xref: /freebsd-src/sys/contrib/openzfs/module/zfs/spa.c (revision 2aa3ef285a23d802f0bd6c7281612e16834e9b68)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Copyright (c) 2011, 2020 by Delphix. All rights reserved.
25  * Copyright (c) 2018, Nexenta Systems, Inc.  All rights reserved.
26  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
27  * Copyright 2013 Saso Kiselkov. All rights reserved.
28  * Copyright (c) 2014 Integros [integros.com]
29  * Copyright 2016 Toomas Soome <tsoome@me.com>
30  * Copyright (c) 2016 Actifio, Inc. All rights reserved.
31  * Copyright 2018 Joyent, Inc.
32  * Copyright (c) 2017, 2019, Datto Inc. All rights reserved.
33  * Copyright 2017 Joyent, Inc.
34  * Copyright (c) 2017, Intel Corporation.
35  */
36 
37 /*
38  * SPA: Storage Pool Allocator
39  *
40  * This file contains all the routines used when modifying on-disk SPA state.
41  * This includes opening, importing, destroying, exporting a pool, and syncing a
42  * pool.
43  */
44 
45 #include <sys/zfs_context.h>
46 #include <sys/fm/fs/zfs.h>
47 #include <sys/spa_impl.h>
48 #include <sys/zio.h>
49 #include <sys/zio_checksum.h>
50 #include <sys/dmu.h>
51 #include <sys/dmu_tx.h>
52 #include <sys/zap.h>
53 #include <sys/zil.h>
54 #include <sys/ddt.h>
55 #include <sys/vdev_impl.h>
56 #include <sys/vdev_removal.h>
57 #include <sys/vdev_indirect_mapping.h>
58 #include <sys/vdev_indirect_births.h>
59 #include <sys/vdev_initialize.h>
60 #include <sys/vdev_rebuild.h>
61 #include <sys/vdev_trim.h>
62 #include <sys/vdev_disk.h>
63 #include <sys/vdev_draid.h>
64 #include <sys/metaslab.h>
65 #include <sys/metaslab_impl.h>
66 #include <sys/mmp.h>
67 #include <sys/uberblock_impl.h>
68 #include <sys/txg.h>
69 #include <sys/avl.h>
70 #include <sys/bpobj.h>
71 #include <sys/dmu_traverse.h>
72 #include <sys/dmu_objset.h>
73 #include <sys/unique.h>
74 #include <sys/dsl_pool.h>
75 #include <sys/dsl_dataset.h>
76 #include <sys/dsl_dir.h>
77 #include <sys/dsl_prop.h>
78 #include <sys/dsl_synctask.h>
79 #include <sys/fs/zfs.h>
80 #include <sys/arc.h>
81 #include <sys/callb.h>
82 #include <sys/systeminfo.h>
83 #include <sys/spa_boot.h>
84 #include <sys/zfs_ioctl.h>
85 #include <sys/dsl_scan.h>
86 #include <sys/zfeature.h>
87 #include <sys/dsl_destroy.h>
88 #include <sys/zvol.h>
89 
90 #ifdef	_KERNEL
91 #include <sys/fm/protocol.h>
92 #include <sys/fm/util.h>
93 #include <sys/callb.h>
94 #include <sys/zone.h>
95 #include <sys/vmsystm.h>
96 #endif	/* _KERNEL */
97 
98 #include "zfs_prop.h"
99 #include "zfs_comutil.h"
100 
101 /*
102  * The interval, in seconds, at which failed configuration cache file writes
103  * should be retried.
104  */
105 int zfs_ccw_retry_interval = 300;
106 
107 typedef enum zti_modes {
108 	ZTI_MODE_FIXED,			/* value is # of threads (min 1) */
109 	ZTI_MODE_BATCH,			/* cpu-intensive; value is ignored */
110 	ZTI_MODE_NULL,			/* don't create a taskq */
111 	ZTI_NMODES
112 } zti_modes_t;
113 
114 #define	ZTI_P(n, q)	{ ZTI_MODE_FIXED, (n), (q) }
115 #define	ZTI_PCT(n)	{ ZTI_MODE_ONLINE_PERCENT, (n), 1 }
116 #define	ZTI_BATCH	{ ZTI_MODE_BATCH, 0, 1 }
117 #define	ZTI_NULL	{ ZTI_MODE_NULL, 0, 0 }
118 
119 #define	ZTI_N(n)	ZTI_P(n, 1)
120 #define	ZTI_ONE		ZTI_N(1)
121 
122 typedef struct zio_taskq_info {
123 	zti_modes_t zti_mode;
124 	uint_t zti_value;
125 	uint_t zti_count;
126 } zio_taskq_info_t;
127 
128 static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = {
129 	"iss", "iss_h", "int", "int_h"
130 };
131 
132 /*
133  * This table defines the taskq settings for each ZFS I/O type. When
134  * initializing a pool, we use this table to create an appropriately sized
135  * taskq. Some operations are low volume and therefore have a small, static
136  * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE
137  * macros. Other operations process a large amount of data; the ZTI_BATCH
138  * macro causes us to create a taskq oriented for throughput. Some operations
139  * are so high frequency and short-lived that the taskq itself can become a
140  * point of lock contention. The ZTI_P(#, #) macro indicates that we need an
141  * additional degree of parallelism specified by the number of threads per-
142  * taskq and the number of taskqs; when dispatching an event in this case, the
143  * particular taskq is chosen at random.
144  *
145  * The different taskq priorities are to handle the different contexts (issue
146  * and interrupt) and then to reserve threads for ZIO_PRIORITY_NOW I/Os that
147  * need to be handled with minimum delay.
148  */
149 const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = {
150 	/* ISSUE	ISSUE_HIGH	INTR		INTR_HIGH */
151 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* NULL */
152 	{ ZTI_N(8),	ZTI_NULL,	ZTI_P(12, 8),	ZTI_NULL }, /* READ */
153 	{ ZTI_BATCH,	ZTI_N(5),	ZTI_P(12, 8),	ZTI_N(5) }, /* WRITE */
154 	{ ZTI_P(12, 8),	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* FREE */
155 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* CLAIM */
156 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* IOCTL */
157 	{ ZTI_N(4),	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* TRIM */
158 };
159 
160 static void spa_sync_version(void *arg, dmu_tx_t *tx);
161 static void spa_sync_props(void *arg, dmu_tx_t *tx);
162 static boolean_t spa_has_active_shared_spare(spa_t *spa);
163 static int spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport);
164 static void spa_vdev_resilver_done(spa_t *spa);
165 
166 uint_t		zio_taskq_batch_pct = 75;	/* 1 thread per cpu in pset */
167 boolean_t	zio_taskq_sysdc = B_TRUE;	/* use SDC scheduling class */
168 uint_t		zio_taskq_basedc = 80;		/* base duty cycle */
169 
170 boolean_t	spa_create_process = B_TRUE;	/* no process ==> no sysdc */
171 
172 /*
173  * Report any spa_load_verify errors found, but do not fail spa_load.
174  * This is used by zdb to analyze non-idle pools.
175  */
176 boolean_t	spa_load_verify_dryrun = B_FALSE;
177 
178 /*
179  * This (illegal) pool name is used when temporarily importing a spa_t in order
180  * to get the vdev stats associated with the imported devices.
181  */
182 #define	TRYIMPORT_NAME	"$import"
183 
184 /*
185  * For debugging purposes: print out vdev tree during pool import.
186  */
187 int		spa_load_print_vdev_tree = B_FALSE;
188 
189 /*
190  * A non-zero value for zfs_max_missing_tvds means that we allow importing
191  * pools with missing top-level vdevs. This is strictly intended for advanced
192  * pool recovery cases since missing data is almost inevitable. Pools with
193  * missing devices can only be imported read-only for safety reasons, and their
194  * fail-mode will be automatically set to "continue".
195  *
196  * With 1 missing vdev we should be able to import the pool and mount all
197  * datasets. User data that was not modified after the missing device has been
198  * added should be recoverable. This means that snapshots created prior to the
199  * addition of that device should be completely intact.
200  *
201  * With 2 missing vdevs, some datasets may fail to mount since there are
202  * dataset statistics that are stored as regular metadata. Some data might be
203  * recoverable if those vdevs were added recently.
204  *
205  * With 3 or more missing vdevs, the pool is severely damaged and MOS entries
206  * may be missing entirely. Chances of data recovery are very low. Note that
207  * there are also risks of performing an inadvertent rewind as we might be
208  * missing all the vdevs with the latest uberblocks.
209  */
210 unsigned long	zfs_max_missing_tvds = 0;
211 
212 /*
213  * The parameters below are similar to zfs_max_missing_tvds but are only
214  * intended for a preliminary open of the pool with an untrusted config which
215  * might be incomplete or out-dated.
216  *
217  * We are more tolerant for pools opened from a cachefile since we could have
218  * an out-dated cachefile where a device removal was not registered.
219  * We could have set the limit arbitrarily high but in the case where devices
220  * are really missing we would want to return the proper error codes; we chose
221  * SPA_DVAS_PER_BP - 1 so that some copies of the MOS would still be available
222  * and we get a chance to retrieve the trusted config.
223  */
224 uint64_t	zfs_max_missing_tvds_cachefile = SPA_DVAS_PER_BP - 1;
225 
226 /*
227  * In the case where config was assembled by scanning device paths (/dev/dsks
228  * by default) we are less tolerant since all the existing devices should have
229  * been detected and we want spa_load to return the right error codes.
230  */
231 uint64_t	zfs_max_missing_tvds_scan = 0;
232 
233 /*
234  * Debugging aid that pauses spa_sync() towards the end.
235  */
236 boolean_t	zfs_pause_spa_sync = B_FALSE;
237 
238 /*
239  * Variables to indicate the livelist condense zthr func should wait at certain
240  * points for the livelist to be removed - used to test condense/destroy races
241  */
242 int zfs_livelist_condense_zthr_pause = 0;
243 int zfs_livelist_condense_sync_pause = 0;
244 
245 /*
246  * Variables to track whether or not condense cancellation has been
247  * triggered in testing.
248  */
249 int zfs_livelist_condense_sync_cancel = 0;
250 int zfs_livelist_condense_zthr_cancel = 0;
251 
252 /*
253  * Variable to track whether or not extra ALLOC blkptrs were added to a
254  * livelist entry while it was being condensed (caused by the way we track
255  * remapped blkptrs in dbuf_remap_impl)
256  */
257 int zfs_livelist_condense_new_alloc = 0;
258 
259 /*
260  * ==========================================================================
261  * SPA properties routines
262  * ==========================================================================
263  */
264 
265 /*
266  * Add a (source=src, propname=propval) list to an nvlist.
267  */
268 static void
269 spa_prop_add_list(nvlist_t *nvl, zpool_prop_t prop, char *strval,
270     uint64_t intval, zprop_source_t src)
271 {
272 	const char *propname = zpool_prop_to_name(prop);
273 	nvlist_t *propval;
274 
275 	VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
276 	VERIFY(nvlist_add_uint64(propval, ZPROP_SOURCE, src) == 0);
277 
278 	if (strval != NULL)
279 		VERIFY(nvlist_add_string(propval, ZPROP_VALUE, strval) == 0);
280 	else
281 		VERIFY(nvlist_add_uint64(propval, ZPROP_VALUE, intval) == 0);
282 
283 	VERIFY(nvlist_add_nvlist(nvl, propname, propval) == 0);
284 	nvlist_free(propval);
285 }
286 
287 /*
288  * Get property values from the spa configuration.
289  */
290 static void
291 spa_prop_get_config(spa_t *spa, nvlist_t **nvp)
292 {
293 	vdev_t *rvd = spa->spa_root_vdev;
294 	dsl_pool_t *pool = spa->spa_dsl_pool;
295 	uint64_t size, alloc, cap, version;
296 	const zprop_source_t src = ZPROP_SRC_NONE;
297 	spa_config_dirent_t *dp;
298 	metaslab_class_t *mc = spa_normal_class(spa);
299 
300 	ASSERT(MUTEX_HELD(&spa->spa_props_lock));
301 
302 	if (rvd != NULL) {
303 		alloc = metaslab_class_get_alloc(mc);
304 		alloc += metaslab_class_get_alloc(spa_special_class(spa));
305 		alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
306 		alloc += metaslab_class_get_alloc(spa_embedded_log_class(spa));
307 
308 		size = metaslab_class_get_space(mc);
309 		size += metaslab_class_get_space(spa_special_class(spa));
310 		size += metaslab_class_get_space(spa_dedup_class(spa));
311 		size += metaslab_class_get_space(spa_embedded_log_class(spa));
312 
313 		spa_prop_add_list(*nvp, ZPOOL_PROP_NAME, spa_name(spa), 0, src);
314 		spa_prop_add_list(*nvp, ZPOOL_PROP_SIZE, NULL, size, src);
315 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALLOCATED, NULL, alloc, src);
316 		spa_prop_add_list(*nvp, ZPOOL_PROP_FREE, NULL,
317 		    size - alloc, src);
318 		spa_prop_add_list(*nvp, ZPOOL_PROP_CHECKPOINT, NULL,
319 		    spa->spa_checkpoint_info.sci_dspace, src);
320 
321 		spa_prop_add_list(*nvp, ZPOOL_PROP_FRAGMENTATION, NULL,
322 		    metaslab_class_fragmentation(mc), src);
323 		spa_prop_add_list(*nvp, ZPOOL_PROP_EXPANDSZ, NULL,
324 		    metaslab_class_expandable_space(mc), src);
325 		spa_prop_add_list(*nvp, ZPOOL_PROP_READONLY, NULL,
326 		    (spa_mode(spa) == SPA_MODE_READ), src);
327 
328 		cap = (size == 0) ? 0 : (alloc * 100 / size);
329 		spa_prop_add_list(*nvp, ZPOOL_PROP_CAPACITY, NULL, cap, src);
330 
331 		spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL,
332 		    ddt_get_pool_dedup_ratio(spa), src);
333 
334 		spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL,
335 		    rvd->vdev_state, src);
336 
337 		version = spa_version(spa);
338 		if (version == zpool_prop_default_numeric(ZPOOL_PROP_VERSION)) {
339 			spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL,
340 			    version, ZPROP_SRC_DEFAULT);
341 		} else {
342 			spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL,
343 			    version, ZPROP_SRC_LOCAL);
344 		}
345 		spa_prop_add_list(*nvp, ZPOOL_PROP_LOAD_GUID,
346 		    NULL, spa_load_guid(spa), src);
347 	}
348 
349 	if (pool != NULL) {
350 		/*
351 		 * The $FREE directory was introduced in SPA_VERSION_DEADLISTS,
352 		 * when opening pools before this version freedir will be NULL.
353 		 */
354 		if (pool->dp_free_dir != NULL) {
355 			spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL,
356 			    dsl_dir_phys(pool->dp_free_dir)->dd_used_bytes,
357 			    src);
358 		} else {
359 			spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING,
360 			    NULL, 0, src);
361 		}
362 
363 		if (pool->dp_leak_dir != NULL) {
364 			spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL,
365 			    dsl_dir_phys(pool->dp_leak_dir)->dd_used_bytes,
366 			    src);
367 		} else {
368 			spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED,
369 			    NULL, 0, src);
370 		}
371 	}
372 
373 	spa_prop_add_list(*nvp, ZPOOL_PROP_GUID, NULL, spa_guid(spa), src);
374 
375 	if (spa->spa_comment != NULL) {
376 		spa_prop_add_list(*nvp, ZPOOL_PROP_COMMENT, spa->spa_comment,
377 		    0, ZPROP_SRC_LOCAL);
378 	}
379 
380 	if (spa->spa_root != NULL)
381 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root,
382 		    0, ZPROP_SRC_LOCAL);
383 
384 	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) {
385 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
386 		    MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE);
387 	} else {
388 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
389 		    SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE);
390 	}
391 
392 	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE)) {
393 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL,
394 		    DNODE_MAX_SIZE, ZPROP_SRC_NONE);
395 	} else {
396 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL,
397 		    DNODE_MIN_SIZE, ZPROP_SRC_NONE);
398 	}
399 
400 	if ((dp = list_head(&spa->spa_config_list)) != NULL) {
401 		if (dp->scd_path == NULL) {
402 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
403 			    "none", 0, ZPROP_SRC_LOCAL);
404 		} else if (strcmp(dp->scd_path, spa_config_path) != 0) {
405 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
406 			    dp->scd_path, 0, ZPROP_SRC_LOCAL);
407 		}
408 	}
409 }
410 
411 /*
412  * Get zpool property values.
413  */
414 int
415 spa_prop_get(spa_t *spa, nvlist_t **nvp)
416 {
417 	objset_t *mos = spa->spa_meta_objset;
418 	zap_cursor_t zc;
419 	zap_attribute_t za;
420 	dsl_pool_t *dp;
421 	int err;
422 
423 	err = nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP);
424 	if (err)
425 		return (err);
426 
427 	dp = spa_get_dsl(spa);
428 	dsl_pool_config_enter(dp, FTAG);
429 	mutex_enter(&spa->spa_props_lock);
430 
431 	/*
432 	 * Get properties from the spa config.
433 	 */
434 	spa_prop_get_config(spa, nvp);
435 
436 	/* If no pool property object, no more prop to get. */
437 	if (mos == NULL || spa->spa_pool_props_object == 0)
438 		goto out;
439 
440 	/*
441 	 * Get properties from the MOS pool property object.
442 	 */
443 	for (zap_cursor_init(&zc, mos, spa->spa_pool_props_object);
444 	    (err = zap_cursor_retrieve(&zc, &za)) == 0;
445 	    zap_cursor_advance(&zc)) {
446 		uint64_t intval = 0;
447 		char *strval = NULL;
448 		zprop_source_t src = ZPROP_SRC_DEFAULT;
449 		zpool_prop_t prop;
450 
451 		if ((prop = zpool_name_to_prop(za.za_name)) == ZPOOL_PROP_INVAL)
452 			continue;
453 
454 		switch (za.za_integer_length) {
455 		case 8:
456 			/* integer property */
457 			if (za.za_first_integer !=
458 			    zpool_prop_default_numeric(prop))
459 				src = ZPROP_SRC_LOCAL;
460 
461 			if (prop == ZPOOL_PROP_BOOTFS) {
462 				dsl_dataset_t *ds = NULL;
463 
464 				err = dsl_dataset_hold_obj(dp,
465 				    za.za_first_integer, FTAG, &ds);
466 				if (err != 0)
467 					break;
468 
469 				strval = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN,
470 				    KM_SLEEP);
471 				dsl_dataset_name(ds, strval);
472 				dsl_dataset_rele(ds, FTAG);
473 			} else {
474 				strval = NULL;
475 				intval = za.za_first_integer;
476 			}
477 
478 			spa_prop_add_list(*nvp, prop, strval, intval, src);
479 
480 			if (strval != NULL)
481 				kmem_free(strval, ZFS_MAX_DATASET_NAME_LEN);
482 
483 			break;
484 
485 		case 1:
486 			/* string property */
487 			strval = kmem_alloc(za.za_num_integers, KM_SLEEP);
488 			err = zap_lookup(mos, spa->spa_pool_props_object,
489 			    za.za_name, 1, za.za_num_integers, strval);
490 			if (err) {
491 				kmem_free(strval, za.za_num_integers);
492 				break;
493 			}
494 			spa_prop_add_list(*nvp, prop, strval, 0, src);
495 			kmem_free(strval, za.za_num_integers);
496 			break;
497 
498 		default:
499 			break;
500 		}
501 	}
502 	zap_cursor_fini(&zc);
503 out:
504 	mutex_exit(&spa->spa_props_lock);
505 	dsl_pool_config_exit(dp, FTAG);
506 	if (err && err != ENOENT) {
507 		nvlist_free(*nvp);
508 		*nvp = NULL;
509 		return (err);
510 	}
511 
512 	return (0);
513 }
514 
515 /*
516  * Validate the given pool properties nvlist and modify the list
517  * for the property values to be set.
518  */
519 static int
520 spa_prop_validate(spa_t *spa, nvlist_t *props)
521 {
522 	nvpair_t *elem;
523 	int error = 0, reset_bootfs = 0;
524 	uint64_t objnum = 0;
525 	boolean_t has_feature = B_FALSE;
526 
527 	elem = NULL;
528 	while ((elem = nvlist_next_nvpair(props, elem)) != NULL) {
529 		uint64_t intval;
530 		char *strval, *slash, *check, *fname;
531 		const char *propname = nvpair_name(elem);
532 		zpool_prop_t prop = zpool_name_to_prop(propname);
533 
534 		switch (prop) {
535 		case ZPOOL_PROP_INVAL:
536 			if (!zpool_prop_feature(propname)) {
537 				error = SET_ERROR(EINVAL);
538 				break;
539 			}
540 
541 			/*
542 			 * Sanitize the input.
543 			 */
544 			if (nvpair_type(elem) != DATA_TYPE_UINT64) {
545 				error = SET_ERROR(EINVAL);
546 				break;
547 			}
548 
549 			if (nvpair_value_uint64(elem, &intval) != 0) {
550 				error = SET_ERROR(EINVAL);
551 				break;
552 			}
553 
554 			if (intval != 0) {
555 				error = SET_ERROR(EINVAL);
556 				break;
557 			}
558 
559 			fname = strchr(propname, '@') + 1;
560 			if (zfeature_lookup_name(fname, NULL) != 0) {
561 				error = SET_ERROR(EINVAL);
562 				break;
563 			}
564 
565 			has_feature = B_TRUE;
566 			break;
567 
568 		case ZPOOL_PROP_VERSION:
569 			error = nvpair_value_uint64(elem, &intval);
570 			if (!error &&
571 			    (intval < spa_version(spa) ||
572 			    intval > SPA_VERSION_BEFORE_FEATURES ||
573 			    has_feature))
574 				error = SET_ERROR(EINVAL);
575 			break;
576 
577 		case ZPOOL_PROP_DELEGATION:
578 		case ZPOOL_PROP_AUTOREPLACE:
579 		case ZPOOL_PROP_LISTSNAPS:
580 		case ZPOOL_PROP_AUTOEXPAND:
581 		case ZPOOL_PROP_AUTOTRIM:
582 			error = nvpair_value_uint64(elem, &intval);
583 			if (!error && intval > 1)
584 				error = SET_ERROR(EINVAL);
585 			break;
586 
587 		case ZPOOL_PROP_MULTIHOST:
588 			error = nvpair_value_uint64(elem, &intval);
589 			if (!error && intval > 1)
590 				error = SET_ERROR(EINVAL);
591 
592 			if (!error) {
593 				uint32_t hostid = zone_get_hostid(NULL);
594 				if (hostid)
595 					spa->spa_hostid = hostid;
596 				else
597 					error = SET_ERROR(ENOTSUP);
598 			}
599 
600 			break;
601 
602 		case ZPOOL_PROP_BOOTFS:
603 			/*
604 			 * If the pool version is less than SPA_VERSION_BOOTFS,
605 			 * or the pool is still being created (version == 0),
606 			 * the bootfs property cannot be set.
607 			 */
608 			if (spa_version(spa) < SPA_VERSION_BOOTFS) {
609 				error = SET_ERROR(ENOTSUP);
610 				break;
611 			}
612 
613 			/*
614 			 * Make sure the vdev config is bootable
615 			 */
616 			if (!vdev_is_bootable(spa->spa_root_vdev)) {
617 				error = SET_ERROR(ENOTSUP);
618 				break;
619 			}
620 
621 			reset_bootfs = 1;
622 
623 			error = nvpair_value_string(elem, &strval);
624 
625 			if (!error) {
626 				objset_t *os;
627 
628 				if (strval == NULL || strval[0] == '\0') {
629 					objnum = zpool_prop_default_numeric(
630 					    ZPOOL_PROP_BOOTFS);
631 					break;
632 				}
633 
634 				error = dmu_objset_hold(strval, FTAG, &os);
635 				if (error != 0)
636 					break;
637 
638 				/* Must be ZPL. */
639 				if (dmu_objset_type(os) != DMU_OST_ZFS) {
640 					error = SET_ERROR(ENOTSUP);
641 				} else {
642 					objnum = dmu_objset_id(os);
643 				}
644 				dmu_objset_rele(os, FTAG);
645 			}
646 			break;
647 
648 		case ZPOOL_PROP_FAILUREMODE:
649 			error = nvpair_value_uint64(elem, &intval);
650 			if (!error && intval > ZIO_FAILURE_MODE_PANIC)
651 				error = SET_ERROR(EINVAL);
652 
653 			/*
654 			 * This is a special case which only occurs when
655 			 * the pool has completely failed. This allows
656 			 * the user to change the in-core failmode property
657 			 * without syncing it out to disk (I/Os might
658 			 * currently be blocked). We do this by returning
659 			 * EIO to the caller (spa_prop_set) to trick it
660 			 * into thinking we encountered a property validation
661 			 * error.
662 			 */
663 			if (!error && spa_suspended(spa)) {
664 				spa->spa_failmode = intval;
665 				error = SET_ERROR(EIO);
666 			}
667 			break;
668 
669 		case ZPOOL_PROP_CACHEFILE:
670 			if ((error = nvpair_value_string(elem, &strval)) != 0)
671 				break;
672 
673 			if (strval[0] == '\0')
674 				break;
675 
676 			if (strcmp(strval, "none") == 0)
677 				break;
678 
679 			if (strval[0] != '/') {
680 				error = SET_ERROR(EINVAL);
681 				break;
682 			}
683 
684 			slash = strrchr(strval, '/');
685 			ASSERT(slash != NULL);
686 
687 			if (slash[1] == '\0' || strcmp(slash, "/.") == 0 ||
688 			    strcmp(slash, "/..") == 0)
689 				error = SET_ERROR(EINVAL);
690 			break;
691 
692 		case ZPOOL_PROP_COMMENT:
693 			if ((error = nvpair_value_string(elem, &strval)) != 0)
694 				break;
695 			for (check = strval; *check != '\0'; check++) {
696 				if (!isprint(*check)) {
697 					error = SET_ERROR(EINVAL);
698 					break;
699 				}
700 			}
701 			if (strlen(strval) > ZPROP_MAX_COMMENT)
702 				error = SET_ERROR(E2BIG);
703 			break;
704 
705 		default:
706 			break;
707 		}
708 
709 		if (error)
710 			break;
711 	}
712 
713 	(void) nvlist_remove_all(props,
714 	    zpool_prop_to_name(ZPOOL_PROP_DEDUPDITTO));
715 
716 	if (!error && reset_bootfs) {
717 		error = nvlist_remove(props,
718 		    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), DATA_TYPE_STRING);
719 
720 		if (!error) {
721 			error = nvlist_add_uint64(props,
722 			    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), objnum);
723 		}
724 	}
725 
726 	return (error);
727 }
728 
729 void
730 spa_configfile_set(spa_t *spa, nvlist_t *nvp, boolean_t need_sync)
731 {
732 	char *cachefile;
733 	spa_config_dirent_t *dp;
734 
735 	if (nvlist_lookup_string(nvp, zpool_prop_to_name(ZPOOL_PROP_CACHEFILE),
736 	    &cachefile) != 0)
737 		return;
738 
739 	dp = kmem_alloc(sizeof (spa_config_dirent_t),
740 	    KM_SLEEP);
741 
742 	if (cachefile[0] == '\0')
743 		dp->scd_path = spa_strdup(spa_config_path);
744 	else if (strcmp(cachefile, "none") == 0)
745 		dp->scd_path = NULL;
746 	else
747 		dp->scd_path = spa_strdup(cachefile);
748 
749 	list_insert_head(&spa->spa_config_list, dp);
750 	if (need_sync)
751 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
752 }
753 
754 int
755 spa_prop_set(spa_t *spa, nvlist_t *nvp)
756 {
757 	int error;
758 	nvpair_t *elem = NULL;
759 	boolean_t need_sync = B_FALSE;
760 
761 	if ((error = spa_prop_validate(spa, nvp)) != 0)
762 		return (error);
763 
764 	while ((elem = nvlist_next_nvpair(nvp, elem)) != NULL) {
765 		zpool_prop_t prop = zpool_name_to_prop(nvpair_name(elem));
766 
767 		if (prop == ZPOOL_PROP_CACHEFILE ||
768 		    prop == ZPOOL_PROP_ALTROOT ||
769 		    prop == ZPOOL_PROP_READONLY)
770 			continue;
771 
772 		if (prop == ZPOOL_PROP_VERSION || prop == ZPOOL_PROP_INVAL) {
773 			uint64_t ver;
774 
775 			if (prop == ZPOOL_PROP_VERSION) {
776 				VERIFY(nvpair_value_uint64(elem, &ver) == 0);
777 			} else {
778 				ASSERT(zpool_prop_feature(nvpair_name(elem)));
779 				ver = SPA_VERSION_FEATURES;
780 				need_sync = B_TRUE;
781 			}
782 
783 			/* Save time if the version is already set. */
784 			if (ver == spa_version(spa))
785 				continue;
786 
787 			/*
788 			 * In addition to the pool directory object, we might
789 			 * create the pool properties object, the features for
790 			 * read object, the features for write object, or the
791 			 * feature descriptions object.
792 			 */
793 			error = dsl_sync_task(spa->spa_name, NULL,
794 			    spa_sync_version, &ver,
795 			    6, ZFS_SPACE_CHECK_RESERVED);
796 			if (error)
797 				return (error);
798 			continue;
799 		}
800 
801 		need_sync = B_TRUE;
802 		break;
803 	}
804 
805 	if (need_sync) {
806 		return (dsl_sync_task(spa->spa_name, NULL, spa_sync_props,
807 		    nvp, 6, ZFS_SPACE_CHECK_RESERVED));
808 	}
809 
810 	return (0);
811 }
812 
813 /*
814  * If the bootfs property value is dsobj, clear it.
815  */
816 void
817 spa_prop_clear_bootfs(spa_t *spa, uint64_t dsobj, dmu_tx_t *tx)
818 {
819 	if (spa->spa_bootfs == dsobj && spa->spa_pool_props_object != 0) {
820 		VERIFY(zap_remove(spa->spa_meta_objset,
821 		    spa->spa_pool_props_object,
822 		    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), tx) == 0);
823 		spa->spa_bootfs = 0;
824 	}
825 }
826 
827 /*ARGSUSED*/
828 static int
829 spa_change_guid_check(void *arg, dmu_tx_t *tx)
830 {
831 	uint64_t *newguid __maybe_unused = arg;
832 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
833 	vdev_t *rvd = spa->spa_root_vdev;
834 	uint64_t vdev_state;
835 
836 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
837 		int error = (spa_has_checkpoint(spa)) ?
838 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
839 		return (SET_ERROR(error));
840 	}
841 
842 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
843 	vdev_state = rvd->vdev_state;
844 	spa_config_exit(spa, SCL_STATE, FTAG);
845 
846 	if (vdev_state != VDEV_STATE_HEALTHY)
847 		return (SET_ERROR(ENXIO));
848 
849 	ASSERT3U(spa_guid(spa), !=, *newguid);
850 
851 	return (0);
852 }
853 
854 static void
855 spa_change_guid_sync(void *arg, dmu_tx_t *tx)
856 {
857 	uint64_t *newguid = arg;
858 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
859 	uint64_t oldguid;
860 	vdev_t *rvd = spa->spa_root_vdev;
861 
862 	oldguid = spa_guid(spa);
863 
864 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
865 	rvd->vdev_guid = *newguid;
866 	rvd->vdev_guid_sum += (*newguid - oldguid);
867 	vdev_config_dirty(rvd);
868 	spa_config_exit(spa, SCL_STATE, FTAG);
869 
870 	spa_history_log_internal(spa, "guid change", tx, "old=%llu new=%llu",
871 	    (u_longlong_t)oldguid, (u_longlong_t)*newguid);
872 }
873 
874 /*
875  * Change the GUID for the pool.  This is done so that we can later
876  * re-import a pool built from a clone of our own vdevs.  We will modify
877  * the root vdev's guid, our own pool guid, and then mark all of our
878  * vdevs dirty.  Note that we must make sure that all our vdevs are
879  * online when we do this, or else any vdevs that weren't present
880  * would be orphaned from our pool.  We are also going to issue a
881  * sysevent to update any watchers.
882  */
883 int
884 spa_change_guid(spa_t *spa)
885 {
886 	int error;
887 	uint64_t guid;
888 
889 	mutex_enter(&spa->spa_vdev_top_lock);
890 	mutex_enter(&spa_namespace_lock);
891 	guid = spa_generate_guid(NULL);
892 
893 	error = dsl_sync_task(spa->spa_name, spa_change_guid_check,
894 	    spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED);
895 
896 	if (error == 0) {
897 		spa_write_cachefile(spa, B_FALSE, B_TRUE);
898 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID);
899 	}
900 
901 	mutex_exit(&spa_namespace_lock);
902 	mutex_exit(&spa->spa_vdev_top_lock);
903 
904 	return (error);
905 }
906 
907 /*
908  * ==========================================================================
909  * SPA state manipulation (open/create/destroy/import/export)
910  * ==========================================================================
911  */
912 
913 static int
914 spa_error_entry_compare(const void *a, const void *b)
915 {
916 	const spa_error_entry_t *sa = (const spa_error_entry_t *)a;
917 	const spa_error_entry_t *sb = (const spa_error_entry_t *)b;
918 	int ret;
919 
920 	ret = memcmp(&sa->se_bookmark, &sb->se_bookmark,
921 	    sizeof (zbookmark_phys_t));
922 
923 	return (TREE_ISIGN(ret));
924 }
925 
926 /*
927  * Utility function which retrieves copies of the current logs and
928  * re-initializes them in the process.
929  */
930 void
931 spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub)
932 {
933 	ASSERT(MUTEX_HELD(&spa->spa_errlist_lock));
934 
935 	bcopy(&spa->spa_errlist_last, last, sizeof (avl_tree_t));
936 	bcopy(&spa->spa_errlist_scrub, scrub, sizeof (avl_tree_t));
937 
938 	avl_create(&spa->spa_errlist_scrub,
939 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
940 	    offsetof(spa_error_entry_t, se_avl));
941 	avl_create(&spa->spa_errlist_last,
942 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
943 	    offsetof(spa_error_entry_t, se_avl));
944 }
945 
946 static void
947 spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
948 {
949 	const zio_taskq_info_t *ztip = &zio_taskqs[t][q];
950 	enum zti_modes mode = ztip->zti_mode;
951 	uint_t value = ztip->zti_value;
952 	uint_t count = ztip->zti_count;
953 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
954 	uint_t flags = 0;
955 	boolean_t batch = B_FALSE;
956 
957 	if (mode == ZTI_MODE_NULL) {
958 		tqs->stqs_count = 0;
959 		tqs->stqs_taskq = NULL;
960 		return;
961 	}
962 
963 	ASSERT3U(count, >, 0);
964 
965 	tqs->stqs_count = count;
966 	tqs->stqs_taskq = kmem_alloc(count * sizeof (taskq_t *), KM_SLEEP);
967 
968 	switch (mode) {
969 	case ZTI_MODE_FIXED:
970 		ASSERT3U(value, >=, 1);
971 		value = MAX(value, 1);
972 		flags |= TASKQ_DYNAMIC;
973 		break;
974 
975 	case ZTI_MODE_BATCH:
976 		batch = B_TRUE;
977 		flags |= TASKQ_THREADS_CPU_PCT;
978 		value = MIN(zio_taskq_batch_pct, 100);
979 		break;
980 
981 	default:
982 		panic("unrecognized mode for %s_%s taskq (%u:%u) in "
983 		    "spa_activate()",
984 		    zio_type_name[t], zio_taskq_types[q], mode, value);
985 		break;
986 	}
987 
988 	for (uint_t i = 0; i < count; i++) {
989 		taskq_t *tq;
990 		char name[32];
991 
992 		(void) snprintf(name, sizeof (name), "%s_%s",
993 		    zio_type_name[t], zio_taskq_types[q]);
994 
995 		if (zio_taskq_sysdc && spa->spa_proc != &p0) {
996 			if (batch)
997 				flags |= TASKQ_DC_BATCH;
998 
999 			tq = taskq_create_sysdc(name, value, 50, INT_MAX,
1000 			    spa->spa_proc, zio_taskq_basedc, flags);
1001 		} else {
1002 			pri_t pri = maxclsyspri;
1003 			/*
1004 			 * The write issue taskq can be extremely CPU
1005 			 * intensive.  Run it at slightly less important
1006 			 * priority than the other taskqs.
1007 			 *
1008 			 * Under Linux and FreeBSD this means incrementing
1009 			 * the priority value as opposed to platforms like
1010 			 * illumos where it should be decremented.
1011 			 *
1012 			 * On FreeBSD, if priorities divided by four (RQ_PPQ)
1013 			 * are equal then a difference between them is
1014 			 * insignificant.
1015 			 */
1016 			if (t == ZIO_TYPE_WRITE && q == ZIO_TASKQ_ISSUE) {
1017 #if defined(__linux__)
1018 				pri++;
1019 #elif defined(__FreeBSD__)
1020 				pri += 4;
1021 #else
1022 #error "unknown OS"
1023 #endif
1024 			}
1025 			tq = taskq_create_proc(name, value, pri, 50,
1026 			    INT_MAX, spa->spa_proc, flags);
1027 		}
1028 
1029 		tqs->stqs_taskq[i] = tq;
1030 	}
1031 }
1032 
1033 static void
1034 spa_taskqs_fini(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
1035 {
1036 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1037 
1038 	if (tqs->stqs_taskq == NULL) {
1039 		ASSERT3U(tqs->stqs_count, ==, 0);
1040 		return;
1041 	}
1042 
1043 	for (uint_t i = 0; i < tqs->stqs_count; i++) {
1044 		ASSERT3P(tqs->stqs_taskq[i], !=, NULL);
1045 		taskq_destroy(tqs->stqs_taskq[i]);
1046 	}
1047 
1048 	kmem_free(tqs->stqs_taskq, tqs->stqs_count * sizeof (taskq_t *));
1049 	tqs->stqs_taskq = NULL;
1050 }
1051 
1052 /*
1053  * Dispatch a task to the appropriate taskq for the ZFS I/O type and priority.
1054  * Note that a type may have multiple discrete taskqs to avoid lock contention
1055  * on the taskq itself. In that case we choose which taskq at random by using
1056  * the low bits of gethrtime().
1057  */
1058 void
1059 spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
1060     task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent)
1061 {
1062 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1063 	taskq_t *tq;
1064 
1065 	ASSERT3P(tqs->stqs_taskq, !=, NULL);
1066 	ASSERT3U(tqs->stqs_count, !=, 0);
1067 
1068 	if (tqs->stqs_count == 1) {
1069 		tq = tqs->stqs_taskq[0];
1070 	} else {
1071 		tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count];
1072 	}
1073 
1074 	taskq_dispatch_ent(tq, func, arg, flags, ent);
1075 }
1076 
1077 /*
1078  * Same as spa_taskq_dispatch_ent() but block on the task until completion.
1079  */
1080 void
1081 spa_taskq_dispatch_sync(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
1082     task_func_t *func, void *arg, uint_t flags)
1083 {
1084 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1085 	taskq_t *tq;
1086 	taskqid_t id;
1087 
1088 	ASSERT3P(tqs->stqs_taskq, !=, NULL);
1089 	ASSERT3U(tqs->stqs_count, !=, 0);
1090 
1091 	if (tqs->stqs_count == 1) {
1092 		tq = tqs->stqs_taskq[0];
1093 	} else {
1094 		tq = tqs->stqs_taskq[((uint64_t)gethrtime()) % tqs->stqs_count];
1095 	}
1096 
1097 	id = taskq_dispatch(tq, func, arg, flags);
1098 	if (id)
1099 		taskq_wait_id(tq, id);
1100 }
1101 
1102 static void
1103 spa_create_zio_taskqs(spa_t *spa)
1104 {
1105 	for (int t = 0; t < ZIO_TYPES; t++) {
1106 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1107 			spa_taskqs_init(spa, t, q);
1108 		}
1109 	}
1110 }
1111 
1112 /*
1113  * Disabled until spa_thread() can be adapted for Linux.
1114  */
1115 #undef HAVE_SPA_THREAD
1116 
1117 #if defined(_KERNEL) && defined(HAVE_SPA_THREAD)
1118 static void
1119 spa_thread(void *arg)
1120 {
1121 	psetid_t zio_taskq_psrset_bind = PS_NONE;
1122 	callb_cpr_t cprinfo;
1123 
1124 	spa_t *spa = arg;
1125 	user_t *pu = PTOU(curproc);
1126 
1127 	CALLB_CPR_INIT(&cprinfo, &spa->spa_proc_lock, callb_generic_cpr,
1128 	    spa->spa_name);
1129 
1130 	ASSERT(curproc != &p0);
1131 	(void) snprintf(pu->u_psargs, sizeof (pu->u_psargs),
1132 	    "zpool-%s", spa->spa_name);
1133 	(void) strlcpy(pu->u_comm, pu->u_psargs, sizeof (pu->u_comm));
1134 
1135 	/* bind this thread to the requested psrset */
1136 	if (zio_taskq_psrset_bind != PS_NONE) {
1137 		pool_lock();
1138 		mutex_enter(&cpu_lock);
1139 		mutex_enter(&pidlock);
1140 		mutex_enter(&curproc->p_lock);
1141 
1142 		if (cpupart_bind_thread(curthread, zio_taskq_psrset_bind,
1143 		    0, NULL, NULL) == 0)  {
1144 			curthread->t_bind_pset = zio_taskq_psrset_bind;
1145 		} else {
1146 			cmn_err(CE_WARN,
1147 			    "Couldn't bind process for zfs pool \"%s\" to "
1148 			    "pset %d\n", spa->spa_name, zio_taskq_psrset_bind);
1149 		}
1150 
1151 		mutex_exit(&curproc->p_lock);
1152 		mutex_exit(&pidlock);
1153 		mutex_exit(&cpu_lock);
1154 		pool_unlock();
1155 	}
1156 
1157 	if (zio_taskq_sysdc) {
1158 		sysdc_thread_enter(curthread, 100, 0);
1159 	}
1160 
1161 	spa->spa_proc = curproc;
1162 	spa->spa_did = curthread->t_did;
1163 
1164 	spa_create_zio_taskqs(spa);
1165 
1166 	mutex_enter(&spa->spa_proc_lock);
1167 	ASSERT(spa->spa_proc_state == SPA_PROC_CREATED);
1168 
1169 	spa->spa_proc_state = SPA_PROC_ACTIVE;
1170 	cv_broadcast(&spa->spa_proc_cv);
1171 
1172 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
1173 	while (spa->spa_proc_state == SPA_PROC_ACTIVE)
1174 		cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1175 	CALLB_CPR_SAFE_END(&cprinfo, &spa->spa_proc_lock);
1176 
1177 	ASSERT(spa->spa_proc_state == SPA_PROC_DEACTIVATE);
1178 	spa->spa_proc_state = SPA_PROC_GONE;
1179 	spa->spa_proc = &p0;
1180 	cv_broadcast(&spa->spa_proc_cv);
1181 	CALLB_CPR_EXIT(&cprinfo);	/* drops spa_proc_lock */
1182 
1183 	mutex_enter(&curproc->p_lock);
1184 	lwp_exit();
1185 }
1186 #endif
1187 
1188 /*
1189  * Activate an uninitialized pool.
1190  */
1191 static void
1192 spa_activate(spa_t *spa, spa_mode_t mode)
1193 {
1194 	ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED);
1195 
1196 	spa->spa_state = POOL_STATE_ACTIVE;
1197 	spa->spa_mode = mode;
1198 
1199 	spa->spa_normal_class = metaslab_class_create(spa, zfs_metaslab_ops);
1200 	spa->spa_log_class = metaslab_class_create(spa, zfs_metaslab_ops);
1201 	spa->spa_embedded_log_class =
1202 	    metaslab_class_create(spa, zfs_metaslab_ops);
1203 	spa->spa_special_class = metaslab_class_create(spa, zfs_metaslab_ops);
1204 	spa->spa_dedup_class = metaslab_class_create(spa, zfs_metaslab_ops);
1205 
1206 	/* Try to create a covering process */
1207 	mutex_enter(&spa->spa_proc_lock);
1208 	ASSERT(spa->spa_proc_state == SPA_PROC_NONE);
1209 	ASSERT(spa->spa_proc == &p0);
1210 	spa->spa_did = 0;
1211 
1212 #ifdef HAVE_SPA_THREAD
1213 	/* Only create a process if we're going to be around a while. */
1214 	if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) {
1215 		if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri,
1216 		    NULL, 0) == 0) {
1217 			spa->spa_proc_state = SPA_PROC_CREATED;
1218 			while (spa->spa_proc_state == SPA_PROC_CREATED) {
1219 				cv_wait(&spa->spa_proc_cv,
1220 				    &spa->spa_proc_lock);
1221 			}
1222 			ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1223 			ASSERT(spa->spa_proc != &p0);
1224 			ASSERT(spa->spa_did != 0);
1225 		} else {
1226 #ifdef _KERNEL
1227 			cmn_err(CE_WARN,
1228 			    "Couldn't create process for zfs pool \"%s\"\n",
1229 			    spa->spa_name);
1230 #endif
1231 		}
1232 	}
1233 #endif /* HAVE_SPA_THREAD */
1234 	mutex_exit(&spa->spa_proc_lock);
1235 
1236 	/* If we didn't create a process, we need to create our taskqs. */
1237 	if (spa->spa_proc == &p0) {
1238 		spa_create_zio_taskqs(spa);
1239 	}
1240 
1241 	for (size_t i = 0; i < TXG_SIZE; i++) {
1242 		spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL,
1243 		    ZIO_FLAG_CANFAIL);
1244 	}
1245 
1246 	list_create(&spa->spa_config_dirty_list, sizeof (vdev_t),
1247 	    offsetof(vdev_t, vdev_config_dirty_node));
1248 	list_create(&spa->spa_evicting_os_list, sizeof (objset_t),
1249 	    offsetof(objset_t, os_evicting_node));
1250 	list_create(&spa->spa_state_dirty_list, sizeof (vdev_t),
1251 	    offsetof(vdev_t, vdev_state_dirty_node));
1252 
1253 	txg_list_create(&spa->spa_vdev_txg_list, spa,
1254 	    offsetof(struct vdev, vdev_txg_node));
1255 
1256 	avl_create(&spa->spa_errlist_scrub,
1257 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1258 	    offsetof(spa_error_entry_t, se_avl));
1259 	avl_create(&spa->spa_errlist_last,
1260 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1261 	    offsetof(spa_error_entry_t, se_avl));
1262 
1263 	spa_keystore_init(&spa->spa_keystore);
1264 
1265 	/*
1266 	 * This taskq is used to perform zvol-minor-related tasks
1267 	 * asynchronously. This has several advantages, including easy
1268 	 * resolution of various deadlocks.
