1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22/*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
25 * Copyright 2017 Nexenta Systems, Inc.
26 * Copyright (c) 2014 Integros [integros.com]
27 * Copyright 2016 Toomas Soome <tsoome@me.com>
28 * Copyright 2017 Joyent, Inc.
29 */
30
31#include <sys/zfs_context.h>
32#include <sys/fm/fs/zfs.h>
33#include <sys/spa.h>
34#include <sys/spa_impl.h>
35#include <sys/dmu.h>
36#include <sys/dmu_tx.h>
37#include <sys/vdev_impl.h>
38#include <sys/uberblock_impl.h>
39#include <sys/metaslab.h>
40#include <sys/metaslab_impl.h>
41#include <sys/space_map.h>
42#include <sys/space_reftree.h>
43#include <sys/zio.h>
44#include <sys/zap.h>
45#include <sys/fs/zfs.h>
46#include <sys/arc.h>
47#include <sys/zil.h>
48#include <sys/dsl_scan.h>
49#include <sys/abd.h>
50
51/*
52 * Virtual device management.
53 */
54
55static vdev_ops_t *vdev_ops_table[] = {
56	&vdev_root_ops,
57	&vdev_raidz_ops,
58	&vdev_mirror_ops,
59	&vdev_replacing_ops,
60	&vdev_spare_ops,
61	&vdev_disk_ops,
62	&vdev_file_ops,
63	&vdev_missing_ops,
64	&vdev_hole_ops,
65	NULL
66};
67
68/* maximum scrub/resilver I/O queue per leaf vdev */
69int zfs_scrub_limit = 10;
70
71/*
72 * When a vdev is added, it will be divided into approximately (but no
73 * more than) this number of metaslabs.
74 */
75int metaslabs_per_vdev = 200;
76
77/*
78 * Given a vdev type, return the appropriate ops vector.
79 */
80static vdev_ops_t *
81vdev_getops(const char *type)
82{
83	vdev_ops_t *ops, **opspp;
84
85	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
86		if (strcmp(ops->vdev_op_type, type) == 0)
87			break;
88
89	return (ops);
90}
91
92/*
93 * Default asize function: return the MAX of psize with the asize of
94 * all children.  This is what's used by anything other than RAID-Z.
95 */
96uint64_t
97vdev_default_asize(vdev_t *vd, uint64_t psize)
98{
99	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
100	uint64_t csize;
101
102	for (int c = 0; c < vd->vdev_children; c++) {
103		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
104		asize = MAX(asize, csize);
105	}
106
107	return (asize);
108}
109
110/*
111 * Get the minimum allocatable size. We define the allocatable size as
112 * the vdev's asize rounded to the nearest metaslab. This allows us to
113 * replace or attach devices which don't have the same physical size but
114 * can still satisfy the same number of allocations.
115 */
116uint64_t
117vdev_get_min_asize(vdev_t *vd)
118{
119	vdev_t *pvd = vd->vdev_parent;
120
121	/*
122	 * If our parent is NULL (inactive spare or cache) or is the root,
123	 * just return our own asize.
124	 */
125	if (pvd == NULL)
126		return (vd->vdev_asize);
127
128	/*
129	 * The top-level vdev just returns the allocatable size rounded
130	 * to the nearest metaslab.
131	 */
132	if (vd == vd->vdev_top)
133		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
134
135	/*
136	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
137	 * so each child must provide at least 1/Nth of its asize.
138	 */
139	if (pvd->vdev_ops == &vdev_raidz_ops)
140		return ((pvd->vdev_min_asize + pvd->vdev_children - 1) /
141		    pvd->vdev_children);
142
143	return (pvd->vdev_min_asize);
144}
145
146void
147vdev_set_min_asize(vdev_t *vd)
148{
149	vd->vdev_min_asize = vdev_get_min_asize(vd);
150
151	for (int c = 0; c < vd->vdev_children; c++)
152		vdev_set_min_asize(vd->vdev_child[c]);
153}
154
155vdev_t *
156vdev_lookup_top(spa_t *spa, uint64_t vdev)
157{
158	vdev_t *rvd = spa->spa_root_vdev;
159
160	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
161
162	if (vdev < rvd->vdev_children) {
163		ASSERT(rvd->vdev_child[vdev] != NULL);
164		return (rvd->vdev_child[vdev]);
165	}
166
167	return (NULL);
168}
169
170vdev_t *
171vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
172{
173	vdev_t *mvd;
174
175	if (vd->vdev_guid == guid)
176		return (vd);
177
178	for (int c = 0; c < vd->vdev_children; c++)
179		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
180		    NULL)
181			return (mvd);
182
183	return (NULL);
184}
185
186static int
187vdev_count_leaves_impl(vdev_t *vd)
188{
189	int n = 0;
190
191	if (vd->vdev_ops->vdev_op_leaf)
192		return (1);
193
194	for (int c = 0; c < vd->vdev_children; c++)
195		n += vdev_count_leaves_impl(vd->vdev_child[c]);
196
197	return (n);
198}
199
200int
201vdev_count_leaves(spa_t *spa)
202{
203	return (vdev_count_leaves_impl(spa->spa_root_vdev));
204}
205
206void
207vdev_add_child(vdev_t *pvd, vdev_t *cvd)
208{
209	size_t oldsize, newsize;
210	uint64_t id = cvd->vdev_id;
211	vdev_t **newchild;
212	spa_t *spa = cvd->vdev_spa;
213
214	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
215	ASSERT(cvd->vdev_parent == NULL);
216
217	cvd->vdev_parent = pvd;
218
219	if (pvd == NULL)
220		return;
221
222	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
223
224	oldsize = pvd->vdev_children * sizeof (vdev_t *);
225	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
226	newsize = pvd->vdev_children * sizeof (vdev_t *);
227
228	newchild = kmem_zalloc(newsize, KM_SLEEP);
229	if (pvd->vdev_child != NULL) {
230		bcopy(pvd->vdev_child, newchild, oldsize);
231		kmem_free(pvd->vdev_child, oldsize);
232	}
233
234	pvd->vdev_child = newchild;
235	pvd->vdev_child[id] = cvd;
236
237	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
238	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
239
240	/*
241	 * Walk up all ancestors to update guid sum.
242	 */
243	for (; pvd != NULL; pvd = pvd->vdev_parent)
244		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
245}
246
247void
248vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
249{
250	int c;
251	uint_t id = cvd->vdev_id;
252
253	ASSERT(cvd->vdev_parent == pvd);
254
255	if (pvd == NULL)
256		return;
257
258	ASSERT(id < pvd->vdev_children);
259	ASSERT(pvd->vdev_child[id] == cvd);
260
261	pvd->vdev_child[id] = NULL;
262	cvd->vdev_parent = NULL;
263
264	for (c = 0; c < pvd->vdev_children; c++)
265		if (pvd->vdev_child[c])
266			break;
267
268	if (c == pvd->vdev_children) {
269		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
270		pvd->vdev_child = NULL;
271		pvd->vdev_children = 0;
272	}
273
274	/*
275	 * Walk up all ancestors to update guid sum.
276	 */
277	for (; pvd != NULL; pvd = pvd->vdev_parent)
278		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
279}
280
281/*
282 * Remove any holes in the child array.
283 */
284void
285vdev_compact_children(vdev_t *pvd)
286{
287	vdev_t **newchild, *cvd;
288	int oldc = pvd->vdev_children;
289	int newc;
290
291	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
292
293	for (int c = newc = 0; c < oldc; c++)
294		if (pvd->vdev_child[c])
295			newc++;
296
297	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
298
299	for (int c = newc = 0; c < oldc; c++) {
300		if ((cvd = pvd->vdev_child[c]) != NULL) {
301			newchild[newc] = cvd;
302			cvd->vdev_id = newc++;
303		}
304	}
305
306	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
307	pvd->vdev_child = newchild;
308	pvd->vdev_children = newc;
309}
310
311/*
312 * Allocate and minimally initialize a vdev_t.
313 */
314vdev_t *
315vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
316{
317	vdev_t *vd;
318
319	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
320
321	if (spa->spa_root_vdev == NULL) {
322		ASSERT(ops == &vdev_root_ops);
323		spa->spa_root_vdev = vd;
324		spa->spa_load_guid = spa_generate_guid(NULL);
325	}
326
327	if (guid == 0 && ops != &vdev_hole_ops) {
328		if (spa->spa_root_vdev == vd) {
329			/*
330			 * The root vdev's guid will also be the pool guid,
331			 * which must be unique among all pools.
332			 */
333			guid = spa_generate_guid(NULL);
334		} else {
335			/*
336			 * Any other vdev's guid must be unique within the pool.
337			 */
338			guid = spa_generate_guid(spa);
339		}
340		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
341	}
342
343	vd->vdev_spa = spa;
344	vd->vdev_id = id;
345	vd->vdev_guid = guid;
346	vd->vdev_guid_sum = guid;
347	vd->vdev_ops = ops;
348	vd->vdev_state = VDEV_STATE_CLOSED;
349	vd->vdev_ishole = (ops == &vdev_hole_ops);
350
351	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
352	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
353	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
354	mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
355	for (int t = 0; t < DTL_TYPES; t++) {
356		vd->vdev_dtl[t] = range_tree_create(NULL, NULL,
357		    &vd->vdev_dtl_lock);
358	}
359	txg_list_create(&vd->vdev_ms_list, spa,
360	    offsetof(struct metaslab, ms_txg_node));
361	txg_list_create(&vd->vdev_dtl_list, spa,
362	    offsetof(struct vdev, vdev_dtl_node));
363	vd->vdev_stat.vs_timestamp = gethrtime();
364	vdev_queue_init(vd);
365	vdev_cache_init(vd);
366
367	return (vd);
368}
369
370/*
371 * Allocate a new vdev.  The 'alloctype' is used to control whether we are
372 * creating a new vdev or loading an existing one - the behavior is slightly
373 * different for each case.
374 */
375int
376vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
377    int alloctype)
378{
379	vdev_ops_t *ops;
380	char *type;
381	uint64_t guid = 0, islog, nparity;
382	vdev_t *vd;
383
384	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
385
386	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
387		return (SET_ERROR(EINVAL));
388
389	if ((ops = vdev_getops(type)) == NULL)
390		return (SET_ERROR(EINVAL));
391
392	/*
393	 * If this is a load, get the vdev guid from the nvlist.
394	 * Otherwise, vdev_alloc_common() will generate one for us.
395	 */
396	if (alloctype == VDEV_ALLOC_LOAD) {
397		uint64_t label_id;
398
399		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
400		    label_id != id)
401			return (SET_ERROR(EINVAL));
402
403		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
404			return (SET_ERROR(EINVAL));
405	} else if (alloctype == VDEV_ALLOC_SPARE) {
406		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
407			return (SET_ERROR(EINVAL));
408	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
409		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
410			return (SET_ERROR(EINVAL));
411	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
412		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
413			return (SET_ERROR(EINVAL));
414	}
415
416	/*
417	 * The first allocated vdev must be of type 'root'.
418	 */
419	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
420		return (SET_ERROR(EINVAL));
421
422	/*
423	 * Determine whether we're a log vdev.
424	 */
425	islog = 0;
426	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
427	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
428		return (SET_ERROR(ENOTSUP));
429
430	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
431		return (SET_ERROR(ENOTSUP));
432
433	/*
434	 * Set the nparity property for RAID-Z vdevs.
435	 */
436	nparity = -1ULL;
437	if (ops == &vdev_raidz_ops) {
438		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
439		    &nparity) == 0) {
440			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
441				return (SET_ERROR(EINVAL));
442			/*
443			 * Previous versions could only support 1 or 2 parity
444			 * device.
445			 */
446			if (nparity > 1 &&
447			    spa_version(spa) < SPA_VERSION_RAIDZ2)
448				return (SET_ERROR(ENOTSUP));
449			if (nparity > 2 &&
450			    spa_version(spa) < SPA_VERSION_RAIDZ3)
451				return (SET_ERROR(ENOTSUP));
452		} else {
453			/*
454			 * We require the parity to be specified for SPAs that
455			 * support multiple parity levels.
456			 */
457			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
458				return (SET_ERROR(EINVAL));
459			/*
460			 * Otherwise, we default to 1 parity device for RAID-Z.
461			 */
462			nparity = 1;
463		}
464	} else {
465		nparity = 0;
466	}
467	ASSERT(nparity != -1ULL);
468
469	vd = vdev_alloc_common(spa, id, guid, ops);
470
471	vd->vdev_islog = islog;
472	vd->vdev_nparity = nparity;
473
474	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
475		vd->vdev_path = spa_strdup(vd->vdev_path);
476	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
477		vd->vdev_devid = spa_strdup(vd->vdev_devid);
478	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
479	    &vd->vdev_physpath) == 0)
480		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
481	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
482		vd->vdev_fru = spa_strdup(vd->vdev_fru);
483
484	/*
485	 * Set the whole_disk property.  If it's not specified, leave the value
486	 * as -1.
487	 */
488	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
489	    &vd->vdev_wholedisk) != 0)
490		vd->vdev_wholedisk = -1ULL;
491
492	/*
493	 * Look for the 'not present' flag.  This will only be set if the device
494	 * was not present at the time of import.
495	 */
496	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
497	    &vd->vdev_not_present);
498
499	/*
500	 * Get the alignment requirement.
