xref: /freebsd/sys/contrib/openzfs/module/zfs/spa.c (revision 9247238cc4b8835892a47701136b0fd073f8d67c)
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 https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
25  * Copyright (c) 2018, Nexenta Systems, Inc.  All rights reserved.
26  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27  * Copyright 2013 Saso Kiselkov. All rights reserved.
28  * Copyright (c) 2014 Integros [integros.com]
29  * Copyright 2016 Toomas Soome <tsoome@me.com>
30  * Copyright (c) 2016 Actifio, Inc. All rights reserved.
31  * Copyright 2018 Joyent, Inc.
32  * Copyright (c) 2017, 2019, Datto Inc. All rights reserved.
33  * Copyright 2017 Joyent, Inc.
34  * Copyright (c) 2017, Intel Corporation.
35  * Copyright (c) 2021, Colm Buckley <colm@tuatha.org>
36  * Copyright (c) 2023 Hewlett Packard Enterprise Development LP.
37  */
38 
39 /*
40  * SPA: Storage Pool Allocator
41  *
42  * This file contains all the routines used when modifying on-disk SPA state.
43  * This includes opening, importing, destroying, exporting a pool, and syncing a
44  * pool.
45  */
46 
47 #include <sys/zfs_context.h>
48 #include <sys/fm/fs/zfs.h>
49 #include <sys/spa_impl.h>
50 #include <sys/zio.h>
51 #include <sys/zio_checksum.h>
52 #include <sys/dmu.h>
53 #include <sys/dmu_tx.h>
54 #include <sys/zap.h>
55 #include <sys/zil.h>
56 #include <sys/brt.h>
57 #include <sys/ddt.h>
58 #include <sys/vdev_impl.h>
59 #include <sys/vdev_removal.h>
60 #include <sys/vdev_indirect_mapping.h>
61 #include <sys/vdev_indirect_births.h>
62 #include <sys/vdev_initialize.h>
63 #include <sys/vdev_rebuild.h>
64 #include <sys/vdev_trim.h>
65 #include <sys/vdev_disk.h>
66 #include <sys/vdev_draid.h>
67 #include <sys/metaslab.h>
68 #include <sys/metaslab_impl.h>
69 #include <sys/mmp.h>
70 #include <sys/uberblock_impl.h>
71 #include <sys/txg.h>
72 #include <sys/avl.h>
73 #include <sys/bpobj.h>
74 #include <sys/dmu_traverse.h>
75 #include <sys/dmu_objset.h>
76 #include <sys/unique.h>
77 #include <sys/dsl_pool.h>
78 #include <sys/dsl_dataset.h>
79 #include <sys/dsl_dir.h>
80 #include <sys/dsl_prop.h>
81 #include <sys/dsl_synctask.h>
82 #include <sys/fs/zfs.h>
83 #include <sys/arc.h>
84 #include <sys/callb.h>
85 #include <sys/systeminfo.h>
86 #include <sys/zfs_ioctl.h>
87 #include <sys/dsl_scan.h>
88 #include <sys/zfeature.h>
89 #include <sys/dsl_destroy.h>
90 #include <sys/zvol.h>
91 
92 #ifdef	_KERNEL
93 #include <sys/fm/protocol.h>
94 #include <sys/fm/util.h>
95 #include <sys/callb.h>
96 #include <sys/zone.h>
97 #include <sys/vmsystm.h>
98 #endif	/* _KERNEL */
99 
100 #include "zfs_prop.h"
101 #include "zfs_comutil.h"
102 
103 /*
104  * The interval, in seconds, at which failed configuration cache file writes
105  * should be retried.
106  */
107 int zfs_ccw_retry_interval = 300;
108 
109 typedef enum zti_modes {
110 	ZTI_MODE_FIXED,			/* value is # of threads (min 1) */
111 	ZTI_MODE_BATCH,			/* cpu-intensive; value is ignored */
112 	ZTI_MODE_SCALE,			/* Taskqs scale with CPUs. */
113 	ZTI_MODE_NULL,			/* don't create a taskq */
114 	ZTI_NMODES
115 } zti_modes_t;
116 
117 #define	ZTI_P(n, q)	{ ZTI_MODE_FIXED, (n), (q) }
118 #define	ZTI_PCT(n)	{ ZTI_MODE_ONLINE_PERCENT, (n), 1 }
119 #define	ZTI_BATCH	{ ZTI_MODE_BATCH, 0, 1 }
120 #define	ZTI_SCALE	{ ZTI_MODE_SCALE, 0, 1 }
121 #define	ZTI_NULL	{ ZTI_MODE_NULL, 0, 0 }
122 
123 #define	ZTI_N(n)	ZTI_P(n, 1)
124 #define	ZTI_ONE		ZTI_N(1)
125 
126 typedef struct zio_taskq_info {
127 	zti_modes_t zti_mode;
128 	uint_t zti_value;
129 	uint_t zti_count;
130 } zio_taskq_info_t;
131 
132 static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = {
133 	"iss", "iss_h", "int", "int_h"
134 };
135 
136 /*
137  * This table defines the taskq settings for each ZFS I/O type. When
138  * initializing a pool, we use this table to create an appropriately sized
139  * taskq. Some operations are low volume and therefore have a small, static
140  * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE
141  * macros. Other operations process a large amount of data; the ZTI_BATCH
142  * macro causes us to create a taskq oriented for throughput. Some operations
143  * are so high frequency and short-lived that the taskq itself can become a
144  * point of lock contention. The ZTI_P(#, #) macro indicates that we need an
145  * additional degree of parallelism specified by the number of threads per-
146  * taskq and the number of taskqs; when dispatching an event in this case, the
147  * particular taskq is chosen at random. ZTI_SCALE is similar to ZTI_BATCH,
148  * but with number of taskqs also scaling with number of CPUs.
149  *
150  * The different taskq priorities are to handle the different contexts (issue
151  * and interrupt) and then to reserve threads for ZIO_PRIORITY_NOW I/Os that
152  * need to be handled with minimum delay.
153  */
154 static const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = {
155 	/* ISSUE	ISSUE_HIGH	INTR		INTR_HIGH */
156 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* NULL */
157 	{ ZTI_N(8),	ZTI_NULL,	ZTI_SCALE,	ZTI_NULL }, /* READ */
158 	{ ZTI_BATCH,	ZTI_N(5),	ZTI_SCALE,	ZTI_N(5) }, /* WRITE */
159 	{ ZTI_SCALE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* FREE */
160 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* CLAIM */
161 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* IOCTL */
162 	{ ZTI_N(4),	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* TRIM */
163 };
164 
165 static void spa_sync_version(void *arg, dmu_tx_t *tx);
166 static void spa_sync_props(void *arg, dmu_tx_t *tx);
167 static boolean_t spa_has_active_shared_spare(spa_t *spa);
168 static int spa_load_impl(spa_t *spa, spa_import_type_t type,
169     const char **ereport);
170 static void spa_vdev_resilver_done(spa_t *spa);
171 
172 static uint_t	zio_taskq_batch_pct = 80;	  /* 1 thread per cpu in pset */
173 static uint_t	zio_taskq_batch_tpq;		  /* threads per taskq */
174 static const boolean_t	zio_taskq_sysdc = B_TRUE; /* use SDC scheduling class */
175 static const uint_t	zio_taskq_basedc = 80;	  /* base duty cycle */
176 
177 static const boolean_t spa_create_process = B_TRUE; /* no process => no sysdc */
178 
179 /*
180  * Report any spa_load_verify errors found, but do not fail spa_load.
181  * This is used by zdb to analyze non-idle pools.
182  */
183 boolean_t	spa_load_verify_dryrun = B_FALSE;
184 
185 /*
186  * Allow read spacemaps in case of readonly import (spa_mode == SPA_MODE_READ).
187  * This is used by zdb for spacemaps verification.
188  */
189 boolean_t	spa_mode_readable_spacemaps = B_FALSE;
190 
191 /*
192  * This (illegal) pool name is used when temporarily importing a spa_t in order
193  * to get the vdev stats associated with the imported devices.
194  */
195 #define	TRYIMPORT_NAME	"$import"
196 
197 /*
198  * For debugging purposes: print out vdev tree during pool import.
199  */
200 static int		spa_load_print_vdev_tree = B_FALSE;
201 
202 /*
203  * A non-zero value for zfs_max_missing_tvds means that we allow importing
204  * pools with missing top-level vdevs. This is strictly intended for advanced
205  * pool recovery cases since missing data is almost inevitable. Pools with
206  * missing devices can only be imported read-only for safety reasons, and their
207  * fail-mode will be automatically set to "continue".
208  *
209  * With 1 missing vdev we should be able to import the pool and mount all
210  * datasets. User data that was not modified after the missing device has been
211  * added should be recoverable. This means that snapshots created prior to the
212  * addition of that device should be completely intact.
213  *
214  * With 2 missing vdevs, some datasets may fail to mount since there are
215  * dataset statistics that are stored as regular metadata. Some data might be
216  * recoverable if those vdevs were added recently.
217  *
218  * With 3 or more missing vdevs, the pool is severely damaged and MOS entries
219  * may be missing entirely. Chances of data recovery are very low. Note that
220  * there are also risks of performing an inadvertent rewind as we might be
221  * missing all the vdevs with the latest uberblocks.
222  */
223 uint64_t	zfs_max_missing_tvds = 0;
224 
225 /*
226  * The parameters below are similar to zfs_max_missing_tvds but are only
227  * intended for a preliminary open of the pool with an untrusted config which
228  * might be incomplete or out-dated.
229  *
230  * We are more tolerant for pools opened from a cachefile since we could have
231  * an out-dated cachefile where a device removal was not registered.
232  * We could have set the limit arbitrarily high but in the case where devices
233  * are really missing we would want to return the proper error codes; we chose
234  * SPA_DVAS_PER_BP - 1 so that some copies of the MOS would still be available
235  * and we get a chance to retrieve the trusted config.
236  */
237 uint64_t	zfs_max_missing_tvds_cachefile = SPA_DVAS_PER_BP - 1;
238 
239 /*
240  * In the case where config was assembled by scanning device paths (/dev/dsks
241  * by default) we are less tolerant since all the existing devices should have
242  * been detected and we want spa_load to return the right error codes.
243  */
244 uint64_t	zfs_max_missing_tvds_scan = 0;
245 
246 /*
247  * Debugging aid that pauses spa_sync() towards the end.
248  */
249 static const boolean_t	zfs_pause_spa_sync = B_FALSE;
250 
251 /*
252  * Variables to indicate the livelist condense zthr func should wait at certain
253  * points for the livelist to be removed - used to test condense/destroy races
254  */
255 static int zfs_livelist_condense_zthr_pause = 0;
256 static int zfs_livelist_condense_sync_pause = 0;
257 
258 /*
259  * Variables to track whether or not condense cancellation has been
260  * triggered in testing.
261  */
262 static int zfs_livelist_condense_sync_cancel = 0;
263 static int zfs_livelist_condense_zthr_cancel = 0;
264 
265 /*
266  * Variable to track whether or not extra ALLOC blkptrs were added to a
267  * livelist entry while it was being condensed (caused by the way we track
268  * remapped blkptrs in dbuf_remap_impl)
269  */
270 static int zfs_livelist_condense_new_alloc = 0;
271 
272 /*
273  * ==========================================================================
274  * SPA properties routines
275  * ==========================================================================
276  */
277 
278 /*
279  * Add a (source=src, propname=propval) list to an nvlist.
280  */
281 static void
282 spa_prop_add_list(nvlist_t *nvl, zpool_prop_t prop, const char *strval,
283     uint64_t intval, zprop_source_t src)
284 {
285 	const char *propname = zpool_prop_to_name(prop);
286 	nvlist_t *propval;
287 
288 	propval = fnvlist_alloc();
289 	fnvlist_add_uint64(propval, ZPROP_SOURCE, src);
290 
291 	if (strval != NULL)
292 		fnvlist_add_string(propval, ZPROP_VALUE, strval);
293 	else
294 		fnvlist_add_uint64(propval, ZPROP_VALUE, intval);
295 
296 	fnvlist_add_nvlist(nvl, propname, propval);
297 	nvlist_free(propval);
298 }
299 
300 /*
301  * Add a user property (source=src, propname=propval) to an nvlist.
302  */
303 static void
304 spa_prop_add_user(nvlist_t *nvl, const char *propname, char *strval,
305     zprop_source_t src)
306 {
307 	nvlist_t *propval;
308 
309 	VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
310 	VERIFY(nvlist_add_uint64(propval, ZPROP_SOURCE, src) == 0);
311 	VERIFY(nvlist_add_string(propval, ZPROP_VALUE, strval) == 0);
312 	VERIFY(nvlist_add_nvlist(nvl, propname, propval) == 0);
313 	nvlist_free(propval);
314 }
315 
316 /*
317  * Get property values from the spa configuration.
318  */
319 static void
320 spa_prop_get_config(spa_t *spa, nvlist_t **nvp)
321 {
322 	vdev_t *rvd = spa->spa_root_vdev;
323 	dsl_pool_t *pool = spa->spa_dsl_pool;
324 	uint64_t size, alloc, cap, version;
325 	const zprop_source_t src = ZPROP_SRC_NONE;
326 	spa_config_dirent_t *dp;
327 	metaslab_class_t *mc = spa_normal_class(spa);
328 
329 	ASSERT(MUTEX_HELD(&spa->spa_props_lock));
330 
331 	if (rvd != NULL) {
332 		alloc = metaslab_class_get_alloc(mc);
333 		alloc += metaslab_class_get_alloc(spa_special_class(spa));
334 		alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
335 		alloc += metaslab_class_get_alloc(spa_embedded_log_class(spa));
336 
337 		size = metaslab_class_get_space(mc);
338 		size += metaslab_class_get_space(spa_special_class(spa));
339 		size += metaslab_class_get_space(spa_dedup_class(spa));
340 		size += metaslab_class_get_space(spa_embedded_log_class(spa));
341 
342 		spa_prop_add_list(*nvp, ZPOOL_PROP_NAME, spa_name(spa), 0, src);
343 		spa_prop_add_list(*nvp, ZPOOL_PROP_SIZE, NULL, size, src);
344 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALLOCATED, NULL, alloc, src);
345 		spa_prop_add_list(*nvp, ZPOOL_PROP_FREE, NULL,
346 		    size - alloc, src);
347 		spa_prop_add_list(*nvp, ZPOOL_PROP_CHECKPOINT, NULL,
348 		    spa->spa_checkpoint_info.sci_dspace, src);
349 
350 		spa_prop_add_list(*nvp, ZPOOL_PROP_FRAGMENTATION, NULL,
351 		    metaslab_class_fragmentation(mc), src);
352 		spa_prop_add_list(*nvp, ZPOOL_PROP_EXPANDSZ, NULL,
353 		    metaslab_class_expandable_space(mc), src);
354 		spa_prop_add_list(*nvp, ZPOOL_PROP_READONLY, NULL,
355 		    (spa_mode(spa) == SPA_MODE_READ), src);
356 
357 		cap = (size == 0) ? 0 : (alloc * 100 / size);
358 		spa_prop_add_list(*nvp, ZPOOL_PROP_CAPACITY, NULL, cap, src);
359 
360 		spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL,
361 		    ddt_get_pool_dedup_ratio(spa), src);
362 		spa_prop_add_list(*nvp, ZPOOL_PROP_BCLONEUSED, NULL,
363 		    brt_get_used(spa), src);
364 		spa_prop_add_list(*nvp, ZPOOL_PROP_BCLONESAVED, NULL,
365 		    brt_get_saved(spa), src);
366 		spa_prop_add_list(*nvp, ZPOOL_PROP_BCLONERATIO, NULL,
367 		    brt_get_ratio(spa), src);
368 
369 		spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL,
370 		    rvd->vdev_state, src);
371 
372 		version = spa_version(spa);
373 		if (version == zpool_prop_default_numeric(ZPOOL_PROP_VERSION)) {
374 			spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL,
375 			    version, ZPROP_SRC_DEFAULT);
376 		} else {
377 			spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL,
378 			    version, ZPROP_SRC_LOCAL);
379 		}
380 		spa_prop_add_list(*nvp, ZPOOL_PROP_LOAD_GUID,
381 		    NULL, spa_load_guid(spa), src);
382 	}
383 
384 	if (pool != NULL) {
385 		/*
386 		 * The $FREE directory was introduced in SPA_VERSION_DEADLISTS,
387 		 * when opening pools before this version freedir will be NULL.
388 		 */
389 		if (pool->dp_free_dir != NULL) {
390 			spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL,
391 			    dsl_dir_phys(pool->dp_free_dir)->dd_used_bytes,
392 			    src);
393 		} else {
394 			spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING,
395 			    NULL, 0, src);
396 		}
397 
398 		if (pool->dp_leak_dir != NULL) {
399 			spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL,
400 			    dsl_dir_phys(pool->dp_leak_dir)->dd_used_bytes,
401 			    src);
402 		} else {
403 			spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED,
404 			    NULL, 0, src);
405 		}
406 	}
407 
408 	spa_prop_add_list(*nvp, ZPOOL_PROP_GUID, NULL, spa_guid(spa), src);
409 
410 	if (spa->spa_comment != NULL) {
411 		spa_prop_add_list(*nvp, ZPOOL_PROP_COMMENT, spa->spa_comment,
412 		    0, ZPROP_SRC_LOCAL);
413 	}
414 
415 	if (spa->spa_compatibility != NULL) {
416 		spa_prop_add_list(*nvp, ZPOOL_PROP_COMPATIBILITY,
417 		    spa->spa_compatibility, 0, ZPROP_SRC_LOCAL);
418 	}
419 
420 	if (spa->spa_root != NULL)
421 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root,
422 		    0, ZPROP_SRC_LOCAL);
423 
424 	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) {
425 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
426 		    MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE);
427 	} else {
428 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
429 		    SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE);
430 	}
431 
432 	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE)) {
433 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL,
434 		    DNODE_MAX_SIZE, ZPROP_SRC_NONE);
435 	} else {
436 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL,
437 		    DNODE_MIN_SIZE, ZPROP_SRC_NONE);
438 	}
439 
440 	if ((dp = list_head(&spa->spa_config_list)) != NULL) {
441 		if (dp->scd_path == NULL) {
442 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
443 			    "none", 0, ZPROP_SRC_LOCAL);
444 		} else if (strcmp(dp->scd_path, spa_config_path) != 0) {
445 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
446 			    dp->scd_path, 0, ZPROP_SRC_LOCAL);
447 		}
448 	}
449 }
450 
451 /*
452  * Get zpool property values.
453  */
454 int
455 spa_prop_get(spa_t *spa, nvlist_t **nvp)
456 {
457 	objset_t *mos = spa->spa_meta_objset;
458 	zap_cursor_t zc;
459 	zap_attribute_t za;
460 	dsl_pool_t *dp;
461 	int err;
462 
463 	err = nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP);
464 	if (err)
465 		return (err);
466 
467 	dp = spa_get_dsl(spa);
468 	dsl_pool_config_enter(dp, FTAG);
469 	mutex_enter(&spa->spa_props_lock);
470 
471 	/*
472 	 * Get properties from the spa config.
473 	 */
474 	spa_prop_get_config(spa, nvp);
475 
476 	/* If no pool property object, no more prop to get. */
477 	if (mos == NULL || spa->spa_pool_props_object == 0)
478 		goto out;
479 
480 	/*
481 	 * Get properties from the MOS pool property object.
482 	 */
483 	for (zap_cursor_init(&zc, mos, spa->spa_pool_props_object);
484 	    (err = zap_cursor_retrieve(&zc, &za)) == 0;
485 	    zap_cursor_advance(&zc)) {
486 		uint64_t intval = 0;
487 		char *strval = NULL;
488 		zprop_source_t src = ZPROP_SRC_DEFAULT;
489 		zpool_prop_t prop;
490 
491 		if ((prop = zpool_name_to_prop(za.za_name)) ==
492 		    ZPOOL_PROP_INVAL && !zfs_prop_user(za.za_name))
493 			continue;
494 
495 		switch (za.za_integer_length) {
496 		case 8:
497 			/* integer property */
498 			if (za.za_first_integer !=
499 			    zpool_prop_default_numeric(prop))
500 				src = ZPROP_SRC_LOCAL;
501 
502 			if (prop == ZPOOL_PROP_BOOTFS) {
503 				dsl_dataset_t *ds = NULL;
504 
505 				err = dsl_dataset_hold_obj(dp,
506 				    za.za_first_integer, FTAG, &ds);
507 				if (err != 0)
508 					break;
509 
510 				strval = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN,
511 				    KM_SLEEP);
512 				dsl_dataset_name(ds, strval);
513 				dsl_dataset_rele(ds, FTAG);
514 			} else {
515 				strval = NULL;
516 				intval = za.za_first_integer;
517 			}
518 
519 			spa_prop_add_list(*nvp, prop, strval, intval, src);
520 
521 			if (strval != NULL)
522 				kmem_free(strval, ZFS_MAX_DATASET_NAME_LEN);
523 
524 			break;
525 
526 		case 1:
527 			/* string property */
528 			strval = kmem_alloc(za.za_num_integers, KM_SLEEP);
529 			err = zap_lookup(mos, spa->spa_pool_props_object,
530 			    za.za_name, 1, za.za_num_integers, strval);
531 			if (err) {
532 				kmem_free(strval, za.za_num_integers);
533 				break;
534 			}
535 			if (prop != ZPOOL_PROP_INVAL) {
536 				spa_prop_add_list(*nvp, prop, strval, 0, src);
537 			} else {
538 				src = ZPROP_SRC_LOCAL;
539 				spa_prop_add_user(*nvp, za.za_name, strval,
540 				    src);
541 			}
542 			kmem_free(strval, za.za_num_integers);
543 			break;
544 
545 		default:
546 			break;
547 		}
548 	}
549 	zap_cursor_fini(&zc);
550 out:
551 	mutex_exit(&spa->spa_props_lock);
552 	dsl_pool_config_exit(dp, FTAG);
553 	if (err && err != ENOENT) {
554 		nvlist_free(*nvp);
555 		*nvp = NULL;
556 		return (err);
557 	}
558 
559 	return (0);
560 }
561 
562 /*
563  * Validate the given pool properties nvlist and modify the list
564  * for the property values to be set.
565  */
566 static int
567 spa_prop_validate(spa_t *spa, nvlist_t *props)
568 {
569 	nvpair_t *elem;
570 	int error = 0, reset_bootfs = 0;
571 	uint64_t objnum = 0;
572 	boolean_t has_feature = B_FALSE;
573 
574 	elem = NULL;
575 	while ((elem = nvlist_next_nvpair(props, elem)) != NULL) {
576 		uint64_t intval;
577 		const char *strval, *slash, *check, *fname;
578 		const char *propname = nvpair_name(elem);
579 		zpool_prop_t prop = zpool_name_to_prop(propname);
580 
581 		switch (prop) {
582 		case ZPOOL_PROP_INVAL:
583 			/*
584 			 * Sanitize the input.
585 			 */
586 			if (zfs_prop_user(propname)) {
587 				if (strlen(propname) >= ZAP_MAXNAMELEN) {
588 					error = SET_ERROR(ENAMETOOLONG);
589 					break;
590 				}
591 
592 				if (strlen(fnvpair_value_string(elem)) >=
593 				    ZAP_MAXVALUELEN) {
594 					error = SET_ERROR(E2BIG);
595 					break;
596 				}
597 			} else if (zpool_prop_feature(propname)) {
598 				if (nvpair_type(elem) != DATA_TYPE_UINT64) {
599 					error = SET_ERROR(EINVAL);
600 					break;
601 				}
602 
603 				if (nvpair_value_uint64(elem, &intval) != 0) {
604 					error = SET_ERROR(EINVAL);
605 					break;
606 				}
607 
608 				if (intval != 0) {
609 					error = SET_ERROR(EINVAL);
610 					break;
611 				}
612 
613 				fname = strchr(propname, '@') + 1;
614 				if (zfeature_lookup_name(fname, NULL) != 0) {
615 					error = SET_ERROR(EINVAL);
616 					break;
617 				}
618 
619 				has_feature = B_TRUE;
620 			} else {
621 				error = SET_ERROR(EINVAL);
622 				break;
623 			}
624 			break;
625 
626 		case ZPOOL_PROP_VERSION:
627 			error = nvpair_value_uint64(elem, &intval);
628 			if (!error &&
629 			    (intval < spa_version(spa) ||
630 			    intval > SPA_VERSION_BEFORE_FEATURES ||
631 			    has_feature))
632 				error = SET_ERROR(EINVAL);
633 			break;
634 
635 		case ZPOOL_PROP_DELEGATION:
636 		case ZPOOL_PROP_AUTOREPLACE:
637 		case ZPOOL_PROP_LISTSNAPS:
638 		case ZPOOL_PROP_AUTOEXPAND:
639 		case ZPOOL_PROP_AUTOTRIM:
640 			error = nvpair_value_uint64(elem, &intval);
641 			if (!error && intval > 1)
642 				error = SET_ERROR(EINVAL);
643 			break;
644 
645 		case ZPOOL_PROP_MULTIHOST:
646 			error = nvpair_value_uint64(elem, &intval);
647 			if (!error && intval > 1)
648 				error = SET_ERROR(EINVAL);
649 
650 			if (!error) {
651 				uint32_t hostid = zone_get_hostid(NULL);
652 				if (hostid)
653 					spa->spa_hostid = hostid;
654 				else
655 					error = SET_ERROR(ENOTSUP);
656 			}
657 
658 			break;
659 
660 		case ZPOOL_PROP_BOOTFS:
661 			/*
662 			 * If the pool version is less than SPA_VERSION_BOOTFS,
663 			 * or the pool is still being created (version == 0),
664 			 * the bootfs property cannot be set.
665 			 */
666 			if (spa_version(spa) < SPA_VERSION_BOOTFS) {
667 				error = SET_ERROR(ENOTSUP);
668 				break;
669 			}
670 
671 			/*
672 			 * Make sure the vdev config is bootable
673 			 */
674 			if (!vdev_is_bootable(spa->spa_root_vdev)) {
675 				error = SET_ERROR(ENOTSUP);
676 				break;
677 			}
678 
679 			reset_bootfs = 1;
680 
681 			error = nvpair_value_string(elem, &strval);
682 
683 			if (!error) {
684 				objset_t *os;
685 
686 				if (strval == NULL || strval[0] == '\0') {
687 					objnum = zpool_prop_default_numeric(
688 					    ZPOOL_PROP_BOOTFS);
689 					break;
690 				}
691 
692 				error = dmu_objset_hold(strval, FTAG, &os);
693 				if (error != 0)
694 					break;
695 
696 				/* Must be ZPL. */
697 				if (dmu_objset_type(os) != DMU_OST_ZFS) {
698 					error = SET_ERROR(ENOTSUP);
699 				} else {
700 					objnum = dmu_objset_id(os);
701 				}
702 				dmu_objset_rele(os, FTAG);
703 			}
704 			break;
705 
706 		case ZPOOL_PROP_FAILUREMODE:
707 			error = nvpair_value_uint64(elem, &intval);
708 			if (!error && intval > ZIO_FAILURE_MODE_PANIC)
709 				error = SET_ERROR(EINVAL);
710 
711 			/*
712 			 * This is a special case which only occurs when
713 			 * the pool has completely failed. This allows
714 			 * the user to change the in-core failmode property
715 			 * without syncing it out to disk (I/Os might
716 			 * currently be blocked). We do this by returning
717 			 * EIO to the caller (spa_prop_set) to trick it
718 			 * into thinking we encountered a property validation
719 			 * error.
720 			 */
721 			if (!error && spa_suspended(spa)) {
722 				spa->spa_failmode = intval;
723 				error = SET_ERROR(EIO);
724 			}
725 			break;
726 
727 		case ZPOOL_PROP_CACHEFILE:
728 			if ((error = nvpair_value_string(elem, &strval)) != 0)
729 				break;
730 
731 			if (strval[0] == '\0')
732 				break;
733 
734 			if (strcmp(strval, "none") == 0)
735 				break;
736 
737 			if (strval[0] != '/') {
738 				error = SET_ERROR(EINVAL);
739 				break;
740 			}
741 
742 			slash = strrchr(strval, '/');
743 			ASSERT(slash != NULL);
744 
745 			if (slash[1] == '\0' || strcmp(slash, "/.") == 0 ||
746 			    strcmp(slash, "/..") == 0)
747 				error = SET_ERROR(EINVAL);
748 			break;
749 
750 		case ZPOOL_PROP_COMMENT:
751 			if ((error = nvpair_value_string(elem, &strval)) != 0)
752 				break;
753 			for (check = strval; *check != '\0'; check++) {
754 				if (!isprint(*check)) {
755 					error = SET_ERROR(EINVAL);
756 					break;
757 				}
758 			}
759 			if (strlen(strval) > ZPROP_MAX_COMMENT)
760 				error = SET_ERROR(E2BIG);
761 			break;
762 
763 		default:
764 			break;
765 		}
766 
767 		if (error)
768 			break;
769 	}
770 
771 	(void) nvlist_remove_all(props,
772 	    zpool_prop_to_name(ZPOOL_PROP_DEDUPDITTO));
773 
774 	if (!error && reset_bootfs) {
775 		error = nvlist_remove(props,
776 		    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), DATA_TYPE_STRING);
777 
778 		if (!error) {
779 			error = nvlist_add_uint64(props,
780 			    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), objnum);
781 		}
782 	}
783 
784 	return (error);
785 }
786 
787 void
788 spa_configfile_set(spa_t *spa, nvlist_t *nvp, boolean_t need_sync)
789 {
790 	const char *cachefile;
791 	spa_config_dirent_t *dp;
792 
793 	if (nvlist_lookup_string(nvp, zpool_prop_to_name(ZPOOL_PROP_CACHEFILE),
794 	    &cachefile) != 0)
795 		return;
796 
797 	dp = kmem_alloc(sizeof (spa_config_dirent_t),
798 	    KM_SLEEP);
799 
800 	if (cachefile[0] == '\0')
801 		dp->scd_path = spa_strdup(spa_config_path);
802 	else if (strcmp(cachefile, "none") == 0)
803 		dp->scd_path = NULL;
804 	else
805 		dp->scd_path = spa_strdup(cachefile);
806 
807 	list_insert_head(&spa->spa_config_list, dp);
808 	if (need_sync)
809 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
810 }
811 
812 int
813 spa_prop_set(spa_t *spa, nvlist_t *nvp)
814 {
815 	int error;
816 	nvpair_t *elem = NULL;
817 	boolean_t need_sync = B_FALSE;
818 
819 	if ((error = spa_prop_validate(spa, nvp)) != 0)
820 		return (error);
821 
822 	while ((elem = nvlist_next_nvpair(nvp, elem)) != NULL) {
823 		zpool_prop_t prop = zpool_name_to_prop(nvpair_name(elem));
824 
825 		if (prop == ZPOOL_PROP_CACHEFILE ||
826 		    prop == ZPOOL_PROP_ALTROOT ||
827 		    prop == ZPOOL_PROP_READONLY)
828 			continue;
829 
830 		if (prop == ZPOOL_PROP_INVAL &&
831 		    zfs_prop_user(nvpair_name(elem))) {
832 			need_sync = B_TRUE;
833 			break;
834 		}
835 
836 		if (prop == ZPOOL_PROP_VERSION || prop == ZPOOL_PROP_INVAL) {
837 			uint64_t ver = 0;
838 
839 			if (prop == ZPOOL_PROP_VERSION) {
840 				VERIFY(nvpair_value_uint64(elem, &ver) == 0);
841 			} else {
842 				ASSERT(zpool_prop_feature(nvpair_name(elem)));
843 				ver = SPA_VERSION_FEATURES;
844 				need_sync = B_TRUE;
845 			}
846 
847 			/* Save time if the version is already set. */
848 			if (ver == spa_version(spa))
849 				continue;
850 
851 			/*
852 			 * In addition to the pool directory object, we might
853 			 * create the pool properties object, the features for
854 			 * read object, the features for write object, or the
855 			 * feature descriptions object.
856 			 */
857 			error = dsl_sync_task(spa->spa_name, NULL,
858 			    spa_sync_version, &ver,
859 			    6, ZFS_SPACE_CHECK_RESERVED);
860 			if (error)
861 				return (error);
862 			continue;
863 		}
864 
865 		need_sync = B_TRUE;
866 		break;
867 	}
868 
869 	if (need_sync) {
870 		return (dsl_sync_task(spa->spa_name, NULL, spa_sync_props,
871 		    nvp, 6, ZFS_SPACE_CHECK_RESERVED));
872 	}
873 
874 	return (0);
875 }
876 
877 /*
878  * If the bootfs property value is dsobj, clear it.
879  */
880 void
881 spa_prop_clear_bootfs(spa_t *spa, uint64_t dsobj, dmu_tx_t *tx)
882 {
883 	if (spa->spa_bootfs == dsobj && spa->spa_pool_props_object != 0) {
884 		VERIFY(zap_remove(spa->spa_meta_objset,
885 		    spa->spa_pool_props_object,
886 		    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), tx) == 0);
887 		spa->spa_bootfs = 0;
888 	}
889 }
890 
891 static int
892 spa_change_guid_check(void *arg, dmu_tx_t *tx)
893 {
894 	uint64_t *newguid __maybe_unused = arg;
895 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
896 	vdev_t *rvd = spa->spa_root_vdev;
897 	uint64_t vdev_state;
898 
899 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
900 		int error = (spa_has_checkpoint(spa)) ?
901 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
902 		return (SET_ERROR(error));
903 	}
904 
905 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
906 	vdev_state = rvd->vdev_state;
907 	spa_config_exit(spa, SCL_STATE, FTAG);
908 
909 	if (vdev_state != VDEV_STATE_HEALTHY)
910 		return (SET_ERROR(ENXIO));
911 
912 	ASSERT3U(spa_guid(spa), !=, *newguid);
913 
914 	return (0);
915 }
916 
917 static void
918 spa_change_guid_sync(void *arg, dmu_tx_t *tx)
919 {
920 	uint64_t *newguid = arg;
921 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
922 	uint64_t oldguid;
923 	vdev_t *rvd = spa->spa_root_vdev;
924 
925 	oldguid = spa_guid(spa);
926 
927 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
928 	rvd->vdev_guid = *newguid;
929 	rvd->vdev_guid_sum += (*newguid - oldguid);
930 	vdev_config_dirty(rvd);
931 	spa_config_exit(spa, SCL_STATE, FTAG);
932 
933 	spa_history_log_internal(spa, "guid change", tx, "old=%llu new=%llu",
934 	    (u_longlong_t)oldguid, (u_longlong_t)*newguid);
935 }
936 
937 /*
938  * Change the GUID for the pool.  This is done so that we can later
939  * re-import a pool built from a clone of our own vdevs.  We will modify
940  * the root vdev's guid, our own pool guid, and then mark all of our
941  * vdevs dirty.  Note that we must make sure that all our vdevs are
942  * online when we do this, or else any vdevs that weren't present
943  * would be orphaned from our pool.  We are also going to issue a
944  * sysevent to update any watchers.
945  */
946 int
947 spa_change_guid(spa_t *spa)
948 {
949 	int error;
950 	uint64_t guid;
951 
952 	mutex_enter(&spa->spa_vdev_top_lock);
953 	mutex_enter(&spa_namespace_lock);
954 	guid = spa_generate_guid(NULL);
955 
956 	error = dsl_sync_task(spa->spa_name, spa_change_guid_check,
957 	    spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED);
958 
959 	if (error == 0) {
960 		/*
961 		 * Clear the kobj flag from all the vdevs to allow
962 		 * vdev_cache_process_kobj_evt() to post events to all the
963 		 * vdevs since GUID is updated.
964 		 */
965 		vdev_clear_kobj_evt(spa->spa_root_vdev);
966 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++)
967 			vdev_clear_kobj_evt(spa->spa_l2cache.sav_vdevs[i]);
968 
969 		spa_write_cachefile(spa, B_FALSE, B_TRUE, B_TRUE);
970 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID);
971 	}
972 
973 	mutex_exit(&spa_namespace_lock);
974 	mutex_exit(&spa->spa_vdev_top_lock);
975 
976 	return (error);
977 }
978 
979 /*
980  * ==========================================================================
981  * SPA state manipulation (open/create/destroy/import/export)
982  * ==========================================================================
983  */
984 
985 static int
986 spa_error_entry_compare(const void *a, const void *b)
987 {
988 	const spa_error_entry_t *sa = (const spa_error_entry_t *)a;
989 	const spa_error_entry_t *sb = (const spa_error_entry_t *)b;
990 	int ret;
991 
992 	ret = memcmp(&sa->se_bookmark, &sb->se_bookmark,
993 	    sizeof (zbookmark_phys_t));
994 
995 	return (TREE_ISIGN(ret));
996 }
997 
998 /*
999  * Utility function which retrieves copies of the current logs and
1000  * re-initializes them in the process.
1001  */
1002 void
1003 spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub)
1004 {
1005 	ASSERT(MUTEX_HELD(&spa->spa_errlist_lock));
1006 
1007 	memcpy(last, &spa->spa_errlist_last, sizeof (avl_tree_t));
1008 	memcpy(scrub, &spa->spa_errlist_scrub, sizeof (avl_tree_t));
1009 
1010 	avl_create(&spa->spa_errlist_scrub,
1011 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1012 	    offsetof(spa_error_entry_t, se_avl));
1013 	avl_create(&spa->spa_errlist_last,
1014 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1015 	    offsetof(spa_error_entry_t, se_avl));
1016 }
1017 
1018 static void
1019 spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
1020 {
1021 	const zio_taskq_info_t *ztip = &zio_taskqs[t][q];
1022 	enum zti_modes mode = ztip->zti_mode;
1023 	uint_t value = ztip->zti_value;
1024 	uint_t count = ztip->zti_count;
1025 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1026 	uint_t cpus, flags = TASKQ_DYNAMIC;
1027 	boolean_t batch = B_FALSE;
1028 
1029 	switch (mode) {
1030 	case ZTI_MODE_FIXED:
1031 		ASSERT3U(value, >, 0);
1032 		break;
1033 
1034 	case ZTI_MODE_BATCH:
1035 		batch = B_TRUE;
1036 		flags |= TASKQ_THREADS_CPU_PCT;
1037 		value = MIN(zio_taskq_batch_pct, 100);
1038 		break;
1039 
1040 	case ZTI_MODE_SCALE:
1041 		flags |= TASKQ_THREADS_CPU_PCT;
1042 		/*
1043 		 * We want more taskqs to reduce lock contention, but we want
1044 		 * less for better request ordering and CPU utilization.
1045 		 */
1046 		cpus = MAX(1, boot_ncpus * zio_taskq_batch_pct / 100);
1047 		if (zio_taskq_batch_tpq > 0) {
1048 			count = MAX(1, (cpus + zio_taskq_batch_tpq / 2) /
1049 			    zio_taskq_batch_tpq);
1050 		} else {
1051 			/*
1052 			 * Prefer 6 threads per taskq, but no more taskqs
1053 			 * than threads in them on large systems. For 80%:
1054 			 *
1055 			 *                 taskq   taskq   total
1056 			 * cpus    taskqs  percent threads threads
1057 			 * ------- ------- ------- ------- -------
1058 			 * 1       1       80%     1       1
1059 			 * 2       1       80%     1       1
1060 			 * 4       1       80%     3       3
1061 			 * 8       2       40%     3       6
1062 			 * 16      3       27%     4       12
1063 			 * 32      5       16%     5       25
1064 			 * 64      7       11%     7       49
1065 			 * 128     10      8%      10      100
1066 			 * 256     14      6%      15      210
1067 			 */
1068 			count = 1 + cpus / 6;
1069 			while (count * count > cpus)
1070 				count--;
1071 		}
1072 		/* Limit each taskq within 100% to not trigger assertion. */
1073 		count = MAX(count, (zio_taskq_batch_pct + 99) / 100);
1074 		value = (zio_taskq_batch_pct + count / 2) / count;
1075 		break;
1076 
1077 	case ZTI_MODE_NULL:
1078 		tqs->stqs_count = 0;
1079 		tqs->stqs_taskq = NULL;
1080 		return;
1081 
1082 	default:
1083 		panic("unrecognized mode for %s_%s taskq (%u:%u) in "
1084 		    "spa_activate()",
1085 		    zio_type_name[t], zio_taskq_types[q], mode, value);
1086 		break;
1087 	}
1088 
1089 	ASSERT3U(count, >, 0);
1090 	tqs->stqs_count = count;
1091 	tqs->stqs_taskq = kmem_alloc(count * sizeof (taskq_t *), KM_SLEEP);
1092 
1093 	for (uint_t i = 0; i < count; i++) {
1094 		taskq_t *tq;
1095 		char name[32];
1096 
1097 		if (count > 1)
1098 			(void) snprintf(name, sizeof (name), "%s_%s_%u",
1099 			    zio_type_name[t], zio_taskq_types[q], i);
1100 		else
1101 			(void) snprintf(name, sizeof (name), "%s_%s",
1102 			    zio_type_name[t], zio_taskq_types[q]);
1103 
1104 		if (zio_taskq_sysdc && spa->spa_proc != &p0) {
1105 			if (batch)
1106 				flags |= TASKQ_DC_BATCH;
1107 
1108 			(void) zio_taskq_basedc;
1109 			tq = taskq_create_sysdc(name, value, 50, INT_MAX,
1110 			    spa->spa_proc, zio_taskq_basedc, flags);
1111 		} else {
1112 			pri_t pri = maxclsyspri;
1113 			/*
1114 			 * The write issue taskq can be extremely CPU
1115 			 * intensive.  Run it at slightly less important
1116 			 * priority than the other taskqs.
1117 			 *
1118 			 * Under Linux and FreeBSD this means incrementing
1119 			 * the priority value as opposed to platforms like
1120 			 * illumos where it should be decremented.
1121 			 *
1122 			 * On FreeBSD, if priorities divided by four (RQ_PPQ)
1123 			 * are equal then a difference between them is
1124 			 * insignificant.
1125 			 */
1126 			if (t == ZIO_TYPE_WRITE && q == ZIO_TASKQ_ISSUE) {
1127 #if defined(__linux__)
1128 				pri++;
1129 #elif defined(__FreeBSD__)
1130 				pri += 4;
1131 #else
1132 #error "unknown OS"
1133 #endif
1134 			}
1135 			tq = taskq_create_proc(name, value, pri, 50,
1136 			    INT_MAX, spa->spa_proc, flags);
1137 		}
1138 
1139 		tqs->stqs_taskq[i] = tq;
1140 	}
1141 }
1142 
1143 static void
1144 spa_taskqs_fini(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
1145 {
1146 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1147 
1148 	if (tqs->stqs_taskq == NULL) {
1149 		ASSERT3U(tqs->stqs_count, ==, 0);
1150 		return;
1151 	}
1152 
1153 	for (uint_t i = 0; i < tqs->stqs_count; i++) {
1154 		ASSERT3P(tqs->stqs_taskq[i], !=, NULL);
1155 		taskq_destroy(tqs->stqs_taskq[i]);
1156 	}
1157 
1158 	kmem_free(tqs->stqs_taskq, tqs->stqs_count * sizeof (taskq_t *));
1159 	tqs->stqs_taskq = NULL;
1160 }
1161 
1162 /*
1163  * Dispatch a task to the appropriate taskq for the ZFS I/O type and priority.
1164  * Note that a type may have multiple discrete taskqs to avoid lock contention
1165  * on the taskq itself. In that case we choose which taskq at random by using
1166  * the low bits of gethrtime().
1167  */
1168 void
1169 spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
1170     task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent)
1171 {
1172 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1173 	taskq_t *tq;
1174 
1175 	ASSERT3P(tqs->stqs_taskq, !=, NULL);
1176 	ASSERT3U(tqs->stqs_count, !=, 0);
1177 
1178 	if (tqs->stqs_count == 1) {
1179 		tq = tqs->stqs_taskq[0];
1180 	} else {
1181 		tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count];
1182 	}
1183 
1184 	taskq_dispatch_ent(tq, func, arg, flags, ent);
1185 }
1186 
1187 /*
1188  * Same as spa_taskq_dispatch_ent() but block on the task until completion.
1189  */
1190 void
1191 spa_taskq_dispatch_sync(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
1192     task_func_t *func, void *arg, uint_t flags)
1193 {
1194 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1195 	taskq_t *tq;
1196 	taskqid_t id;
1197 
1198 	ASSERT3P(tqs->stqs_taskq, !=, NULL);
1199 	ASSERT3U(tqs->stqs_count, !=, 0);
1200 
1201 	if (tqs->stqs_count == 1) {
1202 		tq = tqs->stqs_taskq[0];
1203 	} else {
1204 		tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count];
1205 	}
1206 
1207 	id = taskq_dispatch(tq, func, arg, flags);
1208 	if (id)
1209 		taskq_wait_id(tq, id);
1210 }
1211 
1212 static void
1213 spa_create_zio_taskqs(spa_t *spa)
1214 {
1215 	for (int t = 0; t < ZIO_TYPES; t++) {
1216 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1217 			spa_taskqs_init(spa, t, q);
1218 		}
1219 	}
1220 }
1221 
1222 /*
1223  * Disabled until spa_thread() can be adapted for Linux.
