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, 2015 by Delphix. All rights reserved.
25 * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
26 * Copyright (c) 2014 Integros [integros.com]
27 */
28
29#include <sys/zfs_context.h>
30#include <sys/fm/fs/zfs.h>
31#include <sys/spa.h>
32#include <sys/spa_impl.h>
33#include <sys/dmu.h>
34#include <sys/dmu_tx.h>
35#include <sys/vdev_impl.h>
36#include <sys/uberblock_impl.h>
37#include <sys/metaslab.h>
38#include <sys/metaslab_impl.h>
39#include <sys/space_map.h>
40#include <sys/space_reftree.h>
41#include <sys/zio.h>
42#include <sys/zap.h>
43#include <sys/fs/zfs.h>
44#include <sys/arc.h>
45#include <sys/zil.h>
46#include <sys/dsl_scan.h>
47
48/*
49 * Virtual device management.
50 */
51
52static vdev_ops_t *vdev_ops_table[] = {
53	&vdev_root_ops,
54	&vdev_raidz_ops,
55	&vdev_mirror_ops,
56	&vdev_replacing_ops,
57	&vdev_spare_ops,
58	&vdev_disk_ops,
59	&vdev_file_ops,
60	&vdev_missing_ops,
61	&vdev_hole_ops,
62	NULL
63};
64
65/* maximum scrub/resilver I/O queue per leaf vdev */
66int zfs_scrub_limit = 10;
67
68/*
69 * When a vdev is added, it will be divided into approximately (but no
70 * more than) this number of metaslabs.
71 */
72int metaslabs_per_vdev = 200;
73
74/*
75 * Given a vdev type, return the appropriate ops vector.
76 */
77static vdev_ops_t *
78vdev_getops(const char *type)
79{
80	vdev_ops_t *ops, **opspp;
81
82	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
83		if (strcmp(ops->vdev_op_type, type) == 0)
84			break;
85
86	return (ops);
87}
88
89/*
90 * Default asize function: return the MAX of psize with the asize of
91 * all children.  This is what's used by anything other than RAID-Z.
92 */
93uint64_t
94vdev_default_asize(vdev_t *vd, uint64_t psize)
95{
96	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
97	uint64_t csize;
98
99	for (int c = 0; c < vd->vdev_children; c++) {
100		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
101		asize = MAX(asize, csize);
102	}
103
104	return (asize);
105}
106
107/*
108 * Get the minimum allocatable size. We define the allocatable size as
109 * the vdev's asize rounded to the nearest metaslab. This allows us to
110 * replace or attach devices which don't have the same physical size but
111 * can still satisfy the same number of allocations.
112 */
113uint64_t
114vdev_get_min_asize(vdev_t *vd)
115{
116	vdev_t *pvd = vd->vdev_parent;
117
118	/*
119	 * If our parent is NULL (inactive spare or cache) or is the root,
120	 * just return our own asize.
121	 */
122	if (pvd == NULL)
123		return (vd->vdev_asize);
124
125	/*
126	 * The top-level vdev just returns the allocatable size rounded
127	 * to the nearest metaslab.
128	 */
129	if (vd == vd->vdev_top)
130		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
131
132	/*
133	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
134	 * so each child must provide at least 1/Nth of its asize.
135	 */
136	if (pvd->vdev_ops == &vdev_raidz_ops)
137		return (pvd->vdev_min_asize / pvd->vdev_children);
138
139	return (pvd->vdev_min_asize);
140}
141
142void
143vdev_set_min_asize(vdev_t *vd)
144{
145	vd->vdev_min_asize = vdev_get_min_asize(vd);
146
147	for (int c = 0; c < vd->vdev_children; c++)
148		vdev_set_min_asize(vd->vdev_child[c]);
149}
150
151vdev_t *
152vdev_lookup_top(spa_t *spa, uint64_t vdev)
153{
154	vdev_t *rvd = spa->spa_root_vdev;
155
156	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
157
158	if (vdev < rvd->vdev_children) {
159		ASSERT(rvd->vdev_child[vdev] != NULL);
160		return (rvd->vdev_child[vdev]);
161	}
162
163	return (NULL);
164}
165
166vdev_t *
167vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
168{
169	vdev_t *mvd;
170
171	if (vd->vdev_guid == guid)
172		return (vd);
173
174	for (int c = 0; c < vd->vdev_children; c++)
175		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
176		    NULL)
177			return (mvd);
178
179	return (NULL);
180}
181
182static int
183vdev_count_leaves_impl(vdev_t *vd)
184{
185	int n = 0;
186
187	if (vd->vdev_ops->vdev_op_leaf)
188		return (1);
189
190	for (int c = 0; c < vd->vdev_children; c++)
191		n += vdev_count_leaves_impl(vd->vdev_child[c]);
192
193	return (n);
194}
195
196int
197vdev_count_leaves(spa_t *spa)
198{
199	return (vdev_count_leaves_impl(spa->spa_root_vdev));
200}
201
202void
203vdev_add_child(vdev_t *pvd, vdev_t *cvd)
204{
205	size_t oldsize, newsize;
206	uint64_t id = cvd->vdev_id;
207	vdev_t **newchild;
208	spa_t *spa = cvd->vdev_spa;
209
210	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
211	ASSERT(cvd->vdev_parent == NULL);
212
213	cvd->vdev_parent = pvd;
214
215	if (pvd == NULL)
216		return;
217
218	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
219
220	oldsize = pvd->vdev_children * sizeof (vdev_t *);
221	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
222	newsize = pvd->vdev_children * sizeof (vdev_t *);
223
224	newchild = kmem_zalloc(newsize, KM_SLEEP);
225	if (pvd->vdev_child != NULL) {
226		bcopy(pvd->vdev_child, newchild, oldsize);
227		kmem_free(pvd->vdev_child, oldsize);
228	}
229
230	pvd->vdev_child = newchild;
231	pvd->vdev_child[id] = cvd;
232
233	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
234	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
235
236	/*
237	 * Walk up all ancestors to update guid sum.
238	 */
239	for (; pvd != NULL; pvd = pvd->vdev_parent)
240		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
241}
242
243void
244vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
245{
246	int c;
247	uint_t id = cvd->vdev_id;
248
249	ASSERT(cvd->vdev_parent == pvd);
250
251	if (pvd == NULL)
252		return;
253
254	ASSERT(id < pvd->vdev_children);
255	ASSERT(pvd->vdev_child[id] == cvd);
256
257	pvd->vdev_child[id] = NULL;
258	cvd->vdev_parent = NULL;
259
260	for (c = 0; c < pvd->vdev_children; c++)
261		if (pvd->vdev_child[c])
262			break;
263
264	if (c == pvd->vdev_children) {
265		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
266		pvd->vdev_child = NULL;
267		pvd->vdev_children = 0;
268	}
269
270	/*
271	 * Walk up all ancestors to update guid sum.
272	 */
273	for (; pvd != NULL; pvd = pvd->vdev_parent)
274		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
275}
276
277/*
278 * Remove any holes in the child array.
279 */
280void
281vdev_compact_children(vdev_t *pvd)
282{
283	vdev_t **newchild, *cvd;
284	int oldc = pvd->vdev_children;
285	int newc;
286
287	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
288
289	for (int c = newc = 0; c < oldc; c++)
290		if (pvd->vdev_child[c])
291			newc++;
292
293	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
294
295	for (int c = newc = 0; c < oldc; c++) {
296		if ((cvd = pvd->vdev_child[c]) != NULL) {
297			newchild[newc] = cvd;
298			cvd->vdev_id = newc++;
299		}
300	}
301
302	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
303	pvd->vdev_child = newchild;
304	pvd->vdev_children = newc;
305}
306
307/*
308 * Allocate and minimally initialize a vdev_t.
309 */
310vdev_t *
311vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
312{
313	vdev_t *vd;
314
315	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
316
317	if (spa->spa_root_vdev == NULL) {
318		ASSERT(ops == &vdev_root_ops);
319		spa->spa_root_vdev = vd;
320		spa->spa_load_guid = spa_generate_guid(NULL);
321	}
322
323	if (guid == 0 && ops != &vdev_hole_ops) {
324		if (spa->spa_root_vdev == vd) {
325			/*
326			 * The root vdev's guid will also be the pool guid,
327			 * which must be unique among all pools.
328			 */
329			guid = spa_generate_guid(NULL);
330		} else {
331			/*
332			 * Any other vdev's guid must be unique within the pool.
333			 */
334			guid = spa_generate_guid(spa);
335		}
336		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
337	}
338
339	vd->vdev_spa = spa;
340	vd->vdev_id = id;
341	vd->vdev_guid = guid;
342	vd->vdev_guid_sum = guid;
343	vd->vdev_ops = ops;
344	vd->vdev_state = VDEV_STATE_CLOSED;
345	vd->vdev_ishole = (ops == &vdev_hole_ops);
346
347	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
348	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
349	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
350	for (int t = 0; t < DTL_TYPES; t++) {
351		vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
352		    &vd->vdev_dtl_lock);
353	}
354	txg_list_create(&vd->vdev_ms_list,
355	    offsetof(struct metaslab, ms_txg_node));
356	txg_list_create(&vd->vdev_dtl_list,
357	    offsetof(struct vdev, vdev_dtl_node));
358	vd->vdev_stat.vs_timestamp = gethrtime();
359	vdev_queue_init(vd);
360	vdev_cache_init(vd);
361
362	return (vd);
363}
364
365/*
366 * Allocate a new vdev.  The 'alloctype' is used to control whether we are
367 * creating a new vdev or loading an existing one - the behavior is slightly
368 * different for each case.
369 */
370int
371vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
372    int alloctype)
373{
374	vdev_ops_t *ops;
375	char *type;
376	uint64_t guid = 0, islog, nparity;
377	vdev_t *vd;
378
379	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
380
381	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
382		return (SET_ERROR(EINVAL));
383
384	if ((ops = vdev_getops(type)) == NULL)
385		return (SET_ERROR(EINVAL));
386
387	/*
388	 * If this is a load, get the vdev guid from the nvlist.
389	 * Otherwise, vdev_alloc_common() will generate one for us.
390	 */
391	if (alloctype == VDEV_ALLOC_LOAD) {
392		uint64_t label_id;
393
394		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
395		    label_id != id)
396			return (SET_ERROR(EINVAL));
397
398		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
399			return (SET_ERROR(EINVAL));
400	} else if (alloctype == VDEV_ALLOC_SPARE) {
401		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
402			return (SET_ERROR(EINVAL));
403	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
404		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
405			return (SET_ERROR(EINVAL));
406	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
407		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
408			return (SET_ERROR(EINVAL));
409	}
410
411	/*
412	 * The first allocated vdev must be of type 'root'.
413	 */
414	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
415		return (SET_ERROR(EINVAL));
416
417	/*
418	 * Determine whether we're a log vdev.
419	 */
420	islog = 0;
421	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
422	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
423		return (SET_ERROR(ENOTSUP));
424
425	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
426		return (SET_ERROR(ENOTSUP));
427
428	/*
429	 * Set the nparity property for RAID-Z vdevs.
430	 */
431	nparity = -1ULL;
432	if (ops == &vdev_raidz_ops) {
433		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
434		    &nparity) == 0) {
435			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
436				return (SET_ERROR(EINVAL));
437			/*
438			 * Previous versions could only support 1 or 2 parity
439			 * device.
440			 */
441			if (nparity > 1 &&
442			    spa_version(spa) < SPA_VERSION_RAIDZ2)
443				return (SET_ERROR(ENOTSUP));
444			if (nparity > 2 &&
445			    spa_version(spa) < SPA_VERSION_RAIDZ3)
446				return (SET_ERROR(ENOTSUP));
447		} else {
448			/*
449			 * We require the parity to be specified for SPAs that
450			 * support multiple parity levels.
451			 */
452			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
453				return (SET_ERROR(EINVAL));
454			/*
455			 * Otherwise, we default to 1 parity device for RAID-Z.
456			 */
457			nparity = 1;
458		}
459	} else {
460		nparity = 0;
461	}
462	ASSERT(nparity != -1ULL);
463
464	vd = vdev_alloc_common(spa, id, guid, ops);
465
466	vd->vdev_islog = islog;
467	vd->vdev_nparity = nparity;
468
469	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
470		vd->vdev_path = spa_strdup(vd->vdev_path);
471	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
472		vd->vdev_devid = spa_strdup(vd->vdev_devid);
473	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
474	    &vd->vdev_physpath) == 0)
475		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
476	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
477		vd->vdev_fru = spa_strdup(vd->vdev_fru);
478
479	/*
480	 * Set the whole_disk property.  If it's not specified, leave the value
481	 * as -1.
482	 */
483	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
484	    &vd->vdev_wholedisk) != 0)
485		vd->vdev_wholedisk = -1ULL;
486
487	/*
488	 * Look for the 'not present' flag.  This will only be set if the device
489	 * was not present at the time of import.
