xref: /illumos-gate/usr/src/uts/common/fs/zfs/spa.c (revision d8109ce4330e1b8ad6c29f9fccacec969066bb9d)
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, 2019 by Delphix. All rights reserved.
25  * Copyright (c) 2015, 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 2018 Joyent, Inc.
31  * Copyright (c) 2017, Intel Corporation.
32  * Copyright (c) 2017 Datto Inc.
33  * Copyright 2018 OmniOS Community Edition (OmniOSce) Association.
34  */
35 
36 /*
37  * SPA: Storage Pool Allocator
38  *
39  * This file contains all the routines used when modifying on-disk SPA state.
40  * This includes opening, importing, destroying, exporting a pool, and syncing a
41  * pool.
42  */
43 
44 #include <sys/zfs_context.h>
45 #include <sys/fm/fs/zfs.h>
46 #include <sys/spa_impl.h>
47 #include <sys/zio.h>
48 #include <sys/zio_checksum.h>
49 #include <sys/dmu.h>
50 #include <sys/dmu_tx.h>
51 #include <sys/zap.h>
52 #include <sys/zil.h>
53 #include <sys/ddt.h>
54 #include <sys/vdev_impl.h>
55 #include <sys/vdev_removal.h>
56 #include <sys/vdev_indirect_mapping.h>
57 #include <sys/vdev_indirect_births.h>
58 #include <sys/vdev_initialize.h>
59 #include <sys/metaslab.h>
60 #include <sys/metaslab_impl.h>
61 #include <sys/mmp.h>
62 #include <sys/uberblock_impl.h>
63 #include <sys/txg.h>
64 #include <sys/avl.h>
65 #include <sys/bpobj.h>
66 #include <sys/dmu_traverse.h>
67 #include <sys/dmu_objset.h>
68 #include <sys/unique.h>
69 #include <sys/dsl_pool.h>
70 #include <sys/dsl_dataset.h>
71 #include <sys/dsl_dir.h>
72 #include <sys/dsl_prop.h>
73 #include <sys/dsl_synctask.h>
74 #include <sys/fs/zfs.h>
75 #include <sys/arc.h>
76 #include <sys/callb.h>
77 #include <sys/systeminfo.h>
78 #include <sys/spa_boot.h>
79 #include <sys/zfs_ioctl.h>
80 #include <sys/dsl_scan.h>
81 #include <sys/zfeature.h>
82 #include <sys/dsl_destroy.h>
83 #include <sys/abd.h>
84 
85 #ifdef	_KERNEL
86 #include <sys/bootprops.h>
87 #include <sys/callb.h>
88 #include <sys/cpupart.h>
89 #include <sys/pool.h>
90 #include <sys/sysdc.h>
91 #include <sys/zone.h>
92 #endif	/* _KERNEL */
93 
94 #include "zfs_prop.h"
95 #include "zfs_comutil.h"
96 
97 /*
98  * The interval, in seconds, at which failed configuration cache file writes
99  * should be retried.
100  */
101 int zfs_ccw_retry_interval = 300;
102 
103 typedef enum zti_modes {
104 	ZTI_MODE_FIXED,			/* value is # of threads (min 1) */
105 	ZTI_MODE_BATCH,			/* cpu-intensive; value is ignored */
106 	ZTI_MODE_NULL,			/* don't create a taskq */
107 	ZTI_NMODES
108 } zti_modes_t;
109 
110 #define	ZTI_P(n, q)	{ ZTI_MODE_FIXED, (n), (q) }
111 #define	ZTI_BATCH	{ ZTI_MODE_BATCH, 0, 1 }
112 #define	ZTI_NULL	{ ZTI_MODE_NULL, 0, 0 }
113 
114 #define	ZTI_N(n)	ZTI_P(n, 1)
115 #define	ZTI_ONE		ZTI_N(1)
116 
117 typedef struct zio_taskq_info {
118 	zti_modes_t zti_mode;
119 	uint_t zti_value;
120 	uint_t zti_count;
121 } zio_taskq_info_t;
122 
123 static const char *const zio_taskq_types[ZIO_TASKQ_TYPES] = {
124 	"issue", "issue_high", "intr", "intr_high"
125 };
126 
127 /*
128  * This table defines the taskq settings for each ZFS I/O type. When
129  * initializing a pool, we use this table to create an appropriately sized
130  * taskq. Some operations are low volume and therefore have a small, static
131  * number of threads assigned to their taskqs using the ZTI_N(#) or ZTI_ONE
132  * macros. Other operations process a large amount of data; the ZTI_BATCH
133  * macro causes us to create a taskq oriented for throughput. Some operations
134  * are so high frequency and short-lived that the taskq itself can become a a
135  * point of lock contention. The ZTI_P(#, #) macro indicates that we need an
136  * additional degree of parallelism specified by the number of threads per-
137  * taskq and the number of taskqs; when dispatching an event in this case, the
138  * particular taskq is chosen at random.
139  *
140  * The different taskq priorities are to handle the different contexts (issue
141  * and interrupt) and then to reserve threads for ZIO_PRIORITY_NOW I/Os that
142  * need to be handled with minimum delay.
143  */
144 const zio_taskq_info_t zio_taskqs[ZIO_TYPES][ZIO_TASKQ_TYPES] = {
145 	/* ISSUE	ISSUE_HIGH	INTR		INTR_HIGH */
146 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* NULL */
147 	{ ZTI_N(8),	ZTI_NULL,	ZTI_P(12, 8),	ZTI_NULL }, /* READ */
148 	{ ZTI_BATCH,	ZTI_N(5),	ZTI_N(8),	ZTI_N(5) }, /* WRITE */
149 	{ ZTI_P(12, 8),	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* FREE */
150 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* CLAIM */
151 	{ ZTI_ONE,	ZTI_NULL,	ZTI_ONE,	ZTI_NULL }, /* IOCTL */
152 };
153 
154 static void spa_sync_version(void *arg, dmu_tx_t *tx);
155 static void spa_sync_props(void *arg, dmu_tx_t *tx);
156 static boolean_t spa_has_active_shared_spare(spa_t *spa);
157 static int spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport);
158 static void spa_vdev_resilver_done(spa_t *spa);
159 
160 uint_t		zio_taskq_batch_pct = 75;	/* 1 thread per cpu in pset */
161 id_t		zio_taskq_psrset_bind = PS_NONE;
162 boolean_t	zio_taskq_sysdc = B_TRUE;	/* use SDC scheduling class */
163 uint_t		zio_taskq_basedc = 80;		/* base duty cycle */
164 
165 boolean_t	spa_create_process = B_TRUE;	/* no process ==> no sysdc */
166 extern int	zfs_sync_pass_deferred_free;
167 
168 /*
169  * Report any spa_load_verify errors found, but do not fail spa_load.
170  * This is used by zdb to analyze non-idle pools.
171  */
172 boolean_t	spa_load_verify_dryrun = B_FALSE;
173 
174 /*
175  * This (illegal) pool name is used when temporarily importing a spa_t in order
176  * to get the vdev stats associated with the imported devices.
177  */
178 #define	TRYIMPORT_NAME	"$import"
179 
180 /*
181  * For debugging purposes: print out vdev tree during pool import.
182  */
183 boolean_t	spa_load_print_vdev_tree = B_FALSE;
184 
185 /*
186  * A non-zero value for zfs_max_missing_tvds means that we allow importing
187  * pools with missing top-level vdevs. This is strictly intended for advanced
188  * pool recovery cases since missing data is almost inevitable. Pools with
189  * missing devices can only be imported read-only for safety reasons, and their
190  * fail-mode will be automatically set to "continue".
191  *
192  * With 1 missing vdev we should be able to import the pool and mount all
193  * datasets. User data that was not modified after the missing device has been
194  * added should be recoverable. This means that snapshots created prior to the
195  * addition of that device should be completely intact.
196  *
197  * With 2 missing vdevs, some datasets may fail to mount since there are
198  * dataset statistics that are stored as regular metadata. Some data might be
199  * recoverable if those vdevs were added recently.
200  *
201  * With 3 or more missing vdevs, the pool is severely damaged and MOS entries
202  * may be missing entirely. Chances of data recovery are very low. Note that
203  * there are also risks of performing an inadvertent rewind as we might be
204  * missing all the vdevs with the latest uberblocks.
205  */
206 uint64_t	zfs_max_missing_tvds = 0;
207 
208 /*
209  * The parameters below are similar to zfs_max_missing_tvds but are only
210  * intended for a preliminary open of the pool with an untrusted config which
211  * might be incomplete or out-dated.
212  *
213  * We are more tolerant for pools opened from a cachefile since we could have
214  * an out-dated cachefile where a device removal was not registered.
215  * We could have set the limit arbitrarily high but in the case where devices
216  * are really missing we would want to return the proper error codes; we chose
217  * SPA_DVAS_PER_BP - 1 so that some copies of the MOS would still be available
218  * and we get a chance to retrieve the trusted config.
219  */
220 uint64_t	zfs_max_missing_tvds_cachefile = SPA_DVAS_PER_BP - 1;
221 
222 /*
223  * In the case where config was assembled by scanning device paths (/dev/dsks
224  * by default) we are less tolerant since all the existing devices should have
225  * been detected and we want spa_load to return the right error codes.
226  */
227 uint64_t	zfs_max_missing_tvds_scan = 0;
228 
229 /*
230  * Debugging aid that pauses spa_sync() towards the end.
231  */
232 boolean_t	zfs_pause_spa_sync = B_FALSE;
233 
234 /*
235  * ==========================================================================
236  * SPA properties routines
237  * ==========================================================================
238  */
239 
240 /*
241  * Add a (source=src, propname=propval) list to an nvlist.
242  */
243 static void
244 spa_prop_add_list(nvlist_t *nvl, zpool_prop_t prop, char *strval,
245     uint64_t intval, zprop_source_t src)
246 {
247 	const char *propname = zpool_prop_to_name(prop);
248 	nvlist_t *propval;
249 
250 	VERIFY(nvlist_alloc(&propval, NV_UNIQUE_NAME, KM_SLEEP) == 0);
251 	VERIFY(nvlist_add_uint64(propval, ZPROP_SOURCE, src) == 0);
252 
253 	if (strval != NULL)
254 		VERIFY(nvlist_add_string(propval, ZPROP_VALUE, strval) == 0);
255 	else
256 		VERIFY(nvlist_add_uint64(propval, ZPROP_VALUE, intval) == 0);
257 
258 	VERIFY(nvlist_add_nvlist(nvl, propname, propval) == 0);
259 	nvlist_free(propval);
260 }
261 
262 /*
263  * Get property values from the spa configuration.
264  */
265 static void
266 spa_prop_get_config(spa_t *spa, nvlist_t **nvp)
267 {
268 	vdev_t *rvd = spa->spa_root_vdev;
269 	dsl_pool_t *pool = spa->spa_dsl_pool;
270 	uint64_t size, alloc, cap, version;
271 	zprop_source_t src = ZPROP_SRC_NONE;
272 	spa_config_dirent_t *dp;
273 	metaslab_class_t *mc = spa_normal_class(spa);
274 
275 	ASSERT(MUTEX_HELD(&spa->spa_props_lock));
276 
277 	if (rvd != NULL) {
278 		alloc = metaslab_class_get_alloc(mc);
279 		alloc += metaslab_class_get_alloc(spa_special_class(spa));
280 		alloc += metaslab_class_get_alloc(spa_dedup_class(spa));
281 
282 		size = metaslab_class_get_space(mc);
283 		size += metaslab_class_get_space(spa_special_class(spa));
284 		size += metaslab_class_get_space(spa_dedup_class(spa));
285 
286 		spa_prop_add_list(*nvp, ZPOOL_PROP_NAME, spa_name(spa), 0, src);
287 		spa_prop_add_list(*nvp, ZPOOL_PROP_SIZE, NULL, size, src);
288 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALLOCATED, NULL, alloc, src);
289 		spa_prop_add_list(*nvp, ZPOOL_PROP_FREE, NULL,
290 		    size - alloc, src);
291 		spa_prop_add_list(*nvp, ZPOOL_PROP_CHECKPOINT, NULL,
292 		    spa->spa_checkpoint_info.sci_dspace, src);
293 
294 		spa_prop_add_list(*nvp, ZPOOL_PROP_FRAGMENTATION, NULL,
295 		    metaslab_class_fragmentation(mc), src);
296 		spa_prop_add_list(*nvp, ZPOOL_PROP_EXPANDSZ, NULL,
297 		    metaslab_class_expandable_space(mc), src);
298 		spa_prop_add_list(*nvp, ZPOOL_PROP_READONLY, NULL,
299 		    (spa_mode(spa) == FREAD), src);
300 
301 		cap = (size == 0) ? 0 : (alloc * 100 / size);
302 		spa_prop_add_list(*nvp, ZPOOL_PROP_CAPACITY, NULL, cap, src);
303 
304 		spa_prop_add_list(*nvp, ZPOOL_PROP_DEDUPRATIO, NULL,
305 		    ddt_get_pool_dedup_ratio(spa), src);
306 
307 		spa_prop_add_list(*nvp, ZPOOL_PROP_HEALTH, NULL,
308 		    rvd->vdev_state, src);
309 
310 		version = spa_version(spa);
311 		if (version == zpool_prop_default_numeric(ZPOOL_PROP_VERSION))
312 			src = ZPROP_SRC_DEFAULT;
313 		else
314 			src = ZPROP_SRC_LOCAL;
315 		spa_prop_add_list(*nvp, ZPOOL_PROP_VERSION, NULL, version, src);
316 	}
317 
318 	if (pool != NULL) {
319 		/*
320 		 * The $FREE directory was introduced in SPA_VERSION_DEADLISTS,
321 		 * when opening pools before this version freedir will be NULL.
322 		 */
323 		if (pool->dp_free_dir != NULL) {
324 			spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING, NULL,
325 			    dsl_dir_phys(pool->dp_free_dir)->dd_used_bytes,
326 			    src);
327 		} else {
328 			spa_prop_add_list(*nvp, ZPOOL_PROP_FREEING,
329 			    NULL, 0, src);
330 		}
331 
332 		if (pool->dp_leak_dir != NULL) {
333 			spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED, NULL,
334 			    dsl_dir_phys(pool->dp_leak_dir)->dd_used_bytes,
335 			    src);
336 		} else {
337 			spa_prop_add_list(*nvp, ZPOOL_PROP_LEAKED,
338 			    NULL, 0, src);
339 		}
340 	}
341 
342 	spa_prop_add_list(*nvp, ZPOOL_PROP_GUID, NULL, spa_guid(spa), src);
343 
344 	if (spa->spa_comment != NULL) {
345 		spa_prop_add_list(*nvp, ZPOOL_PROP_COMMENT, spa->spa_comment,
346 		    0, ZPROP_SRC_LOCAL);
347 	}
348 
349 	if (spa->spa_root != NULL)
350 		spa_prop_add_list(*nvp, ZPOOL_PROP_ALTROOT, spa->spa_root,
351 		    0, ZPROP_SRC_LOCAL);
352 
353 	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_BLOCKS)) {
354 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
355 		    MIN(zfs_max_recordsize, SPA_MAXBLOCKSIZE), ZPROP_SRC_NONE);
356 	} else {
357 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXBLOCKSIZE, NULL,
358 		    SPA_OLD_MAXBLOCKSIZE, ZPROP_SRC_NONE);
359 	}
360 
361 	if (spa_feature_is_enabled(spa, SPA_FEATURE_LARGE_DNODE)) {
362 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL,
363 		    DNODE_MAX_SIZE, ZPROP_SRC_NONE);
364 	} else {
365 		spa_prop_add_list(*nvp, ZPOOL_PROP_MAXDNODESIZE, NULL,
366 		    DNODE_MIN_SIZE, ZPROP_SRC_NONE);
367 	}
368 
369 	if ((dp = list_head(&spa->spa_config_list)) != NULL) {
370 		if (dp->scd_path == NULL) {
371 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
372 			    "none", 0, ZPROP_SRC_LOCAL);
373 		} else if (strcmp(dp->scd_path, spa_config_path) != 0) {
374 			spa_prop_add_list(*nvp, ZPOOL_PROP_CACHEFILE,
375 			    dp->scd_path, 0, ZPROP_SRC_LOCAL);
376 		}
377 	}
378 }
379 
380 /*
381  * Get zpool property values.
382  */
383 int
384 spa_prop_get(spa_t *spa, nvlist_t **nvp)
385 {
386 	objset_t *mos = spa->spa_meta_objset;
387 	zap_cursor_t zc;
388 	zap_attribute_t za;
389 	int err;
390 
391 	VERIFY(nvlist_alloc(nvp, NV_UNIQUE_NAME, KM_SLEEP) == 0);
392 
393 	mutex_enter(&spa->spa_props_lock);
394 
395 	/*
396 	 * Get properties from the spa config.
397 	 */
398 	spa_prop_get_config(spa, nvp);
399 
400 	/* If no pool property object, no more prop to get. */
401 	if (mos == NULL || spa->spa_pool_props_object == 0) {
402 		mutex_exit(&spa->spa_props_lock);
403 		return (0);
404 	}
405 
406 	/*
407 	 * Get properties from the MOS pool property object.
408 	 */
409 	for (zap_cursor_init(&zc, mos, spa->spa_pool_props_object);
410 	    (err = zap_cursor_retrieve(&zc, &za)) == 0;
411 	    zap_cursor_advance(&zc)) {
412 		uint64_t intval = 0;
413 		char *strval = NULL;
414 		zprop_source_t src = ZPROP_SRC_DEFAULT;
415 		zpool_prop_t prop;
416 
417 		if ((prop = zpool_name_to_prop(za.za_name)) == ZPOOL_PROP_INVAL)
418 			continue;
419 
420 		switch (za.za_integer_length) {
421 		case 8:
422 			/* integer property */
423 			if (za.za_first_integer !=
424 			    zpool_prop_default_numeric(prop))
425 				src = ZPROP_SRC_LOCAL;
426 
427 			if (prop == ZPOOL_PROP_BOOTFS) {
428 				dsl_pool_t *dp;
429 				dsl_dataset_t *ds = NULL;
430 
431 				dp = spa_get_dsl(spa);
432 				dsl_pool_config_enter(dp, FTAG);
433 				err = dsl_dataset_hold_obj(dp,
434 				    za.za_first_integer, FTAG, &ds);
435 				if (err != 0) {
436 					dsl_pool_config_exit(dp, FTAG);
437 					break;
438 				}
439 
440 				strval = kmem_alloc(ZFS_MAX_DATASET_NAME_LEN,
441 				    KM_SLEEP);
442 				dsl_dataset_name(ds, strval);
443 				dsl_dataset_rele(ds, FTAG);
444 				dsl_pool_config_exit(dp, FTAG);
445 			} else {
446 				strval = NULL;
447 				intval = za.za_first_integer;
448 			}
449 
450 			spa_prop_add_list(*nvp, prop, strval, intval, src);
451 
452 			if (strval != NULL)
453 				kmem_free(strval, ZFS_MAX_DATASET_NAME_LEN);
454 
455 			break;
456 
457 		case 1:
458 			/* string property */
459 			strval = kmem_alloc(za.za_num_integers, KM_SLEEP);
460 			err = zap_lookup(mos, spa->spa_pool_props_object,
461 			    za.za_name, 1, za.za_num_integers, strval);
462 			if (err) {
463 				kmem_free(strval, za.za_num_integers);
464 				break;
465 			}
466 			spa_prop_add_list(*nvp, prop, strval, 0, src);
467 			kmem_free(strval, za.za_num_integers);
468 			break;
469 
470 		default:
471 			break;
472 		}
473 	}
474 	zap_cursor_fini(&zc);
475 	mutex_exit(&spa->spa_props_lock);
476 out:
477 	if (err && err != ENOENT) {
478 		nvlist_free(*nvp);
479 		*nvp = NULL;
480 		return (err);
481 	}
482 
483 	return (0);
484 }
485 
486 /*
487  * Validate the given pool properties nvlist and modify the list
488  * for the property values to be set.
489  */
490 static int
491 spa_prop_validate(spa_t *spa, nvlist_t *props)
492 {
493 	nvpair_t *elem;
494 	int error = 0, reset_bootfs = 0;
495 	uint64_t objnum = 0;
496 	boolean_t has_feature = B_FALSE;
497 
498 	elem = NULL;
499 	while ((elem = nvlist_next_nvpair(props, elem)) != NULL) {
500 		uint64_t intval;
501 		char *strval, *slash, *check, *fname;
502 		const char *propname = nvpair_name(elem);
503 		zpool_prop_t prop = zpool_name_to_prop(propname);
504 
505 		switch (prop) {
506 		case ZPOOL_PROP_INVAL:
507 			if (!zpool_prop_feature(propname)) {
508 				error = SET_ERROR(EINVAL);
509 				break;
510 			}
511 
512 			/*
513 			 * Sanitize the input.
514 			 */
515 			if (nvpair_type(elem) != DATA_TYPE_UINT64) {
516 				error = SET_ERROR(EINVAL);
517 				break;
518 			}
519 
520 			if (nvpair_value_uint64(elem, &intval) != 0) {
521 				error = SET_ERROR(EINVAL);
522 				break;
523 			}
524 
525 			if (intval != 0) {
526 				error = SET_ERROR(EINVAL);
527 				break;
528 			}
529 
530 			fname = strchr(propname, '@') + 1;
531 			if (zfeature_lookup_name(fname, NULL) != 0) {
532 				error = SET_ERROR(EINVAL);
533 				break;
534 			}
535 
536 			has_feature = B_TRUE;
537 			break;
538 
539 		case ZPOOL_PROP_VERSION:
540 			error = nvpair_value_uint64(elem, &intval);
541 			if (!error &&
542 			    (intval < spa_version(spa) ||
543 			    intval > SPA_VERSION_BEFORE_FEATURES ||
544 			    has_feature))
545 				error = SET_ERROR(EINVAL);
546 			break;
547 
548 		case ZPOOL_PROP_DELEGATION:
549 		case ZPOOL_PROP_AUTOREPLACE:
550 		case ZPOOL_PROP_LISTSNAPS:
551 		case ZPOOL_PROP_AUTOEXPAND:
552 			error = nvpair_value_uint64(elem, &intval);
553 			if (!error && intval > 1)
554 				error = SET_ERROR(EINVAL);
555 			break;
556 
557 		case ZPOOL_PROP_MULTIHOST:
558 			error = nvpair_value_uint64(elem, &intval);
559 			if (!error && intval > 1)
560 				error = SET_ERROR(EINVAL);
561 
562 			if (!error && !spa_get_hostid())
563 				error = SET_ERROR(ENOTSUP);
564 
565 			break;
566 
567 		case ZPOOL_PROP_BOOTFS:
568 			/*
569 			 * If the pool version is less than SPA_VERSION_BOOTFS,
570 			 * or the pool is still being created (version == 0),
571 			 * the bootfs property cannot be set.
572 			 */
573 			if (spa_version(spa) < SPA_VERSION_BOOTFS) {
574 				error = SET_ERROR(ENOTSUP);
575 				break;
576 			}
577 
578 			/*
579 			 * Make sure the vdev config is bootable
580 			 */
581 			if (!vdev_is_bootable(spa->spa_root_vdev)) {
582 				error = SET_ERROR(ENOTSUP);
583 				break;
584 			}
585 
586 			reset_bootfs = 1;
587 
588 			error = nvpair_value_string(elem, &strval);
589 
590 			if (!error) {
591 				objset_t *os;
592 				uint64_t propval;
593 
594 				if (strval == NULL || strval[0] == '\0') {
595 					objnum = zpool_prop_default_numeric(
596 					    ZPOOL_PROP_BOOTFS);
597 					break;
598 				}
599 
600 				error = dmu_objset_hold(strval, FTAG, &os);
601 				if (error != 0)
602 					break;
603 
604 				/*
605 				 * Must be ZPL, and its property settings
606 				 * must be supported.
607 				 */
608 
609 				if (dmu_objset_type(os) != DMU_OST_ZFS) {
610 					error = SET_ERROR(ENOTSUP);
611 				} else if ((error =
612 				    dsl_prop_get_int_ds(dmu_objset_ds(os),
613 				    zfs_prop_to_name(ZFS_PROP_COMPRESSION),
614 				    &propval)) == 0 &&
615 				    !BOOTFS_COMPRESS_VALID(propval)) {
616 					error = SET_ERROR(ENOTSUP);
617 				} else {
618 					objnum = dmu_objset_id(os);
619 				}
620 				dmu_objset_rele(os, FTAG);
621 			}
622 			break;
623 
624 		case ZPOOL_PROP_FAILUREMODE:
625 			error = nvpair_value_uint64(elem, &intval);
626 			if (!error && (intval < ZIO_FAILURE_MODE_WAIT ||
627 			    intval > ZIO_FAILURE_MODE_PANIC))
628 				error = SET_ERROR(EINVAL);
629 
630 			/*
631 			 * This is a special case which only occurs when
632 			 * the pool has completely failed. This allows
633 			 * the user to change the in-core failmode property
634 			 * without syncing it out to disk (I/Os might
635 			 * currently be blocked). We do this by returning
636 			 * EIO to the caller (spa_prop_set) to trick it
637 			 * into thinking we encountered a property validation
638 			 * error.
639 			 */
640 			if (!error && spa_suspended(spa)) {
641 				spa->spa_failmode = intval;
642 				error = SET_ERROR(EIO);
643 			}
644 			break;
645 
646 		case ZPOOL_PROP_CACHEFILE:
647 			if ((error = nvpair_value_string(elem, &strval)) != 0)
648 				break;
649 
650 			if (strval[0] == '\0')
651 				break;
652 
653 			if (strcmp(strval, "none") == 0)
654 				break;
655 
656 			if (strval[0] != '/') {
657 				error = SET_ERROR(EINVAL);
658 				break;
659 			}
660 
661 			slash = strrchr(strval, '/');
662 			ASSERT(slash != NULL);
663 
664 			if (slash[1] == '\0' || strcmp(slash, "/.") == 0 ||
665 			    strcmp(slash, "/..") == 0)
666 				error = SET_ERROR(EINVAL);
667 			break;
668 
669 		case ZPOOL_PROP_COMMENT:
670 			if ((error = nvpair_value_string(elem, &strval)) != 0)
671 				break;
672 			for (check = strval; *check != '\0'; check++) {
673 				/*
674 				 * The kernel doesn't have an easy isprint()
675 				 * check.  For this kernel check, we merely
676 				 * check ASCII apart from DEL.  Fix this if
677 				 * there is an easy-to-use kernel isprint().
678 				 */
679 				if (*check >= 0x7f) {
680 					error = SET_ERROR(EINVAL);
681 					break;
682 				}
683 			}
684 			if (strlen(strval) > ZPROP_MAX_COMMENT)
685 				error = E2BIG;
686 			break;
687 
688 		case ZPOOL_PROP_DEDUPDITTO:
689 			if (spa_version(spa) < SPA_VERSION_DEDUP)
690 				error = SET_ERROR(ENOTSUP);
691 			else
692 				error = nvpair_value_uint64(elem, &intval);
693 			if (error == 0 &&
694 			    intval != 0 && intval < ZIO_DEDUPDITTO_MIN)
695 				error = SET_ERROR(EINVAL);
696 			break;
697 		}
698 
699 		if (error)
700 			break;
701 	}
702 
703 	if (!error && reset_bootfs) {
704 		error = nvlist_remove(props,
705 		    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), DATA_TYPE_STRING);
706 
707 		if (!error) {
708 			error = nvlist_add_uint64(props,
709 			    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), objnum);
710 		}
711 	}
712 
713 	return (error);
714 }
715 
716 void
717 spa_configfile_set(spa_t *spa, nvlist_t *nvp, boolean_t need_sync)
718 {
719 	char *cachefile;
720 	spa_config_dirent_t *dp;
721 
722 	if (nvlist_lookup_string(nvp, zpool_prop_to_name(ZPOOL_PROP_CACHEFILE),
723 	    &cachefile) != 0)
724 		return;
725 
726 	dp = kmem_alloc(sizeof (spa_config_dirent_t),
727 	    KM_SLEEP);
728 
729 	if (cachefile[0] == '\0')
730 		dp->scd_path = spa_strdup(spa_config_path);
731 	else if (strcmp(cachefile, "none") == 0)
732 		dp->scd_path = NULL;
733 	else
734 		dp->scd_path = spa_strdup(cachefile);
735 
736 	list_insert_head(&spa->spa_config_list, dp);
737 	if (need_sync)
738 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
739 }
740 
741 int
742 spa_prop_set(spa_t *spa, nvlist_t *nvp)
743 {
744 	int error;
745 	nvpair_t *elem = NULL;
746 	boolean_t need_sync = B_FALSE;
747 
748 	if ((error = spa_prop_validate(spa, nvp)) != 0)
749 		return (error);
750 
751 	while ((elem = nvlist_next_nvpair(nvp, elem)) != NULL) {
752 		zpool_prop_t prop = zpool_name_to_prop(nvpair_name(elem));
753 
754 		if (prop == ZPOOL_PROP_CACHEFILE ||
755 		    prop == ZPOOL_PROP_ALTROOT ||
756 		    prop == ZPOOL_PROP_READONLY)
757 			continue;
758 
759 		if (prop == ZPOOL_PROP_VERSION || prop == ZPOOL_PROP_INVAL) {
760 			uint64_t ver;
761 
762 			if (prop == ZPOOL_PROP_VERSION) {
763 				VERIFY(nvpair_value_uint64(elem, &ver) == 0);
764 			} else {
765 				ASSERT(zpool_prop_feature(nvpair_name(elem)));
766 				ver = SPA_VERSION_FEATURES;
767 				need_sync = B_TRUE;
768 			}
769 
770 			/* Save time if the version is already set. */
771 			if (ver == spa_version(spa))
772 				continue;
773 
774 			/*
775 			 * In addition to the pool directory object, we might
776 			 * create the pool properties object, the features for
777 			 * read object, the features for write object, or the
778 			 * feature descriptions object.
779 			 */
780 			error = dsl_sync_task(spa->spa_name, NULL,
781 			    spa_sync_version, &ver,
782 			    6, ZFS_SPACE_CHECK_RESERVED);
783 			if (error)
784 				return (error);
785 			continue;
786 		}
787 
788 		need_sync = B_TRUE;
789 		break;
790 	}
791 
792 	if (need_sync) {
793 		return (dsl_sync_task(spa->spa_name, NULL, spa_sync_props,
794 		    nvp, 6, ZFS_SPACE_CHECK_RESERVED));
795 	}
796 
797 	return (0);
798 }
799 
800 /*
801  * If the bootfs property value is dsobj, clear it.
802  */
803 void
804 spa_prop_clear_bootfs(spa_t *spa, uint64_t dsobj, dmu_tx_t *tx)
805 {
806 	if (spa->spa_bootfs == dsobj && spa->spa_pool_props_object != 0) {
807 		VERIFY(zap_remove(spa->spa_meta_objset,
808 		    spa->spa_pool_props_object,
809 		    zpool_prop_to_name(ZPOOL_PROP_BOOTFS), tx) == 0);
810 		spa->spa_bootfs = 0;
811 	}
812 }
813 
814 /*ARGSUSED*/
815 static int
816 spa_change_guid_check(void *arg, dmu_tx_t *tx)
817 {
818 	uint64_t *newguid = arg;
819 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
820 	vdev_t *rvd = spa->spa_root_vdev;
821 	uint64_t vdev_state;
822 
823 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
824 		int error = (spa_has_checkpoint(spa)) ?
825 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
826 		return (SET_ERROR(error));
827 	}
828 
829 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
830 	vdev_state = rvd->vdev_state;
831 	spa_config_exit(spa, SCL_STATE, FTAG);
832 
833 	if (vdev_state != VDEV_STATE_HEALTHY)
834 		return (SET_ERROR(ENXIO));
835 
836 	ASSERT3U(spa_guid(spa), !=, *newguid);
837 
838 	return (0);
839 }
840 
841 static void
842 spa_change_guid_sync(void *arg, dmu_tx_t *tx)
843 {
844 	uint64_t *newguid = arg;
845 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
846 	uint64_t oldguid;
847 	vdev_t *rvd = spa->spa_root_vdev;
848 
849 	oldguid = spa_guid(spa);
850 
851 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
852 	rvd->vdev_guid = *newguid;
853 	rvd->vdev_guid_sum += (*newguid - oldguid);
854 	vdev_config_dirty(rvd);
855 	spa_config_exit(spa, SCL_STATE, FTAG);
856 
857 	spa_history_log_internal(spa, "guid change", tx, "old=%llu new=%llu",
858 	    oldguid, *newguid);
859 }
860 
861 /*
862  * Change the GUID for the pool.  This is done so that we can later
863  * re-import a pool built from a clone of our own vdevs.  We will modify
864  * the root vdev's guid, our own pool guid, and then mark all of our
865  * vdevs dirty.  Note that we must make sure that all our vdevs are
866  * online when we do this, or else any vdevs that weren't present
867  * would be orphaned from our pool.  We are also going to issue a
868  * sysevent to update any watchers.
869  */
870 int
871 spa_change_guid(spa_t *spa)
872 {
873 	int error;
874 	uint64_t guid;
875 
876 	mutex_enter(&spa->spa_vdev_top_lock);
877 	mutex_enter(&spa_namespace_lock);
878 	guid = spa_generate_guid(NULL);
879 
880 	error = dsl_sync_task(spa->spa_name, spa_change_guid_check,
881 	    spa_change_guid_sync, &guid, 5, ZFS_SPACE_CHECK_RESERVED);
882 
883 	if (error == 0) {
884 		spa_write_cachefile(spa, B_FALSE, B_TRUE);
885 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_REGUID);
886 	}
887 
888 	mutex_exit(&spa_namespace_lock);
889 	mutex_exit(&spa->spa_vdev_top_lock);
890 
891 	return (error);
892 }
893 
894 /*
895  * ==========================================================================
896  * SPA state manipulation (open/create/destroy/import/export)
897  * ==========================================================================
898  */
899 
900 static int
901 spa_error_entry_compare(const void *a, const void *b)
902 {
903 	const spa_error_entry_t *sa = (const spa_error_entry_t *)a;
904 	const spa_error_entry_t *sb = (const spa_error_entry_t *)b;
905 	int ret;
906 
907 	ret = memcmp(&sa->se_bookmark, &sb->se_bookmark,
908 	    sizeof (zbookmark_phys_t));
909 
910 	return (AVL_ISIGN(ret));
911 }
912 
913 /*
914  * Utility function which retrieves copies of the current logs and
915  * re-initializes them in the process.
916  */
917 void
918 spa_get_errlists(spa_t *spa, avl_tree_t *last, avl_tree_t *scrub)
919 {
920 	ASSERT(MUTEX_HELD(&spa->spa_errlist_lock));
921 
922 	bcopy(&spa->spa_errlist_last, last, sizeof (avl_tree_t));
923 	bcopy(&spa->spa_errlist_scrub, scrub, sizeof (avl_tree_t));
924 
925 	avl_create(&spa->spa_errlist_scrub,
926 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
927 	    offsetof(spa_error_entry_t, se_avl));
928 	avl_create(&spa->spa_errlist_last,
929 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
930 	    offsetof(spa_error_entry_t, se_avl));
931 }
932 
933 static void
934 spa_taskqs_init(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
935 {
936 	const zio_taskq_info_t *ztip = &zio_taskqs[t][q];
937 	enum zti_modes mode = ztip->zti_mode;
938 	uint_t value = ztip->zti_value;
939 	uint_t count = ztip->zti_count;
940 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
941 	char name[32];
942 	uint_t flags = 0;
943 	boolean_t batch = B_FALSE;
944 
945 	if (mode == ZTI_MODE_NULL) {
946 		tqs->stqs_count = 0;
947 		tqs->stqs_taskq = NULL;
948 		return;
949 	}
950 
951 	ASSERT3U(count, >, 0);
952 
953 	tqs->stqs_count = count;
954 	tqs->stqs_taskq = kmem_alloc(count * sizeof (taskq_t *), KM_SLEEP);
955 
956 	switch (mode) {
957 	case ZTI_MODE_FIXED:
958 		ASSERT3U(value, >=, 1);
959 		value = MAX(value, 1);
960 		break;
961 
962 	case ZTI_MODE_BATCH:
963 		batch = B_TRUE;
964 		flags |= TASKQ_THREADS_CPU_PCT;
965 		value = zio_taskq_batch_pct;
966 		break;
967 
968 	default:
969 		panic("unrecognized mode for %s_%s taskq (%u:%u) in "
970 		    "spa_activate()",
971 		    zio_type_name[t], zio_taskq_types[q], mode, value);
972 		break;
973 	}
974 
975 	for (uint_t i = 0; i < count; i++) {
976 		taskq_t *tq;
977 
978 		if (count > 1) {
979 			(void) snprintf(name, sizeof (name), "%s_%s_%u",
980 			    zio_type_name[t], zio_taskq_types[q], i);
981 		} else {
982 			(void) snprintf(name, sizeof (name), "%s_%s",
983 			    zio_type_name[t], zio_taskq_types[q]);
984 		}
985 
986 		if (zio_taskq_sysdc && spa->spa_proc != &p0) {
987 			if (batch)
988 				flags |= TASKQ_DC_BATCH;
989 
990 			tq = taskq_create_sysdc(name, value, 50, INT_MAX,
991 			    spa->spa_proc, zio_taskq_basedc, flags);
992 		} else {
993 			pri_t pri = maxclsyspri;
994 			/*
995 			 * The write issue taskq can be extremely CPU
996 			 * intensive.  Run it at slightly lower priority
997 			 * than the other taskqs.
998 			 */
999 			if (t == ZIO_TYPE_WRITE && q == ZIO_TASKQ_ISSUE)
1000 				pri--;
1001 
1002 			tq = taskq_create_proc(name, value, pri, 50,
1003 			    INT_MAX, spa->spa_proc, flags);
1004 		}
1005 
1006 		tqs->stqs_taskq[i] = tq;
1007 	}
1008 }
1009 
1010 static void
1011 spa_taskqs_fini(spa_t *spa, zio_type_t t, zio_taskq_type_t q)
1012 {
1013 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1014 
1015 	if (tqs->stqs_taskq == NULL) {
1016 		ASSERT0(tqs->stqs_count);
1017 		return;
1018 	}
1019 
1020 	for (uint_t i = 0; i < tqs->stqs_count; i++) {
1021 		ASSERT3P(tqs->stqs_taskq[i], !=, NULL);
1022 		taskq_destroy(tqs->stqs_taskq[i]);
1023 	}
1024 
1025 	kmem_free(tqs->stqs_taskq, tqs->stqs_count * sizeof (taskq_t *));
1026 	tqs->stqs_taskq = NULL;
1027 }
1028 
1029 /*
1030  * Dispatch a task to the appropriate taskq for the ZFS I/O type and priority.
1031  * Note that a type may have multiple discrete taskqs to avoid lock contention
1032  * on the taskq itself. In that case we choose which taskq at random by using
1033  * the low bits of gethrtime().
1034  */
1035 void
1036 spa_taskq_dispatch_ent(spa_t *spa, zio_type_t t, zio_taskq_type_t q,
1037     task_func_t *func, void *arg, uint_t flags, taskq_ent_t *ent)
1038 {
1039 	spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q];
1040 	taskq_t *tq;
1041 
1042 	ASSERT3P(tqs->stqs_taskq, !=, NULL);
1043 	ASSERT3U(tqs->stqs_count, !=, 0);
1044 
1045 	if (tqs->stqs_count == 1) {
1046 		tq = tqs->stqs_taskq[0];
1047 	} else {
1048 		tq = tqs->stqs_taskq[gethrtime() % tqs->stqs_count];
1049 	}
1050 
1051 	taskq_dispatch_ent(tq, func, arg, flags, ent);
1052 }
1053 
1054 static void
1055 spa_create_zio_taskqs(spa_t *spa)
1056 {
1057 	for (int t = 0; t < ZIO_TYPES; t++) {
1058 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1059 			spa_taskqs_init(spa, t, q);
1060 		}
1061 	}
1062 }
1063 
1064 #ifdef _KERNEL
1065 static void
1066 spa_thread(void *arg)
1067 {
1068 	callb_cpr_t cprinfo;
1069 
1070 	spa_t *spa = arg;
1071 	user_t *pu = PTOU(curproc);
1072 
1073 	CALLB_CPR_INIT(&cprinfo, &spa->spa_proc_lock, callb_generic_cpr,
1074 	    spa->spa_name);
1075 
1076 	ASSERT(curproc != &p0);
1077 	(void) snprintf(pu->u_psargs, sizeof (pu->u_psargs),
1078 	    "zpool-%s", spa->spa_name);
1079 	(void) strlcpy(pu->u_comm, pu->u_psargs, sizeof (pu->u_comm));
1080 
1081 	/* bind this thread to the requested psrset */
1082 	if (zio_taskq_psrset_bind != PS_NONE) {
1083 		pool_lock();
1084 		mutex_enter(&cpu_lock);
1085 		mutex_enter(&pidlock);
1086 		mutex_enter(&curproc->p_lock);
1087 
1088 		if (cpupart_bind_thread(curthread, zio_taskq_psrset_bind,
1089 		    0, NULL, NULL) == 0)  {
1090 			curthread->t_bind_pset = zio_taskq_psrset_bind;
1091 		} else {
1092 			cmn_err(CE_WARN,
1093 			    "Couldn't bind process for zfs pool \"%s\" to "
1094 			    "pset %d\n", spa->spa_name, zio_taskq_psrset_bind);
1095 		}
1096 
1097 		mutex_exit(&curproc->p_lock);
1098 		mutex_exit(&pidlock);
1099 		mutex_exit(&cpu_lock);
1100 		pool_unlock();
1101 	}
1102 
1103 	if (zio_taskq_sysdc) {
1104 		sysdc_thread_enter(curthread, 100, 0);
1105 	}
1106 
1107 	spa->spa_proc = curproc;
1108 	spa->spa_did = curthread->t_did;
1109 
1110 	spa_create_zio_taskqs(spa);
1111 
1112 	mutex_enter(&spa->spa_proc_lock);
1113 	ASSERT(spa->spa_proc_state == SPA_PROC_CREATED);
1114 
1115 	spa->spa_proc_state = SPA_PROC_ACTIVE;
1116 	cv_broadcast(&spa->spa_proc_cv);
1117 
1118 	CALLB_CPR_SAFE_BEGIN(&cprinfo);
1119 	while (spa->spa_proc_state == SPA_PROC_ACTIVE)
1120 		cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1121 	CALLB_CPR_SAFE_END(&cprinfo, &spa->spa_proc_lock);
1122 
1123 	ASSERT(spa->spa_proc_state == SPA_PROC_DEACTIVATE);
1124 	spa->spa_proc_state = SPA_PROC_GONE;
1125 	spa->spa_proc = &p0;
1126 	cv_broadcast(&spa->spa_proc_cv);
1127 	CALLB_CPR_EXIT(&cprinfo);	/* drops spa_proc_lock */
1128 
1129 	mutex_enter(&curproc->p_lock);
1130 	lwp_exit();
1131 }
1132 #endif
1133 
1134 /*
1135  * Activate an uninitialized pool.
