1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22/*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
26 * Copyright (c) 2014 Integros [integros.com]
27 */
28
29#include <sys/zfs_context.h>
30#include <sys/fm/fs/zfs.h>
31#include <sys/spa.h>
32#include <sys/spa_impl.h>
33#include <sys/dmu.h>
34#include <sys/dmu_tx.h>
35#include <sys/vdev_impl.h>
36#include <sys/uberblock_impl.h>
37#include <sys/metaslab.h>
38#include <sys/metaslab_impl.h>
39#include <sys/space_map.h>
40#include <sys/space_reftree.h>
41#include <sys/zio.h>
42#include <sys/zap.h>
43#include <sys/fs/zfs.h>
44#include <sys/arc.h>
45#include <sys/zil.h>
46#include <sys/dsl_scan.h>
47
48/*
49 * Virtual device management.
50 */
51
52static vdev_ops_t *vdev_ops_table[] = {
53	&vdev_root_ops,
54	&vdev_raidz_ops,
55	&vdev_mirror_ops,
56	&vdev_replacing_ops,
57	&vdev_spare_ops,
58	&vdev_disk_ops,
59	&vdev_file_ops,
60	&vdev_missing_ops,
61	&vdev_hole_ops,
62	NULL
63};
64
65/* maximum scrub/resilver I/O queue per leaf vdev */
66int zfs_scrub_limit = 10;
67
68/*
69 * When a vdev is added, it will be divided into approximately (but no
70 * more than) this number of metaslabs.
71 */
72int metaslabs_per_vdev = 200;
73
74/*
75 * Given a vdev type, return the appropriate ops vector.
76 */
77static vdev_ops_t *
78vdev_getops(const char *type)
79{
80	vdev_ops_t *ops, **opspp;
81
82	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
83		if (strcmp(ops->vdev_op_type, type) == 0)
84			break;
85
86	return (ops);
87}
88
89/*
90 * Default asize function: return the MAX of psize with the asize of
91 * all children.  This is what's used by anything other than RAID-Z.
92 */
93uint64_t
94vdev_default_asize(vdev_t *vd, uint64_t psize)
95{
96	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
97	uint64_t csize;
98
99	for (int c = 0; c < vd->vdev_children; c++) {
100		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
101		asize = MAX(asize, csize);
102	}
103
104	return (asize);
105}
106
107/*
108 * Get the minimum allocatable size. We define the allocatable size as
109 * the vdev's asize rounded to the nearest metaslab. This allows us to
110 * replace or attach devices which don't have the same physical size but
111 * can still satisfy the same number of allocations.
112 */
113uint64_t
114vdev_get_min_asize(vdev_t *vd)
115{
116	vdev_t *pvd = vd->vdev_parent;
117
118	/*
119	 * If our parent is NULL (inactive spare or cache) or is the root,
120	 * just return our own asize.
121	 */
122	if (pvd == NULL)
123		return (vd->vdev_asize);
124
125	/*
126	 * The top-level vdev just returns the allocatable size rounded
127	 * to the nearest metaslab.
128	 */
129	if (vd == vd->vdev_top)
130		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
131
132	/*
133	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
134	 * so each child must provide at least 1/Nth of its asize.
135	 */
136	if (pvd->vdev_ops == &vdev_raidz_ops)
137		return (pvd->vdev_min_asize / pvd->vdev_children);
138
139	return (pvd->vdev_min_asize);
140}
141
142void
143vdev_set_min_asize(vdev_t *vd)
144{
145	vd->vdev_min_asize = vdev_get_min_asize(vd);
146
147	for (int c = 0; c < vd->vdev_children; c++)
148		vdev_set_min_asize(vd->vdev_child[c]);
149}
150
151vdev_t *
152vdev_lookup_top(spa_t *spa, uint64_t vdev)
153{
154	vdev_t *rvd = spa->spa_root_vdev;
155
156	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
157
158	if (vdev < rvd->vdev_children) {
159		ASSERT(rvd->vdev_child[vdev] != NULL);
160		return (rvd->vdev_child[vdev]);
161	}
162
163	return (NULL);
164}
165
166vdev_t *
167vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
168{
169	vdev_t *mvd;
170
171	if (vd->vdev_guid == guid)
172		return (vd);
173
174	for (int c = 0; c < vd->vdev_children; c++)
175		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
176		    NULL)
177			return (mvd);
178
179	return (NULL);
180}
181
182static int
183vdev_count_leaves_impl(vdev_t *vd)
184{
185	int n = 0;
186
187	if (vd->vdev_ops->vdev_op_leaf)
188		return (1);
189
190	for (int c = 0; c < vd->vdev_children; c++)
191		n += vdev_count_leaves_impl(vd->vdev_child[c]);
192
193	return (n);
194}
195
196int
197vdev_count_leaves(spa_t *spa)
198{
199	return (vdev_count_leaves_impl(spa->spa_root_vdev));
200}
201
202void
203vdev_add_child(vdev_t *pvd, vdev_t *cvd)
204{
205	size_t oldsize, newsize;
206	uint64_t id = cvd->vdev_id;
207	vdev_t **newchild;
208	spa_t *spa = cvd->vdev_spa;
209
210	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
211	ASSERT(cvd->vdev_parent == NULL);
212
213	cvd->vdev_parent = pvd;
214
215	if (pvd == NULL)
216		return;
217
218	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
219
220	oldsize = pvd->vdev_children * sizeof (vdev_t *);
221	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
222	newsize = pvd->vdev_children * sizeof (vdev_t *);
223
224	newchild = kmem_zalloc(newsize, KM_SLEEP);
225	if (pvd->vdev_child != NULL) {
226		bcopy(pvd->vdev_child, newchild, oldsize);
227		kmem_free(pvd->vdev_child, oldsize);
228	}
229
230	pvd->vdev_child = newchild;
231	pvd->vdev_child[id] = cvd;
232
233	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
234	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
235
236	/*
237	 * Walk up all ancestors to update guid sum.
238	 */
239	for (; pvd != NULL; pvd = pvd->vdev_parent)
240		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
241}
242
243void
244vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
245{
246	int c;
247	uint_t id = cvd->vdev_id;
248
249	ASSERT(cvd->vdev_parent == pvd);
250
251	if (pvd == NULL)
252		return;
253
254	ASSERT(id < pvd->vdev_children);
255	ASSERT(pvd->vdev_child[id] == cvd);
256
257	pvd->vdev_child[id] = NULL;
258	cvd->vdev_parent = NULL;
259
260	for (c = 0; c < pvd->vdev_children; c++)
261		if (pvd->vdev_child[c])
262			break;
263
264	if (c == pvd->vdev_children) {
265		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
266		pvd->vdev_child = NULL;
267		pvd->vdev_children = 0;
268	}
269
270	/*
271	 * Walk up all ancestors to update guid sum.
272	 */
273	for (; pvd != NULL; pvd = pvd->vdev_parent)
274		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
275}
276
277/*
278 * Remove any holes in the child array.
279 */
280void
281vdev_compact_children(vdev_t *pvd)
282{
283	vdev_t **newchild, *cvd;
284	int oldc = pvd->vdev_children;
285	int newc;
286
287	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
288
289	for (int c = newc = 0; c < oldc; c++)
290		if (pvd->vdev_child[c])
291			newc++;
292
293	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
294
295	for (int c = newc = 0; c < oldc; c++) {
296		if ((cvd = pvd->vdev_child[c]) != NULL) {
297			newchild[newc] = cvd;
298			cvd->vdev_id = newc++;
299		}
300	}
301
302	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
303	pvd->vdev_child = newchild;
304	pvd->vdev_children = newc;
305}
306
307/*
308 * Allocate and minimally initialize a vdev_t.
309 */
310vdev_t *
311vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
312{
313	vdev_t *vd;
314
315	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
316
317	if (spa->spa_root_vdev == NULL) {
318		ASSERT(ops == &vdev_root_ops);
319		spa->spa_root_vdev = vd;
320		spa->spa_load_guid = spa_generate_guid(NULL);
321	}
322
323	if (guid == 0 && ops != &vdev_hole_ops) {
324		if (spa->spa_root_vdev == vd) {
325			/*
326			 * The root vdev's guid will also be the pool guid,
327			 * which must be unique among all pools.
328			 */
329			guid = spa_generate_guid(NULL);
330		} else {
331			/*
332			 * Any other vdev's guid must be unique within the pool.
333			 */
334			guid = spa_generate_guid(spa);
335		}
336		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
337	}
338
339	vd->vdev_spa = spa;
340	vd->vdev_id = id;
341	vd->vdev_guid = guid;
342	vd->vdev_guid_sum = guid;
343	vd->vdev_ops = ops;
344	vd->vdev_state = VDEV_STATE_CLOSED;
345	vd->vdev_ishole = (ops == &vdev_hole_ops);
346
347	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
348	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
349	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
350	mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
351	for (int t = 0; t < DTL_TYPES; t++) {
352		vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
353		    &vd->vdev_dtl_lock);
354	}
355	txg_list_create(&vd->vdev_ms_list,
356	    offsetof(struct metaslab, ms_txg_node));
357	txg_list_create(&vd->vdev_dtl_list,
358	    offsetof(struct vdev, vdev_dtl_node));
359	vd->vdev_stat.vs_timestamp = gethrtime();
360	vdev_queue_init(vd);
361	vdev_cache_init(vd);
362
363	return (vd);
364}
365
366/*
367 * Allocate a new vdev.  The 'alloctype' is used to control whether we are
368 * creating a new vdev or loading an existing one - the behavior is slightly
369 * different for each case.
370 */
371int
372vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
373    int alloctype)
374{
375	vdev_ops_t *ops;
376	char *type;
377	uint64_t guid = 0, islog, nparity;
378	vdev_t *vd;
379
380	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
381
382	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
383		return (SET_ERROR(EINVAL));
384
385	if ((ops = vdev_getops(type)) == NULL)
386		return (SET_ERROR(EINVAL));
387
388	/*
389	 * If this is a load, get the vdev guid from the nvlist.
390	 * Otherwise, vdev_alloc_common() will generate one for us.
391	 */
392	if (alloctype == VDEV_ALLOC_LOAD) {
393		uint64_t label_id;
394
395		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
396		    label_id != id)
397			return (SET_ERROR(EINVAL));
398
399		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
400			return (SET_ERROR(EINVAL));
401	} else if (alloctype == VDEV_ALLOC_SPARE) {
402		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
403			return (SET_ERROR(EINVAL));
404	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
405		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
406			return (SET_ERROR(EINVAL));
407	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
408		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
409			return (SET_ERROR(EINVAL));
410	}
411
412	/*
413	 * The first allocated vdev must be of type 'root'.
414	 */
415	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
416		return (SET_ERROR(EINVAL));
417
418	/*
419	 * Determine whether we're a log vdev.
420	 */
421	islog = 0;
422	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
423	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
424		return (SET_ERROR(ENOTSUP));
425
426	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
427		return (SET_ERROR(ENOTSUP));
428
429	/*
430	 * Set the nparity property for RAID-Z vdevs.
431	 */
432	nparity = -1ULL;
433	if (ops == &vdev_raidz_ops) {
434		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
435		    &nparity) == 0) {
436			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
437				return (SET_ERROR(EINVAL));
438			/*
439			 * Previous versions could only support 1 or 2 parity
440			 * device.
441			 */
442			if (nparity > 1 &&
443			    spa_version(spa) < SPA_VERSION_RAIDZ2)
444				return (SET_ERROR(ENOTSUP));
445			if (nparity > 2 &&
446			    spa_version(spa) < SPA_VERSION_RAIDZ3)
447				return (SET_ERROR(ENOTSUP));
448		} else {
449			/*
450			 * We require the parity to be specified for SPAs that
451			 * support multiple parity levels.
452			 */
453			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
454				return (SET_ERROR(EINVAL));
455			/*
456			 * Otherwise, we default to 1 parity device for RAID-Z.
457			 */
458			nparity = 1;
459		}
460	} else {
461		nparity = 0;
462	}
463	ASSERT(nparity != -1ULL);
464
465	vd = vdev_alloc_common(spa, id, guid, ops);
466
467	vd->vdev_islog = islog;
468	vd->vdev_nparity = nparity;
469
470	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
471		vd->vdev_path = spa_strdup(vd->vdev_path);
472	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
473		vd->vdev_devid = spa_strdup(vd->vdev_devid);
474	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
475	    &vd->vdev_physpath) == 0)
476		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
477	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
478		vd->vdev_fru = spa_strdup(vd->vdev_fru);
479
480	/*
481	 * Set the whole_disk property.  If it's not specified, leave the value
482	 * as -1.
483	 */
484	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
485	    &vd->vdev_wholedisk) != 0)
486		vd->vdev_wholedisk = -1ULL;
487
488	/*
489	 * Look for the 'not present' flag.  This will only be set if the device
490	 * was not present at the time of import.
491	 */
492	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
493	    &vd->vdev_not_present);
494
495	/*
496	 * Get the alignment requirement.
