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 */ 36 37 /* 38 * SPA: Storage Pool Allocator 39 * 40 * This file contains all the routines used when modifying on-disk SPA state. 41 * This includes opening, importing, destroying, exporting a pool, and syncing a 42 * pool. 43 */ 44 45 #include <sys/zfs_context.h> 46 #include <sys/fm/fs/zfs.h> 47 #include <sys/spa_impl.h> 48 #include <sys/zio.h> 49 #include <sys/zio_checksum.h> 50 #include <sys/dmu.h> 51 #include <sys/dmu_tx.h> 52 #include <sys/zap.h> 53 #include <sys/zil.h> 54 #include <sys/ddt.h> 55 #include <sys/vdev_impl.h> 56 #include <sys/vdev_removal.h> 57 #include <sys/vdev_indirect_mapping.h> 58 #include <sys/vdev_indirect_births.h> 59 #include <sys/vdev_initialize.h> 60 #include <sys/vdev_rebuild.h> 61 #include <sys/vdev_trim.h> 62 #include <sys/vdev_disk.h> 63 #include <sys/vdev_draid.h> 64 #include <sys/metaslab.h> 65 #include <sys/metaslab_impl.h> 66 #include <sys/mmp.h> 67 #include <sys/uberblock_impl.h> 68 #include <sys/txg.h> 69 #include <sys/avl.h> 70 #include <sys/bpobj.h> 71 #include <sys/dmu_traverse.h> 72 #include <sys/dmu_objset.h> 73 #include <sys/unique.h> 74 #include <sys/dsl_pool.h> 75 #include <sys/dsl_dataset.h> 76 #include <sys/dsl_dir.h> 77 #include <sys/dsl_prop.h> 78 #include <sys/dsl_synctask.h> 79 #include <sys/fs/zfs.h> 80 #include <sys/arc.h> 81 #include <sys/callb.h> 82 #include <sys/systeminfo.h> 83 #include <sys/spa_boot.h> 84 #include <sys/zfs_ioctl.h> 85 #include <sys/dsl_scan.h> 86 #include <sys/zfeature.h> 87 #include <sys/dsl_destroy.h> 88 #include <sys/zvol.h> 89 90 #ifdef _KERNEL 91 #include <sys/fm/protocol.h> 92 #include <sys/fm/util.h> 93 #include <sys/callb.h> 94 #include <sys/zone.h> 95 #include <sys/vmsystm.h> 96 #endif /* _KERNEL */ 97 98 #include "zfs_prop.h" 99 #include "zfs_comutil.h" 100 101 /* 102 * The interval, in seconds, at which failed configuration cache file writes 103 * should be retried. 104 */ 105 int zfs_ccw_retry_interval = 300; 106 107 typedef enum zti_modes { 108 ZTI_MODE_FIXED, /* value is # of threads (min 1) */ 109 ZTI_MODE_BATCH, /* cpu-intensive; value is ignored */ 110 ZTI_MODE_NULL, /* don't create a taskq */ 111 ZTI_NMODES 112 } zti_modes_t; 113 114 #define ZTI_P(n, q) { ZTI_MODE_FIXED, (n), (q) } 115 #define ZTI_PCT(n) { ZTI_MODE_ONLINE_PERCENT, (n), 1 } 116 #define ZTI_BATCH { ZTI_MODE_BATCH, 0, 1 } 117 #define ZTI_NULL { ZTI_MODE_NULL, 0, 0 } 118 119 #define ZTI_N(n) ZTI_P(n, 1) 120 #define ZTI_ONE ZTI_N(1) 121 122 typedef struct zio_taskq_info { 123 zti_modes_t zti_mode; 124 uint_t zti_value; 125 uint_t zti_count; 126 } zio_taskq_info_t; 127 128 static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = { 129 "iss", "iss_h", "int", "int_h" 130 }; 131 132 /* 133 * This table defines the taskq settings for each ZFS I/O type. When 134 * initializing a pool, we use this table to create an appropriately sized 135 * taskq. Some operations are low volume and therefore have a small, static 136 * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE 137 * macros. Other operations process a large amount of data; the ZTI_BATCH 138 * macro causes us to create a taskq oriented for throughput. Some operations 139 * are so high frequency and short-lived that the taskq itself can become a 140 * point of lock contention. The ZTI_P(#, #) macro indicates that we need an 141 * additional degree of parallelism specified by the number of threads per- 142 * taskq and the number of taskqs; when dispatching an event in this case, the 143 * particular taskq is chosen at random. 144 * 145 * The different taskq priorities are to handle the different contexts (issue 146 * and interrupt) and then to reserve threads for ZIO_PRIORITY_NOW I/Os that 147 * need to be handled with minimum delay. 148 */ 149 const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = { 150 /* ISSUE ISSUE_HIGH INTR INTR_HIGH */ 151 { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* NULL */ 152 { ZTI_N(8), ZTI_NULL, ZTI_P(12, 8), ZTI_NULL }, /* READ */ 153 { ZTI_BATCH, ZTI_N(5), ZTI_P(12, 8), ZTI_N(5) }, /* WRITE */ 154 { ZTI_P(12, 8), ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* FREE */ 155 { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* CLAIM */ 156 { ZTI_ONE, ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* IOCTL */ 157 { ZTI_N(4), ZTI_NULL, ZTI_ONE, ZTI_NULL }, /* TRIM */ 158 }; 159 160 static void spa_sync_version(void *arg, dmu_tx_t *tx); 161 static void spa_sync_props(void *arg, dmu_tx_t *tx); 162 static boolean_t spa_has_active_shared_spare(spa_t *spa); 163 static int spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport); 164 static void spa_vdev_resilver_done(spa_t *spa); 165 166 uint_t zio_taskq_batch_pct = 75; /* 1 thread per cpu in pset */ 167 boolean_t zio_taskq_sysdc = B_TRUE; /* use SDC scheduling class */ 168 uint_t zio_taskq_basedc = 80; /* base duty cycle */ 169 170 boolean_t spa_create_process = B_TRUE; /* no process ==> no sysdc */ 171 172 /* 173 * Report any spa_load_verify errors found, but do not fail spa_load. 174 * This is used by zdb to analyze non-idle pools. 175 */ 176 boolean_t spa_load_verify_dryrun = B_FALSE; 177 178 /* 179 * This (illegal) pool name is used when temporarily importing a spa_t in order 180 * to get the vdev stats associated with the imported devices. 181 */ 182 #define TRYIMPORT_NAME "$import" 183 184 /* 185 * For debugging purposes: print out vdev tree during pool import. 186 */ 187 int spa_load_print_vdev_tree = B_FALSE; 188 189 /* 190 * A non-zero value for zfs_max_missing_tvds means that we allow importing 191 * pools with missing top-level vdevs. This is strictly intended for advanced 192 * pool recovery cases since missing data is almost inevitable. Pools with 193 * missing devices can only be imported read-only for safety reasons, and their 194 * fail-mode will be automatically set to "continue". 195 * 196 * With 1 missing vdev we should be able to import the pool and mount all 197 * datasets. User data that was not modified after the missing device has been 198 * added should be recoverable. This means that snapshots created prior to the 199 * addition of that device should be completely intact. 200 * 201 * With 2 missing vdevs, some datasets may fail to mount since there are 202 * dataset statistics that are stored as regular metadata. Some data might be 203 * recoverable if those vdevs were added recently. 204 * 205 * With 3 or more missing vdevs, the pool is severely damaged and MOS entries 206 * may be missing entirely. Chances of data recovery are very low. Note that 207 * there are also risks of performing an inadvertent rewind as we might be 208 * missing all the vdevs with the latest uberblocks. 209 */ 210 unsigned long zfs_max_missing_tvds = 0; 211 212 /* 213 * The parameters below are similar to zfs_max_missing_tvds but are only 214 * intended for a preliminary open of the pool with an untrusted config which 215 * might be incomplete or out-dated. 216 * 217 * We are more tolerant for pools opened from a cachefile since we could have 218 * an out-dated cachefile where a device removal was not registered. 219 * We could have set the limit arbitrarily high but in the case where devices 220 * are really missing we would want to return the proper error codes; we chose 221 * SPA_DVAS_PER_BP - 1 so that some copies of the MOS would still be available 222 * and we get a chance to retrieve the trusted config. 223 */ 224 uint64_t zfs_max_missing_tvds_cachefile = SPA_DVAS_PER_BP - 1; 225 226 /* 227 * In the case where config was assembled by scanning device paths (/dev/dsks 228 * by default) we are less tolerant since all the existing devices should have 229 * been detected and we want spa_load to return the right error codes. 230 */ 231 uint64_t zfs_max_missing_tvds_scan = 0; 232 233 /* 234 * Debugging aid that pauses spa_sync() towards the end. 235 */ 236 boolean_t zfs_pause_spa_sync = B_FALSE; 237 238 /* 239 * Variables to indicate the livelist condense zthr func should wait at certain 240 * points for the livelist to be removed - used to test condense/destroy races 241 */ 242 int zfs_livelist_condense_zthr_pause = 0; 243 int zfs_livelist_condense_sync_pause = 0; 244 245 /* 246 * Variables to track whether or not condense cancellation has been 247 * triggered in testing. 248 */ 249 int zfs_livelist_condense_sync_cancel = 0; 250 int zfs_livelist_condense_zthr_cancel = 0; 251 252 /* 253 * Variable to track whether or not extra ALLOC blkptrs were added to a 254 * livelist entry while it was being condensed (caused by the way we track 255 * remapped blkptrs in dbuf_remap_impl) 256 */ 257 int zfs_livelist_condense_new_alloc = 0; 258 259 /* 260 * ========================================================================== 261 * SPA properties routines 262 * ========================================================================== 263 */ 264 265 /* 266 * Add a (source=src, propname=propval) list to an nvlist. 267 */ 268 static void 269 spa_prop_add_list(nvlist_t *nvl, zpool_prop_t prop, char *strval, 270 uint64_t intval, zprop_source_t src) 271 { 272 const char *propname = zpool_prop_to_name(prop); 273 nvlist_t *propval; 274 275 VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0); 276 VERIFY(nvlist_add_uint64(propval, ZPROP_SOURCE, src) == 0); 277 278 if (strval != NULL) 279 VERIFY(nvlist_add_string(propval, ZPROP_VALUE, strval) == 0); 280 else 281 VERIFY(nvlist_add_uint64(propval, ZPROP_VALUE, intval) == 0); 282 283 VERIFY(nvlist_add_nvlist(nvl, propname, propval) == 0); 284 nvlist_free(propval); 285 } 286 287 /* 288 * Get property values from the spa configuration. 289 */ 290 static void 291 spa_prop_get_config(spa_t *spa, nvlist_t **nvp) 292 { 293 vdev_t *rvd = spa->spa_root_vdev; 294 dsl_pool_t *pool = spa->spa_dsl_pool; 295 uint64_t size, alloc, cap, version; 296 const zprop_source_t src = ZPROP_SRC_NONE; 297 spa_config_dirent_t *dp; 298 metaslab_class_t *mc = spa_normal_class(spa); 299 300 ASSERT(MUTEX_HELD(&spa->spa_props_lock)); 301 302 if (rvd != NULL) { 303 alloc = metaslab_class_get_alloc(mc); 304 alloc += metaslab_class_get_alloc(spa_special_class(spa)); 305 alloc += metaslab_class_get_alloc(spa_dedup_class(spa)); 306 307 size = metaslab_class_get_space(mc); 308 size += metaslab_class_get_space(spa_special_class(spa)); 309 size += metaslab_class_get_space(spa_dedup_class(spa)); 310 311 spa_prop_add_list(*nvp, ZPOOL_PROP_NAME, spa_name(spa), 0, src); 312 spa_prop_add_list(*nvp, ZPOOL_PROP_SIZE, NULL, size, src); 313 spa_prop_add_list(*nvp, ZPOOL_PROP_ALLOCATED, NULL, alloc, src); 314 spa_prop_add_list(*nvp, ZPOOL_PROP_FREE, NULL, 315 size - alloc, src); 316 spa_prop_add_list(*nvp, ZPOOL_PROP_CHECKPOINT, NULL, 317 spa->spa_checkpoint_info.sci_dspace, src); 318 319 spa_prop_add_list(*nvp, ZPOOL_PROP_FRAGMENTATION, NULL, 320 metaslab_class_fragmentation(mc), src); 321 spa_prop_add_list(*nvp, ZPOOL_PROP_EXPANDSZ, NULL, 322 metaslab_class_expandable_space(mc), src); 323 spa_prop_add_list(*nvp, ZPOOL_PROP_READONLY, NULL, 324 (spa_mode(spa) == SPA_MODE_READ), src); 325 326 cap = (size == 0) ? 0 : (alloc * 100 / size); 327 spa_prop_add_list(*nvp, ZPOOL_PROP_CAPACITY, NULL, cap, src); 328 329 spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL, 330 ddt_get_pool_dedup_ratio(spa), src); 331 332 spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL, 333 rvd->vdev_state, src); 334 335 version = spa_version(spa); 336 if (version == zpool_prop_default_numeric(ZPOOL_PROP_VERSION)) { 337 spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, 338 version, ZPROP_SRC_DEFAULT); 339 } else { 340 spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, 341 version, ZPROP_SRC_LOCAL); 342 } 343 spa_prop_add_list(*nvp, ZPOOL_PROP_LOAD_GUID, 344 NULL, spa_load_guid(spa), src); 345 } 346 347 if (pool != NULL) { 348 /* 349 * The $FREE directory was introduced in SPA_VERSION_DEADLISTS, 350 * when opening pools before this version freedir will be NULL. 351 */ 352 if (pool->dp_free_dir != NULL) { 353 spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL, 354 dsl_dir_phys(pool->dp_free_dir)->dd_used_bytes, 355 src); 356 } else { 357 spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, 358 NULL, 0, src); 359 } 360 361 if (pool->dp_leak_dir != NULL) { 362 spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL, 363 dsl_dir_phys(pool->dp_leak_dir)->dd_used_bytes, 364 src); 365 } else { 366 spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, 367 NULL, 0, src); 368 } 369 } 370 371 spa_prop_add_list(*nvp, ZPOOL_PROP_GUID, NULL, spa_guid(spa), src); 372 373 if (spa->spa_comment != NULL) { 374 spa_prop_add_list(*nvp, ZPOOL_PROP_COMMENT, spa->spa_comment, 375 0, ZPROP_SRC_LOCAL); 376 } 377 378 if (spa->spa_root != NULL) 379 spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root, 380 0, ZPROP_SRC_LOCAL); 381 382 if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) { 383 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL, 384 MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE); 385 } else { 386 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL, 387 SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE); 388 } 389 390 if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE)) { 391 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL, 392 DNODE_MAX_SIZE, ZPROP_SRC_NONE); 393 } else { 394 spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL, 395 DNODE_MIN_SIZE, ZPROP_SRC_NONE); 396 } 397 398 if ((dp = list_head(&spa->spa_config_list)) != NULL) { 399 if (dp->scd_path == NULL) { 400 spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE, 401 "none", 0, ZPROP_SRC_LOCAL); 402 } else if (strcmp(dp->scd_path, spa_config_path) != 0) { 403 spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE, 404 dp->scd_path, 0, ZPROP_SRC_LOCAL); 405 } 406 } 407 } 408 409 /* 410 * Get zpool property values. 411 */ 412 int 413 spa_prop_get(spa_t *spa, nvlist_t **nvp) 414 { 415 objset_t *mos = spa->spa_meta_objset; 416 zap_cursor_t zc; 417 zap_attribute_t za; 418 dsl_pool_t *dp; 419 int err; 420 421 err = nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP); 422 if (err) 423 return (err); 424 425 dp = spa_get_dsl(spa); 426 dsl_pool_config_enter(dp, FTAG); 427 mutex_enter(&spa->spa_props_lock); 428 429 /* 430 * Get properties from the spa config. 431 */ 432 spa_prop_get_config(spa, nvp); 433 434 /* If no pool property object, no more prop to get. */ 435 if (mos == NULL || spa->spa_pool_props_object == 0) 436 goto out; 437 438 /* 439 * Get properties from the MOS pool property object. 440 */ 441 for (zap_cursor_init(&zc, mos, spa->spa_pool_props_object); 442 (err = zap_cursor_retrieve(&zc, &za)) == 0; 443 zap_cursor_advance(&zc)) { 444 uint64_t intval = 0; 445 char *strval = NULL; 446 zprop_source_t src = ZPROP_SRC_DEFAULT; 447 zpool_prop_t prop; 448 449 if ((prop = zpool_name_to_prop(za.za_name)) == ZPOOL_PROP_INVAL) 450 continue; 451 452 switch (za.za_integer_length) { 453 case 8: 454 /* integer property */ 455 if (za.za_first_integer != 456 zpool_prop_default_numeric(prop)) 457 src = ZPROP_SRC_LOCAL; 458 459 if (prop == ZPOOL_PROP_BOOTFS) { 460 dsl_dataset_t *ds = NULL; 461 462 err = dsl_dataset_hold_obj(dp, 463 za.za_first_integer, FTAG, &ds); 464 if (err != 0) 465 break; 466 467 strval = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN, 468 KM_SLEEP); 469 dsl_dataset_name(ds, strval); 470 dsl_dataset_rele(ds, FTAG); 471 } else { 472 strval = NULL; 473 intval = za.za_first_integer; 474 } 475 476 spa_prop_add_list(*nvp, prop, strval, intval, src); 477 478 if (strval != NULL) 479 kmem_free(strval, ZFS_MAX_DATASET_NAME_LEN); 480 481 break; 482 483 case 1: 484 /* string property */ 485 strval = kmem_alloc(za.za_num_integers, KM_SLEEP); 486 err = zap_lookup(mos, spa->spa_pool_props_object, 487 za.za_name, 1, za.za_num_integers, strval); 488 if (err) { 489 kmem_free(strval, za.za_num_integers); 490 break; 491 } 492 spa_prop_add_list(*nvp, prop, strval, 0, src); 493 kmem_free(strval, za.za_num_integers); 494 break; 495 496 default: 497 break; 498 } 499 } 500 zap_cursor_fini(&zc); 501 out: 502 mutex_exit(&spa->spa_props_lock); 503 dsl_pool_config_exit(dp, FTAG); 504 if (err && err != ENOENT) { 505 nvlist_free(*nvp); 506 *nvp = NULL; 507 return (err); 508 } 509 510 return (0); 511 } 512 513 /* 514 * Validate the given pool properties nvlist and modify the list 515 * for the property values to be set. 516 */ 517 static int 518 spa_prop_validate(spa_t *spa, nvlist_t *props) 519 { 520 nvpair_t *elem; 521 int error = 0, reset_bootfs = 0; 522 uint64_t objnum = 0; 523 boolean_t has_feature = B_FALSE; 524 525 elem = NULL; 526 while ((elem = nvlist_next_nvpair(props, elem)) != NULL) { 527 uint64_t intval; 528 char *strval, *slash, *check, *fname; 529 const char *propname = nvpair_name(elem); 530 zpool_prop_t prop = zpool_name_to_prop(propname); 531 532 switch (prop) { 533 case ZPOOL_PROP_INVAL: 534 if (!zpool_prop_feature(propname)) { 535 error = SET_ERROR(EINVAL); 536 break; 537 } 538 539 /* 540 * Sanitize the input. 541 */ 542 if (nvpair_type(elem) != DATA_TYPE_UINT64) { 543 error = SET_ERROR(EINVAL); 544 break; 545 } 546 547 if (nvpair_value_uint64(elem, &intval) != 0) { 548 error = SET_ERROR(EINVAL); 549 break; 550 } 551 552 if (intval != 0) { 553 error = SET_ERROR(EINVAL); 554 break; 555 } 556 557 fname = strchr(propname, '@') + 1; 558 if (zfeature_lookup_name(fname, NULL) != 0) { 559 error = SET_ERROR(EINVAL); 560 break; 561 } 562 563 has_feature = B_TRUE; 564 break; 565 566 case ZPOOL_PROP_VERSION: 567 error = nvpair_value_uint64(elem, &intval); 568 if (!error && 569 (intval < spa_version(spa) || 570 intval > SPA_VERSION_BEFORE_FEATURES || 571 has_feature)) 572 error = SET_ERROR(EINVAL); 573 break; 574 575 case ZPOOL_PROP_DELEGATION: 576 case ZPOOL_PROP_AUTOREPLACE: 577 case ZPOOL_PROP_LISTSNAPS: 578 case ZPOOL_PROP_AUTOEXPAND: 579 case ZPOOL_PROP_AUTOTRIM: 580 error = nvpair_value_uint64(elem, &intval); 581 if (!error && intval > 1) 582 error = SET_ERROR(EINVAL); 583 break; 584 585 case ZPOOL_PROP_MULTIHOST: 586 error = nvpair_value_uint64(elem, &intval); 587 if (!error && intval > 1) 588 error = SET_ERROR(EINVAL); 589 590 if (!error) { 591 uint32_t hostid = zone_get_hostid(NULL); 592 if (hostid) 593 spa->spa_hostid = hostid; 594 else 595 error = SET_ERROR(ENOTSUP); 596 } 597 598 break; 599 600 case ZPOOL_PROP_BOOTFS: 601 /* 602 * If the pool version is less than SPA_VERSION_BOOTFS, 603 * or the pool is still being created (version == 0), 604 * the bootfs property cannot be set. 605 */ 606 if (spa_version(spa) < SPA_VERSION_BOOTFS) { 607 error = SET_ERROR(ENOTSUP); 608 break; 609 } 610 611 /* 612 * Make sure the vdev config is bootable 613 */ 614 if (!vdev_is_bootable(spa->spa_root_vdev)) { 615 error = SET_ERROR(ENOTSUP); 616 break; 617 } 618 619 reset_bootfs = 1; 620 621 error = nvpair_value_string(elem, &strval); 622 623 if (!error) { 624 objset_t *os; 625 626 if (strval == NULL || strval[0] == '\0') { 627 objnum = zpool_prop_default_numeric( 628 ZPOOL_PROP_BOOTFS); 629 break; 630 } 631 632 error = dmu_objset_hold(strval, FTAG, &os); 633 if (error != 0) 634 break; 635 636 /* Must be ZPL. */ 637 if (dmu_objset_type(os) != DMU_OST_ZFS) { 638 error = SET_ERROR(ENOTSUP); 639 } else { 640 objnum = dmu_objset_id(os); 641 } 642 dmu_objset_rele(os, FTAG); 643 } 644 break; 645 646 case ZPOOL_PROP_FAILUREMODE: 647 error = nvpair_value_uint64(elem, &intval); 648 if (!error && intval > ZIO_FAILURE_MODE_PANIC) 649 error = SET_ERROR(EINVAL); 650 651 /* 652 * This is a special case which only occurs when 653 * the pool has completely failed. This allows 654 * the user to change the in-core failmode property 655 * without syncing it out to disk (I/Os might 656 * currently be blocked). We do this by returning 657 * EIO to the caller (spa_prop_set) to trick it 658 * into thinking we encountered a property validation 659 * error. 660 */ 661 if (!error && spa_suspended(spa)) { 662 spa->spa_failmode = intval; 663 error = SET_ERROR(EIO); 664 } 665 break; 666 667 case ZPOOL_PROP_CACHEFILE: 668 if ((error = nvpair_value_string(elem, &strval)) != 0) 669 break; 670 671 if (strval[0] == '\0') 672 break; 673 674 if (strcmp(strval, "none") == 0) 675 break; 676 677 if (strval[0] != '/') { 678 error = SET_ERROR(EINVAL); 679 break; 680 } 681 682 slash = strrchr(strval, '/'); 683 ASSERT(slash != NULL); 684 685 if (slash[1] == '\0' || strcmp(slash, "/.") == 0 || 686 strcmp(slash, "/..") == 0) 687 error = SET_ERROR(EINVAL); 688 break; 689 690 case ZPOOL_PROP_COMMENT: 691 if ((error = nvpair_value_string(elem, &strval)) != 0) 692 break; 693 for (check = strval; *check != '\0'; check++) { 694 if (!isprint(*check)) { 695 error = SET_ERROR(EINVAL); 696 break; 697 } 698 } 699 if (strlen(strval) > ZPROP_MAX_COMMENT) 700 error = SET_ERROR(E2BIG); 701 break; 702 703 default: 704 break; 705 } 706 707 if (error) 708 break; 709 } 710 711 (void) nvlist_remove_all(props, 712 zpool_prop_to_name(ZPOOL_PROP_DEDUPDITTO)); 713 714 if (!error && reset_bootfs) { 715 error = nvlist_remove(props, 716 zpool_prop_to_name(ZPOOL_PROP_BOOTFS), DATA_TYPE_STRING); 717 718 if (!error) { 719 error = nvlist_add_uint64(props, 720 zpool_prop_to_name(ZPOOL_PROP_BOOTFS), objnum); 721 } 722 } 723 724 return (error); 725 } 726 727 void 728 spa_configfile_set(spa_t *spa, nvlist_t *nvp, boolean_t need_sync) 729 { 730 char *cachefile; 731 spa_config_dirent_t *dp; 732 733 if (nvlist_lookup_string(nvp, zpool_prop_to_name(ZPOOL_PROP_CACHEFILE), 734 &cachefile) != 0) 735 return; 736 737 dp = kmem_alloc(sizeof (spa_config_dirent_t), 738 KM_SLEEP); 739 740 if (cachefile[0] == '\0') 741 dp->scd_path = spa_strdup(spa_config_path); 742 else if (strcmp(cachefile, "none") == 0) 743 dp->scd_path = NULL; 744 else 745 dp->scd_path = spa_strdup(cachefile); 746 747 list_insert_head(&spa->spa_config_list, dp); 748 if (need_sync) 749 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 750 } 751 752 int 753 spa_prop_set(spa_t *spa, nvlist_t *nvp) 754 { 755 int error; 756 nvpair_t *elem = NULL; 757 boolean_t need_sync = B_FALSE; 758 759 if ((error = spa_prop_validate(spa, nvp)) != 0) 760 return (error); 761 762 while ((elem = nvlist_next_nvpair(nvp, elem)) != NULL) { 763 zpool_prop_t prop = zpool_name_to_prop(nvpair_name(elem)); 764 765 if (prop == ZPOOL_PROP_CACHEFILE || 766 prop == ZPOOL_PROP_ALTROOT || 767 prop == ZPOOL_PROP_READONLY) 768 continue; 769 770 if (prop == ZPOOL_PROP_VERSION || prop == ZPOOL_PROP_INVAL) { 771 uint64_t ver; 772 773 if (prop == ZPOOL_PROP_VERSION) { 774 VERIFY(nvpair_value_uint64(elem, &ver) == 0); 775 } else { 776 ASSERT(zpool_prop_feature(nvpair_name(elem))); 777 ver = SPA_VERSION_FEATURES; 778 need_sync = B_TRUE; 779 } 780 781 /* Save time if the version is already set. */ 782 if (ver == spa_version(spa)) 783 continue; 784 785 /* 786 * In addition to the pool directory object, we might 787 * create the pool properties object, the features for 788 * read object, the features for write object, or the 789 * feature descriptions object. 790 */ 791 error = dsl_sync_task(spa->spa_name, NULL, 792 spa_sync_version, &ver, 793 6, ZFS_SPACE_CHECK_RESERVED); 794 if (error) 795 return (error); 796 continue; 797 } 798 799 need_sync = B_TRUE; 800 break; 801 } 802 803 if (need_sync) { 804 return (dsl_sync_task(spa->spa_name, NULL, spa_sync_props, 805 nvp, 6, ZFS_SPACE_CHECK_RESERVED)); 806 } 807 808 return (0); 809 } 810 811 /* 812 * If the bootfs property value is dsobj, clear it. 813 */ 814 void 815 spa_prop_clear_bootfs(spa_t *spa, uint64_t dsobj, dmu_tx_t *tx) 816 { 817 if (spa->spa_bootfs == dsobj && spa->spa_pool_props_object != 0) { 818 VERIFY(zap_remove(spa->spa_meta_objset, 819 spa->spa_pool_props_object, 820 zpool_prop_to_name(ZPOOL_PROP_BOOTFS), tx) == 0); 821 spa->spa_bootfs = 0; 822 } 823 } 824 825 /*ARGSUSED*/ 826 static int 827 spa_change_guid_check(void *arg, dmu_tx_t *tx) 828 { 829 uint64_t *newguid __maybe_unused = arg; 830 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 831 vdev_t *rvd = spa->spa_root_vdev; 832 uint64_t vdev_state; 833 834 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 835 int error = (spa_has_checkpoint(spa)) ? 836 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 837 return (SET_ERROR(error)); 838 } 839 840 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 841 vdev_state = rvd->vdev_state; 842 spa_config_exit(spa, SCL_STATE, FTAG); 843 844 if (vdev_state != VDEV_STATE_HEALTHY) 845 return (SET_ERROR(ENXIO)); 846 847 ASSERT3U(spa_guid(spa), !=, *newguid); 848 849 return (0); 850 } 851 852 static void 853 spa_change_guid_sync(void *arg, dmu_tx_t *tx) 854 { 855 uint64_t *newguid = arg; 856 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 857 uint64_t oldguid; 858 vdev_t *rvd = spa->spa_root_vdev; 859 860 oldguid = spa_guid(spa); 861 862 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 863 rvd->vdev_guid = *newguid; 864 rvd->vdev_guid_sum += (*newguid - oldguid); 865 vdev_config_dirty(rvd); 866 spa_config_exit(spa, SCL_STATE, FTAG); 867 868 spa_history_log_internal(spa, "guid change", tx, "old=%llu new=%llu", 869 (u_longlong_t)oldguid, (u_longlong_t)*newguid); 870 } 871 872 /* 873 * Change the GUID for the pool. This is done so that we can later 874 * re-import a pool built from a clone of our own vdevs. We will modify 875 * the root vdev's guid, our own pool guid, and then mark all of our 876 * vdevs dirty. Note that we must make sure that all our vdevs are 877 * online when we do this, or else any vdevs that weren't present 878 * would be orphaned from our pool. We are also going to issue a 879 * sysevent to update any watchers. 880 */ 881 int 882 spa_change_guid(spa_t *spa) 883 { 884 int error; 885 uint64_t guid; 886 887 mutex_enter(&spa->spa_vdev_top_lock); 888 mutex_enter(&spa_namespace_lock); 889 guid = spa_generate_guid(NULL); 890 891 error = dsl_sync_task(spa->spa_name, spa_change_guid_check, 892 spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED); 893 894 if (error == 0) { 895 spa_write_cachefile(spa, B_FALSE, B_TRUE); 896 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID); 897 } 898 899 mutex_exit(&spa_namespace_lock); 900 mutex_exit(&spa->spa_vdev_top_lock); 901 902 return (error); 903 } 904 905 /* 906 * ========================================================================== 907 * SPA state manipulation (open/create/destroy/import/export) 908 * ========================================================================== 909 */ 910 911 static int 912 spa_error_entry_compare(const void *a, const void *b) 913 { 914 const spa_error_entry_t *sa = (const spa_error_entry_t *)a; 915 const spa_error_entry_t *sb = (const spa_error_entry_t *)b; 916 int ret; 917 918 ret = memcmp(&sa->se_bookmark, &sb->se_bookmark, 919 sizeof (zbookmark_phys_t)); 920 921 return (TREE_ISIGN(ret)); 922 } 923 924 /* 925 * Utility function which retrieves copies of the current logs and 926 * re-initializes them in the process. 927 */ 928 void 929 spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub) 930 { 931 ASSERT(MUTEX_HELD(&spa->spa_errlist_lock)); 932 933 bcopy(&spa->spa_errlist_last, last, sizeof (avl_tree_t)); 934 bcopy(&spa->spa_errlist_scrub, scrub, sizeof (avl_tree_t)); 935 936 avl_create(&spa->spa_errlist_scrub, 937 spa_error_entry_compare, sizeof (spa_error_entry_t), 938 offsetof(spa_error_entry_t, se_avl)); 939 avl_create(&spa->spa_errlist_last, 940 spa_error_entry_compare, sizeof (spa_error_entry_t), 941 offsetof(spa_error_entry_t, se_avl)); 942 } 943 944 static void 945 spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q) 946 { 947 const zio_taskq_info_t *ztip = &zio_taskqs[t][q]; 948 enum zti_modes mode = ztip->zti_mode; 949 uint_t value = ztip->zti_value; 950 uint_t count = ztip->zti_count; 951 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 952 uint_t flags = 0; 953 boolean_t batch = B_FALSE; 954 955 if (mode == ZTI_MODE_NULL) { 956 tqs->stqs_count = 0; 957 tqs->stqs_taskq = NULL; 958 return; 959 } 960 961 ASSERT3U(count, >, 0); 962 963 tqs->stqs_count = count; 964 tqs->stqs_taskq = kmem_alloc(count * sizeof (taskq_t *), KM_SLEEP); 965 966 switch (mode) { 967 case ZTI_MODE_FIXED: 968 ASSERT3U(value, >=, 1); 969 value = MAX(value, 1); 970 flags |= TASKQ_DYNAMIC; 971 break; 972 973 case ZTI_MODE_BATCH: 974 batch = B_TRUE; 975 flags |= TASKQ_THREADS_CPU_PCT; 976 value = MIN(zio_taskq_batch_pct, 100); 977 break; 978 979 default: 980 panic("unrecognized mode for %s_%s taskq (%u:%u) in " 981 "spa_activate()", 982 zio_type_name[t], zio_taskq_types[q], mode, value); 983 break; 984 } 985 986 for (uint_t i = 0; i < count; i++) { 987 taskq_t *tq; 988 char name[32]; 989 990 (void) snprintf(name, sizeof (name), "%s_%s", 991 zio_type_name[t], zio_taskq_types[q]); 992 993 if (zio_taskq_sysdc && spa->spa_proc != &p0) { 994 if (batch) 995 flags |= TASKQ_DC_BATCH; 996 997 tq = taskq_create_sysdc(name, value, 50, INT_MAX, 998 spa->spa_proc, zio_taskq_basedc, flags); 999 } else { 1000 pri_t pri = maxclsyspri; 1001 /* 1002 * The write issue taskq can be extremely CPU 1003 * intensive. Run it at slightly less important 1004 * priority than the other taskqs. 1005 * 1006 * Under Linux and FreeBSD this means incrementing 1007 * the priority value as opposed to platforms like 1008 * illumos where it should be decremented. 1009 * 1010 * On FreeBSD, if priorities divided by four (RQ_PPQ) 1011 * are equal then a difference between them is 1012 * insignificant. 1013 */ 1014 if (t == ZIO_TYPE_WRITE && q == ZIO_TASKQ_ISSUE) { 1015 #if defined(__linux__) 1016 pri++; 1017 #elif defined(__FreeBSD__) 1018 pri += 4; 1019 #else 1020 #error "unknown OS" 1021 #endif 1022 } 1023 tq = taskq_create_proc(name, value, pri, 50, 1024 INT_MAX, spa->spa_proc, flags); 1025 } 1026 1027 tqs->stqs_taskq[i] = tq; 1028 } 1029 } 1030 1031 static void 1032 spa_taskqs_fini(spa_t *spa, zio_type_t t, zio_taskq_type_t q) 1033 { 1034 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1035 1036 if (tqs->stqs_taskq == NULL) { 1037 ASSERT3U(tqs->stqs_count, ==, 0); 1038 return; 1039 } 1040 1041 for (uint_t i = 0; i < tqs->stqs_count; i++) { 1042 ASSERT3P(tqs->stqs_taskq[i], !=, NULL); 1043 taskq_destroy(tqs->stqs_taskq[i]); 1044 } 1045 1046 kmem_free(tqs->stqs_taskq, tqs->stqs_count * sizeof (taskq_t *)); 1047 tqs->stqs_taskq = NULL; 1048 } 1049 1050 /* 1051 * Dispatch a task to the appropriate taskq for the ZFS I/O type and priority. 1052 * Note that a type may have multiple discrete taskqs to avoid lock contention 1053 * on the taskq itself. In that case we choose which taskq at random by using 1054 * the low bits of gethrtime(). 1055 */ 1056 void 1057 spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q, 1058 task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent) 1059 { 1060 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1061 taskq_t *tq; 1062 1063 ASSERT3P(tqs->stqs_taskq, !=, NULL); 1064 ASSERT3U(tqs->stqs_count, !=, 0); 1065 1066 if (tqs->stqs_count == 1) { 1067 tq = tqs->stqs_taskq[0]; 1068 } else { 1069 tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count]; 1070 } 1071 1072 taskq_dispatch_ent(tq, func, arg, flags, ent); 1073 } 1074 1075 /* 1076 * Same as spa_taskq_dispatch_ent() but block on the task until completion. 1077 */ 1078 void 1079 spa_taskq_dispatch_sync(spa_t *spa, zio_type_t t, zio_taskq_type_t q, 1080 task_func_t *func, void *arg, uint_t flags) 1081 { 1082 spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; 1083 taskq_t *tq; 1084 taskqid_t id; 1085 1086 ASSERT3P(tqs->stqs_taskq, !=, NULL); 1087 ASSERT3U(tqs->stqs_count, !=, 0); 1088 1089 if (tqs->stqs_count == 1) { 1090 tq = tqs->stqs_taskq[0]; 1091 } else { 1092 tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count]; 1093 } 1094 1095 id = taskq_dispatch(tq, func, arg, flags); 1096 if (id) 1097 taskq_wait_id(tq, id); 1098 } 1099 1100 static void 1101 spa_create_zio_taskqs(spa_t *spa) 1102 { 1103 for (int t = 0; t < ZIO_TYPES; t++) { 1104 for (int q = 0; q < ZIO_TASKQ_TYPES; q++) { 1105 spa_taskqs_init(spa, t, q); 1106 } 1107 } 1108 } 1109 1110 /* 1111 * Disabled until spa_thread() can be adapted for Linux. 1112 */ 1113 #undef HAVE_SPA_THREAD 1114 1115 #if defined(_KERNEL) && defined(HAVE_SPA_THREAD) 1116 static void 1117 spa_thread(void *arg) 1118 { 1119 psetid_t zio_taskq_psrset_bind = PS_NONE; 1120 callb_cpr_t cprinfo; 1121 1122 spa_t *spa = arg; 1123 user_t *pu = PTOU(curproc); 1124 1125 CALLB_CPR_INIT(&cprinfo, &spa->spa_proc_lock, callb_generic_cpr, 1126 spa->spa_name); 1127 1128 ASSERT(curproc != &p0); 1129 (void) snprintf(pu->u_psargs, sizeof (pu->u_psargs), 1130 "zpool-%s", spa->spa_name); 1131 (void) strlcpy(pu->u_comm, pu->u_psargs, sizeof (pu->u_comm)); 1132 1133 /* bind this thread to the requested psrset */ 1134 if (zio_taskq_psrset_bind != PS_NONE) { 1135 pool_lock(); 1136 mutex_enter(&cpu_lock); 1137 mutex_enter(&pidlock); 1138 mutex_enter(&curproc->p_lock); 1139 1140 if (cpupart_bind_thread(curthread, zio_taskq_psrset_bind, 1141 0, NULL, NULL) == 0) { 1142 curthread->t_bind_pset = zio_taskq_psrset_bind; 1143 } else { 1144 cmn_err(CE_WARN, 1145 "Couldn't bind process for zfs pool \"%s\" to " 1146 "pset %d\n", spa->spa_name, zio_taskq_psrset_bind); 1147 } 1148 1149 mutex_exit(&curproc->p_lock); 1150 mutex_exit(&pidlock); 1151 mutex_exit(&cpu_lock); 1152 pool_unlock(); 1153 } 1154 1155 if (zio_taskq_sysdc) { 1156 sysdc_thread_enter(curthread, 100, 0); 1157 } 1158 1159 spa->spa_proc = curproc; 1160 spa->spa_did = curthread->t_did; 1161 1162 spa_create_zio_taskqs(spa); 1163 1164 mutex_enter(&spa->spa_proc_lock); 1165 ASSERT(spa->spa_proc_state == SPA_PROC_CREATED); 1166 1167 spa->spa_proc_state = SPA_PROC_ACTIVE; 1168 cv_broadcast(&spa->spa_proc_cv); 1169 1170 CALLB_CPR_SAFE_BEGIN(&cprinfo); 1171 while (spa->spa_proc_state == SPA_PROC_ACTIVE) 1172 cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); 1173 CALLB_CPR_SAFE_END(&cprinfo, &spa->spa_proc_lock); 1174 1175 ASSERT(spa->spa_proc_state == SPA_PROC_DEACTIVATE); 1176 spa->spa_proc_state = SPA_PROC_GONE; 1177 spa->spa_proc = &p0; 1178 cv_broadcast(&spa->spa_proc_cv); 1179 CALLB_CPR_EXIT(&cprinfo); /* drops spa_proc_lock */ 1180 1181 mutex_enter(&curproc->p_lock); 1182 lwp_exit(); 1183 } 1184 #endif 1185 1186 /* 1187 * Activate an uninitialized pool. 1188 */ 1189 static void 1190 spa_activate(spa_t *spa, spa_mode_t mode) 1191 { 1192 ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED); 1193 1194 spa->spa_state = POOL_STATE_ACTIVE; 1195 spa->spa_mode = mode; 1196 1197 spa->spa_normal_class = metaslab_class_create(spa, zfs_metaslab_ops); 1198 spa->spa_log_class = metaslab_class_create(spa, zfs_metaslab_ops); 1199 spa->spa_special_class = metaslab_class_create(spa, zfs_metaslab_ops); 1200 spa->spa_dedup_class = metaslab_class_create(spa, zfs_metaslab_ops); 1201 1202 /* Try to create a covering process */ 1203 mutex_enter(&spa->spa_proc_lock); 1204 ASSERT(spa->spa_proc_state == SPA_PROC_NONE); 1205 ASSERT(spa->spa_proc == &p0); 1206 spa->spa_did = 0; 1207 1208 #ifdef HAVE_SPA_THREAD 1209 /* Only create a process if we're going to be around a while. */ 1210 if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) { 1211 if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri, 1212 NULL, 0) == 0) { 1213 spa->spa_proc_state = SPA_PROC_CREATED; 1214 while (spa->spa_proc_state == SPA_PROC_CREATED) { 1215 cv_wait(&spa->spa_proc_cv, 1216 &spa->spa_proc_lock); 1217 } 1218 ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE); 1219 ASSERT(spa->spa_proc != &p0); 1220 ASSERT(spa->spa_did != 0); 1221 } else { 1222 #ifdef _KERNEL 1223 cmn_err(CE_WARN, 1224 "Couldn't create process for zfs pool \"%s\"\n", 1225 spa->spa_name); 1226 #endif 1227 } 1228 } 1229 #endif /* HAVE_SPA_THREAD */ 1230 mutex_exit(&spa->spa_proc_lock); 1231 1232 /* If we didn't create a process, we need to create our taskqs. */ 1233 if (spa->spa_proc == &p0) { 1234 spa_create_zio_taskqs(spa); 1235 } 1236 1237 for (size_t i = 0; i < TXG_SIZE; i++) { 1238 spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL, 1239 ZIO_FLAG_CANFAIL); 1240 } 1241 1242 list_create(&spa->spa_config_dirty_list, sizeof (vdev_t), 1243 offsetof(vdev_t, vdev_config_dirty_node)); 1244 list_create(&spa->spa_evicting_os_list, sizeof (objset_t), 1245 offsetof(objset_t, os_evicting_node)); 1246 list_create(&spa->spa_state_dirty_list, sizeof (vdev_t), 1247 offsetof(vdev_t, vdev_state_dirty_node)); 1248 1249 txg_list_create(&spa->spa_vdev_txg_list, spa, 1250 offsetof(struct vdev, vdev_txg_node)); 1251 1252 avl_create(&spa->spa_errlist_scrub, 1253 spa_error_entry_compare, sizeof (spa_error_entry_t), 1254 offsetof(spa_error_entry_t, se_avl)); 1255 avl_create(&spa->spa_errlist_last, 1256 spa_error_entry_compare, sizeof (spa_error_entry_t), 1257 offsetof(spa_error_entry_t, se_avl)); 1258 1259 spa_keystore_init(&spa->spa_keystore); 1260 1261 /* 1262 * This taskq is used to perform zvol-minor-related tasks 1263 * asynchronously. This has several advantages, including easy 1264 * resolution of various deadlocks. 1265 * 1266 * The taskq must be single threaded to ensure tasks are always 1267 * processed in the order in which they were dispatched. 1268 * 1269 * A taskq per pool allows one to keep the pools independent. 1270 * This way if one pool is suspended, it will not impact another. 1271 * 1272 * The preferred location to dispatch a zvol minor task is a sync 1273 * task. In this context, there is easy access to the spa_t and minimal 1274 * error handling is required because the sync task must succeed. 1275 */ 1276 spa->spa_zvol_taskq = taskq_create("z_zvol", 1, defclsyspri, 1277 1, INT_MAX, 0); 1278 1279 /* 1280 * Taskq dedicated to prefetcher threads: this is used to prevent the 1281 * pool traverse code from monopolizing the global (and limited) 1282 * system_taskq by inappropriately scheduling long running tasks on it. 1283 */ 1284 spa->spa_prefetch_taskq = taskq_create("z_prefetch", 100, 1285 defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 1286 1287 /* 1288 * The taskq to upgrade datasets in this pool. Currently used by 1289 * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA. 1290 */ 1291 spa->spa_upgrade_taskq = taskq_create("z_upgrade", 100, 1292 defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT); 1293 } 1294 1295 /* 1296 * Opposite of spa_activate(). 