1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22/*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
25 * Copyright 2017 Nexenta Systems, Inc.
26 * Copyright (c) 2014 Integros [integros.com]
27 * Copyright 2016 Toomas Soome <tsoome@me.com>
28 * Copyright 2017 Joyent, Inc.
29 */
30
31#include <sys/zfs_context.h>
32#include <sys/fm/fs/zfs.h>
33#include <sys/spa.h>
34#include <sys/spa_impl.h>
35#include <sys/bpobj.h>
36#include <sys/dmu.h>
37#include <sys/dmu_tx.h>
38#include <sys/dsl_dir.h>
39#include <sys/vdev_impl.h>
40#include <sys/uberblock_impl.h>
41#include <sys/metaslab.h>
42#include <sys/metaslab_impl.h>
43#include <sys/space_map.h>
44#include <sys/space_reftree.h>
45#include <sys/zio.h>
46#include <sys/zap.h>
47#include <sys/fs/zfs.h>
48#include <sys/arc.h>
49#include <sys/zil.h>
50#include <sys/dsl_scan.h>
51#include <sys/abd.h>
52
53/*
54 * Virtual device management.
55 */
56
57static vdev_ops_t *vdev_ops_table[] = {
58	&vdev_root_ops,
59	&vdev_raidz_ops,
60	&vdev_mirror_ops,
61	&vdev_replacing_ops,
62	&vdev_spare_ops,
63	&vdev_disk_ops,
64	&vdev_file_ops,
65	&vdev_missing_ops,
66	&vdev_hole_ops,
67	&vdev_indirect_ops,
68	NULL
69};
70
71/* maximum scrub/resilver I/O queue per leaf vdev */
72int zfs_scrub_limit = 10;
73
74/*
75 * When a vdev is added, it will be divided into approximately (but no
76 * more than) this number of metaslabs.
77 */
78int metaslabs_per_vdev = 200;
79
80boolean_t vdev_validate_skip = B_FALSE;
81
82/*PRINTFLIKE2*/
83void
84vdev_dbgmsg(vdev_t *vd, const char *fmt, ...)
85{
86	va_list adx;
87	char buf[256];
88
89	va_start(adx, fmt);
90	(void) vsnprintf(buf, sizeof (buf), fmt, adx);
91	va_end(adx);
92
93	if (vd->vdev_path != NULL) {
94		zfs_dbgmsg("%s vdev '%s': %s", vd->vdev_ops->vdev_op_type,
95		    vd->vdev_path, buf);
96	} else {
97		zfs_dbgmsg("%s-%llu vdev (guid %llu): %s",
98		    vd->vdev_ops->vdev_op_type,
99		    (u_longlong_t)vd->vdev_id,
100		    (u_longlong_t)vd->vdev_guid, buf);
101	}
102}
103
104void
105vdev_dbgmsg_print_tree(vdev_t *vd, int indent)
106{
107	char state[20];
108
109	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops) {
110		zfs_dbgmsg("%*svdev %u: %s", indent, "", vd->vdev_id,
111		    vd->vdev_ops->vdev_op_type);
112		return;
113	}
114
115	switch (vd->vdev_state) {
116	case VDEV_STATE_UNKNOWN:
117		(void) snprintf(state, sizeof (state), "unknown");
118		break;
119	case VDEV_STATE_CLOSED:
120		(void) snprintf(state, sizeof (state), "closed");
121		break;
122	case VDEV_STATE_OFFLINE:
123		(void) snprintf(state, sizeof (state), "offline");
124		break;
125	case VDEV_STATE_REMOVED:
126		(void) snprintf(state, sizeof (state), "removed");
127		break;
128	case VDEV_STATE_CANT_OPEN:
129		(void) snprintf(state, sizeof (state), "can't open");
130		break;
131	case VDEV_STATE_FAULTED:
132		(void) snprintf(state, sizeof (state), "faulted");
133		break;
134	case VDEV_STATE_DEGRADED:
135		(void) snprintf(state, sizeof (state), "degraded");
136		break;
137	case VDEV_STATE_HEALTHY:
138		(void) snprintf(state, sizeof (state), "healthy");
139		break;
140	default:
141		(void) snprintf(state, sizeof (state), "<state %u>",
142		    (uint_t)vd->vdev_state);
143	}
144
145	zfs_dbgmsg("%*svdev %u: %s%s, guid: %llu, path: %s, %s", indent,
146	    "", vd->vdev_id, vd->vdev_ops->vdev_op_type,
147	    vd->vdev_islog ? " (log)" : "",
148	    (u_longlong_t)vd->vdev_guid,
149	    vd->vdev_path ? vd->vdev_path : "N/A", state);
150
151	for (uint64_t i = 0; i < vd->vdev_children; i++)
152		vdev_dbgmsg_print_tree(vd->vdev_child[i], indent + 2);
153}
154
155/*
156 * Given a vdev type, return the appropriate ops vector.
157 */
158static vdev_ops_t *
159vdev_getops(const char *type)
160{
161	vdev_ops_t *ops, **opspp;
162
163	for (opspp = vdev_ops_table; (ops = *opspp) != NULL; opspp++)
164		if (strcmp(ops->vdev_op_type, type) == 0)
165			break;
166
167	return (ops);
168}
169
170/*
171 * Default asize function: return the MAX of psize with the asize of
172 * all children.  This is what's used by anything other than RAID-Z.
173 */
174uint64_t
175vdev_default_asize(vdev_t *vd, uint64_t psize)
176{
177	uint64_t asize = P2ROUNDUP(psize, 1ULL << vd->vdev_top->vdev_ashift);
178	uint64_t csize;
179
180	for (int c = 0; c < vd->vdev_children; c++) {
181		csize = vdev_psize_to_asize(vd->vdev_child[c], psize);
182		asize = MAX(asize, csize);
183	}
184
185	return (asize);
186}
187
188/*
189 * Get the minimum allocatable size. We define the allocatable size as
190 * the vdev's asize rounded to the nearest metaslab. This allows us to
191 * replace or attach devices which don't have the same physical size but
192 * can still satisfy the same number of allocations.
193 */
194uint64_t
195vdev_get_min_asize(vdev_t *vd)
196{
197	vdev_t *pvd = vd->vdev_parent;
198
199	/*
200	 * If our parent is NULL (inactive spare or cache) or is the root,
201	 * just return our own asize.
202	 */
203	if (pvd == NULL)
204		return (vd->vdev_asize);
205
206	/*
207	 * The top-level vdev just returns the allocatable size rounded
208	 * to the nearest metaslab.
209	 */
210	if (vd == vd->vdev_top)
211		return (P2ALIGN(vd->vdev_asize, 1ULL << vd->vdev_ms_shift));
212
213	/*
214	 * The allocatable space for a raidz vdev is N * sizeof(smallest child),
215	 * so each child must provide at least 1/Nth of its asize.
216	 */
217	if (pvd->vdev_ops == &vdev_raidz_ops)
218		return ((pvd->vdev_min_asize + pvd->vdev_children - 1) /
219		    pvd->vdev_children);
220
221	return (pvd->vdev_min_asize);
222}
223
224void
225vdev_set_min_asize(vdev_t *vd)
226{
227	vd->vdev_min_asize = vdev_get_min_asize(vd);
228
229	for (int c = 0; c < vd->vdev_children; c++)
230		vdev_set_min_asize(vd->vdev_child[c]);
231}
232
233vdev_t *
234vdev_lookup_top(spa_t *spa, uint64_t vdev)
235{
236	vdev_t *rvd = spa->spa_root_vdev;
237
238	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
239
240	if (vdev < rvd->vdev_children) {
241		ASSERT(rvd->vdev_child[vdev] != NULL);
242		return (rvd->vdev_child[vdev]);
243	}
244
245	return (NULL);
246}
247
248vdev_t *
249vdev_lookup_by_guid(vdev_t *vd, uint64_t guid)
250{
251	vdev_t *mvd;
252
253	if (vd->vdev_guid == guid)
254		return (vd);
255
256	for (int c = 0; c < vd->vdev_children; c++)
257		if ((mvd = vdev_lookup_by_guid(vd->vdev_child[c], guid)) !=
258		    NULL)
259			return (mvd);
260
261	return (NULL);
262}
263
264static int
265vdev_count_leaves_impl(vdev_t *vd)
266{
267	int n = 0;
268
269	if (vd->vdev_ops->vdev_op_leaf)
270		return (1);
271
272	for (int c = 0; c < vd->vdev_children; c++)
273		n += vdev_count_leaves_impl(vd->vdev_child[c]);
274
275	return (n);
276}
277
278int
279vdev_count_leaves(spa_t *spa)
280{
281	return (vdev_count_leaves_impl(spa->spa_root_vdev));
282}
283
284void
285vdev_add_child(vdev_t *pvd, vdev_t *cvd)
286{
287	size_t oldsize, newsize;
288	uint64_t id = cvd->vdev_id;
289	vdev_t **newchild;
290	spa_t *spa = cvd->vdev_spa;
291
292	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
293	ASSERT(cvd->vdev_parent == NULL);
294
295	cvd->vdev_parent = pvd;
296
297	if (pvd == NULL)
298		return;
299
300	ASSERT(id >= pvd->vdev_children || pvd->vdev_child[id] == NULL);
301
302	oldsize = pvd->vdev_children * sizeof (vdev_t *);
303	pvd->vdev_children = MAX(pvd->vdev_children, id + 1);
304	newsize = pvd->vdev_children * sizeof (vdev_t *);
305
306	newchild = kmem_zalloc(newsize, KM_SLEEP);
307	if (pvd->vdev_child != NULL) {
308		bcopy(pvd->vdev_child, newchild, oldsize);
309		kmem_free(pvd->vdev_child, oldsize);
310	}
311
312	pvd->vdev_child = newchild;
313	pvd->vdev_child[id] = cvd;
314
315	cvd->vdev_top = (pvd->vdev_top ? pvd->vdev_top: cvd);
316	ASSERT(cvd->vdev_top->vdev_parent->vdev_parent == NULL);
317
318	/*
319	 * Walk up all ancestors to update guid sum.
320	 */
321	for (; pvd != NULL; pvd = pvd->vdev_parent)
322		pvd->vdev_guid_sum += cvd->vdev_guid_sum;
323}
324
325void
326vdev_remove_child(vdev_t *pvd, vdev_t *cvd)
327{
328	int c;
329	uint_t id = cvd->vdev_id;
330
331	ASSERT(cvd->vdev_parent == pvd);
332
333	if (pvd == NULL)
334		return;
335
336	ASSERT(id < pvd->vdev_children);
337	ASSERT(pvd->vdev_child[id] == cvd);
338
339	pvd->vdev_child[id] = NULL;
340	cvd->vdev_parent = NULL;
341
342	for (c = 0; c < pvd->vdev_children; c++)
343		if (pvd->vdev_child[c])
344			break;
345
346	if (c == pvd->vdev_children) {
347		kmem_free(pvd->vdev_child, c * sizeof (vdev_t *));
348		pvd->vdev_child = NULL;
349		pvd->vdev_children = 0;
350	}
351
352	/*
353	 * Walk up all ancestors to update guid sum.
354	 */
355	for (; pvd != NULL; pvd = pvd->vdev_parent)
356		pvd->vdev_guid_sum -= cvd->vdev_guid_sum;
357}
358
359/*
360 * Remove any holes in the child array.
361 */
362void
363vdev_compact_children(vdev_t *pvd)
364{
365	vdev_t **newchild, *cvd;
366	int oldc = pvd->vdev_children;
367	int newc;
368
369	ASSERT(spa_config_held(pvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
370
371	for (int c = newc = 0; c < oldc; c++)
372		if (pvd->vdev_child[c])
373			newc++;
374
375	newchild = kmem_alloc(newc * sizeof (vdev_t *), KM_SLEEP);
376
377	for (int c = newc = 0; c < oldc; c++) {
378		if ((cvd = pvd->vdev_child[c]) != NULL) {
379			newchild[newc] = cvd;
380			cvd->vdev_id = newc++;
381		}
382	}
383
384	kmem_free(pvd->vdev_child, oldc * sizeof (vdev_t *));
385	pvd->vdev_child = newchild;
386	pvd->vdev_children = newc;
387}
388
389/*
390 * Allocate and minimally initialize a vdev_t.
391 */
392vdev_t *
393vdev_alloc_common(spa_t *spa, uint_t id, uint64_t guid, vdev_ops_t *ops)
394{
395	vdev_t *vd;
396	vdev_indirect_config_t *vic;
397
398	vd = kmem_zalloc(sizeof (vdev_t), KM_SLEEP);
399	vic = &vd->vdev_indirect_config;
400
401	if (spa->spa_root_vdev == NULL) {
402		ASSERT(ops == &vdev_root_ops);
403		spa->spa_root_vdev = vd;
404		spa->spa_load_guid = spa_generate_guid(NULL);
405	}
406
407	if (guid == 0 && ops != &vdev_hole_ops) {
408		if (spa->spa_root_vdev == vd) {
409			/*
410			 * The root vdev's guid will also be the pool guid,
411			 * which must be unique among all pools.
412			 */
413			guid = spa_generate_guid(NULL);
414		} else {
415			/*
416			 * Any other vdev's guid must be unique within the pool.
417			 */
418			guid = spa_generate_guid(spa);
419		}
420		ASSERT(!spa_guid_exists(spa_guid(spa), guid));
421	}
422
423	vd->vdev_spa = spa;
424	vd->vdev_id = id;
425	vd->vdev_guid = guid;
426	vd->vdev_guid_sum = guid;
427	vd->vdev_ops = ops;
428	vd->vdev_state = VDEV_STATE_CLOSED;
429	vd->vdev_ishole = (ops == &vdev_hole_ops);
430	vic->vic_prev_indirect_vdev = UINT64_MAX;
431
432	rw_init(&vd->vdev_indirect_rwlock, NULL, RW_DEFAULT, NULL);
433	mutex_init(&vd->vdev_obsolete_lock, NULL, MUTEX_DEFAULT, NULL);
434	vd->vdev_obsolete_segments = range_tree_create(NULL, NULL);
435
436	mutex_init(&vd->vdev_dtl_lock, NULL, MUTEX_DEFAULT, NULL);
437	mutex_init(&vd->vdev_stat_lock, NULL, MUTEX_DEFAULT, NULL);
438	mutex_init(&vd->vdev_probe_lock, NULL, MUTEX_DEFAULT, NULL);
439	mutex_init(&vd->vdev_queue_lock, NULL, MUTEX_DEFAULT, NULL);
440	for (int t = 0; t < DTL_TYPES; t++) {
441		vd->vdev_dtl[t] = range_tree_create(NULL, NULL);
442	}
443	txg_list_create(&vd->vdev_ms_list, spa,
444	    offsetof(struct metaslab, ms_txg_node));
445	txg_list_create(&vd->vdev_dtl_list, spa,
446	    offsetof(struct vdev, vdev_dtl_node));
447	vd->vdev_stat.vs_timestamp = gethrtime();
448	vdev_queue_init(vd);
449	vdev_cache_init(vd);
450
451	return (vd);
452}
453
454/*
455 * Allocate a new vdev.  The 'alloctype' is used to control whether we are
456 * creating a new vdev or loading an existing one - the behavior is slightly
457 * different for each case.
458 */
459int
460vdev_alloc(spa_t *spa, vdev_t **vdp, nvlist_t *nv, vdev_t *parent, uint_t id,
461    int alloctype)
462{
463	vdev_ops_t *ops;
464	char *type;
465	uint64_t guid = 0, islog, nparity;
466	vdev_t *vd;
467	vdev_indirect_config_t *vic;
468
469	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
470
471	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_TYPE, &type) != 0)
472		return (SET_ERROR(EINVAL));
473
474	if ((ops = vdev_getops(type)) == NULL)
475		return (SET_ERROR(EINVAL));
476
477	/*
478	 * If this is a load, get the vdev guid from the nvlist.
479	 * Otherwise, vdev_alloc_common() will generate one for us.
480	 */
481	if (alloctype == VDEV_ALLOC_LOAD) {
482		uint64_t label_id;
483
484		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ID, &label_id) ||
485		    label_id != id)
486			return (SET_ERROR(EINVAL));
487
488		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
489			return (SET_ERROR(EINVAL));
490	} else if (alloctype == VDEV_ALLOC_SPARE) {
491		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
492			return (SET_ERROR(EINVAL));
493	} else if (alloctype == VDEV_ALLOC_L2CACHE) {
494		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
495			return (SET_ERROR(EINVAL));
496	} else if (alloctype == VDEV_ALLOC_ROOTPOOL) {
497		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_GUID, &guid) != 0)
498			return (SET_ERROR(EINVAL));
499	}
500
501	/*
502	 * The first allocated vdev must be of type 'root'.
