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 * Copyright 2016 Toomas Soome <tsoome@me.com>
28 */
29
30#include <sys/zfs_context.h>
31#include <sys/fm/fs/zfs.h>
32#include <sys/spa.h>
33#include <sys/spa_impl.h>
34#include <sys/dmu.h>
35#include <sys/dmu_tx.h>
36#include <sys/vdev_impl.h>
37#include <sys/uberblock_impl.h>
38#include <sys/metaslab.h>
39#include <sys/metaslab_impl.h>
40#include <sys/space_map.h>
41#include <sys/space_reftree.h>
42#include <sys/zio.h>
43#include <sys/zap.h>
44#include <sys/fs/zfs.h>
45#include <sys/arc.h>
46#include <sys/zil.h>
47#include <sys/dsl_scan.h>
48
49/*
50 * Virtual device management.
51 */
52
53static vdev_ops_t *vdev_ops_table[] = {
54	&vdev_root_ops,
55	&vdev_raidz_ops,
56	&vdev_mirror_ops,
57	&vdev_replacing_ops,
58	&vdev_spare_ops,
59	&vdev_disk_ops,
60	&vdev_file_ops,
61	&vdev_missing_ops,
62	&vdev_hole_ops,
63	NULL
64};
65
66/* maximum scrub/resilver I/O queue per leaf vdev */
67int zfs_scrub_limit = 10;
68
69/*
70 * When a vdev is added, it will be divided into approximately (but no
71 * more than) this number of metaslabs.
72 */
73int metaslabs_per_vdev = 200;
74
75/*
76 * Given a vdev type, return the appropriate ops vector.
77 */
78static vdev_ops_t *
79vdev_getops(const char *type)
80{
81	vdev_ops_t *ops, **opspp;
82
83	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
84		if (strcmp(ops->vdev_op_type, type) == 0)
85			break;
86
87	return (ops);
88}
89
90/*
91 * Default asize function: return the MAX of psize with the asize of
92 * all children.  This is what's used by anything other than RAID-Z.
93 */
94uint64_t
95vdev_default_asize(vdev_t *vd, uint64_t psize)
96{
97	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
98	uint64_t csize;
99
100	for (int c = 0; c < vd->vdev_children; c++) {
101		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
102		asize = MAX(asize, csize);
103	}
104
105	return (asize);
106}
107
108/*
109 * Get the minimum allocatable size. We define the allocatable size as
110 * the vdev's asize rounded to the nearest metaslab. This allows us to
111 * replace or attach devices which don't have the same physical size but
112 * can still satisfy the same number of allocations.
113 */
114uint64_t
115vdev_get_min_asize(vdev_t *vd)
116{
117	vdev_t *pvd = vd->vdev_parent;
118
119	/*
120	 * If our parent is NULL (inactive spare or cache) or is the root,
121	 * just return our own asize.
122	 */
123	if (pvd == NULL)
124		return (vd->vdev_asize);
125
126	/*
127	 * The top-level vdev just returns the allocatable size rounded
128	 * to the nearest metaslab.
129	 */
130	if (vd == vd->vdev_top)
131		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
132
133	/*
134	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
135	 * so each child must provide at least 1/Nth of its asize.
136	 */
137	if (pvd->vdev_ops == &vdev_raidz_ops)
138		return (pvd->vdev_min_asize / pvd->vdev_children);
139
140	return (pvd->vdev_min_asize);
141}
142
143void
144vdev_set_min_asize(vdev_t *vd)
145{
146	vd->vdev_min_asize = vdev_get_min_asize(vd);
147
148	for (int c = 0; c < vd->vdev_children; c++)
149		vdev_set_min_asize(vd->vdev_child[c]);
150}
151
152vdev_t *
153vdev_lookup_top(spa_t *spa, uint64_t vdev)
154{
155	vdev_t *rvd = spa->spa_root_vdev;
156
157	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
158
159	if (vdev < rvd->vdev_children) {
160		ASSERT(rvd->vdev_child[vdev] != NULL);
161		return (rvd->vdev_child[vdev]);
162	}
163
164	return (NULL);
165}
166
167vdev_t *
168vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
169{
170	vdev_t *mvd;
171
172	if (vd->vdev_guid == guid)
173		return (vd);
174
175	for (int c = 0; c < vd->vdev_children; c++)
176		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
177		    NULL)
178			return (mvd);
179
180	return (NULL);
181}
182
183static int
184vdev_count_leaves_impl(vdev_t *vd)
185{
186	int n = 0;
187
188	if (vd->vdev_ops->vdev_op_leaf)
189		return (1);
190
191	for (int c = 0; c < vd->vdev_children; c++)
192		n += vdev_count_leaves_impl(vd->vdev_child[c]);
193
194	return (n);
195}
196
197int
198vdev_count_leaves(spa_t *spa)
199{
200	return (vdev_count_leaves_impl(spa->spa_root_vdev));
201}
202
203void
204vdev_add_child(vdev_t *pvd, vdev_t *cvd)
205{
206	size_t oldsize, newsize;
207	uint64_t id = cvd->vdev_id;
208	vdev_t **newchild;
209	spa_t *spa = cvd->vdev_spa;
210
211	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
212	ASSERT(cvd->vdev_parent == NULL);
213
214	cvd->vdev_parent = pvd;
215
216	if (pvd == NULL)
217		return;
218
219	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
220
221	oldsize = pvd->vdev_children * sizeof (vdev_t *);
222	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
223	newsize = pvd->vdev_children * sizeof (vdev_t *);
224
225	newchild = kmem_zalloc(newsize, KM_SLEEP);
226	if (pvd->vdev_child != NULL) {
227		bcopy(pvd->vdev_child, newchild, oldsize);
228		kmem_free(pvd->vdev_child, oldsize);
229	}
230
231	pvd->vdev_child = newchild;
232	pvd->vdev_child[id] = cvd;
233
234	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
235	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
236
237	/*
238	 * Walk up all ancestors to update guid sum.
239	 */
240	for (; pvd != NULL; pvd = pvd->vdev_parent)
241		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
242}
243
244void
245vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
246{
247	int c;
248	uint_t id = cvd->vdev_id;
249
250	ASSERT(cvd->vdev_parent == pvd);
251
252	if (pvd == NULL)
253		return;
254
255	ASSERT(id < pvd->vdev_children);
256	ASSERT(pvd->vdev_child[id] == cvd);
257
258	pvd->vdev_child[id] = NULL;
259	cvd->vdev_parent = NULL;
260
261	for (c = 0; c < pvd->vdev_children; c++)
262		if (pvd->vdev_child[c])
263			break;
264
265	if (c == pvd->vdev_children) {
266		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
267		pvd->vdev_child = NULL;
268		pvd->vdev_children = 0;
269	}
270
271	/*
272	 * Walk up all ancestors to update guid sum.
273	 */
274	for (; pvd != NULL; pvd = pvd->vdev_parent)
275		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
276}
277
278/*
279 * Remove any holes in the child array.
280 */
281void
282vdev_compact_children(vdev_t *pvd)
283{
284	vdev_t **newchild, *cvd;
285	int oldc = pvd->vdev_children;
286	int newc;
287
288	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
289
290	for (int c = newc = 0; c < oldc; c++)
291		if (pvd->vdev_child[c])
292			newc++;
293
294	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
295
296	for (int c = newc = 0; c < oldc; c++) {
297		if ((cvd = pvd->vdev_child[c]) != NULL) {
298			newchild[newc] = cvd;
299			cvd->vdev_id = newc++;
300		}
301	}
302
303	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
304	pvd->vdev_child = newchild;
305	pvd->vdev_children = newc;
306}
307
308/*
309 * Allocate and minimally initialize a vdev_t.
310 */
311vdev_t *
312vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
313{
314	vdev_t *vd;
315
316	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
317
318	if (spa->spa_root_vdev == NULL) {
319		ASSERT(ops == &vdev_root_ops);
320		spa->spa_root_vdev = vd;
321		spa->spa_load_guid = spa_generate_guid(NULL);
322	}
323
324	if (guid == 0 && ops != &vdev_hole_ops) {
325		if (spa->spa_root_vdev == vd) {
326			/*
327			 * The root vdev's guid will also be the pool guid,
328			 * which must be unique among all pools.
329			 */
330			guid = spa_generate_guid(NULL);
331		} else {
332			/*
333			 * Any other vdev's guid must be unique within the pool.
334			 */
335			guid = spa_generate_guid(spa);
336		}
337		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
338	}
339
340	vd->vdev_spa = spa;
341	vd->vdev_id = id;
342	vd->vdev_guid = guid;
343	vd->vdev_guid_sum = guid;
344	vd->vdev_ops = ops;
345	vd->vdev_state = VDEV_STATE_CLOSED;
346	vd->vdev_ishole = (ops == &vdev_hole_ops);
347
348	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
349	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
350	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
351	mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
352	for (int t = 0; t < DTL_TYPES; t++) {
353		vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
354		    &vd->vdev_dtl_lock);
355	}
356	txg_list_create(&vd->vdev_ms_list,
357	    offsetof(struct metaslab, ms_txg_node));
358	txg_list_create(&vd->vdev_dtl_list,
359	    offsetof(struct vdev, vdev_dtl_node));
360	vd->vdev_stat.vs_timestamp = gethrtime();
361	vdev_queue_init(vd);
362	vdev_cache_init(vd);
363
364	return (vd);
365}
366
367/*
368 * Allocate a new vdev.  The 'alloctype' is used to control whether we are
369 * creating a new vdev or loading an existing one - the behavior is slightly
370 * different for each case.
371 */
372int
373vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
374    int alloctype)
375{
376	vdev_ops_t *ops;
377	char *type;
378	uint64_t guid = 0, islog, nparity;
379	vdev_t *vd;
380
381	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
382
383	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
384		return (SET_ERROR(EINVAL));
385
386	if ((ops = vdev_getops(type)) == NULL)
387		return (SET_ERROR(EINVAL));
388
389	/*
390	 * If this is a load, get the vdev guid from the nvlist.
391	 * Otherwise, vdev_alloc_common() will generate one for us.
392	 */
393	if (alloctype == VDEV_ALLOC_LOAD) {
394		uint64_t label_id;
395
396		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
397		    label_id != id)
398			return (SET_ERROR(EINVAL));
399
400		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
401			return (SET_ERROR(EINVAL));
402	} else if (alloctype == VDEV_ALLOC_SPARE) {
403		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
404			return (SET_ERROR(EINVAL));
405	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
406		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
407			return (SET_ERROR(EINVAL));
408	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
409		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
410			return (SET_ERROR(EINVAL));
411	}
412
413	/*
414	 * The first allocated vdev must be of type 'root'.
415	 */
416	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
417		return (SET_ERROR(EINVAL));
418
419	/*
420	 * Determine whether we're a log vdev.
421	 */
422	islog = 0;
423	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
424	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
425		return (SET_ERROR(ENOTSUP));
426
427	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
428		return (SET_ERROR(ENOTSUP));
429
430	/*
431	 * Set the nparity property for RAID-Z vdevs.
432	 */
433	nparity = -1ULL;
434	if (ops == &vdev_raidz_ops) {
435		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
436		    &nparity) == 0) {
437			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
438				return (SET_ERROR(EINVAL));
439			/*
440			 * Previous versions could only support 1 or 2 parity
441			 * device.
442			 */
443			if (nparity > 1 &&
444			    spa_version(spa) < SPA_VERSION_RAIDZ2)
445				return (SET_ERROR(ENOTSUP));
446			if (nparity > 2 &&
447			    spa_version(spa) < SPA_VERSION_RAIDZ3)
448				return (SET_ERROR(ENOTSUP));
449		} else {
450			/*
451			 * We require the parity to be specified for SPAs that
452			 * support multiple parity levels.
453			 */
454			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
455				return (SET_ERROR(EINVAL));
456			/*
457			 * Otherwise, we default to 1 parity device for RAID-Z.
458			 */
459			nparity = 1;
460		}
461	} else {
462		nparity = 0;
463	}
464	ASSERT(nparity != -1ULL);
465
466	vd = vdev_alloc_common(spa, id, guid, ops);
467
468	vd->vdev_islog = islog;
469	vd->vdev_nparity = nparity;
470
471	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
472		vd->vdev_path = spa_strdup(vd->vdev_path);
473	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
474		vd->vdev_devid = spa_strdup(vd->vdev_devid);
475	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
476	    &vd->vdev_physpath) == 0)
477		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
478	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
479		vd->vdev_fru = spa_strdup(vd->vdev_fru);
480
481	/*
482	 * Set the whole_disk property.  If it's not specified, leave the value
483	 * as -1.
484	 */
485	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
486	    &vd->vdev_wholedisk) != 0)
487		vd->vdev_wholedisk = -1ULL;
488
489	/*
490	 * Look for the 'not present' flag.  This will only be set if the device
491	 * was not present at the time of import.
492	 */
493	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
494	    &vd->vdev_not_present);
495
496	/*
497	 * Get the alignment requirement.