1297 */ 1298 static void 1299 spa_deactivate(spa_t *spa) 1300 { 1301 ASSERT(spa->spa_sync_on == B_FALSE); 1302 ASSERT(spa->spa_dsl_pool == NULL); 1303 ASSERT(spa->spa_root_vdev == NULL); 1304 ASSERT(spa->spa_async_zio_root == NULL); 1305 ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED); 1306 1307 spa_evicting_os_wait(spa); 1308 1309 if (spa->spa_zvol_taskq) { 1310 taskq_destroy(spa->spa_zvol_taskq); 1311 spa->spa_zvol_taskq = NULL; 1312 } 1313 1314 if (spa->spa_prefetch_taskq) { 1315 taskq_destroy(spa->spa_prefetch_taskq); 1316 spa->spa_prefetch_taskq = NULL; 1317 } 1318 1319 if (spa->spa_upgrade_taskq) { 1320 taskq_destroy(spa->spa_upgrade_taskq); 1321 spa->spa_upgrade_taskq = NULL; 1322 } 1323 1324 txg_list_destroy(&spa->spa_vdev_txg_list); 1325 1326 list_destroy(&spa->spa_config_dirty_list); 1327 list_destroy(&spa->spa_evicting_os_list); 1328 list_destroy(&spa->spa_state_dirty_list); 1329 1330 taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); 1331 1332 for (int t = 0; t < ZIO_TYPES; t++) { 1333 for (int q = 0; q < ZIO_TASKQ_TYPES; q++) { 1334 spa_taskqs_fini(spa, t, q); 1335 } 1336 } 1337 1338 for (size_t i = 0; i < TXG_SIZE; i++) { 1339 ASSERT3P(spa->spa_txg_zio[i], !=, NULL); 1340 VERIFY0(zio_wait(spa->spa_txg_zio[i])); 1341 spa->spa_txg_zio[i] = NULL; 1342 } 1343 1344 metaslab_class_destroy(spa->spa_normal_class); 1345 spa->spa_normal_class = NULL; 1346 1347 metaslab_class_destroy(spa->spa_log_class); 1348 spa->spa_log_class = NULL; 1349 1350 metaslab_class_destroy(spa->spa_special_class); 1351 spa->spa_special_class = NULL; 1352 1353 metaslab_class_destroy(spa->spa_dedup_class); 1354 spa->spa_dedup_class = NULL; 1355 1356 /* 1357 * If this was part of an import or the open otherwise failed, we may 1358 * still have errors left in the queues. Empty them just in case. 1359 */ 1360 spa_errlog_drain(spa); 1361 avl_destroy(&spa->spa_errlist_scrub); 1362 avl_destroy(&spa->spa_errlist_last); 1363 1364 spa_keystore_fini(&spa->spa_keystore); 1365 1366 spa->spa_state = POOL_STATE_UNINITIALIZED; 1367 1368 mutex_enter(&spa->spa_proc_lock); 1369 if (spa->spa_proc_state != SPA_PROC_NONE) { 1370 ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE); 1371 spa->spa_proc_state = SPA_PROC_DEACTIVATE; 1372 cv_broadcast(&spa->spa_proc_cv); 1373 while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) { 1374 ASSERT(spa->spa_proc != &p0); 1375 cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock); 1376 } 1377 ASSERT(spa->spa_proc_state == SPA_PROC_GONE); 1378 spa->spa_proc_state = SPA_PROC_NONE; 1379 } 1380 ASSERT(spa->spa_proc == &p0); 1381 mutex_exit(&spa->spa_proc_lock); 1382 1383 /* 1384 * We want to make sure spa_thread() has actually exited the ZFS 1385 * module, so that the module can't be unloaded out from underneath 1386 * it. 1387 */ 1388 if (spa->spa_did != 0) { 1389 thread_join(spa->spa_did); 1390 spa->spa_did = 0; 1391 } 1392 } 1393 1394 /* 1395 * Verify a pool configuration, and construct the vdev tree appropriately. This 1396 * will create all the necessary vdevs in the appropriate layout, with each vdev 1397 * in the CLOSED state. This will prep the pool before open/creation/import. 1398 * All vdev validation is done by the vdev_alloc() routine. 1399 */ 1400 int 1401 spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, 1402 uint_t id, int atype) 1403 { 1404 nvlist_t **child; 1405 uint_t children; 1406 int error; 1407 1408 if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0) 1409 return (error); 1410 1411 if ((*vdp)->vdev_ops->vdev_op_leaf) 1412 return (0); 1413 1414 error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN, 1415 &child, &children); 1416 1417 if (error == ENOENT) 1418 return (0); 1419 1420 if (error) { 1421 vdev_free(*vdp); 1422 *vdp = NULL; 1423 return (SET_ERROR(EINVAL)); 1424 } 1425 1426 for (int c = 0; c < children; c++) { 1427 vdev_t *vd; 1428 if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c, 1429 atype)) != 0) { 1430 vdev_free(*vdp); 1431 *vdp = NULL; 1432 return (error); 1433 } 1434 } 1435 1436 ASSERT(*vdp != NULL); 1437 1438 return (0); 1439 } 1440 1441 static boolean_t 1442 spa_should_flush_logs_on_unload(spa_t *spa) 1443 { 1444 if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) 1445 return (B_FALSE); 1446 1447 if (!spa_writeable(spa)) 1448 return (B_FALSE); 1449 1450 if (!spa->spa_sync_on) 1451 return (B_FALSE); 1452 1453 if (spa_state(spa) != POOL_STATE_EXPORTED) 1454 return (B_FALSE); 1455 1456 if (zfs_keep_log_spacemaps_at_export) 1457 return (B_FALSE); 1458 1459 return (B_TRUE); 1460 } 1461 1462 /* 1463 * Opens a transaction that will set the flag that will instruct 1464 * spa_sync to attempt to flush all the metaslabs for that txg. 1465 */ 1466 static void 1467 spa_unload_log_sm_flush_all(spa_t *spa) 1468 { 1469 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 1470 VERIFY0(dmu_tx_assign(tx, TXG_WAIT)); 1471 1472 ASSERT3U(spa->spa_log_flushall_txg, ==, 0); 1473 spa->spa_log_flushall_txg = dmu_tx_get_txg(tx); 1474 1475 dmu_tx_commit(tx); 1476 txg_wait_synced(spa_get_dsl(spa), spa->spa_log_flushall_txg); 1477 } 1478 1479 static void 1480 spa_unload_log_sm_metadata(spa_t *spa) 1481 { 1482 void *cookie = NULL; 1483 spa_log_sm_t *sls; 1484 while ((sls = avl_destroy_nodes(&spa->spa_sm_logs_by_txg, 1485 &cookie)) != NULL) { 1486 VERIFY0(sls->sls_mscount); 1487 kmem_free(sls, sizeof (spa_log_sm_t)); 1488 } 1489 1490 for (log_summary_entry_t *e = list_head(&spa->spa_log_summary); 1491 e != NULL; e = list_head(&spa->spa_log_summary)) { 1492 VERIFY0(e->lse_mscount); 1493 list_remove(&spa->spa_log_summary, e); 1494 kmem_free(e, sizeof (log_summary_entry_t)); 1495 } 1496 1497 spa->spa_unflushed_stats.sus_nblocks = 0; 1498 spa->spa_unflushed_stats.sus_memused = 0; 1499 spa->spa_unflushed_stats.sus_blocklimit = 0; 1500 } 1501 1502 static void 1503 spa_destroy_aux_threads(spa_t *spa) 1504 { 1505 if (spa->spa_condense_zthr != NULL) { 1506 zthr_destroy(spa->spa_condense_zthr); 1507 spa->spa_condense_zthr = NULL; 1508 } 1509 if (spa->spa_checkpoint_discard_zthr != NULL) { 1510 zthr_destroy(spa->spa_checkpoint_discard_zthr); 1511 spa->spa_checkpoint_discard_zthr = NULL; 1512 } 1513 if (spa->spa_livelist_delete_zthr != NULL) { 1514 zthr_destroy(spa->spa_livelist_delete_zthr); 1515 spa->spa_livelist_delete_zthr = NULL; 1516 } 1517 if (spa->spa_livelist_condense_zthr != NULL) { 1518 zthr_destroy(spa->spa_livelist_condense_zthr); 1519 spa->spa_livelist_condense_zthr = NULL; 1520 } 1521 } 1522 1523 /* 1524 * Opposite of spa_load(). 1525 */ 1526 static void 1527 spa_unload(spa_t *spa) 1528 { 1529 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 1530 ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED); 1531 1532 spa_import_progress_remove(spa_guid(spa)); 1533 spa_load_note(spa, "UNLOADING"); 1534 1535 spa_wake_waiters(spa); 1536 1537 /* 1538 * If the log space map feature is enabled and the pool is getting 1539 * exported (but not destroyed), we want to spend some time flushing 1540 * as many metaslabs as we can in an attempt to destroy log space 1541 * maps and save import time. 1542 */ 1543 if (spa_should_flush_logs_on_unload(spa)) 1544 spa_unload_log_sm_flush_all(spa); 1545 1546 /* 1547 * Stop async tasks. 1548 */ 1549 spa_async_suspend(spa); 1550 1551 if (spa->spa_root_vdev) { 1552 vdev_t *root_vdev = spa->spa_root_vdev; 1553 vdev_initialize_stop_all(root_vdev, VDEV_INITIALIZE_ACTIVE); 1554 vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE); 1555 vdev_autotrim_stop_all(spa); 1556 vdev_rebuild_stop_all(spa); 1557 } 1558 1559 /* 1560 * Stop syncing. 1561 */ 1562 if (spa->spa_sync_on) { 1563 txg_sync_stop(spa->spa_dsl_pool); 1564 spa->spa_sync_on = B_FALSE; 1565 } 1566 1567 /* 1568 * This ensures that there is no async metaslab prefetching 1569 * while we attempt to unload the spa. 1570 */ 1571 if (spa->spa_root_vdev != NULL) { 1572 for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++) { 1573 vdev_t *vc = spa->spa_root_vdev->vdev_child[c]; 1574 if (vc->vdev_mg != NULL) 1575 taskq_wait(vc->vdev_mg->mg_taskq); 1576 } 1577 } 1578 1579 if (spa->spa_mmp.mmp_thread) 1580 mmp_thread_stop(spa); 1581 1582 /* 1583 * Wait for any outstanding async I/O to complete. 1584 */ 1585 if (spa->spa_async_zio_root != NULL) { 1586 for (int i = 0; i < max_ncpus; i++) 1587 (void) zio_wait(spa->spa_async_zio_root[i]); 1588 kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *)); 1589 spa->spa_async_zio_root = NULL; 1590 } 1591 1592 if (spa->spa_vdev_removal != NULL) { 1593 spa_vdev_removal_destroy(spa->spa_vdev_removal); 1594 spa->spa_vdev_removal = NULL; 1595 } 1596 1597 spa_destroy_aux_threads(spa); 1598 1599 spa_condense_fini(spa); 1600 1601 bpobj_close(&spa->spa_deferred_bpobj); 1602 1603 spa_config_enter(spa, SCL_ALL, spa, RW_WRITER); 1604 1605 /* 1606 * Close all vdevs. 1607 */ 1608 if (spa->spa_root_vdev) 1609 vdev_free(spa->spa_root_vdev); 1610 ASSERT(spa->spa_root_vdev == NULL); 1611 1612 /* 1613 * Close the dsl pool. 1614 */ 1615 if (spa->spa_dsl_pool) { 1616 dsl_pool_close(spa->spa_dsl_pool); 1617 spa->spa_dsl_pool = NULL; 1618 spa->spa_meta_objset = NULL; 1619 } 1620 1621 ddt_unload(spa); 1622 spa_unload_log_sm_metadata(spa); 1623 1624 /* 1625 * Drop and purge level 2 cache 1626 */ 1627 spa_l2cache_drop(spa); 1628 1629 for (int i = 0; i < spa->spa_spares.sav_count; i++) 1630 vdev_free(spa->spa_spares.sav_vdevs[i]); 1631 if (spa->spa_spares.sav_vdevs) { 1632 kmem_free(spa->spa_spares.sav_vdevs, 1633 spa->spa_spares.sav_count * sizeof (void *)); 1634 spa->spa_spares.sav_vdevs = NULL; 1635 } 1636 if (spa->spa_spares.sav_config) { 1637 nvlist_free(spa->spa_spares.sav_config); 1638 spa->spa_spares.sav_config = NULL; 1639 } 1640 spa->spa_spares.sav_count = 0; 1641 1642 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) { 1643 vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]); 1644 vdev_free(spa->spa_l2cache.sav_vdevs[i]); 1645 } 1646 if (spa->spa_l2cache.sav_vdevs) { 1647 kmem_free(spa->spa_l2cache.sav_vdevs, 1648 spa->spa_l2cache.sav_count * sizeof (void *)); 1649 spa->spa_l2cache.sav_vdevs = NULL; 1650 } 1651 if (spa->spa_l2cache.sav_config) { 1652 nvlist_free(spa->spa_l2cache.sav_config); 1653 spa->spa_l2cache.sav_config = NULL; 1654 } 1655 spa->spa_l2cache.sav_count = 0; 1656 1657 spa->spa_async_suspended = 0; 1658 1659 spa->spa_indirect_vdevs_loaded = B_FALSE; 1660 1661 if (spa->spa_comment != NULL) { 1662 spa_strfree(spa->spa_comment); 1663 spa->spa_comment = NULL; 1664 } 1665 1666 spa_config_exit(spa, SCL_ALL, spa); 1667 } 1668 1669 /* 1670 * Load (or re-load) the current list of vdevs describing the active spares for 1671 * this pool. When this is called, we have some form of basic information in 1672 * 'spa_spares.sav_config'. We parse this into vdevs, try to open them, and 1673 * then re-generate a more complete list including status information. 1674 */ 1675 void 1676 spa_load_spares(spa_t *spa) 1677 { 1678 nvlist_t **spares; 1679 uint_t nspares; 1680 int i; 1681 vdev_t *vd, *tvd; 1682 1683 #ifndef _KERNEL 1684 /* 1685 * zdb opens both the current state of the pool and the 1686 * checkpointed state (if present), with a different spa_t. 1687 * 1688 * As spare vdevs are shared among open pools, we skip loading 1689 * them when we load the checkpointed state of the pool. 1690 */ 1691 if (!spa_writeable(spa)) 1692 return; 1693 #endif 1694 1695 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 1696 1697 /* 1698 * First, close and free any existing spare vdevs. 1699 */ 1700 for (i = 0; i < spa->spa_spares.sav_count; i++) { 1701 vd = spa->spa_spares.sav_vdevs[i]; 1702 1703 /* Undo the call to spa_activate() below */ 1704 if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid, 1705 B_FALSE)) != NULL && tvd->vdev_isspare) 1706 spa_spare_remove(tvd); 1707 vdev_close(vd); 1708 vdev_free(vd); 1709 } 1710 1711 if (spa->spa_spares.sav_vdevs) 1712 kmem_free(spa->spa_spares.sav_vdevs, 1713 spa->spa_spares.sav_count * sizeof (void *)); 1714 1715 if (spa->spa_spares.sav_config == NULL) 1716 nspares = 0; 1717 else 1718 VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, 1719 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0); 1720 1721 spa->spa_spares.sav_count = (int)nspares; 1722 spa->spa_spares.sav_vdevs = NULL; 1723 1724 if (nspares == 0) 1725 return; 1726 1727 /* 1728 * Construct the array of vdevs, opening them to get status in the 1729 * process. For each spare, there is potentially two different vdev_t 1730 * structures associated with it: one in the list of spares (used only 1731 * for basic validation purposes) and one in the active vdev 1732 * configuration (if it's spared in). During this phase we open and 1733 * validate each vdev on the spare list. If the vdev also exists in the 1734 * active configuration, then we also mark this vdev as an active spare. 1735 */ 1736 spa->spa_spares.sav_vdevs = kmem_zalloc(nspares * sizeof (void *), 1737 KM_SLEEP); 1738 for (i = 0; i < spa->spa_spares.sav_count; i++) { 1739 VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0, 1740 VDEV_ALLOC_SPARE) == 0); 1741 ASSERT(vd != NULL); 1742 1743 spa->spa_spares.sav_vdevs[i] = vd; 1744 1745 if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid, 1746 B_FALSE)) != NULL) { 1747 if (!tvd->vdev_isspare) 1748 spa_spare_add(tvd); 1749 1750 /* 1751 * We only mark the spare active if we were successfully 1752 * able to load the vdev. Otherwise, importing a pool 1753 * with a bad active spare would result in strange 1754 * behavior, because multiple pool would think the spare 1755 * is actively in use. 1756 * 1757 * There is a vulnerability here to an equally bizarre 1758 * circumstance, where a dead active spare is later 1759 * brought back to life (onlined or otherwise). Given 1760 * the rarity of this scenario, and the extra complexity 1761 * it adds, we ignore the possibility. 1762 */ 1763 if (!vdev_is_dead(tvd)) 1764 spa_spare_activate(tvd); 1765 } 1766 1767 vd->vdev_top = vd; 1768 vd->vdev_aux = &spa->spa_spares; 1769 1770 if (vdev_open(vd) != 0) 1771 continue; 1772 1773 if (vdev_validate_aux(vd) == 0) 1774 spa_spare_add(vd); 1775 } 1776 1777 /* 1778 * Recompute the stashed list of spares, with status information 1779 * this time. 1780 */ 1781 VERIFY(nvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES, 1782 DATA_TYPE_NVLIST_ARRAY) == 0); 1783 1784 spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *), 1785 KM_SLEEP); 1786 for (i = 0; i < spa->spa_spares.sav_count; i++) 1787 spares[i] = vdev_config_generate(spa, 1788 spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE); 1789 VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config, 1790 ZPOOL_CONFIG_SPARES, spares, spa->spa_spares.sav_count) == 0); 1791 for (i = 0; i < spa->spa_spares.sav_count; i++) 1792 nvlist_free(spares[i]); 1793 kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *)); 1794 } 1795 1796 /* 1797 * Load (or re-load) the current list of vdevs describing the active l2cache for 1798 * this pool. When this is called, we have some form of basic information in 1799 * 'spa_l2cache.sav_config'. We parse this into vdevs, try to open them, and 1800 * then re-generate a more complete list including status information. 1801 * Devices which are already active have their details maintained, and are 1802 * not re-opened. 1803 */ 1804 void 1805 spa_load_l2cache(spa_t *spa) 1806 { 1807 nvlist_t **l2cache = NULL; 1808 uint_t nl2cache; 1809 int i, j, oldnvdevs; 1810 uint64_t guid; 1811 vdev_t *vd, **oldvdevs, **newvdevs; 1812 spa_aux_vdev_t *sav = &spa->spa_l2cache; 1813 1814 #ifndef _KERNEL 1815 /* 1816 * zdb opens both the current state of the pool and the 1817 * checkpointed state (if present), with a different spa_t. 1818 * 1819 * As L2 caches are part of the ARC which is shared among open 1820 * pools, we skip loading them when we load the checkpointed 1821 * state of the pool. 1822 */ 1823 if (!spa_writeable(spa)) 1824 return; 1825 #endif 1826 1827 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 1828 1829 oldvdevs = sav->sav_vdevs; 1830 oldnvdevs = sav->sav_count; 1831 sav->sav_vdevs = NULL; 1832 sav->sav_count = 0; 1833 1834 if (sav->sav_config == NULL) { 1835 nl2cache = 0; 1836 newvdevs = NULL; 1837 goto out; 1838 } 1839 1840 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, 1841 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0); 1842 newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP); 1843 1844 /* 1845 * Process new nvlist of vdevs. 1846 */ 1847 for (i = 0; i < nl2cache; i++) { 1848 VERIFY(nvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID, 1849 &guid) == 0); 1850 1851 newvdevs[i] = NULL; 1852 for (j = 0; j < oldnvdevs; j++) { 1853 vd = oldvdevs[j]; 1854 if (vd != NULL && guid == vd->vdev_guid) { 1855 /* 1856 * Retain previous vdev for add/remove ops. 1857 */ 1858 newvdevs[i] = vd; 1859 oldvdevs[j] = NULL; 1860 break; 1861 } 1862 } 1863 1864 if (newvdevs[i] == NULL) { 1865 /* 1866 * Create new vdev 1867 */ 1868 VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0, 1869 VDEV_ALLOC_L2CACHE) == 0); 1870 ASSERT(vd != NULL); 1871 newvdevs[i] = vd; 1872 1873 /* 1874 * Commit this vdev as an l2cache device, 1875 * even if it fails to open. 1876 */ 1877 spa_l2cache_add(vd); 1878 1879 vd->vdev_top = vd; 1880 vd->vdev_aux = sav; 1881 1882 spa_l2cache_activate(vd); 1883 1884 if (vdev_open(vd) != 0) 1885 continue; 1886 1887 (void) vdev_validate_aux(vd); 1888 1889 if (!vdev_is_dead(vd)) 1890 l2arc_add_vdev(spa, vd); 1891 1892 /* 1893 * Upon cache device addition to a pool or pool 1894 * creation with a cache device or if the header 1895 * of the device is invalid we issue an async 1896 * TRIM command for the whole device which will 1897 * execute if l2arc_trim_ahead > 0. 1898 */ 1899 spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM); 1900 } 1901 } 1902 1903 sav->sav_vdevs = newvdevs; 1904 sav->sav_count = (int)nl2cache; 1905 1906 /* 1907 * Recompute the stashed list of l2cache devices, with status 1908 * information this time. 1909 */ 1910 VERIFY(nvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE, 1911 DATA_TYPE_NVLIST_ARRAY) == 0); 1912 1913 if (sav->sav_count > 0) 1914 l2cache = kmem_alloc(sav->sav_count * sizeof (void *), 1915 KM_SLEEP); 1916 for (i = 0; i < sav->sav_count; i++) 1917 l2cache[i] = vdev_config_generate(spa, 1918 sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE); 1919 VERIFY(nvlist_add_nvlist_array(sav->sav_config, 1920 ZPOOL_CONFIG_L2CACHE, l2cache, sav->sav_count) == 0); 1921 1922 out: 1923 /* 1924 * Purge vdevs that were dropped 1925 */ 1926 for (i = 0; i < oldnvdevs; i++) { 1927 uint64_t pool; 1928 1929 vd = oldvdevs[i]; 1930 if (vd != NULL) { 1931 ASSERT(vd->vdev_isl2cache); 1932 1933 if (spa_l2cache_exists(vd->vdev_guid, &pool) && 1934 pool != 0ULL && l2arc_vdev_present(vd)) 1935 l2arc_remove_vdev(vd); 1936 vdev_clear_stats(vd); 1937 vdev_free(vd); 1938 } 1939 } 1940 1941 if (oldvdevs) 1942 kmem_free(oldvdevs, oldnvdevs * sizeof (void *)); 1943 1944 for (i = 0; i < sav->sav_count; i++) 1945 nvlist_free(l2cache[i]); 1946 if (sav->sav_count) 1947 kmem_free(l2cache, sav->sav_count * sizeof (void *)); 1948 } 1949 1950 static int 1951 load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value) 1952 { 1953 dmu_buf_t *db; 1954 char *packed = NULL; 1955 size_t nvsize = 0; 1956 int error; 1957 *value = NULL; 1958 1959 error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db); 1960 if (error) 1961 return (error); 1962 1963 nvsize = *(uint64_t *)db->db_data; 1964 dmu_buf_rele(db, FTAG); 1965 1966 packed = vmem_alloc(nvsize, KM_SLEEP); 1967 error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed, 1968 DMU_READ_PREFETCH); 1969 if (error == 0) 1970 error = nvlist_unpack(packed, nvsize, value, 0); 1971 vmem_free(packed, nvsize); 1972 1973 return (error); 1974 } 1975 1976 /* 1977 * Concrete top-level vdevs that are not missing and are not logs. At every 1978 * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds. 1979 */ 1980 static uint64_t 1981 spa_healthy_core_tvds(spa_t *spa) 1982 { 1983 vdev_t *rvd = spa->spa_root_vdev; 1984 uint64_t tvds = 0; 1985 1986 for (uint64_t i = 0; i < rvd->vdev_children; i++) { 1987 vdev_t *vd = rvd->vdev_child[i]; 1988 if (vd->vdev_islog) 1989 continue; 1990 if (vdev_is_concrete(vd) && !vdev_is_dead(vd)) 1991 tvds++; 1992 } 1993 1994 return (tvds); 1995 } 1996 1997 /* 1998 * Checks to see if the given vdev could not be opened, in which case we post a 1999 * sysevent to notify the autoreplace code that the device has been removed. 2000 */ 2001 static void 2002 spa_check_removed(vdev_t *vd) 2003 { 2004 for (uint64_t c = 0; c < vd->vdev_children; c++) 2005 spa_check_removed(vd->vdev_child[c]); 2006 2007 if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) && 2008 vdev_is_concrete(vd)) { 2009 zfs_post_autoreplace(vd->vdev_spa, vd); 2010 spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK); 2011 } 2012 } 2013 2014 static int 2015 spa_check_for_missing_logs(spa_t *spa) 2016 { 2017 vdev_t *rvd = spa->spa_root_vdev; 2018 2019 /* 2020 * If we're doing a normal import, then build up any additional 2021 * diagnostic information about missing log devices. 2022 * We'll pass this up to the user for further processing. 2023 */ 2024 if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) { 2025 nvlist_t **child, *nv; 2026 uint64_t idx = 0; 2027 2028 child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t *), 2029 KM_SLEEP); 2030 VERIFY(nvlist_alloc(&nv, NV_UNIQUE_NAME, KM_SLEEP) == 0); 2031 2032 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 2033 vdev_t *tvd = rvd->vdev_child[c]; 2034 2035 /* 2036 * We consider a device as missing only if it failed 2037 * to open (i.e. offline or faulted is not considered 2038 * as missing). 2039 */ 2040 if (tvd->vdev_islog && 2041 tvd->vdev_state == VDEV_STATE_CANT_OPEN) { 2042 child[idx++] = vdev_config_generate(spa, tvd, 2043 B_FALSE, VDEV_CONFIG_MISSING); 2044 } 2045 } 2046 2047 if (idx > 0) { 2048 fnvlist_add_nvlist_array(nv, 2049 ZPOOL_CONFIG_CHILDREN, child, idx); 2050 fnvlist_add_nvlist(spa->spa_load_info, 2051 ZPOOL_CONFIG_MISSING_DEVICES, nv); 2052 2053 for (uint64_t i = 0; i < idx; i++) 2054 nvlist_free(child[i]); 2055 } 2056 nvlist_free(nv); 2057 kmem_free(child, rvd->vdev_children * sizeof (char **)); 2058 2059 if (idx > 0) { 2060 spa_load_failed(spa, "some log devices are missing"); 2061 vdev_dbgmsg_print_tree(rvd, 2); 2062 return (SET_ERROR(ENXIO)); 2063 } 2064 } else { 2065 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 2066 vdev_t *tvd = rvd->vdev_child[c]; 2067 2068 if (tvd->vdev_islog && 2069 tvd->vdev_state == VDEV_STATE_CANT_OPEN) { 2070 spa_set_log_state(spa, SPA_LOG_CLEAR); 2071 spa_load_note(spa, "some log devices are " 2072 "missing, ZIL is dropped."); 2073 vdev_dbgmsg_print_tree(rvd, 2); 2074 break; 2075 } 2076 } 2077 } 2078 2079 return (0); 2080 } 2081 2082 /* 2083 * Check for missing log devices 2084 */ 2085 static boolean_t 2086 spa_check_logs(spa_t *spa) 2087 { 2088 boolean_t rv = B_FALSE; 2089 dsl_pool_t *dp = spa_get_dsl(spa); 2090 2091 switch (spa->spa_log_state) { 2092 default: 2093 break; 2094 case SPA_LOG_MISSING: 2095 /* need to recheck in case slog has been restored */ 2096 case SPA_LOG_UNKNOWN: 2097 rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 2098 zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0); 2099 if (rv) 2100 spa_set_log_state(spa, SPA_LOG_MISSING); 2101 break; 2102 } 2103 return (rv); 2104 } 2105 2106 static boolean_t 2107 spa_passivate_log(spa_t *spa) 2108 { 2109 vdev_t *rvd = spa->spa_root_vdev; 2110 boolean_t slog_found = B_FALSE; 2111 2112 ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER)); 2113 2114 for (int c = 0; c < rvd->vdev_children; c++) { 2115 vdev_t *tvd = rvd->vdev_child[c]; 2116 metaslab_group_t *mg = tvd->vdev_mg; 2117 2118 if (tvd->vdev_islog) { 2119 metaslab_group_passivate(mg); 2120 slog_found = B_TRUE; 2121 } 2122 } 2123 2124 return (slog_found); 2125 } 2126 2127 static void 2128 spa_activate_log(spa_t *spa) 2129 { 2130 vdev_t *rvd = spa->spa_root_vdev; 2131 2132 ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER)); 2133 2134 for (int c = 0; c < rvd->vdev_children; c++) { 2135 vdev_t *tvd = rvd->vdev_child[c]; 2136 metaslab_group_t *mg = tvd->vdev_mg; 2137 2138 if (tvd->vdev_islog) 2139 metaslab_group_activate(mg); 2140 } 2141 } 2142 2143 int 2144 spa_reset_logs(spa_t *spa) 2145 { 2146 int error; 2147 2148 error = dmu_objset_find(spa_name(spa), zil_reset, 2149 NULL, DS_FIND_CHILDREN); 2150 if (error == 0) { 2151 /* 2152 * We successfully offlined the log device, sync out the 2153 * current txg so that the "stubby" block can be removed 2154 * by zil_sync(). 2155 */ 2156 txg_wait_synced(spa->spa_dsl_pool, 0); 2157 } 2158 return (error); 2159 } 2160 2161 static void 2162 spa_aux_check_removed(spa_aux_vdev_t *sav) 2163 { 2164 for (int i = 0; i < sav->sav_count; i++) 2165 spa_check_removed(sav->sav_vdevs[i]); 2166 } 2167 2168 void 2169 spa_claim_notify(zio_t *zio) 2170 { 2171 spa_t *spa = zio->io_spa; 2172 2173 if (zio->io_error) 2174 return; 2175 2176 mutex_enter(&spa->spa_props_lock); /* any mutex will do */ 2177 if (spa->spa_claim_max_txg < zio->io_bp->blk_birth) 2178 spa->spa_claim_max_txg = zio->io_bp->blk_birth; 2179 mutex_exit(&spa->spa_props_lock); 2180 } 2181 2182 typedef struct spa_load_error { 2183 uint64_t sle_meta_count; 2184 uint64_t sle_data_count; 2185 } spa_load_error_t; 2186 2187 static void 2188 spa_load_verify_done(zio_t *zio) 2189 { 2190 blkptr_t *bp = zio->io_bp; 2191 spa_load_error_t *sle = zio->io_private; 2192 dmu_object_type_t type = BP_GET_TYPE(bp); 2193 int error = zio->io_error; 2194 spa_t *spa = zio->io_spa; 2195 2196 abd_free(zio->io_abd); 2197 if (error) { 2198 if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) && 2199 type != DMU_OT_INTENT_LOG) 2200 atomic_inc_64(&sle->sle_meta_count); 2201 else 2202 atomic_inc_64(&sle->sle_data_count); 2203 } 2204 2205 mutex_enter(&spa->spa_scrub_lock); 2206 spa->spa_load_verify_bytes -= BP_GET_PSIZE(bp); 2207 cv_broadcast(&spa->spa_scrub_io_cv); 2208 mutex_exit(&spa->spa_scrub_lock); 2209 } 2210 2211 /* 2212 * Maximum number of inflight bytes is the log2 fraction of the arc size. 2213 * By default, we set it to 1/16th of the arc. 2214 */ 2215 int spa_load_verify_shift = 4; 2216 int spa_load_verify_metadata = B_TRUE; 2217 int spa_load_verify_data = B_TRUE; 2218 2219 /*ARGSUSED*/ 2220 static int 2221 spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp, 2222 const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg) 2223 { 2224 if (zb->zb_level == ZB_DNODE_LEVEL || BP_IS_HOLE(bp) || 2225 BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp)) 2226 return (0); 2227 /* 2228 * Note: normally this routine will not be called if 2229 * spa_load_verify_metadata is not set. However, it may be useful 2230 * to manually set the flag after the traversal has begun. 2231 */ 2232 if (!spa_load_verify_metadata) 2233 return (0); 2234 if (!BP_IS_METADATA(bp) && !spa_load_verify_data) 2235 return (0); 2236 2237 uint64_t maxinflight_bytes = 2238 arc_target_bytes() >> spa_load_verify_shift; 2239 zio_t *rio = arg; 2240 size_t size = BP_GET_PSIZE(bp); 2241 2242 mutex_enter(&spa->spa_scrub_lock); 2243 while (spa->spa_load_verify_bytes >= maxinflight_bytes) 2244 cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock); 2245 spa->spa_load_verify_bytes += size; 2246 mutex_exit(&spa->spa_scrub_lock); 2247 2248 zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size, 2249 spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB, 2250 ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL | 2251 ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb)); 2252 return (0); 2253 } 2254 2255 /* ARGSUSED */ 2256 static int 2257 verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg) 2258 { 2259 if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN) 2260 return (SET_ERROR(ENAMETOOLONG)); 2261 2262 return (0); 2263 } 2264 2265 static int 2266 spa_load_verify(spa_t *spa) 2267 { 2268 zio_t *rio; 2269 spa_load_error_t sle = { 0 }; 2270 zpool_load_policy_t policy; 2271 boolean_t verify_ok = B_FALSE; 2272 int error = 0; 2273 2274 zpool_get_load_policy(spa->spa_config, &policy); 2275 2276 if (policy.zlp_rewind & ZPOOL_NEVER_REWIND) 2277 return (0); 2278 2279 dsl_pool_config_enter(spa->spa_dsl_pool, FTAG); 2280 error = dmu_objset_find_dp(spa->spa_dsl_pool, 2281 spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL, 2282 DS_FIND_CHILDREN); 2283 dsl_pool_config_exit(spa->spa_dsl_pool, FTAG); 2284 if (error != 0) 2285 return (error); 2286 2287 rio = zio_root(spa, NULL, &sle, 2288 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE); 2289 2290 if (spa_load_verify_metadata) { 2291 if (spa->spa_extreme_rewind) { 2292 spa_load_note(spa, "performing a complete scan of the " 2293 "pool since extreme rewind is on. This may take " 2294 "a very long time.\n (spa_load_verify_data=%u, " 2295 "spa_load_verify_metadata=%u)", 2296 spa_load_verify_data, spa_load_verify_metadata); 2297 } 2298 2299 error = traverse_pool(spa, spa->spa_verify_min_txg, 2300 TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA | 2301 TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio); 2302 } 2303 2304 (void) zio_wait(rio); 2305 ASSERT0(spa->spa_load_verify_bytes); 2306 2307 spa->spa_load_meta_errors = sle.sle_meta_count; 2308 spa->spa_load_data_errors = sle.sle_data_count; 2309 2310 if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) { 2311 spa_load_note(spa, "spa_load_verify found %llu metadata errors " 2312 "and %llu data errors", (u_longlong_t)sle.sle_meta_count, 2313 (u_longlong_t)sle.sle_data_count); 2314 } 2315 2316 if (spa_load_verify_dryrun || 2317 (!error && sle.sle_meta_count <= policy.zlp_maxmeta && 2318 sle.sle_data_count <= policy.zlp_maxdata)) { 2319 int64_t loss = 0; 2320 2321 verify_ok = B_TRUE; 2322 spa->spa_load_txg = spa->spa_uberblock.ub_txg; 2323 spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp; 2324 2325 loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts; 2326 VERIFY(nvlist_add_uint64(spa->spa_load_info, 2327 ZPOOL_CONFIG_LOAD_TIME, spa->spa_load_txg_ts) == 0); 2328 VERIFY(nvlist_add_int64(spa->spa_load_info, 2329 ZPOOL_CONFIG_REWIND_TIME, loss) == 0); 2330 VERIFY(nvlist_add_uint64(spa->spa_load_info, 2331 ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count) == 0); 2332 } else { 2333 spa->spa_load_max_txg = spa->spa_uberblock.ub_txg; 2334 } 2335 2336 if (spa_load_verify_dryrun) 2337 return (0); 2338 2339 if (error) { 2340 if (error != ENXIO && error != EIO) 2341 error = SET_ERROR(EIO); 2342 return (error); 2343 } 2344 2345 return (verify_ok ? 0 : EIO); 2346 } 2347 2348 /* 2349 * Find a value in the pool props object. 2350 */ 2351 static void 2352 spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val) 2353 { 2354 (void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object, 2355 zpool_prop_to_name(prop), sizeof (uint64_t), 1, val); 2356 } 2357 2358 /* 2359 * Find a value in the pool directory object. 2360 */ 2361 static int 2362 spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent) 2363 { 2364 int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 2365 name, sizeof (uint64_t), 1, val); 2366 2367 if (error != 0 && (error != ENOENT || log_enoent)) { 2368 spa_load_failed(spa, "couldn't get '%s' value in MOS directory " 2369 "[error=%d]", name, error); 2370 } 2371 2372 return (error); 2373 } 2374 2375 static int 2376 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err) 2377 { 2378 vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux); 2379 return (SET_ERROR(err)); 2380 } 2381 2382 boolean_t 2383 spa_livelist_delete_check(spa_t *spa) 2384 { 2385 return (spa->spa_livelists_to_delete != 0); 2386 } 2387 2388 /* ARGSUSED */ 2389 static boolean_t 2390 spa_livelist_delete_cb_check(void *arg, zthr_t *z) 2391 { 2392 spa_t *spa = arg; 2393 return (spa_livelist_delete_check(spa)); 2394 } 2395 2396 static int 2397 delete_blkptr_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 2398 { 2399 spa_t *spa = arg; 2400 zio_free(spa, tx->tx_txg, bp); 2401 dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD, 2402 -bp_get_dsize_sync(spa, bp), 2403 -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx); 2404 return (0); 2405 } 2406 2407 static int 2408 dsl_get_next_livelist_obj(objset_t *os, uint64_t zap_obj, uint64_t *llp) 2409 { 2410 int err; 2411 zap_cursor_t zc; 2412 zap_attribute_t za; 2413 zap_cursor_init(&zc, os, zap_obj); 2414 err = zap_cursor_retrieve(&zc, &za); 2415 zap_cursor_fini(&zc); 2416 if (err == 0) 2417 *llp = za.za_first_integer; 2418 return (err); 2419 } 2420 2421 /* 2422 * Components of livelist deletion that must be performed in syncing 2423 * context: freeing block pointers and updating the pool-wide data 2424 * structures to indicate how much work is left to do 2425 */ 2426 typedef struct sublist_delete_arg { 2427 spa_t *spa; 2428 dsl_deadlist_t *ll; 2429 uint64_t key; 2430 bplist_t *to_free; 2431 } sublist_delete_arg_t; 2432 2433 static void 2434 sublist_delete_sync(void *arg, dmu_tx_t *tx) 2435 { 2436 sublist_delete_arg_t *sda = arg; 2437 spa_t *spa = sda->spa; 2438 dsl_deadlist_t *ll = sda->ll; 2439 uint64_t key = sda->key; 2440 bplist_t *to_free = sda->to_free; 2441 2442 bplist_iterate(to_free, delete_blkptr_cb, spa, tx); 2443 dsl_deadlist_remove_entry(ll, key, tx); 2444 } 2445 2446 typedef struct livelist_delete_arg { 2447 spa_t *spa; 2448 uint64_t ll_obj; 2449 uint64_t zap_obj; 2450 } livelist_delete_arg_t; 2451 2452 static void 2453 livelist_delete_sync(void *arg, dmu_tx_t *tx) 2454 { 2455 livelist_delete_arg_t *lda = arg; 2456 spa_t *spa = lda->spa; 2457 uint64_t ll_obj = lda->ll_obj; 2458 uint64_t zap_obj = lda->zap_obj; 2459 objset_t *mos = spa->spa_meta_objset; 2460 uint64_t count; 2461 2462 /* free the livelist and decrement the feature count */ 2463 VERIFY0(zap_remove_int(mos, zap_obj, ll_obj, tx)); 2464 dsl_deadlist_free(mos, ll_obj, tx); 2465 spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx); 2466 VERIFY0(zap_count(mos, zap_obj, &count)); 2467 if (count == 0) { 2468 /* no more livelists to delete */ 2469 VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT, 2470 DMU_POOL_DELETED_CLONES, tx)); 2471 VERIFY0(zap_destroy(mos, zap_obj, tx)); 2472 spa->spa_livelists_to_delete = 0; 2473 spa_notify_waiters(spa); 2474 } 2475 } 2476 2477 /* 2478 * Load in the value for the livelist to be removed and open it. Then, 2479 * load its first sublist and determine which block pointers should actually 2480 * be freed. Then, call a synctask which performs the actual frees and updates 2481 * the pool-wide livelist data. 2482 */ 2483 /* ARGSUSED */ 2484 static void 2485 spa_livelist_delete_cb(void *arg, zthr_t *z) 2486 { 2487 spa_t *spa = arg; 2488 uint64_t ll_obj = 0, count; 2489 objset_t *mos = spa->spa_meta_objset; 2490 uint64_t zap_obj = spa->spa_livelists_to_delete; 2491 /* 2492 * Determine the next livelist to delete. This function should only 2493 * be called if there is at least one deleted clone. 2494 */ 2495 VERIFY0(dsl_get_next_livelist_obj(mos, zap_obj, &ll_obj)); 2496 VERIFY0(zap_count(mos, ll_obj, &count)); 2497 if (count > 0) { 2498 dsl_deadlist_t *ll; 2499 dsl_deadlist_entry_t *dle; 2500 bplist_t to_free; 2501 ll = kmem_zalloc(sizeof (dsl_deadlist_t), KM_SLEEP); 2502 dsl_deadlist_open(ll, mos, ll_obj); 2503 dle = dsl_deadlist_first(ll); 2504 ASSERT3P(dle, !=, NULL); 2505 bplist_create(&to_free); 2506 int err = dsl_process_sub_livelist(&dle->dle_bpobj, &to_free, 2507 z, NULL); 2508 if (err == 0) { 2509 sublist_delete_arg_t sync_arg = { 2510 .spa = spa, 2511 .ll = ll, 2512 .key = dle->dle_mintxg, 2513 .to_free = &to_free 2514 }; 2515 zfs_dbgmsg("deleting sublist (id %llu) from" 2516 " livelist %llu, %d remaining", 2517 dle->dle_bpobj.bpo_object, ll_obj, count - 1); 2518 VERIFY0(dsl_sync_task(spa_name(spa), NULL, 2519 sublist_delete_sync, &sync_arg, 0, 2520 ZFS_SPACE_CHECK_DESTROY)); 2521 } else { 2522 VERIFY3U(err, ==, EINTR); 2523 } 2524 bplist_clear(&to_free); 2525 bplist_destroy(&to_free); 2526 dsl_deadlist_close(ll); 2527 kmem_free(ll, sizeof (dsl_deadlist_t)); 2528 } else { 2529 livelist_delete_arg_t sync_arg = { 2530 .spa = spa, 2531 .ll_obj = ll_obj, 2532 .zap_obj = zap_obj 2533 }; 2534 zfs_dbgmsg("deletion of livelist %llu completed", ll_obj); 2535 VERIFY0(dsl_sync_task(spa_name(spa), NULL, livelist_delete_sync, 2536 &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY)); 2537 } 2538 } 2539 2540 static void 2541 spa_start_livelist_destroy_thread(spa_t *spa) 2542 { 2543 ASSERT3P(spa->spa_livelist_delete_zthr, ==, NULL); 2544 spa->spa_livelist_delete_zthr = 2545 zthr_create("z_livelist_destroy", 2546 spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa); 2547 } 2548 2549 typedef struct livelist_new_arg { 2550 bplist_t *allocs; 2551 bplist_t *frees; 2552 } livelist_new_arg_t; 2553 2554 static int 2555 livelist_track_new_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 2556 dmu_tx_t *tx) 2557 { 2558 ASSERT(tx == NULL); 2559 livelist_new_arg_t *lna = arg; 2560 if (bp_freed) { 2561 bplist_append(lna->frees, bp); 2562 } else { 2563 bplist_append(lna->allocs, bp); 2564 zfs_livelist_condense_new_alloc++; 2565 } 2566 return (0); 2567 } 2568 2569 typedef struct livelist_condense_arg { 2570 spa_t *spa; 2571 bplist_t to_keep; 2572 uint64_t first_size; 2573 uint64_t next_size; 2574 } livelist_condense_arg_t; 2575 2576 static void 2577 spa_livelist_condense_sync(void *arg, dmu_tx_t *tx) 2578 { 2579 livelist_condense_arg_t *lca = arg; 2580 spa_t *spa = lca->spa; 2581 bplist_t new_frees; 2582 dsl_dataset_t *ds = spa->spa_to_condense.ds; 2583 2584 /* Have we been cancelled? */ 2585 if (spa->spa_to_condense.cancelled) { 2586 zfs_livelist_condense_sync_cancel++; 2587 goto out; 2588 } 2589 2590 dsl_deadlist_entry_t *first = spa->spa_to_condense.first; 2591 dsl_deadlist_entry_t *next = spa->spa_to_condense.next; 2592 dsl_deadlist_t *ll = &ds->ds_dir->dd_livelist; 2593 2594 /* 2595 * It's possible that the livelist was changed while the zthr was 2596 * running. Therefore, we need to check for new blkptrs in the two 2597 * entries being condensed and continue to track them in the livelist. 2598 * Because of the way we handle remapped blkptrs (see dbuf_remap_impl), 2599 * it's possible that the newly added blkptrs are FREEs or ALLOCs so 2600 * we need to sort them into two different bplists. 