1224  */
1225 #undef HAVE_SPA_THREAD
1226 
1227 #if defined(_KERNEL) && defined(HAVE_SPA_THREAD)
1228 static void
1229 spa_thread(void *arg)
1230 {
1231 	psetid_t zio_taskq_psrset_bind = PS_NONE;
1232 	callb_cpr_t cprinfo;
1233 
1234 	spa_t *spa = arg;
1235 	user_t *pu = PTOU(curproc);
1236 
1237 	CALLB_CPR_INIT(&cprinfo, &spa->spa_proc_lock, callb_generic_cpr,
1238 	    spa->spa_name);
1239 
1240 	ASSERT(curproc != &p0);
1241 	(void) snprintf(pu->u_psargs, sizeof (pu->u_psargs),
1242 	    "zpool-%s", spa->spa_name);
1243 	(void) strlcpy(pu->u_comm, pu->u_psargs, sizeof (pu->u_comm));
1244 
1245 	/* bind this thread to the requested psrset */
1246 	if (zio_taskq_psrset_bind != PS_NONE) {
1247 		pool_lock();
1248 		mutex_enter(&cpu_lock);
1249 		mutex_enter(&pidlock);
1250 		mutex_enter(&curproc->p_lock);
1251 
1252 		if (cpupart_bind_thread(curthread, zio_taskq_psrset_bind,
1253 		    0, NULL, NULL) == 0)  {
1254 			curthread->t_bind_pset = zio_taskq_psrset_bind;
1255 		} else {
1256 			cmn_err(CE_WARN,
1257 			    "Couldn't bind process for zfs pool \"%s\" to "
1258 			    "pset %d\n", spa->spa_name, zio_taskq_psrset_bind);
1259 		}
1260 
1261 		mutex_exit(&curproc->p_lock);
1262 		mutex_exit(&pidlock);
1263 		mutex_exit(&cpu_lock);
1264 		pool_unlock();
1265 	}
1266 
1267 	if (zio_taskq_sysdc) {
1268 		sysdc_thread_enter(curthread, 100, 0);
1269 	}
1270 
1271 	spa->spa_proc = curproc;
1272 	spa->spa_did = curthread->t_did;
1273 
1274 	spa_create_zio_taskqs(spa);
1275 
1276 	mutex_enter(&spa->spa_proc_lock);
1277 	ASSERT(spa->spa_proc_state == SPA_PROC_CREATED);
1278 
1279 	spa->spa_proc_state = SPA_PROC_ACTIVE;
1280 	cv_broadcast(&spa->spa_proc_cv);
1281 
1282 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
1283 	while (spa->spa_proc_state == SPA_PROC_ACTIVE)
1284 		cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1285 	CALLB_CPR_SAFE_END(&cprinfo, &spa->spa_proc_lock);
1286 
1287 	ASSERT(spa->spa_proc_state == SPA_PROC_DEACTIVATE);
1288 	spa->spa_proc_state = SPA_PROC_GONE;
1289 	spa->spa_proc = &p0;
1290 	cv_broadcast(&spa->spa_proc_cv);
1291 	CALLB_CPR_EXIT(&cprinfo);	/* drops spa_proc_lock */
1292 
1293 	mutex_enter(&curproc->p_lock);
1294 	lwp_exit();
1295 }
1296 #endif
1297 
1298 /*
1299  * Activate an uninitialized pool.
1300  */
1301 static void
1302 spa_activate(spa_t *spa, spa_mode_t mode)
1303 {
1304 	ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED);
1305 
1306 	spa->spa_state = POOL_STATE_ACTIVE;
1307 	spa->spa_mode = mode;
1308 	spa->spa_read_spacemaps = spa_mode_readable_spacemaps;
1309 
1310 	spa->spa_normal_class = metaslab_class_create(spa, &zfs_metaslab_ops);
1311 	spa->spa_log_class = metaslab_class_create(spa, &zfs_metaslab_ops);
1312 	spa->spa_embedded_log_class =
1313 	    metaslab_class_create(spa, &zfs_metaslab_ops);
1314 	spa->spa_special_class = metaslab_class_create(spa, &zfs_metaslab_ops);
1315 	spa->spa_dedup_class = metaslab_class_create(spa, &zfs_metaslab_ops);
1316 
1317 	/* Try to create a covering process */
1318 	mutex_enter(&spa->spa_proc_lock);
1319 	ASSERT(spa->spa_proc_state == SPA_PROC_NONE);
1320 	ASSERT(spa->spa_proc == &p0);
1321 	spa->spa_did = 0;
1322 
1323 	(void) spa_create_process;
1324 #ifdef HAVE_SPA_THREAD
1325 	/* Only create a process if we're going to be around a while. */
1326 	if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) {
1327 		if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri,
1328 		    NULL, 0) == 0) {
1329 			spa->spa_proc_state = SPA_PROC_CREATED;
1330 			while (spa->spa_proc_state == SPA_PROC_CREATED) {
1331 				cv_wait(&spa->spa_proc_cv,
1332 				    &spa->spa_proc_lock);
1333 			}
1334 			ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1335 			ASSERT(spa->spa_proc != &p0);
1336 			ASSERT(spa->spa_did != 0);
1337 		} else {
1338 #ifdef _KERNEL
1339 			cmn_err(CE_WARN,
1340 			    "Couldn't create process for zfs pool \"%s\"\n",
1341 			    spa->spa_name);
1342 #endif
1343 		}
1344 	}
1345 #endif /* HAVE_SPA_THREAD */
1346 	mutex_exit(&spa->spa_proc_lock);
1347 
1348 	/* If we didn't create a process, we need to create our taskqs. */
1349 	if (spa->spa_proc == &p0) {
1350 		spa_create_zio_taskqs(spa);
1351 	}
1352 
1353 	for (size_t i = 0; i < TXG_SIZE; i++) {
1354 		spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL,
1355 		    ZIO_FLAG_CANFAIL);
1356 	}
1357 
1358 	list_create(&spa->spa_config_dirty_list, sizeof (vdev_t),
1359 	    offsetof(vdev_t, vdev_config_dirty_node));
1360 	list_create(&spa->spa_evicting_os_list, sizeof (objset_t),
1361 	    offsetof(objset_t, os_evicting_node));
1362 	list_create(&spa->spa_state_dirty_list, sizeof (vdev_t),
1363 	    offsetof(vdev_t, vdev_state_dirty_node));
1364 
1365 	txg_list_create(&spa->spa_vdev_txg_list, spa,
1366 	    offsetof(struct vdev, vdev_txg_node));
1367 
1368 	avl_create(&spa->spa_errlist_scrub,
1369 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1370 	    offsetof(spa_error_entry_t, se_avl));
1371 	avl_create(&spa->spa_errlist_last,
1372 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1373 	    offsetof(spa_error_entry_t, se_avl));
1374 	avl_create(&spa->spa_errlist_healed,
1375 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1376 	    offsetof(spa_error_entry_t, se_avl));
1377 
1378 	spa_activate_os(spa);
1379 
1380 	spa_keystore_init(&spa->spa_keystore);
1381 
1382 	/*
1383 	 * This taskq is used to perform zvol-minor-related tasks
1384 	 * asynchronously. This has several advantages, including easy
1385 	 * resolution of various deadlocks.
1386 	 *
1387 	 * The taskq must be single threaded to ensure tasks are always
1388 	 * processed in the order in which they were dispatched.
1389 	 *
1390 	 * A taskq per pool allows one to keep the pools independent.
1391 	 * This way if one pool is suspended, it will not impact another.
1392 	 *
1393 	 * The preferred location to dispatch a zvol minor task is a sync
1394 	 * task. In this context, there is easy access to the spa_t and minimal
1395 	 * error handling is required because the sync task must succeed.
1396 	 */
1397 	spa->spa_zvol_taskq = taskq_create("z_zvol", 1, defclsyspri,
1398 	    1, INT_MAX, 0);
1399 
1400 	/*
1401 	 * Taskq dedicated to prefetcher threads: this is used to prevent the
1402 	 * pool traverse code from monopolizing the global (and limited)
1403 	 * system_taskq by inappropriately scheduling long running tasks on it.
1404 	 */
1405 	spa->spa_prefetch_taskq = taskq_create("z_prefetch", 100,
1406 	    defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT);
1407 
1408 	/*
1409 	 * The taskq to upgrade datasets in this pool. Currently used by
1410 	 * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA.
1411 	 */
1412 	spa->spa_upgrade_taskq = taskq_create("z_upgrade", 100,
1413 	    defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT);
1414 }
1415 
1416 /*
1417  * Opposite of spa_activate().
1418  */
1419 static void
1420 spa_deactivate(spa_t *spa)
1421 {
1422 	ASSERT(spa->spa_sync_on == B_FALSE);
1423 	ASSERT(spa->spa_dsl_pool == NULL);
1424 	ASSERT(spa->spa_root_vdev == NULL);
1425 	ASSERT(spa->spa_async_zio_root == NULL);
1426 	ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED);
1427 
1428 	spa_evicting_os_wait(spa);
1429 
1430 	if (spa->spa_zvol_taskq) {
1431 		taskq_destroy(spa->spa_zvol_taskq);
1432 		spa->spa_zvol_taskq = NULL;
1433 	}
1434 
1435 	if (spa->spa_prefetch_taskq) {
1436 		taskq_destroy(spa->spa_prefetch_taskq);
1437 		spa->spa_prefetch_taskq = NULL;
1438 	}
1439 
1440 	if (spa->spa_upgrade_taskq) {
1441 		taskq_destroy(spa->spa_upgrade_taskq);
1442 		spa->spa_upgrade_taskq = NULL;
1443 	}
1444 
1445 	txg_list_destroy(&spa->spa_vdev_txg_list);
1446 
1447 	list_destroy(&spa->spa_config_dirty_list);
1448 	list_destroy(&spa->spa_evicting_os_list);
1449 	list_destroy(&spa->spa_state_dirty_list);
1450 
1451 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
1452 
1453 	for (int t = 0; t < ZIO_TYPES; t++) {
1454 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1455 			spa_taskqs_fini(spa, t, q);
1456 		}
1457 	}
1458 
1459 	for (size_t i = 0; i < TXG_SIZE; i++) {
1460 		ASSERT3P(spa->spa_txg_zio[i], !=, NULL);
1461 		VERIFY0(zio_wait(spa->spa_txg_zio[i]));
1462 		spa->spa_txg_zio[i] = NULL;
1463 	}
1464 
1465 	metaslab_class_destroy(spa->spa_normal_class);
1466 	spa->spa_normal_class = NULL;
1467 
1468 	metaslab_class_destroy(spa->spa_log_class);
1469 	spa->spa_log_class = NULL;
1470 
1471 	metaslab_class_destroy(spa->spa_embedded_log_class);
1472 	spa->spa_embedded_log_class = NULL;
1473 
1474 	metaslab_class_destroy(spa->spa_special_class);
1475 	spa->spa_special_class = NULL;
1476 
1477 	metaslab_class_destroy(spa->spa_dedup_class);
1478 	spa->spa_dedup_class = NULL;
1479 
1480 	/*
1481 	 * If this was part of an import or the open otherwise failed, we may
1482 	 * still have errors left in the queues.  Empty them just in case.
1483 	 */
1484 	spa_errlog_drain(spa);
1485 	avl_destroy(&spa->spa_errlist_scrub);
1486 	avl_destroy(&spa->spa_errlist_last);
1487 	avl_destroy(&spa->spa_errlist_healed);
1488 
1489 	spa_keystore_fini(&spa->spa_keystore);
1490 
1491 	spa->spa_state = POOL_STATE_UNINITIALIZED;
1492 
1493 	mutex_enter(&spa->spa_proc_lock);
1494 	if (spa->spa_proc_state != SPA_PROC_NONE) {
1495 		ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1496 		spa->spa_proc_state = SPA_PROC_DEACTIVATE;
1497 		cv_broadcast(&spa->spa_proc_cv);
1498 		while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) {
1499 			ASSERT(spa->spa_proc != &p0);
1500 			cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1501 		}
1502 		ASSERT(spa->spa_proc_state == SPA_PROC_GONE);
1503 		spa->spa_proc_state = SPA_PROC_NONE;
1504 	}
1505 	ASSERT(spa->spa_proc == &p0);
1506 	mutex_exit(&spa->spa_proc_lock);
1507 
1508 	/*
1509 	 * We want to make sure spa_thread() has actually exited the ZFS
1510 	 * module, so that the module can't be unloaded out from underneath
1511 	 * it.
1512 	 */
1513 	if (spa->spa_did != 0) {
1514 		thread_join(spa->spa_did);
1515 		spa->spa_did = 0;
1516 	}
1517 
1518 	spa_deactivate_os(spa);
1519 
1520 }
1521 
1522 /*
1523  * Verify a pool configuration, and construct the vdev tree appropriately.  This
1524  * will create all the necessary vdevs in the appropriate layout, with each vdev
1525  * in the CLOSED state.  This will prep the pool before open/creation/import.
1526  * All vdev validation is done by the vdev_alloc() routine.
1527  */
1528 int
1529 spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent,
1530     uint_t id, int atype)
1531 {
1532 	nvlist_t **child;
1533 	uint_t children;
1534 	int error;
1535 
1536 	if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0)
1537 		return (error);
1538 
1539 	if ((*vdp)->vdev_ops->vdev_op_leaf)
1540 		return (0);
1541 
1542 	error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1543 	    &child, &children);
1544 
1545 	if (error == ENOENT)
1546 		return (0);
1547 
1548 	if (error) {
1549 		vdev_free(*vdp);
1550 		*vdp = NULL;
1551 		return (SET_ERROR(EINVAL));
1552 	}
1553 
1554 	for (int c = 0; c < children; c++) {
1555 		vdev_t *vd;
1556 		if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c,
1557 		    atype)) != 0) {
1558 			vdev_free(*vdp);
1559 			*vdp = NULL;
1560 			return (error);
1561 		}
1562 	}
1563 
1564 	ASSERT(*vdp != NULL);
1565 
1566 	return (0);
1567 }
1568 
1569 static boolean_t
1570 spa_should_flush_logs_on_unload(spa_t *spa)
1571 {
1572 	if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))
1573 		return (B_FALSE);
1574 
1575 	if (!spa_writeable(spa))
1576 		return (B_FALSE);
1577 
1578 	if (!spa->spa_sync_on)
1579 		return (B_FALSE);
1580 
1581 	if (spa_state(spa) != POOL_STATE_EXPORTED)
1582 		return (B_FALSE);
1583 
1584 	if (zfs_keep_log_spacemaps_at_export)
1585 		return (B_FALSE);
1586 
1587 	return (B_TRUE);
1588 }
1589 
1590 /*
1591  * Opens a transaction that will set the flag that will instruct
1592  * spa_sync to attempt to flush all the metaslabs for that txg.
1593  */
1594 static void
1595 spa_unload_log_sm_flush_all(spa_t *spa)
1596 {
1597 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1598 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1599 
1600 	ASSERT3U(spa->spa_log_flushall_txg, ==, 0);
1601 	spa->spa_log_flushall_txg = dmu_tx_get_txg(tx);
1602 
1603 	dmu_tx_commit(tx);
1604 	txg_wait_synced(spa_get_dsl(spa), spa->spa_log_flushall_txg);
1605 }
1606 
1607 static void
1608 spa_unload_log_sm_metadata(spa_t *spa)
1609 {
1610 	void *cookie = NULL;
1611 	spa_log_sm_t *sls;
1612 	log_summary_entry_t *e;
1613 
1614 	while ((sls = avl_destroy_nodes(&spa->spa_sm_logs_by_txg,
1615 	    &cookie)) != NULL) {
1616 		VERIFY0(sls->sls_mscount);
1617 		kmem_free(sls, sizeof (spa_log_sm_t));
1618 	}
1619 
1620 	while ((e = list_remove_head(&spa->spa_log_summary)) != NULL) {
1621 		VERIFY0(e->lse_mscount);
1622 		kmem_free(e, sizeof (log_summary_entry_t));
1623 	}
1624 
1625 	spa->spa_unflushed_stats.sus_nblocks = 0;
1626 	spa->spa_unflushed_stats.sus_memused = 0;
1627 	spa->spa_unflushed_stats.sus_blocklimit = 0;
1628 }
1629 
1630 static void
1631 spa_destroy_aux_threads(spa_t *spa)
1632 {
1633 	if (spa->spa_condense_zthr != NULL) {
1634 		zthr_destroy(spa->spa_condense_zthr);
1635 		spa->spa_condense_zthr = NULL;
1636 	}
1637 	if (spa->spa_checkpoint_discard_zthr != NULL) {
1638 		zthr_destroy(spa->spa_checkpoint_discard_zthr);
1639 		spa->spa_checkpoint_discard_zthr = NULL;
1640 	}
1641 	if (spa->spa_livelist_delete_zthr != NULL) {
1642 		zthr_destroy(spa->spa_livelist_delete_zthr);
1643 		spa->spa_livelist_delete_zthr = NULL;
1644 	}
1645 	if (spa->spa_livelist_condense_zthr != NULL) {
1646 		zthr_destroy(spa->spa_livelist_condense_zthr);
1647 		spa->spa_livelist_condense_zthr = NULL;
1648 	}
1649 }
1650 
1651 /*
1652  * Opposite of spa_load().
1653  */
1654 static void
1655 spa_unload(spa_t *spa)
1656 {
1657 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1658 	ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED);
1659 
1660 	spa_import_progress_remove(spa_guid(spa));
1661 	spa_load_note(spa, "UNLOADING");
1662 
1663 	spa_wake_waiters(spa);
1664 
1665 	/*
1666 	 * If we have set the spa_final_txg, we have already performed the
1667 	 * tasks below in spa_export_common(). We should not redo it here since
1668 	 * we delay the final TXGs beyond what spa_final_txg is set at.
1669 	 */
1670 	if (spa->spa_final_txg == UINT64_MAX) {
1671 		/*
1672 		 * If the log space map feature is enabled and the pool is
1673 		 * getting exported (but not destroyed), we want to spend some
1674 		 * time flushing as many metaslabs as we can in an attempt to
1675 		 * destroy log space maps and save import time.
1676 		 */
1677 		if (spa_should_flush_logs_on_unload(spa))
1678 			spa_unload_log_sm_flush_all(spa);
1679 
1680 		/*
1681 		 * Stop async tasks.
1682 		 */
1683 		spa_async_suspend(spa);
1684 
1685 		if (spa->spa_root_vdev) {
1686 			vdev_t *root_vdev = spa->spa_root_vdev;
1687 			vdev_initialize_stop_all(root_vdev,
1688 			    VDEV_INITIALIZE_ACTIVE);
1689 			vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE);
1690 			vdev_autotrim_stop_all(spa);
1691 			vdev_rebuild_stop_all(spa);
1692 		}
1693 	}
1694 
1695 	/*
1696 	 * Stop syncing.
1697 	 */
1698 	if (spa->spa_sync_on) {
1699 		txg_sync_stop(spa->spa_dsl_pool);
1700 		spa->spa_sync_on = B_FALSE;
1701 	}
1702 
1703 	/*
1704 	 * This ensures that there is no async metaslab prefetching
1705 	 * while we attempt to unload the spa.
1706 	 */
1707 	if (spa->spa_root_vdev != NULL) {
1708 		for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++) {
1709 			vdev_t *vc = spa->spa_root_vdev->vdev_child[c];
1710 			if (vc->vdev_mg != NULL)
1711 				taskq_wait(vc->vdev_mg->mg_taskq);
1712 		}
1713 	}
1714 
1715 	if (spa->spa_mmp.mmp_thread)
1716 		mmp_thread_stop(spa);
1717 
1718 	/*
1719 	 * Wait for any outstanding async I/O to complete.
1720 	 */
1721 	if (spa->spa_async_zio_root != NULL) {
1722 		for (int i = 0; i < max_ncpus; i++)
1723 			(void) zio_wait(spa->spa_async_zio_root[i]);
1724 		kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *));
1725 		spa->spa_async_zio_root = NULL;
1726 	}
1727 
1728 	if (spa->spa_vdev_removal != NULL) {
1729 		spa_vdev_removal_destroy(spa->spa_vdev_removal);
1730 		spa->spa_vdev_removal = NULL;
1731 	}
1732 
1733 	spa_destroy_aux_threads(spa);
1734 
1735 	spa_condense_fini(spa);
1736 
1737 	bpobj_close(&spa->spa_deferred_bpobj);
1738 
1739 	spa_config_enter(spa, SCL_ALL, spa, RW_WRITER);
1740 
1741 	/*
1742 	 * Close all vdevs.
1743 	 */
1744 	if (spa->spa_root_vdev)
1745 		vdev_free(spa->spa_root_vdev);
1746 	ASSERT(spa->spa_root_vdev == NULL);
1747 
1748 	/*
1749 	 * Close the dsl pool.
1750 	 */
1751 	if (spa->spa_dsl_pool) {
1752 		dsl_pool_close(spa->spa_dsl_pool);
1753 		spa->spa_dsl_pool = NULL;
1754 		spa->spa_meta_objset = NULL;
1755 	}
1756 
1757 	ddt_unload(spa);
1758 	brt_unload(spa);
1759 	spa_unload_log_sm_metadata(spa);
1760 
1761 	/*
1762 	 * Drop and purge level 2 cache
1763 	 */
1764 	spa_l2cache_drop(spa);
1765 
1766 	if (spa->spa_spares.sav_vdevs) {
1767 		for (int i = 0; i < spa->spa_spares.sav_count; i++)
1768 			vdev_free(spa->spa_spares.sav_vdevs[i]);
1769 		kmem_free(spa->spa_spares.sav_vdevs,
1770 		    spa->spa_spares.sav_count * sizeof (void *));
1771 		spa->spa_spares.sav_vdevs = NULL;
1772 	}
1773 	if (spa->spa_spares.sav_config) {
1774 		nvlist_free(spa->spa_spares.sav_config);
1775 		spa->spa_spares.sav_config = NULL;
1776 	}
1777 	spa->spa_spares.sav_count = 0;
1778 
1779 	if (spa->spa_l2cache.sav_vdevs) {
1780 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
1781 			vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]);
1782 			vdev_free(spa->spa_l2cache.sav_vdevs[i]);
1783 		}
1784 		kmem_free(spa->spa_l2cache.sav_vdevs,
1785 		    spa->spa_l2cache.sav_count * sizeof (void *));
1786 		spa->spa_l2cache.sav_vdevs = NULL;
1787 	}
1788 	if (spa->spa_l2cache.sav_config) {
1789 		nvlist_free(spa->spa_l2cache.sav_config);
1790 		spa->spa_l2cache.sav_config = NULL;
1791 	}
1792 	spa->spa_l2cache.sav_count = 0;
1793 
1794 	spa->spa_async_suspended = 0;
1795 
1796 	spa->spa_indirect_vdevs_loaded = B_FALSE;
1797 
1798 	if (spa->spa_comment != NULL) {
1799 		spa_strfree(spa->spa_comment);
1800 		spa->spa_comment = NULL;
1801 	}
1802 	if (spa->spa_compatibility != NULL) {
1803 		spa_strfree(spa->spa_compatibility);
1804 		spa->spa_compatibility = NULL;
1805 	}
1806 
1807 	spa_config_exit(spa, SCL_ALL, spa);
1808 }
1809 
1810 /*
1811  * Load (or re-load) the current list of vdevs describing the active spares for
1812  * this pool.  When this is called, we have some form of basic information in
1813  * 'spa_spares.sav_config'.  We parse this into vdevs, try to open them, and
1814  * then re-generate a more complete list including status information.
1815  */
1816 void
1817 spa_load_spares(spa_t *spa)
1818 {
1819 	nvlist_t **spares;
1820 	uint_t nspares;
1821 	int i;
1822 	vdev_t *vd, *tvd;
1823 
1824 #ifndef _KERNEL
1825 	/*
1826 	 * zdb opens both the current state of the pool and the
1827 	 * checkpointed state (if present), with a different spa_t.
1828 	 *
1829 	 * As spare vdevs are shared among open pools, we skip loading
1830 	 * them when we load the checkpointed state of the pool.
1831 	 */
1832 	if (!spa_writeable(spa))
1833 		return;
1834 #endif
1835 
1836 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1837 
1838 	/*
1839 	 * First, close and free any existing spare vdevs.
1840 	 */
1841 	if (spa->spa_spares.sav_vdevs) {
1842 		for (i = 0; i < spa->spa_spares.sav_count; i++) {
1843 			vd = spa->spa_spares.sav_vdevs[i];
1844 
1845 			/* Undo the call to spa_activate() below */
1846 			if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1847 			    B_FALSE)) != NULL && tvd->vdev_isspare)
1848 				spa_spare_remove(tvd);
1849 			vdev_close(vd);
1850 			vdev_free(vd);
1851 		}
1852 
1853 		kmem_free(spa->spa_spares.sav_vdevs,
1854 		    spa->spa_spares.sav_count * sizeof (void *));
1855 	}
1856 
1857 	if (spa->spa_spares.sav_config == NULL)
1858 		nspares = 0;
1859 	else
1860 		VERIFY0(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
1861 		    ZPOOL_CONFIG_SPARES, &spares, &nspares));
1862 
1863 	spa->spa_spares.sav_count = (int)nspares;
1864 	spa->spa_spares.sav_vdevs = NULL;
1865 
1866 	if (nspares == 0)
1867 		return;
1868 
1869 	/*
1870 	 * Construct the array of vdevs, opening them to get status in the
1871 	 * process.   For each spare, there is potentially two different vdev_t
1872 	 * structures associated with it: one in the list of spares (used only
1873 	 * for basic validation purposes) and one in the active vdev
1874 	 * configuration (if it's spared in).  During this phase we open and
1875 	 * validate each vdev on the spare list.  If the vdev also exists in the
1876 	 * active configuration, then we also mark this vdev as an active spare.
1877 	 */
1878 	spa->spa_spares.sav_vdevs = kmem_zalloc(nspares * sizeof (void *),
1879 	    KM_SLEEP);
1880 	for (i = 0; i < spa->spa_spares.sav_count; i++) {
1881 		VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0,
1882 		    VDEV_ALLOC_SPARE) == 0);
1883 		ASSERT(vd != NULL);
1884 
1885 		spa->spa_spares.sav_vdevs[i] = vd;
1886 
1887 		if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1888 		    B_FALSE)) != NULL) {
1889 			if (!tvd->vdev_isspare)
1890 				spa_spare_add(tvd);
1891 
1892 			/*
1893 			 * We only mark the spare active if we were successfully
1894 			 * able to load the vdev.  Otherwise, importing a pool
1895 			 * with a bad active spare would result in strange
1896 			 * behavior, because multiple pool would think the spare
1897 			 * is actively in use.
1898 			 *
1899 			 * There is a vulnerability here to an equally bizarre
1900 			 * circumstance, where a dead active spare is later
1901 			 * brought back to life (onlined or otherwise).  Given
1902 			 * the rarity of this scenario, and the extra complexity
1903 			 * it adds, we ignore the possibility.
1904 			 */
1905 			if (!vdev_is_dead(tvd))
1906 				spa_spare_activate(tvd);
1907 		}
1908 
1909 		vd->vdev_top = vd;
1910 		vd->vdev_aux = &spa->spa_spares;
1911 
1912 		if (vdev_open(vd) != 0)
1913 			continue;
1914 
1915 		if (vdev_validate_aux(vd) == 0)
1916 			spa_spare_add(vd);
1917 	}
1918 
1919 	/*
1920 	 * Recompute the stashed list of spares, with status information
1921 	 * this time.
1922 	 */
1923 	fnvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES);
1924 
1925 	spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *),
1926 	    KM_SLEEP);
1927 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1928 		spares[i] = vdev_config_generate(spa,
1929 		    spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE);
1930 	fnvlist_add_nvlist_array(spa->spa_spares.sav_config,
1931 	    ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares,
1932 	    spa->spa_spares.sav_count);
1933 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1934 		nvlist_free(spares[i]);
1935 	kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *));
1936 }
1937 
1938 /*
1939  * Load (or re-load) the current list of vdevs describing the active l2cache for
1940  * this pool.  When this is called, we have some form of basic information in
1941  * 'spa_l2cache.sav_config'.  We parse this into vdevs, try to open them, and
1942  * then re-generate a more complete list including status information.
1943  * Devices which are already active have their details maintained, and are
1944  * not re-opened.
1945  */
1946 void
1947 spa_load_l2cache(spa_t *spa)
1948 {
1949 	nvlist_t **l2cache = NULL;
1950 	uint_t nl2cache;
1951 	int i, j, oldnvdevs;
1952 	uint64_t guid;
1953 	vdev_t *vd, **oldvdevs, **newvdevs;
1954 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
1955 
1956 #ifndef _KERNEL
1957 	/*
1958 	 * zdb opens both the current state of the pool and the
1959 	 * checkpointed state (if present), with a different spa_t.
1960 	 *
1961 	 * As L2 caches are part of the ARC which is shared among open
1962 	 * pools, we skip loading them when we load the checkpointed
1963 	 * state of the pool.
1964 	 */
1965 	if (!spa_writeable(spa))
1966 		return;
1967 #endif
1968 
1969 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1970 
1971 	oldvdevs = sav->sav_vdevs;
1972 	oldnvdevs = sav->sav_count;
1973 	sav->sav_vdevs = NULL;
1974 	sav->sav_count = 0;
1975 
1976 	if (sav->sav_config == NULL) {
1977 		nl2cache = 0;
1978 		newvdevs = NULL;
1979 		goto out;
1980 	}
1981 
1982 	VERIFY0(nvlist_lookup_nvlist_array(sav->sav_config,
1983 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache));
1984 	newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP);
1985 
1986 	/*
1987 	 * Process new nvlist of vdevs.
1988 	 */
1989 	for (i = 0; i < nl2cache; i++) {
1990 		guid = fnvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID);
1991 
1992 		newvdevs[i] = NULL;
1993 		for (j = 0; j < oldnvdevs; j++) {
1994 			vd = oldvdevs[j];
1995 			if (vd != NULL && guid == vd->vdev_guid) {
1996 				/*
1997 				 * Retain previous vdev for add/remove ops.
1998 				 */
1999 				newvdevs[i] = vd;
2000 				oldvdevs[j] = NULL;
2001 				break;
2002 			}
2003 		}
2004 
2005 		if (newvdevs[i] == NULL) {
2006 			/*
2007 			 * Create new vdev
2008 			 */
2009 			VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0,
2010 			    VDEV_ALLOC_L2CACHE) == 0);
2011 			ASSERT(vd != NULL);
2012 			newvdevs[i] = vd;
2013 
2014 			/*
2015 			 * Commit this vdev as an l2cache device,
2016 			 * even if it fails to open.
2017 			 */
2018 			spa_l2cache_add(vd);
2019 
2020 			vd->vdev_top = vd;
2021 			vd->vdev_aux = sav;
2022 
2023 			spa_l2cache_activate(vd);
2024 
2025 			if (vdev_open(vd) != 0)
2026 				continue;
2027 
2028 			(void) vdev_validate_aux(vd);
2029 
2030 			if (!vdev_is_dead(vd))
2031 				l2arc_add_vdev(spa, vd);
2032 
2033 			/*
2034 			 * Upon cache device addition to a pool or pool
2035 			 * creation with a cache device or if the header
2036 			 * of the device is invalid we issue an async
2037 			 * TRIM command for the whole device which will
2038 			 * execute if l2arc_trim_ahead > 0.
2039 			 */
2040 			spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM);
2041 		}
2042 	}
2043 
2044 	sav->sav_vdevs = newvdevs;
2045 	sav->sav_count = (int)nl2cache;
2046 
2047 	/*
2048 	 * Recompute the stashed list of l2cache devices, with status
2049 	 * information this time.
2050 	 */
2051 	fnvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE);
2052 
2053 	if (sav->sav_count > 0)
2054 		l2cache = kmem_alloc(sav->sav_count * sizeof (void *),
2055 		    KM_SLEEP);
2056 	for (i = 0; i < sav->sav_count; i++)
2057 		l2cache[i] = vdev_config_generate(spa,
2058 		    sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE);
2059 	fnvlist_add_nvlist_array(sav->sav_config, ZPOOL_CONFIG_L2CACHE,
2060 	    (const nvlist_t * const *)l2cache, sav->sav_count);
2061 
2062 out:
2063 	/*
2064 	 * Purge vdevs that were dropped
2065 	 */
2066 	if (oldvdevs) {
2067 		for (i = 0; i < oldnvdevs; i++) {
2068 			uint64_t pool;
2069 
2070 			vd = oldvdevs[i];
2071 			if (vd != NULL) {
2072 				ASSERT(vd->vdev_isl2cache);
2073 
2074 				if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
2075 				    pool != 0ULL && l2arc_vdev_present(vd))
2076 					l2arc_remove_vdev(vd);
2077 				vdev_clear_stats(vd);
2078 				vdev_free(vd);
2079 			}
2080 		}
2081 
2082 		kmem_free(oldvdevs, oldnvdevs * sizeof (void *));
2083 	}
2084 
2085 	for (i = 0; i < sav->sav_count; i++)
2086 		nvlist_free(l2cache[i]);
2087 	if (sav->sav_count)
2088 		kmem_free(l2cache, sav->sav_count * sizeof (void *));
2089 }
2090 
2091 static int
2092 load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value)
2093 {
2094 	dmu_buf_t *db;
2095 	char *packed = NULL;
2096 	size_t nvsize = 0;
2097 	int error;
2098 	*value = NULL;
2099 
2100 	error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db);
2101 	if (error)
2102 		return (error);
2103 
2104 	nvsize = *(uint64_t *)db->db_data;
2105 	dmu_buf_rele(db, FTAG);
2106 
2107 	packed = vmem_alloc(nvsize, KM_SLEEP);
2108 	error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed,
2109 	    DMU_READ_PREFETCH);
2110 	if (error == 0)
2111 		error = nvlist_unpack(packed, nvsize, value, 0);
2112 	vmem_free(packed, nvsize);
2113 
2114 	return (error);
2115 }
2116 
2117 /*
2118  * Concrete top-level vdevs that are not missing and are not logs. At every
2119  * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds.
2120  */
2121 static uint64_t
2122 spa_healthy_core_tvds(spa_t *spa)
2123 {
2124 	vdev_t *rvd = spa->spa_root_vdev;
2125 	uint64_t tvds = 0;
2126 
2127 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
2128 		vdev_t *vd = rvd->vdev_child[i];
2129 		if (vd->vdev_islog)
2130 			continue;
2131 		if (vdev_is_concrete(vd) && !vdev_is_dead(vd))
2132 			tvds++;
2133 	}
2134 
2135 	return (tvds);
2136 }
2137 
2138 /*
2139  * Checks to see if the given vdev could not be opened, in which case we post a
2140  * sysevent to notify the autoreplace code that the device has been removed.
2141  */
2142 static void
2143 spa_check_removed(vdev_t *vd)
2144 {
2145 	for (uint64_t c = 0; c < vd->vdev_children; c++)
2146 		spa_check_removed(vd->vdev_child[c]);
2147 
2148 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) &&
2149 	    vdev_is_concrete(vd)) {
2150 		zfs_post_autoreplace(vd->vdev_spa, vd);
2151 		spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK);
2152 	}
2153 }
2154 
2155 static int
2156 spa_check_for_missing_logs(spa_t *spa)
2157 {
2158 	vdev_t *rvd = spa->spa_root_vdev;
2159 
2160 	/*
2161 	 * If we're doing a normal import, then build up any additional
2162 	 * diagnostic information about missing log devices.
2163 	 * We'll pass this up to the user for further processing.
2164 	 */
2165 	if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) {
2166 		nvlist_t **child, *nv;
2167 		uint64_t idx = 0;
2168 
2169 		child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t *),
2170 		    KM_SLEEP);
2171 		nv = fnvlist_alloc();
2172 
2173 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
2174 			vdev_t *tvd = rvd->vdev_child[c];
2175 
2176 			/*
2177 			 * We consider a device as missing only if it failed
2178 			 * to open (i.e. offline or faulted is not considered
2179 			 * as missing).
2180 			 */
2181 			if (tvd->vdev_islog &&
2182 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
2183 				child[idx++] = vdev_config_generate(spa, tvd,
2184 				    B_FALSE, VDEV_CONFIG_MISSING);
2185 			}
2186 		}
2187 
2188 		if (idx > 0) {
2189 			fnvlist_add_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
2190 			    (const nvlist_t * const *)child, idx);
2191 			fnvlist_add_nvlist(spa->spa_load_info,
2192 			    ZPOOL_CONFIG_MISSING_DEVICES, nv);
2193 
2194 			for (uint64_t i = 0; i < idx; i++)
2195 				nvlist_free(child[i]);
2196 		}
2197 		nvlist_free(nv);
2198 		kmem_free(child, rvd->vdev_children * sizeof (char **));
2199 
2200 		if (idx > 0) {
2201 			spa_load_failed(spa, "some log devices are missing");
2202 			vdev_dbgmsg_print_tree(rvd, 2);
2203 			return (SET_ERROR(ENXIO));
2204 		}
2205 	} else {
2206 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
2207 			vdev_t *tvd = rvd->vdev_child[c];
2208 
2209 			if (tvd->vdev_islog &&
2210 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
2211 				spa_set_log_state(spa, SPA_LOG_CLEAR);
2212 				spa_load_note(spa, "some log devices are "
2213 				    "missing, ZIL is dropped.");
2214 				vdev_dbgmsg_print_tree(rvd, 2);
2215 				break;
2216 			}
2217 		}
2218 	}
2219 
2220 	return (0);
2221 }
2222 
2223 /*
2224  * Check for missing log devices
2225  */
2226 static boolean_t
2227 spa_check_logs(spa_t *spa)
2228 {
2229 	boolean_t rv = B_FALSE;
2230 	dsl_pool_t *dp = spa_get_dsl(spa);
2231 
2232 	switch (spa->spa_log_state) {
2233 	default:
2234 		break;
2235 	case SPA_LOG_MISSING:
2236 		/* need to recheck in case slog has been restored */
2237 	case SPA_LOG_UNKNOWN:
2238 		rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2239 		    zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0);
2240 		if (rv)
2241 			spa_set_log_state(spa, SPA_LOG_MISSING);
2242 		break;
2243 	}
2244 	return (rv);
2245 }
2246 
2247 /*
2248  * Passivate any log vdevs (note, does not apply to embedded log metaslabs).
2249  */
2250 static boolean_t
2251 spa_passivate_log(spa_t *spa)
2252 {
2253 	vdev_t *rvd = spa->spa_root_vdev;
2254 	boolean_t slog_found = B_FALSE;
2255 
2256 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
2257 
2258 	for (int c = 0; c < rvd->vdev_children; c++) {
2259 		vdev_t *tvd = rvd->vdev_child[c];
2260 
2261 		if (tvd->vdev_islog) {
2262 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
2263 			metaslab_group_passivate(tvd->vdev_mg);
2264 			slog_found = B_TRUE;
2265 		}
2266 	}
2267 
2268 	return (slog_found);
2269 }
2270 
2271 /*
2272  * Activate any log vdevs (note, does not apply to embedded log metaslabs).
2273  */
2274 static void
2275 spa_activate_log(spa_t *spa)
2276 {
2277 	vdev_t *rvd = spa->spa_root_vdev;
2278 
2279 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
2280 
2281 	for (int c = 0; c < rvd->vdev_children; c++) {
2282 		vdev_t *tvd = rvd->vdev_child[c];
2283 
2284 		if (tvd->vdev_islog) {
2285 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
2286 			metaslab_group_activate(tvd->vdev_mg);
2287 		}
2288 	}
2289 }
2290 
2291 int
2292 spa_reset_logs(spa_t *spa)
2293 {
2294 	int error;
2295 
2296 	error = dmu_objset_find(spa_name(spa), zil_reset,
2297 	    NULL, DS_FIND_CHILDREN);
2298 	if (error == 0) {
2299 		/*
2300 		 * We successfully offlined the log device, sync out the
2301 		 * current txg so that the "stubby" block can be removed
2302 		 * by zil_sync().
2303 		 */
2304 		txg_wait_synced(spa->spa_dsl_pool, 0);
2305 	}
2306 	return (error);
2307 }
2308 
2309 static void
2310 spa_aux_check_removed(spa_aux_vdev_t *sav)
2311 {
2312 	for (int i = 0; i < sav->sav_count; i++)
2313 		spa_check_removed(sav->sav_vdevs[i]);
2314 }
2315 
2316 void
2317 spa_claim_notify(zio_t *zio)
2318 {
2319 	spa_t *spa = zio->io_spa;
2320 
2321 	if (zio->io_error)
2322 		return;
2323 
2324 	mutex_enter(&spa->spa_props_lock);	/* any mutex will do */
2325 	if (spa->spa_claim_max_txg < zio->io_bp->blk_birth)
2326 		spa->spa_claim_max_txg = zio->io_bp->blk_birth;
2327 	mutex_exit(&spa->spa_props_lock);
2328 }
2329 
2330 typedef struct spa_load_error {
2331 	boolean_t	sle_verify_data;
2332 	uint64_t	sle_meta_count;
2333 	uint64_t	sle_data_count;
2334 } spa_load_error_t;
2335 
2336 static void
2337 spa_load_verify_done(zio_t *zio)
2338 {
2339 	blkptr_t *bp = zio->io_bp;
2340 	spa_load_error_t *sle = zio->io_private;
2341 	dmu_object_type_t type = BP_GET_TYPE(bp);
2342 	int error = zio->io_error;
2343 	spa_t *spa = zio->io_spa;
2344 
2345 	abd_free(zio->io_abd);
2346 	if (error) {
2347 		if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) &&
2348 		    type != DMU_OT_INTENT_LOG)
2349 			atomic_inc_64(&sle->sle_meta_count);
2350 		else
2351 			atomic_inc_64(&sle->sle_data_count);
2352 	}
2353 
2354 	mutex_enter(&spa->spa_scrub_lock);
2355 	spa->spa_load_verify_bytes -= BP_GET_PSIZE(bp);
2356 	cv_broadcast(&spa->spa_scrub_io_cv);
2357 	mutex_exit(&spa->spa_scrub_lock);
2358 }
2359 
2360 /*
2361  * Maximum number of inflight bytes is the log2 fraction of the arc size.
2362  * By default, we set it to 1/16th of the arc.
2363  */
2364 static uint_t spa_load_verify_shift = 4;
2365 static int spa_load_verify_metadata = B_TRUE;
2366 static int spa_load_verify_data = B_TRUE;
2367 
2368 static int
2369 spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
2370     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
2371 {
2372 	zio_t *rio = arg;
2373 	spa_load_error_t *sle = rio->io_private;
2374 
2375 	(void) zilog, (void) dnp;
2376 
2377 	/*
2378 	 * Note: normally this routine will not be called if
2379 	 * spa_load_verify_metadata is not set.  However, it may be useful
2380 	 * to manually set the flag after the traversal has begun.
2381 	 */
2382 	if (!spa_load_verify_metadata)
2383 		return (0);
2384 
2385 	/*
2386 	 * Sanity check the block pointer in order to detect obvious damage
2387 	 * before using the contents in subsequent checks or in zio_read().
2388 	 * When damaged consider it to be a metadata error since we cannot
2389 	 * trust the BP_GET_TYPE and BP_GET_LEVEL values.
2390 	 */
2391 	if (!zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
2392 		atomic_inc_64(&sle->sle_meta_count);
2393 		return (0);
2394 	}
2395 
2396 	if (zb->zb_level == ZB_DNODE_LEVEL || BP_IS_HOLE(bp) ||
2397 	    BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp))
2398 		return (0);
2399 
2400 	if (!BP_IS_METADATA(bp) &&
2401 	    (!spa_load_verify_data || !sle->sle_verify_data))
2402 		return (0);
2403 
2404 	uint64_t maxinflight_bytes =
2405 	    arc_target_bytes() >> spa_load_verify_shift;
2406 	size_t size = BP_GET_PSIZE(bp);
2407 
2408 	mutex_enter(&spa->spa_scrub_lock);
2409 	while (spa->spa_load_verify_bytes >= maxinflight_bytes)
2410 		cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2411 	spa->spa_load_verify_bytes += size;
2412 	mutex_exit(&spa->spa_scrub_lock);
2413 
2414 	zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size,
2415 	    spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB,
2416 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL |
2417 	    ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb));
2418 	return (0);
2419 }
2420 
2421 static int
2422 verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
2423 {
2424 	(void) dp, (void) arg;
2425 
2426 	if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN)
2427 		return (SET_ERROR(ENAMETOOLONG));
2428 
2429 	return (0);
2430 }
2431 
2432 static int
2433 spa_load_verify(spa_t *spa)
2434 {
2435 	zio_t *rio;
2436 	spa_load_error_t sle = { 0 };
2437 	zpool_load_policy_t policy;
2438 	boolean_t verify_ok = B_FALSE;
2439 	int error = 0;
2440 
2441 	zpool_get_load_policy(spa->spa_config, &policy);
2442 
2443 	if (policy.zlp_rewind & ZPOOL_NEVER_REWIND ||
2444 	    policy.zlp_maxmeta == UINT64_MAX)
2445 		return (0);
2446 
2447 	dsl_pool_config_enter(spa->spa_dsl_pool, FTAG);
2448 	error = dmu_objset_find_dp(spa->spa_dsl_pool,
2449 	    spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL,
2450 	    DS_FIND_CHILDREN);
2451 	dsl_pool_config_exit(spa->spa_dsl_pool, FTAG);
2452 	if (error != 0)
2453 		return (error);
2454 
2455 	/*
2456 	 * Verify data only if we are rewinding or error limit was set.
2457 	 * Otherwise nothing except dbgmsg care about it to waste time.
2458 	 */
2459 	sle.sle_verify_data = (policy.zlp_rewind & ZPOOL_REWIND_MASK) ||
2460 	    (policy.zlp_maxdata < UINT64_MAX);
2461 
2462 	rio = zio_root(spa, NULL, &sle,
2463 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
2464 
2465 	if (spa_load_verify_metadata) {
2466 		if (spa->spa_extreme_rewind) {
2467 			spa_load_note(spa, "performing a complete scan of the "
2468 			    "pool since extreme rewind is on. This may take "
2469 			    "a very long time.\n  (spa_load_verify_data=%u, "
2470 			    "spa_load_verify_metadata=%u)",
2471 			    spa_load_verify_data, spa_load_verify_metadata);
2472 		}
2473 
2474 		error = traverse_pool(spa, spa->spa_verify_min_txg,
2475 		    TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA |
2476 		    TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio);
2477 	}
2478 
2479 	(void) zio_wait(rio);
2480 	ASSERT0(spa->spa_load_verify_bytes);
2481 
2482 	spa->spa_load_meta_errors = sle.sle_meta_count;
2483 	spa->spa_load_data_errors = sle.sle_data_count;
2484 
2485 	if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) {
2486 		spa_load_note(spa, "spa_load_verify found %llu metadata errors "
2487 		    "and %llu data errors", (u_longlong_t)sle.sle_meta_count,
2488 		    (u_longlong_t)sle.sle_data_count);
2489 	}
2490 
2491 	if (spa_load_verify_dryrun ||
2492 	    (!error && sle.sle_meta_count <= policy.zlp_maxmeta &&
2493 	    sle.sle_data_count <= policy.zlp_maxdata)) {
2494 		int64_t loss = 0;
2495 
2496 		verify_ok = B_TRUE;
2497 		spa->spa_load_txg = spa->spa_uberblock.ub_txg;
2498 		spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp;
2499 
2500 		loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts;
2501 		fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_LOAD_TIME,
2502 		    spa->spa_load_txg_ts);
2503 		fnvlist_add_int64(spa->spa_load_info, ZPOOL_CONFIG_REWIND_TIME,
2504 		    loss);
2505 		fnvlist_add_uint64(spa->spa_load_info,
2506 		    ZPOOL_CONFIG_LOAD_META_ERRORS, sle.sle_meta_count);
2507 		fnvlist_add_uint64(spa->spa_load_info,
2508 		    ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count);
2509 	} else {
2510 		spa->spa_load_max_txg = spa->spa_uberblock.ub_txg;
2511 	}
2512 
2513 	if (spa_load_verify_dryrun)
2514 		return (0);
2515 
2516 	if (error) {
2517 		if (error != ENXIO && error != EIO)
2518 			error = SET_ERROR(EIO);
2519 		return (error);
2520 	}
2521 
2522 	return (verify_ok ? 0 : EIO);
2523 }
2524 
2525 /*
2526  * Find a value in the pool props object.
