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