501	 */
502	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
503
504	/*
505	 * Retrieve the vdev creation time.
506	 */
507	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
508	    &vd->vdev_crtxg);
509
510	/*
511	 * If we're a top-level vdev, try to load the allocation parameters.
512	 */
513	if (parent && !parent->vdev_parent &&
514	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
515		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
516		    &vd->vdev_ms_array);
517		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
518		    &vd->vdev_ms_shift);
519		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
520		    &vd->vdev_asize);
521		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
522		    &vd->vdev_removing);
523		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
524		    &vd->vdev_top_zap);
525	} else {
526		ASSERT0(vd->vdev_top_zap);
527	}
528
529	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
530		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
531		    alloctype == VDEV_ALLOC_ADD ||
532		    alloctype == VDEV_ALLOC_SPLIT ||
533		    alloctype == VDEV_ALLOC_ROOTPOOL);
534		vd->vdev_mg = metaslab_group_create(islog ?
535		    spa_log_class(spa) : spa_normal_class(spa), vd);
536	}
537
538	if (vd->vdev_ops->vdev_op_leaf &&
539	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
540		(void) nvlist_lookup_uint64(nv,
541		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
542	} else {
543		ASSERT0(vd->vdev_leaf_zap);
544	}
545
546	/*
547	 * If we're a leaf vdev, try to load the DTL object and other state.
548	 */
549
550	if (vd->vdev_ops->vdev_op_leaf &&
551	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
552	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
553		if (alloctype == VDEV_ALLOC_LOAD) {
554			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
555			    &vd->vdev_dtl_object);
556			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
557			    &vd->vdev_unspare);
558		}
559
560		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
561			uint64_t spare = 0;
562
563			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
564			    &spare) == 0 && spare)
565				spa_spare_add(vd);
566		}
567
568		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
569		    &vd->vdev_offline);
570
571		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
572		    &vd->vdev_resilver_txg);
573
574		/*
575		 * When importing a pool, we want to ignore the persistent fault
576		 * state, as the diagnosis made on another system may not be
577		 * valid in the current context.  Local vdevs will
578		 * remain in the faulted state.
579		 */
580		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
581			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
582			    &vd->vdev_faulted);
583			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
584			    &vd->vdev_degraded);
585			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
586			    &vd->vdev_removed);
587
588			if (vd->vdev_faulted || vd->vdev_degraded) {
589				char *aux;
590
591				vd->vdev_label_aux =
592				    VDEV_AUX_ERR_EXCEEDED;
593				if (nvlist_lookup_string(nv,
594				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
595				    strcmp(aux, "external") == 0)
596					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
597			}
598		}
599	}
600
601	/*
602	 * Add ourselves to the parent's list of children.
603	 */
604	vdev_add_child(parent, vd);
605
606	*vdp = vd;
607
608	return (0);
609}
610
611void
612vdev_free(vdev_t *vd)
613{
614	spa_t *spa = vd->vdev_spa;
615
616	/*
617	 * vdev_free() implies closing the vdev first.  This is simpler than
618	 * trying to ensure complicated semantics for all callers.
619	 */
620	vdev_close(vd);
621
622	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
623	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
624
625	/*
626	 * Free all children.
627	 */
628	for (int c = 0; c < vd->vdev_children; c++)
629		vdev_free(vd->vdev_child[c]);
630
631	ASSERT(vd->vdev_child == NULL);
632	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
633
634	/*
635	 * Discard allocation state.
636	 */
637	if (vd->vdev_mg != NULL) {
638		vdev_metaslab_fini(vd);
639		metaslab_group_destroy(vd->vdev_mg);
640	}
641
642	ASSERT0(vd->vdev_stat.vs_space);
643	ASSERT0(vd->vdev_stat.vs_dspace);
644	ASSERT0(vd->vdev_stat.vs_alloc);
645
646	/*
647	 * Remove this vdev from its parent's child list.
648	 */
649	vdev_remove_child(vd->vdev_parent, vd);
650
651	ASSERT(vd->vdev_parent == NULL);
652
653	/*
654	 * Clean up vdev structure.
655	 */
656	vdev_queue_fini(vd);
657	vdev_cache_fini(vd);
658
659	if (vd->vdev_path)
660		spa_strfree(vd->vdev_path);
661	if (vd->vdev_devid)
662		spa_strfree(vd->vdev_devid);
663	if (vd->vdev_physpath)
664		spa_strfree(vd->vdev_physpath);
665	if (vd->vdev_fru)
666		spa_strfree(vd->vdev_fru);
667
668	if (vd->vdev_isspare)
669		spa_spare_remove(vd);
670	if (vd->vdev_isl2cache)
671		spa_l2cache_remove(vd);
672
673	txg_list_destroy(&vd->vdev_ms_list);
674	txg_list_destroy(&vd->vdev_dtl_list);
675
676	mutex_enter(&vd->vdev_dtl_lock);
677	space_map_close(vd->vdev_dtl_sm);
678	for (int t = 0; t < DTL_TYPES; t++) {
679		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
680		range_tree_destroy(vd->vdev_dtl[t]);
681	}
682	mutex_exit(&vd->vdev_dtl_lock);
683
684	mutex_destroy(&vd->vdev_queue_lock);
685	mutex_destroy(&vd->vdev_dtl_lock);
686	mutex_destroy(&vd->vdev_stat_lock);
687	mutex_destroy(&vd->vdev_probe_lock);
688
689	if (vd == spa->spa_root_vdev)
690		spa->spa_root_vdev = NULL;
691
692	kmem_free(vd, sizeof (vdev_t));
693}
694
695/*
696 * Transfer top-level vdev state from svd to tvd.
697 */
698static void
699vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
700{
701	spa_t *spa = svd->vdev_spa;
702	metaslab_t *msp;
703	vdev_t *vd;
704	int t;
705
706	ASSERT(tvd == tvd->vdev_top);
707
708	tvd->vdev_ms_array = svd->vdev_ms_array;
709	tvd->vdev_ms_shift = svd->vdev_ms_shift;
710	tvd->vdev_ms_count = svd->vdev_ms_count;
711	tvd->vdev_top_zap = svd->vdev_top_zap;
712
713	svd->vdev_ms_array = 0;
714	svd->vdev_ms_shift = 0;
715	svd->vdev_ms_count = 0;
716	svd->vdev_top_zap = 0;
717
718	if (tvd->vdev_mg)
719		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
720	tvd->vdev_mg = svd->vdev_mg;
721	tvd->vdev_ms = svd->vdev_ms;
722
723	svd->vdev_mg = NULL;
724	svd->vdev_ms = NULL;
725
726	if (tvd->vdev_mg != NULL)
727		tvd->vdev_mg->mg_vd = tvd;
728
729	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
730	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
731	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
732
733	svd->vdev_stat.vs_alloc = 0;
734	svd->vdev_stat.vs_space = 0;
735	svd->vdev_stat.vs_dspace = 0;
736
737	for (t = 0; t < TXG_SIZE; t++) {
738		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
739			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
740		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
741			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
742		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
743			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
744	}
745
746	if (list_link_active(&svd->vdev_config_dirty_node)) {
747		vdev_config_clean(svd);
748		vdev_config_dirty(tvd);
749	}
750
751	if (list_link_active(&svd->vdev_state_dirty_node)) {
752		vdev_state_clean(svd);
753		vdev_state_dirty(tvd);
754	}
755
756	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
757	svd->vdev_deflate_ratio = 0;
758
759	tvd->vdev_islog = svd->vdev_islog;
760	svd->vdev_islog = 0;
761}
762
763static void
764vdev_top_update(vdev_t *tvd, vdev_t *vd)
765{
766	if (vd == NULL)
767		return;
768
769	vd->vdev_top = tvd;
770
771	for (int c = 0; c < vd->vdev_children; c++)
772		vdev_top_update(tvd, vd->vdev_child[c]);
773}
774
775/*
776 * Add a mirror/replacing vdev above an existing vdev.
777 */
778vdev_t *
779vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
780{
781	spa_t *spa = cvd->vdev_spa;
782	vdev_t *pvd = cvd->vdev_parent;
783	vdev_t *mvd;
784
785	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
786
787	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
788
789	mvd->vdev_asize = cvd->vdev_asize;
790	mvd->vdev_min_asize = cvd->vdev_min_asize;
791	mvd->vdev_max_asize = cvd->vdev_max_asize;
792	mvd->vdev_ashift = cvd->vdev_ashift;
793	mvd->vdev_state = cvd->vdev_state;
794	mvd->vdev_crtxg = cvd->vdev_crtxg;
795
796	vdev_remove_child(pvd, cvd);
797	vdev_add_child(pvd, mvd);
798	cvd->vdev_id = mvd->vdev_children;
799	vdev_add_child(mvd, cvd);
800	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
801
802	if (mvd == mvd->vdev_top)
803		vdev_top_transfer(cvd, mvd);
804
805	return (mvd);
806}
807
808/*
809 * Remove a 1-way mirror/replacing vdev from the tree.
810 */
811void
812vdev_remove_parent(vdev_t *cvd)
813{
814	vdev_t *mvd = cvd->vdev_parent;
815	vdev_t *pvd = mvd->vdev_parent;
816
817	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
818
819	ASSERT(mvd->vdev_children == 1);
820	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
821	    mvd->vdev_ops == &vdev_replacing_ops ||
822	    mvd->vdev_ops == &vdev_spare_ops);
823	cvd->vdev_ashift = mvd->vdev_ashift;
824
825	vdev_remove_child(mvd, cvd);
826	vdev_remove_child(pvd, mvd);
827
828	/*
829	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
830	 * Otherwise, we could have detached an offline device, and when we
831	 * go to import the pool we'll think we have two top-level vdevs,
832	 * instead of a different version of the same top-level vdev.
833	 */
834	if (mvd->vdev_top == mvd) {
835		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
836		cvd->vdev_orig_guid = cvd->vdev_guid;
837		cvd->vdev_guid += guid_delta;
838		cvd->vdev_guid_sum += guid_delta;
839	}
840	cvd->vdev_id = mvd->vdev_id;
841	vdev_add_child(pvd, cvd);
842	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
843
844	if (cvd == cvd->vdev_top)
845		vdev_top_transfer(mvd, cvd);
846
847	ASSERT(mvd->vdev_children == 0);
848	vdev_free(mvd);
849}
850
851int
852vdev_metaslab_init(vdev_t *vd, uint64_t txg)
853{
854	spa_t *spa = vd->vdev_spa;
855	objset_t *mos = spa->spa_meta_objset;
856	uint64_t m;
857	uint64_t oldc = vd->vdev_ms_count;
858	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
859	metaslab_t **mspp;
860	int error;
861
862	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
863
864	/*
865	 * This vdev is not being allocated from yet or is a hole.
866	 */
867	if (vd->vdev_ms_shift == 0)
868		return (0);
869
870	ASSERT(!vd->vdev_ishole);
871
872	/*
873	 * Compute the raidz-deflation ratio.  Note, we hard-code
874	 * in 128k (1 << 17) because it is the "typical" blocksize.
875	 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
876	 * otherwise it would inconsistently account for existing bp's.
877	 */
878	vd->vdev_deflate_ratio = (1 << 17) /
879	    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
880
881	ASSERT(oldc <= newc);
882
883	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
884
885	if (oldc != 0) {
886		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
887		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
888	}
889
890	vd->vdev_ms = mspp;
891	vd->vdev_ms_count = newc;
892
893	for (m = oldc; m < newc; m++) {
894		uint64_t object = 0;
895
896		if (txg == 0) {
897			error = dmu_read(mos, vd->vdev_ms_array,
898			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
899			    DMU_READ_PREFETCH);
900			if (error)
901				return (error);
902		}
903
904		error = metaslab_init(vd->vdev_mg, m, object, txg,
905		    &(vd->vdev_ms[m]));
906		if (error)
907			return (error);
908	}
909
910	if (txg == 0)
911		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
912
913	/*
914	 * If the vdev is being removed we don't activate
915	 * the metaslabs since we want to ensure that no new
916	 * allocations are performed on this device.