2527  */
2528 static void
2529 spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val)
2530 {
2531 	(void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object,
2532 	    zpool_prop_to_name(prop), sizeof (uint64_t), 1, val);
2533 }
2534 
2535 /*
2536  * Find a value in the pool directory object.
2537  */
2538 static int
2539 spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent)
2540 {
2541 	int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
2542 	    name, sizeof (uint64_t), 1, val);
2543 
2544 	if (error != 0 && (error != ENOENT || log_enoent)) {
2545 		spa_load_failed(spa, "couldn't get '%s' value in MOS directory "
2546 		    "[error=%d]", name, error);
2547 	}
2548 
2549 	return (error);
2550 }
2551 
2552 static int
2553 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err)
2554 {
2555 	vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux);
2556 	return (SET_ERROR(err));
2557 }
2558 
2559 boolean_t
2560 spa_livelist_delete_check(spa_t *spa)
2561 {
2562 	return (spa->spa_livelists_to_delete != 0);
2563 }
2564 
2565 static boolean_t
2566 spa_livelist_delete_cb_check(void *arg, zthr_t *z)
2567 {
2568 	(void) z;
2569 	spa_t *spa = arg;
2570 	return (spa_livelist_delete_check(spa));
2571 }
2572 
2573 static int
2574 delete_blkptr_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
2575 {
2576 	spa_t *spa = arg;
2577 	zio_free(spa, tx->tx_txg, bp);
2578 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
2579 	    -bp_get_dsize_sync(spa, bp),
2580 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
2581 	return (0);
2582 }
2583 
2584 static int
2585 dsl_get_next_livelist_obj(objset_t *os, uint64_t zap_obj, uint64_t *llp)
2586 {
2587 	int err;
2588 	zap_cursor_t zc;
2589 	zap_attribute_t za;
2590 	zap_cursor_init(&zc, os, zap_obj);
2591 	err = zap_cursor_retrieve(&zc, &za);
2592 	zap_cursor_fini(&zc);
2593 	if (err == 0)
2594 		*llp = za.za_first_integer;
2595 	return (err);
2596 }
2597 
2598 /*
2599  * Components of livelist deletion that must be performed in syncing
2600  * context: freeing block pointers and updating the pool-wide data
2601  * structures to indicate how much work is left to do
2602  */
2603 typedef struct sublist_delete_arg {
2604 	spa_t *spa;
2605 	dsl_deadlist_t *ll;
2606 	uint64_t key;
2607 	bplist_t *to_free;
2608 } sublist_delete_arg_t;
2609 
2610 static void
2611 sublist_delete_sync(void *arg, dmu_tx_t *tx)
2612 {
2613 	sublist_delete_arg_t *sda = arg;
2614 	spa_t *spa = sda->spa;
2615 	dsl_deadlist_t *ll = sda->ll;
2616 	uint64_t key = sda->key;
2617 	bplist_t *to_free = sda->to_free;
2618 
2619 	bplist_iterate(to_free, delete_blkptr_cb, spa, tx);
2620 	dsl_deadlist_remove_entry(ll, key, tx);
2621 }
2622 
2623 typedef struct livelist_delete_arg {
2624 	spa_t *spa;
2625 	uint64_t ll_obj;
2626 	uint64_t zap_obj;
2627 } livelist_delete_arg_t;
2628 
2629 static void
2630 livelist_delete_sync(void *arg, dmu_tx_t *tx)
2631 {
2632 	livelist_delete_arg_t *lda = arg;
2633 	spa_t *spa = lda->spa;
2634 	uint64_t ll_obj = lda->ll_obj;
2635 	uint64_t zap_obj = lda->zap_obj;
2636 	objset_t *mos = spa->spa_meta_objset;
2637 	uint64_t count;
2638 
2639 	/* free the livelist and decrement the feature count */
2640 	VERIFY0(zap_remove_int(mos, zap_obj, ll_obj, tx));
2641 	dsl_deadlist_free(mos, ll_obj, tx);
2642 	spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx);
2643 	VERIFY0(zap_count(mos, zap_obj, &count));
2644 	if (count == 0) {
2645 		/* no more livelists to delete */
2646 		VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT,
2647 		    DMU_POOL_DELETED_CLONES, tx));
2648 		VERIFY0(zap_destroy(mos, zap_obj, tx));
2649 		spa->spa_livelists_to_delete = 0;
2650 		spa_notify_waiters(spa);
2651 	}
2652 }
2653 
2654 /*
2655  * Load in the value for the livelist to be removed and open it. Then,
2656  * load its first sublist and determine which block pointers should actually
2657  * be freed. Then, call a synctask which performs the actual frees and updates
2658  * the pool-wide livelist data.
2659  */
2660 static void
2661 spa_livelist_delete_cb(void *arg, zthr_t *z)
2662 {
2663 	spa_t *spa = arg;
2664 	uint64_t ll_obj = 0, count;
2665 	objset_t *mos = spa->spa_meta_objset;
2666 	uint64_t zap_obj = spa->spa_livelists_to_delete;
2667 	/*
2668 	 * Determine the next livelist to delete. This function should only
2669 	 * be called if there is at least one deleted clone.
2670 	 */
2671 	VERIFY0(dsl_get_next_livelist_obj(mos, zap_obj, &ll_obj));
2672 	VERIFY0(zap_count(mos, ll_obj, &count));
2673 	if (count > 0) {
2674 		dsl_deadlist_t *ll;
2675 		dsl_deadlist_entry_t *dle;
2676 		bplist_t to_free;
2677 		ll = kmem_zalloc(sizeof (dsl_deadlist_t), KM_SLEEP);
2678 		dsl_deadlist_open(ll, mos, ll_obj);
2679 		dle = dsl_deadlist_first(ll);
2680 		ASSERT3P(dle, !=, NULL);
2681 		bplist_create(&to_free);
2682 		int err = dsl_process_sub_livelist(&dle->dle_bpobj, &to_free,
2683 		    z, NULL);
2684 		if (err == 0) {
2685 			sublist_delete_arg_t sync_arg = {
2686 			    .spa = spa,
2687 			    .ll = ll,
2688 			    .key = dle->dle_mintxg,
2689 			    .to_free = &to_free
2690 			};
2691 			zfs_dbgmsg("deleting sublist (id %llu) from"
2692 			    " livelist %llu, %lld remaining",
2693 			    (u_longlong_t)dle->dle_bpobj.bpo_object,
2694 			    (u_longlong_t)ll_obj, (longlong_t)count - 1);
2695 			VERIFY0(dsl_sync_task(spa_name(spa), NULL,
2696 			    sublist_delete_sync, &sync_arg, 0,
2697 			    ZFS_SPACE_CHECK_DESTROY));
2698 		} else {
2699 			VERIFY3U(err, ==, EINTR);
2700 		}
2701 		bplist_clear(&to_free);
2702 		bplist_destroy(&to_free);
2703 		dsl_deadlist_close(ll);
2704 		kmem_free(ll, sizeof (dsl_deadlist_t));
2705 	} else {
2706 		livelist_delete_arg_t sync_arg = {
2707 		    .spa = spa,
2708 		    .ll_obj = ll_obj,
2709 		    .zap_obj = zap_obj
2710 		};
2711 		zfs_dbgmsg("deletion of livelist %llu completed",
2712 		    (u_longlong_t)ll_obj);
2713 		VERIFY0(dsl_sync_task(spa_name(spa), NULL, livelist_delete_sync,
2714 		    &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY));
2715 	}
2716 }
2717 
2718 static void
2719 spa_start_livelist_destroy_thread(spa_t *spa)
2720 {
2721 	ASSERT3P(spa->spa_livelist_delete_zthr, ==, NULL);
2722 	spa->spa_livelist_delete_zthr =
2723 	    zthr_create("z_livelist_destroy",
2724 	    spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa,
2725 	    minclsyspri);
2726 }
2727 
2728 typedef struct livelist_new_arg {
2729 	bplist_t *allocs;
2730 	bplist_t *frees;
2731 } livelist_new_arg_t;
2732 
2733 static int
2734 livelist_track_new_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
2735     dmu_tx_t *tx)
2736 {
2737 	ASSERT(tx == NULL);
2738 	livelist_new_arg_t *lna = arg;
2739 	if (bp_freed) {
2740 		bplist_append(lna->frees, bp);
2741 	} else {
2742 		bplist_append(lna->allocs, bp);
2743 		zfs_livelist_condense_new_alloc++;
2744 	}
2745 	return (0);
2746 }
2747 
2748 typedef struct livelist_condense_arg {
2749 	spa_t *spa;
2750 	bplist_t to_keep;
2751 	uint64_t first_size;
2752 	uint64_t next_size;
2753 } livelist_condense_arg_t;
2754 
2755 static void
2756 spa_livelist_condense_sync(void *arg, dmu_tx_t *tx)
2757 {
2758 	livelist_condense_arg_t *lca = arg;
2759 	spa_t *spa = lca->spa;
2760 	bplist_t new_frees;
2761 	dsl_dataset_t *ds = spa->spa_to_condense.ds;
2762 
2763 	/* Have we been cancelled? */
2764 	if (spa->spa_to_condense.cancelled) {
2765 		zfs_livelist_condense_sync_cancel++;
2766 		goto out;
2767 	}
2768 
2769 	dsl_deadlist_entry_t *first = spa->spa_to_condense.first;
2770 	dsl_deadlist_entry_t *next = spa->spa_to_condense.next;
2771 	dsl_deadlist_t *ll = &ds->ds_dir->dd_livelist;
2772 
2773 	/*
2774 	 * It's possible that the livelist was changed while the zthr was
2775 	 * running. Therefore, we need to check for new blkptrs in the two
2776 	 * entries being condensed and continue to track them in the livelist.
2777 	 * Because of the way we handle remapped blkptrs (see dbuf_remap_impl),
2778 	 * it's possible that the newly added blkptrs are FREEs or ALLOCs so
2779 	 * we need to sort them into two different bplists.
2780 	 */
2781 	uint64_t first_obj = first->dle_bpobj.bpo_object;
2782 	uint64_t next_obj = next->dle_bpobj.bpo_object;
2783 	uint64_t cur_first_size = first->dle_bpobj.bpo_phys->bpo_num_blkptrs;
2784 	uint64_t cur_next_size = next->dle_bpobj.bpo_phys->bpo_num_blkptrs;
2785 
2786 	bplist_create(&new_frees);
2787 	livelist_new_arg_t new_bps = {
2788 	    .allocs = &lca->to_keep,
2789 	    .frees = &new_frees,
2790 	};
2791 
2792 	if (cur_first_size > lca->first_size) {
2793 		VERIFY0(livelist_bpobj_iterate_from_nofree(&first->dle_bpobj,
2794 		    livelist_track_new_cb, &new_bps, lca->first_size));
2795 	}
2796 	if (cur_next_size > lca->next_size) {
2797 		VERIFY0(livelist_bpobj_iterate_from_nofree(&next->dle_bpobj,
2798 		    livelist_track_new_cb, &new_bps, lca->next_size));
2799 	}
2800 
2801 	dsl_deadlist_clear_entry(first, ll, tx);
2802 	ASSERT(bpobj_is_empty(&first->dle_bpobj));
2803 	dsl_deadlist_remove_entry(ll, next->dle_mintxg, tx);
2804 
2805 	bplist_iterate(&lca->to_keep, dsl_deadlist_insert_alloc_cb, ll, tx);
2806 	bplist_iterate(&new_frees, dsl_deadlist_insert_free_cb, ll, tx);
2807 	bplist_destroy(&new_frees);
2808 
2809 	char dsname[ZFS_MAX_DATASET_NAME_LEN];
2810 	dsl_dataset_name(ds, dsname);
2811 	zfs_dbgmsg("txg %llu condensing livelist of %s (id %llu), bpobj %llu "
2812 	    "(%llu blkptrs) and bpobj %llu (%llu blkptrs) -> bpobj %llu "
2813 	    "(%llu blkptrs)", (u_longlong_t)tx->tx_txg, dsname,
2814 	    (u_longlong_t)ds->ds_object, (u_longlong_t)first_obj,
2815 	    (u_longlong_t)cur_first_size, (u_longlong_t)next_obj,
2816 	    (u_longlong_t)cur_next_size,
2817 	    (u_longlong_t)first->dle_bpobj.bpo_object,
2818 	    (u_longlong_t)first->dle_bpobj.bpo_phys->bpo_num_blkptrs);
2819 out:
2820 	dmu_buf_rele(ds->ds_dbuf, spa);
2821 	spa->spa_to_condense.ds = NULL;
2822 	bplist_clear(&lca->to_keep);
2823 	bplist_destroy(&lca->to_keep);
2824 	kmem_free(lca, sizeof (livelist_condense_arg_t));
2825 	spa->spa_to_condense.syncing = B_FALSE;
2826 }
2827 
2828 static void
2829 spa_livelist_condense_cb(void *arg, zthr_t *t)
2830 {
2831 	while (zfs_livelist_condense_zthr_pause &&
2832 	    !(zthr_has_waiters(t) || zthr_iscancelled(t)))
2833 		delay(1);
2834 
2835 	spa_t *spa = arg;
2836 	dsl_deadlist_entry_t *first = spa->spa_to_condense.first;
2837 	dsl_deadlist_entry_t *next = spa->spa_to_condense.next;
2838 	uint64_t first_size, next_size;
2839 
2840 	livelist_condense_arg_t *lca =
2841 	    kmem_alloc(sizeof (livelist_condense_arg_t), KM_SLEEP);
2842 	bplist_create(&lca->to_keep);
2843 
2844 	/*
2845 	 * Process the livelists (matching FREEs and ALLOCs) in open context
2846 	 * so we have minimal work in syncing context to condense.
2847 	 *
2848 	 * We save bpobj sizes (first_size and next_size) to use later in
2849 	 * syncing context to determine if entries were added to these sublists
2850 	 * while in open context. This is possible because the clone is still
2851 	 * active and open for normal writes and we want to make sure the new,
2852 	 * unprocessed blockpointers are inserted into the livelist normally.
2853 	 *
2854 	 * Note that dsl_process_sub_livelist() both stores the size number of
2855 	 * blockpointers and iterates over them while the bpobj's lock held, so
2856 	 * the sizes returned to us are consistent which what was actually
2857 	 * processed.
2858 	 */
2859 	int err = dsl_process_sub_livelist(&first->dle_bpobj, &lca->to_keep, t,
2860 	    &first_size);
2861 	if (err == 0)
2862 		err = dsl_process_sub_livelist(&next->dle_bpobj, &lca->to_keep,
2863 		    t, &next_size);
2864 
2865 	if (err == 0) {
2866 		while (zfs_livelist_condense_sync_pause &&
2867 		    !(zthr_has_waiters(t) || zthr_iscancelled(t)))
2868 			delay(1);
2869 
2870 		dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
2871 		dmu_tx_mark_netfree(tx);
2872 		dmu_tx_hold_space(tx, 1);
2873 		err = dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE);
2874 		if (err == 0) {
2875 			/*
2876 			 * Prevent the condense zthr restarting before
2877 			 * the synctask completes.
2878 			 */
2879 			spa->spa_to_condense.syncing = B_TRUE;
2880 			lca->spa = spa;
2881 			lca->first_size = first_size;
2882 			lca->next_size = next_size;
2883 			dsl_sync_task_nowait(spa_get_dsl(spa),
2884 			    spa_livelist_condense_sync, lca, tx);
2885 			dmu_tx_commit(tx);
2886 			return;
2887 		}
2888 	}
2889 	/*
2890 	 * Condensing can not continue: either it was externally stopped or
2891 	 * we were unable to assign to a tx because the pool has run out of
2892 	 * space. In the second case, we'll just end up trying to condense
2893 	 * again in a later txg.
2894 	 */
2895 	ASSERT(err != 0);
2896 	bplist_clear(&lca->to_keep);
2897 	bplist_destroy(&lca->to_keep);
2898 	kmem_free(lca, sizeof (livelist_condense_arg_t));
2899 	dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf, spa);
2900 	spa->spa_to_condense.ds = NULL;
2901 	if (err == EINTR)
2902 		zfs_livelist_condense_zthr_cancel++;
2903 }
2904 
2905 /*
2906  * Check that there is something to condense but that a condense is not
2907  * already in progress and that condensing has not been cancelled.
2908  */
2909 static boolean_t
2910 spa_livelist_condense_cb_check(void *arg, zthr_t *z)
2911 {
2912 	(void) z;
2913 	spa_t *spa = arg;
2914 	if ((spa->spa_to_condense.ds != NULL) &&
2915 	    (spa->spa_to_condense.syncing == B_FALSE) &&
2916 	    (spa->spa_to_condense.cancelled == B_FALSE)) {
2917 		return (B_TRUE);
2918 	}
2919 	return (B_FALSE);
2920 }
2921 
2922 static void
2923 spa_start_livelist_condensing_thread(spa_t *spa)
2924 {
2925 	spa->spa_to_condense.ds = NULL;
2926 	spa->spa_to_condense.first = NULL;
2927 	spa->spa_to_condense.next = NULL;
2928 	spa->spa_to_condense.syncing = B_FALSE;
2929 	spa->spa_to_condense.cancelled = B_FALSE;
2930 
2931 	ASSERT3P(spa->spa_livelist_condense_zthr, ==, NULL);
2932 	spa->spa_livelist_condense_zthr =
2933 	    zthr_create("z_livelist_condense",
2934 	    spa_livelist_condense_cb_check,
2935 	    spa_livelist_condense_cb, spa, minclsyspri);
2936 }
2937 
2938 static void
2939 spa_spawn_aux_threads(spa_t *spa)
2940 {
2941 	ASSERT(spa_writeable(spa));
2942 
2943 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
2944 
2945 	spa_start_indirect_condensing_thread(spa);
2946 	spa_start_livelist_destroy_thread(spa);
2947 	spa_start_livelist_condensing_thread(spa);
2948 
2949 	ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL);
2950 	spa->spa_checkpoint_discard_zthr =
2951 	    zthr_create("z_checkpoint_discard",
2952 	    spa_checkpoint_discard_thread_check,
2953 	    spa_checkpoint_discard_thread, spa, minclsyspri);
2954 }
2955 
2956 /*
2957  * Fix up config after a partly-completed split.  This is done with the
2958  * ZPOOL_CONFIG_SPLIT nvlist.  Both the splitting pool and the split-off
2959  * pool have that entry in their config, but only the splitting one contains
2960  * a list of all the guids of the vdevs that are being split off.
2961  *
2962  * This function determines what to do with that list: either rejoin
2963  * all the disks to the pool, or complete the splitting process.  To attempt
2964  * the rejoin, each disk that is offlined is marked online again, and
2965  * we do a reopen() call.  If the vdev label for every disk that was
2966  * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL)
2967  * then we call vdev_split() on each disk, and complete the split.
2968  *
2969  * Otherwise we leave the config alone, with all the vdevs in place in
2970  * the original pool.
2971  */
2972 static void
2973 spa_try_repair(spa_t *spa, nvlist_t *config)
2974 {
2975 	uint_t extracted;
2976 	uint64_t *glist;
2977 	uint_t i, gcount;
2978 	nvlist_t *nvl;
2979 	vdev_t **vd;
2980 	boolean_t attempt_reopen;
2981 
2982 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0)
2983 		return;
2984 
2985 	/* check that the config is complete */
2986 	if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST,
2987 	    &glist, &gcount) != 0)
2988 		return;
2989 
2990 	vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP);
2991 
2992 	/* attempt to online all the vdevs & validate */
2993 	attempt_reopen = B_TRUE;
2994 	for (i = 0; i < gcount; i++) {
2995 		if (glist[i] == 0)	/* vdev is hole */
2996 			continue;
2997 
2998 		vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE);
2999 		if (vd[i] == NULL) {
3000 			/*
3001 			 * Don't bother attempting to reopen the disks;
3002 			 * just do the split.
3003 			 */
3004 			attempt_reopen = B_FALSE;
3005 		} else {
3006 			/* attempt to re-online it */
3007 			vd[i]->vdev_offline = B_FALSE;
3008 		}
3009 	}
3010 
3011 	if (attempt_reopen) {
3012 		vdev_reopen(spa->spa_root_vdev);
3013 
3014 		/* check each device to see what state it's in */
3015 		for (extracted = 0, i = 0; i < gcount; i++) {
3016 			if (vd[i] != NULL &&
3017 			    vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL)
3018 				break;
3019 			++extracted;
3020 		}
3021 	}
3022 
3023 	/*
3024 	 * If every disk has been moved to the new pool, or if we never
3025 	 * even attempted to look at them, then we split them off for
3026 	 * good.
3027 	 */
3028 	if (!attempt_reopen || gcount == extracted) {
3029 		for (i = 0; i < gcount; i++)
3030 			if (vd[i] != NULL)
3031 				vdev_split(vd[i]);
3032 		vdev_reopen(spa->spa_root_vdev);
3033 	}
3034 
3035 	kmem_free(vd, gcount * sizeof (vdev_t *));
3036 }
3037 
3038 static int
3039 spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type)
3040 {
3041 	const char *ereport = FM_EREPORT_ZFS_POOL;
3042 	int error;
3043 
3044 	spa->spa_load_state = state;
3045 	(void) spa_import_progress_set_state(spa_guid(spa),
3046 	    spa_load_state(spa));
3047 
3048 	gethrestime(&spa->spa_loaded_ts);
3049 	error = spa_load_impl(spa, type, &ereport);
3050 
3051 	/*
3052 	 * Don't count references from objsets that are already closed
3053 	 * and are making their way through the eviction process.
3054 	 */
3055 	spa_evicting_os_wait(spa);
3056 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
3057 	if (error) {
3058 		if (error != EEXIST) {
3059 			spa->spa_loaded_ts.tv_sec = 0;
3060 			spa->spa_loaded_ts.tv_nsec = 0;
3061 		}
3062 		if (error != EBADF) {
3063 			(void) zfs_ereport_post(ereport, spa,
3064 			    NULL, NULL, NULL, 0);
3065 		}
3066 	}
3067 	spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE;
3068 	spa->spa_ena = 0;
3069 
3070 	(void) spa_import_progress_set_state(spa_guid(spa),
3071 	    spa_load_state(spa));
3072 
3073 	return (error);
3074 }
3075 
3076 #ifdef ZFS_DEBUG
3077 /*
3078  * Count the number of per-vdev ZAPs associated with all of the vdevs in the
3079  * vdev tree rooted in the given vd, and ensure that each ZAP is present in the
3080  * spa's per-vdev ZAP list.
3081  */
3082 static uint64_t
3083 vdev_count_verify_zaps(vdev_t *vd)
3084 {
3085 	spa_t *spa = vd->vdev_spa;
3086 	uint64_t total = 0;
3087 
3088 	if (spa_feature_is_active(vd->vdev_spa, SPA_FEATURE_AVZ_V2) &&
3089 	    vd->vdev_root_zap != 0) {
3090 		total++;
3091 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
3092 		    spa->spa_all_vdev_zaps, vd->vdev_root_zap));
3093 	}
3094 	if (vd->vdev_top_zap != 0) {
3095 		total++;
3096 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
3097 		    spa->spa_all_vdev_zaps, vd->vdev_top_zap));
3098 	}
3099 	if (vd->vdev_leaf_zap != 0) {
3100 		total++;
3101 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
3102 		    spa->spa_all_vdev_zaps, vd->vdev_leaf_zap));
3103 	}
3104 
3105 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3106 		total += vdev_count_verify_zaps(vd->vdev_child[i]);
3107 	}
3108 
3109 	return (total);
3110 }
3111 #else
3112 #define	vdev_count_verify_zaps(vd) ((void) sizeof (vd), 0)
3113 #endif
3114 
3115 /*
3116  * Determine whether the activity check is required.
3117  */
3118 static boolean_t
3119 spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
3120     nvlist_t *config)
3121 {
3122 	uint64_t state = 0;
3123 	uint64_t hostid = 0;
3124 	uint64_t tryconfig_txg = 0;
3125 	uint64_t tryconfig_timestamp = 0;
3126 	uint16_t tryconfig_mmp_seq = 0;
3127 	nvlist_t *nvinfo;
3128 
3129 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
3130 		nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO);
3131 		(void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG,
3132 		    &tryconfig_txg);
3133 		(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
3134 		    &tryconfig_timestamp);
3135 		(void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ,
3136 		    &tryconfig_mmp_seq);
3137 	}
3138 
3139 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state);
3140 
3141 	/*
3142 	 * Disable the MMP activity check - This is used by zdb which
3143 	 * is intended to be used on potentially active pools.
3144 	 */
3145 	if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP)
3146 		return (B_FALSE);
3147 
3148 	/*
3149 	 * Skip the activity check when the MMP feature is disabled.
3150 	 */
3151 	if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0)
3152 		return (B_FALSE);
3153 
3154 	/*
3155 	 * If the tryconfig_ values are nonzero, they are the results of an
3156 	 * earlier tryimport.  If they all match the uberblock we just found,
3157 	 * then the pool has not changed and we return false so we do not test
3158 	 * a second time.
3159 	 */
3160 	if (tryconfig_txg && tryconfig_txg == ub->ub_txg &&
3161 	    tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp &&
3162 	    tryconfig_mmp_seq && tryconfig_mmp_seq ==
3163 	    (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0))
3164 		return (B_FALSE);
3165 
3166 	/*
3167 	 * Allow the activity check to be skipped when importing the pool
3168 	 * on the same host which last imported it.  Since the hostid from
3169 	 * configuration may be stale use the one read from the label.
3170 	 */
3171 	if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID))
3172 		hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID);
3173 
3174 	if (hostid == spa_get_hostid(spa))
3175 		return (B_FALSE);
3176 
3177 	/*
3178 	 * Skip the activity test when the pool was cleanly exported.
3179 	 */
3180 	if (state != POOL_STATE_ACTIVE)
3181 		return (B_FALSE);
3182 
3183 	return (B_TRUE);
3184 }
3185 
3186 /*
3187  * Nanoseconds the activity check must watch for changes on-disk.
3188  */
3189 static uint64_t
3190 spa_activity_check_duration(spa_t *spa, uberblock_t *ub)
3191 {
3192 	uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1);
3193 	uint64_t multihost_interval = MSEC2NSEC(
3194 	    MMP_INTERVAL_OK(zfs_multihost_interval));
3195 	uint64_t import_delay = MAX(NANOSEC, import_intervals *
3196 	    multihost_interval);
3197 
3198 	/*
3199 	 * Local tunables determine a minimum duration except for the case
3200 	 * where we know when the remote host will suspend the pool if MMP
3201 	 * writes do not land.
3202 	 *
3203 	 * See Big Theory comment at the top of mmp.c for the reasoning behind
3204 	 * these cases and times.
3205 	 */
3206 
3207 	ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100);
3208 
3209 	if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
3210 	    MMP_FAIL_INT(ub) > 0) {
3211 
3212 		/* MMP on remote host will suspend pool after failed writes */
3213 		import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) *
3214 		    MMP_IMPORT_SAFETY_FACTOR / 100;
3215 
3216 		zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp "
3217 		    "mmp_fails=%llu ub_mmp mmp_interval=%llu "
3218 		    "import_intervals=%llu", (u_longlong_t)import_delay,
3219 		    (u_longlong_t)MMP_FAIL_INT(ub),
3220 		    (u_longlong_t)MMP_INTERVAL(ub),
3221 		    (u_longlong_t)import_intervals);
3222 
3223 	} else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
3224 	    MMP_FAIL_INT(ub) == 0) {
3225 
3226 		/* MMP on remote host will never suspend pool */
3227 		import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) +
3228 		    ub->ub_mmp_delay) * import_intervals);
3229 
3230 		zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp "
3231 		    "mmp_interval=%llu ub_mmp_delay=%llu "
3232 		    "import_intervals=%llu", (u_longlong_t)import_delay,
3233 		    (u_longlong_t)MMP_INTERVAL(ub),
3234 		    (u_longlong_t)ub->ub_mmp_delay,
3235 		    (u_longlong_t)import_intervals);
3236 
3237 	} else if (MMP_VALID(ub)) {
3238 		/*
3239 		 * zfs-0.7 compatibility case
3240 		 */
3241 
3242 		import_delay = MAX(import_delay, (multihost_interval +
3243 		    ub->ub_mmp_delay) * import_intervals);
3244 
3245 		zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu "
3246 		    "import_intervals=%llu leaves=%u",
3247 		    (u_longlong_t)import_delay,
3248 		    (u_longlong_t)ub->ub_mmp_delay,
3249 		    (u_longlong_t)import_intervals,
3250 		    vdev_count_leaves(spa));
3251 	} else {
3252 		/* Using local tunings is the only reasonable option */
3253 		zfs_dbgmsg("pool last imported on non-MMP aware "
3254 		    "host using import_delay=%llu multihost_interval=%llu "
3255 		    "import_intervals=%llu", (u_longlong_t)import_delay,
3256 		    (u_longlong_t)multihost_interval,
3257 		    (u_longlong_t)import_intervals);
3258 	}
3259 
3260 	return (import_delay);
3261 }
3262 
3263 /*
3264  * Perform the import activity check.  If the user canceled the import or
3265  * we detected activity then fail.
3266  */
3267 static int
3268 spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
3269 {
3270 	uint64_t txg = ub->ub_txg;
3271 	uint64_t timestamp = ub->ub_timestamp;
3272 	uint64_t mmp_config = ub->ub_mmp_config;
3273 	uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0;
3274 	uint64_t import_delay;
3275 	hrtime_t import_expire;
3276 	nvlist_t *mmp_label = NULL;
3277 	vdev_t *rvd = spa->spa_root_vdev;
3278 	kcondvar_t cv;
3279 	kmutex_t mtx;
3280 	int error = 0;
3281 
3282 	cv_init(&cv, NULL, CV_DEFAULT, NULL);
3283 	mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL);
3284 	mutex_enter(&mtx);
3285 
3286 	/*
3287 	 * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed
3288 	 * during the earlier tryimport.  If the txg recorded there is 0 then
3289 	 * the pool is known to be active on another host.
3290 	 *
3291 	 * Otherwise, the pool might be in use on another host.  Check for
3292 	 * changes in the uberblocks on disk if necessary.
3293 	 */
3294 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
3295 		nvlist_t *nvinfo = fnvlist_lookup_nvlist(config,
3296 		    ZPOOL_CONFIG_LOAD_INFO);
3297 
3298 		if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) &&
3299 		    fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) {
3300 			vdev_uberblock_load(rvd, ub, &mmp_label);
3301 			error = SET_ERROR(EREMOTEIO);
3302 			goto out;
3303 		}
3304 	}
3305 
3306 	import_delay = spa_activity_check_duration(spa, ub);
3307 
3308 	/* Add a small random factor in case of simultaneous imports (0-25%) */
3309 	import_delay += import_delay * random_in_range(250) / 1000;
3310 
3311 	import_expire = gethrtime() + import_delay;
3312 
3313 	while (gethrtime() < import_expire) {
3314 		(void) spa_import_progress_set_mmp_check(spa_guid(spa),
3315 		    NSEC2SEC(import_expire - gethrtime()));
3316 
3317 		vdev_uberblock_load(rvd, ub, &mmp_label);
3318 
3319 		if (txg != ub->ub_txg || timestamp != ub->ub_timestamp ||
3320 		    mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) {
3321 			zfs_dbgmsg("multihost activity detected "
3322 			    "txg %llu ub_txg  %llu "
3323 			    "timestamp %llu ub_timestamp  %llu "
3324 			    "mmp_config %#llx ub_mmp_config %#llx",
3325 			    (u_longlong_t)txg, (u_longlong_t)ub->ub_txg,
3326 			    (u_longlong_t)timestamp,
3327 			    (u_longlong_t)ub->ub_timestamp,
3328 			    (u_longlong_t)mmp_config,
3329 			    (u_longlong_t)ub->ub_mmp_config);
3330 
3331 			error = SET_ERROR(EREMOTEIO);
3332 			break;
3333 		}
3334 
3335 		if (mmp_label) {
3336 			nvlist_free(mmp_label);
3337 			mmp_label = NULL;
3338 		}
3339 
3340 		error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz);
3341 		if (error != -1) {
3342 			error = SET_ERROR(EINTR);
3343 			break;
3344 		}
3345 		error = 0;
3346 	}
3347 
3348 out:
3349 	mutex_exit(&mtx);
3350 	mutex_destroy(&mtx);
3351 	cv_destroy(&cv);
3352 
3353 	/*
3354 	 * If the pool is determined to be active store the status in the
3355 	 * spa->spa_load_info nvlist.  If the remote hostname or hostid are
3356 	 * available from configuration read from disk store them as well.
3357 	 * This allows 'zpool import' to generate a more useful message.
3358 	 *
3359 	 * ZPOOL_CONFIG_MMP_STATE    - observed pool status (mandatory)
3360 	 * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool
3361 	 * ZPOOL_CONFIG_MMP_HOSTID   - hostid from the active pool
3362 	 */
3363 	if (error == EREMOTEIO) {
3364 		const char *hostname = "<unknown>";
3365 		uint64_t hostid = 0;
3366 
3367 		if (mmp_label) {
3368 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) {
3369 				hostname = fnvlist_lookup_string(mmp_label,
3370 				    ZPOOL_CONFIG_HOSTNAME);
3371 				fnvlist_add_string(spa->spa_load_info,
3372 				    ZPOOL_CONFIG_MMP_HOSTNAME, hostname);
3373 			}
3374 
3375 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) {
3376 				hostid = fnvlist_lookup_uint64(mmp_label,
3377 				    ZPOOL_CONFIG_HOSTID);
3378 				fnvlist_add_uint64(spa->spa_load_info,
3379 				    ZPOOL_CONFIG_MMP_HOSTID, hostid);
3380 			}
3381 		}
3382 
3383 		fnvlist_add_uint64(spa->spa_load_info,
3384 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE);
3385 		fnvlist_add_uint64(spa->spa_load_info,
3386 		    ZPOOL_CONFIG_MMP_TXG, 0);
3387 
3388 		error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO);
3389 	}
3390 
3391 	if (mmp_label)
3392 		nvlist_free(mmp_label);
3393 
3394 	return (error);
3395 }
3396 
3397 static int
3398 spa_verify_host(spa_t *spa, nvlist_t *mos_config)
3399 {
3400 	uint64_t hostid;
3401 	const char *hostname;
3402 	uint64_t myhostid = 0;
3403 
3404 	if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config,
3405 	    ZPOOL_CONFIG_HOSTID, &hostid) == 0) {
3406 		hostname = fnvlist_lookup_string(mos_config,
3407 		    ZPOOL_CONFIG_HOSTNAME);
3408 
3409 		myhostid = zone_get_hostid(NULL);
3410 
3411 		if (hostid != 0 && myhostid != 0 && hostid != myhostid) {
3412 			cmn_err(CE_WARN, "pool '%s' could not be "
3413 			    "loaded as it was last accessed by "
3414 			    "another system (host: %s hostid: 0x%llx). "
3415 			    "See: https://openzfs.github.io/openzfs-docs/msg/"
3416 			    "ZFS-8000-EY",
3417 			    spa_name(spa), hostname, (u_longlong_t)hostid);
3418 			spa_load_failed(spa, "hostid verification failed: pool "
3419 			    "last accessed by host: %s (hostid: 0x%llx)",
3420 			    hostname, (u_longlong_t)hostid);
3421 			return (SET_ERROR(EBADF));
3422 		}
3423 	}
3424 
3425 	return (0);
3426 }
3427 
3428 static int
3429 spa_ld_parse_config(spa_t *spa, spa_import_type_t type)
3430 {
3431 	int error = 0;
3432 	nvlist_t *nvtree, *nvl, *config = spa->spa_config;
3433 	int parse;
3434 	vdev_t *rvd;
3435 	uint64_t pool_guid;
3436 	const char *comment;
3437 	const char *compatibility;
3438 
3439 	/*
3440 	 * Versioning wasn't explicitly added to the label until later, so if
3441 	 * it's not present treat it as the initial version.
3442 	 */
3443 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION,
3444 	    &spa->spa_ubsync.ub_version) != 0)
3445 		spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL;
3446 
3447 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) {
3448 		spa_load_failed(spa, "invalid config provided: '%s' missing",
3449 		    ZPOOL_CONFIG_POOL_GUID);
3450 		return (SET_ERROR(EINVAL));
3451 	}
3452 
3453 	/*
3454 	 * If we are doing an import, ensure that the pool is not already
3455 	 * imported by checking if its pool guid already exists in the
3456 	 * spa namespace.
3457 	 *
3458 	 * The only case that we allow an already imported pool to be
3459 	 * imported again, is when the pool is checkpointed and we want to
3460 	 * look at its checkpointed state from userland tools like zdb.
3461 	 */
3462 #ifdef _KERNEL
3463 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
3464 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
3465 	    spa_guid_exists(pool_guid, 0)) {
3466 #else
3467 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
3468 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
3469 	    spa_guid_exists(pool_guid, 0) &&
3470 	    !spa_importing_readonly_checkpoint(spa)) {
3471 #endif
3472 		spa_load_failed(spa, "a pool with guid %llu is already open",
3473 		    (u_longlong_t)pool_guid);
3474 		return (SET_ERROR(EEXIST));
3475 	}
3476 
3477 	spa->spa_config_guid = pool_guid;
3478 
3479 	nvlist_free(spa->spa_load_info);
3480 	spa->spa_load_info = fnvlist_alloc();
3481 
3482 	ASSERT(spa->spa_comment == NULL);
3483 	if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0)
3484 		spa->spa_comment = spa_strdup(comment);
3485 
3486 	ASSERT(spa->spa_compatibility == NULL);
3487 	if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMPATIBILITY,
3488 	    &compatibility) == 0)
3489 		spa->spa_compatibility = spa_strdup(compatibility);
3490 
3491 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
3492 	    &spa->spa_config_txg);
3493 
3494 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0)
3495 		spa->spa_config_splitting = fnvlist_dup(nvl);
3496 
3497 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) {
3498 		spa_load_failed(spa, "invalid config provided: '%s' missing",
3499 		    ZPOOL_CONFIG_VDEV_TREE);
3500 		return (SET_ERROR(EINVAL));
3501 	}
3502 
3503 	/*
3504 	 * Create "The Godfather" zio to hold all async IOs
3505 	 */
3506 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
3507 	    KM_SLEEP);
3508 	for (int i = 0; i < max_ncpus; i++) {
3509 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
3510 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3511 		    ZIO_FLAG_GODFATHER);
3512 	}
3513 
3514 	/*
3515 	 * Parse the configuration into a vdev tree.  We explicitly set the
3516 	 * value that will be returned by spa_version() since parsing the
3517 	 * configuration requires knowing the version number.
3518 	 */
3519 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3520 	parse = (type == SPA_IMPORT_EXISTING ?
3521 	    VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT);
3522 	error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse);
3523 	spa_config_exit(spa, SCL_ALL, FTAG);
3524 
3525 	if (error != 0) {
3526 		spa_load_failed(spa, "unable to parse config [error=%d]",
3527 		    error);
3528 		return (error);
3529 	}
3530 
3531 	ASSERT(spa->spa_root_vdev == rvd);
3532 	ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
3533 	ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT);
3534 
3535 	if (type != SPA_IMPORT_ASSEMBLE) {
3536 		ASSERT(spa_guid(spa) == pool_guid);
3537 	}
3538 
3539 	return (0);
3540 }
3541 
3542 /*
3543  * Recursively open all vdevs in the vdev tree. This function is called twice:
3544  * first with the untrusted config, then with the trusted config.
3545  */
3546 static int
3547 spa_ld_open_vdevs(spa_t *spa)
3548 {
3549 	int error = 0;
3550 
3551 	/*
3552 	 * spa_missing_tvds_allowed defines how many top-level vdevs can be
3553 	 * missing/unopenable for the root vdev to be still considered openable.
3554 	 */
3555 	if (spa->spa_trust_config) {
3556 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds;
3557 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) {
3558 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile;
3559 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) {
3560 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan;
3561 	} else {
3562 		spa->spa_missing_tvds_allowed = 0;
3563 	}
3564 
3565 	spa->spa_missing_tvds_allowed =
3566 	    MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed);
3567 
3568 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3569 	error = vdev_open(spa->spa_root_vdev);
3570 	spa_config_exit(spa, SCL_ALL, FTAG);
3571 
3572 	if (spa->spa_missing_tvds != 0) {
3573 		spa_load_note(spa, "vdev tree has %lld missing top-level "
3574 		    "vdevs.", (u_longlong_t)spa->spa_missing_tvds);
3575 		if (spa->spa_trust_config && (spa->spa_mode & SPA_MODE_WRITE)) {
3576 			/*
3577 			 * Although theoretically we could allow users to open
3578 			 * incomplete pools in RW mode, we'd need to add a lot
3579 			 * of extra logic (e.g. adjust pool space to account
3580 			 * for missing vdevs).
3581 			 * This limitation also prevents users from accidentally
3582 			 * opening the pool in RW mode during data recovery and
3583 			 * damaging it further.
3584 			 */
3585 			spa_load_note(spa, "pools with missing top-level "
3586 			    "vdevs can only be opened in read-only mode.");
3587 			error = SET_ERROR(ENXIO);
3588 		} else {
3589 			spa_load_note(spa, "current settings allow for maximum "
3590 			    "%lld missing top-level vdevs at this stage.",
3591 			    (u_longlong_t)spa->spa_missing_tvds_allowed);
3592 		}
3593 	}
3594 	if (error != 0) {
3595 		spa_load_failed(spa, "unable to open vdev tree [error=%d]",
3596 		    error);
3597 	}
3598 	if (spa->spa_missing_tvds != 0 || error != 0)
3599 		vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2);
3600 
3601 	return (error);
3602 }
3603 
3604 /*
3605  * We need to validate the vdev labels against the configuration that
3606  * we have in hand. This function is called twice: first with an untrusted
3607  * config, then with a trusted config. The validation is more strict when the
3608  * config is trusted.
3609  */
3610 static int
3611 spa_ld_validate_vdevs(spa_t *spa)
3612 {
3613 	int error = 0;
3614 	vdev_t *rvd = spa->spa_root_vdev;
3615 
3616 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3617 	error = vdev_validate(rvd);
3618 	spa_config_exit(spa, SCL_ALL, FTAG);
3619 
3620 	if (error != 0) {
3621 		spa_load_failed(spa, "vdev_validate failed [error=%d]", error);
3622 		return (error);
3623 	}
3624 
3625 	if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) {
3626 		spa_load_failed(spa, "cannot open vdev tree after invalidating "
3627 		    "some vdevs");
3628 		vdev_dbgmsg_print_tree(rvd, 2);
3629 		return (SET_ERROR(ENXIO));
3630 	}
3631 
3632 	return (0);
3633 }
3634 
3635 static void
3636 spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub)
3637 {
3638 	spa->spa_state = POOL_STATE_ACTIVE;
3639 	spa->spa_ubsync = spa->spa_uberblock;
3640 	spa->spa_verify_min_txg = spa->spa_extreme_rewind ?
3641 	    TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1;
3642 	spa->spa_first_txg = spa->spa_last_ubsync_txg ?
3643 	    spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1;
3644 	spa->spa_claim_max_txg = spa->spa_first_txg;
3645 	spa->spa_prev_software_version = ub->ub_software_version;
3646 }
3647 
3648 static int
3649 spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type)
3650 {
3651 	vdev_t *rvd = spa->spa_root_vdev;
3652 	nvlist_t *label;
3653 	uberblock_t *ub = &spa->spa_uberblock;
3654 	boolean_t activity_check = B_FALSE;
3655 
3656 	/*
3657 	 * If we are opening the checkpointed state of the pool by
3658 	 * rewinding to it, at this point we will have written the
3659 	 * checkpointed uberblock to the vdev labels, so searching
3660 	 * the labels will find the right uberblock.  However, if
3661 	 * we are opening the checkpointed state read-only, we have
3662 	 * not modified the labels. Therefore, we must ignore the
3663 	 * labels and continue using the spa_uberblock that was set
3664 	 * by spa_ld_checkpoint_rewind.
3665 	 *
3666 	 * Note that it would be fine to ignore the labels when
3667 	 * rewinding (opening writeable) as well. However, if we
3668 	 * crash just after writing the labels, we will end up
3669 	 * searching the labels. Doing so in the common case means
3670 	 * that this code path gets exercised normally, rather than
3671 	 * just in the edge case.
3672 	 */
3673 	if (ub->ub_checkpoint_txg != 0 &&
3674 	    spa_importing_readonly_checkpoint(spa)) {
3675 		spa_ld_select_uberblock_done(spa, ub);
3676 		return (0);
3677 	}
3678 
3679 	/*
3680 	 * Find the best uberblock.
3681 	 */
3682 	vdev_uberblock_load(rvd, ub, &label);
3683 
3684 	/*
3685 	 * If we weren't able to find a single valid uberblock, return failure.
3686 	 */
3687 	if (ub->ub_txg == 0) {
3688 		nvlist_free(label);
3689 		spa_load_failed(spa, "no valid uberblock found");
3690 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO));
3691 	}
3692 
3693 	if (spa->spa_load_max_txg != UINT64_MAX) {
3694 		(void) spa_import_progress_set_max_txg(spa_guid(spa),
3695 		    (u_longlong_t)spa->spa_load_max_txg);
3696 	}
3697 	spa_load_note(spa, "using uberblock with txg=%llu",
3698 	    (u_longlong_t)ub->ub_txg);
3699 
3700 
3701 	/*
3702 	 * For pools which have the multihost property on determine if the
3703 	 * pool is truly inactive and can be safely imported.  Prevent
3704 	 * hosts which don't have a hostid set from importing the pool.
3705 	 */
3706 	activity_check = spa_activity_check_required(spa, ub, label,
3707 	    spa->spa_config);
3708 	if (activity_check) {
3709 		if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay &&
3710 		    spa_get_hostid(spa) == 0) {
3711 			nvlist_free(label);
3712 			fnvlist_add_uint64(spa->spa_load_info,
3713 			    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
3714 			return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
3715 		}
3716 
3717 		int error = spa_activity_check(spa, ub, spa->spa_config);
3718 		if (error) {
3719 			nvlist_free(label);
3720 			return (error);
3721 		}
3722 
3723 		fnvlist_add_uint64(spa->spa_load_info,
3724 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE);
3725 		fnvlist_add_uint64(spa->spa_load_info,
3726 		    ZPOOL_CONFIG_MMP_TXG, ub->ub_txg);
3727 		fnvlist_add_uint16(spa->spa_load_info,
3728 		    ZPOOL_CONFIG_MMP_SEQ,
3729 		    (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0));
3730 	}
3731 
3732 	/*
3733 	 * If the pool has an unsupported version we can't open it.
