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 2017 Nexenta Systems, Inc.
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#include <sys/abd.h>
49
50/*
51 * Virtual device management.
52 */
53
54static vdev_ops_t *vdev_ops_table[] = {
55	&vdev_root_ops,
56	&vdev_raidz_ops,
57	&vdev_mirror_ops,
58	&vdev_replacing_ops,
59	&vdev_spare_ops,
60	&vdev_disk_ops,
61	&vdev_file_ops,
62	&vdev_missing_ops,
63	&vdev_hole_ops,
64	NULL
65};
66
67/* maximum scrub/resilver I/O queue per leaf vdev */
68int zfs_scrub_limit = 10;
69
70/*
71 * When a vdev is added, it will be divided into approximately (but no
72 * more than) this number of metaslabs.
73 */
74int metaslabs_per_vdev = 200;
75
76/*
77 * Given a vdev type, return the appropriate ops vector.
78 */
79static vdev_ops_t *
80vdev_getops(const char *type)
81{
82	vdev_ops_t *ops, **opspp;
83
84	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
85		if (strcmp(ops->vdev_op_type, type) == 0)
86			break;
87
88	return (ops);
89}
90
91/*
92 * Default asize function: return the MAX of psize with the asize of
93 * all children.  This is what's used by anything other than RAID-Z.
94 */
95uint64_t
96vdev_default_asize(vdev_t *vd, uint64_t psize)
97{
98	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
99	uint64_t csize;
100
101	for (int c = 0; c < vd->vdev_children; c++) {
102		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
103		asize = MAX(asize, csize);
104	}
105
106	return (asize);
107}
108
109/*
110 * Get the minimum allocatable size. We define the allocatable size as
111 * the vdev's asize rounded to the nearest metaslab. This allows us to
112 * replace or attach devices which don't have the same physical size but
113 * can still satisfy the same number of allocations.
114 */
115uint64_t
116vdev_get_min_asize(vdev_t *vd)
117{
118	vdev_t *pvd = vd->vdev_parent;
119
120	/*
121	 * If our parent is NULL (inactive spare or cache) or is the root,
122	 * just return our own asize.
123	 */
124	if (pvd == NULL)
125		return (vd->vdev_asize);
126
127	/*
128	 * The top-level vdev just returns the allocatable size rounded
129	 * to the nearest metaslab.
130	 */
131	if (vd == vd->vdev_top)
132		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
133
134	/*
135	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
136	 * so each child must provide at least 1/Nth of its asize.
137	 */
138	if (pvd->vdev_ops == &vdev_raidz_ops)
139		return ((pvd->vdev_min_asize + pvd->vdev_children - 1) /
140		    pvd->vdev_children);
141
142	return (pvd->vdev_min_asize);
143}
144
145void
146vdev_set_min_asize(vdev_t *vd)
147{
148	vd->vdev_min_asize = vdev_get_min_asize(vd);
149
150	for (int c = 0; c < vd->vdev_children; c++)
151		vdev_set_min_asize(vd->vdev_child[c]);
152}
153
154vdev_t *
155vdev_lookup_top(spa_t *spa, uint64_t vdev)
156{
157	vdev_t *rvd = spa->spa_root_vdev;
158
159	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
160
161	if (vdev < rvd->vdev_children) {
162		ASSERT(rvd->vdev_child[vdev] != NULL);
163		return (rvd->vdev_child[vdev]);
164	}
165
166	return (NULL);
167}
168
169vdev_t *
170vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
171{
172	vdev_t *mvd;
173
174	if (vd->vdev_guid == guid)
175		return (vd);
176
177	for (int c = 0; c < vd->vdev_children; c++)
178		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
179		    NULL)
180			return (mvd);
181
182	return (NULL);
183}
184
185static int
186vdev_count_leaves_impl(vdev_t *vd)
187{
188	int n = 0;
189
190	if (vd->vdev_ops->vdev_op_leaf)
191		return (1);
192
193	for (int c = 0; c < vd->vdev_children; c++)
194		n += vdev_count_leaves_impl(vd->vdev_child[c]);
195
196	return (n);
197}
198
199int
200vdev_count_leaves(spa_t *spa)
201{
202	return (vdev_count_leaves_impl(spa->spa_root_vdev));
203}
204
205void
206vdev_add_child(vdev_t *pvd, vdev_t *cvd)
207{
208	size_t oldsize, newsize;
209	uint64_t id = cvd->vdev_id;
210	vdev_t **newchild;
211	spa_t *spa = cvd->vdev_spa;
212
213	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
214	ASSERT(cvd->vdev_parent == NULL);
215
216	cvd->vdev_parent = pvd;
217
218	if (pvd == NULL)
219		return;
220
221	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
222
223	oldsize = pvd->vdev_children * sizeof (vdev_t *);
224	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
225	newsize = pvd->vdev_children * sizeof (vdev_t *);
226
227	newchild = kmem_zalloc(newsize, KM_SLEEP);
228	if (pvd->vdev_child != NULL) {
229		bcopy(pvd->vdev_child, newchild, oldsize);
230		kmem_free(pvd->vdev_child, oldsize);
231	}
232
233	pvd->vdev_child = newchild;
234	pvd->vdev_child[id] = cvd;
235
236	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
237	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
238
239	/*
240	 * Walk up all ancestors to update guid sum.
241	 */
242	for (; pvd != NULL; pvd = pvd->vdev_parent)
243		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
244}
245
246void
247vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
248{
249	int c;
250	uint_t id = cvd->vdev_id;
251
252	ASSERT(cvd->vdev_parent == pvd);
253
254	if (pvd == NULL)
255		return;
256
257	ASSERT(id < pvd->vdev_children);
258	ASSERT(pvd->vdev_child[id] == cvd);
259
260	pvd->vdev_child[id] = NULL;
261	cvd->vdev_parent = NULL;
262
263	for (c = 0; c < pvd->vdev_children; c++)
264		if (pvd->vdev_child[c])
265			break;
266
267	if (c == pvd->vdev_children) {
268		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
269		pvd->vdev_child = NULL;
270		pvd->vdev_children = 0;
271	}
272
273	/*
274	 * Walk up all ancestors to update guid sum.
275	 */
276	for (; pvd != NULL; pvd = pvd->vdev_parent)
277		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
278}
279
280/*
281 * Remove any holes in the child array.
282 */
283void
284vdev_compact_children(vdev_t *pvd)
285{
286	vdev_t **newchild, *cvd;
287	int oldc = pvd->vdev_children;
288	int newc;
289
290	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
291
292	for (int c = newc = 0; c < oldc; c++)
293		if (pvd->vdev_child[c])
294			newc++;
295
296	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
297
298	for (int c = newc = 0; c < oldc; c++) {
299		if ((cvd = pvd->vdev_child[c]) != NULL) {
300			newchild[newc] = cvd;
301			cvd->vdev_id = newc++;
302		}
303	}
304
305	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
306	pvd->vdev_child = newchild;
307	pvd->vdev_children = newc;
308}
309
310/*
311 * Allocate and minimally initialize a vdev_t.
312 */
313vdev_t *
314vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
315{
316	vdev_t *vd;
317
318	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
319
320	if (spa->spa_root_vdev == NULL) {
321		ASSERT(ops == &vdev_root_ops);
322		spa->spa_root_vdev = vd;
323		spa->spa_load_guid = spa_generate_guid(NULL);
324	}
325
326	if (guid == 0 && ops != &vdev_hole_ops) {
327		if (spa->spa_root_vdev == vd) {
328			/*
329			 * The root vdev's guid will also be the pool guid,
330			 * which must be unique among all pools.
331			 */
332			guid = spa_generate_guid(NULL);
333		} else {
334			/*
335			 * Any other vdev's guid must be unique within the pool.
336			 */
337			guid = spa_generate_guid(spa);
338		}
339		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
340	}
341
342	vd->vdev_spa = spa;
343	vd->vdev_id = id;
344	vd->vdev_guid = guid;
345	vd->vdev_guid_sum = guid;
346	vd->vdev_ops = ops;
347	vd->vdev_state = VDEV_STATE_CLOSED;
348	vd->vdev_ishole = (ops == &vdev_hole_ops);
349
350	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
351	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
352	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
353	mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
354	for (int t = 0; t < DTL_TYPES; t++) {
355		vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
356		    &vd->vdev_dtl_lock);
357	}
358	txg_list_create(&vd->vdev_ms_list, spa,
359	    offsetof(struct metaslab, ms_txg_node));
360	txg_list_create(&vd->vdev_dtl_list, spa,
361	    offsetof(struct vdev, vdev_dtl_node));
362	vd->vdev_stat.vs_timestamp = gethrtime();
363	vdev_queue_init(vd);
364	vdev_cache_init(vd);
365
366	return (vd);
367}
368
369/*
370 * Allocate a new vdev.  The 'alloctype' is used to control whether we are
371 * creating a new vdev or loading an existing one - the behavior is slightly
372 * different for each case.
373 */
374int
375vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
376    int alloctype)
377{
378	vdev_ops_t *ops;
379	char *type;
380	uint64_t guid = 0, islog, nparity;
381	vdev_t *vd;
382
383	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
384
385	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
386		return (SET_ERROR(EINVAL));
387
388	if ((ops = vdev_getops(type)) == NULL)
389		return (SET_ERROR(EINVAL));
390
391	/*
392	 * If this is a load, get the vdev guid from the nvlist.
393	 * Otherwise, vdev_alloc_common() will generate one for us.
394	 */
395	if (alloctype == VDEV_ALLOC_LOAD) {
396		uint64_t label_id;
397
398		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
399		    label_id != id)
400			return (SET_ERROR(EINVAL));
401
402		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
403			return (SET_ERROR(EINVAL));
404	} else if (alloctype == VDEV_ALLOC_SPARE) {
405		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
406			return (SET_ERROR(EINVAL));
407	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
408		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
409			return (SET_ERROR(EINVAL));
410	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
411		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
412			return (SET_ERROR(EINVAL));
413	}
414
415	/*
416	 * The first allocated vdev must be of type 'root'.
417	 */
418	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
419		return (SET_ERROR(EINVAL));
420
421	/*
422	 * Determine whether we're a log vdev.
423	 */
424	islog = 0;
425	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
426	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
427		return (SET_ERROR(ENOTSUP));
428
429	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
430		return (SET_ERROR(ENOTSUP));
431
432	/*
433	 * Set the nparity property for RAID-Z vdevs.
434	 */
435	nparity = -1ULL;
436	if (ops == &vdev_raidz_ops) {
437		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
438		    &nparity) == 0) {
439			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
440				return (SET_ERROR(EINVAL));
441			/*
442			 * Previous versions could only support 1 or 2 parity
443			 * device.
444			 */
445			if (nparity > 1 &&
446			    spa_version(spa) < SPA_VERSION_RAIDZ2)
447				return (SET_ERROR(ENOTSUP));
448			if (nparity > 2 &&
449			    spa_version(spa) < SPA_VERSION_RAIDZ3)
450				return (SET_ERROR(ENOTSUP));
451		} else {
452			/*
453			 * We require the parity to be specified for SPAs that
454			 * support multiple parity levels.
455			 */
456			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
457				return (SET_ERROR(EINVAL));
458			/*
459			 * Otherwise, we default to 1 parity device for RAID-Z.
460			 */
461			nparity = 1;
462		}
463	} else {
464		nparity = 0;
465	}
466	ASSERT(nparity != -1ULL);
467
468	vd = vdev_alloc_common(spa, id, guid, ops);
469
470	vd->vdev_islog = islog;
471	vd->vdev_nparity = nparity;
472
473	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
474		vd->vdev_path = spa_strdup(vd->vdev_path);
475	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
476		vd->vdev_devid = spa_strdup(vd->vdev_devid);
477	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
478	    &vd->vdev_physpath) == 0)
479		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
480	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
481		vd->vdev_fru = spa_strdup(vd->vdev_fru);
482
483	/*
484	 * Set the whole_disk property.  If it's not specified, leave the value
485	 * as -1.
486	 */
487	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
488	    &vd->vdev_wholedisk) != 0)
489		vd->vdev_wholedisk = -1ULL;
490
491	/*
492	 * Look for the 'not present' flag.  This will only be set if the device
493	 * was not present at the time of import.
494	 */
495	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
496	    &vd->vdev_not_present);
497
498	/*
499	 * Get the alignment requirement.
500	 */
501	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
502
503	/*
504	 * Retrieve the vdev creation time.
505	 */
506	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
507	    &vd->vdev_crtxg);
508
509	/*
510	 * If we're a top-level vdev, try to load the allocation parameters.