1269 	 *
1270 	 * The taskq must be single threaded to ensure tasks are always
1271 	 * processed in the order in which they were dispatched.
1272 	 *
1273 	 * A taskq per pool allows one to keep the pools independent.
1274 	 * This way if one pool is suspended, it will not impact another.
1275 	 *
1276 	 * The preferred location to dispatch a zvol minor task is a sync
1277 	 * task. In this context, there is easy access to the spa_t and minimal
1278 	 * error handling is required because the sync task must succeed.
1279 	 */
1280 	spa->spa_zvol_taskq = taskq_create("z_zvol", 1, defclsyspri,
1281 	    1, INT_MAX, 0);
1282 
1283 	/*
1284 	 * Taskq dedicated to prefetcher threads: this is used to prevent the
1285 	 * pool traverse code from monopolizing the global (and limited)
1286 	 * system_taskq by inappropriately scheduling long running tasks on it.
1287 	 */
1288 	spa->spa_prefetch_taskq = taskq_create("z_prefetch", 100,
1289 	    defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT);
1290 
1291 	/*
1292 	 * The taskq to upgrade datasets in this pool. Currently used by
1293 	 * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA.
1294 	 */
1295 	spa->spa_upgrade_taskq = taskq_create("z_upgrade", 100,
1296 	    defclsyspri, 1, INT_MAX, TASKQ_DYNAMIC | TASKQ_THREADS_CPU_PCT);
1297 }
1298 
1299 /*
1300  * Opposite of spa_activate().
1301  */
1302 static void
1303 spa_deactivate(spa_t *spa)
1304 {
1305 	ASSERT(spa->spa_sync_on == B_FALSE);
1306 	ASSERT(spa->spa_dsl_pool == NULL);
1307 	ASSERT(spa->spa_root_vdev == NULL);
1308 	ASSERT(spa->spa_async_zio_root == NULL);
1309 	ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED);
1310 
1311 	spa_evicting_os_wait(spa);
1312 
1313 	if (spa->spa_zvol_taskq) {
1314 		taskq_destroy(spa->spa_zvol_taskq);
1315 		spa->spa_zvol_taskq = NULL;
1316 	}
1317 
1318 	if (spa->spa_prefetch_taskq) {
1319 		taskq_destroy(spa->spa_prefetch_taskq);
1320 		spa->spa_prefetch_taskq = NULL;
1321 	}
1322 
1323 	if (spa->spa_upgrade_taskq) {
1324 		taskq_destroy(spa->spa_upgrade_taskq);
1325 		spa->spa_upgrade_taskq = NULL;
1326 	}
1327 
1328 	txg_list_destroy(&spa->spa_vdev_txg_list);
1329 
1330 	list_destroy(&spa->spa_config_dirty_list);
1331 	list_destroy(&spa->spa_evicting_os_list);
1332 	list_destroy(&spa->spa_state_dirty_list);
1333 
1334 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
1335 
1336 	for (int t = 0; t < ZIO_TYPES; t++) {
1337 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1338 			spa_taskqs_fini(spa, t, q);
1339 		}
1340 	}
1341 
1342 	for (size_t i = 0; i < TXG_SIZE; i++) {
1343 		ASSERT3P(spa->spa_txg_zio[i], !=, NULL);
1344 		VERIFY0(zio_wait(spa->spa_txg_zio[i]));
1345 		spa->spa_txg_zio[i] = NULL;
1346 	}
1347 
1348 	metaslab_class_destroy(spa->spa_normal_class);
1349 	spa->spa_normal_class = NULL;
1350 
1351 	metaslab_class_destroy(spa->spa_log_class);
1352 	spa->spa_log_class = NULL;
1353 
1354 	metaslab_class_destroy(spa->spa_embedded_log_class);
1355 	spa->spa_embedded_log_class = NULL;
1356 
1357 	metaslab_class_destroy(spa->spa_special_class);
1358 	spa->spa_special_class = NULL;
1359 
1360 	metaslab_class_destroy(spa->spa_dedup_class);
1361 	spa->spa_dedup_class = NULL;
1362 
1363 	/*
1364 	 * If this was part of an import or the open otherwise failed, we may
1365 	 * still have errors left in the queues.  Empty them just in case.
1366 	 */
1367 	spa_errlog_drain(spa);
1368 	avl_destroy(&spa->spa_errlist_scrub);
1369 	avl_destroy(&spa->spa_errlist_last);
1370 
1371 	spa_keystore_fini(&spa->spa_keystore);
1372 
1373 	spa->spa_state = POOL_STATE_UNINITIALIZED;
1374 
1375 	mutex_enter(&spa->spa_proc_lock);
1376 	if (spa->spa_proc_state != SPA_PROC_NONE) {
1377 		ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1378 		spa->spa_proc_state = SPA_PROC_DEACTIVATE;
1379 		cv_broadcast(&spa->spa_proc_cv);
1380 		while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) {
1381 			ASSERT(spa->spa_proc != &p0);
1382 			cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1383 		}
1384 		ASSERT(spa->spa_proc_state == SPA_PROC_GONE);
1385 		spa->spa_proc_state = SPA_PROC_NONE;
1386 	}
1387 	ASSERT(spa->spa_proc == &p0);
1388 	mutex_exit(&spa->spa_proc_lock);
1389 
1390 	/*
1391 	 * We want to make sure spa_thread() has actually exited the ZFS
1392 	 * module, so that the module can't be unloaded out from underneath
1393 	 * it.
1394 	 */
1395 	if (spa->spa_did != 0) {
1396 		thread_join(spa->spa_did);
1397 		spa->spa_did = 0;
1398 	}
1399 }
1400 
1401 /*
1402  * Verify a pool configuration, and construct the vdev tree appropriately.  This
1403  * will create all the necessary vdevs in the appropriate layout, with each vdev
1404  * in the CLOSED state.  This will prep the pool before open/creation/import.
1405  * All vdev validation is done by the vdev_alloc() routine.
1406  */
1407 int
1408 spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent,
1409     uint_t id, int atype)
1410 {
1411 	nvlist_t **child;
1412 	uint_t children;
1413 	int error;
1414 
1415 	if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0)
1416 		return (error);
1417 
1418 	if ((*vdp)->vdev_ops->vdev_op_leaf)
1419 		return (0);
1420 
1421 	error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1422 	    &child, &children);
1423 
1424 	if (error == ENOENT)
1425 		return (0);
1426 
1427 	if (error) {
1428 		vdev_free(*vdp);
1429 		*vdp = NULL;
1430 		return (SET_ERROR(EINVAL));
1431 	}
1432 
1433 	for (int c = 0; c < children; c++) {
1434 		vdev_t *vd;
1435 		if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c,
1436 		    atype)) != 0) {
1437 			vdev_free(*vdp);
1438 			*vdp = NULL;
1439 			return (error);
1440 		}
1441 	}
1442 
1443 	ASSERT(*vdp != NULL);
1444 
1445 	return (0);
1446 }
1447 
1448 static boolean_t
1449 spa_should_flush_logs_on_unload(spa_t *spa)
1450 {
1451 	if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))
1452 		return (B_FALSE);
1453 
1454 	if (!spa_writeable(spa))
1455 		return (B_FALSE);
1456 
1457 	if (!spa->spa_sync_on)
1458 		return (B_FALSE);
1459 
1460 	if (spa_state(spa) != POOL_STATE_EXPORTED)
1461 		return (B_FALSE);
1462 
1463 	if (zfs_keep_log_spacemaps_at_export)
1464 		return (B_FALSE);
1465 
1466 	return (B_TRUE);
1467 }
1468 
1469 /*
1470  * Opens a transaction that will set the flag that will instruct
1471  * spa_sync to attempt to flush all the metaslabs for that txg.
1472  */
1473 static void
1474 spa_unload_log_sm_flush_all(spa_t *spa)
1475 {
1476 	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
1477 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1478 
1479 	ASSERT3U(spa->spa_log_flushall_txg, ==, 0);
1480 	spa->spa_log_flushall_txg = dmu_tx_get_txg(tx);
1481 
1482 	dmu_tx_commit(tx);
1483 	txg_wait_synced(spa_get_dsl(spa), spa->spa_log_flushall_txg);
1484 }
1485 
1486 static void
1487 spa_unload_log_sm_metadata(spa_t *spa)
1488 {
1489 	void *cookie = NULL;
1490 	spa_log_sm_t *sls;
1491 	while ((sls = avl_destroy_nodes(&spa->spa_sm_logs_by_txg,
1492 	    &cookie)) != NULL) {
1493 		VERIFY0(sls->sls_mscount);
1494 		kmem_free(sls, sizeof (spa_log_sm_t));
1495 	}
1496 
1497 	for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
1498 	    e != NULL; e = list_head(&spa->spa_log_summary)) {
1499 		VERIFY0(e->lse_mscount);
1500 		list_remove(&spa->spa_log_summary, e);
1501 		kmem_free(e, sizeof (log_summary_entry_t));
1502 	}
1503 
1504 	spa->spa_unflushed_stats.sus_nblocks = 0;
1505 	spa->spa_unflushed_stats.sus_memused = 0;
1506 	spa->spa_unflushed_stats.sus_blocklimit = 0;
1507 }
1508 
1509 static void
1510 spa_destroy_aux_threads(spa_t *spa)
1511 {
1512 	if (spa->spa_condense_zthr != NULL) {
1513 		zthr_destroy(spa->spa_condense_zthr);
1514 		spa->spa_condense_zthr = NULL;
1515 	}
1516 	if (spa->spa_checkpoint_discard_zthr != NULL) {
1517 		zthr_destroy(spa->spa_checkpoint_discard_zthr);
1518 		spa->spa_checkpoint_discard_zthr = NULL;
1519 	}
1520 	if (spa->spa_livelist_delete_zthr != NULL) {
1521 		zthr_destroy(spa->spa_livelist_delete_zthr);
1522 		spa->spa_livelist_delete_zthr = NULL;
1523 	}
1524 	if (spa->spa_livelist_condense_zthr != NULL) {
1525 		zthr_destroy(spa->spa_livelist_condense_zthr);
1526 		spa->spa_livelist_condense_zthr = NULL;
1527 	}
1528 }
1529 
1530 /*
1531  * Opposite of spa_load().
1532  */
1533 static void
1534 spa_unload(spa_t *spa)
1535 {
1536 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1537 	ASSERT(spa_state(spa) != POOL_STATE_UNINITIALIZED);
1538 
1539 	spa_import_progress_remove(spa_guid(spa));
1540 	spa_load_note(spa, "UNLOADING");
1541 
1542 	spa_wake_waiters(spa);
1543 
1544 	/*
1545 	 * If the log space map feature is enabled and the pool is getting
1546 	 * exported (but not destroyed), we want to spend some time flushing
1547 	 * as many metaslabs as we can in an attempt to destroy log space
1548 	 * maps and save import time.
1549 	 */
1550 	if (spa_should_flush_logs_on_unload(spa))
1551 		spa_unload_log_sm_flush_all(spa);
1552 
1553 	/*
1554 	 * Stop async tasks.
1555 	 */
1556 	spa_async_suspend(spa);
1557 
1558 	if (spa->spa_root_vdev) {
1559 		vdev_t *root_vdev = spa->spa_root_vdev;
1560 		vdev_initialize_stop_all(root_vdev, VDEV_INITIALIZE_ACTIVE);
1561 		vdev_trim_stop_all(root_vdev, VDEV_TRIM_ACTIVE);
1562 		vdev_autotrim_stop_all(spa);
1563 		vdev_rebuild_stop_all(spa);
1564 	}
1565 
1566 	/*
1567 	 * Stop syncing.
1568 	 */
1569 	if (spa->spa_sync_on) {
1570 		txg_sync_stop(spa->spa_dsl_pool);
1571 		spa->spa_sync_on = B_FALSE;
1572 	}
1573 
1574 	/*
1575 	 * This ensures that there is no async metaslab prefetching
1576 	 * while we attempt to unload the spa.
1577 	 */
1578 	if (spa->spa_root_vdev != NULL) {
1579 		for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++) {
1580 			vdev_t *vc = spa->spa_root_vdev->vdev_child[c];
1581 			if (vc->vdev_mg != NULL)
1582 				taskq_wait(vc->vdev_mg->mg_taskq);
1583 		}
1584 	}
1585 
1586 	if (spa->spa_mmp.mmp_thread)
1587 		mmp_thread_stop(spa);
1588 
1589 	/*
1590 	 * Wait for any outstanding async I/O to complete.
1591 	 */
1592 	if (spa->spa_async_zio_root != NULL) {
1593 		for (int i = 0; i < max_ncpus; i++)
1594 			(void) zio_wait(spa->spa_async_zio_root[i]);
1595 		kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *));
1596 		spa->spa_async_zio_root = NULL;
1597 	}
1598 
1599 	if (spa->spa_vdev_removal != NULL) {
1600 		spa_vdev_removal_destroy(spa->spa_vdev_removal);
1601 		spa->spa_vdev_removal = NULL;
1602 	}
1603 
1604 	spa_destroy_aux_threads(spa);
1605 
1606 	spa_condense_fini(spa);
1607 
1608 	bpobj_close(&spa->spa_deferred_bpobj);
1609 
1610 	spa_config_enter(spa, SCL_ALL, spa, RW_WRITER);
1611 
1612 	/*
1613 	 * Close all vdevs.
1614 	 */
1615 	if (spa->spa_root_vdev)
1616 		vdev_free(spa->spa_root_vdev);
1617 	ASSERT(spa->spa_root_vdev == NULL);
1618 
1619 	/*
1620 	 * Close the dsl pool.
1621 	 */
1622 	if (spa->spa_dsl_pool) {
1623 		dsl_pool_close(spa->spa_dsl_pool);
1624 		spa->spa_dsl_pool = NULL;
1625 		spa->spa_meta_objset = NULL;
1626 	}
1627 
1628 	ddt_unload(spa);
1629 	spa_unload_log_sm_metadata(spa);
1630 
1631 	/*
1632 	 * Drop and purge level 2 cache
1633 	 */
1634 	spa_l2cache_drop(spa);
1635 
1636 	for (int i = 0; i < spa->spa_spares.sav_count; i++)
1637 		vdev_free(spa->spa_spares.sav_vdevs[i]);
1638 	if (spa->spa_spares.sav_vdevs) {
1639 		kmem_free(spa->spa_spares.sav_vdevs,
1640 		    spa->spa_spares.sav_count * sizeof (void *));
1641 		spa->spa_spares.sav_vdevs = NULL;
1642 	}
1643 	if (spa->spa_spares.sav_config) {
1644 		nvlist_free(spa->spa_spares.sav_config);
1645 		spa->spa_spares.sav_config = NULL;
1646 	}
1647 	spa->spa_spares.sav_count = 0;
1648 
1649 	for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
1650 		vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]);
1651 		vdev_free(spa->spa_l2cache.sav_vdevs[i]);
1652 	}
1653 	if (spa->spa_l2cache.sav_vdevs) {
1654 		kmem_free(spa->spa_l2cache.sav_vdevs,
1655 		    spa->spa_l2cache.sav_count * sizeof (void *));
1656 		spa->spa_l2cache.sav_vdevs = NULL;
1657 	}
1658 	if (spa->spa_l2cache.sav_config) {
1659 		nvlist_free(spa->spa_l2cache.sav_config);
1660 		spa->spa_l2cache.sav_config = NULL;
1661 	}
1662 	spa->spa_l2cache.sav_count = 0;
1663 
1664 	spa->spa_async_suspended = 0;
1665 
1666 	spa->spa_indirect_vdevs_loaded = B_FALSE;
1667 
1668 	if (spa->spa_comment != NULL) {
1669 		spa_strfree(spa->spa_comment);
1670 		spa->spa_comment = NULL;
1671 	}
1672 
1673 	spa_config_exit(spa, SCL_ALL, spa);
1674 }
1675 
1676 /*
1677  * Load (or re-load) the current list of vdevs describing the active spares for
1678  * this pool.  When this is called, we have some form of basic information in
1679  * 'spa_spares.sav_config'.  We parse this into vdevs, try to open them, and
1680  * then re-generate a more complete list including status information.
1681  */
1682 void
1683 spa_load_spares(spa_t *spa)
1684 {
1685 	nvlist_t **spares;
1686 	uint_t nspares;
1687 	int i;
1688 	vdev_t *vd, *tvd;
1689 
1690 #ifndef _KERNEL
1691 	/*
1692 	 * zdb opens both the current state of the pool and the
1693 	 * checkpointed state (if present), with a different spa_t.
1694 	 *
1695 	 * As spare vdevs are shared among open pools, we skip loading
1696 	 * them when we load the checkpointed state of the pool.
1697 	 */
1698 	if (!spa_writeable(spa))
1699 		return;
1700 #endif
1701 
1702 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1703 
1704 	/*
1705 	 * First, close and free any existing spare vdevs.
1706 	 */
1707 	for (i = 0; i < spa->spa_spares.sav_count; i++) {
1708 		vd = spa->spa_spares.sav_vdevs[i];
1709 
1710 		/* Undo the call to spa_activate() below */
1711 		if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1712 		    B_FALSE)) != NULL && tvd->vdev_isspare)
1713 			spa_spare_remove(tvd);
1714 		vdev_close(vd);
1715 		vdev_free(vd);
1716 	}
1717 
1718 	if (spa->spa_spares.sav_vdevs)
1719 		kmem_free(spa->spa_spares.sav_vdevs,
1720 		    spa->spa_spares.sav_count * sizeof (void *));
1721 
1722 	if (spa->spa_spares.sav_config == NULL)
1723 		nspares = 0;
1724 	else
1725 		VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
1726 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0);
1727 
1728 	spa->spa_spares.sav_count = (int)nspares;
1729 	spa->spa_spares.sav_vdevs = NULL;
1730 
1731 	if (nspares == 0)
1732 		return;
1733 
1734 	/*
1735 	 * Construct the array of vdevs, opening them to get status in the
1736 	 * process.   For each spare, there is potentially two different vdev_t
1737 	 * structures associated with it: one in the list of spares (used only
1738 	 * for basic validation purposes) and one in the active vdev
1739 	 * configuration (if it's spared in).  During this phase we open and
1740 	 * validate each vdev on the spare list.  If the vdev also exists in the
1741 	 * active configuration, then we also mark this vdev as an active spare.
1742 	 */
1743 	spa->spa_spares.sav_vdevs = kmem_zalloc(nspares * sizeof (void *),
1744 	    KM_SLEEP);
1745 	for (i = 0; i < spa->spa_spares.sav_count; i++) {
1746 		VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0,
1747 		    VDEV_ALLOC_SPARE) == 0);
1748 		ASSERT(vd != NULL);
1749 
1750 		spa->spa_spares.sav_vdevs[i] = vd;
1751 
1752 		if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1753 		    B_FALSE)) != NULL) {
1754 			if (!tvd->vdev_isspare)
1755 				spa_spare_add(tvd);
1756 
1757 			/*
1758 			 * We only mark the spare active if we were successfully
1759 			 * able to load the vdev.  Otherwise, importing a pool
1760 			 * with a bad active spare would result in strange
1761 			 * behavior, because multiple pool would think the spare
1762 			 * is actively in use.
1763 			 *
1764 			 * There is a vulnerability here to an equally bizarre
1765 			 * circumstance, where a dead active spare is later
1766 			 * brought back to life (onlined or otherwise).  Given
1767 			 * the rarity of this scenario, and the extra complexity
1768 			 * it adds, we ignore the possibility.
1769 			 */
1770 			if (!vdev_is_dead(tvd))
1771 				spa_spare_activate(tvd);
1772 		}
1773 
1774 		vd->vdev_top = vd;
1775 		vd->vdev_aux = &spa->spa_spares;
1776 
1777 		if (vdev_open(vd) != 0)
1778 			continue;
1779 
1780 		if (vdev_validate_aux(vd) == 0)
1781 			spa_spare_add(vd);
1782 	}
1783 
1784 	/*
1785 	 * Recompute the stashed list of spares, with status information
1786 	 * this time.
1787 	 */
1788 	VERIFY(nvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES,
1789 	    DATA_TYPE_NVLIST_ARRAY) == 0);
1790 
1791 	spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *),
1792 	    KM_SLEEP);
1793 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1794 		spares[i] = vdev_config_generate(spa,
1795 		    spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE);
1796 	VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config,
1797 	    ZPOOL_CONFIG_SPARES, spares, spa->spa_spares.sav_count) == 0);
1798 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1799 		nvlist_free(spares[i]);
1800 	kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *));
1801 }
1802 
1803 /*
1804  * Load (or re-load) the current list of vdevs describing the active l2cache for
1805  * this pool.  When this is called, we have some form of basic information in
1806  * 'spa_l2cache.sav_config'.  We parse this into vdevs, try to open them, and
1807  * then re-generate a more complete list including status information.
1808  * Devices which are already active have their details maintained, and are
1809  * not re-opened.
1810  */
1811 void
1812 spa_load_l2cache(spa_t *spa)
1813 {
1814 	nvlist_t **l2cache = NULL;
1815 	uint_t nl2cache;
1816 	int i, j, oldnvdevs;
1817 	uint64_t guid;
1818 	vdev_t *vd, **oldvdevs, **newvdevs;
1819 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
1820 
1821 #ifndef _KERNEL
1822 	/*
1823 	 * zdb opens both the current state of the pool and the
1824 	 * checkpointed state (if present), with a different spa_t.
1825 	 *
1826 	 * As L2 caches are part of the ARC which is shared among open
1827 	 * pools, we skip loading them when we load the checkpointed
1828 	 * state of the pool.
1829 	 */
1830 	if (!spa_writeable(spa))
1831 		return;
1832 #endif
1833 
1834 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1835 
1836 	oldvdevs = sav->sav_vdevs;
1837 	oldnvdevs = sav->sav_count;
1838 	sav->sav_vdevs = NULL;
1839 	sav->sav_count = 0;
1840 
1841 	if (sav->sav_config == NULL) {
1842 		nl2cache = 0;
1843 		newvdevs = NULL;
1844 		goto out;
1845 	}
1846 
1847 	VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
1848 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0);
1849 	newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP);
1850 
1851 	/*
1852 	 * Process new nvlist of vdevs.
1853 	 */
1854 	for (i = 0; i < nl2cache; i++) {
1855 		VERIFY(nvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID,
1856 		    &guid) == 0);
1857 
1858 		newvdevs[i] = NULL;
1859 		for (j = 0; j < oldnvdevs; j++) {
1860 			vd = oldvdevs[j];
1861 			if (vd != NULL && guid == vd->vdev_guid) {
1862 				/*
1863 				 * Retain previous vdev for add/remove ops.
1864 				 */
1865 				newvdevs[i] = vd;
1866 				oldvdevs[j] = NULL;
1867 				break;
1868 			}
1869 		}
1870 
1871 		if (newvdevs[i] == NULL) {
1872 			/*
1873 			 * Create new vdev
1874 			 */
1875 			VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0,
1876 			    VDEV_ALLOC_L2CACHE) == 0);
1877 			ASSERT(vd != NULL);
1878 			newvdevs[i] = vd;
1879 
1880 			/*
1881 			 * Commit this vdev as an l2cache device,
1882 			 * even if it fails to open.
1883 			 */
1884 			spa_l2cache_add(vd);
1885 
1886 			vd->vdev_top = vd;
1887 			vd->vdev_aux = sav;
1888 
1889 			spa_l2cache_activate(vd);
1890 
1891 			if (vdev_open(vd) != 0)
1892 				continue;
1893 
1894 			(void) vdev_validate_aux(vd);
1895 
1896 			if (!vdev_is_dead(vd))
1897 				l2arc_add_vdev(spa, vd);
1898 
1899 			/*
1900 			 * Upon cache device addition to a pool or pool
1901 			 * creation with a cache device or if the header
1902 			 * of the device is invalid we issue an async
1903 			 * TRIM command for the whole device which will
1904 			 * execute if l2arc_trim_ahead > 0.
1905 			 */
1906 			spa_async_request(spa, SPA_ASYNC_L2CACHE_TRIM);
1907 		}
1908 	}
1909 
1910 	sav->sav_vdevs = newvdevs;
1911 	sav->sav_count = (int)nl2cache;
1912 
1913 	/*
1914 	 * Recompute the stashed list of l2cache devices, with status
1915 	 * information this time.
1916 	 */
1917 	VERIFY(nvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE,
1918 	    DATA_TYPE_NVLIST_ARRAY) == 0);
1919 
1920 	if (sav->sav_count > 0)
1921 		l2cache = kmem_alloc(sav->sav_count * sizeof (void *),
1922 		    KM_SLEEP);
1923 	for (i = 0; i < sav->sav_count; i++)
1924 		l2cache[i] = vdev_config_generate(spa,
1925 		    sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE);
1926 	VERIFY(nvlist_add_nvlist_array(sav->sav_config,
1927 	    ZPOOL_CONFIG_L2CACHE, l2cache, sav->sav_count) == 0);
1928 
1929 out:
1930 	/*
1931 	 * Purge vdevs that were dropped
1932 	 */
1933 	for (i = 0; i < oldnvdevs; i++) {
1934 		uint64_t pool;
1935 
1936 		vd = oldvdevs[i];
1937 		if (vd != NULL) {
1938 			ASSERT(vd->vdev_isl2cache);
1939 
1940 			if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
1941 			    pool != 0ULL && l2arc_vdev_present(vd))
1942 				l2arc_remove_vdev(vd);
1943 			vdev_clear_stats(vd);
1944 			vdev_free(vd);
1945 		}
1946 	}
1947 
1948 	if (oldvdevs)
1949 		kmem_free(oldvdevs, oldnvdevs * sizeof (void *));
1950 
1951 	for (i = 0; i < sav->sav_count; i++)
1952 		nvlist_free(l2cache[i]);
1953 	if (sav->sav_count)
1954 		kmem_free(l2cache, sav->sav_count * sizeof (void *));
1955 }
1956 
1957 static int
1958 load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value)
1959 {
1960 	dmu_buf_t *db;
1961 	char *packed = NULL;
1962 	size_t nvsize = 0;
1963 	int error;
1964 	*value = NULL;
1965 
1966 	error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db);
1967 	if (error)
1968 		return (error);
1969 
1970 	nvsize = *(uint64_t *)db->db_data;
1971 	dmu_buf_rele(db, FTAG);
1972 
1973 	packed = vmem_alloc(nvsize, KM_SLEEP);
1974 	error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed,
1975 	    DMU_READ_PREFETCH);
1976 	if (error == 0)
1977 		error = nvlist_unpack(packed, nvsize, value, 0);
1978 	vmem_free(packed, nvsize);
1979 
1980 	return (error);
1981 }
1982 
1983 /*
1984  * Concrete top-level vdevs that are not missing and are not logs. At every
1985  * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds.
1986  */
1987 static uint64_t
1988 spa_healthy_core_tvds(spa_t *spa)
1989 {
1990 	vdev_t *rvd = spa->spa_root_vdev;
1991 	uint64_t tvds = 0;
1992 
1993 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1994 		vdev_t *vd = rvd->vdev_child[i];
1995 		if (vd->vdev_islog)
1996 			continue;
1997 		if (vdev_is_concrete(vd) && !vdev_is_dead(vd))
1998 			tvds++;
1999 	}
2000 
2001 	return (tvds);
2002 }
2003 
2004 /*
2005  * Checks to see if the given vdev could not be opened, in which case we post a
2006  * sysevent to notify the autoreplace code that the device has been removed.
2007  */
2008 static void
2009 spa_check_removed(vdev_t *vd)
2010 {
2011 	for (uint64_t c = 0; c < vd->vdev_children; c++)
2012 		spa_check_removed(vd->vdev_child[c]);
2013 
2014 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) &&
2015 	    vdev_is_concrete(vd)) {
2016 		zfs_post_autoreplace(vd->vdev_spa, vd);
2017 		spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK);
2018 	}
2019 }
2020 
2021 static int
2022 spa_check_for_missing_logs(spa_t *spa)
2023 {
2024 	vdev_t *rvd = spa->spa_root_vdev;
2025 
2026 	/*
2027 	 * If we're doing a normal import, then build up any additional
2028 	 * diagnostic information about missing log devices.
2029 	 * We'll pass this up to the user for further processing.
2030 	 */
2031 	if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) {
2032 		nvlist_t **child, *nv;
2033 		uint64_t idx = 0;
2034 
2035 		child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t *),
2036 		    KM_SLEEP);
2037 		VERIFY(nvlist_alloc(&nv, NV_UNIQUE_NAME, KM_SLEEP) == 0);
2038 
2039 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
2040 			vdev_t *tvd = rvd->vdev_child[c];
2041 
2042 			/*
2043 			 * We consider a device as missing only if it failed
2044 			 * to open (i.e. offline or faulted is not considered
2045 			 * as missing).
2046 			 */
2047 			if (tvd->vdev_islog &&
2048 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
2049 				child[idx++] = vdev_config_generate(spa, tvd,
2050 				    B_FALSE, VDEV_CONFIG_MISSING);
2051 			}
2052 		}
2053 
2054 		if (idx > 0) {
2055 			fnvlist_add_nvlist_array(nv,
2056 			    ZPOOL_CONFIG_CHILDREN, child, idx);
2057 			fnvlist_add_nvlist(spa->spa_load_info,
2058 			    ZPOOL_CONFIG_MISSING_DEVICES, nv);
2059 
2060 			for (uint64_t i = 0; i < idx; i++)
2061 				nvlist_free(child[i]);
2062 		}
2063 		nvlist_free(nv);
2064 		kmem_free(child, rvd->vdev_children * sizeof (char **));
2065 
2066 		if (idx > 0) {
2067 			spa_load_failed(spa, "some log devices are missing");
2068 			vdev_dbgmsg_print_tree(rvd, 2);
2069 			return (SET_ERROR(ENXIO));
2070 		}
2071 	} else {
2072 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
2073 			vdev_t *tvd = rvd->vdev_child[c];
2074 
2075 			if (tvd->vdev_islog &&
2076 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
2077 				spa_set_log_state(spa, SPA_LOG_CLEAR);
2078 				spa_load_note(spa, "some log devices are "
2079 				    "missing, ZIL is dropped.");
2080 				vdev_dbgmsg_print_tree(rvd, 2);
2081 				break;
2082 			}
2083 		}
2084 	}
2085 
2086 	return (0);
2087 }
2088 
2089 /*
2090  * Check for missing log devices
2091  */
2092 static boolean_t
2093 spa_check_logs(spa_t *spa)
2094 {
2095 	boolean_t rv = B_FALSE;
2096 	dsl_pool_t *dp = spa_get_dsl(spa);
2097 
2098 	switch (spa->spa_log_state) {
2099 	default:
2100 		break;
2101 	case SPA_LOG_MISSING:
2102 		/* need to recheck in case slog has been restored */
2103 	case SPA_LOG_UNKNOWN:
2104 		rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
2105 		    zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0);
2106 		if (rv)
2107 			spa_set_log_state(spa, SPA_LOG_MISSING);
2108 		break;
2109 	}
2110 	return (rv);
2111 }
2112 
2113 /*
2114  * Passivate any log vdevs (note, does not apply to embedded log metaslabs).
2115  */
2116 static boolean_t
2117 spa_passivate_log(spa_t *spa)
2118 {
2119 	vdev_t *rvd = spa->spa_root_vdev;
2120 	boolean_t slog_found = B_FALSE;
2121 
2122 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
2123 
2124 	for (int c = 0; c < rvd->vdev_children; c++) {
2125 		vdev_t *tvd = rvd->vdev_child[c];
2126 
2127 		if (tvd->vdev_islog) {
2128 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
2129 			metaslab_group_passivate(tvd->vdev_mg);
2130 			slog_found = B_TRUE;
2131 		}
2132 	}
2133 
2134 	return (slog_found);
2135 }
2136 
2137 /*
2138  * Activate any log vdevs (note, does not apply to embedded log metaslabs).
2139  */
2140 static void
2141 spa_activate_log(spa_t *spa)
2142 {
2143 	vdev_t *rvd = spa->spa_root_vdev;
2144 
2145 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
2146 
2147 	for (int c = 0; c < rvd->vdev_children; c++) {
2148 		vdev_t *tvd = rvd->vdev_child[c];
2149 
2150 		if (tvd->vdev_islog) {
2151 			ASSERT3P(tvd->vdev_log_mg, ==, NULL);
2152 			metaslab_group_activate(tvd->vdev_mg);
2153 		}
2154 	}
2155 }
2156 
2157 int
2158 spa_reset_logs(spa_t *spa)
2159 {
2160 	int error;
2161 
2162 	error = dmu_objset_find(spa_name(spa), zil_reset,
2163 	    NULL, DS_FIND_CHILDREN);
2164 	if (error == 0) {
2165 		/*
2166 		 * We successfully offlined the log device, sync out the
2167 		 * current txg so that the "stubby" block can be removed
2168 		 * by zil_sync().
2169 		 */
2170 		txg_wait_synced(spa->spa_dsl_pool, 0);
2171 	}
2172 	return (error);
2173 }
2174 
2175 static void
2176 spa_aux_check_removed(spa_aux_vdev_t *sav)
2177 {
2178 	for (int i = 0; i < sav->sav_count; i++)
2179 		spa_check_removed(sav->sav_vdevs[i]);
2180 }
2181 
2182 void
2183 spa_claim_notify(zio_t *zio)
2184 {
2185 	spa_t *spa = zio->io_spa;
2186 
2187 	if (zio->io_error)
2188 		return;
2189 
2190 	mutex_enter(&spa->spa_props_lock);	/* any mutex will do */
2191 	if (spa->spa_claim_max_txg < zio->io_bp->blk_birth)
2192 		spa->spa_claim_max_txg = zio->io_bp->blk_birth;
2193 	mutex_exit(&spa->spa_props_lock);
2194 }
2195 
2196 typedef struct spa_load_error {
2197 	uint64_t	sle_meta_count;
2198 	uint64_t	sle_data_count;
2199 } spa_load_error_t;
2200 
2201 static void
2202 spa_load_verify_done(zio_t *zio)
2203 {
2204 	blkptr_t *bp = zio->io_bp;
2205 	spa_load_error_t *sle = zio->io_private;
2206 	dmu_object_type_t type = BP_GET_TYPE(bp);
2207 	int error = zio->io_error;
2208 	spa_t *spa = zio->io_spa;
2209 
2210 	abd_free(zio->io_abd);
2211 	if (error) {
2212 		if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) &&
2213 		    type != DMU_OT_INTENT_LOG)
2214 			atomic_inc_64(&sle->sle_meta_count);
2215 		else
2216 			atomic_inc_64(&sle->sle_data_count);
2217 	}
2218 
2219 	mutex_enter(&spa->spa_scrub_lock);
2220 	spa->spa_load_verify_bytes -= BP_GET_PSIZE(bp);
2221 	cv_broadcast(&spa->spa_scrub_io_cv);
2222 	mutex_exit(&spa->spa_scrub_lock);
2223 }
2224 
2225 /*
2226  * Maximum number of inflight bytes is the log2 fraction of the arc size.
2227  * By default, we set it to 1/16th of the arc.
2228  */
2229 int spa_load_verify_shift = 4;
2230 int spa_load_verify_metadata = B_TRUE;
2231 int spa_load_verify_data = B_TRUE;
2232 
2233 /*ARGSUSED*/
2234 static int
2235 spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
2236     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
2237 {
2238 	if (zb->zb_level == ZB_DNODE_LEVEL || BP_IS_HOLE(bp) ||
2239 	    BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp))
2240 		return (0);
2241 	/*
2242 	 * Note: normally this routine will not be called if
2243 	 * spa_load_verify_metadata is not set.  However, it may be useful
2244 	 * to manually set the flag after the traversal has begun.
2245 	 */
2246 	if (!spa_load_verify_metadata)
2247 		return (0);
2248 	if (!BP_IS_METADATA(bp) && !spa_load_verify_data)
2249 		return (0);
2250 
2251 	uint64_t maxinflight_bytes =
2252 	    arc_target_bytes() >> spa_load_verify_shift;
2253 	zio_t *rio = arg;
2254 	size_t size = BP_GET_PSIZE(bp);
2255 
2256 	mutex_enter(&spa->spa_scrub_lock);
2257 	while (spa->spa_load_verify_bytes >= maxinflight_bytes)
2258 		cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2259 	spa->spa_load_verify_bytes += size;
2260 	mutex_exit(&spa->spa_scrub_lock);
2261 
2262 	zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size,
2263 	    spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB,
2264 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL |
2265 	    ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb));
2266 	return (0);
2267 }
2268 
2269 /* ARGSUSED */
2270 static int
2271 verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
2272 {
2273 	if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN)
2274 		return (SET_ERROR(ENAMETOOLONG));
2275 
2276 	return (0);
2277 }
2278 
2279 static int
2280 spa_load_verify(spa_t *spa)
2281 {
2282 	zio_t *rio;
2283 	spa_load_error_t sle = { 0 };
2284 	zpool_load_policy_t policy;
2285 	boolean_t verify_ok = B_FALSE;
2286 	int error = 0;
2287 
2288 	zpool_get_load_policy(spa->spa_config, &policy);
2289 
2290 	if (policy.zlp_rewind & ZPOOL_NEVER_REWIND)
2291 		return (0);
2292 
2293 	dsl_pool_config_enter(spa->spa_dsl_pool, FTAG);
2294 	error = dmu_objset_find_dp(spa->spa_dsl_pool,
2295 	    spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL,
2296 	    DS_FIND_CHILDREN);
2297 	dsl_pool_config_exit(spa->spa_dsl_pool, FTAG);
2298 	if (error != 0)
2299 		return (error);
2300 
2301 	rio = zio_root(spa, NULL, &sle,
2302 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
2303 
2304 	if (spa_load_verify_metadata) {
2305 		if (spa->spa_extreme_rewind) {
2306 			spa_load_note(spa, "performing a complete scan of the "
2307 			    "pool since extreme rewind is on. This may take "
2308 			    "a very long time.\n  (spa_load_verify_data=%u, "
2309 			    "spa_load_verify_metadata=%u)",
2310 			    spa_load_verify_data, spa_load_verify_metadata);
2311 		}
2312 
2313 		error = traverse_pool(spa, spa->spa_verify_min_txg,
2314 		    TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA |
2315 		    TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio);
2316 	}
2317 
2318 	(void) zio_wait(rio);
2319 	ASSERT0(spa->spa_load_verify_bytes);
2320 
2321 	spa->spa_load_meta_errors = sle.sle_meta_count;
2322 	spa->spa_load_data_errors = sle.sle_data_count;
2323 
2324 	if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) {
2325 		spa_load_note(spa, "spa_load_verify found %llu metadata errors "
2326 		    "and %llu data errors", (u_longlong_t)sle.sle_meta_count,
2327 		    (u_longlong_t)sle.sle_data_count);
2328 	}
2329 
2330 	if (spa_load_verify_dryrun ||
2331 	    (!error && sle.sle_meta_count <= policy.zlp_maxmeta &&
2332 	    sle.sle_data_count <= policy.zlp_maxdata)) {
2333 		int64_t loss = 0;
2334 
2335 		verify_ok = B_TRUE;
2336 		spa->spa_load_txg = spa->spa_uberblock.ub_txg;
2337 		spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp;
2338 
2339 		loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts;
2340 		VERIFY(nvlist_add_uint64(spa->spa_load_info,
2341 		    ZPOOL_CONFIG_LOAD_TIME, spa->spa_load_txg_ts) == 0);
2342 		VERIFY(nvlist_add_int64(spa->spa_load_info,
2343 		    ZPOOL_CONFIG_REWIND_TIME, loss) == 0);
2344 		VERIFY(nvlist_add_uint64(spa->spa_load_info,
2345 		    ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count) == 0);
2346 	} else {
2347 		spa->spa_load_max_txg = spa->spa_uberblock.ub_txg;
2348 	}
2349 
2350 	if (spa_load_verify_dryrun)
2351 		return (0);
2352 
2353 	if (error) {
2354 		if (error != ENXIO && error != EIO)
2355 			error = SET_ERROR(EIO);
2356 		return (error);
2357 	}
2358 
2359 	return (verify_ok ? 0 : EIO);
2360 }
2361 
2362 /*
2363  * Find a value in the pool props object.
2364  */
2365 static void
2366 spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val)
2367 {
2368 	(void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object,
2369 	    zpool_prop_to_name(prop), sizeof (uint64_t), 1, val);
2370 }
2371 
2372 /*
2373  * Find a value in the pool directory object.
2374  */
2375 static int
2376 spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent)
2377 {
2378 	int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
2379 	    name, sizeof (uint64_t), 1, val);
2380 
2381 	if (error != 0 && (error != ENOENT || log_enoent)) {
2382 		spa_load_failed(spa, "couldn't get '%s' value in MOS directory "
2383 		    "[error=%d]", name, error);
2384 	}
2385 
2386 	return (error);
2387 }
2388 
2389 static int
2390 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err)
2391 {
2392 	vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux);
2393 	return (SET_ERROR(err));
2394 }
2395 
2396 boolean_t
2397 spa_livelist_delete_check(spa_t *spa)
2398 {
2399 	return (spa->spa_livelists_to_delete != 0);
2400 }
2401 
2402 /* ARGSUSED */
2403 static boolean_t
2404 spa_livelist_delete_cb_check(void *arg, zthr_t *z)
2405 {
2406 	spa_t *spa = arg;
2407 	return (spa_livelist_delete_check(spa));
2408 }
2409 
2410 static int
2411 delete_blkptr_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
2412 {
2413 	spa_t *spa = arg;
2414 	zio_free(spa, tx->tx_txg, bp);
2415 	dsl_dir_diduse_space(tx->tx_pool->dp_free_dir, DD_USED_HEAD,
2416 	    -bp_get_dsize_sync(spa, bp),
2417 	    -BP_GET_PSIZE(bp), -BP_GET_UCSIZE(bp), tx);
2418 	return (0);
2419 }
2420 
2421 static int
2422 dsl_get_next_livelist_obj(objset_t *os, uint64_t zap_obj, uint64_t *llp)
2423 {
2424 	int err;
2425 	zap_cursor_t zc;
2426 	zap_attribute_t za;
2427 	zap_cursor_init(&zc, os, zap_obj);
2428 	err = zap_cursor_retrieve(&zc, &za);
2429 	zap_cursor_fini(&zc);
2430 	if (err == 0)
2431 		*llp = za.za_first_integer;
2432 	return (err);
2433 }
2434 
2435 /*
2436  * Components of livelist deletion that must be performed in syncing
2437  * context: freeing block pointers and updating the pool-wide data
2438  * structures to indicate how much work is left to do
2439  */
2440 typedef struct sublist_delete_arg {
2441 	spa_t *spa;
2442 	dsl_deadlist_t *ll;
2443 	uint64_t key;
2444 	bplist_t *to_free;
2445 } sublist_delete_arg_t;
2446 
2447 static void
2448 sublist_delete_sync(void *arg, dmu_tx_t *tx)
2449 {
2450 	sublist_delete_arg_t *sda = arg;
2451 	spa_t *spa = sda->spa;
2452 	dsl_deadlist_t *ll = sda->ll;
2453 	uint64_t key = sda->key;
2454 	bplist_t *to_free = sda->to_free;
2455 
2456 	bplist_iterate(to_free, delete_blkptr_cb, spa, tx);
2457 	dsl_deadlist_remove_entry(ll, key, tx);
2458 }
2459 
2460 typedef struct livelist_delete_arg {
2461 	spa_t *spa;
2462 	uint64_t ll_obj;
2463 	uint64_t zap_obj;
2464 } livelist_delete_arg_t;
2465 
2466 static void
2467 livelist_delete_sync(void *arg, dmu_tx_t *tx)
2468 {
2469 	livelist_delete_arg_t *lda = arg;
2470 	spa_t *spa = lda->spa;
2471 	uint64_t ll_obj = lda->ll_obj;
2472 	uint64_t zap_obj = lda->zap_obj;
2473 	objset_t *mos = spa->spa_meta_objset;
2474 	uint64_t count;
2475 
2476 	/* free the livelist and decrement the feature count */
2477 	VERIFY0(zap_remove_int(mos, zap_obj, ll_obj, tx));
2478 	dsl_deadlist_free(mos, ll_obj, tx);
2479 	spa_feature_decr(spa, SPA_FEATURE_LIVELIST, tx);
2480 	VERIFY0(zap_count(mos, zap_obj, &count));
2481 	if (count == 0) {
2482 		/* no more livelists to delete */
2483 		VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT,
2484 		    DMU_POOL_DELETED_CLONES, tx));
2485 		VERIFY0(zap_destroy(mos, zap_obj, tx));
2486 		spa->spa_livelists_to_delete = 0;
2487 		spa_notify_waiters(spa);
2488 	}
2489 }
2490 
2491 /*
2492  * Load in the value for the livelist to be removed and open it. Then,
2493  * load its first sublist and determine which block pointers should actually
2494  * be freed. Then, call a synctask which performs the actual frees and updates
2495  * the pool-wide livelist data.