503	 */
504	if (ops != &vdev_root_ops && spa->spa_root_vdev == NULL)
505		return (SET_ERROR(EINVAL));
506
507	/*
508	 * Determine whether we're a log vdev.
509	 */
510	islog = 0;
511	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_LOG, &islog);
512	if (islog && spa_version(spa) < SPA_VERSION_SLOGS)
513		return (SET_ERROR(ENOTSUP));
514
515	if (ops == &vdev_hole_ops && spa_version(spa) < SPA_VERSION_HOLES)
516		return (SET_ERROR(ENOTSUP));
517
518	/*
519	 * Set the nparity property for RAID-Z vdevs.
520	 */
521	nparity = -1ULL;
522	if (ops == &vdev_raidz_ops) {
523		if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NPARITY,
524		    &nparity) == 0) {
525			if (nparity == 0 || nparity > VDEV_RAIDZ_MAXPARITY)
526				return (SET_ERROR(EINVAL));
527			/*
528			 * Previous versions could only support 1 or 2 parity
529			 * device.
530			 */
531			if (nparity > 1 &&
532			    spa_version(spa) < SPA_VERSION_RAIDZ2)
533				return (SET_ERROR(ENOTSUP));
534			if (nparity > 2 &&
535			    spa_version(spa) < SPA_VERSION_RAIDZ3)
536				return (SET_ERROR(ENOTSUP));
537		} else {
538			/*
539			 * We require the parity to be specified for SPAs that
540			 * support multiple parity levels.
541			 */
542			if (spa_version(spa) >= SPA_VERSION_RAIDZ2)
543				return (SET_ERROR(EINVAL));
544			/*
545			 * Otherwise, we default to 1 parity device for RAID-Z.
546			 */
547			nparity = 1;
548		}
549	} else {
550		nparity = 0;
551	}
552	ASSERT(nparity != -1ULL);
553
554	vd = vdev_alloc_common(spa, id, guid, ops);
555	vic = &vd->vdev_indirect_config;
556
557	vd->vdev_islog = islog;
558	vd->vdev_nparity = nparity;
559
560	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PATH, &vd->vdev_path) == 0)
561		vd->vdev_path = spa_strdup(vd->vdev_path);
562	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_DEVID, &vd->vdev_devid) == 0)
563		vd->vdev_devid = spa_strdup(vd->vdev_devid);
564	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_PHYS_PATH,
565	    &vd->vdev_physpath) == 0)
566		vd->vdev_physpath = spa_strdup(vd->vdev_physpath);
567	if (nvlist_lookup_string(nv, ZPOOL_CONFIG_FRU, &vd->vdev_fru) == 0)
568		vd->vdev_fru = spa_strdup(vd->vdev_fru);
569
570	/*
571	 * Set the whole_disk property.  If it's not specified, leave the value
572	 * as -1.
573	 */
574	if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_WHOLE_DISK,
575	    &vd->vdev_wholedisk) != 0)
576		vd->vdev_wholedisk = -1ULL;
577
578	ASSERT0(vic->vic_mapping_object);
579	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_OBJECT,
580	    &vic->vic_mapping_object);
581	ASSERT0(vic->vic_births_object);
582	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_INDIRECT_BIRTHS,
583	    &vic->vic_births_object);
584	ASSERT3U(vic->vic_prev_indirect_vdev, ==, UINT64_MAX);
585	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_PREV_INDIRECT_VDEV,
586	    &vic->vic_prev_indirect_vdev);
587
588	/*
589	 * Look for the 'not present' flag.  This will only be set if the device
590	 * was not present at the time of import.
591	 */
592	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_NOT_PRESENT,
593	    &vd->vdev_not_present);
594
595	/*
596	 * Get the alignment requirement.
597	 */
598	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASHIFT, &vd->vdev_ashift);
599
600	/*
601	 * Retrieve the vdev creation time.
602	 */
603	(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_CREATE_TXG,
604	    &vd->vdev_crtxg);
605
606	/*
607	 * If we're a top-level vdev, try to load the allocation parameters.
608	 */
609	if (parent && !parent->vdev_parent &&
610	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
611		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_ARRAY,
612		    &vd->vdev_ms_array);
613		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_METASLAB_SHIFT,
614		    &vd->vdev_ms_shift);
615		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_ASIZE,
616		    &vd->vdev_asize);
617		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVING,
618		    &vd->vdev_removing);
619		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_VDEV_TOP_ZAP,
620		    &vd->vdev_top_zap);
621	} else {
622		ASSERT0(vd->vdev_top_zap);
623	}
624
625	if (parent && !parent->vdev_parent && alloctype != VDEV_ALLOC_ATTACH) {
626		ASSERT(alloctype == VDEV_ALLOC_LOAD ||
627		    alloctype == VDEV_ALLOC_ADD ||
628		    alloctype == VDEV_ALLOC_SPLIT ||
629		    alloctype == VDEV_ALLOC_ROOTPOOL);
630		vd->vdev_mg = metaslab_group_create(islog ?
631		    spa_log_class(spa) : spa_normal_class(spa), vd);
632	}
633
634	if (vd->vdev_ops->vdev_op_leaf &&
635	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_SPLIT)) {
636		(void) nvlist_lookup_uint64(nv,
637		    ZPOOL_CONFIG_VDEV_LEAF_ZAP, &vd->vdev_leaf_zap);
638	} else {
639		ASSERT0(vd->vdev_leaf_zap);
640	}
641
642	/*
643	 * If we're a leaf vdev, try to load the DTL object and other state.
644	 */
645
646	if (vd->vdev_ops->vdev_op_leaf &&
647	    (alloctype == VDEV_ALLOC_LOAD || alloctype == VDEV_ALLOC_L2CACHE ||
648	    alloctype == VDEV_ALLOC_ROOTPOOL)) {
649		if (alloctype == VDEV_ALLOC_LOAD) {
650			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DTL,
651			    &vd->vdev_dtl_object);
652			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_UNSPARE,
653			    &vd->vdev_unspare);
654		}
655
656		if (alloctype == VDEV_ALLOC_ROOTPOOL) {
657			uint64_t spare = 0;
658
659			if (nvlist_lookup_uint64(nv, ZPOOL_CONFIG_IS_SPARE,
660			    &spare) == 0 && spare)
661				spa_spare_add(vd);
662		}
663
664		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_OFFLINE,
665		    &vd->vdev_offline);
666
667		(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_RESILVER_TXG,
668		    &vd->vdev_resilver_txg);
669
670		/*
671		 * When importing a pool, we want to ignore the persistent fault
672		 * state, as the diagnosis made on another system may not be
673		 * valid in the current context.  Local vdevs will
674		 * remain in the faulted state.
675		 */
676		if (spa_load_state(spa) == SPA_LOAD_OPEN) {
677			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_FAULTED,
678			    &vd->vdev_faulted);
679			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_DEGRADED,
680			    &vd->vdev_degraded);
681			(void) nvlist_lookup_uint64(nv, ZPOOL_CONFIG_REMOVED,
682			    &vd->vdev_removed);
683
684			if (vd->vdev_faulted || vd->vdev_degraded) {
685				char *aux;
686
687				vd->vdev_label_aux =
688				    VDEV_AUX_ERR_EXCEEDED;
689				if (nvlist_lookup_string(nv,
690				    ZPOOL_CONFIG_AUX_STATE, &aux) == 0 &&
691				    strcmp(aux, "external") == 0)
692					vd->vdev_label_aux = VDEV_AUX_EXTERNAL;
693			}
694		}
695	}
696
697	/*
698	 * Add ourselves to the parent's list of children.
699	 */
700	vdev_add_child(parent, vd);
701
702	*vdp = vd;
703
704	return (0);
705}
706
707void
708vdev_free(vdev_t *vd)
709{
710	spa_t *spa = vd->vdev_spa;
711
712	/*
713	 * vdev_free() implies closing the vdev first.  This is simpler than
714	 * trying to ensure complicated semantics for all callers.
715	 */
716	vdev_close(vd);
717
718	ASSERT(!list_link_active(&vd->vdev_config_dirty_node));
719	ASSERT(!list_link_active(&vd->vdev_state_dirty_node));
720
721	/*
722	 * Free all children.
723	 */
724	for (int c = 0; c < vd->vdev_children; c++)
725		vdev_free(vd->vdev_child[c]);
726
727	ASSERT(vd->vdev_child == NULL);
728	ASSERT(vd->vdev_guid_sum == vd->vdev_guid);
729
730	/*
731	 * Discard allocation state.
732	 */
733	if (vd->vdev_mg != NULL) {
734		vdev_metaslab_fini(vd);
735		metaslab_group_destroy(vd->vdev_mg);
736	}
737
738	ASSERT0(vd->vdev_stat.vs_space);
739	ASSERT0(vd->vdev_stat.vs_dspace);
740	ASSERT0(vd->vdev_stat.vs_alloc);
741
742	/*
743	 * Remove this vdev from its parent's child list.
744	 */
745	vdev_remove_child(vd->vdev_parent, vd);
746
747	ASSERT(vd->vdev_parent == NULL);
748
749	/*
750	 * Clean up vdev structure.
751	 */
752	vdev_queue_fini(vd);
753	vdev_cache_fini(vd);
754
755	if (vd->vdev_path)
756		spa_strfree(vd->vdev_path);
757	if (vd->vdev_devid)
758		spa_strfree(vd->vdev_devid);
759	if (vd->vdev_physpath)
760		spa_strfree(vd->vdev_physpath);
761	if (vd->vdev_fru)
762		spa_strfree(vd->vdev_fru);
763
764	if (vd->vdev_isspare)
765		spa_spare_remove(vd);
766	if (vd->vdev_isl2cache)
767		spa_l2cache_remove(vd);
768
769	txg_list_destroy(&vd->vdev_ms_list);
770	txg_list_destroy(&vd->vdev_dtl_list);
771
772	mutex_enter(&vd->vdev_dtl_lock);
773	space_map_close(vd->vdev_dtl_sm);
774	for (int t = 0; t < DTL_TYPES; t++) {
775		range_tree_vacate(vd->vdev_dtl[t], NULL, NULL);
776		range_tree_destroy(vd->vdev_dtl[t]);
777	}
778	mutex_exit(&vd->vdev_dtl_lock);
779
780	EQUIV(vd->vdev_indirect_births != NULL,
781	    vd->vdev_indirect_mapping != NULL);
782	if (vd->vdev_indirect_births != NULL) {
783		vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
784		vdev_indirect_births_close(vd->vdev_indirect_births);
785	}
786
787	if (vd->vdev_obsolete_sm != NULL) {
788		ASSERT(vd->vdev_removing ||
789		    vd->vdev_ops == &vdev_indirect_ops);
790		space_map_close(vd->vdev_obsolete_sm);
791		vd->vdev_obsolete_sm = NULL;
792	}
793	range_tree_destroy(vd->vdev_obsolete_segments);
794	rw_destroy(&vd->vdev_indirect_rwlock);
795	mutex_destroy(&vd->vdev_obsolete_lock);
796
797	mutex_destroy(&vd->vdev_queue_lock);
798	mutex_destroy(&vd->vdev_dtl_lock);
799	mutex_destroy(&vd->vdev_stat_lock);
800	mutex_destroy(&vd->vdev_probe_lock);
801
802	if (vd == spa->spa_root_vdev)
803		spa->spa_root_vdev = NULL;
804
805	kmem_free(vd, sizeof (vdev_t));
806}
807
808/*
809 * Transfer top-level vdev state from svd to tvd.
810 */
811static void
812vdev_top_transfer(vdev_t *svd, vdev_t *tvd)
813{
814	spa_t *spa = svd->vdev_spa;
815	metaslab_t *msp;
816	vdev_t *vd;
817	int t;
818
819	ASSERT(tvd == tvd->vdev_top);
820
821	tvd->vdev_ms_array = svd->vdev_ms_array;
822	tvd->vdev_ms_shift = svd->vdev_ms_shift;
823	tvd->vdev_ms_count = svd->vdev_ms_count;
824	tvd->vdev_top_zap = svd->vdev_top_zap;
825
826	svd->vdev_ms_array = 0;
827	svd->vdev_ms_shift = 0;
828	svd->vdev_ms_count = 0;
829	svd->vdev_top_zap = 0;
830
831	if (tvd->vdev_mg)
832		ASSERT3P(tvd->vdev_mg, ==, svd->vdev_mg);
833	tvd->vdev_mg = svd->vdev_mg;
834	tvd->vdev_ms = svd->vdev_ms;
835
836	svd->vdev_mg = NULL;
837	svd->vdev_ms = NULL;
838
839	if (tvd->vdev_mg != NULL)
840		tvd->vdev_mg->mg_vd = tvd;
841
842	tvd->vdev_stat.vs_alloc = svd->vdev_stat.vs_alloc;
843	tvd->vdev_stat.vs_space = svd->vdev_stat.vs_space;
844	tvd->vdev_stat.vs_dspace = svd->vdev_stat.vs_dspace;
845
846	svd->vdev_stat.vs_alloc = 0;
847	svd->vdev_stat.vs_space = 0;
848	svd->vdev_stat.vs_dspace = 0;
849
850	for (t = 0; t < TXG_SIZE; t++) {
851		while ((msp = txg_list_remove(&svd->vdev_ms_list, t)) != NULL)
852			(void) txg_list_add(&tvd->vdev_ms_list, msp, t);
853		while ((vd = txg_list_remove(&svd->vdev_dtl_list, t)) != NULL)
854			(void) txg_list_add(&tvd->vdev_dtl_list, vd, t);
855		if (txg_list_remove_this(&spa->spa_vdev_txg_list, svd, t))
856			(void) txg_list_add(&spa->spa_vdev_txg_list, tvd, t);
857	}
858
859	if (list_link_active(&svd->vdev_config_dirty_node)) {
860		vdev_config_clean(svd);
861		vdev_config_dirty(tvd);
862	}
863
864	if (list_link_active(&svd->vdev_state_dirty_node)) {
865		vdev_state_clean(svd);
866		vdev_state_dirty(tvd);
867	}
868
869	tvd->vdev_deflate_ratio = svd->vdev_deflate_ratio;
870	svd->vdev_deflate_ratio = 0;
871
872	tvd->vdev_islog = svd->vdev_islog;
873	svd->vdev_islog = 0;
874}
875
876static void
877vdev_top_update(vdev_t *tvd, vdev_t *vd)
878{
879	if (vd == NULL)
880		return;
881
882	vd->vdev_top = tvd;
883
884	for (int c = 0; c < vd->vdev_children; c++)
885		vdev_top_update(tvd, vd->vdev_child[c]);
886}
887
888/*
889 * Add a mirror/replacing vdev above an existing vdev.
890 */
891vdev_t *
892vdev_add_parent(vdev_t *cvd, vdev_ops_t *ops)
893{
894	spa_t *spa = cvd->vdev_spa;
895	vdev_t *pvd = cvd->vdev_parent;
896	vdev_t *mvd;
897
898	ASSERT(spa_config_held(spa, SCL_ALL, RW_WRITER) == SCL_ALL);
899
900	mvd = vdev_alloc_common(spa, cvd->vdev_id, 0, ops);
901
902	mvd->vdev_asize = cvd->vdev_asize;
903	mvd->vdev_min_asize = cvd->vdev_min_asize;
904	mvd->vdev_max_asize = cvd->vdev_max_asize;
905	mvd->vdev_psize = cvd->vdev_psize;
906	mvd->vdev_ashift = cvd->vdev_ashift;
907	mvd->vdev_state = cvd->vdev_state;
908	mvd->vdev_crtxg = cvd->vdev_crtxg;
909
910	vdev_remove_child(pvd, cvd);
911	vdev_add_child(pvd, mvd);
912	cvd->vdev_id = mvd->vdev_children;
913	vdev_add_child(mvd, cvd);
914	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
915
916	if (mvd == mvd->vdev_top)
917		vdev_top_transfer(cvd, mvd);
918
919	return (mvd);
920}
921
922/*
923 * Remove a 1-way mirror/replacing vdev from the tree.
