xref: /freebsd/sys/contrib/openzfs/module/zfs/spa.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
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 extern metaslab_ops_t *metaslab_allocator(spa_t *spa);
1299 
1300 /*
1301  * Activate an uninitialized pool.
1302  */
1303 static void
1304 spa_activate(spa_t *spa, spa_mode_t mode)
1305 {
1306 	metaslab_ops_t *msp = metaslab_allocator(spa);
1307 	ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED);
1308 
1309 	spa->spa_state = POOL_STATE_ACTIVE;
1310 	spa->spa_mode = mode;
1311 	spa->spa_read_spacemaps = spa_mode_readable_spacemaps;
1312 
1313 	spa->spa_normal_class = metaslab_class_create(spa, msp);
1314 	spa->spa_log_class = metaslab_class_create(spa, msp);
1315 	spa->spa_embedded_log_class = metaslab_class_create(spa, msp);
1316 	spa->spa_special_class = metaslab_class_create(spa, msp);
1317 	spa->spa_dedup_class = metaslab_class_create(spa, msp);
1318 
1319 	/* Try to create a covering process */
1320 	mutex_enter(&spa->spa_proc_lock);
1321 	ASSERT(spa->spa_proc_state == SPA_PROC_NONE);
1322 	ASSERT(spa->spa_proc == &p0);
1323 	spa->spa_did = 0;
1324 
1325 	(void) spa_create_process;
1326 #ifdef HAVE_SPA_THREAD
1327 	/* Only create a process if we're going to be around a while. */
1328 	if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) {
1329 		if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri,
1330 		    NULL, 0) == 0) {
1331 			spa->spa_proc_state = SPA_PROC_CREATED;
1332 			while (spa->spa_proc_state == SPA_PROC_CREATED) {
1333 				cv_wait(&spa->spa_proc_cv,
1334 				    &spa->spa_proc_lock);
1335 			}
1336 			ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1337 			ASSERT(spa->spa_proc != &p0);
1338 			ASSERT(spa->spa_did != 0);
1339 		} else {
1340 #ifdef _KERNEL
1341 			cmn_err(CE_WARN,
1342 			    "Couldn't create process for zfs pool \"%s\"\n",
1343 			    spa->spa_name);
1344 #endif
1345 		}
1346 	}
1347 #endif /* HAVE_SPA_THREAD */
1348 	mutex_exit(&spa->spa_proc_lock);
1349 
1350 	/* If we didn't create a process, we need to create our taskqs. */
1351 	if (spa->spa_proc == &p0) {
1352 		spa_create_zio_taskqs(spa);
1353 	}
1354 
1355 	for (size_t i = 0; i < TXG_SIZE; i++) {
1356 		spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL,
1357 		    ZIO_FLAG_CANFAIL);
1358 	}
1359 
1360 	list_create(&spa->spa_config_dirty_list, sizeof (vdev_t),
1361 	    offsetof(vdev_t, vdev_config_dirty_node));
1362 	list_create(&spa->spa_evicting_os_list, sizeof (objset_t),
1363 	    offsetof(objset_t, os_evicting_node));
1364 	list_create(&spa->spa_state_dirty_list, sizeof (vdev_t),
1365 	    offsetof(vdev_t, vdev_state_dirty_node));
1366 
1367 	txg_list_create(&spa->spa_vdev_txg_list, spa,
1368 	    offsetof(struct vdev, vdev_txg_node));
1369 
1370 	avl_create(&spa->spa_errlist_scrub,
1371 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1372 	    offsetof(spa_error_entry_t, se_avl));
1373 	avl_create(&spa->spa_errlist_last,
1374 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1375 	    offsetof(spa_error_entry_t, se_avl));
1376 	avl_create(&spa->spa_errlist_healed,
1377 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1378 	    offsetof(spa_error_entry_t, se_avl));
1379 
1380 	spa_activate_os(spa);
1381 
1382 	spa_keystore_init(&spa->spa_keystore);
1383 
1384 	/*
1385 	 * This taskq is used to perform zvol-minor-related tasks
1386 	 * asynchronously. This has several advantages, including easy
1387 	 * resolution of various deadlocks.
1388 	 *
1389 	 * The taskq must be single threaded to ensure tasks are always
1390 	 * processed in the order in which they were dispatched.
1391 	 *
1392 	 * A taskq per pool allows one to keep the pools independent.
1393 	 * This way if one pool is suspended, it will not impact another.
1394 	 *
1395 	 * The preferred location to dispatch a zvol minor task is a sync
1396 	 * task. In this context, there is easy access to the spa_t and minimal
1397 	 * error handling is required because the sync task must succeed.
1398 	 */
1399 	spa->spa_zvol_taskq = taskq_create("z_zvol", 1, defclsyspri,
1400 	    1, INT_MAX, 0);
1401 
1402 	/*
1403 	 * Taskq dedicated to prefetcher threads: this is used to prevent the
1404 	 * pool traverse code from monopolizing the global (and limited)
1405 	 * system_taskq by inappropriately scheduling long running tasks on it.
1406 	 */
1407 	spa->spa_prefetch_taskq = taskq_create("z_prefetch", 100,
1408 	    defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT);
1409 
1410 	/*
1411 	 * The taskq to upgrade datasets in this pool. Currently used by
1412 	 * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA.
1413 	 */
1414 	spa->spa_upgrade_taskq = taskq_create("z_upgrade", 100,
1415 	    defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT);
1416 }
1417 
1418 /*
1419  * Opposite of spa_activate().
1420  */
1421 static void
1422 spa_deactivate(spa_t *spa)
1423 {
1424 	ASSERT(spa->spa_sync_on == B_FALSE);
1425 	ASSERT(spa->spa_dsl_pool == NULL);
1426 	ASSERT(spa->spa_root_vdev == NULL);
1427 	ASSERT(spa->spa_async_zio_root == NULL);
1428 	ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED);
1429 
1430 	spa_evicting_os_wait(spa);
1431 
1432 	if (spa->spa_zvol_taskq) {
1433 		taskq_destroy(spa->spa_zvol_taskq);
1434 		spa->spa_zvol_taskq = NULL;
1435 	}
1436 
1437 	if (spa->spa_prefetch_taskq) {
1438 		taskq_destroy(spa->spa_prefetch_taskq);
1439 		spa->spa_prefetch_taskq = NULL;
1440 	}
1441 
1442 	if (spa->spa_upgrade_taskq) {
1443 		taskq_destroy(spa->spa_upgrade_taskq);
1444 		spa->spa_upgrade_taskq = NULL;
1445 	}
1446 
1447 	txg_list_destroy(&spa->spa_vdev_txg_list);
1448 
1449 	list_destroy(&spa->spa_config_dirty_list);
1450 	list_destroy(&spa->spa_evicting_os_list);
1451 	list_destroy(&spa->spa_state_dirty_list);
1452 
1453 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
1454 
1455 	for (int t = 0; t < ZIO_TYPES; t++) {
1456 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1457 			spa_taskqs_fini(spa, t, q);
1458 		}
1459 	}
1460 
1461 	for (size_t i = 0; i < TXG_SIZE; i++) {
1462 		ASSERT3P(spa->spa_txg_zio[i], !=, NULL);
1463 		VERIFY0(zio_wait(spa->spa_txg_zio[i]));
1464 		spa->spa_txg_zio[i] = NULL;
1465 	}
1466 
1467 	metaslab_class_destroy(spa->spa_normal_class);
1468 	spa->spa_normal_class = NULL;
1469 
1470 	metaslab_class_destroy(spa->spa_log_class);
1471 	spa->spa_log_class = NULL;
1472 
1473 	metaslab_class_destroy(spa->spa_embedded_log_class);
1474 	spa->spa_embedded_log_class = NULL;
1475 
1476 	metaslab_class_destroy(spa->spa_special_class);
1477 	spa->spa_special_class = NULL;
1478 
1479 	metaslab_class_destroy(spa->spa_dedup_class);
1480 	spa->spa_dedup_class = NULL;
1481 
1482 	/*
1483 	 * If this was part of an import or the open otherwise failed, we may
1484 	 * still have errors left in the queues.  Empty them just in case.
1485 	 */
1486 	spa_errlog_drain(spa);
1487 	avl_destroy(&spa->spa_errlist_scrub);
1488 	avl_destroy(&spa->spa_errlist_last);
1489 	avl_destroy(&spa->spa_errlist_healed);
1490 
1491 	spa_keystore_fini(&spa->spa_keystore);
1492 
1493 	spa->spa_state = POOL_STATE_UNINITIALIZED;
1494 
1495 	mutex_enter(&spa->spa_proc_lock);
1496 	if (spa->spa_proc_state != SPA_PROC_NONE) {
1497 		ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1498 		spa->spa_proc_state = SPA_PROC_DEACTIVATE;
1499 		cv_broadcast(&spa->spa_proc_cv);
1500 		while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) {
1501 			ASSERT(spa->spa_proc != &p0);
1502 			cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1503 		}
1504 		ASSERT(spa->spa_proc_state == SPA_PROC_GONE);
1505 		spa->spa_proc_state = SPA_PROC_NONE;
1506 	}
1507 	ASSERT(spa->spa_proc == &p0);
1508 	mutex_exit(&spa->spa_proc_lock);
1509 
1510 	/*
1511 	 * We want to make sure spa_thread() has actually exited the ZFS
1512 	 * module, so that the module can't be unloaded out from underneath
1513 	 * it.
1514 	 */
1515 	if (spa->spa_did != 0) {
1516 		thread_join(spa->spa_did);
1517 		spa->spa_did = 0;
1518 	}
1519 
1520 	spa_deactivate_os(spa);
1521 
1522 }
1523 
1524 /*
1525  * Verify a pool configuration, and construct the vdev tree appropriately.  This
1526  * will create all the necessary vdevs in the appropriate layout, with each vdev
1527  * in the CLOSED state.  This will prep the pool before open/creation/import.
1528  * All vdev validation is done by the vdev_alloc() routine.
1529  */
1530 int
1531 spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent,
1532     uint_t id, int atype)
1533 {
1534 	nvlist_t **child;
1535 	uint_t children;
1536 	int error;
1537 
1538 	if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0)
1539 		return (error);
1540 
1541 	if ((*vdp)->vdev_ops->vdev_op_leaf)
1542 		return (0);
1543 
1544 	error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1545 	    &child, &children);
1546 
1547 	if (error == ENOENT)
1548 		return (0);
1549 
1550 	if (error) {
1551 		vdev_free(*vdp);
1552 		*vdp = NULL;
1553 		return (SET_ERROR(EINVAL));
1554 	}
1555 
1556 	for (int c = 0; c < children; c++) {
1557 		vdev_t *vd;
1558 		if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c,
1559 		    atype)) != 0) {
1560 			vdev_free(*vdp);
1561 			*vdp = NULL;
1562 			return (error);
1563 		}
1564 	}
1565 
1566 	ASSERT(*vdp != NULL);
1567 
1568 	return (0);
1569 }
1570 
1571 static boolean_t
1572 spa_should_flush_logs_on_unload(spa_t *spa)
1573 {
1574 	if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))
1575 		return (B_FALSE);
1576 
1577 	if (!spa_writeable(spa))
1578 		return (B_FALSE);
1579 
1580 	if (!spa->spa_sync_on)
1581 		return (B_FALSE);
1582 
1583 	if (spa_state(spa) != POOL_STATE_EXPORTED)
1584 		return (B_FALSE);
1585 
1586 	if (zfs_keep_log_spacemaps_at_export)
1587 		return (B_FALSE);
1588 
1589 	return (B_TRUE);
1590 }
1591 
1592 /*
1593  * Opens a transaction that will set the flag that will instruct
1594  * spa_sync to attempt to flush all the metaslabs for that txg.
1595  */
1596 static void
1597 spa_unload_log_sm_flush_all(spa_t *spa)
1598 {
1599 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1600 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1601 
1602 	ASSERT3U(spa->spa_log_flushall_txg, ==, 0);
1603 	spa->spa_log_flushall_txg = dmu_tx_get_txg(tx);
1604 
1605 	dmu_tx_commit(tx);
1606 	txg_wait_synced(spa_get_dsl(spa), spa->spa_log_flushall_txg);
1607 }
1608 
1609 static void
1610 spa_unload_log_sm_metadata(spa_t *spa)
1611 {
1612 	void *cookie = NULL;
1613 	spa_log_sm_t *sls;
1614 	log_summary_entry_t *e;
1615 
1616 	while ((sls = avl_destroy_nodes(&spa->spa_sm_logs_by_txg,
1617 	    &cookie)) != NULL) {
1618 		VERIFY0(sls->sls_mscount);
1619 		kmem_free(sls, sizeof (spa_log_sm_t));
1620 	}
1621 
1622 	while ((e = list_remove_head(&spa->spa_log_summary)) != NULL) {
1623 		VERIFY0(e->lse_mscount);
1624 		kmem_free(e, sizeof (log_summary_entry_t));
1625 	}
1626 
1627 	spa->spa_unflushed_stats.sus_nblocks = 0;
1628 	spa->spa_unflushed_stats.sus_memused = 0;
1629 	spa->spa_unflushed_stats.sus_blocklimit = 0;
1630 }
1631 
1632 static void
1633 spa_destroy_aux_threads(spa_t *spa)
1634 {
1635 	if (spa->spa_condense_zthr != NULL) {
1636 		zthr_destroy(spa->spa_condense_zthr);
1637 		spa->spa_condense_zthr = NULL;
1638 	}
1639 	if (spa->spa_checkpoint_discard_zthr != NULL) {
1640 		zthr_destroy(spa->spa_checkpoint_discard_zthr);
1641 		spa->spa_checkpoint_discard_zthr = NULL;
1642 	}
1643 	if (spa->spa_livelist_delete_zthr != NULL) {
1644 		zthr_destroy(spa->spa_livelist_delete_zthr);
1645 		spa->spa_livelist_delete_zthr = NULL;
1646 	}
1647 	if (spa->spa_livelist_condense_zthr != NULL) {
1648 		zthr_destroy(spa->spa_livelist_condense_zthr);
1649 		spa->spa_livelist_condense_zthr = NULL;
1650 	}
1651 }
1652 
1653 /*
1654  * Opposite of spa_load().
1655  */
1656 static void
1657 spa_unload(spa_t *spa)
1658 {
1659 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1660 	ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED);
1661 
1662 	spa_import_progress_remove(spa_guid(spa));
1663 	spa_load_note(spa, "UNLOADING");
1664 
1665 	spa_wake_waiters(spa);
1666 
1667 	/*
1668 	 * If we have set the spa_final_txg, we have already performed the
1669 	 * tasks below in spa_export_common(). We should not redo it here since
1670 	 * we delay the final TXGs beyond what spa_final_txg is set at.
1671 	 */
1672 	if (spa->spa_final_txg == UINT64_MAX) {
1673 		/*
1674 		 * If the log space map feature is enabled and the pool is
1675 		 * getting exported (but not destroyed), we want to spend some
1676 		 * time flushing as many metaslabs as we can in an attempt to
1677 		 * destroy log space maps and save import time.
1678 		 */
1679 		if (spa_should_flush_logs_on_unload(spa))
1680 			spa_unload_log_sm_flush_all(spa);
1681 
1682 		/*
1683 		 * Stop async tasks.
1684 		 */
1685 		spa_async_suspend(spa);
1686 
1687 		if (spa->spa_root_vdev) {
1688 			vdev_t *root_vdev = spa->spa_root_vdev;
1689 			vdev_initialize_stop_all(root_vdev,
1690 			    VDEV_INITIALIZE_ACTIVE);
1691 			vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE);
1692 			vdev_autotrim_stop_all(spa);
1693 			vdev_rebuild_stop_all(spa);
1694 		}
1695 	}
1696 
1697 	/*
1698 	 * Stop syncing.
1699 	 */
1700 	if (spa->spa_sync_on) {
1701 		txg_sync_stop(spa->spa_dsl_pool);
1702 		spa->spa_sync_on = B_FALSE;
1703 	}
1704 
1705 	/*
1706 	 * This ensures that there is no async metaslab prefetching
1707 	 * while we attempt to unload the spa.
1708 	 */
1709 	if (spa->spa_root_vdev != NULL) {
1710 		for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++) {
1711 			vdev_t *vc = spa->spa_root_vdev->vdev_child[c];
1712 			if (vc->vdev_mg != NULL)
1713 				taskq_wait(vc->vdev_mg->mg_taskq);
1714 		}
1715 	}
1716 
1717 	if (spa->spa_mmp.mmp_thread)
1718 		mmp_thread_stop(spa);
1719 
1720 	/*
1721 	 * Wait for any outstanding async I/O to complete.
1722 	 */
1723 	if (spa->spa_async_zio_root != NULL) {
1724 		for (int i = 0; i < max_ncpus; i++)
1725 			(void) zio_wait(spa->spa_async_zio_root[i]);
1726 		kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *));
1727 		spa->spa_async_zio_root = NULL;
1728 	}
1729 
1730 	if (spa->spa_vdev_removal != NULL) {
1731 		spa_vdev_removal_destroy(spa->spa_vdev_removal);
1732 		spa->spa_vdev_removal = NULL;
1733 	}
1734 
1735 	spa_destroy_aux_threads(spa);
1736 
1737 	spa_condense_fini(spa);
1738 
1739 	bpobj_close(&spa->spa_deferred_bpobj);
1740 
1741 	spa_config_enter(spa, SCL_ALL, spa, RW_WRITER);
1742 
1743 	/*
1744 	 * Close all vdevs.
1745 	 */
1746 	if (spa->spa_root_vdev)
1747 		vdev_free(spa->spa_root_vdev);
1748 	ASSERT(spa->spa_root_vdev == NULL);
1749 
1750 	/*
1751 	 * Close the dsl pool.
1752 	 */
1753 	if (spa->spa_dsl_pool) {
1754 		dsl_pool_close(spa->spa_dsl_pool);
1755 		spa->spa_dsl_pool = NULL;
1756 		spa->spa_meta_objset = NULL;
1757 	}
1758 
1759 	ddt_unload(spa);
1760 	brt_unload(spa);
1761 	spa_unload_log_sm_metadata(spa);
1762 
1763 	/*
1764 	 * Drop and purge level 2 cache
1765 	 */
1766 	spa_l2cache_drop(spa);
1767 
1768 	if (spa->spa_spares.sav_vdevs) {
1769 		for (int i = 0; i < spa->spa_spares.sav_count; i++)
1770 			vdev_free(spa->spa_spares.sav_vdevs[i]);
1771 		kmem_free(spa->spa_spares.sav_vdevs,
1772 		    spa->spa_spares.sav_count * sizeof (void *));
1773 		spa->spa_spares.sav_vdevs = NULL;
1774 	}
1775 	if (spa->spa_spares.sav_config) {
1776 		nvlist_free(spa->spa_spares.sav_config);
1777 		spa->spa_spares.sav_config = NULL;
1778 	}
1779 	spa->spa_spares.sav_count = 0;
1780 
1781 	if (spa->spa_l2cache.sav_vdevs) {
1782 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
1783 			vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]);
1784 			vdev_free(spa->spa_l2cache.sav_vdevs[i]);
1785 		}
1786 		kmem_free(spa->spa_l2cache.sav_vdevs,
1787 		    spa->spa_l2cache.sav_count * sizeof (void *));
1788 		spa->spa_l2cache.sav_vdevs = NULL;
1789 	}
1790 	if (spa->spa_l2cache.sav_config) {
1791 		nvlist_free(spa->spa_l2cache.sav_config);
1792 		spa->spa_l2cache.sav_config = NULL;
1793 	}
1794 	spa->spa_l2cache.sav_count = 0;
1795 
1796 	spa->spa_async_suspended = 0;
1797 
1798 	spa->spa_indirect_vdevs_loaded = B_FALSE;
1799 
1800 	if (spa->spa_comment != NULL) {
1801 		spa_strfree(spa->spa_comment);
1802 		spa->spa_comment = NULL;
1803 	}
1804 	if (spa->spa_compatibility != NULL) {
1805 		spa_strfree(spa->spa_compatibility);
1806 		spa->spa_compatibility = NULL;
1807 	}
1808 
1809 	spa_config_exit(spa, SCL_ALL, spa);
1810 }
1811 
1812 /*
1813  * Load (or re-load) the current list of vdevs describing the active spares for
1814  * this pool.  When this is called, we have some form of basic information in
1815  * 'spa_spares.sav_config'.  We parse this into vdevs, try to open them, and
1816  * then re-generate a more complete list including status information.
1817  */
1818 void
1819 spa_load_spares(spa_t *spa)
1820 {
1821 	nvlist_t **spares;
1822 	uint_t nspares;
1823 	int i;
1824 	vdev_t *vd, *tvd;
1825 
1826 #ifndef _KERNEL
1827 	/*
1828 	 * zdb opens both the current state of the pool and the
1829 	 * checkpointed state (if present), with a different spa_t.
1830 	 *
1831 	 * As spare vdevs are shared among open pools, we skip loading
1832 	 * them when we load the checkpointed state of the pool.
1833 	 */
1834 	if (!spa_writeable(spa))
1835 		return;
1836 #endif
1837 
1838 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1839 
1840 	/*
1841 	 * First, close and free any existing spare vdevs.
1842 	 */
1843 	if (spa->spa_spares.sav_vdevs) {
1844 		for (i = 0; i < spa->spa_spares.sav_count; i++) {
1845 			vd = spa->spa_spares.sav_vdevs[i];
1846 
1847 			/* Undo the call to spa_activate() below */
1848 			if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1849 			    B_FALSE)) != NULL && tvd->vdev_isspare)
1850 				spa_spare_remove(tvd);
1851 			vdev_close(vd);
1852 			vdev_free(vd);
1853 		}
1854 
1855 		kmem_free(spa->spa_spares.sav_vdevs,
1856 		    spa->spa_spares.sav_count * sizeof (void *));
1857 	}
1858 
1859 	if (spa->spa_spares.sav_config == NULL)
1860 		nspares = 0;
1861 	else
1862 		VERIFY0(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
1863 		    ZPOOL_CONFIG_SPARES, &spares, &nspares));
1864 
1865 	spa->spa_spares.sav_count = (int)nspares;
1866 	spa->spa_spares.sav_vdevs = NULL;
1867 
1868 	if (nspares == 0)
1869 		return;
1870 
1871 	/*
1872 	 * Construct the array of vdevs, opening them to get status in the
1873 	 * process.   For each spare, there is potentially two different vdev_t
1874 	 * structures associated with it: one in the list of spares (used only
1875 	 * for basic validation purposes) and one in the active vdev
1876 	 * configuration (if it's spared in).  During this phase we open and
1877 	 * validate each vdev on the spare list.  If the vdev also exists in the
1878 	 * active configuration, then we also mark this vdev as an active spare.
1879 	 */
1880 	spa->spa_spares.sav_vdevs = kmem_zalloc(nspares * sizeof (void *),
1881 	    KM_SLEEP);
1882 	for (i = 0; i < spa->spa_spares.sav_count; i++) {
1883 		VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0,
1884 		    VDEV_ALLOC_SPARE) == 0);
1885 		ASSERT(vd != NULL);
1886 
1887 		spa->spa_spares.sav_vdevs[i] = vd;
1888 
1889 		if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1890 		    B_FALSE)) != NULL) {
1891 			if (!tvd->vdev_isspare)
1892 				spa_spare_add(tvd);
1893 
1894 			/*
1895 			 * We only mark the spare active if we were successfully
1896 			 * able to load the vdev.  Otherwise, importing a pool
1897 			 * with a bad active spare would result in strange
1898 			 * behavior, because multiple pool would think the spare
1899 			 * is actively in use.
1900 			 *
1901 			 * There is a vulnerability here to an equally bizarre
1902 			 * circumstance, where a dead active spare is later
1903 			 * brought back to life (onlined or otherwise).  Given
1904 			 * the rarity of this scenario, and the extra complexity
1905 			 * it adds, we ignore the possibility.
1906 			 */
1907 			if (!vdev_is_dead(tvd))
1908 				spa_spare_activate(tvd);
1909 		}
1910 
1911 		vd->vdev_top = vd;
1912 		vd->vdev_aux = &spa->spa_spares;
1913 
1914 		if (vdev_open(vd) != 0)
1915 			continue;
1916 
1917 		if (vdev_validate_aux(vd) == 0)
1918 			spa_spare_add(vd);
1919 	}
1920 
1921 	/*
1922 	 * Recompute the stashed list of spares, with status information
1923 	 * this time.
1924 	 */
1925 	fnvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES);
1926 
1927 	spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *),
1928 	    KM_SLEEP);
1929 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1930 		spares[i] = vdev_config_generate(spa,
1931 		    spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE);
1932 	fnvlist_add_nvlist_array(spa->spa_spares.sav_config,
1933 	    ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares,
1934 	    spa->spa_spares.sav_count);
1935 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1936 		nvlist_free(spares[i]);
1937 	kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *));
1938 }
1939 
1940 /*
1941  * Load (or re-load) the current list of vdevs describing the active l2cache for
1942  * this pool.  When this is called, we have some form of basic information in
1943  * 'spa_l2cache.sav_config'.  We parse this into vdevs, try to open them, and
1944  * then re-generate a more complete list including status information.
1945  * Devices which are already active have their details maintained, and are
1946  * not re-opened.
1947  */
1948 void
1949 spa_load_l2cache(spa_t *spa)
1950 {
1951 	nvlist_t **l2cache = NULL;
1952 	uint_t nl2cache;
1953 	int i, j, oldnvdevs;
1954 	uint64_t guid;
1955 	vdev_t *vd, **oldvdevs, **newvdevs;
1956 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
1957 
1958 #ifndef _KERNEL
1959 	/*
1960 	 * zdb opens both the current state of the pool and the
1961 	 * checkpointed state (if present), with a different spa_t.
1962 	 *
1963 	 * As L2 caches are part of the ARC which is shared among open
1964 	 * pools, we skip loading them when we load the checkpointed
1965 	 * state of the pool.
1966 	 */
1967 	if (!spa_writeable(spa))
1968 		return;
1969 #endif
1970 
1971 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1972 
1973 	oldvdevs = sav->sav_vdevs;
1974 	oldnvdevs = sav->sav_count;
1975 	sav->sav_vdevs = NULL;
1976 	sav->sav_count = 0;
1977 
1978 	if (sav->sav_config == NULL) {
1979 		nl2cache = 0;
1980 		newvdevs = NULL;
1981 		goto out;
1982 	}
1983 
1984 	VERIFY0(nvlist_lookup_nvlist_array(sav->sav_config,
1985 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache));
1986 	newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP);
1987 
1988 	/*
1989 	 * Process new nvlist of vdevs.
1990 	 */
1991 	for (i = 0; i < nl2cache; i++) {
1992 		guid = fnvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID);
1993 
1994 		newvdevs[i] = NULL;
1995 		for (j = 0; j < oldnvdevs; j++) {
1996 			vd = oldvdevs[j];
1997 			if (vd != NULL && guid == vd->vdev_guid) {
1998 				/*
1999 				 * Retain previous vdev for add/remove ops.
2000 				 */
2001 				newvdevs[i] = vd;
2002 				oldvdevs[j] = NULL;
2003 				break;
2004 			}
2005 		}
2006 
2007 		if (newvdevs[i] == NULL) {
2008 			/*
2009 			 * Create new vdev
2010 			 */
2011 			VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0,
2012 			    VDEV_ALLOC_L2CACHE) == 0);
2013 			ASSERT(vd != NULL);
2014 			newvdevs[i] = vd;
2015 
2016 			/*
2017 			 * Commit this vdev as an l2cache device,
2018 			 * even if it fails to open.
2019 			 */
2020 			spa_l2cache_add(vd);
2021 
2022 			vd->vdev_top = vd;
2023 			vd->vdev_aux = sav;
2024 
2025 			spa_l2cache_activate(vd);
2026 
2027 			if (vdev_open(vd) != 0)
2028 				continue;
2029 
2030 			(void) vdev_validate_aux(vd);
2031 
2032 			if (!vdev_is_dead(vd))
2033 				l2arc_add_vdev(spa, vd);
2034 
2035 			/*
2036 			 * Upon cache device addition to a pool or pool
2037 			 * creation with a cache device or if the header
2038 			 * of the device is invalid we issue an async
2039 			 * TRIM command for the whole device which will
2040 			 * execute if l2arc_trim_ahead > 0.
2041 			 */
2042 			spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM);
2043 		}
2044 	}
2045 
2046 	sav->sav_vdevs = newvdevs;
2047 	sav->sav_count = (int)nl2cache;
2048 
2049 	/*
2050 	 * Recompute the stashed list of l2cache devices, with status
2051 	 * information this time.
2052 	 */
2053 	fnvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE);
2054 
2055 	if (sav->sav_count > 0)
2056 		l2cache = kmem_alloc(sav->sav_count * sizeof (void *),
2057 		    KM_SLEEP);
2058 	for (i = 0; i < sav->sav_count; i++)
2059 		l2cache[i] = vdev_config_generate(spa,
2060 		    sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE);
2061 	fnvlist_add_nvlist_array(sav->sav_config, ZPOOL_CONFIG_L2CACHE,
2062 	    (const nvlist_t * const *)l2cache, sav->sav_count);
2063 
2064 out:
2065 	/*
2066 	 * Purge vdevs that were dropped
2067 	 */
2068 	if (oldvdevs) {
2069 		for (i = 0; i < oldnvdevs; i++) {
2070 			uint64_t pool;
2071 
2072 			vd = oldvdevs[i];
2073 			if (vd != NULL) {
2074 				ASSERT(vd->vdev_isl2cache);
2075 
2076 				if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
2077 				    pool != 0ULL && l2arc_vdev_present(vd))
2078 					l2arc_remove_vdev(vd);
2079 				vdev_clear_stats(vd);
2080 				vdev_free(vd);
2081 			}
2082 		}
2083 
2084 		kmem_free(oldvdevs, oldnvdevs * sizeof (void *));
2085 	}
2086 
2087 	for (i = 0; i < sav->sav_count; i++)
2088 		nvlist_free(l2cache[i]);
2089 	if (sav->sav_count)
2090 		kmem_free(l2cache, sav->sav_count * sizeof (void *));
2091 }
2092 
2093 static int
2094 load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value)
2095 {
2096 	dmu_buf_t *db;
2097 	char *packed = NULL;
2098 	size_t nvsize = 0;
2099 	int error;
2100 	*value = NULL;
2101 
2102 	error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db);
2103 	if (error)
2104 		return (error);
2105 
2106 	nvsize = *(uint64_t *)db->db_data;
2107 	dmu_buf_rele(db, FTAG);
2108 
2109 	packed = vmem_alloc(nvsize, KM_SLEEP);
2110 	error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed,
2111 	    DMU_READ_PREFETCH);
2112 	if (error == 0)
2113 		error = nvlist_unpack(packed, nvsize, value, 0);
2114 	vmem_free(packed, nvsize);
2115 
2116 	return (error);
2117 }
2118 
2119 /*
2120  * Concrete top-level vdevs that are not missing and are not logs. At every
2121  * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds.
2122  */
2123 static uint64_t
2124 spa_healthy_core_tvds(spa_t *spa)
2125 {
2126 	vdev_t *rvd = spa->spa_root_vdev;
2127 	uint64_t tvds = 0;
2128 
2129 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
2130 		vdev_t *vd = rvd->vdev_child[i];
2131 		if (vd->vdev_islog)
2132 			continue;
2133 		if (vdev_is_concrete(vd) && !vdev_is_dead(vd))
2134 			tvds++;
2135 	}
2136 
2137 	return (tvds);
2138 }
2139 
2140 /*
2141  * Checks to see if the given vdev could not be opened, in which case we post a
2142  * sysevent to notify the autoreplace code that the device has been removed.
2143  */
2144 static void
2145 spa_check_removed(vdev_t *vd)
2146 {
2147 	for (uint64_t c = 0; c < vd->vdev_children; c++)
2148 		spa_check_removed(vd->vdev_child[c]);
2149 
2150 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) &&
2151 	    vdev_is_concrete(vd)) {
2152 		zfs_post_autoreplace(vd->vdev_spa, vd);
2153 		spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK);
2154 	}
2155 }
2156 
2157 static int
2158 spa_check_for_missing_logs(spa_t *spa)
2159 {
2160 	vdev_t *rvd = spa->spa_root_vdev;
2161 
2162 	/*
2163 	 * If we're doing a normal import, then build up any additional
2164 	 * diagnostic information about missing log devices.
2165 	 * We'll pass this up to the user for further processing.
2166 	 */
2167 	if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) {
2168 		nvlist_t **child, *nv;
2169 		uint64_t idx = 0;
2170 
2171 		child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t *),
2172 		    KM_SLEEP);
2173 		nv = fnvlist_alloc();
2174 
2175 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
2176 			vdev_t *tvd = rvd->vdev_child[c];
2177 
2178 			/*
2179 			 * We consider a device as missing only if it failed
2180 			 * to open (i.e. offline or faulted is not considered
2181 			 * as missing).
2182 			 */
2183 			if (tvd->vdev_islog &&
2184 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
2185 				child[idx++] = vdev_config_generate(spa, tvd,
2186 				    B_FALSE, VDEV_CONFIG_MISSING);
2187 			}
2188 		}
2189 
2190 		if (idx > 0) {
2191 			fnvlist_add_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
2192 			    (const nvlist_t * const *)child, idx);
2193 			fnvlist_add_nvlist(spa->spa_load_info,
2194 			    ZPOOL_CONFIG_MISSING_DEVICES, nv);
2195 
2196 			for (uint64_t i = 0; i < idx; i++)
2197 				nvlist_free(child[i]);
2198 		}
2199 		nvlist_free(nv);
2200 		kmem_free(child, rvd->vdev_children * sizeof (char **));
2201 
2202 		if (idx > 0) {
2203 			spa_load_failed(spa, "some log devices are missing");
2204 			vdev_dbgmsg_print_tree(rvd, 2);
2205 			return (SET_ERROR(ENXIO));
2206 		}
2207 	} else {
2208 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
2209 			vdev_t *tvd = rvd->vdev_child[c];
2210 
2211 			if (tvd->vdev_islog &&
2212 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
2213 				spa_set_log_state(spa, SPA_LOG_CLEAR);
2214 				spa_load_note(spa, "some log devices are "
2215 				    "missing, ZIL is dropped.");
2216 				vdev_dbgmsg_print_tree(rvd, 2);
2217 				break;
2218 			}
2219 		}
2220 	}
2221 
2222 	return (0);
2223 }
2224 
2225 /*
2226  * Check for missing log devices
2227  */
2228 static boolean_t
2229 spa_check_logs(spa_t *spa)
2230 {
2231 	boolean_t rv = B_FALSE;
2232 	dsl_pool_t *dp = spa_get_dsl(spa);
2233 
2234 	switch (spa->spa_log_state) {
2235 	default:
2236 		break;
2237 	case SPA_LOG_MISSING:
2238 		/* need to recheck in case slog has been restored */
2239 	case SPA_LOG_UNKNOWN:
2240 		rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2241 		    zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0);
2242 		if (rv)
2243 			spa_set_log_state(spa, SPA_LOG_MISSING);
2244 		break;
2245 	}
2246 	return (rv);
2247 }
2248 
2249 /*
2250  * Passivate any log vdevs (note, does not apply to embedded log metaslabs).
2251  */
2252 static boolean_t
2253 spa_passivate_log(spa_t *spa)
2254 {
2255 	vdev_t *rvd = spa->spa_root_vdev;
2256 	boolean_t slog_found = B_FALSE;
2257 
2258 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
2259 
2260 	for (int c = 0; c < rvd->vdev_children; c++) {
2261 		vdev_t *tvd = rvd->vdev_child[c];
2262 
2263 		if (tvd->vdev_islog) {
2264 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
2265 			metaslab_group_passivate(tvd->vdev_mg);
2266 			slog_found = B_TRUE;
2267 		}
2268 	}
2269 
2270 	return (slog_found);
2271 }
2272 
2273 /*
2274  * Activate any log vdevs (note, does not apply to embedded log metaslabs).
2275  */
2276 static void
2277 spa_activate_log(spa_t *spa)
2278 {
2279 	vdev_t *rvd = spa->spa_root_vdev;
2280 
2281 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
2282 
2283 	for (int c = 0; c < rvd->vdev_children; c++) {
2284 		vdev_t *tvd = rvd->vdev_child[c];
2285 
2286 		if (tvd->vdev_islog) {
2287 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
2288 			metaslab_group_activate(tvd->vdev_mg);
2289 		}
2290 	}
2291 }
2292 
2293 int
2294 spa_reset_logs(spa_t *spa)
2295 {
2296 	int error;
2297 
2298 	error = dmu_objset_find(spa_name(spa), zil_reset,
2299 	    NULL, DS_FIND_CHILDREN);
2300 	if (error == 0) {
2301 		/*
2302 		 * We successfully offlined the log device, sync out the
2303 		 * current txg so that the "stubby" block can be removed
2304 		 * by zil_sync().
2305 		 */
2306 		txg_wait_synced(spa->spa_dsl_pool, 0);
2307 	}
2308 	return (error);
2309 }
2310 
2311 static void
2312 spa_aux_check_removed(spa_aux_vdev_t *sav)
2313 {
2314 	for (int i = 0; i < sav->sav_count; i++)
2315 		spa_check_removed(sav->sav_vdevs[i]);
2316 }
2317 
2318 void
2319 spa_claim_notify(zio_t *zio)
2320 {
2321 	spa_t *spa = zio->io_spa;
2322 
2323 	if (zio->io_error)
2324 		return;
2325 
2326 	mutex_enter(&spa->spa_props_lock);	/* any mutex will do */
2327 	if (spa->spa_claim_max_txg < zio->io_bp->blk_birth)
2328 		spa->spa_claim_max_txg = zio->io_bp->blk_birth;
2329 	mutex_exit(&spa->spa_props_lock);
2330 }
2331 
2332 typedef struct spa_load_error {
2333 	boolean_t	sle_verify_data;
2334 	uint64_t	sle_meta_count;
2335 	uint64_t	sle_data_count;
2336 } spa_load_error_t;
2337 
2338 static void
2339 spa_load_verify_done(zio_t *zio)
2340 {
2341 	blkptr_t *bp = zio->io_bp;
2342 	spa_load_error_t *sle = zio->io_private;
2343 	dmu_object_type_t type = BP_GET_TYPE(bp);
2344 	int error = zio->io_error;
2345 	spa_t *spa = zio->io_spa;
2346 
2347 	abd_free(zio->io_abd);
2348 	if (error) {
2349 		if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) &&
2350 		    type != DMU_OT_INTENT_LOG)
2351 			atomic_inc_64(&sle->sle_meta_count);
2352 		else
2353 			atomic_inc_64(&sle->sle_data_count);
2354 	}
2355 
2356 	mutex_enter(&spa->spa_scrub_lock);
2357 	spa->spa_load_verify_bytes -= BP_GET_PSIZE(bp);
2358 	cv_broadcast(&spa->spa_scrub_io_cv);
2359 	mutex_exit(&spa->spa_scrub_lock);
2360 }
2361 
2362 /*
2363  * Maximum number of inflight bytes is the log2 fraction of the arc size.
2364  * By default, we set it to 1/16th of the arc.
2365  */
2366 static uint_t spa_load_verify_shift = 4;
2367 static int spa_load_verify_metadata = B_TRUE;
2368 static int spa_load_verify_data = B_TRUE;
2369 
2370 static int
2371 spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
2372     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
2373 {
2374 	zio_t *rio = arg;
2375 	spa_load_error_t *sle = rio->io_private;
2376 
2377 	(void) zilog, (void) dnp;
2378 
2379 	/*
2380 	 * Note: normally this routine will not be called if
2381 	 * spa_load_verify_metadata is not set.  However, it may be useful
2382 	 * to manually set the flag after the traversal has begun.
2383 	 */
2384 	if (!spa_load_verify_metadata)
2385 		return (0);
2386 
2387 	/*
2388 	 * Sanity check the block pointer in order to detect obvious damage
2389 	 * before using the contents in subsequent checks or in zio_read().
2390 	 * When damaged consider it to be a metadata error since we cannot
2391 	 * trust the BP_GET_TYPE and BP_GET_LEVEL values.