1136  */
1137 static void
1138 spa_activate(spa_t *spa, int mode)
1139 {
1140 	ASSERT(spa->spa_state == POOL_STATE_UNINITIALIZED);
1141 
1142 	spa->spa_state = POOL_STATE_ACTIVE;
1143 	spa->spa_mode = mode;
1144 
1145 	spa->spa_normal_class = metaslab_class_create(spa, zfs_metaslab_ops);
1146 	spa->spa_log_class = metaslab_class_create(spa, zfs_metaslab_ops);
1147 	spa->spa_special_class = metaslab_class_create(spa, zfs_metaslab_ops);
1148 	spa->spa_dedup_class = metaslab_class_create(spa, zfs_metaslab_ops);
1149 
1150 	/* Try to create a covering process */
1151 	mutex_enter(&spa->spa_proc_lock);
1152 	ASSERT(spa->spa_proc_state == SPA_PROC_NONE);
1153 	ASSERT(spa->spa_proc == &p0);
1154 	spa->spa_did = 0;
1155 
1156 	/* Only create a process if we're going to be around a while. */
1157 	if (spa_create_process && strcmp(spa->spa_name, TRYIMPORT_NAME) != 0) {
1158 		if (newproc(spa_thread, (caddr_t)spa, syscid, maxclsyspri,
1159 		    NULL, 0) == 0) {
1160 			spa->spa_proc_state = SPA_PROC_CREATED;
1161 			while (spa->spa_proc_state == SPA_PROC_CREATED) {
1162 				cv_wait(&spa->spa_proc_cv,
1163 				    &spa->spa_proc_lock);
1164 			}
1165 			ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1166 			ASSERT(spa->spa_proc != &p0);
1167 			ASSERT(spa->spa_did != 0);
1168 		} else {
1169 #ifdef _KERNEL
1170 			cmn_err(CE_WARN,
1171 			    "Couldn't create process for zfs pool \"%s\"\n",
1172 			    spa->spa_name);
1173 #endif
1174 		}
1175 	}
1176 	mutex_exit(&spa->spa_proc_lock);
1177 
1178 	/* If we didn't create a process, we need to create our taskqs. */
1179 	if (spa->spa_proc == &p0) {
1180 		spa_create_zio_taskqs(spa);
1181 	}
1182 
1183 	for (size_t i = 0; i < TXG_SIZE; i++) {
1184 		spa->spa_txg_zio[i] = zio_root(spa, NULL, NULL,
1185 		    ZIO_FLAG_CANFAIL);
1186 	}
1187 
1188 	list_create(&spa->spa_config_dirty_list, sizeof (vdev_t),
1189 	    offsetof(vdev_t, vdev_config_dirty_node));
1190 	list_create(&spa->spa_evicting_os_list, sizeof (objset_t),
1191 	    offsetof(objset_t, os_evicting_node));
1192 	list_create(&spa->spa_state_dirty_list, sizeof (vdev_t),
1193 	    offsetof(vdev_t, vdev_state_dirty_node));
1194 
1195 	txg_list_create(&spa->spa_vdev_txg_list, spa,
1196 	    offsetof(struct vdev, vdev_txg_node));
1197 
1198 	avl_create(&spa->spa_errlist_scrub,
1199 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1200 	    offsetof(spa_error_entry_t, se_avl));
1201 	avl_create(&spa->spa_errlist_last,
1202 	    spa_error_entry_compare, sizeof (spa_error_entry_t),
1203 	    offsetof(spa_error_entry_t, se_avl));
1204 
1205 	spa_keystore_init(&spa->spa_keystore);
1206 
1207 	/*
1208 	 * The taskq to upgrade datasets in this pool. Currently used by
1209 	 * feature SPA_FEATURE_USEROBJ_ACCOUNTING/SPA_FEATURE_PROJECT_QUOTA.
1210 	 */
1211 	spa->spa_upgrade_taskq = taskq_create("z_upgrade", boot_ncpus,
1212 	    minclsyspri, 1, INT_MAX, TASKQ_DYNAMIC);
1213 }
1214 
1215 /*
1216  * Opposite of spa_activate().
1217  */
1218 static void
1219 spa_deactivate(spa_t *spa)
1220 {
1221 	ASSERT(spa->spa_sync_on == B_FALSE);
1222 	ASSERT(spa->spa_dsl_pool == NULL);
1223 	ASSERT(spa->spa_root_vdev == NULL);
1224 	ASSERT(spa->spa_async_zio_root == NULL);
1225 	ASSERT(spa->spa_state != POOL_STATE_UNINITIALIZED);
1226 
1227 	spa_evicting_os_wait(spa);
1228 
1229 	if (spa->spa_upgrade_taskq) {
1230 		taskq_destroy(spa->spa_upgrade_taskq);
1231 		spa->spa_upgrade_taskq = NULL;
1232 	}
1233 
1234 	txg_list_destroy(&spa->spa_vdev_txg_list);
1235 
1236 	list_destroy(&spa->spa_config_dirty_list);
1237 	list_destroy(&spa->spa_evicting_os_list);
1238 	list_destroy(&spa->spa_state_dirty_list);
1239 
1240 	for (int t = 0; t < ZIO_TYPES; t++) {
1241 		for (int q = 0; q < ZIO_TASKQ_TYPES; q++) {
1242 			spa_taskqs_fini(spa, t, q);
1243 		}
1244 	}
1245 
1246 	for (size_t i = 0; i < TXG_SIZE; i++) {
1247 		ASSERT3P(spa->spa_txg_zio[i], !=, NULL);
1248 		VERIFY0(zio_wait(spa->spa_txg_zio[i]));
1249 		spa->spa_txg_zio[i] = NULL;
1250 	}
1251 
1252 	metaslab_class_destroy(spa->spa_normal_class);
1253 	spa->spa_normal_class = NULL;
1254 
1255 	metaslab_class_destroy(spa->spa_log_class);
1256 	spa->spa_log_class = NULL;
1257 
1258 	metaslab_class_destroy(spa->spa_special_class);
1259 	spa->spa_special_class = NULL;
1260 
1261 	metaslab_class_destroy(spa->spa_dedup_class);
1262 	spa->spa_dedup_class = NULL;
1263 
1264 	/*
1265 	 * If this was part of an import or the open otherwise failed, we may
1266 	 * still have errors left in the queues.  Empty them just in case.
1267 	 */
1268 	spa_errlog_drain(spa);
1269 	avl_destroy(&spa->spa_errlist_scrub);
1270 	avl_destroy(&spa->spa_errlist_last);
1271 
1272 	spa_keystore_fini(&spa->spa_keystore);
1273 
1274 	spa->spa_state = POOL_STATE_UNINITIALIZED;
1275 
1276 	mutex_enter(&spa->spa_proc_lock);
1277 	if (spa->spa_proc_state != SPA_PROC_NONE) {
1278 		ASSERT(spa->spa_proc_state == SPA_PROC_ACTIVE);
1279 		spa->spa_proc_state = SPA_PROC_DEACTIVATE;
1280 		cv_broadcast(&spa->spa_proc_cv);
1281 		while (spa->spa_proc_state == SPA_PROC_DEACTIVATE) {
1282 			ASSERT(spa->spa_proc != &p0);
1283 			cv_wait(&spa->spa_proc_cv, &spa->spa_proc_lock);
1284 		}
1285 		ASSERT(spa->spa_proc_state == SPA_PROC_GONE);
1286 		spa->spa_proc_state = SPA_PROC_NONE;
1287 	}
1288 	ASSERT(spa->spa_proc == &p0);
1289 	mutex_exit(&spa->spa_proc_lock);
1290 
1291 	/*
1292 	 * We want to make sure spa_thread() has actually exited the ZFS
1293 	 * module, so that the module can't be unloaded out from underneath
1294 	 * it.
1295 	 */
1296 	if (spa->spa_did != 0) {
1297 		thread_join(spa->spa_did);
1298 		spa->spa_did = 0;
1299 	}
1300 }
1301 
1302 /*
1303  * Verify a pool configuration, and construct the vdev tree appropriately.  This
1304  * will create all the necessary vdevs in the appropriate layout, with each vdev
1305  * in the CLOSED state.  This will prep the pool before open/creation/import.
1306  * All vdev validation is done by the vdev_alloc() routine.
1307  */
1308 static int
1309 spa_config_parse(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent,
1310     uint_t id, int atype)
1311 {
1312 	nvlist_t **child;
1313 	uint_t children;
1314 	int error;
1315 
1316 	if ((error = vdev_alloc(spa, vdp, nv, parent, id, atype)) != 0)
1317 		return (error);
1318 
1319 	if ((*vdp)->vdev_ops->vdev_op_leaf)
1320 		return (0);
1321 
1322 	error = nvlist_lookup_nvlist_array(nv, ZPOOL_CONFIG_CHILDREN,
1323 	    &child, &children);
1324 
1325 	if (error == ENOENT)
1326 		return (0);
1327 
1328 	if (error) {
1329 		vdev_free(*vdp);
1330 		*vdp = NULL;
1331 		return (SET_ERROR(EINVAL));
1332 	}
1333 
1334 	for (int c = 0; c < children; c++) {
1335 		vdev_t *vd;
1336 		if ((error = spa_config_parse(spa, &vd, child[c], *vdp, c,
1337 		    atype)) != 0) {
1338 			vdev_free(*vdp);
1339 			*vdp = NULL;
1340 			return (error);
1341 		}
1342 	}
1343 
1344 	ASSERT(*vdp != NULL);
1345 
1346 	return (0);
1347 }
1348 
1349 /*
1350  * Opposite of spa_load().
1351  */
1352 static void
1353 spa_unload(spa_t *spa)
1354 {
1355 	int i;
1356 
1357 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
1358 
1359 	spa_load_note(spa, "UNLOADING");
1360 
1361 	/*
1362 	 * Stop async tasks.
1363 	 */
1364 	spa_async_suspend(spa);
1365 
1366 	if (spa->spa_root_vdev) {
1367 		vdev_initialize_stop_all(spa->spa_root_vdev,
1368 		    VDEV_INITIALIZE_ACTIVE);
1369 	}
1370 
1371 	/*
1372 	 * Stop syncing.
1373 	 */
1374 	if (spa->spa_sync_on) {
1375 		txg_sync_stop(spa->spa_dsl_pool);
1376 		spa->spa_sync_on = B_FALSE;
1377 	}
1378 
1379 	/*
1380 	 * Even though vdev_free() also calls vdev_metaslab_fini, we need
1381 	 * to call it earlier, before we wait for async i/o to complete.
1382 	 * This ensures that there is no async metaslab prefetching, by
1383 	 * calling taskq_wait(mg_taskq).
1384 	 */
1385 	if (spa->spa_root_vdev != NULL) {
1386 		spa_config_enter(spa, SCL_ALL, spa, RW_WRITER);
1387 		for (int c = 0; c < spa->spa_root_vdev->vdev_children; c++)
1388 			vdev_metaslab_fini(spa->spa_root_vdev->vdev_child[c]);
1389 		spa_config_exit(spa, SCL_ALL, spa);
1390 	}
1391 
1392 	if (spa->spa_mmp.mmp_thread)
1393 		mmp_thread_stop(spa);
1394 
1395 	/*
1396 	 * Wait for any outstanding async I/O to complete.
1397 	 */
1398 	if (spa->spa_async_zio_root != NULL) {
1399 		for (int i = 0; i < max_ncpus; i++)
1400 			(void) zio_wait(spa->spa_async_zio_root[i]);
1401 		kmem_free(spa->spa_async_zio_root, max_ncpus * sizeof (void *));
1402 		spa->spa_async_zio_root = NULL;
1403 	}
1404 
1405 	if (spa->spa_vdev_removal != NULL) {
1406 		spa_vdev_removal_destroy(spa->spa_vdev_removal);
1407 		spa->spa_vdev_removal = NULL;
1408 	}
1409 
1410 	if (spa->spa_condense_zthr != NULL) {
1411 		zthr_destroy(spa->spa_condense_zthr);
1412 		spa->spa_condense_zthr = NULL;
1413 	}
1414 
1415 	if (spa->spa_checkpoint_discard_zthr != NULL) {
1416 		zthr_destroy(spa->spa_checkpoint_discard_zthr);
1417 		spa->spa_checkpoint_discard_zthr = NULL;
1418 	}
1419 
1420 	spa_condense_fini(spa);
1421 
1422 	bpobj_close(&spa->spa_deferred_bpobj);
1423 
1424 	spa_config_enter(spa, SCL_ALL, spa, RW_WRITER);
1425 
1426 	/*
1427 	 * Close all vdevs.
1428 	 */
1429 	if (spa->spa_root_vdev)
1430 		vdev_free(spa->spa_root_vdev);
1431 	ASSERT(spa->spa_root_vdev == NULL);
1432 
1433 	/*
1434 	 * Close the dsl pool.
1435 	 */
1436 	if (spa->spa_dsl_pool) {
1437 		dsl_pool_close(spa->spa_dsl_pool);
1438 		spa->spa_dsl_pool = NULL;
1439 		spa->spa_meta_objset = NULL;
1440 	}
1441 
1442 	ddt_unload(spa);
1443 
1444 	/*
1445 	 * Drop and purge level 2 cache
1446 	 */
1447 	spa_l2cache_drop(spa);
1448 
1449 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1450 		vdev_free(spa->spa_spares.sav_vdevs[i]);
1451 	if (spa->spa_spares.sav_vdevs) {
1452 		kmem_free(spa->spa_spares.sav_vdevs,
1453 		    spa->spa_spares.sav_count * sizeof (void *));
1454 		spa->spa_spares.sav_vdevs = NULL;
1455 	}
1456 	if (spa->spa_spares.sav_config) {
1457 		nvlist_free(spa->spa_spares.sav_config);
1458 		spa->spa_spares.sav_config = NULL;
1459 	}
1460 	spa->spa_spares.sav_count = 0;
1461 
1462 	for (i = 0; i < spa->spa_l2cache.sav_count; i++) {
1463 		vdev_clear_stats(spa->spa_l2cache.sav_vdevs[i]);
1464 		vdev_free(spa->spa_l2cache.sav_vdevs[i]);
1465 	}
1466 	if (spa->spa_l2cache.sav_vdevs) {
1467 		kmem_free(spa->spa_l2cache.sav_vdevs,
1468 		    spa->spa_l2cache.sav_count * sizeof (void *));
1469 		spa->spa_l2cache.sav_vdevs = NULL;
1470 	}
1471 	if (spa->spa_l2cache.sav_config) {
1472 		nvlist_free(spa->spa_l2cache.sav_config);
1473 		spa->spa_l2cache.sav_config = NULL;
1474 	}
1475 	spa->spa_l2cache.sav_count = 0;
1476 
1477 	spa->spa_async_suspended = 0;
1478 
1479 	spa->spa_indirect_vdevs_loaded = B_FALSE;
1480 
1481 	if (spa->spa_comment != NULL) {
1482 		spa_strfree(spa->spa_comment);
1483 		spa->spa_comment = NULL;
1484 	}
1485 
1486 	spa_config_exit(spa, SCL_ALL, spa);
1487 }
1488 
1489 /*
1490  * Load (or re-load) the current list of vdevs describing the active spares for
1491  * this pool.  When this is called, we have some form of basic information in
1492  * 'spa_spares.sav_config'.  We parse this into vdevs, try to open them, and
1493  * then re-generate a more complete list including status information.
1494  */
1495 void
1496 spa_load_spares(spa_t *spa)
1497 {
1498 	nvlist_t **spares;
1499 	uint_t nspares;
1500 	int i;
1501 	vdev_t *vd, *tvd;
1502 
1503 #ifndef _KERNEL
1504 	/*
1505 	 * zdb opens both the current state of the pool and the
1506 	 * checkpointed state (if present), with a different spa_t.
1507 	 *
1508 	 * As spare vdevs are shared among open pools, we skip loading
1509 	 * them when we load the checkpointed state of the pool.
1510 	 */
1511 	if (!spa_writeable(spa))
1512 		return;
1513 #endif
1514 
1515 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1516 
1517 	/*
1518 	 * First, close and free any existing spare vdevs.
1519 	 */
1520 	for (i = 0; i < spa->spa_spares.sav_count; i++) {
1521 		vd = spa->spa_spares.sav_vdevs[i];
1522 
1523 		/* Undo the call to spa_activate() below */
1524 		if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1525 		    B_FALSE)) != NULL && tvd->vdev_isspare)
1526 			spa_spare_remove(tvd);
1527 		vdev_close(vd);
1528 		vdev_free(vd);
1529 	}
1530 
1531 	if (spa->spa_spares.sav_vdevs)
1532 		kmem_free(spa->spa_spares.sav_vdevs,
1533 		    spa->spa_spares.sav_count * sizeof (void *));
1534 
1535 	if (spa->spa_spares.sav_config == NULL)
1536 		nspares = 0;
1537 	else
1538 		VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
1539 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0);
1540 
1541 	spa->spa_spares.sav_count = (int)nspares;
1542 	spa->spa_spares.sav_vdevs = NULL;
1543 
1544 	if (nspares == 0)
1545 		return;
1546 
1547 	/*
1548 	 * Construct the array of vdevs, opening them to get status in the
1549 	 * process.   For each spare, there is potentially two different vdev_t
1550 	 * structures associated with it: one in the list of spares (used only
1551 	 * for basic validation purposes) and one in the active vdev
1552 	 * configuration (if it's spared in).  During this phase we open and
1553 	 * validate each vdev on the spare list.  If the vdev also exists in the
1554 	 * active configuration, then we also mark this vdev as an active spare.
1555 	 */
1556 	spa->spa_spares.sav_vdevs = kmem_alloc(nspares * sizeof (void *),
1557 	    KM_SLEEP);
1558 	for (i = 0; i < spa->spa_spares.sav_count; i++) {
1559 		VERIFY(spa_config_parse(spa, &vd, spares[i], NULL, 0,
1560 		    VDEV_ALLOC_SPARE) == 0);
1561 		ASSERT(vd != NULL);
1562 
1563 		spa->spa_spares.sav_vdevs[i] = vd;
1564 
1565 		if ((tvd = spa_lookup_by_guid(spa, vd->vdev_guid,
1566 		    B_FALSE)) != NULL) {
1567 			if (!tvd->vdev_isspare)
1568 				spa_spare_add(tvd);
1569 
1570 			/*
1571 			 * We only mark the spare active if we were successfully
1572 			 * able to load the vdev.  Otherwise, importing a pool
1573 			 * with a bad active spare would result in strange
1574 			 * behavior, because multiple pool would think the spare
1575 			 * is actively in use.
1576 			 *
1577 			 * There is a vulnerability here to an equally bizarre
1578 			 * circumstance, where a dead active spare is later
1579 			 * brought back to life (onlined or otherwise).  Given
1580 			 * the rarity of this scenario, and the extra complexity
1581 			 * it adds, we ignore the possibility.
1582 			 */
1583 			if (!vdev_is_dead(tvd))
1584 				spa_spare_activate(tvd);
1585 		}
1586 
1587 		vd->vdev_top = vd;
1588 		vd->vdev_aux = &spa->spa_spares;
1589 
1590 		if (vdev_open(vd) != 0)
1591 			continue;
1592 
1593 		if (vdev_validate_aux(vd) == 0)
1594 			spa_spare_add(vd);
1595 	}
1596 
1597 	/*
1598 	 * Recompute the stashed list of spares, with status information
1599 	 * this time.
1600 	 */
1601 	VERIFY(nvlist_remove(spa->spa_spares.sav_config, ZPOOL_CONFIG_SPARES,
1602 	    DATA_TYPE_NVLIST_ARRAY) == 0);
1603 
1604 	spares = kmem_alloc(spa->spa_spares.sav_count * sizeof (void *),
1605 	    KM_SLEEP);
1606 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1607 		spares[i] = vdev_config_generate(spa,
1608 		    spa->spa_spares.sav_vdevs[i], B_TRUE, VDEV_CONFIG_SPARE);
1609 	VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config,
1610 	    ZPOOL_CONFIG_SPARES, spares, spa->spa_spares.sav_count) == 0);
1611 	for (i = 0; i < spa->spa_spares.sav_count; i++)
1612 		nvlist_free(spares[i]);
1613 	kmem_free(spares, spa->spa_spares.sav_count * sizeof (void *));
1614 }
1615 
1616 /*
1617  * Load (or re-load) the current list of vdevs describing the active l2cache for
1618  * this pool.  When this is called, we have some form of basic information in
1619  * 'spa_l2cache.sav_config'.  We parse this into vdevs, try to open them, and
1620  * then re-generate a more complete list including status information.
1621  * Devices which are already active have their details maintained, and are
1622  * not re-opened.
1623  */
1624 void
1625 spa_load_l2cache(spa_t *spa)
1626 {
1627 	nvlist_t **l2cache;
1628 	uint_t nl2cache;
1629 	int i, j, oldnvdevs;
1630 	uint64_t guid;
1631 	vdev_t *vd, **oldvdevs, **newvdevs;
1632 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
1633 
1634 #ifndef _KERNEL
1635 	/*
1636 	 * zdb opens both the current state of the pool and the
1637 	 * checkpointed state (if present), with a different spa_t.
1638 	 *
1639 	 * As L2 caches are part of the ARC which is shared among open
1640 	 * pools, we skip loading them when we load the checkpointed
1641 	 * state of the pool.
1642 	 */
1643 	if (!spa_writeable(spa))
1644 		return;
1645 #endif
1646 
1647 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
1648 
1649 	if (sav->sav_config != NULL) {
1650 		VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
1651 		    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0);
1652 		newvdevs = kmem_alloc(nl2cache * sizeof (void *), KM_SLEEP);
1653 	} else {
1654 		nl2cache = 0;
1655 		newvdevs = NULL;
1656 	}
1657 
1658 	oldvdevs = sav->sav_vdevs;
1659 	oldnvdevs = sav->sav_count;
1660 	sav->sav_vdevs = NULL;
1661 	sav->sav_count = 0;
1662 
1663 	/*
1664 	 * Process new nvlist of vdevs.
1665 	 */
1666 	for (i = 0; i < nl2cache; i++) {
1667 		VERIFY(nvlist_lookup_uint64(l2cache[i], ZPOOL_CONFIG_GUID,
1668 		    &guid) == 0);
1669 
1670 		newvdevs[i] = NULL;
1671 		for (j = 0; j < oldnvdevs; j++) {
1672 			vd = oldvdevs[j];
1673 			if (vd != NULL && guid == vd->vdev_guid) {
1674 				/*
1675 				 * Retain previous vdev for add/remove ops.
1676 				 */
1677 				newvdevs[i] = vd;
1678 				oldvdevs[j] = NULL;
1679 				break;
1680 			}
1681 		}
1682 
1683 		if (newvdevs[i] == NULL) {
1684 			/*
1685 			 * Create new vdev
1686 			 */
1687 			VERIFY(spa_config_parse(spa, &vd, l2cache[i], NULL, 0,
1688 			    VDEV_ALLOC_L2CACHE) == 0);
1689 			ASSERT(vd != NULL);
1690 			newvdevs[i] = vd;
1691 
1692 			/*
1693 			 * Commit this vdev as an l2cache device,
1694 			 * even if it fails to open.
1695 			 */
1696 			spa_l2cache_add(vd);
1697 
1698 			vd->vdev_top = vd;
1699 			vd->vdev_aux = sav;
1700 
1701 			spa_l2cache_activate(vd);
1702 
1703 			if (vdev_open(vd) != 0)
1704 				continue;
1705 
1706 			(void) vdev_validate_aux(vd);
1707 
1708 			if (!vdev_is_dead(vd))
1709 				l2arc_add_vdev(spa, vd);
1710 		}
1711 	}
1712 
1713 	/*
1714 	 * Purge vdevs that were dropped
1715 	 */
1716 	for (i = 0; i < oldnvdevs; i++) {
1717 		uint64_t pool;
1718 
1719 		vd = oldvdevs[i];
1720 		if (vd != NULL) {
1721 			ASSERT(vd->vdev_isl2cache);
1722 
1723 			if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
1724 			    pool != 0ULL && l2arc_vdev_present(vd))
1725 				l2arc_remove_vdev(vd);
1726 			vdev_clear_stats(vd);
1727 			vdev_free(vd);
1728 		}
1729 	}
1730 
1731 	if (oldvdevs)
1732 		kmem_free(oldvdevs, oldnvdevs * sizeof (void *));
1733 
1734 	if (sav->sav_config == NULL)
1735 		goto out;
1736 
1737 	sav->sav_vdevs = newvdevs;
1738 	sav->sav_count = (int)nl2cache;
1739 
1740 	/*
1741 	 * Recompute the stashed list of l2cache devices, with status
1742 	 * information this time.
1743 	 */
1744 	VERIFY(nvlist_remove(sav->sav_config, ZPOOL_CONFIG_L2CACHE,
1745 	    DATA_TYPE_NVLIST_ARRAY) == 0);
1746 
1747 	l2cache = kmem_alloc(sav->sav_count * sizeof (void *), KM_SLEEP);
1748 	for (i = 0; i < sav->sav_count; i++)
1749 		l2cache[i] = vdev_config_generate(spa,
1750 		    sav->sav_vdevs[i], B_TRUE, VDEV_CONFIG_L2CACHE);
1751 	VERIFY(nvlist_add_nvlist_array(sav->sav_config,
1752 	    ZPOOL_CONFIG_L2CACHE, l2cache, sav->sav_count) == 0);
1753 out:
1754 	for (i = 0; i < sav->sav_count; i++)
1755 		nvlist_free(l2cache[i]);
1756 	if (sav->sav_count)
1757 		kmem_free(l2cache, sav->sav_count * sizeof (void *));
1758 }
1759 
1760 static int
1761 load_nvlist(spa_t *spa, uint64_t obj, nvlist_t **value)
1762 {
1763 	dmu_buf_t *db;
1764 	char *packed = NULL;
1765 	size_t nvsize = 0;
1766 	int error;
1767 	*value = NULL;
1768 
1769 	error = dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db);
1770 	if (error != 0)
1771 		return (error);
1772 
1773 	nvsize = *(uint64_t *)db->db_data;
1774 	dmu_buf_rele(db, FTAG);
1775 
1776 	packed = kmem_alloc(nvsize, KM_SLEEP);
1777 	error = dmu_read(spa->spa_meta_objset, obj, 0, nvsize, packed,
1778 	    DMU_READ_PREFETCH);
1779 	if (error == 0)
1780 		error = nvlist_unpack(packed, nvsize, value, 0);
1781 	kmem_free(packed, nvsize);
1782 
1783 	return (error);
1784 }
1785 
1786 /*
1787  * Concrete top-level vdevs that are not missing and are not logs. At every
1788  * spa_sync we write new uberblocks to at least SPA_SYNC_MIN_VDEVS core tvds.
1789  */
1790 static uint64_t
1791 spa_healthy_core_tvds(spa_t *spa)
1792 {
1793 	vdev_t *rvd = spa->spa_root_vdev;
1794 	uint64_t tvds = 0;
1795 
1796 	for (uint64_t i = 0; i < rvd->vdev_children; i++) {
1797 		vdev_t *vd = rvd->vdev_child[i];
1798 		if (vd->vdev_islog)
1799 			continue;
1800 		if (vdev_is_concrete(vd) && !vdev_is_dead(vd))
1801 			tvds++;
1802 	}
1803 
1804 	return (tvds);
1805 }
1806 
1807 /*
1808  * Checks to see if the given vdev could not be opened, in which case we post a
1809  * sysevent to notify the autoreplace code that the device has been removed.
1810  */
1811 static void
1812 spa_check_removed(vdev_t *vd)
1813 {
1814 	for (uint64_t c = 0; c < vd->vdev_children; c++)
1815 		spa_check_removed(vd->vdev_child[c]);
1816 
1817 	if (vd->vdev_ops->vdev_op_leaf && vdev_is_dead(vd) &&
1818 	    vdev_is_concrete(vd)) {
1819 		zfs_post_autoreplace(vd->vdev_spa, vd);
1820 		spa_event_notify(vd->vdev_spa, vd, NULL, ESC_ZFS_VDEV_CHECK);
1821 	}
1822 }
1823 
1824 static int
1825 spa_check_for_missing_logs(spa_t *spa)
1826 {
1827 	vdev_t *rvd = spa->spa_root_vdev;
1828 
1829 	/*
1830 	 * If we're doing a normal import, then build up any additional
1831 	 * diagnostic information about missing log devices.
1832 	 * We'll pass this up to the user for further processing.
1833 	 */
1834 	if (!(spa->spa_import_flags & ZFS_IMPORT_MISSING_LOG)) {
1835 		nvlist_t **child, *nv;
1836 		uint64_t idx = 0;
1837 
1838 		child = kmem_alloc(rvd->vdev_children * sizeof (nvlist_t **),
1839 		    KM_SLEEP);
1840 		VERIFY(nvlist_alloc(&nv, NV_UNIQUE_NAME, KM_SLEEP) == 0);
1841 
1842 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
1843 			vdev_t *tvd = rvd->vdev_child[c];
1844 
1845 			/*
1846 			 * We consider a device as missing only if it failed
1847 			 * to open (i.e. offline or faulted is not considered
1848 			 * as missing).
1849 			 */
1850 			if (tvd->vdev_islog &&
1851 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
1852 				child[idx++] = vdev_config_generate(spa, tvd,
1853 				    B_FALSE, VDEV_CONFIG_MISSING);
1854 			}
1855 		}
1856 
1857 		if (idx > 0) {
1858 			fnvlist_add_nvlist_array(nv,
1859 			    ZPOOL_CONFIG_CHILDREN, child, idx);
1860 			fnvlist_add_nvlist(spa->spa_load_info,
1861 			    ZPOOL_CONFIG_MISSING_DEVICES, nv);
1862 
1863 			for (uint64_t i = 0; i < idx; i++)
1864 				nvlist_free(child[i]);
1865 		}
1866 		nvlist_free(nv);
1867 		kmem_free(child, rvd->vdev_children * sizeof (char **));
1868 
1869 		if (idx > 0) {
1870 			spa_load_failed(spa, "some log devices are missing");
1871 			vdev_dbgmsg_print_tree(rvd, 2);
1872 			return (SET_ERROR(ENXIO));
1873 		}
1874 	} else {
1875 		for (uint64_t c = 0; c < rvd->vdev_children; c++) {
1876 			vdev_t *tvd = rvd->vdev_child[c];
1877 
1878 			if (tvd->vdev_islog &&
1879 			    tvd->vdev_state == VDEV_STATE_CANT_OPEN) {
1880 				spa_set_log_state(spa, SPA_LOG_CLEAR);
1881 				spa_load_note(spa, "some log devices are "
1882 				    "missing, ZIL is dropped.");
1883 				vdev_dbgmsg_print_tree(rvd, 2);
1884 				break;
1885 			}
1886 		}
1887 	}
1888 
1889 	return (0);
1890 }
1891 
1892 /*
1893  * Check for missing log devices
1894  */
1895 static boolean_t
1896 spa_check_logs(spa_t *spa)
1897 {
1898 	boolean_t rv = B_FALSE;
1899 	dsl_pool_t *dp = spa_get_dsl(spa);
1900 
1901 	switch (spa->spa_log_state) {
1902 	case SPA_LOG_MISSING:
1903 		/* need to recheck in case slog has been restored */
1904 	case SPA_LOG_UNKNOWN:
1905 		rv = (dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
1906 		    zil_check_log_chain, NULL, DS_FIND_CHILDREN) != 0);
1907 		if (rv)
1908 			spa_set_log_state(spa, SPA_LOG_MISSING);
1909 		break;
1910 	}
1911 	return (rv);
1912 }
1913 
1914 static boolean_t
1915 spa_passivate_log(spa_t *spa)
1916 {
1917 	vdev_t *rvd = spa->spa_root_vdev;
1918 	boolean_t slog_found = B_FALSE;
1919 
1920 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
1921 
1922 	if (!spa_has_slogs(spa))
1923 		return (B_FALSE);
1924 
1925 	for (int c = 0; c < rvd->vdev_children; c++) {
1926 		vdev_t *tvd = rvd->vdev_child[c];
1927 		metaslab_group_t *mg = tvd->vdev_mg;
1928 
1929 		if (tvd->vdev_islog) {
1930 			metaslab_group_passivate(mg);
1931 			slog_found = B_TRUE;
1932 		}
1933 	}
1934 
1935 	return (slog_found);
1936 }
1937 
1938 static void
1939 spa_activate_log(spa_t *spa)
1940 {
1941 	vdev_t *rvd = spa->spa_root_vdev;
1942 
1943 	ASSERT(spa_config_held(spa, SCL_ALLOC, RW_WRITER));
1944 
1945 	for (int c = 0; c < rvd->vdev_children; c++) {
1946 		vdev_t *tvd = rvd->vdev_child[c];
1947 		metaslab_group_t *mg = tvd->vdev_mg;
1948 
1949 		if (tvd->vdev_islog)
1950 			metaslab_group_activate(mg);
1951 	}
1952 }
1953 
1954 int
1955 spa_reset_logs(spa_t *spa)
1956 {
1957 	int error;
1958 
1959 	error = dmu_objset_find(spa_name(spa), zil_reset,
1960 	    NULL, DS_FIND_CHILDREN);
1961 	if (error == 0) {
1962 		/*
1963 		 * We successfully offlined the log device, sync out the
1964 		 * current txg so that the "stubby" block can be removed
1965 		 * by zil_sync().
1966 		 */
1967 		txg_wait_synced(spa->spa_dsl_pool, 0);
1968 	}
1969 	return (error);
1970 }
1971 
1972 static void
1973 spa_aux_check_removed(spa_aux_vdev_t *sav)
1974 {
1975 	for (int i = 0; i < sav->sav_count; i++)
1976 		spa_check_removed(sav->sav_vdevs[i]);
1977 }
1978 
1979 void
1980 spa_claim_notify(zio_t *zio)
1981 {
1982 	spa_t *spa = zio->io_spa;
1983 
1984 	if (zio->io_error)
1985 		return;
1986 
1987 	mutex_enter(&spa->spa_props_lock);	/* any mutex will do */
1988 	if (spa->spa_claim_max_txg < zio->io_bp->blk_birth)
1989 		spa->spa_claim_max_txg = zio->io_bp->blk_birth;
1990 	mutex_exit(&spa->spa_props_lock);
1991 }
1992 
1993 typedef struct spa_load_error {
1994 	uint64_t	sle_meta_count;
1995 	uint64_t	sle_data_count;
1996 } spa_load_error_t;
1997 
1998 static void
1999 spa_load_verify_done(zio_t *zio)
2000 {
2001 	blkptr_t *bp = zio->io_bp;
2002 	spa_load_error_t *sle = zio->io_private;
2003 	dmu_object_type_t type = BP_GET_TYPE(bp);
2004 	int error = zio->io_error;
2005 	spa_t *spa = zio->io_spa;
2006 
2007 	abd_free(zio->io_abd);
2008 	if (error) {
2009 		if ((BP_GET_LEVEL(bp) != 0 || DMU_OT_IS_METADATA(type)) &&
2010 		    type != DMU_OT_INTENT_LOG)
2011 			atomic_inc_64(&sle->sle_meta_count);
2012 		else
2013 			atomic_inc_64(&sle->sle_data_count);
2014 	}
2015 
2016 	mutex_enter(&spa->spa_scrub_lock);
2017 	spa->spa_load_verify_ios--;
2018 	cv_broadcast(&spa->spa_scrub_io_cv);
2019 	mutex_exit(&spa->spa_scrub_lock);
2020 }
2021 
2022 /*
2023  * Maximum number of concurrent scrub i/os to create while verifying
2024  * a pool while importing it.
2025  */
2026 int spa_load_verify_maxinflight = 10000;
2027 boolean_t spa_load_verify_metadata = B_TRUE;
2028 boolean_t spa_load_verify_data = B_TRUE;
2029 
2030 /*ARGSUSED*/
2031 static int
2032 spa_load_verify_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
2033     const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
2034 {
2035 	if (bp == NULL || BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2036 		return (0);
2037 	/*
2038 	 * Note: normally this routine will not be called if
2039 	 * spa_load_verify_metadata is not set.  However, it may be useful
2040 	 * to manually set the flag after the traversal has begun.
2041 	 */
2042 	if (!spa_load_verify_metadata)
2043 		return (0);
2044 	if (!BP_IS_METADATA(bp) && !spa_load_verify_data)
2045 		return (0);
2046 
2047 	zio_t *rio = arg;
2048 	size_t size = BP_GET_PSIZE(bp);
2049 
2050 	mutex_enter(&spa->spa_scrub_lock);
2051 	while (spa->spa_load_verify_ios >= spa_load_verify_maxinflight)
2052 		cv_wait(&spa->spa_scrub_io_cv, &spa->spa_scrub_lock);
2053 	spa->spa_load_verify_ios++;
2054 	mutex_exit(&spa->spa_scrub_lock);
2055 
2056 	zio_nowait(zio_read(rio, spa, bp, abd_alloc_for_io(size, B_FALSE), size,
2057 	    spa_load_verify_done, rio->io_private, ZIO_PRIORITY_SCRUB,
2058 	    ZIO_FLAG_SPECULATIVE | ZIO_FLAG_CANFAIL |
2059 	    ZIO_FLAG_SCRUB | ZIO_FLAG_RAW, zb));
2060 	return (0);
2061 }
2062 
2063 /* ARGSUSED */
2064 int
2065 verify_dataset_name_len(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
2066 {
2067 	if (dsl_dataset_namelen(ds) >= ZFS_MAX_DATASET_NAME_LEN)
2068 		return (SET_ERROR(ENAMETOOLONG));
2069 
2070 	return (0);
2071 }
2072 
2073 static int
2074 spa_load_verify(spa_t *spa)
2075 {
2076 	zio_t *rio;
2077 	spa_load_error_t sle = { 0 };
2078 	zpool_load_policy_t policy;
2079 	boolean_t verify_ok = B_FALSE;
2080 	int error = 0;
2081 
2082 	zpool_get_load_policy(spa->spa_config, &policy);
2083 
2084 	if (policy.zlp_rewind & ZPOOL_NEVER_REWIND)
2085 		return (0);
2086 
2087 	dsl_pool_config_enter(spa->spa_dsl_pool, FTAG);
2088 	error = dmu_objset_find_dp(spa->spa_dsl_pool,
2089 	    spa->spa_dsl_pool->dp_root_dir_obj, verify_dataset_name_len, NULL,
2090 	    DS_FIND_CHILDREN);
2091 	dsl_pool_config_exit(spa->spa_dsl_pool, FTAG);
2092 	if (error != 0)
2093 		return (error);
2094 
2095 	rio = zio_root(spa, NULL, &sle,
2096 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE);
2097 
2098 	if (spa_load_verify_metadata) {
2099 		if (spa->spa_extreme_rewind) {
2100 			spa_load_note(spa, "performing a complete scan of the "
2101 			    "pool since extreme rewind is on. This may take "
2102 			    "a very long time.\n  (spa_load_verify_data=%u, "
2103 			    "spa_load_verify_metadata=%u)",
2104 			    spa_load_verify_data, spa_load_verify_metadata);
2105 		}
2106 		error = traverse_pool(spa, spa->spa_verify_min_txg,
2107 		    TRAVERSE_PRE | TRAVERSE_PREFETCH_METADATA |
2108 		    TRAVERSE_NO_DECRYPT, spa_load_verify_cb, rio);
2109 	}
2110 
2111 	(void) zio_wait(rio);
2112 
2113 	spa->spa_load_meta_errors = sle.sle_meta_count;
2114 	spa->spa_load_data_errors = sle.sle_data_count;
2115 
2116 	if (sle.sle_meta_count != 0 || sle.sle_data_count != 0) {
2117 		spa_load_note(spa, "spa_load_verify found %llu metadata errors "
2118 		    "and %llu data errors", (u_longlong_t)sle.sle_meta_count,
2119 		    (u_longlong_t)sle.sle_data_count);
2120 	}
2121 
2122 	if (spa_load_verify_dryrun ||
2123 	    (!error && sle.sle_meta_count <= policy.zlp_maxmeta &&
2124 	    sle.sle_data_count <= policy.zlp_maxdata)) {
2125 		int64_t loss = 0;
2126 
2127 		verify_ok = B_TRUE;
2128 		spa->spa_load_txg = spa->spa_uberblock.ub_txg;
2129 		spa->spa_load_txg_ts = spa->spa_uberblock.ub_timestamp;
2130 
2131 		loss = spa->spa_last_ubsync_txg_ts - spa->spa_load_txg_ts;
2132 		VERIFY(nvlist_add_uint64(spa->spa_load_info,
2133 		    ZPOOL_CONFIG_LOAD_TIME, spa->spa_load_txg_ts) == 0);
2134 		VERIFY(nvlist_add_int64(spa->spa_load_info,
2135 		    ZPOOL_CONFIG_REWIND_TIME, loss) == 0);
2136 		VERIFY(nvlist_add_uint64(spa->spa_load_info,
2137 		    ZPOOL_CONFIG_LOAD_DATA_ERRORS, sle.sle_data_count) == 0);
2138 	} else {
2139 		spa->spa_load_max_txg = spa->spa_uberblock.ub_txg;
2140 	}
2141 
2142 	if (spa_load_verify_dryrun)
2143 		return (0);
2144 
2145 	if (error) {
2146 		if (error != ENXIO && error != EIO)
2147 			error = SET_ERROR(EIO);
2148 		return (error);
2149 	}
2150 
2151 	return (verify_ok ? 0 : EIO);
2152 }
2153 
2154 /*
2155  * Find a value in the pool props object.
2156  */
2157 static void
2158 spa_prop_find(spa_t *spa, zpool_prop_t prop, uint64_t *val)
2159 {
2160 	(void) zap_lookup(spa->spa_meta_objset, spa->spa_pool_props_object,
2161 	    zpool_prop_to_name(prop), sizeof (uint64_t), 1, val);
2162 }
2163 
2164 /*
2165  * Find a value in the pool directory object.
2166  */
2167 static int
2168 spa_dir_prop(spa_t *spa, const char *name, uint64_t *val, boolean_t log_enoent)
2169 {
2170 	int error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
2171 	    name, sizeof (uint64_t), 1, val);
2172 
2173 	if (error != 0 && (error != ENOENT || log_enoent)) {
2174 		spa_load_failed(spa, "couldn't get '%s' value in MOS directory "
2175 		    "[error=%d]", name, error);
2176 	}
2177 
2178 	return (error);
2179 }
2180 
2181 static int
2182 spa_vdev_err(vdev_t *vdev, vdev_aux_t aux, int err)
2183 {
2184 	vdev_set_state(vdev, B_TRUE, VDEV_STATE_CANT_OPEN, aux);
2185 	return (SET_ERROR(err));
2186 }
2187 
2188 static void
2189 spa_spawn_aux_threads(spa_t *spa)
2190 {
2191 	ASSERT(spa_writeable(spa));
2192 
2193 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
2194 
2195 	spa_start_indirect_condensing_thread(spa);
2196 
2197 	ASSERT3P(spa->spa_checkpoint_discard_zthr, ==, NULL);
2198 	spa->spa_checkpoint_discard_zthr =
2199 	    zthr_create(spa_checkpoint_discard_thread_check,
2200 	    spa_checkpoint_discard_thread, spa);
2201 }
2202 
2203 /*
2204  * Fix up config after a partly-completed split.  This is done with the
2205  * ZPOOL_CONFIG_SPLIT nvlist.  Both the splitting pool and the split-off
2206  * pool have that entry in their config, but only the splitting one contains
2207  * a list of all the guids of the vdevs that are being split off.