917	 */
918	if (oldc == 0 && !vd->vdev_removing)
919		metaslab_group_activate(vd->vdev_mg);
920
921	if (txg == 0)
922		spa_config_exit(spa, SCL_ALLOC, FTAG);
923
924	return (0);
925}
926
927void
928vdev_metaslab_fini(vdev_t *vd)
929{
930	uint64_t m;
931	uint64_t count = vd->vdev_ms_count;
932
933	if (vd->vdev_ms != NULL) {
934		metaslab_group_passivate(vd->vdev_mg);
935		for (m = 0; m < count; m++) {
936			metaslab_t *msp = vd->vdev_ms[m];
937
938			if (msp != NULL)
939				metaslab_fini(msp);
940		}
941		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
942		vd->vdev_ms = NULL;
943	}
944}
945
946typedef struct vdev_probe_stats {
947	boolean_t	vps_readable;
948	boolean_t	vps_writeable;
949	int		vps_flags;
950} vdev_probe_stats_t;
951
952static void
953vdev_probe_done(zio_t *zio)
954{
955	spa_t *spa = zio->io_spa;
956	vdev_t *vd = zio->io_vd;
957	vdev_probe_stats_t *vps = zio->io_private;
958
959	ASSERT(vd->vdev_probe_zio != NULL);
960
961	if (zio->io_type == ZIO_TYPE_READ) {
962		if (zio->io_error == 0)
963			vps->vps_readable = 1;
964		if (zio->io_error == 0 && spa_writeable(spa)) {
965			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
966			    zio->io_offset, zio->io_size, zio->io_abd,
967			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
968			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
969		} else {
970			abd_free(zio->io_abd);
971		}
972	} else if (zio->io_type == ZIO_TYPE_WRITE) {
973		if (zio->io_error == 0)
974			vps->vps_writeable = 1;
975		abd_free(zio->io_abd);
976	} else if (zio->io_type == ZIO_TYPE_NULL) {
977		zio_t *pio;
978
979		vd->vdev_cant_read |= !vps->vps_readable;
980		vd->vdev_cant_write |= !vps->vps_writeable;
981
982		if (vdev_readable(vd) &&
983		    (vdev_writeable(vd) || !spa_writeable(spa))) {
984			zio->io_error = 0;
985		} else {
986			ASSERT(zio->io_error != 0);
987			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
988			    spa, vd, NULL, 0, 0);
989			zio->io_error = SET_ERROR(ENXIO);
990		}
991
992		mutex_enter(&vd->vdev_probe_lock);
993		ASSERT(vd->vdev_probe_zio == zio);
994		vd->vdev_probe_zio = NULL;
995		mutex_exit(&vd->vdev_probe_lock);
996
997		zio_link_t *zl = NULL;
998		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
999			if (!vdev_accessible(vd, pio))
1000				pio->io_error = SET_ERROR(ENXIO);
1001
1002		kmem_free(vps, sizeof (*vps));
1003	}
1004}
1005
1006/*
1007 * Determine whether this device is accessible.
1008 *
1009 * Read and write to several known locations: the pad regions of each
1010 * vdev label but the first, which we leave alone in case it contains
1011 * a VTOC.
1012 */
1013zio_t *
1014vdev_probe(vdev_t *vd, zio_t *zio)
1015{
1016	spa_t *spa = vd->vdev_spa;
1017	vdev_probe_stats_t *vps = NULL;
1018	zio_t *pio;
1019
1020	ASSERT(vd->vdev_ops->vdev_op_leaf);
1021
1022	/*
1023	 * Don't probe the probe.
1024	 */
1025	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1026		return (NULL);
1027
1028	/*
1029	 * To prevent 'probe storms' when a device fails, we create
1030	 * just one probe i/o at a time.  All zios that want to probe
1031	 * this vdev will become parents of the probe io.
1032	 */
1033	mutex_enter(&vd->vdev_probe_lock);
1034
1035	if ((pio = vd->vdev_probe_zio) == NULL) {
1036		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1037
1038		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1039		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1040		    ZIO_FLAG_TRYHARD;
1041
1042		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1043			/*
1044			 * vdev_cant_read and vdev_cant_write can only
1045			 * transition from TRUE to FALSE when we have the
1046			 * SCL_ZIO lock as writer; otherwise they can only
1047			 * transition from FALSE to TRUE.  This ensures that
1048			 * any zio looking at these values can assume that
1049			 * failures persist for the life of the I/O.  That's
1050			 * important because when a device has intermittent
1051			 * connectivity problems, we want to ensure that
1052			 * they're ascribed to the device (ENXIO) and not
1053			 * the zio (EIO).
1054			 *
1055			 * Since we hold SCL_ZIO as writer here, clear both
1056			 * values so the probe can reevaluate from first
1057			 * principles.
1058			 */
1059			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1060			vd->vdev_cant_read = B_FALSE;
1061			vd->vdev_cant_write = B_FALSE;
1062		}
1063
1064		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1065		    vdev_probe_done, vps,
1066		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1067
1068		/*
1069		 * We can't change the vdev state in this context, so we
1070		 * kick off an async task to do it on our behalf.
1071		 */
1072		if (zio != NULL) {
1073			vd->vdev_probe_wanted = B_TRUE;
1074			spa_async_request(spa, SPA_ASYNC_PROBE);
1075		}
1076	}
1077
1078	if (zio != NULL)
1079		zio_add_child(zio, pio);
1080
1081	mutex_exit(&vd->vdev_probe_lock);
1082
1083	if (vps == NULL) {
1084		ASSERT(zio != NULL);
1085		return (NULL);
1086	}
1087
1088	for (int l = 1; l < VDEV_LABELS; l++) {
1089		zio_nowait(zio_read_phys(pio, vd,
1090		    vdev_label_offset(vd->vdev_psize, l,
1091		    offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE,
1092		    abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1093		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1094		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1095	}
1096
1097	if (zio == NULL)
1098		return (pio);
1099
1100	zio_nowait(pio);
1101	return (NULL);
1102}
1103
1104static void
1105vdev_open_child(void *arg)
1106{
1107	vdev_t *vd = arg;
1108
1109	vd->vdev_open_thread = curthread;
1110	vd->vdev_open_error = vdev_open(vd);
1111	vd->vdev_open_thread = NULL;
1112}
1113
1114boolean_t
1115vdev_uses_zvols(vdev_t *vd)
1116{
1117	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1118	    strlen(ZVOL_DIR)) == 0)
1119		return (B_TRUE);
1120	for (int c = 0; c < vd->vdev_children; c++)
1121		if (vdev_uses_zvols(vd->vdev_child[c]))
1122			return (B_TRUE);
1123	return (B_FALSE);
1124}
1125
1126void
1127vdev_open_children(vdev_t *vd)
1128{
1129	taskq_t *tq;
1130	int children = vd->vdev_children;
1131
1132	/*
1133	 * in order to handle pools on top of zvols, do the opens
1134	 * in a single thread so that the same thread holds the
1135	 * spa_namespace_lock
1136	 */
1137	if (vdev_uses_zvols(vd)) {
1138		for (int c = 0; c < children; c++)
1139			vd->vdev_child[c]->vdev_open_error =
1140			    vdev_open(vd->vdev_child[c]);
1141		return;
1142	}
1143	tq = taskq_create("vdev_open", children, minclsyspri,
1144	    children, children, TASKQ_PREPOPULATE);
1145
1146	for (int c = 0; c < children; c++)
1147		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1148		    TQ_SLEEP) != NULL);
1149
1150	taskq_destroy(tq);
1151}
1152
1153/*
1154 * Prepare a virtual device for access.
1155 */
1156int
1157vdev_open(vdev_t *vd)
1158{
1159	spa_t *spa = vd->vdev_spa;
1160	int error;
1161	uint64_t osize = 0;
1162	uint64_t max_osize = 0;
1163	uint64_t asize, max_asize, psize;
1164	uint64_t ashift = 0;
1165
1166	ASSERT(vd->vdev_open_thread == curthread ||
1167	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1168	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1169	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1170	    vd->vdev_state == VDEV_STATE_OFFLINE);
1171
1172	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1173	vd->vdev_cant_read = B_FALSE;
1174	vd->vdev_cant_write = B_FALSE;
1175	vd->vdev_min_asize = vdev_get_min_asize(vd);
1176
1177	/*
1178	 * If this vdev is not removed, check its fault status.  If it's
1179	 * faulted, bail out of the open.
1180	 */
1181	if (!vd->vdev_removed && vd->vdev_faulted) {
1182		ASSERT(vd->vdev_children == 0);
1183		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1184		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1185		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1186		    vd->vdev_label_aux);
1187		return (SET_ERROR(ENXIO));
1188	} else if (vd->vdev_offline) {
1189		ASSERT(vd->vdev_children == 0);
1190		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1191		return (SET_ERROR(ENXIO));
1192	}
1193
1194	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1195
1196	/*
1197	 * Reset the vdev_reopening flag so that we actually close
1198	 * the vdev on error.
1199	 */
1200	vd->vdev_reopening = B_FALSE;
1201	if (zio_injection_enabled && error == 0)
1202		error = zio_handle_device_injection(vd, NULL, ENXIO);
1203
1204	if (error) {
1205		if (vd->vdev_removed &&
1206		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1207			vd->vdev_removed = B_FALSE;
1208
1209		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1210		    vd->vdev_stat.vs_aux);
1211		return (error);
1212	}
1213
1214	vd->vdev_removed = B_FALSE;
1215
1216	/*
1217	 * Recheck the faulted flag now that we have confirmed that
1218	 * the vdev is accessible.  If we're faulted, bail.
1219	 */
1220	if (vd->vdev_faulted) {
1221		ASSERT(vd->vdev_children == 0);
1222		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1223		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1224		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1225		    vd->vdev_label_aux);
1226		return (SET_ERROR(ENXIO));
1227	}
1228
1229	if (vd->vdev_degraded) {
1230		ASSERT(vd->vdev_children == 0);
1231		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1232		    VDEV_AUX_ERR_EXCEEDED);
1233	} else {
1234		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1235	}
1236
1237	/*
1238	 * For hole or missing vdevs we just return success.
1239	 */
1240	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1241		return (0);
1242
1243	for (int c = 0; c < vd->vdev_children; c++) {
1244		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1245			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1246			    VDEV_AUX_NONE);
1247			break;
1248		}
1249	}
1250
1251	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1252	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1253
1254	if (vd->vdev_children == 0) {
1255		if (osize < SPA_MINDEVSIZE) {
1256			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1257			    VDEV_AUX_TOO_SMALL);
1258			return (SET_ERROR(EOVERFLOW));
1259		}
1260		psize = osize;
1261		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1262		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1263		    VDEV_LABEL_END_SIZE);
1264	} else {
1265		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1266		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1267			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1268			    VDEV_AUX_TOO_SMALL);
1269			return (SET_ERROR(EOVERFLOW));
1270		}
1271		psize = 0;
1272		asize = osize;
1273		max_asize = max_osize;
1274	}
1275
1276	vd->vdev_psize = psize;
1277
1278	/*
1279	 * Make sure the allocatable size hasn't shrunk too much.
1280	 */
1281	if (asize < vd->vdev_min_asize) {
1282		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1283		    VDEV_AUX_BAD_LABEL);
1284		return (SET_ERROR(EINVAL));
1285	}
1286
1287	if (vd->vdev_asize == 0) {
1288		/*
1289		 * This is the first-ever open, so use the computed values.
1290		 * For testing purposes, a higher ashift can be requested.
1291		 */
1292		vd->vdev_asize = asize;
1293		vd->vdev_max_asize = max_asize;
1294		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1295	} else {
1296		/*
1297		 * Detect if the alignment requirement has increased.
1298		 * We don't want to make the pool unavailable, just
1299		 * issue a warning instead.
1300		 */
1301		if (ashift > vd->vdev_top->vdev_ashift &&
1302		    vd->vdev_ops->vdev_op_leaf) {
1303			cmn_err(CE_WARN,
1304			    "Disk, '%s', has a block alignment that is "
1305			    "larger than the pool's alignment\n",
1306			    vd->vdev_path);
1307		}
1308		vd->vdev_max_asize = max_asize;
1309	}
1310
1311	/*
1312	 * If all children are healthy we update asize if either:
1313	 * The asize has increased, due to a device expansion caused by dynamic
1314	 * LUN growth or vdev replacement, and automatic expansion is enabled;
1315	 * making the additional space available.
1316	 *
1317	 * The asize has decreased, due to a device shrink usually caused by a
1318	 * vdev replace with a smaller device. This ensures that calculations
1319	 * based of max_asize and asize e.g. esize are always valid. It's safe
1320	 * to do this as we've already validated that asize is greater than
1321	 * vdev_min_asize.
1322	 */
1323	if (vd->vdev_state == VDEV_STATE_HEALTHY &&
1324	    ((asize > vd->vdev_asize &&
1325	    (vd->vdev_expanding || spa->spa_autoexpand)) ||
1326	    (asize < vd->vdev_asize)))
1327		vd->vdev_asize = asize;
1328
1329	vdev_set_min_asize(vd);
1330
1331	/*
1332	 * Ensure we can issue some IO before declaring the
1333	 * vdev open for business.
1334	 */
1335	if (vd->vdev_ops->vdev_op_leaf &&
1336	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1337		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1338		    VDEV_AUX_ERR_EXCEEDED);
1339		return (error);
1340	}
1341
1342	/*
1343	 * Track the min and max ashift values for normal data devices.
1344	 */
1345	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1346	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1347		if (vd->vdev_ashift > spa->spa_max_ashift)
1348			spa->spa_max_ashift = vd->vdev_ashift;
1349		if (vd->vdev_ashift < spa->spa_min_ashift)
1350			spa->spa_min_ashift = vd->vdev_ashift;
1351	}
1352
1353	/*
1354	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1355	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1356	 * since this would just restart the scrub we are already doing.
1357	 */
1358	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1359	    vdev_resilver_needed(vd, NULL, NULL))
1360		spa_async_request(spa, SPA_ASYNC_RESILVER);
1361
1362	return (0);
1363}
1364
1365/*
1366 * Called once the vdevs are all opened, this routine validates the label
1367 * contents.  This needs to be done before vdev_load() so that we don't
1368 * inadvertently do repair I/Os to the wrong device.
1369 *
1370 * If 'strict' is false ignore the spa guid check. This is necessary because
1371 * if the machine crashed during a re-guid the new guid might have been written
1372 * to all of the vdev labels, but not the cached config. The strict check
1373 * will be performed when the pool is opened again using the mos config.
