xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev.c (revision b037f3db)
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 
57 static 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 */
72 int 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  */
78 int metaslabs_per_vdev = 200;
79 
80 boolean_t vdev_validate_skip = B_FALSE;
81 
82 /*PRINTFLIKE2*/
83 void
84 vdev_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 
104 void
105 vdev_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  */
158 static vdev_ops_t *
159 vdev_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  */
174 uint64_t
175 vdev_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  */
194 uint64_t
195 vdev_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 
224 void
225 vdev_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 
233 vdev_t *
234 vdev_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 
248 vdev_t *
249 vdev_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 
264 static int
265 vdev_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 
278 int
279 vdev_count_leaves(spa_t *spa)
280 {
281 	return (vdev_count_leaves_impl(spa->spa_root_vdev));
282 }
283 
284 void
285 vdev_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 
325 void
326 vdev_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  */
362 void
363 vdev_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  */
392 vdev_t *
393 vdev_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  */
459 int
460 vdev_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 
707 void
708 vdev_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  */
811 static void
812 vdev_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 
876 static void
877 vdev_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  */
891 vdev_t *
892 vdev_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  */
925 void
926 vdev_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 
965 int
966 vdev_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 
1043 void
1044 vdev_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 
1064 typedef struct vdev_probe_stats {
1065 	boolean_t	vps_readable;
1066 	boolean_t	vps_writeable;
1067 	int		vps_flags;
1068 } vdev_probe_stats_t;
1069 
1070 static void
1071 vdev_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  */
1132 zio_t *
1133 vdev_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 
1223 static void
1224 vdev_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 
1233 boolean_t
1234 vdev_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 
1245 void
1246 vdev_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  */
1278 static void
1279 vdev_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  */
1290 int
1291 vdev_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 	/*
1482 	 * Track the min and max ashift values for normal data devices.
1483 	 */
1484 	if (vd->vdev_top == vd && vd->vdev_ashift != 0 &&
1485 	    !vd->vdev_islog && vd->vdev_aux == NULL) {
1486 		if (vd->vdev_ashift > spa->spa_max_ashift)
1487 			spa->spa_max_ashift = vd->vdev_ashift;
1488 		if (vd->vdev_ashift < spa->spa_min_ashift)
1489 			spa->spa_min_ashift = vd->vdev_ashift;
1490 	}
1491 
1492 	/*
1493 	 * If a leaf vdev has a DTL, and seems healthy, then kick off a
1494 	 * resilver.  But don't do this if we are doing a reopen for a scrub,
1495 	 * since this would just restart the scrub we are already doing.
1496 	 */
1497 	if (vd->vdev_ops->vdev_op_leaf && !spa->spa_scrub_reopen &&
1498 	    vdev_resilver_needed(vd, NULL, NULL))
1499 		spa_async_request(spa, SPA_ASYNC_RESILVER);
1500 
1501 	return (0);
1502 }
1503 
1504 /*
1505  * Called once the vdevs are all opened, this routine validates the label
1506  * contents. This needs to be done before vdev_load() so that we don't
1507  * inadvertently do repair I/Os to the wrong device.
1508  *
1509  * This function will only return failure if one of the vdevs indicates that it
1510  * has since been destroyed or exported.  This is only possible if
1511  * /etc/zfs/zpool.cache was readonly at the time.  Otherwise, the vdev state
1512  * will be updated but the function will return 0.
1513  */
1514 int
1515 vdev_validate(vdev_t *vd)
1516 {
1517 	spa_t *spa = vd->vdev_spa;
1518 	nvlist_t *label;
1519 	uint64_t guid = 0, aux_guid = 0, top_guid;
1520 	uint64_t state;
1521 	nvlist_t *nvl;
1522 	uint64_t txg;
1523 
1524 	if (vdev_validate_skip)
1525 		return (0);
1526 
1527 	for (uint64_t c = 0; c < vd->vdev_children; c++)
1528 		if (vdev_validate(vd->vdev_child[c]) != 0)
1529 			return (SET_ERROR(EBADF));
1530 
1531 	/*
1532 	 * If the device has already failed, or was marked offline, don't do
1533 	 * any further validation.  Otherwise, label I/O will fail and we will
1534 	 * overwrite the previous state.
1535 	 */
1536 	if (!vd->vdev_ops->vdev_op_leaf || !vdev_readable(vd))
1537 		return (0);
1538 
1539 	/*
1540 	 * If we are performing an extreme rewind, we allow for a label that
1541 	 * was modified at a point after the current txg.
1542 	 */
1543 	if (spa->spa_extreme_rewind || spa_last_synced_txg(spa) == 0)
1544 		txg = UINT64_MAX;
1545 	else
1546 		txg = spa_last_synced_txg(spa);
1547 
1548 	if ((label = vdev_label_read_config(vd, txg)) == NULL) {
1549 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
1550 		    VDEV_AUX_BAD_LABEL);
1551 		vdev_dbgmsg(vd, "vdev_validate: failed reading config");
1552 		return (0);
1553 	}
1554 
1555 	/*
1556 	 * Determine if this vdev has been split off into another
1557 	 * pool.  If so, then refuse to open it.
1558 	 */
1559 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_SPLIT_GUID,
1560 	    &aux_guid) == 0 && aux_guid == spa_guid(spa)) {
1561 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1562 		    VDEV_AUX_SPLIT_POOL);
1563 		nvlist_free(label);
1564 		vdev_dbgmsg(vd, "vdev_validate: vdev split into other pool");
1565 		return (0);
1566 	}
1567 
1568 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_GUID, &guid) != 0) {
1569 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1570 		    VDEV_AUX_CORRUPT_DATA);
1571 		nvlist_free(label);
1572 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1573 		    ZPOOL_CONFIG_POOL_GUID);
1574 		return (0);
1575 	}
1576 
1577 	/*
1578 	 * If config is not trusted then ignore the spa guid check. This is
1579 	 * necessary because if the machine crashed during a re-guid the new
1580 	 * guid might have been written to all of the vdev labels, but not the
1581 	 * cached config. The check will be performed again once we have the
1582 	 * trusted config from the MOS.
1583 	 */
1584 	if (spa->spa_trust_config && guid != spa_guid(spa)) {
1585 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1586 		    VDEV_AUX_CORRUPT_DATA);
1587 		nvlist_free(label);
1588 		vdev_dbgmsg(vd, "vdev_validate: vdev label pool_guid doesn't "
1589 		    "match config (%llu != %llu)", (u_longlong_t)guid,
1590 		    (u_longlong_t)spa_guid(spa));
1591 		return (0);
1592 	}
1593 
1594 	if (nvlist_lookup_nvlist(label, ZPOOL_CONFIG_VDEV_TREE, &nvl)
1595 	    != 0 || nvlist_lookup_uint64(nvl, ZPOOL_CONFIG_ORIG_GUID,
1596 	    &aux_guid) != 0)
1597 		aux_guid = 0;
1598 
1599 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0) {
1600 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1601 		    VDEV_AUX_CORRUPT_DATA);
1602 		nvlist_free(label);
1603 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1604 		    ZPOOL_CONFIG_GUID);
1605 		return (0);
1606 	}
1607 
1608 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_TOP_GUID, &top_guid)
1609 	    != 0) {
1610 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1611 		    VDEV_AUX_CORRUPT_DATA);
1612 		nvlist_free(label);
1613 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1614 		    ZPOOL_CONFIG_TOP_GUID);
1615 		return (0);
1616 	}
1617 
1618 	/*
1619 	 * If this vdev just became a top-level vdev because its sibling was
1620 	 * detached, it will have adopted the parent's vdev guid -- but the
1621 	 * label may or may not be on disk yet. Fortunately, either version
1622 	 * of the label will have the same top guid, so if we're a top-level
1623 	 * vdev, we can safely compare to that instead.
1624 	 * However, if the config comes from a cachefile that failed to update
1625 	 * after the detach, a top-level vdev will appear as a non top-level
1626 	 * vdev in the config. Also relax the constraints if we perform an
1627 	 * extreme rewind.
1628 	 *
1629 	 * If we split this vdev off instead, then we also check the
1630 	 * original pool's guid. We don't want to consider the vdev
1631 	 * corrupt if it is partway through a split operation.
1632 	 */
1633 	if (vd->vdev_guid != guid && vd->vdev_guid != aux_guid) {
1634 		boolean_t mismatch = B_FALSE;
1635 		if (spa->spa_trust_config && !spa->spa_extreme_rewind) {
1636 			if (vd != vd->vdev_top || vd->vdev_guid != top_guid)
1637 				mismatch = B_TRUE;
1638 		} else {
1639 			if (vd->vdev_guid != top_guid &&
1640 			    vd->vdev_top->vdev_guid != guid)
1641 				mismatch = B_TRUE;
1642 		}
1643 
1644 		if (mismatch) {
1645 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1646 			    VDEV_AUX_CORRUPT_DATA);
1647 			nvlist_free(label);
1648 			vdev_dbgmsg(vd, "vdev_validate: config guid "
1649 			    "doesn't match label guid");
1650 			vdev_dbgmsg(vd, "CONFIG: guid %llu, top_guid %llu",
1651 			    (u_longlong_t)vd->vdev_guid,
1652 			    (u_longlong_t)vd->vdev_top->vdev_guid);
1653 			vdev_dbgmsg(vd, "LABEL: guid %llu, top_guid %llu, "
1654 			    "aux_guid %llu", (u_longlong_t)guid,
1655 			    (u_longlong_t)top_guid, (u_longlong_t)aux_guid);
1656 			return (0);
1657 		}
1658 	}
1659 
1660 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE,
1661 	    &state) != 0) {
1662 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
1663 		    VDEV_AUX_CORRUPT_DATA);
1664 		nvlist_free(label);
1665 		vdev_dbgmsg(vd, "vdev_validate: '%s' missing from label",
1666 		    ZPOOL_CONFIG_POOL_STATE);
1667 		return (0);
1668 	}
1669 
1670 	nvlist_free(label);
1671 
1672 	/*
1673 	 * If this is a verbatim import, no need to check the
1674 	 * state of the pool.
1675 	 */
1676 	if (!(spa->spa_import_flags & ZFS_IMPORT_VERBATIM) &&
1677 	    spa_load_state(spa) == SPA_LOAD_OPEN &&
1678 	    state != POOL_STATE_ACTIVE) {
1679 		vdev_dbgmsg(vd, "vdev_validate: invalid pool state (%llu) "
1680 		    "for spa %s", (u_longlong_t)state, spa->spa_name);
1681 		return (SET_ERROR(EBADF));
1682 	}
1683 
1684 	/*
1685 	 * If we were able to open and validate a vdev that was
1686 	 * previously marked permanently unavailable, clear that state
1687 	 * now.
1688 	 */
1689 	if (vd->vdev_not_present)
1690 		vd->vdev_not_present = 0;
1691 
1692 	return (0);
1693 }
1694 
1695 static void
1696 vdev_copy_path_impl(vdev_t *svd, vdev_t *dvd)
1697 {
1698 	if (svd->vdev_path != NULL && dvd->vdev_path != NULL) {
1699 		if (strcmp(svd->vdev_path, dvd->vdev_path) != 0) {
1700 			zfs_dbgmsg("vdev_copy_path: vdev %llu: path changed "
1701 			    "from '%s' to '%s'", (u_longlong_t)dvd->vdev_guid,
1702 			    dvd->vdev_path, svd->vdev_path);
1703 			spa_strfree(dvd->vdev_path);
1704 			dvd->vdev_path = spa_strdup(svd->vdev_path);
1705 		}
1706 	} else if (svd->vdev_path != NULL) {
1707 		dvd->vdev_path = spa_strdup(svd->vdev_path);
1708 		zfs_dbgmsg("vdev_copy_path: vdev %llu: path set to '%s'",
1709 		    (u_longlong_t)dvd->vdev_guid, dvd->vdev_path);
1710 	}
1711 }
1712 
1713 /*
1714  * Recursively copy vdev paths from one vdev to another. Source and destination
1715  * vdev trees must have same geometry otherwise return error. Intended to copy
1716  * paths from userland config into MOS config.