2208  *
2209  * This function determines what to do with that list: either rejoin
2210  * all the disks to the pool, or complete the splitting process.  To attempt
2211  * the rejoin, each disk that is offlined is marked online again, and
2212  * we do a reopen() call.  If the vdev label for every disk that was
2213  * marked online indicates it was successfully split off (VDEV_AUX_SPLIT_POOL)
2214  * then we call vdev_split() on each disk, and complete the split.
2215  *
2216  * Otherwise we leave the config alone, with all the vdevs in place in
2217  * the original pool.
2218  */
2219 static void
2220 spa_try_repair(spa_t *spa, nvlist_t *config)
2221 {
2222 	uint_t extracted;
2223 	uint64_t *glist;
2224 	uint_t i, gcount;
2225 	nvlist_t *nvl;
2226 	vdev_t **vd;
2227 	boolean_t attempt_reopen;
2228 
2229 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) != 0)
2230 		return;
2231 
2232 	/* check that the config is complete */
2233 	if (nvlist_lookup_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST,
2234 	    &glist, &gcount) != 0)
2235 		return;
2236 
2237 	vd = kmem_zalloc(gcount * sizeof (vdev_t *), KM_SLEEP);
2238 
2239 	/* attempt to online all the vdevs & validate */
2240 	attempt_reopen = B_TRUE;
2241 	for (i = 0; i < gcount; i++) {
2242 		if (glist[i] == 0)	/* vdev is hole */
2243 			continue;
2244 
2245 		vd[i] = spa_lookup_by_guid(spa, glist[i], B_FALSE);
2246 		if (vd[i] == NULL) {
2247 			/*
2248 			 * Don't bother attempting to reopen the disks;
2249 			 * just do the split.
2250 			 */
2251 			attempt_reopen = B_FALSE;
2252 		} else {
2253 			/* attempt to re-online it */
2254 			vd[i]->vdev_offline = B_FALSE;
2255 		}
2256 	}
2257 
2258 	if (attempt_reopen) {
2259 		vdev_reopen(spa->spa_root_vdev);
2260 
2261 		/* check each device to see what state it's in */
2262 		for (extracted = 0, i = 0; i < gcount; i++) {
2263 			if (vd[i] != NULL &&
2264 			    vd[i]->vdev_stat.vs_aux != VDEV_AUX_SPLIT_POOL)
2265 				break;
2266 			++extracted;
2267 		}
2268 	}
2269 
2270 	/*
2271 	 * If every disk has been moved to the new pool, or if we never
2272 	 * even attempted to look at them, then we split them off for
2273 	 * good.
2274 	 */
2275 	if (!attempt_reopen || gcount == extracted) {
2276 		for (i = 0; i < gcount; i++)
2277 			if (vd[i] != NULL)
2278 				vdev_split(vd[i]);
2279 		vdev_reopen(spa->spa_root_vdev);
2280 	}
2281 
2282 	kmem_free(vd, gcount * sizeof (vdev_t *));
2283 }
2284 
2285 static int
2286 spa_load(spa_t *spa, spa_load_state_t state, spa_import_type_t type)
2287 {
2288 	char *ereport = FM_EREPORT_ZFS_POOL;
2289 	int error;
2290 
2291 	spa->spa_load_state = state;
2292 
2293 	gethrestime(&spa->spa_loaded_ts);
2294 	error = spa_load_impl(spa, type, &ereport);
2295 
2296 	/*
2297 	 * Don't count references from objsets that are already closed
2298 	 * and are making their way through the eviction process.
2299 	 */
2300 	spa_evicting_os_wait(spa);
2301 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
2302 	if (error) {
2303 		if (error != EEXIST) {
2304 			spa->spa_loaded_ts.tv_sec = 0;
2305 			spa->spa_loaded_ts.tv_nsec = 0;
2306 		}
2307 		if (error != EBADF) {
2308 			zfs_ereport_post(ereport, spa, NULL, NULL, NULL, 0, 0);
2309 		}
2310 	}
2311 	spa->spa_load_state = error ? SPA_LOAD_ERROR : SPA_LOAD_NONE;
2312 	spa->spa_ena = 0;
2313 
2314 	return (error);
2315 }
2316 
2317 /*
2318  * Count the number of per-vdev ZAPs associated with all of the vdevs in the
2319  * vdev tree rooted in the given vd, and ensure that each ZAP is present in the
2320  * spa's per-vdev ZAP list.
2321  */
2322 static uint64_t
2323 vdev_count_verify_zaps(vdev_t *vd)
2324 {
2325 	spa_t *spa = vd->vdev_spa;
2326 	uint64_t total = 0;
2327 	if (vd->vdev_top_zap != 0) {
2328 		total++;
2329 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
2330 		    spa->spa_all_vdev_zaps, vd->vdev_top_zap));
2331 	}
2332 	if (vd->vdev_leaf_zap != 0) {
2333 		total++;
2334 		ASSERT0(zap_lookup_int(spa->spa_meta_objset,
2335 		    spa->spa_all_vdev_zaps, vd->vdev_leaf_zap));
2336 	}
2337 
2338 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2339 		total += vdev_count_verify_zaps(vd->vdev_child[i]);
2340 	}
2341 
2342 	return (total);
2343 }
2344 
2345 /*
2346  * Determine whether the activity check is required.
2347  */
2348 static boolean_t
2349 spa_activity_check_required(spa_t *spa, uberblock_t *ub, nvlist_t *label,
2350     nvlist_t *config)
2351 {
2352 	uint64_t state = 0;
2353 	uint64_t hostid = 0;
2354 	uint64_t tryconfig_txg = 0;
2355 	uint64_t tryconfig_timestamp = 0;
2356 	nvlist_t *nvinfo;
2357 
2358 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
2359 		nvinfo = fnvlist_lookup_nvlist(config, ZPOOL_CONFIG_LOAD_INFO);
2360 		(void) nvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG,
2361 		    &tryconfig_txg);
2362 		(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
2363 		    &tryconfig_timestamp);
2364 	}
2365 
2366 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_STATE, &state);
2367 
2368 	/*
2369 	 * Disable the MMP activity check - This is used by zdb which
2370 	 * is intended to be used on potentially active pools.
2371 	 */
2372 	if (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP)
2373 		return (B_FALSE);
2374 
2375 	/*
2376 	 * Skip the activity check when the MMP feature is disabled.
2377 	 */
2378 	if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay == 0)
2379 		return (B_FALSE);
2380 	/*
2381 	 * If the tryconfig_* values are nonzero, they are the results of an
2382 	 * earlier tryimport.  If they match the uberblock we just found, then
2383 	 * the pool has not changed and we return false so we do not test a
2384 	 * second time.
2385 	 */
2386 	if (tryconfig_txg && tryconfig_txg == ub->ub_txg &&
2387 	    tryconfig_timestamp && tryconfig_timestamp == ub->ub_timestamp)
2388 		return (B_FALSE);
2389 
2390 	/*
2391 	 * Allow the activity check to be skipped when importing the pool
2392 	 * on the same host which last imported it.  Since the hostid from
2393 	 * configuration may be stale use the one read from the label.
2394 	 */
2395 	if (nvlist_exists(label, ZPOOL_CONFIG_HOSTID))
2396 		hostid = fnvlist_lookup_uint64(label, ZPOOL_CONFIG_HOSTID);
2397 
2398 	if (hostid == spa_get_hostid())
2399 		return (B_FALSE);
2400 
2401 	/*
2402 	 * Skip the activity test when the pool was cleanly exported.
2403 	 */
2404 	if (state != POOL_STATE_ACTIVE)
2405 		return (B_FALSE);
2406 
2407 	return (B_TRUE);
2408 }
2409 
2410 /*
2411  * Perform the import activity check.  If the user canceled the import or
2412  * we detected activity then fail.
2413  */
2414 static int
2415 spa_activity_check(spa_t *spa, uberblock_t *ub, nvlist_t *config)
2416 {
2417 	uint64_t import_intervals = MAX(zfs_multihost_import_intervals, 1);
2418 	uint64_t txg = ub->ub_txg;
2419 	uint64_t timestamp = ub->ub_timestamp;
2420 	uint64_t import_delay = NANOSEC;
2421 	hrtime_t import_expire;
2422 	nvlist_t *mmp_label = NULL;
2423 	vdev_t *rvd = spa->spa_root_vdev;
2424 	kcondvar_t cv;
2425 	kmutex_t mtx;
2426 	int error = 0;
2427 
2428 	cv_init(&cv, NULL, CV_DEFAULT, NULL);
2429 	mutex_init(&mtx, NULL, MUTEX_DEFAULT, NULL);
2430 	mutex_enter(&mtx);
2431 
2432 	/*
2433 	 * If ZPOOL_CONFIG_MMP_TXG is present an activity check was performed
2434 	 * during the earlier tryimport.  If the txg recorded there is 0 then
2435 	 * the pool is known to be active on another host.
2436 	 *
2437 	 * Otherwise, the pool might be in use on another node.  Check for
2438 	 * changes in the uberblocks on disk if necessary.
2439 	 */
2440 	if (nvlist_exists(config, ZPOOL_CONFIG_LOAD_INFO)) {
2441 		nvlist_t *nvinfo = fnvlist_lookup_nvlist(config,
2442 		    ZPOOL_CONFIG_LOAD_INFO);
2443 
2444 		if (nvlist_exists(nvinfo, ZPOOL_CONFIG_MMP_TXG) &&
2445 		    fnvlist_lookup_uint64(nvinfo, ZPOOL_CONFIG_MMP_TXG) == 0) {
2446 			vdev_uberblock_load(rvd, ub, &mmp_label);
2447 			error = SET_ERROR(EREMOTEIO);
2448 			goto out;
2449 		}
2450 	}
2451 
2452 	/*
2453 	 * Preferentially use the zfs_multihost_interval from the node which
2454 	 * last imported the pool.  This value is stored in an MMP uberblock as.
2455 	 *
2456 	 * ub_mmp_delay * vdev_count_leaves() == zfs_multihost_interval
2457 	 */
2458 	if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay)
2459 		import_delay = MAX(import_delay, import_intervals *
2460 		    ub->ub_mmp_delay * MAX(vdev_count_leaves(spa), 1));
2461 
2462 	/* Apply a floor using the local default values. */
2463 	import_delay = MAX(import_delay, import_intervals *
2464 	    MSEC2NSEC(MAX(zfs_multihost_interval, MMP_MIN_INTERVAL)));
2465 
2466 	zfs_dbgmsg("import_delay=%llu ub_mmp_delay=%llu import_intervals=%u "
2467 	    "leaves=%u", import_delay, ub->ub_mmp_delay, import_intervals,
2468 	    vdev_count_leaves(spa));
2469 
2470 	/* Add a small random factor in case of simultaneous imports (0-25%) */
2471 	import_expire = gethrtime() + import_delay +
2472 	    (import_delay * spa_get_random(250) / 1000);
2473 
2474 	while (gethrtime() < import_expire) {
2475 		vdev_uberblock_load(rvd, ub, &mmp_label);
2476 
2477 		if (txg != ub->ub_txg || timestamp != ub->ub_timestamp) {
2478 			error = SET_ERROR(EREMOTEIO);
2479 			break;
2480 		}
2481 
2482 		if (mmp_label) {
2483 			nvlist_free(mmp_label);
2484 			mmp_label = NULL;
2485 		}
2486 
2487 		error = cv_timedwait_sig(&cv, &mtx, ddi_get_lbolt() + hz);
2488 		if (error != -1) {
2489 			error = SET_ERROR(EINTR);
2490 			break;
2491 		}
2492 		error = 0;
2493 	}
2494 
2495 out:
2496 	mutex_exit(&mtx);
2497 	mutex_destroy(&mtx);
2498 	cv_destroy(&cv);
2499 
2500 	/*
2501 	 * If the pool is determined to be active store the status in the
2502 	 * spa->spa_load_info nvlist.  If the remote hostname or hostid are
2503 	 * available from configuration read from disk store them as well.
2504 	 * This allows 'zpool import' to generate a more useful message.
2505 	 *
2506 	 * ZPOOL_CONFIG_MMP_STATE    - observed pool status (mandatory)
2507 	 * ZPOOL_CONFIG_MMP_HOSTNAME - hostname from the active pool
2508 	 * ZPOOL_CONFIG_MMP_HOSTID   - hostid from the active pool
2509 	 */
2510 	if (error == EREMOTEIO) {
2511 		char *hostname = "<unknown>";
2512 		uint64_t hostid = 0;
2513 
2514 		if (mmp_label) {
2515 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTNAME)) {
2516 				hostname = fnvlist_lookup_string(mmp_label,
2517 				    ZPOOL_CONFIG_HOSTNAME);
2518 				fnvlist_add_string(spa->spa_load_info,
2519 				    ZPOOL_CONFIG_MMP_HOSTNAME, hostname);
2520 			}
2521 
2522 			if (nvlist_exists(mmp_label, ZPOOL_CONFIG_HOSTID)) {
2523 				hostid = fnvlist_lookup_uint64(mmp_label,
2524 				    ZPOOL_CONFIG_HOSTID);
2525 				fnvlist_add_uint64(spa->spa_load_info,
2526 				    ZPOOL_CONFIG_MMP_HOSTID, hostid);
2527 			}
2528 		}
2529 
2530 		fnvlist_add_uint64(spa->spa_load_info,
2531 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_ACTIVE);
2532 		fnvlist_add_uint64(spa->spa_load_info,
2533 		    ZPOOL_CONFIG_MMP_TXG, 0);
2534 
2535 		error = spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO);
2536 	}
2537 
2538 	if (mmp_label)
2539 		nvlist_free(mmp_label);
2540 
2541 	return (error);
2542 }
2543 
2544 static int
2545 spa_verify_host(spa_t *spa, nvlist_t *mos_config)
2546 {
2547 	uint64_t hostid;
2548 	char *hostname;
2549 	uint64_t myhostid = 0;
2550 
2551 	if (!spa_is_root(spa) && nvlist_lookup_uint64(mos_config,
2552 	    ZPOOL_CONFIG_HOSTID, &hostid) == 0) {
2553 		hostname = fnvlist_lookup_string(mos_config,
2554 		    ZPOOL_CONFIG_HOSTNAME);
2555 
2556 		myhostid = zone_get_hostid(NULL);
2557 
2558 		if (hostid != 0 && myhostid != 0 && hostid != myhostid) {
2559 			cmn_err(CE_WARN, "pool '%s' could not be "
2560 			    "loaded as it was last accessed by "
2561 			    "another system (host: %s hostid: 0x%llx). "
2562 			    "See: http://illumos.org/msg/ZFS-8000-EY",
2563 			    spa_name(spa), hostname, (u_longlong_t)hostid);
2564 			spa_load_failed(spa, "hostid verification failed: pool "
2565 			    "last accessed by host: %s (hostid: 0x%llx)",
2566 			    hostname, (u_longlong_t)hostid);
2567 			return (SET_ERROR(EBADF));
2568 		}
2569 	}
2570 
2571 	return (0);
2572 }
2573 
2574 static int
2575 spa_ld_parse_config(spa_t *spa, spa_import_type_t type)
2576 {
2577 	int error = 0;
2578 	nvlist_t *nvtree, *nvl, *config = spa->spa_config;
2579 	int parse;
2580 	vdev_t *rvd;
2581 	uint64_t pool_guid;
2582 	char *comment;
2583 
2584 	/*
2585 	 * Versioning wasn't explicitly added to the label until later, so if
2586 	 * it's not present treat it as the initial version.
2587 	 */
2588 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION,
2589 	    &spa->spa_ubsync.ub_version) != 0)
2590 		spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL;
2591 
2592 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID, &pool_guid)) {
2593 		spa_load_failed(spa, "invalid config provided: '%s' missing",
2594 		    ZPOOL_CONFIG_POOL_GUID);
2595 		return (SET_ERROR(EINVAL));
2596 	}
2597 
2598 	/*
2599 	 * If we are doing an import, ensure that the pool is not already
2600 	 * imported by checking if its pool guid already exists in the
2601 	 * spa namespace.
2602 	 *
2603 	 * The only case that we allow an already imported pool to be
2604 	 * imported again, is when the pool is checkpointed and we want to
2605 	 * look at its checkpointed state from userland tools like zdb.
2606 	 */
2607 #ifdef _KERNEL
2608 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
2609 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
2610 	    spa_guid_exists(pool_guid, 0)) {
2611 #else
2612 	if ((spa->spa_load_state == SPA_LOAD_IMPORT ||
2613 	    spa->spa_load_state == SPA_LOAD_TRYIMPORT) &&
2614 	    spa_guid_exists(pool_guid, 0) &&
2615 	    !spa_importing_readonly_checkpoint(spa)) {
2616 #endif
2617 		spa_load_failed(spa, "a pool with guid %llu is already open",
2618 		    (u_longlong_t)pool_guid);
2619 		return (SET_ERROR(EEXIST));
2620 	}
2621 
2622 	spa->spa_config_guid = pool_guid;
2623 
2624 	nvlist_free(spa->spa_load_info);
2625 	spa->spa_load_info = fnvlist_alloc();
2626 
2627 	ASSERT(spa->spa_comment == NULL);
2628 	if (nvlist_lookup_string(config, ZPOOL_CONFIG_COMMENT, &comment) == 0)
2629 		spa->spa_comment = spa_strdup(comment);
2630 
2631 	(void) nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG,
2632 	    &spa->spa_config_txg);
2633 
2634 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_SPLIT, &nvl) == 0)
2635 		spa->spa_config_splitting = fnvlist_dup(nvl);
2636 
2637 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvtree)) {
2638 		spa_load_failed(spa, "invalid config provided: '%s' missing",
2639 		    ZPOOL_CONFIG_VDEV_TREE);
2640 		return (SET_ERROR(EINVAL));
2641 	}
2642 
2643 	/*
2644 	 * Create "The Godfather" zio to hold all async IOs
2645 	 */
2646 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
2647 	    KM_SLEEP);
2648 	for (int i = 0; i < max_ncpus; i++) {
2649 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
2650 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
2651 		    ZIO_FLAG_GODFATHER);
2652 	}
2653 
2654 	/*
2655 	 * Parse the configuration into a vdev tree.  We explicitly set the
2656 	 * value that will be returned by spa_version() since parsing the
2657 	 * configuration requires knowing the version number.
2658 	 */
2659 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
2660 	parse = (type == SPA_IMPORT_EXISTING ?
2661 	    VDEV_ALLOC_LOAD : VDEV_ALLOC_SPLIT);
2662 	error = spa_config_parse(spa, &rvd, nvtree, NULL, 0, parse);
2663 	spa_config_exit(spa, SCL_ALL, FTAG);
2664 
2665 	if (error != 0) {
2666 		spa_load_failed(spa, "unable to parse config [error=%d]",
2667 		    error);
2668 		return (error);
2669 	}
2670 
2671 	ASSERT(spa->spa_root_vdev == rvd);
2672 	ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT);
2673 	ASSERT3U(spa->spa_max_ashift, <=, SPA_MAXBLOCKSHIFT);
2674 
2675 	if (type != SPA_IMPORT_ASSEMBLE) {
2676 		ASSERT(spa_guid(spa) == pool_guid);
2677 	}
2678 
2679 	return (0);
2680 }
2681 
2682 /*
2683  * Recursively open all vdevs in the vdev tree. This function is called twice:
2684  * first with the untrusted config, then with the trusted config.
2685  */
2686 static int
2687 spa_ld_open_vdevs(spa_t *spa)
2688 {
2689 	int error = 0;
2690 
2691 	/*
2692 	 * spa_missing_tvds_allowed defines how many top-level vdevs can be
2693 	 * missing/unopenable for the root vdev to be still considered openable.
2694 	 */
2695 	if (spa->spa_trust_config) {
2696 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds;
2697 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_CACHEFILE) {
2698 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_cachefile;
2699 	} else if (spa->spa_config_source == SPA_CONFIG_SRC_SCAN) {
2700 		spa->spa_missing_tvds_allowed = zfs_max_missing_tvds_scan;
2701 	} else {
2702 		spa->spa_missing_tvds_allowed = 0;
2703 	}
2704 
2705 	spa->spa_missing_tvds_allowed =
2706 	    MAX(zfs_max_missing_tvds, spa->spa_missing_tvds_allowed);
2707 
2708 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
2709 	error = vdev_open(spa->spa_root_vdev);
2710 	spa_config_exit(spa, SCL_ALL, FTAG);
2711 
2712 	if (spa->spa_missing_tvds != 0) {
2713 		spa_load_note(spa, "vdev tree has %lld missing top-level "
2714 		    "vdevs.", (u_longlong_t)spa->spa_missing_tvds);
2715 		if (spa->spa_trust_config && (spa->spa_mode & FWRITE)) {
2716 			/*
2717 			 * Although theoretically we could allow users to open
2718 			 * incomplete pools in RW mode, we'd need to add a lot
2719 			 * of extra logic (e.g. adjust pool space to account
2720 			 * for missing vdevs).
2721 			 * This limitation also prevents users from accidentally
2722 			 * opening the pool in RW mode during data recovery and
2723 			 * damaging it further.
2724 			 */
2725 			spa_load_note(spa, "pools with missing top-level "
2726 			    "vdevs can only be opened in read-only mode.");
2727 			error = SET_ERROR(ENXIO);
2728 		} else {
2729 			spa_load_note(spa, "current settings allow for maximum "
2730 			    "%lld missing top-level vdevs at this stage.",
2731 			    (u_longlong_t)spa->spa_missing_tvds_allowed);
2732 		}
2733 	}
2734 	if (error != 0) {
2735 		spa_load_failed(spa, "unable to open vdev tree [error=%d]",
2736 		    error);
2737 	}
2738 	if (spa->spa_missing_tvds != 0 || error != 0)
2739 		vdev_dbgmsg_print_tree(spa->spa_root_vdev, 2);
2740 
2741 	return (error);
2742 }
2743 
2744 /*
2745  * We need to validate the vdev labels against the configuration that
2746  * we have in hand. This function is called twice: first with an untrusted
2747  * config, then with a trusted config. The validation is more strict when the
2748  * config is trusted.
2749  */
2750 static int
2751 spa_ld_validate_vdevs(spa_t *spa)
2752 {
2753 	int error = 0;
2754 	vdev_t *rvd = spa->spa_root_vdev;
2755 
2756 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
2757 	error = vdev_validate(rvd);
2758 	spa_config_exit(spa, SCL_ALL, FTAG);
2759 
2760 	if (error != 0) {
2761 		spa_load_failed(spa, "vdev_validate failed [error=%d]", error);
2762 		return (error);
2763 	}
2764 
2765 	if (rvd->vdev_state <= VDEV_STATE_CANT_OPEN) {
2766 		spa_load_failed(spa, "cannot open vdev tree after invalidating "
2767 		    "some vdevs");
2768 		vdev_dbgmsg_print_tree(rvd, 2);
2769 		return (SET_ERROR(ENXIO));
2770 	}
2771 
2772 	return (0);
2773 }
2774 
2775 static void
2776 spa_ld_select_uberblock_done(spa_t *spa, uberblock_t *ub)
2777 {
2778 	spa->spa_state = POOL_STATE_ACTIVE;
2779 	spa->spa_ubsync = spa->spa_uberblock;
2780 	spa->spa_verify_min_txg = spa->spa_extreme_rewind ?
2781 	    TXG_INITIAL - 1 : spa_last_synced_txg(spa) - TXG_DEFER_SIZE - 1;
2782 	spa->spa_first_txg = spa->spa_last_ubsync_txg ?
2783 	    spa->spa_last_ubsync_txg : spa_last_synced_txg(spa) + 1;
2784 	spa->spa_claim_max_txg = spa->spa_first_txg;
2785 	spa->spa_prev_software_version = ub->ub_software_version;
2786 }
2787 
2788 static int
2789 spa_ld_select_uberblock(spa_t *spa, spa_import_type_t type)
2790 {
2791 	vdev_t *rvd = spa->spa_root_vdev;
2792 	nvlist_t *label;
2793 	uberblock_t *ub = &spa->spa_uberblock;
2794 	boolean_t activity_check = B_FALSE;
2795 
2796 	/*
2797 	 * If we are opening the checkpointed state of the pool by
2798 	 * rewinding to it, at this point we will have written the
2799 	 * checkpointed uberblock to the vdev labels, so searching
2800 	 * the labels will find the right uberblock.  However, if
2801 	 * we are opening the checkpointed state read-only, we have
2802 	 * not modified the labels. Therefore, we must ignore the
2803 	 * labels and continue using the spa_uberblock that was set
2804 	 * by spa_ld_checkpoint_rewind.
2805 	 *
2806 	 * Note that it would be fine to ignore the labels when
2807 	 * rewinding (opening writeable) as well. However, if we
2808 	 * crash just after writing the labels, we will end up
2809 	 * searching the labels. Doing so in the common case means
2810 	 * that this code path gets exercised normally, rather than
2811 	 * just in the edge case.
2812 	 */
2813 	if (ub->ub_checkpoint_txg != 0 &&
2814 	    spa_importing_readonly_checkpoint(spa)) {
2815 		spa_ld_select_uberblock_done(spa, ub);
2816 		return (0);
2817 	}
2818 
2819 	/*
2820 	 * Find the best uberblock.
2821 	 */
2822 	vdev_uberblock_load(rvd, ub, &label);
2823 
2824 	/*
2825 	 * If we weren't able to find a single valid uberblock, return failure.
2826 	 */
2827 	if (ub->ub_txg == 0) {
2828 		nvlist_free(label);
2829 		spa_load_failed(spa, "no valid uberblock found");
2830 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, ENXIO));
2831 	}
2832 
2833 	spa_load_note(spa, "using uberblock with txg=%llu",
2834 	    (u_longlong_t)ub->ub_txg);
2835 
2836 	/*
2837 	 * For pools which have the multihost property on determine if the
2838 	 * pool is truly inactive and can be safely imported.  Prevent
2839 	 * hosts which don't have a hostid set from importing the pool.
2840 	 */
2841 	activity_check = spa_activity_check_required(spa, ub, label,
2842 	    spa->spa_config);
2843 	if (activity_check) {
2844 		if (ub->ub_mmp_magic == MMP_MAGIC && ub->ub_mmp_delay &&
2845 		    spa_get_hostid() == 0) {
2846 			nvlist_free(label);
2847 			fnvlist_add_uint64(spa->spa_load_info,
2848 			    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
2849 			return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
2850 		}
2851 
2852 		int error = spa_activity_check(spa, ub, spa->spa_config);
2853 		if (error) {
2854 			nvlist_free(label);
2855 			return (error);
2856 		}
2857 
2858 		fnvlist_add_uint64(spa->spa_load_info,
2859 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_INACTIVE);
2860 		fnvlist_add_uint64(spa->spa_load_info,
2861 		    ZPOOL_CONFIG_MMP_TXG, ub->ub_txg);
2862 	}
2863 
2864 	/*
2865 	 * If the pool has an unsupported version we can't open it.
2866 	 */
2867 	if (!SPA_VERSION_IS_SUPPORTED(ub->ub_version)) {
2868 		nvlist_free(label);
2869 		spa_load_failed(spa, "version %llu is not supported",
2870 		    (u_longlong_t)ub->ub_version);
2871 		return (spa_vdev_err(rvd, VDEV_AUX_VERSION_NEWER, ENOTSUP));
2872 	}
2873 
2874 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
2875 		nvlist_t *features;
2876 
2877 		/*
2878 		 * If we weren't able to find what's necessary for reading the
2879 		 * MOS in the label, return failure.
2880 		 */
2881 		if (label == NULL) {
2882 			spa_load_failed(spa, "label config unavailable");
2883 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
2884 			    ENXIO));
2885 		}
2886 
2887 		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_FEATURES_FOR_READ,
2888 		    &features) != 0) {
2889 			nvlist_free(label);
2890 			spa_load_failed(spa, "invalid label: '%s' missing",
2891 			    ZPOOL_CONFIG_FEATURES_FOR_READ);
2892 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
2893 			    ENXIO));
2894 		}
2895 
2896 		/*
2897 		 * Update our in-core representation with the definitive values
2898 		 * from the label.
2899 		 */
2900 		nvlist_free(spa->spa_label_features);
2901 		VERIFY(nvlist_dup(features, &spa->spa_label_features, 0) == 0);
2902 	}
2903 
2904 	nvlist_free(label);
2905 
2906 	/*
2907 	 * Look through entries in the label nvlist's features_for_read. If
2908 	 * there is a feature listed there which we don't understand then we
2909 	 * cannot open a pool.
2910 	 */
2911 	if (ub->ub_version >= SPA_VERSION_FEATURES) {
2912 		nvlist_t *unsup_feat;
2913 
2914 		VERIFY(nvlist_alloc(&unsup_feat, NV_UNIQUE_NAME, KM_SLEEP) ==
2915 		    0);
2916 
2917 		for (nvpair_t *nvp = nvlist_next_nvpair(spa->spa_label_features,
2918 		    NULL); nvp != NULL;
2919 		    nvp = nvlist_next_nvpair(spa->spa_label_features, nvp)) {
2920 			if (!zfeature_is_supported(nvpair_name(nvp))) {
2921 				VERIFY(nvlist_add_string(unsup_feat,
2922 				    nvpair_name(nvp), "") == 0);
2923 			}
2924 		}
2925 
2926 		if (!nvlist_empty(unsup_feat)) {
2927 			VERIFY(nvlist_add_nvlist(spa->spa_load_info,
2928 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat) == 0);
2929 			nvlist_free(unsup_feat);
2930 			spa_load_failed(spa, "some features are unsupported");
2931 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
2932 			    ENOTSUP));
2933 		}
2934 
2935 		nvlist_free(unsup_feat);
2936 	}
2937 
2938 	if (type != SPA_IMPORT_ASSEMBLE && spa->spa_config_splitting) {
2939 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
2940 		spa_try_repair(spa, spa->spa_config);
2941 		spa_config_exit(spa, SCL_ALL, FTAG);
2942 		nvlist_free(spa->spa_config_splitting);
2943 		spa->spa_config_splitting = NULL;
2944 	}
2945 
2946 	/*
2947 	 * Initialize internal SPA structures.
2948 	 */
2949 	spa_ld_select_uberblock_done(spa, ub);
2950 
2951 	return (0);
2952 }
2953 
2954 static int
2955 spa_ld_open_rootbp(spa_t *spa)
2956 {
2957 	int error = 0;
2958 	vdev_t *rvd = spa->spa_root_vdev;
2959 
2960 	error = dsl_pool_init(spa, spa->spa_first_txg, &spa->spa_dsl_pool);
2961 	if (error != 0) {
2962 		spa_load_failed(spa, "unable to open rootbp in dsl_pool_init "
2963 		    "[error=%d]", error);
2964 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
2965 	}
2966 	spa->spa_meta_objset = spa->spa_dsl_pool->dp_meta_objset;
2967 
2968 	return (0);
2969 }
2970 
2971 static int
2972 spa_ld_trusted_config(spa_t *spa, spa_import_type_t type,
2973     boolean_t reloading)
2974 {
2975 	vdev_t *mrvd, *rvd = spa->spa_root_vdev;
2976 	nvlist_t *nv, *mos_config, *policy;
2977 	int error = 0, copy_error;
2978 	uint64_t healthy_tvds, healthy_tvds_mos;
2979 	uint64_t mos_config_txg;
2980 
2981 	if (spa_dir_prop(spa, DMU_POOL_CONFIG, &spa->spa_config_object, B_TRUE)
2982 	    != 0)
2983 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
2984 
2985 	/*
2986 	 * If we're assembling a pool from a split, the config provided is
2987 	 * already trusted so there is nothing to do.
2988 	 */
2989 	if (type == SPA_IMPORT_ASSEMBLE)
2990 		return (0);
2991 
2992 	healthy_tvds = spa_healthy_core_tvds(spa);
2993 
2994 	if (load_nvlist(spa, spa->spa_config_object, &mos_config)
2995 	    != 0) {
2996 		spa_load_failed(spa, "unable to retrieve MOS config");
2997 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
2998 	}
2999 
3000 	/*
3001 	 * If we are doing an open, pool owner wasn't verified yet, thus do
3002 	 * the verification here.
3003 	 */
3004 	if (spa->spa_load_state == SPA_LOAD_OPEN) {
3005 		error = spa_verify_host(spa, mos_config);
3006 		if (error != 0) {
3007 			nvlist_free(mos_config);
3008 			return (error);
3009 		}
3010 	}
3011 
3012 	nv = fnvlist_lookup_nvlist(mos_config, ZPOOL_CONFIG_VDEV_TREE);
3013 
3014 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3015 
3016 	/*
3017 	 * Build a new vdev tree from the trusted config
3018 	 */
3019 	VERIFY(spa_config_parse(spa, &mrvd, nv, NULL, 0, VDEV_ALLOC_LOAD) == 0);
3020 
3021 	/*
3022 	 * Vdev paths in the MOS may be obsolete. If the untrusted config was
3023 	 * obtained by scanning /dev/dsk, then it will have the right vdev
3024 	 * paths. We update the trusted MOS config with this information.
3025 	 * We first try to copy the paths with vdev_copy_path_strict, which
3026 	 * succeeds only when both configs have exactly the same vdev tree.
3027 	 * If that fails, we fall back to a more flexible method that has a
3028 	 * best effort policy.
3029 	 */
3030 	copy_error = vdev_copy_path_strict(rvd, mrvd);
3031 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3032 		spa_load_note(spa, "provided vdev tree:");
3033 		vdev_dbgmsg_print_tree(rvd, 2);
3034 		spa_load_note(spa, "MOS vdev tree:");
3035 		vdev_dbgmsg_print_tree(mrvd, 2);
3036 	}
3037 	if (copy_error != 0) {
3038 		spa_load_note(spa, "vdev_copy_path_strict failed, falling "
3039 		    "back to vdev_copy_path_relaxed");
3040 		vdev_copy_path_relaxed(rvd, mrvd);
3041 	}
3042 
3043 	vdev_close(rvd);
3044 	vdev_free(rvd);
3045 	spa->spa_root_vdev = mrvd;
3046 	rvd = mrvd;
3047 	spa_config_exit(spa, SCL_ALL, FTAG);
3048 
3049 	/*
3050 	 * We will use spa_config if we decide to reload the spa or if spa_load
3051 	 * fails and we rewind. We must thus regenerate the config using the
3052 	 * MOS information with the updated paths. ZPOOL_LOAD_POLICY is used to
3053 	 * pass settings on how to load the pool and is not stored in the MOS.
3054 	 * We copy it over to our new, trusted config.
3055 	 */
3056 	mos_config_txg = fnvlist_lookup_uint64(mos_config,
3057 	    ZPOOL_CONFIG_POOL_TXG);
3058 	nvlist_free(mos_config);
3059 	mos_config = spa_config_generate(spa, NULL, mos_config_txg, B_FALSE);
3060 	if (nvlist_lookup_nvlist(spa->spa_config, ZPOOL_LOAD_POLICY,
3061 	    &policy) == 0)
3062 		fnvlist_add_nvlist(mos_config, ZPOOL_LOAD_POLICY, policy);
3063 	spa_config_set(spa, mos_config);
3064 	spa->spa_config_source = SPA_CONFIG_SRC_MOS;
3065 
3066 	/*
3067 	 * Now that we got the config from the MOS, we should be more strict
3068 	 * in checking blkptrs and can make assumptions about the consistency
3069 	 * of the vdev tree. spa_trust_config must be set to true before opening
3070 	 * vdevs in order for them to be writeable.
3071 	 */
3072 	spa->spa_trust_config = B_TRUE;
3073 
3074 	/*
3075 	 * Open and validate the new vdev tree
3076 	 */
3077 	error = spa_ld_open_vdevs(spa);
3078 	if (error != 0)
3079 		return (error);
3080 
3081 	error = spa_ld_validate_vdevs(spa);
3082 	if (error != 0)
3083 		return (error);
3084 
3085 	if (copy_error != 0 || spa_load_print_vdev_tree) {
3086 		spa_load_note(spa, "final vdev tree:");
3087 		vdev_dbgmsg_print_tree(rvd, 2);
3088 	}
3089 
3090 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT &&
3091 	    !spa->spa_extreme_rewind && zfs_max_missing_tvds == 0) {
3092 		/*
3093 		 * Sanity check to make sure that we are indeed loading the
3094 		 * latest uberblock. If we missed SPA_SYNC_MIN_VDEVS tvds
3095 		 * in the config provided and they happened to be the only ones
3096 		 * to have the latest uberblock, we could involuntarily perform
3097 		 * an extreme rewind.
3098 		 */
3099 		healthy_tvds_mos = spa_healthy_core_tvds(spa);
3100 		if (healthy_tvds_mos - healthy_tvds >=
3101 		    SPA_SYNC_MIN_VDEVS) {
3102 			spa_load_note(spa, "config provided misses too many "
3103 			    "top-level vdevs compared to MOS (%lld vs %lld). ",
3104 			    (u_longlong_t)healthy_tvds,
3105 			    (u_longlong_t)healthy_tvds_mos);
3106 			spa_load_note(spa, "vdev tree:");
3107 			vdev_dbgmsg_print_tree(rvd, 2);
3108 			if (reloading) {
3109 				spa_load_failed(spa, "config was already "
3110 				    "provided from MOS. Aborting.");
3111 				return (spa_vdev_err(rvd,
3112 				    VDEV_AUX_CORRUPT_DATA, EIO));
3113 			}
3114 			spa_load_note(spa, "spa must be reloaded using MOS "
3115 			    "config");
3116 			return (SET_ERROR(EAGAIN));
3117 		}
3118 	}
3119 
3120 	error = spa_check_for_missing_logs(spa);
3121 	if (error != 0)
3122 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM, ENXIO));
3123 
3124 	if (rvd->vdev_guid_sum != spa->spa_uberblock.ub_guid_sum) {
3125 		spa_load_failed(spa, "uberblock guid sum doesn't match MOS "
3126 		    "guid sum (%llu != %llu)",
3127 		    (u_longlong_t)spa->spa_uberblock.ub_guid_sum,
3128 		    (u_longlong_t)rvd->vdev_guid_sum);
3129 		return (spa_vdev_err(rvd, VDEV_AUX_BAD_GUID_SUM,
3130 		    ENXIO));
3131 	}
3132 
3133 	return (0);
3134 }
3135 
3136 static int
3137 spa_ld_open_indirect_vdev_metadata(spa_t *spa)
3138 {
3139 	int error = 0;
3140 	vdev_t *rvd = spa->spa_root_vdev;
3141 
3142 	/*
3143 	 * Everything that we read before spa_remove_init() must be stored
3144 	 * on concreted vdevs.  Therefore we do this as early as possible.
3145 	 */
3146 	error = spa_remove_init(spa);
3147 	if (error != 0) {
3148 		spa_load_failed(spa, "spa_remove_init failed [error=%d]",
3149 		    error);
3150 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3151 	}
3152 
3153 	/*
3154 	 * Retrieve information needed to condense indirect vdev mappings.
3155 	 */
3156 	error = spa_condense_init(spa);
3157 	if (error != 0) {
3158 		spa_load_failed(spa, "spa_condense_init failed [error=%d]",
3159 		    error);
3160 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
3161 	}
3162 
3163 	return (0);
3164 }
3165 
3166 static int
3167 spa_ld_check_features(spa_t *spa, boolean_t *missing_feat_writep)
3168 {
3169 	int error = 0;
3170 	vdev_t *rvd = spa->spa_root_vdev;
3171 
3172 	if (spa_version(spa) >= SPA_VERSION_FEATURES) {
3173 		boolean_t missing_feat_read = B_FALSE;
3174 		nvlist_t *unsup_feat, *enabled_feat;
3175 
3176 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_READ,
3177 		    &spa->spa_feat_for_read_obj, B_TRUE) != 0) {
3178 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3179 		}
3180 
3181 		if (spa_dir_prop(spa, DMU_POOL_FEATURES_FOR_WRITE,
3182 		    &spa->spa_feat_for_write_obj, B_TRUE) != 0) {
3183 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3184 		}
3185 
3186 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_DESCRIPTIONS,
3187 		    &spa->spa_feat_desc_obj, B_TRUE) != 0) {
3188 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3189 		}
3190 
3191 		enabled_feat = fnvlist_alloc();
3192 		unsup_feat = fnvlist_alloc();
3193 
3194 		if (!spa_features_check(spa, B_FALSE,
3195 		    unsup_feat, enabled_feat))
3196 			missing_feat_read = B_TRUE;
3197 
3198 		if (spa_writeable(spa) ||
3199 		    spa->spa_load_state == SPA_LOAD_TRYIMPORT) {
3200 			if (!spa_features_check(spa, B_TRUE,
3201 			    unsup_feat, enabled_feat)) {
3202 				*missing_feat_writep = B_TRUE;
3203 			}
3204 		}
3205 
3206 		fnvlist_add_nvlist(spa->spa_load_info,
3207 		    ZPOOL_CONFIG_ENABLED_FEAT, enabled_feat);
3208 
3209 		if (!nvlist_empty(unsup_feat)) {
3210 			fnvlist_add_nvlist(spa->spa_load_info,
3211 			    ZPOOL_CONFIG_UNSUP_FEAT, unsup_feat);
3212 		}
3213 
3214 		fnvlist_free(enabled_feat);
3215 		fnvlist_free(unsup_feat);
3216 
3217 		if (!missing_feat_read) {
3218 			fnvlist_add_boolean(spa->spa_load_info,
3219 			    ZPOOL_CONFIG_CAN_RDONLY);
3220 		}
3221 
3222 		/*
3223 		 * If the state is SPA_LOAD_TRYIMPORT, our objective is
3224 		 * twofold: to determine whether the pool is available for
3225 		 * import in read-write mode and (if it is not) whether the
3226 		 * pool is available for import in read-only mode. If the pool
3227 		 * is available for import in read-write mode, it is displayed
3228 		 * as available in userland; if it is not available for import
3229 		 * in read-only mode, it is displayed as unavailable in
3230 		 * userland. If the pool is available for import in read-only
3231 		 * mode but not read-write mode, it is displayed as unavailable
3232 		 * in userland with a special note that the pool is actually
3233 		 * available for open in read-only mode.
3234 		 *
3235 		 * As a result, if the state is SPA_LOAD_TRYIMPORT and we are
3236 		 * missing a feature for write, we must first determine whether
3237 		 * the pool can be opened read-only before returning to
3238 		 * userland in order to know whether to display the
3239 		 * abovementioned note.
3240 		 */
3241 		if (missing_feat_read || (*missing_feat_writep &&
3242 		    spa_writeable(spa))) {
3243 			spa_load_failed(spa, "pool uses unsupported features");
3244 			return (spa_vdev_err(rvd, VDEV_AUX_UNSUP_FEAT,
3245 			    ENOTSUP));
3246 		}
3247 
3248 		/*
3249 		 * Load refcounts for ZFS features from disk into an in-memory
3250 		 * cache during SPA initialization.