3734 	 */
3735 	if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) {
3736 		nvlist_free(label);
3737 		spa_load_failed(spa, "version %llu is not supported",
3738 		    (u_longlong_t)ub->ub_version);
3739 		return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP));
3740 	}
3741 
3742 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
3743 		nvlist_t *features;
3744 
3745 		/*
3746 		 * If we weren't able to find what's necessary for reading the
3747 		 * MOS in the label, return failure.
3748 		 */
3749 		if (label == NULL) {
3750 			spa_load_failed(spa, "label config unavailable");
3751 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3752 			    ENXIO));
3753 		}
3754 
3755 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ,
3756 		    &features) != 0) {
3757 			nvlist_free(label);
3758 			spa_load_failed(spa, "invalid label: '%s' missing",
3759 			    ZPOOL_CONFIG_FEATURES_FOR_READ);
3760 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3761 			    ENXIO));
3762 		}
3763 
3764 		/*
3765 		 * Update our in-core representation with the definitive values
3766 		 * from the label.
3767 		 */
3768 		nvlist_free(spa->spa_label_features);
3769 		spa->spa_label_features = fnvlist_dup(features);
3770 	}
3771 
3772 	nvlist_free(label);
3773 
3774 	/*
3775 	 * Look through entries in the label nvlist's features_for_read. If
3776 	 * there is a feature listed there which we don't understand then we
3777 	 * cannot open a pool.
3778 	 */
3779 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
3780 		nvlist_t *unsup_feat;
3781 
3782 		unsup_feat = fnvlist_alloc();
3783 
3784 		for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features,
3785 		    NULL); nvp != NULL;
3786 		    nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) {
3787 			if (!zfeature_is_supported(nvpair_name(nvp))) {
3788 				fnvlist_add_string(unsup_feat,
3789 				    nvpair_name(nvp), "");
3790 			}
3791 		}
3792 
3793 		if (!nvlist_empty(unsup_feat)) {
3794 			fnvlist_add_nvlist(spa->spa_load_info,
3795 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat);
3796 			nvlist_free(unsup_feat);
3797 			spa_load_failed(spa, "some features are unsupported");
3798 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
3799 			    ENOTSUP));
3800 		}
3801 
3802 		nvlist_free(unsup_feat);
3803 	}
3804 
3805 	if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) {
3806 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3807 		spa_try_repair(spa, spa->spa_config);
3808 		spa_config_exit(spa, SCL_ALL, FTAG);
3809 		nvlist_free(spa->spa_config_splitting);
3810 		spa->spa_config_splitting = NULL;
3811 	}
3812 
3813 	/*
3814 	 * Initialize internal SPA structures.
3815 	 */
3816 	spa_ld_select_uberblock_done(spa, ub);
3817 
3818 	return (0);
3819 }
3820 
3821 static int
3822 spa_ld_open_rootbp(spa_t *spa)
3823 {
3824 	int error = 0;
3825 	vdev_t *rvd = spa->spa_root_vdev;
3826 
3827 	error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool);
3828 	if (error != 0) {
3829 		spa_load_failed(spa, "unable to open rootbp in dsl_pool_init "
3830 		    "[error=%d]", error);
3831 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3832 	}
3833 	spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset;
3834 
3835 	return (0);
3836 }
3837 
3838 static int
3839 spa_ld_trusted_config(spa_t *spa, spa_import_type_t type,
3840     boolean_t reloading)
3841 {
3842 	vdev_t *mrvd, *rvd = spa->spa_root_vdev;
3843 	nvlist_t *nv, *mos_config, *policy;
3844 	int error = 0, copy_error;
3845 	uint64_t healthy_tvds, healthy_tvds_mos;
3846 	uint64_t mos_config_txg;
3847 
3848 	if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE)
3849 	    != 0)
3850 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3851 
3852 	/*
3853 	 * If we're assembling a pool from a split, the config provided is
3854 	 * already trusted so there is nothing to do.
3855 	 */
3856 	if (type == SPA_IMPORT_ASSEMBLE)
3857 		return (0);
3858 
3859 	healthy_tvds = spa_healthy_core_tvds(spa);
3860 
3861 	if (load_nvlist(spa, spa->spa_config_object, &mos_config)
3862 	    != 0) {
3863 		spa_load_failed(spa, "unable to retrieve MOS config");
3864 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3865 	}
3866 
3867 	/*
3868 	 * If we are doing an open, pool owner wasn't verified yet, thus do
3869 	 * the verification here.
3870 	 */
3871 	if (spa->spa_load_state == SPA_LOAD_OPEN) {
3872 		error = spa_verify_host(spa, mos_config);
3873 		if (error != 0) {
3874 			nvlist_free(mos_config);
3875 			return (error);
3876 		}
3877 	}
3878 
3879 	nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE);
3880 
3881 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3882 
3883 	/*
3884 	 * Build a new vdev tree from the trusted config
3885 	 */
3886 	error = spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD);
3887 	if (error != 0) {
3888 		nvlist_free(mos_config);
3889 		spa_config_exit(spa, SCL_ALL, FTAG);
3890 		spa_load_failed(spa, "spa_config_parse failed [error=%d]",
3891 		    error);
3892 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
3893 	}
3894 
3895 	/*
3896 	 * Vdev paths in the MOS may be obsolete. If the untrusted config was
3897 	 * obtained by scanning /dev/dsk, then it will have the right vdev
3898 	 * paths. We update the trusted MOS config with this information.
3899 	 * We first try to copy the paths with vdev_copy_path_strict, which
3900 	 * succeeds only when both configs have exactly the same vdev tree.
3901 	 * If that fails, we fall back to a more flexible method that has a
3902 	 * best effort policy.
3903 	 */
3904 	copy_error = vdev_copy_path_strict(rvd, mrvd);
3905 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3906 		spa_load_note(spa, "provided vdev tree:");
3907 		vdev_dbgmsg_print_tree(rvd, 2);
3908 		spa_load_note(spa, "MOS vdev tree:");
3909 		vdev_dbgmsg_print_tree(mrvd, 2);
3910 	}
3911 	if (copy_error != 0) {
3912 		spa_load_note(spa, "vdev_copy_path_strict failed, falling "
3913 		    "back to vdev_copy_path_relaxed");
3914 		vdev_copy_path_relaxed(rvd, mrvd);
3915 	}
3916 
3917 	vdev_close(rvd);
3918 	vdev_free(rvd);
3919 	spa->spa_root_vdev = mrvd;
3920 	rvd = mrvd;
3921 	spa_config_exit(spa, SCL_ALL, FTAG);
3922 
3923 	/*
3924 	 * We will use spa_config if we decide to reload the spa or if spa_load
3925 	 * fails and we rewind. We must thus regenerate the config using the
3926 	 * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to
3927 	 * pass settings on how to load the pool and is not stored in the MOS.
3928 	 * We copy it over to our new, trusted config.
3929 	 */
3930 	mos_config_txg = fnvlist_lookup_uint64(mos_config,
3931 	    ZPOOL_CONFIG_POOL_TXG);
3932 	nvlist_free(mos_config);
3933 	mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE);
3934 	if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY,
3935 	    &policy) == 0)
3936 		fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy);
3937 	spa_config_set(spa, mos_config);
3938 	spa->spa_config_source = SPA_CONFIG_SRC_MOS;
3939 
3940 	/*
3941 	 * Now that we got the config from the MOS, we should be more strict
3942 	 * in checking blkptrs and can make assumptions about the consistency
3943 	 * of the vdev tree. spa_trust_config must be set to true before opening
3944 	 * vdevs in order for them to be writeable.
3945 	 */
3946 	spa->spa_trust_config = B_TRUE;
3947 
3948 	/*
3949 	 * Open and validate the new vdev tree
3950 	 */
3951 	error = spa_ld_open_vdevs(spa);
3952 	if (error != 0)
3953 		return (error);
3954 
3955 	error = spa_ld_validate_vdevs(spa);
3956 	if (error != 0)
3957 		return (error);
3958 
3959 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3960 		spa_load_note(spa, "final vdev tree:");
3961 		vdev_dbgmsg_print_tree(rvd, 2);
3962 	}
3963 
3964 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT &&
3965 	    !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) {
3966 		/*
3967 		 * Sanity check to make sure that we are indeed loading the
3968 		 * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds
3969 		 * in the config provided and they happened to be the only ones
3970 		 * to have the latest uberblock, we could involuntarily perform
3971 		 * an extreme rewind.
3972 		 */
3973 		healthy_tvds_mos = spa_healthy_core_tvds(spa);
3974 		if (healthy_tvds_mos - healthy_tvds >=
3975 		    SPA_SYNC_MIN_VDEVS) {
3976 			spa_load_note(spa, "config provided misses too many "
3977 			    "top-level vdevs compared to MOS (%lld vs %lld). ",
3978 			    (u_longlong_t)healthy_tvds,
3979 			    (u_longlong_t)healthy_tvds_mos);
3980 			spa_load_note(spa, "vdev tree:");
3981 			vdev_dbgmsg_print_tree(rvd, 2);
3982 			if (reloading) {
3983 				spa_load_failed(spa, "config was already "
3984 				    "provided from MOS. Aborting.");
3985 				return (spa_vdev_err(rvd,
3986 				    VDEV_AUX_CORRUPT_DATA, EIO));
3987 			}
3988 			spa_load_note(spa, "spa must be reloaded using MOS "
3989 			    "config");
3990 			return (SET_ERROR(EAGAIN));
3991 		}
3992 	}
3993 
3994 	error = spa_check_for_missing_logs(spa);
3995 	if (error != 0)
3996 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO));
3997 
3998 	if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) {
3999 		spa_load_failed(spa, "uberblock guid sum doesn't match MOS "
4000 		    "guid sum (%llu != %llu)",
4001 		    (u_longlong_t)spa->spa_uberblock.ub_guid_sum,
4002 		    (u_longlong_t)rvd->vdev_guid_sum);
4003 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM,
4004 		    ENXIO));
4005 	}
4006 
4007 	return (0);
4008 }
4009 
4010 static int
4011 spa_ld_open_indirect_vdev_metadata(spa_t *spa)
4012 {
4013 	int error = 0;
4014 	vdev_t *rvd = spa->spa_root_vdev;
4015 
4016 	/*
4017 	 * Everything that we read before spa_remove_init() must be stored
4018 	 * on concreted vdevs.  Therefore we do this as early as possible.
4019 	 */
4020 	error = spa_remove_init(spa);
4021 	if (error != 0) {
4022 		spa_load_failed(spa, "spa_remove_init failed [error=%d]",
4023 		    error);
4024 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4025 	}
4026 
4027 	/*
4028 	 * Retrieve information needed to condense indirect vdev mappings.
4029 	 */
4030 	error = spa_condense_init(spa);
4031 	if (error != 0) {
4032 		spa_load_failed(spa, "spa_condense_init failed [error=%d]",
4033 		    error);
4034 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4035 	}
4036 
4037 	return (0);
4038 }
4039 
4040 static int
4041 spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep)
4042 {
4043 	int error = 0;
4044 	vdev_t *rvd = spa->spa_root_vdev;
4045 
4046 	if (spa_version(spa) >= SPA_VERSION_FEATURES) {
4047 		boolean_t missing_feat_read = B_FALSE;
4048 		nvlist_t *unsup_feat, *enabled_feat;
4049 
4050 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ,
4051 		    &spa->spa_feat_for_read_obj, B_TRUE) != 0) {
4052 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4053 		}
4054 
4055 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE,
4056 		    &spa->spa_feat_for_write_obj, B_TRUE) != 0) {
4057 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4058 		}
4059 
4060 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS,
4061 		    &spa->spa_feat_desc_obj, B_TRUE) != 0) {
4062 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4063 		}
4064 
4065 		enabled_feat = fnvlist_alloc();
4066 		unsup_feat = fnvlist_alloc();
4067 
4068 		if (!spa_features_check(spa, B_FALSE,
4069 		    unsup_feat, enabled_feat))
4070 			missing_feat_read = B_TRUE;
4071 
4072 		if (spa_writeable(spa) ||
4073 		    spa->spa_load_state == SPA_LOAD_TRYIMPORT) {
4074 			if (!spa_features_check(spa, B_TRUE,
4075 			    unsup_feat, enabled_feat)) {
4076 				*missing_feat_writep = B_TRUE;
4077 			}
4078 		}
4079 
4080 		fnvlist_add_nvlist(spa->spa_load_info,
4081 		    ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat);
4082 
4083 		if (!nvlist_empty(unsup_feat)) {
4084 			fnvlist_add_nvlist(spa->spa_load_info,
4085 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat);
4086 		}
4087 
4088 		fnvlist_free(enabled_feat);
4089 		fnvlist_free(unsup_feat);
4090 
4091 		if (!missing_feat_read) {
4092 			fnvlist_add_boolean(spa->spa_load_info,
4093 			    ZPOOL_CONFIG_CAN_RDONLY);
4094 		}
4095 
4096 		/*
4097 		 * If the state is SPA_LOAD_TRYIMPORT, our objective is
4098 		 * twofold: to determine whether the pool is available for
4099 		 * import in read-write mode and (if it is not) whether the
4100 		 * pool is available for import in read-only mode. If the pool
4101 		 * is available for import in read-write mode, it is displayed
4102 		 * as available in userland; if it is not available for import
4103 		 * in read-only mode, it is displayed as unavailable in
4104 		 * userland. If the pool is available for import in read-only
4105 		 * mode but not read-write mode, it is displayed as unavailable
4106 		 * in userland with a special note that the pool is actually
4107 		 * available for open in read-only mode.
4108 		 *
4109 		 * As a result, if the state is SPA_LOAD_TRYIMPORT and we are
4110 		 * missing a feature for write, we must first determine whether
4111 		 * the pool can be opened read-only before returning to
4112 		 * userland in order to know whether to display the
4113 		 * abovementioned note.
4114 		 */
4115 		if (missing_feat_read || (*missing_feat_writep &&
4116 		    spa_writeable(spa))) {
4117 			spa_load_failed(spa, "pool uses unsupported features");
4118 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
4119 			    ENOTSUP));
4120 		}
4121 
4122 		/*
4123 		 * Load refcounts for ZFS features from disk into an in-memory
4124 		 * cache during SPA initialization.
4125 		 */
4126 		for (spa_feature_t i = 0; i < SPA_FEATURES; i++) {
4127 			uint64_t refcount;
4128 
4129 			error = feature_get_refcount_from_disk(spa,
4130 			    &spa_feature_table[i], &refcount);
4131 			if (error == 0) {
4132 				spa->spa_feat_refcount_cache[i] = refcount;
4133 			} else if (error == ENOTSUP) {
4134 				spa->spa_feat_refcount_cache[i] =
4135 				    SPA_FEATURE_DISABLED;
4136 			} else {
4137 				spa_load_failed(spa, "error getting refcount "
4138 				    "for feature %s [error=%d]",
4139 				    spa_feature_table[i].fi_guid, error);
4140 				return (spa_vdev_err(rvd,
4141 				    VDEV_AUX_CORRUPT_DATA, EIO));
4142 			}
4143 		}
4144 	}
4145 
4146 	if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) {
4147 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG,
4148 		    &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0)
4149 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4150 	}
4151 
4152 	/*
4153 	 * Encryption was added before bookmark_v2, even though bookmark_v2
4154 	 * is now a dependency. If this pool has encryption enabled without
4155 	 * bookmark_v2, trigger an errata message.
4156 	 */
4157 	if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) &&
4158 	    !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
4159 		spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION;
4160 	}
4161 
4162 	return (0);
4163 }
4164 
4165 static int
4166 spa_ld_load_special_directories(spa_t *spa)
4167 {
4168 	int error = 0;
4169 	vdev_t *rvd = spa->spa_root_vdev;
4170 
4171 	spa->spa_is_initializing = B_TRUE;
4172 	error = dsl_pool_open(spa->spa_dsl_pool);
4173 	spa->spa_is_initializing = B_FALSE;
4174 	if (error != 0) {
4175 		spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error);
4176 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4177 	}
4178 
4179 	return (0);
4180 }
4181 
4182 static int
4183 spa_ld_get_props(spa_t *spa)
4184 {
4185 	int error = 0;
4186 	uint64_t obj;
4187 	vdev_t *rvd = spa->spa_root_vdev;
4188 
4189 	/* Grab the checksum salt from the MOS. */
4190 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4191 	    DMU_POOL_CHECKSUM_SALT, 1,
4192 	    sizeof (spa->spa_cksum_salt.zcs_bytes),
4193 	    spa->spa_cksum_salt.zcs_bytes);
4194 	if (error == ENOENT) {
4195 		/* Generate a new salt for subsequent use */
4196 		(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
4197 		    sizeof (spa->spa_cksum_salt.zcs_bytes));
4198 	} else if (error != 0) {
4199 		spa_load_failed(spa, "unable to retrieve checksum salt from "
4200 		    "MOS [error=%d]", error);
4201 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4202 	}
4203 
4204 	if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0)
4205 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4206 	error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj);
4207 	if (error != 0) {
4208 		spa_load_failed(spa, "error opening deferred-frees bpobj "
4209 		    "[error=%d]", error);
4210 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4211 	}
4212 
4213 	/*
4214 	 * Load the bit that tells us to use the new accounting function
4215 	 * (raid-z deflation).  If we have an older pool, this will not
4216 	 * be present.
4217 	 */
4218 	error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE);
4219 	if (error != 0 && error != ENOENT)
4220 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4221 
4222 	error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION,
4223 	    &spa->spa_creation_version, B_FALSE);
4224 	if (error != 0 && error != ENOENT)
4225 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4226 
4227 	/*
4228 	 * Load the persistent error log.  If we have an older pool, this will
4229 	 * not be present.
4230 	 */
4231 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last,
4232 	    B_FALSE);
4233 	if (error != 0 && error != ENOENT)
4234 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4235 
4236 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB,
4237 	    &spa->spa_errlog_scrub, B_FALSE);
4238 	if (error != 0 && error != ENOENT)
4239 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4240 
4241 	/*
4242 	 * Load the livelist deletion field. If a livelist is queued for
4243 	 * deletion, indicate that in the spa
4244 	 */
4245 	error = spa_dir_prop(spa, DMU_POOL_DELETED_CLONES,
4246 	    &spa->spa_livelists_to_delete, B_FALSE);
4247 	if (error != 0 && error != ENOENT)
4248 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4249 
4250 	/*
4251 	 * Load the history object.  If we have an older pool, this
4252 	 * will not be present.
4253 	 */
4254 	error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE);
4255 	if (error != 0 && error != ENOENT)
4256 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4257 
4258 	/*
4259 	 * Load the per-vdev ZAP map. If we have an older pool, this will not
4260 	 * be present; in this case, defer its creation to a later time to
4261 	 * avoid dirtying the MOS this early / out of sync context. See
4262 	 * spa_sync_config_object.
4263 	 */
4264 
4265 	/* The sentinel is only available in the MOS config. */
4266 	nvlist_t *mos_config;
4267 	if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) {
4268 		spa_load_failed(spa, "unable to retrieve MOS config");
4269 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4270 	}
4271 
4272 	error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP,
4273 	    &spa->spa_all_vdev_zaps, B_FALSE);
4274 
4275 	if (error == ENOENT) {
4276 		VERIFY(!nvlist_exists(mos_config,
4277 		    ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
4278 		spa->spa_avz_action = AVZ_ACTION_INITIALIZE;
4279 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
4280 	} else if (error != 0) {
4281 		nvlist_free(mos_config);
4282 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4283 	} else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) {
4284 		/*
4285 		 * An older version of ZFS overwrote the sentinel value, so
4286 		 * we have orphaned per-vdev ZAPs in the MOS. Defer their
4287 		 * destruction to later; see spa_sync_config_object.
4288 		 */
4289 		spa->spa_avz_action = AVZ_ACTION_DESTROY;
4290 		/*
4291 		 * We're assuming that no vdevs have had their ZAPs created
4292 		 * before this. Better be sure of it.
4293 		 */
4294 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
4295 	}
4296 	nvlist_free(mos_config);
4297 
4298 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
4299 
4300 	error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object,
4301 	    B_FALSE);
4302 	if (error && error != ENOENT)
4303 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4304 
4305 	if (error == 0) {
4306 		uint64_t autoreplace = 0;
4307 
4308 		spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs);
4309 		spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace);
4310 		spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation);
4311 		spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode);
4312 		spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand);
4313 		spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost);
4314 		spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim);
4315 		spa->spa_autoreplace = (autoreplace != 0);
4316 	}
4317 
4318 	/*
4319 	 * If we are importing a pool with missing top-level vdevs,
4320 	 * we enforce that the pool doesn't panic or get suspended on
4321 	 * error since the likelihood of missing data is extremely high.
4322 	 */
4323 	if (spa->spa_missing_tvds > 0 &&
4324 	    spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE &&
4325 	    spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4326 		spa_load_note(spa, "forcing failmode to 'continue' "
4327 		    "as some top level vdevs are missing");
4328 		spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE;
4329 	}
4330 
4331 	return (0);
4332 }
4333 
4334 static int
4335 spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type)
4336 {
4337 	int error = 0;
4338 	vdev_t *rvd = spa->spa_root_vdev;
4339 
4340 	/*
4341 	 * If we're assembling the pool from the split-off vdevs of
4342 	 * an existing pool, we don't want to attach the spares & cache
4343 	 * devices.
4344 	 */
4345 
4346 	/*
4347 	 * Load any hot spares for this pool.
4348 	 */
4349 	error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object,
4350 	    B_FALSE);
4351 	if (error != 0 && error != ENOENT)
4352 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4353 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
4354 		ASSERT(spa_version(spa) >= SPA_VERSION_SPARES);
4355 		if (load_nvlist(spa, spa->spa_spares.sav_object,
4356 		    &spa->spa_spares.sav_config) != 0) {
4357 			spa_load_failed(spa, "error loading spares nvlist");
4358 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4359 		}
4360 
4361 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4362 		spa_load_spares(spa);
4363 		spa_config_exit(spa, SCL_ALL, FTAG);
4364 	} else if (error == 0) {
4365 		spa->spa_spares.sav_sync = B_TRUE;
4366 	}
4367 
4368 	/*
4369 	 * Load any level 2 ARC devices for this pool.
4370 	 */
4371 	error = spa_dir_prop(spa, DMU_POOL_L2CACHE,
4372 	    &spa->spa_l2cache.sav_object, B_FALSE);
4373 	if (error != 0 && error != ENOENT)
4374 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4375 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
4376 		ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE);
4377 		if (load_nvlist(spa, spa->spa_l2cache.sav_object,
4378 		    &spa->spa_l2cache.sav_config) != 0) {
4379 			spa_load_failed(spa, "error loading l2cache nvlist");
4380 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4381 		}
4382 
4383 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4384 		spa_load_l2cache(spa);
4385 		spa_config_exit(spa, SCL_ALL, FTAG);
4386 	} else if (error == 0) {
4387 		spa->spa_l2cache.sav_sync = B_TRUE;
4388 	}
4389 
4390 	return (0);
4391 }
4392 
4393 static int
4394 spa_ld_load_vdev_metadata(spa_t *spa)
4395 {
4396 	int error = 0;
4397 	vdev_t *rvd = spa->spa_root_vdev;
4398 
4399 	/*
4400 	 * If the 'multihost' property is set, then never allow a pool to
4401 	 * be imported when the system hostid is zero.  The exception to
4402 	 * this rule is zdb which is always allowed to access pools.
4403 	 */
4404 	if (spa_multihost(spa) && spa_get_hostid(spa) == 0 &&
4405 	    (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) {
4406 		fnvlist_add_uint64(spa->spa_load_info,
4407 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
4408 		return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
4409 	}
4410 
4411 	/*
4412 	 * If the 'autoreplace' property is set, then post a resource notifying
4413 	 * the ZFS DE that it should not issue any faults for unopenable
4414 	 * devices.  We also iterate over the vdevs, and post a sysevent for any
4415 	 * unopenable vdevs so that the normal autoreplace handler can take
4416 	 * over.
4417 	 */
4418 	if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4419 		spa_check_removed(spa->spa_root_vdev);
4420 		/*
4421 		 * For the import case, this is done in spa_import(), because
4422 		 * at this point we're using the spare definitions from
4423 		 * the MOS config, not necessarily from the userland config.
4424 		 */
4425 		if (spa->spa_load_state != SPA_LOAD_IMPORT) {
4426 			spa_aux_check_removed(&spa->spa_spares);
4427 			spa_aux_check_removed(&spa->spa_l2cache);
4428 		}
4429 	}
4430 
4431 	/*
4432 	 * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc.
4433 	 */
4434 	error = vdev_load(rvd);
4435 	if (error != 0) {
4436 		spa_load_failed(spa, "vdev_load failed [error=%d]", error);
4437 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4438 	}
4439 
4440 	error = spa_ld_log_spacemaps(spa);
4441 	if (error != 0) {
4442 		spa_load_failed(spa, "spa_ld_log_spacemaps failed [error=%d]",
4443 		    error);
4444 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4445 	}
4446 
4447 	/*
4448 	 * Propagate the leaf DTLs we just loaded all the way up the vdev tree.
4449 	 */
4450 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4451 	vdev_dtl_reassess(rvd, 0, 0, B_FALSE, B_FALSE);
4452 	spa_config_exit(spa, SCL_ALL, FTAG);
4453 
4454 	return (0);
4455 }
4456 
4457 static int
4458 spa_ld_load_dedup_tables(spa_t *spa)
4459 {
4460 	int error = 0;
4461 	vdev_t *rvd = spa->spa_root_vdev;
4462 
4463 	error = ddt_load(spa);
4464 	if (error != 0) {
4465 		spa_load_failed(spa, "ddt_load failed [error=%d]", error);
4466 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4467 	}
4468 
4469 	return (0);
4470 }
4471 
4472 static int
4473 spa_ld_load_brt(spa_t *spa)
4474 {
4475 	int error = 0;
4476 	vdev_t *rvd = spa->spa_root_vdev;
4477 
4478 	error = brt_load(spa);
4479 	if (error != 0) {
4480 		spa_load_failed(spa, "brt_load failed [error=%d]", error);
4481 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4482 	}
4483 
4484 	return (0);
4485 }
4486 
4487 static int
4488 spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, const char **ereport)
4489 {
4490 	vdev_t *rvd = spa->spa_root_vdev;
4491 
4492 	if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) {
4493 		boolean_t missing = spa_check_logs(spa);
4494 		if (missing) {
4495 			if (spa->spa_missing_tvds != 0) {
4496 				spa_load_note(spa, "spa_check_logs failed "
4497 				    "so dropping the logs");
4498 			} else {
4499 				*ereport = FM_EREPORT_ZFS_LOG_REPLAY;
4500 				spa_load_failed(spa, "spa_check_logs failed");
4501 				return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG,
4502 				    ENXIO));
4503 			}
4504 		}
4505 	}
4506 
4507 	return (0);
4508 }
4509 
4510 static int
4511 spa_ld_verify_pool_data(spa_t *spa)
4512 {
4513 	int error = 0;
4514 	vdev_t *rvd = spa->spa_root_vdev;
4515 
4516 	/*
4517 	 * We've successfully opened the pool, verify that we're ready
4518 	 * to start pushing transactions.
4519 	 */
4520 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4521 		error = spa_load_verify(spa);
4522 		if (error != 0) {
4523 			spa_load_failed(spa, "spa_load_verify failed "
4524 			    "[error=%d]", error);
4525 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
4526 			    error));
4527 		}
4528 	}
4529 
4530 	return (0);
4531 }
4532 
4533 static void
4534 spa_ld_claim_log_blocks(spa_t *spa)
4535 {
4536 	dmu_tx_t *tx;
4537 	dsl_pool_t *dp = spa_get_dsl(spa);
4538 
4539 	/*
4540 	 * Claim log blocks that haven't been committed yet.
4541 	 * This must all happen in a single txg.
4542 	 * Note: spa_claim_max_txg is updated by spa_claim_notify(),
4543 	 * invoked from zil_claim_log_block()'s i/o done callback.
4544 	 * Price of rollback is that we abandon the log.
4545 	 */
4546 	spa->spa_claiming = B_TRUE;
4547 
4548 	tx = dmu_tx_create_assigned(dp, spa_first_txg(spa));
4549 	(void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
4550 	    zil_claim, tx, DS_FIND_CHILDREN);
4551 	dmu_tx_commit(tx);
4552 
4553 	spa->spa_claiming = B_FALSE;
4554 
4555 	spa_set_log_state(spa, SPA_LOG_GOOD);
4556 }
4557 
4558 static void
4559 spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg,
4560     boolean_t update_config_cache)
4561 {
4562 	vdev_t *rvd = spa->spa_root_vdev;
4563 	int need_update = B_FALSE;
4564 
4565 	/*
4566 	 * If the config cache is stale, or we have uninitialized
4567 	 * metaslabs (see spa_vdev_add()), then update the config.
4568 	 *
4569 	 * If this is a verbatim import, trust the current
4570 	 * in-core spa_config and update the disk labels.
4571 	 */
4572 	if (update_config_cache || config_cache_txg != spa->spa_config_txg ||
4573 	    spa->spa_load_state == SPA_LOAD_IMPORT ||
4574 	    spa->spa_load_state == SPA_LOAD_RECOVER ||
4575 	    (spa->spa_import_flags & ZFS_IMPORT_VERBATIM))
4576 		need_update = B_TRUE;
4577 
4578 	for (int c = 0; c < rvd->vdev_children; c++)
4579 		if (rvd->vdev_child[c]->vdev_ms_array == 0)
4580 			need_update = B_TRUE;
4581 
4582 	/*
4583 	 * Update the config cache asynchronously in case we're the
4584 	 * root pool, in which case the config cache isn't writable yet.
4585 	 */
4586 	if (need_update)
4587 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
4588 }
4589 
4590 static void
4591 spa_ld_prepare_for_reload(spa_t *spa)
4592 {
4593 	spa_mode_t mode = spa->spa_mode;
4594 	int async_suspended = spa->spa_async_suspended;
4595 
4596 	spa_unload(spa);
4597 	spa_deactivate(spa);
4598 	spa_activate(spa, mode);
4599 
4600 	/*
4601 	 * We save the value of spa_async_suspended as it gets reset to 0 by
4602 	 * spa_unload(). We want to restore it back to the original value before
4603 	 * returning as we might be calling spa_async_resume() later.
4604 	 */
4605 	spa->spa_async_suspended = async_suspended;
4606 }
4607 
4608 static int
4609 spa_ld_read_checkpoint_txg(spa_t *spa)
4610 {
4611 	uberblock_t checkpoint;
4612 	int error = 0;
4613 
4614 	ASSERT0(spa->spa_checkpoint_txg);
4615 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4616 
4617 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4618 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
4619 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
4620 
4621 	if (error == ENOENT)
4622 		return (0);
4623 
4624 	if (error != 0)
4625 		return (error);
4626 
4627 	ASSERT3U(checkpoint.ub_txg, !=, 0);
4628 	ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0);
4629 	ASSERT3U(checkpoint.ub_timestamp, !=, 0);
4630 	spa->spa_checkpoint_txg = checkpoint.ub_txg;
4631 	spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp;
4632 
4633 	return (0);
4634 }
4635 
4636 static int
4637 spa_ld_mos_init(spa_t *spa, spa_import_type_t type)
4638 {
4639 	int error = 0;
4640 
4641 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4642 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
4643 
4644 	/*
4645 	 * Never trust the config that is provided unless we are assembling
4646 	 * a pool following a split.
4647 	 * This means don't trust blkptrs and the vdev tree in general. This
4648 	 * also effectively puts the spa in read-only mode since
4649 	 * spa_writeable() checks for spa_trust_config to be true.
4650 	 * We will later load a trusted config from the MOS.
4651 	 */
4652 	if (type != SPA_IMPORT_ASSEMBLE)
4653 		spa->spa_trust_config = B_FALSE;
4654 
4655 	/*
4656 	 * Parse the config provided to create a vdev tree.
4657 	 */
4658 	error = spa_ld_parse_config(spa, type);
4659 	if (error != 0)
4660 		return (error);
4661 
4662 	spa_import_progress_add(spa);
4663 
4664 	/*
4665 	 * Now that we have the vdev tree, try to open each vdev. This involves
4666 	 * opening the underlying physical device, retrieving its geometry and
4667 	 * probing the vdev with a dummy I/O. The state of each vdev will be set
4668 	 * based on the success of those operations. After this we'll be ready
4669 	 * to read from the vdevs.
4670 	 */
4671 	error = spa_ld_open_vdevs(spa);
4672 	if (error != 0)
4673 		return (error);
4674 
4675 	/*
4676 	 * Read the label of each vdev and make sure that the GUIDs stored
4677 	 * there match the GUIDs in the config provided.
4678 	 * If we're assembling a new pool that's been split off from an
4679 	 * existing pool, the labels haven't yet been updated so we skip
4680 	 * validation for now.
4681 	 */
4682 	if (type != SPA_IMPORT_ASSEMBLE) {
4683 		error = spa_ld_validate_vdevs(spa);
4684 		if (error != 0)
4685 			return (error);
4686 	}
4687 
4688 	/*
4689 	 * Read all vdev labels to find the best uberblock (i.e. latest,
4690 	 * unless spa_load_max_txg is set) and store it in spa_uberblock. We
4691 	 * get the list of features required to read blkptrs in the MOS from
4692 	 * the vdev label with the best uberblock and verify that our version
4693 	 * of zfs supports them all.
4694 	 */
4695 	error = spa_ld_select_uberblock(spa, type);
4696 	if (error != 0)
4697 		return (error);
4698 
4699 	/*
4700 	 * Pass that uberblock to the dsl_pool layer which will open the root
4701 	 * blkptr. This blkptr points to the latest version of the MOS and will
4702 	 * allow us to read its contents.
4703 	 */
4704 	error = spa_ld_open_rootbp(spa);
4705 	if (error != 0)
4706 		return (error);
4707 
4708 	return (0);
4709 }
4710 
4711 static int
4712 spa_ld_checkpoint_rewind(spa_t *spa)
4713 {
4714 	uberblock_t checkpoint;
4715 	int error = 0;
4716 
4717 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4718 	ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4719 
4720 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4721 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
4722 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
4723 
4724 	if (error != 0) {
4725 		spa_load_failed(spa, "unable to retrieve checkpointed "
4726 		    "uberblock from the MOS config [error=%d]", error);
4727 
4728 		if (error == ENOENT)
4729 			error = ZFS_ERR_NO_CHECKPOINT;
4730 
4731 		return (error);
4732 	}
4733 
4734 	ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg);
4735 	ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg);
4736 
4737 	/*
4738 	 * We need to update the txg and timestamp of the checkpointed
4739 	 * uberblock to be higher than the latest one. This ensures that
4740 	 * the checkpointed uberblock is selected if we were to close and
4741 	 * reopen the pool right after we've written it in the vdev labels.
4742 	 * (also see block comment in vdev_uberblock_compare)
4743 	 */
4744 	checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1;
4745 	checkpoint.ub_timestamp = gethrestime_sec();
4746 
4747 	/*
4748 	 * Set current uberblock to be the checkpointed uberblock.
4749 	 */
4750 	spa->spa_uberblock = checkpoint;
4751 
4752 	/*
4753 	 * If we are doing a normal rewind, then the pool is open for
4754 	 * writing and we sync the "updated" checkpointed uberblock to
4755 	 * disk. Once this is done, we've basically rewound the whole
4756 	 * pool and there is no way back.
4757 	 *
4758 	 * There are cases when we don't want to attempt and sync the
4759 	 * checkpointed uberblock to disk because we are opening a
4760 	 * pool as read-only. Specifically, verifying the checkpointed
4761 	 * state with zdb, and importing the checkpointed state to get
4762 	 * a "preview" of its content.
4763 	 */
4764 	if (spa_writeable(spa)) {
4765 		vdev_t *rvd = spa->spa_root_vdev;
4766 
4767 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4768 		vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
4769 		int svdcount = 0;
4770 		int children = rvd->vdev_children;
4771 		int c0 = random_in_range(children);
4772 
4773 		for (int c = 0; c < children; c++) {
4774 			vdev_t *vd = rvd->vdev_child[(c0 + c) % children];
4775 
4776 			/* Stop when revisiting the first vdev */
4777 			if (c > 0 && svd[0] == vd)
4778 				break;
4779 
4780 			if (vd->vdev_ms_array == 0 || vd->vdev_islog ||
4781 			    !vdev_is_concrete(vd))
4782 				continue;
4783 
4784 			svd[svdcount++] = vd;
4785 			if (svdcount == SPA_SYNC_MIN_VDEVS)
4786 				break;
4787 		}
4788 		error = vdev_config_sync(svd, svdcount, spa->spa_first_txg);
4789 		if (error == 0)
4790 			spa->spa_last_synced_guid = rvd->vdev_guid;
4791 		spa_config_exit(spa, SCL_ALL, FTAG);
4792 
4793 		if (error != 0) {
4794 			spa_load_failed(spa, "failed to write checkpointed "
4795 			    "uberblock to the vdev labels [error=%d]", error);
4796 			return (error);
4797 		}
4798 	}
4799 
4800 	return (0);
4801 }
4802 
4803 static int
4804 spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type,
4805     boolean_t *update_config_cache)
4806 {
4807 	int error;
4808 
4809 	/*
4810 	 * Parse the config for pool, open and validate vdevs,
4811 	 * select an uberblock, and use that uberblock to open
4812 	 * the MOS.
4813 	 */
4814 	error = spa_ld_mos_init(spa, type);
4815 	if (error != 0)
4816 		return (error);
4817 
4818 	/*
4819 	 * Retrieve the trusted config stored in the MOS and use it to create
4820 	 * a new, exact version of the vdev tree, then reopen all vdevs.
4821 	 */
4822 	error = spa_ld_trusted_config(spa, type, B_FALSE);
4823 	if (error == EAGAIN) {
4824 		if (update_config_cache != NULL)
4825 			*update_config_cache = B_TRUE;
4826 
4827 		/*
4828 		 * Redo the loading process with the trusted config if it is
4829 		 * too different from the untrusted config.
4830 		 */
4831 		spa_ld_prepare_for_reload(spa);
4832 		spa_load_note(spa, "RELOADING");
4833 		error = spa_ld_mos_init(spa, type);
4834 		if (error != 0)
4835 			return (error);
4836 
4837 		error = spa_ld_trusted_config(spa, type, B_TRUE);
4838 		if (error != 0)
4839 			return (error);
4840 
4841 	} else if (error != 0) {
4842 		return (error);
4843 	}
4844 
4845 	return (0);
4846 }
4847 
4848 /*
4849  * Load an existing storage pool, using the config provided. This config
4850  * describes which vdevs are part of the pool and is later validated against
4851  * partial configs present in each vdev's label and an entire copy of the
4852  * config stored in the MOS.
4853  */
4854 static int
4855 spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
4856 {
4857 	int error = 0;
4858 	boolean_t missing_feat_write = B_FALSE;
4859 	boolean_t checkpoint_rewind =
4860 	    (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4861 	boolean_t update_config_cache = B_FALSE;
4862 
4863 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4864 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
4865 
4866 	spa_load_note(spa, "LOADING");
4867 
4868 	error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache);
4869 	if (error != 0)
4870 		return (error);
4871 
4872 	/*
4873 	 * If we are rewinding to the checkpoint then we need to repeat
4874 	 * everything we've done so far in this function but this time
4875 	 * selecting the checkpointed uberblock and using that to open
4876 	 * the MOS.
4877 	 */
4878 	if (checkpoint_rewind) {
4879 		/*
4880 		 * If we are rewinding to the checkpoint update config cache
4881 		 * anyway.
4882 		 */
4883 		update_config_cache = B_TRUE;
4884 
4885 		/*
4886 		 * Extract the checkpointed uberblock from the current MOS
4887 		 * and use this as the pool's uberblock from now on. If the
4888 		 * pool is imported as writeable we also write the checkpoint
4889 		 * uberblock to the labels, making the rewind permanent.
4890 		 */
4891 		error = spa_ld_checkpoint_rewind(spa);
4892 		if (error != 0)
4893 			return (error);
4894 
4895 		/*
4896 		 * Redo the loading process again with the
4897 		 * checkpointed uberblock.
4898 		 */
4899 		spa_ld_prepare_for_reload(spa);
4900 		spa_load_note(spa, "LOADING checkpointed uberblock");
4901 		error = spa_ld_mos_with_trusted_config(spa, type, NULL);
4902 		if (error != 0)
4903 			return (error);
4904 	}
4905 
4906 	/*
4907 	 * Retrieve the checkpoint txg if the pool has a checkpoint.
4908 	 */
4909 	error = spa_ld_read_checkpoint_txg(spa);
4910 	if (error != 0)
4911 		return (error);
4912 
4913 	/*
4914 	 * Retrieve the mapping of indirect vdevs. Those vdevs were removed
4915 	 * from the pool and their contents were re-mapped to other vdevs. Note
4916 	 * that everything that we read before this step must have been
4917 	 * rewritten on concrete vdevs after the last device removal was
4918 	 * initiated. Otherwise we could be reading from indirect vdevs before
4919 	 * we have loaded their mappings.
4920 	 */
4921 	error = spa_ld_open_indirect_vdev_metadata(spa);
4922 	if (error != 0)
4923 		return (error);
4924 
4925 	/*
4926 	 * Retrieve the full list of active features from the MOS and check if
4927 	 * they are all supported.
4928 	 */
4929 	error = spa_ld_check_features(spa, &missing_feat_write);
4930 	if (error != 0)
4931 		return (error);
4932 
4933 	/*
4934 	 * Load several special directories from the MOS needed by the dsl_pool
4935 	 * layer.
4936 	 */
4937 	error = spa_ld_load_special_directories(spa);
4938 	if (error != 0)
4939 		return (error);
4940 
4941 	/*
4942 	 * Retrieve pool properties from the MOS.
4943 	 */
4944 	error = spa_ld_get_props(spa);
4945 	if (error != 0)
4946 		return (error);
4947 
4948 	/*
4949 	 * Retrieve the list of auxiliary devices - cache devices and spares -
4950 	 * and open them.
4951 	 */
4952 	error = spa_ld_open_aux_vdevs(spa, type);
4953 	if (error != 0)
4954 		return (error);
4955 
4956 	/*
4957 	 * Load the metadata for all vdevs. Also check if unopenable devices
4958 	 * should be autoreplaced.
4959 	 */
4960 	error = spa_ld_load_vdev_metadata(spa);
4961 	if (error != 0)
4962 		return (error);
4963 
4964 	error = spa_ld_load_dedup_tables(spa);
4965 	if (error != 0)
4966 		return (error);
4967 
4968 	error = spa_ld_load_brt(spa);
4969 	if (error != 0)
4970 		return (error);
4971 
4972 	/*
4973 	 * Verify the logs now to make sure we don't have any unexpected errors
4974 	 * when we claim log blocks later.
4975 	 */
4976 	error = spa_ld_verify_logs(spa, type, ereport);
4977 	if (error != 0)
4978 		return (error);
4979 
4980 	if (missing_feat_write) {
4981 		ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT);
4982 
4983 		/*
4984 		 * At this point, we know that we can open the pool in
4985 		 * read-only mode but not read-write mode. We now have enough
4986 		 * information and can return to userland.
4987 		 */
4988 		return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT,
4989 		    ENOTSUP));
4990 	}
4991 
4992 	/*
4993 	 * Traverse the last txgs to make sure the pool was left off in a safe
4994 	 * state. When performing an extreme rewind, we verify the whole pool,
4995 	 * which can take a very long time.
4996 	 */
4997 	error = spa_ld_verify_pool_data(spa);
4998 	if (error != 0)
4999 		return (error);
5000 
5001 	/*
5002 	 * Calculate the deflated space for the pool. This must be done before
5003 	 * we write anything to the pool because we'd need to update the space
5004 	 * accounting using the deflated sizes.
5005 	 */
5006 	spa_update_dspace(spa);
5007 
5008 	/*
5009 	 * We have now retrieved all the information we needed to open the
5010 	 * pool. If we are importing the pool in read-write mode, a few
5011 	 * additional steps must be performed to finish the import.
5012 	 */
5013 	if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER ||
5014 	    spa->spa_load_max_txg == UINT64_MAX)) {
5015 		uint64_t config_cache_txg = spa->spa_config_txg;
5016 
5017 		ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT);
5018 
5019 		/*
5020 		 * In case of a checkpoint rewind, log the original txg
5021 		 * of the checkpointed uberblock.
5022 		 */
5023 		if (checkpoint_rewind) {
5024 			spa_history_log_internal(spa, "checkpoint rewind",
5025 			    NULL, "rewound state to txg=%llu",
5026 			    (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg);
5027 		}
5028 
5029 		/*
5030 		 * Traverse the ZIL and claim all blocks.
5031 		 */
5032 		spa_ld_claim_log_blocks(spa);
5033 
5034 		/*
5035 		 * Kick-off the syncing thread.
5036 		 */
5037 		spa->spa_sync_on = B_TRUE;
5038 		txg_sync_start(spa->spa_dsl_pool);
5039 		mmp_thread_start(spa);
5040 
5041 		/*
5042 		 * Wait for all claims to sync.  We sync up to the highest
5043 		 * claimed log block birth time so that claimed log blocks
5044 		 * don't appear to be from the future.  spa_claim_max_txg
5045 		 * will have been set for us by ZIL traversal operations
5046 		 * performed above.
5047 		 */
5048 		txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg);
5049 
5050 		/*
5051 		 * Check if we need to request an update of the config. On the
5052 		 * next sync, we would update the config stored in vdev labels
5053 		 * and the cachefile (by default /etc/zfs/zpool.cache).
5054 		 */
5055 		spa_ld_check_for_config_update(spa, config_cache_txg,
5056 		    update_config_cache);
5057 
5058 		/*
5059 		 * Check if a rebuild was in progress and if so resume it.
5060 		 * Then check all DTLs to see if anything needs resilvering.
5061 		 * The resilver will be deferred if a rebuild was started.
5062 		 */
5063 		if (vdev_rebuild_active(spa->spa_root_vdev)) {
5064 			vdev_rebuild_restart(spa);
5065 		} else if (!dsl_scan_resilvering(spa->spa_dsl_pool) &&
5066 		    vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
5067 			spa_async_request(spa, SPA_ASYNC_RESILVER);
5068 		}
5069 
5070 		/*
5071 		 * Log the fact that we booted up (so that we can detect if
5072 		 * we rebooted in the middle of an operation).
5073 		 */
5074 		spa_history_log_version(spa, "open", NULL);
5075 
5076 		spa_restart_removal(spa);
5077 		spa_spawn_aux_threads(spa);
5078 
5079 		/*
5080 		 * Delete any inconsistent datasets.
5081 		 *
5082 		 * Note:
5083 		 * Since we may be issuing deletes for clones here,
5084 		 * we make sure to do so after we've spawned all the
5085 		 * auxiliary threads above (from which the livelist
5086 		 * deletion zthr is part of).