490	 */
491	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
492	    &vd->vdev_not_present);
493
494	/*
495	 * Get the alignment requirement.
496	 */
497	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
498
499	/*
500	 * Retrieve the vdev creation time.
501	 */
502	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
503	    &vd->vdev_crtxg);
504
505	/*
506	 * If we're a top-level vdev, try to load the allocation parameters.
507	 */
508	if (parent && !parent->vdev_parent &&
509	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
510		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
511		    &vd->vdev_ms_array);
512		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
513		    &vd->vdev_ms_shift);
514		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
515		    &vd->vdev_asize);
516		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
517		    &vd->vdev_removing);
518		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
519		    &vd->vdev_top_zap);
520	} else {
521		ASSERT0(vd->vdev_top_zap);
522	}
523
524	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
525		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
526		    alloctype == VDEV_ALLOC_ADD ||
527		    alloctype == VDEV_ALLOC_SPLIT ||
528		    alloctype == VDEV_ALLOC_ROOTPOOL);
529		vd->vdev_mg = metaslab_group_create(islog ?
530		    spa_log_class(spa) : spa_normal_class(spa), vd);
531	}
532
533	if (vd->vdev_ops->vdev_op_leaf &&
534	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
535		(void) nvlist_lookup_uint64(nv,
536		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
537	} else {
538		ASSERT0(vd->vdev_leaf_zap);
539	}
540
541	/*
542	 * If we're a leaf vdev, try to load the DTL object and other state.
543	 */
544
545	if (vd->vdev_ops->vdev_op_leaf &&
546	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
547	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
548		if (alloctype == VDEV_ALLOC_LOAD) {
549			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
550			    &vd->vdev_dtl_object);
551			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
552			    &vd->vdev_unspare);
553		}
554
555		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
556			uint64_t spare = 0;
557
558			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
559			    &spare) == 0 && spare)
560				spa_spare_add(vd);
561		}
562
563		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
564		    &vd->vdev_offline);
565
566		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
567		    &vd->vdev_resilver_txg);
568
569		/*
570		 * When importing a pool, we want to ignore the persistent fault
571		 * state, as the diagnosis made on another system may not be
572		 * valid in the current context.  Local vdevs will
573		 * remain in the faulted state.
574		 */
575		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
576			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
577			    &vd->vdev_faulted);
578			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
579			    &vd->vdev_degraded);
580			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
581			    &vd->vdev_removed);
582
583			if (vd->vdev_faulted || vd->vdev_degraded) {
584				char *aux;
585
586				vd->vdev_label_aux =
587				    VDEV_AUX_ERR_EXCEEDED;
588				if (nvlist_lookup_string(nv,
589				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
590				    strcmp(aux, "external") == 0)
591					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
592			}
593		}
594	}
595
596	/*
597	 * Add ourselves to the parent's list of children.
598	 */
599	vdev_add_child(parent, vd);
600
601	*vdp = vd;
602
603	return (0);
604}
605
606void
607vdev_free(vdev_t *vd)
608{
609	spa_t *spa = vd->vdev_spa;
610
611	/*
612	 * vdev_free() implies closing the vdev first.  This is simpler than
613	 * trying to ensure complicated semantics for all callers.
614	 */
615	vdev_close(vd);
616
617	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
618	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
619
620	/*
621	 * Free all children.
622	 */
623	for (int c = 0; c < vd->vdev_children; c++)
624		vdev_free(vd->vdev_child[c]);
625
626	ASSERT(vd->vdev_child == NULL);
627	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
628
629	/*
630	 * Discard allocation state.
631	 */
632	if (vd->vdev_mg != NULL) {
633		vdev_metaslab_fini(vd);
634		metaslab_group_destroy(vd->vdev_mg);
635	}
636
637	ASSERT0(vd->vdev_stat.vs_space);
638	ASSERT0(vd->vdev_stat.vs_dspace);
639	ASSERT0(vd->vdev_stat.vs_alloc);
640
641	/*
642	 * Remove this vdev from its parent's child list.
643	 */
644	vdev_remove_child(vd->vdev_parent, vd);
645
646	ASSERT(vd->vdev_parent == NULL);
647
648	/*
649	 * Clean up vdev structure.
650	 */
651	vdev_queue_fini(vd);
652	vdev_cache_fini(vd);
653
654	if (vd->vdev_path)
655		spa_strfree(vd->vdev_path);
656	if (vd->vdev_devid)
657		spa_strfree(vd->vdev_devid);
658	if (vd->vdev_physpath)
659		spa_strfree(vd->vdev_physpath);
660	if (vd->vdev_fru)
661		spa_strfree(vd->vdev_fru);
662
663	if (vd->vdev_isspare)
664		spa_spare_remove(vd);
665	if (vd->vdev_isl2cache)
666		spa_l2cache_remove(vd);
667
668	txg_list_destroy(&vd->vdev_ms_list);
669	txg_list_destroy(&vd->vdev_dtl_list);
670
671	mutex_enter(&vd->vdev_dtl_lock);
672	space_map_close(vd->vdev_dtl_sm);
673	for (int t = 0; t < DTL_TYPES; t++) {
674		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
675		range_tree_destroy(vd->vdev_dtl[t]);
676	}
677	mutex_exit(&vd->vdev_dtl_lock);
678
679	mutex_destroy(&vd->vdev_dtl_lock);
680	mutex_destroy(&vd->vdev_stat_lock);
681	mutex_destroy(&vd->vdev_probe_lock);
682
683	if (vd == spa->spa_root_vdev)
684		spa->spa_root_vdev = NULL;
685
686	kmem_free(vd, sizeof (vdev_t));
687}
688
689/*
690 * Transfer top-level vdev state from svd to tvd.
691 */
692static void
693vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
694{
695	spa_t *spa = svd->vdev_spa;
696	metaslab_t *msp;
697	vdev_t *vd;
698	int t;
699
700	ASSERT(tvd == tvd->vdev_top);
701
702	tvd->vdev_ms_array = svd->vdev_ms_array;
703	tvd->vdev_ms_shift = svd->vdev_ms_shift;
704	tvd->vdev_ms_count = svd->vdev_ms_count;
705	tvd->vdev_top_zap = svd->vdev_top_zap;
706
707	svd->vdev_ms_array = 0;
708	svd->vdev_ms_shift = 0;
709	svd->vdev_ms_count = 0;
710	svd->vdev_top_zap = 0;
711
712	if (tvd->vdev_mg)
713		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
714	tvd->vdev_mg = svd->vdev_mg;
715	tvd->vdev_ms = svd->vdev_ms;
716
717	svd->vdev_mg = NULL;
718	svd->vdev_ms = NULL;
719
720	if (tvd->vdev_mg != NULL)
721		tvd->vdev_mg->mg_vd = tvd;
722
723	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
724	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
725	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
726
727	svd->vdev_stat.vs_alloc = 0;
728	svd->vdev_stat.vs_space = 0;
729	svd->vdev_stat.vs_dspace = 0;
730
731	for (t = 0; t < TXG_SIZE; t++) {
732		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
733			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
734		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
735			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
736		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
737			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
738	}
739
740	if (list_link_active(&svd->vdev_config_dirty_node)) {
741		vdev_config_clean(svd);
742		vdev_config_dirty(tvd);
743	}
744
745	if (list_link_active(&svd->vdev_state_dirty_node)) {
746		vdev_state_clean(svd);
747		vdev_state_dirty(tvd);
748	}
749
750	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
751	svd->vdev_deflate_ratio = 0;
752
753	tvd->vdev_islog = svd->vdev_islog;
754	svd->vdev_islog = 0;
755}
756
757static void
758vdev_top_update(vdev_t *tvd, vdev_t *vd)
759{
760	if (vd == NULL)
761		return;
762
763	vd->vdev_top = tvd;
764
765	for (int c = 0; c < vd->vdev_children; c++)
766		vdev_top_update(tvd, vd->vdev_child[c]);
767}
768
769/*
770 * Add a mirror/replacing vdev above an existing vdev.
771 */
772vdev_t *
773vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
774{
775	spa_t *spa = cvd->vdev_spa;
776	vdev_t *pvd = cvd->vdev_parent;
777	vdev_t *mvd;
778
779	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
780
781	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
782
783	mvd->vdev_asize = cvd->vdev_asize;
784	mvd->vdev_min_asize = cvd->vdev_min_asize;
785	mvd->vdev_max_asize = cvd->vdev_max_asize;
786	mvd->vdev_ashift = cvd->vdev_ashift;
787	mvd->vdev_state = cvd->vdev_state;
788	mvd->vdev_crtxg = cvd->vdev_crtxg;
789
790	vdev_remove_child(pvd, cvd);
791	vdev_add_child(pvd, mvd);
792	cvd->vdev_id = mvd->vdev_children;
793	vdev_add_child(mvd, cvd);
794	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
795
796	if (mvd == mvd->vdev_top)
797		vdev_top_transfer(cvd, mvd);
798
799	return (mvd);
800}
801
802/*
803 * Remove a 1-way mirror/replacing vdev from the tree.
804 */
805void
806vdev_remove_parent(vdev_t *cvd)
807{
808	vdev_t *mvd = cvd->vdev_parent;
809	vdev_t *pvd = mvd->vdev_parent;
810
811	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
812
813	ASSERT(mvd->vdev_children == 1);
814	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
815	    mvd->vdev_ops == &vdev_replacing_ops ||
816	    mvd->vdev_ops == &vdev_spare_ops);
817	cvd->vdev_ashift = mvd->vdev_ashift;
818
819	vdev_remove_child(mvd, cvd);
820	vdev_remove_child(pvd, mvd);
821
822	/*
823	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
824	 * Otherwise, we could have detached an offline device, and when we
825	 * go to import the pool we'll think we have two top-level vdevs,
826	 * instead of a different version of the same top-level vdev.
827	 */
828	if (mvd->vdev_top == mvd) {
829		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
830		cvd->vdev_orig_guid = cvd->vdev_guid;
831		cvd->vdev_guid += guid_delta;
832		cvd->vdev_guid_sum += guid_delta;
833	}
834	cvd->vdev_id = mvd->vdev_id;
835	vdev_add_child(pvd, cvd);
836	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
837
838	if (cvd == cvd->vdev_top)
839		vdev_top_transfer(mvd, cvd);
840
841	ASSERT(mvd->vdev_children == 0);
842	vdev_free(mvd);
843}
844
845int
846vdev_metaslab_init(vdev_t *vd, uint64_t txg)
847{
848	spa_t *spa = vd->vdev_spa;
849	objset_t *mos = spa->spa_meta_objset;
850	uint64_t m;
851	uint64_t oldc = vd->vdev_ms_count;
852	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
853	metaslab_t **mspp;
854	int error;
855
856	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
857
858	/*
859	 * This vdev is not being allocated from yet or is a hole.
860	 */
861	if (vd->vdev_ms_shift == 0)
862		return (0);
863
864	ASSERT(!vd->vdev_ishole);
865
866	/*
867	 * Compute the raidz-deflation ratio.  Note, we hard-code
868	 * in 128k (1 << 17) because it is the "typical" blocksize.
869	 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
870	 * otherwise it would inconsistently account for existing bp's.
871	 */
872	vd->vdev_deflate_ratio = (1 << 17) /
873	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
874
875	ASSERT(oldc <= newc);
876
877	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
878
879	if (oldc != 0) {
880		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
881		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
882	}
883
884	vd->vdev_ms = mspp;
885	vd->vdev_ms_count = newc;
886
887	for (m = oldc; m < newc; m++) {
888		uint64_t object = 0;
889
890		if (txg == 0) {
891			error = dmu_read(mos, vd->vdev_ms_array,
892			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
893			    DMU_READ_PREFETCH);
894			if (error)
895				return (error);
896		}
897
898		error = metaslab_init(vd->vdev_mg, m, object, txg,
899		    &(vd->vdev_ms[m]));
900		if (error)
901			return (error);
902	}
903
904	if (txg == 0)
905		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
906
907	/*
908	 * If the vdev is being removed we don't activate
909	 * the metaslabs since we want to ensure that no new
910	 * allocations are performed on this device.