2392 	 */
2393 	if (!zfs_blkptr_verify(spa, bp, BLK_CONFIG_NEEDED, BLK_VERIFY_LOG)) {
2394 		atomic_inc_64(&sle->sle_meta_count);
2395 		return (0);
2396 	}
2397 
2398 	if (zb->zb_level == ZB_DNODE_LEVEL || BP_IS_HOLE(bp) ||
2399 	    BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp))
2400 		return (0);
2401 
2402 	if (!BP_IS_METADATA(bp) &&
2403 	    (!spa_load_verify_data || !sle->sle_verify_data))
2404 		return (0);
2405 
2406 	uint64_t maxinflight_bytes =
2407 	    arc_target_bytes() >> spa_load_verify_shift;
2408 	size_t size = BP_GET_PSIZE(bp);
2409 
2410 	mutex_enter(&spa->spa_scrub_lock);
2411 	while (spa->spa_load_verify_bytes >= maxinflight_bytes)
2412 		cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2413 	spa->spa_load_verify_bytes += size;
2414 	mutex_exit(&spa->spa_scrub_lock);
2415 
2416 	zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size,
2417 	    spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB,
2418 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL |
2419 	    ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb));
2420 	return (0);
2421 }
2422 
2423 static int
2424 verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
2425 {
2426 	(void) dp, (void) arg;
2427 
2428 	if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN)
2429 		return (SET_ERROR(ENAMETOOLONG));
2430 
2431 	return (0);
2432 }
2433 
2434 static int
2435 spa_load_verify(spa_t *spa)
2436 {
2437 	zio_t *rio;
2438 	spa_load_error_t sle = { 0 };
2439 	zpool_load_policy_t policy;
2440 	boolean_t verify_ok = B_FALSE;
2441 	int error = 0;
2442 
2443 	zpool_get_load_policy(spa->spa_config, &policy);
2444 
2445 	if (policy.zlp_rewind & ZPOOL_NEVER_REWIND ||
2446 	    policy.zlp_maxmeta == UINT64_MAX)
2447 		return (0);
2448 
2449 	dsl_pool_config_enter(spa->spa_dsl_pool, FTAG);
2450 	error = dmu_objset_find_dp(spa->spa_dsl_pool,
2451 	    spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL,
2452 	    DS_FIND_CHILDREN);
2453 	dsl_pool_config_exit(spa->spa_dsl_pool, FTAG);
2454 	if (error != 0)
2455 		return (error);
2456 
2457 	/*
2458 	 * Verify data only if we are rewinding or error limit was set.
2459 	 * Otherwise nothing except dbgmsg care about it to waste time.
2460 	 */
2461 	sle.sle_verify_data = (policy.zlp_rewind & ZPOOL_REWIND_MASK) ||
2462 	    (policy.zlp_maxdata < UINT64_MAX);
2463 
2464 	rio = zio_root(spa, NULL, &sle,
2465 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
2466 
2467 	if (spa_load_verify_metadata) {
2468 		if (spa->spa_extreme_rewind) {
2469 			spa_load_note(spa, "performing a complete scan of the "
2470 			    "pool since extreme rewind is on. This may take "
2471 			    "a very long time.\n  (spa_load_verify_data=%u, "
2472 			    "spa_load_verify_metadata=%u)",
2473 			    spa_load_verify_data, spa_load_verify_metadata);
2474 		}
2475 
2476 		error = traverse_pool(spa, spa->spa_verify_min_txg,
2477 		    TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA |
2478 		    TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio);
2479 	}
2480 
2481 	(void) zio_wait(rio);
2482 	ASSERT0(spa->spa_load_verify_bytes);
2483 
2484 	spa->spa_load_meta_errors = sle.sle_meta_count;
2485 	spa->spa_load_data_errors = sle.sle_data_count;
2486 
2487 	if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) {
2488 		spa_load_note(spa, "spa_load_verify found %llu metadata errors "
2489 		    "and %llu data errors", (u_longlong_t)sle.sle_meta_count,
2490 		    (u_longlong_t)sle.sle_data_count);
2491 	}
2492 
2493 	if (spa_load_verify_dryrun ||
2494 	    (!error && sle.sle_meta_count <= policy.zlp_maxmeta &&
2495 	    sle.sle_data_count <= policy.zlp_maxdata)) {
2496 		int64_t loss = 0;
2497 
2498 		verify_ok = B_TRUE;
2499 		spa->spa_load_txg = spa->spa_uberblock.ub_txg;
2500 		spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp;
2501 
2502 		loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts;
2503 		fnvlist_add_uint64(spa->spa_load_info, ZPOOL_CONFIG_LOAD_TIME,
2504 		    spa->spa_load_txg_ts);
2505 		fnvlist_add_int64(spa->spa_load_info, ZPOOL_CONFIG_REWIND_TIME,
2506 		    loss);
2507 		fnvlist_add_uint64(spa->spa_load_info,
2508 		    ZPOOL_CONFIG_LOAD_META_ERRORS, sle.sle_meta_count);
2509 		fnvlist_add_uint64(spa->spa_load_info,
2510 		    ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count);
2511 	} else {
2512 		spa->spa_load_max_txg = spa->spa_uberblock.ub_txg;
2513 	}
2514 
2515 	if (spa_load_verify_dryrun)
2516 		return (0);
2517 
2518 	if (error) {
2519 		if (error != ENXIO && error != EIO)
2520 			error = SET_ERROR(EIO);
2521 		return (error);
2522 	}
2523 
2524 	return (verify_ok ? 0 : EIO);
2525 }
2526 
2527 /*
2528  * Find a value in the pool props object.
2529  */
2530 static void
2531 spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val)
2532 {
2533 	(void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object,
2534 	    zpool_prop_to_name(prop), sizeof (uint64_t), 1, val);
2535 }
2536 
2537 /*
2538  * Find a value in the pool directory object.
2539  */
2540 static int
2541 spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent)
2542 {
2543 	int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
2544 	    name, sizeof (uint64_t), 1, val);
2545 
2546 	if (error != 0 && (error != ENOENT || log_enoent)) {
2547 		spa_load_failed(spa, "couldn't get '%s' value in MOS directory "
2548 		    "[error=%d]", name, error);
2549 	}
2550 
2551 	return (error);
2552 }
2553 
2554 static int
2555 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err)
2556 {
2557 	vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux);
2558 	return (SET_ERROR(err));
2559 }
2560 
2561 boolean_t
2562 spa_livelist_delete_check(spa_t *spa)
2563 {
2564 	return (spa->spa_livelists_to_delete != 0);
2565 }
2566 
2567 static boolean_t
2568 spa_livelist_delete_cb_check(void *arg, zthr_t *z)
2569 {
2570 	(void) z;
2571 	spa_t *spa = arg;
2572 	return (spa_livelist_delete_check(spa));
2573 }
2574 
2575 static int
2576 delete_blkptr_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
2577 {
2578 	spa_t *spa = arg;
2579 	zio_free(spa, tx->tx_txg, bp);
2580 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
2581 	    -bp_get_dsize_sync(spa, bp),
2582 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
2583 	return (0);
2584 }
2585 
2586 static int
2587 dsl_get_next_livelist_obj(objset_t *os, uint64_t zap_obj, uint64_t *llp)
2588 {
2589 	int err;
2590 	zap_cursor_t zc;
2591 	zap_attribute_t za;
2592 	zap_cursor_init(&zc, os, zap_obj);
2593 	err = zap_cursor_retrieve(&zc, &za);
2594 	zap_cursor_fini(&zc);
2595 	if (err == 0)
2596 		*llp = za.za_first_integer;
2597 	return (err);
2598 }
2599 
2600 /*
2601  * Components of livelist deletion that must be performed in syncing
2602  * context: freeing block pointers and updating the pool-wide data
2603  * structures to indicate how much work is left to do
2604  */
2605 typedef struct sublist_delete_arg {
2606 	spa_t *spa;
2607 	dsl_deadlist_t *ll;
2608 	uint64_t key;
2609 	bplist_t *to_free;
2610 } sublist_delete_arg_t;
2611 
2612 static void
2613 sublist_delete_sync(void *arg, dmu_tx_t *tx)
2614 {
2615 	sublist_delete_arg_t *sda = arg;
2616 	spa_t *spa = sda->spa;
2617 	dsl_deadlist_t *ll = sda->ll;
2618 	uint64_t key = sda->key;
2619 	bplist_t *to_free = sda->to_free;
2620 
2621 	bplist_iterate(to_free, delete_blkptr_cb, spa, tx);
2622 	dsl_deadlist_remove_entry(ll, key, tx);
2623 }
2624 
2625 typedef struct livelist_delete_arg {
2626 	spa_t *spa;
2627 	uint64_t ll_obj;
2628 	uint64_t zap_obj;
2629 } livelist_delete_arg_t;
2630 
2631 static void
2632 livelist_delete_sync(void *arg, dmu_tx_t *tx)
2633 {
2634 	livelist_delete_arg_t *lda = arg;
2635 	spa_t *spa = lda->spa;
2636 	uint64_t ll_obj = lda->ll_obj;
2637 	uint64_t zap_obj = lda->zap_obj;
2638 	objset_t *mos = spa->spa_meta_objset;
2639 	uint64_t count;
2640 
2641 	/* free the livelist and decrement the feature count */
2642 	VERIFY0(zap_remove_int(mos, zap_obj, ll_obj, tx));
2643 	dsl_deadlist_free(mos, ll_obj, tx);
2644 	spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx);
2645 	VERIFY0(zap_count(mos, zap_obj, &count));
2646 	if (count == 0) {
2647 		/* no more livelists to delete */
2648 		VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT,
2649 		    DMU_POOL_DELETED_CLONES, tx));
2650 		VERIFY0(zap_destroy(mos, zap_obj, tx));
2651 		spa->spa_livelists_to_delete = 0;
2652 		spa_notify_waiters(spa);
2653 	}
2654 }
2655 
2656 /*
2657  * Load in the value for the livelist to be removed and open it. Then,
2658  * load its first sublist and determine which block pointers should actually
2659  * be freed. Then, call a synctask which performs the actual frees and updates
2660  * the pool-wide livelist data.
2661  */
2662 static void
2663 spa_livelist_delete_cb(void *arg, zthr_t *z)
2664 {
2665 	spa_t *spa = arg;
2666 	uint64_t ll_obj = 0, count;
2667 	objset_t *mos = spa->spa_meta_objset;
2668 	uint64_t zap_obj = spa->spa_livelists_to_delete;
2669 	/*
2670 	 * Determine the next livelist to delete. This function should only
2671 	 * be called if there is at least one deleted clone.
2672 	 */
2673 	VERIFY0(dsl_get_next_livelist_obj(mos, zap_obj, &ll_obj));
2674 	VERIFY0(zap_count(mos, ll_obj, &count));
2675 	if (count > 0) {
2676 		dsl_deadlist_t *ll;
2677 		dsl_deadlist_entry_t *dle;
2678 		bplist_t to_free;
2679 		ll = kmem_zalloc(sizeof (dsl_deadlist_t), KM_SLEEP);
2680 		dsl_deadlist_open(ll, mos, ll_obj);
2681 		dle = dsl_deadlist_first(ll);
2682 		ASSERT3P(dle, !=, NULL);
2683 		bplist_create(&to_free);
2684 		int err = dsl_process_sub_livelist(&dle->dle_bpobj, &to_free,
2685 		    z, NULL);
2686 		if (err == 0) {
2687 			sublist_delete_arg_t sync_arg = {
2688 			    .spa = spa,
2689 			    .ll = ll,
2690 			    .key = dle->dle_mintxg,
2691 			    .to_free = &to_free
2692 			};
2693 			zfs_dbgmsg("deleting sublist (id %llu) from"
2694 			    " livelist %llu, %lld remaining",
2695 			    (u_longlong_t)dle->dle_bpobj.bpo_object,
2696 			    (u_longlong_t)ll_obj, (longlong_t)count - 1);
2697 			VERIFY0(dsl_sync_task(spa_name(spa), NULL,
2698 			    sublist_delete_sync, &sync_arg, 0,
2699 			    ZFS_SPACE_CHECK_DESTROY));
2700 		} else {
2701 			VERIFY3U(err, ==, EINTR);
2702 		}
2703 		bplist_clear(&to_free);
2704 		bplist_destroy(&to_free);
2705 		dsl_deadlist_close(ll);
2706 		kmem_free(ll, sizeof (dsl_deadlist_t));
2707 	} else {
2708 		livelist_delete_arg_t sync_arg = {
2709 		    .spa = spa,
2710 		    .ll_obj = ll_obj,
2711 		    .zap_obj = zap_obj
2712 		};
2713 		zfs_dbgmsg("deletion of livelist %llu completed",
2714 		    (u_longlong_t)ll_obj);
2715 		VERIFY0(dsl_sync_task(spa_name(spa), NULL, livelist_delete_sync,
2716 		    &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY));
2717 	}
2718 }
2719 
2720 static void
2721 spa_start_livelist_destroy_thread(spa_t *spa)
2722 {
2723 	ASSERT3P(spa->spa_livelist_delete_zthr, ==, NULL);
2724 	spa->spa_livelist_delete_zthr =
2725 	    zthr_create("z_livelist_destroy",
2726 	    spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa,
2727 	    minclsyspri);
2728 }
2729 
2730 typedef struct livelist_new_arg {
2731 	bplist_t *allocs;
2732 	bplist_t *frees;
2733 } livelist_new_arg_t;
2734 
2735 static int
2736 livelist_track_new_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
2737     dmu_tx_t *tx)
2738 {
2739 	ASSERT(tx == NULL);
2740 	livelist_new_arg_t *lna = arg;
2741 	if (bp_freed) {
2742 		bplist_append(lna->frees, bp);
2743 	} else {
2744 		bplist_append(lna->allocs, bp);
2745 		zfs_livelist_condense_new_alloc++;
2746 	}
2747 	return (0);
2748 }
2749 
2750 typedef struct livelist_condense_arg {
2751 	spa_t *spa;
2752 	bplist_t to_keep;
2753 	uint64_t first_size;
2754 	uint64_t next_size;
2755 } livelist_condense_arg_t;
2756 
2757 static void
2758 spa_livelist_condense_sync(void *arg, dmu_tx_t *tx)
2759 {
2760 	livelist_condense_arg_t *lca = arg;
2761 	spa_t *spa = lca->spa;
2762 	bplist_t new_frees;
2763 	dsl_dataset_t *ds = spa->spa_to_condense.ds;
2764 
2765 	/* Have we been cancelled? */
2766 	if (spa->spa_to_condense.cancelled) {
2767 		zfs_livelist_condense_sync_cancel++;
2768 		goto out;
2769 	}
2770 
2771 	dsl_deadlist_entry_t *first = spa->spa_to_condense.first;
2772 	dsl_deadlist_entry_t *next = spa->spa_to_condense.next;
2773 	dsl_deadlist_t *ll = &ds->ds_dir->dd_livelist;
2774 
2775 	/*
2776 	 * It's possible that the livelist was changed while the zthr was
2777 	 * running. Therefore, we need to check for new blkptrs in the two
2778 	 * entries being condensed and continue to track them in the livelist.
2779 	 * Because of the way we handle remapped blkptrs (see dbuf_remap_impl),
2780 	 * it's possible that the newly added blkptrs are FREEs or ALLOCs so
2781 	 * we need to sort them into two different bplists.
2782 	 */
2783 	uint64_t first_obj = first->dle_bpobj.bpo_object;
2784 	uint64_t next_obj = next->dle_bpobj.bpo_object;
2785 	uint64_t cur_first_size = first->dle_bpobj.bpo_phys->bpo_num_blkptrs;
2786 	uint64_t cur_next_size = next->dle_bpobj.bpo_phys->bpo_num_blkptrs;
2787 
2788 	bplist_create(&new_frees);
2789 	livelist_new_arg_t new_bps = {
2790 	    .allocs = &lca->to_keep,
2791 	    .frees = &new_frees,
2792 	};
2793 
2794 	if (cur_first_size > lca->first_size) {
2795 		VERIFY0(livelist_bpobj_iterate_from_nofree(&first->dle_bpobj,
2796 		    livelist_track_new_cb, &new_bps, lca->first_size));
2797 	}
2798 	if (cur_next_size > lca->next_size) {
2799 		VERIFY0(livelist_bpobj_iterate_from_nofree(&next->dle_bpobj,
2800 		    livelist_track_new_cb, &new_bps, lca->next_size));
2801 	}
2802 
2803 	dsl_deadlist_clear_entry(first, ll, tx);
2804 	ASSERT(bpobj_is_empty(&first->dle_bpobj));
2805 	dsl_deadlist_remove_entry(ll, next->dle_mintxg, tx);
2806 
2807 	bplist_iterate(&lca->to_keep, dsl_deadlist_insert_alloc_cb, ll, tx);
2808 	bplist_iterate(&new_frees, dsl_deadlist_insert_free_cb, ll, tx);
2809 	bplist_destroy(&new_frees);
2810 
2811 	char dsname[ZFS_MAX_DATASET_NAME_LEN];
2812 	dsl_dataset_name(ds, dsname);
2813 	zfs_dbgmsg("txg %llu condensing livelist of %s (id %llu), bpobj %llu "
2814 	    "(%llu blkptrs) and bpobj %llu (%llu blkptrs) -> bpobj %llu "
2815 	    "(%llu blkptrs)", (u_longlong_t)tx->tx_txg, dsname,
2816 	    (u_longlong_t)ds->ds_object, (u_longlong_t)first_obj,
2817 	    (u_longlong_t)cur_first_size, (u_longlong_t)next_obj,
2818 	    (u_longlong_t)cur_next_size,
2819 	    (u_longlong_t)first->dle_bpobj.bpo_object,
2820 	    (u_longlong_t)first->dle_bpobj.bpo_phys->bpo_num_blkptrs);
2821 out:
2822 	dmu_buf_rele(ds->ds_dbuf, spa);
2823 	spa->spa_to_condense.ds = NULL;
2824 	bplist_clear(&lca->to_keep);
2825 	bplist_destroy(&lca->to_keep);
2826 	kmem_free(lca, sizeof (livelist_condense_arg_t));
2827 	spa->spa_to_condense.syncing = B_FALSE;
2828 }
2829 
2830 static void
2831 spa_livelist_condense_cb(void *arg, zthr_t *t)
2832 {
2833 	while (zfs_livelist_condense_zthr_pause &&
2834 	    !(zthr_has_waiters(t) || zthr_iscancelled(t)))
2835 		delay(1);
2836 
2837 	spa_t *spa = arg;
2838 	dsl_deadlist_entry_t *first = spa->spa_to_condense.first;
2839 	dsl_deadlist_entry_t *next = spa->spa_to_condense.next;
2840 	uint64_t first_size, next_size;
2841 
2842 	livelist_condense_arg_t *lca =
2843 	    kmem_alloc(sizeof (livelist_condense_arg_t), KM_SLEEP);
2844 	bplist_create(&lca->to_keep);
2845 
2846 	/*
2847 	 * Process the livelists (matching FREEs and ALLOCs) in open context
2848 	 * so we have minimal work in syncing context to condense.
2849 	 *
2850 	 * We save bpobj sizes (first_size and next_size) to use later in
2851 	 * syncing context to determine if entries were added to these sublists
2852 	 * while in open context. This is possible because the clone is still
2853 	 * active and open for normal writes and we want to make sure the new,
2854 	 * unprocessed blockpointers are inserted into the livelist normally.
2855 	 *
2856 	 * Note that dsl_process_sub_livelist() both stores the size number of
2857 	 * blockpointers and iterates over them while the bpobj's lock held, so
2858 	 * the sizes returned to us are consistent which what was actually
2859 	 * processed.
2860 	 */
2861 	int err = dsl_process_sub_livelist(&first->dle_bpobj, &lca->to_keep, t,
2862 	    &first_size);
2863 	if (err == 0)
2864 		err = dsl_process_sub_livelist(&next->dle_bpobj, &lca->to_keep,
2865 		    t, &next_size);
2866 
2867 	if (err == 0) {
2868 		while (zfs_livelist_condense_sync_pause &&
2869 		    !(zthr_has_waiters(t) || zthr_iscancelled(t)))
2870 			delay(1);
2871 
2872 		dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
2873 		dmu_tx_mark_netfree(tx);
2874 		dmu_tx_hold_space(tx, 1);
2875 		err = dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE);
2876 		if (err == 0) {
2877 			/*
2878 			 * Prevent the condense zthr restarting before
2879 			 * the synctask completes.
2880 			 */
2881 			spa->spa_to_condense.syncing = B_TRUE;
2882 			lca->spa = spa;
2883 			lca->first_size = first_size;
2884 			lca->next_size = next_size;
2885 			dsl_sync_task_nowait(spa_get_dsl(spa),
2886 			    spa_livelist_condense_sync, lca, tx);
2887 			dmu_tx_commit(tx);
2888 			return;
2889 		}
2890 	}
2891 	/*
2892 	 * Condensing can not continue: either it was externally stopped or
2893 	 * we were unable to assign to a tx because the pool has run out of
2894 	 * space. In the second case, we'll just end up trying to condense
2895 	 * again in a later txg.
2896 	 */
2897 	ASSERT(err != 0);
2898 	bplist_clear(&lca->to_keep);
2899 	bplist_destroy(&lca->to_keep);
2900 	kmem_free(lca, sizeof (livelist_condense_arg_t));
2901 	dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf, spa);
2902 	spa->spa_to_condense.ds = NULL;
2903 	if (err == EINTR)
2904 		zfs_livelist_condense_zthr_cancel++;
2905 }
2906 
2907 /*
2908  * Check that there is something to condense but that a condense is not
2909  * already in progress and that condensing has not been cancelled.
2910  */
2911 static boolean_t
2912 spa_livelist_condense_cb_check(void *arg, zthr_t *z)
2913 {
2914 	(void) z;
2915 	spa_t *spa = arg;
2916 	if ((spa->spa_to_condense.ds != NULL) &&
2917 	    (spa->spa_to_condense.syncing == B_FALSE) &&
2918 	    (spa->spa_to_condense.cancelled == B_FALSE)) {
2919 		return (B_TRUE);
2920 	}
2921 	return (B_FALSE);
2922 }
2923 
2924 static void
2925 spa_start_livelist_condensing_thread(spa_t *spa)
2926 {
2927 	spa->spa_to_condense.ds = NULL;
2928 	spa->spa_to_condense.first = NULL;
2929 	spa->spa_to_condense.next = NULL;
2930 	spa->spa_to_condense.syncing = B_FALSE;
2931 	spa->spa_to_condense.cancelled = B_FALSE;
2932 
2933 	ASSERT3P(spa->spa_livelist_condense_zthr, ==, NULL);
2934 	spa->spa_livelist_condense_zthr =
2935 	    zthr_create("z_livelist_condense",
2936 	    spa_livelist_condense_cb_check,
2937 	    spa_livelist_condense_cb, spa, minclsyspri);
2938 }
2939 
2940 static void
2941 spa_spawn_aux_threads(spa_t *spa)
2942 {
2943 	ASSERT(spa_writeable(spa));
2944 
2945 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
2946 
2947 	spa_start_indirect_condensing_thread(spa);
2948 	spa_start_livelist_destroy_thread(spa);
2949 	spa_start_livelist_condensing_thread(spa);
2950 
2951 	ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL);
2952 	spa->spa_checkpoint_discard_zthr =
2953 	    zthr_create("z_checkpoint_discard",
2954 	    spa_checkpoint_discard_thread_check,
2955 	    spa_checkpoint_discard_thread, spa, minclsyspri);
2956 }
2957 
2958 /*
2959  * Fix up config after a partly-completed split.  This is done with the
2960  * ZPOOL_CONFIG_SPLIT nvlist.  Both the splitting pool and the split-off
2961  * pool have that entry in their config, but only the splitting one contains
2962  * a list of all the guids of the vdevs that are being split off.
2963  *
2964  * This function determines what to do with that list: either rejoin
2965  * all the disks to the pool, or complete the splitting process.  To attempt
2966  * the rejoin, each disk that is offlined is marked online again, and
2967  * we do a reopen() call.  If the vdev label for every disk that was
2968  * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL)
2969  * then we call vdev_split() on each disk, and complete the split.
2970  *
2971  * Otherwise we leave the config alone, with all the vdevs in place in
2972  * the original pool.
2973  */
2974 static void
2975 spa_try_repair(spa_t *spa, nvlist_t *config)
2976 {
2977 	uint_t extracted;
2978 	uint64_t *glist;
2979 	uint_t i, gcount;
2980 	nvlist_t *nvl;
2981 	vdev_t **vd;
2982 	boolean_t attempt_reopen;
2983 
2984 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0)
2985 		return;
2986 
2987 	/* check that the config is complete */
2988 	if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST,
2989 	    &glist, &gcount) != 0)
2990 		return;
2991 
2992 	vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP);
2993 
2994 	/* attempt to online all the vdevs & validate */
2995 	attempt_reopen = B_TRUE;
2996 	for (i = 0; i < gcount; i++) {
2997 		if (glist[i] == 0)	/* vdev is hole */
2998 			continue;
2999 
3000 		vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE);
3001 		if (vd[i] == NULL) {
3002 			/*
3003 			 * Don't bother attempting to reopen the disks;
3004 			 * just do the split.
3005 			 */
3006 			attempt_reopen = B_FALSE;
3007 		} else {
3008 			/* attempt to re-online it */
3009 			vd[i]->vdev_offline = B_FALSE;
3010 		}
3011 	}
3012 
3013 	if (attempt_reopen) {
3014 		vdev_reopen(spa->spa_root_vdev);
3015 
3016 		/* check each device to see what state it's in */
3017 		for (extracted = 0, i = 0; i < gcount; i++) {
3018 			if (vd[i] != NULL &&
3019 			    vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL)
3020 				break;
3021 			++extracted;
3022 		}
3023 	}
3024 
3025 	/*
3026 	 * If every disk has been moved to the new pool, or if we never
3027 	 * even attempted to look at them, then we split them off for
3028 	 * good.
3029 	 */
3030 	if (!attempt_reopen || gcount == extracted) {
3031 		for (i = 0; i < gcount; i++)
3032 			if (vd[i] != NULL)
3033 				vdev_split(vd[i]);
3034 		vdev_reopen(spa->spa_root_vdev);
3035 	}
3036 
3037 	kmem_free(vd, gcount * sizeof (vdev_t *));
3038 }
3039 
3040 static int
3041 spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type)
3042 {
3043 	const char *ereport = FM_EREPORT_ZFS_POOL;
3044 	int error;
3045 
3046 	spa->spa_load_state = state;
3047 	(void) spa_import_progress_set_state(spa_guid(spa),
3048 	    spa_load_state(spa));
3049 
3050 	gethrestime(&spa->spa_loaded_ts);
3051 	error = spa_load_impl(spa, type, &ereport);
3052 
3053 	/*
3054 	 * Don't count references from objsets that are already closed
3055 	 * and are making their way through the eviction process.
3056 	 */
3057 	spa_evicting_os_wait(spa);
3058 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
3059 	if (error) {
3060 		if (error != EEXIST) {
3061 			spa->spa_loaded_ts.tv_sec = 0;
3062 			spa->spa_loaded_ts.tv_nsec = 0;
3063 		}
3064 		if (error != EBADF) {
3065 			(void) zfs_ereport_post(ereport, spa,
3066 			    NULL, NULL, NULL, 0);
3067 		}
3068 	}
3069 	spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE;
3070 	spa->spa_ena = 0;
3071 
3072 	(void) spa_import_progress_set_state(spa_guid(spa),
3073 	    spa_load_state(spa));
3074 
3075 	return (error);
3076 }
3077 
3078 #ifdef ZFS_DEBUG
3079 /*
3080  * Count the number of per-vdev ZAPs associated with all of the vdevs in the
3081  * vdev tree rooted in the given vd, and ensure that each ZAP is present in the
3082  * spa's per-vdev ZAP list.
3083  */
3084 static uint64_t
3085 vdev_count_verify_zaps(vdev_t *vd)
3086 {
3087 	spa_t *spa = vd->vdev_spa;
3088 	uint64_t total = 0;
3089 
3090 	if (spa_feature_is_active(vd->vdev_spa, SPA_FEATURE_AVZ_V2) &&
3091 	    vd->vdev_root_zap != 0) {
3092 		total++;
3093 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
3094 		    spa->spa_all_vdev_zaps, vd->vdev_root_zap));
3095 	}
3096 	if (vd->vdev_top_zap != 0) {
3097 		total++;
3098 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
3099 		    spa->spa_all_vdev_zaps, vd->vdev_top_zap));
3100 	}
3101 	if (vd->vdev_leaf_zap != 0) {
3102 		total++;
3103 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
3104 		    spa->spa_all_vdev_zaps, vd->vdev_leaf_zap));
3105 	}
3106 
3107 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3108 		total += vdev_count_verify_zaps(vd->vdev_child[i]);
3109 	}
3110 
3111 	return (total);
3112 }
3113 #else
3114 #define	vdev_count_verify_zaps(vd) ((void) sizeof (vd), 0)
3115 #endif
3116 
3117 /*
3118  * Determine whether the activity check is required.
3119  */
3120 static boolean_t
3121 spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
3122     nvlist_t *config)
3123 {
3124 	uint64_t state = 0;
3125 	uint64_t hostid = 0;
3126 	uint64_t tryconfig_txg = 0;
3127 	uint64_t tryconfig_timestamp = 0;
3128 	uint16_t tryconfig_mmp_seq = 0;
3129 	nvlist_t *nvinfo;
3130 
3131 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
3132 		nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO);
3133 		(void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG,
3134 		    &tryconfig_txg);
3135 		(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
3136 		    &tryconfig_timestamp);
3137 		(void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ,
3138 		    &tryconfig_mmp_seq);
3139 	}
3140 
3141 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state);
3142 
3143 	/*
3144 	 * Disable the MMP activity check - This is used by zdb which
3145 	 * is intended to be used on potentially active pools.
3146 	 */
3147 	if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP)
3148 		return (B_FALSE);
3149 
3150 	/*
3151 	 * Skip the activity check when the MMP feature is disabled.
3152 	 */
3153 	if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0)
3154 		return (B_FALSE);
3155 
3156 	/*
3157 	 * If the tryconfig_ values are nonzero, they are the results of an
3158 	 * earlier tryimport.  If they all match the uberblock we just found,
3159 	 * then the pool has not changed and we return false so we do not test
3160 	 * a second time.
3161 	 */
3162 	if (tryconfig_txg && tryconfig_txg == ub->ub_txg &&
3163 	    tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp &&
3164 	    tryconfig_mmp_seq && tryconfig_mmp_seq ==
3165 	    (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0))
3166 		return (B_FALSE);
3167 
3168 	/*
3169 	 * Allow the activity check to be skipped when importing the pool
3170 	 * on the same host which last imported it.  Since the hostid from
3171 	 * configuration may be stale use the one read from the label.
3172 	 */
3173 	if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID))
3174 		hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID);
3175 
3176 	if (hostid == spa_get_hostid(spa))
3177 		return (B_FALSE);
3178 
3179 	/*
3180 	 * Skip the activity test when the pool was cleanly exported.
3181 	 */
3182 	if (state != POOL_STATE_ACTIVE)
3183 		return (B_FALSE);
3184 
3185 	return (B_TRUE);
3186 }
3187 
3188 /*
3189  * Nanoseconds the activity check must watch for changes on-disk.
3190  */
3191 static uint64_t
3192 spa_activity_check_duration(spa_t *spa, uberblock_t *ub)
3193 {
3194 	uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1);
3195 	uint64_t multihost_interval = MSEC2NSEC(
3196 	    MMP_INTERVAL_OK(zfs_multihost_interval));
3197 	uint64_t import_delay = MAX(NANOSEC, import_intervals *
3198 	    multihost_interval);
3199 
3200 	/*
3201 	 * Local tunables determine a minimum duration except for the case
3202 	 * where we know when the remote host will suspend the pool if MMP
3203 	 * writes do not land.
3204 	 *
3205 	 * See Big Theory comment at the top of mmp.c for the reasoning behind
3206 	 * these cases and times.
3207 	 */
3208 
3209 	ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100);
3210 
3211 	if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
3212 	    MMP_FAIL_INT(ub) > 0) {
3213 
3214 		/* MMP on remote host will suspend pool after failed writes */
3215 		import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) *
3216 		    MMP_IMPORT_SAFETY_FACTOR / 100;
3217 
3218 		zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp "
3219 		    "mmp_fails=%llu ub_mmp mmp_interval=%llu "
3220 		    "import_intervals=%llu", (u_longlong_t)import_delay,
3221 		    (u_longlong_t)MMP_FAIL_INT(ub),
3222 		    (u_longlong_t)MMP_INTERVAL(ub),
3223 		    (u_longlong_t)import_intervals);
3224 
3225 	} else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
3226 	    MMP_FAIL_INT(ub) == 0) {
3227 
3228 		/* MMP on remote host will never suspend pool */
3229 		import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) +
3230 		    ub->ub_mmp_delay) * import_intervals);
3231 
3232 		zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp "
3233 		    "mmp_interval=%llu ub_mmp_delay=%llu "
3234 		    "import_intervals=%llu", (u_longlong_t)import_delay,
3235 		    (u_longlong_t)MMP_INTERVAL(ub),
3236 		    (u_longlong_t)ub->ub_mmp_delay,
3237 		    (u_longlong_t)import_intervals);
3238 
3239 	} else if (MMP_VALID(ub)) {
3240 		/*
3241 		 * zfs-0.7 compatibility case
3242 		 */
3243 
3244 		import_delay = MAX(import_delay, (multihost_interval +
3245 		    ub->ub_mmp_delay) * import_intervals);
3246 
3247 		zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu "
3248 		    "import_intervals=%llu leaves=%u",
3249 		    (u_longlong_t)import_delay,
3250 		    (u_longlong_t)ub->ub_mmp_delay,
3251 		    (u_longlong_t)import_intervals,
3252 		    vdev_count_leaves(spa));
3253 	} else {
3254 		/* Using local tunings is the only reasonable option */
3255 		zfs_dbgmsg("pool last imported on non-MMP aware "
3256 		    "host using import_delay=%llu multihost_interval=%llu "
3257 		    "import_intervals=%llu", (u_longlong_t)import_delay,
3258 		    (u_longlong_t)multihost_interval,
3259 		    (u_longlong_t)import_intervals);
3260 	}
3261 
3262 	return (import_delay);
3263 }
3264 
3265 /*
3266  * Perform the import activity check.  If the user canceled the import or
3267  * we detected activity then fail.
3268  */
3269 static int
3270 spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
3271 {
3272 	uint64_t txg = ub->ub_txg;
3273 	uint64_t timestamp = ub->ub_timestamp;
3274 	uint64_t mmp_config = ub->ub_mmp_config;
3275 	uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0;
3276 	uint64_t import_delay;
3277 	hrtime_t import_expire;
3278 	nvlist_t *mmp_label = NULL;
3279 	vdev_t *rvd = spa->spa_root_vdev;
3280 	kcondvar_t cv;
3281 	kmutex_t mtx;
3282 	int error = 0;
3283 
3284 	cv_init(&cv, NULL, CV_DEFAULT, NULL);
3285 	mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL);
3286 	mutex_enter(&mtx);
3287 
3288 	/*
3289 	 * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed
3290 	 * during the earlier tryimport.  If the txg recorded there is 0 then
3291 	 * the pool is known to be active on another host.
3292 	 *
3293 	 * Otherwise, the pool might be in use on another host.  Check for
3294 	 * changes in the uberblocks on disk if necessary.
3295 	 */
3296 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
3297 		nvlist_t *nvinfo = fnvlist_lookup_nvlist(config,
3298 		    ZPOOL_CONFIG_LOAD_INFO);
3299 
3300 		if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) &&
3301 		    fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) {
3302 			vdev_uberblock_load(rvd, ub, &mmp_label);
3303 			error = SET_ERROR(EREMOTEIO);
3304 			goto out;
3305 		}
3306 	}
3307 
3308 	import_delay = spa_activity_check_duration(spa, ub);
3309 
3310 	/* Add a small random factor in case of simultaneous imports (0-25%) */
3311 	import_delay += import_delay * random_in_range(250) / 1000;
3312 
3313 	import_expire = gethrtime() + import_delay;
3314 
3315 	while (gethrtime() < import_expire) {
3316 		(void) spa_import_progress_set_mmp_check(spa_guid(spa),
3317 		    NSEC2SEC(import_expire - gethrtime()));
3318 
3319 		vdev_uberblock_load(rvd, ub, &mmp_label);
3320 
3321 		if (txg != ub->ub_txg || timestamp != ub->ub_timestamp ||
3322 		    mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) {
3323 			zfs_dbgmsg("multihost activity detected "
3324 			    "txg %llu ub_txg  %llu "
3325 			    "timestamp %llu ub_timestamp  %llu "
3326 			    "mmp_config %#llx ub_mmp_config %#llx",
3327 			    (u_longlong_t)txg, (u_longlong_t)ub->ub_txg,
3328 			    (u_longlong_t)timestamp,
3329 			    (u_longlong_t)ub->ub_timestamp,
3330 			    (u_longlong_t)mmp_config,
3331 			    (u_longlong_t)ub->ub_mmp_config);
3332 
3333 			error = SET_ERROR(EREMOTEIO);
3334 			break;
3335 		}
3336 
3337 		if (mmp_label) {
3338 			nvlist_free(mmp_label);
3339 			mmp_label = NULL;
3340 		}
3341 
3342 		error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz);
3343 		if (error != -1) {
3344 			error = SET_ERROR(EINTR);
3345 			break;
3346 		}
3347 		error = 0;
3348 	}
3349 
3350 out:
3351 	mutex_exit(&mtx);
3352 	mutex_destroy(&mtx);
3353 	cv_destroy(&cv);
3354 
3355 	/*
3356 	 * If the pool is determined to be active store the status in the
3357 	 * spa->spa_load_info nvlist.  If the remote hostname or hostid are
3358 	 * available from configuration read from disk store them as well.
3359 	 * This allows 'zpool import' to generate a more useful message.
3360 	 *
3361 	 * ZPOOL_CONFIG_MMP_STATE    - observed pool status (mandatory)
3362 	 * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool
3363 	 * ZPOOL_CONFIG_MMP_HOSTID   - hostid from the active pool
3364 	 */
3365 	if (error == EREMOTEIO) {
3366 		const char *hostname = "<unknown>";
3367 		uint64_t hostid = 0;
3368 
3369 		if (mmp_label) {
3370 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) {
3371 				hostname = fnvlist_lookup_string(mmp_label,
3372 				    ZPOOL_CONFIG_HOSTNAME);
3373 				fnvlist_add_string(spa->spa_load_info,
3374 				    ZPOOL_CONFIG_MMP_HOSTNAME, hostname);
3375 			}
3376 
3377 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) {
3378 				hostid = fnvlist_lookup_uint64(mmp_label,
3379 				    ZPOOL_CONFIG_HOSTID);
3380 				fnvlist_add_uint64(spa->spa_load_info,
3381 				    ZPOOL_CONFIG_MMP_HOSTID, hostid);
3382 			}
3383 		}
3384 
3385 		fnvlist_add_uint64(spa->spa_load_info,
3386 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE);
3387 		fnvlist_add_uint64(spa->spa_load_info,
3388 		    ZPOOL_CONFIG_MMP_TXG, 0);
3389 
3390 		error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO);
3391 	}
3392 
3393 	if (mmp_label)
3394 		nvlist_free(mmp_label);
3395 
3396 	return (error);
3397 }
3398 
3399 static int
3400 spa_verify_host(spa_t *spa, nvlist_t *mos_config)
3401 {
3402 	uint64_t hostid;
3403 	const char *hostname;
3404 	uint64_t myhostid = 0;
3405 
3406 	if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config,
3407 	    ZPOOL_CONFIG_HOSTID, &hostid) == 0) {
3408 		hostname = fnvlist_lookup_string(mos_config,
3409 		    ZPOOL_CONFIG_HOSTNAME);
3410 
3411 		myhostid = zone_get_hostid(NULL);
3412 
3413 		if (hostid != 0 && myhostid != 0 && hostid != myhostid) {
3414 			cmn_err(CE_WARN, "pool '%s' could not be "
3415 			    "loaded as it was last accessed by "
3416 			    "another system (host: %s hostid: 0x%llx). "
3417 			    "See: https://openzfs.github.io/openzfs-docs/msg/"
3418 			    "ZFS-8000-EY",
3419 			    spa_name(spa), hostname, (u_longlong_t)hostid);
3420 			spa_load_failed(spa, "hostid verification failed: pool "
3421 			    "last accessed by host: %s (hostid: 0x%llx)",
3422 			    hostname, (u_longlong_t)hostid);
3423 			return (SET_ERROR(EBADF));
3424 		}
3425 	}
3426 
3427 	return (0);
3428 }
3429 
3430 static int
3431 spa_ld_parse_config(spa_t *spa, spa_import_type_t type)
3432 {
3433 	int error = 0;
3434 	nvlist_t *nvtree, *nvl, *config = spa->spa_config;
3435 	int parse;
3436 	vdev_t *rvd;
3437 	uint64_t pool_guid;
3438 	const char *comment;
3439 	const char *compatibility;
3440 
3441 	/*
3442 	 * Versioning wasn't explicitly added to the label until later, so if
3443 	 * it's not present treat it as the initial version.
