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