1717  */
1718 int
1719 vdev_copy_path_strict(vdev_t *svd, vdev_t *dvd)
1720 {
1721 	if ((svd->vdev_ops == &vdev_missing_ops) ||
1722 	    (svd->vdev_ishole && dvd->vdev_ishole) ||
1723 	    (dvd->vdev_ops == &vdev_indirect_ops))
1724 		return (0);
1725 
1726 	if (svd->vdev_ops != dvd->vdev_ops) {
1727 		vdev_dbgmsg(svd, "vdev_copy_path: vdev type mismatch: %s != %s",
1728 		    svd->vdev_ops->vdev_op_type, dvd->vdev_ops->vdev_op_type);
1729 		return (SET_ERROR(EINVAL));
1730 	}
1731 
1732 	if (svd->vdev_guid != dvd->vdev_guid) {
1733 		vdev_dbgmsg(svd, "vdev_copy_path: guids mismatch (%llu != "
1734 		    "%llu)", (u_longlong_t)svd->vdev_guid,
1735 		    (u_longlong_t)dvd->vdev_guid);
1736 		return (SET_ERROR(EINVAL));
1737 	}
1738 
1739 	if (svd->vdev_children != dvd->vdev_children) {
1740 		vdev_dbgmsg(svd, "vdev_copy_path: children count mismatch: "
1741 		    "%llu != %llu", (u_longlong_t)svd->vdev_children,
1742 		    (u_longlong_t)dvd->vdev_children);
1743 		return (SET_ERROR(EINVAL));
1744 	}
1745 
1746 	for (uint64_t i = 0; i < svd->vdev_children; i++) {
1747 		int error = vdev_copy_path_strict(svd->vdev_child[i],
1748 		    dvd->vdev_child[i]);
1749 		if (error != 0)
1750 			return (error);
1751 	}
1752 
1753 	if (svd->vdev_ops->vdev_op_leaf)
1754 		vdev_copy_path_impl(svd, dvd);
1755 
1756 	return (0);
1757 }
1758 
1759 static void
1760 vdev_copy_path_search(vdev_t *stvd, vdev_t *dvd)
1761 {
1762 	ASSERT(stvd->vdev_top == stvd);
1763 	ASSERT3U(stvd->vdev_id, ==, dvd->vdev_top->vdev_id);
1764 
1765 	for (uint64_t i = 0; i < dvd->vdev_children; i++) {
1766 		vdev_copy_path_search(stvd, dvd->vdev_child[i]);
1767 	}
1768 
1769 	if (!dvd->vdev_ops->vdev_op_leaf || !vdev_is_concrete(dvd))
1770 		return;
1771 
1772 	/*
1773 	 * The idea here is that while a vdev can shift positions within
1774 	 * a top vdev (when replacing, attaching mirror, etc.) it cannot
1775 	 * step outside of it.
1776 	 */
1777 	vdev_t *vd = vdev_lookup_by_guid(stvd, dvd->vdev_guid);
1778 
1779 	if (vd == NULL || vd->vdev_ops != dvd->vdev_ops)
1780 		return;
1781 
1782 	ASSERT(vd->vdev_ops->vdev_op_leaf);
1783 
1784 	vdev_copy_path_impl(vd, dvd);
1785 }
1786 
1787 /*
1788  * Recursively copy vdev paths from one root vdev to another. Source and
1789  * destination vdev trees may differ in geometry. For each destination leaf
1790  * vdev, search a vdev with the same guid and top vdev id in the source.
1791  * Intended to copy paths from userland config into MOS config.
1792  */
1793 void
1794 vdev_copy_path_relaxed(vdev_t *srvd, vdev_t *drvd)
1795 {
1796 	uint64_t children = MIN(srvd->vdev_children, drvd->vdev_children);
1797 	ASSERT(srvd->vdev_ops == &vdev_root_ops);
1798 	ASSERT(drvd->vdev_ops == &vdev_root_ops);
1799 
1800 	for (uint64_t i = 0; i < children; i++) {
1801 		vdev_copy_path_search(srvd->vdev_child[i],
1802 		    drvd->vdev_child[i]);
1803 	}
1804 }
1805 
1806 /*
1807  * Close a virtual device.
1808  */
1809 void
1810 vdev_close(vdev_t *vd)
1811 {
1812 	spa_t *spa = vd->vdev_spa;
1813 	vdev_t *pvd = vd->vdev_parent;
1814 
1815 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1816 
1817 	/*
1818 	 * If our parent is reopening, then we are as well, unless we are
1819 	 * going offline.
1820 	 */
1821 	if (pvd != NULL && pvd->vdev_reopening)
1822 		vd->vdev_reopening = (pvd->vdev_reopening && !vd->vdev_offline);
1823 
1824 	vd->vdev_ops->vdev_op_close(vd);
1825 
1826 	vdev_cache_purge(vd);
1827 
1828 	/*
1829 	 * We record the previous state before we close it, so that if we are
1830 	 * doing a reopen(), we don't generate FMA ereports if we notice that
1831 	 * it's still faulted.
1832 	 */
1833 	vd->vdev_prevstate = vd->vdev_state;
1834 
1835 	if (vd->vdev_offline)
1836 		vd->vdev_state = VDEV_STATE_OFFLINE;
1837 	else
1838 		vd->vdev_state = VDEV_STATE_CLOSED;
1839 	vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
1840 }
1841 
1842 void
1843 vdev_hold(vdev_t *vd)
1844 {
1845 	spa_t *spa = vd->vdev_spa;
1846 
1847 	ASSERT(spa_is_root(spa));
1848 	if (spa->spa_state == POOL_STATE_UNINITIALIZED)
1849 		return;
1850 
1851 	for (int c = 0; c < vd->vdev_children; c++)
1852 		vdev_hold(vd->vdev_child[c]);
1853 
1854 	if (vd->vdev_ops->vdev_op_leaf)
1855 		vd->vdev_ops->vdev_op_hold(vd);
1856 }
1857 
1858 void
1859 vdev_rele(vdev_t *vd)
1860 {
1861 	spa_t *spa = vd->vdev_spa;
1862 
1863 	ASSERT(spa_is_root(spa));
1864 	for (int c = 0; c < vd->vdev_children; c++)
1865 		vdev_rele(vd->vdev_child[c]);
1866 
1867 	if (vd->vdev_ops->vdev_op_leaf)
1868 		vd->vdev_ops->vdev_op_rele(vd);
1869 }
1870 
1871 /*
1872  * Reopen all interior vdevs and any unopened leaves.  We don't actually
1873  * reopen leaf vdevs which had previously been opened as they might deadlock
1874  * on the spa_config_lock.  Instead we only obtain the leaf's physical size.
1875  * If the leaf has never been opened then open it, as usual.
1876  */
1877 void
1878 vdev_reopen(vdev_t *vd)
1879 {
1880 	spa_t *spa = vd->vdev_spa;
1881 
1882 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
1883 
1884 	/* set the reopening flag unless we're taking the vdev offline */
1885 	vd->vdev_reopening = !vd->vdev_offline;
1886 	vdev_close(vd);
1887 	(void) vdev_open(vd);
1888 
1889 	/*
1890 	 * Call vdev_validate() here to make sure we have the same device.
1891 	 * Otherwise, a device with an invalid label could be successfully
1892 	 * opened in response to vdev_reopen().
1893 	 */
1894 	if (vd->vdev_aux) {
1895 		(void) vdev_validate_aux(vd);
1896 		if (vdev_readable(vd) && vdev_writeable(vd) &&
1897 		    vd->vdev_aux == &spa->spa_l2cache &&
1898 		    !l2arc_vdev_present(vd))
1899 			l2arc_add_vdev(spa, vd);
1900 	} else {
1901 		(void) vdev_validate(vd);
1902 	}
1903 
1904 	/*
1905 	 * Reassess parent vdev's health.
1906 	 */
1907 	vdev_propagate_state(vd);
1908 }
1909 
1910 int
1911 vdev_create(vdev_t *vd, uint64_t txg, boolean_t isreplacing)
1912 {
1913 	int error;
1914 
1915 	/*
1916 	 * Normally, partial opens (e.g. of a mirror) are allowed.
1917 	 * For a create, however, we want to fail the request if
1918 	 * there are any components we can't open.
1919 	 */
1920 	error = vdev_open(vd);
1921 
1922 	if (error || vd->vdev_state != VDEV_STATE_HEALTHY) {
1923 		vdev_close(vd);
1924 		return (error ? error : ENXIO);
1925 	}
1926 
1927 	/*
1928 	 * Recursively load DTLs and initialize all labels.
1929 	 */
1930 	if ((error = vdev_dtl_load(vd)) != 0 ||
1931 	    (error = vdev_label_init(vd, txg, isreplacing ?
1932 	    VDEV_LABEL_REPLACE : VDEV_LABEL_CREATE)) != 0) {
1933 		vdev_close(vd);
1934 		return (error);
1935 	}
1936 
1937 	return (0);
1938 }
1939 
1940 void
1941 vdev_metaslab_set_size(vdev_t *vd)
1942 {
1943 	/*
1944 	 * Aim for roughly metaslabs_per_vdev (default 200) metaslabs per vdev.
1945 	 */
1946 	vd->vdev_ms_shift = highbit64(vd->vdev_asize / metaslabs_per_vdev);
1947 	vd->vdev_ms_shift = MAX(vd->vdev_ms_shift, SPA_MAXBLOCKSHIFT);
1948 }
1949 
1950 void
1951 vdev_dirty(vdev_t *vd, int flags, void *arg, uint64_t txg)
1952 {
1953 	ASSERT(vd == vd->vdev_top);
1954 	/* indirect vdevs don't have metaslabs or dtls */
1955 	ASSERT(vdev_is_concrete(vd) || flags == 0);
1956 	ASSERT(ISP2(flags));
1957 	ASSERT(spa_writeable(vd->vdev_spa));
1958 
1959 	if (flags & VDD_METASLAB)
1960 		(void) txg_list_add(&vd->vdev_ms_list, arg, txg);
1961 
1962 	if (flags & VDD_DTL)
1963 		(void) txg_list_add(&vd->vdev_dtl_list, arg, txg);
1964 
1965 	(void) txg_list_add(&vd->vdev_spa->spa_vdev_txg_list, vd, txg);
1966 }
1967 
1968 void
1969 vdev_dirty_leaves(vdev_t *vd, int flags, uint64_t txg)
1970 {
1971 	for (int c = 0; c < vd->vdev_children; c++)
1972 		vdev_dirty_leaves(vd->vdev_child[c], flags, txg);
1973 
1974 	if (vd->vdev_ops->vdev_op_leaf)
1975 		vdev_dirty(vd->vdev_top, flags, vd, txg);
1976 }
1977 
1978 /*
1979  * DTLs.
1980  *
1981  * A vdev's DTL (dirty time log) is the set of transaction groups for which
1982  * the vdev has less than perfect replication.  There are four kinds of DTL:
1983  *
1984  * DTL_MISSING: txgs for which the vdev has no valid copies of the data
1985  *
1986  * DTL_PARTIAL: txgs for which data is available, but not fully replicated
1987  *
1988  * DTL_SCRUB: the txgs that could not be repaired by the last scrub; upon
1989  *	scrub completion, DTL_SCRUB replaces DTL_MISSING in the range of
1990  *	txgs that was scrubbed.
1991  *
1992  * DTL_OUTAGE: txgs which cannot currently be read, whether due to
1993  *	persistent errors or just some device being offline.
1994  *	Unlike the other three, the DTL_OUTAGE map is not generally
1995  *	maintained; it's only computed when needed, typically to
1996  *	determine whether a device can be detached.
1997  *
1998  * For leaf vdevs, DTL_MISSING and DTL_PARTIAL are identical: the device
1999  * either has the data or it doesn't.
2000  *
2001  * For interior vdevs such as mirror and RAID-Z the picture is more complex.
2002  * A vdev's DTL_PARTIAL is the union of its children's DTL_PARTIALs, because
2003  * if any child is less than fully replicated, then so is its parent.
2004  * A vdev's DTL_MISSING is a modified union of its children's DTL_MISSINGs,
2005  * comprising only those txgs which appear in 'maxfaults' or more children;
2006  * those are the txgs we don't have enough replication to read.  For example,
2007  * double-parity RAID-Z can tolerate up to two missing devices (maxfaults == 2);
2008  * thus, its DTL_MISSING consists of the set of txgs that appear in more than
2009  * two child DTL_MISSING maps.
2010  *
2011  * It should be clear from the above that to compute the DTLs and outage maps
2012  * for all vdevs, it suffices to know just the leaf vdevs' DTL_MISSING maps.
2013  * Therefore, that is all we keep on disk.  When loading the pool, or after
2014  * a configuration change, we generate all other DTLs from first principles.
2015  */
2016 void
2017 vdev_dtl_dirty(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2018 {
2019 	range_tree_t *rt = vd->vdev_dtl[t];
2020 
2021 	ASSERT(t < DTL_TYPES);
2022 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2023 	ASSERT(spa_writeable(vd->vdev_spa));
2024 
2025 	mutex_enter(&vd->vdev_dtl_lock);
2026 	if (!range_tree_contains(rt, txg, size))
2027 		range_tree_add(rt, txg, size);
2028 	mutex_exit(&vd->vdev_dtl_lock);
2029 }
2030 
2031 boolean_t
2032 vdev_dtl_contains(vdev_t *vd, vdev_dtl_type_t t, uint64_t txg, uint64_t size)
2033 {
2034 	range_tree_t *rt = vd->vdev_dtl[t];
2035 	boolean_t dirty = B_FALSE;
2036 
2037 	ASSERT(t < DTL_TYPES);
2038 	ASSERT(vd != vd->vdev_spa->spa_root_vdev);
2039 
2040 	/*
2041 	 * While we are loading the pool, the DTLs have not been loaded yet.
2042 	 * Ignore the DTLs and try all devices.  This avoids a recursive
2043 	 * mutex enter on the vdev_dtl_lock, and also makes us try hard
2044 	 * when loading the pool (relying on the checksum to ensure that
2045 	 * we get the right data -- note that we while loading, we are
2046 	 * only reading the MOS, which is always checksummed).
