spa_add_spares(spa, config); 6328 spa_add_l2cache(spa, config); 6329 spa_config_exit(spa, SCL_CONFIG, FTAG); 6330 } 6331 6332 spa_unload(spa); 6333 spa_deactivate(spa); 6334 spa_remove(spa); 6335 mutex_exit(&spa_namespace_lock); 6336 6337 return (config); 6338 } 6339 6340 /* 6341 * Pool export/destroy 6342 * 6343 * The act of destroying or exporting a pool is very simple. We make sure there 6344 * is no more pending I/O and any references to the pool are gone. Then, we 6345 * update the pool state and sync all the labels to disk, removing the 6346 * configuration from the cache afterwards. If the 'hardforce' flag is set, then 6347 * we don't sync the labels or remove the configuration cache. 6348 */ 6349 static int 6350 spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig, 6351 boolean_t force, boolean_t hardforce) 6352 { 6353 int error; 6354 spa_t *spa; 6355 6356 if (oldconfig) 6357 *oldconfig = NULL; 6358 6359 if (!(spa_mode_global & SPA_MODE_WRITE)) 6360 return (SET_ERROR(EROFS)); 6361 6362 mutex_enter(&spa_namespace_lock); 6363 if ((spa = spa_lookup(pool)) == NULL) { 6364 mutex_exit(&spa_namespace_lock); 6365 return (SET_ERROR(ENOENT)); 6366 } 6367 6368 if (spa->spa_is_exporting) { 6369 /* the pool is being exported by another thread */ 6370 mutex_exit(&spa_namespace_lock); 6371 return (SET_ERROR(ZFS_ERR_EXPORT_IN_PROGRESS)); 6372 } 6373 spa->spa_is_exporting = B_TRUE; 6374 6375 /* 6376 * Put a hold on the pool, drop the namespace lock, stop async tasks, 6377 * reacquire the namespace lock, and see if we can export. 6378 */ 6379 spa_open_ref(spa, FTAG); 6380 mutex_exit(&spa_namespace_lock); 6381 spa_async_suspend(spa); 6382 if (spa->spa_zvol_taskq) { 6383 zvol_remove_minors(spa, spa_name(spa), B_TRUE); 6384 taskq_wait(spa->spa_zvol_taskq); 6385 } 6386 mutex_enter(&spa_namespace_lock); 6387 spa_close(spa, FTAG); 6388 6389 if (spa->spa_state == POOL_STATE_UNINITIALIZED) 6390 goto export_spa; 6391 /* 6392 * The pool will be in core if it's openable, in which case we can 6393 * modify its state. Objsets may be open only because they're dirty, 6394 * so we have to force it to sync before checking spa_refcnt. 6395 */ 6396 if (spa->spa_sync_on) { 6397 txg_wait_synced(spa->spa_dsl_pool, 0); 6398 spa_evicting_os_wait(spa); 6399 } 6400 6401 /* 6402 * A pool cannot be exported or destroyed if there are active 6403 * references. If we are resetting a pool, allow references by 6404 * fault injection handlers. 6405 */ 6406 if (!spa_refcount_zero(spa) || (spa->spa_inject_ref != 0)) { 6407 error = SET_ERROR(EBUSY); 6408 goto fail; 6409 } 6410 6411 if (spa->spa_sync_on) { 6412 /* 6413 * A pool cannot be exported if it has an active shared spare. 6414 * This is to prevent other pools stealing the active spare 6415 * from an exported pool. At user's own will, such pool can 6416 * be forcedly exported. 6417 */ 6418 if (!force && new_state == POOL_STATE_EXPORTED && 6419 spa_has_active_shared_spare(spa)) { 6420 error = SET_ERROR(EXDEV); 6421 goto fail; 6422 } 6423 6424 /* 6425 * We're about to export or destroy this pool. Make sure 6426 * we stop all initialization and trim activity here before 6427 * we set the spa_final_txg. This will ensure that all 6428 * dirty data resulting from the initialization is 6429 * committed to disk before we unload the pool. 6430 */ 6431 if (spa->spa_root_vdev != NULL) { 6432 vdev_t *rvd = spa->spa_root_vdev; 6433 vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE); 6434 vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE); 6435 vdev_autotrim_stop_all(spa); 6436 vdev_rebuild_stop_all(spa); 6437 } 6438 6439 /* 6440 * We want this to be reflected on every label, 6441 * so mark them all dirty. spa_unload() will do the 6442 * final sync that pushes these changes out. 6443 */ 6444 if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) { 6445 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6446 spa->spa_state = new_state; 6447 vdev_config_dirty(spa->spa_root_vdev); 6448 spa_config_exit(spa, SCL_ALL, FTAG); 6449 } 6450 6451 /* 6452 * If the log space map feature is enabled and the pool is 6453 * getting exported (but not destroyed), we want to spend some 6454 * time flushing as many metaslabs as we can in an attempt to 6455 * destroy log space maps and save import time. This has to be 6456 * done before we set the spa_final_txg, otherwise 6457 * spa_sync() -> spa_flush_metaslabs() may dirty the final TXGs. 6458 * spa_should_flush_logs_on_unload() should be called after 6459 * spa_state has been set to the new_state. 6460 */ 6461 if (spa_should_flush_logs_on_unload(spa)) 6462 spa_unload_log_sm_flush_all(spa); 6463 6464 if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) { 6465 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6466 spa->spa_final_txg = spa_last_synced_txg(spa) + 6467 TXG_DEFER_SIZE + 1; 6468 spa_config_exit(spa, SCL_ALL, FTAG); 6469 } 6470 } 6471 6472 export_spa: 6473 spa_export_os(spa); 6474 6475 if (new_state == POOL_STATE_DESTROYED) 6476 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY); 6477 else if (new_state == POOL_STATE_EXPORTED) 6478 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_EXPORT); 6479 6480 if (spa->spa_state != POOL_STATE_UNINITIALIZED) { 6481 spa_unload(spa); 6482 spa_deactivate(spa); 6483 } 6484 6485 if (oldconfig && spa->spa_config) 6486 *oldconfig = fnvlist_dup(spa->spa_config); 6487 6488 if (new_state != POOL_STATE_UNINITIALIZED) { 6489 if (!hardforce) 6490 spa_write_cachefile(spa, B_TRUE, B_TRUE); 6491 spa_remove(spa); 6492 } else { 6493 /* 6494 * If spa_remove() is not called for this spa_t and 6495 * there is any possibility that it can be reused, 6496 * we make sure to reset the exporting flag. 6497 */ 6498 spa->spa_is_exporting = B_FALSE; 6499 } 6500 6501 mutex_exit(&spa_namespace_lock); 6502 return (0); 6503 6504 fail: 6505 spa->spa_is_exporting = B_FALSE; 6506 spa_async_resume(spa); 6507 mutex_exit(&spa_namespace_lock); 6508 return (error); 6509 } 6510 6511 /* 6512 * Destroy a storage pool. 6513 */ 6514 int 6515 spa_destroy(const char *pool) 6516 { 6517 return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL, 6518 B_FALSE, B_FALSE)); 6519 } 6520 6521 /* 6522 * Export a storage pool. 6523 */ 6524 int 6525 spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force, 6526 boolean_t hardforce) 6527 { 6528 return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig, 6529 force, hardforce)); 6530 } 6531 6532 /* 6533 * Similar to spa_export(), this unloads the spa_t without actually removing it 6534 * from the namespace in any way. 6535 */ 6536 int 6537 spa_reset(const char *pool) 6538 { 6539 return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL, 6540 B_FALSE, B_FALSE)); 6541 } 6542 6543 /* 6544 * ========================================================================== 6545 * Device manipulation 6546 * ========================================================================== 6547 */ 6548 6549 /* 6550 * This is called as a synctask to increment the draid feature flag 6551 */ 6552 static void 6553 spa_draid_feature_incr(void *arg, dmu_tx_t *tx) 6554 { 6555 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 6556 int draid = (int)(uintptr_t)arg; 6557 6558 for (int c = 0; c < draid; c++) 6559 spa_feature_incr(spa, SPA_FEATURE_DRAID, tx); 6560 } 6561 6562 /* 6563 * Add a device to a storage pool. 6564 */ 6565 int 6566 spa_vdev_add(spa_t *spa, nvlist_t *nvroot) 6567 { 6568 uint64_t txg, ndraid = 0; 6569 int error; 6570 vdev_t *rvd = spa->spa_root_vdev; 6571 vdev_t *vd, *tvd; 6572 nvlist_t **spares, **l2cache; 6573 uint_t nspares, nl2cache; 6574 6575 ASSERT(spa_writeable(spa)); 6576 6577 txg = spa_vdev_enter(spa); 6578 6579 if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0, 6580 VDEV_ALLOC_ADD)) != 0) 6581 return (spa_vdev_exit(spa, NULL, txg, error)); 6582 6583 spa->spa_pending_vdev = vd; /* spa_vdev_exit() will clear this */ 6584 6585 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, 6586 &nspares) != 0) 6587 nspares = 0; 6588 6589 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache, 6590 &nl2cache) != 0) 6591 nl2cache = 0; 6592 6593 if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0) 6594 return (spa_vdev_exit(spa, vd, txg, EINVAL)); 6595 6596 if (vd->vdev_children != 0 && 6597 (error = vdev_create(vd, txg, B_FALSE)) != 0) { 6598 return (spa_vdev_exit(spa, vd, txg, error)); 6599 } 6600 6601 /* 6602 * The virtual dRAID spares must be added after vdev tree is created 6603 * and the vdev guids are generated. The guid of their associated 6604 * dRAID is stored in the config and used when opening the spare. 6605 */ 6606 if ((error = vdev_draid_spare_create(nvroot, vd, &ndraid, 6607 rvd->vdev_children)) == 0) { 6608 if (ndraid > 0 && nvlist_lookup_nvlist_array(nvroot, 6609 ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0) 6610 nspares = 0; 6611 } else { 6612 return (spa_vdev_exit(spa, vd, txg, error)); 6613 } 6614 6615 /* 6616 * We must validate the spares and l2cache devices after checking the 6617 * children. Otherwise, vdev_inuse() will blindly overwrite the spare. 6618 */ 6619 if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0) 6620 return (spa_vdev_exit(spa, vd, txg, error)); 6621 6622 /* 6623 * If we are in the middle of a device removal, we can only add 6624 * devices which match the existing devices in the pool. 6625 * If we are in the middle of a removal, or have some indirect 6626 * vdevs, we can not add raidz or dRAID top levels. 6627 */ 6628 if (spa->spa_vdev_removal != NULL || 6629 spa->spa_removing_phys.sr_prev_indirect_vdev != -1) { 6630 for (int c = 0; c < vd->vdev_children; c++) { 6631 tvd = vd->vdev_child[c]; 6632 if (spa->spa_vdev_removal != NULL && 6633 tvd->vdev_ashift != spa->spa_max_ashift) { 6634 return (spa_vdev_exit(spa, vd, txg, EINVAL)); 6635 } 6636 /* Fail if top level vdev is raidz or a dRAID */ 6637 if (vdev_get_nparity(tvd) != 0) 6638 return (spa_vdev_exit(spa, vd, txg, EINVAL)); 6639 6640 /* 6641 * Need the top level mirror to be 6642 * a mirror of leaf vdevs only 6643 */ 6644 if (tvd->vdev_ops == &vdev_mirror_ops) { 6645 for (uint64_t cid = 0; 6646 cid < tvd->vdev_children; cid++) { 6647 vdev_t *cvd = tvd->vdev_child[cid]; 6648 if (!cvd->vdev_ops->vdev_op_leaf) { 6649 return (spa_vdev_exit(spa, vd, 6650 txg, EINVAL)); 6651 } 6652 } 6653 } 6654 } 6655 } 6656 6657 for (int c = 0; c < vd->vdev_children; c++) { 6658 tvd = vd->vdev_child[c]; 6659 vdev_remove_child(vd, tvd); 6660 tvd->vdev_id = rvd->vdev_children; 6661 vdev_add_child(rvd, tvd); 6662 vdev_config_dirty(tvd); 6663 } 6664 6665 if (nspares != 0) { 6666 spa_set_aux_vdevs(&spa->spa_spares, spares, nspares, 6667 ZPOOL_CONFIG_SPARES); 6668 spa_load_spares(spa); 6669 spa->spa_spares.sav_sync = B_TRUE; 6670 } 6671 6672 if (nl2cache != 0) { 6673 spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache, 6674 ZPOOL_CONFIG_L2CACHE); 6675 spa_load_l2cache(spa); 6676 spa->spa_l2cache.sav_sync = B_TRUE; 6677 } 6678 6679 /* 6680 * We can't increment a feature while holding spa_vdev so we 6681 * have to do it in a synctask. 6682 */ 6683 if (ndraid != 0) { 6684 dmu_tx_t *tx; 6685 6686 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 6687 dsl_sync_task_nowait(spa->spa_dsl_pool, spa_draid_feature_incr, 6688 (void *)(uintptr_t)ndraid, tx); 6689 dmu_tx_commit(tx); 6690 } 6691 6692 /* 6693 * We have to be careful when adding new vdevs to an existing pool. 6694 * If other threads start allocating from these vdevs before we 6695 * sync the config cache, and we lose power, then upon reboot we may 6696 * fail to open the pool because there are DVAs that the config cache 6697 * can't translate. Therefore, we first add the vdevs without 6698 * initializing metaslabs; sync the config cache (via spa_vdev_exit()); 6699 * and then let spa_config_update() initialize the new metaslabs. 6700 * 6701 * spa_load() checks for added-but-not-initialized vdevs, so that 6702 * if we lose power at any point in this sequence, the remaining 6703 * steps will be completed the next time we load the pool. 6704 */ 6705 (void) spa_vdev_exit(spa, vd, txg, 0); 6706 6707 mutex_enter(&spa_namespace_lock); 6708 spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); 6709 spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD); 6710 mutex_exit(&spa_namespace_lock); 6711 6712 return (0); 6713 } 6714 6715 /* 6716 * Attach a device to a mirror. The arguments are the path to any device 6717 * in the mirror, and the nvroot for the new device. If the path specifies 6718 * a device that is not mirrored, we automatically insert the mirror vdev. 6719 * 6720 * If 'replacing' is specified, the new device is intended to replace the 6721 * existing device; in this case the two devices are made into their own 6722 * mirror using the 'replacing' vdev, which is functionally identical to 6723 * the mirror vdev (it actually reuses all the same ops) but has a few 6724 * extra rules: you can't attach to it after it's been created, and upon 6725 * completion of resilvering, the first disk (the one being replaced) 6726 * is automatically detached. 6727 * 6728 * If 'rebuild' is specified, then sequential reconstruction (a.ka. rebuild) 6729 * should be performed instead of traditional healing reconstruction. From 6730 * an administrators perspective these are both resilver operations. 6731 */ 6732 int 6733 spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing, 6734 int rebuild) 6735 { 6736 uint64_t txg, dtl_max_txg; 6737 vdev_t *rvd = spa->spa_root_vdev; 6738 vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd; 6739 vdev_ops_t *pvops; 6740 char *oldvdpath, *newvdpath; 6741 int newvd_isspare; 6742 int error; 6743 6744 ASSERT(spa_writeable(spa)); 6745 6746 txg = spa_vdev_enter(spa); 6747 6748 oldvd = spa_lookup_by_guid(spa, guid, B_FALSE); 6749 6750 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 6751 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 6752 error = (spa_has_checkpoint(spa)) ? 