498	 */
499	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
500
501	/*
502	 * Retrieve the vdev creation time.
503	 */
504	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
505	    &vd->vdev_crtxg);
506
507	/*
508	 * If we're a top-level vdev, try to load the allocation parameters.
509	 */
510	if (parent && !parent->vdev_parent &&
511	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
512		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
513		    &vd->vdev_ms_array);
514		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
515		    &vd->vdev_ms_shift);
516		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
517		    &vd->vdev_asize);
518		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
519		    &vd->vdev_removing);
520		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
521		    &vd->vdev_top_zap);
522	} else {
523		ASSERT0(vd->vdev_top_zap);
524	}
525
526	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
527		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
528		    alloctype == VDEV_ALLOC_ADD ||
529		    alloctype == VDEV_ALLOC_SPLIT ||
530		    alloctype == VDEV_ALLOC_ROOTPOOL);
531		vd->vdev_mg = metaslab_group_create(islog ?
532		    spa_log_class(spa) : spa_normal_class(spa), vd);
533	}
534
535	if (vd->vdev_ops->vdev_op_leaf &&
536	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
537		(void) nvlist_lookup_uint64(nv,
538		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
539	} else {
540		ASSERT0(vd->vdev_leaf_zap);
541	}
542
543	/*
544	 * If we're a leaf vdev, try to load the DTL object and other state.
545	 */
546
547	if (vd->vdev_ops->vdev_op_leaf &&
548	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
549	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
550		if (alloctype == VDEV_ALLOC_LOAD) {
551			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
552			    &vd->vdev_dtl_object);
553			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
554			    &vd->vdev_unspare);
555		}
556
557		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
558			uint64_t spare = 0;
559
560			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
561			    &spare) == 0 && spare)
562				spa_spare_add(vd);
563		}
564
565		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
566		    &vd->vdev_offline);
567
568		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
569		    &vd->vdev_resilver_txg);
570
571		/*
572		 * When importing a pool, we want to ignore the persistent fault
573		 * state, as the diagnosis made on another system may not be
574		 * valid in the current context.  Local vdevs will
575		 * remain in the faulted state.
576		 */
577		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
578			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
579			    &vd->vdev_faulted);
580			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
581			    &vd->vdev_degraded);
582			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
583			    &vd->vdev_removed);
584
585			if (vd->vdev_faulted || vd->vdev_degraded) {
586				char *aux;
587
588				vd->vdev_label_aux =
589				    VDEV_AUX_ERR_EXCEEDED;
590				if (nvlist_lookup_string(nv,
591				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
592				    strcmp(aux, "external") == 0)
593					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
594			}
595		}
596	}
597
598	/*
599	 * Add ourselves to the parent's list of children.
600	 */
601	vdev_add_child(parent, vd);
602
603	*vdp = vd;
604
605	return (0);
606}
607
608void
609vdev_free(vdev_t *vd)
610{
611	spa_t *spa = vd->vdev_spa;
612
613	/*
614	 * vdev_free() implies closing the vdev first.  This is simpler than
615	 * trying to ensure complicated semantics for all callers.
616	 */
617	vdev_close(vd);
618
619	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
620	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
621
622	/*
623	 * Free all children.
624	 */
625	for (int c = 0; c < vd->vdev_children; c++)
626		vdev_free(vd->vdev_child[c]);
627
628	ASSERT(vd->vdev_child == NULL);
629	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
630
631	/*
632	 * Discard allocation state.
633	 */
634	if (vd->vdev_mg != NULL) {
635		vdev_metaslab_fini(vd);
636		metaslab_group_destroy(vd->vdev_mg);
637	}
638
639	ASSERT0(vd->vdev_stat.vs_space);
640	ASSERT0(vd->vdev_stat.vs_dspace);
641	ASSERT0(vd->vdev_stat.vs_alloc);
642
643	/*
644	 * Remove this vdev from its parent's child list.
645	 */
646	vdev_remove_child(vd->vdev_parent, vd);
647
648	ASSERT(vd->vdev_parent == NULL);
649
650	/*
651	 * Clean up vdev structure.
652	 */
653	vdev_queue_fini(vd);
654	vdev_cache_fini(vd);
655
656	if (vd->vdev_path)
657		spa_strfree(vd->vdev_path);
658	if (vd->vdev_devid)
659		spa_strfree(vd->vdev_devid);
660	if (vd->vdev_physpath)
661		spa_strfree(vd->vdev_physpath);
662	if (vd->vdev_fru)
663		spa_strfree(vd->vdev_fru);
664
665	if (vd->vdev_isspare)
666		spa_spare_remove(vd);
667	if (vd->vdev_isl2cache)
668		spa_l2cache_remove(vd);
669
670	txg_list_destroy(&vd->vdev_ms_list);
671	txg_list_destroy(&vd->vdev_dtl_list);
672
673	mutex_enter(&vd->vdev_dtl_lock);
674	space_map_close(vd->vdev_dtl_sm);
675	for (int t = 0; t < DTL_TYPES; t++) {
676		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
677		range_tree_destroy(vd->vdev_dtl[t]);
678	}
679	mutex_exit(&vd->vdev_dtl_lock);
680
681	mutex_destroy(&vd->vdev_queue_lock);
682	mutex_destroy(&vd->vdev_dtl_lock);
683	mutex_destroy(&vd->vdev_stat_lock);
684	mutex_destroy(&vd->vdev_probe_lock);
685
686	if (vd == spa->spa_root_vdev)
687		spa->spa_root_vdev = NULL;
688
689	kmem_free(vd, sizeof (vdev_t));
690}
691
692/*
693 * Transfer top-level vdev state from svd to tvd.
694 */
695static void
696vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
697{
698	spa_t *spa = svd->vdev_spa;
699	metaslab_t *msp;
700	vdev_t *vd;
701	int t;
702
703	ASSERT(tvd == tvd->vdev_top);
704
705	tvd->vdev_ms_array = svd->vdev_ms_array;
706	tvd->vdev_ms_shift = svd->vdev_ms_shift;
707	tvd->vdev_ms_count = svd->vdev_ms_count;
708	tvd->vdev_top_zap = svd->vdev_top_zap;
709
710	svd->vdev_ms_array = 0;
711	svd->vdev_ms_shift = 0;
712	svd->vdev_ms_count = 0;
713	svd->vdev_top_zap = 0;
714
715	if (tvd->vdev_mg)
716		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
717	tvd->vdev_mg = svd->vdev_mg;
718	tvd->vdev_ms = svd->vdev_ms;
719
720	svd->vdev_mg = NULL;
721	svd->vdev_ms = NULL;
722
723	if (tvd->vdev_mg != NULL)
724		tvd->vdev_mg->mg_vd = tvd;
725
726	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
727	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
728	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
729
730	svd->vdev_stat.vs_alloc = 0;
731	svd->vdev_stat.vs_space = 0;
732	svd->vdev_stat.vs_dspace = 0;
733
734	for (t = 0; t < TXG_SIZE; t++) {
735		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
736			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
737		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
738			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
739		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
740			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
741	}
742
743	if (list_link_active(&svd->vdev_config_dirty_node)) {
744		vdev_config_clean(svd);
745		vdev_config_dirty(tvd);
746	}
747
748	if (list_link_active(&svd->vdev_state_dirty_node)) {
749		vdev_state_clean(svd);
750		vdev_state_dirty(tvd);
751	}
752
753	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
754	svd->vdev_deflate_ratio = 0;
755
756	tvd->vdev_islog = svd->vdev_islog;
757	svd->vdev_islog = 0;
758}
759
760static void
761vdev_top_update(vdev_t *tvd, vdev_t *vd)
762{
763	if (vd == NULL)
764		return;
765
766	vd->vdev_top = tvd;
767
768	for (int c = 0; c < vd->vdev_children; c++)
769		vdev_top_update(tvd, vd->vdev_child[c]);
770}
771
772/*
773 * Add a mirror/replacing vdev above an existing vdev.
774 */
775vdev_t *
776vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
777{
778	spa_t *spa = cvd->vdev_spa;
779	vdev_t *pvd = cvd->vdev_parent;
780	vdev_t *mvd;
781
782	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
783
784	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
785
786	mvd->vdev_asize = cvd->vdev_asize;
787	mvd->vdev_min_asize = cvd->vdev_min_asize;
788	mvd->vdev_max_asize = cvd->vdev_max_asize;
789	mvd->vdev_ashift = cvd->vdev_ashift;
790	mvd->vdev_state = cvd->vdev_state;
791	mvd->vdev_crtxg = cvd->vdev_crtxg;
792
793	vdev_remove_child(pvd, cvd);
794	vdev_add_child(pvd, mvd);
795	cvd->vdev_id = mvd->vdev_children;
796	vdev_add_child(mvd, cvd);
797	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
798
799	if (mvd == mvd->vdev_top)
800		vdev_top_transfer(cvd, mvd);
801
802	return (mvd);
803}
804
805/*
806 * Remove a 1-way mirror/replacing vdev from the tree.
807 */
808void
809vdev_remove_parent(vdev_t *cvd)
810{
811	vdev_t *mvd = cvd->vdev_parent;
812	vdev_t *pvd = mvd->vdev_parent;
813
814	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
815
816	ASSERT(mvd->vdev_children == 1);
817	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
818	    mvd->vdev_ops == &vdev_replacing_ops ||
819	    mvd->vdev_ops == &vdev_spare_ops);
820	cvd->vdev_ashift = mvd->vdev_ashift;
821
822	vdev_remove_child(mvd, cvd);
823	vdev_remove_child(pvd, mvd);
824
825	/*
826	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
827	 * Otherwise, we could have detached an offline device, and when we
828	 * go to import the pool we'll think we have two top-level vdevs,
829	 * instead of a different version of the same top-level vdev.
830	 */
831	if (mvd->vdev_top == mvd) {
832		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
833		cvd->vdev_orig_guid = cvd->vdev_guid;
834		cvd->vdev_guid += guid_delta;
835		cvd->vdev_guid_sum += guid_delta;
836	}
837	cvd->vdev_id = mvd->vdev_id;
838	vdev_add_child(pvd, cvd);
839	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
840
841	if (cvd == cvd->vdev_top)
842		vdev_top_transfer(mvd, cvd);
843
844	ASSERT(mvd->vdev_children == 0);
845	vdev_free(mvd);
846}
847
848int
849vdev_metaslab_init(vdev_t *vd, uint64_t txg)
850{
851	spa_t *spa = vd->vdev_spa;
852	objset_t *mos = spa->spa_meta_objset;
853	uint64_t m;
854	uint64_t oldc = vd->vdev_ms_count;
855	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
856	metaslab_t **mspp;
857	int error;
858
859	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
860
861	/*
862	 * This vdev is not being allocated from yet or is a hole.
863	 */
864	if (vd->vdev_ms_shift == 0)
865		return (0);
866
867	ASSERT(!vd->vdev_ishole);
868
869	/*
870	 * Compute the raidz-deflation ratio.  Note, we hard-code
871	 * in 128k (1 << 17) because it is the "typical" blocksize.
872	 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
873	 * otherwise it would inconsistently account for existing bp's.
874	 */
875	vd->vdev_deflate_ratio = (1 << 17) /
876	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
877
878	ASSERT(oldc <= newc);
879
880	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
881
882	if (oldc != 0) {
883		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
884		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
885	}
886
887	vd->vdev_ms = mspp;
888	vd->vdev_ms_count = newc;
889
890	for (m = oldc; m < newc; m++) {
891		uint64_t object = 0;
892
893		if (txg == 0) {
894			error = dmu_read(mos, vd->vdev_ms_array,
895			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
896			    DMU_READ_PREFETCH);
897			if (error)
898				return (error);
899		}
900
901		error = metaslab_init(vd->vdev_mg, m, object, txg,
902		    &(vd->vdev_ms[m]));
903		if (error)
904			return (error);
905	}
906
907	if (txg == 0)
908		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
909
910	/*
911	 * If the vdev is being removed we don't activate
912	 * the metaslabs since we want to ensure that no new
913	 * allocations are performed on this device.