2496  */
2497 /* ARGSUSED */
2498 static void
2499 spa_livelist_delete_cb(void *arg, zthr_t *z)
2500 {
2501 	spa_t *spa = arg;
2502 	uint64_t ll_obj = 0, count;
2503 	objset_t *mos = spa->spa_meta_objset;
2504 	uint64_t zap_obj = spa->spa_livelists_to_delete;
2505 	/*
2506 	 * Determine the next livelist to delete. This function should only
2507 	 * be called if there is at least one deleted clone.
2508 	 */
2509 	VERIFY0(dsl_get_next_livelist_obj(mos, zap_obj, &ll_obj));
2510 	VERIFY0(zap_count(mos, ll_obj, &count));
2511 	if (count > 0) {
2512 		dsl_deadlist_t *ll;
2513 		dsl_deadlist_entry_t *dle;
2514 		bplist_t to_free;
2515 		ll = kmem_zalloc(sizeof (dsl_deadlist_t), KM_SLEEP);
2516 		dsl_deadlist_open(ll, mos, ll_obj);
2517 		dle = dsl_deadlist_first(ll);
2518 		ASSERT3P(dle, !=, NULL);
2519 		bplist_create(&to_free);
2520 		int err = dsl_process_sub_livelist(&dle->dle_bpobj, &to_free,
2521 		    z, NULL);
2522 		if (err == 0) {
2523 			sublist_delete_arg_t sync_arg = {
2524 			    .spa = spa,
2525 			    .ll = ll,
2526 			    .key = dle->dle_mintxg,
2527 			    .to_free = &to_free
2528 			};
2529 			zfs_dbgmsg("deleting sublist (id %llu) from"
2530 			    " livelist %llu, %d remaining",
2531 			    dle->dle_bpobj.bpo_object, ll_obj, count - 1);
2532 			VERIFY0(dsl_sync_task(spa_name(spa), NULL,
2533 			    sublist_delete_sync, &sync_arg, 0,
2534 			    ZFS_SPACE_CHECK_DESTROY));
2535 		} else {
2536 			VERIFY3U(err, ==, EINTR);
2537 		}
2538 		bplist_clear(&to_free);
2539 		bplist_destroy(&to_free);
2540 		dsl_deadlist_close(ll);
2541 		kmem_free(ll, sizeof (dsl_deadlist_t));
2542 	} else {
2543 		livelist_delete_arg_t sync_arg = {
2544 		    .spa = spa,
2545 		    .ll_obj = ll_obj,
2546 		    .zap_obj = zap_obj
2547 		};
2548 		zfs_dbgmsg("deletion of livelist %llu completed", ll_obj);
2549 		VERIFY0(dsl_sync_task(spa_name(spa), NULL, livelist_delete_sync,
2550 		    &sync_arg, 0, ZFS_SPACE_CHECK_DESTROY));
2551 	}
2552 }
2553 
2554 static void
2555 spa_start_livelist_destroy_thread(spa_t *spa)
2556 {
2557 	ASSERT3P(spa->spa_livelist_delete_zthr, ==, NULL);
2558 	spa->spa_livelist_delete_zthr =
2559 	    zthr_create("z_livelist_destroy",
2560 	    spa_livelist_delete_cb_check, spa_livelist_delete_cb, spa);
2561 }
2562 
2563 typedef struct livelist_new_arg {
2564 	bplist_t *allocs;
2565 	bplist_t *frees;
2566 } livelist_new_arg_t;
2567 
2568 static int
2569 livelist_track_new_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
2570     dmu_tx_t *tx)
2571 {
2572 	ASSERT(tx == NULL);
2573 	livelist_new_arg_t *lna = arg;
2574 	if (bp_freed) {
2575 		bplist_append(lna->frees, bp);
2576 	} else {
2577 		bplist_append(lna->allocs, bp);
2578 		zfs_livelist_condense_new_alloc++;
2579 	}
2580 	return (0);
2581 }
2582 
2583 typedef struct livelist_condense_arg {
2584 	spa_t *spa;
2585 	bplist_t to_keep;
2586 	uint64_t first_size;
2587 	uint64_t next_size;
2588 } livelist_condense_arg_t;
2589 
2590 static void
2591 spa_livelist_condense_sync(void *arg, dmu_tx_t *tx)
2592 {
2593 	livelist_condense_arg_t *lca = arg;
2594 	spa_t *spa = lca->spa;
2595 	bplist_t new_frees;
2596 	dsl_dataset_t *ds = spa->spa_to_condense.ds;
2597 
2598 	/* Have we been cancelled? */
2599 	if (spa->spa_to_condense.cancelled) {
2600 		zfs_livelist_condense_sync_cancel++;
2601 		goto out;
2602 	}
2603 
2604 	dsl_deadlist_entry_t *first = spa->spa_to_condense.first;
2605 	dsl_deadlist_entry_t *next = spa->spa_to_condense.next;
2606 	dsl_deadlist_t *ll = &ds->ds_dir->dd_livelist;
2607 
2608 	/*
2609 	 * It's possible that the livelist was changed while the zthr was
2610 	 * running. Therefore, we need to check for new blkptrs in the two
2611 	 * entries being condensed and continue to track them in the livelist.
2612 	 * Because of the way we handle remapped blkptrs (see dbuf_remap_impl),
2613 	 * it's possible that the newly added blkptrs are FREEs or ALLOCs so
2614 	 * we need to sort them into two different bplists.
2615 	 */
2616 	uint64_t first_obj = first->dle_bpobj.bpo_object;
2617 	uint64_t next_obj = next->dle_bpobj.bpo_object;
2618 	uint64_t cur_first_size = first->dle_bpobj.bpo_phys->bpo_num_blkptrs;
2619 	uint64_t cur_next_size = next->dle_bpobj.bpo_phys->bpo_num_blkptrs;
2620 
2621 	bplist_create(&new_frees);
2622 	livelist_new_arg_t new_bps = {
2623 	    .allocs = &lca->to_keep,
2624 	    .frees = &new_frees,
2625 	};
2626 
2627 	if (cur_first_size > lca->first_size) {
2628 		VERIFY0(livelist_bpobj_iterate_from_nofree(&first->dle_bpobj,
2629 		    livelist_track_new_cb, &new_bps, lca->first_size));
2630 	}
2631 	if (cur_next_size > lca->next_size) {
2632 		VERIFY0(livelist_bpobj_iterate_from_nofree(&next->dle_bpobj,
2633 		    livelist_track_new_cb, &new_bps, lca->next_size));
2634 	}
2635 
2636 	dsl_deadlist_clear_entry(first, ll, tx);
2637 	ASSERT(bpobj_is_empty(&first->dle_bpobj));
2638 	dsl_deadlist_remove_entry(ll, next->dle_mintxg, tx);
2639 
2640 	bplist_iterate(&lca->to_keep, dsl_deadlist_insert_alloc_cb, ll, tx);
2641 	bplist_iterate(&new_frees, dsl_deadlist_insert_free_cb, ll, tx);
2642 	bplist_destroy(&new_frees);
2643 
2644 	char dsname[ZFS_MAX_DATASET_NAME_LEN];
2645 	dsl_dataset_name(ds, dsname);
2646 	zfs_dbgmsg("txg %llu condensing livelist of %s (id %llu), bpobj %llu "
2647 	    "(%llu blkptrs) and bpobj %llu (%llu blkptrs) -> bpobj %llu "
2648 	    "(%llu blkptrs)", tx->tx_txg, dsname, ds->ds_object, first_obj,
2649 	    cur_first_size, next_obj, cur_next_size,
2650 	    first->dle_bpobj.bpo_object,
2651 	    first->dle_bpobj.bpo_phys->bpo_num_blkptrs);
2652 out:
2653 	dmu_buf_rele(ds->ds_dbuf, spa);
2654 	spa->spa_to_condense.ds = NULL;
2655 	bplist_clear(&lca->to_keep);
2656 	bplist_destroy(&lca->to_keep);
2657 	kmem_free(lca, sizeof (livelist_condense_arg_t));
2658 	spa->spa_to_condense.syncing = B_FALSE;
2659 }
2660 
2661 static void
2662 spa_livelist_condense_cb(void *arg, zthr_t *t)
2663 {
2664 	while (zfs_livelist_condense_zthr_pause &&
2665 	    !(zthr_has_waiters(t) || zthr_iscancelled(t)))
2666 		delay(1);
2667 
2668 	spa_t *spa = arg;
2669 	dsl_deadlist_entry_t *first = spa->spa_to_condense.first;
2670 	dsl_deadlist_entry_t *next = spa->spa_to_condense.next;
2671 	uint64_t first_size, next_size;
2672 
2673 	livelist_condense_arg_t *lca =
2674 	    kmem_alloc(sizeof (livelist_condense_arg_t), KM_SLEEP);
2675 	bplist_create(&lca->to_keep);
2676 
2677 	/*
2678 	 * Process the livelists (matching FREEs and ALLOCs) in open context
2679 	 * so we have minimal work in syncing context to condense.
2680 	 *
2681 	 * We save bpobj sizes (first_size and next_size) to use later in
2682 	 * syncing context to determine if entries were added to these sublists
2683 	 * while in open context. This is possible because the clone is still
2684 	 * active and open for normal writes and we want to make sure the new,
2685 	 * unprocessed blockpointers are inserted into the livelist normally.
2686 	 *
2687 	 * Note that dsl_process_sub_livelist() both stores the size number of
2688 	 * blockpointers and iterates over them while the bpobj's lock held, so
2689 	 * the sizes returned to us are consistent which what was actually
2690 	 * processed.
2691 	 */
2692 	int err = dsl_process_sub_livelist(&first->dle_bpobj, &lca->to_keep, t,
2693 	    &first_size);
2694 	if (err == 0)
2695 		err = dsl_process_sub_livelist(&next->dle_bpobj, &lca->to_keep,
2696 		    t, &next_size);
2697 
2698 	if (err == 0) {
2699 		while (zfs_livelist_condense_sync_pause &&
2700 		    !(zthr_has_waiters(t) || zthr_iscancelled(t)))
2701 			delay(1);
2702 
2703 		dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
2704 		dmu_tx_mark_netfree(tx);
2705 		dmu_tx_hold_space(tx, 1);
2706 		err = dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE);
2707 		if (err == 0) {
2708 			/*
2709 			 * Prevent the condense zthr restarting before
2710 			 * the synctask completes.
2711 			 */
2712 			spa->spa_to_condense.syncing = B_TRUE;
2713 			lca->spa = spa;
2714 			lca->first_size = first_size;
2715 			lca->next_size = next_size;
2716 			dsl_sync_task_nowait(spa_get_dsl(spa),
2717 			    spa_livelist_condense_sync, lca, tx);
2718 			dmu_tx_commit(tx);
2719 			return;
2720 		}
2721 	}
2722 	/*
2723 	 * Condensing can not continue: either it was externally stopped or
2724 	 * we were unable to assign to a tx because the pool has run out of
2725 	 * space. In the second case, we'll just end up trying to condense
2726 	 * again in a later txg.
2727 	 */
2728 	ASSERT(err != 0);
2729 	bplist_clear(&lca->to_keep);
2730 	bplist_destroy(&lca->to_keep);
2731 	kmem_free(lca, sizeof (livelist_condense_arg_t));
2732 	dmu_buf_rele(spa->spa_to_condense.ds->ds_dbuf, spa);
2733 	spa->spa_to_condense.ds = NULL;
2734 	if (err == EINTR)
2735 		zfs_livelist_condense_zthr_cancel++;
2736 }
2737 
2738 /* ARGSUSED */
2739 /*
2740  * Check that there is something to condense but that a condense is not
2741  * already in progress and that condensing has not been cancelled.
2742  */
2743 static boolean_t
2744 spa_livelist_condense_cb_check(void *arg, zthr_t *z)
2745 {
2746 	spa_t *spa = arg;
2747 	if ((spa->spa_to_condense.ds != NULL) &&
2748 	    (spa->spa_to_condense.syncing == B_FALSE) &&
2749 	    (spa->spa_to_condense.cancelled == B_FALSE)) {
2750 		return (B_TRUE);
2751 	}
2752 	return (B_FALSE);
2753 }
2754 
2755 static void
2756 spa_start_livelist_condensing_thread(spa_t *spa)
2757 {
2758 	spa->spa_to_condense.ds = NULL;
2759 	spa->spa_to_condense.first = NULL;
2760 	spa->spa_to_condense.next = NULL;
2761 	spa->spa_to_condense.syncing = B_FALSE;
2762 	spa->spa_to_condense.cancelled = B_FALSE;
2763 
2764 	ASSERT3P(spa->spa_livelist_condense_zthr, ==, NULL);
2765 	spa->spa_livelist_condense_zthr =
2766 	    zthr_create("z_livelist_condense",
2767 	    spa_livelist_condense_cb_check,
2768 	    spa_livelist_condense_cb, spa);
2769 }
2770 
2771 static void
2772 spa_spawn_aux_threads(spa_t *spa)
2773 {
2774 	ASSERT(spa_writeable(spa));
2775 
2776 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
2777 
2778 	spa_start_indirect_condensing_thread(spa);
2779 	spa_start_livelist_destroy_thread(spa);
2780 	spa_start_livelist_condensing_thread(spa);
2781 
2782 	ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL);
2783 	spa->spa_checkpoint_discard_zthr =
2784 	    zthr_create("z_checkpoint_discard",
2785 	    spa_checkpoint_discard_thread_check,
2786 	    spa_checkpoint_discard_thread, spa);
2787 }
2788 
2789 /*
2790  * Fix up config after a partly-completed split.  This is done with the
2791  * ZPOOL_CONFIG_SPLIT nvlist.  Both the splitting pool and the split-off
2792  * pool have that entry in their config, but only the splitting one contains
2793  * a list of all the guids of the vdevs that are being split off.
2794  *
2795  * This function determines what to do with that list: either rejoin
2796  * all the disks to the pool, or complete the splitting process.  To attempt
2797  * the rejoin, each disk that is offlined is marked online again, and
2798  * we do a reopen() call.  If the vdev label for every disk that was
2799  * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL)
2800  * then we call vdev_split() on each disk, and complete the split.
2801  *
2802  * Otherwise we leave the config alone, with all the vdevs in place in
2803  * the original pool.
2804  */
2805 static void
2806 spa_try_repair(spa_t *spa, nvlist_t *config)
2807 {
2808 	uint_t extracted;
2809 	uint64_t *glist;
2810 	uint_t i, gcount;
2811 	nvlist_t *nvl;
2812 	vdev_t **vd;
2813 	boolean_t attempt_reopen;
2814 
2815 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0)
2816 		return;
2817 
2818 	/* check that the config is complete */
2819 	if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST,
2820 	    &glist, &gcount) != 0)
2821 		return;
2822 
2823 	vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP);
2824 
2825 	/* attempt to online all the vdevs & validate */
2826 	attempt_reopen = B_TRUE;
2827 	for (i = 0; i < gcount; i++) {
2828 		if (glist[i] == 0)	/* vdev is hole */
2829 			continue;
2830 
2831 		vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE);
2832 		if (vd[i] == NULL) {
2833 			/*
2834 			 * Don't bother attempting to reopen the disks;
2835 			 * just do the split.
2836 			 */
2837 			attempt_reopen = B_FALSE;
2838 		} else {
2839 			/* attempt to re-online it */
2840 			vd[i]->vdev_offline = B_FALSE;
2841 		}
2842 	}
2843 
2844 	if (attempt_reopen) {
2845 		vdev_reopen(spa->spa_root_vdev);
2846 
2847 		/* check each device to see what state it's in */
2848 		for (extracted = 0, i = 0; i < gcount; i++) {
2849 			if (vd[i] != NULL &&
2850 			    vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL)
2851 				break;
2852 			++extracted;
2853 		}
2854 	}
2855 
2856 	/*
2857 	 * If every disk has been moved to the new pool, or if we never
2858 	 * even attempted to look at them, then we split them off for
2859 	 * good.
2860 	 */
2861 	if (!attempt_reopen || gcount == extracted) {
2862 		for (i = 0; i < gcount; i++)
2863 			if (vd[i] != NULL)
2864 				vdev_split(vd[i]);
2865 		vdev_reopen(spa->spa_root_vdev);
2866 	}
2867 
2868 	kmem_free(vd, gcount * sizeof (vdev_t *));
2869 }
2870 
2871 static int
2872 spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type)
2873 {
2874 	char *ereport = FM_EREPORT_ZFS_POOL;
2875 	int error;
2876 
2877 	spa->spa_load_state = state;
2878 	(void) spa_import_progress_set_state(spa_guid(spa),
2879 	    spa_load_state(spa));
2880 
2881 	gethrestime(&spa->spa_loaded_ts);
2882 	error = spa_load_impl(spa, type, &ereport);
2883 
2884 	/*
2885 	 * Don't count references from objsets that are already closed
2886 	 * and are making their way through the eviction process.
2887 	 */
2888 	spa_evicting_os_wait(spa);
2889 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
2890 	if (error) {
2891 		if (error != EEXIST) {
2892 			spa->spa_loaded_ts.tv_sec = 0;
2893 			spa->spa_loaded_ts.tv_nsec = 0;
2894 		}
2895 		if (error != EBADF) {
2896 			(void) zfs_ereport_post(ereport, spa,
2897 			    NULL, NULL, NULL, 0);
2898 		}
2899 	}
2900 	spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE;
2901 	spa->spa_ena = 0;
2902 
2903 	(void) spa_import_progress_set_state(spa_guid(spa),
2904 	    spa_load_state(spa));
2905 
2906 	return (error);
2907 }
2908 
2909 #ifdef ZFS_DEBUG
2910 /*
2911  * Count the number of per-vdev ZAPs associated with all of the vdevs in the
2912  * vdev tree rooted in the given vd, and ensure that each ZAP is present in the
2913  * spa's per-vdev ZAP list.
2914  */
2915 static uint64_t
2916 vdev_count_verify_zaps(vdev_t *vd)
2917 {
2918 	spa_t *spa = vd->vdev_spa;
2919 	uint64_t total = 0;
2920 
2921 	if (vd->vdev_top_zap != 0) {
2922 		total++;
2923 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
2924 		    spa->spa_all_vdev_zaps, vd->vdev_top_zap));
2925 	}
2926 	if (vd->vdev_leaf_zap != 0) {
2927 		total++;
2928 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
2929 		    spa->spa_all_vdev_zaps, vd->vdev_leaf_zap));
2930 	}
2931 
2932 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2933 		total += vdev_count_verify_zaps(vd->vdev_child[i]);
2934 	}
2935 
2936 	return (total);
2937 }
2938 #endif
2939 
2940 /*
2941  * Determine whether the activity check is required.
2942  */
2943 static boolean_t
2944 spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
2945     nvlist_t *config)
2946 {
2947 	uint64_t state = 0;
2948 	uint64_t hostid = 0;
2949 	uint64_t tryconfig_txg = 0;
2950 	uint64_t tryconfig_timestamp = 0;
2951 	uint16_t tryconfig_mmp_seq = 0;
2952 	nvlist_t *nvinfo;
2953 
2954 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
2955 		nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO);
2956 		(void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG,
2957 		    &tryconfig_txg);
2958 		(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
2959 		    &tryconfig_timestamp);
2960 		(void) nvlist_lookup_uint16(nvinfo, ZPOOL_CONFIG_MMP_SEQ,
2961 		    &tryconfig_mmp_seq);
2962 	}
2963 
2964 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state);
2965 
2966 	/*
2967 	 * Disable the MMP activity check - This is used by zdb which
2968 	 * is intended to be used on potentially active pools.
2969 	 */
2970 	if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP)
2971 		return (B_FALSE);
2972 
2973 	/*
2974 	 * Skip the activity check when the MMP feature is disabled.
2975 	 */
2976 	if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0)
2977 		return (B_FALSE);
2978 
2979 	/*
2980 	 * If the tryconfig_ values are nonzero, they are the results of an
2981 	 * earlier tryimport.  If they all match the uberblock we just found,
2982 	 * then the pool has not changed and we return false so we do not test
2983 	 * a second time.
2984 	 */
2985 	if (tryconfig_txg && tryconfig_txg == ub->ub_txg &&
2986 	    tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp &&
2987 	    tryconfig_mmp_seq && tryconfig_mmp_seq ==
2988 	    (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0))
2989 		return (B_FALSE);
2990 
2991 	/*
2992 	 * Allow the activity check to be skipped when importing the pool
2993 	 * on the same host which last imported it.  Since the hostid from
2994 	 * configuration may be stale use the one read from the label.
2995 	 */
2996 	if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID))
2997 		hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID);
2998 
2999 	if (hostid == spa_get_hostid(spa))
3000 		return (B_FALSE);
3001 
3002 	/*
3003 	 * Skip the activity test when the pool was cleanly exported.
3004 	 */
3005 	if (state != POOL_STATE_ACTIVE)
3006 		return (B_FALSE);
3007 
3008 	return (B_TRUE);
3009 }
3010 
3011 /*
3012  * Nanoseconds the activity check must watch for changes on-disk.
3013  */
3014 static uint64_t
3015 spa_activity_check_duration(spa_t *spa, uberblock_t *ub)
3016 {
3017 	uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1);
3018 	uint64_t multihost_interval = MSEC2NSEC(
3019 	    MMP_INTERVAL_OK(zfs_multihost_interval));
3020 	uint64_t import_delay = MAX(NANOSEC, import_intervals *
3021 	    multihost_interval);
3022 
3023 	/*
3024 	 * Local tunables determine a minimum duration except for the case
3025 	 * where we know when the remote host will suspend the pool if MMP
3026 	 * writes do not land.
3027 	 *
3028 	 * See Big Theory comment at the top of mmp.c for the reasoning behind
3029 	 * these cases and times.
3030 	 */
3031 
3032 	ASSERT(MMP_IMPORT_SAFETY_FACTOR >= 100);
3033 
3034 	if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
3035 	    MMP_FAIL_INT(ub) > 0) {
3036 
3037 		/* MMP on remote host will suspend pool after failed writes */
3038 		import_delay = MMP_FAIL_INT(ub) * MSEC2NSEC(MMP_INTERVAL(ub)) *
3039 		    MMP_IMPORT_SAFETY_FACTOR / 100;
3040 
3041 		zfs_dbgmsg("fail_intvals>0 import_delay=%llu ub_mmp "
3042 		    "mmp_fails=%llu ub_mmp mmp_interval=%llu "
3043 		    "import_intervals=%u", import_delay, MMP_FAIL_INT(ub),
3044 		    MMP_INTERVAL(ub), import_intervals);
3045 
3046 	} else if (MMP_INTERVAL_VALID(ub) && MMP_FAIL_INT_VALID(ub) &&
3047 	    MMP_FAIL_INT(ub) == 0) {
3048 
3049 		/* MMP on remote host will never suspend pool */
3050 		import_delay = MAX(import_delay, (MSEC2NSEC(MMP_INTERVAL(ub)) +
3051 		    ub->ub_mmp_delay) * import_intervals);
3052 
3053 		zfs_dbgmsg("fail_intvals=0 import_delay=%llu ub_mmp "
3054 		    "mmp_interval=%llu ub_mmp_delay=%llu "
3055 		    "import_intervals=%u", import_delay, MMP_INTERVAL(ub),
3056 		    ub->ub_mmp_delay, import_intervals);
3057 
3058 	} else if (MMP_VALID(ub)) {
3059 		/*
3060 		 * zfs-0.7 compatibility case
3061 		 */
3062 
3063 		import_delay = MAX(import_delay, (multihost_interval +
3064 		    ub->ub_mmp_delay) * import_intervals);
3065 
3066 		zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu "
3067 		    "import_intervals=%u leaves=%u", import_delay,
3068 		    ub->ub_mmp_delay, import_intervals,
3069 		    vdev_count_leaves(spa));
3070 	} else {
3071 		/* Using local tunings is the only reasonable option */
3072 		zfs_dbgmsg("pool last imported on non-MMP aware "
3073 		    "host using import_delay=%llu multihost_interval=%llu "
3074 		    "import_intervals=%u", import_delay, multihost_interval,
3075 		    import_intervals);
3076 	}
3077 
3078 	return (import_delay);
3079 }
3080 
3081 /*
3082  * Perform the import activity check.  If the user canceled the import or
3083  * we detected activity then fail.
3084  */
3085 static int
3086 spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
3087 {
3088 	uint64_t txg = ub->ub_txg;
3089 	uint64_t timestamp = ub->ub_timestamp;
3090 	uint64_t mmp_config = ub->ub_mmp_config;
3091 	uint16_t mmp_seq = MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0;
3092 	uint64_t import_delay;
3093 	hrtime_t import_expire;
3094 	nvlist_t *mmp_label = NULL;
3095 	vdev_t *rvd = spa->spa_root_vdev;
3096 	kcondvar_t cv;
3097 	kmutex_t mtx;
3098 	int error = 0;
3099 
3100 	cv_init(&cv, NULL, CV_DEFAULT, NULL);
3101 	mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL);
3102 	mutex_enter(&mtx);
3103 
3104 	/*
3105 	 * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed
3106 	 * during the earlier tryimport.  If the txg recorded there is 0 then
3107 	 * the pool is known to be active on another host.
3108 	 *
3109 	 * Otherwise, the pool might be in use on another host.  Check for
3110 	 * changes in the uberblocks on disk if necessary.
3111 	 */
3112 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
3113 		nvlist_t *nvinfo = fnvlist_lookup_nvlist(config,
3114 		    ZPOOL_CONFIG_LOAD_INFO);
3115 
3116 		if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) &&
3117 		    fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) {
3118 			vdev_uberblock_load(rvd, ub, &mmp_label);
3119 			error = SET_ERROR(EREMOTEIO);
3120 			goto out;
3121 		}
3122 	}
3123 
3124 	import_delay = spa_activity_check_duration(spa, ub);
3125 
3126 	/* Add a small random factor in case of simultaneous imports (0-25%) */
3127 	import_delay += import_delay * spa_get_random(250) / 1000;
3128 
3129 	import_expire = gethrtime() + import_delay;
3130 
3131 	while (gethrtime() < import_expire) {
3132 		(void) spa_import_progress_set_mmp_check(spa_guid(spa),
3133 		    NSEC2SEC(import_expire - gethrtime()));
3134 
3135 		vdev_uberblock_load(rvd, ub, &mmp_label);
3136 
3137 		if (txg != ub->ub_txg || timestamp != ub->ub_timestamp ||
3138 		    mmp_seq != (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0)) {
3139 			zfs_dbgmsg("multihost activity detected "
3140 			    "txg %llu ub_txg  %llu "
3141 			    "timestamp %llu ub_timestamp  %llu "
3142 			    "mmp_config %#llx ub_mmp_config %#llx",
3143 			    txg, ub->ub_txg, timestamp, ub->ub_timestamp,
3144 			    mmp_config, ub->ub_mmp_config);
3145 
3146 			error = SET_ERROR(EREMOTEIO);
3147 			break;
3148 		}
3149 
3150 		if (mmp_label) {
3151 			nvlist_free(mmp_label);
3152 			mmp_label = NULL;
3153 		}
3154 
3155 		error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz);
3156 		if (error != -1) {
3157 			error = SET_ERROR(EINTR);
3158 			break;
3159 		}
3160 		error = 0;
3161 	}
3162 
3163 out:
3164 	mutex_exit(&mtx);
3165 	mutex_destroy(&mtx);
3166 	cv_destroy(&cv);
3167 
3168 	/*
3169 	 * If the pool is determined to be active store the status in the
3170 	 * spa->spa_load_info nvlist.  If the remote hostname or hostid are
3171 	 * available from configuration read from disk store them as well.
3172 	 * This allows 'zpool import' to generate a more useful message.
3173 	 *
3174 	 * ZPOOL_CONFIG_MMP_STATE    - observed pool status (mandatory)
3175 	 * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool
3176 	 * ZPOOL_CONFIG_MMP_HOSTID   - hostid from the active pool
3177 	 */
3178 	if (error == EREMOTEIO) {
3179 		char *hostname = "<unknown>";
3180 		uint64_t hostid = 0;
3181 
3182 		if (mmp_label) {
3183 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) {
3184 				hostname = fnvlist_lookup_string(mmp_label,
3185 				    ZPOOL_CONFIG_HOSTNAME);
3186 				fnvlist_add_string(spa->spa_load_info,
3187 				    ZPOOL_CONFIG_MMP_HOSTNAME, hostname);
3188 			}
3189 
3190 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) {
3191 				hostid = fnvlist_lookup_uint64(mmp_label,
3192 				    ZPOOL_CONFIG_HOSTID);
3193 				fnvlist_add_uint64(spa->spa_load_info,
3194 				    ZPOOL_CONFIG_MMP_HOSTID, hostid);
3195 			}
3196 		}
3197 
3198 		fnvlist_add_uint64(spa->spa_load_info,
3199 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE);
3200 		fnvlist_add_uint64(spa->spa_load_info,
3201 		    ZPOOL_CONFIG_MMP_TXG, 0);
3202 
3203 		error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO);
3204 	}
3205 
3206 	if (mmp_label)
3207 		nvlist_free(mmp_label);
3208 
3209 	return (error);
3210 }
3211 
3212 static int
3213 spa_verify_host(spa_t *spa, nvlist_t *mos_config)
3214 {
3215 	uint64_t hostid;
3216 	char *hostname;
3217 	uint64_t myhostid = 0;
3218 
3219 	if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config,
3220 	    ZPOOL_CONFIG_HOSTID, &hostid) == 0) {
3221 		hostname = fnvlist_lookup_string(mos_config,
3222 		    ZPOOL_CONFIG_HOSTNAME);
3223 
3224 		myhostid = zone_get_hostid(NULL);
3225 
3226 		if (hostid != 0 && myhostid != 0 && hostid != myhostid) {
3227 			cmn_err(CE_WARN, "pool '%s' could not be "
3228 			    "loaded as it was last accessed by "
3229 			    "another system (host: %s hostid: 0x%llx). "
3230 			    "See: https://openzfs.github.io/openzfs-docs/msg/"
3231 			    "ZFS-8000-EY",
3232 			    spa_name(spa), hostname, (u_longlong_t)hostid);
3233 			spa_load_failed(spa, "hostid verification failed: pool "
3234 			    "last accessed by host: %s (hostid: 0x%llx)",
3235 			    hostname, (u_longlong_t)hostid);
3236 			return (SET_ERROR(EBADF));
3237 		}
3238 	}
3239 
3240 	return (0);
3241 }
3242 
3243 static int
3244 spa_ld_parse_config(spa_t *spa, spa_import_type_t type)
3245 {
3246 	int error = 0;
3247 	nvlist_t *nvtree, *nvl, *config = spa->spa_config;
3248 	int parse;
3249 	vdev_t *rvd;
3250 	uint64_t pool_guid;
3251 	char *comment;
3252 
3253 	/*
3254 	 * Versioning wasn't explicitly added to the label until later, so if
3255 	 * it's not present treat it as the initial version.
3256 	 */
3257 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION,
3258 	    &spa->spa_ubsync.ub_version) != 0)
3259 		spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL;
3260 
3261 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) {
3262 		spa_load_failed(spa, "invalid config provided: '%s' missing",
3263 		    ZPOOL_CONFIG_POOL_GUID);
3264 		return (SET_ERROR(EINVAL));
3265 	}
3266 
3267 	/*
3268 	 * If we are doing an import, ensure that the pool is not already
3269 	 * imported by checking if its pool guid already exists in the
3270 	 * spa namespace.
3271 	 *
3272 	 * The only case that we allow an already imported pool to be
3273 	 * imported again, is when the pool is checkpointed and we want to
3274 	 * look at its checkpointed state from userland tools like zdb.
3275 	 */
3276 #ifdef _KERNEL
3277 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
3278 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
3279 	    spa_guid_exists(pool_guid, 0)) {
3280 #else
3281 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
3282 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
3283 	    spa_guid_exists(pool_guid, 0) &&
3284 	    !spa_importing_readonly_checkpoint(spa)) {
3285 #endif
3286 		spa_load_failed(spa, "a pool with guid %llu is already open",
3287 		    (u_longlong_t)pool_guid);
3288 		return (SET_ERROR(EEXIST));
3289 	}
3290 
3291 	spa->spa_config_guid = pool_guid;
3292 
3293 	nvlist_free(spa->spa_load_info);
3294 	spa->spa_load_info = fnvlist_alloc();
3295 
3296 	ASSERT(spa->spa_comment == NULL);
3297 	if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0)
3298 		spa->spa_comment = spa_strdup(comment);
3299 
3300 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
3301 	    &spa->spa_config_txg);
3302 
3303 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0)
3304 		spa->spa_config_splitting = fnvlist_dup(nvl);
3305 
3306 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) {
3307 		spa_load_failed(spa, "invalid config provided: '%s' missing",
3308 		    ZPOOL_CONFIG_VDEV_TREE);
3309 		return (SET_ERROR(EINVAL));
3310 	}
3311 
3312 	/*
3313 	 * Create "The Godfather" zio to hold all async IOs
3314 	 */
3315 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
3316 	    KM_SLEEP);
3317 	for (int i = 0; i < max_ncpus; i++) {
3318 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
3319 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
3320 		    ZIO_FLAG_GODFATHER);
3321 	}
3322 
3323 	/*
3324 	 * Parse the configuration into a vdev tree.  We explicitly set the
3325 	 * value that will be returned by spa_version() since parsing the
3326 	 * configuration requires knowing the version number.
3327 	 */
3328 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3329 	parse = (type == SPA_IMPORT_EXISTING ?
3330 	    VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT);
3331 	error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse);
3332 	spa_config_exit(spa, SCL_ALL, FTAG);
3333 
3334 	if (error != 0) {
3335 		spa_load_failed(spa, "unable to parse config [error=%d]",
3336 		    error);
3337 		return (error);
3338 	}
3339 
3340 	ASSERT(spa->spa_root_vdev == rvd);
3341 	ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
3342 	ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT);
3343 
3344 	if (type != SPA_IMPORT_ASSEMBLE) {
3345 		ASSERT(spa_guid(spa) == pool_guid);
3346 	}
3347 
3348 	return (0);
3349 }
3350 
3351 /*
3352  * Recursively open all vdevs in the vdev tree. This function is called twice:
3353  * first with the untrusted config, then with the trusted config.
3354  */
3355 static int
3356 spa_ld_open_vdevs(spa_t *spa)
3357 {
3358 	int error = 0;
3359 
3360 	/*
3361 	 * spa_missing_tvds_allowed defines how many top-level vdevs can be
3362 	 * missing/unopenable for the root vdev to be still considered openable.
3363 	 */
3364 	if (spa->spa_trust_config) {
3365 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds;
3366 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) {
3367 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile;
3368 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) {
3369 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan;
3370 	} else {
3371 		spa->spa_missing_tvds_allowed = 0;
3372 	}
3373 
3374 	spa->spa_missing_tvds_allowed =
3375 	    MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed);
3376 
3377 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3378 	error = vdev_open(spa->spa_root_vdev);
3379 	spa_config_exit(spa, SCL_ALL, FTAG);
3380 
3381 	if (spa->spa_missing_tvds != 0) {
3382 		spa_load_note(spa, "vdev tree has %lld missing top-level "
3383 		    "vdevs.", (u_longlong_t)spa->spa_missing_tvds);
3384 		if (spa->spa_trust_config && (spa->spa_mode & SPA_MODE_WRITE)) {
3385 			/*
3386 			 * Although theoretically we could allow users to open
3387 			 * incomplete pools in RW mode, we'd need to add a lot
3388 			 * of extra logic (e.g. adjust pool space to account
3389 			 * for missing vdevs).
3390 			 * This limitation also prevents users from accidentally
3391 			 * opening the pool in RW mode during data recovery and
3392 			 * damaging it further.
3393 			 */
3394 			spa_load_note(spa, "pools with missing top-level "
3395 			    "vdevs can only be opened in read-only mode.");
3396 			error = SET_ERROR(ENXIO);
3397 		} else {
3398 			spa_load_note(spa, "current settings allow for maximum "
3399 			    "%lld missing top-level vdevs at this stage.",
3400 			    (u_longlong_t)spa->spa_missing_tvds_allowed);
3401 		}
3402 	}
3403 	if (error != 0) {
3404 		spa_load_failed(spa, "unable to open vdev tree [error=%d]",
3405 		    error);
3406 	}
3407 	if (spa->spa_missing_tvds != 0 || error != 0)
3408 		vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2);
3409 
3410 	return (error);
3411 }
3412 
3413 /*
3414  * We need to validate the vdev labels against the configuration that
3415  * we have in hand. This function is called twice: first with an untrusted
3416  * config, then with a trusted config. The validation is more strict when the
3417  * config is trusted.
3418  */
3419 static int
3420 spa_ld_validate_vdevs(spa_t *spa)
3421 {
3422 	int error = 0;
3423 	vdev_t *rvd = spa->spa_root_vdev;
3424 
3425 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3426 	error = vdev_validate(rvd);
3427 	spa_config_exit(spa, SCL_ALL, FTAG);
3428 
3429 	if (error != 0) {
3430 		spa_load_failed(spa, "vdev_validate failed [error=%d]", error);
3431 		return (error);
3432 	}
3433 
3434 	if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) {
3435 		spa_load_failed(spa, "cannot open vdev tree after invalidating "
3436 		    "some vdevs");
3437 		vdev_dbgmsg_print_tree(rvd, 2);
3438 		return (SET_ERROR(ENXIO));
3439 	}
3440 
3441 	return (0);
3442 }
3443 
3444 static void
3445 spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub)
3446 {
3447 	spa->spa_state = POOL_STATE_ACTIVE;
3448 	spa->spa_ubsync = spa->spa_uberblock;
3449 	spa->spa_verify_min_txg = spa->spa_extreme_rewind ?
3450 	    TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1;
3451 	spa->spa_first_txg = spa->spa_last_ubsync_txg ?
3452 	    spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1;
3453 	spa->spa_claim_max_txg = spa->spa_first_txg;
3454 	spa->spa_prev_software_version = ub->ub_software_version;
3455 }
3456 
3457 static int
3458 spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type)
3459 {
3460 	vdev_t *rvd = spa->spa_root_vdev;
3461 	nvlist_t *label;
3462 	uberblock_t *ub = &spa->spa_uberblock;
3463 	boolean_t activity_check = B_FALSE;
3464 
3465 	/*
3466 	 * If we are opening the checkpointed state of the pool by
3467 	 * rewinding to it, at this point we will have written the
3468 	 * checkpointed uberblock to the vdev labels, so searching
3469 	 * the labels will find the right uberblock.  However, if
3470 	 * we are opening the checkpointed state read-only, we have
3471 	 * not modified the labels. Therefore, we must ignore the
3472 	 * labels and continue using the spa_uberblock that was set
3473 	 * by spa_ld_checkpoint_rewind.
3474 	 *
3475 	 * Note that it would be fine to ignore the labels when
3476 	 * rewinding (opening writeable) as well. However, if we
3477 	 * crash just after writing the labels, we will end up
3478 	 * searching the labels. Doing so in the common case means
3479 	 * that this code path gets exercised normally, rather than
3480 	 * just in the edge case.
3481 	 */
3482 	if (ub->ub_checkpoint_txg != 0 &&
3483 	    spa_importing_readonly_checkpoint(spa)) {
3484 		spa_ld_select_uberblock_done(spa, ub);
3485 		return (0);
3486 	}
3487 
3488 	/*
3489 	 * Find the best uberblock.
3490 	 */
3491 	vdev_uberblock_load(rvd, ub, &label);
3492 
3493 	/*
3494 	 * If we weren't able to find a single valid uberblock, return failure.
3495 	 */
3496 	if (ub->ub_txg == 0) {
3497 		nvlist_free(label);
3498 		spa_load_failed(spa, "no valid uberblock found");
3499 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO));
3500 	}
3501 
3502 	if (spa->spa_load_max_txg != UINT64_MAX) {
3503 		(void) spa_import_progress_set_max_txg(spa_guid(spa),
3504 		    (u_longlong_t)spa->spa_load_max_txg);
3505 	}
3506 	spa_load_note(spa, "using uberblock with txg=%llu",
3507 	    (u_longlong_t)ub->ub_txg);
3508 
3509 
3510 	/*
3511 	 * For pools which have the multihost property on determine if the
3512 	 * pool is truly inactive and can be safely imported.  Prevent
3513 	 * hosts which don't have a hostid set from importing the pool.
3514 	 */
3515 	activity_check = spa_activity_check_required(spa, ub, label,
3516 	    spa->spa_config);
3517 	if (activity_check) {
3518 		if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay &&
3519 		    spa_get_hostid(spa) == 0) {
3520 			nvlist_free(label);
3521 			fnvlist_add_uint64(spa->spa_load_info,
3522 			    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
3523 			return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
3524 		}
3525 
3526 		int error = spa_activity_check(spa, ub, spa->spa_config);
3527 		if (error) {
3528 			nvlist_free(label);
3529 			return (error);
3530 		}
3531 
3532 		fnvlist_add_uint64(spa->spa_load_info,
3533 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE);
3534 		fnvlist_add_uint64(spa->spa_load_info,
3535 		    ZPOOL_CONFIG_MMP_TXG, ub->ub_txg);
3536 		fnvlist_add_uint16(spa->spa_load_info,
3537 		    ZPOOL_CONFIG_MMP_SEQ,
3538 		    (MMP_SEQ_VALID(ub) ? MMP_SEQ(ub) : 0));
3539 	}
3540 
3541 	/*
3542 	 * If the pool has an unsupported version we can't open it.
3543 	 */
3544 	if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) {
3545 		nvlist_free(label);
3546 		spa_load_failed(spa, "version %llu is not supported",
3547 		    (u_longlong_t)ub->ub_version);
3548 		return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP));
3549 	}
3550 
3551 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
3552 		nvlist_t *features;
3553 
3554 		/*
3555 		 * If we weren't able to find what's necessary for reading the
3556 		 * MOS in the label, return failure.
3557 		 */
3558 		if (label == NULL) {
3559 			spa_load_failed(spa, "label config unavailable");
3560 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3561 			    ENXIO));
3562 		}
3563 
3564 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ,
3565 		    &features) != 0) {
3566 			nvlist_free(label);
3567 			spa_load_failed(spa, "invalid label: '%s' missing",
3568 			    ZPOOL_CONFIG_FEATURES_FOR_READ);
3569 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3570 			    ENXIO));
3571 		}
3572 
3573 		/*
3574 		 * Update our in-core representation with the definitive values
3575 		 * from the label.
3576 		 */
3577 		nvlist_free(spa->spa_label_features);
3578 		VERIFY(nvlist_dup(features, &spa->spa_label_features, 0) == 0);
3579 	}
3580 
3581 	nvlist_free(label);
3582 
3583 	/*
3584 	 * Look through entries in the label nvlist's features_for_read. If
3585 	 * there is a feature listed there which we don't understand then we
3586 	 * cannot open a pool.
3587 	 */
3588 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
3589 		nvlist_t *unsup_feat;
3590 
3591 		VERIFY(nvlist_alloc(&unsup_feat, NV_UNIQUE_NAME, KM_SLEEP) ==
3592 		    0);
3593 
3594 		for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features,
3595 		    NULL); nvp != NULL;
3596 		    nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) {
3597 			if (!zfeature_is_supported(nvpair_name(nvp))) {
3598 				VERIFY(nvlist_add_string(unsup_feat,
3599 				    nvpair_name(nvp), "") == 0);
3600 			}
3601 		}
3602 
3603 		if (!nvlist_empty(unsup_feat)) {
3604 			VERIFY(nvlist_add_nvlist(spa->spa_load_info,
3605 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat) == 0);
3606 			nvlist_free(unsup_feat);
3607 			spa_load_failed(spa, "some features are unsupported");
3608 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
3609 			    ENOTSUP));
3610 		}
3611 
3612 		nvlist_free(unsup_feat);
3613 	}
3614 
3615 	if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) {
3616 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3617 		spa_try_repair(spa, spa->spa_config);
3618 		spa_config_exit(spa, SCL_ALL, FTAG);
3619 		nvlist_free(spa->spa_config_splitting);
3620 		spa->spa_config_splitting = NULL;
3621 	}
3622 
3623 	/*
3624 	 * Initialize internal SPA structures.
3625 	 */
3626 	spa_ld_select_uberblock_done(spa, ub);
3627 
3628 	return (0);
3629 }
3630 
3631 static int
3632 spa_ld_open_rootbp(spa_t *spa)
3633 {
3634 	int error = 0;
3635 	vdev_t *rvd = spa->spa_root_vdev;
3636 
3637 	error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool);
3638 	if (error != 0) {
3639 		spa_load_failed(spa, "unable to open rootbp in dsl_pool_init "
3640 		    "[error=%d]", error);
3641 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3642 	}
3643 	spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset;
3644 
3645 	return (0);
3646 }
3647 
3648 static int
3649 spa_ld_trusted_config(spa_t *spa, spa_import_type_t type,
3650     boolean_t reloading)
3651 {
3652 	vdev_t *mrvd, *rvd = spa->spa_root_vdev;
3653 	nvlist_t *nv, *mos_config, *policy;
3654 	int error = 0, copy_error;
3655 	uint64_t healthy_tvds, healthy_tvds_mos;
3656 	uint64_t mos_config_txg;
3657 
3658 	if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE)
3659 	    != 0)
3660 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3661 
3662 	/*
3663 	 * If we're assembling a pool from a split, the config provided is
3664 	 * already trusted so there is nothing to do.