497	 */
498	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
499
500	/*
501	 * Retrieve the vdev creation time.
502	 */
503	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
504	    &vd->vdev_crtxg);
505
506	/*
507	 * If we're a top-level vdev, try to load the allocation parameters.
508	 */
509	if (parent && !parent->vdev_parent &&
510	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
511		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
512		    &vd->vdev_ms_array);
513		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
514		    &vd->vdev_ms_shift);
515		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
516		    &vd->vdev_asize);
517		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
518		    &vd->vdev_removing);
519		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
520		    &vd->vdev_top_zap);
521	} else {
522		ASSERT0(vd->vdev_top_zap);
523	}
524
525	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
526		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
527		    alloctype == VDEV_ALLOC_ADD ||
528		    alloctype == VDEV_ALLOC_SPLIT ||
529		    alloctype == VDEV_ALLOC_ROOTPOOL);
530		vd->vdev_mg = metaslab_group_create(islog ?
531		    spa_log_class(spa) : spa_normal_class(spa), vd);
532	}
533
534	if (vd->vdev_ops->vdev_op_leaf &&
535	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
536		(void) nvlist_lookup_uint64(nv,
537		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
538	} else {
539		ASSERT0(vd->vdev_leaf_zap);
540	}
541
542	/*
543	 * If we're a leaf vdev, try to load the DTL object and other state.
544	 */
545
546	if (vd->vdev_ops->vdev_op_leaf &&
547	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
548	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
549		if (alloctype == VDEV_ALLOC_LOAD) {
550			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
551			    &vd->vdev_dtl_object);
552			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
553			    &vd->vdev_unspare);
554		}
555
556		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
557			uint64_t spare = 0;
558
559			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
560			    &spare) == 0 && spare)
561				spa_spare_add(vd);
562		}
563
564		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
565		    &vd->vdev_offline);
566
567		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
568		    &vd->vdev_resilver_txg);
569
570		/*
571		 * When importing a pool, we want to ignore the persistent fault
572		 * state, as the diagnosis made on another system may not be
573		 * valid in the current context.  Local vdevs will
574		 * remain in the faulted state.
575		 */
576		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
577			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
578			    &vd->vdev_faulted);
579			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
580			    &vd->vdev_degraded);
581			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
582			    &vd->vdev_removed);
583
584			if (vd->vdev_faulted || vd->vdev_degraded) {
585				char *aux;
586
587				vd->vdev_label_aux =
588				    VDEV_AUX_ERR_EXCEEDED;
589				if (nvlist_lookup_string(nv,
590				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
591				    strcmp(aux, "external") == 0)
592					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
593			}
594		}
595	}
596
597	/*
598	 * Add ourselves to the parent's list of children.
599	 */
600	vdev_add_child(parent, vd);
601
602	*vdp = vd;
603
604	return (0);
605}
606
607void
608vdev_free(vdev_t *vd)
609{
610	spa_t *spa = vd->vdev_spa;
611
612	/*
613	 * vdev_free() implies closing the vdev first.  This is simpler than
614	 * trying to ensure complicated semantics for all callers.
615	 */
616	vdev_close(vd);
617
618	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
619	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
620
621	/*
622	 * Free all children.
623	 */
624	for (int c = 0; c < vd->vdev_children; c++)
625		vdev_free(vd->vdev_child[c]);
626
627	ASSERT(vd->vdev_child == NULL);
628	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
629
630	/*
631	 * Discard allocation state.
632	 */
633	if (vd->vdev_mg != NULL) {
634		vdev_metaslab_fini(vd);
635		metaslab_group_destroy(vd->vdev_mg);
636	}
637
638	ASSERT0(vd->vdev_stat.vs_space);
639	ASSERT0(vd->vdev_stat.vs_dspace);
640	ASSERT0(vd->vdev_stat.vs_alloc);
641
642	/*
643	 * Remove this vdev from its parent's child list.
644	 */
645	vdev_remove_child(vd->vdev_parent, vd);
646
647	ASSERT(vd->vdev_parent == NULL);
648
649	/*
650	 * Clean up vdev structure.
651	 */
652	vdev_queue_fini(vd);
653	vdev_cache_fini(vd);
654
655	if (vd->vdev_path)
656		spa_strfree(vd->vdev_path);
657	if (vd->vdev_devid)
658		spa_strfree(vd->vdev_devid);
659	if (vd->vdev_physpath)
660		spa_strfree(vd->vdev_physpath);
661	if (vd->vdev_fru)
662		spa_strfree(vd->vdev_fru);
663
664	if (vd->vdev_isspare)
665		spa_spare_remove(vd);
666	if (vd->vdev_isl2cache)
667		spa_l2cache_remove(vd);
668
669	txg_list_destroy(&vd->vdev_ms_list);
670	txg_list_destroy(&vd->vdev_dtl_list);
671
672	mutex_enter(&vd->vdev_dtl_lock);
673	space_map_close(vd->vdev_dtl_sm);
674	for (int t = 0; t < DTL_TYPES; t++) {
675		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
676		range_tree_destroy(vd->vdev_dtl[t]);
677	}
678	mutex_exit(&vd->vdev_dtl_lock);
679
680	mutex_destroy(&vd->vdev_queue_lock);
681	mutex_destroy(&vd->vdev_dtl_lock);
682	mutex_destroy(&vd->vdev_stat_lock);
683	mutex_destroy(&vd->vdev_probe_lock);
684
685	if (vd == spa->spa_root_vdev)
686		spa->spa_root_vdev = NULL;
687
688	kmem_free(vd, sizeof (vdev_t));
689}
690
691/*
692 * Transfer top-level vdev state from svd to tvd.
693 */
694static void
695vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
696{
697	spa_t *spa = svd->vdev_spa;
698	metaslab_t *msp;
699	vdev_t *vd;
700	int t;
701
702	ASSERT(tvd == tvd->vdev_top);
703
704	tvd->vdev_ms_array = svd->vdev_ms_array;
705	tvd->vdev_ms_shift = svd->vdev_ms_shift;
706	tvd->vdev_ms_count = svd->vdev_ms_count;
707	tvd->vdev_top_zap = svd->vdev_top_zap;
708
709	svd->vdev_ms_array = 0;
710	svd->vdev_ms_shift = 0;
711	svd->vdev_ms_count = 0;
712	svd->vdev_top_zap = 0;
713
714	if (tvd->vdev_mg)
715		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
716	tvd->vdev_mg = svd->vdev_mg;
717	tvd->vdev_ms = svd->vdev_ms;
718
719	svd->vdev_mg = NULL;
720	svd->vdev_ms = NULL;
721
722	if (tvd->vdev_mg != NULL)
723		tvd->vdev_mg->mg_vd = tvd;
724
725	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
726	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
727	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
728
729	svd->vdev_stat.vs_alloc = 0;
730	svd->vdev_stat.vs_space = 0;
731	svd->vdev_stat.vs_dspace = 0;
732
733	for (t = 0; t < TXG_SIZE; t++) {
734		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
735			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
736		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
737			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
738		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
739			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
740	}
741
742	if (list_link_active(&svd->vdev_config_dirty_node)) {
743		vdev_config_clean(svd);
744		vdev_config_dirty(tvd);
745	}
746
747	if (list_link_active(&svd->vdev_state_dirty_node)) {
748		vdev_state_clean(svd);
749		vdev_state_dirty(tvd);
750	}
751
752	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
753	svd->vdev_deflate_ratio = 0;
754
755	tvd->vdev_islog = svd->vdev_islog;
756	svd->vdev_islog = 0;
757}
758
759static void
760vdev_top_update(vdev_t *tvd, vdev_t *vd)
761{
762	if (vd == NULL)
763		return;
764
765	vd->vdev_top = tvd;
766
767	for (int c = 0; c < vd->vdev_children; c++)
768		vdev_top_update(tvd, vd->vdev_child[c]);
769}
770
771/*
772 * Add a mirror/replacing vdev above an existing vdev.
773 */
774vdev_t *
775vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
776{
777	spa_t *spa = cvd->vdev_spa;
778	vdev_t *pvd = cvd->vdev_parent;
779	vdev_t *mvd;
780
781	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
782
783	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
784
785	mvd->vdev_asize = cvd->vdev_asize;
786	mvd->vdev_min_asize = cvd->vdev_min_asize;
787	mvd->vdev_max_asize = cvd->vdev_max_asize;
788	mvd->vdev_ashift = cvd->vdev_ashift;
789	mvd->vdev_state = cvd->vdev_state;
790	mvd->vdev_crtxg = cvd->vdev_crtxg;
791
792	vdev_remove_child(pvd, cvd);
793	vdev_add_child(pvd, mvd);
794	cvd->vdev_id = mvd->vdev_children;
795	vdev_add_child(mvd, cvd);
796	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
797
798	if (mvd == mvd->vdev_top)
799		vdev_top_transfer(cvd, mvd);
800
801	return (mvd);
802}
803
804/*
805 * Remove a 1-way mirror/replacing vdev from the tree.
806 */
807void
808vdev_remove_parent(vdev_t *cvd)
809{
810	vdev_t *mvd = cvd->vdev_parent;
811	vdev_t *pvd = mvd->vdev_parent;
812
813	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
814
815	ASSERT(mvd->vdev_children == 1);
816	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
817	    mvd->vdev_ops == &vdev_replacing_ops ||
818	    mvd->vdev_ops == &vdev_spare_ops);
819	cvd->vdev_ashift = mvd->vdev_ashift;
820
821	vdev_remove_child(mvd, cvd);
822	vdev_remove_child(pvd, mvd);
823
824	/*
825	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
826	 * Otherwise, we could have detached an offline device, and when we
827	 * go to import the pool we'll think we have two top-level vdevs,
828	 * instead of a different version of the same top-level vdev.
829	 */
830	if (mvd->vdev_top == mvd) {
831		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
832		cvd->vdev_orig_guid = cvd->vdev_guid;
833		cvd->vdev_guid += guid_delta;
834		cvd->vdev_guid_sum += guid_delta;
835	}
836	cvd->vdev_id = mvd->vdev_id;
837	vdev_add_child(pvd, cvd);
838	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
839
840	if (cvd == cvd->vdev_top)
841		vdev_top_transfer(mvd, cvd);
842
843	ASSERT(mvd->vdev_children == 0);
844	vdev_free(mvd);
845}
846
847int
848vdev_metaslab_init(vdev_t *vd, uint64_t txg)
849{
850	spa_t *spa = vd->vdev_spa;
851	objset_t *mos = spa->spa_meta_objset;
852	uint64_t m;
853	uint64_t oldc = vd->vdev_ms_count;
854	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
855	metaslab_t **mspp;
856	int error;
857
858	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
859
860	/*
861	 * This vdev is not being allocated from yet or is a hole.
862	 */
863	if (vd->vdev_ms_shift == 0)
864		return (0);
865
866	ASSERT(!vd->vdev_ishole);
867
868	/*
869	 * Compute the raidz-deflation ratio.  Note, we hard-code
870	 * in 128k (1 << 17) because it is the "typical" blocksize.
871	 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
872	 * otherwise it would inconsistently account for existing bp's.
873	 */
874	vd->vdev_deflate_ratio = (1 << 17) /
875	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
876
877	ASSERT(oldc <= newc);
878
879	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
880
881	if (oldc != 0) {
882		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
883		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
884	}
885
886	vd->vdev_ms = mspp;
887	vd->vdev_ms_count = newc;
888
889	for (m = oldc; m < newc; m++) {
890		uint64_t object = 0;
891
892		if (txg == 0) {
893			error = dmu_read(mos, vd->vdev_ms_array,
894			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
895			    DMU_READ_PREFETCH);
896			if (error)
897				return (error);
898		}
899
900		error = metaslab_init(vd->vdev_mg, m, object, txg,
901		    &(vd->vdev_ms[m]));
902		if (error)
903			return (error);
904	}
905
906	if (txg == 0)
907		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
908
909	/*
910	 * If the vdev is being removed we don't activate
911	 * the metaslabs since we want to ensure that no new
912	 * allocations are performed on this device.