914	 */
915	if (oldc == 0 && !vd->vdev_removing)
916		metaslab_group_activate(vd->vdev_mg);
917
918	if (txg == 0)
919		spa_config_exit(spa, SCL_ALLOC, FTAG);
920
921	return (0);
922}
923
924void
925vdev_metaslab_fini(vdev_t *vd)
926{
927	uint64_t m;
928	uint64_t count = vd->vdev_ms_count;
929
930	if (vd->vdev_ms != NULL) {
931		metaslab_group_passivate(vd->vdev_mg);
932		for (m = 0; m < count; m++) {
933			metaslab_t *msp = vd->vdev_ms[m];
934
935			if (msp != NULL)
936				metaslab_fini(msp);
937		}
938		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
939		vd->vdev_ms = NULL;
940	}
941}
942
943typedef struct vdev_probe_stats {
944	boolean_t	vps_readable;
945	boolean_t	vps_writeable;
946	int		vps_flags;
947} vdev_probe_stats_t;
948
949static void
950vdev_probe_done(zio_t *zio)
951{
952	spa_t *spa = zio->io_spa;
953	vdev_t *vd = zio->io_vd;
954	vdev_probe_stats_t *vps = zio->io_private;
955
956	ASSERT(vd->vdev_probe_zio != NULL);
957
958	if (zio->io_type == ZIO_TYPE_READ) {
959		if (zio->io_error == 0)
960			vps->vps_readable = 1;
961		if (zio->io_error == 0 && spa_writeable(spa)) {
962			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
963			    zio->io_offset, zio->io_size, zio->io_data,
964			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
965			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
966		} else {
967			zio_buf_free(zio->io_data, zio->io_size);
968		}
969	} else if (zio->io_type == ZIO_TYPE_WRITE) {
970		if (zio->io_error == 0)
971			vps->vps_writeable = 1;
972		zio_buf_free(zio->io_data, zio->io_size);
973	} else if (zio->io_type == ZIO_TYPE_NULL) {
974		zio_t *pio;
975
976		vd->vdev_cant_read |= !vps->vps_readable;
977		vd->vdev_cant_write |= !vps->vps_writeable;
978
979		if (vdev_readable(vd) &&
980		    (vdev_writeable(vd) || !spa_writeable(spa))) {
981			zio->io_error = 0;
982		} else {
983			ASSERT(zio->io_error != 0);
984			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
985			    spa, vd, NULL, 0, 0);
986			zio->io_error = SET_ERROR(ENXIO);
987		}
988
989		mutex_enter(&vd->vdev_probe_lock);
990		ASSERT(vd->vdev_probe_zio == zio);
991		vd->vdev_probe_zio = NULL;
992		mutex_exit(&vd->vdev_probe_lock);
993
994		zio_link_t *zl = NULL;
995		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
996			if (!vdev_accessible(vd, pio))
997				pio->io_error = SET_ERROR(ENXIO);
998
999		kmem_free(vps, sizeof (*vps));
1000	}
1001}
1002
1003/*
1004 * Determine whether this device is accessible.
1005 *
1006 * Read and write to several known locations: the pad regions of each
1007 * vdev label but the first, which we leave alone in case it contains
1008 * a VTOC.
1009 */
1010zio_t *
1011vdev_probe(vdev_t *vd, zio_t *zio)
1012{
1013	spa_t *spa = vd->vdev_spa;
1014	vdev_probe_stats_t *vps = NULL;
1015	zio_t *pio;
1016
1017	ASSERT(vd->vdev_ops->vdev_op_leaf);
1018
1019	/*
1020	 * Don't probe the probe.
1021	 */
1022	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1023		return (NULL);
1024
1025	/*
1026	 * To prevent 'probe storms' when a device fails, we create
1027	 * just one probe i/o at a time.  All zios that want to probe
1028	 * this vdev will become parents of the probe io.
1029	 */
1030	mutex_enter(&vd->vdev_probe_lock);
1031
1032	if ((pio = vd->vdev_probe_zio) == NULL) {
1033		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1034
1035		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1036		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1037		    ZIO_FLAG_TRYHARD;
1038
1039		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1040			/*
1041			 * vdev_cant_read and vdev_cant_write can only
1042			 * transition from TRUE to FALSE when we have the
1043			 * SCL_ZIO lock as writer; otherwise they can only
1044			 * transition from FALSE to TRUE.  This ensures that
1045			 * any zio looking at these values can assume that
1046			 * failures persist for the life of the I/O.  That's
1047			 * important because when a device has intermittent
1048			 * connectivity problems, we want to ensure that
1049			 * they're ascribed to the device (ENXIO) and not
1050			 * the zio (EIO).
1051			 *
1052			 * Since we hold SCL_ZIO as writer here, clear both
1053			 * values so the probe can reevaluate from first
1054			 * principles.
1055			 */
1056			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1057			vd->vdev_cant_read = B_FALSE;
1058			vd->vdev_cant_write = B_FALSE;
1059		}
1060
1061		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1062		    vdev_probe_done, vps,
1063		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1064
1065		/*
1066		 * We can't change the vdev state in this context, so we
1067		 * kick off an async task to do it on our behalf.
1068		 */
1069		if (zio != NULL) {
1070			vd->vdev_probe_wanted = B_TRUE;
1071			spa_async_request(spa, SPA_ASYNC_PROBE);
1072		}
1073	}
1074
1075	if (zio != NULL)
1076		zio_add_child(zio, pio);
1077
1078	mutex_exit(&vd->vdev_probe_lock);
1079
1080	if (vps == NULL) {
1081		ASSERT(zio != NULL);
1082		return (NULL);
1083	}
1084
1085	for (int l = 1; l < VDEV_LABELS; l++) {
1086		zio_nowait(zio_read_phys(pio, vd,
1087		    vdev_label_offset(vd->vdev_psize, l,
1088		    offsetof(vdev_label_t, vl_pad2)),
1089		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1090		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1091		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1092	}
1093
1094	if (zio == NULL)
1095		return (pio);
1096
1097	zio_nowait(pio);
1098	return (NULL);
1099}
1100
1101static void
1102vdev_open_child(void *arg)
1103{
1104	vdev_t *vd = arg;
1105
1106	vd->vdev_open_thread = curthread;
1107	vd->vdev_open_error = vdev_open(vd);
1108	vd->vdev_open_thread = NULL;
1109}
1110
1111boolean_t
1112vdev_uses_zvols(vdev_t *vd)
1113{
1114	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1115	    strlen(ZVOL_DIR)) == 0)
1116		return (B_TRUE);
1117	for (int c = 0; c < vd->vdev_children; c++)
1118		if (vdev_uses_zvols(vd->vdev_child[c]))
1119			return (B_TRUE);
1120	return (B_FALSE);
1121}
1122
1123void
1124vdev_open_children(vdev_t *vd)
1125{
1126	taskq_t *tq;
1127	int children = vd->vdev_children;
1128
1129	/*
1130	 * in order to handle pools on top of zvols, do the opens
1131	 * in a single thread so that the same thread holds the
1132	 * spa_namespace_lock
1133	 */
1134	if (vdev_uses_zvols(vd)) {
1135		for (int c = 0; c < children; c++)
1136			vd->vdev_child[c]->vdev_open_error =
1137			    vdev_open(vd->vdev_child[c]);
1138		return;
1139	}
1140	tq = taskq_create("vdev_open", children, minclsyspri,
1141	    children, children, TASKQ_PREPOPULATE);
1142
1143	for (int c = 0; c < children; c++)
1144		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1145		    TQ_SLEEP) != NULL);
1146
1147	taskq_destroy(tq);
1148}
1149
1150/*
1151 * Prepare a virtual device for access.
1152 */
1153int
1154vdev_open(vdev_t *vd)
1155{
1156	spa_t *spa = vd->vdev_spa;
1157	int error;
1158	uint64_t osize = 0;
1159	uint64_t max_osize = 0;
1160	uint64_t asize, max_asize, psize;
1161	uint64_t ashift = 0;
1162
1163	ASSERT(vd->vdev_open_thread == curthread ||
1164	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1165	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1166	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1167	    vd->vdev_state == VDEV_STATE_OFFLINE);
1168
1169	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1170	vd->vdev_cant_read = B_FALSE;
1171	vd->vdev_cant_write = B_FALSE;
1172	vd->vdev_min_asize = vdev_get_min_asize(vd);
1173
1174	/*
1175	 * If this vdev is not removed, check its fault status.  If it's
1176	 * faulted, bail out of the open.
1177	 */
1178	if (!vd->vdev_removed && vd->vdev_faulted) {
1179		ASSERT(vd->vdev_children == 0);
1180		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1181		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1182		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1183		    vd->vdev_label_aux);
1184		return (SET_ERROR(ENXIO));
1185	} else if (vd->vdev_offline) {
1186		ASSERT(vd->vdev_children == 0);
1187		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1188		return (SET_ERROR(ENXIO));
1189	}
1190
1191	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1192
1193	/*
1194	 * Reset the vdev_reopening flag so that we actually close
1195	 * the vdev on error.
1196	 */
1197	vd->vdev_reopening = B_FALSE;
1198	if (zio_injection_enabled && error == 0)
1199		error = zio_handle_device_injection(vd, NULL, ENXIO);
1200
1201	if (error) {
1202		if (vd->vdev_removed &&
1203		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1204			vd->vdev_removed = B_FALSE;
1205
1206		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1207		    vd->vdev_stat.vs_aux);
1208		return (error);
1209	}
1210
1211	vd->vdev_removed = B_FALSE;
1212
1213	/*
1214	 * Recheck the faulted flag now that we have confirmed that
1215	 * the vdev is accessible.  If we're faulted, bail.
1216	 */
1217	if (vd->vdev_faulted) {
1218		ASSERT(vd->vdev_children == 0);
1219		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1220		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1221		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1222		    vd->vdev_label_aux);
1223		return (SET_ERROR(ENXIO));
1224	}
1225
1226	if (vd->vdev_degraded) {
1227		ASSERT(vd->vdev_children == 0);
1228		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1229		    VDEV_AUX_ERR_EXCEEDED);
1230	} else {
1231		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1232	}
1233
1234	/*
1235	 * For hole or missing vdevs we just return success.
1236	 */
1237	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1238		return (0);
1239
1240	for (int c = 0; c < vd->vdev_children; c++) {
1241		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1242			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1243			    VDEV_AUX_NONE);
1244			break;
1245		}
1246	}
1247
1248	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1249	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1250
1251	if (vd->vdev_children == 0) {
1252		if (osize < SPA_MINDEVSIZE) {
1253			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1254			    VDEV_AUX_TOO_SMALL);
1255			return (SET_ERROR(EOVERFLOW));
1256		}
1257		psize = osize;
1258		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1259		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1260		    VDEV_LABEL_END_SIZE);
1261	} else {
1262		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1263		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1264			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1265			    VDEV_AUX_TOO_SMALL);
1266			return (SET_ERROR(EOVERFLOW));
1267		}
1268		psize = 0;
1269		asize = osize;
1270		max_asize = max_osize;
1271	}
1272
1273	vd->vdev_psize = psize;
1274
1275	/*
1276	 * Make sure the allocatable size hasn't shrunk.
1277	 */
1278	if (asize < vd->vdev_min_asize) {
1279		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1280		    VDEV_AUX_BAD_LABEL);
1281		return (SET_ERROR(EINVAL));
1282	}
1283
1284	if (vd->vdev_asize == 0) {
1285		/*
1286		 * This is the first-ever open, so use the computed values.
1287		 * For testing purposes, a higher ashift can be requested.
1288		 */
1289		vd->vdev_asize = asize;
1290		vd->vdev_max_asize = max_asize;
1291		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1292	} else {
1293		/*
1294		 * Detect if the alignment requirement has increased.
1295		 * We don't want to make the pool unavailable, just
1296		 * issue a warning instead.
1297		 */
1298		if (ashift > vd->vdev_top->vdev_ashift &&
1299		    vd->vdev_ops->vdev_op_leaf) {
1300			cmn_err(CE_WARN,
1301			    "Disk, '%s', has a block alignment that is "
1302			    "larger than the pool's alignment\n",
1303			    vd->vdev_path);
1304		}
1305		vd->vdev_max_asize = max_asize;
1306	}
1307
1308	/*
1309	 * If all children are healthy and the asize has increased,
1310	 * then we've experienced dynamic LUN growth.  If automatic
1311	 * expansion is enabled then use the additional space.
1312	 */
1313	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1314	    (vd->vdev_expanding || spa->spa_autoexpand))
1315		vd->vdev_asize = asize;
1316
1317	vdev_set_min_asize(vd);
1318
1319	/*
1320	 * Ensure we can issue some IO before declaring the
1321	 * vdev open for business.
1322	 */
1323	if (vd->vdev_ops->vdev_op_leaf &&
1324	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1325		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1326		    VDEV_AUX_ERR_EXCEEDED);
1327		return (error);
1328	}
1329
1330	/*
1331	 * Track the min and max ashift values for normal data devices.
1332	 */
1333	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1334	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1335		if (vd->vdev_ashift > spa->spa_max_ashift)
1336			spa->spa_max_ashift = vd->vdev_ashift;
1337		if (vd->vdev_ashift < spa->spa_min_ashift)
1338			spa->spa_min_ashift = vd->vdev_ashift;
1339	}
1340
1341	/*
1342	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1343	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1344	 * since this would just restart the scrub we are already doing.
1345	 */
1346	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1347	    vdev_resilver_needed(vd, NULL, NULL))
1348		spa_async_request(spa, SPA_ASYNC_RESILVER);
1349
1350	return (0);
1351}
1352
1353/*
1354 * Called once the vdevs are all opened, this routine validates the label
1355 * contents.  This needs to be done before vdev_load() so that we don't
1356 * inadvertently do repair I/Os to the wrong device.