2047 	 */
2048 	if (vd->vdev_spa->spa_load_state != SPA_LOAD_NONE)
2049 		return (B_FALSE);
2050 
2051 	mutex_enter(&vd->vdev_dtl_lock);
2052 	if (range_tree_space(rt) != 0)
2053 		dirty = range_tree_contains(rt, txg, size);
2054 	mutex_exit(&vd->vdev_dtl_lock);
2055 
2056 	return (dirty);
2057 }
2058 
2059 boolean_t
2060 vdev_dtl_empty(vdev_t *vd, vdev_dtl_type_t t)
2061 {
2062 	range_tree_t *rt = vd->vdev_dtl[t];
2063 	boolean_t empty;
2064 
2065 	mutex_enter(&vd->vdev_dtl_lock);
2066 	empty = (range_tree_space(rt) == 0);
2067 	mutex_exit(&vd->vdev_dtl_lock);
2068 
2069 	return (empty);
2070 }
2071 
2072 /*
2073  * Returns the lowest txg in the DTL range.
2074  */
2075 static uint64_t
2076 vdev_dtl_min(vdev_t *vd)
2077 {
2078 	range_seg_t *rs;
2079 
2080 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2081 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2082 	ASSERT0(vd->vdev_children);
2083 
2084 	rs = avl_first(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2085 	return (rs->rs_start - 1);
2086 }
2087 
2088 /*
2089  * Returns the highest txg in the DTL.
2090  */
2091 static uint64_t
2092 vdev_dtl_max(vdev_t *vd)
2093 {
2094 	range_seg_t *rs;
2095 
2096 	ASSERT(MUTEX_HELD(&vd->vdev_dtl_lock));
2097 	ASSERT3U(range_tree_space(vd->vdev_dtl[DTL_MISSING]), !=, 0);
2098 	ASSERT0(vd->vdev_children);
2099 
2100 	rs = avl_last(&vd->vdev_dtl[DTL_MISSING]->rt_root);
2101 	return (rs->rs_end);
2102 }
2103 
2104 /*
2105  * Determine if a resilvering vdev should remove any DTL entries from
2106  * its range. If the vdev was resilvering for the entire duration of the
2107  * scan then it should excise that range from its DTLs. Otherwise, this
2108  * vdev is considered partially resilvered and should leave its DTL
2109  * entries intact. The comment in vdev_dtl_reassess() describes how we
2110  * excise the DTLs.
2111  */
2112 static boolean_t
2113 vdev_dtl_should_excise(vdev_t *vd)
2114 {
2115 	spa_t *spa = vd->vdev_spa;
2116 	dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2117 
2118 	ASSERT0(scn->scn_phys.scn_errors);
2119 	ASSERT0(vd->vdev_children);
2120 
2121 	if (vd->vdev_state < VDEV_STATE_DEGRADED)
2122 		return (B_FALSE);
2123 
2124 	if (vd->vdev_resilver_txg == 0 ||
2125 	    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0)
2126 		return (B_TRUE);
2127 
2128 	/*
2129 	 * When a resilver is initiated the scan will assign the scn_max_txg
2130 	 * value to the highest txg value that exists in all DTLs. If this
2131 	 * device's max DTL is not part of this scan (i.e. it is not in
2132 	 * the range (scn_min_txg, scn_max_txg] then it is not eligible
2133 	 * for excision.
2134 	 */
2135 	if (vdev_dtl_max(vd) <= scn->scn_phys.scn_max_txg) {
2136 		ASSERT3U(scn->scn_phys.scn_min_txg, <=, vdev_dtl_min(vd));
2137 		ASSERT3U(scn->scn_phys.scn_min_txg, <, vd->vdev_resilver_txg);
2138 		ASSERT3U(vd->vdev_resilver_txg, <=, scn->scn_phys.scn_max_txg);
2139 		return (B_TRUE);
2140 	}
2141 	return (B_FALSE);
2142 }
2143 
2144 /*
2145  * Reassess DTLs after a config change or scrub completion.
2146  */
2147 void
2148 vdev_dtl_reassess(vdev_t *vd, uint64_t txg, uint64_t scrub_txg, int scrub_done)
2149 {
2150 	spa_t *spa = vd->vdev_spa;
2151 	avl_tree_t reftree;
2152 	int minref;
2153 
2154 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
2155 
2156 	for (int c = 0; c < vd->vdev_children; c++)
2157 		vdev_dtl_reassess(vd->vdev_child[c], txg,
2158 		    scrub_txg, scrub_done);
2159 
2160 	if (vd == spa->spa_root_vdev || !vdev_is_concrete(vd) || vd->vdev_aux)
2161 		return;
2162 
2163 	if (vd->vdev_ops->vdev_op_leaf) {
2164 		dsl_scan_t *scn = spa->spa_dsl_pool->dp_scan;
2165 
2166 		mutex_enter(&vd->vdev_dtl_lock);
2167 
2168 		/*
2169 		 * If we've completed a scan cleanly then determine
2170 		 * if this vdev should remove any DTLs. We only want to
2171 		 * excise regions on vdevs that were available during
2172 		 * the entire duration of this scan.
2173 		 */
2174 		if (scrub_txg != 0 &&
2175 		    (spa->spa_scrub_started ||
2176 		    (scn != NULL && scn->scn_phys.scn_errors == 0)) &&
2177 		    vdev_dtl_should_excise(vd)) {
2178 			/*
2179 			 * We completed a scrub up to scrub_txg.  If we
2180 			 * did it without rebooting, then the scrub dtl
2181 			 * will be valid, so excise the old region and
2182 			 * fold in the scrub dtl.  Otherwise, leave the
2183 			 * dtl as-is if there was an error.
2184 			 *
2185 			 * There's little trick here: to excise the beginning
2186 			 * of the DTL_MISSING map, we put it into a reference
2187 			 * tree and then add a segment with refcnt -1 that
2188 			 * covers the range [0, scrub_txg).  This means
2189 			 * that each txg in that range has refcnt -1 or 0.
2190 			 * We then add DTL_SCRUB with a refcnt of 2, so that
2191 			 * entries in the range [0, scrub_txg) will have a
2192 			 * positive refcnt -- either 1 or 2.  We then convert
2193 			 * the reference tree into the new DTL_MISSING map.
2194 			 */
2195 			space_reftree_create(&reftree);
2196 			space_reftree_add_map(&reftree,
2197 			    vd->vdev_dtl[DTL_MISSING], 1);
2198 			space_reftree_add_seg(&reftree, 0, scrub_txg, -1);
2199 			space_reftree_add_map(&reftree,
2200 			    vd->vdev_dtl[DTL_SCRUB], 2);
2201 			space_reftree_generate_map(&reftree,
2202 			    vd->vdev_dtl[DTL_MISSING], 1);
2203 			space_reftree_destroy(&reftree);
2204 		}
2205 		range_tree_vacate(vd->vdev_dtl[DTL_PARTIAL], NULL, NULL);
2206 		range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2207 		    range_tree_add, vd->vdev_dtl[DTL_PARTIAL]);
2208 		if (scrub_done)
2209 			range_tree_vacate(vd->vdev_dtl[DTL_SCRUB], NULL, NULL);
2210 		range_tree_vacate(vd->vdev_dtl[DTL_OUTAGE], NULL, NULL);
2211 		if (!vdev_readable(vd))
2212 			range_tree_add(vd->vdev_dtl[DTL_OUTAGE], 0, -1ULL);
2213 		else
2214 			range_tree_walk(vd->vdev_dtl[DTL_MISSING],
2215 			    range_tree_add, vd->vdev_dtl[DTL_OUTAGE]);
2216 
2217 		/*
2218 		 * If the vdev was resilvering and no longer has any
2219 		 * DTLs then reset its resilvering flag.
2220 		 */
2221 		if (vd->vdev_resilver_txg != 0 &&
2222 		    range_tree_space(vd->vdev_dtl[DTL_MISSING]) == 0 &&
2223 		    range_tree_space(vd->vdev_dtl[DTL_OUTAGE]) == 0)
2224 			vd->vdev_resilver_txg = 0;
2225 
2226 		mutex_exit(&vd->vdev_dtl_lock);
2227 
2228 		if (txg != 0)
2229 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, txg);
2230 		return;
2231 	}
2232 
2233 	mutex_enter(&vd->vdev_dtl_lock);
2234 	for (int t = 0; t < DTL_TYPES; t++) {
2235 		/* account for child's outage in parent's missing map */
2236 		int s = (t == DTL_MISSING) ? DTL_OUTAGE: t;
2237 		if (t == DTL_SCRUB)
2238 			continue;			/* leaf vdevs only */
2239 		if (t == DTL_PARTIAL)
2240 			minref = 1;			/* i.e. non-zero */
2241 		else if (vd->vdev_nparity != 0)
2242 			minref = vd->vdev_nparity + 1;	/* RAID-Z */
2243 		else
2244 			minref = vd->vdev_children;	/* any kind of mirror */
2245 		space_reftree_create(&reftree);
2246 		for (int c = 0; c < vd->vdev_children; c++) {
2247 			vdev_t *cvd = vd->vdev_child[c];
2248 			mutex_enter(&cvd->vdev_dtl_lock);
2249 			space_reftree_add_map(&reftree, cvd->vdev_dtl[s], 1);
2250 			mutex_exit(&cvd->vdev_dtl_lock);
2251 		}
2252 		space_reftree_generate_map(&reftree, vd->vdev_dtl[t], minref);
2253 		space_reftree_destroy(&reftree);
2254 	}
2255 	mutex_exit(&vd->vdev_dtl_lock);
2256 }
2257 
2258 int
2259 vdev_dtl_load(vdev_t *vd)
2260 {
2261 	spa_t *spa = vd->vdev_spa;
2262 	objset_t *mos = spa->spa_meta_objset;
2263 	int error = 0;
2264 
2265 	if (vd->vdev_ops->vdev_op_leaf && vd->vdev_dtl_object != 0) {
2266 		ASSERT(vdev_is_concrete(vd));
2267 
2268 		error = space_map_open(&vd->vdev_dtl_sm, mos,
2269 		    vd->vdev_dtl_object, 0, -1ULL, 0);
2270 		if (error)
2271 			return (error);
2272 		ASSERT(vd->vdev_dtl_sm != NULL);
2273 
2274 		mutex_enter(&vd->vdev_dtl_lock);
2275 
2276 		/*
2277 		 * Now that we've opened the space_map we need to update
2278 		 * the in-core DTL.
2279 		 */
2280 		space_map_update(vd->vdev_dtl_sm);
2281 
2282 		error = space_map_load(vd->vdev_dtl_sm,
2283 		    vd->vdev_dtl[DTL_MISSING], SM_ALLOC);
2284 		mutex_exit(&vd->vdev_dtl_lock);
2285 
2286 		return (error);
2287 	}
2288 
2289 	for (int c = 0; c < vd->vdev_children; c++) {
2290 		error = vdev_dtl_load(vd->vdev_child[c]);
2291 		if (error != 0)
2292 			break;
2293 	}
2294 
2295 	return (error);
2296 }
2297 
2298 void
2299 vdev_destroy_unlink_zap(vdev_t *vd, uint64_t zapobj, dmu_tx_t *tx)
2300 {
2301 	spa_t *spa = vd->vdev_spa;
2302 
2303 	VERIFY0(zap_destroy(spa->spa_meta_objset, zapobj, tx));
2304 	VERIFY0(zap_remove_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2305 	    zapobj, tx));
2306 }
2307 
2308 uint64_t
2309 vdev_create_link_zap(vdev_t *vd, dmu_tx_t *tx)
2310 {
2311 	spa_t *spa = vd->vdev_spa;
2312 	uint64_t zap = zap_create(spa->spa_meta_objset, DMU_OTN_ZAP_METADATA,
2313 	    DMU_OT_NONE, 0, tx);
2314 
2315 	ASSERT(zap != 0);
2316 	VERIFY0(zap_add_int(spa->spa_meta_objset, spa->spa_all_vdev_zaps,
2317 	    zap, tx));
2318 
2319 	return (zap);
2320 }
2321 
2322 void
2323 vdev_construct_zaps(vdev_t *vd, dmu_tx_t *tx)
2324 {
2325 	if (vd->vdev_ops != &vdev_hole_ops &&
2326 	    vd->vdev_ops != &vdev_missing_ops &&
2327 	    vd->vdev_ops != &vdev_root_ops &&
2328 	    !vd->vdev_top->vdev_removing) {
2329 		if (vd->vdev_ops->vdev_op_leaf && vd->vdev_leaf_zap == 0) {
2330 			vd->vdev_leaf_zap = vdev_create_link_zap(vd, tx);
2331 		}
2332 		if (vd == vd->vdev_top && vd->vdev_top_zap == 0) {
2333 			vd->vdev_top_zap = vdev_create_link_zap(vd, tx);
2334 		}
2335 	}
2336 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
2337 		vdev_construct_zaps(vd->vdev_child[i], tx);
2338 	}
2339 }
2340 
2341 void
2342 vdev_dtl_sync(vdev_t *vd, uint64_t txg)
2343 {
2344 	spa_t *spa = vd->vdev_spa;
2345 	range_tree_t *rt = vd->vdev_dtl[DTL_MISSING];
2346 	objset_t *mos = spa->spa_meta_objset;
2347 	range_tree_t *rtsync;
2348 	dmu_tx_t *tx;
2349 	uint64_t object = space_map_object(vd->vdev_dtl_sm);
2350 
2351 	ASSERT(vdev_is_concrete(vd));
2352 	ASSERT(vd->vdev_ops->vdev_op_leaf);
2353 
2354 	tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2355 
2356 	if (vd->vdev_detached || vd->vdev_top->vdev_removing) {
2357 		mutex_enter(&vd->vdev_dtl_lock);
2358 		space_map_free(vd->vdev_dtl_sm, tx);
2359 		space_map_close(vd->vdev_dtl_sm);
2360 		vd->vdev_dtl_sm = NULL;
2361 		mutex_exit(&vd->vdev_dtl_lock);
2362 
2363 		/*
2364 		 * We only destroy the leaf ZAP for detached leaves or for
2365 		 * removed log devices. Removed data devices handle leaf ZAP
2366 		 * cleanup later, once cancellation is no longer possible.