3444 	 */
3445 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION,
3446 	    &spa->spa_ubsync.ub_version) != 0)
3447 		spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL;
3448 
3449 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) {
3450 		spa_load_failed(spa, "invalid config provided: '%s' missing",
3451 		    ZPOOL_CONFIG_POOL_GUID);
3452 		return (SET_ERROR(EINVAL));
3453 	}
3454 
3455 	/*
3456 	 * If we are doing an import, ensure that the pool is not already
3457 	 * imported by checking if its pool guid already exists in the
3458 	 * spa namespace.
3459 	 *
3460 	 * The only case that we allow an already imported pool to be
3461 	 * imported again, is when the pool is checkpointed and we want to
3462 	 * look at its checkpointed state from userland tools like zdb.
3463 	 */
3464 #ifdef _KERNEL
3465 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
3466 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
3467 	    spa_guid_exists(pool_guid, 0)) {
3468 #else
3469 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
3470 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
3471 	    spa_guid_exists(pool_guid, 0) &&
3472 	    !spa_importing_readonly_checkpoint(spa)) {
3473 #endif
3474 		spa_load_failed(spa, "a pool with guid %llu is already open",
3475 		    (u_longlong_t)pool_guid);
3476 		return (SET_ERROR(EEXIST));
3477 	}
3478 
3479 	spa->spa_config_guid = pool_guid;
3480 
3481 	nvlist_free(spa->spa_load_info);
3482 	spa->spa_load_info = fnvlist_alloc();
3483 
3484 	ASSERT(spa->spa_comment == NULL);
3485 	if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0)
3486 		spa->spa_comment = spa_strdup(comment);
3487 
3488 	ASSERT(spa->spa_compatibility == NULL);
3489 	if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMPATIBILITY,
3490 	    &compatibility) == 0)
3491 		spa->spa_compatibility = spa_strdup(compatibility);
3492 
3493 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
3494 	    &spa->spa_config_txg);
3495 
3496 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0)
3497 		spa->spa_config_splitting = fnvlist_dup(nvl);
3498 
3499 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) {
3500 		spa_load_failed(spa, "invalid config provided: '%s' missing",
3501 		    ZPOOL_CONFIG_VDEV_TREE);
3502 		return (SET_ERROR(EINVAL));
3503 	}
3504 
3505 	/*
3506 	 * Create "The Godfather" zio to hold all async IOs
3507 	 */
3508 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
3509 	    KM_SLEEP);
3510 	for (int i = 0; i < max_ncpus; i++) {
3511 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
3512 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3513 		    ZIO_FLAG_GODFATHER);
3514 	}
3515 
3516 	/*
3517 	 * Parse the configuration into a vdev tree.  We explicitly set the
3518 	 * value that will be returned by spa_version() since parsing the
3519 	 * configuration requires knowing the version number.
3520 	 */
3521 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3522 	parse = (type == SPA_IMPORT_EXISTING ?
3523 	    VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT);
3524 	error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse);
3525 	spa_config_exit(spa, SCL_ALL, FTAG);
3526 
3527 	if (error != 0) {
3528 		spa_load_failed(spa, "unable to parse config [error=%d]",
3529 		    error);
3530 		return (error);
3531 	}
3532 
3533 	ASSERT(spa->spa_root_vdev == rvd);
3534 	ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
3535 	ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT);
3536 
3537 	if (type != SPA_IMPORT_ASSEMBLE) {
3538 		ASSERT(spa_guid(spa) == pool_guid);
3539 	}
3540 
3541 	return (0);
3542 }
3543 
3544 /*
3545  * Recursively open all vdevs in the vdev tree. This function is called twice:
3546  * first with the untrusted config, then with the trusted config.
3547  */
3548 static int
3549 spa_ld_open_vdevs(spa_t *spa)
3550 {
3551 	int error = 0;
3552 
3553 	/*
3554 	 * spa_missing_tvds_allowed defines how many top-level vdevs can be
3555 	 * missing/unopenable for the root vdev to be still considered openable.
3556 	 */
3557 	if (spa->spa_trust_config) {
3558 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds;
3559 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) {
3560 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile;
3561 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) {
3562 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan;
3563 	} else {
3564 		spa->spa_missing_tvds_allowed = 0;
3565 	}
3566 
3567 	spa->spa_missing_tvds_allowed =
3568 	    MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed);
3569 
3570 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3571 	error = vdev_open(spa->spa_root_vdev);
3572 	spa_config_exit(spa, SCL_ALL, FTAG);
3573 
3574 	if (spa->spa_missing_tvds != 0) {
3575 		spa_load_note(spa, "vdev tree has %lld missing top-level "
3576 		    "vdevs.", (u_longlong_t)spa->spa_missing_tvds);
3577 		if (spa->spa_trust_config && (spa->spa_mode & SPA_MODE_WRITE)) {
3578 			/*
3579 			 * Although theoretically we could allow users to open
3580 			 * incomplete pools in RW mode, we'd need to add a lot
3581 			 * of extra logic (e.g. adjust pool space to account
3582 			 * for missing vdevs).
3583 			 * This limitation also prevents users from accidentally
3584 			 * opening the pool in RW mode during data recovery and
3585 			 * damaging it further.
3586 			 */
3587 			spa_load_note(spa, "pools with missing top-level "
3588 			    "vdevs can only be opened in read-only mode.");
3589 			error = SET_ERROR(ENXIO);
3590 		} else {
3591 			spa_load_note(spa, "current settings allow for maximum "
3592 			    "%lld missing top-level vdevs at this stage.",
3593 			    (u_longlong_t)spa->spa_missing_tvds_allowed);
3594 		}
3595 	}
3596 	if (error != 0) {
3597 		spa_load_failed(spa, "unable to open vdev tree [error=%d]",
3598 		    error);
3599 	}
3600 	if (spa->spa_missing_tvds != 0 || error != 0)
3601 		vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2);
3602 
3603 	return (error);
3604 }
3605 
3606 /*
3607  * We need to validate the vdev labels against the configuration that
3608  * we have in hand. This function is called twice: first with an untrusted
3609  * config, then with a trusted config. The validation is more strict when the
3610  * config is trusted.
3611  */
3612 static int
3613 spa_ld_validate_vdevs(spa_t *spa)
3614 {
3615 	int error = 0;
3616 	vdev_t *rvd = spa->spa_root_vdev;
3617 
3618 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3619 	error = vdev_validate(rvd);
3620 	spa_config_exit(spa, SCL_ALL, FTAG);
3621 
3622 	if (error != 0) {
3623 		spa_load_failed(spa, "vdev_validate failed [error=%d]", error);
3624 		return (error);
3625 	}
3626 
3627 	if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) {
3628 		spa_load_failed(spa, "cannot open vdev tree after invalidating "
3629 		    "some vdevs");
3630 		vdev_dbgmsg_print_tree(rvd, 2);
3631 		return (SET_ERROR(ENXIO));
3632 	}
3633 
3634 	return (0);
3635 }
3636 
3637 static void
3638 spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub)
3639 {
3640 	spa->spa_state = POOL_STATE_ACTIVE;
3641 	spa->spa_ubsync = spa->spa_uberblock;
3642 	spa->spa_verify_min_txg = spa->spa_extreme_rewind ?
3643 	    TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1;
3644 	spa->spa_first_txg = spa->spa_last_ubsync_txg ?
3645 	    spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1;
3646 	spa->spa_claim_max_txg = spa->spa_first_txg;
3647 	spa->spa_prev_software_version = ub->ub_software_version;
3648 }
3649 
3650 static int
3651 spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type)
3652 {
3653 	vdev_t *rvd = spa->spa_root_vdev;
3654 	nvlist_t *label;
3655 	uberblock_t *ub = &spa->spa_uberblock;
3656 	boolean_t activity_check = B_FALSE;
3657 
3658 	/*
3659 	 * If we are opening the checkpointed state of the pool by
3660 	 * rewinding to it, at this point we will have written the
3661 	 * checkpointed uberblock to the vdev labels, so searching
3662 	 * the labels will find the right uberblock.  However, if
3663 	 * we are opening the checkpointed state read-only, we have
3664 	 * not modified the labels. Therefore, we must ignore the
3665 	 * labels and continue using the spa_uberblock that was set
3666 	 * by spa_ld_checkpoint_rewind.
3667 	 *
3668 	 * Note that it would be fine to ignore the labels when
3669 	 * rewinding (opening writeable) as well. However, if we
3670 	 * crash just after writing the labels, we will end up
3671 	 * searching the labels. Doing so in the common case means
3672 	 * that this code path gets exercised normally, rather than
3673 	 * just in the edge case.
3674 	 */
3675 	if (ub->ub_checkpoint_txg != 0 &&
3676 	    spa_importing_readonly_checkpoint(spa)) {
3677 		spa_ld_select_uberblock_done(spa, ub);
3678 		return (0);
3679 	}
3680 
3681 	/*
3682 	 * Find the best uberblock.
3683 	 */
3684 	vdev_uberblock_load(rvd, ub, &label);
3685 
3686 	/*
3687 	 * If we weren't able to find a single valid uberblock, return failure.
3688 	 */
3689 	if (ub->ub_txg == 0) {
3690 		nvlist_free(label);
3691 		spa_load_failed(spa, "no valid uberblock found");
3692 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO));
3693 	}
3694 
3695 	if (spa->spa_load_max_txg != UINT64_MAX) {
3696 		(void) spa_import_progress_set_max_txg(spa_guid(spa),
3697 		    (u_longlong_t)spa->spa_load_max_txg);
3698 	}
3699 	spa_load_note(spa, "using uberblock with txg=%llu",
3700 	    (u_longlong_t)ub->ub_txg);
3701 
3702 
3703 	/*
3704 	 * For pools which have the multihost property on determine if the
3705 	 * pool is truly inactive and can be safely imported.  Prevent
3706 	 * hosts which don't have a hostid set from importing the pool.
3707 	 */
3708 	activity_check = spa_activity_check_required(spa, ub, label,
3709 	    spa->spa_config);
3710 	if (activity_check) {
3711 		if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay &&
3712 		    spa_get_hostid(spa) == 0) {
3713 			nvlist_free(label);
3714 			fnvlist_add_uint64(spa->spa_load_info,
3715 			    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
3716 			return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
3717 		}
3718 
3719 		int error = spa_activity_check(spa, ub, spa->spa_config);
3720 		if (error) {
3721 			nvlist_free(label);
3722 			return (error);
3723 		}
3724 
3725 		fnvlist_add_uint64(spa->spa_load_info,
3726 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE);
3727 		fnvlist_add_uint64(spa->spa_load_info,
3728 		    ZPOOL_CONFIG_MMP_TXG, ub->ub_txg);
3729 		fnvlist_add_uint16(spa->spa_load_info,
3730 		    ZPOOL_CONFIG_MMP_SEQ,
3731 		    (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0));
3732 	}
3733 
3734 	/*
3735 	 * If the pool has an unsupported version we can't open it.
3736 	 */
3737 	if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) {
3738 		nvlist_free(label);
3739 		spa_load_failed(spa, "version %llu is not supported",
3740 		    (u_longlong_t)ub->ub_version);
3741 		return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP));
3742 	}
3743 
3744 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
3745 		nvlist_t *features;
3746 
3747 		/*
3748 		 * If we weren't able to find what's necessary for reading the
3749 		 * MOS in the label, return failure.
3750 		 */
3751 		if (label == NULL) {
3752 			spa_load_failed(spa, "label config unavailable");
3753 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3754 			    ENXIO));
3755 		}
3756 
3757 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ,
3758 		    &features) != 0) {
3759 			nvlist_free(label);
3760 			spa_load_failed(spa, "invalid label: '%s' missing",
3761 			    ZPOOL_CONFIG_FEATURES_FOR_READ);
3762 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3763 			    ENXIO));
3764 		}
3765 
3766 		/*
3767 		 * Update our in-core representation with the definitive values
3768 		 * from the label.
3769 		 */
3770 		nvlist_free(spa->spa_label_features);
3771 		spa->spa_label_features = fnvlist_dup(features);
3772 	}
3773 
3774 	nvlist_free(label);
3775 
3776 	/*
3777 	 * Look through entries in the label nvlist's features_for_read. If
3778 	 * there is a feature listed there which we don't understand then we
3779 	 * cannot open a pool.
3780 	 */
3781 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
3782 		nvlist_t *unsup_feat;
3783 
3784 		unsup_feat = fnvlist_alloc();
3785 
3786 		for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features,
3787 		    NULL); nvp != NULL;
3788 		    nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) {
3789 			if (!zfeature_is_supported(nvpair_name(nvp))) {
3790 				fnvlist_add_string(unsup_feat,
3791 				    nvpair_name(nvp), "");
3792 			}
3793 		}
3794 
3795 		if (!nvlist_empty(unsup_feat)) {
3796 			fnvlist_add_nvlist(spa->spa_load_info,
3797 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat);
3798 			nvlist_free(unsup_feat);
3799 			spa_load_failed(spa, "some features are unsupported");
3800 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
3801 			    ENOTSUP));
3802 		}
3803 
3804 		nvlist_free(unsup_feat);
3805 	}
3806 
3807 	if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) {
3808 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3809 		spa_try_repair(spa, spa->spa_config);
3810 		spa_config_exit(spa, SCL_ALL, FTAG);
3811 		nvlist_free(spa->spa_config_splitting);
3812 		spa->spa_config_splitting = NULL;
3813 	}
3814 
3815 	/*
3816 	 * Initialize internal SPA structures.
3817 	 */
3818 	spa_ld_select_uberblock_done(spa, ub);
3819 
3820 	return (0);
3821 }
3822 
3823 static int
3824 spa_ld_open_rootbp(spa_t *spa)
3825 {
3826 	int error = 0;
3827 	vdev_t *rvd = spa->spa_root_vdev;
3828 
3829 	error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool);
3830 	if (error != 0) {
3831 		spa_load_failed(spa, "unable to open rootbp in dsl_pool_init "
3832 		    "[error=%d]", error);
3833 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3834 	}
3835 	spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset;
3836 
3837 	return (0);
3838 }
3839 
3840 static int
3841 spa_ld_trusted_config(spa_t *spa, spa_import_type_t type,
3842     boolean_t reloading)
3843 {
3844 	vdev_t *mrvd, *rvd = spa->spa_root_vdev;
3845 	nvlist_t *nv, *mos_config, *policy;
3846 	int error = 0, copy_error;
3847 	uint64_t healthy_tvds, healthy_tvds_mos;
3848 	uint64_t mos_config_txg;
3849 
3850 	if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE)
3851 	    != 0)
3852 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3853 
3854 	/*
3855 	 * If we're assembling a pool from a split, the config provided is
3856 	 * already trusted so there is nothing to do.
3857 	 */
3858 	if (type == SPA_IMPORT_ASSEMBLE)
3859 		return (0);
3860 
3861 	healthy_tvds = spa_healthy_core_tvds(spa);
3862 
3863 	if (load_nvlist(spa, spa->spa_config_object, &mos_config)
3864 	    != 0) {
3865 		spa_load_failed(spa, "unable to retrieve MOS config");
3866 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3867 	}
3868 
3869 	/*
3870 	 * If we are doing an open, pool owner wasn't verified yet, thus do
3871 	 * the verification here.
3872 	 */
3873 	if (spa->spa_load_state == SPA_LOAD_OPEN) {
3874 		error = spa_verify_host(spa, mos_config);
3875 		if (error != 0) {
3876 			nvlist_free(mos_config);
3877 			return (error);
3878 		}
3879 	}
3880 
3881 	nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE);
3882 
3883 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3884 
3885 	/*
3886 	 * Build a new vdev tree from the trusted config
3887 	 */
3888 	error = spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD);
3889 	if (error != 0) {
3890 		nvlist_free(mos_config);
3891 		spa_config_exit(spa, SCL_ALL, FTAG);
3892 		spa_load_failed(spa, "spa_config_parse failed [error=%d]",
3893 		    error);
3894 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
3895 	}
3896 
3897 	/*
3898 	 * Vdev paths in the MOS may be obsolete. If the untrusted config was
3899 	 * obtained by scanning /dev/dsk, then it will have the right vdev
3900 	 * paths. We update the trusted MOS config with this information.
3901 	 * We first try to copy the paths with vdev_copy_path_strict, which
3902 	 * succeeds only when both configs have exactly the same vdev tree.
3903 	 * If that fails, we fall back to a more flexible method that has a
3904 	 * best effort policy.
3905 	 */
3906 	copy_error = vdev_copy_path_strict(rvd, mrvd);
3907 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3908 		spa_load_note(spa, "provided vdev tree:");
3909 		vdev_dbgmsg_print_tree(rvd, 2);
3910 		spa_load_note(spa, "MOS vdev tree:");
3911 		vdev_dbgmsg_print_tree(mrvd, 2);
3912 	}
3913 	if (copy_error != 0) {
3914 		spa_load_note(spa, "vdev_copy_path_strict failed, falling "
3915 		    "back to vdev_copy_path_relaxed");
3916 		vdev_copy_path_relaxed(rvd, mrvd);
3917 	}
3918 
3919 	vdev_close(rvd);
3920 	vdev_free(rvd);
3921 	spa->spa_root_vdev = mrvd;
3922 	rvd = mrvd;
3923 	spa_config_exit(spa, SCL_ALL, FTAG);
3924 
3925 	/*
3926 	 * We will use spa_config if we decide to reload the spa or if spa_load
3927 	 * fails and we rewind. We must thus regenerate the config using the
3928 	 * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to
3929 	 * pass settings on how to load the pool and is not stored in the MOS.
3930 	 * We copy it over to our new, trusted config.
3931 	 */
3932 	mos_config_txg = fnvlist_lookup_uint64(mos_config,
3933 	    ZPOOL_CONFIG_POOL_TXG);
3934 	nvlist_free(mos_config);
3935 	mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE);
3936 	if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY,
3937 	    &policy) == 0)
3938 		fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy);
3939 	spa_config_set(spa, mos_config);
3940 	spa->spa_config_source = SPA_CONFIG_SRC_MOS;
3941 
3942 	/*
3943 	 * Now that we got the config from the MOS, we should be more strict
3944 	 * in checking blkptrs and can make assumptions about the consistency
3945 	 * of the vdev tree. spa_trust_config must be set to true before opening
3946 	 * vdevs in order for them to be writeable.
3947 	 */
3948 	spa->spa_trust_config = B_TRUE;
3949 
3950 	/*
3951 	 * Open and validate the new vdev tree
3952 	 */
3953 	error = spa_ld_open_vdevs(spa);
3954 	if (error != 0)
3955 		return (error);
3956 
3957 	error = spa_ld_validate_vdevs(spa);
3958 	if (error != 0)
3959 		return (error);
3960 
3961 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3962 		spa_load_note(spa, "final vdev tree:");
3963 		vdev_dbgmsg_print_tree(rvd, 2);
3964 	}
3965 
3966 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT &&
3967 	    !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) {
3968 		/*
3969 		 * Sanity check to make sure that we are indeed loading the
3970 		 * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds
3971 		 * in the config provided and they happened to be the only ones
3972 		 * to have the latest uberblock, we could involuntarily perform
3973 		 * an extreme rewind.
3974 		 */
3975 		healthy_tvds_mos = spa_healthy_core_tvds(spa);
3976 		if (healthy_tvds_mos - healthy_tvds >=
3977 		    SPA_SYNC_MIN_VDEVS) {
3978 			spa_load_note(spa, "config provided misses too many "
3979 			    "top-level vdevs compared to MOS (%lld vs %lld). ",
3980 			    (u_longlong_t)healthy_tvds,
3981 			    (u_longlong_t)healthy_tvds_mos);
3982 			spa_load_note(spa, "vdev tree:");
3983 			vdev_dbgmsg_print_tree(rvd, 2);
3984 			if (reloading) {
3985 				spa_load_failed(spa, "config was already "
3986 				    "provided from MOS. Aborting.");
3987 				return (spa_vdev_err(rvd,
3988 				    VDEV_AUX_CORRUPT_DATA, EIO));
3989 			}
3990 			spa_load_note(spa, "spa must be reloaded using MOS "
3991 			    "config");
3992 			return (SET_ERROR(EAGAIN));
3993 		}
3994 	}
3995 
3996 	error = spa_check_for_missing_logs(spa);
3997 	if (error != 0)
3998 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO));
3999 
4000 	if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) {
4001 		spa_load_failed(spa, "uberblock guid sum doesn't match MOS "
4002 		    "guid sum (%llu != %llu)",
4003 		    (u_longlong_t)spa->spa_uberblock.ub_guid_sum,
4004 		    (u_longlong_t)rvd->vdev_guid_sum);
4005 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM,
4006 		    ENXIO));
4007 	}
4008 
4009 	return (0);
4010 }
4011 
4012 static int
4013 spa_ld_open_indirect_vdev_metadata(spa_t *spa)
4014 {
4015 	int error = 0;
4016 	vdev_t *rvd = spa->spa_root_vdev;
4017 
4018 	/*
4019 	 * Everything that we read before spa_remove_init() must be stored
4020 	 * on concreted vdevs.  Therefore we do this as early as possible.
4021 	 */
4022 	error = spa_remove_init(spa);
4023 	if (error != 0) {
4024 		spa_load_failed(spa, "spa_remove_init failed [error=%d]",
4025 		    error);
4026 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4027 	}
4028 
4029 	/*
4030 	 * Retrieve information needed to condense indirect vdev mappings.
4031 	 */
4032 	error = spa_condense_init(spa);
4033 	if (error != 0) {
4034 		spa_load_failed(spa, "spa_condense_init failed [error=%d]",
4035 		    error);
4036 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4037 	}
4038 
4039 	return (0);
4040 }
4041 
4042 static int
4043 spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep)
4044 {
4045 	int error = 0;
4046 	vdev_t *rvd = spa->spa_root_vdev;
4047 
4048 	if (spa_version(spa) >= SPA_VERSION_FEATURES) {
4049 		boolean_t missing_feat_read = B_FALSE;
4050 		nvlist_t *unsup_feat, *enabled_feat;
4051 
4052 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ,
4053 		    &spa->spa_feat_for_read_obj, B_TRUE) != 0) {
4054 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4055 		}
4056 
4057 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE,
4058 		    &spa->spa_feat_for_write_obj, B_TRUE) != 0) {
4059 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4060 		}
4061 
4062 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS,
4063 		    &spa->spa_feat_desc_obj, B_TRUE) != 0) {
4064 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4065 		}
4066 
4067 		enabled_feat = fnvlist_alloc();
4068 		unsup_feat = fnvlist_alloc();
4069 
4070 		if (!spa_features_check(spa, B_FALSE,
4071 		    unsup_feat, enabled_feat))
4072 			missing_feat_read = B_TRUE;
4073 
4074 		if (spa_writeable(spa) ||
4075 		    spa->spa_load_state == SPA_LOAD_TRYIMPORT) {
4076 			if (!spa_features_check(spa, B_TRUE,
4077 			    unsup_feat, enabled_feat)) {
4078 				*missing_feat_writep = B_TRUE;
4079 			}
4080 		}
4081 
4082 		fnvlist_add_nvlist(spa->spa_load_info,
4083 		    ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat);
4084 
4085 		if (!nvlist_empty(unsup_feat)) {
4086 			fnvlist_add_nvlist(spa->spa_load_info,
4087 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat);
4088 		}
4089 
4090 		fnvlist_free(enabled_feat);
4091 		fnvlist_free(unsup_feat);
4092 
4093 		if (!missing_feat_read) {
4094 			fnvlist_add_boolean(spa->spa_load_info,
4095 			    ZPOOL_CONFIG_CAN_RDONLY);
4096 		}
4097 
4098 		/*
4099 		 * If the state is SPA_LOAD_TRYIMPORT, our objective is
4100 		 * twofold: to determine whether the pool is available for
4101 		 * import in read-write mode and (if it is not) whether the
4102 		 * pool is available for import in read-only mode. If the pool
4103 		 * is available for import in read-write mode, it is displayed
4104 		 * as available in userland; if it is not available for import
4105 		 * in read-only mode, it is displayed as unavailable in
4106 		 * userland. If the pool is available for import in read-only
4107 		 * mode but not read-write mode, it is displayed as unavailable
4108 		 * in userland with a special note that the pool is actually
4109 		 * available for open in read-only mode.
4110 		 *
4111 		 * As a result, if the state is SPA_LOAD_TRYIMPORT and we are
4112 		 * missing a feature for write, we must first determine whether
4113 		 * the pool can be opened read-only before returning to
4114 		 * userland in order to know whether to display the
4115 		 * abovementioned note.
4116 		 */
4117 		if (missing_feat_read || (*missing_feat_writep &&
4118 		    spa_writeable(spa))) {
4119 			spa_load_failed(spa, "pool uses unsupported features");
4120 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
4121 			    ENOTSUP));
4122 		}
4123 
4124 		/*
4125 		 * Load refcounts for ZFS features from disk into an in-memory
4126 		 * cache during SPA initialization.
4127 		 */
4128 		for (spa_feature_t i = 0; i < SPA_FEATURES; i++) {
4129 			uint64_t refcount;
4130 
4131 			error = feature_get_refcount_from_disk(spa,
4132 			    &spa_feature_table[i], &refcount);
4133 			if (error == 0) {
4134 				spa->spa_feat_refcount_cache[i] = refcount;
4135 			} else if (error == ENOTSUP) {
4136 				spa->spa_feat_refcount_cache[i] =
4137 				    SPA_FEATURE_DISABLED;
4138 			} else {
4139 				spa_load_failed(spa, "error getting refcount "
4140 				    "for feature %s [error=%d]",
4141 				    spa_feature_table[i].fi_guid, error);
4142 				return (spa_vdev_err(rvd,
4143 				    VDEV_AUX_CORRUPT_DATA, EIO));
4144 			}
4145 		}
4146 	}
4147 
4148 	if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) {
4149 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG,
4150 		    &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0)
4151 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4152 	}
4153 
4154 	/*
4155 	 * Encryption was added before bookmark_v2, even though bookmark_v2
4156 	 * is now a dependency. If this pool has encryption enabled without
4157 	 * bookmark_v2, trigger an errata message.
4158 	 */
4159 	if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) &&
4160 	    !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
4161 		spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION;
4162 	}
4163 
4164 	return (0);
4165 }
4166 
4167 static int
4168 spa_ld_load_special_directories(spa_t *spa)
4169 {
4170 	int error = 0;
4171 	vdev_t *rvd = spa->spa_root_vdev;
4172 
4173 	spa->spa_is_initializing = B_TRUE;
4174 	error = dsl_pool_open(spa->spa_dsl_pool);
4175 	spa->spa_is_initializing = B_FALSE;
4176 	if (error != 0) {
4177 		spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error);
4178 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4179 	}
4180 
4181 	return (0);
4182 }
4183 
4184 static int
4185 spa_ld_get_props(spa_t *spa)
4186 {
4187 	int error = 0;
4188 	uint64_t obj;
4189 	vdev_t *rvd = spa->spa_root_vdev;
4190 
4191 	/* Grab the checksum salt from the MOS. */
4192 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4193 	    DMU_POOL_CHECKSUM_SALT, 1,
4194 	    sizeof (spa->spa_cksum_salt.zcs_bytes),
4195 	    spa->spa_cksum_salt.zcs_bytes);
4196 	if (error == ENOENT) {
4197 		/* Generate a new salt for subsequent use */
4198 		(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
4199 		    sizeof (spa->spa_cksum_salt.zcs_bytes));
4200 	} else if (error != 0) {
4201 		spa_load_failed(spa, "unable to retrieve checksum salt from "
4202 		    "MOS [error=%d]", error);
4203 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4204 	}
4205 
4206 	if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0)
4207 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4208 	error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj);
4209 	if (error != 0) {
4210 		spa_load_failed(spa, "error opening deferred-frees bpobj "
4211 		    "[error=%d]", error);
4212 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4213 	}
4214 
4215 	/*
4216 	 * Load the bit that tells us to use the new accounting function
4217 	 * (raid-z deflation).  If we have an older pool, this will not
4218 	 * be present.
4219 	 */
4220 	error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE);
4221 	if (error != 0 && error != ENOENT)
4222 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4223 
4224 	error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION,
4225 	    &spa->spa_creation_version, B_FALSE);
4226 	if (error != 0 && error != ENOENT)
4227 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4228 
4229 	/*
4230 	 * Load the persistent error log.  If we have an older pool, this will
4231 	 * not be present.
4232 	 */
4233 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last,
4234 	    B_FALSE);
4235 	if (error != 0 && error != ENOENT)
4236 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4237 
4238 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB,
4239 	    &spa->spa_errlog_scrub, B_FALSE);
4240 	if (error != 0 && error != ENOENT)
4241 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4242 
4243 	/*
4244 	 * Load the livelist deletion field. If a livelist is queued for
4245 	 * deletion, indicate that in the spa
4246 	 */
4247 	error = spa_dir_prop(spa, DMU_POOL_DELETED_CLONES,
4248 	    &spa->spa_livelists_to_delete, B_FALSE);
4249 	if (error != 0 && error != ENOENT)
4250 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4251 
4252 	/*
4253 	 * Load the history object.  If we have an older pool, this
4254 	 * will not be present.
4255 	 */
4256 	error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE);
4257 	if (error != 0 && error != ENOENT)
4258 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4259 
4260 	/*
4261 	 * Load the per-vdev ZAP map. If we have an older pool, this will not
4262 	 * be present; in this case, defer its creation to a later time to
4263 	 * avoid dirtying the MOS this early / out of sync context. See
4264 	 * spa_sync_config_object.
4265 	 */
4266 
4267 	/* The sentinel is only available in the MOS config. */
4268 	nvlist_t *mos_config;
4269 	if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) {
4270 		spa_load_failed(spa, "unable to retrieve MOS config");
4271 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4272 	}
4273 
4274 	error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP,
4275 	    &spa->spa_all_vdev_zaps, B_FALSE);
4276 
4277 	if (error == ENOENT) {
4278 		VERIFY(!nvlist_exists(mos_config,
4279 		    ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
4280 		spa->spa_avz_action = AVZ_ACTION_INITIALIZE;
4281 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
4282 	} else if (error != 0) {
4283 		nvlist_free(mos_config);
4284 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4285 	} else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) {
4286 		/*
4287 		 * An older version of ZFS overwrote the sentinel value, so
4288 		 * we have orphaned per-vdev ZAPs in the MOS. Defer their
4289 		 * destruction to later; see spa_sync_config_object.
4290 		 */
4291 		spa->spa_avz_action = AVZ_ACTION_DESTROY;
4292 		/*
4293 		 * We're assuming that no vdevs have had their ZAPs created
4294 		 * before this. Better be sure of it.
4295 		 */
4296 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
4297 	}
4298 	nvlist_free(mos_config);
4299 
4300 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
4301 
4302 	error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object,
4303 	    B_FALSE);
4304 	if (error && error != ENOENT)
4305 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4306 
4307 	if (error == 0) {
4308 		uint64_t autoreplace = 0;
4309 
4310 		spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs);
4311 		spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace);
4312 		spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation);
4313 		spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode);
4314 		spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand);
4315 		spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost);
4316 		spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim);
4317 		spa->spa_autoreplace = (autoreplace != 0);
4318 	}
4319 
4320 	/*
4321 	 * If we are importing a pool with missing top-level vdevs,
4322 	 * we enforce that the pool doesn't panic or get suspended on
4323 	 * error since the likelihood of missing data is extremely high.
4324 	 */
4325 	if (spa->spa_missing_tvds > 0 &&
4326 	    spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE &&
4327 	    spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4328 		spa_load_note(spa, "forcing failmode to 'continue' "
4329 		    "as some top level vdevs are missing");
4330 		spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE;
4331 	}
4332 
4333 	return (0);
4334 }
4335 
4336 static int
4337 spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type)
4338 {
4339 	int error = 0;
4340 	vdev_t *rvd = spa->spa_root_vdev;
4341 
4342 	/*
4343 	 * If we're assembling the pool from the split-off vdevs of
4344 	 * an existing pool, we don't want to attach the spares & cache
4345 	 * devices.
4346 	 */
4347 
4348 	/*
4349 	 * Load any hot spares for this pool.
4350 	 */
4351 	error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object,
4352 	    B_FALSE);
4353 	if (error != 0 && error != ENOENT)
4354 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4355 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
4356 		ASSERT(spa_version(spa) >= SPA_VERSION_SPARES);
4357 		if (load_nvlist(spa, spa->spa_spares.sav_object,
4358 		    &spa->spa_spares.sav_config) != 0) {
4359 			spa_load_failed(spa, "error loading spares nvlist");
4360 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4361 		}
4362 
4363 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4364 		spa_load_spares(spa);
4365 		spa_config_exit(spa, SCL_ALL, FTAG);
4366 	} else if (error == 0) {
4367 		spa->spa_spares.sav_sync = B_TRUE;
4368 	}
4369 
4370 	/*
4371 	 * Load any level 2 ARC devices for this pool.
4372 	 */
4373 	error = spa_dir_prop(spa, DMU_POOL_L2CACHE,
4374 	    &spa->spa_l2cache.sav_object, B_FALSE);
4375 	if (error != 0 && error != ENOENT)
4376 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4377 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
4378 		ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE);
4379 		if (load_nvlist(spa, spa->spa_l2cache.sav_object,
4380 		    &spa->spa_l2cache.sav_config) != 0) {
4381 			spa_load_failed(spa, "error loading l2cache nvlist");
4382 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4383 		}
4384 
4385 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4386 		spa_load_l2cache(spa);
4387 		spa_config_exit(spa, SCL_ALL, FTAG);
4388 	} else if (error == 0) {
4389 		spa->spa_l2cache.sav_sync = B_TRUE;
4390 	}
4391 
4392 	return (0);
4393 }
4394 
4395 static int
4396 spa_ld_load_vdev_metadata(spa_t *spa)
4397 {
4398 	int error = 0;
4399 	vdev_t *rvd = spa->spa_root_vdev;
4400 
4401 	/*
4402 	 * If the 'multihost' property is set, then never allow a pool to
4403 	 * be imported when the system hostid is zero.  The exception to
4404 	 * this rule is zdb which is always allowed to access pools.
4405 	 */
4406 	if (spa_multihost(spa) && spa_get_hostid(spa) == 0 &&
4407 	    (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) {
4408 		fnvlist_add_uint64(spa->spa_load_info,
4409 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
4410 		return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
4411 	}
4412 
4413 	/*
4414 	 * If the 'autoreplace' property is set, then post a resource notifying
4415 	 * the ZFS DE that it should not issue any faults for unopenable
4416 	 * devices.  We also iterate over the vdevs, and post a sysevent for any
4417 	 * unopenable vdevs so that the normal autoreplace handler can take
4418 	 * over.
4419 	 */
4420 	if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4421 		spa_check_removed(spa->spa_root_vdev);
4422 		/*
4423 		 * For the import case, this is done in spa_import(), because
4424 		 * at this point we're using the spare definitions from
4425 		 * the MOS config, not necessarily from the userland config.
4426 		 */
4427 		if (spa->spa_load_state != SPA_LOAD_IMPORT) {
4428 			spa_aux_check_removed(&spa->spa_spares);
4429 			spa_aux_check_removed(&spa->spa_l2cache);
4430 		}
4431 	}
4432 
4433 	/*
4434 	 * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc.
4435 	 */
4436 	error = vdev_load(rvd);
4437 	if (error != 0) {
4438 		spa_load_failed(spa, "vdev_load failed [error=%d]", error);
4439 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4440 	}
4441 
4442 	error = spa_ld_log_spacemaps(spa);
4443 	if (error != 0) {
4444 		spa_load_failed(spa, "spa_ld_log_spacemaps failed [error=%d]",
4445 		    error);
4446 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4447 	}
4448 
4449 	/*
4450 	 * Propagate the leaf DTLs we just loaded all the way up the vdev tree.
4451 	 */
4452 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4453 	vdev_dtl_reassess(rvd, 0, 0, B_FALSE, B_FALSE);
4454 	spa_config_exit(spa, SCL_ALL, FTAG);
4455 
4456 	return (0);
4457 }
4458 
4459 static int
4460 spa_ld_load_dedup_tables(spa_t *spa)
4461 {
4462 	int error = 0;
4463 	vdev_t *rvd = spa->spa_root_vdev;
4464 
4465 	error = ddt_load(spa);
4466 	if (error != 0) {
4467 		spa_load_failed(spa, "ddt_load failed [error=%d]", error);
4468 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4469 	}
4470 
4471 	return (0);
4472 }
4473 
4474 static int
4475 spa_ld_load_brt(spa_t *spa)
4476 {
4477 	int error = 0;
4478 	vdev_t *rvd = spa->spa_root_vdev;
4479 
4480 	error = brt_load(spa);
4481 	if (error != 0) {
4482 		spa_load_failed(spa, "brt_load failed [error=%d]", error);
4483 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4484 	}
4485 
4486 	return (0);
4487 }
4488 
4489 static int
4490 spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, const char **ereport)
4491 {
4492 	vdev_t *rvd = spa->spa_root_vdev;
4493 
4494 	if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) {
4495 		boolean_t missing = spa_check_logs(spa);
4496 		if (missing) {
4497 			if (spa->spa_missing_tvds != 0) {
4498 				spa_load_note(spa, "spa_check_logs failed "
4499 				    "so dropping the logs");
4500 			} else {
4501 				*ereport = FM_EREPORT_ZFS_LOG_REPLAY;
4502 				spa_load_failed(spa, "spa_check_logs failed");
4503 				return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG,
4504 				    ENXIO));
4505 			}
4506 		}
4507 	}
4508 
4509 	return (0);
4510 }
4511 
4512 static int
4513 spa_ld_verify_pool_data(spa_t *spa)
4514 {
4515 	int error = 0;
4516 	vdev_t *rvd = spa->spa_root_vdev;
4517 
4518 	/*
4519 	 * We've successfully opened the pool, verify that we're ready
4520 	 * to start pushing transactions.
4521 	 */
4522 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4523 		error = spa_load_verify(spa);
4524 		if (error != 0) {
4525 			spa_load_failed(spa, "spa_load_verify failed "
4526 			    "[error=%d]", error);
4527 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
4528 			    error));
4529 		}
4530 	}
4531 
4532 	return (0);
4533 }
4534 
4535 static void
4536 spa_ld_claim_log_blocks(spa_t *spa)
4537 {
4538 	dmu_tx_t *tx;
4539 	dsl_pool_t *dp = spa_get_dsl(spa);
4540 
4541 	/*
4542 	 * Claim log blocks that haven't been committed yet.
4543 	 * This must all happen in a single txg.
4544 	 * Note: spa_claim_max_txg is updated by spa_claim_notify(),
4545 	 * invoked from zil_claim_log_block()'s i/o done callback.
4546 	 * Price of rollback is that we abandon the log.
4547 	 */
4548 	spa->spa_claiming = B_TRUE;
4549 
4550 	tx = dmu_tx_create_assigned(dp, spa_first_txg(spa));
4551 	(void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
4552 	    zil_claim, tx, DS_FIND_CHILDREN);
4553 	dmu_tx_commit(tx);
4554 
4555 	spa->spa_claiming = B_FALSE;
4556 
4557 	spa_set_log_state(spa, SPA_LOG_GOOD);
4558 }
4559 
4560 static void
4561 spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg,
4562     boolean_t update_config_cache)
4563 {
4564 	vdev_t *rvd = spa->spa_root_vdev;
4565 	int need_update = B_FALSE;
4566 
4567 	/*
4568 	 * If the config cache is stale, or we have uninitialized
4569 	 * metaslabs (see spa_vdev_add()), then update the config.
4570 	 *
4571 	 * If this is a verbatim import, trust the current
4572 	 * in-core spa_config and update the disk labels.
4573 	 */
4574 	if (update_config_cache || config_cache_txg != spa->spa_config_txg ||
4575 	    spa->spa_load_state == SPA_LOAD_IMPORT ||
4576 	    spa->spa_load_state == SPA_LOAD_RECOVER ||
4577 	    (spa->spa_import_flags & ZFS_IMPORT_VERBATIM))
4578 		need_update = B_TRUE;
4579 
4580 	for (int c = 0; c < rvd->vdev_children; c++)
4581 		if (rvd->vdev_child[c]->vdev_ms_array == 0)
4582 			need_update = B_TRUE;
4583 
4584 	/*
4585 	 * Update the config cache asynchronously in case we're the
4586 	 * root pool, in which case the config cache isn't writable yet.
4587 	 */
4588 	if (need_update)
4589 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
4590 }
4591 
4592 static void
4593 spa_ld_prepare_for_reload(spa_t *spa)
4594 {
4595 	spa_mode_t mode = spa->spa_mode;
4596 	int async_suspended = spa->spa_async_suspended;
4597 
4598 	spa_unload(spa);
4599 	spa_deactivate(spa);
4600 	spa_activate(spa, mode);
4601 
4602 	/*
4603 	 * We save the value of spa_async_suspended as it gets reset to 0 by
4604 	 * spa_unload(). We want to restore it back to the original value before
4605 	 * returning as we might be calling spa_async_resume() later.