1357 *
1358 * If 'strict' is false ignore the spa guid check. This is necessary because
1359 * if the machine crashed during a re-guid the new guid might have been written
1360 * to all of the vdev labels, but not the cached config. The strict check
1361 * will be performed when the pool is opened again using the mos config.
1362 *
1363 * This function will only return failure if one of the vdevs indicates that it
1364 * has since been destroyed or exported.  This is only possible if
1365 * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1366 * will be updated but the function will return 0.
1367 */
1368int
1369vdev_validate(vdev_t *vd, boolean_t strict)
1370{
1371	spa_t *spa = vd->vdev_spa;
1372	nvlist_t *label;
1373	uint64_t guid = 0, top_guid;
1374	uint64_t state;
1375
1376	for (int c = 0; c < vd->vdev_children; c++)
1377		if (vdev_validate(vd->vdev_child[c], strict) != 0)
1378			return (SET_ERROR(EBADF));
1379
1380	/*
1381	 * If the device has already failed, or was marked offline, don't do
1382	 * any further validation.  Otherwise, label I/O will fail and we will
1383	 * overwrite the previous state.
1384	 */
1385	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1386		uint64_t aux_guid = 0;
1387		nvlist_t *nvl;
1388		uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1389		    spa_last_synced_txg(spa) : -1ULL;
1390
1391		if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1392			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1393			    VDEV_AUX_BAD_LABEL);
1394			return (0);
1395		}
1396
1397		/*
1398		 * Determine if this vdev has been split off into another
1399		 * pool.  If so, then refuse to open it.
1400		 */
1401		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1402		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1403			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1404			    VDEV_AUX_SPLIT_POOL);
1405			nvlist_free(label);
1406			return (0);
1407		}
1408
1409		if (strict && (nvlist_lookup_uint64(label,
1410		    ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1411		    guid != spa_guid(spa))) {
1412			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1413			    VDEV_AUX_CORRUPT_DATA);
1414			nvlist_free(label);
1415			return (0);
1416		}
1417
1418		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1419		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1420		    &aux_guid) != 0)
1421			aux_guid = 0;
1422
1423		/*
1424		 * If this vdev just became a top-level vdev because its
1425		 * sibling was detached, it will have adopted the parent's
1426		 * vdev guid -- but the label may or may not be on disk yet.
1427		 * Fortunately, either version of the label will have the
1428		 * same top guid, so if we're a top-level vdev, we can
1429		 * safely compare to that instead.
1430		 *
1431		 * If we split this vdev off instead, then we also check the
1432		 * original pool's guid.  We don't want to consider the vdev
1433		 * corrupt if it is partway through a split operation.
1434		 */
1435		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1436		    &guid) != 0 ||
1437		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1438		    &top_guid) != 0 ||
1439		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1440		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1441			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1442			    VDEV_AUX_CORRUPT_DATA);
1443			nvlist_free(label);
1444			return (0);
1445		}
1446
1447		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1448		    &state) != 0) {
1449			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1450			    VDEV_AUX_CORRUPT_DATA);
1451			nvlist_free(label);
1452			return (0);
1453		}
1454
1455		nvlist_free(label);
1456
1457		/*
1458		 * If this is a verbatim import, no need to check the
1459		 * state of the pool.
1460		 */
1461		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1462		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1463		    state != POOL_STATE_ACTIVE)
1464			return (SET_ERROR(EBADF));
1465
1466		/*
1467		 * If we were able to open and validate a vdev that was
1468		 * previously marked permanently unavailable, clear that state
1469		 * now.
1470		 */
1471		if (vd->vdev_not_present)
1472			vd->vdev_not_present = 0;
1473	}
1474
1475	return (0);
1476}
1477
1478/*
1479 * Close a virtual device.
1480 */
1481void
1482vdev_close(vdev_t *vd)
1483{
1484	spa_t *spa = vd->vdev_spa;
1485	vdev_t *pvd = vd->vdev_parent;
1486
1487	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1488
1489	/*
1490	 * If our parent is reopening, then we are as well, unless we are
1491	 * going offline.
1492	 */
1493	if (pvd != NULL && pvd->vdev_reopening)
1494		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1495
1496	vd->vdev_ops->vdev_op_close(vd);
1497
1498	vdev_cache_purge(vd);
1499
1500	/*
1501	 * We record the previous state before we close it, so that if we are
1502	 * doing a reopen(), we don't generate FMA ereports if we notice that
1503	 * it's still faulted.
1504	 */
1505	vd->vdev_prevstate = vd->vdev_state;
1506
1507	if (vd->vdev_offline)
1508		vd->vdev_state = VDEV_STATE_OFFLINE;
1509	else
1510		vd->vdev_state = VDEV_STATE_CLOSED;
1511	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1512}
1513
1514void
1515vdev_hold(vdev_t *vd)
1516{
1517	spa_t *spa = vd->vdev_spa;
1518
1519	ASSERT(spa_is_root(spa));
1520	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1521		return;
1522
1523	for (int c = 0; c < vd->vdev_children; c++)
1524		vdev_hold(vd->vdev_child[c]);
1525
1526	if (vd->vdev_ops->vdev_op_leaf)
1527		vd->vdev_ops->vdev_op_hold(vd);
1528}
1529
1530void
1531vdev_rele(vdev_t *vd)
1532{
1533	spa_t *spa = vd->vdev_spa;
1534
1535	ASSERT(spa_is_root(spa));
1536	for (int c = 0; c < vd->vdev_children; c++)
1537		vdev_rele(vd->vdev_child[c]);
1538
1539	if (vd->vdev_ops->vdev_op_leaf)
1540		vd->vdev_ops->vdev_op_rele(vd);
1541}
1542
1543/*
1544 * Reopen all interior vdevs and any unopened leaves.  We don't actually
1545 * reopen leaf vdevs which had previously been opened as they might deadlock
1546 * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1547 * If the leaf has never been opened then open it, as usual.
1548 */
1549void
1550vdev_reopen(vdev_t *vd)
1551{
1552	spa_t *spa = vd->vdev_spa;
1553
1554	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1555
1556	/* set the reopening flag unless we're taking the vdev offline */
1557	vd->vdev_reopening = !vd->vdev_offline;
1558	vdev_close(vd);
1559	(void) vdev_open(vd);
1560
1561	/*
1562	 * Call vdev_validate() here to make sure we have the same device.
1563	 * Otherwise, a device with an invalid label could be successfully
1564	 * opened in response to vdev_reopen().
1565	 */
1566	if (vd->vdev_aux) {
1567		(void) vdev_validate_aux(vd);
1568		if (vdev_readable(vd) && vdev_writeable(vd) &&
1569		    vd->vdev_aux == &spa->spa_l2cache &&
1570		    !l2arc_vdev_present(vd))
1571			l2arc_add_vdev(spa, vd);
1572	} else {
1573		(void) vdev_validate(vd, B_TRUE);
1574	}
1575
1576	/*
1577	 * Reassess parent vdev's health.
1578	 */
1579	vdev_propagate_state(vd);
1580}
1581
1582int
1583vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1584{
1585	int error;
1586
1587	/*
1588	 * Normally, partial opens (e.g. of a mirror) are allowed.
1589	 * For a create, however, we want to fail the request if
1590	 * there are any components we can't open.
1591	 */
1592	error = vdev_open(vd);
1593
1594	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1595		vdev_close(vd);
1596		return (error ? error : ENXIO);
1597	}
1598
1599	/*
1600	 * Recursively load DTLs and initialize all labels.
1601	 */
1602	if ((error = vdev_dtl_load(vd)) != 0 ||
1603	    (error = vdev_label_init(vd, txg, isreplacing ?
1604	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1605		vdev_close(vd);
1606		return (error);
1607	}
1608
1609	return (0);
1610}
1611
1612void
1613vdev_metaslab_set_size(vdev_t *vd)
1614{
1615	/*
1616	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1617	 */
1618	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1619	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1620}
1621
1622void
1623vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1624{
1625	ASSERT(vd == vd->vdev_top);
1626	ASSERT(!vd->vdev_ishole);
1627	ASSERT(ISP2(flags));
1628	ASSERT(spa_writeable(vd->vdev_spa));
1629
1630	if (flags & VDD_METASLAB)
1631		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1632
1633	if (flags & VDD_DTL)
1634		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1635
1636	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1637}
1638
1639void
1640vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1641{
1642	for (int c = 0; c < vd->vdev_children; c++)
1643		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1644
1645	if (vd->vdev_ops->vdev_op_leaf)
1646		vdev_dirty(vd->vdev_top, flags, vd, txg);
1647}
1648
1649/*
1650 * DTLs.
1651 *
1652 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1653 * the vdev has less than perfect replication.  There are four kinds of DTL:
1654 *
1655 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1656 *
1657 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1658 *
1659 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1660 *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1661 *	txgs that was scrubbed.
1662 *
1663 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1664 *	persistent errors or just some device being offline.
1665 *	Unlike the other three, the DTL_OUTAGE map is not generally
1666 *	maintained; it's only computed when needed, typically to
1667 *	determine whether a device can be detached.
1668 *
1669 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1670 * either has the data or it doesn't.
1671 *
1672 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1673 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1674 * if any child is less than fully replicated, then so is its parent.
1675 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1676 * comprising only those txgs which appear in 'maxfaults' or more children;
1677 * those are the txgs we don't have enough replication to read.  For example,
1678 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1679 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1680 * two child DTL_MISSING maps.
1681 *
1682 * It should be clear from the above that to compute the DTLs and outage maps
1683 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1684 * Therefore, that is all we keep on disk.  When loading the pool, or after
1685 * a configuration change, we generate all other DTLs from first principles.
1686 */
1687void
1688vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1689{
1690	range_tree_t *rt = vd->vdev_dtl[t];
1691
1692	ASSERT(t < DTL_TYPES);
1693	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1694	ASSERT(spa_writeable(vd->vdev_spa));
1695
1696	mutex_enter(rt->rt_lock);
1697	if (!range_tree_contains(rt, txg, size))
1698		range_tree_add(rt, txg, size);
1699	mutex_exit(rt->rt_lock);
1700}
1701
1702boolean_t
1703vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1704{
1705	range_tree_t *rt = vd->vdev_dtl[t];
1706	boolean_t dirty = B_FALSE;
1707
1708	ASSERT(t < DTL_TYPES);
1709	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1710
1711	mutex_enter(rt->rt_lock);
1712	if (range_tree_space(rt) != 0)
1713		dirty = range_tree_contains(rt, txg, size);
1714	mutex_exit(rt->rt_lock);
1715
1716	return (dirty);
1717}
1718
1719boolean_t
1720vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1721{
1722	range_tree_t *rt = vd->vdev_dtl[t];
1723	boolean_t empty;
1724
1725	mutex_enter(rt->rt_lock);
1726	empty = (range_tree_space(rt) == 0);
1727	mutex_exit(rt->rt_lock);
1728
1729	return (empty);
1730}
1731
1732/*
1733 * Returns the lowest txg in the DTL range.
1734 */
1735static uint64_t
1736vdev_dtl_min(vdev_t *vd)
1737{
1738	range_seg_t *rs;
1739
1740	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1741	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1742	ASSERT0(vd->vdev_children);
1743
1744	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1745	return (rs->rs_start - 1);
1746}
1747
1748/*
1749 * Returns the highest txg in the DTL.
1750 */
1751static uint64_t
1752vdev_dtl_max(vdev_t *vd)
1753{
1754	range_seg_t *rs;
1755
1756	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1757	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1758	ASSERT0(vd->vdev_children);
1759
1760	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1761	return (rs->rs_end);
1762}
1763
1764/*
1765 * Determine if a resilvering vdev should remove any DTL entries from
1766 * its range. If the vdev was resilvering for the entire duration of the
1767 * scan then it should excise that range from its DTLs. Otherwise, this
1768 * vdev is considered partially resilvered and should leave its DTL
1769 * entries intact. The comment in vdev_dtl_reassess() describes how we
1770 * excise the DTLs.
1771 */
1772static boolean_t
1773vdev_dtl_should_excise(vdev_t *vd)
1774{
1775	spa_t *spa = vd->vdev_spa;
1776	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1777
1778	ASSERT0(scn->scn_phys.scn_errors);
1779	ASSERT0(vd->vdev_children);
1780
1781	if (vd->vdev_resilver_txg == 0 ||
1782	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1783		return (B_TRUE);
1784
1785	/*
1786	 * When a resilver is initiated the scan will assign the scn_max_txg
1787	 * value to the highest txg value that exists in all DTLs. If this
1788	 * device's max DTL is not part of this scan (i.e. it is not in
1789	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1790	 * for excision.
1791	 */
1792	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1793		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1794		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1795		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1796		return (B_TRUE);
1797	}
1798	return (B_FALSE);
1799}
1800
1801/*
1802 * Reassess DTLs after a config change or scrub completion.