2367 		 */
2368 		if (vd->vdev_leaf_zap != 0 && (vd->vdev_detached ||
2369 		    vd->vdev_top->vdev_islog)) {
2370 			vdev_destroy_unlink_zap(vd, vd->vdev_leaf_zap, tx);
2371 			vd->vdev_leaf_zap = 0;
2372 		}
2373 
2374 		dmu_tx_commit(tx);
2375 		return;
2376 	}
2377 
2378 	if (vd->vdev_dtl_sm == NULL) {
2379 		uint64_t new_object;
2380 
2381 		new_object = space_map_alloc(mos, tx);
2382 		VERIFY3U(new_object, !=, 0);
2383 
2384 		VERIFY0(space_map_open(&vd->vdev_dtl_sm, mos, new_object,
2385 		    0, -1ULL, 0));
2386 		ASSERT(vd->vdev_dtl_sm != NULL);
2387 	}
2388 
2389 	rtsync = range_tree_create(NULL, NULL);
2390 
2391 	mutex_enter(&vd->vdev_dtl_lock);
2392 	range_tree_walk(rt, range_tree_add, rtsync);
2393 	mutex_exit(&vd->vdev_dtl_lock);
2394 
2395 	space_map_truncate(vd->vdev_dtl_sm, tx);
2396 	space_map_write(vd->vdev_dtl_sm, rtsync, SM_ALLOC, tx);
2397 	range_tree_vacate(rtsync, NULL, NULL);
2398 
2399 	range_tree_destroy(rtsync);
2400 
2401 	/*
2402 	 * If the object for the space map has changed then dirty
2403 	 * the top level so that we update the config.
2404 	 */
2405 	if (object != space_map_object(vd->vdev_dtl_sm)) {
2406 		vdev_dbgmsg(vd, "txg %llu, spa %s, DTL old object %llu, "
2407 		    "new object %llu", (u_longlong_t)txg, spa_name(spa),
2408 		    (u_longlong_t)object,
2409 		    (u_longlong_t)space_map_object(vd->vdev_dtl_sm));
2410 		vdev_config_dirty(vd->vdev_top);
2411 	}
2412 
2413 	dmu_tx_commit(tx);
2414 
2415 	mutex_enter(&vd->vdev_dtl_lock);
2416 	space_map_update(vd->vdev_dtl_sm);
2417 	mutex_exit(&vd->vdev_dtl_lock);
2418 }
2419 
2420 /*
2421  * Determine whether the specified vdev can be offlined/detached/removed
2422  * without losing data.
2423  */
2424 boolean_t
2425 vdev_dtl_required(vdev_t *vd)
2426 {
2427 	spa_t *spa = vd->vdev_spa;
2428 	vdev_t *tvd = vd->vdev_top;
2429 	uint8_t cant_read = vd->vdev_cant_read;
2430 	boolean_t required;
2431 
2432 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
2433 
2434 	if (vd == spa->spa_root_vdev || vd == tvd)
2435 		return (B_TRUE);
2436 
2437 	/*
2438 	 * Temporarily mark the device as unreadable, and then determine
2439 	 * whether this results in any DTL outages in the top-level vdev.
2440 	 * If not, we can safely offline/detach/remove the device.
2441 	 */
2442 	vd->vdev_cant_read = B_TRUE;
2443 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2444 	required = !vdev_dtl_empty(tvd, DTL_OUTAGE);
2445 	vd->vdev_cant_read = cant_read;
2446 	vdev_dtl_reassess(tvd, 0, 0, B_FALSE);
2447 
2448 	if (!required && zio_injection_enabled)
2449 		required = !!zio_handle_device_injection(vd, NULL, ECHILD);
2450 
2451 	return (required);
2452 }
2453 
2454 /*
2455  * Determine if resilver is needed, and if so the txg range.
2456  */
2457 boolean_t
2458 vdev_resilver_needed(vdev_t *vd, uint64_t *minp, uint64_t *maxp)
2459 {
2460 	boolean_t needed = B_FALSE;
2461 	uint64_t thismin = UINT64_MAX;
2462 	uint64_t thismax = 0;
2463 
2464 	if (vd->vdev_children == 0) {
2465 		mutex_enter(&vd->vdev_dtl_lock);
2466 		if (range_tree_space(vd->vdev_dtl[DTL_MISSING]) != 0 &&
2467 		    vdev_writeable(vd)) {
2468 
2469 			thismin = vdev_dtl_min(vd);
2470 			thismax = vdev_dtl_max(vd);
2471 			needed = B_TRUE;
2472 		}
2473 		mutex_exit(&vd->vdev_dtl_lock);
2474 	} else {
2475 		for (int c = 0; c < vd->vdev_children; c++) {
2476 			vdev_t *cvd = vd->vdev_child[c];
2477 			uint64_t cmin, cmax;
2478 
2479 			if (vdev_resilver_needed(cvd, &cmin, &cmax)) {
2480 				thismin = MIN(thismin, cmin);
2481 				thismax = MAX(thismax, cmax);
2482 				needed = B_TRUE;
2483 			}
2484 		}
2485 	}
2486 
2487 	if (needed && minp) {
2488 		*minp = thismin;
2489 		*maxp = thismax;
2490 	}
2491 	return (needed);
2492 }
2493 
2494 int
2495 vdev_load(vdev_t *vd)
2496 {
2497 	int error = 0;
2498 	/*
2499 	 * Recursively load all children.
2500 	 */
2501 	for (int c = 0; c < vd->vdev_children; c++) {
2502 		error = vdev_load(vd->vdev_child[c]);
2503 		if (error != 0) {
2504 			return (error);
2505 		}
2506 	}
2507 
2508 	vdev_set_deflate_ratio(vd);
2509 
2510 	/*
2511 	 * If this is a top-level vdev, initialize its metaslabs.
2512 	 */
2513 	if (vd == vd->vdev_top && vdev_is_concrete(vd)) {
2514 		if (vd->vdev_ashift == 0 || vd->vdev_asize == 0) {
2515 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2516 			    VDEV_AUX_CORRUPT_DATA);
2517 			vdev_dbgmsg(vd, "vdev_load: invalid size. ashift=%llu, "
2518 			    "asize=%llu", (u_longlong_t)vd->vdev_ashift,
2519 			    (u_longlong_t)vd->vdev_asize);
2520 			return (SET_ERROR(ENXIO));
2521 		} else if ((error = vdev_metaslab_init(vd, 0)) != 0) {
2522 			vdev_dbgmsg(vd, "vdev_load: metaslab_init failed "
2523 			    "[error=%d]", error);
2524 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2525 			    VDEV_AUX_CORRUPT_DATA);
2526 			return (error);
2527 		}
2528 	}
2529 
2530 	/*
2531 	 * If this is a leaf vdev, load its DTL.
2532 	 */
2533 	if (vd->vdev_ops->vdev_op_leaf && (error = vdev_dtl_load(vd)) != 0) {
2534 		vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2535 		    VDEV_AUX_CORRUPT_DATA);
2536 		vdev_dbgmsg(vd, "vdev_load: vdev_dtl_load failed "
2537 		    "[error=%d]", error);
2538 		return (error);
2539 	}
2540 
2541 	uint64_t obsolete_sm_object = vdev_obsolete_sm_object(vd);
2542 	if (obsolete_sm_object != 0) {
2543 		objset_t *mos = vd->vdev_spa->spa_meta_objset;
2544 		ASSERT(vd->vdev_asize != 0);
2545 		ASSERT(vd->vdev_obsolete_sm == NULL);
2546 
2547 		if ((error = space_map_open(&vd->vdev_obsolete_sm, mos,
2548 		    obsolete_sm_object, 0, vd->vdev_asize, 0))) {
2549 			vdev_set_state(vd, B_FALSE, VDEV_STATE_CANT_OPEN,
2550 			    VDEV_AUX_CORRUPT_DATA);
2551 			vdev_dbgmsg(vd, "vdev_load: space_map_open failed for "
2552 			    "obsolete spacemap (obj %llu) [error=%d]",
2553 			    (u_longlong_t)obsolete_sm_object, error);
2554 			return (error);
2555 		}
2556 		space_map_update(vd->vdev_obsolete_sm);
2557 	}
2558 
2559 	return (0);
2560 }
2561 
2562 /*
2563  * The special vdev case is used for hot spares and l2cache devices.  Its
2564  * sole purpose it to set the vdev state for the associated vdev.  To do this,
2565  * we make sure that we can open the underlying device, then try to read the
2566  * label, and make sure that the label is sane and that it hasn't been
2567  * repurposed to another pool.
2568  */
2569 int
2570 vdev_validate_aux(vdev_t *vd)
2571 {
2572 	nvlist_t *label;
2573 	uint64_t guid, version;
2574 	uint64_t state;
2575 
2576 	if (!vdev_readable(vd))
2577 		return (0);
2578 
2579 	if ((label = vdev_label_read_config(vd, -1ULL)) == NULL) {
2580 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2581 		    VDEV_AUX_CORRUPT_DATA);
2582 		return (-1);
2583 	}
2584 
2585 	if (nvlist_lookup_uint64(label, ZPOOL_CONFIG_VERSION, &version) != 0 ||
2586 	    !SPA_VERSION_IS_SUPPORTED(version) ||
2587 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_GUID, &guid) != 0 ||
2588 	    guid != vd->vdev_guid ||
2589 	    nvlist_lookup_uint64(label, ZPOOL_CONFIG_POOL_STATE, &state) != 0) {
2590 		vdev_set_state(vd, B_TRUE, VDEV_STATE_CANT_OPEN,
2591 		    VDEV_AUX_CORRUPT_DATA);
2592 		nvlist_free(label);
2593 		return (-1);
2594 	}
2595 
2596 	/*
2597 	 * We don't actually check the pool state here.  If it's in fact in
2598 	 * use by another pool, we update this fact on the fly when requested.
2599 	 */
2600 	nvlist_free(label);
2601 	return (0);
2602 }
2603 
2604 /*
2605  * Free the objects used to store this vdev's spacemaps, and the array
2606  * that points to them.
2607  */
2608 void
2609 vdev_destroy_spacemaps(vdev_t *vd, dmu_tx_t *tx)
2610 {
2611 	if (vd->vdev_ms_array == 0)
2612 		return;
2613 
2614 	objset_t *mos = vd->vdev_spa->spa_meta_objset;
2615 	uint64_t array_count = vd->vdev_asize >> vd->vdev_ms_shift;
2616 	size_t array_bytes = array_count * sizeof (uint64_t);
2617 	uint64_t *smobj_array = kmem_alloc(array_bytes, KM_SLEEP);
2618 	VERIFY0(dmu_read(mos, vd->vdev_ms_array, 0,
2619 	    array_bytes, smobj_array, 0));
2620 
2621 	for (uint64_t i = 0; i < array_count; i++) {
2622 		uint64_t smobj = smobj_array[i];
2623 		if (smobj == 0)
2624 			continue;
2625 
2626 		space_map_free_obj(mos, smobj, tx);
2627 	}
2628 
2629 	kmem_free(smobj_array, array_bytes);
2630 	VERIFY0(dmu_object_free(mos, vd->vdev_ms_array, tx));
2631 	vd->vdev_ms_array = 0;
2632 }
2633 
2634 static void
2635 vdev_remove_empty(vdev_t *vd, uint64_t txg)
2636 {
2637 	spa_t *spa = vd->vdev_spa;
2638 	dmu_tx_t *tx;
2639 
2640 	ASSERT(vd == vd->vdev_top);
2641 	ASSERT3U(txg, ==, spa_syncing_txg(spa));
2642 
2643 	if (vd->vdev_ms != NULL) {
2644 		metaslab_group_t *mg = vd->vdev_mg;
2645 
2646 		metaslab_group_histogram_verify(mg);
2647 		metaslab_class_histogram_verify(mg->mg_class);
2648 
2649 		for (int m = 0; m < vd->vdev_ms_count; m++) {
2650 			metaslab_t *msp = vd->vdev_ms[m];
2651 
2652 			if (msp == NULL || msp->ms_sm == NULL)
2653 				continue;
2654 
2655 			mutex_enter(&msp->ms_lock);
2656 			/*
2657 			 * If the metaslab was not loaded when the vdev
2658 			 * was removed then the histogram accounting may
2659 			 * not be accurate. Update the histogram information
2660 			 * here so that we ensure that the metaslab group
2661 			 * and metaslab class are up-to-date.