3665 	 */
3666 	if (type == SPA_IMPORT_ASSEMBLE)
3667 		return (0);
3668 
3669 	healthy_tvds = spa_healthy_core_tvds(spa);
3670 
3671 	if (load_nvlist(spa, spa->spa_config_object, &mos_config)
3672 	    != 0) {
3673 		spa_load_failed(spa, "unable to retrieve MOS config");
3674 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3675 	}
3676 
3677 	/*
3678 	 * If we are doing an open, pool owner wasn't verified yet, thus do
3679 	 * the verification here.
3680 	 */
3681 	if (spa->spa_load_state == SPA_LOAD_OPEN) {
3682 		error = spa_verify_host(spa, mos_config);
3683 		if (error != 0) {
3684 			nvlist_free(mos_config);
3685 			return (error);
3686 		}
3687 	}
3688 
3689 	nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE);
3690 
3691 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3692 
3693 	/*
3694 	 * Build a new vdev tree from the trusted config
3695 	 */
3696 	error = spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD);
3697 	if (error != 0) {
3698 		nvlist_free(mos_config);
3699 		spa_config_exit(spa, SCL_ALL, FTAG);
3700 		spa_load_failed(spa, "spa_config_parse failed [error=%d]",
3701 		    error);
3702 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
3703 	}
3704 
3705 	/*
3706 	 * Vdev paths in the MOS may be obsolete. If the untrusted config was
3707 	 * obtained by scanning /dev/dsk, then it will have the right vdev
3708 	 * paths. We update the trusted MOS config with this information.
3709 	 * We first try to copy the paths with vdev_copy_path_strict, which
3710 	 * succeeds only when both configs have exactly the same vdev tree.
3711 	 * If that fails, we fall back to a more flexible method that has a
3712 	 * best effort policy.
3713 	 */
3714 	copy_error = vdev_copy_path_strict(rvd, mrvd);
3715 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3716 		spa_load_note(spa, "provided vdev tree:");
3717 		vdev_dbgmsg_print_tree(rvd, 2);
3718 		spa_load_note(spa, "MOS vdev tree:");
3719 		vdev_dbgmsg_print_tree(mrvd, 2);
3720 	}
3721 	if (copy_error != 0) {
3722 		spa_load_note(spa, "vdev_copy_path_strict failed, falling "
3723 		    "back to vdev_copy_path_relaxed");
3724 		vdev_copy_path_relaxed(rvd, mrvd);
3725 	}
3726 
3727 	vdev_close(rvd);
3728 	vdev_free(rvd);
3729 	spa->spa_root_vdev = mrvd;
3730 	rvd = mrvd;
3731 	spa_config_exit(spa, SCL_ALL, FTAG);
3732 
3733 	/*
3734 	 * We will use spa_config if we decide to reload the spa or if spa_load
3735 	 * fails and we rewind. We must thus regenerate the config using the
3736 	 * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to
3737 	 * pass settings on how to load the pool and is not stored in the MOS.
3738 	 * We copy it over to our new, trusted config.
3739 	 */
3740 	mos_config_txg = fnvlist_lookup_uint64(mos_config,
3741 	    ZPOOL_CONFIG_POOL_TXG);
3742 	nvlist_free(mos_config);
3743 	mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE);
3744 	if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY,
3745 	    &policy) == 0)
3746 		fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy);
3747 	spa_config_set(spa, mos_config);
3748 	spa->spa_config_source = SPA_CONFIG_SRC_MOS;
3749 
3750 	/*
3751 	 * Now that we got the config from the MOS, we should be more strict
3752 	 * in checking blkptrs and can make assumptions about the consistency
3753 	 * of the vdev tree. spa_trust_config must be set to true before opening
3754 	 * vdevs in order for them to be writeable.
3755 	 */
3756 	spa->spa_trust_config = B_TRUE;
3757 
3758 	/*
3759 	 * Open and validate the new vdev tree
3760 	 */
3761 	error = spa_ld_open_vdevs(spa);
3762 	if (error != 0)
3763 		return (error);
3764 
3765 	error = spa_ld_validate_vdevs(spa);
3766 	if (error != 0)
3767 		return (error);
3768 
3769 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3770 		spa_load_note(spa, "final vdev tree:");
3771 		vdev_dbgmsg_print_tree(rvd, 2);
3772 	}
3773 
3774 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT &&
3775 	    !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) {
3776 		/*
3777 		 * Sanity check to make sure that we are indeed loading the
3778 		 * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds
3779 		 * in the config provided and they happened to be the only ones
3780 		 * to have the latest uberblock, we could involuntarily perform
3781 		 * an extreme rewind.
3782 		 */
3783 		healthy_tvds_mos = spa_healthy_core_tvds(spa);
3784 		if (healthy_tvds_mos - healthy_tvds >=
3785 		    SPA_SYNC_MIN_VDEVS) {
3786 			spa_load_note(spa, "config provided misses too many "
3787 			    "top-level vdevs compared to MOS (%lld vs %lld). ",
3788 			    (u_longlong_t)healthy_tvds,
3789 			    (u_longlong_t)healthy_tvds_mos);
3790 			spa_load_note(spa, "vdev tree:");
3791 			vdev_dbgmsg_print_tree(rvd, 2);
3792 			if (reloading) {
3793 				spa_load_failed(spa, "config was already "
3794 				    "provided from MOS. Aborting.");
3795 				return (spa_vdev_err(rvd,
3796 				    VDEV_AUX_CORRUPT_DATA, EIO));
3797 			}
3798 			spa_load_note(spa, "spa must be reloaded using MOS "
3799 			    "config");
3800 			return (SET_ERROR(EAGAIN));
3801 		}
3802 	}
3803 
3804 	error = spa_check_for_missing_logs(spa);
3805 	if (error != 0)
3806 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO));
3807 
3808 	if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) {
3809 		spa_load_failed(spa, "uberblock guid sum doesn't match MOS "
3810 		    "guid sum (%llu != %llu)",
3811 		    (u_longlong_t)spa->spa_uberblock.ub_guid_sum,
3812 		    (u_longlong_t)rvd->vdev_guid_sum);
3813 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM,
3814 		    ENXIO));
3815 	}
3816 
3817 	return (0);
3818 }
3819 
3820 static int
3821 spa_ld_open_indirect_vdev_metadata(spa_t *spa)
3822 {
3823 	int error = 0;
3824 	vdev_t *rvd = spa->spa_root_vdev;
3825 
3826 	/*
3827 	 * Everything that we read before spa_remove_init() must be stored
3828 	 * on concreted vdevs.  Therefore we do this as early as possible.
3829 	 */
3830 	error = spa_remove_init(spa);
3831 	if (error != 0) {
3832 		spa_load_failed(spa, "spa_remove_init failed [error=%d]",
3833 		    error);
3834 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3835 	}
3836 
3837 	/*
3838 	 * Retrieve information needed to condense indirect vdev mappings.
3839 	 */
3840 	error = spa_condense_init(spa);
3841 	if (error != 0) {
3842 		spa_load_failed(spa, "spa_condense_init failed [error=%d]",
3843 		    error);
3844 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
3845 	}
3846 
3847 	return (0);
3848 }
3849 
3850 static int
3851 spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep)
3852 {
3853 	int error = 0;
3854 	vdev_t *rvd = spa->spa_root_vdev;
3855 
3856 	if (spa_version(spa) >= SPA_VERSION_FEATURES) {
3857 		boolean_t missing_feat_read = B_FALSE;
3858 		nvlist_t *unsup_feat, *enabled_feat;
3859 
3860 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ,
3861 		    &spa->spa_feat_for_read_obj, B_TRUE) != 0) {
3862 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3863 		}
3864 
3865 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE,
3866 		    &spa->spa_feat_for_write_obj, B_TRUE) != 0) {
3867 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3868 		}
3869 
3870 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS,
3871 		    &spa->spa_feat_desc_obj, B_TRUE) != 0) {
3872 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3873 		}
3874 
3875 		enabled_feat = fnvlist_alloc();
3876 		unsup_feat = fnvlist_alloc();
3877 
3878 		if (!spa_features_check(spa, B_FALSE,
3879 		    unsup_feat, enabled_feat))
3880 			missing_feat_read = B_TRUE;
3881 
3882 		if (spa_writeable(spa) ||
3883 		    spa->spa_load_state == SPA_LOAD_TRYIMPORT) {
3884 			if (!spa_features_check(spa, B_TRUE,
3885 			    unsup_feat, enabled_feat)) {
3886 				*missing_feat_writep = B_TRUE;
3887 			}
3888 		}
3889 
3890 		fnvlist_add_nvlist(spa->spa_load_info,
3891 		    ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat);
3892 
3893 		if (!nvlist_empty(unsup_feat)) {
3894 			fnvlist_add_nvlist(spa->spa_load_info,
3895 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat);
3896 		}
3897 
3898 		fnvlist_free(enabled_feat);
3899 		fnvlist_free(unsup_feat);
3900 
3901 		if (!missing_feat_read) {
3902 			fnvlist_add_boolean(spa->spa_load_info,
3903 			    ZPOOL_CONFIG_CAN_RDONLY);
3904 		}
3905 
3906 		/*
3907 		 * If the state is SPA_LOAD_TRYIMPORT, our objective is
3908 		 * twofold: to determine whether the pool is available for
3909 		 * import in read-write mode and (if it is not) whether the
3910 		 * pool is available for import in read-only mode. If the pool
3911 		 * is available for import in read-write mode, it is displayed
3912 		 * as available in userland; if it is not available for import
3913 		 * in read-only mode, it is displayed as unavailable in
3914 		 * userland. If the pool is available for import in read-only
3915 		 * mode but not read-write mode, it is displayed as unavailable
3916 		 * in userland with a special note that the pool is actually
3917 		 * available for open in read-only mode.
3918 		 *
3919 		 * As a result, if the state is SPA_LOAD_TRYIMPORT and we are
3920 		 * missing a feature for write, we must first determine whether
3921 		 * the pool can be opened read-only before returning to
3922 		 * userland in order to know whether to display the
3923 		 * abovementioned note.
3924 		 */
3925 		if (missing_feat_read || (*missing_feat_writep &&
3926 		    spa_writeable(spa))) {
3927 			spa_load_failed(spa, "pool uses unsupported features");
3928 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
3929 			    ENOTSUP));
3930 		}
3931 
3932 		/*
3933 		 * Load refcounts for ZFS features from disk into an in-memory
3934 		 * cache during SPA initialization.
3935 		 */
3936 		for (spa_feature_t i = 0; i < SPA_FEATURES; i++) {
3937 			uint64_t refcount;
3938 
3939 			error = feature_get_refcount_from_disk(spa,
3940 			    &spa_feature_table[i], &refcount);
3941 			if (error == 0) {
3942 				spa->spa_feat_refcount_cache[i] = refcount;
3943 			} else if (error == ENOTSUP) {
3944 				spa->spa_feat_refcount_cache[i] =
3945 				    SPA_FEATURE_DISABLED;
3946 			} else {
3947 				spa_load_failed(spa, "error getting refcount "
3948 				    "for feature %s [error=%d]",
3949 				    spa_feature_table[i].fi_guid, error);
3950 				return (spa_vdev_err(rvd,
3951 				    VDEV_AUX_CORRUPT_DATA, EIO));
3952 			}
3953 		}
3954 	}
3955 
3956 	if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) {
3957 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG,
3958 		    &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0)
3959 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3960 	}
3961 
3962 	/*
3963 	 * Encryption was added before bookmark_v2, even though bookmark_v2
3964 	 * is now a dependency. If this pool has encryption enabled without
3965 	 * bookmark_v2, trigger an errata message.
3966 	 */
3967 	if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) &&
3968 	    !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
3969 		spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION;
3970 	}
3971 
3972 	return (0);
3973 }
3974 
3975 static int
3976 spa_ld_load_special_directories(spa_t *spa)
3977 {
3978 	int error = 0;
3979 	vdev_t *rvd = spa->spa_root_vdev;
3980 
3981 	spa->spa_is_initializing = B_TRUE;
3982 	error = dsl_pool_open(spa->spa_dsl_pool);
3983 	spa->spa_is_initializing = B_FALSE;
3984 	if (error != 0) {
3985 		spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error);
3986 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3987 	}
3988 
3989 	return (0);
3990 }
3991 
3992 static int
3993 spa_ld_get_props(spa_t *spa)
3994 {
3995 	int error = 0;
3996 	uint64_t obj;
3997 	vdev_t *rvd = spa->spa_root_vdev;
3998 
3999 	/* Grab the checksum salt from the MOS. */
4000 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4001 	    DMU_POOL_CHECKSUM_SALT, 1,
4002 	    sizeof (spa->spa_cksum_salt.zcs_bytes),
4003 	    spa->spa_cksum_salt.zcs_bytes);
4004 	if (error == ENOENT) {
4005 		/* Generate a new salt for subsequent use */
4006 		(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
4007 		    sizeof (spa->spa_cksum_salt.zcs_bytes));
4008 	} else if (error != 0) {
4009 		spa_load_failed(spa, "unable to retrieve checksum salt from "
4010 		    "MOS [error=%d]", error);
4011 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4012 	}
4013 
4014 	if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0)
4015 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4016 	error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj);
4017 	if (error != 0) {
4018 		spa_load_failed(spa, "error opening deferred-frees bpobj "
4019 		    "[error=%d]", error);
4020 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4021 	}
4022 
4023 	/*
4024 	 * Load the bit that tells us to use the new accounting function
4025 	 * (raid-z deflation).  If we have an older pool, this will not
4026 	 * be present.
4027 	 */
4028 	error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE);
4029 	if (error != 0 && error != ENOENT)
4030 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4031 
4032 	error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION,
4033 	    &spa->spa_creation_version, B_FALSE);
4034 	if (error != 0 && error != ENOENT)
4035 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4036 
4037 	/*
4038 	 * Load the persistent error log.  If we have an older pool, this will
4039 	 * not be present.
4040 	 */
4041 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last,
4042 	    B_FALSE);
4043 	if (error != 0 && error != ENOENT)
4044 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4045 
4046 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB,
4047 	    &spa->spa_errlog_scrub, B_FALSE);
4048 	if (error != 0 && error != ENOENT)
4049 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4050 
4051 	/*
4052 	 * Load the livelist deletion field. If a livelist is queued for
4053 	 * deletion, indicate that in the spa
4054 	 */
4055 	error = spa_dir_prop(spa, DMU_POOL_DELETED_CLONES,
4056 	    &spa->spa_livelists_to_delete, B_FALSE);
4057 	if (error != 0 && error != ENOENT)
4058 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4059 
4060 	/*
4061 	 * Load the history object.  If we have an older pool, this
4062 	 * will not be present.
4063 	 */
4064 	error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE);
4065 	if (error != 0 && error != ENOENT)
4066 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4067 
4068 	/*
4069 	 * Load the per-vdev ZAP map. If we have an older pool, this will not
4070 	 * be present; in this case, defer its creation to a later time to
4071 	 * avoid dirtying the MOS this early / out of sync context. See
4072 	 * spa_sync_config_object.
4073 	 */
4074 
4075 	/* The sentinel is only available in the MOS config. */
4076 	nvlist_t *mos_config;
4077 	if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) {
4078 		spa_load_failed(spa, "unable to retrieve MOS config");
4079 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4080 	}
4081 
4082 	error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP,
4083 	    &spa->spa_all_vdev_zaps, B_FALSE);
4084 
4085 	if (error == ENOENT) {
4086 		VERIFY(!nvlist_exists(mos_config,
4087 		    ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
4088 		spa->spa_avz_action = AVZ_ACTION_INITIALIZE;
4089 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
4090 	} else if (error != 0) {
4091 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4092 	} else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) {
4093 		/*
4094 		 * An older version of ZFS overwrote the sentinel value, so
4095 		 * we have orphaned per-vdev ZAPs in the MOS. Defer their
4096 		 * destruction to later; see spa_sync_config_object.
4097 		 */
4098 		spa->spa_avz_action = AVZ_ACTION_DESTROY;
4099 		/*
4100 		 * We're assuming that no vdevs have had their ZAPs created
4101 		 * before this. Better be sure of it.
4102 		 */
4103 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
4104 	}
4105 	nvlist_free(mos_config);
4106 
4107 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
4108 
4109 	error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object,
4110 	    B_FALSE);
4111 	if (error && error != ENOENT)
4112 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4113 
4114 	if (error == 0) {
4115 		uint64_t autoreplace;
4116 
4117 		spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs);
4118 		spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace);
4119 		spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation);
4120 		spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode);
4121 		spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand);
4122 		spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost);
4123 		spa_prop_find(spa, ZPOOL_PROP_AUTOTRIM, &spa->spa_autotrim);
4124 		spa->spa_autoreplace = (autoreplace != 0);
4125 	}
4126 
4127 	/*
4128 	 * If we are importing a pool with missing top-level vdevs,
4129 	 * we enforce that the pool doesn't panic or get suspended on
4130 	 * error since the likelihood of missing data is extremely high.
4131 	 */
4132 	if (spa->spa_missing_tvds > 0 &&
4133 	    spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE &&
4134 	    spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4135 		spa_load_note(spa, "forcing failmode to 'continue' "
4136 		    "as some top level vdevs are missing");
4137 		spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE;
4138 	}
4139 
4140 	return (0);
4141 }
4142 
4143 static int
4144 spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type)
4145 {
4146 	int error = 0;
4147 	vdev_t *rvd = spa->spa_root_vdev;
4148 
4149 	/*
4150 	 * If we're assembling the pool from the split-off vdevs of
4151 	 * an existing pool, we don't want to attach the spares & cache
4152 	 * devices.
4153 	 */
4154 
4155 	/*
4156 	 * Load any hot spares for this pool.
4157 	 */
4158 	error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object,
4159 	    B_FALSE);
4160 	if (error != 0 && error != ENOENT)
4161 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4162 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
4163 		ASSERT(spa_version(spa) >= SPA_VERSION_SPARES);
4164 		if (load_nvlist(spa, spa->spa_spares.sav_object,
4165 		    &spa->spa_spares.sav_config) != 0) {
4166 			spa_load_failed(spa, "error loading spares nvlist");
4167 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4168 		}
4169 
4170 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4171 		spa_load_spares(spa);
4172 		spa_config_exit(spa, SCL_ALL, FTAG);
4173 	} else if (error == 0) {
4174 		spa->spa_spares.sav_sync = B_TRUE;
4175 	}
4176 
4177 	/*
4178 	 * Load any level 2 ARC devices for this pool.
4179 	 */
4180 	error = spa_dir_prop(spa, DMU_POOL_L2CACHE,
4181 	    &spa->spa_l2cache.sav_object, B_FALSE);
4182 	if (error != 0 && error != ENOENT)
4183 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4184 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
4185 		ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE);
4186 		if (load_nvlist(spa, spa->spa_l2cache.sav_object,
4187 		    &spa->spa_l2cache.sav_config) != 0) {
4188 			spa_load_failed(spa, "error loading l2cache nvlist");
4189 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4190 		}
4191 
4192 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4193 		spa_load_l2cache(spa);
4194 		spa_config_exit(spa, SCL_ALL, FTAG);
4195 	} else if (error == 0) {
4196 		spa->spa_l2cache.sav_sync = B_TRUE;
4197 	}
4198 
4199 	return (0);
4200 }
4201 
4202 static int
4203 spa_ld_load_vdev_metadata(spa_t *spa)
4204 {
4205 	int error = 0;
4206 	vdev_t *rvd = spa->spa_root_vdev;
4207 
4208 	/*
4209 	 * If the 'multihost' property is set, then never allow a pool to
4210 	 * be imported when the system hostid is zero.  The exception to
4211 	 * this rule is zdb which is always allowed to access pools.
4212 	 */
4213 	if (spa_multihost(spa) && spa_get_hostid(spa) == 0 &&
4214 	    (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) {
4215 		fnvlist_add_uint64(spa->spa_load_info,
4216 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
4217 		return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
4218 	}
4219 
4220 	/*
4221 	 * If the 'autoreplace' property is set, then post a resource notifying
4222 	 * the ZFS DE that it should not issue any faults for unopenable
4223 	 * devices.  We also iterate over the vdevs, and post a sysevent for any
4224 	 * unopenable vdevs so that the normal autoreplace handler can take
4225 	 * over.
4226 	 */
4227 	if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4228 		spa_check_removed(spa->spa_root_vdev);
4229 		/*
4230 		 * For the import case, this is done in spa_import(), because
4231 		 * at this point we're using the spare definitions from
4232 		 * the MOS config, not necessarily from the userland config.
4233 		 */
4234 		if (spa->spa_load_state != SPA_LOAD_IMPORT) {
4235 			spa_aux_check_removed(&spa->spa_spares);
4236 			spa_aux_check_removed(&spa->spa_l2cache);
4237 		}
4238 	}
4239 
4240 	/*
4241 	 * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc.
4242 	 */
4243 	error = vdev_load(rvd);
4244 	if (error != 0) {
4245 		spa_load_failed(spa, "vdev_load failed [error=%d]", error);
4246 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4247 	}
4248 
4249 	error = spa_ld_log_spacemaps(spa);
4250 	if (error != 0) {
4251 		spa_load_failed(spa, "spa_ld_log_sm_data failed [error=%d]",
4252 		    error);
4253 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
4254 	}
4255 
4256 	/*
4257 	 * Propagate the leaf DTLs we just loaded all the way up the vdev tree.
4258 	 */
4259 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4260 	vdev_dtl_reassess(rvd, 0, 0, B_FALSE, B_FALSE);
4261 	spa_config_exit(spa, SCL_ALL, FTAG);
4262 
4263 	return (0);
4264 }
4265 
4266 static int
4267 spa_ld_load_dedup_tables(spa_t *spa)
4268 {
4269 	int error = 0;
4270 	vdev_t *rvd = spa->spa_root_vdev;
4271 
4272 	error = ddt_load(spa);
4273 	if (error != 0) {
4274 		spa_load_failed(spa, "ddt_load failed [error=%d]", error);
4275 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
4276 	}
4277 
4278 	return (0);
4279 }
4280 
4281 static int
4282 spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, char **ereport)
4283 {
4284 	vdev_t *rvd = spa->spa_root_vdev;
4285 
4286 	if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) {
4287 		boolean_t missing = spa_check_logs(spa);
4288 		if (missing) {
4289 			if (spa->spa_missing_tvds != 0) {
4290 				spa_load_note(spa, "spa_check_logs failed "
4291 				    "so dropping the logs");
4292 			} else {
4293 				*ereport = FM_EREPORT_ZFS_LOG_REPLAY;
4294 				spa_load_failed(spa, "spa_check_logs failed");
4295 				return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG,
4296 				    ENXIO));
4297 			}
4298 		}
4299 	}
4300 
4301 	return (0);
4302 }
4303 
4304 static int
4305 spa_ld_verify_pool_data(spa_t *spa)
4306 {
4307 	int error = 0;
4308 	vdev_t *rvd = spa->spa_root_vdev;
4309 
4310 	/*
4311 	 * We've successfully opened the pool, verify that we're ready
4312 	 * to start pushing transactions.
4313 	 */
4314 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
4315 		error = spa_load_verify(spa);
4316 		if (error != 0) {
4317 			spa_load_failed(spa, "spa_load_verify failed "
4318 			    "[error=%d]", error);
4319 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
4320 			    error));
4321 		}
4322 	}
4323 
4324 	return (0);
4325 }
4326 
4327 static void
4328 spa_ld_claim_log_blocks(spa_t *spa)
4329 {
4330 	dmu_tx_t *tx;
4331 	dsl_pool_t *dp = spa_get_dsl(spa);
4332 
4333 	/*
4334 	 * Claim log blocks that haven't been committed yet.
4335 	 * This must all happen in a single txg.
4336 	 * Note: spa_claim_max_txg is updated by spa_claim_notify(),
4337 	 * invoked from zil_claim_log_block()'s i/o done callback.
4338 	 * Price of rollback is that we abandon the log.
4339 	 */
4340 	spa->spa_claiming = B_TRUE;
4341 
4342 	tx = dmu_tx_create_assigned(dp, spa_first_txg(spa));
4343 	(void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
4344 	    zil_claim, tx, DS_FIND_CHILDREN);
4345 	dmu_tx_commit(tx);
4346 
4347 	spa->spa_claiming = B_FALSE;
4348 
4349 	spa_set_log_state(spa, SPA_LOG_GOOD);
4350 }
4351 
4352 static void
4353 spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg,
4354     boolean_t update_config_cache)
4355 {
4356 	vdev_t *rvd = spa->spa_root_vdev;
4357 	int need_update = B_FALSE;
4358 
4359 	/*
4360 	 * If the config cache is stale, or we have uninitialized
4361 	 * metaslabs (see spa_vdev_add()), then update the config.
4362 	 *
4363 	 * If this is a verbatim import, trust the current
4364 	 * in-core spa_config and update the disk labels.
4365 	 */
4366 	if (update_config_cache || config_cache_txg != spa->spa_config_txg ||
4367 	    spa->spa_load_state == SPA_LOAD_IMPORT ||
4368 	    spa->spa_load_state == SPA_LOAD_RECOVER ||
4369 	    (spa->spa_import_flags & ZFS_IMPORT_VERBATIM))
4370 		need_update = B_TRUE;
4371 
4372 	for (int c = 0; c < rvd->vdev_children; c++)
4373 		if (rvd->vdev_child[c]->vdev_ms_array == 0)
4374 			need_update = B_TRUE;
4375 
4376 	/*
4377 	 * Update the config cache asynchronously in case we're the
4378 	 * root pool, in which case the config cache isn't writable yet.
4379 	 */
4380 	if (need_update)
4381 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
4382 }
4383 
4384 static void
4385 spa_ld_prepare_for_reload(spa_t *spa)
4386 {
4387 	spa_mode_t mode = spa->spa_mode;
4388 	int async_suspended = spa->spa_async_suspended;
4389 
4390 	spa_unload(spa);
4391 	spa_deactivate(spa);
4392 	spa_activate(spa, mode);
4393 
4394 	/*
4395 	 * We save the value of spa_async_suspended as it gets reset to 0 by
4396 	 * spa_unload(). We want to restore it back to the original value before
4397 	 * returning as we might be calling spa_async_resume() later.
4398 	 */
4399 	spa->spa_async_suspended = async_suspended;
4400 }
4401 
4402 static int
4403 spa_ld_read_checkpoint_txg(spa_t *spa)
4404 {
4405 	uberblock_t checkpoint;
4406 	int error = 0;
4407 
4408 	ASSERT0(spa->spa_checkpoint_txg);
4409 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4410 
4411 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4412 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
4413 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
4414 
4415 	if (error == ENOENT)
4416 		return (0);
4417 
4418 	if (error != 0)
4419 		return (error);
4420 
4421 	ASSERT3U(checkpoint.ub_txg, !=, 0);
4422 	ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0);
4423 	ASSERT3U(checkpoint.ub_timestamp, !=, 0);
4424 	spa->spa_checkpoint_txg = checkpoint.ub_txg;
4425 	spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp;
4426 
4427 	return (0);
4428 }
4429 
4430 static int
4431 spa_ld_mos_init(spa_t *spa, spa_import_type_t type)
4432 {
4433 	int error = 0;
4434 
4435 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4436 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
4437 
4438 	/*
4439 	 * Never trust the config that is provided unless we are assembling
4440 	 * a pool following a split.
4441 	 * This means don't trust blkptrs and the vdev tree in general. This
4442 	 * also effectively puts the spa in read-only mode since
4443 	 * spa_writeable() checks for spa_trust_config to be true.
4444 	 * We will later load a trusted config from the MOS.
4445 	 */
4446 	if (type != SPA_IMPORT_ASSEMBLE)
4447 		spa->spa_trust_config = B_FALSE;
4448 
4449 	/*
4450 	 * Parse the config provided to create a vdev tree.
4451 	 */
4452 	error = spa_ld_parse_config(spa, type);
4453 	if (error != 0)
4454 		return (error);
4455 
4456 	spa_import_progress_add(spa);
4457 
4458 	/*
4459 	 * Now that we have the vdev tree, try to open each vdev. This involves
4460 	 * opening the underlying physical device, retrieving its geometry and
4461 	 * probing the vdev with a dummy I/O. The state of each vdev will be set
4462 	 * based on the success of those operations. After this we'll be ready
4463 	 * to read from the vdevs.
4464 	 */
4465 	error = spa_ld_open_vdevs(spa);
4466 	if (error != 0)
4467 		return (error);
4468 
4469 	/*
4470 	 * Read the label of each vdev and make sure that the GUIDs stored
4471 	 * there match the GUIDs in the config provided.
4472 	 * If we're assembling a new pool that's been split off from an
4473 	 * existing pool, the labels haven't yet been updated so we skip
4474 	 * validation for now.
4475 	 */
4476 	if (type != SPA_IMPORT_ASSEMBLE) {
4477 		error = spa_ld_validate_vdevs(spa);
4478 		if (error != 0)
4479 			return (error);
4480 	}
4481 
4482 	/*
4483 	 * Read all vdev labels to find the best uberblock (i.e. latest,
4484 	 * unless spa_load_max_txg is set) and store it in spa_uberblock. We
4485 	 * get the list of features required to read blkptrs in the MOS from
4486 	 * the vdev label with the best uberblock and verify that our version
4487 	 * of zfs supports them all.
4488 	 */
4489 	error = spa_ld_select_uberblock(spa, type);
4490 	if (error != 0)
4491 		return (error);
4492 
4493 	/*
4494 	 * Pass that uberblock to the dsl_pool layer which will open the root
4495 	 * blkptr. This blkptr points to the latest version of the MOS and will
4496 	 * allow us to read its contents.
4497 	 */
4498 	error = spa_ld_open_rootbp(spa);
4499 	if (error != 0)
4500 		return (error);
4501 
4502 	return (0);
4503 }
4504 
4505 static int
4506 spa_ld_checkpoint_rewind(spa_t *spa)
4507 {
4508 	uberblock_t checkpoint;
4509 	int error = 0;
4510 
4511 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4512 	ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4513 
4514 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
4515 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
4516 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
4517 
4518 	if (error != 0) {
4519 		spa_load_failed(spa, "unable to retrieve checkpointed "
4520 		    "uberblock from the MOS config [error=%d]", error);
4521 
4522 		if (error == ENOENT)
4523 			error = ZFS_ERR_NO_CHECKPOINT;
4524 
4525 		return (error);
4526 	}
4527 
4528 	ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg);
4529 	ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg);
4530 
4531 	/*
4532 	 * We need to update the txg and timestamp of the checkpointed
4533 	 * uberblock to be higher than the latest one. This ensures that
4534 	 * the checkpointed uberblock is selected if we were to close and
4535 	 * reopen the pool right after we've written it in the vdev labels.
4536 	 * (also see block comment in vdev_uberblock_compare)
4537 	 */
4538 	checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1;
4539 	checkpoint.ub_timestamp = gethrestime_sec();
4540 
4541 	/*
4542 	 * Set current uberblock to be the checkpointed uberblock.
4543 	 */
4544 	spa->spa_uberblock = checkpoint;
4545 
4546 	/*
4547 	 * If we are doing a normal rewind, then the pool is open for
4548 	 * writing and we sync the "updated" checkpointed uberblock to
4549 	 * disk. Once this is done, we've basically rewound the whole
4550 	 * pool and there is no way back.
4551 	 *
4552 	 * There are cases when we don't want to attempt and sync the
4553 	 * checkpointed uberblock to disk because we are opening a
4554 	 * pool as read-only. Specifically, verifying the checkpointed
4555 	 * state with zdb, and importing the checkpointed state to get
4556 	 * a "preview" of its content.
4557 	 */
4558 	if (spa_writeable(spa)) {
4559 		vdev_t *rvd = spa->spa_root_vdev;
4560 
4561 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4562 		vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
4563 		int svdcount = 0;
4564 		int children = rvd->vdev_children;
4565 		int c0 = spa_get_random(children);
4566 
4567 		for (int c = 0; c < children; c++) {
4568 			vdev_t *vd = rvd->vdev_child[(c0 + c) % children];
4569 
4570 			/* Stop when revisiting the first vdev */
4571 			if (c > 0 && svd[0] == vd)
4572 				break;
4573 
4574 			if (vd->vdev_ms_array == 0 || vd->vdev_islog ||
4575 			    !vdev_is_concrete(vd))
4576 				continue;
4577 
4578 			svd[svdcount++] = vd;
4579 			if (svdcount == SPA_SYNC_MIN_VDEVS)
4580 				break;
4581 		}
4582 		error = vdev_config_sync(svd, svdcount, spa->spa_first_txg);
4583 		if (error == 0)
4584 			spa->spa_last_synced_guid = rvd->vdev_guid;
4585 		spa_config_exit(spa, SCL_ALL, FTAG);
4586 
4587 		if (error != 0) {
4588 			spa_load_failed(spa, "failed to write checkpointed "
4589 			    "uberblock to the vdev labels [error=%d]", error);
4590 			return (error);
4591 		}
4592 	}
4593 
4594 	return (0);
4595 }
4596 
4597 static int
4598 spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type,
4599     boolean_t *update_config_cache)
4600 {
4601 	int error;
4602 
4603 	/*
4604 	 * Parse the config for pool, open and validate vdevs,
4605 	 * select an uberblock, and use that uberblock to open
4606 	 * the MOS.
4607 	 */
4608 	error = spa_ld_mos_init(spa, type);
4609 	if (error != 0)
4610 		return (error);
4611 
4612 	/*
4613 	 * Retrieve the trusted config stored in the MOS and use it to create
4614 	 * a new, exact version of the vdev tree, then reopen all vdevs.
4615 	 */
4616 	error = spa_ld_trusted_config(spa, type, B_FALSE);
4617 	if (error == EAGAIN) {
4618 		if (update_config_cache != NULL)
4619 			*update_config_cache = B_TRUE;
4620 
4621 		/*
4622 		 * Redo the loading process with the trusted config if it is
4623 		 * too different from the untrusted config.
4624 		 */
4625 		spa_ld_prepare_for_reload(spa);
4626 		spa_load_note(spa, "RELOADING");
4627 		error = spa_ld_mos_init(spa, type);
4628 		if (error != 0)
4629 			return (error);
4630 
4631 		error = spa_ld_trusted_config(spa, type, B_TRUE);
4632 		if (error != 0)
4633 			return (error);
4634 
4635 	} else if (error != 0) {
4636 		return (error);
4637 	}
4638 
4639 	return (0);
4640 }
4641 
4642 /*
4643  * Load an existing storage pool, using the config provided. This config
4644  * describes which vdevs are part of the pool and is later validated against
4645  * partial configs present in each vdev's label and an entire copy of the
4646  * config stored in the MOS.
4647  */
4648 static int
4649 spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport)
4650 {
4651 	int error = 0;
4652 	boolean_t missing_feat_write = B_FALSE;
4653 	boolean_t checkpoint_rewind =
4654 	    (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4655 	boolean_t update_config_cache = B_FALSE;
4656 
4657 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
4658 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
4659 
4660 	spa_load_note(spa, "LOADING");
4661 
4662 	error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache);
4663 	if (error != 0)
4664 		return (error);
4665 
4666 	/*
4667 	 * If we are rewinding to the checkpoint then we need to repeat
4668 	 * everything we've done so far in this function but this time
4669 	 * selecting the checkpointed uberblock and using that to open
4670 	 * the MOS.
4671 	 */
4672 	if (checkpoint_rewind) {
4673 		/*
4674 		 * If we are rewinding to the checkpoint update config cache
4675 		 * anyway.
4676 		 */
4677 		update_config_cache = B_TRUE;
4678 
4679 		/*
4680 		 * Extract the checkpointed uberblock from the current MOS
4681 		 * and use this as the pool's uberblock from now on. If the
4682 		 * pool is imported as writeable we also write the checkpoint
4683 		 * uberblock to the labels, making the rewind permanent.
4684 		 */
4685 		error = spa_ld_checkpoint_rewind(spa);
4686 		if (error != 0)
4687 			return (error);
4688 
4689 		/*
4690 		 * Redo the loading process again with the
4691 		 * checkpointed uberblock.
4692 		 */
4693 		spa_ld_prepare_for_reload(spa);
4694 		spa_load_note(spa, "LOADING checkpointed uberblock");
4695 		error = spa_ld_mos_with_trusted_config(spa, type, NULL);
4696 		if (error != 0)
4697 			return (error);
4698 	}
4699 
4700 	/*
4701 	 * Retrieve the checkpoint txg if the pool has a checkpoint.
4702 	 */
4703 	error = spa_ld_read_checkpoint_txg(spa);
4704 	if (error != 0)
4705 		return (error);
4706 
4707 	/*
4708 	 * Retrieve the mapping of indirect vdevs. Those vdevs were removed
4709 	 * from the pool and their contents were re-mapped to other vdevs. Note
4710 	 * that everything that we read before this step must have been
4711 	 * rewritten on concrete vdevs after the last device removal was
4712 	 * initiated. Otherwise we could be reading from indirect vdevs before
4713 	 * we have loaded their mappings.
4714 	 */
4715 	error = spa_ld_open_indirect_vdev_metadata(spa);
4716 	if (error != 0)
4717 		return (error);
4718 
4719 	/*
4720 	 * Retrieve the full list of active features from the MOS and check if
4721 	 * they are all supported.
4722 	 */
4723 	error = spa_ld_check_features(spa, &missing_feat_write);
4724 	if (error != 0)
4725 		return (error);
4726 
4727 	/*
4728 	 * Load several special directories from the MOS needed by the dsl_pool
4729 	 * layer.
4730 	 */
4731 	error = spa_ld_load_special_directories(spa);
4732 	if (error != 0)
4733 		return (error);
4734 
4735 	/*
4736 	 * Retrieve pool properties from the MOS.
4737 	 */
4738 	error = spa_ld_get_props(spa);
4739 	if (error != 0)
4740 		return (error);
4741 
4742 	/*
4743 	 * Retrieve the list of auxiliary devices - cache devices and spares -
4744 	 * and open them.
4745 	 */
4746 	error = spa_ld_open_aux_vdevs(spa, type);
4747 	if (error != 0)
4748 		return (error);
4749 
4750 	/*
4751 	 * Load the metadata for all vdevs. Also check if unopenable devices
4752 	 * should be autoreplaced.
4753 	 */
4754 	error = spa_ld_load_vdev_metadata(spa);
4755 	if (error != 0)
4756 		return (error);
4757 
4758 	error = spa_ld_load_dedup_tables(spa);
4759 	if (error != 0)
4760 		return (error);
4761 
4762 	/*
4763 	 * Verify the logs now to make sure we don't have any unexpected errors
4764 	 * when we claim log blocks later.
4765 	 */
4766 	error = spa_ld_verify_logs(spa, type, ereport);
4767 	if (error != 0)
4768 		return (error);
4769 
4770 	if (missing_feat_write) {
4771 		ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT);
4772 
4773 		/*
4774 		 * At this point, we know that we can open the pool in
4775 		 * read-only mode but not read-write mode. We now have enough
4776 		 * information and can return to userland.
4777 		 */
4778 		return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT,
4779 		    ENOTSUP));
4780 	}
4781 
4782 	/*
4783 	 * Traverse the last txgs to make sure the pool was left off in a safe
4784 	 * state. When performing an extreme rewind, we verify the whole pool,
4785 	 * which can take a very long time.
4786 	 */
4787 	error = spa_ld_verify_pool_data(spa);
4788 	if (error != 0)
4789 		return (error);
4790 
4791 	/*
4792 	 * Calculate the deflated space for the pool. This must be done before
4793 	 * we write anything to the pool because we'd need to update the space
4794 	 * accounting using the deflated sizes.
4795 	 */
4796 	spa_update_dspace(spa);
4797 
4798 	/*
4799 	 * We have now retrieved all the information we needed to open the
4800 	 * pool. If we are importing the pool in read-write mode, a few
4801 	 * additional steps must be performed to finish the import.
4802 	 */
4803 	if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER ||
4804 	    spa->spa_load_max_txg == UINT64_MAX)) {
4805 		uint64_t config_cache_txg = spa->spa_config_txg;
4806 
4807 		ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT);
4808 
4809 		/*
4810 		 * In case of a checkpoint rewind, log the original txg
4811 		 * of the checkpointed uberblock.
4812 		 */
4813 		if (checkpoint_rewind) {
4814 			spa_history_log_internal(spa, "checkpoint rewind",
4815 			    NULL, "rewound state to txg=%llu",
4816 			    (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg);
4817 		}
4818 
4819 		/*
4820 		 * Traverse the ZIL and claim all blocks.
4821 		 */
4822 		spa_ld_claim_log_blocks(spa);
4823 
4824 		/*
4825 		 * Kick-off the syncing thread.
4826 		 */
4827 		spa->spa_sync_on = B_TRUE;
4828 		txg_sync_start(spa->spa_dsl_pool);
4829 		mmp_thread_start(spa);
4830 
4831 		/*
4832 		 * Wait for all claims to sync.  We sync up to the highest
4833 		 * claimed log block birth time so that claimed log blocks
4834 		 * don't appear to be from the future.  spa_claim_max_txg
4835 		 * will have been set for us by ZIL traversal operations
4836 		 * performed above.
4837 		 */
4838 		txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg);
4839 
4840 		/*
4841 		 * Check if we need to request an update of the config. On the
4842 		 * next sync, we would update the config stored in vdev labels
4843 		 * and the cachefile (by default /etc/zfs/zpool.cache).
4844 		 */
4845 		spa_ld_check_for_config_update(spa, config_cache_txg,
4846 		    update_config_cache);
4847 
4848 		/*
4849 		 * Check if a rebuild was in progress and if so resume it.
4850 		 * Then check all DTLs to see if anything needs resilvering.
4851 		 * The resilver will be deferred if a rebuild was started.
4852 		 */
4853 		if (vdev_rebuild_active(spa->spa_root_vdev)) {
4854 			vdev_rebuild_restart(spa);
4855 		} else if (!dsl_scan_resilvering(spa->spa_dsl_pool) &&
4856 		    vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
4857 			spa_async_request(spa, SPA_ASYNC_RESILVER);
4858 		}
4859 
4860 		/*
4861 		 * Log the fact that we booted up (so that we can detect if
4862 		 * we rebooted in the middle of an operation).
4863 		 */
4864 		spa_history_log_version(spa, "open", NULL);
4865 
4866 		spa_restart_removal(spa);
4867 		spa_spawn_aux_threads(spa);
4868 
4869 		/*
4870 		 * Delete any inconsistent datasets.
4871 		 *
4872 		 * Note:
4873 		 * Since we may be issuing deletes for clones here,
4874 		 * we make sure to do so after we've spawned all the
4875 		 * auxiliary threads above (from which the livelist
4876 		 * deletion zthr is part of).
4877 		 */
4878 		(void) dmu_objset_find(spa_name(spa),
4879 		    dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN);
4880 
4881 		/*
4882 		 * Clean up any stale temporary dataset userrefs.
4883 		 */
4884 		dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool);
4885 
4886 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
4887 		vdev_initialize_restart(spa->spa_root_vdev);
4888 		vdev_trim_restart(spa->spa_root_vdev);
4889 		vdev_autotrim_restart(spa);
4890 		spa_config_exit(spa, SCL_CONFIG, FTAG);
4891 	}
4892 
4893 	spa_import_progress_remove(spa_guid(spa));
4894 	spa_async_request(spa, SPA_ASYNC_L2CACHE_REBUILD);
4895 
4896 	spa_load_note(spa, "LOADED");
4897 
4898 	return (0);
4899 }
4900 
4901 static int
4902 spa_load_retry(spa_t *spa, spa_load_state_t state)
4903 {
4904 	spa_mode_t mode = spa->spa_mode;
4905 
4906 	spa_unload(spa);
4907 	spa_deactivate(spa);
4908 
4909 	spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1;
4910 
4911 	spa_activate(spa, mode);
4912 	spa_async_suspend(spa);
4913 
4914 	spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu",
4915 	    (u_longlong_t)spa->spa_load_max_txg);
4916 
4917 	return (spa_load(spa, state, SPA_IMPORT_EXISTING));
4918 }
4919 
4920 /*
4921  * If spa_load() fails this function will try loading prior txg's. If
4922  * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool
4923  * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this
4924  * function will not rewind the pool and will return the same error as
4925  * spa_load().
4926  */
4927 static int
4928 spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request,
4929     int rewind_flags)
4930 {
4931 	nvlist_t *loadinfo = NULL;
4932 	nvlist_t *config = NULL;
4933 	int load_error, rewind_error;
4934 	uint64_t safe_rewind_txg;
4935 	uint64_t min_txg;
4936 
4937 	if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) {
4938 		spa->spa_load_max_txg = spa->spa_load_txg;
4939 		spa_set_log_state(spa, SPA_LOG_CLEAR);
4940 	} else {
4941 		spa->spa_load_max_txg = max_request;
4942 		if (max_request != UINT64_MAX)
4943 			spa->spa_extreme_rewind = B_TRUE;
4944 	}
4945 
4946 	load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING);
4947 	if (load_error == 0)
4948 		return (0);
4949 	if (load_error == ZFS_ERR_NO_CHECKPOINT) {
4950 		/*
4951 		 * When attempting checkpoint-rewind on a pool with no
4952 		 * checkpoint, we should not attempt to load uberblocks
4953 		 * from previous txgs when spa_load fails.
4954 		 */
4955 		ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4956 		spa_import_progress_remove(spa_guid(spa));
4957 		return (load_error);
4958 	}
4959 
4960 	if (spa->spa_root_vdev != NULL)
4961 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
4962 
4963 	spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg;
4964 	spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp;
4965 
4966 	if (rewind_flags & ZPOOL_NEVER_REWIND) {
4967 		nvlist_free(config);
4968 		spa_import_progress_remove(spa_guid(spa));
4969 		return (load_error);
4970 	}
4971 
4972 	if (state == SPA_LOAD_RECOVER) {
4973 		/* Price of rolling back is discarding txgs, including log */
4974 		spa_set_log_state(spa, SPA_LOG_CLEAR);
4975 	} else {
4976 		/*
4977 		 * If we aren't rolling back save the load info from our first
4978 		 * import attempt so that we can restore it after attempting
4979 		 * to rewind.