1803 */
1804void
1805vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1806{
1807	spa_t *spa = vd->vdev_spa;
1808	avl_tree_t reftree;
1809	int minref;
1810
1811	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1812
1813	for (int c = 0; c < vd->vdev_children; c++)
1814		vdev_dtl_reassess(vd->vdev_child[c], txg,
1815		    scrub_txg, scrub_done);
1816
1817	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1818		return;
1819
1820	if (vd->vdev_ops->vdev_op_leaf) {
1821		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1822
1823		mutex_enter(&vd->vdev_dtl_lock);
1824
1825		/*
1826		 * If we've completed a scan cleanly then determine
1827		 * if this vdev should remove any DTLs. We only want to
1828		 * excise regions on vdevs that were available during
1829		 * the entire duration of this scan.
1830		 */
1831		if (scrub_txg != 0 &&
1832		    (spa->spa_scrub_started ||
1833		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1834		    vdev_dtl_should_excise(vd)) {
1835			/*
1836			 * We completed a scrub up to scrub_txg.  If we
1837			 * did it without rebooting, then the scrub dtl
1838			 * will be valid, so excise the old region and
1839			 * fold in the scrub dtl.  Otherwise, leave the
1840			 * dtl as-is if there was an error.
1841			 *
1842			 * There's little trick here: to excise the beginning
1843			 * of the DTL_MISSING map, we put it into a reference
1844			 * tree and then add a segment with refcnt -1 that
1845			 * covers the range [0, scrub_txg).  This means
1846			 * that each txg in that range has refcnt -1 or 0.
1847			 * We then add DTL_SCRUB with a refcnt of 2, so that
1848			 * entries in the range [0, scrub_txg) will have a
1849			 * positive refcnt -- either 1 or 2.  We then convert
1850			 * the reference tree into the new DTL_MISSING map.
1851			 */
1852			space_reftree_create(&reftree);
1853			space_reftree_add_map(&reftree,
1854			    vd->vdev_dtl[DTL_MISSING], 1);
1855			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1856			space_reftree_add_map(&reftree,
1857			    vd->vdev_dtl[DTL_SCRUB], 2);
1858			space_reftree_generate_map(&reftree,
1859			    vd->vdev_dtl[DTL_MISSING], 1);
1860			space_reftree_destroy(&reftree);
1861		}
1862		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1863		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1864		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1865		if (scrub_done)
1866			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1867		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1868		if (!vdev_readable(vd))
1869			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1870		else
1871			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1872			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1873
1874		/*
1875		 * If the vdev was resilvering and no longer has any
1876		 * DTLs then reset its resilvering flag.
1877		 */
1878		if (vd->vdev_resilver_txg != 0 &&
1879		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1880		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1881			vd->vdev_resilver_txg = 0;
1882
1883		mutex_exit(&vd->vdev_dtl_lock);
1884
1885		if (txg != 0)
1886			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1887		return;
1888	}
1889
1890	mutex_enter(&vd->vdev_dtl_lock);
1891	for (int t = 0; t < DTL_TYPES; t++) {
1892		/* account for child's outage in parent's missing map */
1893		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1894		if (t == DTL_SCRUB)
1895			continue;			/* leaf vdevs only */
1896		if (t == DTL_PARTIAL)
1897			minref = 1;			/* i.e. non-zero */
1898		else if (vd->vdev_nparity != 0)
1899			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1900		else
1901			minref = vd->vdev_children;	/* any kind of mirror */
1902		space_reftree_create(&reftree);
1903		for (int c = 0; c < vd->vdev_children; c++) {
1904			vdev_t *cvd = vd->vdev_child[c];
1905			mutex_enter(&cvd->vdev_dtl_lock);
1906			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1907			mutex_exit(&cvd->vdev_dtl_lock);
1908		}
1909		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1910		space_reftree_destroy(&reftree);
1911	}
1912	mutex_exit(&vd->vdev_dtl_lock);
1913}
1914
1915int
1916vdev_dtl_load(vdev_t *vd)
1917{
1918	spa_t *spa = vd->vdev_spa;
1919	objset_t *mos = spa->spa_meta_objset;
1920	int error = 0;
1921
1922	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1923		ASSERT(!vd->vdev_ishole);
1924
1925		error = space_map_open(&vd->vdev_dtl_sm, mos,
1926		    vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1927		if (error)
1928			return (error);
1929		ASSERT(vd->vdev_dtl_sm != NULL);
1930
1931		mutex_enter(&vd->vdev_dtl_lock);
1932
1933		/*
1934		 * Now that we've opened the space_map we need to update
1935		 * the in-core DTL.
1936		 */
1937		space_map_update(vd->vdev_dtl_sm);
1938
1939		error = space_map_load(vd->vdev_dtl_sm,
1940		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1941		mutex_exit(&vd->vdev_dtl_lock);
1942
1943		return (error);
1944	}
1945
1946	for (int c = 0; c < vd->vdev_children; c++) {
1947		error = vdev_dtl_load(vd->vdev_child[c]);
1948		if (error != 0)
1949			break;
1950	}
1951
1952	return (error);
1953}
1954
1955void
1956vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
1957{
1958	spa_t *spa = vd->vdev_spa;
1959
1960	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
1961	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1962	    zapobj, tx));
1963}
1964
1965uint64_t
1966vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
1967{
1968	spa_t *spa = vd->vdev_spa;
1969	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
1970	    DMU_OT_NONE, 0, tx);
1971
1972	ASSERT(zap != 0);
1973	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1974	    zap, tx));
1975
1976	return (zap);
1977}
1978
1979void
1980vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
1981{
1982	if (vd->vdev_ops != &vdev_hole_ops &&
1983	    vd->vdev_ops != &vdev_missing_ops &&
1984	    vd->vdev_ops != &vdev_root_ops &&
1985	    !vd->vdev_top->vdev_removing) {
1986		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
1987			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
1988		}
1989		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
1990			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
1991		}
1992	}
1993	for (uint64_t i = 0; i < vd->vdev_children; i++) {
1994		vdev_construct_zaps(vd->vdev_child[i], tx);
1995	}
1996}
1997
1998void
1999vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2000{
2001	spa_t *spa = vd->vdev_spa;
2002	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2003	objset_t *mos = spa->spa_meta_objset;
2004	range_tree_t *rtsync;
2005	kmutex_t rtlock;
2006	dmu_tx_t *tx;
2007	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2008
2009	ASSERT(!vd->vdev_ishole);
2010	ASSERT(vd->vdev_ops->vdev_op_leaf);
2011
2012	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2013
2014	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2015		mutex_enter(&vd->vdev_dtl_lock);
2016		space_map_free(vd->vdev_dtl_sm, tx);
2017		space_map_close(vd->vdev_dtl_sm);
2018		vd->vdev_dtl_sm = NULL;
2019		mutex_exit(&vd->vdev_dtl_lock);
2020
2021		/*
2022		 * We only destroy the leaf ZAP for detached leaves or for
2023		 * removed log devices. Removed data devices handle leaf ZAP
2024		 * cleanup later, once cancellation is no longer possible.
2025		 */
2026		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2027		    vd->vdev_top->vdev_islog)) {
2028			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2029			vd->vdev_leaf_zap = 0;
2030		}
2031
2032		dmu_tx_commit(tx);
2033		return;
2034	}
2035
2036	if (vd->vdev_dtl_sm == NULL) {
2037		uint64_t new_object;
2038
2039		new_object = space_map_alloc(mos, tx);
2040		VERIFY3U(new_object, !=, 0);
2041
2042		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2043		    0, -1ULL, 0, &vd->vdev_dtl_lock));
2044		ASSERT(vd->vdev_dtl_sm != NULL);
2045	}
2046
2047	mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
2048
2049	rtsync = range_tree_create(NULL, NULL, &rtlock);
2050
2051	mutex_enter(&rtlock);
2052
2053	mutex_enter(&vd->vdev_dtl_lock);
2054	range_tree_walk(rt, range_tree_add, rtsync);
2055	mutex_exit(&vd->vdev_dtl_lock);
2056
2057	space_map_truncate(vd->vdev_dtl_sm, tx);
2058	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2059	range_tree_vacate(rtsync, NULL, NULL);
2060
2061	range_tree_destroy(rtsync);
2062
2063	mutex_exit(&rtlock);
2064	mutex_destroy(&rtlock);
2065
2066	/*
2067	 * If the object for the space map has changed then dirty
2068	 * the top level so that we update the config.
2069	 */
2070	if (object != space_map_object(vd->vdev_dtl_sm)) {
2071		zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
2072		    "new object %llu", txg, spa_name(spa), object,
2073		    space_map_object(vd->vdev_dtl_sm));
2074		vdev_config_dirty(vd->vdev_top);
2075	}
2076
2077	dmu_tx_commit(tx);
2078
2079	mutex_enter(&vd->vdev_dtl_lock);
2080	space_map_update(vd->vdev_dtl_sm);
2081	mutex_exit(&vd->vdev_dtl_lock);
2082}
2083
2084/*
2085 * Determine whether the specified vdev can be offlined/detached/removed
2086 * without losing data.
2087 */
2088boolean_t
2089vdev_dtl_required(vdev_t *vd)
2090{
2091	spa_t *spa = vd->vdev_spa;
2092	vdev_t *tvd = vd->vdev_top;
2093	uint8_t cant_read = vd->vdev_cant_read;
2094	boolean_t required;
2095
2096	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2097
2098	if (vd == spa->spa_root_vdev || vd == tvd)
2099		return (B_TRUE);
2100
2101	/*
2102	 * Temporarily mark the device as unreadable, and then determine
2103	 * whether this results in any DTL outages in the top-level vdev.
2104	 * If not, we can safely offline/detach/remove the device.
2105	 */
2106	vd->vdev_cant_read = B_TRUE;
2107	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2108	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2109	vd->vdev_cant_read = cant_read;
2110	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2111
2112	if (!required && zio_injection_enabled)
2113		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2114
2115	return (required);
2116}
2117
2118/*
2119 * Determine if resilver is needed, and if so the txg range.
2120 */
2121boolean_t
2122vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2123{
2124	boolean_t needed = B_FALSE;
2125	uint64_t thismin = UINT64_MAX;
2126	uint64_t thismax = 0;
2127
2128	if (vd->vdev_children == 0) {
2129		mutex_enter(&vd->vdev_dtl_lock);
2130		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2131		    vdev_writeable(vd)) {
2132
2133			thismin = vdev_dtl_min(vd);
2134			thismax = vdev_dtl_max(vd);
2135			needed = B_TRUE;
2136		}
2137		mutex_exit(&vd->vdev_dtl_lock);
2138	} else {
2139		for (int c = 0; c < vd->vdev_children; c++) {
2140			vdev_t *cvd = vd->vdev_child[c];
2141			uint64_t cmin, cmax;
2142
2143			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2144				thismin = MIN(thismin, cmin);
2145				thismax = MAX(thismax, cmax);
2146				needed = B_TRUE;
2147			}
2148		}
2149	}
2150
2151	if (needed && minp) {
2152		*minp = thismin;
2153		*maxp = thismax;
2154	}
2155	return (needed);
2156}
2157
2158void
2159vdev_load(vdev_t *vd)
2160{
2161	/*
2162	 * Recursively load all children.
2163	 */
2164	for (int c = 0; c < vd->vdev_children; c++)
2165		vdev_load(vd->vdev_child[c]);
2166
2167	/*
2168	 * If this is a top-level vdev, initialize its metaslabs.
2169	 */
2170	if (vd == vd->vdev_top && !vd->vdev_ishole &&
2171	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2172	    vdev_metaslab_init(vd, 0) != 0))
2173		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2174		    VDEV_AUX_CORRUPT_DATA);
2175
2176	/*
2177	 * If this is a leaf vdev, load its DTL.
2178	 */
2179	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2180		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2181		    VDEV_AUX_CORRUPT_DATA);
2182}
2183
2184/*
2185 * The special vdev case is used for hot spares and l2cache devices.  Its
2186 * sole purpose it to set the vdev state for the associated vdev.  To do this,
2187 * we make sure that we can open the underlying device, then try to read the
2188 * label, and make sure that the label is sane and that it hasn't been
2189 * repurposed to another pool.
2190 */
2191int
2192vdev_validate_aux(vdev_t *vd)
2193{
2194	nvlist_t *label;
2195	uint64_t guid, version;
2196	uint64_t state;
2197
2198	if (!vdev_readable(vd))
2199		return (0);
2200
2201	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2202		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2203		    VDEV_AUX_CORRUPT_DATA);
2204		return (-1);
2205	}
2206
2207	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2208	    !SPA_VERSION_IS_SUPPORTED(version) ||
2209	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2210	    guid != vd->vdev_guid ||
2211	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2212		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2213		    VDEV_AUX_CORRUPT_DATA);
2214		nvlist_free(label);
2215		return (-1);
2216	}
2217
2218	/*
2219	 * We don't actually check the pool state here.  If it's in fact in
2220	 * use by another pool, we update this fact on the fly when requested.