511	 */
512	if (parent && !parent->vdev_parent &&
513	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
514		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
515		    &vd->vdev_ms_array);
516		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
517		    &vd->vdev_ms_shift);
518		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
519		    &vd->vdev_asize);
520		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
521		    &vd->vdev_removing);
522		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
523		    &vd->vdev_top_zap);
524	} else {
525		ASSERT0(vd->vdev_top_zap);
526	}
527
528	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
529		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
530		    alloctype == VDEV_ALLOC_ADD ||
531		    alloctype == VDEV_ALLOC_SPLIT ||
532		    alloctype == VDEV_ALLOC_ROOTPOOL);
533		vd->vdev_mg = metaslab_group_create(islog ?
534		    spa_log_class(spa) : spa_normal_class(spa), vd);
535	}
536
537	if (vd->vdev_ops->vdev_op_leaf &&
538	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
539		(void) nvlist_lookup_uint64(nv,
540		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
541	} else {
542		ASSERT0(vd->vdev_leaf_zap);
543	}
544
545	/*
546	 * If we're a leaf vdev, try to load the DTL object and other state.
547	 */
548
549	if (vd->vdev_ops->vdev_op_leaf &&
550	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
551	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
552		if (alloctype == VDEV_ALLOC_LOAD) {
553			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
554			    &vd->vdev_dtl_object);
555			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
556			    &vd->vdev_unspare);
557		}
558
559		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
560			uint64_t spare = 0;
561
562			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
563			    &spare) == 0 && spare)
564				spa_spare_add(vd);
565		}
566
567		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
568		    &vd->vdev_offline);
569
570		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
571		    &vd->vdev_resilver_txg);
572
573		/*
574		 * When importing a pool, we want to ignore the persistent fault
575		 * state, as the diagnosis made on another system may not be
576		 * valid in the current context.  Local vdevs will
577		 * remain in the faulted state.
578		 */
579		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
580			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
581			    &vd->vdev_faulted);
582			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
583			    &vd->vdev_degraded);
584			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
585			    &vd->vdev_removed);
586
587			if (vd->vdev_faulted || vd->vdev_degraded) {
588				char *aux;
589
590				vd->vdev_label_aux =
591				    VDEV_AUX_ERR_EXCEEDED;
592				if (nvlist_lookup_string(nv,
593				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
594				    strcmp(aux, "external") == 0)
595					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
596			}
597		}
598	}
599
600	/*
601	 * Add ourselves to the parent's list of children.
602	 */
603	vdev_add_child(parent, vd);
604
605	*vdp = vd;
606
607	return (0);
608}
609
610void
611vdev_free(vdev_t *vd)
612{
613	spa_t *spa = vd->vdev_spa;
614
615	/*
616	 * vdev_free() implies closing the vdev first.  This is simpler than
617	 * trying to ensure complicated semantics for all callers.
618	 */
619	vdev_close(vd);
620
621	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
622	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
623
624	/*
625	 * Free all children.
626	 */
627	for (int c = 0; c < vd->vdev_children; c++)
628		vdev_free(vd->vdev_child[c]);
629
630	ASSERT(vd->vdev_child == NULL);
631	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
632
633	/*
634	 * Discard allocation state.
635	 */
636	if (vd->vdev_mg != NULL) {
637		vdev_metaslab_fini(vd);
638		metaslab_group_destroy(vd->vdev_mg);
639	}
640
641	ASSERT0(vd->vdev_stat.vs_space);
642	ASSERT0(vd->vdev_stat.vs_dspace);
643	ASSERT0(vd->vdev_stat.vs_alloc);
644
645	/*
646	 * Remove this vdev from its parent's child list.
647	 */
648	vdev_remove_child(vd->vdev_parent, vd);
649
650	ASSERT(vd->vdev_parent == NULL);
651
652	/*
653	 * Clean up vdev structure.
654	 */
655	vdev_queue_fini(vd);
656	vdev_cache_fini(vd);
657
658	if (vd->vdev_path)
659		spa_strfree(vd->vdev_path);
660	if (vd->vdev_devid)
661		spa_strfree(vd->vdev_devid);
662	if (vd->vdev_physpath)
663		spa_strfree(vd->vdev_physpath);
664	if (vd->vdev_fru)
665		spa_strfree(vd->vdev_fru);
666
667	if (vd->vdev_isspare)
668		spa_spare_remove(vd);
669	if (vd->vdev_isl2cache)
670		spa_l2cache_remove(vd);
671
672	txg_list_destroy(&vd->vdev_ms_list);
673	txg_list_destroy(&vd->vdev_dtl_list);
674
675	mutex_enter(&vd->vdev_dtl_lock);
676	space_map_close(vd->vdev_dtl_sm);
677	for (int t = 0; t < DTL_TYPES; t++) {
678		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
679		range_tree_destroy(vd->vdev_dtl[t]);
680	}
681	mutex_exit(&vd->vdev_dtl_lock);
682
683	mutex_destroy(&vd->vdev_queue_lock);
684	mutex_destroy(&vd->vdev_dtl_lock);
685	mutex_destroy(&vd->vdev_stat_lock);
686	mutex_destroy(&vd->vdev_probe_lock);
687
688	if (vd == spa->spa_root_vdev)
689		spa->spa_root_vdev = NULL;
690
691	kmem_free(vd, sizeof (vdev_t));
692}
693
694/*
695 * Transfer top-level vdev state from svd to tvd.
696 */
697static void
698vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
699{
700	spa_t *spa = svd->vdev_spa;
701	metaslab_t *msp;
702	vdev_t *vd;
703	int t;
704
705	ASSERT(tvd == tvd->vdev_top);
706
707	tvd->vdev_ms_array = svd->vdev_ms_array;
708	tvd->vdev_ms_shift = svd->vdev_ms_shift;
709	tvd->vdev_ms_count = svd->vdev_ms_count;
710	tvd->vdev_top_zap = svd->vdev_top_zap;
711
712	svd->vdev_ms_array = 0;
713	svd->vdev_ms_shift = 0;
714	svd->vdev_ms_count = 0;
715	svd->vdev_top_zap = 0;
716
717	if (tvd->vdev_mg)
718		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
719	tvd->vdev_mg = svd->vdev_mg;
720	tvd->vdev_ms = svd->vdev_ms;
721
722	svd->vdev_mg = NULL;
723	svd->vdev_ms = NULL;
724
725	if (tvd->vdev_mg != NULL)
726		tvd->vdev_mg->mg_vd = tvd;
727
728	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
729	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
730	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
731
732	svd->vdev_stat.vs_alloc = 0;
733	svd->vdev_stat.vs_space = 0;
734	svd->vdev_stat.vs_dspace = 0;
735
736	for (t = 0; t < TXG_SIZE; t++) {
737		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
738			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
739		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
740			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
741		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
742			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
743	}
744
745	if (list_link_active(&svd->vdev_config_dirty_node)) {
746		vdev_config_clean(svd);
747		vdev_config_dirty(tvd);
748	}
749
750	if (list_link_active(&svd->vdev_state_dirty_node)) {
751		vdev_state_clean(svd);
752		vdev_state_dirty(tvd);
753	}
754
755	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
756	svd->vdev_deflate_ratio = 0;
757
758	tvd->vdev_islog = svd->vdev_islog;
759	svd->vdev_islog = 0;
760}
761
762static void
763vdev_top_update(vdev_t *tvd, vdev_t *vd)
764{
765	if (vd == NULL)
766		return;
767
768	vd->vdev_top = tvd;
769
770	for (int c = 0; c < vd->vdev_children; c++)
771		vdev_top_update(tvd, vd->vdev_child[c]);
772}
773
774/*
775 * Add a mirror/replacing vdev above an existing vdev.
776 */
777vdev_t *
778vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
779{
780	spa_t *spa = cvd->vdev_spa;
781	vdev_t *pvd = cvd->vdev_parent;
782	vdev_t *mvd;
783
784	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
785
786	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
787
788	mvd->vdev_asize = cvd->vdev_asize;
789	mvd->vdev_min_asize = cvd->vdev_min_asize;
790	mvd->vdev_max_asize = cvd->vdev_max_asize;
791	mvd->vdev_ashift = cvd->vdev_ashift;
792	mvd->vdev_state = cvd->vdev_state;
793	mvd->vdev_crtxg = cvd->vdev_crtxg;
794
795	vdev_remove_child(pvd, cvd);
796	vdev_add_child(pvd, mvd);
797	cvd->vdev_id = mvd->vdev_children;
798	vdev_add_child(mvd, cvd);
799	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
800
801	if (mvd == mvd->vdev_top)
802		vdev_top_transfer(cvd, mvd);
803
804	return (mvd);
805}
806
807/*
808 * Remove a 1-way mirror/replacing vdev from the tree.
809 */
810void
811vdev_remove_parent(vdev_t *cvd)
812{
813	vdev_t *mvd = cvd->vdev_parent;
814	vdev_t *pvd = mvd->vdev_parent;
815
816	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
817
818	ASSERT(mvd->vdev_children == 1);
819	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
820	    mvd->vdev_ops == &vdev_replacing_ops ||
821	    mvd->vdev_ops == &vdev_spare_ops);
822	cvd->vdev_ashift = mvd->vdev_ashift;
823
824	vdev_remove_child(mvd, cvd);
825	vdev_remove_child(pvd, mvd);
826
827	/*
828	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
829	 * Otherwise, we could have detached an offline device, and when we
830	 * go to import the pool we'll think we have two top-level vdevs,
831	 * instead of a different version of the same top-level vdev.
832	 */
833	if (mvd->vdev_top == mvd) {
834		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
835		cvd->vdev_orig_guid = cvd->vdev_guid;
836		cvd->vdev_guid += guid_delta;
837		cvd->vdev_guid_sum += guid_delta;
838	}
839	cvd->vdev_id = mvd->vdev_id;
840	vdev_add_child(pvd, cvd);
841	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
842
843	if (cvd == cvd->vdev_top)
844		vdev_top_transfer(mvd, cvd);
845
846	ASSERT(mvd->vdev_children == 0);
847	vdev_free(mvd);
848}
849
850int
851vdev_metaslab_init(vdev_t *vd, uint64_t txg)
852{
853	spa_t *spa = vd->vdev_spa;
854	objset_t *mos = spa->spa_meta_objset;
855	uint64_t m;
856	uint64_t oldc = vd->vdev_ms_count;
857	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
858	metaslab_t **mspp;
859	int error;
860
861	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
862
863	/*
864	 * This vdev is not being allocated from yet or is a hole.
865	 */
866	if (vd->vdev_ms_shift == 0)
867		return (0);
868
869	ASSERT(!vd->vdev_ishole);
870
871	/*
872	 * Compute the raidz-deflation ratio.  Note, we hard-code
873	 * in 128k (1 << 17) because it is the "typical" blocksize.
874	 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
875	 * otherwise it would inconsistently account for existing bp's.
876	 */
877	vd->vdev_deflate_ratio = (1 << 17) /
878	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
879
880	ASSERT(oldc <= newc);
881
882	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
883
884	if (oldc != 0) {
885		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
886		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
887	}
888
889	vd->vdev_ms = mspp;
890	vd->vdev_ms_count = newc;
891
892	for (m = oldc; m < newc; m++) {
893		uint64_t object = 0;
894
895		if (txg == 0) {
896			error = dmu_read(mos, vd->vdev_ms_array,
897			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
898			    DMU_READ_PREFETCH);
899			if (error)
900				return (error);
901		}
902
903		error = metaslab_init(vd->vdev_mg, m, object, txg,
904		    &(vd->vdev_ms[m]));
905		if (error)
906			return (error);
907	}
908
909	if (txg == 0)
910		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
911
912	/*
913	 * If the vdev is being removed we don't activate
914	 * the metaslabs since we want to ensure that no new
915	 * allocations are performed on this device.