4980 		 */
4981 		loadinfo = spa->spa_load_info;
4982 		spa->spa_load_info = fnvlist_alloc();
4983 	}
4984 
4985 	spa->spa_load_max_txg = spa->spa_last_ubsync_txg;
4986 	safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE;
4987 	min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ?
4988 	    TXG_INITIAL : safe_rewind_txg;
4989 
4990 	/*
4991 	 * Continue as long as we're finding errors, we're still within
4992 	 * the acceptable rewind range, and we're still finding uberblocks
4993 	 */
4994 	while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg &&
4995 	    spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) {
4996 		if (spa->spa_load_max_txg < safe_rewind_txg)
4997 			spa->spa_extreme_rewind = B_TRUE;
4998 		rewind_error = spa_load_retry(spa, state);
4999 	}
5000 
5001 	spa->spa_extreme_rewind = B_FALSE;
5002 	spa->spa_load_max_txg = UINT64_MAX;
5003 
5004 	if (config && (rewind_error || state != SPA_LOAD_RECOVER))
5005 		spa_config_set(spa, config);
5006 	else
5007 		nvlist_free(config);
5008 
5009 	if (state == SPA_LOAD_RECOVER) {
5010 		ASSERT3P(loadinfo, ==, NULL);
5011 		spa_import_progress_remove(spa_guid(spa));
5012 		return (rewind_error);
5013 	} else {
5014 		/* Store the rewind info as part of the initial load info */
5015 		fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO,
5016 		    spa->spa_load_info);
5017 
5018 		/* Restore the initial load info */
5019 		fnvlist_free(spa->spa_load_info);
5020 		spa->spa_load_info = loadinfo;
5021 
5022 		spa_import_progress_remove(spa_guid(spa));
5023 		return (load_error);
5024 	}
5025 }
5026 
5027 /*
5028  * Pool Open/Import
5029  *
5030  * The import case is identical to an open except that the configuration is sent
5031  * down from userland, instead of grabbed from the configuration cache.  For the
5032  * case of an open, the pool configuration will exist in the
5033  * POOL_STATE_UNINITIALIZED state.
5034  *
5035  * The stats information (gen/count/ustats) is used to gather vdev statistics at
5036  * the same time open the pool, without having to keep around the spa_t in some
5037  * ambiguous state.
5038  */
5039 static int
5040 spa_open_common(const char *pool, spa_t **spapp, void *tag, nvlist_t *nvpolicy,
5041     nvlist_t **config)
5042 {
5043 	spa_t *spa;
5044 	spa_load_state_t state = SPA_LOAD_OPEN;
5045 	int error;
5046 	int locked = B_FALSE;
5047 	int firstopen = B_FALSE;
5048 
5049 	*spapp = NULL;
5050 
5051 	/*
5052 	 * As disgusting as this is, we need to support recursive calls to this
5053 	 * function because dsl_dir_open() is called during spa_load(), and ends
5054 	 * up calling spa_open() again.  The real fix is to figure out how to
5055 	 * avoid dsl_dir_open() calling this in the first place.
5056 	 */
5057 	if (MUTEX_NOT_HELD(&spa_namespace_lock)) {
5058 		mutex_enter(&spa_namespace_lock);
5059 		locked = B_TRUE;
5060 	}
5061 
5062 	if ((spa = spa_lookup(pool)) == NULL) {
5063 		if (locked)
5064 			mutex_exit(&spa_namespace_lock);
5065 		return (SET_ERROR(ENOENT));
5066 	}
5067 
5068 	if (spa->spa_state == POOL_STATE_UNINITIALIZED) {
5069 		zpool_load_policy_t policy;
5070 
5071 		firstopen = B_TRUE;
5072 
5073 		zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config,
5074 		    &policy);
5075 		if (policy.zlp_rewind & ZPOOL_DO_REWIND)
5076 			state = SPA_LOAD_RECOVER;
5077 
5078 		spa_activate(spa, spa_mode_global);
5079 
5080 		if (state != SPA_LOAD_RECOVER)
5081 			spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
5082 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
5083 
5084 		zfs_dbgmsg("spa_open_common: opening %s", pool);
5085 		error = spa_load_best(spa, state, policy.zlp_txg,
5086 		    policy.zlp_rewind);
5087 
5088 		if (error == EBADF) {
5089 			/*
5090 			 * If vdev_validate() returns failure (indicated by
5091 			 * EBADF), it indicates that one of the vdevs indicates
5092 			 * that the pool has been exported or destroyed.  If
5093 			 * this is the case, the config cache is out of sync and
5094 			 * we should remove the pool from the namespace.
5095 			 */
5096 			spa_unload(spa);
5097 			spa_deactivate(spa);
5098 			spa_write_cachefile(spa, B_TRUE, B_TRUE);
5099 			spa_remove(spa);
5100 			if (locked)
5101 				mutex_exit(&spa_namespace_lock);
5102 			return (SET_ERROR(ENOENT));
5103 		}
5104 
5105 		if (error) {
5106 			/*
5107 			 * We can't open the pool, but we still have useful
5108 			 * information: the state of each vdev after the
5109 			 * attempted vdev_open().  Return this to the user.
5110 			 */
5111 			if (config != NULL && spa->spa_config) {
5112 				VERIFY(nvlist_dup(spa->spa_config, config,
5113 				    KM_SLEEP) == 0);
5114 				VERIFY(nvlist_add_nvlist(*config,
5115 				    ZPOOL_CONFIG_LOAD_INFO,
5116 				    spa->spa_load_info) == 0);
5117 			}
5118 			spa_unload(spa);
5119 			spa_deactivate(spa);
5120 			spa->spa_last_open_failed = error;
5121 			if (locked)
5122 				mutex_exit(&spa_namespace_lock);
5123 			*spapp = NULL;
5124 			return (error);
5125 		}
5126 	}
5127 
5128 	spa_open_ref(spa, tag);
5129 
5130 	if (config != NULL)
5131 		*config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
5132 
5133 	/*
5134 	 * If we've recovered the pool, pass back any information we
5135 	 * gathered while doing the load.
5136 	 */
5137 	if (state == SPA_LOAD_RECOVER) {
5138 		VERIFY(nvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO,
5139 		    spa->spa_load_info) == 0);
5140 	}
5141 
5142 	if (locked) {
5143 		spa->spa_last_open_failed = 0;
5144 		spa->spa_last_ubsync_txg = 0;
5145 		spa->spa_load_txg = 0;
5146 		mutex_exit(&spa_namespace_lock);
5147 	}
5148 
5149 	if (firstopen)
5150 		zvol_create_minors_recursive(spa_name(spa));
5151 
5152 	*spapp = spa;
5153 
5154 	return (0);
5155 }
5156 
5157 int
5158 spa_open_rewind(const char *name, spa_t **spapp, void *tag, nvlist_t *policy,
5159     nvlist_t **config)
5160 {
5161 	return (spa_open_common(name, spapp, tag, policy, config));
5162 }
5163 
5164 int
5165 spa_open(const char *name, spa_t **spapp, void *tag)
5166 {
5167 	return (spa_open_common(name, spapp, tag, NULL, NULL));
5168 }
5169 
5170 /*
5171  * Lookup the given spa_t, incrementing the inject count in the process,
5172  * preventing it from being exported or destroyed.
5173  */
5174 spa_t *
5175 spa_inject_addref(char *name)
5176 {
5177 	spa_t *spa;
5178 
5179 	mutex_enter(&spa_namespace_lock);
5180 	if ((spa = spa_lookup(name)) == NULL) {
5181 		mutex_exit(&spa_namespace_lock);
5182 		return (NULL);
5183 	}
5184 	spa->spa_inject_ref++;
5185 	mutex_exit(&spa_namespace_lock);
5186 
5187 	return (spa);
5188 }
5189 
5190 void
5191 spa_inject_delref(spa_t *spa)
5192 {
5193 	mutex_enter(&spa_namespace_lock);
5194 	spa->spa_inject_ref--;
5195 	mutex_exit(&spa_namespace_lock);
5196 }
5197 
5198 /*
5199  * Add spares device information to the nvlist.
5200  */
5201 static void
5202 spa_add_spares(spa_t *spa, nvlist_t *config)
5203 {
5204 	nvlist_t **spares;
5205 	uint_t i, nspares;
5206 	nvlist_t *nvroot;
5207 	uint64_t guid;
5208 	vdev_stat_t *vs;
5209 	uint_t vsc;
5210 	uint64_t pool;
5211 
5212 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5213 
5214 	if (spa->spa_spares.sav_count == 0)
5215 		return;
5216 
5217 	VERIFY(nvlist_lookup_nvlist(config,
5218 	    ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0);
5219 	VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
5220 	    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0);
5221 	if (nspares != 0) {
5222 		VERIFY(nvlist_add_nvlist_array(nvroot,
5223 		    ZPOOL_CONFIG_SPARES, spares, nspares) == 0);
5224 		VERIFY(nvlist_lookup_nvlist_array(nvroot,
5225 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0);
5226 
5227 		/*
5228 		 * Go through and find any spares which have since been
5229 		 * repurposed as an active spare.  If this is the case, update
5230 		 * their status appropriately.
5231 		 */
5232 		for (i = 0; i < nspares; i++) {
5233 			VERIFY(nvlist_lookup_uint64(spares[i],
5234 			    ZPOOL_CONFIG_GUID, &guid) == 0);
5235 			if (spa_spare_exists(guid, &pool, NULL) &&
5236 			    pool != 0ULL) {
5237 				VERIFY(nvlist_lookup_uint64_array(
5238 				    spares[i], ZPOOL_CONFIG_VDEV_STATS,
5239 				    (uint64_t **)&vs, &vsc) == 0);
5240 				vs->vs_state = VDEV_STATE_CANT_OPEN;
5241 				vs->vs_aux = VDEV_AUX_SPARED;
5242 			}
5243 		}
5244 	}
5245 }
5246 
5247 /*
5248  * Add l2cache device information to the nvlist, including vdev stats.
5249  */
5250 static void
5251 spa_add_l2cache(spa_t *spa, nvlist_t *config)
5252 {
5253 	nvlist_t **l2cache;
5254 	uint_t i, j, nl2cache;
5255 	nvlist_t *nvroot;
5256 	uint64_t guid;
5257 	vdev_t *vd;
5258 	vdev_stat_t *vs;
5259 	uint_t vsc;
5260 
5261 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5262 
5263 	if (spa->spa_l2cache.sav_count == 0)
5264 		return;
5265 
5266 	VERIFY(nvlist_lookup_nvlist(config,
5267 	    ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0);
5268 	VERIFY(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
5269 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0);
5270 	if (nl2cache != 0) {
5271 		VERIFY(nvlist_add_nvlist_array(nvroot,
5272 		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0);
5273 		VERIFY(nvlist_lookup_nvlist_array(nvroot,
5274 		    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0);
5275 
5276 		/*
5277 		 * Update level 2 cache device stats.
5278 		 */
5279 
5280 		for (i = 0; i < nl2cache; i++) {
5281 			VERIFY(nvlist_lookup_uint64(l2cache[i],
5282 			    ZPOOL_CONFIG_GUID, &guid) == 0);
5283 
5284 			vd = NULL;
5285 			for (j = 0; j < spa->spa_l2cache.sav_count; j++) {
5286 				if (guid ==
5287 				    spa->spa_l2cache.sav_vdevs[j]->vdev_guid) {
5288 					vd = spa->spa_l2cache.sav_vdevs[j];
5289 					break;
5290 				}
5291 			}
5292 			ASSERT(vd != NULL);
5293 
5294 			VERIFY(nvlist_lookup_uint64_array(l2cache[i],
5295 			    ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc)
5296 			    == 0);
5297 			vdev_get_stats(vd, vs);
5298 			vdev_config_generate_stats(vd, l2cache[i]);
5299 
5300 		}
5301 	}
5302 }
5303 
5304 static void
5305 spa_feature_stats_from_disk(spa_t *spa, nvlist_t *features)
5306 {
5307 	zap_cursor_t zc;
5308 	zap_attribute_t za;
5309 
5310 	if (spa->spa_feat_for_read_obj != 0) {
5311 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
5312 		    spa->spa_feat_for_read_obj);
5313 		    zap_cursor_retrieve(&zc, &za) == 0;
5314 		    zap_cursor_advance(&zc)) {
5315 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
5316 			    za.za_num_integers == 1);
5317 			VERIFY0(nvlist_add_uint64(features, za.za_name,
5318 			    za.za_first_integer));
5319 		}
5320 		zap_cursor_fini(&zc);
5321 	}
5322 
5323 	if (spa->spa_feat_for_write_obj != 0) {
5324 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
5325 		    spa->spa_feat_for_write_obj);
5326 		    zap_cursor_retrieve(&zc, &za) == 0;
5327 		    zap_cursor_advance(&zc)) {
5328 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
5329 			    za.za_num_integers == 1);
5330 			VERIFY0(nvlist_add_uint64(features, za.za_name,
5331 			    za.za_first_integer));
5332 		}
5333 		zap_cursor_fini(&zc);
5334 	}
5335 }
5336 
5337 static void
5338 spa_feature_stats_from_cache(spa_t *spa, nvlist_t *features)
5339 {
5340 	int i;
5341 
5342 	for (i = 0; i < SPA_FEATURES; i++) {
5343 		zfeature_info_t feature = spa_feature_table[i];
5344 		uint64_t refcount;
5345 
5346 		if (feature_get_refcount(spa, &feature, &refcount) != 0)
5347 			continue;
5348 
5349 		VERIFY0(nvlist_add_uint64(features, feature.fi_guid, refcount));
5350 	}
5351 }
5352 
5353 /*
5354  * Store a list of pool features and their reference counts in the
5355  * config.
5356  *
5357  * The first time this is called on a spa, allocate a new nvlist, fetch
5358  * the pool features and reference counts from disk, then save the list
5359  * in the spa. In subsequent calls on the same spa use the saved nvlist
5360  * and refresh its values from the cached reference counts.  This
5361  * ensures we don't block here on I/O on a suspended pool so 'zpool
5362  * clear' can resume the pool.
5363  */
5364 static void
5365 spa_add_feature_stats(spa_t *spa, nvlist_t *config)
5366 {
5367 	nvlist_t *features;
5368 
5369 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
5370 
5371 	mutex_enter(&spa->spa_feat_stats_lock);
5372 	features = spa->spa_feat_stats;
5373 
5374 	if (features != NULL) {
5375 		spa_feature_stats_from_cache(spa, features);
5376 	} else {
5377 		VERIFY0(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP));
5378 		spa->spa_feat_stats = features;
5379 		spa_feature_stats_from_disk(spa, features);
5380 	}
5381 
5382 	VERIFY0(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS,
5383 	    features));
5384 
5385 	mutex_exit(&spa->spa_feat_stats_lock);
5386 }
5387 
5388 int
5389 spa_get_stats(const char *name, nvlist_t **config,
5390     char *altroot, size_t buflen)
5391 {
5392 	int error;
5393 	spa_t *spa;
5394 
5395 	*config = NULL;
5396 	error = spa_open_common(name, &spa, FTAG, NULL, config);
5397 
5398 	if (spa != NULL) {
5399 		/*
5400 		 * This still leaves a window of inconsistency where the spares
5401 		 * or l2cache devices could change and the config would be
5402 		 * self-inconsistent.
5403 		 */
5404 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
5405 
5406 		if (*config != NULL) {
5407 			uint64_t loadtimes[2];
5408 
5409 			loadtimes[0] = spa->spa_loaded_ts.tv_sec;
5410 			loadtimes[1] = spa->spa_loaded_ts.tv_nsec;
5411 			VERIFY(nvlist_add_uint64_array(*config,
5412 			    ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2) == 0);
5413 
5414 			VERIFY(nvlist_add_uint64(*config,
5415 			    ZPOOL_CONFIG_ERRCOUNT,
5416 			    spa_get_errlog_size(spa)) == 0);
5417 
5418 			if (spa_suspended(spa)) {
5419 				VERIFY(nvlist_add_uint64(*config,
5420 				    ZPOOL_CONFIG_SUSPENDED,
5421 				    spa->spa_failmode) == 0);
5422 				VERIFY(nvlist_add_uint64(*config,
5423 				    ZPOOL_CONFIG_SUSPENDED_REASON,
5424 				    spa->spa_suspended) == 0);
5425 			}
5426 
5427 			spa_add_spares(spa, *config);
5428 			spa_add_l2cache(spa, *config);
5429 			spa_add_feature_stats(spa, *config);
5430 		}
5431 	}
5432 
5433 	/*
5434 	 * We want to get the alternate root even for faulted pools, so we cheat
5435 	 * and call spa_lookup() directly.
5436 	 */
5437 	if (altroot) {
5438 		if (spa == NULL) {
5439 			mutex_enter(&spa_namespace_lock);
5440 			spa = spa_lookup(name);
5441 			if (spa)
5442 				spa_altroot(spa, altroot, buflen);
5443 			else
5444 				altroot[0] = '\0';
5445 			spa = NULL;
5446 			mutex_exit(&spa_namespace_lock);
5447 		} else {
5448 			spa_altroot(spa, altroot, buflen);
5449 		}
5450 	}
5451 
5452 	if (spa != NULL) {
5453 		spa_config_exit(spa, SCL_CONFIG, FTAG);
5454 		spa_close(spa, FTAG);
5455 	}
5456 
5457 	return (error);
5458 }
5459 
5460 /*
5461  * Validate that the auxiliary device array is well formed.  We must have an
5462  * array of nvlists, each which describes a valid leaf vdev.  If this is an
5463  * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be
5464  * specified, as long as they are well-formed.
5465  */
5466 static int
5467 spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode,
5468     spa_aux_vdev_t *sav, const char *config, uint64_t version,
5469     vdev_labeltype_t label)
5470 {
5471 	nvlist_t **dev;
5472 	uint_t i, ndev;
5473 	vdev_t *vd;
5474 	int error;
5475 
5476 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5477 
5478 	/*
5479 	 * It's acceptable to have no devs specified.
5480 	 */
5481 	if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0)
5482 		return (0);
5483 
5484 	if (ndev == 0)
5485 		return (SET_ERROR(EINVAL));
5486 
5487 	/*
5488 	 * Make sure the pool is formatted with a version that supports this
5489 	 * device type.
5490 	 */
5491 	if (spa_version(spa) < version)
5492 		return (SET_ERROR(ENOTSUP));
5493 
5494 	/*
5495 	 * Set the pending device list so we correctly handle device in-use
5496 	 * checking.
5497 	 */
5498 	sav->sav_pending = dev;
5499 	sav->sav_npending = ndev;
5500 
5501 	for (i = 0; i < ndev; i++) {
5502 		if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0,
5503 		    mode)) != 0)
5504 			goto out;
5505 
5506 		if (!vd->vdev_ops->vdev_op_leaf) {
5507 			vdev_free(vd);
5508 			error = SET_ERROR(EINVAL);
5509 			goto out;
5510 		}
5511 
5512 		vd->vdev_top = vd;
5513 
5514 		if ((error = vdev_open(vd)) == 0 &&
5515 		    (error = vdev_label_init(vd, crtxg, label)) == 0) {
5516 			VERIFY(nvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID,
5517 			    vd->vdev_guid) == 0);
5518 		}
5519 
5520 		vdev_free(vd);
5521 
5522 		if (error &&
5523 		    (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE))
5524 			goto out;
5525 		else
5526 			error = 0;
5527 	}
5528 
5529 out:
5530 	sav->sav_pending = NULL;
5531 	sav->sav_npending = 0;
5532 	return (error);
5533 }
5534 
5535 static int
5536 spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode)
5537 {
5538 	int error;
5539 
5540 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
5541 
5542 	if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode,
5543 	    &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES,
5544 	    VDEV_LABEL_SPARE)) != 0) {
5545 		return (error);
5546 	}
5547 
5548 	return (spa_validate_aux_devs(spa, nvroot, crtxg, mode,
5549 	    &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE,
5550 	    VDEV_LABEL_L2CACHE));
5551 }
5552 
5553 static void
5554 spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs,
5555     const char *config)
5556 {
5557 	int i;
5558 
5559 	if (sav->sav_config != NULL) {
5560 		nvlist_t **olddevs;
5561 		uint_t oldndevs;
5562 		nvlist_t **newdevs;
5563 
5564 		/*
5565 		 * Generate new dev list by concatenating with the
5566 		 * current dev list.
5567 		 */
5568 		VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, config,
5569 		    &olddevs, &oldndevs) == 0);
5570 
5571 		newdevs = kmem_alloc(sizeof (void *) *
5572 		    (ndevs + oldndevs), KM_SLEEP);
5573 		for (i = 0; i < oldndevs; i++)
5574 			VERIFY(nvlist_dup(olddevs[i], &newdevs[i],
5575 			    KM_SLEEP) == 0);
5576 		for (i = 0; i < ndevs; i++)
5577 			VERIFY(nvlist_dup(devs[i], &newdevs[i + oldndevs],
5578 			    KM_SLEEP) == 0);
5579 
5580 		VERIFY(nvlist_remove(sav->sav_config, config,
5581 		    DATA_TYPE_NVLIST_ARRAY) == 0);
5582 
5583 		VERIFY(nvlist_add_nvlist_array(sav->sav_config,
5584 		    config, newdevs, ndevs + oldndevs) == 0);
5585 		for (i = 0; i < oldndevs + ndevs; i++)
5586 			nvlist_free(newdevs[i]);
5587 		kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *));
5588 	} else {
5589 		/*
5590 		 * Generate a new dev list.
5591 		 */
5592 		VERIFY(nvlist_alloc(&sav->sav_config, NV_UNIQUE_NAME,
5593 		    KM_SLEEP) == 0);
5594 		VERIFY(nvlist_add_nvlist_array(sav->sav_config, config,
5595 		    devs, ndevs) == 0);
5596 	}
5597 }
5598 
5599 /*
5600  * Stop and drop level 2 ARC devices
5601  */
5602 void
5603 spa_l2cache_drop(spa_t *spa)
5604 {
5605 	vdev_t *vd;
5606 	int i;
5607 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
5608 
5609 	for (i = 0; i < sav->sav_count; i++) {
5610 		uint64_t pool;
5611 
5612 		vd = sav->sav_vdevs[i];
5613 		ASSERT(vd != NULL);
5614 
5615 		if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
5616 		    pool != 0ULL && l2arc_vdev_present(vd))
5617 			l2arc_remove_vdev(vd);
5618 	}
5619 }
5620 
5621 /*
5622  * Verify encryption parameters for spa creation. If we are encrypting, we must
5623  * have the encryption feature flag enabled.
5624  */
5625 static int
5626 spa_create_check_encryption_params(dsl_crypto_params_t *dcp,
5627     boolean_t has_encryption)
5628 {
5629 	if (dcp->cp_crypt != ZIO_CRYPT_OFF &&
5630 	    dcp->cp_crypt != ZIO_CRYPT_INHERIT &&
5631 	    !has_encryption)
5632 		return (SET_ERROR(ENOTSUP));
5633 
5634 	return (dmu_objset_create_crypt_check(NULL, dcp, NULL));
5635 }
5636 
5637 /*
5638  * Pool Creation
5639  */
5640 int
5641 spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
5642     nvlist_t *zplprops, dsl_crypto_params_t *dcp)
5643 {
5644 	spa_t *spa;
5645 	char *altroot = NULL;
5646 	vdev_t *rvd;
5647 	dsl_pool_t *dp;
5648 	dmu_tx_t *tx;
5649 	int error = 0;
5650 	uint64_t txg = TXG_INITIAL;
5651 	nvlist_t **spares, **l2cache;
5652 	uint_t nspares, nl2cache;
5653 	uint64_t version, obj, ndraid = 0;
5654 	boolean_t has_features;
5655 	boolean_t has_encryption;
5656 	boolean_t has_allocclass;
5657 	spa_feature_t feat;
5658 	char *feat_name;
5659 	char *poolname;
5660 	nvlist_t *nvl;
5661 
5662 	if (props == NULL ||
5663 	    nvlist_lookup_string(props, "tname", &poolname) != 0)
5664 		poolname = (char *)pool;
5665 
5666 	/*
5667 	 * If this pool already exists, return failure.
5668 	 */
5669 	mutex_enter(&spa_namespace_lock);
5670 	if (spa_lookup(poolname) != NULL) {
5671 		mutex_exit(&spa_namespace_lock);
5672 		return (SET_ERROR(EEXIST));
5673 	}
5674 
5675 	/*
5676 	 * Allocate a new spa_t structure.
5677 	 */
5678 	nvl = fnvlist_alloc();
5679 	fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool);
5680 	(void) nvlist_lookup_string(props,
5681 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
5682 	spa = spa_add(poolname, nvl, altroot);
5683 	fnvlist_free(nvl);
5684 	spa_activate(spa, spa_mode_global);
5685 
5686 	if (props && (error = spa_prop_validate(spa, props))) {
5687 		spa_deactivate(spa);
5688 		spa_remove(spa);
5689 		mutex_exit(&spa_namespace_lock);
5690 		return (error);
5691 	}
5692 
5693 	/*
5694 	 * Temporary pool names should never be written to disk.
5695 	 */
5696 	if (poolname != pool)
5697 		spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME;
5698 
5699 	has_features = B_FALSE;
5700 	has_encryption = B_FALSE;
5701 	has_allocclass = B_FALSE;
5702 	for (nvpair_t *elem = nvlist_next_nvpair(props, NULL);
5703 	    elem != NULL; elem = nvlist_next_nvpair(props, elem)) {
5704 		if (zpool_prop_feature(nvpair_name(elem))) {
5705 			has_features = B_TRUE;
5706 
5707 			feat_name = strchr(nvpair_name(elem), '@') + 1;
5708 			VERIFY0(zfeature_lookup_name(feat_name, &feat));
5709 			if (feat == SPA_FEATURE_ENCRYPTION)
5710 				has_encryption = B_TRUE;
5711 			if (feat == SPA_FEATURE_ALLOCATION_CLASSES)
5712 				has_allocclass = B_TRUE;
5713 		}
5714 	}
5715 
5716 	/* verify encryption params, if they were provided */
5717 	if (dcp != NULL) {
5718 		error = spa_create_check_encryption_params(dcp, has_encryption);
5719 		if (error != 0) {
5720 			spa_deactivate(spa);
5721 			spa_remove(spa);
5722 			mutex_exit(&spa_namespace_lock);
5723 			return (error);
5724 		}
5725 	}
5726 	if (!has_allocclass && zfs_special_devs(nvroot, NULL)) {
5727 		spa_deactivate(spa);
5728 		spa_remove(spa);
5729 		mutex_exit(&spa_namespace_lock);
5730 		return (ENOTSUP);
5731 	}
5732 
5733 	if (has_features || nvlist_lookup_uint64(props,
5734 	    zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) {
5735 		version = SPA_VERSION;
5736 	}
5737 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
5738 
5739 	spa->spa_first_txg = txg;
5740 	spa->spa_uberblock.ub_txg = txg - 1;
5741 	spa->spa_uberblock.ub_version = version;
5742 	spa->spa_ubsync = spa->spa_uberblock;
5743 	spa->spa_load_state = SPA_LOAD_CREATE;
5744 	spa->spa_removing_phys.sr_state = DSS_NONE;
5745 	spa->spa_removing_phys.sr_removing_vdev = -1;
5746 	spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
5747 	spa->spa_indirect_vdevs_loaded = B_TRUE;
5748 
5749 	/*
5750 	 * Create "The Godfather" zio to hold all async IOs
5751 	 */
5752 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
5753 	    KM_SLEEP);
5754 	for (int i = 0; i < max_ncpus; i++) {
5755 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
5756 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
5757 		    ZIO_FLAG_GODFATHER);
5758 	}
5759 
5760 	/*
5761 	 * Create the root vdev.
5762 	 */
5763 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5764 
5765 	error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD);
5766 
5767 	ASSERT(error != 0 || rvd != NULL);
5768 	ASSERT(error != 0 || spa->spa_root_vdev == rvd);
5769 
5770 	if (error == 0 && !zfs_allocatable_devs(nvroot))
5771 		error = SET_ERROR(EINVAL);
5772 
5773 	if (error == 0 &&
5774 	    (error = vdev_create(rvd, txg, B_FALSE)) == 0 &&
5775 	    (error = vdev_draid_spare_create(nvroot, rvd, &ndraid, 0)) == 0 &&
5776 	    (error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) == 0) {
5777 		/*
5778 		 * instantiate the metaslab groups (this will dirty the vdevs)
5779 		 * we can no longer error exit past this point
5780 		 */
5781 		for (int c = 0; error == 0 && c < rvd->vdev_children; c++) {
5782 			vdev_t *vd = rvd->vdev_child[c];
5783 
5784 			vdev_metaslab_set_size(vd);
5785 			vdev_expand(vd, txg);
5786 		}
5787 	}
5788 
5789 	spa_config_exit(spa, SCL_ALL, FTAG);
5790 
5791 	if (error != 0) {
5792 		spa_unload(spa);
5793 		spa_deactivate(spa);
5794 		spa_remove(spa);
5795 		mutex_exit(&spa_namespace_lock);
5796 		return (error);
5797 	}
5798 
5799 	/*
5800 	 * Get the list of spares, if specified.
5801 	 */
5802 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5803 	    &spares, &nspares) == 0) {
5804 		VERIFY(nvlist_alloc(&spa->spa_spares.sav_config, NV_UNIQUE_NAME,
5805 		    KM_SLEEP) == 0);
5806 		VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config,
5807 		    ZPOOL_CONFIG_SPARES, spares, nspares) == 0);
5808 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5809 		spa_load_spares(spa);
5810 		spa_config_exit(spa, SCL_ALL, FTAG);
5811 		spa->spa_spares.sav_sync = B_TRUE;
5812 	}
5813 
5814 	/*
5815 	 * Get the list of level 2 cache devices, if specified.
5816 	 */
5817 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5818 	    &l2cache, &nl2cache) == 0) {
5819 		VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config,
5820 		    NV_UNIQUE_NAME, KM_SLEEP) == 0);
5821 		VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
5822 		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0);
5823 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5824 		spa_load_l2cache(spa);
5825 		spa_config_exit(spa, SCL_ALL, FTAG);
5826 		spa->spa_l2cache.sav_sync = B_TRUE;
5827 	}
5828 
5829 	spa->spa_is_initializing = B_TRUE;
5830 	spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg);
5831 	spa->spa_is_initializing = B_FALSE;
5832 
5833 	/*
5834 	 * Create DDTs (dedup tables).
5835 	 */
5836 	ddt_create(spa);
5837 
5838 	spa_update_dspace(spa);
5839 
5840 	tx = dmu_tx_create_assigned(dp, txg);
5841 
5842 	/*
5843 	 * Create the pool's history object.
5844 	 */
5845 	if (version >= SPA_VERSION_ZPOOL_HISTORY && !spa->spa_history)
5846 		spa_history_create_obj(spa, tx);
5847 
5848 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE);
5849 	spa_history_log_version(spa, "create", tx);
5850 
5851 	/*
5852 	 * Create the pool config object.
5853 	 */
5854 	spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset,
5855 	    DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE,
5856 	    DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx);
5857 
5858 	if (zap_add(spa->spa_meta_objset,
5859 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG,
5860 	    sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) {
5861 		cmn_err(CE_PANIC, "failed to add pool config");
5862 	}
5863 
5864 	if (zap_add(spa->spa_meta_objset,
5865 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION,
5866 	    sizeof (uint64_t), 1, &version, tx) != 0) {
5867 		cmn_err(CE_PANIC, "failed to add pool version");
5868 	}
5869 
5870 	/* Newly created pools with the right version are always deflated. */
5871 	if (version >= SPA_VERSION_RAIDZ_DEFLATE) {
5872 		spa->spa_deflate = TRUE;
5873 		if (zap_add(spa->spa_meta_objset,
5874 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
5875 		    sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) {
5876 			cmn_err(CE_PANIC, "failed to add deflate");
5877 		}
5878 	}
5879 
5880 	/*
5881 	 * Create the deferred-free bpobj.  Turn off compression
5882 	 * because sync-to-convergence takes longer if the blocksize
5883 	 * keeps changing.
5884 	 */
5885 	obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx);
5886 	dmu_object_set_compress(spa->spa_meta_objset, obj,
5887 	    ZIO_COMPRESS_OFF, tx);
5888 	if (zap_add(spa->spa_meta_objset,
5889 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ,
5890 	    sizeof (uint64_t), 1, &obj, tx) != 0) {
5891 		cmn_err(CE_PANIC, "failed to add bpobj");
5892 	}
5893 	VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj,
5894 	    spa->spa_meta_objset, obj));
5895 
5896 	/*
5897 	 * Generate some random noise for salted checksums to operate on.
5898 	 */
5899 	(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
5900 	    sizeof (spa->spa_cksum_salt.zcs_bytes));
5901 
5902 	/*
5903 	 * Set pool properties.
5904 	 */
5905 	spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS);
5906 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
5907 	spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE);
5908 	spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND);
5909 	spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST);
5910 	spa->spa_autotrim = zpool_prop_default_numeric(ZPOOL_PROP_AUTOTRIM);
5911 
5912 	if (props != NULL) {
5913 		spa_configfile_set(spa, props, B_FALSE);
5914 		spa_sync_props(props, tx);
5915 	}
5916 
5917 	for (int i = 0; i < ndraid; i++)
5918 		spa_feature_incr(spa, SPA_FEATURE_DRAID, tx);
5919 
5920 	dmu_tx_commit(tx);
5921 
5922 	spa->spa_sync_on = B_TRUE;
5923 	txg_sync_start(dp);
5924 	mmp_thread_start(spa);
5925 	txg_wait_synced(dp, txg);
5926 
5927 	spa_spawn_aux_threads(spa);
5928 
5929 	spa_write_cachefile(spa, B_FALSE, B_TRUE);
5930 
5931 	/*
5932 	 * Don't count references from objsets that are already closed
5933 	 * and are making their way through the eviction process.
5934 	 */
5935 	spa_evicting_os_wait(spa);
5936 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
5937 	spa->spa_load_state = SPA_LOAD_NONE;
5938 
5939 	mutex_exit(&spa_namespace_lock);
5940 
5941 	return (0);
5942 }
5943 
5944 /*
5945  * Import a non-root pool into the system.
5946  */
5947 int
5948 spa_import(char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags)
5949 {
5950 	spa_t *spa;
5951 	char *altroot = NULL;
5952 	spa_load_state_t state = SPA_LOAD_IMPORT;
5953 	zpool_load_policy_t policy;
5954 	spa_mode_t mode = spa_mode_global;
5955 	uint64_t readonly = B_FALSE;
5956 	int error;
5957 	nvlist_t *nvroot;
5958 	nvlist_t **spares, **l2cache;
5959 	uint_t nspares, nl2cache;
5960 
5961 	/*
5962 	 * If a pool with this name exists, return failure.
5963 	 */
5964 	mutex_enter(&spa_namespace_lock);
5965 	if (spa_lookup(pool) != NULL) {
5966 		mutex_exit(&spa_namespace_lock);
5967 		return (SET_ERROR(EEXIST));
5968 	}
5969 
5970 	/*
5971 	 * Create and initialize the spa structure.
5972 	 */
5973 	(void) nvlist_lookup_string(props,
5974 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
5975 	(void) nvlist_lookup_uint64(props,
5976 	    zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly);
5977 	if (readonly)
5978 		mode = SPA_MODE_READ;
5979 	spa = spa_add(pool, config, altroot);
5980 	spa->spa_import_flags = flags;
5981 
5982 	/*
5983 	 * Verbatim import - Take a pool and insert it into the namespace
5984 	 * as if it had been loaded at boot.
5985 	 */
5986 	if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) {
5987 		if (props != NULL)
5988 			spa_configfile_set(spa, props, B_FALSE);
5989 
5990 		spa_write_cachefile(spa, B_FALSE, B_TRUE);
5991 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
5992 		zfs_dbgmsg("spa_import: verbatim import of %s", pool);
5993 		mutex_exit(&spa_namespace_lock);
5994 		return (0);
5995 	}
5996 
5997 	spa_activate(spa, mode);
5998 
5999 	/*
6000 	 * Don't start async tasks until we know everything is healthy.
6001 	 */
6002 	spa_async_suspend(spa);
6003 
6004 	zpool_get_load_policy(config, &policy);
6005 	if (policy.zlp_rewind & ZPOOL_DO_REWIND)
6006 		state = SPA_LOAD_RECOVER;
6007 
6008 	spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT;
6009 
6010 	if (state != SPA_LOAD_RECOVER) {
6011 		spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
6012 		zfs_dbgmsg("spa_import: importing %s", pool);
6013 	} else {
6014 		zfs_dbgmsg("spa_import: importing %s, max_txg=%lld "
6015 		    "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg);
6016 	}
6017 	error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind);
6018 
6019 	/*
6020 	 * Propagate anything learned while loading the pool and pass it
6021 	 * back to caller (i.e. rewind info, missing devices, etc).
6022 	 */
6023 	VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO,
6024 	    spa->spa_load_info) == 0);
6025 
6026 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6027 	/*
6028 	 * Toss any existing sparelist, as it doesn't have any validity
6029 	 * anymore, and conflicts with spa_has_spare().
6030 	 */
6031 	if (spa->spa_spares.sav_config) {
6032 		nvlist_free(spa->spa_spares.sav_config);
6033 		spa->spa_spares.sav_config = NULL;
6034 		spa_load_spares(spa);
6035 	}
6036 	if (spa->spa_l2cache.sav_config) {
6037 		nvlist_free(spa->spa_l2cache.sav_config);
6038 		spa->spa_l2cache.sav_config = NULL;
6039 		spa_load_l2cache(spa);
6040 	}
6041 
6042 	VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
6043 	    &nvroot) == 0);
6044 	spa_config_exit(spa, SCL_ALL, FTAG);
6045 
6046 	if (props != NULL)
6047 		spa_configfile_set(spa, props, B_FALSE);
6048 
6049 	if (error != 0 || (props && spa_writeable(spa) &&
6050 	    (error = spa_prop_set(spa, props)))) {
6051 		spa_unload(spa);
6052 		spa_deactivate(spa);
6053 		spa_remove(spa);
6054 		mutex_exit(&spa_namespace_lock);
6055 		return (error);
6056 	}
6057 
6058 	spa_async_resume(spa);
6059 
6060 	/*
6061 	 * Override any spares and level 2 cache devices as specified by
6062 	 * the user, as these may have correct device names/devids, etc.
6063 	 */
6064 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
6065 	    &spares, &nspares) == 0) {
6066 		if (spa->spa_spares.sav_config)
6067 			VERIFY(nvlist_remove(spa->spa_spares.sav_config,
6068 			    ZPOOL_CONFIG_SPARES, DATA_TYPE_NVLIST_ARRAY) == 0);
6069 		else
6070 			VERIFY(nvlist_alloc(&spa->spa_spares.sav_config,
6071 			    NV_UNIQUE_NAME, KM_SLEEP) == 0);
6072 		VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config,
6073 		    ZPOOL_CONFIG_SPARES, spares, nspares) == 0);
6074 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6075 		spa_load_spares(spa);
6076 		spa_config_exit(spa, SCL_ALL, FTAG);
6077 		spa->spa_spares.sav_sync = B_TRUE;
6078 	}
6079 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
6080 	    &l2cache, &nl2cache) == 0) {
6081 		if (spa->spa_l2cache.sav_config)
6082 			VERIFY(nvlist_remove(spa->spa_l2cache.sav_config,
6083 			    ZPOOL_CONFIG_L2CACHE, DATA_TYPE_NVLIST_ARRAY) == 0);
6084 		else
6085 			VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config,
6086 			    NV_UNIQUE_NAME, KM_SLEEP) == 0);
6087 		VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
6088 		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0);
6089 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6090 		spa_load_l2cache(spa);
6091 		spa_config_exit(spa, SCL_ALL, FTAG);
6092 		spa->spa_l2cache.sav_sync = B_TRUE;
6093 	}
6094 
6095 	/*
6096 	 * Check for any removed devices.
6097 	 */
6098 	if (spa->spa_autoreplace) {
6099 		spa_aux_check_removed(&spa->spa_spares);
6100 		spa_aux_check_removed(&spa->spa_l2cache);
6101 	}
6102 
6103 	if (spa_writeable(spa)) {
6104 		/*
6105 		 * Update the config cache to include the newly-imported pool.
6106 		 */
6107 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
6108 	}
6109 
6110 	/*
6111 	 * It's possible that the pool was expanded while it was exported.
6112 	 * We kick off an async task to handle this for us.
6113 	 */
6114 	spa_async_request(spa, SPA_ASYNC_AUTOEXPAND);
6115 
6116 	spa_history_log_version(spa, "import", NULL);
6117 
6118 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
6119 
6120 	mutex_exit(&spa_namespace_lock);
6121 
6122 	zvol_create_minors_recursive(pool);
6123 
6124 	return (0);
6125 }
6126 
6127 nvlist_t *
6128 spa_tryimport(nvlist_t *tryconfig)
6129 {
6130 	nvlist_t *config = NULL;
6131 	char *poolname, *cachefile;
6132 	spa_t *spa;
6133 	uint64_t state;
6134 	int error;
6135 	zpool_load_policy_t policy;
6136 
6137 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname))
6138 		return (NULL);
6139 
6140 	if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state))
6141 		return (NULL);
6142 
6143 	/*
6144 	 * Create and initialize the spa structure.
6145 	 */
6146 	mutex_enter(&spa_namespace_lock);
6147 	spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL);
6148 	spa_activate(spa, SPA_MODE_READ);
6149 
6150 	/*
6151 	 * Rewind pool if a max txg was provided.
6152 	 */
6153 	zpool_get_load_policy(spa->spa_config, &policy);
6154 	if (policy.zlp_txg != UINT64_MAX) {
6155 		spa->spa_load_max_txg = policy.zlp_txg;
6156 		spa->spa_extreme_rewind = B_TRUE;
6157 		zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld",
6158 		    poolname, (longlong_t)policy.zlp_txg);
6159 	} else {
6160 		zfs_dbgmsg("spa_tryimport: importing %s", poolname);
6161 	}
6162 
6163 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile)
6164 	    == 0) {
6165 		zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile);
6166 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
6167 	} else {
6168 		spa->spa_config_source = SPA_CONFIG_SRC_SCAN;
6169 	}
6170 
6171 	error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING);
6172 
6173 	/*
6174 	 * If 'tryconfig' was at least parsable, return the current config.
6175 	 */
6176 	if (spa->spa_root_vdev != NULL) {
6177 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
6178 		VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME,
6179 		    poolname) == 0);
6180 		VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE,
6181 		    state) == 0);
6182 		VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
6183 		    spa->spa_uberblock.ub_timestamp) == 0);
6184 		VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO,
6185 		    spa->spa_load_info) == 0);
6186 		VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_ERRATA,
6187 		    spa->spa_errata) == 0);
6188 
6189 		/*
6190 		 * If the bootfs property exists on this pool then we
6191 		 * copy it out so that external consumers can tell which
6192 		 * pools are bootable.
6193 		 */
6194 		if ((!error || error == EEXIST) && spa->spa_bootfs) {
6195 			char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
6196 
6197 			/*
6198 			 * We have to play games with the name since the
6199 			 * pool was opened as TRYIMPORT_NAME.
6200 			 */
6201 			if (dsl_dsobj_to_dsname(spa_name(spa),
6202 			    spa->spa_bootfs, tmpname) == 0) {
6203 				char *cp;
6204 				char *dsname;
6205 
6206 				dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
6207 
6208 				cp = strchr(tmpname, '/');
6209 				if (cp == NULL) {
6210 					(void) strlcpy(dsname, tmpname,
6211 					    MAXPATHLEN);
6212 				} else {
6213 					(void) snprintf(dsname, MAXPATHLEN,
6214 					    "%s/%s", poolname, ++cp);
6215 				}
6216 				VERIFY(nvlist_add_string(config,
6217 				    ZPOOL_CONFIG_BOOTFS, dsname) == 0);
6218 				kmem_free(dsname, MAXPATHLEN);
6219 			}
6220 			kmem_free(tmpname, MAXPATHLEN);
6221 		}
6222 
6223 		/*
6224 		 * Add the list of hot spares and level 2 cache devices.
6225 		 */
6226 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
6227 		spa_add_spares(spa, config);
6228 		spa_add_l2cache(spa, config);
6229 		spa_config_exit(spa, SCL_CONFIG, FTAG);
6230 	}
6231 
6232 	spa_unload(spa);
6233 	spa_deactivate(spa);
6234 	spa_remove(spa);
6235 	mutex_exit(&spa_namespace_lock);
6236 
6237 	return (config);
6238 }
6239 
6240 /*
6241  * Pool export/destroy
6242  *
6243  * The act of destroying or exporting a pool is very simple.  We make sure there
6244  * is no more pending I/O and any references to the pool are gone.  Then, we
6245  * update the pool state and sync all the labels to disk, removing the
6246  * configuration from the cache afterwards. If the 'hardforce' flag is set, then
6247  * we don't sync the labels or remove the configuration cache.