3251 		 */
3252 		for (spa_feature_t i = 0; i < SPA_FEATURES; i++) {
3253 			uint64_t refcount;
3254 
3255 			error = feature_get_refcount_from_disk(spa,
3256 			    &spa_feature_table[i], &refcount);
3257 			if (error == 0) {
3258 				spa->spa_feat_refcount_cache[i] = refcount;
3259 			} else if (error == ENOTSUP) {
3260 				spa->spa_feat_refcount_cache[i] =
3261 				    SPA_FEATURE_DISABLED;
3262 			} else {
3263 				spa_load_failed(spa, "error getting refcount "
3264 				    "for feature %s [error=%d]",
3265 				    spa_feature_table[i].fi_guid, error);
3266 				return (spa_vdev_err(rvd,
3267 				    VDEV_AUX_CORRUPT_DATA, EIO));
3268 			}
3269 		}
3270 	}
3271 
3272 	if (spa_feature_is_active(spa, SPA_FEATURE_ENABLED_TXG)) {
3273 		if (spa_dir_prop(spa, DMU_POOL_FEATURE_ENABLED_TXG,
3274 		    &spa->spa_feat_enabled_txg_obj, B_TRUE) != 0)
3275 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3276 	}
3277 
3278 	/*
3279 	 * Encryption was added before bookmark_v2, even though bookmark_v2
3280 	 * is now a dependency. If this pool has encryption enabled without
3281 	 * bookmark_v2, trigger an errata message.
3282 	 */
3283 	if (spa_feature_is_enabled(spa, SPA_FEATURE_ENCRYPTION) &&
3284 	    !spa_feature_is_enabled(spa, SPA_FEATURE_BOOKMARK_V2)) {
3285 		spa->spa_errata = ZPOOL_ERRATA_ZOL_8308_ENCRYPTION;
3286 	}
3287 
3288 	return (0);
3289 }
3290 
3291 static int
3292 spa_ld_load_special_directories(spa_t *spa)
3293 {
3294 	int error = 0;
3295 	vdev_t *rvd = spa->spa_root_vdev;
3296 
3297 	spa->spa_is_initializing = B_TRUE;
3298 	error = dsl_pool_open(spa->spa_dsl_pool);
3299 	spa->spa_is_initializing = B_FALSE;
3300 	if (error != 0) {
3301 		spa_load_failed(spa, "dsl_pool_open failed [error=%d]", error);
3302 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3303 	}
3304 
3305 	return (0);
3306 }
3307 
3308 static int
3309 spa_ld_get_props(spa_t *spa)
3310 {
3311 	int error = 0;
3312 	uint64_t obj;
3313 	vdev_t *rvd = spa->spa_root_vdev;
3314 
3315 	/* Grab the secret checksum salt from the MOS. */
3316 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3317 	    DMU_POOL_CHECKSUM_SALT, 1,
3318 	    sizeof (spa->spa_cksum_salt.zcs_bytes),
3319 	    spa->spa_cksum_salt.zcs_bytes);
3320 	if (error == ENOENT) {
3321 		/* Generate a new salt for subsequent use */
3322 		(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
3323 		    sizeof (spa->spa_cksum_salt.zcs_bytes));
3324 	} else if (error != 0) {
3325 		spa_load_failed(spa, "unable to retrieve checksum salt from "
3326 		    "MOS [error=%d]", error);
3327 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3328 	}
3329 
3330 	if (spa_dir_prop(spa, DMU_POOL_SYNC_BPOBJ, &obj, B_TRUE) != 0)
3331 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3332 	error = bpobj_open(&spa->spa_deferred_bpobj, spa->spa_meta_objset, obj);
3333 	if (error != 0) {
3334 		spa_load_failed(spa, "error opening deferred-frees bpobj "
3335 		    "[error=%d]", error);
3336 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3337 	}
3338 
3339 	/*
3340 	 * Load the bit that tells us to use the new accounting function
3341 	 * (raid-z deflation).  If we have an older pool, this will not
3342 	 * be present.
3343 	 */
3344 	error = spa_dir_prop(spa, DMU_POOL_DEFLATE, &spa->spa_deflate, B_FALSE);
3345 	if (error != 0 && error != ENOENT)
3346 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3347 
3348 	error = spa_dir_prop(spa, DMU_POOL_CREATION_VERSION,
3349 	    &spa->spa_creation_version, B_FALSE);
3350 	if (error != 0 && error != ENOENT)
3351 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3352 
3353 	/*
3354 	 * Load the persistent error log.  If we have an older pool, this will
3355 	 * not be present.
3356 	 */
3357 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_LAST, &spa->spa_errlog_last,
3358 	    B_FALSE);
3359 	if (error != 0 && error != ENOENT)
3360 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3361 
3362 	error = spa_dir_prop(spa, DMU_POOL_ERRLOG_SCRUB,
3363 	    &spa->spa_errlog_scrub, B_FALSE);
3364 	if (error != 0 && error != ENOENT)
3365 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3366 
3367 	/*
3368 	 * Load the history object.  If we have an older pool, this
3369 	 * will not be present.
3370 	 */
3371 	error = spa_dir_prop(spa, DMU_POOL_HISTORY, &spa->spa_history, B_FALSE);
3372 	if (error != 0 && error != ENOENT)
3373 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3374 
3375 	/*
3376 	 * Load the per-vdev ZAP map. If we have an older pool, this will not
3377 	 * be present; in this case, defer its creation to a later time to
3378 	 * avoid dirtying the MOS this early / out of sync context. See
3379 	 * spa_sync_config_object.
3380 	 */
3381 
3382 	/* The sentinel is only available in the MOS config. */
3383 	nvlist_t *mos_config;
3384 	if (load_nvlist(spa, spa->spa_config_object, &mos_config) != 0) {
3385 		spa_load_failed(spa, "unable to retrieve MOS config");
3386 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3387 	}
3388 
3389 	error = spa_dir_prop(spa, DMU_POOL_VDEV_ZAP_MAP,
3390 	    &spa->spa_all_vdev_zaps, B_FALSE);
3391 
3392 	if (error == ENOENT) {
3393 		VERIFY(!nvlist_exists(mos_config,
3394 		    ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
3395 		spa->spa_avz_action = AVZ_ACTION_INITIALIZE;
3396 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
3397 	} else if (error != 0) {
3398 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3399 	} else if (!nvlist_exists(mos_config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS)) {
3400 		/*
3401 		 * An older version of ZFS overwrote the sentinel value, so
3402 		 * we have orphaned per-vdev ZAPs in the MOS. Defer their
3403 		 * destruction to later; see spa_sync_config_object.
3404 		 */
3405 		spa->spa_avz_action = AVZ_ACTION_DESTROY;
3406 		/*
3407 		 * We're assuming that no vdevs have had their ZAPs created
3408 		 * before this. Better be sure of it.
3409 		 */
3410 		ASSERT0(vdev_count_verify_zaps(spa->spa_root_vdev));
3411 	}
3412 	nvlist_free(mos_config);
3413 
3414 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
3415 
3416 	error = spa_dir_prop(spa, DMU_POOL_PROPS, &spa->spa_pool_props_object,
3417 	    B_FALSE);
3418 	if (error && error != ENOENT)
3419 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3420 
3421 	if (error == 0) {
3422 		uint64_t autoreplace;
3423 
3424 		spa_prop_find(spa, ZPOOL_PROP_BOOTFS, &spa->spa_bootfs);
3425 		spa_prop_find(spa, ZPOOL_PROP_AUTOREPLACE, &autoreplace);
3426 		spa_prop_find(spa, ZPOOL_PROP_DELEGATION, &spa->spa_delegation);
3427 		spa_prop_find(spa, ZPOOL_PROP_FAILUREMODE, &spa->spa_failmode);
3428 		spa_prop_find(spa, ZPOOL_PROP_AUTOEXPAND, &spa->spa_autoexpand);
3429 		spa_prop_find(spa, ZPOOL_PROP_MULTIHOST, &spa->spa_multihost);
3430 		spa_prop_find(spa, ZPOOL_PROP_DEDUPDITTO,
3431 		    &spa->spa_dedup_ditto);
3432 
3433 		spa->spa_autoreplace = (autoreplace != 0);
3434 	}
3435 
3436 	/*
3437 	 * If we are importing a pool with missing top-level vdevs,
3438 	 * we enforce that the pool doesn't panic or get suspended on
3439 	 * error since the likelihood of missing data is extremely high.
3440 	 */
3441 	if (spa->spa_missing_tvds > 0 &&
3442 	    spa->spa_failmode != ZIO_FAILURE_MODE_CONTINUE &&
3443 	    spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
3444 		spa_load_note(spa, "forcing failmode to 'continue' "
3445 		    "as some top level vdevs are missing");
3446 		spa->spa_failmode = ZIO_FAILURE_MODE_CONTINUE;
3447 	}
3448 
3449 	return (0);
3450 }
3451 
3452 static int
3453 spa_ld_open_aux_vdevs(spa_t *spa, spa_import_type_t type)
3454 {
3455 	int error = 0;
3456 	vdev_t *rvd = spa->spa_root_vdev;
3457 
3458 	/*
3459 	 * If we're assembling the pool from the split-off vdevs of
3460 	 * an existing pool, we don't want to attach the spares & cache
3461 	 * devices.
3462 	 */
3463 
3464 	/*
3465 	 * Load any hot spares for this pool.
3466 	 */
3467 	error = spa_dir_prop(spa, DMU_POOL_SPARES, &spa->spa_spares.sav_object,
3468 	    B_FALSE);
3469 	if (error != 0 && error != ENOENT)
3470 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3471 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
3472 		ASSERT(spa_version(spa) >= SPA_VERSION_SPARES);
3473 		if (load_nvlist(spa, spa->spa_spares.sav_object,
3474 		    &spa->spa_spares.sav_config) != 0) {
3475 			spa_load_failed(spa, "error loading spares nvlist");
3476 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3477 		}
3478 
3479 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3480 		spa_load_spares(spa);
3481 		spa_config_exit(spa, SCL_ALL, FTAG);
3482 	} else if (error == 0) {
3483 		spa->spa_spares.sav_sync = B_TRUE;
3484 	}
3485 
3486 	/*
3487 	 * Load any level 2 ARC devices for this pool.
3488 	 */
3489 	error = spa_dir_prop(spa, DMU_POOL_L2CACHE,
3490 	    &spa->spa_l2cache.sav_object, B_FALSE);
3491 	if (error != 0 && error != ENOENT)
3492 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3493 	if (error == 0 && type != SPA_IMPORT_ASSEMBLE) {
3494 		ASSERT(spa_version(spa) >= SPA_VERSION_L2CACHE);
3495 		if (load_nvlist(spa, spa->spa_l2cache.sav_object,
3496 		    &spa->spa_l2cache.sav_config) != 0) {
3497 			spa_load_failed(spa, "error loading l2cache nvlist");
3498 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3499 		}
3500 
3501 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3502 		spa_load_l2cache(spa);
3503 		spa_config_exit(spa, SCL_ALL, FTAG);
3504 	} else if (error == 0) {
3505 		spa->spa_l2cache.sav_sync = B_TRUE;
3506 	}
3507 
3508 	return (0);
3509 }
3510 
3511 static int
3512 spa_ld_load_vdev_metadata(spa_t *spa)
3513 {
3514 	int error = 0;
3515 	vdev_t *rvd = spa->spa_root_vdev;
3516 
3517 	/*
3518 	 * If the 'multihost' property is set, then never allow a pool to
3519 	 * be imported when the system hostid is zero.  The exception to
3520 	 * this rule is zdb which is always allowed to access pools.
3521 	 */
3522 	if (spa_multihost(spa) && spa_get_hostid() == 0 &&
3523 	    (spa->spa_import_flags & ZFS_IMPORT_SKIP_MMP) == 0) {
3524 		fnvlist_add_uint64(spa->spa_load_info,
3525 		    ZPOOL_CONFIG_MMP_STATE, MMP_STATE_NO_HOSTID);
3526 		return (spa_vdev_err(rvd, VDEV_AUX_ACTIVE, EREMOTEIO));
3527 	}
3528 
3529 	/*
3530 	 * If the 'autoreplace' property is set, then post a resource notifying
3531 	 * the ZFS DE that it should not issue any faults for unopenable
3532 	 * devices.  We also iterate over the vdevs, and post a sysevent for any
3533 	 * unopenable vdevs so that the normal autoreplace handler can take
3534 	 * over.
3535 	 */
3536 	if (spa->spa_autoreplace && spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
3537 		spa_check_removed(spa->spa_root_vdev);
3538 		/*
3539 		 * For the import case, this is done in spa_import(), because
3540 		 * at this point we're using the spare definitions from
3541 		 * the MOS config, not necessarily from the userland config.
3542 		 */
3543 		if (spa->spa_load_state != SPA_LOAD_IMPORT) {
3544 			spa_aux_check_removed(&spa->spa_spares);
3545 			spa_aux_check_removed(&spa->spa_l2cache);
3546 		}
3547 	}
3548 
3549 	/*
3550 	 * Load the vdev metadata such as metaslabs, DTLs, spacemap object, etc.
3551 	 */
3552 	error = vdev_load(rvd);
3553 	if (error != 0) {
3554 		spa_load_failed(spa, "vdev_load failed [error=%d]", error);
3555 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, error));
3556 	}
3557 
3558 	/*
3559 	 * Propagate the leaf DTLs we just loaded all the way up the vdev tree.
3560 	 */
3561 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3562 	vdev_dtl_reassess(rvd, 0, 0, B_FALSE);
3563 	spa_config_exit(spa, SCL_ALL, FTAG);
3564 
3565 	return (0);
3566 }
3567 
3568 static int
3569 spa_ld_load_dedup_tables(spa_t *spa)
3570 {
3571 	int error = 0;
3572 	vdev_t *rvd = spa->spa_root_vdev;
3573 
3574 	error = ddt_load(spa);
3575 	if (error != 0) {
3576 		spa_load_failed(spa, "ddt_load failed [error=%d]", error);
3577 		return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA, EIO));
3578 	}
3579 
3580 	return (0);
3581 }
3582 
3583 static int
3584 spa_ld_verify_logs(spa_t *spa, spa_import_type_t type, char **ereport)
3585 {
3586 	vdev_t *rvd = spa->spa_root_vdev;
3587 
3588 	if (type != SPA_IMPORT_ASSEMBLE && spa_writeable(spa)) {
3589 		boolean_t missing = spa_check_logs(spa);
3590 		if (missing) {
3591 			if (spa->spa_missing_tvds != 0) {
3592 				spa_load_note(spa, "spa_check_logs failed "
3593 				    "so dropping the logs");
3594 			} else {
3595 				*ereport = FM_EREPORT_ZFS_LOG_REPLAY;
3596 				spa_load_failed(spa, "spa_check_logs failed");
3597 				return (spa_vdev_err(rvd, VDEV_AUX_BAD_LOG,
3598 				    ENXIO));
3599 			}
3600 		}
3601 	}
3602 
3603 	return (0);
3604 }
3605 
3606 static int
3607 spa_ld_verify_pool_data(spa_t *spa)
3608 {
3609 	int error = 0;
3610 	vdev_t *rvd = spa->spa_root_vdev;
3611 
3612 	/*
3613 	 * We've successfully opened the pool, verify that we're ready
3614 	 * to start pushing transactions.
3615 	 */
3616 	if (spa->spa_load_state != SPA_LOAD_TRYIMPORT) {
3617 		error = spa_load_verify(spa);
3618 		if (error != 0) {
3619 			spa_load_failed(spa, "spa_load_verify failed "
3620 			    "[error=%d]", error);
3621 			return (spa_vdev_err(rvd, VDEV_AUX_CORRUPT_DATA,
3622 			    error));
3623 		}
3624 	}
3625 
3626 	return (0);
3627 }
3628 
3629 static void
3630 spa_ld_claim_log_blocks(spa_t *spa)
3631 {
3632 	dmu_tx_t *tx;
3633 	dsl_pool_t *dp = spa_get_dsl(spa);
3634 
3635 	/*
3636 	 * Claim log blocks that haven't been committed yet.
3637 	 * This must all happen in a single txg.
3638 	 * Note: spa_claim_max_txg is updated by spa_claim_notify(),
3639 	 * invoked from zil_claim_log_block()'s i/o done callback.
3640 	 * Price of rollback is that we abandon the log.
3641 	 */
3642 	spa->spa_claiming = B_TRUE;
3643 
3644 	tx = dmu_tx_create_assigned(dp, spa_first_txg(spa));
3645 	(void) dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
3646 	    zil_claim, tx, DS_FIND_CHILDREN);
3647 	dmu_tx_commit(tx);
3648 
3649 	spa->spa_claiming = B_FALSE;
3650 
3651 	spa_set_log_state(spa, SPA_LOG_GOOD);
3652 }
3653 
3654 static void
3655 spa_ld_check_for_config_update(spa_t *spa, uint64_t config_cache_txg,
3656     boolean_t update_config_cache)
3657 {
3658 	vdev_t *rvd = spa->spa_root_vdev;
3659 	int need_update = B_FALSE;
3660 
3661 	/*
3662 	 * If the config cache is stale, or we have uninitialized
3663 	 * metaslabs (see spa_vdev_add()), then update the config.
3664 	 *
3665 	 * If this is a verbatim import, trust the current
3666 	 * in-core spa_config and update the disk labels.
3667 	 */
3668 	if (update_config_cache || config_cache_txg != spa->spa_config_txg ||
3669 	    spa->spa_load_state == SPA_LOAD_IMPORT ||
3670 	    spa->spa_load_state == SPA_LOAD_RECOVER ||
3671 	    (spa->spa_import_flags & ZFS_IMPORT_VERBATIM))
3672 		need_update = B_TRUE;
3673 
3674 	for (int c = 0; c < rvd->vdev_children; c++)
3675 		if (rvd->vdev_child[c]->vdev_ms_array == 0)
3676 			need_update = B_TRUE;
3677 
3678 	/*
3679 	 * Update the config cache asychronously in case we're the
3680 	 * root pool, in which case the config cache isn't writable yet.
3681 	 */
3682 	if (need_update)
3683 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
3684 }
3685 
3686 static void
3687 spa_ld_prepare_for_reload(spa_t *spa)
3688 {
3689 	int mode = spa->spa_mode;
3690 	int async_suspended = spa->spa_async_suspended;
3691 
3692 	spa_unload(spa);
3693 	spa_deactivate(spa);
3694 	spa_activate(spa, mode);
3695 
3696 	/*
3697 	 * We save the value of spa_async_suspended as it gets reset to 0 by
3698 	 * spa_unload(). We want to restore it back to the original value before
3699 	 * returning as we might be calling spa_async_resume() later.
3700 	 */
3701 	spa->spa_async_suspended = async_suspended;
3702 }
3703 
3704 static int
3705 spa_ld_read_checkpoint_txg(spa_t *spa)
3706 {
3707 	uberblock_t checkpoint;
3708 	int error = 0;
3709 
3710 	ASSERT0(spa->spa_checkpoint_txg);
3711 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
3712 
3713 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3714 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
3715 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
3716 
3717 	if (error == ENOENT)
3718 		return (0);
3719 
3720 	if (error != 0)
3721 		return (error);
3722 
3723 	ASSERT3U(checkpoint.ub_txg, !=, 0);
3724 	ASSERT3U(checkpoint.ub_checkpoint_txg, !=, 0);
3725 	ASSERT3U(checkpoint.ub_timestamp, !=, 0);
3726 	spa->spa_checkpoint_txg = checkpoint.ub_txg;
3727 	spa->spa_checkpoint_info.sci_timestamp = checkpoint.ub_timestamp;
3728 
3729 	return (0);
3730 }
3731 
3732 static int
3733 spa_ld_mos_init(spa_t *spa, spa_import_type_t type)
3734 {
3735 	int error = 0;
3736 
3737 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
3738 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
3739 
3740 	/*
3741 	 * Never trust the config that is provided unless we are assembling
3742 	 * a pool following a split.
3743 	 * This means don't trust blkptrs and the vdev tree in general. This
3744 	 * also effectively puts the spa in read-only mode since
3745 	 * spa_writeable() checks for spa_trust_config to be true.
3746 	 * We will later load a trusted config from the MOS.
3747 	 */
3748 	if (type != SPA_IMPORT_ASSEMBLE)
3749 		spa->spa_trust_config = B_FALSE;
3750 
3751 	/*
3752 	 * Parse the config provided to create a vdev tree.
3753 	 */
3754 	error = spa_ld_parse_config(spa, type);
3755 	if (error != 0)
3756 		return (error);
3757 
3758 	/*
3759 	 * Now that we have the vdev tree, try to open each vdev. This involves
3760 	 * opening the underlying physical device, retrieving its geometry and
3761 	 * probing the vdev with a dummy I/O. The state of each vdev will be set
3762 	 * based on the success of those operations. After this we'll be ready
3763 	 * to read from the vdevs.
3764 	 */
3765 	error = spa_ld_open_vdevs(spa);
3766 	if (error != 0)
3767 		return (error);
3768 
3769 	/*
3770 	 * Read the label of each vdev and make sure that the GUIDs stored
3771 	 * there match the GUIDs in the config provided.
3772 	 * If we're assembling a new pool that's been split off from an
3773 	 * existing pool, the labels haven't yet been updated so we skip
3774 	 * validation for now.
3775 	 */
3776 	if (type != SPA_IMPORT_ASSEMBLE) {
3777 		error = spa_ld_validate_vdevs(spa);
3778 		if (error != 0)
3779 			return (error);
3780 	}
3781 
3782 	/*
3783 	 * Read all vdev labels to find the best uberblock (i.e. latest,
3784 	 * unless spa_load_max_txg is set) and store it in spa_uberblock. We
3785 	 * get the list of features required to read blkptrs in the MOS from
3786 	 * the vdev label with the best uberblock and verify that our version
3787 	 * of zfs supports them all.
3788 	 */
3789 	error = spa_ld_select_uberblock(spa, type);
3790 	if (error != 0)
3791 		return (error);
3792 
3793 	/*
3794 	 * Pass that uberblock to the dsl_pool layer which will open the root
3795 	 * blkptr. This blkptr points to the latest version of the MOS and will
3796 	 * allow us to read its contents.
3797 	 */
3798 	error = spa_ld_open_rootbp(spa);
3799 	if (error != 0)
3800 		return (error);
3801 
3802 	return (0);
3803 }
3804 
3805 static int
3806 spa_ld_checkpoint_rewind(spa_t *spa)
3807 {
3808 	uberblock_t checkpoint;
3809 	int error = 0;
3810 
3811 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
3812 	ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
3813 
3814 	error = zap_lookup(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
3815 	    DMU_POOL_ZPOOL_CHECKPOINT, sizeof (uint64_t),
3816 	    sizeof (uberblock_t) / sizeof (uint64_t), &checkpoint);
3817 
3818 	if (error != 0) {
3819 		spa_load_failed(spa, "unable to retrieve checkpointed "
3820 		    "uberblock from the MOS config [error=%d]", error);
3821 
3822 		if (error == ENOENT)
3823 			error = ZFS_ERR_NO_CHECKPOINT;
3824 
3825 		return (error);
3826 	}
3827 
3828 	ASSERT3U(checkpoint.ub_txg, <, spa->spa_uberblock.ub_txg);
3829 	ASSERT3U(checkpoint.ub_txg, ==, checkpoint.ub_checkpoint_txg);
3830 
3831 	/*
3832 	 * We need to update the txg and timestamp of the checkpointed
3833 	 * uberblock to be higher than the latest one. This ensures that
3834 	 * the checkpointed uberblock is selected if we were to close and
3835 	 * reopen the pool right after we've written it in the vdev labels.
3836 	 * (also see block comment in vdev_uberblock_compare)
3837 	 */
3838 	checkpoint.ub_txg = spa->spa_uberblock.ub_txg + 1;
3839 	checkpoint.ub_timestamp = gethrestime_sec();
3840 
3841 	/*
3842 	 * Set current uberblock to be the checkpointed uberblock.
3843 	 */
3844 	spa->spa_uberblock = checkpoint;
3845 
3846 	/*
3847 	 * If we are doing a normal rewind, then the pool is open for
3848 	 * writing and we sync the "updated" checkpointed uberblock to
3849 	 * disk. Once this is done, we've basically rewound the whole
3850 	 * pool and there is no way back.
3851 	 *
3852 	 * There are cases when we don't want to attempt and sync the
3853 	 * checkpointed uberblock to disk because we are opening a
3854 	 * pool as read-only. Specifically, verifying the checkpointed
3855 	 * state with zdb, and importing the checkpointed state to get
3856 	 * a "preview" of its content.
3857 	 */
3858 	if (spa_writeable(spa)) {
3859 		vdev_t *rvd = spa->spa_root_vdev;
3860 
3861 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
3862 		vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
3863 		int svdcount = 0;
3864 		int children = rvd->vdev_children;
3865 		int c0 = spa_get_random(children);
3866 
3867 		for (int c = 0; c < children; c++) {
3868 			vdev_t *vd = rvd->vdev_child[(c0 + c) % children];
3869 
3870 			/* Stop when revisiting the first vdev */
3871 			if (c > 0 && svd[0] == vd)
3872 				break;
3873 
3874 			if (vd->vdev_ms_array == 0 || vd->vdev_islog ||
3875 			    !vdev_is_concrete(vd))
3876 				continue;
3877 
3878 			svd[svdcount++] = vd;
3879 			if (svdcount == SPA_SYNC_MIN_VDEVS)
3880 				break;
3881 		}
3882 		error = vdev_config_sync(svd, svdcount, spa->spa_first_txg);
3883 		if (error == 0)
3884 			spa->spa_last_synced_guid = rvd->vdev_guid;
3885 		spa_config_exit(spa, SCL_ALL, FTAG);
3886 
3887 		if (error != 0) {
3888 			spa_load_failed(spa, "failed to write checkpointed "
3889 			    "uberblock to the vdev labels [error=%d]", error);
3890 			return (error);
3891 		}
3892 	}
3893 
3894 	return (0);
3895 }
3896 
3897 static int
3898 spa_ld_mos_with_trusted_config(spa_t *spa, spa_import_type_t type,
3899     boolean_t *update_config_cache)
3900 {
3901 	int error;
3902 
3903 	/*
3904 	 * Parse the config for pool, open and validate vdevs,
3905 	 * select an uberblock, and use that uberblock to open
3906 	 * the MOS.
3907 	 */
3908 	error = spa_ld_mos_init(spa, type);
3909 	if (error != 0)
3910 		return (error);
3911 
3912 	/*
3913 	 * Retrieve the trusted config stored in the MOS and use it to create
3914 	 * a new, exact version of the vdev tree, then reopen all vdevs.
3915 	 */
3916 	error = spa_ld_trusted_config(spa, type, B_FALSE);
3917 	if (error == EAGAIN) {
3918 		if (update_config_cache != NULL)
3919 			*update_config_cache = B_TRUE;
3920 
3921 		/*
3922 		 * Redo the loading process with the trusted config if it is
3923 		 * too different from the untrusted config.
3924 		 */
3925 		spa_ld_prepare_for_reload(spa);
3926 		spa_load_note(spa, "RELOADING");
3927 		error = spa_ld_mos_init(spa, type);
3928 		if (error != 0)
3929 			return (error);
3930 
3931 		error = spa_ld_trusted_config(spa, type, B_TRUE);
3932 		if (error != 0)
3933 			return (error);
3934 
3935 	} else if (error != 0) {
3936 		return (error);
3937 	}
3938 
3939 	return (0);
3940 }
3941 
3942 /*
3943  * Load an existing storage pool, using the config provided. This config
3944  * describes which vdevs are part of the pool and is later validated against
3945  * partial configs present in each vdev's label and an entire copy of the
3946  * config stored in the MOS.
3947  */
3948 static int
3949 spa_load_impl(spa_t *spa, spa_import_type_t type, char **ereport)
3950 {
3951 	int error = 0;
3952 	boolean_t missing_feat_write = B_FALSE;
3953 	boolean_t checkpoint_rewind =
3954 	    (spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
3955 	boolean_t update_config_cache = B_FALSE;
3956 
3957 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
3958 	ASSERT(spa->spa_config_source != SPA_CONFIG_SRC_NONE);
3959 
3960 	spa_load_note(spa, "LOADING");
3961 
3962 	error = spa_ld_mos_with_trusted_config(spa, type, &update_config_cache);
3963 	if (error != 0)
3964 		return (error);
3965 
3966 	/*
3967 	 * If we are rewinding to the checkpoint then we need to repeat
3968 	 * everything we've done so far in this function but this time
3969 	 * selecting the checkpointed uberblock and using that to open
3970 	 * the MOS.
3971 	 */
3972 	if (checkpoint_rewind) {
3973 		/*
3974 		 * If we are rewinding to the checkpoint update config cache
3975 		 * anyway.
3976 		 */
3977 		update_config_cache = B_TRUE;
3978 
3979 		/*
3980 		 * Extract the checkpointed uberblock from the current MOS
3981 		 * and use this as the pool's uberblock from now on. If the
3982 		 * pool is imported as writeable we also write the checkpoint
3983 		 * uberblock to the labels, making the rewind permanent.
3984 		 */
3985 		error = spa_ld_checkpoint_rewind(spa);
3986 		if (error != 0)
3987 			return (error);
3988 
3989 		/*
3990 		 * Redo the loading process process again with the
3991 		 * checkpointed uberblock.
3992 		 */
3993 		spa_ld_prepare_for_reload(spa);
3994 		spa_load_note(spa, "LOADING checkpointed uberblock");
3995 		error = spa_ld_mos_with_trusted_config(spa, type, NULL);
3996 		if (error != 0)
3997 			return (error);
3998 	}
3999 
4000 	/*
4001 	 * Retrieve the checkpoint txg if the pool has a checkpoint.
4002 	 */
4003 	error = spa_ld_read_checkpoint_txg(spa);
4004 	if (error != 0)
4005 		return (error);
4006 
4007 	/*
4008 	 * Retrieve the mapping of indirect vdevs. Those vdevs were removed
4009 	 * from the pool and their contents were re-mapped to other vdevs. Note
4010 	 * that everything that we read before this step must have been
4011 	 * rewritten on concrete vdevs after the last device removal was
4012 	 * initiated. Otherwise we could be reading from indirect vdevs before
4013 	 * we have loaded their mappings.
4014 	 */
4015 	error = spa_ld_open_indirect_vdev_metadata(spa);
4016 	if (error != 0)
4017 		return (error);
4018 
4019 	/*
4020 	 * Retrieve the full list of active features from the MOS and check if
4021 	 * they are all supported.
4022 	 */
4023 	error = spa_ld_check_features(spa, &missing_feat_write);
4024 	if (error != 0)
4025 		return (error);
4026 
4027 	/*
4028 	 * Load several special directories from the MOS needed by the dsl_pool
4029 	 * layer.
4030 	 */
4031 	error = spa_ld_load_special_directories(spa);
4032 	if (error != 0)
4033 		return (error);
4034 
4035 	/*
4036 	 * Retrieve pool properties from the MOS.
4037 	 */
4038 	error = spa_ld_get_props(spa);
4039 	if (error != 0)
4040 		return (error);
4041 
4042 	/*
4043 	 * Retrieve the list of auxiliary devices - cache devices and spares -
4044 	 * and open them.
4045 	 */
4046 	error = spa_ld_open_aux_vdevs(spa, type);
4047 	if (error != 0)
4048 		return (error);
4049 
4050 	/*
4051 	 * Load the metadata for all vdevs. Also check if unopenable devices
4052 	 * should be autoreplaced.
4053 	 */
4054 	error = spa_ld_load_vdev_metadata(spa);
4055 	if (error != 0)
4056 		return (error);
4057 
4058 	error = spa_ld_load_dedup_tables(spa);
4059 	if (error != 0)
4060 		return (error);
4061 
4062 	/*
4063 	 * Verify the logs now to make sure we don't have any unexpected errors
4064 	 * when we claim log blocks later.
4065 	 */
4066 	error = spa_ld_verify_logs(spa, type, ereport);
4067 	if (error != 0)
4068 		return (error);
4069 
4070 	if (missing_feat_write) {
4071 		ASSERT(spa->spa_load_state == SPA_LOAD_TRYIMPORT);
4072 
4073 		/*
4074 		 * At this point, we know that we can open the pool in
4075 		 * read-only mode but not read-write mode. We now have enough
4076 		 * information and can return to userland.
4077 		 */
4078 		return (spa_vdev_err(spa->spa_root_vdev, VDEV_AUX_UNSUP_FEAT,
4079 		    ENOTSUP));
4080 	}
4081 
4082 	/*
4083 	 * Traverse the last txgs to make sure the pool was left off in a safe
4084 	 * state. When performing an extreme rewind, we verify the whole pool,
4085 	 * which can take a very long time.
4086 	 */
4087 	error = spa_ld_verify_pool_data(spa);
4088 	if (error != 0)
4089 		return (error);
4090 
4091 	/*
4092 	 * Calculate the deflated space for the pool. This must be done before
4093 	 * we write anything to the pool because we'd need to update the space
4094 	 * accounting using the deflated sizes.
4095 	 */
4096 	spa_update_dspace(spa);
4097 
4098 	/*
4099 	 * We have now retrieved all the information we needed to open the
4100 	 * pool. If we are importing the pool in read-write mode, a few
4101 	 * additional steps must be performed to finish the import.
4102 	 */
4103 	if (spa_writeable(spa) && (spa->spa_load_state == SPA_LOAD_RECOVER ||
4104 	    spa->spa_load_max_txg == UINT64_MAX)) {
4105 		uint64_t config_cache_txg = spa->spa_config_txg;
4106 
4107 		ASSERT(spa->spa_load_state != SPA_LOAD_TRYIMPORT);
4108 
4109 		/*
4110 		 * In case of a checkpoint rewind, log the original txg
4111 		 * of the checkpointed uberblock.
4112 		 */
4113 		if (checkpoint_rewind) {
4114 			spa_history_log_internal(spa, "checkpoint rewind",
4115 			    NULL, "rewound state to txg=%llu",
4116 			    (u_longlong_t)spa->spa_uberblock.ub_checkpoint_txg);
4117 		}
4118 
4119 		/*
4120 		 * Traverse the ZIL and claim all blocks.
4121 		 */
4122 		spa_ld_claim_log_blocks(spa);
4123 
4124 		/*
4125 		 * Kick-off the syncing thread.
4126 		 */
4127 		spa->spa_sync_on = B_TRUE;
4128 		txg_sync_start(spa->spa_dsl_pool);
4129 		mmp_thread_start(spa);
4130 
4131 		/*
4132 		 * Wait for all claims to sync.  We sync up to the highest
4133 		 * claimed log block birth time so that claimed log blocks
4134 		 * don't appear to be from the future.  spa_claim_max_txg
4135 		 * will have been set for us by ZIL traversal operations
4136 		 * performed above.
4137 		 */
4138 		txg_wait_synced(spa->spa_dsl_pool, spa->spa_claim_max_txg);
4139 
4140 		/*
4141 		 * Check if we need to request an update of the config. On the
4142 		 * next sync, we would update the config stored in vdev labels
4143 		 * and the cachefile (by default /etc/zfs/zpool.cache).
4144 		 */
4145 		spa_ld_check_for_config_update(spa, config_cache_txg,
4146 		    update_config_cache);
4147 
4148 		/*
4149 		 * Check all DTLs to see if anything needs resilvering.
4150 		 */
4151 		if (!dsl_scan_resilvering(spa->spa_dsl_pool) &&
4152 		    vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL))
4153 			spa_async_request(spa, SPA_ASYNC_RESILVER);
4154 
4155 		/*
4156 		 * Log the fact that we booted up (so that we can detect if
4157 		 * we rebooted in the middle of an operation).
4158 		 */
4159 		spa_history_log_version(spa, "open");
4160 
4161 		spa_restart_removal(spa);
4162 		spa_spawn_aux_threads(spa);
4163 
4164 		/*
4165 		 * Delete any inconsistent datasets.
4166 		 *
4167 		 * Note:
4168 		 * Since we may be issuing deletes for clones here,
4169 		 * we make sure to do so after we've spawned all the
4170 		 * auxiliary threads above (from which the livelist
4171 		 * deletion zthr is part of).
4172 		 */
4173 		(void) dmu_objset_find(spa_name(spa),
4174 		    dsl_destroy_inconsistent, NULL, DS_FIND_CHILDREN);
4175 
4176 		/*
4177 		 * Clean up any stale temporary dataset userrefs.
4178 		 */
4179 		dsl_pool_clean_tmp_userrefs(spa->spa_dsl_pool);
4180 
4181 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
4182 		vdev_initialize_restart(spa->spa_root_vdev);
4183 		spa_config_exit(spa, SCL_CONFIG, FTAG);
4184 	}
4185 
4186 	spa_load_note(spa, "LOADED");
4187 
4188 	return (0);
4189 }
4190 
4191 static int
4192 spa_load_retry(spa_t *spa, spa_load_state_t state)
4193 {
4194 	int mode = spa->spa_mode;
4195 
4196 	spa_unload(spa);
4197 	spa_deactivate(spa);
4198 
4199 	spa->spa_load_max_txg = spa->spa_uberblock.ub_txg - 1;
4200 
4201 	spa_activate(spa, mode);
4202 	spa_async_suspend(spa);
4203 
4204 	spa_load_note(spa, "spa_load_retry: rewind, max txg: %llu",
4205 	    (u_longlong_t)spa->spa_load_max_txg);
4206 
4207 	return (spa_load(spa, state, SPA_IMPORT_EXISTING));
4208 }
4209 
4210 /*
4211  * If spa_load() fails this function will try loading prior txg's. If
4212  * 'state' is SPA_LOAD_RECOVER and one of these loads succeeds the pool
4213  * will be rewound to that txg. If 'state' is not SPA_LOAD_RECOVER this
4214  * function will not rewind the pool and will return the same error as
4215  * spa_load().
4216  */
4217 static int
4218 spa_load_best(spa_t *spa, spa_load_state_t state, uint64_t max_request,
4219     int rewind_flags)
4220 {
4221 	nvlist_t *loadinfo = NULL;
4222 	nvlist_t *config = NULL;
4223 	int load_error, rewind_error;
4224 	uint64_t safe_rewind_txg;
4225 	uint64_t min_txg;
4226 
4227 	if (spa->spa_load_txg && state == SPA_LOAD_RECOVER) {
4228 		spa->spa_load_max_txg = spa->spa_load_txg;
4229 		spa_set_log_state(spa, SPA_LOG_CLEAR);
4230 	} else {
4231 		spa->spa_load_max_txg = max_request;
4232 		if (max_request != UINT64_MAX)
4233 			spa->spa_extreme_rewind = B_TRUE;
4234 	}
4235 
4236 	load_error = rewind_error = spa_load(spa, state, SPA_IMPORT_EXISTING);
4237 	if (load_error == 0)
4238 		return (0);
4239 	if (load_error == ZFS_ERR_NO_CHECKPOINT) {
4240 		/*
4241 		 * When attempting checkpoint-rewind on a pool with no
4242 		 * checkpoint, we should not attempt to load uberblocks
4243 		 * from previous txgs when spa_load fails.
4244 		 */
4245 		ASSERT(spa->spa_import_flags & ZFS_IMPORT_CHECKPOINT);
4246 		return (load_error);
4247 	}
4248 
4249 	if (spa->spa_root_vdev != NULL)
4250 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
4251 
4252 	spa->spa_last_ubsync_txg = spa->spa_uberblock.ub_txg;
4253 	spa->spa_last_ubsync_txg_ts = spa->spa_uberblock.ub_timestamp;
4254 
4255 	if (rewind_flags & ZPOOL_NEVER_REWIND) {
4256 		nvlist_free(config);
4257 		return (load_error);
4258 	}
4259 
4260 	if (state == SPA_LOAD_RECOVER) {
4261 		/* Price of rolling back is discarding txgs, including log */
4262 		spa_set_log_state(spa, SPA_LOG_CLEAR);
4263 	} else {
4264 		/*
4265 		 * If we aren't rolling back save the load info from our first
4266 		 * import attempt so that we can restore it after attempting
4267 		 * to rewind.
4268 		 */
4269 		loadinfo = spa->spa_load_info;
4270 		spa->spa_load_info = fnvlist_alloc();
4271 	}
4272 
4273 	spa->spa_load_max_txg = spa->spa_last_ubsync_txg;
4274 	safe_rewind_txg = spa->spa_last_ubsync_txg - TXG_DEFER_SIZE;
4275 	min_txg = (rewind_flags & ZPOOL_EXTREME_REWIND) ?
4276 	    TXG_INITIAL : safe_rewind_txg;
4277 
4278 	/*
4279 	 * Continue as long as we're finding errors, we're still within
4280 	 * the acceptable rewind range, and we're still finding uberblocks
4281 	 */
4282 	while (rewind_error && spa->spa_uberblock.ub_txg >= min_txg &&
4283 	    spa->spa_uberblock.ub_txg <= spa->spa_load_max_txg) {
4284 		if (spa->spa_load_max_txg < safe_rewind_txg)
4285 			spa->spa_extreme_rewind = B_TRUE;
4286 		rewind_error = spa_load_retry(spa, state);
4287 	}
4288 
4289 	spa->spa_extreme_rewind = B_FALSE;
4290 	spa->spa_load_max_txg = UINT64_MAX;
4291 
4292 	if (config && (rewind_error || state != SPA_LOAD_RECOVER))
4293 		spa_config_set(spa, config);
4294 	else
4295 		nvlist_free(config);
4296 
4297 	if (state == SPA_LOAD_RECOVER) {
4298 		ASSERT3P(loadinfo, ==, NULL);
4299 		return (rewind_error);
4300 	} else {
4301 		/* Store the rewind info as part of the initial load info */
4302 		fnvlist_add_nvlist(loadinfo, ZPOOL_CONFIG_REWIND_INFO,
4303 		    spa->spa_load_info);
4304 
4305 		/* Restore the initial load info */
4306 		fnvlist_free(spa->spa_load_info);
4307 		spa->spa_load_info = loadinfo;
4308 
4309 		return (load_error);
4310 	}
4311 }
4312 
4313 /*
4314  * Pool Open/Import
4315  *
4316  * The import case is identical to an open except that the configuration is sent
4317  * down from userland, instead of grabbed from the configuration cache.  For the
4318  * case of an open, the pool configuration will exist in the
4319  * POOL_STATE_UNINITIALIZED state.
4320  *
4321  * The stats information (gen/count/ustats) is used to gather vdev statistics at
4322  * the same time open the pool, without having to keep around the spa_t in some
4323  * ambiguous state.
4324  */
4325 static int
4326 spa_open_common(const char *pool, spa_t **spapp, void *tag, nvlist_t *nvpolicy,
4327     nvlist_t **config)
4328 {
4329 	spa_t *spa;
4330 	spa_load_state_t state = SPA_LOAD_OPEN;
4331 	int error;
4332 	int locked = B_FALSE;
4333 
4334 	*spapp = NULL;
4335 
4336 	/*
4337 	 * As disgusting as this is, we need to support recursive calls to this
4338 	 * function because dsl_dir_open() is called during spa_load(), and ends
4339 	 * up calling spa_open() again.  The real fix is to figure out how to
4340 	 * avoid dsl_dir_open() calling this in the first place.