924 */
925void
926vdev_remove_parent(vdev_t *cvd)
927{
928	vdev_t *mvd = cvd->vdev_parent;
929	vdev_t *pvd = mvd->vdev_parent;
930
931	ASSERT(spa_config_held(cvd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
932
933	ASSERT(mvd->vdev_children == 1);
934	ASSERT(mvd->vdev_ops == &vdev_mirror_ops ||
935	    mvd->vdev_ops == &vdev_replacing_ops ||
936	    mvd->vdev_ops == &vdev_spare_ops);
937	cvd->vdev_ashift = mvd->vdev_ashift;
938
939	vdev_remove_child(mvd, cvd);
940	vdev_remove_child(pvd, mvd);
941
942	/*
943	 * If cvd will replace mvd as a top-level vdev, preserve mvd's guid.
944	 * Otherwise, we could have detached an offline device, and when we
945	 * go to import the pool we'll think we have two top-level vdevs,
946	 * instead of a different version of the same top-level vdev.
947	 */
948	if (mvd->vdev_top == mvd) {
949		uint64_t guid_delta = mvd->vdev_guid - cvd->vdev_guid;
950		cvd->vdev_orig_guid = cvd->vdev_guid;
951		cvd->vdev_guid += guid_delta;
952		cvd->vdev_guid_sum += guid_delta;
953	}
954	cvd->vdev_id = mvd->vdev_id;
955	vdev_add_child(pvd, cvd);
956	vdev_top_update(cvd->vdev_top, cvd->vdev_top);
957
958	if (cvd == cvd->vdev_top)
959		vdev_top_transfer(mvd, cvd);
960
961	ASSERT(mvd->vdev_children == 0);
962	vdev_free(mvd);
963}
964
965int
966vdev_metaslab_init(vdev_t *vd, uint64_t txg)
967{
968	spa_t *spa = vd->vdev_spa;
969	objset_t *mos = spa->spa_meta_objset;
970	uint64_t m;
971	uint64_t oldc = vd->vdev_ms_count;
972	uint64_t newc = vd->vdev_asize >> vd->vdev_ms_shift;
973	metaslab_t **mspp;
974	int error;
975
976	ASSERT(txg == 0 || spa_config_held(spa, SCL_ALLOC, RW_WRITER));
977
978	/*
979	 * This vdev is not being allocated from yet or is a hole.
980	 */
981	if (vd->vdev_ms_shift == 0)
982		return (0);
983
984	ASSERT(!vd->vdev_ishole);
985
986	ASSERT(oldc <= newc);
987
988	mspp = kmem_zalloc(newc * sizeof (*mspp), KM_SLEEP);
989
990	if (oldc != 0) {
991		bcopy(vd->vdev_ms, mspp, oldc * sizeof (*mspp));
992		kmem_free(vd->vdev_ms, oldc * sizeof (*mspp));
993	}
994
995	vd->vdev_ms = mspp;
996	vd->vdev_ms_count = newc;
997
998	for (m = oldc; m < newc; m++) {
999		uint64_t object = 0;
1000
1001		/*
1002		 * vdev_ms_array may be 0 if we are creating the "fake"
1003		 * metaslabs for an indirect vdev for zdb's leak detection.
1004		 * See zdb_leak_init().
1005		 */
1006		if (txg == 0 && vd->vdev_ms_array != 0) {
1007			error = dmu_read(mos, vd->vdev_ms_array,
1008			    m * sizeof (uint64_t), sizeof (uint64_t), &object,
1009			    DMU_READ_PREFETCH);
1010			if (error != 0) {
1011				vdev_dbgmsg(vd, "unable to read the metaslab "
1012				    "array [error=%d]", error);
1013				return (error);
1014			}
1015		}
1016
1017		error = metaslab_init(vd->vdev_mg, m, object, txg,
1018		    &(vd->vdev_ms[m]));
1019		if (error != 0) {
1020			vdev_dbgmsg(vd, "metaslab_init failed [error=%d]",
1021			    error);
1022			return (error);
1023		}
1024	}
1025
1026	if (txg == 0)
1027		spa_config_enter(spa, SCL_ALLOC, FTAG, RW_WRITER);
1028
1029	/*
1030	 * If the vdev is being removed we don't activate
1031	 * the metaslabs since we want to ensure that no new
1032	 * allocations are performed on this device.
1033	 */
1034	if (oldc == 0 && !vd->vdev_removing)
1035		metaslab_group_activate(vd->vdev_mg);
1036
1037	if (txg == 0)
1038		spa_config_exit(spa, SCL_ALLOC, FTAG);
1039
1040	return (0);
1041}
1042
1043void
1044vdev_metaslab_fini(vdev_t *vd)
1045{
1046	if (vd->vdev_ms != NULL) {
1047		uint64_t count = vd->vdev_ms_count;
1048
1049		metaslab_group_passivate(vd->vdev_mg);
1050		for (uint64_t m = 0; m < count; m++) {
1051			metaslab_t *msp = vd->vdev_ms[m];
1052
1053			if (msp != NULL)
1054				metaslab_fini(msp);
1055		}
1056		kmem_free(vd->vdev_ms, count * sizeof (metaslab_t *));
1057		vd->vdev_ms = NULL;
1058
1059		vd->vdev_ms_count = 0;
1060	}
1061	ASSERT0(vd->vdev_ms_count);
1062}
1063
1064typedef struct vdev_probe_stats {
1065	boolean_t	vps_readable;
1066	boolean_t	vps_writeable;
1067	int		vps_flags;
1068} vdev_probe_stats_t;
1069
1070static void
1071vdev_probe_done(zio_t *zio)
1072{
1073	spa_t *spa = zio->io_spa;
1074	vdev_t *vd = zio->io_vd;
1075	vdev_probe_stats_t *vps = zio->io_private;
1076
1077	ASSERT(vd->vdev_probe_zio != NULL);
1078
1079	if (zio->io_type == ZIO_TYPE_READ) {
1080		if (zio->io_error == 0)
1081			vps->vps_readable = 1;
1082		if (zio->io_error == 0 && spa_writeable(spa)) {
1083			zio_nowait(zio_write_phys(vd->vdev_probe_zio, vd,
1084			    zio->io_offset, zio->io_size, zio->io_abd,
1085			    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1086			    ZIO_PRIORITY_SYNC_WRITE, vps->vps_flags, B_TRUE));
1087		} else {
1088			abd_free(zio->io_abd);
1089		}
1090	} else if (zio->io_type == ZIO_TYPE_WRITE) {
1091		if (zio->io_error == 0)
1092			vps->vps_writeable = 1;
1093		abd_free(zio->io_abd);
1094	} else if (zio->io_type == ZIO_TYPE_NULL) {
1095		zio_t *pio;
1096
1097		vd->vdev_cant_read |= !vps->vps_readable;
1098		vd->vdev_cant_write |= !vps->vps_writeable;
1099
1100		if (vdev_readable(vd) &&
1101		    (vdev_writeable(vd) || !spa_writeable(spa))) {
1102			zio->io_error = 0;
1103		} else {
1104			ASSERT(zio->io_error != 0);
1105			vdev_dbgmsg(vd, "failed probe");
1106			zfs_ereport_post(FM_EREPORT_ZFS_PROBE_FAILURE,
1107			    spa, vd, NULL, 0, 0);
1108			zio->io_error = SET_ERROR(ENXIO);
1109		}
1110
1111		mutex_enter(&vd->vdev_probe_lock);
1112		ASSERT(vd->vdev_probe_zio == zio);
1113		vd->vdev_probe_zio = NULL;
1114		mutex_exit(&vd->vdev_probe_lock);
1115
1116		zio_link_t *zl = NULL;
1117		while ((pio = zio_walk_parents(zio, &zl)) != NULL)
1118			if (!vdev_accessible(vd, pio))
1119				pio->io_error = SET_ERROR(ENXIO);
1120
1121		kmem_free(vps, sizeof (*vps));
1122	}
1123}
1124
1125/*
1126 * Determine whether this device is accessible.
1127 *
1128 * Read and write to several known locations: the pad regions of each
1129 * vdev label but the first, which we leave alone in case it contains
1130 * a VTOC.
1131 */
1132zio_t *
1133vdev_probe(vdev_t *vd, zio_t *zio)
1134{
1135	spa_t *spa = vd->vdev_spa;
1136	vdev_probe_stats_t *vps = NULL;
1137	zio_t *pio;
1138
1139	ASSERT(vd->vdev_ops->vdev_op_leaf);
1140
1141	/*
1142	 * Don't probe the probe.
1143	 */
1144	if (zio && (zio->io_flags & ZIO_FLAG_PROBE))
1145		return (NULL);
1146
1147	/*
1148	 * To prevent 'probe storms' when a device fails, we create
1149	 * just one probe i/o at a time.  All zios that want to probe
1150	 * this vdev will become parents of the probe io.
1151	 */
1152	mutex_enter(&vd->vdev_probe_lock);
1153
1154	if ((pio = vd->vdev_probe_zio) == NULL) {
1155		vps = kmem_zalloc(sizeof (*vps), KM_SLEEP);
1156
1157		vps->vps_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_PROBE |
1158		    ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE |
1159		    ZIO_FLAG_TRYHARD;
1160
1161		if (spa_config_held(spa, SCL_ZIO, RW_WRITER)) {
1162			/*
1163			 * vdev_cant_read and vdev_cant_write can only
1164			 * transition from TRUE to FALSE when we have the
1165			 * SCL_ZIO lock as writer; otherwise they can only
1166			 * transition from FALSE to TRUE.  This ensures that
1167			 * any zio looking at these values can assume that
1168			 * failures persist for the life of the I/O.  That's
1169			 * important because when a device has intermittent
1170			 * connectivity problems, we want to ensure that
1171			 * they're ascribed to the device (ENXIO) and not
1172			 * the zio (EIO).
1173			 *
1174			 * Since we hold SCL_ZIO as writer here, clear both
1175			 * values so the probe can reevaluate from first
1176			 * principles.
1177			 */
1178			vps->vps_flags |= ZIO_FLAG_CONFIG_WRITER;
1179			vd->vdev_cant_read = B_FALSE;
1180			vd->vdev_cant_write = B_FALSE;
1181		}
1182
1183		vd->vdev_probe_zio = pio = zio_null(NULL, spa, vd,
1184		    vdev_probe_done, vps,
1185		    vps->vps_flags | ZIO_FLAG_DONT_PROPAGATE);
1186
1187		/*
1188		 * We can't change the vdev state in this context, so we
1189		 * kick off an async task to do it on our behalf.
1190		 */
1191		if (zio != NULL) {
1192			vd->vdev_probe_wanted = B_TRUE;
1193			spa_async_request(spa, SPA_ASYNC_PROBE);
1194		}
1195	}
1196
1197	if (zio != NULL)
1198		zio_add_child(zio, pio);
1199
1200	mutex_exit(&vd->vdev_probe_lock);
1201
1202	if (vps == NULL) {
1203		ASSERT(zio != NULL);
1204		return (NULL);
1205	}
1206
1207	for (int l = 1; l < VDEV_LABELS; l++) {
1208		zio_nowait(zio_read_phys(pio, vd,
1209		    vdev_label_offset(vd->vdev_psize, l,
1210		    offsetof(vdev_label_t, vl_pad2)), VDEV_PAD_SIZE,
1211		    abd_alloc_for_io(VDEV_PAD_SIZE, B_TRUE),
1212		    ZIO_CHECKSUM_OFF, vdev_probe_done, vps,
1213		    ZIO_PRIORITY_SYNC_READ, vps->vps_flags, B_TRUE));
1214	}
1215
1216	if (zio == NULL)
1217		return (pio);
1218
1219	zio_nowait(pio);
1220	return (NULL);
1221}
1222
1223static void
1224vdev_open_child(void *arg)
1225{
1226	vdev_t *vd = arg;
1227
1228	vd->vdev_open_thread = curthread;
1229	vd->vdev_open_error = vdev_open(vd);
1230	vd->vdev_open_thread = NULL;
1231}
1232
1233boolean_t
1234vdev_uses_zvols(vdev_t *vd)
1235{
1236	if (vd->vdev_path && strncmp(vd->vdev_path, ZVOL_DIR,
1237	    strlen(ZVOL_DIR)) == 0)
1238		return (B_TRUE);
1239	for (int c = 0; c < vd->vdev_children; c++)
1240		if (vdev_uses_zvols(vd->vdev_child[c]))
1241			return (B_TRUE);
1242	return (B_FALSE);
1243}
1244
1245void
1246vdev_open_children(vdev_t *vd)
1247{
1248	taskq_t *tq;
1249	int children = vd->vdev_children;
1250
1251	/*
1252	 * in order to handle pools on top of zvols, do the opens
1253	 * in a single thread so that the same thread holds the
1254	 * spa_namespace_lock
1255	 */
1256	if (vdev_uses_zvols(vd)) {
1257		for (int c = 0; c < children; c++)
1258			vd->vdev_child[c]->vdev_open_error =
1259			    vdev_open(vd->vdev_child[c]);
1260		return;
1261	}
1262	tq = taskq_create("vdev_open", children, minclsyspri,
1263	    children, children, TASKQ_PREPOPULATE);
1264
1265	for (int c = 0; c < children; c++)
1266		VERIFY(taskq_dispatch(tq, vdev_open_child, vd->vdev_child[c],
1267		    TQ_SLEEP) != NULL);
1268
1269	taskq_destroy(tq);
1270}
1271
1272/*
1273 * Compute the raidz-deflation ratio.  Note, we hard-code
1274 * in 128k (1 << 17) because it is the "typical" blocksize.
1275 * Even though SPA_MAXBLOCKSIZE changed, this algorithm can not change,
1276 * otherwise it would inconsistently account for existing bp's.
1277 */
1278static void
1279vdev_set_deflate_ratio(vdev_t *vd)
1280{
1281	if (vd == vd->vdev_top && !vd->vdev_ishole && vd->vdev_ashift != 0) {
1282		vd->vdev_deflate_ratio = (1 << 17) /
1283		    (vdev_psize_to_asize(vd, 1 << 17) >> SPA_MINBLOCKSHIFT);
1284	}
1285}
1286
1287/*
1288 * Prepare a virtual device for access.
1289 */
1290int
1291vdev_open(vdev_t *vd)
1292{
1293	spa_t *spa = vd->vdev_spa;
1294	int error;
1295	uint64_t osize = 0;
1296	uint64_t max_osize = 0;
1297	uint64_t asize, max_asize, psize;
1298	uint64_t ashift = 0;
1299
1300	ASSERT(vd->vdev_open_thread == curthread ||
1301	    spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1302	ASSERT(vd->vdev_state == VDEV_STATE_CLOSED ||
1303	    vd->vdev_state == VDEV_STATE_CANT_OPEN ||
1304	    vd->vdev_state == VDEV_STATE_OFFLINE);
1305
1306	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1307	vd->vdev_cant_read = B_FALSE;
1308	vd->vdev_cant_write = B_FALSE;
1309	vd->vdev_min_asize = vdev_get_min_asize(vd);
1310
1311	/*
1312	 * If this vdev is not removed, check its fault status.  If it's
1313	 * faulted, bail out of the open.
1314	 */
1315	if (!vd->vdev_removed && vd->vdev_faulted) {
1316		ASSERT(vd->vdev_children == 0);
1317		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1318		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1319		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1320		    vd->vdev_label_aux);
1321		return (SET_ERROR(ENXIO));
1322	} else if (vd->vdev_offline) {
1323		ASSERT(vd->vdev_children == 0);
1324		vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE, VDEV_AUX_NONE);
1325		return (SET_ERROR(ENXIO));
1326	}
1327
1328	error = vd->vdev_ops->vdev_op_open(vd, &osize, &max_osize, &ashift);
1329
1330	/*
1331	 * Reset the vdev_reopening flag so that we actually close
1332	 * the vdev on error.
1333	 */
1334	vd->vdev_reopening = B_FALSE;
1335	if (zio_injection_enabled && error == 0)
1336		error = zio_handle_device_injection(vd, NULL, ENXIO);
1337
1338	if (error) {
1339		if (vd->vdev_removed &&
1340		    vd->vdev_stat.vs_aux != VDEV_AUX_OPEN_FAILED)
1341			vd->vdev_removed = B_FALSE;
1342
1343		if (vd->vdev_stat.vs_aux == VDEV_AUX_CHILDREN_OFFLINE) {
1344			vdev_set_state(vd, B_TRUE, VDEV_STATE_OFFLINE,
1345			    vd->vdev_stat.vs_aux);
1346		} else {
1347			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1348			    vd->vdev_stat.vs_aux);
1349		}
1350		return (error);
1351	}
1352
1353	vd->vdev_removed = B_FALSE;
1354
1355	/*
1356	 * Recheck the faulted flag now that we have confirmed that
1357	 * the vdev is accessible.  If we're faulted, bail.