913	 */
914	if (oldc == 0 && !vd->vdev_removing)
915		metaslab_group_activate(vd->vdev_mg);
916
917	if (txg == 0)
918		spa_config_exit(spa, SCL_ALLOC, FTAG);
919
920	return (0);
921}
922
923void
924vdev_metaslab_fini(vdev_t *vd)
925{
926	uint64_t m;
927	uint64_t count = vd->vdev_ms_count;
928
929	if (vd->vdev_ms != NULL) {
930		metaslab_group_passivate(vd->vdev_mg);
931		for (m = 0; m < count; m++) {
932			metaslab_t *msp = vd->vdev_ms[m];
933
934			if (msp != NULL)
935				metaslab_fini(msp);
936		}
937		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
938		vd->vdev_ms = NULL;
939	}
940}
941
942typedef struct vdev_probe_stats {
943	boolean_t	vps_readable;
944	boolean_t	vps_writeable;
945	int		vps_flags;
946} vdev_probe_stats_t;
947
948static void
949vdev_probe_done(zio_t *zio)
950{
951	spa_t *spa = zio->io_spa;
952	vdev_t *vd = zio->io_vd;
953	vdev_probe_stats_t *vps = zio->io_private;
954
955	ASSERT(vd->vdev_probe_zio != NULL);
956
957	if (zio->io_type == ZIO_TYPE_READ) {
958		if (zio->io_error == 0)
959			vps->vps_readable = 1;
960		if (zio->io_error == 0 && spa_writeable(spa)) {
961			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
962			    zio->io_offset, zio->io_size, zio->io_data,
963			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
964			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
965		} else {
966			zio_buf_free(zio->io_data, zio->io_size);
967		}
968	} else if (zio->io_type == ZIO_TYPE_WRITE) {
969		if (zio->io_error == 0)
970			vps->vps_writeable = 1;
971		zio_buf_free(zio->io_data, zio->io_size);
972	} else if (zio->io_type == ZIO_TYPE_NULL) {
973		zio_t *pio;
974
975		vd->vdev_cant_read |= !vps->vps_readable;
976		vd->vdev_cant_write |= !vps->vps_writeable;
977
978		if (vdev_readable(vd) &&
979		    (vdev_writeable(vd) || !spa_writeable(spa))) {
980			zio->io_error = 0;
981		} else {
982			ASSERT(zio->io_error != 0);
983			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
984			    spa, vd, NULL, 0, 0);
985			zio->io_error = SET_ERROR(ENXIO);
986		}
987
988		mutex_enter(&vd->vdev_probe_lock);
989		ASSERT(vd->vdev_probe_zio == zio);
990		vd->vdev_probe_zio = NULL;
991		mutex_exit(&vd->vdev_probe_lock);
992
993		zio_link_t *zl = NULL;
994		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
995			if (!vdev_accessible(vd, pio))
996				pio->io_error = SET_ERROR(ENXIO);
997
998		kmem_free(vps, sizeof (*vps));
999	}
1000}
1001
1002/*
1003 * Determine whether this device is accessible.
1004 *
1005 * Read and write to several known locations: the pad regions of each
1006 * vdev label but the first, which we leave alone in case it contains
1007 * a VTOC.
1008 */
1009zio_t *
1010vdev_probe(vdev_t *vd, zio_t *zio)
1011{
1012	spa_t *spa = vd->vdev_spa;
1013	vdev_probe_stats_t *vps = NULL;
1014	zio_t *pio;
1015
1016	ASSERT(vd->vdev_ops->vdev_op_leaf);
1017
1018	/*
1019	 * Don't probe the probe.
1020	 */
1021	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1022		return (NULL);
1023
1024	/*
1025	 * To prevent 'probe storms' when a device fails, we create
1026	 * just one probe i/o at a time.  All zios that want to probe
1027	 * this vdev will become parents of the probe io.
1028	 */
1029	mutex_enter(&vd->vdev_probe_lock);
1030
1031	if ((pio = vd->vdev_probe_zio) == NULL) {
1032		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1033
1034		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1035		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1036		    ZIO_FLAG_TRYHARD;
1037
1038		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1039			/*
1040			 * vdev_cant_read and vdev_cant_write can only
1041			 * transition from TRUE to FALSE when we have the
1042			 * SCL_ZIO lock as writer; otherwise they can only
1043			 * transition from FALSE to TRUE.  This ensures that
1044			 * any zio looking at these values can assume that
1045			 * failures persist for the life of the I/O.  That's
1046			 * important because when a device has intermittent
1047			 * connectivity problems, we want to ensure that
1048			 * they're ascribed to the device (ENXIO) and not
1049			 * the zio (EIO).
1050			 *
1051			 * Since we hold SCL_ZIO as writer here, clear both
1052			 * values so the probe can reevaluate from first
1053			 * principles.
1054			 */
1055			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1056			vd->vdev_cant_read = B_FALSE;
1057			vd->vdev_cant_write = B_FALSE;
1058		}
1059
1060		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1061		    vdev_probe_done, vps,
1062		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1063
1064		/*
1065		 * We can't change the vdev state in this context, so we
1066		 * kick off an async task to do it on our behalf.
1067		 */
1068		if (zio != NULL) {
1069			vd->vdev_probe_wanted = B_TRUE;
1070			spa_async_request(spa, SPA_ASYNC_PROBE);
1071		}
1072	}
1073
1074	if (zio != NULL)
1075		zio_add_child(zio, pio);
1076
1077	mutex_exit(&vd->vdev_probe_lock);
1078
1079	if (vps == NULL) {
1080		ASSERT(zio != NULL);
1081		return (NULL);
1082	}
1083
1084	for (int l = 1; l < VDEV_LABELS; l++) {
1085		zio_nowait(zio_read_phys(pio, vd,
1086		    vdev_label_offset(vd->vdev_psize, l,
1087		    offsetof(vdev_label_t, vl_pad2)),
1088		    VDEV_PAD_SIZE, zio_buf_alloc(VDEV_PAD_SIZE),
1089		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1090		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1091	}
1092
1093	if (zio == NULL)
1094		return (pio);
1095
1096	zio_nowait(pio);
1097	return (NULL);
1098}
1099
1100static void
1101vdev_open_child(void *arg)
1102{
1103	vdev_t *vd = arg;
1104
1105	vd->vdev_open_thread = curthread;
1106	vd->vdev_open_error = vdev_open(vd);
1107	vd->vdev_open_thread = NULL;
1108}
1109
1110boolean_t
1111vdev_uses_zvols(vdev_t *vd)
1112{
1113	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1114	    strlen(ZVOL_DIR)) == 0)
1115		return (B_TRUE);
1116	for (int c = 0; c < vd->vdev_children; c++)
1117		if (vdev_uses_zvols(vd->vdev_child[c]))
1118			return (B_TRUE);
1119	return (B_FALSE);
1120}
1121
1122void
1123vdev_open_children(vdev_t *vd)
1124{
1125	taskq_t *tq;
1126	int children = vd->vdev_children;
1127
1128	/*
1129	 * in order to handle pools on top of zvols, do the opens
1130	 * in a single thread so that the same thread holds the
1131	 * spa_namespace_lock
1132	 */
1133	if (vdev_uses_zvols(vd)) {
1134		for (int c = 0; c < children; c++)
1135			vd->vdev_child[c]->vdev_open_error =
1136			    vdev_open(vd->vdev_child[c]);
1137		return;
1138	}
1139	tq = taskq_create("vdev_open", children, minclsyspri,
1140	    children, children, TASKQ_PREPOPULATE);
1141
1142	for (int c = 0; c < children; c++)
1143		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1144		    TQ_SLEEP) != NULL);
1145
1146	taskq_destroy(tq);
1147}
1148
1149/*
1150 * Prepare a virtual device for access.
1151 */
1152int
1153vdev_open(vdev_t *vd)
1154{
1155	spa_t *spa = vd->vdev_spa;
1156	int error;
1157	uint64_t osize = 0;
1158	uint64_t max_osize = 0;
1159	uint64_t asize, max_asize, psize;
1160	uint64_t ashift = 0;
1161
1162	ASSERT(vd->vdev_open_thread == curthread ||
1163	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1164	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1165	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1166	    vd->vdev_state == VDEV_STATE_OFFLINE);
1167
1168	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1169	vd->vdev_cant_read = B_FALSE;
1170	vd->vdev_cant_write = B_FALSE;
1171	vd->vdev_min_asize = vdev_get_min_asize(vd);
1172
1173	/*
1174	 * If this vdev is not removed, check its fault status.  If it's
1175	 * faulted, bail out of the open.
1176	 */
1177	if (!vd->vdev_removed && vd->vdev_faulted) {
1178		ASSERT(vd->vdev_children == 0);
1179		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1180		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1181		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1182		    vd->vdev_label_aux);
1183		return (SET_ERROR(ENXIO));
1184	} else if (vd->vdev_offline) {
1185		ASSERT(vd->vdev_children == 0);
1186		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1187		return (SET_ERROR(ENXIO));
1188	}
1189
1190	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1191
1192	/*
1193	 * Reset the vdev_reopening flag so that we actually close
1194	 * the vdev on error.
1195	 */
1196	vd->vdev_reopening = B_FALSE;
1197	if (zio_injection_enabled && error == 0)
1198		error = zio_handle_device_injection(vd, NULL, ENXIO);
1199
1200	if (error) {
1201		if (vd->vdev_removed &&
1202		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1203			vd->vdev_removed = B_FALSE;
1204
1205		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1206		    vd->vdev_stat.vs_aux);
1207		return (error);
1208	}
1209
1210	vd->vdev_removed = B_FALSE;
1211
1212	/*
1213	 * Recheck the faulted flag now that we have confirmed that
1214	 * the vdev is accessible.  If we're faulted, bail.
1215	 */
1216	if (vd->vdev_faulted) {
1217		ASSERT(vd->vdev_children == 0);
1218		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1219		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1220		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1221		    vd->vdev_label_aux);
1222		return (SET_ERROR(ENXIO));
1223	}
1224
1225	if (vd->vdev_degraded) {
1226		ASSERT(vd->vdev_children == 0);
1227		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1228		    VDEV_AUX_ERR_EXCEEDED);
1229	} else {
1230		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1231	}
1232
1233	/*
1234	 * For hole or missing vdevs we just return success.
1235	 */
1236	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1237		return (0);
1238
1239	for (int c = 0; c < vd->vdev_children; c++) {
1240		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1241			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1242			    VDEV_AUX_NONE);
1243			break;
1244		}
1245	}
1246
1247	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1248	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1249
1250	if (vd->vdev_children == 0) {
1251		if (osize < SPA_MINDEVSIZE) {
1252			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1253			    VDEV_AUX_TOO_SMALL);
1254			return (SET_ERROR(EOVERFLOW));
1255		}
1256		psize = osize;
1257		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1258		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1259		    VDEV_LABEL_END_SIZE);
1260	} else {
1261		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1262		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1263			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1264			    VDEV_AUX_TOO_SMALL);
1265			return (SET_ERROR(EOVERFLOW));
1266		}
1267		psize = 0;
1268		asize = osize;
1269		max_asize = max_osize;
1270	}
1271
1272	vd->vdev_psize = psize;
1273
1274	/*
1275	 * Make sure the allocatable size hasn't shrunk.
1276	 */
1277	if (asize < vd->vdev_min_asize) {
1278		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1279		    VDEV_AUX_BAD_LABEL);
1280		return (SET_ERROR(EINVAL));
1281	}
1282
1283	if (vd->vdev_asize == 0) {
1284		/*
1285		 * This is the first-ever open, so use the computed values.
1286		 * For testing purposes, a higher ashift can be requested.
1287		 */
1288		vd->vdev_asize = asize;
1289		vd->vdev_max_asize = max_asize;
1290		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1291	} else {
1292		/*
1293		 * Detect if the alignment requirement has increased.
1294		 * We don't want to make the pool unavailable, just
1295		 * issue a warning instead.
1296		 */
1297		if (ashift > vd->vdev_top->vdev_ashift &&
1298		    vd->vdev_ops->vdev_op_leaf) {
1299			cmn_err(CE_WARN,
1300			    "Disk, '%s', has a block alignment that is "
1301			    "larger than the pool's alignment\n",
1302			    vd->vdev_path);
1303		}
1304		vd->vdev_max_asize = max_asize;
1305	}
1306
1307	/*
1308	 * If all children are healthy and the asize has increased,
1309	 * then we've experienced dynamic LUN growth.  If automatic
1310	 * expansion is enabled then use the additional space.
1311	 */
1312	if (vd->vdev_state == VDEV_STATE_HEALTHY && asize > vd->vdev_asize &&
1313	    (vd->vdev_expanding || spa->spa_autoexpand))
1314		vd->vdev_asize = asize;
1315
1316	vdev_set_min_asize(vd);
1317
1318	/*
1319	 * Ensure we can issue some IO before declaring the
1320	 * vdev open for business.
1321	 */
1322	if (vd->vdev_ops->vdev_op_leaf &&
1323	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1324		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1325		    VDEV_AUX_ERR_EXCEEDED);
1326		return (error);
1327	}
1328
1329	/*
1330	 * Track the min and max ashift values for normal data devices.
1331	 */
1332	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1333	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1334		if (vd->vdev_ashift > spa->spa_max_ashift)
1335			spa->spa_max_ashift = vd->vdev_ashift;
1336		if (vd->vdev_ashift < spa->spa_min_ashift)
1337			spa->spa_min_ashift = vd->vdev_ashift;
1338	}
1339
1340	/*
1341	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1342	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1343	 * since this would just restart the scrub we are already doing.
1344	 */
1345	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1346	    vdev_resilver_needed(vd, NULL, NULL))
1347		spa_async_request(spa, SPA_ASYNC_RESILVER);
1348
1349	return (0);
1350}
1351
1352/*
1353 * Called once the vdevs are all opened, this routine validates the label
1354 * contents.  This needs to be done before vdev_load() so that we don't
1355 * inadvertently do repair I/Os to the wrong device.