4341 	 */
4342 	if (mutex_owner(&spa_namespace_lock) != curthread) {
4343 		mutex_enter(&spa_namespace_lock);
4344 		locked = B_TRUE;
4345 	}
4346 
4347 	if ((spa = spa_lookup(pool)) == NULL) {
4348 		if (locked)
4349 			mutex_exit(&spa_namespace_lock);
4350 		return (SET_ERROR(ENOENT));
4351 	}
4352 
4353 	if (spa->spa_state == POOL_STATE_UNINITIALIZED) {
4354 		zpool_load_policy_t policy;
4355 
4356 		zpool_get_load_policy(nvpolicy ? nvpolicy : spa->spa_config,
4357 		    &policy);
4358 		if (policy.zlp_rewind & ZPOOL_DO_REWIND)
4359 			state = SPA_LOAD_RECOVER;
4360 
4361 		spa_activate(spa, spa_mode_global);
4362 
4363 		if (state != SPA_LOAD_RECOVER)
4364 			spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
4365 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
4366 
4367 		zfs_dbgmsg("spa_open_common: opening %s", pool);
4368 		error = spa_load_best(spa, state, policy.zlp_txg,
4369 		    policy.zlp_rewind);
4370 
4371 		if (error == EBADF) {
4372 			/*
4373 			 * If vdev_validate() returns failure (indicated by
4374 			 * EBADF), it indicates that one of the vdevs indicates
4375 			 * that the pool has been exported or destroyed.  If
4376 			 * this is the case, the config cache is out of sync and
4377 			 * we should remove the pool from the namespace.
4378 			 */
4379 			spa_unload(spa);
4380 			spa_deactivate(spa);
4381 			spa_write_cachefile(spa, B_TRUE, B_TRUE);
4382 			spa_remove(spa);
4383 			if (locked)
4384 				mutex_exit(&spa_namespace_lock);
4385 			return (SET_ERROR(ENOENT));
4386 		}
4387 
4388 		if (error) {
4389 			/*
4390 			 * We can't open the pool, but we still have useful
4391 			 * information: the state of each vdev after the
4392 			 * attempted vdev_open().  Return this to the user.
4393 			 */
4394 			if (config != NULL && spa->spa_config) {
4395 				VERIFY(nvlist_dup(spa->spa_config, config,
4396 				    KM_SLEEP) == 0);
4397 				VERIFY(nvlist_add_nvlist(*config,
4398 				    ZPOOL_CONFIG_LOAD_INFO,
4399 				    spa->spa_load_info) == 0);
4400 			}
4401 			spa_unload(spa);
4402 			spa_deactivate(spa);
4403 			spa->spa_last_open_failed = error;
4404 			if (locked)
4405 				mutex_exit(&spa_namespace_lock);
4406 			*spapp = NULL;
4407 			return (error);
4408 		}
4409 	}
4410 
4411 	spa_open_ref(spa, tag);
4412 
4413 	if (config != NULL)
4414 		*config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
4415 
4416 	/*
4417 	 * If we've recovered the pool, pass back any information we
4418 	 * gathered while doing the load.
4419 	 */
4420 	if (state == SPA_LOAD_RECOVER) {
4421 		VERIFY(nvlist_add_nvlist(*config, ZPOOL_CONFIG_LOAD_INFO,
4422 		    spa->spa_load_info) == 0);
4423 	}
4424 
4425 	if (locked) {
4426 		spa->spa_last_open_failed = 0;
4427 		spa->spa_last_ubsync_txg = 0;
4428 		spa->spa_load_txg = 0;
4429 		mutex_exit(&spa_namespace_lock);
4430 	}
4431 
4432 	*spapp = spa;
4433 
4434 	return (0);
4435 }
4436 
4437 int
4438 spa_open_rewind(const char *name, spa_t **spapp, void *tag, nvlist_t *policy,
4439     nvlist_t **config)
4440 {
4441 	return (spa_open_common(name, spapp, tag, policy, config));
4442 }
4443 
4444 int
4445 spa_open(const char *name, spa_t **spapp, void *tag)
4446 {
4447 	return (spa_open_common(name, spapp, tag, NULL, NULL));
4448 }
4449 
4450 /*
4451  * Lookup the given spa_t, incrementing the inject count in the process,
4452  * preventing it from being exported or destroyed.
4453  */
4454 spa_t *
4455 spa_inject_addref(char *name)
4456 {
4457 	spa_t *spa;
4458 
4459 	mutex_enter(&spa_namespace_lock);
4460 	if ((spa = spa_lookup(name)) == NULL) {
4461 		mutex_exit(&spa_namespace_lock);
4462 		return (NULL);
4463 	}
4464 	spa->spa_inject_ref++;
4465 	mutex_exit(&spa_namespace_lock);
4466 
4467 	return (spa);
4468 }
4469 
4470 void
4471 spa_inject_delref(spa_t *spa)
4472 {
4473 	mutex_enter(&spa_namespace_lock);
4474 	spa->spa_inject_ref--;
4475 	mutex_exit(&spa_namespace_lock);
4476 }
4477 
4478 /*
4479  * Add spares device information to the nvlist.
4480  */
4481 static void
4482 spa_add_spares(spa_t *spa, nvlist_t *config)
4483 {
4484 	nvlist_t **spares;
4485 	uint_t i, nspares;
4486 	nvlist_t *nvroot;
4487 	uint64_t guid;
4488 	vdev_stat_t *vs;
4489 	uint_t vsc;
4490 	uint64_t pool;
4491 
4492 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
4493 
4494 	if (spa->spa_spares.sav_count == 0)
4495 		return;
4496 
4497 	VERIFY(nvlist_lookup_nvlist(config,
4498 	    ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0);
4499 	VERIFY(nvlist_lookup_nvlist_array(spa->spa_spares.sav_config,
4500 	    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0);
4501 	if (nspares != 0) {
4502 		VERIFY(nvlist_add_nvlist_array(nvroot,
4503 		    ZPOOL_CONFIG_SPARES, spares, nspares) == 0);
4504 		VERIFY(nvlist_lookup_nvlist_array(nvroot,
4505 		    ZPOOL_CONFIG_SPARES, &spares, &nspares) == 0);
4506 
4507 		/*
4508 		 * Go through and find any spares which have since been
4509 		 * repurposed as an active spare.  If this is the case, update
4510 		 * their status appropriately.
4511 		 */
4512 		for (i = 0; i < nspares; i++) {
4513 			VERIFY(nvlist_lookup_uint64(spares[i],
4514 			    ZPOOL_CONFIG_GUID, &guid) == 0);
4515 			if (spa_spare_exists(guid, &pool, NULL) &&
4516 			    pool != 0ULL) {
4517 				VERIFY(nvlist_lookup_uint64_array(
4518 				    spares[i], ZPOOL_CONFIG_VDEV_STATS,
4519 				    (uint64_t **)&vs, &vsc) == 0);
4520 				vs->vs_state = VDEV_STATE_CANT_OPEN;
4521 				vs->vs_aux = VDEV_AUX_SPARED;
4522 			}
4523 		}
4524 	}
4525 }
4526 
4527 /*
4528  * Add l2cache device information to the nvlist, including vdev stats.
4529  */
4530 static void
4531 spa_add_l2cache(spa_t *spa, nvlist_t *config)
4532 {
4533 	nvlist_t **l2cache;
4534 	uint_t i, j, nl2cache;
4535 	nvlist_t *nvroot;
4536 	uint64_t guid;
4537 	vdev_t *vd;
4538 	vdev_stat_t *vs;
4539 	uint_t vsc;
4540 
4541 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
4542 
4543 	if (spa->spa_l2cache.sav_count == 0)
4544 		return;
4545 
4546 	VERIFY(nvlist_lookup_nvlist(config,
4547 	    ZPOOL_CONFIG_VDEV_TREE, &nvroot) == 0);
4548 	VERIFY(nvlist_lookup_nvlist_array(spa->spa_l2cache.sav_config,
4549 	    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0);
4550 	if (nl2cache != 0) {
4551 		VERIFY(nvlist_add_nvlist_array(nvroot,
4552 		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0);
4553 		VERIFY(nvlist_lookup_nvlist_array(nvroot,
4554 		    ZPOOL_CONFIG_L2CACHE, &l2cache, &nl2cache) == 0);
4555 
4556 		/*
4557 		 * Update level 2 cache device stats.
4558 		 */
4559 
4560 		for (i = 0; i < nl2cache; i++) {
4561 			VERIFY(nvlist_lookup_uint64(l2cache[i],
4562 			    ZPOOL_CONFIG_GUID, &guid) == 0);
4563 
4564 			vd = NULL;
4565 			for (j = 0; j < spa->spa_l2cache.sav_count; j++) {
4566 				if (guid ==
4567 				    spa->spa_l2cache.sav_vdevs[j]->vdev_guid) {
4568 					vd = spa->spa_l2cache.sav_vdevs[j];
4569 					break;
4570 				}
4571 			}
4572 			ASSERT(vd != NULL);
4573 
4574 			VERIFY(nvlist_lookup_uint64_array(l2cache[i],
4575 			    ZPOOL_CONFIG_VDEV_STATS, (uint64_t **)&vs, &vsc)
4576 			    == 0);
4577 			vdev_get_stats(vd, vs);
4578 		}
4579 	}
4580 }
4581 
4582 static void
4583 spa_add_feature_stats(spa_t *spa, nvlist_t *config)
4584 {
4585 	nvlist_t *features;
4586 	zap_cursor_t zc;
4587 	zap_attribute_t za;
4588 
4589 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_READER));
4590 	VERIFY(nvlist_alloc(&features, NV_UNIQUE_NAME, KM_SLEEP) == 0);
4591 
4592 	if (spa->spa_feat_for_read_obj != 0) {
4593 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
4594 		    spa->spa_feat_for_read_obj);
4595 		    zap_cursor_retrieve(&zc, &za) == 0;
4596 		    zap_cursor_advance(&zc)) {
4597 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
4598 			    za.za_num_integers == 1);
4599 			VERIFY3U(0, ==, nvlist_add_uint64(features, za.za_name,
4600 			    za.za_first_integer));
4601 		}
4602 		zap_cursor_fini(&zc);
4603 	}
4604 
4605 	if (spa->spa_feat_for_write_obj != 0) {
4606 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
4607 		    spa->spa_feat_for_write_obj);
4608 		    zap_cursor_retrieve(&zc, &za) == 0;
4609 		    zap_cursor_advance(&zc)) {
4610 			ASSERT(za.za_integer_length == sizeof (uint64_t) &&
4611 			    za.za_num_integers == 1);
4612 			VERIFY3U(0, ==, nvlist_add_uint64(features, za.za_name,
4613 			    za.za_first_integer));
4614 		}
4615 		zap_cursor_fini(&zc);
4616 	}
4617 
4618 	VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_FEATURE_STATS,
4619 	    features) == 0);
4620 	nvlist_free(features);
4621 }
4622 
4623 int
4624 spa_get_stats(const char *name, nvlist_t **config,
4625     char *altroot, size_t buflen)
4626 {
4627 	int error;
4628 	spa_t *spa;
4629 
4630 	*config = NULL;
4631 	error = spa_open_common(name, &spa, FTAG, NULL, config);
4632 
4633 	if (spa != NULL) {
4634 		/*
4635 		 * This still leaves a window of inconsistency where the spares
4636 		 * or l2cache devices could change and the config would be
4637 		 * self-inconsistent.
4638 		 */
4639 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
4640 
4641 		if (*config != NULL) {
4642 			uint64_t loadtimes[2];
4643 
4644 			loadtimes[0] = spa->spa_loaded_ts.tv_sec;
4645 			loadtimes[1] = spa->spa_loaded_ts.tv_nsec;
4646 			VERIFY(nvlist_add_uint64_array(*config,
4647 			    ZPOOL_CONFIG_LOADED_TIME, loadtimes, 2) == 0);
4648 
4649 			VERIFY(nvlist_add_uint64(*config,
4650 			    ZPOOL_CONFIG_ERRCOUNT,
4651 			    spa_get_errlog_size(spa)) == 0);
4652 
4653 			if (spa_suspended(spa)) {
4654 				VERIFY(nvlist_add_uint64(*config,
4655 				    ZPOOL_CONFIG_SUSPENDED,
4656 				    spa->spa_failmode) == 0);
4657 				VERIFY(nvlist_add_uint64(*config,
4658 				    ZPOOL_CONFIG_SUSPENDED_REASON,
4659 				    spa->spa_suspended) == 0);
4660 			}
4661 
4662 			spa_add_spares(spa, *config);
4663 			spa_add_l2cache(spa, *config);
4664 			spa_add_feature_stats(spa, *config);
4665 		}
4666 	}
4667 
4668 	/*
4669 	 * We want to get the alternate root even for faulted pools, so we cheat
4670 	 * and call spa_lookup() directly.
4671 	 */
4672 	if (altroot) {
4673 		if (spa == NULL) {
4674 			mutex_enter(&spa_namespace_lock);
4675 			spa = spa_lookup(name);
4676 			if (spa)
4677 				spa_altroot(spa, altroot, buflen);
4678 			else
4679 				altroot[0] = '\0';
4680 			spa = NULL;
4681 			mutex_exit(&spa_namespace_lock);
4682 		} else {
4683 			spa_altroot(spa, altroot, buflen);
4684 		}
4685 	}
4686 
4687 	if (spa != NULL) {
4688 		spa_config_exit(spa, SCL_CONFIG, FTAG);
4689 		spa_close(spa, FTAG);
4690 	}
4691 
4692 	return (error);
4693 }
4694 
4695 /*
4696  * Validate that the auxiliary device array is well formed.  We must have an
4697  * array of nvlists, each which describes a valid leaf vdev.  If this is an
4698  * import (mode is VDEV_ALLOC_SPARE), then we allow corrupted spares to be
4699  * specified, as long as they are well-formed.
4700  */
4701 static int
4702 spa_validate_aux_devs(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode,
4703     spa_aux_vdev_t *sav, const char *config, uint64_t version,
4704     vdev_labeltype_t label)
4705 {
4706 	nvlist_t **dev;
4707 	uint_t i, ndev;
4708 	vdev_t *vd;
4709 	int error;
4710 
4711 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
4712 
4713 	/*
4714 	 * It's acceptable to have no devs specified.
4715 	 */
4716 	if (nvlist_lookup_nvlist_array(nvroot, config, &dev, &ndev) != 0)
4717 		return (0);
4718 
4719 	if (ndev == 0)
4720 		return (SET_ERROR(EINVAL));
4721 
4722 	/*
4723 	 * Make sure the pool is formatted with a version that supports this
4724 	 * device type.
4725 	 */
4726 	if (spa_version(spa) < version)
4727 		return (SET_ERROR(ENOTSUP));
4728 
4729 	/*
4730 	 * Set the pending device list so we correctly handle device in-use
4731 	 * checking.
4732 	 */
4733 	sav->sav_pending = dev;
4734 	sav->sav_npending = ndev;
4735 
4736 	for (i = 0; i < ndev; i++) {
4737 		if ((error = spa_config_parse(spa, &vd, dev[i], NULL, 0,
4738 		    mode)) != 0)
4739 			goto out;
4740 
4741 		if (!vd->vdev_ops->vdev_op_leaf) {
4742 			vdev_free(vd);
4743 			error = SET_ERROR(EINVAL);
4744 			goto out;
4745 		}
4746 
4747 		vd->vdev_top = vd;
4748 
4749 		if ((error = vdev_open(vd)) == 0 &&
4750 		    (error = vdev_label_init(vd, crtxg, label)) == 0) {
4751 			VERIFY(nvlist_add_uint64(dev[i], ZPOOL_CONFIG_GUID,
4752 			    vd->vdev_guid) == 0);
4753 		}
4754 
4755 		vdev_free(vd);
4756 
4757 		if (error &&
4758 		    (mode != VDEV_ALLOC_SPARE && mode != VDEV_ALLOC_L2CACHE))
4759 			goto out;
4760 		else
4761 			error = 0;
4762 	}
4763 
4764 out:
4765 	sav->sav_pending = NULL;
4766 	sav->sav_npending = 0;
4767 	return (error);
4768 }
4769 
4770 static int
4771 spa_validate_aux(spa_t *spa, nvlist_t *nvroot, uint64_t crtxg, int mode)
4772 {
4773 	int error;
4774 
4775 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
4776 
4777 	if ((error = spa_validate_aux_devs(spa, nvroot, crtxg, mode,
4778 	    &spa->spa_spares, ZPOOL_CONFIG_SPARES, SPA_VERSION_SPARES,
4779 	    VDEV_LABEL_SPARE)) != 0) {
4780 		return (error);
4781 	}
4782 
4783 	return (spa_validate_aux_devs(spa, nvroot, crtxg, mode,
4784 	    &spa->spa_l2cache, ZPOOL_CONFIG_L2CACHE, SPA_VERSION_L2CACHE,
4785 	    VDEV_LABEL_L2CACHE));
4786 }
4787 
4788 static void
4789 spa_set_aux_vdevs(spa_aux_vdev_t *sav, nvlist_t **devs, int ndevs,
4790     const char *config)
4791 {
4792 	int i;
4793 
4794 	if (sav->sav_config != NULL) {
4795 		nvlist_t **olddevs;
4796 		uint_t oldndevs;
4797 		nvlist_t **newdevs;
4798 
4799 		/*
4800 		 * Generate new dev list by concatentating with the
4801 		 * current dev list.
4802 		 */
4803 		VERIFY(nvlist_lookup_nvlist_array(sav->sav_config, config,
4804 		    &olddevs, &oldndevs) == 0);
4805 
4806 		newdevs = kmem_alloc(sizeof (void *) *
4807 		    (ndevs + oldndevs), KM_SLEEP);
4808 		for (i = 0; i < oldndevs; i++)
4809 			VERIFY(nvlist_dup(olddevs[i], &newdevs[i],
4810 			    KM_SLEEP) == 0);
4811 		for (i = 0; i < ndevs; i++)
4812 			VERIFY(nvlist_dup(devs[i], &newdevs[i + oldndevs],
4813 			    KM_SLEEP) == 0);
4814 
4815 		VERIFY(nvlist_remove(sav->sav_config, config,
4816 		    DATA_TYPE_NVLIST_ARRAY) == 0);
4817 
4818 		VERIFY(nvlist_add_nvlist_array(sav->sav_config,
4819 		    config, newdevs, ndevs + oldndevs) == 0);
4820 		for (i = 0; i < oldndevs + ndevs; i++)
4821 			nvlist_free(newdevs[i]);
4822 		kmem_free(newdevs, (oldndevs + ndevs) * sizeof (void *));
4823 	} else {
4824 		/*
4825 		 * Generate a new dev list.
4826 		 */
4827 		VERIFY(nvlist_alloc(&sav->sav_config, NV_UNIQUE_NAME,
4828 		    KM_SLEEP) == 0);
4829 		VERIFY(nvlist_add_nvlist_array(sav->sav_config, config,
4830 		    devs, ndevs) == 0);
4831 	}
4832 }
4833 
4834 /*
4835  * Stop and drop level 2 ARC devices
4836  */
4837 void
4838 spa_l2cache_drop(spa_t *spa)
4839 {
4840 	vdev_t *vd;
4841 	int i;
4842 	spa_aux_vdev_t *sav = &spa->spa_l2cache;
4843 
4844 	for (i = 0; i < sav->sav_count; i++) {
4845 		uint64_t pool;
4846 
4847 		vd = sav->sav_vdevs[i];
4848 		ASSERT(vd != NULL);
4849 
4850 		if (spa_l2cache_exists(vd->vdev_guid, &pool) &&
4851 		    pool != 0ULL && l2arc_vdev_present(vd))
4852 			l2arc_remove_vdev(vd);
4853 	}
4854 }
4855 
4856 /*
4857  * Verify encryption parameters for spa creation. If we are encrypting, we must
4858  * have the encryption feature flag enabled.
4859  */
4860 static int
4861 spa_create_check_encryption_params(dsl_crypto_params_t *dcp,
4862     boolean_t has_encryption)
4863 {
4864 	if (dcp->cp_crypt != ZIO_CRYPT_OFF &&
4865 	    dcp->cp_crypt != ZIO_CRYPT_INHERIT &&
4866 	    !has_encryption)
4867 		return (SET_ERROR(ENOTSUP));
4868 
4869 	return (dmu_objset_create_crypt_check(NULL, dcp, NULL));
4870 }
4871 
4872 /*
4873  * Pool Creation
4874  */
4875 int
4876 spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
4877     nvlist_t *zplprops, dsl_crypto_params_t *dcp)
4878 {
4879 	spa_t *spa;
4880 	char *altroot = NULL;
4881 	vdev_t *rvd;
4882 	dsl_pool_t *dp;
4883 	dmu_tx_t *tx;
4884 	int error = 0;
4885 	uint64_t txg = TXG_INITIAL;
4886 	nvlist_t **spares, **l2cache;
4887 	uint_t nspares, nl2cache;
4888 	uint64_t version, obj;
4889 	boolean_t has_features;
4890 	char *poolname;
4891 	nvlist_t *nvl;
4892 	boolean_t has_encryption;
4893 	spa_feature_t feat;
4894 	char *feat_name;
4895 
4896 	if (props == NULL ||
4897 	    nvlist_lookup_string(props,
4898 	    zpool_prop_to_name(ZPOOL_PROP_TNAME), &poolname) != 0)
4899 		poolname = (char *)pool;
4900 
4901 	/*
4902 	 * If this pool already exists, return failure.
4903 	 */
4904 	mutex_enter(&spa_namespace_lock);
4905 	if (spa_lookup(poolname) != NULL) {
4906 		mutex_exit(&spa_namespace_lock);
4907 		return (SET_ERROR(EEXIST));
4908 	}
4909 
4910 	/*
4911 	 * Allocate a new spa_t structure.
4912 	 */
4913 	nvl = fnvlist_alloc();
4914 	fnvlist_add_string(nvl, ZPOOL_CONFIG_POOL_NAME, pool);
4915 	(void) nvlist_lookup_string(props,
4916 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
4917 	spa = spa_add(poolname, nvl, altroot);
4918 	fnvlist_free(nvl);
4919 	spa_activate(spa, spa_mode_global);
4920 
4921 	if (props && (error = spa_prop_validate(spa, props))) {
4922 		spa_deactivate(spa);
4923 		spa_remove(spa);
4924 		mutex_exit(&spa_namespace_lock);
4925 		return (error);
4926 	}
4927 
4928 	/*
4929 	 * Temporary pool names should never be written to disk.
4930 	 */
4931 	if (poolname != pool)
4932 		spa->spa_import_flags |= ZFS_IMPORT_TEMP_NAME;
4933 
4934 	has_features = B_FALSE;
4935 	has_encryption = B_FALSE;
4936 	for (nvpair_t *elem = nvlist_next_nvpair(props, NULL);
4937 	    elem != NULL; elem = nvlist_next_nvpair(props, elem)) {
4938 		if (zpool_prop_feature(nvpair_name(elem))) {
4939 			has_features = B_TRUE;
4940 			feat_name = strchr(nvpair_name(elem), '@') + 1;
4941 			VERIFY0(zfeature_lookup_name(feat_name, &feat));
4942 			if (feat == SPA_FEATURE_ENCRYPTION)
4943 				has_encryption = B_TRUE;
4944 		}
4945 	}
4946 
4947 	/* verify encryption params, if they were provided */
4948 	if (dcp != NULL) {
4949 		error = spa_create_check_encryption_params(dcp, has_encryption);
4950 		if (error != 0) {
4951 			spa_deactivate(spa);
4952 			spa_remove(spa);
4953 			mutex_exit(&spa_namespace_lock);
4954 			return (error);
4955 		}
4956 	}
4957 
4958 	if (has_features || nvlist_lookup_uint64(props,
4959 	    zpool_prop_to_name(ZPOOL_PROP_VERSION), &version) != 0) {
4960 		version = SPA_VERSION;
4961 	}
4962 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
4963 
4964 	spa->spa_first_txg = txg;
4965 	spa->spa_uberblock.ub_txg = txg - 1;
4966 	spa->spa_uberblock.ub_version = version;
4967 	spa->spa_ubsync = spa->spa_uberblock;
4968 	spa->spa_load_state = SPA_LOAD_CREATE;
4969 	spa->spa_removing_phys.sr_state = DSS_NONE;
4970 	spa->spa_removing_phys.sr_removing_vdev = -1;
4971 	spa->spa_removing_phys.sr_prev_indirect_vdev = -1;
4972 	spa->spa_indirect_vdevs_loaded = B_TRUE;
4973 
4974 	/*
4975 	 * Create "The Godfather" zio to hold all async IOs
4976 	 */
4977 	spa->spa_async_zio_root = kmem_alloc(max_ncpus * sizeof (void *),
4978 	    KM_SLEEP);
4979 	for (int i = 0; i < max_ncpus; i++) {
4980 		spa->spa_async_zio_root[i] = zio_root(spa, NULL, NULL,
4981 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE |
4982 		    ZIO_FLAG_GODFATHER);
4983 	}
4984 
4985 	/*
4986 	 * Create the root vdev.
4987 	 */
4988 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
4989 
4990 	error = spa_config_parse(spa, &rvd, nvroot, NULL, 0, VDEV_ALLOC_ADD);
4991 
4992 	ASSERT(error != 0 || rvd != NULL);
4993 	ASSERT(error != 0 || spa->spa_root_vdev == rvd);
4994 
4995 	if (error == 0 && !zfs_allocatable_devs(nvroot))
4996 		error = SET_ERROR(EINVAL);
4997 
4998 	if (error == 0 &&
4999 	    (error = vdev_create(rvd, txg, B_FALSE)) == 0 &&
5000 	    (error = spa_validate_aux(spa, nvroot, txg,
5001 	    VDEV_ALLOC_ADD)) == 0) {
5002 		/*
5003 		 * instantiate the metaslab groups (this will dirty the vdevs)
5004 		 * we can no longer error exit past this point
5005 		 */
5006 		for (int c = 0; error == 0 && c < rvd->vdev_children; c++) {
5007 			vdev_t *vd = rvd->vdev_child[c];
5008 
5009 			vdev_metaslab_set_size(vd);
5010 			vdev_expand(vd, txg);
5011 		}
5012 	}
5013 
5014 	spa_config_exit(spa, SCL_ALL, FTAG);
5015 
5016 	if (error != 0) {
5017 		spa_unload(spa);
5018 		spa_deactivate(spa);
5019 		spa_remove(spa);
5020 		mutex_exit(&spa_namespace_lock);
5021 		return (error);
5022 	}
5023 
5024 	/*
5025 	 * Get the list of spares, if specified.
5026 	 */
5027 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5028 	    &spares, &nspares) == 0) {
5029 		VERIFY(nvlist_alloc(&spa->spa_spares.sav_config, NV_UNIQUE_NAME,
5030 		    KM_SLEEP) == 0);
5031 		VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config,
5032 		    ZPOOL_CONFIG_SPARES, spares, nspares) == 0);
5033 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5034 		spa_load_spares(spa);
5035 		spa_config_exit(spa, SCL_ALL, FTAG);
5036 		spa->spa_spares.sav_sync = B_TRUE;
5037 	}
5038 
5039 	/*
5040 	 * Get the list of level 2 cache devices, if specified.
5041 	 */
5042 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5043 	    &l2cache, &nl2cache) == 0) {
5044 		VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config,
5045 		    NV_UNIQUE_NAME, KM_SLEEP) == 0);
5046 		VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
5047 		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0);
5048 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5049 		spa_load_l2cache(spa);
5050 		spa_config_exit(spa, SCL_ALL, FTAG);
5051 		spa->spa_l2cache.sav_sync = B_TRUE;
5052 	}
5053 
5054 	spa->spa_is_initializing = B_TRUE;
5055 	spa->spa_dsl_pool = dp = dsl_pool_create(spa, zplprops, dcp, txg);
5056 	spa->spa_is_initializing = B_FALSE;
5057 
5058 	/*
5059 	 * Create DDTs (dedup tables).
5060 	 */
5061 	ddt_create(spa);
5062 
5063 	spa_update_dspace(spa);
5064 
5065 	tx = dmu_tx_create_assigned(dp, txg);
5066 
5067 	/*
5068 	 * Create the pool config object.
5069 	 */
5070 	spa->spa_config_object = dmu_object_alloc(spa->spa_meta_objset,
5071 	    DMU_OT_PACKED_NVLIST, SPA_CONFIG_BLOCKSIZE,
5072 	    DMU_OT_PACKED_NVLIST_SIZE, sizeof (uint64_t), tx);
5073 
5074 	if (zap_add(spa->spa_meta_objset,
5075 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CONFIG,
5076 	    sizeof (uint64_t), 1, &spa->spa_config_object, tx) != 0) {
5077 		cmn_err(CE_PANIC, "failed to add pool config");
5078 	}
5079 
5080 	if (zap_add(spa->spa_meta_objset,
5081 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CREATION_VERSION,
5082 	    sizeof (uint64_t), 1, &version, tx) != 0) {
5083 		cmn_err(CE_PANIC, "failed to add pool version");
5084 	}
5085 
5086 	/* Newly created pools with the right version are always deflated. */
5087 	if (version >= SPA_VERSION_RAIDZ_DEFLATE) {
5088 		spa->spa_deflate = TRUE;
5089 		if (zap_add(spa->spa_meta_objset,
5090 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
5091 		    sizeof (uint64_t), 1, &spa->spa_deflate, tx) != 0) {
5092 			cmn_err(CE_PANIC, "failed to add deflate");
5093 		}
5094 	}
5095 
5096 	/*
5097 	 * Create the deferred-free bpobj.  Turn off compression
5098 	 * because sync-to-convergence takes longer if the blocksize
5099 	 * keeps changing.
5100 	 */
5101 	obj = bpobj_alloc(spa->spa_meta_objset, 1 << 14, tx);
5102 	dmu_object_set_compress(spa->spa_meta_objset, obj,
5103 	    ZIO_COMPRESS_OFF, tx);
5104 	if (zap_add(spa->spa_meta_objset,
5105 	    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_SYNC_BPOBJ,
5106 	    sizeof (uint64_t), 1, &obj, tx) != 0) {
5107 		cmn_err(CE_PANIC, "failed to add bpobj");
5108 	}
5109 	VERIFY3U(0, ==, bpobj_open(&spa->spa_deferred_bpobj,
5110 	    spa->spa_meta_objset, obj));
5111 
5112 	/*
5113 	 * Create the pool's history object.
5114 	 */
5115 	if (version >= SPA_VERSION_ZPOOL_HISTORY)
5116 		spa_history_create_obj(spa, tx);
5117 
5118 	/*
5119 	 * Generate some random noise for salted checksums to operate on.
5120 	 */
5121 	(void) random_get_pseudo_bytes(spa->spa_cksum_salt.zcs_bytes,
5122 	    sizeof (spa->spa_cksum_salt.zcs_bytes));
5123 
5124 	/*
5125 	 * Set pool properties.
5126 	 */
5127 	spa->spa_bootfs = zpool_prop_default_numeric(ZPOOL_PROP_BOOTFS);
5128 	spa->spa_delegation = zpool_prop_default_numeric(ZPOOL_PROP_DELEGATION);
5129 	spa->spa_failmode = zpool_prop_default_numeric(ZPOOL_PROP_FAILUREMODE);
5130 	spa->spa_autoexpand = zpool_prop_default_numeric(ZPOOL_PROP_AUTOEXPAND);
5131 	spa->spa_multihost = zpool_prop_default_numeric(ZPOOL_PROP_MULTIHOST);
5132 
5133 	if (props != NULL) {
5134 		spa_configfile_set(spa, props, B_FALSE);
5135 		spa_sync_props(props, tx);
5136 	}
5137 
5138 	dmu_tx_commit(tx);
5139 
5140 	spa->spa_sync_on = B_TRUE;
5141 	txg_sync_start(spa->spa_dsl_pool);
5142 	mmp_thread_start(spa);
5143 
5144 	/*
5145 	 * We explicitly wait for the first transaction to complete so that our
5146 	 * bean counters are appropriately updated.
5147 	 */
5148 	txg_wait_synced(spa->spa_dsl_pool, txg);
5149 
5150 	spa_spawn_aux_threads(spa);
5151 
5152 	spa_write_cachefile(spa, B_FALSE, B_TRUE);
5153 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_CREATE);
5154 
5155 	spa_history_log_version(spa, "create");
5156 
5157 	/*
5158 	 * Don't count references from objsets that are already closed
5159 	 * and are making their way through the eviction process.
5160 	 */
5161 	spa_evicting_os_wait(spa);
5162 	spa->spa_minref = zfs_refcount_count(&spa->spa_refcount);
5163 	spa->spa_load_state = SPA_LOAD_NONE;
5164 
5165 	mutex_exit(&spa_namespace_lock);
5166 
5167 	return (0);
5168 }
5169 
5170 #ifdef _KERNEL
5171 /*
5172  * Get the root pool information from the root disk, then import the root pool
5173  * during the system boot up time.
5174  */
5175 extern int vdev_disk_read_rootlabel(char *, char *, nvlist_t **);
5176 
5177 static nvlist_t *
5178 spa_generate_rootconf(char *devpath, char *devid, uint64_t *guid)
5179 {
5180 	nvlist_t *config;
5181 	nvlist_t *nvtop, *nvroot;
5182 	uint64_t pgid;
5183 
5184 	if (vdev_disk_read_rootlabel(devpath, devid, &config) != 0)
5185 		return (NULL);
5186 
5187 	/*
5188 	 * Add this top-level vdev to the child array.
5189 	 */
5190 	VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
5191 	    &nvtop) == 0);
5192 	VERIFY(nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_GUID,
5193 	    &pgid) == 0);
5194 	VERIFY(nvlist_lookup_uint64(config, ZPOOL_CONFIG_GUID, guid) == 0);
5195 
5196 	/*
5197 	 * Put this pool's top-level vdevs into a root vdev.
5198 	 */
5199 	VERIFY(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, KM_SLEEP) == 0);
5200 	VERIFY(nvlist_add_string(nvroot, ZPOOL_CONFIG_TYPE,
5201 	    VDEV_TYPE_ROOT) == 0);
5202 	VERIFY(nvlist_add_uint64(nvroot, ZPOOL_CONFIG_ID, 0ULL) == 0);
5203 	VERIFY(nvlist_add_uint64(nvroot, ZPOOL_CONFIG_GUID, pgid) == 0);
5204 	VERIFY(nvlist_add_nvlist_array(nvroot, ZPOOL_CONFIG_CHILDREN,
5205 	    &nvtop, 1) == 0);
5206 
5207 	/*
5208 	 * Replace the existing vdev_tree with the new root vdev in
5209 	 * this pool's configuration (remove the old, add the new).
5210 	 */
5211 	VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, nvroot) == 0);
5212 	nvlist_free(nvroot);
5213 	return (config);
5214 }
5215 
5216 /*
5217  * Walk the vdev tree and see if we can find a device with "better"
5218  * configuration. A configuration is "better" if the label on that
5219  * device has a more recent txg.
5220  */
5221 static void
5222 spa_alt_rootvdev(vdev_t *vd, vdev_t **avd, uint64_t *txg)
5223 {
5224 	for (int c = 0; c < vd->vdev_children; c++)
5225 		spa_alt_rootvdev(vd->vdev_child[c], avd, txg);
5226 
5227 	if (vd->vdev_ops->vdev_op_leaf) {
5228 		nvlist_t *label;
5229 		uint64_t label_txg;
5230 
5231 		if (vdev_disk_read_rootlabel(vd->vdev_physpath, vd->vdev_devid,
5232 		    &label) != 0)
5233 			return;
5234 
5235 		VERIFY(nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_TXG,
5236 		    &label_txg) == 0);
5237 
5238 		/*
5239 		 * Do we have a better boot device?
5240 		 */
5241 		if (label_txg > *txg) {
5242 			*txg = label_txg;
5243 			*avd = vd;
5244 		}
5245 		nvlist_free(label);
5246 	}
5247 }
5248 
5249 /*
5250  * Import a root pool.
5251  *
5252  * For x86. devpath_list will consist of devid and/or physpath name of
5253  * the vdev (e.g. "id1,sd@SSEAGATE..." or "/pci@1f,0/ide@d/disk@0,0:a").
5254  * The GRUB "findroot" command will return the vdev we should boot.
5255  *
5256  * For Sparc, devpath_list consists the physpath name of the booting device
5257  * no matter the rootpool is a single device pool or a mirrored pool.
5258  * e.g.
5259  *	"/pci@1f,0/ide@d/disk@0,0:a"
5260  */
5261 int
5262 spa_import_rootpool(char *devpath, char *devid)
5263 {
5264 	spa_t *spa;
5265 	vdev_t *rvd, *bvd, *avd = NULL;
5266 	nvlist_t *config, *nvtop;
5267 	uint64_t guid, txg;
5268 	char *pname;
5269 	int error;
5270 
5271 	/*
5272 	 * Read the label from the boot device and generate a configuration.
5273 	 */
5274 	config = spa_generate_rootconf(devpath, devid, &guid);
5275 #if defined(_OBP) && defined(_KERNEL)
5276 	if (config == NULL) {
5277 		if (strstr(devpath, "/iscsi/ssd") != NULL) {
5278 			/* iscsi boot */
5279 			get_iscsi_bootpath_phy(devpath);
5280 			config = spa_generate_rootconf(devpath, devid, &guid);
5281 		}
5282 	}
5283 #endif
5284 	if (config == NULL) {
5285 		cmn_err(CE_NOTE, "Cannot read the pool label from '%s'",
5286 		    devpath);
5287 		return (SET_ERROR(EIO));
5288 	}
5289 
5290 	VERIFY(nvlist_lookup_string(config, ZPOOL_CONFIG_POOL_NAME,
5291 	    &pname) == 0);
5292 	VERIFY(nvlist_lookup_uint64(config, ZPOOL_CONFIG_POOL_TXG, &txg) == 0);
5293 
5294 	mutex_enter(&spa_namespace_lock);
5295 	if ((spa = spa_lookup(pname)) != NULL) {
5296 		/*
5297 		 * Remove the existing root pool from the namespace so that we
5298 		 * can replace it with the correct config we just read in.
5299 		 */
5300 		spa_remove(spa);
5301 	}
5302 
5303 	spa = spa_add(pname, config, NULL);
5304 	spa->spa_is_root = B_TRUE;
5305 	spa->spa_import_flags = ZFS_IMPORT_VERBATIM;
5306 	if (nvlist_lookup_uint64(config, ZPOOL_CONFIG_VERSION,
5307 	    &spa->spa_ubsync.ub_version) != 0)
5308 		spa->spa_ubsync.ub_version = SPA_VERSION_INITIAL;
5309 
5310 	/*
5311 	 * Build up a vdev tree based on the boot device's label config.
5312 	 */
5313 	VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
5314 	    &nvtop) == 0);
5315 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5316 	error = spa_config_parse(spa, &rvd, nvtop, NULL, 0,
5317 	    VDEV_ALLOC_ROOTPOOL);
5318 	spa_config_exit(spa, SCL_ALL, FTAG);
5319 	if (error) {
5320 		mutex_exit(&spa_namespace_lock);
5321 		nvlist_free(config);
5322 		cmn_err(CE_NOTE, "Can not parse the config for pool '%s'",
5323 		    pname);
5324 		return (error);
5325 	}
5326 
5327 	/*
5328 	 * Get the boot vdev.
5329 	 */
5330 	if ((bvd = vdev_lookup_by_guid(rvd, guid)) == NULL) {
5331 		cmn_err(CE_NOTE, "Can not find the boot vdev for guid %llu",
5332 		    (u_longlong_t)guid);
5333 		error = SET_ERROR(ENOENT);
5334 		goto out;
5335 	}
5336 
5337 	/*
5338 	 * Determine if there is a better boot device.
5339 	 */
5340 	avd = bvd;
5341 	spa_alt_rootvdev(rvd, &avd, &txg);
5342 	if (avd != bvd) {
5343 		cmn_err(CE_NOTE, "The boot device is 'degraded'. Please "
5344 		    "try booting from '%s'", avd->vdev_path);
5345 		error = SET_ERROR(EINVAL);
5346 		goto out;
5347 	}
5348 
5349 	/*
5350 	 * If the boot device is part of a spare vdev then ensure that
5351 	 * we're booting off the active spare.
5352 	 */
5353 	if (bvd->vdev_parent->vdev_ops == &vdev_spare_ops &&
5354 	    !bvd->vdev_isspare) {
5355 		cmn_err(CE_NOTE, "The boot device is currently spared. Please "
5356 		    "try booting from '%s'",
5357 		    bvd->vdev_parent->
5358 		    vdev_child[bvd->vdev_parent->vdev_children - 1]->vdev_path);
5359 		error = SET_ERROR(EINVAL);
5360 		goto out;
5361 	}
5362 
5363 	error = 0;
5364 out:
5365 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5366 	vdev_free(rvd);
5367 	spa_config_exit(spa, SCL_ALL, FTAG);
5368 	mutex_exit(&spa_namespace_lock);
5369 
5370 	nvlist_free(config);
5371 	return (error);
5372 }
5373 
5374 #endif
5375 
5376 /*
5377  * Import a non-root pool into the system.
5378  */
5379 int
5380 spa_import(const char *pool, nvlist_t *config, nvlist_t *props, uint64_t flags)
5381 {
5382 	spa_t *spa;
5383 	char *altroot = NULL;
5384 	spa_load_state_t state = SPA_LOAD_IMPORT;
5385 	zpool_load_policy_t policy;
5386 	uint64_t mode = spa_mode_global;
5387 	uint64_t readonly = B_FALSE;
5388 	int error;
5389 	nvlist_t *nvroot;
5390 	nvlist_t **spares, **l2cache;
5391 	uint_t nspares, nl2cache;
5392 
5393 	/*
5394 	 * If a pool with this name exists, return failure.
5395 	 */
5396 	mutex_enter(&spa_namespace_lock);
5397 	if (spa_lookup(pool) != NULL) {
5398 		mutex_exit(&spa_namespace_lock);
5399 		return (SET_ERROR(EEXIST));
5400 	}
5401 
5402 	/*
5403 	 * Create and initialize the spa structure.
5404 	 */
5405 	(void) nvlist_lookup_string(props,
5406 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
5407 	(void) nvlist_lookup_uint64(props,
5408 	    zpool_prop_to_name(ZPOOL_PROP_READONLY), &readonly);
5409 	if (readonly)
5410 		mode = FREAD;
5411 	spa = spa_add(pool, config, altroot);
5412 	spa->spa_import_flags = flags;
5413 
5414 	/*
5415 	 * Verbatim import - Take a pool and insert it into the namespace
5416 	 * as if it had been loaded at boot.
5417 	 */
5418 	if (spa->spa_import_flags & ZFS_IMPORT_VERBATIM) {
5419 		if (props != NULL)
5420 			spa_configfile_set(spa, props, B_FALSE);
5421 
5422 		spa_write_cachefile(spa, B_FALSE, B_TRUE);
5423 		spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
5424 		zfs_dbgmsg("spa_import: verbatim import of %s", pool);
5425 		mutex_exit(&spa_namespace_lock);
5426 		return (0);
5427 	}
5428 
5429 	spa_activate(spa, mode);
5430 
5431 	/*
5432 	 * Don't start async tasks until we know everything is healthy.