1358	 */
1359	if (vd->vdev_faulted) {
1360		ASSERT(vd->vdev_children == 0);
1361		ASSERT(vd->vdev_label_aux == VDEV_AUX_ERR_EXCEEDED ||
1362		    vd->vdev_label_aux == VDEV_AUX_EXTERNAL);
1363		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1364		    vd->vdev_label_aux);
1365		return (SET_ERROR(ENXIO));
1366	}
1367
1368	if (vd->vdev_degraded) {
1369		ASSERT(vd->vdev_children == 0);
1370		vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1371		    VDEV_AUX_ERR_EXCEEDED);
1372	} else {
1373		vdev_set_state(vd, B_TRUE, VDEV_STATE_HEALTHY, 0);
1374	}
1375
1376	/*
1377	 * For hole or missing vdevs we just return success.
1378	 */
1379	if (vd->vdev_ishole || vd->vdev_ops == &vdev_missing_ops)
1380		return (0);
1381
1382	for (int c = 0; c < vd->vdev_children; c++) {
1383		if (vd->vdev_child[c]->vdev_state != VDEV_STATE_HEALTHY) {
1384			vdev_set_state(vd, B_TRUE, VDEV_STATE_DEGRADED,
1385			    VDEV_AUX_NONE);
1386			break;
1387		}
1388	}
1389
1390	osize = P2ALIGN(osize, (uint64_t)sizeof (vdev_label_t));
1391	max_osize = P2ALIGN(max_osize, (uint64_t)sizeof (vdev_label_t));
1392
1393	if (vd->vdev_children == 0) {
1394		if (osize < SPA_MINDEVSIZE) {
1395			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1396			    VDEV_AUX_TOO_SMALL);
1397			return (SET_ERROR(EOVERFLOW));
1398		}
1399		psize = osize;
1400		asize = osize - (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE);
1401		max_asize = max_osize - (VDEV_LABEL_START_SIZE +
1402		    VDEV_LABEL_END_SIZE);
1403	} else {
1404		if (vd->vdev_parent != NULL && osize < SPA_MINDEVSIZE -
1405		    (VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE)) {
1406			vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1407			    VDEV_AUX_TOO_SMALL);
1408			return (SET_ERROR(EOVERFLOW));
1409		}
1410		psize = 0;
1411		asize = osize;
1412		max_asize = max_osize;
1413	}
1414
1415	vd->vdev_psize = psize;
1416
1417	/*
1418	 * Make sure the allocatable size hasn't shrunk too much.
1419	 */
1420	if (asize < vd->vdev_min_asize) {
1421		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1422		    VDEV_AUX_BAD_LABEL);
1423		return (SET_ERROR(EINVAL));
1424	}
1425
1426	if (vd->vdev_asize == 0) {
1427		/*
1428		 * This is the first-ever open, so use the computed values.
1429		 * For testing purposes, a higher ashift can be requested.
1430		 */
1431		vd->vdev_asize = asize;
1432		vd->vdev_max_asize = max_asize;
1433		vd->vdev_ashift = MAX(ashift, vd->vdev_ashift);
1434	} else {
1435		/*
1436		 * Detect if the alignment requirement has increased.
1437		 * We don't want to make the pool unavailable, just
1438		 * issue a warning instead.
1439		 */
1440		if (ashift > vd->vdev_top->vdev_ashift &&
1441		    vd->vdev_ops->vdev_op_leaf) {
1442			cmn_err(CE_WARN,
1443			    "Disk, '%s', has a block alignment that is "
1444			    "larger than the pool's alignment\n",
1445			    vd->vdev_path);
1446		}
1447		vd->vdev_max_asize = max_asize;
1448	}
1449
1450	/*
1451	 * If all children are healthy we update asize if either:
1452	 * The asize has increased, due to a device expansion caused by dynamic
1453	 * LUN growth or vdev replacement, and automatic expansion is enabled;
1454	 * making the additional space available.
1455	 *
1456	 * The asize has decreased, due to a device shrink usually caused by a
1457	 * vdev replace with a smaller device. This ensures that calculations
1458	 * based of max_asize and asize e.g. esize are always valid. It's safe
1459	 * to do this as we've already validated that asize is greater than
1460	 * vdev_min_asize.
1461	 */
1462	if (vd->vdev_state == VDEV_STATE_HEALTHY &&
1463	    ((asize > vd->vdev_asize &&
1464	    (vd->vdev_expanding || spa->spa_autoexpand)) ||
1465	    (asize < vd->vdev_asize)))
1466		vd->vdev_asize = asize;
1467
1468	vdev_set_min_asize(vd);
1469
1470	/*
1471	 * Ensure we can issue some IO before declaring the
1472	 * vdev open for business.
1473	 */
1474	if (vd->vdev_ops->vdev_op_leaf &&
1475	    (error = zio_wait(vdev_probe(vd, NULL))) != 0) {
1476		vdev_set_state(vd, B_TRUE, VDEV_STATE_FAULTED,
1477		    VDEV_AUX_ERR_EXCEEDED);
1478		return (error);
1479	}
1480
1481	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1482	    !vd->vdev_isl2cache && !vd->vdev_islog) {
1483		if (vd->vdev_ashift > spa->spa_max_ashift)
1484			spa->spa_max_ashift = vd->vdev_ashift;
1485		if (vd->vdev_ashift < spa->spa_min_ashift)
1486			spa->spa_min_ashift = vd->vdev_ashift;
1487	}
1488
1489	/*
1490	 * Track the min and max ashift values for normal data devices.
1491	 */
1492	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1493	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1494		if (vd->vdev_ashift > spa->spa_max_ashift)
1495			spa->spa_max_ashift = vd->vdev_ashift;
1496		if (vd->vdev_ashift < spa->spa_min_ashift)
1497			spa->spa_min_ashift = vd->vdev_ashift;
1498	}
1499
1500	/*
1501	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1502	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1503	 * since this would just restart the scrub we are already doing.
1504	 */
1505	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1506	    vdev_resilver_needed(vd, NULL, NULL))
1507		spa_async_request(spa, SPA_ASYNC_RESILVER);
1508
1509	return (0);
1510}
1511
1512/*
1513 * Called once the vdevs are all opened, this routine validates the label
1514 * contents. This needs to be done before vdev_load() so that we don't
1515 * inadvertently do repair I/Os to the wrong device.
1516 *
1517 * This function will only return failure if one of the vdevs indicates that it
1518 * has since been destroyed or exported.  This is only possible if
1519 * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1520 * will be updated but the function will return 0.
1521 */
1522int
1523vdev_validate(vdev_t *vd)
1524{
1525	spa_t *spa = vd->vdev_spa;
1526	nvlist_t *label;
1527	uint64_t guid = 0, aux_guid = 0, top_guid;
1528	uint64_t state;
1529	nvlist_t *nvl;
1530	uint64_t txg;
1531
1532	if (vdev_validate_skip)
1533		return (0);
1534
1535	for (uint64_t c = 0; c < vd->vdev_children; c++)
1536		if (vdev_validate(vd->vdev_child[c]) != 0)
1537			return (SET_ERROR(EBADF));
1538
1539	/*
1540	 * If the device has already failed, or was marked offline, don't do
1541	 * any further validation.  Otherwise, label I/O will fail and we will
1542	 * overwrite the previous state.
1543	 */
1544	if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd))
1545		return (0);
1546
1547	/*
1548	 * If we are performing an extreme rewind, we allow for a label that
1549	 * was modified at a point after the current txg.
1550	 */
1551	if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0)
1552		txg = UINT64_MAX;
1553	else
1554		txg = spa_last_synced_txg(spa);
1555
1556	if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1557		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1558		    VDEV_AUX_BAD_LABEL);
1559		vdev_dbgmsg(vd, "vdev_validate: failed reading config");
1560		return (0);
1561	}
1562
1563	/*
1564	 * Determine if this vdev has been split off into another
1565	 * pool.  If so, then refuse to open it.
1566	 */
1567	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1568	    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1569		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1570		    VDEV_AUX_SPLIT_POOL);
1571		nvlist_free(label);
1572		vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool");
1573		return (0);
1574	}
1575
1576	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) {
1577		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1578		    VDEV_AUX_CORRUPT_DATA);
1579		nvlist_free(label);
1580		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1581		    ZPOOL_CONFIG_POOL_GUID);
1582		return (0);
1583	}
1584
1585	/*
1586	 * If config is not trusted then ignore the spa guid check. This is
1587	 * necessary because if the machine crashed during a re-guid the new
1588	 * guid might have been written to all of the vdev labels, but not the
1589	 * cached config. The check will be performed again once we have the
1590	 * trusted config from the MOS.
1591	 */
1592	if (spa->spa_trust_config && guid != spa_guid(spa)) {
1593		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1594		    VDEV_AUX_CORRUPT_DATA);
1595		nvlist_free(label);
1596		vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't "
1597		    "match config (%llu != %llu)", (u_longlong_t)guid,
1598		    (u_longlong_t)spa_guid(spa));
1599		return (0);
1600	}
1601
1602	if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1603	    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1604	    &aux_guid) != 0)
1605		aux_guid = 0;
1606
1607	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) {
1608		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1609		    VDEV_AUX_CORRUPT_DATA);
1610		nvlist_free(label);
1611		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1612		    ZPOOL_CONFIG_GUID);
1613		return (0);
1614	}
1615
1616	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid)
1617	    != 0) {
1618		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1619		    VDEV_AUX_CORRUPT_DATA);
1620		nvlist_free(label);
1621		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1622		    ZPOOL_CONFIG_TOP_GUID);
1623		return (0);
1624	}
1625
1626	/*
1627	 * If this vdev just became a top-level vdev because its sibling was
1628	 * detached, it will have adopted the parent's vdev guid -- but the
1629	 * label may or may not be on disk yet. Fortunately, either version
1630	 * of the label will have the same top guid, so if we're a top-level
1631	 * vdev, we can safely compare to that instead.
1632	 * However, if the config comes from a cachefile that failed to update
1633	 * after the detach, a top-level vdev will appear as a non top-level
1634	 * vdev in the config. Also relax the constraints if we perform an
1635	 * extreme rewind.
1636	 *
1637	 * If we split this vdev off instead, then we also check the
1638	 * original pool's guid. We don't want to consider the vdev
1639	 * corrupt if it is partway through a split operation.
1640	 */
1641	if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) {
1642		boolean_t mismatch = B_FALSE;
1643		if (spa->spa_trust_config && !spa->spa_extreme_rewind) {
1644			if (vd != vd->vdev_top || vd->vdev_guid != top_guid)
1645				mismatch = B_TRUE;
1646		} else {
1647			if (vd->vdev_guid != top_guid &&
1648			    vd->vdev_top->vdev_guid != guid)
1649				mismatch = B_TRUE;
1650		}
1651
1652		if (mismatch) {
1653			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1654			    VDEV_AUX_CORRUPT_DATA);
1655			nvlist_free(label);
1656			vdev_dbgmsg(vd, "vdev_validate: config guid "
1657			    "doesn't match label guid");
1658			vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu",
1659			    (u_longlong_t)vd->vdev_guid,
1660			    (u_longlong_t)vd->vdev_top->vdev_guid);
1661			vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, "
1662			    "aux_guid %llu", (u_longlong_t)guid,
1663			    (u_longlong_t)top_guid, (u_longlong_t)aux_guid);
1664			return (0);
1665		}
1666	}
1667
1668	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1669	    &state) != 0) {
1670		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1671		    VDEV_AUX_CORRUPT_DATA);
1672		nvlist_free(label);
1673		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1674		    ZPOOL_CONFIG_POOL_STATE);
1675		return (0);
1676	}
1677
1678	nvlist_free(label);
1679
1680	/*
1681	 * If this is a verbatim import, no need to check the
1682	 * state of the pool.
1683	 */
1684	if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1685	    spa_load_state(spa) == SPA_LOAD_OPEN &&
1686	    state != POOL_STATE_ACTIVE) {
1687		vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) "
1688		    "for spa %s", (u_longlong_t)state, spa->spa_name);
1689		return (SET_ERROR(EBADF));
1690	}
1691
1692	/*
1693	 * If we were able to open and validate a vdev that was
1694	 * previously marked permanently unavailable, clear that state
1695	 * now.
1696	 */
1697	if (vd->vdev_not_present)
1698		vd->vdev_not_present = 0;
1699
1700	return (0);
1701}
1702
1703static void
1704vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd)
1705{
1706	if (svd->vdev_path != NULL && dvd->vdev_path != NULL) {
1707		if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) {
1708			zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed "
1709			    "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
1710			    dvd->vdev_path, svd->vdev_path);
1711			spa_strfree(dvd->vdev_path);
1712			dvd->vdev_path = spa_strdup(svd->vdev_path);
1713		}
1714	} else if (svd->vdev_path != NULL) {
1715		dvd->vdev_path = spa_strdup(svd->vdev_path);
1716		zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'",
1717		    (u_longlong_t)dvd->vdev_guid, dvd->vdev_path);
1718	}
1719}
1720
1721/*
1722 * Recursively copy vdev paths from one vdev to another. Source and destination
1723 * vdev trees must have same geometry otherwise return error. Intended to copy
1724 * paths from userland config into MOS config.
1725 */
1726int
1727vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd)
1728{
1729	if ((svd->vdev_ops == &vdev_missing_ops) ||
1730	    (svd->vdev_ishole && dvd->vdev_ishole) ||
1731	    (dvd->vdev_ops == &vdev_indirect_ops))
1732		return (0);
1733
1734	if (svd->vdev_ops != dvd->vdev_ops) {
1735		vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s",
1736		    svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type);
1737		return (SET_ERROR(EINVAL));
1738	}
1739
1740	if (svd->vdev_guid != dvd->vdev_guid) {
1741		vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != "
1742		    "%llu)", (u_longlong_t)svd->vdev_guid,
1743		    (u_longlong_t)dvd->vdev_guid);
1744		return (SET_ERROR(EINVAL));
1745	}
1746
1747	if (svd->vdev_children != dvd->vdev_children) {
1748		vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: "
1749		    "%llu != %llu", (u_longlong_t)svd->vdev_children,
1750		    (u_longlong_t)dvd->vdev_children);
1751		return (SET_ERROR(EINVAL));
1752	}
1753
1754	for (uint64_t i = 0; i < svd->vdev_children; i++) {
1755		int error = vdev_copy_path_strict(svd->vdev_child[i],
1756		    dvd->vdev_child[i]);
1757		if (error != 0)
1758			return (error);
1759	}
1760
1761	if (svd->vdev_ops->vdev_op_leaf)
1762		vdev_copy_path_impl(svd, dvd);
1763
1764	return (0);
1765}
1766
1767static void
1768vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd)
1769{
1770	ASSERT(stvd->vdev_top == stvd);
1771	ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id);
1772
1773	for (uint64_t i = 0; i < dvd->vdev_children; i++) {
1774		vdev_copy_path_search(stvd, dvd->vdev_child[i]);
1775	}
1776
1777	if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd))
1778		return;
1779
1780	/*
1781	 * The idea here is that while a vdev can shift positions within
1782	 * a top vdev (when replacing, attaching mirror, etc.) it cannot
1783	 * step outside of it.
1784	 */
1785	vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid);
1786
1787	if (vd == NULL || vd->vdev_ops != dvd->vdev_ops)
1788		return;
1789
1790	ASSERT(vd->vdev_ops->vdev_op_leaf);
1791
1792	vdev_copy_path_impl(vd, dvd);
1793}
1794
1795/*
1796 * Recursively copy vdev paths from one root vdev to another. Source and
1797 * destination vdev trees may differ in geometry. For each destination leaf
1798 * vdev, search a vdev with the same guid and top vdev id in the source.
1799 * Intended to copy paths from userland config into MOS config.
1800 */
1801void
1802vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd)
1803{
1804	uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children);
1805	ASSERT(srvd->vdev_ops == &vdev_root_ops);
1806	ASSERT(drvd->vdev_ops == &vdev_root_ops);
1807
1808	for (uint64_t i = 0; i < children; i++) {
1809		vdev_copy_path_search(srvd->vdev_child[i],
1810		    drvd->vdev_child[i]);
1811	}
1812}
1813
1814/*
1815 * Close a virtual device.
1816 */
1817void
1818vdev_close(vdev_t *vd)
1819{
1820	spa_t *spa = vd->vdev_spa;
1821	vdev_t *pvd = vd->vdev_parent;
1822
1823	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1824
1825	/*
1826	 * If our parent is reopening, then we are as well, unless we are
1827	 * going offline.