1374 *
1375 * This function will only return failure if one of the vdevs indicates that it
1376 * has since been destroyed or exported.  This is only possible if
1377 * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1378 * will be updated but the function will return 0.
1379 */
1380int
1381vdev_validate(vdev_t *vd, boolean_t strict)
1382{
1383	spa_t *spa = vd->vdev_spa;
1384	nvlist_t *label;
1385	uint64_t guid = 0, top_guid;
1386	uint64_t state;
1387
1388	for (int c = 0; c < vd->vdev_children; c++)
1389		if (vdev_validate(vd->vdev_child[c], strict) != 0)
1390			return (SET_ERROR(EBADF));
1391
1392	/*
1393	 * If the device has already failed, or was marked offline, don't do
1394	 * any further validation.  Otherwise, label I/O will fail and we will
1395	 * overwrite the previous state.
1396	 */
1397	if (vd->vdev_ops->vdev_op_leaf && vdev_readable(vd)) {
1398		uint64_t aux_guid = 0;
1399		nvlist_t *nvl;
1400		uint64_t txg = spa_last_synced_txg(spa) != 0 ?
1401		    spa_last_synced_txg(spa) : -1ULL;
1402
1403		if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1404			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1405			    VDEV_AUX_BAD_LABEL);
1406			return (0);
1407		}
1408
1409		/*
1410		 * Determine if this vdev has been split off into another
1411		 * pool.  If so, then refuse to open it.
1412		 */
1413		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1414		    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1415			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1416			    VDEV_AUX_SPLIT_POOL);
1417			nvlist_free(label);
1418			return (0);
1419		}
1420
1421		if (strict && (nvlist_lookup_uint64(label,
1422		    ZPOOL_CONFIG_POOL_GUID, &guid) != 0 ||
1423		    guid != spa_guid(spa))) {
1424			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1425			    VDEV_AUX_CORRUPT_DATA);
1426			nvlist_free(label);
1427			return (0);
1428		}
1429
1430		if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1431		    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1432		    &aux_guid) != 0)
1433			aux_guid = 0;
1434
1435		/*
1436		 * If this vdev just became a top-level vdev because its
1437		 * sibling was detached, it will have adopted the parent's
1438		 * vdev guid -- but the label may or may not be on disk yet.
1439		 * Fortunately, either version of the label will have the
1440		 * same top guid, so if we're a top-level vdev, we can
1441		 * safely compare to that instead.
1442		 *
1443		 * If we split this vdev off instead, then we also check the
1444		 * original pool's guid.  We don't want to consider the vdev
1445		 * corrupt if it is partway through a split operation.
1446		 */
1447		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID,
1448		    &guid) != 0 ||
1449		    nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID,
1450		    &top_guid) != 0 ||
1451		    ((vd->vdev_guid != guid && vd->vdev_guid != aux_guid) &&
1452		    (vd->vdev_guid != top_guid || vd != vd->vdev_top))) {
1453			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1454			    VDEV_AUX_CORRUPT_DATA);
1455			nvlist_free(label);
1456			return (0);
1457		}
1458
1459		if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1460		    &state) != 0) {
1461			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1462			    VDEV_AUX_CORRUPT_DATA);
1463			nvlist_free(label);
1464			return (0);
1465		}
1466
1467		nvlist_free(label);
1468
1469		/*
1470		 * If this is a verbatim import, no need to check the
1471		 * state of the pool.
1472		 */
1473		if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1474		    spa_load_state(spa) == SPA_LOAD_OPEN &&
1475		    state != POOL_STATE_ACTIVE)
1476			return (SET_ERROR(EBADF));
1477
1478		/*
1479		 * If we were able to open and validate a vdev that was
1480		 * previously marked permanently unavailable, clear that state
1481		 * now.
1482		 */
1483		if (vd->vdev_not_present)
1484			vd->vdev_not_present = 0;
1485	}
1486
1487	return (0);
1488}
1489
1490/*
1491 * Close a virtual device.
1492 */
1493void
1494vdev_close(vdev_t *vd)
1495{
1496	spa_t *spa = vd->vdev_spa;
1497	vdev_t *pvd = vd->vdev_parent;
1498
1499	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1500
1501	/*
1502	 * If our parent is reopening, then we are as well, unless we are
1503	 * going offline.
1504	 */
1505	if (pvd != NULL && pvd->vdev_reopening)
1506		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1507
1508	vd->vdev_ops->vdev_op_close(vd);
1509
1510	vdev_cache_purge(vd);
1511
1512	/*
1513	 * We record the previous state before we close it, so that if we are
1514	 * doing a reopen(), we don't generate FMA ereports if we notice that
1515	 * it's still faulted.
1516	 */
1517	vd->vdev_prevstate = vd->vdev_state;
1518
1519	if (vd->vdev_offline)
1520		vd->vdev_state = VDEV_STATE_OFFLINE;
1521	else
1522		vd->vdev_state = VDEV_STATE_CLOSED;
1523	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1524}
1525
1526void
1527vdev_hold(vdev_t *vd)
1528{
1529	spa_t *spa = vd->vdev_spa;
1530
1531	ASSERT(spa_is_root(spa));
1532	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1533		return;
1534
1535	for (int c = 0; c < vd->vdev_children; c++)
1536		vdev_hold(vd->vdev_child[c]);
1537
1538	if (vd->vdev_ops->vdev_op_leaf)
1539		vd->vdev_ops->vdev_op_hold(vd);
1540}
1541
1542void
1543vdev_rele(vdev_t *vd)
1544{
1545	spa_t *spa = vd->vdev_spa;
1546
1547	ASSERT(spa_is_root(spa));
1548	for (int c = 0; c < vd->vdev_children; c++)
1549		vdev_rele(vd->vdev_child[c]);
1550
1551	if (vd->vdev_ops->vdev_op_leaf)
1552		vd->vdev_ops->vdev_op_rele(vd);
1553}
1554
1555/*
1556 * Reopen all interior vdevs and any unopened leaves.  We don't actually
1557 * reopen leaf vdevs which had previously been opened as they might deadlock
1558 * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1559 * If the leaf has never been opened then open it, as usual.
1560 */
1561void
1562vdev_reopen(vdev_t *vd)
1563{
1564	spa_t *spa = vd->vdev_spa;
1565
1566	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1567
1568	/* set the reopening flag unless we're taking the vdev offline */
1569	vd->vdev_reopening = !vd->vdev_offline;
1570	vdev_close(vd);
1571	(void) vdev_open(vd);
1572
1573	/*
1574	 * Call vdev_validate() here to make sure we have the same device.
1575	 * Otherwise, a device with an invalid label could be successfully
1576	 * opened in response to vdev_reopen().
1577	 */
1578	if (vd->vdev_aux) {
1579		(void) vdev_validate_aux(vd);
1580		if (vdev_readable(vd) && vdev_writeable(vd) &&
1581		    vd->vdev_aux == &spa->spa_l2cache &&
1582		    !l2arc_vdev_present(vd))
1583			l2arc_add_vdev(spa, vd);
1584	} else {
1585		(void) vdev_validate(vd, B_TRUE);
1586	}
1587
1588	/*
1589	 * Reassess parent vdev's health.
1590	 */
1591	vdev_propagate_state(vd);
1592}
1593
1594int
1595vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1596{
1597	int error;
1598
1599	/*
1600	 * Normally, partial opens (e.g. of a mirror) are allowed.
1601	 * For a create, however, we want to fail the request if
1602	 * there are any components we can't open.
1603	 */
1604	error = vdev_open(vd);
1605
1606	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1607		vdev_close(vd);
1608		return (error ? error : ENXIO);
1609	}
1610
1611	/*
1612	 * Recursively load DTLs and initialize all labels.
1613	 */
1614	if ((error = vdev_dtl_load(vd)) != 0 ||
1615	    (error = vdev_label_init(vd, txg, isreplacing ?
1616	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1617		vdev_close(vd);
1618		return (error);
1619	}
1620
1621	return (0);
1622}
1623
1624void
1625vdev_metaslab_set_size(vdev_t *vd)
1626{
1627	/*
1628	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1629	 */
1630	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1631	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1632}
1633
1634void
1635vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1636{
1637	ASSERT(vd == vd->vdev_top);
1638	ASSERT(!vd->vdev_ishole);
1639	ASSERT(ISP2(flags));
1640	ASSERT(spa_writeable(vd->vdev_spa));
1641
1642	if (flags & VDD_METASLAB)
1643		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1644
1645	if (flags & VDD_DTL)
1646		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1647
1648	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1649}
1650
1651void
1652vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1653{
1654	for (int c = 0; c < vd->vdev_children; c++)
1655		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1656
1657	if (vd->vdev_ops->vdev_op_leaf)
1658		vdev_dirty(vd->vdev_top, flags, vd, txg);
1659}
1660
1661/*
1662 * DTLs.
1663 *
1664 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1665 * the vdev has less than perfect replication.  There are four kinds of DTL:
1666 *
1667 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1668 *
1669 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1670 *
1671 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1672 *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1673 *	txgs that was scrubbed.
1674 *
1675 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1676 *	persistent errors or just some device being offline.
1677 *	Unlike the other three, the DTL_OUTAGE map is not generally
1678 *	maintained; it's only computed when needed, typically to
1679 *	determine whether a device can be detached.
1680 *
1681 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1682 * either has the data or it doesn't.
1683 *
1684 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
1685 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
1686 * if any child is less than fully replicated, then so is its parent.
1687 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
1688 * comprising only those txgs which appear in 'maxfaults' or more children;
1689 * those are the txgs we don't have enough replication to read.  For example,
1690 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
1691 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
1692 * two child DTL_MISSING maps.
1693 *
1694 * It should be clear from the above that to compute the DTLs and outage maps
1695 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
1696 * Therefore, that is all we keep on disk.  When loading the pool, or after
1697 * a configuration change, we generate all other DTLs from first principles.
1698 */
1699void
1700vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1701{
1702	range_tree_t *rt = vd->vdev_dtl[t];
1703
1704	ASSERT(t < DTL_TYPES);
1705	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1706	ASSERT(spa_writeable(vd->vdev_spa));
1707
1708	mutex_enter(rt->rt_lock);
1709	if (!range_tree_contains(rt, txg, size))
1710		range_tree_add(rt, txg, size);
1711	mutex_exit(rt->rt_lock);
1712}
1713
1714boolean_t
1715vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
1716{
1717	range_tree_t *rt = vd->vdev_dtl[t];
1718	boolean_t dirty = B_FALSE;
1719
1720	ASSERT(t < DTL_TYPES);
1721	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
1722
1723	mutex_enter(rt->rt_lock);
1724	if (range_tree_space(rt) != 0)
1725		dirty = range_tree_contains(rt, txg, size);
1726	mutex_exit(rt->rt_lock);
1727
1728	return (dirty);
1729}
1730
1731boolean_t
1732vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
1733{
1734	range_tree_t *rt = vd->vdev_dtl[t];
1735	boolean_t empty;
1736
1737	mutex_enter(rt->rt_lock);
1738	empty = (range_tree_space(rt) == 0);
1739	mutex_exit(rt->rt_lock);
1740
1741	return (empty);
1742}
1743
1744/*
1745 * Returns the lowest txg in the DTL range.
1746 */
1747static uint64_t
1748vdev_dtl_min(vdev_t *vd)
1749{
1750	range_seg_t *rs;
1751
1752	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1753	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1754	ASSERT0(vd->vdev_children);
1755
1756	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1757	return (rs->rs_start - 1);
1758}
1759
1760/*
1761 * Returns the highest txg in the DTL.
1762 */
1763static uint64_t
1764vdev_dtl_max(vdev_t *vd)
1765{
1766	range_seg_t *rs;
1767
1768	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
1769	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
1770	ASSERT0(vd->vdev_children);
1771
1772	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
1773	return (rs->rs_end);
1774}
1775
1776/*
1777 * Determine if a resilvering vdev should remove any DTL entries from
1778 * its range. If the vdev was resilvering for the entire duration of the
1779 * scan then it should excise that range from its DTLs. Otherwise, this
1780 * vdev is considered partially resilvered and should leave its DTL
1781 * entries intact. The comment in vdev_dtl_reassess() describes how we
1782 * excise the DTLs.
1783 */
1784static boolean_t
1785vdev_dtl_should_excise(vdev_t *vd)
1786{
1787	spa_t *spa = vd->vdev_spa;
1788	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1789
1790	ASSERT0(scn->scn_phys.scn_errors);
1791	ASSERT0(vd->vdev_children);
1792
1793	if (vd->vdev_state < VDEV_STATE_DEGRADED)
1794		return (B_FALSE);
1795
1796	if (vd->vdev_resilver_txg == 0 ||
1797	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
1798		return (B_TRUE);
1799
1800	/*
1801	 * When a resilver is initiated the scan will assign the scn_max_txg
1802	 * value to the highest txg value that exists in all DTLs. If this
1803	 * device's max DTL is not part of this scan (i.e. it is not in
1804	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
1805	 * for excision.
1806	 */
1807	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
1808		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
1809		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
1810		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
1811		return (B_TRUE);
1812	}
1813	return (B_FALSE);
1814}
1815
1816/*
1817 * Reassess DTLs after a config change or scrub completion.