4606 	 */
4607 	spa->spa_async_suspended = async_suspended;
4608 }
4609 
4610 static int
4611 spa_ld_read_checkpoint_txg(spa_t *spa)
4612 {
4613 	uberblock_t checkpoint;
4614 	int error = 0;
4615 
4616 	ASSERT0(spa->spa_checkpoint_txg);
4617 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4618 
4619 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4620 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
4621 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
4622 
4623 	if (error == ENOENT)
4624 		return (0);
4625 
4626 	if (error != 0)
4627 		return (error);
4628 
4629 	ASSERT3U(checkpoint.ub_txg, !=, 0);
4630 	ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0);
4631 	ASSERT3U(checkpoint.ub_timestamp, !=, 0);
4632 	spa->spa_checkpoint_txg = checkpoint.ub_txg;
4633 	spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp;
4634 
4635 	return (0);
4636 }
4637 
4638 static int
4639 spa_ld_mos_init(spa_t *spa, spa_import_type_t type)
4640 {
4641 	int error = 0;
4642 
4643 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4644 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
4645 
4646 	/*
4647 	 * Never trust the config that is provided unless we are assembling
4648 	 * a pool following a split.
4649 	 * This means don't trust blkptrs and the vdev tree in general. This
4650 	 * also effectively puts the spa in read-only mode since
4651 	 * spa_writeable() checks for spa_trust_config to be true.
4652 	 * We will later load a trusted config from the MOS.
4653 	 */
4654 	if (type != SPA_IMPORT_ASSEMBLE)
4655 		spa->spa_trust_config = B_FALSE;
4656 
4657 	/*
4658 	 * Parse the config provided to create a vdev tree.
4659 	 */
4660 	error = spa_ld_parse_config(spa, type);
4661 	if (error != 0)
4662 		return (error);
4663 
4664 	spa_import_progress_add(spa);
4665 
4666 	/*
4667 	 * Now that we have the vdev tree, try to open each vdev. This involves
4668 	 * opening the underlying physical device, retrieving its geometry and
4669 	 * probing the vdev with a dummy I/O. The state of each vdev will be set
4670 	 * based on the success of those operations. After this we'll be ready
4671 	 * to read from the vdevs.
4672 	 */
4673 	error = spa_ld_open_vdevs(spa);
4674 	if (error != 0)
4675 		return (error);
4676 
4677 	/*
4678 	 * Read the label of each vdev and make sure that the GUIDs stored
4679 	 * there match the GUIDs in the config provided.
4680 	 * If we're assembling a new pool that's been split off from an
4681 	 * existing pool, the labels haven't yet been updated so we skip
4682 	 * validation for now.
4683 	 */
4684 	if (type != SPA_IMPORT_ASSEMBLE) {
4685 		error = spa_ld_validate_vdevs(spa);
4686 		if (error != 0)
4687 			return (error);
4688 	}
4689 
4690 	/*
4691 	 * Read all vdev labels to find the best uberblock (i.e. latest,
4692 	 * unless spa_load_max_txg is set) and store it in spa_uberblock. We
4693 	 * get the list of features required to read blkptrs in the MOS from
4694 	 * the vdev label with the best uberblock and verify that our version
4695 	 * of zfs supports them all.
4696 	 */
4697 	error = spa_ld_select_uberblock(spa, type);
4698 	if (error != 0)
4699 		return (error);
4700 
4701 	/*
4702 	 * Pass that uberblock to the dsl_pool layer which will open the root
4703 	 * blkptr. This blkptr points to the latest version of the MOS and will
4704 	 * allow us to read its contents.
4705 	 */
4706 	error = spa_ld_open_rootbp(spa);
4707 	if (error != 0)
4708 		return (error);
4709 
4710 	return (0);
4711 }
4712 
4713 static int
4714 spa_ld_checkpoint_rewind(spa_t *spa)
4715 {
4716 	uberblock_t checkpoint;
4717 	int error = 0;
4718 
4719 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4720 	ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4721 
4722 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4723 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
4724 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
4725 
4726 	if (error != 0) {
4727 		spa_load_failed(spa, "unable to retrieve checkpointed "
4728 		    "uberblock from the MOS config [error=%d]", error);
4729 
4730 		if (error == ENOENT)
4731 			error = ZFS_ERR_NO_CHECKPOINT;
4732 
4733 		return (error);
4734 	}
4735 
4736 	ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg);
4737 	ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg);
4738 
4739 	/*
4740 	 * We need to update the txg and timestamp of the checkpointed
4741 	 * uberblock to be higher than the latest one. This ensures that
4742 	 * the checkpointed uberblock is selected if we were to close and
4743 	 * reopen the pool right after we've written it in the vdev labels.
4744 	 * (also see block comment in vdev_uberblock_compare)
4745 	 */
4746 	checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1;
4747 	checkpoint.ub_timestamp = gethrestime_sec();
4748 
4749 	/*
4750 	 * Set current uberblock to be the checkpointed uberblock.
4751 	 */
4752 	spa->spa_uberblock = checkpoint;
4753 
4754 	/*
4755 	 * If we are doing a normal rewind, then the pool is open for
4756 	 * writing and we sync the "updated" checkpointed uberblock to
4757 	 * disk. Once this is done, we've basically rewound the whole
4758 	 * pool and there is no way back.
4759 	 *
4760 	 * There are cases when we don't want to attempt and sync the
4761 	 * checkpointed uberblock to disk because we are opening a
4762 	 * pool as read-only. Specifically, verifying the checkpointed
4763 	 * state with zdb, and importing the checkpointed state to get
4764 	 * a "preview" of its content.
4765 	 */
4766 	if (spa_writeable(spa)) {
4767 		vdev_t *rvd = spa->spa_root_vdev;
4768 
4769 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4770 		vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
4771 		int svdcount = 0;
4772 		int children = rvd->vdev_children;
4773 		int c0 = random_in_range(children);
4774 
4775 		for (int c = 0; c < children; c++) {
4776 			vdev_t *vd = rvd->vdev_child[(c0 + c) % children];
4777 
4778 			/* Stop when revisiting the first vdev */
4779 			if (c > 0 && svd[0] == vd)
4780 				break;
4781 
4782 			if (vd->vdev_ms_array == 0 || vd->vdev_islog ||
4783 			    !vdev_is_concrete(vd))
4784 				continue;
4785 
4786 			svd[svdcount++] = vd;
4787 			if (svdcount == SPA_SYNC_MIN_VDEVS)
4788 				break;
4789 		}
4790 		error = vdev_config_sync(svd, svdcount, spa->spa_first_txg);
4791 		if (error == 0)
4792 			spa->spa_last_synced_guid = rvd->vdev_guid;
4793 		spa_config_exit(spa, SCL_ALL, FTAG);
4794 
4795 		if (error != 0) {
4796 			spa_load_failed(spa, "failed to write checkpointed "
4797 			    "uberblock to the vdev labels [error=%d]", error);
4798 			return (error);
4799 		}
4800 	}
4801 
4802 	return (0);
4803 }
4804 
4805 static int
4806 spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type,
4807     boolean_t *update_config_cache)
4808 {
4809 	int error;
4810 
4811 	/*
4812 	 * Parse the config for pool, open and validate vdevs,
4813 	 * select an uberblock, and use that uberblock to open
4814 	 * the MOS.
4815 	 */
4816 	error = spa_ld_mos_init(spa, type);
4817 	if (error != 0)
4818 		return (error);
4819 
4820 	/*
4821 	 * Retrieve the trusted config stored in the MOS and use it to create
4822 	 * a new, exact version of the vdev tree, then reopen all vdevs.
4823 	 */
4824 	error = spa_ld_trusted_config(spa, type, B_FALSE);
4825 	if (error == EAGAIN) {
4826 		if (update_config_cache != NULL)
4827 			*update_config_cache = B_TRUE;
4828 
4829 		/*
4830 		 * Redo the loading process with the trusted config if it is
4831 		 * too different from the untrusted config.
4832 		 */
4833 		spa_ld_prepare_for_reload(spa);
4834 		spa_load_note(spa, "RELOADING");
4835 		error = spa_ld_mos_init(spa, type);
4836 		if (error != 0)
4837 			return (error);
4838 
4839 		error = spa_ld_trusted_config(spa, type, B_TRUE);
4840 		if (error != 0)
4841 			return (error);
4842 
4843 	} else if (error != 0) {
4844 		return (error);
4845 	}
4846 
4847 	return (0);
4848 }
4849 
4850 /*
4851  * Load an existing storage pool, using the config provided. This config
4852  * describes which vdevs are part of the pool and is later validated against
4853  * partial configs present in each vdev's label and an entire copy of the
4854  * config stored in the MOS.
4855  */
4856 static int
4857 spa_load_impl(spa_t *spa, spa_import_type_t type, const char **ereport)
4858 {
4859 	int error = 0;
4860 	boolean_t missing_feat_write = B_FALSE;
4861 	boolean_t checkpoint_rewind =
4862 	    (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4863 	boolean_t update_config_cache = B_FALSE;
4864 
4865 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4866 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
4867 
4868 	spa_load_note(spa, "LOADING");
4869 
4870 	error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache);
4871 	if (error != 0)
4872 		return (error);
4873 
4874 	/*
4875 	 * If we are rewinding to the checkpoint then we need to repeat
4876 	 * everything we've done so far in this function but this time
4877 	 * selecting the checkpointed uberblock and using that to open
4878 	 * the MOS.
4879 	 */
4880 	if (checkpoint_rewind) {
4881 		/*
4882 		 * If we are rewinding to the checkpoint update config cache
4883 		 * anyway.
4884 		 */
4885 		update_config_cache = B_TRUE;
4886 
4887 		/*
4888 		 * Extract the checkpointed uberblock from the current MOS
4889 		 * and use this as the pool's uberblock from now on. If the
4890 		 * pool is imported as writeable we also write the checkpoint
4891 		 * uberblock to the labels, making the rewind permanent.
4892 		 */
4893 		error = spa_ld_checkpoint_rewind(spa);
4894 		if (error != 0)
4895 			return (error);
4896 
4897 		/*
4898 		 * Redo the loading process again with the
4899 		 * checkpointed uberblock.
4900 		 */
4901 		spa_ld_prepare_for_reload(spa);
4902 		spa_load_note(spa, "LOADING checkpointed uberblock");
4903 		error = spa_ld_mos_with_trusted_config(spa, type, NULL);
4904 		if (error != 0)
4905 			return (error);
4906 	}
4907 
4908 	/*
4909 	 * Retrieve the checkpoint txg if the pool has a checkpoint.
4910 	 */
4911 	error = spa_ld_read_checkpoint_txg(spa);
4912 	if (error != 0)
4913 		return (error);
4914 
4915 	/*
4916 	 * Retrieve the mapping of indirect vdevs. Those vdevs were removed
4917 	 * from the pool and their contents were re-mapped to other vdevs. Note
4918 	 * that everything that we read before this step must have been
4919 	 * rewritten on concrete vdevs after the last device removal was
4920 	 * initiated. Otherwise we could be reading from indirect vdevs before
4921 	 * we have loaded their mappings.
4922 	 */
4923 	error = spa_ld_open_indirect_vdev_metadata(spa);
4924 	if (error != 0)
4925 		return (error);
4926 
4927 	/*
4928 	 * Retrieve the full list of active features from the MOS and check if
4929 	 * they are all supported.
4930 	 */
4931 	error = spa_ld_check_features(spa, &missing_feat_write);
4932 	if (error != 0)
4933 		return (error);
4934 
4935 	/*
4936 	 * Load several special directories from the MOS needed by the dsl_pool
4937 	 * layer.
4938 	 */
4939 	error = spa_ld_load_special_directories(spa);
4940 	if (error != 0)
4941 		return (error);
4942 
4943 	/*
4944 	 * Retrieve pool properties from the MOS.
4945 	 */
4946 	error = spa_ld_get_props(spa);
4947 	if (error != 0)
4948 		return (error);
4949 
4950 	/*
4951 	 * Retrieve the list of auxiliary devices - cache devices and spares -
4952 	 * and open them.
4953 	 */
4954 	error = spa_ld_open_aux_vdevs(spa, type);
4955 	if (error != 0)
4956 		return (error);
4957 
4958 	/*
4959 	 * Load the metadata for all vdevs. Also check if unopenable devices
4960 	 * should be autoreplaced.
4961 	 */
4962 	error = spa_ld_load_vdev_metadata(spa);
4963 	if (error != 0)
4964 		return (error);
4965 
4966 	error = spa_ld_load_dedup_tables(spa);
4967 	if (error != 0)
4968 		return (error);
4969 
4970 	error = spa_ld_load_brt(spa);
4971 	if (error != 0)
4972 		return (error);
4973 
4974 	/*
4975 	 * Verify the logs now to make sure we don't have any unexpected errors
4976 	 * when we claim log blocks later.
4977 	 */
4978 	error = spa_ld_verify_logs(spa, type, ereport);
4979 	if (error != 0)
4980 		return (error);
4981 
4982 	if (missing_feat_write) {
4983 		ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT);
4984 
4985 		/*
4986 		 * At this point, we know that we can open the pool in
4987 		 * read-only mode but not read-write mode. We now have enough
4988 		 * information and can return to userland.
4989 		 */
4990 		return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT,
4991 		    ENOTSUP));
4992 	}
4993 
4994 	/*
4995 	 * Traverse the last txgs to make sure the pool was left off in a safe
4996 	 * state. When performing an extreme rewind, we verify the whole pool,
4997 	 * which can take a very long time.
4998 	 */
4999 	error = spa_ld_verify_pool_data(spa);
5000 	if (error != 0)
5001 		return (error);
5002 
5003 	/*
5004 	 * Calculate the deflated space for the pool. This must be done before
5005 	 * we write anything to the pool because we'd need to update the space
5006 	 * accounting using the deflated sizes.
5007 	 */
5008 	spa_update_dspace(spa);
5009 
5010 	/*
5011 	 * We have now retrieved all the information we needed to open the
5012 	 * pool. If we are importing the pool in read-write mode, a few
5013 	 * additional steps must be performed to finish the import.
5014 	 */
5015 	if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER ||
5016 	    spa->spa_load_max_txg == UINT64_MAX)) {
5017 		uint64_t config_cache_txg = spa->spa_config_txg;
5018 
5019 		ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT);
5020 
5021 		/*
5022 		 * In case of a checkpoint rewind, log the original txg
5023 		 * of the checkpointed uberblock.
5024 		 */
5025 		if (checkpoint_rewind) {
5026 			spa_history_log_internal(spa, "checkpoint rewind",
5027 			    NULL, "rewound state to txg=%llu",
5028 			    (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg);
5029 		}
5030 
5031 		/*
5032 		 * Traverse the ZIL and claim all blocks.
5033 		 */
5034 		spa_ld_claim_log_blocks(spa);
5035 
5036 		/*
5037 		 * Kick-off the syncing thread.
5038 		 */
5039 		spa->spa_sync_on = B_TRUE;
5040 		txg_sync_start(spa->spa_dsl_pool);
5041 		mmp_thread_start(spa);
5042 
5043 		/*
5044 		 * Wait for all claims to sync.  We sync up to the highest
5045 		 * claimed log block birth time so that claimed log blocks
5046 		 * don't appear to be from the future.  spa_claim_max_txg
5047 		 * will have been set for us by ZIL traversal operations
5048 		 * performed above.
5049 		 */
5050 		txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg);
5051 
5052 		/*
5053 		 * Check if we need to request an update of the config. On the
5054 		 * next sync, we would update the config stored in vdev labels
5055 		 * and the cachefile (by default /etc/zfs/zpool.cache).
5056 		 */
5057 		spa_ld_check_for_config_update(spa, config_cache_txg,
5058 		    update_config_cache);
5059 
5060 		/*
5061 		 * Check if a rebuild was in progress and if so resume it.
5062 		 * Then check all DTLs to see if anything needs resilvering.
5063 		 * The resilver will be deferred if a rebuild was started.
5064 		 */
5065 		if (vdev_rebuild_active(spa->spa_root_vdev)) {
5066 			vdev_rebuild_restart(spa);
5067 		} else if (!dsl_scan_resilvering(spa->spa_dsl_pool) &&
5068 		    vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
5069 			spa_async_request(spa, SPA_ASYNC_RESILVER);
5070 		}
5071 
5072 		/*
5073 		 * Log the fact that we booted up (so that we can detect if
5074 		 * we rebooted in the middle of an operation).
5075 		 */
5076 		spa_history_log_version(spa, "open", NULL);
5077 
5078 		spa_restart_removal(spa);
5079 		spa_spawn_aux_threads(spa);
5080 
5081 		/*
5082 		 * Delete any inconsistent datasets.
5083 		 *
5084 		 * Note:
5085 		 * Since we may be issuing deletes for clones here,
5086 		 * we make sure to do so after we've spawned all the
5087 		 * auxiliary threads above (from which the livelist
5088 		 * deletion zthr is part of).
5089 		 */
5090 		(void) dmu_objset_find(spa_name(spa),
5091 		    dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN);
5092 
5093 		/*
5094 		 * Clean up any stale temporary dataset userrefs.
5095 		 */
5096 		dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool);
5097 
5098 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
5099 		vdev_initialize_restart(spa->spa_root_vdev);
5100 		vdev_trim_restart(spa->spa_root_vdev);
5101 		vdev_autotrim_restart(spa);
5102 		spa_config_exit(spa, SCL_CONFIG, FTAG);
5103 	}
5104 
5105 	spa_import_progress_remove(spa_guid(spa));
5106 	spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD);
5107 
5108 	spa_load_note(spa, "LOADED");
5109 
5110 	return (0);
5111 }
5112 
5113 static int
5114 spa_load_retry(spa_t *spa, spa_load_state_t state)
5115 {
5116 	spa_mode_t mode = spa->spa_mode;
5117 
5118 	spa_unload(spa);
5119 	spa_deactivate(spa);
5120 
5121 	spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1;
5122 
5123 	spa_activate(spa, mode);
5124 	spa_async_suspend(spa);
5125 
5126 	spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu",
5127 	    (u_longlong_t)spa->spa_load_max_txg);
5128 
5129 	return (spa_load(spa, state, SPA_IMPORT_EXISTING));
5130 }
5131 
5132 /*
5133  * If spa_load() fails this function will try loading prior txg's. If
5134  * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool
5135  * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this
5136  * function will not rewind the pool and will return the same error as
5137  * spa_load().
5138  */
5139 static int
5140 spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request,
5141     int rewind_flags)
5142 {
5143 	nvlist_t *loadinfo = NULL;
5144 	nvlist_t *config = NULL;
5145 	int load_error, rewind_error;
5146 	uint64_t safe_rewind_txg;
5147 	uint64_t min_txg;
5148 
5149 	if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) {
5150 		spa->spa_load_max_txg = spa->spa_load_txg;
5151 		spa_set_log_state(spa, SPA_LOG_CLEAR);
5152 	} else {
5153 		spa->spa_load_max_txg = max_request;
5154 		if (max_request != UINT64_MAX)
5155 			spa->spa_extreme_rewind = B_TRUE;
5156 	}
5157 
5158 	load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING);
5159 	if (load_error == 0)
5160 		return (0);
5161 	if (load_error == ZFS_ERR_NO_CHECKPOINT) {
5162 		/*
5163 		 * When attempting checkpoint-rewind on a pool with no
5164 		 * checkpoint, we should not attempt to load uberblocks
5165 		 * from previous txgs when spa_load fails.
5166 		 */
5167 		ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
5168 		spa_import_progress_remove(spa_guid(spa));
5169 		return (load_error);
5170 	}
5171 
5172 	if (spa->spa_root_vdev != NULL)
5173 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
5174 
5175 	spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg;
5176 	spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp;
5177 
5178 	if (rewind_flags & ZPOOL_NEVER_REWIND) {
5179 		nvlist_free(config);
5180 		spa_import_progress_remove(spa_guid(spa));
5181 		return (load_error);
5182 	}
5183 
5184 	if (state == SPA_LOAD_RECOVER) {
5185 		/* Price of rolling back is discarding txgs, including log */
5186 		spa_set_log_state(spa, SPA_LOG_CLEAR);
5187 	} else {
5188 		/*
5189 		 * If we aren't rolling back save the load info from our first
5190 		 * import attempt so that we can restore it after attempting
5191 		 * to rewind.
5192 		 */
5193 		loadinfo = spa->spa_load_info;
5194 		spa->spa_load_info = fnvlist_alloc();
5195 	}
5196 
5197 	spa->spa_load_max_txg = spa->spa_last_ubsync_txg;
5198 	safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE;
5199 	min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ?
5200 	    TXG_INITIAL : safe_rewind_txg;
5201 
5202 	/*
5203 	 * Continue as long as we're finding errors, we're still within
5204 	 * the acceptable rewind range, and we're still finding uberblocks
5205 	 */
5206 	while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg &&
5207 	    spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) {
5208 		if (spa->spa_load_max_txg < safe_rewind_txg)
5209 			spa->spa_extreme_rewind = B_TRUE;
5210 		rewind_error = spa_load_retry(spa, state);
5211 	}
5212 
5213 	spa->spa_extreme_rewind = B_FALSE;
5214 	spa->spa_load_max_txg = UINT64_MAX;
5215 
5216 	if (config && (rewind_error || state != SPA_LOAD_RECOVER))
5217 		spa_config_set(spa, config);
5218 	else
5219 		nvlist_free(config);
5220 
5221 	if (state == SPA_LOAD_RECOVER) {
5222 		ASSERT3P(loadinfo, ==, NULL);
5223 		spa_import_progress_remove(spa_guid(spa));
5224 		return (rewind_error);
5225 	} else {
5226 		/* Store the rewind info as part of the initial load info */
5227 		fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO,
5228 		    spa->spa_load_info);
5229 
5230 		/* Restore the initial load info */
5231 		fnvlist_free(spa->spa_load_info);
5232 		spa->spa_load_info = loadinfo;
5233 
5234 		spa_import_progress_remove(spa_guid(spa));
5235 		return (load_error);
5236 	}
5237 }
5238 
5239 /*
5240  * Pool Open/Import
5241  *
5242  * The import case is identical to an open except that the configuration is sent
5243  * down from userland, instead of grabbed from the configuration cache.  For the
5244  * case of an open, the pool configuration will exist in the
5245  * POOL_STATE_UNINITIALIZED state.
5246  *
5247  * The stats information (gen/count/ustats) is used to gather vdev statistics at
5248  * the same time open the pool, without having to keep around the spa_t in some
5249  * ambiguous state.
5250  */
5251 static int
5252 spa_open_common(const char *pool, spa_t **spapp, const void *tag,
5253     nvlist_t *nvpolicy, nvlist_t **config)
5254 {
5255 	spa_t *spa;
5256 	spa_load_state_t state = SPA_LOAD_OPEN;
5257 	int error;
5258 	int locked = B_FALSE;
5259 	int firstopen = B_FALSE;
5260 
5261 	*spapp = NULL;
5262 
5263 	/*
5264 	 * As disgusting as this is, we need to support recursive calls to this
5265 	 * function because dsl_dir_open() is called during spa_load(), and ends
5266 	 * up calling spa_open() again.  The real fix is to figure out how to
5267 	 * avoid dsl_dir_open() calling this in the first place.
5268 	 */
5269 	if (MUTEX_NOT_HELD(&spa_namespace_lock)) {
5270 		mutex_enter(&spa_namespace_lock);
5271 		locked = B_TRUE;
5272 	}
5273 
5274 	if ((spa = spa_lookup(pool)) == NULL) {
5275 		if (locked)
5276 			mutex_exit(&spa_namespace_lock);
5277 		return (SET_ERROR(ENOENT));
5278 	}
5279 
5280 	if (spa->spa_state == POOL_STATE_UNINITIALIZED) {
5281 		zpool_load_policy_t policy;
5282 
5283 		firstopen = B_TRUE;
5284 
5285 		zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config,
5286 		    &policy);
5287 		if (policy.zlp_rewind & ZPOOL_DO_REWIND)
5288 			state = SPA_LOAD_RECOVER;
5289 
5290 		spa_activate(spa, spa_mode_global);
5291 
5292 		if (state != SPA_LOAD_RECOVER)
5293 			spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
5294 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
5295 
5296 		zfs_dbgmsg("spa_open_common: opening %s", pool);
5297 		error = spa_load_best(spa, state, policy.zlp_txg,
5298 		    policy.zlp_rewind);
5299 
5300 		if (error == EBADF) {
5301 			/*
5302 			 * If vdev_validate() returns failure (indicated by
5303 			 * EBADF), it indicates that one of the vdevs indicates
5304 			 * that the pool has been exported or destroyed.  If
5305 			 * this is the case, the config cache is out of sync and
5306 			 * we should remove the pool from the namespace.
5307 			 */
5308 			spa_unload(spa);
5309 			spa_deactivate(spa);
5310 			spa_write_cachefile(spa, B_TRUE, B_TRUE, B_FALSE);
5311 			spa_remove(spa);
5312 			if (locked)
5313 				mutex_exit(&spa_namespace_lock);
5314 			return (SET_ERROR(ENOENT));
5315 		}
5316 
5317 		if (error) {
5318 			/*
5319 			 * We can't open the pool, but we still have useful
5320 			 * information: the state of each vdev after the
5321 			 * attempted vdev_open().  Return this to the user.
5322 			 */
5323 			if (config != NULL && spa->spa_config) {
5324 				*config = fnvlist_dup(spa->spa_config);
5325 				fnvlist_add_nvlist(*config,
5326 				    ZPOOL_CONFIG_LOAD_INFO,
5327 				    spa->spa_load_info);
5328 			}
5329 			spa_unload(spa);
5330 			spa_deactivate(spa);
5331 			spa->spa_last_open_failed = error;
5332 			if (locked)
5333 				mutex_exit(&spa_namespace_lock);
5334 			*spapp = NULL;
5335 			return (error);
5336 		}
5337 	}
5338 
5339 	spa_open_ref(spa, tag);
5340 
5341 	if (config != NULL)
5342 		*config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
5343 
5344 	/*
5345 	 * If we've recovered the pool, pass back any information we
5346 	 * gathered while doing the load.
5347 	 */
5348 	if (state == SPA_LOAD_RECOVER && config != NULL) {
5349 		fnvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO,
5350 		    spa->spa_load_info);
5351 	}
5352 
5353 	if (locked) {
5354 		spa->spa_last_open_failed = 0;
5355 		spa->spa_last_ubsync_txg = 0;
5356 		spa->spa_load_txg = 0;
5357 		mutex_exit(&spa_namespace_lock);
5358 	}
5359 
5360 	if (firstopen)
5361 		zvol_create_minors_recursive(spa_name(spa));
5362 
5363 	*spapp = spa;
5364 
5365 	return (0);
5366 }
5367 
5368 int
5369 spa_open_rewind(const char *name, spa_t **spapp, const void *tag,
5370     nvlist_t *policy, nvlist_t **config)
5371 {
5372 	return (spa_open_common(name, spapp, tag, policy, config));
5373 }
5374 
5375 int
5376 spa_open(const char *name, spa_t **spapp, const void *tag)
5377 {
5378 	return (spa_open_common(name, spapp, tag, NULL, NULL));
5379 }
5380 
5381 /*
5382  * Lookup the given spa_t, incrementing the inject count in the process,
5383  * preventing it from being exported or destroyed.
5384  */
5385 spa_t *
5386 spa_inject_addref(char *name)
5387 {
5388 	spa_t *spa;
5389 
5390 	mutex_enter(&spa_namespace_lock);
5391 	if ((spa = spa_lookup(name)) == NULL) {
5392 		mutex_exit(&spa_namespace_lock);
5393 		return (NULL);
5394 	}
5395 	spa->spa_inject_ref++;
5396 	mutex_exit(&spa_namespace_lock);
5397 
5398 	return (spa);
5399 }
5400 
5401 void
5402 spa_inject_delref(spa_t *spa)
5403 {
5404 	mutex_enter(&spa_namespace_lock);
5405 	spa->spa_inject_ref--;
5406 	mutex_exit(&spa_namespace_lock);
5407 }
5408 
5409 /*
5410  * Add spares device information to the nvlist.
5411  */
5412 static void
5413 spa_add_spares(spa_t *spa, nvlist_t *config)
5414 {
5415 	nvlist_t **spares;
5416 	uint_t i, nspares;
5417 	nvlist_t *nvroot;
5418 	uint64_t guid;
5419 	vdev_stat_t *vs;
5420 	uint_t vsc;
5421 	uint64_t pool;
5422 
5423 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5424 
5425 	if (spa->spa_spares.sav_count == 0)
5426 		return;
5427 
5428 	nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE);
5429 	VERIFY0(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
5430 	    ZPOOL_CONFIG_SPARES, &spares, &nspares));
5431 	if (nspares != 0) {
5432 		fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5433 		    (const nvlist_t * const *)spares, nspares);
5434 		VERIFY0(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5435 		    &spares, &nspares));
5436 
5437 		/*
5438 		 * Go through and find any spares which have since been
5439 		 * repurposed as an active spare.  If this is the case, update
5440 		 * their status appropriately.
5441 		 */
5442 		for (i = 0; i < nspares; i++) {
5443 			guid = fnvlist_lookup_uint64(spares[i],
5444 			    ZPOOL_CONFIG_GUID);
5445 			VERIFY0(nvlist_lookup_uint64_array(spares[i],
5446 			    ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc));
5447 			if (spa_spare_exists(guid, &pool, NULL) &&
5448 			    pool != 0ULL) {
5449 				vs->vs_state = VDEV_STATE_CANT_OPEN;
5450 				vs->vs_aux = VDEV_AUX_SPARED;
5451 			} else {
5452 				vs->vs_state =
5453 				    spa->spa_spares.sav_vdevs[i]->vdev_state;
5454 			}
5455 		}
5456 	}
5457 }
5458 
5459 /*
5460  * Add l2cache device information to the nvlist, including vdev stats.
5461  */
5462 static void
5463 spa_add_l2cache(spa_t *spa, nvlist_t *config)
5464 {
5465 	nvlist_t **l2cache;
5466 	uint_t i, j, nl2cache;
5467 	nvlist_t *nvroot;
5468 	uint64_t guid;
5469 	vdev_t *vd;
5470 	vdev_stat_t *vs;
5471 	uint_t vsc;
5472 
5473 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5474 
5475 	if (spa->spa_l2cache.sav_count == 0)
5476 		return;
5477 
5478 	nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE);
5479 	VERIFY0(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
5480 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache));
5481 	if (nl2cache != 0) {
5482 		fnvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5483 		    (const nvlist_t * const *)l2cache, nl2cache);
5484 		VERIFY0(nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5485 		    &l2cache, &nl2cache));
5486 
5487 		/*
5488 		 * Update level 2 cache device stats.
5489 		 */
5490 
5491 		for (i = 0; i < nl2cache; i++) {
5492 			guid = fnvlist_lookup_uint64(l2cache[i],
5493 			    ZPOOL_CONFIG_GUID);
5494 
5495 			vd = NULL;
5496 			for (j = 0; j < spa->spa_l2cache.sav_count; j++) {
5497 				if (guid ==
5498 				    spa->spa_l2cache.sav_vdevs[j]->vdev_guid) {
5499 					vd = spa->spa_l2cache.sav_vdevs[j];
5500 					break;
5501 				}
5502 			}
5503 			ASSERT(vd != NULL);
5504 
5505 			VERIFY0(nvlist_lookup_uint64_array(l2cache[i],
5506 			    ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc));
5507 			vdev_get_stats(vd, vs);
5508 			vdev_config_generate_stats(vd, l2cache[i]);
5509 
5510 		}
5511 	}
5512 }
5513 
5514 static void
5515 spa_feature_stats_from_disk(spa_t *spa, nvlist_t *features)
5516 {
5517 	zap_cursor_t zc;
5518 	zap_attribute_t za;
5519 
5520 	if (spa->spa_feat_for_read_obj != 0) {
5521 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
5522 		    spa->spa_feat_for_read_obj);
5523 		    zap_cursor_retrieve(&zc, &za) == 0;
5524 		    zap_cursor_advance(&zc)) {
5525 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
5526 			    za.za_num_integers == 1);
5527 			VERIFY0(nvlist_add_uint64(features, za.za_name,
5528 			    za.za_first_integer));
5529 		}
5530 		zap_cursor_fini(&zc);
5531 	}
5532 
5533 	if (spa->spa_feat_for_write_obj != 0) {
5534 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
5535 		    spa->spa_feat_for_write_obj);
5536 		    zap_cursor_retrieve(&zc, &za) == 0;
5537 		    zap_cursor_advance(&zc)) {
5538 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
5539 			    za.za_num_integers == 1);
5540 			VERIFY0(nvlist_add_uint64(features, za.za_name,
5541 			    za.za_first_integer));
5542 		}
5543 		zap_cursor_fini(&zc);
5544 	}
5545 }
5546 
5547 static void
5548 spa_feature_stats_from_cache(spa_t *spa, nvlist_t *features)
5549 {
5550 	int i;
5551 
5552 	for (i = 0; i < SPA_FEATURES; i++) {
5553 		zfeature_info_t feature = spa_feature_table[i];
5554 		uint64_t refcount;
5555 
5556 		if (feature_get_refcount(spa, &feature, &refcount) != 0)
5557 			continue;
5558 
5559 		VERIFY0(nvlist_add_uint64(features, feature.fi_guid, refcount));
5560 	}
5561 }
5562 
5563 /*
5564  * Store a list of pool features and their reference counts in the
5565  * config.
5566  *
5567  * The first time this is called on a spa, allocate a new nvlist, fetch
5568  * the pool features and reference counts from disk, then save the list
5569  * in the spa. In subsequent calls on the same spa use the saved nvlist
5570  * and refresh its values from the cached reference counts.  This
5571  * ensures we don't block here on I/O on a suspended pool so 'zpool
5572  * clear' can resume the pool.
5573  */
5574 static void
5575 spa_add_feature_stats(spa_t *spa, nvlist_t *config)
5576 {
5577 	nvlist_t *features;
5578 
5579 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5580 
5581 	mutex_enter(&spa->spa_feat_stats_lock);
5582 	features = spa->spa_feat_stats;
5583 
5584 	if (features != NULL) {
5585 		spa_feature_stats_from_cache(spa, features);
5586 	} else {
5587 		VERIFY0(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP));
5588 		spa->spa_feat_stats = features;
5589 		spa_feature_stats_from_disk(spa, features);
5590 	}
5591 
5592 	VERIFY0(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS,
5593 	    features));
5594 
5595 	mutex_exit(&spa->spa_feat_stats_lock);
5596 }
5597 
5598 int
5599 spa_get_stats(const char *name, nvlist_t **config,
5600     char *altroot, size_t buflen)
5601 {
5602 	int error;
5603 	spa_t *spa;
5604 
5605 	*config = NULL;
5606 	error = spa_open_common(name, &spa, FTAG, NULL, config);
5607 
5608 	if (spa != NULL) {
5609 		/*
5610 		 * This still leaves a window of inconsistency where the spares
5611 		 * or l2cache devices could change and the config would be
5612 		 * self-inconsistent.
5613 		 */
5614 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
5615 
5616 		if (*config != NULL) {
5617 			uint64_t loadtimes[2];
5618 
5619 			loadtimes[0] = spa->spa_loaded_ts.tv_sec;
5620 			loadtimes[1] = spa->spa_loaded_ts.tv_nsec;
5621 			fnvlist_add_uint64_array(*config,
5622 			    ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2);
5623 
5624 			fnvlist_add_uint64(*config,
5625 			    ZPOOL_CONFIG_ERRCOUNT,
5626 			    spa_approx_errlog_size(spa));
5627 
5628 			if (spa_suspended(spa)) {
5629 				fnvlist_add_uint64(*config,
5630 				    ZPOOL_CONFIG_SUSPENDED,
5631 				    spa->spa_failmode);
5632 				fnvlist_add_uint64(*config,
5633 				    ZPOOL_CONFIG_SUSPENDED_REASON,
5634 				    spa->spa_suspended);
5635 			}
5636 
5637 			spa_add_spares(spa, *config);
5638 			spa_add_l2cache(spa, *config);
5639 			spa_add_feature_stats(spa, *config);
5640 		}
5641 	}
5642 
5643 	/*
5644 	 * We want to get the alternate root even for faulted pools, so we cheat
5645 	 * and call spa_lookup() directly.
5646 	 */
5647 	if (altroot) {
5648 		if (spa == NULL) {
5649 			mutex_enter(&spa_namespace_lock);
5650 			spa = spa_lookup(name);
5651 			if (spa)
5652 				spa_altroot(spa, altroot, buflen);
5653 			else
5654 				altroot[0] = '\0';
5655 			spa = NULL;
5656 			mutex_exit(&spa_namespace_lock);
5657 		} else {
5658 			spa_altroot(spa, altroot, buflen);
5659 		}
5660 	}
5661 
5662 	if (spa != NULL) {
5663 		spa_config_exit(spa, SCL_CONFIG, FTAG);
5664 		spa_close(spa, FTAG);
5665 	}
5666 
5667 	return (error);
5668 }
5669 
5670 /*
5671  * Validate that the auxiliary device array is well formed.  We must have an
5672  * array of nvlists, each which describes a valid leaf vdev.  If this is an
5673  * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be
5674  * specified, as long as they are well-formed.
5675  */
5676 static int
5677 spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode,
5678     spa_aux_vdev_t *sav, const char *config, uint64_t version,
5679     vdev_labeltype_t label)
5680 {
5681 	nvlist_t **dev;
5682 	uint_t i, ndev;
5683 	vdev_t *vd;
5684 	int error;
5685 
5686 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5687 
5688 	/*
5689 	 * It's acceptable to have no devs specified.
5690 	 */
5691 	if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0)
5692 		return (0);
5693 
5694 	if (ndev == 0)
5695 		return (SET_ERROR(EINVAL));
5696 
5697 	/*
5698 	 * Make sure the pool is formatted with a version that supports this
5699 	 * device type.
5700 	 */
5701 	if (spa_version(spa) < version)
5702 		return (SET_ERROR(ENOTSUP));
5703 
5704 	/*
5705 	 * Set the pending device list so we correctly handle device in-use
5706 	 * checking.
5707 	 */
5708 	sav->sav_pending = dev;
5709 	sav->sav_npending = ndev;
5710 
5711 	for (i = 0; i < ndev; i++) {
5712 		if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0,
5713 		    mode)) != 0)
5714 			goto out;
5715 
5716 		if (!vd->vdev_ops->vdev_op_leaf) {
5717 			vdev_free(vd);
5718 			error = SET_ERROR(EINVAL);
5719 			goto out;
5720 		}
5721 
5722 		vd->vdev_top = vd;
5723 
5724 		if ((error = vdev_open(vd)) == 0 &&
5725 		    (error = vdev_label_init(vd, crtxg, label)) == 0) {
5726 			fnvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID,
5727 			    vd->vdev_guid);
5728 		}
5729 
5730 		vdev_free(vd);
5731 
5732 		if (error &&
5733 		    (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE))
5734 			goto out;
5735 		else
5736 			error = 0;
5737 	}
5738 
5739 out:
5740 	sav->sav_pending = NULL;
5741 	sav->sav_npending = 0;
5742 	return (error);
5743 }
5744 
5745 static int
5746 spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode)
5747 {
5748 	int error;
5749 
5750 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5751 
5752 	if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode,
5753 	    &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES,
5754 	    VDEV_LABEL_SPARE)) != 0) {
5755 		return (error);
5756 	}
5757 
5758 	return (spa_validate_aux_devs(spa, nvroot, crtxg, mode,
5759 	    &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE,
5760 	    VDEV_LABEL_L2CACHE));
5761 }
5762 
5763 static void
5764 spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs,
5765     const char *config)
5766 {
5767 	int i;
5768 
5769 	if (sav->sav_config != NULL) {
5770 		nvlist_t **olddevs;
5771 		uint_t oldndevs;
5772 		nvlist_t **newdevs;
5773 
5774 		/*
5775 		 * Generate new dev list by concatenating with the
5776 		 * current dev list.
5777 		 */
5778 		VERIFY0(nvlist_lookup_nvlist_array(sav->sav_config, config,
5779 		    &olddevs, &oldndevs));
5780 
5781 		newdevs = kmem_alloc(sizeof (void *) *
5782 		    (ndevs + oldndevs), KM_SLEEP);
5783 		for (i = 0; i < oldndevs; i++)
5784 			newdevs[i] = fnvlist_dup(olddevs[i]);
5785 		for (i = 0; i < ndevs; i++)
5786 			newdevs[i + oldndevs] = fnvlist_dup(devs[i]);
5787 
5788 		fnvlist_remove(sav->sav_config, config);
5789 
5790 		fnvlist_add_nvlist_array(sav->sav_config, config,
5791 		    (const nvlist_t * const *)newdevs, ndevs + oldndevs);
5792 		for (i = 0; i < oldndevs + ndevs; i++)
5793 			nvlist_free(newdevs[i]);
5794 		kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *));
5795 	} else {
5796 		/*
5797 		 * Generate a new dev list.