1828	 */
1829	if (pvd != NULL && pvd->vdev_reopening)
1830		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1831
1832	vd->vdev_ops->vdev_op_close(vd);
1833
1834	vdev_cache_purge(vd);
1835
1836	/*
1837	 * We record the previous state before we close it, so that if we are
1838	 * doing a reopen(), we don't generate FMA ereports if we notice that
1839	 * it's still faulted.
1840	 */
1841	vd->vdev_prevstate = vd->vdev_state;
1842
1843	if (vd->vdev_offline)
1844		vd->vdev_state = VDEV_STATE_OFFLINE;
1845	else
1846		vd->vdev_state = VDEV_STATE_CLOSED;
1847	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1848}
1849
1850void
1851vdev_hold(vdev_t *vd)
1852{
1853	spa_t *spa = vd->vdev_spa;
1854
1855	ASSERT(spa_is_root(spa));
1856	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1857		return;
1858
1859	for (int c = 0; c < vd->vdev_children; c++)
1860		vdev_hold(vd->vdev_child[c]);
1861
1862	if (vd->vdev_ops->vdev_op_leaf)
1863		vd->vdev_ops->vdev_op_hold(vd);
1864}
1865
1866void
1867vdev_rele(vdev_t *vd)
1868{
1869	spa_t *spa = vd->vdev_spa;
1870
1871	ASSERT(spa_is_root(spa));
1872	for (int c = 0; c < vd->vdev_children; c++)
1873		vdev_rele(vd->vdev_child[c]);
1874
1875	if (vd->vdev_ops->vdev_op_leaf)
1876		vd->vdev_ops->vdev_op_rele(vd);
1877}
1878
1879/*
1880 * Reopen all interior vdevs and any unopened leaves.  We don't actually
1881 * reopen leaf vdevs which had previously been opened as they might deadlock
1882 * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1883 * If the leaf has never been opened then open it, as usual.
1884 */
1885void
1886vdev_reopen(vdev_t *vd)
1887{
1888	spa_t *spa = vd->vdev_spa;
1889
1890	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1891
1892	/* set the reopening flag unless we're taking the vdev offline */
1893	vd->vdev_reopening = !vd->vdev_offline;
1894	vdev_close(vd);
1895	(void) vdev_open(vd);
1896
1897	/*
1898	 * Call vdev_validate() here to make sure we have the same device.
1899	 * Otherwise, a device with an invalid label could be successfully
1900	 * opened in response to vdev_reopen().
1901	 */
1902	if (vd->vdev_aux) {
1903		(void) vdev_validate_aux(vd);
1904		if (vdev_readable(vd) && vdev_writeable(vd) &&
1905		    vd->vdev_aux == &spa->spa_l2cache &&
1906		    !l2arc_vdev_present(vd))
1907			l2arc_add_vdev(spa, vd);
1908	} else {
1909		(void) vdev_validate(vd);
1910	}
1911
1912	/*
1913	 * Reassess parent vdev's health.
1914	 */
1915	vdev_propagate_state(vd);
1916}
1917
1918int
1919vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1920{
1921	int error;
1922
1923	/*
1924	 * Normally, partial opens (e.g. of a mirror) are allowed.
1925	 * For a create, however, we want to fail the request if
1926	 * there are any components we can't open.
1927	 */
1928	error = vdev_open(vd);
1929
1930	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1931		vdev_close(vd);
1932		return (error ? error : ENXIO);
1933	}
1934
1935	/*
1936	 * Recursively load DTLs and initialize all labels.
1937	 */
1938	if ((error = vdev_dtl_load(vd)) != 0 ||
1939	    (error = vdev_label_init(vd, txg, isreplacing ?
1940	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1941		vdev_close(vd);
1942		return (error);
1943	}
1944
1945	return (0);
1946}
1947
1948void
1949vdev_metaslab_set_size(vdev_t *vd)
1950{
1951	/*
1952	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1953	 */
1954	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1955	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1956}
1957
1958void
1959vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1960{
1961	ASSERT(vd == vd->vdev_top);
1962	/* indirect vdevs don't have metaslabs or dtls */
1963	ASSERT(vdev_is_concrete(vd) || flags == 0);
1964	ASSERT(ISP2(flags));
1965	ASSERT(spa_writeable(vd->vdev_spa));
1966
1967	if (flags & VDD_METASLAB)
1968		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1969
1970	if (flags & VDD_DTL)
1971		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1972
1973	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1974}
1975
1976void
1977vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1978{
1979	for (int c = 0; c < vd->vdev_children; c++)
1980		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1981
1982	if (vd->vdev_ops->vdev_op_leaf)
1983		vdev_dirty(vd->vdev_top, flags, vd, txg);
1984}
1985
1986/*
1987 * DTLs.
1988 *
1989 * A vdev's DTL (dirty time log) is the set of transaction groups for which
1990 * the vdev has less than perfect replication.  There are four kinds of DTL:
1991 *
1992 * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1993 *
1994 * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1995 *
1996 * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1997 *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1998 *	txgs that was scrubbed.
1999 *
2000 * DTL_OUTAGE: txgs which cannot currently be read, whether due to
2001 *	persistent errors or just some device being offline.
2002 *	Unlike the other three, the DTL_OUTAGE map is not generally
2003 *	maintained; it's only computed when needed, typically to
2004 *	determine whether a device can be detached.
2005 *
2006 * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
2007 * either has the data or it doesn't.
2008 *
2009 * For interior vdevs such as mirror and RAID-Z the picture is more complex.
2010 * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
2011 * if any child is less than fully replicated, then so is its parent.
2012 * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
2013 * comprising only those txgs which appear in 'maxfaults' or more children;
2014 * those are the txgs we don't have enough replication to read.  For example,
2015 * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
2016 * thus, its DTL_MISSING consists of the set of txgs that appear in more than
2017 * two child DTL_MISSING maps.
2018 *
2019 * It should be clear from the above that to compute the DTLs and outage maps
2020 * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
2021 * Therefore, that is all we keep on disk.  When loading the pool, or after
2022 * a configuration change, we generate all other DTLs from first principles.
2023 */
2024void
2025vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2026{
2027	range_tree_t *rt = vd->vdev_dtl[t];
2028
2029	ASSERT(t < DTL_TYPES);
2030	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2031	ASSERT(spa_writeable(vd->vdev_spa));
2032
2033	mutex_enter(&vd->vdev_dtl_lock);
2034	if (!range_tree_contains(rt, txg, size))
2035		range_tree_add(rt, txg, size);
2036	mutex_exit(&vd->vdev_dtl_lock);
2037}
2038
2039boolean_t
2040vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2041{
2042	range_tree_t *rt = vd->vdev_dtl[t];
2043	boolean_t dirty = B_FALSE;
2044
2045	ASSERT(t < DTL_TYPES);
2046	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2047
2048	/*
2049	 * While we are loading the pool, the DTLs have not been loaded yet.
2050	 * Ignore the DTLs and try all devices.  This avoids a recursive
2051	 * mutex enter on the vdev_dtl_lock, and also makes us try hard
2052	 * when loading the pool (relying on the checksum to ensure that
2053	 * we get the right data -- note that we while loading, we are
2054	 * only reading the MOS, which is always checksummed).
2055	 */
2056	if (vd->vdev_spa->spa_load_state != SPA_LOAD_NONE)
2057		return (B_FALSE);
2058
2059	mutex_enter(&vd->vdev_dtl_lock);
2060	if (range_tree_space(rt) != 0)
2061		dirty = range_tree_contains(rt, txg, size);
2062	mutex_exit(&vd->vdev_dtl_lock);
2063
2064	return (dirty);
2065}
2066
2067boolean_t
2068vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
2069{
2070	range_tree_t *rt = vd->vdev_dtl[t];
2071	boolean_t empty;
2072
2073	mutex_enter(&vd->vdev_dtl_lock);
2074	empty = (range_tree_space(rt) == 0);
2075	mutex_exit(&vd->vdev_dtl_lock);
2076
2077	return (empty);
2078}
2079
2080/*
2081 * Returns the lowest txg in the DTL range.
2082 */
2083static uint64_t
2084vdev_dtl_min(vdev_t *vd)
2085{
2086	range_seg_t *rs;
2087
2088	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2089	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2090	ASSERT0(vd->vdev_children);
2091
2092	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2093	return (rs->rs_start - 1);
2094}
2095
2096/*
2097 * Returns the highest txg in the DTL.
2098 */
2099static uint64_t
2100vdev_dtl_max(vdev_t *vd)
2101{
2102	range_seg_t *rs;
2103
2104	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2105	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2106	ASSERT0(vd->vdev_children);
2107
2108	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2109	return (rs->rs_end);
2110}
2111
2112/*
2113 * Determine if a resilvering vdev should remove any DTL entries from
2114 * its range. If the vdev was resilvering for the entire duration of the
2115 * scan then it should excise that range from its DTLs. Otherwise, this
2116 * vdev is considered partially resilvered and should leave its DTL
2117 * entries intact. The comment in vdev_dtl_reassess() describes how we
2118 * excise the DTLs.
2119 */
2120static boolean_t
2121vdev_dtl_should_excise(vdev_t *vd)
2122{
2123	spa_t *spa = vd->vdev_spa;
2124	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2125
2126	ASSERT0(scn->scn_phys.scn_errors);
2127	ASSERT0(vd->vdev_children);
2128
2129	if (vd->vdev_state < VDEV_STATE_DEGRADED)
2130		return (B_FALSE);
2131
2132	if (vd->vdev_resilver_txg == 0 ||
2133	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
2134		return (B_TRUE);
2135
2136	/*
2137	 * When a resilver is initiated the scan will assign the scn_max_txg
2138	 * value to the highest txg value that exists in all DTLs. If this
2139	 * device's max DTL is not part of this scan (i.e. it is not in
2140	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
2141	 * for excision.
2142	 */
2143	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
2144		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
2145		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
2146		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
2147		return (B_TRUE);
2148	}
2149	return (B_FALSE);
2150}
2151
2152/*
2153 * Reassess DTLs after a config change or scrub completion.
2154 */
2155void
2156vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
2157{
2158	spa_t *spa = vd->vdev_spa;
2159	avl_tree_t reftree;
2160	int minref;
2161
2162	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2163
2164	for (int c = 0; c < vd->vdev_children; c++)
2165		vdev_dtl_reassess(vd->vdev_child[c], txg,
2166		    scrub_txg, scrub_done);
2167
2168	if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux)
2169		return;
2170
2171	if (vd->vdev_ops->vdev_op_leaf) {
2172		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2173
2174		mutex_enter(&vd->vdev_dtl_lock);
2175
2176		/*
2177		 * If we've completed a scan cleanly then determine
2178		 * if this vdev should remove any DTLs. We only want to
2179		 * excise regions on vdevs that were available during
2180		 * the entire duration of this scan.
2181		 */
2182		if (scrub_txg != 0 &&
2183		    (spa->spa_scrub_started ||
2184		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
2185		    vdev_dtl_should_excise(vd)) {
2186			/*
2187			 * We completed a scrub up to scrub_txg.  If we
2188			 * did it without rebooting, then the scrub dtl
2189			 * will be valid, so excise the old region and
2190			 * fold in the scrub dtl.  Otherwise, leave the
2191			 * dtl as-is if there was an error.
2192			 *
2193			 * There's little trick here: to excise the beginning
2194			 * of the DTL_MISSING map, we put it into a reference
2195			 * tree and then add a segment with refcnt -1 that
2196			 * covers the range [0, scrub_txg).  This means
2197			 * that each txg in that range has refcnt -1 or 0.
2198			 * We then add DTL_SCRUB with a refcnt of 2, so that
2199			 * entries in the range [0, scrub_txg) will have a
2200			 * positive refcnt -- either 1 or 2.  We then convert
2201			 * the reference tree into the new DTL_MISSING map.
2202			 */
2203			space_reftree_create(&reftree);
2204			space_reftree_add_map(&reftree,
2205			    vd->vdev_dtl[DTL_MISSING], 1);
2206			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
2207			space_reftree_add_map(&reftree,
2208			    vd->vdev_dtl[DTL_SCRUB], 2);
2209			space_reftree_generate_map(&reftree,
2210			    vd->vdev_dtl[DTL_MISSING], 1);
2211			space_reftree_destroy(&reftree);
2212		}
2213		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
2214		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2215		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
2216		if (scrub_done)
2217			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
2218		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
2219		if (!vdev_readable(vd))
2220			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
2221		else
2222			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2223			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
2224
2225		/*
2226		 * If the vdev was resilvering and no longer has any
2227		 * DTLs then reset its resilvering flag.
2228		 */
2229		if (vd->vdev_resilver_txg != 0 &&
2230		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
2231		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
2232			vd->vdev_resilver_txg = 0;
2233
2234		mutex_exit(&vd->vdev_dtl_lock);
2235
2236		if (txg != 0)
2237			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
2238		return;
2239	}
2240
2241	mutex_enter(&vd->vdev_dtl_lock);
2242	for (int t = 0; t < DTL_TYPES; t++) {
2243		/* account for child's outage in parent's missing map */
2244		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
2245		if (t == DTL_SCRUB)
2246			continue;			/* leaf vdevs only */
2247		if (t == DTL_PARTIAL)
2248			minref = 1;			/* i.e. non-zero */
2249		else if (vd->vdev_nparity != 0)
2250			minref = vd->vdev_nparity + 1;	/* RAID-Z */
2251		else
2252			minref = vd->vdev_children;	/* any kind of mirror */
2253		space_reftree_create(&reftree);
2254		for (int c = 0; c < vd->vdev_children; c++) {
2255			vdev_t *cvd = vd->vdev_child[c];
2256			mutex_enter(&cvd->vdev_dtl_lock);
2257			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
2258			mutex_exit(&cvd->vdev_dtl_lock);
2259		}
2260		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
2261		space_reftree_destroy(&reftree);
2262	}
2263	mutex_exit(&vd->vdev_dtl_lock);
2264}
2265
2266int
2267vdev_dtl_load(vdev_t *vd)
2268{
2269	spa_t *spa = vd->vdev_spa;
2270	objset_t *mos = spa->spa_meta_objset;
2271	int error = 0;
2272
2273	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
2274		ASSERT(vdev_is_concrete(vd));
2275
2276		error = space_map_open(&vd->vdev_dtl_sm, mos,
2277		    vd->vdev_dtl_object, 0, -1ULL, 0);
2278		if (error)
2279			return (error);
2280		ASSERT(vd->vdev_dtl_sm != NULL);
2281
2282		mutex_enter(&vd->vdev_dtl_lock);
2283
2284		/*
2285		 * Now that we've opened the space_map we need to update
2286		 * the in-core DTL.
2287		 */
2288		space_map_update(vd->vdev_dtl_sm);
2289
2290		error = space_map_load(vd->vdev_dtl_sm,
2291		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
2292		mutex_exit(&vd->vdev_dtl_lock);
2293
2294		return (error);
2295	}
2296
2297	for (int c = 0; c < vd->vdev_children; c++) {
2298		error = vdev_dtl_load(vd->vdev_child[c]);
2299		if (error != 0)
2300			break;
2301	}
2302
2303	return (error);
2304}
2305
2306void
2307vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
2308{
2309	spa_t *spa = vd->vdev_spa;
2310
2311	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
2312	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2313	    zapobj, tx));
2314}
2315
2316uint64_t
2317vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
2318{
2319	spa_t *spa = vd->vdev_spa;
2320	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
2321	    DMU_OT_NONE, 0, tx);
2322
2323	ASSERT(zap != 0);
2324	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2325	    zap, tx));
2326
2327	return (zap);
2328}
2329
2330void
2331vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
2332{
2333	if (vd->vdev_ops != &vdev_hole_ops &&
2334	    vd->vdev_ops != &vdev_missing_ops &&
2335	    vd->vdev_ops != &vdev_root_ops &&
2336	    !vd->vdev_top->vdev_removing) {
2337		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2338			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2339		}
2340		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2341			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2342		}
2343	}
2344	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2345		vdev_construct_zaps(vd->vdev_child[i], tx);
2346	}
2347}
2348
2349void
2350vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2351{
2352	spa_t *spa = vd->vdev_spa;
2353	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2354	objset_t *mos = spa->spa_meta_objset;
2355	range_tree_t *rtsync;
2356	dmu_tx_t *tx;
2357	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2358
2359	ASSERT(vdev_is_concrete(vd));
2360	ASSERT(vd->vdev_ops->vdev_op_leaf);
2361
2362	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2363
2364	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2365		mutex_enter(&vd->vdev_dtl_lock);
2366		space_map_free(vd->vdev_dtl_sm, tx);
2367		space_map_close(vd->vdev_dtl_sm);
2368		vd->vdev_dtl_sm = NULL;
2369		mutex_exit(&vd->vdev_dtl_lock);
2370
2371		/*
2372		 * We only destroy the leaf ZAP for detached leaves or for
2373		 * removed log devices. Removed data devices handle leaf ZAP
2374		 * cleanup later, once cancellation is no longer possible.