911	 */
912	if (oldc == 0 && !vd->vdev_removing)
913		metaslab_group_activate(vd->vdev_mg);
914
915	if (txg == 0)
916		spa_config_exit(spa, SCL_ALLOC, FTAG);
917
918	return (0);
919}
920
921void
922vdev_metaslab_fini(vdev_t *vd)
923{
924	uint64_t m;
925	uint64_t count = vd->vdev_ms_count;
926
927	if (vd->vdev_ms != NULL) {
928		metaslab_group_passivate(vd->vdev_mg);
929		for (m = 0; m < count; m++) {
930			metaslab_t *msp = vd->vdev_ms[m];
931
932			if (msp != NULL)
933				metaslab_fini(msp);
934		}
935		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
936		vd->vdev_ms = NULL;
937	}
938}
939
940typedef struct vdev_probe_stats {
941	boolean_t	vps_readable;
942	boolean_t	vps_writeable;
943	int		vps_flags;
944} vdev_probe_stats_t;
945
946static void
947vdev_probe_done(zio_t *zio)
948{
949	spa_t *spa = zio->io_spa;
950	vdev_t *vd = zio->io_vd;
951	vdev_probe_stats_t *vps = zio->io_private;
952
953	ASSERT(vd->vdev_probe_zio != NULL);
954
955	if (zio->io_type == ZIO_TYPE_READ) {
956		if (zio->io_error == 0)
957			vps->vps_readable = 1;
958		if (zio->io_error == 0 && spa_writeable(spa)) {
959			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
960			    zio->io_offset, zio->io_size, zio->io_data,
961			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
962			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
963		} else {
964			zio_buf_free(zio->io_data, zio->io_size);
965		}
966	} else if (zio->io_type == ZIO_TYPE_WRITE) {
967		if (zio->io_error == 0)
968			vps->vps_writeable = 1;
969		zio_buf_free(zio->io_data, zio->io_size);
970	} else if (zio->io_type == ZIO_TYPE_NULL) {
971		zio_t *pio;
972
973		vd->vdev_cant_read |= !vps->vps_readable;
974		vd->vdev_cant_write |= !vps->vps_writeable;
975
976		if (vdev_readable(vd) &&
977		    (vdev_writeable(vd) || !spa_writeable(spa))) {
978			zio->io_error = 0;
979		} else {
980			ASSERT(zio->io_error != 0);
981			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
982			    spa, vd, NULL, 0, 0);
983			zio->io_error = SET_ERROR(ENXIO);
984		}
985
986		mutex_enter(&vd->vdev_probe_lock);
987		ASSERT(vd->vdev_probe_zio == zio);
988		vd->vdev_probe_zio = NULL;
989		mutex_exit(&vd->vdev_probe_lock);
990
991		while ((pio = zio_walk_parents(zio)) != NULL)
992			if (!vdev_accessible(vd, pio))
993				pio->io_error = SET_ERROR(ENXIO);
994
995		kmem_free(vps, sizeof (*vps));
996	}
997}
998
999/*
1000 * Determine whether this device is accessible.
1001 *
1002 * Read and write to several known locations: the pad regions of each
1003 * vdev label but the first, which we leave alone in case it contains
1004 * a VTOC.
1005 */
1006zio_t *
1007vdev_probe(vdev_t *vd, zio_t *zio)
1008{
1009	spa_t *spa = vd->vdev_spa;
1010	vdev_probe_stats_t *vps = NULL;
1011	zio_t *pio;
1012
1013	ASSERT(vd->vdev_ops->vdev_op_leaf);
1014
1015	/*
1016	 * Don't probe the probe.
1017	 */
1018	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1019		return (NULL);
1020
1021	/*
1022	 * To prevent 'probe storms' when a device fails, we create
1023	 * just one probe i/o at a time.  All zios that want to probe
1024	 * this vdev will become parents of the probe io.
1025	 */
1026	mutex_enter(&vd->vdev_probe_lock);
1027
1028	if ((pio = vd->vdev_probe_zio) == NULL) {
1029		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1030
1031		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1032		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1033		    ZIO_FLAG_TRYHARD;
1034
1035		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1036			/*
1037			 * vdev_cant_read and vdev_cant_write can only
1038			 * transition from TRUE to FALSE when we have the
1039			 * SCL_ZIO lock as writer; otherwise they can only
1040			 * transition from FALSE to TRUE.  This ensures that
1041			 * any zio looking at these values can assume that
1042			 * failures persist for the life of the I/O.  That's
1043			 * important because when a device has intermittent
1044			 * connectivity problems, we want to ensure that
1045			 * they're ascribed to the device (ENXIO) and not
1046			 * the zio (EIO).
1047			 *
1048			 * Since we hold SCL_ZIO as writer here, clear both
1049			 * values so the probe can reevaluate from first
1050			 * principles.
1051			 */
1052			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1053			vd->vdev_cant_read = B_FALSE;
1054			vd->vdev_cant_write = B_FALSE;
1055		}
1056
1057		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1058		    vdev_probe_done, vps,
1059		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1060
1061		/*
1062		 * We can't change the vdev state in this context, so we
1063		 * kick off an async task to do it on our behalf.
1064		 */
1065		if (zio != NULL) {
1066			vd->vdev_probe_wanted = B_TRUE;
1067			spa_async_request(spa, SPA_ASYNC_PROBE);
1068		}
1069	}
1070
1071	if (zio != NULL)
1072		zio_add_child(zio, pio);
1073
1074	mutex_exit(&vd->vdev_probe_lock);
1075
1076	if (vps == NULL) {
1077		ASSERT(zio != NULL);
1078		return (NULL);
1079	}
1080
1081	for (int l = 1; l < VDEV_LABELS; l++) {
1082		zio_nowait(zio_read_phys(pio, vd,
1083		    vdev_label_offset(vd->vdev_psize, l,
1084		    offsetof(vdev_label_t, vl_pad2)),
1085		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1086		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1087		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1088	}
1089
1090	if (zio == NULL)
1091		return (pio);
1092
1093	zio_nowait(pio);
1094	return (NULL);
1095}
1096
1097static void
1098vdev_open_child(void *arg)
1099{
1100	vdev_t *vd = arg;
1101
1102	vd->vdev_open_thread = curthread;
1103	vd->vdev_open_error = vdev_open(vd);
1104	vd->vdev_open_thread = NULL;
1105}
1106
1107boolean_t
1108vdev_uses_zvols(vdev_t *vd)
1109{
1110	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1111	    strlen(ZVOL_DIR)) == 0)
1112		return (B_TRUE);
1113	for (int c = 0; c < vd->vdev_children; c++)
1114		if (vdev_uses_zvols(vd->vdev_child[c]))
1115			return (B_TRUE);
1116	return (B_FALSE);
1117}
1118
1119void
1120vdev_open_children(vdev_t *vd)
1121{
1122	taskq_t *tq;
1123	int children = vd->vdev_children;
1124
1125	/*
1126	 * in order to handle pools on top of zvols, do the opens
1127	 * in a single thread so that the same thread holds the
1128	 * spa_namespace_lock
1129	 */
1130	if (vdev_uses_zvols(vd)) {
1131		for (int c = 0; c < children; c++)
1132			vd->vdev_child[c]->vdev_open_error =
1133			    vdev_open(vd->vdev_child[c]);
1134		return;
1135	}
1136	tq = taskq_create("vdev_open", children, minclsyspri,
1137	    children, children, TASKQ_PREPOPULATE);
1138
1139	for (int c = 0; c < children; c++)
1140		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1141		    TQ_SLEEP) != NULL);
1142
1143	taskq_destroy(tq);
1144}
1145
1146/*
1147 * Prepare a virtual device for access.
1148 */
1149int
1150vdev_open(vdev_t *vd)
1151{
1152	spa_t *spa = vd->vdev_spa;
1153	int error;
1154	uint64_t osize = 0;
1155	uint64_t max_osize = 0;
1156	uint64_t asize, max_asize, psize;
1157	uint64_t ashift = 0;
1158
1159	ASSERT(vd->vdev_open_thread == curthread ||
1160	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1161	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1162	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1163	    vd->vdev_state == VDEV_STATE_OFFLINE);
1164
1165	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1166	vd->vdev_cant_read = B_FALSE;
1167	vd->vdev_cant_write = B_FALSE;
1168	vd->vdev_min_asize = vdev_get_min_asize(vd);
1169
1170	/*
1171	 * If this vdev is not removed, check its fault status.  If it's
1172	 * faulted, bail out of the open.
1173	 */
1174	if (!vd->vdev_removed && vd->vdev_faulted) {
1175		ASSERT(vd->vdev_children == 0);
1176		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1177		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1178		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1179		    vd->vdev_label_aux);
1180		return (SET_ERROR(ENXIO));
1181	} else if (vd->vdev_offline) {
1182		ASSERT(vd->vdev_children == 0);
1183		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1184		return (SET_ERROR(ENXIO));
1185	}
1186
1187	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1188
1189	/*
1190	 * Reset the vdev_reopening flag so that we actually close
1191	 * the vdev on error.
1192	 */
1193	vd->vdev_reopening = B_FALSE;
1194	if (zio_injection_enabled && error == 0)
1195		error = zio_handle_device_injection(vd, NULL, ENXIO);
1196
1197	if (error) {
1198		if (vd->vdev_removed &&
1199		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1200			vd->vdev_removed = B_FALSE;
1201
1202		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1203		    vd->vdev_stat.vs_aux);
1204		return (error);
1205	}
1206
1207	vd->vdev_removed = B_FALSE;
1208
1209	/*
1210	 * Recheck the faulted flag now that we have confirmed that
1211	 * the vdev is accessible.  If we're faulted, bail.
1212	 */
1213	if (vd->vdev_faulted) {
1214		ASSERT(vd->vdev_children == 0);
1215		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1216		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1217		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1218		    vd->vdev_label_aux);
1219		return (SET_ERROR(ENXIO));
1220	}
1221
1222	if (vd->vdev_degraded) {
1223		ASSERT(vd->vdev_children == 0);
1224		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1225		    VDEV_AUX_ERR_EXCEEDED);
1226	} else {
1227		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1228	}
1229
1230	/*
1231	 * For hole or missing vdevs we just return success.
1232	 */
1233	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1234		return (0);
1235
1236	for (int c = 0; c < vd->vdev_children; c++) {
1237		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1238			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1239			    VDEV_AUX_NONE);
1240			break;
1241		}
1242	}
1243
1244	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1245	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1246
1247	if (vd->vdev_children == 0) {
1248		if (osize < SPA_MINDEVSIZE) {
1249			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1250			    VDEV_AUX_TOO_SMALL);
1251			return (SET_ERROR(EOVERFLOW));
1252		}
1253		psize = osize;
1254		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1255		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1256		    VDEV_LABEL_END_SIZE);
1257	} else {
1258		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1259		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1260			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1261			    VDEV_AUX_TOO_SMALL);
1262			return (SET_ERROR(EOVERFLOW));
1263		}
1264		psize = 0;
1265		asize = osize;
1266		max_asize = max_osize;
1267	}
1268
1269	vd->vdev_psize = psize;
1270
1271	/*
1272	 * Make sure the allocatable size hasn't shrunk.
1273	 */
1274	if (asize < vd->vdev_min_asize) {
1275		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1276		    VDEV_AUX_BAD_LABEL);
1277		return (SET_ERROR(EINVAL));
1278	}
1279
1280	if (vd->vdev_asize == 0) {
1281		/*
1282		 * This is the first-ever open, so use the computed values.
1283		 * For testing purposes, a higher ashift can be requested.
1284		 */
1285		vd->vdev_asize = asize;
1286		vd->vdev_max_asize = max_asize;
1287		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1288	} else {
1289		/*
1290		 * Detect if the alignment requirement has increased.
1291		 * We don't want to make the pool unavailable, just
1292		 * issue a warning instead.
1293		 */
1294		if (ashift > vd->vdev_top->vdev_ashift &&
1295		    vd->vdev_ops->vdev_op_leaf) {
1296			cmn_err(CE_WARN,
1297			    "Disk, '%s', has a block alignment that is "
1298			    "larger than the pool's alignment\n",
1299			    vd->vdev_path);
1300		}
1301		vd->vdev_max_asize = max_asize;
1302	}
1303
1304	/*
1305	 * If all children are healthy and the asize has increased,
1306	 * then we've experienced dynamic LUN growth.  If automatic
1307	 * expansion is enabled then use the additional space.
1308	 */
1309	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1310	    (vd->vdev_expanding || spa->spa_autoexpand))
1311		vd->vdev_asize = asize;
1312
1313	vdev_set_min_asize(vd);
1314
1315	/*
1316	 * Ensure we can issue some IO before declaring the
1317	 * vdev open for business.
1318	 */
1319	if (vd->vdev_ops->vdev_op_leaf &&
1320	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1321		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1322		    VDEV_AUX_ERR_EXCEEDED);
1323		return (error);
1324	}
1325
1326	/*
1327	 * Track the min and max ashift values for normal data devices.
1328	 */
1329	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1330	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1331		if (vd->vdev_ashift > spa->spa_max_ashift)
1332			spa->spa_max_ashift = vd->vdev_ashift;
1333		if (vd->vdev_ashift < spa->spa_min_ashift)
1334			spa->spa_min_ashift = vd->vdev_ashift;
1335	}
1336
1337	/*
1338	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1339	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1340	 * since this would just restart the scrub we are already doing.
1341	 */
1342	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1343	    vdev_resilver_needed(vd, NULL, NULL))
1344		spa_async_request(spa, SPA_ASYNC_RESILVER);
1345
1346	return (0);
1347}
1348
1349/*
1350 * Called once the vdevs are all opened, this routine validates the label
1351 * contents.  This needs to be done before vdev_load() so that we don't
1352 * inadvertently do repair I/Os to the wrong device.