916	 */
917	if (oldc == 0 && !vd->vdev_removing)
918		metaslab_group_activate(vd->vdev_mg);
919
920	if (txg == 0)
921		spa_config_exit(spa, SCL_ALLOC, FTAG);
922
923	return (0);
924}
925
926void
927vdev_metaslab_fini(vdev_t *vd)
928{
929	uint64_t m;
930	uint64_t count = vd->vdev_ms_count;
931
932	if (vd->vdev_ms != NULL) {
933		metaslab_group_passivate(vd->vdev_mg);
934		for (m = 0; m < count; m++) {
935			metaslab_t *msp = vd->vdev_ms[m];
936
937			if (msp != NULL)
938				metaslab_fini(msp);
939		}
940		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
941		vd->vdev_ms = NULL;
942	}
943}
944
945typedef struct vdev_probe_stats {
946	boolean_t	vps_readable;
947	boolean_t	vps_writeable;
948	int		vps_flags;
949} vdev_probe_stats_t;
950
951static void
952vdev_probe_done(zio_t *zio)
953{
954	spa_t *spa = zio->io_spa;
955	vdev_t *vd = zio->io_vd;
956	vdev_probe_stats_t *vps = zio->io_private;
957
958	ASSERT(vd->vdev_probe_zio != NULL);
959
960	if (zio->io_type == ZIO_TYPE_READ) {
961		if (zio->io_error == 0)
962			vps->vps_readable = 1;
963		if (zio->io_error == 0 && spa_writeable(spa)) {
964			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
965			    zio->io_offset, zio->io_size, zio->io_abd,
966			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
967			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
968		} else {
969			abd_free(zio->io_abd);
970		}
971	} else if (zio->io_type == ZIO_TYPE_WRITE) {
972		if (zio->io_error == 0)
973			vps->vps_writeable = 1;
974		abd_free(zio->io_abd);
975	} else if (zio->io_type == ZIO_TYPE_NULL) {
976		zio_t *pio;
977
978		vd->vdev_cant_read |= !vps->vps_readable;
979		vd->vdev_cant_write |= !vps->vps_writeable;
980
981		if (vdev_readable(vd) &&
982		    (vdev_writeable(vd) || !spa_writeable(spa))) {
983			zio->io_error = 0;
984		} else {
985			ASSERT(zio->io_error != 0);
986			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
987			    spa, vd, NULL, 0, 0);
988			zio->io_error = SET_ERROR(ENXIO);
989		}
990
991		mutex_enter(&vd->vdev_probe_lock);
992		ASSERT(vd->vdev_probe_zio == zio);
993		vd->vdev_probe_zio = NULL;
994		mutex_exit(&vd->vdev_probe_lock);
995
996		zio_link_t *zl = NULL;
997		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
998			if (!vdev_accessible(vd, pio))
999				pio->io_error = SET_ERROR(ENXIO);
1000
1001		kmem_free(vps, sizeof (*vps));
1002	}
1003}
1004
1005/*
1006 * Determine whether this device is accessible.
1007 *
1008 * Read and write to several known locations: the pad regions of each
1009 * vdev label but the first, which we leave alone in case it contains
1010 * a VTOC.
1011 */
1012zio_t *
1013vdev_probe(vdev_t *vd, zio_t *zio)
1014{
1015	spa_t *spa = vd->vdev_spa;
1016	vdev_probe_stats_t *vps = NULL;
1017	zio_t *pio;
1018
1019	ASSERT(vd->vdev_ops->vdev_op_leaf);
1020
1021	/*
1022	 * Don't probe the probe.
1023	 */
1024	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1025		return (NULL);
1026
1027	/*
1028	 * To prevent 'probe storms' when a device fails, we create
1029	 * just one probe i/o at a time.  All zios that want to probe
1030	 * this vdev will become parents of the probe io.
1031	 */
1032	mutex_enter(&vd->vdev_probe_lock);
1033
1034	if ((pio = vd->vdev_probe_zio) == NULL) {
1035		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1036
1037		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1038		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1039		    ZIO_FLAG_TRYHARD;
1040
1041		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1042			/*
1043			 * vdev_cant_read and vdev_cant_write can only
1044			 * transition from TRUE to FALSE when we have the
1045			 * SCL_ZIO lock as writer; otherwise they can only
1046			 * transition from FALSE to TRUE.  This ensures that
1047			 * any zio looking at these values can assume that
1048			 * failures persist for the life of the I/O.  That's
1049			 * important because when a device has intermittent
1050			 * connectivity problems, we want to ensure that
1051			 * they're ascribed to the device (ENXIO) and not
1052			 * the zio (EIO).
1053			 *
1054			 * Since we hold SCL_ZIO as writer here, clear both
1055			 * values so the probe can reevaluate from first
1056			 * principles.
1057			 */
1058			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1059			vd->vdev_cant_read = B_FALSE;
1060			vd->vdev_cant_write = B_FALSE;
1061		}
1062
1063		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1064		    vdev_probe_done, vps,
1065		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1066
1067		/*
1068		 * We can't change the vdev state in this context, so we
1069		 * kick off an async task to do it on our behalf.
1070		 */
1071		if (zio != NULL) {
1072			vd->vdev_probe_wanted = B_TRUE;
1073			spa_async_request(spa, SPA_ASYNC_PROBE);
1074		}
1075	}
1076
1077	if (zio != NULL)
1078		zio_add_child(zio, pio);
1079
1080	mutex_exit(&vd->vdev_probe_lock);
1081
1082	if (vps == NULL) {
1083		ASSERT(zio != NULL);
1084		return (NULL);
1085	}
1086
1087	for (int l = 1; l < VDEV_LABELS; l++) {
1088		zio_nowait(zio_read_phys(pio, vd,
1089		    vdev_label_offset(vd->vdev_psize, l,
1090		    offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE,
1091		    abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1092		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1093		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1094	}
1095
1096	if (zio == NULL)
1097		return (pio);
1098
1099	zio_nowait(pio);
1100	return (NULL);
1101}
1102
1103static void
1104vdev_open_child(void *arg)
1105{
1106	vdev_t *vd = arg;
1107
1108	vd->vdev_open_thread = curthread;
1109	vd->vdev_open_error = vdev_open(vd);
1110	vd->vdev_open_thread = NULL;
1111}
1112
1113boolean_t
1114vdev_uses_zvols(vdev_t *vd)
1115{
1116	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1117	    strlen(ZVOL_DIR)) == 0)
1118		return (B_TRUE);
1119	for (int c = 0; c < vd->vdev_children; c++)
1120		if (vdev_uses_zvols(vd->vdev_child[c]))
1121			return (B_TRUE);
1122	return (B_FALSE);
1123}
1124
1125void
1126vdev_open_children(vdev_t *vd)
1127{
1128	taskq_t *tq;
1129	int children = vd->vdev_children;
1130
1131	/*
1132	 * in order to handle pools on top of zvols, do the opens
1133	 * in a single thread so that the same thread holds the
1134	 * spa_namespace_lock
1135	 */
1136	if (vdev_uses_zvols(vd)) {
1137		for (int c = 0; c < children; c++)
1138			vd->vdev_child[c]->vdev_open_error =
1139			    vdev_open(vd->vdev_child[c]);
1140		return;
1141	}
1142	tq = taskq_create("vdev_open", children, minclsyspri,
1143	    children, children, TASKQ_PREPOPULATE);
1144
1145	for (int c = 0; c < children; c++)
1146		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1147		    TQ_SLEEP) != NULL);
1148
1149	taskq_destroy(tq);
1150}
1151
1152/*
1153 * Prepare a virtual device for access.
1154 */
1155int
1156vdev_open(vdev_t *vd)
1157{
1158	spa_t *spa = vd->vdev_spa;
1159	int error;
1160	uint64_t osize = 0;
1161	uint64_t max_osize = 0;
1162	uint64_t asize, max_asize, psize;
1163	uint64_t ashift = 0;
1164
1165	ASSERT(vd->vdev_open_thread == curthread ||
1166	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1167	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1168	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1169	    vd->vdev_state == VDEV_STATE_OFFLINE);
1170
1171	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1172	vd->vdev_cant_read = B_FALSE;
1173	vd->vdev_cant_write = B_FALSE;
1174	vd->vdev_min_asize = vdev_get_min_asize(vd);
1175
1176	/*
1177	 * If this vdev is not removed, check its fault status.  If it's
1178	 * faulted, bail out of the open.
1179	 */
1180	if (!vd->vdev_removed && vd->vdev_faulted) {
1181		ASSERT(vd->vdev_children == 0);
1182		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1183		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1184		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1185		    vd->vdev_label_aux);
1186		return (SET_ERROR(ENXIO));
1187	} else if (vd->vdev_offline) {
1188		ASSERT(vd->vdev_children == 0);
1189		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1190		return (SET_ERROR(ENXIO));
1191	}
1192
1193	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1194
1195	/*
1196	 * Reset the vdev_reopening flag so that we actually close
1197	 * the vdev on error.
1198	 */
1199	vd->vdev_reopening = B_FALSE;
1200	if (zio_injection_enabled && error == 0)
1201		error = zio_handle_device_injection(vd, NULL, ENXIO);
1202
1203	if (error) {
1204		if (vd->vdev_removed &&
1205		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1206			vd->vdev_removed = B_FALSE;
1207
1208		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1209		    vd->vdev_stat.vs_aux);
1210		return (error);
1211	}
1212
1213	vd->vdev_removed = B_FALSE;
1214
1215	/*
1216	 * Recheck the faulted flag now that we have confirmed that
1217	 * the vdev is accessible.  If we're faulted, bail.
1218	 */
1219	if (vd->vdev_faulted) {
1220		ASSERT(vd->vdev_children == 0);
1221		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1222		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1223		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1224		    vd->vdev_label_aux);
1225		return (SET_ERROR(ENXIO));
1226	}
1227
1228	if (vd->vdev_degraded) {
1229		ASSERT(vd->vdev_children == 0);
1230		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1231		    VDEV_AUX_ERR_EXCEEDED);
1232	} else {
1233		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1234	}
1235
1236	/*
1237	 * For hole or missing vdevs we just return success.
1238	 */
1239	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1240		return (0);
1241
1242	for (int c = 0; c < vd->vdev_children; c++) {
1243		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1244			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1245			    VDEV_AUX_NONE);
1246			break;
1247		}
1248	}
1249
1250	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1251	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1252
1253	if (vd->vdev_children == 0) {
1254		if (osize < SPA_MINDEVSIZE) {
1255			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1256			    VDEV_AUX_TOO_SMALL);
1257			return (SET_ERROR(EOVERFLOW));
1258		}
1259		psize = osize;
1260		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1261		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1262		    VDEV_LABEL_END_SIZE);
1263	} else {
1264		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1265		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1266			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1267			    VDEV_AUX_TOO_SMALL);
1268			return (SET_ERROR(EOVERFLOW));
1269		}
1270		psize = 0;
1271		asize = osize;
1272		max_asize = max_osize;
1273	}
1274
1275	vd->vdev_psize = psize;
1276
1277	/*
1278	 * Make sure the allocatable size hasn't shrunk too much.
1279	 */
1280	if (asize < vd->vdev_min_asize) {
1281		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1282		    VDEV_AUX_BAD_LABEL);
1283		return (SET_ERROR(EINVAL));
1284	}
1285
1286	if (vd->vdev_asize == 0) {
1287		/*
1288		 * This is the first-ever open, so use the computed values.
1289		 * For testing purposes, a higher ashift can be requested.
1290		 */
1291		vd->vdev_asize = asize;
1292		vd->vdev_max_asize = max_asize;
1293		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1294	} else {
1295		/*
1296		 * Detect if the alignment requirement has increased.
1297		 * We don't want to make the pool unavailable, just
1298		 * issue a warning instead.
1299		 */
1300		if (ashift > vd->vdev_top->vdev_ashift &&
1301		    vd->vdev_ops->vdev_op_leaf) {
1302			cmn_err(CE_WARN,
1303			    "Disk, '%s', has a block alignment that is "
1304			    "larger than the pool's alignment\n",
1305			    vd->vdev_path);
1306		}
1307		vd->vdev_max_asize = max_asize;
1308	}
1309
1310	/*
1311	 * If all children are healthy we update asize if either:
1312	 * The asize has increased, due to a device expansion caused by dynamic
1313	 * LUN growth or vdev replacement, and automatic expansion is enabled;
1314	 * making the additional space available.
1315	 *
1316	 * The asize has decreased, due to a device shrink usually caused by a
1317	 * vdev replace with a smaller device. This ensures that calculations
1318	 * based of max_asize and asize e.g. esize are always valid. It's safe
1319	 * to do this as we've already validated that asize is greater than
1320	 * vdev_min_asize.
1321	 */
1322	if (vd->vdev_state == VDEV_STATE_HEALTHY &&
1323	    ((asize > vd->vdev_asize &&
1324	    (vd->vdev_expanding || spa->spa_autoexpand)) ||
1325	    (asize < vd->vdev_asize)))
1326		vd->vdev_asize = asize;
1327
1328	vdev_set_min_asize(vd);
1329
1330	/*
1331	 * Ensure we can issue some IO before declaring the
1332	 * vdev open for business.
1333	 */
1334	if (vd->vdev_ops->vdev_op_leaf &&
1335	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1336		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1337		    VDEV_AUX_ERR_EXCEEDED);
1338		return (error);
1339	}
1340
1341	/*
1342	 * Track the min and max ashift values for normal data devices.
1343	 */
1344	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1345	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1346		if (vd->vdev_ashift > spa->spa_max_ashift)
1347			spa->spa_max_ashift = vd->vdev_ashift;
1348		if (vd->vdev_ashift < spa->spa_min_ashift)
1349			spa->spa_min_ashift = vd->vdev_ashift;
1350	}
1351
1352	/*
1353	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1354	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1355	 * since this would just restart the scrub we are already doing.
1356	 */
1357	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1358	    vdev_resilver_needed(vd, NULL, NULL))
1359		spa_async_request(spa, SPA_ASYNC_RESILVER);
1360
1361	return (0);
1362}
1363
1364/*
1365 * Called once the vdevs are all opened, this routine validates the label
1366 * contents.  This needs to be done before vdev_load() so that we don't
1367 * inadvertently do repair I/Os to the wrong device.