6753 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 6754 return (spa_vdev_exit(spa, NULL, txg, error)); 6755 } 6756 6757 if (rebuild) { 6758 if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REBUILD)) 6759 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6760 6761 if (dsl_scan_resilvering(spa_get_dsl(spa))) 6762 return (spa_vdev_exit(spa, NULL, txg, 6763 ZFS_ERR_RESILVER_IN_PROGRESS)); 6764 } else { 6765 if (vdev_rebuild_active(rvd)) 6766 return (spa_vdev_exit(spa, NULL, txg, 6767 ZFS_ERR_REBUILD_IN_PROGRESS)); 6768 } 6769 6770 if (spa->spa_vdev_removal != NULL) 6771 return (spa_vdev_exit(spa, NULL, txg, EBUSY)); 6772 6773 if (oldvd == NULL) 6774 return (spa_vdev_exit(spa, NULL, txg, ENODEV)); 6775 6776 if (!oldvd->vdev_ops->vdev_op_leaf) 6777 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6778 6779 pvd = oldvd->vdev_parent; 6780 6781 if ((error = spa_config_parse(spa, &newrootvd, nvroot, NULL, 0, 6782 VDEV_ALLOC_ATTACH)) != 0) 6783 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 6784 6785 if (newrootvd->vdev_children != 1) 6786 return (spa_vdev_exit(spa, newrootvd, txg, EINVAL)); 6787 6788 newvd = newrootvd->vdev_child[0]; 6789 6790 if (!newvd->vdev_ops->vdev_op_leaf) 6791 return (spa_vdev_exit(spa, newrootvd, txg, EINVAL)); 6792 6793 if ((error = vdev_create(newrootvd, txg, replacing)) != 0) 6794 return (spa_vdev_exit(spa, newrootvd, txg, error)); 6795 6796 /* 6797 * Spares can't replace logs 6798 */ 6799 if (oldvd->vdev_top->vdev_islog && newvd->vdev_isspare) 6800 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6801 6802 /* 6803 * A dRAID spare can only replace a child of its parent dRAID vdev. 6804 */ 6805 if (newvd->vdev_ops == &vdev_draid_spare_ops && 6806 oldvd->vdev_top != vdev_draid_spare_get_parent(newvd)) { 6807 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6808 } 6809 6810 if (rebuild) { 6811 /* 6812 * For rebuilds, the top vdev must support reconstruction 6813 * using only space maps. This means the only allowable 6814 * vdevs types are the root vdev, a mirror, or dRAID. 6815 */ 6816 tvd = pvd; 6817 if (pvd->vdev_top != NULL) 6818 tvd = pvd->vdev_top; 6819 6820 if (tvd->vdev_ops != &vdev_mirror_ops && 6821 tvd->vdev_ops != &vdev_root_ops && 6822 tvd->vdev_ops != &vdev_draid_ops) { 6823 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6824 } 6825 } 6826 6827 if (!replacing) { 6828 /* 6829 * For attach, the only allowable parent is a mirror or the root 6830 * vdev. 6831 */ 6832 if (pvd->vdev_ops != &vdev_mirror_ops && 6833 pvd->vdev_ops != &vdev_root_ops) 6834 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6835 6836 pvops = &vdev_mirror_ops; 6837 } else { 6838 /* 6839 * Active hot spares can only be replaced by inactive hot 6840 * spares. 6841 */ 6842 if (pvd->vdev_ops == &vdev_spare_ops && 6843 oldvd->vdev_isspare && 6844 !spa_has_spare(spa, newvd->vdev_guid)) 6845 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6846 6847 /* 6848 * If the source is a hot spare, and the parent isn't already a 6849 * spare, then we want to create a new hot spare. Otherwise, we 6850 * want to create a replacing vdev. The user is not allowed to 6851 * attach to a spared vdev child unless the 'isspare' state is 6852 * the same (spare replaces spare, non-spare replaces 6853 * non-spare). 6854 */ 6855 if (pvd->vdev_ops == &vdev_replacing_ops && 6856 spa_version(spa) < SPA_VERSION_MULTI_REPLACE) { 6857 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6858 } else if (pvd->vdev_ops == &vdev_spare_ops && 6859 newvd->vdev_isspare != oldvd->vdev_isspare) { 6860 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6861 } 6862 6863 if (newvd->vdev_isspare) 6864 pvops = &vdev_spare_ops; 6865 else 6866 pvops = &vdev_replacing_ops; 6867 } 6868 6869 /* 6870 * Make sure the new device is big enough. 6871 */ 6872 if (newvd->vdev_asize < vdev_get_min_asize(oldvd)) 6873 return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW)); 6874 6875 /* 6876 * The new device cannot have a higher alignment requirement 6877 * than the top-level vdev. 6878 */ 6879 if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift) 6880 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6881 6882 /* 6883 * If this is an in-place replacement, update oldvd's path and devid 6884 * to make it distinguishable from newvd, and unopenable from now on. 6885 */ 6886 if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) { 6887 spa_strfree(oldvd->vdev_path); 6888 oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5, 6889 KM_SLEEP); 6890 (void) snprintf(oldvd->vdev_path, strlen(newvd->vdev_path) + 5, 6891 "%s/%s", newvd->vdev_path, "old"); 6892 if (oldvd->vdev_devid != NULL) { 6893 spa_strfree(oldvd->vdev_devid); 6894 oldvd->vdev_devid = NULL; 6895 } 6896 } 6897 6898 /* 6899 * If the parent is not a mirror, or if we're replacing, insert the new 6900 * mirror/replacing/spare vdev above oldvd. 6901 */ 6902 if (pvd->vdev_ops != pvops) 6903 pvd = vdev_add_parent(oldvd, pvops); 6904 6905 ASSERT(pvd->vdev_top->vdev_parent == rvd); 6906 ASSERT(pvd->vdev_ops == pvops); 6907 ASSERT(oldvd->vdev_parent == pvd); 6908 6909 /* 6910 * Extract the new device from its root and add it to pvd. 6911 */ 6912 vdev_remove_child(newrootvd, newvd); 6913 newvd->vdev_id = pvd->vdev_children; 6914 newvd->vdev_crtxg = oldvd->vdev_crtxg; 6915 vdev_add_child(pvd, newvd); 6916 6917 /* 6918 * Reevaluate the parent vdev state. 6919 */ 6920 vdev_propagate_state(pvd); 6921 6922 tvd = newvd->vdev_top; 6923 ASSERT(pvd->vdev_top == tvd); 6924 ASSERT(tvd->vdev_parent == rvd); 6925 6926 vdev_config_dirty(tvd); 6927 6928 /* 6929 * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account 6930 * for any dmu_sync-ed blocks. It will propagate upward when 6931 * spa_vdev_exit() calls vdev_dtl_reassess(). 6932 */ 6933 dtl_max_txg = txg + TXG_CONCURRENT_STATES; 6934 6935 vdev_dtl_dirty(newvd, DTL_MISSING, 6936 TXG_INITIAL, dtl_max_txg - TXG_INITIAL); 6937 6938 if (newvd->vdev_isspare) { 6939 spa_spare_activate(newvd); 6940 spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE); 6941 } 6942 6943 oldvdpath = spa_strdup(oldvd->vdev_path); 6944 newvdpath = spa_strdup(newvd->vdev_path); 6945 newvd_isspare = newvd->vdev_isspare; 6946 6947 /* 6948 * Mark newvd's DTL dirty in this txg. 6949 */ 6950 vdev_dirty(tvd, VDD_DTL, newvd, txg); 6951 6952 /* 6953 * Schedule the resilver or rebuild to restart in the future. We do 6954 * this to ensure that dmu_sync-ed blocks have been stitched into the 6955 * respective datasets. 6956 */ 6957 if (rebuild) { 6958 newvd->vdev_rebuild_txg = txg; 6959 6960 vdev_rebuild(tvd); 6961 } else { 6962 newvd->vdev_resilver_txg = txg; 6963 6964 if (dsl_scan_resilvering(spa_get_dsl(spa)) && 6965 spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) { 6966 vdev_defer_resilver(newvd); 6967 } else { 6968 dsl_scan_restart_resilver(spa->spa_dsl_pool, 6969 dtl_max_txg); 6970 } 6971 } 6972 6973 if (spa->spa_bootfs) 6974 spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH); 6975 6976 spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH); 6977 6978 /* 6979 * Commit the config 6980 */ 6981 (void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0); 6982 6983 spa_history_log_internal(spa, "vdev attach", NULL, 6984 "%s vdev=%s %s vdev=%s", 6985 replacing && newvd_isspare ? "spare in" : 6986 replacing ? "replace" : "attach", newvdpath, 6987 replacing ? "for" : "to", oldvdpath); 6988 6989 spa_strfree(oldvdpath); 6990 spa_strfree(newvdpath); 6991 6992 return (0); 6993 } 6994 6995 /* 6996 * Detach a device from a mirror or replacing vdev. 6997 * 6998 * If 'replace_done' is specified, only detach if the parent 6999 * is a replacing vdev. 7000 */ 7001 int 7002 spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done) 7003 { 7004 uint64_t txg; 7005 int error; 7006 vdev_t *rvd __maybe_unused = spa->spa_root_vdev; 7007 vdev_t *vd, *pvd, *cvd, *tvd; 7008 boolean_t unspare = B_FALSE; 7009 uint64_t unspare_guid = 0; 7010 char *vdpath; 7011 7012 ASSERT(spa_writeable(spa)); 7013 7014 txg = spa_vdev_detach_enter(spa, guid); 7015 7016 vd = spa_lookup_by_guid(spa, guid, B_FALSE); 7017 7018 /* 7019 * Besides being called directly from the userland through the 7020 * ioctl interface, spa_vdev_detach() can be potentially called 7021 * at the end of spa_vdev_resilver_done(). 7022 * 7023 * In the regular case, when we have a checkpoint this shouldn't 7024 * happen as we never empty the DTLs of a vdev during the scrub 7025 * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done() 7026 * should never get here when we have a checkpoint. 7027 * 7028 * That said, even in a case when we checkpoint the pool exactly 7029 * as spa_vdev_resilver_done() calls this function everything 7030 * should be fine as the resilver will return right away. 7031 */ 7032 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7033 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 7034 error = (spa_has_checkpoint(spa)) ? 7035 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 7036 return (spa_vdev_exit(spa, NULL, txg, error)); 7037 } 7038 7039 if (vd == NULL) 7040 return (spa_vdev_exit(spa, NULL, txg, ENODEV)); 7041 7042 if (!vd->vdev_ops->vdev_op_leaf) 7043 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 7044 7045 pvd = vd->vdev_parent; 7046 7047 /* 7048 * If the parent/child relationship is not as expected, don't do it. 7049 * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing 7050 * vdev that's replacing B with C. The user's intent in replacing 7051 * is to go from M(A,B) to M(A,C). If the user decides to cancel 7052 * the replace by detaching C, the expected behavior is to end up 7053 * M(A,B). But suppose that right after deciding to detach C, 7054 * the replacement of B completes. We would have M(A,C), and then 7055 * ask to detach C, which would leave us with just A -- not what 7056 * the user wanted. To prevent this, we make sure that the 7057 * parent/child relationship hasn't changed -- in this example, 7058 * that C's parent is still the replacing vdev R. 7059 */ 7060 if (pvd->vdev_guid != pguid && pguid != 0) 7061 return (spa_vdev_exit(spa, NULL, txg, EBUSY)); 7062 7063 /* 7064 * Only 'replacing' or 'spare' vdevs can be replaced. 7065 */ 7066 if (replace_done && pvd->vdev_ops != &vdev_replacing_ops && 7067 pvd->vdev_ops != &vdev_spare_ops) 7068 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 7069 7070 ASSERT(pvd->vdev_ops != &vdev_spare_ops || 7071 spa_version(spa) >= SPA_VERSION_SPARES); 7072 7073 /* 7074 * Only mirror, replacing, and spare vdevs support detach. 7075 */ 7076 if (pvd->vdev_ops != &vdev_replacing_ops && 7077 pvd->vdev_ops != &vdev_mirror_ops && 7078 pvd->vdev_ops != &vdev_spare_ops) 7079 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 7080 7081 /* 7082 * If this device has the only valid copy of some data, 7083 * we cannot safely detach it. 7084 */ 7085 if (vdev_dtl_required(vd)) 7086 return (spa_vdev_exit(spa, NULL, txg, EBUSY)); 7087 7088 ASSERT(pvd->vdev_children >= 2); 7089 7090 /* 7091 * If we are detaching the second disk from a replacing vdev, then 7092 * check to see if we changed the original vdev's path to have "/old" 7093 * at the end in spa_vdev_attach(). If so, undo that change now. 7094 */ 7095 if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 && 7096 vd->vdev_path != NULL) { 7097 size_t len = strlen(vd->vdev_path); 7098 7099 for (int c = 0; c < pvd->vdev_children; c++) { 7100 cvd = pvd->vdev_child[c]; 7101 7102 if (cvd == vd || cvd->vdev_path == NULL) 7103 continue; 7104 7105 if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 && 7106 strcmp(cvd->vdev_path + len, "/old") == 0) { 7107 spa_strfree(cvd->vdev_path); 7108 cvd->vdev_path = spa_strdup(vd->vdev_path); 7109 break; 7110 } 7111 } 7112 } 7113 7114 /* 7115 * If we are detaching the original disk from a normal spare, then it 7116 * implies that the spare should become a real disk, and be removed 7117 * from the active spare list for the pool. dRAID spares on the 7118 * other hand are coupled to the pool and thus should never be removed 7119 * from the spares list. 7120 */ 7121 if (pvd->vdev_ops == &vdev_spare_ops && vd->vdev_id == 0) { 7122 vdev_t *last_cvd = pvd->vdev_child[pvd->vdev_children - 1]; 7123 7124 if (last_cvd->vdev_isspare && 7125 last_cvd->vdev_ops != &vdev_draid_spare_ops) { 7126 unspare = B_TRUE; 7127 } 7128 } 7129 7130 /* 7131 * Erase the disk labels so the disk can be used for other things. 7132 * This must be done after all other error cases are handled, 7133 * but before we disembowel vd (so we can still do I/O to it). 7134 * But if we can't do it, don't treat the error as fatal -- 7135 * it may be that the unwritability of the disk is the reason 7136 * it's being detached! 7137 */ 7138 error = vdev_label_init(vd, 0, VDEV_LABEL_REMOVE); 7139 7140 /* 7141 * Remove vd from its parent and compact the parent's children. 7142 */ 7143 vdev_remove_child(pvd, vd); 7144 vdev_compact_children(pvd); 7145 7146 /* 7147 * Remember one of the remaining children so we can get tvd below. 7148 */ 7149 cvd = pvd->vdev_child[pvd->vdev_children - 1]; 7150 7151 /* 7152 * If we need to remove the remaining child from the list of hot spares, 7153 * do it now, marking the vdev as no longer a spare in the process. 7154 * We must do this before vdev_remove_parent(), because that can 7155 * change the GUID if it creates a new toplevel GUID. For a similar 7156 * reason, we must remove the spare now, in the same txg as the detach; 7157 * otherwise someone could attach a new sibling, change the GUID, and 7158 * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail. 7159 */ 7160 if (unspare) { 7161 ASSERT(cvd->vdev_isspare); 7162 spa_spare_remove(cvd); 7163 unspare_guid = cvd->vdev_guid; 7164 (void) spa_vdev_remove(spa, unspare_guid, B_TRUE); 7165 cvd->vdev_unspare = B_TRUE; 7166 } 7167 7168 /* 7169 * If the parent mirror/replacing vdev only has one child, 7170 * the parent is no longer needed. Remove it from the tree. 7171 */ 7172 if (pvd->vdev_children == 1) { 7173 if (pvd->vdev_ops == &vdev_spare_ops) 7174 cvd->vdev_unspare = B_FALSE; 7175 vdev_remove_parent(cvd); 7176 } 7177 7178 /* 7179 * We don't set tvd until now because the parent we just removed 7180 * may have been the previous top-level vdev. 7181 */ 7182 tvd = cvd->vdev_top; 7183 ASSERT(tvd->vdev_parent == rvd); 7184 7185 /* 7186 * Reevaluate the parent vdev state. 7187 */ 7188 vdev_propagate_state(cvd); 7189 7190 /* 7191 * If the 'autoexpand' property is set on the pool then automatically 7192 * try to expand the size of the pool. For example if the device we 7193 * just detached was smaller than the others, it may be possible to 7194 * add metaslabs (i.e. grow the pool). We need to reopen the vdev 7195 * first so that we can obtain the updated sizes of the leaf vdevs. 7196 */ 7197 if (spa->spa_autoexpand) { 7198 vdev_reopen(tvd); 7199 vdev_expand(tvd, txg); 7200 } 7201 7202 vdev_config_dirty(tvd); 7203 7204 /* 7205 * Mark vd's DTL as dirty in this txg. vdev_dtl_sync() will see that 7206 * vd->vdev_detached is set and free vd's DTL object in syncing context. 7207 * But first make sure we're not on any *other* txg's DTL list, to 7208 * prevent vd from being accessed after it's freed. 