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