1368 *
1369 * If 'strict' is false ignore the spa guid check. This is necessary because
1370 * if the machine crashed during a re-guid the new guid might have been written
1371 * to all of the vdev labels, but not the cached config. The strict check
1372 * will be performed when the pool is opened again using the mos config.
1373 *
1374 * This function will only return failure if one of the vdevs indicates that it
1375 * has since been destroyed or exported.  This is only possible if
1376 * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1377 * will be updated but the function will return 0.
1378 */
1379int
1380vdev_validate(vdev_t *vd, boolean_t strict)
1381{
1382	spa_t *spa = vd->vdev_spa;
1383	nvlist_t *label;
1384	uint64_t guid = 0, top_guid;
1385	uint64_t state;
1386
1387	for (int c = 0; c < vd->vdev_children; c++)
1388		if (vdev_validate(vd->vdev_child[c], strict) != 0)
1389			return (SET_ERROR(EBADF));
1390
1391	/*
1392	 * If the device has already failed, or was marked offline, don't do
1393	 * any further validation.  Otherwise, label I/O will fail and we will
1394	 * overwrite the previous state.
1395	 */
1396	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1397		uint64_t aux_guid = 0;
1398		nvlist_t *nvl;
1399		uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1400		    spa_last_synced_txg(spa) : -1ULL;
1401
1402		if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1403			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1404			    VDEV_AUX_BAD_LABEL);
1405			return (0);
1406		}
1407
1408		/*
1409		 * Determine if this vdev has been split off into another
1410		 * pool.  If so, then refuse to open it.
1411		 */
1412		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1413		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1414			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1415			    VDEV_AUX_SPLIT_POOL);
1416			nvlist_free(label);
1417			return (0);
1418		}
1419
1420		if (strict && (nvlist_lookup_uint64(label,
1421		    ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1422		    guid != spa_guid(spa))) {
1423			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1424			    VDEV_AUX_CORRUPT_DATA);
1425			nvlist_free(label);
1426			return (0);
1427		}
1428
1429		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1430		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1431		    &aux_guid) != 0)
1432			aux_guid = 0;
1433
1434		/*
1435		 * If this vdev just became a top-level vdev because its
1436		 * sibling was detached, it will have adopted the parent's
1437		 * vdev guid -- but the label may or may not be on disk yet.
1438		 * Fortunately, either version of the label will have the
1439		 * same top guid, so if we're a top-level vdev, we can
1440		 * safely compare to that instead.
1441		 *
1442		 * If we split this vdev off instead, then we also check the
1443		 * original pool's guid.  We don't want to consider the vdev
1444		 * corrupt if it is partway through a split operation.
1445		 */
1446		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1447		    &guid) != 0 ||
1448		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1449		    &top_guid) != 0 ||
1450		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1451		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1452			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1453			    VDEV_AUX_CORRUPT_DATA);
1454			nvlist_free(label);
1455			return (0);
1456		}
1457
1458		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1459		    &state) != 0) {
1460			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1461			    VDEV_AUX_CORRUPT_DATA);
1462			nvlist_free(label);
1463			return (0);
1464		}
1465
1466		nvlist_free(label);
1467
1468		/*
1469		 * If this is a verbatim import, no need to check the
1470		 * state of the pool.
1471		 */
1472		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1473		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1474		    state != POOL_STATE_ACTIVE)
1475			return (SET_ERROR(EBADF));
1476
1477		/*
1478		 * If we were able to open and validate a vdev that was
1479		 * previously marked permanently unavailable, clear that state
1480		 * now.
1481		 */
1482		if (vd->vdev_not_present)
1483			vd->vdev_not_present = 0;
1484	}
1485
1486	return (0);
1487}
1488
1489/*
1490 * Close a virtual device.
1491 */
1492void
1493vdev_close(vdev_t *vd)
1494{
1495	spa_t *spa = vd->vdev_spa;
1496	vdev_t *pvd = vd->vdev_parent;
1497
1498	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1499
1500	/*
1501	 * If our parent is reopening, then we are as well, unless we are
1502	 * going offline.
1503	 */
1504	if (pvd != NULL && pvd->vdev_reopening)
1505		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1506
1507	vd->vdev_ops->vdev_op_close(vd);
1508
1509	vdev_cache_purge(vd);
1510
1511	/*
1512	 * We record the previous state before we close it, so that if we are
1513	 * doing a reopen(), we don't generate FMA ereports if we notice that
1514	 * it's still faulted.
1515	 */
1516	vd->vdev_prevstate = vd->vdev_state;
1517
1518	if (vd->vdev_offline)
1519		vd->vdev_state = VDEV_STATE_OFFLINE;
1520	else
1521		vd->vdev_state = VDEV_STATE_CLOSED;
1522	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1523}
1524
1525void
1526vdev_hold(vdev_t *vd)
1527{
1528	spa_t *spa = vd->vdev_spa;
1529
1530	ASSERT(spa_is_root(spa));
1531	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1532		return;
1533
1534	for (int c = 0; c < vd->vdev_children; c++)
1535		vdev_hold(vd->vdev_child[c]);
1536
1537	if (vd->vdev_ops->vdev_op_leaf)
1538		vd->vdev_ops->vdev_op_hold(vd);
1539}
1540
1541void
1542vdev_rele(vdev_t *vd)
1543{
1544	spa_t *spa = vd->vdev_spa;
1545
1546	ASSERT(spa_is_root(spa));
1547	for (int c = 0; c < vd->vdev_children; c++)
1548		vdev_rele(vd->vdev_child[c]);
1549
1550	if (vd->vdev_ops->vdev_op_leaf)
1551		vd->vdev_ops->vdev_op_rele(vd);
1552}
1553
1554/*
1555 * Reopen all interior vdevs and any unopened leaves.  We don't actually
1556 * reopen leaf vdevs which had previously been opened as they might deadlock
1557 * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1558 * If the leaf has never been opened then open it, as usual.
1559 */
1560void
1561vdev_reopen(vdev_t *vd)
1562{
1563	spa_t *spa = vd->vdev_spa;
1564
1565	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1566
1567	/* set the reopening flag unless we're taking the vdev offline */
1568	vd->vdev_reopening = !vd->vdev_offline;
1569	vdev_close(vd);
1570	(void) vdev_open(vd);
1571
1572	/*
1573	 * Call vdev_validate() here to make sure we have the same device.
1574	 * Otherwise, a device with an invalid label could be successfully
1575	 * opened in response to vdev_reopen().
1576	 */
1577	if (vd->vdev_aux) {
1578		(void) vdev_validate_aux(vd);
1579		if (vdev_readable(vd) && vdev_writeable(vd) &&
1580		    vd->vdev_aux == &spa->spa_l2cache &&
1581		    !l2arc_vdev_present(vd))
1582			l2arc_add_vdev(spa, vd);
1583	} else {
1584		(void) vdev_validate(vd, B_TRUE);
1585	}
1586
1587	/*
1588	 * Reassess parent vdev's health.
1589	 */
1590	vdev_propagate_state(vd);
1591}
1592
1593int
1594vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1595{
1596	int error;
1597
1598	/*
1599	 * Normally, partial opens (e.g. of a mirror) are allowed.
1600	 * For a create, however, we want to fail the request if
1601	 * there are any components we can't open.
1602	 */
1603	error = vdev_open(vd);
1604
1605	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1606		vdev_close(vd);
1607		return (error ? error : ENXIO);
1608	}
1609
1610	/*
1611	 * Recursively load DTLs and initialize all labels.
1612	 */
1613	if ((error = vdev_dtl_load(vd)) != 0 ||
1614	    (error = vdev_label_init(vd, txg, isreplacing ?
1615	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1616		vdev_close(vd);
1617		return (error);
1618	}
1619
1620	return (0);
1621}
1622
1623void
1624vdev_metaslab_set_size(vdev_t *vd)
1625{
1626	/*
1627	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1628	 */
1629	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1630	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1631}
1632
1633void
1634vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1635{
1636	ASSERT(vd == vd->vdev_top);
1637	ASSERT(!vd->vdev_ishole);
1638	ASSERT(ISP2(flags));
1639	ASSERT(spa_writeable(vd->vdev_spa));
1640
1641	if (flags & VDD_METASLAB)
1642		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1643
1644	if (flags & VDD_DTL)
1645		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1646
1647	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1648}
1649
1650void
1651vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1652{
1653	for (int c = 0; c < vd->vdev_children; c++)
1654		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1655
1656	if (vd->vdev_ops->vdev_op_leaf)
1657		vdev_dirty(vd->vdev_top, flags, vd, txg);
1658}
1659
1660/*
1661 * DTLs.
1662 *
1663 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1664 * the vdev has less than perfect replication.  There are four kinds of DTL:
1665 *
1666 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1667 *
1668 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1669 *
1670 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1671 *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1672 *	txgs that was scrubbed.
1673 *
1674 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1675 *	persistent errors or just some device being offline.
1676 *	Unlike the other three, the DTL_OUTAGE map is not generally
1677 *	maintained; it's only computed when needed, typically to
1678 *	determine whether a device can be detached.
1679 *
1680 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1681 * either has the data or it doesn't.
1682 *
1683 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1684 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1685 * if any child is less than fully replicated, then so is its parent.
1686 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1687 * comprising only those txgs which appear in 'maxfaults' or more children;
1688 * those are the txgs we don't have enough replication to read.  For example,
1689 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1690 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1691 * two child DTL_MISSING maps.
1692 *
1693 * It should be clear from the above that to compute the DTLs and outage maps
1694 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1695 * Therefore, that is all we keep on disk.  When loading the pool, or after
1696 * a configuration change, we generate all other DTLs from first principles.
1697 */
1698void
1699vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1700{
1701	range_tree_t *rt = vd->vdev_dtl[t];
1702
1703	ASSERT(t < DTL_TYPES);
1704	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1705	ASSERT(spa_writeable(vd->vdev_spa));
1706
1707	mutex_enter(rt->rt_lock);
1708	if (!range_tree_contains(rt, txg, size))
1709		range_tree_add(rt, txg, size);
1710	mutex_exit(rt->rt_lock);
1711}
1712
1713boolean_t
1714vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1715{
1716	range_tree_t *rt = vd->vdev_dtl[t];
1717	boolean_t dirty = B_FALSE;
1718
1719	ASSERT(t < DTL_TYPES);
1720	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1721
1722	mutex_enter(rt->rt_lock);
1723	if (range_tree_space(rt) != 0)
1724		dirty = range_tree_contains(rt, txg, size);
1725	mutex_exit(rt->rt_lock);
1726
1727	return (dirty);
1728}
1729
1730boolean_t
1731vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1732{
1733	range_tree_t *rt = vd->vdev_dtl[t];
1734	boolean_t empty;
1735
1736	mutex_enter(rt->rt_lock);
1737	empty = (range_tree_space(rt) == 0);
1738	mutex_exit(rt->rt_lock);
1739
1740	return (empty);
1741}
1742
1743/*
1744 * Returns the lowest txg in the DTL range.
1745 */
1746static uint64_t
1747vdev_dtl_min(vdev_t *vd)
1748{
1749	range_seg_t *rs;
1750
1751	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1752	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1753	ASSERT0(vd->vdev_children);
1754
1755	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1756	return (rs->rs_start - 1);
1757}
1758
1759/*
1760 * Returns the highest txg in the DTL.
1761 */
1762static uint64_t
1763vdev_dtl_max(vdev_t *vd)
1764{
1765	range_seg_t *rs;
1766
1767	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1768	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1769	ASSERT0(vd->vdev_children);
1770
1771	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1772	return (rs->rs_end);
1773}
1774
1775/*
1776 * Determine if a resilvering vdev should remove any DTL entries from
1777 * its range. If the vdev was resilvering for the entire duration of the
1778 * scan then it should excise that range from its DTLs. Otherwise, this
1779 * vdev is considered partially resilvered and should leave its DTL
1780 * entries intact. The comment in vdev_dtl_reassess() describes how we
1781 * excise the DTLs.
1782 */
1783static boolean_t
1784vdev_dtl_should_excise(vdev_t *vd)
1785{
1786	spa_t *spa = vd->vdev_spa;
1787	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1788
1789	ASSERT0(scn->scn_phys.scn_errors);
1790	ASSERT0(vd->vdev_children);
1791
1792	if (vd->vdev_state < VDEV_STATE_DEGRADED)
1793		return (B_FALSE);
1794
1795	if (vd->vdev_resilver_txg == 0 ||
1796	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1797		return (B_TRUE);
1798
1799	/*
1800	 * When a resilver is initiated the scan will assign the scn_max_txg
1801	 * value to the highest txg value that exists in all DTLs. If this
1802	 * device's max DTL is not part of this scan (i.e. it is not in
1803	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1804	 * for excision.
1805	 */
1806	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1807		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1808		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1809		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1810		return (B_TRUE);
1811	}
1812	return (B_FALSE);
1813}
1814
1815/*
1816 * Reassess DTLs after a config change or scrub completion.