1353 *
1354 * If 'strict' is false ignore the spa guid check. This is necessary because
1355 * if the machine crashed during a re-guid the new guid might have been written
1356 * to all of the vdev labels, but not the cached config. The strict check
1357 * will be performed when the pool is opened again using the mos config.
1358 *
1359 * This function will only return failure if one of the vdevs indicates that it
1360 * has since been destroyed or exported.  This is only possible if
1361 * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1362 * will be updated but the function will return 0.
1363 */
1364int
1365vdev_validate(vdev_t *vd, boolean_t strict)
1366{
1367	spa_t *spa = vd->vdev_spa;
1368	nvlist_t *label;
1369	uint64_t guid = 0, top_guid;
1370	uint64_t state;
1371
1372	for (int c = 0; c < vd->vdev_children; c++)
1373		if (vdev_validate(vd->vdev_child[c], strict) != 0)
1374			return (SET_ERROR(EBADF));
1375
1376	/*
1377	 * If the device has already failed, or was marked offline, don't do
1378	 * any further validation.  Otherwise, label I/O will fail and we will
1379	 * overwrite the previous state.
1380	 */
1381	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1382		uint64_t aux_guid = 0;
1383		nvlist_t *nvl;
1384		uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1385		    spa_last_synced_txg(spa) : -1ULL;
1386
1387		if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1388			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1389			    VDEV_AUX_BAD_LABEL);
1390			return (0);
1391		}
1392
1393		/*
1394		 * Determine if this vdev has been split off into another
1395		 * pool.  If so, then refuse to open it.
1396		 */
1397		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1398		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1399			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1400			    VDEV_AUX_SPLIT_POOL);
1401			nvlist_free(label);
1402			return (0);
1403		}
1404
1405		if (strict && (nvlist_lookup_uint64(label,
1406		    ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1407		    guid != spa_guid(spa))) {
1408			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1409			    VDEV_AUX_CORRUPT_DATA);
1410			nvlist_free(label);
1411			return (0);
1412		}
1413
1414		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1415		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1416		    &aux_guid) != 0)
1417			aux_guid = 0;
1418
1419		/*
1420		 * If this vdev just became a top-level vdev because its
1421		 * sibling was detached, it will have adopted the parent's
1422		 * vdev guid -- but the label may or may not be on disk yet.
1423		 * Fortunately, either version of the label will have the
1424		 * same top guid, so if we're a top-level vdev, we can
1425		 * safely compare to that instead.
1426		 *
1427		 * If we split this vdev off instead, then we also check the
1428		 * original pool's guid.  We don't want to consider the vdev
1429		 * corrupt if it is partway through a split operation.
1430		 */
1431		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1432		    &guid) != 0 ||
1433		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1434		    &top_guid) != 0 ||
1435		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1436		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1437			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1438			    VDEV_AUX_CORRUPT_DATA);
1439			nvlist_free(label);
1440			return (0);
1441		}
1442
1443		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1444		    &state) != 0) {
1445			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1446			    VDEV_AUX_CORRUPT_DATA);
1447			nvlist_free(label);
1448			return (0);
1449		}
1450
1451		nvlist_free(label);
1452
1453		/*
1454		 * If this is a verbatim import, no need to check the
1455		 * state of the pool.
1456		 */
1457		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1458		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1459		    state != POOL_STATE_ACTIVE)
1460			return (SET_ERROR(EBADF));
1461
1462		/*
1463		 * If we were able to open and validate a vdev that was
1464		 * previously marked permanently unavailable, clear that state
1465		 * now.
1466		 */
1467		if (vd->vdev_not_present)
1468			vd->vdev_not_present = 0;
1469	}
1470
1471	return (0);
1472}
1473
1474/*
1475 * Close a virtual device.
1476 */
1477void
1478vdev_close(vdev_t *vd)
1479{
1480	spa_t *spa = vd->vdev_spa;
1481	vdev_t *pvd = vd->vdev_parent;
1482
1483	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1484
1485	/*
1486	 * If our parent is reopening, then we are as well, unless we are
1487	 * going offline.
1488	 */
1489	if (pvd != NULL && pvd->vdev_reopening)
1490		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1491
1492	vd->vdev_ops->vdev_op_close(vd);
1493
1494	vdev_cache_purge(vd);
1495
1496	/*
1497	 * We record the previous state before we close it, so that if we are
1498	 * doing a reopen(), we don't generate FMA ereports if we notice that
1499	 * it's still faulted.
1500	 */
1501	vd->vdev_prevstate = vd->vdev_state;
1502
1503	if (vd->vdev_offline)
1504		vd->vdev_state = VDEV_STATE_OFFLINE;
1505	else
1506		vd->vdev_state = VDEV_STATE_CLOSED;
1507	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1508}
1509
1510void
1511vdev_hold(vdev_t *vd)
1512{
1513	spa_t *spa = vd->vdev_spa;
1514
1515	ASSERT(spa_is_root(spa));
1516	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1517		return;
1518
1519	for (int c = 0; c < vd->vdev_children; c++)
1520		vdev_hold(vd->vdev_child[c]);
1521
1522	if (vd->vdev_ops->vdev_op_leaf)
1523		vd->vdev_ops->vdev_op_hold(vd);
1524}
1525
1526void
1527vdev_rele(vdev_t *vd)
1528{
1529	spa_t *spa = vd->vdev_spa;
1530
1531	ASSERT(spa_is_root(spa));
1532	for (int c = 0; c < vd->vdev_children; c++)
1533		vdev_rele(vd->vdev_child[c]);
1534
1535	if (vd->vdev_ops->vdev_op_leaf)
1536		vd->vdev_ops->vdev_op_rele(vd);
1537}
1538
1539/*
1540 * Reopen all interior vdevs and any unopened leaves.  We don't actually
1541 * reopen leaf vdevs which had previously been opened as they might deadlock
1542 * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1543 * If the leaf has never been opened then open it, as usual.
1544 */
1545void
1546vdev_reopen(vdev_t *vd)
1547{
1548	spa_t *spa = vd->vdev_spa;
1549
1550	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1551
1552	/* set the reopening flag unless we're taking the vdev offline */
1553	vd->vdev_reopening = !vd->vdev_offline;
1554	vdev_close(vd);
1555	(void) vdev_open(vd);
1556
1557	/*
1558	 * Call vdev_validate() here to make sure we have the same device.
1559	 * Otherwise, a device with an invalid label could be successfully
1560	 * opened in response to vdev_reopen().
1561	 */
1562	if (vd->vdev_aux) {
1563		(void) vdev_validate_aux(vd);
1564		if (vdev_readable(vd) && vdev_writeable(vd) &&
1565		    vd->vdev_aux == &spa->spa_l2cache &&
1566		    !l2arc_vdev_present(vd))
1567			l2arc_add_vdev(spa, vd);
1568	} else {
1569		(void) vdev_validate(vd, B_TRUE);
1570	}
1571
1572	/*
1573	 * Reassess parent vdev's health.
1574	 */
1575	vdev_propagate_state(vd);
1576}
1577
1578int
1579vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1580{
1581	int error;
1582
1583	/*
1584	 * Normally, partial opens (e.g. of a mirror) are allowed.
1585	 * For a create, however, we want to fail the request if
1586	 * there are any components we can't open.
1587	 */
1588	error = vdev_open(vd);
1589
1590	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1591		vdev_close(vd);
1592		return (error ? error : ENXIO);
1593	}
1594
1595	/*
1596	 * Recursively load DTLs and initialize all labels.
1597	 */
1598	if ((error = vdev_dtl_load(vd)) != 0 ||
1599	    (error = vdev_label_init(vd, txg, isreplacing ?
1600	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1601		vdev_close(vd);
1602		return (error);
1603	}
1604
1605	return (0);
1606}
1607
1608void
1609vdev_metaslab_set_size(vdev_t *vd)
1610{
1611	/*
1612	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1613	 */
1614	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1615	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1616}
1617
1618void
1619vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1620{
1621	ASSERT(vd == vd->vdev_top);
1622	ASSERT(!vd->vdev_ishole);
1623	ASSERT(ISP2(flags));
1624	ASSERT(spa_writeable(vd->vdev_spa));
1625
1626	if (flags & VDD_METASLAB)
1627		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1628
1629	if (flags & VDD_DTL)
1630		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1631
1632	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1633}
1634
1635void
1636vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1637{
1638	for (int c = 0; c < vd->vdev_children; c++)
1639		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1640
1641	if (vd->vdev_ops->vdev_op_leaf)
1642		vdev_dirty(vd->vdev_top, flags, vd, txg);
1643}
1644
1645/*
1646 * DTLs.
1647 *
1648 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1649 * the vdev has less than perfect replication.  There are four kinds of DTL:
1650 *
1651 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1652 *
1653 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1654 *
1655 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1656 *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1657 *	txgs that was scrubbed.
1658 *
1659 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1660 *	persistent errors or just some device being offline.
1661 *	Unlike the other three, the DTL_OUTAGE map is not generally
1662 *	maintained; it's only computed when needed, typically to
1663 *	determine whether a device can be detached.
1664 *
1665 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1666 * either has the data or it doesn't.
1667 *
1668 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1669 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1670 * if any child is less than fully replicated, then so is its parent.
1671 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1672 * comprising only those txgs which appear in 'maxfaults' or more children;
1673 * those are the txgs we don't have enough replication to read.  For example,
1674 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1675 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1676 * two child DTL_MISSING maps.
1677 *
1678 * It should be clear from the above that to compute the DTLs and outage maps
1679 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1680 * Therefore, that is all we keep on disk.  When loading the pool, or after
1681 * a configuration change, we generate all other DTLs from first principles.
1682 */
1683void
1684vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1685{
1686	range_tree_t *rt = vd->vdev_dtl[t];
1687
1688	ASSERT(t < DTL_TYPES);
1689	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1690	ASSERT(spa_writeable(vd->vdev_spa));
1691
1692	mutex_enter(rt->rt_lock);
1693	if (!range_tree_contains(rt, txg, size))
1694		range_tree_add(rt, txg, size);
1695	mutex_exit(rt->rt_lock);
1696}
1697
1698boolean_t
1699vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1700{
1701	range_tree_t *rt = vd->vdev_dtl[t];
1702	boolean_t dirty = B_FALSE;
1703
1704	ASSERT(t < DTL_TYPES);
1705	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1706
1707	mutex_enter(rt->rt_lock);
1708	if (range_tree_space(rt) != 0)
1709		dirty = range_tree_contains(rt, txg, size);
1710	mutex_exit(rt->rt_lock);
1711
1712	return (dirty);
1713}
1714
1715boolean_t
1716vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1717{
1718	range_tree_t *rt = vd->vdev_dtl[t];
1719	boolean_t empty;
1720
1721	mutex_enter(rt->rt_lock);
1722	empty = (range_tree_space(rt) == 0);
1723	mutex_exit(rt->rt_lock);
1724
1725	return (empty);
1726}
1727
1728/*
1729 * Returns the lowest txg in the DTL range.
1730 */
1731static uint64_t
1732vdev_dtl_min(vdev_t *vd)
1733{
1734	range_seg_t *rs;
1735
1736	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1737	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1738	ASSERT0(vd->vdev_children);
1739
1740	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1741	return (rs->rs_start - 1);
1742}
1743
1744/*
1745 * Returns the highest txg in the DTL.
1746 */
1747static uint64_t
1748vdev_dtl_max(vdev_t *vd)
1749{
1750	range_seg_t *rs;
1751
1752	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1753	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1754	ASSERT0(vd->vdev_children);
1755
1756	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1757	return (rs->rs_end);
1758}
1759
1760/*
1761 * Determine if a resilvering vdev should remove any DTL entries from
1762 * its range. If the vdev was resilvering for the entire duration of the
1763 * scan then it should excise that range from its DTLs. Otherwise, this
1764 * vdev is considered partially resilvered and should leave its DTL
1765 * entries intact. The comment in vdev_dtl_reassess() describes how we
1766 * excise the DTLs.
1767 */
1768static boolean_t
1769vdev_dtl_should_excise(vdev_t *vd)
1770{
1771	spa_t *spa = vd->vdev_spa;
1772	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1773
1774	ASSERT0(scn->scn_phys.scn_errors);
1775	ASSERT0(vd->vdev_children);
1776
1777	if (vd->vdev_resilver_txg == 0 ||
1778	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1779		return (B_TRUE);
1780
1781	/*
1782	 * When a resilver is initiated the scan will assign the scn_max_txg
1783	 * value to the highest txg value that exists in all DTLs. If this
1784	 * device's max DTL is not part of this scan (i.e. it is not in
1785	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1786	 * for excision.