1356 *
1357 * If 'strict' is false ignore the spa guid check. This is necessary because
1358 * if the machine crashed during a re-guid the new guid might have been written
1359 * to all of the vdev labels, but not the cached config. The strict check
1360 * will be performed when the pool is opened again using the mos config.
1361 *
1362 * This function will only return failure if one of the vdevs indicates that it
1363 * has since been destroyed or exported.  This is only possible if
1364 * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1365 * will be updated but the function will return 0.
1366 */
1367int
1368vdev_validate(vdev_t *vd, boolean_t strict)
1369{
1370	spa_t *spa = vd->vdev_spa;
1371	nvlist_t *label;
1372	uint64_t guid = 0, top_guid;
1373	uint64_t state;
1374
1375	for (int c = 0; c < vd->vdev_children; c++)
1376		if (vdev_validate(vd->vdev_child[c], strict) != 0)
1377			return (SET_ERROR(EBADF));
1378
1379	/*
1380	 * If the device has already failed, or was marked offline, don't do
1381	 * any further validation.  Otherwise, label I/O will fail and we will
1382	 * overwrite the previous state.
1383	 */
1384	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1385		uint64_t aux_guid = 0;
1386		nvlist_t *nvl;
1387		uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1388		    spa_last_synced_txg(spa) : -1ULL;
1389
1390		if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1391			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1392			    VDEV_AUX_BAD_LABEL);
1393			return (0);
1394		}
1395
1396		/*
1397		 * Determine if this vdev has been split off into another
1398		 * pool.  If so, then refuse to open it.
1399		 */
1400		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1401		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1402			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1403			    VDEV_AUX_SPLIT_POOL);
1404			nvlist_free(label);
1405			return (0);
1406		}
1407
1408		if (strict && (nvlist_lookup_uint64(label,
1409		    ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1410		    guid != spa_guid(spa))) {
1411			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1412			    VDEV_AUX_CORRUPT_DATA);
1413			nvlist_free(label);
1414			return (0);
1415		}
1416
1417		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1418		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1419		    &aux_guid) != 0)
1420			aux_guid = 0;
1421
1422		/*
1423		 * If this vdev just became a top-level vdev because its
1424		 * sibling was detached, it will have adopted the parent's
1425		 * vdev guid -- but the label may or may not be on disk yet.
1426		 * Fortunately, either version of the label will have the
1427		 * same top guid, so if we're a top-level vdev, we can
1428		 * safely compare to that instead.
1429		 *
1430		 * If we split this vdev off instead, then we also check the
1431		 * original pool's guid.  We don't want to consider the vdev
1432		 * corrupt if it is partway through a split operation.
1433		 */
1434		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1435		    &guid) != 0 ||
1436		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1437		    &top_guid) != 0 ||
1438		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1439		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1440			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1441			    VDEV_AUX_CORRUPT_DATA);
1442			nvlist_free(label);
1443			return (0);
1444		}
1445
1446		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1447		    &state) != 0) {
1448			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1449			    VDEV_AUX_CORRUPT_DATA);
1450			nvlist_free(label);
1451			return (0);
1452		}
1453
1454		nvlist_free(label);
1455
1456		/*
1457		 * If this is a verbatim import, no need to check the
1458		 * state of the pool.
1459		 */
1460		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1461		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1462		    state != POOL_STATE_ACTIVE)
1463			return (SET_ERROR(EBADF));
1464
1465		/*
1466		 * If we were able to open and validate a vdev that was
1467		 * previously marked permanently unavailable, clear that state
1468		 * now.
1469		 */
1470		if (vd->vdev_not_present)
1471			vd->vdev_not_present = 0;
1472	}
1473
1474	return (0);
1475}
1476
1477/*
1478 * Close a virtual device.
1479 */
1480void
1481vdev_close(vdev_t *vd)
1482{
1483	spa_t *spa = vd->vdev_spa;
1484	vdev_t *pvd = vd->vdev_parent;
1485
1486	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1487
1488	/*
1489	 * If our parent is reopening, then we are as well, unless we are
1490	 * going offline.
1491	 */
1492	if (pvd != NULL && pvd->vdev_reopening)
1493		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1494
1495	vd->vdev_ops->vdev_op_close(vd);
1496
1497	vdev_cache_purge(vd);
1498
1499	/*
1500	 * We record the previous state before we close it, so that if we are
1501	 * doing a reopen(), we don't generate FMA ereports if we notice that
1502	 * it's still faulted.
1503	 */
1504	vd->vdev_prevstate = vd->vdev_state;
1505
1506	if (vd->vdev_offline)
1507		vd->vdev_state = VDEV_STATE_OFFLINE;
1508	else
1509		vd->vdev_state = VDEV_STATE_CLOSED;
1510	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1511}
1512
1513void
1514vdev_hold(vdev_t *vd)
1515{
1516	spa_t *spa = vd->vdev_spa;
1517
1518	ASSERT(spa_is_root(spa));
1519	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1520		return;
1521
1522	for (int c = 0; c < vd->vdev_children; c++)
1523		vdev_hold(vd->vdev_child[c]);
1524
1525	if (vd->vdev_ops->vdev_op_leaf)
1526		vd->vdev_ops->vdev_op_hold(vd);
1527}
1528
1529void
1530vdev_rele(vdev_t *vd)
1531{
1532	spa_t *spa = vd->vdev_spa;
1533
1534	ASSERT(spa_is_root(spa));
1535	for (int c = 0; c < vd->vdev_children; c++)
1536		vdev_rele(vd->vdev_child[c]);
1537
1538	if (vd->vdev_ops->vdev_op_leaf)
1539		vd->vdev_ops->vdev_op_rele(vd);
1540}
1541
1542/*
1543 * Reopen all interior vdevs and any unopened leaves.  We don't actually
1544 * reopen leaf vdevs which had previously been opened as they might deadlock
1545 * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1546 * If the leaf has never been opened then open it, as usual.
1547 */
1548void
1549vdev_reopen(vdev_t *vd)
1550{
1551	spa_t *spa = vd->vdev_spa;
1552
1553	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1554
1555	/* set the reopening flag unless we're taking the vdev offline */
1556	vd->vdev_reopening = !vd->vdev_offline;
1557	vdev_close(vd);
1558	(void) vdev_open(vd);
1559
1560	/*
1561	 * Call vdev_validate() here to make sure we have the same device.
1562	 * Otherwise, a device with an invalid label could be successfully
1563	 * opened in response to vdev_reopen().
1564	 */
1565	if (vd->vdev_aux) {
1566		(void) vdev_validate_aux(vd);
1567		if (vdev_readable(vd) && vdev_writeable(vd) &&
1568		    vd->vdev_aux == &spa->spa_l2cache &&
1569		    !l2arc_vdev_present(vd))
1570			l2arc_add_vdev(spa, vd);
1571	} else {
1572		(void) vdev_validate(vd, B_TRUE);
1573	}
1574
1575	/*
1576	 * Reassess parent vdev's health.
1577	 */
1578	vdev_propagate_state(vd);
1579}
1580
1581int
1582vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1583{
1584	int error;
1585
1586	/*
1587	 * Normally, partial opens (e.g. of a mirror) are allowed.
1588	 * For a create, however, we want to fail the request if
1589	 * there are any components we can't open.
1590	 */
1591	error = vdev_open(vd);
1592
1593	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1594		vdev_close(vd);
1595		return (error ? error : ENXIO);
1596	}
1597
1598	/*
1599	 * Recursively load DTLs and initialize all labels.
1600	 */
1601	if ((error = vdev_dtl_load(vd)) != 0 ||
1602	    (error = vdev_label_init(vd, txg, isreplacing ?
1603	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1604		vdev_close(vd);
1605		return (error);
1606	}
1607
1608	return (0);
1609}
1610
1611void
1612vdev_metaslab_set_size(vdev_t *vd)
1613{
1614	/*
1615	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1616	 */
1617	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1618	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1619}
1620
1621void
1622vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1623{
1624	ASSERT(vd == vd->vdev_top);
1625	ASSERT(!vd->vdev_ishole);
1626	ASSERT(ISP2(flags));
1627	ASSERT(spa_writeable(vd->vdev_spa));
1628
1629	if (flags & VDD_METASLAB)
1630		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1631
1632	if (flags & VDD_DTL)
1633		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1634
1635	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1636}
1637
1638void
1639vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1640{
1641	for (int c = 0; c < vd->vdev_children; c++)
1642		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1643
1644	if (vd->vdev_ops->vdev_op_leaf)
1645		vdev_dirty(vd->vdev_top, flags, vd, txg);
1646}
1647
1648/*
1649 * DTLs.
1650 *
1651 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1652 * the vdev has less than perfect replication.  There are four kinds of DTL:
1653 *
1654 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1655 *
1656 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1657 *
1658 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1659 *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1660 *	txgs that was scrubbed.
1661 *
1662 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1663 *	persistent errors or just some device being offline.
1664 *	Unlike the other three, the DTL_OUTAGE map is not generally
1665 *	maintained; it's only computed when needed, typically to
1666 *	determine whether a device can be detached.
1667 *
1668 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1669 * either has the data or it doesn't.
1670 *
1671 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1672 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1673 * if any child is less than fully replicated, then so is its parent.
1674 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1675 * comprising only those txgs which appear in 'maxfaults' or more children;
1676 * those are the txgs we don't have enough replication to read.  For example,
1677 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1678 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1679 * two child DTL_MISSING maps.
1680 *
1681 * It should be clear from the above that to compute the DTLs and outage maps
1682 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1683 * Therefore, that is all we keep on disk.  When loading the pool, or after
1684 * a configuration change, we generate all other DTLs from first principles.
1685 */
1686void
1687vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1688{
1689	range_tree_t *rt = vd->vdev_dtl[t];
1690
1691	ASSERT(t < DTL_TYPES);
1692	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1693	ASSERT(spa_writeable(vd->vdev_spa));
1694
1695	mutex_enter(rt->rt_lock);
1696	if (!range_tree_contains(rt, txg, size))
1697		range_tree_add(rt, txg, size);
1698	mutex_exit(rt->rt_lock);
1699}
1700
1701boolean_t
1702vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1703{
1704	range_tree_t *rt = vd->vdev_dtl[t];
1705	boolean_t dirty = B_FALSE;
1706
1707	ASSERT(t < DTL_TYPES);
1708	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1709
1710	mutex_enter(rt->rt_lock);
1711	if (range_tree_space(rt) != 0)
1712		dirty = range_tree_contains(rt, txg, size);
1713	mutex_exit(rt->rt_lock);
1714
1715	return (dirty);
1716}
1717
1718boolean_t
1719vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1720{
1721	range_tree_t *rt = vd->vdev_dtl[t];
1722	boolean_t empty;
1723
1724	mutex_enter(rt->rt_lock);
1725	empty = (range_tree_space(rt) == 0);
1726	mutex_exit(rt->rt_lock);
1727
1728	return (empty);
1729}
1730
1731/*
1732 * Returns the lowest txg in the DTL range.
1733 */
1734static uint64_t
1735vdev_dtl_min(vdev_t *vd)
1736{
1737	range_seg_t *rs;
1738
1739	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1740	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1741	ASSERT0(vd->vdev_children);
1742
1743	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1744	return (rs->rs_start - 1);
1745}
1746
1747/*
1748 * Returns the highest txg in the DTL.
1749 */
1750static uint64_t
1751vdev_dtl_max(vdev_t *vd)
1752{
1753	range_seg_t *rs;
1754
1755	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1756	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1757	ASSERT0(vd->vdev_children);
1758
1759	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1760	return (rs->rs_end);
1761}
1762
1763/*
1764 * Determine if a resilvering vdev should remove any DTL entries from
1765 * its range. If the vdev was resilvering for the entire duration of the
1766 * scan then it should excise that range from its DTLs. Otherwise, this
1767 * vdev is considered partially resilvered and should leave its DTL
1768 * entries intact. The comment in vdev_dtl_reassess() describes how we
1769 * excise the DTLs.
1770 */
1771static boolean_t
1772vdev_dtl_should_excise(vdev_t *vd)
1773{
1774	spa_t *spa = vd->vdev_spa;
1775	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1776
1777	ASSERT0(scn->scn_phys.scn_errors);
1778	ASSERT0(vd->vdev_children);
1779
1780	if (vd->vdev_resilver_txg == 0 ||
1781	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1782		return (B_TRUE);
1783
1784	/*
1785	 * When a resilver is initiated the scan will assign the scn_max_txg
1786	 * value to the highest txg value that exists in all DTLs. If this
1787	 * device's max DTL is not part of this scan (i.e. it is not in
1788	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1789	 * for excision.
1790	 */
1791	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1792		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1793		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1794		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1795		return (B_TRUE);
1796	}
1797	return (B_FALSE);
1798}
1799
1800/*
1801 * Reassess DTLs after a config change or scrub completion.