2601 */ 2602 uint64_t first_obj = first->dle_bpobj.bpo_object; 2603 uint64_t next_obj = next->dle_bpobj.bpo_object; 2604 uint64_t cur_first_size = first->dle_bpobj.bpo_phys->bpo_num_blkptrs; 2605 uint64_t cur_next_size = next->dle_bpobj.bpo_phys->bpo_num_blkptrs; 2606 2607 bplist_create(&new_frees); 2608 livelist_new_arg_t new_bps = { 2609 .allocs = &lca->to_keep, 2610 .frees = &new_frees, 2611 }; 2612 2613 if (cur_first_size > lca->first_size) { 2614 VERIFY0(livelist_bpobj_iterate_from_nofree(&first->dle_bpobj, 2615 livelist_track_new_cb, &new_bps, lca->first_size)); 2616 } 2617 if (cur_next_size > lca->next_size) { 2618 VERIFY0(livelist_bpobj_iterate_from_nofree(&next->dle_bpobj, 2619 livelist_track_new_cb, &new_bps, lca->next_size)); 2620 } 2621 2622 dsl_deadlist_clear_entry(first, ll, tx); 2623 ASSERT(bpobj_is_empty(&first->dle_bpobj)); 2624 dsl_deadlist_remove_entry(ll, next->dle_mintxg, tx); 2625 2626 bplist_iterate(&lca->to_keep, dsl_deadlist_insert_alloc_cb, ll, tx); 2627 bplist_iterate(&new_frees, dsl_deadlist_insert_free_cb, ll, tx); 2628 bplist_destroy(&new_frees); 2629 2630 char dsname[ZFS_MAX_DATASET_NAME_LEN]; 2631 dsl_dataset_name(ds, dsname); 2632 zfs_dbgmsg("txg %llu condensing livelist of %s (id %llu), bpobj %llu " 2633 "(%llu blkptrs) and bpobj %llu (%llu blkptrs) -> bpobj %llu " 2634 "(%llu blkptrs)", tx->tx_txg, dsname, ds->ds_object, first_obj, 2635 cur_first_size, next_obj, cur_next_size, 2636 first->dle_bpobj.bpo_object, 2637 first->dle_bpobj.bpo_phys->bpo_num_blkptrs); 2638 out: 2639 dmu_buf_rele(ds->ds_dbuf, spa); 2640 spa->spa_to_condense.ds = NULL; 2641 bplist_clear(&lca->to_keep); 2642 bplist_destroy(&lca->to_keep); 2643 kmem_free(lca, sizeof (livelist_condense_arg_t)); 2644 spa->spa_to_condense.syncing = B_FALSE; 2645 } 2646 2647 static void 2648 spa_livelist_condense_cb(void *arg, zthr_t *t) 2649 { 2650 while (zfs_livelist_condense_zthr_pause && 2651 !(zthr_has_waiters(t) || zthr_iscancelled(t))) 2652 delay(1); 2653 2654 spa_t *spa = arg; 2655 dsl_deadlist_entry_t *first = spa->spa_to_condense.first; 2656 dsl_deadlist_entry_t *next = spa->spa_to_condense.next; 2657 uint64_t first_size, next_size; 2658 2659 livelist_condense_arg_t *lca = 2660 kmem_alloc(sizeof (livelist_condense_arg_t), KM_SLEEP); 2661 bplist_create(&lca->to_keep); 2662 2663 /* 2664 * Process the livelists (matching FREEs and ALLOCs) in open context 2665 * so we have minimal work in syncing context to condense. 2666 * 2667 * We save bpobj sizes (first_size and next_size) to use later in 2668 * syncing context to determine if entries were added to these sublists 2669 * while in open context. This is possible because the clone is still 2670 * active and open for normal writes and we want to make sure the new, 2671 * unprocessed blockpointers are inserted into the livelist normally. 2672 * 2673 * Note that dsl_process_sub_livelist() both stores the size number of 2674 * blockpointers and iterates over them while the bpobj's lock held, so 2675 * the sizes returned to us are consistent which what was actually 2676 * processed. 2677 */ 2678 int err = dsl_process_sub_livelist(&first->dle_bpobj, &lca->to_keep, t, 2679 &first_size); 2680 if (err == 0) 2681 err = dsl_process_sub_livelist(&next->dle_bpobj, &lca->to_keep, 2682 t, &next_size); 2683 2684 if (err == 0) { 2685 while (zfs_livelist_condense_sync_pause && 2686 !(zthr_has_waiters(t) || zthr_iscancelled(t))) 2687 delay(1); 2688 2689 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 2690 dmu_tx_mark_netfree(tx); 2691 dmu_tx_hold_space(tx, 1); 2692 err = dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE); 2693 if (err == 0) { 2694 /* 2695 * Prevent the condense zthr restarting before 2696 * the synctask completes. 2697 */ 2698 spa->spa_to_condense.syncing = B_TRUE; 2699 lca->spa = spa; 2700 lca->first_size = first_size; 2701 lca->next_size = next_size; 2702 dsl_sync_task_nowait(spa_get_dsl(spa), 2703 spa_livelist_condense_sync, lca, tx); 2704 dmu_tx_commit(tx); 2705 return; 2706 } 2707 } 2708 /* 2709 * Condensing can not continue: either it was externally stopped or 2710 * we were unable to assign to a tx because the pool has run out of 2711 * space. In the second case, we'll just end up trying to condense 2712 * again in a later txg. 2713 */ 2714 ASSERT(err != 0); 2715 bplist_clear(&lca->to_keep); 2716 bplist_destroy(&lca->to_keep); 2717 kmem_free(lca, sizeof (livelist_condense_arg_t)); 2718 dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf, spa); 2719 spa->spa_to_condense.ds = NULL; 2720 if (err == EINTR) 2721 zfs_livelist_condense_zthr_cancel++; 2722 } 2723 2724 /* ARGSUSED */ 2725 /* 2726 * Check that there is something to condense but that a condense is not 2727 * already in progress and that condensing has not been cancelled. 2728 */ 2729 static boolean_t 2730 spa_livelist_condense_cb_check(void *arg, zthr_t *z) 2731 { 2732 spa_t *spa = arg; 2733 if ((spa->spa_to_condense.ds != NULL) && 2734 (spa->spa_to_condense.syncing == B_FALSE) && 2735 (spa->spa_to_condense.cancelled == B_FALSE)) { 2736 return (B_TRUE); 2737 } 2738 return (B_FALSE); 2739 } 2740 2741 static void 2742 spa_start_livelist_condensing_thread(spa_t *spa) 2743 { 2744 spa->spa_to_condense.ds = NULL; 2745 spa->spa_to_condense.first = NULL; 2746 spa->spa_to_condense.next = NULL; 2747 spa->spa_to_condense.syncing = B_FALSE; 2748 spa->spa_to_condense.cancelled = B_FALSE; 2749 2750 ASSERT3P(spa->spa_livelist_condense_zthr, ==, NULL); 2751 spa->spa_livelist_condense_zthr = 2752 zthr_create("z_livelist_condense", 2753 spa_livelist_condense_cb_check, 2754 spa_livelist_condense_cb, spa); 2755 } 2756 2757 static void 2758 spa_spawn_aux_threads(spa_t *spa) 2759 { 2760 ASSERT(spa_writeable(spa)); 2761 2762 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 2763 2764 spa_start_indirect_condensing_thread(spa); 2765 spa_start_livelist_destroy_thread(spa); 2766 spa_start_livelist_condensing_thread(spa); 2767 2768 ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL); 2769 spa->spa_checkpoint_discard_zthr = 2770 zthr_create("z_checkpoint_discard", 2771 spa_checkpoint_discard_thread_check, 2772 spa_checkpoint_discard_thread, spa); 2773 } 2774 2775 /* 2776 * Fix up config after a partly-completed split. This is done with the 2777 * ZPOOL_CONFIG_SPLIT nvlist. Both the splitting pool and the split-off 2778 * pool have that entry in their config, but only the splitting one contains 2779 * a list of all the guids of the vdevs that are being split off. 2780 * 2781 * This function determines what to do with that list: either rejoin 2782 * all the disks to the pool, or complete the splitting process. To attempt 2783 * the rejoin, each disk that is offlined is marked online again, and 2784 * we do a reopen() call. If the vdev label for every disk that was 2785 * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL) 2786 * then we call vdev_split() on each disk, and complete the split. 2787 * 2788 * Otherwise we leave the config alone, with all the vdevs in place in 2789 * the original pool. 2790 */ 2791 static void 2792 spa_try_repair(spa_t *spa, nvlist_t *config) 2793 { 2794 uint_t extracted; 2795 uint64_t *glist; 2796 uint_t i, gcount; 2797 nvlist_t *nvl; 2798 vdev_t **vd; 2799 boolean_t attempt_reopen; 2800 2801 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0) 2802 return; 2803 2804 /* check that the config is complete */ 2805 if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, 2806 &glist, &gcount) != 0) 2807 return; 2808 2809 vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP); 2810 2811 /* attempt to online all the vdevs & validate */ 2812 attempt_reopen = B_TRUE; 2813 for (i = 0; i < gcount; i++) { 2814 if (glist[i] == 0) /* vdev is hole */ 2815 continue; 2816 2817 vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE); 2818 if (vd[i] == NULL) { 2819 /* 2820 * Don't bother attempting to reopen the disks; 2821 * just do the split. 2822 */ 2823 attempt_reopen = B_FALSE; 2824 } else { 2825 /* attempt to re-online it */ 2826 vd[i]->vdev_offline = B_FALSE; 2827 } 2828 } 2829 2830 if (attempt_reopen) { 2831 vdev_reopen(spa->spa_root_vdev); 2832 2833 /* check each device to see what state it's in */ 2834 for (extracted = 0, i = 0; i < gcount; i++) { 2835 if (vd[i] != NULL && 2836 vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL) 2837 break; 2838 ++extracted; 2839 } 2840 } 2841 2842 /* 2843 * If every disk has been moved to the new pool, or if we never 2844 * even attempted to look at them, then we split them off for 2845 * good. 2846 */ 2847 if (!attempt_reopen || gcount == extracted) { 2848 for (i = 0; i < gcount; i++) 2849 if (vd[i] != NULL) 2850 vdev_split(vd[i]); 2851 vdev_reopen(spa->spa_root_vdev); 2852 } 2853 2854 kmem_free(vd, gcount * sizeof (vdev_t *)); 2855 } 2856 2857 static int 2858 spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type) 2859 { 2860 char *ereport = FM_EREPORT_ZFS_POOL; 2861 int error; 2862 2863 spa->spa_load_state = state; 2864 (void) spa_import_progress_set_state(spa_guid(spa), 2865 spa_load_state(spa)); 2866 2867 gethrestime(&spa->spa_loaded_ts); 2868 error = spa_load_impl(spa, type, &ereport); 2869 2870 /* 2871 * Don't count references from objsets that are already closed 2872 * and are making their way through the eviction process. 2873 */ 2874 spa_evicting_os_wait(spa); 2875 spa->spa_minref = zfs_refcount_count(&spa->spa_refcount); 2876 if (error) { 2877 if (error != EEXIST) { 2878 spa->spa_loaded_ts.tv_sec = 0; 2879 spa->spa_loaded_ts.tv_nsec = 0; 2880 } 2881 if (error != EBADF) { 2882 (void) zfs_ereport_post(ereport, spa, 2883 NULL, NULL, NULL, 0); 2884 } 2885 } 2886 spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE; 2887 spa->spa_ena = 0; 2888 2889 (void) spa_import_progress_set_state(spa_guid(spa), 2890 spa_load_state(spa)); 2891 2892 return (error); 2893 } 2894 2895 #ifdef ZFS_DEBUG 2896 /* 2897 * Count the number of per-vdev ZAPs associated with all of the vdevs in the 2898 * vdev tree rooted in the given vd, and ensure that each ZAP is present in the 2899 * spa's per-vdev ZAP list. 2900 */ 2901 static uint64_t 2902 vdev_count_verify_zaps(vdev_t *vd) 2903 { 2904 spa_t *spa = vd->vdev_spa; 2905 uint64_t total = 0; 2906 2907 if (vd->vdev_top_zap != 0) { 2908 total++; 2909 ASSERT0(zap_lookup_int(spa->spa_meta_objset, 2910 spa->spa_all_vdev_zaps, vd->vdev_top_zap)); 2911 } 2912 if (vd->vdev_leaf_zap != 0) { 2913 total++; 2914 ASSERT0(zap_lookup_int(spa->spa_meta_objset, 2915 spa->spa_all_vdev_zaps, vd->vdev_leaf_zap)); 2916 } 2917 2918 for (uint64_t i = 0; i < vd->vdev_children; i++) { 2919 total += vdev_count_verify_zaps(vd->vdev_child[i]); 2920 } 2921 2922 return (total); 2923 } 2924 #endif 2925 2926 /* 2927 * Determine whether the activity check is required. 2928 */ 2929 static boolean_t 2930 spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label, 2931 nvlist_t *config) 2932 { 2933 uint64_t state = 0; 2934 uint64_t hostid = 0; 2935 uint64_t tryconfig_txg = 0; 2936 uint64_t tryconfig_timestamp = 0; 2937 uint16_t tryconfig_mmp_seq = 0; 2938 nvlist_t *nvinfo; 2939 2940 if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) { 2941 nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO); 2942 (void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG, 2943 &tryconfig_txg); 2944 (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP, 2945 &tryconfig_timestamp); 2946 (void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ, 2947 &tryconfig_mmp_seq); 2948 } 2949 2950 (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state); 2951 2952 /* 2953 * Disable the MMP activity check - This is used by zdb which 2954 * is intended to be used on potentially active pools. 2955 */ 2956 if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) 2957 return (B_FALSE); 2958 2959 /* 2960 * Skip the activity check when the MMP feature is disabled. 2961 */ 2962 if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0) 2963 return (B_FALSE); 2964 2965 /* 2966 * If the tryconfig_ values are nonzero, they are the results of an 2967 * earlier tryimport. If they all match the uberblock we just found, 2968 * then the pool has not changed and we return false so we do not test 2969 * a second time. 2970 */ 2971 if (tryconfig_txg && tryconfig_txg == ub->ub_txg && 2972 tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp && 2973 tryconfig_mmp_seq && tryconfig_mmp_seq == 2974 (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) 2975 return (B_FALSE); 2976 2977 /* 2978 * Allow the activity check to be skipped when importing the pool 2979 * on the same host which last imported it. Since the hostid from 2980 * configuration may be stale use the one read from the label. 2981 */ 2982 if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID)) 2983 hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID); 2984 2985 if (hostid == spa_get_hostid(spa)) 2986 return (B_FALSE); 2987 2988 /* 2989 * Skip the activity test when the pool was cleanly exported. 2990 */ 2991 if (state != POOL_STATE_ACTIVE) 2992 return (B_FALSE); 2993 2994 return (B_TRUE); 2995 } 2996 2997 /* 2998 * Nanoseconds the activity check must watch for changes on-disk. 2999 */ 3000 static uint64_t 3001 spa_activity_check_duration(spa_t *spa, uberblock_t *ub) 3002 { 3003 uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1); 3004 uint64_t multihost_interval = MSEC2NSEC( 3005 MMP_INTERVAL_OK(zfs_multihost_interval)); 3006 uint64_t import_delay = MAX(NANOSEC, import_intervals * 3007 multihost_interval); 3008 3009 /* 3010 * Local tunables determine a minimum duration except for the case 3011 * where we know when the remote host will suspend the pool if MMP 3012 * writes do not land. 3013 * 3014 * See Big Theory comment at the top of mmp.c for the reasoning behind 3015 * these cases and times. 3016 */ 3017 3018 ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100); 3019 3020 if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) && 3021 MMP_FAIL_INT(ub) > 0) { 3022 3023 /* MMP on remote host will suspend pool after failed writes */ 3024 import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) * 3025 MMP_IMPORT_SAFETY_FACTOR / 100; 3026 3027 zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp " 3028 "mmp_fails=%llu ub_mmp mmp_interval=%llu " 3029 "import_intervals=%u", import_delay, MMP_FAIL_INT(ub), 3030 MMP_INTERVAL(ub), import_intervals); 3031 3032 } else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) && 3033 MMP_FAIL_INT(ub) == 0) { 3034 3035 /* MMP on remote host will never suspend pool */ 3036 import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) + 3037 ub->ub_mmp_delay) * import_intervals); 3038 3039 zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp " 3040 "mmp_interval=%llu ub_mmp_delay=%llu " 3041 "import_intervals=%u", import_delay, MMP_INTERVAL(ub), 3042 ub->ub_mmp_delay, import_intervals); 3043 3044 } else if (MMP_VALID(ub)) { 3045 /* 3046 * zfs-0.7 compatibility case 3047 */ 3048 3049 import_delay = MAX(import_delay, (multihost_interval + 3050 ub->ub_mmp_delay) * import_intervals); 3051 3052 zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu " 3053 "import_intervals=%u leaves=%u", import_delay, 3054 ub->ub_mmp_delay, import_intervals, 3055 vdev_count_leaves(spa)); 3056 } else { 3057 /* Using local tunings is the only reasonable option */ 3058 zfs_dbgmsg("pool last imported on non-MMP aware " 3059 "host using import_delay=%llu multihost_interval=%llu " 3060 "import_intervals=%u", import_delay, multihost_interval, 3061 import_intervals); 3062 } 3063 3064 return (import_delay); 3065 } 3066 3067 /* 3068 * Perform the import activity check. If the user canceled the import or 3069 * we detected activity then fail. 3070 */ 3071 static int 3072 spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config) 3073 { 3074 uint64_t txg = ub->ub_txg; 3075 uint64_t timestamp = ub->ub_timestamp; 3076 uint64_t mmp_config = ub->ub_mmp_config; 3077 uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0; 3078 uint64_t import_delay; 3079 hrtime_t import_expire; 3080 nvlist_t *mmp_label = NULL; 3081 vdev_t *rvd = spa->spa_root_vdev; 3082 kcondvar_t cv; 3083 kmutex_t mtx; 3084 int error = 0; 3085 3086 cv_init(&cv, NULL, CV_DEFAULT, NULL); 3087 mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL); 3088 mutex_enter(&mtx); 3089 3090 /* 3091 * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed 3092 * during the earlier tryimport. If the txg recorded there is 0 then 3093 * the pool is known to be active on another host. 3094 * 3095 * Otherwise, the pool might be in use on another host. Check for 3096 * changes in the uberblocks on disk if necessary. 3097 */ 3098 if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) { 3099 nvlist_t *nvinfo = fnvlist_lookup_nvlist(config, 3100 ZPOOL_CONFIG_LOAD_INFO); 3101 3102 if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) && 3103 fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) { 3104 vdev_uberblock_load(rvd, ub, &mmp_label); 3105 error = SET_ERROR(EREMOTEIO); 3106 goto out; 3107 } 3108 } 3109 3110 import_delay = spa_activity_check_duration(spa, ub); 3111 3112 /* Add a small random factor in case of simultaneous imports (0-25%) */ 3113 import_delay += import_delay * spa_get_random(250) / 1000; 3114 3115 import_expire = gethrtime() + import_delay; 3116 3117 while (gethrtime() < import_expire) { 3118 (void) spa_import_progress_set_mmp_check(spa_guid(spa), 3119 NSEC2SEC(import_expire - gethrtime())); 3120 3121 vdev_uberblock_load(rvd, ub, &mmp_label); 3122 3123 if (txg != ub->ub_txg || timestamp != ub->ub_timestamp || 3124 mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) { 3125 zfs_dbgmsg("multihost activity detected " 3126 "txg %llu ub_txg %llu " 3127 "timestamp %llu ub_timestamp %llu " 3128 "mmp_config %#llx ub_mmp_config %#llx", 3129 txg, ub->ub_txg, timestamp, ub->ub_timestamp, 3130 mmp_config, ub->ub_mmp_config); 3131 3132 error = SET_ERROR(EREMOTEIO); 3133 break; 3134 } 3135 3136 if (mmp_label) { 3137 nvlist_free(mmp_label); 3138 mmp_label = NULL; 3139 } 3140 3141 error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz); 3142 if (error != -1) { 3143 error = SET_ERROR(EINTR); 3144 break; 3145 } 3146 error = 0; 3147 } 3148 3149 out: 3150 mutex_exit(&mtx); 3151 mutex_destroy(&mtx); 3152 cv_destroy(&cv); 3153 3154 /* 3155 * If the pool is determined to be active store the status in the 3156 * spa->spa_load_info nvlist. If the remote hostname or hostid are 3157 * available from configuration read from disk store them as well. 3158 * This allows 'zpool import' to generate a more useful message. 3159 * 3160 * ZPOOL_CONFIG_MMP_STATE - observed pool status (mandatory) 3161 * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool 3162 * ZPOOL_CONFIG_MMP_HOSTID - hostid from the active pool 3163 */ 3164 if (error == EREMOTEIO) { 3165 char *hostname = "<unknown>"; 3166 uint64_t hostid = 0; 3167 3168 if (mmp_label) { 3169 if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) { 3170 hostname = fnvlist_lookup_string(mmp_label, 3171 ZPOOL_CONFIG_HOSTNAME); 3172 fnvlist_add_string(spa->spa_load_info, 3173 ZPOOL_CONFIG_MMP_HOSTNAME, hostname); 3174 } 3175 3176 if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) { 3177 hostid = fnvlist_lookup_uint64(mmp_label, 3178 ZPOOL_CONFIG_HOSTID); 3179 fnvlist_add_uint64(spa->spa_load_info, 3180 ZPOOL_CONFIG_MMP_HOSTID, hostid); 3181 } 3182 } 3183 3184 fnvlist_add_uint64(spa->spa_load_info, 3185 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE); 3186 fnvlist_add_uint64(spa->spa_load_info, 3187 ZPOOL_CONFIG_MMP_TXG, 0); 3188 3189 error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO); 3190 } 3191 3192 if (mmp_label) 3193 nvlist_free(mmp_label); 3194 3195 return (error); 3196 } 3197 3198 static int 3199 spa_verify_host(spa_t *spa, nvlist_t *mos_config) 3200 { 3201 uint64_t hostid; 3202 char *hostname; 3203 uint64_t myhostid = 0; 3204 3205 if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config, 3206 ZPOOL_CONFIG_HOSTID, &hostid) == 0) { 3207 hostname = fnvlist_lookup_string(mos_config, 3208 ZPOOL_CONFIG_HOSTNAME); 3209 3210 myhostid = zone_get_hostid(NULL); 3211 3212 if (hostid != 0 && myhostid != 0 && hostid != myhostid) { 3213 cmn_err(CE_WARN, "pool '%s' could not be " 3214 "loaded as it was last accessed by " 3215 "another system (host: %s hostid: 0x%llx). " 3216 "See: https://openzfs.github.io/openzfs-docs/msg/" 3217 "ZFS-8000-EY", 3218 spa_name(spa), hostname, (u_longlong_t)hostid); 3219 spa_load_failed(spa, "hostid verification failed: pool " 3220 "last accessed by host: %s (hostid: 0x%llx)", 3221 hostname, (u_longlong_t)hostid); 3222 return (SET_ERROR(EBADF)); 3223 } 3224 } 3225 3226 return (0); 3227 } 3228 3229 static int 3230 spa_ld_parse_config(spa_t *spa, spa_import_type_t type) 3231 { 3232 int error = 0; 3233 nvlist_t *nvtree, *nvl, *config = spa->spa_config; 3234 int parse; 3235 vdev_t *rvd; 3236 uint64_t pool_guid; 3237 char *comment; 3238 3239 /* 3240 * Versioning wasn't explicitly added to the label until later, so if 3241 * it's not present treat it as the initial version. 3242 */ 3243 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION, 3244 &spa->spa_ubsync.ub_version) != 0) 3245 spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL; 3246 3247 if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) { 3248 spa_load_failed(spa, "invalid config provided: '%s' missing", 3249 ZPOOL_CONFIG_POOL_GUID); 3250 return (SET_ERROR(EINVAL)); 3251 } 3252 3253 /* 3254 * If we are doing an import, ensure that the pool is not already 3255 * imported by checking if its pool guid already exists in the 3256 * spa namespace. 3257 * 3258 * The only case that we allow an already imported pool to be 3259 * imported again, is when the pool is checkpointed and we want to 3260 * look at its checkpointed state from userland tools like zdb. 3261 */ 3262 #ifdef _KERNEL 3263 if ((spa->spa_load_state == SPA_LOAD_IMPORT || 3264 spa->spa_load_state == SPA_LOAD_TRYIMPORT) && 3265 spa_guid_exists(pool_guid, 0)) { 3266 #else 3267 if ((spa->spa_load_state == SPA_LOAD_IMPORT || 3268 spa->spa_load_state == SPA_LOAD_TRYIMPORT) && 3269 spa_guid_exists(pool_guid, 0) && 3270 !spa_importing_readonly_checkpoint(spa)) { 3271 #endif 3272 spa_load_failed(spa, "a pool with guid %llu is already open", 3273 (u_longlong_t)pool_guid); 3274 return (SET_ERROR(EEXIST)); 3275 } 3276 3277 spa->spa_config_guid = pool_guid; 3278 3279 nvlist_free(spa->spa_load_info); 3280 spa->spa_load_info = fnvlist_alloc(); 3281 3282 ASSERT(spa->spa_comment == NULL); 3283 if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0) 3284 spa->spa_comment = spa_strdup(comment); 3285 3286 (void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, 3287 &spa->spa_config_txg); 3288 3289 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0) 3290 spa->spa_config_splitting = fnvlist_dup(nvl); 3291 3292 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) { 3293 spa_load_failed(spa, "invalid config provided: '%s' missing", 3294 ZPOOL_CONFIG_VDEV_TREE); 3295 return (SET_ERROR(EINVAL)); 3296 } 3297 3298 /* 3299 * Create "The Godfather" zio to hold all async IOs 3300 */ 3301 spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *), 3302 KM_SLEEP); 3303 for (int i = 0; i < max_ncpus; i++) { 3304 spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL, 3305 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 3306 ZIO_FLAG_GODFATHER); 3307 } 3308 3309 /* 3310 * Parse the configuration into a vdev tree. We explicitly set the 3311 * value that will be returned by spa_version() since parsing the 3312 * configuration requires knowing the version number. 3313 */ 3314 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3315 parse = (type == SPA_IMPORT_EXISTING ? 3316 VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT); 3317 error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse); 3318 spa_config_exit(spa, SCL_ALL, FTAG); 3319 3320 if (error != 0) { 3321 spa_load_failed(spa, "unable to parse config [error=%d]", 3322 error); 3323 return (error); 3324 } 3325 3326 ASSERT(spa->spa_root_vdev == rvd); 3327 ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); 3328 ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT); 3329 3330 if (type != SPA_IMPORT_ASSEMBLE) { 3331 ASSERT(spa_guid(spa) == pool_guid); 3332 } 3333 3334 return (0); 3335 } 3336 3337 /* 3338 * Recursively open all vdevs in the vdev tree. This function is called twice: 3339 * first with the untrusted config, then with the trusted config. 3340 */ 3341 static int 3342 spa_ld_open_vdevs(spa_t *spa) 3343 { 3344 int error = 0; 3345 3346 /* 3347 * spa_missing_tvds_allowed defines how many top-level vdevs can be 3348 * missing/unopenable for the root vdev to be still considered openable. 3349 */ 3350 if (spa->spa_trust_config) { 3351 spa->spa_missing_tvds_allowed = zfs_max_missing_tvds; 3352 } else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) { 3353 spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile; 3354 } else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) { 3355 spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan; 3356 } else { 3357 spa->spa_missing_tvds_allowed = 0; 3358 } 3359 3360 spa->spa_missing_tvds_allowed = 3361 MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed); 3362 3363 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3364 error = vdev_open(spa->spa_root_vdev); 3365 spa_config_exit(spa, SCL_ALL, FTAG); 3366 3367 if (spa->spa_missing_tvds != 0) { 3368 spa_load_note(spa, "vdev tree has %lld missing top-level " 3369 "vdevs.", (u_longlong_t)spa->spa_missing_tvds); 3370 if (spa->spa_trust_config && (spa->spa_mode & SPA_MODE_WRITE)) { 3371 /* 3372 * Although theoretically we could allow users to open 3373 * incomplete pools in RW mode, we'd need to add a lot 3374 * of extra logic (e.g. adjust pool space to account 3375 * for missing vdevs). 3376 * This limitation also prevents users from accidentally 3377 * opening the pool in RW mode during data recovery and 3378 * damaging it further. 3379 */ 3380 spa_load_note(spa, "pools with missing top-level " 3381 "vdevs can only be opened in read-only mode."); 3382 error = SET_ERROR(ENXIO); 3383 } else { 3384 spa_load_note(spa, "current settings allow for maximum " 3385 "%lld missing top-level vdevs at this stage.", 3386 (u_longlong_t)spa->spa_missing_tvds_allowed); 3387 } 3388 } 3389 if (error != 0) { 3390 spa_load_failed(spa, "unable to open vdev tree [error=%d]", 3391 error); 3392 } 3393 if (spa->spa_missing_tvds != 0 || error != 0) 3394 vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2); 3395 3396 return (error); 3397 } 3398 3399 /* 3400 * We need to validate the vdev labels against the configuration that 3401 * we have in hand. This function is called twice: first with an untrusted 3402 * config, then with a trusted config. The validation is more strict when the 3403 * config is trusted. 3404 */ 3405 static int 3406 spa_ld_validate_vdevs(spa_t *spa) 3407 { 3408 int error = 0; 3409 vdev_t *rvd = spa->spa_root_vdev; 3410 3411 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3412 error = vdev_validate(rvd); 3413 spa_config_exit(spa, SCL_ALL, FTAG); 3414 3415 if (error != 0) { 3416 spa_load_failed(spa, "vdev_validate failed [error=%d]", error); 3417 return (error); 3418 } 3419 3420 if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) { 3421 spa_load_failed(spa, "cannot open vdev tree after invalidating " 3422 "some vdevs"); 3423 vdev_dbgmsg_print_tree(rvd, 2); 3424 return (SET_ERROR(ENXIO)); 3425 } 3426 3427 return (0); 3428 } 3429 3430 static void 3431 spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub) 3432 { 3433 spa->spa_state = POOL_STATE_ACTIVE; 3434 spa->spa_ubsync = spa->spa_uberblock; 3435 spa->spa_verify_min_txg = spa->spa_extreme_rewind ? 3436 TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1; 3437 spa->spa_first_txg = spa->spa_last_ubsync_txg ? 3438 spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1; 3439 spa->spa_claim_max_txg = spa->spa_first_txg; 3440 spa->spa_prev_software_version = ub->ub_software_version; 3441 } 3442 3443 static int 3444 spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type) 3445 { 3446 vdev_t *rvd = spa->spa_root_vdev; 3447 nvlist_t *label; 3448 uberblock_t *ub = &spa->spa_uberblock; 3449 boolean_t activity_check = B_FALSE; 3450 3451 /* 3452 * If we are opening the checkpointed state of the pool by 3453 * rewinding to it, at this point we will have written the 3454 * checkpointed uberblock to the vdev labels, so searching 3455 * the labels will find the right uberblock. However, if 3456 * we are opening the checkpointed state read-only, we have 3457 * not modified the labels. Therefore, we must ignore the 3458 * labels and continue using the spa_uberblock that was set 3459 * by spa_ld_checkpoint_rewind. 3460 * 3461 * Note that it would be fine to ignore the labels when 3462 * rewinding (opening writeable) as well. However, if we 3463 * crash just after writing the labels, we will end up 3464 * searching the labels. Doing so in the common case means 3465 * that this code path gets exercised normally, rather than 3466 * just in the edge case. 3467 */ 3468 if (ub->ub_checkpoint_txg != 0 && 3469 spa_importing_readonly_checkpoint(spa)) { 3470 spa_ld_select_uberblock_done(spa, ub); 3471 return (0); 3472 } 3473 3474 /* 3475 * Find the best uberblock. 3476 */ 3477 vdev_uberblock_load(rvd, ub, &label); 3478 3479 /* 3480 * If we weren't able to find a single valid uberblock, return failure. 3481 */ 3482 if (ub->ub_txg == 0) { 3483 nvlist_free(label); 3484 spa_load_failed(spa, "no valid uberblock found"); 3485 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO)); 3486 } 3487 3488 if (spa->spa_load_max_txg != UINT64_MAX) { 3489 (void) spa_import_progress_set_max_txg(spa_guid(spa), 3490 (u_longlong_t)spa->spa_load_max_txg); 3491 } 3492 spa_load_note(spa, "using uberblock with txg=%llu", 3493 (u_longlong_t)ub->ub_txg); 3494 3495 3496 /* 3497 * For pools which have the multihost property on determine if the 3498 * pool is truly inactive and can be safely imported. Prevent 3499 * hosts which don't have a hostid set from importing the pool. 3500 */ 3501 activity_check = spa_activity_check_required(spa, ub, label, 3502 spa->spa_config); 3503 if (activity_check) { 3504 if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay && 3505 spa_get_hostid(spa) == 0) { 3506 nvlist_free(label); 3507 fnvlist_add_uint64(spa->spa_load_info, 3508 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID); 3509 return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO)); 3510 } 3511 3512 int error = spa_activity_check(spa, ub, spa->spa_config); 3513 if (error) { 3514 nvlist_free(label); 3515 return (error); 3516 } 3517 3518 fnvlist_add_uint64(spa->spa_load_info, 3519 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE); 3520 fnvlist_add_uint64(spa->spa_load_info, 3521 ZPOOL_CONFIG_MMP_TXG, ub->ub_txg); 3522 fnvlist_add_uint16(spa->spa_load_info, 3523 ZPOOL_CONFIG_MMP_SEQ, 3524 (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)); 3525 } 3526 3527 /* 3528 * If the pool has an unsupported version we can't open it. 3529 */ 3530 if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) { 3531 nvlist_free(label); 3532 spa_load_failed(spa, "version %llu is not supported", 3533 (u_longlong_t)ub->ub_version); 3534 return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP)); 3535 } 3536 3537 if (ub->ub_version >= SPA_VERSION_FEATURES) { 3538 nvlist_t *features; 3539 3540 /* 3541 * If we weren't able to find what's necessary for reading the 3542 * MOS in the label, return failure. 3543 */ 3544 if (label == NULL) { 3545 spa_load_failed(spa, "label config unavailable"); 3546 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, 3547 ENXIO)); 3548 } 3549 3550 if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ, 3551 &features) != 0) { 3552 nvlist_free(label); 3553 spa_load_failed(spa, "invalid label: '%s' missing", 3554 ZPOOL_CONFIG_FEATURES_FOR_READ); 3555 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, 3556 ENXIO)); 3557 } 3558 3559 /* 3560 * Update our in-core representation with the definitive values 3561 * from the label. 3562 */ 3563 nvlist_free(spa->spa_label_features); 3564 VERIFY(nvlist_dup(features, &spa->spa_label_features, 0) == 0); 3565 } 3566 3567 nvlist_free(label); 3568 3569 /* 3570 * Look through entries in the label nvlist's features_for_read. If 3571 * there is a feature listed there which we don't understand then we 3572 * cannot open a pool. 3573 */ 3574 if (ub->ub_version >= SPA_VERSION_FEATURES) { 3575 nvlist_t *unsup_feat; 3576 3577 VERIFY(nvlist_alloc(&unsup_feat, NV_UNIQUE_NAME, KM_SLEEP) == 3578 0); 3579 3580 for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features, 3581 NULL); nvp != NULL; 3582 nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) { 3583 if (!zfeature_is_supported(nvpair_name(nvp))) { 3584 VERIFY(nvlist_add_string(unsup_feat, 3585 nvpair_name(nvp), "") == 0); 3586 } 3587 } 3588 3589 if (!nvlist_empty(unsup_feat)) { 3590 VERIFY(nvlist_add_nvlist(spa->spa_load_info, 3591 ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat) == 0); 3592 nvlist_free(unsup_feat); 3593 spa_load_failed(spa, "some features are unsupported"); 3594 return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT, 3595 ENOTSUP)); 3596 } 3597 3598 nvlist_free(unsup_feat); 3599 } 3600 3601 if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) { 3602 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3603 spa_try_repair(spa, spa->spa_config); 3604 spa_config_exit(spa, SCL_ALL, FTAG); 3605 nvlist_free(spa->spa_config_splitting); 3606 spa->spa_config_splitting = NULL; 3607 } 3608 3609 /* 3610 * Initialize internal SPA structures. 3611 */ 3612 spa_ld_select_uberblock_done(spa, ub); 3613 3614 return (0); 3615 } 3616 3617 static int 3618 spa_ld_open_rootbp(spa_t *spa) 3619 { 3620 int error = 0; 3621 vdev_t *rvd = spa->spa_root_vdev; 3622 3623 error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool); 3624 if (error != 0) { 3625 spa_load_failed(spa, "unable to open rootbp in dsl_pool_init " 3626 "[error=%d]", error); 3627 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3628 } 3629 spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset; 3630 3631 return (0); 3632 } 3633 3634 static int 3635 spa_ld_trusted_config(spa_t *spa, spa_import_type_t type, 3636 boolean_t reloading) 3637 { 3638 vdev_t *mrvd, *rvd = spa->spa_root_vdev; 3639 nvlist_t *nv, *mos_config, *policy; 3640 int error = 0, copy_error; 3641 uint64_t healthy_tvds, healthy_tvds_mos; 3642 uint64_t mos_config_txg; 3643 3644 if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE) 3645 != 0) 3646 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3647 3648 /* 3649 * If we're assembling a pool from a split, the config provided is 3650 * already trusted so there is nothing to do. 3651 */ 3652 if (type == SPA_IMPORT_ASSEMBLE) 3653 return (0); 3654 3655 healthy_tvds = spa_healthy_core_tvds(spa); 3656 3657 if (load_nvlist(spa, spa->spa_config_object, &mos_config) 3658 != 0) { 3659 spa_load_failed(spa, "unable to retrieve MOS config"); 3660 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3661 } 3662 3663 /* 3664 * If we are doing an open, pool owner wasn't verified yet, thus do 3665 * the verification here. 3666 */ 3667 if (spa->spa_load_state == SPA_LOAD_OPEN) { 3668 error = spa_verify_host(spa, mos_config); 3669 if (error != 0) { 3670 nvlist_free(mos_config); 3671 return (error); 3672 } 3673 } 3674 3675 nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE); 3676 3677 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 3678 3679 /* 3680 * Build a new vdev tree from the trusted config 3681 */ 3682 error = spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD); 3683 if (error != 0) { 3684 nvlist_free(mos_config); 3685 spa_config_exit(spa, SCL_ALL, FTAG); 3686 spa_load_failed(spa, "spa_config_parse failed [error=%d]", 3687 error); 3688 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 3689 } 3690 3691 /* 3692 * Vdev paths in the MOS may be obsolete. If the untrusted config was 3693 * obtained by scanning /dev/dsk, then it will have the right vdev 3694 * paths. We update the trusted MOS config with this information. 3695 * We first try to copy the paths with vdev_copy_path_strict, which 3696 * succeeds only when both configs have exactly the same vdev tree. 3697 * If that fails, we fall back to a more flexible method that has a 3698 * best effort policy. 3699 */ 3700 copy_error = vdev_copy_path_strict(rvd, mrvd); 3701 if (copy_error != 0 || spa_load_print_vdev_tree) { 3702 spa_load_note(spa, "provided vdev tree:"); 3703 vdev_dbgmsg_print_tree(rvd, 2); 3704 spa_load_note(spa, "MOS vdev tree:"); 3705 vdev_dbgmsg_print_tree(mrvd, 2); 3706 } 3707 if (copy_error != 0) { 3708 spa_load_note(spa, "vdev_copy_path_strict failed, falling " 3709 "back to vdev_copy_path_relaxed"); 3710 vdev_copy_path_relaxed(rvd, mrvd); 3711 } 3712 3713 vdev_close(rvd); 3714 vdev_free(rvd); 3715 spa->spa_root_vdev = mrvd; 3716 rvd = mrvd; 3717 spa_config_exit(spa, SCL_ALL, FTAG); 3718 3719 /* 3720 * We will use spa_config if we decide to reload the spa or if spa_load 3721 * fails and we rewind. We must thus regenerate the config using the 3722 * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to 3723 * pass settings on how to load the pool and is not stored in the MOS. 3724 * We copy it over to our new, trusted config. 3725 */ 3726 mos_config_txg = fnvlist_lookup_uint64(mos_config, 3727 ZPOOL_CONFIG_POOL_TXG); 3728 nvlist_free(mos_config); 3729 mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE); 3730 if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY, 3731 &policy) == 0) 3732 fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy); 3733 spa_config_set(spa, mos_config); 3734 spa->spa_config_source = SPA_CONFIG_SRC_MOS; 3735 3736 /* 3737 * Now that we got the config from the MOS, we should be more strict 3738 * in checking blkptrs and can make assumptions about the consistency 3739 * of the vdev tree. spa_trust_config must be set to true before opening 3740 * vdevs in order for them to be writeable. 3741 */ 3742 spa->spa_trust_config = B_TRUE; 3743 3744 /* 3745 * Open and validate the new vdev tree 3746 */ 3747 error = spa_ld_open_vdevs(spa); 3748 if (error != 0) 3749 return (error); 3750 3751 error = spa_ld_validate_vdevs(spa); 3752 if (error != 0) 3753 return (error); 3754 3755 if (copy_error != 0 || spa_load_print_vdev_tree) { 3756 spa_load_note(spa, "final vdev tree:"); 3757 vdev_dbgmsg_print_tree(rvd, 2); 3758 } 3759 3760 if (spa->spa_load_state != SPA_LOAD_TRYIMPORT && 3761 !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) { 3762 /* 3763 * Sanity check to make sure that we are indeed loading the 3764 * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds 3765 * in the config provided and they happened to be the only ones 3766 * to have the latest uberblock, we could involuntarily perform 3767 * an extreme rewind. 