5433 	 */
5434 	spa_async_suspend(spa);
5435 
5436 	zpool_get_load_policy(config, &policy);
5437 	if (policy.zlp_rewind & ZPOOL_DO_REWIND)
5438 		state = SPA_LOAD_RECOVER;
5439 
5440 	spa->spa_config_source = SPA_CONFIG_SRC_TRYIMPORT;
5441 
5442 	if (state != SPA_LOAD_RECOVER) {
5443 		spa->spa_last_ubsync_txg = spa->spa_load_txg = 0;
5444 		zfs_dbgmsg("spa_import: importing %s", pool);
5445 	} else {
5446 		zfs_dbgmsg("spa_import: importing %s, max_txg=%lld "
5447 		    "(RECOVERY MODE)", pool, (longlong_t)policy.zlp_txg);
5448 	}
5449 	error = spa_load_best(spa, state, policy.zlp_txg, policy.zlp_rewind);
5450 
5451 	/*
5452 	 * Propagate anything learned while loading the pool and pass it
5453 	 * back to caller (i.e. rewind info, missing devices, etc).
5454 	 */
5455 	VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO,
5456 	    spa->spa_load_info) == 0);
5457 
5458 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5459 	/*
5460 	 * Toss any existing sparelist, as it doesn't have any validity
5461 	 * anymore, and conflicts with spa_has_spare().
5462 	 */
5463 	if (spa->spa_spares.sav_config) {
5464 		nvlist_free(spa->spa_spares.sav_config);
5465 		spa->spa_spares.sav_config = NULL;
5466 		spa_load_spares(spa);
5467 	}
5468 	if (spa->spa_l2cache.sav_config) {
5469 		nvlist_free(spa->spa_l2cache.sav_config);
5470 		spa->spa_l2cache.sav_config = NULL;
5471 		spa_load_l2cache(spa);
5472 	}
5473 
5474 	VERIFY(nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE,
5475 	    &nvroot) == 0);
5476 	if (error == 0)
5477 		error = spa_validate_aux(spa, nvroot, -1ULL,
5478 		    VDEV_ALLOC_SPARE);
5479 	if (error == 0)
5480 		error = spa_validate_aux(spa, nvroot, -1ULL,
5481 		    VDEV_ALLOC_L2CACHE);
5482 	spa_config_exit(spa, SCL_ALL, FTAG);
5483 
5484 	if (props != NULL)
5485 		spa_configfile_set(spa, props, B_FALSE);
5486 
5487 	if (error != 0 || (props && spa_writeable(spa) &&
5488 	    (error = spa_prop_set(spa, props)))) {
5489 		spa_unload(spa);
5490 		spa_deactivate(spa);
5491 		spa_remove(spa);
5492 		mutex_exit(&spa_namespace_lock);
5493 		return (error);
5494 	}
5495 
5496 	spa_async_resume(spa);
5497 
5498 	/*
5499 	 * Override any spares and level 2 cache devices as specified by
5500 	 * the user, as these may have correct device names/devids, etc.
5501 	 */
5502 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES,
5503 	    &spares, &nspares) == 0) {
5504 		if (spa->spa_spares.sav_config)
5505 			VERIFY(nvlist_remove(spa->spa_spares.sav_config,
5506 			    ZPOOL_CONFIG_SPARES, DATA_TYPE_NVLIST_ARRAY) == 0);
5507 		else
5508 			VERIFY(nvlist_alloc(&spa->spa_spares.sav_config,
5509 			    NV_UNIQUE_NAME, KM_SLEEP) == 0);
5510 		VERIFY(nvlist_add_nvlist_array(spa->spa_spares.sav_config,
5511 		    ZPOOL_CONFIG_SPARES, spares, nspares) == 0);
5512 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5513 		spa_load_spares(spa);
5514 		spa_config_exit(spa, SCL_ALL, FTAG);
5515 		spa->spa_spares.sav_sync = B_TRUE;
5516 	}
5517 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE,
5518 	    &l2cache, &nl2cache) == 0) {
5519 		if (spa->spa_l2cache.sav_config)
5520 			VERIFY(nvlist_remove(spa->spa_l2cache.sav_config,
5521 			    ZPOOL_CONFIG_L2CACHE, DATA_TYPE_NVLIST_ARRAY) == 0);
5522 		else
5523 			VERIFY(nvlist_alloc(&spa->spa_l2cache.sav_config,
5524 			    NV_UNIQUE_NAME, KM_SLEEP) == 0);
5525 		VERIFY(nvlist_add_nvlist_array(spa->spa_l2cache.sav_config,
5526 		    ZPOOL_CONFIG_L2CACHE, l2cache, nl2cache) == 0);
5527 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5528 		spa_load_l2cache(spa);
5529 		spa_config_exit(spa, SCL_ALL, FTAG);
5530 		spa->spa_l2cache.sav_sync = B_TRUE;
5531 	}
5532 
5533 	/*
5534 	 * Check for any removed devices.
5535 	 */
5536 	if (spa->spa_autoreplace) {
5537 		spa_aux_check_removed(&spa->spa_spares);
5538 		spa_aux_check_removed(&spa->spa_l2cache);
5539 	}
5540 
5541 	if (spa_writeable(spa)) {
5542 		/*
5543 		 * Update the config cache to include the newly-imported pool.
5544 		 */
5545 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
5546 	}
5547 
5548 	/*
5549 	 * It's possible that the pool was expanded while it was exported.
5550 	 * We kick off an async task to handle this for us.
5551 	 */
5552 	spa_async_request(spa, SPA_ASYNC_AUTOEXPAND);
5553 
5554 	spa_history_log_version(spa, "import");
5555 
5556 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_IMPORT);
5557 
5558 	mutex_exit(&spa_namespace_lock);
5559 
5560 	return (0);
5561 }
5562 
5563 nvlist_t *
5564 spa_tryimport(nvlist_t *tryconfig)
5565 {
5566 	nvlist_t *config = NULL;
5567 	char *poolname, *cachefile;
5568 	spa_t *spa;
5569 	uint64_t state;
5570 	int error;
5571 	zpool_load_policy_t policy;
5572 
5573 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_POOL_NAME, &poolname))
5574 		return (NULL);
5575 
5576 	if (nvlist_lookup_uint64(tryconfig, ZPOOL_CONFIG_POOL_STATE, &state))
5577 		return (NULL);
5578 
5579 	/*
5580 	 * Create and initialize the spa structure.
5581 	 */
5582 	mutex_enter(&spa_namespace_lock);
5583 	spa = spa_add(TRYIMPORT_NAME, tryconfig, NULL);
5584 	spa_activate(spa, FREAD);
5585 
5586 	/*
5587 	 * Rewind pool if a max txg was provided.
5588 	 */
5589 	zpool_get_load_policy(spa->spa_config, &policy);
5590 	if (policy.zlp_txg != UINT64_MAX) {
5591 		spa->spa_load_max_txg = policy.zlp_txg;
5592 		spa->spa_extreme_rewind = B_TRUE;
5593 		zfs_dbgmsg("spa_tryimport: importing %s, max_txg=%lld",
5594 		    poolname, (longlong_t)policy.zlp_txg);
5595 	} else {
5596 		zfs_dbgmsg("spa_tryimport: importing %s", poolname);
5597 	}
5598 
5599 	if (nvlist_lookup_string(tryconfig, ZPOOL_CONFIG_CACHEFILE, &cachefile)
5600 	    == 0) {
5601 		zfs_dbgmsg("spa_tryimport: using cachefile '%s'", cachefile);
5602 		spa->spa_config_source = SPA_CONFIG_SRC_CACHEFILE;
5603 	} else {
5604 		spa->spa_config_source = SPA_CONFIG_SRC_SCAN;
5605 	}
5606 
5607 	error = spa_load(spa, SPA_LOAD_TRYIMPORT, SPA_IMPORT_EXISTING);
5608 
5609 	/*
5610 	 * If 'tryconfig' was at least parsable, return the current config.
5611 	 */
5612 	if (spa->spa_root_vdev != NULL) {
5613 		config = spa_config_generate(spa, NULL, -1ULL, B_TRUE);
5614 		VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME,
5615 		    poolname) == 0);
5616 		VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE,
5617 		    state) == 0);
5618 		VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_TIMESTAMP,
5619 		    spa->spa_uberblock.ub_timestamp) == 0);
5620 		VERIFY(nvlist_add_nvlist(config, ZPOOL_CONFIG_LOAD_INFO,
5621 		    spa->spa_load_info) == 0);
5622 
5623 		/*
5624 		 * If the bootfs property exists on this pool then we
5625 		 * copy it out so that external consumers can tell which
5626 		 * pools are bootable.
5627 		 */
5628 		if ((!error || error == EEXIST) && spa->spa_bootfs) {
5629 			char *tmpname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
5630 
5631 			/*
5632 			 * We have to play games with the name since the
5633 			 * pool was opened as TRYIMPORT_NAME.
5634 			 */
5635 			if (dsl_dsobj_to_dsname(spa_name(spa),
5636 			    spa->spa_bootfs, tmpname) == 0) {
5637 				char *cp;
5638 				char *dsname = kmem_alloc(MAXPATHLEN, KM_SLEEP);
5639 
5640 				cp = strchr(tmpname, '/');
5641 				if (cp == NULL) {
5642 					(void) strlcpy(dsname, tmpname,
5643 					    MAXPATHLEN);
5644 				} else {
5645 					(void) snprintf(dsname, MAXPATHLEN,
5646 					    "%s/%s", poolname, ++cp);
5647 				}
5648 				VERIFY(nvlist_add_string(config,
5649 				    ZPOOL_CONFIG_BOOTFS, dsname) == 0);
5650 				kmem_free(dsname, MAXPATHLEN);
5651 			}
5652 			kmem_free(tmpname, MAXPATHLEN);
5653 		}
5654 
5655 		/*
5656 		 * Add the list of hot spares and level 2 cache devices.
5657 		 */
5658 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
5659 		spa_add_spares(spa, config);
5660 		spa_add_l2cache(spa, config);
5661 		spa_config_exit(spa, SCL_CONFIG, FTAG);
5662 	}
5663 
5664 	spa_unload(spa);
5665 	spa_deactivate(spa);
5666 	spa_remove(spa);
5667 	mutex_exit(&spa_namespace_lock);
5668 
5669 	return (config);
5670 }
5671 
5672 /*
5673  * Pool export/destroy
5674  *
5675  * The act of destroying or exporting a pool is very simple.  We make sure there
5676  * is no more pending I/O and any references to the pool are gone.  Then, we
5677  * update the pool state and sync all the labels to disk, removing the
5678  * configuration from the cache afterwards. If the 'hardforce' flag is set, then
5679  * we don't sync the labels or remove the configuration cache.
5680  */
5681 static int
5682 spa_export_common(char *pool, int new_state, nvlist_t **oldconfig,
5683     boolean_t force, boolean_t hardforce)
5684 {
5685 	spa_t *spa;
5686 
5687 	if (oldconfig)
5688 		*oldconfig = NULL;
5689 
5690 	if (!(spa_mode_global & FWRITE))
5691 		return (SET_ERROR(EROFS));
5692 
5693 	mutex_enter(&spa_namespace_lock);
5694 	if ((spa = spa_lookup(pool)) == NULL) {
5695 		mutex_exit(&spa_namespace_lock);
5696 		return (SET_ERROR(ENOENT));
5697 	}
5698 
5699 	/*
5700 	 * Put a hold on the pool, drop the namespace lock, stop async tasks,
5701 	 * reacquire the namespace lock, and see if we can export.
5702 	 */
5703 	spa_open_ref(spa, FTAG);
5704 	mutex_exit(&spa_namespace_lock);
5705 	spa_async_suspend(spa);
5706 	mutex_enter(&spa_namespace_lock);
5707 	spa_close(spa, FTAG);
5708 
5709 	/*
5710 	 * The pool will be in core if it's openable,
5711 	 * in which case we can modify its state.
5712 	 */
5713 	if (spa->spa_state != POOL_STATE_UNINITIALIZED && spa->spa_sync_on) {
5714 
5715 		/*
5716 		 * Objsets may be open only because they're dirty, so we
5717 		 * have to force it to sync before checking spa_refcnt.
5718 		 */
5719 		txg_wait_synced(spa->spa_dsl_pool, 0);
5720 		spa_evicting_os_wait(spa);
5721 
5722 		/*
5723 		 * A pool cannot be exported or destroyed if there are active
5724 		 * references.  If we are resetting a pool, allow references by
5725 		 * fault injection handlers.
5726 		 */
5727 		if (!spa_refcount_zero(spa) ||
5728 		    (spa->spa_inject_ref != 0 &&
5729 		    new_state != POOL_STATE_UNINITIALIZED)) {
5730 			spa_async_resume(spa);
5731 			mutex_exit(&spa_namespace_lock);
5732 			return (SET_ERROR(EBUSY));
5733 		}
5734 
5735 		/*
5736 		 * A pool cannot be exported if it has an active shared spare.
5737 		 * This is to prevent other pools stealing the active spare
5738 		 * from an exported pool. At user's own will, such pool can
5739 		 * be forcedly exported.
5740 		 */
5741 		if (!force && new_state == POOL_STATE_EXPORTED &&
5742 		    spa_has_active_shared_spare(spa)) {
5743 			spa_async_resume(spa);
5744 			mutex_exit(&spa_namespace_lock);
5745 			return (SET_ERROR(EXDEV));
5746 		}
5747 
5748 		/*
5749 		 * We're about to export or destroy this pool. Make sure
5750 		 * we stop all initializtion activity here before we
5751 		 * set the spa_final_txg. This will ensure that all
5752 		 * dirty data resulting from the initialization is
5753 		 * committed to disk before we unload the pool.
5754 		 */
5755 		if (spa->spa_root_vdev != NULL) {
5756 			vdev_initialize_stop_all(spa->spa_root_vdev,
5757 			    VDEV_INITIALIZE_ACTIVE);
5758 		}
5759 
5760 		/*
5761 		 * We want this to be reflected on every label,
5762 		 * so mark them all dirty.  spa_unload() will do the
5763 		 * final sync that pushes these changes out.
5764 		 */
5765 		if (new_state != POOL_STATE_UNINITIALIZED && !hardforce) {
5766 			spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
5767 			spa->spa_state = new_state;
5768 			spa->spa_final_txg = spa_last_synced_txg(spa) +
5769 			    TXG_DEFER_SIZE + 1;
5770 			vdev_config_dirty(spa->spa_root_vdev);
5771 			spa_config_exit(spa, SCL_ALL, FTAG);
5772 		}
5773 	}
5774 
5775 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_POOL_DESTROY);
5776 
5777 	if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
5778 		spa_unload(spa);
5779 		spa_deactivate(spa);
5780 	}
5781 
5782 	if (oldconfig && spa->spa_config)
5783 		VERIFY(nvlist_dup(spa->spa_config, oldconfig, 0) == 0);
5784 
5785 	if (new_state != POOL_STATE_UNINITIALIZED) {
5786 		if (!hardforce)
5787 			spa_write_cachefile(spa, B_TRUE, B_TRUE);
5788 		spa_remove(spa);
5789 	}
5790 	mutex_exit(&spa_namespace_lock);
5791 
5792 	return (0);
5793 }
5794 
5795 /*
5796  * Destroy a storage pool.
5797  */
5798 int
5799 spa_destroy(char *pool)
5800 {
5801 	return (spa_export_common(pool, POOL_STATE_DESTROYED, NULL,
5802 	    B_FALSE, B_FALSE));
5803 }
5804 
5805 /*
5806  * Export a storage pool.
5807  */
5808 int
5809 spa_export(char *pool, nvlist_t **oldconfig, boolean_t force,
5810     boolean_t hardforce)
5811 {
5812 	return (spa_export_common(pool, POOL_STATE_EXPORTED, oldconfig,
5813 	    force, hardforce));
5814 }
5815 
5816 /*
5817  * Similar to spa_export(), this unloads the spa_t without actually removing it
5818  * from the namespace in any way.
5819  */
5820 int
5821 spa_reset(char *pool)
5822 {
5823 	return (spa_export_common(pool, POOL_STATE_UNINITIALIZED, NULL,
5824 	    B_FALSE, B_FALSE));
5825 }
5826 
5827 /*
5828  * ==========================================================================
5829  * Device manipulation
5830  * ==========================================================================
5831  */
5832 
5833 /*
5834  * Add a device to a storage pool.
5835  */
5836 int
5837 spa_vdev_add(spa_t *spa, nvlist_t *nvroot)
5838 {
5839 	uint64_t txg, id;
5840 	int error;
5841 	vdev_t *rvd = spa->spa_root_vdev;
5842 	vdev_t *vd, *tvd;
5843 	nvlist_t **spares, **l2cache;
5844 	uint_t nspares, nl2cache;
5845 
5846 	ASSERT(spa_writeable(spa));
5847 
5848 	txg = spa_vdev_enter(spa);
5849 
5850 	if ((error = spa_config_parse(spa, &vd, nvroot, NULL, 0,
5851 	    VDEV_ALLOC_ADD)) != 0)
5852 		return (spa_vdev_exit(spa, NULL, txg, error));
5853 
5854 	spa->spa_pending_vdev = vd;	/* spa_vdev_exit() will clear this */
5855 
5856 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_SPARES, &spares,
5857 	    &nspares) != 0)
5858 		nspares = 0;
5859 
5860 	if (nvlist_lookup_nvlist_array(nvroot, ZPOOL_CONFIG_L2CACHE, &l2cache,
5861 	    &nl2cache) != 0)
5862 		nl2cache = 0;
5863 
5864 	if (vd->vdev_children == 0 && nspares == 0 && nl2cache == 0)
5865 		return (spa_vdev_exit(spa, vd, txg, EINVAL));
5866 
5867 	if (vd->vdev_children != 0 &&
5868 	    (error = vdev_create(vd, txg, B_FALSE)) != 0)
5869 		return (spa_vdev_exit(spa, vd, txg, error));
5870 
5871 	/*
5872 	 * We must validate the spares and l2cache devices after checking the
5873 	 * children.  Otherwise, vdev_inuse() will blindly overwrite the spare.
5874 	 */
5875 	if ((error = spa_validate_aux(spa, nvroot, txg, VDEV_ALLOC_ADD)) != 0)
5876 		return (spa_vdev_exit(spa, vd, txg, error));
5877 
5878 	/*
5879 	 * If we are in the middle of a device removal, we can only add
5880 	 * devices which match the existing devices in the pool.
5881 	 * If we are in the middle of a removal, or have some indirect
5882 	 * vdevs, we can not add raidz toplevels.
5883 	 */
5884 	if (spa->spa_vdev_removal != NULL ||
5885 	    spa->spa_removing_phys.sr_prev_indirect_vdev != -1) {
5886 		for (int c = 0; c < vd->vdev_children; c++) {
5887 			tvd = vd->vdev_child[c];
5888 			if (spa->spa_vdev_removal != NULL &&
5889 			    tvd->vdev_ashift != spa->spa_max_ashift) {
5890 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
5891 			}
5892 			/* Fail if top level vdev is raidz */
5893 			if (tvd->vdev_ops == &vdev_raidz_ops) {
5894 				return (spa_vdev_exit(spa, vd, txg, EINVAL));
5895 			}
5896 			/*
5897 			 * Need the top level mirror to be
5898 			 * a mirror of leaf vdevs only
5899 			 */
5900 			if (tvd->vdev_ops == &vdev_mirror_ops) {
5901 				for (uint64_t cid = 0;
5902 				    cid < tvd->vdev_children; cid++) {
5903 					vdev_t *cvd = tvd->vdev_child[cid];
5904 					if (!cvd->vdev_ops->vdev_op_leaf) {
5905 						return (spa_vdev_exit(spa, vd,
5906 						    txg, EINVAL));
5907 					}
5908 				}
5909 			}
5910 		}
5911 	}
5912 
5913 	for (int c = 0; c < vd->vdev_children; c++) {
5914 
5915 		/*
5916 		 * Set the vdev id to the first hole, if one exists.
5917 		 */
5918 		for (id = 0; id < rvd->vdev_children; id++) {
5919 			if (rvd->vdev_child[id]->vdev_ishole) {
5920 				vdev_free(rvd->vdev_child[id]);
5921 				break;
5922 			}
5923 		}
5924 		tvd = vd->vdev_child[c];
5925 		vdev_remove_child(vd, tvd);
5926 		tvd->vdev_id = id;
5927 		vdev_add_child(rvd, tvd);
5928 		vdev_config_dirty(tvd);
5929 	}
5930 
5931 	if (nspares != 0) {
5932 		spa_set_aux_vdevs(&spa->spa_spares, spares, nspares,
5933 		    ZPOOL_CONFIG_SPARES);
5934 		spa_load_spares(spa);
5935 		spa->spa_spares.sav_sync = B_TRUE;
5936 	}
5937 
5938 	if (nl2cache != 0) {
5939 		spa_set_aux_vdevs(&spa->spa_l2cache, l2cache, nl2cache,
5940 		    ZPOOL_CONFIG_L2CACHE);
5941 		spa_load_l2cache(spa);
5942 		spa->spa_l2cache.sav_sync = B_TRUE;
5943 	}
5944 
5945 	/*
5946 	 * We have to be careful when adding new vdevs to an existing pool.
5947 	 * If other threads start allocating from these vdevs before we
5948 	 * sync the config cache, and we lose power, then upon reboot we may
5949 	 * fail to open the pool because there are DVAs that the config cache
5950 	 * can't translate.  Therefore, we first add the vdevs without
5951 	 * initializing metaslabs; sync the config cache (via spa_vdev_exit());
5952 	 * and then let spa_config_update() initialize the new metaslabs.
5953 	 *
5954 	 * spa_load() checks for added-but-not-initialized vdevs, so that
5955 	 * if we lose power at any point in this sequence, the remaining
5956 	 * steps will be completed the next time we load the pool.
5957 	 */
5958 	(void) spa_vdev_exit(spa, vd, txg, 0);
5959 
5960 	mutex_enter(&spa_namespace_lock);
5961 	spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
5962 	spa_event_notify(spa, NULL, NULL, ESC_ZFS_VDEV_ADD);
5963 	mutex_exit(&spa_namespace_lock);
5964 
5965 	return (0);
5966 }
5967 
5968 /*
5969  * Attach a device to a mirror.  The arguments are the path to any device
5970  * in the mirror, and the nvroot for the new device.  If the path specifies
5971  * a device that is not mirrored, we automatically insert the mirror vdev.
5972  *
5973  * If 'replacing' is specified, the new device is intended to replace the
5974  * existing device; in this case the two devices are made into their own
5975  * mirror using the 'replacing' vdev, which is functionally identical to
5976  * the mirror vdev (it actually reuses all the same ops) but has a few
5977  * extra rules: you can't attach to it after it's been created, and upon
5978  * completion of resilvering, the first disk (the one being replaced)
5979  * is automatically detached.
5980  */
5981 int
5982 spa_vdev_attach(spa_t *spa, uint64_t guid, nvlist_t *nvroot, int replacing)
5983 {
5984 	uint64_t txg, dtl_max_txg;
5985 	vdev_t *rvd = spa->spa_root_vdev;
5986 	vdev_t *oldvd, *newvd, *newrootvd, *pvd, *tvd;
5987 	vdev_ops_t *pvops;
5988 	char *oldvdpath, *newvdpath;
5989 	int newvd_isspare;
5990 	int error;
5991 
5992 	ASSERT(spa_writeable(spa));
5993 
5994 	txg = spa_vdev_enter(spa);
5995 
5996 	oldvd = spa_lookup_by_guid(spa, guid, B_FALSE);
5997 
5998 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
5999 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6000 		error = (spa_has_checkpoint(spa)) ?
6001 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6002 		return (spa_vdev_exit(spa, NULL, txg, error));
6003 	}
6004 
6005 	if (spa->spa_vdev_removal != NULL)
6006 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6007 
6008 	if (oldvd == NULL)
6009 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
6010 
6011 	if (!oldvd->vdev_ops->vdev_op_leaf)
6012 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6013 
6014 	pvd = oldvd->vdev_parent;
6015 
6016 	if ((error = spa_config_parse(spa, &newrootvd, nvroot, NULL, 0,
6017 	    VDEV_ALLOC_ATTACH)) != 0)
6018 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6019 
6020 	if (newrootvd->vdev_children != 1)
6021 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6022 
6023 	newvd = newrootvd->vdev_child[0];
6024 
6025 	if (!newvd->vdev_ops->vdev_op_leaf)
6026 		return (spa_vdev_exit(spa, newrootvd, txg, EINVAL));
6027 
6028 	if ((error = vdev_create(newrootvd, txg, replacing)) != 0)
6029 		return (spa_vdev_exit(spa, newrootvd, txg, error));
6030 
6031 	/*
6032 	 * Spares can't replace logs
6033 	 */
6034 	if (oldvd->vdev_top->vdev_islog && newvd->vdev_isspare)
6035 		return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6036 
6037 	if (!replacing) {
6038 		/*
6039 		 * For attach, the only allowable parent is a mirror or the root
6040 		 * vdev.
6041 		 */
6042 		if (pvd->vdev_ops != &vdev_mirror_ops &&
6043 		    pvd->vdev_ops != &vdev_root_ops)
6044 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6045 
6046 		pvops = &vdev_mirror_ops;
6047 	} else {
6048 		/*
6049 		 * Active hot spares can only be replaced by inactive hot
6050 		 * spares.
6051 		 */
6052 		if (pvd->vdev_ops == &vdev_spare_ops &&
6053 		    oldvd->vdev_isspare &&
6054 		    !spa_has_spare(spa, newvd->vdev_guid))
6055 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6056 
6057 		/*
6058 		 * If the source is a hot spare, and the parent isn't already a
6059 		 * spare, then we want to create a new hot spare.  Otherwise, we
6060 		 * want to create a replacing vdev.  The user is not allowed to
6061 		 * attach to a spared vdev child unless the 'isspare' state is
6062 		 * the same (spare replaces spare, non-spare replaces
6063 		 * non-spare).
6064 		 */
6065 		if (pvd->vdev_ops == &vdev_replacing_ops &&
6066 		    spa_version(spa) < SPA_VERSION_MULTI_REPLACE) {
6067 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6068 		} else if (pvd->vdev_ops == &vdev_spare_ops &&
6069 		    newvd->vdev_isspare != oldvd->vdev_isspare) {
6070 			return (spa_vdev_exit(spa, newrootvd, txg, ENOTSUP));
6071 		}
6072 
6073 		if (newvd->vdev_isspare)
6074 			pvops = &vdev_spare_ops;
6075 		else
6076 			pvops = &vdev_replacing_ops;
6077 	}
6078 
6079 	/*
6080 	 * Make sure the new device is big enough.
6081 	 */
6082 	if (newvd->vdev_asize < vdev_get_min_asize(oldvd))
6083 		return (spa_vdev_exit(spa, newrootvd, txg, EOVERFLOW));
6084 
6085 	/*
6086 	 * The new device cannot have a higher alignment requirement
6087 	 * than the top-level vdev.
6088 	 */
6089 	if (newvd->vdev_ashift > oldvd->vdev_top->vdev_ashift)
6090 		return (spa_vdev_exit(spa, newrootvd, txg, EDOM));
6091 
6092 	/*
6093 	 * If this is an in-place replacement, update oldvd's path and devid
6094 	 * to make it distinguishable from newvd, and unopenable from now on.
6095 	 */
6096 	if (strcmp(oldvd->vdev_path, newvd->vdev_path) == 0) {
6097 		spa_strfree(oldvd->vdev_path);
6098 		oldvd->vdev_path = kmem_alloc(strlen(newvd->vdev_path) + 5,
6099 		    KM_SLEEP);
6100 		(void) sprintf(oldvd->vdev_path, "%s/%s",
6101 		    newvd->vdev_path, "old");
6102 		if (oldvd->vdev_devid != NULL) {
6103 			spa_strfree(oldvd->vdev_devid);
6104 			oldvd->vdev_devid = NULL;
6105 		}
6106 	}
6107 
6108 	/* mark the device being resilvered */
6109 	newvd->vdev_resilver_txg = txg;
6110 
6111 	/*
6112 	 * If the parent is not a mirror, or if we're replacing, insert the new
6113 	 * mirror/replacing/spare vdev above oldvd.
6114 	 */
6115 	if (pvd->vdev_ops != pvops)
6116 		pvd = vdev_add_parent(oldvd, pvops);
6117 
6118 	ASSERT(pvd->vdev_top->vdev_parent == rvd);
6119 	ASSERT(pvd->vdev_ops == pvops);
6120 	ASSERT(oldvd->vdev_parent == pvd);
6121 
6122 	/*
6123 	 * Extract the new device from its root and add it to pvd.
6124 	 */
6125 	vdev_remove_child(newrootvd, newvd);
6126 	newvd->vdev_id = pvd->vdev_children;
6127 	newvd->vdev_crtxg = oldvd->vdev_crtxg;
6128 	vdev_add_child(pvd, newvd);
6129 
6130 	tvd = newvd->vdev_top;
6131 	ASSERT(pvd->vdev_top == tvd);
6132 	ASSERT(tvd->vdev_parent == rvd);
6133 
6134 	vdev_config_dirty(tvd);
6135 
6136 	/*
6137 	 * Set newvd's DTL to [TXG_INITIAL, dtl_max_txg) so that we account
6138 	 * for any dmu_sync-ed blocks.  It will propagate upward when
6139 	 * spa_vdev_exit() calls vdev_dtl_reassess().
6140 	 */
6141 	dtl_max_txg = txg + TXG_CONCURRENT_STATES;
6142 
6143 	vdev_dtl_dirty(newvd, DTL_MISSING, TXG_INITIAL,
6144 	    dtl_max_txg - TXG_INITIAL);
6145 
6146 	if (newvd->vdev_isspare) {
6147 		spa_spare_activate(newvd);
6148 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_SPARE);
6149 	}
6150 
6151 	oldvdpath = spa_strdup(oldvd->vdev_path);
6152 	newvdpath = spa_strdup(newvd->vdev_path);
6153 	newvd_isspare = newvd->vdev_isspare;
6154 
6155 	/*
6156 	 * Mark newvd's DTL dirty in this txg.
6157 	 */
6158 	vdev_dirty(tvd, VDD_DTL, newvd, txg);
6159 
6160 	/*
6161 	 * Schedule the resilver to restart in the future. We do this to
6162 	 * ensure that dmu_sync-ed blocks have been stitched into the
6163 	 * respective datasets. We do not do this if resilvers have been
6164 	 * deferred.
6165 	 */
6166 	if (dsl_scan_resilvering(spa_get_dsl(spa)) &&
6167 	    spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER))
6168 		vdev_set_deferred_resilver(spa, newvd);
6169 	else
6170 		dsl_resilver_restart(spa->spa_dsl_pool, dtl_max_txg);
6171 
6172 	if (spa->spa_bootfs)
6173 		spa_event_notify(spa, newvd, NULL, ESC_ZFS_BOOTFS_VDEV_ATTACH);
6174 
6175 	spa_event_notify(spa, newvd, NULL, ESC_ZFS_VDEV_ATTACH);
6176 
6177 	/*
6178 	 * Commit the config
6179 	 */
6180 	(void) spa_vdev_exit(spa, newrootvd, dtl_max_txg, 0);
6181 
6182 	spa_history_log_internal(spa, "vdev attach", NULL,
6183 	    "%s vdev=%s %s vdev=%s",
6184 	    replacing && newvd_isspare ? "spare in" :
6185 	    replacing ? "replace" : "attach", newvdpath,
6186 	    replacing ? "for" : "to", oldvdpath);
6187 
6188 	spa_strfree(oldvdpath);
6189 	spa_strfree(newvdpath);
6190 
6191 	return (0);
6192 }
6193 
6194 /*
6195  * Detach a device from a mirror or replacing vdev.
6196  *
6197  * If 'replace_done' is specified, only detach if the parent
6198  * is a replacing vdev.
6199  */
6200 int
6201 spa_vdev_detach(spa_t *spa, uint64_t guid, uint64_t pguid, int replace_done)
6202 {
6203 	uint64_t txg;
6204 	int error;
6205 	vdev_t *rvd = spa->spa_root_vdev;
6206 	vdev_t *vd, *pvd, *cvd, *tvd;
6207 	boolean_t unspare = B_FALSE;
6208 	uint64_t unspare_guid = 0;
6209 	char *vdpath;
6210 
6211 	ASSERT(spa_writeable(spa));
6212 
6213 	txg = spa_vdev_enter(spa);
6214 
6215 	vd = spa_lookup_by_guid(spa, guid, B_FALSE);
6216 
6217 	/*
6218 	 * Besides being called directly from the userland through the
6219 	 * ioctl interface, spa_vdev_detach() can be potentially called
6220 	 * at the end of spa_vdev_resilver_done().
6221 	 *
6222 	 * In the regular case, when we have a checkpoint this shouldn't
6223 	 * happen as we never empty the DTLs of a vdev during the scrub
6224 	 * [see comment in dsl_scan_done()]. Thus spa_vdev_resilvering_done()
6225 	 * should never get here when we have a checkpoint.
6226 	 *
6227 	 * That said, even in a case when we checkpoint the pool exactly
6228 	 * as spa_vdev_resilver_done() calls this function everything
6229 	 * should be fine as the resilver will return right away.
6230 	 */
6231 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
6232 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6233 		error = (spa_has_checkpoint(spa)) ?
6234 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6235 		return (spa_vdev_exit(spa, NULL, txg, error));
6236 	}
6237 
6238 	if (vd == NULL)
6239 		return (spa_vdev_exit(spa, NULL, txg, ENODEV));
6240 
6241 	if (!vd->vdev_ops->vdev_op_leaf)
6242 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6243 
6244 	pvd = vd->vdev_parent;
6245 
6246 	/*
6247 	 * If the parent/child relationship is not as expected, don't do it.
6248 	 * Consider M(A,R(B,C)) -- that is, a mirror of A with a replacing
6249 	 * vdev that's replacing B with C.  The user's intent in replacing
6250 	 * is to go from M(A,B) to M(A,C).  If the user decides to cancel
6251 	 * the replace by detaching C, the expected behavior is to end up
6252 	 * M(A,B).  But suppose that right after deciding to detach C,
6253 	 * the replacement of B completes.  We would have M(A,C), and then
6254 	 * ask to detach C, which would leave us with just A -- not what
6255 	 * the user wanted.  To prevent this, we make sure that the
6256 	 * parent/child relationship hasn't changed -- in this example,
6257 	 * that C's parent is still the replacing vdev R.
6258 	 */
6259 	if (pvd->vdev_guid != pguid && pguid != 0)
6260 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6261 
6262 	/*
6263 	 * Only 'replacing' or 'spare' vdevs can be replaced.
6264 	 */
6265 	if (replace_done && pvd->vdev_ops != &vdev_replacing_ops &&
6266 	    pvd->vdev_ops != &vdev_spare_ops)
6267 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6268 
6269 	ASSERT(pvd->vdev_ops != &vdev_spare_ops ||
6270 	    spa_version(spa) >= SPA_VERSION_SPARES);
6271 
6272 	/*
6273 	 * Only mirror, replacing, and spare vdevs support detach.
6274 	 */
6275 	if (pvd->vdev_ops != &vdev_replacing_ops &&
6276 	    pvd->vdev_ops != &vdev_mirror_ops &&
6277 	    pvd->vdev_ops != &vdev_spare_ops)
6278 		return (spa_vdev_exit(spa, NULL, txg, ENOTSUP));
6279 
6280 	/*
6281 	 * If this device has the only valid copy of some data,
6282 	 * we cannot safely detach it.
6283 	 */
6284 	if (vdev_dtl_required(vd))
6285 		return (spa_vdev_exit(spa, NULL, txg, EBUSY));
6286 
6287 	ASSERT(pvd->vdev_children >= 2);
6288 
6289 	/*
6290 	 * If we are detaching the second disk from a replacing vdev, then
6291 	 * check to see if we changed the original vdev's path to have "/old"
6292 	 * at the end in spa_vdev_attach().  If so, undo that change now.
6293 	 */
6294 	if (pvd->vdev_ops == &vdev_replacing_ops && vd->vdev_id > 0 &&
6295 	    vd->vdev_path != NULL) {
6296 		size_t len = strlen(vd->vdev_path);
6297 
6298 		for (int c = 0; c < pvd->vdev_children; c++) {
6299 			cvd = pvd->vdev_child[c];
6300 
6301 			if (cvd == vd || cvd->vdev_path == NULL)
6302 				continue;
6303 
6304 			if (strncmp(cvd->vdev_path, vd->vdev_path, len) == 0 &&
6305 			    strcmp(cvd->vdev_path + len, "/old") == 0) {
6306 				spa_strfree(cvd->vdev_path);
6307 				cvd->vdev_path = spa_strdup(vd->vdev_path);
6308 				break;
6309 			}
6310 		}
6311 	}
6312 
6313 	/*
6314 	 * If we are detaching the original disk from a spare, then it implies
6315 	 * that the spare should become a real disk, and be removed from the
6316 	 * active spare list for the pool.
6317 	 */
6318 	if (pvd->vdev_ops == &vdev_spare_ops &&
6319 	    vd->vdev_id == 0 &&
6320 	    pvd->vdev_child[pvd->vdev_children - 1]->vdev_isspare)
6321 		unspare = B_TRUE;
6322 
6323 	/*
6324 	 * Erase the disk labels so the disk can be used for other things.
6325 	 * This must be done after all other error cases are handled,
6326 	 * but before we disembowel vd (so we can still do I/O to it).
6327 	 * But if we can't do it, don't treat the error as fatal --
6328 	 * it may be that the unwritability of the disk is the reason
6329 	 * it's being detached!
6330 	 */
6331 	error = vdev_label_init(vd, 0, VDEV_LABEL_REMOVE);
6332 
6333 	/*
6334 	 * Remove vd from its parent and compact the parent's children.
6335 	 */
6336 	vdev_remove_child(pvd, vd);
6337 	vdev_compact_children(pvd);
6338 
6339 	/*
6340 	 * Remember one of the remaining children so we can get tvd below.
6341 	 */
6342 	cvd = pvd->vdev_child[pvd->vdev_children - 1];
6343 
6344 	/*
6345 	 * If we need to remove the remaining child from the list of hot spares,
6346 	 * do it now, marking the vdev as no longer a spare in the process.
6347 	 * We must do this before vdev_remove_parent(), because that can
6348 	 * change the GUID if it creates a new toplevel GUID.  For a similar
6349 	 * reason, we must remove the spare now, in the same txg as the detach;
6350 	 * otherwise someone could attach a new sibling, change the GUID, and
6351 	 * the subsequent attempt to spa_vdev_remove(unspare_guid) would fail.
6352 	 */
6353 	if (unspare) {
6354 		ASSERT(cvd->vdev_isspare);
6355 		spa_spare_remove(cvd);
6356 		unspare_guid = cvd->vdev_guid;
6357 		(void) spa_vdev_remove(spa, unspare_guid, B_TRUE);
6358 		cvd->vdev_unspare = B_TRUE;
6359 	}
6360 
6361 	/*
6362 	 * If the parent mirror/replacing vdev only has one child,
6363 	 * the parent is no longer needed.  Remove it from the tree.
6364 	 */
6365 	if (pvd->vdev_children == 1) {
6366 		if (pvd->vdev_ops == &vdev_spare_ops)
6367 			cvd->vdev_unspare = B_FALSE;
6368 		vdev_remove_parent(cvd);
6369 	}
6370 
6371 
6372 	/*
6373 	 * We don't set tvd until now because the parent we just removed
6374 	 * may have been the previous top-level vdev.
6375 	 */
6376 	tvd = cvd->vdev_top;
6377 	ASSERT(tvd->vdev_parent == rvd);
6378 
6379 	/*
6380 	 * Reevaluate the parent vdev state.
6381 	 */
6382 	vdev_propagate_state(cvd);
6383 
6384 	/*
6385 	 * If the 'autoexpand' property is set on the pool then automatically
6386 	 * try to expand the size of the pool. For example if the device we
6387 	 * just detached was smaller than the others, it may be possible to
6388 	 * add metaslabs (i.e. grow the pool). We need to reopen the vdev
6389 	 * first so that we can obtain the updated sizes of the leaf vdevs.
6390 	 */
6391 	if (spa->spa_autoexpand) {
6392 		vdev_reopen(tvd);
6393 		vdev_expand(tvd, txg);
6394 	}
6395 
6396 	vdev_config_dirty(tvd);
6397 
6398 	/*
6399 	 * Mark vd's DTL as dirty in this txg.  vdev_dtl_sync() will see that
6400 	 * vd->vdev_detached is set and free vd's DTL object in syncing context.
6401 	 * But first make sure we're not on any *other* txg's DTL list, to
6402 	 * prevent vd from being accessed after it's freed.
6403 	 */
6404 	vdpath = spa_strdup(vd->vdev_path);
6405 	for (int t = 0; t < TXG_SIZE; t++)
6406 		(void) txg_list_remove_this(&tvd->vdev_dtl_list, vd, t);
6407 	vd->vdev_detached = B_TRUE;
6408 	vdev_dirty(tvd, VDD_DTL, vd, txg);
6409 
6410 	spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_REMOVE);
6411 
6412 	/* hang on to the spa before we release the lock */
6413 	spa_open_ref(spa, FTAG);
6414 
6415 	error = spa_vdev_exit(spa, vd, txg, 0);
6416 
6417 	spa_history_log_internal(spa, "detach", NULL,
6418 	    "vdev=%s", vdpath);
6419 	spa_strfree(vdpath);
6420 
6421 	/*
6422 	 * If this was the removal of the original device in a hot spare vdev,
6423 	 * then we want to go through and remove the device from the hot spare
6424 	 * list of every other pool.
6425 	 */
6426 	if (unspare) {
6427 		spa_t *altspa = NULL;
6428 
6429 		mutex_enter(&spa_namespace_lock);
6430 		while ((altspa = spa_next(altspa)) != NULL) {
6431 			if (altspa->spa_state != POOL_STATE_ACTIVE ||
6432 			    altspa == spa)
6433 				continue;
6434 
6435 			spa_open_ref(altspa, FTAG);
6436 			mutex_exit(&spa_namespace_lock);
6437 			(void) spa_vdev_remove(altspa, unspare_guid, B_TRUE);
6438 			mutex_enter(&spa_namespace_lock);
6439 			spa_close(altspa, FTAG);
6440 		}
6441 		mutex_exit(&spa_namespace_lock);
6442 
6443 		/* search the rest of the vdevs for spares to remove */
6444 		spa_vdev_resilver_done(spa);
6445 	}
6446 
6447 	/* all done with the spa; OK to release */
6448 	mutex_enter(&spa_namespace_lock);
6449 	spa_close(spa, FTAG);
6450 	mutex_exit(&spa_namespace_lock);
6451 
6452 	return (error);
6453 }
6454 
6455 int
6456 spa_vdev_initialize(spa_t *spa, uint64_t guid, uint64_t cmd_type)
6457 {
6458 	/*
6459 	 * We hold the namespace lock through the whole function
6460 	 * to prevent any changes to the pool while we're starting or
6461 	 * stopping initialization. The config and state locks are held so that
6462 	 * we can properly assess the vdev state before we commit to
6463 	 * the initializing operation.