7209 */ 7210 vdpath = spa_strdup(vd->vdev_path ? vd->vdev_path : "none"); 7211 for (int t = 0; t < TXG_SIZE; t++) 7212 (void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t); 7213 vd->vdev_detached = B_TRUE; 7214 vdev_dirty(tvd, VDD_DTL, vd, txg); 7215 7216 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE); 7217 spa_notify_waiters(spa); 7218 7219 /* hang on to the spa before we release the lock */ 7220 spa_open_ref(spa, FTAG); 7221 7222 error = spa_vdev_exit(spa, vd, txg, 0); 7223 7224 spa_history_log_internal(spa, "detach", NULL, 7225 "vdev=%s", vdpath); 7226 spa_strfree(vdpath); 7227 7228 /* 7229 * If this was the removal of the original device in a hot spare vdev, 7230 * then we want to go through and remove the device from the hot spare 7231 * list of every other pool. 7232 */ 7233 if (unspare) { 7234 spa_t *altspa = NULL; 7235 7236 mutex_enter(&spa_namespace_lock); 7237 while ((altspa = spa_next(altspa)) != NULL) { 7238 if (altspa->spa_state != POOL_STATE_ACTIVE || 7239 altspa == spa) 7240 continue; 7241 7242 spa_open_ref(altspa, FTAG); 7243 mutex_exit(&spa_namespace_lock); 7244 (void) spa_vdev_remove(altspa, unspare_guid, B_TRUE); 7245 mutex_enter(&spa_namespace_lock); 7246 spa_close(altspa, FTAG); 7247 } 7248 mutex_exit(&spa_namespace_lock); 7249 7250 /* search the rest of the vdevs for spares to remove */ 7251 spa_vdev_resilver_done(spa); 7252 } 7253 7254 /* all done with the spa; OK to release */ 7255 mutex_enter(&spa_namespace_lock); 7256 spa_close(spa, FTAG); 7257 mutex_exit(&spa_namespace_lock); 7258 7259 return (error); 7260 } 7261 7262 static int 7263 spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type, 7264 list_t *vd_list) 7265 { 7266 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7267 7268 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 7269 7270 /* Look up vdev and ensure it's a leaf. */ 7271 vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE); 7272 if (vd == NULL || vd->vdev_detached) { 7273 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7274 return (SET_ERROR(ENODEV)); 7275 } else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) { 7276 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7277 return (SET_ERROR(EINVAL)); 7278 } else if (!vdev_writeable(vd)) { 7279 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7280 return (SET_ERROR(EROFS)); 7281 } 7282 mutex_enter(&vd->vdev_initialize_lock); 7283 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7284 7285 /* 7286 * When we activate an initialize action we check to see 7287 * if the vdev_initialize_thread is NULL. We do this instead 7288 * of using the vdev_initialize_state since there might be 7289 * a previous initialization process which has completed but 7290 * the thread is not exited. 7291 */ 7292 if (cmd_type == POOL_INITIALIZE_START && 7293 (vd->vdev_initialize_thread != NULL || 7294 vd->vdev_top->vdev_removing)) { 7295 mutex_exit(&vd->vdev_initialize_lock); 7296 return (SET_ERROR(EBUSY)); 7297 } else if (cmd_type == POOL_INITIALIZE_CANCEL && 7298 (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE && 7299 vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) { 7300 mutex_exit(&vd->vdev_initialize_lock); 7301 return (SET_ERROR(ESRCH)); 7302 } else if (cmd_type == POOL_INITIALIZE_SUSPEND && 7303 vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) { 7304 mutex_exit(&vd->vdev_initialize_lock); 7305 return (SET_ERROR(ESRCH)); 7306 } 7307 7308 switch (cmd_type) { 7309 case POOL_INITIALIZE_START: 7310 vdev_initialize(vd); 7311 break; 7312 case POOL_INITIALIZE_CANCEL: 7313 vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, vd_list); 7314 break; 7315 case POOL_INITIALIZE_SUSPEND: 7316 vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED, vd_list); 7317 break; 7318 default: 7319 panic("invalid cmd_type %llu", (unsigned long long)cmd_type); 7320 } 7321 mutex_exit(&vd->vdev_initialize_lock); 7322 7323 return (0); 7324 } 7325 7326 int 7327 spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 7328 nvlist_t *vdev_errlist) 7329 { 7330 int total_errors = 0; 7331 list_t vd_list; 7332 7333 list_create(&vd_list, sizeof (vdev_t), 7334 offsetof(vdev_t, vdev_initialize_node)); 7335 7336 /* 7337 * We hold the namespace lock through the whole function 7338 * to prevent any changes to the pool while we're starting or 7339 * stopping initialization. The config and state locks are held so that 7340 * we can properly assess the vdev state before we commit to 7341 * the initializing operation. 7342 */ 7343 mutex_enter(&spa_namespace_lock); 7344 7345 for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL); 7346 pair != NULL; pair = nvlist_next_nvpair(nv, pair)) { 7347 uint64_t vdev_guid = fnvpair_value_uint64(pair); 7348 7349 int error = spa_vdev_initialize_impl(spa, vdev_guid, cmd_type, 7350 &vd_list); 7351 if (error != 0) { 7352 char guid_as_str[MAXNAMELEN]; 7353 7354 (void) snprintf(guid_as_str, sizeof (guid_as_str), 7355 "%llu", (unsigned long long)vdev_guid); 7356 fnvlist_add_int64(vdev_errlist, guid_as_str, error); 7357 total_errors++; 7358 } 7359 } 7360 7361 /* Wait for all initialize threads to stop. */ 7362 vdev_initialize_stop_wait(spa, &vd_list); 7363 7364 /* Sync out the initializing state */ 7365 txg_wait_synced(spa->spa_dsl_pool, 0); 7366 mutex_exit(&spa_namespace_lock); 7367 7368 list_destroy(&vd_list); 7369 7370 return (total_errors); 7371 } 7372 7373 static int 7374 spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type, 7375 uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list) 7376 { 7377 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7378 7379 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 7380 7381 /* Look up vdev and ensure it's a leaf. */ 7382 vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE); 7383 if (vd == NULL || vd->vdev_detached) { 7384 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7385 return (SET_ERROR(ENODEV)); 7386 } else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) { 7387 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7388 return (SET_ERROR(EINVAL)); 7389 } else if (!vdev_writeable(vd)) { 7390 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7391 return (SET_ERROR(EROFS)); 7392 } else if (!vd->vdev_has_trim) { 7393 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7394 return (SET_ERROR(EOPNOTSUPP)); 7395 } else if (secure && !vd->vdev_has_securetrim) { 7396 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7397 return (SET_ERROR(EOPNOTSUPP)); 7398 } 7399 mutex_enter(&vd->vdev_trim_lock); 7400 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7401 7402 /* 7403 * When we activate a TRIM action we check to see if the 7404 * vdev_trim_thread is NULL. We do this instead of using the 7405 * vdev_trim_state since there might be a previous TRIM process 7406 * which has completed but the thread is not exited. 7407 */ 7408 if (cmd_type == POOL_TRIM_START && 7409 (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) { 7410 mutex_exit(&vd->vdev_trim_lock); 7411 return (SET_ERROR(EBUSY)); 7412 } else if (cmd_type == POOL_TRIM_CANCEL && 7413 (vd->vdev_trim_state != VDEV_TRIM_ACTIVE && 7414 vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) { 7415 mutex_exit(&vd->vdev_trim_lock); 7416 return (SET_ERROR(ESRCH)); 7417 } else if (cmd_type == POOL_TRIM_SUSPEND && 7418 vd->vdev_trim_state != VDEV_TRIM_ACTIVE) { 7419 mutex_exit(&vd->vdev_trim_lock); 7420 return (SET_ERROR(ESRCH)); 7421 } 7422 7423 switch (cmd_type) { 7424 case POOL_TRIM_START: 7425 vdev_trim(vd, rate, partial, secure); 7426 break; 7427 case POOL_TRIM_CANCEL: 7428 vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list); 7429 break; 7430 case POOL_TRIM_SUSPEND: 7431 vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list); 7432 break; 7433 default: 7434 panic("invalid cmd_type %llu", (unsigned long long)cmd_type); 7435 } 7436 mutex_exit(&vd->vdev_trim_lock); 7437 7438 return (0); 7439 } 7440 7441 /* 7442 * Initiates a manual TRIM for the requested vdevs. This kicks off individual 7443 * TRIM threads for each child vdev. These threads pass over all of the free 7444 * space in the vdev's metaslabs and issues TRIM commands for that space. 7445 */ 7446 int 7447 spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate, 7448 boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist) 7449 { 7450 int total_errors = 0; 7451 list_t vd_list; 7452 7453 list_create(&vd_list, sizeof (vdev_t), 7454 offsetof(vdev_t, vdev_trim_node)); 7455 7456 /* 7457 * We hold the namespace lock through the whole function 7458 * to prevent any changes to the pool while we're starting or 7459 * stopping TRIM. The config and state locks are held so that 7460 * we can properly assess the vdev state before we commit to 7461 * the TRIM operation. 7462 */ 7463 mutex_enter(&spa_namespace_lock); 7464 7465 for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL); 7466 pair != NULL; pair = nvlist_next_nvpair(nv, pair)) { 7467 uint64_t vdev_guid = fnvpair_value_uint64(pair); 7468 7469 int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type, 7470 rate, partial, secure, &vd_list); 7471 if (error != 0) { 7472 char guid_as_str[MAXNAMELEN]; 7473 7474 (void) snprintf(guid_as_str, sizeof (guid_as_str), 7475 "%llu", (unsigned long long)vdev_guid); 7476 fnvlist_add_int64(vdev_errlist, guid_as_str, error); 7477 total_errors++; 7478 } 7479 } 7480 7481 /* Wait for all TRIM threads to stop. */ 7482 vdev_trim_stop_wait(spa, &vd_list); 7483 7484 /* Sync out the TRIM state */ 7485 txg_wait_synced(spa->spa_dsl_pool, 0); 7486 mutex_exit(&spa_namespace_lock); 7487 7488 list_destroy(&vd_list); 7489 7490 return (total_errors); 7491 } 7492 7493 /* 7494 * Split a set of devices from their mirrors, and create a new pool from them. 7495 */ 7496 int 7497 spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config, 7498 nvlist_t *props, boolean_t exp) 7499 { 7500 int error = 0; 7501 uint64_t txg, *glist; 7502 spa_t *newspa; 7503 uint_t c, children, lastlog; 7504 nvlist_t **child, *nvl, *tmp; 7505 dmu_tx_t *tx; 7506 char *altroot = NULL; 7507 vdev_t *rvd, **vml = NULL; /* vdev modify list */ 7508 boolean_t activate_slog; 7509 7510 ASSERT(spa_writeable(spa)); 7511 7512 txg = spa_vdev_enter(spa); 7513 7514 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7515 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 7516 error = (spa_has_checkpoint(spa)) ? 7517 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 7518 return (spa_vdev_exit(spa, NULL, txg, error)); 7519 } 7520 7521 /* clear the log and flush everything up to now */ 7522 activate_slog = spa_passivate_log(spa); 7523 (void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG); 7524 error = spa_reset_logs(spa); 7525 txg = spa_vdev_config_enter(spa); 7526 7527 if (activate_slog) 7528 spa_activate_log(spa); 7529 7530 if (error != 0) 7531 return (spa_vdev_exit(spa, NULL, txg, error)); 7532 7533 /* check new spa name before going any further */ 7534 if (spa_lookup(newname) != NULL) 7535 return (spa_vdev_exit(spa, NULL, txg, EEXIST)); 7536 7537 /* 7538 * scan through all the children to ensure they're all mirrors 7539 */ 7540 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 || 7541 nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child, 7542 &children) != 0) 7543 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 7544 7545 /* first, check to ensure we've got the right child count */ 7546 rvd = spa->spa_root_vdev; 7547 lastlog = 0; 7548 for (c = 0; c < rvd->vdev_children; c++) { 7549 vdev_t *vd = rvd->vdev_child[c]; 7550 7551 /* don't count the holes & logs as children */ 7552 if (vd->vdev_islog || (vd->vdev_ops != &vdev_indirect_ops && 7553 !vdev_is_concrete(vd))) { 7554 if (lastlog == 0) 7555 lastlog = c; 7556 continue; 7557 } 7558 7559 lastlog = 0; 7560 } 7561 if (children != (lastlog != 0 ? lastlog : rvd->vdev_children)) 7562 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 7563 7564 /* next, ensure no spare or cache devices are part of the split */ 7565 if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 || 7566 nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0) 7567 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 7568 7569 vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP); 7570 glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP); 7571 7572 /* then, loop over each vdev and validate it */ 7573 for (c = 0; c < children; c++) { 7574 uint64_t is_hole = 0; 7575 7576 (void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE, 7577 &is_hole); 7578 7579 if (is_hole != 0) { 7580 if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole || 7581 spa->spa_root_vdev->vdev_child[c]->vdev_islog) { 7582 continue; 7583 } else { 7584 error = SET_ERROR(EINVAL); 7585 break; 7586 } 7587 } 7588 7589 /* deal with indirect vdevs */ 7590 if (spa->spa_root_vdev->vdev_child[c]->vdev_ops == 7591 &vdev_indirect_ops) 7592 continue; 7593 7594 /* which disk is going to be split? */ 7595 if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID, 7596 &glist[c]) != 0) { 7597 error = SET_ERROR(EINVAL); 7598 break; 7599 } 7600 7601 /* look it up in the spa */ 7602 vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE); 7603 if (vml[c] == NULL) { 7604 error = SET_ERROR(ENODEV); 7605 break; 7606 } 7607 7608 /* make sure there's nothing stopping the split */ 7609 if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops || 7610 vml[c]->vdev_islog || 7611 !vdev_is_concrete(vml[c]) || 7612 vml[c]->vdev_isspare || 7613 vml[c]->vdev_isl2cache || 7614 !vdev_writeable(vml[c]) || 7615 vml[c]->vdev_children != 0 || 7616 vml[c]->vdev_state != VDEV_STATE_HEALTHY || 7617 c != spa->spa_root_vdev->vdev_child[c]->vdev_id) { 7618 error = SET_ERROR(EINVAL); 7619 break; 7620 } 7621 7622 if (vdev_dtl_required(vml[c]) || 7623 vdev_resilver_needed(vml[c], NULL, NULL)) { 7624 error = SET_ERROR(EBUSY); 7625 break; 7626 } 7627 7628 /* we need certain info from the top level */ 7629 fnvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY, 7630 vml[c]->vdev_top->vdev_ms_array); 7631 fnvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT, 7632 vml[c]->vdev_top->vdev_ms_shift); 7633 fnvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE, 7634 vml[c]->vdev_top->vdev_asize); 7635 fnvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT, 7636 vml[c]->vdev_top->vdev_ashift); 7637 7638 /* transfer per-vdev ZAPs */ 7639 ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0); 7640 VERIFY0(nvlist_add_uint64(child[c], 7641 ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap)); 7642 7643 ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0); 7644 VERIFY0(nvlist_add_uint64(child[c], 7645 ZPOOL_CONFIG_VDEV_TOP_ZAP, 7646 vml[c]->vdev_parent->vdev_top_zap)); 7647 } 7648 7649 if (error != 0) { 7650 kmem_free(vml, children * sizeof (vdev_t *)); 7651 kmem_free(glist, children * sizeof (uint64_t)); 7652 return (spa_vdev_exit(spa, NULL, txg, error)); 7653 } 7654 7655 /* stop writers from using the disks */ 7656 for (c = 0; c < children; c++) { 7657 if (vml[c] != NULL) 7658 vml[c]->vdev_offline = B_TRUE; 7659 } 7660 vdev_reopen(spa->spa_root_vdev); 7661 7662 /* 7663 * Temporarily record the splitting vdevs in the spa config. This 7664 * will disappear once the config is regenerated. 