1817 */
1818void
1819vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1820{
1821	spa_t *spa = vd->vdev_spa;
1822	avl_tree_t reftree;
1823	int minref;
1824
1825	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1826
1827	for (int c = 0; c < vd->vdev_children; c++)
1828		vdev_dtl_reassess(vd->vdev_child[c], txg,
1829		    scrub_txg, scrub_done);
1830
1831	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1832		return;
1833
1834	if (vd->vdev_ops->vdev_op_leaf) {
1835		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1836
1837		mutex_enter(&vd->vdev_dtl_lock);
1838
1839		/*
1840		 * If we've completed a scan cleanly then determine
1841		 * if this vdev should remove any DTLs. We only want to
1842		 * excise regions on vdevs that were available during
1843		 * the entire duration of this scan.
1844		 */
1845		if (scrub_txg != 0 &&
1846		    (spa->spa_scrub_started ||
1847		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1848		    vdev_dtl_should_excise(vd)) {
1849			/*
1850			 * We completed a scrub up to scrub_txg.  If we
1851			 * did it without rebooting, then the scrub dtl
1852			 * will be valid, so excise the old region and
1853			 * fold in the scrub dtl.  Otherwise, leave the
1854			 * dtl as-is if there was an error.
1855			 *
1856			 * There's little trick here: to excise the beginning
1857			 * of the DTL_MISSING map, we put it into a reference
1858			 * tree and then add a segment with refcnt -1 that
1859			 * covers the range [0, scrub_txg).  This means
1860			 * that each txg in that range has refcnt -1 or 0.
1861			 * We then add DTL_SCRUB with a refcnt of 2, so that
1862			 * entries in the range [0, scrub_txg) will have a
1863			 * positive refcnt -- either 1 or 2.  We then convert
1864			 * the reference tree into the new DTL_MISSING map.
1865			 */
1866			space_reftree_create(&reftree);
1867			space_reftree_add_map(&reftree,
1868			    vd->vdev_dtl[DTL_MISSING], 1);
1869			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1870			space_reftree_add_map(&reftree,
1871			    vd->vdev_dtl[DTL_SCRUB], 2);
1872			space_reftree_generate_map(&reftree,
1873			    vd->vdev_dtl[DTL_MISSING], 1);
1874			space_reftree_destroy(&reftree);
1875		}
1876		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1877		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1878		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1879		if (scrub_done)
1880			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1881		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1882		if (!vdev_readable(vd))
1883			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1884		else
1885			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1886			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1887
1888		/*
1889		 * If the vdev was resilvering and no longer has any
1890		 * DTLs then reset its resilvering flag.
1891		 */
1892		if (vd->vdev_resilver_txg != 0 &&
1893		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1894		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1895			vd->vdev_resilver_txg = 0;
1896
1897		mutex_exit(&vd->vdev_dtl_lock);
1898
1899		if (txg != 0)
1900			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1901		return;
1902	}
1903
1904	mutex_enter(&vd->vdev_dtl_lock);
1905	for (int t = 0; t < DTL_TYPES; t++) {
1906		/* account for child's outage in parent's missing map */
1907		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1908		if (t == DTL_SCRUB)
1909			continue;			/* leaf vdevs only */
1910		if (t == DTL_PARTIAL)
1911			minref = 1;			/* i.e. non-zero */
1912		else if (vd->vdev_nparity != 0)
1913			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1914		else
1915			minref = vd->vdev_children;	/* any kind of mirror */
1916		space_reftree_create(&reftree);
1917		for (int c = 0; c < vd->vdev_children; c++) {
1918			vdev_t *cvd = vd->vdev_child[c];
1919			mutex_enter(&cvd->vdev_dtl_lock);
1920			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1921			mutex_exit(&cvd->vdev_dtl_lock);
1922		}
1923		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1924		space_reftree_destroy(&reftree);
1925	}
1926	mutex_exit(&vd->vdev_dtl_lock);
1927}
1928
1929int
1930vdev_dtl_load(vdev_t *vd)
1931{
1932	spa_t *spa = vd->vdev_spa;
1933	objset_t *mos = spa->spa_meta_objset;
1934	int error = 0;
1935
1936	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1937		ASSERT(!vd->vdev_ishole);
1938
1939		error = space_map_open(&vd->vdev_dtl_sm, mos,
1940		    vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1941		if (error)
1942			return (error);
1943		ASSERT(vd->vdev_dtl_sm != NULL);
1944
1945		mutex_enter(&vd->vdev_dtl_lock);
1946
1947		/*
1948		 * Now that we've opened the space_map we need to update
1949		 * the in-core DTL.
1950		 */
1951		space_map_update(vd->vdev_dtl_sm);
1952
1953		error = space_map_load(vd->vdev_dtl_sm,
1954		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1955		mutex_exit(&vd->vdev_dtl_lock);
1956
1957		return (error);
1958	}
1959
1960	for (int c = 0; c < vd->vdev_children; c++) {
1961		error = vdev_dtl_load(vd->vdev_child[c]);
1962		if (error != 0)
1963			break;
1964	}
1965
1966	return (error);
1967}
1968
1969void
1970vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
1971{
1972	spa_t *spa = vd->vdev_spa;
1973
1974	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
1975	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1976	    zapobj, tx));
1977}
1978
1979uint64_t
1980vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
1981{
1982	spa_t *spa = vd->vdev_spa;
1983	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
1984	    DMU_OT_NONE, 0, tx);
1985
1986	ASSERT(zap != 0);
1987	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1988	    zap, tx));
1989
1990	return (zap);
1991}
1992
1993void
1994vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
1995{
1996	if (vd->vdev_ops != &vdev_hole_ops &&
1997	    vd->vdev_ops != &vdev_missing_ops &&
1998	    vd->vdev_ops != &vdev_root_ops &&
1999	    !vd->vdev_top->vdev_removing) {
2000		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2001			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2002		}
2003		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2004			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2005		}
2006	}
2007	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2008		vdev_construct_zaps(vd->vdev_child[i], tx);
2009	}
2010}
2011
2012void
2013vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2014{
2015	spa_t *spa = vd->vdev_spa;
2016	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2017	objset_t *mos = spa->spa_meta_objset;
2018	range_tree_t *rtsync;
2019	kmutex_t rtlock;
2020	dmu_tx_t *tx;
2021	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2022
2023	ASSERT(!vd->vdev_ishole);
2024	ASSERT(vd->vdev_ops->vdev_op_leaf);
2025
2026	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2027
2028	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2029		mutex_enter(&vd->vdev_dtl_lock);
2030		space_map_free(vd->vdev_dtl_sm, tx);
2031		space_map_close(vd->vdev_dtl_sm);
2032		vd->vdev_dtl_sm = NULL;
2033		mutex_exit(&vd->vdev_dtl_lock);
2034
2035		/*
2036		 * We only destroy the leaf ZAP for detached leaves or for
2037		 * removed log devices. Removed data devices handle leaf ZAP
2038		 * cleanup later, once cancellation is no longer possible.
2039		 */
2040		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2041		    vd->vdev_top->vdev_islog)) {
2042			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2043			vd->vdev_leaf_zap = 0;
2044		}
2045
2046		dmu_tx_commit(tx);
2047		return;
2048	}
2049
2050	if (vd->vdev_dtl_sm == NULL) {
2051		uint64_t new_object;
2052
2053		new_object = space_map_alloc(mos, tx);
2054		VERIFY3U(new_object, !=, 0);
2055
2056		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2057		    0, -1ULL, 0, &vd->vdev_dtl_lock));
2058		ASSERT(vd->vdev_dtl_sm != NULL);
2059	}
2060
2061	mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
2062
2063	rtsync = range_tree_create(NULL, NULL, &rtlock);
2064
2065	mutex_enter(&rtlock);
2066
2067	mutex_enter(&vd->vdev_dtl_lock);
2068	range_tree_walk(rt, range_tree_add, rtsync);
2069	mutex_exit(&vd->vdev_dtl_lock);
2070
2071	space_map_truncate(vd->vdev_dtl_sm, tx);
2072	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2073	range_tree_vacate(rtsync, NULL, NULL);
2074
2075	range_tree_destroy(rtsync);
2076
2077	mutex_exit(&rtlock);
2078	mutex_destroy(&rtlock);
2079
2080	/*
2081	 * If the object for the space map has changed then dirty
2082	 * the top level so that we update the config.
2083	 */
2084	if (object != space_map_object(vd->vdev_dtl_sm)) {
2085		zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
2086		    "new object %llu", txg, spa_name(spa), object,
2087		    space_map_object(vd->vdev_dtl_sm));
2088		vdev_config_dirty(vd->vdev_top);
2089	}
2090
2091	dmu_tx_commit(tx);
2092
2093	mutex_enter(&vd->vdev_dtl_lock);
2094	space_map_update(vd->vdev_dtl_sm);
2095	mutex_exit(&vd->vdev_dtl_lock);
2096}
2097
2098/*
2099 * Determine whether the specified vdev can be offlined/detached/removed
2100 * without losing data.
2101 */
2102boolean_t
2103vdev_dtl_required(vdev_t *vd)
2104{
2105	spa_t *spa = vd->vdev_spa;
2106	vdev_t *tvd = vd->vdev_top;
2107	uint8_t cant_read = vd->vdev_cant_read;
2108	boolean_t required;
2109
2110	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2111
2112	if (vd == spa->spa_root_vdev || vd == tvd)
2113		return (B_TRUE);
2114
2115	/*
2116	 * Temporarily mark the device as unreadable, and then determine
2117	 * whether this results in any DTL outages in the top-level vdev.
2118	 * If not, we can safely offline/detach/remove the device.
2119	 */
2120	vd->vdev_cant_read = B_TRUE;
2121	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2122	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2123	vd->vdev_cant_read = cant_read;
2124	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2125
2126	if (!required && zio_injection_enabled)
2127		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2128
2129	return (required);
2130}
2131
2132/*
2133 * Determine if resilver is needed, and if so the txg range.
2134 */
2135boolean_t
2136vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2137{
2138	boolean_t needed = B_FALSE;
2139	uint64_t thismin = UINT64_MAX;
2140	uint64_t thismax = 0;
2141
2142	if (vd->vdev_children == 0) {
2143		mutex_enter(&vd->vdev_dtl_lock);
2144		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2145		    vdev_writeable(vd)) {
2146
2147			thismin = vdev_dtl_min(vd);
2148			thismax = vdev_dtl_max(vd);
2149			needed = B_TRUE;
2150		}
2151		mutex_exit(&vd->vdev_dtl_lock);
2152	} else {
2153		for (int c = 0; c < vd->vdev_children; c++) {
2154			vdev_t *cvd = vd->vdev_child[c];
2155			uint64_t cmin, cmax;
2156
2157			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2158				thismin = MIN(thismin, cmin);
2159				thismax = MAX(thismax, cmax);
2160				needed = B_TRUE;
2161			}
2162		}
2163	}
2164
2165	if (needed && minp) {
2166		*minp = thismin;
2167		*maxp = thismax;
2168	}
2169	return (needed);
2170}
2171
2172void
2173vdev_load(vdev_t *vd)
2174{
2175	/*
2176	 * Recursively load all children.
2177	 */
2178	for (int c = 0; c < vd->vdev_children; c++)
2179		vdev_load(vd->vdev_child[c]);
2180
2181	/*
2182	 * If this is a top-level vdev, initialize its metaslabs.
2183	 */
2184	if (vd == vd->vdev_top && !vd->vdev_ishole &&
2185	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2186	    vdev_metaslab_init(vd, 0) != 0))
2187		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2188		    VDEV_AUX_CORRUPT_DATA);
2189
2190	/*
2191	 * If this is a leaf vdev, load its DTL.
2192	 */
2193	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2194		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2195		    VDEV_AUX_CORRUPT_DATA);
2196}
2197
2198/*
2199 * The special vdev case is used for hot spares and l2cache devices.  Its
2200 * sole purpose it to set the vdev state for the associated vdev.  To do this,
2201 * we make sure that we can open the underlying device, then try to read the
2202 * label, and make sure that the label is sane and that it hasn't been
2203 * repurposed to another pool.
2204 */
2205int
2206vdev_validate_aux(vdev_t *vd)
2207{
2208	nvlist_t *label;
2209	uint64_t guid, version;
2210	uint64_t state;
2211
2212	if (!vdev_readable(vd))
2213		return (0);
2214
2215	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2216		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2217		    VDEV_AUX_CORRUPT_DATA);
2218		return (-1);
2219	}
2220
2221	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2222	    !SPA_VERSION_IS_SUPPORTED(version) ||
2223	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2224	    guid != vd->vdev_guid ||
2225	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2226		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2227		    VDEV_AUX_CORRUPT_DATA);
2228		nvlist_free(label);
2229		return (-1);
2230	}
2231
2232	/*
2233	 * We don't actually check the pool state here.  If it's in fact in
2234	 * use by another pool, we update this fact on the fly when requested.