1802 */
1803void
1804vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1805{
1806	spa_t *spa = vd->vdev_spa;
1807	avl_tree_t reftree;
1808	int minref;
1809
1810	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1811
1812	for (int c = 0; c < vd->vdev_children; c++)
1813		vdev_dtl_reassess(vd->vdev_child[c], txg,
1814		    scrub_txg, scrub_done);
1815
1816	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1817		return;
1818
1819	if (vd->vdev_ops->vdev_op_leaf) {
1820		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1821
1822		mutex_enter(&vd->vdev_dtl_lock);
1823
1824		/*
1825		 * If we've completed a scan cleanly then determine
1826		 * if this vdev should remove any DTLs. We only want to
1827		 * excise regions on vdevs that were available during
1828		 * the entire duration of this scan.
1829		 */
1830		if (scrub_txg != 0 &&
1831		    (spa->spa_scrub_started ||
1832		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1833		    vdev_dtl_should_excise(vd)) {
1834			/*
1835			 * We completed a scrub up to scrub_txg.  If we
1836			 * did it without rebooting, then the scrub dtl
1837			 * will be valid, so excise the old region and
1838			 * fold in the scrub dtl.  Otherwise, leave the
1839			 * dtl as-is if there was an error.
1840			 *
1841			 * There's little trick here: to excise the beginning
1842			 * of the DTL_MISSING map, we put it into a reference
1843			 * tree and then add a segment with refcnt -1 that
1844			 * covers the range [0, scrub_txg).  This means
1845			 * that each txg in that range has refcnt -1 or 0.
1846			 * We then add DTL_SCRUB with a refcnt of 2, so that
1847			 * entries in the range [0, scrub_txg) will have a
1848			 * positive refcnt -- either 1 or 2.  We then convert
1849			 * the reference tree into the new DTL_MISSING map.
1850			 */
1851			space_reftree_create(&reftree);
1852			space_reftree_add_map(&reftree,
1853			    vd->vdev_dtl[DTL_MISSING], 1);
1854			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1855			space_reftree_add_map(&reftree,
1856			    vd->vdev_dtl[DTL_SCRUB], 2);
1857			space_reftree_generate_map(&reftree,
1858			    vd->vdev_dtl[DTL_MISSING], 1);
1859			space_reftree_destroy(&reftree);
1860		}
1861		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1862		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1863		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1864		if (scrub_done)
1865			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1866		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1867		if (!vdev_readable(vd))
1868			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1869		else
1870			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1871			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1872
1873		/*
1874		 * If the vdev was resilvering and no longer has any
1875		 * DTLs then reset its resilvering flag.
1876		 */
1877		if (vd->vdev_resilver_txg != 0 &&
1878		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1879		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1880			vd->vdev_resilver_txg = 0;
1881
1882		mutex_exit(&vd->vdev_dtl_lock);
1883
1884		if (txg != 0)
1885			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1886		return;
1887	}
1888
1889	mutex_enter(&vd->vdev_dtl_lock);
1890	for (int t = 0; t < DTL_TYPES; t++) {
1891		/* account for child's outage in parent's missing map */
1892		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1893		if (t == DTL_SCRUB)
1894			continue;			/* leaf vdevs only */
1895		if (t == DTL_PARTIAL)
1896			minref = 1;			/* i.e. non-zero */
1897		else if (vd->vdev_nparity != 0)
1898			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1899		else
1900			minref = vd->vdev_children;	/* any kind of mirror */
1901		space_reftree_create(&reftree);
1902		for (int c = 0; c < vd->vdev_children; c++) {
1903			vdev_t *cvd = vd->vdev_child[c];
1904			mutex_enter(&cvd->vdev_dtl_lock);
1905			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1906			mutex_exit(&cvd->vdev_dtl_lock);
1907		}
1908		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1909		space_reftree_destroy(&reftree);
1910	}
1911	mutex_exit(&vd->vdev_dtl_lock);
1912}
1913
1914int
1915vdev_dtl_load(vdev_t *vd)
1916{
1917	spa_t *spa = vd->vdev_spa;
1918	objset_t *mos = spa->spa_meta_objset;
1919	int error = 0;
1920
1921	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1922		ASSERT(!vd->vdev_ishole);
1923
1924		error = space_map_open(&vd->vdev_dtl_sm, mos,
1925		    vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1926		if (error)
1927			return (error);
1928		ASSERT(vd->vdev_dtl_sm != NULL);
1929
1930		mutex_enter(&vd->vdev_dtl_lock);
1931
1932		/*
1933		 * Now that we've opened the space_map we need to update
1934		 * the in-core DTL.
1935		 */
1936		space_map_update(vd->vdev_dtl_sm);
1937
1938		error = space_map_load(vd->vdev_dtl_sm,
1939		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1940		mutex_exit(&vd->vdev_dtl_lock);
1941
1942		return (error);
1943	}
1944
1945	for (int c = 0; c < vd->vdev_children; c++) {
1946		error = vdev_dtl_load(vd->vdev_child[c]);
1947		if (error != 0)
1948			break;
1949	}
1950
1951	return (error);
1952}
1953
1954void
1955vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
1956{
1957	spa_t *spa = vd->vdev_spa;
1958
1959	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
1960	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1961	    zapobj, tx));
1962}
1963
1964uint64_t
1965vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
1966{
1967	spa_t *spa = vd->vdev_spa;
1968	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
1969	    DMU_OT_NONE, 0, tx);
1970
1971	ASSERT(zap != 0);
1972	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1973	    zap, tx));
1974
1975	return (zap);
1976}
1977
1978void
1979vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
1980{
1981	if (vd->vdev_ops != &vdev_hole_ops &&
1982	    vd->vdev_ops != &vdev_missing_ops &&
1983	    vd->vdev_ops != &vdev_root_ops &&
1984	    !vd->vdev_top->vdev_removing) {
1985		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
1986			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
1987		}
1988		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
1989			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
1990		}
1991	}
1992	for (uint64_t i = 0; i < vd->vdev_children; i++) {
1993		vdev_construct_zaps(vd->vdev_child[i], tx);
1994	}
1995}
1996
1997void
1998vdev_dtl_sync(vdev_t *vd, uint64_t txg)
1999{
2000	spa_t *spa = vd->vdev_spa;
2001	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2002	objset_t *mos = spa->spa_meta_objset;
2003	range_tree_t *rtsync;
2004	kmutex_t rtlock;
2005	dmu_tx_t *tx;
2006	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2007
2008	ASSERT(!vd->vdev_ishole);
2009	ASSERT(vd->vdev_ops->vdev_op_leaf);
2010
2011	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2012
2013	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2014		mutex_enter(&vd->vdev_dtl_lock);
2015		space_map_free(vd->vdev_dtl_sm, tx);
2016		space_map_close(vd->vdev_dtl_sm);
2017		vd->vdev_dtl_sm = NULL;
2018		mutex_exit(&vd->vdev_dtl_lock);
2019
2020		/*
2021		 * We only destroy the leaf ZAP for detached leaves or for
2022		 * removed log devices. Removed data devices handle leaf ZAP
2023		 * cleanup later, once cancellation is no longer possible.
2024		 */
2025		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2026		    vd->vdev_top->vdev_islog)) {
2027			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2028			vd->vdev_leaf_zap = 0;
2029		}
2030
2031		dmu_tx_commit(tx);
2032		return;
2033	}
2034
2035	if (vd->vdev_dtl_sm == NULL) {
2036		uint64_t new_object;
2037
2038		new_object = space_map_alloc(mos, tx);
2039		VERIFY3U(new_object, !=, 0);
2040
2041		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2042		    0, -1ULL, 0, &vd->vdev_dtl_lock));
2043		ASSERT(vd->vdev_dtl_sm != NULL);
2044	}
2045
2046	mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
2047
2048	rtsync = range_tree_create(NULL, NULL, &rtlock);
2049
2050	mutex_enter(&rtlock);
2051
2052	mutex_enter(&vd->vdev_dtl_lock);
2053	range_tree_walk(rt, range_tree_add, rtsync);
2054	mutex_exit(&vd->vdev_dtl_lock);
2055
2056	space_map_truncate(vd->vdev_dtl_sm, tx);
2057	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2058	range_tree_vacate(rtsync, NULL, NULL);
2059
2060	range_tree_destroy(rtsync);
2061
2062	mutex_exit(&rtlock);
2063	mutex_destroy(&rtlock);
2064
2065	/*
2066	 * If the object for the space map has changed then dirty
2067	 * the top level so that we update the config.
2068	 */
2069	if (object != space_map_object(vd->vdev_dtl_sm)) {
2070		zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
2071		    "new object %llu", txg, spa_name(spa), object,
2072		    space_map_object(vd->vdev_dtl_sm));
2073		vdev_config_dirty(vd->vdev_top);
2074	}
2075
2076	dmu_tx_commit(tx);
2077
2078	mutex_enter(&vd->vdev_dtl_lock);
2079	space_map_update(vd->vdev_dtl_sm);
2080	mutex_exit(&vd->vdev_dtl_lock);
2081}
2082
2083/*
2084 * Determine whether the specified vdev can be offlined/detached/removed
2085 * without losing data.
2086 */
2087boolean_t
2088vdev_dtl_required(vdev_t *vd)
2089{
2090	spa_t *spa = vd->vdev_spa;
2091	vdev_t *tvd = vd->vdev_top;
2092	uint8_t cant_read = vd->vdev_cant_read;
2093	boolean_t required;
2094
2095	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2096
2097	if (vd == spa->spa_root_vdev || vd == tvd)
2098		return (B_TRUE);
2099
2100	/*
2101	 * Temporarily mark the device as unreadable, and then determine
2102	 * whether this results in any DTL outages in the top-level vdev.
2103	 * If not, we can safely offline/detach/remove the device.
2104	 */
2105	vd->vdev_cant_read = B_TRUE;
2106	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2107	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2108	vd->vdev_cant_read = cant_read;
2109	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2110
2111	if (!required && zio_injection_enabled)
2112		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2113
2114	return (required);
2115}
2116
2117/*
2118 * Determine if resilver is needed, and if so the txg range.
2119 */
2120boolean_t
2121vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2122{
2123	boolean_t needed = B_FALSE;
2124	uint64_t thismin = UINT64_MAX;
2125	uint64_t thismax = 0;
2126
2127	if (vd->vdev_children == 0) {
2128		mutex_enter(&vd->vdev_dtl_lock);
2129		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2130		    vdev_writeable(vd)) {
2131
2132			thismin = vdev_dtl_min(vd);
2133			thismax = vdev_dtl_max(vd);
2134			needed = B_TRUE;
2135		}
2136		mutex_exit(&vd->vdev_dtl_lock);
2137	} else {
2138		for (int c = 0; c < vd->vdev_children; c++) {
2139			vdev_t *cvd = vd->vdev_child[c];
2140			uint64_t cmin, cmax;
2141
2142			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2143				thismin = MIN(thismin, cmin);
2144				thismax = MAX(thismax, cmax);
2145				needed = B_TRUE;
2146			}
2147		}
2148	}
2149
2150	if (needed && minp) {
2151		*minp = thismin;
2152		*maxp = thismax;
2153	}
2154	return (needed);
2155}
2156
2157void
2158vdev_load(vdev_t *vd)
2159{
2160	/*
2161	 * Recursively load all children.
2162	 */
2163	for (int c = 0; c < vd->vdev_children; c++)
2164		vdev_load(vd->vdev_child[c]);
2165
2166	/*
2167	 * If this is a top-level vdev, initialize its metaslabs.
2168	 */
2169	if (vd == vd->vdev_top && !vd->vdev_ishole &&
2170	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2171	    vdev_metaslab_init(vd, 0) != 0))
2172		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2173		    VDEV_AUX_CORRUPT_DATA);
2174
2175	/*
2176	 * If this is a leaf vdev, load its DTL.
2177	 */
2178	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2179		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2180		    VDEV_AUX_CORRUPT_DATA);
2181}
2182
2183/*
2184 * The special vdev case is used for hot spares and l2cache devices.  Its
2185 * sole purpose it to set the vdev state for the associated vdev.  To do this,
2186 * we make sure that we can open the underlying device, then try to read the
2187 * label, and make sure that the label is sane and that it hasn't been
2188 * repurposed to another pool.
2189 */
2190int
2191vdev_validate_aux(vdev_t *vd)
2192{
2193	nvlist_t *label;
2194	uint64_t guid, version;
2195	uint64_t state;
2196
2197	if (!vdev_readable(vd))
2198		return (0);
2199
2200	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2201		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2202		    VDEV_AUX_CORRUPT_DATA);
2203		return (-1);
2204	}
2205
2206	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2207	    !SPA_VERSION_IS_SUPPORTED(version) ||
2208	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2209	    guid != vd->vdev_guid ||
2210	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2211		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2212		    VDEV_AUX_CORRUPT_DATA);
2213		nvlist_free(label);
2214		return (-1);
2215	}
2216
2217	/*
2218	 * We don't actually check the pool state here.  If it's in fact in
2219	 * use by another pool, we update this fact on the fly when requested.