7665 */ 7666 nvl = fnvlist_alloc(); 7667 fnvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, glist, children); 7668 kmem_free(glist, children * sizeof (uint64_t)); 7669 7670 mutex_enter(&spa->spa_props_lock); 7671 fnvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT, nvl); 7672 mutex_exit(&spa->spa_props_lock); 7673 spa->spa_config_splitting = nvl; 7674 vdev_config_dirty(spa->spa_root_vdev); 7675 7676 /* configure and create the new pool */ 7677 fnvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname); 7678 fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, 7679 exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE); 7680 fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, spa_version(spa)); 7681 fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG, spa->spa_config_txg); 7682 fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID, 7683 spa_generate_guid(NULL)); 7684 VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)); 7685 (void) nvlist_lookup_string(props, 7686 zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); 7687 7688 /* add the new pool to the namespace */ 7689 newspa = spa_add(newname, config, altroot); 7690 newspa->spa_avz_action = AVZ_ACTION_REBUILD; 7691 newspa->spa_config_txg = spa->spa_config_txg; 7692 spa_set_log_state(newspa, SPA_LOG_CLEAR); 7693 7694 /* release the spa config lock, retaining the namespace lock */ 7695 spa_vdev_config_exit(spa, NULL, txg, 0, FTAG); 7696 7697 if (zio_injection_enabled) 7698 zio_handle_panic_injection(spa, FTAG, 1); 7699 7700 spa_activate(newspa, spa_mode_global); 7701 spa_async_suspend(newspa); 7702 7703 /* 7704 * Temporarily stop the initializing and TRIM activity. We set the 7705 * state to ACTIVE so that we know to resume initializing or TRIM 7706 * once the split has completed. 7707 */ 7708 list_t vd_initialize_list; 7709 list_create(&vd_initialize_list, sizeof (vdev_t), 7710 offsetof(vdev_t, vdev_initialize_node)); 7711 7712 list_t vd_trim_list; 7713 list_create(&vd_trim_list, sizeof (vdev_t), 7714 offsetof(vdev_t, vdev_trim_node)); 7715 7716 for (c = 0; c < children; c++) { 7717 if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) { 7718 mutex_enter(&vml[c]->vdev_initialize_lock); 7719 vdev_initialize_stop(vml[c], 7720 VDEV_INITIALIZE_ACTIVE, &vd_initialize_list); 7721 mutex_exit(&vml[c]->vdev_initialize_lock); 7722 7723 mutex_enter(&vml[c]->vdev_trim_lock); 7724 vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list); 7725 mutex_exit(&vml[c]->vdev_trim_lock); 7726 } 7727 } 7728 7729 vdev_initialize_stop_wait(spa, &vd_initialize_list); 7730 vdev_trim_stop_wait(spa, &vd_trim_list); 7731 7732 list_destroy(&vd_initialize_list); 7733 list_destroy(&vd_trim_list); 7734 7735 newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT; 7736 newspa->spa_is_splitting = B_TRUE; 7737 7738 /* create the new pool from the disks of the original pool */ 7739 error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE); 7740 if (error) 7741 goto out; 7742 7743 /* if that worked, generate a real config for the new pool */ 7744 if (newspa->spa_root_vdev != NULL) { 7745 newspa->spa_config_splitting = fnvlist_alloc(); 7746 fnvlist_add_uint64(newspa->spa_config_splitting, 7747 ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa)); 7748 spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL, 7749 B_TRUE)); 7750 } 7751 7752 /* set the props */ 7753 if (props != NULL) { 7754 spa_configfile_set(newspa, props, B_FALSE); 7755 error = spa_prop_set(newspa, props); 7756 if (error) 7757 goto out; 7758 } 7759 7760 /* flush everything */ 7761 txg = spa_vdev_config_enter(newspa); 7762 vdev_config_dirty(newspa->spa_root_vdev); 7763 (void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG); 7764 7765 if (zio_injection_enabled) 7766 zio_handle_panic_injection(spa, FTAG, 2); 7767 7768 spa_async_resume(newspa); 7769 7770 /* finally, update the original pool's config */ 7771 txg = spa_vdev_config_enter(spa); 7772 tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 7773 error = dmu_tx_assign(tx, TXG_WAIT); 7774 if (error != 0) 7775 dmu_tx_abort(tx); 7776 for (c = 0; c < children; c++) { 7777 if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) { 7778 vdev_t *tvd = vml[c]->vdev_top; 7779 7780 /* 7781 * Need to be sure the detachable VDEV is not 7782 * on any *other* txg's DTL list to prevent it 7783 * from being accessed after it's freed. 7784 */ 7785 for (int t = 0; t < TXG_SIZE; t++) { 7786 (void) txg_list_remove_this( 7787 &tvd->vdev_dtl_list, vml[c], t); 7788 } 7789 7790 vdev_split(vml[c]); 7791 if (error == 0) 7792 spa_history_log_internal(spa, "detach", tx, 7793 "vdev=%s", vml[c]->vdev_path); 7794 7795 vdev_free(vml[c]); 7796 } 7797 } 7798 spa->spa_avz_action = AVZ_ACTION_REBUILD; 7799 vdev_config_dirty(spa->spa_root_vdev); 7800 spa->spa_config_splitting = NULL; 7801 nvlist_free(nvl); 7802 if (error == 0) 7803 dmu_tx_commit(tx); 7804 (void) spa_vdev_exit(spa, NULL, txg, 0); 7805 7806 if (zio_injection_enabled) 7807 zio_handle_panic_injection(spa, FTAG, 3); 7808 7809 /* split is complete; log a history record */ 7810 spa_history_log_internal(newspa, "split", NULL, 7811 "from pool %s", spa_name(spa)); 7812 7813 newspa->spa_is_splitting = B_FALSE; 7814 kmem_free(vml, children * sizeof (vdev_t *)); 7815 7816 /* if we're not going to mount the filesystems in userland, export */ 7817 if (exp) 7818 error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL, 7819 B_FALSE, B_FALSE); 7820 7821 return (error); 7822 7823 out: 7824 spa_unload(newspa); 7825 spa_deactivate(newspa); 7826 spa_remove(newspa); 7827 7828 txg = spa_vdev_config_enter(spa); 7829 7830 /* re-online all offlined disks */ 7831 for (c = 0; c < children; c++) { 7832 if (vml[c] != NULL) 7833 vml[c]->vdev_offline = B_FALSE; 7834 } 7835 7836 /* restart initializing or trimming disks as necessary */ 7837 spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART); 7838 spa_async_request(spa, SPA_ASYNC_TRIM_RESTART); 7839 spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART); 7840 7841 vdev_reopen(spa->spa_root_vdev); 7842 7843 nvlist_free(spa->spa_config_splitting); 7844 spa->spa_config_splitting = NULL; 7845 (void) spa_vdev_exit(spa, NULL, txg, error); 7846 7847 kmem_free(vml, children * sizeof (vdev_t *)); 7848 return (error); 7849 } 7850 7851 /* 7852 * Find any device that's done replacing, or a vdev marked 'unspare' that's 7853 * currently spared, so we can detach it. 7854 */ 7855 static vdev_t * 7856 spa_vdev_resilver_done_hunt(vdev_t *vd) 7857 { 7858 vdev_t *newvd, *oldvd; 7859 7860 for (int c = 0; c < vd->vdev_children; c++) { 7861 oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]); 7862 if (oldvd != NULL) 7863 return (oldvd); 7864 } 7865 7866 /* 7867 * Check for a completed replacement. We always consider the first 7868 * vdev in the list to be the oldest vdev, and the last one to be 7869 * the newest (see spa_vdev_attach() for how that works). In 7870 * the case where the newest vdev is faulted, we will not automatically 7871 * remove it after a resilver completes. This is OK as it will require 7872 * user intervention to determine which disk the admin wishes to keep. 7873 */ 7874 if (vd->vdev_ops == &vdev_replacing_ops) { 7875 ASSERT(vd->vdev_children > 1); 7876 7877 newvd = vd->vdev_child[vd->vdev_children - 1]; 7878 oldvd = vd->vdev_child[0]; 7879 7880 if (vdev_dtl_empty(newvd, DTL_MISSING) && 7881 vdev_dtl_empty(newvd, DTL_OUTAGE) && 7882 !vdev_dtl_required(oldvd)) 7883 return (oldvd); 7884 } 7885 7886 /* 7887 * Check for a completed resilver with the 'unspare' flag set. 7888 * Also potentially update faulted state. 7889 */ 7890 if (vd->vdev_ops == &vdev_spare_ops) { 7891 vdev_t *first = vd->vdev_child[0]; 7892 vdev_t *last = vd->vdev_child[vd->vdev_children - 1]; 7893 7894 if (last->vdev_unspare) { 7895 oldvd = first; 7896 newvd = last; 7897 } else if (first->vdev_unspare) { 7898 oldvd = last; 7899 newvd = first; 7900 } else { 7901 oldvd = NULL; 7902 } 7903 7904 if (oldvd != NULL && 7905 vdev_dtl_empty(newvd, DTL_MISSING) && 7906 vdev_dtl_empty(newvd, DTL_OUTAGE) && 7907 !vdev_dtl_required(oldvd)) 7908 return (oldvd); 7909 7910 vdev_propagate_state(vd); 7911 7912 /* 7913 * If there are more than two spares attached to a disk, 7914 * and those spares are not required, then we want to 7915 * attempt to free them up now so that they can be used 7916 * by other pools. Once we're back down to a single 7917 * disk+spare, we stop removing them. 7918 */ 7919 if (vd->vdev_children > 2) { 7920 newvd = vd->vdev_child[1]; 7921 7922 if (newvd->vdev_isspare && last->vdev_isspare && 7923 vdev_dtl_empty(last, DTL_MISSING) && 7924 vdev_dtl_empty(last, DTL_OUTAGE) && 7925 !vdev_dtl_required(newvd)) 7926 return (newvd); 7927 } 7928 } 7929 7930 return (NULL); 7931 } 7932 7933 static void 7934 spa_vdev_resilver_done(spa_t *spa) 7935 { 7936 vdev_t *vd, *pvd, *ppvd; 7937 uint64_t guid, sguid, pguid, ppguid; 7938 7939 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 7940 7941 while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) { 7942 pvd = vd->vdev_parent; 7943 ppvd = pvd->vdev_parent; 7944 guid = vd->vdev_guid; 7945 pguid = pvd->vdev_guid; 7946 ppguid = ppvd->vdev_guid; 7947 sguid = 0; 7948 /* 7949 * If we have just finished replacing a hot spared device, then 7950 * we need to detach the parent's first child (the original hot 7951 * spare) as well. 7952 */ 7953 if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 && 7954 ppvd->vdev_children == 2) { 7955 ASSERT(pvd->vdev_ops == &vdev_replacing_ops); 7956 sguid = ppvd->vdev_child[1]->vdev_guid; 7957 } 7958 ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd)); 7959 7960 spa_config_exit(spa, SCL_ALL, FTAG); 7961 if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0) 7962 return; 7963 if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0) 7964 return; 7965 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 7966 } 7967 7968 spa_config_exit(spa, SCL_ALL, FTAG); 7969 7970 /* 7971 * If a detach was not performed above replace waiters will not have 7972 * been notified. In which case we must do so now. 7973 */ 7974 spa_notify_waiters(spa); 7975 } 7976 7977 /* 7978 * Update the stored path or FRU for this vdev. 7979 */ 7980 static int 7981 spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value, 7982 boolean_t ispath) 7983 { 7984 vdev_t *vd; 7985 boolean_t sync = B_FALSE; 7986 7987 ASSERT(spa_writeable(spa)); 7988 7989 spa_vdev_state_enter(spa, SCL_ALL); 7990 7991 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) 7992 return (spa_vdev_state_exit(spa, NULL, ENOENT)); 7993 7994 if (!vd->vdev_ops->vdev_op_leaf) 7995 return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); 7996 7997 if (ispath) { 7998 if (strcmp(value, vd->vdev_path) != 0) { 7999 spa_strfree(vd->vdev_path); 8000 vd->vdev_path = spa_strdup(value); 8001 sync = B_TRUE; 8002 } 8003 } else { 8004 if (vd->vdev_fru == NULL) { 8005 vd->vdev_fru = spa_strdup(value); 8006 sync = B_TRUE; 8007 } else if (strcmp(value, vd->vdev_fru) != 0) { 8008 spa_strfree(vd->vdev_fru); 8009 vd->vdev_fru = spa_strdup(value); 8010 sync = B_TRUE; 8011 } 8012 } 8013 8014 return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0)); 8015 } 8016 8017 int 8018 spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath) 8019 { 8020 return (spa_vdev_set_common(spa, guid, newpath, B_TRUE)); 8021 } 8022 8023 int 8024 spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru) 8025 { 8026 return (spa_vdev_set_common(spa, guid, newfru, B_FALSE)); 8027 } 8028 8029 /* 8030 * ========================================================================== 8031 * SPA Scanning 8032 * ========================================================================== 8033 */ 8034 int 8035 spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd) 8036 { 8037 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); 8038 8039 if (dsl_scan_resilvering(spa->spa_dsl_pool)) 8040 return (SET_ERROR(EBUSY)); 8041 8042 return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd)); 8043 } 8044 8045 int 8046 spa_scan_stop(spa_t *spa) 8047 { 8048 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); 8049 if (dsl_scan_resilvering(spa->spa_dsl_pool)) 8050 return (SET_ERROR(EBUSY)); 8051 return (dsl_scan_cancel(spa->spa_dsl_pool)); 8052 } 8053 8054 int 8055 spa_scan(spa_t *spa, pool_scan_func_t func) 8056 { 8057 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); 8058 8059 if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE) 8060 return (SET_ERROR(ENOTSUP)); 8061 8062 if (func == POOL_SCAN_RESILVER && 8063 !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) 8064 return (SET_ERROR(ENOTSUP)); 8065 8066 /* 8067 * If a resilver was requested, but there is no DTL on a 8068 * writeable leaf device, we have nothing to do. 8069 */ 8070 if (func == POOL_SCAN_RESILVER && 8071 !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) { 8072 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); 8073 return (0); 8074 } 8075 8076 return (dsl_scan(spa->spa_dsl_pool, func)); 8077 } 8078 8079 /* 8080 * ========================================================================== 8081 * SPA async task processing 8082 * ========================================================================== 8083 */ 8084 8085 static void 8086 spa_async_remove(spa_t *spa, vdev_t *vd) 8087 { 8088 if (vd->vdev_remove_wanted) { 8089 vd->vdev_remove_wanted = B_FALSE; 8090 vd->vdev_delayed_close = B_FALSE; 8091 vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE); 8092 8093 /* 8094 * We want to clear the stats, but we don't want to do a full 8095 * vdev_clear() as that will cause us to throw away 8096 * degraded/faulted state as well as attempt to reopen the 8097 * device, all of which is a waste. 8098 */ 8099 vd->vdev_stat.vs_read_errors = 0; 8100 vd->vdev_stat.vs_write_errors = 0; 8101 vd->vdev_stat.vs_checksum_errors = 0; 8102 8103 vdev_state_dirty(vd->vdev_top); 8104 8105 /* Tell userspace that the vdev is gone. */ 8106 zfs_post_remove(spa, vd); 8107 } 8108 8109 for (int c = 0; c < vd->vdev_children; c++) 8110 spa_async_remove(spa, vd->vdev_child[c]); 8111 } 8112 8113 static void 8114 spa_async_probe(spa_t *spa, vdev_t *vd) 8115 { 8116 if (vd->vdev_probe_wanted) { 8117 vd->vdev_probe_wanted = B_FALSE; 8118 vdev_reopen(vd); /* vdev_open() does the actual probe */ 8119 } 8120 8121 for (int c = 0; c < vd->vdev_children; c++) 8122 spa_async_probe(spa, vd->vdev_child[c]); 8123 } 8124 8125 static void 8126 spa_async_autoexpand(spa_t *spa, vdev_t *vd) 8127 { 8128 if (!spa->spa_autoexpand) 8129 return; 8130 8131 for (int c = 0; c < vd->vdev_children; c++) { 8132 vdev_t *cvd = vd->vdev_child[c]; 8133 spa_async_autoexpand(spa, cvd); 8134 } 8135 8136 if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL) 8137 return; 8138 8139 spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_AUTOEXPAND); 8140 } 8141 8142 static __attribute__((noreturn)) void 8143 spa_async_thread(void *arg) 8144 { 8145 spa_t *spa = (spa_t *)arg; 8146 dsl_pool_t *dp = spa->spa_dsl_pool; 8147 int tasks; 8148 8149 ASSERT(spa->spa_sync_on); 8150 8151 mutex_enter(&spa->spa_async_lock); 8152 tasks = spa->spa_async_tasks; 8153 spa->spa_async_tasks = 0; 8154 mutex_exit(&spa->spa_async_lock); 8155 8156 /* 8157 * See if the config needs to be updated. 