6248  */
6249 static int
6250 spa_export_common(const char *pool, int new_state, nvlist_t **oldconfig,
6251     boolean_t force, boolean_t hardforce)
6252 {
6253 	int error;
6254 	spa_t *spa;
6255 
6256 	if (oldconfig)
6257 		*oldconfig = NULL;
6258 
6259 	if (!(spa_mode_global & SPA_MODE_WRITE))
6260 		return (SET_ERROR(EROFS));
6261 
6262 	mutex_enter(&spa_namespace_lock);
6263 	if ((spa = spa_lookup(pool)) == NULL) {
6264 		mutex_exit(&spa_namespace_lock);
6265 		return (SET_ERROR(ENOENT));
6266 	}
6267 
6268 	if (spa->spa_is_exporting) {
6269 		/* the pool is being exported by another thread */
6270 		mutex_exit(&spa_namespace_lock);
6271 		return (SET_ERROR(ZFS_ERR_EXPORT_IN_PROGRESS));
6272 	}
6273 	spa->spa_is_exporting = B_TRUE;
6274 
6275 	/*
6276 	 * Put a hold on the pool, drop the namespace lock, stop async tasks,
6277 	 * reacquire the namespace lock, and see if we can export.
6278 	 */
6279 	spa_open_ref(spa, FTAG);
6280 	mutex_exit(&spa_namespace_lock);
6281 	spa_async_suspend(spa);
6282 	if (spa->spa_zvol_taskq) {
6283 		zvol_remove_minors(spa, spa_name(spa), B_TRUE);
6284 		taskq_wait(spa->spa_zvol_taskq);
6285 	}
6286 	mutex_enter(&spa_namespace_lock);
6287 	spa_close(spa, FTAG);
6288 
6289 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
6290 		goto export_spa;
6291 	/*
6292 	 * The pool will be in core if it's openable, in which case we can
6293 	 * modify its state.  Objsets may be open only because they're dirty,
6294 	 * so we have to force it to sync before checking spa_refcnt.
6295 	 */
6296 	if (spa->spa_sync_on) {
6297 		txg_wait_synced(spa->spa_dsl_pool, 0);
6298 		spa_evicting_os_wait(spa);
6299 	}
6300 
6301 	/*
6302 	 * A pool cannot be exported or destroyed if there are active
6303 	 * references.  If we are resetting a pool, allow references by
6304 	 * fault injection handlers.
6305 	 */
6306 	if (!spa_refcount_zero(spa) || (spa->spa_inject_ref != 0)) {
6307 		error = SET_ERROR(EBUSY);
6308 		goto fail;
6309 	}
6310 
6311 	if (spa->spa_sync_on) {
6312 		/*
6313 		 * A pool cannot be exported if it has an active shared spare.
6314 		 * This is to prevent other pools stealing the active spare
6315 		 * from an exported pool. At user's own will, such pool can
6316 		 * be forcedly exported.
6317 		 */
6318 		if (!force && new_state == POOL_STATE_EXPORTED &&
6319 		    spa_has_active_shared_spare(spa)) {
6320 			error = SET_ERROR(EXDEV);
6321 			goto fail;
6322 		}
6323 
6324 		/*
6325 		 * We're about to export or destroy this pool. Make sure
6326 		 * we stop all initialization and trim activity here before
6327 		 * we set the spa_final_txg. This will ensure that all
6328 		 * dirty data resulting from the initialization is
6329 		 * committed to disk before we unload the pool.
6330 		 */
6331 		if (spa->spa_root_vdev != NULL) {
6332 			vdev_t *rvd = spa->spa_root_vdev;
6333 			vdev_initialize_stop_all(rvd, VDEV_INITIALIZE_ACTIVE);
6334 			vdev_trim_stop_all(rvd, VDEV_TRIM_ACTIVE);
6335 			vdev_autotrim_stop_all(spa);
6336 			vdev_rebuild_stop_all(spa);
6337 		}
6338 
6339 		/*
6340 		 * We want this to be reflected on every label,
6341 		 * so mark them all dirty.  spa_unload() will do the
6342 		 * final sync that pushes these changes out.
6343 		 */
6344 		if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) {
6345 			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6346 			spa->spa_state = new_state;
6347 			spa->spa_final_txg = spa_last_synced_txg(spa) +
6348 			    TXG_DEFER_SIZE + 1;
6349 			vdev_config_dirty(spa->spa_root_vdev);
6350 			spa_config_exit(spa, SCL_ALL, FTAG);
6351 		}
6352 	}
6353 
6354 export_spa:
6355 	if (new_state == POOL_STATE_DESTROYED)
6356 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY);
6357 	else if (new_state == POOL_STATE_EXPORTED)
6358 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_EXPORT);
6359 
6360 	if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
6361 		spa_unload(spa);
6362 		spa_deactivate(spa);
6363 	}
6364 
6365 	if (oldconfig && spa->spa_config)
6366 		VERIFY(nvlist_dup(spa->spa_config, oldconfig, 0) == 0);
6367 
6368 	if (new_state != POOL_STATE_UNINITIALIZED) {
6369 		if (!hardforce)
6370 			spa_write_cachefile(spa, B_TRUE, B_TRUE);
6371 		spa_remove(spa);
6372 	} else {
6373 		/*
6374 		 * If spa_remove() is not called for this spa_t and
6375 		 * there is any possibility that it can be reused,
6376 		 * we make sure to reset the exporting flag.
6377 		 */
6378 		spa->spa_is_exporting = B_FALSE;
6379 	}
6380 
6381 	mutex_exit(&spa_namespace_lock);
6382 	return (0);
6383 
6384 fail:
6385 	spa->spa_is_exporting = B_FALSE;
6386 	spa_async_resume(spa);
6387 	mutex_exit(&spa_namespace_lock);
6388 	return (error);
6389 }
6390 
6391 /*
6392  * Destroy a storage pool.
6393  */
6394 int
6395 spa_destroy(const char *pool)
6396 {
6397 	return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL,
6398 	    B_FALSE, B_FALSE));
6399 }
6400 
6401 /*
6402  * Export a storage pool.
6403  */
6404 int
6405 spa_export(const char *pool, nvlist_t **oldconfig, boolean_t force,
6406     boolean_t hardforce)
6407 {
6408 	return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig,
6409 	    force, hardforce));
6410 }
6411 
6412 /*
6413  * Similar to spa_export(), this unloads the spa_t without actually removing it
6414  * from the namespace in any way.
6415  */
6416 int
6417 spa_reset(const char *pool)
6418 {
6419 	return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL,
6420 	    B_FALSE, B_FALSE));
6421 }
6422 
6423 /*
6424  * ==========================================================================
6425  * Device manipulation
6426  * ==========================================================================
6427  */
6428 
6429 /*
6430  * This is called as a synctask to increment the draid feature flag
6431  */
6432 static void
6433 spa_draid_feature_incr(void *arg, dmu_tx_t *tx)
6434 {
6435 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
6436 	int draid = (int)(uintptr_t)arg;
6437 
6438 	for (int c = 0; c < draid; c++)
6439 		spa_feature_incr(spa, SPA_FEATURE_DRAID, tx);
6440 }
6441 
6442 /*
6443  * Add a device to a storage pool.
6444  */
6445 int
6446 spa_vdev_add(spa_t *spa, nvlist_t *nvroot)
6447 {
6448 	uint64_t txg, ndraid = 0;
6449 	int error;
6450 	vdev_t *rvd = spa->spa_root_vdev;
6451 	vdev_t *vd, *tvd;
6452 	nvlist_t **spares, **l2cache;
6453 	uint_t nspares, nl2cache;
6454 
6455 	ASSERT(spa_writeable(spa));
6456 
6457 	txg = spa_vdev_enter(spa);
6458 
6459 	if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0,
6460 	    VDEV_ALLOC_ADD)) != 0)
6461 		return (spa_vdev_exit(spa, NULL, txg, error));
6462 
6463 	spa->spa_pending_vdev = vd;	/* spa_vdev_exit() will clear this */
6464 
6465 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares,
6466 	    &nspares) != 0)
6467 		nspares = 0;
6468 
6469 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache,
6470 	    &nl2cache) != 0)
6471 		nl2cache = 0;
6472 
6473 	if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0)
6474 		return (spa_vdev_exit(spa, vd, txg, EINVAL));
6475 
6476 	if (vd->vdev_children != 0 &&
6477 	    (error = vdev_create(vd, txg, B_FALSE)) != 0) {
6478 		return (spa_vdev_exit(spa, vd, txg, error));
6479 	}
6480 
6481 	/*
6482 	 * The virtual dRAID spares must be added after vdev tree is created
6483 	 * and the vdev guids are generated.  The guid of their assoicated
6484 	 * dRAID is stored in the config and used when opening the spare.
6485 	 */
6486 	if ((error = vdev_draid_spare_create(nvroot, vd, &ndraid,
6487 	    rvd->vdev_children)) == 0) {
6488 		if (ndraid > 0 && nvlist_lookup_nvlist_array(nvroot,
6489 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) != 0)
6490 			nspares = 0;
6491 	} else {
6492 		return (spa_vdev_exit(spa, vd, txg, error));
6493 	}
6494 
6495 	/*
6496 	 * We must validate the spares and l2cache devices after checking the
6497 	 * children.  Otherwise, vdev_inuse() will blindly overwrite the spare.
6498 	 */
6499 	if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0)
6500 		return (spa_vdev_exit(spa, vd, txg, error));
6501 
6502 	/*
6503 	 * If we are in the middle of a device removal, we can only add
6504 	 * devices which match the existing devices in the pool.
6505 	 * If we are in the middle of a removal, or have some indirect
6506 	 * vdevs, we can not add raidz or dRAID top levels.
6507 	 */
6508 	if (spa->spa_vdev_removal != NULL ||
6509 	    spa->spa_removing_phys.sr_prev_indirect_vdev != -1) {
6510 		for (int c = 0; c < vd->vdev_children; c++) {
6511 			tvd = vd->vdev_child[c];
6512 			if (spa->spa_vdev_removal != NULL &&
6513 			    tvd->vdev_ashift != spa->spa_max_ashift) {
6514 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
6515 			}
6516 			/* Fail if top level vdev is raidz or a dRAID */
6517 			if (vdev_get_nparity(tvd) != 0)
6518 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
6519 
6520 			/*
6521 			 * Need the top level mirror to be
6522 			 * a mirror of leaf vdevs only
6523 			 */
6524 			if (tvd->vdev_ops == &vdev_mirror_ops) {
6525 				for (uint64_t cid = 0;
6526 				    cid < tvd->vdev_children; cid++) {
6527 					vdev_t *cvd = tvd->vdev_child[cid];
6528 					if (!cvd->vdev_ops->vdev_op_leaf) {
6529 						return (spa_vdev_exit(spa, vd,
6530 						    txg, EINVAL));
6531 					}
6532 				}
6533 			}
6534 		}
6535 	}
6536 
6537 	for (int c = 0; c < vd->vdev_children; c++) {
6538 		tvd = vd->vdev_child[c];
6539 		vdev_remove_child(vd, tvd);
6540 		tvd->vdev_id = rvd->vdev_children;
6541 		vdev_add_child(rvd, tvd);
6542 		vdev_config_dirty(tvd);
6543 	}
6544 
6545 	if (nspares != 0) {
6546 		spa_set_aux_vdevs(&spa->spa_spares, spares, nspares,
6547 		    ZPOOL_CONFIG_SPARES);
6548 		spa_load_spares(spa);
6549 		spa->spa_spares.sav_sync = B_TRUE;
6550 	}
6551 
6552 	if (nl2cache != 0) {
6553 		spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache,
6554 		    ZPOOL_CONFIG_L2CACHE);
6555 		spa_load_l2cache(spa);
6556 		spa->spa_l2cache.sav_sync = B_TRUE;
6557 	}
6558 
6559 	/*
6560 	 * We can't increment a feature while holding spa_vdev so we
6561 	 * have to do it in a synctask.
6562 	 */
6563 	if (ndraid != 0) {
6564 		dmu_tx_t *tx;
6565 
6566 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
6567 		dsl_sync_task_nowait(spa->spa_dsl_pool, spa_draid_feature_incr,
6568 		    (void *)(uintptr_t)ndraid, tx);
6569 		dmu_tx_commit(tx);
6570 	}
6571 
6572 	/*
6573 	 * We have to be careful when adding new vdevs to an existing pool.
6574 	 * If other threads start allocating from these vdevs before we
6575 	 * sync the config cache, and we lose power, then upon reboot we may
6576 	 * fail to open the pool because there are DVAs that the config cache
6577 	 * can't translate.  Therefore, we first add the vdevs without
6578 	 * initializing metaslabs; sync the config cache (via spa_vdev_exit());
6579 	 * and then let spa_config_update() initialize the new metaslabs.
6580 	 *
6581 	 * spa_load() checks for added-but-not-initialized vdevs, so that
6582 	 * if we lose power at any point in this sequence, the remaining
6583 	 * steps will be completed the next time we load the pool.
6584 	 */
6585 	(void) spa_vdev_exit(spa, vd, txg, 0);
6586 
6587 	mutex_enter(&spa_namespace_lock);
6588 	spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
6589 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD);
6590 	mutex_exit(&spa_namespace_lock);
6591 
6592 	return (0);
6593 }
6594 
6595 /*
6596  * Attach a device to a mirror.  The arguments are the path to any device
6597  * in the mirror, and the nvroot for the new device.  If the path specifies
6598  * a device that is not mirrored, we automatically insert the mirror vdev.
6599  *
6600  * If 'replacing' is specified, the new device is intended to replace the
6601  * existing device; in this case the two devices are made into their own
6602  * mirror using the 'replacing' vdev, which is functionally identical to
6603  * the mirror vdev (it actually reuses all the same ops) but has a few
6604  * extra rules: you can't attach to it after it's been created, and upon
6605  * completion of resilvering, the first disk (the one being replaced)
6606  * is automatically detached.
6607  *
6608  * If 'rebuild' is specified, then sequential reconstruction (a.ka. rebuild)
6609  * should be performed instead of traditional healing reconstruction.  From
6610  * an administrators perspective these are both resilver operations.
6611  */
6612 int
6613 spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing,
6614     int rebuild)
6615 {
6616 	uint64_t txg, dtl_max_txg;
6617 	vdev_t *rvd = spa->spa_root_vdev;
6618 	vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd;
6619 	vdev_ops_t *pvops;
6620 	char *oldvdpath, *newvdpath;
6621 	int newvd_isspare;
6622 	int error;
6623 
6624 	ASSERT(spa_writeable(spa));
6625 
6626 	txg = spa_vdev_enter(spa);
6627 
6628 	oldvd = spa_lookup_by_guid(spa, guid, B_FALSE);
6629 
6630 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
6631 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6632 		error = (spa_has_checkpoint(spa)) ?
6633 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6634 		return (spa_vdev_exit(spa, NULL, txg, error));
6635 	}
6636 
6637 	if (rebuild) {
6638 		if (!spa_feature_is_enabled(spa, SPA_FEATURE_DEVICE_REBUILD))
6639 			return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6640 
6641 		if (dsl_scan_resilvering(spa_get_dsl(spa)))
6642 			return (spa_vdev_exit(spa, NULL, txg,
6643 			    ZFS_ERR_RESILVER_IN_PROGRESS));
6644 	} else {
6645 		if (vdev_rebuild_active(rvd))
6646 			return (spa_vdev_exit(spa, NULL, txg,
6647 			    ZFS_ERR_REBUILD_IN_PROGRESS));
6648 	}
6649 
6650 	if (spa->spa_vdev_removal != NULL)
6651 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6652 
6653 	if (oldvd == NULL)
6654 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
6655 
6656 	if (!oldvd->vdev_ops->vdev_op_leaf)
6657 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6658 
6659 	pvd = oldvd->vdev_parent;
6660 
6661 	if ((error = spa_config_parse(spa, &newrootvd, nvroot, NULL, 0,
6662 	    VDEV_ALLOC_ATTACH)) != 0)
6663 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6664 
6665 	if (newrootvd->vdev_children != 1)
6666 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6667 
6668 	newvd = newrootvd->vdev_child[0];
6669 
6670 	if (!newvd->vdev_ops->vdev_op_leaf)
6671 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6672 
6673 	if ((error = vdev_create(newrootvd, txg, replacing)) != 0)
6674 		return (spa_vdev_exit(spa, newrootvd, txg, error));
6675 
6676 	/*
6677 	 * Spares can't replace logs
6678 	 */
6679 	if (oldvd->vdev_top->vdev_islog && newvd->vdev_isspare)
6680 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6681 
6682 	/*
6683 	 * A dRAID spare can only replace a child of its parent dRAID vdev.
6684 	 */
6685 	if (newvd->vdev_ops == &vdev_draid_spare_ops &&
6686 	    oldvd->vdev_top != vdev_draid_spare_get_parent(newvd)) {
6687 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6688 	}
6689 
6690 	if (rebuild) {
6691 		/*
6692 		 * For rebuilds, the top vdev must support reconstruction
6693 		 * using only space maps.  This means the only allowable
6694 		 * vdevs types are the root vdev, a mirror, or dRAID.
6695 		 */
6696 		tvd = pvd;
6697 		if (pvd->vdev_top != NULL)
6698 			tvd = pvd->vdev_top;
6699 
6700 		if (tvd->vdev_ops != &vdev_mirror_ops &&
6701 		    tvd->vdev_ops != &vdev_root_ops &&
6702 		    tvd->vdev_ops != &vdev_draid_ops) {
6703 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6704 		}
6705 	}
6706 
6707 	if (!replacing) {
6708 		/*
6709 		 * For attach, the only allowable parent is a mirror or the root
6710 		 * vdev.
6711 		 */
6712 		if (pvd->vdev_ops != &vdev_mirror_ops &&
6713 		    pvd->vdev_ops != &vdev_root_ops)
6714 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6715 
6716 		pvops = &vdev_mirror_ops;
6717 	} else {
6718 		/*
6719 		 * Active hot spares can only be replaced by inactive hot
6720 		 * spares.
6721 		 */
6722 		if (pvd->vdev_ops == &vdev_spare_ops &&
6723 		    oldvd->vdev_isspare &&
6724 		    !spa_has_spare(spa, newvd->vdev_guid))
6725 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6726 
6727 		/*
6728 		 * If the source is a hot spare, and the parent isn't already a
6729 		 * spare, then we want to create a new hot spare.  Otherwise, we
6730 		 * want to create a replacing vdev.  The user is not allowed to
6731 		 * attach to a spared vdev child unless the 'isspare' state is
6732 		 * the same (spare replaces spare, non-spare replaces
6733 		 * non-spare).
6734 		 */
6735 		if (pvd->vdev_ops == &vdev_replacing_ops &&
6736 		    spa_version(spa) < SPA_VERSION_MULTI_REPLACE) {
6737 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6738 		} else if (pvd->vdev_ops == &vdev_spare_ops &&
6739 		    newvd->vdev_isspare != oldvd->vdev_isspare) {
6740 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6741 		}
6742 
6743 		if (newvd->vdev_isspare)
6744 			pvops = &vdev_spare_ops;
6745 		else
6746 			pvops = &vdev_replacing_ops;
6747 	}
6748 
6749 	/*
6750 	 * Make sure the new device is big enough.
6751 	 */
6752 	if (newvd->vdev_asize < vdev_get_min_asize(oldvd))
6753 		return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW));
6754 
6755 	/*
6756 	 * The new device cannot have a higher alignment requirement
6757 	 * than the top-level vdev.
6758 	 */
6759 	if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift)
6760 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6761 
6762 	/*
6763 	 * If this is an in-place replacement, update oldvd's path and devid
6764 	 * to make it distinguishable from newvd, and unopenable from now on.
6765 	 */
6766 	if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) {
6767 		spa_strfree(oldvd->vdev_path);
6768 		oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5,
6769 		    KM_SLEEP);
6770 		(void) snprintf(oldvd->vdev_path, strlen(newvd->vdev_path) + 5,
6771 		    "%s/%s", newvd->vdev_path, "old");
6772 		if (oldvd->vdev_devid != NULL) {
6773 			spa_strfree(oldvd->vdev_devid);
6774 			oldvd->vdev_devid = NULL;
6775 		}
6776 	}
6777 
6778 	/*
6779 	 * If the parent is not a mirror, or if we're replacing, insert the new
6780 	 * mirror/replacing/spare vdev above oldvd.
6781 	 */
6782 	if (pvd->vdev_ops != pvops)
6783 		pvd = vdev_add_parent(oldvd, pvops);
6784 
6785 	ASSERT(pvd->vdev_top->vdev_parent == rvd);
6786 	ASSERT(pvd->vdev_ops == pvops);
6787 	ASSERT(oldvd->vdev_parent == pvd);
6788 
6789 	/*
6790 	 * Extract the new device from its root and add it to pvd.
6791 	 */
6792 	vdev_remove_child(newrootvd, newvd);
6793 	newvd->vdev_id = pvd->vdev_children;
6794 	newvd->vdev_crtxg = oldvd->vdev_crtxg;
6795 	vdev_add_child(pvd, newvd);
6796 
6797 	/*
6798 	 * Reevaluate the parent vdev state.
6799 	 */
6800 	vdev_propagate_state(pvd);
6801 
6802 	tvd = newvd->vdev_top;
6803 	ASSERT(pvd->vdev_top == tvd);
6804 	ASSERT(tvd->vdev_parent == rvd);
6805 
6806 	vdev_config_dirty(tvd);
6807 
6808 	/*
6809 	 * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account
6810 	 * for any dmu_sync-ed blocks.  It will propagate upward when
6811 	 * spa_vdev_exit() calls vdev_dtl_reassess().
6812 	 */
6813 	dtl_max_txg = txg + TXG_CONCURRENT_STATES;
6814 
6815 	vdev_dtl_dirty(newvd, DTL_MISSING,
6816 	    TXG_INITIAL, dtl_max_txg - TXG_INITIAL);
6817 
6818 	if (newvd->vdev_isspare) {
6819 		spa_spare_activate(newvd);
6820 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE);
6821 	}
6822 
6823 	oldvdpath = spa_strdup(oldvd->vdev_path);
6824 	newvdpath = spa_strdup(newvd->vdev_path);
6825 	newvd_isspare = newvd->vdev_isspare;
6826 
6827 	/*
6828 	 * Mark newvd's DTL dirty in this txg.
6829 	 */
6830 	vdev_dirty(tvd, VDD_DTL, newvd, txg);
6831 
6832 	/*
6833 	 * Schedule the resilver or rebuild to restart in the future. We do
6834 	 * this to ensure that dmu_sync-ed blocks have been stitched into the
6835 	 * respective datasets.
6836 	 */
6837 	if (rebuild) {
6838 		newvd->vdev_rebuild_txg = txg;
6839 
6840 		vdev_rebuild(tvd);
6841 	} else {
6842 		newvd->vdev_resilver_txg = txg;
6843 
6844 		if (dsl_scan_resilvering(spa_get_dsl(spa)) &&
6845 		    spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER)) {
6846 			vdev_defer_resilver(newvd);
6847 		} else {
6848 			dsl_scan_restart_resilver(spa->spa_dsl_pool,
6849 			    dtl_max_txg);
6850 		}
6851 	}
6852 
6853 	if (spa->spa_bootfs)
6854 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH);
6855 
6856 	spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH);
6857 
6858 	/*
6859 	 * Commit the config
6860 	 */
6861 	(void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0);
6862 
6863 	spa_history_log_internal(spa, "vdev attach", NULL,
6864 	    "%s vdev=%s %s vdev=%s",
6865 	    replacing && newvd_isspare ? "spare in" :
6866 	    replacing ? "replace" : "attach", newvdpath,
6867 	    replacing ? "for" : "to", oldvdpath);
6868 
6869 	spa_strfree(oldvdpath);
6870 	spa_strfree(newvdpath);
6871 
6872 	return (0);
6873 }
6874 
6875 /*
6876  * Detach a device from a mirror or replacing vdev.
6877  *
6878  * If 'replace_done' is specified, only detach if the parent
6879  * is a replacing vdev.
6880  */
6881 int
6882 spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done)
6883 {
6884 	uint64_t txg;
6885 	int error;
6886 	vdev_t *rvd __maybe_unused = spa->spa_root_vdev;
6887 	vdev_t *vd, *pvd, *cvd, *tvd;
6888 	boolean_t unspare = B_FALSE;
6889 	uint64_t unspare_guid = 0;
6890 	char *vdpath;
6891 
6892 	ASSERT(spa_writeable(spa));
6893 
6894 	txg = spa_vdev_detach_enter(spa, guid);
6895 
6896 	vd = spa_lookup_by_guid(spa, guid, B_FALSE);
6897 
6898 	/*
6899 	 * Besides being called directly from the userland through the
6900 	 * ioctl interface, spa_vdev_detach() can be potentially called
6901 	 * at the end of spa_vdev_resilver_done().
6902 	 *
6903 	 * In the regular case, when we have a checkpoint this shouldn't
6904 	 * happen as we never empty the DTLs of a vdev during the scrub
6905 	 * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done()
6906 	 * should never get here when we have a checkpoint.
6907 	 *
6908 	 * That said, even in a case when we checkpoint the pool exactly
6909 	 * as spa_vdev_resilver_done() calls this function everything
6910 	 * should be fine as the resilver will return right away.
6911 	 */
6912 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
6913 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6914 		error = (spa_has_checkpoint(spa)) ?
6915 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6916 		return (spa_vdev_exit(spa, NULL, txg, error));
6917 	}
6918 
6919 	if (vd == NULL)
6920 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
6921 
6922 	if (!vd->vdev_ops->vdev_op_leaf)
6923 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6924 
6925 	pvd = vd->vdev_parent;
6926 
6927 	/*
6928 	 * If the parent/child relationship is not as expected, don't do it.
6929 	 * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing
6930 	 * vdev that's replacing B with C.  The user's intent in replacing
6931 	 * is to go from M(A,B) to M(A,C).  If the user decides to cancel
6932 	 * the replace by detaching C, the expected behavior is to end up
6933 	 * M(A,B).  But suppose that right after deciding to detach C,
6934 	 * the replacement of B completes.  We would have M(A,C), and then
6935 	 * ask to detach C, which would leave us with just A -- not what
6936 	 * the user wanted.  To prevent this, we make sure that the
6937 	 * parent/child relationship hasn't changed -- in this example,
6938 	 * that C's parent is still the replacing vdev R.
6939 	 */
6940 	if (pvd->vdev_guid != pguid && pguid != 0)
6941 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6942 
6943 	/*
6944 	 * Only 'replacing' or 'spare' vdevs can be replaced.
6945 	 */
6946 	if (replace_done && pvd->vdev_ops != &vdev_replacing_ops &&
6947 	    pvd->vdev_ops != &vdev_spare_ops)
6948 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6949 
6950 	ASSERT(pvd->vdev_ops != &vdev_spare_ops ||
6951 	    spa_version(spa) >= SPA_VERSION_SPARES);
6952 
6953 	/*
6954 	 * Only mirror, replacing, and spare vdevs support detach.
6955 	 */
6956 	if (pvd->vdev_ops != &vdev_replacing_ops &&
6957 	    pvd->vdev_ops != &vdev_mirror_ops &&
6958 	    pvd->vdev_ops != &vdev_spare_ops)
6959 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6960 
6961 	/*
6962 	 * If this device has the only valid copy of some data,
6963 	 * we cannot safely detach it.
6964 	 */
6965 	if (vdev_dtl_required(vd))
6966 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6967 
6968 	ASSERT(pvd->vdev_children >= 2);
6969 
6970 	/*
6971 	 * If we are detaching the second disk from a replacing vdev, then
6972 	 * check to see if we changed the original vdev's path to have "/old"
6973 	 * at the end in spa_vdev_attach().  If so, undo that change now.
6974 	 */
6975 	if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 &&
6976 	    vd->vdev_path != NULL) {
6977 		size_t len = strlen(vd->vdev_path);
6978 
6979 		for (int c = 0; c < pvd->vdev_children; c++) {
6980 			cvd = pvd->vdev_child[c];
6981 
6982 			if (cvd == vd || cvd->vdev_path == NULL)
6983 				continue;
6984 
6985 			if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 &&
6986 			    strcmp(cvd->vdev_path + len, "/old") == 0) {
6987 				spa_strfree(cvd->vdev_path);
6988 				cvd->vdev_path = spa_strdup(vd->vdev_path);
6989 				break;
6990 			}
6991 		}
6992 	}
6993 
6994 	/*
6995 	 * If we are detaching the original disk from a normal spare, then it
6996 	 * implies that the spare should become a real disk, and be removed
6997 	 * from the active spare list for the pool.  dRAID spares on the
6998 	 * other hand are coupled to the pool and thus should never be removed
6999 	 * from the spares list.
7000 	 */
7001 	if (pvd->vdev_ops == &vdev_spare_ops && vd->vdev_id == 0) {
7002 		vdev_t *last_cvd = pvd->vdev_child[pvd->vdev_children - 1];
7003 
7004 		if (last_cvd->vdev_isspare &&
7005 		    last_cvd->vdev_ops != &vdev_draid_spare_ops) {
7006 			unspare = B_TRUE;
7007 		}
7008 	}
7009 
7010 	/*
7011 	 * Erase the disk labels so the disk can be used for other things.
7012 	 * This must be done after all other error cases are handled,
7013 	 * but before we disembowel vd (so we can still do I/O to it).
7014 	 * But if we can't do it, don't treat the error as fatal --
7015 	 * it may be that the unwritability of the disk is the reason
7016 	 * it's being detached!
7017 	 */
7018 	error = vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
7019 
7020 	/*
7021 	 * Remove vd from its parent and compact the parent's children.
7022 	 */
7023 	vdev_remove_child(pvd, vd);
7024 	vdev_compact_children(pvd);
7025 
7026 	/*
7027 	 * Remember one of the remaining children so we can get tvd below.
7028 	 */
7029 	cvd = pvd->vdev_child[pvd->vdev_children - 1];
7030 
7031 	/*
7032 	 * If we need to remove the remaining child from the list of hot spares,
7033 	 * do it now, marking the vdev as no longer a spare in the process.
7034 	 * We must do this before vdev_remove_parent(), because that can
7035 	 * change the GUID if it creates a new toplevel GUID.  For a similar
7036 	 * reason, we must remove the spare now, in the same txg as the detach;
7037 	 * otherwise someone could attach a new sibling, change the GUID, and
7038 	 * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail.
7039 	 */
7040 	if (unspare) {
7041 		ASSERT(cvd->vdev_isspare);
7042 		spa_spare_remove(cvd);
7043 		unspare_guid = cvd->vdev_guid;
7044 		(void) spa_vdev_remove(spa, unspare_guid, B_TRUE);
7045 		cvd->vdev_unspare = B_TRUE;
7046 	}
7047 
7048 	/*
7049 	 * If the parent mirror/replacing vdev only has one child,
7050 	 * the parent is no longer needed.  Remove it from the tree.
7051 	 */
7052 	if (pvd->vdev_children == 1) {
7053 		if (pvd->vdev_ops == &vdev_spare_ops)
7054 			cvd->vdev_unspare = B_FALSE;
7055 		vdev_remove_parent(cvd);
7056 	}
7057 
7058 	/*
7059 	 * We don't set tvd until now because the parent we just removed
7060 	 * may have been the previous top-level vdev.
7061 	 */
7062 	tvd = cvd->vdev_top;
7063 	ASSERT(tvd->vdev_parent == rvd);
7064 
7065 	/*
7066 	 * Reevaluate the parent vdev state.
7067 	 */
7068 	vdev_propagate_state(cvd);
7069 
7070 	/*
7071 	 * If the 'autoexpand' property is set on the pool then automatically
7072 	 * try to expand the size of the pool. For example if the device we
7073 	 * just detached was smaller than the others, it may be possible to
7074 	 * add metaslabs (i.e. grow the pool). We need to reopen the vdev
7075 	 * first so that we can obtain the updated sizes of the leaf vdevs.
7076 	 */
7077 	if (spa->spa_autoexpand) {
7078 		vdev_reopen(tvd);
7079 		vdev_expand(tvd, txg);
7080 	}
7081 
7082 	vdev_config_dirty(tvd);
7083 
7084 	/*
7085 	 * Mark vd's DTL as dirty in this txg.  vdev_dtl_sync() will see that
7086 	 * vd->vdev_detached is set and free vd's DTL object in syncing context.
7087 	 * But first make sure we're not on any *other* txg's DTL list, to
7088 	 * prevent vd from being accessed after it's freed.
7089 	 */
7090 	vdpath = spa_strdup(vd->vdev_path ? vd->vdev_path : "none");
7091 	for (int t = 0; t < TXG_SIZE; t++)
7092 		(void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t);
7093 	vd->vdev_detached = B_TRUE;
7094 	vdev_dirty(tvd, VDD_DTL, vd, txg);
7095 
7096 	spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE);
7097 	spa_notify_waiters(spa);
7098 
7099 	/* hang on to the spa before we release the lock */
7100 	spa_open_ref(spa, FTAG);
7101 
7102 	error = spa_vdev_exit(spa, vd, txg, 0);
7103 
7104 	spa_history_log_internal(spa, "detach", NULL,
7105 	    "vdev=%s", vdpath);
7106 	spa_strfree(vdpath);
7107 
7108 	/*
7109 	 * If this was the removal of the original device in a hot spare vdev,
7110 	 * then we want to go through and remove the device from the hot spare
7111 	 * list of every other pool.
7112 	 */
7113 	if (unspare) {
7114 		spa_t *altspa = NULL;
7115 
7116 		mutex_enter(&spa_namespace_lock);
7117 		while ((altspa = spa_next(altspa)) != NULL) {
7118 			if (altspa->spa_state != POOL_STATE_ACTIVE ||
7119 			    altspa == spa)
7120 				continue;
7121 
7122 			spa_open_ref(altspa, FTAG);
7123 			mutex_exit(&spa_namespace_lock);
7124 			(void) spa_vdev_remove(altspa, unspare_guid, B_TRUE);
7125 			mutex_enter(&spa_namespace_lock);
7126 			spa_close(altspa, FTAG);
7127 		}
7128 		mutex_exit(&spa_namespace_lock);
7129 
7130 		/* search the rest of the vdevs for spares to remove */
7131 		spa_vdev_resilver_done(spa);
7132 	}
7133 
7134 	/* all done with the spa; OK to release */
7135 	mutex_enter(&spa_namespace_lock);
7136 	spa_close(spa, FTAG);
7137 	mutex_exit(&spa_namespace_lock);
7138 
7139 	return (error);
7140 }
7141 
7142 static int
7143 spa_vdev_initialize_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
7144     list_t *vd_list)
7145 {
7146 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7147 
7148 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
7149 
7150 	/* Look up vdev and ensure it's a leaf. */
7151 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7152 	if (vd == NULL || vd->vdev_detached) {
7153 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7154 		return (SET_ERROR(ENODEV));
7155 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
7156 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7157 		return (SET_ERROR(EINVAL));
7158 	} else if (!vdev_writeable(vd)) {
7159 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7160 		return (SET_ERROR(EROFS));
7161 	}
7162 	mutex_enter(&vd->vdev_initialize_lock);
7163 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7164 
7165 	/*
7166 	 * When we activate an initialize action we check to see
7167 	 * if the vdev_initialize_thread is NULL. We do this instead
7168 	 * of using the vdev_initialize_state since there might be
7169 	 * a previous initialization process which has completed but
7170 	 * the thread is not exited.
7171 	 */
7172 	if (cmd_type == POOL_INITIALIZE_START &&
7173 	    (vd->vdev_initialize_thread != NULL ||
7174 	    vd->vdev_top->vdev_removing)) {
7175 		mutex_exit(&vd->vdev_initialize_lock);
7176 		return (SET_ERROR(EBUSY));
7177 	} else if (cmd_type == POOL_INITIALIZE_CANCEL &&
7178 	    (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE &&
7179 	    vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) {
7180 		mutex_exit(&vd->vdev_initialize_lock);
7181 		return (SET_ERROR(ESRCH));
7182 	} else if (cmd_type == POOL_INITIALIZE_SUSPEND &&
7183 	    vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) {
7184 		mutex_exit(&vd->vdev_initialize_lock);
7185 		return (SET_ERROR(ESRCH));
7186 	}
7187 
7188 	switch (cmd_type) {
7189 	case POOL_INITIALIZE_START:
7190 		vdev_initialize(vd);
7191 		break;
7192 	case POOL_INITIALIZE_CANCEL:
7193 		vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED, vd_list);
7194 		break;
7195 	case POOL_INITIALIZE_SUSPEND:
7196 		vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED, vd_list);
7197 		break;
7198 	default:
7199 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
7200 	}
7201 	mutex_exit(&vd->vdev_initialize_lock);
7202 
7203 	return (0);
7204 }
7205 
7206 int
7207 spa_vdev_initialize(spa_t *spa, nvlist_t *nv, uint64_t cmd_type,
7208     nvlist_t *vdev_errlist)
7209 {
7210 	int total_errors = 0;
7211 	list_t vd_list;
7212 
7213 	list_create(&vd_list, sizeof (vdev_t),
7214 	    offsetof(vdev_t, vdev_initialize_node));
7215 
7216 	/*
7217 	 * We hold the namespace lock through the whole function
7218 	 * to prevent any changes to the pool while we're starting or
7219 	 * stopping initialization. The config and state locks are held so that
7220 	 * we can properly assess the vdev state before we commit to
7221 	 * the initializing operation.
7222 	 */
7223 	mutex_enter(&spa_namespace_lock);
7224 
7225 	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
7226 	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
7227 		uint64_t vdev_guid = fnvpair_value_uint64(pair);
7228 
7229 		int error = spa_vdev_initialize_impl(spa, vdev_guid, cmd_type,
7230 		    &vd_list);
7231 		if (error != 0) {
7232 			char guid_as_str[MAXNAMELEN];
7233 
7234 			(void) snprintf(guid_as_str, sizeof (guid_as_str),
7235 			    "%llu", (unsigned long long)vdev_guid);
7236 			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
7237 			total_errors++;
7238 		}
7239 	}
7240 
7241 	/* Wait for all initialize threads to stop. */
7242 	vdev_initialize_stop_wait(spa, &vd_list);
7243 
7244 	/* Sync out the initializing state */
7245 	txg_wait_synced(spa->spa_dsl_pool, 0);
7246 	mutex_exit(&spa_namespace_lock);
7247 
7248 	list_destroy(&vd_list);
7249 
7250 	return (total_errors);
7251 }
7252 
7253 static int
7254 spa_vdev_trim_impl(spa_t *spa, uint64_t guid, uint64_t cmd_type,
7255     uint64_t rate, boolean_t partial, boolean_t secure, list_t *vd_list)
7256 {
7257 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7258 
7259 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
7260 
7261 	/* Look up vdev and ensure it's a leaf. */
7262 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
7263 	if (vd == NULL || vd->vdev_detached) {
7264 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7265 		return (SET_ERROR(ENODEV));
7266 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
7267 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7268 		return (SET_ERROR(EINVAL));
7269 	} else if (!vdev_writeable(vd)) {
7270 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7271 		return (SET_ERROR(EROFS));
7272 	} else if (!vd->vdev_has_trim) {
7273 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7274 		return (SET_ERROR(EOPNOTSUPP));
7275 	} else if (secure && !vd->vdev_has_securetrim) {
7276 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7277 		return (SET_ERROR(EOPNOTSUPP));
7278 	}
7279 	mutex_enter(&vd->vdev_trim_lock);
7280 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
7281 
7282 	/*
7283 	 * When we activate a TRIM action we check to see if the
7284 	 * vdev_trim_thread is NULL. We do this instead of using the
7285 	 * vdev_trim_state since there might be a previous TRIM process
7286 	 * which has completed but the thread is not exited.
7287 	 */
7288 	if (cmd_type == POOL_TRIM_START &&
7289 	    (vd->vdev_trim_thread != NULL || vd->vdev_top->vdev_removing)) {
7290 		mutex_exit(&vd->vdev_trim_lock);
7291 		return (SET_ERROR(EBUSY));
7292 	} else if (cmd_type == POOL_TRIM_CANCEL &&
7293 	    (vd->vdev_trim_state != VDEV_TRIM_ACTIVE &&
7294 	    vd->vdev_trim_state != VDEV_TRIM_SUSPENDED)) {
7295 		mutex_exit(&vd->vdev_trim_lock);
7296 		return (SET_ERROR(ESRCH));
7297 	} else if (cmd_type == POOL_TRIM_SUSPEND &&
7298 	    vd->vdev_trim_state != VDEV_TRIM_ACTIVE) {
7299 		mutex_exit(&vd->vdev_trim_lock);
7300 		return (SET_ERROR(ESRCH));
7301 	}
7302 
7303 	switch (cmd_type) {
7304 	case POOL_TRIM_START:
7305 		vdev_trim(vd, rate, partial, secure);
7306 		break;
7307 	case POOL_TRIM_CANCEL:
7308 		vdev_trim_stop(vd, VDEV_TRIM_CANCELED, vd_list);
7309 		break;
7310 	case POOL_TRIM_SUSPEND:
7311 		vdev_trim_stop(vd, VDEV_TRIM_SUSPENDED, vd_list);
7312 		break;
7313 	default:
7314 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
7315 	}
7316 	mutex_exit(&vd->vdev_trim_lock);
7317 
7318 	return (0);
7319 }
7320 
7321 /*
7322  * Initiates a manual TRIM for the requested vdevs. This kicks off individual
7323  * TRIM threads for each child vdev.  These threads pass over all of the free
7324  * space in the vdev's metaslabs and issues TRIM commands for that space.
7325  */
7326 int
7327 spa_vdev_trim(spa_t *spa, nvlist_t *nv, uint64_t cmd_type, uint64_t rate,
7328     boolean_t partial, boolean_t secure, nvlist_t *vdev_errlist)
7329 {
7330 	int total_errors = 0;
7331 	list_t vd_list;
7332 
7333 	list_create(&vd_list, sizeof (vdev_t),
7334 	    offsetof(vdev_t, vdev_trim_node));
7335 
7336 	/*
7337 	 * We hold the namespace lock through the whole function
7338 	 * to prevent any changes to the pool while we're starting or
7339 	 * stopping TRIM. The config and state locks are held so that
7340 	 * we can properly assess the vdev state before we commit to
7341 	 * the TRIM operation.
7342 	 */
7343 	mutex_enter(&spa_namespace_lock);
7344 
7345 	for (nvpair_t *pair = nvlist_next_nvpair(nv, NULL);
7346 	    pair != NULL; pair = nvlist_next_nvpair(nv, pair)) {
7347 		uint64_t vdev_guid = fnvpair_value_uint64(pair);
7348 
7349 		int error = spa_vdev_trim_impl(spa, vdev_guid, cmd_type,
7350 		    rate, partial, secure, &vd_list);
7351 		if (error != 0) {
7352 			char guid_as_str[MAXNAMELEN];
7353 
7354 			(void) snprintf(guid_as_str, sizeof (guid_as_str),
7355 			    "%llu", (unsigned long long)vdev_guid);
7356 			fnvlist_add_int64(vdev_errlist, guid_as_str, error);
7357 			total_errors++;
7358 		}
7359 	}
7360 
7361 	/* Wait for all TRIM threads to stop. */
7362 	vdev_trim_stop_wait(spa, &vd_list);
7363 
7364 	/* Sync out the TRIM state */
7365 	txg_wait_synced(spa->spa_dsl_pool, 0);
7366 	mutex_exit(&spa_namespace_lock);
7367 
7368 	list_destroy(&vd_list);
7369 
7370 	return (total_errors);
7371 }
7372 
7373 /*
7374  * Split a set of devices from their mirrors, and create a new pool from them.
7375  */
7376 int
7377 spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
7378     nvlist_t *props, boolean_t exp)
7379 {
7380 	int error = 0;
7381 	uint64_t txg, *glist;
7382 	spa_t *newspa;
7383 	uint_t c, children, lastlog;
7384 	nvlist_t **child, *nvl, *tmp;
7385 	dmu_tx_t *tx;
7386 	char *altroot = NULL;
7387 	vdev_t *rvd, **vml = NULL;			/* vdev modify list */
7388 	boolean_t activate_slog;
7389 
7390 	ASSERT(spa_writeable(spa));
7391 
7392 	txg = spa_vdev_enter(spa);
7393 
7394 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
7395 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
7396 		error = (spa_has_checkpoint(spa)) ?