1787	 */
1788	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1789		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1790		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1791		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1792		return (B_TRUE);
1793	}
1794	return (B_FALSE);
1795}
1796
1797/*
1798 * Reassess DTLs after a config change or scrub completion.
1799 */
1800void
1801vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1802{
1803	spa_t *spa = vd->vdev_spa;
1804	avl_tree_t reftree;
1805	int minref;
1806
1807	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1808
1809	for (int c = 0; c < vd->vdev_children; c++)
1810		vdev_dtl_reassess(vd->vdev_child[c], txg,
1811		    scrub_txg, scrub_done);
1812
1813	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1814		return;
1815
1816	if (vd->vdev_ops->vdev_op_leaf) {
1817		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1818
1819		mutex_enter(&vd->vdev_dtl_lock);
1820
1821		/*
1822		 * If we've completed a scan cleanly then determine
1823		 * if this vdev should remove any DTLs. We only want to
1824		 * excise regions on vdevs that were available during
1825		 * the entire duration of this scan.
1826		 */
1827		if (scrub_txg != 0 &&
1828		    (spa->spa_scrub_started ||
1829		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1830		    vdev_dtl_should_excise(vd)) {
1831			/*
1832			 * We completed a scrub up to scrub_txg.  If we
1833			 * did it without rebooting, then the scrub dtl
1834			 * will be valid, so excise the old region and
1835			 * fold in the scrub dtl.  Otherwise, leave the
1836			 * dtl as-is if there was an error.
1837			 *
1838			 * There's little trick here: to excise the beginning
1839			 * of the DTL_MISSING map, we put it into a reference
1840			 * tree and then add a segment with refcnt -1 that
1841			 * covers the range [0, scrub_txg).  This means
1842			 * that each txg in that range has refcnt -1 or 0.
1843			 * We then add DTL_SCRUB with a refcnt of 2, so that
1844			 * entries in the range [0, scrub_txg) will have a
1845			 * positive refcnt -- either 1 or 2.  We then convert
1846			 * the reference tree into the new DTL_MISSING map.
1847			 */
1848			space_reftree_create(&reftree);
1849			space_reftree_add_map(&reftree,
1850			    vd->vdev_dtl[DTL_MISSING], 1);
1851			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1852			space_reftree_add_map(&reftree,
1853			    vd->vdev_dtl[DTL_SCRUB], 2);
1854			space_reftree_generate_map(&reftree,
1855			    vd->vdev_dtl[DTL_MISSING], 1);
1856			space_reftree_destroy(&reftree);
1857		}
1858		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1859		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1860		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1861		if (scrub_done)
1862			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1863		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1864		if (!vdev_readable(vd))
1865			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1866		else
1867			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1868			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1869
1870		/*
1871		 * If the vdev was resilvering and no longer has any
1872		 * DTLs then reset its resilvering flag.
1873		 */
1874		if (vd->vdev_resilver_txg != 0 &&
1875		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1876		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1877			vd->vdev_resilver_txg = 0;
1878
1879		mutex_exit(&vd->vdev_dtl_lock);
1880
1881		if (txg != 0)
1882			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1883		return;
1884	}
1885
1886	mutex_enter(&vd->vdev_dtl_lock);
1887	for (int t = 0; t < DTL_TYPES; t++) {
1888		/* account for child's outage in parent's missing map */
1889		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1890		if (t == DTL_SCRUB)
1891			continue;			/* leaf vdevs only */
1892		if (t == DTL_PARTIAL)
1893			minref = 1;			/* i.e. non-zero */
1894		else if (vd->vdev_nparity != 0)
1895			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1896		else
1897			minref = vd->vdev_children;	/* any kind of mirror */
1898		space_reftree_create(&reftree);
1899		for (int c = 0; c < vd->vdev_children; c++) {
1900			vdev_t *cvd = vd->vdev_child[c];
1901			mutex_enter(&cvd->vdev_dtl_lock);
1902			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1903			mutex_exit(&cvd->vdev_dtl_lock);
1904		}
1905		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1906		space_reftree_destroy(&reftree);
1907	}
1908	mutex_exit(&vd->vdev_dtl_lock);
1909}
1910
1911int
1912vdev_dtl_load(vdev_t *vd)
1913{
1914	spa_t *spa = vd->vdev_spa;
1915	objset_t *mos = spa->spa_meta_objset;
1916	int error = 0;
1917
1918	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1919		ASSERT(!vd->vdev_ishole);
1920
1921		error = space_map_open(&vd->vdev_dtl_sm, mos,
1922		    vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1923		if (error)
1924			return (error);
1925		ASSERT(vd->vdev_dtl_sm != NULL);
1926
1927		mutex_enter(&vd->vdev_dtl_lock);
1928
1929		/*
1930		 * Now that we've opened the space_map we need to update
1931		 * the in-core DTL.
1932		 */
1933		space_map_update(vd->vdev_dtl_sm);
1934
1935		error = space_map_load(vd->vdev_dtl_sm,
1936		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1937		mutex_exit(&vd->vdev_dtl_lock);
1938
1939		return (error);
1940	}
1941
1942	for (int c = 0; c < vd->vdev_children; c++) {
1943		error = vdev_dtl_load(vd->vdev_child[c]);
1944		if (error != 0)
1945			break;
1946	}
1947
1948	return (error);
1949}
1950
1951void
1952vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
1953{
1954	spa_t *spa = vd->vdev_spa;
1955
1956	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
1957	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1958	    zapobj, tx));
1959}
1960
1961uint64_t
1962vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
1963{
1964	spa_t *spa = vd->vdev_spa;
1965	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
1966	    DMU_OT_NONE, 0, tx);
1967
1968	ASSERT(zap != 0);
1969	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1970	    zap, tx));
1971
1972	return (zap);
1973}
1974
1975void
1976vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
1977{
1978	if (vd->vdev_ops != &vdev_hole_ops &&
1979	    vd->vdev_ops != &vdev_missing_ops &&
1980	    vd->vdev_ops != &vdev_root_ops &&
1981	    !vd->vdev_top->vdev_removing) {
1982		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
1983			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
1984		}
1985		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
1986			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
1987		}
1988	}
1989	for (uint64_t i = 0; i < vd->vdev_children; i++) {
1990		vdev_construct_zaps(vd->vdev_child[i], tx);
1991	}
1992}
1993
1994void
1995vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1996{
1997	spa_t *spa = vd->vdev_spa;
1998	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
1999	objset_t *mos = spa->spa_meta_objset;
2000	range_tree_t *rtsync;
2001	kmutex_t rtlock;
2002	dmu_tx_t *tx;
2003	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2004
2005	ASSERT(!vd->vdev_ishole);
2006	ASSERT(vd->vdev_ops->vdev_op_leaf);
2007
2008	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2009
2010	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2011		mutex_enter(&vd->vdev_dtl_lock);
2012		space_map_free(vd->vdev_dtl_sm, tx);
2013		space_map_close(vd->vdev_dtl_sm);
2014		vd->vdev_dtl_sm = NULL;
2015		mutex_exit(&vd->vdev_dtl_lock);
2016
2017		/*
2018		 * We only destroy the leaf ZAP for detached leaves or for
2019		 * removed log devices. Removed data devices handle leaf ZAP
2020		 * cleanup later, once cancellation is no longer possible.
2021		 */
2022		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2023		    vd->vdev_top->vdev_islog)) {
2024			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2025			vd->vdev_leaf_zap = 0;
2026		}
2027
2028		dmu_tx_commit(tx);
2029		return;
2030	}
2031
2032	if (vd->vdev_dtl_sm == NULL) {
2033		uint64_t new_object;
2034
2035		new_object = space_map_alloc(mos, tx);
2036		VERIFY3U(new_object, !=, 0);
2037
2038		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2039		    0, -1ULL, 0, &vd->vdev_dtl_lock));
2040		ASSERT(vd->vdev_dtl_sm != NULL);
2041	}
2042
2043	mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
2044
2045	rtsync = range_tree_create(NULL, NULL, &rtlock);
2046
2047	mutex_enter(&rtlock);
2048
2049	mutex_enter(&vd->vdev_dtl_lock);
2050	range_tree_walk(rt, range_tree_add, rtsync);
2051	mutex_exit(&vd->vdev_dtl_lock);
2052
2053	space_map_truncate(vd->vdev_dtl_sm, tx);
2054	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2055	range_tree_vacate(rtsync, NULL, NULL);
2056
2057	range_tree_destroy(rtsync);
2058
2059	mutex_exit(&rtlock);
2060	mutex_destroy(&rtlock);
2061
2062	/*
2063	 * If the object for the space map has changed then dirty
2064	 * the top level so that we update the config.
2065	 */
2066	if (object != space_map_object(vd->vdev_dtl_sm)) {
2067		zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
2068		    "new object %llu", txg, spa_name(spa), object,
2069		    space_map_object(vd->vdev_dtl_sm));
2070		vdev_config_dirty(vd->vdev_top);
2071	}
2072
2073	dmu_tx_commit(tx);
2074
2075	mutex_enter(&vd->vdev_dtl_lock);
2076	space_map_update(vd->vdev_dtl_sm);
2077	mutex_exit(&vd->vdev_dtl_lock);
2078}
2079
2080/*
2081 * Determine whether the specified vdev can be offlined/detached/removed
2082 * without losing data.
2083 */
2084boolean_t
2085vdev_dtl_required(vdev_t *vd)
2086{
2087	spa_t *spa = vd->vdev_spa;
2088	vdev_t *tvd = vd->vdev_top;
2089	uint8_t cant_read = vd->vdev_cant_read;
2090	boolean_t required;
2091
2092	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2093
2094	if (vd == spa->spa_root_vdev || vd == tvd)
2095		return (B_TRUE);
2096
2097	/*
2098	 * Temporarily mark the device as unreadable, and then determine
2099	 * whether this results in any DTL outages in the top-level vdev.
2100	 * If not, we can safely offline/detach/remove the device.
2101	 */
2102	vd->vdev_cant_read = B_TRUE;
2103	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2104	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2105	vd->vdev_cant_read = cant_read;
2106	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2107
2108	if (!required && zio_injection_enabled)
2109		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2110
2111	return (required);
2112}
2113
2114/*
2115 * Determine if resilver is needed, and if so the txg range.
2116 */
2117boolean_t
2118vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2119{
2120	boolean_t needed = B_FALSE;
2121	uint64_t thismin = UINT64_MAX;
2122	uint64_t thismax = 0;
2123
2124	if (vd->vdev_children == 0) {
2125		mutex_enter(&vd->vdev_dtl_lock);
2126		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2127		    vdev_writeable(vd)) {
2128
2129			thismin = vdev_dtl_min(vd);
2130			thismax = vdev_dtl_max(vd);
2131			needed = B_TRUE;
2132		}
2133		mutex_exit(&vd->vdev_dtl_lock);
2134	} else {
2135		for (int c = 0; c < vd->vdev_children; c++) {
2136			vdev_t *cvd = vd->vdev_child[c];
2137			uint64_t cmin, cmax;
2138
2139			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2140				thismin = MIN(thismin, cmin);
2141				thismax = MAX(thismax, cmax);
2142				needed = B_TRUE;
2143			}
2144		}
2145	}
2146
2147	if (needed && minp) {
2148		*minp = thismin;
2149		*maxp = thismax;
2150	}
2151	return (needed);
2152}
2153
2154void
2155vdev_load(vdev_t *vd)
2156{
2157	/*
2158	 * Recursively load all children.
2159	 */
2160	for (int c = 0; c < vd->vdev_children; c++)
2161		vdev_load(vd->vdev_child[c]);
2162
2163	/*
2164	 * If this is a top-level vdev, initialize its metaslabs.
2165	 */
2166	if (vd == vd->vdev_top && !vd->vdev_ishole &&
2167	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2168	    vdev_metaslab_init(vd, 0) != 0))
2169		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2170		    VDEV_AUX_CORRUPT_DATA);
2171
2172	/*
2173	 * If this is a leaf vdev, load its DTL.
2174	 */
2175	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2176		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2177		    VDEV_AUX_CORRUPT_DATA);
2178}
2179
2180/*
2181 * The special vdev case is used for hot spares and l2cache devices.  Its
2182 * sole purpose it to set the vdev state for the associated vdev.  To do this,
2183 * we make sure that we can open the underlying device, then try to read the
2184 * label, and make sure that the label is sane and that it hasn't been
2185 * repurposed to another pool.
2186 */
2187int
2188vdev_validate_aux(vdev_t *vd)
2189{
2190	nvlist_t *label;
2191	uint64_t guid, version;
2192	uint64_t state;
2193
2194	if (!vdev_readable(vd))
2195		return (0);
2196
2197	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2198		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2199		    VDEV_AUX_CORRUPT_DATA);
2200		return (-1);
2201	}
2202
2203	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2204	    !SPA_VERSION_IS_SUPPORTED(version) ||
2205	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2206	    guid != vd->vdev_guid ||
2207	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2208		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2209		    VDEV_AUX_CORRUPT_DATA);
2210		nvlist_free(label);
2211		return (-1);
2212	}
2213
2214	/*
2215	 * We don't actually check the pool state here.  If it's in fact in
2216	 * use by another pool, we update this fact on the fly when requested.