5087 		 */
5088 		(void) dmu_objset_find(spa_name(spa),
5089 		    dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN);
5090 
5091 		/*
5092 		 * Clean up any stale temporary dataset userrefs.
5093 		 */
5094 		dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool);
5095 
5096 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
5097 		vdev_initialize_restart(spa->spa_root_vdev);
5098 		vdev_trim_restart(spa->spa_root_vdev);
5099 		vdev_autotrim_restart(spa);
5100 		spa_config_exit(spa, SCL_CONFIG, FTAG);
5101 	}
5102 
5103 	spa_import_progress_remove(spa_guid(spa));
5104 	spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD);
5105 
5106 	spa_load_note(spa, "LOADED");
5107 
5108 	return (0);
5109 }
5110 
5111 static int
5112 spa_load_retry(spa_t *spa, spa_load_state_t state)
5113 {
5114 	spa_mode_t mode = spa->spa_mode;
5115 
5116 	spa_unload(spa);
5117 	spa_deactivate(spa);
5118 
5119 	spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1;
5120 
5121 	spa_activate(spa, mode);
5122 	spa_async_suspend(spa);
5123 
5124 	spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu",
5125 	    (u_longlong_t)spa->spa_load_max_txg);
5126 
5127 	return (spa_load(spa, state, SPA_IMPORT_EXISTING));
5128 }
5129 
5130 /*
5131  * If spa_load() fails this function will try loading prior txg's. If
5132  * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool
5133  * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this
5134  * function will not rewind the pool and will return the same error as
5135  * spa_load().
5136  */
5137 static int
5138 spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request,
5139     int rewind_flags)
5140 {
5141 	nvlist_t *loadinfo = NULL;
5142 	nvlist_t *config = NULL;
5143 	int load_error, rewind_error;
5144 	uint64_t safe_rewind_txg;
5145 	uint64_t min_txg;
5146 
5147 	if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) {
5148 		spa->spa_load_max_txg = spa->spa_load_txg;
5149 		spa_set_log_state(spa, SPA_LOG_CLEAR);
5150 	} else {
5151 		spa->spa_load_max_txg = max_request;
5152 		if (max_request != UINT64_MAX)
5153 			spa->spa_extreme_rewind = B_TRUE;
5154 	}
5155 
5156 	load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING);
5157 	if (load_error == 0)
5158 		return (0);
5159 	if (load_error == ZFS_ERR_NO_CHECKPOINT) {
5160 		/*
5161 		 * When attempting checkpoint-rewind on a pool with no
5162 		 * checkpoint, we should not attempt to load uberblocks
5163 		 * from previous txgs when spa_load fails.
5164 		 */
5165 		ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
5166 		spa_import_progress_remove(spa_guid(spa));
5167 		return (load_error);
5168 	}
5169 
5170 	if (spa->spa_root_vdev != NULL)
5171 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
5172 
5173 	spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg;
5174 	spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp;
5175 
5176 	if (rewind_flags & ZPOOL_NEVER_REWIND) {
5177 		nvlist_free(config);
5178 		spa_import_progress_remove(spa_guid(spa));
5179 		return (load_error);
5180 	}
5181 
5182 	if (state == SPA_LOAD_RECOVER) {
5183 		/* Price of rolling back is discarding txgs, including log */
5184 		spa_set_log_state(spa, SPA_LOG_CLEAR);
5185 	} else {
5186 		/*
5187 		 * If we aren't rolling back save the load info from our first
5188 		 * import attempt so that we can restore it after attempting
5189 		 * to rewind.
5190 		 */
5191 		loadinfo = spa->spa_load_info;
5192 		spa->spa_load_info = fnvlist_alloc();
5193 	}
5194 
5195 	spa->spa_load_max_txg = spa->spa_last_ubsync_txg;
5196 	safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE;
5197 	min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ?
5198 	    TXG_INITIAL : safe_rewind_txg;
5199 
5200 	/*
5201 	 * Continue as long as we're finding errors, we're still within
5202 	 * the acceptable rewind range, and we're still finding uberblocks
5203 	 */
5204 	while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg &&
5205 	    spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) {
5206 		if (spa->spa_load_max_txg < safe_rewind_txg)
5207 			spa->spa_extreme_rewind = B_TRUE;
5208 		rewind_error = spa_load_retry(spa, state);
5209 	}
5210 
5211 	spa->spa_extreme_rewind = B_FALSE;
5212 	spa->spa_load_max_txg = UINT64_MAX;
5213 
5214 	if (config && (rewind_error || state != SPA_LOAD_RECOVER))
5215 		spa_config_set(spa, config);
5216 	else
5217 		nvlist_free(config);
5218 
5219 	if (state == SPA_LOAD_RECOVER) {
5220 		ASSERT3P(loadinfo, ==, NULL);
5221 		spa_import_progress_remove(spa_guid(spa));
5222 		return (rewind_error);
5223 	} else {
5224 		/* Store the rewind info as part of the initial load info */
5225 		fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO,
5226 		    spa->spa_load_info);
5227 
5228 		/* Restore the initial load info */
5229 		fnvlist_free(spa->spa_load_info);
5230 		spa->spa_load_info = loadinfo;
5231 
5232 		spa_import_progress_remove(spa_guid(spa));
5233 		return (load_error);
5234 	}
5235 }
5236 
5237 /*
5238  * Pool Open/Import
5239  *
5240  * The import case is identical to an open except that the configuration is sent
5241  * down from userland, instead of grabbed from the configuration cache.  For the
5242  * case of an open, the pool configuration will exist in the
5243  * POOL_STATE_UNINITIALIZED state.
5244  *
5245  * The stats information (gen/count/ustats) is used to gather vdev statistics at
5246  * the same time open the pool, without having to keep around the spa_t in some
5247  * ambiguous state.
5248  */
5249 static int
5250 spa_open_common(const char *pool, spa_t **spapp, const void *tag,
5251     nvlist_t *nvpolicy, nvlist_t **config)
5252 {
5253 	spa_t *spa;
5254 	spa_load_state_t state = SPA_LOAD_OPEN;
5255 	int error;
5256 	int locked = B_FALSE;
5257 	int firstopen = B_FALSE;
5258 
5259 	*spapp = NULL;
5260 
5261 	/*
5262 	 * As disgusting as this is, we need to support recursive calls to this
5263 	 * function because dsl_dir_open() is called during spa_load(), and ends
5264 	 * up calling spa_open() again.  The real fix is to figure out how to
5265 	 * avoid dsl_dir_open() calling this in the first place.
5266 	 */
5267 	if (MUTEX_NOT_HELD(&spa_namespace_lock)) {
5268 		mutex_enter(&spa_namespace_lock);
5269 		locked = B_TRUE;
5270 	}
5271 
5272 	if ((spa = spa_lookup(pool)) == NULL) {
5273 		if (locked)
5274 			mutex_exit(&spa_namespace_lock);
5275 		return (SET_ERROR(ENOENT));
5276 	}
5277 
5278 	if (spa->spa_state == POOL_STATE_UNINITIALIZED) {
5279 		zpool_load_policy_t policy;
5280 
5281 		firstopen = B_TRUE;
5282 
5283 		zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config,
5284 		    &policy);
5285 		if (policy.zlp_rewind & ZPOOL_DO_REWIND)
5286 			state = SPA_LOAD_RECOVER;
5287 
5288 		spa_activate(spa, spa_mode_global);
5289 
5290 		if (state != SPA_LOAD_RECOVER)
5291 			spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
5292 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
5293 
5294 		zfs_dbgmsg("spa_open_common: opening %s", pool);
5295 		error = spa_load_best(spa, state, policy.zlp_txg,
5296 		    policy.zlp_rewind);
5297 
5298 		if (error == EBADF) {
5299 			/*
5300 			 * If vdev_validate() returns failure (indicated by
5301 			 * EBADF), it indicates that one of the vdevs indicates
5302 			 * that the pool has been exported or destroyed.  If
5303 			 * this is the case, the config cache is out of sync and
5304 			 * we should remove the pool from the namespace.
5305 			 */
5306 			spa_unload(spa);
5307 			spa_deactivate(spa);
5308 			spa_write_cachefile(spa, B_TRUE, B_TRUE, B_FALSE);
5309 			spa_remove(spa);
5310 			if (locked)
5311 				mutex_exit(&spa_namespace_lock);
5312 			return (SET_ERROR(ENOENT));
5313 		}
5314 
5315 		if (error) {
5316 			/*
5317 			 * We can't open the pool, but we still have useful
5318 			 * information: the state of each vdev after the
5319 			 * attempted vdev_open().  Return this to the user.
5320 			 */
5321 			if (config != NULL && spa->spa_config) {
5322 				*config = fnvlist_dup(spa->spa_config);
5323 				fnvlist_add_nvlist(*config,
5324 				    ZPOOL_CONFIG_LOAD_INFO,
5325 				    spa->spa_load_info);
5326 			}
5327 			spa_unload(spa);
5328 			spa_deactivate(spa);
5329 			spa->spa_last_open_failed = error;
5330 			if (locked)
5331 				mutex_exit(&spa_namespace_lock);
5332 			*spapp = NULL;
5333 			return (error);
5334 		}
5335 	}
5336 
5337 	spa_open_ref(spa, tag);
5338 
5339 	if (config != NULL)
5340 		*config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
5341 
5342 	/*
5343 	 * If we've recovered the pool, pass back any information we
5344 	 * gathered while doing the load.
5345 	 */
5346 	if (state == SPA_LOAD_RECOVER && config != NULL) {
5347 		fnvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO,
5348 		    spa->spa_load_info);
5349 	}
5350 
5351 	if (locked) {
5352 		spa->spa_last_open_failed = 0;
5353 		spa->spa_last_ubsync_txg = 0;
5354 		spa->spa_load_txg = 0;
5355 		mutex_exit(&spa_namespace_lock);
5356 	}
5357 
5358 	if (firstopen)
5359 		zvol_create_minors_recursive(spa_name(spa));
5360 
5361 	*spapp = spa;
5362 
5363 	return (0);
5364 }
5365 
5366 int
5367 spa_open_rewind(const char *name, spa_t **spapp, const void *tag,
5368     nvlist_t *policy, nvlist_t **config)
5369 {
5370 	return (spa_open_common(name, spapp, tag, policy, config));
5371 }
5372 
5373 int
5374 spa_open(const char *name, spa_t **spapp, const void *tag)
5375 {
5376 	return (spa_open_common(name, spapp, tag, NULL, NULL));
5377 }
5378 
5379 /*
5380  * Lookup the given spa_t, incrementing the inject count in the process,
5381  * preventing it from being exported or destroyed.
5382  */
5383 spa_t *
5384 spa_inject_addref(char *name)
5385 {
5386 	spa_t *spa;
5387 
5388 	mutex_enter(&spa_namespace_lock);
5389 	if ((spa = spa_lookup(name)) == NULL) {
5390 		mutex_exit(&spa_namespace_lock);
5391 		return (NULL);
5392 	}
5393 	spa->spa_inject_ref++;
5394 	mutex_exit(&spa_namespace_lock);
5395 
5396 	return (spa);
5397 }
5398 
5399 void
5400 spa_inject_delref(spa_t *spa)
5401 {
5402 	mutex_enter(&spa_namespace_lock);
5403 	spa->spa_inject_ref--;
5404 	mutex_exit(&spa_namespace_lock);
5405 }
5406 
5407 /*
5408  * Add spares device information to the nvlist.
5409  */
5410 static void
5411 spa_add_spares(spa_t *spa, nvlist_t *config)
5412 {
5413 	nvlist_t **spares;
5414 	uint_t i, nspares;
5415 	nvlist_t *nvroot;
5416 	uint64_t guid;
5417 	vdev_stat_t *vs;
5418 	uint_t vsc;
5419 	uint64_t pool;
5420 
5421 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5422 
5423 	if (spa->spa_spares.sav_count == 0)
5424 		return;
5425 
5426 	nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE);
5427 	VERIFY0(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
5428 	    ZPOOL_CONFIG_SPARES, &spares, &nspares));
5429 	if (nspares != 0) {
5430 		fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5431 		    (const nvlist_t * const *)spares, nspares);
5432 		VERIFY0(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5433 		    &spares, &nspares));
5434 
5435 		/*
5436 		 * Go through and find any spares which have since been
5437 		 * repurposed as an active spare.  If this is the case, update
5438 		 * their status appropriately.
5439 		 */
5440 		for (i = 0; i < nspares; i++) {
5441 			guid = fnvlist_lookup_uint64(spares[i],
5442 			    ZPOOL_CONFIG_GUID);
5443 			VERIFY0(nvlist_lookup_uint64_array(spares[i],
5444 			    ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc));
5445 			if (spa_spare_exists(guid, &pool, NULL) &&
5446 			    pool != 0ULL) {
5447 				vs->vs_state = VDEV_STATE_CANT_OPEN;
5448 				vs->vs_aux = VDEV_AUX_SPARED;
5449 			} else {
5450 				vs->vs_state =
5451 				    spa->spa_spares.sav_vdevs[i]->vdev_state;
5452 			}
5453 		}
5454 	}
5455 }
5456 
5457 /*
5458  * Add l2cache device information to the nvlist, including vdev stats.
5459  */
5460 static void
5461 spa_add_l2cache(spa_t *spa, nvlist_t *config)
5462 {
5463 	nvlist_t **l2cache;
5464 	uint_t i, j, nl2cache;
5465 	nvlist_t *nvroot;
5466 	uint64_t guid;
5467 	vdev_t *vd;
5468 	vdev_stat_t *vs;
5469 	uint_t vsc;
5470 
5471 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5472 
5473 	if (spa->spa_l2cache.sav_count == 0)
5474 		return;
5475 
5476 	nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE);
5477 	VERIFY0(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
5478 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache));
5479 	if (nl2cache != 0) {
5480 		fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5481 		    (const nvlist_t * const *)l2cache, nl2cache);
5482 		VERIFY0(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5483 		    &l2cache, &nl2cache));
5484 
5485 		/*
5486 		 * Update level 2 cache device stats.
5487 		 */
5488 
5489 		for (i = 0; i < nl2cache; i++) {
5490 			guid = fnvlist_lookup_uint64(l2cache[i],
5491 			    ZPOOL_CONFIG_GUID);
5492 
5493 			vd = NULL;
5494 			for (j = 0; j < spa->spa_l2cache.sav_count; j++) {
5495 				if (guid ==
5496 				    spa->spa_l2cache.sav_vdevs[j]->vdev_guid) {
5497 					vd = spa->spa_l2cache.sav_vdevs[j];
5498 					break;
5499 				}
5500 			}
5501 			ASSERT(vd != NULL);
5502 
5503 			VERIFY0(nvlist_lookup_uint64_array(l2cache[i],
5504 			    ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc));
5505 			vdev_get_stats(vd, vs);
5506 			vdev_config_generate_stats(vd, l2cache[i]);
5507 
5508 		}
5509 	}
5510 }
5511 
5512 static void
5513 spa_feature_stats_from_disk(spa_t *spa, nvlist_t *features)
5514 {
5515 	zap_cursor_t zc;
5516 	zap_attribute_t za;
5517 
5518 	if (spa->spa_feat_for_read_obj != 0) {
5519 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
5520 		    spa->spa_feat_for_read_obj);
5521 		    zap_cursor_retrieve(&zc, &za) == 0;
5522 		    zap_cursor_advance(&zc)) {
5523 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
5524 			    za.za_num_integers == 1);
5525 			VERIFY0(nvlist_add_uint64(features, za.za_name,
5526 			    za.za_first_integer));
5527 		}
5528 		zap_cursor_fini(&zc);
5529 	}
5530 
5531 	if (spa->spa_feat_for_write_obj != 0) {
5532 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
5533 		    spa->spa_feat_for_write_obj);
5534 		    zap_cursor_retrieve(&zc, &za) == 0;
5535 		    zap_cursor_advance(&zc)) {
5536 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
5537 			    za.za_num_integers == 1);
5538 			VERIFY0(nvlist_add_uint64(features, za.za_name,
5539 			    za.za_first_integer));
5540 		}
5541 		zap_cursor_fini(&zc);
5542 	}
5543 }
5544 
5545 static void
5546 spa_feature_stats_from_cache(spa_t *spa, nvlist_t *features)
5547 {
5548 	int i;
5549 
5550 	for (i = 0; i < SPA_FEATURES; i++) {
5551 		zfeature_info_t feature = spa_feature_table[i];
5552 		uint64_t refcount;
5553 
5554 		if (feature_get_refcount(spa, &feature, &refcount) != 0)
5555 			continue;
5556 
5557 		VERIFY0(nvlist_add_uint64(features, feature.fi_guid, refcount));
5558 	}
5559 }
5560 
5561 /*
5562  * Store a list of pool features and their reference counts in the
5563  * config.
5564  *
5565  * The first time this is called on a spa, allocate a new nvlist, fetch
5566  * the pool features and reference counts from disk, then save the list
5567  * in the spa. In subsequent calls on the same spa use the saved nvlist
5568  * and refresh its values from the cached reference counts.  This
5569  * ensures we don't block here on I/O on a suspended pool so 'zpool
5570  * clear' can resume the pool.
5571  */
5572 static void
5573 spa_add_feature_stats(spa_t *spa, nvlist_t *config)
5574 {
5575 	nvlist_t *features;
5576 
5577 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5578 
5579 	mutex_enter(&spa->spa_feat_stats_lock);
5580 	features = spa->spa_feat_stats;
5581 
5582 	if (features != NULL) {
5583 		spa_feature_stats_from_cache(spa, features);
5584 	} else {
5585 		VERIFY0(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP));
5586 		spa->spa_feat_stats = features;
5587 		spa_feature_stats_from_disk(spa, features);
5588 	}
5589 
5590 	VERIFY0(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS,
5591 	    features));
5592 
5593 	mutex_exit(&spa->spa_feat_stats_lock);
5594 }
5595 
5596 int
5597 spa_get_stats(const char *name, nvlist_t **config,
5598     char *altroot, size_t buflen)
5599 {
5600 	int error;
5601 	spa_t *spa;
5602 
5603 	*config = NULL;
5604 	error = spa_open_common(name, &spa, FTAG, NULL, config);
5605 
5606 	if (spa != NULL) {
5607 		/*
5608 		 * This still leaves a window of inconsistency where the spares
5609 		 * or l2cache devices could change and the config would be
5610 		 * self-inconsistent.
5611 		 */
5612 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
5613 
5614 		if (*config != NULL) {
5615 			uint64_t loadtimes[2];
5616 
5617 			loadtimes[0] = spa->spa_loaded_ts.tv_sec;
5618 			loadtimes[1] = spa->spa_loaded_ts.tv_nsec;
5619 			fnvlist_add_uint64_array(*config,
5620 			    ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2);
5621 
5622 			fnvlist_add_uint64(*config,
5623 			    ZPOOL_CONFIG_ERRCOUNT,
5624 			    spa_approx_errlog_size(spa));
5625 
5626 			if (spa_suspended(spa)) {
5627 				fnvlist_add_uint64(*config,
5628 				    ZPOOL_CONFIG_SUSPENDED,
5629 				    spa->spa_failmode);
5630 				fnvlist_add_uint64(*config,
5631 				    ZPOOL_CONFIG_SUSPENDED_REASON,
5632 				    spa->spa_suspended);
5633 			}
5634 
5635 			spa_add_spares(spa, *config);
5636 			spa_add_l2cache(spa, *config);
5637 			spa_add_feature_stats(spa, *config);
5638 		}
5639 	}
5640 
5641 	/*
5642 	 * We want to get the alternate root even for faulted pools, so we cheat
5643 	 * and call spa_lookup() directly.
5644 	 */
5645 	if (altroot) {
5646 		if (spa == NULL) {
5647 			mutex_enter(&spa_namespace_lock);
5648 			spa = spa_lookup(name);
5649 			if (spa)
5650 				spa_altroot(spa, altroot, buflen);
5651 			else
5652 				altroot[0] = '\0';
5653 			spa = NULL;
5654 			mutex_exit(&spa_namespace_lock);
5655 		} else {
5656 			spa_altroot(spa, altroot, buflen);
5657 		}
5658 	}
5659 
5660 	if (spa != NULL) {
5661 		spa_config_exit(spa, SCL_CONFIG, FTAG);
5662 		spa_close(spa, FTAG);
5663 	}
5664 
5665 	return (error);
5666 }
5667 
5668 /*
5669  * Validate that the auxiliary device array is well formed.  We must have an
5670  * array of nvlists, each which describes a valid leaf vdev.  If this is an
5671  * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be
5672  * specified, as long as they are well-formed.
5673  */
5674 static int
5675 spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode,
5676     spa_aux_vdev_t *sav, const char *config, uint64_t version,
5677     vdev_labeltype_t label)
5678 {
5679 	nvlist_t **dev;
5680 	uint_t i, ndev;
5681 	vdev_t *vd;
5682 	int error;
5683 
5684 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5685 
5686 	/*
5687 	 * It's acceptable to have no devs specified.
5688 	 */
5689 	if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0)
5690 		return (0);
5691 
5692 	if (ndev == 0)
5693 		return (SET_ERROR(EINVAL));
5694 
5695 	/*
5696 	 * Make sure the pool is formatted with a version that supports this
5697 	 * device type.
5698 	 */
5699 	if (spa_version(spa) < version)
5700 		return (SET_ERROR(ENOTSUP));
5701 
5702 	/*
5703 	 * Set the pending device list so we correctly handle device in-use
5704 	 * checking.
5705 	 */
5706 	sav->sav_pending = dev;
5707 	sav->sav_npending = ndev;
5708 
5709 	for (i = 0; i < ndev; i++) {
5710 		if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0,
5711 		    mode)) != 0)
5712 			goto out;
5713 
5714 		if (!vd->vdev_ops->vdev_op_leaf) {
5715 			vdev_free(vd);
5716 			error = SET_ERROR(EINVAL);
5717 			goto out;
5718 		}
5719 
5720 		vd->vdev_top = vd;
5721 
5722 		if ((error = vdev_open(vd)) == 0 &&
5723 		    (error = vdev_label_init(vd, crtxg, label)) == 0) {
5724 			fnvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID,
5725 			    vd->vdev_guid);
5726 		}
5727 
5728 		vdev_free(vd);
5729 
5730 		if (error &&
5731 		    (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE))
5732 			goto out;
5733 		else
5734 			error = 0;
5735 	}
5736 
5737 out:
5738 	sav->sav_pending = NULL;
5739 	sav->sav_npending = 0;
5740 	return (error);
5741 }
5742 
5743 static int
5744 spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode)
5745 {
5746 	int error;
5747 
5748 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5749 
5750 	if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode,
5751 	    &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES,
5752 	    VDEV_LABEL_SPARE)) != 0) {
5753 		return (error);
5754 	}
5755 
5756 	return (spa_validate_aux_devs(spa, nvroot, crtxg, mode,
5757 	    &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE,
5758 	    VDEV_LABEL_L2CACHE));
5759 }
5760 
5761 static void
5762 spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs,
5763     const char *config)
5764 {
5765 	int i;
5766 
5767 	if (sav->sav_config != NULL) {
5768 		nvlist_t **olddevs;
5769 		uint_t oldndevs;
5770 		nvlist_t **newdevs;
5771 
5772 		/*
5773 		 * Generate new dev list by concatenating with the
5774 		 * current dev list.
5775 		 */
5776 		VERIFY0(nvlist_lookup_nvlist_array(sav->sav_config, config,
5777 		    &olddevs, &oldndevs));
5778 
5779 		newdevs = kmem_alloc(sizeof (void *) *
5780 		    (ndevs + oldndevs), KM_SLEEP);
5781 		for (i = 0; i < oldndevs; i++)
5782 			newdevs[i] = fnvlist_dup(olddevs[i]);
5783 		for (i = 0; i < ndevs; i++)
5784 			newdevs[i + oldndevs] = fnvlist_dup(devs[i]);
5785 
5786 		fnvlist_remove(sav->sav_config, config);
5787 
5788 		fnvlist_add_nvlist_array(sav->sav_config, config,
5789 		    (const nvlist_t * const *)newdevs, ndevs + oldndevs);
5790 		for (i = 0; i < oldndevs + ndevs; i++)
5791 			nvlist_free(newdevs[i]);
5792 		kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *));
5793 	} else {
5794 		/*
5795 		 * Generate a new dev list.
5796 		 */
5797 		sav->sav_config = fnvlist_alloc();
5798 		fnvlist_add_nvlist_array(sav->sav_config, config,
5799 		    (const nvlist_t * const *)devs, ndevs);
5800 	}
5801 }
5802 
5803 /*
5804  * Stop and drop level 2 ARC devices
5805  */
5806 void
5807 spa_l2cache_drop(spa_t *spa)
5808 {
5809 	vdev_t *vd;
5810 	int i;
5811 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
5812 
5813 	for (i = 0; i < sav->sav_count; i++) {
5814 		uint64_t pool;
5815 
5816 		vd = sav->sav_vdevs[i];
5817 		ASSERT(vd != NULL);
5818 
5819 		if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
5820 		    pool != 0ULL && l2arc_vdev_present(vd))
5821 			l2arc_remove_vdev(vd);
5822 	}
5823 }
5824 
5825 /*
5826  * Verify encryption parameters for spa creation. If we are encrypting, we must
5827  * have the encryption feature flag enabled.
5828  */
5829 static int
5830 spa_create_check_encryption_params(dsl_crypto_params_t *dcp,
5831     boolean_t has_encryption)
5832 {
5833 	if (dcp->cp_crypt != ZIO_CRYPT_OFF &&
5834 	    dcp->cp_crypt != ZIO_CRYPT_INHERIT &&
5835 	    !has_encryption)
5836 		return (SET_ERROR(ENOTSUP));
5837 
5838 	return (dmu_objset_create_crypt_check(NULL, dcp, NULL));
5839 }
5840 
5841 /*
5842  * Pool Creation
5843  */
5844 int
5845 spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
5846     nvlist_t *zplprops, dsl_crypto_params_t *dcp)
5847 {
5848 	spa_t *spa;
5849 	const char *altroot = NULL;
5850 	vdev_t *rvd;
5851 	dsl_pool_t *dp;
5852 	dmu_tx_t *tx;
5853 	int error = 0;
5854 	uint64_t txg = TXG_INITIAL;
5855 	nvlist_t **spares, **l2cache;
5856 	uint_t nspares, nl2cache;
5857 	uint64_t version, obj, ndraid = 0;
5858 	boolean_t has_features;
5859 	boolean_t has_encryption;
5860 	boolean_t has_allocclass;
5861 	spa_feature_t feat;
5862 	const char *feat_name;
5863 	const char *poolname;
5864 	nvlist_t *nvl;
5865 
5866 	if (props == NULL ||
5867 	    nvlist_lookup_string(props, "tname", &poolname) != 0)
5868 		poolname = (char *)pool;
5869 
5870 	/*
5871 	 * If this pool already exists, return failure.
5872 	 */
5873 	mutex_enter(&spa_namespace_lock);
5874 	if (spa_lookup(poolname) != NULL) {
5875 		mutex_exit(&spa_namespace_lock);
5876 		return (SET_ERROR(EEXIST));
5877 	}
5878 
5879 	/*
5880 	 * Allocate a new spa_t structure.
5881 	 */
5882 	nvl = fnvlist_alloc();
5883 	fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool);
5884 	(void) nvlist_lookup_string(props,
5885 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
5886 	spa = spa_add(poolname, nvl, altroot);
5887 	fnvlist_free(nvl);
5888 	spa_activate(spa, spa_mode_global);
5889 
5890 	if (props && (error = spa_prop_validate(spa, props))) {
5891 		spa_deactivate(spa);
5892 		spa_remove(spa);
5893 		mutex_exit(&spa_namespace_lock);
5894 		return (error);
5895 	}
5896 
5897 	/*
5898 	 * Temporary pool names should never be written to disk.
5899 	 */
5900 	if (poolname != pool)
5901 		spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME;
5902 
5903 	has_features = B_FALSE;
5904 	has_encryption = B_FALSE;
5905 	has_allocclass = B_FALSE;
5906 	for (nvpair_t *elem = nvlist_next_nvpair(props, NULL);
5907 	    elem != NULL; elem = nvlist_next_nvpair(props, elem)) {
5908 		if (zpool_prop_feature(nvpair_name(elem))) {
5909 			has_features = B_TRUE;
5910 
5911 			feat_name = strchr(nvpair_name(elem), '@') + 1;
5912 			VERIFY0(zfeature_lookup_name(feat_name, &feat));
5913 			if (feat == SPA_FEATURE_ENCRYPTION)
5914 				has_encryption = B_TRUE;
5915 			if (feat == SPA_FEATURE_ALLOCATION_CLASSES)
5916 				has_allocclass = B_TRUE;
5917 		}
5918 	}
5919 
5920 	/* verify encryption params, if they were provided */
5921 	if (dcp != NULL) {
5922 		error = spa_create_check_encryption_params(dcp, has_encryption);
5923 		if (error != 0) {
5924 			spa_deactivate(spa);
5925 			spa_remove(spa);
5926 			mutex_exit(&spa_namespace_lock);
5927 			return (error);
5928 		}
5929 	}
5930 	if (!has_allocclass && zfs_special_devs(nvroot, NULL)) {
5931 		spa_deactivate(spa);
5932 		spa_remove(spa);
5933 		mutex_exit(&spa_namespace_lock);
5934 		return (ENOTSUP);
5935 	}
5936 
5937 	if (has_features || nvlist_lookup_uint64(props,
5938 	    zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) {
5939 		version = SPA_VERSION;
5940 	}
5941 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
5942 
5943 	spa->spa_first_txg = txg;
5944 	spa->spa_uberblock.ub_txg = txg - 1;
5945 	spa->spa_uberblock.ub_version = version;
5946 	spa->spa_ubsync = spa->spa_uberblock;
5947 	spa->spa_load_state = SPA_LOAD_CREATE;
5948 	spa->spa_removing_phys.sr_state = DSS_NONE;
5949 	spa->spa_removing_phys.sr_removing_vdev = -1;
5950 	spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
5951 	spa->spa_indirect_vdevs_loaded = B_TRUE;
5952 
5953 	/*
5954 	 * Create "The Godfather" zio to hold all async IOs
5955 	 */
5956 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
5957 	    KM_SLEEP);
5958 	for (int i = 0; i < max_ncpus; i++) {
5959 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
5960 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
5961 		    ZIO_FLAG_GODFATHER);
5962 	}
5963 
5964 	/*
5965 	 * Create the root vdev.
5966 	 */
5967 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5968 
5969 	error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD);
5970 
5971 	ASSERT(error != 0 || rvd != NULL);
5972 	ASSERT(error != 0 || spa->spa_root_vdev == rvd);
5973 
5974 	if (error == 0 && !zfs_allocatable_devs(nvroot))
5975 		error = SET_ERROR(EINVAL);
5976 
5977 	if (error == 0 &&
5978 	    (error = vdev_create(rvd, txg, B_FALSE)) == 0 &&
5979 	    (error = vdev_draid_spare_create(nvroot, rvd, &ndraid, 0)) == 0 &&
5980 	    (error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) == 0) {
5981 		/*
5982 		 * instantiate the metaslab groups (this will dirty the vdevs)
5983 		 * we can no longer error exit past this point
5984 		 */
5985 		for (int c = 0; error == 0 && c < rvd->vdev_children; c++) {
5986 			vdev_t *vd = rvd->vdev_child[c];
5987 
5988 			vdev_metaslab_set_size(vd);
5989 			vdev_expand(vd, txg);
5990 		}
5991 	}
5992 
5993 	spa_config_exit(spa, SCL_ALL, FTAG);
5994 
5995 	if (error != 0) {
5996 		spa_unload(spa);
5997 		spa_deactivate(spa);
5998 		spa_remove(spa);
5999 		mutex_exit(&spa_namespace_lock);
6000 		return (error);
6001 	}
6002 
6003 	/*
6004 	 * Get the list of spares, if specified.
6005 	 */
6006 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
6007 	    &spares, &nspares) == 0) {
6008 		spa->spa_spares.sav_config = fnvlist_alloc();
6009 		fnvlist_add_nvlist_array(spa->spa_spares.sav_config,
6010 		    ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares,
6011 		    nspares);
6012 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6013 		spa_load_spares(spa);
6014 		spa_config_exit(spa, SCL_ALL, FTAG);
6015 		spa->spa_spares.sav_sync = B_TRUE;
6016 	}
6017 
6018 	/*
6019 	 * Get the list of level 2 cache devices, if specified.
6020 	 */
6021 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
6022 	    &l2cache, &nl2cache) == 0) {
6023 		VERIFY0(nvlist_alloc(&spa->spa_l2cache.sav_config,
6024 		    NV_UNIQUE_NAME, KM_SLEEP));
6025 		fnvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
6026 		    ZPOOL_CONFIG_L2CACHE, (const nvlist_t * const *)l2cache,
6027 		    nl2cache);
6028 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6029 		spa_load_l2cache(spa);
6030 		spa_config_exit(spa, SCL_ALL, FTAG);
6031 		spa->spa_l2cache.sav_sync = B_TRUE;
6032 	}
6033 
6034 	spa->spa_is_initializing = B_TRUE;
6035 	spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg);
6036 	spa->spa_is_initializing = B_FALSE;
6037 
6038 	/*
6039 	 * Create DDTs (dedup tables).
6040 	 */
6041 	ddt_create(spa);
6042 	/*
6043 	 * Create BRT table and BRT table object.
6044 	 */
6045 	brt_create(spa);
6046 
6047 	spa_update_dspace(spa);
6048 
6049 	tx = dmu_tx_create_assigned(dp, txg);
6050 
6051 	/*
6052 	 * Create the pool's history object.
6053 	 */
6054 	if (version >= SPA_VERSION_ZPOOL_HISTORY && !spa->spa_history)
6055 		spa_history_create_obj(spa, tx);
6056 
6057 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE);
6058 	spa_history_log_version(spa, "create", tx);
6059 
6060 	/*
6061 	 * Create the pool config object.
6062 	 */
6063 	spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset,
6064 	    DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE,
6065 	    DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx);
6066 
6067 	if (zap_add(spa->spa_meta_objset,
6068 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG,
6069 	    sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) {
6070 		cmn_err(CE_PANIC, "failed to add pool config");
6071 	}
6072 
6073 	if (zap_add(spa->spa_meta_objset,
6074 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION,
6075 	    sizeof (uint64_t), 1, &version, tx) != 0) {
6076 		cmn_err(CE_PANIC, "failed to add pool version");
6077 	}
6078 
6079 	/* Newly created pools with the right version are always deflated. */
6080 	if (version >= SPA_VERSION_RAIDZ_DEFLATE) {
6081 		spa->spa_deflate = TRUE;
6082 		if (zap_add(spa->spa_meta_objset,
6083 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
6084 		    sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) {
6085 			cmn_err(CE_PANIC, "failed to add deflate");
6086 		}
6087 	}
6088 
6089 	/*
6090 	 * Create the deferred-free bpobj.  Turn off compression
6091 	 * because sync-to-convergence takes longer if the blocksize
6092 	 * keeps changing.
6093 	 */
6094 	obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx);
6095 	dmu_object_set_compress(spa->spa_meta_objset, obj,
6096 	    ZIO_COMPRESS_OFF, tx);
6097 	if (zap_add(spa->spa_meta_objset,
6098 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ,
6099 	    sizeof (uint64_t), 1, &obj, tx) != 0) {
6100 		cmn_err(CE_PANIC, "failed to add bpobj");
6101 	}
6102 	VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj,
6103 	    spa->spa_meta_objset, obj));
6104 
6105 	/*
6106 	 * Generate some random noise for salted checksums to operate on.
6107 	 */
6108 	(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
6109 	    sizeof (spa->spa_cksum_salt.zcs_bytes));
6110 
6111 	/*
6112 	 * Set pool properties.
6113 	 */
6114 	spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS);
6115 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
6116 	spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE);
6117 	spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND);
6118 	spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST);
6119 	spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM);
6120 
6121 	if (props != NULL) {
6122 		spa_configfile_set(spa, props, B_FALSE);
6123 		spa_sync_props(props, tx);
6124 	}
6125 
6126 	for (int i = 0; i < ndraid; i++)
6127 		spa_feature_incr(spa, SPA_FEATURE_DRAID, tx);
6128 
6129 	dmu_tx_commit(tx);
6130 
6131 	spa->spa_sync_on = B_TRUE;
6132 	txg_sync_start(dp);
6133 	mmp_thread_start(spa);
6134 	txg_wait_synced(dp, txg);
6135 
6136 	spa_spawn_aux_threads(spa);
6137 
6138 	spa_write_cachefile(spa, B_FALSE, B_TRUE, B_TRUE);
6139 
6140 	/*
6141 	 * Don't count references from objsets that are already closed
6142 	 * and are making their way through the eviction process.
6143 	 */
6144 	spa_evicting_os_wait(spa);
6145 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
6146 	spa->spa_load_state = SPA_LOAD_NONE;
6147 
6148 	spa_import_os(spa);
6149 
6150 	mutex_exit(&spa_namespace_lock);
6151 
6152 	return (0);
6153 }
6154 
6155 /*
6156  * Import a non-root pool into the system.
6157  */
6158 int
6159 spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags)
6160 {
6161 	spa_t *spa;
6162 	const char *altroot = NULL;
6163 	spa_load_state_t state = SPA_LOAD_IMPORT;
6164 	zpool_load_policy_t policy;
6165 	spa_mode_t mode = spa_mode_global;
6166 	uint64_t readonly = B_FALSE;
6167 	int error;
6168 	nvlist_t *nvroot;
6169 	nvlist_t **spares, **l2cache;
6170 	uint_t nspares, nl2cache;
6171 
6172 	/*
6173 	 * If a pool with this name exists, return failure.
6174 	 */
6175 	mutex_enter(&spa_namespace_lock);
6176 	if (spa_lookup(pool) != NULL) {
6177 		mutex_exit(&spa_namespace_lock);
6178 		return (SET_ERROR(EEXIST));
6179 	}
6180 
6181 	/*
6182 	 * Create and initialize the spa structure.
6183 	 */
6184 	(void) nvlist_lookup_string(props,
6185 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
6186 	(void) nvlist_lookup_uint64(props,
6187 	    zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly);
6188 	if (readonly)
6189 		mode = SPA_MODE_READ;
6190 	spa = spa_add(pool, config, altroot);
6191 	spa->spa_import_flags = flags;
6192 
6193 	/*
6194 	 * Verbatim import - Take a pool and insert it into the namespace
6195 	 * as if it had been loaded at boot.
6196 	 */
6197 	if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) {
6198 		if (props != NULL)
6199 			spa_configfile_set(spa, props, B_FALSE);
6200 
6201 		spa_write_cachefile(spa, B_FALSE, B_TRUE, B_FALSE);
6202 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
6203 		zfs_dbgmsg("spa_import: verbatim import of %s", pool);
6204 		mutex_exit(&spa_namespace_lock);
6205 		return (0);
6206 	}
6207 
6208 	spa_activate(spa, mode);
6209 
6210 	/*
6211 	 * Don't start async tasks until we know everything is healthy.
6212 	 */
6213 	spa_async_suspend(spa);
6214 
6215 	zpool_get_load_policy(config, &policy);
6216 	if (policy.zlp_rewind & ZPOOL_DO_REWIND)
6217 		state = SPA_LOAD_RECOVER;
6218 
6219 	spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT;
6220 
6221 	if (state != SPA_LOAD_RECOVER) {
6222 		spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
6223 		zfs_dbgmsg("spa_import: importing %s", pool);
6224 	} else {
6225 		zfs_dbgmsg("spa_import: importing %s, max_txg=%lld "
6226 		    "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg);
6227 	}
6228 	error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind);
6229 
6230 	/*
6231 	 * Propagate anything learned while loading the pool and pass it
6232 	 * back to caller (i.e. rewind info, missing devices, etc).
6233 	 */
6234 	fnvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info);
6235 
6236 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6237 	/*
6238 	 * Toss any existing sparelist, as it doesn't have any validity
6239 	 * anymore, and conflicts with spa_has_spare().
6240 	 */
6241 	if (spa->spa_spares.sav_config) {
6242 		nvlist_free(spa->spa_spares.sav_config);
6243 		spa->spa_spares.sav_config = NULL;
6244 		spa_load_spares(spa);
6245 	}
6246 	if (spa->spa_l2cache.sav_config) {
6247 		nvlist_free(spa->spa_l2cache.sav_config);
6248 		spa->spa_l2cache.sav_config = NULL;
6249 		spa_load_l2cache(spa);
6250 	}
6251 
6252 	nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE);
6253 	spa_config_exit(spa, SCL_ALL, FTAG);
6254 
6255 	if (props != NULL)
6256 		spa_configfile_set(spa, props, B_FALSE);
6257 
6258 	if (error != 0 || (props && spa_writeable(spa) &&
6259 	    (error = spa_prop_set(spa, props)))) {
6260 		spa_unload(spa);
6261 		spa_deactivate(spa);
6262 		spa_remove(spa);
6263 		mutex_exit(&spa_namespace_lock);
6264 		return (error);
6265 	}
6266 
6267 	spa_async_resume(spa);
6268 
6269 	/*
6270 	 * Override any spares and level 2 cache devices as specified by
6271 	 * the user, as these may have correct device names/devids, etc.
6272 	 */
6273 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
6274 	    &spares, &nspares) == 0) {
6275 		if (spa->spa_spares.sav_config)
6276 			fnvlist_remove(spa->spa_spares.sav_config,
6277 			    ZPOOL_CONFIG_SPARES);
6278 		else
6279 			spa->spa_spares.sav_config = fnvlist_alloc();
6280 		fnvlist_add_nvlist_array(spa->spa_spares.sav_config,
6281 		    ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares,
6282 		    nspares);
6283 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6284 		spa_load_spares(spa);
6285 		spa_config_exit(spa, SCL_ALL, FTAG);
6286 		spa->spa_spares.sav_sync = B_TRUE;
6287 	}
6288 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
6289 	    &l2cache, &nl2cache) == 0) {
6290 		if (spa->spa_l2cache.sav_config)
6291 			fnvlist_remove(spa->spa_l2cache.sav_config,
6292 			    ZPOOL_CONFIG_L2CACHE);
6293 		else
6294 			spa->spa_l2cache.sav_config = fnvlist_alloc();
6295 		fnvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
6296 		    ZPOOL_CONFIG_L2CACHE, (const nvlist_t * const *)l2cache,
6297 		    nl2cache);
6298 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6299 		spa_load_l2cache(spa);
6300 		spa_config_exit(spa, SCL_ALL, FTAG);
6301 		spa->spa_l2cache.sav_sync = B_TRUE;
6302 	}
6303 
6304 	/*
6305 	 * Check for any removed devices.
6306 	 */
6307 	if (spa->spa_autoreplace) {
6308 		spa_aux_check_removed(&spa->spa_spares);
6309 		spa_aux_check_removed(&spa->spa_l2cache);
6310 	}
6311 
6312 	if (spa_writeable(spa)) {
6313 		/*
6314 		 * Update the config cache to include the newly-imported pool.
6315 		 */
6316 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
6317 	}
6318 
6319 	/*
6320 	 * It's possible that the pool was expanded while it was exported.
6321 	 * We kick off an async task to handle this for us.
6322 	 */
6323 	spa_async_request(spa, SPA_ASYNC_AUTOEXPAND);
6324 
6325 	spa_history_log_version(spa, "import", NULL);
6326 
6327 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
6328 
6329 	mutex_exit(&spa_namespace_lock);
6330 
6331 	zvol_create_minors_recursive(pool);
6332 
6333 	spa_import_os(spa);
6334 
6335 	return (0);
6336 }
6337 
6338 nvlist_t *
6339 spa_tryimport(nvlist_t *tryconfig)
6340 {
6341 	nvlist_t *config = NULL;
6342 	const char *poolname, *cachefile;
6343 	spa_t *spa;
6344 	uint64_t state;
6345 	int error;
6346 	zpool_load_policy_t policy;
6347 
6348 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname))
6349 		return (NULL);
6350 
6351 	if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state))
6352 		return (NULL);
6353 
6354 	/*
6355 	 * Create and initialize the spa structure.
6356 	 */
6357 	mutex_enter(&spa_namespace_lock);
6358 	spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL);
6359 	spa_activate(spa, SPA_MODE_READ);
6360 
6361 	/*
6362 	 * Rewind pool if a max txg was provided.
6363 	 */
6364 	zpool_get_load_policy(spa->spa_config, &policy);
6365 	if (policy.zlp_txg != UINT64_MAX) {
6366 		spa->spa_load_max_txg = policy.zlp_txg;
6367 		spa->spa_extreme_rewind = B_TRUE;
6368 		zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld",
6369 		    poolname, (longlong_t)policy.zlp_txg);
6370 	} else {
6371 		zfs_dbgmsg("spa_tryimport: importing %s", poolname);
6372 	}
6373 
6374 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile)
6375 	    == 0) {
6376 		zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile);
6377 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
6378 	} else {
6379 		spa->spa_config_source = SPA_CONFIG_SRC_SCAN;
6380 	}
6381 
6382 	/*
6383 	 * spa_import() relies on a pool config fetched by spa_try_import()
6384 	 * for spare/cache devices. Import flags are not passed to
6385 	 * spa_tryimport(), which makes it return early due to a missing log
6386 	 * device and missing retrieving the cache device and spare eventually.
6387 	 * Passing ZFS_IMPORT_MISSING_LOG to spa_tryimport() makes it fetch
6388 	 * the correct configuration regardless of the missing log device.
6389 	 */
6390 	spa->spa_import_flags |= ZFS_IMPORT_MISSING_LOG;
6391 
6392 	error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING);
6393 
6394 	/*
6395 	 * If 'tryconfig' was at least parsable, return the current config.
6396 	 */
6397 	if (spa->spa_root_vdev != NULL) {
6398 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
6399 		fnvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, poolname);
6400 		fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, state);
6401 		fnvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
6402 		    spa->spa_uberblock.ub_timestamp);
6403 		fnvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO,
6404 		    spa->spa_load_info);
6405 		fnvlist_add_uint64(config, ZPOOL_CONFIG_ERRATA,
6406 		    spa->spa_errata);
6407 
6408 		/*
6409 		 * If the bootfs property exists on this pool then we
6410 		 * copy it out so that external consumers can tell which
6411 		 * pools are bootable.
6412 		 */
6413 		if ((!error || error == EEXIST) && spa->spa_bootfs) {
6414 			char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
6415 
6416 			/*
6417 			 * We have to play games with the name since the
6418 			 * pool was opened as TRYIMPORT_NAME.