7397 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
7398 		return (spa_vdev_exit(spa, NULL, txg, error));
7399 	}
7400 
7401 	/* clear the log and flush everything up to now */
7402 	activate_slog = spa_passivate_log(spa);
7403 	(void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
7404 	error = spa_reset_logs(spa);
7405 	txg = spa_vdev_config_enter(spa);
7406 
7407 	if (activate_slog)
7408 		spa_activate_log(spa);
7409 
7410 	if (error != 0)
7411 		return (spa_vdev_exit(spa, NULL, txg, error));
7412 
7413 	/* check new spa name before going any further */
7414 	if (spa_lookup(newname) != NULL)
7415 		return (spa_vdev_exit(spa, NULL, txg, EEXIST));
7416 
7417 	/*
7418 	 * scan through all the children to ensure they're all mirrors
7419 	 */
7420 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 ||
7421 	    nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child,
7422 	    &children) != 0)
7423 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7424 
7425 	/* first, check to ensure we've got the right child count */
7426 	rvd = spa->spa_root_vdev;
7427 	lastlog = 0;
7428 	for (c = 0; c < rvd->vdev_children; c++) {
7429 		vdev_t *vd = rvd->vdev_child[c];
7430 
7431 		/* don't count the holes & logs as children */
7432 		if (vd->vdev_islog || (vd->vdev_ops != &vdev_indirect_ops &&
7433 		    !vdev_is_concrete(vd))) {
7434 			if (lastlog == 0)
7435 				lastlog = c;
7436 			continue;
7437 		}
7438 
7439 		lastlog = 0;
7440 	}
7441 	if (children != (lastlog != 0 ? lastlog : rvd->vdev_children))
7442 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7443 
7444 	/* next, ensure no spare or cache devices are part of the split */
7445 	if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 ||
7446 	    nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0)
7447 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
7448 
7449 	vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP);
7450 	glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP);
7451 
7452 	/* then, loop over each vdev and validate it */
7453 	for (c = 0; c < children; c++) {
7454 		uint64_t is_hole = 0;
7455 
7456 		(void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE,
7457 		    &is_hole);
7458 
7459 		if (is_hole != 0) {
7460 			if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole ||
7461 			    spa->spa_root_vdev->vdev_child[c]->vdev_islog) {
7462 				continue;
7463 			} else {
7464 				error = SET_ERROR(EINVAL);
7465 				break;
7466 			}
7467 		}
7468 
7469 		/* deal with indirect vdevs */
7470 		if (spa->spa_root_vdev->vdev_child[c]->vdev_ops ==
7471 		    &vdev_indirect_ops)
7472 			continue;
7473 
7474 		/* which disk is going to be split? */
7475 		if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID,
7476 		    &glist[c]) != 0) {
7477 			error = SET_ERROR(EINVAL);
7478 			break;
7479 		}
7480 
7481 		/* look it up in the spa */
7482 		vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE);
7483 		if (vml[c] == NULL) {
7484 			error = SET_ERROR(ENODEV);
7485 			break;
7486 		}
7487 
7488 		/* make sure there's nothing stopping the split */
7489 		if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops ||
7490 		    vml[c]->vdev_islog ||
7491 		    !vdev_is_concrete(vml[c]) ||
7492 		    vml[c]->vdev_isspare ||
7493 		    vml[c]->vdev_isl2cache ||
7494 		    !vdev_writeable(vml[c]) ||
7495 		    vml[c]->vdev_children != 0 ||
7496 		    vml[c]->vdev_state != VDEV_STATE_HEALTHY ||
7497 		    c != spa->spa_root_vdev->vdev_child[c]->vdev_id) {
7498 			error = SET_ERROR(EINVAL);
7499 			break;
7500 		}
7501 
7502 		if (vdev_dtl_required(vml[c]) ||
7503 		    vdev_resilver_needed(vml[c], NULL, NULL)) {
7504 			error = SET_ERROR(EBUSY);
7505 			break;
7506 		}
7507 
7508 		/* we need certain info from the top level */
7509 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY,
7510 		    vml[c]->vdev_top->vdev_ms_array) == 0);
7511 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT,
7512 		    vml[c]->vdev_top->vdev_ms_shift) == 0);
7513 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE,
7514 		    vml[c]->vdev_top->vdev_asize) == 0);
7515 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT,
7516 		    vml[c]->vdev_top->vdev_ashift) == 0);
7517 
7518 		/* transfer per-vdev ZAPs */
7519 		ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0);
7520 		VERIFY0(nvlist_add_uint64(child[c],
7521 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap));
7522 
7523 		ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0);
7524 		VERIFY0(nvlist_add_uint64(child[c],
7525 		    ZPOOL_CONFIG_VDEV_TOP_ZAP,
7526 		    vml[c]->vdev_parent->vdev_top_zap));
7527 	}
7528 
7529 	if (error != 0) {
7530 		kmem_free(vml, children * sizeof (vdev_t *));
7531 		kmem_free(glist, children * sizeof (uint64_t));
7532 		return (spa_vdev_exit(spa, NULL, txg, error));
7533 	}
7534 
7535 	/* stop writers from using the disks */
7536 	for (c = 0; c < children; c++) {
7537 		if (vml[c] != NULL)
7538 			vml[c]->vdev_offline = B_TRUE;
7539 	}
7540 	vdev_reopen(spa->spa_root_vdev);
7541 
7542 	/*
7543 	 * Temporarily record the splitting vdevs in the spa config.  This
7544 	 * will disappear once the config is regenerated.
7545 	 */
7546 	VERIFY(nvlist_alloc(&nvl, NV_UNIQUE_NAME, KM_SLEEP) == 0);
7547 	VERIFY(nvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST,
7548 	    glist, children) == 0);
7549 	kmem_free(glist, children * sizeof (uint64_t));
7550 
7551 	mutex_enter(&spa->spa_props_lock);
7552 	VERIFY(nvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT,
7553 	    nvl) == 0);
7554 	mutex_exit(&spa->spa_props_lock);
7555 	spa->spa_config_splitting = nvl;
7556 	vdev_config_dirty(spa->spa_root_vdev);
7557 
7558 	/* configure and create the new pool */
7559 	VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname) == 0);
7560 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE,
7561 	    exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE) == 0);
7562 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_VERSION,
7563 	    spa_version(spa)) == 0);
7564 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG,
7565 	    spa->spa_config_txg) == 0);
7566 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID,
7567 	    spa_generate_guid(NULL)) == 0);
7568 	VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
7569 	(void) nvlist_lookup_string(props,
7570 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
7571 
7572 	/* add the new pool to the namespace */
7573 	newspa = spa_add(newname, config, altroot);
7574 	newspa->spa_avz_action = AVZ_ACTION_REBUILD;
7575 	newspa->spa_config_txg = spa->spa_config_txg;
7576 	spa_set_log_state(newspa, SPA_LOG_CLEAR);
7577 
7578 	/* release the spa config lock, retaining the namespace lock */
7579 	spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
7580 
7581 	if (zio_injection_enabled)
7582 		zio_handle_panic_injection(spa, FTAG, 1);
7583 
7584 	spa_activate(newspa, spa_mode_global);
7585 	spa_async_suspend(newspa);
7586 
7587 	/*
7588 	 * Temporarily stop the initializing and TRIM activity.  We set the
7589 	 * state to ACTIVE so that we know to resume initializing or TRIM
7590 	 * once the split has completed.
7591 	 */
7592 	list_t vd_initialize_list;
7593 	list_create(&vd_initialize_list, sizeof (vdev_t),
7594 	    offsetof(vdev_t, vdev_initialize_node));
7595 
7596 	list_t vd_trim_list;
7597 	list_create(&vd_trim_list, sizeof (vdev_t),
7598 	    offsetof(vdev_t, vdev_trim_node));
7599 
7600 	for (c = 0; c < children; c++) {
7601 		if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) {
7602 			mutex_enter(&vml[c]->vdev_initialize_lock);
7603 			vdev_initialize_stop(vml[c],
7604 			    VDEV_INITIALIZE_ACTIVE, &vd_initialize_list);
7605 			mutex_exit(&vml[c]->vdev_initialize_lock);
7606 
7607 			mutex_enter(&vml[c]->vdev_trim_lock);
7608 			vdev_trim_stop(vml[c], VDEV_TRIM_ACTIVE, &vd_trim_list);
7609 			mutex_exit(&vml[c]->vdev_trim_lock);
7610 		}
7611 	}
7612 
7613 	vdev_initialize_stop_wait(spa, &vd_initialize_list);
7614 	vdev_trim_stop_wait(spa, &vd_trim_list);
7615 
7616 	list_destroy(&vd_initialize_list);
7617 	list_destroy(&vd_trim_list);
7618 
7619 	newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT;
7620 	newspa->spa_is_splitting = B_TRUE;
7621 
7622 	/* create the new pool from the disks of the original pool */
7623 	error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE);
7624 	if (error)
7625 		goto out;
7626 
7627 	/* if that worked, generate a real config for the new pool */
7628 	if (newspa->spa_root_vdev != NULL) {
7629 		VERIFY(nvlist_alloc(&newspa->spa_config_splitting,
7630 		    NV_UNIQUE_NAME, KM_SLEEP) == 0);
7631 		VERIFY(nvlist_add_uint64(newspa->spa_config_splitting,
7632 		    ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa)) == 0);
7633 		spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL,
7634 		    B_TRUE));
7635 	}
7636 
7637 	/* set the props */
7638 	if (props != NULL) {
7639 		spa_configfile_set(newspa, props, B_FALSE);
7640 		error = spa_prop_set(newspa, props);
7641 		if (error)
7642 			goto out;
7643 	}
7644 
7645 	/* flush everything */
7646 	txg = spa_vdev_config_enter(newspa);
7647 	vdev_config_dirty(newspa->spa_root_vdev);
7648 	(void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG);
7649 
7650 	if (zio_injection_enabled)
7651 		zio_handle_panic_injection(spa, FTAG, 2);
7652 
7653 	spa_async_resume(newspa);
7654 
7655 	/* finally, update the original pool's config */
7656 	txg = spa_vdev_config_enter(spa);
7657 	tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
7658 	error = dmu_tx_assign(tx, TXG_WAIT);
7659 	if (error != 0)
7660 		dmu_tx_abort(tx);
7661 	for (c = 0; c < children; c++) {
7662 		if (vml[c] != NULL && vml[c]->vdev_ops != &vdev_indirect_ops) {
7663 			vdev_t *tvd = vml[c]->vdev_top;
7664 
7665 			/*
7666 			 * Need to be sure the detachable VDEV is not
7667 			 * on any *other* txg's DTL list to prevent it
7668 			 * from being accessed after it's freed.
7669 			 */
7670 			for (int t = 0; t < TXG_SIZE; t++) {
7671 				(void) txg_list_remove_this(
7672 				    &tvd->vdev_dtl_list, vml[c], t);
7673 			}
7674 
7675 			vdev_split(vml[c]);
7676 			if (error == 0)
7677 				spa_history_log_internal(spa, "detach", tx,
7678 				    "vdev=%s", vml[c]->vdev_path);
7679 
7680 			vdev_free(vml[c]);
7681 		}
7682 	}
7683 	spa->spa_avz_action = AVZ_ACTION_REBUILD;
7684 	vdev_config_dirty(spa->spa_root_vdev);
7685 	spa->spa_config_splitting = NULL;
7686 	nvlist_free(nvl);
7687 	if (error == 0)
7688 		dmu_tx_commit(tx);
7689 	(void) spa_vdev_exit(spa, NULL, txg, 0);
7690 
7691 	if (zio_injection_enabled)
7692 		zio_handle_panic_injection(spa, FTAG, 3);
7693 
7694 	/* split is complete; log a history record */
7695 	spa_history_log_internal(newspa, "split", NULL,
7696 	    "from pool %s", spa_name(spa));
7697 
7698 	newspa->spa_is_splitting = B_FALSE;
7699 	kmem_free(vml, children * sizeof (vdev_t *));
7700 
7701 	/* if we're not going to mount the filesystems in userland, export */
7702 	if (exp)
7703 		error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL,
7704 		    B_FALSE, B_FALSE);
7705 
7706 	return (error);
7707 
7708 out:
7709 	spa_unload(newspa);
7710 	spa_deactivate(newspa);
7711 	spa_remove(newspa);
7712 
7713 	txg = spa_vdev_config_enter(spa);
7714 
7715 	/* re-online all offlined disks */
7716 	for (c = 0; c < children; c++) {
7717 		if (vml[c] != NULL)
7718 			vml[c]->vdev_offline = B_FALSE;
7719 	}
7720 
7721 	/* restart initializing or trimming disks as necessary */
7722 	spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
7723 	spa_async_request(spa, SPA_ASYNC_TRIM_RESTART);
7724 	spa_async_request(spa, SPA_ASYNC_AUTOTRIM_RESTART);
7725 
7726 	vdev_reopen(spa->spa_root_vdev);
7727 
7728 	nvlist_free(spa->spa_config_splitting);
7729 	spa->spa_config_splitting = NULL;
7730 	(void) spa_vdev_exit(spa, NULL, txg, error);
7731 
7732 	kmem_free(vml, children * sizeof (vdev_t *));
7733 	return (error);
7734 }
7735 
7736 /*
7737  * Find any device that's done replacing, or a vdev marked 'unspare' that's
7738  * currently spared, so we can detach it.
7739  */
7740 static vdev_t *
7741 spa_vdev_resilver_done_hunt(vdev_t *vd)
7742 {
7743 	vdev_t *newvd, *oldvd;
7744 
7745 	for (int c = 0; c < vd->vdev_children; c++) {
7746 		oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]);
7747 		if (oldvd != NULL)
7748 			return (oldvd);
7749 	}
7750 
7751 	/*
7752 	 * Check for a completed replacement.  We always consider the first
7753 	 * vdev in the list to be the oldest vdev, and the last one to be
7754 	 * the newest (see spa_vdev_attach() for how that works).  In
7755 	 * the case where the newest vdev is faulted, we will not automatically
7756 	 * remove it after a resilver completes.  This is OK as it will require
7757 	 * user intervention to determine which disk the admin wishes to keep.
7758 	 */
7759 	if (vd->vdev_ops == &vdev_replacing_ops) {
7760 		ASSERT(vd->vdev_children > 1);
7761 
7762 		newvd = vd->vdev_child[vd->vdev_children - 1];
7763 		oldvd = vd->vdev_child[0];
7764 
7765 		if (vdev_dtl_empty(newvd, DTL_MISSING) &&
7766 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
7767 		    !vdev_dtl_required(oldvd))
7768 			return (oldvd);
7769 	}
7770 
7771 	/*
7772 	 * Check for a completed resilver with the 'unspare' flag set.
7773 	 * Also potentially update faulted state.
7774 	 */
7775 	if (vd->vdev_ops == &vdev_spare_ops) {
7776 		vdev_t *first = vd->vdev_child[0];
7777 		vdev_t *last = vd->vdev_child[vd->vdev_children - 1];
7778 
7779 		if (last->vdev_unspare) {
7780 			oldvd = first;
7781 			newvd = last;
7782 		} else if (first->vdev_unspare) {
7783 			oldvd = last;
7784 			newvd = first;
7785 		} else {
7786 			oldvd = NULL;
7787 		}
7788 
7789 		if (oldvd != NULL &&
7790 		    vdev_dtl_empty(newvd, DTL_MISSING) &&
7791 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
7792 		    !vdev_dtl_required(oldvd))
7793 			return (oldvd);
7794 
7795 		vdev_propagate_state(vd);
7796 
7797 		/*
7798 		 * If there are more than two spares attached to a disk,
7799 		 * and those spares are not required, then we want to
7800 		 * attempt to free them up now so that they can be used
7801 		 * by other pools.  Once we're back down to a single
7802 		 * disk+spare, we stop removing them.
7803 		 */
7804 		if (vd->vdev_children > 2) {
7805 			newvd = vd->vdev_child[1];
7806 
7807 			if (newvd->vdev_isspare && last->vdev_isspare &&
7808 			    vdev_dtl_empty(last, DTL_MISSING) &&
7809 			    vdev_dtl_empty(last, DTL_OUTAGE) &&
7810 			    !vdev_dtl_required(newvd))
7811 				return (newvd);
7812 		}
7813 	}
7814 
7815 	return (NULL);
7816 }
7817 
7818 static void
7819 spa_vdev_resilver_done(spa_t *spa)
7820 {
7821 	vdev_t *vd, *pvd, *ppvd;
7822 	uint64_t guid, sguid, pguid, ppguid;
7823 
7824 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
7825 
7826 	while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) {
7827 		pvd = vd->vdev_parent;
7828 		ppvd = pvd->vdev_parent;
7829 		guid = vd->vdev_guid;
7830 		pguid = pvd->vdev_guid;
7831 		ppguid = ppvd->vdev_guid;
7832 		sguid = 0;
7833 		/*
7834 		 * If we have just finished replacing a hot spared device, then
7835 		 * we need to detach the parent's first child (the original hot
7836 		 * spare) as well.
7837 		 */
7838 		if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 &&
7839 		    ppvd->vdev_children == 2) {
7840 			ASSERT(pvd->vdev_ops == &vdev_replacing_ops);
7841 			sguid = ppvd->vdev_child[1]->vdev_guid;
7842 		}
7843 		ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd));
7844 
7845 		spa_config_exit(spa, SCL_ALL, FTAG);
7846 		if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0)
7847 			return;
7848 		if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0)
7849 			return;
7850 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
7851 	}
7852 
7853 	spa_config_exit(spa, SCL_ALL, FTAG);
7854 
7855 	/*
7856 	 * If a detach was not performed above replace waiters will not have
7857 	 * been notified.  In which case we must do so now.
7858 	 */
7859 	spa_notify_waiters(spa);
7860 }
7861 
7862 /*
7863  * Update the stored path or FRU for this vdev.
7864  */
7865 static int
7866 spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value,
7867     boolean_t ispath)
7868 {
7869 	vdev_t *vd;
7870 	boolean_t sync = B_FALSE;
7871 
7872 	ASSERT(spa_writeable(spa));
7873 
7874 	spa_vdev_state_enter(spa, SCL_ALL);
7875 
7876 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
7877 		return (spa_vdev_state_exit(spa, NULL, ENOENT));
7878 
7879 	if (!vd->vdev_ops->vdev_op_leaf)
7880 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
7881 
7882 	if (ispath) {
7883 		if (strcmp(value, vd->vdev_path) != 0) {
7884 			spa_strfree(vd->vdev_path);
7885 			vd->vdev_path = spa_strdup(value);
7886 			sync = B_TRUE;
7887 		}
7888 	} else {
7889 		if (vd->vdev_fru == NULL) {
7890 			vd->vdev_fru = spa_strdup(value);
7891 			sync = B_TRUE;
7892 		} else if (strcmp(value, vd->vdev_fru) != 0) {
7893 			spa_strfree(vd->vdev_fru);
7894 			vd->vdev_fru = spa_strdup(value);
7895 			sync = B_TRUE;
7896 		}
7897 	}
7898 
7899 	return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0));
7900 }
7901 
7902 int
7903 spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath)
7904 {
7905 	return (spa_vdev_set_common(spa, guid, newpath, B_TRUE));
7906 }
7907 
7908 int
7909 spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru)
7910 {
7911 	return (spa_vdev_set_common(spa, guid, newfru, B_FALSE));
7912 }
7913 
7914 /*
7915  * ==========================================================================
7916  * SPA Scanning
7917  * ==========================================================================
7918  */
7919 int
7920 spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd)
7921 {
7922 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
7923 
7924 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
7925 		return (SET_ERROR(EBUSY));
7926 
7927 	return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd));
7928 }
7929 
7930 int
7931 spa_scan_stop(spa_t *spa)
7932 {
7933 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
7934 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
7935 		return (SET_ERROR(EBUSY));
7936 	return (dsl_scan_cancel(spa->spa_dsl_pool));
7937 }
7938 
7939 int
7940 spa_scan(spa_t *spa, pool_scan_func_t func)
7941 {
7942 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
7943 
7944 	if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE)
7945 		return (SET_ERROR(ENOTSUP));
7946 
7947 	if (func == POOL_SCAN_RESILVER &&
7948 	    !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER))
7949 		return (SET_ERROR(ENOTSUP));
7950 
7951 	/*
7952 	 * If a resilver was requested, but there is no DTL on a
7953 	 * writeable leaf device, we have nothing to do.
7954 	 */
7955 	if (func == POOL_SCAN_RESILVER &&
7956 	    !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
7957 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
7958 		return (0);
7959 	}
7960 
7961 	return (dsl_scan(spa->spa_dsl_pool, func));
7962 }
7963 
7964 /*
7965  * ==========================================================================
7966  * SPA async task processing
7967  * ==========================================================================
7968  */
7969 
7970 static void
7971 spa_async_remove(spa_t *spa, vdev_t *vd)
7972 {
7973 	if (vd->vdev_remove_wanted) {
7974 		vd->vdev_remove_wanted = B_FALSE;
7975 		vd->vdev_delayed_close = B_FALSE;
7976 		vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE);
7977 
7978 		/*
7979 		 * We want to clear the stats, but we don't want to do a full
7980 		 * vdev_clear() as that will cause us to throw away
7981 		 * degraded/faulted state as well as attempt to reopen the
7982 		 * device, all of which is a waste.
7983 		 */
7984 		vd->vdev_stat.vs_read_errors = 0;
7985 		vd->vdev_stat.vs_write_errors = 0;
7986 		vd->vdev_stat.vs_checksum_errors = 0;
7987 
7988 		vdev_state_dirty(vd->vdev_top);
7989 
7990 		/* Tell userspace that the vdev is gone. */
7991 		zfs_post_remove(spa, vd);
7992 	}
7993 
7994 	for (int c = 0; c < vd->vdev_children; c++)
7995 		spa_async_remove(spa, vd->vdev_child[c]);
7996 }
7997 
7998 static void
7999 spa_async_probe(spa_t *spa, vdev_t *vd)
8000 {
8001 	if (vd->vdev_probe_wanted) {
8002 		vd->vdev_probe_wanted = B_FALSE;
8003 		vdev_reopen(vd);	/* vdev_open() does the actual probe */
8004 	}
8005 
8006 	for (int c = 0; c < vd->vdev_children; c++)
8007 		spa_async_probe(spa, vd->vdev_child[c]);
8008 }
8009 
8010 static void
8011 spa_async_autoexpand(spa_t *spa, vdev_t *vd)
8012 {
8013 	if (!spa->spa_autoexpand)
8014 		return;
8015 
8016 	for (int c = 0; c < vd->vdev_children; c++) {
8017 		vdev_t *cvd = vd->vdev_child[c];
8018 		spa_async_autoexpand(spa, cvd);
8019 	}
8020 
8021 	if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL)
8022 		return;
8023 
8024 	spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_AUTOEXPAND);
8025 }
8026 
8027 static void
8028 spa_async_thread(void *arg)
8029 {
8030 	spa_t *spa = (spa_t *)arg;
8031 	dsl_pool_t *dp = spa->spa_dsl_pool;
8032 	int tasks;
8033 
8034 	ASSERT(spa->spa_sync_on);
8035 
8036 	mutex_enter(&spa->spa_async_lock);
8037 	tasks = spa->spa_async_tasks;
8038 	spa->spa_async_tasks = 0;
8039 	mutex_exit(&spa->spa_async_lock);
8040 
8041 	/*
8042 	 * See if the config needs to be updated.
8043 	 */
8044 	if (tasks & SPA_ASYNC_CONFIG_UPDATE) {
8045 		uint64_t old_space, new_space;
8046 
8047 		mutex_enter(&spa_namespace_lock);
8048 		old_space = metaslab_class_get_space(spa_normal_class(spa));
8049 		old_space += metaslab_class_get_space(spa_special_class(spa));
8050 		old_space += metaslab_class_get_space(spa_dedup_class(spa));
8051 		old_space += metaslab_class_get_space(
8052 		    spa_embedded_log_class(spa));
8053 
8054 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
8055 
8056 		new_space = metaslab_class_get_space(spa_normal_class(spa));
8057 		new_space += metaslab_class_get_space(spa_special_class(spa));
8058 		new_space += metaslab_class_get_space(spa_dedup_class(spa));
8059 		new_space += metaslab_class_get_space(
8060 		    spa_embedded_log_class(spa));
8061 		mutex_exit(&spa_namespace_lock);
8062 
8063 		/*
8064 		 * If the pool grew as a result of the config update,
8065 		 * then log an internal history event.
8066 		 */
8067 		if (new_space != old_space) {
8068 			spa_history_log_internal(spa, "vdev online", NULL,
8069 			    "pool '%s' size: %llu(+%llu)",
8070 			    spa_name(spa), (u_longlong_t)new_space,
8071 			    (u_longlong_t)(new_space - old_space));
8072 		}
8073 	}
8074 
8075 	/*
8076 	 * See if any devices need to be marked REMOVED.
8077 	 */
8078 	if (tasks & SPA_ASYNC_REMOVE) {
8079 		spa_vdev_state_enter(spa, SCL_NONE);
8080 		spa_async_remove(spa, spa->spa_root_vdev);
8081 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++)
8082 			spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]);
8083 		for (int i = 0; i < spa->spa_spares.sav_count; i++)
8084 			spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]);
8085 		(void) spa_vdev_state_exit(spa, NULL, 0);
8086 	}
8087 
8088 	if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) {
8089 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8090 		spa_async_autoexpand(spa, spa->spa_root_vdev);
8091 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8092 	}
8093 
8094 	/*
8095 	 * See if any devices need to be probed.
8096 	 */
8097 	if (tasks & SPA_ASYNC_PROBE) {
8098 		spa_vdev_state_enter(spa, SCL_NONE);
8099 		spa_async_probe(spa, spa->spa_root_vdev);
8100 		(void) spa_vdev_state_exit(spa, NULL, 0);
8101 	}
8102 
8103 	/*
8104 	 * If any devices are done replacing, detach them.
8105 	 */
8106 	if (tasks & SPA_ASYNC_RESILVER_DONE ||
8107 	    tasks & SPA_ASYNC_REBUILD_DONE) {
8108 		spa_vdev_resilver_done(spa);
8109 	}
8110 
8111 	/*
8112 	 * Kick off a resilver.
8113 	 */
8114 	if (tasks & SPA_ASYNC_RESILVER &&
8115 	    !vdev_rebuild_active(spa->spa_root_vdev) &&
8116 	    (!dsl_scan_resilvering(dp) ||
8117 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER)))
8118 		dsl_scan_restart_resilver(dp, 0);
8119 
8120 	if (tasks & SPA_ASYNC_INITIALIZE_RESTART) {
8121 		mutex_enter(&spa_namespace_lock);
8122 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8123 		vdev_initialize_restart(spa->spa_root_vdev);
8124 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8125 		mutex_exit(&spa_namespace_lock);
8126 	}
8127 
8128 	if (tasks & SPA_ASYNC_TRIM_RESTART) {
8129 		mutex_enter(&spa_namespace_lock);
8130 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8131 		vdev_trim_restart(spa->spa_root_vdev);
8132 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8133 		mutex_exit(&spa_namespace_lock);
8134 	}
8135 
8136 	if (tasks & SPA_ASYNC_AUTOTRIM_RESTART) {
8137 		mutex_enter(&spa_namespace_lock);
8138 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8139 		vdev_autotrim_restart(spa);
8140 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8141 		mutex_exit(&spa_namespace_lock);
8142 	}
8143 
8144 	/*
8145 	 * Kick off L2 cache whole device TRIM.
8146 	 */
8147 	if (tasks & SPA_ASYNC_L2CACHE_TRIM) {
8148 		mutex_enter(&spa_namespace_lock);
8149 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8150 		vdev_trim_l2arc(spa);
8151 		spa_config_exit(spa, SCL_CONFIG, FTAG);
8152 		mutex_exit(&spa_namespace_lock);
8153 	}
8154 
8155 	/*
8156 	 * Kick off L2 cache rebuilding.
8157 	 */
8158 	if (tasks & SPA_ASYNC_L2CACHE_REBUILD) {
8159 		mutex_enter(&spa_namespace_lock);
8160 		spa_config_enter(spa, SCL_L2ARC, FTAG, RW_READER);
8161 		l2arc_spa_rebuild_start(spa);
8162 		spa_config_exit(spa, SCL_L2ARC, FTAG);
8163 		mutex_exit(&spa_namespace_lock);
8164 	}
8165 
8166 	/*
8167 	 * Let the world know that we're done.
8168 	 */
8169 	mutex_enter(&spa->spa_async_lock);
8170 	spa->spa_async_thread = NULL;
8171 	cv_broadcast(&spa->spa_async_cv);
8172 	mutex_exit(&spa->spa_async_lock);
8173 	thread_exit();
8174 }
8175 
8176 void
8177 spa_async_suspend(spa_t *spa)
8178 {
8179 	mutex_enter(&spa->spa_async_lock);
8180 	spa->spa_async_suspended++;
8181 	while (spa->spa_async_thread != NULL)
8182 		cv_wait(&spa->spa_async_cv, &spa->spa_async_lock);
8183 	mutex_exit(&spa->spa_async_lock);
8184 
8185 	spa_vdev_remove_suspend(spa);
8186 
8187 	zthr_t *condense_thread = spa->spa_condense_zthr;
8188 	if (condense_thread != NULL)
8189 		zthr_cancel(condense_thread);
8190 
8191 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
8192 	if (discard_thread != NULL)
8193 		zthr_cancel(discard_thread);
8194 
8195 	zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr;
8196 	if (ll_delete_thread != NULL)
8197 		zthr_cancel(ll_delete_thread);
8198 
8199 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
8200 	if (ll_condense_thread != NULL)
8201 		zthr_cancel(ll_condense_thread);
8202 }
8203 
8204 void
8205 spa_async_resume(spa_t *spa)
8206 {
8207 	mutex_enter(&spa->spa_async_lock);
8208 	ASSERT(spa->spa_async_suspended != 0);
8209 	spa->spa_async_suspended--;
8210 	mutex_exit(&spa->spa_async_lock);
8211 	spa_restart_removal(spa);
8212 
8213 	zthr_t *condense_thread = spa->spa_condense_zthr;
8214 	if (condense_thread != NULL)
8215 		zthr_resume(condense_thread);
8216 
8217 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
8218 	if (discard_thread != NULL)
8219 		zthr_resume(discard_thread);
8220 
8221 	zthr_t *ll_delete_thread = spa->spa_livelist_delete_zthr;
8222 	if (ll_delete_thread != NULL)
8223 		zthr_resume(ll_delete_thread);
8224 
8225 	zthr_t *ll_condense_thread = spa->spa_livelist_condense_zthr;
8226 	if (ll_condense_thread != NULL)
8227 		zthr_resume(ll_condense_thread);
8228 }
8229 
8230 static boolean_t
8231 spa_async_tasks_pending(spa_t *spa)
8232 {
8233 	uint_t non_config_tasks;
8234 	uint_t config_task;
8235 	boolean_t config_task_suspended;
8236 
8237 	non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE;
8238 	config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE;
8239 	if (spa->spa_ccw_fail_time == 0) {
8240 		config_task_suspended = B_FALSE;
8241 	} else {
8242 		config_task_suspended =
8243 		    (gethrtime() - spa->spa_ccw_fail_time) <
8244 		    ((hrtime_t)zfs_ccw_retry_interval * NANOSEC);
8245 	}
8246 
8247 	return (non_config_tasks || (config_task && !config_task_suspended));
8248 }
8249 
8250 static void
8251 spa_async_dispatch(spa_t *spa)
8252 {
8253 	mutex_enter(&spa->spa_async_lock);
8254 	if (spa_async_tasks_pending(spa) &&
8255 	    !spa->spa_async_suspended &&
8256 	    spa->spa_async_thread == NULL)
8257 		spa->spa_async_thread = thread_create(NULL, 0,
8258 		    spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri);
8259 	mutex_exit(&spa->spa_async_lock);
8260 }
8261 
8262 void
8263 spa_async_request(spa_t *spa, int task)
8264 {
8265 	zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task);
8266 	mutex_enter(&spa->spa_async_lock);
8267 	spa->spa_async_tasks |= task;
8268 	mutex_exit(&spa->spa_async_lock);
8269 }
8270 
8271 int
8272 spa_async_tasks(spa_t *spa)
8273 {
8274 	return (spa->spa_async_tasks);
8275 }
8276 
8277 /*
8278  * ==========================================================================
8279  * SPA syncing routines
8280  * ==========================================================================
8281  */
8282 
8283 
8284 static int
8285 bpobj_enqueue_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
8286     dmu_tx_t *tx)
8287 {
8288 	bpobj_t *bpo = arg;
8289 	bpobj_enqueue(bpo, bp, bp_freed, tx);
8290 	return (0);
8291 }
8292 
8293 int
8294 bpobj_enqueue_alloc_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8295 {
8296 	return (bpobj_enqueue_cb(arg, bp, B_FALSE, tx));
8297 }
8298 
8299 int
8300 bpobj_enqueue_free_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8301 {
8302 	return (bpobj_enqueue_cb(arg, bp, B_TRUE, tx));
8303 }
8304 
8305 static int
8306 spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
8307 {
8308 	zio_t *pio = arg;
8309 
8310 	zio_nowait(zio_free_sync(pio, pio->io_spa, dmu_tx_get_txg(tx), bp,
8311 	    pio->io_flags));
8312 	return (0);
8313 }
8314 
8315 static int
8316 bpobj_spa_free_sync_cb(void *arg, const blkptr_t *bp, boolean_t bp_freed,
8317     dmu_tx_t *tx)
8318 {
8319 	ASSERT(!bp_freed);
8320 	return (spa_free_sync_cb(arg, bp, tx));
8321 }
8322 
8323 /*
8324  * Note: this simple function is not inlined to make it easier to dtrace the
8325  * amount of time spent syncing frees.
8326  */
8327 static void
8328 spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx)
8329 {
8330 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
8331 	bplist_iterate(bpl, spa_free_sync_cb, zio, tx);
8332 	VERIFY(zio_wait(zio) == 0);
8333 }
8334 
8335 /*
8336  * Note: this simple function is not inlined to make it easier to dtrace the
8337  * amount of time spent syncing deferred frees.
8338  */
8339 static void
8340 spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx)
8341 {
8342 	if (spa_sync_pass(spa) != 1)
8343 		return;
8344 
8345 	/*
8346 	 * Note:
8347 	 * If the log space map feature is active, we stop deferring
8348 	 * frees to the next TXG and therefore running this function
8349 	 * would be considered a no-op as spa_deferred_bpobj should
8350 	 * not have any entries.
8351 	 *
8352 	 * That said we run this function anyway (instead of returning
8353 	 * immediately) for the edge-case scenario where we just
8354 	 * activated the log space map feature in this TXG but we have
8355 	 * deferred frees from the previous TXG.
8356 	 */
8357 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
8358 	VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj,
8359 	    bpobj_spa_free_sync_cb, zio, tx), ==, 0);
8360 	VERIFY0(zio_wait(zio));
8361 }
8362 
8363 static void
8364 spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx)
8365 {
8366 	char *packed = NULL;
8367 	size_t bufsize;
8368 	size_t nvsize = 0;
8369 	dmu_buf_t *db;
8370 
8371 	VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0);
8372 
8373 	/*
8374 	 * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration
8375 	 * information.  This avoids the dmu_buf_will_dirty() path and
8376 	 * saves us a pre-read to get data we don't actually care about.
8377 	 */
8378 	bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE);
8379 	packed = vmem_alloc(bufsize, KM_SLEEP);
8380 
8381 	VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR,
8382 	    KM_SLEEP) == 0);
8383 	bzero(packed + nvsize, bufsize - nvsize);
8384 
8385 	dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx);
8386 
8387 	vmem_free(packed, bufsize);
8388 
8389 	VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db));
8390 	dmu_buf_will_dirty(db, tx);
8391 	*(uint64_t *)db->db_data = nvsize;
8392 	dmu_buf_rele(db, FTAG);
8393 }
8394 
8395 static void
8396 spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx,
8397     const char *config, const char *entry)
8398 {
8399 	nvlist_t *nvroot;
8400 	nvlist_t **list;
8401 	int i;
8402 
8403 	if (!sav->sav_sync)
8404 		return;
8405 
8406 	/*
8407 	 * Update the MOS nvlist describing the list of available devices.
8408 	 * spa_validate_aux() will have already made sure this nvlist is
8409 	 * valid and the vdevs are labeled appropriately.
8410 	 */
8411 	if (sav->sav_object == 0) {
8412 		sav->sav_object = dmu_object_alloc(spa->spa_meta_objset,
8413 		    DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE,
8414 		    sizeof (uint64_t), tx);
8415 		VERIFY(zap_update(spa->spa_meta_objset,
8416 		    DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1,
8417 		    &sav->sav_object, tx) == 0);
8418 	}
8419 
8420 	VERIFY(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, KM_SLEEP) == 0);
8421 	if (sav->sav_count == 0) {
8422 		VERIFY(nvlist_add_nvlist_array(nvroot, config, NULL, 0) == 0);
8423 	} else {
8424 		list = kmem_alloc(sav->sav_count*sizeof (void *), KM_SLEEP);
8425 		for (i = 0; i < sav->sav_count; i++)
8426 			list[i] = vdev_config_generate(spa, sav->sav_vdevs[i],
8427 			    B_FALSE, VDEV_CONFIG_L2CACHE);
8428 		VERIFY(nvlist_add_nvlist_array(nvroot, config, list,
8429 		    sav->sav_count) == 0);
8430 		for (i = 0; i < sav->sav_count; i++)
8431 			nvlist_free(list[i]);
8432 		kmem_free(list, sav->sav_count * sizeof (void *));
8433 	}
8434 
8435 	spa_sync_nvlist(spa, sav->sav_object, nvroot, tx);
8436 	nvlist_free(nvroot);
8437 
8438 	sav->sav_sync = B_FALSE;
8439 }
8440 
8441 /*
8442  * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t.
8443  * The all-vdev ZAP must be empty.
8444  */
8445 static void
8446 spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx)
8447 {
8448 	spa_t *spa = vd->vdev_spa;
8449 
8450 	if (vd->vdev_top_zap != 0) {
8451 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8452 		    vd->vdev_top_zap, tx));
8453 	}
8454 	if (vd->vdev_leaf_zap != 0) {
8455 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
8456 		    vd->vdev_leaf_zap, tx));
8457 	}
8458 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
8459 		spa_avz_build(vd->vdev_child[i], avz, tx);
8460 	}
8461 }
8462 
8463 static void
8464 spa_sync_config_object(spa_t *spa, dmu_tx_t *tx)
8465 {
8466 	nvlist_t *config;
8467 
8468 	/*
8469 	 * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS,
8470 	 * its config may not be dirty but we still need to build per-vdev ZAPs.
8471 	 * Similarly, if the pool is being assembled (e.g. after a split), we
8472 	 * need to rebuild the AVZ although the config may not be dirty.
8473 	 */
8474 	if (list_is_empty(&spa->spa_config_dirty_list) &&
8475 	    spa->spa_avz_action == AVZ_ACTION_NONE)
8476 		return;
8477 
8478 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
8479 
8480 	ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE ||
8481 	    spa->spa_avz_action == AVZ_ACTION_INITIALIZE ||
8482 	    spa->spa_all_vdev_zaps != 0);
8483 
8484 	if (spa->spa_avz_action == AVZ_ACTION_REBUILD) {
8485 		/* Make and build the new AVZ */
8486 		uint64_t new_avz = zap_create(spa->spa_meta_objset,
8487 		    DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx);
8488 		spa_avz_build(spa->spa_root_vdev, new_avz, tx);
8489 
8490 		/* Diff old AVZ with new one */
8491 		zap_cursor_t zc;
8492 		zap_attribute_t za;
8493 
8494 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
8495 		    spa->spa_all_vdev_zaps);
8496 		    zap_cursor_retrieve(&zc, &za) == 0;
8497 		    zap_cursor_advance(&zc)) {
8498 			uint64_t vdzap = za.za_first_integer;
8499 			if (zap_lookup_int(spa->spa_meta_objset, new_avz,
8500 			    vdzap) == ENOENT) {
8501 				/*
8502 				 * ZAP is listed in old AVZ but not in new one;
8503 				 * destroy it
8504 				 */
8505 				VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap,
8506 				    tx));
8507 			}
8508 		}
8509 
8510 		zap_cursor_fini(&zc);
8511 
8512 		/* Destroy the old AVZ */
8513 		VERIFY0(zap_destroy(spa->spa_meta_objset,
8514 		    spa->spa_all_vdev_zaps, tx));
8515 
8516 		/* Replace the old AVZ in the dir obj with the new one */
8517 		VERIFY0(zap_update(spa->spa_meta_objset,
8518 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP,
8519 		    sizeof (new_avz), 1, &new_avz, tx));
8520 
8521 		spa->spa_all_vdev_zaps = new_avz;
8522 	} else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) {
8523 		zap_cursor_t zc;
8524 		zap_attribute_t za;
8525 
8526 		/* Walk through the AVZ and destroy all listed ZAPs */
8527 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
8528 		    spa->spa_all_vdev_zaps);
8529 		    zap_cursor_retrieve(&zc, &za) == 0;
8530 		    zap_cursor_advance(&zc)) {
8531 			uint64_t zap = za.za_first_integer;
8532 			VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx));
8533 		}
8534 
8535 		zap_cursor_fini(&zc);
8536 
8537 		/* Destroy and unlink the AVZ itself */
8538 		VERIFY0(zap_destroy(spa->spa_meta_objset,
8539 		    spa->spa_all_vdev_zaps, tx));
8540 		VERIFY0(zap_remove(spa->spa_meta_objset,
8541 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx));
8542 		spa->spa_all_vdev_zaps = 0;
8543 	}
8544 
8545 	if (spa->spa_all_vdev_zaps == 0) {
8546 		spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset,
8547 		    DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
8548 		    DMU_POOL_VDEV_ZAP_MAP, tx);
8549 	}
8550 	spa->spa_avz_action = AVZ_ACTION_NONE;
8551 
8552 	/* Create ZAPs for vdevs that don't have them. */
8553 	vdev_construct_zaps(spa->spa_root_vdev, tx);
8554 
8555 	config = spa_config_generate(spa, spa->spa_root_vdev,
8556 	    dmu_tx_get_txg(tx), B_FALSE);
8557 
8558 	/*
8559 	 * If we're upgrading the spa version then make sure that
8560 	 * the config object gets updated with the correct version.
8561 	 */
8562 	if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version)
8563 		fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION,
8564 		    spa->spa_uberblock.ub_version);
8565 
8566 	spa_config_exit(spa, SCL_STATE, FTAG);
8567 
8568 	nvlist_free(spa->spa_config_syncing);
8569 	spa->spa_config_syncing = config;
8570 
8571 	spa_sync_nvlist(spa, spa->spa_config_object, config, tx);
8572 }
8573 
8574 static void
8575 spa_sync_version(void *arg, dmu_tx_t *tx)
8576 {
8577 	uint64_t *versionp = arg;
8578 	uint64_t version = *versionp;
8579 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
8580 
8581 	/*
8582 	 * Setting the version is special cased when first creating the pool.
8583 	 */
8584 	ASSERT(tx->tx_txg != TXG_INITIAL);
8585 
8586 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
8587 	ASSERT(version >= spa_version(spa));
8588 
8589 	spa->spa_uberblock.ub_version = version;
8590 	vdev_config_dirty(spa->spa_root_vdev);
8591 	spa_history_log_internal(spa, "set", tx, "version=%lld",
8592 	    (longlong_t)version);
8593 }
8594 
8595 /*
8596  * Set zpool properties.
8597  */
8598 static void
8599 spa_sync_props(void *arg, dmu_tx_t *tx)
8600 {
8601 	nvlist_t *nvp = arg;
8602 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
8603 	objset_t *mos = spa->spa_meta_objset;
8604 	nvpair_t *elem = NULL;
8605 
8606 	mutex_enter(&spa->spa_props_lock);
8607 
8608 	while ((elem = nvlist_next_nvpair(nvp, elem))) {
8609 		uint64_t intval;
8610 		char *strval, *fname;
8611 		zpool_prop_t prop;
8612 		const char *propname;
8613 		zprop_type_t proptype;
8614 		spa_feature_t fid;
8615 
8616 		switch (prop = zpool_name_to_prop(nvpair_name(elem))) {
8617 		case ZPOOL_PROP_INVAL:
8618 			/*
8619 			 * We checked this earlier in spa_prop_validate().
8620 			 */
8621 			ASSERT(zpool_prop_feature(nvpair_name(elem)));
8622 
8623 			fname = strchr(nvpair_name(elem), '@') + 1;
8624 			VERIFY0(zfeature_lookup_name(fname, &fid));
8625 
8626 			spa_feature_enable(spa, fid, tx);
8627 			spa_history_log_internal(spa, "set", tx,
8628 			    "%s=enabled", nvpair_name(elem));
8629 			break;
8630 
8631 		case ZPOOL_PROP_VERSION:
8632 			intval = fnvpair_value_uint64(elem);
8633 			/*
8634 			 * The version is synced separately before other
8635 			 * properties and should be correct by now.
8636 			 */
8637 			ASSERT3U(spa_version(spa), >=, intval);
8638 			break;
8639 
8640 		case ZPOOL_PROP_ALTROOT:
8641 			/*
8642 			 * 'altroot' is a non-persistent property. It should
8643 			 * have been set temporarily at creation or import time.
8644 			 */
8645 			ASSERT(spa->spa_root != NULL);
8646 			break;
8647 
8648 		case ZPOOL_PROP_READONLY:
8649 		case ZPOOL_PROP_CACHEFILE:
8650 			/*
8651 			 * 'readonly' and 'cachefile' are also non-persistent
8652 			 * properties.
8653 			 */
8654 			break;
8655 		case ZPOOL_PROP_COMMENT:
8656 			strval = fnvpair_value_string(elem);
8657 			if (spa->spa_comment != NULL)
8658 				spa_strfree(spa->spa_comment);
8659 			spa->spa_comment = spa_strdup(strval);
8660 			/*
8661 			 * We need to dirty the configuration on all the vdevs
8662 			 * so that their labels get updated.  It's unnecessary
8663 			 * to do this for pool creation since the vdev's
8664 			 * configuration has already been dirtied.
8665 			 */
8666 			if (tx->tx_txg != TXG_INITIAL)
8667 				vdev_config_dirty(spa->spa_root_vdev);
8668 			spa_history_log_internal(spa, "set", tx,
8669 			    "%s=%s", nvpair_name(elem), strval);
8670 			break;
8671 		default:
8672 			/*
8673 			 * Set pool property values in the poolprops mos object.