5798 		 */
5799 		sav->sav_config = fnvlist_alloc();
5800 		fnvlist_add_nvlist_array(sav->sav_config, config,
5801 		    (const nvlist_t * const *)devs, ndevs);
5802 	}
5803 }
5804 
5805 /*
5806  * Stop and drop level 2 ARC devices
5807  */
5808 void
5809 spa_l2cache_drop(spa_t *spa)
5810 {
5811 	vdev_t *vd;
5812 	int i;
5813 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
5814 
5815 	for (i = 0; i < sav->sav_count; i++) {
5816 		uint64_t pool;
5817 
5818 		vd = sav->sav_vdevs[i];
5819 		ASSERT(vd != NULL);
5820 
5821 		if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
5822 		    pool != 0ULL && l2arc_vdev_present(vd))
5823 			l2arc_remove_vdev(vd);
5824 	}
5825 }
5826 
5827 /*
5828  * Verify encryption parameters for spa creation. If we are encrypting, we must
5829  * have the encryption feature flag enabled.
5830  */
5831 static int
5832 spa_create_check_encryption_params(dsl_crypto_params_t *dcp,
5833     boolean_t has_encryption)
5834 {
5835 	if (dcp->cp_crypt != ZIO_CRYPT_OFF &&
5836 	    dcp->cp_crypt != ZIO_CRYPT_INHERIT &&
5837 	    !has_encryption)
5838 		return (SET_ERROR(ENOTSUP));
5839 
5840 	return (dmu_objset_create_crypt_check(NULL, dcp, NULL));
5841 }
5842 
5843 /*
5844  * Pool Creation
5845  */
5846 int
5847 spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
5848     nvlist_t *zplprops, dsl_crypto_params_t *dcp)
5849 {
5850 	spa_t *spa;
5851 	const char *altroot = NULL;
5852 	vdev_t *rvd;
5853 	dsl_pool_t *dp;
5854 	dmu_tx_t *tx;
5855 	int error = 0;
5856 	uint64_t txg = TXG_INITIAL;
5857 	nvlist_t **spares, **l2cache;
5858 	uint_t nspares, nl2cache;
5859 	uint64_t version, obj, ndraid = 0;
5860 	boolean_t has_features;
5861 	boolean_t has_encryption;
5862 	boolean_t has_allocclass;
5863 	spa_feature_t feat;
5864 	const char *feat_name;
5865 	const char *poolname;
5866 	nvlist_t *nvl;
5867 
5868 	if (props == NULL ||
5869 	    nvlist_lookup_string(props, "tname", &poolname) != 0)
5870 		poolname = (char *)pool;
5871 
5872 	/*
5873 	 * If this pool already exists, return failure.
5874 	 */
5875 	mutex_enter(&spa_namespace_lock);
5876 	if (spa_lookup(poolname) != NULL) {
5877 		mutex_exit(&spa_namespace_lock);
5878 		return (SET_ERROR(EEXIST));
5879 	}
5880 
5881 	/*
5882 	 * Allocate a new spa_t structure.
5883 	 */
5884 	nvl = fnvlist_alloc();
5885 	fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool);
5886 	(void) nvlist_lookup_string(props,
5887 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
5888 	spa = spa_add(poolname, nvl, altroot);
5889 	fnvlist_free(nvl);
5890 	spa_activate(spa, spa_mode_global);
5891 
5892 	if (props && (error = spa_prop_validate(spa, props))) {
5893 		spa_deactivate(spa);
5894 		spa_remove(spa);
5895 		mutex_exit(&spa_namespace_lock);
5896 		return (error);
5897 	}
5898 
5899 	/*
5900 	 * Temporary pool names should never be written to disk.
5901 	 */
5902 	if (poolname != pool)
5903 		spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME;
5904 
5905 	has_features = B_FALSE;
5906 	has_encryption = B_FALSE;
5907 	has_allocclass = B_FALSE;
5908 	for (nvpair_t *elem = nvlist_next_nvpair(props, NULL);
5909 	    elem != NULL; elem = nvlist_next_nvpair(props, elem)) {
5910 		if (zpool_prop_feature(nvpair_name(elem))) {
5911 			has_features = B_TRUE;
5912 
5913 			feat_name = strchr(nvpair_name(elem), '@') + 1;
5914 			VERIFY0(zfeature_lookup_name(feat_name, &feat));
5915 			if (feat == SPA_FEATURE_ENCRYPTION)
5916 				has_encryption = B_TRUE;
5917 			if (feat == SPA_FEATURE_ALLOCATION_CLASSES)
5918 				has_allocclass = B_TRUE;
5919 		}
5920 	}
5921 
5922 	/* verify encryption params, if they were provided */
5923 	if (dcp != NULL) {
5924 		error = spa_create_check_encryption_params(dcp, has_encryption);
5925 		if (error != 0) {
5926 			spa_deactivate(spa);
5927 			spa_remove(spa);
5928 			mutex_exit(&spa_namespace_lock);
5929 			return (error);
5930 		}
5931 	}
5932 	if (!has_allocclass && zfs_special_devs(nvroot, NULL)) {
5933 		spa_deactivate(spa);
5934 		spa_remove(spa);
5935 		mutex_exit(&spa_namespace_lock);
5936 		return (ENOTSUP);
5937 	}
5938 
5939 	if (has_features || nvlist_lookup_uint64(props,
5940 	    zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) {
5941 		version = SPA_VERSION;
5942 	}
5943 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
5944 
5945 	spa->spa_first_txg = txg;
5946 	spa->spa_uberblock.ub_txg = txg - 1;
5947 	spa->spa_uberblock.ub_version = version;
5948 	spa->spa_ubsync = spa->spa_uberblock;
5949 	spa->spa_load_state = SPA_LOAD_CREATE;
5950 	spa->spa_removing_phys.sr_state = DSS_NONE;
5951 	spa->spa_removing_phys.sr_removing_vdev = -1;
5952 	spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
5953 	spa->spa_indirect_vdevs_loaded = B_TRUE;
5954 
5955 	/*
5956 	 * Create "The Godfather" zio to hold all async IOs
5957 	 */
5958 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
5959 	    KM_SLEEP);
5960 	for (int i = 0; i < max_ncpus; i++) {
5961 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
5962 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
5963 		    ZIO_FLAG_GODFATHER);
5964 	}
5965 
5966 	/*
5967 	 * Create the root vdev.
5968 	 */
5969 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5970 
5971 	error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD);
5972 
5973 	ASSERT(error != 0 || rvd != NULL);
5974 	ASSERT(error != 0 || spa->spa_root_vdev == rvd);
5975 
5976 	if (error == 0 && !zfs_allocatable_devs(nvroot))
5977 		error = SET_ERROR(EINVAL);
5978 
5979 	if (error == 0 &&
5980 	    (error = vdev_create(rvd, txg, B_FALSE)) == 0 &&
5981 	    (error = vdev_draid_spare_create(nvroot, rvd, &ndraid, 0)) == 0 &&
5982 	    (error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) == 0) {
5983 		/*
5984 		 * instantiate the metaslab groups (this will dirty the vdevs)
5985 		 * we can no longer error exit past this point
5986 		 */
5987 		for (int c = 0; error == 0 && c < rvd->vdev_children; c++) {
5988 			vdev_t *vd = rvd->vdev_child[c];
5989 
5990 			vdev_metaslab_set_size(vd);
5991 			vdev_expand(vd, txg);
5992 		}
5993 	}
5994 
5995 	spa_config_exit(spa, SCL_ALL, FTAG);
5996 
5997 	if (error != 0) {
5998 		spa_unload(spa);
5999 		spa_deactivate(spa);
6000 		spa_remove(spa);
6001 		mutex_exit(&spa_namespace_lock);
6002 		return (error);
6003 	}
6004 
6005 	/*
6006 	 * Get the list of spares, if specified.
6007 	 */
6008 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
6009 	    &spares, &nspares) == 0) {
6010 		spa->spa_spares.sav_config = fnvlist_alloc();
6011 		fnvlist_add_nvlist_array(spa->spa_spares.sav_config,
6012 		    ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares,
6013 		    nspares);
6014 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6015 		spa_load_spares(spa);
6016 		spa_config_exit(spa, SCL_ALL, FTAG);
6017 		spa->spa_spares.sav_sync = B_TRUE;
6018 	}
6019 
6020 	/*
6021 	 * Get the list of level 2 cache devices, if specified.
6022 	 */
6023 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
6024 	    &l2cache, &nl2cache) == 0) {
6025 		VERIFY0(nvlist_alloc(&spa->spa_l2cache.sav_config,
6026 		    NV_UNIQUE_NAME, KM_SLEEP));
6027 		fnvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
6028 		    ZPOOL_CONFIG_L2CACHE, (const nvlist_t * const *)l2cache,
6029 		    nl2cache);
6030 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6031 		spa_load_l2cache(spa);
6032 		spa_config_exit(spa, SCL_ALL, FTAG);
6033 		spa->spa_l2cache.sav_sync = B_TRUE;
6034 	}
6035 
6036 	spa->spa_is_initializing = B_TRUE;
6037 	spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg);
6038 	spa->spa_is_initializing = B_FALSE;
6039 
6040 	/*
6041 	 * Create DDTs (dedup tables).
6042 	 */
6043 	ddt_create(spa);
6044 	/*
6045 	 * Create BRT table and BRT table object.
6046 	 */
6047 	brt_create(spa);
6048 
6049 	spa_update_dspace(spa);
6050 
6051 	tx = dmu_tx_create_assigned(dp, txg);
6052 
6053 	/*
6054 	 * Create the pool's history object.
6055 	 */
6056 	if (version >= SPA_VERSION_ZPOOL_HISTORY && !spa->spa_history)
6057 		spa_history_create_obj(spa, tx);
6058 
6059 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE);
6060 	spa_history_log_version(spa, "create", tx);
6061 
6062 	/*
6063 	 * Create the pool config object.
6064 	 */
6065 	spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset,
6066 	    DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE,
6067 	    DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx);
6068 
6069 	if (zap_add(spa->spa_meta_objset,
6070 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG,
6071 	    sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) {
6072 		cmn_err(CE_PANIC, "failed to add pool config");
6073 	}
6074 
6075 	if (zap_add(spa->spa_meta_objset,
6076 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION,
6077 	    sizeof (uint64_t), 1, &version, tx) != 0) {
6078 		cmn_err(CE_PANIC, "failed to add pool version");
6079 	}
6080 
6081 	/* Newly created pools with the right version are always deflated. */
6082 	if (version >= SPA_VERSION_RAIDZ_DEFLATE) {
6083 		spa->spa_deflate = TRUE;
6084 		if (zap_add(spa->spa_meta_objset,
6085 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
6086 		    sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) {
6087 			cmn_err(CE_PANIC, "failed to add deflate");
6088 		}
6089 	}
6090 
6091 	/*
6092 	 * Create the deferred-free bpobj.  Turn off compression
6093 	 * because sync-to-convergence takes longer if the blocksize
6094 	 * keeps changing.
6095 	 */
6096 	obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx);
6097 	dmu_object_set_compress(spa->spa_meta_objset, obj,
6098 	    ZIO_COMPRESS_OFF, tx);
6099 	if (zap_add(spa->spa_meta_objset,
6100 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ,
6101 	    sizeof (uint64_t), 1, &obj, tx) != 0) {
6102 		cmn_err(CE_PANIC, "failed to add bpobj");
6103 	}
6104 	VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj,
6105 	    spa->spa_meta_objset, obj));
6106 
6107 	/*
6108 	 * Generate some random noise for salted checksums to operate on.
6109 	 */
6110 	(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
6111 	    sizeof (spa->spa_cksum_salt.zcs_bytes));
6112 
6113 	/*
6114 	 * Set pool properties.
6115 	 */
6116 	spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS);
6117 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
6118 	spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE);
6119 	spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND);
6120 	spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST);
6121 	spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM);
6122 
6123 	if (props != NULL) {
6124 		spa_configfile_set(spa, props, B_FALSE);
6125 		spa_sync_props(props, tx);
6126 	}
6127 
6128 	for (int i = 0; i < ndraid; i++)
6129 		spa_feature_incr(spa, SPA_FEATURE_DRAID, tx);
6130 
6131 	dmu_tx_commit(tx);
6132 
6133 	spa->spa_sync_on = B_TRUE;
6134 	txg_sync_start(dp);
6135 	mmp_thread_start(spa);
6136 	txg_wait_synced(dp, txg);
6137 
6138 	spa_spawn_aux_threads(spa);
6139 
6140 	spa_write_cachefile(spa, B_FALSE, B_TRUE, B_TRUE);
6141 
6142 	/*
6143 	 * Don't count references from objsets that are already closed
6144 	 * and are making their way through the eviction process.
6145 	 */
6146 	spa_evicting_os_wait(spa);
6147 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
6148 	spa->spa_load_state = SPA_LOAD_NONE;
6149 
6150 	spa_import_os(spa);
6151 
6152 	mutex_exit(&spa_namespace_lock);
6153 
6154 	return (0);
6155 }
6156 
6157 /*
6158  * Import a non-root pool into the system.
6159  */
6160 int
6161 spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags)
6162 {
6163 	spa_t *spa;
6164 	const char *altroot = NULL;
6165 	spa_load_state_t state = SPA_LOAD_IMPORT;
6166 	zpool_load_policy_t policy;
6167 	spa_mode_t mode = spa_mode_global;
6168 	uint64_t readonly = B_FALSE;
6169 	int error;
6170 	nvlist_t *nvroot;
6171 	nvlist_t **spares, **l2cache;
6172 	uint_t nspares, nl2cache;
6173 
6174 	/*
6175 	 * If a pool with this name exists, return failure.
6176 	 */
6177 	mutex_enter(&spa_namespace_lock);
6178 	if (spa_lookup(pool) != NULL) {
6179 		mutex_exit(&spa_namespace_lock);
6180 		return (SET_ERROR(EEXIST));
6181 	}
6182 
6183 	/*
6184 	 * Create and initialize the spa structure.
6185 	 */
6186 	(void) nvlist_lookup_string(props,
6187 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
6188 	(void) nvlist_lookup_uint64(props,
6189 	    zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly);
6190 	if (readonly)
6191 		mode = SPA_MODE_READ;
6192 	spa = spa_add(pool, config, altroot);
6193 	spa->spa_import_flags = flags;
6194 
6195 	/*
6196 	 * Verbatim import - Take a pool and insert it into the namespace
6197 	 * as if it had been loaded at boot.
6198 	 */
6199 	if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) {
6200 		if (props != NULL)
6201 			spa_configfile_set(spa, props, B_FALSE);
6202 
6203 		spa_write_cachefile(spa, B_FALSE, B_TRUE, B_FALSE);
6204 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
6205 		zfs_dbgmsg("spa_import: verbatim import of %s", pool);
6206 		mutex_exit(&spa_namespace_lock);
6207 		return (0);
6208 	}
6209 
6210 	spa_activate(spa, mode);
6211 
6212 	/*
6213 	 * Don't start async tasks until we know everything is healthy.
6214 	 */
6215 	spa_async_suspend(spa);
6216 
6217 	zpool_get_load_policy(config, &policy);
6218 	if (policy.zlp_rewind & ZPOOL_DO_REWIND)
6219 		state = SPA_LOAD_RECOVER;
6220 
6221 	spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT;
6222 
6223 	if (state != SPA_LOAD_RECOVER) {
6224 		spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
6225 		zfs_dbgmsg("spa_import: importing %s", pool);
6226 	} else {
6227 		zfs_dbgmsg("spa_import: importing %s, max_txg=%lld "
6228 		    "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg);
6229 	}
6230 	error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind);
6231 
6232 	/*
6233 	 * Propagate anything learned while loading the pool and pass it
6234 	 * back to caller (i.e. rewind info, missing devices, etc).
6235 	 */
6236 	fnvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO, spa->spa_load_info);
6237 
6238 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6239 	/*
6240 	 * Toss any existing sparelist, as it doesn't have any validity
6241 	 * anymore, and conflicts with spa_has_spare().
6242 	 */
6243 	if (spa->spa_spares.sav_config) {
6244 		nvlist_free(spa->spa_spares.sav_config);
6245 		spa->spa_spares.sav_config = NULL;
6246 		spa_load_spares(spa);
6247 	}
6248 	if (spa->spa_l2cache.sav_config) {
6249 		nvlist_free(spa->spa_l2cache.sav_config);
6250 		spa->spa_l2cache.sav_config = NULL;
6251 		spa_load_l2cache(spa);
6252 	}
6253 
6254 	nvroot = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE);
6255 	spa_config_exit(spa, SCL_ALL, FTAG);
6256 
6257 	if (props != NULL)
6258 		spa_configfile_set(spa, props, B_FALSE);
6259 
6260 	if (error != 0 || (props && spa_writeable(spa) &&
6261 	    (error = spa_prop_set(spa, props)))) {
6262 		spa_unload(spa);
6263 		spa_deactivate(spa);
6264 		spa_remove(spa);
6265 		mutex_exit(&spa_namespace_lock);
6266 		return (error);
6267 	}
6268 
6269 	spa_async_resume(spa);
6270 
6271 	/*
6272 	 * Override any spares and level 2 cache devices as specified by
6273 	 * the user, as these may have correct device names/devids, etc.
6274 	 */
6275 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
6276 	    &spares, &nspares) == 0) {
6277 		if (spa->spa_spares.sav_config)
6278 			fnvlist_remove(spa->spa_spares.sav_config,
6279 			    ZPOOL_CONFIG_SPARES);
6280 		else
6281 			spa->spa_spares.sav_config = fnvlist_alloc();
6282 		fnvlist_add_nvlist_array(spa->spa_spares.sav_config,
6283 		    ZPOOL_CONFIG_SPARES, (const nvlist_t * const *)spares,
6284 		    nspares);
6285 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6286 		spa_load_spares(spa);
6287 		spa_config_exit(spa, SCL_ALL, FTAG);
6288 		spa->spa_spares.sav_sync = B_TRUE;
6289 	}
6290 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
6291 	    &l2cache, &nl2cache) == 0) {
6292 		if (spa->spa_l2cache.sav_config)
6293 			fnvlist_remove(spa->spa_l2cache.sav_config,
6294 			    ZPOOL_CONFIG_L2CACHE);
6295 		else
6296 			spa->spa_l2cache.sav_config = fnvlist_alloc();
6297 		fnvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
6298 		    ZPOOL_CONFIG_L2CACHE, (const nvlist_t * const *)l2cache,
6299 		    nl2cache);
6300 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6301 		spa_load_l2cache(spa);
6302 		spa_config_exit(spa, SCL_ALL, FTAG);
6303 		spa->spa_l2cache.sav_sync = B_TRUE;
6304 	}
6305 
6306 	/*
6307 	 * Check for any removed devices.
6308 	 */
6309 	if (spa->spa_autoreplace) {
6310 		spa_aux_check_removed(&spa->spa_spares);
6311 		spa_aux_check_removed(&spa->spa_l2cache);
6312 	}
6313 
6314 	if (spa_writeable(spa)) {
6315 		/*
6316 		 * Update the config cache to include the newly-imported pool.
6317 		 */
6318 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
6319 	}
6320 
6321 	/*
6322 	 * It's possible that the pool was expanded while it was exported.
6323 	 * We kick off an async task to handle this for us.
6324 	 */
6325 	spa_async_request(spa, SPA_ASYNC_AUTOEXPAND);
6326 
6327 	spa_history_log_version(spa, "import", NULL);
6328 
6329 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
6330 
6331 	mutex_exit(&spa_namespace_lock);
6332 
6333 	zvol_create_minors_recursive(pool);
6334 
6335 	spa_import_os(spa);
6336 
6337 	return (0);
6338 }
6339 
6340 nvlist_t *
6341 spa_tryimport(nvlist_t *tryconfig)
6342 {
6343 	nvlist_t *config = NULL;
6344 	const char *poolname, *cachefile;
6345 	spa_t *spa;
6346 	uint64_t state;
6347 	int error;
6348 	zpool_load_policy_t policy;
6349 
6350 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname))
6351 		return (NULL);
6352 
6353 	if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state))
6354 		return (NULL);
6355 
6356 	/*
6357 	 * Create and initialize the spa structure.
6358 	 */
6359 	mutex_enter(&spa_namespace_lock);
6360 	spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL);
6361 	spa_activate(spa, SPA_MODE_READ);
6362 
6363 	/*
6364 	 * Rewind pool if a max txg was provided.
6365 	 */
6366 	zpool_get_load_policy(spa->spa_config, &policy);
6367 	if (policy.zlp_txg != UINT64_MAX) {
6368 		spa->spa_load_max_txg = policy.zlp_txg;
6369 		spa->spa_extreme_rewind = B_TRUE;
6370 		zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld",
6371 		    poolname, (longlong_t)policy.zlp_txg);
6372 	} else {
6373 		zfs_dbgmsg("spa_tryimport: importing %s", poolname);
6374 	}
6375 
6376 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile)
6377 	    == 0) {
6378 		zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile);
6379 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
6380 	} else {
6381 		spa->spa_config_source = SPA_CONFIG_SRC_SCAN;
6382 	}
6383 
6384 	/*
6385 	 * spa_import() relies on a pool config fetched by spa_try_import()
6386 	 * for spare/cache devices. Import flags are not passed to
6387 	 * spa_tryimport(), which makes it return early due to a missing log
6388 	 * device and missing retrieving the cache device and spare eventually.
6389 	 * Passing ZFS_IMPORT_MISSING_LOG to spa_tryimport() makes it fetch
6390 	 * the correct configuration regardless of the missing log device.
6391 	 */
6392 	spa->spa_import_flags |= ZFS_IMPORT_MISSING_LOG;
6393 
6394 	error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING);
6395 
6396 	/*
6397 	 * If 'tryconfig' was at least parsable, return the current config.
6398 	 */
6399 	if (spa->spa_root_vdev != NULL) {
6400 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
6401 		fnvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, poolname);
6402 		fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE, state);
6403 		fnvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
6404 		    spa->spa_uberblock.ub_timestamp);
6405 		fnvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO,
6406 		    spa->spa_load_info);
6407 		fnvlist_add_uint64(config, ZPOOL_CONFIG_ERRATA,
6408 		    spa->spa_errata);
6409 
6410 		/*
6411 		 * If the bootfs property exists on this pool then we
6412 		 * copy it out so that external consumers can tell which
6413 		 * pools are bootable.
6414 		 */
6415 		if ((!error || error == EEXIST) && spa->spa_bootfs) {
6416 			char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
6417 
6418 			/*
6419 			 * We have to play games with the name since the
6420 			 * pool was opened as TRYIMPORT_NAME.
6421 			 */
6422 			if (dsl_dsobj_to_dsname(spa_name(spa),
6423 			    spa->spa_bootfs, tmpname) == 0) {
6424 				char *cp;
6425 				char *dsname;
6426 
6427 				dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
6428 
6429 				cp = strchr(tmpname, '/');
6430 				if (cp == NULL) {
6431 					(void) strlcpy(dsname, tmpname,
6432 					    MAXPATHLEN);
6433 				} else {
6434 					(void) snprintf(dsname, MAXPATHLEN,
6435 					    "%s/%s", poolname, ++cp);
6436 				}
6437 				fnvlist_add_string(config, ZPOOL_CONFIG_BOOTFS,
6438 				    dsname);
6439 				kmem_free(dsname, MAXPATHLEN);
6440 			}
6441 			kmem_free(tmpname, MAXPATHLEN);
6442 		}
6443 
6444 		/*
6445 		 * Add the list of hot spares and level 2 cache devices.
6446 		 */
6447 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
6448 		spa_add_spares(spa, config);
6449 		spa_add_l2cache(spa, config);
6450 		spa_config_exit(spa, SCL_CONFIG, FTAG);
6451 	}
6452 
6453 	spa_unload(spa);
6454 	spa_deactivate(spa);
6455 	spa_remove(spa);
6456 	mutex_exit(&spa_namespace_lock);
6457 
6458 	return (config);
6459 }
6460 
6461 /*
6462  * Pool export/destroy
6463  *
6464  * The act of destroying or exporting a pool is very simple.  We make sure there
6465  * is no more pending I/O and any references to the pool are gone.  Then, we
6466  * update the pool state and sync all the labels to disk, removing the
6467  * configuration from the cache afterwards. If the 'hardforce' flag is set, then
6468  * we don't sync the labels or remove the configuration cache.
6469  */
6470 static int
6471 spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
6472     boolean_t force, boolean_t hardforce)
6473 {
6474 	int error;
6475 	spa_t *spa;
6476 
6477 	if (oldconfig)
6478 		*oldconfig = NULL;
6479 
6480 	if (!(spa_mode_global & SPA_MODE_WRITE))
6481 		return (SET_ERROR(EROFS));
6482 
6483 	mutex_enter(&spa_namespace_lock);
6484 	if ((spa = spa_lookup(pool)) == NULL) {
6485 		mutex_exit(&spa_namespace_lock);
6486 		return (SET_ERROR(ENOENT));
6487 	}
6488 
6489 	if (spa->spa_is_exporting) {
6490 		/* the pool is being exported by another thread */
6491 		mutex_exit(&spa_namespace_lock);
6492 		return (SET_ERROR(ZFS_ERR_EXPORT_IN_PROGRESS));
6493 	}
6494 	spa->spa_is_exporting = B_TRUE;
6495 
6496 	/*
6497 	 * Put a hold on the pool, drop the namespace lock, stop async tasks,
6498 	 * reacquire the namespace lock, and see if we can export.
6499 	 */
6500 	spa_open_ref(spa, FTAG);
6501 	mutex_exit(&spa_namespace_lock);
6502 	spa_async_suspend(spa);
6503 	if (spa->spa_zvol_taskq) {
6504 		zvol_remove_minors(spa, spa_name(spa), B_TRUE);
6505 		taskq_wait(spa->spa_zvol_taskq);
6506 	}
6507 	mutex_enter(&spa_namespace_lock);
6508 	spa_close(spa, FTAG);
6509 
6510 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
6511 		goto export_spa;
6512 	/*
6513 	 * The pool will be in core if it's openable, in which case we can
6514 	 * modify its state.  Objsets may be open only because they're dirty,
6515 	 * so we have to force it to sync before checking spa_refcnt.
6516 	 */
6517 	if (spa->spa_sync_on) {
6518 		txg_wait_synced(spa->spa_dsl_pool, 0);
6519 		spa_evicting_os_wait(spa);
6520 	}
6521 
6522 	/*
6523 	 * A pool cannot be exported or destroyed if there are active
6524 	 * references.  If we are resetting a pool, allow references by
6525 	 * fault injection handlers.
6526 	 */
6527 	if (!spa_refcount_zero(spa) || (spa->spa_inject_ref != 0)) {
6528 		error = SET_ERROR(EBUSY);
6529 		goto fail;
6530 	}
6531 
6532 	if (spa->spa_sync_on) {
6533 		vdev_t *rvd = spa->spa_root_vdev;
6534 		/*
6535 		 * A pool cannot be exported if it has an active shared spare.
6536 		 * This is to prevent other pools stealing the active spare
6537 		 * from an exported pool. At user's own will, such pool can
6538 		 * be forcedly exported.
6539 		 */
6540 		if (!force && new_state == POOL_STATE_EXPORTED &&
6541 		    spa_has_active_shared_spare(spa)) {
6542 			error = SET_ERROR(EXDEV);
6543 			goto fail;
6544 		}
6545 
6546 		/*
6547 		 * We're about to export or destroy this pool. Make sure
6548 		 * we stop all initialization and trim activity here before
6549 		 * we set the spa_final_txg. This will ensure that all
6550 		 * dirty data resulting from the initialization is
6551 		 * committed to disk before we unload the pool.
6552 		 */
6553 		vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE);
6554 		vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE);
6555 		vdev_autotrim_stop_all(spa);
6556 		vdev_rebuild_stop_all(spa);
6557 
6558 		/*
6559 		 * We want this to be reflected on every label,
6560 		 * so mark them all dirty.  spa_unload() will do the
6561 		 * final sync that pushes these changes out.
6562 		 */
6563 		if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) {
6564 			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6565 			spa->spa_state = new_state;
6566 			vdev_config_dirty(rvd);
6567 			spa_config_exit(spa, SCL_ALL, FTAG);
6568 		}
6569 
6570 		/*
6571 		 * If the log space map feature is enabled and the pool is
6572 		 * getting exported (but not destroyed), we want to spend some
6573 		 * time flushing as many metaslabs as we can in an attempt to
6574 		 * destroy log space maps and save import time. This has to be
6575 		 * done before we set the spa_final_txg, otherwise
6576 		 * spa_sync() -> spa_flush_metaslabs() may dirty the final TXGs.
6577 		 * spa_should_flush_logs_on_unload() should be called after
6578 		 * spa_state has been set to the new_state.
6579 		 */
6580 		if (spa_should_flush_logs_on_unload(spa))
6581 			spa_unload_log_sm_flush_all(spa);
6582 
6583 		if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) {
6584 			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6585 			spa->spa_final_txg = spa_last_synced_txg(spa) +
6586 			    TXG_DEFER_SIZE + 1;
6587 			spa_config_exit(spa, SCL_ALL, FTAG);
6588 		}
6589 	}
6590 
6591 export_spa:
6592 	spa_export_os(spa);
6593 
6594 	if (new_state == POOL_STATE_DESTROYED)
6595 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY);
6596 	else if (new_state == POOL_STATE_EXPORTED)
6597 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_EXPORT);
6598 
6599 	if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
6600 		spa_unload(spa);
6601 		spa_deactivate(spa);
6602 	}
6603 
6604 	if (oldconfig && spa->spa_config)
6605 		*oldconfig = fnvlist_dup(spa->spa_config);
6606 
6607 	if (new_state != POOL_STATE_UNINITIALIZED) {
6608 		if (!hardforce)
6609 			spa_write_cachefile(spa, B_TRUE, B_TRUE, B_FALSE);
6610 		spa_remove(spa);
6611 	} else {
6612 		/*
6613 		 * If spa_remove() is not called for this spa_t and
6614 		 * there is any possibility that it can be reused,
6615 		 * we make sure to reset the exporting flag.
6616 		 */
6617 		spa->spa_is_exporting = B_FALSE;
6618 	}
6619 
6620 	mutex_exit(&spa_namespace_lock);
6621 	return (0);
6622 
6623 fail:
6624 	spa->spa_is_exporting = B_FALSE;
6625 	spa_async_resume(spa);
6626 	mutex_exit(&spa_namespace_lock);
6627 	return (error);
6628 }
6629 
6630 /*
6631  * Destroy a storage pool.
6632  */
6633 int
6634 spa_destroy(const char *pool)
6635 {
6636 	return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL,
6637 	    B_FALSE, B_FALSE));
6638 }
6639 
6640 /*
6641  * Export a storage pool.
6642  */
6643 int
6644 spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
6645     boolean_t hardforce)
6646 {
6647 	return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig,
6648 	    force, hardforce));
6649 }
6650 
6651 /*
6652  * Similar to spa_export(), this unloads the spa_t without actually removing it
6653  * from the namespace in any way.
6654  */
6655 int
6656 spa_reset(const char *pool)
6657 {
6658 	return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL,
6659 	    B_FALSE, B_FALSE));
6660 }
6661 
6662 /*
6663  * ==========================================================================
6664  * Device manipulation
6665  * ==========================================================================
6666  */
6667 
6668 /*
6669  * This is called as a synctask to increment the draid feature flag
6670  */
6671 static void
6672 spa_draid_feature_incr(void *arg, dmu_tx_t *tx)
6673 {
6674 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
6675 	int draid = (int)(uintptr_t)arg;
6676 
6677 	for (int c = 0; c < draid; c++)
6678 		spa_feature_incr(spa, SPA_FEATURE_DRAID, tx);
6679 }
6680 
6681 /*
6682  * Add a device to a storage pool.
6683  */
6684 int
6685 spa_vdev_add(spa_t *spa, nvlist_t *nvroot)
6686 {
6687 	uint64_t txg, ndraid = 0;
6688 	int error;
6689 	vdev_t *rvd = spa->spa_root_vdev;
6690 	vdev_t *vd, *tvd;
6691 	nvlist_t **spares, **l2cache;
6692 	uint_t nspares, nl2cache;
6693 
6694 	ASSERT(spa_writeable(spa));
6695 
6696 	txg = spa_vdev_enter(spa);
6697 
6698 	if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0,
6699 	    VDEV_ALLOC_ADD)) != 0)
6700 		return (spa_vdev_exit(spa, NULL, txg, error));
6701 
6702 	spa->spa_pending_vdev = vd;	/* spa_vdev_exit() will clear this */
6703 
6704 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares,
6705 	    &nspares) != 0)
6706 		nspares = 0;
6707 
6708 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache,
6709 	    &nl2cache) != 0)
6710 		nl2cache = 0;
6711 
6712 	if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0)
6713 		return (spa_vdev_exit(spa, vd, txg, EINVAL));
6714 
6715 	if (vd->vdev_children != 0 &&
6716 	    (error = vdev_create(vd, txg, B_FALSE)) != 0) {
6717 		return (spa_vdev_exit(spa, vd, txg, error));
6718 	}
6719 
6720 	/*
6721 	 * The virtual dRAID spares must be added after vdev tree is created
6722 	 * and the vdev guids are generated.  The guid of their associated
6723 	 * dRAID is stored in the config and used when opening the spare.
6724 	 */
6725 	if ((error = vdev_draid_spare_create(nvroot, vd, &ndraid,
6726 	    rvd->vdev_children)) == 0) {
6727 		if (ndraid > 0 && nvlist_lookup_nvlist_array(nvroot,
6728 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0)
6729 			nspares = 0;
6730 	} else {
6731 		return (spa_vdev_exit(spa, vd, txg, error));
6732 	}
6733 
6734 	/*
6735 	 * We must validate the spares and l2cache devices after checking the
6736 	 * children.  Otherwise, vdev_inuse() will blindly overwrite the spare.
6737 	 */
6738 	if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0)
6739 		return (spa_vdev_exit(spa, vd, txg, error));
6740 
6741 	/*
6742 	 * If we are in the middle of a device removal, we can only add
6743 	 * devices which match the existing devices in the pool.
6744 	 * If we are in the middle of a removal, or have some indirect
6745 	 * vdevs, we can not add raidz or dRAID top levels.
6746 	 */
6747 	if (spa->spa_vdev_removal != NULL ||
6748 	    spa->spa_removing_phys.sr_prev_indirect_vdev != -1) {
6749 		for (int c = 0; c < vd->vdev_children; c++) {
6750 			tvd = vd->vdev_child[c];
6751 			if (spa->spa_vdev_removal != NULL &&
6752 			    tvd->vdev_ashift != spa->spa_max_ashift) {
6753 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
6754 			}
6755 			/* Fail if top level vdev is raidz or a dRAID */
6756 			if (vdev_get_nparity(tvd) != 0)
6757 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
6758 
6759 			/*
6760 			 * Need the top level mirror to be
6761 			 * a mirror of leaf vdevs only
6762 			 */
6763 			if (tvd->vdev_ops == &vdev_mirror_ops) {
6764 				for (uint64_t cid = 0;
6765 				    cid < tvd->vdev_children; cid++) {
6766 					vdev_t *cvd = tvd->vdev_child[cid];
6767 					if (!cvd->vdev_ops->vdev_op_leaf) {
6768 						return (spa_vdev_exit(spa, vd,
6769 						    txg, EINVAL));
6770 					}
6771 				}
6772 			}
6773 		}
6774 	}
6775 
6776 	for (int c = 0; c < vd->vdev_children; c++) {
6777 		tvd = vd->vdev_child[c];
6778 		vdev_remove_child(vd, tvd);
6779 		tvd->vdev_id = rvd->vdev_children;
6780 		vdev_add_child(rvd, tvd);
6781 		vdev_config_dirty(tvd);
6782 	}
6783 
6784 	if (nspares != 0) {
6785 		spa_set_aux_vdevs(&spa->spa_spares, spares, nspares,
6786 		    ZPOOL_CONFIG_SPARES);
6787 		spa_load_spares(spa);
6788 		spa->spa_spares.sav_sync = B_TRUE;
6789 	}
6790 
6791 	if (nl2cache != 0) {
6792 		spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache,
6793 		    ZPOOL_CONFIG_L2CACHE);
6794 		spa_load_l2cache(spa);
6795 		spa->spa_l2cache.sav_sync = B_TRUE;
6796 	}
6797 
6798 	/*
6799 	 * We can't increment a feature while holding spa_vdev so we
6800 	 * have to do it in a synctask.
6801 	 */
6802 	if (ndraid != 0) {
6803 		dmu_tx_t *tx;
6804 
6805 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
6806 		dsl_sync_task_nowait(spa->spa_dsl_pool, spa_draid_feature_incr,
6807 		    (void *)(uintptr_t)ndraid, tx);
6808 		dmu_tx_commit(tx);
6809 	}
6810 
6811 	/*
6812 	 * We have to be careful when adding new vdevs to an existing pool.
6813 	 * If other threads start allocating from these vdevs before we
6814 	 * sync the config cache, and we lose power, then upon reboot we may
6815 	 * fail to open the pool because there are DVAs that the config cache
6816 	 * can't translate.  Therefore, we first add the vdevs without
6817 	 * initializing metaslabs; sync the config cache (via spa_vdev_exit());
6818 	 * and then let spa_config_update() initialize the new metaslabs.
6819 	 *
6820 	 * spa_load() checks for added-but-not-initialized vdevs, so that
6821 	 * if we lose power at any point in this sequence, the remaining
6822 	 * steps will be completed the next time we load the pool.
6823 	 */
6824 	(void) spa_vdev_exit(spa, vd, txg, 0);
6825 
6826 	mutex_enter(&spa_namespace_lock);
6827 	spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
6828 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD);
6829 	mutex_exit(&spa_namespace_lock);
6830 
6831 	return (0);
6832 }
6833 
6834 /*
6835  * Attach a device to a mirror.  The arguments are the path to any device
6836  * in the mirror, and the nvroot for the new device.  If the path specifies
6837  * a device that is not mirrored, we automatically insert the mirror vdev.
6838  *
6839  * If 'replacing' is specified, the new device is intended to replace the
6840  * existing device; in this case the two devices are made into their own
6841  * mirror using the 'replacing' vdev, which is functionally identical to
6842  * the mirror vdev (it actually reuses all the same ops) but has a few
6843  * extra rules: you can't attach to it after it's been created, and upon
6844  * completion of resilvering, the first disk (the one being replaced)
6845  * is automatically detached.
6846  *
6847  * If 'rebuild' is specified, then sequential reconstruction (a.ka. rebuild)
6848  * should be performed instead of traditional healing reconstruction.  From
6849  * an administrators perspective these are both resilver operations.
6850  */
6851 int
6852 spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing,
6853     int rebuild)
6854 {
6855 	uint64_t txg, dtl_max_txg;
6856 	vdev_t *rvd = spa->spa_root_vdev;
6857 	vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd;
6858 	vdev_ops_t *pvops;
6859 	char *oldvdpath, *newvdpath;
6860 	int newvd_isspare;
6861 	int error;
6862 
6863 	ASSERT(spa_writeable(spa));
6864 
6865 	txg = spa_vdev_enter(spa);
6866 
6867 	oldvd = spa_lookup_by_guid(spa, guid, B_FALSE);
6868 
6869 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
6870 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6871 		error = (spa_has_checkpoint(spa)) ?
6872 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6873 		return (spa_vdev_exit(spa, NULL, txg, error));
6874 	}
6875 
6876 	if (rebuild) {
6877 		if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REBUILD))
6878 			return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6879 
6880 		if (dsl_scan_resilvering(spa_get_dsl(spa)) ||
6881 		    dsl_scan_resilver_scheduled(spa_get_dsl(spa))) {
6882 			return (spa_vdev_exit(spa, NULL, txg,
6883 			    ZFS_ERR_RESILVER_IN_PROGRESS));
6884 		}
6885 	} else {
6886 		if (vdev_rebuild_active(rvd))
6887 			return (spa_vdev_exit(spa, NULL, txg,
6888 			    ZFS_ERR_REBUILD_IN_PROGRESS));
6889 	}
6890 
6891 	if (spa->spa_vdev_removal != NULL)
6892 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6893 
6894 	if (oldvd == NULL)
6895 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
6896 
6897 	if (!oldvd->vdev_ops->vdev_op_leaf)
6898 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6899 
6900 	pvd = oldvd->vdev_parent;
6901 
6902 	if (spa_config_parse(spa, &newrootvd, nvroot, NULL, 0,
6903 	    VDEV_ALLOC_ATTACH) != 0)
6904 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6905 
6906 	if (newrootvd->vdev_children != 1)
6907 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6908 
6909 	newvd = newrootvd->vdev_child[0];
6910 
6911 	if (!newvd->vdev_ops->vdev_op_leaf)
6912 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6913 
6914 	if ((error = vdev_create(newrootvd, txg, replacing)) != 0)
6915 		return (spa_vdev_exit(spa, newrootvd, txg, error));
6916 
6917 	/*
6918 	 * log, dedup and special vdevs should not be replaced by spares.