3768 */ 3769 healthy_tvds_mos = spa_healthy_core_tvds(spa); 3770 if (healthy_tvds_mos - healthy_tvds >= 3771 SPA_SYNC_MIN_VDEVS) { 3772 spa_load_note(spa, "config provided misses too many " 3773 "top-level vdevs compared to MOS (%lld vs %lld). ", 3774 (u_longlong_t)healthy_tvds, 3775 (u_longlong_t)healthy_tvds_mos); 3776 spa_load_note(spa, "vdev tree:"); 3777 vdev_dbgmsg_print_tree(rvd, 2); 3778 if (reloading) { 3779 spa_load_failed(spa, "config was already " 3780 "provided from MOS. Aborting."); 3781 return (spa_vdev_err(rvd, 3782 VDEV_AUX_CORRUPT_DATA, EIO)); 3783 } 3784 spa_load_note(spa, "spa must be reloaded using MOS " 3785 "config"); 3786 return (SET_ERROR(EAGAIN)); 3787 } 3788 } 3789 3790 error = spa_check_for_missing_logs(spa); 3791 if (error != 0) 3792 return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO)); 3793 3794 if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) { 3795 spa_load_failed(spa, "uberblock guid sum doesn't match MOS " 3796 "guid sum (%llu != %llu)", 3797 (u_longlong_t)spa->spa_uberblock.ub_guid_sum, 3798 (u_longlong_t)rvd->vdev_guid_sum); 3799 return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, 3800 ENXIO)); 3801 } 3802 3803 return (0); 3804 } 3805 3806 static int 3807 spa_ld_open_indirect_vdev_metadata(spa_t *spa) 3808 { 3809 int error = 0; 3810 vdev_t *rvd = spa->spa_root_vdev; 3811 3812 /* 3813 * Everything that we read before spa_remove_init() must be stored 3814 * on concreted vdevs. Therefore we do this as early as possible. 3815 */ 3816 error = spa_remove_init(spa); 3817 if (error != 0) { 3818 spa_load_failed(spa, "spa_remove_init failed [error=%d]", 3819 error); 3820 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3821 } 3822 3823 /* 3824 * Retrieve information needed to condense indirect vdev mappings. 3825 */ 3826 error = spa_condense_init(spa); 3827 if (error != 0) { 3828 spa_load_failed(spa, "spa_condense_init failed [error=%d]", 3829 error); 3830 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 3831 } 3832 3833 return (0); 3834 } 3835 3836 static int 3837 spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep) 3838 { 3839 int error = 0; 3840 vdev_t *rvd = spa->spa_root_vdev; 3841 3842 if (spa_version(spa) >= SPA_VERSION_FEATURES) { 3843 boolean_t missing_feat_read = B_FALSE; 3844 nvlist_t *unsup_feat, *enabled_feat; 3845 3846 if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ, 3847 &spa->spa_feat_for_read_obj, B_TRUE) != 0) { 3848 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3849 } 3850 3851 if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE, 3852 &spa->spa_feat_for_write_obj, B_TRUE) != 0) { 3853 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3854 } 3855 3856 if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS, 3857 &spa->spa_feat_desc_obj, B_TRUE) != 0) { 3858 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3859 } 3860 3861 enabled_feat = fnvlist_alloc(); 3862 unsup_feat = fnvlist_alloc(); 3863 3864 if (!spa_features_check(spa, B_FALSE, 3865 unsup_feat, enabled_feat)) 3866 missing_feat_read = B_TRUE; 3867 3868 if (spa_writeable(spa) || 3869 spa->spa_load_state == SPA_LOAD_TRYIMPORT) { 3870 if (!spa_features_check(spa, B_TRUE, 3871 unsup_feat, enabled_feat)) { 3872 *missing_feat_writep = B_TRUE; 3873 } 3874 } 3875 3876 fnvlist_add_nvlist(spa->spa_load_info, 3877 ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat); 3878 3879 if (!nvlist_empty(unsup_feat)) { 3880 fnvlist_add_nvlist(spa->spa_load_info, 3881 ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat); 3882 } 3883 3884 fnvlist_free(enabled_feat); 3885 fnvlist_free(unsup_feat); 3886 3887 if (!missing_feat_read) { 3888 fnvlist_add_boolean(spa->spa_load_info, 3889 ZPOOL_CONFIG_CAN_RDONLY); 3890 } 3891 3892 /* 3893 * If the state is SPA_LOAD_TRYIMPORT, our objective is 3894 * twofold: to determine whether the pool is available for 3895 * import in read-write mode and (if it is not) whether the 3896 * pool is available for import in read-only mode. If the pool 3897 * is available for import in read-write mode, it is displayed 3898 * as available in userland; if it is not available for import 3899 * in read-only mode, it is displayed as unavailable in 3900 * userland. If the pool is available for import in read-only 3901 * mode but not read-write mode, it is displayed as unavailable 3902 * in userland with a special note that the pool is actually 3903 * available for open in read-only mode. 3904 * 3905 * As a result, if the state is SPA_LOAD_TRYIMPORT and we are 3906 * missing a feature for write, we must first determine whether 3907 * the pool can be opened read-only before returning to 3908 * userland in order to know whether to display the 3909 * abovementioned note. 3910 */ 3911 if (missing_feat_read || (*missing_feat_writep && 3912 spa_writeable(spa))) { 3913 spa_load_failed(spa, "pool uses unsupported features"); 3914 return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT, 3915 ENOTSUP)); 3916 } 3917 3918 /* 3919 * Load refcounts for ZFS features from disk into an in-memory 3920 * cache during SPA initialization. 3921 */ 3922 for (spa_feature_t i = 0; i < SPA_FEATURES; i++) { 3923 uint64_t refcount; 3924 3925 error = feature_get_refcount_from_disk(spa, 3926 &spa_feature_table[i], &refcount); 3927 if (error == 0) { 3928 spa->spa_feat_refcount_cache[i] = refcount; 3929 } else if (error == ENOTSUP) { 3930 spa->spa_feat_refcount_cache[i] = 3931 SPA_FEATURE_DISABLED; 3932 } else { 3933 spa_load_failed(spa, "error getting refcount " 3934 "for feature %s [error=%d]", 3935 spa_feature_table[i].fi_guid, error); 3936 return (spa_vdev_err(rvd, 3937 VDEV_AUX_CORRUPT_DATA, EIO)); 3938 } 3939 } 3940 } 3941 3942 if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) { 3943 if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG, 3944 &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0) 3945 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3946 } 3947 3948 /* 3949 * Encryption was added before bookmark_v2, even though bookmark_v2 3950 * is now a dependency. If this pool has encryption enabled without 3951 * bookmark_v2, trigger an errata message. 3952 */ 3953 if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) && 3954 !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) { 3955 spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION; 3956 } 3957 3958 return (0); 3959 } 3960 3961 static int 3962 spa_ld_load_special_directories(spa_t *spa) 3963 { 3964 int error = 0; 3965 vdev_t *rvd = spa->spa_root_vdev; 3966 3967 spa->spa_is_initializing = B_TRUE; 3968 error = dsl_pool_open(spa->spa_dsl_pool); 3969 spa->spa_is_initializing = B_FALSE; 3970 if (error != 0) { 3971 spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error); 3972 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3973 } 3974 3975 return (0); 3976 } 3977 3978 static int 3979 spa_ld_get_props(spa_t *spa) 3980 { 3981 int error = 0; 3982 uint64_t obj; 3983 vdev_t *rvd = spa->spa_root_vdev; 3984 3985 /* Grab the checksum salt from the MOS. */ 3986 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 3987 DMU_POOL_CHECKSUM_SALT, 1, 3988 sizeof (spa->spa_cksum_salt.zcs_bytes), 3989 spa->spa_cksum_salt.zcs_bytes); 3990 if (error == ENOENT) { 3991 /* Generate a new salt for subsequent use */ 3992 (void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes, 3993 sizeof (spa->spa_cksum_salt.zcs_bytes)); 3994 } else if (error != 0) { 3995 spa_load_failed(spa, "unable to retrieve checksum salt from " 3996 "MOS [error=%d]", error); 3997 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 3998 } 3999 4000 if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0) 4001 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4002 error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj); 4003 if (error != 0) { 4004 spa_load_failed(spa, "error opening deferred-frees bpobj " 4005 "[error=%d]", error); 4006 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4007 } 4008 4009 /* 4010 * Load the bit that tells us to use the new accounting function 4011 * (raid-z deflation). If we have an older pool, this will not 4012 * be present. 4013 */ 4014 error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE); 4015 if (error != 0 && error != ENOENT) 4016 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4017 4018 error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION, 4019 &spa->spa_creation_version, B_FALSE); 4020 if (error != 0 && error != ENOENT) 4021 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4022 4023 /* 4024 * Load the persistent error log. If we have an older pool, this will 4025 * not be present. 4026 */ 4027 error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last, 4028 B_FALSE); 4029 if (error != 0 && error != ENOENT) 4030 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4031 4032 error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB, 4033 &spa->spa_errlog_scrub, B_FALSE); 4034 if (error != 0 && error != ENOENT) 4035 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4036 4037 /* 4038 * Load the livelist deletion field. If a livelist is queued for 4039 * deletion, indicate that in the spa 4040 */ 4041 error = spa_dir_prop(spa, DMU_POOL_DELETED_CLONES, 4042 &spa->spa_livelists_to_delete, B_FALSE); 4043 if (error != 0 && error != ENOENT) 4044 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4045 4046 /* 4047 * Load the history object. If we have an older pool, this 4048 * will not be present. 4049 */ 4050 error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE); 4051 if (error != 0 && error != ENOENT) 4052 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4053 4054 /* 4055 * Load the per-vdev ZAP map. If we have an older pool, this will not 4056 * be present; in this case, defer its creation to a later time to 4057 * avoid dirtying the MOS this early / out of sync context. See 4058 * spa_sync_config_object. 4059 */ 4060 4061 /* The sentinel is only available in the MOS config. */ 4062 nvlist_t *mos_config; 4063 if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) { 4064 spa_load_failed(spa, "unable to retrieve MOS config"); 4065 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4066 } 4067 4068 error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP, 4069 &spa->spa_all_vdev_zaps, B_FALSE); 4070 4071 if (error == ENOENT) { 4072 VERIFY(!nvlist_exists(mos_config, 4073 ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)); 4074 spa->spa_avz_action = AVZ_ACTION_INITIALIZE; 4075 ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev)); 4076 } else if (error != 0) { 4077 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4078 } else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) { 4079 /* 4080 * An older version of ZFS overwrote the sentinel value, so 4081 * we have orphaned per-vdev ZAPs in the MOS. Defer their 4082 * destruction to later; see spa_sync_config_object. 4083 */ 4084 spa->spa_avz_action = AVZ_ACTION_DESTROY; 4085 /* 4086 * We're assuming that no vdevs have had their ZAPs created 4087 * before this. Better be sure of it. 4088 */ 4089 ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev)); 4090 } 4091 nvlist_free(mos_config); 4092 4093 spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION); 4094 4095 error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object, 4096 B_FALSE); 4097 if (error && error != ENOENT) 4098 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4099 4100 if (error == 0) { 4101 uint64_t autoreplace; 4102 4103 spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs); 4104 spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace); 4105 spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation); 4106 spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode); 4107 spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand); 4108 spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost); 4109 spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim); 4110 spa->spa_autoreplace = (autoreplace != 0); 4111 } 4112 4113 /* 4114 * If we are importing a pool with missing top-level vdevs, 4115 * we enforce that the pool doesn't panic or get suspended on 4116 * error since the likelihood of missing data is extremely high. 4117 */ 4118 if (spa->spa_missing_tvds > 0 && 4119 spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE && 4120 spa->spa_load_state != SPA_LOAD_TRYIMPORT) { 4121 spa_load_note(spa, "forcing failmode to 'continue' " 4122 "as some top level vdevs are missing"); 4123 spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE; 4124 } 4125 4126 return (0); 4127 } 4128 4129 static int 4130 spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type) 4131 { 4132 int error = 0; 4133 vdev_t *rvd = spa->spa_root_vdev; 4134 4135 /* 4136 * If we're assembling the pool from the split-off vdevs of 4137 * an existing pool, we don't want to attach the spares & cache 4138 * devices. 4139 */ 4140 4141 /* 4142 * Load any hot spares for this pool. 4143 */ 4144 error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object, 4145 B_FALSE); 4146 if (error != 0 && error != ENOENT) 4147 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4148 if (error == 0 && type != SPA_IMPORT_ASSEMBLE) { 4149 ASSERT(spa_version(spa) >= SPA_VERSION_SPARES); 4150 if (load_nvlist(spa, spa->spa_spares.sav_object, 4151 &spa->spa_spares.sav_config) != 0) { 4152 spa_load_failed(spa, "error loading spares nvlist"); 4153 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4154 } 4155 4156 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4157 spa_load_spares(spa); 4158 spa_config_exit(spa, SCL_ALL, FTAG); 4159 } else if (error == 0) { 4160 spa->spa_spares.sav_sync = B_TRUE; 4161 } 4162 4163 /* 4164 * Load any level 2 ARC devices for this pool. 4165 */ 4166 error = spa_dir_prop(spa, DMU_POOL_L2CACHE, 4167 &spa->spa_l2cache.sav_object, B_FALSE); 4168 if (error != 0 && error != ENOENT) 4169 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4170 if (error == 0 && type != SPA_IMPORT_ASSEMBLE) { 4171 ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE); 4172 if (load_nvlist(spa, spa->spa_l2cache.sav_object, 4173 &spa->spa_l2cache.sav_config) != 0) { 4174 spa_load_failed(spa, "error loading l2cache nvlist"); 4175 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4176 } 4177 4178 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4179 spa_load_l2cache(spa); 4180 spa_config_exit(spa, SCL_ALL, FTAG); 4181 } else if (error == 0) { 4182 spa->spa_l2cache.sav_sync = B_TRUE; 4183 } 4184 4185 return (0); 4186 } 4187 4188 static int 4189 spa_ld_load_vdev_metadata(spa_t *spa) 4190 { 4191 int error = 0; 4192 vdev_t *rvd = spa->spa_root_vdev; 4193 4194 /* 4195 * If the 'multihost' property is set, then never allow a pool to 4196 * be imported when the system hostid is zero. The exception to 4197 * this rule is zdb which is always allowed to access pools. 4198 */ 4199 if (spa_multihost(spa) && spa_get_hostid(spa) == 0 && 4200 (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) { 4201 fnvlist_add_uint64(spa->spa_load_info, 4202 ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID); 4203 return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO)); 4204 } 4205 4206 /* 4207 * If the 'autoreplace' property is set, then post a resource notifying 4208 * the ZFS DE that it should not issue any faults for unopenable 4209 * devices. We also iterate over the vdevs, and post a sysevent for any 4210 * unopenable vdevs so that the normal autoreplace handler can take 4211 * over. 4212 */ 4213 if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) { 4214 spa_check_removed(spa->spa_root_vdev); 4215 /* 4216 * For the import case, this is done in spa_import(), because 4217 * at this point we're using the spare definitions from 4218 * the MOS config, not necessarily from the userland config. 4219 */ 4220 if (spa->spa_load_state != SPA_LOAD_IMPORT) { 4221 spa_aux_check_removed(&spa->spa_spares); 4222 spa_aux_check_removed(&spa->spa_l2cache); 4223 } 4224 } 4225 4226 /* 4227 * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc. 4228 */ 4229 error = vdev_load(rvd); 4230 if (error != 0) { 4231 spa_load_failed(spa, "vdev_load failed [error=%d]", error); 4232 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 4233 } 4234 4235 error = spa_ld_log_spacemaps(spa); 4236 if (error != 0) { 4237 spa_load_failed(spa, "spa_ld_log_sm_data failed [error=%d]", 4238 error); 4239 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error)); 4240 } 4241 4242 /* 4243 * Propagate the leaf DTLs we just loaded all the way up the vdev tree. 4244 */ 4245 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4246 vdev_dtl_reassess(rvd, 0, 0, B_FALSE, B_FALSE); 4247 spa_config_exit(spa, SCL_ALL, FTAG); 4248 4249 return (0); 4250 } 4251 4252 static int 4253 spa_ld_load_dedup_tables(spa_t *spa) 4254 { 4255 int error = 0; 4256 vdev_t *rvd = spa->spa_root_vdev; 4257 4258 error = ddt_load(spa); 4259 if (error != 0) { 4260 spa_load_failed(spa, "ddt_load failed [error=%d]", error); 4261 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO)); 4262 } 4263 4264 return (0); 4265 } 4266 4267 static int 4268 spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, char **ereport) 4269 { 4270 vdev_t *rvd = spa->spa_root_vdev; 4271 4272 if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) { 4273 boolean_t missing = spa_check_logs(spa); 4274 if (missing) { 4275 if (spa->spa_missing_tvds != 0) { 4276 spa_load_note(spa, "spa_check_logs failed " 4277 "so dropping the logs"); 4278 } else { 4279 *ereport = FM_EREPORT_ZFS_LOG_REPLAY; 4280 spa_load_failed(spa, "spa_check_logs failed"); 4281 return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG, 4282 ENXIO)); 4283 } 4284 } 4285 } 4286 4287 return (0); 4288 } 4289 4290 static int 4291 spa_ld_verify_pool_data(spa_t *spa) 4292 { 4293 int error = 0; 4294 vdev_t *rvd = spa->spa_root_vdev; 4295 4296 /* 4297 * We've successfully opened the pool, verify that we're ready 4298 * to start pushing transactions. 4299 */ 4300 if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) { 4301 error = spa_load_verify(spa); 4302 if (error != 0) { 4303 spa_load_failed(spa, "spa_load_verify failed " 4304 "[error=%d]", error); 4305 return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, 4306 error)); 4307 } 4308 } 4309 4310 return (0); 4311 } 4312 4313 static void 4314 spa_ld_claim_log_blocks(spa_t *spa) 4315 { 4316 dmu_tx_t *tx; 4317 dsl_pool_t *dp = spa_get_dsl(spa); 4318 4319 /* 4320 * Claim log blocks that haven't been committed yet. 4321 * This must all happen in a single txg. 4322 * Note: spa_claim_max_txg is updated by spa_claim_notify(), 4323 * invoked from zil_claim_log_block()'s i/o done callback. 4324 * Price of rollback is that we abandon the log. 4325 */ 4326 spa->spa_claiming = B_TRUE; 4327 4328 tx = dmu_tx_create_assigned(dp, spa_first_txg(spa)); 4329 (void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj, 4330 zil_claim, tx, DS_FIND_CHILDREN); 4331 dmu_tx_commit(tx); 4332 4333 spa->spa_claiming = B_FALSE; 4334 4335 spa_set_log_state(spa, SPA_LOG_GOOD); 4336 } 4337 4338 static void 4339 spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg, 4340 boolean_t update_config_cache) 4341 { 4342 vdev_t *rvd = spa->spa_root_vdev; 4343 int need_update = B_FALSE; 4344 4345 /* 4346 * If the config cache is stale, or we have uninitialized 4347 * metaslabs (see spa_vdev_add()), then update the config. 4348 * 4349 * If this is a verbatim import, trust the current 4350 * in-core spa_config and update the disk labels. 4351 */ 4352 if (update_config_cache || config_cache_txg != spa->spa_config_txg || 4353 spa->spa_load_state == SPA_LOAD_IMPORT || 4354 spa->spa_load_state == SPA_LOAD_RECOVER || 4355 (spa->spa_import_flags & ZFS_IMPORT_VERBATIM)) 4356 need_update = B_TRUE; 4357 4358 for (int c = 0; c < rvd->vdev_children; c++) 4359 if (rvd->vdev_child[c]->vdev_ms_array == 0) 4360 need_update = B_TRUE; 4361 4362 /* 4363 * Update the config cache asynchronously in case we're the 4364 * root pool, in which case the config cache isn't writable yet. 4365 */ 4366 if (need_update) 4367 spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE); 4368 } 4369 4370 static void 4371 spa_ld_prepare_for_reload(spa_t *spa) 4372 { 4373 spa_mode_t mode = spa->spa_mode; 4374 int async_suspended = spa->spa_async_suspended; 4375 4376 spa_unload(spa); 4377 spa_deactivate(spa); 4378 spa_activate(spa, mode); 4379 4380 /* 4381 * We save the value of spa_async_suspended as it gets reset to 0 by 4382 * spa_unload(). We want to restore it back to the original value before 4383 * returning as we might be calling spa_async_resume() later. 4384 */ 4385 spa->spa_async_suspended = async_suspended; 4386 } 4387 4388 static int 4389 spa_ld_read_checkpoint_txg(spa_t *spa) 4390 { 4391 uberblock_t checkpoint; 4392 int error = 0; 4393 4394 ASSERT0(spa->spa_checkpoint_txg); 4395 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4396 4397 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 4398 DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t), 4399 sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint); 4400 4401 if (error == ENOENT) 4402 return (0); 4403 4404 if (error != 0) 4405 return (error); 4406 4407 ASSERT3U(checkpoint.ub_txg, !=, 0); 4408 ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0); 4409 ASSERT3U(checkpoint.ub_timestamp, !=, 0); 4410 spa->spa_checkpoint_txg = checkpoint.ub_txg; 4411 spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp; 4412 4413 return (0); 4414 } 4415 4416 static int 4417 spa_ld_mos_init(spa_t *spa, spa_import_type_t type) 4418 { 4419 int error = 0; 4420 4421 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4422 ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE); 4423 4424 /* 4425 * Never trust the config that is provided unless we are assembling 4426 * a pool following a split. 4427 * This means don't trust blkptrs and the vdev tree in general. This 4428 * also effectively puts the spa in read-only mode since 4429 * spa_writeable() checks for spa_trust_config to be true. 4430 * We will later load a trusted config from the MOS. 4431 */ 4432 if (type != SPA_IMPORT_ASSEMBLE) 4433 spa->spa_trust_config = B_FALSE; 4434 4435 /* 4436 * Parse the config provided to create a vdev tree. 4437 */ 4438 error = spa_ld_parse_config(spa, type); 4439 if (error != 0) 4440 return (error); 4441 4442 spa_import_progress_add(spa); 4443 4444 /* 4445 * Now that we have the vdev tree, try to open each vdev. This involves 4446 * opening the underlying physical device, retrieving its geometry and 4447 * probing the vdev with a dummy I/O. The state of each vdev will be set 4448 * based on the success of those operations. After this we'll be ready 4449 * to read from the vdevs. 4450 */ 4451 error = spa_ld_open_vdevs(spa); 4452 if (error != 0) 4453 return (error); 4454 4455 /* 4456 * Read the label of each vdev and make sure that the GUIDs stored 4457 * there match the GUIDs in the config provided. 4458 * If we're assembling a new pool that's been split off from an 4459 * existing pool, the labels haven't yet been updated so we skip 4460 * validation for now. 4461 */ 4462 if (type != SPA_IMPORT_ASSEMBLE) { 4463 error = spa_ld_validate_vdevs(spa); 4464 if (error != 0) 4465 return (error); 4466 } 4467 4468 /* 4469 * Read all vdev labels to find the best uberblock (i.e. latest, 4470 * unless spa_load_max_txg is set) and store it in spa_uberblock. We 4471 * get the list of features required to read blkptrs in the MOS from 4472 * the vdev label with the best uberblock and verify that our version 4473 * of zfs supports them all. 4474 */ 4475 error = spa_ld_select_uberblock(spa, type); 4476 if (error != 0) 4477 return (error); 4478 4479 /* 4480 * Pass that uberblock to the dsl_pool layer which will open the root 4481 * blkptr. This blkptr points to the latest version of the MOS and will 4482 * allow us to read its contents. 4483 */ 4484 error = spa_ld_open_rootbp(spa); 4485 if (error != 0) 4486 return (error); 4487 4488 return (0); 4489 } 4490 4491 static int 4492 spa_ld_checkpoint_rewind(spa_t *spa) 4493 { 4494 uberblock_t checkpoint; 4495 int error = 0; 4496 4497 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4498 ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); 4499 4500 error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 4501 DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t), 4502 sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint); 4503 4504 if (error != 0) { 4505 spa_load_failed(spa, "unable to retrieve checkpointed " 4506 "uberblock from the MOS config [error=%d]", error); 4507 4508 if (error == ENOENT) 4509 error = ZFS_ERR_NO_CHECKPOINT; 4510 4511 return (error); 4512 } 4513 4514 ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg); 4515 ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg); 4516 4517 /* 4518 * We need to update the txg and timestamp of the checkpointed 4519 * uberblock to be higher than the latest one. This ensures that 4520 * the checkpointed uberblock is selected if we were to close and 4521 * reopen the pool right after we've written it in the vdev labels. 4522 * (also see block comment in vdev_uberblock_compare) 4523 */ 4524 checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1; 4525 checkpoint.ub_timestamp = gethrestime_sec(); 4526 4527 /* 4528 * Set current uberblock to be the checkpointed uberblock. 4529 */ 4530 spa->spa_uberblock = checkpoint; 4531 4532 /* 4533 * If we are doing a normal rewind, then the pool is open for 4534 * writing and we sync the "updated" checkpointed uberblock to 4535 * disk. Once this is done, we've basically rewound the whole 4536 * pool and there is no way back. 4537 * 4538 * There are cases when we don't want to attempt and sync the 4539 * checkpointed uberblock to disk because we are opening a 4540 * pool as read-only. Specifically, verifying the checkpointed 4541 * state with zdb, and importing the checkpointed state to get 4542 * a "preview" of its content. 4543 */ 4544 if (spa_writeable(spa)) { 4545 vdev_t *rvd = spa->spa_root_vdev; 4546 4547 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 4548 vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL }; 4549 int svdcount = 0; 4550 int children = rvd->vdev_children; 4551 int c0 = spa_get_random(children); 4552 4553 for (int c = 0; c < children; c++) { 4554 vdev_t *vd = rvd->vdev_child[(c0 + c) % children]; 4555 4556 /* Stop when revisiting the first vdev */ 4557 if (c > 0 && svd[0] == vd) 4558 break; 4559 4560 if (vd->vdev_ms_array == 0 || vd->vdev_islog || 4561 !vdev_is_concrete(vd)) 4562 continue; 4563 4564 svd[svdcount++] = vd; 4565 if (svdcount == SPA_SYNC_MIN_VDEVS) 4566 break; 4567 } 4568 error = vdev_config_sync(svd, svdcount, spa->spa_first_txg); 4569 if (error == 0) 4570 spa->spa_last_synced_guid = rvd->vdev_guid; 4571 spa_config_exit(spa, SCL_ALL, FTAG); 4572 4573 if (error != 0) { 4574 spa_load_failed(spa, "failed to write checkpointed " 4575 "uberblock to the vdev labels [error=%d]", error); 4576 return (error); 4577 } 4578 } 4579 4580 return (0); 4581 } 4582 4583 static int 4584 spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type, 4585 boolean_t *update_config_cache) 4586 { 4587 int error; 4588 4589 /* 4590 * Parse the config for pool, open and validate vdevs, 4591 * select an uberblock, and use that uberblock to open 4592 * the MOS. 4593 */ 4594 error = spa_ld_mos_init(spa, type); 4595 if (error != 0) 4596 return (error); 4597 4598 /* 4599 * Retrieve the trusted config stored in the MOS and use it to create 4600 * a new, exact version of the vdev tree, then reopen all vdevs. 4601 */ 4602 error = spa_ld_trusted_config(spa, type, B_FALSE); 4603 if (error == EAGAIN) { 4604 if (update_config_cache != NULL) 4605 *update_config_cache = B_TRUE; 4606 4607 /* 4608 * Redo the loading process with the trusted config if it is 4609 * too different from the untrusted config. 4610 */ 4611 spa_ld_prepare_for_reload(spa); 4612 spa_load_note(spa, "RELOADING"); 4613 error = spa_ld_mos_init(spa, type); 4614 if (error != 0) 4615 return (error); 4616 4617 error = spa_ld_trusted_config(spa, type, B_TRUE); 4618 if (error != 0) 4619 return (error); 4620 4621 } else if (error != 0) { 4622 return (error); 4623 } 4624 4625 return (0); 4626 } 4627 4628 /* 4629 * Load an existing storage pool, using the config provided. This config 4630 * describes which vdevs are part of the pool and is later validated against 4631 * partial configs present in each vdev's label and an entire copy of the 4632 * config stored in the MOS. 4633 */ 4634 static int 4635 spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport) 4636 { 4637 int error = 0; 4638 boolean_t missing_feat_write = B_FALSE; 4639 boolean_t checkpoint_rewind = 4640 (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); 4641 boolean_t update_config_cache = B_FALSE; 4642 4643 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 4644 ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE); 4645 4646 spa_load_note(spa, "LOADING"); 4647 4648 error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache); 4649 if (error != 0) 4650 return (error); 4651 4652 /* 4653 * If we are rewinding to the checkpoint then we need to repeat 4654 * everything we've done so far in this function but this time 4655 * selecting the checkpointed uberblock and using that to open 4656 * the MOS. 4657 */ 4658 if (checkpoint_rewind) { 4659 /* 4660 * If we are rewinding to the checkpoint update config cache 4661 * anyway. 4662 */ 4663 update_config_cache = B_TRUE; 4664 4665 /* 4666 * Extract the checkpointed uberblock from the current MOS 4667 * and use this as the pool's uberblock from now on. If the 4668 * pool is imported as writeable we also write the checkpoint 4669 * uberblock to the labels, making the rewind permanent. 4670 */ 4671 error = spa_ld_checkpoint_rewind(spa); 4672 if (error != 0) 4673 return (error); 4674 4675 /* 4676 * Redo the loading process again with the 4677 * checkpointed uberblock. 4678 */ 4679 spa_ld_prepare_for_reload(spa); 4680 spa_load_note(spa, "LOADING checkpointed uberblock"); 4681 error = spa_ld_mos_with_trusted_config(spa, type, NULL); 4682 if (error != 0) 4683 return (error); 4684 } 4685 4686 /* 4687 * Retrieve the checkpoint txg if the pool has a checkpoint. 4688 */ 4689 error = spa_ld_read_checkpoint_txg(spa); 4690 if (error != 0) 4691 return (error); 4692 4693 /* 4694 * Retrieve the mapping of indirect vdevs. Those vdevs were removed 4695 * from the pool and their contents were re-mapped to other vdevs. Note 4696 * that everything that we read before this step must have been 4697 * rewritten on concrete vdevs after the last device removal was 4698 * initiated. Otherwise we could be reading from indirect vdevs before 4699 * we have loaded their mappings. 4700 */ 4701 error = spa_ld_open_indirect_vdev_metadata(spa); 4702 if (error != 0) 4703 return (error); 4704 4705 /* 4706 * Retrieve the full list of active features from the MOS and check if 4707 * they are all supported. 4708 */ 4709 error = spa_ld_check_features(spa, &missing_feat_write); 4710 if (error != 0) 4711 return (error); 4712 4713 /* 4714 * Load several special directories from the MOS needed by the dsl_pool 4715 * layer. 4716 */ 4717 error = spa_ld_load_special_directories(spa); 4718 if (error != 0) 4719 return (error); 4720 4721 /* 4722 * Retrieve pool properties from the MOS. 4723 */ 4724 error = spa_ld_get_props(spa); 4725 if (error != 0) 4726 return (error); 4727 4728 /* 4729 * Retrieve the list of auxiliary devices - cache devices and spares - 4730 * and open them. 4731 */ 4732 error = spa_ld_open_aux_vdevs(spa, type); 4733 if (error != 0) 4734 return (error); 4735 4736 /* 4737 * Load the metadata for all vdevs. Also check if unopenable devices 4738 * should be autoreplaced. 4739 */ 4740 error = spa_ld_load_vdev_metadata(spa); 4741 if (error != 0) 4742 return (error); 4743 4744 error = spa_ld_load_dedup_tables(spa); 4745 if (error != 0) 4746 return (error); 4747 4748 /* 4749 * Verify the logs now to make sure we don't have any unexpected errors 4750 * when we claim log blocks later. 4751 */ 4752 error = spa_ld_verify_logs(spa, type, ereport); 4753 if (error != 0) 4754 return (error); 4755 4756 if (missing_feat_write) { 4757 ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT); 4758 4759 /* 4760 * At this point, we know that we can open the pool in 4761 * read-only mode but not read-write mode. We now have enough 4762 * information and can return to userland. 4763 */ 4764 return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT, 4765 ENOTSUP)); 4766 } 4767 4768 /* 4769 * Traverse the last txgs to make sure the pool was left off in a safe 4770 * state. When performing an extreme rewind, we verify the whole pool, 4771 * which can take a very long time. 4772 */ 4773 error = spa_ld_verify_pool_data(spa); 4774 if (error != 0) 4775 return (error); 4776 4777 /* 4778 * Calculate the deflated space for the pool. This must be done before 4779 * we write anything to the pool because we'd need to update the space 4780 * accounting using the deflated sizes. 4781 */ 4782 spa_update_dspace(spa); 4783 4784 /* 4785 * We have now retrieved all the information we needed to open the 4786 * pool. If we are importing the pool in read-write mode, a few 4787 * additional steps must be performed to finish the import. 4788 */ 4789 if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER || 4790 spa->spa_load_max_txg == UINT64_MAX)) { 4791 uint64_t config_cache_txg = spa->spa_config_txg; 4792 4793 ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT); 4794 4795 /* 4796 * In case of a checkpoint rewind, log the original txg 4797 * of the checkpointed uberblock. 4798 */ 4799 if (checkpoint_rewind) { 4800 spa_history_log_internal(spa, "checkpoint rewind", 4801 NULL, "rewound state to txg=%llu", 4802 (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg); 4803 } 4804 4805 /* 4806 * Traverse the ZIL and claim all blocks. 4807 */ 4808 spa_ld_claim_log_blocks(spa); 4809 4810 /* 4811 * Kick-off the syncing thread. 4812 */ 4813 spa->spa_sync_on = B_TRUE; 4814 txg_sync_start(spa->spa_dsl_pool); 4815 mmp_thread_start(spa); 4816 4817 /* 4818 * Wait for all claims to sync. We sync up to the highest 4819 * claimed log block birth time so that claimed log blocks 4820 * don't appear to be from the future. spa_claim_max_txg 4821 * will have been set for us by ZIL traversal operations 4822 * performed above. 4823 */ 4824 txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg); 4825 4826 /* 4827 * Check if we need to request an update of the config. On the 4828 * next sync, we would update the config stored in vdev labels 4829 * and the cachefile (by default /etc/zfs/zpool.cache). 4830 */ 4831 spa_ld_check_for_config_update(spa, config_cache_txg, 4832 update_config_cache); 4833 4834 /* 4835 * Check if a rebuild was in progress and if so resume it. 4836 * Then check all DTLs to see if anything needs resilvering. 4837 * The resilver will be deferred if a rebuild was started. 4838 */ 4839 if (vdev_rebuild_active(spa->spa_root_vdev)) { 4840 vdev_rebuild_restart(spa); 4841 } else if (!dsl_scan_resilvering(spa->spa_dsl_pool) && 4842 vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) { 4843 spa_async_request(spa, SPA_ASYNC_RESILVER); 4844 } 4845 4846 /* 4847 * Log the fact that we booted up (so that we can detect if 4848 * we rebooted in the middle of an operation). 4849 */ 4850 spa_history_log_version(spa, "open", NULL); 4851 4852 spa_restart_removal(spa); 4853 spa_spawn_aux_threads(spa); 4854 4855 /* 4856 * Delete any inconsistent datasets. 4857 * 4858 * Note: 4859 * Since we may be issuing deletes for clones here, 4860 * we make sure to do so after we've spawned all the 4861 * auxiliary threads above (from which the livelist 4862 * deletion zthr is part of). 4863 */ 4864 (void) dmu_objset_find(spa_name(spa), 4865 dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN); 4866 4867 /* 4868 * Clean up any stale temporary dataset userrefs. 4869 */ 4870 dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool); 4871 4872 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 4873 vdev_initialize_restart(spa->spa_root_vdev); 4874 vdev_trim_restart(spa->spa_root_vdev); 4875 vdev_autotrim_restart(spa); 4876 spa_config_exit(spa, SCL_CONFIG, FTAG); 4877 } 4878 4879 spa_import_progress_remove(spa_guid(spa)); 4880 spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD); 4881 4882 spa_load_note(spa, "LOADED"); 4883 4884 return (0); 4885 } 4886 4887 static int 4888 spa_load_retry(spa_t *spa, spa_load_state_t state) 4889 { 4890 spa_mode_t mode = spa->spa_mode; 4891 4892 spa_unload(spa); 4893 spa_deactivate(spa); 4894 4895 spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1; 4896 4897 spa_activate(spa, mode); 4898 spa_async_suspend(spa); 4899 4900 spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu", 4901 (u_longlong_t)spa->spa_load_max_txg); 4902 4903 return (spa_load(spa, state, SPA_IMPORT_EXISTING)); 4904 } 4905 4906 /* 4907 * If spa_load() fails this function will try loading prior txg's. If 4908 * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool 4909 * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this 4910 * function will not rewind the pool and will return the same error as 4911 * spa_load(). 4912 */ 4913 static int 4914 spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request, 4915 int rewind_flags) 4916 { 4917 nvlist_t *loadinfo = NULL; 4918 nvlist_t *config = NULL; 4919 int load_error, rewind_error; 4920 uint64_t safe_rewind_txg; 4921 uint64_t min_txg; 4922 4923 if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) { 4924 spa->spa_load_max_txg = spa->spa_load_txg; 4925 spa_set_log_state(spa, SPA_LOG_CLEAR); 4926 } else { 4927 spa->spa_load_max_txg = max_request; 4928 if (max_request != UINT64_MAX) 4929 spa->spa_extreme_rewind = B_TRUE; 4930 } 4931 4932 load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING); 4933 if (load_error == 0) 4934 return (0); 4935 if (load_error == ZFS_ERR_NO_CHECKPOINT) { 4936 /* 4937 * When attempting checkpoint-rewind on a pool with no 4938 * checkpoint, we should not attempt to load uberblocks 4939 * from previous txgs when spa_load fails. 4940 */ 4941 ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT); 4942 spa_import_progress_remove(spa_guid(spa)); 4943 return (load_error); 4944 } 4945 4946 if (spa->spa_root_vdev != NULL) 4947 config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); 4948 4949 spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg; 4950 spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp; 4951 4952 if (rewind_flags & ZPOOL_NEVER_REWIND) { 4953 nvlist_free(config); 4954 spa_import_progress_remove(spa_guid(spa)); 4955 return (load_error); 4956 } 4957 4958 if (state == SPA_LOAD_RECOVER) { 4959 /* Price of rolling back is discarding txgs, including log */ 4960 spa_set_log_state(spa, SPA_LOG_CLEAR); 4961 } else { 4962 /* 4963 * If we aren't rolling back save the load info from our first 4964 * import attempt so that we can restore it after attempting 4965 * to rewind. 