8674 			 */
8675 			if (spa->spa_pool_props_object == 0) {
8676 				spa->spa_pool_props_object =
8677 				    zap_create_link(mos, DMU_OT_POOL_PROPS,
8678 				    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS,
8679 				    tx);
8680 			}
8681 
8682 			/* normalize the property name */
8683 			propname = zpool_prop_to_name(prop);
8684 			proptype = zpool_prop_get_type(prop);
8685 
8686 			if (nvpair_type(elem) == DATA_TYPE_STRING) {
8687 				ASSERT(proptype == PROP_TYPE_STRING);
8688 				strval = fnvpair_value_string(elem);
8689 				VERIFY0(zap_update(mos,
8690 				    spa->spa_pool_props_object, propname,
8691 				    1, strlen(strval) + 1, strval, tx));
8692 				spa_history_log_internal(spa, "set", tx,
8693 				    "%s=%s", nvpair_name(elem), strval);
8694 			} else if (nvpair_type(elem) == DATA_TYPE_UINT64) {
8695 				intval = fnvpair_value_uint64(elem);
8696 
8697 				if (proptype == PROP_TYPE_INDEX) {
8698 					const char *unused;
8699 					VERIFY0(zpool_prop_index_to_string(
8700 					    prop, intval, &unused));
8701 				}
8702 				VERIFY0(zap_update(mos,
8703 				    spa->spa_pool_props_object, propname,
8704 				    8, 1, &intval, tx));
8705 				spa_history_log_internal(spa, "set", tx,
8706 				    "%s=%lld", nvpair_name(elem),
8707 				    (longlong_t)intval);
8708 			} else {
8709 				ASSERT(0); /* not allowed */
8710 			}
8711 
8712 			switch (prop) {
8713 			case ZPOOL_PROP_DELEGATION:
8714 				spa->spa_delegation = intval;
8715 				break;
8716 			case ZPOOL_PROP_BOOTFS:
8717 				spa->spa_bootfs = intval;
8718 				break;
8719 			case ZPOOL_PROP_FAILUREMODE:
8720 				spa->spa_failmode = intval;
8721 				break;
8722 			case ZPOOL_PROP_AUTOTRIM:
8723 				spa->spa_autotrim = intval;
8724 				spa_async_request(spa,
8725 				    SPA_ASYNC_AUTOTRIM_RESTART);
8726 				break;
8727 			case ZPOOL_PROP_AUTOEXPAND:
8728 				spa->spa_autoexpand = intval;
8729 				if (tx->tx_txg != TXG_INITIAL)
8730 					spa_async_request(spa,
8731 					    SPA_ASYNC_AUTOEXPAND);
8732 				break;
8733 			case ZPOOL_PROP_MULTIHOST:
8734 				spa->spa_multihost = intval;
8735 				break;
8736 			default:
8737 				break;
8738 			}
8739 		}
8740 
8741 	}
8742 
8743 	mutex_exit(&spa->spa_props_lock);
8744 }
8745 
8746 /*
8747  * Perform one-time upgrade on-disk changes.  spa_version() does not
8748  * reflect the new version this txg, so there must be no changes this
8749  * txg to anything that the upgrade code depends on after it executes.
8750  * Therefore this must be called after dsl_pool_sync() does the sync
8751  * tasks.
8752  */
8753 static void
8754 spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx)
8755 {
8756 	if (spa_sync_pass(spa) != 1)
8757 		return;
8758 
8759 	dsl_pool_t *dp = spa->spa_dsl_pool;
8760 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
8761 
8762 	if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN &&
8763 	    spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) {
8764 		dsl_pool_create_origin(dp, tx);
8765 
8766 		/* Keeping the origin open increases spa_minref */
8767 		spa->spa_minref += 3;
8768 	}
8769 
8770 	if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES &&
8771 	    spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) {
8772 		dsl_pool_upgrade_clones(dp, tx);
8773 	}
8774 
8775 	if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES &&
8776 	    spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) {
8777 		dsl_pool_upgrade_dir_clones(dp, tx);
8778 
8779 		/* Keeping the freedir open increases spa_minref */
8780 		spa->spa_minref += 3;
8781 	}
8782 
8783 	if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES &&
8784 	    spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
8785 		spa_feature_create_zap_objects(spa, tx);
8786 	}
8787 
8788 	/*
8789 	 * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable
8790 	 * when possibility to use lz4 compression for metadata was added
8791 	 * Old pools that have this feature enabled must be upgraded to have
8792 	 * this feature active
8793 	 */
8794 	if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
8795 		boolean_t lz4_en = spa_feature_is_enabled(spa,
8796 		    SPA_FEATURE_LZ4_COMPRESS);
8797 		boolean_t lz4_ac = spa_feature_is_active(spa,
8798 		    SPA_FEATURE_LZ4_COMPRESS);
8799 
8800 		if (lz4_en && !lz4_ac)
8801 			spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx);
8802 	}
8803 
8804 	/*
8805 	 * If we haven't written the salt, do so now.  Note that the
8806 	 * feature may not be activated yet, but that's fine since
8807 	 * the presence of this ZAP entry is backwards compatible.
8808 	 */
8809 	if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
8810 	    DMU_POOL_CHECKSUM_SALT) == ENOENT) {
8811 		VERIFY0(zap_add(spa->spa_meta_objset,
8812 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1,
8813 		    sizeof (spa->spa_cksum_salt.zcs_bytes),
8814 		    spa->spa_cksum_salt.zcs_bytes, tx));
8815 	}
8816 
8817 	rrw_exit(&dp->dp_config_rwlock, FTAG);
8818 }
8819 
8820 static void
8821 vdev_indirect_state_sync_verify(vdev_t *vd)
8822 {
8823 	vdev_indirect_mapping_t *vim __maybe_unused = vd->vdev_indirect_mapping;
8824 	vdev_indirect_births_t *vib __maybe_unused = vd->vdev_indirect_births;
8825 
8826 	if (vd->vdev_ops == &vdev_indirect_ops) {
8827 		ASSERT(vim != NULL);
8828 		ASSERT(vib != NULL);
8829 	}
8830 
8831 	uint64_t obsolete_sm_object = 0;
8832 	ASSERT0(vdev_obsolete_sm_object(vd, &obsolete_sm_object));
8833 	if (obsolete_sm_object != 0) {
8834 		ASSERT(vd->vdev_obsolete_sm != NULL);
8835 		ASSERT(vd->vdev_removing ||
8836 		    vd->vdev_ops == &vdev_indirect_ops);
8837 		ASSERT(vdev_indirect_mapping_num_entries(vim) > 0);
8838 		ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0);
8839 		ASSERT3U(obsolete_sm_object, ==,
8840 		    space_map_object(vd->vdev_obsolete_sm));
8841 		ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=,
8842 		    space_map_allocated(vd->vdev_obsolete_sm));
8843 	}
8844 	ASSERT(vd->vdev_obsolete_segments != NULL);
8845 
8846 	/*
8847 	 * Since frees / remaps to an indirect vdev can only
8848 	 * happen in syncing context, the obsolete segments
8849 	 * tree must be empty when we start syncing.
8850 	 */
8851 	ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
8852 }
8853 
8854 /*
8855  * Set the top-level vdev's max queue depth. Evaluate each top-level's
8856  * async write queue depth in case it changed. The max queue depth will
8857  * not change in the middle of syncing out this txg.
8858  */
8859 static void
8860 spa_sync_adjust_vdev_max_queue_depth(spa_t *spa)
8861 {
8862 	ASSERT(spa_writeable(spa));
8863 
8864 	vdev_t *rvd = spa->spa_root_vdev;
8865 	uint32_t max_queue_depth = zfs_vdev_async_write_max_active *
8866 	    zfs_vdev_queue_depth_pct / 100;
8867 	metaslab_class_t *normal = spa_normal_class(spa);
8868 	metaslab_class_t *special = spa_special_class(spa);
8869 	metaslab_class_t *dedup = spa_dedup_class(spa);
8870 
8871 	uint64_t slots_per_allocator = 0;
8872 	for (int c = 0; c < rvd->vdev_children; c++) {
8873 		vdev_t *tvd = rvd->vdev_child[c];
8874 
8875 		metaslab_group_t *mg = tvd->vdev_mg;
8876 		if (mg == NULL || !metaslab_group_initialized(mg))
8877 			continue;
8878 
8879 		metaslab_class_t *mc = mg->mg_class;
8880 		if (mc != normal && mc != special && mc != dedup)
8881 			continue;
8882 
8883 		/*
8884 		 * It is safe to do a lock-free check here because only async
8885 		 * allocations look at mg_max_alloc_queue_depth, and async
8886 		 * allocations all happen from spa_sync().
8887 		 */
8888 		for (int i = 0; i < mg->mg_allocators; i++) {
8889 			ASSERT0(zfs_refcount_count(
8890 			    &(mg->mg_allocator[i].mga_alloc_queue_depth)));
8891 		}
8892 		mg->mg_max_alloc_queue_depth = max_queue_depth;
8893 
8894 		for (int i = 0; i < mg->mg_allocators; i++) {
8895 			mg->mg_allocator[i].mga_cur_max_alloc_queue_depth =
8896 			    zfs_vdev_def_queue_depth;
8897 		}
8898 		slots_per_allocator += zfs_vdev_def_queue_depth;
8899 	}
8900 
8901 	for (int i = 0; i < spa->spa_alloc_count; i++) {
8902 		ASSERT0(zfs_refcount_count(&normal->mc_allocator[i].
8903 		    mca_alloc_slots));
8904 		ASSERT0(zfs_refcount_count(&special->mc_allocator[i].
8905 		    mca_alloc_slots));
8906 		ASSERT0(zfs_refcount_count(&dedup->mc_allocator[i].
8907 		    mca_alloc_slots));
8908 		normal->mc_allocator[i].mca_alloc_max_slots =
8909 		    slots_per_allocator;
8910 		special->mc_allocator[i].mca_alloc_max_slots =
8911 		    slots_per_allocator;
8912 		dedup->mc_allocator[i].mca_alloc_max_slots =
8913 		    slots_per_allocator;
8914 	}
8915 	normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
8916 	special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
8917 	dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
8918 }
8919 
8920 static void
8921 spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx)
8922 {
8923 	ASSERT(spa_writeable(spa));
8924 
8925 	vdev_t *rvd = spa->spa_root_vdev;
8926 	for (int c = 0; c < rvd->vdev_children; c++) {
8927 		vdev_t *vd = rvd->vdev_child[c];
8928 		vdev_indirect_state_sync_verify(vd);
8929 
8930 		if (vdev_indirect_should_condense(vd)) {
8931 			spa_condense_indirect_start_sync(vd, tx);
8932 			break;
8933 		}
8934 	}
8935 }
8936 
8937 static void
8938 spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx)
8939 {
8940 	objset_t *mos = spa->spa_meta_objset;
8941 	dsl_pool_t *dp = spa->spa_dsl_pool;
8942 	uint64_t txg = tx->tx_txg;
8943 	bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK];
8944 
8945 	do {
8946 		int pass = ++spa->spa_sync_pass;
8947 
8948 		spa_sync_config_object(spa, tx);
8949 		spa_sync_aux_dev(spa, &spa->spa_spares, tx,
8950 		    ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES);
8951 		spa_sync_aux_dev(spa, &spa->spa_l2cache, tx,
8952 		    ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE);
8953 		spa_errlog_sync(spa, txg);
8954 		dsl_pool_sync(dp, txg);
8955 
8956 		if (pass < zfs_sync_pass_deferred_free ||
8957 		    spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) {
8958 			/*
8959 			 * If the log space map feature is active we don't
8960 			 * care about deferred frees and the deferred bpobj
8961 			 * as the log space map should effectively have the
8962 			 * same results (i.e. appending only to one object).
8963 			 */
8964 			spa_sync_frees(spa, free_bpl, tx);
8965 		} else {
8966 			/*
8967 			 * We can not defer frees in pass 1, because
8968 			 * we sync the deferred frees later in pass 1.
8969 			 */
8970 			ASSERT3U(pass, >, 1);
8971 			bplist_iterate(free_bpl, bpobj_enqueue_alloc_cb,
8972 			    &spa->spa_deferred_bpobj, tx);
8973 		}
8974 
8975 		ddt_sync(spa, txg);
8976 		dsl_scan_sync(dp, tx);
8977 		svr_sync(spa, tx);
8978 		spa_sync_upgrades(spa, tx);
8979 
8980 		spa_flush_metaslabs(spa, tx);
8981 
8982 		vdev_t *vd = NULL;
8983 		while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg))
8984 		    != NULL)
8985 			vdev_sync(vd, txg);
8986 
8987 		/*
8988 		 * Note: We need to check if the MOS is dirty because we could
8989 		 * have marked the MOS dirty without updating the uberblock
8990 		 * (e.g. if we have sync tasks but no dirty user data). We need
8991 		 * to check the uberblock's rootbp because it is updated if we
8992 		 * have synced out dirty data (though in this case the MOS will
8993 		 * most likely also be dirty due to second order effects, we
8994 		 * don't want to rely on that here).
8995 		 */
8996 		if (pass == 1 &&
8997 		    spa->spa_uberblock.ub_rootbp.blk_birth < txg &&
8998 		    !dmu_objset_is_dirty(mos, txg)) {
8999 			/*
9000 			 * Nothing changed on the first pass, therefore this
9001 			 * TXG is a no-op. Avoid syncing deferred frees, so
9002 			 * that we can keep this TXG as a no-op.
9003 			 */
9004 			ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
9005 			ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
9006 			ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg));
9007 			ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg));
9008 			break;
9009 		}
9010 
9011 		spa_sync_deferred_frees(spa, tx);
9012 	} while (dmu_objset_is_dirty(mos, txg));
9013 }
9014 
9015 /*
9016  * Rewrite the vdev configuration (which includes the uberblock) to
9017  * commit the transaction group.
9018  *
9019  * If there are no dirty vdevs, we sync the uberblock to a few random
9020  * top-level vdevs that are known to be visible in the config cache
9021  * (see spa_vdev_add() for a complete description). If there *are* dirty
9022  * vdevs, sync the uberblock to all vdevs.
9023  */
9024 static void
9025 spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx)
9026 {
9027 	vdev_t *rvd = spa->spa_root_vdev;
9028 	uint64_t txg = tx->tx_txg;
9029 
9030 	for (;;) {
9031 		int error = 0;
9032 
9033 		/*
9034 		 * We hold SCL_STATE to prevent vdev open/close/etc.
9035 		 * while we're attempting to write the vdev labels.
9036 		 */
9037 		spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
9038 
9039 		if (list_is_empty(&spa->spa_config_dirty_list)) {
9040 			vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
9041 			int svdcount = 0;
9042 			int children = rvd->vdev_children;
9043 			int c0 = spa_get_random(children);
9044 
9045 			for (int c = 0; c < children; c++) {
9046 				vdev_t *vd =
9047 				    rvd->vdev_child[(c0 + c) % children];
9048 
9049 				/* Stop when revisiting the first vdev */
9050 				if (c > 0 && svd[0] == vd)
9051 					break;
9052 
9053 				if (vd->vdev_ms_array == 0 ||
9054 				    vd->vdev_islog ||
9055 				    !vdev_is_concrete(vd))
9056 					continue;
9057 
9058 				svd[svdcount++] = vd;
9059 				if (svdcount == SPA_SYNC_MIN_VDEVS)
9060 					break;
9061 			}
9062 			error = vdev_config_sync(svd, svdcount, txg);
9063 		} else {
9064 			error = vdev_config_sync(rvd->vdev_child,
9065 			    rvd->vdev_children, txg);
9066 		}
9067 
9068 		if (error == 0)
9069 			spa->spa_last_synced_guid = rvd->vdev_guid;
9070 
9071 		spa_config_exit(spa, SCL_STATE, FTAG);
9072 
9073 		if (error == 0)
9074 			break;
9075 		zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR);
9076 		zio_resume_wait(spa);
9077 	}
9078 }
9079 
9080 /*
9081  * Sync the specified transaction group.  New blocks may be dirtied as
9082  * part of the process, so we iterate until it converges.
9083  */
9084 void
9085 spa_sync(spa_t *spa, uint64_t txg)
9086 {
9087 	vdev_t *vd = NULL;
9088 
9089 	VERIFY(spa_writeable(spa));
9090 
9091 	/*
9092 	 * Wait for i/os issued in open context that need to complete
9093 	 * before this txg syncs.
9094 	 */
9095 	(void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]);
9096 	spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL,
9097 	    ZIO_FLAG_CANFAIL);
9098 
9099 	/*
9100 	 * Lock out configuration changes.
9101 	 */
9102 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
9103 
9104 	spa->spa_syncing_txg = txg;
9105 	spa->spa_sync_pass = 0;
9106 
9107 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9108 		mutex_enter(&spa->spa_alloc_locks[i]);
9109 		VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i]));
9110 		mutex_exit(&spa->spa_alloc_locks[i]);
9111 	}
9112 
9113 	/*
9114 	 * If there are any pending vdev state changes, convert them
9115 	 * into config changes that go out with this transaction group.
9116 	 */
9117 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
9118 	while (list_head(&spa->spa_state_dirty_list) != NULL) {
9119 		/*
9120 		 * We need the write lock here because, for aux vdevs,
9121 		 * calling vdev_config_dirty() modifies sav_config.
9122 		 * This is ugly and will become unnecessary when we
9123 		 * eliminate the aux vdev wart by integrating all vdevs
9124 		 * into the root vdev tree.
9125 		 */
9126 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9127 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER);
9128 		while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) {
9129 			vdev_state_clean(vd);
9130 			vdev_config_dirty(vd);
9131 		}
9132 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9133 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9134 	}
9135 	spa_config_exit(spa, SCL_STATE, FTAG);
9136 
9137 	dsl_pool_t *dp = spa->spa_dsl_pool;
9138 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
9139 
9140 	spa->spa_sync_starttime = gethrtime();
9141 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
9142 	spa->spa_deadman_tqid = taskq_dispatch_delay(system_delay_taskq,
9143 	    spa_deadman, spa, TQ_SLEEP, ddi_get_lbolt() +
9144 	    NSEC_TO_TICK(spa->spa_deadman_synctime));
9145 
9146 	/*
9147 	 * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg,
9148 	 * set spa_deflate if we have no raid-z vdevs.
9149 	 */
9150 	if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE &&
9151 	    spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) {
9152 		vdev_t *rvd = spa->spa_root_vdev;
9153 
9154 		int i;
9155 		for (i = 0; i < rvd->vdev_children; i++) {
9156 			vd = rvd->vdev_child[i];
9157 			if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE)
9158 				break;
9159 		}
9160 		if (i == rvd->vdev_children) {
9161 			spa->spa_deflate = TRUE;
9162 			VERIFY0(zap_add(spa->spa_meta_objset,
9163 			    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
9164 			    sizeof (uint64_t), 1, &spa->spa_deflate, tx));
9165 		}
9166 	}
9167 
9168 	spa_sync_adjust_vdev_max_queue_depth(spa);
9169 
9170 	spa_sync_condense_indirect(spa, tx);
9171 
9172 	spa_sync_iterate_to_convergence(spa, tx);
9173 
9174 #ifdef ZFS_DEBUG
9175 	if (!list_is_empty(&spa->spa_config_dirty_list)) {
9176 	/*
9177 	 * Make sure that the number of ZAPs for all the vdevs matches
9178 	 * the number of ZAPs in the per-vdev ZAP list. This only gets
9179 	 * called if the config is dirty; otherwise there may be
9180 	 * outstanding AVZ operations that weren't completed in
9181 	 * spa_sync_config_object.
9182 	 */
9183 		uint64_t all_vdev_zap_entry_count;
9184 		ASSERT0(zap_count(spa->spa_meta_objset,
9185 		    spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count));
9186 		ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==,
9187 		    all_vdev_zap_entry_count);
9188 	}
9189 #endif
9190 
9191 	if (spa->spa_vdev_removal != NULL) {
9192 		ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]);
9193 	}
9194 
9195 	spa_sync_rewrite_vdev_config(spa, tx);
9196 	dmu_tx_commit(tx);
9197 
9198 	taskq_cancel_id(system_delay_taskq, spa->spa_deadman_tqid);
9199 	spa->spa_deadman_tqid = 0;
9200 
9201 	/*
9202 	 * Clear the dirty config list.
9203 	 */
9204 	while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL)
9205 		vdev_config_clean(vd);
9206 
9207 	/*
9208 	 * Now that the new config has synced transactionally,
9209 	 * let it become visible to the config cache.
9210 	 */
9211 	if (spa->spa_config_syncing != NULL) {
9212 		spa_config_set(spa, spa->spa_config_syncing);
9213 		spa->spa_config_txg = txg;
9214 		spa->spa_config_syncing = NULL;
9215 	}
9216 
9217 	dsl_pool_sync_done(dp, txg);
9218 
9219 	for (int i = 0; i < spa->spa_alloc_count; i++) {
9220 		mutex_enter(&spa->spa_alloc_locks[i]);
9221 		VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i]));
9222 		mutex_exit(&spa->spa_alloc_locks[i]);
9223 	}
9224 
9225 	/*
9226 	 * Update usable space statistics.
9227 	 */
9228 	while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)))
9229 	    != NULL)
9230 		vdev_sync_done(vd, txg);
9231 
9232 	metaslab_class_evict_old(spa->spa_normal_class, txg);
9233 	metaslab_class_evict_old(spa->spa_log_class, txg);
9234 
9235 	spa_sync_close_syncing_log_sm(spa);
9236 
9237 	spa_update_dspace(spa);
9238 
9239 	/*
9240 	 * It had better be the case that we didn't dirty anything
9241 	 * since vdev_config_sync().
9242 	 */
9243 	ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
9244 	ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
9245 	ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg));
9246 
9247 	while (zfs_pause_spa_sync)
9248 		delay(1);
9249 
9250 	spa->spa_sync_pass = 0;
9251 
9252 	/*
9253 	 * Update the last synced uberblock here. We want to do this at
9254 	 * the end of spa_sync() so that consumers of spa_last_synced_txg()
9255 	 * will be guaranteed that all the processing associated with
9256 	 * that txg has been completed.
9257 	 */
9258 	spa->spa_ubsync = spa->spa_uberblock;
9259 	spa_config_exit(spa, SCL_CONFIG, FTAG);
9260 
9261 	spa_handle_ignored_writes(spa);
9262 
9263 	/*
9264 	 * If any async tasks have been requested, kick them off.
9265 	 */
9266 	spa_async_dispatch(spa);
9267 }
9268 
9269 /*
9270  * Sync all pools.  We don't want to hold the namespace lock across these
9271  * operations, so we take a reference on the spa_t and drop the lock during the
9272  * sync.
9273  */
9274 void
9275 spa_sync_allpools(void)
9276 {
9277 	spa_t *spa = NULL;
9278 	mutex_enter(&spa_namespace_lock);
9279 	while ((spa = spa_next(spa)) != NULL) {
9280 		if (spa_state(spa) != POOL_STATE_ACTIVE ||
9281 		    !spa_writeable(spa) || spa_suspended(spa))
9282 			continue;
9283 		spa_open_ref(spa, FTAG);
9284 		mutex_exit(&spa_namespace_lock);
9285 		txg_wait_synced(spa_get_dsl(spa), 0);
9286 		mutex_enter(&spa_namespace_lock);
9287 		spa_close(spa, FTAG);
9288 	}
9289 	mutex_exit(&spa_namespace_lock);
9290 }
9291 
9292 /*
9293  * ==========================================================================
9294  * Miscellaneous routines
9295  * ==========================================================================
9296  */
9297 
9298 /*
9299  * Remove all pools in the system.
9300  */
9301 void
9302 spa_evict_all(void)
9303 {
9304 	spa_t *spa;
9305 
9306 	/*
9307 	 * Remove all cached state.  All pools should be closed now,
9308 	 * so every spa in the AVL tree should be unreferenced.
9309 	 */
9310 	mutex_enter(&spa_namespace_lock);
9311 	while ((spa = spa_next(NULL)) != NULL) {
9312 		/*
9313 		 * Stop async tasks.  The async thread may need to detach
9314 		 * a device that's been replaced, which requires grabbing
9315 		 * spa_namespace_lock, so we must drop it here.
9316 		 */
9317 		spa_open_ref(spa, FTAG);
9318 		mutex_exit(&spa_namespace_lock);
9319 		spa_async_suspend(spa);
9320 		mutex_enter(&spa_namespace_lock);
9321 		spa_close(spa, FTAG);
9322 
9323 		if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
9324 			spa_unload(spa);
9325 			spa_deactivate(spa);
9326 		}
9327 		spa_remove(spa);
9328 	}
9329 	mutex_exit(&spa_namespace_lock);
9330 }
9331 
9332 vdev_t *
9333 spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux)
9334 {
9335 	vdev_t *vd;
9336 	int i;
9337 
9338 	if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL)
9339 		return (vd);
9340 
9341 	if (aux) {
9342 		for (i = 0; i < spa->spa_l2cache.sav_count; i++) {
9343 			vd = spa->spa_l2cache.sav_vdevs[i];
9344 			if (vd->vdev_guid == guid)
9345 				return (vd);
9346 		}
9347 
9348 		for (i = 0; i < spa->spa_spares.sav_count; i++) {
9349 			vd = spa->spa_spares.sav_vdevs[i];
9350 			if (vd->vdev_guid == guid)
9351 				return (vd);
9352 		}
9353 	}
9354 
9355 	return (NULL);
9356 }
9357 
9358 void
9359 spa_upgrade(spa_t *spa, uint64_t version)
9360 {
9361 	ASSERT(spa_writeable(spa));
9362 
9363 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
9364 
9365 	/*
9366 	 * This should only be called for a non-faulted pool, and since a
9367 	 * future version would result in an unopenable pool, this shouldn't be
9368 	 * possible.
9369 	 */
9370 	ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version));
9371 	ASSERT3U(version, >=, spa->spa_uberblock.ub_version);
9372 
9373 	spa->spa_uberblock.ub_version = version;
9374 	vdev_config_dirty(spa->spa_root_vdev);
9375 
9376 	spa_config_exit(spa, SCL_ALL, FTAG);
9377 
9378 	txg_wait_synced(spa_get_dsl(spa), 0);
9379 }
9380 
9381 boolean_t
9382 spa_has_spare(spa_t *spa, uint64_t guid)
9383 {
9384 	int i;
9385 	uint64_t spareguid;
9386 	spa_aux_vdev_t *sav = &spa->spa_spares;
9387 
9388 	for (i = 0; i < sav->sav_count; i++)
9389 		if (sav->sav_vdevs[i]->vdev_guid == guid)
9390 			return (B_TRUE);
9391 
9392 	for (i = 0; i < sav->sav_npending; i++) {
9393 		if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID,
9394 		    &spareguid) == 0 && spareguid == guid)
9395 			return (B_TRUE);
9396 	}
9397 
9398 	return (B_FALSE);
9399 }
9400 
9401 /*
9402  * Check if a pool has an active shared spare device.
9403  * Note: reference count of an active spare is 2, as a spare and as a replace
9404  */
9405 static boolean_t
9406 spa_has_active_shared_spare(spa_t *spa)
9407 {
9408 	int i, refcnt;
9409 	uint64_t pool;
9410 	spa_aux_vdev_t *sav = &spa->spa_spares;
9411 
9412 	for (i = 0; i < sav->sav_count; i++) {
9413 		if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool,
9414 		    &refcnt) && pool != 0ULL && pool == spa_guid(spa) &&
9415 		    refcnt > 2)
9416 			return (B_TRUE);
9417 	}
9418 
9419 	return (B_FALSE);
9420 }
9421 
9422 uint64_t
9423 spa_total_metaslabs(spa_t *spa)
9424 {
9425 	vdev_t *rvd = spa->spa_root_vdev;
9426 
9427 	uint64_t m = 0;
9428 	for (uint64_t c = 0; c < rvd->vdev_children; c++) {
9429 		vdev_t *vd = rvd->vdev_child[c];
9430 		if (!vdev_is_concrete(vd))
9431 			continue;
9432 		m += vd->vdev_ms_count;
9433 	}
9434 	return (m);
9435 }
9436 
9437 /*
9438  * Notify any waiting threads that some activity has switched from being in-
9439  * progress to not-in-progress so that the thread can wake up and determine
9440  * whether it is finished waiting.
9441  */
9442 void
9443 spa_notify_waiters(spa_t *spa)
9444 {
9445 	/*
9446 	 * Acquiring spa_activities_lock here prevents the cv_broadcast from
9447 	 * happening between the waiting thread's check and cv_wait.
9448 	 */
9449 	mutex_enter(&spa->spa_activities_lock);
9450 	cv_broadcast(&spa->spa_activities_cv);
9451 	mutex_exit(&spa->spa_activities_lock);
9452 }
9453 
9454 /*
9455  * Notify any waiting threads that the pool is exporting, and then block until
9456  * they are finished using the spa_t.
9457  */
9458 void
9459 spa_wake_waiters(spa_t *spa)
9460 {
9461 	mutex_enter(&spa->spa_activities_lock);
9462 	spa->spa_waiters_cancel = B_TRUE;
9463 	cv_broadcast(&spa->spa_activities_cv);
9464 	while (spa->spa_waiters != 0)
9465 		cv_wait(&spa->spa_waiters_cv, &spa->spa_activities_lock);
9466 	spa->spa_waiters_cancel = B_FALSE;
9467 	mutex_exit(&spa->spa_activities_lock);
9468 }
9469 
9470 /* Whether the vdev or any of its descendants are being initialized/trimmed. */
9471 static boolean_t
9472 spa_vdev_activity_in_progress_impl(vdev_t *vd, zpool_wait_activity_t activity)
9473 {
9474 	spa_t *spa = vd->vdev_spa;
9475 
9476 	ASSERT(spa_config_held(spa, SCL_CONFIG | SCL_STATE, RW_READER));
9477 	ASSERT(MUTEX_HELD(&spa->spa_activities_lock));
9478 	ASSERT(activity == ZPOOL_WAIT_INITIALIZE ||
9479 	    activity == ZPOOL_WAIT_TRIM);
9480 
9481 	kmutex_t *lock = activity == ZPOOL_WAIT_INITIALIZE ?
9482 	    &vd->vdev_initialize_lock : &vd->vdev_trim_lock;
9483 
9484 	mutex_exit(&spa->spa_activities_lock);
9485 	mutex_enter(lock);
9486 	mutex_enter(&spa->spa_activities_lock);
9487 
9488 	boolean_t in_progress = (activity == ZPOOL_WAIT_INITIALIZE) ?
9489 	    (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE) :
9490 	    (vd->vdev_trim_state == VDEV_TRIM_ACTIVE);
9491 	mutex_exit(lock);
9492 
9493 	if (in_progress)
9494 		return (B_TRUE);
9495 
9496 	for (int i = 0; i < vd->vdev_children; i++) {
9497 		if (spa_vdev_activity_in_progress_impl(vd->vdev_child[i],
9498 		    activity))
9499 			return (B_TRUE);
9500 	}
9501 
9502 	return (B_FALSE);
9503 }
9504 
9505 /*
9506  * If use_guid is true, this checks whether the vdev specified by guid is
9507  * being initialized/trimmed. Otherwise, it checks whether any vdev in the pool
9508  * is being initialized/trimmed. The caller must hold the config lock and
9509  * spa_activities_lock.
9510  */
9511 static int
9512 spa_vdev_activity_in_progress(spa_t *spa, boolean_t use_guid, uint64_t guid,
9513     zpool_wait_activity_t activity, boolean_t *in_progress)
9514 {
9515 	mutex_exit(&spa->spa_activities_lock);
9516 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9517 	mutex_enter(&spa->spa_activities_lock);
9518 
9519 	vdev_t *vd;
9520 	if (use_guid) {
9521 		vd = spa_lookup_by_guid(spa, guid, B_FALSE);
9522 		if (vd == NULL || !vd->vdev_ops->vdev_op_leaf) {
9523 			spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9524 			return (EINVAL);
9525 		}
9526 	} else {
9527 		vd = spa->spa_root_vdev;
9528 	}
9529 
9530 	*in_progress = spa_vdev_activity_in_progress_impl(vd, activity);
9531 
9532 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9533 	return (0);
9534 }
9535 
9536 /*
9537  * Locking for waiting threads
9538  * ---------------------------
9539  *
9540  * Waiting threads need a way to check whether a given activity is in progress,
9541  * and then, if it is, wait for it to complete. Each activity will have some
9542  * in-memory representation of the relevant on-disk state which can be used to
9543  * determine whether or not the activity is in progress. The in-memory state and
9544  * the locking used to protect it will be different for each activity, and may
9545  * not be suitable for use with a cvar (e.g., some state is protected by the
9546  * config lock). To allow waiting threads to wait without any races, another
9547  * lock, spa_activities_lock, is used.
9548  *
9549  * When the state is checked, both the activity-specific lock (if there is one)
9550  * and spa_activities_lock are held. In some cases, the activity-specific lock
9551  * is acquired explicitly (e.g. the config lock). In others, the locking is
9552  * internal to some check (e.g. bpobj_is_empty). After checking, the waiting
9553  * thread releases the activity-specific lock and, if the activity is in
9554  * progress, then cv_waits using spa_activities_lock.
9555  *
9556  * The waiting thread is woken when another thread, one completing some
9557  * activity, updates the state of the activity and then calls
9558  * spa_notify_waiters, which will cv_broadcast. This 'completing' thread only
9559  * needs to hold its activity-specific lock when updating the state, and this
9560  * lock can (but doesn't have to) be dropped before calling spa_notify_waiters.
9561  *
9562  * Because spa_notify_waiters acquires spa_activities_lock before broadcasting,
9563  * and because it is held when the waiting thread checks the state of the
9564  * activity, it can never be the case that the completing thread both updates
9565  * the activity state and cv_broadcasts in between the waiting thread's check
9566  * and cv_wait. Thus, a waiting thread can never miss a wakeup.
9567  *
9568  * In order to prevent deadlock, when the waiting thread does its check, in some
9569  * cases it will temporarily drop spa_activities_lock in order to acquire the
9570  * activity-specific lock. The order in which spa_activities_lock and the
9571  * activity specific lock are acquired in the waiting thread is determined by
9572  * the order in which they are acquired in the completing thread; if the
9573  * completing thread calls spa_notify_waiters with the activity-specific lock
9574  * held, then the waiting thread must also acquire the activity-specific lock
9575  * first.
9576  */
9577 
9578 static int
9579 spa_activity_in_progress(spa_t *spa, zpool_wait_activity_t activity,
9580     boolean_t use_tag, uint64_t tag, boolean_t *in_progress)
9581 {
9582 	int error = 0;
9583 
9584 	ASSERT(MUTEX_HELD(&spa->spa_activities_lock));
9585 
9586 	switch (activity) {
9587 	case ZPOOL_WAIT_CKPT_DISCARD:
9588 		*in_progress =
9589 		    (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT) &&
9590 		    zap_contains(spa_meta_objset(spa),
9591 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_ZPOOL_CHECKPOINT) ==
9592 		    ENOENT);
9593 		break;
9594 	case ZPOOL_WAIT_FREE:
9595 		*in_progress = ((spa_version(spa) >= SPA_VERSION_DEADLISTS &&
9596 		    !bpobj_is_empty(&spa->spa_dsl_pool->dp_free_bpobj)) ||
9597 		    spa_feature_is_active(spa, SPA_FEATURE_ASYNC_DESTROY) ||
9598 		    spa_livelist_delete_check(spa));
9599 		break;
9600 	case ZPOOL_WAIT_INITIALIZE:
9601 	case ZPOOL_WAIT_TRIM:
9602 		error = spa_vdev_activity_in_progress(spa, use_tag, tag,
9603 		    activity, in_progress);
9604 		break;
9605 	case ZPOOL_WAIT_REPLACE:
9606 		mutex_exit(&spa->spa_activities_lock);
9607 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
9608 		mutex_enter(&spa->spa_activities_lock);
9609 
9610 		*in_progress = vdev_replace_in_progress(spa->spa_root_vdev);
9611 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
9612 		break;
9613 	case ZPOOL_WAIT_REMOVE:
9614 		*in_progress = (spa->spa_removing_phys.sr_state ==
9615 		    DSS_SCANNING);
9616 		break;
9617 	case ZPOOL_WAIT_RESILVER:
9618 		if ((*in_progress = vdev_rebuild_active(spa->spa_root_vdev)))
9619 			break;
9620 		/* fall through */
9621 	case ZPOOL_WAIT_SCRUB:
9622 	{
9623 		boolean_t scanning, paused, is_scrub;
9624 		dsl_scan_t *scn =  spa->spa_dsl_pool->dp_scan;
9625 
9626 		is_scrub = (scn->scn_phys.scn_func == POOL_SCAN_SCRUB);
9627 		scanning = (scn->scn_phys.scn_state == DSS_SCANNING);
9628 		paused = dsl_scan_is_paused_scrub(scn);
9629 		*in_progress = (scanning && !paused &&
9630 		    is_scrub == (activity == ZPOOL_WAIT_SCRUB));
9631 		break;
9632 	}
9633 	default:
9634 		panic("unrecognized value for activity %d", activity);
9635 	}
9636 
9637 	return (error);
9638 }
9639 
9640 static int
9641 spa_wait_common(const char *pool, zpool_wait_activity_t activity,
9642     boolean_t use_tag, uint64_t tag, boolean_t *waited)
9643 {
9644 	/*
9645 	 * The tag is used to distinguish between instances of an activity.
9646 	 * 'initialize' and 'trim' are the only activities that we use this for.
9647 	 * The other activities can only have a single instance in progress in a
9648 	 * pool at one time, making the tag unnecessary.
9649 	 *
9650 	 * There can be multiple devices being replaced at once, but since they
9651 	 * all finish once resilvering finishes, we don't bother keeping track
9652 	 * of them individually, we just wait for them all to finish.
9653 	 */
9654 	if (use_tag && activity != ZPOOL_WAIT_INITIALIZE &&
9655 	    activity != ZPOOL_WAIT_TRIM)
9656 		return (EINVAL);
9657 
9658 	if (activity < 0 || activity >= ZPOOL_WAIT_NUM_ACTIVITIES)
9659 		return (EINVAL);
9660 
9661 	spa_t *spa;
9662 	int error = spa_open(pool, &spa, FTAG);
9663 	if (error != 0)
9664 		return (error);
9665 
9666 	/*
9667 	 * Increment the spa's waiter count so that we can call spa_close and
9668 	 * still ensure that the spa_t doesn't get freed before this thread is
9669 	 * finished with it when the pool is exported. We want to call spa_close
9670 	 * before we start waiting because otherwise the additional ref would
9671 	 * prevent the pool from being exported or destroyed throughout the
9672 	 * potentially long wait.
9673 	 */
9674 	mutex_enter(&spa->spa_activities_lock);
9675 	spa->spa_waiters++;
9676 	spa_close(spa, FTAG);
9677 
9678 	*waited = B_FALSE;
9679 	for (;;) {
9680 		boolean_t in_progress;
9681 		error = spa_activity_in_progress(spa, activity, use_tag, tag,
9682 		    &in_progress);
9683 
9684 		if (error || !in_progress || spa->spa_waiters_cancel)
9685 			break;
9686 
9687 		*waited = B_TRUE;
9688 
9689 		if (cv_wait_sig(&spa->spa_activities_cv,
9690 		    &spa->spa_activities_lock) == 0) {
9691 			error = EINTR;
9692 			break;
9693 		}
9694 	}
9695 
9696 	spa->spa_waiters--;
9697 	cv_signal(&spa->spa_waiters_cv);
9698 	mutex_exit(&spa->spa_activities_lock);
9699 
9700 	return (error);
9701 }
9702 
9703 /*
9704  * Wait for a particular instance of the specified activity to complete, where
9705  * the instance is identified by 'tag'
9706  */
9707 int
9708 spa_wait_tag(const char *pool, zpool_wait_activity_t activity, uint64_t tag,
9709     boolean_t *waited)
9710 {
9711 	return (spa_wait_common(pool, activity, B_TRUE, tag, waited));
9712 }
9713 
9714 /*
9715  * Wait for all instances of the specified activity complete
9716  */
9717 int
9718 spa_wait(const char *pool, zpool_wait_activity_t activity, boolean_t *waited)
9719 {
9720 
9721 	return (spa_wait_common(pool, activity, B_FALSE, 0, waited));
9722 }
9723 
9724 sysevent_t *
9725 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
9726 {
9727 	sysevent_t *ev = NULL;
9728 #ifdef _KERNEL
9729 	nvlist_t *resource;
9730 
9731 	resource = zfs_event_create(spa, vd, FM_SYSEVENT_CLASS, name, hist_nvl);
9732 	if (resource) {
9733 		ev = kmem_alloc(sizeof (sysevent_t), KM_SLEEP);
9734 		ev->resource = resource;
9735 	}
9736 #endif
9737 	return (ev);
9738 }
9739 
9740 void
9741 spa_event_post(sysevent_t *ev)
9742 {
9743 #ifdef _KERNEL
9744 	if (ev) {
9745 		zfs_zevent_post(ev->resource, NULL, zfs_zevent_post_cb);
9746 		kmem_free(ev, sizeof (*ev));
9747 	}
9748 #endif
9749 }
9750 
9751 /*
9752  * Post a zevent corresponding to the given sysevent.   The 'name' must be one
9753  * of the event definitions in sys/sysevent/eventdefs.h.  The payload will be
9754  * filled in from the spa and (optionally) the vdev.  This doesn't do anything
9755  * in the userland libzpool, as we don't want consumers to misinterpret ztest
9756  * or zdb as real changes.
9757  */
9758 void
9759 spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
9760 {
9761 	spa_event_post(spa_event_create(spa, vd, hist_nvl, name));
9762 }
9763 
9764 /* state manipulation functions */
9765 EXPORT_SYMBOL(spa_open);
9766 EXPORT_SYMBOL(spa_open_rewind);
9767 EXPORT_SYMBOL(spa_get_stats);
9768 EXPORT_SYMBOL(spa_create);
9769 EXPORT_SYMBOL(spa_import);
9770 EXPORT_SYMBOL(spa_tryimport);
9771 EXPORT_SYMBOL(spa_destroy);
9772 EXPORT_SYMBOL(spa_export);
9773 EXPORT_SYMBOL(spa_reset);
9774 EXPORT_SYMBOL(spa_async_request);
9775 EXPORT_SYMBOL(spa_async_suspend);
9776 EXPORT_SYMBOL(spa_async_resume);
9777 EXPORT_SYMBOL(spa_inject_addref);
9778 EXPORT_SYMBOL(spa_inject_delref);
9779 EXPORT_SYMBOL(spa_scan_stat_init);
9780 EXPORT_SYMBOL(spa_scan_get_stats);
9781 
9782 /* device manipulation */
9783 EXPORT_SYMBOL(spa_vdev_add);
9784 EXPORT_SYMBOL(spa_vdev_attach);
9785 EXPORT_SYMBOL(spa_vdev_detach);
9786 EXPORT_SYMBOL(spa_vdev_setpath);
9787 EXPORT_SYMBOL(spa_vdev_setfru);
9788 EXPORT_SYMBOL(spa_vdev_split_mirror);
9789 
9790 /* spare statech is global across all pools) */
9791 EXPORT_SYMBOL(spa_spare_add);
9792 EXPORT_SYMBOL(spa_spare_remove);
9793 EXPORT_SYMBOL(spa_spare_exists);
9794 EXPORT_SYMBOL(spa_spare_activate);
9795 
9796 /* L2ARC statech is global across all pools) */
9797 EXPORT_SYMBOL(spa_l2cache_add);
9798 EXPORT_SYMBOL(spa_l2cache_remove);
9799 EXPORT_SYMBOL(spa_l2cache_exists);
9800 EXPORT_SYMBOL(spa_l2cache_activate);
9801 EXPORT_SYMBOL(spa_l2cache_drop);
9802 
9803 /* scanning */
9804 EXPORT_SYMBOL(spa_scan);
9805 EXPORT_SYMBOL(spa_scan_stop);
9806 
9807 /* spa syncing */
9808 EXPORT_SYMBOL(spa_sync); /* only for DMU use */
9809 EXPORT_SYMBOL(spa_sync_allpools);
9810 
9811 /* properties */
9812 EXPORT_SYMBOL(spa_prop_set);
9813 EXPORT_SYMBOL(spa_prop_get);
9814 EXPORT_SYMBOL(spa_prop_clear_bootfs);
9815 
9816 /* asynchronous event notification */
9817 EXPORT_SYMBOL(spa_event_notify);
9818 
9819 /* BEGIN CSTYLED */
9820 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_shift, INT, ZMOD_RW,
9821 	"log2(fraction of arc that can be used by inflight I/Os when "
9822 	"verifying pool during import");
9823 
9824 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_metadata, INT, ZMOD_RW,
9825 	"Set to traverse metadata on pool import");
9826 
9827 ZFS_MODULE_PARAM(zfs_spa, spa_, load_verify_data, INT, ZMOD_RW,
9828 	"Set to traverse data on pool import");
9829 
9830 ZFS_MODULE_PARAM(zfs_spa, spa_, load_print_vdev_tree, INT, ZMOD_RW,
9831 	"Print vdev tree to zfs_dbgmsg during pool import");
9832 
9833 ZFS_MODULE_PARAM(zfs_zio, zio_, taskq_batch_pct, UINT, ZMOD_RD,
9834 	"Percentage of CPUs to run an IO worker thread");
9835 
9836 ZFS_MODULE_PARAM(zfs, zfs_, max_missing_tvds, ULONG, ZMOD_RW,
9837 	"Allow importing pool with up to this number of missing top-level "
9838 	"vdevs (in read-only mode)");
9839 
9840 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_pause, INT, ZMOD_RW,
9841 	"Set the livelist condense zthr to pause");
9842 
9843 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_pause, INT, ZMOD_RW,
9844 	"Set the livelist condense synctask to pause");
9845 
9846 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, sync_cancel, INT, ZMOD_RW,
9847 	"Whether livelist condensing was canceled in the synctask");
9848 
9849 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, zthr_cancel, INT, ZMOD_RW,
9850 	"Whether livelist condensing was canceled in the zthr function");
9851 
9852 ZFS_MODULE_PARAM(zfs_livelist_condense, zfs_livelist_condense_, new_alloc, INT, ZMOD_RW,
9853 	"Whether extra ALLOC blkptrs were added to a livelist entry while it "
9854 	"was being condensed");
9855 /* END CSTYLED */
9856