6419 			 */
6420 			if (dsl_dsobj_to_dsname(spa_name(spa),
6421 			    spa->spa_bootfs, tmpname) == 0) {
6422 				char *cp;
6423 				char *dsname;
6424 
6425 				dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
6426 
6427 				cp = strchr(tmpname, '/');
6428 				if (cp == NULL) {
6429 					(void) strlcpy(dsname, tmpname,
6430 					    MAXPATHLEN);
6431 				} else {
6432 					(void) snprintf(dsname, MAXPATHLEN,
6433 					    "%s/%s", poolname, ++cp);
6434 				}
6435 				fnvlist_add_string(config, ZPOOL_CONFIG_BOOTFS,
6436 				    dsname);
6437 				kmem_free(dsname, MAXPATHLEN);
6438 			}
6439 			kmem_free(tmpname, MAXPATHLEN);
6440 		}
6441 
6442 		/*
6443 		 * Add the list of hot spares and level 2 cache devices.
6444 		 */
6445 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
6446 		spa_add_spares(spa, config);
6447 		spa_add_l2cache(spa, config);
6448 		spa_config_exit(spa, SCL_CONFIG, FTAG);
6449 	}
6450 
6451 	spa_unload(spa);
6452 	spa_deactivate(spa);
6453 	spa_remove(spa);
6454 	mutex_exit(&spa_namespace_lock);
6455 
6456 	return (config);
6457 }
6458 
6459 /*
6460  * Pool export/destroy
6461  *
6462  * The act of destroying or exporting a pool is very simple.  We make sure there
6463  * is no more pending I/O and any references to the pool are gone.  Then, we
6464  * update the pool state and sync all the labels to disk, removing the
6465  * configuration from the cache afterwards. If the 'hardforce' flag is set, then
6466  * we don't sync the labels or remove the configuration cache.
6467  */
6468 static int
6469 spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
6470     boolean_t force, boolean_t hardforce)
6471 {
6472 	int error;
6473 	spa_t *spa;
6474 
6475 	if (oldconfig)
6476 		*oldconfig = NULL;
6477 
6478 	if (!(spa_mode_global & SPA_MODE_WRITE))
6479 		return (SET_ERROR(EROFS));
6480 
6481 	mutex_enter(&spa_namespace_lock);
6482 	if ((spa = spa_lookup(pool)) == NULL) {
6483 		mutex_exit(&spa_namespace_lock);
6484 		return (SET_ERROR(ENOENT));
6485 	}
6486 
6487 	if (spa->spa_is_exporting) {
6488 		/* the pool is being exported by another thread */
6489 		mutex_exit(&spa_namespace_lock);
6490 		return (SET_ERROR(ZFS_ERR_EXPORT_IN_PROGRESS));
6491 	}
6492 	spa->spa_is_exporting = B_TRUE;
6493 
6494 	/*
6495 	 * Put a hold on the pool, drop the namespace lock, stop async tasks,
6496 	 * reacquire the namespace lock, and see if we can export.
6497 	 */
6498 	spa_open_ref(spa, FTAG);
6499 	mutex_exit(&spa_namespace_lock);
6500 	spa_async_suspend(spa);
6501 	if (spa->spa_zvol_taskq) {
6502 		zvol_remove_minors(spa, spa_name(spa), B_TRUE);
6503 		taskq_wait(spa->spa_zvol_taskq);
6504 	}
6505 	mutex_enter(&spa_namespace_lock);
6506 	spa_close(spa, FTAG);
6507 
6508 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
6509 		goto export_spa;
6510 	/*
6511 	 * The pool will be in core if it's openable, in which case we can
6512 	 * modify its state.  Objsets may be open only because they're dirty,
6513 	 * so we have to force it to sync before checking spa_refcnt.
6514 	 */
6515 	if (spa->spa_sync_on) {
6516 		txg_wait_synced(spa->spa_dsl_pool, 0);
6517 		spa_evicting_os_wait(spa);
6518 	}
6519 
6520 	/*
6521 	 * A pool cannot be exported or destroyed if there are active
6522 	 * references.  If we are resetting a pool, allow references by
6523 	 * fault injection handlers.
6524 	 */
6525 	if (!spa_refcount_zero(spa) || (spa->spa_inject_ref != 0)) {
6526 		error = SET_ERROR(EBUSY);
6527 		goto fail;
6528 	}
6529 
6530 	if (spa->spa_sync_on) {
6531 		vdev_t *rvd = spa->spa_root_vdev;
6532 		/*
6533 		 * A pool cannot be exported if it has an active shared spare.
6534 		 * This is to prevent other pools stealing the active spare
6535 		 * from an exported pool. At user's own will, such pool can
6536 		 * be forcedly exported.
6537 		 */
6538 		if (!force && new_state == POOL_STATE_EXPORTED &&
6539 		    spa_has_active_shared_spare(spa)) {
6540 			error = SET_ERROR(EXDEV);
6541 			goto fail;
6542 		}
6543 
6544 		/*
6545 		 * We're about to export or destroy this pool. Make sure
6546 		 * we stop all initialization and trim activity here before
6547 		 * we set the spa_final_txg. This will ensure that all
6548 		 * dirty data resulting from the initialization is
6549 		 * committed to disk before we unload the pool.
6550 		 */
6551 		vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE);
6552 		vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE);
6553 		vdev_autotrim_stop_all(spa);
6554 		vdev_rebuild_stop_all(spa);
6555 
6556 		/*
6557 		 * We want this to be reflected on every label,
6558 		 * so mark them all dirty.  spa_unload() will do the
6559 		 * final sync that pushes these changes out.
6560 		 */
6561 		if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) {
6562 			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6563 			spa->spa_state = new_state;
6564 			vdev_config_dirty(rvd);
6565 			spa_config_exit(spa, SCL_ALL, FTAG);
6566 		}
6567 
6568 		/*
6569 		 * If the log space map feature is enabled and the pool is
6570 		 * getting exported (but not destroyed), we want to spend some
6571 		 * time flushing as many metaslabs as we can in an attempt to
6572 		 * destroy log space maps and save import time. This has to be
6573 		 * done before we set the spa_final_txg, otherwise
6574 		 * spa_sync() -> spa_flush_metaslabs() may dirty the final TXGs.
6575 		 * spa_should_flush_logs_on_unload() should be called after
6576 		 * spa_state has been set to the new_state.
6577 		 */
6578 		if (spa_should_flush_logs_on_unload(spa))
6579 			spa_unload_log_sm_flush_all(spa);
6580 
6581 		if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) {
6582 			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6583 			spa->spa_final_txg = spa_last_synced_txg(spa) +
6584 			    TXG_DEFER_SIZE + 1;
6585 			spa_config_exit(spa, SCL_ALL, FTAG);
6586 		}
6587 	}
6588 
6589 export_spa:
6590 	spa_export_os(spa);
6591 
6592 	if (new_state == POOL_STATE_DESTROYED)
6593 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY);
6594 	else if (new_state == POOL_STATE_EXPORTED)
6595 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_EXPORT);
6596 
6597 	if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
6598 		spa_unload(spa);
6599 		spa_deactivate(spa);
6600 	}
6601 
6602 	if (oldconfig && spa->spa_config)
6603 		*oldconfig = fnvlist_dup(spa->spa_config);
6604 
6605 	if (new_state != POOL_STATE_UNINITIALIZED) {
6606 		if (!hardforce)
6607 			spa_write_cachefile(spa, B_TRUE, B_TRUE, B_FALSE);
6608 		spa_remove(spa);
6609 	} else {
6610 		/*
6611 		 * If spa_remove() is not called for this spa_t and
6612 		 * there is any possibility that it can be reused,
6613 		 * we make sure to reset the exporting flag.
6614 		 */
6615 		spa->spa_is_exporting = B_FALSE;
6616 	}
6617 
6618 	mutex_exit(&spa_namespace_lock);
6619 	return (0);
6620 
6621 fail:
6622 	spa->spa_is_exporting = B_FALSE;
6623 	spa_async_resume(spa);
6624 	mutex_exit(&spa_namespace_lock);
6625 	return (error);
6626 }
6627 
6628 /*
6629  * Destroy a storage pool.
6630  */
6631 int
6632 spa_destroy(const char *pool)
6633 {
6634 	return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL,
6635 	    B_FALSE, B_FALSE));
6636 }
6637 
6638 /*
6639  * Export a storage pool.
6640  */
6641 int
6642 spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
6643     boolean_t hardforce)
6644 {
6645 	return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig,
6646 	    force, hardforce));
6647 }
6648 
6649 /*
6650  * Similar to spa_export(), this unloads the spa_t without actually removing it
6651  * from the namespace in any way.
6652  */
6653 int
6654 spa_reset(const char *pool)
6655 {
6656 	return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL,
6657 	    B_FALSE, B_FALSE));
6658 }
6659 
6660 /*
6661  * ==========================================================================
6662  * Device manipulation
6663  * ==========================================================================
6664  */
6665 
6666 /*
6667  * This is called as a synctask to increment the draid feature flag
6668  */
6669 static void
6670 spa_draid_feature_incr(void *arg, dmu_tx_t *tx)
6671 {
6672 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
6673 	int draid = (int)(uintptr_t)arg;
6674 
6675 	for (int c = 0; c < draid; c++)
6676 		spa_feature_incr(spa, SPA_FEATURE_DRAID, tx);
6677 }
6678 
6679 /*
6680  * Add a device to a storage pool.
6681  */
6682 int
6683 spa_vdev_add(spa_t *spa, nvlist_t *nvroot)
6684 {
6685 	uint64_t txg, ndraid = 0;
6686 	int error;
6687 	vdev_t *rvd = spa->spa_root_vdev;
6688 	vdev_t *vd, *tvd;
6689 	nvlist_t **spares, **l2cache;
6690 	uint_t nspares, nl2cache;
6691 
6692 	ASSERT(spa_writeable(spa));
6693 
6694 	txg = spa_vdev_enter(spa);
6695 
6696 	if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0,
6697 	    VDEV_ALLOC_ADD)) != 0)
6698 		return (spa_vdev_exit(spa, NULL, txg, error));
6699 
6700 	spa->spa_pending_vdev = vd;	/* spa_vdev_exit() will clear this */
6701 
6702 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares,
6703 	    &nspares) != 0)
6704 		nspares = 0;
6705 
6706 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache,
6707 	    &nl2cache) != 0)
6708 		nl2cache = 0;
6709 
6710 	if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0)
6711 		return (spa_vdev_exit(spa, vd, txg, EINVAL));
6712 
6713 	if (vd->vdev_children != 0 &&
6714 	    (error = vdev_create(vd, txg, B_FALSE)) != 0) {
6715 		return (spa_vdev_exit(spa, vd, txg, error));
6716 	}
6717 
6718 	/*
6719 	 * The virtual dRAID spares must be added after vdev tree is created
6720 	 * and the vdev guids are generated.  The guid of their associated
6721 	 * dRAID is stored in the config and used when opening the spare.
6722 	 */
6723 	if ((error = vdev_draid_spare_create(nvroot, vd, &ndraid,
6724 	    rvd->vdev_children)) == 0) {
6725 		if (ndraid > 0 && nvlist_lookup_nvlist_array(nvroot,
6726 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0)
6727 			nspares = 0;
6728 	} else {
6729 		return (spa_vdev_exit(spa, vd, txg, error));
6730 	}
6731 
6732 	/*
6733 	 * We must validate the spares and l2cache devices after checking the
6734 	 * children.  Otherwise, vdev_inuse() will blindly overwrite the spare.
6735 	 */
6736 	if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0)
6737 		return (spa_vdev_exit(spa, vd, txg, error));
6738 
6739 	/*
6740 	 * If we are in the middle of a device removal, we can only add
6741 	 * devices which match the existing devices in the pool.
6742 	 * If we are in the middle of a removal, or have some indirect
6743 	 * vdevs, we can not add raidz or dRAID top levels.
6744 	 */
6745 	if (spa->spa_vdev_removal != NULL ||
6746 	    spa->spa_removing_phys.sr_prev_indirect_vdev != -1) {
6747 		for (int c = 0; c < vd->vdev_children; c++) {
6748 			tvd = vd->vdev_child[c];
6749 			if (spa->spa_vdev_removal != NULL &&
6750 			    tvd->vdev_ashift != spa->spa_max_ashift) {
6751 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
6752 			}
6753 			/* Fail if top level vdev is raidz or a dRAID */
6754 			if (vdev_get_nparity(tvd) != 0)
6755 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
6756 
6757 			/*
6758 			 * Need the top level mirror to be
6759 			 * a mirror of leaf vdevs only
6760 			 */
6761 			if (tvd->vdev_ops == &vdev_mirror_ops) {
6762 				for (uint64_t cid = 0;
6763 				    cid < tvd->vdev_children; cid++) {
6764 					vdev_t *cvd = tvd->vdev_child[cid];
6765 					if (!cvd->vdev_ops->vdev_op_leaf) {
6766 						return (spa_vdev_exit(spa, vd,
6767 						    txg, EINVAL));
6768 					}
6769 				}
6770 			}
6771 		}
6772 	}
6773 
6774 	for (int c = 0; c < vd->vdev_children; c++) {
6775 		tvd = vd->vdev_child[c];
6776 		vdev_remove_child(vd, tvd);
6777 		tvd->vdev_id = rvd->vdev_children;
6778 		vdev_add_child(rvd, tvd);
6779 		vdev_config_dirty(tvd);
6780 	}
6781 
6782 	if (nspares != 0) {
6783 		spa_set_aux_vdevs(&spa->spa_spares, spares, nspares,
6784 		    ZPOOL_CONFIG_SPARES);
6785 		spa_load_spares(spa);
6786 		spa->spa_spares.sav_sync = B_TRUE;
6787 	}
6788 
6789 	if (nl2cache != 0) {
6790 		spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache,
6791 		    ZPOOL_CONFIG_L2CACHE);
6792 		spa_load_l2cache(spa);
6793 		spa->spa_l2cache.sav_sync = B_TRUE;
6794 	}
6795 
6796 	/*
6797 	 * We can't increment a feature while holding spa_vdev so we
6798 	 * have to do it in a synctask.
6799 	 */
6800 	if (ndraid != 0) {
6801 		dmu_tx_t *tx;
6802 
6803 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
6804 		dsl_sync_task_nowait(spa->spa_dsl_pool, spa_draid_feature_incr,
6805 		    (void *)(uintptr_t)ndraid, tx);
6806 		dmu_tx_commit(tx);
6807 	}
6808 
6809 	/*
6810 	 * We have to be careful when adding new vdevs to an existing pool.
6811 	 * If other threads start allocating from these vdevs before we
6812 	 * sync the config cache, and we lose power, then upon reboot we may
6813 	 * fail to open the pool because there are DVAs that the config cache
6814 	 * can't translate.  Therefore, we first add the vdevs without
6815 	 * initializing metaslabs; sync the config cache (via spa_vdev_exit());
6816 	 * and then let spa_config_update() initialize the new metaslabs.
6817 	 *
6818 	 * spa_load() checks for added-but-not-initialized vdevs, so that
6819 	 * if we lose power at any point in this sequence, the remaining
6820 	 * steps will be completed the next time we load the pool.
6821 	 */
6822 	(void) spa_vdev_exit(spa, vd, txg, 0);
6823 
6824 	mutex_enter(&spa_namespace_lock);
6825 	spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
6826 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD);
6827 	mutex_exit(&spa_namespace_lock);
6828 
6829 	return (0);
6830 }
6831 
6832 /*
6833  * Attach a device to a mirror.  The arguments are the path to any device
6834  * in the mirror, and the nvroot for the new device.  If the path specifies
6835  * a device that is not mirrored, we automatically insert the mirror vdev.
6836  *
6837  * If 'replacing' is specified, the new device is intended to replace the
6838  * existing device; in this case the two devices are made into their own
6839  * mirror using the 'replacing' vdev, which is functionally identical to
6840  * the mirror vdev (it actually reuses all the same ops) but has a few
6841  * extra rules: you can't attach to it after it's been created, and upon
6842  * completion of resilvering, the first disk (the one being replaced)
6843  * is automatically detached.
6844  *
6845  * If 'rebuild' is specified, then sequential reconstruction (a.ka. rebuild)
6846  * should be performed instead of traditional healing reconstruction.  From
6847  * an administrators perspective these are both resilver operations.
6848  */
6849 int
6850 spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing,
6851     int rebuild)
6852 {
6853 	uint64_t txg, dtl_max_txg;
6854 	vdev_t *rvd = spa->spa_root_vdev;
6855 	vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd;
6856 	vdev_ops_t *pvops;
6857 	char *oldvdpath, *newvdpath;
6858 	int newvd_isspare;
6859 	int error;
6860 
6861 	ASSERT(spa_writeable(spa));
6862 
6863 	txg = spa_vdev_enter(spa);
6864 
6865 	oldvd = spa_lookup_by_guid(spa, guid, B_FALSE);
6866 
6867 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
6868 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6869 		error = (spa_has_checkpoint(spa)) ?
6870 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6871 		return (spa_vdev_exit(spa, NULL, txg, error));
6872 	}
6873 
6874 	if (rebuild) {
6875 		if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REBUILD))
6876 			return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6877 
6878 		if (dsl_scan_resilvering(spa_get_dsl(spa)) ||
6879 		    dsl_scan_resilver_scheduled(spa_get_dsl(spa))) {
6880 			return (spa_vdev_exit(spa, NULL, txg,
6881 			    ZFS_ERR_RESILVER_IN_PROGRESS));
6882 		}
6883 	} else {
6884 		if (vdev_rebuild_active(rvd))
6885 			return (spa_vdev_exit(spa, NULL, txg,
6886 			    ZFS_ERR_REBUILD_IN_PROGRESS));
6887 	}
6888 
6889 	if (spa->spa_vdev_removal != NULL)
6890 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6891 
6892 	if (oldvd == NULL)
6893 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
6894 
6895 	if (!oldvd->vdev_ops->vdev_op_leaf)
6896 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6897 
6898 	pvd = oldvd->vdev_parent;
6899 
6900 	if (spa_config_parse(spa, &newrootvd, nvroot, NULL, 0,
6901 	    VDEV_ALLOC_ATTACH) != 0)
6902 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6903 
6904 	if (newrootvd->vdev_children != 1)
6905 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6906 
6907 	newvd = newrootvd->vdev_child[0];
6908 
6909 	if (!newvd->vdev_ops->vdev_op_leaf)
6910 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6911 
6912 	if ((error = vdev_create(newrootvd, txg, replacing)) != 0)
6913 		return (spa_vdev_exit(spa, newrootvd, txg, error));
6914 
6915 	/*
6916 	 * log, dedup and special vdevs should not be replaced by spares.
6917 	 */
6918 	if ((oldvd->vdev_top->vdev_alloc_bias != VDEV_BIAS_NONE ||
6919 	    oldvd->vdev_top->vdev_islog) && newvd->vdev_isspare) {
6920 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6921 	}
6922 
6923 	/*
6924 	 * A dRAID spare can only replace a child of its parent dRAID vdev.
6925 	 */
6926 	if (newvd->vdev_ops == &vdev_draid_spare_ops &&
6927 	    oldvd->vdev_top != vdev_draid_spare_get_parent(newvd)) {
6928 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6929 	}
6930 
6931 	if (rebuild) {
6932 		/*
6933 		 * For rebuilds, the top vdev must support reconstruction
6934 		 * using only space maps.  This means the only allowable
6935 		 * vdevs types are the root vdev, a mirror, or dRAID.
6936 		 */
6937 		tvd = pvd;
6938 		if (pvd->vdev_top != NULL)
6939 			tvd = pvd->vdev_top;
6940 
6941 		if (tvd->vdev_ops != &vdev_mirror_ops &&
6942 		    tvd->vdev_ops != &vdev_root_ops &&
6943 		    tvd->vdev_ops != &vdev_draid_ops) {
6944 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6945 		}
6946 	}
6947 
6948 	if (!replacing) {
6949 		/*
6950 		 * For attach, the only allowable parent is a mirror or the root
6951 		 * vdev.
6952 		 */
6953 		if (pvd->vdev_ops != &vdev_mirror_ops &&
6954 		    pvd->vdev_ops != &vdev_root_ops)
6955 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6956 
6957 		pvops = &vdev_mirror_ops;
6958 	} else {
6959 		/*
6960 		 * Active hot spares can only be replaced by inactive hot
6961 		 * spares.
6962 		 */
6963 		if (pvd->vdev_ops == &vdev_spare_ops &&
6964 		    oldvd->vdev_isspare &&
6965 		    !spa_has_spare(spa, newvd->vdev_guid))
6966 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6967 
6968 		/*
6969 		 * If the source is a hot spare, and the parent isn't already a
6970 		 * spare, then we want to create a new hot spare.  Otherwise, we
6971 		 * want to create a replacing vdev.  The user is not allowed to
6972 		 * attach to a spared vdev child unless the 'isspare' state is
6973 		 * the same (spare replaces spare, non-spare replaces
6974 		 * non-spare).
6975 		 */
6976 		if (pvd->vdev_ops == &vdev_replacing_ops &&
6977 		    spa_version(spa) < SPA_VERSION_MULTI_REPLACE) {
6978 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6979 		} else if (pvd->vdev_ops == &vdev_spare_ops &&
6980 		    newvd->vdev_isspare != oldvd->vdev_isspare) {
6981 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6982 		}
6983 
6984 		if (newvd->vdev_isspare)
6985 			pvops = &vdev_spare_ops;
6986 		else
6987 			pvops = &vdev_replacing_ops;
6988 	}
6989 
6990 	/*
6991 	 * Make sure the new device is big enough.
6992 	 */
6993 	if (newvd->vdev_asize < vdev_get_min_asize(oldvd))
6994 		return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW));
6995 
6996 	/*
6997 	 * The new device cannot have a higher alignment requirement
6998 	 * than the top-level vdev.
6999 	 */
7000 	if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift)
7001 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
7002 
7003 	/*
7004 	 * If this is an in-place replacement, update oldvd's path and devid
7005 	 * to make it distinguishable from newvd, and unopenable from now on.
7006 	 */
7007 	if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) {
7008 		spa_strfree(oldvd->vdev_path);
7009 		oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5,
7010 		    KM_SLEEP);
7011 		(void) snprintf(oldvd->vdev_path, strlen(newvd->vdev_path) + 5,
7012 		    "%s/%s", newvd->vdev_path, "old");
7013 		if (oldvd->vdev_devid != NULL) {
7014 			spa_strfree(oldvd->vdev_devid);
7015 			oldvd->vdev_devid = NULL;
7016 		}
7017 	}
7018 
7019 	/*
7020 	 * If the parent is not a mirror, or if we're replacing, insert the new
7021 	 * mirror/replacing/spare vdev above oldvd.
7022 	 */
7023 	if (pvd->vdev_ops != pvops)
7024 		pvd = vdev_add_parent(oldvd, pvops);
7025 
7026 	ASSERT(pvd->vdev_top->vdev_parent == rvd);
7027 	ASSERT(pvd->vdev_ops == pvops);
7028 	ASSERT(oldvd->vdev_parent == pvd);
7029 
7030 	/*
7031 	 * Extract the new device from its root and add it to pvd.
7032 	 */
7033 	vdev_remove_child(newrootvd, newvd);
7034 	newvd->vdev_id = pvd->vdev_children;
7035 	newvd->vdev_crtxg = oldvd->vdev_crtxg;
7036 	vdev_add_child(pvd, newvd);
7037 
7038 	/*
7039 	 * Reevaluate the parent vdev state.
7040 	 */
7041 	vdev_propagate_state(pvd);
7042 
7043 	tvd = newvd->vdev_top;
7044 	ASSERT(pvd->vdev_top == tvd);
7045 	ASSERT(tvd->vdev_parent == rvd);
7046 
7047 	vdev_config_dirty(tvd);
7048 
7049 	/*
7050 	 * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account
7051 	 * for any dmu_sync-ed blocks.  It will propagate upward when
7052 	 * spa_vdev_exit() calls vdev_dtl_reassess().
7053 	 */
7054 	dtl_max_txg = txg + TXG_CONCURRENT_STATES;
7055 
7056 	vdev_dtl_dirty(newvd, DTL_MISSING,
7057 	    TXG_INITIAL, dtl_max_txg - TXG_INITIAL);
7058 
7059 	if (newvd->vdev_isspare) {
7060 		spa_spare_activate(newvd);
7061 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE);
7062 	}
7063 
7064 	oldvdpath = spa_strdup(oldvd->vdev_path);
7065 	newvdpath = spa_strdup(newvd->vdev_path);
7066 	newvd_isspare = newvd->vdev_isspare;
7067 
7068 	/*
7069 	 * Mark newvd's DTL dirty in this txg.
7070 	 */
7071 	vdev_dirty(tvd, VDD_DTL, newvd, txg);
7072 
7073 	/*
7074 	 * Schedule the resilver or rebuild to restart in the future. We do
7075 	 * this to ensure that dmu_sync-ed blocks have been stitched into the
7076 	 * respective datasets.
7077 	 */
7078 	if (rebuild) {
7079 		newvd->vdev_rebuild_txg = txg;
7080 
7081 		vdev_rebuild(tvd);
7082 	} else {
7083 		newvd->vdev_resilver_txg = txg;
7084 
7085 		if (dsl_scan_resilvering(spa_get_dsl(spa)) &&
7086 		    spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) {
7087 			vdev_defer_resilver(newvd);
7088 		} else {
7089 			dsl_scan_restart_resilver(spa->spa_dsl_pool,
7090 			    dtl_max_txg);
7091 		}
7092 	}
7093 
7094 	if (spa->spa_bootfs)
7095 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH);
7096 
7097 	spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH);
7098 
7099 	/*
7100 	 * Commit the config
7101 	 */
7102 	(void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0);
7103 
7104 	spa_history_log_internal(spa, "vdev attach", NULL,
7105 	    "%s vdev=%s %s vdev=%s",
7106 	    replacing && newvd_isspare ? "spare in" :
7107 	    replacing ? "replace" : "attach", newvdpath,
7108 	    replacing ? "for" : "to", oldvdpath);
7109 
7110 	spa_strfree(oldvdpath);
7111 	spa_strfree(newvdpath);
7112 
7113 	return (0);
7114 }
7115 
7116 /*
7117  * Detach a device from a mirror or replacing vdev.
7118  *
7119  * If 'replace_done' is specified, only detach if the parent
7120  * is a replacing or a spare vdev.
7121  */
7122 int
7123 spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done)
7124 {
7125 	uint64_t txg;
7126 	int error;
7127 	vdev_t *rvd __maybe_unused = spa->spa_root_vdev;
7128 	vdev_t *vd, *pvd, *cvd, *tvd;
7129 	boolean_t unspare = B_FALSE;
7130 	uint64_t unspare_guid = 0;
7131 	char *vdpath;
7132 
7133 	ASSERT(spa_writeable(spa));
7134 
7135 	txg = spa_vdev_detach_enter(spa, guid);
7136 
7137 	vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7138 
7139 	/*
7140 	 * Besides being called directly from the userland through the
7141 	 * ioctl interface, spa_vdev_detach() can be potentially called
7142 	 * at the end of spa_vdev_resilver_done().
7143 	 *
7144 	 * In the regular case, when we have a checkpoint this shouldn't
7145 	 * happen as we never empty the DTLs of a vdev during the scrub
7146 	 * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done()
7147 	 * should never get here when we have a checkpoint.
7148 	 *
7149 	 * That said, even in a case when we checkpoint the pool exactly
7150 	 * as spa_vdev_resilver_done() calls this function everything
7151 	 * should be fine as the resilver will return right away.
7152 	 */
7153 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7154 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
7155 		error = (spa_has_checkpoint(spa)) ?
7156 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
7157 		return (spa_vdev_exit(spa, NULL, txg, error));
7158 	}
7159 
7160 	if (vd == NULL)
7161 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
7162 
7163 	if (!vd->vdev_ops->vdev_op_leaf)
7164 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
7165 
7166 	pvd = vd->vdev_parent;
7167 
7168 	/*
7169 	 * If the parent/child relationship is not as expected, don't do it.
7170 	 * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing
7171 	 * vdev that's replacing B with C.  The user's intent in replacing
7172 	 * is to go from M(A,B) to M(A,C).  If the user decides to cancel
7173 	 * the replace by detaching C, the expected behavior is to end up
7174 	 * M(A,B).  But suppose that right after deciding to detach C,
7175 	 * the replacement of B completes.  We would have M(A,C), and then
7176 	 * ask to detach C, which would leave us with just A -- not what
7177 	 * the user wanted.  To prevent this, we make sure that the
7178 	 * parent/child relationship hasn't changed -- in this example,
7179 	 * that C's parent is still the replacing vdev R.
7180 	 */
7181 	if (pvd->vdev_guid != pguid && pguid != 0)
7182 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
7183 
7184 	/*
7185 	 * Only 'replacing' or 'spare' vdevs can be replaced.
7186 	 */
7187 	if (replace_done && pvd->vdev_ops != &vdev_replacing_ops &&
7188 	    pvd->vdev_ops != &vdev_spare_ops)
7189 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
7190 
7191 	ASSERT(pvd->vdev_ops != &vdev_spare_ops ||
7192 	    spa_version(spa) >= SPA_VERSION_SPARES);
7193 
7194 	/*
7195 	 * Only mirror, replacing, and spare vdevs support detach.
7196 	 */
7197 	if (pvd->vdev_ops != &vdev_replacing_ops &&
7198 	    pvd->vdev_ops != &vdev_mirror_ops &&
7199 	    pvd->vdev_ops != &vdev_spare_ops)
7200 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
7201 
7202 	/*
7203 	 * If this device has the only valid copy of some data,
7204 	 * we cannot safely detach it.
7205 	 */
7206 	if (vdev_dtl_required(vd))
7207 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
7208 
7209 	ASSERT(pvd->vdev_children >= 2);
7210 
7211 	/*
7212 	 * If we are detaching the second disk from a replacing vdev, then
7213 	 * check to see if we changed the original vdev's path to have "/old"
7214 	 * at the end in spa_vdev_attach().  If so, undo that change now.
7215 	 */
7216 	if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 &&
7217 	    vd->vdev_path != NULL) {
7218 		size_t len = strlen(vd->vdev_path);
7219 
7220 		for (int c = 0; c < pvd->vdev_children; c++) {
7221 			cvd = pvd->vdev_child[c];
7222 
7223 			if (cvd == vd || cvd->vdev_path == NULL)
7224 				continue;
7225 
7226 			if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 &&
7227 			    strcmp(cvd->vdev_path + len, "/old") == 0) {
7228 				spa_strfree(cvd->vdev_path);
7229 				cvd->vdev_path = spa_strdup(vd->vdev_path);
7230 				break;
7231 			}
7232 		}
7233 	}
7234 
7235 	/*
7236 	 * If we are detaching the original disk from a normal spare, then it
7237 	 * implies that the spare should become a real disk, and be removed
7238 	 * from the active spare list for the pool.  dRAID spares on the
7239 	 * other hand are coupled to the pool and thus should never be removed
7240 	 * from the spares list.
7241 	 */
7242 	if (pvd->vdev_ops == &vdev_spare_ops && vd->vdev_id == 0) {
7243 		vdev_t *last_cvd = pvd->vdev_child[pvd->vdev_children - 1];
7244 
7245 		if (last_cvd->vdev_isspare &&
7246 		    last_cvd->vdev_ops != &vdev_draid_spare_ops) {
7247 			unspare = B_TRUE;
7248 		}
7249 	}
7250 
7251 	/*
7252 	 * Erase the disk labels so the disk can be used for other things.
7253 	 * This must be done after all other error cases are handled,
7254 	 * but before we disembowel vd (so we can still do I/O to it).
7255 	 * But if we can't do it, don't treat the error as fatal --
7256 	 * it may be that the unwritability of the disk is the reason
7257 	 * it's being detached!
7258 	 */
7259 	(void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
7260 
7261 	/*
7262 	 * Remove vd from its parent and compact the parent's children.
7263 	 */
7264 	vdev_remove_child(pvd, vd);
7265 	vdev_compact_children(pvd);
7266 
7267 	/*
7268 	 * Remember one of the remaining children so we can get tvd below.
7269 	 */
7270 	cvd = pvd->vdev_child[pvd->vdev_children - 1];
7271 
7272 	/*
7273 	 * If we need to remove the remaining child from the list of hot spares,
7274 	 * do it now, marking the vdev as no longer a spare in the process.
7275 	 * We must do this before vdev_remove_parent(), because that can
7276 	 * change the GUID if it creates a new toplevel GUID.  For a similar
7277 	 * reason, we must remove the spare now, in the same txg as the detach;
7278 	 * otherwise someone could attach a new sibling, change the GUID, and
7279 	 * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail.
7280 	 */
7281 	if (unspare) {
7282 		ASSERT(cvd->vdev_isspare);
7283 		spa_spare_remove(cvd);
7284 		unspare_guid = cvd->vdev_guid;
7285 		(void) spa_vdev_remove(spa, unspare_guid, B_TRUE);
7286 		cvd->vdev_unspare = B_TRUE;
7287 	}
7288 
7289 	/*
7290 	 * If the parent mirror/replacing vdev only has one child,
7291 	 * the parent is no longer needed.  Remove it from the tree.
7292 	 */
7293 	if (pvd->vdev_children == 1) {
7294 		if (pvd->vdev_ops == &vdev_spare_ops)
7295 			cvd->vdev_unspare = B_FALSE;
7296 		vdev_remove_parent(cvd);
7297 	}
7298 
7299 	/*
7300 	 * We don't set tvd until now because the parent we just removed
7301 	 * may have been the previous top-level vdev.
7302 	 */
7303 	tvd = cvd->vdev_top;
7304 	ASSERT(tvd->vdev_parent == rvd);
7305 
7306 	/*
7307 	 * Reevaluate the parent vdev state.
7308 	 */
7309 	vdev_propagate_state(cvd);
7310 
7311 	/*
7312 	 * If the 'autoexpand' property is set on the pool then automatically
7313 	 * try to expand the size of the pool. For example if the device we
7314 	 * just detached was smaller than the others, it may be possible to
7315 	 * add metaslabs (i.e. grow the pool). We need to reopen the vdev
7316 	 * first so that we can obtain the updated sizes of the leaf vdevs.
7317 	 */
7318 	if (spa->spa_autoexpand) {
7319 		vdev_reopen(tvd);
7320 		vdev_expand(tvd, txg);
7321 	}
7322 
7323 	vdev_config_dirty(tvd);
7324 
7325 	/*
7326 	 * Mark vd's DTL as dirty in this txg.  vdev_dtl_sync() will see that
7327 	 * vd->vdev_detached is set and free vd's DTL object in syncing context.
7328 	 * But first make sure we're not on any *other* txg's DTL list, to
7329 	 * prevent vd from being accessed after it's freed.
7330 	 */
7331 	vdpath = spa_strdup(vd->vdev_path ? vd->vdev_path : "none");
7332 	for (int t = 0; t < TXG_SIZE; t++)
7333 		(void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t);
7334 	vd->vdev_detached = B_TRUE;
7335 	vdev_dirty(tvd, VDD_DTL, vd, txg);
7336 
7337 	spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE);
7338 	spa_notify_waiters(spa);
7339 
7340 	/* hang on to the spa before we release the lock */
7341 	spa_open_ref(spa, FTAG);
7342 
7343 	error = spa_vdev_exit(spa, vd, txg, 0);
7344 
7345 	spa_history_log_internal(spa, "detach", NULL,
7346 	    "vdev=%s", vdpath);
7347 	spa_strfree(vdpath);
7348 
7349 	/*
7350 	 * If this was the removal of the original device in a hot spare vdev,
7351 	 * then we want to go through and remove the device from the hot spare
7352 	 * list of every other pool.
7353 	 */
7354 	if (unspare) {
7355 		spa_t *altspa = NULL;
7356 
7357 		mutex_enter(&spa_namespace_lock);
7358 		while ((altspa = spa_next(altspa)) != NULL) {
7359 			if (altspa->spa_state != POOL_STATE_ACTIVE ||
7360 			    altspa == spa)
7361 				continue;
7362 
7363 			spa_open_ref(altspa, FTAG);
7364 			mutex_exit(&spa_namespace_lock);
7365 			(void) spa_vdev_remove(altspa, unspare_guid, B_TRUE);
7366 			mutex_enter(&spa_namespace_lock);
7367 			spa_close(altspa, FTAG);
7368 		}
7369 		mutex_exit(&spa_namespace_lock);
7370 
7371 		/* search the rest of the vdevs for spares to remove */
7372 		spa_vdev_resilver_done(spa);
7373 	}
7374 
7375 	/* all done with the spa; OK to release */
7376 	mutex_enter(&spa_namespace_lock);
7377 	spa_close(spa, FTAG);
7378 	mutex_exit(&spa_namespace_lock);
7379 
7380 	return (error);
7381 }
7382 
7383 static int
7384 spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
7385     list_t *vd_list)
7386 {
7387 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7388 
7389 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
7390 
7391 	/* Look up vdev and ensure it's a leaf. */
7392 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7393 	if (vd == NULL || vd->vdev_detached) {
7394 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7395 		return (SET_ERROR(ENODEV));
7396 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
7397 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7398 		return (SET_ERROR(EINVAL));
7399 	} else if (!vdev_writeable(vd)) {
7400 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7401 		return (SET_ERROR(EROFS));
7402 	}
7403 	mutex_enter(&vd->vdev_initialize_lock);
7404 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7405 
7406 	/*
7407 	 * When we activate an initialize action we check to see
7408 	 * if the vdev_initialize_thread is NULL. We do this instead
7409 	 * of using the vdev_initialize_state since there might be
7410 	 * a previous initialization process which has completed but
7411 	 * the thread is not exited.
7412 	 */
7413 	if (cmd_type == POOL_INITIALIZE_START &&
7414 	    (vd->vdev_initialize_thread != NULL ||
7415 	    vd->vdev_top->vdev_removing)) {
7416 		mutex_exit(&vd->vdev_initialize_lock);
7417 		return (SET_ERROR(EBUSY));
7418 	} else if (cmd_type == POOL_INITIALIZE_CANCEL &&
7419 	    (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE &&
7420 	    vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) {
7421 		mutex_exit(&vd->vdev_initialize_lock);
7422 		return (SET_ERROR(ESRCH));
7423 	} else if (cmd_type == POOL_INITIALIZE_SUSPEND &&
7424 	    vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) {
7425 		mutex_exit(&vd->vdev_initialize_lock);
7426 		return (SET_ERROR(ESRCH));
7427 	} else if (cmd_type == POOL_INITIALIZE_UNINIT &&
7428 	    vd->vdev_initialize_thread != NULL) {
7429 		mutex_exit(&vd->vdev_initialize_lock);
7430 		return (SET_ERROR(EBUSY));
7431 	}
7432 
7433 	switch (cmd_type) {
7434 	case POOL_INITIALIZE_START:
7435 		vdev_initialize(vd);
7436 		break;
7437 	case POOL_INITIALIZE_CANCEL:
7438 		vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, vd_list);
7439 		break;
7440 	case POOL_INITIALIZE_SUSPEND:
7441 		vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED, vd_list);
7442 		break;
7443 	case POOL_INITIALIZE_UNINIT:
7444 		vdev_uninitialize(vd);
7445 		break;
7446 	default:
7447 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
7448 	}
7449 	mutex_exit(&vd->vdev_initialize_lock);
7450 
7451 	return (0);
7452 }
7453 
7454 int
7455 spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
7456     nvlist_t *vdev_errlist)
7457 {
7458 	int total_errors = 0;
7459 	list_t vd_list;
7460 
7461 	list_create(&vd_list, sizeof (vdev_t),
7462 	    offsetof(vdev_t, vdev_initialize_node));
7463 
7464 	/*
7465 	 * We hold the namespace lock through the whole function
7466 	 * to prevent any changes to the pool while we're starting or
7467 	 * stopping initialization. The config and state locks are held so that
7468 	 * we can properly assess the vdev state before we commit to
7469 	 * the initializing operation.
7470 	 */
7471 	mutex_enter(&spa_namespace_lock);
7472 
7473 	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
7474 	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
7475 		uint64_t vdev_guid = fnvpair_value_uint64(pair);
7476 
7477 		int error = spa_vdev_initialize_impl(spa, vdev_guid, cmd_type,
7478 		    &vd_list);
7479 		if (error != 0) {
7480 			char guid_as_str[MAXNAMELEN];
7481 
7482 			(void) snprintf(guid_as_str, sizeof (guid_as_str),
7483 			    "%llu", (unsigned long long)vdev_guid);
7484 			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
7485 			total_errors++;
7486 		}
7487 	}
7488 
7489 	/* Wait for all initialize threads to stop. */
7490 	vdev_initialize_stop_wait(spa, &vd_list);
7491 
7492 	/* Sync out the initializing state */
7493 	txg_wait_synced(spa->spa_dsl_pool, 0);
7494 	mutex_exit(&spa_namespace_lock);
7495 
7496 	list_destroy(&vd_list);
7497 
7498 	return (total_errors);
7499 }
7500 
7501 static int
7502 spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
7503     uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list)
7504 {
7505 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7506 
7507 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
7508 
7509 	/* Look up vdev and ensure it's a leaf. */
7510 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7511 	if (vd == NULL || vd->vdev_detached) {
7512 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7513 		return (SET_ERROR(ENODEV));
7514 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
7515 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7516 		return (SET_ERROR(EINVAL));
7517 	} else if (!vdev_writeable(vd)) {
7518 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7519 		return (SET_ERROR(EROFS));
7520 	} else if (!vd->vdev_has_trim) {
7521 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7522 		return (SET_ERROR(EOPNOTSUPP));
7523 	} else if (secure && !vd->vdev_has_securetrim) {
7524 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7525 		return (SET_ERROR(EOPNOTSUPP));
7526 	}
7527 	mutex_enter(&vd->vdev_trim_lock);
7528 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7529 
7530 	/*
7531 	 * When we activate a TRIM action we check to see if the
7532 	 * vdev_trim_thread is NULL. We do this instead of using the
7533 	 * vdev_trim_state since there might be a previous TRIM process
7534 	 * which has completed but the thread is not exited.
7535 	 */
7536 	if (cmd_type == POOL_TRIM_START &&
7537 	    (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) {
7538 		mutex_exit(&vd->vdev_trim_lock);
7539 		return (SET_ERROR(EBUSY));
7540 	} else if (cmd_type == POOL_TRIM_CANCEL &&
7541 	    (vd->vdev_trim_state != VDEV_TRIM_ACTIVE &&
7542 	    vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) {
7543 		mutex_exit(&vd->vdev_trim_lock);
7544 		return (SET_ERROR(ESRCH));
7545 	} else if (cmd_type == POOL_TRIM_SUSPEND &&
7546 	    vd->vdev_trim_state != VDEV_TRIM_ACTIVE) {
7547 		mutex_exit(&vd->vdev_trim_lock);
7548 		return (SET_ERROR(ESRCH));
7549 	}
7550 
7551 	switch (cmd_type) {
7552 	case POOL_TRIM_START:
7553 		vdev_trim(vd, rate, partial, secure);
7554 		break;
7555 	case POOL_TRIM_CANCEL:
7556 		vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list);
7557 		break;
7558 	case POOL_TRIM_SUSPEND:
7559 		vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list);
7560 		break;
7561 	default:
7562 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
7563 	}
7564 	mutex_exit(&vd->vdev_trim_lock);
7565 
7566 	return (0);
7567 }
7568 
7569 /*
7570  * Initiates a manual TRIM for the requested vdevs. This kicks off individual
7571  * TRIM threads for each child vdev.  These threads pass over all of the free
7572  * space in the vdev's metaslabs and issues TRIM commands for that space.
7573  */
7574 int
7575 spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate,
7576     boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist)
7577 {
7578 	int total_errors = 0;
7579 	list_t vd_list;
7580 
7581 	list_create(&vd_list, sizeof (vdev_t),
7582 	    offsetof(vdev_t, vdev_trim_node));
7583 
7584 	/*
7585 	 * We hold the namespace lock through the whole function
7586 	 * to prevent any changes to the pool while we're starting or
7587 	 * stopping TRIM. The config and state locks are held so that
7588 	 * we can properly assess the vdev state before we commit to
7589 	 * the TRIM operation.
7590 	 */
7591 	mutex_enter(&spa_namespace_lock);
7592 
7593 	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
7594 	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
7595 		uint64_t vdev_guid = fnvpair_value_uint64(pair);
7596 
7597 		int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type,
7598 		    rate, partial, secure, &vd_list);
7599 		if (error != 0) {
7600 			char guid_as_str[MAXNAMELEN];
7601 
7602 			(void) snprintf(guid_as_str, sizeof (guid_as_str),
7603 			    "%llu", (unsigned long long)vdev_guid);
7604 			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
7605 			total_errors++;
7606 		}
7607 	}
7608 
7609 	/* Wait for all TRIM threads to stop. */
7610 	vdev_trim_stop_wait(spa, &vd_list);
7611 
7612 	/* Sync out the TRIM state */
7613 	txg_wait_synced(spa->spa_dsl_pool, 0);
7614 	mutex_exit(&spa_namespace_lock);
7615 
7616 	list_destroy(&vd_list);
7617 
7618 	return (total_errors);
7619 }
7620 
7621 /*
7622  * Split a set of devices from their mirrors, and create a new pool from them.
7623  */
7624 int
7625 spa_vdev_split_mirror(spa_t *spa, const char *newname, nvlist_t *config,
7626     nvlist_t *props, boolean_t exp)
7627 {
7628 	int error = 0;
7629 	uint64_t txg, *glist;
7630 	spa_t *newspa;
7631 	uint_t c, children, lastlog;
7632 	nvlist_t **child, *nvl, *tmp;
7633 	dmu_tx_t *tx;
7634 	const char *altroot = NULL;
7635 	vdev_t *rvd, **vml = NULL;			/* vdev modify list */
7636 	boolean_t activate_slog;
7637 
7638 	ASSERT(spa_writeable(spa));
7639 
7640 	txg = spa_vdev_enter(spa);
7641 
7642 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7643 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
7644 		error = (spa_has_checkpoint(spa)) ?