2220	 */
2221	nvlist_free(label);
2222	return (0);
2223}
2224
2225void
2226vdev_remove(vdev_t *vd, uint64_t txg)
2227{
2228	spa_t *spa = vd->vdev_spa;
2229	objset_t *mos = spa->spa_meta_objset;
2230	dmu_tx_t *tx;
2231
2232	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2233	ASSERT(vd == vd->vdev_top);
2234	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2235
2236	if (vd->vdev_ms != NULL) {
2237		metaslab_group_t *mg = vd->vdev_mg;
2238
2239		metaslab_group_histogram_verify(mg);
2240		metaslab_class_histogram_verify(mg->mg_class);
2241
2242		for (int m = 0; m < vd->vdev_ms_count; m++) {
2243			metaslab_t *msp = vd->vdev_ms[m];
2244
2245			if (msp == NULL || msp->ms_sm == NULL)
2246				continue;
2247
2248			mutex_enter(&msp->ms_lock);
2249			/*
2250			 * If the metaslab was not loaded when the vdev
2251			 * was removed then the histogram accounting may
2252			 * not be accurate. Update the histogram information
2253			 * here so that we ensure that the metaslab group
2254			 * and metaslab class are up-to-date.
2255			 */
2256			metaslab_group_histogram_remove(mg, msp);
2257
2258			VERIFY0(space_map_allocated(msp->ms_sm));
2259			space_map_free(msp->ms_sm, tx);
2260			space_map_close(msp->ms_sm);
2261			msp->ms_sm = NULL;
2262			mutex_exit(&msp->ms_lock);
2263		}
2264
2265		metaslab_group_histogram_verify(mg);
2266		metaslab_class_histogram_verify(mg->mg_class);
2267		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2268			ASSERT0(mg->mg_histogram[i]);
2269
2270	}
2271
2272	if (vd->vdev_ms_array) {
2273		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2274		vd->vdev_ms_array = 0;
2275	}
2276
2277	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2278		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2279		vd->vdev_top_zap = 0;
2280	}
2281	dmu_tx_commit(tx);
2282}
2283
2284void
2285vdev_sync_done(vdev_t *vd, uint64_t txg)
2286{
2287	metaslab_t *msp;
2288	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2289
2290	ASSERT(!vd->vdev_ishole);
2291
2292	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2293		metaslab_sync_done(msp, txg);
2294
2295	if (reassess)
2296		metaslab_sync_reassess(vd->vdev_mg);
2297}
2298
2299void
2300vdev_sync(vdev_t *vd, uint64_t txg)
2301{
2302	spa_t *spa = vd->vdev_spa;
2303	vdev_t *lvd;
2304	metaslab_t *msp;
2305	dmu_tx_t *tx;
2306
2307	ASSERT(!vd->vdev_ishole);
2308
2309	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2310		ASSERT(vd == vd->vdev_top);
2311		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2312		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2313		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2314		ASSERT(vd->vdev_ms_array != 0);
2315		vdev_config_dirty(vd);
2316		dmu_tx_commit(tx);
2317	}
2318
2319	/*
2320	 * Remove the metadata associated with this vdev once it's empty.
2321	 */
2322	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2323		vdev_remove(vd, txg);
2324
2325	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2326		metaslab_sync(msp, txg);
2327		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2328	}
2329
2330	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2331		vdev_dtl_sync(lvd, txg);
2332
2333	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2334}
2335
2336uint64_t
2337vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2338{
2339	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2340}
2341
2342/*
2343 * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2344 * not be opened, and no I/O is attempted.
2345 */
2346int
2347vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2348{
2349	vdev_t *vd, *tvd;
2350
2351	spa_vdev_state_enter(spa, SCL_NONE);
2352
2353	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2354		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2355
2356	if (!vd->vdev_ops->vdev_op_leaf)
2357		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2358
2359	tvd = vd->vdev_top;
2360
2361	/*
2362	 * We don't directly use the aux state here, but if we do a
2363	 * vdev_reopen(), we need this value to be present to remember why we
2364	 * were faulted.
2365	 */
2366	vd->vdev_label_aux = aux;
2367
2368	/*
2369	 * Faulted state takes precedence over degraded.
2370	 */
2371	vd->vdev_delayed_close = B_FALSE;
2372	vd->vdev_faulted = 1ULL;
2373	vd->vdev_degraded = 0ULL;
2374	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2375
2376	/*
2377	 * If this device has the only valid copy of the data, then
2378	 * back off and simply mark the vdev as degraded instead.
2379	 */
2380	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2381		vd->vdev_degraded = 1ULL;
2382		vd->vdev_faulted = 0ULL;
2383
2384		/*
2385		 * If we reopen the device and it's not dead, only then do we
2386		 * mark it degraded.
2387		 */
2388		vdev_reopen(tvd);
2389
2390		if (vdev_readable(vd))
2391			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2392	}
2393
2394	return (spa_vdev_state_exit(spa, vd, 0));
2395}
2396
2397/*
2398 * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2399 * user that something is wrong.  The vdev continues to operate as normal as far
2400 * as I/O is concerned.
2401 */
2402int
2403vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2404{
2405	vdev_t *vd;
2406
2407	spa_vdev_state_enter(spa, SCL_NONE);
2408
2409	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2410		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2411
2412	if (!vd->vdev_ops->vdev_op_leaf)
2413		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2414
2415	/*
2416	 * If the vdev is already faulted, then don't do anything.
2417	 */
2418	if (vd->vdev_faulted || vd->vdev_degraded)
2419		return (spa_vdev_state_exit(spa, NULL, 0));
2420
2421	vd->vdev_degraded = 1ULL;
2422	if (!vdev_is_dead(vd))
2423		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2424		    aux);
2425
2426	return (spa_vdev_state_exit(spa, vd, 0));
2427}
2428
2429/*
2430 * Online the given vdev.
2431 *
2432 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2433 * spare device should be detached when the device finishes resilvering.
2434 * Second, the online should be treated like a 'test' online case, so no FMA
2435 * events are generated if the device fails to open.
2436 */
2437int
2438vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2439{
2440	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2441	boolean_t postevent = B_FALSE;
2442
2443	spa_vdev_state_enter(spa, SCL_NONE);
2444
2445	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2446		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2447
2448	if (!vd->vdev_ops->vdev_op_leaf)
2449		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2450
2451	postevent =
2452	    (vd->vdev_offline == B_TRUE || vd->vdev_tmpoffline == B_TRUE) ?
2453	    B_TRUE : B_FALSE;
2454
2455	tvd = vd->vdev_top;
2456	vd->vdev_offline = B_FALSE;
2457	vd->vdev_tmpoffline = B_FALSE;
2458	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2459	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2460
2461	/* XXX - L2ARC 1.0 does not support expansion */
2462	if (!vd->vdev_aux) {
2463		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2464			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2465	}
2466
2467	vdev_reopen(tvd);
2468	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2469
2470	if (!vd->vdev_aux) {
2471		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2472			pvd->vdev_expanding = B_FALSE;
2473	}
2474
2475	if (newstate)
2476		*newstate = vd->vdev_state;
2477	if ((flags & ZFS_ONLINE_UNSPARE) &&
2478	    !vdev_is_dead(vd) && vd->vdev_parent &&
2479	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2480	    vd->vdev_parent->vdev_child[0] == vd)
2481		vd->vdev_unspare = B_TRUE;
2482
2483	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2484
2485		/* XXX - L2ARC 1.0 does not support expansion */
2486		if (vd->vdev_aux)
2487			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2488		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2489	}
2490
2491	if (postevent)
2492		spa_event_notify(spa, vd, ESC_ZFS_VDEV_ONLINE);
2493
2494	return (spa_vdev_state_exit(spa, vd, 0));
2495}
2496
2497static int
2498vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2499{
2500	vdev_t *vd, *tvd;
2501	int error = 0;
2502	uint64_t generation;
2503	metaslab_group_t *mg;
2504
2505top:
2506	spa_vdev_state_enter(spa, SCL_ALLOC);
2507
2508	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2509		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2510
2511	if (!vd->vdev_ops->vdev_op_leaf)
2512		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2513
2514	tvd = vd->vdev_top;
2515	mg = tvd->vdev_mg;
2516	generation = spa->spa_config_generation + 1;
2517
2518	/*
2519	 * If the device isn't already offline, try to offline it.
2520	 */
2521	if (!vd->vdev_offline) {
2522		/*
2523		 * If this device has the only valid copy of some data,
2524		 * don't allow it to be offlined. Log devices are always
2525		 * expendable.
2526		 */
2527		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2528		    vdev_dtl_required(vd))
2529			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2530
2531		/*
2532		 * If the top-level is a slog and it has had allocations
2533		 * then proceed.  We check that the vdev's metaslab group
2534		 * is not NULL since it's possible that we may have just
2535		 * added this vdev but not yet initialized its metaslabs.
2536		 */
2537		if (tvd->vdev_islog && mg != NULL) {
2538			/*
2539			 * Prevent any future allocations.
2540			 */
2541			metaslab_group_passivate(mg);
2542			(void) spa_vdev_state_exit(spa, vd, 0);
2543
2544			error = spa_offline_log(spa);
2545
2546			spa_vdev_state_enter(spa, SCL_ALLOC);
2547
2548			/*
2549			 * Check to see if the config has changed.
2550			 */
2551			if (error || generation != spa->spa_config_generation) {
2552				metaslab_group_activate(mg);
2553				if (error)
2554					return (spa_vdev_state_exit(spa,
2555					    vd, error));
2556				(void) spa_vdev_state_exit(spa, vd, 0);
2557				goto top;
2558			}
2559			ASSERT0(tvd->vdev_stat.vs_alloc);
2560		}
2561
2562		/*
2563		 * Offline this device and reopen its top-level vdev.
2564		 * If the top-level vdev is a log device then just offline
2565		 * it. Otherwise, if this action results in the top-level
2566		 * vdev becoming unusable, undo it and fail the request.
2567		 */
2568		vd->vdev_offline = B_TRUE;
2569		vdev_reopen(tvd);
2570
2571		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2572		    vdev_is_dead(tvd)) {
2573			vd->vdev_offline = B_FALSE;
2574			vdev_reopen(tvd);
2575			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2576		}
2577
2578		/*
2579		 * Add the device back into the metaslab rotor so that
2580		 * once we online the device it's open for business.
2581		 */
2582		if (tvd->vdev_islog && mg != NULL)
2583			metaslab_group_activate(mg);
2584	}
2585
2586	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2587
2588	return (spa_vdev_state_exit(spa, vd, 0));
2589}
2590
2591int
2592vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2593{
2594	int error;
2595
2596	mutex_enter(&spa->spa_vdev_top_lock);
2597	error = vdev_offline_locked(spa, guid, flags);
2598	mutex_exit(&spa->spa_vdev_top_lock);
2599
2600	return (error);
2601}
2602
2603/*
2604 * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2605 * vdev_offline(), we assume the spa config is locked.  We also clear all
2606 * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2607 */
2608void
2609vdev_clear(spa_t *spa, vdev_t *vd)
2610{
2611	vdev_t *rvd = spa->spa_root_vdev;
2612
2613	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2614
2615	if (vd == NULL)
2616		vd = rvd;
2617
2618	vd->vdev_stat.vs_read_errors = 0;
2619	vd->vdev_stat.vs_write_errors = 0;
2620	vd->vdev_stat.vs_checksum_errors = 0;
2621
2622	for (int c = 0; c < vd->vdev_children; c++)
2623		vdev_clear(spa, vd->vdev_child[c]);
2624
2625	/*
2626	 * If we're in the FAULTED state or have experienced failed I/O, then
2627	 * clear the persistent state and attempt to reopen the device.  We
2628	 * also mark the vdev config dirty, so that the new faulted state is
2629	 * written out to disk.
2630	 */
2631	if (vd->vdev_faulted || vd->vdev_degraded ||
2632	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2633
2634		/*
2635		 * When reopening in reponse to a clear event, it may be due to
2636		 * a fmadm repair request.  In this case, if the device is
2637		 * still broken, we want to still post the ereport again.
2638		 */
2639		vd->vdev_forcefault = B_TRUE;
2640
2641		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2642		vd->vdev_cant_read = B_FALSE;
2643		vd->vdev_cant_write = B_FALSE;
2644
2645		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2646
2647		vd->vdev_forcefault = B_FALSE;
2648
2649		if (vd != rvd && vdev_writeable(vd->vdev_top))
2650			vdev_state_dirty(vd->vdev_top);
2651
2652		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2653			spa_async_request(spa, SPA_ASYNC_RESILVER);
2654
2655		spa_event_notify(spa, vd, ESC_ZFS_VDEV_CLEAR);
2656	}
2657
2658	/*
2659	 * When clearing a FMA-diagnosed fault, we always want to
2660	 * unspare the device, as we assume that the original spare was
2661	 * done in response to the FMA fault.