2235	 */
2236	nvlist_free(label);
2237	return (0);
2238}
2239
2240void
2241vdev_remove(vdev_t *vd, uint64_t txg)
2242{
2243	spa_t *spa = vd->vdev_spa;
2244	objset_t *mos = spa->spa_meta_objset;
2245	dmu_tx_t *tx;
2246
2247	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2248	ASSERT(vd == vd->vdev_top);
2249	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2250
2251	if (vd->vdev_ms != NULL) {
2252		metaslab_group_t *mg = vd->vdev_mg;
2253
2254		metaslab_group_histogram_verify(mg);
2255		metaslab_class_histogram_verify(mg->mg_class);
2256
2257		for (int m = 0; m < vd->vdev_ms_count; m++) {
2258			metaslab_t *msp = vd->vdev_ms[m];
2259
2260			if (msp == NULL || msp->ms_sm == NULL)
2261				continue;
2262
2263			mutex_enter(&msp->ms_lock);
2264			/*
2265			 * If the metaslab was not loaded when the vdev
2266			 * was removed then the histogram accounting may
2267			 * not be accurate. Update the histogram information
2268			 * here so that we ensure that the metaslab group
2269			 * and metaslab class are up-to-date.
2270			 */
2271			metaslab_group_histogram_remove(mg, msp);
2272
2273			VERIFY0(space_map_allocated(msp->ms_sm));
2274			space_map_free(msp->ms_sm, tx);
2275			space_map_close(msp->ms_sm);
2276			msp->ms_sm = NULL;
2277			mutex_exit(&msp->ms_lock);
2278		}
2279
2280		metaslab_group_histogram_verify(mg);
2281		metaslab_class_histogram_verify(mg->mg_class);
2282		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2283			ASSERT0(mg->mg_histogram[i]);
2284
2285	}
2286
2287	if (vd->vdev_ms_array) {
2288		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2289		vd->vdev_ms_array = 0;
2290	}
2291
2292	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2293		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2294		vd->vdev_top_zap = 0;
2295	}
2296	dmu_tx_commit(tx);
2297}
2298
2299void
2300vdev_sync_done(vdev_t *vd, uint64_t txg)
2301{
2302	metaslab_t *msp;
2303	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2304
2305	ASSERT(!vd->vdev_ishole);
2306
2307	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2308		metaslab_sync_done(msp, txg);
2309
2310	if (reassess)
2311		metaslab_sync_reassess(vd->vdev_mg);
2312}
2313
2314void
2315vdev_sync(vdev_t *vd, uint64_t txg)
2316{
2317	spa_t *spa = vd->vdev_spa;
2318	vdev_t *lvd;
2319	metaslab_t *msp;
2320	dmu_tx_t *tx;
2321
2322	ASSERT(!vd->vdev_ishole);
2323
2324	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2325		ASSERT(vd == vd->vdev_top);
2326		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2327		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2328		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2329		ASSERT(vd->vdev_ms_array != 0);
2330		vdev_config_dirty(vd);
2331		dmu_tx_commit(tx);
2332	}
2333
2334	/*
2335	 * Remove the metadata associated with this vdev once it's empty.
2336	 */
2337	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2338		vdev_remove(vd, txg);
2339
2340	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2341		metaslab_sync(msp, txg);
2342		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2343	}
2344
2345	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2346		vdev_dtl_sync(lvd, txg);
2347
2348	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2349}
2350
2351uint64_t
2352vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2353{
2354	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2355}
2356
2357/*
2358 * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2359 * not be opened, and no I/O is attempted.
2360 */
2361int
2362vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2363{
2364	vdev_t *vd, *tvd;
2365
2366	spa_vdev_state_enter(spa, SCL_NONE);
2367
2368	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2369		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2370
2371	if (!vd->vdev_ops->vdev_op_leaf)
2372		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2373
2374	tvd = vd->vdev_top;
2375
2376	/*
2377	 * We don't directly use the aux state here, but if we do a
2378	 * vdev_reopen(), we need this value to be present to remember why we
2379	 * were faulted.
2380	 */
2381	vd->vdev_label_aux = aux;
2382
2383	/*
2384	 * Faulted state takes precedence over degraded.
2385	 */
2386	vd->vdev_delayed_close = B_FALSE;
2387	vd->vdev_faulted = 1ULL;
2388	vd->vdev_degraded = 0ULL;
2389	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2390
2391	/*
2392	 * If this device has the only valid copy of the data, then
2393	 * back off and simply mark the vdev as degraded instead.
2394	 */
2395	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2396		vd->vdev_degraded = 1ULL;
2397		vd->vdev_faulted = 0ULL;
2398
2399		/*
2400		 * If we reopen the device and it's not dead, only then do we
2401		 * mark it degraded.
2402		 */
2403		vdev_reopen(tvd);
2404
2405		if (vdev_readable(vd))
2406			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2407	}
2408
2409	return (spa_vdev_state_exit(spa, vd, 0));
2410}
2411
2412/*
2413 * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2414 * user that something is wrong.  The vdev continues to operate as normal as far
2415 * as I/O is concerned.
2416 */
2417int
2418vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2419{
2420	vdev_t *vd;
2421
2422	spa_vdev_state_enter(spa, SCL_NONE);
2423
2424	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2425		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2426
2427	if (!vd->vdev_ops->vdev_op_leaf)
2428		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2429
2430	/*
2431	 * If the vdev is already faulted, then don't do anything.
2432	 */
2433	if (vd->vdev_faulted || vd->vdev_degraded)
2434		return (spa_vdev_state_exit(spa, NULL, 0));
2435
2436	vd->vdev_degraded = 1ULL;
2437	if (!vdev_is_dead(vd))
2438		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2439		    aux);
2440
2441	return (spa_vdev_state_exit(spa, vd, 0));
2442}
2443
2444/*
2445 * Online the given vdev.
2446 *
2447 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2448 * spare device should be detached when the device finishes resilvering.
2449 * Second, the online should be treated like a 'test' online case, so no FMA
2450 * events are generated if the device fails to open.
2451 */
2452int
2453vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2454{
2455	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2456	boolean_t wasoffline;
2457	vdev_state_t oldstate;
2458
2459	spa_vdev_state_enter(spa, SCL_NONE);
2460
2461	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2462		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2463
2464	if (!vd->vdev_ops->vdev_op_leaf)
2465		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2466
2467	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
2468	oldstate = vd->vdev_state;
2469
2470	tvd = vd->vdev_top;
2471	vd->vdev_offline = B_FALSE;
2472	vd->vdev_tmpoffline = B_FALSE;
2473	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2474	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2475
2476	/* XXX - L2ARC 1.0 does not support expansion */
2477	if (!vd->vdev_aux) {
2478		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2479			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2480	}
2481
2482	vdev_reopen(tvd);
2483	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2484
2485	if (!vd->vdev_aux) {
2486		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2487			pvd->vdev_expanding = B_FALSE;
2488	}
2489
2490	if (newstate)
2491		*newstate = vd->vdev_state;
2492	if ((flags & ZFS_ONLINE_UNSPARE) &&
2493	    !vdev_is_dead(vd) && vd->vdev_parent &&
2494	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2495	    vd->vdev_parent->vdev_child[0] == vd)
2496		vd->vdev_unspare = B_TRUE;
2497
2498	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2499
2500		/* XXX - L2ARC 1.0 does not support expansion */
2501		if (vd->vdev_aux)
2502			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2503		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2504	}
2505
2506	if (wasoffline ||
2507	    (oldstate < VDEV_STATE_DEGRADED &&
2508	    vd->vdev_state >= VDEV_STATE_DEGRADED))
2509		spa_event_notify(spa, vd, ESC_ZFS_VDEV_ONLINE);
2510
2511	return (spa_vdev_state_exit(spa, vd, 0));
2512}
2513
2514static int
2515vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2516{
2517	vdev_t *vd, *tvd;
2518	int error = 0;
2519	uint64_t generation;
2520	metaslab_group_t *mg;
2521
2522top:
2523	spa_vdev_state_enter(spa, SCL_ALLOC);
2524
2525	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2526		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2527
2528	if (!vd->vdev_ops->vdev_op_leaf)
2529		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2530
2531	tvd = vd->vdev_top;
2532	mg = tvd->vdev_mg;
2533	generation = spa->spa_config_generation + 1;
2534
2535	/*
2536	 * If the device isn't already offline, try to offline it.
2537	 */
2538	if (!vd->vdev_offline) {
2539		/*
2540		 * If this device has the only valid copy of some data,
2541		 * don't allow it to be offlined. Log devices are always
2542		 * expendable.
2543		 */
2544		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2545		    vdev_dtl_required(vd))
2546			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2547
2548		/*
2549		 * If the top-level is a slog and it has had allocations
2550		 * then proceed.  We check that the vdev's metaslab group
2551		 * is not NULL since it's possible that we may have just
2552		 * added this vdev but not yet initialized its metaslabs.
2553		 */
2554		if (tvd->vdev_islog && mg != NULL) {
2555			/*
2556			 * Prevent any future allocations.
2557			 */
2558			metaslab_group_passivate(mg);
2559			(void) spa_vdev_state_exit(spa, vd, 0);
2560
2561			error = spa_offline_log(spa);
2562
2563			spa_vdev_state_enter(spa, SCL_ALLOC);
2564
2565			/*
2566			 * Check to see if the config has changed.
2567			 */
2568			if (error || generation != spa->spa_config_generation) {
2569				metaslab_group_activate(mg);
2570				if (error)
2571					return (spa_vdev_state_exit(spa,
2572					    vd, error));
2573				(void) spa_vdev_state_exit(spa, vd, 0);
2574				goto top;
2575			}
2576			ASSERT0(tvd->vdev_stat.vs_alloc);
2577		}
2578
2579		/*
2580		 * Offline this device and reopen its top-level vdev.
2581		 * If the top-level vdev is a log device then just offline
2582		 * it. Otherwise, if this action results in the top-level
2583		 * vdev becoming unusable, undo it and fail the request.
2584		 */
2585		vd->vdev_offline = B_TRUE;
2586		vdev_reopen(tvd);
2587
2588		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2589		    vdev_is_dead(tvd)) {
2590			vd->vdev_offline = B_FALSE;
2591			vdev_reopen(tvd);
2592			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2593		}
2594
2595		/*
2596		 * Add the device back into the metaslab rotor so that
2597		 * once we online the device it's open for business.
2598		 */
2599		if (tvd->vdev_islog && mg != NULL)
2600			metaslab_group_activate(mg);
2601	}
2602
2603	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2604
2605	return (spa_vdev_state_exit(spa, vd, 0));
2606}
2607
2608int
2609vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2610{
2611	int error;
2612
2613	mutex_enter(&spa->spa_vdev_top_lock);
2614	error = vdev_offline_locked(spa, guid, flags);
2615	mutex_exit(&spa->spa_vdev_top_lock);
2616
2617	return (error);
2618}
2619
2620/*
2621 * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2622 * vdev_offline(), we assume the spa config is locked.  We also clear all
2623 * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2624 */
2625void
2626vdev_clear(spa_t *spa, vdev_t *vd)
2627{
2628	vdev_t *rvd = spa->spa_root_vdev;
2629
2630	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2631
2632	if (vd == NULL)
2633		vd = rvd;
2634
2635	vd->vdev_stat.vs_read_errors = 0;
2636	vd->vdev_stat.vs_write_errors = 0;
2637	vd->vdev_stat.vs_checksum_errors = 0;
2638
2639	for (int c = 0; c < vd->vdev_children; c++)
2640		vdev_clear(spa, vd->vdev_child[c]);
2641
2642	/*
2643	 * If we're in the FAULTED state or have experienced failed I/O, then
2644	 * clear the persistent state and attempt to reopen the device.  We
2645	 * also mark the vdev config dirty, so that the new faulted state is
2646	 * written out to disk.
2647	 */
2648	if (vd->vdev_faulted || vd->vdev_degraded ||
2649	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2650
2651		/*
2652		 * When reopening in reponse to a clear event, it may be due to
2653		 * a fmadm repair request.  In this case, if the device is
2654		 * still broken, we want to still post the ereport again.
2655		 */
2656		vd->vdev_forcefault = B_TRUE;
2657
2658		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2659		vd->vdev_cant_read = B_FALSE;
2660		vd->vdev_cant_write = B_FALSE;
2661
2662		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2663
2664		vd->vdev_forcefault = B_FALSE;
2665
2666		if (vd != rvd && vdev_writeable(vd->vdev_top))
2667			vdev_state_dirty(vd->vdev_top);
2668
2669		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2670			spa_async_request(spa, SPA_ASYNC_RESILVER);
2671
2672		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2673	}
2674
2675	/*
2676	 * When clearing a FMA-diagnosed fault, we always want to
2677	 * unspare the device, as we assume that the original spare was
2678	 * done in response to the FMA fault.