6919 	 */
6920 	if ((oldvd->vdev_top->vdev_alloc_bias != VDEV_BIAS_NONE ||
6921 	    oldvd->vdev_top->vdev_islog) && newvd->vdev_isspare) {
6922 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6923 	}
6924 
6925 	/*
6926 	 * A dRAID spare can only replace a child of its parent dRAID vdev.
6927 	 */
6928 	if (newvd->vdev_ops == &vdev_draid_spare_ops &&
6929 	    oldvd->vdev_top != vdev_draid_spare_get_parent(newvd)) {
6930 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6931 	}
6932 
6933 	if (rebuild) {
6934 		/*
6935 		 * For rebuilds, the top vdev must support reconstruction
6936 		 * using only space maps.  This means the only allowable
6937 		 * vdevs types are the root vdev, a mirror, or dRAID.
6938 		 */
6939 		tvd = pvd;
6940 		if (pvd->vdev_top != NULL)
6941 			tvd = pvd->vdev_top;
6942 
6943 		if (tvd->vdev_ops != &vdev_mirror_ops &&
6944 		    tvd->vdev_ops != &vdev_root_ops &&
6945 		    tvd->vdev_ops != &vdev_draid_ops) {
6946 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6947 		}
6948 	}
6949 
6950 	if (!replacing) {
6951 		/*
6952 		 * For attach, the only allowable parent is a mirror or the root
6953 		 * vdev.
6954 		 */
6955 		if (pvd->vdev_ops != &vdev_mirror_ops &&
6956 		    pvd->vdev_ops != &vdev_root_ops)
6957 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6958 
6959 		pvops = &vdev_mirror_ops;
6960 	} else {
6961 		/*
6962 		 * Active hot spares can only be replaced by inactive hot
6963 		 * spares.
6964 		 */
6965 		if (pvd->vdev_ops == &vdev_spare_ops &&
6966 		    oldvd->vdev_isspare &&
6967 		    !spa_has_spare(spa, newvd->vdev_guid))
6968 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6969 
6970 		/*
6971 		 * If the source is a hot spare, and the parent isn't already a
6972 		 * spare, then we want to create a new hot spare.  Otherwise, we
6973 		 * want to create a replacing vdev.  The user is not allowed to
6974 		 * attach to a spared vdev child unless the 'isspare' state is
6975 		 * the same (spare replaces spare, non-spare replaces
6976 		 * non-spare).
6977 		 */
6978 		if (pvd->vdev_ops == &vdev_replacing_ops &&
6979 		    spa_version(spa) < SPA_VERSION_MULTI_REPLACE) {
6980 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6981 		} else if (pvd->vdev_ops == &vdev_spare_ops &&
6982 		    newvd->vdev_isspare != oldvd->vdev_isspare) {
6983 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6984 		}
6985 
6986 		if (newvd->vdev_isspare)
6987 			pvops = &vdev_spare_ops;
6988 		else
6989 			pvops = &vdev_replacing_ops;
6990 	}
6991 
6992 	/*
6993 	 * Make sure the new device is big enough.
6994 	 */
6995 	if (newvd->vdev_asize < vdev_get_min_asize(oldvd))
6996 		return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW));
6997 
6998 	/*
6999 	 * The new device cannot have a higher alignment requirement
7000 	 * than the top-level vdev.
7001 	 */
7002 	if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift)
7003 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
7004 
7005 	/*
7006 	 * If this is an in-place replacement, update oldvd's path and devid
7007 	 * to make it distinguishable from newvd, and unopenable from now on.
7008 	 */
7009 	if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) {
7010 		spa_strfree(oldvd->vdev_path);
7011 		oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5,
7012 		    KM_SLEEP);
7013 		(void) snprintf(oldvd->vdev_path, strlen(newvd->vdev_path) + 5,
7014 		    "%s/%s", newvd->vdev_path, "old");
7015 		if (oldvd->vdev_devid != NULL) {
7016 			spa_strfree(oldvd->vdev_devid);
7017 			oldvd->vdev_devid = NULL;
7018 		}
7019 	}
7020 
7021 	/*
7022 	 * If the parent is not a mirror, or if we're replacing, insert the new
7023 	 * mirror/replacing/spare vdev above oldvd.
7024 	 */
7025 	if (pvd->vdev_ops != pvops)
7026 		pvd = vdev_add_parent(oldvd, pvops);
7027 
7028 	ASSERT(pvd->vdev_top->vdev_parent == rvd);
7029 	ASSERT(pvd->vdev_ops == pvops);
7030 	ASSERT(oldvd->vdev_parent == pvd);
7031 
7032 	/*
7033 	 * Extract the new device from its root and add it to pvd.
7034 	 */
7035 	vdev_remove_child(newrootvd, newvd);
7036 	newvd->vdev_id = pvd->vdev_children;
7037 	newvd->vdev_crtxg = oldvd->vdev_crtxg;
7038 	vdev_add_child(pvd, newvd);
7039 
7040 	/*
7041 	 * Reevaluate the parent vdev state.
7042 	 */
7043 	vdev_propagate_state(pvd);
7044 
7045 	tvd = newvd->vdev_top;
7046 	ASSERT(pvd->vdev_top == tvd);
7047 	ASSERT(tvd->vdev_parent == rvd);
7048 
7049 	vdev_config_dirty(tvd);
7050 
7051 	/*
7052 	 * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account
7053 	 * for any dmu_sync-ed blocks.  It will propagate upward when
7054 	 * spa_vdev_exit() calls vdev_dtl_reassess().
7055 	 */
7056 	dtl_max_txg = txg + TXG_CONCURRENT_STATES;
7057 
7058 	vdev_dtl_dirty(newvd, DTL_MISSING,
7059 	    TXG_INITIAL, dtl_max_txg - TXG_INITIAL);
7060 
7061 	if (newvd->vdev_isspare) {
7062 		spa_spare_activate(newvd);
7063 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE);
7064 	}
7065 
7066 	oldvdpath = spa_strdup(oldvd->vdev_path);
7067 	newvdpath = spa_strdup(newvd->vdev_path);
7068 	newvd_isspare = newvd->vdev_isspare;
7069 
7070 	/*
7071 	 * Mark newvd's DTL dirty in this txg.
7072 	 */
7073 	vdev_dirty(tvd, VDD_DTL, newvd, txg);
7074 
7075 	/*
7076 	 * Schedule the resilver or rebuild to restart in the future. We do
7077 	 * this to ensure that dmu_sync-ed blocks have been stitched into the
7078 	 * respective datasets.
7079 	 */
7080 	if (rebuild) {
7081 		newvd->vdev_rebuild_txg = txg;
7082 
7083 		vdev_rebuild(tvd);
7084 	} else {
7085 		newvd->vdev_resilver_txg = txg;
7086 
7087 		if (dsl_scan_resilvering(spa_get_dsl(spa)) &&
7088 		    spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) {
7089 			vdev_defer_resilver(newvd);
7090 		} else {
7091 			dsl_scan_restart_resilver(spa->spa_dsl_pool,
7092 			    dtl_max_txg);
7093 		}
7094 	}
7095 
7096 	if (spa->spa_bootfs)
7097 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH);
7098 
7099 	spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH);
7100 
7101 	/*
7102 	 * Commit the config
7103 	 */
7104 	(void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0);
7105 
7106 	spa_history_log_internal(spa, "vdev attach", NULL,
7107 	    "%s vdev=%s %s vdev=%s",
7108 	    replacing && newvd_isspare ? "spare in" :
7109 	    replacing ? "replace" : "attach", newvdpath,
7110 	    replacing ? "for" : "to", oldvdpath);
7111 
7112 	spa_strfree(oldvdpath);
7113 	spa_strfree(newvdpath);
7114 
7115 	return (0);
7116 }
7117 
7118 /*
7119  * Detach a device from a mirror or replacing vdev.
7120  *
7121  * If 'replace_done' is specified, only detach if the parent
7122  * is a replacing or a spare vdev.
7123  */
7124 int
7125 spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done)
7126 {
7127 	uint64_t txg;
7128 	int error;
7129 	vdev_t *rvd __maybe_unused = spa->spa_root_vdev;
7130 	vdev_t *vd, *pvd, *cvd, *tvd;
7131 	boolean_t unspare = B_FALSE;
7132 	uint64_t unspare_guid = 0;
7133 	char *vdpath;
7134 
7135 	ASSERT(spa_writeable(spa));
7136 
7137 	txg = spa_vdev_detach_enter(spa, guid);
7138 
7139 	vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7140 
7141 	/*
7142 	 * Besides being called directly from the userland through the
7143 	 * ioctl interface, spa_vdev_detach() can be potentially called
7144 	 * at the end of spa_vdev_resilver_done().
7145 	 *
7146 	 * In the regular case, when we have a checkpoint this shouldn't
7147 	 * happen as we never empty the DTLs of a vdev during the scrub
7148 	 * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done()
7149 	 * should never get here when we have a checkpoint.
7150 	 *
7151 	 * That said, even in a case when we checkpoint the pool exactly
7152 	 * as spa_vdev_resilver_done() calls this function everything
7153 	 * should be fine as the resilver will return right away.
7154 	 */
7155 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7156 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
7157 		error = (spa_has_checkpoint(spa)) ?
7158 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
7159 		return (spa_vdev_exit(spa, NULL, txg, error));
7160 	}
7161 
7162 	if (vd == NULL)
7163 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
7164 
7165 	if (!vd->vdev_ops->vdev_op_leaf)
7166 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
7167 
7168 	pvd = vd->vdev_parent;
7169 
7170 	/*
7171 	 * If the parent/child relationship is not as expected, don't do it.
7172 	 * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing
7173 	 * vdev that's replacing B with C.  The user's intent in replacing
7174 	 * is to go from M(A,B) to M(A,C).  If the user decides to cancel
7175 	 * the replace by detaching C, the expected behavior is to end up
7176 	 * M(A,B).  But suppose that right after deciding to detach C,
7177 	 * the replacement of B completes.  We would have M(A,C), and then
7178 	 * ask to detach C, which would leave us with just A -- not what
7179 	 * the user wanted.  To prevent this, we make sure that the
7180 	 * parent/child relationship hasn't changed -- in this example,
7181 	 * that C's parent is still the replacing vdev R.
7182 	 */
7183 	if (pvd->vdev_guid != pguid && pguid != 0)
7184 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
7185 
7186 	/*
7187 	 * Only 'replacing' or 'spare' vdevs can be replaced.
7188 	 */
7189 	if (replace_done && pvd->vdev_ops != &vdev_replacing_ops &&
7190 	    pvd->vdev_ops != &vdev_spare_ops)
7191 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
7192 
7193 	ASSERT(pvd->vdev_ops != &vdev_spare_ops ||
7194 	    spa_version(spa) >= SPA_VERSION_SPARES);
7195 
7196 	/*
7197 	 * Only mirror, replacing, and spare vdevs support detach.
7198 	 */
7199 	if (pvd->vdev_ops != &vdev_replacing_ops &&
7200 	    pvd->vdev_ops != &vdev_mirror_ops &&
7201 	    pvd->vdev_ops != &vdev_spare_ops)
7202 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
7203 
7204 	/*
7205 	 * If this device has the only valid copy of some data,
7206 	 * we cannot safely detach it.
7207 	 */
7208 	if (vdev_dtl_required(vd))
7209 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
7210 
7211 	ASSERT(pvd->vdev_children >= 2);
7212 
7213 	/*
7214 	 * If we are detaching the second disk from a replacing vdev, then
7215 	 * check to see if we changed the original vdev's path to have "/old"
7216 	 * at the end in spa_vdev_attach().  If so, undo that change now.
7217 	 */
7218 	if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 &&
7219 	    vd->vdev_path != NULL) {
7220 		size_t len = strlen(vd->vdev_path);
7221 
7222 		for (int c = 0; c < pvd->vdev_children; c++) {
7223 			cvd = pvd->vdev_child[c];
7224 
7225 			if (cvd == vd || cvd->vdev_path == NULL)
7226 				continue;
7227 
7228 			if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 &&
7229 			    strcmp(cvd->vdev_path + len, "/old") == 0) {
7230 				spa_strfree(cvd->vdev_path);
7231 				cvd->vdev_path = spa_strdup(vd->vdev_path);
7232 				break;
7233 			}
7234 		}
7235 	}
7236 
7237 	/*
7238 	 * If we are detaching the original disk from a normal spare, then it
7239 	 * implies that the spare should become a real disk, and be removed
7240 	 * from the active spare list for the pool.  dRAID spares on the
7241 	 * other hand are coupled to the pool and thus should never be removed
7242 	 * from the spares list.
7243 	 */
7244 	if (pvd->vdev_ops == &vdev_spare_ops && vd->vdev_id == 0) {
7245 		vdev_t *last_cvd = pvd->vdev_child[pvd->vdev_children - 1];
7246 
7247 		if (last_cvd->vdev_isspare &&
7248 		    last_cvd->vdev_ops != &vdev_draid_spare_ops) {
7249 			unspare = B_TRUE;
7250 		}
7251 	}
7252 
7253 	/*
7254 	 * Erase the disk labels so the disk can be used for other things.
7255 	 * This must be done after all other error cases are handled,
7256 	 * but before we disembowel vd (so we can still do I/O to it).
7257 	 * But if we can't do it, don't treat the error as fatal --
7258 	 * it may be that the unwritability of the disk is the reason
7259 	 * it's being detached!
7260 	 */
7261 	(void) vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
7262 
7263 	/*
7264 	 * Remove vd from its parent and compact the parent's children.
7265 	 */
7266 	vdev_remove_child(pvd, vd);
7267 	vdev_compact_children(pvd);
7268 
7269 	/*
7270 	 * Remember one of the remaining children so we can get tvd below.
7271 	 */
7272 	cvd = pvd->vdev_child[pvd->vdev_children - 1];
7273 
7274 	/*
7275 	 * If we need to remove the remaining child from the list of hot spares,
7276 	 * do it now, marking the vdev as no longer a spare in the process.
7277 	 * We must do this before vdev_remove_parent(), because that can
7278 	 * change the GUID if it creates a new toplevel GUID.  For a similar
7279 	 * reason, we must remove the spare now, in the same txg as the detach;
7280 	 * otherwise someone could attach a new sibling, change the GUID, and
7281 	 * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail.
7282 	 */
7283 	if (unspare) {
7284 		ASSERT(cvd->vdev_isspare);
7285 		spa_spare_remove(cvd);
7286 		unspare_guid = cvd->vdev_guid;
7287 		(void) spa_vdev_remove(spa, unspare_guid, B_TRUE);
7288 		cvd->vdev_unspare = B_TRUE;
7289 	}
7290 
7291 	/*
7292 	 * If the parent mirror/replacing vdev only has one child,
7293 	 * the parent is no longer needed.  Remove it from the tree.
7294 	 */
7295 	if (pvd->vdev_children == 1) {
7296 		if (pvd->vdev_ops == &vdev_spare_ops)
7297 			cvd->vdev_unspare = B_FALSE;
7298 		vdev_remove_parent(cvd);
7299 	}
7300 
7301 	/*
7302 	 * We don't set tvd until now because the parent we just removed
7303 	 * may have been the previous top-level vdev.
7304 	 */
7305 	tvd = cvd->vdev_top;
7306 	ASSERT(tvd->vdev_parent == rvd);
7307 
7308 	/*
7309 	 * Reevaluate the parent vdev state.
7310 	 */
7311 	vdev_propagate_state(cvd);
7312 
7313 	/*
7314 	 * If the 'autoexpand' property is set on the pool then automatically
7315 	 * try to expand the size of the pool. For example if the device we
7316 	 * just detached was smaller than the others, it may be possible to
7317 	 * add metaslabs (i.e. grow the pool). We need to reopen the vdev
7318 	 * first so that we can obtain the updated sizes of the leaf vdevs.
7319 	 */
7320 	if (spa->spa_autoexpand) {
7321 		vdev_reopen(tvd);
7322 		vdev_expand(tvd, txg);
7323 	}
7324 
7325 	vdev_config_dirty(tvd);
7326 
7327 	/*
7328 	 * Mark vd's DTL as dirty in this txg.  vdev_dtl_sync() will see that
7329 	 * vd->vdev_detached is set and free vd's DTL object in syncing context.
7330 	 * But first make sure we're not on any *other* txg's DTL list, to
7331 	 * prevent vd from being accessed after it's freed.
7332 	 */
7333 	vdpath = spa_strdup(vd->vdev_path ? vd->vdev_path : "none");
7334 	for (int t = 0; t < TXG_SIZE; t++)
7335 		(void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t);
7336 	vd->vdev_detached = B_TRUE;
7337 	vdev_dirty(tvd, VDD_DTL, vd, txg);
7338 
7339 	spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE);
7340 	spa_notify_waiters(spa);
7341 
7342 	/* hang on to the spa before we release the lock */
7343 	spa_open_ref(spa, FTAG);
7344 
7345 	error = spa_vdev_exit(spa, vd, txg, 0);
7346 
7347 	spa_history_log_internal(spa, "detach", NULL,
7348 	    "vdev=%s", vdpath);
7349 	spa_strfree(vdpath);
7350 
7351 	/*
7352 	 * If this was the removal of the original device in a hot spare vdev,
7353 	 * then we want to go through and remove the device from the hot spare
7354 	 * list of every other pool.
7355 	 */
7356 	if (unspare) {
7357 		spa_t *altspa = NULL;
7358 
7359 		mutex_enter(&spa_namespace_lock);
7360 		while ((altspa = spa_next(altspa)) != NULL) {
7361 			if (altspa->spa_state != POOL_STATE_ACTIVE ||
7362 			    altspa == spa)
7363 				continue;
7364 
7365 			spa_open_ref(altspa, FTAG);
7366 			mutex_exit(&spa_namespace_lock);
7367 			(void) spa_vdev_remove(altspa, unspare_guid, B_TRUE);
7368 			mutex_enter(&spa_namespace_lock);
7369 			spa_close(altspa, FTAG);
7370 		}
7371 		mutex_exit(&spa_namespace_lock);
7372 
7373 		/* search the rest of the vdevs for spares to remove */
7374 		spa_vdev_resilver_done(spa);
7375 	}
7376 
7377 	/* all done with the spa; OK to release */
7378 	mutex_enter(&spa_namespace_lock);
7379 	spa_close(spa, FTAG);
7380 	mutex_exit(&spa_namespace_lock);
7381 
7382 	return (error);
7383 }
7384 
7385 static int
7386 spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
7387     list_t *vd_list)
7388 {
7389 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7390 
7391 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
7392 
7393 	/* Look up vdev and ensure it's a leaf. */
7394 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7395 	if (vd == NULL || vd->vdev_detached) {
7396 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7397 		return (SET_ERROR(ENODEV));
7398 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
7399 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7400 		return (SET_ERROR(EINVAL));
7401 	} else if (!vdev_writeable(vd)) {
7402 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7403 		return (SET_ERROR(EROFS));
7404 	}
7405 	mutex_enter(&vd->vdev_initialize_lock);
7406 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7407 
7408 	/*
7409 	 * When we activate an initialize action we check to see
7410 	 * if the vdev_initialize_thread is NULL. We do this instead
7411 	 * of using the vdev_initialize_state since there might be
7412 	 * a previous initialization process which has completed but
7413 	 * the thread is not exited.
7414 	 */
7415 	if (cmd_type == POOL_INITIALIZE_START &&
7416 	    (vd->vdev_initialize_thread != NULL ||
7417 	    vd->vdev_top->vdev_removing)) {
7418 		mutex_exit(&vd->vdev_initialize_lock);
7419 		return (SET_ERROR(EBUSY));
7420 	} else if (cmd_type == POOL_INITIALIZE_CANCEL &&
7421 	    (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE &&
7422 	    vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) {
7423 		mutex_exit(&vd->vdev_initialize_lock);
7424 		return (SET_ERROR(ESRCH));
7425 	} else if (cmd_type == POOL_INITIALIZE_SUSPEND &&
7426 	    vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) {
7427 		mutex_exit(&vd->vdev_initialize_lock);
7428 		return (SET_ERROR(ESRCH));
7429 	} else if (cmd_type == POOL_INITIALIZE_UNINIT &&
7430 	    vd->vdev_initialize_thread != NULL) {
7431 		mutex_exit(&vd->vdev_initialize_lock);
7432 		return (SET_ERROR(EBUSY));
7433 	}
7434 
7435 	switch (cmd_type) {
7436 	case POOL_INITIALIZE_START:
7437 		vdev_initialize(vd);
7438 		break;
7439 	case POOL_INITIALIZE_CANCEL:
7440 		vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, vd_list);
7441 		break;
7442 	case POOL_INITIALIZE_SUSPEND:
7443 		vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED, vd_list);
7444 		break;
7445 	case POOL_INITIALIZE_UNINIT:
7446 		vdev_uninitialize(vd);
7447 		break;
7448 	default:
7449 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
7450 	}
7451 	mutex_exit(&vd->vdev_initialize_lock);
7452 
7453 	return (0);
7454 }
7455 
7456 int
7457 spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
7458     nvlist_t *vdev_errlist)
7459 {
7460 	int total_errors = 0;
7461 	list_t vd_list;
7462 
7463 	list_create(&vd_list, sizeof (vdev_t),
7464 	    offsetof(vdev_t, vdev_initialize_node));
7465 
7466 	/*
7467 	 * We hold the namespace lock through the whole function
7468 	 * to prevent any changes to the pool while we're starting or
7469 	 * stopping initialization. The config and state locks are held so that
7470 	 * we can properly assess the vdev state before we commit to
7471 	 * the initializing operation.
7472 	 */
7473 	mutex_enter(&spa_namespace_lock);
7474 
7475 	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
7476 	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
7477 		uint64_t vdev_guid = fnvpair_value_uint64(pair);
7478 
7479 		int error = spa_vdev_initialize_impl(spa, vdev_guid, cmd_type,
7480 		    &vd_list);
7481 		if (error != 0) {
7482 			char guid_as_str[MAXNAMELEN];
7483 
7484 			(void) snprintf(guid_as_str, sizeof (guid_as_str),
7485 			    "%llu", (unsigned long long)vdev_guid);
7486 			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
7487 			total_errors++;
7488 		}
7489 	}
7490 
7491 	/* Wait for all initialize threads to stop. */
7492 	vdev_initialize_stop_wait(spa, &vd_list);
7493 
7494 	/* Sync out the initializing state */
7495 	txg_wait_synced(spa->spa_dsl_pool, 0);
7496 	mutex_exit(&spa_namespace_lock);
7497 
7498 	list_destroy(&vd_list);
7499 
7500 	return (total_errors);
7501 }
7502 
7503 static int
7504 spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
7505     uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list)
7506 {
7507 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7508 
7509 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
7510 
7511 	/* Look up vdev and ensure it's a leaf. */
7512 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7513 	if (vd == NULL || vd->vdev_detached) {
7514 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7515 		return (SET_ERROR(ENODEV));
7516 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
7517 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7518 		return (SET_ERROR(EINVAL));
7519 	} else if (!vdev_writeable(vd)) {
7520 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7521 		return (SET_ERROR(EROFS));
7522 	} else if (!vd->vdev_has_trim) {
7523 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7524 		return (SET_ERROR(EOPNOTSUPP));
7525 	} else if (secure && !vd->vdev_has_securetrim) {
7526 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7527 		return (SET_ERROR(EOPNOTSUPP));
7528 	}
7529 	mutex_enter(&vd->vdev_trim_lock);
7530 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7531 
7532 	/*
7533 	 * When we activate a TRIM action we check to see if the
7534 	 * vdev_trim_thread is NULL. We do this instead of using the
7535 	 * vdev_trim_state since there might be a previous TRIM process
7536 	 * which has completed but the thread is not exited.
7537 	 */
7538 	if (cmd_type == POOL_TRIM_START &&
7539 	    (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) {
7540 		mutex_exit(&vd->vdev_trim_lock);
7541 		return (SET_ERROR(EBUSY));
7542 	} else if (cmd_type == POOL_TRIM_CANCEL &&
7543 	    (vd->vdev_trim_state != VDEV_TRIM_ACTIVE &&
7544 	    vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) {
7545 		mutex_exit(&vd->vdev_trim_lock);
7546 		return (SET_ERROR(ESRCH));
7547 	} else if (cmd_type == POOL_TRIM_SUSPEND &&
7548 	    vd->vdev_trim_state != VDEV_TRIM_ACTIVE) {
7549 		mutex_exit(&vd->vdev_trim_lock);
7550 		return (SET_ERROR(ESRCH));
7551 	}
7552 
7553 	switch (cmd_type) {
7554 	case POOL_TRIM_START:
7555 		vdev_trim(vd, rate, partial, secure);
7556 		break;
7557 	case POOL_TRIM_CANCEL:
7558 		vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list);
7559 		break;
7560 	case POOL_TRIM_SUSPEND:
7561 		vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list);
7562 		break;
7563 	default:
7564 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
7565 	}
7566 	mutex_exit(&vd->vdev_trim_lock);
7567 
7568 	return (0);
7569 }
7570 
7571 /*
7572  * Initiates a manual TRIM for the requested vdevs. This kicks off individual
7573  * TRIM threads for each child vdev.  These threads pass over all of the free
7574  * space in the vdev's metaslabs and issues TRIM commands for that space.
7575  */
7576 int
7577 spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate,
7578     boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist)
7579 {
7580 	int total_errors = 0;
7581 	list_t vd_list;
7582 
7583 	list_create(&vd_list, sizeof (vdev_t),
7584 	    offsetof(vdev_t, vdev_trim_node));
7585 
7586 	/*
7587 	 * We hold the namespace lock through the whole function
7588 	 * to prevent any changes to the pool while we're starting or
7589 	 * stopping TRIM. The config and state locks are held so that
7590 	 * we can properly assess the vdev state before we commit to
7591 	 * the TRIM operation.
7592 	 */
7593 	mutex_enter(&spa_namespace_lock);
7594 
7595 	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
7596 	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
7597 		uint64_t vdev_guid = fnvpair_value_uint64(pair);
7598 
7599 		int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type,
7600 		    rate, partial, secure, &vd_list);
7601 		if (error != 0) {
7602 			char guid_as_str[MAXNAMELEN];
7603 
7604 			(void) snprintf(guid_as_str, sizeof (guid_as_str),
7605 			    "%llu", (unsigned long long)vdev_guid);
7606 			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
7607 			total_errors++;
7608 		}
7609 	}
7610 
7611 	/* Wait for all TRIM threads to stop. */
7612 	vdev_trim_stop_wait(spa, &vd_list);
7613 
7614 	/* Sync out the TRIM state */
7615 	txg_wait_synced(spa->spa_dsl_pool, 0);
7616 	mutex_exit(&spa_namespace_lock);
7617 
7618 	list_destroy(&vd_list);
7619 
7620 	return (total_errors);
7621 }
7622 
7623 /*
7624  * Split a set of devices from their mirrors, and create a new pool from them.
7625  */
7626 int
7627 spa_vdev_split_mirror(spa_t *spa, const char *newname, nvlist_t *config,
7628     nvlist_t *props, boolean_t exp)
7629 {
7630 	int error = 0;
7631 	uint64_t txg, *glist;
7632 	spa_t *newspa;
7633 	uint_t c, children, lastlog;
7634 	nvlist_t **child, *nvl, *tmp;
7635 	dmu_tx_t *tx;
7636 	const char *altroot = NULL;
7637 	vdev_t *rvd, **vml = NULL;			/* vdev modify list */
7638 	boolean_t activate_slog;
7639 
7640 	ASSERT(spa_writeable(spa));
7641 
7642 	txg = spa_vdev_enter(spa);
7643 
7644 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7645 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
7646 		error = (spa_has_checkpoint(spa)) ?
7647 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
7648 		return (spa_vdev_exit(spa, NULL, txg, error));
7649 	}
7650 
7651 	/* clear the log and flush everything up to now */
7652 	activate_slog = spa_passivate_log(spa);
7653 	(void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
7654 	error = spa_reset_logs(spa);
7655 	txg = spa_vdev_config_enter(spa);
7656 
7657 	if (activate_slog)
7658 		spa_activate_log(spa);
7659 
7660 	if (error != 0)
7661 		return (spa_vdev_exit(spa, NULL, txg, error));
7662 
7663 	/* check new spa name before going any further */
7664 	if (spa_lookup(newname) != NULL)
7665 		return (spa_vdev_exit(spa, NULL, txg, EEXIST));
7666 
7667 	/*
7668 	 * scan through all the children to ensure they're all mirrors
7669 	 */
7670 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 ||
7671 	    nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child,
7672 	    &children) != 0)
7673 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7674 
7675 	/* first, check to ensure we've got the right child count */
7676 	rvd = spa->spa_root_vdev;
7677 	lastlog = 0;
7678 	for (c = 0; c < rvd->vdev_children; c++) {
7679 		vdev_t *vd = rvd->vdev_child[c];
7680 
7681 		/* don't count the holes & logs as children */
7682 		if (vd->vdev_islog || (vd->vdev_ops != &vdev_indirect_ops &&
7683 		    !vdev_is_concrete(vd))) {
7684 			if (lastlog == 0)
7685 				lastlog = c;
7686 			continue;
7687 		}
7688 
7689 		lastlog = 0;
7690 	}
7691 	if (children != (lastlog != 0 ? lastlog : rvd->vdev_children))
7692 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7693 
7694 	/* next, ensure no spare or cache devices are part of the split */
7695 	if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 ||
7696 	    nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0)
7697 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7698 
7699 	vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP);
7700 	glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP);
7701 
7702 	/* then, loop over each vdev and validate it */
7703 	for (c = 0; c < children; c++) {
7704 		uint64_t is_hole = 0;
7705 
7706 		(void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE,
7707 		    &is_hole);
7708 
7709 		if (is_hole != 0) {
7710 			if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole ||
7711 			    spa->spa_root_vdev->vdev_child[c]->vdev_islog) {
7712 				continue;
7713 			} else {
7714 				error = SET_ERROR(EINVAL);
7715 				break;
7716 			}
7717 		}
7718 
7719 		/* deal with indirect vdevs */
7720 		if (spa->spa_root_vdev->vdev_child[c]->vdev_ops ==
7721 		    &vdev_indirect_ops)
7722 			continue;
7723 
7724 		/* which disk is going to be split? */
7725 		if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID,
7726 		    &glist[c]) != 0) {
7727 			error = SET_ERROR(EINVAL);
7728 			break;
7729 		}
7730 
7731 		/* look it up in the spa */
7732 		vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE);
7733 		if (vml[c] == NULL) {
7734 			error = SET_ERROR(ENODEV);
7735 			break;
7736 		}
7737 
7738 		/* make sure there's nothing stopping the split */
7739 		if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops ||
7740 		    vml[c]->vdev_islog ||
7741 		    !vdev_is_concrete(vml[c]) ||
7742 		    vml[c]->vdev_isspare ||
7743 		    vml[c]->vdev_isl2cache ||
7744 		    !vdev_writeable(vml[c]) ||
7745 		    vml[c]->vdev_children != 0 ||
7746 		    vml[c]->vdev_state != VDEV_STATE_HEALTHY ||
7747 		    c != spa->spa_root_vdev->vdev_child[c]->vdev_id) {
7748 			error = SET_ERROR(EINVAL);
7749 			break;
7750 		}
7751 
7752 		if (vdev_dtl_required(vml[c]) ||
7753 		    vdev_resilver_needed(vml[c], NULL, NULL)) {
7754 			error = SET_ERROR(EBUSY);
7755 			break;
7756 		}
7757 
7758 		/* we need certain info from the top level */
7759 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY,
7760 		    vml[c]->vdev_top->vdev_ms_array);
7761 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT,
7762 		    vml[c]->vdev_top->vdev_ms_shift);
7763 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE,
7764 		    vml[c]->vdev_top->vdev_asize);
7765 		fnvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT,
7766 		    vml[c]->vdev_top->vdev_ashift);
7767 
7768 		/* transfer per-vdev ZAPs */
7769 		ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0);
7770 		VERIFY0(nvlist_add_uint64(child[c],
7771 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap));
7772 
7773 		ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0);
7774 		VERIFY0(nvlist_add_uint64(child[c],
7775 		    ZPOOL_CONFIG_VDEV_TOP_ZAP,
7776 		    vml[c]->vdev_parent->vdev_top_zap));
7777 	}
7778 
7779 	if (error != 0) {
7780 		kmem_free(vml, children * sizeof (vdev_t *));
7781 		kmem_free(glist, children * sizeof (uint64_t));
7782 		return (spa_vdev_exit(spa, NULL, txg, error));
7783 	}
7784 
7785 	/* stop writers from using the disks */
7786 	for (c = 0; c < children; c++) {
7787 		if (vml[c] != NULL)
7788 			vml[c]->vdev_offline = B_TRUE;
7789 	}
7790 	vdev_reopen(spa->spa_root_vdev);
7791 
7792 	/*
7793 	 * Temporarily record the splitting vdevs in the spa config.  This
7794 	 * will disappear once the config is regenerated.
7795 	 */
7796 	nvl = fnvlist_alloc();
7797 	fnvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST, glist, children);
7798 	kmem_free(glist, children * sizeof (uint64_t));
7799 
7800 	mutex_enter(&spa->spa_props_lock);
7801 	fnvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT, nvl);
7802 	mutex_exit(&spa->spa_props_lock);
7803 	spa->spa_config_splitting = nvl;
7804 	vdev_config_dirty(spa->spa_root_vdev);
7805 
7806 	/* configure and create the new pool */
7807 	fnvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname);
7808 	fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE,
7809 	    exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE);
7810 	fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION, spa_version(spa));
7811 	fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG, spa->spa_config_txg);
7812 	fnvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID,
7813 	    spa_generate_guid(NULL));
7814 	VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
7815 	(void) nvlist_lookup_string(props,
7816 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
7817 
7818 	/* add the new pool to the namespace */
7819 	newspa = spa_add(newname, config, altroot);
7820 	newspa->spa_avz_action = AVZ_ACTION_REBUILD;
7821 	newspa->spa_config_txg = spa->spa_config_txg;
7822 	spa_set_log_state(newspa, SPA_LOG_CLEAR);
7823 
7824 	/* release the spa config lock, retaining the namespace lock */
7825 	spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
7826 
7827 	if (zio_injection_enabled)
7828 		zio_handle_panic_injection(spa, FTAG, 1);
7829 
7830 	spa_activate(newspa, spa_mode_global);
7831 	spa_async_suspend(newspa);
7832 
7833 	/*
7834 	 * Temporarily stop the initializing and TRIM activity.  We set the
7835 	 * state to ACTIVE so that we know to resume initializing or TRIM
7836 	 * once the split has completed.
7837 	 */
7838 	list_t vd_initialize_list;
7839 	list_create(&vd_initialize_list, sizeof (vdev_t),
7840 	    offsetof(vdev_t, vdev_initialize_node));
7841 
7842 	list_t vd_trim_list;
7843 	list_create(&vd_trim_list, sizeof (vdev_t),
7844 	    offsetof(vdev_t, vdev_trim_node));
7845 
7846 	for (c = 0; c < children; c++) {
7847 		if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) {
7848 			mutex_enter(&vml[c]->vdev_initialize_lock);
7849 			vdev_initialize_stop(vml[c],
7850 			    VDEV_INITIALIZE_ACTIVE, &vd_initialize_list);
7851 			mutex_exit(&vml[c]->vdev_initialize_lock);
7852 
7853 			mutex_enter(&vml[c]->vdev_trim_lock);
7854 			vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list);
7855 			mutex_exit(&vml[c]->vdev_trim_lock);
7856 		}
7857 	}
7858 
7859 	vdev_initialize_stop_wait(spa, &vd_initialize_list);
7860 	vdev_trim_stop_wait(spa, &vd_trim_list);
7861 
7862 	list_destroy(&vd_initialize_list);
7863 	list_destroy(&vd_trim_list);
7864 
7865 	newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT;
7866 	newspa->spa_is_splitting = B_TRUE;
7867 
7868 	/* create the new pool from the disks of the original pool */
7869 	error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE);
7870 	if (error)
7871 		goto out;
7872 
7873 	/* if that worked, generate a real config for the new pool */
7874 	if (newspa->spa_root_vdev != NULL) {
7875 		newspa->spa_config_splitting = fnvlist_alloc();
7876 		fnvlist_add_uint64(newspa->spa_config_splitting,
7877 		    ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa));
7878 		spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL,
7879 		    B_TRUE));
7880 	}
7881 
7882 	/* set the props */
7883 	if (props != NULL) {
7884 		spa_configfile_set(newspa, props, B_FALSE);
7885 		error = spa_prop_set(newspa, props);
7886 		if (error)
7887 			goto out;
7888 	}
7889 
7890 	/* flush everything */
7891 	txg = spa_vdev_config_enter(newspa);
7892 	vdev_config_dirty(newspa->spa_root_vdev);
7893 	(void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG);
7894 
7895 	if (zio_injection_enabled)
7896 		zio_handle_panic_injection(spa, FTAG, 2);
7897 
7898 	spa_async_resume(newspa);
7899 
7900 	/* finally, update the original pool's config */
7901 	txg = spa_vdev_config_enter(spa);
7902 	tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
7903 	error = dmu_tx_assign(tx, TXG_WAIT);
7904 	if (error != 0)
7905 		dmu_tx_abort(tx);
7906 	for (c = 0; c < children; c++) {
7907 		if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) {
7908 			vdev_t *tvd = vml[c]->vdev_top;
7909 
7910 			/*
7911 			 * Need to be sure the detachable VDEV is not
7912 			 * on any *other* txg's DTL list to prevent it
7913 			 * from being accessed after it's freed.
7914 			 */
7915 			for (int t = 0; t < TXG_SIZE; t++) {
7916 				(void) txg_list_remove_this(
7917 				    &tvd->vdev_dtl_list, vml[c], t);
7918 			}
7919 
7920 			vdev_split(vml[c]);
7921 			if (error == 0)
7922 				spa_history_log_internal(spa, "detach", tx,
7923 				    "vdev=%s", vml[c]->vdev_path);
7924 
7925 			vdev_free(vml[c]);
7926 		}
7927 	}
7928 	spa->spa_avz_action = AVZ_ACTION_REBUILD;
7929 	vdev_config_dirty(spa->spa_root_vdev);
7930 	spa->spa_config_splitting = NULL;
7931 	nvlist_free(nvl);
7932 	if (error == 0)
7933 		dmu_tx_commit(tx);
7934 	(void) spa_vdev_exit(spa, NULL, txg, 0);
7935 
7936 	if (zio_injection_enabled)
7937 		zio_handle_panic_injection(spa, FTAG, 3);
7938 
7939 	/* split is complete; log a history record */
7940 	spa_history_log_internal(newspa, "split", NULL,
7941 	    "from pool %s", spa_name(spa));
7942 
7943 	newspa->spa_is_splitting = B_FALSE;
7944 	kmem_free(vml, children * sizeof (vdev_t *));
7945 
7946 	/* if we're not going to mount the filesystems in userland, export */
7947 	if (exp)
7948 		error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL,
7949 		    B_FALSE, B_FALSE);
7950 
7951 	return (error);
7952 
7953 out:
7954 	spa_unload(newspa);
7955 	spa_deactivate(newspa);
7956 	spa_remove(newspa);
7957 
7958 	txg = spa_vdev_config_enter(spa);
7959 
7960 	/* re-online all offlined disks */
7961 	for (c = 0; c < children; c++) {
7962 		if (vml[c] != NULL)
7963 			vml[c]->vdev_offline = B_FALSE;
7964 	}
7965 
7966 	/* restart initializing or trimming disks as necessary */
7967 	spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
7968 	spa_async_request(spa, SPA_ASYNC_TRIM_RESTART);
7969 	spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART);
7970 
7971 	vdev_reopen(spa->spa_root_vdev);
7972 
7973 	nvlist_free(spa->spa_config_splitting);
7974 	spa->spa_config_splitting = NULL;
7975 	(void) spa_vdev_exit(spa, NULL, txg, error);
7976 
7977 	kmem_free(vml, children * sizeof (vdev_t *));
7978 	return (error);
7979 }
7980 
7981 /*
7982  * Find any device that's done replacing, or a vdev marked 'unspare' that's
7983  * currently spared, so we can detach it.