4966 */ 4967 loadinfo = spa->spa_load_info; 4968 spa->spa_load_info = fnvlist_alloc(); 4969 } 4970 4971 spa->spa_load_max_txg = spa->spa_last_ubsync_txg; 4972 safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE; 4973 min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ? 4974 TXG_INITIAL : safe_rewind_txg; 4975 4976 /* 4977 * Continue as long as we're finding errors, we're still within 4978 * the acceptable rewind range, and we're still finding uberblocks 4979 */ 4980 while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg && 4981 spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) { 4982 if (spa->spa_load_max_txg < safe_rewind_txg) 4983 spa->spa_extreme_rewind = B_TRUE; 4984 rewind_error = spa_load_retry(spa, state); 4985 } 4986 4987 spa->spa_extreme_rewind = B_FALSE; 4988 spa->spa_load_max_txg = UINT64_MAX; 4989 4990 if (config && (rewind_error || state != SPA_LOAD_RECOVER)) 4991 spa_config_set(spa, config); 4992 else 4993 nvlist_free(config); 4994 4995 if (state == SPA_LOAD_RECOVER) { 4996 ASSERT3P(loadinfo, ==, NULL); 4997 spa_import_progress_remove(spa_guid(spa)); 4998 return (rewind_error); 4999 } else { 5000 /* Store the rewind info as part of the initial load info */ 5001 fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO, 5002 spa->spa_load_info); 5003 5004 /* Restore the initial load info */ 5005 fnvlist_free(spa->spa_load_info); 5006 spa->spa_load_info = loadinfo; 5007 5008 spa_import_progress_remove(spa_guid(spa)); 5009 return (load_error); 5010 } 5011 } 5012 5013 /* 5014 * Pool Open/Import 5015 * 5016 * The import case is identical to an open except that the configuration is sent 5017 * down from userland, instead of grabbed from the configuration cache. For the 5018 * case of an open, the pool configuration will exist in the 5019 * POOL_STATE_UNINITIALIZED state. 5020 * 5021 * The stats information (gen/count/ustats) is used to gather vdev statistics at 5022 * the same time open the pool, without having to keep around the spa_t in some 5023 * ambiguous state. 5024 */ 5025 static int 5026 spa_open_common(const char *pool, spa_t **spapp, void *tag, nvlist_t *nvpolicy, 5027 nvlist_t **config) 5028 { 5029 spa_t *spa; 5030 spa_load_state_t state = SPA_LOAD_OPEN; 5031 int error; 5032 int locked = B_FALSE; 5033 int firstopen = B_FALSE; 5034 5035 *spapp = NULL; 5036 5037 /* 5038 * As disgusting as this is, we need to support recursive calls to this 5039 * function because dsl_dir_open() is called during spa_load(), and ends 5040 * up calling spa_open() again. The real fix is to figure out how to 5041 * avoid dsl_dir_open() calling this in the first place. 5042 */ 5043 if (MUTEX_NOT_HELD(&spa_namespace_lock)) { 5044 mutex_enter(&spa_namespace_lock); 5045 locked = B_TRUE; 5046 } 5047 5048 if ((spa = spa_lookup(pool)) == NULL) { 5049 if (locked) 5050 mutex_exit(&spa_namespace_lock); 5051 return (SET_ERROR(ENOENT)); 5052 } 5053 5054 if (spa->spa_state == POOL_STATE_UNINITIALIZED) { 5055 zpool_load_policy_t policy; 5056 5057 firstopen = B_TRUE; 5058 5059 zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config, 5060 &policy); 5061 if (policy.zlp_rewind & ZPOOL_DO_REWIND) 5062 state = SPA_LOAD_RECOVER; 5063 5064 spa_activate(spa, spa_mode_global); 5065 5066 if (state != SPA_LOAD_RECOVER) 5067 spa->spa_last_ubsync_txg = spa->spa_load_txg = 0; 5068 spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE; 5069 5070 zfs_dbgmsg("spa_open_common: opening %s", pool); 5071 error = spa_load_best(spa, state, policy.zlp_txg, 5072 policy.zlp_rewind); 5073 5074 if (error == EBADF) { 5075 /* 5076 * If vdev_validate() returns failure (indicated by 5077 * EBADF), it indicates that one of the vdevs indicates 5078 * that the pool has been exported or destroyed. If 5079 * this is the case, the config cache is out of sync and 5080 * we should remove the pool from the namespace. 5081 */ 5082 spa_unload(spa); 5083 spa_deactivate(spa); 5084 spa_write_cachefile(spa, B_TRUE, B_TRUE); 5085 spa_remove(spa); 5086 if (locked) 5087 mutex_exit(&spa_namespace_lock); 5088 return (SET_ERROR(ENOENT)); 5089 } 5090 5091 if (error) { 5092 /* 5093 * We can't open the pool, but we still have useful 5094 * information: the state of each vdev after the 5095 * attempted vdev_open(). Return this to the user. 5096 */ 5097 if (config != NULL && spa->spa_config) { 5098 VERIFY(nvlist_dup(spa->spa_config, config, 5099 KM_SLEEP) == 0); 5100 VERIFY(nvlist_add_nvlist(*config, 5101 ZPOOL_CONFIG_LOAD_INFO, 5102 spa->spa_load_info) == 0); 5103 } 5104 spa_unload(spa); 5105 spa_deactivate(spa); 5106 spa->spa_last_open_failed = error; 5107 if (locked) 5108 mutex_exit(&spa_namespace_lock); 5109 *spapp = NULL; 5110 return (error); 5111 } 5112 } 5113 5114 spa_open_ref(spa, tag); 5115 5116 if (config != NULL) 5117 *config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); 5118 5119 /* 5120 * If we've recovered the pool, pass back any information we 5121 * gathered while doing the load. 5122 */ 5123 if (state == SPA_LOAD_RECOVER) { 5124 VERIFY(nvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO, 5125 spa->spa_load_info) == 0); 5126 } 5127 5128 if (locked) { 5129 spa->spa_last_open_failed = 0; 5130 spa->spa_last_ubsync_txg = 0; 5131 spa->spa_load_txg = 0; 5132 mutex_exit(&spa_namespace_lock); 5133 } 5134 5135 if (firstopen) 5136 zvol_create_minors_recursive(spa_name(spa)); 5137 5138 *spapp = spa; 5139 5140 return (0); 5141 } 5142 5143 int 5144 spa_open_rewind(const char *name, spa_t **spapp, void *tag, nvlist_t *policy, 5145 nvlist_t **config) 5146 { 5147 return (spa_open_common(name, spapp, tag, policy, config)); 5148 } 5149 5150 int 5151 spa_open(const char *name, spa_t **spapp, void *tag) 5152 { 5153 return (spa_open_common(name, spapp, tag, NULL, NULL)); 5154 } 5155 5156 /* 5157 * Lookup the given spa_t, incrementing the inject count in the process, 5158 * preventing it from being exported or destroyed. 5159 */ 5160 spa_t * 5161 spa_inject_addref(char *name) 5162 { 5163 spa_t *spa; 5164 5165 mutex_enter(&spa_namespace_lock); 5166 if ((spa = spa_lookup(name)) == NULL) { 5167 mutex_exit(&spa_namespace_lock); 5168 return (NULL); 5169 } 5170 spa->spa_inject_ref++; 5171 mutex_exit(&spa_namespace_lock); 5172 5173 return (spa); 5174 } 5175 5176 void 5177 spa_inject_delref(spa_t *spa) 5178 { 5179 mutex_enter(&spa_namespace_lock); 5180 spa->spa_inject_ref--; 5181 mutex_exit(&spa_namespace_lock); 5182 } 5183 5184 /* 5185 * Add spares device information to the nvlist. 5186 */ 5187 static void 5188 spa_add_spares(spa_t *spa, nvlist_t *config) 5189 { 5190 nvlist_t **spares; 5191 uint_t i, nspares; 5192 nvlist_t *nvroot; 5193 uint64_t guid; 5194 vdev_stat_t *vs; 5195 uint_t vsc; 5196 uint64_t pool; 5197 5198 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 5199 5200 if (spa->spa_spares.sav_count == 0) 5201 return; 5202 5203 VERIFY(nvlist_lookup_nvlist(config, 5204 ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0); 5205 VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config, 5206 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0); 5207 if (nspares != 0) { 5208 VERIFY(nvlist_add_nvlist_array(nvroot, 5209 ZPOOL_CONFIG_SPARES, spares, nspares) == 0); 5210 VERIFY(nvlist_lookup_nvlist_array(nvroot, 5211 ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0); 5212 5213 /* 5214 * Go through and find any spares which have since been 5215 * repurposed as an active spare. If this is the case, update 5216 * their status appropriately. 5217 */ 5218 for (i = 0; i < nspares; i++) { 5219 VERIFY(nvlist_lookup_uint64(spares[i], 5220 ZPOOL_CONFIG_GUID, &guid) == 0); 5221 if (spa_spare_exists(guid, &pool, NULL) && 5222 pool != 0ULL) { 5223 VERIFY(nvlist_lookup_uint64_array( 5224 spares[i], ZPOOL_CONFIG_VDEV_STATS, 5225 (uint64_t **)&vs, &vsc) == 0); 5226 vs->vs_state = VDEV_STATE_CANT_OPEN; 5227 vs->vs_aux = VDEV_AUX_SPARED; 5228 } 5229 } 5230 } 5231 } 5232 5233 /* 5234 * Add l2cache device information to the nvlist, including vdev stats. 5235 */ 5236 static void 5237 spa_add_l2cache(spa_t *spa, nvlist_t *config) 5238 { 5239 nvlist_t **l2cache; 5240 uint_t i, j, nl2cache; 5241 nvlist_t *nvroot; 5242 uint64_t guid; 5243 vdev_t *vd; 5244 vdev_stat_t *vs; 5245 uint_t vsc; 5246 5247 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 5248 5249 if (spa->spa_l2cache.sav_count == 0) 5250 return; 5251 5252 VERIFY(nvlist_lookup_nvlist(config, 5253 ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0); 5254 VERIFY(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config, 5255 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0); 5256 if (nl2cache != 0) { 5257 VERIFY(nvlist_add_nvlist_array(nvroot, 5258 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0); 5259 VERIFY(nvlist_lookup_nvlist_array(nvroot, 5260 ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0); 5261 5262 /* 5263 * Update level 2 cache device stats. 5264 */ 5265 5266 for (i = 0; i < nl2cache; i++) { 5267 VERIFY(nvlist_lookup_uint64(l2cache[i], 5268 ZPOOL_CONFIG_GUID, &guid) == 0); 5269 5270 vd = NULL; 5271 for (j = 0; j < spa->spa_l2cache.sav_count; j++) { 5272 if (guid == 5273 spa->spa_l2cache.sav_vdevs[j]->vdev_guid) { 5274 vd = spa->spa_l2cache.sav_vdevs[j]; 5275 break; 5276 } 5277 } 5278 ASSERT(vd != NULL); 5279 5280 VERIFY(nvlist_lookup_uint64_array(l2cache[i], 5281 ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc) 5282 == 0); 5283 vdev_get_stats(vd, vs); 5284 vdev_config_generate_stats(vd, l2cache[i]); 5285 5286 } 5287 } 5288 } 5289 5290 static void 5291 spa_feature_stats_from_disk(spa_t *spa, nvlist_t *features) 5292 { 5293 zap_cursor_t zc; 5294 zap_attribute_t za; 5295 5296 if (spa->spa_feat_for_read_obj != 0) { 5297 for (zap_cursor_init(&zc, spa->spa_meta_objset, 5298 spa->spa_feat_for_read_obj); 5299 zap_cursor_retrieve(&zc, &za) == 0; 5300 zap_cursor_advance(&zc)) { 5301 ASSERT(za.za_integer_length == sizeof (uint64_t) && 5302 za.za_num_integers == 1); 5303 VERIFY0(nvlist_add_uint64(features, za.za_name, 5304 za.za_first_integer)); 5305 } 5306 zap_cursor_fini(&zc); 5307 } 5308 5309 if (spa->spa_feat_for_write_obj != 0) { 5310 for (zap_cursor_init(&zc, spa->spa_meta_objset, 5311 spa->spa_feat_for_write_obj); 5312 zap_cursor_retrieve(&zc, &za) == 0; 5313 zap_cursor_advance(&zc)) { 5314 ASSERT(za.za_integer_length == sizeof (uint64_t) && 5315 za.za_num_integers == 1); 5316 VERIFY0(nvlist_add_uint64(features, za.za_name, 5317 za.za_first_integer)); 5318 } 5319 zap_cursor_fini(&zc); 5320 } 5321 } 5322 5323 static void 5324 spa_feature_stats_from_cache(spa_t *spa, nvlist_t *features) 5325 { 5326 int i; 5327 5328 for (i = 0; i < SPA_FEATURES; i++) { 5329 zfeature_info_t feature = spa_feature_table[i]; 5330 uint64_t refcount; 5331 5332 if (feature_get_refcount(spa, &feature, &refcount) != 0) 5333 continue; 5334 5335 VERIFY0(nvlist_add_uint64(features, feature.fi_guid, refcount)); 5336 } 5337 } 5338 5339 /* 5340 * Store a list of pool features and their reference counts in the 5341 * config. 5342 * 5343 * The first time this is called on a spa, allocate a new nvlist, fetch 5344 * the pool features and reference counts from disk, then save the list 5345 * in the spa. In subsequent calls on the same spa use the saved nvlist 5346 * and refresh its values from the cached reference counts. This 5347 * ensures we don't block here on I/O on a suspended pool so 'zpool 5348 * clear' can resume the pool. 5349 */ 5350 static void 5351 spa_add_feature_stats(spa_t *spa, nvlist_t *config) 5352 { 5353 nvlist_t *features; 5354 5355 ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER)); 5356 5357 mutex_enter(&spa->spa_feat_stats_lock); 5358 features = spa->spa_feat_stats; 5359 5360 if (features != NULL) { 5361 spa_feature_stats_from_cache(spa, features); 5362 } else { 5363 VERIFY0(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP)); 5364 spa->spa_feat_stats = features; 5365 spa_feature_stats_from_disk(spa, features); 5366 } 5367 5368 VERIFY0(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS, 5369 features)); 5370 5371 mutex_exit(&spa->spa_feat_stats_lock); 5372 } 5373 5374 int 5375 spa_get_stats(const char *name, nvlist_t **config, 5376 char *altroot, size_t buflen) 5377 { 5378 int error; 5379 spa_t *spa; 5380 5381 *config = NULL; 5382 error = spa_open_common(name, &spa, FTAG, NULL, config); 5383 5384 if (spa != NULL) { 5385 /* 5386 * This still leaves a window of inconsistency where the spares 5387 * or l2cache devices could change and the config would be 5388 * self-inconsistent. 5389 */ 5390 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 5391 5392 if (*config != NULL) { 5393 uint64_t loadtimes[2]; 5394 5395 loadtimes[0] = spa->spa_loaded_ts.tv_sec; 5396 loadtimes[1] = spa->spa_loaded_ts.tv_nsec; 5397 VERIFY(nvlist_add_uint64_array(*config, 5398 ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2) == 0); 5399 5400 VERIFY(nvlist_add_uint64(*config, 5401 ZPOOL_CONFIG_ERRCOUNT, 5402 spa_get_errlog_size(spa)) == 0); 5403 5404 if (spa_suspended(spa)) { 5405 VERIFY(nvlist_add_uint64(*config, 5406 ZPOOL_CONFIG_SUSPENDED, 5407 spa->spa_failmode) == 0); 5408 VERIFY(nvlist_add_uint64(*config, 5409 ZPOOL_CONFIG_SUSPENDED_REASON, 5410 spa->spa_suspended) == 0); 5411 } 5412 5413 spa_add_spares(spa, *config); 5414 spa_add_l2cache(spa, *config); 5415 spa_add_feature_stats(spa, *config); 5416 } 5417 } 5418 5419 /* 5420 * We want to get the alternate root even for faulted pools, so we cheat 5421 * and call spa_lookup() directly. 5422 */ 5423 if (altroot) { 5424 if (spa == NULL) { 5425 mutex_enter(&spa_namespace_lock); 5426 spa = spa_lookup(name); 5427 if (spa) 5428 spa_altroot(spa, altroot, buflen); 5429 else 5430 altroot[0] = '\0'; 5431 spa = NULL; 5432 mutex_exit(&spa_namespace_lock); 5433 } else { 5434 spa_altroot(spa, altroot, buflen); 5435 } 5436 } 5437 5438 if (spa != NULL) { 5439 spa_config_exit(spa, SCL_CONFIG, FTAG); 5440 spa_close(spa, FTAG); 5441 } 5442 5443 return (error); 5444 } 5445 5446 /* 5447 * Validate that the auxiliary device array is well formed. We must have an 5448 * array of nvlists, each which describes a valid leaf vdev. If this is an 5449 * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be 5450 * specified, as long as they are well-formed. 5451 */ 5452 static int 5453 spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode, 5454 spa_aux_vdev_t *sav, const char *config, uint64_t version, 5455 vdev_labeltype_t label) 5456 { 5457 nvlist_t **dev; 5458 uint_t i, ndev; 5459 vdev_t *vd; 5460 int error; 5461 5462 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 5463 5464 /* 5465 * It's acceptable to have no devs specified. 5466 */ 5467 if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0) 5468 return (0); 5469 5470 if (ndev == 0) 5471 return (SET_ERROR(EINVAL)); 5472 5473 /* 5474 * Make sure the pool is formatted with a version that supports this 5475 * device type. 5476 */ 5477 if (spa_version(spa) < version) 5478 return (SET_ERROR(ENOTSUP)); 5479 5480 /* 5481 * Set the pending device list so we correctly handle device in-use 5482 * checking. 5483 */ 5484 sav->sav_pending = dev; 5485 sav->sav_npending = ndev; 5486 5487 for (i = 0; i < ndev; i++) { 5488 if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0, 5489 mode)) != 0) 5490 goto out; 5491 5492 if (!vd->vdev_ops->vdev_op_leaf) { 5493 vdev_free(vd); 5494 error = SET_ERROR(EINVAL); 5495 goto out; 5496 } 5497 5498 vd->vdev_top = vd; 5499 5500 if ((error = vdev_open(vd)) == 0 && 5501 (error = vdev_label_init(vd, crtxg, label)) == 0) { 5502 VERIFY(nvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID, 5503 vd->vdev_guid) == 0); 5504 } 5505 5506 vdev_free(vd); 5507 5508 if (error && 5509 (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE)) 5510 goto out; 5511 else 5512 error = 0; 5513 } 5514 5515 out: 5516 sav->sav_pending = NULL; 5517 sav->sav_npending = 0; 5518 return (error); 5519 } 5520 5521 static int 5522 spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode) 5523 { 5524 int error; 5525 5526 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL); 5527 5528 if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode, 5529 &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES, 5530 VDEV_LABEL_SPARE)) != 0) { 5531 return (error); 5532 } 5533 5534 return (spa_validate_aux_devs(spa, nvroot, crtxg, mode, 5535 &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE, 5536 VDEV_LABEL_L2CACHE)); 5537 } 5538 5539 static void 5540 spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs, 5541 const char *config) 5542 { 5543 int i; 5544 5545 if (sav->sav_config != NULL) { 5546 nvlist_t **olddevs; 5547 uint_t oldndevs; 5548 nvlist_t **newdevs; 5549 5550 /* 5551 * Generate new dev list by concatenating with the 5552 * current dev list. 5553 */ 5554 VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, config, 5555 &olddevs, &oldndevs) == 0); 5556 5557 newdevs = kmem_alloc(sizeof (void *) * 5558 (ndevs + oldndevs), KM_SLEEP); 5559 for (i = 0; i < oldndevs; i++) 5560 VERIFY(nvlist_dup(olddevs[i], &newdevs[i], 5561 KM_SLEEP) == 0); 5562 for (i = 0; i < ndevs; i++) 5563 VERIFY(nvlist_dup(devs[i], &newdevs[i + oldndevs], 5564 KM_SLEEP) == 0); 5565 5566 VERIFY(nvlist_remove(sav->sav_config, config, 5567 DATA_TYPE_NVLIST_ARRAY) == 0); 5568 5569 VERIFY(nvlist_add_nvlist_array(sav->sav_config, 5570 config, newdevs, ndevs + oldndevs) == 0); 5571 for (i = 0; i < oldndevs + ndevs; i++) 5572 nvlist_free(newdevs[i]); 5573 kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *)); 5574 } else { 5575 /* 5576 * Generate a new dev list. 5577 */ 5578 VERIFY(nvlist_alloc(&sav->sav_config, NV_UNIQUE_NAME, 5579 KM_SLEEP) == 0); 5580 VERIFY(nvlist_add_nvlist_array(sav->sav_config, config, 5581 devs, ndevs) == 0); 5582 } 5583 } 5584 5585 /* 5586 * Stop and drop level 2 ARC devices 5587 */ 5588 void 5589 spa_l2cache_drop(spa_t *spa) 5590 { 5591 vdev_t *vd; 5592 int i; 5593 spa_aux_vdev_t *sav = &spa->spa_l2cache; 5594 5595 for (i = 0; i < sav->sav_count; i++) { 5596 uint64_t pool; 5597 5598 vd = sav->sav_vdevs[i]; 5599 ASSERT(vd != NULL); 5600 5601 if (spa_l2cache_exists(vd->vdev_guid, &pool) && 5602 pool != 0ULL && l2arc_vdev_present(vd)) 5603 l2arc_remove_vdev(vd); 5604 } 5605 } 5606 5607 /* 5608 * Verify encryption parameters for spa creation. If we are encrypting, we must 5609 * have the encryption feature flag enabled. 5610 */ 5611 static int 5612 spa_create_check_encryption_params(dsl_crypto_params_t *dcp, 5613 boolean_t has_encryption) 5614 { 5615 if (dcp->cp_crypt != ZIO_CRYPT_OFF && 5616 dcp->cp_crypt != ZIO_CRYPT_INHERIT && 5617 !has_encryption) 5618 return (SET_ERROR(ENOTSUP)); 5619 5620 return (dmu_objset_create_crypt_check(NULL, dcp, NULL)); 5621 } 5622 5623 /* 5624 * Pool Creation 5625 */ 5626 int 5627 spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props, 5628 nvlist_t *zplprops, dsl_crypto_params_t *dcp) 5629 { 5630 spa_t *spa; 5631 char *altroot = NULL; 5632 vdev_t *rvd; 5633 dsl_pool_t *dp; 5634 dmu_tx_t *tx; 5635 int error = 0; 5636 uint64_t txg = TXG_INITIAL; 5637 nvlist_t **spares, **l2cache; 5638 uint_t nspares, nl2cache; 5639 uint64_t version, obj, ndraid = 0; 5640 boolean_t has_features; 5641 boolean_t has_encryption; 5642 boolean_t has_allocclass; 5643 spa_feature_t feat; 5644 char *feat_name; 5645 char *poolname; 5646 nvlist_t *nvl; 5647 5648 if (props == NULL || 5649 nvlist_lookup_string(props, "tname", &poolname) != 0) 5650 poolname = (char *)pool; 5651 5652 /* 5653 * If this pool already exists, return failure. 5654 */ 5655 mutex_enter(&spa_namespace_lock); 5656 if (spa_lookup(poolname) != NULL) { 5657 mutex_exit(&spa_namespace_lock); 5658 return (SET_ERROR(EEXIST)); 5659 } 5660 5661 /* 5662 * Allocate a new spa_t structure. 5663 */ 5664 nvl = fnvlist_alloc(); 5665 fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool); 5666 (void) nvlist_lookup_string(props, 5667 zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); 5668 spa = spa_add(poolname, nvl, altroot); 5669 fnvlist_free(nvl); 5670 spa_activate(spa, spa_mode_global); 5671 5672 if (props && (error = spa_prop_validate(spa, props))) { 5673 spa_deactivate(spa); 5674 spa_remove(spa); 5675 mutex_exit(&spa_namespace_lock); 5676 return (error); 5677 } 5678 5679 /* 5680 * Temporary pool names should never be written to disk. 5681 */ 5682 if (poolname != pool) 5683 spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME; 5684 5685 has_features = B_FALSE; 5686 has_encryption = B_FALSE; 5687 has_allocclass = B_FALSE; 5688 for (nvpair_t *elem = nvlist_next_nvpair(props, NULL); 5689 elem != NULL; elem = nvlist_next_nvpair(props, elem)) { 5690 if (zpool_prop_feature(nvpair_name(elem))) { 5691 has_features = B_TRUE; 5692 5693 feat_name = strchr(nvpair_name(elem), '@') + 1; 5694 VERIFY0(zfeature_lookup_name(feat_name, &feat)); 5695 if (feat == SPA_FEATURE_ENCRYPTION) 5696 has_encryption = B_TRUE; 5697 if (feat == SPA_FEATURE_ALLOCATION_CLASSES) 5698 has_allocclass = B_TRUE; 5699 } 5700 } 5701 5702 /* verify encryption params, if they were provided */ 5703 if (dcp != NULL) { 5704 error = spa_create_check_encryption_params(dcp, has_encryption); 5705 if (error != 0) { 5706 spa_deactivate(spa); 5707 spa_remove(spa); 5708 mutex_exit(&spa_namespace_lock); 5709 return (error); 5710 } 5711 } 5712 if (!has_allocclass && zfs_special_devs(nvroot, NULL)) { 5713 spa_deactivate(spa); 5714 spa_remove(spa); 5715 mutex_exit(&spa_namespace_lock); 5716 return (ENOTSUP); 5717 } 5718 5719 if (has_features || nvlist_lookup_uint64(props, 5720 zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) { 5721 version = SPA_VERSION; 5722 } 5723 ASSERT(SPA_VERSION_IS_SUPPORTED(version)); 5724 5725 spa->spa_first_txg = txg; 5726 spa->spa_uberblock.ub_txg = txg - 1; 5727 spa->spa_uberblock.ub_version = version; 5728 spa->spa_ubsync = spa->spa_uberblock; 5729 spa->spa_load_state = SPA_LOAD_CREATE; 5730 spa->spa_removing_phys.sr_state = DSS_NONE; 5731 spa->spa_removing_phys.sr_removing_vdev = -1; 5732 spa->spa_removing_phys.sr_prev_indirect_vdev = -1; 5733 spa->spa_indirect_vdevs_loaded = B_TRUE; 5734 5735 /* 5736 * Create "The Godfather" zio to hold all async IOs 5737 */ 5738 spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *), 5739 KM_SLEEP); 5740 for (int i = 0; i < max_ncpus; i++) { 5741 spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL, 5742 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | 5743 ZIO_FLAG_GODFATHER); 5744 } 5745 5746 /* 5747 * Create the root vdev. 5748 */ 5749 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 5750 5751 error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD); 5752 5753 ASSERT(error != 0 || rvd != NULL); 5754 ASSERT(error != 0 || spa->spa_root_vdev == rvd); 5755 5756 if (error == 0 && !zfs_allocatable_devs(nvroot)) 5757 error = SET_ERROR(EINVAL); 5758 5759 if (error == 0 && 5760 (error = vdev_create(rvd, txg, B_FALSE)) == 0 && 5761 (error = vdev_draid_spare_create(nvroot, rvd, &ndraid, 0)) == 0 && 5762 (error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) == 0) { 5763 /* 5764 * instantiate the metaslab groups (this will dirty the vdevs) 5765 * we can no longer error exit past this point 5766 */ 5767 for (int c = 0; error == 0 && c < rvd->vdev_children; c++) { 5768 vdev_t *vd = rvd->vdev_child[c]; 5769 5770 vdev_metaslab_set_size(vd); 5771 vdev_expand(vd, txg); 5772 } 5773 } 5774 5775 spa_config_exit(spa, SCL_ALL, FTAG); 5776 5777 if (error != 0) { 5778 spa_unload(spa); 5779 spa_deactivate(spa); 5780 spa_remove(spa); 5781 mutex_exit(&spa_namespace_lock); 5782 return (error); 5783 } 5784 5785 /* 5786 * Get the list of spares, if specified. 5787 */ 5788 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 5789 &spares, &nspares) == 0) { 5790 VERIFY(nvlist_alloc(&spa->spa_spares.sav_config, NV_UNIQUE_NAME, 5791 KM_SLEEP) == 0); 5792 VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config, 5793 ZPOOL_CONFIG_SPARES, spares, nspares) == 0); 5794 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 5795 spa_load_spares(spa); 5796 spa_config_exit(spa, SCL_ALL, FTAG); 5797 spa->spa_spares.sav_sync = B_TRUE; 5798 } 5799 5800 /* 5801 * Get the list of level 2 cache devices, if specified. 5802 */ 5803 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, 5804 &l2cache, &nl2cache) == 0) { 5805 VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config, 5806 NV_UNIQUE_NAME, KM_SLEEP) == 0); 5807 VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config, 5808 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0); 5809 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 5810 spa_load_l2cache(spa); 5811 spa_config_exit(spa, SCL_ALL, FTAG); 5812 spa->spa_l2cache.sav_sync = B_TRUE; 5813 } 5814 5815 spa->spa_is_initializing = B_TRUE; 5816 spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg); 5817 spa->spa_is_initializing = B_FALSE; 5818 5819 /* 5820 * Create DDTs (dedup tables). 5821 */ 5822 ddt_create(spa); 5823 5824 spa_update_dspace(spa); 5825 5826 tx = dmu_tx_create_assigned(dp, txg); 5827 5828 /* 5829 * Create the pool's history object. 5830 */ 5831 if (version >= SPA_VERSION_ZPOOL_HISTORY && !spa->spa_history) 5832 spa_history_create_obj(spa, tx); 5833 5834 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE); 5835 spa_history_log_version(spa, "create", tx); 5836 5837 /* 5838 * Create the pool config object. 5839 */ 5840 spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset, 5841 DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE, 5842 DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx); 5843 5844 if (zap_add(spa->spa_meta_objset, 5845 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG, 5846 sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) { 5847 cmn_err(CE_PANIC, "failed to add pool config"); 5848 } 5849 5850 if (zap_add(spa->spa_meta_objset, 5851 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION, 5852 sizeof (uint64_t), 1, &version, tx) != 0) { 5853 cmn_err(CE_PANIC, "failed to add pool version"); 5854 } 5855 5856 /* Newly created pools with the right version are always deflated. */ 5857 if (version >= SPA_VERSION_RAIDZ_DEFLATE) { 5858 spa->spa_deflate = TRUE; 5859 if (zap_add(spa->spa_meta_objset, 5860 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE, 5861 sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) { 5862 cmn_err(CE_PANIC, "failed to add deflate"); 5863 } 5864 } 5865 5866 /* 5867 * Create the deferred-free bpobj. Turn off compression 5868 * because sync-to-convergence takes longer if the blocksize 5869 * keeps changing. 5870 */ 5871 obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx); 5872 dmu_object_set_compress(spa->spa_meta_objset, obj, 5873 ZIO_COMPRESS_OFF, tx); 5874 if (zap_add(spa->spa_meta_objset, 5875 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ, 5876 sizeof (uint64_t), 1, &obj, tx) != 0) { 5877 cmn_err(CE_PANIC, "failed to add bpobj"); 5878 } 5879 VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj, 5880 spa->spa_meta_objset, obj)); 5881 5882 /* 5883 * Generate some random noise for salted checksums to operate on. 5884 */ 5885 (void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes, 5886 sizeof (spa->spa_cksum_salt.zcs_bytes)); 5887 5888 /* 5889 * Set pool properties. 5890 */ 5891 spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS); 5892 spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION); 5893 spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE); 5894 spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND); 5895 spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST); 5896 spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM); 5897 5898 if (props != NULL) { 5899 spa_configfile_set(spa, props, B_FALSE); 5900 spa_sync_props(props, tx); 5901 } 5902 5903 for (int i = 0; i < ndraid; i++) 5904 spa_feature_incr(spa, SPA_FEATURE_DRAID, tx); 5905 5906 dmu_tx_commit(tx); 5907 5908 spa->spa_sync_on = B_TRUE; 5909 txg_sync_start(dp); 5910 mmp_thread_start(spa); 5911 txg_wait_synced(dp, txg); 5912 5913 spa_spawn_aux_threads(spa); 5914 5915 spa_write_cachefile(spa, B_FALSE, B_TRUE); 5916 5917 /* 5918 * Don't count references from objsets that are already closed 5919 * and are making their way through the eviction process. 5920 */ 5921 spa_evicting_os_wait(spa); 5922 spa->spa_minref = zfs_refcount_count(&spa->spa_refcount); 5923 spa->spa_load_state = SPA_LOAD_NONE; 5924 5925 mutex_exit(&spa_namespace_lock); 5926 5927 return (0); 5928 } 5929 5930 /* 5931 * Import a non-root pool into the system. 5932 */ 5933 int 5934 spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags) 5935 { 5936 spa_t *spa; 5937 char *altroot = NULL; 5938 spa_load_state_t state = SPA_LOAD_IMPORT; 5939 zpool_load_policy_t policy; 5940 spa_mode_t mode = spa_mode_global; 5941 uint64_t readonly = B_FALSE; 5942 int error; 5943 nvlist_t *nvroot; 5944 nvlist_t **spares, **l2cache; 5945 uint_t nspares, nl2cache; 5946 5947 /* 5948 * If a pool with this name exists, return failure. 5949 */ 5950 mutex_enter(&spa_namespace_lock); 5951 if (spa_lookup(pool) != NULL) { 5952 mutex_exit(&spa_namespace_lock); 5953 return (SET_ERROR(EEXIST)); 5954 } 5955 5956 /* 5957 * Create and initialize the spa structure. 5958 */ 5959 (void) nvlist_lookup_string(props, 5960 zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); 5961 (void) nvlist_lookup_uint64(props, 5962 zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly); 5963 if (readonly) 5964 mode = SPA_MODE_READ; 5965 spa = spa_add(pool, config, altroot); 5966 spa->spa_import_flags = flags; 5967 5968 /* 5969 * Verbatim import - Take a pool and insert it into the namespace 5970 * as if it had been loaded at boot. 5971 */ 5972 if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) { 5973 if (props != NULL) 5974 spa_configfile_set(spa, props, B_FALSE); 5975 5976 spa_write_cachefile(spa, B_FALSE, B_TRUE); 5977 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT); 5978 zfs_dbgmsg("spa_import: verbatim import of %s", pool); 5979 mutex_exit(&spa_namespace_lock); 5980 return (0); 5981 } 5982 5983 spa_activate(spa, mode); 5984 5985 /* 5986 * Don't start async tasks until we know everything is healthy. 5987 */ 5988 spa_async_suspend(spa); 5989 5990 zpool_get_load_policy(config, &policy); 5991 if (policy.zlp_rewind & ZPOOL_DO_REWIND) 5992 state = SPA_LOAD_RECOVER; 5993 5994 spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT; 5995 5996 if (state != SPA_LOAD_RECOVER) { 5997 spa->spa_last_ubsync_txg = spa->spa_load_txg = 0; 5998 zfs_dbgmsg("spa_import: importing %s", pool); 5999 } else { 6000 zfs_dbgmsg("spa_import: importing %s, max_txg=%lld " 6001 "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg); 6002 } 6003 error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind); 6004 6005 /* 6006 * Propagate anything learned while loading the pool and pass it 6007 * back to caller (i.e. rewind info, missing devices, etc). 6008 */ 6009 VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, 6010 spa->spa_load_info) == 0); 6011 6012 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6013 /* 6014 * Toss any existing sparelist, as it doesn't have any validity 6015 * anymore, and conflicts with spa_has_spare(). 6016 */ 6017 if (spa->spa_spares.sav_config) { 6018 nvlist_free(spa->spa_spares.sav_config); 6019 spa->spa_spares.sav_config = NULL; 6020 spa_load_spares(spa); 6021 } 6022 if (spa->spa_l2cache.sav_config) { 6023 nvlist_free(spa->spa_l2cache.sav_config); 6024 spa->spa_l2cache.sav_config = NULL; 6025 spa_load_l2cache(spa); 6026 } 6027 6028 VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, 6029 &nvroot) == 0); 6030 spa_config_exit(spa, SCL_ALL, FTAG); 6031 6032 if (props != NULL) 6033 spa_configfile_set(spa, props, B_FALSE); 6034 6035 if (error != 0 || (props && spa_writeable(spa) && 6036 (error = spa_prop_set(spa, props)))) { 6037 spa_unload(spa); 6038 spa_deactivate(spa); 6039 spa_remove(spa); 6040 mutex_exit(&spa_namespace_lock); 6041 return (error); 6042 } 6043 6044 spa_async_resume(spa); 6045 6046 /* 6047 * Override any spares and level 2 cache devices as specified by 6048 * the user, as these may have correct device names/devids, etc. 6049 */ 6050 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, 6051 &spares, &nspares) == 0) { 6052 if (spa->spa_spares.sav_config) 6053 VERIFY(nvlist_remove(spa->spa_spares.sav_config, 6054 ZPOOL_CONFIG_SPARES, DATA_TYPE_NVLIST_ARRAY) == 0); 6055 else 6056 VERIFY(nvlist_alloc(&spa->spa_spares.sav_config, 6057 NV_UNIQUE_NAME, KM_SLEEP) == 0); 6058 VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config, 6059 ZPOOL_CONFIG_SPARES, spares, nspares) == 0); 6060 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6061 spa_load_spares(spa); 6062 spa_config_exit(spa, SCL_ALL, FTAG); 6063 spa->spa_spares.sav_sync = B_TRUE; 6064 } 6065 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, 6066 &l2cache, &nl2cache) == 0) { 6067 if (spa->spa_l2cache.sav_config) 6068 VERIFY(nvlist_remove(spa->spa_l2cache.sav_config, 6069 ZPOOL_CONFIG_L2CACHE, DATA_TYPE_NVLIST_ARRAY) == 0); 6070 else 6071 VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config, 6072 NV_UNIQUE_NAME, KM_SLEEP) == 0); 6073 VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config, 6074 ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0); 6075 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6076 spa_load_l2cache(spa); 6077 spa_config_exit(spa, SCL_ALL, FTAG); 6078 spa->spa_l2cache.sav_sync = B_TRUE; 6079 } 6080 6081 /* 6082 * Check for any removed devices. 6083 */ 6084 if (spa->spa_autoreplace) { 6085 spa_aux_check_removed(&spa->spa_spares); 6086 spa_aux_check_removed(&spa->spa_l2cache); 6087 } 6088 6089 if (spa_writeable(spa)) { 6090 /* 6091 * Update the config cache to include the newly-imported pool. 6092 */ 6093 spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); 6094 } 6095 6096 /* 6097 * It's possible that the pool was expanded while it was exported. 6098 * We kick off an async task to handle this for us. 6099 */ 6100 spa_async_request(spa, SPA_ASYNC_AUTOEXPAND); 6101 6102 spa_history_log_version(spa, "import", NULL); 6103 6104 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT); 6105 6106 mutex_exit(&spa_namespace_lock); 6107 6108 zvol_create_minors_recursive(pool); 6109 6110 return (0); 6111 } 6112 6113 nvlist_t * 6114 spa_tryimport(nvlist_t *tryconfig) 6115 { 6116 nvlist_t *config = NULL; 6117 char *poolname, *cachefile; 6118 spa_t *spa; 6119 uint64_t state; 6120 int error; 6121 zpool_load_policy_t policy; 6122 6123 if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname)) 6124 return (NULL); 6125 6126 if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state)) 6127 return (NULL); 6128 6129 /* 6130 * Create and initialize the spa structure. 6131 */ 6132 mutex_enter(&spa_namespace_lock); 6133 spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL); 6134 spa_activate(spa, SPA_MODE_READ); 6135 6136 /* 6137 * Rewind pool if a max txg was provided. 6138 */ 6139 zpool_get_load_policy(spa->spa_config, &policy); 6140 if (policy.zlp_txg != UINT64_MAX) { 6141 spa->spa_load_max_txg = policy.zlp_txg; 6142 spa->spa_extreme_rewind = B_TRUE; 6143 zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld", 6144 poolname, (longlong_t)policy.zlp_txg); 6145 } else { 6146 zfs_dbgmsg("spa_tryimport: importing %s", poolname); 6147 } 6148 6149 if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile) 6150 == 0) { 6151 zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile); 6152 spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE; 6153 } else { 6154 spa->spa_config_source = SPA_CONFIG_SRC_SCAN; 6155 } 6156 6157 error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING); 6158 6159 /* 6160 * If 'tryconfig' was at least parsable, return the current config. 6161 */ 6162 if (spa->spa_root_vdev != NULL) { 6163 config = spa_config_generate(spa, NULL, -1ULL, B_TRUE); 6164 VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, 6165 poolname) == 0); 6166 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, 6167 state) == 0); 6168 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP, 6169 spa->spa_uberblock.ub_timestamp) == 0); 6170 VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, 6171 spa->spa_load_info) == 0); 6172 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_ERRATA, 6173 spa->spa_errata) == 0); 6174 6175 /* 6176 * If the bootfs property exists on this pool then we 6177 * copy it out so that external consumers can tell which 6178 * pools are bootable. 6179 */ 6180 if ((!error || error == EEXIST) && spa->spa_bootfs) { 6181 char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP); 6182 6183 /* 6184 * We have to play games with the name since the 6185 * pool was opened as TRYIMPORT_NAME. 6186 */ 6187 if (dsl_dsobj_to_dsname(spa_name(spa), 6188 spa->spa_bootfs, tmpname) == 0) { 6189 char *cp; 6190 char *dsname; 6191 6192 dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP); 6193 6194 cp = strchr(tmpname, '/'); 6195 if (cp == NULL) { 6196 (void) strlcpy(dsname, tmpname, 6197 MAXPATHLEN); 6198 } else { 6199 (void) snprintf(dsname, MAXPATHLEN, 6200 "%s/%s", poolname, ++cp); 6201 } 6202 VERIFY(nvlist_add_string(config, 6203 ZPOOL_CONFIG_BOOTFS, dsname) == 0); 6204 kmem_free(dsname, MAXPATHLEN); 6205 } 6206 kmem_free(tmpname, MAXPATHLEN); 6207 } 6208 6209 /* 6210 * Add the list of hot spares and level 2 cache devices. 6211 */ 6212 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 6213 spa_add_spares(spa, config); 6214 spa_add_l2cache(spa, config); 6215 spa_config_exit(spa, SCL_CONFIG, FTAG); 6216 } 6217 6218 spa_unload(spa); 6219 spa_deactivate(spa); 6220 spa_remove(spa); 6221 mutex_exit(&spa_namespace_lock); 6222 6223 return (config); 6224 } 6225 6226 /* 6227 * Pool export/destroy 6228 * 6229 * The act of destroying or exporting a pool is very simple. We make sure there 6230 * is no more pending I/O and any references to the pool are gone. Then, we 6231 * update the pool state and sync all the labels to disk, removing the 6232 * configuration from the cache afterwards. If the 'hardforce' flag is set, then 6233 * we don't sync the labels or remove the configuration cache. 6234 */ 6235 static int 6236 spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig, 6237 boolean_t force, boolean_t hardforce) 6238 { 6239 spa_t *spa; 6240 6241 if (oldconfig) 6242 *oldconfig = NULL; 6243 6244 if (!(spa_mode_global & SPA_MODE_WRITE)) 6245 return (SET_ERROR(EROFS)); 6246 6247 mutex_enter(&spa_namespace_lock); 6248 if ((spa = spa_lookup(pool)) == NULL) { 6249 mutex_exit(&spa_namespace_lock); 6250 return (SET_ERROR(ENOENT)); 6251 } 6252 6253 if (spa->spa_is_exporting) { 6254 /* the pool is being exported by another thread */ 6255 mutex_exit(&spa_namespace_lock); 6256 return (SET_ERROR(ZFS_ERR_EXPORT_IN_PROGRESS)); 6257 } 6258 spa->spa_is_exporting = B_TRUE; 6259 6260 /* 6261 * Put a hold on the pool, drop the namespace lock, stop async tasks, 6262 * reacquire the namespace lock, and see if we can export. 