2679	 */
2680	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2681	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2682	    vd->vdev_parent->vdev_child[0] == vd)
2683		vd->vdev_unspare = B_TRUE;
2684}
2685
2686boolean_t
2687vdev_is_dead(vdev_t *vd)
2688{
2689	/*
2690	 * Holes and missing devices are always considered "dead".
2691	 * This simplifies the code since we don't have to check for
2692	 * these types of devices in the various code paths.
2693	 * Instead we rely on the fact that we skip over dead devices
2694	 * before issuing I/O to them.
2695	 */
2696	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2697	    vd->vdev_ops == &vdev_missing_ops);
2698}
2699
2700boolean_t
2701vdev_readable(vdev_t *vd)
2702{
2703	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2704}
2705
2706boolean_t
2707vdev_writeable(vdev_t *vd)
2708{
2709	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2710}
2711
2712boolean_t
2713vdev_allocatable(vdev_t *vd)
2714{
2715	uint64_t state = vd->vdev_state;
2716
2717	/*
2718	 * We currently allow allocations from vdevs which may be in the
2719	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2720	 * fails to reopen then we'll catch it later when we're holding
2721	 * the proper locks.  Note that we have to get the vdev state
2722	 * in a local variable because although it changes atomically,
2723	 * we're asking two separate questions about it.
2724	 */
2725	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2726	    !vd->vdev_cant_write && !vd->vdev_ishole &&
2727	    vd->vdev_mg->mg_initialized);
2728}
2729
2730boolean_t
2731vdev_accessible(vdev_t *vd, zio_t *zio)
2732{
2733	ASSERT(zio->io_vd == vd);
2734
2735	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2736		return (B_FALSE);
2737
2738	if (zio->io_type == ZIO_TYPE_READ)
2739		return (!vd->vdev_cant_read);
2740
2741	if (zio->io_type == ZIO_TYPE_WRITE)
2742		return (!vd->vdev_cant_write);
2743
2744	return (B_TRUE);
2745}
2746
2747/*
2748 * Get statistics for the given vdev.
2749 */
2750void
2751vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2752{
2753	spa_t *spa = vd->vdev_spa;
2754	vdev_t *rvd = spa->spa_root_vdev;
2755	vdev_t *tvd = vd->vdev_top;
2756
2757	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2758
2759	mutex_enter(&vd->vdev_stat_lock);
2760	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2761	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2762	vs->vs_state = vd->vdev_state;
2763	vs->vs_rsize = vdev_get_min_asize(vd);
2764	if (vd->vdev_ops->vdev_op_leaf)
2765		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2766	/*
2767	 * Report expandable space on top-level, non-auxillary devices only.
2768	 * The expandable space is reported in terms of metaslab sized units
2769	 * since that determines how much space the pool can expand.
2770	 */
2771	if (vd->vdev_aux == NULL && tvd != NULL) {
2772		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
2773		    spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
2774	}
2775	if (vd->vdev_aux == NULL && vd == vd->vdev_top && !vd->vdev_ishole) {
2776		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
2777	}
2778
2779	/*
2780	 * If we're getting stats on the root vdev, aggregate the I/O counts
2781	 * over all top-level vdevs (i.e. the direct children of the root).
2782	 */
2783	if (vd == rvd) {
2784		for (int c = 0; c < rvd->vdev_children; c++) {
2785			vdev_t *cvd = rvd->vdev_child[c];
2786			vdev_stat_t *cvs = &cvd->vdev_stat;
2787
2788			for (int t = 0; t < ZIO_TYPES; t++) {
2789				vs->vs_ops[t] += cvs->vs_ops[t];
2790				vs->vs_bytes[t] += cvs->vs_bytes[t];
2791			}
2792			cvs->vs_scan_removing = cvd->vdev_removing;
2793		}
2794	}
2795	mutex_exit(&vd->vdev_stat_lock);
2796}
2797
2798void
2799vdev_clear_stats(vdev_t *vd)
2800{
2801	mutex_enter(&vd->vdev_stat_lock);
2802	vd->vdev_stat.vs_space = 0;
2803	vd->vdev_stat.vs_dspace = 0;
2804	vd->vdev_stat.vs_alloc = 0;
2805	mutex_exit(&vd->vdev_stat_lock);
2806}
2807
2808void
2809vdev_scan_stat_init(vdev_t *vd)
2810{
2811	vdev_stat_t *vs = &vd->vdev_stat;
2812
2813	for (int c = 0; c < vd->vdev_children; c++)
2814		vdev_scan_stat_init(vd->vdev_child[c]);
2815
2816	mutex_enter(&vd->vdev_stat_lock);
2817	vs->vs_scan_processed = 0;
2818	mutex_exit(&vd->vdev_stat_lock);
2819}
2820
2821void
2822vdev_stat_update(zio_t *zio, uint64_t psize)
2823{
2824	spa_t *spa = zio->io_spa;
2825	vdev_t *rvd = spa->spa_root_vdev;
2826	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2827	vdev_t *pvd;
2828	uint64_t txg = zio->io_txg;
2829	vdev_stat_t *vs = &vd->vdev_stat;
2830	zio_type_t type = zio->io_type;
2831	int flags = zio->io_flags;
2832
2833	/*
2834	 * If this i/o is a gang leader, it didn't do any actual work.
2835	 */
2836	if (zio->io_gang_tree)
2837		return;
2838
2839	if (zio->io_error == 0) {
2840		/*
2841		 * If this is a root i/o, don't count it -- we've already
2842		 * counted the top-level vdevs, and vdev_get_stats() will
2843		 * aggregate them when asked.  This reduces contention on
2844		 * the root vdev_stat_lock and implicitly handles blocks
2845		 * that compress away to holes, for which there is no i/o.
2846		 * (Holes never create vdev children, so all the counters
2847		 * remain zero, which is what we want.)
2848		 *
2849		 * Note: this only applies to successful i/o (io_error == 0)
2850		 * because unlike i/o counts, errors are not additive.
2851		 * When reading a ditto block, for example, failure of
2852		 * one top-level vdev does not imply a root-level error.
2853		 */
2854		if (vd == rvd)
2855			return;
2856
2857		ASSERT(vd == zio->io_vd);
2858
2859		if (flags & ZIO_FLAG_IO_BYPASS)
2860			return;
2861
2862		mutex_enter(&vd->vdev_stat_lock);
2863
2864		if (flags & ZIO_FLAG_IO_REPAIR) {
2865			if (flags & ZIO_FLAG_SCAN_THREAD) {
2866				dsl_scan_phys_t *scn_phys =
2867				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2868				uint64_t *processed = &scn_phys->scn_processed;
2869
2870				/* XXX cleanup? */
2871				if (vd->vdev_ops->vdev_op_leaf)
2872					atomic_add_64(processed, psize);
2873				vs->vs_scan_processed += psize;
2874			}
2875
2876			if (flags & ZIO_FLAG_SELF_HEAL)
2877				vs->vs_self_healed += psize;
2878		}
2879
2880		vs->vs_ops[type]++;
2881		vs->vs_bytes[type] += psize;
2882
2883		mutex_exit(&vd->vdev_stat_lock);
2884		return;
2885	}
2886
2887	if (flags & ZIO_FLAG_SPECULATIVE)
2888		return;
2889
2890	/*
2891	 * If this is an I/O error that is going to be retried, then ignore the
2892	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2893	 * hard errors, when in reality they can happen for any number of
2894	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2895	 */
2896	if (zio->io_error == EIO &&
2897	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2898		return;
2899
2900	/*
2901	 * Intent logs writes won't propagate their error to the root
2902	 * I/O so don't mark these types of failures as pool-level
2903	 * errors.
2904	 */
2905	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2906		return;
2907
2908	mutex_enter(&vd->vdev_stat_lock);
2909	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2910		if (zio->io_error == ECKSUM)
2911			vs->vs_checksum_errors++;
2912		else
2913			vs->vs_read_errors++;
2914	}
2915	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2916		vs->vs_write_errors++;
2917	mutex_exit(&vd->vdev_stat_lock);
2918
2919	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2920	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2921	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2922	    spa->spa_claiming)) {
2923		/*
2924		 * This is either a normal write (not a repair), or it's
2925		 * a repair induced by the scrub thread, or it's a repair
2926		 * made by zil_claim() during spa_load() in the first txg.
2927		 * In the normal case, we commit the DTL change in the same
2928		 * txg as the block was born.  In the scrub-induced repair
2929		 * case, we know that scrubs run in first-pass syncing context,
2930		 * so we commit the DTL change in spa_syncing_txg(spa).
2931		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2932		 *
2933		 * We currently do not make DTL entries for failed spontaneous
2934		 * self-healing writes triggered by normal (non-scrubbing)
2935		 * reads, because we have no transactional context in which to
2936		 * do so -- and it's not clear that it'd be desirable anyway.
2937		 */
2938		if (vd->vdev_ops->vdev_op_leaf) {
2939			uint64_t commit_txg = txg;
2940			if (flags & ZIO_FLAG_SCAN_THREAD) {
2941				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2942				ASSERT(spa_sync_pass(spa) == 1);
2943				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2944				commit_txg = spa_syncing_txg(spa);
2945			} else if (spa->spa_claiming) {
2946				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2947				commit_txg = spa_first_txg(spa);
2948			}
2949			ASSERT(commit_txg >= spa_syncing_txg(spa));
2950			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2951				return;
2952			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2953				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2954			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2955		}
2956		if (vd != rvd)
2957			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2958	}
2959}
2960
2961/*
2962 * Update the in-core space usage stats for this vdev, its metaslab class,
2963 * and the root vdev.
2964 */
2965void
2966vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2967    int64_t space_delta)
2968{
2969	int64_t dspace_delta = space_delta;
2970	spa_t *spa = vd->vdev_spa;
2971	vdev_t *rvd = spa->spa_root_vdev;
2972	metaslab_group_t *mg = vd->vdev_mg;
2973	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2974
2975	ASSERT(vd == vd->vdev_top);
2976
2977	/*
2978	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2979	 * factor.  We must calculate this here and not at the root vdev
2980	 * because the root vdev's psize-to-asize is simply the max of its
2981	 * childrens', thus not accurate enough for us.
2982	 */
2983	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2984	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2985	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2986	    vd->vdev_deflate_ratio;
2987
2988	mutex_enter(&vd->vdev_stat_lock);
2989	vd->vdev_stat.vs_alloc += alloc_delta;
2990	vd->vdev_stat.vs_space += space_delta;
2991	vd->vdev_stat.vs_dspace += dspace_delta;
2992	mutex_exit(&vd->vdev_stat_lock);
2993
2994	if (mc == spa_normal_class(spa)) {
2995		mutex_enter(&rvd->vdev_stat_lock);
2996		rvd->vdev_stat.vs_alloc += alloc_delta;
2997		rvd->vdev_stat.vs_space += space_delta;
2998		rvd->vdev_stat.vs_dspace += dspace_delta;
2999		mutex_exit(&rvd->vdev_stat_lock);
3000	}
3001
3002	if (mc != NULL) {
3003		ASSERT(rvd == vd->vdev_parent);
3004		ASSERT(vd->vdev_ms_count != 0);
3005
3006		metaslab_class_space_update(mc,
3007		    alloc_delta, defer_delta, space_delta, dspace_delta);
3008	}
3009}
3010
3011/*
3012 * Mark a top-level vdev's config as dirty, placing it on the dirty list
3013 * so that it will be written out next time the vdev configuration is synced.
3014 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3015 */
3016void
3017vdev_config_dirty(vdev_t *vd)
3018{
3019	spa_t *spa = vd->vdev_spa;
3020	vdev_t *rvd = spa->spa_root_vdev;
3021	int c;
3022
3023	ASSERT(spa_writeable(spa));
3024
3025	/*
3026	 * If this is an aux vdev (as with l2cache and spare devices), then we
3027	 * update the vdev config manually and set the sync flag.
3028	 */
3029	if (vd->vdev_aux != NULL) {
3030		spa_aux_vdev_t *sav = vd->vdev_aux;
3031		nvlist_t **aux;
3032		uint_t naux;
3033
3034		for (c = 0; c < sav->sav_count; c++) {
3035			if (sav->sav_vdevs[c] == vd)
3036				break;
3037		}
3038
3039		if (c == sav->sav_count) {
3040			/*
3041			 * We're being removed.  There's nothing more to do.
3042			 */
3043			ASSERT(sav->sav_sync == B_TRUE);
3044			return;
3045		}
3046
3047		sav->sav_sync = B_TRUE;
3048
3049		if (nvlist_lookup_nvlist_array(sav->sav_config,
3050		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3051			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3052			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3053		}
3054
3055		ASSERT(c < naux);
3056
3057		/*
3058		 * Setting the nvlist in the middle if the array is a little
3059		 * sketchy, but it will work.
3060		 */
3061		nvlist_free(aux[c]);
3062		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3063
3064		return;
3065	}
3066
3067	/*
3068	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3069	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3070	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3071	 * so this is sufficient to ensure mutual exclusion.