8158 */ 8159 if (tasks & SPA_ASYNC_CONFIG_UPDATE) { 8160 uint64_t old_space, new_space; 8161 8162 mutex_enter(&spa_namespace_lock); 8163 old_space = metaslab_class_get_space(spa_normal_class(spa)); 8164 old_space += metaslab_class_get_space(spa_special_class(spa)); 8165 old_space += metaslab_class_get_space(spa_dedup_class(spa)); 8166 old_space += metaslab_class_get_space( 8167 spa_embedded_log_class(spa)); 8168 8169 spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); 8170 8171 new_space = metaslab_class_get_space(spa_normal_class(spa)); 8172 new_space += metaslab_class_get_space(spa_special_class(spa)); 8173 new_space += metaslab_class_get_space(spa_dedup_class(spa)); 8174 new_space += metaslab_class_get_space( 8175 spa_embedded_log_class(spa)); 8176 mutex_exit(&spa_namespace_lock); 8177 8178 /* 8179 * If the pool grew as a result of the config update, 8180 * then log an internal history event. 8181 */ 8182 if (new_space != old_space) { 8183 spa_history_log_internal(spa, "vdev online", NULL, 8184 "pool '%s' size: %llu(+%llu)", 8185 spa_name(spa), (u_longlong_t)new_space, 8186 (u_longlong_t)(new_space - old_space)); 8187 } 8188 } 8189 8190 /* 8191 * See if any devices need to be marked REMOVED. 8192 */ 8193 if (tasks & SPA_ASYNC_REMOVE) { 8194 spa_vdev_state_enter(spa, SCL_NONE); 8195 spa_async_remove(spa, spa->spa_root_vdev); 8196 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) 8197 spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]); 8198 for (int i = 0; i < spa->spa_spares.sav_count; i++) 8199 spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]); 8200 (void) spa_vdev_state_exit(spa, NULL, 0); 8201 } 8202 8203 if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) { 8204 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8205 spa_async_autoexpand(spa, spa->spa_root_vdev); 8206 spa_config_exit(spa, SCL_CONFIG, FTAG); 8207 } 8208 8209 /* 8210 * See if any devices need to be probed. 8211 */ 8212 if (tasks & SPA_ASYNC_PROBE) { 8213 spa_vdev_state_enter(spa, SCL_NONE); 8214 spa_async_probe(spa, spa->spa_root_vdev); 8215 (void) spa_vdev_state_exit(spa, NULL, 0); 8216 } 8217 8218 /* 8219 * If any devices are done replacing, detach them. 8220 */ 8221 if (tasks & SPA_ASYNC_RESILVER_DONE || 8222 tasks & SPA_ASYNC_REBUILD_DONE) { 8223 spa_vdev_resilver_done(spa); 8224 } 8225 8226 /* 8227 * Kick off a resilver. 8228 */ 8229 if (tasks & SPA_ASYNC_RESILVER && 8230 !vdev_rebuild_active(spa->spa_root_vdev) && 8231 (!dsl_scan_resilvering(dp) || 8232 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))) 8233 dsl_scan_restart_resilver(dp, 0); 8234 8235 if (tasks & SPA_ASYNC_INITIALIZE_RESTART) { 8236 mutex_enter(&spa_namespace_lock); 8237 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8238 vdev_initialize_restart(spa->spa_root_vdev); 8239 spa_config_exit(spa, SCL_CONFIG, FTAG); 8240 mutex_exit(&spa_namespace_lock); 8241 } 8242 8243 if (tasks & SPA_ASYNC_TRIM_RESTART) { 8244 mutex_enter(&spa_namespace_lock); 8245 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8246 vdev_trim_restart(spa->spa_root_vdev); 8247 spa_config_exit(spa, SCL_CONFIG, FTAG); 8248 mutex_exit(&spa_namespace_lock); 8249 } 8250 8251 if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) { 8252 mutex_enter(&spa_namespace_lock); 8253 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8254 vdev_autotrim_restart(spa); 8255 spa_config_exit(spa, SCL_CONFIG, FTAG); 8256 mutex_exit(&spa_namespace_lock); 8257 } 8258 8259 /* 8260 * Kick off L2 cache whole device TRIM. 8261 */ 8262 if (tasks & SPA_ASYNC_L2CACHE_TRIM) { 8263 mutex_enter(&spa_namespace_lock); 8264 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8265 vdev_trim_l2arc(spa); 8266 spa_config_exit(spa, SCL_CONFIG, FTAG); 8267 mutex_exit(&spa_namespace_lock); 8268 } 8269 8270 /* 8271 * Kick off L2 cache rebuilding. 8272 */ 8273 if (tasks & SPA_ASYNC_L2CACHE_REBUILD) { 8274 mutex_enter(&spa_namespace_lock); 8275 spa_config_enter(spa, SCL_L2ARC, FTAG, RW_READER); 8276 l2arc_spa_rebuild_start(spa); 8277 spa_config_exit(spa, SCL_L2ARC, FTAG); 8278 mutex_exit(&spa_namespace_lock); 8279 } 8280 8281 /* 8282 * Let the world know that we're done. 8283 */ 8284 mutex_enter(&spa->spa_async_lock); 8285 spa->spa_async_thread = NULL; 8286 cv_broadcast(&spa->spa_async_cv); 8287 mutex_exit(&spa->spa_async_lock); 8288 thread_exit(); 8289 } 8290 8291 void 8292 spa_async_suspend(spa_t *spa) 8293 { 8294 mutex_enter(&spa->spa_async_lock); 8295 spa->spa_async_suspended++; 8296 while (spa->spa_async_thread != NULL) 8297 cv_wait(&spa->spa_async_cv, &spa->spa_async_lock); 8298 mutex_exit(&spa->spa_async_lock); 8299 8300 spa_vdev_remove_suspend(spa); 8301 8302 zthr_t *condense_thread = spa->spa_condense_zthr; 8303 if (condense_thread != NULL) 8304 zthr_cancel(condense_thread); 8305 8306 zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr; 8307 if (discard_thread != NULL) 8308 zthr_cancel(discard_thread); 8309 8310 zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr; 8311 if (ll_delete_thread != NULL) 8312 zthr_cancel(ll_delete_thread); 8313 8314 zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr; 8315 if (ll_condense_thread != NULL) 8316 zthr_cancel(ll_condense_thread); 8317 } 8318 8319 void 8320 spa_async_resume(spa_t *spa) 8321 { 8322 mutex_enter(&spa->spa_async_lock); 8323 ASSERT(spa->spa_async_suspended != 0); 8324 spa->spa_async_suspended--; 8325 mutex_exit(&spa->spa_async_lock); 8326 spa_restart_removal(spa); 8327 8328 zthr_t *condense_thread = spa->spa_condense_zthr; 8329 if (condense_thread != NULL) 8330 zthr_resume(condense_thread); 8331 8332 zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr; 8333 if (discard_thread != NULL) 8334 zthr_resume(discard_thread); 8335 8336 zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr; 8337 if (ll_delete_thread != NULL) 8338 zthr_resume(ll_delete_thread); 8339 8340 zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr; 8341 if (ll_condense_thread != NULL) 8342 zthr_resume(ll_condense_thread); 8343 } 8344 8345 static boolean_t 8346 spa_async_tasks_pending(spa_t *spa) 8347 { 8348 uint_t non_config_tasks; 8349 uint_t config_task; 8350 boolean_t config_task_suspended; 8351 8352 non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE; 8353 config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE; 8354 if (spa->spa_ccw_fail_time == 0) { 8355 config_task_suspended = B_FALSE; 8356 } else { 8357 config_task_suspended = 8358 (gethrtime() - spa->spa_ccw_fail_time) < 8359 ((hrtime_t)zfs_ccw_retry_interval * NANOSEC); 8360 } 8361 8362 return (non_config_tasks || (config_task && !config_task_suspended)); 8363 } 8364 8365 static void 8366 spa_async_dispatch(spa_t *spa) 8367 { 8368 mutex_enter(&spa->spa_async_lock); 8369 if (spa_async_tasks_pending(spa) && 8370 !spa->spa_async_suspended && 8371 spa->spa_async_thread == NULL) 8372 spa->spa_async_thread = thread_create(NULL, 0, 8373 spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri); 8374 mutex_exit(&spa->spa_async_lock); 8375 } 8376 8377 void 8378 spa_async_request(spa_t *spa, int task) 8379 { 8380 zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task); 8381 mutex_enter(&spa->spa_async_lock); 8382 spa->spa_async_tasks |= task; 8383 mutex_exit(&spa->spa_async_lock); 8384 } 8385 8386 int 8387 spa_async_tasks(spa_t *spa) 8388 { 8389 return (spa->spa_async_tasks); 8390 } 8391 8392 /* 8393 * ========================================================================== 8394 * SPA syncing routines 8395 * ========================================================================== 8396 */ 8397 8398 8399 static int 8400 bpobj_enqueue_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 8401 dmu_tx_t *tx) 8402 { 8403 bpobj_t *bpo = arg; 8404 bpobj_enqueue(bpo, bp, bp_freed, tx); 8405 return (0); 8406 } 8407 8408 int 8409 bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 8410 { 8411 return (bpobj_enqueue_cb(arg, bp, B_FALSE, tx)); 8412 } 8413 8414 int 8415 bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 8416 { 8417 return (bpobj_enqueue_cb(arg, bp, B_TRUE, tx)); 8418 } 8419 8420 static int 8421 spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 8422 { 8423 zio_t *pio = arg; 8424 8425 zio_nowait(zio_free_sync(pio, pio->io_spa, dmu_tx_get_txg(tx), bp, 8426 pio->io_flags)); 8427 return (0); 8428 } 8429 8430 static int 8431 bpobj_spa_free_sync_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 8432 dmu_tx_t *tx) 8433 { 8434 ASSERT(!bp_freed); 8435 return (spa_free_sync_cb(arg, bp, tx)); 8436 } 8437 8438 /* 8439 * Note: this simple function is not inlined to make it easier to dtrace the 8440 * amount of time spent syncing frees. 8441 */ 8442 static void 8443 spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx) 8444 { 8445 zio_t *zio = zio_root(spa, NULL, NULL, 0); 8446 bplist_iterate(bpl, spa_free_sync_cb, zio, tx); 8447 VERIFY(zio_wait(zio) == 0); 8448 } 8449 8450 /* 8451 * Note: this simple function is not inlined to make it easier to dtrace the 8452 * amount of time spent syncing deferred frees. 8453 */ 8454 static void 8455 spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx) 8456 { 8457 if (spa_sync_pass(spa) != 1) 8458 return; 8459 8460 /* 8461 * Note: 8462 * If the log space map feature is active, we stop deferring 8463 * frees to the next TXG and therefore running this function 8464 * would be considered a no-op as spa_deferred_bpobj should 8465 * not have any entries. 8466 * 8467 * That said we run this function anyway (instead of returning 8468 * immediately) for the edge-case scenario where we just 8469 * activated the log space map feature in this TXG but we have 8470 * deferred frees from the previous TXG. 8471 */ 8472 zio_t *zio = zio_root(spa, NULL, NULL, 0); 8473 VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj, 8474 bpobj_spa_free_sync_cb, zio, tx), ==, 0); 8475 VERIFY0(zio_wait(zio)); 8476 } 8477 8478 static void 8479 spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx) 8480 { 8481 char *packed = NULL; 8482 size_t bufsize; 8483 size_t nvsize = 0; 8484 dmu_buf_t *db; 8485 8486 VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0); 8487 8488 /* 8489 * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration 8490 * information. This avoids the dmu_buf_will_dirty() path and 8491 * saves us a pre-read to get data we don't actually care about. 8492 */ 8493 bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE); 8494 packed = vmem_alloc(bufsize, KM_SLEEP); 8495 8496 VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR, 8497 KM_SLEEP) == 0); 8498 memset(packed + nvsize, 0, bufsize - nvsize); 8499 8500 dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx); 8501 8502 vmem_free(packed, bufsize); 8503 8504 VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db)); 8505 dmu_buf_will_dirty(db, tx); 8506 *(uint64_t *)db->db_data = nvsize; 8507 dmu_buf_rele(db, FTAG); 8508 } 8509 8510 static void 8511 spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx, 8512 const char *config, const char *entry) 8513 { 8514 nvlist_t *nvroot; 8515 nvlist_t **list; 8516 int i; 8517 8518 if (!sav->sav_sync) 8519 return; 8520 8521 /* 8522 * Update the MOS nvlist describing the list of available devices. 8523 * spa_validate_aux() will have already made sure this nvlist is 8524 * valid and the vdevs are labeled appropriately. 8525 */ 8526 if (sav->sav_object == 0) { 8527 sav->sav_object = dmu_object_alloc(spa->spa_meta_objset, 8528 DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE, 8529 sizeof (uint64_t), tx); 8530 VERIFY(zap_update(spa->spa_meta_objset, 8531 DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1, 8532 &sav->sav_object, tx) == 0); 8533 } 8534 8535 nvroot = fnvlist_alloc(); 8536 if (sav->sav_count == 0) { 8537 fnvlist_add_nvlist_array(nvroot, config, 8538 (const nvlist_t * const *)NULL, 0); 8539 } else { 8540 list = kmem_alloc(sav->sav_count*sizeof (void *), KM_SLEEP); 8541 for (i = 0; i < sav->sav_count; i++) 8542 list[i] = vdev_config_generate(spa, sav->sav_vdevs[i], 8543 B_FALSE, VDEV_CONFIG_L2CACHE); 8544 fnvlist_add_nvlist_array(nvroot, config, 8545 (const nvlist_t * const *)list, sav->sav_count); 8546 for (i = 0; i < sav->sav_count; i++) 8547 nvlist_free(list[i]); 8548 kmem_free(list, sav->sav_count * sizeof (void *)); 8549 } 8550 8551 spa_sync_nvlist(spa, sav->sav_object, nvroot, tx); 8552 nvlist_free(nvroot); 8553 8554 sav->sav_sync = B_FALSE; 8555 } 8556 8557 /* 8558 * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t. 8559 * The all-vdev ZAP must be empty. 8560 */ 8561 static void 8562 spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx) 8563 { 8564 spa_t *spa = vd->vdev_spa; 8565 8566 if (vd->vdev_top_zap != 0) { 8567 VERIFY0(zap_add_int(spa->spa_meta_objset, avz, 8568 vd->vdev_top_zap, tx)); 8569 } 8570 if (vd->vdev_leaf_zap != 0) { 8571 VERIFY0(zap_add_int(spa->spa_meta_objset, avz, 8572 vd->vdev_leaf_zap, tx)); 8573 } 8574 for (uint64_t i = 0; i < vd->vdev_children; i++) { 8575 spa_avz_build(vd->vdev_child[i], avz, tx); 8576 } 8577 } 8578 8579 static void 8580 spa_sync_config_object(spa_t *spa, dmu_tx_t *tx) 8581 { 8582 nvlist_t *config; 8583 8584 /* 8585 * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS, 8586 * its config may not be dirty but we still need to build per-vdev ZAPs. 8587 * Similarly, if the pool is being assembled (e.g. after a split), we 8588 * need to rebuild the AVZ although the config may not be dirty. 8589 */ 8590 if (list_is_empty(&spa->spa_config_dirty_list) && 8591 spa->spa_avz_action == AVZ_ACTION_NONE) 8592 return; 8593 8594 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 8595 8596 ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE || 8597 spa->spa_avz_action == AVZ_ACTION_INITIALIZE || 8598 spa->spa_all_vdev_zaps != 0); 8599 8600 if (spa->spa_avz_action == AVZ_ACTION_REBUILD) { 8601 /* Make and build the new AVZ */ 8602 uint64_t new_avz = zap_create(spa->spa_meta_objset, 8603 DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx); 8604 spa_avz_build(spa->spa_root_vdev, new_avz, tx); 8605 8606 /* Diff old AVZ with new one */ 8607 zap_cursor_t zc; 8608 zap_attribute_t za; 8609 8610 for (zap_cursor_init(&zc, spa->spa_meta_objset, 8611 spa->spa_all_vdev_zaps); 8612 zap_cursor_retrieve(&zc, &za) == 0; 8613 zap_cursor_advance(&zc)) { 8614 uint64_t vdzap = za.za_first_integer; 8615 if (zap_lookup_int(spa->spa_meta_objset, new_avz, 8616 vdzap) == ENOENT) { 8617 /* 8618 * ZAP is listed in old AVZ but not in new one; 8619 * destroy it 8620 */ 8621 VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap, 8622 tx)); 8623 } 8624 } 8625 8626 zap_cursor_fini(&zc); 8627 8628 /* Destroy the old AVZ */ 8629 VERIFY0(zap_destroy(spa->spa_meta_objset, 8630 spa->spa_all_vdev_zaps, tx)); 8631 8632 /* Replace the old AVZ in the dir obj with the new one */ 8633 VERIFY0(zap_update(spa->spa_meta_objset, 8634 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, 8635 sizeof (new_avz), 1, &new_avz, tx)); 8636 8637 spa->spa_all_vdev_zaps = new_avz; 8638 } else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) { 8639 zap_cursor_t zc; 8640 zap_attribute_t za; 8641 8642 /* Walk through the AVZ and destroy all listed ZAPs */ 8643 for (zap_cursor_init(&zc, spa->spa_meta_objset, 8644 spa->spa_all_vdev_zaps); 8645 zap_cursor_retrieve(&zc, &za) == 0; 8646 zap_cursor_advance(&zc)) { 8647 uint64_t zap = za.za_first_integer; 8648 VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx)); 8649 } 8650 8651 zap_cursor_fini(&zc); 8652 8653 /* Destroy and unlink the AVZ itself */ 8654 VERIFY0(zap_destroy(spa->spa_meta_objset, 8655 spa->spa_all_vdev_zaps, tx)); 8656 VERIFY0(zap_remove(spa->spa_meta_objset, 8657 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx)); 8658 spa->spa_all_vdev_zaps = 0; 8659 } 8660 8661 if (spa->spa_all_vdev_zaps == 0) { 8662 spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset, 8663 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT, 8664 DMU_POOL_VDEV_ZAP_MAP, tx); 8665 } 8666 spa->spa_avz_action = AVZ_ACTION_NONE; 8667 8668 /* Create ZAPs for vdevs that don't have them. */ 8669 vdev_construct_zaps(spa->spa_root_vdev, tx); 8670 8671 config = spa_config_generate(spa, spa->spa_root_vdev, 8672 dmu_tx_get_txg(tx), B_FALSE); 8673 8674 /* 8675 * If we're upgrading the spa version then make sure that 8676 * the config object gets updated with the correct version. 