2221	 */
2222	nvlist_free(label);
2223	return (0);
2224}
2225
2226void
2227vdev_remove(vdev_t *vd, uint64_t txg)
2228{
2229	spa_t *spa = vd->vdev_spa;
2230	objset_t *mos = spa->spa_meta_objset;
2231	dmu_tx_t *tx;
2232
2233	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2234	ASSERT(vd == vd->vdev_top);
2235	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2236
2237	if (vd->vdev_ms != NULL) {
2238		metaslab_group_t *mg = vd->vdev_mg;
2239
2240		metaslab_group_histogram_verify(mg);
2241		metaslab_class_histogram_verify(mg->mg_class);
2242
2243		for (int m = 0; m < vd->vdev_ms_count; m++) {
2244			metaslab_t *msp = vd->vdev_ms[m];
2245
2246			if (msp == NULL || msp->ms_sm == NULL)
2247				continue;
2248
2249			mutex_enter(&msp->ms_lock);
2250			/*
2251			 * If the metaslab was not loaded when the vdev
2252			 * was removed then the histogram accounting may
2253			 * not be accurate. Update the histogram information
2254			 * here so that we ensure that the metaslab group
2255			 * and metaslab class are up-to-date.
2256			 */
2257			metaslab_group_histogram_remove(mg, msp);
2258
2259			VERIFY0(space_map_allocated(msp->ms_sm));
2260			space_map_free(msp->ms_sm, tx);
2261			space_map_close(msp->ms_sm);
2262			msp->ms_sm = NULL;
2263			mutex_exit(&msp->ms_lock);
2264		}
2265
2266		metaslab_group_histogram_verify(mg);
2267		metaslab_class_histogram_verify(mg->mg_class);
2268		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2269			ASSERT0(mg->mg_histogram[i]);
2270
2271	}
2272
2273	if (vd->vdev_ms_array) {
2274		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2275		vd->vdev_ms_array = 0;
2276	}
2277
2278	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2279		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2280		vd->vdev_top_zap = 0;
2281	}
2282	dmu_tx_commit(tx);
2283}
2284
2285void
2286vdev_sync_done(vdev_t *vd, uint64_t txg)
2287{
2288	metaslab_t *msp;
2289	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2290
2291	ASSERT(!vd->vdev_ishole);
2292
2293	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2294		metaslab_sync_done(msp, txg);
2295
2296	if (reassess)
2297		metaslab_sync_reassess(vd->vdev_mg);
2298}
2299
2300void
2301vdev_sync(vdev_t *vd, uint64_t txg)
2302{
2303	spa_t *spa = vd->vdev_spa;
2304	vdev_t *lvd;
2305	metaslab_t *msp;
2306	dmu_tx_t *tx;
2307
2308	ASSERT(!vd->vdev_ishole);
2309
2310	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2311		ASSERT(vd == vd->vdev_top);
2312		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2313		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2314		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2315		ASSERT(vd->vdev_ms_array != 0);
2316		vdev_config_dirty(vd);
2317		dmu_tx_commit(tx);
2318	}
2319
2320	/*
2321	 * Remove the metadata associated with this vdev once it's empty.
2322	 */
2323	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2324		vdev_remove(vd, txg);
2325
2326	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2327		metaslab_sync(msp, txg);
2328		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2329	}
2330
2331	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2332		vdev_dtl_sync(lvd, txg);
2333
2334	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2335}
2336
2337uint64_t
2338vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2339{
2340	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2341}
2342
2343/*
2344 * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2345 * not be opened, and no I/O is attempted.
2346 */
2347int
2348vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2349{
2350	vdev_t *vd, *tvd;
2351
2352	spa_vdev_state_enter(spa, SCL_NONE);
2353
2354	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2355		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2356
2357	if (!vd->vdev_ops->vdev_op_leaf)
2358		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2359
2360	tvd = vd->vdev_top;
2361
2362	/*
2363	 * We don't directly use the aux state here, but if we do a
2364	 * vdev_reopen(), we need this value to be present to remember why we
2365	 * were faulted.
2366	 */
2367	vd->vdev_label_aux = aux;
2368
2369	/*
2370	 * Faulted state takes precedence over degraded.
2371	 */
2372	vd->vdev_delayed_close = B_FALSE;
2373	vd->vdev_faulted = 1ULL;
2374	vd->vdev_degraded = 0ULL;
2375	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2376
2377	/*
2378	 * If this device has the only valid copy of the data, then
2379	 * back off and simply mark the vdev as degraded instead.
2380	 */
2381	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2382		vd->vdev_degraded = 1ULL;
2383		vd->vdev_faulted = 0ULL;
2384
2385		/*
2386		 * If we reopen the device and it's not dead, only then do we
2387		 * mark it degraded.
2388		 */
2389		vdev_reopen(tvd);
2390
2391		if (vdev_readable(vd))
2392			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2393	}
2394
2395	return (spa_vdev_state_exit(spa, vd, 0));
2396}
2397
2398/*
2399 * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2400 * user that something is wrong.  The vdev continues to operate as normal as far
2401 * as I/O is concerned.
2402 */
2403int
2404vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2405{
2406	vdev_t *vd;
2407
2408	spa_vdev_state_enter(spa, SCL_NONE);
2409
2410	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2411		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2412
2413	if (!vd->vdev_ops->vdev_op_leaf)
2414		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2415
2416	/*
2417	 * If the vdev is already faulted, then don't do anything.
2418	 */
2419	if (vd->vdev_faulted || vd->vdev_degraded)
2420		return (spa_vdev_state_exit(spa, NULL, 0));
2421
2422	vd->vdev_degraded = 1ULL;
2423	if (!vdev_is_dead(vd))
2424		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2425		    aux);
2426
2427	return (spa_vdev_state_exit(spa, vd, 0));
2428}
2429
2430/*
2431 * Online the given vdev.
2432 *
2433 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2434 * spare device should be detached when the device finishes resilvering.
2435 * Second, the online should be treated like a 'test' online case, so no FMA
2436 * events are generated if the device fails to open.
2437 */
2438int
2439vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2440{
2441	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2442	boolean_t postevent = B_FALSE;
2443
2444	spa_vdev_state_enter(spa, SCL_NONE);
2445
2446	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2447		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2448
2449	if (!vd->vdev_ops->vdev_op_leaf)
2450		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2451
2452	postevent =
2453	    (vd->vdev_offline == B_TRUE || vd->vdev_tmpoffline == B_TRUE) ?
2454	    B_TRUE : B_FALSE;
2455
2456	tvd = vd->vdev_top;
2457	vd->vdev_offline = B_FALSE;
2458	vd->vdev_tmpoffline = B_FALSE;
2459	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2460	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2461
2462	/* XXX - L2ARC 1.0 does not support expansion */
2463	if (!vd->vdev_aux) {
2464		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2465			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2466	}
2467
2468	vdev_reopen(tvd);
2469	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2470
2471	if (!vd->vdev_aux) {
2472		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2473			pvd->vdev_expanding = B_FALSE;
2474	}
2475
2476	if (newstate)
2477		*newstate = vd->vdev_state;
2478	if ((flags & ZFS_ONLINE_UNSPARE) &&
2479	    !vdev_is_dead(vd) && vd->vdev_parent &&
2480	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2481	    vd->vdev_parent->vdev_child[0] == vd)
2482		vd->vdev_unspare = B_TRUE;
2483
2484	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2485
2486		/* XXX - L2ARC 1.0 does not support expansion */
2487		if (vd->vdev_aux)
2488			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2489		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2490	}
2491
2492	if (postevent)
2493		spa_event_notify(spa, vd, ESC_ZFS_VDEV_ONLINE);
2494
2495	return (spa_vdev_state_exit(spa, vd, 0));
2496}
2497
2498static int
2499vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2500{
2501	vdev_t *vd, *tvd;
2502	int error = 0;
2503	uint64_t generation;
2504	metaslab_group_t *mg;
2505
2506top:
2507	spa_vdev_state_enter(spa, SCL_ALLOC);
2508
2509	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2510		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2511
2512	if (!vd->vdev_ops->vdev_op_leaf)
2513		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2514
2515	tvd = vd->vdev_top;
2516	mg = tvd->vdev_mg;
2517	generation = spa->spa_config_generation + 1;
2518
2519	/*
2520	 * If the device isn't already offline, try to offline it.
2521	 */
2522	if (!vd->vdev_offline) {
2523		/*
2524		 * If this device has the only valid copy of some data,
2525		 * don't allow it to be offlined. Log devices are always
2526		 * expendable.
2527		 */
2528		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2529		    vdev_dtl_required(vd))
2530			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2531
2532		/*
2533		 * If the top-level is a slog and it has had allocations
2534		 * then proceed.  We check that the vdev's metaslab group
2535		 * is not NULL since it's possible that we may have just
2536		 * added this vdev but not yet initialized its metaslabs.
2537		 */
2538		if (tvd->vdev_islog && mg != NULL) {
2539			/*
2540			 * Prevent any future allocations.
2541			 */
2542			metaslab_group_passivate(mg);
2543			(void) spa_vdev_state_exit(spa, vd, 0);
2544
2545			error = spa_offline_log(spa);
2546
2547			spa_vdev_state_enter(spa, SCL_ALLOC);
2548
2549			/*
2550			 * Check to see if the config has changed.
2551			 */
2552			if (error || generation != spa->spa_config_generation) {
2553				metaslab_group_activate(mg);
2554				if (error)
2555					return (spa_vdev_state_exit(spa,
2556					    vd, error));
2557				(void) spa_vdev_state_exit(spa, vd, 0);
2558				goto top;
2559			}
2560			ASSERT0(tvd->vdev_stat.vs_alloc);
2561		}
2562
2563		/*
2564		 * Offline this device and reopen its top-level vdev.
2565		 * If the top-level vdev is a log device then just offline
2566		 * it. Otherwise, if this action results in the top-level
2567		 * vdev becoming unusable, undo it and fail the request.
2568		 */
2569		vd->vdev_offline = B_TRUE;
2570		vdev_reopen(tvd);
2571
2572		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2573		    vdev_is_dead(tvd)) {
2574			vd->vdev_offline = B_FALSE;
2575			vdev_reopen(tvd);
2576			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2577		}
2578
2579		/*
2580		 * Add the device back into the metaslab rotor so that
2581		 * once we online the device it's open for business.
2582		 */
2583		if (tvd->vdev_islog && mg != NULL)
2584			metaslab_group_activate(mg);
2585	}
2586
2587	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2588
2589	return (spa_vdev_state_exit(spa, vd, 0));
2590}
2591
2592int
2593vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2594{
2595	int error;
2596
2597	mutex_enter(&spa->spa_vdev_top_lock);
2598	error = vdev_offline_locked(spa, guid, flags);
2599	mutex_exit(&spa->spa_vdev_top_lock);
2600
2601	return (error);
2602}
2603
2604/*
2605 * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2606 * vdev_offline(), we assume the spa config is locked.  We also clear all
2607 * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2608 */
2609void
2610vdev_clear(spa_t *spa, vdev_t *vd)
2611{
2612	vdev_t *rvd = spa->spa_root_vdev;
2613
2614	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2615
2616	if (vd == NULL)
2617		vd = rvd;
2618
2619	vd->vdev_stat.vs_read_errors = 0;
2620	vd->vdev_stat.vs_write_errors = 0;
2621	vd->vdev_stat.vs_checksum_errors = 0;
2622
2623	for (int c = 0; c < vd->vdev_children; c++)
2624		vdev_clear(spa, vd->vdev_child[c]);
2625
2626	/*
2627	 * If we're in the FAULTED state or have experienced failed I/O, then
2628	 * clear the persistent state and attempt to reopen the device.  We
2629	 * also mark the vdev config dirty, so that the new faulted state is
2630	 * written out to disk.
2631	 */
2632	if (vd->vdev_faulted || vd->vdev_degraded ||
2633	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2634
2635		/*
2636		 * When reopening in reponse to a clear event, it may be due to
2637		 * a fmadm repair request.  In this case, if the device is
2638		 * still broken, we want to still post the ereport again.
2639		 */
2640		vd->vdev_forcefault = B_TRUE;
2641
2642		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2643		vd->vdev_cant_read = B_FALSE;
2644		vd->vdev_cant_write = B_FALSE;
2645
2646		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2647
2648		vd->vdev_forcefault = B_FALSE;
2649
2650		if (vd != rvd && vdev_writeable(vd->vdev_top))
2651			vdev_state_dirty(vd->vdev_top);
2652
2653		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2654			spa_async_request(spa, SPA_ASYNC_RESILVER);
2655
2656		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2657	}
2658
2659	/*
2660	 * When clearing a FMA-diagnosed fault, we always want to
2661	 * unspare the device, as we assume that the original spare was
2662	 * done in response to the FMA fault.