3072	 */
3073	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3074	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3075	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3076
3077	if (vd == rvd) {
3078		for (c = 0; c < rvd->vdev_children; c++)
3079			vdev_config_dirty(rvd->vdev_child[c]);
3080	} else {
3081		ASSERT(vd == vd->vdev_top);
3082
3083		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3084		    !vd->vdev_ishole)
3085			list_insert_head(&spa->spa_config_dirty_list, vd);
3086	}
3087}
3088
3089void
3090vdev_config_clean(vdev_t *vd)
3091{
3092	spa_t *spa = vd->vdev_spa;
3093
3094	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3095	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3096	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3097
3098	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3099	list_remove(&spa->spa_config_dirty_list, vd);
3100}
3101
3102/*
3103 * Mark a top-level vdev's state as dirty, so that the next pass of
3104 * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3105 * the state changes from larger config changes because they require
3106 * much less locking, and are often needed for administrative actions.
3107 */
3108void
3109vdev_state_dirty(vdev_t *vd)
3110{
3111	spa_t *spa = vd->vdev_spa;
3112
3113	ASSERT(spa_writeable(spa));
3114	ASSERT(vd == vd->vdev_top);
3115
3116	/*
3117	 * The state list is protected by the SCL_STATE lock.  The caller
3118	 * must either hold SCL_STATE as writer, or must be the sync thread
3119	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3120	 * so this is sufficient to ensure mutual exclusion.
3121	 */
3122	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3123	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3124	    spa_config_held(spa, SCL_STATE, RW_READER)));
3125
3126	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
3127		list_insert_head(&spa->spa_state_dirty_list, vd);
3128}
3129
3130void
3131vdev_state_clean(vdev_t *vd)
3132{
3133	spa_t *spa = vd->vdev_spa;
3134
3135	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3136	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3137	    spa_config_held(spa, SCL_STATE, RW_READER)));
3138
3139	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3140	list_remove(&spa->spa_state_dirty_list, vd);
3141}
3142
3143/*
3144 * Propagate vdev state up from children to parent.
3145 */
3146void
3147vdev_propagate_state(vdev_t *vd)
3148{
3149	spa_t *spa = vd->vdev_spa;
3150	vdev_t *rvd = spa->spa_root_vdev;
3151	int degraded = 0, faulted = 0;
3152	int corrupted = 0;
3153	vdev_t *child;
3154
3155	if (vd->vdev_children > 0) {
3156		for (int c = 0; c < vd->vdev_children; c++) {
3157			child = vd->vdev_child[c];
3158
3159			/*
3160			 * Don't factor holes into the decision.
3161			 */
3162			if (child->vdev_ishole)
3163				continue;
3164
3165			if (!vdev_readable(child) ||
3166			    (!vdev_writeable(child) && spa_writeable(spa))) {
3167				/*
3168				 * Root special: if there is a top-level log
3169				 * device, treat the root vdev as if it were
3170				 * degraded.
3171				 */
3172				if (child->vdev_islog && vd == rvd)
3173					degraded++;
3174				else
3175					faulted++;
3176			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3177				degraded++;
3178			}
3179
3180			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3181				corrupted++;
3182		}
3183
3184		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3185
3186		/*
3187		 * Root special: if there is a top-level vdev that cannot be
3188		 * opened due to corrupted metadata, then propagate the root
3189		 * vdev's aux state as 'corrupt' rather than 'insufficient
3190		 * replicas'.
3191		 */
3192		if (corrupted && vd == rvd &&
3193		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3194			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3195			    VDEV_AUX_CORRUPT_DATA);
3196	}
3197
3198	if (vd->vdev_parent)
3199		vdev_propagate_state(vd->vdev_parent);
3200}
3201
3202/*
3203 * Set a vdev's state.  If this is during an open, we don't update the parent
3204 * state, because we're in the process of opening children depth-first.
3205 * Otherwise, we propagate the change to the parent.
3206 *
3207 * If this routine places a device in a faulted state, an appropriate ereport is
3208 * generated.
3209 */
3210void
3211vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3212{
3213	uint64_t save_state;
3214	spa_t *spa = vd->vdev_spa;
3215
3216	if (state == vd->vdev_state) {
3217		vd->vdev_stat.vs_aux = aux;
3218		return;
3219	}
3220
3221	save_state = vd->vdev_state;
3222
3223	vd->vdev_state = state;
3224	vd->vdev_stat.vs_aux = aux;
3225
3226	/*
3227	 * If we are setting the vdev state to anything but an open state, then
3228	 * always close the underlying device unless the device has requested
3229	 * a delayed close (i.e. we're about to remove or fault the device).
3230	 * Otherwise, we keep accessible but invalid devices open forever.
3231	 * We don't call vdev_close() itself, because that implies some extra
3232	 * checks (offline, etc) that we don't want here.  This is limited to
3233	 * leaf devices, because otherwise closing the device will affect other
3234	 * children.
3235	 */
3236	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3237	    vd->vdev_ops->vdev_op_leaf)
3238		vd->vdev_ops->vdev_op_close(vd);
3239
3240	/*
3241	 * If we have brought this vdev back into service, we need
3242	 * to notify fmd so that it can gracefully repair any outstanding
3243	 * cases due to a missing device.  We do this in all cases, even those
3244	 * that probably don't correlate to a repaired fault.  This is sure to
3245	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3246	 * this is a transient state it's OK, as the retire agent will
3247	 * double-check the state of the vdev before repairing it.
3248	 */
3249	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3250	    vd->vdev_prevstate != state)
3251		zfs_post_state_change(spa, vd);
3252
3253	if (vd->vdev_removed &&
3254	    state == VDEV_STATE_CANT_OPEN &&
3255	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3256		/*
3257		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3258		 * device was previously marked removed and someone attempted to
3259		 * reopen it.  If this failed due to a nonexistent device, then
3260		 * keep the device in the REMOVED state.  We also let this be if
3261		 * it is one of our special test online cases, which is only
3262		 * attempting to online the device and shouldn't generate an FMA
3263		 * fault.
3264		 */
3265		vd->vdev_state = VDEV_STATE_REMOVED;
3266		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3267	} else if (state == VDEV_STATE_REMOVED) {
3268		vd->vdev_removed = B_TRUE;
3269	} else if (state == VDEV_STATE_CANT_OPEN) {
3270		/*
3271		 * If we fail to open a vdev during an import or recovery, we
3272		 * mark it as "not available", which signifies that it was
3273		 * never there to begin with.  Failure to open such a device
3274		 * is not considered an error.
3275		 */
3276		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3277		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3278		    vd->vdev_ops->vdev_op_leaf)
3279			vd->vdev_not_present = 1;
3280
3281		/*
3282		 * Post the appropriate ereport.  If the 'prevstate' field is
3283		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3284		 * that this is part of a vdev_reopen().  In this case, we don't
3285		 * want to post the ereport if the device was already in the
3286		 * CANT_OPEN state beforehand.
3287		 *
3288		 * If the 'checkremove' flag is set, then this is an attempt to
3289		 * online the device in response to an insertion event.  If we
3290		 * hit this case, then we have detected an insertion event for a
3291		 * faulted or offline device that wasn't in the removed state.
3292		 * In this scenario, we don't post an ereport because we are
3293		 * about to replace the device, or attempt an online with
3294		 * vdev_forcefault, which will generate the fault for us.
3295		 */
3296		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3297		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3298		    vd != spa->spa_root_vdev) {
3299			const char *class;
3300
3301			switch (aux) {
3302			case VDEV_AUX_OPEN_FAILED:
3303				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3304				break;
3305			case VDEV_AUX_CORRUPT_DATA:
3306				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3307				break;
3308			case VDEV_AUX_NO_REPLICAS:
3309				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3310				break;
3311			case VDEV_AUX_BAD_GUID_SUM:
3312				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3313				break;
3314			case VDEV_AUX_TOO_SMALL:
3315				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3316				break;
3317			case VDEV_AUX_BAD_LABEL:
3318				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3319				break;
3320			default:
3321				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3322			}
3323
3324			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3325		}
3326
3327		/* Erase any notion of persistent removed state */
3328		vd->vdev_removed = B_FALSE;
3329	} else {
3330		vd->vdev_removed = B_FALSE;
3331	}
3332
3333	if (!isopen && vd->vdev_parent)
3334		vdev_propagate_state(vd->vdev_parent);
3335}
3336
3337/*
3338 * Check the vdev configuration to ensure that it's capable of supporting
3339 * a root pool. We do not support partial configuration.
3340 * In addition, only a single top-level vdev is allowed.
3341 */
3342boolean_t
3343vdev_is_bootable(vdev_t *vd)
3344{
3345	if (!vd->vdev_ops->vdev_op_leaf) {
3346		char *vdev_type = vd->vdev_ops->vdev_op_type;
3347
3348		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3349		    vd->vdev_children > 1) {
3350			return (B_FALSE);
3351		} else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3352			return (B_FALSE);
3353		}
3354	}
3355
3356	for (int c = 0; c < vd->vdev_children; c++) {
3357		if (!vdev_is_bootable(vd->vdev_child[c]))
3358			return (B_FALSE);
3359	}
3360	return (B_TRUE);
3361}
3362
3363/*
3364 * Load the state from the original vdev tree (ovd) which
3365 * we've retrieved from the MOS config object. If the original
3366 * vdev was offline or faulted then we transfer that state to the
3367 * device in the current vdev tree (nvd).
3368 */
3369void
3370vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3371{
3372	spa_t *spa = nvd->vdev_spa;
3373
3374	ASSERT(nvd->vdev_top->vdev_islog);
3375	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3376	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3377
3378	for (int c = 0; c < nvd->vdev_children; c++)
3379		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3380
3381	if (nvd->vdev_ops->vdev_op_leaf) {
3382		/*
3383		 * Restore the persistent vdev state
3384		 */
3385		nvd->vdev_offline = ovd->vdev_offline;
3386		nvd->vdev_faulted = ovd->vdev_faulted;
3387		nvd->vdev_degraded = ovd->vdev_degraded;
3388		nvd->vdev_removed = ovd->vdev_removed;
3389	}
3390}
3391
3392/*
3393 * Determine if a log device has valid content.  If the vdev was
3394 * removed or faulted in the MOS config then we know that
3395 * the content on the log device has already been written to the pool.
3396 */
3397boolean_t
3398vdev_log_state_valid(vdev_t *vd)
3399{
3400	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3401	    !vd->vdev_removed)
3402		return (B_TRUE);
3403
3404	for (int c = 0; c < vd->vdev_children; c++)
3405		if (vdev_log_state_valid(vd->vdev_child[c]))
3406			return (B_TRUE);
3407
3408	return (B_FALSE);
3409}
3410
3411/*
3412 * Expand a vdev if possible.
3413 */
3414void
3415vdev_expand(vdev_t *vd, uint64_t txg)
3416{
3417	ASSERT(vd->vdev_top == vd);
3418	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3419
3420	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3421		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3422		vdev_config_dirty(vd);
3423	}
3424}
3425
3426/*
3427 * Split a vdev.
3428 */
3429void
3430vdev_split(vdev_t *vd)
3431{
3432	vdev_t *cvd, *pvd = vd->vdev_parent;
3433
3434	vdev_remove_child(pvd, vd);
3435	vdev_compact_children(pvd);
3436
3437	cvd = pvd->vdev_child[0];
3438	if (pvd->vdev_children == 1) {
3439		vdev_remove_parent(cvd);
3440		cvd->vdev_splitting = B_TRUE;
3441	}
3442	vdev_propagate_state(cvd);
3443}
3444
3445void
3446vdev_deadman(vdev_t *vd)
3447{
3448	for (int c = 0; c < vd->vdev_children; c++) {
3449		vdev_t *cvd = vd->vdev_child[c];
3450
3451		vdev_deadman(cvd);
3452	}
3453
3454	if (vd->vdev_ops->vdev_op_leaf) {
3455		vdev_queue_t *vq = &vd->vdev_queue;
3456
3457		mutex_enter(&vq->vq_lock);
3458		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3459			spa_t *spa = vd->vdev_spa;
3460			zio_t *fio;
3461			uint64_t delta;
3462
3463			/*
3464			 * Look at the head of all the pending queues,
3465			 * if any I/O has been outstanding for longer than
3466			 * the spa_deadman_synctime we panic the system.
3467			 */
3468			fio = avl_first(&vq->vq_active_tree);
3469			delta = gethrtime() - fio->io_timestamp;
3470			if (delta > spa_deadman_synctime(spa)) {
3471				zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3472				    "delta %lluns, last io %lluns",
3473				    fio->io_timestamp, delta,
3474				    vq->vq_io_complete_ts);
3475				fm_panic("I/O to pool '%s' appears to be "
3476				    "hung.", spa_name(spa));
3477			}
3478		}
3479		mutex_exit(&vq->vq_lock);
3480	}
3481}
3482