8677 */ 8678 if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version) 8679 fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, 8680 spa->spa_uberblock.ub_version); 8681 8682 spa_config_exit(spa, SCL_STATE, FTAG); 8683 8684 nvlist_free(spa->spa_config_syncing); 8685 spa->spa_config_syncing = config; 8686 8687 spa_sync_nvlist(spa, spa->spa_config_object, config, tx); 8688 } 8689 8690 static void 8691 spa_sync_version(void *arg, dmu_tx_t *tx) 8692 { 8693 uint64_t *versionp = arg; 8694 uint64_t version = *versionp; 8695 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 8696 8697 /* 8698 * Setting the version is special cased when first creating the pool. 8699 */ 8700 ASSERT(tx->tx_txg != TXG_INITIAL); 8701 8702 ASSERT(SPA_VERSION_IS_SUPPORTED(version)); 8703 ASSERT(version >= spa_version(spa)); 8704 8705 spa->spa_uberblock.ub_version = version; 8706 vdev_config_dirty(spa->spa_root_vdev); 8707 spa_history_log_internal(spa, "set", tx, "version=%lld", 8708 (longlong_t)version); 8709 } 8710 8711 /* 8712 * Set zpool properties. 8713 */ 8714 static void 8715 spa_sync_props(void *arg, dmu_tx_t *tx) 8716 { 8717 nvlist_t *nvp = arg; 8718 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 8719 objset_t *mos = spa->spa_meta_objset; 8720 nvpair_t *elem = NULL; 8721 8722 mutex_enter(&spa->spa_props_lock); 8723 8724 while ((elem = nvlist_next_nvpair(nvp, elem))) { 8725 uint64_t intval; 8726 char *strval, *fname; 8727 zpool_prop_t prop; 8728 const char *propname; 8729 zprop_type_t proptype; 8730 spa_feature_t fid; 8731 8732 switch (prop = zpool_name_to_prop(nvpair_name(elem))) { 8733 case ZPOOL_PROP_INVAL: 8734 /* 8735 * We checked this earlier in spa_prop_validate(). 8736 */ 8737 ASSERT(zpool_prop_feature(nvpair_name(elem))); 8738 8739 fname = strchr(nvpair_name(elem), '@') + 1; 8740 VERIFY0(zfeature_lookup_name(fname, &fid)); 8741 8742 spa_feature_enable(spa, fid, tx); 8743 spa_history_log_internal(spa, "set", tx, 8744 "%s=enabled", nvpair_name(elem)); 8745 break; 8746 8747 case ZPOOL_PROP_VERSION: 8748 intval = fnvpair_value_uint64(elem); 8749 /* 8750 * The version is synced separately before other 8751 * properties and should be correct by now. 8752 */ 8753 ASSERT3U(spa_version(spa), >=, intval); 8754 break; 8755 8756 case ZPOOL_PROP_ALTROOT: 8757 /* 8758 * 'altroot' is a non-persistent property. It should 8759 * have been set temporarily at creation or import time. 8760 */ 8761 ASSERT(spa->spa_root != NULL); 8762 break; 8763 8764 case ZPOOL_PROP_READONLY: 8765 case ZPOOL_PROP_CACHEFILE: 8766 /* 8767 * 'readonly' and 'cachefile' are also non-persistent 8768 * properties. 8769 */ 8770 break; 8771 case ZPOOL_PROP_COMMENT: 8772 strval = fnvpair_value_string(elem); 8773 if (spa->spa_comment != NULL) 8774 spa_strfree(spa->spa_comment); 8775 spa->spa_comment = spa_strdup(strval); 8776 /* 8777 * We need to dirty the configuration on all the vdevs 8778 * so that their labels get updated. We also need to 8779 * update the cache file to keep it in sync with the 8780 * MOS version. It's unnecessary to do this for pool 8781 * creation since the vdev's configuration has already 8782 * been dirtied. 8783 */ 8784 if (tx->tx_txg != TXG_INITIAL) { 8785 vdev_config_dirty(spa->spa_root_vdev); 8786 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 8787 } 8788 spa_history_log_internal(spa, "set", tx, 8789 "%s=%s", nvpair_name(elem), strval); 8790 break; 8791 case ZPOOL_PROP_COMPATIBILITY: 8792 strval = fnvpair_value_string(elem); 8793 if (spa->spa_compatibility != NULL) 8794 spa_strfree(spa->spa_compatibility); 8795 spa->spa_compatibility = spa_strdup(strval); 8796 /* 8797 * Dirty the configuration on vdevs as above. 8798 */ 8799 if (tx->tx_txg != TXG_INITIAL) { 8800 vdev_config_dirty(spa->spa_root_vdev); 8801 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 8802 } 8803 8804 spa_history_log_internal(spa, "set", tx, 8805 "%s=%s", nvpair_name(elem), strval); 8806 break; 8807 8808 default: 8809 /* 8810 * Set pool property values in the poolprops mos object. 8811 */ 8812 if (spa->spa_pool_props_object == 0) { 8813 spa->spa_pool_props_object = 8814 zap_create_link(mos, DMU_OT_POOL_PROPS, 8815 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS, 8816 tx); 8817 } 8818 8819 /* normalize the property name */ 8820 propname = zpool_prop_to_name(prop); 8821 proptype = zpool_prop_get_type(prop); 8822 8823 if (nvpair_type(elem) == DATA_TYPE_STRING) { 8824 ASSERT(proptype == PROP_TYPE_STRING); 8825 strval = fnvpair_value_string(elem); 8826 VERIFY0(zap_update(mos, 8827 spa->spa_pool_props_object, propname, 8828 1, strlen(strval) + 1, strval, tx)); 8829 spa_history_log_internal(spa, "set", tx, 8830 "%s=%s", nvpair_name(elem), strval); 8831 } else if (nvpair_type(elem) == DATA_TYPE_UINT64) { 8832 intval = fnvpair_value_uint64(elem); 8833 8834 if (proptype == PROP_TYPE_INDEX) { 8835 const char *unused; 8836 VERIFY0(zpool_prop_index_to_string( 8837 prop, intval, &unused)); 8838 } 8839 VERIFY0(zap_update(mos, 8840 spa->spa_pool_props_object, propname, 8841 8, 1, &intval, tx)); 8842 spa_history_log_internal(spa, "set", tx, 8843 "%s=%lld", nvpair_name(elem), 8844 (longlong_t)intval); 8845 } else { 8846 ASSERT(0); /* not allowed */ 8847 } 8848 8849 switch (prop) { 8850 case ZPOOL_PROP_DELEGATION: 8851 spa->spa_delegation = intval; 8852 break; 8853 case ZPOOL_PROP_BOOTFS: 8854 spa->spa_bootfs = intval; 8855 break; 8856 case ZPOOL_PROP_FAILUREMODE: 8857 spa->spa_failmode = intval; 8858 break; 8859 case ZPOOL_PROP_AUTOTRIM: 8860 spa->spa_autotrim = intval; 8861 spa_async_request(spa, 8862 SPA_ASYNC_AUTOTRIM_RESTART); 8863 break; 8864 case ZPOOL_PROP_AUTOEXPAND: 8865 spa->spa_autoexpand = intval; 8866 if (tx->tx_txg != TXG_INITIAL) 8867 spa_async_request(spa, 8868 SPA_ASYNC_AUTOEXPAND); 8869 break; 8870 case ZPOOL_PROP_MULTIHOST: 8871 spa->spa_multihost = intval; 8872 break; 8873 default: 8874 break; 8875 } 8876 } 8877 8878 } 8879 8880 mutex_exit(&spa->spa_props_lock); 8881 } 8882 8883 /* 8884 * Perform one-time upgrade on-disk changes. spa_version() does not 8885 * reflect the new version this txg, so there must be no changes this 8886 * txg to anything that the upgrade code depends on after it executes. 8887 * Therefore this must be called after dsl_pool_sync() does the sync 8888 * tasks. 8889 */ 8890 static void 8891 spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx) 8892 { 8893 if (spa_sync_pass(spa) != 1) 8894 return; 8895 8896 dsl_pool_t *dp = spa->spa_dsl_pool; 8897 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 8898 8899 if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN && 8900 spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) { 8901 dsl_pool_create_origin(dp, tx); 8902 8903 /* Keeping the origin open increases spa_minref */ 8904 spa->spa_minref += 3; 8905 } 8906 8907 if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES && 8908 spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) { 8909 dsl_pool_upgrade_clones(dp, tx); 8910 } 8911 8912 if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES && 8913 spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) { 8914 dsl_pool_upgrade_dir_clones(dp, tx); 8915 8916 /* Keeping the freedir open increases spa_minref */ 8917 spa->spa_minref += 3; 8918 } 8919 8920 if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES && 8921 spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) { 8922 spa_feature_create_zap_objects(spa, tx); 8923 } 8924 8925 /* 8926 * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable 8927 * when possibility to use lz4 compression for metadata was added 8928 * Old pools that have this feature enabled must be upgraded to have 8929 * this feature active 8930 */ 8931 if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) { 8932 boolean_t lz4_en = spa_feature_is_enabled(spa, 8933 SPA_FEATURE_LZ4_COMPRESS); 8934 boolean_t lz4_ac = spa_feature_is_active(spa, 8935 SPA_FEATURE_LZ4_COMPRESS); 8936 8937 if (lz4_en && !lz4_ac) 8938 spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx); 8939 } 8940 8941 /* 8942 * If we haven't written the salt, do so now. Note that the 8943 * feature may not be activated yet, but that's fine since 8944 * the presence of this ZAP entry is backwards compatible. 8945 */ 8946 if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 8947 DMU_POOL_CHECKSUM_SALT) == ENOENT) { 8948 VERIFY0(zap_add(spa->spa_meta_objset, 8949 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1, 8950 sizeof (spa->spa_cksum_salt.zcs_bytes), 8951 spa->spa_cksum_salt.zcs_bytes, tx)); 8952 } 8953 8954 rrw_exit(&dp->dp_config_rwlock, FTAG); 8955 } 8956 8957 static void 8958 vdev_indirect_state_sync_verify(vdev_t *vd) 8959 { 8960 vdev_indirect_mapping_t *vim __maybe_unused = vd->vdev_indirect_mapping; 8961 vdev_indirect_births_t *vib __maybe_unused = vd->vdev_indirect_births; 8962 8963 if (vd->vdev_ops == &vdev_indirect_ops) { 8964 ASSERT(vim != NULL); 8965 ASSERT(vib != NULL); 8966 } 8967 8968 uint64_t obsolete_sm_object = 0; 8969 ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object)); 8970 if (obsolete_sm_object != 0) { 8971 ASSERT(vd->vdev_obsolete_sm != NULL); 8972 ASSERT(vd->vdev_removing || 8973 vd->vdev_ops == &vdev_indirect_ops); 8974 ASSERT(vdev_indirect_mapping_num_entries(vim) > 0); 8975 ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0); 8976 ASSERT3U(obsolete_sm_object, ==, 8977 space_map_object(vd->vdev_obsolete_sm)); 8978 ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=, 8979 space_map_allocated(vd->vdev_obsolete_sm)); 8980 } 8981 ASSERT(vd->vdev_obsolete_segments != NULL); 8982 8983 /* 8984 * Since frees / remaps to an indirect vdev can only 8985 * happen in syncing context, the obsolete segments 8986 * tree must be empty when we start syncing. 8987 */ 8988 ASSERT0(range_tree_space(vd->vdev_obsolete_segments)); 8989 } 8990 8991 /* 8992 * Set the top-level vdev's max queue depth. Evaluate each top-level's 8993 * async write queue depth in case it changed. The max queue depth will 8994 * not change in the middle of syncing out this txg. 8995 */ 8996 static void 8997 spa_sync_adjust_vdev_max_queue_depth(spa_t *spa) 8998 { 8999 ASSERT(spa_writeable(spa)); 9000 9001 vdev_t *rvd = spa->spa_root_vdev; 9002 uint32_t max_queue_depth = zfs_vdev_async_write_max_active * 9003 zfs_vdev_queue_depth_pct / 100; 9004 metaslab_class_t *normal = spa_normal_class(spa); 9005 metaslab_class_t *special = spa_special_class(spa); 9006 metaslab_class_t *dedup = spa_dedup_class(spa); 9007 9008 uint64_t slots_per_allocator = 0; 9009 for (int c = 0; c < rvd->vdev_children; c++) { 9010 vdev_t *tvd = rvd->vdev_child[c]; 9011 9012 metaslab_group_t *mg = tvd->vdev_mg; 9013 if (mg == NULL || !metaslab_group_initialized(mg)) 9014 continue; 9015 9016 metaslab_class_t *mc = mg->mg_class; 9017 if (mc != normal && mc != special && mc != dedup) 9018 continue; 9019 9020 /* 9021 * It is safe to do a lock-free check here because only async 9022 * allocations look at mg_max_alloc_queue_depth, and async 9023 * allocations all happen from spa_sync(). 9024 */ 9025 for (int i = 0; i < mg->mg_allocators; i++) { 9026 ASSERT0(zfs_refcount_count( 9027 &(mg->mg_allocator[i].mga_alloc_queue_depth))); 9028 } 9029 mg->mg_max_alloc_queue_depth = max_queue_depth; 9030 9031 for (int i = 0; i < mg->mg_allocators; i++) { 9032 mg->mg_allocator[i].mga_cur_max_alloc_queue_depth = 9033 zfs_vdev_def_queue_depth; 9034 } 9035 slots_per_allocator += zfs_vdev_def_queue_depth; 9036 } 9037 9038 for (int i = 0; i < spa->spa_alloc_count; i++) { 9039 ASSERT0(zfs_refcount_count(&normal->mc_allocator[i]. 9040 mca_alloc_slots)); 9041 ASSERT0(zfs_refcount_count(&special->mc_allocator[i]. 9042 mca_alloc_slots)); 9043 ASSERT0(zfs_refcount_count(&dedup->mc_allocator[i]. 9044 mca_alloc_slots)); 9045 normal->mc_allocator[i].mca_alloc_max_slots = 9046 slots_per_allocator; 9047 special->mc_allocator[i].mca_alloc_max_slots = 9048 slots_per_allocator; 9049 dedup->mc_allocator[i].mca_alloc_max_slots = 9050 slots_per_allocator; 9051 } 9052 normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; 9053 special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; 9054 dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; 9055 } 9056 9057 static void 9058 spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx) 9059 { 9060 ASSERT(spa_writeable(spa)); 9061 9062 vdev_t *rvd = spa->spa_root_vdev; 9063 for (int c = 0; c < rvd->vdev_children; c++) { 9064 vdev_t *vd = rvd->vdev_child[c]; 9065 vdev_indirect_state_sync_verify(vd); 9066 9067 if (vdev_indirect_should_condense(vd)) { 9068 spa_condense_indirect_start_sync(vd, tx); 9069 break; 9070 } 9071 } 9072 } 9073 9074 static void 9075 spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx) 9076 { 9077 objset_t *mos = spa->spa_meta_objset; 9078 dsl_pool_t *dp = spa->spa_dsl_pool; 9079 uint64_t txg = tx->tx_txg; 9080 bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK]; 9081 9082 do { 9083 int pass = ++spa->spa_sync_pass; 9084 9085 spa_sync_config_object(spa, tx); 9086 spa_sync_aux_dev(spa, &spa->spa_spares, tx, 9087 ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES); 9088 spa_sync_aux_dev(spa, &spa->spa_l2cache, tx, 9089 ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE); 9090 spa_errlog_sync(spa, txg); 9091 dsl_pool_sync(dp, txg); 9092 9093 if (pass < zfs_sync_pass_deferred_free || 9094 spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) { 9095 /* 9096 * If the log space map feature is active we don't 9097 * care about deferred frees and the deferred bpobj 9098 * as the log space map should effectively have the 9099 * same results (i.e. appending only to one object). 