2375		 */
2376		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2377		    vd->vdev_top->vdev_islog)) {
2378			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2379			vd->vdev_leaf_zap = 0;
2380		}
2381
2382		dmu_tx_commit(tx);
2383		return;
2384	}
2385
2386	if (vd->vdev_dtl_sm == NULL) {
2387		uint64_t new_object;
2388
2389		new_object = space_map_alloc(mos, tx);
2390		VERIFY3U(new_object, !=, 0);
2391
2392		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2393		    0, -1ULL, 0));
2394		ASSERT(vd->vdev_dtl_sm != NULL);
2395	}
2396
2397	rtsync = range_tree_create(NULL, NULL);
2398
2399	mutex_enter(&vd->vdev_dtl_lock);
2400	range_tree_walk(rt, range_tree_add, rtsync);
2401	mutex_exit(&vd->vdev_dtl_lock);
2402
2403	space_map_truncate(vd->vdev_dtl_sm, tx);
2404	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2405	range_tree_vacate(rtsync, NULL, NULL);
2406
2407	range_tree_destroy(rtsync);
2408
2409	/*
2410	 * If the object for the space map has changed then dirty
2411	 * the top level so that we update the config.
2412	 */
2413	if (object != space_map_object(vd->vdev_dtl_sm)) {
2414		vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, "
2415		    "new object %llu", (u_longlong_t)txg, spa_name(spa),
2416		    (u_longlong_t)object,
2417		    (u_longlong_t)space_map_object(vd->vdev_dtl_sm));
2418		vdev_config_dirty(vd->vdev_top);
2419	}
2420
2421	dmu_tx_commit(tx);
2422
2423	mutex_enter(&vd->vdev_dtl_lock);
2424	space_map_update(vd->vdev_dtl_sm);
2425	mutex_exit(&vd->vdev_dtl_lock);
2426}
2427
2428/*
2429 * Determine whether the specified vdev can be offlined/detached/removed
2430 * without losing data.
2431 */
2432boolean_t
2433vdev_dtl_required(vdev_t *vd)
2434{
2435	spa_t *spa = vd->vdev_spa;
2436	vdev_t *tvd = vd->vdev_top;
2437	uint8_t cant_read = vd->vdev_cant_read;
2438	boolean_t required;
2439
2440	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2441
2442	if (vd == spa->spa_root_vdev || vd == tvd)
2443		return (B_TRUE);
2444
2445	/*
2446	 * Temporarily mark the device as unreadable, and then determine
2447	 * whether this results in any DTL outages in the top-level vdev.
2448	 * If not, we can safely offline/detach/remove the device.
2449	 */
2450	vd->vdev_cant_read = B_TRUE;
2451	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2452	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2453	vd->vdev_cant_read = cant_read;
2454	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2455
2456	if (!required && zio_injection_enabled)
2457		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2458
2459	return (required);
2460}
2461
2462/*
2463 * Determine if resilver is needed, and if so the txg range.
2464 */
2465boolean_t
2466vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2467{
2468	boolean_t needed = B_FALSE;
2469	uint64_t thismin = UINT64_MAX;
2470	uint64_t thismax = 0;
2471
2472	if (vd->vdev_children == 0) {
2473		mutex_enter(&vd->vdev_dtl_lock);
2474		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2475		    vdev_writeable(vd)) {
2476
2477			thismin = vdev_dtl_min(vd);
2478			thismax = vdev_dtl_max(vd);
2479			needed = B_TRUE;
2480		}
2481		mutex_exit(&vd->vdev_dtl_lock);
2482	} else {
2483		for (int c = 0; c < vd->vdev_children; c++) {
2484			vdev_t *cvd = vd->vdev_child[c];
2485			uint64_t cmin, cmax;
2486
2487			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2488				thismin = MIN(thismin, cmin);
2489				thismax = MAX(thismax, cmax);
2490				needed = B_TRUE;
2491			}
2492		}
2493	}
2494
2495	if (needed && minp) {
2496		*minp = thismin;
2497		*maxp = thismax;
2498	}
2499	return (needed);
2500}
2501
2502int
2503vdev_load(vdev_t *vd)
2504{
2505	int error = 0;
2506	/*
2507	 * Recursively load all children.
2508	 */
2509	for (int c = 0; c < vd->vdev_children; c++) {
2510		error = vdev_load(vd->vdev_child[c]);
2511		if (error != 0) {
2512			return (error);
2513		}
2514	}
2515
2516	vdev_set_deflate_ratio(vd);
2517
2518	/*
2519	 * If this is a top-level vdev, initialize its metaslabs.
2520	 */
2521	if (vd == vd->vdev_top && vdev_is_concrete(vd)) {
2522		if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) {
2523			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2524			    VDEV_AUX_CORRUPT_DATA);
2525			vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, "
2526			    "asize=%llu", (u_longlong_t)vd->vdev_ashift,
2527			    (u_longlong_t)vd->vdev_asize);
2528			return (SET_ERROR(ENXIO));
2529		} else if ((error = vdev_metaslab_init(vd, 0)) != 0) {
2530			vdev_dbgmsg(vd, "vdev_load: metaslab_init failed "
2531			    "[error=%d]", error);
2532			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2533			    VDEV_AUX_CORRUPT_DATA);
2534			return (error);
2535		}
2536	}
2537
2538	/*
2539	 * If this is a leaf vdev, load its DTL.
2540	 */
2541	if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) {
2542		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2543		    VDEV_AUX_CORRUPT_DATA);
2544		vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed "
2545		    "[error=%d]", error);
2546		return (error);
2547	}
2548
2549	uint64_t obsolete_sm_object = vdev_obsolete_sm_object(vd);
2550	if (obsolete_sm_object != 0) {
2551		objset_t *mos = vd->vdev_spa->spa_meta_objset;
2552		ASSERT(vd->vdev_asize != 0);
2553		ASSERT(vd->vdev_obsolete_sm == NULL);
2554
2555		if ((error = space_map_open(&vd->vdev_obsolete_sm, mos,
2556		    obsolete_sm_object, 0, vd->vdev_asize, 0))) {
2557			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2558			    VDEV_AUX_CORRUPT_DATA);
2559			vdev_dbgmsg(vd, "vdev_load: space_map_open failed for "
2560			    "obsolete spacemap (obj %llu) [error=%d]",
2561			    (u_longlong_t)obsolete_sm_object, error);
2562			return (error);
2563		}
2564		space_map_update(vd->vdev_obsolete_sm);
2565	}
2566
2567	return (0);
2568}
2569
2570/*
2571 * The special vdev case is used for hot spares and l2cache devices.  Its
2572 * sole purpose it to set the vdev state for the associated vdev.  To do this,
2573 * we make sure that we can open the underlying device, then try to read the
2574 * label, and make sure that the label is sane and that it hasn't been
2575 * repurposed to another pool.
2576 */
2577int
2578vdev_validate_aux(vdev_t *vd)
2579{
2580	nvlist_t *label;
2581	uint64_t guid, version;
2582	uint64_t state;
2583
2584	if (!vdev_readable(vd))
2585		return (0);
2586
2587	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2588		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2589		    VDEV_AUX_CORRUPT_DATA);
2590		return (-1);
2591	}
2592
2593	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2594	    !SPA_VERSION_IS_SUPPORTED(version) ||
2595	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2596	    guid != vd->vdev_guid ||
2597	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2598		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2599		    VDEV_AUX_CORRUPT_DATA);
2600		nvlist_free(label);
2601		return (-1);
2602	}
2603
2604	/*
2605	 * We don't actually check the pool state here.  If it's in fact in
2606	 * use by another pool, we update this fact on the fly when requested.
2607	 */
2608	nvlist_free(label);
2609	return (0);
2610}
2611
2612/*
2613 * Free the objects used to store this vdev's spacemaps, and the array
2614 * that points to them.
2615 */
2616void
2617vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx)
2618{
2619	if (vd->vdev_ms_array == 0)
2620		return;
2621
2622	objset_t *mos = vd->vdev_spa->spa_meta_objset;
2623	uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift;
2624	size_t array_bytes = array_count * sizeof (uint64_t);
2625	uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP);
2626	VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0,
2627	    array_bytes, smobj_array, 0));
2628
2629	for (uint64_t i = 0; i < array_count; i++) {
2630		uint64_t smobj = smobj_array[i];
2631		if (smobj == 0)
2632			continue;
2633
2634		space_map_free_obj(mos, smobj, tx);
2635	}
2636
2637	kmem_free(smobj_array, array_bytes);
2638	VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx));
2639	vd->vdev_ms_array = 0;
2640}
2641
2642static void
2643vdev_remove_empty(vdev_t *vd, uint64_t txg)
2644{
2645	spa_t *spa = vd->vdev_spa;
2646	dmu_tx_t *tx;
2647
2648	ASSERT(vd == vd->vdev_top);
2649	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2650
2651	if (vd->vdev_ms != NULL) {
2652		metaslab_group_t *mg = vd->vdev_mg;
2653
2654		metaslab_group_histogram_verify(mg);
2655		metaslab_class_histogram_verify(mg->mg_class);
2656
2657		for (int m = 0; m < vd->vdev_ms_count; m++) {
2658			metaslab_t *msp = vd->vdev_ms[m];
2659
2660			if (msp == NULL || msp->ms_sm == NULL)
2661				continue;
2662
2663			mutex_enter(&msp->ms_lock);
2664			/*
2665			 * If the metaslab was not loaded when the vdev
2666			 * was removed then the histogram accounting may
2667			 * not be accurate. Update the histogram information
2668			 * here so that we ensure that the metaslab group
2669			 * and metaslab class are up-to-date.
2670			 */
2671			metaslab_group_histogram_remove(mg, msp);
2672
2673			VERIFY0(space_map_allocated(msp->ms_sm));
2674			space_map_close(msp->ms_sm);
2675			msp->ms_sm = NULL;
2676			mutex_exit(&msp->ms_lock);
2677		}
2678
2679		metaslab_group_histogram_verify(mg);
2680		metaslab_class_histogram_verify(mg->mg_class);
2681		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2682			ASSERT0(mg->mg_histogram[i]);
2683	}
2684
2685	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2686	vdev_destroy_spacemaps(vd, tx);
2687
2688	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2689		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2690		vd->vdev_top_zap = 0;
2691	}
2692	dmu_tx_commit(tx);
2693}
2694
2695void
2696vdev_sync_done(vdev_t *vd, uint64_t txg)
2697{
2698	metaslab_t *msp;
2699	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2700
2701	ASSERT(vdev_is_concrete(vd));
2702
2703	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2704		metaslab_sync_done(msp, txg);
2705
2706	if (reassess)
2707		metaslab_sync_reassess(vd->vdev_mg);
2708}
2709
2710void
2711vdev_sync(vdev_t *vd, uint64_t txg)
2712{
2713	spa_t *spa = vd->vdev_spa;
2714	vdev_t *lvd;
2715	metaslab_t *msp;
2716	dmu_tx_t *tx;
2717
2718	if (range_tree_space(vd->vdev_obsolete_segments) > 0) {
2719		dmu_tx_t *tx;
2720
2721		ASSERT(vd->vdev_removing ||
2722		    vd->vdev_ops == &vdev_indirect_ops);
2723
2724		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2725		vdev_indirect_sync_obsolete(vd, tx);
2726		dmu_tx_commit(tx);
2727
2728		/*
2729		 * If the vdev is indirect, it can't have dirty
2730		 * metaslabs or DTLs.
2731		 */
2732		if (vd->vdev_ops == &vdev_indirect_ops) {
2733			ASSERT(txg_list_empty(&vd->vdev_ms_list, txg));
2734			ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg));
2735			return;
2736		}
2737	}
2738
2739	ASSERT(vdev_is_concrete(vd));
2740
2741	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 &&
2742	    !vd->vdev_removing) {
2743		ASSERT(vd == vd->vdev_top);
2744		ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
2745		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2746		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2747		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2748		ASSERT(vd->vdev_ms_array != 0);
2749		vdev_config_dirty(vd);
2750		dmu_tx_commit(tx);
2751	}
2752
2753	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2754		metaslab_sync(msp, txg);
2755		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2756	}
2757
2758	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2759		vdev_dtl_sync(lvd, txg);
2760
2761	/*
2762	 * Remove the metadata associated with this vdev once it's empty.
2763	 * Note that this is typically used for log/cache device removal;
2764	 * we don't empty toplevel vdevs when removing them.  But if
2765	 * a toplevel happens to be emptied, this is not harmful.
2766	 */
2767	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing) {
2768		vdev_remove_empty(vd, txg);
2769	}
2770
2771	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2772}
2773
2774uint64_t
2775vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2776{
2777	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2778}
2779
2780/*
2781 * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2782 * not be opened, and no I/O is attempted.
2783 */
2784int
2785vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2786{
2787	vdev_t *vd, *tvd;
2788
2789	spa_vdev_state_enter(spa, SCL_NONE);
2790
2791	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2792		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2793
2794	if (!vd->vdev_ops->vdev_op_leaf)
2795		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2796
2797	tvd = vd->vdev_top;
2798
2799	/*
2800	 * We don't directly use the aux state here, but if we do a
2801	 * vdev_reopen(), we need this value to be present to remember why we
2802	 * were faulted.
2803	 */
2804	vd->vdev_label_aux = aux;
2805
2806	/*
2807	 * Faulted state takes precedence over degraded.
2808	 */
2809	vd->vdev_delayed_close = B_FALSE;
2810	vd->vdev_faulted = 1ULL;
2811	vd->vdev_degraded = 0ULL;
2812	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2813
2814	/*
2815	 * If this device has the only valid copy of the data, then
2816	 * back off and simply mark the vdev as degraded instead.
2817	 */
2818	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2819		vd->vdev_degraded = 1ULL;
2820		vd->vdev_faulted = 0ULL;
2821
2822		/*
2823		 * If we reopen the device and it's not dead, only then do we
2824		 * mark it degraded.
2825		 */
2826		vdev_reopen(tvd);
2827
2828		if (vdev_readable(vd))
2829			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2830	}
2831
2832	return (spa_vdev_state_exit(spa, vd, 0));
2833}
2834
2835/*
2836 * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2837 * user that something is wrong.  The vdev continues to operate as normal as far
2838 * as I/O is concerned.
2839 */
2840int
2841vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2842{
2843	vdev_t *vd;
2844
2845	spa_vdev_state_enter(spa, SCL_NONE);
2846
2847	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2848		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2849
2850	if (!vd->vdev_ops->vdev_op_leaf)
2851		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2852
2853	/*
2854	 * If the vdev is already faulted, then don't do anything.
2855	 */
2856	if (vd->vdev_faulted || vd->vdev_degraded)
2857		return (spa_vdev_state_exit(spa, NULL, 0));
2858
2859	vd->vdev_degraded = 1ULL;
2860	if (!vdev_is_dead(vd))
2861		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2862		    aux);
2863
2864	return (spa_vdev_state_exit(spa, vd, 0));
2865}
2866
2867/*
2868 * Online the given vdev.
2869 *
2870 * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2871 * spare device should be detached when the device finishes resilvering.
2872 * Second, the online should be treated like a 'test' online case, so no FMA
2873 * events are generated if the device fails to open.