7645 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
7646 		return (spa_vdev_exit(spa, NULL, txg, error));
7647 	}
7648 
7649 	/* clear the log and flush everything up to now */
7650 	activate_slog = spa_passivate_log(spa);
7651 	(void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
7652 	error = spa_reset_logs(spa);
7653 	txg = spa_vdev_config_enter(spa);
7654 
7655 	if (activate_slog)
7656 		spa_activate_log(spa);
7657 
7658 	if (error != 0)
7659 		return (spa_vdev_exit(spa, NULL, txg, error));
7660 
7661 	/* check new spa name before going any further */
7662 	if (spa_lookup(newname) != NULL)
7663 		return (spa_vdev_exit(spa, NULL, txg, EEXIST));
7664 
7665 	/*
7666 	 * scan through all the children to ensure they're all mirrors
7667 	 */
7668 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 ||
7669 	    nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child,
7670 	    &children) != 0)
7671 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7672 
7673 	/* first, check to ensure we've got the right child count */
7674 	rvd = spa->spa_root_vdev;
7675 	lastlog = 0;
7676 	for (c = 0; c < rvd->vdev_children; c++) {
7677 		vdev_t *vd = rvd->vdev_child[c];
7678 
7679 		/* don't count the holes & logs as children */
7680 		if (vd->vdev_islog || (vd->vdev_ops != &vdev_indirect_ops &&
7681 		    !vdev_is_concrete(vd))) {
7682 			if (lastlog == 0)
7683 				lastlog = c;
7684 			continue;
7685 		}
7686 
7687 		lastlog = 0;
7688 	}
7689 	if (children != (lastlog != 0 ? lastlog : rvd->vdev_children))
7690 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7691 
7692 	/* next, ensure no spare or cache devices are part of the split */
7693 	if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 ||
7694 	    nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0)
7695 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7696 
7697 	vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP);
7698 	glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP);
7699 
7700 	/* then, loop over each vdev and validate it */
7701 	for (c = 0; c < children; c++) {
7702 		uint64_t is_hole = 0;
7703 
7704 		(void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE,
7705 		    &is_hole);
7706 
7707 		if (is_hole != 0) {
7708 			if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole ||
7709 			    spa->spa_root_vdev->vdev_child[c]->vdev_islog) {
7710 				continue;
7711 			} else {
7712 				error = SET_ERROR(EINVAL);
7713 				break;
7714 			}
7715 		}
7716 
7717 		/* deal with indirect vdevs */
7718 		if (spa->spa_root_vdev->vdev_child[c]->vdev_ops ==
7719 		    &vdev_indirect_ops)
7720 			continue;
7721 
7722 		/* which disk is going to be split? */
7723 		if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID,
7724 		    &glist[c]) != 0) {
7725 			error = SET_ERROR(EINVAL);
7726 			break;
7727 		}
7728 
7729 		/* look it up in the spa */
7730 		vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE);
7731 		if (vml[c] == NULL) {
7732 			error = SET_ERROR(ENODEV);
7733 			break;
7734 		}
7735 
7736 		/* make sure there's nothing stopping the split */
7737 		if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops ||
7738 		    vml[c]->vdev_islog ||
7739 		    !vdev_is_concrete(vml[c]) ||
7740 		    vml[c]->vdev_isspare ||
7741 		    vml[c]->vdev_isl2cache ||
7742 		    !vdev_writeable(vml[c]) ||
7743 		    vml[c]->vdev_children != 0 ||
7744 		    vml[c]->vdev_state != VDEV_STATE_HEALTHY ||
7745 		    c != spa->spa_root_vdev->vdev_child[c]->vdev_id) {
7746 			error = SET_ERROR(EINVAL);
7747 			break;
7748 		}
7749 
7750 		if (vdev_dtl_required(vml[c]) ||
7751 		    vdev_resilver_needed(vml[c], NULL, NULL)) {
7752 			error = SET_ERROR(EBUSY);
7753 			break;
7754 		}
7755 
7756 		/* we need certain info from the top level */
7757 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY,
7758 		    vml[c]->vdev_top->vdev_ms_array);
7759 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT,
7760 		    vml[c]->vdev_top->vdev_ms_shift);
7761 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE,
7762 		    vml[c]->vdev_top->vdev_asize);
7763 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT,
7764 		    vml[c]->vdev_top->vdev_ashift);
7765 
7766 		/* transfer per-vdev ZAPs */
7767 		ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0);
7768 		VERIFY0(nvlist_add_uint64(child[c],
7769 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap));
7770 
7771 		ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0);
7772 		VERIFY0(nvlist_add_uint64(child[c],
7773 		    ZPOOL_CONFIG_VDEV_TOP_ZAP,
7774 		    vml[c]->vdev_parent->vdev_top_zap));
7775 	}
7776 
7777 	if (error != 0) {
7778 		kmem_free(vml, children * sizeof (vdev_t *));
7779 		kmem_free(glist, children * sizeof (uint64_t));
7780 		return (spa_vdev_exit(spa, NULL, txg, error));
7781 	}
7782 
7783 	/* stop writers from using the disks */
7784 	for (c = 0; c < children; c++) {
7785 		if (vml[c] != NULL)
7786 			vml[c]->vdev_offline = B_TRUE;
7787 	}
7788 	vdev_reopen(spa->spa_root_vdev);
7789 
7790 	/*
7791 	 * Temporarily record the splitting vdevs in the spa config.  This
7792 	 * will disappear once the config is regenerated.
7793 	 */
7794 	nvl = fnvlist_alloc();
7795 	fnvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, glist, children);
7796 	kmem_free(glist, children * sizeof (uint64_t));
7797 
7798 	mutex_enter(&spa->spa_props_lock);
7799 	fnvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT, nvl);
7800 	mutex_exit(&spa->spa_props_lock);
7801 	spa->spa_config_splitting = nvl;
7802 	vdev_config_dirty(spa->spa_root_vdev);
7803 
7804 	/* configure and create the new pool */
7805 	fnvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname);
7806 	fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE,
7807 	    exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE);
7808 	fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, spa_version(spa));
7809 	fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG, spa->spa_config_txg);
7810 	fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID,
7811 	    spa_generate_guid(NULL));
7812 	VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
7813 	(void) nvlist_lookup_string(props,
7814 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
7815 
7816 	/* add the new pool to the namespace */
7817 	newspa = spa_add(newname, config, altroot);
7818 	newspa->spa_avz_action = AVZ_ACTION_REBUILD;
7819 	newspa->spa_config_txg = spa->spa_config_txg;
7820 	spa_set_log_state(newspa, SPA_LOG_CLEAR);
7821 
7822 	/* release the spa config lock, retaining the namespace lock */
7823 	spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
7824 
7825 	if (zio_injection_enabled)
7826 		zio_handle_panic_injection(spa, FTAG, 1);
7827 
7828 	spa_activate(newspa, spa_mode_global);
7829 	spa_async_suspend(newspa);
7830 
7831 	/*
7832 	 * Temporarily stop the initializing and TRIM activity.  We set the
7833 	 * state to ACTIVE so that we know to resume initializing or TRIM
7834 	 * once the split has completed.
7835 	 */
7836 	list_t vd_initialize_list;
7837 	list_create(&vd_initialize_list, sizeof (vdev_t),
7838 	    offsetof(vdev_t, vdev_initialize_node));
7839 
7840 	list_t vd_trim_list;
7841 	list_create(&vd_trim_list, sizeof (vdev_t),
7842 	    offsetof(vdev_t, vdev_trim_node));
7843 
7844 	for (c = 0; c < children; c++) {
7845 		if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) {
7846 			mutex_enter(&vml[c]->vdev_initialize_lock);
7847 			vdev_initialize_stop(vml[c],
7848 			    VDEV_INITIALIZE_ACTIVE, &vd_initialize_list);
7849 			mutex_exit(&vml[c]->vdev_initialize_lock);
7850 
7851 			mutex_enter(&vml[c]->vdev_trim_lock);
7852 			vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list);
7853 			mutex_exit(&vml[c]->vdev_trim_lock);
7854 		}
7855 	}
7856 
7857 	vdev_initialize_stop_wait(spa, &vd_initialize_list);
7858 	vdev_trim_stop_wait(spa, &vd_trim_list);
7859 
7860 	list_destroy(&vd_initialize_list);
7861 	list_destroy(&vd_trim_list);
7862 
7863 	newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT;
7864 	newspa->spa_is_splitting = B_TRUE;
7865 
7866 	/* create the new pool from the disks of the original pool */
7867 	error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE);
7868 	if (error)
7869 		goto out;
7870 
7871 	/* if that worked, generate a real config for the new pool */
7872 	if (newspa->spa_root_vdev != NULL) {
7873 		newspa->spa_config_splitting = fnvlist_alloc();
7874 		fnvlist_add_uint64(newspa->spa_config_splitting,
7875 		    ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa));
7876 		spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL,
7877 		    B_TRUE));
7878 	}
7879 
7880 	/* set the props */
7881 	if (props != NULL) {
7882 		spa_configfile_set(newspa, props, B_FALSE);
7883 		error = spa_prop_set(newspa, props);
7884 		if (error)
7885 			goto out;
7886 	}
7887 
7888 	/* flush everything */
7889 	txg = spa_vdev_config_enter(newspa);
7890 	vdev_config_dirty(newspa->spa_root_vdev);
7891 	(void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG);
7892 
7893 	if (zio_injection_enabled)
7894 		zio_handle_panic_injection(spa, FTAG, 2);
7895 
7896 	spa_async_resume(newspa);
7897 
7898 	/* finally, update the original pool's config */
7899 	txg = spa_vdev_config_enter(spa);
7900 	tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
7901 	error = dmu_tx_assign(tx, TXG_WAIT);
7902 	if (error != 0)
7903 		dmu_tx_abort(tx);
7904 	for (c = 0; c < children; c++) {
7905 		if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) {
7906 			vdev_t *tvd = vml[c]->vdev_top;
7907 
7908 			/*
7909 			 * Need to be sure the detachable VDEV is not
7910 			 * on any *other* txg's DTL list to prevent it
7911 			 * from being accessed after it's freed.
7912 			 */
7913 			for (int t = 0; t < TXG_SIZE; t++) {
7914 				(void) txg_list_remove_this(
7915 				    &tvd->vdev_dtl_list, vml[c], t);
7916 			}
7917 
7918 			vdev_split(vml[c]);
7919 			if (error == 0)
7920 				spa_history_log_internal(spa, "detach", tx,
7921 				    "vdev=%s", vml[c]->vdev_path);
7922 
7923 			vdev_free(vml[c]);
7924 		}
7925 	}
7926 	spa->spa_avz_action = AVZ_ACTION_REBUILD;
7927 	vdev_config_dirty(spa->spa_root_vdev);
7928 	spa->spa_config_splitting = NULL;
7929 	nvlist_free(nvl);
7930 	if (error == 0)
7931 		dmu_tx_commit(tx);
7932 	(void) spa_vdev_exit(spa, NULL, txg, 0);
7933 
7934 	if (zio_injection_enabled)
7935 		zio_handle_panic_injection(spa, FTAG, 3);
7936 
7937 	/* split is complete; log a history record */
7938 	spa_history_log_internal(newspa, "split", NULL,
7939 	    "from pool %s", spa_name(spa));
7940 
7941 	newspa->spa_is_splitting = B_FALSE;
7942 	kmem_free(vml, children * sizeof (vdev_t *));
7943 
7944 	/* if we're not going to mount the filesystems in userland, export */
7945 	if (exp)
7946 		error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL,
7947 		    B_FALSE, B_FALSE);
7948 
7949 	return (error);
7950 
7951 out:
7952 	spa_unload(newspa);
7953 	spa_deactivate(newspa);
7954 	spa_remove(newspa);
7955 
7956 	txg = spa_vdev_config_enter(spa);
7957 
7958 	/* re-online all offlined disks */
7959 	for (c = 0; c < children; c++) {
7960 		if (vml[c] != NULL)
7961 			vml[c]->vdev_offline = B_FALSE;
7962 	}
7963 
7964 	/* restart initializing or trimming disks as necessary */
7965 	spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
7966 	spa_async_request(spa, SPA_ASYNC_TRIM_RESTART);
7967 	spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART);
7968 
7969 	vdev_reopen(spa->spa_root_vdev);
7970 
7971 	nvlist_free(spa->spa_config_splitting);
7972 	spa->spa_config_splitting = NULL;
7973 	(void) spa_vdev_exit(spa, NULL, txg, error);
7974 
7975 	kmem_free(vml, children * sizeof (vdev_t *));
7976 	return (error);
7977 }
7978 
7979 /*
7980  * Find any device that's done replacing, or a vdev marked 'unspare' that's
7981  * currently spared, so we can detach it.
7982  */
7983 static vdev_t *
7984 spa_vdev_resilver_done_hunt(vdev_t *vd)
7985 {
7986 	vdev_t *newvd, *oldvd;
7987 
7988 	for (int c = 0; c < vd->vdev_children; c++) {
7989 		oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]);
7990 		if (oldvd != NULL)
7991 			return (oldvd);
7992 	}
7993 
7994 	/*
7995 	 * Check for a completed replacement.  We always consider the first
7996 	 * vdev in the list to be the oldest vdev, and the last one to be
7997 	 * the newest (see spa_vdev_attach() for how that works).  In
7998 	 * the case where the newest vdev is faulted, we will not automatically
7999 	 * remove it after a resilver completes.  This is OK as it will require
8000 	 * user intervention to determine which disk the admin wishes to keep.
8001 	 */
8002 	if (vd->vdev_ops == &vdev_replacing_ops) {
8003 		ASSERT(vd->vdev_children > 1);
8004 
8005 		newvd = vd->vdev_child[vd->vdev_children - 1];
8006 		oldvd = vd->vdev_child[0];
8007 
8008 		if (vdev_dtl_empty(newvd, DTL_MISSING) &&
8009 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
8010 		    !vdev_dtl_required(oldvd))
8011 			return (oldvd);
8012 	}
8013 
8014 	/*
8015 	 * Check for a completed resilver with the 'unspare' flag set.
8016 	 * Also potentially update faulted state.
8017 	 */
8018 	if (vd->vdev_ops == &vdev_spare_ops) {
8019 		vdev_t *first = vd->vdev_child[0];
8020 		vdev_t *last = vd->vdev_child[vd->vdev_children - 1];
8021 
8022 		if (last->vdev_unspare) {
8023 			oldvd = first;
8024 			newvd = last;
8025 		} else if (first->vdev_unspare) {
8026 			oldvd = last;
8027 			newvd = first;
8028 		} else {
8029 			oldvd = NULL;
8030 		}
8031 
8032 		if (oldvd != NULL &&
8033 		    vdev_dtl_empty(newvd, DTL_MISSING) &&
8034 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
8035 		    !vdev_dtl_required(oldvd))
8036 			return (oldvd);
8037 
8038 		vdev_propagate_state(vd);
8039 
8040 		/*
8041 		 * If there are more than two spares attached to a disk,
8042 		 * and those spares are not required, then we want to
8043 		 * attempt to free them up now so that they can be used
8044 		 * by other pools.  Once we're back down to a single
8045 		 * disk+spare, we stop removing them.
8046 		 */
8047 		if (vd->vdev_children > 2) {
8048 			newvd = vd->vdev_child[1];
8049 
8050 			if (newvd->vdev_isspare && last->vdev_isspare &&
8051 			    vdev_dtl_empty(last, DTL_MISSING) &&
8052 			    vdev_dtl_empty(last, DTL_OUTAGE) &&
8053 			    !vdev_dtl_required(newvd))
8054 				return (newvd);
8055 		}
8056 	}
8057 
8058 	return (NULL);
8059 }
8060 
8061 static void
8062 spa_vdev_resilver_done(spa_t *spa)
8063 {
8064 	vdev_t *vd, *pvd, *ppvd;
8065 	uint64_t guid, sguid, pguid, ppguid;
8066 
8067 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
8068 
8069 	while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) {
8070 		pvd = vd->vdev_parent;
8071 		ppvd = pvd->vdev_parent;
8072 		guid = vd->vdev_guid;
8073 		pguid = pvd->vdev_guid;
8074 		ppguid = ppvd->vdev_guid;
8075 		sguid = 0;
8076 		/*
8077 		 * If we have just finished replacing a hot spared device, then
8078 		 * we need to detach the parent's first child (the original hot
8079 		 * spare) as well.
8080 		 */
8081 		if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 &&
8082 		    ppvd->vdev_children == 2) {
8083 			ASSERT(pvd->vdev_ops == &vdev_replacing_ops);
8084 			sguid = ppvd->vdev_child[1]->vdev_guid;
8085 		}
8086 		ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd));
8087 
8088 		spa_config_exit(spa, SCL_ALL, FTAG);
8089 		if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0)
8090 			return;
8091 		if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0)
8092 			return;
8093 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
8094 	}
8095 
8096 	spa_config_exit(spa, SCL_ALL, FTAG);
8097 
8098 	/*
8099 	 * If a detach was not performed above replace waiters will not have
8100 	 * been notified.  In which case we must do so now.
8101 	 */
8102 	spa_notify_waiters(spa);
8103 }
8104 
8105 /*
8106  * Update the stored path or FRU for this vdev.
8107  */
8108 static int
8109 spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value,
8110     boolean_t ispath)
8111 {
8112 	vdev_t *vd;
8113 	boolean_t sync = B_FALSE;
8114 
8115 	ASSERT(spa_writeable(spa));
8116 
8117 	spa_vdev_state_enter(spa, SCL_ALL);
8118 
8119 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
8120 		return (spa_vdev_state_exit(spa, NULL, ENOENT));
8121 
8122 	if (!vd->vdev_ops->vdev_op_leaf)
8123 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
8124 
8125 	if (ispath) {
8126 		if (strcmp(value, vd->vdev_path) != 0) {
8127 			spa_strfree(vd->vdev_path);
8128 			vd->vdev_path = spa_strdup(value);
8129 			sync = B_TRUE;
8130 		}
8131 	} else {
8132 		if (vd->vdev_fru == NULL) {
8133 			vd->vdev_fru = spa_strdup(value);
8134 			sync = B_TRUE;
8135 		} else if (strcmp(value, vd->vdev_fru) != 0) {
8136 			spa_strfree(vd->vdev_fru);
8137 			vd->vdev_fru = spa_strdup(value);
8138 			sync = B_TRUE;
8139 		}
8140 	}
8141 
8142 	return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0));
8143 }
8144 
8145 int
8146 spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath)
8147 {
8148 	return (spa_vdev_set_common(spa, guid, newpath, B_TRUE));
8149 }
8150 
8151 int
8152 spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru)
8153 {
8154 	return (spa_vdev_set_common(spa, guid, newfru, B_FALSE));
8155 }
8156 
8157 /*
8158  * ==========================================================================
8159  * SPA Scanning
8160  * ==========================================================================
8161  */
8162 int
8163 spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd)
8164 {
8165 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
8166 
8167 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
8168 		return (SET_ERROR(EBUSY));
8169 
8170 	return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd));
8171 }
8172 
8173 int
8174 spa_scan_stop(spa_t *spa)
8175 {
8176 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
8177 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
8178 		return (SET_ERROR(EBUSY));
8179 
8180 	return (dsl_scan_cancel(spa->spa_dsl_pool));
8181 }
8182 
8183 int
8184 spa_scan(spa_t *spa, pool_scan_func_t func)
8185 {
8186 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
8187 
8188 	if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE)
8189 		return (SET_ERROR(ENOTSUP));
8190 
8191 	if (func == POOL_SCAN_RESILVER &&
8192 	    !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER))
8193 		return (SET_ERROR(ENOTSUP));
8194 
8195 	/*
8196 	 * If a resilver was requested, but there is no DTL on a
8197 	 * writeable leaf device, we have nothing to do.
8198 	 */
8199 	if (func == POOL_SCAN_RESILVER &&
8200 	    !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
8201 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
8202 		return (0);
8203 	}
8204 
8205 	if (func == POOL_SCAN_ERRORSCRUB &&
8206 	    !spa_feature_is_enabled(spa, SPA_FEATURE_HEAD_ERRLOG))
8207 		return (SET_ERROR(ENOTSUP));
8208 
8209 	return (dsl_scan(spa->spa_dsl_pool, func));
8210 }
8211 
8212 /*
8213  * ==========================================================================
8214  * SPA async task processing
8215  * ==========================================================================
8216  */
8217 
8218 static void
8219 spa_async_remove(spa_t *spa, vdev_t *vd)
8220 {
8221 	if (vd->vdev_remove_wanted) {
8222 		vd->vdev_remove_wanted = B_FALSE;
8223 		vd->vdev_delayed_close = B_FALSE;
8224 		vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE);
8225 
8226 		/*
8227 		 * We want to clear the stats, but we don't want to do a full
8228 		 * vdev_clear() as that will cause us to throw away
8229 		 * degraded/faulted state as well as attempt to reopen the
8230 		 * device, all of which is a waste.
8231 		 */
8232 		vd->vdev_stat.vs_read_errors = 0;
8233 		vd->vdev_stat.vs_write_errors = 0;
8234 		vd->vdev_stat.vs_checksum_errors = 0;
8235 
8236 		vdev_state_dirty(vd->vdev_top);
8237 
8238 		/* Tell userspace that the vdev is gone. */
8239 		zfs_post_remove(spa, vd);
8240 	}
8241 
8242 	for (int c = 0; c < vd->vdev_children; c++)
8243 		spa_async_remove(spa, vd->vdev_child[c]);
8244 }
8245 
8246 static void
8247 spa_async_probe(spa_t *spa, vdev_t *vd)
8248 {
8249 	if (vd->vdev_probe_wanted) {
8250 		vd->vdev_probe_wanted = B_FALSE;
8251 		vdev_reopen(vd);	/* vdev_open() does the actual probe */
8252 	}
8253 
8254 	for (int c = 0; c < vd->vdev_children; c++)
8255 		spa_async_probe(spa, vd->vdev_child[c]);
8256 }
8257 
8258 static void
8259 spa_async_autoexpand(spa_t *spa, vdev_t *vd)
8260 {
8261 	if (!spa->spa_autoexpand)
8262 		return;
8263 
8264 	for (int c = 0; c < vd->vdev_children; c++) {
8265 		vdev_t *cvd = vd->vdev_child[c];
8266 		spa_async_autoexpand(spa, cvd);
8267 	}
8268 
8269 	if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL)
8270 		return;
8271 
8272 	spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_AUTOEXPAND);
8273 }
8274 
8275 static __attribute__((noreturn)) void
8276 spa_async_thread(void *arg)
8277 {
8278 	spa_t *spa = (spa_t *)arg;
8279 	dsl_pool_t *dp = spa->spa_dsl_pool;
8280 	int tasks;
8281 
8282 	ASSERT(spa->spa_sync_on);
8283 
8284 	mutex_enter(&spa->spa_async_lock);
8285 	tasks = spa->spa_async_tasks;
8286 	spa->spa_async_tasks = 0;
8287 	mutex_exit(&spa->spa_async_lock);
8288 
8289 	/*
8290 	 * See if the config needs to be updated.
8291 	 */
8292 	if (tasks & SPA_ASYNC_CONFIG_UPDATE) {
8293 		uint64_t old_space, new_space;
8294 
8295 		mutex_enter(&spa_namespace_lock);
8296 		old_space = metaslab_class_get_space(spa_normal_class(spa));
8297 		old_space += metaslab_class_get_space(spa_special_class(spa));
8298 		old_space += metaslab_class_get_space(spa_dedup_class(spa));
8299 		old_space += metaslab_class_get_space(
8300 		    spa_embedded_log_class(spa));
8301 
8302 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
8303 
8304 		new_space = metaslab_class_get_space(spa_normal_class(spa));
8305 		new_space += metaslab_class_get_space(spa_special_class(spa));
8306 		new_space += metaslab_class_get_space(spa_dedup_class(spa));
8307 		new_space += metaslab_class_get_space(
8308 		    spa_embedded_log_class(spa));
8309 		mutex_exit(&spa_namespace_lock);
8310 
8311 		/*
8312 		 * If the pool grew as a result of the config update,
8313 		 * then log an internal history event.
8314 		 */
8315 		if (new_space != old_space) {
8316 			spa_history_log_internal(spa, "vdev online", NULL,
8317 			    "pool '%s' size: %llu(+%llu)",
8318 			    spa_name(spa), (u_longlong_t)new_space,
8319 			    (u_longlong_t)(new_space - old_space));
8320 		}
8321 	}
8322 
8323 	/*
8324 	 * See if any devices need to be marked REMOVED.
8325 	 */
8326 	if (tasks & SPA_ASYNC_REMOVE) {
8327 		spa_vdev_state_enter(spa, SCL_NONE);
8328 		spa_async_remove(spa, spa->spa_root_vdev);
8329 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++)
8330 			spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]);
8331 		for (int i = 0; i < spa->spa_spares.sav_count; i++)
8332 			spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]);
8333 		(void) spa_vdev_state_exit(spa, NULL, 0);
8334 	}
8335 
8336 	if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) {
8337 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8338 		spa_async_autoexpand(spa, spa->spa_root_vdev);
8339 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8340 	}
8341 
8342 	/*
8343 	 * See if any devices need to be probed.
8344 	 */
8345 	if (tasks & SPA_ASYNC_PROBE) {
8346 		spa_vdev_state_enter(spa, SCL_NONE);
8347 		spa_async_probe(spa, spa->spa_root_vdev);
8348 		(void) spa_vdev_state_exit(spa, NULL, 0);
8349 	}
8350 
8351 	/*
8352 	 * If any devices are done replacing, detach them.
8353 	 */
8354 	if (tasks & SPA_ASYNC_RESILVER_DONE ||
8355 	    tasks & SPA_ASYNC_REBUILD_DONE ||
8356 	    tasks & SPA_ASYNC_DETACH_SPARE) {
8357 		spa_vdev_resilver_done(spa);
8358 	}
8359 
8360 	/*
8361 	 * Kick off a resilver.
8362 	 */
8363 	if (tasks & SPA_ASYNC_RESILVER &&
8364 	    !vdev_rebuild_active(spa->spa_root_vdev) &&
8365 	    (!dsl_scan_resilvering(dp) ||
8366 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER)))
8367 		dsl_scan_restart_resilver(dp, 0);
8368 
8369 	if (tasks & SPA_ASYNC_INITIALIZE_RESTART) {
8370 		mutex_enter(&spa_namespace_lock);
8371 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8372 		vdev_initialize_restart(spa->spa_root_vdev);
8373 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8374 		mutex_exit(&spa_namespace_lock);
8375 	}
8376 
8377 	if (tasks & SPA_ASYNC_TRIM_RESTART) {
8378 		mutex_enter(&spa_namespace_lock);
8379 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8380 		vdev_trim_restart(spa->spa_root_vdev);
8381 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8382 		mutex_exit(&spa_namespace_lock);
8383 	}
8384 
8385 	if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) {
8386 		mutex_enter(&spa_namespace_lock);
8387 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8388 		vdev_autotrim_restart(spa);
8389 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8390 		mutex_exit(&spa_namespace_lock);
8391 	}
8392 
8393 	/*
8394 	 * Kick off L2 cache whole device TRIM.
8395 	 */
8396 	if (tasks & SPA_ASYNC_L2CACHE_TRIM) {
8397 		mutex_enter(&spa_namespace_lock);
8398 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8399 		vdev_trim_l2arc(spa);
8400 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8401 		mutex_exit(&spa_namespace_lock);
8402 	}
8403 
8404 	/*
8405 	 * Kick off L2 cache rebuilding.
8406 	 */
8407 	if (tasks & SPA_ASYNC_L2CACHE_REBUILD) {
8408 		mutex_enter(&spa_namespace_lock);
8409 		spa_config_enter(spa, SCL_L2ARC, FTAG, RW_READER);
8410 		l2arc_spa_rebuild_start(spa);
8411 		spa_config_exit(spa, SCL_L2ARC, FTAG);
8412 		mutex_exit(&spa_namespace_lock);
8413 	}
8414 
8415 	/*
8416 	 * Let the world know that we're done.
8417 	 */
8418 	mutex_enter(&spa->spa_async_lock);
8419 	spa->spa_async_thread = NULL;
8420 	cv_broadcast(&spa->spa_async_cv);
8421 	mutex_exit(&spa->spa_async_lock);
8422 	thread_exit();
8423 }
8424 
8425 void
8426 spa_async_suspend(spa_t *spa)
8427 {
8428 	mutex_enter(&spa->spa_async_lock);
8429 	spa->spa_async_suspended++;
8430 	while (spa->spa_async_thread != NULL)
8431 		cv_wait(&spa->spa_async_cv, &spa->spa_async_lock);
8432 	mutex_exit(&spa->spa_async_lock);
8433 
8434 	spa_vdev_remove_suspend(spa);
8435 
8436 	zthr_t *condense_thread = spa->spa_condense_zthr;
8437 	if (condense_thread != NULL)
8438 		zthr_cancel(condense_thread);
8439 
8440 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
8441 	if (discard_thread != NULL)
8442 		zthr_cancel(discard_thread);
8443 
8444 	zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr;
8445 	if (ll_delete_thread != NULL)
8446 		zthr_cancel(ll_delete_thread);
8447 
8448 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
8449 	if (ll_condense_thread != NULL)
8450 		zthr_cancel(ll_condense_thread);
8451 }
8452 
8453 void
8454 spa_async_resume(spa_t *spa)
8455 {
8456 	mutex_enter(&spa->spa_async_lock);
8457 	ASSERT(spa->spa_async_suspended != 0);
8458 	spa->spa_async_suspended--;
8459 	mutex_exit(&spa->spa_async_lock);
8460 	spa_restart_removal(spa);
8461 
8462 	zthr_t *condense_thread = spa->spa_condense_zthr;
8463 	if (condense_thread != NULL)
8464 		zthr_resume(condense_thread);
8465 
8466 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
8467 	if (discard_thread != NULL)
8468 		zthr_resume(discard_thread);
8469 
8470 	zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr;
8471 	if (ll_delete_thread != NULL)
8472 		zthr_resume(ll_delete_thread);
8473 
8474 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
8475 	if (ll_condense_thread != NULL)
8476 		zthr_resume(ll_condense_thread);
8477 }
8478 
8479 static boolean_t
8480 spa_async_tasks_pending(spa_t *spa)
8481 {
8482 	uint_t non_config_tasks;
8483 	uint_t config_task;
8484 	boolean_t config_task_suspended;
8485 
8486 	non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE;
8487 	config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE;
8488 	if (spa->spa_ccw_fail_time == 0) {
8489 		config_task_suspended = B_FALSE;
8490 	} else {
8491 		config_task_suspended =
8492 		    (gethrtime() - spa->spa_ccw_fail_time) <
8493 		    ((hrtime_t)zfs_ccw_retry_interval * NANOSEC);
8494 	}
8495 
8496 	return (non_config_tasks || (config_task && !config_task_suspended));
8497 }
8498 
8499 static void
8500 spa_async_dispatch(spa_t *spa)
8501 {
8502 	mutex_enter(&spa->spa_async_lock);
8503 	if (spa_async_tasks_pending(spa) &&
8504 	    !spa->spa_async_suspended &&
8505 	    spa->spa_async_thread == NULL)
8506 		spa->spa_async_thread = thread_create(NULL, 0,
8507 		    spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri);
8508 	mutex_exit(&spa->spa_async_lock);
8509 }
8510 
8511 void
8512 spa_async_request(spa_t *spa, int task)
8513 {
8514 	zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task);
8515 	mutex_enter(&spa->spa_async_lock);
8516 	spa->spa_async_tasks |= task;
8517 	mutex_exit(&spa->spa_async_lock);
8518 }
8519 
8520 int
8521 spa_async_tasks(spa_t *spa)
8522 {
8523 	return (spa->spa_async_tasks);
8524 }
8525 
8526 /*
8527  * ==========================================================================
8528  * SPA syncing routines
8529  * ==========================================================================
8530  */
8531 
8532 
8533 static int
8534 bpobj_enqueue_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
8535     dmu_tx_t *tx)
8536 {
8537 	bpobj_t *bpo = arg;
8538 	bpobj_enqueue(bpo, bp, bp_freed, tx);
8539 	return (0);
8540 }
8541 
8542 int
8543 bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8544 {
8545 	return (bpobj_enqueue_cb(arg, bp, B_FALSE, tx));
8546 }
8547 
8548 int
8549 bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8550 {
8551 	return (bpobj_enqueue_cb(arg, bp, B_TRUE, tx));
8552 }
8553 
8554 static int
8555 spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8556 {
8557 	zio_t *pio = arg;
8558 
8559 	zio_nowait(zio_free_sync(pio, pio->io_spa, dmu_tx_get_txg(tx), bp,
8560 	    pio->io_flags));
8561 	return (0);
8562 }
8563 
8564 static int
8565 bpobj_spa_free_sync_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
8566     dmu_tx_t *tx)
8567 {
8568 	ASSERT(!bp_freed);
8569 	return (spa_free_sync_cb(arg, bp, tx));
8570 }
8571 
8572 /*
8573  * Note: this simple function is not inlined to make it easier to dtrace the
8574  * amount of time spent syncing frees.
8575  */
8576 static void
8577 spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx)
8578 {
8579 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
8580 	bplist_iterate(bpl, spa_free_sync_cb, zio, tx);
8581 	VERIFY(zio_wait(zio) == 0);
8582 }
8583 
8584 /*
8585  * Note: this simple function is not inlined to make it easier to dtrace the
8586  * amount of time spent syncing deferred frees.
8587  */
8588 static void
8589 spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx)
8590 {
8591 	if (spa_sync_pass(spa) != 1)
8592 		return;
8593 
8594 	/*
8595 	 * Note:
8596 	 * If the log space map feature is active, we stop deferring
8597 	 * frees to the next TXG and therefore running this function
8598 	 * would be considered a no-op as spa_deferred_bpobj should
8599 	 * not have any entries.
8600 	 *
8601 	 * That said we run this function anyway (instead of returning
8602 	 * immediately) for the edge-case scenario where we just
8603 	 * activated the log space map feature in this TXG but we have
8604 	 * deferred frees from the previous TXG.
8605 	 */
8606 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
8607 	VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj,
8608 	    bpobj_spa_free_sync_cb, zio, tx), ==, 0);
8609 	VERIFY0(zio_wait(zio));
8610 }
8611 
8612 static void
8613 spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx)
8614 {
8615 	char *packed = NULL;
8616 	size_t bufsize;
8617 	size_t nvsize = 0;
8618 	dmu_buf_t *db;
8619 
8620 	VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0);
8621 
8622 	/*
8623 	 * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration
8624 	 * information.  This avoids the dmu_buf_will_dirty() path and
8625 	 * saves us a pre-read to get data we don't actually care about.
8626 	 */
8627 	bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE);
8628 	packed = vmem_alloc(bufsize, KM_SLEEP);
8629 
8630 	VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR,
8631 	    KM_SLEEP) == 0);
8632 	memset(packed + nvsize, 0, bufsize - nvsize);
8633 
8634 	dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx);
8635 
8636 	vmem_free(packed, bufsize);
8637 
8638 	VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db));
8639 	dmu_buf_will_dirty(db, tx);
8640 	*(uint64_t *)db->db_data = nvsize;
8641 	dmu_buf_rele(db, FTAG);
8642 }
8643 
8644 static void
8645 spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx,
8646     const char *config, const char *entry)
8647 {
8648 	nvlist_t *nvroot;
8649 	nvlist_t **list;
8650 	int i;
8651 
8652 	if (!sav->sav_sync)
8653 		return;
8654 
8655 	/*
8656 	 * Update the MOS nvlist describing the list of available devices.
8657 	 * spa_validate_aux() will have already made sure this nvlist is
8658 	 * valid and the vdevs are labeled appropriately.
8659 	 */
8660 	if (sav->sav_object == 0) {
8661 		sav->sav_object = dmu_object_alloc(spa->spa_meta_objset,
8662 		    DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE,
8663 		    sizeof (uint64_t), tx);
8664 		VERIFY(zap_update(spa->spa_meta_objset,
8665 		    DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1,
8666 		    &sav->sav_object, tx) == 0);
8667 	}
8668 
8669 	nvroot = fnvlist_alloc();
8670 	if (sav->sav_count == 0) {
8671 		fnvlist_add_nvlist_array(nvroot, config,
8672 		    (const nvlist_t * const *)NULL, 0);
8673 	} else {
8674 		list = kmem_alloc(sav->sav_count*sizeof (void *), KM_SLEEP);
8675 		for (i = 0; i < sav->sav_count; i++)
8676 			list[i] = vdev_config_generate(spa, sav->sav_vdevs[i],
8677 			    B_FALSE, VDEV_CONFIG_L2CACHE);
8678 		fnvlist_add_nvlist_array(nvroot, config,
8679 		    (const nvlist_t * const *)list, sav->sav_count);
8680 		for (i = 0; i < sav->sav_count; i++)
8681 			nvlist_free(list[i]);
8682 		kmem_free(list, sav->sav_count * sizeof (void *));
8683 	}
8684 
8685 	spa_sync_nvlist(spa, sav->sav_object, nvroot, tx);
8686 	nvlist_free(nvroot);
8687 
8688 	sav->sav_sync = B_FALSE;
8689 }
8690 
8691 /*
8692  * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t.
8693  * The all-vdev ZAP must be empty.
8694  */
8695 static void
8696 spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx)
8697 {
8698 	spa_t *spa = vd->vdev_spa;
8699 
8700 	if (vd->vdev_root_zap != 0 &&
8701 	    spa_feature_is_active(spa, SPA_FEATURE_AVZ_V2)) {
8702 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8703 		    vd->vdev_root_zap, tx));
8704 	}
8705 	if (vd->vdev_top_zap != 0) {
8706 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8707 		    vd->vdev_top_zap, tx));
8708 	}
8709 	if (vd->vdev_leaf_zap != 0) {
8710 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8711 		    vd->vdev_leaf_zap, tx));
8712 	}
8713 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
8714 		spa_avz_build(vd->vdev_child[i], avz, tx);
8715 	}
8716 }
8717 
8718 static void
8719 spa_sync_config_object(spa_t *spa, dmu_tx_t *tx)
8720 {
8721 	nvlist_t *config;
8722 
8723 	/*
8724 	 * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS,
8725 	 * its config may not be dirty but we still need to build per-vdev ZAPs.
8726 	 * Similarly, if the pool is being assembled (e.g. after a split), we
8727 	 * need to rebuild the AVZ although the config may not be dirty.
8728 	 */
8729 	if (list_is_empty(&spa->spa_config_dirty_list) &&
8730 	    spa->spa_avz_action == AVZ_ACTION_NONE)
8731 		return;
8732 
8733 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
8734 
8735 	ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE ||
8736 	    spa->spa_avz_action == AVZ_ACTION_INITIALIZE ||
8737 	    spa->spa_all_vdev_zaps != 0);
8738 
8739 	if (spa->spa_avz_action == AVZ_ACTION_REBUILD) {
8740 		/* Make and build the new AVZ */
8741 		uint64_t new_avz = zap_create(spa->spa_meta_objset,
8742 		    DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx);
8743 		spa_avz_build(spa->spa_root_vdev, new_avz, tx);
8744 
8745 		/* Diff old AVZ with new one */
8746 		zap_cursor_t zc;
8747 		zap_attribute_t za;
8748 
8749 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
8750 		    spa->spa_all_vdev_zaps);
8751 		    zap_cursor_retrieve(&zc, &za) == 0;
8752 		    zap_cursor_advance(&zc)) {
8753 			uint64_t vdzap = za.za_first_integer;
8754 			if (zap_lookup_int(spa->spa_meta_objset, new_avz,
8755 			    vdzap) == ENOENT) {
8756 				/*
8757 				 * ZAP is listed in old AVZ but not in new one;
8758 				 * destroy it
8759 				 */
8760 				VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap,
8761 				    tx));
8762 			}
8763 		}
8764 
8765 		zap_cursor_fini(&zc);
8766 
8767 		/* Destroy the old AVZ */
8768 		VERIFY0(zap_destroy(spa->spa_meta_objset,
8769 		    spa->spa_all_vdev_zaps, tx));
8770 
8771 		/* Replace the old AVZ in the dir obj with the new one */
8772 		VERIFY0(zap_update(spa->spa_meta_objset,
8773 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP,
8774 		    sizeof (new_avz), 1, &new_avz, tx));
8775 
8776 		spa->spa_all_vdev_zaps = new_avz;
8777 	} else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) {
8778 		zap_cursor_t zc;
8779 		zap_attribute_t za;
8780 
8781 		/* Walk through the AVZ and destroy all listed ZAPs */
8782 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
8783 		    spa->spa_all_vdev_zaps);
8784 		    zap_cursor_retrieve(&zc, &za) == 0;
8785 		    zap_cursor_advance(&zc)) {
8786 			uint64_t zap = za.za_first_integer;
8787 			VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx));
8788 		}
8789 
8790 		zap_cursor_fini(&zc);
8791 
8792 		/* Destroy and unlink the AVZ itself */
8793 		VERIFY0(zap_destroy(spa->spa_meta_objset,
8794 		    spa->spa_all_vdev_zaps, tx));
8795 		VERIFY0(zap_remove(spa->spa_meta_objset,
8796 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx));
8797 		spa->spa_all_vdev_zaps = 0;
8798 	}
8799 
8800 	if (spa->spa_all_vdev_zaps == 0) {
8801 		spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset,
8802 		    DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
8803 		    DMU_POOL_VDEV_ZAP_MAP, tx);
8804 	}
8805 	spa->spa_avz_action = AVZ_ACTION_NONE;
8806 
8807 	/* Create ZAPs for vdevs that don't have them. */
8808 	vdev_construct_zaps(spa->spa_root_vdev, tx);
8809 
8810 	config = spa_config_generate(spa, spa->spa_root_vdev,
8811 	    dmu_tx_get_txg(tx), B_FALSE);
8812 
8813 	/*
8814 	 * If we're upgrading the spa version then make sure that
8815 	 * the config object gets updated with the correct version.
8816 	 */
8817 	if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version)
8818 		fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION,
8819 		    spa->spa_uberblock.ub_version);
8820 
8821 	spa_config_exit(spa, SCL_STATE, FTAG);
8822 
8823 	nvlist_free(spa->spa_config_syncing);
8824 	spa->spa_config_syncing = config;
8825 
8826 	spa_sync_nvlist(spa, spa->spa_config_object, config, tx);
8827 }
8828 
8829 static void
8830 spa_sync_version(void *arg, dmu_tx_t *tx)
8831 {
8832 	uint64_t *versionp = arg;
8833 	uint64_t version = *versionp;
8834 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
8835 
8836 	/*
8837 	 * Setting the version is special cased when first creating the pool.
8838 	 */
8839 	ASSERT(tx->tx_txg != TXG_INITIAL);
8840 
8841 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
8842 	ASSERT(version >= spa_version(spa));
8843 
8844 	spa->spa_uberblock.ub_version = version;
8845 	vdev_config_dirty(spa->spa_root_vdev);
8846 	spa_history_log_internal(spa, "set", tx, "version=%lld",
8847 	    (longlong_t)version);
8848 }
8849 
8850 /*
8851  * Set zpool properties.
8852  */
8853 static void
8854 spa_sync_props(void *arg, dmu_tx_t *tx)
8855 {
8856 	nvlist_t *nvp = arg;
8857 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
8858 	objset_t *mos = spa->spa_meta_objset;
8859 	nvpair_t *elem = NULL;
8860 
8861 	mutex_enter(&spa->spa_props_lock);
8862 
8863 	while ((elem = nvlist_next_nvpair(nvp, elem))) {
8864 		uint64_t intval;
8865 		const char *strval, *fname;
8866 		zpool_prop_t prop;
8867 		const char *propname;
8868 		const char *elemname = nvpair_name(elem);
8869 		zprop_type_t proptype;
8870 		spa_feature_t fid;
8871 
8872 		switch (prop = zpool_name_to_prop(elemname)) {
8873 		case ZPOOL_PROP_VERSION:
8874 			intval = fnvpair_value_uint64(elem);
8875 			/*
8876 			 * The version is synced separately before other
8877 			 * properties and should be correct by now.
8878 			 */
8879 			ASSERT3U(spa_version(spa), >=, intval);
8880 			break;
8881 
8882 		case ZPOOL_PROP_ALTROOT:
8883 			/*
8884 			 * 'altroot' is a non-persistent property. It should
8885 			 * have been set temporarily at creation or import time.
8886 			 */
8887 			ASSERT(spa->spa_root != NULL);
8888 			break;
8889 
8890 		case ZPOOL_PROP_READONLY:
8891 		case ZPOOL_PROP_CACHEFILE:
8892 			/*
8893 			 * 'readonly' and 'cachefile' are also non-persistent
8894 			 * properties.
8895 			 */
8896 			break;
8897 		case ZPOOL_PROP_COMMENT:
8898 			strval = fnvpair_value_string(elem);
8899 			if (spa->spa_comment != NULL)
8900 				spa_strfree(spa->spa_comment);
8901 			spa->spa_comment = spa_strdup(strval);
8902 			/*
8903 			 * We need to dirty the configuration on all the vdevs
8904 			 * so that their labels get updated.  We also need to
8905 			 * update the cache file to keep it in sync with the
8906 			 * MOS version. It's unnecessary to do this for pool
8907 			 * creation since the vdev's configuration has already
8908 			 * been dirtied.
8909 			 */
8910 			if (tx->tx_txg != TXG_INITIAL) {
8911 				vdev_config_dirty(spa->spa_root_vdev);
8912 				spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
8913 			}
8914 			spa_history_log_internal(spa, "set", tx,
8915 			    "%s=%s", elemname, strval);
8916 			break;
8917 		case ZPOOL_PROP_COMPATIBILITY:
8918 			strval = fnvpair_value_string(elem);
8919 			if (spa->spa_compatibility != NULL)
8920 				spa_strfree(spa->spa_compatibility);
8921 			spa->spa_compatibility = spa_strdup(strval);
8922 			/*
8923 			 * Dirty the configuration on vdevs as above.
8924 			 */
8925 			if (tx->tx_txg != TXG_INITIAL) {
8926 				vdev_config_dirty(spa->spa_root_vdev);
8927 				spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
8928 			}
8929 
8930 			spa_history_log_internal(spa, "set", tx,
8931 			    "%s=%s", nvpair_name(elem), strval);
8932 			break;
8933 
8934 		case ZPOOL_PROP_INVAL:
8935 			if (zpool_prop_feature(elemname)) {
8936 				fname = strchr(elemname, '@') + 1;
8937 				VERIFY0(zfeature_lookup_name(fname, &fid));
8938 
8939 				spa_feature_enable(spa, fid, tx);
8940 				spa_history_log_internal(spa, "set", tx,
8941 				    "%s=enabled", elemname);
8942 				break;
8943 			} else if (!zfs_prop_user(elemname)) {
8944 				ASSERT(zpool_prop_feature(elemname));
8945 				break;
8946 			}
8947 			zfs_fallthrough;
8948 		default:
8949 			/*
8950 			 * Set pool property values in the poolprops mos object.