1818 */
1819void
1820vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
1821{
1822	spa_t *spa = vd->vdev_spa;
1823	avl_tree_t reftree;
1824	int minref;
1825
1826	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1827
1828	for (int c = 0; c < vd->vdev_children; c++)
1829		vdev_dtl_reassess(vd->vdev_child[c], txg,
1830		    scrub_txg, scrub_done);
1831
1832	if (vd == spa->spa_root_vdev || vd->vdev_ishole || vd->vdev_aux)
1833		return;
1834
1835	if (vd->vdev_ops->vdev_op_leaf) {
1836		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
1837
1838		mutex_enter(&vd->vdev_dtl_lock);
1839
1840		/*
1841		 * If we've completed a scan cleanly then determine
1842		 * if this vdev should remove any DTLs. We only want to
1843		 * excise regions on vdevs that were available during
1844		 * the entire duration of this scan.
1845		 */
1846		if (scrub_txg != 0 &&
1847		    (spa->spa_scrub_started ||
1848		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
1849		    vdev_dtl_should_excise(vd)) {
1850			/*
1851			 * We completed a scrub up to scrub_txg.  If we
1852			 * did it without rebooting, then the scrub dtl
1853			 * will be valid, so excise the old region and
1854			 * fold in the scrub dtl.  Otherwise, leave the
1855			 * dtl as-is if there was an error.
1856			 *
1857			 * There's little trick here: to excise the beginning
1858			 * of the DTL_MISSING map, we put it into a reference
1859			 * tree and then add a segment with refcnt -1 that
1860			 * covers the range [0, scrub_txg).  This means
1861			 * that each txg in that range has refcnt -1 or 0.
1862			 * We then add DTL_SCRUB with a refcnt of 2, so that
1863			 * entries in the range [0, scrub_txg) will have a
1864			 * positive refcnt -- either 1 or 2.  We then convert
1865			 * the reference tree into the new DTL_MISSING map.
1866			 */
1867			space_reftree_create(&reftree);
1868			space_reftree_add_map(&reftree,
1869			    vd->vdev_dtl[DTL_MISSING], 1);
1870			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
1871			space_reftree_add_map(&reftree,
1872			    vd->vdev_dtl[DTL_SCRUB], 2);
1873			space_reftree_generate_map(&reftree,
1874			    vd->vdev_dtl[DTL_MISSING], 1);
1875			space_reftree_destroy(&reftree);
1876		}
1877		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
1878		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1879		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
1880		if (scrub_done)
1881			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
1882		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
1883		if (!vdev_readable(vd))
1884			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
1885		else
1886			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
1887			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
1888
1889		/*
1890		 * If the vdev was resilvering and no longer has any
1891		 * DTLs then reset its resilvering flag.
1892		 */
1893		if (vd->vdev_resilver_txg != 0 &&
1894		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
1895		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
1896			vd->vdev_resilver_txg = 0;
1897
1898		mutex_exit(&vd->vdev_dtl_lock);
1899
1900		if (txg != 0)
1901			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
1902		return;
1903	}
1904
1905	mutex_enter(&vd->vdev_dtl_lock);
1906	for (int t = 0; t < DTL_TYPES; t++) {
1907		/* account for child's outage in parent's missing map */
1908		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
1909		if (t == DTL_SCRUB)
1910			continue;			/* leaf vdevs only */
1911		if (t == DTL_PARTIAL)
1912			minref = 1;			/* i.e. non-zero */
1913		else if (vd->vdev_nparity != 0)
1914			minref = vd->vdev_nparity + 1;	/* RAID-Z */
1915		else
1916			minref = vd->vdev_children;	/* any kind of mirror */
1917		space_reftree_create(&reftree);
1918		for (int c = 0; c < vd->vdev_children; c++) {
1919			vdev_t *cvd = vd->vdev_child[c];
1920			mutex_enter(&cvd->vdev_dtl_lock);
1921			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
1922			mutex_exit(&cvd->vdev_dtl_lock);
1923		}
1924		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
1925		space_reftree_destroy(&reftree);
1926	}
1927	mutex_exit(&vd->vdev_dtl_lock);
1928}
1929
1930int
1931vdev_dtl_load(vdev_t *vd)
1932{
1933	spa_t *spa = vd->vdev_spa;
1934	objset_t *mos = spa->spa_meta_objset;
1935	int error = 0;
1936
1937	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
1938		ASSERT(!vd->vdev_ishole);
1939
1940		error = space_map_open(&vd->vdev_dtl_sm, mos,
1941		    vd->vdev_dtl_object, 0, -1ULL, 0, &vd->vdev_dtl_lock);
1942		if (error)
1943			return (error);
1944		ASSERT(vd->vdev_dtl_sm != NULL);
1945
1946		mutex_enter(&vd->vdev_dtl_lock);
1947
1948		/*
1949		 * Now that we've opened the space_map we need to update
1950		 * the in-core DTL.
1951		 */
1952		space_map_update(vd->vdev_dtl_sm);
1953
1954		error = space_map_load(vd->vdev_dtl_sm,
1955		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
1956		mutex_exit(&vd->vdev_dtl_lock);
1957
1958		return (error);
1959	}
1960
1961	for (int c = 0; c < vd->vdev_children; c++) {
1962		error = vdev_dtl_load(vd->vdev_child[c]);
1963		if (error != 0)
1964			break;
1965	}
1966
1967	return (error);
1968}
1969
1970void
1971vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
1972{
1973	spa_t *spa = vd->vdev_spa;
1974
1975	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
1976	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1977	    zapobj, tx));
1978}
1979
1980uint64_t
1981vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
1982{
1983	spa_t *spa = vd->vdev_spa;
1984	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
1985	    DMU_OT_NONE, 0, tx);
1986
1987	ASSERT(zap != 0);
1988	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
1989	    zap, tx));
1990
1991	return (zap);
1992}
1993
1994void
1995vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
1996{
1997	if (vd->vdev_ops != &vdev_hole_ops &&
1998	    vd->vdev_ops != &vdev_missing_ops &&
1999	    vd->vdev_ops != &vdev_root_ops &&
2000	    !vd->vdev_top->vdev_removing) {
2001		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2002			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2003		}
2004		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2005			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2006		}
2007	}
2008	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2009		vdev_construct_zaps(vd->vdev_child[i], tx);
2010	}
2011}
2012
2013void
2014vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2015{
2016	spa_t *spa = vd->vdev_spa;
2017	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2018	objset_t *mos = spa->spa_meta_objset;
2019	range_tree_t *rtsync;
2020	kmutex_t rtlock;
2021	dmu_tx_t *tx;
2022	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2023
2024	ASSERT(!vd->vdev_ishole);
2025	ASSERT(vd->vdev_ops->vdev_op_leaf);
2026
2027	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2028
2029	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2030		mutex_enter(&vd->vdev_dtl_lock);
2031		space_map_free(vd->vdev_dtl_sm, tx);
2032		space_map_close(vd->vdev_dtl_sm);
2033		vd->vdev_dtl_sm = NULL;
2034		mutex_exit(&vd->vdev_dtl_lock);
2035
2036		/*
2037		 * We only destroy the leaf ZAP for detached leaves or for
2038		 * removed log devices. Removed data devices handle leaf ZAP
2039		 * cleanup later, once cancellation is no longer possible.
2040		 */
2041		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2042		    vd->vdev_top->vdev_islog)) {
2043			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2044			vd->vdev_leaf_zap = 0;
2045		}
2046
2047		dmu_tx_commit(tx);
2048		return;
2049	}
2050
2051	if (vd->vdev_dtl_sm == NULL) {
2052		uint64_t new_object;
2053
2054		new_object = space_map_alloc(mos, tx);
2055		VERIFY3U(new_object, !=, 0);
2056
2057		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2058		    0, -1ULL, 0, &vd->vdev_dtl_lock));
2059		ASSERT(vd->vdev_dtl_sm != NULL);
2060	}
2061
2062	mutex_init(&rtlock, NULL, MUTEX_DEFAULT, NULL);
2063
2064	rtsync = range_tree_create(NULL, NULL, &rtlock);
2065
2066	mutex_enter(&rtlock);
2067
2068	mutex_enter(&vd->vdev_dtl_lock);
2069	range_tree_walk(rt, range_tree_add, rtsync);
2070	mutex_exit(&vd->vdev_dtl_lock);
2071
2072	space_map_truncate(vd->vdev_dtl_sm, tx);
2073	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2074	range_tree_vacate(rtsync, NULL, NULL);
2075
2076	range_tree_destroy(rtsync);
2077
2078	mutex_exit(&rtlock);
2079	mutex_destroy(&rtlock);
2080
2081	/*
2082	 * If the object for the space map has changed then dirty
2083	 * the top level so that we update the config.
2084	 */
2085	if (object != space_map_object(vd->vdev_dtl_sm)) {
2086		zfs_dbgmsg("txg %llu, spa %s, DTL old object %llu, "
2087		    "new object %llu", txg, spa_name(spa), object,
2088		    space_map_object(vd->vdev_dtl_sm));
2089		vdev_config_dirty(vd->vdev_top);
2090	}
2091
2092	dmu_tx_commit(tx);
2093
2094	mutex_enter(&vd->vdev_dtl_lock);
2095	space_map_update(vd->vdev_dtl_sm);
2096	mutex_exit(&vd->vdev_dtl_lock);
2097}
2098
2099/*
2100 * Determine whether the specified vdev can be offlined/detached/removed
2101 * without losing data.
2102 */
2103boolean_t
2104vdev_dtl_required(vdev_t *vd)
2105{
2106	spa_t *spa = vd->vdev_spa;
2107	vdev_t *tvd = vd->vdev_top;
2108	uint8_t cant_read = vd->vdev_cant_read;
2109	boolean_t required;
2110
2111	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2112
2113	if (vd == spa->spa_root_vdev || vd == tvd)
2114		return (B_TRUE);
2115
2116	/*
2117	 * Temporarily mark the device as unreadable, and then determine
2118	 * whether this results in any DTL outages in the top-level vdev.
2119	 * If not, we can safely offline/detach/remove the device.
2120	 */
2121	vd->vdev_cant_read = B_TRUE;
2122	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2123	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2124	vd->vdev_cant_read = cant_read;
2125	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2126
2127	if (!required && zio_injection_enabled)
2128		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2129
2130	return (required);
2131}
2132
2133/*
2134 * Determine if resilver is needed, and if so the txg range.
2135 */
2136boolean_t
2137vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2138{
2139	boolean_t needed = B_FALSE;
2140	uint64_t thismin = UINT64_MAX;
2141	uint64_t thismax = 0;
2142
2143	if (vd->vdev_children == 0) {
2144		mutex_enter(&vd->vdev_dtl_lock);
2145		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2146		    vdev_writeable(vd)) {
2147
2148			thismin = vdev_dtl_min(vd);
2149			thismax = vdev_dtl_max(vd);
2150			needed = B_TRUE;
2151		}
2152		mutex_exit(&vd->vdev_dtl_lock);
2153	} else {
2154		for (int c = 0; c < vd->vdev_children; c++) {
2155			vdev_t *cvd = vd->vdev_child[c];
2156			uint64_t cmin, cmax;
2157
2158			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2159				thismin = MIN(thismin, cmin);
2160				thismax = MAX(thismax, cmax);
2161				needed = B_TRUE;
2162			}
2163		}
2164	}
2165
2166	if (needed && minp) {
2167		*minp = thismin;
2168		*maxp = thismax;
2169	}
2170	return (needed);
2171}
2172
2173void
2174vdev_load(vdev_t *vd)
2175{
2176	/*
2177	 * Recursively load all children.
2178	 */
2179	for (int c = 0; c < vd->vdev_children; c++)
2180		vdev_load(vd->vdev_child[c]);
2181
2182	/*
2183	 * If this is a top-level vdev, initialize its metaslabs.
2184	 */
2185	if (vd == vd->vdev_top && !vd->vdev_ishole &&
2186	    (vd->vdev_ashift == 0 || vd->vdev_asize == 0 ||
2187	    vdev_metaslab_init(vd, 0) != 0))
2188		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2189		    VDEV_AUX_CORRUPT_DATA);
2190
2191	/*
2192	 * If this is a leaf vdev, load its DTL.
2193	 */
2194	if (vd->vdev_ops->vdev_op_leaf && vdev_dtl_load(vd) != 0)
2195		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2196		    VDEV_AUX_CORRUPT_DATA);
2197}
2198
2199/*
2200 * The special vdev case is used for hot spares and l2cache devices.  Its
2201 * sole purpose it to set the vdev state for the associated vdev.  To do this,
2202 * we make sure that we can open the underlying device, then try to read the
2203 * label, and make sure that the label is sane and that it hasn't been
2204 * repurposed to another pool.
2205 */
2206int
2207vdev_validate_aux(vdev_t *vd)
2208{
2209	nvlist_t *label;
2210	uint64_t guid, version;
2211	uint64_t state;
2212
2213	if (!vdev_readable(vd))
2214		return (0);
2215
2216	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2217		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2218		    VDEV_AUX_CORRUPT_DATA);
2219		return (-1);
2220	}
2221
2222	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2223	    !SPA_VERSION_IS_SUPPORTED(version) ||
2224	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2225	    guid != vd->vdev_guid ||
2226	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2227		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2228		    VDEV_AUX_CORRUPT_DATA);
2229		nvlist_free(label);
2230		return (-1);
2231	}
2232
2233	/*
2234	 * We don't actually check the pool state here.  If it's in fact in
2235	 * use by another pool, we update this fact on the fly when requested.