7984  */
7985 static vdev_t *
7986 spa_vdev_resilver_done_hunt(vdev_t *vd)
7987 {
7988 	vdev_t *newvd, *oldvd;
7989 
7990 	for (int c = 0; c < vd->vdev_children; c++) {
7991 		oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]);
7992 		if (oldvd != NULL)
7993 			return (oldvd);
7994 	}
7995 
7996 	/*
7997 	 * Check for a completed replacement.  We always consider the first
7998 	 * vdev in the list to be the oldest vdev, and the last one to be
7999 	 * the newest (see spa_vdev_attach() for how that works).  In
8000 	 * the case where the newest vdev is faulted, we will not automatically
8001 	 * remove it after a resilver completes.  This is OK as it will require
8002 	 * user intervention to determine which disk the admin wishes to keep.
8003 	 */
8004 	if (vd->vdev_ops == &vdev_replacing_ops) {
8005 		ASSERT(vd->vdev_children > 1);
8006 
8007 		newvd = vd->vdev_child[vd->vdev_children - 1];
8008 		oldvd = vd->vdev_child[0];
8009 
8010 		if (vdev_dtl_empty(newvd, DTL_MISSING) &&
8011 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
8012 		    !vdev_dtl_required(oldvd))
8013 			return (oldvd);
8014 	}
8015 
8016 	/*
8017 	 * Check for a completed resilver with the 'unspare' flag set.
8018 	 * Also potentially update faulted state.
8019 	 */
8020 	if (vd->vdev_ops == &vdev_spare_ops) {
8021 		vdev_t *first = vd->vdev_child[0];
8022 		vdev_t *last = vd->vdev_child[vd->vdev_children - 1];
8023 
8024 		if (last->vdev_unspare) {
8025 			oldvd = first;
8026 			newvd = last;
8027 		} else if (first->vdev_unspare) {
8028 			oldvd = last;
8029 			newvd = first;
8030 		} else {
8031 			oldvd = NULL;
8032 		}
8033 
8034 		if (oldvd != NULL &&
8035 		    vdev_dtl_empty(newvd, DTL_MISSING) &&
8036 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
8037 		    !vdev_dtl_required(oldvd))
8038 			return (oldvd);
8039 
8040 		vdev_propagate_state(vd);
8041 
8042 		/*
8043 		 * If there are more than two spares attached to a disk,
8044 		 * and those spares are not required, then we want to
8045 		 * attempt to free them up now so that they can be used
8046 		 * by other pools.  Once we're back down to a single
8047 		 * disk+spare, we stop removing them.
8048 		 */
8049 		if (vd->vdev_children > 2) {
8050 			newvd = vd->vdev_child[1];
8051 
8052 			if (newvd->vdev_isspare && last->vdev_isspare &&
8053 			    vdev_dtl_empty(last, DTL_MISSING) &&
8054 			    vdev_dtl_empty(last, DTL_OUTAGE) &&
8055 			    !vdev_dtl_required(newvd))
8056 				return (newvd);
8057 		}
8058 	}
8059 
8060 	return (NULL);
8061 }
8062 
8063 static void
8064 spa_vdev_resilver_done(spa_t *spa)
8065 {
8066 	vdev_t *vd, *pvd, *ppvd;
8067 	uint64_t guid, sguid, pguid, ppguid;
8068 
8069 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
8070 
8071 	while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) {
8072 		pvd = vd->vdev_parent;
8073 		ppvd = pvd->vdev_parent;
8074 		guid = vd->vdev_guid;
8075 		pguid = pvd->vdev_guid;
8076 		ppguid = ppvd->vdev_guid;
8077 		sguid = 0;
8078 		/*
8079 		 * If we have just finished replacing a hot spared device, then
8080 		 * we need to detach the parent's first child (the original hot
8081 		 * spare) as well.
8082 		 */
8083 		if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 &&
8084 		    ppvd->vdev_children == 2) {
8085 			ASSERT(pvd->vdev_ops == &vdev_replacing_ops);
8086 			sguid = ppvd->vdev_child[1]->vdev_guid;
8087 		}
8088 		ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd));
8089 
8090 		spa_config_exit(spa, SCL_ALL, FTAG);
8091 		if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0)
8092 			return;
8093 		if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0)
8094 			return;
8095 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
8096 	}
8097 
8098 	spa_config_exit(spa, SCL_ALL, FTAG);
8099 
8100 	/*
8101 	 * If a detach was not performed above replace waiters will not have
8102 	 * been notified.  In which case we must do so now.
8103 	 */
8104 	spa_notify_waiters(spa);
8105 }
8106 
8107 /*
8108  * Update the stored path or FRU for this vdev.
8109  */
8110 static int
8111 spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value,
8112     boolean_t ispath)
8113 {
8114 	vdev_t *vd;
8115 	boolean_t sync = B_FALSE;
8116 
8117 	ASSERT(spa_writeable(spa));
8118 
8119 	spa_vdev_state_enter(spa, SCL_ALL);
8120 
8121 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
8122 		return (spa_vdev_state_exit(spa, NULL, ENOENT));
8123 
8124 	if (!vd->vdev_ops->vdev_op_leaf)
8125 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
8126 
8127 	if (ispath) {
8128 		if (strcmp(value, vd->vdev_path) != 0) {
8129 			spa_strfree(vd->vdev_path);
8130 			vd->vdev_path = spa_strdup(value);
8131 			sync = B_TRUE;
8132 		}
8133 	} else {
8134 		if (vd->vdev_fru == NULL) {
8135 			vd->vdev_fru = spa_strdup(value);
8136 			sync = B_TRUE;
8137 		} else if (strcmp(value, vd->vdev_fru) != 0) {
8138 			spa_strfree(vd->vdev_fru);
8139 			vd->vdev_fru = spa_strdup(value);
8140 			sync = B_TRUE;
8141 		}
8142 	}
8143 
8144 	return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0));
8145 }
8146 
8147 int
8148 spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath)
8149 {
8150 	return (spa_vdev_set_common(spa, guid, newpath, B_TRUE));
8151 }
8152 
8153 int
8154 spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru)
8155 {
8156 	return (spa_vdev_set_common(spa, guid, newfru, B_FALSE));
8157 }
8158 
8159 /*
8160  * ==========================================================================
8161  * SPA Scanning
8162  * ==========================================================================
8163  */
8164 int
8165 spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd)
8166 {
8167 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
8168 
8169 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
8170 		return (SET_ERROR(EBUSY));
8171 
8172 	return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd));
8173 }
8174 
8175 int
8176 spa_scan_stop(spa_t *spa)
8177 {
8178 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
8179 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
8180 		return (SET_ERROR(EBUSY));
8181 
8182 	return (dsl_scan_cancel(spa->spa_dsl_pool));
8183 }
8184 
8185 int
8186 spa_scan(spa_t *spa, pool_scan_func_t func)
8187 {
8188 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
8189 
8190 	if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE)
8191 		return (SET_ERROR(ENOTSUP));
8192 
8193 	if (func == POOL_SCAN_RESILVER &&
8194 	    !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER))
8195 		return (SET_ERROR(ENOTSUP));
8196 
8197 	/*
8198 	 * If a resilver was requested, but there is no DTL on a
8199 	 * writeable leaf device, we have nothing to do.
8200 	 */
8201 	if (func == POOL_SCAN_RESILVER &&
8202 	    !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
8203 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
8204 		return (0);
8205 	}
8206 
8207 	if (func == POOL_SCAN_ERRORSCRUB &&
8208 	    !spa_feature_is_enabled(spa, SPA_FEATURE_HEAD_ERRLOG))
8209 		return (SET_ERROR(ENOTSUP));
8210 
8211 	return (dsl_scan(spa->spa_dsl_pool, func));
8212 }
8213 
8214 /*
8215  * ==========================================================================
8216  * SPA async task processing
8217  * ==========================================================================
8218  */
8219 
8220 static void
8221 spa_async_remove(spa_t *spa, vdev_t *vd)
8222 {
8223 	if (vd->vdev_remove_wanted) {
8224 		vd->vdev_remove_wanted = B_FALSE;
8225 		vd->vdev_delayed_close = B_FALSE;
8226 		vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE);
8227 
8228 		/*
8229 		 * We want to clear the stats, but we don't want to do a full
8230 		 * vdev_clear() as that will cause us to throw away
8231 		 * degraded/faulted state as well as attempt to reopen the
8232 		 * device, all of which is a waste.
8233 		 */
8234 		vd->vdev_stat.vs_read_errors = 0;
8235 		vd->vdev_stat.vs_write_errors = 0;
8236 		vd->vdev_stat.vs_checksum_errors = 0;
8237 
8238 		vdev_state_dirty(vd->vdev_top);
8239 
8240 		/* Tell userspace that the vdev is gone. */
8241 		zfs_post_remove(spa, vd);
8242 	}
8243 
8244 	for (int c = 0; c < vd->vdev_children; c++)
8245 		spa_async_remove(spa, vd->vdev_child[c]);
8246 }
8247 
8248 static void
8249 spa_async_probe(spa_t *spa, vdev_t *vd)
8250 {
8251 	if (vd->vdev_probe_wanted) {
8252 		vd->vdev_probe_wanted = B_FALSE;
8253 		vdev_reopen(vd);	/* vdev_open() does the actual probe */
8254 	}
8255 
8256 	for (int c = 0; c < vd->vdev_children; c++)
8257 		spa_async_probe(spa, vd->vdev_child[c]);
8258 }
8259 
8260 static void
8261 spa_async_autoexpand(spa_t *spa, vdev_t *vd)
8262 {
8263 	if (!spa->spa_autoexpand)
8264 		return;
8265 
8266 	for (int c = 0; c < vd->vdev_children; c++) {
8267 		vdev_t *cvd = vd->vdev_child[c];
8268 		spa_async_autoexpand(spa, cvd);
8269 	}
8270 
8271 	if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL)
8272 		return;
8273 
8274 	spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_AUTOEXPAND);
8275 }
8276 
8277 static __attribute__((noreturn)) void
8278 spa_async_thread(void *arg)
8279 {
8280 	spa_t *spa = (spa_t *)arg;
8281 	dsl_pool_t *dp = spa->spa_dsl_pool;
8282 	int tasks;
8283 
8284 	ASSERT(spa->spa_sync_on);
8285 
8286 	mutex_enter(&spa->spa_async_lock);
8287 	tasks = spa->spa_async_tasks;
8288 	spa->spa_async_tasks = 0;
8289 	mutex_exit(&spa->spa_async_lock);
8290 
8291 	/*
8292 	 * See if the config needs to be updated.
8293 	 */
8294 	if (tasks & SPA_ASYNC_CONFIG_UPDATE) {
8295 		uint64_t old_space, new_space;
8296 
8297 		mutex_enter(&spa_namespace_lock);
8298 		old_space = metaslab_class_get_space(spa_normal_class(spa));
8299 		old_space += metaslab_class_get_space(spa_special_class(spa));
8300 		old_space += metaslab_class_get_space(spa_dedup_class(spa));
8301 		old_space += metaslab_class_get_space(
8302 		    spa_embedded_log_class(spa));
8303 
8304 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
8305 
8306 		new_space = metaslab_class_get_space(spa_normal_class(spa));
8307 		new_space += metaslab_class_get_space(spa_special_class(spa));
8308 		new_space += metaslab_class_get_space(spa_dedup_class(spa));
8309 		new_space += metaslab_class_get_space(
8310 		    spa_embedded_log_class(spa));
8311 		mutex_exit(&spa_namespace_lock);
8312 
8313 		/*
8314 		 * If the pool grew as a result of the config update,
8315 		 * then log an internal history event.
8316 		 */
8317 		if (new_space != old_space) {
8318 			spa_history_log_internal(spa, "vdev online", NULL,
8319 			    "pool '%s' size: %llu(+%llu)",
8320 			    spa_name(spa), (u_longlong_t)new_space,
8321 			    (u_longlong_t)(new_space - old_space));
8322 		}
8323 	}
8324 
8325 	/*
8326 	 * See if any devices need to be marked REMOVED.
8327 	 */
8328 	if (tasks & SPA_ASYNC_REMOVE) {
8329 		spa_vdev_state_enter(spa, SCL_NONE);
8330 		spa_async_remove(spa, spa->spa_root_vdev);
8331 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++)
8332 			spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]);
8333 		for (int i = 0; i < spa->spa_spares.sav_count; i++)
8334 			spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]);
8335 		(void) spa_vdev_state_exit(spa, NULL, 0);
8336 	}
8337 
8338 	if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) {
8339 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8340 		spa_async_autoexpand(spa, spa->spa_root_vdev);
8341 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8342 	}
8343 
8344 	/*
8345 	 * See if any devices need to be probed.
8346 	 */
8347 	if (tasks & SPA_ASYNC_PROBE) {
8348 		spa_vdev_state_enter(spa, SCL_NONE);
8349 		spa_async_probe(spa, spa->spa_root_vdev);
8350 		(void) spa_vdev_state_exit(spa, NULL, 0);
8351 	}
8352 
8353 	/*
8354 	 * If any devices are done replacing, detach them.
8355 	 */
8356 	if (tasks & SPA_ASYNC_RESILVER_DONE ||
8357 	    tasks & SPA_ASYNC_REBUILD_DONE ||
8358 	    tasks & SPA_ASYNC_DETACH_SPARE) {
8359 		spa_vdev_resilver_done(spa);
8360 	}
8361 
8362 	/*
8363 	 * Kick off a resilver.
8364 	 */
8365 	if (tasks & SPA_ASYNC_RESILVER &&
8366 	    !vdev_rebuild_active(spa->spa_root_vdev) &&
8367 	    (!dsl_scan_resilvering(dp) ||
8368 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER)))
8369 		dsl_scan_restart_resilver(dp, 0);
8370 
8371 	if (tasks & SPA_ASYNC_INITIALIZE_RESTART) {
8372 		mutex_enter(&spa_namespace_lock);
8373 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8374 		vdev_initialize_restart(spa->spa_root_vdev);
8375 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8376 		mutex_exit(&spa_namespace_lock);
8377 	}
8378 
8379 	if (tasks & SPA_ASYNC_TRIM_RESTART) {
8380 		mutex_enter(&spa_namespace_lock);
8381 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8382 		vdev_trim_restart(spa->spa_root_vdev);
8383 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8384 		mutex_exit(&spa_namespace_lock);
8385 	}
8386 
8387 	if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) {
8388 		mutex_enter(&spa_namespace_lock);
8389 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8390 		vdev_autotrim_restart(spa);
8391 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8392 		mutex_exit(&spa_namespace_lock);
8393 	}
8394 
8395 	/*
8396 	 * Kick off L2 cache whole device TRIM.
8397 	 */
8398 	if (tasks & SPA_ASYNC_L2CACHE_TRIM) {
8399 		mutex_enter(&spa_namespace_lock);
8400 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8401 		vdev_trim_l2arc(spa);
8402 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8403 		mutex_exit(&spa_namespace_lock);
8404 	}
8405 
8406 	/*
8407 	 * Kick off L2 cache rebuilding.
8408 	 */
8409 	if (tasks & SPA_ASYNC_L2CACHE_REBUILD) {
8410 		mutex_enter(&spa_namespace_lock);
8411 		spa_config_enter(spa, SCL_L2ARC, FTAG, RW_READER);
8412 		l2arc_spa_rebuild_start(spa);
8413 		spa_config_exit(spa, SCL_L2ARC, FTAG);
8414 		mutex_exit(&spa_namespace_lock);
8415 	}
8416 
8417 	/*
8418 	 * Let the world know that we're done.
8419 	 */
8420 	mutex_enter(&spa->spa_async_lock);
8421 	spa->spa_async_thread = NULL;
8422 	cv_broadcast(&spa->spa_async_cv);
8423 	mutex_exit(&spa->spa_async_lock);
8424 	thread_exit();
8425 }
8426 
8427 void
8428 spa_async_suspend(spa_t *spa)
8429 {
8430 	mutex_enter(&spa->spa_async_lock);
8431 	spa->spa_async_suspended++;
8432 	while (spa->spa_async_thread != NULL)
8433 		cv_wait(&spa->spa_async_cv, &spa->spa_async_lock);
8434 	mutex_exit(&spa->spa_async_lock);
8435 
8436 	spa_vdev_remove_suspend(spa);
8437 
8438 	zthr_t *condense_thread = spa->spa_condense_zthr;
8439 	if (condense_thread != NULL)
8440 		zthr_cancel(condense_thread);
8441 
8442 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
8443 	if (discard_thread != NULL)
8444 		zthr_cancel(discard_thread);
8445 
8446 	zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr;
8447 	if (ll_delete_thread != NULL)
8448 		zthr_cancel(ll_delete_thread);
8449 
8450 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
8451 	if (ll_condense_thread != NULL)
8452 		zthr_cancel(ll_condense_thread);
8453 }
8454 
8455 void
8456 spa_async_resume(spa_t *spa)
8457 {
8458 	mutex_enter(&spa->spa_async_lock);
8459 	ASSERT(spa->spa_async_suspended != 0);
8460 	spa->spa_async_suspended--;
8461 	mutex_exit(&spa->spa_async_lock);
8462 	spa_restart_removal(spa);
8463 
8464 	zthr_t *condense_thread = spa->spa_condense_zthr;
8465 	if (condense_thread != NULL)
8466 		zthr_resume(condense_thread);
8467 
8468 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
8469 	if (discard_thread != NULL)
8470 		zthr_resume(discard_thread);
8471 
8472 	zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr;
8473 	if (ll_delete_thread != NULL)
8474 		zthr_resume(ll_delete_thread);
8475 
8476 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
8477 	if (ll_condense_thread != NULL)
8478 		zthr_resume(ll_condense_thread);
8479 }
8480 
8481 static boolean_t
8482 spa_async_tasks_pending(spa_t *spa)
8483 {
8484 	uint_t non_config_tasks;
8485 	uint_t config_task;
8486 	boolean_t config_task_suspended;
8487 
8488 	non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE;
8489 	config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE;
8490 	if (spa->spa_ccw_fail_time == 0) {
8491 		config_task_suspended = B_FALSE;
8492 	} else {
8493 		config_task_suspended =
8494 		    (gethrtime() - spa->spa_ccw_fail_time) <
8495 		    ((hrtime_t)zfs_ccw_retry_interval * NANOSEC);
8496 	}
8497 
8498 	return (non_config_tasks || (config_task && !config_task_suspended));
8499 }
8500 
8501 static void
8502 spa_async_dispatch(spa_t *spa)
8503 {
8504 	mutex_enter(&spa->spa_async_lock);
8505 	if (spa_async_tasks_pending(spa) &&
8506 	    !spa->spa_async_suspended &&
8507 	    spa->spa_async_thread == NULL)
8508 		spa->spa_async_thread = thread_create(NULL, 0,
8509 		    spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri);
8510 	mutex_exit(&spa->spa_async_lock);
8511 }
8512 
8513 void
8514 spa_async_request(spa_t *spa, int task)
8515 {
8516 	zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task);
8517 	mutex_enter(&spa->spa_async_lock);
8518 	spa->spa_async_tasks |= task;
8519 	mutex_exit(&spa->spa_async_lock);
8520 }
8521 
8522 int
8523 spa_async_tasks(spa_t *spa)
8524 {
8525 	return (spa->spa_async_tasks);
8526 }
8527 
8528 /*
8529  * ==========================================================================
8530  * SPA syncing routines
8531  * ==========================================================================
8532  */
8533 
8534 
8535 static int
8536 bpobj_enqueue_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
8537     dmu_tx_t *tx)
8538 {
8539 	bpobj_t *bpo = arg;
8540 	bpobj_enqueue(bpo, bp, bp_freed, tx);
8541 	return (0);
8542 }
8543 
8544 int
8545 bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8546 {
8547 	return (bpobj_enqueue_cb(arg, bp, B_FALSE, tx));
8548 }
8549 
8550 int
8551 bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8552 {
8553 	return (bpobj_enqueue_cb(arg, bp, B_TRUE, tx));
8554 }
8555 
8556 static int
8557 spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8558 {
8559 	zio_t *pio = arg;
8560 
8561 	zio_nowait(zio_free_sync(pio, pio->io_spa, dmu_tx_get_txg(tx), bp,
8562 	    pio->io_flags));
8563 	return (0);
8564 }
8565 
8566 static int
8567 bpobj_spa_free_sync_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
8568     dmu_tx_t *tx)
8569 {
8570 	ASSERT(!bp_freed);
8571 	return (spa_free_sync_cb(arg, bp, tx));
8572 }
8573 
8574 /*
8575  * Note: this simple function is not inlined to make it easier to dtrace the
8576  * amount of time spent syncing frees.
8577  */
8578 static void
8579 spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx)
8580 {
8581 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
8582 	bplist_iterate(bpl, spa_free_sync_cb, zio, tx);
8583 	VERIFY(zio_wait(zio) == 0);
8584 }
8585 
8586 /*
8587  * Note: this simple function is not inlined to make it easier to dtrace the
8588  * amount of time spent syncing deferred frees.
8589  */
8590 static void
8591 spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx)
8592 {
8593 	if (spa_sync_pass(spa) != 1)
8594 		return;
8595 
8596 	/*
8597 	 * Note:
8598 	 * If the log space map feature is active, we stop deferring
8599 	 * frees to the next TXG and therefore running this function
8600 	 * would be considered a no-op as spa_deferred_bpobj should
8601 	 * not have any entries.
8602 	 *
8603 	 * That said we run this function anyway (instead of returning
8604 	 * immediately) for the edge-case scenario where we just
8605 	 * activated the log space map feature in this TXG but we have
8606 	 * deferred frees from the previous TXG.
8607 	 */
8608 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
8609 	VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj,
8610 	    bpobj_spa_free_sync_cb, zio, tx), ==, 0);
8611 	VERIFY0(zio_wait(zio));
8612 }
8613 
8614 static void
8615 spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx)
8616 {
8617 	char *packed = NULL;
8618 	size_t bufsize;
8619 	size_t nvsize = 0;
8620 	dmu_buf_t *db;
8621 
8622 	VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0);
8623 
8624 	/*
8625 	 * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration
8626 	 * information.  This avoids the dmu_buf_will_dirty() path and
8627 	 * saves us a pre-read to get data we don't actually care about.
8628 	 */
8629 	bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE);
8630 	packed = vmem_alloc(bufsize, KM_SLEEP);
8631 
8632 	VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR,
8633 	    KM_SLEEP) == 0);
8634 	memset(packed + nvsize, 0, bufsize - nvsize);
8635 
8636 	dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx);
8637 
8638 	vmem_free(packed, bufsize);
8639 
8640 	VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db));
8641 	dmu_buf_will_dirty(db, tx);
8642 	*(uint64_t *)db->db_data = nvsize;
8643 	dmu_buf_rele(db, FTAG);
8644 }
8645 
8646 static void
8647 spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx,
8648     const char *config, const char *entry)
8649 {
8650 	nvlist_t *nvroot;
8651 	nvlist_t **list;
8652 	int i;
8653 
8654 	if (!sav->sav_sync)
8655 		return;
8656 
8657 	/*
8658 	 * Update the MOS nvlist describing the list of available devices.
8659 	 * spa_validate_aux() will have already made sure this nvlist is
8660 	 * valid and the vdevs are labeled appropriately.
8661 	 */
8662 	if (sav->sav_object == 0) {
8663 		sav->sav_object = dmu_object_alloc(spa->spa_meta_objset,
8664 		    DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE,
8665 		    sizeof (uint64_t), tx);
8666 		VERIFY(zap_update(spa->spa_meta_objset,
8667 		    DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1,
8668 		    &sav->sav_object, tx) == 0);
8669 	}
8670 
8671 	nvroot = fnvlist_alloc();
8672 	if (sav->sav_count == 0) {
8673 		fnvlist_add_nvlist_array(nvroot, config,
8674 		    (const nvlist_t * const *)NULL, 0);
8675 	} else {
8676 		list = kmem_alloc(sav->sav_count*sizeof (void *), KM_SLEEP);
8677 		for (i = 0; i < sav->sav_count; i++)
8678 			list[i] = vdev_config_generate(spa, sav->sav_vdevs[i],
8679 			    B_FALSE, VDEV_CONFIG_L2CACHE);
8680 		fnvlist_add_nvlist_array(nvroot, config,
8681 		    (const nvlist_t * const *)list, sav->sav_count);
8682 		for (i = 0; i < sav->sav_count; i++)
8683 			nvlist_free(list[i]);
8684 		kmem_free(list, sav->sav_count * sizeof (void *));
8685 	}
8686 
8687 	spa_sync_nvlist(spa, sav->sav_object, nvroot, tx);
8688 	nvlist_free(nvroot);
8689 
8690 	sav->sav_sync = B_FALSE;
8691 }
8692 
8693 /*
8694  * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t.
8695  * The all-vdev ZAP must be empty.
8696  */
8697 static void
8698 spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx)
8699 {
8700 	spa_t *spa = vd->vdev_spa;
8701 
8702 	if (vd->vdev_root_zap != 0 &&
8703 	    spa_feature_is_active(spa, SPA_FEATURE_AVZ_V2)) {
8704 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8705 		    vd->vdev_root_zap, tx));
8706 	}
8707 	if (vd->vdev_top_zap != 0) {
8708 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8709 		    vd->vdev_top_zap, tx));
8710 	}
8711 	if (vd->vdev_leaf_zap != 0) {
8712 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8713 		    vd->vdev_leaf_zap, tx));
8714 	}
8715 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
8716 		spa_avz_build(vd->vdev_child[i], avz, tx);
8717 	}
8718 }
8719 
8720 static void
8721 spa_sync_config_object(spa_t *spa, dmu_tx_t *tx)
8722 {
8723 	nvlist_t *config;
8724 
8725 	/*
8726 	 * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS,
8727 	 * its config may not be dirty but we still need to build per-vdev ZAPs.
8728 	 * Similarly, if the pool is being assembled (e.g. after a split), we
8729 	 * need to rebuild the AVZ although the config may not be dirty.
8730 	 */
8731 	if (list_is_empty(&spa->spa_config_dirty_list) &&
8732 	    spa->spa_avz_action == AVZ_ACTION_NONE)
8733 		return;
8734 
8735 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
8736 
8737 	ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE ||
8738 	    spa->spa_avz_action == AVZ_ACTION_INITIALIZE ||
8739 	    spa->spa_all_vdev_zaps != 0);
8740 
8741 	if (spa->spa_avz_action == AVZ_ACTION_REBUILD) {
8742 		/* Make and build the new AVZ */
8743 		uint64_t new_avz = zap_create(spa->spa_meta_objset,
8744 		    DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx);
8745 		spa_avz_build(spa->spa_root_vdev, new_avz, tx);
8746 
8747 		/* Diff old AVZ with new one */
8748 		zap_cursor_t zc;
8749 		zap_attribute_t za;
8750 
8751 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
8752 		    spa->spa_all_vdev_zaps);
8753 		    zap_cursor_retrieve(&zc, &za) == 0;
8754 		    zap_cursor_advance(&zc)) {
8755 			uint64_t vdzap = za.za_first_integer;
8756 			if (zap_lookup_int(spa->spa_meta_objset, new_avz,
8757 			    vdzap) == ENOENT) {
8758 				/*
8759 				 * ZAP is listed in old AVZ but not in new one;
8760 				 * destroy it
8761 				 */
8762 				VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap,
8763 				    tx));
8764 			}
8765 		}
8766 
8767 		zap_cursor_fini(&zc);
8768 
8769 		/* Destroy the old AVZ */
8770 		VERIFY0(zap_destroy(spa->spa_meta_objset,
8771 		    spa->spa_all_vdev_zaps, tx));
8772 
8773 		/* Replace the old AVZ in the dir obj with the new one */
8774 		VERIFY0(zap_update(spa->spa_meta_objset,
8775 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP,
8776 		    sizeof (new_avz), 1, &new_avz, tx));
8777 
8778 		spa->spa_all_vdev_zaps = new_avz;
8779 	} else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) {
8780 		zap_cursor_t zc;
8781 		zap_attribute_t za;
8782 
8783 		/* Walk through the AVZ and destroy all listed ZAPs */
8784 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
8785 		    spa->spa_all_vdev_zaps);
8786 		    zap_cursor_retrieve(&zc, &za) == 0;
8787 		    zap_cursor_advance(&zc)) {
8788 			uint64_t zap = za.za_first_integer;
8789 			VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx));
8790 		}
8791 
8792 		zap_cursor_fini(&zc);
8793 
8794 		/* Destroy and unlink the AVZ itself */
8795 		VERIFY0(zap_destroy(spa->spa_meta_objset,
8796 		    spa->spa_all_vdev_zaps, tx));
8797 		VERIFY0(zap_remove(spa->spa_meta_objset,
8798 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx));
8799 		spa->spa_all_vdev_zaps = 0;
8800 	}
8801 
8802 	if (spa->spa_all_vdev_zaps == 0) {
8803 		spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset,
8804 		    DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
8805 		    DMU_POOL_VDEV_ZAP_MAP, tx);
8806 	}
8807 	spa->spa_avz_action = AVZ_ACTION_NONE;
8808 
8809 	/* Create ZAPs for vdevs that don't have them. */
8810 	vdev_construct_zaps(spa->spa_root_vdev, tx);
8811 
8812 	config = spa_config_generate(spa, spa->spa_root_vdev,
8813 	    dmu_tx_get_txg(tx), B_FALSE);
8814 
8815 	/*
8816 	 * If we're upgrading the spa version then make sure that
8817 	 * the config object gets updated with the correct version.
8818 	 */
8819 	if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version)
8820 		fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION,
8821 		    spa->spa_uberblock.ub_version);
8822 
8823 	spa_config_exit(spa, SCL_STATE, FTAG);
8824 
8825 	nvlist_free(spa->spa_config_syncing);
8826 	spa->spa_config_syncing = config;
8827 
8828 	spa_sync_nvlist(spa, spa->spa_config_object, config, tx);
8829 }
8830 
8831 static void
8832 spa_sync_version(void *arg, dmu_tx_t *tx)
8833 {
8834 	uint64_t *versionp = arg;
8835 	uint64_t version = *versionp;
8836 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
8837 
8838 	/*
8839 	 * Setting the version is special cased when first creating the pool.
8840 	 */
8841 	ASSERT(tx->tx_txg != TXG_INITIAL);
8842 
8843 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
8844 	ASSERT(version >= spa_version(spa));
8845 
8846 	spa->spa_uberblock.ub_version = version;
8847 	vdev_config_dirty(spa->spa_root_vdev);
8848 	spa_history_log_internal(spa, "set", tx, "version=%lld",
8849 	    (longlong_t)version);
8850 }
8851 
8852 /*
8853  * Set zpool properties.
8854  */
8855 static void
8856 spa_sync_props(void *arg, dmu_tx_t *tx)
8857 {
8858 	nvlist_t *nvp = arg;
8859 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
8860 	objset_t *mos = spa->spa_meta_objset;
8861 	nvpair_t *elem = NULL;
8862 
8863 	mutex_enter(&spa->spa_props_lock);
8864 
8865 	while ((elem = nvlist_next_nvpair(nvp, elem))) {
8866 		uint64_t intval;
8867 		const char *strval, *fname;
8868 		zpool_prop_t prop;
8869 		const char *propname;
8870 		const char *elemname = nvpair_name(elem);
8871 		zprop_type_t proptype;
8872 		spa_feature_t fid;
8873 
8874 		switch (prop = zpool_name_to_prop(elemname)) {
8875 		case ZPOOL_PROP_VERSION:
8876 			intval = fnvpair_value_uint64(elem);
8877 			/*
8878 			 * The version is synced separately before other
8879 			 * properties and should be correct by now.
8880 			 */
8881 			ASSERT3U(spa_version(spa), >=, intval);
8882 			break;
8883 
8884 		case ZPOOL_PROP_ALTROOT:
8885 			/*
8886 			 * 'altroot' is a non-persistent property. It should
8887 			 * have been set temporarily at creation or import time.
8888 			 */
8889 			ASSERT(spa->spa_root != NULL);
8890 			break;
8891 
8892 		case ZPOOL_PROP_READONLY:
8893 		case ZPOOL_PROP_CACHEFILE:
8894 			/*
8895 			 * 'readonly' and 'cachefile' are also non-persistent
8896 			 * properties.
8897 			 */
8898 			break;
8899 		case ZPOOL_PROP_COMMENT:
8900 			strval = fnvpair_value_string(elem);
8901 			if (spa->spa_comment != NULL)
8902 				spa_strfree(spa->spa_comment);
8903 			spa->spa_comment = spa_strdup(strval);
8904 			/*
8905 			 * We need to dirty the configuration on all the vdevs
8906 			 * so that their labels get updated.  We also need to
8907 			 * update the cache file to keep it in sync with the
8908 			 * MOS version. It's unnecessary to do this for pool
8909 			 * creation since the vdev's configuration has already
8910 			 * been dirtied.
8911 			 */
8912 			if (tx->tx_txg != TXG_INITIAL) {
8913 				vdev_config_dirty(spa->spa_root_vdev);
8914 				spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
8915 			}
8916 			spa_history_log_internal(spa, "set", tx,
8917 			    "%s=%s", elemname, strval);
8918 			break;
8919 		case ZPOOL_PROP_COMPATIBILITY:
8920 			strval = fnvpair_value_string(elem);
8921 			if (spa->spa_compatibility != NULL)
8922 				spa_strfree(spa->spa_compatibility);
8923 			spa->spa_compatibility = spa_strdup(strval);
8924 			/*
8925 			 * Dirty the configuration on vdevs as above.
8926 			 */
8927 			if (tx->tx_txg != TXG_INITIAL) {
8928 				vdev_config_dirty(spa->spa_root_vdev);
8929 				spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
8930 			}
8931 
8932 			spa_history_log_internal(spa, "set", tx,
8933 			    "%s=%s", nvpair_name(elem), strval);
8934 			break;
8935 
8936 		case ZPOOL_PROP_INVAL:
8937 			if (zpool_prop_feature(elemname)) {
8938 				fname = strchr(elemname, '@') + 1;
8939 				VERIFY0(zfeature_lookup_name(fname, &fid));
8940 
8941 				spa_feature_enable(spa, fid, tx);
8942 				spa_history_log_internal(spa, "set", tx,
8943 				    "%s=enabled", elemname);
8944 				break;
8945 			} else if (!zfs_prop_user(elemname)) {
8946 				ASSERT(zpool_prop_feature(elemname));
8947 				break;
8948 			}
8949 			zfs_fallthrough;
8950 		default:
8951 			/*
8952 			 * Set pool property values in the poolprops mos object.
8953 			 */
8954 			if (spa->spa_pool_props_object == 0) {
8955 				spa->spa_pool_props_object =
8956 				    zap_create_link(mos, DMU_OT_POOL_PROPS,
8957 				    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS,
8958 				    tx);
8959 			}
8960 
8961 			/* normalize the property name */
8962 			if (prop == ZPOOL_PROP_INVAL) {
8963 				propname = elemname;
8964 				proptype = PROP_TYPE_STRING;
8965 			} else {
8966 				propname = zpool_prop_to_name(prop);
8967 				proptype = zpool_prop_get_type(prop);
8968 			}
8969 
8970 			if (nvpair_type(elem) == DATA_TYPE_STRING) {
8971 				ASSERT(proptype == PROP_TYPE_STRING);
8972 				strval = fnvpair_value_string(elem);
8973 				VERIFY0(zap_update(mos,
8974 				    spa->spa_pool_props_object, propname,
8975 				    1, strlen(strval) + 1, strval, tx));
8976 				spa_history_log_internal(spa, "set", tx,
8977 				    "%s=%s", elemname, strval);
8978 			} else if (nvpair_type(elem) == DATA_TYPE_UINT64) {
8979 				intval = fnvpair_value_uint64(elem);
8980 
8981 				if (proptype == PROP_TYPE_INDEX) {
8982 					const char *unused;
8983 					VERIFY0(zpool_prop_index_to_string(
8984 					    prop, intval, &unused));
8985 				}
8986 				VERIFY0(zap_update(mos,
8987 				    spa->spa_pool_props_object, propname,
8988 				    8, 1, &intval, tx));
8989 				spa_history_log_internal(spa, "set", tx,
8990 				    "%s=%lld", elemname,
8991 				    (longlong_t)intval);
8992 
8993 				switch (prop) {
8994 				case ZPOOL_PROP_DELEGATION:
8995 					spa->spa_delegation = intval;
8996 					break;
8997 				case ZPOOL_PROP_BOOTFS:
8998 					spa->spa_bootfs = intval;
8999 					break;
9000 				case ZPOOL_PROP_FAILUREMODE:
9001 					spa->spa_failmode = intval;
9002 					break;
9003 				case ZPOOL_PROP_AUTOTRIM:
9004 					spa->spa_autotrim = intval;
9005 					spa_async_request(spa,
9006 					    SPA_ASYNC_AUTOTRIM_RESTART);
9007 					break;
9008 				case ZPOOL_PROP_AUTOEXPAND:
9009 					spa->spa_autoexpand = intval;
9010 					if (tx->tx_txg != TXG_INITIAL)
9011 						spa_async_request(spa,
9012 						    SPA_ASYNC_AUTOEXPAND);
9013 					break;
9014 				case ZPOOL_PROP_MULTIHOST:
9015 					spa->spa_multihost = intval;
9016 					break;
9017 				default:
9018 					break;
9019 				}
9020 			} else {
9021 				ASSERT(0); /* not allowed */
9022 			}
9023 		}
9024 
9025 	}
9026 
9027 	mutex_exit(&spa->spa_props_lock);
9028 }
9029 
9030 /*
9031  * Perform one-time upgrade on-disk changes.  spa_version() does not
9032  * reflect the new version this txg, so there must be no changes this
9033  * txg to anything that the upgrade code depends on after it executes.
9034  * Therefore this must be called after dsl_pool_sync() does the sync
9035  * tasks.
9036  */
9037 static void
9038 spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx)
9039 {
9040 	if (spa_sync_pass(spa) != 1)
9041 		return;
9042 
9043 	dsl_pool_t *dp = spa->spa_dsl_pool;
9044 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
9045 
9046 	if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN &&
9047 	    spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) {
9048 		dsl_pool_create_origin(dp, tx);
9049 
9050 		/* Keeping the origin open increases spa_minref */
9051 		spa->spa_minref += 3;
9052 	}
9053 
9054 	if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES &&
9055 	    spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) {
9056 		dsl_pool_upgrade_clones(dp, tx);
9057 	}
9058 
9059 	if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES &&
9060 	    spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) {
9061 		dsl_pool_upgrade_dir_clones(dp, tx);
9062 
9063 		/* Keeping the freedir open increases spa_minref */
9064 		spa->spa_minref += 3;
9065 	}
9066 
9067 	if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES &&
9068 	    spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
9069 		spa_feature_create_zap_objects(spa, tx);
9070 	}
9071 
9072 	/*
9073 	 * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable
9074 	 * when possibility to use lz4 compression for metadata was added
9075 	 * Old pools that have this feature enabled must be upgraded to have
9076 	 * this feature active
9077 	 */
9078 	if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
9079 		boolean_t lz4_en = spa_feature_is_enabled(spa,
9080 		    SPA_FEATURE_LZ4_COMPRESS);
9081 		boolean_t lz4_ac = spa_feature_is_active(spa,
9082 		    SPA_FEATURE_LZ4_COMPRESS);
9083 
9084 		if (lz4_en && !lz4_ac)
9085 			spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx);
9086 	}
9087 
9088 	/*
9089 	 * If we haven't written the salt, do so now.  Note that the
9090 	 * feature may not be activated yet, but that's fine since
9091 	 * the presence of this ZAP entry is backwards compatible.
9092 	 */
9093 	if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
9094 	    DMU_POOL_CHECKSUM_SALT) == ENOENT) {
9095 		VERIFY0(zap_add(spa->spa_meta_objset,
9096 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1,
9097 		    sizeof (spa->spa_cksum_salt.zcs_bytes),
9098 		    spa->spa_cksum_salt.zcs_bytes, tx));
9099 	}
9100 
9101 	rrw_exit(&dp->dp_config_rwlock, FTAG);
9102 }
9103 
9104 static void
9105 vdev_indirect_state_sync_verify(vdev_t *vd)
9106 {
9107 	vdev_indirect_mapping_t *vim __maybe_unused = vd->vdev_indirect_mapping;
9108 	vdev_indirect_births_t *vib __maybe_unused = vd->vdev_indirect_births;
9109 
9110 	if (vd->vdev_ops == &vdev_indirect_ops) {
9111 		ASSERT(vim != NULL);
9112 		ASSERT(vib != NULL);
9113 	}
9114 
9115 	uint64_t obsolete_sm_object = 0;
9116 	ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object));
9117 	if (obsolete_sm_object != 0) {
9118 		ASSERT(vd->vdev_obsolete_sm != NULL);
9119 		ASSERT(vd->vdev_removing ||
9120 		    vd->vdev_ops == &vdev_indirect_ops);
9121 		ASSERT(vdev_indirect_mapping_num_entries(vim) > 0);
9122 		ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0);
9123 		ASSERT3U(obsolete_sm_object, ==,
9124 		    space_map_object(vd->vdev_obsolete_sm));
9125 		ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=,
9126 		    space_map_allocated(vd->vdev_obsolete_sm));
9127 	}
9128 	ASSERT(vd->vdev_obsolete_segments != NULL);
9129 
9130 	/*
9131 	 * Since frees / remaps to an indirect vdev can only
9132 	 * happen in syncing context, the obsolete segments
9133 	 * tree must be empty when we start syncing.