9100 */ 9101 spa_sync_frees(spa, free_bpl, tx); 9102 } else { 9103 /* 9104 * We can not defer frees in pass 1, because 9105 * we sync the deferred frees later in pass 1. 9106 */ 9107 ASSERT3U(pass, >, 1); 9108 bplist_iterate(free_bpl, bpobj_enqueue_alloc_cb, 9109 &spa->spa_deferred_bpobj, tx); 9110 } 9111 9112 ddt_sync(spa, txg); 9113 dsl_scan_sync(dp, tx); 9114 svr_sync(spa, tx); 9115 spa_sync_upgrades(spa, tx); 9116 9117 spa_flush_metaslabs(spa, tx); 9118 9119 vdev_t *vd = NULL; 9120 while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg)) 9121 != NULL) 9122 vdev_sync(vd, txg); 9123 9124 /* 9125 * Note: We need to check if the MOS is dirty because we could 9126 * have marked the MOS dirty without updating the uberblock 9127 * (e.g. if we have sync tasks but no dirty user data). We need 9128 * to check the uberblock's rootbp because it is updated if we 9129 * have synced out dirty data (though in this case the MOS will 9130 * most likely also be dirty due to second order effects, we 9131 * don't want to rely on that here). 9132 */ 9133 if (pass == 1 && 9134 spa->spa_uberblock.ub_rootbp.blk_birth < txg && 9135 !dmu_objset_is_dirty(mos, txg)) { 9136 /* 9137 * Nothing changed on the first pass, therefore this 9138 * TXG is a no-op. Avoid syncing deferred frees, so 9139 * that we can keep this TXG as a no-op. 9140 */ 9141 ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg)); 9142 ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg)); 9143 ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg)); 9144 ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg)); 9145 break; 9146 } 9147 9148 spa_sync_deferred_frees(spa, tx); 9149 } while (dmu_objset_is_dirty(mos, txg)); 9150 } 9151 9152 /* 9153 * Rewrite the vdev configuration (which includes the uberblock) to 9154 * commit the transaction group. 9155 * 9156 * If there are no dirty vdevs, we sync the uberblock to a few random 9157 * top-level vdevs that are known to be visible in the config cache 9158 * (see spa_vdev_add() for a complete description). If there *are* dirty 9159 * vdevs, sync the uberblock to all vdevs. 9160 */ 9161 static void 9162 spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx) 9163 { 9164 vdev_t *rvd = spa->spa_root_vdev; 9165 uint64_t txg = tx->tx_txg; 9166 9167 for (;;) { 9168 int error = 0; 9169 9170 /* 9171 * We hold SCL_STATE to prevent vdev open/close/etc. 9172 * while we're attempting to write the vdev labels. 9173 */ 9174 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 9175 9176 if (list_is_empty(&spa->spa_config_dirty_list)) { 9177 vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL }; 9178 int svdcount = 0; 9179 int children = rvd->vdev_children; 9180 int c0 = random_in_range(children); 9181 9182 for (int c = 0; c < children; c++) { 9183 vdev_t *vd = 9184 rvd->vdev_child[(c0 + c) % children]; 9185 9186 /* Stop when revisiting the first vdev */ 9187 if (c > 0 && svd[0] == vd) 9188 break; 9189 9190 if (vd->vdev_ms_array == 0 || 9191 vd->vdev_islog || 9192 !vdev_is_concrete(vd)) 9193 continue; 9194 9195 svd[svdcount++] = vd; 9196 if (svdcount == SPA_SYNC_MIN_VDEVS) 9197 break; 9198 } 9199 error = vdev_config_sync(svd, svdcount, txg); 9200 } else { 9201 error = vdev_config_sync(rvd->vdev_child, 9202 rvd->vdev_children, txg); 9203 } 9204 9205 if (error == 0) 9206 spa->spa_last_synced_guid = rvd->vdev_guid; 9207 9208 spa_config_exit(spa, SCL_STATE, FTAG); 9209 9210 if (error == 0) 9211 break; 9212 zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR); 9213 zio_resume_wait(spa); 9214 } 9215 } 9216 9217 /* 9218 * Sync the specified transaction group. New blocks may be dirtied as 9219 * part of the process, so we iterate until it converges. 9220 */ 9221 void 9222 spa_sync(spa_t *spa, uint64_t txg) 9223 { 9224 vdev_t *vd = NULL; 9225 9226 VERIFY(spa_writeable(spa)); 9227 9228 /* 9229 * Wait for i/os issued in open context that need to complete 9230 * before this txg syncs. 9231 */ 9232 (void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]); 9233 spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL, 9234 ZIO_FLAG_CANFAIL); 9235 9236 /* 9237 * Lock out configuration changes. 9238 */ 9239 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 9240 9241 spa->spa_syncing_txg = txg; 9242 spa->spa_sync_pass = 0; 9243 9244 for (int i = 0; i < spa->spa_alloc_count; i++) { 9245 mutex_enter(&spa->spa_allocs[i].spaa_lock); 9246 VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree)); 9247 mutex_exit(&spa->spa_allocs[i].spaa_lock); 9248 } 9249 9250 /* 9251 * If there are any pending vdev state changes, convert them 9252 * into config changes that go out with this transaction group. 9253 */ 9254 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 9255 while (list_head(&spa->spa_state_dirty_list) != NULL) { 9256 /* 9257 * We need the write lock here because, for aux vdevs, 9258 * calling vdev_config_dirty() modifies sav_config. 9259 * This is ugly and will become unnecessary when we 9260 * eliminate the aux vdev wart by integrating all vdevs 9261 * into the root vdev tree. 9262 */ 9263 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9264 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER); 9265 while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) { 9266 vdev_state_clean(vd); 9267 vdev_config_dirty(vd); 9268 } 9269 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9270 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 9271 } 9272 spa_config_exit(spa, SCL_STATE, FTAG); 9273 9274 dsl_pool_t *dp = spa->spa_dsl_pool; 9275 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 9276 9277 spa->spa_sync_starttime = gethrtime(); 9278 taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); 9279 spa->spa_deadman_tqid = taskq_dispatch_delay(system_delay_taskq, 9280 spa_deadman, spa, TQ_SLEEP, ddi_get_lbolt() + 9281 NSEC_TO_TICK(spa->spa_deadman_synctime)); 9282 9283 /* 9284 * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg, 9285 * set spa_deflate if we have no raid-z vdevs. 9286 */ 9287 if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE && 9288 spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) { 9289 vdev_t *rvd = spa->spa_root_vdev; 9290 9291 int i; 9292 for (i = 0; i < rvd->vdev_children; i++) { 9293 vd = rvd->vdev_child[i]; 9294 if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE) 9295 break; 9296 } 9297 if (i == rvd->vdev_children) { 9298 spa->spa_deflate = TRUE; 9299 VERIFY0(zap_add(spa->spa_meta_objset, 9300 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE, 9301 sizeof (uint64_t), 1, &spa->spa_deflate, tx)); 9302 } 9303 } 9304 9305 spa_sync_adjust_vdev_max_queue_depth(spa); 9306 9307 spa_sync_condense_indirect(spa, tx); 9308 9309 spa_sync_iterate_to_convergence(spa, tx); 9310 9311 #ifdef ZFS_DEBUG 9312 if (!list_is_empty(&spa->spa_config_dirty_list)) { 9313 /* 9314 * Make sure that the number of ZAPs for all the vdevs matches 9315 * the number of ZAPs in the per-vdev ZAP list. This only gets 9316 * called if the config is dirty; otherwise there may be 9317 * outstanding AVZ operations that weren't completed in 9318 * spa_sync_config_object. 9319 */ 9320 uint64_t all_vdev_zap_entry_count; 9321 ASSERT0(zap_count(spa->spa_meta_objset, 9322 spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count)); 9323 ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==, 9324 all_vdev_zap_entry_count); 9325 } 9326 #endif 9327 9328 if (spa->spa_vdev_removal != NULL) { 9329 ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]); 9330 } 9331 9332 spa_sync_rewrite_vdev_config(spa, tx); 9333 dmu_tx_commit(tx); 9334 9335 taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); 9336 spa->spa_deadman_tqid = 0; 9337 9338 /* 9339 * Clear the dirty config list. 9340 */ 9341 while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL) 9342 vdev_config_clean(vd); 9343 9344 /* 9345 * Now that the new config has synced transactionally, 9346 * let it become visible to the config cache. 9347 */ 9348 if (spa->spa_config_syncing != NULL) { 9349 spa_config_set(spa, spa->spa_config_syncing); 9350 spa->spa_config_txg = txg; 9351 spa->spa_config_syncing = NULL; 9352 } 9353 9354 dsl_pool_sync_done(dp, txg); 9355 9356 for (int i = 0; i < spa->spa_alloc_count; i++) { 9357 mutex_enter(&spa->spa_allocs[i].spaa_lock); 9358 VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree)); 9359 mutex_exit(&spa->spa_allocs[i].spaa_lock); 9360 } 9361 9362 /* 9363 * Update usable space statistics. 9364 */ 9365 while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg))) 9366 != NULL) 9367 vdev_sync_done(vd, txg); 9368 9369 metaslab_class_evict_old(spa->spa_normal_class, txg); 9370 metaslab_class_evict_old(spa->spa_log_class, txg); 9371 9372 spa_sync_close_syncing_log_sm(spa); 9373 9374 spa_update_dspace(spa); 9375 9376 /* 9377 * It had better be the case that we didn't dirty anything 9378 * since vdev_config_sync(). 9379 */ 9380 ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg)); 9381 ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg)); 9382 ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg)); 9383 9384 while (zfs_pause_spa_sync) 9385 delay(1); 9386 9387 spa->spa_sync_pass = 0; 9388 9389 /* 9390 * Update the last synced uberblock here. We want to do this at 9391 * the end of spa_sync() so that consumers of spa_last_synced_txg() 9392 * will be guaranteed that all the processing associated with 9393 * that txg has been completed. 9394 */ 9395 spa->spa_ubsync = spa->spa_uberblock; 9396 spa_config_exit(spa, SCL_CONFIG, FTAG); 9397 9398 spa_handle_ignored_writes(spa); 9399 9400 /* 9401 * If any async tasks have been requested, kick them off. 9402 */ 9403 spa_async_dispatch(spa); 9404 } 9405 9406 /* 9407 * Sync all pools. We don't want to hold the namespace lock across these 9408 * operations, so we take a reference on the spa_t and drop the lock during the 9409 * sync. 9410 */ 9411 void 9412 spa_sync_allpools(void) 9413 { 9414 spa_t *spa = NULL; 9415 mutex_enter(&spa_namespace_lock); 9416 while ((spa = spa_next(spa)) != NULL) { 9417 if (spa_state(spa) != POOL_STATE_ACTIVE || 9418 !spa_writeable(spa) || spa_suspended(spa)) 9419 continue; 9420 spa_open_ref(spa, FTAG); 9421 mutex_exit(&spa_namespace_lock); 9422 txg_wait_synced(spa_get_dsl(spa), 0); 9423 mutex_enter(&spa_namespace_lock); 9424 spa_close(spa, FTAG); 9425 } 9426 mutex_exit(&spa_namespace_lock); 9427 } 9428 9429 /* 9430 * ========================================================================== 9431 * Miscellaneous routines 9432 * ========================================================================== 9433 */ 9434 9435 /* 9436 * Remove all pools in the system. 9437 */ 9438 void 9439 spa_evict_all(void) 9440 { 9441 spa_t *spa; 9442 9443 /* 9444 * Remove all cached state. All pools should be closed now, 9445 * so every spa in the AVL tree should be unreferenced. 9446 */ 9447 mutex_enter(&spa_namespace_lock); 9448 while ((spa = spa_next(NULL)) != NULL) { 9449 /* 9450 * Stop async tasks. The async thread may need to detach 9451 * a device that's been replaced, which requires grabbing 9452 * spa_namespace_lock, so we must drop it here. 9453 */ 9454 spa_open_ref(spa, FTAG); 9455 mutex_exit(&spa_namespace_lock); 9456 spa_async_suspend(spa); 9457 mutex_enter(&spa_namespace_lock); 9458 spa_close(spa, FTAG); 9459 9460 if (spa->spa_state != POOL_STATE_UNINITIALIZED) { 9461 spa_unload(spa); 9462 spa_deactivate(spa); 9463 } 9464 spa_remove(spa); 9465 } 9466 mutex_exit(&spa_namespace_lock); 9467 } 9468 9469 vdev_t * 9470 spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux) 9471 { 9472 vdev_t *vd; 9473 int i; 9474 9475 if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL) 9476 return (vd); 9477 9478 if (aux) { 9479 for (i = 0; i < spa->spa_l2cache.sav_count; i++) { 9480 vd = spa->spa_l2cache.sav_vdevs[i]; 9481 if (vd->vdev_guid == guid) 9482 return (vd); 9483 } 9484 9485 for (i = 0; i < spa->spa_spares.sav_count; i++) { 9486 vd = spa->spa_spares.sav_vdevs[i]; 9487 if (vd->vdev_guid == guid) 9488 return (vd); 9489 } 9490 } 9491 9492 return (NULL); 9493 } 9494 9495 void 9496 spa_upgrade(spa_t *spa, uint64_t version) 9497 { 9498 ASSERT(spa_writeable(spa)); 9499 9500 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 9501 9502 /* 9503 * This should only be called for a non-faulted pool, and since a 9504 * future version would result in an unopenable pool, this shouldn't be 9505 * possible. 9506 */ 9507 ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version)); 9508 ASSERT3U(version, >=, spa->spa_uberblock.ub_version); 9509 9510 spa->spa_uberblock.ub_version = version; 9511 vdev_config_dirty(spa->spa_root_vdev); 9512 9513 spa_config_exit(spa, SCL_ALL, FTAG); 9514 9515 txg_wait_synced(spa_get_dsl(spa), 0); 9516 } 9517 9518 static boolean_t 9519 spa_has_aux_vdev(spa_t *spa, uint64_t guid, spa_aux_vdev_t *sav) 9520 { 9521 (void) spa; 9522 int i; 9523 uint64_t vdev_guid; 9524 9525 for (i = 0; i < sav->sav_count; i++) 9526 if (sav->sav_vdevs[i]->vdev_guid == guid) 9527 return (B_TRUE); 9528 9529 for (i = 0; i < sav->sav_npending; i++) { 9530 if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID, 9531 &vdev_guid) == 0 && vdev_guid == guid) 9532 return (B_TRUE); 9533 } 9534 9535 return (B_FALSE); 9536 } 9537 9538 boolean_t 9539 spa_has_l2cache(spa_t *spa, uint64_t guid) 9540 { 9541 return (spa_has_aux_vdev(spa, guid, &spa->spa_l2cache)); 9542 } 9543 9544 boolean_t 9545 spa_has_spare(spa_t *spa, uint64_t guid) 9546 { 9547 return (spa_has_aux_vdev(spa, guid, &spa->spa_spares)); 9548 } 9549 9550 /* 9551 * Check if a pool has an active shared spare device. 9552 * Note: reference count of an active spare is 2, as a spare and as a replace 9553 */ 9554 static boolean_t 9555 spa_has_active_shared_spare(spa_t *spa) 9556 { 9557 int i, refcnt; 9558 uint64_t pool; 9559 spa_aux_vdev_t *sav = &spa->spa_spares; 9560 9561 for (i = 0; i < sav->sav_count; i++) { 9562 if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool, 9563 &refcnt) && pool != 0ULL && pool == spa_guid(spa) && 9564 refcnt > 2) 9565 return (B_TRUE); 9566 } 9567 9568 return (B_FALSE); 9569 } 9570 9571 uint64_t 9572 spa_total_metaslabs(spa_t *spa) 9573 { 9574 vdev_t *rvd = spa->spa_root_vdev; 9575 9576 uint64_t m = 0; 9577 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 9578 vdev_t *vd = rvd->vdev_child[c]; 9579 if (!vdev_is_concrete(vd)) 9580 continue; 9581 m += vd->vdev_ms_count; 9582 } 9583 return (m); 9584 } 9585 9586 /* 9587 * Notify any waiting threads that some activity has switched from being in- 9588 * progress to not-in-progress so that the thread can wake up and determine 9589 * whether it is finished waiting. 