2236	 */
2237	nvlist_free(label);
2238	return (0);
2239}
2240
2241void
2242vdev_remove(vdev_t *vd, uint64_t txg)
2243{
2244	spa_t *spa = vd->vdev_spa;
2245	objset_t *mos = spa->spa_meta_objset;
2246	dmu_tx_t *tx;
2247
2248	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2249	ASSERT(vd == vd->vdev_top);
2250	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2251
2252	if (vd->vdev_ms != NULL) {
2253		metaslab_group_t *mg = vd->vdev_mg;
2254
2255		metaslab_group_histogram_verify(mg);
2256		metaslab_class_histogram_verify(mg->mg_class);
2257
2258		for (int m = 0; m < vd->vdev_ms_count; m++) {
2259			metaslab_t *msp = vd->vdev_ms[m];
2260
2261			if (msp == NULL || msp->ms_sm == NULL)
2262				continue;
2263
2264			mutex_enter(&msp->ms_lock);
2265			/*
2266			 * If the metaslab was not loaded when the vdev
2267			 * was removed then the histogram accounting may
2268			 * not be accurate. Update the histogram information
2269			 * here so that we ensure that the metaslab group
2270			 * and metaslab class are up-to-date.
2271			 */
2272			metaslab_group_histogram_remove(mg, msp);
2273
2274			VERIFY0(space_map_allocated(msp->ms_sm));
2275			space_map_free(msp->ms_sm, tx);
2276			space_map_close(msp->ms_sm);
2277			msp->ms_sm = NULL;
2278			mutex_exit(&msp->ms_lock);
2279		}
2280
2281		metaslab_group_histogram_verify(mg);
2282		metaslab_class_histogram_verify(mg->mg_class);
2283		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2284			ASSERT0(mg->mg_histogram[i]);
2285
2286	}
2287
2288	if (vd->vdev_ms_array) {
2289		(void) dmu_object_free(mos, vd->vdev_ms_array, tx);
2290		vd->vdev_ms_array = 0;
2291	}
2292
2293	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2294		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2295		vd->vdev_top_zap = 0;
2296	}
2297	dmu_tx_commit(tx);
2298}
2299
2300void
2301vdev_sync_done(vdev_t *vd, uint64_t txg)
2302{
2303	metaslab_t *msp;
2304	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2305
2306	ASSERT(!vd->vdev_ishole);
2307
2308	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2309		metaslab_sync_done(msp, txg);
2310
2311	if (reassess)
2312		metaslab_sync_reassess(vd->vdev_mg);
2313}
2314
2315void
2316vdev_sync(vdev_t *vd, uint64_t txg)
2317{
2318	spa_t *spa = vd->vdev_spa;
2319	vdev_t *lvd;
2320	metaslab_t *msp;
2321	dmu_tx_t *tx;
2322
2323	ASSERT(!vd->vdev_ishole);
2324
2325	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0) {
2326		ASSERT(vd == vd->vdev_top);
2327		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2328		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2329		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2330		ASSERT(vd->vdev_ms_array != 0);
2331		vdev_config_dirty(vd);
2332		dmu_tx_commit(tx);
2333	}
2334
2335	/*
2336	 * Remove the metadata associated with this vdev once it's empty.
2337	 */
2338	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing)
2339		vdev_remove(vd, txg);
2340
2341	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2342		metaslab_sync(msp, txg);
2343		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2344	}
2345
2346	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2347		vdev_dtl_sync(lvd, txg);
2348
2349	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2350}
2351
2352uint64_t
2353vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2354{
2355	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2356}
2357
2358/*
2359 * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2360 * not be opened, and no I/O is attempted.
2361 */
2362int
2363vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2364{
2365	vdev_t *vd, *tvd;
2366
2367	spa_vdev_state_enter(spa, SCL_NONE);
2368
2369	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2370		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2371
2372	if (!vd->vdev_ops->vdev_op_leaf)
2373		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2374
2375	tvd = vd->vdev_top;
2376
2377	/*
2378	 * We don't directly use the aux state here, but if we do a
2379	 * vdev_reopen(), we need this value to be present to remember why we
2380	 * were faulted.
2381	 */
2382	vd->vdev_label_aux = aux;
2383
2384	/*
2385	 * Faulted state takes precedence over degraded.
2386	 */
2387	vd->vdev_delayed_close = B_FALSE;
2388	vd->vdev_faulted = 1ULL;
2389	vd->vdev_degraded = 0ULL;
2390	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2391
2392	/*
2393	 * If this device has the only valid copy of the data, then
2394	 * back off and simply mark the vdev as degraded instead.
2395	 */
2396	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2397		vd->vdev_degraded = 1ULL;
2398		vd->vdev_faulted = 0ULL;
2399
2400		/*
2401		 * If we reopen the device and it's not dead, only then do we
2402		 * mark it degraded.
2403		 */
2404		vdev_reopen(tvd);
2405
2406		if (vdev_readable(vd))
2407			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2408	}
2409
2410	return (spa_vdev_state_exit(spa, vd, 0));
2411}
2412
2413/*
2414 * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2415 * user that something is wrong.  The vdev continues to operate as normal as far
2416 * as I/O is concerned.
2417 */
2418int
2419vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2420{
2421	vdev_t *vd;
2422
2423	spa_vdev_state_enter(spa, SCL_NONE);
2424
2425	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2426		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2427
2428	if (!vd->vdev_ops->vdev_op_leaf)
2429		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2430
2431	/*
2432	 * If the vdev is already faulted, then don't do anything.
2433	 */
2434	if (vd->vdev_faulted || vd->vdev_degraded)
2435		return (spa_vdev_state_exit(spa, NULL, 0));
2436
2437	vd->vdev_degraded = 1ULL;
2438	if (!vdev_is_dead(vd))
2439		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2440		    aux);
2441
2442	return (spa_vdev_state_exit(spa, vd, 0));
2443}
2444
2445/*
2446 * Online the given vdev.
2447 *
2448 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2449 * spare device should be detached when the device finishes resilvering.
2450 * Second, the online should be treated like a 'test' online case, so no FMA
2451 * events are generated if the device fails to open.
2452 */
2453int
2454vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2455{
2456	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2457	boolean_t wasoffline;
2458	vdev_state_t oldstate;
2459
2460	spa_vdev_state_enter(spa, SCL_NONE);
2461
2462	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2463		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2464
2465	if (!vd->vdev_ops->vdev_op_leaf)
2466		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2467
2468	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
2469	oldstate = vd->vdev_state;
2470
2471	tvd = vd->vdev_top;
2472	vd->vdev_offline = B_FALSE;
2473	vd->vdev_tmpoffline = B_FALSE;
2474	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2475	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2476
2477	/* XXX - L2ARC 1.0 does not support expansion */
2478	if (!vd->vdev_aux) {
2479		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2480			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2481	}
2482
2483	vdev_reopen(tvd);
2484	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2485
2486	if (!vd->vdev_aux) {
2487		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2488			pvd->vdev_expanding = B_FALSE;
2489	}
2490
2491	if (newstate)
2492		*newstate = vd->vdev_state;
2493	if ((flags & ZFS_ONLINE_UNSPARE) &&
2494	    !vdev_is_dead(vd) && vd->vdev_parent &&
2495	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2496	    vd->vdev_parent->vdev_child[0] == vd)
2497		vd->vdev_unspare = B_TRUE;
2498
2499	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2500
2501		/* XXX - L2ARC 1.0 does not support expansion */
2502		if (vd->vdev_aux)
2503			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2504		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2505	}
2506
2507	if (wasoffline ||
2508	    (oldstate < VDEV_STATE_DEGRADED &&
2509	    vd->vdev_state >= VDEV_STATE_DEGRADED))
2510		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
2511
2512	return (spa_vdev_state_exit(spa, vd, 0));
2513}
2514
2515static int
2516vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2517{
2518	vdev_t *vd, *tvd;
2519	int error = 0;
2520	uint64_t generation;
2521	metaslab_group_t *mg;
2522
2523top:
2524	spa_vdev_state_enter(spa, SCL_ALLOC);
2525
2526	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2527		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2528
2529	if (!vd->vdev_ops->vdev_op_leaf)
2530		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2531
2532	tvd = vd->vdev_top;
2533	mg = tvd->vdev_mg;
2534	generation = spa->spa_config_generation + 1;
2535
2536	/*
2537	 * If the device isn't already offline, try to offline it.
2538	 */
2539	if (!vd->vdev_offline) {
2540		/*
2541		 * If this device has the only valid copy of some data,
2542		 * don't allow it to be offlined. Log devices are always
2543		 * expendable.
2544		 */
2545		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2546		    vdev_dtl_required(vd))
2547			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2548
2549		/*
2550		 * If the top-level is a slog and it has had allocations
2551		 * then proceed.  We check that the vdev's metaslab group
2552		 * is not NULL since it's possible that we may have just
2553		 * added this vdev but not yet initialized its metaslabs.
2554		 */
2555		if (tvd->vdev_islog && mg != NULL) {
2556			/*
2557			 * Prevent any future allocations.
2558			 */
2559			metaslab_group_passivate(mg);
2560			(void) spa_vdev_state_exit(spa, vd, 0);
2561
2562			error = spa_offline_log(spa);
2563
2564			spa_vdev_state_enter(spa, SCL_ALLOC);
2565
2566			/*
2567			 * Check to see if the config has changed.
2568			 */
2569			if (error || generation != spa->spa_config_generation) {
2570				metaslab_group_activate(mg);
2571				if (error)
2572					return (spa_vdev_state_exit(spa,
2573					    vd, error));
2574				(void) spa_vdev_state_exit(spa, vd, 0);
2575				goto top;
2576			}
2577			ASSERT0(tvd->vdev_stat.vs_alloc);
2578		}
2579
2580		/*
2581		 * Offline this device and reopen its top-level vdev.
2582		 * If the top-level vdev is a log device then just offline
2583		 * it. Otherwise, if this action results in the top-level
2584		 * vdev becoming unusable, undo it and fail the request.
2585		 */
2586		vd->vdev_offline = B_TRUE;
2587		vdev_reopen(tvd);
2588
2589		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2590		    vdev_is_dead(tvd)) {
2591			vd->vdev_offline = B_FALSE;
2592			vdev_reopen(tvd);
2593			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2594		}
2595
2596		/*
2597		 * Add the device back into the metaslab rotor so that
2598		 * once we online the device it's open for business.
2599		 */
2600		if (tvd->vdev_islog && mg != NULL)
2601			metaslab_group_activate(mg);
2602	}
2603
2604	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
2605
2606	return (spa_vdev_state_exit(spa, vd, 0));
2607}
2608
2609int
2610vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
2611{
2612	int error;
2613
2614	mutex_enter(&spa->spa_vdev_top_lock);
2615	error = vdev_offline_locked(spa, guid, flags);
2616	mutex_exit(&spa->spa_vdev_top_lock);
2617
2618	return (error);
2619}
2620
2621/*
2622 * Clear the error counts associated with this vdev.  Unlike vdev_online() and
2623 * vdev_offline(), we assume the spa config is locked.  We also clear all
2624 * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
2625 */
2626void
2627vdev_clear(spa_t *spa, vdev_t *vd)
2628{
2629	vdev_t *rvd = spa->spa_root_vdev;
2630
2631	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2632
2633	if (vd == NULL)
2634		vd = rvd;
2635
2636	vd->vdev_stat.vs_read_errors = 0;
2637	vd->vdev_stat.vs_write_errors = 0;
2638	vd->vdev_stat.vs_checksum_errors = 0;
2639
2640	for (int c = 0; c < vd->vdev_children; c++)
2641		vdev_clear(spa, vd->vdev_child[c]);
2642
2643	/*
2644	 * If we're in the FAULTED state or have experienced failed I/O, then
2645	 * clear the persistent state and attempt to reopen the device.  We
2646	 * also mark the vdev config dirty, so that the new faulted state is
2647	 * written out to disk.
2648	 */
2649	if (vd->vdev_faulted || vd->vdev_degraded ||
2650	    !vdev_readable(vd) || !vdev_writeable(vd)) {
2651
2652		/*
2653		 * When reopening in reponse to a clear event, it may be due to
2654		 * a fmadm repair request.  In this case, if the device is
2655		 * still broken, we want to still post the ereport again.
2656		 */
2657		vd->vdev_forcefault = B_TRUE;
2658
2659		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
2660		vd->vdev_cant_read = B_FALSE;
2661		vd->vdev_cant_write = B_FALSE;
2662
2663		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
2664
2665		vd->vdev_forcefault = B_FALSE;
2666
2667		if (vd != rvd && vdev_writeable(vd->vdev_top))
2668			vdev_state_dirty(vd->vdev_top);
2669
2670		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
2671			spa_async_request(spa, SPA_ASYNC_RESILVER);
2672
2673		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
2674	}
2675
2676	/*
2677	 * When clearing a FMA-diagnosed fault, we always want to
2678	 * unspare the device, as we assume that the original spare was
2679	 * done in response to the FMA fault.
2680	 */
2681	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
2682	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2683	    vd->vdev_parent->vdev_child[0] == vd)
2684		vd->vdev_unspare = B_TRUE;
2685}
2686
2687boolean_t
2688vdev_is_dead(vdev_t *vd)
2689{
2690	/*
2691	 * Holes and missing devices are always considered "dead".