6464 	 */
6465 	mutex_enter(&spa_namespace_lock);
6466 	spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
6467 
6468 	/* Look up vdev and ensure it's a leaf. */
6469 	vdev_t *vd = spa_lookup_by_guid(spa, guid, B_FALSE);
6470 	if (vd == NULL || vd->vdev_detached) {
6471 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
6472 		mutex_exit(&spa_namespace_lock);
6473 		return (SET_ERROR(ENODEV));
6474 	} else if (!vd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(vd)) {
6475 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
6476 		mutex_exit(&spa_namespace_lock);
6477 		return (SET_ERROR(EINVAL));
6478 	} else if (!vdev_writeable(vd)) {
6479 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
6480 		mutex_exit(&spa_namespace_lock);
6481 		return (SET_ERROR(EROFS));
6482 	}
6483 	mutex_enter(&vd->vdev_initialize_lock);
6484 	spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
6485 
6486 	/*
6487 	 * When we activate an initialize action we check to see
6488 	 * if the vdev_initialize_thread is NULL. We do this instead
6489 	 * of using the vdev_initialize_state since there might be
6490 	 * a previous initialization process which has completed but
6491 	 * the thread is not exited.
6492 	 */
6493 	if (cmd_type == POOL_INITIALIZE_DO &&
6494 	    (vd->vdev_initialize_thread != NULL ||
6495 	    vd->vdev_top->vdev_removing)) {
6496 		mutex_exit(&vd->vdev_initialize_lock);
6497 		mutex_exit(&spa_namespace_lock);
6498 		return (SET_ERROR(EBUSY));
6499 	} else if (cmd_type == POOL_INITIALIZE_CANCEL &&
6500 	    (vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE &&
6501 	    vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED)) {
6502 		mutex_exit(&vd->vdev_initialize_lock);
6503 		mutex_exit(&spa_namespace_lock);
6504 		return (SET_ERROR(ESRCH));
6505 	} else if (cmd_type == POOL_INITIALIZE_SUSPEND &&
6506 	    vd->vdev_initialize_state != VDEV_INITIALIZE_ACTIVE) {
6507 		mutex_exit(&vd->vdev_initialize_lock);
6508 		mutex_exit(&spa_namespace_lock);
6509 		return (SET_ERROR(ESRCH));
6510 	}
6511 
6512 	switch (cmd_type) {
6513 	case POOL_INITIALIZE_DO:
6514 		vdev_initialize(vd);
6515 		break;
6516 	case POOL_INITIALIZE_CANCEL:
6517 		vdev_initialize_stop(vd, VDEV_INITIALIZE_CANCELED);
6518 		break;
6519 	case POOL_INITIALIZE_SUSPEND:
6520 		vdev_initialize_stop(vd, VDEV_INITIALIZE_SUSPENDED);
6521 		break;
6522 	default:
6523 		panic("invalid cmd_type %llu", (unsigned long long)cmd_type);
6524 	}
6525 	mutex_exit(&vd->vdev_initialize_lock);
6526 
6527 	/* Sync out the initializing state */
6528 	txg_wait_synced(spa->spa_dsl_pool, 0);
6529 	mutex_exit(&spa_namespace_lock);
6530 
6531 	return (0);
6532 }
6533 
6534 
6535 /*
6536  * Split a set of devices from their mirrors, and create a new pool from them.
6537  */
6538 int
6539 spa_vdev_split_mirror(spa_t *spa, char *newname, nvlist_t *config,
6540     nvlist_t *props, boolean_t exp)
6541 {
6542 	int error = 0;
6543 	uint64_t txg, *glist;
6544 	spa_t *newspa;
6545 	uint_t c, children, lastlog;
6546 	nvlist_t **child, *nvl, *tmp;
6547 	dmu_tx_t *tx;
6548 	char *altroot = NULL;
6549 	vdev_t *rvd, **vml = NULL;			/* vdev modify list */
6550 	boolean_t activate_slog;
6551 
6552 	ASSERT(spa_writeable(spa));
6553 
6554 	txg = spa_vdev_enter(spa);
6555 
6556 	ASSERT(MUTEX_HELD(&spa_namespace_lock));
6557 	if (spa_feature_is_active(spa, SPA_FEATURE_POOL_CHECKPOINT)) {
6558 		error = (spa_has_checkpoint(spa)) ?
6559 		    ZFS_ERR_CHECKPOINT_EXISTS : ZFS_ERR_DISCARDING_CHECKPOINT;
6560 		return (spa_vdev_exit(spa, NULL, txg, error));
6561 	}
6562 
6563 	/* clear the log and flush everything up to now */
6564 	activate_slog = spa_passivate_log(spa);
6565 	(void) spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
6566 	error = spa_reset_logs(spa);
6567 	txg = spa_vdev_config_enter(spa);
6568 
6569 	if (activate_slog)
6570 		spa_activate_log(spa);
6571 
6572 	if (error != 0)
6573 		return (spa_vdev_exit(spa, NULL, txg, error));
6574 
6575 	/* check new spa name before going any further */
6576 	if (spa_lookup(newname) != NULL)
6577 		return (spa_vdev_exit(spa, NULL, txg, EEXIST));
6578 
6579 	/*
6580 	 * scan through all the children to ensure they're all mirrors
6581 	 */
6582 	if (nvlist_lookup_nvlist(config, ZPOOL_CONFIG_VDEV_TREE, &nvl) != 0 ||
6583 	    nvlist_lookup_nvlist_array(nvl, ZPOOL_CONFIG_CHILDREN, &child,
6584 	    &children) != 0)
6585 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6586 
6587 	/* first, check to ensure we've got the right child count */
6588 	rvd = spa->spa_root_vdev;
6589 	lastlog = 0;
6590 	for (c = 0; c < rvd->vdev_children; c++) {
6591 		vdev_t *vd = rvd->vdev_child[c];
6592 
6593 		/* don't count the holes & logs as children */
6594 		if (vd->vdev_islog || !vdev_is_concrete(vd)) {
6595 			if (lastlog == 0)
6596 				lastlog = c;
6597 			continue;
6598 		}
6599 
6600 		lastlog = 0;
6601 	}
6602 	if (children != (lastlog != 0 ? lastlog : rvd->vdev_children))
6603 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6604 
6605 	/* next, ensure no spare or cache devices are part of the split */
6606 	if (nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_SPARES, &tmp) == 0 ||
6607 	    nvlist_lookup_nvlist(nvl, ZPOOL_CONFIG_L2CACHE, &tmp) == 0)
6608 		return (spa_vdev_exit(spa, NULL, txg, EINVAL));
6609 
6610 	vml = kmem_zalloc(children * sizeof (vdev_t *), KM_SLEEP);
6611 	glist = kmem_zalloc(children * sizeof (uint64_t), KM_SLEEP);
6612 
6613 	/* then, loop over each vdev and validate it */
6614 	for (c = 0; c < children; c++) {
6615 		uint64_t is_hole = 0;
6616 
6617 		(void) nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_IS_HOLE,
6618 		    &is_hole);
6619 
6620 		if (is_hole != 0) {
6621 			if (spa->spa_root_vdev->vdev_child[c]->vdev_ishole ||
6622 			    spa->spa_root_vdev->vdev_child[c]->vdev_islog) {
6623 				continue;
6624 			} else {
6625 				error = SET_ERROR(EINVAL);
6626 				break;
6627 			}
6628 		}
6629 
6630 		/* which disk is going to be split? */
6631 		if (nvlist_lookup_uint64(child[c], ZPOOL_CONFIG_GUID,
6632 		    &glist[c]) != 0) {
6633 			error = SET_ERROR(EINVAL);
6634 			break;
6635 		}
6636 
6637 		/* look it up in the spa */
6638 		vml[c] = spa_lookup_by_guid(spa, glist[c], B_FALSE);
6639 		if (vml[c] == NULL) {
6640 			error = SET_ERROR(ENODEV);
6641 			break;
6642 		}
6643 
6644 		/* make sure there's nothing stopping the split */
6645 		if (vml[c]->vdev_parent->vdev_ops != &vdev_mirror_ops ||
6646 		    vml[c]->vdev_islog ||
6647 		    !vdev_is_concrete(vml[c]) ||
6648 		    vml[c]->vdev_isspare ||
6649 		    vml[c]->vdev_isl2cache ||
6650 		    !vdev_writeable(vml[c]) ||
6651 		    vml[c]->vdev_children != 0 ||
6652 		    vml[c]->vdev_state != VDEV_STATE_HEALTHY ||
6653 		    c != spa->spa_root_vdev->vdev_child[c]->vdev_id) {
6654 			error = SET_ERROR(EINVAL);
6655 			break;
6656 		}
6657 
6658 		if (vdev_dtl_required(vml[c])) {
6659 			error = SET_ERROR(EBUSY);
6660 			break;
6661 		}
6662 
6663 		/* we need certain info from the top level */
6664 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_ARRAY,
6665 		    vml[c]->vdev_top->vdev_ms_array) == 0);
6666 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_METASLAB_SHIFT,
6667 		    vml[c]->vdev_top->vdev_ms_shift) == 0);
6668 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASIZE,
6669 		    vml[c]->vdev_top->vdev_asize) == 0);
6670 		VERIFY(nvlist_add_uint64(child[c], ZPOOL_CONFIG_ASHIFT,
6671 		    vml[c]->vdev_top->vdev_ashift) == 0);
6672 
6673 		/* transfer per-vdev ZAPs */
6674 		ASSERT3U(vml[c]->vdev_leaf_zap, !=, 0);
6675 		VERIFY0(nvlist_add_uint64(child[c],
6676 		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, vml[c]->vdev_leaf_zap));
6677 
6678 		ASSERT3U(vml[c]->vdev_top->vdev_top_zap, !=, 0);
6679 		VERIFY0(nvlist_add_uint64(child[c],
6680 		    ZPOOL_CONFIG_VDEV_TOP_ZAP,
6681 		    vml[c]->vdev_parent->vdev_top_zap));
6682 	}
6683 
6684 	if (error != 0) {
6685 		kmem_free(vml, children * sizeof (vdev_t *));
6686 		kmem_free(glist, children * sizeof (uint64_t));
6687 		return (spa_vdev_exit(spa, NULL, txg, error));
6688 	}
6689 
6690 	/* stop writers from using the disks */
6691 	for (c = 0; c < children; c++) {
6692 		if (vml[c] != NULL)
6693 			vml[c]->vdev_offline = B_TRUE;
6694 	}
6695 	vdev_reopen(spa->spa_root_vdev);
6696 
6697 	/*
6698 	 * Temporarily record the splitting vdevs in the spa config.  This
6699 	 * will disappear once the config is regenerated.
6700 	 */
6701 	VERIFY(nvlist_alloc(&nvl, NV_UNIQUE_NAME, KM_SLEEP) == 0);
6702 	VERIFY(nvlist_add_uint64_array(nvl, ZPOOL_CONFIG_SPLIT_LIST,
6703 	    glist, children) == 0);
6704 	kmem_free(glist, children * sizeof (uint64_t));
6705 
6706 	mutex_enter(&spa->spa_props_lock);
6707 	VERIFY(nvlist_add_nvlist(spa->spa_config, ZPOOL_CONFIG_SPLIT,
6708 	    nvl) == 0);
6709 	mutex_exit(&spa->spa_props_lock);
6710 	spa->spa_config_splitting = nvl;
6711 	vdev_config_dirty(spa->spa_root_vdev);
6712 
6713 	/* configure and create the new pool */
6714 	VERIFY(nvlist_add_string(config, ZPOOL_CONFIG_POOL_NAME, newname) == 0);
6715 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_STATE,
6716 	    exp ? POOL_STATE_EXPORTED : POOL_STATE_ACTIVE) == 0);
6717 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_VERSION,
6718 	    spa_version(spa)) == 0);
6719 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_TXG,
6720 	    spa->spa_config_txg) == 0);
6721 	VERIFY(nvlist_add_uint64(config, ZPOOL_CONFIG_POOL_GUID,
6722 	    spa_generate_guid(NULL)) == 0);
6723 	VERIFY0(nvlist_add_boolean(config, ZPOOL_CONFIG_HAS_PER_VDEV_ZAPS));
6724 	(void) nvlist_lookup_string(props,
6725 	    zpool_prop_to_name(ZPOOL_PROP_ALTROOT), &altroot);
6726 
6727 	/* add the new pool to the namespace */
6728 	newspa = spa_add(newname, config, altroot);
6729 	newspa->spa_avz_action = AVZ_ACTION_REBUILD;
6730 	newspa->spa_config_txg = spa->spa_config_txg;
6731 	spa_set_log_state(newspa, SPA_LOG_CLEAR);
6732 
6733 	/* release the spa config lock, retaining the namespace lock */
6734 	spa_vdev_config_exit(spa, NULL, txg, 0, FTAG);
6735 
6736 	if (zio_injection_enabled)
6737 		zio_handle_panic_injection(spa, FTAG, 1);
6738 
6739 	spa_activate(newspa, spa_mode_global);
6740 	spa_async_suspend(newspa);
6741 
6742 	for (c = 0; c < children; c++) {
6743 		if (vml[c] != NULL) {
6744 			/*
6745 			 * Temporarily stop the initializing activity. We set
6746 			 * the state to ACTIVE so that we know to resume
6747 			 * the initializing once the split has completed.
6748 			 */
6749 			mutex_enter(&vml[c]->vdev_initialize_lock);
6750 			vdev_initialize_stop(vml[c], VDEV_INITIALIZE_ACTIVE);
6751 			mutex_exit(&vml[c]->vdev_initialize_lock);
6752 		}
6753 	}
6754 
6755 	newspa->spa_config_source = SPA_CONFIG_SRC_SPLIT;
6756 
6757 	/* create the new pool from the disks of the original pool */
6758 	error = spa_load(newspa, SPA_LOAD_IMPORT, SPA_IMPORT_ASSEMBLE);
6759 	if (error)
6760 		goto out;
6761 
6762 	/* if that worked, generate a real config for the new pool */
6763 	if (newspa->spa_root_vdev != NULL) {
6764 		VERIFY(nvlist_alloc(&newspa->spa_config_splitting,
6765 		    NV_UNIQUE_NAME, KM_SLEEP) == 0);
6766 		VERIFY(nvlist_add_uint64(newspa->spa_config_splitting,
6767 		    ZPOOL_CONFIG_SPLIT_GUID, spa_guid(spa)) == 0);
6768 		spa_config_set(newspa, spa_config_generate(newspa, NULL, -1ULL,
6769 		    B_TRUE));
6770 	}
6771 
6772 	/* set the props */
6773 	if (props != NULL) {
6774 		spa_configfile_set(newspa, props, B_FALSE);
6775 		error = spa_prop_set(newspa, props);
6776 		if (error)
6777 			goto out;
6778 	}
6779 
6780 	/* flush everything */
6781 	txg = spa_vdev_config_enter(newspa);
6782 	vdev_config_dirty(newspa->spa_root_vdev);
6783 	(void) spa_vdev_config_exit(newspa, NULL, txg, 0, FTAG);
6784 
6785 	if (zio_injection_enabled)
6786 		zio_handle_panic_injection(spa, FTAG, 2);
6787 
6788 	spa_async_resume(newspa);
6789 
6790 	/* finally, update the original pool's config */
6791 	txg = spa_vdev_config_enter(spa);
6792 	tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
6793 	error = dmu_tx_assign(tx, TXG_WAIT);
6794 	if (error != 0)
6795 		dmu_tx_abort(tx);
6796 	for (c = 0; c < children; c++) {
6797 		if (vml[c] != NULL) {
6798 			vdev_split(vml[c]);
6799 			if (error == 0)
6800 				spa_history_log_internal(spa, "detach", tx,
6801 				    "vdev=%s", vml[c]->vdev_path);
6802 
6803 			vdev_free(vml[c]);
6804 		}
6805 	}
6806 	spa->spa_avz_action = AVZ_ACTION_REBUILD;
6807 	vdev_config_dirty(spa->spa_root_vdev);
6808 	spa->spa_config_splitting = NULL;
6809 	nvlist_free(nvl);
6810 	if (error == 0)
6811 		dmu_tx_commit(tx);
6812 	(void) spa_vdev_exit(spa, NULL, txg, 0);
6813 
6814 	if (zio_injection_enabled)
6815 		zio_handle_panic_injection(spa, FTAG, 3);
6816 
6817 	/* split is complete; log a history record */
6818 	spa_history_log_internal(newspa, "split", NULL,
6819 	    "from pool %s", spa_name(spa));
6820 
6821 	kmem_free(vml, children * sizeof (vdev_t *));
6822 
6823 	/* if we're not going to mount the filesystems in userland, export */
6824 	if (exp)
6825 		error = spa_export_common(newname, POOL_STATE_EXPORTED, NULL,
6826 		    B_FALSE, B_FALSE);
6827 
6828 	return (error);
6829 
6830 out:
6831 	spa_unload(newspa);
6832 	spa_deactivate(newspa);
6833 	spa_remove(newspa);
6834 
6835 	txg = spa_vdev_config_enter(spa);
6836 
6837 	/* re-online all offlined disks */
6838 	for (c = 0; c < children; c++) {
6839 		if (vml[c] != NULL)
6840 			vml[c]->vdev_offline = B_FALSE;
6841 	}
6842 
6843 	/* restart initializing disks as necessary */
6844 	spa_async_request(spa, SPA_ASYNC_INITIALIZE_RESTART);
6845 
6846 	vdev_reopen(spa->spa_root_vdev);
6847 
6848 	nvlist_free(spa->spa_config_splitting);
6849 	spa->spa_config_splitting = NULL;
6850 	(void) spa_vdev_exit(spa, NULL, txg, error);
6851 
6852 	kmem_free(vml, children * sizeof (vdev_t *));
6853 	return (error);
6854 }
6855 
6856 /*
6857  * Find any device that's done replacing, or a vdev marked 'unspare' that's
6858  * currently spared, so we can detach it.
6859  */
6860 static vdev_t *
6861 spa_vdev_resilver_done_hunt(vdev_t *vd)
6862 {
6863 	vdev_t *newvd, *oldvd;
6864 
6865 	for (int c = 0; c < vd->vdev_children; c++) {
6866 		oldvd = spa_vdev_resilver_done_hunt(vd->vdev_child[c]);
6867 		if (oldvd != NULL)
6868 			return (oldvd);
6869 	}
6870 
6871 	/*
6872 	 * Check for a completed replacement.  We always consider the first
6873 	 * vdev in the list to be the oldest vdev, and the last one to be
6874 	 * the newest (see spa_vdev_attach() for how that works).  In
6875 	 * the case where the newest vdev is faulted, we will not automatically
6876 	 * remove it after a resilver completes.  This is OK as it will require
6877 	 * user intervention to determine which disk the admin wishes to keep.
6878 	 */
6879 	if (vd->vdev_ops == &vdev_replacing_ops) {
6880 		ASSERT(vd->vdev_children > 1);
6881 
6882 		newvd = vd->vdev_child[vd->vdev_children - 1];
6883 		oldvd = vd->vdev_child[0];
6884 
6885 		if (vdev_dtl_empty(newvd, DTL_MISSING) &&
6886 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
6887 		    !vdev_dtl_required(oldvd))
6888 			return (oldvd);
6889 	}
6890 
6891 	/*
6892 	 * Check for a completed resilver with the 'unspare' flag set.
6893 	 * Also potentially update faulted state.
6894 	 */
6895 	if (vd->vdev_ops == &vdev_spare_ops) {
6896 		vdev_t *first = vd->vdev_child[0];
6897 		vdev_t *last = vd->vdev_child[vd->vdev_children - 1];
6898 
6899 		if (last->vdev_unspare) {
6900 			oldvd = first;
6901 			newvd = last;
6902 		} else if (first->vdev_unspare) {
6903 			oldvd = last;
6904 			newvd = first;
6905 		} else {
6906 			oldvd = NULL;
6907 		}
6908 
6909 		if (oldvd != NULL &&
6910 		    vdev_dtl_empty(newvd, DTL_MISSING) &&
6911 		    vdev_dtl_empty(newvd, DTL_OUTAGE) &&
6912 		    !vdev_dtl_required(oldvd))
6913 			return (oldvd);
6914 
6915 		vdev_propagate_state(vd);
6916 
6917 		/*
6918 		 * If there are more than two spares attached to a disk,
6919 		 * and those spares are not required, then we want to
6920 		 * attempt to free them up now so that they can be used
6921 		 * by other pools.  Once we're back down to a single
6922 		 * disk+spare, we stop removing them.
6923 		 */
6924 		if (vd->vdev_children > 2) {
6925 			newvd = vd->vdev_child[1];
6926 
6927 			if (newvd->vdev_isspare && last->vdev_isspare &&
6928 			    vdev_dtl_empty(last, DTL_MISSING) &&
6929 			    vdev_dtl_empty(last, DTL_OUTAGE) &&
6930 			    !vdev_dtl_required(newvd))
6931 				return (newvd);
6932 		}
6933 	}
6934 
6935 	return (NULL);
6936 }
6937 
6938 static void
6939 spa_vdev_resilver_done(spa_t *spa)
6940 {
6941 	vdev_t *vd, *pvd, *ppvd;
6942 	uint64_t guid, sguid, pguid, ppguid;
6943 
6944 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6945 
6946 	while ((vd = spa_vdev_resilver_done_hunt(spa->spa_root_vdev)) != NULL) {
6947 		pvd = vd->vdev_parent;
6948 		ppvd = pvd->vdev_parent;
6949 		guid = vd->vdev_guid;
6950 		pguid = pvd->vdev_guid;
6951 		ppguid = ppvd->vdev_guid;
6952 		sguid = 0;
6953 		/*
6954 		 * If we have just finished replacing a hot spared device, then
6955 		 * we need to detach the parent's first child (the original hot
6956 		 * spare) as well.
6957 		 */
6958 		if (ppvd->vdev_ops == &vdev_spare_ops && pvd->vdev_id == 0 &&
6959 		    ppvd->vdev_children == 2) {
6960 			ASSERT(pvd->vdev_ops == &vdev_replacing_ops);
6961 			sguid = ppvd->vdev_child[1]->vdev_guid;
6962 		}
6963 		ASSERT(vd->vdev_resilver_txg == 0 || !vdev_dtl_required(vd));
6964 
6965 		spa_config_exit(spa, SCL_ALL, FTAG);
6966 		if (spa_vdev_detach(spa, guid, pguid, B_TRUE) != 0)
6967 			return;
6968 		if (sguid && spa_vdev_detach(spa, sguid, ppguid, B_TRUE) != 0)
6969 			return;
6970 		spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
6971 	}
6972 
6973 	spa_config_exit(spa, SCL_ALL, FTAG);
6974 }
6975 
6976 /*
6977  * Update the stored path or FRU for this vdev.
6978  */
6979 int
6980 spa_vdev_set_common(spa_t *spa, uint64_t guid, const char *value,
6981     boolean_t ispath)
6982 {
6983 	vdev_t *vd;
6984 	boolean_t sync = B_FALSE;
6985 
6986 	ASSERT(spa_writeable(spa));
6987 
6988 	spa_vdev_state_enter(spa, SCL_ALL);
6989 
6990 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
6991 		return (spa_vdev_state_exit(spa, NULL, ENOENT));
6992 
6993 	if (!vd->vdev_ops->vdev_op_leaf)
6994 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
6995 
6996 	if (ispath) {
6997 		if (strcmp(value, vd->vdev_path) != 0) {
6998 			spa_strfree(vd->vdev_path);
6999 			vd->vdev_path = spa_strdup(value);
7000 			sync = B_TRUE;
7001 		}
7002 	} else {
7003 		if (vd->vdev_fru == NULL) {
7004 			vd->vdev_fru = spa_strdup(value);
7005 			sync = B_TRUE;
7006 		} else if (strcmp(value, vd->vdev_fru) != 0) {
7007 			spa_strfree(vd->vdev_fru);
7008 			vd->vdev_fru = spa_strdup(value);
7009 			sync = B_TRUE;
7010 		}
7011 	}
7012 
7013 	return (spa_vdev_state_exit(spa, sync ? vd : NULL, 0));
7014 }
7015 
7016 int
7017 spa_vdev_setpath(spa_t *spa, uint64_t guid, const char *newpath)
7018 {
7019 	return (spa_vdev_set_common(spa, guid, newpath, B_TRUE));
7020 }
7021 
7022 int
7023 spa_vdev_setfru(spa_t *spa, uint64_t guid, const char *newfru)
7024 {
7025 	return (spa_vdev_set_common(spa, guid, newfru, B_FALSE));
7026 }
7027 
7028 /*
7029  * ==========================================================================
7030  * SPA Scanning
7031  * ==========================================================================
7032  */
7033 int
7034 spa_scrub_pause_resume(spa_t *spa, pool_scrub_cmd_t cmd)
7035 {
7036 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
7037 
7038 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
7039 		return (SET_ERROR(EBUSY));
7040 
7041 	return (dsl_scrub_set_pause_resume(spa->spa_dsl_pool, cmd));
7042 }
7043 
7044 int
7045 spa_scan_stop(spa_t *spa)
7046 {
7047 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
7048 	if (dsl_scan_resilvering(spa->spa_dsl_pool))
7049 		return (SET_ERROR(EBUSY));
7050 	return (dsl_scan_cancel(spa->spa_dsl_pool));
7051 }
7052 
7053 int
7054 spa_scan(spa_t *spa, pool_scan_func_t func)
7055 {
7056 	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == 0);
7057 
7058 	if (func >= POOL_SCAN_FUNCS || func == POOL_SCAN_NONE)
7059 		return (SET_ERROR(ENOTSUP));
7060 
7061 	if (func == POOL_SCAN_RESILVER &&
7062 	    !spa_feature_is_enabled(spa, SPA_FEATURE_RESILVER_DEFER))
7063 		return (SET_ERROR(ENOTSUP));
7064 
7065 	/*
7066 	 * If a resilver was requested, but there is no DTL on a
7067 	 * writeable leaf device, we have nothing to do.
7068 	 */
7069 	if (func == POOL_SCAN_RESILVER &&
7070 	    !vdev_resilver_needed(spa->spa_root_vdev, NULL, NULL)) {
7071 		spa_async_request(spa, SPA_ASYNC_RESILVER_DONE);
7072 		return (0);
7073 	}
7074 
7075 	return (dsl_scan(spa->spa_dsl_pool, func));
7076 }
7077 
7078 /*
7079  * ==========================================================================
7080  * SPA async task processing
7081  * ==========================================================================
7082  */
7083 
7084 static void
7085 spa_async_remove(spa_t *spa, vdev_t *vd)
7086 {
7087 	if (vd->vdev_remove_wanted) {
7088 		vd->vdev_remove_wanted = B_FALSE;
7089 		vd->vdev_delayed_close = B_FALSE;
7090 		vdev_set_state(vd, B_FALSE, VDEV_STATE_REMOVED, VDEV_AUX_NONE);
7091 
7092 		/*
7093 		 * We want to clear the stats, but we don't want to do a full
7094 		 * vdev_clear() as that will cause us to throw away
7095 		 * degraded/faulted state as well as attempt to reopen the
7096 		 * device, all of which is a waste.
7097 		 */
7098 		vd->vdev_stat.vs_read_errors = 0;
7099 		vd->vdev_stat.vs_write_errors = 0;
7100 		vd->vdev_stat.vs_checksum_errors = 0;
7101 
7102 		vdev_state_dirty(vd->vdev_top);
7103 	}
7104 
7105 	for (int c = 0; c < vd->vdev_children; c++)
7106 		spa_async_remove(spa, vd->vdev_child[c]);
7107 }
7108 
7109 static void
7110 spa_async_probe(spa_t *spa, vdev_t *vd)
7111 {
7112 	if (vd->vdev_probe_wanted) {
7113 		vd->vdev_probe_wanted = B_FALSE;
7114 		vdev_reopen(vd);	/* vdev_open() does the actual probe */
7115 	}
7116 
7117 	for (int c = 0; c < vd->vdev_children; c++)
7118 		spa_async_probe(spa, vd->vdev_child[c]);
7119 }
7120 
7121 static void
7122 spa_async_autoexpand(spa_t *spa, vdev_t *vd)
7123 {
7124 	sysevent_id_t eid;
7125 	nvlist_t *attr;
7126 	char *physpath;
7127 
7128 	if (!spa->spa_autoexpand)
7129 		return;
7130 
7131 	for (int c = 0; c < vd->vdev_children; c++) {
7132 		vdev_t *cvd = vd->vdev_child[c];
7133 		spa_async_autoexpand(spa, cvd);
7134 	}
7135 
7136 	if (!vd->vdev_ops->vdev_op_leaf || vd->vdev_physpath == NULL)
7137 		return;
7138 
7139 	physpath = kmem_zalloc(MAXPATHLEN, KM_SLEEP);
7140 	(void) snprintf(physpath, MAXPATHLEN, "/devices%s", vd->vdev_physpath);
7141 
7142 	VERIFY(nvlist_alloc(&attr, NV_UNIQUE_NAME, KM_SLEEP) == 0);
7143 	VERIFY(nvlist_add_string(attr, DEV_PHYS_PATH, physpath) == 0);
7144 
7145 	(void) ddi_log_sysevent(zfs_dip, SUNW_VENDOR, EC_DEV_STATUS,
7146 	    ESC_DEV_DLE, attr, &eid, DDI_SLEEP);
7147 
7148 	nvlist_free(attr);
7149 	kmem_free(physpath, MAXPATHLEN);
7150 }
7151 
7152 static void
7153 spa_async_thread(void *arg)
7154 {
7155 	spa_t *spa = (spa_t *)arg;
7156 	dsl_pool_t *dp = spa->spa_dsl_pool;
7157 	int tasks;
7158 
7159 	ASSERT(spa->spa_sync_on);
7160 
7161 	mutex_enter(&spa->spa_async_lock);
7162 	tasks = spa->spa_async_tasks;
7163 	spa->spa_async_tasks = 0;
7164 	mutex_exit(&spa->spa_async_lock);
7165 
7166 	/*
7167 	 * See if the config needs to be updated.
7168 	 */
7169 	if (tasks & SPA_ASYNC_CONFIG_UPDATE) {
7170 		uint64_t old_space, new_space;
7171 
7172 		mutex_enter(&spa_namespace_lock);
7173 		old_space = metaslab_class_get_space(spa_normal_class(spa));
7174 		old_space += metaslab_class_get_space(spa_special_class(spa));
7175 		old_space += metaslab_class_get_space(spa_dedup_class(spa));
7176 
7177 		spa_config_update(spa, SPA_CONFIG_UPDATE_POOL);
7178 
7179 		new_space = metaslab_class_get_space(spa_normal_class(spa));
7180 		new_space += metaslab_class_get_space(spa_special_class(spa));
7181 		new_space += metaslab_class_get_space(spa_dedup_class(spa));
7182 		mutex_exit(&spa_namespace_lock);
7183 
7184 		/*
7185 		 * If the pool grew as a result of the config update,
7186 		 * then log an internal history event.
7187 		 */
7188 		if (new_space != old_space) {
7189 			spa_history_log_internal(spa, "vdev online", NULL,
7190 			    "pool '%s' size: %llu(+%llu)",
7191 			    spa_name(spa), new_space, new_space - old_space);
7192 		}
7193 	}
7194 
7195 	/*
7196 	 * See if any devices need to be marked REMOVED.
7197 	 */
7198 	if (tasks & SPA_ASYNC_REMOVE) {
7199 		spa_vdev_state_enter(spa, SCL_NONE);
7200 		spa_async_remove(spa, spa->spa_root_vdev);
7201 		for (int i = 0; i < spa->spa_l2cache.sav_count; i++)
7202 			spa_async_remove(spa, spa->spa_l2cache.sav_vdevs[i]);
7203 		for (int i = 0; i < spa->spa_spares.sav_count; i++)
7204 			spa_async_remove(spa, spa->spa_spares.sav_vdevs[i]);
7205 		(void) spa_vdev_state_exit(spa, NULL, 0);
7206 	}
7207 
7208 	if ((tasks & SPA_ASYNC_AUTOEXPAND) && !spa_suspended(spa)) {
7209 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
7210 		spa_async_autoexpand(spa, spa->spa_root_vdev);
7211 		spa_config_exit(spa, SCL_CONFIG, FTAG);
7212 	}
7213 
7214 	/*
7215 	 * See if any devices need to be probed.
7216 	 */
7217 	if (tasks & SPA_ASYNC_PROBE) {
7218 		spa_vdev_state_enter(spa, SCL_NONE);
7219 		spa_async_probe(spa, spa->spa_root_vdev);
7220 		(void) spa_vdev_state_exit(spa, NULL, 0);
7221 	}
7222 
7223 	/*
7224 	 * If any devices are done replacing, detach them.
7225 	 */
7226 	if (tasks & SPA_ASYNC_RESILVER_DONE)
7227 		spa_vdev_resilver_done(spa);
7228 
7229 	/*
7230 	 * Kick off a resilver.
7231 	 */
7232 	if (tasks & SPA_ASYNC_RESILVER &&
7233 	    (!dsl_scan_resilvering(dp) ||
7234 	    !spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_RESILVER_DEFER)))
7235 		dsl_resilver_restart(dp, 0);
7236 
7237 	if (tasks & SPA_ASYNC_INITIALIZE_RESTART) {
7238 		mutex_enter(&spa_namespace_lock);
7239 		spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
7240 		vdev_initialize_restart(spa->spa_root_vdev);
7241 		spa_config_exit(spa, SCL_CONFIG, FTAG);
7242 		mutex_exit(&spa_namespace_lock);
7243 	}
7244 
7245 	/*
7246 	 * Let the world know that we're done.
7247 	 */
7248 	mutex_enter(&spa->spa_async_lock);
7249 	spa->spa_async_thread = NULL;
7250 	cv_broadcast(&spa->spa_async_cv);
7251 	mutex_exit(&spa->spa_async_lock);
7252 	thread_exit();
7253 }
7254 
7255 void
7256 spa_async_suspend(spa_t *spa)
7257 {
7258 	mutex_enter(&spa->spa_async_lock);
7259 	spa->spa_async_suspended++;
7260 	while (spa->spa_async_thread != NULL)
7261 		cv_wait(&spa->spa_async_cv, &spa->spa_async_lock);
7262 	mutex_exit(&spa->spa_async_lock);
7263 
7264 	spa_vdev_remove_suspend(spa);
7265 
7266 	zthr_t *condense_thread = spa->spa_condense_zthr;
7267 	if (condense_thread != NULL)
7268 		zthr_cancel(condense_thread);
7269 
7270 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
7271 	if (discard_thread != NULL)
7272 		zthr_cancel(discard_thread);
7273 }
7274 
7275 void
7276 spa_async_resume(spa_t *spa)
7277 {
7278 	mutex_enter(&spa->spa_async_lock);
7279 	ASSERT(spa->spa_async_suspended != 0);
7280 	spa->spa_async_suspended--;
7281 	mutex_exit(&spa->spa_async_lock);
7282 	spa_restart_removal(spa);
7283 
7284 	zthr_t *condense_thread = spa->spa_condense_zthr;
7285 	if (condense_thread != NULL)
7286 		zthr_resume(condense_thread);
7287 
7288 	zthr_t *discard_thread = spa->spa_checkpoint_discard_zthr;
7289 	if (discard_thread != NULL)
7290 		zthr_resume(discard_thread);
7291 }
7292 
7293 static boolean_t
7294 spa_async_tasks_pending(spa_t *spa)
7295 {
7296 	uint_t non_config_tasks;
7297 	uint_t config_task;
7298 	boolean_t config_task_suspended;
7299 
7300 	non_config_tasks = spa->spa_async_tasks & ~SPA_ASYNC_CONFIG_UPDATE;
7301 	config_task = spa->spa_async_tasks & SPA_ASYNC_CONFIG_UPDATE;
7302 	if (spa->spa_ccw_fail_time == 0) {
7303 		config_task_suspended = B_FALSE;
7304 	} else {
7305 		config_task_suspended =
7306 		    (gethrtime() - spa->spa_ccw_fail_time) <
7307 		    (zfs_ccw_retry_interval * NANOSEC);
7308 	}
7309 
7310 	return (non_config_tasks || (config_task && !config_task_suspended));
7311 }
7312 
7313 static void
7314 spa_async_dispatch(spa_t *spa)
7315 {
7316 	mutex_enter(&spa->spa_async_lock);
7317 	if (spa_async_tasks_pending(spa) &&
7318 	    !spa->spa_async_suspended &&
7319 	    spa->spa_async_thread == NULL &&
7320 	    rootdir != NULL)
7321 		spa->spa_async_thread = thread_create(NULL, 0,
7322 		    spa_async_thread, spa, 0, &p0, TS_RUN, maxclsyspri);
7323 	mutex_exit(&spa->spa_async_lock);
7324 }
7325 
7326 void
7327 spa_async_request(spa_t *spa, int task)
7328 {
7329 	zfs_dbgmsg("spa=%s async request task=%u", spa->spa_name, task);
7330 	mutex_enter(&spa->spa_async_lock);
7331 	spa->spa_async_tasks |= task;
7332 	mutex_exit(&spa->spa_async_lock);
7333 }
7334 
7335 /*
7336  * ==========================================================================
7337  * SPA syncing routines
7338  * ==========================================================================
7339  */
7340 
7341 static int
7342 bpobj_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
7343 {
7344 	bpobj_t *bpo = arg;
7345 	bpobj_enqueue(bpo, bp, tx);
7346 	return (0);
7347 }
7348 
7349 static int
7350 spa_free_sync_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
7351 {
7352 	zio_t *zio = arg;
7353 
7354 	zio_nowait(zio_free_sync(zio, zio->io_spa, dmu_tx_get_txg(tx), bp,
7355 	    zio->io_flags));
7356 	return (0);
7357 }
7358 
7359 /*
7360  * Note: this simple function is not inlined to make it easier to dtrace the
7361  * amount of time spent syncing frees.
7362  */
7363 static void
7364 spa_sync_frees(spa_t *spa, bplist_t *bpl, dmu_tx_t *tx)
7365 {
7366 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
7367 	bplist_iterate(bpl, spa_free_sync_cb, zio, tx);
7368 	VERIFY(zio_wait(zio) == 0);
7369 }
7370 
7371 /*
7372  * Note: this simple function is not inlined to make it easier to dtrace the
7373  * amount of time spent syncing deferred frees.
7374  */
7375 static void
7376 spa_sync_deferred_frees(spa_t *spa, dmu_tx_t *tx)
7377 {
7378 	if (spa_sync_pass(spa) != 1)
7379 		return;
7380 
7381 	zio_t *zio = zio_root(spa, NULL, NULL, 0);
7382 	VERIFY3U(bpobj_iterate(&spa->spa_deferred_bpobj,
7383 	    spa_free_sync_cb, zio, tx), ==, 0);
7384 	VERIFY0(zio_wait(zio));
7385 }
7386 
7387 
7388 static void
7389 spa_sync_nvlist(spa_t *spa, uint64_t obj, nvlist_t *nv, dmu_tx_t *tx)
7390 {
7391 	char *packed = NULL;
7392 	size_t bufsize;
7393 	size_t nvsize = 0;
7394 	dmu_buf_t *db;
7395 
7396 	VERIFY(nvlist_size(nv, &nvsize, NV_ENCODE_XDR) == 0);
7397 
7398 	/*
7399 	 * Write full (SPA_CONFIG_BLOCKSIZE) blocks of configuration
7400 	 * information.  This avoids the dmu_buf_will_dirty() path and
7401 	 * saves us a pre-read to get data we don't actually care about.
7402 	 */
7403 	bufsize = P2ROUNDUP((uint64_t)nvsize, SPA_CONFIG_BLOCKSIZE);
7404 	packed = kmem_alloc(bufsize, KM_SLEEP);
7405 
7406 	VERIFY(nvlist_pack(nv, &packed, &nvsize, NV_ENCODE_XDR,
7407 	    KM_SLEEP) == 0);
7408 	bzero(packed + nvsize, bufsize - nvsize);
7409 
7410 	dmu_write(spa->spa_meta_objset, obj, 0, bufsize, packed, tx);
7411 
7412 	kmem_free(packed, bufsize);
7413 
7414 	VERIFY(0 == dmu_bonus_hold(spa->spa_meta_objset, obj, FTAG, &db));
7415 	dmu_buf_will_dirty(db, tx);
7416 	*(uint64_t *)db->db_data = nvsize;
7417 	dmu_buf_rele(db, FTAG);
7418 }
7419 
7420 static void
7421 spa_sync_aux_dev(spa_t *spa, spa_aux_vdev_t *sav, dmu_tx_t *tx,
7422     const char *config, const char *entry)
7423 {
7424 	nvlist_t *nvroot;
7425 	nvlist_t **list;
7426 	int i;
7427 
7428 	if (!sav->sav_sync)
7429 		return;
7430 
7431 	/*
7432 	 * Update the MOS nvlist describing the list of available devices.
7433 	 * spa_validate_aux() will have already made sure this nvlist is
7434 	 * valid and the vdevs are labeled appropriately.
7435 	 */
7436 	if (sav->sav_object == 0) {
7437 		sav->sav_object = dmu_object_alloc(spa->spa_meta_objset,
7438 		    DMU_OT_PACKED_NVLIST, 1 << 14, DMU_OT_PACKED_NVLIST_SIZE,
7439 		    sizeof (uint64_t), tx);
7440 		VERIFY(zap_update(spa->spa_meta_objset,
7441 		    DMU_POOL_DIRECTORY_OBJECT, entry, sizeof (uint64_t), 1,
7442 		    &sav->sav_object, tx) == 0);
7443 	}
7444 
7445 	VERIFY(nvlist_alloc(&nvroot, NV_UNIQUE_NAME, KM_SLEEP) == 0);
7446 	if (sav->sav_count == 0) {
7447 		VERIFY(nvlist_add_nvlist_array(nvroot, config, NULL, 0) == 0);
7448 	} else {
7449 		list = kmem_alloc(sav->sav_count * sizeof (void *), KM_SLEEP);
7450 		for (i = 0; i < sav->sav_count; i++)
7451 			list[i] = vdev_config_generate(spa, sav->sav_vdevs[i],
7452 			    B_FALSE, VDEV_CONFIG_L2CACHE);
7453 		VERIFY(nvlist_add_nvlist_array(nvroot, config, list,
7454 		    sav->sav_count) == 0);
7455 		for (i = 0; i < sav->sav_count; i++)
7456 			nvlist_free(list[i]);
7457 		kmem_free(list, sav->sav_count * sizeof (void *));
7458 	}
7459 
7460 	spa_sync_nvlist(spa, sav->sav_object, nvroot, tx);
7461 	nvlist_free(nvroot);
7462 
7463 	sav->sav_sync = B_FALSE;
7464 }
7465 
7466 /*
7467  * Rebuild spa's all-vdev ZAP from the vdev ZAPs indicated in each vdev_t.
7468  * The all-vdev ZAP must be empty.