2874 */
2875int
2876vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2877{
2878	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2879	boolean_t wasoffline;
2880	vdev_state_t oldstate;
2881
2882	spa_vdev_state_enter(spa, SCL_NONE);
2883
2884	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2885		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2886
2887	if (!vd->vdev_ops->vdev_op_leaf)
2888		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2889
2890	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
2891	oldstate = vd->vdev_state;
2892
2893	tvd = vd->vdev_top;
2894	vd->vdev_offline = B_FALSE;
2895	vd->vdev_tmpoffline = B_FALSE;
2896	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2897	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2898
2899	/* XXX - L2ARC 1.0 does not support expansion */
2900	if (!vd->vdev_aux) {
2901		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2902			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2903	}
2904
2905	vdev_reopen(tvd);
2906	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2907
2908	if (!vd->vdev_aux) {
2909		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2910			pvd->vdev_expanding = B_FALSE;
2911	}
2912
2913	if (newstate)
2914		*newstate = vd->vdev_state;
2915	if ((flags & ZFS_ONLINE_UNSPARE) &&
2916	    !vdev_is_dead(vd) && vd->vdev_parent &&
2917	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2918	    vd->vdev_parent->vdev_child[0] == vd)
2919		vd->vdev_unspare = B_TRUE;
2920
2921	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2922
2923		/* XXX - L2ARC 1.0 does not support expansion */
2924		if (vd->vdev_aux)
2925			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2926		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2927	}
2928
2929	if (wasoffline ||
2930	    (oldstate < VDEV_STATE_DEGRADED &&
2931	    vd->vdev_state >= VDEV_STATE_DEGRADED))
2932		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
2933
2934	return (spa_vdev_state_exit(spa, vd, 0));
2935}
2936
2937static int
2938vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2939{
2940	vdev_t *vd, *tvd;
2941	int error = 0;
2942	uint64_t generation;
2943	metaslab_group_t *mg;
2944
2945top:
2946	spa_vdev_state_enter(spa, SCL_ALLOC);
2947
2948	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2949		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2950
2951	if (!vd->vdev_ops->vdev_op_leaf)
2952		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2953
2954	tvd = vd->vdev_top;
2955	mg = tvd->vdev_mg;
2956	generation = spa->spa_config_generation + 1;
2957
2958	/*
2959	 * If the device isn't already offline, try to offline it.
2960	 */
2961	if (!vd->vdev_offline) {
2962		/*
2963		 * If this device has the only valid copy of some data,
2964		 * don't allow it to be offlined. Log devices are always
2965		 * expendable.
2966		 */
2967		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2968		    vdev_dtl_required(vd))
2969			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2970
2971		/*
2972		 * If the top-level is a slog and it has had allocations
2973		 * then proceed.  We check that the vdev's metaslab group
2974		 * is not NULL since it's possible that we may have just
2975		 * added this vdev but not yet initialized its metaslabs.
2976		 */
2977		if (tvd->vdev_islog && mg != NULL) {
2978			/*
2979			 * Prevent any future allocations.
2980			 */
2981			metaslab_group_passivate(mg);
2982			(void) spa_vdev_state_exit(spa, vd, 0);
2983
2984			error = spa_reset_logs(spa);
2985
2986			spa_vdev_state_enter(spa, SCL_ALLOC);
2987
2988			/*
2989			 * Check to see if the config has changed.
2990			 */
2991			if (error || generation != spa->spa_config_generation) {
2992				metaslab_group_activate(mg);
2993				if (error)
2994					return (spa_vdev_state_exit(spa,
2995					    vd, error));
2996				(void) spa_vdev_state_exit(spa, vd, 0);
2997				goto top;
2998			}
2999			ASSERT0(tvd->vdev_stat.vs_alloc);
3000		}
3001
3002		/*
3003		 * Offline this device and reopen its top-level vdev.
3004		 * If the top-level vdev is a log device then just offline
3005		 * it. Otherwise, if this action results in the top-level
3006		 * vdev becoming unusable, undo it and fail the request.
3007		 */
3008		vd->vdev_offline = B_TRUE;
3009		vdev_reopen(tvd);
3010
3011		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3012		    vdev_is_dead(tvd)) {
3013			vd->vdev_offline = B_FALSE;
3014			vdev_reopen(tvd);
3015			return (spa_vdev_state_exit(spa, NULL, EBUSY));
3016		}
3017
3018		/*
3019		 * Add the device back into the metaslab rotor so that
3020		 * once we online the device it's open for business.
3021		 */
3022		if (tvd->vdev_islog && mg != NULL)
3023			metaslab_group_activate(mg);
3024	}
3025
3026	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
3027
3028	return (spa_vdev_state_exit(spa, vd, 0));
3029}
3030
3031int
3032vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
3033{
3034	int error;
3035
3036	mutex_enter(&spa->spa_vdev_top_lock);
3037	error = vdev_offline_locked(spa, guid, flags);
3038	mutex_exit(&spa->spa_vdev_top_lock);
3039
3040	return (error);
3041}
3042
3043/*
3044 * Clear the error counts associated with this vdev.  Unlike vdev_online() and
3045 * vdev_offline(), we assume the spa config is locked.  We also clear all
3046 * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
3047 */
3048void
3049vdev_clear(spa_t *spa, vdev_t *vd)
3050{
3051	vdev_t *rvd = spa->spa_root_vdev;
3052
3053	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3054
3055	if (vd == NULL)
3056		vd = rvd;
3057
3058	vd->vdev_stat.vs_read_errors = 0;
3059	vd->vdev_stat.vs_write_errors = 0;
3060	vd->vdev_stat.vs_checksum_errors = 0;
3061
3062	for (int c = 0; c < vd->vdev_children; c++)
3063		vdev_clear(spa, vd->vdev_child[c]);
3064
3065	/*
3066	 * It makes no sense to "clear" an indirect vdev.
3067	 */
3068	if (!vdev_is_concrete(vd))
3069		return;
3070
3071	/*
3072	 * If we're in the FAULTED state or have experienced failed I/O, then
3073	 * clear the persistent state and attempt to reopen the device.  We
3074	 * also mark the vdev config dirty, so that the new faulted state is
3075	 * written out to disk.
3076	 */
3077	if (vd->vdev_faulted || vd->vdev_degraded ||
3078	    !vdev_readable(vd) || !vdev_writeable(vd)) {
3079
3080		/*
3081		 * When reopening in reponse to a clear event, it may be due to
3082		 * a fmadm repair request.  In this case, if the device is
3083		 * still broken, we want to still post the ereport again.
3084		 */
3085		vd->vdev_forcefault = B_TRUE;
3086
3087		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
3088		vd->vdev_cant_read = B_FALSE;
3089		vd->vdev_cant_write = B_FALSE;
3090
3091		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
3092
3093		vd->vdev_forcefault = B_FALSE;
3094
3095		if (vd != rvd && vdev_writeable(vd->vdev_top))
3096			vdev_state_dirty(vd->vdev_top);
3097
3098		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
3099			spa_async_request(spa, SPA_ASYNC_RESILVER);
3100
3101		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
3102	}
3103
3104	/*
3105	 * When clearing a FMA-diagnosed fault, we always want to
3106	 * unspare the device, as we assume that the original spare was
3107	 * done in response to the FMA fault.
3108	 */
3109	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
3110	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3111	    vd->vdev_parent->vdev_child[0] == vd)
3112		vd->vdev_unspare = B_TRUE;
3113}
3114
3115boolean_t
3116vdev_is_dead(vdev_t *vd)
3117{
3118	/*
3119	 * Holes and missing devices are always considered "dead".
3120	 * This simplifies the code since we don't have to check for
3121	 * these types of devices in the various code paths.
3122	 * Instead we rely on the fact that we skip over dead devices
3123	 * before issuing I/O to them.
3124	 */
3125	return (vd->vdev_state < VDEV_STATE_DEGRADED ||
3126	    vd->vdev_ops == &vdev_hole_ops ||
3127	    vd->vdev_ops == &vdev_missing_ops);
3128}
3129
3130boolean_t
3131vdev_readable(vdev_t *vd)
3132{
3133	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
3134}
3135
3136boolean_t
3137vdev_writeable(vdev_t *vd)
3138{
3139	return (!vdev_is_dead(vd) && !vd->vdev_cant_write &&
3140	    vdev_is_concrete(vd));
3141}
3142
3143boolean_t
3144vdev_allocatable(vdev_t *vd)
3145{
3146	uint64_t state = vd->vdev_state;
3147
3148	/*
3149	 * We currently allow allocations from vdevs which may be in the
3150	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
3151	 * fails to reopen then we'll catch it later when we're holding
3152	 * the proper locks.  Note that we have to get the vdev state
3153	 * in a local variable because although it changes atomically,
3154	 * we're asking two separate questions about it.
3155	 */
3156	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
3157	    !vd->vdev_cant_write && vdev_is_concrete(vd) &&
3158	    vd->vdev_mg->mg_initialized);
3159}
3160
3161boolean_t
3162vdev_accessible(vdev_t *vd, zio_t *zio)
3163{
3164	ASSERT(zio->io_vd == vd);
3165
3166	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
3167		return (B_FALSE);
3168
3169	if (zio->io_type == ZIO_TYPE_READ)
3170		return (!vd->vdev_cant_read);
3171
3172	if (zio->io_type == ZIO_TYPE_WRITE)
3173		return (!vd->vdev_cant_write);
3174
3175	return (B_TRUE);
3176}
3177
3178/*
3179 * Get statistics for the given vdev.
3180 */
3181void
3182vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
3183{
3184	spa_t *spa = vd->vdev_spa;
3185	vdev_t *rvd = spa->spa_root_vdev;
3186	vdev_t *tvd = vd->vdev_top;
3187
3188	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
3189
3190	mutex_enter(&vd->vdev_stat_lock);
3191	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
3192	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
3193	vs->vs_state = vd->vdev_state;
3194	vs->vs_rsize = vdev_get_min_asize(vd);
3195	if (vd->vdev_ops->vdev_op_leaf)
3196		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
3197	/*
3198	 * Report expandable space on top-level, non-auxillary devices only.
3199	 * The expandable space is reported in terms of metaslab sized units
3200	 * since that determines how much space the pool can expand.
3201	 */
3202	if (vd->vdev_aux == NULL && tvd != NULL) {
3203		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
3204		    spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
3205	}
3206	if (vd->vdev_aux == NULL && vd == vd->vdev_top &&
3207	    vdev_is_concrete(vd)) {
3208		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
3209	}
3210
3211	/*
3212	 * If we're getting stats on the root vdev, aggregate the I/O counts
3213	 * over all top-level vdevs (i.e. the direct children of the root).
3214	 */
3215	if (vd == rvd) {
3216		for (int c = 0; c < rvd->vdev_children; c++) {
3217			vdev_t *cvd = rvd->vdev_child[c];
3218			vdev_stat_t *cvs = &cvd->vdev_stat;
3219
3220			for (int t = 0; t < ZIO_TYPES; t++) {
3221				vs->vs_ops[t] += cvs->vs_ops[t];
3222				vs->vs_bytes[t] += cvs->vs_bytes[t];
3223			}
3224			cvs->vs_scan_removing = cvd->vdev_removing;
3225		}
3226	}
3227	mutex_exit(&vd->vdev_stat_lock);
3228}
3229
3230void
3231vdev_clear_stats(vdev_t *vd)
3232{
3233	mutex_enter(&vd->vdev_stat_lock);
3234	vd->vdev_stat.vs_space = 0;
3235	vd->vdev_stat.vs_dspace = 0;
3236	vd->vdev_stat.vs_alloc = 0;
3237	mutex_exit(&vd->vdev_stat_lock);
3238}
3239
3240void
3241vdev_scan_stat_init(vdev_t *vd)
3242{
3243	vdev_stat_t *vs = &vd->vdev_stat;
3244
3245	for (int c = 0; c < vd->vdev_children; c++)
3246		vdev_scan_stat_init(vd->vdev_child[c]);
3247
3248	mutex_enter(&vd->vdev_stat_lock);
3249	vs->vs_scan_processed = 0;
3250	mutex_exit(&vd->vdev_stat_lock);
3251}
3252
3253void
3254vdev_stat_update(zio_t *zio, uint64_t psize)
3255{
3256	spa_t *spa = zio->io_spa;
3257	vdev_t *rvd = spa->spa_root_vdev;
3258	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
3259	vdev_t *pvd;
3260	uint64_t txg = zio->io_txg;
3261	vdev_stat_t *vs = &vd->vdev_stat;
3262	zio_type_t type = zio->io_type;
3263	int flags = zio->io_flags;
3264
3265	/*
3266	 * If this i/o is a gang leader, it didn't do any actual work.
3267	 */
3268	if (zio->io_gang_tree)
3269		return;
3270
3271	if (zio->io_error == 0) {
3272		/*
3273		 * If this is a root i/o, don't count it -- we've already
3274		 * counted the top-level vdevs, and vdev_get_stats() will
3275		 * aggregate them when asked.  This reduces contention on
3276		 * the root vdev_stat_lock and implicitly handles blocks
3277		 * that compress away to holes, for which there is no i/o.
3278		 * (Holes never create vdev children, so all the counters
3279		 * remain zero, which is what we want.)
3280		 *
3281		 * Note: this only applies to successful i/o (io_error == 0)
3282		 * because unlike i/o counts, errors are not additive.
3283		 * When reading a ditto block, for example, failure of
3284		 * one top-level vdev does not imply a root-level error.
3285		 */
3286		if (vd == rvd)
3287			return;
3288
3289		ASSERT(vd == zio->io_vd);
3290
3291		if (flags & ZIO_FLAG_IO_BYPASS)
3292			return;
3293
3294		mutex_enter(&vd->vdev_stat_lock);
3295
3296		if (flags & ZIO_FLAG_IO_REPAIR) {
3297			if (flags & ZIO_FLAG_SCAN_THREAD) {
3298				dsl_scan_phys_t *scn_phys =
3299				    &spa->spa_dsl_pool->dp_scan->scn_phys;
3300				uint64_t *processed = &scn_phys->scn_processed;
3301
3302				/* XXX cleanup? */
3303				if (vd->vdev_ops->vdev_op_leaf)
3304					atomic_add_64(processed, psize);
3305				vs->vs_scan_processed += psize;
3306			}
3307
3308			if (flags & ZIO_FLAG_SELF_HEAL)
3309				vs->vs_self_healed += psize;
3310		}
3311
3312		vs->vs_ops[type]++;
3313		vs->vs_bytes[type] += psize;
3314
3315		mutex_exit(&vd->vdev_stat_lock);
3316		return;
3317	}
3318
3319	if (flags & ZIO_FLAG_SPECULATIVE)
3320		return;
3321
3322	/*
3323	 * If this is an I/O error that is going to be retried, then ignore the
3324	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
3325	 * hard errors, when in reality they can happen for any number of
3326	 * innocuous reasons (bus resets, MPxIO link failure, etc).
3327	 */
3328	if (zio->io_error == EIO &&
3329	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
3330		return;
3331
3332	/*
3333	 * Intent logs writes won't propagate their error to the root
3334	 * I/O so don't mark these types of failures as pool-level
3335	 * errors.
3336	 */
3337	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
3338		return;
3339
3340	mutex_enter(&vd->vdev_stat_lock);
3341	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
3342		if (zio->io_error == ECKSUM)
3343			vs->vs_checksum_errors++;
3344		else
3345			vs->vs_read_errors++;
3346	}
3347	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
3348		vs->vs_write_errors++;
3349	mutex_exit(&vd->vdev_stat_lock);
3350
3351	if (spa->spa_load_state == SPA_LOAD_NONE &&
3352	    type == ZIO_TYPE_WRITE && txg != 0 &&
3353	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
3354	    (flags & ZIO_FLAG_SCAN_THREAD) ||
3355	    spa->spa_claiming)) {
3356		/*
3357		 * This is either a normal write (not a repair), or it's
3358		 * a repair induced by the scrub thread, or it's a repair
3359		 * made by zil_claim() during spa_load() in the first txg.
3360		 * In the normal case, we commit the DTL change in the same
3361		 * txg as the block was born.  In the scrub-induced repair
3362		 * case, we know that scrubs run in first-pass syncing context,
3363		 * so we commit the DTL change in spa_syncing_txg(spa).
3364		 * In the zil_claim() case, we commit in spa_first_txg(spa).
3365		 *
3366		 * We currently do not make DTL entries for failed spontaneous
3367		 * self-healing writes triggered by normal (non-scrubbing)
3368		 * reads, because we have no transactional context in which to
3369		 * do so -- and it's not clear that it'd be desirable anyway.
3370		 */
3371		if (vd->vdev_ops->vdev_op_leaf) {
3372			uint64_t commit_txg = txg;
3373			if (flags & ZIO_FLAG_SCAN_THREAD) {
3374				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3375				ASSERT(spa_sync_pass(spa) == 1);
3376				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
3377				commit_txg = spa_syncing_txg(spa);
3378			} else if (spa->spa_claiming) {
3379				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3380				commit_txg = spa_first_txg(spa);
3381			}
3382			ASSERT(commit_txg >= spa_syncing_txg(spa));
3383			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
3384				return;
3385			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3386				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
3387			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
3388		}
3389		if (vd != rvd)
3390			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
3391	}
3392}
3393
3394/*
3395 * Update the in-core space usage stats for this vdev, its metaslab class,
3396 * and the root vdev.