2662 			 */
2663 			metaslab_group_histogram_remove(mg, msp);
2664 
2665 			VERIFY0(space_map_allocated(msp->ms_sm));
2666 			space_map_close(msp->ms_sm);
2667 			msp->ms_sm = NULL;
2668 			mutex_exit(&msp->ms_lock);
2669 		}
2670 
2671 		metaslab_group_histogram_verify(mg);
2672 		metaslab_class_histogram_verify(mg->mg_class);
2673 		for (int i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
2674 			ASSERT0(mg->mg_histogram[i]);
2675 	}
2676 
2677 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2678 	vdev_destroy_spacemaps(vd, tx);
2679 
2680 	if (vd->vdev_islog && vd->vdev_top_zap != 0) {
2681 		vdev_destroy_unlink_zap(vd, vd->vdev_top_zap, tx);
2682 		vd->vdev_top_zap = 0;
2683 	}
2684 	dmu_tx_commit(tx);
2685 }
2686 
2687 void
2688 vdev_sync_done(vdev_t *vd, uint64_t txg)
2689 {
2690 	metaslab_t *msp;
2691 	boolean_t reassess = !txg_list_empty(&vd->vdev_ms_list, TXG_CLEAN(txg));
2692 
2693 	ASSERT(vdev_is_concrete(vd));
2694 
2695 	while (msp = txg_list_remove(&vd->vdev_ms_list, TXG_CLEAN(txg)))
2696 		metaslab_sync_done(msp, txg);
2697 
2698 	if (reassess)
2699 		metaslab_sync_reassess(vd->vdev_mg);
2700 }
2701 
2702 void
2703 vdev_sync(vdev_t *vd, uint64_t txg)
2704 {
2705 	spa_t *spa = vd->vdev_spa;
2706 	vdev_t *lvd;
2707 	metaslab_t *msp;
2708 	dmu_tx_t *tx;
2709 
2710 	if (range_tree_space(vd->vdev_obsolete_segments) > 0) {
2711 		dmu_tx_t *tx;
2712 
2713 		ASSERT(vd->vdev_removing ||
2714 		    vd->vdev_ops == &vdev_indirect_ops);
2715 
2716 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2717 		vdev_indirect_sync_obsolete(vd, tx);
2718 		dmu_tx_commit(tx);
2719 
2720 		/*
2721 		 * If the vdev is indirect, it can't have dirty
2722 		 * metaslabs or DTLs.
2723 		 */
2724 		if (vd->vdev_ops == &vdev_indirect_ops) {
2725 			ASSERT(txg_list_empty(&vd->vdev_ms_list, txg));
2726 			ASSERT(txg_list_empty(&vd->vdev_dtl_list, txg));
2727 			return;
2728 		}
2729 	}
2730 
2731 	ASSERT(vdev_is_concrete(vd));
2732 
2733 	if (vd->vdev_ms_array == 0 && vd->vdev_ms_shift != 0 &&
2734 	    !vd->vdev_removing) {
2735 		ASSERT(vd == vd->vdev_top);
2736 		ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
2737 		tx = dmu_tx_create_assigned(spa->spa_dsl_pool, txg);
2738 		vd->vdev_ms_array = dmu_object_alloc(spa->spa_meta_objset,
2739 		    DMU_OT_OBJECT_ARRAY, 0, DMU_OT_NONE, 0, tx);
2740 		ASSERT(vd->vdev_ms_array != 0);
2741 		vdev_config_dirty(vd);
2742 		dmu_tx_commit(tx);
2743 	}
2744 
2745 	while ((msp = txg_list_remove(&vd->vdev_ms_list, txg)) != NULL) {
2746 		metaslab_sync(msp, txg);
2747 		(void) txg_list_add(&vd->vdev_ms_list, msp, TXG_CLEAN(txg));
2748 	}
2749 
2750 	while ((lvd = txg_list_remove(&vd->vdev_dtl_list, txg)) != NULL)
2751 		vdev_dtl_sync(lvd, txg);
2752 
2753 	/*
2754 	 * Remove the metadata associated with this vdev once it's empty.
2755 	 * Note that this is typically used for log/cache device removal;
2756 	 * we don't empty toplevel vdevs when removing them.  But if
2757 	 * a toplevel happens to be emptied, this is not harmful.
2758 	 */
2759 	if (vd->vdev_stat.vs_alloc == 0 && vd->vdev_removing) {
2760 		vdev_remove_empty(vd, txg);
2761 	}
2762 
2763 	(void) txg_list_add(&spa->spa_vdev_txg_list, vd, TXG_CLEAN(txg));
2764 }
2765 
2766 uint64_t
2767 vdev_psize_to_asize(vdev_t *vd, uint64_t psize)
2768 {
2769 	return (vd->vdev_ops->vdev_op_asize(vd, psize));
2770 }
2771 
2772 /*
2773  * Mark the given vdev faulted.  A faulted vdev behaves as if the device could
2774  * not be opened, and no I/O is attempted.
2775  */
2776 int
2777 vdev_fault(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2778 {
2779 	vdev_t *vd, *tvd;
2780 
2781 	spa_vdev_state_enter(spa, SCL_NONE);
2782 
2783 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2784 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2785 
2786 	if (!vd->vdev_ops->vdev_op_leaf)
2787 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2788 
2789 	tvd = vd->vdev_top;
2790 
2791 	/*
2792 	 * We don't directly use the aux state here, but if we do a
2793 	 * vdev_reopen(), we need this value to be present to remember why we
2794 	 * were faulted.
2795 	 */
2796 	vd->vdev_label_aux = aux;
2797 
2798 	/*
2799 	 * Faulted state takes precedence over degraded.
2800 	 */
2801 	vd->vdev_delayed_close = B_FALSE;
2802 	vd->vdev_faulted = 1ULL;
2803 	vd->vdev_degraded = 0ULL;
2804 	vdev_set_state(vd, B_FALSE, VDEV_STATE_FAULTED, aux);
2805 
2806 	/*
2807 	 * If this device has the only valid copy of the data, then
2808 	 * back off and simply mark the vdev as degraded instead.
2809 	 */
2810 	if (!tvd->vdev_islog && vd->vdev_aux == NULL && vdev_dtl_required(vd)) {
2811 		vd->vdev_degraded = 1ULL;
2812 		vd->vdev_faulted = 0ULL;
2813 
2814 		/*
2815 		 * If we reopen the device and it's not dead, only then do we
2816 		 * mark it degraded.
2817 		 */
2818 		vdev_reopen(tvd);
2819 
2820 		if (vdev_readable(vd))
2821 			vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED, aux);
2822 	}
2823 
2824 	return (spa_vdev_state_exit(spa, vd, 0));
2825 }
2826 
2827 /*
2828  * Mark the given vdev degraded.  A degraded vdev is purely an indication to the
2829  * user that something is wrong.  The vdev continues to operate as normal as far
2830  * as I/O is concerned.
2831  */
2832 int
2833 vdev_degrade(spa_t *spa, uint64_t guid, vdev_aux_t aux)
2834 {
2835 	vdev_t *vd;
2836 
2837 	spa_vdev_state_enter(spa, SCL_NONE);
2838 
2839 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2840 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2841 
2842 	if (!vd->vdev_ops->vdev_op_leaf)
2843 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2844 
2845 	/*
2846 	 * If the vdev is already faulted, then don't do anything.
2847 	 */
2848 	if (vd->vdev_faulted || vd->vdev_degraded)
2849 		return (spa_vdev_state_exit(spa, NULL, 0));
2850 
2851 	vd->vdev_degraded = 1ULL;
2852 	if (!vdev_is_dead(vd))
2853 		vdev_set_state(vd, B_FALSE, VDEV_STATE_DEGRADED,
2854 		    aux);
2855 
2856 	return (spa_vdev_state_exit(spa, vd, 0));
2857 }
2858 
2859 /*
2860  * Online the given vdev.
2861  *
2862  * If 'ZFS_ONLINE_UNSPARE' is set, it implies two things.  First, any attached
2863  * spare device should be detached when the device finishes resilvering.
2864  * Second, the online should be treated like a 'test' online case, so no FMA
2865  * events are generated if the device fails to open.
2866  */
2867 int
2868 vdev_online(spa_t *spa, uint64_t guid, uint64_t flags, vdev_state_t *newstate)
2869 {
2870 	vdev_t *vd, *tvd, *pvd, *rvd = spa->spa_root_vdev;
2871 	boolean_t wasoffline;
2872 	vdev_state_t oldstate;
2873 
2874 	spa_vdev_state_enter(spa, SCL_NONE);
2875 
2876 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2877 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2878 
2879 	if (!vd->vdev_ops->vdev_op_leaf)
2880 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2881 
2882 	wasoffline = (vd->vdev_offline || vd->vdev_tmpoffline);
2883 	oldstate = vd->vdev_state;
2884 
2885 	tvd = vd->vdev_top;
2886 	vd->vdev_offline = B_FALSE;
2887 	vd->vdev_tmpoffline = B_FALSE;
2888 	vd->vdev_checkremove = !!(flags & ZFS_ONLINE_CHECKREMOVE);
2889 	vd->vdev_forcefault = !!(flags & ZFS_ONLINE_FORCEFAULT);
2890 
2891 	/* XXX - L2ARC 1.0 does not support expansion */
2892 	if (!vd->vdev_aux) {
2893 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2894 			pvd->vdev_expanding = !!(flags & ZFS_ONLINE_EXPAND);
2895 	}
2896 
2897 	vdev_reopen(tvd);
2898 	vd->vdev_checkremove = vd->vdev_forcefault = B_FALSE;
2899 
2900 	if (!vd->vdev_aux) {
2901 		for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
2902 			pvd->vdev_expanding = B_FALSE;
2903 	}
2904 
2905 	if (newstate)
2906 		*newstate = vd->vdev_state;
2907 	if ((flags & ZFS_ONLINE_UNSPARE) &&
2908 	    !vdev_is_dead(vd) && vd->vdev_parent &&
2909 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
2910 	    vd->vdev_parent->vdev_child[0] == vd)
2911 		vd->vdev_unspare = B_TRUE;
2912 
2913 	if ((flags & ZFS_ONLINE_EXPAND) || spa->spa_autoexpand) {
2914 
2915 		/* XXX - L2ARC 1.0 does not support expansion */
2916 		if (vd->vdev_aux)
2917 			return (spa_vdev_state_exit(spa, vd, ENOTSUP));
2918 		spa_async_request(spa, SPA_ASYNC_CONFIG_UPDATE);
2919 	}
2920 
2921 	if (wasoffline ||
2922 	    (oldstate < VDEV_STATE_DEGRADED &&
2923 	    vd->vdev_state >= VDEV_STATE_DEGRADED))
2924 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_ONLINE);
2925 
2926 	return (spa_vdev_state_exit(spa, vd, 0));
2927 }
2928 
2929 static int
2930 vdev_offline_locked(spa_t *spa, uint64_t guid, uint64_t flags)
2931 {
2932 	vdev_t *vd, *tvd;
2933 	int error = 0;
2934 	uint64_t generation;
2935 	metaslab_group_t *mg;
2936 
2937 top:
2938 	spa_vdev_state_enter(spa, SCL_ALLOC);
2939 
2940 	if ((vd = spa_lookup_by_guid(spa, guid, B_TRUE)) == NULL)
2941 		return (spa_vdev_state_exit(spa, NULL, ENODEV));
2942 
2943 	if (!vd->vdev_ops->vdev_op_leaf)
2944 		return (spa_vdev_state_exit(spa, NULL, ENOTSUP));
2945 
2946 	tvd = vd->vdev_top;
2947 	mg = tvd->vdev_mg;
2948 	generation = spa->spa_config_generation + 1;
2949 
2950 	/*
2951 	 * If the device isn't already offline, try to offline it.
2952 	 */
2953 	if (!vd->vdev_offline) {
2954 		/*
2955 		 * If this device has the only valid copy of some data,
2956 		 * don't allow it to be offlined. Log devices are always
2957 		 * expendable.
2958 		 */
2959 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
2960 		    vdev_dtl_required(vd))
2961 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
2962 
2963 		/*
2964 		 * If the top-level is a slog and it has had allocations
2965 		 * then proceed.  We check that the vdev's metaslab group
2966 		 * is not NULL since it's possible that we may have just
2967 		 * added this vdev but not yet initialized its metaslabs.
2968 		 */
2969 		if (tvd->vdev_islog && mg != NULL) {
2970 			/*
2971 			 * Prevent any future allocations.
2972 			 */
2973 			metaslab_group_passivate(mg);
2974 			(void) spa_vdev_state_exit(spa, vd, 0);
2975 
2976 			error = spa_reset_logs(spa);
2977 
2978 			spa_vdev_state_enter(spa, SCL_ALLOC);
2979 
2980 			/*
2981 			 * Check to see if the config has changed.
2982 			 */
2983 			if (error || generation != spa->spa_config_generation) {
2984 				metaslab_group_activate(mg);
2985 				if (error)
2986 					return (spa_vdev_state_exit(spa,
2987 					    vd, error));
2988 				(void) spa_vdev_state_exit(spa, vd, 0);
2989 				goto top;
2990 			}
2991 			ASSERT0(tvd->vdev_stat.vs_alloc);
2992 		}
2993 
2994 		/*
2995 		 * Offline this device and reopen its top-level vdev.