2217	 */
2218	nvlist_free(label);
2219	return (0);
2220}
2221
2222void
2223vdev_remove(vdev_t *vd, uint64_t txg)
2224{
2225	spa_t *spa = vd->vdev_spa;
2226	objset_t *mos = spa->spa_meta_objset;
2227	dmu_tx_t *tx;
2228
2229	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2230	ASSERT(vd == vd->vdev_top);
2231	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2232
2233	if (vd->vdev_ms != NULL) {
2234		metaslab_group_t *mg = vd->vdev_mg;
2235
2236		metaslab_group_histogram_verify(mg);
2237		metaslab_class_histogram_verify(mg->mg_class);
2238
2239		for (int m = 0; m < vd->vdev_ms_count; m++) {
2240			metaslab_t *msp = vd->vdev_ms[m];
2241
2242			if (msp == NULL || msp->ms_sm == NULL)
2243				continue;
2244
2245			mutex_enter(&msp->ms_lock);
2246			/*
2247			 * If the metaslab was not loaded when the vdev
2248			 * was removed then the histogram accounting may
2249			 * not be accurate. Update the histogram information
2250			 * here so that we ensure that the metaslab group
2251			 * and metaslab class are up-to-date.
2252			 */
2253			metaslab_group_histogram_remove(mg, msp);
2254
2255			VERIFY0(space_map_allocated(msp->ms_sm));
2256			space_map_free(msp->ms_sm, tx);
2257			space_map_close(msp->ms_sm);
2258			msp->ms_sm = NULL;
2259			mutex_exit(&msp->ms_lock);
2260		}
2261
2262		metaslab_group_histogram_verify(mg);
2263		metaslab_class_histogram_verify(mg->mg_class);
2264		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2265			ASSERT0(mg->mg_histogram[i]);
2266
2267	}
2268
2269	if (vd->vdev_ms_array) {
2270		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2271		vd->vdev_ms_array = 0;
2272	}
2273
2274	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2275		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2276		vd->vdev_top_zap = 0;
2277	}
2278	dmu_tx_commit(tx);
2279}
2280
2281void
2282vdev_sync_done(vdev_t *vd, uint64_t txg)
2283{
2284	metaslab_t *msp;
2285	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2286
2287	ASSERT(!vd->vdev_ishole);
2288
2289	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2290		metaslab_sync_done(msp, txg);
2291
2292	if (reassess)
2293		metaslab_sync_reassess(vd->vdev_mg);
2294}
2295
2296void
2297vdev_sync(vdev_t *vd, uint64_t txg)
2298{
2299	spa_t *spa = vd->vdev_spa;
2300	vdev_t *lvd;
2301	metaslab_t *msp;
2302	dmu_tx_t *tx;
2303
2304	ASSERT(!vd->vdev_ishole);
2305
2306	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2307		ASSERT(vd == vd->vdev_top);
2308		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2309		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2310		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2311		ASSERT(vd->vdev_ms_array != 0);
2312		vdev_config_dirty(vd);
2313		dmu_tx_commit(tx);
2314	}
2315
2316	/*
2317	 * Remove the metadata associated with this vdev once it's empty.
2318	 */
2319	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2320		vdev_remove(vd, txg);
2321
2322	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2323		metaslab_sync(msp, txg);
2324		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2325	}
2326
2327	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2328		vdev_dtl_sync(lvd, txg);
2329
2330	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2331}
2332
2333uint64_t
2334vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2335{
2336	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2337}
2338
2339/*
2340 * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2341 * not be opened, and no I/O is attempted.
2342 */
2343int
2344vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2345{
2346	vdev_t *vd, *tvd;
2347
2348	spa_vdev_state_enter(spa, SCL_NONE);
2349
2350	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2351		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2352
2353	if (!vd->vdev_ops->vdev_op_leaf)
2354		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2355
2356	tvd = vd->vdev_top;
2357
2358	/*
2359	 * We don't directly use the aux state here, but if we do a
2360	 * vdev_reopen(), we need this value to be present to remember why we
2361	 * were faulted.
2362	 */
2363	vd->vdev_label_aux = aux;
2364
2365	/*
2366	 * Faulted state takes precedence over degraded.
2367	 */
2368	vd->vdev_delayed_close = B_FALSE;
2369	vd->vdev_faulted = 1ULL;
2370	vd->vdev_degraded = 0ULL;
2371	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2372
2373	/*
2374	 * If this device has the only valid copy of the data, then
2375	 * back off and simply mark the vdev as degraded instead.
2376	 */
2377	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2378		vd->vdev_degraded = 1ULL;
2379		vd->vdev_faulted = 0ULL;
2380
2381		/*
2382		 * If we reopen the device and it's not dead, only then do we
2383		 * mark it degraded.
2384		 */
2385		vdev_reopen(tvd);
2386
2387		if (vdev_readable(vd))
2388			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2389	}
2390
2391	return (spa_vdev_state_exit(spa, vd, 0));
2392}
2393
2394/*
2395 * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2396 * user that something is wrong.  The vdev continues to operate as normal as far
2397 * as I/O is concerned.
2398 */
2399int
2400vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2401{
2402	vdev_t *vd;
2403
2404	spa_vdev_state_enter(spa, SCL_NONE);
2405
2406	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2407		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2408
2409	if (!vd->vdev_ops->vdev_op_leaf)
2410		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2411
2412	/*
2413	 * If the vdev is already faulted, then don't do anything.
2414	 */
2415	if (vd->vdev_faulted || vd->vdev_degraded)
2416		return (spa_vdev_state_exit(spa, NULL, 0));
2417
2418	vd->vdev_degraded = 1ULL;
2419	if (!vdev_is_dead(vd))
2420		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2421		    aux);
2422
2423	return (spa_vdev_state_exit(spa, vd, 0));
2424}
2425
2426/*
2427 * Online the given vdev.
2428 *
2429 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2430 * spare device should be detached when the device finishes resilvering.
2431 * Second, the online should be treated like a 'test' online case, so no FMA
2432 * events are generated if the device fails to open.
2433 */
2434int
2435vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2436{
2437	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2438	boolean_t postevent = B_FALSE;
2439
2440	spa_vdev_state_enter(spa, SCL_NONE);
2441
2442	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2443		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2444
2445	if (!vd->vdev_ops->vdev_op_leaf)
2446		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2447
2448	postevent =
2449	    (vd->vdev_offline == B_TRUE || vd->vdev_tmpoffline == B_TRUE) ?
2450	    B_TRUE : B_FALSE;
2451
2452	tvd = vd->vdev_top;
2453	vd->vdev_offline = B_FALSE;
2454	vd->vdev_tmpoffline = B_FALSE;
2455	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2456	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2457
2458	/* XXX - L2ARC 1.0 does not support expansion */
2459	if (!vd->vdev_aux) {
2460		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2461			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2462	}
2463
2464	vdev_reopen(tvd);
2465	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2466
2467	if (!vd->vdev_aux) {
2468		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2469			pvd->vdev_expanding = B_FALSE;
2470	}
2471
2472	if (newstate)
2473		*newstate = vd->vdev_state;
2474	if ((flags & ZFS_ONLINE_UNSPARE) &&
2475	    !vdev_is_dead(vd) && vd->vdev_parent &&
2476	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2477	    vd->vdev_parent->vdev_child[0] == vd)
2478		vd->vdev_unspare = B_TRUE;
2479
2480	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2481
2482		/* XXX - L2ARC 1.0 does not support expansion */
2483		if (vd->vdev_aux)
2484			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2485		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2486	}
2487
2488	if (postevent)
2489		spa_event_notify(spa, vd, ESC_ZFS_VDEV_ONLINE);
2490
2491	return (spa_vdev_state_exit(spa, vd, 0));
2492}
2493
2494static int
2495vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2496{
2497	vdev_t *vd, *tvd;
2498	int error = 0;
2499	uint64_t generation;
2500	metaslab_group_t *mg;
2501
2502top:
2503	spa_vdev_state_enter(spa, SCL_ALLOC);
2504
2505	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2506		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2507
2508	if (!vd->vdev_ops->vdev_op_leaf)
2509		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2510
2511	tvd = vd->vdev_top;
2512	mg = tvd->vdev_mg;
2513	generation = spa->spa_config_generation + 1;
2514
2515	/*
2516	 * If the device isn't already offline, try to offline it.
2517	 */
2518	if (!vd->vdev_offline) {
2519		/*
2520		 * If this device has the only valid copy of some data,
2521		 * don't allow it to be offlined. Log devices are always
2522		 * expendable.
2523		 */
2524		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2525		    vdev_dtl_required(vd))
2526			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2527
2528		/*
2529		 * If the top-level is a slog and it has had allocations
2530		 * then proceed.  We check that the vdev's metaslab group
2531		 * is not NULL since it's possible that we may have just
2532		 * added this vdev but not yet initialized its metaslabs.
2533		 */
2534		if (tvd->vdev_islog && mg != NULL) {
2535			/*
2536			 * Prevent any future allocations.
2537			 */
2538			metaslab_group_passivate(mg);
2539			(void) spa_vdev_state_exit(spa, vd, 0);
2540
2541			error = spa_offline_log(spa);
2542
2543			spa_vdev_state_enter(spa, SCL_ALLOC);
2544
2545			/*
2546			 * Check to see if the config has changed.
2547			 */
2548			if (error || generation != spa->spa_config_generation) {
2549				metaslab_group_activate(mg);
2550				if (error)
2551					return (spa_vdev_state_exit(spa,
2552					    vd, error));
2553				(void) spa_vdev_state_exit(spa, vd, 0);
2554				goto top;
2555			}
2556			ASSERT0(tvd->vdev_stat.vs_alloc);
2557		}
2558
2559		/*
2560		 * Offline this device and reopen its top-level vdev.
2561		 * If the top-level vdev is a log device then just offline
2562		 * it. Otherwise, if this action results in the top-level
2563		 * vdev becoming unusable, undo it and fail the request.
2564		 */
2565		vd->vdev_offline = B_TRUE;
2566		vdev_reopen(tvd);
2567
2568		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2569		    vdev_is_dead(tvd)) {
2570			vd->vdev_offline = B_FALSE;
2571			vdev_reopen(tvd);
2572			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2573		}
2574
2575		/*
2576		 * Add the device back into the metaslab rotor so that
2577		 * once we online the device it's open for business.
2578		 */
2579		if (tvd->vdev_islog && mg != NULL)
2580			metaslab_group_activate(mg);
2581	}
2582
2583	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2584
2585	return (spa_vdev_state_exit(spa, vd, 0));
2586}
2587
2588int
2589vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2590{
2591	int error;
2592
2593	mutex_enter(&spa->spa_vdev_top_lock);
2594	error = vdev_offline_locked(spa, guid, flags);
2595	mutex_exit(&spa->spa_vdev_top_lock);
2596
2597	return (error);
2598}
2599
2600/*
2601 * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2602 * vdev_offline(), we assume the spa config is locked.  We also clear all
2603 * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2604 */
2605void
2606vdev_clear(spa_t *spa, vdev_t *vd)
2607{
2608	vdev_t *rvd = spa->spa_root_vdev;
2609
2610	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2611
2612	if (vd == NULL)
2613		vd = rvd;
2614
2615	vd->vdev_stat.vs_read_errors = 0;
2616	vd->vdev_stat.vs_write_errors = 0;
2617	vd->vdev_stat.vs_checksum_errors = 0;
2618
2619	for (int c = 0; c < vd->vdev_children; c++)
2620		vdev_clear(spa, vd->vdev_child[c]);
2621
2622	/*
2623	 * If we're in the FAULTED state or have experienced failed I/O, then
2624	 * clear the persistent state and attempt to reopen the device.  We
2625	 * also mark the vdev config dirty, so that the new faulted state is
2626	 * written out to disk.
2627	 */
2628	if (vd->vdev_faulted || vd->vdev_degraded ||
2629	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2630
2631		/*
2632		 * When reopening in reponse to a clear event, it may be due to
2633		 * a fmadm repair request.  In this case, if the device is
2634		 * still broken, we want to still post the ereport again.
2635		 */
2636		vd->vdev_forcefault = B_TRUE;
2637
2638		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2639		vd->vdev_cant_read = B_FALSE;
2640		vd->vdev_cant_write = B_FALSE;
2641
2642		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2643
2644		vd->vdev_forcefault = B_FALSE;
2645
2646		if (vd != rvd && vdev_writeable(vd->vdev_top))
2647			vdev_state_dirty(vd->vdev_top);
2648
2649		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2650			spa_async_request(spa, SPA_ASYNC_RESILVER);
2651
2652		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2653	}
2654
2655	/*
2656	 * When clearing a FMA-diagnosed fault, we always want to
2657	 * unspare the device, as we assume that the original spare was
2658	 * done in response to the FMA fault.