6263 */ 6264 spa_open_ref(spa, FTAG); 6265 mutex_exit(&spa_namespace_lock); 6266 spa_async_suspend(spa); 6267 if (spa->spa_zvol_taskq) { 6268 zvol_remove_minors(spa, spa_name(spa), B_TRUE); 6269 taskq_wait(spa->spa_zvol_taskq); 6270 } 6271 mutex_enter(&spa_namespace_lock); 6272 spa_close(spa, FTAG); 6273 6274 if (spa->spa_state == POOL_STATE_UNINITIALIZED) 6275 goto export_spa; 6276 /* 6277 * The pool will be in core if it's openable, in which case we can 6278 * modify its state. Objsets may be open only because they're dirty, 6279 * so we have to force it to sync before checking spa_refcnt. 6280 */ 6281 if (spa->spa_sync_on) { 6282 txg_wait_synced(spa->spa_dsl_pool, 0); 6283 spa_evicting_os_wait(spa); 6284 } 6285 6286 /* 6287 * A pool cannot be exported or destroyed if there are active 6288 * references. If we are resetting a pool, allow references by 6289 * fault injection handlers. 6290 */ 6291 if (!spa_refcount_zero(spa) || 6292 (spa->spa_inject_ref != 0 && 6293 new_state != POOL_STATE_UNINITIALIZED)) { 6294 spa_async_resume(spa); 6295 spa->spa_is_exporting = B_FALSE; 6296 mutex_exit(&spa_namespace_lock); 6297 return (SET_ERROR(EBUSY)); 6298 } 6299 6300 if (spa->spa_sync_on) { 6301 /* 6302 * A pool cannot be exported if it has an active shared spare. 6303 * This is to prevent other pools stealing the active spare 6304 * from an exported pool. At user's own will, such pool can 6305 * be forcedly exported. 6306 */ 6307 if (!force && new_state == POOL_STATE_EXPORTED && 6308 spa_has_active_shared_spare(spa)) { 6309 spa_async_resume(spa); 6310 spa->spa_is_exporting = B_FALSE; 6311 mutex_exit(&spa_namespace_lock); 6312 return (SET_ERROR(EXDEV)); 6313 } 6314 6315 /* 6316 * We're about to export or destroy this pool. Make sure 6317 * we stop all initialization and trim activity here before 6318 * we set the spa_final_txg. This will ensure that all 6319 * dirty data resulting from the initialization is 6320 * committed to disk before we unload the pool. 6321 */ 6322 if (spa->spa_root_vdev != NULL) { 6323 vdev_t *rvd = spa->spa_root_vdev; 6324 vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE); 6325 vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE); 6326 vdev_autotrim_stop_all(spa); 6327 vdev_rebuild_stop_all(spa); 6328 } 6329 6330 /* 6331 * We want this to be reflected on every label, 6332 * so mark them all dirty. spa_unload() will do the 6333 * final sync that pushes these changes out. 6334 */ 6335 if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) { 6336 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 6337 spa->spa_state = new_state; 6338 spa->spa_final_txg = spa_last_synced_txg(spa) + 6339 TXG_DEFER_SIZE + 1; 6340 vdev_config_dirty(spa->spa_root_vdev); 6341 spa_config_exit(spa, SCL_ALL, FTAG); 6342 } 6343 } 6344 6345 export_spa: 6346 if (new_state == POOL_STATE_DESTROYED) 6347 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY); 6348 else if (new_state == POOL_STATE_EXPORTED) 6349 spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_EXPORT); 6350 6351 if (spa->spa_state != POOL_STATE_UNINITIALIZED) { 6352 spa_unload(spa); 6353 spa_deactivate(spa); 6354 } 6355 6356 if (oldconfig && spa->spa_config) 6357 VERIFY(nvlist_dup(spa->spa_config, oldconfig, 0) == 0); 6358 6359 if (new_state != POOL_STATE_UNINITIALIZED) { 6360 if (!hardforce) 6361 spa_write_cachefile(spa, B_TRUE, B_TRUE); 6362 spa_remove(spa); 6363 } else { 6364 /* 6365 * If spa_remove() is not called for this spa_t and 6366 * there is any possibility that it can be reused, 6367 * we make sure to reset the exporting flag. 6368 */ 6369 spa->spa_is_exporting = B_FALSE; 6370 } 6371 6372 mutex_exit(&spa_namespace_lock); 6373 return (0); 6374 } 6375 6376 /* 6377 * Destroy a storage pool. 6378 */ 6379 int 6380 spa_destroy(const char *pool) 6381 { 6382 return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL, 6383 B_FALSE, B_FALSE)); 6384 } 6385 6386 /* 6387 * Export a storage pool. 6388 */ 6389 int 6390 spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force, 6391 boolean_t hardforce) 6392 { 6393 return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig, 6394 force, hardforce)); 6395 } 6396 6397 /* 6398 * Similar to spa_export(), this unloads the spa_t without actually removing it 6399 * from the namespace in any way. 6400 */ 6401 int 6402 spa_reset(const char *pool) 6403 { 6404 return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL, 6405 B_FALSE, B_FALSE)); 6406 } 6407 6408 /* 6409 * ========================================================================== 6410 * Device manipulation 6411 * ========================================================================== 6412 */ 6413 6414 /* 6415 * This is called as a synctask to increment the draid feature flag 6416 */ 6417 static void 6418 spa_draid_feature_incr(void *arg, dmu_tx_t *tx) 6419 { 6420 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 6421 int draid = (int)(uintptr_t)arg; 6422 6423 for (int c = 0; c < draid; c++) 6424 spa_feature_incr(spa, SPA_FEATURE_DRAID, tx); 6425 } 6426 6427 /* 6428 * Add a device to a storage pool. 6429 */ 6430 int 6431 spa_vdev_add(spa_t *spa, nvlist_t *nvroot) 6432 { 6433 uint64_t txg, ndraid = 0; 6434 int error; 6435 vdev_t *rvd = spa->spa_root_vdev; 6436 vdev_t *vd, *tvd; 6437 nvlist_t **spares, **l2cache; 6438 uint_t nspares, nl2cache; 6439 6440 ASSERT(spa_writeable(spa)); 6441 6442 txg = spa_vdev_enter(spa); 6443 6444 if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0, 6445 VDEV_ALLOC_ADD)) != 0) 6446 return (spa_vdev_exit(spa, NULL, txg, error)); 6447 6448 spa->spa_pending_vdev = vd; /* spa_vdev_exit() will clear this */ 6449 6450 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares, 6451 &nspares) != 0) 6452 nspares = 0; 6453 6454 if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache, 6455 &nl2cache) != 0) 6456 nl2cache = 0; 6457 6458 if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0) 6459 return (spa_vdev_exit(spa, vd, txg, EINVAL)); 6460 6461 if (vd->vdev_children != 0 && 6462 (error = vdev_create(vd, txg, B_FALSE)) != 0) { 6463 return (spa_vdev_exit(spa, vd, txg, error)); 6464 } 6465 6466 /* 6467 * The virtual dRAID spares must be added after vdev tree is created 6468 * and the vdev guids are generated. The guid of their assoicated 6469 * dRAID is stored in the config and used when opening the spare. 6470 */ 6471 if ((error = vdev_draid_spare_create(nvroot, vd, &ndraid, 6472 rvd->vdev_children)) == 0) { 6473 if (ndraid > 0 && nvlist_lookup_nvlist_array(nvroot, 6474 ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0) 6475 nspares = 0; 6476 } else { 6477 return (spa_vdev_exit(spa, vd, txg, error)); 6478 } 6479 6480 /* 6481 * We must validate the spares and l2cache devices after checking the 6482 * children. Otherwise, vdev_inuse() will blindly overwrite the spare. 6483 */ 6484 if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0) 6485 return (spa_vdev_exit(spa, vd, txg, error)); 6486 6487 /* 6488 * If we are in the middle of a device removal, we can only add 6489 * devices which match the existing devices in the pool. 6490 * If we are in the middle of a removal, or have some indirect 6491 * vdevs, we can not add raidz or dRAID top levels. 6492 */ 6493 if (spa->spa_vdev_removal != NULL || 6494 spa->spa_removing_phys.sr_prev_indirect_vdev != -1) { 6495 for (int c = 0; c < vd->vdev_children; c++) { 6496 tvd = vd->vdev_child[c]; 6497 if (spa->spa_vdev_removal != NULL && 6498 tvd->vdev_ashift != spa->spa_max_ashift) { 6499 return (spa_vdev_exit(spa, vd, txg, EINVAL)); 6500 } 6501 /* Fail if top level vdev is raidz or a dRAID */ 6502 if (vdev_get_nparity(tvd) != 0) 6503 return (spa_vdev_exit(spa, vd, txg, EINVAL)); 6504 6505 /* 6506 * Need the top level mirror to be 6507 * a mirror of leaf vdevs only 6508 */ 6509 if (tvd->vdev_ops == &vdev_mirror_ops) { 6510 for (uint64_t cid = 0; 6511 cid < tvd->vdev_children; cid++) { 6512 vdev_t *cvd = tvd->vdev_child[cid]; 6513 if (!cvd->vdev_ops->vdev_op_leaf) { 6514 return (spa_vdev_exit(spa, vd, 6515 txg, EINVAL)); 6516 } 6517 } 6518 } 6519 } 6520 } 6521 6522 for (int c = 0; c < vd->vdev_children; c++) { 6523 tvd = vd->vdev_child[c]; 6524 vdev_remove_child(vd, tvd); 6525 tvd->vdev_id = rvd->vdev_children; 6526 vdev_add_child(rvd, tvd); 6527 vdev_config_dirty(tvd); 6528 } 6529 6530 if (nspares != 0) { 6531 spa_set_aux_vdevs(&spa->spa_spares, spares, nspares, 6532 ZPOOL_CONFIG_SPARES); 6533 spa_load_spares(spa); 6534 spa->spa_spares.sav_sync = B_TRUE; 6535 } 6536 6537 if (nl2cache != 0) { 6538 spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache, 6539 ZPOOL_CONFIG_L2CACHE); 6540 spa_load_l2cache(spa); 6541 spa->spa_l2cache.sav_sync = B_TRUE; 6542 } 6543 6544 /* 6545 * We can't increment a feature while holding spa_vdev so we 6546 * have to do it in a synctask. 6547 */ 6548 if (ndraid != 0) { 6549 dmu_tx_t *tx; 6550 6551 tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg); 6552 dsl_sync_task_nowait(spa->spa_dsl_pool, spa_draid_feature_incr, 6553 (void *)(uintptr_t)ndraid, tx); 6554 dmu_tx_commit(tx); 6555 } 6556 6557 /* 6558 * We have to be careful when adding new vdevs to an existing pool. 6559 * If other threads start allocating from these vdevs before we 6560 * sync the config cache, and we lose power, then upon reboot we may 6561 * fail to open the pool because there are DVAs that the config cache 6562 * can't translate. Therefore, we first add the vdevs without 6563 * initializing metaslabs; sync the config cache (via spa_vdev_exit()); 6564 * and then let spa_config_update() initialize the new metaslabs. 6565 * 6566 * spa_load() checks for added-but-not-initialized vdevs, so that 6567 * if we lose power at any point in this sequence, the remaining 6568 * steps will be completed the next time we load the pool. 6569 */ 6570 (void) spa_vdev_exit(spa, vd, txg, 0); 6571 6572 mutex_enter(&spa_namespace_lock); 6573 spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); 6574 spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD); 6575 mutex_exit(&spa_namespace_lock); 6576 6577 return (0); 6578 } 6579 6580 /* 6581 * Attach a device to a mirror. The arguments are the path to any device 6582 * in the mirror, and the nvroot for the new device. If the path specifies 6583 * a device that is not mirrored, we automatically insert the mirror vdev. 6584 * 6585 * If 'replacing' is specified, the new device is intended to replace the 6586 * existing device; in this case the two devices are made into their own 6587 * mirror using the 'replacing' vdev, which is functionally identical to 6588 * the mirror vdev (it actually reuses all the same ops) but has a few 6589 * extra rules: you can't attach to it after it's been created, and upon 6590 * completion of resilvering, the first disk (the one being replaced) 6591 * is automatically detached. 6592 * 6593 * If 'rebuild' is specified, then sequential reconstruction (a.ka. rebuild) 6594 * should be performed instead of traditional healing reconstruction. From 6595 * an administrators perspective these are both resilver operations. 6596 */ 6597 int 6598 spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing, 6599 int rebuild) 6600 { 6601 uint64_t txg, dtl_max_txg; 6602 vdev_t *rvd = spa->spa_root_vdev; 6603 vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd; 6604 vdev_ops_t *pvops; 6605 char *oldvdpath, *newvdpath; 6606 int newvd_isspare; 6607 int error; 6608 6609 ASSERT(spa_writeable(spa)); 6610 6611 txg = spa_vdev_enter(spa); 6612 6613 oldvd = spa_lookup_by_guid(spa, guid, B_FALSE); 6614 6615 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 6616 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 6617 error = (spa_has_checkpoint(spa)) ? 6618 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 6619 return (spa_vdev_exit(spa, NULL, txg, error)); 6620 } 6621 6622 if (rebuild) { 6623 if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REBUILD)) 6624 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6625 6626 if (dsl_scan_resilvering(spa_get_dsl(spa))) 6627 return (spa_vdev_exit(spa, NULL, txg, 6628 ZFS_ERR_RESILVER_IN_PROGRESS)); 6629 } else { 6630 if (vdev_rebuild_active(rvd)) 6631 return (spa_vdev_exit(spa, NULL, txg, 6632 ZFS_ERR_REBUILD_IN_PROGRESS)); 6633 } 6634 6635 if (spa->spa_vdev_removal != NULL) 6636 return (spa_vdev_exit(spa, NULL, txg, EBUSY)); 6637 6638 if (oldvd == NULL) 6639 return (spa_vdev_exit(spa, NULL, txg, ENODEV)); 6640 6641 if (!oldvd->vdev_ops->vdev_op_leaf) 6642 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6643 6644 pvd = oldvd->vdev_parent; 6645 6646 if ((error = spa_config_parse(spa, &newrootvd, nvroot, NULL, 0, 6647 VDEV_ALLOC_ATTACH)) != 0) 6648 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 6649 6650 if (newrootvd->vdev_children != 1) 6651 return (spa_vdev_exit(spa, newrootvd, txg, EINVAL)); 6652 6653 newvd = newrootvd->vdev_child[0]; 6654 6655 if (!newvd->vdev_ops->vdev_op_leaf) 6656 return (spa_vdev_exit(spa, newrootvd, txg, EINVAL)); 6657 6658 if ((error = vdev_create(newrootvd, txg, replacing)) != 0) 6659 return (spa_vdev_exit(spa, newrootvd, txg, error)); 6660 6661 /* 6662 * Spares can't replace logs 6663 */ 6664 if (oldvd->vdev_top->vdev_islog && newvd->vdev_isspare) 6665 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6666 6667 /* 6668 * A dRAID spare can only replace a child of its parent dRAID vdev. 6669 */ 6670 if (newvd->vdev_ops == &vdev_draid_spare_ops && 6671 oldvd->vdev_top != vdev_draid_spare_get_parent(newvd)) { 6672 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6673 } 6674 6675 if (rebuild) { 6676 /* 6677 * For rebuilds, the top vdev must support reconstruction 6678 * using only space maps. This means the only allowable 6679 * vdevs types are the root vdev, a mirror, or dRAID. 6680 */ 6681 tvd = pvd; 6682 if (pvd->vdev_top != NULL) 6683 tvd = pvd->vdev_top; 6684 6685 if (tvd->vdev_ops != &vdev_mirror_ops && 6686 tvd->vdev_ops != &vdev_root_ops && 6687 tvd->vdev_ops != &vdev_draid_ops) { 6688 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6689 } 6690 } 6691 6692 if (!replacing) { 6693 /* 6694 * For attach, the only allowable parent is a mirror or the root 6695 * vdev. 6696 */ 6697 if (pvd->vdev_ops != &vdev_mirror_ops && 6698 pvd->vdev_ops != &vdev_root_ops) 6699 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6700 6701 pvops = &vdev_mirror_ops; 6702 } else { 6703 /* 6704 * Active hot spares can only be replaced by inactive hot 6705 * spares. 6706 */ 6707 if (pvd->vdev_ops == &vdev_spare_ops && 6708 oldvd->vdev_isspare && 6709 !spa_has_spare(spa, newvd->vdev_guid)) 6710 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6711 6712 /* 6713 * If the source is a hot spare, and the parent isn't already a 6714 * spare, then we want to create a new hot spare. Otherwise, we 6715 * want to create a replacing vdev. The user is not allowed to 6716 * attach to a spared vdev child unless the 'isspare' state is 6717 * the same (spare replaces spare, non-spare replaces 6718 * non-spare). 6719 */ 6720 if (pvd->vdev_ops == &vdev_replacing_ops && 6721 spa_version(spa) < SPA_VERSION_MULTI_REPLACE) { 6722 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6723 } else if (pvd->vdev_ops == &vdev_spare_ops && 6724 newvd->vdev_isspare != oldvd->vdev_isspare) { 6725 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6726 } 6727 6728 if (newvd->vdev_isspare) 6729 pvops = &vdev_spare_ops; 6730 else 6731 pvops = &vdev_replacing_ops; 6732 } 6733 6734 /* 6735 * Make sure the new device is big enough. 6736 */ 6737 if (newvd->vdev_asize < vdev_get_min_asize(oldvd)) 6738 return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW)); 6739 6740 /* 6741 * The new device cannot have a higher alignment requirement 6742 * than the top-level vdev. 6743 */ 6744 if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift) 6745 return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP)); 6746 6747 /* 6748 * If this is an in-place replacement, update oldvd's path and devid 6749 * to make it distinguishable from newvd, and unopenable from now on. 6750 */ 6751 if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) { 6752 spa_strfree(oldvd->vdev_path); 6753 oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5, 6754 KM_SLEEP); 6755 (void) snprintf(oldvd->vdev_path, strlen(newvd->vdev_path) + 5, 6756 "%s/%s", newvd->vdev_path, "old"); 6757 if (oldvd->vdev_devid != NULL) { 6758 spa_strfree(oldvd->vdev_devid); 6759 oldvd->vdev_devid = NULL; 6760 } 6761 } 6762 6763 /* 6764 * If the parent is not a mirror, or if we're replacing, insert the new 6765 * mirror/replacing/spare vdev above oldvd. 6766 */ 6767 if (pvd->vdev_ops != pvops) 6768 pvd = vdev_add_parent(oldvd, pvops); 6769 6770 ASSERT(pvd->vdev_top->vdev_parent == rvd); 6771 ASSERT(pvd->vdev_ops == pvops); 6772 ASSERT(oldvd->vdev_parent == pvd); 6773 6774 /* 6775 * Extract the new device from its root and add it to pvd. 6776 */ 6777 vdev_remove_child(newrootvd, newvd); 6778 newvd->vdev_id = pvd->vdev_children; 6779 newvd->vdev_crtxg = oldvd->vdev_crtxg; 6780 vdev_add_child(pvd, newvd); 6781 6782 /* 6783 * Reevaluate the parent vdev state. 6784 */ 6785 vdev_propagate_state(pvd); 6786 6787 tvd = newvd->vdev_top; 6788 ASSERT(pvd->vdev_top == tvd); 6789 ASSERT(tvd->vdev_parent == rvd); 6790 6791 vdev_config_dirty(tvd); 6792 6793 /* 6794 * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account 6795 * for any dmu_sync-ed blocks. It will propagate upward when 6796 * spa_vdev_exit() calls vdev_dtl_reassess(). 6797 */ 6798 dtl_max_txg = txg + TXG_CONCURRENT_STATES; 6799 6800 vdev_dtl_dirty(newvd, DTL_MISSING, 6801 TXG_INITIAL, dtl_max_txg - TXG_INITIAL); 6802 6803 if (newvd->vdev_isspare) { 6804 spa_spare_activate(newvd); 6805 spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE); 6806 } 6807 6808 oldvdpath = spa_strdup(oldvd->vdev_path); 6809 newvdpath = spa_strdup(newvd->vdev_path); 6810 newvd_isspare = newvd->vdev_isspare; 6811 6812 /* 6813 * Mark newvd's DTL dirty in this txg. 6814 */ 6815 vdev_dirty(tvd, VDD_DTL, newvd, txg); 6816 6817 /* 6818 * Schedule the resilver or rebuild to restart in the future. We do 6819 * this to ensure that dmu_sync-ed blocks have been stitched into the 6820 * respective datasets. 6821 */ 6822 if (rebuild) { 6823 newvd->vdev_rebuild_txg = txg; 6824 6825 vdev_rebuild(tvd); 6826 } else { 6827 newvd->vdev_resilver_txg = txg; 6828 6829 if (dsl_scan_resilvering(spa_get_dsl(spa)) && 6830 spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) { 6831 vdev_defer_resilver(newvd); 6832 } else { 6833 dsl_scan_restart_resilver(spa->spa_dsl_pool, 6834 dtl_max_txg); 6835 } 6836 } 6837 6838 if (spa->spa_bootfs) 6839 spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH); 6840 6841 spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH); 6842 6843 /* 6844 * Commit the config 6845 */ 6846 (void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0); 6847 6848 spa_history_log_internal(spa, "vdev attach", NULL, 6849 "%s vdev=%s %s vdev=%s", 6850 replacing && newvd_isspare ? "spare in" : 6851 replacing ? "replace" : "attach", newvdpath, 6852 replacing ? "for" : "to", oldvdpath); 6853 6854 spa_strfree(oldvdpath); 6855 spa_strfree(newvdpath); 6856 6857 return (0); 6858 } 6859 6860 /* 6861 * Detach a device from a mirror or replacing vdev. 6862 * 6863 * If 'replace_done' is specified, only detach if the parent 6864 * is a replacing vdev. 6865 */ 6866 int 6867 spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done) 6868 { 6869 uint64_t txg; 6870 int error; 6871 vdev_t *rvd __maybe_unused = spa->spa_root_vdev; 6872 vdev_t *vd, *pvd, *cvd, *tvd; 6873 boolean_t unspare = B_FALSE; 6874 uint64_t unspare_guid = 0; 6875 char *vdpath; 6876 6877 ASSERT(spa_writeable(spa)); 6878 6879 txg = spa_vdev_detach_enter(spa, guid); 6880 6881 vd = spa_lookup_by_guid(spa, guid, B_FALSE); 6882 6883 /* 6884 * Besides being called directly from the userland through the 6885 * ioctl interface, spa_vdev_detach() can be potentially called 6886 * at the end of spa_vdev_resilver_done(). 6887 * 6888 * In the regular case, when we have a checkpoint this shouldn't 6889 * happen as we never empty the DTLs of a vdev during the scrub 6890 * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done() 6891 * should never get here when we have a checkpoint. 6892 * 6893 * That said, even in a case when we checkpoint the pool exactly 6894 * as spa_vdev_resilver_done() calls this function everything 6895 * should be fine as the resilver will return right away. 6896 */ 6897 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 6898 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 6899 error = (spa_has_checkpoint(spa)) ? 6900 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 6901 return (spa_vdev_exit(spa, NULL, txg, error)); 6902 } 6903 6904 if (vd == NULL) 6905 return (spa_vdev_exit(spa, NULL, txg, ENODEV)); 6906 6907 if (!vd->vdev_ops->vdev_op_leaf) 6908 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6909 6910 pvd = vd->vdev_parent; 6911 6912 /* 6913 * If the parent/child relationship is not as expected, don't do it. 6914 * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing 6915 * vdev that's replacing B with C. The user's intent in replacing 6916 * is to go from M(A,B) to M(A,C). If the user decides to cancel 6917 * the replace by detaching C, the expected behavior is to end up 6918 * M(A,B). But suppose that right after deciding to detach C, 6919 * the replacement of B completes. We would have M(A,C), and then 6920 * ask to detach C, which would leave us with just A -- not what 6921 * the user wanted. To prevent this, we make sure that the 6922 * parent/child relationship hasn't changed -- in this example, 6923 * that C's parent is still the replacing vdev R. 6924 */ 6925 if (pvd->vdev_guid != pguid && pguid != 0) 6926 return (spa_vdev_exit(spa, NULL, txg, EBUSY)); 6927 6928 /* 6929 * Only 'replacing' or 'spare' vdevs can be replaced. 6930 */ 6931 if (replace_done && pvd->vdev_ops != &vdev_replacing_ops && 6932 pvd->vdev_ops != &vdev_spare_ops) 6933 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6934 6935 ASSERT(pvd->vdev_ops != &vdev_spare_ops || 6936 spa_version(spa) >= SPA_VERSION_SPARES); 6937 6938 /* 6939 * Only mirror, replacing, and spare vdevs support detach. 6940 */ 6941 if (pvd->vdev_ops != &vdev_replacing_ops && 6942 pvd->vdev_ops != &vdev_mirror_ops && 6943 pvd->vdev_ops != &vdev_spare_ops) 6944 return (spa_vdev_exit(spa, NULL, txg, ENOTSUP)); 6945 6946 /* 6947 * If this device has the only valid copy of some data, 6948 * we cannot safely detach it. 6949 */ 6950 if (vdev_dtl_required(vd)) 6951 return (spa_vdev_exit(spa, NULL, txg, EBUSY)); 6952 6953 ASSERT(pvd->vdev_children >= 2); 6954 6955 /* 6956 * If we are detaching the second disk from a replacing vdev, then 6957 * check to see if we changed the original vdev's path to have "/old" 6958 * at the end in spa_vdev_attach(). If so, undo that change now. 6959 */ 6960 if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 && 6961 vd->vdev_path != NULL) { 6962 size_t len = strlen(vd->vdev_path); 6963 6964 for (int c = 0; c < pvd->vdev_children; c++) { 6965 cvd = pvd->vdev_child[c]; 6966 6967 if (cvd == vd || cvd->vdev_path == NULL) 6968 continue; 6969 6970 if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 && 6971 strcmp(cvd->vdev_path + len, "/old") == 0) { 6972 spa_strfree(cvd->vdev_path); 6973 cvd->vdev_path = spa_strdup(vd->vdev_path); 6974 break; 6975 } 6976 } 6977 } 6978 6979 /* 6980 * If we are detaching the original disk from a normal spare, then it 6981 * implies that the spare should become a real disk, and be removed 6982 * from the active spare list for the pool. dRAID spares on the 6983 * other hand are coupled to the pool and thus should never be removed 6984 * from the spares list. 6985 */ 6986 if (pvd->vdev_ops == &vdev_spare_ops && vd->vdev_id == 0) { 6987 vdev_t *last_cvd = pvd->vdev_child[pvd->vdev_children - 1]; 6988 6989 if (last_cvd->vdev_isspare && 6990 last_cvd->vdev_ops != &vdev_draid_spare_ops) { 6991 unspare = B_TRUE; 6992 } 6993 } 6994 6995 /* 6996 * Erase the disk labels so the disk can be used for other things. 6997 * This must be done after all other error cases are handled, 6998 * but before we disembowel vd (so we can still do I/O to it). 6999 * But if we can't do it, don't treat the error as fatal -- 7000 * it may be that the unwritability of the disk is the reason 7001 * it's being detached! 7002 */ 7003 error = vdev_label_init(vd, 0, VDEV_LABEL_REMOVE); 7004 7005 /* 7006 * Remove vd from its parent and compact the parent's children. 7007 */ 7008 vdev_remove_child(pvd, vd); 7009 vdev_compact_children(pvd); 7010 7011 /* 7012 * Remember one of the remaining children so we can get tvd below. 7013 */ 7014 cvd = pvd->vdev_child[pvd->vdev_children - 1]; 7015 7016 /* 7017 * If we need to remove the remaining child from the list of hot spares, 7018 * do it now, marking the vdev as no longer a spare in the process. 7019 * We must do this before vdev_remove_parent(), because that can 7020 * change the GUID if it creates a new toplevel GUID. For a similar 7021 * reason, we must remove the spare now, in the same txg as the detach; 7022 * otherwise someone could attach a new sibling, change the GUID, and 7023 * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail. 7024 */ 7025 if (unspare) { 7026 ASSERT(cvd->vdev_isspare); 7027 spa_spare_remove(cvd); 7028 unspare_guid = cvd->vdev_guid; 7029 (void) spa_vdev_remove(spa, unspare_guid, B_TRUE); 7030 cvd->vdev_unspare = B_TRUE; 7031 } 7032 7033 /* 7034 * If the parent mirror/replacing vdev only has one child, 7035 * the parent is no longer needed. Remove it from the tree. 7036 */ 7037 if (pvd->vdev_children == 1) { 7038 if (pvd->vdev_ops == &vdev_spare_ops) 7039 cvd->vdev_unspare = B_FALSE; 7040 vdev_remove_parent(cvd); 7041 } 7042 7043 /* 7044 * We don't set tvd until now because the parent we just removed 7045 * may have been the previous top-level vdev. 7046 */ 7047 tvd = cvd->vdev_top; 7048 ASSERT(tvd->vdev_parent == rvd); 7049 7050 /* 7051 * Reevaluate the parent vdev state. 7052 */ 7053 vdev_propagate_state(cvd); 7054 7055 /* 7056 * If the 'autoexpand' property is set on the pool then automatically 7057 * try to expand the size of the pool. For example if the device we 7058 * just detached was smaller than the others, it may be possible to 7059 * add metaslabs (i.e. grow the pool). We need to reopen the vdev 7060 * first so that we can obtain the updated sizes of the leaf vdevs. 7061 */ 7062 if (spa->spa_autoexpand) { 7063 vdev_reopen(tvd); 7064 vdev_expand(tvd, txg); 7065 } 7066 7067 vdev_config_dirty(tvd); 7068 7069 /* 7070 * Mark vd's DTL as dirty in this txg. vdev_dtl_sync() will see that 7071 * vd->vdev_detached is set and free vd's DTL object in syncing context. 7072 * But first make sure we're not on any *other* txg's DTL list, to 7073 * prevent vd from being accessed after it's freed. 7074 */ 7075 vdpath = spa_strdup(vd->vdev_path ? vd->vdev_path : "none"); 7076 for (int t = 0; t < TXG_SIZE; t++) 7077 (void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t); 7078 vd->vdev_detached = B_TRUE; 7079 vdev_dirty(tvd, VDD_DTL, vd, txg); 7080 7081 spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE); 7082 spa_notify_waiters(spa); 7083 7084 /* hang on to the spa before we release the lock */ 7085 spa_open_ref(spa, FTAG); 7086 7087 error = spa_vdev_exit(spa, vd, txg, 0); 7088 7089 spa_history_log_internal(spa, "detach", NULL, 7090 "vdev=%s", vdpath); 7091 spa_strfree(vdpath); 7092 7093 /* 7094 * If this was the removal of the original device in a hot spare vdev, 7095 * then we want to go through and remove the device from the hot spare 7096 * list of every other pool. 7097 */ 7098 if (unspare) { 7099 spa_t *altspa = NULL; 7100 7101 mutex_enter(&spa_namespace_lock); 7102 while ((altspa = spa_next(altspa)) != NULL) { 7103 if (altspa->spa_state != POOL_STATE_ACTIVE || 7104 altspa == spa) 7105 continue; 7106 7107 spa_open_ref(altspa, FTAG); 7108 mutex_exit(&spa_namespace_lock); 7109 (void) spa_vdev_remove(altspa, unspare_guid, B_TRUE); 7110 mutex_enter(&spa_namespace_lock); 7111 spa_close(altspa, FTAG); 7112 } 7113 mutex_exit(&spa_namespace_lock); 7114 7115 /* search the rest of the vdevs for spares to remove */ 7116 spa_vdev_resilver_done(spa); 7117 } 7118 7119 /* all done with the spa; OK to release */ 7120 mutex_enter(&spa_namespace_lock); 7121 spa_close(spa, FTAG); 7122 mutex_exit(&spa_namespace_lock); 7123 7124 return (error); 7125 } 7126 7127 static int 7128 spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type, 7129 list_t *vd_list) 7130 { 7131 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7132 7133 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 7134 7135 /* Look up vdev and ensure it's a leaf. */ 7136 vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE); 7137 if (vd == NULL || vd->vdev_detached) { 7138 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7139 return (SET_ERROR(ENODEV)); 7140 } else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) { 7141 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7142 return (SET_ERROR(EINVAL)); 7143 } else if (!vdev_writeable(vd)) { 7144 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7145 return (SET_ERROR(EROFS)); 7146 } 7147 mutex_enter(&vd->vdev_initialize_lock); 7148 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7149 7150 /* 7151 * When we activate an initialize action we check to see 7152 * if the vdev_initialize_thread is NULL. We do this instead 7153 * of using the vdev_initialize_state since there might be 7154 * a previous initialization process which has completed but 7155 * the thread is not exited. 7156 */ 7157 if (cmd_type == POOL_INITIALIZE_START && 7158 (vd->vdev_initialize_thread != NULL || 7159 vd->vdev_top->vdev_removing)) { 7160 mutex_exit(&vd->vdev_initialize_lock); 7161 return (SET_ERROR(EBUSY)); 7162 } else if (cmd_type == POOL_INITIALIZE_CANCEL && 7163 (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE && 7164 vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) { 7165 mutex_exit(&vd->vdev_initialize_lock); 7166 return (SET_ERROR(ESRCH)); 7167 } else if (cmd_type == POOL_INITIALIZE_SUSPEND && 7168 vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) { 7169 mutex_exit(&vd->vdev_initialize_lock); 7170 return (SET_ERROR(ESRCH)); 7171 } 7172 7173 switch (cmd_type) { 7174 case POOL_INITIALIZE_START: 7175 vdev_initialize(vd); 7176 break; 7177 case POOL_INITIALIZE_CANCEL: 7178 vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, vd_list); 7179 break; 7180 case POOL_INITIALIZE_SUSPEND: 7181 vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED, vd_list); 7182 break; 7183 default: 7184 panic("invalid cmd_type %llu", (unsigned long long)cmd_type); 7185 } 7186 mutex_exit(&vd->vdev_initialize_lock); 7187 7188 return (0); 7189 } 7190 7191 int 7192 spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, 7193 nvlist_t *vdev_errlist) 7194 { 7195 int total_errors = 0; 7196 list_t vd_list; 7197 7198 list_create(&vd_list, sizeof (vdev_t), 7199 offsetof(vdev_t, vdev_initialize_node)); 7200 7201 /* 7202 * We hold the namespace lock through the whole function 7203 * to prevent any changes to the pool while we're starting or 7204 * stopping initialization. The config and state locks are held so that 7205 * we can properly assess the vdev state before we commit to 7206 * the initializing operation. 7207 */ 7208 mutex_enter(&spa_namespace_lock); 7209 7210 for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL); 7211 pair != NULL; pair = nvlist_next_nvpair(nv, pair)) { 7212 uint64_t vdev_guid = fnvpair_value_uint64(pair); 7213 7214 int error = spa_vdev_initialize_impl(spa, vdev_guid, cmd_type, 7215 &vd_list); 7216 if (error != 0) { 7217 char guid_as_str[MAXNAMELEN]; 7218 7219 (void) snprintf(guid_as_str, sizeof (guid_as_str), 7220 "%llu", (unsigned long long)vdev_guid); 7221 fnvlist_add_int64(vdev_errlist, guid_as_str, error); 7222 total_errors++; 7223 } 7224 } 7225 7226 /* Wait for all initialize threads to stop. */ 7227 vdev_initialize_stop_wait(spa, &vd_list); 7228 7229 /* Sync out the initializing state */ 7230 txg_wait_synced(spa->spa_dsl_pool, 0); 7231 mutex_exit(&spa_namespace_lock); 7232 7233 list_destroy(&vd_list); 7234 7235 return (total_errors); 7236 } 7237 7238 static int 7239 spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type, 7240 uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list) 7241 { 7242 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7243 7244 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 7245 7246 /* Look up vdev and ensure it's a leaf. */ 7247 vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE); 7248 if (vd == NULL || vd->vdev_detached) { 7249 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7250 return (SET_ERROR(ENODEV)); 7251 } else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) { 7252 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7253 return (SET_ERROR(EINVAL)); 7254 } else if (!vdev_writeable(vd)) { 7255 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7256 return (SET_ERROR(EROFS)); 7257 } else if (!vd->vdev_has_trim) { 7258 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7259 return (SET_ERROR(EOPNOTSUPP)); 7260 } else if (secure && !vd->vdev_has_securetrim) { 7261 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7262 return (SET_ERROR(EOPNOTSUPP)); 7263 } 7264 mutex_enter(&vd->vdev_trim_lock); 7265 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 7266 7267 /* 7268 * When we activate a TRIM action we check to see if the 7269 * vdev_trim_thread is NULL. We do this instead of using the 7270 * vdev_trim_state since there might be a previous TRIM process 7271 * which has completed but the thread is not exited. 7272 */ 7273 if (cmd_type == POOL_TRIM_START && 7274 (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) { 7275 mutex_exit(&vd->vdev_trim_lock); 7276 return (SET_ERROR(EBUSY)); 7277 } else if (cmd_type == POOL_TRIM_CANCEL && 7278 (vd->vdev_trim_state != VDEV_TRIM_ACTIVE && 7279 vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) { 7280 mutex_exit(&vd->vdev_trim_lock); 7281 return (SET_ERROR(ESRCH)); 7282 } else if (cmd_type == POOL_TRIM_SUSPEND && 7283 vd->vdev_trim_state != VDEV_TRIM_ACTIVE) { 7284 mutex_exit(&vd->vdev_trim_lock); 7285 return (SET_ERROR(ESRCH)); 7286 } 7287 7288 switch (cmd_type) { 7289 case POOL_TRIM_START: 7290 vdev_trim(vd, rate, partial, secure); 7291 break; 7292 case POOL_TRIM_CANCEL: 7293 vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list); 7294 break; 7295 case POOL_TRIM_SUSPEND: 7296 vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list); 7297 break; 7298 default: 7299 panic("invalid cmd_type %llu", (unsigned long long)cmd_type); 7300 } 7301 mutex_exit(&vd->vdev_trim_lock); 7302 7303 return (0); 7304 } 7305 7306 /* 7307 * Initiates a manual TRIM for the requested vdevs. This kicks off individual 7308 * TRIM threads for each child vdev. These threads pass over all of the free 7309 * space in the vdev's metaslabs and issues TRIM commands for that space. 7310 */ 7311 int 7312 spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate, 7313 boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist) 7314 { 7315 int total_errors = 0; 7316 list_t vd_list; 7317 7318 list_create(&vd_list, sizeof (vdev_t), 7319 offsetof(vdev_t, vdev_trim_node)); 7320 7321 /* 7322 * We hold the namespace lock through the whole function 7323 * to prevent any changes to the pool while we're starting or 7324 * stopping TRIM. The config and state locks are held so that 7325 * we can properly assess the vdev state before we commit to 7326 * the TRIM operation. 7327 */ 7328 mutex_enter(&spa_namespace_lock); 7329 7330 for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL); 7331 pair != NULL; pair = nvlist_next_nvpair(nv, pair)) { 7332 uint64_t vdev_guid = fnvpair_value_uint64(pair); 7333 7334 int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type, 7335 rate, partial, secure, &vd_list); 7336 if (error != 0) { 7337 char guid_as_str[MAXNAMELEN]; 7338 7339 (void) snprintf(guid_as_str, sizeof (guid_as_str), 7340 "%llu", (unsigned long long)vdev_guid); 7341 fnvlist_add_int64(vdev_errlist, guid_as_str, error); 7342 total_errors++; 7343 } 7344 } 7345 7346 /* Wait for all TRIM threads to stop. */ 7347 vdev_trim_stop_wait(spa, &vd_list); 7348 7349 /* Sync out the TRIM state */ 7350 txg_wait_synced(spa->spa_dsl_pool, 0); 7351 mutex_exit(&spa_namespace_lock); 7352 7353 list_destroy(&vd_list); 7354 7355 return (total_errors); 7356 } 7357 7358 /* 7359 * Split a set of devices from their mirrors, and create a new pool from them. 7360 */ 7361 int 7362 spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config, 7363 nvlist_t *props, boolean_t exp) 7364 { 7365 int error = 0; 7366 uint64_t txg, *glist; 7367 spa_t *newspa; 7368 uint_t c, children, lastlog; 7369 nvlist_t **child, *nvl, *tmp; 7370 dmu_tx_t *tx; 7371 char *altroot = NULL; 7372 vdev_t *rvd, **vml = NULL; /* vdev modify list */ 7373 boolean_t activate_slog; 7374 7375 ASSERT(spa_writeable(spa)); 7376 7377 txg = spa_vdev_enter(spa); 7378 7379 ASSERT(MUTEX_HELD(&spa_namespace_lock)); 7380 if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) { 7381 error = (spa_has_checkpoint(spa)) ? 