2996 		 * If the top-level vdev is a log device then just offline
2997 		 * it. Otherwise, if this action results in the top-level
2998 		 * vdev becoming unusable, undo it and fail the request.
2999 		 */
3000 		vd->vdev_offline = B_TRUE;
3001 		vdev_reopen(tvd);
3002 
3003 		if (!tvd->vdev_islog && vd->vdev_aux == NULL &&
3004 		    vdev_is_dead(tvd)) {
3005 			vd->vdev_offline = B_FALSE;
3006 			vdev_reopen(tvd);
3007 			return (spa_vdev_state_exit(spa, NULL, EBUSY));
3008 		}
3009 
3010 		/*
3011 		 * Add the device back into the metaslab rotor so that
3012 		 * once we online the device it's open for business.
3013 		 */
3014 		if (tvd->vdev_islog && mg != NULL)
3015 			metaslab_group_activate(mg);
3016 	}
3017 
3018 	vd->vdev_tmpoffline = !!(flags & ZFS_OFFLINE_TEMPORARY);
3019 
3020 	return (spa_vdev_state_exit(spa, vd, 0));
3021 }
3022 
3023 int
3024 vdev_offline(spa_t *spa, uint64_t guid, uint64_t flags)
3025 {
3026 	int error;
3027 
3028 	mutex_enter(&spa->spa_vdev_top_lock);
3029 	error = vdev_offline_locked(spa, guid, flags);
3030 	mutex_exit(&spa->spa_vdev_top_lock);
3031 
3032 	return (error);
3033 }
3034 
3035 /*
3036  * Clear the error counts associated with this vdev.  Unlike vdev_online() and
3037  * vdev_offline(), we assume the spa config is locked.  We also clear all
3038  * children.  If 'vd' is NULL, then the user wants to clear all vdevs.
3039  */
3040 void
3041 vdev_clear(spa_t *spa, vdev_t *vd)
3042 {
3043 	vdev_t *rvd = spa->spa_root_vdev;
3044 
3045 	ASSERT(spa_config_held(spa, SCL_STATE_ALL, RW_WRITER) == SCL_STATE_ALL);
3046 
3047 	if (vd == NULL)
3048 		vd = rvd;
3049 
3050 	vd->vdev_stat.vs_read_errors = 0;
3051 	vd->vdev_stat.vs_write_errors = 0;
3052 	vd->vdev_stat.vs_checksum_errors = 0;
3053 
3054 	for (int c = 0; c < vd->vdev_children; c++)
3055 		vdev_clear(spa, vd->vdev_child[c]);
3056 
3057 	/*
3058 	 * It makes no sense to "clear" an indirect vdev.
3059 	 */
3060 	if (!vdev_is_concrete(vd))
3061 		return;
3062 
3063 	/*
3064 	 * If we're in the FAULTED state or have experienced failed I/O, then
3065 	 * clear the persistent state and attempt to reopen the device.  We
3066 	 * also mark the vdev config dirty, so that the new faulted state is
3067 	 * written out to disk.
3068 	 */
3069 	if (vd->vdev_faulted || vd->vdev_degraded ||
3070 	    !vdev_readable(vd) || !vdev_writeable(vd)) {
3071 
3072 		/*
3073 		 * When reopening in reponse to a clear event, it may be due to
3074 		 * a fmadm repair request.  In this case, if the device is
3075 		 * still broken, we want to still post the ereport again.
3076 		 */
3077 		vd->vdev_forcefault = B_TRUE;
3078 
3079 		vd->vdev_faulted = vd->vdev_degraded = 0ULL;
3080 		vd->vdev_cant_read = B_FALSE;
3081 		vd->vdev_cant_write = B_FALSE;
3082 
3083 		vdev_reopen(vd == rvd ? rvd : vd->vdev_top);
3084 
3085 		vd->vdev_forcefault = B_FALSE;
3086 
3087 		if (vd != rvd && vdev_writeable(vd->vdev_top))
3088 			vdev_state_dirty(vd->vdev_top);
3089 
3090 		if (vd->vdev_aux == NULL && !vdev_is_dead(vd))
3091 			spa_async_request(spa, SPA_ASYNC_RESILVER);
3092 
3093 		spa_event_notify(spa, vd, NULL, ESC_ZFS_VDEV_CLEAR);
3094 	}
3095 
3096 	/*
3097 	 * When clearing a FMA-diagnosed fault, we always want to
3098 	 * unspare the device, as we assume that the original spare was
3099 	 * done in response to the FMA fault.
3100 	 */
3101 	if (!vdev_is_dead(vd) && vd->vdev_parent != NULL &&
3102 	    vd->vdev_parent->vdev_ops == &vdev_spare_ops &&
3103 	    vd->vdev_parent->vdev_child[0] == vd)
3104 		vd->vdev_unspare = B_TRUE;
3105 }
3106 
3107 boolean_t
3108 vdev_is_dead(vdev_t *vd)
3109 {
3110 	/*
3111 	 * Holes and missing devices are always considered "dead".
3112 	 * This simplifies the code since we don't have to check for
3113 	 * these types of devices in the various code paths.
3114 	 * Instead we rely on the fact that we skip over dead devices
3115 	 * before issuing I/O to them.
3116 	 */
3117 	return (vd->vdev_state < VDEV_STATE_DEGRADED ||
3118 	    vd->vdev_ops == &vdev_hole_ops ||
3119 	    vd->vdev_ops == &vdev_missing_ops);
3120 }
3121 
3122 boolean_t
3123 vdev_readable(vdev_t *vd)
3124 {
3125 	return (!vdev_is_dead(vd) && !vd->vdev_cant_read);
3126 }
3127 
3128 boolean_t
3129 vdev_writeable(vdev_t *vd)
3130 {
3131 	return (!vdev_is_dead(vd) && !vd->vdev_cant_write &&
3132 	    vdev_is_concrete(vd));
3133 }
3134 
3135 boolean_t
3136 vdev_allocatable(vdev_t *vd)
3137 {
3138 	uint64_t state = vd->vdev_state;
3139 
3140 	/*
3141 	 * We currently allow allocations from vdevs which may be in the
3142 	 * process of reopening (i.e. VDEV_STATE_CLOSED). If the device
3143 	 * fails to reopen then we'll catch it later when we're holding
3144 	 * the proper locks.  Note that we have to get the vdev state
3145 	 * in a local variable because although it changes atomically,
3146 	 * we're asking two separate questions about it.
3147 	 */
3148 	return (!(state < VDEV_STATE_DEGRADED && state != VDEV_STATE_CLOSED) &&
3149 	    !vd->vdev_cant_write && vdev_is_concrete(vd) &&
3150 	    vd->vdev_mg->mg_initialized);
3151 }
3152 
3153 boolean_t
3154 vdev_accessible(vdev_t *vd, zio_t *zio)
3155 {
3156 	ASSERT(zio->io_vd == vd);
3157 
3158 	if (vdev_is_dead(vd) || vd->vdev_remove_wanted)
3159 		return (B_FALSE);
3160 
3161 	if (zio->io_type == ZIO_TYPE_READ)
3162 		return (!vd->vdev_cant_read);
3163 
3164 	if (zio->io_type == ZIO_TYPE_WRITE)
3165 		return (!vd->vdev_cant_write);
3166 
3167 	return (B_TRUE);
3168 }
3169 
3170 /*
3171  * Get statistics for the given vdev.
3172  */
3173 void
3174 vdev_get_stats(vdev_t *vd, vdev_stat_t *vs)
3175 {
3176 	spa_t *spa = vd->vdev_spa;
3177 	vdev_t *rvd = spa->spa_root_vdev;
3178 	vdev_t *tvd = vd->vdev_top;
3179 
3180 	ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
3181 
3182 	mutex_enter(&vd->vdev_stat_lock);
3183 	bcopy(&vd->vdev_stat, vs, sizeof (*vs));
3184 	vs->vs_timestamp = gethrtime() - vs->vs_timestamp;
3185 	vs->vs_state = vd->vdev_state;
3186 	vs->vs_rsize = vdev_get_min_asize(vd);
3187 	if (vd->vdev_ops->vdev_op_leaf)
3188 		vs->vs_rsize += VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
3189 	/*
3190 	 * Report expandable space on top-level, non-auxillary devices only.
3191 	 * The expandable space is reported in terms of metaslab sized units
3192 	 * since that determines how much space the pool can expand.
3193 	 */
3194 	if (vd->vdev_aux == NULL && tvd != NULL) {
3195 		vs->vs_esize = P2ALIGN(vd->vdev_max_asize - vd->vdev_asize -
3196 		    spa->spa_bootsize, 1ULL << tvd->vdev_ms_shift);
3197 	}
3198 	if (vd->vdev_aux == NULL && vd == vd->vdev_top &&
3199 	    vdev_is_concrete(vd)) {
3200 		vs->vs_fragmentation = vd->vdev_mg->mg_fragmentation;
3201 	}
3202 
3203 	/*
3204 	 * If we're getting stats on the root vdev, aggregate the I/O counts
3205 	 * over all top-level vdevs (i.e. the direct children of the root).
3206 	 */
3207 	if (vd == rvd) {
3208 		for (int c = 0; c < rvd->vdev_children; c++) {
3209 			vdev_t *cvd = rvd->vdev_child[c];
3210 			vdev_stat_t *cvs = &cvd->vdev_stat;
3211 
3212 			for (int t = 0; t < ZIO_TYPES; t++) {
3213 				vs->vs_ops[t] += cvs->vs_ops[t];
3214 				vs->vs_bytes[t] += cvs->vs_bytes[t];
3215 			}
3216 			cvs->vs_scan_removing = cvd->vdev_removing;
3217 		}
3218 	}
3219 	mutex_exit(&vd->vdev_stat_lock);
3220 }
3221 
3222 void
3223 vdev_clear_stats(vdev_t *vd)
3224 {
3225 	mutex_enter(&vd->vdev_stat_lock);
3226 	vd->vdev_stat.vs_space = 0;
3227 	vd->vdev_stat.vs_dspace = 0;
3228 	vd->vdev_stat.vs_alloc = 0;
3229 	mutex_exit(&vd->vdev_stat_lock);
3230 }
3231 
3232 void
3233 vdev_scan_stat_init(vdev_t *vd)
3234 {
3235 	vdev_stat_t *vs = &vd->vdev_stat;
3236 
3237 	for (int c = 0; c < vd->vdev_children; c++)
3238 		vdev_scan_stat_init(vd->vdev_child[c]);
3239 
3240 	mutex_enter(&vd->vdev_stat_lock);
3241 	vs->vs_scan_processed = 0;
3242 	mutex_exit(&vd->vdev_stat_lock);
3243 }
3244 
3245 void
3246 vdev_stat_update(zio_t *zio, uint64_t psize)
3247 {
3248 	spa_t *spa = zio->io_spa;
3249 	vdev_t *rvd = spa->spa_root_vdev;
3250 	vdev_t *vd = zio->io_vd ? zio->io_vd : rvd;
3251 	vdev_t *pvd;
3252 	uint64_t txg = zio->io_txg;
3253 	vdev_stat_t *vs = &vd->vdev_stat;
3254 	zio_type_t type = zio->io_type;
3255 	int flags = zio->io_flags;
3256 
3257 	/*
3258 	 * If this i/o is a gang leader, it didn't do any actual work.
3259 	 */
3260 	if (zio->io_gang_tree)
3261 		return;
3262 
3263 	if (zio->io_error == 0) {
3264 		/*
3265 		 * If this is a root i/o, don't count it -- we've already
3266 		 * counted the top-level vdevs, and vdev_get_stats() will
3267 		 * aggregate them when asked.  This reduces contention on
3268 		 * the root vdev_stat_lock and implicitly handles blocks
3269 		 * that compress away to holes, for which there is no i/o.
3270 		 * (Holes never create vdev children, so all the counters
3271 		 * remain zero, which is what we want.)
3272 		 *
3273 		 * Note: this only applies to successful i/o (io_error == 0)
3274 		 * because unlike i/o counts, errors are not additive.
3275 		 * When reading a ditto block, for example, failure of
3276 		 * one top-level vdev does not imply a root-level error.
3277 		 */
3278 		if (vd == rvd)
3279 			return;
3280 
3281 		ASSERT(vd == zio->io_vd);
3282 
3283 		if (flags & ZIO_FLAG_IO_BYPASS)
3284 			return;
3285 
3286 		mutex_enter(&vd->vdev_stat_lock);
3287 
3288 		if (flags & ZIO_FLAG_IO_REPAIR) {
3289 			if (flags & ZIO_FLAG_SCAN_THREAD) {
3290 				dsl_scan_phys_t *scn_phys =
3291 				    &spa->spa_dsl_pool->dp_scan->scn_phys;
3292 				uint64_t *processed = &scn_phys->scn_processed;
3293 
3294 				/* XXX cleanup? */
3295 				if (vd->vdev_ops->vdev_op_leaf)
3296 					atomic_add_64(processed, psize);
3297 				vs->vs_scan_processed += psize;
3298 			}
3299 
3300 			if (flags & ZIO_FLAG_SELF_HEAL)
3301 				vs->vs_self_healed += psize;
3302 		}
3303 
3304 		vs->vs_ops[type]++;
3305 		vs->vs_bytes[type] += psize;
3306 
3307 		mutex_exit(&vd->vdev_stat_lock);
3308 		return;
3309 	}
3310 
3311 	if (flags & ZIO_FLAG_SPECULATIVE)
3312 		return;
3313 
3314 	/*
3315 	 * If this is an I/O error that is going to be retried, then ignore the
3316 	 * error.  Otherwise, the user may interpret B_FAILFAST I/O errors as
3317 	 * hard errors, when in reality they can happen for any number of
3318 	 * innocuous reasons (bus resets, MPxIO link failure, etc).