2659	 */
2660	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2661	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2662	    vd->vdev_parent->vdev_child[0] == vd)
2663		vd->vdev_unspare = B_TRUE;
2664}
2665
2666boolean_t
2667vdev_is_dead(vdev_t *vd)
2668{
2669	/*
2670	 * Holes and missing devices are always considered "dead".
2671	 * This simplifies the code since we don't have to check for
2672	 * these types of devices in the various code paths.
2673	 * Instead we rely on the fact that we skip over dead devices
2674	 * before issuing I/O to them.
2675	 */
2676	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2677	    vd->vdev_ops == &vdev_missing_ops);
2678}
2679
2680boolean_t
2681vdev_readable(vdev_t *vd)
2682{
2683	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2684}
2685
2686boolean_t
2687vdev_writeable(vdev_t *vd)
2688{
2689	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2690}
2691
2692boolean_t
2693vdev_allocatable(vdev_t *vd)
2694{
2695	uint64_t state = vd->vdev_state;
2696
2697	/*
2698	 * We currently allow allocations from vdevs which may be in the
2699	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2700	 * fails to reopen then we'll catch it later when we're holding
2701	 * the proper locks.  Note that we have to get the vdev state
2702	 * in a local variable because although it changes atomically,
2703	 * we're asking two separate questions about it.
2704	 */
2705	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2706	    !vd->vdev_cant_write && !vd->vdev_ishole);
2707}
2708
2709boolean_t
2710vdev_accessible(vdev_t *vd, zio_t *zio)
2711{
2712	ASSERT(zio->io_vd == vd);
2713
2714	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2715		return (B_FALSE);
2716
2717	if (zio->io_type == ZIO_TYPE_READ)
2718		return (!vd->vdev_cant_read);
2719
2720	if (zio->io_type == ZIO_TYPE_WRITE)
2721		return (!vd->vdev_cant_write);
2722
2723	return (B_TRUE);
2724}
2725
2726/*
2727 * Get statistics for the given vdev.
2728 */
2729void
2730vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2731{
2732	spa_t *spa = vd->vdev_spa;
2733	vdev_t *rvd = spa->spa_root_vdev;
2734
2735	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2736
2737	mutex_enter(&vd->vdev_stat_lock);
2738	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2739	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2740	vs->vs_state = vd->vdev_state;
2741	vs->vs_rsize = vdev_get_min_asize(vd);
2742	if (vd->vdev_ops->vdev_op_leaf)
2743		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2744	vs->vs_esize = vd->vdev_max_asize - vd->vdev_asize;
2745	if (vd->vdev_aux == NULL && vd == vd->vdev_top && !vd->vdev_ishole) {
2746		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
2747	}
2748
2749	/*
2750	 * If we're getting stats on the root vdev, aggregate the I/O counts
2751	 * over all top-level vdevs (i.e. the direct children of the root).
2752	 */
2753	if (vd == rvd) {
2754		for (int c = 0; c < rvd->vdev_children; c++) {
2755			vdev_t *cvd = rvd->vdev_child[c];
2756			vdev_stat_t *cvs = &cvd->vdev_stat;
2757
2758			for (int t = 0; t < ZIO_TYPES; t++) {
2759				vs->vs_ops[t] += cvs->vs_ops[t];
2760				vs->vs_bytes[t] += cvs->vs_bytes[t];
2761			}
2762			cvs->vs_scan_removing = cvd->vdev_removing;
2763		}
2764	}
2765	mutex_exit(&vd->vdev_stat_lock);
2766}
2767
2768void
2769vdev_clear_stats(vdev_t *vd)
2770{
2771	mutex_enter(&vd->vdev_stat_lock);
2772	vd->vdev_stat.vs_space = 0;
2773	vd->vdev_stat.vs_dspace = 0;
2774	vd->vdev_stat.vs_alloc = 0;
2775	mutex_exit(&vd->vdev_stat_lock);
2776}
2777
2778void
2779vdev_scan_stat_init(vdev_t *vd)
2780{
2781	vdev_stat_t *vs = &vd->vdev_stat;
2782
2783	for (int c = 0; c < vd->vdev_children; c++)
2784		vdev_scan_stat_init(vd->vdev_child[c]);
2785
2786	mutex_enter(&vd->vdev_stat_lock);
2787	vs->vs_scan_processed = 0;
2788	mutex_exit(&vd->vdev_stat_lock);
2789}
2790
2791void
2792vdev_stat_update(zio_t *zio, uint64_t psize)
2793{
2794	spa_t *spa = zio->io_spa;
2795	vdev_t *rvd = spa->spa_root_vdev;
2796	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2797	vdev_t *pvd;
2798	uint64_t txg = zio->io_txg;
2799	vdev_stat_t *vs = &vd->vdev_stat;
2800	zio_type_t type = zio->io_type;
2801	int flags = zio->io_flags;
2802
2803	/*
2804	 * If this i/o is a gang leader, it didn't do any actual work.
2805	 */
2806	if (zio->io_gang_tree)
2807		return;
2808
2809	if (zio->io_error == 0) {
2810		/*
2811		 * If this is a root i/o, don't count it -- we've already
2812		 * counted the top-level vdevs, and vdev_get_stats() will
2813		 * aggregate them when asked.  This reduces contention on
2814		 * the root vdev_stat_lock and implicitly handles blocks
2815		 * that compress away to holes, for which there is no i/o.
2816		 * (Holes never create vdev children, so all the counters
2817		 * remain zero, which is what we want.)
2818		 *
2819		 * Note: this only applies to successful i/o (io_error == 0)
2820		 * because unlike i/o counts, errors are not additive.
2821		 * When reading a ditto block, for example, failure of
2822		 * one top-level vdev does not imply a root-level error.
2823		 */
2824		if (vd == rvd)
2825			return;
2826
2827		ASSERT(vd == zio->io_vd);
2828
2829		if (flags & ZIO_FLAG_IO_BYPASS)
2830			return;
2831
2832		mutex_enter(&vd->vdev_stat_lock);
2833
2834		if (flags & ZIO_FLAG_IO_REPAIR) {
2835			if (flags & ZIO_FLAG_SCAN_THREAD) {
2836				dsl_scan_phys_t *scn_phys =
2837				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2838				uint64_t *processed = &scn_phys->scn_processed;
2839
2840				/* XXX cleanup? */
2841				if (vd->vdev_ops->vdev_op_leaf)
2842					atomic_add_64(processed, psize);
2843				vs->vs_scan_processed += psize;
2844			}
2845
2846			if (flags & ZIO_FLAG_SELF_HEAL)
2847				vs->vs_self_healed += psize;
2848		}
2849
2850		vs->vs_ops[type]++;
2851		vs->vs_bytes[type] += psize;
2852
2853		mutex_exit(&vd->vdev_stat_lock);
2854		return;
2855	}
2856
2857	if (flags & ZIO_FLAG_SPECULATIVE)
2858		return;
2859
2860	/*
2861	 * If this is an I/O error that is going to be retried, then ignore the
2862	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2863	 * hard errors, when in reality they can happen for any number of
2864	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2865	 */
2866	if (zio->io_error == EIO &&
2867	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2868		return;
2869
2870	/*
2871	 * Intent logs writes won't propagate their error to the root
2872	 * I/O so don't mark these types of failures as pool-level
2873	 * errors.
2874	 */
2875	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2876		return;
2877
2878	mutex_enter(&vd->vdev_stat_lock);
2879	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2880		if (zio->io_error == ECKSUM)
2881			vs->vs_checksum_errors++;
2882		else
2883			vs->vs_read_errors++;
2884	}
2885	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2886		vs->vs_write_errors++;
2887	mutex_exit(&vd->vdev_stat_lock);
2888
2889	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2890	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2891	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2892	    spa->spa_claiming)) {
2893		/*
2894		 * This is either a normal write (not a repair), or it's
2895		 * a repair induced by the scrub thread, or it's a repair
2896		 * made by zil_claim() during spa_load() in the first txg.
2897		 * In the normal case, we commit the DTL change in the same
2898		 * txg as the block was born.  In the scrub-induced repair
2899		 * case, we know that scrubs run in first-pass syncing context,
2900		 * so we commit the DTL change in spa_syncing_txg(spa).
2901		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2902		 *
2903		 * We currently do not make DTL entries for failed spontaneous
2904		 * self-healing writes triggered by normal (non-scrubbing)
2905		 * reads, because we have no transactional context in which to
2906		 * do so -- and it's not clear that it'd be desirable anyway.
2907		 */
2908		if (vd->vdev_ops->vdev_op_leaf) {
2909			uint64_t commit_txg = txg;
2910			if (flags & ZIO_FLAG_SCAN_THREAD) {
2911				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2912				ASSERT(spa_sync_pass(spa) == 1);
2913				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2914				commit_txg = spa_syncing_txg(spa);
2915			} else if (spa->spa_claiming) {
2916				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2917				commit_txg = spa_first_txg(spa);
2918			}
2919			ASSERT(commit_txg >= spa_syncing_txg(spa));
2920			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2921				return;
2922			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2923				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2924			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2925		}
2926		if (vd != rvd)
2927			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2928	}
2929}
2930
2931/*
2932 * Update the in-core space usage stats for this vdev, its metaslab class,
2933 * and the root vdev.
2934 */
2935void
2936vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2937    int64_t space_delta)
2938{
2939	int64_t dspace_delta = space_delta;
2940	spa_t *spa = vd->vdev_spa;
2941	vdev_t *rvd = spa->spa_root_vdev;
2942	metaslab_group_t *mg = vd->vdev_mg;
2943	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2944
2945	ASSERT(vd == vd->vdev_top);
2946
2947	/*
2948	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2949	 * factor.  We must calculate this here and not at the root vdev
2950	 * because the root vdev's psize-to-asize is simply the max of its
2951	 * childrens', thus not accurate enough for us.
2952	 */
2953	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2954	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2955	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2956	    vd->vdev_deflate_ratio;
2957
2958	mutex_enter(&vd->vdev_stat_lock);
2959	vd->vdev_stat.vs_alloc += alloc_delta;
2960	vd->vdev_stat.vs_space += space_delta;
2961	vd->vdev_stat.vs_dspace += dspace_delta;
2962	mutex_exit(&vd->vdev_stat_lock);
2963
2964	if (mc == spa_normal_class(spa)) {
2965		mutex_enter(&rvd->vdev_stat_lock);
2966		rvd->vdev_stat.vs_alloc += alloc_delta;
2967		rvd->vdev_stat.vs_space += space_delta;
2968		rvd->vdev_stat.vs_dspace += dspace_delta;
2969		mutex_exit(&rvd->vdev_stat_lock);
2970	}
2971
2972	if (mc != NULL) {
2973		ASSERT(rvd == vd->vdev_parent);
2974		ASSERT(vd->vdev_ms_count != 0);
2975
2976		metaslab_class_space_update(mc,
2977		    alloc_delta, defer_delta, space_delta, dspace_delta);
2978	}
2979}
2980
2981/*
2982 * Mark a top-level vdev's config as dirty, placing it on the dirty list
2983 * so that it will be written out next time the vdev configuration is synced.
2984 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2985 */
2986void
2987vdev_config_dirty(vdev_t *vd)
2988{
2989	spa_t *spa = vd->vdev_spa;
2990	vdev_t *rvd = spa->spa_root_vdev;
2991	int c;
2992
2993	ASSERT(spa_writeable(spa));
2994
2995	/*
2996	 * If this is an aux vdev (as with l2cache and spare devices), then we
2997	 * update the vdev config manually and set the sync flag.
2998	 */
2999	if (vd->vdev_aux != NULL) {
3000		spa_aux_vdev_t *sav = vd->vdev_aux;
3001		nvlist_t **aux;
3002		uint_t naux;
3003
3004		for (c = 0; c < sav->sav_count; c++) {
3005			if (sav->sav_vdevs[c] == vd)
3006				break;
3007		}
3008
3009		if (c == sav->sav_count) {
3010			/*
3011			 * We're being removed.  There's nothing more to do.
3012			 */
3013			ASSERT(sav->sav_sync == B_TRUE);
3014			return;
3015		}
3016
3017		sav->sav_sync = B_TRUE;
3018
3019		if (nvlist_lookup_nvlist_array(sav->sav_config,
3020		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3021			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3022			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3023		}
3024
3025		ASSERT(c < naux);
3026
3027		/*
3028		 * Setting the nvlist in the middle if the array is a little
3029		 * sketchy, but it will work.
3030		 */
3031		nvlist_free(aux[c]);
3032		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3033
3034		return;
3035	}
3036
3037	/*
3038	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3039	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3040	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3041	 * so this is sufficient to ensure mutual exclusion.