2663	 */
2664	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2665	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2666	    vd->vdev_parent->vdev_child[0] == vd)
2667		vd->vdev_unspare = B_TRUE;
2668}
2669
2670boolean_t
2671vdev_is_dead(vdev_t *vd)
2672{
2673	/*
2674	 * Holes and missing devices are always considered "dead".
2675	 * This simplifies the code since we don't have to check for
2676	 * these types of devices in the various code paths.
2677	 * Instead we rely on the fact that we skip over dead devices
2678	 * before issuing I/O to them.
2679	 */
2680	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2681	    vd->vdev_ops == &vdev_missing_ops);
2682}
2683
2684boolean_t
2685vdev_readable(vdev_t *vd)
2686{
2687	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2688}
2689
2690boolean_t
2691vdev_writeable(vdev_t *vd)
2692{
2693	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2694}
2695
2696boolean_t
2697vdev_allocatable(vdev_t *vd)
2698{
2699	uint64_t state = vd->vdev_state;
2700
2701	/*
2702	 * We currently allow allocations from vdevs which may be in the
2703	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2704	 * fails to reopen then we'll catch it later when we're holding
2705	 * the proper locks.  Note that we have to get the vdev state
2706	 * in a local variable because although it changes atomically,
2707	 * we're asking two separate questions about it.
2708	 */
2709	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2710	    !vd->vdev_cant_write && !vd->vdev_ishole &&
2711	    vd->vdev_mg->mg_initialized);
2712}
2713
2714boolean_t
2715vdev_accessible(vdev_t *vd, zio_t *zio)
2716{
2717	ASSERT(zio->io_vd == vd);
2718
2719	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2720		return (B_FALSE);
2721
2722	if (zio->io_type == ZIO_TYPE_READ)
2723		return (!vd->vdev_cant_read);
2724
2725	if (zio->io_type == ZIO_TYPE_WRITE)
2726		return (!vd->vdev_cant_write);
2727
2728	return (B_TRUE);
2729}
2730
2731/*
2732 * Get statistics for the given vdev.
2733 */
2734void
2735vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2736{
2737	spa_t *spa = vd->vdev_spa;
2738	vdev_t *rvd = spa->spa_root_vdev;
2739	vdev_t *tvd = vd->vdev_top;
2740
2741	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2742
2743	mutex_enter(&vd->vdev_stat_lock);
2744	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2745	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2746	vs->vs_state = vd->vdev_state;
2747	vs->vs_rsize = vdev_get_min_asize(vd);
2748	if (vd->vdev_ops->vdev_op_leaf)
2749		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2750	/*
2751	 * Report expandable space on top-level, non-auxillary devices only.
2752	 * The expandable space is reported in terms of metaslab sized units
2753	 * since that determines how much space the pool can expand.
2754	 */
2755	if (vd->vdev_aux == NULL && tvd != NULL) {
2756		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize,
2757		    1ULL << tvd->vdev_ms_shift);
2758	}
2759	if (vd->vdev_aux == NULL && vd == vd->vdev_top && !vd->vdev_ishole) {
2760		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
2761	}
2762
2763	/*
2764	 * If we're getting stats on the root vdev, aggregate the I/O counts
2765	 * over all top-level vdevs (i.e. the direct children of the root).
2766	 */
2767	if (vd == rvd) {
2768		for (int c = 0; c < rvd->vdev_children; c++) {
2769			vdev_t *cvd = rvd->vdev_child[c];
2770			vdev_stat_t *cvs = &cvd->vdev_stat;
2771
2772			for (int t = 0; t < ZIO_TYPES; t++) {
2773				vs->vs_ops[t] += cvs->vs_ops[t];
2774				vs->vs_bytes[t] += cvs->vs_bytes[t];
2775			}
2776			cvs->vs_scan_removing = cvd->vdev_removing;
2777		}
2778	}
2779	mutex_exit(&vd->vdev_stat_lock);
2780}
2781
2782void
2783vdev_clear_stats(vdev_t *vd)
2784{
2785	mutex_enter(&vd->vdev_stat_lock);
2786	vd->vdev_stat.vs_space = 0;
2787	vd->vdev_stat.vs_dspace = 0;
2788	vd->vdev_stat.vs_alloc = 0;
2789	mutex_exit(&vd->vdev_stat_lock);
2790}
2791
2792void
2793vdev_scan_stat_init(vdev_t *vd)
2794{
2795	vdev_stat_t *vs = &vd->vdev_stat;
2796
2797	for (int c = 0; c < vd->vdev_children; c++)
2798		vdev_scan_stat_init(vd->vdev_child[c]);
2799
2800	mutex_enter(&vd->vdev_stat_lock);
2801	vs->vs_scan_processed = 0;
2802	mutex_exit(&vd->vdev_stat_lock);
2803}
2804
2805void
2806vdev_stat_update(zio_t *zio, uint64_t psize)
2807{
2808	spa_t *spa = zio->io_spa;
2809	vdev_t *rvd = spa->spa_root_vdev;
2810	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2811	vdev_t *pvd;
2812	uint64_t txg = zio->io_txg;
2813	vdev_stat_t *vs = &vd->vdev_stat;
2814	zio_type_t type = zio->io_type;
2815	int flags = zio->io_flags;
2816
2817	/*
2818	 * If this i/o is a gang leader, it didn't do any actual work.
2819	 */
2820	if (zio->io_gang_tree)
2821		return;
2822
2823	if (zio->io_error == 0) {
2824		/*
2825		 * If this is a root i/o, don't count it -- we've already
2826		 * counted the top-level vdevs, and vdev_get_stats() will
2827		 * aggregate them when asked.  This reduces contention on
2828		 * the root vdev_stat_lock and implicitly handles blocks
2829		 * that compress away to holes, for which there is no i/o.
2830		 * (Holes never create vdev children, so all the counters
2831		 * remain zero, which is what we want.)
2832		 *
2833		 * Note: this only applies to successful i/o (io_error == 0)
2834		 * because unlike i/o counts, errors are not additive.
2835		 * When reading a ditto block, for example, failure of
2836		 * one top-level vdev does not imply a root-level error.
2837		 */
2838		if (vd == rvd)
2839			return;
2840
2841		ASSERT(vd == zio->io_vd);
2842
2843		if (flags & ZIO_FLAG_IO_BYPASS)
2844			return;
2845
2846		mutex_enter(&vd->vdev_stat_lock);
2847
2848		if (flags & ZIO_FLAG_IO_REPAIR) {
2849			if (flags & ZIO_FLAG_SCAN_THREAD) {
2850				dsl_scan_phys_t *scn_phys =
2851				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2852				uint64_t *processed = &scn_phys->scn_processed;
2853
2854				/* XXX cleanup? */
2855				if (vd->vdev_ops->vdev_op_leaf)
2856					atomic_add_64(processed, psize);
2857				vs->vs_scan_processed += psize;
2858			}
2859
2860			if (flags & ZIO_FLAG_SELF_HEAL)
2861				vs->vs_self_healed += psize;
2862		}
2863
2864		vs->vs_ops[type]++;
2865		vs->vs_bytes[type] += psize;
2866
2867		mutex_exit(&vd->vdev_stat_lock);
2868		return;
2869	}
2870
2871	if (flags & ZIO_FLAG_SPECULATIVE)
2872		return;
2873
2874	/*
2875	 * If this is an I/O error that is going to be retried, then ignore the
2876	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2877	 * hard errors, when in reality they can happen for any number of
2878	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2879	 */
2880	if (zio->io_error == EIO &&
2881	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2882		return;
2883
2884	/*
2885	 * Intent logs writes won't propagate their error to the root
2886	 * I/O so don't mark these types of failures as pool-level
2887	 * errors.
2888	 */
2889	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2890		return;
2891
2892	mutex_enter(&vd->vdev_stat_lock);
2893	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2894		if (zio->io_error == ECKSUM)
2895			vs->vs_checksum_errors++;
2896		else
2897			vs->vs_read_errors++;
2898	}
2899	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2900		vs->vs_write_errors++;
2901	mutex_exit(&vd->vdev_stat_lock);
2902
2903	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2904	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2905	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2906	    spa->spa_claiming)) {
2907		/*
2908		 * This is either a normal write (not a repair), or it's
2909		 * a repair induced by the scrub thread, or it's a repair
2910		 * made by zil_claim() during spa_load() in the first txg.
2911		 * In the normal case, we commit the DTL change in the same
2912		 * txg as the block was born.  In the scrub-induced repair
2913		 * case, we know that scrubs run in first-pass syncing context,
2914		 * so we commit the DTL change in spa_syncing_txg(spa).
2915		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2916		 *
2917		 * We currently do not make DTL entries for failed spontaneous
2918		 * self-healing writes triggered by normal (non-scrubbing)
2919		 * reads, because we have no transactional context in which to
2920		 * do so -- and it's not clear that it'd be desirable anyway.
2921		 */
2922		if (vd->vdev_ops->vdev_op_leaf) {
2923			uint64_t commit_txg = txg;
2924			if (flags & ZIO_FLAG_SCAN_THREAD) {
2925				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2926				ASSERT(spa_sync_pass(spa) == 1);
2927				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2928				commit_txg = spa_syncing_txg(spa);
2929			} else if (spa->spa_claiming) {
2930				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2931				commit_txg = spa_first_txg(spa);
2932			}
2933			ASSERT(commit_txg >= spa_syncing_txg(spa));
2934			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2935				return;
2936			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2937				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2938			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2939		}
2940		if (vd != rvd)
2941			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2942	}
2943}
2944
2945/*
2946 * Update the in-core space usage stats for this vdev, its metaslab class,
2947 * and the root vdev.
2948 */
2949void
2950vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2951    int64_t space_delta)
2952{
2953	int64_t dspace_delta = space_delta;
2954	spa_t *spa = vd->vdev_spa;
2955	vdev_t *rvd = spa->spa_root_vdev;
2956	metaslab_group_t *mg = vd->vdev_mg;
2957	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2958
2959	ASSERT(vd == vd->vdev_top);
2960
2961	/*
2962	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2963	 * factor.  We must calculate this here and not at the root vdev
2964	 * because the root vdev's psize-to-asize is simply the max of its
2965	 * childrens', thus not accurate enough for us.
2966	 */
2967	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2968	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2969	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2970	    vd->vdev_deflate_ratio;
2971
2972	mutex_enter(&vd->vdev_stat_lock);
2973	vd->vdev_stat.vs_alloc += alloc_delta;
2974	vd->vdev_stat.vs_space += space_delta;
2975	vd->vdev_stat.vs_dspace += dspace_delta;
2976	mutex_exit(&vd->vdev_stat_lock);
2977
2978	if (mc == spa_normal_class(spa)) {
2979		mutex_enter(&rvd->vdev_stat_lock);
2980		rvd->vdev_stat.vs_alloc += alloc_delta;
2981		rvd->vdev_stat.vs_space += space_delta;
2982		rvd->vdev_stat.vs_dspace += dspace_delta;
2983		mutex_exit(&rvd->vdev_stat_lock);
2984	}
2985
2986	if (mc != NULL) {
2987		ASSERT(rvd == vd->vdev_parent);
2988		ASSERT(vd->vdev_ms_count != 0);
2989
2990		metaslab_class_space_update(mc,
2991		    alloc_delta, defer_delta, space_delta, dspace_delta);
2992	}
2993}
2994
2995/*
2996 * Mark a top-level vdev's config as dirty, placing it on the dirty list
2997 * so that it will be written out next time the vdev configuration is synced.
2998 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2999 */
3000void
3001vdev_config_dirty(vdev_t *vd)
3002{
3003	spa_t *spa = vd->vdev_spa;
3004	vdev_t *rvd = spa->spa_root_vdev;
3005	int c;
3006
3007	ASSERT(spa_writeable(spa));
3008
3009	/*
3010	 * If this is an aux vdev (as with l2cache and spare devices), then we
3011	 * update the vdev config manually and set the sync flag.
3012	 */
3013	if (vd->vdev_aux != NULL) {
3014		spa_aux_vdev_t *sav = vd->vdev_aux;
3015		nvlist_t **aux;
3016		uint_t naux;
3017
3018		for (c = 0; c < sav->sav_count; c++) {
3019			if (sav->sav_vdevs[c] == vd)
3020				break;
3021		}
3022
3023		if (c == sav->sav_count) {
3024			/*
3025			 * We're being removed.  There's nothing more to do.
3026			 */
3027			ASSERT(sav->sav_sync == B_TRUE);
3028			return;
3029		}
3030
3031		sav->sav_sync = B_TRUE;
3032
3033		if (nvlist_lookup_nvlist_array(sav->sav_config,
3034		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3035			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3036			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3037		}
3038
3039		ASSERT(c < naux);
3040
3041		/*
3042		 * Setting the nvlist in the middle if the array is a little
3043		 * sketchy, but it will work.
3044		 */
3045		nvlist_free(aux[c]);
3046		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3047
3048		return;
3049	}
3050
3051	/*
3052	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3053	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3054	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3055	 * so this is sufficient to ensure mutual exclusion.