3397 */
3398void
3399vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
3400    int64_t space_delta)
3401{
3402	int64_t dspace_delta = space_delta;
3403	spa_t *spa = vd->vdev_spa;
3404	vdev_t *rvd = spa->spa_root_vdev;
3405	metaslab_group_t *mg = vd->vdev_mg;
3406	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
3407
3408	ASSERT(vd == vd->vdev_top);
3409
3410	/*
3411	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
3412	 * factor.  We must calculate this here and not at the root vdev
3413	 * because the root vdev's psize-to-asize is simply the max of its
3414	 * childrens', thus not accurate enough for us.
3415	 */
3416	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
3417	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
3418	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
3419	    vd->vdev_deflate_ratio;
3420
3421	mutex_enter(&vd->vdev_stat_lock);
3422	vd->vdev_stat.vs_alloc += alloc_delta;
3423	vd->vdev_stat.vs_space += space_delta;
3424	vd->vdev_stat.vs_dspace += dspace_delta;
3425	mutex_exit(&vd->vdev_stat_lock);
3426
3427	if (mc == spa_normal_class(spa)) {
3428		mutex_enter(&rvd->vdev_stat_lock);
3429		rvd->vdev_stat.vs_alloc += alloc_delta;
3430		rvd->vdev_stat.vs_space += space_delta;
3431		rvd->vdev_stat.vs_dspace += dspace_delta;
3432		mutex_exit(&rvd->vdev_stat_lock);
3433	}
3434
3435	if (mc != NULL) {
3436		ASSERT(rvd == vd->vdev_parent);
3437		ASSERT(vd->vdev_ms_count != 0);
3438
3439		metaslab_class_space_update(mc,
3440		    alloc_delta, defer_delta, space_delta, dspace_delta);
3441	}
3442}
3443
3444/*
3445 * Mark a top-level vdev's config as dirty, placing it on the dirty list
3446 * so that it will be written out next time the vdev configuration is synced.
3447 * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3448 */
3449void
3450vdev_config_dirty(vdev_t *vd)
3451{
3452	spa_t *spa = vd->vdev_spa;
3453	vdev_t *rvd = spa->spa_root_vdev;
3454	int c;
3455
3456	ASSERT(spa_writeable(spa));
3457
3458	/*
3459	 * If this is an aux vdev (as with l2cache and spare devices), then we
3460	 * update the vdev config manually and set the sync flag.
3461	 */
3462	if (vd->vdev_aux != NULL) {
3463		spa_aux_vdev_t *sav = vd->vdev_aux;
3464		nvlist_t **aux;
3465		uint_t naux;
3466
3467		for (c = 0; c < sav->sav_count; c++) {
3468			if (sav->sav_vdevs[c] == vd)
3469				break;
3470		}
3471
3472		if (c == sav->sav_count) {
3473			/*
3474			 * We're being removed.  There's nothing more to do.
3475			 */
3476			ASSERT(sav->sav_sync == B_TRUE);
3477			return;
3478		}
3479
3480		sav->sav_sync = B_TRUE;
3481
3482		if (nvlist_lookup_nvlist_array(sav->sav_config,
3483		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3484			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3485			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3486		}
3487
3488		ASSERT(c < naux);
3489
3490		/*
3491		 * Setting the nvlist in the middle if the array is a little
3492		 * sketchy, but it will work.
3493		 */
3494		nvlist_free(aux[c]);
3495		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3496
3497		return;
3498	}
3499
3500	/*
3501	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3502	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3503	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3504	 * so this is sufficient to ensure mutual exclusion.
3505	 */
3506	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3507	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3508	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3509
3510	if (vd == rvd) {
3511		for (c = 0; c < rvd->vdev_children; c++)
3512			vdev_config_dirty(rvd->vdev_child[c]);
3513	} else {
3514		ASSERT(vd == vd->vdev_top);
3515
3516		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3517		    vdev_is_concrete(vd)) {
3518			list_insert_head(&spa->spa_config_dirty_list, vd);
3519		}
3520	}
3521}
3522
3523void
3524vdev_config_clean(vdev_t *vd)
3525{
3526	spa_t *spa = vd->vdev_spa;
3527
3528	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3529	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3530	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3531
3532	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3533	list_remove(&spa->spa_config_dirty_list, vd);
3534}
3535
3536/*
3537 * Mark a top-level vdev's state as dirty, so that the next pass of
3538 * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3539 * the state changes from larger config changes because they require
3540 * much less locking, and are often needed for administrative actions.
3541 */
3542void
3543vdev_state_dirty(vdev_t *vd)
3544{
3545	spa_t *spa = vd->vdev_spa;
3546
3547	ASSERT(spa_writeable(spa));
3548	ASSERT(vd == vd->vdev_top);
3549
3550	/*
3551	 * The state list is protected by the SCL_STATE lock.  The caller
3552	 * must either hold SCL_STATE as writer, or must be the sync thread
3553	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3554	 * so this is sufficient to ensure mutual exclusion.
3555	 */
3556	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3557	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3558	    spa_config_held(spa, SCL_STATE, RW_READER)));
3559
3560	if (!list_link_active(&vd->vdev_state_dirty_node) &&
3561	    vdev_is_concrete(vd))
3562		list_insert_head(&spa->spa_state_dirty_list, vd);
3563}
3564
3565void
3566vdev_state_clean(vdev_t *vd)
3567{
3568	spa_t *spa = vd->vdev_spa;
3569
3570	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3571	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3572	    spa_config_held(spa, SCL_STATE, RW_READER)));
3573
3574	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3575	list_remove(&spa->spa_state_dirty_list, vd);
3576}
3577
3578/*
3579 * Propagate vdev state up from children to parent.
3580 */
3581void
3582vdev_propagate_state(vdev_t *vd)
3583{
3584	spa_t *spa = vd->vdev_spa;
3585	vdev_t *rvd = spa->spa_root_vdev;
3586	int degraded = 0, faulted = 0;
3587	int corrupted = 0;
3588	vdev_t *child;
3589
3590	if (vd->vdev_children > 0) {
3591		for (int c = 0; c < vd->vdev_children; c++) {
3592			child = vd->vdev_child[c];
3593
3594			/*
3595			 * Don't factor holes or indirect vdevs into the
3596			 * decision.
3597			 */
3598			if (!vdev_is_concrete(child))
3599				continue;
3600
3601			if (!vdev_readable(child) ||
3602			    (!vdev_writeable(child) && spa_writeable(spa))) {
3603				/*
3604				 * Root special: if there is a top-level log
3605				 * device, treat the root vdev as if it were
3606				 * degraded.
3607				 */
3608				if (child->vdev_islog && vd == rvd)
3609					degraded++;
3610				else
3611					faulted++;
3612			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3613				degraded++;
3614			}
3615
3616			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3617				corrupted++;
3618		}
3619
3620		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3621
3622		/*
3623		 * Root special: if there is a top-level vdev that cannot be
3624		 * opened due to corrupted metadata, then propagate the root
3625		 * vdev's aux state as 'corrupt' rather than 'insufficient
3626		 * replicas'.
3627		 */
3628		if (corrupted && vd == rvd &&
3629		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3630			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3631			    VDEV_AUX_CORRUPT_DATA);
3632	}
3633
3634	if (vd->vdev_parent)
3635		vdev_propagate_state(vd->vdev_parent);
3636}
3637
3638/*
3639 * Set a vdev's state.  If this is during an open, we don't update the parent
3640 * state, because we're in the process of opening children depth-first.
3641 * Otherwise, we propagate the change to the parent.
3642 *
3643 * If this routine places a device in a faulted state, an appropriate ereport is
3644 * generated.
3645 */
3646void
3647vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3648{
3649	uint64_t save_state;
3650	spa_t *spa = vd->vdev_spa;
3651
3652	if (state == vd->vdev_state) {
3653		vd->vdev_stat.vs_aux = aux;
3654		return;
3655	}
3656
3657	save_state = vd->vdev_state;
3658
3659	vd->vdev_state = state;
3660	vd->vdev_stat.vs_aux = aux;
3661
3662	/*
3663	 * If we are setting the vdev state to anything but an open state, then
3664	 * always close the underlying device unless the device has requested
3665	 * a delayed close (i.e. we're about to remove or fault the device).
3666	 * Otherwise, we keep accessible but invalid devices open forever.
3667	 * We don't call vdev_close() itself, because that implies some extra
3668	 * checks (offline, etc) that we don't want here.  This is limited to
3669	 * leaf devices, because otherwise closing the device will affect other
3670	 * children.
3671	 */
3672	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3673	    vd->vdev_ops->vdev_op_leaf)
3674		vd->vdev_ops->vdev_op_close(vd);
3675
3676	/*
3677	 * If we have brought this vdev back into service, we need
3678	 * to notify fmd so that it can gracefully repair any outstanding
3679	 * cases due to a missing device.  We do this in all cases, even those
3680	 * that probably don't correlate to a repaired fault.  This is sure to
3681	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3682	 * this is a transient state it's OK, as the retire agent will
3683	 * double-check the state of the vdev before repairing it.
3684	 */
3685	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3686	    vd->vdev_prevstate != state)
3687		zfs_post_state_change(spa, vd);
3688
3689	if (vd->vdev_removed &&
3690	    state == VDEV_STATE_CANT_OPEN &&
3691	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3692		/*
3693		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3694		 * device was previously marked removed and someone attempted to
3695		 * reopen it.  If this failed due to a nonexistent device, then
3696		 * keep the device in the REMOVED state.  We also let this be if
3697		 * it is one of our special test online cases, which is only
3698		 * attempting to online the device and shouldn't generate an FMA
3699		 * fault.
3700		 */
3701		vd->vdev_state = VDEV_STATE_REMOVED;
3702		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3703	} else if (state == VDEV_STATE_REMOVED) {
3704		vd->vdev_removed = B_TRUE;
3705	} else if (state == VDEV_STATE_CANT_OPEN) {
3706		/*
3707		 * If we fail to open a vdev during an import or recovery, we
3708		 * mark it as "not available", which signifies that it was
3709		 * never there to begin with.  Failure to open such a device
3710		 * is not considered an error.
3711		 */
3712		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3713		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3714		    vd->vdev_ops->vdev_op_leaf)
3715			vd->vdev_not_present = 1;
3716
3717		/*
3718		 * Post the appropriate ereport.  If the 'prevstate' field is
3719		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3720		 * that this is part of a vdev_reopen().  In this case, we don't
3721		 * want to post the ereport if the device was already in the
3722		 * CANT_OPEN state beforehand.
3723		 *
3724		 * If the 'checkremove' flag is set, then this is an attempt to
3725		 * online the device in response to an insertion event.  If we
3726		 * hit this case, then we have detected an insertion event for a
3727		 * faulted or offline device that wasn't in the removed state.
3728		 * In this scenario, we don't post an ereport because we are
3729		 * about to replace the device, or attempt an online with
3730		 * vdev_forcefault, which will generate the fault for us.
3731		 */
3732		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3733		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3734		    vd != spa->spa_root_vdev) {
3735			const char *class;
3736
3737			switch (aux) {
3738			case VDEV_AUX_OPEN_FAILED:
3739				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3740				break;
3741			case VDEV_AUX_CORRUPT_DATA:
3742				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3743				break;
3744			case VDEV_AUX_NO_REPLICAS:
3745				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3746				break;
3747			case VDEV_AUX_BAD_GUID_SUM:
3748				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3749				break;
3750			case VDEV_AUX_TOO_SMALL:
3751				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3752				break;
3753			case VDEV_AUX_BAD_LABEL:
3754				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3755				break;
3756			default:
3757				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3758			}
3759
3760			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3761		}
3762
3763		/* Erase any notion of persistent removed state */
3764		vd->vdev_removed = B_FALSE;
3765	} else {
3766		vd->vdev_removed = B_FALSE;
3767	}
3768
3769	if (!isopen && vd->vdev_parent)
3770		vdev_propagate_state(vd->vdev_parent);
3771}
3772
3773boolean_t
3774vdev_children_are_offline(vdev_t *vd)
3775{
3776	ASSERT(!vd->vdev_ops->vdev_op_leaf);
3777
3778	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3779		if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE)
3780			return (B_FALSE);
3781	}
3782
3783	return (B_TRUE);
3784}
3785
3786/*
3787 * Check the vdev configuration to ensure that it's capable of supporting
3788 * a root pool. We do not support partial configuration.
3789 * In addition, only a single top-level vdev is allowed.
3790 */
3791boolean_t
3792vdev_is_bootable(vdev_t *vd)
3793{
3794	if (!vd->vdev_ops->vdev_op_leaf) {
3795		char *vdev_type = vd->vdev_ops->vdev_op_type;
3796
3797		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3798		    vd->vdev_children > 1) {
3799			return (B_FALSE);
3800		} else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0 ||
3801		    strcmp(vdev_type, VDEV_TYPE_INDIRECT) == 0) {
3802			return (B_FALSE);
3803		}
3804	}
3805
3806	for (int c = 0; c < vd->vdev_children; c++) {
3807		if (!vdev_is_bootable(vd->vdev_child[c]))
3808			return (B_FALSE);
3809	}
3810	return (B_TRUE);
3811}
3812
3813boolean_t
3814vdev_is_concrete(vdev_t *vd)
3815{
3816	vdev_ops_t *ops = vd->vdev_ops;
3817	if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops ||
3818	    ops == &vdev_missing_ops || ops == &vdev_root_ops) {
3819		return (B_FALSE);
3820	} else {
3821		return (B_TRUE);
3822	}
3823}
3824
3825/*
3826 * Determine if a log device has valid content.  If the vdev was
3827 * removed or faulted in the MOS config then we know that
3828 * the content on the log device has already been written to the pool.
3829 */
3830boolean_t
3831vdev_log_state_valid(vdev_t *vd)
3832{
3833	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3834	    !vd->vdev_removed)
3835		return (B_TRUE);
3836
3837	for (int c = 0; c < vd->vdev_children; c++)
3838		if (vdev_log_state_valid(vd->vdev_child[c]))
3839			return (B_TRUE);
3840
3841	return (B_FALSE);
3842}
3843
3844/*
3845 * Expand a vdev if possible.
3846 */
3847void
3848vdev_expand(vdev_t *vd, uint64_t txg)
3849{
3850	ASSERT(vd->vdev_top == vd);
3851	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3852
3853	vdev_set_deflate_ratio(vd);
3854
3855	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count &&
3856	    vdev_is_concrete(vd)) {
3857		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3858		vdev_config_dirty(vd);
3859	}
3860}
3861
3862/*
3863 * Split a vdev.
3864 */
3865void
3866vdev_split(vdev_t *vd)
3867{
3868	vdev_t *cvd, *pvd = vd->vdev_parent;
3869
3870	vdev_remove_child(pvd, vd);
3871	vdev_compact_children(pvd);
3872
3873	cvd = pvd->vdev_child[0];
3874	if (pvd->vdev_children == 1) {
3875		vdev_remove_parent(cvd);
3876		cvd->vdev_splitting = B_TRUE;
3877	}
3878	vdev_propagate_state(cvd);
3879}
3880
3881void
3882vdev_deadman(vdev_t *vd)
3883{
3884	for (int c = 0; c < vd->vdev_children; c++) {
3885		vdev_t *cvd = vd->vdev_child[c];
3886
3887		vdev_deadman(cvd);
3888	}
3889
3890	if (vd->vdev_ops->vdev_op_leaf) {
3891		vdev_queue_t *vq = &vd->vdev_queue;
3892
3893		mutex_enter(&vq->vq_lock);
3894		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3895			spa_t *spa = vd->vdev_spa;
3896			zio_t *fio;
3897			uint64_t delta;
3898
3899			/*
3900			 * Look at the head of all the pending queues,
3901			 * if any I/O has been outstanding for longer than
3902			 * the spa_deadman_synctime we panic the system.
3903			 */
3904			fio = avl_first(&vq->vq_active_tree);
3905			delta = gethrtime() - fio->io_timestamp;
3906			if (delta > spa_deadman_synctime(spa)) {
3907				vdev_dbgmsg(vd, "SLOW IO: zio timestamp "
3908				    "%lluns, delta %lluns, last io %lluns",
3909				    fio->io_timestamp, (u_longlong_t)delta,
3910				    vq->vq_io_complete_ts);
3911				fm_panic("I/O to pool '%s' appears to be "
3912				    "hung.", spa_name(spa));
3913			}
3914		}
3915		mutex_exit(&vq->vq_lock);
3916	}
3917}
3918