2692	 * This simplifies the code since we don't have to check for
2693	 * these types of devices in the various code paths.
2694	 * Instead we rely on the fact that we skip over dead devices
2695	 * before issuing I/O to them.
2696	 */
2697	return (vd->vdev_state < VDEV_STATE_DEGRADED || vd->vdev_ishole ||
2698	    vd->vdev_ops == &vdev_missing_ops);
2699}
2700
2701boolean_t
2702vdev_readable(vdev_t *vd)
2703{
2704	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
2705}
2706
2707boolean_t
2708vdev_writeable(vdev_t *vd)
2709{
2710	return (!vdev_is_dead(vd) && !vd->vdev_cant_write);
2711}
2712
2713boolean_t
2714vdev_allocatable(vdev_t *vd)
2715{
2716	uint64_t state = vd->vdev_state;
2717
2718	/*
2719	 * We currently allow allocations from vdevs which may be in the
2720	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
2721	 * fails to reopen then we'll catch it later when we're holding
2722	 * the proper locks.  Note that we have to get the vdev state
2723	 * in a local variable because although it changes atomically,
2724	 * we're asking two separate questions about it.
2725	 */
2726	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
2727	    !vd->vdev_cant_write && !vd->vdev_ishole &&
2728	    vd->vdev_mg->mg_initialized);
2729}
2730
2731boolean_t
2732vdev_accessible(vdev_t *vd, zio_t *zio)
2733{
2734	ASSERT(zio->io_vd == vd);
2735
2736	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
2737		return (B_FALSE);
2738
2739	if (zio->io_type == ZIO_TYPE_READ)
2740		return (!vd->vdev_cant_read);
2741
2742	if (zio->io_type == ZIO_TYPE_WRITE)
2743		return (!vd->vdev_cant_write);
2744
2745	return (B_TRUE);
2746}
2747
2748/*
2749 * Get statistics for the given vdev.
2750 */
2751void
2752vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
2753{
2754	spa_t *spa = vd->vdev_spa;
2755	vdev_t *rvd = spa->spa_root_vdev;
2756	vdev_t *tvd = vd->vdev_top;
2757
2758	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2759
2760	mutex_enter(&vd->vdev_stat_lock);
2761	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
2762	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
2763	vs->vs_state = vd->vdev_state;
2764	vs->vs_rsize = vdev_get_min_asize(vd);
2765	if (vd->vdev_ops->vdev_op_leaf)
2766		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
2767	/*
2768	 * Report expandable space on top-level, non-auxillary devices only.
2769	 * The expandable space is reported in terms of metaslab sized units
2770	 * since that determines how much space the pool can expand.
2771	 */
2772	if (vd->vdev_aux == NULL && tvd != NULL) {
2773		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
2774		    spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
2775	}
2776	if (vd->vdev_aux == NULL && vd == vd->vdev_top && !vd->vdev_ishole) {
2777		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
2778	}
2779
2780	/*
2781	 * If we're getting stats on the root vdev, aggregate the I/O counts
2782	 * over all top-level vdevs (i.e. the direct children of the root).
2783	 */
2784	if (vd == rvd) {
2785		for (int c = 0; c < rvd->vdev_children; c++) {
2786			vdev_t *cvd = rvd->vdev_child[c];
2787			vdev_stat_t *cvs = &cvd->vdev_stat;
2788
2789			for (int t = 0; t < ZIO_TYPES; t++) {
2790				vs->vs_ops[t] += cvs->vs_ops[t];
2791				vs->vs_bytes[t] += cvs->vs_bytes[t];
2792			}
2793			cvs->vs_scan_removing = cvd->vdev_removing;
2794		}
2795	}
2796	mutex_exit(&vd->vdev_stat_lock);
2797}
2798
2799void
2800vdev_clear_stats(vdev_t *vd)
2801{
2802	mutex_enter(&vd->vdev_stat_lock);
2803	vd->vdev_stat.vs_space = 0;
2804	vd->vdev_stat.vs_dspace = 0;
2805	vd->vdev_stat.vs_alloc = 0;
2806	mutex_exit(&vd->vdev_stat_lock);
2807}
2808
2809void
2810vdev_scan_stat_init(vdev_t *vd)
2811{
2812	vdev_stat_t *vs = &vd->vdev_stat;
2813
2814	for (int c = 0; c < vd->vdev_children; c++)
2815		vdev_scan_stat_init(vd->vdev_child[c]);
2816
2817	mutex_enter(&vd->vdev_stat_lock);
2818	vs->vs_scan_processed = 0;
2819	mutex_exit(&vd->vdev_stat_lock);
2820}
2821
2822void
2823vdev_stat_update(zio_t *zio, uint64_t psize)
2824{
2825	spa_t *spa = zio->io_spa;
2826	vdev_t *rvd = spa->spa_root_vdev;
2827	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
2828	vdev_t *pvd;
2829	uint64_t txg = zio->io_txg;
2830	vdev_stat_t *vs = &vd->vdev_stat;
2831	zio_type_t type = zio->io_type;
2832	int flags = zio->io_flags;
2833
2834	/*
2835	 * If this i/o is a gang leader, it didn't do any actual work.
2836	 */
2837	if (zio->io_gang_tree)
2838		return;
2839
2840	if (zio->io_error == 0) {
2841		/*
2842		 * If this is a root i/o, don't count it -- we've already
2843		 * counted the top-level vdevs, and vdev_get_stats() will
2844		 * aggregate them when asked.  This reduces contention on
2845		 * the root vdev_stat_lock and implicitly handles blocks
2846		 * that compress away to holes, for which there is no i/o.
2847		 * (Holes never create vdev children, so all the counters
2848		 * remain zero, which is what we want.)
2849		 *
2850		 * Note: this only applies to successful i/o (io_error == 0)
2851		 * because unlike i/o counts, errors are not additive.
2852		 * When reading a ditto block, for example, failure of
2853		 * one top-level vdev does not imply a root-level error.
2854		 */
2855		if (vd == rvd)
2856			return;
2857
2858		ASSERT(vd == zio->io_vd);
2859
2860		if (flags & ZIO_FLAG_IO_BYPASS)
2861			return;
2862
2863		mutex_enter(&vd->vdev_stat_lock);
2864
2865		if (flags & ZIO_FLAG_IO_REPAIR) {
2866			if (flags & ZIO_FLAG_SCAN_THREAD) {
2867				dsl_scan_phys_t *scn_phys =
2868				    &spa->spa_dsl_pool->dp_scan->scn_phys;
2869				uint64_t *processed = &scn_phys->scn_processed;
2870
2871				/* XXX cleanup? */
2872				if (vd->vdev_ops->vdev_op_leaf)
2873					atomic_add_64(processed, psize);
2874				vs->vs_scan_processed += psize;
2875			}
2876
2877			if (flags & ZIO_FLAG_SELF_HEAL)
2878				vs->vs_self_healed += psize;
2879		}
2880
2881		vs->vs_ops[type]++;
2882		vs->vs_bytes[type] += psize;
2883
2884		mutex_exit(&vd->vdev_stat_lock);
2885		return;
2886	}
2887
2888	if (flags & ZIO_FLAG_SPECULATIVE)
2889		return;
2890
2891	/*
2892	 * If this is an I/O error that is going to be retried, then ignore the
2893	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
2894	 * hard errors, when in reality they can happen for any number of
2895	 * innocuous reasons (bus resets, MPxIO link failure, etc).
2896	 */
2897	if (zio->io_error == EIO &&
2898	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
2899		return;
2900
2901	/*
2902	 * Intent logs writes won't propagate their error to the root
2903	 * I/O so don't mark these types of failures as pool-level
2904	 * errors.
2905	 */
2906	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
2907		return;
2908
2909	mutex_enter(&vd->vdev_stat_lock);
2910	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
2911		if (zio->io_error == ECKSUM)
2912			vs->vs_checksum_errors++;
2913		else
2914			vs->vs_read_errors++;
2915	}
2916	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
2917		vs->vs_write_errors++;
2918	mutex_exit(&vd->vdev_stat_lock);
2919
2920	if (type == ZIO_TYPE_WRITE && txg != 0 &&
2921	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
2922	    (flags & ZIO_FLAG_SCAN_THREAD) ||
2923	    spa->spa_claiming)) {
2924		/*
2925		 * This is either a normal write (not a repair), or it's
2926		 * a repair induced by the scrub thread, or it's a repair
2927		 * made by zil_claim() during spa_load() in the first txg.
2928		 * In the normal case, we commit the DTL change in the same
2929		 * txg as the block was born.  In the scrub-induced repair
2930		 * case, we know that scrubs run in first-pass syncing context,
2931		 * so we commit the DTL change in spa_syncing_txg(spa).
2932		 * In the zil_claim() case, we commit in spa_first_txg(spa).
2933		 *
2934		 * We currently do not make DTL entries for failed spontaneous
2935		 * self-healing writes triggered by normal (non-scrubbing)
2936		 * reads, because we have no transactional context in which to
2937		 * do so -- and it's not clear that it'd be desirable anyway.
2938		 */
2939		if (vd->vdev_ops->vdev_op_leaf) {
2940			uint64_t commit_txg = txg;
2941			if (flags & ZIO_FLAG_SCAN_THREAD) {
2942				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2943				ASSERT(spa_sync_pass(spa) == 1);
2944				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
2945				commit_txg = spa_syncing_txg(spa);
2946			} else if (spa->spa_claiming) {
2947				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
2948				commit_txg = spa_first_txg(spa);
2949			}
2950			ASSERT(commit_txg >= spa_syncing_txg(spa));
2951			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
2952				return;
2953			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2954				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
2955			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
2956		}
2957		if (vd != rvd)
2958			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
2959	}
2960}
2961
2962/*
2963 * Update the in-core space usage stats for this vdev, its metaslab class,
2964 * and the root vdev.
2965 */
2966void
2967vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
2968    int64_t space_delta)
2969{
2970	int64_t dspace_delta = space_delta;
2971	spa_t *spa = vd->vdev_spa;
2972	vdev_t *rvd = spa->spa_root_vdev;
2973	metaslab_group_t *mg = vd->vdev_mg;
2974	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
2975
2976	ASSERT(vd == vd->vdev_top);
2977
2978	/*
2979	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
2980	 * factor.  We must calculate this here and not at the root vdev
2981	 * because the root vdev's psize-to-asize is simply the max of its
2982	 * childrens', thus not accurate enough for us.
2983	 */
2984	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
2985	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
2986	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
2987	    vd->vdev_deflate_ratio;
2988
2989	mutex_enter(&vd->vdev_stat_lock);
2990	vd->vdev_stat.vs_alloc += alloc_delta;
2991	vd->vdev_stat.vs_space += space_delta;
2992	vd->vdev_stat.vs_dspace += dspace_delta;
2993	mutex_exit(&vd->vdev_stat_lock);
2994
2995	if (mc == spa_normal_class(spa)) {
2996		mutex_enter(&rvd->vdev_stat_lock);
2997		rvd->vdev_stat.vs_alloc += alloc_delta;
2998		rvd->vdev_stat.vs_space += space_delta;
2999		rvd->vdev_stat.vs_dspace += dspace_delta;
3000		mutex_exit(&rvd->vdev_stat_lock);
3001	}
3002
3003	if (mc != NULL) {
3004		ASSERT(rvd == vd->vdev_parent);
3005		ASSERT(vd->vdev_ms_count != 0);
3006
3007		metaslab_class_space_update(mc,
3008		    alloc_delta, defer_delta, space_delta, dspace_delta);
3009	}
3010}
3011
3012/*
3013 * Mark a top-level vdev's config as dirty, placing it on the dirty list
3014 * so that it will be written out next time the vdev configuration is synced.
3015 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3016 */
3017void
3018vdev_config_dirty(vdev_t *vd)
3019{
3020	spa_t *spa = vd->vdev_spa;
3021	vdev_t *rvd = spa->spa_root_vdev;
3022	int c;
3023
3024	ASSERT(spa_writeable(spa));
3025
3026	/*
3027	 * If this is an aux vdev (as with l2cache and spare devices), then we
3028	 * update the vdev config manually and set the sync flag.
3029	 */
3030	if (vd->vdev_aux != NULL) {
3031		spa_aux_vdev_t *sav = vd->vdev_aux;
3032		nvlist_t **aux;
3033		uint_t naux;
3034
3035		for (c = 0; c < sav->sav_count; c++) {
3036			if (sav->sav_vdevs[c] == vd)
3037				break;
3038		}
3039
3040		if (c == sav->sav_count) {
3041			/*
3042			 * We're being removed.  There's nothing more to do.
3043			 */
3044			ASSERT(sav->sav_sync == B_TRUE);
3045			return;
3046		}
3047
3048		sav->sav_sync = B_TRUE;
3049
3050		if (nvlist_lookup_nvlist_array(sav->sav_config,
3051		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3052			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3053			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3054		}
3055
3056		ASSERT(c < naux);
3057
3058		/*
3059		 * Setting the nvlist in the middle if the array is a little
3060		 * sketchy, but it will work.
3061		 */
3062		nvlist_free(aux[c]);
3063		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3064
3065		return;
3066	}
3067
3068	/*
3069	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3070	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3071	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3072	 * so this is sufficient to ensure mutual exclusion.