9134 	 */
9135 	ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
9136 }
9137 
9138 /*
9139  * Set the top-level vdev's max queue depth. Evaluate each top-level's
9140  * async write queue depth in case it changed. The max queue depth will
9141  * not change in the middle of syncing out this txg.
9142  */
9143 static void
9144 spa_sync_adjust_vdev_max_queue_depth(spa_t *spa)
9145 {
9146 	ASSERT(spa_writeable(spa));
9147 
9148 	vdev_t *rvd = spa->spa_root_vdev;
9149 	uint32_t max_queue_depth = zfs_vdev_async_write_max_active *
9150 	    zfs_vdev_queue_depth_pct / 100;
9151 	metaslab_class_t *normal = spa_normal_class(spa);
9152 	metaslab_class_t *special = spa_special_class(spa);
9153 	metaslab_class_t *dedup = spa_dedup_class(spa);
9154 
9155 	uint64_t slots_per_allocator = 0;
9156 	for (int c = 0; c < rvd->vdev_children; c++) {
9157 		vdev_t *tvd = rvd->vdev_child[c];
9158 
9159 		metaslab_group_t *mg = tvd->vdev_mg;
9160 		if (mg == NULL || !metaslab_group_initialized(mg))
9161 			continue;
9162 
9163 		metaslab_class_t *mc = mg->mg_class;
9164 		if (mc != normal && mc != special && mc != dedup)
9165 			continue;
9166 
9167 		/*
9168 		 * It is safe to do a lock-free check here because only async
9169 		 * allocations look at mg_max_alloc_queue_depth, and async
9170 		 * allocations all happen from spa_sync().
9171 		 */
9172 		for (int i = 0; i < mg->mg_allocators; i++) {
9173 			ASSERT0(zfs_refcount_count(
9174 			    &(mg->mg_allocator[i].mga_alloc_queue_depth)));
9175 		}
9176 		mg->mg_max_alloc_queue_depth = max_queue_depth;
9177 
9178 		for (int i = 0; i < mg->mg_allocators; i++) {
9179 			mg->mg_allocator[i].mga_cur_max_alloc_queue_depth =
9180 			    zfs_vdev_def_queue_depth;
9181 		}
9182 		slots_per_allocator += zfs_vdev_def_queue_depth;
9183 	}
9184 
9185 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9186 		ASSERT0(zfs_refcount_count(&normal->mc_allocator[i].
9187 		    mca_alloc_slots));
9188 		ASSERT0(zfs_refcount_count(&special->mc_allocator[i].
9189 		    mca_alloc_slots));
9190 		ASSERT0(zfs_refcount_count(&dedup->mc_allocator[i].
9191 		    mca_alloc_slots));
9192 		normal->mc_allocator[i].mca_alloc_max_slots =
9193 		    slots_per_allocator;
9194 		special->mc_allocator[i].mca_alloc_max_slots =
9195 		    slots_per_allocator;
9196 		dedup->mc_allocator[i].mca_alloc_max_slots =
9197 		    slots_per_allocator;
9198 	}
9199 	normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
9200 	special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
9201 	dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
9202 }
9203 
9204 static void
9205 spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx)
9206 {
9207 	ASSERT(spa_writeable(spa));
9208 
9209 	vdev_t *rvd = spa->spa_root_vdev;
9210 	for (int c = 0; c < rvd->vdev_children; c++) {
9211 		vdev_t *vd = rvd->vdev_child[c];
9212 		vdev_indirect_state_sync_verify(vd);
9213 
9214 		if (vdev_indirect_should_condense(vd)) {
9215 			spa_condense_indirect_start_sync(vd, tx);
9216 			break;
9217 		}
9218 	}
9219 }
9220 
9221 static void
9222 spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx)
9223 {
9224 	objset_t *mos = spa->spa_meta_objset;
9225 	dsl_pool_t *dp = spa->spa_dsl_pool;
9226 	uint64_t txg = tx->tx_txg;
9227 	bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK];
9228 
9229 	do {
9230 		int pass = ++spa->spa_sync_pass;
9231 
9232 		spa_sync_config_object(spa, tx);
9233 		spa_sync_aux_dev(spa, &spa->spa_spares, tx,
9234 		    ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES);
9235 		spa_sync_aux_dev(spa, &spa->spa_l2cache, tx,
9236 		    ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE);
9237 		spa_errlog_sync(spa, txg);
9238 		dsl_pool_sync(dp, txg);
9239 
9240 		if (pass < zfs_sync_pass_deferred_free ||
9241 		    spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) {
9242 			/*
9243 			 * If the log space map feature is active we don't
9244 			 * care about deferred frees and the deferred bpobj
9245 			 * as the log space map should effectively have the
9246 			 * same results (i.e. appending only to one object).
9247 			 */
9248 			spa_sync_frees(spa, free_bpl, tx);
9249 		} else {
9250 			/*
9251 			 * We can not defer frees in pass 1, because
9252 			 * we sync the deferred frees later in pass 1.
9253 			 */
9254 			ASSERT3U(pass, >, 1);
9255 			bplist_iterate(free_bpl, bpobj_enqueue_alloc_cb,
9256 			    &spa->spa_deferred_bpobj, tx);
9257 		}
9258 
9259 		brt_sync(spa, txg);
9260 		ddt_sync(spa, txg);
9261 		dsl_scan_sync(dp, tx);
9262 		dsl_errorscrub_sync(dp, tx);
9263 		svr_sync(spa, tx);
9264 		spa_sync_upgrades(spa, tx);
9265 
9266 		spa_flush_metaslabs(spa, tx);
9267 
9268 		vdev_t *vd = NULL;
9269 		while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg))
9270 		    != NULL)
9271 			vdev_sync(vd, txg);
9272 
9273 		/*
9274 		 * Note: We need to check if the MOS is dirty because we could
9275 		 * have marked the MOS dirty without updating the uberblock
9276 		 * (e.g. if we have sync tasks but no dirty user data). We need
9277 		 * to check the uberblock's rootbp because it is updated if we
9278 		 * have synced out dirty data (though in this case the MOS will
9279 		 * most likely also be dirty due to second order effects, we
9280 		 * don't want to rely on that here).
9281 		 */
9282 		if (pass == 1 &&
9283 		    spa->spa_uberblock.ub_rootbp.blk_birth < txg &&
9284 		    !dmu_objset_is_dirty(mos, txg)) {
9285 			/*
9286 			 * Nothing changed on the first pass, therefore this
9287 			 * TXG is a no-op. Avoid syncing deferred frees, so
9288 			 * that we can keep this TXG as a no-op.
9289 			 */
9290 			ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
9291 			ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
9292 			ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg));
9293 			ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg));
9294 			break;
9295 		}
9296 
9297 		spa_sync_deferred_frees(spa, tx);
9298 	} while (dmu_objset_is_dirty(mos, txg));
9299 }
9300 
9301 /*
9302  * Rewrite the vdev configuration (which includes the uberblock) to
9303  * commit the transaction group.
9304  *
9305  * If there are no dirty vdevs, we sync the uberblock to a few random
9306  * top-level vdevs that are known to be visible in the config cache
9307  * (see spa_vdev_add() for a complete description). If there *are* dirty
9308  * vdevs, sync the uberblock to all vdevs.
9309  */
9310 static void
9311 spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx)
9312 {
9313 	vdev_t *rvd = spa->spa_root_vdev;
9314 	uint64_t txg = tx->tx_txg;
9315 
9316 	for (;;) {
9317 		int error = 0;
9318 
9319 		/*
9320 		 * We hold SCL_STATE to prevent vdev open/close/etc.
9321 		 * while we're attempting to write the vdev labels.
9322 		 */
9323 		spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
9324 
9325 		if (list_is_empty(&spa->spa_config_dirty_list)) {
9326 			vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
9327 			int svdcount = 0;
9328 			int children = rvd->vdev_children;
9329 			int c0 = random_in_range(children);
9330 
9331 			for (int c = 0; c < children; c++) {
9332 				vdev_t *vd =
9333 				    rvd->vdev_child[(c0 + c) % children];
9334 
9335 				/* Stop when revisiting the first vdev */
9336 				if (c > 0 && svd[0] == vd)
9337 					break;
9338 
9339 				if (vd->vdev_ms_array == 0 ||
9340 				    vd->vdev_islog ||
9341 				    !vdev_is_concrete(vd))
9342 					continue;
9343 
9344 				svd[svdcount++] = vd;
9345 				if (svdcount == SPA_SYNC_MIN_VDEVS)
9346 					break;
9347 			}
9348 			error = vdev_config_sync(svd, svdcount, txg);
9349 		} else {
9350 			error = vdev_config_sync(rvd->vdev_child,
9351 			    rvd->vdev_children, txg);
9352 		}
9353 
9354 		if (error == 0)
9355 			spa->spa_last_synced_guid = rvd->vdev_guid;
9356 
9357 		spa_config_exit(spa, SCL_STATE, FTAG);
9358 
9359 		if (error == 0)
9360 			break;
9361 		zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR);
9362 		zio_resume_wait(spa);
9363 	}
9364 }
9365 
9366 /*
9367  * Sync the specified transaction group.  New blocks may be dirtied as
9368  * part of the process, so we iterate until it converges.
9369  */
9370 void
9371 spa_sync(spa_t *spa, uint64_t txg)
9372 {
9373 	vdev_t *vd = NULL;
9374 
9375 	VERIFY(spa_writeable(spa));
9376 
9377 	/*
9378 	 * Wait for i/os issued in open context that need to complete
9379 	 * before this txg syncs.
9380 	 */
9381 	(void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]);
9382 	spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL,
9383 	    ZIO_FLAG_CANFAIL);
9384 
9385 	/*
9386 	 * Now that there can be no more cloning in this transaction group,
9387 	 * but we are still before issuing frees, we can process pending BRT
9388 	 * updates.
9389 	 */
9390 	brt_pending_apply(spa, txg);
9391 
9392 	/*
9393 	 * Lock out configuration changes.
9394 	 */
9395 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
9396 
9397 	spa->spa_syncing_txg = txg;
9398 	spa->spa_sync_pass = 0;
9399 
9400 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9401 		mutex_enter(&spa->spa_allocs[i].spaa_lock);
9402 		VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree));
9403 		mutex_exit(&spa->spa_allocs[i].spaa_lock);
9404 	}
9405 
9406 	/*
9407 	 * If there are any pending vdev state changes, convert them
9408 	 * into config changes that go out with this transaction group.
9409 	 */
9410 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
9411 	while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) {
9412 		/* Avoid holding the write lock unless actually necessary */
9413 		if (vd->vdev_aux == NULL) {
9414 			vdev_state_clean(vd);
9415 			vdev_config_dirty(vd);
9416 			continue;
9417 		}
9418 		/*
9419 		 * We need the write lock here because, for aux vdevs,
9420 		 * calling vdev_config_dirty() modifies sav_config.
9421 		 * This is ugly and will become unnecessary when we
9422 		 * eliminate the aux vdev wart by integrating all vdevs
9423 		 * into the root vdev tree.
9424 		 */
9425 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9426 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER);
9427 		while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) {
9428 			vdev_state_clean(vd);
9429 			vdev_config_dirty(vd);
9430 		}
9431 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9432 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9433 	}
9434 	spa_config_exit(spa, SCL_STATE, FTAG);
9435 
9436 	dsl_pool_t *dp = spa->spa_dsl_pool;
9437 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
9438 
9439 	spa->spa_sync_starttime = gethrtime();
9440 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
9441 	spa->spa_deadman_tqid = taskq_dispatch_delay(system_delay_taskq,
9442 	    spa_deadman, spa, TQ_SLEEP, ddi_get_lbolt() +
9443 	    NSEC_TO_TICK(spa->spa_deadman_synctime));
9444 
9445 	/*
9446 	 * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg,
9447 	 * set spa_deflate if we have no raid-z vdevs.
9448 	 */
9449 	if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE &&
9450 	    spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) {
9451 		vdev_t *rvd = spa->spa_root_vdev;
9452 
9453 		int i;
9454 		for (i = 0; i < rvd->vdev_children; i++) {
9455 			vd = rvd->vdev_child[i];
9456 			if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE)
9457 				break;
9458 		}
9459 		if (i == rvd->vdev_children) {
9460 			spa->spa_deflate = TRUE;
9461 			VERIFY0(zap_add(spa->spa_meta_objset,
9462 			    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
9463 			    sizeof (uint64_t), 1, &spa->spa_deflate, tx));
9464 		}
9465 	}
9466 
9467 	spa_sync_adjust_vdev_max_queue_depth(spa);
9468 
9469 	spa_sync_condense_indirect(spa, tx);
9470 
9471 	spa_sync_iterate_to_convergence(spa, tx);
9472 
9473 #ifdef ZFS_DEBUG
9474 	if (!list_is_empty(&spa->spa_config_dirty_list)) {
9475 	/*
9476 	 * Make sure that the number of ZAPs for all the vdevs matches
9477 	 * the number of ZAPs in the per-vdev ZAP list. This only gets
9478 	 * called if the config is dirty; otherwise there may be
9479 	 * outstanding AVZ operations that weren't completed in
9480 	 * spa_sync_config_object.
9481 	 */
9482 		uint64_t all_vdev_zap_entry_count;
9483 		ASSERT0(zap_count(spa->spa_meta_objset,
9484 		    spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count));
9485 		ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==,
9486 		    all_vdev_zap_entry_count);
9487 	}
9488 #endif
9489 
9490 	if (spa->spa_vdev_removal != NULL) {
9491 		ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]);
9492 	}
9493 
9494 	spa_sync_rewrite_vdev_config(spa, tx);
9495 	dmu_tx_commit(tx);
9496 
9497 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
9498 	spa->spa_deadman_tqid = 0;
9499 
9500 	/*
9501 	 * Clear the dirty config list.
9502 	 */
9503 	while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL)
9504 		vdev_config_clean(vd);
9505 
9506 	/*
9507 	 * Now that the new config has synced transactionally,
9508 	 * let it become visible to the config cache.
9509 	 */
9510 	if (spa->spa_config_syncing != NULL) {
9511 		spa_config_set(spa, spa->spa_config_syncing);
9512 		spa->spa_config_txg = txg;
9513 		spa->spa_config_syncing = NULL;
9514 	}
9515 
9516 	dsl_pool_sync_done(dp, txg);
9517 
9518 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9519 		mutex_enter(&spa->spa_allocs[i].spaa_lock);
9520 		VERIFY0(avl_numnodes(&spa->spa_allocs[i].spaa_tree));
9521 		mutex_exit(&spa->spa_allocs[i].spaa_lock);
9522 	}
9523 
9524 	/*
9525 	 * Update usable space statistics.
9526 	 */
9527 	while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)))
9528 	    != NULL)
9529 		vdev_sync_done(vd, txg);
9530 
9531 	metaslab_class_evict_old(spa->spa_normal_class, txg);
9532 	metaslab_class_evict_old(spa->spa_log_class, txg);
9533 
9534 	spa_sync_close_syncing_log_sm(spa);
9535 
9536 	spa_update_dspace(spa);
9537 
9538 	if (spa_get_autotrim(spa) == SPA_AUTOTRIM_ON)
9539 		vdev_autotrim_kick(spa);
9540 
9541 	/*
9542 	 * It had better be the case that we didn't dirty anything
9543 	 * since vdev_config_sync().
9544 	 */
9545 	ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
9546 	ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
9547 	ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg));
9548 
9549 	while (zfs_pause_spa_sync)
9550 		delay(1);
9551 
9552 	spa->spa_sync_pass = 0;
9553 
9554 	/*
9555 	 * Update the last synced uberblock here. We want to do this at
9556 	 * the end of spa_sync() so that consumers of spa_last_synced_txg()
9557 	 * will be guaranteed that all the processing associated with
9558 	 * that txg has been completed.
9559 	 */
9560 	spa->spa_ubsync = spa->spa_uberblock;
9561 	spa_config_exit(spa, SCL_CONFIG, FTAG);
9562 
9563 	spa_handle_ignored_writes(spa);
9564 
9565 	/*
9566 	 * If any async tasks have been requested, kick them off.
9567 	 */
9568 	spa_async_dispatch(spa);
9569 }
9570 
9571 /*
9572  * Sync all pools.  We don't want to hold the namespace lock across these
9573  * operations, so we take a reference on the spa_t and drop the lock during the
9574  * sync.
9575  */
9576 void
9577 spa_sync_allpools(void)
9578 {
9579 	spa_t *spa = NULL;
9580 	mutex_enter(&spa_namespace_lock);
9581 	while ((spa = spa_next(spa)) != NULL) {
9582 		if (spa_state(spa) != POOL_STATE_ACTIVE ||
9583 		    !spa_writeable(spa) || spa_suspended(spa))
9584 			continue;
9585 		spa_open_ref(spa, FTAG);
9586 		mutex_exit(&spa_namespace_lock);
9587 		txg_wait_synced(spa_get_dsl(spa), 0);
9588 		mutex_enter(&spa_namespace_lock);
9589 		spa_close(spa, FTAG);
9590 	}
9591 	mutex_exit(&spa_namespace_lock);
9592 }
9593 
9594 /*
9595  * ==========================================================================
9596  * Miscellaneous routines
9597  * ==========================================================================
9598  */
9599 
9600 /*
9601  * Remove all pools in the system.
9602  */
9603 void
9604 spa_evict_all(void)
9605 {
9606 	spa_t *spa;
9607 
9608 	/*
9609 	 * Remove all cached state.  All pools should be closed now,
9610 	 * so every spa in the AVL tree should be unreferenced.
9611 	 */
9612 	mutex_enter(&spa_namespace_lock);
9613 	while ((spa = spa_next(NULL)) != NULL) {
9614 		/*
9615 		 * Stop async tasks.  The async thread may need to detach
9616 		 * a device that's been replaced, which requires grabbing
9617 		 * spa_namespace_lock, so we must drop it here.
9618 		 */
9619 		spa_open_ref(spa, FTAG);
9620 		mutex_exit(&spa_namespace_lock);
9621 		spa_async_suspend(spa);
9622 		mutex_enter(&spa_namespace_lock);
9623 		spa_close(spa, FTAG);
9624 
9625 		if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
9626 			spa_unload(spa);
9627 			spa_deactivate(spa);
9628 		}
9629 		spa_remove(spa);
9630 	}
9631 	mutex_exit(&spa_namespace_lock);
9632 }
9633 
9634 vdev_t *
9635 spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux)
9636 {
9637 	vdev_t *vd;
9638 	int i;
9639 
9640 	if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL)
9641 		return (vd);
9642 
9643 	if (aux) {
9644 		for (i = 0; i < spa->spa_l2cache.sav_count; i++) {
9645 			vd = spa->spa_l2cache.sav_vdevs[i];
9646 			if (vd->vdev_guid == guid)
9647 				return (vd);
9648 		}
9649 
9650 		for (i = 0; i < spa->spa_spares.sav_count; i++) {
9651 			vd = spa->spa_spares.sav_vdevs[i];
9652 			if (vd->vdev_guid == guid)
9653 				return (vd);
9654 		}
9655 	}
9656 
9657 	return (NULL);
9658 }
9659 
9660 void
9661 spa_upgrade(spa_t *spa, uint64_t version)
9662 {
9663 	ASSERT(spa_writeable(spa));
9664 
9665 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
9666 
9667 	/*
9668 	 * This should only be called for a non-faulted pool, and since a
9669 	 * future version would result in an unopenable pool, this shouldn't be
9670 	 * possible.
9671 	 */
9672 	ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version));
9673 	ASSERT3U(version, >=, spa->spa_uberblock.ub_version);
9674 
9675 	spa->spa_uberblock.ub_version = version;
9676 	vdev_config_dirty(spa->spa_root_vdev);
9677 
9678 	spa_config_exit(spa, SCL_ALL, FTAG);
9679 
9680 	txg_wait_synced(spa_get_dsl(spa), 0);
9681 }
9682 
9683 static boolean_t
9684 spa_has_aux_vdev(spa_t *spa, uint64_t guid, spa_aux_vdev_t *sav)
9685 {
9686 	(void) spa;
9687 	int i;
9688 	uint64_t vdev_guid;
9689 
9690 	for (i = 0; i < sav->sav_count; i++)
9691 		if (sav->sav_vdevs[i]->vdev_guid == guid)
9692 			return (B_TRUE);
9693 
9694 	for (i = 0; i < sav->sav_npending; i++) {
9695 		if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID,
9696 		    &vdev_guid) == 0 && vdev_guid == guid)
9697 			return (B_TRUE);
9698 	}
9699 
9700 	return (B_FALSE);
9701 }
9702 
9703 boolean_t
9704 spa_has_l2cache(spa_t *spa, uint64_t guid)
9705 {
9706 	return (spa_has_aux_vdev(spa, guid, &spa->spa_l2cache));
9707 }
9708 
9709 boolean_t
9710 spa_has_spare(spa_t *spa, uint64_t guid)
9711 {
9712 	return (spa_has_aux_vdev(spa, guid, &spa->spa_spares));
9713 }
9714 
9715 /*
9716  * Check if a pool has an active shared spare device.
9717  * Note: reference count of an active spare is 2, as a spare and as a replace
9718  */
9719 static boolean_t
9720 spa_has_active_shared_spare(spa_t *spa)
9721 {
9722 	int i, refcnt;
9723 	uint64_t pool;
9724 	spa_aux_vdev_t *sav = &spa->spa_spares;
9725 
9726 	for (i = 0; i < sav->sav_count; i++) {
9727 		if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool,
9728 		    &refcnt) && pool != 0ULL && pool == spa_guid(spa) &&
9729 		    refcnt > 2)
9730 			return (B_TRUE);
9731 	}
9732 
9733 	return (B_FALSE);
9734 }
9735 
9736 uint64_t
9737 spa_total_metaslabs(spa_t *spa)
9738 {
9739 	vdev_t *rvd = spa->spa_root_vdev;
9740 
9741 	uint64_t m = 0;
9742 	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
9743 		vdev_t *vd = rvd->vdev_child[c];
9744 		if (!vdev_is_concrete(vd))
9745 			continue;
9746 		m += vd->vdev_ms_count;
9747 	}
9748 	return (m);
9749 }
9750 
9751 /*
9752  * Notify any waiting threads that some activity has switched from being in-
9753  * progress to not-in-progress so that the thread can wake up and determine
9754  * whether it is finished waiting.
9755  */
9756 void
9757 spa_notify_waiters(spa_t *spa)
9758 {
9759 	/*
9760 	 * Acquiring spa_activities_lock here prevents the cv_broadcast from
9761 	 * happening between the waiting thread's check and cv_wait.
9762 	 */
9763 	mutex_enter(&spa->spa_activities_lock);
9764 	cv_broadcast(&spa->spa_activities_cv);
9765 	mutex_exit(&spa->spa_activities_lock);
9766 }
9767 
9768 /*
9769  * Notify any waiting threads that the pool is exporting, and then block until
9770  * they are finished using the spa_t.
9771  */
9772 void
9773 spa_wake_waiters(spa_t *spa)
9774 {
9775 	mutex_enter(&spa->spa_activities_lock);
9776 	spa->spa_waiters_cancel = B_TRUE;
9777 	cv_broadcast(&spa->spa_activities_cv);
9778 	while (spa->spa_waiters != 0)
9779 		cv_wait(&spa->spa_waiters_cv, &spa->spa_activities_lock);
9780 	spa->spa_waiters_cancel = B_FALSE;
9781 	mutex_exit(&spa->spa_activities_lock);
9782 }
9783 
9784 /* Whether the vdev or any of its descendants are being initialized/trimmed. */
9785 static boolean_t
9786 spa_vdev_activity_in_progress_impl(vdev_t *vd, zpool_wait_activity_t activity)
9787 {
9788 	spa_t *spa = vd->vdev_spa;
9789 
9790 	ASSERT(spa_config_held(spa, SCL_CONFIG | SCL_STATE, RW_READER));
9791 	ASSERT(MUTEX_HELD(&spa->spa_activities_lock));
9792 	ASSERT(activity == ZPOOL_WAIT_INITIALIZE ||
9793 	    activity == ZPOOL_WAIT_TRIM);
9794 
9795 	kmutex_t *lock = activity == ZPOOL_WAIT_INITIALIZE ?
9796 	    &vd->vdev_initialize_lock : &vd->vdev_trim_lock;
9797 
9798 	mutex_exit(&spa->spa_activities_lock);
9799 	mutex_enter(lock);
9800 	mutex_enter(&spa->spa_activities_lock);
9801 
9802 	boolean_t in_progress = (activity == ZPOOL_WAIT_INITIALIZE) ?
9803 	    (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) :
9804 	    (vd->vdev_trim_state == VDEV_TRIM_ACTIVE);
9805 	mutex_exit(lock);
9806 
9807 	if (in_progress)
9808 		return (B_TRUE);
9809 
9810 	for (int i = 0; i < vd->vdev_children; i++) {
9811 		if (spa_vdev_activity_in_progress_impl(vd->vdev_child[i],
9812 		    activity))
9813 			return (B_TRUE);
9814 	}
9815 
9816 	return (B_FALSE);
9817 }
9818 
9819 /*
9820  * If use_guid is true, this checks whether the vdev specified by guid is
9821  * being initialized/trimmed. Otherwise, it checks whether any vdev in the pool
9822  * is being initialized/trimmed. The caller must hold the config lock and
9823  * spa_activities_lock.
9824  */
9825 static int
9826 spa_vdev_activity_in_progress(spa_t *spa, boolean_t use_guid, uint64_t guid,
9827     zpool_wait_activity_t activity, boolean_t *in_progress)
9828 {
9829 	mutex_exit(&spa->spa_activities_lock);
9830 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9831 	mutex_enter(&spa->spa_activities_lock);
9832 
9833 	vdev_t *vd;
9834 	if (use_guid) {
9835 		vd = spa_lookup_by_guid(spa, guid, B_FALSE);
9836 		if (vd == NULL || !vd->vdev_ops->vdev_op_leaf) {
9837 			spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9838 			return (EINVAL);
9839 		}
9840 	} else {
9841 		vd = spa->spa_root_vdev;
9842 	}
9843 
9844 	*in_progress = spa_vdev_activity_in_progress_impl(vd, activity);
9845 
9846 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9847 	return (0);
9848 }
9849 
9850 /*
9851  * Locking for waiting threads
9852  * ---------------------------
9853  *
9854  * Waiting threads need a way to check whether a given activity is in progress,
9855  * and then, if it is, wait for it to complete. Each activity will have some
9856  * in-memory representation of the relevant on-disk state which can be used to
9857  * determine whether or not the activity is in progress. The in-memory state and
9858  * the locking used to protect it will be different for each activity, and may
9859  * not be suitable for use with a cvar (e.g., some state is protected by the
9860  * config lock). To allow waiting threads to wait without any races, another
9861  * lock, spa_activities_lock, is used.
9862  *
9863  * When the state is checked, both the activity-specific lock (if there is one)
9864  * and spa_activities_lock are held. In some cases, the activity-specific lock
9865  * is acquired explicitly (e.g. the config lock). In others, the locking is
9866  * internal to some check (e.g. bpobj_is_empty). After checking, the waiting
9867  * thread releases the activity-specific lock and, if the activity is in
9868  * progress, then cv_waits using spa_activities_lock.
9869  *
9870  * The waiting thread is woken when another thread, one completing some
9871  * activity, updates the state of the activity and then calls
9872  * spa_notify_waiters, which will cv_broadcast. This 'completing' thread only
9873  * needs to hold its activity-specific lock when updating the state, and this
9874  * lock can (but doesn't have to) be dropped before calling spa_notify_waiters.
9875  *
9876  * Because spa_notify_waiters acquires spa_activities_lock before broadcasting,
9877  * and because it is held when the waiting thread checks the state of the
9878  * activity, it can never be the case that the completing thread both updates
9879  * the activity state and cv_broadcasts in between the waiting thread's check
9880  * and cv_wait. Thus, a waiting thread can never miss a wakeup.
9881  *
9882  * In order to prevent deadlock, when the waiting thread does its check, in some
9883  * cases it will temporarily drop spa_activities_lock in order to acquire the
9884  * activity-specific lock. The order in which spa_activities_lock and the
9885  * activity specific lock are acquired in the waiting thread is determined by
9886  * the order in which they are acquired in the completing thread; if the
9887  * completing thread calls spa_notify_waiters with the activity-specific lock
9888  * held, then the waiting thread must also acquire the activity-specific lock
9889  * first.
9890  */
9891 
9892 static int
9893 spa_activity_in_progress(spa_t *spa, zpool_wait_activity_t activity,
9894     boolean_t use_tag, uint64_t tag, boolean_t *in_progress)
9895 {
9896 	int error = 0;
9897 
9898 	ASSERT(MUTEX_HELD(&spa->spa_activities_lock));
9899 
9900 	switch (activity) {
9901 	case ZPOOL_WAIT_CKPT_DISCARD:
9902 		*in_progress =
9903 		    (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT) &&
9904 		    zap_contains(spa_meta_objset(spa),
9905 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT) ==
9906 		    ENOENT);
9907 		break;
9908 	case ZPOOL_WAIT_FREE:
9909 		*in_progress = ((spa_version(spa) >= SPA_VERSION_DEADLISTS &&
9910 		    !bpobj_is_empty(&spa->spa_dsl_pool->dp_free_bpobj)) ||
9911 		    spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY) ||
9912 		    spa_livelist_delete_check(spa));
9913 		break;
9914 	case ZPOOL_WAIT_INITIALIZE:
9915 	case ZPOOL_WAIT_TRIM:
9916 		error = spa_vdev_activity_in_progress(spa, use_tag, tag,
9917 		    activity, in_progress);
9918 		break;
9919 	case ZPOOL_WAIT_REPLACE:
9920 		mutex_exit(&spa->spa_activities_lock);
9921 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9922 		mutex_enter(&spa->spa_activities_lock);
9923 
9924 		*in_progress = vdev_replace_in_progress(spa->spa_root_vdev);
9925 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9926 		break;
9927 	case ZPOOL_WAIT_REMOVE:
9928 		*in_progress = (spa->spa_removing_phys.sr_state ==
9929 		    DSS_SCANNING);
9930 		break;
9931 	case ZPOOL_WAIT_RESILVER:
9932 		if ((*in_progress = vdev_rebuild_active(spa->spa_root_vdev)))
9933 			break;
9934 		zfs_fallthrough;
9935 	case ZPOOL_WAIT_SCRUB:
9936 	{
9937 		boolean_t scanning, paused, is_scrub;
9938 		dsl_scan_t *scn =  spa->spa_dsl_pool->dp_scan;
9939 
9940 		is_scrub = (scn->scn_phys.scn_func == POOL_SCAN_SCRUB);
9941 		scanning = (scn->scn_phys.scn_state == DSS_SCANNING);
9942 		paused = dsl_scan_is_paused_scrub(scn);
9943 		*in_progress = (scanning && !paused &&
9944 		    is_scrub == (activity == ZPOOL_WAIT_SCRUB));
9945 		break;
9946 	}
9947 	default:
9948 		panic("unrecognized value for activity %d", activity);
9949 	}
9950 
9951 	return (error);
9952 }
9953 
9954 static int
9955 spa_wait_common(const char *pool, zpool_wait_activity_t activity,
9956     boolean_t use_tag, uint64_t tag, boolean_t *waited)
9957 {
9958 	/*
9959 	 * The tag is used to distinguish between instances of an activity.
9960 	 * 'initialize' and 'trim' are the only activities that we use this for.
9961 	 * The other activities can only have a single instance in progress in a
9962 	 * pool at one time, making the tag unnecessary.
9963 	 *
9964 	 * There can be multiple devices being replaced at once, but since they
9965 	 * all finish once resilvering finishes, we don't bother keeping track
9966 	 * of them individually, we just wait for them all to finish.
9967 	 */
9968 	if (use_tag && activity != ZPOOL_WAIT_INITIALIZE &&
9969 	    activity != ZPOOL_WAIT_TRIM)
9970 		return (EINVAL);
9971 
9972 	if (activity < 0 || activity >= ZPOOL_WAIT_NUM_ACTIVITIES)
9973 		return (EINVAL);
9974 
9975 	spa_t *spa;
9976 	int error = spa_open(pool, &spa, FTAG);
9977 	if (error != 0)
9978 		return (error);
9979 
9980 	/*
9981 	 * Increment the spa's waiter count so that we can call spa_close and
9982 	 * still ensure that the spa_t doesn't get freed before this thread is
9983 	 * finished with it when the pool is exported. We want to call spa_close
9984 	 * before we start waiting because otherwise the additional ref would
9985 	 * prevent the pool from being exported or destroyed throughout the
9986 	 * potentially long wait.
9987 	 */
9988 	mutex_enter(&spa->spa_activities_lock);
9989 	spa->spa_waiters++;
9990 	spa_close(spa, FTAG);
9991 
9992 	*waited = B_FALSE;
9993 	for (;;) {
9994 		boolean_t in_progress;
9995 		error = spa_activity_in_progress(spa, activity, use_tag, tag,
9996 		    &in_progress);
9997 
9998 		if (error || !in_progress || spa->spa_waiters_cancel)
9999 			break;
10000 
10001 		*waited = B_TRUE;
10002 
10003 		if (cv_wait_sig(&spa->spa_activities_cv,
10004 		    &spa->spa_activities_lock) == 0) {
10005 			error = EINTR;
10006 			break;
10007 		}
10008 	}
10009 
10010 	spa->spa_waiters--;
10011 	cv_signal(&spa->spa_waiters_cv);
10012 	mutex_exit(&spa->spa_activities_lock);
10013 
10014 	return (error);
10015 }
10016 
10017 /*
10018  * Wait for a particular instance of the specified activity to complete, where
10019  * the instance is identified by 'tag'
10020  */
10021 int
10022 spa_wait_tag(const char *pool, zpool_wait_activity_t activity, uint64_t tag,
10023     boolean_t *waited)
10024 {
10025 	return (spa_wait_common(pool, activity, B_TRUE, tag, waited));
10026 }
10027 
10028 /*
10029  * Wait for all instances of the specified activity complete
10030  */
10031 int
10032 spa_wait(const char *pool, zpool_wait_activity_t activity, boolean_t *waited)
10033 {
10034 
10035 	return (spa_wait_common(pool, activity, B_FALSE, 0, waited));
10036 }
10037 
10038 sysevent_t *
10039 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
10040 {
10041 	sysevent_t *ev = NULL;
10042 #ifdef _KERNEL
10043 	nvlist_t *resource;
10044 
10045 	resource = zfs_event_create(spa, vd, FM_SYSEVENT_CLASS, name, hist_nvl);
10046 	if (resource) {
10047 		ev = kmem_alloc(sizeof (sysevent_t), KM_SLEEP);
10048 		ev->resource = resource;
10049 	}
10050 #else
10051 	(void) spa, (void) vd, (void) hist_nvl, (void) name;
10052 #endif
10053 	return (ev);
10054 }
10055 
10056 void
10057 spa_event_post(sysevent_t *ev)
10058 {
10059 #ifdef _KERNEL
10060 	if (ev) {
10061 		zfs_zevent_post(ev->resource, NULL, zfs_zevent_post_cb);
10062 		kmem_free(ev, sizeof (*ev));
10063 	}
10064 #else
10065 	(void) ev;
10066 #endif
10067 }
10068 
10069 /*
10070  * Post a zevent corresponding to the given sysevent.   The 'name' must be one
10071  * of the event definitions in sys/sysevent/eventdefs.h.  The payload will be
10072  * filled in from the spa and (optionally) the vdev.  This doesn't do anything
10073  * in the userland libzpool, as we don't want consumers to misinterpret ztest
10074  * or zdb as real changes.
10075  */
10076 void
10077 spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
10078 {
10079 	spa_event_post(spa_event_create(spa, vd, hist_nvl, name));
10080 }
10081 
10082 /* state manipulation functions */
10083 EXPORT_SYMBOL(spa_open);
10084 EXPORT_SYMBOL(spa_open_rewind);
10085 EXPORT_SYMBOL(spa_get_stats);
10086 EXPORT_SYMBOL(spa_create);
10087 EXPORT_SYMBOL(spa_import);
10088 EXPORT_SYMBOL(spa_tryimport);
10089 EXPORT_SYMBOL(spa_destroy);
10090 EXPORT_SYMBOL(spa_export);
10091 EXPORT_SYMBOL(spa_reset);
10092 EXPORT_SYMBOL(spa_async_request);
10093 EXPORT_SYMBOL(spa_async_suspend);
10094 EXPORT_SYMBOL(spa_async_resume);
10095 EXPORT_SYMBOL(spa_inject_addref);
10096 EXPORT_SYMBOL(spa_inject_delref);
10097 EXPORT_SYMBOL(spa_scan_stat_init);
10098 EXPORT_SYMBOL(spa_scan_get_stats);
10099 
10100 /* device manipulation */
10101 EXPORT_SYMBOL(spa_vdev_add);
10102 EXPORT_SYMBOL(spa_vdev_attach);
10103 EXPORT_SYMBOL(spa_vdev_detach);
10104 EXPORT_SYMBOL(spa_vdev_setpath);
10105 EXPORT_SYMBOL(spa_vdev_setfru);
10106 EXPORT_SYMBOL(spa_vdev_split_mirror);
10107 
10108 /* spare statech is global across all pools) */
10109 EXPORT_SYMBOL(spa_spare_add);
10110 EXPORT_SYMBOL(spa_spare_remove);
10111 EXPORT_SYMBOL(spa_spare_exists);
10112 EXPORT_SYMBOL(spa_spare_activate);
10113 
10114 /* L2ARC statech is global across all pools) */
10115 EXPORT_SYMBOL(spa_l2cache_add);
10116 EXPORT_SYMBOL(spa_l2cache_remove);
10117 EXPORT_SYMBOL(spa_l2cache_exists);
10118 EXPORT_SYMBOL(spa_l2cache_activate);
10119 EXPORT_SYMBOL(spa_l2cache_drop);
10120 
10121 /* scanning */
10122 EXPORT_SYMBOL(spa_scan);
10123 EXPORT_SYMBOL(spa_scan_stop);
10124 
10125 /* spa syncing */
10126 EXPORT_SYMBOL(spa_sync); /* only for DMU use */
10127 EXPORT_SYMBOL(spa_sync_allpools);
10128 
10129 /* properties */
10130 EXPORT_SYMBOL(spa_prop_set);
10131 EXPORT_SYMBOL(spa_prop_get);
10132 EXPORT_SYMBOL(spa_prop_clear_bootfs);
10133 
10134 /* asynchronous event notification */
10135 EXPORT_SYMBOL(spa_event_notify);
10136 
10137 /* BEGIN CSTYLED */
10138 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_shift, UINT, ZMOD_RW,
10139 	"log2 fraction of arc that can be used by inflight I/Os when "
10140 	"verifying pool during import");
10141 /* END CSTYLED */
10142 
10143 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_metadata, INT, ZMOD_RW,
10144 	"Set to traverse metadata on pool import");
10145 
10146 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW,
10147 	"Set to traverse data on pool import");
10148 
10149 ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW,
10150 	"Print vdev tree to zfs_dbgmsg during pool import");
10151 
10152 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD,
10153 	"Percentage of CPUs to run an IO worker thread");
10154 
10155 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_tpq, UINT, ZMOD_RD,
10156 	"Number of threads per IO worker taskqueue");
10157 
10158 /* BEGIN CSTYLED */
10159 ZFS_MODULE_PARAM(zfs, zfs_, max_missing_tvds, U64, ZMOD_RW,
10160 	"Allow importing pool with up to this number of missing top-level "
10161 	"vdevs (in read-only mode)");
10162 /* END CSTYLED */
10163 
10164 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_pause, INT,
10165 	ZMOD_RW, "Set the livelist condense zthr to pause");
10166 
10167 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_pause, INT,
10168 	ZMOD_RW, "Set the livelist condense synctask to pause");
10169 
10170 /* BEGIN CSTYLED */
10171 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_cancel,
10172 	INT, ZMOD_RW,
10173 	"Whether livelist condensing was canceled in the synctask");
10174 
10175 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_cancel,
10176 	INT, ZMOD_RW,
10177 	"Whether livelist condensing was canceled in the zthr function");
10178 
10179 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT,
10180 	ZMOD_RW,
10181 	"Whether extra ALLOC blkptrs were added to a livelist entry while it "
10182 	"was being condensed");
10183 /* END CSTYLED */
10184