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