3073	 */
3074	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3075	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3076	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3077
3078	if (vd == rvd) {
3079		for (c = 0; c < rvd->vdev_children; c++)
3080			vdev_config_dirty(rvd->vdev_child[c]);
3081	} else {
3082		ASSERT(vd == vd->vdev_top);
3083
3084		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3085		    !vd->vdev_ishole)
3086			list_insert_head(&spa->spa_config_dirty_list, vd);
3087	}
3088}
3089
3090void
3091vdev_config_clean(vdev_t *vd)
3092{
3093	spa_t *spa = vd->vdev_spa;
3094
3095	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3096	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3097	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3098
3099	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3100	list_remove(&spa->spa_config_dirty_list, vd);
3101}
3102
3103/*
3104 * Mark a top-level vdev's state as dirty, so that the next pass of
3105 * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3106 * the state changes from larger config changes because they require
3107 * much less locking, and are often needed for administrative actions.
3108 */
3109void
3110vdev_state_dirty(vdev_t *vd)
3111{
3112	spa_t *spa = vd->vdev_spa;
3113
3114	ASSERT(spa_writeable(spa));
3115	ASSERT(vd == vd->vdev_top);
3116
3117	/*
3118	 * The state list is protected by the SCL_STATE lock.  The caller
3119	 * must either hold SCL_STATE as writer, or must be the sync thread
3120	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3121	 * so this is sufficient to ensure mutual exclusion.
3122	 */
3123	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3124	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3125	    spa_config_held(spa, SCL_STATE, RW_READER)));
3126
3127	if (!list_link_active(&vd->vdev_state_dirty_node) && !vd->vdev_ishole)
3128		list_insert_head(&spa->spa_state_dirty_list, vd);
3129}
3130
3131void
3132vdev_state_clean(vdev_t *vd)
3133{
3134	spa_t *spa = vd->vdev_spa;
3135
3136	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3137	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3138	    spa_config_held(spa, SCL_STATE, RW_READER)));
3139
3140	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3141	list_remove(&spa->spa_state_dirty_list, vd);
3142}
3143
3144/*
3145 * Propagate vdev state up from children to parent.
3146 */
3147void
3148vdev_propagate_state(vdev_t *vd)
3149{
3150	spa_t *spa = vd->vdev_spa;
3151	vdev_t *rvd = spa->spa_root_vdev;
3152	int degraded = 0, faulted = 0;
3153	int corrupted = 0;
3154	vdev_t *child;
3155
3156	if (vd->vdev_children > 0) {
3157		for (int c = 0; c < vd->vdev_children; c++) {
3158			child = vd->vdev_child[c];
3159
3160			/*
3161			 * Don't factor holes into the decision.
3162			 */
3163			if (child->vdev_ishole)
3164				continue;
3165
3166			if (!vdev_readable(child) ||
3167			    (!vdev_writeable(child) && spa_writeable(spa))) {
3168				/*
3169				 * Root special: if there is a top-level log
3170				 * device, treat the root vdev as if it were
3171				 * degraded.
3172				 */
3173				if (child->vdev_islog && vd == rvd)
3174					degraded++;
3175				else
3176					faulted++;
3177			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3178				degraded++;
3179			}
3180
3181			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3182				corrupted++;
3183		}
3184
3185		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3186
3187		/*
3188		 * Root special: if there is a top-level vdev that cannot be
3189		 * opened due to corrupted metadata, then propagate the root
3190		 * vdev's aux state as 'corrupt' rather than 'insufficient
3191		 * replicas'.
3192		 */
3193		if (corrupted && vd == rvd &&
3194		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3195			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3196			    VDEV_AUX_CORRUPT_DATA);
3197	}
3198
3199	if (vd->vdev_parent)
3200		vdev_propagate_state(vd->vdev_parent);
3201}
3202
3203/*
3204 * Set a vdev's state.  If this is during an open, we don't update the parent
3205 * state, because we're in the process of opening children depth-first.
3206 * Otherwise, we propagate the change to the parent.
3207 *
3208 * If this routine places a device in a faulted state, an appropriate ereport is
3209 * generated.
3210 */
3211void
3212vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3213{
3214	uint64_t save_state;
3215	spa_t *spa = vd->vdev_spa;
3216
3217	if (state == vd->vdev_state) {
3218		vd->vdev_stat.vs_aux = aux;
3219		return;
3220	}
3221
3222	save_state = vd->vdev_state;
3223
3224	vd->vdev_state = state;
3225	vd->vdev_stat.vs_aux = aux;
3226
3227	/*
3228	 * If we are setting the vdev state to anything but an open state, then
3229	 * always close the underlying device unless the device has requested
3230	 * a delayed close (i.e. we're about to remove or fault the device).
3231	 * Otherwise, we keep accessible but invalid devices open forever.
3232	 * We don't call vdev_close() itself, because that implies some extra
3233	 * checks (offline, etc) that we don't want here.  This is limited to
3234	 * leaf devices, because otherwise closing the device will affect other
3235	 * children.
3236	 */
3237	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3238	    vd->vdev_ops->vdev_op_leaf)
3239		vd->vdev_ops->vdev_op_close(vd);
3240
3241	/*
3242	 * If we have brought this vdev back into service, we need
3243	 * to notify fmd so that it can gracefully repair any outstanding
3244	 * cases due to a missing device.  We do this in all cases, even those
3245	 * that probably don't correlate to a repaired fault.  This is sure to
3246	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3247	 * this is a transient state it's OK, as the retire agent will
3248	 * double-check the state of the vdev before repairing it.
3249	 */
3250	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3251	    vd->vdev_prevstate != state)
3252		zfs_post_state_change(spa, vd);
3253
3254	if (vd->vdev_removed &&
3255	    state == VDEV_STATE_CANT_OPEN &&
3256	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3257		/*
3258		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3259		 * device was previously marked removed and someone attempted to
3260		 * reopen it.  If this failed due to a nonexistent device, then
3261		 * keep the device in the REMOVED state.  We also let this be if
3262		 * it is one of our special test online cases, which is only
3263		 * attempting to online the device and shouldn't generate an FMA
3264		 * fault.
3265		 */
3266		vd->vdev_state = VDEV_STATE_REMOVED;
3267		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3268	} else if (state == VDEV_STATE_REMOVED) {
3269		vd->vdev_removed = B_TRUE;
3270	} else if (state == VDEV_STATE_CANT_OPEN) {
3271		/*
3272		 * If we fail to open a vdev during an import or recovery, we
3273		 * mark it as "not available", which signifies that it was
3274		 * never there to begin with.  Failure to open such a device
3275		 * is not considered an error.
3276		 */
3277		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3278		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3279		    vd->vdev_ops->vdev_op_leaf)
3280			vd->vdev_not_present = 1;
3281
3282		/*
3283		 * Post the appropriate ereport.  If the 'prevstate' field is
3284		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3285		 * that this is part of a vdev_reopen().  In this case, we don't
3286		 * want to post the ereport if the device was already in the
3287		 * CANT_OPEN state beforehand.
3288		 *
3289		 * If the 'checkremove' flag is set, then this is an attempt to
3290		 * online the device in response to an insertion event.  If we
3291		 * hit this case, then we have detected an insertion event for a
3292		 * faulted or offline device that wasn't in the removed state.
3293		 * In this scenario, we don't post an ereport because we are
3294		 * about to replace the device, or attempt an online with
3295		 * vdev_forcefault, which will generate the fault for us.
3296		 */
3297		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3298		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3299		    vd != spa->spa_root_vdev) {
3300			const char *class;
3301
3302			switch (aux) {
3303			case VDEV_AUX_OPEN_FAILED:
3304				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3305				break;
3306			case VDEV_AUX_CORRUPT_DATA:
3307				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3308				break;
3309			case VDEV_AUX_NO_REPLICAS:
3310				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3311				break;
3312			case VDEV_AUX_BAD_GUID_SUM:
3313				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3314				break;
3315			case VDEV_AUX_TOO_SMALL:
3316				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3317				break;
3318			case VDEV_AUX_BAD_LABEL:
3319				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3320				break;
3321			default:
3322				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3323			}
3324
3325			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3326		}
3327
3328		/* Erase any notion of persistent removed state */
3329		vd->vdev_removed = B_FALSE;
3330	} else {
3331		vd->vdev_removed = B_FALSE;
3332	}
3333
3334	if (!isopen && vd->vdev_parent)
3335		vdev_propagate_state(vd->vdev_parent);
3336}
3337
3338/*
3339 * Check the vdev configuration to ensure that it's capable of supporting
3340 * a root pool. We do not support partial configuration.
3341 * In addition, only a single top-level vdev is allowed.
3342 */
3343boolean_t
3344vdev_is_bootable(vdev_t *vd)
3345{
3346	if (!vd->vdev_ops->vdev_op_leaf) {
3347		char *vdev_type = vd->vdev_ops->vdev_op_type;
3348
3349		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3350		    vd->vdev_children > 1) {
3351			return (B_FALSE);
3352		} else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0) {
3353			return (B_FALSE);
3354		}
3355	}
3356
3357	for (int c = 0; c < vd->vdev_children; c++) {
3358		if (!vdev_is_bootable(vd->vdev_child[c]))
3359			return (B_FALSE);
3360	}
3361	return (B_TRUE);
3362}
3363
3364/*
3365 * Load the state from the original vdev tree (ovd) which
3366 * we've retrieved from the MOS config object. If the original
3367 * vdev was offline or faulted then we transfer that state to the
3368 * device in the current vdev tree (nvd).
3369 */
3370void
3371vdev_load_log_state(vdev_t *nvd, vdev_t *ovd)
3372{
3373	spa_t *spa = nvd->vdev_spa;
3374
3375	ASSERT(nvd->vdev_top->vdev_islog);
3376	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3377	ASSERT3U(nvd->vdev_guid, ==, ovd->vdev_guid);
3378
3379	for (int c = 0; c < nvd->vdev_children; c++)
3380		vdev_load_log_state(nvd->vdev_child[c], ovd->vdev_child[c]);
3381
3382	if (nvd->vdev_ops->vdev_op_leaf) {
3383		/*
3384		 * Restore the persistent vdev state
3385		 */
3386		nvd->vdev_offline = ovd->vdev_offline;
3387		nvd->vdev_faulted = ovd->vdev_faulted;
3388		nvd->vdev_degraded = ovd->vdev_degraded;
3389		nvd->vdev_removed = ovd->vdev_removed;
3390	}
3391}
3392
3393/*
3394 * Determine if a log device has valid content.  If the vdev was
3395 * removed or faulted in the MOS config then we know that
3396 * the content on the log device has already been written to the pool.
3397 */
3398boolean_t
3399vdev_log_state_valid(vdev_t *vd)
3400{
3401	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3402	    !vd->vdev_removed)
3403		return (B_TRUE);
3404
3405	for (int c = 0; c < vd->vdev_children; c++)
3406		if (vdev_log_state_valid(vd->vdev_child[c]))
3407			return (B_TRUE);
3408
3409	return (B_FALSE);
3410}
3411
3412/*
3413 * Expand a vdev if possible.
3414 */
3415void
3416vdev_expand(vdev_t *vd, uint64_t txg)
3417{
3418	ASSERT(vd->vdev_top == vd);
3419	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3420
3421	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count) {
3422		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3423		vdev_config_dirty(vd);
3424	}
3425}
3426
3427/*
3428 * Split a vdev.
3429 */
3430void
3431vdev_split(vdev_t *vd)
3432{
3433	vdev_t *cvd, *pvd = vd->vdev_parent;
3434
3435	vdev_remove_child(pvd, vd);
3436	vdev_compact_children(pvd);
3437
3438	cvd = pvd->vdev_child[0];
3439	if (pvd->vdev_children == 1) {
3440		vdev_remove_parent(cvd);
3441		cvd->vdev_splitting = B_TRUE;
3442	}
3443	vdev_propagate_state(cvd);
3444}
3445
3446void
3447vdev_deadman(vdev_t *vd)
3448{
3449	for (int c = 0; c < vd->vdev_children; c++) {
3450		vdev_t *cvd = vd->vdev_child[c];
3451
3452		vdev_deadman(cvd);
3453	}
3454
3455	if (vd->vdev_ops->vdev_op_leaf) {
3456		vdev_queue_t *vq = &vd->vdev_queue;
3457
3458		mutex_enter(&vq->vq_lock);
3459		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3460			spa_t *spa = vd->vdev_spa;
3461			zio_t *fio;
3462			uint64_t delta;
3463
3464			/*
3465			 * Look at the head of all the pending queues,
3466			 * if any I/O has been outstanding for longer than
3467			 * the spa_deadman_synctime we panic the system.
3468			 */
3469			fio = avl_first(&vq->vq_active_tree);
3470			delta = gethrtime() - fio->io_timestamp;
3471			if (delta > spa_deadman_synctime(spa)) {
3472				zfs_dbgmsg("SLOW IO: zio timestamp %lluns, "
3473				    "delta %lluns, last io %lluns",
3474				    fio->io_timestamp, delta,
3475				    vq->vq_io_complete_ts);
3476				fm_panic("I/O to pool '%s' appears to be "
3477				    "hung.", spa_name(spa));
3478			}
3479		}
3480		mutex_exit(&vq->vq_lock);
3481	}
3482}
3483