2662	 */
2663	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2664	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2665	    vd->vdev_parent->vdev_child[0] == vd)
2666		vd->vdev_unspare = B_TRUE;
2667}
2668
2669boolean_t
2670vdev_is_dead(vdev_t *vd)
2671{
2672	/*
2673	 * Holes and missing devices are always considered "dead".
2674	 * This simplifies the code since we don't have to check for
2675	 * these types of devices in the various code paths.
2676	 * Instead we rely on the fact that we skip over dead devices
2677	 * before issuing I/O to them.
2678	 */
2679	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2680	    vd->vdev_ops == &vdev_missing_ops);
2681}
2682
2683boolean_t
2684vdev_readable(vdev_t *vd)
2685{
2686	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2687}
2688
2689boolean_t
2690vdev_writeable(vdev_t *vd)
2691{
2692	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2693}
2694
2695boolean_t
2696vdev_allocatable(vdev_t *vd)
2697{
2698	uint64_t state = vd->vdev_state;
2699
2700	/*
2701	 * We currently allow allocations from vdevs which may be in the
2702	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2703	 * fails to reopen then we'll catch it later when we're holding
2704	 * the proper locks.  Note that we have to get the vdev state
2705	 * in a local variable because although it changes atomically,
2706	 * we're asking two separate questions about it.
2707	 */
2708	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2709	    !vd->vdev_cant_write && !vd->vdev_ishole &&
2710	    vd->vdev_mg->mg_initialized);
2711}
2712
2713boolean_t
2714vdev_accessible(vdev_t *vd, zio_t *zio)
2715{
2716	ASSERT(zio->io_vd == vd);
2717
2718	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2719		return (B_FALSE);
2720
2721	if (zio->io_type == ZIO_TYPE_READ)
2722		return (!vd->vdev_cant_read);
2723
2724	if (zio->io_type == ZIO_TYPE_WRITE)
2725		return (!vd->vdev_cant_write);
2726
2727	return (B_TRUE);
2728}
2729
2730/*
2731 * Get statistics for the given vdev.
2732 */
2733void
2734vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2735{
2736	spa_t *spa = vd->vdev_spa;
2737	vdev_t *rvd = spa->spa_root_vdev;
2738	vdev_t *tvd = vd->vdev_top;
2739
2740	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2741
2742	mutex_enter(&vd->vdev_stat_lock);
2743	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2744	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2745	vs->vs_state = vd->vdev_state;
2746	vs->vs_rsize = vdev_get_min_asize(vd);
2747	if (vd->vdev_ops->vdev_op_leaf)
2748		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2749	/*
2750	 * Report expandable space on top-level, non-auxillary devices only.
2751	 * The expandable space is reported in terms of metaslab sized units
2752	 * since that determines how much space the pool can expand.
2753	 */
2754	if (vd->vdev_aux == NULL && tvd != NULL) {
2755		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize,
2756		    1ULL << tvd->vdev_ms_shift);
2757	}
2758	if (vd->vdev_aux == NULL && vd == vd->vdev_top && !vd->vdev_ishole) {
2759		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
2760	}
2761
2762	/*
2763	 * If we're getting stats on the root vdev, aggregate the I/O counts
2764	 * over all top-level vdevs (i.e. the direct children of the root).
2765	 */
2766	if (vd == rvd) {
2767		for (int c = 0; c < rvd->vdev_children; c++) {
2768			vdev_t *cvd = rvd->vdev_child[c];
2769			vdev_stat_t *cvs = &cvd->vdev_stat;
2770
2771			for (int t = 0; t < ZIO_TYPES; t++) {
2772				vs->vs_ops[t] += cvs->vs_ops[t];
2773				vs->vs_bytes[t] += cvs->vs_bytes[t];
2774			}
2775			cvs->vs_scan_removing = cvd->vdev_removing;
2776		}
2777	}
2778	mutex_exit(&vd->vdev_stat_lock);
2779}
2780
2781void
2782vdev_clear_stats(vdev_t *vd)
2783{
2784	mutex_enter(&vd->vdev_stat_lock);
2785	vd->vdev_stat.vs_space = 0;
2786	vd->vdev_stat.vs_dspace = 0;
2787	vd->vdev_stat.vs_alloc = 0;
2788	mutex_exit(&vd->vdev_stat_lock);
2789}
2790
2791void
2792vdev_scan_stat_init(vdev_t *vd)
2793{
2794	vdev_stat_t *vs = &vd->vdev_stat;
2795
2796	for (int c = 0; c < vd->vdev_children; c++)
2797		vdev_scan_stat_init(vd->vdev_child[c]);
2798
2799	mutex_enter(&vd->vdev_stat_lock);
2800	vs->vs_scan_processed = 0;
2801	mutex_exit(&vd->vdev_stat_lock);
2802}
2803
2804void
2805vdev_stat_update(zio_t *zio, uint64_t psize)
2806{
2807	spa_t *spa = zio->io_spa;
2808	vdev_t *rvd = spa->spa_root_vdev;
2809	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2810	vdev_t *pvd;
2811	uint64_t txg = zio->io_txg;
2812	vdev_stat_t *vs = &vd->vdev_stat;
2813	zio_type_t type = zio->io_type;
2814	int flags = zio->io_flags;
2815
2816	/*
2817	 * If this i/o is a gang leader, it didn't do any actual work.
2818	 */
2819	if (zio->io_gang_tree)
2820		return;
2821
2822	if (zio->io_error == 0) {
2823		/*
2824		 * If this is a root i/o, don't count it -- we've already
2825		 * counted the top-level vdevs, and vdev_get_stats() will
2826		 * aggregate them when asked.  This reduces contention on
2827		 * the root vdev_stat_lock and implicitly handles blocks
2828		 * that compress away to holes, for which there is no i/o.
2829		 * (Holes never create vdev children, so all the counters
2830		 * remain zero, which is what we want.)
2831		 *
2832		 * Note: this only applies to successful i/o (io_error == 0)
2833		 * because unlike i/o counts, errors are not additive.
2834		 * When reading a ditto block, for example, failure of
2835		 * one top-level vdev does not imply a root-level error.
2836		 */
2837		if (vd == rvd)
2838			return;
2839
2840		ASSERT(vd == zio->io_vd);
2841
2842		if (flags & ZIO_FLAG_IO_BYPASS)
2843			return;
2844
2845		mutex_enter(&vd->vdev_stat_lock);
2846
2847		if (flags & ZIO_FLAG_IO_REPAIR) {
2848			if (flags & ZIO_FLAG_SCAN_THREAD) {
2849				dsl_scan_phys_t *scn_phys =
2850				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2851				uint64_t *processed = &scn_phys->scn_processed;
2852
2853				/* XXX cleanup? */
2854				if (vd->vdev_ops->vdev_op_leaf)
2855					atomic_add_64(processed, psize);
2856				vs->vs_scan_processed += psize;
2857			}
2858
2859			if (flags & ZIO_FLAG_SELF_HEAL)
2860				vs->vs_self_healed += psize;
2861		}
2862
2863		vs->vs_ops[type]++;
2864		vs->vs_bytes[type] += psize;
2865
2866		mutex_exit(&vd->vdev_stat_lock);
2867		return;
2868	}
2869
2870	if (flags & ZIO_FLAG_SPECULATIVE)
2871		return;
2872
2873	/*
2874	 * If this is an I/O error that is going to be retried, then ignore the
2875	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2876	 * hard errors, when in reality they can happen for any number of
2877	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2878	 */
2879	if (zio->io_error == EIO &&
2880	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2881		return;
2882
2883	/*
2884	 * Intent logs writes won't propagate their error to the root
2885	 * I/O so don't mark these types of failures as pool-level
2886	 * errors.
2887	 */
2888	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2889		return;
2890
2891	mutex_enter(&vd->vdev_stat_lock);
2892	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2893		if (zio->io_error == ECKSUM)
2894			vs->vs_checksum_errors++;
2895		else
2896			vs->vs_read_errors++;
2897	}
2898	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2899		vs->vs_write_errors++;
2900	mutex_exit(&vd->vdev_stat_lock);
2901
2902	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2903	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2904	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2905	    spa->spa_claiming)) {
2906		/*
2907		 * This is either a normal write (not a repair), or it's
2908		 * a repair induced by the scrub thread, or it's a repair
2909		 * made by zil_claim() during spa_load() in the first txg.
2910		 * In the normal case, we commit the DTL change in the same
2911		 * txg as the block was born.  In the scrub-induced repair
2912		 * case, we know that scrubs run in first-pass syncing context,
2913		 * so we commit the DTL change in spa_syncing_txg(spa).
2914		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2915		 *
2916		 * We currently do not make DTL entries for failed spontaneous
2917		 * self-healing writes triggered by normal (non-scrubbing)
2918		 * reads, because we have no transactional context in which to
2919		 * do so -- and it's not clear that it'd be desirable anyway.
2920		 */
2921		if (vd->vdev_ops->vdev_op_leaf) {
2922			uint64_t commit_txg = txg;
2923			if (flags & ZIO_FLAG_SCAN_THREAD) {
2924				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2925				ASSERT(spa_sync_pass(spa) == 1);
2926				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2927				commit_txg = spa_syncing_txg(spa);
2928			} else if (spa->spa_claiming) {
2929				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2930				commit_txg = spa_first_txg(spa);
2931			}
2932			ASSERT(commit_txg >= spa_syncing_txg(spa));
2933			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2934				return;
2935			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2936				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2937			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2938		}
2939		if (vd != rvd)
2940			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2941	}
2942}
2943
2944/*
2945 * Update the in-core space usage stats for this vdev, its metaslab class,
2946 * and the root vdev.
2947 */
2948void
2949vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2950    int64_t space_delta)
2951{
2952	int64_t dspace_delta = space_delta;
2953	spa_t *spa = vd->vdev_spa;
2954	vdev_t *rvd = spa->spa_root_vdev;
2955	metaslab_group_t *mg = vd->vdev_mg;
2956	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2957
2958	ASSERT(vd == vd->vdev_top);
2959
2960	/*
2961	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2962	 * factor.  We must calculate this here and not at the root vdev
2963	 * because the root vdev's psize-to-asize is simply the max of its
2964	 * childrens', thus not accurate enough for us.
2965	 */
2966	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2967	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2968	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2969	    vd->vdev_deflate_ratio;
2970
2971	mutex_enter(&vd->vdev_stat_lock);
2972	vd->vdev_stat.vs_alloc += alloc_delta;
2973	vd->vdev_stat.vs_space += space_delta;
2974	vd->vdev_stat.vs_dspace += dspace_delta;
2975	mutex_exit(&vd->vdev_stat_lock);
2976
2977	if (mc == spa_normal_class(spa)) {
2978		mutex_enter(&rvd->vdev_stat_lock);
2979		rvd->vdev_stat.vs_alloc += alloc_delta;
2980		rvd->vdev_stat.vs_space += space_delta;
2981		rvd->vdev_stat.vs_dspace += dspace_delta;
2982		mutex_exit(&rvd->vdev_stat_lock);
2983	}
2984
2985	if (mc != NULL) {
2986		ASSERT(rvd == vd->vdev_parent);
2987		ASSERT(vd->vdev_ms_count != 0);
2988
2989		metaslab_class_space_update(mc,
2990		    alloc_delta, defer_delta, space_delta, dspace_delta);
2991	}
2992}
2993
2994/*
2995 * Mark a top-level vdev's config as dirty, placing it on the dirty list
2996 * so that it will be written out next time the vdev configuration is synced.
2997 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
2998 */
2999void
3000vdev_config_dirty(vdev_t *vd)
3001{
3002	spa_t *spa = vd->vdev_spa;
3003	vdev_t *rvd = spa->spa_root_vdev;
3004	int c;
3005
3006	ASSERT(spa_writeable(spa));
3007
3008	/*
3009	 * If this is an aux vdev (as with l2cache and spare devices), then we
3010	 * update the vdev config manually and set the sync flag.
3011	 */
3012	if (vd->vdev_aux != NULL) {
3013		spa_aux_vdev_t *sav = vd->vdev_aux;
3014		nvlist_t **aux;
3015		uint_t naux;
3016
3017		for (c = 0; c < sav->sav_count; c++) {
3018			if (sav->sav_vdevs[c] == vd)
3019				break;
3020		}
3021
3022		if (c == sav->sav_count) {
3023			/*
3024			 * We're being removed.  There's nothing more to do.
3025			 */
3026			ASSERT(sav->sav_sync == B_TRUE);
3027			return;
3028		}
3029
3030		sav->sav_sync = B_TRUE;
3031
3032		if (nvlist_lookup_nvlist_array(sav->sav_config,
3033		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3034			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3035			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3036		}
3037
3038		ASSERT(c < naux);
3039
3040		/*
3041		 * Setting the nvlist in the middle if the array is a little
3042		 * sketchy, but it will work.
3043		 */
3044		nvlist_free(aux[c]);
3045		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3046
3047		return;
3048	}
3049
3050	/*
3051	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3052	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3053	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3054	 * so this is sufficient to ensure mutual exclusion.