3042	 */
3043	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3044	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3045	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3046
3047	if (vd == rvd) {
3048		for (c = 0; c < rvd->vdev_children; c++)
3049			vdev_config_dirty(rvd->vdev_child[c]);
3050	} else {
3051		ASSERT(vd == vd->vdev_top);
3052
3053		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3054		    !vd->vdev_ishole)
3055			list_insert_head(&spa->spa_config_dirty_list, vd);
3056	}
3057}
3058
3059void
3060vdev_config_clean(vdev_t *vd)
3061{
3062	spa_t *spa = vd->vdev_spa;
3063
3064	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3065	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3066	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3067
3068	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3069	list_remove(&spa->spa_config_dirty_list, vd);
3070}
3071
3072/*
3073 * Mark a top-level vdev's state as dirty, so that the next pass of
3074 * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3075 * the state changes from larger config changes because they require
3076 * much less locking, and are often needed for administrative actions.
3077 */
3078void
3079vdev_state_dirty(vdev_t *vd)
3080{
3081	spa_t *spa = vd->vdev_spa;
3082
3083	ASSERT(spa_writeable(spa));
3084	ASSERT(vd == vd->vdev_top);
3085
3086	/*
3087	 * The state list is protected by the SCL_STATE lock.  The caller
3088	 * must either hold SCL_STATE as writer, or must be the sync thread
3089	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3090	 * so this is sufficient to ensure mutual exclusion.
3091	 */
3092	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3093	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3094	    spa_config_held(spa, SCL_STATE, RW_READER)));
3095
3096	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
3097		list_insert_head(&spa->spa_state_dirty_list, vd);
3098}
3099
3100void
3101vdev_state_clean(vdev_t *vd)
3102{
3103	spa_t *spa = vd->vdev_spa;
3104
3105	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3106	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3107	    spa_config_held(spa, SCL_STATE, RW_READER)));
3108
3109	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3110	list_remove(&spa->spa_state_dirty_list, vd);
3111}
3112
3113/*
3114 * Propagate vdev state up from children to parent.
3115 */
3116void
3117vdev_propagate_state(vdev_t *vd)
3118{
3119	spa_t *spa = vd->vdev_spa;
3120	vdev_t *rvd = spa->spa_root_vdev;
3121	int degraded = 0, faulted = 0;
3122	int corrupted = 0;
3123	vdev_t *child;
3124
3125	if (vd->vdev_children > 0) {
3126		for (int c = 0; c < vd->vdev_children; c++) {
3127			child = vd->vdev_child[c];
3128
3129			/*
3130			 * Don't factor holes into the decision.
3131			 */
3132			if (child->vdev_ishole)
3133				continue;
3134
3135			if (!vdev_readable(child) ||
3136			    (!vdev_writeable(child) && spa_writeable(spa))) {
3137				/*
3138				 * Root special: if there is a top-level log
3139				 * device, treat the root vdev as if it were
3140				 * degraded.
3141				 */
3142				if (child->vdev_islog && vd == rvd)
3143					degraded++;
3144				else
3145					faulted++;
3146			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3147				degraded++;
3148			}
3149
3150			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3151				corrupted++;
3152		}
3153
3154		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3155
3156		/*
3157		 * Root special: if there is a top-level vdev that cannot be
3158		 * opened due to corrupted metadata, then propagate the root
3159		 * vdev's aux state as 'corrupt' rather than 'insufficient
3160		 * replicas'.
3161		 */
3162		if (corrupted && vd == rvd &&
3163		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3164			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3165			    VDEV_AUX_CORRUPT_DATA);
3166	}
3167
3168	if (vd->vdev_parent)
3169		vdev_propagate_state(vd->vdev_parent);
3170}
3171
3172/*
3173 * Set a vdev's state.  If this is during an open, we don't update the parent
3174 * state, because we're in the process of opening children depth-first.
3175 * Otherwise, we propagate the change to the parent.
3176 *
3177 * If this routine places a device in a faulted state, an appropriate ereport is
3178 * generated.
3179 */
3180void
3181vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3182{
3183	uint64_t save_state;
3184	spa_t *spa = vd->vdev_spa;
3185
3186	if (state == vd->vdev_state) {
3187		vd->vdev_stat.vs_aux = aux;
3188		return;
3189	}
3190
3191	save_state = vd->vdev_state;
3192
3193	vd->vdev_state = state;
3194	vd->vdev_stat.vs_aux = aux;
3195
3196	/*
3197	 * If we are setting the vdev state to anything but an open state, then
3198	 * always close the underlying device unless the device has requested
3199	 * a delayed close (i.e. we're about to remove or fault the device).
3200	 * Otherwise, we keep accessible but invalid devices open forever.
3201	 * We don't call vdev_close() itself, because that implies some extra
3202	 * checks (offline, etc) that we don't want here.  This is limited to
3203	 * leaf devices, because otherwise closing the device will affect other
3204	 * children.
3205	 */
3206	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3207	    vd->vdev_ops->vdev_op_leaf)
3208		vd->vdev_ops->vdev_op_close(vd);
3209
3210	/*
3211	 * If we have brought this vdev back into service, we need
3212	 * to notify fmd so that it can gracefully repair any outstanding
3213	 * cases due to a missing device.  We do this in all cases, even those
3214	 * that probably don't correlate to a repaired fault.  This is sure to
3215	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3216	 * this is a transient state it's OK, as the retire agent will
3217	 * double-check the state of the vdev before repairing it.
3218	 */
3219	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3220	    vd->vdev_prevstate != state)
3221		zfs_post_state_change(spa, vd);
3222
3223	if (vd->vdev_removed &&
3224	    state == VDEV_STATE_CANT_OPEN &&
3225	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3226		/*
3227		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3228		 * device was previously marked removed and someone attempted to
3229		 * reopen it.  If this failed due to a nonexistent device, then
3230		 * keep the device in the REMOVED state.  We also let this be if
3231		 * it is one of our special test online cases, which is only
3232		 * attempting to online the device and shouldn't generate an FMA
3233		 * fault.
3234		 */
3235		vd->vdev_state = VDEV_STATE_REMOVED;
3236		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3237	} else if (state == VDEV_STATE_REMOVED) {
3238		vd->vdev_removed = B_TRUE;
3239	} else if (state == VDEV_STATE_CANT_OPEN) {
3240		/*
3241		 * If we fail to open a vdev during an import or recovery, we
3242		 * mark it as "not available", which signifies that it was
3243		 * never there to begin with.  Failure to open such a device
3244		 * is not considered an error.
3245		 */
3246		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3247		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3248		    vd->vdev_ops->vdev_op_leaf)
3249			vd->vdev_not_present = 1;
3250
3251		/*
3252		 * Post the appropriate ereport.  If the 'prevstate' field is
3253		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3254		 * that this is part of a vdev_reopen().  In this case, we don't
3255		 * want to post the ereport if the device was already in the
3256		 * CANT_OPEN state beforehand.
3257		 *
3258		 * If the 'checkremove' flag is set, then this is an attempt to
3259		 * online the device in response to an insertion event.  If we
3260		 * hit this case, then we have detected an insertion event for a
3261		 * faulted or offline device that wasn't in the removed state.
3262		 * In this scenario, we don't post an ereport because we are
3263		 * about to replace the device, or attempt an online with
3264		 * vdev_forcefault, which will generate the fault for us.
3265		 */
3266		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3267		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3268		    vd != spa->spa_root_vdev) {
3269			const char *class;
3270
3271			switch (aux) {
3272			case VDEV_AUX_OPEN_FAILED:
3273				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3274				break;
3275			case VDEV_AUX_CORRUPT_DATA:
3276				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3277				break;
3278			case VDEV_AUX_NO_REPLICAS:
3279				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3280				break;
3281			case VDEV_AUX_BAD_GUID_SUM:
3282				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3283				break;
3284			case VDEV_AUX_TOO_SMALL:
3285				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3286				break;
3287			case VDEV_AUX_BAD_LABEL:
3288				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3289				break;
3290			default:
3291				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3292			}
3293
3294			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3295		}
3296
3297		/* Erase any notion of persistent removed state */
3298		vd->vdev_removed = B_FALSE;
3299	} else {
3300		vd->vdev_removed = B_FALSE;
3301	}
3302
3303	if (!isopen && vd->vdev_parent)
3304		vdev_propagate_state(vd->vdev_parent);
3305}
3306
3307/*
3308 * Check the vdev configuration to ensure that it's capable of supporting
3309 * a root pool. Currently, we do not support RAID-Z or partial configuration.
3310 * In addition, only a single top-level vdev is allowed and none of the leaves
3311 * can be wholedisks.
3312 */
3313boolean_t
3314vdev_is_bootable(vdev_t *vd)
3315{
3316	if (!vd->vdev_ops->vdev_op_leaf) {
3317		char *vdev_type = vd->vdev_ops->vdev_op_type;
3318
3319		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3320		    vd->vdev_children > 1) {
3321			return (B_FALSE);
3322		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3323		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3324			return (B_FALSE);
3325		}
3326	}
3327
3328	for (int c = 0; c < vd->vdev_children; c++) {
3329		if (!vdev_is_bootable(vd->vdev_child[c]))
3330			return (B_FALSE);
3331	}
3332	return (B_TRUE);
3333}
3334
3335/*
3336 * Load the state from the original vdev tree (ovd) which
3337 * we've retrieved from the MOS config object. If the original
3338 * vdev was offline or faulted then we transfer that state to the
3339 * device in the current vdev tree (nvd).
3340 */
3341void
3342vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3343{
3344	spa_t *spa = nvd->vdev_spa;
3345
3346	ASSERT(nvd->vdev_top->vdev_islog);
3347	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3348	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3349
3350	for (int c = 0; c < nvd->vdev_children; c++)
3351		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3352
3353	if (nvd->vdev_ops->vdev_op_leaf) {
3354		/*
3355		 * Restore the persistent vdev state
3356		 */
3357		nvd->vdev_offline = ovd->vdev_offline;
3358		nvd->vdev_faulted = ovd->vdev_faulted;
3359		nvd->vdev_degraded = ovd->vdev_degraded;
3360		nvd->vdev_removed = ovd->vdev_removed;
3361	}
3362}
3363
3364/*
3365 * Determine if a log device has valid content.  If the vdev was
3366 * removed or faulted in the MOS config then we know that
3367 * the content on the log device has already been written to the pool.
3368 */
3369boolean_t
3370vdev_log_state_valid(vdev_t *vd)
3371{
3372	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3373	    !vd->vdev_removed)
3374		return (B_TRUE);
3375
3376	for (int c = 0; c < vd->vdev_children; c++)
3377		if (vdev_log_state_valid(vd->vdev_child[c]))
3378			return (B_TRUE);
3379
3380	return (B_FALSE);
3381}
3382
3383/*
3384 * Expand a vdev if possible.
3385 */
3386void
3387vdev_expand(vdev_t *vd, uint64_t txg)
3388{
3389	ASSERT(vd->vdev_top == vd);
3390	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3391
3392	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3393		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3394		vdev_config_dirty(vd);
3395	}
3396}
3397
3398/*
3399 * Split a vdev.
3400 */
3401void
3402vdev_split(vdev_t *vd)
3403{
3404	vdev_t *cvd, *pvd = vd->vdev_parent;
3405
3406	vdev_remove_child(pvd, vd);
3407	vdev_compact_children(pvd);
3408
3409	cvd = pvd->vdev_child[0];
3410	if (pvd->vdev_children == 1) {
3411		vdev_remove_parent(cvd);
3412		cvd->vdev_splitting = B_TRUE;
3413	}
3414	vdev_propagate_state(cvd);
3415}
3416
3417void
3418vdev_deadman(vdev_t *vd)
3419{
3420	for (int c = 0; c < vd->vdev_children; c++) {
3421		vdev_t *cvd = vd->vdev_child[c];
3422
3423		vdev_deadman(cvd);
3424	}
3425
3426	if (vd->vdev_ops->vdev_op_leaf) {
3427		vdev_queue_t *vq = &vd->vdev_queue;
3428
3429		mutex_enter(&vq->vq_lock);
3430		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3431			spa_t *spa = vd->vdev_spa;
3432			zio_t *fio;
3433			uint64_t delta;
3434
3435			/*
3436			 * Look at the head of all the pending queues,
3437			 * if any I/O has been outstanding for longer than
3438			 * the spa_deadman_synctime we panic the system.
3439			 */
3440			fio = avl_first(&vq->vq_active_tree);
3441			delta = gethrtime() - fio->io_timestamp;
3442			if (delta > spa_deadman_synctime(spa)) {
3443				zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3444				    "delta %lluns, last io %lluns",
3445				    fio->io_timestamp, delta,
3446				    vq->vq_io_complete_ts);
3447				fm_panic("I/O to pool '%s' appears to be "
3448				    "hung.", spa_name(spa));
3449			}
3450		}
3451		mutex_exit(&vq->vq_lock);
3452	}
3453}
3454