8951 			 */
8952 			if (spa->spa_pool_props_object == 0) {
8953 				spa->spa_pool_props_object =
8954 				    zap_create_link(mos, DMU_OT_POOL_PROPS,
8955 				    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS,
8956 				    tx);
8957 			}
8958 
8959 			/* normalize the property name */
8960 			if (prop == ZPOOL_PROP_INVAL) {
8961 				propname = elemname;
8962 				proptype = PROP_TYPE_STRING;
8963 			} else {
8964 				propname = zpool_prop_to_name(prop);
8965 				proptype = zpool_prop_get_type(prop);
8966 			}
8967 
8968 			if (nvpair_type(elem) == DATA_TYPE_STRING) {
8969 				ASSERT(proptype == PROP_TYPE_STRING);
8970 				strval = fnvpair_value_string(elem);
8971 				VERIFY0(zap_update(mos,
8972 				    spa->spa_pool_props_object, propname,
8973 				    1, strlen(strval) + 1, strval, tx));
8974 				spa_history_log_internal(spa, "set", tx,
8975 				    "%s=%s", elemname, strval);
8976 			} else if (nvpair_type(elem) == DATA_TYPE_UINT64) {
8977 				intval = fnvpair_value_uint64(elem);
8978 
8979 				if (proptype == PROP_TYPE_INDEX) {
8980 					const char *unused;
8981 					VERIFY0(zpool_prop_index_to_string(
8982 					    prop, intval, &unused));
8983 				}
8984 				VERIFY0(zap_update(mos,
8985 				    spa->spa_pool_props_object, propname,
8986 				    8, 1, &intval, tx));
8987 				spa_history_log_internal(spa, "set", tx,
8988 				    "%s=%lld", elemname,
8989 				    (longlong_t)intval);
8990 
8991 				switch (prop) {
8992 				case ZPOOL_PROP_DELEGATION:
8993 					spa->spa_delegation = intval;
8994 					break;
8995 				case ZPOOL_PROP_BOOTFS:
8996 					spa->spa_bootfs = intval;
8997 					break;
8998 				case ZPOOL_PROP_FAILUREMODE:
8999 					spa->spa_failmode = intval;
9000 					break;
9001 				case ZPOOL_PROP_AUTOTRIM:
9002 					spa->spa_autotrim = intval;
9003 					spa_async_request(spa,
9004 					    SPA_ASYNC_AUTOTRIM_RESTART);
9005 					break;
9006 				case ZPOOL_PROP_AUTOEXPAND:
9007 					spa->spa_autoexpand = intval;
9008 					if (tx->tx_txg != TXG_INITIAL)
9009 						spa_async_request(spa,
9010 						    SPA_ASYNC_AUTOEXPAND);
9011 					break;
9012 				case ZPOOL_PROP_MULTIHOST:
9013 					spa->spa_multihost = intval;
9014 					break;
9015 				default:
9016 					break;
9017 				}
9018 			} else {
9019 				ASSERT(0); /* not allowed */
9020 			}
9021 		}
9022 
9023 	}
9024 
9025 	mutex_exit(&spa->spa_props_lock);
9026 }
9027 
9028 /*
9029  * Perform one-time upgrade on-disk changes.  spa_version() does not
9030  * reflect the new version this txg, so there must be no changes this
9031  * txg to anything that the upgrade code depends on after it executes.
9032  * Therefore this must be called after dsl_pool_sync() does the sync
9033  * tasks.
9034  */
9035 static void
9036 spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx)
9037 {
9038 	if (spa_sync_pass(spa) != 1)
9039 		return;
9040 
9041 	dsl_pool_t *dp = spa->spa_dsl_pool;
9042 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
9043 
9044 	if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN &&
9045 	    spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) {
9046 		dsl_pool_create_origin(dp, tx);
9047 
9048 		/* Keeping the origin open increases spa_minref */
9049 		spa->spa_minref += 3;
9050 	}
9051 
9052 	if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES &&
9053 	    spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) {
9054 		dsl_pool_upgrade_clones(dp, tx);
9055 	}
9056 
9057 	if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES &&
9058 	    spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) {
9059 		dsl_pool_upgrade_dir_clones(dp, tx);
9060 
9061 		/* Keeping the freedir open increases spa_minref */
9062 		spa->spa_minref += 3;
9063 	}
9064 
9065 	if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES &&
9066 	    spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
9067 		spa_feature_create_zap_objects(spa, tx);
9068 	}
9069 
9070 	/*
9071 	 * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable
9072 	 * when possibility to use lz4 compression for metadata was added
9073 	 * Old pools that have this feature enabled must be upgraded to have
9074 	 * this feature active
9075 	 */
9076 	if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
9077 		boolean_t lz4_en = spa_feature_is_enabled(spa,
9078 		    SPA_FEATURE_LZ4_COMPRESS);
9079 		boolean_t lz4_ac = spa_feature_is_active(spa,
9080 		    SPA_FEATURE_LZ4_COMPRESS);
9081 
9082 		if (lz4_en && !lz4_ac)
9083 			spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx);
9084 	}
9085 
9086 	/*
9087 	 * If we haven't written the salt, do so now.  Note that the
9088 	 * feature may not be activated yet, but that's fine since
9089 	 * the presence of this ZAP entry is backwards compatible.
9090 	 */
9091 	if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
9092 	    DMU_POOL_CHECKSUM_SALT) == ENOENT) {
9093 		VERIFY0(zap_add(spa->spa_meta_objset,
9094 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1,
9095 		    sizeof (spa->spa_cksum_salt.zcs_bytes),
9096 		    spa->spa_cksum_salt.zcs_bytes, tx));
9097 	}
9098 
9099 	rrw_exit(&dp->dp_config_rwlock, FTAG);
9100 }
9101 
9102 static void
9103 vdev_indirect_state_sync_verify(vdev_t *vd)
9104 {
9105 	vdev_indirect_mapping_t *vim __maybe_unused = vd->vdev_indirect_mapping;
9106 	vdev_indirect_births_t *vib __maybe_unused = vd->vdev_indirect_births;
9107 
9108 	if (vd->vdev_ops == &vdev_indirect_ops) {
9109 		ASSERT(vim != NULL);
9110 		ASSERT(vib != NULL);
9111 	}
9112 
9113 	uint64_t obsolete_sm_object = 0;
9114 	ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object));
9115 	if (obsolete_sm_object != 0) {
9116 		ASSERT(vd->vdev_obsolete_sm != NULL);
9117 		ASSERT(vd->vdev_removing ||
9118 		    vd->vdev_ops == &vdev_indirect_ops);
9119 		ASSERT(vdev_indirect_mapping_num_entries(vim) > 0);
9120 		ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0);
9121 		ASSERT3U(obsolete_sm_object, ==,
9122 		    space_map_object(vd->vdev_obsolete_sm));
9123 		ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=,
9124 		    space_map_allocated(vd->vdev_obsolete_sm));
9125 	}
9126 	ASSERT(vd->vdev_obsolete_segments != NULL);
9127 
9128 	/*
9129 	 * Since frees / remaps to an indirect vdev can only
9130 	 * happen in syncing context, the obsolete segments
9131 	 * tree must be empty when we start syncing.
9132 	 */
9133 	ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
9134 }
9135 
9136 /*
9137  * Set the top-level vdev's max queue depth. Evaluate each top-level's
9138  * async write queue depth in case it changed. The max queue depth will
9139  * not change in the middle of syncing out this txg.
9140  */
9141 static void
9142 spa_sync_adjust_vdev_max_queue_depth(spa_t *spa)
9143 {
9144 	ASSERT(spa_writeable(spa));
9145 
9146 	vdev_t *rvd = spa->spa_root_vdev;
9147 	uint32_t max_queue_depth = zfs_vdev_async_write_max_active *
9148 	    zfs_vdev_queue_depth_pct / 100;
9149 	metaslab_class_t *normal = spa_normal_class(spa);
9150 	metaslab_class_t *special = spa_special_class(spa);
9151 	metaslab_class_t *dedup = spa_dedup_class(spa);
9152 
9153 	uint64_t slots_per_allocator = 0;
9154 	for (int c = 0; c < rvd->vdev_children; c++) {
9155 		vdev_t *tvd = rvd->vdev_child[c];
9156 
9157 		metaslab_group_t *mg = tvd->vdev_mg;
9158 		if (mg == NULL || !metaslab_group_initialized(mg))
9159 			continue;
9160 
9161 		metaslab_class_t *mc = mg->mg_class;
9162 		if (mc != normal && mc != special && mc != dedup)
9163 			continue;
9164 
9165 		/*
9166 		 * It is safe to do a lock-free check here because only async
9167 		 * allocations look at mg_max_alloc_queue_depth, and async
9168 		 * allocations all happen from spa_sync().
9169 		 */
9170 		for (int i = 0; i < mg->mg_allocators; i++) {
9171 			ASSERT0(zfs_refcount_count(
9172 			    &(mg->mg_allocator[i].mga_alloc_queue_depth)));
9173 		}
9174 		mg->mg_max_alloc_queue_depth = max_queue_depth;
9175 
9176 		for (int i = 0; i < mg->mg_allocators; i++) {
9177 			mg->mg_allocator[i].mga_cur_max_alloc_queue_depth =
9178 			    zfs_vdev_def_queue_depth;
9179 		}
9180 		slots_per_allocator += zfs_vdev_def_queue_depth;
9181 	}
9182 
9183 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9184 		ASSERT0(zfs_refcount_count(&normal->mc_allocator[i].
9185 		    mca_alloc_slots));
9186 		ASSERT0(zfs_refcount_count(&special->mc_allocator[i].
9187 		    mca_alloc_slots));
9188 		ASSERT0(zfs_refcount_count(&dedup->mc_allocator[i].
9189 		    mca_alloc_slots));
9190 		normal->mc_allocator[i].mca_alloc_max_slots =
9191 		    slots_per_allocator;
9192 		special->mc_allocator[i].mca_alloc_max_slots =
9193 		    slots_per_allocator;
9194 		dedup->mc_allocator[i].mca_alloc_max_slots =
9195 		    slots_per_allocator;
9196 	}
9197 	normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
9198 	special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
9199 	dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
9200 }
9201 
9202 static void
9203 spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx)
9204 {
9205 	ASSERT(spa_writeable(spa));
9206 
9207 	vdev_t *rvd = spa->spa_root_vdev;
9208 	for (int c = 0; c < rvd->vdev_children; c++) {
9209 		vdev_t *vd = rvd->vdev_child[c];
9210 		vdev_indirect_state_sync_verify(vd);
9211 
9212 		if (vdev_indirect_should_condense(vd)) {
9213 			spa_condense_indirect_start_sync(vd, tx);
9214 			break;
9215 		}
9216 	}
9217 }
9218 
9219 static void
9220 spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx)
9221 {
9222 	objset_t *mos = spa->spa_meta_objset;
9223 	dsl_pool_t *dp = spa->spa_dsl_pool;
9224 	uint64_t txg = tx->tx_txg;
9225 	bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK];
9226 
9227 	do {
9228 		int pass = ++spa->spa_sync_pass;
9229 
9230 		spa_sync_config_object(spa, tx);
9231 		spa_sync_aux_dev(spa, &spa->spa_spares, tx,
9232 		    ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES);
9233 		spa_sync_aux_dev(spa, &spa->spa_l2cache, tx,
9234 		    ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE);
9235 		spa_errlog_sync(spa, txg);
9236 		dsl_pool_sync(dp, txg);
9237 
9238 		if (pass < zfs_sync_pass_deferred_free ||
9239 		    spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) {
9240 			/*
9241 			 * If the log space map feature is active we don't
9242 			 * care about deferred frees and the deferred bpobj
9243 			 * as the log space map should effectively have the
9244 			 * same results (i.e. appending only to one object).
9245 			 */
9246 			spa_sync_frees(spa, free_bpl, tx);
9247 		} else {
9248 			/*
9249 			 * We can not defer frees in pass 1, because
9250 			 * we sync the deferred frees later in pass 1.
9251 			 */
9252 			ASSERT3U(pass, >, 1);
9253 			bplist_iterate(free_bpl, bpobj_enqueue_alloc_cb,
9254 			    &spa->spa_deferred_bpobj, tx);
9255 		}
9256 
9257 		brt_sync(spa, txg);
9258 		ddt_sync(spa, txg);
9259 		dsl_scan_sync(dp, tx);
9260 		dsl_errorscrub_sync(dp, tx);
9261 		svr_sync(spa, tx);
9262 		spa_sync_upgrades(spa, tx);
9263 
9264 		spa_flush_metaslabs(spa, tx);
9265 
9266 		vdev_t *vd = NULL;
9267 		while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg))
9268 		    != NULL)
9269 			vdev_sync(vd, txg);
9270 
9271 		/*
9272 		 * Note: We need to check if the MOS is dirty because we could
9273 		 * have marked the MOS dirty without updating the uberblock
9274 		 * (e.g. if we have sync tasks but no dirty user data). We need
9275 		 * to check the uberblock's rootbp because it is updated if we
9276 		 * have synced out dirty data (though in this case the MOS will
9277 		 * most likely also be dirty due to second order effects, we
9278 		 * don't want to rely on that here).
9279 		 */
9280 		if (pass == 1 &&
9281 		    spa->spa_uberblock.ub_rootbp.blk_birth < txg &&
9282 		    !dmu_objset_is_dirty(mos, txg)) {
9283 			/*
9284 			 * Nothing changed on the first pass, therefore this
9285 			 * TXG is a no-op. Avoid syncing deferred frees, so
9286 			 * that we can keep this TXG as a no-op.
9287 			 */
9288 			ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
9289 			ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
9290 			ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg));
9291 			ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg));
9292 			break;
9293 		}
9294 
9295 		spa_sync_deferred_frees(spa, tx);
9296 	} while (dmu_objset_is_dirty(mos, txg));
9297 }
9298 
9299 /*
9300  * Rewrite the vdev configuration (which includes the uberblock) to
9301  * commit the transaction group.
9302  *
9303  * If there are no dirty vdevs, we sync the uberblock to a few random
9304  * top-level vdevs that are known to be visible in the config cache
9305  * (see spa_vdev_add() for a complete description). If there *are* dirty
9306  * vdevs, sync the uberblock to all vdevs.
9307  */
9308 static void
9309 spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx)
9310 {
9311 	vdev_t *rvd = spa->spa_root_vdev;
9312 	uint64_t txg = tx->tx_txg;
9313 
9314 	for (;;) {
9315 		int error = 0;
9316 
9317 		/*
9318 		 * We hold SCL_STATE to prevent vdev open/close/etc.
9319 		 * while we're attempting to write the vdev labels.
9320 		 */
9321 		spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
9322 
9323 		if (list_is_empty(&spa->spa_config_dirty_list)) {
9324 			vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
9325 			int svdcount = 0;
9326 			int children = rvd->vdev_children;
9327 			int c0 = random_in_range(children);
9328 
9329 			for (int c = 0; c < children; c++) {
9330 				vdev_t *vd =
9331 				    rvd->vdev_child[(c0 + c) % children];
9332 
9333 				/* Stop when revisiting the first vdev */
9334 				if (c > 0 && svd[0] == vd)
9335 					break;
9336 
9337 				if (vd->vdev_ms_array == 0 ||
9338 				    vd->vdev_islog ||
9339 				    !vdev_is_concrete(vd))
9340 					continue;
9341 
9342 				svd[svdcount++] = vd;
9343 				if (svdcount == SPA_SYNC_MIN_VDEVS)
9344 					break;
9345 			}
9346 			error = vdev_config_sync(svd, svdcount, txg);
9347 		} else {
9348 			error = vdev_config_sync(rvd->vdev_child,
9349 			    rvd->vdev_children, txg);
9350 		}
9351 
9352 		if (error == 0)
9353 			spa->spa_last_synced_guid = rvd->vdev_guid;
9354 
9355 		spa_config_exit(spa, SCL_STATE, FTAG);
9356 
9357 		if (error == 0)
9358 			break;
9359 		zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR);
9360 		zio_resume_wait(spa);
9361 	}
9362 }
9363 
9364 /*
9365  * Sync the specified transaction group.  New blocks may be dirtied as
9366  * part of the process, so we iterate until it converges.
9367  */
9368 void
9369 spa_sync(spa_t *spa, uint64_t txg)
9370 {
9371 	vdev_t *vd = NULL;
9372 
9373 	VERIFY(spa_writeable(spa));
9374 
9375 	/*
9376 	 * Wait for i/os issued in open context that need to complete
9377 	 * before this txg syncs.
9378 	 */
9379 	(void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]);
9380 	spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL,
9381 	    ZIO_FLAG_CANFAIL);
9382 
9383 	/*
9384 	 * Now that there can be no more cloning in this transaction group,
9385 	 * but we are still before issuing frees, we can process pending BRT
9386 	 * updates.
9387 	 */
9388 	brt_pending_apply(spa, txg);
9389 
9390 	/*
9391 	 * Lock out configuration changes.
9392 	 */
9393 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
9394 
9395 	spa->spa_syncing_txg = txg;
9396 	spa->spa_sync_pass = 0;
9397 
9398 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9399 		mutex_enter(&spa->spa_allocs[i].spaa_lock);
9400 		VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree));
9401 		mutex_exit(&spa->spa_allocs[i].spaa_lock);
9402 	}
9403 
9404 	/*
9405 	 * If there are any pending vdev state changes, convert them
9406 	 * into config changes that go out with this transaction group.
9407 	 */
9408 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
9409 	while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) {
9410 		/* Avoid holding the write lock unless actually necessary */
9411 		if (vd->vdev_aux == NULL) {
9412 			vdev_state_clean(vd);
9413 			vdev_config_dirty(vd);
9414 			continue;
9415 		}
9416 		/*
9417 		 * We need the write lock here because, for aux vdevs,
9418 		 * calling vdev_config_dirty() modifies sav_config.
9419 		 * This is ugly and will become unnecessary when we
9420 		 * eliminate the aux vdev wart by integrating all vdevs
9421 		 * into the root vdev tree.
9422 		 */
9423 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9424 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER);
9425 		while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) {
9426 			vdev_state_clean(vd);
9427 			vdev_config_dirty(vd);
9428 		}
9429 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9430 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9431 	}
9432 	spa_config_exit(spa, SCL_STATE, FTAG);
9433 
9434 	dsl_pool_t *dp = spa->spa_dsl_pool;
9435 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
9436 
9437 	spa->spa_sync_starttime = gethrtime();
9438 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
9439 	spa->spa_deadman_tqid = taskq_dispatch_delay(system_delay_taskq,
9440 	    spa_deadman, spa, TQ_SLEEP, ddi_get_lbolt() +
9441 	    NSEC_TO_TICK(spa->spa_deadman_synctime));
9442 
9443 	/*
9444 	 * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg,
9445 	 * set spa_deflate if we have no raid-z vdevs.
9446 	 */
9447 	if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE &&
9448 	    spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) {
9449 		vdev_t *rvd = spa->spa_root_vdev;
9450 
9451 		int i;
9452 		for (i = 0; i < rvd->vdev_children; i++) {
9453 			vd = rvd->vdev_child[i];
9454 			if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE)
9455 				break;
9456 		}
9457 		if (i == rvd->vdev_children) {
9458 			spa->spa_deflate = TRUE;
9459 			VERIFY0(zap_add(spa->spa_meta_objset,
9460 			    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
9461 			    sizeof (uint64_t), 1, &spa->spa_deflate, tx));
9462 		}
9463 	}
9464 
9465 	spa_sync_adjust_vdev_max_queue_depth(spa);
9466 
9467 	spa_sync_condense_indirect(spa, tx);
9468 
9469 	spa_sync_iterate_to_convergence(spa, tx);
9470 
9471 #ifdef ZFS_DEBUG
9472 	if (!list_is_empty(&spa->spa_config_dirty_list)) {
9473 	/*
9474 	 * Make sure that the number of ZAPs for all the vdevs matches
9475 	 * the number of ZAPs in the per-vdev ZAP list. This only gets
9476 	 * called if the config is dirty; otherwise there may be
9477 	 * outstanding AVZ operations that weren't completed in
9478 	 * spa_sync_config_object.
9479 	 */
9480 		uint64_t all_vdev_zap_entry_count;
9481 		ASSERT0(zap_count(spa->spa_meta_objset,
9482 		    spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count));
9483 		ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==,
9484 		    all_vdev_zap_entry_count);
9485 	}
9486 #endif
9487 
9488 	if (spa->spa_vdev_removal != NULL) {
9489 		ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]);
9490 	}
9491 
9492 	spa_sync_rewrite_vdev_config(spa, tx);
9493 	dmu_tx_commit(tx);
9494 
9495 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
9496 	spa->spa_deadman_tqid = 0;
9497 
9498 	/*
9499 	 * Clear the dirty config list.
9500 	 */
9501 	while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL)
9502 		vdev_config_clean(vd);
9503 
9504 	/*
9505 	 * Now that the new config has synced transactionally,
9506 	 * let it become visible to the config cache.
9507 	 */
9508 	if (spa->spa_config_syncing != NULL) {
9509 		spa_config_set(spa, spa->spa_config_syncing);
9510 		spa->spa_config_txg = txg;
9511 		spa->spa_config_syncing = NULL;
9512 	}
9513 
9514 	dsl_pool_sync_done(dp, txg);
9515 
9516 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9517 		mutex_enter(&spa->spa_allocs[i].spaa_lock);
9518 		VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree));
9519 		mutex_exit(&spa->spa_allocs[i].spaa_lock);
9520 	}
9521 
9522 	/*
9523 	 * Update usable space statistics.
9524 	 */
9525 	while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)))
9526 	    != NULL)
9527 		vdev_sync_done(vd, txg);
9528 
9529 	metaslab_class_evict_old(spa->spa_normal_class, txg);
9530 	metaslab_class_evict_old(spa->spa_log_class, txg);
9531 
9532 	spa_sync_close_syncing_log_sm(spa);
9533 
9534 	spa_update_dspace(spa);
9535 
9536 	if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON)
9537 		vdev_autotrim_kick(spa);
9538 
9539 	/*
9540 	 * It had better be the case that we didn't dirty anything
9541 	 * since vdev_config_sync().
9542 	 */
9543 	ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
9544 	ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
9545 	ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg));
9546 
9547 	while (zfs_pause_spa_sync)
9548 		delay(1);
9549 
9550 	spa->spa_sync_pass = 0;
9551 
9552 	/*
9553 	 * Update the last synced uberblock here. We want to do this at
9554 	 * the end of spa_sync() so that consumers of spa_last_synced_txg()
9555 	 * will be guaranteed that all the processing associated with
9556 	 * that txg has been completed.
9557 	 */
9558 	spa->spa_ubsync = spa->spa_uberblock;
9559 	spa_config_exit(spa, SCL_CONFIG, FTAG);
9560 
9561 	spa_handle_ignored_writes(spa);
9562 
9563 	/*
9564 	 * If any async tasks have been requested, kick them off.
9565 	 */
9566 	spa_async_dispatch(spa);
9567 }
9568 
9569 /*
9570  * Sync all pools.  We don't want to hold the namespace lock across these
9571  * operations, so we take a reference on the spa_t and drop the lock during the
9572  * sync.
9573  */
9574 void
9575 spa_sync_allpools(void)
9576 {
9577 	spa_t *spa = NULL;
9578 	mutex_enter(&spa_namespace_lock);
9579 	while ((spa = spa_next(spa)) != NULL) {
9580 		if (spa_state(spa) != POOL_STATE_ACTIVE ||
9581 		    !spa_writeable(spa) || spa_suspended(spa))
9582 			continue;
9583 		spa_open_ref(spa, FTAG);
9584 		mutex_exit(&spa_namespace_lock);
9585 		txg_wait_synced(spa_get_dsl(spa), 0);
9586 		mutex_enter(&spa_namespace_lock);
9587 		spa_close(spa, FTAG);
9588 	}
9589 	mutex_exit(&spa_namespace_lock);
9590 }
9591 
9592 /*
9593  * ==========================================================================
9594  * Miscellaneous routines
9595  * ==========================================================================
9596  */
9597 
9598 /*
9599  * Remove all pools in the system.
9600  */
9601 void
9602 spa_evict_all(void)
9603 {
9604 	spa_t *spa;
9605 
9606 	/*
9607 	 * Remove all cached state.  All pools should be closed now,
9608 	 * so every spa in the AVL tree should be unreferenced.
9609 	 */
9610 	mutex_enter(&spa_namespace_lock);
9611 	while ((spa = spa_next(NULL)) != NULL) {
9612 		/*
9613 		 * Stop async tasks.  The async thread may need to detach
9614 		 * a device that's been replaced, which requires grabbing
9615 		 * spa_namespace_lock, so we must drop it here.
9616 		 */
9617 		spa_open_ref(spa, FTAG);
9618 		mutex_exit(&spa_namespace_lock);
9619 		spa_async_suspend(spa);
9620 		mutex_enter(&spa_namespace_lock);
9621 		spa_close(spa, FTAG);
9622 
9623 		if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
9624 			spa_unload(spa);
9625 			spa_deactivate(spa);
9626 		}
9627 		spa_remove(spa);
9628 	}
9629 	mutex_exit(&spa_namespace_lock);
9630 }
9631 
9632 vdev_t *
9633 spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux)
9634 {
9635 	vdev_t *vd;
9636 	int i;
9637 
9638 	if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL)
9639 		return (vd);
9640 
9641 	if (aux) {
9642 		for (i = 0; i < spa->spa_l2cache.sav_count; i++) {
9643 			vd = spa->spa_l2cache.sav_vdevs[i];
9644 			if (vd->vdev_guid == guid)
9645 				return (vd);
9646 		}
9647 
9648 		for (i = 0; i < spa->spa_spares.sav_count; i++) {
9649 			vd = spa->spa_spares.sav_vdevs[i];
9650 			if (vd->vdev_guid == guid)
9651 				return (vd);
9652 		}
9653 	}
9654 
9655 	return (NULL);
9656 }
9657 
9658 void
9659 spa_upgrade(spa_t *spa, uint64_t version)
9660 {
9661 	ASSERT(spa_writeable(spa));
9662 
9663 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
9664 
9665 	/*
9666 	 * This should only be called for a non-faulted pool, and since a
9667 	 * future version would result in an unopenable pool, this shouldn't be
9668 	 * possible.
9669 	 */
9670 	ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version));
9671 	ASSERT3U(version, >=, spa->spa_uberblock.ub_version);
9672 
9673 	spa->spa_uberblock.ub_version = version;
9674 	vdev_config_dirty(spa->spa_root_vdev);
9675 
9676 	spa_config_exit(spa, SCL_ALL, FTAG);
9677 
9678 	txg_wait_synced(spa_get_dsl(spa), 0);
9679 }
9680 
9681 static boolean_t
9682 spa_has_aux_vdev(spa_t *spa, uint64_t guid, spa_aux_vdev_t *sav)
9683 {
9684 	(void) spa;
9685 	int i;
9686 	uint64_t vdev_guid;
9687 
9688 	for (i = 0; i < sav->sav_count; i++)
9689 		if (sav->sav_vdevs[i]->vdev_guid == guid)
9690 			return (B_TRUE);
9691 
9692 	for (i = 0; i < sav->sav_npending; i++) {
9693 		if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID,
9694 		    &vdev_guid) == 0 && vdev_guid == guid)
9695 			return (B_TRUE);
9696 	}
9697 
9698 	return (B_FALSE);
9699 }
9700 
9701 boolean_t
9702 spa_has_l2cache(spa_t *spa, uint64_t guid)
9703 {
9704 	return (spa_has_aux_vdev(spa, guid, &spa->spa_l2cache));
9705 }
9706 
9707 boolean_t
9708 spa_has_spare(spa_t *spa, uint64_t guid)
9709 {
9710 	return (spa_has_aux_vdev(spa, guid, &spa->spa_spares));
9711 }
9712 
9713 /*
9714  * Check if a pool has an active shared spare device.
9715  * Note: reference count of an active spare is 2, as a spare and as a replace
9716  */
9717 static boolean_t
9718 spa_has_active_shared_spare(spa_t *spa)
9719 {
9720 	int i, refcnt;
9721 	uint64_t pool;
9722 	spa_aux_vdev_t *sav = &spa->spa_spares;
9723 
9724 	for (i = 0; i < sav->sav_count; i++) {
9725 		if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool,
9726 		    &refcnt) && pool != 0ULL && pool == spa_guid(spa) &&
9727 		    refcnt > 2)
9728 			return (B_TRUE);
9729 	}
9730 
9731 	return (B_FALSE);
9732 }
9733 
9734 uint64_t
9735 spa_total_metaslabs(spa_t *spa)
9736 {
9737 	vdev_t *rvd = spa->spa_root_vdev;
9738 
9739 	uint64_t m = 0;
9740 	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
9741 		vdev_t *vd = rvd->vdev_child[c];
9742 		if (!vdev_is_concrete(vd))
9743 			continue;
9744 		m += vd->vdev_ms_count;
9745 	}
9746 	return (m);
9747 }
9748 
9749 /*
9750  * Notify any waiting threads that some activity has switched from being in-
9751  * progress to not-in-progress so that the thread can wake up and determine
9752  * whether it is finished waiting.
9753  */
9754 void
9755 spa_notify_waiters(spa_t *spa)
9756 {
9757 	/*
9758 	 * Acquiring spa_activities_lock here prevents the cv_broadcast from
9759 	 * happening between the waiting thread's check and cv_wait.
9760 	 */
9761 	mutex_enter(&spa->spa_activities_lock);
9762 	cv_broadcast(&spa->spa_activities_cv);
9763 	mutex_exit(&spa->spa_activities_lock);
9764 }
9765 
9766 /*
9767  * Notify any waiting threads that the pool is exporting, and then block until
9768  * they are finished using the spa_t.
9769  */
9770 void
9771 spa_wake_waiters(spa_t *spa)
9772 {
9773 	mutex_enter(&spa->spa_activities_lock);
9774 	spa->spa_waiters_cancel = B_TRUE;
9775 	cv_broadcast(&spa->spa_activities_cv);
9776 	while (spa->spa_waiters != 0)
9777 		cv_wait(&spa->spa_waiters_cv, &spa->spa_activities_lock);
9778 	spa->spa_waiters_cancel = B_FALSE;
9779 	mutex_exit(&spa->spa_activities_lock);
9780 }
9781 
9782 /* Whether the vdev or any of its descendants are being initialized/trimmed. */
9783 static boolean_t
9784 spa_vdev_activity_in_progress_impl(vdev_t *vd, zpool_wait_activity_t activity)
9785 {
9786 	spa_t *spa = vd->vdev_spa;
9787 
9788 	ASSERT(spa_config_held(spa, SCL_CONFIG | SCL_STATE, RW_READER));
9789 	ASSERT(MUTEX_HELD(&spa->spa_activities_lock));
9790 	ASSERT(activity == ZPOOL_WAIT_INITIALIZE ||
9791 	    activity == ZPOOL_WAIT_TRIM);
9792 
9793 	kmutex_t *lock = activity == ZPOOL_WAIT_INITIALIZE ?
9794 	    &vd->vdev_initialize_lock : &vd->vdev_trim_lock;
9795 
9796 	mutex_exit(&spa->spa_activities_lock);
9797 	mutex_enter(lock);
9798 	mutex_enter(&spa->spa_activities_lock);
9799 
9800 	boolean_t in_progress = (activity == ZPOOL_WAIT_INITIALIZE) ?
9801 	    (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) :
9802 	    (vd->vdev_trim_state == VDEV_TRIM_ACTIVE);
9803 	mutex_exit(lock);
9804 
9805 	if (in_progress)
9806 		return (B_TRUE);
9807 
9808 	for (int i = 0; i < vd->vdev_children; i++) {
9809 		if (spa_vdev_activity_in_progress_impl(vd->vdev_child[i],
9810 		    activity))
9811 			return (B_TRUE);
9812 	}
9813 
9814 	return (B_FALSE);
9815 }
9816 
9817 /*
9818  * If use_guid is true, this checks whether the vdev specified by guid is
9819  * being initialized/trimmed. Otherwise, it checks whether any vdev in the pool
9820  * is being initialized/trimmed. The caller must hold the config lock and
9821  * spa_activities_lock.
9822  */
9823 static int
9824 spa_vdev_activity_in_progress(spa_t *spa, boolean_t use_guid, uint64_t guid,
9825     zpool_wait_activity_t activity, boolean_t *in_progress)
9826 {
9827 	mutex_exit(&spa->spa_activities_lock);
9828 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9829 	mutex_enter(&spa->spa_activities_lock);
9830 
9831 	vdev_t *vd;
9832 	if (use_guid) {
9833 		vd = spa_lookup_by_guid(spa, guid, B_FALSE);
9834 		if (vd == NULL || !vd->vdev_ops->vdev_op_leaf) {
9835 			spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9836 			return (EINVAL);
9837 		}
9838 	} else {
9839 		vd = spa->spa_root_vdev;
9840 	}
9841 
9842 	*in_progress = spa_vdev_activity_in_progress_impl(vd, activity);
9843 
9844 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9845 	return (0);
9846 }
9847 
9848 /*
9849  * Locking for waiting threads
9850  * ---------------------------
9851  *
9852  * Waiting threads need a way to check whether a given activity is in progress,
9853  * and then, if it is, wait for it to complete. Each activity will have some
9854  * in-memory representation of the relevant on-disk state which can be used to
9855  * determine whether or not the activity is in progress. The in-memory state and
9856  * the locking used to protect it will be different for each activity, and may
9857  * not be suitable for use with a cvar (e.g., some state is protected by the
9858  * config lock). To allow waiting threads to wait without any races, another
9859  * lock, spa_activities_lock, is used.
9860  *
9861  * When the state is checked, both the activity-specific lock (if there is one)
9862  * and spa_activities_lock are held. In some cases, the activity-specific lock
9863  * is acquired explicitly (e.g. the config lock). In others, the locking is
9864  * internal to some check (e.g. bpobj_is_empty). After checking, the waiting
9865  * thread releases the activity-specific lock and, if the activity is in
9866  * progress, then cv_waits using spa_activities_lock.
9867  *
9868  * The waiting thread is woken when another thread, one completing some
9869  * activity, updates the state of the activity and then calls
9870  * spa_notify_waiters, which will cv_broadcast. This 'completing' thread only
9871  * needs to hold its activity-specific lock when updating the state, and this
9872  * lock can (but doesn't have to) be dropped before calling spa_notify_waiters.
9873  *
9874  * Because spa_notify_waiters acquires spa_activities_lock before broadcasting,
9875  * and because it is held when the waiting thread checks the state of the
9876  * activity, it can never be the case that the completing thread both updates
9877  * the activity state and cv_broadcasts in between the waiting thread's check
9878  * and cv_wait. Thus, a waiting thread can never miss a wakeup.
9879  *
9880  * In order to prevent deadlock, when the waiting thread does its check, in some
9881  * cases it will temporarily drop spa_activities_lock in order to acquire the
9882  * activity-specific lock. The order in which spa_activities_lock and the
9883  * activity specific lock are acquired in the waiting thread is determined by
9884  * the order in which they are acquired in the completing thread; if the
9885  * completing thread calls spa_notify_waiters with the activity-specific lock
9886  * held, then the waiting thread must also acquire the activity-specific lock
9887  * first.
9888  */
9889 
9890 static int
9891 spa_activity_in_progress(spa_t *spa, zpool_wait_activity_t activity,
9892     boolean_t use_tag, uint64_t tag, boolean_t *in_progress)
9893 {
9894 	int error = 0;
9895 
9896 	ASSERT(MUTEX_HELD(&spa->spa_activities_lock));
9897 
9898 	switch (activity) {
9899 	case ZPOOL_WAIT_CKPT_DISCARD:
9900 		*in_progress =
9901 		    (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT) &&
9902 		    zap_contains(spa_meta_objset(spa),
9903 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT) ==
9904 		    ENOENT);
9905 		break;
9906 	case ZPOOL_WAIT_FREE:
9907 		*in_progress = ((spa_version(spa) >= SPA_VERSION_DEADLISTS &&
9908 		    !bpobj_is_empty(&spa->spa_dsl_pool->dp_free_bpobj)) ||
9909 		    spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY) ||
9910 		    spa_livelist_delete_check(spa));
9911 		break;
9912 	case ZPOOL_WAIT_INITIALIZE:
9913 	case ZPOOL_WAIT_TRIM:
9914 		error = spa_vdev_activity_in_progress(spa, use_tag, tag,
9915 		    activity, in_progress);
9916 		break;
9917 	case ZPOOL_WAIT_REPLACE:
9918 		mutex_exit(&spa->spa_activities_lock);
9919 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9920 		mutex_enter(&spa->spa_activities_lock);
9921 
9922 		*in_progress = vdev_replace_in_progress(spa->spa_root_vdev);
9923 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9924 		break;
9925 	case ZPOOL_WAIT_REMOVE:
9926 		*in_progress = (spa->spa_removing_phys.sr_state ==
9927 		    DSS_SCANNING);
9928 		break;
9929 	case ZPOOL_WAIT_RESILVER:
9930 		if ((*in_progress = vdev_rebuild_active(spa->spa_root_vdev)))
9931 			break;
9932 		zfs_fallthrough;
9933 	case ZPOOL_WAIT_SCRUB:
9934 	{
9935 		boolean_t scanning, paused, is_scrub;
9936 		dsl_scan_t *scn =  spa->spa_dsl_pool->dp_scan;
9937 
9938 		is_scrub = (scn->scn_phys.scn_func == POOL_SCAN_SCRUB);
9939 		scanning = (scn->scn_phys.scn_state == DSS_SCANNING);
9940 		paused = dsl_scan_is_paused_scrub(scn);
9941 		*in_progress = (scanning && !paused &&
9942 		    is_scrub == (activity == ZPOOL_WAIT_SCRUB));
9943 		break;
9944 	}
9945 	default:
9946 		panic("unrecognized value for activity %d", activity);
9947 	}
9948 
9949 	return (error);
9950 }
9951 
9952 static int
9953 spa_wait_common(const char *pool, zpool_wait_activity_t activity,
9954     boolean_t use_tag, uint64_t tag, boolean_t *waited)
9955 {
9956 	/*
9957 	 * The tag is used to distinguish between instances of an activity.
9958 	 * 'initialize' and 'trim' are the only activities that we use this for.
9959 	 * The other activities can only have a single instance in progress in a
9960 	 * pool at one time, making the tag unnecessary.
9961 	 *
9962 	 * There can be multiple devices being replaced at once, but since they
9963 	 * all finish once resilvering finishes, we don't bother keeping track
9964 	 * of them individually, we just wait for them all to finish.
9965 	 */
9966 	if (use_tag && activity != ZPOOL_WAIT_INITIALIZE &&
9967 	    activity != ZPOOL_WAIT_TRIM)
9968 		return (EINVAL);
9969 
9970 	if (activity < 0 || activity >= ZPOOL_WAIT_NUM_ACTIVITIES)
9971 		return (EINVAL);
9972 
9973 	spa_t *spa;
9974 	int error = spa_open(pool, &spa, FTAG);
9975 	if (error != 0)
9976 		return (error);
9977 
9978 	/*
9979 	 * Increment the spa's waiter count so that we can call spa_close and
9980 	 * still ensure that the spa_t doesn't get freed before this thread is
9981 	 * finished with it when the pool is exported. We want to call spa_close
9982 	 * before we start waiting because otherwise the additional ref would
9983 	 * prevent the pool from being exported or destroyed throughout the
9984 	 * potentially long wait.
9985 	 */
9986 	mutex_enter(&spa->spa_activities_lock);
9987 	spa->spa_waiters++;
9988 	spa_close(spa, FTAG);
9989 
9990 	*waited = B_FALSE;
9991 	for (;;) {
9992 		boolean_t in_progress;
9993 		error = spa_activity_in_progress(spa, activity, use_tag, tag,
9994 		    &in_progress);
9995 
9996 		if (error || !in_progress || spa->spa_waiters_cancel)
9997 			break;
9998 
9999 		*waited = B_TRUE;
10000 
10001 		if (cv_wait_sig(&spa->spa_activities_cv,
10002 		    &spa->spa_activities_lock) == 0) {
10003 			error = EINTR;
10004 			break;
10005 		}
10006 	}
10007 
10008 	spa->spa_waiters--;
10009 	cv_signal(&spa->spa_waiters_cv);
10010 	mutex_exit(&spa->spa_activities_lock);
10011 
10012 	return (error);
10013 }
10014 
10015 /*
10016  * Wait for a particular instance of the specified activity to complete, where
10017  * the instance is identified by 'tag'
10018  */
10019 int
10020 spa_wait_tag(const char *pool, zpool_wait_activity_t activity, uint64_t tag,
10021     boolean_t *waited)
10022 {
10023 	return (spa_wait_common(pool, activity, B_TRUE, tag, waited));
10024 }
10025 
10026 /*
10027  * Wait for all instances of the specified activity complete
10028  */
10029 int
10030 spa_wait(const char *pool, zpool_wait_activity_t activity, boolean_t *waited)
10031 {
10032 
10033 	return (spa_wait_common(pool, activity, B_FALSE, 0, waited));
10034 }
10035 
10036 sysevent_t *
10037 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
10038 {
10039 	sysevent_t *ev = NULL;
10040 #ifdef _KERNEL
10041 	nvlist_t *resource;
10042 
10043 	resource = zfs_event_create(spa, vd, FM_SYSEVENT_CLASS, name, hist_nvl);
10044 	if (resource) {
10045 		ev = kmem_alloc(sizeof (sysevent_t), KM_SLEEP);
10046 		ev->resource = resource;
10047 	}
10048 #else
10049 	(void) spa, (void) vd, (void) hist_nvl, (void) name;
10050 #endif
10051 	return (ev);
10052 }
10053 
10054 void
10055 spa_event_post(sysevent_t *ev)
10056 {
10057 #ifdef _KERNEL
10058 	if (ev) {
10059 		zfs_zevent_post(ev->resource, NULL, zfs_zevent_post_cb);
10060 		kmem_free(ev, sizeof (*ev));
10061 	}
10062 #else
10063 	(void) ev;
10064 #endif
10065 }
10066 
10067 /*
10068  * Post a zevent corresponding to the given sysevent.   The 'name' must be one
10069  * of the event definitions in sys/sysevent/eventdefs.h.  The payload will be
10070  * filled in from the spa and (optionally) the vdev.  This doesn't do anything
10071  * in the userland libzpool, as we don't want consumers to misinterpret ztest
10072  * or zdb as real changes.
10073  */
10074 void
10075 spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
10076 {
10077 	spa_event_post(spa_event_create(spa, vd, hist_nvl, name));
10078 }
10079 
10080 /* state manipulation functions */
10081 EXPORT_SYMBOL(spa_open);
10082 EXPORT_SYMBOL(spa_open_rewind);
10083 EXPORT_SYMBOL(spa_get_stats);
10084 EXPORT_SYMBOL(spa_create);
10085 EXPORT_SYMBOL(spa_import);
10086 EXPORT_SYMBOL(spa_tryimport);
10087 EXPORT_SYMBOL(spa_destroy);
10088 EXPORT_SYMBOL(spa_export);
10089 EXPORT_SYMBOL(spa_reset);
10090 EXPORT_SYMBOL(spa_async_request);
10091 EXPORT_SYMBOL(spa_async_suspend);
10092 EXPORT_SYMBOL(spa_async_resume);
10093 EXPORT_SYMBOL(spa_inject_addref);
10094 EXPORT_SYMBOL(spa_inject_delref);
10095 EXPORT_SYMBOL(spa_scan_stat_init);
10096 EXPORT_SYMBOL(spa_scan_get_stats);
10097 
10098 /* device manipulation */
10099 EXPORT_SYMBOL(spa_vdev_add);
10100 EXPORT_SYMBOL(spa_vdev_attach);
10101 EXPORT_SYMBOL(spa_vdev_detach);
10102 EXPORT_SYMBOL(spa_vdev_setpath);
10103 EXPORT_SYMBOL(spa_vdev_setfru);
10104 EXPORT_SYMBOL(spa_vdev_split_mirror);
10105 
10106 /* spare statech is global across all pools) */
10107 EXPORT_SYMBOL(spa_spare_add);
10108 EXPORT_SYMBOL(spa_spare_remove);
10109 EXPORT_SYMBOL(spa_spare_exists);
10110 EXPORT_SYMBOL(spa_spare_activate);
10111 
10112 /* L2ARC statech is global across all pools) */
10113 EXPORT_SYMBOL(spa_l2cache_add);
10114 EXPORT_SYMBOL(spa_l2cache_remove);
10115 EXPORT_SYMBOL(spa_l2cache_exists);
10116 EXPORT_SYMBOL(spa_l2cache_activate);
10117 EXPORT_SYMBOL(spa_l2cache_drop);
10118 
10119 /* scanning */
10120 EXPORT_SYMBOL(spa_scan);
10121 EXPORT_SYMBOL(spa_scan_stop);
10122 
10123 /* spa syncing */
10124 EXPORT_SYMBOL(spa_sync); /* only for DMU use */
10125 EXPORT_SYMBOL(spa_sync_allpools);
10126 
10127 /* properties */
10128 EXPORT_SYMBOL(spa_prop_set);
10129 EXPORT_SYMBOL(spa_prop_get);
10130 EXPORT_SYMBOL(spa_prop_clear_bootfs);
10131 
10132 /* asynchronous event notification */
10133 EXPORT_SYMBOL(spa_event_notify);
10134 
10135 /* BEGIN CSTYLED */
10136 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_shift, UINT, ZMOD_RW,
10137 	"log2 fraction of arc that can be used by inflight I/Os when "
10138 	"verifying pool during import");
10139 /* END CSTYLED */
10140 
10141 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_metadata, INT, ZMOD_RW,
10142 	"Set to traverse metadata on pool import");
10143 
10144 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW,
10145 	"Set to traverse data on pool import");
10146 
10147 ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW,
10148 	"Print vdev tree to zfs_dbgmsg during pool import");
10149 
10150 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD,
10151 	"Percentage of CPUs to run an IO worker thread");
10152 
10153 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_tpq, UINT, ZMOD_RD,
10154 	"Number of threads per IO worker taskqueue");
10155 
10156 /* BEGIN CSTYLED */
10157 ZFS_MODULE_PARAM(zfs, zfs_, max_missing_tvds, U64, ZMOD_RW,
10158 	"Allow importing pool with up to this number of missing top-level "
10159 	"vdevs (in read-only mode)");
10160 /* END CSTYLED */
10161 
10162 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_pause, INT,
10163 	ZMOD_RW, "Set the livelist condense zthr to pause");
10164 
10165 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_pause, INT,
10166 	ZMOD_RW, "Set the livelist condense synctask to pause");
10167 
10168 /* BEGIN CSTYLED */
10169 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_cancel,
10170 	INT, ZMOD_RW,
10171 	"Whether livelist condensing was canceled in the synctask");
10172 
10173 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_cancel,
10174 	INT, ZMOD_RW,
10175 	"Whether livelist condensing was canceled in the zthr function");
10176 
10177 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT,
10178 	ZMOD_RW,
10179 	"Whether extra ALLOC blkptrs were added to a livelist entry while it "
10180 	"was being condensed");
10181 /* END CSTYLED */
10182