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