3319 	 */
3320 	if (zio->io_error == EIO &&
3321 	    !(zio->io_flags & ZIO_FLAG_IO_RETRY))
3322 		return;
3323 
3324 	/*
3325 	 * Intent logs writes won't propagate their error to the root
3326 	 * I/O so don't mark these types of failures as pool-level
3327 	 * errors.
3328 	 */
3329 	if (zio->io_vd == NULL && (zio->io_flags & ZIO_FLAG_DONT_PROPAGATE))
3330 		return;
3331 
3332 	mutex_enter(&vd->vdev_stat_lock);
3333 	if (type == ZIO_TYPE_READ && !vdev_is_dead(vd)) {
3334 		if (zio->io_error == ECKSUM)
3335 			vs->vs_checksum_errors++;
3336 		else
3337 			vs->vs_read_errors++;
3338 	}
3339 	if (type == ZIO_TYPE_WRITE && !vdev_is_dead(vd))
3340 		vs->vs_write_errors++;
3341 	mutex_exit(&vd->vdev_stat_lock);
3342 
3343 	if (spa->spa_load_state == SPA_LOAD_NONE &&
3344 	    type == ZIO_TYPE_WRITE && txg != 0 &&
3345 	    (!(flags & ZIO_FLAG_IO_REPAIR) ||
3346 	    (flags & ZIO_FLAG_SCAN_THREAD) ||
3347 	    spa->spa_claiming)) {
3348 		/*
3349 		 * This is either a normal write (not a repair), or it's
3350 		 * a repair induced by the scrub thread, or it's a repair
3351 		 * made by zil_claim() during spa_load() in the first txg.
3352 		 * In the normal case, we commit the DTL change in the same
3353 		 * txg as the block was born.  In the scrub-induced repair
3354 		 * case, we know that scrubs run in first-pass syncing context,
3355 		 * so we commit the DTL change in spa_syncing_txg(spa).
3356 		 * In the zil_claim() case, we commit in spa_first_txg(spa).
3357 		 *
3358 		 * We currently do not make DTL entries for failed spontaneous
3359 		 * self-healing writes triggered by normal (non-scrubbing)
3360 		 * reads, because we have no transactional context in which to
3361 		 * do so -- and it's not clear that it'd be desirable anyway.
3362 		 */
3363 		if (vd->vdev_ops->vdev_op_leaf) {
3364 			uint64_t commit_txg = txg;
3365 			if (flags & ZIO_FLAG_SCAN_THREAD) {
3366 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3367 				ASSERT(spa_sync_pass(spa) == 1);
3368 				vdev_dtl_dirty(vd, DTL_SCRUB, txg, 1);
3369 				commit_txg = spa_syncing_txg(spa);
3370 			} else if (spa->spa_claiming) {
3371 				ASSERT(flags & ZIO_FLAG_IO_REPAIR);
3372 				commit_txg = spa_first_txg(spa);
3373 			}
3374 			ASSERT(commit_txg >= spa_syncing_txg(spa));
3375 			if (vdev_dtl_contains(vd, DTL_MISSING, txg, 1))
3376 				return;
3377 			for (pvd = vd; pvd != rvd; pvd = pvd->vdev_parent)
3378 				vdev_dtl_dirty(pvd, DTL_PARTIAL, txg, 1);
3379 			vdev_dirty(vd->vdev_top, VDD_DTL, vd, commit_txg);
3380 		}
3381 		if (vd != rvd)
3382 			vdev_dtl_dirty(vd, DTL_MISSING, txg, 1);
3383 	}
3384 }
3385 
3386 /*
3387  * Update the in-core space usage stats for this vdev, its metaslab class,
3388  * and the root vdev.
3389  */
3390 void
3391 vdev_space_update(vdev_t *vd, int64_t alloc_delta, int64_t defer_delta,
3392     int64_t space_delta)
3393 {
3394 	int64_t dspace_delta = space_delta;
3395 	spa_t *spa = vd->vdev_spa;
3396 	vdev_t *rvd = spa->spa_root_vdev;
3397 	metaslab_group_t *mg = vd->vdev_mg;
3398 	metaslab_class_t *mc = mg ? mg->mg_class : NULL;
3399 
3400 	ASSERT(vd == vd->vdev_top);
3401 
3402 	/*
3403 	 * Apply the inverse of the psize-to-asize (ie. RAID-Z) space-expansion
3404 	 * factor.  We must calculate this here and not at the root vdev
3405 	 * because the root vdev's psize-to-asize is simply the max of its
3406 	 * childrens', thus not accurate enough for us.
3407 	 */
3408 	ASSERT((dspace_delta & (SPA_MINBLOCKSIZE-1)) == 0);
3409 	ASSERT(vd->vdev_deflate_ratio != 0 || vd->vdev_isl2cache);
3410 	dspace_delta = (dspace_delta >> SPA_MINBLOCKSHIFT) *
3411 	    vd->vdev_deflate_ratio;
3412 
3413 	mutex_enter(&vd->vdev_stat_lock);
3414 	vd->vdev_stat.vs_alloc += alloc_delta;
3415 	vd->vdev_stat.vs_space += space_delta;
3416 	vd->vdev_stat.vs_dspace += dspace_delta;
3417 	mutex_exit(&vd->vdev_stat_lock);
3418 
3419 	if (mc == spa_normal_class(spa)) {
3420 		mutex_enter(&rvd->vdev_stat_lock);
3421 		rvd->vdev_stat.vs_alloc += alloc_delta;
3422 		rvd->vdev_stat.vs_space += space_delta;
3423 		rvd->vdev_stat.vs_dspace += dspace_delta;
3424 		mutex_exit(&rvd->vdev_stat_lock);
3425 	}
3426 
3427 	if (mc != NULL) {
3428 		ASSERT(rvd == vd->vdev_parent);
3429 		ASSERT(vd->vdev_ms_count != 0);
3430 
3431 		metaslab_class_space_update(mc,
3432 		    alloc_delta, defer_delta, space_delta, dspace_delta);
3433 	}
3434 }
3435 
3436 /*
3437  * Mark a top-level vdev's config as dirty, placing it on the dirty list
3438  * so that it will be written out next time the vdev configuration is synced.
3439  * If the root vdev is specified (vdev_top == NULL), dirty all top-level vdevs.
3440  */
3441 void
3442 vdev_config_dirty(vdev_t *vd)
3443 {
3444 	spa_t *spa = vd->vdev_spa;
3445 	vdev_t *rvd = spa->spa_root_vdev;
3446 	int c;
3447 
3448 	ASSERT(spa_writeable(spa));
3449 
3450 	/*
3451 	 * If this is an aux vdev (as with l2cache and spare devices), then we
3452 	 * update the vdev config manually and set the sync flag.
3453 	 */
3454 	if (vd->vdev_aux != NULL) {
3455 		spa_aux_vdev_t *sav = vd->vdev_aux;
3456 		nvlist_t **aux;
3457 		uint_t naux;
3458 
3459 		for (c = 0; c < sav->sav_count; c++) {
3460 			if (sav->sav_vdevs[c] == vd)
3461 				break;
3462 		}
3463 
3464 		if (c == sav->sav_count) {
3465 			/*
3466 			 * We're being removed.  There's nothing more to do.
3467 			 */
3468 			ASSERT(sav->sav_sync == B_TRUE);
3469 			return;
3470 		}
3471 
3472 		sav->sav_sync = B_TRUE;
3473 
3474 		if (nvlist_lookup_nvlist_array(sav->sav_config,
3475 		    ZPOOL_CONFIG_L2CACHE, &aux, &naux) != 0) {
3476 			VERIFY(nvlist_lookup_nvlist_array(sav->sav_config,
3477 			    ZPOOL_CONFIG_SPARES, &aux, &naux) == 0);
3478 		}
3479 
3480 		ASSERT(c < naux);
3481 
3482 		/*
3483 		 * Setting the nvlist in the middle if the array is a little
3484 		 * sketchy, but it will work.
3485 		 */
3486 		nvlist_free(aux[c]);
3487 		aux[c] = vdev_config_generate(spa, vd, B_TRUE, 0);
3488 
3489 		return;
3490 	}
3491 
3492 	/*
3493 	 * The dirty list is protected by the SCL_CONFIG lock.  The caller
3494 	 * must either hold SCL_CONFIG as writer, or must be the sync thread
3495 	 * (which holds SCL_CONFIG as reader).  There's only one sync thread,
3496 	 * so this is sufficient to ensure mutual exclusion.
3497 	 */
3498 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3499 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3500 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3501 
3502 	if (vd == rvd) {
3503 		for (c = 0; c < rvd->vdev_children; c++)
3504 			vdev_config_dirty(rvd->vdev_child[c]);
3505 	} else {
3506 		ASSERT(vd == vd->vdev_top);
3507 
3508 		if (!list_link_active(&vd->vdev_config_dirty_node) &&
3509 		    vdev_is_concrete(vd)) {
3510 			list_insert_head(&spa->spa_config_dirty_list, vd);
3511 		}
3512 	}
3513 }
3514 
3515 void
3516 vdev_config_clean(vdev_t *vd)
3517 {
3518 	spa_t *spa = vd->vdev_spa;
3519 
3520 	ASSERT(spa_config_held(spa, SCL_CONFIG, RW_WRITER) ||
3521 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3522 	    spa_config_held(spa, SCL_CONFIG, RW_READER)));
3523 
3524 	ASSERT(list_link_active(&vd->vdev_config_dirty_node));
3525 	list_remove(&spa->spa_config_dirty_list, vd);
3526 }
3527 
3528 /*
3529  * Mark a top-level vdev's state as dirty, so that the next pass of
3530  * spa_sync() can convert this into vdev_config_dirty().  We distinguish
3531  * the state changes from larger config changes because they require
3532  * much less locking, and are often needed for administrative actions.
3533  */
3534 void
3535 vdev_state_dirty(vdev_t *vd)
3536 {
3537 	spa_t *spa = vd->vdev_spa;
3538 
3539 	ASSERT(spa_writeable(spa));
3540 	ASSERT(vd == vd->vdev_top);
3541 
3542 	/*
3543 	 * The state list is protected by the SCL_STATE lock.  The caller
3544 	 * must either hold SCL_STATE as writer, or must be the sync thread
3545 	 * (which holds SCL_STATE as reader).  There's only one sync thread,
3546 	 * so this is sufficient to ensure mutual exclusion.
3547 	 */
3548 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3549 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3550 	    spa_config_held(spa, SCL_STATE, RW_READER)));
3551 
3552 	if (!list_link_active(&vd->vdev_state_dirty_node) &&
3553 	    vdev_is_concrete(vd))
3554 		list_insert_head(&spa->spa_state_dirty_list, vd);
3555 }
3556 
3557 void
3558 vdev_state_clean(vdev_t *vd)
3559 {
3560 	spa_t *spa = vd->vdev_spa;
3561 
3562 	ASSERT(spa_config_held(spa, SCL_STATE, RW_WRITER) ||
3563 	    (dsl_pool_sync_context(spa_get_dsl(spa)) &&
3564 	    spa_config_held(spa, SCL_STATE, RW_READER)));
3565 
3566 	ASSERT(list_link_active(&vd->vdev_state_dirty_node));
3567 	list_remove(&spa->spa_state_dirty_list, vd);
3568 }
3569 
3570 /*
3571  * Propagate vdev state up from children to parent.
3572  */
3573 void
3574 vdev_propagate_state(vdev_t *vd)
3575 {
3576 	spa_t *spa = vd->vdev_spa;
3577 	vdev_t *rvd = spa->spa_root_vdev;
3578 	int degraded = 0, faulted = 0;
3579 	int corrupted = 0;
3580 	vdev_t *child;
3581 
3582 	if (vd->vdev_children > 0) {
3583 		for (int c = 0; c < vd->vdev_children; c++) {
3584 			child = vd->vdev_child[c];
3585 
3586 			/*
3587 			 * Don't factor holes or indirect vdevs into the
3588 			 * decision.
3589 			 */
3590 			if (!vdev_is_concrete(child))
3591 				continue;
3592 
3593 			if (!vdev_readable(child) ||
3594 			    (!vdev_writeable(child) && spa_writeable(spa))) {
3595 				/*
3596 				 * Root special: if there is a top-level log
3597 				 * device, treat the root vdev as if it were
3598 				 * degraded.