7469  */
7470 static void
7471 spa_avz_build(vdev_t *vd, uint64_t avz, dmu_tx_t *tx)
7472 {
7473 	spa_t *spa = vd->vdev_spa;
7474 	if (vd->vdev_top_zap != 0) {
7475 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
7476 		    vd->vdev_top_zap, tx));
7477 	}
7478 	if (vd->vdev_leaf_zap != 0) {
7479 		VERIFY0(zap_add_int(spa->spa_meta_objset, avz,
7480 		    vd->vdev_leaf_zap, tx));
7481 	}
7482 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
7483 		spa_avz_build(vd->vdev_child[i], avz, tx);
7484 	}
7485 }
7486 
7487 static void
7488 spa_sync_config_object(spa_t *spa, dmu_tx_t *tx)
7489 {
7490 	nvlist_t *config;
7491 
7492 	/*
7493 	 * If the pool is being imported from a pre-per-vdev-ZAP version of ZFS,
7494 	 * its config may not be dirty but we still need to build per-vdev ZAPs.
7495 	 * Similarly, if the pool is being assembled (e.g. after a split), we
7496 	 * need to rebuild the AVZ although the config may not be dirty.
7497 	 */
7498 	if (list_is_empty(&spa->spa_config_dirty_list) &&
7499 	    spa->spa_avz_action == AVZ_ACTION_NONE)
7500 		return;
7501 
7502 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
7503 
7504 	ASSERT(spa->spa_avz_action == AVZ_ACTION_NONE ||
7505 	    spa->spa_avz_action == AVZ_ACTION_INITIALIZE ||
7506 	    spa->spa_all_vdev_zaps != 0);
7507 
7508 	if (spa->spa_avz_action == AVZ_ACTION_REBUILD) {
7509 		/* Make and build the new AVZ */
7510 		uint64_t new_avz = zap_create(spa->spa_meta_objset,
7511 		    DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx);
7512 		spa_avz_build(spa->spa_root_vdev, new_avz, tx);
7513 
7514 		/* Diff old AVZ with new one */
7515 		zap_cursor_t zc;
7516 		zap_attribute_t za;
7517 
7518 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
7519 		    spa->spa_all_vdev_zaps);
7520 		    zap_cursor_retrieve(&zc, &za) == 0;
7521 		    zap_cursor_advance(&zc)) {
7522 			uint64_t vdzap = za.za_first_integer;
7523 			if (zap_lookup_int(spa->spa_meta_objset, new_avz,
7524 			    vdzap) == ENOENT) {
7525 				/*
7526 				 * ZAP is listed in old AVZ but not in new one;
7527 				 * destroy it
7528 				 */
7529 				VERIFY0(zap_destroy(spa->spa_meta_objset, vdzap,
7530 				    tx));
7531 			}
7532 		}
7533 
7534 		zap_cursor_fini(&zc);
7535 
7536 		/* Destroy the old AVZ */
7537 		VERIFY0(zap_destroy(spa->spa_meta_objset,
7538 		    spa->spa_all_vdev_zaps, tx));
7539 
7540 		/* Replace the old AVZ in the dir obj with the new one */
7541 		VERIFY0(zap_update(spa->spa_meta_objset,
7542 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP,
7543 		    sizeof (new_avz), 1, &new_avz, tx));
7544 
7545 		spa->spa_all_vdev_zaps = new_avz;
7546 	} else if (spa->spa_avz_action == AVZ_ACTION_DESTROY) {
7547 		zap_cursor_t zc;
7548 		zap_attribute_t za;
7549 
7550 		/* Walk through the AVZ and destroy all listed ZAPs */
7551 		for (zap_cursor_init(&zc, spa->spa_meta_objset,
7552 		    spa->spa_all_vdev_zaps);
7553 		    zap_cursor_retrieve(&zc, &za) == 0;
7554 		    zap_cursor_advance(&zc)) {
7555 			uint64_t zap = za.za_first_integer;
7556 			VERIFY0(zap_destroy(spa->spa_meta_objset, zap, tx));
7557 		}
7558 
7559 		zap_cursor_fini(&zc);
7560 
7561 		/* Destroy and unlink the AVZ itself */
7562 		VERIFY0(zap_destroy(spa->spa_meta_objset,
7563 		    spa->spa_all_vdev_zaps, tx));
7564 		VERIFY0(zap_remove(spa->spa_meta_objset,
7565 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_VDEV_ZAP_MAP, tx));
7566 		spa->spa_all_vdev_zaps = 0;
7567 	}
7568 
7569 	if (spa->spa_all_vdev_zaps == 0) {
7570 		spa->spa_all_vdev_zaps = zap_create_link(spa->spa_meta_objset,
7571 		    DMU_OTN_ZAP_METADATA, DMU_POOL_DIRECTORY_OBJECT,
7572 		    DMU_POOL_VDEV_ZAP_MAP, tx);
7573 	}
7574 	spa->spa_avz_action = AVZ_ACTION_NONE;
7575 
7576 	/* Create ZAPs for vdevs that don't have them. */
7577 	vdev_construct_zaps(spa->spa_root_vdev, tx);
7578 
7579 	config = spa_config_generate(spa, spa->spa_root_vdev,
7580 	    dmu_tx_get_txg(tx), B_FALSE);
7581 
7582 	/*
7583 	 * If we're upgrading the spa version then make sure that
7584 	 * the config object gets updated with the correct version.
7585 	 */
7586 	if (spa->spa_ubsync.ub_version < spa->spa_uberblock.ub_version)
7587 		fnvlist_add_uint64(config, ZPOOL_CONFIG_VERSION,
7588 		    spa->spa_uberblock.ub_version);
7589 
7590 	spa_config_exit(spa, SCL_STATE, FTAG);
7591 
7592 	nvlist_free(spa->spa_config_syncing);
7593 	spa->spa_config_syncing = config;
7594 
7595 	spa_sync_nvlist(spa, spa->spa_config_object, config, tx);
7596 }
7597 
7598 static void
7599 spa_sync_version(void *arg, dmu_tx_t *tx)
7600 {
7601 	uint64_t *versionp = arg;
7602 	uint64_t version = *versionp;
7603 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
7604 
7605 	/*
7606 	 * Setting the version is special cased when first creating the pool.
7607 	 */
7608 	ASSERT(tx->tx_txg != TXG_INITIAL);
7609 
7610 	ASSERT(SPA_VERSION_IS_SUPPORTED(version));
7611 	ASSERT(version >= spa_version(spa));
7612 
7613 	spa->spa_uberblock.ub_version = version;
7614 	vdev_config_dirty(spa->spa_root_vdev);
7615 	spa_history_log_internal(spa, "set", tx, "version=%lld", version);
7616 }
7617 
7618 /*
7619  * Set zpool properties.
7620  */
7621 static void
7622 spa_sync_props(void *arg, dmu_tx_t *tx)
7623 {
7624 	nvlist_t *nvp = arg;
7625 	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
7626 	objset_t *mos = spa->spa_meta_objset;
7627 	nvpair_t *elem = NULL;
7628 
7629 	mutex_enter(&spa->spa_props_lock);
7630 
7631 	while ((elem = nvlist_next_nvpair(nvp, elem))) {
7632 		uint64_t intval;
7633 		char *strval, *fname;
7634 		zpool_prop_t prop;
7635 		const char *propname;
7636 		zprop_type_t proptype;
7637 		spa_feature_t fid;
7638 
7639 		switch (prop = zpool_name_to_prop(nvpair_name(elem))) {
7640 		case ZPOOL_PROP_INVAL:
7641 			/*
7642 			 * We checked this earlier in spa_prop_validate().
7643 			 */
7644 			ASSERT(zpool_prop_feature(nvpair_name(elem)));
7645 
7646 			fname = strchr(nvpair_name(elem), '@') + 1;
7647 			VERIFY0(zfeature_lookup_name(fname, &fid));
7648 
7649 			spa_feature_enable(spa, fid, tx);
7650 			spa_history_log_internal(spa, "set", tx,
7651 			    "%s=enabled", nvpair_name(elem));
7652 			break;
7653 
7654 		case ZPOOL_PROP_VERSION:
7655 			intval = fnvpair_value_uint64(elem);
7656 			/*
7657 			 * The version is synced seperatly before other
7658 			 * properties and should be correct by now.
7659 			 */
7660 			ASSERT3U(spa_version(spa), >=, intval);
7661 			break;
7662 
7663 		case ZPOOL_PROP_ALTROOT:
7664 			/*
7665 			 * 'altroot' is a non-persistent property. It should
7666 			 * have been set temporarily at creation or import time.
7667 			 */
7668 			ASSERT(spa->spa_root != NULL);
7669 			break;
7670 
7671 		case ZPOOL_PROP_READONLY:
7672 		case ZPOOL_PROP_CACHEFILE:
7673 			/*
7674 			 * 'readonly' and 'cachefile' are also non-persisitent
7675 			 * properties.
7676 			 */
7677 			break;
7678 		case ZPOOL_PROP_COMMENT:
7679 			strval = fnvpair_value_string(elem);
7680 			if (spa->spa_comment != NULL)
7681 				spa_strfree(spa->spa_comment);
7682 			spa->spa_comment = spa_strdup(strval);
7683 			/*
7684 			 * We need to dirty the configuration on all the vdevs
7685 			 * so that their labels get updated.  It's unnecessary
7686 			 * to do this for pool creation since the vdev's
7687 			 * configuratoin has already been dirtied.
7688 			 */
7689 			if (tx->tx_txg != TXG_INITIAL)
7690 				vdev_config_dirty(spa->spa_root_vdev);
7691 			spa_history_log_internal(spa, "set", tx,
7692 			    "%s=%s", nvpair_name(elem), strval);
7693 			break;
7694 		default:
7695 			/*
7696 			 * Set pool property values in the poolprops mos object.
7697 			 */
7698 			if (spa->spa_pool_props_object == 0) {
7699 				spa->spa_pool_props_object =
7700 				    zap_create_link(mos, DMU_OT_POOL_PROPS,
7701 				    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_PROPS,
7702 				    tx);
7703 			}
7704 
7705 			/* normalize the property name */
7706 			propname = zpool_prop_to_name(prop);
7707 			proptype = zpool_prop_get_type(prop);
7708 
7709 			if (nvpair_type(elem) == DATA_TYPE_STRING) {
7710 				ASSERT(proptype == PROP_TYPE_STRING);
7711 				strval = fnvpair_value_string(elem);
7712 				VERIFY0(zap_update(mos,
7713 				    spa->spa_pool_props_object, propname,
7714 				    1, strlen(strval) + 1, strval, tx));
7715 				spa_history_log_internal(spa, "set", tx,
7716 				    "%s=%s", nvpair_name(elem), strval);
7717 			} else if (nvpair_type(elem) == DATA_TYPE_UINT64) {
7718 				intval = fnvpair_value_uint64(elem);
7719 
7720 				if (proptype == PROP_TYPE_INDEX) {
7721 					const char *unused;
7722 					VERIFY0(zpool_prop_index_to_string(
7723 					    prop, intval, &unused));
7724 				}
7725 				VERIFY0(zap_update(mos,
7726 				    spa->spa_pool_props_object, propname,
7727 				    8, 1, &intval, tx));
7728 				spa_history_log_internal(spa, "set", tx,
7729 				    "%s=%lld", nvpair_name(elem), intval);
7730 			} else {
7731 				ASSERT(0); /* not allowed */
7732 			}
7733 
7734 			switch (prop) {
7735 			case ZPOOL_PROP_DELEGATION:
7736 				spa->spa_delegation = intval;
7737 				break;
7738 			case ZPOOL_PROP_BOOTFS:
7739 				spa->spa_bootfs = intval;
7740 				break;
7741 			case ZPOOL_PROP_FAILUREMODE:
7742 				spa->spa_failmode = intval;
7743 				break;
7744 			case ZPOOL_PROP_AUTOEXPAND:
7745 				spa->spa_autoexpand = intval;
7746 				if (tx->tx_txg != TXG_INITIAL)
7747 					spa_async_request(spa,
7748 					    SPA_ASYNC_AUTOEXPAND);
7749 				break;
7750 			case ZPOOL_PROP_MULTIHOST:
7751 				spa->spa_multihost = intval;
7752 				break;
7753 			case ZPOOL_PROP_DEDUPDITTO:
7754 				spa->spa_dedup_ditto = intval;
7755 				break;
7756 			default:
7757 				break;
7758 			}
7759 		}
7760 
7761 	}
7762 
7763 	mutex_exit(&spa->spa_props_lock);
7764 }
7765 
7766 /*
7767  * Perform one-time upgrade on-disk changes.  spa_version() does not
7768  * reflect the new version this txg, so there must be no changes this
7769  * txg to anything that the upgrade code depends on after it executes.
7770  * Therefore this must be called after dsl_pool_sync() does the sync
7771  * tasks.
7772  */
7773 static void
7774 spa_sync_upgrades(spa_t *spa, dmu_tx_t *tx)
7775 {
7776 	if (spa_sync_pass(spa) != 1)
7777 		return;
7778 
7779 	dsl_pool_t *dp = spa->spa_dsl_pool;
7780 	rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
7781 
7782 	if (spa->spa_ubsync.ub_version < SPA_VERSION_ORIGIN &&
7783 	    spa->spa_uberblock.ub_version >= SPA_VERSION_ORIGIN) {
7784 		dsl_pool_create_origin(dp, tx);
7785 
7786 		/* Keeping the origin open increases spa_minref */
7787 		spa->spa_minref += 3;
7788 	}
7789 
7790 	if (spa->spa_ubsync.ub_version < SPA_VERSION_NEXT_CLONES &&
7791 	    spa->spa_uberblock.ub_version >= SPA_VERSION_NEXT_CLONES) {
7792 		dsl_pool_upgrade_clones(dp, tx);
7793 	}
7794 
7795 	if (spa->spa_ubsync.ub_version < SPA_VERSION_DIR_CLONES &&
7796 	    spa->spa_uberblock.ub_version >= SPA_VERSION_DIR_CLONES) {
7797 		dsl_pool_upgrade_dir_clones(dp, tx);
7798 
7799 		/* Keeping the freedir open increases spa_minref */
7800 		spa->spa_minref += 3;
7801 	}
7802 
7803 	if (spa->spa_ubsync.ub_version < SPA_VERSION_FEATURES &&
7804 	    spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
7805 		spa_feature_create_zap_objects(spa, tx);
7806 	}
7807 
7808 	/*
7809 	 * LZ4_COMPRESS feature's behaviour was changed to activate_on_enable
7810 	 * when possibility to use lz4 compression for metadata was added
7811 	 * Old pools that have this feature enabled must be upgraded to have
7812 	 * this feature active
7813 	 */
7814 	if (spa->spa_uberblock.ub_version >= SPA_VERSION_FEATURES) {
7815 		boolean_t lz4_en = spa_feature_is_enabled(spa,
7816 		    SPA_FEATURE_LZ4_COMPRESS);
7817 		boolean_t lz4_ac = spa_feature_is_active(spa,
7818 		    SPA_FEATURE_LZ4_COMPRESS);
7819 
7820 		if (lz4_en && !lz4_ac)
7821 			spa_feature_incr(spa, SPA_FEATURE_LZ4_COMPRESS, tx);
7822 	}
7823 
7824 	/*
7825 	 * If we haven't written the salt, do so now.  Note that the
7826 	 * feature may not be activated yet, but that's fine since
7827 	 * the presence of this ZAP entry is backwards compatible.
7828 	 */
7829 	if (zap_contains(spa->spa_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
7830 	    DMU_POOL_CHECKSUM_SALT) == ENOENT) {
7831 		VERIFY0(zap_add(spa->spa_meta_objset,
7832 		    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_CHECKSUM_SALT, 1,
7833 		    sizeof (spa->spa_cksum_salt.zcs_bytes),
7834 		    spa->spa_cksum_salt.zcs_bytes, tx));
7835 	}
7836 
7837 	rrw_exit(&dp->dp_config_rwlock, FTAG);
7838 }
7839 
7840 static void
7841 vdev_indirect_state_sync_verify(vdev_t *vd)
7842 {
7843 	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
7844 	vdev_indirect_births_t *vib = vd->vdev_indirect_births;
7845 
7846 	if (vd->vdev_ops == &vdev_indirect_ops) {
7847 		ASSERT(vim != NULL);
7848 		ASSERT(vib != NULL);
7849 	}
7850 
7851 	if (vdev_obsolete_sm_object(vd) != 0) {
7852 		ASSERT(vd->vdev_obsolete_sm != NULL);
7853 		ASSERT(vd->vdev_removing ||
7854 		    vd->vdev_ops == &vdev_indirect_ops);
7855 		ASSERT(vdev_indirect_mapping_num_entries(vim) > 0);
7856 		ASSERT(vdev_indirect_mapping_bytes_mapped(vim) > 0);
7857 
7858 		ASSERT3U(vdev_obsolete_sm_object(vd), ==,
7859 		    space_map_object(vd->vdev_obsolete_sm));
7860 		ASSERT3U(vdev_indirect_mapping_bytes_mapped(vim), >=,
7861 		    space_map_allocated(vd->vdev_obsolete_sm));
7862 	}
7863 	ASSERT(vd->vdev_obsolete_segments != NULL);
7864 
7865 	/*
7866 	 * Since frees / remaps to an indirect vdev can only
7867 	 * happen in syncing context, the obsolete segments
7868 	 * tree must be empty when we start syncing.
7869 	 */
7870 	ASSERT0(range_tree_space(vd->vdev_obsolete_segments));
7871 }
7872 
7873 /*
7874  * Set the top-level vdev's max queue depth. Evaluate each top-level's
7875  * async write queue depth in case it changed. The max queue depth will
7876  * not change in the middle of syncing out this txg.
7877  */
7878 static void
7879 spa_sync_adjust_vdev_max_queue_depth(spa_t *spa)
7880 {
7881 	ASSERT(spa_writeable(spa));
7882 
7883 	vdev_t *rvd = spa->spa_root_vdev;
7884 	uint32_t max_queue_depth = zfs_vdev_async_write_max_active *
7885 	    zfs_vdev_queue_depth_pct / 100;
7886 	metaslab_class_t *normal = spa_normal_class(spa);
7887 	metaslab_class_t *special = spa_special_class(spa);
7888 	metaslab_class_t *dedup = spa_dedup_class(spa);
7889 
7890 	uint64_t slots_per_allocator = 0;
7891 	for (int c = 0; c < rvd->vdev_children; c++) {
7892 		vdev_t *tvd = rvd->vdev_child[c];
7893 
7894 		metaslab_group_t *mg = tvd->vdev_mg;
7895 		if (mg == NULL || !metaslab_group_initialized(mg))
7896 			continue;
7897 
7898 		metaslab_class_t *mc = mg->mg_class;
7899 		if (mc != normal && mc != special && mc != dedup)
7900 			continue;
7901 
7902 		/*
7903 		 * It is safe to do a lock-free check here because only async
7904 		 * allocations look at mg_max_alloc_queue_depth, and async
7905 		 * allocations all happen from spa_sync().
7906 		 */
7907 		for (int i = 0; i < spa->spa_alloc_count; i++)
7908 			ASSERT0(zfs_refcount_count(
7909 			    &(mg->mg_alloc_queue_depth[i])));
7910 		mg->mg_max_alloc_queue_depth = max_queue_depth;
7911 
7912 		for (int i = 0; i < spa->spa_alloc_count; i++) {
7913 			mg->mg_cur_max_alloc_queue_depth[i] =
7914 			    zfs_vdev_def_queue_depth;
7915 		}
7916 		slots_per_allocator += zfs_vdev_def_queue_depth;
7917 	}
7918 
7919 	for (int i = 0; i < spa->spa_alloc_count; i++) {
7920 		ASSERT0(zfs_refcount_count(&normal->mc_alloc_slots[i]));
7921 		ASSERT0(zfs_refcount_count(&special->mc_alloc_slots[i]));
7922 		ASSERT0(zfs_refcount_count(&dedup->mc_alloc_slots[i]));
7923 		normal->mc_alloc_max_slots[i] = slots_per_allocator;
7924 		special->mc_alloc_max_slots[i] = slots_per_allocator;
7925 		dedup->mc_alloc_max_slots[i] = slots_per_allocator;
7926 	}
7927 	normal->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
7928 	special->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
7929 	dedup->mc_alloc_throttle_enabled = zio_dva_throttle_enabled;
7930 }
7931 
7932 static void
7933 spa_sync_condense_indirect(spa_t *spa, dmu_tx_t *tx)
7934 {
7935 	ASSERT(spa_writeable(spa));
7936 
7937 	vdev_t *rvd = spa->spa_root_vdev;
7938 	for (int c = 0; c < rvd->vdev_children; c++) {
7939 		vdev_t *vd = rvd->vdev_child[c];
7940 		vdev_indirect_state_sync_verify(vd);
7941 
7942 		if (vdev_indirect_should_condense(vd)) {
7943 			spa_condense_indirect_start_sync(vd, tx);
7944 			break;
7945 		}
7946 	}
7947 }
7948 
7949 static void
7950 spa_sync_iterate_to_convergence(spa_t *spa, dmu_tx_t *tx)
7951 {
7952 	objset_t *mos = spa->spa_meta_objset;
7953 	dsl_pool_t *dp = spa->spa_dsl_pool;
7954 	uint64_t txg = tx->tx_txg;
7955 	bplist_t *free_bpl = &spa->spa_free_bplist[txg & TXG_MASK];
7956 
7957 	do {
7958 		int pass = ++spa->spa_sync_pass;
7959 
7960 		spa_sync_config_object(spa, tx);
7961 		spa_sync_aux_dev(spa, &spa->spa_spares, tx,
7962 		    ZPOOL_CONFIG_SPARES, DMU_POOL_SPARES);
7963 		spa_sync_aux_dev(spa, &spa->spa_l2cache, tx,
7964 		    ZPOOL_CONFIG_L2CACHE, DMU_POOL_L2CACHE);
7965 		spa_errlog_sync(spa, txg);
7966 		dsl_pool_sync(dp, txg);
7967 
7968 		if (pass < zfs_sync_pass_deferred_free) {
7969 			spa_sync_frees(spa, free_bpl, tx);
7970 		} else {
7971 			/*
7972 			 * We can not defer frees in pass 1, because
7973 			 * we sync the deferred frees later in pass 1.
7974 			 */
7975 			ASSERT3U(pass, >, 1);
7976 			bplist_iterate(free_bpl, bpobj_enqueue_cb,
7977 			    &spa->spa_deferred_bpobj, tx);
7978 		}
7979 
7980 		ddt_sync(spa, txg);
7981 		dsl_scan_sync(dp, tx);
7982 		svr_sync(spa, tx);
7983 		spa_sync_upgrades(spa, tx);
7984 
7985 		vdev_t *vd = NULL;
7986 		while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, txg))
7987 		    != NULL)
7988 			vdev_sync(vd, txg);
7989 
7990 		/*
7991 		 * Note: We need to check if the MOS is dirty because we could
7992 		 * have marked the MOS dirty without updating the uberblock
7993 		 * (e.g. if we have sync tasks but no dirty user data). We need
7994 		 * to check the uberblock's rootbp because it is updated if we
7995 		 * have synced out dirty data (though in this case the MOS will
7996 		 * most likely also be dirty due to second order effects, we
7997 		 * don't want to rely on that here).
7998 		 */
7999 		if (pass == 1 &&
8000 		    spa->spa_uberblock.ub_rootbp.blk_birth < txg &&
8001 		    !dmu_objset_is_dirty(mos, txg)) {
8002 			/*
8003 			 * Nothing changed on the first pass, therefore this
8004 			 * TXG is a no-op. Avoid syncing deferred frees, so
8005 			 * that we can keep this TXG as a no-op.
8006 			 */
8007 			ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
8008 			ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
8009 			ASSERT(txg_list_empty(&dp->dp_sync_tasks, txg));
8010 			ASSERT(txg_list_empty(&dp->dp_early_sync_tasks, txg));
8011 			break;
8012 		}
8013 
8014 		spa_sync_deferred_frees(spa, tx);
8015 	} while (dmu_objset_is_dirty(mos, txg));
8016 }
8017 
8018 /*
8019  * Rewrite the vdev configuration (which includes the uberblock) to
8020  * commit the transaction group.
8021  *
8022  * If there are no dirty vdevs, we sync the uberblock to a few random
8023  * top-level vdevs that are known to be visible in the config cache
8024  * (see spa_vdev_add() for a complete description). If there *are* dirty
8025  * vdevs, sync the uberblock to all vdevs.
8026  */
8027 static void
8028 spa_sync_rewrite_vdev_config(spa_t *spa, dmu_tx_t *tx)
8029 {
8030 	vdev_t *rvd = spa->spa_root_vdev;
8031 	uint64_t txg = tx->tx_txg;
8032 
8033 	for (;;) {
8034 		int error = 0;
8035 
8036 		/*
8037 		 * We hold SCL_STATE to prevent vdev open/close/etc.
8038 		 * while we're attempting to write the vdev labels.
8039 		 */
8040 		spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
8041 
8042 		if (list_is_empty(&spa->spa_config_dirty_list)) {
8043 			vdev_t *svd[SPA_SYNC_MIN_VDEVS] = { NULL };
8044 			int svdcount = 0;
8045 			int children = rvd->vdev_children;
8046 			int c0 = spa_get_random(children);
8047 
8048 			for (int c = 0; c < children; c++) {
8049 				vdev_t *vd =
8050 				    rvd->vdev_child[(c0 + c) % children];
8051 
8052 				/* Stop when revisiting the first vdev */
8053 				if (c > 0 && svd[0] == vd)
8054 					break;
8055 
8056 				if (vd->vdev_ms_array == 0 ||
8057 				    vd->vdev_islog ||
8058 				    !vdev_is_concrete(vd))
8059 					continue;
8060 
8061 				svd[svdcount++] = vd;
8062 				if (svdcount == SPA_SYNC_MIN_VDEVS)
8063 					break;
8064 			}
8065 			error = vdev_config_sync(svd, svdcount, txg);
8066 		} else {
8067 			error = vdev_config_sync(rvd->vdev_child,
8068 			    rvd->vdev_children, txg);
8069 		}
8070 
8071 		if (error == 0)
8072 			spa->spa_last_synced_guid = rvd->vdev_guid;
8073 
8074 		spa_config_exit(spa, SCL_STATE, FTAG);
8075 
8076 		if (error == 0)
8077 			break;
8078 		zio_suspend(spa, NULL, ZIO_SUSPEND_IOERR);
8079 		zio_resume_wait(spa);
8080 	}
8081 }
8082 
8083 /*
8084  * Sync the specified transaction group.  New blocks may be dirtied as
8085  * part of the process, so we iterate until it converges.
8086  */
8087 void
8088 spa_sync(spa_t *spa, uint64_t txg)
8089 {
8090 	vdev_t *vd = NULL;
8091 
8092 	VERIFY(spa_writeable(spa));
8093 
8094 	/*
8095 	 * Wait for i/os issued in open context that need to complete
8096 	 * before this txg syncs.
8097 	 */
8098 	(void) zio_wait(spa->spa_txg_zio[txg & TXG_MASK]);
8099 	spa->spa_txg_zio[txg & TXG_MASK] = zio_root(spa, NULL, NULL,
8100 	    ZIO_FLAG_CANFAIL);
8101 
8102 	/*
8103 	 * Lock out configuration changes.
8104 	 */
8105 	spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
8106 
8107 	spa->spa_syncing_txg = txg;
8108 	spa->spa_sync_pass = 0;
8109 
8110 	for (int i = 0; i < spa->spa_alloc_count; i++) {
8111 		mutex_enter(&spa->spa_alloc_locks[i]);
8112 		VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i]));
8113 		mutex_exit(&spa->spa_alloc_locks[i]);
8114 	}
8115 
8116 	/*
8117 	 * If there are any pending vdev state changes, convert them
8118 	 * into config changes that go out with this transaction group.
8119 	 */
8120 	spa_config_enter(spa, SCL_STATE, FTAG, RW_READER);
8121 	while (list_head(&spa->spa_state_dirty_list) != NULL) {
8122 		/*
8123 		 * We need the write lock here because, for aux vdevs,
8124 		 * calling vdev_config_dirty() modifies sav_config.
8125 		 * This is ugly and will become unnecessary when we
8126 		 * eliminate the aux vdev wart by integrating all vdevs
8127 		 * into the root vdev tree.
8128 		 */
8129 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
8130 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_WRITER);
8131 		while ((vd = list_head(&spa->spa_state_dirty_list)) != NULL) {
8132 			vdev_state_clean(vd);
8133 			vdev_config_dirty(vd);
8134 		}
8135 		spa_config_exit(spa, SCL_CONFIG | SCL_STATE, FTAG);
8136 		spa_config_enter(spa, SCL_CONFIG | SCL_STATE, FTAG, RW_READER);
8137 	}
8138 	spa_config_exit(spa, SCL_STATE, FTAG);
8139 
8140 	dsl_pool_t *dp = spa->spa_dsl_pool;
8141 	dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
8142 
8143 	spa->spa_sync_starttime = gethrtime();
8144 	VERIFY(cyclic_reprogram(spa->spa_deadman_cycid,
8145 	    spa->spa_sync_starttime + spa->spa_deadman_synctime));
8146 
8147 	/*
8148 	 * If we are upgrading to SPA_VERSION_RAIDZ_DEFLATE this txg,
8149 	 * set spa_deflate if we have no raid-z vdevs.
8150 	 */
8151 	if (spa->spa_ubsync.ub_version < SPA_VERSION_RAIDZ_DEFLATE &&
8152 	    spa->spa_uberblock.ub_version >= SPA_VERSION_RAIDZ_DEFLATE) {
8153 		vdev_t *rvd = spa->spa_root_vdev;
8154 
8155 		int i;
8156 		for (i = 0; i < rvd->vdev_children; i++) {
8157 			vd = rvd->vdev_child[i];
8158 			if (vd->vdev_deflate_ratio != SPA_MINBLOCKSIZE)
8159 				break;
8160 		}
8161 		if (i == rvd->vdev_children) {
8162 			spa->spa_deflate = TRUE;
8163 			VERIFY0(zap_add(spa->spa_meta_objset,
8164 			    DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_DEFLATE,
8165 			    sizeof (uint64_t), 1, &spa->spa_deflate, tx));
8166 		}
8167 	}
8168 
8169 	spa_sync_adjust_vdev_max_queue_depth(spa);
8170 
8171 	spa_sync_condense_indirect(spa, tx);
8172 
8173 	spa_sync_iterate_to_convergence(spa, tx);
8174 
8175 #ifdef ZFS_DEBUG
8176 	if (!list_is_empty(&spa->spa_config_dirty_list)) {
8177 		/*
8178 		 * Make sure that the number of ZAPs for all the vdevs matches
8179 		 * the number of ZAPs in the per-vdev ZAP list. This only gets
8180 		 * called if the config is dirty; otherwise there may be
8181 		 * outstanding AVZ operations that weren't completed in
8182 		 * spa_sync_config_object.
8183 		 */
8184 		uint64_t all_vdev_zap_entry_count;
8185 		ASSERT0(zap_count(spa->spa_meta_objset,
8186 		    spa->spa_all_vdev_zaps, &all_vdev_zap_entry_count));
8187 		ASSERT3U(vdev_count_verify_zaps(spa->spa_root_vdev), ==,
8188 		    all_vdev_zap_entry_count);
8189 	}
8190 #endif
8191 
8192 	if (spa->spa_vdev_removal != NULL) {
8193 		ASSERT0(spa->spa_vdev_removal->svr_bytes_done[txg & TXG_MASK]);
8194 	}
8195 
8196 	spa_sync_rewrite_vdev_config(spa, tx);
8197 	dmu_tx_commit(tx);
8198 
8199 	VERIFY(cyclic_reprogram(spa->spa_deadman_cycid, CY_INFINITY));
8200 
8201 	/*
8202 	 * Clear the dirty config list.
8203 	 */
8204 	while ((vd = list_head(&spa->spa_config_dirty_list)) != NULL)
8205 		vdev_config_clean(vd);
8206 
8207 	/*
8208 	 * Now that the new config has synced transactionally,
8209 	 * let it become visible to the config cache.
8210 	 */
8211 	if (spa->spa_config_syncing != NULL) {
8212 		spa_config_set(spa, spa->spa_config_syncing);
8213 		spa->spa_config_txg = txg;
8214 		spa->spa_config_syncing = NULL;
8215 	}
8216 
8217 	dsl_pool_sync_done(dp, txg);
8218 
8219 	for (int i = 0; i < spa->spa_alloc_count; i++) {
8220 		mutex_enter(&spa->spa_alloc_locks[i]);
8221 		VERIFY0(avl_numnodes(&spa->spa_alloc_trees[i]));
8222 		mutex_exit(&spa->spa_alloc_locks[i]);
8223 	}
8224 
8225 	/*
8226 	 * Update usable space statistics.
8227 	 */
8228 	while ((vd = txg_list_remove(&spa->spa_vdev_txg_list, TXG_CLEAN(txg)))
8229 	    != NULL)
8230 		vdev_sync_done(vd, txg);
8231 
8232 	spa_update_dspace(spa);
8233 
8234 	/*
8235 	 * It had better be the case that we didn't dirty anything
8236 	 * since vdev_config_sync().
8237 	 */
8238 	ASSERT(txg_list_empty(&dp->dp_dirty_datasets, txg));
8239 	ASSERT(txg_list_empty(&dp->dp_dirty_dirs, txg));
8240 	ASSERT(txg_list_empty(&spa->spa_vdev_txg_list, txg));
8241 
8242 	while (zfs_pause_spa_sync)
8243 		delay(1);
8244 
8245 	spa->spa_sync_pass = 0;
8246 
8247 	/*
8248 	 * Update the last synced uberblock here. We want to do this at
8249 	 * the end of spa_sync() so that consumers of spa_last_synced_txg()
8250 	 * will be guaranteed that all the processing associated with
8251 	 * that txg has been completed.
8252 	 */
8253 	spa->spa_ubsync = spa->spa_uberblock;
8254 	spa_config_exit(spa, SCL_CONFIG, FTAG);
8255 
8256 	spa_handle_ignored_writes(spa);
8257 
8258 	/*
8259 	 * If any async tasks have been requested, kick them off.
8260 	 */
8261 	spa_async_dispatch(spa);
8262 }
8263 
8264 /*
8265  * Sync all pools.  We don't want to hold the namespace lock across these
8266  * operations, so we take a reference on the spa_t and drop the lock during the
8267  * sync.
8268  */
8269 void
8270 spa_sync_allpools(void)
8271 {
8272 	spa_t *spa = NULL;
8273 	mutex_enter(&spa_namespace_lock);
8274 	while ((spa = spa_next(spa)) != NULL) {
8275 		if (spa_state(spa) != POOL_STATE_ACTIVE ||
8276 		    !spa_writeable(spa) || spa_suspended(spa))
8277 			continue;
8278 		spa_open_ref(spa, FTAG);
8279 		mutex_exit(&spa_namespace_lock);
8280 		txg_wait_synced(spa_get_dsl(spa), 0);
8281 		mutex_enter(&spa_namespace_lock);
8282 		spa_close(spa, FTAG);
8283 	}
8284 	mutex_exit(&spa_namespace_lock);
8285 }
8286 
8287 /*
8288  * ==========================================================================
8289  * Miscellaneous routines
8290  * ==========================================================================
8291  */
8292 
8293 /*
8294  * Remove all pools in the system.
8295  */
8296 void
8297 spa_evict_all(void)
8298 {
8299 	spa_t *spa;
8300 
8301 	/*
8302 	 * Remove all cached state.  All pools should be closed now,
8303 	 * so every spa in the AVL tree should be unreferenced.
8304 	 */
8305 	mutex_enter(&spa_namespace_lock);
8306 	while ((spa = spa_next(NULL)) != NULL) {
8307 		/*
8308 		 * Stop async tasks.  The async thread may need to detach
8309 		 * a device that's been replaced, which requires grabbing
8310 		 * spa_namespace_lock, so we must drop it here.
8311 		 */
8312 		spa_open_ref(spa, FTAG);
8313 		mutex_exit(&spa_namespace_lock);
8314 		spa_async_suspend(spa);
8315 		mutex_enter(&spa_namespace_lock);
8316 		spa_close(spa, FTAG);
8317 
8318 		if (spa->spa_state != POOL_STATE_UNINITIALIZED) {
8319 			spa_unload(spa);
8320 			spa_deactivate(spa);
8321 		}
8322 		spa_remove(spa);
8323 	}
8324 	mutex_exit(&spa_namespace_lock);
8325 }
8326 
8327 vdev_t *
8328 spa_lookup_by_guid(spa_t *spa, uint64_t guid, boolean_t aux)
8329 {
8330 	vdev_t *vd;
8331 	int i;
8332 
8333 	if ((vd = vdev_lookup_by_guid(spa->spa_root_vdev, guid)) != NULL)
8334 		return (vd);
8335 
8336 	if (aux) {
8337 		for (i = 0; i < spa->spa_l2cache.sav_count; i++) {
8338 			vd = spa->spa_l2cache.sav_vdevs[i];
8339 			if (vd->vdev_guid == guid)
8340 				return (vd);
8341 		}
8342 
8343 		for (i = 0; i < spa->spa_spares.sav_count; i++) {
8344 			vd = spa->spa_spares.sav_vdevs[i];
8345 			if (vd->vdev_guid == guid)
8346 				return (vd);
8347 		}
8348 	}
8349 
8350 	return (NULL);
8351 }
8352 
8353 void
8354 spa_upgrade(spa_t *spa, uint64_t version)
8355 {
8356 	ASSERT(spa_writeable(spa));
8357 
8358 	spa_config_enter(spa, SCL_ALL, FTAG, RW_WRITER);
8359 
8360 	/*
8361 	 * This should only be called for a non-faulted pool, and since a
8362 	 * future version would result in an unopenable pool, this shouldn't be
8363 	 * possible.
8364 	 */
8365 	ASSERT(SPA_VERSION_IS_SUPPORTED(spa->spa_uberblock.ub_version));
8366 	ASSERT3U(version, >=, spa->spa_uberblock.ub_version);
8367 
8368 	spa->spa_uberblock.ub_version = version;
8369 	vdev_config_dirty(spa->spa_root_vdev);
8370 
8371 	spa_config_exit(spa, SCL_ALL, FTAG);
8372 
8373 	txg_wait_synced(spa_get_dsl(spa), 0);
8374 }
8375 
8376 boolean_t
8377 spa_has_spare(spa_t *spa, uint64_t guid)
8378 {
8379 	int i;
8380 	uint64_t spareguid;
8381 	spa_aux_vdev_t *sav = &spa->spa_spares;
8382 
8383 	for (i = 0; i < sav->sav_count; i++)
8384 		if (sav->sav_vdevs[i]->vdev_guid == guid)
8385 			return (B_TRUE);
8386 
8387 	for (i = 0; i < sav->sav_npending; i++) {
8388 		if (nvlist_lookup_uint64(sav->sav_pending[i], ZPOOL_CONFIG_GUID,
8389 		    &spareguid) == 0 && spareguid == guid)
8390 			return (B_TRUE);
8391 	}
8392 
8393 	return (B_FALSE);
8394 }
8395 
8396 /*
8397  * Check if a pool has an active shared spare device.
8398  * Note: reference count of an active spare is 2, as a spare and as a replace
8399  */
8400 static boolean_t
8401 spa_has_active_shared_spare(spa_t *spa)
8402 {
8403 	int i, refcnt;
8404 	uint64_t pool;
8405 	spa_aux_vdev_t *sav = &spa->spa_spares;
8406 
8407 	for (i = 0; i < sav->sav_count; i++) {
8408 		if (spa_spare_exists(sav->sav_vdevs[i]->vdev_guid, &pool,
8409 		    &refcnt) && pool != 0ULL && pool == spa_guid(spa) &&
8410 		    refcnt > 2)
8411 			return (B_TRUE);
8412 	}
8413 
8414 	return (B_FALSE);
8415 }
8416 
8417 sysevent_t *
8418 spa_event_create(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
8419 {
8420 	sysevent_t		*ev = NULL;
8421 #ifdef _KERNEL
8422 	sysevent_attr_list_t	*attr = NULL;
8423 	sysevent_value_t	value;
8424 
8425 	ev = sysevent_alloc(EC_ZFS, (char *)name, SUNW_KERN_PUB "zfs",
8426 	    SE_SLEEP);
8427 	ASSERT(ev != NULL);
8428 
8429 	value.value_type = SE_DATA_TYPE_STRING;
8430 	value.value.sv_string = spa_name(spa);
8431 	if (sysevent_add_attr(&attr, ZFS_EV_POOL_NAME, &value, SE_SLEEP) != 0)
8432 		goto done;
8433 
8434 	value.value_type = SE_DATA_TYPE_UINT64;
8435 	value.value.sv_uint64 = spa_guid(spa);
8436 	if (sysevent_add_attr(&attr, ZFS_EV_POOL_GUID, &value, SE_SLEEP) != 0)
8437 		goto done;
8438 
8439 	if (vd) {
8440 		value.value_type = SE_DATA_TYPE_UINT64;
8441 		value.value.sv_uint64 = vd->vdev_guid;
8442 		if (sysevent_add_attr(&attr, ZFS_EV_VDEV_GUID, &value,
8443 		    SE_SLEEP) != 0)
8444 			goto done;
8445 
8446 		if (vd->vdev_path) {
8447 			value.value_type = SE_DATA_TYPE_STRING;
8448 			value.value.sv_string = vd->vdev_path;
8449 			if (sysevent_add_attr(&attr, ZFS_EV_VDEV_PATH,
8450 			    &value, SE_SLEEP) != 0)
8451 				goto done;
8452 		}
8453 	}
8454 
8455 	if (hist_nvl != NULL) {
8456 		fnvlist_merge((nvlist_t *)attr, hist_nvl);
8457 	}
8458 
8459 	if (sysevent_attach_attributes(ev, attr) != 0)
8460 		goto done;
8461 	attr = NULL;
8462 
8463 done:
8464 	if (attr)
8465 		sysevent_free_attr(attr);
8466 
8467 #endif
8468 	return (ev);
8469 }
8470 
8471 void
8472 spa_event_post(sysevent_t *ev)
8473 {
8474 #ifdef _KERNEL
8475 	sysevent_id_t		eid;
8476 
8477 	(void) log_sysevent(ev, SE_SLEEP, &eid);
8478 	sysevent_free(ev);
8479 #endif
8480 }
8481 
8482 void
8483 spa_event_discard(sysevent_t *ev)
8484 {
8485 #ifdef _KERNEL
8486 	sysevent_free(ev);
8487 #endif
8488 }
8489 
8490 /*
8491  * Post a sysevent corresponding to the given event.  The 'name' must be one of
8492  * the event definitions in sys/sysevent/eventdefs.h.  The payload will be
8493  * filled in from the spa and (optionally) the vdev and history nvl.  This
8494  * doesn't do anything in the userland libzpool, as we don't want consumers to
8495  * misinterpret ztest or zdb as real changes.
8496  */
8497 void
8498 spa_event_notify(spa_t *spa, vdev_t *vd, nvlist_t *hist_nvl, const char *name)
8499 {
8500 	spa_event_post(spa_event_create(spa, vd, hist_nvl, name));
8501 }
8502