9590 */ 9591 void 9592 spa_notify_waiters(spa_t *spa) 9593 { 9594 /* 9595 * Acquiring spa_activities_lock here prevents the cv_broadcast from 9596 * happening between the waiting thread's check and cv_wait. 9597 */ 9598 mutex_enter(&spa->spa_activities_lock); 9599 cv_broadcast(&spa->spa_activities_cv); 9600 mutex_exit(&spa->spa_activities_lock); 9601 } 9602 9603 /* 9604 * Notify any waiting threads that the pool is exporting, and then block until 9605 * they are finished using the spa_t. 9606 */ 9607 void 9608 spa_wake_waiters(spa_t *spa) 9609 { 9610 mutex_enter(&spa->spa_activities_lock); 9611 spa->spa_waiters_cancel = B_TRUE; 9612 cv_broadcast(&spa->spa_activities_cv); 9613 while (spa->spa_waiters != 0) 9614 cv_wait(&spa->spa_waiters_cv, &spa->spa_activities_lock); 9615 spa->spa_waiters_cancel = B_FALSE; 9616 mutex_exit(&spa->spa_activities_lock); 9617 } 9618 9619 /* Whether the vdev or any of its descendants are being initialized/trimmed. */ 9620 static boolean_t 9621 spa_vdev_activity_in_progress_impl(vdev_t *vd, zpool_wait_activity_t activity) 9622 { 9623 spa_t *spa = vd->vdev_spa; 9624 9625 ASSERT(spa_config_held(spa, SCL_CONFIG | SCL_STATE, RW_READER)); 9626 ASSERT(MUTEX_HELD(&spa->spa_activities_lock)); 9627 ASSERT(activity == ZPOOL_WAIT_INITIALIZE || 9628 activity == ZPOOL_WAIT_TRIM); 9629 9630 kmutex_t *lock = activity == ZPOOL_WAIT_INITIALIZE ? 9631 &vd->vdev_initialize_lock : &vd->vdev_trim_lock; 9632 9633 mutex_exit(&spa->spa_activities_lock); 9634 mutex_enter(lock); 9635 mutex_enter(&spa->spa_activities_lock); 9636 9637 boolean_t in_progress = (activity == ZPOOL_WAIT_INITIALIZE) ? 9638 (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) : 9639 (vd->vdev_trim_state == VDEV_TRIM_ACTIVE); 9640 mutex_exit(lock); 9641 9642 if (in_progress) 9643 return (B_TRUE); 9644 9645 for (int i = 0; i < vd->vdev_children; i++) { 9646 if (spa_vdev_activity_in_progress_impl(vd->vdev_child[i], 9647 activity)) 9648 return (B_TRUE); 9649 } 9650 9651 return (B_FALSE); 9652 } 9653 9654 /* 9655 * If use_guid is true, this checks whether the vdev specified by guid is 9656 * being initialized/trimmed. Otherwise, it checks whether any vdev in the pool 9657 * is being initialized/trimmed. The caller must hold the config lock and 9658 * spa_activities_lock. 9659 */ 9660 static int 9661 spa_vdev_activity_in_progress(spa_t *spa, boolean_t use_guid, uint64_t guid, 9662 zpool_wait_activity_t activity, boolean_t *in_progress) 9663 { 9664 mutex_exit(&spa->spa_activities_lock); 9665 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 9666 mutex_enter(&spa->spa_activities_lock); 9667 9668 vdev_t *vd; 9669 if (use_guid) { 9670 vd = spa_lookup_by_guid(spa, guid, B_FALSE); 9671 if (vd == NULL || !vd->vdev_ops->vdev_op_leaf) { 9672 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9673 return (EINVAL); 9674 } 9675 } else { 9676 vd = spa->spa_root_vdev; 9677 } 9678 9679 *in_progress = spa_vdev_activity_in_progress_impl(vd, activity); 9680 9681 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9682 return (0); 9683 } 9684 9685 /* 9686 * Locking for waiting threads 9687 * --------------------------- 9688 * 9689 * Waiting threads need a way to check whether a given activity is in progress, 9690 * and then, if it is, wait for it to complete. Each activity will have some 9691 * in-memory representation of the relevant on-disk state which can be used to 9692 * determine whether or not the activity is in progress. The in-memory state and 9693 * the locking used to protect it will be different for each activity, and may 9694 * not be suitable for use with a cvar (e.g., some state is protected by the 9695 * config lock). To allow waiting threads to wait without any races, another 9696 * lock, spa_activities_lock, is used. 9697 * 9698 * When the state is checked, both the activity-specific lock (if there is one) 9699 * and spa_activities_lock are held. In some cases, the activity-specific lock 9700 * is acquired explicitly (e.g. the config lock). In others, the locking is 9701 * internal to some check (e.g. bpobj_is_empty). After checking, the waiting 9702 * thread releases the activity-specific lock and, if the activity is in 9703 * progress, then cv_waits using spa_activities_lock. 9704 * 9705 * The waiting thread is woken when another thread, one completing some 9706 * activity, updates the state of the activity and then calls 9707 * spa_notify_waiters, which will cv_broadcast. This 'completing' thread only 9708 * needs to hold its activity-specific lock when updating the state, and this 9709 * lock can (but doesn't have to) be dropped before calling spa_notify_waiters. 9710 * 9711 * Because spa_notify_waiters acquires spa_activities_lock before broadcasting, 9712 * and because it is held when the waiting thread checks the state of the 9713 * activity, it can never be the case that the completing thread both updates 9714 * the activity state and cv_broadcasts in between the waiting thread's check 9715 * and cv_wait. Thus, a waiting thread can never miss a wakeup. 9716 * 9717 * In order to prevent deadlock, when the waiting thread does its check, in some 9718 * cases it will temporarily drop spa_activities_lock in order to acquire the 9719 * activity-specific lock. The order in which spa_activities_lock and the 9720 * activity specific lock are acquired in the waiting thread is determined by 9721 * the order in which they are acquired in the completing thread; if the 9722 * completing thread calls spa_notify_waiters with the activity-specific lock 9723 * held, then the waiting thread must also acquire the activity-specific lock 9724 * first. 9725 */ 9726 9727 static int 9728 spa_activity_in_progress(spa_t *spa, zpool_wait_activity_t activity, 9729 boolean_t use_tag, uint64_t tag, boolean_t *in_progress) 9730 { 9731 int error = 0; 9732 9733 ASSERT(MUTEX_HELD(&spa->spa_activities_lock)); 9734 9735 switch (activity) { 9736 case ZPOOL_WAIT_CKPT_DISCARD: 9737 *in_progress = 9738 (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT) && 9739 zap_contains(spa_meta_objset(spa), 9740 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT) == 9741 ENOENT); 9742 break; 9743 case ZPOOL_WAIT_FREE: 9744 *in_progress = ((spa_version(spa) >= SPA_VERSION_DEADLISTS && 9745 !bpobj_is_empty(&spa->spa_dsl_pool->dp_free_bpobj)) || 9746 spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY) || 9747 spa_livelist_delete_check(spa)); 9748 break; 9749 case ZPOOL_WAIT_INITIALIZE: 9750 case ZPOOL_WAIT_TRIM: 9751 error = spa_vdev_activity_in_progress(spa, use_tag, tag, 9752 activity, in_progress); 9753 break; 9754 case ZPOOL_WAIT_REPLACE: 9755 mutex_exit(&spa->spa_activities_lock); 9756 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 9757 mutex_enter(&spa->spa_activities_lock); 9758 9759 *in_progress = vdev_replace_in_progress(spa->spa_root_vdev); 9760 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9761 break; 9762 case ZPOOL_WAIT_REMOVE: 9763 *in_progress = (spa->spa_removing_phys.sr_state == 9764 DSS_SCANNING); 9765 break; 9766 case ZPOOL_WAIT_RESILVER: 9767 if ((*in_progress = vdev_rebuild_active(spa->spa_root_vdev))) 9768 break; 9769 zfs_fallthrough; 9770 case ZPOOL_WAIT_SCRUB: 9771 { 9772 boolean_t scanning, paused, is_scrub; 9773 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan; 9774 9775 is_scrub = (scn->scn_phys.scn_func == POOL_SCAN_SCRUB); 9776 scanning = (scn->scn_phys.scn_state == DSS_SCANNING); 9777 paused = dsl_scan_is_paused_scrub(scn); 9778 *in_progress = (scanning && !paused && 9779 is_scrub == (activity == ZPOOL_WAIT_SCRUB)); 9780 break; 9781 } 9782 default: 9783 panic("unrecognized value for activity %d", activity); 9784 } 9785 9786 return (error); 9787 } 9788 9789 static int 9790 spa_wait_common(const char *pool, zpool_wait_activity_t activity, 9791 boolean_t use_tag, uint64_t tag, boolean_t *waited) 9792 { 9793 /* 9794 * The tag is used to distinguish between instances of an activity. 9795 * 'initialize' and 'trim' are the only activities that we use this for. 9796 * The other activities can only have a single instance in progress in a 9797 * pool at one time, making the tag unnecessary. 9798 * 9799 * There can be multiple devices being replaced at once, but since they 9800 * all finish once resilvering finishes, we don't bother keeping track 9801 * of them individually, we just wait for them all to finish. 9802 */ 9803 if (use_tag && activity != ZPOOL_WAIT_INITIALIZE && 9804 activity != ZPOOL_WAIT_TRIM) 9805 return (EINVAL); 9806 9807 if (activity < 0 || activity >= ZPOOL_WAIT_NUM_ACTIVITIES) 9808 return (EINVAL); 9809 9810 spa_t *spa; 9811 int error = spa_open(pool, &spa, FTAG); 9812 if (error != 0) 9813 return (error); 9814 9815 /* 9816 * Increment the spa's waiter count so that we can call spa_close and 9817 * still ensure that the spa_t doesn't get freed before this thread is 9818 * finished with it when the pool is exported. We want to call spa_close 9819 * before we start waiting because otherwise the additional ref would 9820 * prevent the pool from being exported or destroyed throughout the 9821 * potentially long wait. 9822 */ 9823 mutex_enter(&spa->spa_activities_lock); 9824 spa->spa_waiters++; 9825 spa_close(spa, FTAG); 9826 9827 *waited = B_FALSE; 9828 for (;;) { 9829 boolean_t in_progress; 9830 error = spa_activity_in_progress(spa, activity, use_tag, tag, 9831 &in_progress); 9832 9833 if (error || !in_progress || spa->spa_waiters_cancel) 9834 break; 9835 9836 *waited = B_TRUE; 9837 9838 if (cv_wait_sig(&spa->spa_activities_cv, 9839 &spa->spa_activities_lock) == 0) { 9840 error = EINTR; 9841 break; 9842 } 9843 } 9844 9845 spa->spa_waiters--; 9846 cv_signal(&spa->spa_waiters_cv); 9847 mutex_exit(&spa->spa_activities_lock); 9848 9849 return (error); 9850 } 9851 9852 /* 9853 * Wait for a particular instance of the specified activity to complete, where 9854 * the instance is identified by 'tag' 9855 */ 9856 int 9857 spa_wait_tag(const char *pool, zpool_wait_activity_t activity, uint64_t tag, 9858 boolean_t *waited) 9859 { 9860 return (spa_wait_common(pool, activity, B_TRUE, tag, waited)); 9861 } 9862 9863 /* 9864 * Wait for all instances of the specified activity complete 9865 */ 9866 int 9867 spa_wait(const char *pool, zpool_wait_activity_t activity, boolean_t *waited) 9868 { 9869 9870 return (spa_wait_common(pool, activity, B_FALSE, 0, waited)); 9871 } 9872 9873 sysevent_t * 9874 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name) 9875 { 9876 sysevent_t *ev = NULL; 9877 #ifdef _KERNEL 9878 nvlist_t *resource; 9879 9880 resource = zfs_event_create(spa, vd, FM_SYSEVENT_CLASS, name, hist_nvl); 9881 if (resource) { 9882 ev = kmem_alloc(sizeof (sysevent_t), KM_SLEEP); 9883 ev->resource = resource; 9884 } 9885 #else 9886 (void) spa, (void) vd, (void) hist_nvl, (void) name; 9887 #endif 9888 return (ev); 9889 } 9890 9891 void 9892 spa_event_post(sysevent_t *ev) 9893 { 9894 #ifdef _KERNEL 9895 if (ev) { 9896 zfs_zevent_post(ev->resource, NULL, zfs_zevent_post_cb); 9897 kmem_free(ev, sizeof (*ev)); 9898 } 9899 #else 9900 (void) ev; 9901 #endif 9902 } 9903 9904 /* 9905 * Post a zevent corresponding to the given sysevent. The 'name' must be one 9906 * of the event definitions in sys/sysevent/eventdefs.h. The payload will be 9907 * filled in from the spa and (optionally) the vdev. This doesn't do anything 9908 * in the userland libzpool, as we don't want consumers to misinterpret ztest 9909 * or zdb as real changes. 9910 */ 9911 void 9912 spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name) 9913 { 9914 spa_event_post(spa_event_create(spa, vd, hist_nvl, name)); 9915 } 9916 9917 /* state manipulation functions */ 9918 EXPORT_SYMBOL(spa_open); 9919 EXPORT_SYMBOL(spa_open_rewind); 9920 EXPORT_SYMBOL(spa_get_stats); 9921 EXPORT_SYMBOL(spa_create); 9922 EXPORT_SYMBOL(spa_import); 9923 EXPORT_SYMBOL(spa_tryimport); 9924 EXPORT_SYMBOL(spa_destroy); 9925 EXPORT_SYMBOL(spa_export); 9926 EXPORT_SYMBOL(spa_reset); 9927 EXPORT_SYMBOL(spa_async_request); 9928 EXPORT_SYMBOL(spa_async_suspend); 9929 EXPORT_SYMBOL(spa_async_resume); 9930 EXPORT_SYMBOL(spa_inject_addref); 9931 EXPORT_SYMBOL(spa_inject_delref); 9932 EXPORT_SYMBOL(spa_scan_stat_init); 9933 EXPORT_SYMBOL(spa_scan_get_stats); 9934 9935 /* device manipulation */ 9936 EXPORT_SYMBOL(spa_vdev_add); 9937 EXPORT_SYMBOL(spa_vdev_attach); 9938 EXPORT_SYMBOL(spa_vdev_detach); 9939 EXPORT_SYMBOL(spa_vdev_setpath); 9940 EXPORT_SYMBOL(spa_vdev_setfru); 9941 EXPORT_SYMBOL(spa_vdev_split_mirror); 9942 9943 /* spare statech is global across all pools) */ 9944 EXPORT_SYMBOL(spa_spare_add); 9945 EXPORT_SYMBOL(spa_spare_remove); 9946 EXPORT_SYMBOL(spa_spare_exists); 9947 EXPORT_SYMBOL(spa_spare_activate); 9948 9949 /* L2ARC statech is global across all pools) */ 9950 EXPORT_SYMBOL(spa_l2cache_add); 9951 EXPORT_SYMBOL(spa_l2cache_remove); 9952 EXPORT_SYMBOL(spa_l2cache_exists); 9953 EXPORT_SYMBOL(spa_l2cache_activate); 9954 EXPORT_SYMBOL(spa_l2cache_drop); 9955 9956 /* scanning */ 9957 EXPORT_SYMBOL(spa_scan); 9958 EXPORT_SYMBOL(spa_scan_stop); 9959 9960 /* spa syncing */ 9961 EXPORT_SYMBOL(spa_sync); /* only for DMU use */ 9962 EXPORT_SYMBOL(spa_sync_allpools); 9963 9964 /* properties */ 9965 EXPORT_SYMBOL(spa_prop_set); 9966 EXPORT_SYMBOL(spa_prop_get); 9967 EXPORT_SYMBOL(spa_prop_clear_bootfs); 9968 9969 /* asynchronous event notification */ 9970 EXPORT_SYMBOL(spa_event_notify); 9971 9972 /* BEGIN CSTYLED */ 9973 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_shift, INT, ZMOD_RW, 9974 "log2 fraction of arc that can be used by inflight I/Os when " 9975 "verifying pool during import"); 9976 /* END CSTYLED */ 9977 9978 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_metadata, INT, ZMOD_RW, 9979 "Set to traverse metadata on pool import"); 9980 9981 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW, 9982 "Set to traverse data on pool import"); 9983 9984 ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW, 9985 "Print vdev tree to zfs_dbgmsg during pool import"); 9986 9987 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD, 9988 "Percentage of CPUs to run an IO worker thread"); 9989 9990 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_tpq, UINT, ZMOD_RD, 9991 "Number of threads per IO worker taskqueue"); 9992 9993 /* BEGIN CSTYLED */ 9994 ZFS_MODULE_PARAM(zfs, zfs_, max_missing_tvds, ULONG, ZMOD_RW, 9995 "Allow importing pool with up to this number of missing top-level " 9996 "vdevs (in read-only mode)"); 9997 /* END CSTYLED */ 9998 9999 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_pause, INT, 10000 ZMOD_RW, "Set the livelist condense zthr to pause"); 10001 10002 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_pause, INT, 10003 ZMOD_RW, "Set the livelist condense synctask to pause"); 10004 10005 /* BEGIN CSTYLED */ 10006 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_cancel, 10007 INT, ZMOD_RW, 10008 "Whether livelist condensing was canceled in the synctask"); 10009 10010 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_cancel, 10011 INT, ZMOD_RW, 10012 "Whether livelist condensing was canceled in the zthr function"); 10013 10014 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT, 10015 ZMOD_RW, 10016 "Whether extra ALLOC blkptrs were added to a livelist entry while it " 10017 "was being condensed"); 10018 /* END CSTYLED */ 10019