7382 ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT; 7383 return (spa_vdev_exit(spa, NULL, txg, error)); 7384 } 7385 7386 /* clear the log and flush everything up to now */ 7387 activate_slog = spa_passivate_log(spa); 7388 (void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG); 7389 error = spa_reset_logs(spa); 7390 txg = spa_vdev_config_enter(spa); 7391 7392 if (activate_slog) 7393 spa_activate_log(spa); 7394 7395 if (error != 0) 7396 return (spa_vdev_exit(spa, NULL, txg, error)); 7397 7398 /* check new spa name before going any further */ 7399 if (spa_lookup(newname) != NULL) 7400 return (spa_vdev_exit(spa, NULL, txg, EEXIST)); 7401 7402 /* 7403 * scan through all the children to ensure they're all mirrors 7404 */ 7405 if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 || 7406 nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child, 7407 &children) != 0) 7408 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 7409 7410 /* first, check to ensure we've got the right child count */ 7411 rvd = spa->spa_root_vdev; 7412 lastlog = 0; 7413 for (c = 0; c < rvd->vdev_children; c++) { 7414 vdev_t *vd = rvd->vdev_child[c]; 7415 7416 /* don't count the holes & logs as children */ 7417 if (vd->vdev_islog || (vd->vdev_ops != &vdev_indirect_ops && 7418 !vdev_is_concrete(vd))) { 7419 if (lastlog == 0) 7420 lastlog = c; 7421 continue; 7422 } 7423 7424 lastlog = 0; 7425 } 7426 if (children != (lastlog != 0 ? lastlog : rvd->vdev_children)) 7427 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 7428 7429 /* next, ensure no spare or cache devices are part of the split */ 7430 if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 || 7431 nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0) 7432 return (spa_vdev_exit(spa, NULL, txg, EINVAL)); 7433 7434 vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP); 7435 glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP); 7436 7437 /* then, loop over each vdev and validate it */ 7438 for (c = 0; c < children; c++) { 7439 uint64_t is_hole = 0; 7440 7441 (void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE, 7442 &is_hole); 7443 7444 if (is_hole != 0) { 7445 if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole || 7446 spa->spa_root_vdev->vdev_child[c]->vdev_islog) { 7447 continue; 7448 } else { 7449 error = SET_ERROR(EINVAL); 7450 break; 7451 } 7452 } 7453 7454 /* deal with indirect vdevs */ 7455 if (spa->spa_root_vdev->vdev_child[c]->vdev_ops == 7456 &vdev_indirect_ops) 7457 continue; 7458 7459 /* which disk is going to be split? */ 7460 if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID, 7461 &glist[c]) != 0) { 7462 error = SET_ERROR(EINVAL); 7463 break; 7464 } 7465 7466 /* look it up in the spa */ 7467 vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE); 7468 if (vml[c] == NULL) { 7469 error = SET_ERROR(ENODEV); 7470 break; 7471 } 7472 7473 /* make sure there's nothing stopping the split */ 7474 if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops || 7475 vml[c]->vdev_islog || 7476 !vdev_is_concrete(vml[c]) || 7477 vml[c]->vdev_isspare || 7478 vml[c]->vdev_isl2cache || 7479 !vdev_writeable(vml[c]) || 7480 vml[c]->vdev_children != 0 || 7481 vml[c]->vdev_state != VDEV_STATE_HEALTHY || 7482 c != spa->spa_root_vdev->vdev_child[c]->vdev_id) { 7483 error = SET_ERROR(EINVAL); 7484 break; 7485 } 7486 7487 if (vdev_dtl_required(vml[c]) || 7488 vdev_resilver_needed(vml[c], NULL, NULL)) { 7489 error = SET_ERROR(EBUSY); 7490 break; 7491 } 7492 7493 /* we need certain info from the top level */ 7494 VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY, 7495 vml[c]->vdev_top->vdev_ms_array) == 0); 7496 VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT, 7497 vml[c]->vdev_top->vdev_ms_shift) == 0); 7498 VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE, 7499 vml[c]->vdev_top->vdev_asize) == 0); 7500 VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT, 7501 vml[c]->vdev_top->vdev_ashift) == 0); 7502 7503 /* transfer per-vdev ZAPs */ 7504 ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0); 7505 VERIFY0(nvlist_add_uint64(child[c], 7506 ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap)); 7507 7508 ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0); 7509 VERIFY0(nvlist_add_uint64(child[c], 7510 ZPOOL_CONFIG_VDEV_TOP_ZAP, 7511 vml[c]->vdev_parent->vdev_top_zap)); 7512 } 7513 7514 if (error != 0) { 7515 kmem_free(vml, children * sizeof (vdev_t *)); 7516 kmem_free(glist, children * sizeof (uint64_t)); 7517 return (spa_vdev_exit(spa, NULL, txg, error)); 7518 } 7519 7520 /* stop writers from using the disks */ 7521 for (c = 0; c < children; c++) { 7522 if (vml[c] != NULL) 7523 vml[c]->vdev_offline = B_TRUE; 7524 } 7525 vdev_reopen(spa->spa_root_vdev); 7526 7527 /* 7528 * Temporarily record the splitting vdevs in the spa config. This 7529 * will disappear once the config is regenerated. 7530 */ 7531 VERIFY(nvlist_alloc(&nvl, NV_UNIQUE_NAME, KM_SLEEP) == 0); 7532 VERIFY(nvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, 7533 glist, children) == 0); 7534 kmem_free(glist, children * sizeof (uint64_t)); 7535 7536 mutex_enter(&spa->spa_props_lock); 7537 VERIFY(nvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT, 7538 nvl) == 0); 7539 mutex_exit(&spa->spa_props_lock); 7540 spa->spa_config_splitting = nvl; 7541 vdev_config_dirty(spa->spa_root_vdev); 7542 7543 /* configure and create the new pool */ 7544 VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname) == 0); 7545 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, 7546 exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE) == 0); 7547 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, 7548 spa_version(spa)) == 0); 7549 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG, 7550 spa->spa_config_txg) == 0); 7551 VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID, 7552 spa_generate_guid(NULL)) == 0); 7553 VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)); 7554 (void) nvlist_lookup_string(props, 7555 zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot); 7556 7557 /* add the new pool to the namespace */ 7558 newspa = spa_add(newname, config, altroot); 7559 newspa->spa_avz_action = AVZ_ACTION_REBUILD; 7560 newspa->spa_config_txg = spa->spa_config_txg; 7561 spa_set_log_state(newspa, SPA_LOG_CLEAR); 7562 7563 /* release the spa config lock, retaining the namespace lock */ 7564 spa_vdev_config_exit(spa, NULL, txg, 0, FTAG); 7565 7566 if (zio_injection_enabled) 7567 zio_handle_panic_injection(spa, FTAG, 1); 7568 7569 spa_activate(newspa, spa_mode_global); 7570 spa_async_suspend(newspa); 7571 7572 /* 7573 * Temporarily stop the initializing and TRIM activity. We set the 7574 * state to ACTIVE so that we know to resume initializing or TRIM 7575 * once the split has completed. 7576 */ 7577 list_t vd_initialize_list; 7578 list_create(&vd_initialize_list, sizeof (vdev_t), 7579 offsetof(vdev_t, vdev_initialize_node)); 7580 7581 list_t vd_trim_list; 7582 list_create(&vd_trim_list, sizeof (vdev_t), 7583 offsetof(vdev_t, vdev_trim_node)); 7584 7585 for (c = 0; c < children; c++) { 7586 if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) { 7587 mutex_enter(&vml[c]->vdev_initialize_lock); 7588 vdev_initialize_stop(vml[c], 7589 VDEV_INITIALIZE_ACTIVE, &vd_initialize_list); 7590 mutex_exit(&vml[c]->vdev_initialize_lock); 7591 7592 mutex_enter(&vml[c]->vdev_trim_lock); 7593 vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list); 7594 mutex_exit(&vml[c]->vdev_trim_lock); 7595 } 7596 } 7597 7598 vdev_initialize_stop_wait(spa, &vd_initialize_list); 7599 vdev_trim_stop_wait(spa, &vd_trim_list); 7600 7601 list_destroy(&vd_initialize_list); 7602 list_destroy(&vd_trim_list); 7603 7604 newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT; 7605 newspa->spa_is_splitting = B_TRUE; 7606 7607 /* create the new pool from the disks of the original pool */ 7608 error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE); 7609 if (error) 7610 goto out; 7611 7612 /* if that worked, generate a real config for the new pool */ 7613 if (newspa->spa_root_vdev != NULL) { 7614 VERIFY(nvlist_alloc(&newspa->spa_config_splitting, 7615 NV_UNIQUE_NAME, KM_SLEEP) == 0); 7616 VERIFY(nvlist_add_uint64(newspa->spa_config_splitting, 7617 ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa)) == 0); 7618 spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL, 7619 B_TRUE)); 7620 } 7621 7622 /* set the props */ 7623 if (props != NULL) { 7624 spa_configfile_set(newspa, props, B_FALSE); 7625 error = spa_prop_set(newspa, props); 7626 if (error) 7627 goto out; 7628 } 7629 7630 /* flush everything */ 7631 txg = spa_vdev_config_enter(newspa); 7632 vdev_config_dirty(newspa->spa_root_vdev); 7633 (void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG); 7634 7635 if (zio_injection_enabled) 7636 zio_handle_panic_injection(spa, FTAG, 2); 7637 7638 spa_async_resume(newspa); 7639 7640 /* finally, update the original pool's config */ 7641 txg = spa_vdev_config_enter(spa); 7642 tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir); 7643 error = dmu_tx_assign(tx, TXG_WAIT); 7644 if (error != 0) 7645 dmu_tx_abort(tx); 7646 for (c = 0; c < children; c++) { 7647 if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) { 7648 vdev_t *tvd = vml[c]->vdev_top; 7649 7650 /* 7651 * Need to be sure the detachable VDEV is not 7652 * on any *other* txg's DTL list to prevent it 7653 * from being accessed after it's freed. 7654 */ 7655 for (int t = 0; t < TXG_SIZE; t++) { 7656 (void) txg_list_remove_this( 7657 &tvd->vdev_dtl_list, vml[c], t); 7658 } 7659 7660 vdev_split(vml[c]); 7661 if (error == 0) 7662 spa_history_log_internal(spa, "detach", tx, 7663 "vdev=%s", vml[c]->vdev_path); 7664 7665 vdev_free(vml[c]); 7666 } 7667 } 7668 spa->spa_avz_action = AVZ_ACTION_REBUILD; 7669 vdev_config_dirty(spa->spa_root_vdev); 7670 spa->spa_config_splitting = NULL; 7671 nvlist_free(nvl); 7672 if (error == 0) 7673 dmu_tx_commit(tx); 7674 (void) spa_vdev_exit(spa, NULL, txg, 0); 7675 7676 if (zio_injection_enabled) 7677 zio_handle_panic_injection(spa, FTAG, 3); 7678 7679 /* split is complete; log a history record */ 7680 spa_history_log_internal(newspa, "split", NULL, 7681 "from pool %s", spa_name(spa)); 7682 7683 newspa->spa_is_splitting = B_FALSE; 7684 kmem_free(vml, children * sizeof (vdev_t *)); 7685 7686 /* if we're not going to mount the filesystems in userland, export */ 7687 if (exp) 7688 error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL, 7689 B_FALSE, B_FALSE); 7690 7691 return (error); 7692 7693 out: 7694 spa_unload(newspa); 7695 spa_deactivate(newspa); 7696 spa_remove(newspa); 7697 7698 txg = spa_vdev_config_enter(spa); 7699 7700 /* re-online all offlined disks */ 7701 for (c = 0; c < children; c++) { 7702 if (vml[c] != NULL) 7703 vml[c]->vdev_offline = B_FALSE; 7704 } 7705 7706 /* restart initializing or trimming disks as necessary */ 7707 spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART); 7708 spa_async_request(spa, SPA_ASYNC_TRIM_RESTART); 7709 spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART); 7710 7711 vdev_reopen(spa->spa_root_vdev); 7712 7713 nvlist_free(spa->spa_config_splitting); 7714 spa->spa_config_splitting = NULL; 7715 (void) spa_vdev_exit(spa, NULL, txg, error); 7716 7717 kmem_free(vml, children * sizeof (vdev_t *)); 7718 return (error); 7719 } 7720 7721 /* 7722 * Find any device that's done replacing, or a vdev marked 'unspare' that's 7723 * currently spared, so we can detach it. 7724 */ 7725 static vdev_t * 7726 spa_vdev_resilver_done_hunt(vdev_t *vd) 7727 { 7728 vdev_t *newvd, *oldvd; 7729 7730 for (int c = 0; c < vd->vdev_children; c++) { 7731 oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]); 7732 if (oldvd != NULL) 7733 return (oldvd); 7734 } 7735 7736 /* 7737 * Check for a completed replacement. We always consider the first 7738 * vdev in the list to be the oldest vdev, and the last one to be 7739 * the newest (see spa_vdev_attach() for how that works). In 7740 * the case where the newest vdev is faulted, we will not automatically 7741 * remove it after a resilver completes. This is OK as it will require 7742 * user intervention to determine which disk the admin wishes to keep. 7743 */ 7744 if (vd->vdev_ops == &vdev_replacing_ops) { 7745 ASSERT(vd->vdev_children > 1); 7746 7747 newvd = vd->vdev_child[vd->vdev_children - 1]; 7748 oldvd = vd->vdev_child[0]; 7749 7750 if (vdev_dtl_empty(newvd, DTL_MISSING) && 7751 vdev_dtl_empty(newvd, DTL_OUTAGE) && 7752 !vdev_dtl_required(oldvd)) 7753 return (oldvd); 7754 } 7755 7756 /* 7757 * Check for a completed resilver with the 'unspare' flag set. 7758 * Also potentially update faulted state. 7759 */ 7760 if (vd->vdev_ops == &vdev_spare_ops) { 7761 vdev_t *first = vd->vdev_child[0]; 7762 vdev_t *last = vd->vdev_child[vd->vdev_children - 1]; 7763 7764 if (last->vdev_unspare) { 7765 oldvd = first; 7766 newvd = last; 7767 } else if (first->vdev_unspare) { 7768 oldvd = last; 7769 newvd = first; 7770 } else { 7771 oldvd = NULL; 7772 } 7773 7774 if (oldvd != NULL && 7775 vdev_dtl_empty(newvd, DTL_MISSING) && 7776 vdev_dtl_empty(newvd, DTL_OUTAGE) && 7777 !vdev_dtl_required(oldvd)) 7778 return (oldvd); 7779 7780 vdev_propagate_state(vd); 7781 7782 /* 7783 * If there are more than two spares attached to a disk, 7784 * and those spares are not required, then we want to 7785 * attempt to free them up now so that they can be used 7786 * by other pools. Once we're back down to a single 7787 * disk+spare, we stop removing them. 7788 */ 7789 if (vd->vdev_children > 2) { 7790 newvd = vd->vdev_child[1]; 7791 7792 if (newvd->vdev_isspare && last->vdev_isspare && 7793 vdev_dtl_empty(last, DTL_MISSING) && 7794 vdev_dtl_empty(last, DTL_OUTAGE) && 7795 !vdev_dtl_required(newvd)) 7796 return (newvd); 7797 } 7798 } 7799 7800 return (NULL); 7801 } 7802 7803 static void 7804 spa_vdev_resilver_done(spa_t *spa) 7805 { 7806 vdev_t *vd, *pvd, *ppvd; 7807 uint64_t guid, sguid, pguid, ppguid; 7808 7809 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 7810 7811 while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) { 7812 pvd = vd->vdev_parent; 7813 ppvd = pvd->vdev_parent; 7814 guid = vd->vdev_guid; 7815 pguid = pvd->vdev_guid; 7816 ppguid = ppvd->vdev_guid; 7817 sguid = 0; 7818 /* 7819 * If we have just finished replacing a hot spared device, then 7820 * we need to detach the parent's first child (the original hot 7821 * spare) as well. 7822 */ 7823 if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 && 7824 ppvd->vdev_children == 2) { 7825 ASSERT(pvd->vdev_ops == &vdev_replacing_ops); 7826 sguid = ppvd->vdev_child[1]->vdev_guid; 7827 } 7828 ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd)); 7829 7830 spa_config_exit(spa, SCL_ALL, FTAG); 7831 if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0) 7832 return; 7833 if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0) 7834 return; 7835 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 7836 } 7837 7838 spa_config_exit(spa, SCL_ALL, FTAG); 7839 7840 /* 7841 * If a detach was not performed above replace waiters will not have 7842 * been notified. In which case we must do so now. 7843 */ 7844 spa_notify_waiters(spa); 7845 } 7846 7847 /* 7848 * Update the stored path or FRU for this vdev. 7849 */ 7850 static int 7851 spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value, 7852 boolean_t ispath) 7853 { 7854 vdev_t *vd; 7855 boolean_t sync = B_FALSE; 7856 7857 ASSERT(spa_writeable(spa)); 7858 7859 spa_vdev_state_enter(spa, SCL_ALL); 7860 7861 if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL) 7862 return (spa_vdev_state_exit(spa, NULL, ENOENT)); 7863 7864 if (!vd->vdev_ops->vdev_op_leaf) 7865 return (spa_vdev_state_exit(spa, NULL, ENOTSUP)); 7866 7867 if (ispath) { 7868 if (strcmp(value, vd->vdev_path) != 0) { 7869 spa_strfree(vd->vdev_path); 7870 vd->vdev_path = spa_strdup(value); 7871 sync = B_TRUE; 7872 } 7873 } else { 7874 if (vd->vdev_fru == NULL) { 7875 vd->vdev_fru = spa_strdup(value); 7876 sync = B_TRUE; 7877 } else if (strcmp(value, vd->vdev_fru) != 0) { 7878 spa_strfree(vd->vdev_fru); 7879 vd->vdev_fru = spa_strdup(value); 7880 sync = B_TRUE; 7881 } 7882 } 7883 7884 return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0)); 7885 } 7886 7887 int 7888 spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath) 7889 { 7890 return (spa_vdev_set_common(spa, guid, newpath, B_TRUE)); 7891 } 7892 7893 int 7894 spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru) 7895 { 7896 return (spa_vdev_set_common(spa, guid, newfru, B_FALSE)); 7897 } 7898 7899 /* 7900 * ========================================================================== 7901 * SPA Scanning 7902 * ========================================================================== 7903 */ 7904 int 7905 spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd) 7906 { 7907 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); 7908 7909 if (dsl_scan_resilvering(spa->spa_dsl_pool)) 7910 return (SET_ERROR(EBUSY)); 7911 7912 return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd)); 7913 } 7914 7915 int 7916 spa_scan_stop(spa_t *spa) 7917 { 7918 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); 7919 if (dsl_scan_resilvering(spa->spa_dsl_pool)) 7920 return (SET_ERROR(EBUSY)); 7921 return (dsl_scan_cancel(spa->spa_dsl_pool)); 7922 } 7923 7924 int 7925 spa_scan(spa_t *spa, pool_scan_func_t func) 7926 { 7927 ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0); 7928 7929 if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE) 7930 return (SET_ERROR(ENOTSUP)); 7931 7932 if (func == POOL_SCAN_RESILVER && 7933 !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) 7934 return (SET_ERROR(ENOTSUP)); 7935 7936 /* 7937 * If a resilver was requested, but there is no DTL on a 7938 * writeable leaf device, we have nothing to do. 7939 */ 7940 if (func == POOL_SCAN_RESILVER && 7941 !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) { 7942 spa_async_request(spa, SPA_ASYNC_RESILVER_DONE); 7943 return (0); 7944 } 7945 7946 return (dsl_scan(spa->spa_dsl_pool, func)); 7947 } 7948 7949 /* 7950 * ========================================================================== 7951 * SPA async task processing 7952 * ========================================================================== 7953 */ 7954 7955 static void 7956 spa_async_remove(spa_t *spa, vdev_t *vd) 7957 { 7958 if (vd->vdev_remove_wanted) { 7959 vd->vdev_remove_wanted = B_FALSE; 7960 vd->vdev_delayed_close = B_FALSE; 7961 vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE); 7962 7963 /* 7964 * We want to clear the stats, but we don't want to do a full 7965 * vdev_clear() as that will cause us to throw away 7966 * degraded/faulted state as well as attempt to reopen the 7967 * device, all of which is a waste. 7968 */ 7969 vd->vdev_stat.vs_read_errors = 0; 7970 vd->vdev_stat.vs_write_errors = 0; 7971 vd->vdev_stat.vs_checksum_errors = 0; 7972 7973 vdev_state_dirty(vd->vdev_top); 7974 7975 /* Tell userspace that the vdev is gone. */ 7976 zfs_post_remove(spa, vd); 7977 } 7978 7979 for (int c = 0; c < vd->vdev_children; c++) 7980 spa_async_remove(spa, vd->vdev_child[c]); 7981 } 7982 7983 static void 7984 spa_async_probe(spa_t *spa, vdev_t *vd) 7985 { 7986 if (vd->vdev_probe_wanted) { 7987 vd->vdev_probe_wanted = B_FALSE; 7988 vdev_reopen(vd); /* vdev_open() does the actual probe */ 7989 } 7990 7991 for (int c = 0; c < vd->vdev_children; c++) 7992 spa_async_probe(spa, vd->vdev_child[c]); 7993 } 7994 7995 static void 7996 spa_async_autoexpand(spa_t *spa, vdev_t *vd) 7997 { 7998 if (!spa->spa_autoexpand) 7999 return; 8000 8001 for (int c = 0; c < vd->vdev_children; c++) { 8002 vdev_t *cvd = vd->vdev_child[c]; 8003 spa_async_autoexpand(spa, cvd); 8004 } 8005 8006 if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL) 8007 return; 8008 8009 spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_AUTOEXPAND); 8010 } 8011 8012 static void 8013 spa_async_thread(void *arg) 8014 { 8015 spa_t *spa = (spa_t *)arg; 8016 dsl_pool_t *dp = spa->spa_dsl_pool; 8017 int tasks; 8018 8019 ASSERT(spa->spa_sync_on); 8020 8021 mutex_enter(&spa->spa_async_lock); 8022 tasks = spa->spa_async_tasks; 8023 spa->spa_async_tasks = 0; 8024 mutex_exit(&spa->spa_async_lock); 8025 8026 /* 8027 * See if the config needs to be updated. 8028 */ 8029 if (tasks & SPA_ASYNC_CONFIG_UPDATE) { 8030 uint64_t old_space, new_space; 8031 8032 mutex_enter(&spa_namespace_lock); 8033 old_space = metaslab_class_get_space(spa_normal_class(spa)); 8034 old_space += metaslab_class_get_space(spa_special_class(spa)); 8035 old_space += metaslab_class_get_space(spa_dedup_class(spa)); 8036 8037 spa_config_update(spa, SPA_CONFIG_UPDATE_POOL); 8038 8039 new_space = metaslab_class_get_space(spa_normal_class(spa)); 8040 new_space += metaslab_class_get_space(spa_special_class(spa)); 8041 new_space += metaslab_class_get_space(spa_dedup_class(spa)); 8042 mutex_exit(&spa_namespace_lock); 8043 8044 /* 8045 * If the pool grew as a result of the config update, 8046 * then log an internal history event. 8047 */ 8048 if (new_space != old_space) { 8049 spa_history_log_internal(spa, "vdev online", NULL, 8050 "pool '%s' size: %llu(+%llu)", 8051 spa_name(spa), (u_longlong_t)new_space, 8052 (u_longlong_t)(new_space - old_space)); 8053 } 8054 } 8055 8056 /* 8057 * See if any devices need to be marked REMOVED. 8058 */ 8059 if (tasks & SPA_ASYNC_REMOVE) { 8060 spa_vdev_state_enter(spa, SCL_NONE); 8061 spa_async_remove(spa, spa->spa_root_vdev); 8062 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) 8063 spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]); 8064 for (int i = 0; i < spa->spa_spares.sav_count; i++) 8065 spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]); 8066 (void) spa_vdev_state_exit(spa, NULL, 0); 8067 } 8068 8069 if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) { 8070 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8071 spa_async_autoexpand(spa, spa->spa_root_vdev); 8072 spa_config_exit(spa, SCL_CONFIG, FTAG); 8073 } 8074 8075 /* 8076 * See if any devices need to be probed. 8077 */ 8078 if (tasks & SPA_ASYNC_PROBE) { 8079 spa_vdev_state_enter(spa, SCL_NONE); 8080 spa_async_probe(spa, spa->spa_root_vdev); 8081 (void) spa_vdev_state_exit(spa, NULL, 0); 8082 } 8083 8084 /* 8085 * If any devices are done replacing, detach them. 8086 */ 8087 if (tasks & SPA_ASYNC_RESILVER_DONE || 8088 tasks & SPA_ASYNC_REBUILD_DONE) { 8089 spa_vdev_resilver_done(spa); 8090 } 8091 8092 /* 8093 * Kick off a resilver. 8094 */ 8095 if (tasks & SPA_ASYNC_RESILVER && 8096 !vdev_rebuild_active(spa->spa_root_vdev) && 8097 (!dsl_scan_resilvering(dp) || 8098 !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER))) 8099 dsl_scan_restart_resilver(dp, 0); 8100 8101 if (tasks & SPA_ASYNC_INITIALIZE_RESTART) { 8102 mutex_enter(&spa_namespace_lock); 8103 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8104 vdev_initialize_restart(spa->spa_root_vdev); 8105 spa_config_exit(spa, SCL_CONFIG, FTAG); 8106 mutex_exit(&spa_namespace_lock); 8107 } 8108 8109 if (tasks & SPA_ASYNC_TRIM_RESTART) { 8110 mutex_enter(&spa_namespace_lock); 8111 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8112 vdev_trim_restart(spa->spa_root_vdev); 8113 spa_config_exit(spa, SCL_CONFIG, FTAG); 8114 mutex_exit(&spa_namespace_lock); 8115 } 8116 8117 if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) { 8118 mutex_enter(&spa_namespace_lock); 8119 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8120 vdev_autotrim_restart(spa); 8121 spa_config_exit(spa, SCL_CONFIG, FTAG); 8122 mutex_exit(&spa_namespace_lock); 8123 } 8124 8125 /* 8126 * Kick off L2 cache whole device TRIM. 8127 */ 8128 if (tasks & SPA_ASYNC_L2CACHE_TRIM) { 8129 mutex_enter(&spa_namespace_lock); 8130 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 8131 vdev_trim_l2arc(spa); 8132 spa_config_exit(spa, SCL_CONFIG, FTAG); 8133 mutex_exit(&spa_namespace_lock); 8134 } 8135 8136 /* 8137 * Kick off L2 cache rebuilding. 8138 */ 8139 if (tasks & SPA_ASYNC_L2CACHE_REBUILD) { 8140 mutex_enter(&spa_namespace_lock); 8141 spa_config_enter(spa, SCL_L2ARC, FTAG, RW_READER); 8142 l2arc_spa_rebuild_start(spa); 8143 spa_config_exit(spa, SCL_L2ARC, FTAG); 8144 mutex_exit(&spa_namespace_lock); 8145 } 8146 8147 /* 8148 * Let the world know that we're done. 8149 */ 8150 mutex_enter(&spa->spa_async_lock); 8151 spa->spa_async_thread = NULL; 8152 cv_broadcast(&spa->spa_async_cv); 8153 mutex_exit(&spa->spa_async_lock); 8154 thread_exit(); 8155 } 8156 8157 void 8158 spa_async_suspend(spa_t *spa) 8159 { 8160 mutex_enter(&spa->spa_async_lock); 8161 spa->spa_async_suspended++; 8162 while (spa->spa_async_thread != NULL) 8163 cv_wait(&spa->spa_async_cv, &spa->spa_async_lock); 8164 mutex_exit(&spa->spa_async_lock); 8165 8166 spa_vdev_remove_suspend(spa); 8167 8168 zthr_t *condense_thread = spa->spa_condense_zthr; 8169 if (condense_thread != NULL) 8170 zthr_cancel(condense_thread); 8171 8172 zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr; 8173 if (discard_thread != NULL) 8174 zthr_cancel(discard_thread); 8175 8176 zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr; 8177 if (ll_delete_thread != NULL) 8178 zthr_cancel(ll_delete_thread); 8179 8180 zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr; 8181 if (ll_condense_thread != NULL) 8182 zthr_cancel(ll_condense_thread); 8183 } 8184 8185 void 8186 spa_async_resume(spa_t *spa) 8187 { 8188 mutex_enter(&spa->spa_async_lock); 8189 ASSERT(spa->spa_async_suspended != 0); 8190 spa->spa_async_suspended--; 8191 mutex_exit(&spa->spa_async_lock); 8192 spa_restart_removal(spa); 8193 8194 zthr_t *condense_thread = spa->spa_condense_zthr; 8195 if (condense_thread != NULL) 8196 zthr_resume(condense_thread); 8197 8198 zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr; 8199 if (discard_thread != NULL) 8200 zthr_resume(discard_thread); 8201 8202 zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr; 8203 if (ll_delete_thread != NULL) 8204 zthr_resume(ll_delete_thread); 8205 8206 zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr; 8207 if (ll_condense_thread != NULL) 8208 zthr_resume(ll_condense_thread); 8209 } 8210 8211 static boolean_t 8212 spa_async_tasks_pending(spa_t *spa) 8213 { 8214 uint_t non_config_tasks; 8215 uint_t config_task; 8216 boolean_t config_task_suspended; 8217 8218 non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE; 8219 config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE; 8220 if (spa->spa_ccw_fail_time == 0) { 8221 config_task_suspended = B_FALSE; 8222 } else { 8223 config_task_suspended = 8224 (gethrtime() - spa->spa_ccw_fail_time) < 8225 ((hrtime_t)zfs_ccw_retry_interval * NANOSEC); 8226 } 8227 8228 return (non_config_tasks || (config_task && !config_task_suspended)); 8229 } 8230 8231 static void 8232 spa_async_dispatch(spa_t *spa) 8233 { 8234 mutex_enter(&spa->spa_async_lock); 8235 if (spa_async_tasks_pending(spa) && 8236 !spa->spa_async_suspended && 8237 spa->spa_async_thread == NULL) 8238 spa->spa_async_thread = thread_create(NULL, 0, 8239 spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri); 8240 mutex_exit(&spa->spa_async_lock); 8241 } 8242 8243 void 8244 spa_async_request(spa_t *spa, int task) 8245 { 8246 zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task); 8247 mutex_enter(&spa->spa_async_lock); 8248 spa->spa_async_tasks |= task; 8249 mutex_exit(&spa->spa_async_lock); 8250 } 8251 8252 int 8253 spa_async_tasks(spa_t *spa) 8254 { 8255 return (spa->spa_async_tasks); 8256 } 8257 8258 /* 8259 * ========================================================================== 8260 * SPA syncing routines 8261 * ========================================================================== 8262 */ 8263 8264 8265 static int 8266 bpobj_enqueue_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 8267 dmu_tx_t *tx) 8268 { 8269 bpobj_t *bpo = arg; 8270 bpobj_enqueue(bpo, bp, bp_freed, tx); 8271 return (0); 8272 } 8273 8274 int 8275 bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 8276 { 8277 return (bpobj_enqueue_cb(arg, bp, B_FALSE, tx)); 8278 } 8279 8280 int 8281 bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 8282 { 8283 return (bpobj_enqueue_cb(arg, bp, B_TRUE, tx)); 8284 } 8285 8286 static int 8287 spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx) 8288 { 8289 zio_t *pio = arg; 8290 8291 zio_nowait(zio_free_sync(pio, pio->io_spa, dmu_tx_get_txg(tx), bp, 8292 pio->io_flags)); 8293 return (0); 8294 } 8295 8296 static int 8297 bpobj_spa_free_sync_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed, 8298 dmu_tx_t *tx) 8299 { 8300 ASSERT(!bp_freed); 8301 return (spa_free_sync_cb(arg, bp, tx)); 8302 } 8303 8304 /* 8305 * Note: this simple function is not inlined to make it easier to dtrace the 8306 * amount of time spent syncing frees. 8307 */ 8308 static void 8309 spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx) 8310 { 8311 zio_t *zio = zio_root(spa, NULL, NULL, 0); 8312 bplist_iterate(bpl, spa_free_sync_cb, zio, tx); 8313 VERIFY(zio_wait(zio) == 0); 8314 } 8315 8316 /* 8317 * Note: this simple function is not inlined to make it easier to dtrace the 8318 * amount of time spent syncing deferred frees. 8319 */ 8320 static void 8321 spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx) 8322 { 8323 if (spa_sync_pass(spa) != 1) 8324 return; 8325 8326 /* 8327 * Note: 8328 * If the log space map feature is active, we stop deferring 8329 * frees to the next TXG and therefore running this function 8330 * would be considered a no-op as spa_deferred_bpobj should 8331 * not have any entries. 8332 * 8333 * That said we run this function anyway (instead of returning 8334 * immediately) for the edge-case scenario where we just 8335 * activated the log space map feature in this TXG but we have 8336 * deferred frees from the previous TXG. 8337 */ 8338 zio_t *zio = zio_root(spa, NULL, NULL, 0); 8339 VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj, 8340 bpobj_spa_free_sync_cb, zio, tx), ==, 0); 8341 VERIFY0(zio_wait(zio)); 8342 } 8343 8344 static void 8345 spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx) 8346 { 8347 char *packed = NULL; 8348 size_t bufsize; 8349 size_t nvsize = 0; 8350 dmu_buf_t *db; 8351 8352 VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0); 8353 8354 /* 8355 * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration 8356 * information. This avoids the dmu_buf_will_dirty() path and 8357 * saves us a pre-read to get data we don't actually care about. 8358 */ 8359 bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE); 8360 packed = vmem_alloc(bufsize, KM_SLEEP); 8361 8362 VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR, 8363 KM_SLEEP) == 0); 8364 bzero(packed + nvsize, bufsize - nvsize); 8365 8366 dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx); 8367 8368 vmem_free(packed, bufsize); 8369 8370 VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db)); 8371 dmu_buf_will_dirty(db, tx); 8372 *(uint64_t *)db->db_data = nvsize; 8373 dmu_buf_rele(db, FTAG); 8374 } 8375 8376 static void 8377 spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx, 8378 const char *config, const char *entry) 8379 { 8380 nvlist_t *nvroot; 8381 nvlist_t **list; 8382 int i; 8383 8384 if (!sav->sav_sync) 8385 return; 8386 8387 /* 8388 * Update the MOS nvlist describing the list of available devices. 8389 * spa_validate_aux() will have already made sure this nvlist is 8390 * valid and the vdevs are labeled appropriately. 8391 */ 8392 if (sav->sav_object == 0) { 8393 sav->sav_object = dmu_object_alloc(spa->spa_meta_objset, 8394 DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE, 8395 sizeof (uint64_t), tx); 8396 VERIFY(zap_update(spa->spa_meta_objset, 8397 DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1, 8398 &sav->sav_object, tx) == 0); 8399 } 8400 8401 VERIFY(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, KM_SLEEP) == 0); 8402 if (sav->sav_count == 0) { 8403 VERIFY(nvlist_add_nvlist_array(nvroot, config, NULL, 0) == 0); 8404 } else { 8405 list = kmem_alloc(sav->sav_count*sizeof (void *), KM_SLEEP); 8406 for (i = 0; i < sav->sav_count; i++) 8407 list[i] = vdev_config_generate(spa, sav->sav_vdevs[i], 8408 B_FALSE, VDEV_CONFIG_L2CACHE); 8409 VERIFY(nvlist_add_nvlist_array(nvroot, config, list, 8410 sav->sav_count) == 0); 8411 for (i = 0; i < sav->sav_count; i++) 8412 nvlist_free(list[i]); 8413 kmem_free(list, sav->sav_count * sizeof (void *)); 8414 } 8415 8416 spa_sync_nvlist(spa, sav->sav_object, nvroot, tx); 8417 nvlist_free(nvroot); 8418 8419 sav->sav_sync = B_FALSE; 8420 } 8421 8422 /* 8423 * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t. 8424 * The all-vdev ZAP must be empty. 8425 */ 8426 static void 8427 spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx) 8428 { 8429 spa_t *spa = vd->vdev_spa; 8430 8431 if (vd->vdev_top_zap != 0) { 8432 VERIFY0(zap_add_int(spa->spa_meta_objset, avz, 8433 vd->vdev_top_zap, tx)); 8434 } 8435 if (vd->vdev_leaf_zap != 0) { 8436 VERIFY0(zap_add_int(spa->spa_meta_objset, avz, 8437 vd->vdev_leaf_zap, tx)); 8438 } 8439 for (uint64_t i = 0; i < vd->vdev_children; i++) { 8440 spa_avz_build(vd->vdev_child[i], avz, tx); 8441 } 8442 } 8443 8444 static void 8445 spa_sync_config_object(spa_t *spa, dmu_tx_t *tx) 8446 { 8447 nvlist_t *config; 8448 8449 /* 8450 * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS, 8451 * its config may not be dirty but we still need to build per-vdev ZAPs. 8452 * Similarly, if the pool is being assembled (e.g. after a split), we 8453 * need to rebuild the AVZ although the config may not be dirty. 8454 */ 8455 if (list_is_empty(&spa->spa_config_dirty_list) && 8456 spa->spa_avz_action == AVZ_ACTION_NONE) 8457 return; 8458 8459 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 8460 8461 ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE || 8462 spa->spa_avz_action == AVZ_ACTION_INITIALIZE || 8463 spa->spa_all_vdev_zaps != 0); 8464 8465 if (spa->spa_avz_action == AVZ_ACTION_REBUILD) { 8466 /* Make and build the new AVZ */ 8467 uint64_t new_avz = zap_create(spa->spa_meta_objset, 8468 DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx); 8469 spa_avz_build(spa->spa_root_vdev, new_avz, tx); 8470 8471 /* Diff old AVZ with new one */ 8472 zap_cursor_t zc; 8473 zap_attribute_t za; 8474 8475 for (zap_cursor_init(&zc, spa->spa_meta_objset, 8476 spa->spa_all_vdev_zaps); 8477 zap_cursor_retrieve(&zc, &za) == 0; 8478 zap_cursor_advance(&zc)) { 8479 uint64_t vdzap = za.za_first_integer; 8480 if (zap_lookup_int(spa->spa_meta_objset, new_avz, 8481 vdzap) == ENOENT) { 8482 /* 8483 * ZAP is listed in old AVZ but not in new one; 8484 * destroy it 8485 */ 8486 VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap, 8487 tx)); 8488 } 8489 } 8490 8491 zap_cursor_fini(&zc); 8492 8493 /* Destroy the old AVZ */ 8494 VERIFY0(zap_destroy(spa->spa_meta_objset, 8495 spa->spa_all_vdev_zaps, tx)); 8496 8497 /* Replace the old AVZ in the dir obj with the new one */ 8498 VERIFY0(zap_update(spa->spa_meta_objset, 8499 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, 8500 sizeof (new_avz), 1, &new_avz, tx)); 8501 8502 spa->spa_all_vdev_zaps = new_avz; 8503 } else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) { 8504 zap_cursor_t zc; 8505 zap_attribute_t za; 8506 8507 /* Walk through the AVZ and destroy all listed ZAPs */ 8508 for (zap_cursor_init(&zc, spa->spa_meta_objset, 8509 spa->spa_all_vdev_zaps); 8510 zap_cursor_retrieve(&zc, &za) == 0; 8511 zap_cursor_advance(&zc)) { 8512 uint64_t zap = za.za_first_integer; 8513 VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx)); 8514 } 8515 8516 zap_cursor_fini(&zc); 8517 8518 /* Destroy and unlink the AVZ itself */ 8519 VERIFY0(zap_destroy(spa->spa_meta_objset, 8520 spa->spa_all_vdev_zaps, tx)); 8521 VERIFY0(zap_remove(spa->spa_meta_objset, 8522 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx)); 8523 spa->spa_all_vdev_zaps = 0; 8524 } 8525 8526 if (spa->spa_all_vdev_zaps == 0) { 8527 spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset, 8528 DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT, 8529 DMU_POOL_VDEV_ZAP_MAP, tx); 8530 } 8531 spa->spa_avz_action = AVZ_ACTION_NONE; 8532 8533 /* Create ZAPs for vdevs that don't have them. */ 8534 vdev_construct_zaps(spa->spa_root_vdev, tx); 8535 8536 config = spa_config_generate(spa, spa->spa_root_vdev, 8537 dmu_tx_get_txg(tx), B_FALSE); 8538 8539 /* 8540 * If we're upgrading the spa version then make sure that 8541 * the config object gets updated with the correct version. 8542 */ 8543 if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version) 8544 fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, 8545 spa->spa_uberblock.ub_version); 8546 8547 spa_config_exit(spa, SCL_STATE, FTAG); 8548 8549 nvlist_free(spa->spa_config_syncing); 8550 spa->spa_config_syncing = config; 8551 8552 spa_sync_nvlist(spa, spa->spa_config_object, config, tx); 8553 } 8554 8555 static void 8556 spa_sync_version(void *arg, dmu_tx_t *tx) 8557 { 8558 uint64_t *versionp = arg; 8559 uint64_t version = *versionp; 8560 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 8561 8562 /* 8563 * Setting the version is special cased when first creating the pool. 8564 */ 8565 ASSERT(tx->tx_txg != TXG_INITIAL); 8566 8567 ASSERT(SPA_VERSION_IS_SUPPORTED(version)); 8568 ASSERT(version >= spa_version(spa)); 8569 8570 spa->spa_uberblock.ub_version = version; 8571 vdev_config_dirty(spa->spa_root_vdev); 8572 spa_history_log_internal(spa, "set", tx, "version=%lld", 8573 (longlong_t)version); 8574 } 8575 8576 /* 8577 * Set zpool properties. 8578 */ 8579 static void 8580 spa_sync_props(void *arg, dmu_tx_t *tx) 8581 { 8582 nvlist_t *nvp = arg; 8583 spa_t *spa = dmu_tx_pool(tx)->dp_spa; 8584 objset_t *mos = spa->spa_meta_objset; 8585 nvpair_t *elem = NULL; 8586 8587 mutex_enter(&spa->spa_props_lock); 8588 8589 while ((elem = nvlist_next_nvpair(nvp, elem))) { 8590 uint64_t intval; 8591 char *strval, *fname; 8592 zpool_prop_t prop; 8593 const char *propname; 8594 zprop_type_t proptype; 8595 spa_feature_t fid; 8596 8597 switch (prop = zpool_name_to_prop(nvpair_name(elem))) { 8598 case ZPOOL_PROP_INVAL: 8599 /* 8600 * We checked this earlier in spa_prop_validate(). 8601 */ 8602 ASSERT(zpool_prop_feature(nvpair_name(elem))); 8603 8604 fname = strchr(nvpair_name(elem), '@') + 1; 8605 VERIFY0(zfeature_lookup_name(fname, &fid)); 8606 8607 spa_feature_enable(spa, fid, tx); 8608 spa_history_log_internal(spa, "set", tx, 8609 "%s=enabled", nvpair_name(elem)); 8610 break; 8611 8612 case ZPOOL_PROP_VERSION: 8613 intval = fnvpair_value_uint64(elem); 8614 /* 8615 * The version is synced separately before other 8616 * properties and should be correct by now. 8617 */ 8618 ASSERT3U(spa_version(spa), >=, intval); 8619 break; 8620 8621 case ZPOOL_PROP_ALTROOT: 8622 /* 8623 * 'altroot' is a non-persistent property. It should 8624 * have been set temporarily at creation or import time. 8625 */ 8626 ASSERT(spa->spa_root != NULL); 8627 break; 8628 8629 case ZPOOL_PROP_READONLY: 8630 case ZPOOL_PROP_CACHEFILE: 8631 /* 8632 * 'readonly' and 'cachefile' are also non-persistent 8633 * properties. 8634 */ 8635 break; 8636 case ZPOOL_PROP_COMMENT: 8637 strval = fnvpair_value_string(elem); 8638 if (spa->spa_comment != NULL) 8639 spa_strfree(spa->spa_comment); 8640 spa->spa_comment = spa_strdup(strval); 8641 /* 8642 * We need to dirty the configuration on all the vdevs 8643 * so that their labels get updated. It's unnecessary 8644 * to do this for pool creation since the vdev's 8645 * configuration has already been dirtied. 8646 */ 8647 if (tx->tx_txg != TXG_INITIAL) 8648 vdev_config_dirty(spa->spa_root_vdev); 8649 spa_history_log_internal(spa, "set", tx, 8650 "%s=%s", nvpair_name(elem), strval); 8651 break; 8652 default: 8653 /* 8654 * Set pool property values in the poolprops mos object. 