3056	 */
3057	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3058	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3059	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3060
3061	if (vd == rvd) {
3062		for (c = 0; c < rvd->vdev_children; c++)
3063			vdev_config_dirty(rvd->vdev_child[c]);
3064	} else {
3065		ASSERT(vd == vd->vdev_top);
3066
3067		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3068		    !vd->vdev_ishole)
3069			list_insert_head(&spa->spa_config_dirty_list, vd);
3070	}
3071}
3072
3073void
3074vdev_config_clean(vdev_t *vd)
3075{
3076	spa_t *spa = vd->vdev_spa;
3077
3078	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3079	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3080	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3081
3082	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3083	list_remove(&spa->spa_config_dirty_list, vd);
3084}
3085
3086/*
3087 * Mark a top-level vdev's state as dirty, so that the next pass of
3088 * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3089 * the state changes from larger config changes because they require
3090 * much less locking, and are often needed for administrative actions.
3091 */
3092void
3093vdev_state_dirty(vdev_t *vd)
3094{
3095	spa_t *spa = vd->vdev_spa;
3096
3097	ASSERT(spa_writeable(spa));
3098	ASSERT(vd == vd->vdev_top);
3099
3100	/*
3101	 * The state list is protected by the SCL_STATE lock.  The caller
3102	 * must either hold SCL_STATE as writer, or must be the sync thread
3103	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3104	 * so this is sufficient to ensure mutual exclusion.
3105	 */
3106	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3107	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3108	    spa_config_held(spa, SCL_STATE, RW_READER)));
3109
3110	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
3111		list_insert_head(&spa->spa_state_dirty_list, vd);
3112}
3113
3114void
3115vdev_state_clean(vdev_t *vd)
3116{
3117	spa_t *spa = vd->vdev_spa;
3118
3119	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3120	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3121	    spa_config_held(spa, SCL_STATE, RW_READER)));
3122
3123	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3124	list_remove(&spa->spa_state_dirty_list, vd);
3125}
3126
3127/*
3128 * Propagate vdev state up from children to parent.
3129 */
3130void
3131vdev_propagate_state(vdev_t *vd)
3132{
3133	spa_t *spa = vd->vdev_spa;
3134	vdev_t *rvd = spa->spa_root_vdev;
3135	int degraded = 0, faulted = 0;
3136	int corrupted = 0;
3137	vdev_t *child;
3138
3139	if (vd->vdev_children > 0) {
3140		for (int c = 0; c < vd->vdev_children; c++) {
3141			child = vd->vdev_child[c];
3142
3143			/*
3144			 * Don't factor holes into the decision.
3145			 */
3146			if (child->vdev_ishole)
3147				continue;
3148
3149			if (!vdev_readable(child) ||
3150			    (!vdev_writeable(child) && spa_writeable(spa))) {
3151				/*
3152				 * Root special: if there is a top-level log
3153				 * device, treat the root vdev as if it were
3154				 * degraded.
3155				 */
3156				if (child->vdev_islog && vd == rvd)
3157					degraded++;
3158				else
3159					faulted++;
3160			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3161				degraded++;
3162			}
3163
3164			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3165				corrupted++;
3166		}
3167
3168		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3169
3170		/*
3171		 * Root special: if there is a top-level vdev that cannot be
3172		 * opened due to corrupted metadata, then propagate the root
3173		 * vdev's aux state as 'corrupt' rather than 'insufficient
3174		 * replicas'.
3175		 */
3176		if (corrupted && vd == rvd &&
3177		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3178			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3179			    VDEV_AUX_CORRUPT_DATA);
3180	}
3181
3182	if (vd->vdev_parent)
3183		vdev_propagate_state(vd->vdev_parent);
3184}
3185
3186/*
3187 * Set a vdev's state.  If this is during an open, we don't update the parent
3188 * state, because we're in the process of opening children depth-first.
3189 * Otherwise, we propagate the change to the parent.
3190 *
3191 * If this routine places a device in a faulted state, an appropriate ereport is
3192 * generated.
3193 */
3194void
3195vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3196{
3197	uint64_t save_state;
3198	spa_t *spa = vd->vdev_spa;
3199
3200	if (state == vd->vdev_state) {
3201		vd->vdev_stat.vs_aux = aux;
3202		return;
3203	}
3204
3205	save_state = vd->vdev_state;
3206
3207	vd->vdev_state = state;
3208	vd->vdev_stat.vs_aux = aux;
3209
3210	/*
3211	 * If we are setting the vdev state to anything but an open state, then
3212	 * always close the underlying device unless the device has requested
3213	 * a delayed close (i.e. we're about to remove or fault the device).
3214	 * Otherwise, we keep accessible but invalid devices open forever.
3215	 * We don't call vdev_close() itself, because that implies some extra
3216	 * checks (offline, etc) that we don't want here.  This is limited to
3217	 * leaf devices, because otherwise closing the device will affect other
3218	 * children.
3219	 */
3220	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3221	    vd->vdev_ops->vdev_op_leaf)
3222		vd->vdev_ops->vdev_op_close(vd);
3223
3224	/*
3225	 * If we have brought this vdev back into service, we need
3226	 * to notify fmd so that it can gracefully repair any outstanding
3227	 * cases due to a missing device.  We do this in all cases, even those
3228	 * that probably don't correlate to a repaired fault.  This is sure to
3229	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3230	 * this is a transient state it's OK, as the retire agent will
3231	 * double-check the state of the vdev before repairing it.
3232	 */
3233	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3234	    vd->vdev_prevstate != state)
3235		zfs_post_state_change(spa, vd);
3236
3237	if (vd->vdev_removed &&
3238	    state == VDEV_STATE_CANT_OPEN &&
3239	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3240		/*
3241		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3242		 * device was previously marked removed and someone attempted to
3243		 * reopen it.  If this failed due to a nonexistent device, then
3244		 * keep the device in the REMOVED state.  We also let this be if
3245		 * it is one of our special test online cases, which is only
3246		 * attempting to online the device and shouldn't generate an FMA
3247		 * fault.
3248		 */
3249		vd->vdev_state = VDEV_STATE_REMOVED;
3250		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3251	} else if (state == VDEV_STATE_REMOVED) {
3252		vd->vdev_removed = B_TRUE;
3253	} else if (state == VDEV_STATE_CANT_OPEN) {
3254		/*
3255		 * If we fail to open a vdev during an import or recovery, we
3256		 * mark it as "not available", which signifies that it was
3257		 * never there to begin with.  Failure to open such a device
3258		 * is not considered an error.
3259		 */
3260		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3261		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3262		    vd->vdev_ops->vdev_op_leaf)
3263			vd->vdev_not_present = 1;
3264
3265		/*
3266		 * Post the appropriate ereport.  If the 'prevstate' field is
3267		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3268		 * that this is part of a vdev_reopen().  In this case, we don't
3269		 * want to post the ereport if the device was already in the
3270		 * CANT_OPEN state beforehand.
3271		 *
3272		 * If the 'checkremove' flag is set, then this is an attempt to
3273		 * online the device in response to an insertion event.  If we
3274		 * hit this case, then we have detected an insertion event for a
3275		 * faulted or offline device that wasn't in the removed state.
3276		 * In this scenario, we don't post an ereport because we are
3277		 * about to replace the device, or attempt an online with
3278		 * vdev_forcefault, which will generate the fault for us.
3279		 */
3280		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3281		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3282		    vd != spa->spa_root_vdev) {
3283			const char *class;
3284
3285			switch (aux) {
3286			case VDEV_AUX_OPEN_FAILED:
3287				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3288				break;
3289			case VDEV_AUX_CORRUPT_DATA:
3290				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3291				break;
3292			case VDEV_AUX_NO_REPLICAS:
3293				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3294				break;
3295			case VDEV_AUX_BAD_GUID_SUM:
3296				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3297				break;
3298			case VDEV_AUX_TOO_SMALL:
3299				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3300				break;
3301			case VDEV_AUX_BAD_LABEL:
3302				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3303				break;
3304			default:
3305				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3306			}
3307
3308			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3309		}
3310
3311		/* Erase any notion of persistent removed state */
3312		vd->vdev_removed = B_FALSE;
3313	} else {
3314		vd->vdev_removed = B_FALSE;
3315	}
3316
3317	if (!isopen && vd->vdev_parent)
3318		vdev_propagate_state(vd->vdev_parent);
3319}
3320
3321/*
3322 * Check the vdev configuration to ensure that it's capable of supporting
3323 * a root pool. We do not support partial configuration.
3324 * In addition, only a single top-level vdev is allowed.
3325 */
3326boolean_t
3327vdev_is_bootable(vdev_t *vd)
3328{
3329	if (!vd->vdev_ops->vdev_op_leaf) {
3330		char *vdev_type = vd->vdev_ops->vdev_op_type;
3331
3332		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3333		    vd->vdev_children > 1) {
3334			return (B_FALSE);
3335		} else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3336			return (B_FALSE);
3337		}
3338	}
3339
3340	for (int c = 0; c < vd->vdev_children; c++) {
3341		if (!vdev_is_bootable(vd->vdev_child[c]))
3342			return (B_FALSE);
3343	}
3344	return (B_TRUE);
3345}
3346
3347/*
3348 * Load the state from the original vdev tree (ovd) which
3349 * we've retrieved from the MOS config object. If the original
3350 * vdev was offline or faulted then we transfer that state to the
3351 * device in the current vdev tree (nvd).
3352 */
3353void
3354vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3355{
3356	spa_t *spa = nvd->vdev_spa;
3357
3358	ASSERT(nvd->vdev_top->vdev_islog);
3359	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3360	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3361
3362	for (int c = 0; c < nvd->vdev_children; c++)
3363		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3364
3365	if (nvd->vdev_ops->vdev_op_leaf) {
3366		/*
3367		 * Restore the persistent vdev state
3368		 */
3369		nvd->vdev_offline = ovd->vdev_offline;
3370		nvd->vdev_faulted = ovd->vdev_faulted;
3371		nvd->vdev_degraded = ovd->vdev_degraded;
3372		nvd->vdev_removed = ovd->vdev_removed;
3373	}
3374}
3375
3376/*
3377 * Determine if a log device has valid content.  If the vdev was
3378 * removed or faulted in the MOS config then we know that
3379 * the content on the log device has already been written to the pool.
3380 */
3381boolean_t
3382vdev_log_state_valid(vdev_t *vd)
3383{
3384	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3385	    !vd->vdev_removed)
3386		return (B_TRUE);
3387
3388	for (int c = 0; c < vd->vdev_children; c++)
3389		if (vdev_log_state_valid(vd->vdev_child[c]))
3390			return (B_TRUE);
3391
3392	return (B_FALSE);
3393}
3394
3395/*
3396 * Expand a vdev if possible.
3397 */
3398void
3399vdev_expand(vdev_t *vd, uint64_t txg)
3400{
3401	ASSERT(vd->vdev_top == vd);
3402	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3403
3404	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3405		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3406		vdev_config_dirty(vd);
3407	}
3408}
3409
3410/*
3411 * Split a vdev.
3412 */
3413void
3414vdev_split(vdev_t *vd)
3415{
3416	vdev_t *cvd, *pvd = vd->vdev_parent;
3417
3418	vdev_remove_child(pvd, vd);
3419	vdev_compact_children(pvd);
3420
3421	cvd = pvd->vdev_child[0];
3422	if (pvd->vdev_children == 1) {
3423		vdev_remove_parent(cvd);
3424		cvd->vdev_splitting = B_TRUE;
3425	}
3426	vdev_propagate_state(cvd);
3427}
3428
3429void
3430vdev_deadman(vdev_t *vd)
3431{
3432	for (int c = 0; c < vd->vdev_children; c++) {
3433		vdev_t *cvd = vd->vdev_child[c];
3434
3435		vdev_deadman(cvd);
3436	}
3437
3438	if (vd->vdev_ops->vdev_op_leaf) {
3439		vdev_queue_t *vq = &vd->vdev_queue;
3440
3441		mutex_enter(&vq->vq_lock);
3442		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3443			spa_t *spa = vd->vdev_spa;
3444			zio_t *fio;
3445			uint64_t delta;
3446
3447			/*
3448			 * Look at the head of all the pending queues,
3449			 * if any I/O has been outstanding for longer than
3450			 * the spa_deadman_synctime we panic the system.
3451			 */
3452			fio = avl_first(&vq->vq_active_tree);
3453			delta = gethrtime() - fio->io_timestamp;
3454			if (delta > spa_deadman_synctime(spa)) {
3455				zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3456				    "delta %lluns, last io %lluns",
3457				    fio->io_timestamp, delta,
3458				    vq->vq_io_complete_ts);
3459				fm_panic("I/O to pool '%s' appears to be "
3460				    "hung.", spa_name(spa));
3461			}
3462		}
3463		mutex_exit(&vq->vq_lock);
3464	}
3465}
3466