3055	 */
3056	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3057	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3058	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3059
3060	if (vd == rvd) {
3061		for (c = 0; c < rvd->vdev_children; c++)
3062			vdev_config_dirty(rvd->vdev_child[c]);
3063	} else {
3064		ASSERT(vd == vd->vdev_top);
3065
3066		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3067		    !vd->vdev_ishole)
3068			list_insert_head(&spa->spa_config_dirty_list, vd);
3069	}
3070}
3071
3072void
3073vdev_config_clean(vdev_t *vd)
3074{
3075	spa_t *spa = vd->vdev_spa;
3076
3077	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3078	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3079	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3080
3081	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3082	list_remove(&spa->spa_config_dirty_list, vd);
3083}
3084
3085/*
3086 * Mark a top-level vdev's state as dirty, so that the next pass of
3087 * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3088 * the state changes from larger config changes because they require
3089 * much less locking, and are often needed for administrative actions.
3090 */
3091void
3092vdev_state_dirty(vdev_t *vd)
3093{
3094	spa_t *spa = vd->vdev_spa;
3095
3096	ASSERT(spa_writeable(spa));
3097	ASSERT(vd == vd->vdev_top);
3098
3099	/*
3100	 * The state list is protected by the SCL_STATE lock.  The caller
3101	 * must either hold SCL_STATE as writer, or must be the sync thread
3102	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3103	 * so this is sufficient to ensure mutual exclusion.
3104	 */
3105	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3106	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3107	    spa_config_held(spa, SCL_STATE, RW_READER)));
3108
3109	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
3110		list_insert_head(&spa->spa_state_dirty_list, vd);
3111}
3112
3113void
3114vdev_state_clean(vdev_t *vd)
3115{
3116	spa_t *spa = vd->vdev_spa;
3117
3118	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3119	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3120	    spa_config_held(spa, SCL_STATE, RW_READER)));
3121
3122	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3123	list_remove(&spa->spa_state_dirty_list, vd);
3124}
3125
3126/*
3127 * Propagate vdev state up from children to parent.
3128 */
3129void
3130vdev_propagate_state(vdev_t *vd)
3131{
3132	spa_t *spa = vd->vdev_spa;
3133	vdev_t *rvd = spa->spa_root_vdev;
3134	int degraded = 0, faulted = 0;
3135	int corrupted = 0;
3136	vdev_t *child;
3137
3138	if (vd->vdev_children > 0) {
3139		for (int c = 0; c < vd->vdev_children; c++) {
3140			child = vd->vdev_child[c];
3141
3142			/*
3143			 * Don't factor holes into the decision.
3144			 */
3145			if (child->vdev_ishole)
3146				continue;
3147
3148			if (!vdev_readable(child) ||
3149			    (!vdev_writeable(child) && spa_writeable(spa))) {
3150				/*
3151				 * Root special: if there is a top-level log
3152				 * device, treat the root vdev as if it were
3153				 * degraded.
3154				 */
3155				if (child->vdev_islog && vd == rvd)
3156					degraded++;
3157				else
3158					faulted++;
3159			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3160				degraded++;
3161			}
3162
3163			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3164				corrupted++;
3165		}
3166
3167		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3168
3169		/*
3170		 * Root special: if there is a top-level vdev that cannot be
3171		 * opened due to corrupted metadata, then propagate the root
3172		 * vdev's aux state as 'corrupt' rather than 'insufficient
3173		 * replicas'.
3174		 */
3175		if (corrupted && vd == rvd &&
3176		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3177			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3178			    VDEV_AUX_CORRUPT_DATA);
3179	}
3180
3181	if (vd->vdev_parent)
3182		vdev_propagate_state(vd->vdev_parent);
3183}
3184
3185/*
3186 * Set a vdev's state.  If this is during an open, we don't update the parent
3187 * state, because we're in the process of opening children depth-first.
3188 * Otherwise, we propagate the change to the parent.
3189 *
3190 * If this routine places a device in a faulted state, an appropriate ereport is
3191 * generated.
3192 */
3193void
3194vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3195{
3196	uint64_t save_state;
3197	spa_t *spa = vd->vdev_spa;
3198
3199	if (state == vd->vdev_state) {
3200		vd->vdev_stat.vs_aux = aux;
3201		return;
3202	}
3203
3204	save_state = vd->vdev_state;
3205
3206	vd->vdev_state = state;
3207	vd->vdev_stat.vs_aux = aux;
3208
3209	/*
3210	 * If we are setting the vdev state to anything but an open state, then
3211	 * always close the underlying device unless the device has requested
3212	 * a delayed close (i.e. we're about to remove or fault the device).
3213	 * Otherwise, we keep accessible but invalid devices open forever.
3214	 * We don't call vdev_close() itself, because that implies some extra
3215	 * checks (offline, etc) that we don't want here.  This is limited to
3216	 * leaf devices, because otherwise closing the device will affect other
3217	 * children.
3218	 */
3219	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3220	    vd->vdev_ops->vdev_op_leaf)
3221		vd->vdev_ops->vdev_op_close(vd);
3222
3223	/*
3224	 * If we have brought this vdev back into service, we need
3225	 * to notify fmd so that it can gracefully repair any outstanding
3226	 * cases due to a missing device.  We do this in all cases, even those
3227	 * that probably don't correlate to a repaired fault.  This is sure to
3228	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3229	 * this is a transient state it's OK, as the retire agent will
3230	 * double-check the state of the vdev before repairing it.
3231	 */
3232	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3233	    vd->vdev_prevstate != state)
3234		zfs_post_state_change(spa, vd);
3235
3236	if (vd->vdev_removed &&
3237	    state == VDEV_STATE_CANT_OPEN &&
3238	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3239		/*
3240		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3241		 * device was previously marked removed and someone attempted to
3242		 * reopen it.  If this failed due to a nonexistent device, then
3243		 * keep the device in the REMOVED state.  We also let this be if
3244		 * it is one of our special test online cases, which is only
3245		 * attempting to online the device and shouldn't generate an FMA
3246		 * fault.
3247		 */
3248		vd->vdev_state = VDEV_STATE_REMOVED;
3249		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3250	} else if (state == VDEV_STATE_REMOVED) {
3251		vd->vdev_removed = B_TRUE;
3252	} else if (state == VDEV_STATE_CANT_OPEN) {
3253		/*
3254		 * If we fail to open a vdev during an import or recovery, we
3255		 * mark it as "not available", which signifies that it was
3256		 * never there to begin with.  Failure to open such a device
3257		 * is not considered an error.
3258		 */
3259		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3260		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3261		    vd->vdev_ops->vdev_op_leaf)
3262			vd->vdev_not_present = 1;
3263
3264		/*
3265		 * Post the appropriate ereport.  If the 'prevstate' field is
3266		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3267		 * that this is part of a vdev_reopen().  In this case, we don't
3268		 * want to post the ereport if the device was already in the
3269		 * CANT_OPEN state beforehand.
3270		 *
3271		 * If the 'checkremove' flag is set, then this is an attempt to
3272		 * online the device in response to an insertion event.  If we
3273		 * hit this case, then we have detected an insertion event for a
3274		 * faulted or offline device that wasn't in the removed state.
3275		 * In this scenario, we don't post an ereport because we are
3276		 * about to replace the device, or attempt an online with
3277		 * vdev_forcefault, which will generate the fault for us.
3278		 */
3279		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3280		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3281		    vd != spa->spa_root_vdev) {
3282			const char *class;
3283
3284			switch (aux) {
3285			case VDEV_AUX_OPEN_FAILED:
3286				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3287				break;
3288			case VDEV_AUX_CORRUPT_DATA:
3289				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3290				break;
3291			case VDEV_AUX_NO_REPLICAS:
3292				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3293				break;
3294			case VDEV_AUX_BAD_GUID_SUM:
3295				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3296				break;
3297			case VDEV_AUX_TOO_SMALL:
3298				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3299				break;
3300			case VDEV_AUX_BAD_LABEL:
3301				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3302				break;
3303			default:
3304				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3305			}
3306
3307			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3308		}
3309
3310		/* Erase any notion of persistent removed state */
3311		vd->vdev_removed = B_FALSE;
3312	} else {
3313		vd->vdev_removed = B_FALSE;
3314	}
3315
3316	if (!isopen && vd->vdev_parent)
3317		vdev_propagate_state(vd->vdev_parent);
3318}
3319
3320/*
3321 * Check the vdev configuration to ensure that it's capable of supporting
3322 * a root pool. Currently, we do not support RAID-Z or partial configuration.
3323 * In addition, only a single top-level vdev is allowed and none of the leaves
3324 * can be wholedisks.
3325 */
3326boolean_t
3327vdev_is_bootable(vdev_t *vd)
3328{
3329	if (!vd->vdev_ops->vdev_op_leaf) {
3330		char *vdev_type = vd->vdev_ops->vdev_op_type;
3331
3332		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3333		    vd->vdev_children > 1) {
3334			return (B_FALSE);
3335		} else if (strcmp(vdev_type, VDEV_TYPE_RAIDZ) == 0 ||
3336		    strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3337			return (B_FALSE);
3338		}
3339	}
3340
3341	for (int c = 0; c < vd->vdev_children; c++) {
3342		if (!vdev_is_bootable(vd->vdev_child[c]))
3343			return (B_FALSE);
3344	}
3345	return (B_TRUE);
3346}
3347
3348/*
3349 * Load the state from the original vdev tree (ovd) which
3350 * we've retrieved from the MOS config object. If the original
3351 * vdev was offline or faulted then we transfer that state to the
3352 * device in the current vdev tree (nvd).
3353 */
3354void
3355vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3356{
3357	spa_t *spa = nvd->vdev_spa;
3358
3359	ASSERT(nvd->vdev_top->vdev_islog);
3360	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3361	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3362
3363	for (int c = 0; c < nvd->vdev_children; c++)
3364		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3365
3366	if (nvd->vdev_ops->vdev_op_leaf) {
3367		/*
3368		 * Restore the persistent vdev state
3369		 */
3370		nvd->vdev_offline = ovd->vdev_offline;
3371		nvd->vdev_faulted = ovd->vdev_faulted;
3372		nvd->vdev_degraded = ovd->vdev_degraded;
3373		nvd->vdev_removed = ovd->vdev_removed;
3374	}
3375}
3376
3377/*
3378 * Determine if a log device has valid content.  If the vdev was
3379 * removed or faulted in the MOS config then we know that
3380 * the content on the log device has already been written to the pool.
3381 */
3382boolean_t
3383vdev_log_state_valid(vdev_t *vd)
3384{
3385	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3386	    !vd->vdev_removed)
3387		return (B_TRUE);
3388
3389	for (int c = 0; c < vd->vdev_children; c++)
3390		if (vdev_log_state_valid(vd->vdev_child[c]))
3391			return (B_TRUE);
3392
3393	return (B_FALSE);
3394}
3395
3396/*
3397 * Expand a vdev if possible.
3398 */
3399void
3400vdev_expand(vdev_t *vd, uint64_t txg)
3401{
3402	ASSERT(vd->vdev_top == vd);
3403	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3404
3405	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3406		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3407		vdev_config_dirty(vd);
3408	}
3409}
3410
3411/*
3412 * Split a vdev.
3413 */
3414void
3415vdev_split(vdev_t *vd)
3416{
3417	vdev_t *cvd, *pvd = vd->vdev_parent;
3418
3419	vdev_remove_child(pvd, vd);
3420	vdev_compact_children(pvd);
3421
3422	cvd = pvd->vdev_child[0];
3423	if (pvd->vdev_children == 1) {
3424		vdev_remove_parent(cvd);
3425		cvd->vdev_splitting = B_TRUE;
3426	}
3427	vdev_propagate_state(cvd);
3428}
3429
3430void
3431vdev_deadman(vdev_t *vd)
3432{
3433	for (int c = 0; c < vd->vdev_children; c++) {
3434		vdev_t *cvd = vd->vdev_child[c];
3435
3436		vdev_deadman(cvd);
3437	}
3438
3439	if (vd->vdev_ops->vdev_op_leaf) {
3440		vdev_queue_t *vq = &vd->vdev_queue;
3441
3442		mutex_enter(&vq->vq_lock);
3443		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3444			spa_t *spa = vd->vdev_spa;
3445			zio_t *fio;
3446			uint64_t delta;
3447
3448			/*
3449			 * Look at the head of all the pending queues,
3450			 * if any I/O has been outstanding for longer than
3451			 * the spa_deadman_synctime we panic the system.
3452			 */
3453			fio = avl_first(&vq->vq_active_tree);
3454			delta = gethrtime() - fio->io_timestamp;
3455			if (delta > spa_deadman_synctime(spa)) {
3456				zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3457				    "delta %lluns, last io %lluns",
3458				    fio->io_timestamp, delta,
3459				    vq->vq_io_complete_ts);
3460				fm_panic("I/O to pool '%s' appears to be "
3461				    "hung.", spa_name(spa));
3462			}
3463		}
3464		mutex_exit(&vq->vq_lock);
3465	}
3466}
3467