8655 */ 8656 if (spa->spa_pool_props_object == 0) { 8657 spa->spa_pool_props_object = 8658 zap_create_link(mos, DMU_OT_POOL_PROPS, 8659 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS, 8660 tx); 8661 } 8662 8663 /* normalize the property name */ 8664 propname = zpool_prop_to_name(prop); 8665 proptype = zpool_prop_get_type(prop); 8666 8667 if (nvpair_type(elem) == DATA_TYPE_STRING) { 8668 ASSERT(proptype == PROP_TYPE_STRING); 8669 strval = fnvpair_value_string(elem); 8670 VERIFY0(zap_update(mos, 8671 spa->spa_pool_props_object, propname, 8672 1, strlen(strval) + 1, strval, tx)); 8673 spa_history_log_internal(spa, "set", tx, 8674 "%s=%s", nvpair_name(elem), strval); 8675 } else if (nvpair_type(elem) == DATA_TYPE_UINT64) { 8676 intval = fnvpair_value_uint64(elem); 8677 8678 if (proptype == PROP_TYPE_INDEX) { 8679 const char *unused; 8680 VERIFY0(zpool_prop_index_to_string( 8681 prop, intval, &unused)); 8682 } 8683 VERIFY0(zap_update(mos, 8684 spa->spa_pool_props_object, propname, 8685 8, 1, &intval, tx)); 8686 spa_history_log_internal(spa, "set", tx, 8687 "%s=%lld", nvpair_name(elem), 8688 (longlong_t)intval); 8689 } else { 8690 ASSERT(0); /* not allowed */ 8691 } 8692 8693 switch (prop) { 8694 case ZPOOL_PROP_DELEGATION: 8695 spa->spa_delegation = intval; 8696 break; 8697 case ZPOOL_PROP_BOOTFS: 8698 spa->spa_bootfs = intval; 8699 break; 8700 case ZPOOL_PROP_FAILUREMODE: 8701 spa->spa_failmode = intval; 8702 break; 8703 case ZPOOL_PROP_AUTOTRIM: 8704 spa->spa_autotrim = intval; 8705 spa_async_request(spa, 8706 SPA_ASYNC_AUTOTRIM_RESTART); 8707 break; 8708 case ZPOOL_PROP_AUTOEXPAND: 8709 spa->spa_autoexpand = intval; 8710 if (tx->tx_txg != TXG_INITIAL) 8711 spa_async_request(spa, 8712 SPA_ASYNC_AUTOEXPAND); 8713 break; 8714 case ZPOOL_PROP_MULTIHOST: 8715 spa->spa_multihost = intval; 8716 break; 8717 default: 8718 break; 8719 } 8720 } 8721 8722 } 8723 8724 mutex_exit(&spa->spa_props_lock); 8725 } 8726 8727 /* 8728 * Perform one-time upgrade on-disk changes. spa_version() does not 8729 * reflect the new version this txg, so there must be no changes this 8730 * txg to anything that the upgrade code depends on after it executes. 8731 * Therefore this must be called after dsl_pool_sync() does the sync 8732 * tasks. 8733 */ 8734 static void 8735 spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx) 8736 { 8737 if (spa_sync_pass(spa) != 1) 8738 return; 8739 8740 dsl_pool_t *dp = spa->spa_dsl_pool; 8741 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG); 8742 8743 if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN && 8744 spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) { 8745 dsl_pool_create_origin(dp, tx); 8746 8747 /* Keeping the origin open increases spa_minref */ 8748 spa->spa_minref += 3; 8749 } 8750 8751 if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES && 8752 spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) { 8753 dsl_pool_upgrade_clones(dp, tx); 8754 } 8755 8756 if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES && 8757 spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) { 8758 dsl_pool_upgrade_dir_clones(dp, tx); 8759 8760 /* Keeping the freedir open increases spa_minref */ 8761 spa->spa_minref += 3; 8762 } 8763 8764 if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES && 8765 spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) { 8766 spa_feature_create_zap_objects(spa, tx); 8767 } 8768 8769 /* 8770 * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable 8771 * when possibility to use lz4 compression for metadata was added 8772 * Old pools that have this feature enabled must be upgraded to have 8773 * this feature active 8774 */ 8775 if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) { 8776 boolean_t lz4_en = spa_feature_is_enabled(spa, 8777 SPA_FEATURE_LZ4_COMPRESS); 8778 boolean_t lz4_ac = spa_feature_is_active(spa, 8779 SPA_FEATURE_LZ4_COMPRESS); 8780 8781 if (lz4_en && !lz4_ac) 8782 spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx); 8783 } 8784 8785 /* 8786 * If we haven't written the salt, do so now. Note that the 8787 * feature may not be activated yet, but that's fine since 8788 * the presence of this ZAP entry is backwards compatible. 8789 */ 8790 if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT, 8791 DMU_POOL_CHECKSUM_SALT) == ENOENT) { 8792 VERIFY0(zap_add(spa->spa_meta_objset, 8793 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1, 8794 sizeof (spa->spa_cksum_salt.zcs_bytes), 8795 spa->spa_cksum_salt.zcs_bytes, tx)); 8796 } 8797 8798 rrw_exit(&dp->dp_config_rwlock, FTAG); 8799 } 8800 8801 static void 8802 vdev_indirect_state_sync_verify(vdev_t *vd) 8803 { 8804 vdev_indirect_mapping_t *vim __maybe_unused = vd->vdev_indirect_mapping; 8805 vdev_indirect_births_t *vib __maybe_unused = vd->vdev_indirect_births; 8806 8807 if (vd->vdev_ops == &vdev_indirect_ops) { 8808 ASSERT(vim != NULL); 8809 ASSERT(vib != NULL); 8810 } 8811 8812 uint64_t obsolete_sm_object = 0; 8813 ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object)); 8814 if (obsolete_sm_object != 0) { 8815 ASSERT(vd->vdev_obsolete_sm != NULL); 8816 ASSERT(vd->vdev_removing || 8817 vd->vdev_ops == &vdev_indirect_ops); 8818 ASSERT(vdev_indirect_mapping_num_entries(vim) > 0); 8819 ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0); 8820 ASSERT3U(obsolete_sm_object, ==, 8821 space_map_object(vd->vdev_obsolete_sm)); 8822 ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=, 8823 space_map_allocated(vd->vdev_obsolete_sm)); 8824 } 8825 ASSERT(vd->vdev_obsolete_segments != NULL); 8826 8827 /* 8828 * Since frees / remaps to an indirect vdev can only 8829 * happen in syncing context, the obsolete segments 8830 * tree must be empty when we start syncing. 8831 */ 8832 ASSERT0(range_tree_space(vd->vdev_obsolete_segments)); 8833 } 8834 8835 /* 8836 * Set the top-level vdev's max queue depth. Evaluate each top-level's 8837 * async write queue depth in case it changed. The max queue depth will 8838 * not change in the middle of syncing out this txg. 8839 */ 8840 static void 8841 spa_sync_adjust_vdev_max_queue_depth(spa_t *spa) 8842 { 8843 ASSERT(spa_writeable(spa)); 8844 8845 vdev_t *rvd = spa->spa_root_vdev; 8846 uint32_t max_queue_depth = zfs_vdev_async_write_max_active * 8847 zfs_vdev_queue_depth_pct / 100; 8848 metaslab_class_t *normal = spa_normal_class(spa); 8849 metaslab_class_t *special = spa_special_class(spa); 8850 metaslab_class_t *dedup = spa_dedup_class(spa); 8851 8852 uint64_t slots_per_allocator = 0; 8853 for (int c = 0; c < rvd->vdev_children; c++) { 8854 vdev_t *tvd = rvd->vdev_child[c]; 8855 8856 metaslab_group_t *mg = tvd->vdev_mg; 8857 if (mg == NULL || !metaslab_group_initialized(mg)) 8858 continue; 8859 8860 metaslab_class_t *mc = mg->mg_class; 8861 if (mc != normal && mc != special && mc != dedup) 8862 continue; 8863 8864 /* 8865 * It is safe to do a lock-free check here because only async 8866 * allocations look at mg_max_alloc_queue_depth, and async 8867 * allocations all happen from spa_sync(). 8868 */ 8869 for (int i = 0; i < mg->mg_allocators; i++) { 8870 ASSERT0(zfs_refcount_count( 8871 &(mg->mg_allocator[i].mga_alloc_queue_depth))); 8872 } 8873 mg->mg_max_alloc_queue_depth = max_queue_depth; 8874 8875 for (int i = 0; i < mg->mg_allocators; i++) { 8876 mg->mg_allocator[i].mga_cur_max_alloc_queue_depth = 8877 zfs_vdev_def_queue_depth; 8878 } 8879 slots_per_allocator += zfs_vdev_def_queue_depth; 8880 } 8881 8882 for (int i = 0; i < spa->spa_alloc_count; i++) { 8883 ASSERT0(zfs_refcount_count(&normal->mc_allocator[i]. 8884 mca_alloc_slots)); 8885 ASSERT0(zfs_refcount_count(&special->mc_allocator[i]. 8886 mca_alloc_slots)); 8887 ASSERT0(zfs_refcount_count(&dedup->mc_allocator[i]. 8888 mca_alloc_slots)); 8889 normal->mc_allocator[i].mca_alloc_max_slots = 8890 slots_per_allocator; 8891 special->mc_allocator[i].mca_alloc_max_slots = 8892 slots_per_allocator; 8893 dedup->mc_allocator[i].mca_alloc_max_slots = 8894 slots_per_allocator; 8895 } 8896 normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; 8897 special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; 8898 dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled; 8899 } 8900 8901 static void 8902 spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx) 8903 { 8904 ASSERT(spa_writeable(spa)); 8905 8906 vdev_t *rvd = spa->spa_root_vdev; 8907 for (int c = 0; c < rvd->vdev_children; c++) { 8908 vdev_t *vd = rvd->vdev_child[c]; 8909 vdev_indirect_state_sync_verify(vd); 8910 8911 if (vdev_indirect_should_condense(vd)) { 8912 spa_condense_indirect_start_sync(vd, tx); 8913 break; 8914 } 8915 } 8916 } 8917 8918 static void 8919 spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx) 8920 { 8921 objset_t *mos = spa->spa_meta_objset; 8922 dsl_pool_t *dp = spa->spa_dsl_pool; 8923 uint64_t txg = tx->tx_txg; 8924 bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK]; 8925 8926 do { 8927 int pass = ++spa->spa_sync_pass; 8928 8929 spa_sync_config_object(spa, tx); 8930 spa_sync_aux_dev(spa, &spa->spa_spares, tx, 8931 ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES); 8932 spa_sync_aux_dev(spa, &spa->spa_l2cache, tx, 8933 ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE); 8934 spa_errlog_sync(spa, txg); 8935 dsl_pool_sync(dp, txg); 8936 8937 if (pass < zfs_sync_pass_deferred_free || 8938 spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) { 8939 /* 8940 * If the log space map feature is active we don't 8941 * care about deferred frees and the deferred bpobj 8942 * as the log space map should effectively have the 8943 * same results (i.e. appending only to one object). 8944 */ 8945 spa_sync_frees(spa, free_bpl, tx); 8946 } else { 8947 /* 8948 * We can not defer frees in pass 1, because 8949 * we sync the deferred frees later in pass 1. 8950 */ 8951 ASSERT3U(pass, >, 1); 8952 bplist_iterate(free_bpl, bpobj_enqueue_alloc_cb, 8953 &spa->spa_deferred_bpobj, tx); 8954 } 8955 8956 ddt_sync(spa, txg); 8957 dsl_scan_sync(dp, tx); 8958 svr_sync(spa, tx); 8959 spa_sync_upgrades(spa, tx); 8960 8961 spa_flush_metaslabs(spa, tx); 8962 8963 vdev_t *vd = NULL; 8964 while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg)) 8965 != NULL) 8966 vdev_sync(vd, txg); 8967 8968 /* 8969 * Note: We need to check if the MOS is dirty because we could 8970 * have marked the MOS dirty without updating the uberblock 8971 * (e.g. if we have sync tasks but no dirty user data). We need 8972 * to check the uberblock's rootbp because it is updated if we 8973 * have synced out dirty data (though in this case the MOS will 8974 * most likely also be dirty due to second order effects, we 8975 * don't want to rely on that here). 8976 */ 8977 if (pass == 1 && 8978 spa->spa_uberblock.ub_rootbp.blk_birth < txg && 8979 !dmu_objset_is_dirty(mos, txg)) { 8980 /* 8981 * Nothing changed on the first pass, therefore this 8982 * TXG is a no-op. Avoid syncing deferred frees, so 8983 * that we can keep this TXG as a no-op. 8984 */ 8985 ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg)); 8986 ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg)); 8987 ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg)); 8988 ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg)); 8989 break; 8990 } 8991 8992 spa_sync_deferred_frees(spa, tx); 8993 } while (dmu_objset_is_dirty(mos, txg)); 8994 } 8995 8996 /* 8997 * Rewrite the vdev configuration (which includes the uberblock) to 8998 * commit the transaction group. 8999 * 9000 * If there are no dirty vdevs, we sync the uberblock to a few random 9001 * top-level vdevs that are known to be visible in the config cache 9002 * (see spa_vdev_add() for a complete description). If there *are* dirty 9003 * vdevs, sync the uberblock to all vdevs. 9004 */ 9005 static void 9006 spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx) 9007 { 9008 vdev_t *rvd = spa->spa_root_vdev; 9009 uint64_t txg = tx->tx_txg; 9010 9011 for (;;) { 9012 int error = 0; 9013 9014 /* 9015 * We hold SCL_STATE to prevent vdev open/close/etc. 9016 * while we're attempting to write the vdev labels. 9017 */ 9018 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 9019 9020 if (list_is_empty(&spa->spa_config_dirty_list)) { 9021 vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL }; 9022 int svdcount = 0; 9023 int children = rvd->vdev_children; 9024 int c0 = spa_get_random(children); 9025 9026 for (int c = 0; c < children; c++) { 9027 vdev_t *vd = 9028 rvd->vdev_child[(c0 + c) % children]; 9029 9030 /* Stop when revisiting the first vdev */ 9031 if (c > 0 && svd[0] == vd) 9032 break; 9033 9034 if (vd->vdev_ms_array == 0 || 9035 vd->vdev_islog || 9036 !vdev_is_concrete(vd)) 9037 continue; 9038 9039 svd[svdcount++] = vd; 9040 if (svdcount == SPA_SYNC_MIN_VDEVS) 9041 break; 9042 } 9043 error = vdev_config_sync(svd, svdcount, txg); 9044 } else { 9045 error = vdev_config_sync(rvd->vdev_child, 9046 rvd->vdev_children, txg); 9047 } 9048 9049 if (error == 0) 9050 spa->spa_last_synced_guid = rvd->vdev_guid; 9051 9052 spa_config_exit(spa, SCL_STATE, FTAG); 9053 9054 if (error == 0) 9055 break; 9056 zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR); 9057 zio_resume_wait(spa); 9058 } 9059 } 9060 9061 /* 9062 * Sync the specified transaction group. New blocks may be dirtied as 9063 * part of the process, so we iterate until it converges. 9064 */ 9065 void 9066 spa_sync(spa_t *spa, uint64_t txg) 9067 { 9068 vdev_t *vd = NULL; 9069 9070 VERIFY(spa_writeable(spa)); 9071 9072 /* 9073 * Wait for i/os issued in open context that need to complete 9074 * before this txg syncs. 9075 */ 9076 (void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]); 9077 spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL, 9078 ZIO_FLAG_CANFAIL); 9079 9080 /* 9081 * Lock out configuration changes. 9082 */ 9083 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER); 9084 9085 spa->spa_syncing_txg = txg; 9086 spa->spa_sync_pass = 0; 9087 9088 for (int i = 0; i < spa->spa_alloc_count; i++) { 9089 mutex_enter(&spa->spa_alloc_locks[i]); 9090 VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i])); 9091 mutex_exit(&spa->spa_alloc_locks[i]); 9092 } 9093 9094 /* 9095 * If there are any pending vdev state changes, convert them 9096 * into config changes that go out with this transaction group. 9097 */ 9098 spa_config_enter(spa, SCL_STATE, FTAG, RW_READER); 9099 while (list_head(&spa->spa_state_dirty_list) != NULL) { 9100 /* 9101 * We need the write lock here because, for aux vdevs, 9102 * calling vdev_config_dirty() modifies sav_config. 9103 * This is ugly and will become unnecessary when we 9104 * eliminate the aux vdev wart by integrating all vdevs 9105 * into the root vdev tree. 9106 */ 9107 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9108 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER); 9109 while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) { 9110 vdev_state_clean(vd); 9111 vdev_config_dirty(vd); 9112 } 9113 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9114 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 9115 } 9116 spa_config_exit(spa, SCL_STATE, FTAG); 9117 9118 dsl_pool_t *dp = spa->spa_dsl_pool; 9119 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg); 9120 9121 spa->spa_sync_starttime = gethrtime(); 9122 taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); 9123 spa->spa_deadman_tqid = taskq_dispatch_delay(system_delay_taskq, 9124 spa_deadman, spa, TQ_SLEEP, ddi_get_lbolt() + 9125 NSEC_TO_TICK(spa->spa_deadman_synctime)); 9126 9127 /* 9128 * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg, 9129 * set spa_deflate if we have no raid-z vdevs. 9130 */ 9131 if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE && 9132 spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) { 9133 vdev_t *rvd = spa->spa_root_vdev; 9134 9135 int i; 9136 for (i = 0; i < rvd->vdev_children; i++) { 9137 vd = rvd->vdev_child[i]; 9138 if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE) 9139 break; 9140 } 9141 if (i == rvd->vdev_children) { 9142 spa->spa_deflate = TRUE; 9143 VERIFY0(zap_add(spa->spa_meta_objset, 9144 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE, 9145 sizeof (uint64_t), 1, &spa->spa_deflate, tx)); 9146 } 9147 } 9148 9149 spa_sync_adjust_vdev_max_queue_depth(spa); 9150 9151 spa_sync_condense_indirect(spa, tx); 9152 9153 spa_sync_iterate_to_convergence(spa, tx); 9154 9155 #ifdef ZFS_DEBUG 9156 if (!list_is_empty(&spa->spa_config_dirty_list)) { 9157 /* 9158 * Make sure that the number of ZAPs for all the vdevs matches 9159 * the number of ZAPs in the per-vdev ZAP list. This only gets 9160 * called if the config is dirty; otherwise there may be 9161 * outstanding AVZ operations that weren't completed in 9162 * spa_sync_config_object. 9163 */ 9164 uint64_t all_vdev_zap_entry_count; 9165 ASSERT0(zap_count(spa->spa_meta_objset, 9166 spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count)); 9167 ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==, 9168 all_vdev_zap_entry_count); 9169 } 9170 #endif 9171 9172 if (spa->spa_vdev_removal != NULL) { 9173 ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]); 9174 } 9175 9176 spa_sync_rewrite_vdev_config(spa, tx); 9177 dmu_tx_commit(tx); 9178 9179 taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid); 9180 spa->spa_deadman_tqid = 0; 9181 9182 /* 9183 * Clear the dirty config list. 9184 */ 9185 while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL) 9186 vdev_config_clean(vd); 9187 9188 /* 9189 * Now that the new config has synced transactionally, 9190 * let it become visible to the config cache. 9191 */ 9192 if (spa->spa_config_syncing != NULL) { 9193 spa_config_set(spa, spa->spa_config_syncing); 9194 spa->spa_config_txg = txg; 9195 spa->spa_config_syncing = NULL; 9196 } 9197 9198 dsl_pool_sync_done(dp, txg); 9199 9200 for (int i = 0; i < spa->spa_alloc_count; i++) { 9201 mutex_enter(&spa->spa_alloc_locks[i]); 9202 VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i])); 9203 mutex_exit(&spa->spa_alloc_locks[i]); 9204 } 9205 9206 /* 9207 * Update usable space statistics. 9208 */ 9209 while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg))) 9210 != NULL) 9211 vdev_sync_done(vd, txg); 9212 9213 metaslab_class_evict_old(spa->spa_normal_class, txg); 9214 metaslab_class_evict_old(spa->spa_log_class, txg); 9215 9216 spa_sync_close_syncing_log_sm(spa); 9217 9218 spa_update_dspace(spa); 9219 9220 /* 9221 * It had better be the case that we didn't dirty anything 9222 * since vdev_config_sync(). 9223 */ 9224 ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg)); 9225 ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg)); 9226 ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg)); 9227 9228 while (zfs_pause_spa_sync) 9229 delay(1); 9230 9231 spa->spa_sync_pass = 0; 9232 9233 /* 9234 * Update the last synced uberblock here. We want to do this at 9235 * the end of spa_sync() so that consumers of spa_last_synced_txg() 9236 * will be guaranteed that all the processing associated with 9237 * that txg has been completed. 9238 */ 9239 spa->spa_ubsync = spa->spa_uberblock; 9240 spa_config_exit(spa, SCL_CONFIG, FTAG); 9241 9242 spa_handle_ignored_writes(spa); 9243 9244 /* 9245 * If any async tasks have been requested, kick them off. 9246 */ 9247 spa_async_dispatch(spa); 9248 } 9249 9250 /* 9251 * Sync all pools. We don't want to hold the namespace lock across these 9252 * operations, so we take a reference on the spa_t and drop the lock during the 9253 * sync. 9254 */ 9255 void 9256 spa_sync_allpools(void) 9257 { 9258 spa_t *spa = NULL; 9259 mutex_enter(&spa_namespace_lock); 9260 while ((spa = spa_next(spa)) != NULL) { 9261 if (spa_state(spa) != POOL_STATE_ACTIVE || 9262 !spa_writeable(spa) || spa_suspended(spa)) 9263 continue; 9264 spa_open_ref(spa, FTAG); 9265 mutex_exit(&spa_namespace_lock); 9266 txg_wait_synced(spa_get_dsl(spa), 0); 9267 mutex_enter(&spa_namespace_lock); 9268 spa_close(spa, FTAG); 9269 } 9270 mutex_exit(&spa_namespace_lock); 9271 } 9272 9273 /* 9274 * ========================================================================== 9275 * Miscellaneous routines 9276 * ========================================================================== 9277 */ 9278 9279 /* 9280 * Remove all pools in the system. 9281 */ 9282 void 9283 spa_evict_all(void) 9284 { 9285 spa_t *spa; 9286 9287 /* 9288 * Remove all cached state. All pools should be closed now, 9289 * so every spa in the AVL tree should be unreferenced. 9290 */ 9291 mutex_enter(&spa_namespace_lock); 9292 while ((spa = spa_next(NULL)) != NULL) { 9293 /* 9294 * Stop async tasks. The async thread may need to detach 9295 * a device that's been replaced, which requires grabbing 9296 * spa_namespace_lock, so we must drop it here. 9297 */ 9298 spa_open_ref(spa, FTAG); 9299 mutex_exit(&spa_namespace_lock); 9300 spa_async_suspend(spa); 9301 mutex_enter(&spa_namespace_lock); 9302 spa_close(spa, FTAG); 9303 9304 if (spa->spa_state != POOL_STATE_UNINITIALIZED) { 9305 spa_unload(spa); 9306 spa_deactivate(spa); 9307 } 9308 spa_remove(spa); 9309 } 9310 mutex_exit(&spa_namespace_lock); 9311 } 9312 9313 vdev_t * 9314 spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux) 9315 { 9316 vdev_t *vd; 9317 int i; 9318 9319 if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL) 9320 return (vd); 9321 9322 if (aux) { 9323 for (i = 0; i < spa->spa_l2cache.sav_count; i++) { 9324 vd = spa->spa_l2cache.sav_vdevs[i]; 9325 if (vd->vdev_guid == guid) 9326 return (vd); 9327 } 9328 9329 for (i = 0; i < spa->spa_spares.sav_count; i++) { 9330 vd = spa->spa_spares.sav_vdevs[i]; 9331 if (vd->vdev_guid == guid) 9332 return (vd); 9333 } 9334 } 9335 9336 return (NULL); 9337 } 9338 9339 void 9340 spa_upgrade(spa_t *spa, uint64_t version) 9341 { 9342 ASSERT(spa_writeable(spa)); 9343 9344 spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER); 9345 9346 /* 9347 * This should only be called for a non-faulted pool, and since a 9348 * future version would result in an unopenable pool, this shouldn't be 9349 * possible. 9350 */ 9351 ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version)); 9352 ASSERT3U(version, >=, spa->spa_uberblock.ub_version); 9353 9354 spa->spa_uberblock.ub_version = version; 9355 vdev_config_dirty(spa->spa_root_vdev); 9356 9357 spa_config_exit(spa, SCL_ALL, FTAG); 9358 9359 txg_wait_synced(spa_get_dsl(spa), 0); 9360 } 9361 9362 boolean_t 9363 spa_has_spare(spa_t *spa, uint64_t guid) 9364 { 9365 int i; 9366 uint64_t spareguid; 9367 spa_aux_vdev_t *sav = &spa->spa_spares; 9368 9369 for (i = 0; i < sav->sav_count; i++) 9370 if (sav->sav_vdevs[i]->vdev_guid == guid) 9371 return (B_TRUE); 9372 9373 for (i = 0; i < sav->sav_npending; i++) { 9374 if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID, 9375 &spareguid) == 0 && spareguid == guid) 9376 return (B_TRUE); 9377 } 9378 9379 return (B_FALSE); 9380 } 9381 9382 /* 9383 * Check if a pool has an active shared spare device. 9384 * Note: reference count of an active spare is 2, as a spare and as a replace 9385 */ 9386 static boolean_t 9387 spa_has_active_shared_spare(spa_t *spa) 9388 { 9389 int i, refcnt; 9390 uint64_t pool; 9391 spa_aux_vdev_t *sav = &spa->spa_spares; 9392 9393 for (i = 0; i < sav->sav_count; i++) { 9394 if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool, 9395 &refcnt) && pool != 0ULL && pool == spa_guid(spa) && 9396 refcnt > 2) 9397 return (B_TRUE); 9398 } 9399 9400 return (B_FALSE); 9401 } 9402 9403 uint64_t 9404 spa_total_metaslabs(spa_t *spa) 9405 { 9406 vdev_t *rvd = spa->spa_root_vdev; 9407 9408 uint64_t m = 0; 9409 for (uint64_t c = 0; c < rvd->vdev_children; c++) { 9410 vdev_t *vd = rvd->vdev_child[c]; 9411 if (!vdev_is_concrete(vd)) 9412 continue; 9413 m += vd->vdev_ms_count; 9414 } 9415 return (m); 9416 } 9417 9418 /* 9419 * Notify any waiting threads that some activity has switched from being in- 9420 * progress to not-in-progress so that the thread can wake up and determine 9421 * whether it is finished waiting. 9422 */ 9423 void 9424 spa_notify_waiters(spa_t *spa) 9425 { 9426 /* 9427 * Acquiring spa_activities_lock here prevents the cv_broadcast from 9428 * happening between the waiting thread's check and cv_wait. 9429 */ 9430 mutex_enter(&spa->spa_activities_lock); 9431 cv_broadcast(&spa->spa_activities_cv); 9432 mutex_exit(&spa->spa_activities_lock); 9433 } 9434 9435 /* 9436 * Notify any waiting threads that the pool is exporting, and then block until 9437 * they are finished using the spa_t. 9438 */ 9439 void 9440 spa_wake_waiters(spa_t *spa) 9441 { 9442 mutex_enter(&spa->spa_activities_lock); 9443 spa->spa_waiters_cancel = B_TRUE; 9444 cv_broadcast(&spa->spa_activities_cv); 9445 while (spa->spa_waiters != 0) 9446 cv_wait(&spa->spa_waiters_cv, &spa->spa_activities_lock); 9447 spa->spa_waiters_cancel = B_FALSE; 9448 mutex_exit(&spa->spa_activities_lock); 9449 } 9450 9451 /* Whether the vdev or any of its descendants are being initialized/trimmed. */ 9452 static boolean_t 9453 spa_vdev_activity_in_progress_impl(vdev_t *vd, zpool_wait_activity_t activity) 9454 { 9455 spa_t *spa = vd->vdev_spa; 9456 9457 ASSERT(spa_config_held(spa, SCL_CONFIG | SCL_STATE, RW_READER)); 9458 ASSERT(MUTEX_HELD(&spa->spa_activities_lock)); 9459 ASSERT(activity == ZPOOL_WAIT_INITIALIZE || 9460 activity == ZPOOL_WAIT_TRIM); 9461 9462 kmutex_t *lock = activity == ZPOOL_WAIT_INITIALIZE ? 9463 &vd->vdev_initialize_lock : &vd->vdev_trim_lock; 9464 9465 mutex_exit(&spa->spa_activities_lock); 9466 mutex_enter(lock); 9467 mutex_enter(&spa->spa_activities_lock); 9468 9469 boolean_t in_progress = (activity == ZPOOL_WAIT_INITIALIZE) ? 9470 (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) : 9471 (vd->vdev_trim_state == VDEV_TRIM_ACTIVE); 9472 mutex_exit(lock); 9473 9474 if (in_progress) 9475 return (B_TRUE); 9476 9477 for (int i = 0; i < vd->vdev_children; i++) { 9478 if (spa_vdev_activity_in_progress_impl(vd->vdev_child[i], 9479 activity)) 9480 return (B_TRUE); 9481 } 9482 9483 return (B_FALSE); 9484 } 9485 9486 /* 9487 * If use_guid is true, this checks whether the vdev specified by guid is 9488 * being initialized/trimmed. Otherwise, it checks whether any vdev in the pool 9489 * is being initialized/trimmed. The caller must hold the config lock and 9490 * spa_activities_lock. 9491 */ 9492 static int 9493 spa_vdev_activity_in_progress(spa_t *spa, boolean_t use_guid, uint64_t guid, 9494 zpool_wait_activity_t activity, boolean_t *in_progress) 9495 { 9496 mutex_exit(&spa->spa_activities_lock); 9497 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 9498 mutex_enter(&spa->spa_activities_lock); 9499 9500 vdev_t *vd; 9501 if (use_guid) { 9502 vd = spa_lookup_by_guid(spa, guid, B_FALSE); 9503 if (vd == NULL || !vd->vdev_ops->vdev_op_leaf) { 9504 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9505 return (EINVAL); 9506 } 9507 } else { 9508 vd = spa->spa_root_vdev; 9509 } 9510 9511 *in_progress = spa_vdev_activity_in_progress_impl(vd, activity); 9512 9513 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9514 return (0); 9515 } 9516 9517 /* 9518 * Locking for waiting threads 9519 * --------------------------- 9520 * 9521 * Waiting threads need a way to check whether a given activity is in progress, 9522 * and then, if it is, wait for it to complete. Each activity will have some 9523 * in-memory representation of the relevant on-disk state which can be used to 9524 * determine whether or not the activity is in progress. The in-memory state and 9525 * the locking used to protect it will be different for each activity, and may 9526 * not be suitable for use with a cvar (e.g., some state is protected by the 9527 * config lock). To allow waiting threads to wait without any races, another 9528 * lock, spa_activities_lock, is used. 9529 * 9530 * When the state is checked, both the activity-specific lock (if there is one) 9531 * and spa_activities_lock are held. In some cases, the activity-specific lock 9532 * is acquired explicitly (e.g. the config lock). In others, the locking is 9533 * internal to some check (e.g. bpobj_is_empty). After checking, the waiting 9534 * thread releases the activity-specific lock and, if the activity is in 9535 * progress, then cv_waits using spa_activities_lock. 9536 * 9537 * The waiting thread is woken when another thread, one completing some 9538 * activity, updates the state of the activity and then calls 9539 * spa_notify_waiters, which will cv_broadcast. This 'completing' thread only 9540 * needs to hold its activity-specific lock when updating the state, and this 9541 * lock can (but doesn't have to) be dropped before calling spa_notify_waiters. 9542 * 9543 * Because spa_notify_waiters acquires spa_activities_lock before broadcasting, 9544 * and because it is held when the waiting thread checks the state of the 9545 * activity, it can never be the case that the completing thread both updates 9546 * the activity state and cv_broadcasts in between the waiting thread's check 9547 * and cv_wait. Thus, a waiting thread can never miss a wakeup. 9548 * 9549 * In order to prevent deadlock, when the waiting thread does its check, in some 9550 * cases it will temporarily drop spa_activities_lock in order to acquire the 9551 * activity-specific lock. The order in which spa_activities_lock and the 9552 * activity specific lock are acquired in the waiting thread is determined by 9553 * the order in which they are acquired in the completing thread; if the 9554 * completing thread calls spa_notify_waiters with the activity-specific lock 9555 * held, then the waiting thread must also acquire the activity-specific lock 9556 * first. 9557 */ 9558 9559 static int 9560 spa_activity_in_progress(spa_t *spa, zpool_wait_activity_t activity, 9561 boolean_t use_tag, uint64_t tag, boolean_t *in_progress) 9562 { 9563 int error = 0; 9564 9565 ASSERT(MUTEX_HELD(&spa->spa_activities_lock)); 9566 9567 switch (activity) { 9568 case ZPOOL_WAIT_CKPT_DISCARD: 9569 *in_progress = 9570 (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT) && 9571 zap_contains(spa_meta_objset(spa), 9572 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT) == 9573 ENOENT); 9574 break; 9575 case ZPOOL_WAIT_FREE: 9576 *in_progress = ((spa_version(spa) >= SPA_VERSION_DEADLISTS && 9577 !bpobj_is_empty(&spa->spa_dsl_pool->dp_free_bpobj)) || 9578 spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY) || 9579 spa_livelist_delete_check(spa)); 9580 break; 9581 case ZPOOL_WAIT_INITIALIZE: 9582 case ZPOOL_WAIT_TRIM: 9583 error = spa_vdev_activity_in_progress(spa, use_tag, tag, 9584 activity, in_progress); 9585 break; 9586 case ZPOOL_WAIT_REPLACE: 9587 mutex_exit(&spa->spa_activities_lock); 9588 spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER); 9589 mutex_enter(&spa->spa_activities_lock); 9590 9591 *in_progress = vdev_replace_in_progress(spa->spa_root_vdev); 9592 spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG); 9593 break; 9594 case ZPOOL_WAIT_REMOVE: 9595 *in_progress = (spa->spa_removing_phys.sr_state == 9596 DSS_SCANNING); 9597 break; 9598 case ZPOOL_WAIT_RESILVER: 9599 if ((*in_progress = vdev_rebuild_active(spa->spa_root_vdev))) 9600 break; 9601 /* fall through */ 9602 case ZPOOL_WAIT_SCRUB: 9603 { 9604 boolean_t scanning, paused, is_scrub; 9605 dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan; 9606 9607 is_scrub = (scn->scn_phys.scn_func == POOL_SCAN_SCRUB); 9608 scanning = (scn->scn_phys.scn_state == DSS_SCANNING); 9609 paused = dsl_scan_is_paused_scrub(scn); 9610 *in_progress = (scanning && !paused && 9611 is_scrub == (activity == ZPOOL_WAIT_SCRUB)); 9612 break; 9613 } 9614 default: 9615 panic("unrecognized value for activity %d", activity); 9616 } 9617 9618 return (error); 9619 } 9620 9621 static int 9622 spa_wait_common(const char *pool, zpool_wait_activity_t activity, 9623 boolean_t use_tag, uint64_t tag, boolean_t *waited) 9624 { 9625 /* 9626 * The tag is used to distinguish between instances of an activity. 9627 * 'initialize' and 'trim' are the only activities that we use this for. 9628 * The other activities can only have a single instance in progress in a 9629 * pool at one time, making the tag unnecessary. 9630 * 9631 * There can be multiple devices being replaced at once, but since they 9632 * all finish once resilvering finishes, we don't bother keeping track 9633 * of them individually, we just wait for them all to finish. 9634 */ 9635 if (use_tag && activity != ZPOOL_WAIT_INITIALIZE && 9636 activity != ZPOOL_WAIT_TRIM) 9637 return (EINVAL); 9638 9639 if (activity < 0 || activity >= ZPOOL_WAIT_NUM_ACTIVITIES) 9640 return (EINVAL); 9641 9642 spa_t *spa; 9643 int error = spa_open(pool, &spa, FTAG); 9644 if (error != 0) 9645 return (error); 9646 9647 /* 9648 * Increment the spa's waiter count so that we can call spa_close and 9649 * still ensure that the spa_t doesn't get freed before this thread is 9650 * finished with it when the pool is exported. We want to call spa_close 9651 * before we start waiting because otherwise the additional ref would 9652 * prevent the pool from being exported or destroyed throughout the 9653 * potentially long wait. 9654 */ 9655 mutex_enter(&spa->spa_activities_lock); 9656 spa->spa_waiters++; 9657 spa_close(spa, FTAG); 9658 9659 *waited = B_FALSE; 9660 for (;;) { 9661 boolean_t in_progress; 9662 error = spa_activity_in_progress(spa, activity, use_tag, tag, 9663 &in_progress); 9664 9665 if (error || !in_progress || spa->spa_waiters_cancel) 9666 break; 9667 9668 *waited = B_TRUE; 9669 9670 if (cv_wait_sig(&spa->spa_activities_cv, 9671 &spa->spa_activities_lock) == 0) { 9672 error = EINTR; 9673 break; 9674 } 9675 } 9676 9677 spa->spa_waiters--; 9678 cv_signal(&spa->spa_waiters_cv); 9679 mutex_exit(&spa->spa_activities_lock); 9680 9681 return (error); 9682 } 9683 9684 /* 9685 * Wait for a particular instance of the specified activity to complete, where 9686 * the instance is identified by 'tag' 9687 */ 9688 int 9689 spa_wait_tag(const char *pool, zpool_wait_activity_t activity, uint64_t tag, 9690 boolean_t *waited) 9691 { 9692 return (spa_wait_common(pool, activity, B_TRUE, tag, waited)); 9693 } 9694 9695 /* 9696 * Wait for all instances of the specified activity complete 9697 */ 9698 int 9699 spa_wait(const char *pool, zpool_wait_activity_t activity, boolean_t *waited) 9700 { 9701 9702 return (spa_wait_common(pool, activity, B_FALSE, 0, waited)); 9703 } 9704 9705 sysevent_t * 9706 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name) 9707 { 9708 sysevent_t *ev = NULL; 9709 #ifdef _KERNEL 9710 nvlist_t *resource; 9711 9712 resource = zfs_event_create(spa, vd, FM_SYSEVENT_CLASS, name, hist_nvl); 9713 if (resource) { 9714 ev = kmem_alloc(sizeof (sysevent_t), KM_SLEEP); 9715 ev->resource = resource; 9716 } 9717 #endif 9718 return (ev); 9719 } 9720 9721 void 9722 spa_event_post(sysevent_t *ev) 9723 { 9724 #ifdef _KERNEL 9725 if (ev) { 9726 zfs_zevent_post(ev->resource, NULL, zfs_zevent_post_cb); 9727 kmem_free(ev, sizeof (*ev)); 9728 } 9729 #endif 9730 } 9731 9732 /* 9733 * Post a zevent corresponding to the given sysevent. The 'name' must be one 9734 * of the event definitions in sys/sysevent/eventdefs.h. The payload will be 9735 * filled in from the spa and (optionally) the vdev. This doesn't do anything 9736 * in the userland libzpool, as we don't want consumers to misinterpret ztest 9737 * or zdb as real changes. 9738 */ 9739 void 9740 spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name) 9741 { 9742 spa_event_post(spa_event_create(spa, vd, hist_nvl, name)); 9743 } 9744 9745 /* state manipulation functions */ 9746 EXPORT_SYMBOL(spa_open); 9747 EXPORT_SYMBOL(spa_open_rewind); 9748 EXPORT_SYMBOL(spa_get_stats); 9749 EXPORT_SYMBOL(spa_create); 9750 EXPORT_SYMBOL(spa_import); 9751 EXPORT_SYMBOL(spa_tryimport); 9752 EXPORT_SYMBOL(spa_destroy); 9753 EXPORT_SYMBOL(spa_export); 9754 EXPORT_SYMBOL(spa_reset); 9755 EXPORT_SYMBOL(spa_async_request); 9756 EXPORT_SYMBOL(spa_async_suspend); 9757 EXPORT_SYMBOL(spa_async_resume); 9758 EXPORT_SYMBOL(spa_inject_addref); 9759 EXPORT_SYMBOL(spa_inject_delref); 9760 EXPORT_SYMBOL(spa_scan_stat_init); 9761 EXPORT_SYMBOL(spa_scan_get_stats); 9762 9763 /* device manipulation */ 9764 EXPORT_SYMBOL(spa_vdev_add); 9765 EXPORT_SYMBOL(spa_vdev_attach); 9766 EXPORT_SYMBOL(spa_vdev_detach); 9767 EXPORT_SYMBOL(spa_vdev_setpath); 9768 EXPORT_SYMBOL(spa_vdev_setfru); 9769 EXPORT_SYMBOL(spa_vdev_split_mirror); 9770 9771 /* spare statech is global across all pools) */ 9772 EXPORT_SYMBOL(spa_spare_add); 9773 EXPORT_SYMBOL(spa_spare_remove); 9774 EXPORT_SYMBOL(spa_spare_exists); 9775 EXPORT_SYMBOL(spa_spare_activate); 9776 9777 /* L2ARC statech is global across all pools) */ 9778 EXPORT_SYMBOL(spa_l2cache_add); 9779 EXPORT_SYMBOL(spa_l2cache_remove); 9780 EXPORT_SYMBOL(spa_l2cache_exists); 9781 EXPORT_SYMBOL(spa_l2cache_activate); 9782 EXPORT_SYMBOL(spa_l2cache_drop); 9783 9784 /* scanning */ 9785 EXPORT_SYMBOL(spa_scan); 9786 EXPORT_SYMBOL(spa_scan_stop); 9787 9788 /* spa syncing */ 9789 EXPORT_SYMBOL(spa_sync); /* only for DMU use */ 9790 EXPORT_SYMBOL(spa_sync_allpools); 9791 9792 /* properties */ 9793 EXPORT_SYMBOL(spa_prop_set); 9794 EXPORT_SYMBOL(spa_prop_get); 9795 EXPORT_SYMBOL(spa_prop_clear_bootfs); 9796 9797 /* asynchronous event notification */ 9798 EXPORT_SYMBOL(spa_event_notify); 9799 9800 /* BEGIN CSTYLED */ 9801 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_shift, INT, ZMOD_RW, 9802 "log2(fraction of arc that can be used by inflight I/Os when " 9803 "verifying pool during import"); 9804 9805 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_metadata, INT, ZMOD_RW, 9806 "Set to traverse metadata on pool import"); 9807 9808 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW, 9809 "Set to traverse data on pool import"); 9810 9811 ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW, 9812 "Print vdev tree to zfs_dbgmsg during pool import"); 9813 9814 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD, 9815 "Percentage of CPUs to run an IO worker thread"); 9816 9817 ZFS_MODULE_PARAM(zfs, zfs_, max_missing_tvds, ULONG, ZMOD_RW, 9818 "Allow importing pool with up to this number of missing top-level " 9819 "vdevs (in read-only mode)"); 9820 9821 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_pause, INT, ZMOD_RW, 9822 "Set the livelist condense zthr to pause"); 9823 9824 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_pause, INT, ZMOD_RW, 9825 "Set the livelist condense synctask to pause"); 9826 9827 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_cancel, INT, ZMOD_RW, 9828 "Whether livelist condensing was canceled in the synctask"); 9829 9830 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_cancel, INT, ZMOD_RW, 9831 "Whether livelist condensing was canceled in the zthr function"); 9832 9833 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT, ZMOD_RW, 9834 "Whether extra ALLOC blkptrs were added to a livelist entry while it " 9835 "was being condensed"); 9836 /* END CSTYLED */ 9837