3599 				 */
3600 				if (child->vdev_islog && vd == rvd)
3601 					degraded++;
3602 				else
3603 					faulted++;
3604 			} else if (child->vdev_state <= VDEV_STATE_DEGRADED) {
3605 				degraded++;
3606 			}
3607 
3608 			if (child->vdev_stat.vs_aux == VDEV_AUX_CORRUPT_DATA)
3609 				corrupted++;
3610 		}
3611 
3612 		vd->vdev_ops->vdev_op_state_change(vd, faulted, degraded);
3613 
3614 		/*
3615 		 * Root special: if there is a top-level vdev that cannot be
3616 		 * opened due to corrupted metadata, then propagate the root
3617 		 * vdev's aux state as 'corrupt' rather than 'insufficient
3618 		 * replicas'.
3619 		 */
3620 		if (corrupted && vd == rvd &&
3621 		    rvd->vdev_state == VDEV_STATE_CANT_OPEN)
3622 			vdev_set_state(rvd, B_FALSE, VDEV_STATE_CANT_OPEN,
3623 			    VDEV_AUX_CORRUPT_DATA);
3624 	}
3625 
3626 	if (vd->vdev_parent)
3627 		vdev_propagate_state(vd->vdev_parent);
3628 }
3629 
3630 /*
3631  * Set a vdev's state.  If this is during an open, we don't update the parent
3632  * state, because we're in the process of opening children depth-first.
3633  * Otherwise, we propagate the change to the parent.
3634  *
3635  * If this routine places a device in a faulted state, an appropriate ereport is
3636  * generated.
3637  */
3638 void
3639 vdev_set_state(vdev_t *vd, boolean_t isopen, vdev_state_t state, vdev_aux_t aux)
3640 {
3641 	uint64_t save_state;
3642 	spa_t *spa = vd->vdev_spa;
3643 
3644 	if (state == vd->vdev_state) {
3645 		vd->vdev_stat.vs_aux = aux;
3646 		return;
3647 	}
3648 
3649 	save_state = vd->vdev_state;
3650 
3651 	vd->vdev_state = state;
3652 	vd->vdev_stat.vs_aux = aux;
3653 
3654 	/*
3655 	 * If we are setting the vdev state to anything but an open state, then
3656 	 * always close the underlying device unless the device has requested
3657 	 * a delayed close (i.e. we're about to remove or fault the device).
3658 	 * Otherwise, we keep accessible but invalid devices open forever.
3659 	 * We don't call vdev_close() itself, because that implies some extra
3660 	 * checks (offline, etc) that we don't want here.  This is limited to
3661 	 * leaf devices, because otherwise closing the device will affect other
3662 	 * children.
3663 	 */
3664 	if (!vd->vdev_delayed_close && vdev_is_dead(vd) &&
3665 	    vd->vdev_ops->vdev_op_leaf)
3666 		vd->vdev_ops->vdev_op_close(vd);
3667 
3668 	/*
3669 	 * If we have brought this vdev back into service, we need
3670 	 * to notify fmd so that it can gracefully repair any outstanding
3671 	 * cases due to a missing device.  We do this in all cases, even those
3672 	 * that probably don't correlate to a repaired fault.  This is sure to
3673 	 * catch all cases, and we let the zfs-retire agent sort it out.  If
3674 	 * this is a transient state it's OK, as the retire agent will
3675 	 * double-check the state of the vdev before repairing it.
3676 	 */
3677 	if (state == VDEV_STATE_HEALTHY && vd->vdev_ops->vdev_op_leaf &&
3678 	    vd->vdev_prevstate != state)
3679 		zfs_post_state_change(spa, vd);
3680 
3681 	if (vd->vdev_removed &&
3682 	    state == VDEV_STATE_CANT_OPEN &&
3683 	    (aux == VDEV_AUX_OPEN_FAILED || vd->vdev_checkremove)) {
3684 		/*
3685 		 * If the previous state is set to VDEV_STATE_REMOVED, then this
3686 		 * device was previously marked removed and someone attempted to
3687 		 * reopen it.  If this failed due to a nonexistent device, then
3688 		 * keep the device in the REMOVED state.  We also let this be if
3689 		 * it is one of our special test online cases, which is only
3690 		 * attempting to online the device and shouldn't generate an FMA
3691 		 * fault.
3692 		 */
3693 		vd->vdev_state = VDEV_STATE_REMOVED;
3694 		vd->vdev_stat.vs_aux = VDEV_AUX_NONE;
3695 	} else if (state == VDEV_STATE_REMOVED) {
3696 		vd->vdev_removed = B_TRUE;
3697 	} else if (state == VDEV_STATE_CANT_OPEN) {
3698 		/*
3699 		 * If we fail to open a vdev during an import or recovery, we
3700 		 * mark it as "not available", which signifies that it was
3701 		 * never there to begin with.  Failure to open such a device
3702 		 * is not considered an error.
3703 		 */
3704 		if ((spa_load_state(spa) == SPA_LOAD_IMPORT ||
3705 		    spa_load_state(spa) == SPA_LOAD_RECOVER) &&
3706 		    vd->vdev_ops->vdev_op_leaf)
3707 			vd->vdev_not_present = 1;
3708 
3709 		/*
3710 		 * Post the appropriate ereport.  If the 'prevstate' field is
3711 		 * set to something other than VDEV_STATE_UNKNOWN, it indicates
3712 		 * that this is part of a vdev_reopen().  In this case, we don't
3713 		 * want to post the ereport if the device was already in the
3714 		 * CANT_OPEN state beforehand.
3715 		 *
3716 		 * If the 'checkremove' flag is set, then this is an attempt to
3717 		 * online the device in response to an insertion event.  If we
3718 		 * hit this case, then we have detected an insertion event for a
3719 		 * faulted or offline device that wasn't in the removed state.
3720 		 * In this scenario, we don't post an ereport because we are
3721 		 * about to replace the device, or attempt an online with
3722 		 * vdev_forcefault, which will generate the fault for us.
3723 		 */
3724 		if ((vd->vdev_prevstate != state || vd->vdev_forcefault) &&
3725 		    !vd->vdev_not_present && !vd->vdev_checkremove &&
3726 		    vd != spa->spa_root_vdev) {
3727 			const char *class;
3728 
3729 			switch (aux) {
3730 			case VDEV_AUX_OPEN_FAILED:
3731 				class = FM_EREPORT_ZFS_DEVICE_OPEN_FAILED;
3732 				break;
3733 			case VDEV_AUX_CORRUPT_DATA:
3734 				class = FM_EREPORT_ZFS_DEVICE_CORRUPT_DATA;
3735 				break;
3736 			case VDEV_AUX_NO_REPLICAS:
3737 				class = FM_EREPORT_ZFS_DEVICE_NO_REPLICAS;
3738 				break;
3739 			case VDEV_AUX_BAD_GUID_SUM:
3740 				class = FM_EREPORT_ZFS_DEVICE_BAD_GUID_SUM;
3741 				break;
3742 			case VDEV_AUX_TOO_SMALL:
3743 				class = FM_EREPORT_ZFS_DEVICE_TOO_SMALL;
3744 				break;
3745 			case VDEV_AUX_BAD_LABEL:
3746 				class = FM_EREPORT_ZFS_DEVICE_BAD_LABEL;
3747 				break;
3748 			default:
3749 				class = FM_EREPORT_ZFS_DEVICE_UNKNOWN;
3750 			}
3751 
3752 			zfs_ereport_post(class, spa, vd, NULL, save_state, 0);
3753 		}
3754 
3755 		/* Erase any notion of persistent removed state */
3756 		vd->vdev_removed = B_FALSE;
3757 	} else {
3758 		vd->vdev_removed = B_FALSE;
3759 	}
3760 
3761 	if (!isopen && vd->vdev_parent)
3762 		vdev_propagate_state(vd->vdev_parent);
3763 }
3764 
3765 boolean_t
3766 vdev_children_are_offline(vdev_t *vd)
3767 {
3768 	ASSERT(!vd->vdev_ops->vdev_op_leaf);
3769 
3770 	for (uint64_t i = 0; i < vd->vdev_children; i++) {
3771 		if (vd->vdev_child[i]->vdev_state != VDEV_STATE_OFFLINE)
3772 			return (B_FALSE);
3773 	}
3774 
3775 	return (B_TRUE);
3776 }
3777 
3778 /*
3779  * Check the vdev configuration to ensure that it's capable of supporting
3780  * a root pool. We do not support partial configuration.
3781  * In addition, only a single top-level vdev is allowed.
3782  */
3783 boolean_t
3784 vdev_is_bootable(vdev_t *vd)
3785 {
3786 	if (!vd->vdev_ops->vdev_op_leaf) {
3787 		char *vdev_type = vd->vdev_ops->vdev_op_type;
3788 
3789 		if (strcmp(vdev_type, VDEV_TYPE_ROOT) == 0 &&
3790 		    vd->vdev_children > 1) {
3791 			return (B_FALSE);
3792 		} else if (strcmp(vdev_type, VDEV_TYPE_MISSING) == 0 ||
3793 		    strcmp(vdev_type, VDEV_TYPE_INDIRECT) == 0) {
3794 			return (B_FALSE);
3795 		}
3796 	}
3797 
3798 	for (int c = 0; c < vd->vdev_children; c++) {
3799 		if (!vdev_is_bootable(vd->vdev_child[c]))
3800 			return (B_FALSE);
3801 	}
3802 	return (B_TRUE);
3803 }
3804 
3805 boolean_t
3806 vdev_is_concrete(vdev_t *vd)
3807 {
3808 	vdev_ops_t *ops = vd->vdev_ops;
3809 	if (ops == &vdev_indirect_ops || ops == &vdev_hole_ops ||
3810 	    ops == &vdev_missing_ops || ops == &vdev_root_ops) {
3811 		return (B_FALSE);
3812 	} else {
3813 		return (B_TRUE);
3814 	}
3815 }
3816 
3817 /*
3818  * Determine if a log device has valid content.  If the vdev was
3819  * removed or faulted in the MOS config then we know that
3820  * the content on the log device has already been written to the pool.
3821  */
3822 boolean_t
3823 vdev_log_state_valid(vdev_t *vd)
3824 {
3825 	if (vd->vdev_ops->vdev_op_leaf && !vd->vdev_faulted &&
3826 	    !vd->vdev_removed)
3827 		return (B_TRUE);
3828 
3829 	for (int c = 0; c < vd->vdev_children; c++)
3830 		if (vdev_log_state_valid(vd->vdev_child[c]))
3831 			return (B_TRUE);
3832 
3833 	return (B_FALSE);
3834 }
3835 
3836 /*
3837  * Expand a vdev if possible.
3838  */
3839 void
3840 vdev_expand(vdev_t *vd, uint64_t txg)
3841 {
3842 	ASSERT(vd->vdev_top == vd);
3843 	ASSERT(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER) == SCL_ALL);
3844 
3845 	vdev_set_deflate_ratio(vd);
3846 
3847 	if ((vd->vdev_asize >> vd->vdev_ms_shift) > vd->vdev_ms_count &&
3848 	    vdev_is_concrete(vd)) {
3849 		VERIFY(vdev_metaslab_init(vd, txg) == 0);
3850 		vdev_config_dirty(vd);
3851 	}
3852 }
3853 
3854 /*
3855  * Split a vdev.
3856  */
3857 void
3858 vdev_split(vdev_t *vd)
3859 {
3860 	vdev_t *cvd, *pvd = vd->vdev_parent;
3861 
3862 	vdev_remove_child(pvd, vd);
3863 	vdev_compact_children(pvd);
3864 
3865 	cvd = pvd->vdev_child[0];
3866 	if (pvd->vdev_children == 1) {
3867 		vdev_remove_parent(cvd);
3868 		cvd->vdev_splitting = B_TRUE;
3869 	}
3870 	vdev_propagate_state(cvd);
3871 }
3872 
3873 void
3874 vdev_deadman(vdev_t *vd)
3875 {
3876 	for (int c = 0; c < vd->vdev_children; c++) {
3877 		vdev_t *cvd = vd->vdev_child[c];
3878 
3879 		vdev_deadman(cvd);
3880 	}
3881 
3882 	if (vd->vdev_ops->vdev_op_leaf) {
3883 		vdev_queue_t *vq = &vd->vdev_queue;
3884 
3885 		mutex_enter(&vq->vq_lock);
3886 		if (avl_numnodes(&vq->vq_active_tree) > 0) {
3887 			spa_t *spa = vd->vdev_spa;
3888 			zio_t *fio;
3889 			uint64_t delta;
3890 
3891 			/*
3892 			 * Look at the head of all the pending queues,
3893 			 * if any I/O has been outstanding for longer than
3894 			 * the spa_deadman_synctime we panic the system.
3895 			 */
3896 			fio = avl_first(&vq->vq_active_tree);
3897 			delta = gethrtime() - fio->io_timestamp;
3898 			if (delta > spa_deadman_synctime(spa)) {
3899 				vdev_dbgmsg(vd, "SLOW IO: zio timestamp "
3900 				    "%lluns, delta %lluns, last io %lluns",
3901 				    fio->io_timestamp, (u_longlong_t)delta,
3902 				    vq->vq_io_complete_ts);
3903 				fm_panic("I/O to pool '%s' appears to be "
3904 				    "hung.", spa_name(spa));
3905 			}
3906 		}
3907 		mutex_exit(&vq->vq_lock);
3908 	}
3909 }
3910